1 /*
2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 // no precompiled headers
26 #include "classfile/classLoader.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "code/icBuffer.hpp"
30 #include "code/vtableStubs.hpp"
31 #include "compiler/compileBroker.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "jvm_linux.h"
35 #include "memory/allocation.inline.hpp"
36 #include "memory/filemap.hpp"
37 #include "mutex_linux.inline.hpp"
38 #include "oops/oop.inline.hpp"
39 #include "os_share_linux.hpp"
40 #include "prims/jniFastGetField.hpp"
41 #include "prims/jvm.h"
42 #include "prims/jvm_misc.hpp"
43 #include "runtime/arguments.hpp"
44 #include "runtime/extendedPC.hpp"
45 #include "runtime/globals.hpp"
46 #include "runtime/interfaceSupport.hpp"
47 #include "runtime/init.hpp"
48 #include "runtime/java.hpp"
49 #include "runtime/javaCalls.hpp"
50 #include "runtime/mutexLocker.hpp"
51 #include "runtime/objectMonitor.hpp"
52 #include "runtime/orderAccess.inline.hpp"
53 #include "runtime/osThread.hpp"
54 #include "runtime/perfMemory.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #include "runtime/statSampler.hpp"
57 #include "runtime/stubRoutines.hpp"
58 #include "runtime/thread.inline.hpp"
59 #include "runtime/threadCritical.hpp"
60 #include "runtime/timer.hpp"
61 #include "services/attachListener.hpp"
62 #include "services/memTracker.hpp"
63 #include "services/runtimeService.hpp"
64 #include "utilities/decoder.hpp"
65 #include "utilities/defaultStream.hpp"
66 #include "utilities/events.hpp"
67 #include "utilities/elfFile.hpp"
68 #include "utilities/growableArray.hpp"
69 #include "utilities/vmError.hpp"
70
71 // put OS-includes here
72 # include <sys/types.h>
73 # include <sys/mman.h>
74 # include <sys/stat.h>
75 # include <sys/select.h>
76 # include <pthread.h>
77 # include <signal.h>
78 # include <errno.h>
79 # include <dlfcn.h>
80 # include <stdio.h>
81 # include <unistd.h>
82 # include <sys/resource.h>
83 # include <pthread.h>
84 # include <sys/stat.h>
85 # include <sys/time.h>
86 # include <sys/times.h>
87 # include <sys/utsname.h>
88 # include <sys/socket.h>
89 # include <sys/wait.h>
90 # include <pwd.h>
91 # include <poll.h>
92 # include <semaphore.h>
93 # include <fcntl.h>
94 # include <string.h>
95 # include <syscall.h>
96 # include <sys/sysinfo.h>
97 # include <gnu/libc-version.h>
98 # include <sys/ipc.h>
99 # include <sys/shm.h>
100 # include <link.h>
101 # include <stdint.h>
102 # include <inttypes.h>
103 # include <sys/ioctl.h>
104
105 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
106
107 #ifndef _GNU_SOURCE
108 #define _GNU_SOURCE
109 #include <sched.h>
110 #undef _GNU_SOURCE
111 #else
112 #include <sched.h>
113 #endif
114
115 // if RUSAGE_THREAD for getrusage() has not been defined, do it here. The code calling
116 // getrusage() is prepared to handle the associated failure.
117 #ifndef RUSAGE_THREAD
118 #define RUSAGE_THREAD (1) /* only the calling thread */
119 #endif
120
121 #define MAX_PATH (2 * K)
122
123 #define MAX_SECS 100000000
124
125 // for timer info max values which include all bits
126 #define ALL_64_BITS CONST64(0xFFFFFFFFFFFFFFFF)
127
128 #define LARGEPAGES_BIT (1 << 6)
129 ////////////////////////////////////////////////////////////////////////////////
130 // global variables
131 julong os::Linux::_physical_memory = 0;
132
133 address os::Linux::_initial_thread_stack_bottom = NULL;
134 uintptr_t os::Linux::_initial_thread_stack_size = 0;
135
136 int (*os::Linux::_clock_gettime)(clockid_t, struct timespec *) = NULL;
137 int (*os::Linux::_pthread_getcpuclockid)(pthread_t, clockid_t *) = NULL;
138 Mutex* os::Linux::_createThread_lock = NULL;
139 pthread_t os::Linux::_main_thread;
140 int os::Linux::_page_size = -1;
141 const int os::Linux::_vm_default_page_size = (8 * K);
142 bool os::Linux::_is_floating_stack = false;
143 bool os::Linux::_is_NPTL = false;
144 bool os::Linux::_supports_fast_thread_cpu_time = false;
145 const char * os::Linux::_glibc_version = NULL;
146 const char * os::Linux::_libpthread_version = NULL;
147 pthread_condattr_t os::Linux::_condattr[1];
148
149 static jlong initial_time_count=0;
150
151 static int clock_tics_per_sec = 100;
152
153 // For diagnostics to print a message once. see run_periodic_checks
154 static sigset_t check_signal_done;
155 static bool check_signals = true;
156
157 static pid_t _initial_pid = 0;
158
159 /* Signal number used to suspend/resume a thread */
160
161 /* do not use any signal number less than SIGSEGV, see 4355769 */
162 static int SR_signum = SIGUSR2;
163 sigset_t SR_sigset;
164
165 /* Used to protect dlsym() calls */
166 static pthread_mutex_t dl_mutex;
167
168 // Declarations
169 static void unpackTime(timespec* absTime, bool isAbsolute, jlong time);
170
171 // utility functions
172
173 static int SR_initialize();
174
175 julong os::available_memory() {
176 return Linux::available_memory();
177 }
178
179 julong os::Linux::available_memory() {
180 // values in struct sysinfo are "unsigned long"
181 struct sysinfo si;
182 sysinfo(&si);
183
184 return (julong)si.freeram * si.mem_unit;
185 }
186
187 julong os::physical_memory() {
188 return Linux::physical_memory();
189 }
190
191 ////////////////////////////////////////////////////////////////////////////////
192 // environment support
193
194 bool os::getenv(const char* name, char* buf, int len) {
195 const char* val = ::getenv(name);
196 if (val != NULL && strlen(val) < (size_t)len) {
197 strcpy(buf, val);
198 return true;
199 }
200 if (len > 0) buf[0] = 0; // return a null string
201 return false;
202 }
203
204
205 // Return true if user is running as root.
206
207 bool os::have_special_privileges() {
208 static bool init = false;
209 static bool privileges = false;
210 if (!init) {
211 privileges = (getuid() != geteuid()) || (getgid() != getegid());
212 init = true;
213 }
214 return privileges;
215 }
216
217
218 #ifndef SYS_gettid
219 // i386: 224, ia64: 1105, amd64: 186, sparc 143
220 #ifdef __ia64__
221 #define SYS_gettid 1105
222 #else
223 #ifdef __i386__
224 #define SYS_gettid 224
225 #else
226 #ifdef __amd64__
227 #define SYS_gettid 186
228 #else
229 #ifdef __sparc__
230 #define SYS_gettid 143
231 #else
232 #error define gettid for the arch
233 #endif
234 #endif
235 #endif
236 #endif
237 #endif
238
239 // Cpu architecture string
240 static char cpu_arch[] = HOTSPOT_LIB_ARCH;
241
242 // pid_t gettid()
243 //
244 // Returns the kernel thread id of the currently running thread. Kernel
245 // thread id is used to access /proc.
246 //
247 // (Note that getpid() on LinuxThreads returns kernel thread id too; but
248 // on NPTL, it returns the same pid for all threads, as required by POSIX.)
249 //
250 pid_t os::Linux::gettid() {
251 int rslt = syscall(SYS_gettid);
252 if (rslt == -1) {
253 // old kernel, no NPTL support
254 return getpid();
255 } else {
256 return (pid_t)rslt;
257 }
258 }
259
260 // Most versions of linux have a bug where the number of processors are
261 // determined by looking at the /proc file system. In a chroot environment,
262 // the system call returns 1. This causes the VM to act as if it is
263 // a single processor and elide locking (see is_MP() call).
264 static bool unsafe_chroot_detected = false;
265 static const char *unstable_chroot_error = "/proc file system not found.\n"
266 "Java may be unstable running multithreaded in a chroot "
267 "environment on Linux when /proc filesystem is not mounted.";
268
269 void os::Linux::initialize_system_info() {
270 set_processor_count(sysconf(_SC_NPROCESSORS_CONF));
271 if (processor_count() == 1) {
272 pid_t pid = os::Linux::gettid();
273 char fname[32];
274 jio_snprintf(fname, sizeof(fname), "/proc/%d", pid);
275 FILE *fp = fopen(fname, "r");
276 if (fp == NULL) {
277 unsafe_chroot_detected = true;
278 } else {
279 fclose(fp);
280 }
281 }
282 _physical_memory = (julong)sysconf(_SC_PHYS_PAGES) * (julong)sysconf(_SC_PAGESIZE);
283 assert(processor_count() > 0, "linux error");
284 }
285
286 void os::init_system_properties_values() {
287 // The next steps are taken in the product version:
288 //
289 // Obtain the JAVA_HOME value from the location of libjvm.so.
290 // This library should be located at:
291 // <JAVA_HOME>/jre/lib/<arch>/{client|server}/libjvm.so.
292 //
293 // If "/jre/lib/" appears at the right place in the path, then we
294 // assume libjvm.so is installed in a JDK and we use this path.
295 //
296 // Otherwise exit with message: "Could not create the Java virtual machine."
297 //
298 // The following extra steps are taken in the debugging version:
299 //
300 // If "/jre/lib/" does NOT appear at the right place in the path
301 // instead of exit check for $JAVA_HOME environment variable.
302 //
303 // If it is defined and we are able to locate $JAVA_HOME/jre/lib/<arch>,
304 // then we append a fake suffix "hotspot/libjvm.so" to this path so
305 // it looks like libjvm.so is installed there
306 // <JAVA_HOME>/jre/lib/<arch>/hotspot/libjvm.so.
307 //
308 // Otherwise exit.
309 //
310 // Important note: if the location of libjvm.so changes this
311 // code needs to be changed accordingly.
312
313 // See ld(1):
314 // The linker uses the following search paths to locate required
315 // shared libraries:
316 // 1: ...
317 // ...
318 // 7: The default directories, normally /lib and /usr/lib.
319 #if defined(AMD64) || defined(_LP64) && (defined(SPARC) || defined(PPC) || defined(S390))
320 #define DEFAULT_LIBPATH "/usr/lib64:/lib64:/lib:/usr/lib"
321 #else
322 #define DEFAULT_LIBPATH "/lib:/usr/lib"
323 #endif
324
325 // Base path of extensions installed on the system.
326 #define SYS_EXT_DIR "/usr/java/packages"
327 #define EXTENSIONS_DIR "/lib/ext"
328 #define ENDORSED_DIR "/lib/endorsed"
329
330 // Buffer that fits several sprintfs.
331 // Note that the space for the colon and the trailing null are provided
332 // by the nulls included by the sizeof operator.
333 const size_t bufsize =
334 MAX3((size_t)MAXPATHLEN, // For dll_dir & friends.
335 (size_t)MAXPATHLEN + sizeof(EXTENSIONS_DIR) + sizeof(SYS_EXT_DIR) + sizeof(EXTENSIONS_DIR), // extensions dir
336 (size_t)MAXPATHLEN + sizeof(ENDORSED_DIR)); // endorsed dir
337 char *buf = (char *)NEW_C_HEAP_ARRAY(char, bufsize, mtInternal);
338
339 // sysclasspath, java_home, dll_dir
340 {
341 char *pslash;
342 os::jvm_path(buf, bufsize);
343
344 // Found the full path to libjvm.so.
345 // Now cut the path to <java_home>/jre if we can.
346 *(strrchr(buf, '/')) = '\0'; // Get rid of /libjvm.so.
347 pslash = strrchr(buf, '/');
348 if (pslash != NULL) {
349 *pslash = '\0'; // Get rid of /{client|server|hotspot}.
350 }
351 Arguments::set_dll_dir(buf);
352
353 if (pslash != NULL) {
354 pslash = strrchr(buf, '/');
355 if (pslash != NULL) {
356 *pslash = '\0'; // Get rid of /<arch>.
357 pslash = strrchr(buf, '/');
358 if (pslash != NULL) {
359 *pslash = '\0'; // Get rid of /lib.
360 }
361 }
362 }
363 Arguments::set_java_home(buf);
364 set_boot_path('/', ':');
365 }
366
367 // Where to look for native libraries.
368 //
369 // Note: Due to a legacy implementation, most of the library path
370 // is set in the launcher. This was to accomodate linking restrictions
371 // on legacy Linux implementations (which are no longer supported).
372 // Eventually, all the library path setting will be done here.
373 //
374 // However, to prevent the proliferation of improperly built native
375 // libraries, the new path component /usr/java/packages is added here.
376 // Eventually, all the library path setting will be done here.
377 {
378 // Get the user setting of LD_LIBRARY_PATH, and prepended it. It
379 // should always exist (until the legacy problem cited above is
380 // addressed).
381 const char *v = ::getenv("LD_LIBRARY_PATH");
382 const char *v_colon = ":";
383 if (v == NULL) { v = ""; v_colon = ""; }
384 // That's +1 for the colon and +1 for the trailing '\0'.
385 char *ld_library_path = (char *)NEW_C_HEAP_ARRAY(char,
386 strlen(v) + 1 +
387 sizeof(SYS_EXT_DIR) + sizeof("/lib/") + strlen(cpu_arch) + sizeof(DEFAULT_LIBPATH) + 1,
388 mtInternal);
389 sprintf(ld_library_path, "%s%s" SYS_EXT_DIR "/lib/%s:" DEFAULT_LIBPATH, v, v_colon, cpu_arch);
390 Arguments::set_library_path(ld_library_path);
391 FREE_C_HEAP_ARRAY(char, ld_library_path, mtInternal);
392 }
393
394 // Extensions directories.
395 sprintf(buf, "%s" EXTENSIONS_DIR ":" SYS_EXT_DIR EXTENSIONS_DIR, Arguments::get_java_home());
396 Arguments::set_ext_dirs(buf);
397
398 // Endorsed standards default directory.
399 sprintf(buf, "%s" ENDORSED_DIR, Arguments::get_java_home());
400 Arguments::set_endorsed_dirs(buf);
401
402 FREE_C_HEAP_ARRAY(char, buf, mtInternal);
403
404 #undef DEFAULT_LIBPATH
405 #undef SYS_EXT_DIR
406 #undef EXTENSIONS_DIR
407 #undef ENDORSED_DIR
408 }
409
410 ////////////////////////////////////////////////////////////////////////////////
411 // breakpoint support
412
413 void os::breakpoint() {
414 BREAKPOINT;
415 }
416
417 extern "C" void breakpoint() {
418 // use debugger to set breakpoint here
419 }
420
421 ////////////////////////////////////////////////////////////////////////////////
422 // signal support
423
424 debug_only(static bool signal_sets_initialized = false);
425 static sigset_t unblocked_sigs, vm_sigs, allowdebug_blocked_sigs;
426
427 bool os::Linux::is_sig_ignored(int sig) {
428 struct sigaction oact;
429 sigaction(sig, (struct sigaction*)NULL, &oact);
430 void* ohlr = oact.sa_sigaction ? CAST_FROM_FN_PTR(void*, oact.sa_sigaction)
431 : CAST_FROM_FN_PTR(void*, oact.sa_handler);
432 if (ohlr == CAST_FROM_FN_PTR(void*, SIG_IGN))
433 return true;
434 else
435 return false;
436 }
437
438 void os::Linux::signal_sets_init() {
439 // Should also have an assertion stating we are still single-threaded.
440 assert(!signal_sets_initialized, "Already initialized");
441 // Fill in signals that are necessarily unblocked for all threads in
442 // the VM. Currently, we unblock the following signals:
443 // SHUTDOWN{1,2,3}_SIGNAL: for shutdown hooks support (unless over-ridden
444 // by -Xrs (=ReduceSignalUsage));
445 // BREAK_SIGNAL which is unblocked only by the VM thread and blocked by all
446 // other threads. The "ReduceSignalUsage" boolean tells us not to alter
447 // the dispositions or masks wrt these signals.
448 // Programs embedding the VM that want to use the above signals for their
449 // own purposes must, at this time, use the "-Xrs" option to prevent
450 // interference with shutdown hooks and BREAK_SIGNAL thread dumping.
451 // (See bug 4345157, and other related bugs).
452 // In reality, though, unblocking these signals is really a nop, since
453 // these signals are not blocked by default.
454 sigemptyset(&unblocked_sigs);
455 sigemptyset(&allowdebug_blocked_sigs);
456 sigaddset(&unblocked_sigs, SIGILL);
457 sigaddset(&unblocked_sigs, SIGSEGV);
458 sigaddset(&unblocked_sigs, SIGBUS);
459 sigaddset(&unblocked_sigs, SIGFPE);
460 #if defined(PPC64)
461 sigaddset(&unblocked_sigs, SIGTRAP);
462 #endif
463 sigaddset(&unblocked_sigs, SR_signum);
464
465 if (!ReduceSignalUsage) {
466 if (!os::Linux::is_sig_ignored(SHUTDOWN1_SIGNAL)) {
467 sigaddset(&unblocked_sigs, SHUTDOWN1_SIGNAL);
468 sigaddset(&allowdebug_blocked_sigs, SHUTDOWN1_SIGNAL);
469 }
470 if (!os::Linux::is_sig_ignored(SHUTDOWN2_SIGNAL)) {
471 sigaddset(&unblocked_sigs, SHUTDOWN2_SIGNAL);
472 sigaddset(&allowdebug_blocked_sigs, SHUTDOWN2_SIGNAL);
473 }
474 if (!os::Linux::is_sig_ignored(SHUTDOWN3_SIGNAL)) {
475 sigaddset(&unblocked_sigs, SHUTDOWN3_SIGNAL);
476 sigaddset(&allowdebug_blocked_sigs, SHUTDOWN3_SIGNAL);
477 }
478 }
479 // Fill in signals that are blocked by all but the VM thread.
480 sigemptyset(&vm_sigs);
481 if (!ReduceSignalUsage)
482 sigaddset(&vm_sigs, BREAK_SIGNAL);
483 debug_only(signal_sets_initialized = true);
484
485 }
486
487 // These are signals that are unblocked while a thread is running Java.
488 // (For some reason, they get blocked by default.)
489 sigset_t* os::Linux::unblocked_signals() {
490 assert(signal_sets_initialized, "Not initialized");
491 return &unblocked_sigs;
492 }
493
494 // These are the signals that are blocked while a (non-VM) thread is
495 // running Java. Only the VM thread handles these signals.
496 sigset_t* os::Linux::vm_signals() {
497 assert(signal_sets_initialized, "Not initialized");
498 return &vm_sigs;
499 }
500
501 // These are signals that are blocked during cond_wait to allow debugger in
502 sigset_t* os::Linux::allowdebug_blocked_signals() {
503 assert(signal_sets_initialized, "Not initialized");
504 return &allowdebug_blocked_sigs;
505 }
506
507 void os::Linux::hotspot_sigmask(Thread* thread) {
508
509 //Save caller's signal mask before setting VM signal mask
510 sigset_t caller_sigmask;
511 pthread_sigmask(SIG_BLOCK, NULL, &caller_sigmask);
512
513 OSThread* osthread = thread->osthread();
514 osthread->set_caller_sigmask(caller_sigmask);
515
516 pthread_sigmask(SIG_UNBLOCK, os::Linux::unblocked_signals(), NULL);
517
518 if (!ReduceSignalUsage) {
519 if (thread->is_VM_thread()) {
520 // Only the VM thread handles BREAK_SIGNAL ...
521 pthread_sigmask(SIG_UNBLOCK, vm_signals(), NULL);
522 } else {
523 // ... all other threads block BREAK_SIGNAL
524 pthread_sigmask(SIG_BLOCK, vm_signals(), NULL);
525 }
526 }
527 }
528
529 //////////////////////////////////////////////////////////////////////////////
530 // detecting pthread library
531
532 void os::Linux::libpthread_init() {
533 // Save glibc and pthread version strings. Note that _CS_GNU_LIBC_VERSION
534 // and _CS_GNU_LIBPTHREAD_VERSION are supported in glibc >= 2.3.2. Use a
535 // generic name for earlier versions.
536 // Define macros here so we can build HotSpot on old systems.
537 # ifndef _CS_GNU_LIBC_VERSION
538 # define _CS_GNU_LIBC_VERSION 2
539 # endif
540 # ifndef _CS_GNU_LIBPTHREAD_VERSION
541 # define _CS_GNU_LIBPTHREAD_VERSION 3
542 # endif
543
544 size_t n = confstr(_CS_GNU_LIBC_VERSION, NULL, 0);
545 if (n > 0) {
546 char *str = (char *)malloc(n, mtInternal);
547 confstr(_CS_GNU_LIBC_VERSION, str, n);
548 os::Linux::set_glibc_version(str);
549 } else {
550 // _CS_GNU_LIBC_VERSION is not supported, try gnu_get_libc_version()
551 static char _gnu_libc_version[32];
552 jio_snprintf(_gnu_libc_version, sizeof(_gnu_libc_version),
553 "glibc %s %s", gnu_get_libc_version(), gnu_get_libc_release());
554 os::Linux::set_glibc_version(_gnu_libc_version);
555 }
556
557 n = confstr(_CS_GNU_LIBPTHREAD_VERSION, NULL, 0);
558 if (n > 0) {
559 char *str = (char *)malloc(n, mtInternal);
560 confstr(_CS_GNU_LIBPTHREAD_VERSION, str, n);
561 // Vanilla RH-9 (glibc 2.3.2) has a bug that confstr() always tells
562 // us "NPTL-0.29" even we are running with LinuxThreads. Check if this
563 // is the case. LinuxThreads has a hard limit on max number of threads.
564 // So sysconf(_SC_THREAD_THREADS_MAX) will return a positive value.
565 // On the other hand, NPTL does not have such a limit, sysconf()
566 // will return -1 and errno is not changed. Check if it is really NPTL.
567 if (strcmp(os::Linux::glibc_version(), "glibc 2.3.2") == 0 &&
568 strstr(str, "NPTL") &&
569 sysconf(_SC_THREAD_THREADS_MAX) > 0) {
570 free(str);
571 os::Linux::set_libpthread_version("linuxthreads");
572 } else {
573 os::Linux::set_libpthread_version(str);
574 }
575 } else {
576 // glibc before 2.3.2 only has LinuxThreads.
577 os::Linux::set_libpthread_version("linuxthreads");
578 }
579
580 if (strstr(libpthread_version(), "NPTL")) {
581 os::Linux::set_is_NPTL();
582 } else {
583 os::Linux::set_is_LinuxThreads();
584 }
585
586 // LinuxThreads have two flavors: floating-stack mode, which allows variable
587 // stack size; and fixed-stack mode. NPTL is always floating-stack.
588 if (os::Linux::is_NPTL() || os::Linux::supports_variable_stack_size()) {
589 os::Linux::set_is_floating_stack();
590 }
591 }
592
593 /////////////////////////////////////////////////////////////////////////////
594 // thread stack
595
596 // Force Linux kernel to expand current thread stack. If "bottom" is close
597 // to the stack guard, caller should block all signals.
598 //
599 // MAP_GROWSDOWN:
600 // A special mmap() flag that is used to implement thread stacks. It tells
601 // kernel that the memory region should extend downwards when needed. This
602 // allows early versions of LinuxThreads to only mmap the first few pages
603 // when creating a new thread. Linux kernel will automatically expand thread
604 // stack as needed (on page faults).
605 //
606 // However, because the memory region of a MAP_GROWSDOWN stack can grow on
607 // demand, if a page fault happens outside an already mapped MAP_GROWSDOWN
608 // region, it's hard to tell if the fault is due to a legitimate stack
609 // access or because of reading/writing non-exist memory (e.g. buffer
610 // overrun). As a rule, if the fault happens below current stack pointer,
611 // Linux kernel does not expand stack, instead a SIGSEGV is sent to the
612 // application (see Linux kernel fault.c).
613 //
614 // This Linux feature can cause SIGSEGV when VM bangs thread stack for
615 // stack overflow detection.
616 //
617 // Newer version of LinuxThreads (since glibc-2.2, or, RH-7.x) and NPTL do
618 // not use this flag. However, the stack of initial thread is not created
619 // by pthread, it is still MAP_GROWSDOWN. Also it's possible (though
620 // unlikely) that user code can create a thread with MAP_GROWSDOWN stack
621 // and then attach the thread to JVM.
622 //
623 // To get around the problem and allow stack banging on Linux, we need to
624 // manually expand thread stack after receiving the SIGSEGV.
625 //
626 // There are two ways to expand thread stack to address "bottom", we used
627 // both of them in JVM before 1.5:
628 // 1. adjust stack pointer first so that it is below "bottom", and then
629 // touch "bottom"
630 // 2. mmap() the page in question
631 //
632 // Now alternate signal stack is gone, it's harder to use 2. For instance,
633 // if current sp is already near the lower end of page 101, and we need to
634 // call mmap() to map page 100, it is possible that part of the mmap() frame
635 // will be placed in page 100. When page 100 is mapped, it is zero-filled.
636 // That will destroy the mmap() frame and cause VM to crash.
637 //
638 // The following code works by adjusting sp first, then accessing the "bottom"
639 // page to force a page fault. Linux kernel will then automatically expand the
640 // stack mapping.
641 //
642 // _expand_stack_to() assumes its frame size is less than page size, which
643 // should always be true if the function is not inlined.
644
645 #if __GNUC__ < 3 // gcc 2.x does not support noinline attribute
646 #define NOINLINE
647 #else
648 #define NOINLINE __attribute__ ((noinline))
649 #endif
650
651 static void _expand_stack_to(address bottom) NOINLINE;
652
653 static void _expand_stack_to(address bottom) {
654 address sp;
655 size_t size;
656 volatile char *p;
657
658 // Adjust bottom to point to the largest address within the same page, it
659 // gives us a one-page buffer if alloca() allocates slightly more memory.
660 bottom = (address)align_size_down((uintptr_t)bottom, os::Linux::page_size());
661 bottom += os::Linux::page_size() - 1;
662
663 // sp might be slightly above current stack pointer; if that's the case, we
664 // will alloca() a little more space than necessary, which is OK. Don't use
665 // os::current_stack_pointer(), as its result can be slightly below current
666 // stack pointer, causing us to not alloca enough to reach "bottom".
667 sp = (address)&sp;
668
669 if (sp > bottom) {
670 size = sp - bottom;
671 p = (volatile char *)alloca(size);
672 assert(p != NULL && p <= (volatile char *)bottom, "alloca problem?");
673 p[0] = '\0';
674 }
675 }
676
677 void os::Linux::expand_stack_to(address bottom) {
678 _expand_stack_to(bottom);
679 }
680
681 bool os::Linux::manually_expand_stack(JavaThread * t, address addr) {
682 assert(t!=NULL, "just checking");
683 assert(t->osthread()->expanding_stack(), "expand should be set");
684 assert(t->stack_base() != NULL, "stack_base was not initialized");
685
686 if (addr < t->stack_base() && addr >= t->stack_yellow_zone_base()) {
687 sigset_t mask_all, old_sigset;
688 sigfillset(&mask_all);
689 pthread_sigmask(SIG_SETMASK, &mask_all, &old_sigset);
690 _expand_stack_to(addr);
691 pthread_sigmask(SIG_SETMASK, &old_sigset, NULL);
692 return true;
693 }
694 return false;
695 }
696
697 //////////////////////////////////////////////////////////////////////////////
698 // create new thread
699
700 static address highest_vm_reserved_address();
701
702 // check if it's safe to start a new thread
703 static bool _thread_safety_check(Thread* thread) {
704 if (os::Linux::is_LinuxThreads() && !os::Linux::is_floating_stack()) {
705 // Fixed stack LinuxThreads (SuSE Linux/x86, and some versions of Redhat)
706 // Heap is mmap'ed at lower end of memory space. Thread stacks are
707 // allocated (MAP_FIXED) from high address space. Every thread stack
708 // occupies a fixed size slot (usually 2Mbytes, but user can change
709 // it to other values if they rebuild LinuxThreads).
710 //
711 // Problem with MAP_FIXED is that mmap() can still succeed even part of
712 // the memory region has already been mmap'ed. That means if we have too
713 // many threads and/or very large heap, eventually thread stack will
714 // collide with heap.
715 //
716 // Here we try to prevent heap/stack collision by comparing current
717 // stack bottom with the highest address that has been mmap'ed by JVM
718 // plus a safety margin for memory maps created by native code.
719 //
720 // This feature can be disabled by setting ThreadSafetyMargin to 0
721 //
722 if (ThreadSafetyMargin > 0) {
723 address stack_bottom = os::current_stack_base() - os::current_stack_size();
724
725 // not safe if our stack extends below the safety margin
726 return stack_bottom - ThreadSafetyMargin >= highest_vm_reserved_address();
727 } else {
728 return true;
729 }
730 } else {
731 // Floating stack LinuxThreads or NPTL:
732 // Unlike fixed stack LinuxThreads, thread stacks are not MAP_FIXED. When
733 // there's not enough space left, pthread_create() will fail. If we come
734 // here, that means enough space has been reserved for stack.
735 return true;
736 }
737 }
738
739 // Thread start routine for all newly created threads
740 static void *java_start(Thread *thread) {
741 // Try to randomize the cache line index of hot stack frames.
742 // This helps when threads of the same stack traces evict each other's
743 // cache lines. The threads can be either from the same JVM instance, or
744 // from different JVM instances. The benefit is especially true for
745 // processors with hyperthreading technology.
746 static int counter = 0;
747 int pid = os::current_process_id();
748 alloca(((pid ^ counter++) & 7) * 128);
749
750 ThreadLocalStorage::set_thread(thread);
751
752 OSThread* osthread = thread->osthread();
753 Monitor* sync = osthread->startThread_lock();
754
755 // non floating stack LinuxThreads needs extra check, see above
756 if (!_thread_safety_check(thread)) {
757 // notify parent thread
758 MutexLockerEx ml(sync, Mutex::_no_safepoint_check_flag);
759 osthread->set_state(ZOMBIE);
760 sync->notify_all();
761 return NULL;
762 }
763
764 // thread_id is kernel thread id (similar to Solaris LWP id)
765 osthread->set_thread_id(os::Linux::gettid());
766
767 if (UseNUMA) {
768 int lgrp_id = os::numa_get_group_id();
769 if (lgrp_id != -1) {
770 thread->set_lgrp_id(lgrp_id);
771 }
772 }
773 // initialize signal mask for this thread
774 os::Linux::hotspot_sigmask(thread);
775
776 // initialize floating point control register
777 os::Linux::init_thread_fpu_state();
778
779 // handshaking with parent thread
780 {
781 MutexLockerEx ml(sync, Mutex::_no_safepoint_check_flag);
782
783 // notify parent thread
784 osthread->set_state(INITIALIZED);
785 sync->notify_all();
786
787 // wait until os::start_thread()
788 while (osthread->get_state() == INITIALIZED) {
789 sync->wait(Mutex::_no_safepoint_check_flag);
790 }
791 }
792
793 // call one more level start routine
794 thread->run();
795
796 return 0;
797 }
798
799 bool os::create_thread(Thread* thread, ThreadType thr_type, size_t stack_size) {
800 assert(thread->osthread() == NULL, "caller responsible");
801
802 // Allocate the OSThread object
803 OSThread* osthread = new OSThread(NULL, NULL);
804 if (osthread == NULL) {
805 return false;
806 }
807
808 // set the correct thread state
809 osthread->set_thread_type(thr_type);
810
811 // Initial state is ALLOCATED but not INITIALIZED
812 osthread->set_state(ALLOCATED);
813
814 thread->set_osthread(osthread);
815
816 // init thread attributes
817 pthread_attr_t attr;
818 pthread_attr_init(&attr);
819 pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
820
821 // stack size
822 if (os::Linux::supports_variable_stack_size()) {
823 // calculate stack size if it's not specified by caller
824 if (stack_size == 0) {
825 stack_size = os::Linux::default_stack_size(thr_type);
826
827 switch (thr_type) {
828 case os::java_thread:
829 // Java threads use ThreadStackSize which default value can be
830 // changed with the flag -Xss
831 assert (JavaThread::stack_size_at_create() > 0, "this should be set");
832 stack_size = JavaThread::stack_size_at_create();
833 break;
834 case os::compiler_thread:
835 if (CompilerThreadStackSize > 0) {
836 stack_size = (size_t)(CompilerThreadStackSize * K);
837 break;
838 } // else fall through:
839 // use VMThreadStackSize if CompilerThreadStackSize is not defined
840 case os::vm_thread:
841 case os::pgc_thread:
842 case os::cgc_thread:
843 case os::watcher_thread:
844 if (VMThreadStackSize > 0) stack_size = (size_t)(VMThreadStackSize * K);
845 break;
846 }
847 }
848
849 stack_size = MAX2(stack_size, os::Linux::min_stack_allowed);
850 pthread_attr_setstacksize(&attr, stack_size);
851 } else {
852 // let pthread_create() pick the default value.
853 }
854
855 // glibc guard page
856 pthread_attr_setguardsize(&attr, os::Linux::default_guard_size(thr_type));
857
858 ThreadState state;
859
860 {
861 // Serialize thread creation if we are running with fixed stack LinuxThreads
862 bool lock = os::Linux::is_LinuxThreads() && !os::Linux::is_floating_stack();
863 if (lock) {
864 os::Linux::createThread_lock()->lock_without_safepoint_check();
865 }
866
867 pthread_t tid;
868 int ret = pthread_create(&tid, &attr, (void* (*)(void*)) java_start, thread);
869
870 pthread_attr_destroy(&attr);
871
872 if (ret != 0) {
873 if (PrintMiscellaneous && (Verbose || WizardMode)) {
874 perror("pthread_create()");
875 }
876 // Need to clean up stuff we've allocated so far
877 thread->set_osthread(NULL);
878 delete osthread;
879 if (lock) os::Linux::createThread_lock()->unlock();
880 return false;
881 }
882
883 // Store pthread info into the OSThread
884 osthread->set_pthread_id(tid);
885
886 // Wait until child thread is either initialized or aborted
887 {
888 Monitor* sync_with_child = osthread->startThread_lock();
889 MutexLockerEx ml(sync_with_child, Mutex::_no_safepoint_check_flag);
890 while ((state = osthread->get_state()) == ALLOCATED) {
891 sync_with_child->wait(Mutex::_no_safepoint_check_flag);
892 }
893 }
894
895 if (lock) {
896 os::Linux::createThread_lock()->unlock();
897 }
898 }
899
900 // Aborted due to thread limit being reached
901 if (state == ZOMBIE) {
902 thread->set_osthread(NULL);
903 delete osthread;
904 return false;
905 }
906
907 // The thread is returned suspended (in state INITIALIZED),
908 // and is started higher up in the call chain
909 assert(state == INITIALIZED, "race condition");
910 return true;
911 }
912
913 /////////////////////////////////////////////////////////////////////////////
914 // attach existing thread
915
916 // bootstrap the main thread
917 bool os::create_main_thread(JavaThread* thread) {
918 assert(os::Linux::_main_thread == pthread_self(), "should be called inside main thread");
919 return create_attached_thread(thread);
920 }
921
922 bool os::create_attached_thread(JavaThread* thread) {
923 #ifdef ASSERT
924 thread->verify_not_published();
925 #endif
926
927 // Allocate the OSThread object
928 OSThread* osthread = new OSThread(NULL, NULL);
929
930 if (osthread == NULL) {
931 return false;
932 }
933
934 // Store pthread info into the OSThread
935 osthread->set_thread_id(os::Linux::gettid());
936 osthread->set_pthread_id(::pthread_self());
937
938 // initialize floating point control register
939 os::Linux::init_thread_fpu_state();
940
941 // Initial thread state is RUNNABLE
942 osthread->set_state(RUNNABLE);
943
944 thread->set_osthread(osthread);
945
946 if (UseNUMA) {
947 int lgrp_id = os::numa_get_group_id();
948 if (lgrp_id != -1) {
949 thread->set_lgrp_id(lgrp_id);
950 }
951 }
952
953 if (os::is_primordial_thread()) {
954 // If current thread is primordial thread, its stack is mapped on demand,
955 // see notes about MAP_GROWSDOWN. Here we try to force kernel to map
956 // the entire stack region to avoid SEGV in stack banging.
957 // It is also useful to get around the heap-stack-gap problem on SuSE
958 // kernel (see 4821821 for details). We first expand stack to the top
959 // of yellow zone, then enable stack yellow zone (order is significant,
960 // enabling yellow zone first will crash JVM on SuSE Linux), so there
961 // is no gap between the last two virtual memory regions.
962
963 JavaThread *jt = (JavaThread *)thread;
964 address addr = jt->stack_yellow_zone_base();
965 assert(addr != NULL, "initialization problem?");
966 assert(jt->stack_available(addr) > 0, "stack guard should not be enabled");
967
968 osthread->set_expanding_stack();
969 os::Linux::manually_expand_stack(jt, addr);
970 osthread->clear_expanding_stack();
971 }
972
973 // initialize signal mask for this thread
974 // and save the caller's signal mask
975 os::Linux::hotspot_sigmask(thread);
976
977 return true;
978 }
979
980 void os::pd_start_thread(Thread* thread) {
981 OSThread * osthread = thread->osthread();
982 assert(osthread->get_state() != INITIALIZED, "just checking");
983 Monitor* sync_with_child = osthread->startThread_lock();
984 MutexLockerEx ml(sync_with_child, Mutex::_no_safepoint_check_flag);
985 sync_with_child->notify();
986 }
987
988 // Free Linux resources related to the OSThread
989 void os::free_thread(OSThread* osthread) {
990 assert(osthread != NULL, "osthread not set");
991
992 if (Thread::current()->osthread() == osthread) {
993 // Restore caller's signal mask
994 sigset_t sigmask = osthread->caller_sigmask();
995 pthread_sigmask(SIG_SETMASK, &sigmask, NULL);
996 }
997
998 delete osthread;
999 }
1000
1001 //////////////////////////////////////////////////////////////////////////////
1002 // thread local storage
1003
1004 // Restore the thread pointer if the destructor is called. This is in case
1005 // someone from JNI code sets up a destructor with pthread_key_create to run
1006 // detachCurrentThread on thread death. Unless we restore the thread pointer we
1007 // will hang or crash. When detachCurrentThread is called the key will be set
1008 // to null and we will not be called again. If detachCurrentThread is never
1009 // called we could loop forever depending on the pthread implementation.
1010 static void restore_thread_pointer(void* p) {
1011 Thread* thread = (Thread*) p;
1012 os::thread_local_storage_at_put(ThreadLocalStorage::thread_index(), thread);
1013 }
1014
1015 int os::allocate_thread_local_storage() {
1016 pthread_key_t key;
1017 int rslt = pthread_key_create(&key, restore_thread_pointer);
1018 assert(rslt == 0, "cannot allocate thread local storage");
1019 return (int)key;
1020 }
1021
1022 // Note: This is currently not used by VM, as we don't destroy TLS key
1023 // on VM exit.
1024 void os::free_thread_local_storage(int index) {
1025 int rslt = pthread_key_delete((pthread_key_t)index);
1026 assert(rslt == 0, "invalid index");
1027 }
1028
1029 void os::thread_local_storage_at_put(int index, void* value) {
1030 int rslt = pthread_setspecific((pthread_key_t)index, value);
1031 assert(rslt == 0, "pthread_setspecific failed");
1032 }
1033
1034 extern "C" Thread* get_thread() {
1035 return ThreadLocalStorage::thread();
1036 }
1037
1038 //////////////////////////////////////////////////////////////////////////////
1039 // primordial thread
1040
1041 // Check if current thread is the primordial thread, similar to Solaris thr_main.
1042 bool os::is_primordial_thread(void) {
1043 char dummy;
1044 // If called before init complete, thread stack bottom will be null.
1045 // Can be called if fatal error occurs before initialization.
1046 if (os::Linux::initial_thread_stack_bottom() == NULL) return false;
1047 assert(os::Linux::initial_thread_stack_bottom() != NULL &&
1048 os::Linux::initial_thread_stack_size() != 0,
1049 "os::init did not locate primordial thread's stack region");
1050 if ((address)&dummy >= os::Linux::initial_thread_stack_bottom() &&
1051 (address)&dummy < os::Linux::initial_thread_stack_bottom() +
1052 os::Linux::initial_thread_stack_size()) {
1053 return true;
1054 } else {
1055 return false;
1056 }
1057 }
1058
1059 // Find the virtual memory area that contains addr
1060 static bool find_vma(address addr, address* vma_low, address* vma_high) {
1061 FILE *fp = fopen("/proc/self/maps", "r");
1062 if (fp) {
1063 address low, high;
1064 while (!feof(fp)) {
1065 if (fscanf(fp, "%p-%p", &low, &high) == 2) {
1066 if (low <= addr && addr < high) {
1067 if (vma_low) *vma_low = low;
1068 if (vma_high) *vma_high = high;
1069 fclose (fp);
1070 return true;
1071 }
1072 }
1073 for (;;) {
1074 int ch = fgetc(fp);
1075 if (ch == EOF || ch == (int)'\n') break;
1076 }
1077 }
1078 fclose(fp);
1079 }
1080 return false;
1081 }
1082
1083 // Locate primordial thread stack. This special handling of primordial thread stack
1084 // is needed because pthread_getattr_np() on most (all?) Linux distros returns
1085 // bogus value for the primordial process thread. While the launcher has created
1086 // the VM in a new thread since JDK 6, we still have to allow for the use of the
1087 // JNI invocation API from a primordial thread.
1088 void os::Linux::capture_initial_stack(size_t max_size) {
1089
1090 // max_size is either 0 (which means accept OS default for thread stacks) or
1091 // a user-specified value known to be at least the minimum needed. If we
1092 // are actually on the primordial thread we can make it appear that we have a
1093 // smaller max_size stack by inserting the guard pages at that location. But we
1094 // cannot do anything to emulate a larger stack than what has been provided by
1095 // the OS or threading library. In fact if we try to use a stack greater than
1096 // what is set by rlimit then we will crash the hosting process.
1097
1098 // Maximum stack size is the easy part, get it from RLIMIT_STACK.
1099 // If this is "unlimited" then it will be a huge value.
1100 struct rlimit rlim;
1101 getrlimit(RLIMIT_STACK, &rlim);
1102 size_t stack_size = rlim.rlim_cur;
1103
1104 // 6308388: a bug in ld.so will relocate its own .data section to the
1105 // lower end of primordial stack; reduce ulimit -s value a little bit
1106 // so we won't install guard page on ld.so's data section.
1107 // But ensure we don't underflow the stack size - allow 1 page spare
1108 if (stack_size >= (size_t)(3 * page_size())) {
1109 stack_size -= 2 * page_size();
1110 }
1111
1112 // Try to figure out where the stack base (top) is. This is harder.
1113 //
1114 // When an application is started, glibc saves the initial stack pointer in
1115 // a global variable "__libc_stack_end", which is then used by system
1116 // libraries. __libc_stack_end should be pretty close to stack top. The
1117 // variable is available since the very early days. However, because it is
1118 // a private interface, it could disappear in the future.
1119 //
1120 // Linux kernel saves start_stack information in /proc/<pid>/stat. Similar
1121 // to __libc_stack_end, it is very close to stack top, but isn't the real
1122 // stack top. Note that /proc may not exist if VM is running as a chroot
1123 // program, so reading /proc/<pid>/stat could fail. Also the contents of
1124 // /proc/<pid>/stat could change in the future (though unlikely).
1125 //
1126 // We try __libc_stack_end first. If that doesn't work, look for
1127 // /proc/<pid>/stat. If neither of them works, we use current stack pointer
1128 // as a hint, which should work well in most cases.
1129
1130 uintptr_t stack_start;
1131
1132 // try __libc_stack_end first
1133 uintptr_t *p = (uintptr_t *)dlsym(RTLD_DEFAULT, "__libc_stack_end");
1134 if (p && *p) {
1135 stack_start = *p;
1136 } else {
1137 // see if we can get the start_stack field from /proc/self/stat
1138 FILE *fp;
1139 int pid;
1140 char state;
1141 int ppid;
1142 int pgrp;
1143 int session;
1144 int nr;
1145 int tpgrp;
1146 unsigned long flags;
1147 unsigned long minflt;
1148 unsigned long cminflt;
1149 unsigned long majflt;
1150 unsigned long cmajflt;
1151 unsigned long utime;
1152 unsigned long stime;
1153 long cutime;
1154 long cstime;
1155 long prio;
1156 long nice;
1157 long junk;
1158 long it_real;
1159 uintptr_t start;
1160 uintptr_t vsize;
1161 intptr_t rss;
1162 uintptr_t rsslim;
1163 uintptr_t scodes;
1164 uintptr_t ecode;
1165 int i;
1166
1167 // Figure what the primordial thread stack base is. Code is inspired
1168 // by email from Hans Boehm. /proc/self/stat begins with current pid,
1169 // followed by command name surrounded by parentheses, state, etc.
1170 char stat[2048];
1171 int statlen;
1172
1173 fp = fopen("/proc/self/stat", "r");
1174 if (fp) {
1175 statlen = fread(stat, 1, 2047, fp);
1176 stat[statlen] = '\0';
1177 fclose(fp);
1178
1179 // Skip pid and the command string. Note that we could be dealing with
1180 // weird command names, e.g. user could decide to rename java launcher
1181 // to "java 1.4.2 :)", then the stat file would look like
1182 // 1234 (java 1.4.2 :)) R ... ...
1183 // We don't really need to know the command string, just find the last
1184 // occurrence of ")" and then start parsing from there. See bug 4726580.
1185 char * s = strrchr(stat, ')');
1186
1187 i = 0;
1188 if (s) {
1189 // Skip blank chars
1190 do s++; while (isspace(*s));
1191
1192 #define _UFM UINTX_FORMAT
1193 #define _DFM INTX_FORMAT
1194
1195 /* 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 */
1196 /* 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 */
1197 i = sscanf(s, "%c %d %d %d %d %d %lu %lu %lu %lu %lu %lu %lu %ld %ld %ld %ld %ld %ld " _UFM _UFM _DFM _UFM _UFM _UFM _UFM,
1198 &state, /* 3 %c */
1199 &ppid, /* 4 %d */
1200 &pgrp, /* 5 %d */
1201 &session, /* 6 %d */
1202 &nr, /* 7 %d */
1203 &tpgrp, /* 8 %d */
1204 &flags, /* 9 %lu */
1205 &minflt, /* 10 %lu */
1206 &cminflt, /* 11 %lu */
1207 &majflt, /* 12 %lu */
1208 &cmajflt, /* 13 %lu */
1209 &utime, /* 14 %lu */
1210 &stime, /* 15 %lu */
1211 &cutime, /* 16 %ld */
1212 &cstime, /* 17 %ld */
1213 &prio, /* 18 %ld */
1214 &nice, /* 19 %ld */
1215 &junk, /* 20 %ld */
1216 &it_real, /* 21 %ld */
1217 &start, /* 22 UINTX_FORMAT */
1218 &vsize, /* 23 UINTX_FORMAT */
1219 &rss, /* 24 INTX_FORMAT */
1220 &rsslim, /* 25 UINTX_FORMAT */
1221 &scodes, /* 26 UINTX_FORMAT */
1222 &ecode, /* 27 UINTX_FORMAT */
1223 &stack_start); /* 28 UINTX_FORMAT */
1224 }
1225
1226 #undef _UFM
1227 #undef _DFM
1228
1229 if (i != 28 - 2) {
1230 assert(false, "Bad conversion from /proc/self/stat");
1231 // product mode - assume we are the primordial thread, good luck in the
1232 // embedded case.
1233 warning("Can't detect primordial thread stack location - bad conversion");
1234 stack_start = (uintptr_t) &rlim;
1235 }
1236 } else {
1237 // For some reason we can't open /proc/self/stat (for example, running on
1238 // FreeBSD with a Linux emulator, or inside chroot), this should work for
1239 // most cases, so don't abort:
1240 warning("Can't detect primordial thread stack location - no /proc/self/stat");
1241 stack_start = (uintptr_t) &rlim;
1242 }
1243 }
1244
1245 // Now we have a pointer (stack_start) very close to the stack top, the
1246 // next thing to do is to figure out the exact location of stack top. We
1247 // can find out the virtual memory area that contains stack_start by
1248 // reading /proc/self/maps, it should be the last vma in /proc/self/maps,
1249 // and its upper limit is the real stack top. (again, this would fail if
1250 // running inside chroot, because /proc may not exist.)
1251
1252 uintptr_t stack_top;
1253 address low, high;
1254 if (find_vma((address)stack_start, &low, &high)) {
1255 // success, "high" is the true stack top. (ignore "low", because initial
1256 // thread stack grows on demand, its real bottom is high - RLIMIT_STACK.)
1257 stack_top = (uintptr_t)high;
1258 } else {
1259 // failed, likely because /proc/self/maps does not exist
1260 warning("Can't detect primordial thread stack location - find_vma failed");
1261 // best effort: stack_start is normally within a few pages below the real
1262 // stack top, use it as stack top, and reduce stack size so we won't put
1263 // guard page outside stack.
1264 stack_top = stack_start;
1265 stack_size -= 16 * page_size();
1266 }
1267
1268 // stack_top could be partially down the page so align it
1269 stack_top = align_size_up(stack_top, page_size());
1270
1271 // Allowed stack value is minimum of max_size and what we derived from rlimit
1272 if (max_size > 0) {
1273 _initial_thread_stack_size = MIN2(max_size, stack_size);
1274 } else {
1275 // Accept the rlimit max, but if stack is unlimited then it will be huge, so
1276 // clamp it at 8MB as we do on Solaris
1277 _initial_thread_stack_size = MIN2(stack_size, 8*M);
1278 }
1279
1280 _initial_thread_stack_size = align_size_down(_initial_thread_stack_size, page_size());
1281 _initial_thread_stack_bottom = (address)stack_top - _initial_thread_stack_size;
1282 assert(_initial_thread_stack_bottom < (address)stack_top, "overflow!");
1283 }
1284
1285 ////////////////////////////////////////////////////////////////////////////////
1286 // time support
1287
1288 // Time since start-up in seconds to a fine granularity.
1289 // Used by VMSelfDestructTimer and the MemProfiler.
1290 double os::elapsedTime() {
1291
1292 return ((double)os::elapsed_counter()) / os::elapsed_frequency(); // nanosecond resolution
1293 }
1294
1295 jlong os::elapsed_counter() {
1296 return javaTimeNanos() - initial_time_count;
1297 }
1298
1299 jlong os::elapsed_frequency() {
1300 return NANOSECS_PER_SEC; // nanosecond resolution
1301 }
1302
1303 bool os::supports_vtime() { return true; }
1304 bool os::enable_vtime() { return false; }
1305 bool os::vtime_enabled() { return false; }
1306
1307 double os::elapsedVTime() {
1308 struct rusage usage;
1309 int retval = getrusage(RUSAGE_THREAD, &usage);
1310 if (retval == 0) {
1311 return (double) (usage.ru_utime.tv_sec + usage.ru_stime.tv_sec) + (double) (usage.ru_utime.tv_usec + usage.ru_stime.tv_usec) / (1000 * 1000);
1312 } else {
1313 // better than nothing, but not much
1314 return elapsedTime();
1315 }
1316 }
1317
1318 jlong os::javaTimeMillis() {
1319 timeval time;
1320 int status = gettimeofday(&time, NULL);
1321 assert(status != -1, "linux error");
1322 return jlong(time.tv_sec) * 1000 + jlong(time.tv_usec / 1000);
1323 }
1324
1325 #ifndef CLOCK_MONOTONIC
1326 #define CLOCK_MONOTONIC (1)
1327 #endif
1328
1329 void os::Linux::clock_init() {
1330 // we do dlopen's in this particular order due to bug in linux
1331 // dynamical loader (see 6348968) leading to crash on exit
1332 void* handle = dlopen("librt.so.1", RTLD_LAZY);
1333 if (handle == NULL) {
1334 handle = dlopen("librt.so", RTLD_LAZY);
1335 }
1336
1337 if (handle) {
1338 int (*clock_getres_func)(clockid_t, struct timespec*) =
1339 (int(*)(clockid_t, struct timespec*))dlsym(handle, "clock_getres");
1340 int (*clock_gettime_func)(clockid_t, struct timespec*) =
1341 (int(*)(clockid_t, struct timespec*))dlsym(handle, "clock_gettime");
1342 if (clock_getres_func && clock_gettime_func) {
1343 // See if monotonic clock is supported by the kernel. Note that some
1344 // early implementations simply return kernel jiffies (updated every
1345 // 1/100 or 1/1000 second). It would be bad to use such a low res clock
1346 // for nano time (though the monotonic property is still nice to have).
1347 // It's fixed in newer kernels, however clock_getres() still returns
1348 // 1/HZ. We check if clock_getres() works, but will ignore its reported
1349 // resolution for now. Hopefully as people move to new kernels, this
1350 // won't be a problem.
1351 struct timespec res;
1352 struct timespec tp;
1353 if (clock_getres_func (CLOCK_MONOTONIC, &res) == 0 &&
1354 clock_gettime_func(CLOCK_MONOTONIC, &tp) == 0) {
1355 // yes, monotonic clock is supported
1356 _clock_gettime = clock_gettime_func;
1357 return;
1358 } else {
1359 // close librt if there is no monotonic clock
1360 dlclose(handle);
1361 }
1362 }
1363 }
1364 warning("No monotonic clock was available - timed services may " \
1365 "be adversely affected if the time-of-day clock changes");
1366 }
1367
1368 #ifndef SYS_clock_getres
1369
1370 #if defined(IA32) || defined(AMD64)
1371 #define SYS_clock_getres IA32_ONLY(266) AMD64_ONLY(229)
1372 #define sys_clock_getres(x,y) ::syscall(SYS_clock_getres, x, y)
1373 #else
1374 #warning "SYS_clock_getres not defined for this platform, disabling fast_thread_cpu_time"
1375 #define sys_clock_getres(x,y) -1
1376 #endif
1377
1378 #else
1379 #define sys_clock_getres(x,y) ::syscall(SYS_clock_getres, x, y)
1380 #endif
1381
1382 void os::Linux::fast_thread_clock_init() {
1383 if (!UseLinuxPosixThreadCPUClocks) {
1384 return;
1385 }
1386 clockid_t clockid;
1387 struct timespec tp;
1388 int (*pthread_getcpuclockid_func)(pthread_t, clockid_t *) =
1389 (int(*)(pthread_t, clockid_t *)) dlsym(RTLD_DEFAULT, "pthread_getcpuclockid");
1390
1391 // Switch to using fast clocks for thread cpu time if
1392 // the sys_clock_getres() returns 0 error code.
1393 // Note, that some kernels may support the current thread
1394 // clock (CLOCK_THREAD_CPUTIME_ID) but not the clocks
1395 // returned by the pthread_getcpuclockid().
1396 // If the fast Posix clocks are supported then the sys_clock_getres()
1397 // must return at least tp.tv_sec == 0 which means a resolution
1398 // better than 1 sec. This is extra check for reliability.
1399
1400 if(pthread_getcpuclockid_func &&
1401 pthread_getcpuclockid_func(_main_thread, &clockid) == 0 &&
1402 sys_clock_getres(clockid, &tp) == 0 && tp.tv_sec == 0) {
1403
1404 _supports_fast_thread_cpu_time = true;
1405 _pthread_getcpuclockid = pthread_getcpuclockid_func;
1406 }
1407 }
1408
1409 jlong os::javaTimeNanos() {
1410 if (Linux::supports_monotonic_clock()) {
1411 struct timespec tp;
1412 int status = Linux::clock_gettime(CLOCK_MONOTONIC, &tp);
1413 assert(status == 0, "gettime error");
1414 jlong result = jlong(tp.tv_sec) * (1000 * 1000 * 1000) + jlong(tp.tv_nsec);
1415 return result;
1416 } else {
1417 timeval time;
1418 int status = gettimeofday(&time, NULL);
1419 assert(status != -1, "linux error");
1420 jlong usecs = jlong(time.tv_sec) * (1000 * 1000) + jlong(time.tv_usec);
1421 return 1000 * usecs;
1422 }
1423 }
1424
1425 void os::javaTimeNanos_info(jvmtiTimerInfo *info_ptr) {
1426 if (Linux::supports_monotonic_clock()) {
1427 info_ptr->max_value = ALL_64_BITS;
1428
1429 // CLOCK_MONOTONIC - amount of time since some arbitrary point in the past
1430 info_ptr->may_skip_backward = false; // not subject to resetting or drifting
1431 info_ptr->may_skip_forward = false; // not subject to resetting or drifting
1432 } else {
1433 // gettimeofday - based on time in seconds since the Epoch thus does not wrap
1434 info_ptr->max_value = ALL_64_BITS;
1435
1436 // gettimeofday is a real time clock so it skips
1437 info_ptr->may_skip_backward = true;
1438 info_ptr->may_skip_forward = true;
1439 }
1440
1441 info_ptr->kind = JVMTI_TIMER_ELAPSED; // elapsed not CPU time
1442 }
1443
1444 // Return the real, user, and system times in seconds from an
1445 // arbitrary fixed point in the past.
1446 bool os::getTimesSecs(double* process_real_time,
1447 double* process_user_time,
1448 double* process_system_time) {
1449 struct tms ticks;
1450 clock_t real_ticks = times(&ticks);
1451
1452 if (real_ticks == (clock_t) (-1)) {
1453 return false;
1454 } else {
1455 double ticks_per_second = (double) clock_tics_per_sec;
1456 *process_user_time = ((double) ticks.tms_utime) / ticks_per_second;
1457 *process_system_time = ((double) ticks.tms_stime) / ticks_per_second;
1458 *process_real_time = ((double) real_ticks) / ticks_per_second;
1459
1460 return true;
1461 }
1462 }
1463
1464
1465 char * os::local_time_string(char *buf, size_t buflen) {
1466 struct tm t;
1467 time_t long_time;
1468 time(&long_time);
1469 localtime_r(&long_time, &t);
1470 jio_snprintf(buf, buflen, "%d-%02d-%02d %02d:%02d:%02d",
1471 t.tm_year + 1900, t.tm_mon + 1, t.tm_mday,
1472 t.tm_hour, t.tm_min, t.tm_sec);
1473 return buf;
1474 }
1475
1476 struct tm* os::localtime_pd(const time_t* clock, struct tm* res) {
1477 return localtime_r(clock, res);
1478 }
1479
1480 ////////////////////////////////////////////////////////////////////////////////
1481 // runtime exit support
1482
1483 // Note: os::shutdown() might be called very early during initialization, or
1484 // called from signal handler. Before adding something to os::shutdown(), make
1485 // sure it is async-safe and can handle partially initialized VM.
1486 void os::shutdown() {
1487
1488 // allow PerfMemory to attempt cleanup of any persistent resources
1489 perfMemory_exit();
1490
1491 // needs to remove object in file system
1492 AttachListener::abort();
1493
1494 // flush buffered output, finish log files
1495 ostream_abort();
1496
1497 // Check for abort hook
1498 abort_hook_t abort_hook = Arguments::abort_hook();
1499 if (abort_hook != NULL) {
1500 abort_hook();
1501 }
1502
1503 }
1504
1505 // Note: os::abort() might be called very early during initialization, or
1506 // called from signal handler. Before adding something to os::abort(), make
1507 // sure it is async-safe and can handle partially initialized VM.
1508 void os::abort(bool dump_core) {
1509 os::shutdown();
1510 if (dump_core) {
1511 #ifndef PRODUCT
1512 fdStream out(defaultStream::output_fd());
1513 out.print_raw("Current thread is ");
1514 char buf[16];
1515 jio_snprintf(buf, sizeof(buf), UINTX_FORMAT, os::current_thread_id());
1516 out.print_raw_cr(buf);
1517 out.print_raw_cr("Dumping core ...");
1518 #endif
1519 ::abort(); // dump core
1520 }
1521
1522 ::exit(1);
1523 }
1524
1525 // Die immediately, no exit hook, no abort hook, no cleanup.
1526 void os::die() {
1527 // _exit() on LinuxThreads only kills current thread
1528 ::abort();
1529 }
1530
1531
1532 // This method is a copy of JDK's sysGetLastErrorString
1533 // from src/solaris/hpi/src/system_md.c
1534
1535 size_t os::lasterror(char *buf, size_t len) {
1536
1537 if (errno == 0) return 0;
1538
1539 const char *s = ::strerror(errno);
1540 size_t n = ::strlen(s);
1541 if (n >= len) {
1542 n = len - 1;
1543 }
1544 ::strncpy(buf, s, n);
1545 buf[n] = '\0';
1546 return n;
1547 }
1548
1549 intx os::current_thread_id() { return (intx)pthread_self(); }
1550 int os::current_process_id() {
1551
1552 // Under the old linux thread library, linux gives each thread
1553 // its own process id. Because of this each thread will return
1554 // a different pid if this method were to return the result
1555 // of getpid(2). Linux provides no api that returns the pid
1556 // of the launcher thread for the vm. This implementation
1557 // returns a unique pid, the pid of the launcher thread
1558 // that starts the vm 'process'.
1559
1560 // Under the NPTL, getpid() returns the same pid as the
1561 // launcher thread rather than a unique pid per thread.
1562 // Use gettid() if you want the old pre NPTL behaviour.
1563
1564 // if you are looking for the result of a call to getpid() that
1565 // returns a unique pid for the calling thread, then look at the
1566 // OSThread::thread_id() method in osThread_linux.hpp file
1567
1568 return (int)(_initial_pid ? _initial_pid : getpid());
1569 }
1570
1571 // DLL functions
1572
1573 const char* os::dll_file_extension() { return ".so"; }
1574
1575 // This must be hard coded because it's the system's temporary
1576 // directory not the java application's temp directory, ala java.io.tmpdir.
1577 const char* os::get_temp_directory() { return "/tmp"; }
1578
1579 static bool file_exists(const char* filename) {
1580 struct stat statbuf;
1581 if (filename == NULL || strlen(filename) == 0) {
1582 return false;
1583 }
1584 return os::stat(filename, &statbuf) == 0;
1585 }
1586
1587 bool os::dll_build_name(char* buffer, size_t buflen,
1588 const char* pname, const char* fname) {
1589 bool retval = false;
1590 // Copied from libhpi
1591 const size_t pnamelen = pname ? strlen(pname) : 0;
1592
1593 // Return error on buffer overflow.
1594 if (pnamelen + strlen(fname) + 10 > (size_t) buflen) {
1595 return retval;
1596 }
1597
1598 if (pnamelen == 0) {
1599 snprintf(buffer, buflen, "lib%s.so", fname);
1600 retval = true;
1601 } else if (strchr(pname, *os::path_separator()) != NULL) {
1602 int n;
1603 char** pelements = split_path(pname, &n);
1604 if (pelements == NULL) {
1605 return false;
1606 }
1607 for (int i = 0 ; i < n ; i++) {
1608 // Really shouldn't be NULL, but check can't hurt
1609 if (pelements[i] == NULL || strlen(pelements[i]) == 0) {
1610 continue; // skip the empty path values
1611 }
1612 snprintf(buffer, buflen, "%s/lib%s.so", pelements[i], fname);
1613 if (file_exists(buffer)) {
1614 retval = true;
1615 break;
1616 }
1617 }
1618 // release the storage
1619 for (int i = 0 ; i < n ; i++) {
1620 if (pelements[i] != NULL) {
1621 FREE_C_HEAP_ARRAY(char, pelements[i], mtInternal);
1622 }
1623 }
1624 if (pelements != NULL) {
1625 FREE_C_HEAP_ARRAY(char*, pelements, mtInternal);
1626 }
1627 } else {
1628 snprintf(buffer, buflen, "%s/lib%s.so", pname, fname);
1629 retval = true;
1630 }
1631 return retval;
1632 }
1633
1634 // check if addr is inside libjvm.so
1635 bool os::address_is_in_vm(address addr) {
1636 static address libjvm_base_addr;
1637 Dl_info dlinfo;
1638
1639 if (libjvm_base_addr == NULL) {
1640 if (dladdr(CAST_FROM_FN_PTR(void *, os::address_is_in_vm), &dlinfo) != 0) {
1641 libjvm_base_addr = (address)dlinfo.dli_fbase;
1642 }
1643 assert(libjvm_base_addr !=NULL, "Cannot obtain base address for libjvm");
1644 }
1645
1646 if (dladdr((void *)addr, &dlinfo) != 0) {
1647 if (libjvm_base_addr == (address)dlinfo.dli_fbase) return true;
1648 }
1649
1650 return false;
1651 }
1652
1653 bool os::dll_address_to_function_name(address addr, char *buf,
1654 int buflen, int *offset) {
1655 // buf is not optional, but offset is optional
1656 assert(buf != NULL, "sanity check");
1657
1658 Dl_info dlinfo;
1659
1660 if (dladdr((void*)addr, &dlinfo) != 0) {
1661 // see if we have a matching symbol
1662 if (dlinfo.dli_saddr != NULL && dlinfo.dli_sname != NULL) {
1663 if (!Decoder::demangle(dlinfo.dli_sname, buf, buflen)) {
1664 jio_snprintf(buf, buflen, "%s", dlinfo.dli_sname);
1665 }
1666 if (offset != NULL) *offset = addr - (address)dlinfo.dli_saddr;
1667 return true;
1668 }
1669 // no matching symbol so try for just file info
1670 if (dlinfo.dli_fname != NULL && dlinfo.dli_fbase != NULL) {
1671 if (Decoder::decode((address)(addr - (address)dlinfo.dli_fbase),
1672 buf, buflen, offset, dlinfo.dli_fname)) {
1673 return true;
1674 }
1675 }
1676 }
1677
1678 buf[0] = '\0';
1679 if (offset != NULL) *offset = -1;
1680 return false;
1681 }
1682
1683 struct _address_to_library_name {
1684 address addr; // input : memory address
1685 size_t buflen; // size of fname
1686 char* fname; // output: library name
1687 address base; // library base addr
1688 };
1689
1690 static int address_to_library_name_callback(struct dl_phdr_info *info,
1691 size_t size, void *data) {
1692 int i;
1693 bool found = false;
1694 address libbase = NULL;
1695 struct _address_to_library_name * d = (struct _address_to_library_name *)data;
1696
1697 // iterate through all loadable segments
1698 for (i = 0; i < info->dlpi_phnum; i++) {
1699 address segbase = (address)(info->dlpi_addr + info->dlpi_phdr[i].p_vaddr);
1700 if (info->dlpi_phdr[i].p_type == PT_LOAD) {
1701 // base address of a library is the lowest address of its loaded
1702 // segments.
1703 if (libbase == NULL || libbase > segbase) {
1704 libbase = segbase;
1705 }
1706 // see if 'addr' is within current segment
1707 if (segbase <= d->addr &&
1708 d->addr < segbase + info->dlpi_phdr[i].p_memsz) {
1709 found = true;
1710 }
1711 }
1712 }
1713
1714 // dlpi_name is NULL or empty if the ELF file is executable, return 0
1715 // so dll_address_to_library_name() can fall through to use dladdr() which
1716 // can figure out executable name from argv[0].
1717 if (found && info->dlpi_name && info->dlpi_name[0]) {
1718 d->base = libbase;
1719 if (d->fname) {
1720 jio_snprintf(d->fname, d->buflen, "%s", info->dlpi_name);
1721 }
1722 return 1;
1723 }
1724 return 0;
1725 }
1726
1727 bool os::dll_address_to_library_name(address addr, char* buf,
1728 int buflen, int* offset) {
1729 // buf is not optional, but offset is optional
1730 assert(buf != NULL, "sanity check");
1731
1732 Dl_info dlinfo;
1733 struct _address_to_library_name data;
1734
1735 // There is a bug in old glibc dladdr() implementation that it could resolve
1736 // to wrong library name if the .so file has a base address != NULL. Here
1737 // we iterate through the program headers of all loaded libraries to find
1738 // out which library 'addr' really belongs to. This workaround can be
1739 // removed once the minimum requirement for glibc is moved to 2.3.x.
1740 data.addr = addr;
1741 data.fname = buf;
1742 data.buflen = buflen;
1743 data.base = NULL;
1744 int rslt = dl_iterate_phdr(address_to_library_name_callback, (void *)&data);
1745
1746 if (rslt) {
1747 // buf already contains library name
1748 if (offset) *offset = addr - data.base;
1749 return true;
1750 }
1751 if (dladdr((void*)addr, &dlinfo) != 0) {
1752 if (dlinfo.dli_fname != NULL) {
1753 jio_snprintf(buf, buflen, "%s", dlinfo.dli_fname);
1754 }
1755 if (dlinfo.dli_fbase != NULL && offset != NULL) {
1756 *offset = addr - (address)dlinfo.dli_fbase;
1757 }
1758 return true;
1759 }
1760
1761 buf[0] = '\0';
1762 if (offset) *offset = -1;
1763 return false;
1764 }
1765
1766 // Loads .dll/.so and
1767 // in case of error it checks if .dll/.so was built for the
1768 // same architecture as Hotspot is running on
1769
1770
1771 // Remember the stack's state. The Linux dynamic linker will change
1772 // the stack to 'executable' at most once, so we must safepoint only once.
1773 bool os::Linux::_stack_is_executable = false;
1774
1775 // VM operation that loads a library. This is necessary if stack protection
1776 // of the Java stacks can be lost during loading the library. If we
1777 // do not stop the Java threads, they can stack overflow before the stacks
1778 // are protected again.
1779 class VM_LinuxDllLoad: public VM_Operation {
1780 private:
1781 const char *_filename;
1782 char *_ebuf;
1783 int _ebuflen;
1784 void *_lib;
1785 public:
1786 VM_LinuxDllLoad(const char *fn, char *ebuf, int ebuflen) :
1787 _filename(fn), _ebuf(ebuf), _ebuflen(ebuflen), _lib(NULL) {}
1788 VMOp_Type type() const { return VMOp_LinuxDllLoad; }
1789 void doit() {
1790 _lib = os::Linux::dll_load_in_vmthread(_filename, _ebuf, _ebuflen);
1791 os::Linux::_stack_is_executable = true;
1792 }
1793 void* loaded_library() { return _lib; }
1794 };
1795
1796 void * os::dll_load(const char *filename, char *ebuf, int ebuflen)
1797 {
1798 void * result = NULL;
1799 bool load_attempted = false;
1800
1801 // Check whether the library to load might change execution rights
1802 // of the stack. If they are changed, the protection of the stack
1803 // guard pages will be lost. We need a safepoint to fix this.
1804 //
1805 // See Linux man page execstack(8) for more info.
1806 if (os::uses_stack_guard_pages() && !os::Linux::_stack_is_executable) {
1807 ElfFile ef(filename);
1808 if (!ef.specifies_noexecstack()) {
1809 if (!is_init_completed()) {
1810 os::Linux::_stack_is_executable = true;
1811 // This is OK - No Java threads have been created yet, and hence no
1812 // stack guard pages to fix.
1813 //
1814 // This should happen only when you are building JDK7 using a very
1815 // old version of JDK6 (e.g., with JPRT) and running test_gamma.
1816 //
1817 // Dynamic loader will make all stacks executable after
1818 // this function returns, and will not do that again.
1819 assert(Threads::first() == NULL, "no Java threads should exist yet.");
1820 } else {
1821 warning("You have loaded library %s which might have disabled stack guard. "
1822 "The VM will try to fix the stack guard now.\n"
1823 "It's highly recommended that you fix the library with "
1824 "'execstack -c <libfile>', or link it with '-z noexecstack'.",
1825 filename);
1826
1827 assert(Thread::current()->is_Java_thread(), "must be Java thread");
1828 JavaThread *jt = JavaThread::current();
1829 if (jt->thread_state() != _thread_in_native) {
1830 // This happens when a compiler thread tries to load a hsdis-<arch>.so file
1831 // that requires ExecStack. Cannot enter safe point. Let's give up.
1832 warning("Unable to fix stack guard. Giving up.");
1833 } else {
1834 if (!LoadExecStackDllInVMThread) {
1835 // This is for the case where the DLL has an static
1836 // constructor function that executes JNI code. We cannot
1837 // load such DLLs in the VMThread.
1838 result = os::Linux::dlopen_helper(filename, ebuf, ebuflen);
1839 }
1840
1841 ThreadInVMfromNative tiv(jt);
1842 debug_only(VMNativeEntryWrapper vew;)
1843
1844 VM_LinuxDllLoad op(filename, ebuf, ebuflen);
1845 VMThread::execute(&op);
1846 if (LoadExecStackDllInVMThread) {
1847 result = op.loaded_library();
1848 }
1849 load_attempted = true;
1850 }
1851 }
1852 }
1853 }
1854
1855 if (!load_attempted) {
1856 result = os::Linux::dlopen_helper(filename, ebuf, ebuflen);
1857 }
1858
1859 if (result != NULL) {
1860 // Successful loading
1861 return result;
1862 }
1863
1864 Elf32_Ehdr elf_head;
1865 int diag_msg_max_length=ebuflen-strlen(ebuf);
1866 char* diag_msg_buf=ebuf+strlen(ebuf);
1867
1868 if (diag_msg_max_length==0) {
1869 // No more space in ebuf for additional diagnostics message
1870 return NULL;
1871 }
1872
1873
1874 int file_descriptor= ::open(filename, O_RDONLY | O_NONBLOCK);
1875
1876 if (file_descriptor < 0) {
1877 // Can't open library, report dlerror() message
1878 return NULL;
1879 }
1880
1881 bool failed_to_read_elf_head=
1882 (sizeof(elf_head)!=
1883 (::read(file_descriptor, &elf_head,sizeof(elf_head)))) ;
1884
1885 ::close(file_descriptor);
1886 if (failed_to_read_elf_head) {
1887 // file i/o error - report dlerror() msg
1888 return NULL;
1889 }
1890
1891 typedef struct {
1892 Elf32_Half code; // Actual value as defined in elf.h
1893 Elf32_Half compat_class; // Compatibility of archs at VM's sense
1894 char elf_class; // 32 or 64 bit
1895 char endianess; // MSB or LSB
1896 char* name; // String representation
1897 } arch_t;
1898
1899 #ifndef EM_486
1900 #define EM_486 6 /* Intel 80486 */
1901 #endif
1902
1903 static const arch_t arch_array[]={
1904 {EM_386, EM_386, ELFCLASS32, ELFDATA2LSB, (char*)"IA 32"},
1905 {EM_486, EM_386, ELFCLASS32, ELFDATA2LSB, (char*)"IA 32"},
1906 {EM_IA_64, EM_IA_64, ELFCLASS64, ELFDATA2LSB, (char*)"IA 64"},
1907 {EM_X86_64, EM_X86_64, ELFCLASS64, ELFDATA2LSB, (char*)"AMD 64"},
1908 {EM_SPARC, EM_SPARC, ELFCLASS32, ELFDATA2MSB, (char*)"Sparc 32"},
1909 {EM_SPARC32PLUS, EM_SPARC, ELFCLASS32, ELFDATA2MSB, (char*)"Sparc 32"},
1910 {EM_SPARCV9, EM_SPARCV9, ELFCLASS64, ELFDATA2MSB, (char*)"Sparc v9 64"},
1911 {EM_PPC, EM_PPC, ELFCLASS32, ELFDATA2MSB, (char*)"Power PC 32"},
1912 #if defined(VM_LITTLE_ENDIAN)
1913 {EM_PPC64, EM_PPC64, ELFCLASS64, ELFDATA2LSB, (char*)"Power PC 64"},
1914 #else
1915 {EM_PPC64, EM_PPC64, ELFCLASS64, ELFDATA2MSB, (char*)"Power PC 64"},
1916 #endif
1917 {EM_ARM, EM_ARM, ELFCLASS32, ELFDATA2LSB, (char*)"ARM"},
1918 {EM_S390, EM_S390, ELFCLASSNONE, ELFDATA2MSB, (char*)"IBM System/390"},
1919 {EM_ALPHA, EM_ALPHA, ELFCLASS64, ELFDATA2LSB, (char*)"Alpha"},
1920 {EM_MIPS_RS3_LE, EM_MIPS_RS3_LE, ELFCLASS32, ELFDATA2LSB, (char*)"MIPSel"},
1921 {EM_MIPS, EM_MIPS, ELFCLASS32, ELFDATA2MSB, (char*)"MIPS"},
1922 {EM_PARISC, EM_PARISC, ELFCLASS32, ELFDATA2MSB, (char*)"PARISC"},
1923 {EM_68K, EM_68K, ELFCLASS32, ELFDATA2MSB, (char*)"M68k"}
1924 };
1925
1926 #if (defined IA32)
1927 static Elf32_Half running_arch_code=EM_386;
1928 #elif (defined AMD64)
1929 static Elf32_Half running_arch_code=EM_X86_64;
1930 #elif (defined IA64)
1931 static Elf32_Half running_arch_code=EM_IA_64;
1932 #elif (defined __sparc) && (defined _LP64)
1933 static Elf32_Half running_arch_code=EM_SPARCV9;
1934 #elif (defined __sparc) && (!defined _LP64)
1935 static Elf32_Half running_arch_code=EM_SPARC;
1936 #elif (defined __powerpc64__)
1937 static Elf32_Half running_arch_code=EM_PPC64;
1938 #elif (defined __powerpc__)
1939 static Elf32_Half running_arch_code=EM_PPC;
1940 #elif (defined ARM)
1941 static Elf32_Half running_arch_code=EM_ARM;
1942 #elif (defined S390)
1943 static Elf32_Half running_arch_code=EM_S390;
1944 #elif (defined ALPHA)
1945 static Elf32_Half running_arch_code=EM_ALPHA;
1946 #elif (defined MIPSEL)
1947 static Elf32_Half running_arch_code=EM_MIPS_RS3_LE;
1948 #elif (defined PARISC)
1949 static Elf32_Half running_arch_code=EM_PARISC;
1950 #elif (defined MIPS)
1951 static Elf32_Half running_arch_code=EM_MIPS;
1952 #elif (defined M68K)
1953 static Elf32_Half running_arch_code=EM_68K;
1954 #else
1955 #error Method os::dll_load requires that one of following is defined:\
1956 IA32, AMD64, IA64, __sparc, __powerpc__, ARM, S390, ALPHA, MIPS, MIPSEL, PARISC, M68K
1957 #endif
1958
1959 // Identify compatability class for VM's architecture and library's architecture
1960 // Obtain string descriptions for architectures
1961
1962 arch_t lib_arch={elf_head.e_machine,0,elf_head.e_ident[EI_CLASS], elf_head.e_ident[EI_DATA], NULL};
1963 int running_arch_index=-1;
1964
1965 for (unsigned int i=0 ; i < ARRAY_SIZE(arch_array) ; i++ ) {
1966 if (running_arch_code == arch_array[i].code) {
1967 running_arch_index = i;
1968 }
1969 if (lib_arch.code == arch_array[i].code) {
1970 lib_arch.compat_class = arch_array[i].compat_class;
1971 lib_arch.name = arch_array[i].name;
1972 }
1973 }
1974
1975 assert(running_arch_index != -1,
1976 "Didn't find running architecture code (running_arch_code) in arch_array");
1977 if (running_arch_index == -1) {
1978 // Even though running architecture detection failed
1979 // we may still continue with reporting dlerror() message
1980 return NULL;
1981 }
1982
1983 if (lib_arch.endianess != arch_array[running_arch_index].endianess) {
1984 ::snprintf(diag_msg_buf, diag_msg_max_length-1," (Possible cause: endianness mismatch)");
1985 return NULL;
1986 }
1987
1988 #ifndef S390
1989 if (lib_arch.elf_class != arch_array[running_arch_index].elf_class) {
1990 ::snprintf(diag_msg_buf, diag_msg_max_length-1," (Possible cause: architecture word width mismatch)");
1991 return NULL;
1992 }
1993 #endif // !S390
1994
1995 if (lib_arch.compat_class != arch_array[running_arch_index].compat_class) {
1996 if ( lib_arch.name!=NULL ) {
1997 ::snprintf(diag_msg_buf, diag_msg_max_length-1,
1998 " (Possible cause: can't load %s-bit .so on a %s-bit platform)",
1999 lib_arch.name, arch_array[running_arch_index].name);
2000 } else {
2001 ::snprintf(diag_msg_buf, diag_msg_max_length-1,
2002 " (Possible cause: can't load this .so (machine code=0x%x) on a %s-bit platform)",
2003 lib_arch.code,
2004 arch_array[running_arch_index].name);
2005 }
2006 }
2007
2008 return NULL;
2009 }
2010
2011 void * os::Linux::dlopen_helper(const char *filename, char *ebuf, int ebuflen) {
2012 void * result = ::dlopen(filename, RTLD_LAZY);
2013 if (result == NULL) {
2014 ::strncpy(ebuf, ::dlerror(), ebuflen - 1);
2015 ebuf[ebuflen-1] = '\0';
2016 }
2017 return result;
2018 }
2019
2020 void * os::Linux::dll_load_in_vmthread(const char *filename, char *ebuf, int ebuflen) {
2021 void * result = NULL;
2022 if (LoadExecStackDllInVMThread) {
2023 result = dlopen_helper(filename, ebuf, ebuflen);
2024 }
2025
2026 // Since 7019808, libjvm.so is linked with -noexecstack. If the VM loads a
2027 // library that requires an executable stack, or which does not have this
2028 // stack attribute set, dlopen changes the stack attribute to executable. The
2029 // read protection of the guard pages gets lost.
2030 //
2031 // Need to check _stack_is_executable again as multiple VM_LinuxDllLoad
2032 // may have been queued at the same time.
2033
2034 if (!_stack_is_executable) {
2035 JavaThread *jt = Threads::first();
2036
2037 while (jt) {
2038 if (!jt->stack_guard_zone_unused() && // Stack not yet fully initialized
2039 jt->stack_yellow_zone_enabled()) { // No pending stack overflow exceptions
2040 if (!os::guard_memory((char *) jt->stack_red_zone_base() - jt->stack_red_zone_size(),
2041 jt->stack_yellow_zone_size() + jt->stack_red_zone_size())) {
2042 warning("Attempt to reguard stack yellow zone failed.");
2043 }
2044 }
2045 jt = jt->next();
2046 }
2047 }
2048
2049 return result;
2050 }
2051
2052 /*
2053 * glibc-2.0 libdl is not MT safe. If you are building with any glibc,
2054 * chances are you might want to run the generated bits against glibc-2.0
2055 * libdl.so, so always use locking for any version of glibc.
2056 */
2057 void* os::dll_lookup(void* handle, const char* name) {
2058 pthread_mutex_lock(&dl_mutex);
2059 void* res = dlsym(handle, name);
2060 pthread_mutex_unlock(&dl_mutex);
2061 return res;
2062 }
2063
2064 void* os::get_default_process_handle() {
2065 return (void*)::dlopen(NULL, RTLD_LAZY);
2066 }
2067
2068 static bool _print_ascii_file(const char* filename, outputStream* st) {
2069 int fd = ::open(filename, O_RDONLY);
2070 if (fd == -1) {
2071 return false;
2072 }
2073
2074 char buf[32];
2075 int bytes;
2076 while ((bytes = ::read(fd, buf, sizeof(buf))) > 0) {
2077 st->print_raw(buf, bytes);
2078 }
2079
2080 ::close(fd);
2081
2082 return true;
2083 }
2084
2085 void os::print_dll_info(outputStream *st) {
2086 st->print_cr("Dynamic libraries:");
2087
2088 char fname[32];
2089 pid_t pid = os::Linux::gettid();
2090
2091 jio_snprintf(fname, sizeof(fname), "/proc/%d/maps", pid);
2092
2093 if (!_print_ascii_file(fname, st)) {
2094 st->print("Can not get library information for pid = %d\n", pid);
2095 }
2096 }
2097
2098 void os::print_os_info_brief(outputStream* st) {
2099 os::Linux::print_distro_info(st);
2100
2101 os::Posix::print_uname_info(st);
2102
2103 os::Linux::print_libversion_info(st);
2104
2105 }
2106
2107 void os::print_os_info(outputStream* st) {
2108 st->print("OS:");
2109
2110 os::Linux::print_distro_info(st);
2111
2112 os::Posix::print_uname_info(st);
2113
2114 // Print warning if unsafe chroot environment detected
2115 if (unsafe_chroot_detected) {
2116 st->print("WARNING!! ");
2117 st->print_cr("%s", unstable_chroot_error);
2118 }
2119
2120 os::Linux::print_libversion_info(st);
2121
2122 os::Posix::print_rlimit_info(st);
2123
2124 os::Posix::print_load_average(st);
2125
2126 os::Linux::print_full_memory_info(st);
2127 }
2128
2129 // Try to identify popular distros.
2130 // Most Linux distributions have a /etc/XXX-release file, which contains
2131 // the OS version string. Newer Linux distributions have a /etc/lsb-release
2132 // file that also contains the OS version string. Some have more than one
2133 // /etc/XXX-release file (e.g. Mandrake has both /etc/mandrake-release and
2134 // /etc/redhat-release.), so the order is important.
2135 // Any Linux that is based on Redhat (i.e. Oracle, Mandrake, Sun JDS...) have
2136 // their own specific XXX-release file as well as a redhat-release file.
2137 // Because of this the XXX-release file needs to be searched for before the
2138 // redhat-release file.
2139 // Since Red Hat has a lsb-release file that is not very descriptive the
2140 // search for redhat-release needs to be before lsb-release.
2141 // Since the lsb-release file is the new standard it needs to be searched
2142 // before the older style release files.
2143 // Searching system-release (Red Hat) and os-release (other Linuxes) are a
2144 // next to last resort. The os-release file is a new standard that contains
2145 // distribution information and the system-release file seems to be an old
2146 // standard that has been replaced by the lsb-release and os-release files.
2147 // Searching for the debian_version file is the last resort. It contains
2148 // an informative string like "6.0.6" or "wheezy/sid". Because of this
2149 // "Debian " is printed before the contents of the debian_version file.
2150 void os::Linux::print_distro_info(outputStream* st) {
2151 if (!_print_ascii_file("/etc/oracle-release", st) &&
2152 !_print_ascii_file("/etc/mandriva-release", st) &&
2153 !_print_ascii_file("/etc/mandrake-release", st) &&
2154 !_print_ascii_file("/etc/sun-release", st) &&
2155 !_print_ascii_file("/etc/redhat-release", st) &&
2156 !_print_ascii_file("/etc/lsb-release", st) &&
2157 !_print_ascii_file("/etc/SuSE-release", st) &&
2158 !_print_ascii_file("/etc/turbolinux-release", st) &&
2159 !_print_ascii_file("/etc/gentoo-release", st) &&
2160 !_print_ascii_file("/etc/ltib-release", st) &&
2161 !_print_ascii_file("/etc/angstrom-version", st) &&
2162 !_print_ascii_file("/etc/system-release", st) &&
2163 !_print_ascii_file("/etc/os-release", st)) {
2164
2165 if (file_exists("/etc/debian_version")) {
2166 st->print("Debian ");
2167 _print_ascii_file("/etc/debian_version", st);
2168 } else {
2169 st->print("Linux");
2170 }
2171 }
2172 st->cr();
2173 }
2174
2175 void os::Linux::print_libversion_info(outputStream* st) {
2176 // libc, pthread
2177 st->print("libc:");
2178 st->print("%s ", os::Linux::glibc_version());
2179 st->print("%s ", os::Linux::libpthread_version());
2180 if (os::Linux::is_LinuxThreads()) {
2181 st->print("(%s stack)", os::Linux::is_floating_stack() ? "floating" : "fixed");
2182 }
2183 st->cr();
2184 }
2185
2186 void os::Linux::print_full_memory_info(outputStream* st) {
2187 st->print("\n/proc/meminfo:\n");
2188 _print_ascii_file("/proc/meminfo", st);
2189 st->cr();
2190 }
2191
2192 void os::print_memory_info(outputStream* st) {
2193
2194 st->print("Memory:");
2195 st->print(" %dk page", os::vm_page_size()>>10);
2196
2197 // values in struct sysinfo are "unsigned long"
2198 struct sysinfo si;
2199 sysinfo(&si);
2200
2201 st->print(", physical " UINT64_FORMAT "k",
2202 os::physical_memory() >> 10);
2203 st->print("(" UINT64_FORMAT "k free)",
2204 os::available_memory() >> 10);
2205 st->print(", swap " UINT64_FORMAT "k",
2206 ((jlong)si.totalswap * si.mem_unit) >> 10);
2207 st->print("(" UINT64_FORMAT "k free)",
2208 ((jlong)si.freeswap * si.mem_unit) >> 10);
2209 st->cr();
2210 }
2211
2212 void os::pd_print_cpu_info(outputStream* st) {
2213 st->print("\n/proc/cpuinfo:\n");
2214 if (!_print_ascii_file("/proc/cpuinfo", st)) {
2215 st->print(" <Not Available>");
2216 }
2217 st->cr();
2218 }
2219
2220 void os::print_siginfo(outputStream* st, void* siginfo) {
2221 const siginfo_t* si = (const siginfo_t*)siginfo;
2222
2223 os::Posix::print_siginfo_brief(st, si);
2224 #if INCLUDE_CDS
2225 if (si && (si->si_signo == SIGBUS || si->si_signo == SIGSEGV) &&
2226 UseSharedSpaces) {
2227 FileMapInfo* mapinfo = FileMapInfo::current_info();
2228 if (mapinfo->is_in_shared_space(si->si_addr)) {
2229 st->print("\n\nError accessing class data sharing archive." \
2230 " Mapped file inaccessible during execution, " \
2231 " possible disk/network problem.");
2232 }
2233 }
2234 #endif
2235 st->cr();
2236 }
2237
2238
2239 static void print_signal_handler(outputStream* st, int sig,
2240 char* buf, size_t buflen);
2241
2242 void os::print_signal_handlers(outputStream* st, char* buf, size_t buflen) {
2243 st->print_cr("Signal Handlers:");
2244 print_signal_handler(st, SIGSEGV, buf, buflen);
2245 print_signal_handler(st, SIGBUS , buf, buflen);
2246 print_signal_handler(st, SIGFPE , buf, buflen);
2247 print_signal_handler(st, SIGPIPE, buf, buflen);
2248 print_signal_handler(st, SIGXFSZ, buf, buflen);
2249 print_signal_handler(st, SIGILL , buf, buflen);
2250 print_signal_handler(st, INTERRUPT_SIGNAL, buf, buflen);
2251 print_signal_handler(st, SR_signum, buf, buflen);
2252 print_signal_handler(st, SHUTDOWN1_SIGNAL, buf, buflen);
2253 print_signal_handler(st, SHUTDOWN2_SIGNAL , buf, buflen);
2254 print_signal_handler(st, SHUTDOWN3_SIGNAL , buf, buflen);
2255 print_signal_handler(st, BREAK_SIGNAL, buf, buflen);
2256 #if defined(PPC64)
2257 print_signal_handler(st, SIGTRAP, buf, buflen);
2258 #endif
2259 }
2260
2261 static char saved_jvm_path[MAXPATHLEN] = {0};
2262
2263 // Find the full path to the current module, libjvm.so
2264 void os::jvm_path(char *buf, jint buflen) {
2265 // Error checking.
2266 if (buflen < MAXPATHLEN) {
2267 assert(false, "must use a large-enough buffer");
2268 buf[0] = '\0';
2269 return;
2270 }
2271 // Lazy resolve the path to current module.
2272 if (saved_jvm_path[0] != 0) {
2273 strcpy(buf, saved_jvm_path);
2274 return;
2275 }
2276
2277 char dli_fname[MAXPATHLEN];
2278 bool ret = dll_address_to_library_name(
2279 CAST_FROM_FN_PTR(address, os::jvm_path),
2280 dli_fname, sizeof(dli_fname), NULL);
2281 assert(ret, "cannot locate libjvm");
2282 char *rp = NULL;
2283 if (ret && dli_fname[0] != '\0') {
2284 rp = realpath(dli_fname, buf);
2285 }
2286 if (rp == NULL)
2287 return;
2288
2289 if (Arguments::created_by_gamma_launcher()) {
2290 // Support for the gamma launcher. Typical value for buf is
2291 // "<JAVA_HOME>/jre/lib/<arch>/<vmtype>/libjvm.so". If "/jre/lib/" appears at
2292 // the right place in the string, then assume we are installed in a JDK and
2293 // we're done. Otherwise, check for a JAVA_HOME environment variable and fix
2294 // up the path so it looks like libjvm.so is installed there (append a
2295 // fake suffix hotspot/libjvm.so).
2296 const char *p = buf + strlen(buf) - 1;
2297 for (int count = 0; p > buf && count < 5; ++count) {
2298 for (--p; p > buf && *p != '/'; --p)
2299 /* empty */ ;
2300 }
2301
2302 if (strncmp(p, "/jre/lib/", 9) != 0) {
2303 // Look for JAVA_HOME in the environment.
2304 char* java_home_var = ::getenv("JAVA_HOME");
2305 if (java_home_var != NULL && java_home_var[0] != 0) {
2306 char* jrelib_p;
2307 int len;
2308
2309 // Check the current module name "libjvm.so".
2310 p = strrchr(buf, '/');
2311 assert(strstr(p, "/libjvm") == p, "invalid library name");
2312
2313 rp = realpath(java_home_var, buf);
2314 if (rp == NULL)
2315 return;
2316
2317 // determine if this is a legacy image or modules image
2318 // modules image doesn't have "jre" subdirectory
2319 len = strlen(buf);
2320 assert(len < buflen, "Ran out of buffer room");
2321 jrelib_p = buf + len;
2322 snprintf(jrelib_p, buflen-len, "/jre/lib/%s", cpu_arch);
2323 if (0 != access(buf, F_OK)) {
2324 snprintf(jrelib_p, buflen-len, "/lib/%s", cpu_arch);
2325 }
2326
2327 if (0 == access(buf, F_OK)) {
2328 // Use current module name "libjvm.so"
2329 len = strlen(buf);
2330 snprintf(buf + len, buflen-len, "/hotspot/libjvm.so");
2331 } else {
2332 // Go back to path of .so
2333 rp = realpath(dli_fname, buf);
2334 if (rp == NULL)
2335 return;
2336 }
2337 }
2338 }
2339 }
2340
2341 strncpy(saved_jvm_path, buf, MAXPATHLEN);
2342 }
2343
2344 void os::print_jni_name_prefix_on(outputStream* st, int args_size) {
2345 // no prefix required, not even "_"
2346 }
2347
2348 void os::print_jni_name_suffix_on(outputStream* st, int args_size) {
2349 // no suffix required
2350 }
2351
2352 ////////////////////////////////////////////////////////////////////////////////
2353 // sun.misc.Signal support
2354
2355 static volatile jint sigint_count = 0;
2356
2357 static void
2358 UserHandler(int sig, void *siginfo, void *context) {
2359 // 4511530 - sem_post is serialized and handled by the manager thread. When
2360 // the program is interrupted by Ctrl-C, SIGINT is sent to every thread. We
2361 // don't want to flood the manager thread with sem_post requests.
2362 if (sig == SIGINT && Atomic::add(1, &sigint_count) > 1)
2363 return;
2364
2365 // Ctrl-C is pressed during error reporting, likely because the error
2366 // handler fails to abort. Let VM die immediately.
2367 if (sig == SIGINT && is_error_reported()) {
2368 os::die();
2369 }
2370
2371 os::signal_notify(sig);
2372 }
2373
2374 void* os::user_handler() {
2375 return CAST_FROM_FN_PTR(void*, UserHandler);
2376 }
2377
2378 class Semaphore : public StackObj {
2379 public:
2380 Semaphore();
2381 ~Semaphore();
2382 void signal();
2383 void wait();
2384 bool trywait();
2385 bool timedwait(unsigned int sec, int nsec);
2386 private:
2387 sem_t _semaphore;
2388 };
2389
2390 Semaphore::Semaphore() {
2391 sem_init(&_semaphore, 0, 0);
2392 }
2393
2394 Semaphore::~Semaphore() {
2395 sem_destroy(&_semaphore);
2396 }
2397
2398 void Semaphore::signal() {
2399 sem_post(&_semaphore);
2400 }
2401
2402 void Semaphore::wait() {
2403 sem_wait(&_semaphore);
2404 }
2405
2406 bool Semaphore::trywait() {
2407 return sem_trywait(&_semaphore) == 0;
2408 }
2409
2410 bool Semaphore::timedwait(unsigned int sec, int nsec) {
2411
2412 struct timespec ts;
2413 // Semaphore's are always associated with CLOCK_REALTIME
2414 os::Linux::clock_gettime(CLOCK_REALTIME, &ts);
2415 // see unpackTime for discussion on overflow checking
2416 if (sec >= MAX_SECS) {
2417 ts.tv_sec += MAX_SECS;
2418 ts.tv_nsec = 0;
2419 } else {
2420 ts.tv_sec += sec;
2421 ts.tv_nsec += nsec;
2422 if (ts.tv_nsec >= NANOSECS_PER_SEC) {
2423 ts.tv_nsec -= NANOSECS_PER_SEC;
2424 ++ts.tv_sec; // note: this must be <= max_secs
2425 }
2426 }
2427
2428 while (1) {
2429 int result = sem_timedwait(&_semaphore, &ts);
2430 if (result == 0) {
2431 return true;
2432 } else if (errno == EINTR) {
2433 continue;
2434 } else if (errno == ETIMEDOUT) {
2435 return false;
2436 } else {
2437 return false;
2438 }
2439 }
2440 }
2441
2442 extern "C" {
2443 typedef void (*sa_handler_t)(int);
2444 typedef void (*sa_sigaction_t)(int, siginfo_t *, void *);
2445 }
2446
2447 void* os::signal(int signal_number, void* handler) {
2448 struct sigaction sigAct, oldSigAct;
2449
2450 sigfillset(&(sigAct.sa_mask));
2451 sigAct.sa_flags = SA_RESTART|SA_SIGINFO;
2452 sigAct.sa_handler = CAST_TO_FN_PTR(sa_handler_t, handler);
2453
2454 if (sigaction(signal_number, &sigAct, &oldSigAct)) {
2455 // -1 means registration failed
2456 return (void *)-1;
2457 }
2458
2459 return CAST_FROM_FN_PTR(void*, oldSigAct.sa_handler);
2460 }
2461
2462 void os::signal_raise(int signal_number) {
2463 ::raise(signal_number);
2464 }
2465
2466 /*
2467 * The following code is moved from os.cpp for making this
2468 * code platform specific, which it is by its very nature.
2469 */
2470
2471 // Will be modified when max signal is changed to be dynamic
2472 int os::sigexitnum_pd() {
2473 return NSIG;
2474 }
2475
2476 // a counter for each possible signal value
2477 static volatile jint pending_signals[NSIG+1] = { 0 };
2478
2479 // Linux(POSIX) specific hand shaking semaphore.
2480 static sem_t sig_sem;
2481 static Semaphore sr_semaphore;
2482
2483 void os::signal_init_pd() {
2484 // Initialize signal structures
2485 ::memset((void*)pending_signals, 0, sizeof(pending_signals));
2486
2487 // Initialize signal semaphore
2488 ::sem_init(&sig_sem, 0, 0);
2489 }
2490
2491 void os::signal_notify(int sig) {
2492 Atomic::inc(&pending_signals[sig]);
2493 ::sem_post(&sig_sem);
2494 }
2495
2496 static int check_pending_signals(bool wait) {
2497 Atomic::store(0, &sigint_count);
2498 for (;;) {
2499 for (int i = 0; i < NSIG + 1; i++) {
2500 jint n = pending_signals[i];
2501 if (n > 0 && n == Atomic::cmpxchg(n - 1, &pending_signals[i], n)) {
2502 return i;
2503 }
2504 }
2505 if (!wait) {
2506 return -1;
2507 }
2508 JavaThread *thread = JavaThread::current();
2509 ThreadBlockInVM tbivm(thread);
2510
2511 bool threadIsSuspended;
2512 do {
2513 thread->set_suspend_equivalent();
2514 // cleared by handle_special_suspend_equivalent_condition() or java_suspend_self()
2515 ::sem_wait(&sig_sem);
2516
2517 // were we externally suspended while we were waiting?
2518 threadIsSuspended = thread->handle_special_suspend_equivalent_condition();
2519 if (threadIsSuspended) {
2520 //
2521 // The semaphore has been incremented, but while we were waiting
2522 // another thread suspended us. We don't want to continue running
2523 // while suspended because that would surprise the thread that
2524 // suspended us.
2525 //
2526 ::sem_post(&sig_sem);
2527
2528 thread->java_suspend_self();
2529 }
2530 } while (threadIsSuspended);
2531 }
2532 }
2533
2534 int os::signal_lookup() {
2535 return check_pending_signals(false);
2536 }
2537
2538 int os::signal_wait() {
2539 return check_pending_signals(true);
2540 }
2541
2542 ////////////////////////////////////////////////////////////////////////////////
2543 // Virtual Memory
2544
2545 int os::vm_page_size() {
2546 // Seems redundant as all get out
2547 assert(os::Linux::page_size() != -1, "must call os::init");
2548 return os::Linux::page_size();
2549 }
2550
2551 // Solaris allocates memory by pages.
2552 int os::vm_allocation_granularity() {
2553 assert(os::Linux::page_size() != -1, "must call os::init");
2554 return os::Linux::page_size();
2555 }
2556
2557 // Rationale behind this function:
2558 // current (Mon Apr 25 20:12:18 MSD 2005) oprofile drops samples without executable
2559 // mapping for address (see lookup_dcookie() in the kernel module), thus we cannot get
2560 // samples for JITted code. Here we create private executable mapping over the code cache
2561 // and then we can use standard (well, almost, as mapping can change) way to provide
2562 // info for the reporting script by storing timestamp and location of symbol
2563 void linux_wrap_code(char* base, size_t size) {
2564 static volatile jint cnt = 0;
2565
2566 if (!UseOprofile) {
2567 return;
2568 }
2569
2570 char buf[PATH_MAX+1];
2571 int num = Atomic::add(1, &cnt);
2572
2573 snprintf(buf, sizeof(buf), "%s/hs-vm-%d-%d",
2574 os::get_temp_directory(), os::current_process_id(), num);
2575 unlink(buf);
2576
2577 int fd = ::open(buf, O_CREAT | O_RDWR, S_IRWXU);
2578
2579 if (fd != -1) {
2580 off_t rv = ::lseek(fd, size-2, SEEK_SET);
2581 if (rv != (off_t)-1) {
2582 if (::write(fd, "", 1) == 1) {
2583 mmap(base, size,
2584 PROT_READ|PROT_WRITE|PROT_EXEC,
2585 MAP_PRIVATE|MAP_FIXED|MAP_NORESERVE, fd, 0);
2586 }
2587 }
2588 ::close(fd);
2589 unlink(buf);
2590 }
2591 }
2592
2593 static bool recoverable_mmap_error(int err) {
2594 // See if the error is one we can let the caller handle. This
2595 // list of errno values comes from JBS-6843484. I can't find a
2596 // Linux man page that documents this specific set of errno
2597 // values so while this list currently matches Solaris, it may
2598 // change as we gain experience with this failure mode.
2599 switch (err) {
2600 case EBADF:
2601 case EINVAL:
2602 case ENOTSUP:
2603 // let the caller deal with these errors
2604 return true;
2605
2606 default:
2607 // Any remaining errors on this OS can cause our reserved mapping
2608 // to be lost. That can cause confusion where different data
2609 // structures think they have the same memory mapped. The worst
2610 // scenario is if both the VM and a library think they have the
2611 // same memory mapped.
2612 return false;
2613 }
2614 }
2615
2616 static void warn_fail_commit_memory(char* addr, size_t size, bool exec,
2617 int err) {
2618 warning("INFO: os::commit_memory(" PTR_FORMAT ", " SIZE_FORMAT
2619 ", %d) failed; error='%s' (errno=%d)", addr, size, exec,
2620 strerror(err), err);
2621 }
2622
2623 static void warn_fail_commit_memory(char* addr, size_t size,
2624 size_t alignment_hint, bool exec,
2625 int err) {
2626 warning("INFO: os::commit_memory(" PTR_FORMAT ", " SIZE_FORMAT
2627 ", " SIZE_FORMAT ", %d) failed; error='%s' (errno=%d)", addr, size,
2628 alignment_hint, exec, strerror(err), err);
2629 }
2630
2631 // NOTE: Linux kernel does not really reserve the pages for us.
2632 // All it does is to check if there are enough free pages
2633 // left at the time of mmap(). This could be a potential
2634 // problem.
2635 int os::Linux::commit_memory_impl(char* addr, size_t size, bool exec) {
2636 int prot = exec ? PROT_READ|PROT_WRITE|PROT_EXEC : PROT_READ|PROT_WRITE;
2637 uintptr_t res = (uintptr_t) ::mmap(addr, size, prot,
2638 MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0);
2639 if (res != (uintptr_t) MAP_FAILED) {
2640 if (UseNUMAInterleaving) {
2641 numa_make_global(addr, size);
2642 }
2643 return 0;
2644 }
2645
2646 int err = errno; // save errno from mmap() call above
2647
2648 if (!recoverable_mmap_error(err)) {
2649 warn_fail_commit_memory(addr, size, exec, err);
2650 vm_exit_out_of_memory(size, OOM_MMAP_ERROR, "committing reserved memory.");
2651 }
2652
2653 return err;
2654 }
2655
2656 bool os::pd_commit_memory(char* addr, size_t size, bool exec) {
2657 return os::Linux::commit_memory_impl(addr, size, exec) == 0;
2658 }
2659
2660 void os::pd_commit_memory_or_exit(char* addr, size_t size, bool exec,
2661 const char* mesg) {
2662 assert(mesg != NULL, "mesg must be specified");
2663 int err = os::Linux::commit_memory_impl(addr, size, exec);
2664 if (err != 0) {
2665 // the caller wants all commit errors to exit with the specified mesg:
2666 warn_fail_commit_memory(addr, size, exec, err);
2667 vm_exit_out_of_memory(size, OOM_MMAP_ERROR, mesg);
2668 }
2669 }
2670
2671 // Define MAP_HUGETLB here so we can build HotSpot on old systems.
2672 #ifndef MAP_HUGETLB
2673 #define MAP_HUGETLB 0x40000
2674 #endif
2675
2676 // Define MADV_HUGEPAGE here so we can build HotSpot on old systems.
2677 #ifndef MADV_HUGEPAGE
2678 #define MADV_HUGEPAGE 14
2679 #endif
2680
2681 int os::Linux::commit_memory_impl(char* addr, size_t size,
2682 size_t alignment_hint, bool exec) {
2683 int err = os::Linux::commit_memory_impl(addr, size, exec);
2684 if (err == 0) {
2685 realign_memory(addr, size, alignment_hint);
2686 }
2687 return err;
2688 }
2689
2690 bool os::pd_commit_memory(char* addr, size_t size, size_t alignment_hint,
2691 bool exec) {
2692 return os::Linux::commit_memory_impl(addr, size, alignment_hint, exec) == 0;
2693 }
2694
2695 void os::pd_commit_memory_or_exit(char* addr, size_t size,
2696 size_t alignment_hint, bool exec,
2697 const char* mesg) {
2698 assert(mesg != NULL, "mesg must be specified");
2699 int err = os::Linux::commit_memory_impl(addr, size, alignment_hint, exec);
2700 if (err != 0) {
2701 // the caller wants all commit errors to exit with the specified mesg:
2702 warn_fail_commit_memory(addr, size, alignment_hint, exec, err);
2703 vm_exit_out_of_memory(size, OOM_MMAP_ERROR, mesg);
2704 }
2705 }
2706
2707 void os::pd_realign_memory(char *addr, size_t bytes, size_t alignment_hint) {
2708 if (UseTransparentHugePages && alignment_hint > (size_t)vm_page_size()) {
2709 // We don't check the return value: madvise(MADV_HUGEPAGE) may not
2710 // be supported or the memory may already be backed by huge pages.
2711 ::madvise(addr, bytes, MADV_HUGEPAGE);
2712 }
2713 }
2714
2715 void os::pd_free_memory(char *addr, size_t bytes, size_t alignment_hint) {
2716 // This method works by doing an mmap over an existing mmaping and effectively discarding
2717 // the existing pages. However it won't work for SHM-based large pages that cannot be
2718 // uncommitted at all. We don't do anything in this case to avoid creating a segment with
2719 // small pages on top of the SHM segment. This method always works for small pages, so we
2720 // allow that in any case.
2721 if (alignment_hint <= (size_t)os::vm_page_size() || can_commit_large_page_memory()) {
2722 commit_memory(addr, bytes, alignment_hint, !ExecMem);
2723 }
2724 }
2725
2726 void os::numa_make_global(char *addr, size_t bytes) {
2727 Linux::numa_interleave_memory(addr, bytes);
2728 }
2729
2730 // Define for numa_set_bind_policy(int). Setting the argument to 0 will set the
2731 // bind policy to MPOL_PREFERRED for the current thread.
2732 #define USE_MPOL_PREFERRED 0
2733
2734 void os::numa_make_local(char *addr, size_t bytes, int lgrp_hint) {
2735 // To make NUMA and large pages more robust when both enabled, we need to ease
2736 // the requirements on where the memory should be allocated. MPOL_BIND is the
2737 // default policy and it will force memory to be allocated on the specified
2738 // node. Changing this to MPOL_PREFERRED will prefer to allocate the memory on
2739 // the specified node, but will not force it. Using this policy will prevent
2740 // getting SIGBUS when trying to allocate large pages on NUMA nodes with no
2741 // free large pages.
2742 Linux::numa_set_bind_policy(USE_MPOL_PREFERRED);
2743 Linux::numa_tonode_memory(addr, bytes, lgrp_hint);
2744 }
2745
2746 bool os::numa_topology_changed() { return false; }
2747
2748 size_t os::numa_get_groups_num() {
2749 // Return just the number of nodes in which it's possible to allocate memory
2750 // (in numa terminology, configured nodes).
2751 return Linux::numa_num_configured_nodes();
2752 }
2753
2754 int os::numa_get_group_id() {
2755 int cpu_id = Linux::sched_getcpu();
2756 if (cpu_id != -1) {
2757 int lgrp_id = Linux::get_node_by_cpu(cpu_id);
2758 if (lgrp_id != -1) {
2759 return lgrp_id;
2760 }
2761 }
2762 return 0;
2763 }
2764
2765 int os::Linux::get_existing_num_nodes() {
2766 size_t node;
2767 size_t highest_node_number = Linux::numa_max_node();
2768 int num_nodes = 0;
2769
2770 // Get the total number of nodes in the system including nodes without memory.
2771 for (node = 0; node <= highest_node_number; node++) {
2772 if (isnode_in_existing_nodes(node)) {
2773 num_nodes++;
2774 }
2775 }
2776 return num_nodes;
2777 }
2778
2779 size_t os::numa_get_leaf_groups(int *ids, size_t size) {
2780 size_t highest_node_number = Linux::numa_max_node();
2781 size_t i = 0;
2782
2783 // Map all node ids in which is possible to allocate memory. Also nodes are
2784 // not always consecutively available, i.e. available from 0 to the highest
2785 // node number.
2786 for (size_t node = 0; node <= highest_node_number; node++) {
2787 if (Linux::isnode_in_configured_nodes(node)) {
2788 ids[i++] = node;
2789 }
2790 }
2791 return i;
2792 }
2793
2794 bool os::get_page_info(char *start, page_info* info) {
2795 return false;
2796 }
2797
2798 char *os::scan_pages(char *start, char* end, page_info* page_expected, page_info* page_found) {
2799 return end;
2800 }
2801
2802
2803 int os::Linux::sched_getcpu_syscall(void) {
2804 unsigned int cpu = 0;
2805 int retval = -1;
2806
2807 #if defined(IA32)
2808 # ifndef SYS_getcpu
2809 # define SYS_getcpu 318
2810 # endif
2811 retval = syscall(SYS_getcpu, &cpu, NULL, NULL);
2812 #elif defined(AMD64)
2813 // Unfortunately we have to bring all these macros here from vsyscall.h
2814 // to be able to compile on old linuxes.
2815 # define __NR_vgetcpu 2
2816 # define VSYSCALL_START (-10UL << 20)
2817 # define VSYSCALL_SIZE 1024
2818 # define VSYSCALL_ADDR(vsyscall_nr) (VSYSCALL_START+VSYSCALL_SIZE*(vsyscall_nr))
2819 typedef long (*vgetcpu_t)(unsigned int *cpu, unsigned int *node, unsigned long *tcache);
2820 vgetcpu_t vgetcpu = (vgetcpu_t)VSYSCALL_ADDR(__NR_vgetcpu);
2821 retval = vgetcpu(&cpu, NULL, NULL);
2822 #endif
2823
2824 return (retval == -1) ? retval : cpu;
2825 }
2826
2827 // Something to do with the numa-aware allocator needs these symbols
2828 extern "C" JNIEXPORT void numa_warn(int number, char *where, ...) { }
2829 extern "C" JNIEXPORT void numa_error(char *where) { }
2830 extern "C" JNIEXPORT int fork1() { return fork(); }
2831
2832 // Handle request to load libnuma symbol version 1.1 (API v1). If it fails
2833 // load symbol from base version instead.
2834 void* os::Linux::libnuma_dlsym(void* handle, const char *name) {
2835 void *f = dlvsym(handle, name, "libnuma_1.1");
2836 if (f == NULL) {
2837 f = dlsym(handle, name);
2838 }
2839 return f;
2840 }
2841
2842 // Handle request to load libnuma symbol version 1.2 (API v2) only.
2843 // Return NULL if the symbol is not defined in this particular version.
2844 void* os::Linux::libnuma_v2_dlsym(void* handle, const char* name) {
2845 return dlvsym(handle, name, "libnuma_1.2");
2846 }
2847
2848 bool os::Linux::libnuma_init() {
2849 // sched_getcpu() should be in libc.
2850 set_sched_getcpu(CAST_TO_FN_PTR(sched_getcpu_func_t,
2851 dlsym(RTLD_DEFAULT, "sched_getcpu")));
2852
2853 // If it's not, try a direct syscall.
2854 if (sched_getcpu() == -1)
2855 set_sched_getcpu(CAST_TO_FN_PTR(sched_getcpu_func_t, (void*)&sched_getcpu_syscall));
2856
2857 if (sched_getcpu() != -1) { // Does it work?
2858 void *handle = dlopen("libnuma.so.1", RTLD_LAZY);
2859 if (handle != NULL) {
2860 set_numa_node_to_cpus(CAST_TO_FN_PTR(numa_node_to_cpus_func_t,
2861 libnuma_dlsym(handle, "numa_node_to_cpus")));
2862 set_numa_max_node(CAST_TO_FN_PTR(numa_max_node_func_t,
2863 libnuma_dlsym(handle, "numa_max_node")));
2864 set_numa_num_configured_nodes(CAST_TO_FN_PTR(numa_num_configured_nodes_func_t,
2865 libnuma_dlsym(handle, "numa_num_configured_nodes")));
2866 set_numa_available(CAST_TO_FN_PTR(numa_available_func_t,
2867 libnuma_dlsym(handle, "numa_available")));
2868 set_numa_tonode_memory(CAST_TO_FN_PTR(numa_tonode_memory_func_t,
2869 libnuma_dlsym(handle, "numa_tonode_memory")));
2870 set_numa_interleave_memory(CAST_TO_FN_PTR(numa_interleave_memory_func_t,
2871 libnuma_dlsym(handle, "numa_interleave_memory")));
2872 set_numa_interleave_memory_v2(CAST_TO_FN_PTR(numa_interleave_memory_v2_func_t,
2873 libnuma_v2_dlsym(handle, "numa_interleave_memory")));
2874 set_numa_set_bind_policy(CAST_TO_FN_PTR(numa_set_bind_policy_func_t,
2875 libnuma_dlsym(handle, "numa_set_bind_policy")));
2876 set_numa_bitmask_isbitset(CAST_TO_FN_PTR(numa_bitmask_isbitset_func_t,
2877 libnuma_dlsym(handle, "numa_bitmask_isbitset")));
2878 set_numa_distance(CAST_TO_FN_PTR(numa_distance_func_t,
2879 libnuma_dlsym(handle, "numa_distance")));
2880
2881 if (numa_available() != -1) {
2882 set_numa_all_nodes((unsigned long*)libnuma_dlsym(handle, "numa_all_nodes"));
2883 set_numa_all_nodes_ptr((struct bitmask **)libnuma_dlsym(handle, "numa_all_nodes_ptr"));
2884 set_numa_nodes_ptr((struct bitmask **)libnuma_dlsym(handle, "numa_nodes_ptr"));
2885 // Create an index -> node mapping, since nodes are not always consecutive
2886 _nindex_to_node = new (ResourceObj::C_HEAP, mtInternal) GrowableArray<int>(0, true);
2887 rebuild_nindex_to_node_map();
2888 // Create a cpu -> node mapping
2889 _cpu_to_node = new (ResourceObj::C_HEAP, mtInternal) GrowableArray<int>(0, true);
2890 rebuild_cpu_to_node_map();
2891 return true;
2892 }
2893 }
2894 }
2895 return false;
2896 }
2897
2898 void os::Linux::rebuild_nindex_to_node_map() {
2899 int highest_node_number = Linux::numa_max_node();
2900
2901 nindex_to_node()->clear();
2902 for (int node = 0; node <= highest_node_number; node++) {
2903 if (Linux::isnode_in_existing_nodes(node)) {
2904 nindex_to_node()->append(node);
2905 }
2906 }
2907 }
2908
2909 // rebuild_cpu_to_node_map() constructs a table mapping cpud id to node id.
2910 // The table is later used in get_node_by_cpu().
2911 void os::Linux::rebuild_cpu_to_node_map() {
2912 const size_t NCPUS = 32768; // Since the buffer size computation is very obscure
2913 // in libnuma (possible values are starting from 16,
2914 // and continuing up with every other power of 2, but less
2915 // than the maximum number of CPUs supported by kernel), and
2916 // is a subject to change (in libnuma version 2 the requirements
2917 // are more reasonable) we'll just hardcode the number they use
2918 // in the library.
2919 const size_t BitsPerCLong = sizeof(long) * CHAR_BIT;
2920
2921 size_t cpu_num = processor_count();
2922 size_t cpu_map_size = NCPUS / BitsPerCLong;
2923 size_t cpu_map_valid_size =
2924 MIN2((cpu_num + BitsPerCLong - 1) / BitsPerCLong, cpu_map_size);
2925
2926 cpu_to_node()->clear();
2927 cpu_to_node()->at_grow(cpu_num - 1);
2928
2929 size_t node_num = get_existing_num_nodes();
2930
2931 int distance = 0;
2932 int closest_distance = INT_MAX;
2933 int closest_node = 0;
2934 unsigned long *cpu_map = NEW_C_HEAP_ARRAY(unsigned long, cpu_map_size, mtInternal);
2935 for (size_t i = 0; i < node_num; i++) {
2936 // Check if node is configured (not a memory-less node). If it is not, find
2937 // the closest configured node.
2938 if (!isnode_in_configured_nodes(nindex_to_node()->at(i))) {
2939 closest_distance = INT_MAX;
2940 // Check distance from all remaining nodes in the system. Ignore distance
2941 // from itself and from another non-configured node.
2942 for (size_t m = 0; m < node_num; m++) {
2943 if (m != i && isnode_in_configured_nodes(nindex_to_node()->at(m))) {
2944 distance = numa_distance(nindex_to_node()->at(i), nindex_to_node()->at(m));
2945 // If a closest node is found, update. There is always at least one
2946 // configured node in the system so there is always at least one node
2947 // close.
2948 if (distance != 0 && distance < closest_distance) {
2949 closest_distance = distance;
2950 closest_node = nindex_to_node()->at(m);
2951 }
2952 }
2953 }
2954 } else {
2955 // Current node is already a configured node.
2956 closest_node = nindex_to_node()->at(i);
2957 }
2958
2959 // Get cpus from the original node and map them to the closest node. If node
2960 // is a configured node (not a memory-less node), then original node and
2961 // closest node are the same.
2962 if (numa_node_to_cpus(nindex_to_node()->at(i), cpu_map, cpu_map_size * sizeof(unsigned long)) != -1) {
2963 for (size_t j = 0; j < cpu_map_valid_size; j++) {
2964 if (cpu_map[j] != 0) {
2965 for (size_t k = 0; k < BitsPerCLong; k++) {
2966 if (cpu_map[j] & (1UL << k)) {
2967 cpu_to_node()->at_put(j * BitsPerCLong + k, closest_node);
2968 }
2969 }
2970 }
2971 }
2972 }
2973 }
2974 FREE_C_HEAP_ARRAY(unsigned long, cpu_map, mtInternal);
2975 }
2976
2977 int os::Linux::get_node_by_cpu(int cpu_id) {
2978 if (cpu_to_node() != NULL && cpu_id >= 0 && cpu_id < cpu_to_node()->length()) {
2979 return cpu_to_node()->at(cpu_id);
2980 }
2981 return -1;
2982 }
2983
2984 GrowableArray<int>* os::Linux::_cpu_to_node;
2985 GrowableArray<int>* os::Linux::_nindex_to_node;
2986 os::Linux::sched_getcpu_func_t os::Linux::_sched_getcpu;
2987 os::Linux::numa_node_to_cpus_func_t os::Linux::_numa_node_to_cpus;
2988 os::Linux::numa_max_node_func_t os::Linux::_numa_max_node;
2989 os::Linux::numa_num_configured_nodes_func_t os::Linux::_numa_num_configured_nodes;
2990 os::Linux::numa_available_func_t os::Linux::_numa_available;
2991 os::Linux::numa_tonode_memory_func_t os::Linux::_numa_tonode_memory;
2992 os::Linux::numa_interleave_memory_func_t os::Linux::_numa_interleave_memory;
2993 os::Linux::numa_interleave_memory_v2_func_t os::Linux::_numa_interleave_memory_v2;
2994 os::Linux::numa_set_bind_policy_func_t os::Linux::_numa_set_bind_policy;
2995 os::Linux::numa_bitmask_isbitset_func_t os::Linux::_numa_bitmask_isbitset;
2996 os::Linux::numa_distance_func_t os::Linux::_numa_distance;
2997 unsigned long* os::Linux::_numa_all_nodes;
2998 struct bitmask* os::Linux::_numa_all_nodes_ptr;
2999 struct bitmask* os::Linux::_numa_nodes_ptr;
3000
3001 bool os::pd_uncommit_memory(char* addr, size_t size) {
3002 uintptr_t res = (uintptr_t) ::mmap(addr, size, PROT_NONE,
3003 MAP_PRIVATE|MAP_FIXED|MAP_NORESERVE|MAP_ANONYMOUS, -1, 0);
3004 return res != (uintptr_t) MAP_FAILED;
3005 }
3006
3007 static
3008 address get_stack_commited_bottom(address bottom, size_t size) {
3009 address nbot = bottom;
3010 address ntop = bottom + size;
3011
3012 size_t page_sz = os::vm_page_size();
3013 unsigned pages = size / page_sz;
3014
3015 unsigned char vec[1];
3016 unsigned imin = 1, imax = pages + 1, imid;
3017 int mincore_return_value = 0;
3018
3019 assert(imin <= imax, "Unexpected page size");
3020
3021 while (imin < imax) {
3022 imid = (imax + imin) / 2;
3023 nbot = ntop - (imid * page_sz);
3024
3025 // Use a trick with mincore to check whether the page is mapped or not.
3026 // mincore sets vec to 1 if page resides in memory and to 0 if page
3027 // is swapped output but if page we are asking for is unmapped
3028 // it returns -1,ENOMEM
3029 mincore_return_value = mincore(nbot, page_sz, vec);
3030
3031 if (mincore_return_value == -1) {
3032 // Page is not mapped go up
3033 // to find first mapped page
3034 if (errno != EAGAIN) {
3035 assert(errno == ENOMEM, "Unexpected mincore errno");
3036 imax = imid;
3037 }
3038 } else {
3039 // Page is mapped go down
3040 // to find first not mapped page
3041 imin = imid + 1;
3042 }
3043 }
3044
3045 nbot = nbot + page_sz;
3046
3047 // Adjust stack bottom one page up if last checked page is not mapped
3048 if (mincore_return_value == -1) {
3049 nbot = nbot + page_sz;
3050 }
3051
3052 return nbot;
3053 }
3054
3055
3056 // Linux uses a growable mapping for the stack, and if the mapping for
3057 // the stack guard pages is not removed when we detach a thread the
3058 // stack cannot grow beyond the pages where the stack guard was
3059 // mapped. If at some point later in the process the stack expands to
3060 // that point, the Linux kernel cannot expand the stack any further
3061 // because the guard pages are in the way, and a segfault occurs.
3062 //
3063 // However, it's essential not to split the stack region by unmapping
3064 // a region (leaving a hole) that's already part of the stack mapping,
3065 // so if the stack mapping has already grown beyond the guard pages at
3066 // the time we create them, we have to truncate the stack mapping.
3067 // So, we need to know the extent of the stack mapping when
3068 // create_stack_guard_pages() is called.
3069
3070 // We only need this for stacks that are growable: at the time of
3071 // writing thread stacks don't use growable mappings (i.e. those
3072 // creeated with MAP_GROWSDOWN), and aren't marked "[stack]", so this
3073 // only applies to the main thread.
3074
3075 // If the (growable) stack mapping already extends beyond the point
3076 // where we're going to put our guard pages, truncate the mapping at
3077 // that point by munmap()ping it. This ensures that when we later
3078 // munmap() the guard pages we don't leave a hole in the stack
3079 // mapping. This only affects the main/primordial thread
3080
3081 bool os::pd_create_stack_guard_pages(char* addr, size_t size) {
3082
3083 if (os::is_primordial_thread()) {
3084 // As we manually grow stack up to bottom inside create_attached_thread(),
3085 // it's likely that os::Linux::initial_thread_stack_bottom is mapped and
3086 // we don't need to do anything special.
3087 // Check it first, before calling heavy function.
3088 uintptr_t stack_extent = (uintptr_t) os::Linux::initial_thread_stack_bottom();
3089 unsigned char vec[1];
3090
3091 if (mincore((address)stack_extent, os::vm_page_size(), vec) == -1) {
3092 // Fallback to slow path on all errors, including EAGAIN
3093 stack_extent = (uintptr_t) get_stack_commited_bottom(
3094 os::Linux::initial_thread_stack_bottom(),
3095 (size_t)addr - stack_extent);
3096 }
3097
3098 if (stack_extent < (uintptr_t)addr) {
3099 ::munmap((void*)stack_extent, (uintptr_t)(addr - stack_extent));
3100 }
3101 }
3102
3103 return os::commit_memory(addr, size, !ExecMem);
3104 }
3105
3106 // If this is a growable mapping, remove the guard pages entirely by
3107 // munmap()ping them. If not, just call uncommit_memory(). This only
3108 // affects the main/primordial thread, but guard against future OS changes.
3109 // It's safe to always unmap guard pages for primordial thread because we
3110 // always place it right after end of the mapped region.
3111
3112 bool os::remove_stack_guard_pages(char* addr, size_t size) {
3113 uintptr_t stack_extent, stack_base;
3114
3115 if (os::is_primordial_thread()) {
3116 return ::munmap(addr, size) == 0;
3117 }
3118
3119 return os::uncommit_memory(addr, size);
3120 }
3121
3122 static address _highest_vm_reserved_address = NULL;
3123
3124 // If 'fixed' is true, anon_mmap() will attempt to reserve anonymous memory
3125 // at 'requested_addr'. If there are existing memory mappings at the same
3126 // location, however, they will be overwritten. If 'fixed' is false,
3127 // 'requested_addr' is only treated as a hint, the return value may or
3128 // may not start from the requested address. Unlike Linux mmap(), this
3129 // function returns NULL to indicate failure.
3130 static char* anon_mmap(char* requested_addr, size_t bytes, bool fixed) {
3131 char * addr;
3132 int flags;
3133
3134 flags = MAP_PRIVATE | MAP_NORESERVE | MAP_ANONYMOUS;
3135 if (fixed) {
3136 assert((uintptr_t)requested_addr % os::Linux::page_size() == 0, "unaligned address");
3137 flags |= MAP_FIXED;
3138 }
3139
3140 // Map reserved/uncommitted pages PROT_NONE so we fail early if we
3141 // touch an uncommitted page. Otherwise, the read/write might
3142 // succeed if we have enough swap space to back the physical page.
3143 addr = (char*)::mmap(requested_addr, bytes, PROT_NONE,
3144 flags, -1, 0);
3145
3146 if (addr != MAP_FAILED) {
3147 // anon_mmap() should only get called during VM initialization,
3148 // don't need lock (actually we can skip locking even it can be called
3149 // from multiple threads, because _highest_vm_reserved_address is just a
3150 // hint about the upper limit of non-stack memory regions.)
3151 if ((address)addr + bytes > _highest_vm_reserved_address) {
3152 _highest_vm_reserved_address = (address)addr + bytes;
3153 }
3154 }
3155
3156 return addr == MAP_FAILED ? NULL : addr;
3157 }
3158
3159 // Allocate (using mmap, NO_RESERVE, with small pages) at either a given request address
3160 // (req_addr != NULL) or with a given alignment.
3161 // - bytes shall be a multiple of alignment.
3162 // - req_addr can be NULL. If not NULL, it must be a multiple of alignment.
3163 // - alignment sets the alignment at which memory shall be allocated.
3164 // It must be a multiple of allocation granularity.
3165 // Returns address of memory or NULL. If req_addr was not NULL, will only return
3166 // req_addr or NULL.
3167 static char* anon_mmap_aligned(size_t bytes, size_t alignment, char* req_addr) {
3168
3169 size_t extra_size = bytes;
3170 if (req_addr == NULL && alignment > 0) {
3171 extra_size += alignment;
3172 }
3173
3174 char* start = (char*) ::mmap(req_addr, extra_size, PROT_NONE,
3175 MAP_PRIVATE|MAP_ANONYMOUS|MAP_NORESERVE,
3176 -1, 0);
3177 if (start == MAP_FAILED) {
3178 start = NULL;
3179 } else {
3180 if (req_addr != NULL) {
3181 if (start != req_addr) {
3182 ::munmap(start, extra_size);
3183 start = NULL;
3184 }
3185 } else {
3186 char* const start_aligned = (char*) align_ptr_up(start, alignment);
3187 char* const end_aligned = start_aligned + bytes;
3188 char* const end = start + extra_size;
3189 if (start_aligned > start) {
3190 ::munmap(start, start_aligned - start);
3191 }
3192 if (end_aligned < end) {
3193 ::munmap(end_aligned, end - end_aligned);
3194 }
3195 start = start_aligned;
3196 }
3197 }
3198 return start;
3199 }
3200
3201 // Don't update _highest_vm_reserved_address, because there might be memory
3202 // regions above addr + size. If so, releasing a memory region only creates
3203 // a hole in the address space, it doesn't help prevent heap-stack collision.
3204 //
3205 static int anon_munmap(char * addr, size_t size) {
3206 return ::munmap(addr, size) == 0;
3207 }
3208
3209 char* os::pd_reserve_memory(size_t bytes, char* requested_addr,
3210 size_t alignment_hint) {
3211 return anon_mmap(requested_addr, bytes, (requested_addr != NULL));
3212 }
3213
3214 bool os::pd_release_memory(char* addr, size_t size) {
3215 return anon_munmap(addr, size);
3216 }
3217
3218 static address highest_vm_reserved_address() {
3219 return _highest_vm_reserved_address;
3220 }
3221
3222 static bool linux_mprotect(char* addr, size_t size, int prot) {
3223 // Linux wants the mprotect address argument to be page aligned.
3224 char* bottom = (char*)align_size_down((intptr_t)addr, os::Linux::page_size());
3225
3226 // According to SUSv3, mprotect() should only be used with mappings
3227 // established by mmap(), and mmap() always maps whole pages. Unaligned
3228 // 'addr' likely indicates problem in the VM (e.g. trying to change
3229 // protection of malloc'ed or statically allocated memory). Check the
3230 // caller if you hit this assert.
3231 assert(addr == bottom, "sanity check");
3232
3233 size = align_size_up(pointer_delta(addr, bottom, 1) + size, os::Linux::page_size());
3234 return ::mprotect(bottom, size, prot) == 0;
3235 }
3236
3237 // Set protections specified
3238 bool os::protect_memory(char* addr, size_t bytes, ProtType prot,
3239 bool is_committed) {
3240 unsigned int p = 0;
3241 switch (prot) {
3242 case MEM_PROT_NONE: p = PROT_NONE; break;
3243 case MEM_PROT_READ: p = PROT_READ; break;
3244 case MEM_PROT_RW: p = PROT_READ|PROT_WRITE; break;
3245 case MEM_PROT_RWX: p = PROT_READ|PROT_WRITE|PROT_EXEC; break;
3246 default:
3247 ShouldNotReachHere();
3248 }
3249 // is_committed is unused.
3250 return linux_mprotect(addr, bytes, p);
3251 }
3252
3253 bool os::guard_memory(char* addr, size_t size) {
3254 return linux_mprotect(addr, size, PROT_NONE);
3255 }
3256
3257 bool os::unguard_memory(char* addr, size_t size) {
3258 return linux_mprotect(addr, size, PROT_READ|PROT_WRITE);
3259 }
3260
3261 bool os::Linux::transparent_huge_pages_sanity_check(bool warn, size_t page_size) {
3262 bool result = false;
3263 void *p = mmap(NULL, page_size * 2, PROT_READ|PROT_WRITE,
3264 MAP_ANONYMOUS|MAP_PRIVATE,
3265 -1, 0);
3266 if (p != MAP_FAILED) {
3267 void *aligned_p = align_ptr_up(p, page_size);
3268
3269 result = madvise(aligned_p, page_size, MADV_HUGEPAGE) == 0;
3270
3271 munmap(p, page_size * 2);
3272 }
3273
3274 if (warn && !result) {
3275 warning("TransparentHugePages is not supported by the operating system.");
3276 }
3277
3278 return result;
3279 }
3280
3281 bool os::Linux::hugetlbfs_sanity_check(bool warn, size_t page_size) {
3282 bool result = false;
3283 void *p = mmap(NULL, page_size, PROT_READ|PROT_WRITE,
3284 MAP_ANONYMOUS|MAP_PRIVATE|MAP_HUGETLB,
3285 -1, 0);
3286
3287 if (p != MAP_FAILED) {
3288 // We don't know if this really is a huge page or not.
3289 FILE *fp = fopen("/proc/self/maps", "r");
3290 if (fp) {
3291 while (!feof(fp)) {
3292 char chars[257];
3293 long x = 0;
3294 if (fgets(chars, sizeof(chars), fp)) {
3295 if (sscanf(chars, "%lx-%*x", &x) == 1
3296 && x == (long)p) {
3297 if (strstr (chars, "hugepage")) {
3298 result = true;
3299 break;
3300 }
3301 }
3302 }
3303 }
3304 fclose(fp);
3305 }
3306 munmap(p, page_size);
3307 }
3308
3309 if (warn && !result) {
3310 warning("HugeTLBFS is not supported by the operating system.");
3311 }
3312
3313 return result;
3314 }
3315
3316 /*
3317 * Set the coredump_filter bits to include largepages in core dump (bit 6)
3318 *
3319 * From the coredump_filter documentation:
3320 *
3321 * - (bit 0) anonymous private memory
3322 * - (bit 1) anonymous shared memory
3323 * - (bit 2) file-backed private memory
3324 * - (bit 3) file-backed shared memory
3325 * - (bit 4) ELF header pages in file-backed private memory areas (it is
3326 * effective only if the bit 2 is cleared)
3327 * - (bit 5) hugetlb private memory
3328 * - (bit 6) hugetlb shared memory
3329 */
3330 static void set_coredump_filter(void) {
3331 FILE *f;
3332 long cdm;
3333
3334 if ((f = fopen("/proc/self/coredump_filter", "r+")) == NULL) {
3335 return;
3336 }
3337
3338 if (fscanf(f, "%lx", &cdm) != 1) {
3339 fclose(f);
3340 return;
3341 }
3342
3343 rewind(f);
3344
3345 if ((cdm & LARGEPAGES_BIT) == 0) {
3346 cdm |= LARGEPAGES_BIT;
3347 fprintf(f, "%#lx", cdm);
3348 }
3349
3350 fclose(f);
3351 }
3352
3353 // Large page support
3354
3355 static size_t _large_page_size = 0;
3356
3357 size_t os::Linux::find_large_page_size() {
3358 size_t large_page_size = 0;
3359
3360 // large_page_size on Linux is used to round up heap size. x86 uses either
3361 // 2M or 4M page, depending on whether PAE (Physical Address Extensions)
3362 // mode is enabled. AMD64/EM64T uses 2M page in 64bit mode. IA64 can use
3363 // page as large as 256M.
3364 //
3365 // Here we try to figure out page size by parsing /proc/meminfo and looking
3366 // for a line with the following format:
3367 // Hugepagesize: 2048 kB
3368 //
3369 // If we can't determine the value (e.g. /proc is not mounted, or the text
3370 // format has been changed), we'll use the largest page size supported by
3371 // the processor.
3372
3373 #ifndef ZERO
3374 large_page_size = IA32_ONLY(4 * M) AMD64_ONLY(2 * M) IA64_ONLY(256 * M) SPARC_ONLY(4 * M)
3375 ARM_ONLY(2 * M) PPC_ONLY(4 * M);
3376 #endif // ZERO
3377
3378 FILE *fp = fopen("/proc/meminfo", "r");
3379 if (fp) {
3380 while (!feof(fp)) {
3381 int x = 0;
3382 char buf[16];
3383 if (fscanf(fp, "Hugepagesize: %d", &x) == 1) {
3384 if (x && fgets(buf, sizeof(buf), fp) && strcmp(buf, " kB\n") == 0) {
3385 large_page_size = x * K;
3386 break;
3387 }
3388 } else {
3389 // skip to next line
3390 for (;;) {
3391 int ch = fgetc(fp);
3392 if (ch == EOF || ch == (int)'\n') break;
3393 }
3394 }
3395 }
3396 fclose(fp);
3397 }
3398
3399 if (!FLAG_IS_DEFAULT(LargePageSizeInBytes) && LargePageSizeInBytes != large_page_size) {
3400 warning("Setting LargePageSizeInBytes has no effect on this OS. Large page size is "
3401 SIZE_FORMAT "%s.", byte_size_in_proper_unit(large_page_size),
3402 proper_unit_for_byte_size(large_page_size));
3403 }
3404
3405 return large_page_size;
3406 }
3407
3408 size_t os::Linux::setup_large_page_size() {
3409 _large_page_size = Linux::find_large_page_size();
3410 const size_t default_page_size = (size_t)Linux::page_size();
3411 if (_large_page_size > default_page_size) {
3412 _page_sizes[0] = _large_page_size;
3413 _page_sizes[1] = default_page_size;
3414 _page_sizes[2] = 0;
3415 }
3416
3417 return _large_page_size;
3418 }
3419
3420 bool os::Linux::setup_large_page_type(size_t page_size) {
3421 if (FLAG_IS_DEFAULT(UseHugeTLBFS) &&
3422 FLAG_IS_DEFAULT(UseSHM) &&
3423 FLAG_IS_DEFAULT(UseTransparentHugePages)) {
3424
3425 // The type of large pages has not been specified by the user.
3426
3427 // Try UseHugeTLBFS and then UseSHM.
3428 UseHugeTLBFS = UseSHM = true;
3429
3430 // Don't try UseTransparentHugePages since there are known
3431 // performance issues with it turned on. This might change in the future.
3432 UseTransparentHugePages = false;
3433 }
3434
3435 if (UseTransparentHugePages) {
3436 bool warn_on_failure = !FLAG_IS_DEFAULT(UseTransparentHugePages);
3437 if (transparent_huge_pages_sanity_check(warn_on_failure, page_size)) {
3438 UseHugeTLBFS = false;
3439 UseSHM = false;
3440 return true;
3441 }
3442 UseTransparentHugePages = false;
3443 }
3444
3445 if (UseHugeTLBFS) {
3446 bool warn_on_failure = !FLAG_IS_DEFAULT(UseHugeTLBFS);
3447 if (hugetlbfs_sanity_check(warn_on_failure, page_size)) {
3448 UseSHM = false;
3449 return true;
3450 }
3451 UseHugeTLBFS = false;
3452 }
3453
3454 return UseSHM;
3455 }
3456
3457 void os::large_page_init() {
3458 if (!UseLargePages &&
3459 !UseTransparentHugePages &&
3460 !UseHugeTLBFS &&
3461 !UseSHM) {
3462 // Not using large pages.
3463 return;
3464 }
3465
3466 if (!FLAG_IS_DEFAULT(UseLargePages) && !UseLargePages) {
3467 // The user explicitly turned off large pages.
3468 // Ignore the rest of the large pages flags.
3469 UseTransparentHugePages = false;
3470 UseHugeTLBFS = false;
3471 UseSHM = false;
3472 return;
3473 }
3474
3475 size_t large_page_size = Linux::setup_large_page_size();
3476 UseLargePages = Linux::setup_large_page_type(large_page_size);
3477
3478 set_coredump_filter();
3479 }
3480
3481 #ifndef SHM_HUGETLB
3482 #define SHM_HUGETLB 04000
3483 #endif
3484
3485 #define shm_warning_format(format, ...) \
3486 do { \
3487 if (UseLargePages && \
3488 (!FLAG_IS_DEFAULT(UseLargePages) || \
3489 !FLAG_IS_DEFAULT(UseSHM) || \
3490 !FLAG_IS_DEFAULT(LargePageSizeInBytes))) { \
3491 warning(format, __VA_ARGS__); \
3492 } \
3493 } while (0)
3494
3495 #define shm_warning(str) shm_warning_format("%s", str)
3496
3497 #define shm_warning_with_errno(str) \
3498 do { \
3499 int err = errno; \
3500 shm_warning_format(str " (error = %d)", err); \
3501 } while (0)
3502
3503 static char* shmat_with_alignment(int shmid, size_t bytes, size_t alignment) {
3504 assert(is_size_aligned(bytes, alignment), "Must be divisible by the alignment");
3505
3506 if (!is_size_aligned(alignment, SHMLBA)) {
3507 assert(false, "Code below assumes that alignment is at least SHMLBA aligned");
3508 return NULL;
3509 }
3510
3511 // To ensure that we get 'alignment' aligned memory from shmat,
3512 // we pre-reserve aligned virtual memory and then attach to that.
3513
3514 char* pre_reserved_addr = anon_mmap_aligned(bytes, alignment, NULL);
3515 if (pre_reserved_addr == NULL) {
3516 // Couldn't pre-reserve aligned memory.
3517 shm_warning("Failed to pre-reserve aligned memory for shmat.");
3518 return NULL;
3519 }
3520
3521 // SHM_REMAP is needed to allow shmat to map over an existing mapping.
3522 char* addr = (char*)shmat(shmid, pre_reserved_addr, SHM_REMAP);
3523
3524 if ((intptr_t)addr == -1) {
3525 int err = errno;
3526 shm_warning_with_errno("Failed to attach shared memory.");
3527
3528 assert(err != EACCES, "Unexpected error");
3529 assert(err != EIDRM, "Unexpected error");
3530 assert(err != EINVAL, "Unexpected error");
3531
3532 // Since we don't know if the kernel unmapped the pre-reserved memory area
3533 // we can't unmap it, since that would potentially unmap memory that was
3534 // mapped from other threads.
3535 return NULL;
3536 }
3537
3538 return addr;
3539 }
3540
3541 static char* shmat_at_address(int shmid, char* req_addr) {
3542 if (!is_ptr_aligned(req_addr, SHMLBA)) {
3543 assert(false, "Requested address needs to be SHMLBA aligned");
3544 return NULL;
3545 }
3546
3547 char* addr = (char*)shmat(shmid, req_addr, 0);
3548
3549 if ((intptr_t)addr == -1) {
3550 shm_warning_with_errno("Failed to attach shared memory.");
3551 return NULL;
3552 }
3553
3554 return addr;
3555 }
3556
3557 static char* shmat_large_pages(int shmid, size_t bytes, size_t alignment, char* req_addr) {
3558 // If a req_addr has been provided, we assume that the caller has already aligned the address.
3559 if (req_addr != NULL) {
3560 assert(is_ptr_aligned(req_addr, os::large_page_size()), "Must be divisible by the large page size");
3561 assert(is_ptr_aligned(req_addr, alignment), "Must be divisible by given alignment");
3562 return shmat_at_address(shmid, req_addr);
3563 }
3564
3565 // Since shmid has been setup with SHM_HUGETLB, shmat will automatically
3566 // return large page size aligned memory addresses when req_addr == NULL.
3567 // However, if the alignment is larger than the large page size, we have
3568 // to manually ensure that the memory returned is 'alignment' aligned.
3569 if (alignment > os::large_page_size()) {
3570 assert(is_size_aligned(alignment, os::large_page_size()), "Must be divisible by the large page size");
3571 return shmat_with_alignment(shmid, bytes, alignment);
3572 } else {
3573 return shmat_at_address(shmid, NULL);
3574 }
3575 }
3576
3577 char* os::Linux::reserve_memory_special_shm(size_t bytes, size_t alignment, char* req_addr, bool exec) {
3578 // "exec" is passed in but not used. Creating the shared image for
3579 // the code cache doesn't have an SHM_X executable permission to check.
3580 assert(UseLargePages && UseSHM, "only for SHM large pages");
3581 assert(is_ptr_aligned(req_addr, os::large_page_size()), "Unaligned address");
3582 assert(is_ptr_aligned(req_addr, alignment), "Unaligned address");
3583
3584 if (!is_size_aligned(bytes, os::large_page_size())) {
3585 return NULL; // Fallback to small pages.
3586 }
3587
3588 // Create a large shared memory region to attach to based on size.
3589 // Currently, size is the total size of the heap.
3590 int shmid = shmget(IPC_PRIVATE, bytes, SHM_HUGETLB|IPC_CREAT|SHM_R|SHM_W);
3591 if (shmid == -1) {
3592 // Possible reasons for shmget failure:
3593 // 1. shmmax is too small for Java heap.
3594 // > check shmmax value: cat /proc/sys/kernel/shmmax
3595 // > increase shmmax value: echo "0xffffffff" > /proc/sys/kernel/shmmax
3596 // 2. not enough large page memory.
3597 // > check available large pages: cat /proc/meminfo
3598 // > increase amount of large pages:
3599 // echo new_value > /proc/sys/vm/nr_hugepages
3600 // Note 1: different Linux may use different name for this property,
3601 // e.g. on Redhat AS-3 it is "hugetlb_pool".
3602 // Note 2: it's possible there's enough physical memory available but
3603 // they are so fragmented after a long run that they can't
3604 // coalesce into large pages. Try to reserve large pages when
3605 // the system is still "fresh".
3606 shm_warning_with_errno("Failed to reserve shared memory.");
3607 return NULL;
3608 }
3609
3610 // Attach to the region.
3611 char* addr = shmat_large_pages(shmid, bytes, alignment, req_addr);
3612
3613 // Remove shmid. If shmat() is successful, the actual shared memory segment
3614 // will be deleted when it's detached by shmdt() or when the process
3615 // terminates. If shmat() is not successful this will remove the shared
3616 // segment immediately.
3617 shmctl(shmid, IPC_RMID, NULL);
3618
3619 return addr;
3620 }
3621
3622 static void warn_on_large_pages_failure(char* req_addr, size_t bytes, int error) {
3623 assert(error == ENOMEM, "Only expect to fail if no memory is available");
3624
3625 bool warn_on_failure = UseLargePages &&
3626 (!FLAG_IS_DEFAULT(UseLargePages) ||
3627 !FLAG_IS_DEFAULT(UseHugeTLBFS) ||
3628 !FLAG_IS_DEFAULT(LargePageSizeInBytes));
3629
3630 if (warn_on_failure) {
3631 char msg[128];
3632 jio_snprintf(msg, sizeof(msg), "Failed to reserve large pages memory req_addr: "
3633 PTR_FORMAT " bytes: " SIZE_FORMAT " (errno = %d).", req_addr, bytes, error);
3634 warning("%s", msg);
3635 }
3636 }
3637
3638 char* os::Linux::reserve_memory_special_huge_tlbfs_only(size_t bytes, char* req_addr, bool exec) {
3639 assert(UseLargePages && UseHugeTLBFS, "only for Huge TLBFS large pages");
3640 assert(is_size_aligned(bytes, os::large_page_size()), "Unaligned size");
3641 assert(is_ptr_aligned(req_addr, os::large_page_size()), "Unaligned address");
3642
3643 int prot = exec ? PROT_READ|PROT_WRITE|PROT_EXEC : PROT_READ|PROT_WRITE;
3644 char* addr = (char*)::mmap(req_addr, bytes, prot,
3645 MAP_PRIVATE|MAP_ANONYMOUS|MAP_HUGETLB,
3646 -1, 0);
3647
3648 if (addr == MAP_FAILED) {
3649 warn_on_large_pages_failure(req_addr, bytes, errno);
3650 return NULL;
3651 }
3652
3653 assert(is_ptr_aligned(addr, os::large_page_size()), "Must be");
3654
3655 return addr;
3656 }
3657
3658 // Reserve memory using mmap(MAP_HUGETLB).
3659 // - bytes shall be a multiple of alignment.
3660 // - req_addr can be NULL. If not NULL, it must be a multiple of alignment.
3661 // - alignment sets the alignment at which memory shall be allocated.
3662 // It must be a multiple of allocation granularity.
3663 // Returns address of memory or NULL. If req_addr was not NULL, will only return
3664 // req_addr or NULL.
3665 char* os::Linux::reserve_memory_special_huge_tlbfs_mixed(size_t bytes, size_t alignment, char* req_addr, bool exec) {
3666 size_t large_page_size = os::large_page_size();
3667 assert(bytes >= large_page_size, "Shouldn't allocate large pages for small sizes");
3668
3669 assert(is_ptr_aligned(req_addr, alignment), "Must be");
3670 assert(is_size_aligned(bytes, alignment), "Must be");
3671
3672 // First reserve - but not commit - the address range in small pages.
3673 char* const start = anon_mmap_aligned(bytes, alignment, req_addr);
3674
3675 if (start == NULL) {
3676 return NULL;
3677 }
3678
3679 assert(is_ptr_aligned(start, alignment), "Must be");
3680
3681 char* end = start + bytes;
3682
3683 // Find the regions of the allocated chunk that can be promoted to large pages.
3684 char* lp_start = (char*)align_ptr_up(start, large_page_size);
3685 char* lp_end = (char*)align_ptr_down(end, large_page_size);
3686
3687 size_t lp_bytes = lp_end - lp_start;
3688
3689 assert(is_size_aligned(lp_bytes, large_page_size), "Must be");
3690
3691 if (lp_bytes == 0) {
3692 // The mapped region doesn't even span the start and the end of a large page.
3693 // Fall back to allocate a non-special area.
3694 ::munmap(start, end - start);
3695 return NULL;
3696 }
3697
3698 int prot = exec ? PROT_READ|PROT_WRITE|PROT_EXEC : PROT_READ|PROT_WRITE;
3699
3700 void* result;
3701
3702 // Commit small-paged leading area.
3703 if (start != lp_start) {
3704 result = ::mmap(start, lp_start - start, prot,
3705 MAP_PRIVATE|MAP_ANONYMOUS|MAP_FIXED,
3706 -1, 0);
3707 if (result == MAP_FAILED) {
3708 ::munmap(lp_start, end - lp_start);
3709 return NULL;
3710 }
3711 }
3712
3713 // Commit large-paged area.
3714 result = ::mmap(lp_start, lp_bytes, prot,
3715 MAP_PRIVATE|MAP_ANONYMOUS|MAP_FIXED|MAP_HUGETLB,
3716 -1, 0);
3717 if (result == MAP_FAILED) {
3718 warn_on_large_pages_failure(lp_start, lp_bytes, errno);
3719 // If the mmap above fails, the large pages region will be unmapped and we
3720 // have regions before and after with small pages. Release these regions.
3721 //
3722 // | mapped | unmapped | mapped |
3723 // ^ ^ ^ ^
3724 // start lp_start lp_end end
3725 //
3726 ::munmap(start, lp_start - start);
3727 ::munmap(lp_end, end - lp_end);
3728 return NULL;
3729 }
3730
3731 // Commit small-paged trailing area.
3732 if (lp_end != end) {
3733 result = ::mmap(lp_end, end - lp_end, prot,
3734 MAP_PRIVATE|MAP_ANONYMOUS|MAP_FIXED,
3735 -1, 0);
3736 if (result == MAP_FAILED) {
3737 ::munmap(start, lp_end - start);
3738 return NULL;
3739 }
3740 }
3741
3742 return start;
3743 }
3744
3745 char* os::Linux::reserve_memory_special_huge_tlbfs(size_t bytes, size_t alignment, char* req_addr, bool exec) {
3746 assert(UseLargePages && UseHugeTLBFS, "only for Huge TLBFS large pages");
3747 assert(is_ptr_aligned(req_addr, alignment), "Must be");
3748 assert(is_size_aligned(alignment, os::vm_allocation_granularity()), "Must be");
3749 assert(is_power_of_2(os::large_page_size()), "Must be");
3750 assert(bytes >= os::large_page_size(), "Shouldn't allocate large pages for small sizes");
3751
3752 if (is_size_aligned(bytes, os::large_page_size()) && alignment <= os::large_page_size()) {
3753 return reserve_memory_special_huge_tlbfs_only(bytes, req_addr, exec);
3754 } else {
3755 return reserve_memory_special_huge_tlbfs_mixed(bytes, alignment, req_addr, exec);
3756 }
3757 }
3758
3759 char* os::reserve_memory_special(size_t bytes, size_t alignment, char* req_addr, bool exec) {
3760 assert(UseLargePages, "only for large pages");
3761
3762 char* addr;
3763 if (UseSHM) {
3764 addr = os::Linux::reserve_memory_special_shm(bytes, alignment, req_addr, exec);
3765 } else {
3766 assert(UseHugeTLBFS, "must be");
3767 addr = os::Linux::reserve_memory_special_huge_tlbfs(bytes, alignment, req_addr, exec);
3768 }
3769
3770 if (addr != NULL) {
3771 if (UseNUMAInterleaving) {
3772 numa_make_global(addr, bytes);
3773 }
3774
3775 // The memory is committed
3776 MemTracker::record_virtual_memory_reserve_and_commit((address)addr, bytes, CALLER_PC);
3777 }
3778
3779 return addr;
3780 }
3781
3782 bool os::Linux::release_memory_special_shm(char* base, size_t bytes) {
3783 // detaching the SHM segment will also delete it, see reserve_memory_special_shm()
3784 return shmdt(base) == 0;
3785 }
3786
3787 bool os::Linux::release_memory_special_huge_tlbfs(char* base, size_t bytes) {
3788 return pd_release_memory(base, bytes);
3789 }
3790
3791 bool os::release_memory_special(char* base, size_t bytes) {
3792 bool res;
3793 if (MemTracker::tracking_level() > NMT_minimal) {
3794 Tracker tkr = MemTracker::get_virtual_memory_release_tracker();
3795 res = os::Linux::release_memory_special_impl(base, bytes);
3796 if (res) {
3797 tkr.record((address)base, bytes);
3798 }
3799
3800 } else {
3801 res = os::Linux::release_memory_special_impl(base, bytes);
3802 }
3803 return res;
3804 }
3805
3806 bool os::Linux::release_memory_special_impl(char* base, size_t bytes) {
3807 assert(UseLargePages, "only for large pages");
3808 bool res;
3809
3810 if (UseSHM) {
3811 res = os::Linux::release_memory_special_shm(base, bytes);
3812 } else {
3813 assert(UseHugeTLBFS, "must be");
3814 res = os::Linux::release_memory_special_huge_tlbfs(base, bytes);
3815 }
3816 return res;
3817 }
3818
3819 size_t os::large_page_size() {
3820 return _large_page_size;
3821 }
3822
3823 // With SysV SHM the entire memory region must be allocated as shared
3824 // memory.
3825 // HugeTLBFS allows application to commit large page memory on demand.
3826 // However, when committing memory with HugeTLBFS fails, the region
3827 // that was supposed to be committed will lose the old reservation
3828 // and allow other threads to steal that memory region. Because of this
3829 // behavior we can't commit HugeTLBFS memory.
3830 bool os::can_commit_large_page_memory() {
3831 return UseTransparentHugePages;
3832 }
3833
3834 bool os::can_execute_large_page_memory() {
3835 return UseTransparentHugePages || UseHugeTLBFS;
3836 }
3837
3838 // Reserve memory at an arbitrary address, only if that area is
3839 // available (and not reserved for something else).
3840
3841 char* os::pd_attempt_reserve_memory_at(size_t bytes, char* requested_addr) {
3842 const int max_tries = 10;
3843 char* base[max_tries];
3844 size_t size[max_tries];
3845 const size_t gap = 0x000000;
3846
3847 // Assert only that the size is a multiple of the page size, since
3848 // that's all that mmap requires, and since that's all we really know
3849 // about at this low abstraction level. If we need higher alignment,
3850 // we can either pass an alignment to this method or verify alignment
3851 // in one of the methods further up the call chain. See bug 5044738.
3852 assert(bytes % os::vm_page_size() == 0, "reserving unexpected size block");
3853
3854 // Repeatedly allocate blocks until the block is allocated at the
3855 // right spot. Give up after max_tries. Note that reserve_memory() will
3856 // automatically update _highest_vm_reserved_address if the call is
3857 // successful. The variable tracks the highest memory address every reserved
3858 // by JVM. It is used to detect heap-stack collision if running with
3859 // fixed-stack LinuxThreads. Because here we may attempt to reserve more
3860 // space than needed, it could confuse the collision detecting code. To
3861 // solve the problem, save current _highest_vm_reserved_address and
3862 // calculate the correct value before return.
3863 address old_highest = _highest_vm_reserved_address;
3864
3865 // Linux mmap allows caller to pass an address as hint; give it a try first,
3866 // if kernel honors the hint then we can return immediately.
3867 char * addr = anon_mmap(requested_addr, bytes, false);
3868 if (addr == requested_addr) {
3869 return requested_addr;
3870 }
3871
3872 if (addr != NULL) {
3873 // mmap() is successful but it fails to reserve at the requested address
3874 anon_munmap(addr, bytes);
3875 }
3876
3877 int i;
3878 for (i = 0; i < max_tries; ++i) {
3879 base[i] = reserve_memory(bytes);
3880
3881 if (base[i] != NULL) {
3882 // Is this the block we wanted?
3883 if (base[i] == requested_addr) {
3884 size[i] = bytes;
3885 break;
3886 }
3887
3888 // Does this overlap the block we wanted? Give back the overlapped
3889 // parts and try again.
3890
3891 size_t top_overlap = requested_addr + (bytes + gap) - base[i];
3892 if (top_overlap >= 0 && top_overlap < bytes) {
3893 unmap_memory(base[i], top_overlap);
3894 base[i] += top_overlap;
3895 size[i] = bytes - top_overlap;
3896 } else {
3897 size_t bottom_overlap = base[i] + bytes - requested_addr;
3898 if (bottom_overlap >= 0 && bottom_overlap < bytes) {
3899 unmap_memory(requested_addr, bottom_overlap);
3900 size[i] = bytes - bottom_overlap;
3901 } else {
3902 size[i] = bytes;
3903 }
3904 }
3905 }
3906 }
3907
3908 // Give back the unused reserved pieces.
3909
3910 for (int j = 0; j < i; ++j) {
3911 if (base[j] != NULL) {
3912 unmap_memory(base[j], size[j]);
3913 }
3914 }
3915
3916 if (i < max_tries) {
3917 _highest_vm_reserved_address = MAX2(old_highest, (address)requested_addr + bytes);
3918 return requested_addr;
3919 } else {
3920 _highest_vm_reserved_address = old_highest;
3921 return NULL;
3922 }
3923 }
3924
3925 size_t os::read(int fd, void *buf, unsigned int nBytes) {
3926 return ::read(fd, buf, nBytes);
3927 }
3928
3929 // TODO-FIXME: reconcile Solaris' os::sleep with the linux variation.
3930 // Solaris uses poll(), linux uses park().
3931 // Poll() is likely a better choice, assuming that Thread.interrupt()
3932 // generates a SIGUSRx signal. Note that SIGUSR1 can interfere with
3933 // SIGSEGV, see 4355769.
3934
3935 int os::sleep(Thread* thread, jlong millis, bool interruptible) {
3936 assert(thread == Thread::current(), "thread consistency check");
3937
3938 ParkEvent * const slp = thread->_SleepEvent ;
3939 slp->reset() ;
3940 OrderAccess::fence() ;
3941
3942 if (interruptible) {
3943 jlong prevtime = javaTimeNanos();
3944
3945 for (;;) {
3946 if (os::is_interrupted(thread, true)) {
3947 return OS_INTRPT;
3948 }
3949
3950 jlong newtime = javaTimeNanos();
3951
3952 if (newtime - prevtime < 0) {
3953 // time moving backwards, should only happen if no monotonic clock
3954 // not a guarantee() because JVM should not abort on kernel/glibc bugs
3955 assert(!Linux::supports_monotonic_clock(), "time moving backwards");
3956 } else {
3957 millis -= (newtime - prevtime) / NANOSECS_PER_MILLISEC;
3958 }
3959
3960 if(millis <= 0) {
3961 return OS_OK;
3962 }
3963
3964 prevtime = newtime;
3965
3966 {
3967 assert(thread->is_Java_thread(), "sanity check");
3968 JavaThread *jt = (JavaThread *) thread;
3969 ThreadBlockInVM tbivm(jt);
3970 OSThreadWaitState osts(jt->osthread(), false /* not Object.wait() */);
3971
3972 jt->set_suspend_equivalent();
3973 // cleared by handle_special_suspend_equivalent_condition() or
3974 // java_suspend_self() via check_and_wait_while_suspended()
3975
3976 slp->park(millis);
3977
3978 // were we externally suspended while we were waiting?
3979 jt->check_and_wait_while_suspended();
3980 }
3981 }
3982 } else {
3983 OSThreadWaitState osts(thread->osthread(), false /* not Object.wait() */);
3984 jlong prevtime = javaTimeNanos();
3985
3986 for (;;) {
3987 // It'd be nice to avoid the back-to-back javaTimeNanos() calls on
3988 // the 1st iteration ...
3989 jlong newtime = javaTimeNanos();
3990
3991 if (newtime - prevtime < 0) {
3992 // time moving backwards, should only happen if no monotonic clock
3993 // not a guarantee() because JVM should not abort on kernel/glibc bugs
3994 assert(!Linux::supports_monotonic_clock(), "time moving backwards");
3995 } else {
3996 millis -= (newtime - prevtime) / NANOSECS_PER_MILLISEC;
3997 }
3998
3999 if(millis <= 0) break ;
4000
4001 prevtime = newtime;
4002 slp->park(millis);
4003 }
4004 return OS_OK ;
4005 }
4006 }
4007
4008 //
4009 // Short sleep, direct OS call.
4010 //
4011 // Note: certain versions of Linux CFS scheduler (since 2.6.23) do not guarantee
4012 // sched_yield(2) will actually give up the CPU:
4013 //
4014 // * Alone on this pariticular CPU, keeps running.
4015 // * Before the introduction of "skip_buddy" with "compat_yield" disabled
4016 // (pre 2.6.39).
4017 //
4018 // So calling this with 0 is an alternative.
4019 //
4020 void os::naked_short_sleep(jlong ms) {
4021 struct timespec req;
4022
4023 assert(ms < 1000, "Un-interruptable sleep, short time use only");
4024 req.tv_sec = 0;
4025 if (ms > 0) {
4026 req.tv_nsec = (ms % 1000) * 1000000;
4027 }
4028 else {
4029 req.tv_nsec = 1;
4030 }
4031
4032 nanosleep(&req, NULL);
4033
4034 return;
4035 }
4036
4037 // Sleep forever; naked call to OS-specific sleep; use with CAUTION
4038 void os::infinite_sleep() {
4039 while (true) { // sleep forever ...
4040 ::sleep(100); // ... 100 seconds at a time
4041 }
4042 }
4043
4044 // Used to convert frequent JVM_Yield() to nops
4045 bool os::dont_yield() {
4046 return DontYieldALot;
4047 }
4048
4049 void os::yield() {
4050 sched_yield();
4051 }
4052
4053 os::YieldResult os::NakedYield() { sched_yield(); return os::YIELD_UNKNOWN ;}
4054
4055 void os::yield_all(int attempts) {
4056 // Yields to all threads, including threads with lower priorities
4057 // Threads on Linux are all with same priority. The Solaris style
4058 // os::yield_all() with nanosleep(1ms) is not necessary.
4059 sched_yield();
4060 }
4061
4062 // Called from the tight loops to possibly influence time-sharing heuristics
4063 void os::loop_breaker(int attempts) {
4064 os::yield_all(attempts);
4065 }
4066
4067 ////////////////////////////////////////////////////////////////////////////////
4068 // thread priority support
4069
4070 // Note: Normal Linux applications are run with SCHED_OTHER policy. SCHED_OTHER
4071 // only supports dynamic priority, static priority must be zero. For real-time
4072 // applications, Linux supports SCHED_RR which allows static priority (1-99).
4073 // However, for large multi-threaded applications, SCHED_RR is not only slower
4074 // than SCHED_OTHER, but also very unstable (my volano tests hang hard 4 out
4075 // of 5 runs - Sep 2005).
4076 //
4077 // The following code actually changes the niceness of kernel-thread/LWP. It
4078 // has an assumption that setpriority() only modifies one kernel-thread/LWP,
4079 // not the entire user process, and user level threads are 1:1 mapped to kernel
4080 // threads. It has always been the case, but could change in the future. For
4081 // this reason, the code should not be used as default (ThreadPriorityPolicy=0).
4082 // It is only used when ThreadPriorityPolicy=1 and requires root privilege.
4083
4084 int os::java_to_os_priority[CriticalPriority + 1] = {
4085 19, // 0 Entry should never be used
4086
4087 4, // 1 MinPriority
4088 3, // 2
4089 2, // 3
4090
4091 1, // 4
4092 0, // 5 NormPriority
4093 -1, // 6
4094
4095 -2, // 7
4096 -3, // 8
4097 -4, // 9 NearMaxPriority
4098
4099 -5, // 10 MaxPriority
4100
4101 -5 // 11 CriticalPriority
4102 };
4103
4104 static int prio_init() {
4105 if (ThreadPriorityPolicy == 1) {
4106 // Only root can raise thread priority. Don't allow ThreadPriorityPolicy=1
4107 // if effective uid is not root. Perhaps, a more elegant way of doing
4108 // this is to test CAP_SYS_NICE capability, but that will require libcap.so
4109 if (geteuid() != 0) {
4110 if (!FLAG_IS_DEFAULT(ThreadPriorityPolicy)) {
4111 warning("-XX:ThreadPriorityPolicy requires root privilege on Linux");
4112 }
4113 ThreadPriorityPolicy = 0;
4114 }
4115 }
4116 if (UseCriticalJavaThreadPriority) {
4117 os::java_to_os_priority[MaxPriority] = os::java_to_os_priority[CriticalPriority];
4118 }
4119 return 0;
4120 }
4121
4122 OSReturn os::set_native_priority(Thread* thread, int newpri) {
4123 if ( !UseThreadPriorities || ThreadPriorityPolicy == 0 ) return OS_OK;
4124
4125 int ret = setpriority(PRIO_PROCESS, thread->osthread()->thread_id(), newpri);
4126 return (ret == 0) ? OS_OK : OS_ERR;
4127 }
4128
4129 OSReturn os::get_native_priority(const Thread* const thread, int *priority_ptr) {
4130 if ( !UseThreadPriorities || ThreadPriorityPolicy == 0 ) {
4131 *priority_ptr = java_to_os_priority[NormPriority];
4132 return OS_OK;
4133 }
4134
4135 errno = 0;
4136 *priority_ptr = getpriority(PRIO_PROCESS, thread->osthread()->thread_id());
4137 return (*priority_ptr != -1 || errno == 0 ? OS_OK : OS_ERR);
4138 }
4139
4140 // Hint to the underlying OS that a task switch would not be good.
4141 // Void return because it's a hint and can fail.
4142 void os::hint_no_preempt() {}
4143
4144 ////////////////////////////////////////////////////////////////////////////////
4145 // suspend/resume support
4146
4147 // the low-level signal-based suspend/resume support is a remnant from the
4148 // old VM-suspension that used to be for java-suspension, safepoints etc,
4149 // within hotspot. Now there is a single use-case for this:
4150 // - calling get_thread_pc() on the VMThread by the flat-profiler task
4151 // that runs in the watcher thread.
4152 // The remaining code is greatly simplified from the more general suspension
4153 // code that used to be used.
4154 //
4155 // The protocol is quite simple:
4156 // - suspend:
4157 // - sends a signal to the target thread
4158 // - polls the suspend state of the osthread using a yield loop
4159 // - target thread signal handler (SR_handler) sets suspend state
4160 // and blocks in sigsuspend until continued
4161 // - resume:
4162 // - sets target osthread state to continue
4163 // - sends signal to end the sigsuspend loop in the SR_handler
4164 //
4165 // Note that the SR_lock plays no role in this suspend/resume protocol.
4166 //
4167
4168 static void resume_clear_context(OSThread *osthread) {
4169 osthread->set_ucontext(NULL);
4170 osthread->set_siginfo(NULL);
4171 }
4172
4173 static void suspend_save_context(OSThread *osthread, siginfo_t* siginfo, ucontext_t* context) {
4174 osthread->set_ucontext(context);
4175 osthread->set_siginfo(siginfo);
4176 }
4177
4178 //
4179 // Handler function invoked when a thread's execution is suspended or
4180 // resumed. We have to be careful that only async-safe functions are
4181 // called here (Note: most pthread functions are not async safe and
4182 // should be avoided.)
4183 //
4184 // Note: sigwait() is a more natural fit than sigsuspend() from an
4185 // interface point of view, but sigwait() prevents the signal hander
4186 // from being run. libpthread would get very confused by not having
4187 // its signal handlers run and prevents sigwait()'s use with the
4188 // mutex granting granting signal.
4189 //
4190 // Currently only ever called on the VMThread and JavaThreads (PC sampling)
4191 //
4192 static void SR_handler(int sig, siginfo_t* siginfo, ucontext_t* context) {
4193 // Save and restore errno to avoid confusing native code with EINTR
4194 // after sigsuspend.
4195 int old_errno = errno;
4196
4197 Thread* thread = Thread::current();
4198 OSThread* osthread = thread->osthread();
4199 assert(thread->is_VM_thread() || thread->is_Java_thread(), "Must be VMThread or JavaThread");
4200
4201 os::SuspendResume::State current = osthread->sr.state();
4202 if (current == os::SuspendResume::SR_SUSPEND_REQUEST) {
4203 suspend_save_context(osthread, siginfo, context);
4204
4205 // attempt to switch the state, we assume we had a SUSPEND_REQUEST
4206 os::SuspendResume::State state = osthread->sr.suspended();
4207 if (state == os::SuspendResume::SR_SUSPENDED) {
4208 sigset_t suspend_set; // signals for sigsuspend()
4209
4210 // get current set of blocked signals and unblock resume signal
4211 pthread_sigmask(SIG_BLOCK, NULL, &suspend_set);
4212 sigdelset(&suspend_set, SR_signum);
4213
4214 sr_semaphore.signal();
4215 // wait here until we are resumed
4216 while (1) {
4217 sigsuspend(&suspend_set);
4218
4219 os::SuspendResume::State result = osthread->sr.running();
4220 if (result == os::SuspendResume::SR_RUNNING) {
4221 sr_semaphore.signal();
4222 break;
4223 }
4224 }
4225
4226 } else if (state == os::SuspendResume::SR_RUNNING) {
4227 // request was cancelled, continue
4228 } else {
4229 ShouldNotReachHere();
4230 }
4231
4232 resume_clear_context(osthread);
4233 } else if (current == os::SuspendResume::SR_RUNNING) {
4234 // request was cancelled, continue
4235 } else if (current == os::SuspendResume::SR_WAKEUP_REQUEST) {
4236 // ignore
4237 } else {
4238 // ignore
4239 }
4240
4241 errno = old_errno;
4242 }
4243
4244
4245 static int SR_initialize() {
4246 struct sigaction act;
4247 char *s;
4248 /* Get signal number to use for suspend/resume */
4249 if ((s = ::getenv("_JAVA_SR_SIGNUM")) != 0) {
4250 int sig = ::strtol(s, 0, 10);
4251 if (sig > 0 || sig < _NSIG) {
4252 SR_signum = sig;
4253 }
4254 }
4255
4256 assert(SR_signum > SIGSEGV && SR_signum > SIGBUS,
4257 "SR_signum must be greater than max(SIGSEGV, SIGBUS), see 4355769");
4258
4259 sigemptyset(&SR_sigset);
4260 sigaddset(&SR_sigset, SR_signum);
4261
4262 /* Set up signal handler for suspend/resume */
4263 act.sa_flags = SA_RESTART|SA_SIGINFO;
4264 act.sa_handler = (void (*)(int)) SR_handler;
4265
4266 // SR_signum is blocked by default.
4267 // 4528190 - We also need to block pthread restart signal (32 on all
4268 // supported Linux platforms). Note that LinuxThreads need to block
4269 // this signal for all threads to work properly. So we don't have
4270 // to use hard-coded signal number when setting up the mask.
4271 pthread_sigmask(SIG_BLOCK, NULL, &act.sa_mask);
4272
4273 if (sigaction(SR_signum, &act, 0) == -1) {
4274 return -1;
4275 }
4276
4277 // Save signal flag
4278 os::Linux::set_our_sigflags(SR_signum, act.sa_flags);
4279 return 0;
4280 }
4281
4282 static int sr_notify(OSThread* osthread) {
4283 int status = pthread_kill(osthread->pthread_id(), SR_signum);
4284 assert_status(status == 0, status, "pthread_kill");
4285 return status;
4286 }
4287
4288 // "Randomly" selected value for how long we want to spin
4289 // before bailing out on suspending a thread, also how often
4290 // we send a signal to a thread we want to resume
4291 static const int RANDOMLY_LARGE_INTEGER = 1000000;
4292 static const int RANDOMLY_LARGE_INTEGER2 = 100;
4293
4294 // returns true on success and false on error - really an error is fatal
4295 // but this seems the normal response to library errors
4296 static bool do_suspend(OSThread* osthread) {
4297 assert(osthread->sr.is_running(), "thread should be running");
4298 assert(!sr_semaphore.trywait(), "semaphore has invalid state");
4299
4300 // mark as suspended and send signal
4301 if (osthread->sr.request_suspend() != os::SuspendResume::SR_SUSPEND_REQUEST) {
4302 // failed to switch, state wasn't running?
4303 ShouldNotReachHere();
4304 return false;
4305 }
4306
4307 if (sr_notify(osthread) != 0) {
4308 ShouldNotReachHere();
4309 }
4310
4311 // managed to send the signal and switch to SUSPEND_REQUEST, now wait for SUSPENDED
4312 while (true) {
4313 if (sr_semaphore.timedwait(0, 2 * NANOSECS_PER_MILLISEC)) {
4314 break;
4315 } else {
4316 // timeout
4317 os::SuspendResume::State cancelled = osthread->sr.cancel_suspend();
4318 if (cancelled == os::SuspendResume::SR_RUNNING) {
4319 return false;
4320 } else if (cancelled == os::SuspendResume::SR_SUSPENDED) {
4321 // make sure that we consume the signal on the semaphore as well
4322 sr_semaphore.wait();
4323 break;
4324 } else {
4325 ShouldNotReachHere();
4326 return false;
4327 }
4328 }
4329 }
4330
4331 guarantee(osthread->sr.is_suspended(), "Must be suspended");
4332 return true;
4333 }
4334
4335 static void do_resume(OSThread* osthread) {
4336 assert(osthread->sr.is_suspended(), "thread should be suspended");
4337 assert(!sr_semaphore.trywait(), "invalid semaphore state");
4338
4339 if (osthread->sr.request_wakeup() != os::SuspendResume::SR_WAKEUP_REQUEST) {
4340 // failed to switch to WAKEUP_REQUEST
4341 ShouldNotReachHere();
4342 return;
4343 }
4344
4345 while (true) {
4346 if (sr_notify(osthread) == 0) {
4347 if (sr_semaphore.timedwait(0, 2 * NANOSECS_PER_MILLISEC)) {
4348 if (osthread->sr.is_running()) {
4349 return;
4350 }
4351 }
4352 } else {
4353 ShouldNotReachHere();
4354 }
4355 }
4356
4357 guarantee(osthread->sr.is_running(), "Must be running!");
4358 }
4359
4360 ////////////////////////////////////////////////////////////////////////////////
4361 // interrupt support
4362
4363 void os::interrupt(Thread* thread) {
4364 assert(Thread::current() == thread || Threads_lock->owned_by_self(),
4365 "possibility of dangling Thread pointer");
4366
4367 OSThread* osthread = thread->osthread();
4368
4369 if (!osthread->interrupted()) {
4370 osthread->set_interrupted(true);
4371 // More than one thread can get here with the same value of osthread,
4372 // resulting in multiple notifications. We do, however, want the store
4373 // to interrupted() to be visible to other threads before we execute unpark().
4374 OrderAccess::fence();
4375 ParkEvent * const slp = thread->_SleepEvent ;
4376 if (slp != NULL) slp->unpark() ;
4377 }
4378
4379 // For JSR166. Unpark even if interrupt status already was set
4380 if (thread->is_Java_thread())
4381 ((JavaThread*)thread)->parker()->unpark();
4382
4383 ParkEvent * ev = thread->_ParkEvent ;
4384 if (ev != NULL) ev->unpark() ;
4385
4386 }
4387
4388 bool os::is_interrupted(Thread* thread, bool clear_interrupted) {
4389 assert(Thread::current() == thread || Threads_lock->owned_by_self(),
4390 "possibility of dangling Thread pointer");
4391
4392 OSThread* osthread = thread->osthread();
4393
4394 bool interrupted = osthread->interrupted();
4395
4396 if (interrupted && clear_interrupted) {
4397 osthread->set_interrupted(false);
4398 // consider thread->_SleepEvent->reset() ... optional optimization
4399 }
4400
4401 return interrupted;
4402 }
4403
4404 ///////////////////////////////////////////////////////////////////////////////////
4405 // signal handling (except suspend/resume)
4406
4407 // This routine may be used by user applications as a "hook" to catch signals.
4408 // The user-defined signal handler must pass unrecognized signals to this
4409 // routine, and if it returns true (non-zero), then the signal handler must
4410 // return immediately. If the flag "abort_if_unrecognized" is true, then this
4411 // routine will never retun false (zero), but instead will execute a VM panic
4412 // routine kill the process.
4413 //
4414 // If this routine returns false, it is OK to call it again. This allows
4415 // the user-defined signal handler to perform checks either before or after
4416 // the VM performs its own checks. Naturally, the user code would be making
4417 // a serious error if it tried to handle an exception (such as a null check
4418 // or breakpoint) that the VM was generating for its own correct operation.
4419 //
4420 // This routine may recognize any of the following kinds of signals:
4421 // SIGBUS, SIGSEGV, SIGILL, SIGFPE, SIGQUIT, SIGPIPE, SIGXFSZ, SIGUSR1.
4422 // It should be consulted by handlers for any of those signals.
4423 //
4424 // The caller of this routine must pass in the three arguments supplied
4425 // to the function referred to in the "sa_sigaction" (not the "sa_handler")
4426 // field of the structure passed to sigaction(). This routine assumes that
4427 // the sa_flags field passed to sigaction() includes SA_SIGINFO and SA_RESTART.
4428 //
4429 // Note that the VM will print warnings if it detects conflicting signal
4430 // handlers, unless invoked with the option "-XX:+AllowUserSignalHandlers".
4431 //
4432 extern "C" JNIEXPORT int
4433 JVM_handle_linux_signal(int signo, siginfo_t* siginfo,
4434 void* ucontext, int abort_if_unrecognized);
4435
4436 void signalHandler(int sig, siginfo_t* info, void* uc) {
4437 assert(info != NULL && uc != NULL, "it must be old kernel");
4438 int orig_errno = errno; // Preserve errno value over signal handler.
4439 JVM_handle_linux_signal(sig, info, uc, true);
4440 errno = orig_errno;
4441 }
4442
4443
4444 // This boolean allows users to forward their own non-matching signals
4445 // to JVM_handle_linux_signal, harmlessly.
4446 bool os::Linux::signal_handlers_are_installed = false;
4447
4448 // For signal-chaining
4449 struct sigaction os::Linux::sigact[MAXSIGNUM];
4450 unsigned int os::Linux::sigs = 0;
4451 bool os::Linux::libjsig_is_loaded = false;
4452 typedef struct sigaction *(*get_signal_t)(int);
4453 get_signal_t os::Linux::get_signal_action = NULL;
4454
4455 struct sigaction* os::Linux::get_chained_signal_action(int sig) {
4456 struct sigaction *actp = NULL;
4457
4458 if (libjsig_is_loaded) {
4459 // Retrieve the old signal handler from libjsig
4460 actp = (*get_signal_action)(sig);
4461 }
4462 if (actp == NULL) {
4463 // Retrieve the preinstalled signal handler from jvm
4464 actp = get_preinstalled_handler(sig);
4465 }
4466
4467 return actp;
4468 }
4469
4470 static bool call_chained_handler(struct sigaction *actp, int sig,
4471 siginfo_t *siginfo, void *context) {
4472 // Call the old signal handler
4473 if (actp->sa_handler == SIG_DFL) {
4474 // It's more reasonable to let jvm treat it as an unexpected exception
4475 // instead of taking the default action.
4476 return false;
4477 } else if (actp->sa_handler != SIG_IGN) {
4478 if ((actp->sa_flags & SA_NODEFER) == 0) {
4479 // automaticlly block the signal
4480 sigaddset(&(actp->sa_mask), sig);
4481 }
4482
4483 sa_handler_t hand = NULL;
4484 sa_sigaction_t sa = NULL;
4485 bool siginfo_flag_set = (actp->sa_flags & SA_SIGINFO) != 0;
4486 // retrieve the chained handler
4487 if (siginfo_flag_set) {
4488 sa = actp->sa_sigaction;
4489 } else {
4490 hand = actp->sa_handler;
4491 }
4492
4493 if ((actp->sa_flags & SA_RESETHAND) != 0) {
4494 actp->sa_handler = SIG_DFL;
4495 }
4496
4497 // try to honor the signal mask
4498 sigset_t oset;
4499 pthread_sigmask(SIG_SETMASK, &(actp->sa_mask), &oset);
4500
4501 // call into the chained handler
4502 if (siginfo_flag_set) {
4503 (*sa)(sig, siginfo, context);
4504 } else {
4505 (*hand)(sig);
4506 }
4507
4508 // restore the signal mask
4509 pthread_sigmask(SIG_SETMASK, &oset, 0);
4510 }
4511 // Tell jvm's signal handler the signal is taken care of.
4512 return true;
4513 }
4514
4515 bool os::Linux::chained_handler(int sig, siginfo_t* siginfo, void* context) {
4516 bool chained = false;
4517 // signal-chaining
4518 if (UseSignalChaining) {
4519 struct sigaction *actp = get_chained_signal_action(sig);
4520 if (actp != NULL) {
4521 chained = call_chained_handler(actp, sig, siginfo, context);
4522 }
4523 }
4524 return chained;
4525 }
4526
4527 struct sigaction* os::Linux::get_preinstalled_handler(int sig) {
4528 if ((( (unsigned int)1 << sig ) & sigs) != 0) {
4529 return &sigact[sig];
4530 }
4531 return NULL;
4532 }
4533
4534 void os::Linux::save_preinstalled_handler(int sig, struct sigaction& oldAct) {
4535 assert(sig > 0 && sig < MAXSIGNUM, "vm signal out of expected range");
4536 sigact[sig] = oldAct;
4537 sigs |= (unsigned int)1 << sig;
4538 }
4539
4540 // for diagnostic
4541 int os::Linux::sigflags[MAXSIGNUM];
4542
4543 int os::Linux::get_our_sigflags(int sig) {
4544 assert(sig > 0 && sig < MAXSIGNUM, "vm signal out of expected range");
4545 return sigflags[sig];
4546 }
4547
4548 void os::Linux::set_our_sigflags(int sig, int flags) {
4549 assert(sig > 0 && sig < MAXSIGNUM, "vm signal out of expected range");
4550 sigflags[sig] = flags;
4551 }
4552
4553 void os::Linux::set_signal_handler(int sig, bool set_installed) {
4554 // Check for overwrite.
4555 struct sigaction oldAct;
4556 sigaction(sig, (struct sigaction*)NULL, &oldAct);
4557
4558 void* oldhand = oldAct.sa_sigaction
4559 ? CAST_FROM_FN_PTR(void*, oldAct.sa_sigaction)
4560 : CAST_FROM_FN_PTR(void*, oldAct.sa_handler);
4561 if (oldhand != CAST_FROM_FN_PTR(void*, SIG_DFL) &&
4562 oldhand != CAST_FROM_FN_PTR(void*, SIG_IGN) &&
4563 oldhand != CAST_FROM_FN_PTR(void*, (sa_sigaction_t)signalHandler)) {
4564 if (AllowUserSignalHandlers || !set_installed) {
4565 // Do not overwrite; user takes responsibility to forward to us.
4566 return;
4567 } else if (UseSignalChaining) {
4568 // save the old handler in jvm
4569 save_preinstalled_handler(sig, oldAct);
4570 // libjsig also interposes the sigaction() call below and saves the
4571 // old sigaction on it own.
4572 } else {
4573 fatal(err_msg("Encountered unexpected pre-existing sigaction handler "
4574 "%#lx for signal %d.", (long)oldhand, sig));
4575 }
4576 }
4577
4578 struct sigaction sigAct;
4579 sigfillset(&(sigAct.sa_mask));
4580 sigAct.sa_handler = SIG_DFL;
4581 if (!set_installed) {
4582 sigAct.sa_flags = SA_SIGINFO|SA_RESTART;
4583 } else {
4584 sigAct.sa_sigaction = signalHandler;
4585 sigAct.sa_flags = SA_SIGINFO|SA_RESTART;
4586 }
4587 // Save flags, which are set by ours
4588 assert(sig > 0 && sig < MAXSIGNUM, "vm signal out of expected range");
4589 sigflags[sig] = sigAct.sa_flags;
4590
4591 int ret = sigaction(sig, &sigAct, &oldAct);
4592 assert(ret == 0, "check");
4593
4594 void* oldhand2 = oldAct.sa_sigaction
4595 ? CAST_FROM_FN_PTR(void*, oldAct.sa_sigaction)
4596 : CAST_FROM_FN_PTR(void*, oldAct.sa_handler);
4597 assert(oldhand2 == oldhand, "no concurrent signal handler installation");
4598 }
4599
4600 // install signal handlers for signals that HotSpot needs to
4601 // handle in order to support Java-level exception handling.
4602
4603 void os::Linux::install_signal_handlers() {
4604 if (!signal_handlers_are_installed) {
4605 signal_handlers_are_installed = true;
4606
4607 // signal-chaining
4608 typedef void (*signal_setting_t)();
4609 signal_setting_t begin_signal_setting = NULL;
4610 signal_setting_t end_signal_setting = NULL;
4611 begin_signal_setting = CAST_TO_FN_PTR(signal_setting_t,
4612 dlsym(RTLD_DEFAULT, "JVM_begin_signal_setting"));
4613 if (begin_signal_setting != NULL) {
4614 end_signal_setting = CAST_TO_FN_PTR(signal_setting_t,
4615 dlsym(RTLD_DEFAULT, "JVM_end_signal_setting"));
4616 get_signal_action = CAST_TO_FN_PTR(get_signal_t,
4617 dlsym(RTLD_DEFAULT, "JVM_get_signal_action"));
4618 libjsig_is_loaded = true;
4619 assert(UseSignalChaining, "should enable signal-chaining");
4620 }
4621 if (libjsig_is_loaded) {
4622 // Tell libjsig jvm is setting signal handlers
4623 (*begin_signal_setting)();
4624 }
4625
4626 set_signal_handler(SIGSEGV, true);
4627 set_signal_handler(SIGPIPE, true);
4628 set_signal_handler(SIGBUS, true);
4629 set_signal_handler(SIGILL, true);
4630 set_signal_handler(SIGFPE, true);
4631 #if defined(PPC64)
4632 set_signal_handler(SIGTRAP, true);
4633 #endif
4634 set_signal_handler(SIGXFSZ, true);
4635
4636 if (libjsig_is_loaded) {
4637 // Tell libjsig jvm finishes setting signal handlers
4638 (*end_signal_setting)();
4639 }
4640
4641 // We don't activate signal checker if libjsig is in place, we trust ourselves
4642 // and if UserSignalHandler is installed all bets are off.
4643 // Log that signal checking is off only if -verbose:jni is specified.
4644 if (CheckJNICalls) {
4645 if (libjsig_is_loaded) {
4646 if (PrintJNIResolving) {
4647 tty->print_cr("Info: libjsig is activated, all active signal checking is disabled");
4648 }
4649 check_signals = false;
4650 }
4651 if (AllowUserSignalHandlers) {
4652 if (PrintJNIResolving) {
4653 tty->print_cr("Info: AllowUserSignalHandlers is activated, all active signal checking is disabled");
4654 }
4655 check_signals = false;
4656 }
4657 }
4658 }
4659 }
4660
4661 // This is the fastest way to get thread cpu time on Linux.
4662 // Returns cpu time (user+sys) for any thread, not only for current.
4663 // POSIX compliant clocks are implemented in the kernels 2.6.16+.
4664 // It might work on 2.6.10+ with a special kernel/glibc patch.
4665 // For reference, please, see IEEE Std 1003.1-2004:
4666 // http://www.unix.org/single_unix_specification
4667
4668 jlong os::Linux::fast_thread_cpu_time(clockid_t clockid) {
4669 struct timespec tp;
4670 int rc = os::Linux::clock_gettime(clockid, &tp);
4671 assert(rc == 0, "clock_gettime is expected to return 0 code");
4672
4673 return (tp.tv_sec * NANOSECS_PER_SEC) + tp.tv_nsec;
4674 }
4675
4676 /////
4677 // glibc on Linux platform uses non-documented flag
4678 // to indicate, that some special sort of signal
4679 // trampoline is used.
4680 // We will never set this flag, and we should
4681 // ignore this flag in our diagnostic
4682 #ifdef SIGNIFICANT_SIGNAL_MASK
4683 #undef SIGNIFICANT_SIGNAL_MASK
4684 #endif
4685 #define SIGNIFICANT_SIGNAL_MASK (~0x04000000)
4686
4687 static const char* get_signal_handler_name(address handler,
4688 char* buf, int buflen) {
4689 int offset = 0;
4690 bool found = os::dll_address_to_library_name(handler, buf, buflen, &offset);
4691 if (found) {
4692 // skip directory names
4693 const char *p1, *p2;
4694 p1 = buf;
4695 size_t len = strlen(os::file_separator());
4696 while ((p2 = strstr(p1, os::file_separator())) != NULL) p1 = p2 + len;
4697 jio_snprintf(buf, buflen, "%s+0x%x", p1, offset);
4698 } else {
4699 jio_snprintf(buf, buflen, PTR_FORMAT, handler);
4700 }
4701 return buf;
4702 }
4703
4704 static void print_signal_handler(outputStream* st, int sig,
4705 char* buf, size_t buflen) {
4706 struct sigaction sa;
4707
4708 sigaction(sig, NULL, &sa);
4709
4710 // See comment for SIGNIFICANT_SIGNAL_MASK define
4711 sa.sa_flags &= SIGNIFICANT_SIGNAL_MASK;
4712
4713 st->print("%s: ", os::exception_name(sig, buf, buflen));
4714
4715 address handler = (sa.sa_flags & SA_SIGINFO)
4716 ? CAST_FROM_FN_PTR(address, sa.sa_sigaction)
4717 : CAST_FROM_FN_PTR(address, sa.sa_handler);
4718
4719 if (handler == CAST_FROM_FN_PTR(address, SIG_DFL)) {
4720 st->print("SIG_DFL");
4721 } else if (handler == CAST_FROM_FN_PTR(address, SIG_IGN)) {
4722 st->print("SIG_IGN");
4723 } else {
4724 st->print("[%s]", get_signal_handler_name(handler, buf, buflen));
4725 }
4726
4727 st->print(", sa_mask[0]=");
4728 os::Posix::print_signal_set_short(st, &sa.sa_mask);
4729
4730 address rh = VMError::get_resetted_sighandler(sig);
4731 // May be, handler was resetted by VMError?
4732 if(rh != NULL) {
4733 handler = rh;
4734 sa.sa_flags = VMError::get_resetted_sigflags(sig) & SIGNIFICANT_SIGNAL_MASK;
4735 }
4736
4737 st->print(", sa_flags=");
4738 os::Posix::print_sa_flags(st, sa.sa_flags);
4739
4740 // Check: is it our handler?
4741 if(handler == CAST_FROM_FN_PTR(address, (sa_sigaction_t)signalHandler) ||
4742 handler == CAST_FROM_FN_PTR(address, (sa_sigaction_t)SR_handler)) {
4743 // It is our signal handler
4744 // check for flags, reset system-used one!
4745 if((int)sa.sa_flags != os::Linux::get_our_sigflags(sig)) {
4746 st->print(
4747 ", flags was changed from " PTR32_FORMAT ", consider using jsig library",
4748 os::Linux::get_our_sigflags(sig));
4749 }
4750 }
4751 st->cr();
4752 }
4753
4754
4755 #define DO_SIGNAL_CHECK(sig) \
4756 if (!sigismember(&check_signal_done, sig)) \
4757 os::Linux::check_signal_handler(sig)
4758
4759 // This method is a periodic task to check for misbehaving JNI applications
4760 // under CheckJNI, we can add any periodic checks here
4761
4762 void os::run_periodic_checks() {
4763
4764 if (check_signals == false) return;
4765
4766 // SEGV and BUS if overridden could potentially prevent
4767 // generation of hs*.log in the event of a crash, debugging
4768 // such a case can be very challenging, so we absolutely
4769 // check the following for a good measure:
4770 DO_SIGNAL_CHECK(SIGSEGV);
4771 DO_SIGNAL_CHECK(SIGILL);
4772 DO_SIGNAL_CHECK(SIGFPE);
4773 DO_SIGNAL_CHECK(SIGBUS);
4774 DO_SIGNAL_CHECK(SIGPIPE);
4775 DO_SIGNAL_CHECK(SIGXFSZ);
4776 #if defined(PPC64)
4777 DO_SIGNAL_CHECK(SIGTRAP);
4778 #endif
4779
4780 // ReduceSignalUsage allows the user to override these handlers
4781 // see comments at the very top and jvm_solaris.h
4782 if (!ReduceSignalUsage) {
4783 DO_SIGNAL_CHECK(SHUTDOWN1_SIGNAL);
4784 DO_SIGNAL_CHECK(SHUTDOWN2_SIGNAL);
4785 DO_SIGNAL_CHECK(SHUTDOWN3_SIGNAL);
4786 DO_SIGNAL_CHECK(BREAK_SIGNAL);
4787 }
4788
4789 DO_SIGNAL_CHECK(SR_signum);
4790 DO_SIGNAL_CHECK(INTERRUPT_SIGNAL);
4791 }
4792
4793 typedef int (*os_sigaction_t)(int, const struct sigaction *, struct sigaction *);
4794
4795 static os_sigaction_t os_sigaction = NULL;
4796
4797 void os::Linux::check_signal_handler(int sig) {
4798 char buf[O_BUFLEN];
4799 address jvmHandler = NULL;
4800
4801
4802 struct sigaction act;
4803 if (os_sigaction == NULL) {
4804 // only trust the default sigaction, in case it has been interposed
4805 os_sigaction = (os_sigaction_t)dlsym(RTLD_DEFAULT, "sigaction");
4806 if (os_sigaction == NULL) return;
4807 }
4808
4809 os_sigaction(sig, (struct sigaction*)NULL, &act);
4810
4811
4812 act.sa_flags &= SIGNIFICANT_SIGNAL_MASK;
4813
4814 address thisHandler = (act.sa_flags & SA_SIGINFO)
4815 ? CAST_FROM_FN_PTR(address, act.sa_sigaction)
4816 : CAST_FROM_FN_PTR(address, act.sa_handler) ;
4817
4818
4819 switch(sig) {
4820 case SIGSEGV:
4821 case SIGBUS:
4822 case SIGFPE:
4823 case SIGPIPE:
4824 case SIGILL:
4825 case SIGXFSZ:
4826 jvmHandler = CAST_FROM_FN_PTR(address, (sa_sigaction_t)signalHandler);
4827 break;
4828
4829 case SHUTDOWN1_SIGNAL:
4830 case SHUTDOWN2_SIGNAL:
4831 case SHUTDOWN3_SIGNAL:
4832 case BREAK_SIGNAL:
4833 jvmHandler = (address)user_handler();
4834 break;
4835
4836 case INTERRUPT_SIGNAL:
4837 jvmHandler = CAST_FROM_FN_PTR(address, SIG_DFL);
4838 break;
4839
4840 default:
4841 if (sig == SR_signum) {
4842 jvmHandler = CAST_FROM_FN_PTR(address, (sa_sigaction_t)SR_handler);
4843 } else {
4844 return;
4845 }
4846 break;
4847 }
4848
4849 if (thisHandler != jvmHandler) {
4850 tty->print("Warning: %s handler ", exception_name(sig, buf, O_BUFLEN));
4851 tty->print("expected:%s", get_signal_handler_name(jvmHandler, buf, O_BUFLEN));
4852 tty->print_cr(" found:%s", get_signal_handler_name(thisHandler, buf, O_BUFLEN));
4853 // No need to check this sig any longer
4854 sigaddset(&check_signal_done, sig);
4855 // Running under non-interactive shell, SHUTDOWN2_SIGNAL will be reassigned SIG_IGN
4856 if (sig == SHUTDOWN2_SIGNAL && !isatty(fileno(stdin))) {
4857 tty->print_cr("Running in non-interactive shell, %s handler is replaced by shell",
4858 exception_name(sig, buf, O_BUFLEN));
4859 }
4860 } else if(os::Linux::get_our_sigflags(sig) != 0 && (int)act.sa_flags != os::Linux::get_our_sigflags(sig)) {
4861 tty->print("Warning: %s handler flags ", exception_name(sig, buf, O_BUFLEN));
4862 tty->print("expected:" PTR32_FORMAT, os::Linux::get_our_sigflags(sig));
4863 tty->print_cr(" found:" PTR32_FORMAT, act.sa_flags);
4864 // No need to check this sig any longer
4865 sigaddset(&check_signal_done, sig);
4866 }
4867
4868 // Dump all the signal
4869 if (sigismember(&check_signal_done, sig)) {
4870 print_signal_handlers(tty, buf, O_BUFLEN);
4871 }
4872 }
4873
4874 extern void report_error(char* file_name, int line_no, char* title, char* format, ...);
4875
4876 extern bool signal_name(int signo, char* buf, size_t len);
4877
4878 const char* os::exception_name(int exception_code, char* buf, size_t size) {
4879 if (0 < exception_code && exception_code <= SIGRTMAX) {
4880 // signal
4881 if (!signal_name(exception_code, buf, size)) {
4882 jio_snprintf(buf, size, "SIG%d", exception_code);
4883 }
4884 return buf;
4885 } else {
4886 return NULL;
4887 }
4888 }
4889
4890 // this is called _before_ most of the global arguments have been parsed
4891 void os::init(void) {
4892 char dummy; /* used to get a guess on initial stack address */
4893
4894 // With LinuxThreads the JavaMain thread pid (primordial thread)
4895 // is different than the pid of the java launcher thread.
4896 // So, on Linux, the launcher thread pid is passed to the VM
4897 // via the sun.java.launcher.pid property.
4898 // Use this property instead of getpid() if it was correctly passed.
4899 // See bug 6351349.
4900 pid_t java_launcher_pid = (pid_t) Arguments::sun_java_launcher_pid();
4901
4902 _initial_pid = (java_launcher_pid > 0) ? java_launcher_pid : getpid();
4903
4904 clock_tics_per_sec = sysconf(_SC_CLK_TCK);
4905
4906 init_random(1234567);
4907
4908 ThreadCritical::initialize();
4909
4910 Linux::set_page_size(sysconf(_SC_PAGESIZE));
4911 if (Linux::page_size() == -1) {
4912 fatal(err_msg("os_linux.cpp: os::init: sysconf failed (%s)",
4913 strerror(errno)));
4914 }
4915 init_page_sizes((size_t) Linux::page_size());
4916
4917 Linux::initialize_system_info();
4918
4919 // _main_thread points to the thread that created/loaded the JVM.
4920 Linux::_main_thread = pthread_self();
4921
4922 Linux::clock_init();
4923 initial_time_count = javaTimeNanos();
4924
4925 // pthread_condattr initialization for monotonic clock
4926 int status;
4927 pthread_condattr_t* _condattr = os::Linux::condAttr();
4928 if ((status = pthread_condattr_init(_condattr)) != 0) {
4929 fatal(err_msg("pthread_condattr_init: %s", strerror(status)));
4930 }
4931 // Only set the clock if CLOCK_MONOTONIC is available
4932 if (Linux::supports_monotonic_clock()) {
4933 if ((status = pthread_condattr_setclock(_condattr, CLOCK_MONOTONIC)) != 0) {
4934 if (status == EINVAL) {
4935 warning("Unable to use monotonic clock with relative timed-waits" \
4936 " - changes to the time-of-day clock may have adverse affects");
4937 } else {
4938 fatal(err_msg("pthread_condattr_setclock: %s", strerror(status)));
4939 }
4940 }
4941 }
4942 // else it defaults to CLOCK_REALTIME
4943
4944 pthread_mutex_init(&dl_mutex, NULL);
4945
4946 // If the pagesize of the VM is greater than 8K determine the appropriate
4947 // number of initial guard pages. The user can change this with the
4948 // command line arguments, if needed.
4949 if (vm_page_size() > (int)Linux::vm_default_page_size()) {
4950 StackYellowPages = 1;
4951 StackRedPages = 1;
4952 StackShadowPages = round_to((StackShadowPages*Linux::vm_default_page_size()), vm_page_size()) / vm_page_size();
4953 }
4954 }
4955
4956 // To install functions for atexit system call
4957 extern "C" {
4958 static void perfMemory_exit_helper() {
4959 perfMemory_exit();
4960 }
4961 }
4962
4963 // this is called _after_ the global arguments have been parsed
4964 jint os::init_2(void)
4965 {
4966 Linux::fast_thread_clock_init();
4967
4968 // Allocate a single page and mark it as readable for safepoint polling
4969 address polling_page = (address) ::mmap(NULL, Linux::page_size(), PROT_READ, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
4970 guarantee( polling_page != MAP_FAILED, "os::init_2: failed to allocate polling page" );
4971
4972 os::set_polling_page( polling_page );
4973
4974 #ifndef PRODUCT
4975 if(Verbose && PrintMiscellaneous)
4976 tty->print("[SafePoint Polling address: " INTPTR_FORMAT "]\n", (intptr_t)polling_page);
4977 #endif
4978
4979 if (!UseMembar) {
4980 address mem_serialize_page = (address) ::mmap(NULL, Linux::page_size(), PROT_READ | PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
4981 guarantee( mem_serialize_page != MAP_FAILED, "mmap Failed for memory serialize page");
4982 os::set_memory_serialize_page( mem_serialize_page );
4983
4984 #ifndef PRODUCT
4985 if(Verbose && PrintMiscellaneous)
4986 tty->print("[Memory Serialize Page address: " INTPTR_FORMAT "]\n", (intptr_t)mem_serialize_page);
4987 #endif
4988 }
4989
4990 // initialize suspend/resume support - must do this before signal_sets_init()
4991 if (SR_initialize() != 0) {
4992 perror("SR_initialize failed");
4993 return JNI_ERR;
4994 }
4995
4996 Linux::signal_sets_init();
4997 Linux::install_signal_handlers();
4998
4999 // Check minimum allowable stack size for thread creation and to initialize
5000 // the java system classes, including StackOverflowError - depends on page
5001 // size. Add a page for compiler2 recursion in main thread.
5002 // Add in 2*BytesPerWord times page size to account for VM stack during
5003 // class initialization depending on 32 or 64 bit VM.
5004 os::Linux::min_stack_allowed = MAX2(os::Linux::min_stack_allowed,
5005 (size_t)(StackYellowPages+StackRedPages+StackShadowPages) * Linux::page_size() +
5006 (2*BytesPerWord COMPILER2_PRESENT(+1)) * Linux::vm_default_page_size());
5007
5008 size_t threadStackSizeInBytes = ThreadStackSize * K;
5009 if (threadStackSizeInBytes != 0 &&
5010 threadStackSizeInBytes < os::Linux::min_stack_allowed) {
5011 tty->print_cr("\nThe stack size specified is too small, "
5012 "Specify at least %dk",
5013 os::Linux::min_stack_allowed/ K);
5014 return JNI_ERR;
5015 }
5016
5017 // Make the stack size a multiple of the page size so that
5018 // the yellow/red zones can be guarded.
5019 JavaThread::set_stack_size_at_create(round_to(threadStackSizeInBytes,
5020 vm_page_size()));
5021
5022 Linux::capture_initial_stack(JavaThread::stack_size_at_create());
5023
5024 #if defined(IA32)
5025 workaround_expand_exec_shield_cs_limit();
5026 #endif
5027
5028 Linux::libpthread_init();
5029 if (PrintMiscellaneous && (Verbose || WizardMode)) {
5030 tty->print_cr("[HotSpot is running with %s, %s(%s)]\n",
5031 Linux::glibc_version(), Linux::libpthread_version(),
5032 Linux::is_floating_stack() ? "floating stack" : "fixed stack");
5033 }
5034
5035 if (UseNUMA) {
5036 if (!Linux::libnuma_init()) {
5037 UseNUMA = false;
5038 } else {
5039 if ((Linux::numa_max_node() < 1)) {
5040 // There's only one node(they start from 0), disable NUMA.
5041 UseNUMA = false;
5042 }
5043 }
5044 // With SHM and HugeTLBFS large pages we cannot uncommit a page, so there's no way
5045 // we can make the adaptive lgrp chunk resizing work. If the user specified
5046 // both UseNUMA and UseLargePages (or UseSHM/UseHugeTLBFS) on the command line - warn and
5047 // disable adaptive resizing.
5048 if (UseNUMA && UseLargePages && !can_commit_large_page_memory()) {
5049 if (FLAG_IS_DEFAULT(UseNUMA)) {
5050 UseNUMA = false;
5051 } else {
5052 if (FLAG_IS_DEFAULT(UseLargePages) &&
5053 FLAG_IS_DEFAULT(UseSHM) &&
5054 FLAG_IS_DEFAULT(UseHugeTLBFS)) {
5055 UseLargePages = false;
5056 } else {
5057 warning("UseNUMA is not fully compatible with SHM/HugeTLBFS large pages, disabling adaptive resizing");
5058 UseAdaptiveSizePolicy = false;
5059 UseAdaptiveNUMAChunkSizing = false;
5060 }
5061 }
5062 }
5063 if (!UseNUMA && ForceNUMA) {
5064 UseNUMA = true;
5065 }
5066 }
5067
5068 if (MaxFDLimit) {
5069 // set the number of file descriptors to max. print out error
5070 // if getrlimit/setrlimit fails but continue regardless.
5071 struct rlimit nbr_files;
5072 int status = getrlimit(RLIMIT_NOFILE, &nbr_files);
5073 if (status != 0) {
5074 if (PrintMiscellaneous && (Verbose || WizardMode))
5075 perror("os::init_2 getrlimit failed");
5076 } else {
5077 nbr_files.rlim_cur = nbr_files.rlim_max;
5078 status = setrlimit(RLIMIT_NOFILE, &nbr_files);
5079 if (status != 0) {
5080 if (PrintMiscellaneous && (Verbose || WizardMode))
5081 perror("os::init_2 setrlimit failed");
5082 }
5083 }
5084 }
5085
5086 // Initialize lock used to serialize thread creation (see os::create_thread)
5087 Linux::set_createThread_lock(new Mutex(Mutex::leaf, "createThread_lock", false));
5088
5089 // at-exit methods are called in the reverse order of their registration.
5090 // atexit functions are called on return from main or as a result of a
5091 // call to exit(3C). There can be only 32 of these functions registered
5092 // and atexit() does not set errno.
5093
5094 if (PerfAllowAtExitRegistration) {
5095 // only register atexit functions if PerfAllowAtExitRegistration is set.
5096 // atexit functions can be delayed until process exit time, which
5097 // can be problematic for embedded VM situations. Embedded VMs should
5098 // call DestroyJavaVM() to assure that VM resources are released.
5099
5100 // note: perfMemory_exit_helper atexit function may be removed in
5101 // the future if the appropriate cleanup code can be added to the
5102 // VM_Exit VMOperation's doit method.
5103 if (atexit(perfMemory_exit_helper) != 0) {
5104 warning("os::init_2 atexit(perfMemory_exit_helper) failed");
5105 }
5106 }
5107
5108 // initialize thread priority policy
5109 prio_init();
5110
5111 return JNI_OK;
5112 }
5113
5114 // Mark the polling page as unreadable
5115 void os::make_polling_page_unreadable(void) {
5116 if( !guard_memory((char*)_polling_page, Linux::page_size()) )
5117 fatal("Could not disable polling page");
5118 };
5119
5120 // Mark the polling page as readable
5121 void os::make_polling_page_readable(void) {
5122 if( !linux_mprotect((char *)_polling_page, Linux::page_size(), PROT_READ)) {
5123 fatal("Could not enable polling page");
5124 }
5125 };
5126
5127 static int os_cpu_count(const cpu_set_t* cpus) {
5128 int count = 0;
5129 // only look up to the number of configured processors
5130 for (int i = 0; i < os::processor_count(); i++) {
5131 if (CPU_ISSET(i, cpus)) {
5132 count++;
5133 }
5134 }
5135 return count;
5136 }
5137
5138 // Get the current number of available processors for this process.
5139 // This value can change at any time during a process's lifetime.
5140 // sched_getaffinity gives an accurate answer as it accounts for cpusets.
5141 // If anything goes wrong we fallback to returning the number of online
5142 // processors - which can be greater than the number available to the process.
5143 int os::active_processor_count() {
5144 cpu_set_t cpus; // can represent at most 1024 (CPU_SETSIZE) processors
5145 int cpus_size = sizeof(cpu_set_t);
5146 int cpu_count = 0;
5147
5148 // pid 0 means the current thread - which we have to assume represents the process
5149 if (sched_getaffinity(0, cpus_size, &cpus) == 0) {
5150 cpu_count = os_cpu_count(&cpus);
5151 if (PrintActiveCpus) {
5152 tty->print_cr("active_processor_count: sched_getaffinity processor count: %d", cpu_count);
5153 }
5154 }
5155 else {
5156 cpu_count = ::sysconf(_SC_NPROCESSORS_ONLN);
5157 warning("sched_getaffinity failed (%s)- using online processor count (%d) "
5158 "which may exceed available processors", strerror(errno), cpu_count);
5159 }
5160
5161 assert(cpu_count > 0 && cpu_count <= processor_count(), "sanity check");
5162 return cpu_count;
5163 }
5164
5165 void os::set_native_thread_name(const char *name) {
5166 // Not yet implemented.
5167 return;
5168 }
5169
5170 bool os::distribute_processes(uint length, uint* distribution) {
5171 // Not yet implemented.
5172 return false;
5173 }
5174
5175 bool os::bind_to_processor(uint processor_id) {
5176 // Not yet implemented.
5177 return false;
5178 }
5179
5180 ///
5181
5182 void os::SuspendedThreadTask::internal_do_task() {
5183 if (do_suspend(_thread->osthread())) {
5184 SuspendedThreadTaskContext context(_thread, _thread->osthread()->ucontext());
5185 do_task(context);
5186 do_resume(_thread->osthread());
5187 }
5188 }
5189
5190 class PcFetcher : public os::SuspendedThreadTask {
5191 public:
5192 PcFetcher(Thread* thread) : os::SuspendedThreadTask(thread) {}
5193 ExtendedPC result();
5194 protected:
5195 void do_task(const os::SuspendedThreadTaskContext& context);
5196 private:
5197 ExtendedPC _epc;
5198 };
5199
5200 ExtendedPC PcFetcher::result() {
5201 guarantee(is_done(), "task is not done yet.");
5202 return _epc;
5203 }
5204
5205 void PcFetcher::do_task(const os::SuspendedThreadTaskContext& context) {
5206 Thread* thread = context.thread();
5207 OSThread* osthread = thread->osthread();
5208 if (osthread->ucontext() != NULL) {
5209 _epc = os::Linux::ucontext_get_pc((ucontext_t *) context.ucontext());
5210 } else {
5211 // NULL context is unexpected, double-check this is the VMThread
5212 guarantee(thread->is_VM_thread(), "can only be called for VMThread");
5213 }
5214 }
5215
5216 // Suspends the target using the signal mechanism and then grabs the PC before
5217 // resuming the target. Used by the flat-profiler only
5218 ExtendedPC os::get_thread_pc(Thread* thread) {
5219 // Make sure that it is called by the watcher for the VMThread
5220 assert(Thread::current()->is_Watcher_thread(), "Must be watcher");
5221 assert(thread->is_VM_thread(), "Can only be called for VMThread");
5222
5223 PcFetcher fetcher(thread);
5224 fetcher.run();
5225 return fetcher.result();
5226 }
5227
5228 int os::Linux::safe_cond_timedwait(pthread_cond_t *_cond, pthread_mutex_t *_mutex, const struct timespec *_abstime)
5229 {
5230 if (is_NPTL()) {
5231 return pthread_cond_timedwait(_cond, _mutex, _abstime);
5232 } else {
5233 // 6292965: LinuxThreads pthread_cond_timedwait() resets FPU control
5234 // word back to default 64bit precision if condvar is signaled. Java
5235 // wants 53bit precision. Save and restore current value.
5236 int fpu = get_fpu_control_word();
5237 int status = pthread_cond_timedwait(_cond, _mutex, _abstime);
5238 set_fpu_control_word(fpu);
5239 return status;
5240 }
5241 }
5242
5243 ////////////////////////////////////////////////////////////////////////////////
5244 // debug support
5245
5246 bool os::find(address addr, outputStream* st) {
5247 Dl_info dlinfo;
5248 memset(&dlinfo, 0, sizeof(dlinfo));
5249 if (dladdr(addr, &dlinfo) != 0) {
5250 st->print(PTR_FORMAT ": ", addr);
5251 if (dlinfo.dli_sname != NULL && dlinfo.dli_saddr != NULL) {
5252 st->print("%s+%#x", dlinfo.dli_sname,
5253 addr - (intptr_t)dlinfo.dli_saddr);
5254 } else if (dlinfo.dli_fbase != NULL) {
5255 st->print("<offset %#x>", addr - (intptr_t)dlinfo.dli_fbase);
5256 } else {
5257 st->print("<absolute address>");
5258 }
5259 if (dlinfo.dli_fname != NULL) {
5260 st->print(" in %s", dlinfo.dli_fname);
5261 }
5262 if (dlinfo.dli_fbase != NULL) {
5263 st->print(" at " PTR_FORMAT, dlinfo.dli_fbase);
5264 }
5265 st->cr();
5266
5267 if (Verbose) {
5268 // decode some bytes around the PC
5269 address begin = clamp_address_in_page(addr-40, addr, os::vm_page_size());
5270 address end = clamp_address_in_page(addr+40, addr, os::vm_page_size());
5271 address lowest = (address) dlinfo.dli_sname;
5272 if (!lowest) lowest = (address) dlinfo.dli_fbase;
5273 if (begin < lowest) begin = lowest;
5274 Dl_info dlinfo2;
5275 if (dladdr(end, &dlinfo2) != 0 && dlinfo2.dli_saddr != dlinfo.dli_saddr
5276 && end > dlinfo2.dli_saddr && dlinfo2.dli_saddr > begin)
5277 end = (address) dlinfo2.dli_saddr;
5278 Disassembler::decode(begin, end, st);
5279 }
5280 return true;
5281 }
5282 return false;
5283 }
5284
5285 ////////////////////////////////////////////////////////////////////////////////
5286 // misc
5287
5288 // This does not do anything on Linux. This is basically a hook for being
5289 // able to use structured exception handling (thread-local exception filters)
5290 // on, e.g., Win32.
5291 void
5292 os::os_exception_wrapper(java_call_t f, JavaValue* value, methodHandle* method,
5293 JavaCallArguments* args, Thread* thread) {
5294 f(value, method, args, thread);
5295 }
5296
5297 void os::print_statistics() {
5298 }
5299
5300 int os::message_box(const char* title, const char* message) {
5301 int i;
5302 fdStream err(defaultStream::error_fd());
5303 for (i = 0; i < 78; i++) err.print_raw("=");
5304 err.cr();
5305 err.print_raw_cr(title);
5306 for (i = 0; i < 78; i++) err.print_raw("-");
5307 err.cr();
5308 err.print_raw_cr(message);
5309 for (i = 0; i < 78; i++) err.print_raw("=");
5310 err.cr();
5311
5312 char buf[16];
5313 // Prevent process from exiting upon "read error" without consuming all CPU
5314 while (::read(0, buf, sizeof(buf)) <= 0) { ::sleep(100); }
5315
5316 return buf[0] == 'y' || buf[0] == 'Y';
5317 }
5318
5319 int os::stat(const char *path, struct stat *sbuf) {
5320 char pathbuf[MAX_PATH];
5321 if (strlen(path) > MAX_PATH - 1) {
5322 errno = ENAMETOOLONG;
5323 return -1;
5324 }
5325 os::native_path(strcpy(pathbuf, path));
5326 return ::stat(pathbuf, sbuf);
5327 }
5328
5329 bool os::check_heap(bool force) {
5330 return true;
5331 }
5332
5333 int local_vsnprintf(char* buf, size_t count, const char* format, va_list args) {
5334 return ::vsnprintf(buf, count, format, args);
5335 }
5336
5337 // Is a (classpath) directory empty?
5338 bool os::dir_is_empty(const char* path) {
5339 DIR *dir = NULL;
5340 struct dirent *ptr;
5341
5342 dir = opendir(path);
5343 if (dir == NULL) return true;
5344
5345 /* Scan the directory */
5346 bool result = true;
5347 char buf[sizeof(struct dirent) + MAX_PATH];
5348 while (result && (ptr = ::readdir(dir)) != NULL) {
5349 if (strcmp(ptr->d_name, ".") != 0 && strcmp(ptr->d_name, "..") != 0) {
5350 result = false;
5351 }
5352 }
5353 closedir(dir);
5354 return result;
5355 }
5356
5357 // This code originates from JDK's sysOpen and open64_w
5358 // from src/solaris/hpi/src/system_md.c
5359
5360 #ifndef O_DELETE
5361 #define O_DELETE 0x10000
5362 #endif
5363
5364 // Open a file. Unlink the file immediately after open returns
5365 // if the specified oflag has the O_DELETE flag set.
5366 // O_DELETE is used only in j2se/src/share/native/java/util/zip/ZipFile.c
5367
5368 int os::open(const char *path, int oflag, int mode) {
5369
5370 if (strlen(path) > MAX_PATH - 1) {
5371 errno = ENAMETOOLONG;
5372 return -1;
5373 }
5374 int fd;
5375 int o_delete = (oflag & O_DELETE);
5376 oflag = oflag & ~O_DELETE;
5377
5378 fd = ::open64(path, oflag, mode);
5379 if (fd == -1) return -1;
5380
5381 //If the open succeeded, the file might still be a directory
5382 {
5383 struct stat64 buf64;
5384 int ret = ::fstat64(fd, &buf64);
5385 int st_mode = buf64.st_mode;
5386
5387 if (ret != -1) {
5388 if ((st_mode & S_IFMT) == S_IFDIR) {
5389 errno = EISDIR;
5390 ::close(fd);
5391 return -1;
5392 }
5393 } else {
5394 ::close(fd);
5395 return -1;
5396 }
5397 }
5398
5399 /*
5400 * All file descriptors that are opened in the JVM and not
5401 * specifically destined for a subprocess should have the
5402 * close-on-exec flag set. If we don't set it, then careless 3rd
5403 * party native code might fork and exec without closing all
5404 * appropriate file descriptors (e.g. as we do in closeDescriptors in
5405 * UNIXProcess.c), and this in turn might:
5406 *
5407 * - cause end-of-file to fail to be detected on some file
5408 * descriptors, resulting in mysterious hangs, or
5409 *
5410 * - might cause an fopen in the subprocess to fail on a system
5411 * suffering from bug 1085341.
5412 *
5413 * (Yes, the default setting of the close-on-exec flag is a Unix
5414 * design flaw)
5415 *
5416 * See:
5417 * 1085341: 32-bit stdio routines should support file descriptors >255
5418 * 4843136: (process) pipe file descriptor from Runtime.exec not being closed
5419 * 6339493: (process) Runtime.exec does not close all file descriptors on Solaris 9
5420 */
5421 #ifdef FD_CLOEXEC
5422 {
5423 int flags = ::fcntl(fd, F_GETFD);
5424 if (flags != -1)
5425 ::fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
5426 }
5427 #endif
5428
5429 if (o_delete != 0) {
5430 ::unlink(path);
5431 }
5432 return fd;
5433 }
5434
5435
5436 // create binary file, rewriting existing file if required
5437 int os::create_binary_file(const char* path, bool rewrite_existing) {
5438 int oflags = O_WRONLY | O_CREAT;
5439 if (!rewrite_existing) {
5440 oflags |= O_EXCL;
5441 }
5442 return ::open64(path, oflags, S_IREAD | S_IWRITE);
5443 }
5444
5445 // return current position of file pointer
5446 jlong os::current_file_offset(int fd) {
5447 return (jlong)::lseek64(fd, (off64_t)0, SEEK_CUR);
5448 }
5449
5450 // move file pointer to the specified offset
5451 jlong os::seek_to_file_offset(int fd, jlong offset) {
5452 return (jlong)::lseek64(fd, (off64_t)offset, SEEK_SET);
5453 }
5454
5455 // This code originates from JDK's sysAvailable
5456 // from src/solaris/hpi/src/native_threads/src/sys_api_td.c
5457
5458 int os::available(int fd, jlong *bytes) {
5459 jlong cur, end;
5460 int mode;
5461 struct stat64 buf64;
5462
5463 if (::fstat64(fd, &buf64) >= 0) {
5464 mode = buf64.st_mode;
5465 if (S_ISCHR(mode) || S_ISFIFO(mode) || S_ISSOCK(mode)) {
5466 /*
5467 * XXX: is the following call interruptible? If so, this might
5468 * need to go through the INTERRUPT_IO() wrapper as for other
5469 * blocking, interruptible calls in this file.
5470 */
5471 int n;
5472 if (::ioctl(fd, FIONREAD, &n) >= 0) {
5473 *bytes = n;
5474 return 1;
5475 }
5476 }
5477 }
5478 if ((cur = ::lseek64(fd, 0L, SEEK_CUR)) == -1) {
5479 return 0;
5480 } else if ((end = ::lseek64(fd, 0L, SEEK_END)) == -1) {
5481 return 0;
5482 } else if (::lseek64(fd, cur, SEEK_SET) == -1) {
5483 return 0;
5484 }
5485 *bytes = end - cur;
5486 return 1;
5487 }
5488
5489 int os::socket_available(int fd, jint *pbytes) {
5490 // Linux doc says EINTR not returned, unlike Solaris
5491 int ret = ::ioctl(fd, FIONREAD, pbytes);
5492
5493 //%% note ioctl can return 0 when successful, JVM_SocketAvailable
5494 // is expected to return 0 on failure and 1 on success to the jdk.
5495 return (ret < 0) ? 0 : 1;
5496 }
5497
5498 // Map a block of memory.
5499 char* os::pd_map_memory(int fd, const char* file_name, size_t file_offset,
5500 char *addr, size_t bytes, bool read_only,
5501 bool allow_exec) {
5502 int prot;
5503 int flags = MAP_PRIVATE;
5504
5505 if (read_only) {
5506 prot = PROT_READ;
5507 } else {
5508 prot = PROT_READ | PROT_WRITE;
5509 }
5510
5511 if (allow_exec) {
5512 prot |= PROT_EXEC;
5513 }
5514
5515 if (addr != NULL) {
5516 flags |= MAP_FIXED;
5517 }
5518
5519 char* mapped_address = (char*)mmap(addr, (size_t)bytes, prot, flags,
5520 fd, file_offset);
5521 if (mapped_address == MAP_FAILED) {
5522 return NULL;
5523 }
5524 return mapped_address;
5525 }
5526
5527
5528 // Remap a block of memory.
5529 char* os::pd_remap_memory(int fd, const char* file_name, size_t file_offset,
5530 char *addr, size_t bytes, bool read_only,
5531 bool allow_exec) {
5532 // same as map_memory() on this OS
5533 return os::map_memory(fd, file_name, file_offset, addr, bytes, read_only,
5534 allow_exec);
5535 }
5536
5537
5538 // Unmap a block of memory.
5539 bool os::pd_unmap_memory(char* addr, size_t bytes) {
5540 return munmap(addr, bytes) == 0;
5541 }
5542
5543 static jlong slow_thread_cpu_time(Thread *thread, bool user_sys_cpu_time);
5544
5545 static clockid_t thread_cpu_clockid(Thread* thread) {
5546 pthread_t tid = thread->osthread()->pthread_id();
5547 clockid_t clockid;
5548
5549 // Get thread clockid
5550 int rc = os::Linux::pthread_getcpuclockid(tid, &clockid);
5551 assert(rc == 0, "pthread_getcpuclockid is expected to return 0 code");
5552 return clockid;
5553 }
5554
5555 // current_thread_cpu_time(bool) and thread_cpu_time(Thread*, bool)
5556 // are used by JVM M&M and JVMTI to get user+sys or user CPU time
5557 // of a thread.
5558 //
5559 // current_thread_cpu_time() and thread_cpu_time(Thread*) returns
5560 // the fast estimate available on the platform.
5561
5562 jlong os::current_thread_cpu_time() {
5563 if (os::Linux::supports_fast_thread_cpu_time()) {
5564 return os::Linux::fast_thread_cpu_time(CLOCK_THREAD_CPUTIME_ID);
5565 } else {
5566 // return user + sys since the cost is the same
5567 return slow_thread_cpu_time(Thread::current(), true /* user + sys */);
5568 }
5569 }
5570
5571 jlong os::thread_cpu_time(Thread* thread) {
5572 // consistent with what current_thread_cpu_time() returns
5573 if (os::Linux::supports_fast_thread_cpu_time()) {
5574 return os::Linux::fast_thread_cpu_time(thread_cpu_clockid(thread));
5575 } else {
5576 return slow_thread_cpu_time(thread, true /* user + sys */);
5577 }
5578 }
5579
5580 jlong os::current_thread_cpu_time(bool user_sys_cpu_time) {
5581 if (user_sys_cpu_time && os::Linux::supports_fast_thread_cpu_time()) {
5582 return os::Linux::fast_thread_cpu_time(CLOCK_THREAD_CPUTIME_ID);
5583 } else {
5584 return slow_thread_cpu_time(Thread::current(), user_sys_cpu_time);
5585 }
5586 }
5587
5588 jlong os::thread_cpu_time(Thread *thread, bool user_sys_cpu_time) {
5589 if (user_sys_cpu_time && os::Linux::supports_fast_thread_cpu_time()) {
5590 return os::Linux::fast_thread_cpu_time(thread_cpu_clockid(thread));
5591 } else {
5592 return slow_thread_cpu_time(thread, user_sys_cpu_time);
5593 }
5594 }
5595
5596 //
5597 // -1 on error.
5598 //
5599
5600 PRAGMA_DIAG_PUSH
5601 PRAGMA_FORMAT_NONLITERAL_IGNORED
5602 static jlong slow_thread_cpu_time(Thread *thread, bool user_sys_cpu_time) {
5603 static bool proc_task_unchecked = true;
5604 static const char *proc_stat_path = "/proc/%d/stat";
5605 pid_t tid = thread->osthread()->thread_id();
5606 char *s;
5607 char stat[2048];
5608 int statlen;
5609 char proc_name[64];
5610 int count;
5611 long sys_time, user_time;
5612 char cdummy;
5613 int idummy;
5614 long ldummy;
5615 FILE *fp;
5616
5617 // The /proc/<tid>/stat aggregates per-process usage on
5618 // new Linux kernels 2.6+ where NPTL is supported.
5619 // The /proc/self/task/<tid>/stat still has the per-thread usage.
5620 // See bug 6328462.
5621 // There possibly can be cases where there is no directory
5622 // /proc/self/task, so we check its availability.
5623 if (proc_task_unchecked && os::Linux::is_NPTL()) {
5624 // This is executed only once
5625 proc_task_unchecked = false;
5626 fp = fopen("/proc/self/task", "r");
5627 if (fp != NULL) {
5628 proc_stat_path = "/proc/self/task/%d/stat";
5629 fclose(fp);
5630 }
5631 }
5632
5633 sprintf(proc_name, proc_stat_path, tid);
5634 fp = fopen(proc_name, "r");
5635 if ( fp == NULL ) return -1;
5636 statlen = fread(stat, 1, 2047, fp);
5637 stat[statlen] = '\0';
5638 fclose(fp);
5639
5640 // Skip pid and the command string. Note that we could be dealing with
5641 // weird command names, e.g. user could decide to rename java launcher
5642 // to "java 1.4.2 :)", then the stat file would look like
5643 // 1234 (java 1.4.2 :)) R ... ...
5644 // We don't really need to know the command string, just find the last
5645 // occurrence of ")" and then start parsing from there. See bug 4726580.
5646 s = strrchr(stat, ')');
5647 if (s == NULL ) return -1;
5648
5649 // Skip blank chars
5650 do s++; while (isspace(*s));
5651
5652 count = sscanf(s,"%c %d %d %d %d %d %lu %lu %lu %lu %lu %lu %lu",
5653 &cdummy, &idummy, &idummy, &idummy, &idummy, &idummy,
5654 &ldummy, &ldummy, &ldummy, &ldummy, &ldummy,
5655 &user_time, &sys_time);
5656 if ( count != 13 ) return -1;
5657 if (user_sys_cpu_time) {
5658 return ((jlong)sys_time + (jlong)user_time) * (1000000000 / clock_tics_per_sec);
5659 } else {
5660 return (jlong)user_time * (1000000000 / clock_tics_per_sec);
5661 }
5662 }
5663 PRAGMA_DIAG_POP
5664
5665 void os::current_thread_cpu_time_info(jvmtiTimerInfo *info_ptr) {
5666 info_ptr->max_value = ALL_64_BITS; // will not wrap in less than 64 bits
5667 info_ptr->may_skip_backward = false; // elapsed time not wall time
5668 info_ptr->may_skip_forward = false; // elapsed time not wall time
5669 info_ptr->kind = JVMTI_TIMER_TOTAL_CPU; // user+system time is returned
5670 }
5671
5672 void os::thread_cpu_time_info(jvmtiTimerInfo *info_ptr) {
5673 info_ptr->max_value = ALL_64_BITS; // will not wrap in less than 64 bits
5674 info_ptr->may_skip_backward = false; // elapsed time not wall time
5675 info_ptr->may_skip_forward = false; // elapsed time not wall time
5676 info_ptr->kind = JVMTI_TIMER_TOTAL_CPU; // user+system time is returned
5677 }
5678
5679 bool os::is_thread_cpu_time_supported() {
5680 return true;
5681 }
5682
5683 // System loadavg support. Returns -1 if load average cannot be obtained.
5684 // Linux doesn't yet have a (official) notion of processor sets,
5685 // so just return the system wide load average.
5686 int os::loadavg(double loadavg[], int nelem) {
5687 return ::getloadavg(loadavg, nelem);
5688 }
5689
5690 void os::pause() {
5691 char filename[MAX_PATH];
5692 if (PauseAtStartupFile && PauseAtStartupFile[0]) {
5693 jio_snprintf(filename, MAX_PATH, PauseAtStartupFile);
5694 } else {
5695 jio_snprintf(filename, MAX_PATH, "./vm.paused.%d", current_process_id());
5696 }
5697
5698 int fd = ::open(filename, O_WRONLY | O_CREAT | O_TRUNC, 0666);
5699 if (fd != -1) {
5700 struct stat buf;
5701 ::close(fd);
5702 while (::stat(filename, &buf) == 0) {
5703 (void)::poll(NULL, 0, 100);
5704 }
5705 } else {
5706 jio_fprintf(stderr,
5707 "Could not open pause file '%s', continuing immediately.\n", filename);
5708 }
5709 }
5710
5711
5712 // Refer to the comments in os_solaris.cpp park-unpark.
5713 //
5714 // Beware -- Some versions of NPTL embody a flaw where pthread_cond_timedwait() can
5715 // hang indefinitely. For instance NPTL 0.60 on 2.4.21-4ELsmp is vulnerable.
5716 // For specifics regarding the bug see GLIBC BUGID 261237 :
5717 // http://www.mail-archive.com/debian-glibc@lists.debian.org/msg10837.html.
5718 // Briefly, pthread_cond_timedwait() calls with an expiry time that's not in the future
5719 // will either hang or corrupt the condvar, resulting in subsequent hangs if the condvar
5720 // is used. (The simple C test-case provided in the GLIBC bug report manifests the
5721 // hang). The JVM is vulernable via sleep(), Object.wait(timo), LockSupport.parkNanos()
5722 // and monitorenter when we're using 1-0 locking. All those operations may result in
5723 // calls to pthread_cond_timedwait(). Using LD_ASSUME_KERNEL to use an older version
5724 // of libpthread avoids the problem, but isn't practical.
5725 //
5726 // Possible remedies:
5727 //
5728 // 1. Establish a minimum relative wait time. 50 to 100 msecs seems to work.
5729 // This is palliative and probabilistic, however. If the thread is preempted
5730 // between the call to compute_abstime() and pthread_cond_timedwait(), more
5731 // than the minimum period may have passed, and the abstime may be stale (in the
5732 // past) resultin in a hang. Using this technique reduces the odds of a hang
5733 // but the JVM is still vulnerable, particularly on heavily loaded systems.
5734 //
5735 // 2. Modify park-unpark to use per-thread (per ParkEvent) pipe-pairs instead
5736 // of the usual flag-condvar-mutex idiom. The write side of the pipe is set
5737 // NDELAY. unpark() reduces to write(), park() reduces to read() and park(timo)
5738 // reduces to poll()+read(). This works well, but consumes 2 FDs per extant
5739 // thread.
5740 //
5741 // 3. Embargo pthread_cond_timedwait() and implement a native "chron" thread
5742 // that manages timeouts. We'd emulate pthread_cond_timedwait() by enqueuing
5743 // a timeout request to the chron thread and then blocking via pthread_cond_wait().
5744 // This also works well. In fact it avoids kernel-level scalability impediments
5745 // on certain platforms that don't handle lots of active pthread_cond_timedwait()
5746 // timers in a graceful fashion.
5747 //
5748 // 4. When the abstime value is in the past it appears that control returns
5749 // correctly from pthread_cond_timedwait(), but the condvar is left corrupt.
5750 // Subsequent timedwait/wait calls may hang indefinitely. Given that, we
5751 // can avoid the problem by reinitializing the condvar -- by cond_destroy()
5752 // followed by cond_init() -- after all calls to pthread_cond_timedwait().
5753 // It may be possible to avoid reinitialization by checking the return
5754 // value from pthread_cond_timedwait(). In addition to reinitializing the
5755 // condvar we must establish the invariant that cond_signal() is only called
5756 // within critical sections protected by the adjunct mutex. This prevents
5757 // cond_signal() from "seeing" a condvar that's in the midst of being
5758 // reinitialized or that is corrupt. Sadly, this invariant obviates the
5759 // desirable signal-after-unlock optimization that avoids futile context switching.
5760 //
5761 // I'm also concerned that some versions of NTPL might allocate an auxilliary
5762 // structure when a condvar is used or initialized. cond_destroy() would
5763 // release the helper structure. Our reinitialize-after-timedwait fix
5764 // put excessive stress on malloc/free and locks protecting the c-heap.
5765 //
5766 // We currently use (4). See the WorkAroundNTPLTimedWaitHang flag.
5767 // It may be possible to refine (4) by checking the kernel and NTPL verisons
5768 // and only enabling the work-around for vulnerable environments.
5769
5770 // utility to compute the abstime argument to timedwait:
5771 // millis is the relative timeout time
5772 // abstime will be the absolute timeout time
5773 // TODO: replace compute_abstime() with unpackTime()
5774
5775 static struct timespec* compute_abstime(timespec* abstime, jlong millis) {
5776 if (millis < 0) millis = 0;
5777
5778 jlong seconds = millis / 1000;
5779 millis %= 1000;
5780 if (seconds > 50000000) { // see man cond_timedwait(3T)
5781 seconds = 50000000;
5782 }
5783
5784 if (os::Linux::supports_monotonic_clock()) {
5785 struct timespec now;
5786 int status = os::Linux::clock_gettime(CLOCK_MONOTONIC, &now);
5787 assert_status(status == 0, status, "clock_gettime");
5788 abstime->tv_sec = now.tv_sec + seconds;
5789 long nanos = now.tv_nsec + millis * NANOSECS_PER_MILLISEC;
5790 if (nanos >= NANOSECS_PER_SEC) {
5791 abstime->tv_sec += 1;
5792 nanos -= NANOSECS_PER_SEC;
5793 }
5794 abstime->tv_nsec = nanos;
5795 } else {
5796 struct timeval now;
5797 int status = gettimeofday(&now, NULL);
5798 assert(status == 0, "gettimeofday");
5799 abstime->tv_sec = now.tv_sec + seconds;
5800 long usec = now.tv_usec + millis * 1000;
5801 if (usec >= 1000000) {
5802 abstime->tv_sec += 1;
5803 usec -= 1000000;
5804 }
5805 abstime->tv_nsec = usec * 1000;
5806 }
5807 return abstime;
5808 }
5809
5810
5811 // Test-and-clear _Event, always leaves _Event set to 0, returns immediately.
5812 // Conceptually TryPark() should be equivalent to park(0).
5813
5814 int os::PlatformEvent::TryPark() {
5815 for (;;) {
5816 const int v = _Event ;
5817 guarantee ((v == 0) || (v == 1), "invariant") ;
5818 if (Atomic::cmpxchg (0, &_Event, v) == v) return v ;
5819 }
5820 }
5821
5822 void os::PlatformEvent::park() { // AKA "down()"
5823 // Invariant: Only the thread associated with the Event/PlatformEvent
5824 // may call park().
5825 // TODO: assert that _Assoc != NULL or _Assoc == Self
5826 int v ;
5827 for (;;) {
5828 v = _Event ;
5829 if (Atomic::cmpxchg (v-1, &_Event, v) == v) break ;
5830 }
5831 guarantee (v >= 0, "invariant") ;
5832 if (v == 0) {
5833 // Do this the hard way by blocking ...
5834 int status = pthread_mutex_lock(_mutex);
5835 assert_status(status == 0, status, "mutex_lock");
5836 guarantee (_nParked == 0, "invariant") ;
5837 ++ _nParked ;
5838 while (_Event < 0) {
5839 status = pthread_cond_wait(_cond, _mutex);
5840 // for some reason, under 2.7 lwp_cond_wait() may return ETIME ...
5841 // Treat this the same as if the wait was interrupted
5842 if (status == ETIME) { status = EINTR; }
5843 assert_status(status == 0 || status == EINTR, status, "cond_wait");
5844 }
5845 -- _nParked ;
5846
5847 _Event = 0 ;
5848 status = pthread_mutex_unlock(_mutex);
5849 assert_status(status == 0, status, "mutex_unlock");
5850 // Paranoia to ensure our locked and lock-free paths interact
5851 // correctly with each other.
5852 OrderAccess::fence();
5853 }
5854 guarantee (_Event >= 0, "invariant") ;
5855 }
5856
5857 int os::PlatformEvent::park(jlong millis) {
5858 guarantee (_nParked == 0, "invariant") ;
5859
5860 int v ;
5861 for (;;) {
5862 v = _Event ;
5863 if (Atomic::cmpxchg (v-1, &_Event, v) == v) break ;
5864 }
5865 guarantee (v >= 0, "invariant") ;
5866 if (v != 0) return OS_OK ;
5867
5868 // We do this the hard way, by blocking the thread.
5869 // Consider enforcing a minimum timeout value.
5870 struct timespec abst;
5871 compute_abstime(&abst, millis);
5872
5873 int ret = OS_TIMEOUT;
5874 int status = pthread_mutex_lock(_mutex);
5875 assert_status(status == 0, status, "mutex_lock");
5876 guarantee (_nParked == 0, "invariant") ;
5877 ++_nParked ;
5878
5879 // Object.wait(timo) will return because of
5880 // (a) notification
5881 // (b) timeout
5882 // (c) thread.interrupt
5883 //
5884 // Thread.interrupt and object.notify{All} both call Event::set.
5885 // That is, we treat thread.interrupt as a special case of notification.
5886 // The underlying Solaris implementation, cond_timedwait, admits
5887 // spurious/premature wakeups, but the JLS/JVM spec prevents the
5888 // JVM from making those visible to Java code. As such, we must
5889 // filter out spurious wakeups. We assume all ETIME returns are valid.
5890 //
5891 // TODO: properly differentiate simultaneous notify+interrupt.
5892 // In that case, we should propagate the notify to another waiter.
5893
5894 while (_Event < 0) {
5895 status = os::Linux::safe_cond_timedwait(_cond, _mutex, &abst);
5896 if (status != 0 && WorkAroundNPTLTimedWaitHang) {
5897 pthread_cond_destroy (_cond);
5898 pthread_cond_init (_cond, os::Linux::condAttr()) ;
5899 }
5900 assert_status(status == 0 || status == EINTR ||
5901 status == ETIME || status == ETIMEDOUT,
5902 status, "cond_timedwait");
5903 if (!FilterSpuriousWakeups) break ; // previous semantics
5904 if (status == ETIME || status == ETIMEDOUT) break ;
5905 // We consume and ignore EINTR and spurious wakeups.
5906 }
5907 --_nParked ;
5908 if (_Event >= 0) {
5909 ret = OS_OK;
5910 }
5911 _Event = 0 ;
5912 status = pthread_mutex_unlock(_mutex);
5913 assert_status(status == 0, status, "mutex_unlock");
5914 assert (_nParked == 0, "invariant") ;
5915 // Paranoia to ensure our locked and lock-free paths interact
5916 // correctly with each other.
5917 OrderAccess::fence();
5918 return ret;
5919 }
5920
5921 void os::PlatformEvent::unpark() {
5922 // Transitions for _Event:
5923 // 0 :=> 1
5924 // 1 :=> 1
5925 // -1 :=> either 0 or 1; must signal target thread
5926 // That is, we can safely transition _Event from -1 to either
5927 // 0 or 1. Forcing 1 is slightly more efficient for back-to-back
5928 // unpark() calls.
5929 // See also: "Semaphores in Plan 9" by Mullender & Cox
5930 //
5931 // Note: Forcing a transition from "-1" to "1" on an unpark() means
5932 // that it will take two back-to-back park() calls for the owning
5933 // thread to block. This has the benefit of forcing a spurious return
5934 // from the first park() call after an unpark() call which will help
5935 // shake out uses of park() and unpark() without condition variables.
5936
5937 if (Atomic::xchg(1, &_Event) >= 0) return;
5938
5939 // Wait for the thread associated with the event to vacate
5940 int status = pthread_mutex_lock(_mutex);
5941 assert_status(status == 0, status, "mutex_lock");
5942 int AnyWaiters = _nParked;
5943 assert(AnyWaiters == 0 || AnyWaiters == 1, "invariant");
5944 if (AnyWaiters != 0 && WorkAroundNPTLTimedWaitHang) {
5945 AnyWaiters = 0;
5946 pthread_cond_signal(_cond);
5947 }
5948 status = pthread_mutex_unlock(_mutex);
5949 assert_status(status == 0, status, "mutex_unlock");
5950 if (AnyWaiters != 0) {
5951 status = pthread_cond_signal(_cond);
5952 assert_status(status == 0, status, "cond_signal");
5953 }
5954
5955 // Note that we signal() _after dropping the lock for "immortal" Events.
5956 // This is safe and avoids a common class of futile wakeups. In rare
5957 // circumstances this can cause a thread to return prematurely from
5958 // cond_{timed}wait() but the spurious wakeup is benign and the victim will
5959 // simply re-test the condition and re-park itself.
5960 }
5961
5962
5963 // JSR166
5964 // -------------------------------------------------------
5965
5966 /*
5967 * The solaris and linux implementations of park/unpark are fairly
5968 * conservative for now, but can be improved. They currently use a
5969 * mutex/condvar pair, plus a a count.
5970 * Park decrements count if > 0, else does a condvar wait. Unpark
5971 * sets count to 1 and signals condvar. Only one thread ever waits
5972 * on the condvar. Contention seen when trying to park implies that someone
5973 * is unparking you, so don't wait. And spurious returns are fine, so there
5974 * is no need to track notifications.
5975 */
5976
5977 /*
5978 * This code is common to linux and solaris and will be moved to a
5979 * common place in dolphin.
5980 *
5981 * The passed in time value is either a relative time in nanoseconds
5982 * or an absolute time in milliseconds. Either way it has to be unpacked
5983 * into suitable seconds and nanoseconds components and stored in the
5984 * given timespec structure.
5985 * Given time is a 64-bit value and the time_t used in the timespec is only
5986 * a signed-32-bit value (except on 64-bit Linux) we have to watch for
5987 * overflow if times way in the future are given. Further on Solaris versions
5988 * prior to 10 there is a restriction (see cond_timedwait) that the specified
5989 * number of seconds, in abstime, is less than current_time + 100,000,000.
5990 * As it will be 28 years before "now + 100000000" will overflow we can
5991 * ignore overflow and just impose a hard-limit on seconds using the value
5992 * of "now + 100,000,000". This places a limit on the timeout of about 3.17
5993 * years from "now".
5994 */
5995
5996 static void unpackTime(timespec* absTime, bool isAbsolute, jlong time) {
5997 assert (time > 0, "convertTime");
5998 time_t max_secs = 0;
5999
6000 if (!os::Linux::supports_monotonic_clock() || isAbsolute) {
6001 struct timeval now;
6002 int status = gettimeofday(&now, NULL);
6003 assert(status == 0, "gettimeofday");
6004
6005 max_secs = now.tv_sec + MAX_SECS;
6006
6007 if (isAbsolute) {
6008 jlong secs = time / 1000;
6009 if (secs > max_secs) {
6010 absTime->tv_sec = max_secs;
6011 } else {
6012 absTime->tv_sec = secs;
6013 }
6014 absTime->tv_nsec = (time % 1000) * NANOSECS_PER_MILLISEC;
6015 } else {
6016 jlong secs = time / NANOSECS_PER_SEC;
6017 if (secs >= MAX_SECS) {
6018 absTime->tv_sec = max_secs;
6019 absTime->tv_nsec = 0;
6020 } else {
6021 absTime->tv_sec = now.tv_sec + secs;
6022 absTime->tv_nsec = (time % NANOSECS_PER_SEC) + now.tv_usec*1000;
6023 if (absTime->tv_nsec >= NANOSECS_PER_SEC) {
6024 absTime->tv_nsec -= NANOSECS_PER_SEC;
6025 ++absTime->tv_sec; // note: this must be <= max_secs
6026 }
6027 }
6028 }
6029 } else {
6030 // must be relative using monotonic clock
6031 struct timespec now;
6032 int status = os::Linux::clock_gettime(CLOCK_MONOTONIC, &now);
6033 assert_status(status == 0, status, "clock_gettime");
6034 max_secs = now.tv_sec + MAX_SECS;
6035 jlong secs = time / NANOSECS_PER_SEC;
6036 if (secs >= MAX_SECS) {
6037 absTime->tv_sec = max_secs;
6038 absTime->tv_nsec = 0;
6039 } else {
6040 absTime->tv_sec = now.tv_sec + secs;
6041 absTime->tv_nsec = (time % NANOSECS_PER_SEC) + now.tv_nsec;
6042 if (absTime->tv_nsec >= NANOSECS_PER_SEC) {
6043 absTime->tv_nsec -= NANOSECS_PER_SEC;
6044 ++absTime->tv_sec; // note: this must be <= max_secs
6045 }
6046 }
6047 }
6048 assert(absTime->tv_sec >= 0, "tv_sec < 0");
6049 assert(absTime->tv_sec <= max_secs, "tv_sec > max_secs");
6050 assert(absTime->tv_nsec >= 0, "tv_nsec < 0");
6051 assert(absTime->tv_nsec < NANOSECS_PER_SEC, "tv_nsec >= nanos_per_sec");
6052 }
6053
6054 void Parker::park(bool isAbsolute, jlong time) {
6055 // Ideally we'd do something useful while spinning, such
6056 // as calling unpackTime().
6057
6058 // Optional fast-path check:
6059 // Return immediately if a permit is available.
6060 // We depend on Atomic::xchg() having full barrier semantics
6061 // since we are doing a lock-free update to _counter.
6062 if (Atomic::xchg(0, &_counter) > 0) return;
6063
6064 Thread* thread = Thread::current();
6065 assert(thread->is_Java_thread(), "Must be JavaThread");
6066 JavaThread *jt = (JavaThread *)thread;
6067
6068 // Optional optimization -- avoid state transitions if there's an interrupt pending.
6069 // Check interrupt before trying to wait
6070 if (Thread::is_interrupted(thread, false)) {
6071 return;
6072 }
6073
6074 // Next, demultiplex/decode time arguments
6075 timespec absTime;
6076 if (time < 0 || (isAbsolute && time == 0) ) { // don't wait at all
6077 return;
6078 }
6079 if (time > 0) {
6080 unpackTime(&absTime, isAbsolute, time);
6081 }
6082
6083
6084 // Enter safepoint region
6085 // Beware of deadlocks such as 6317397.
6086 // The per-thread Parker:: mutex is a classic leaf-lock.
6087 // In particular a thread must never block on the Threads_lock while
6088 // holding the Parker:: mutex. If safepoints are pending both the
6089 // the ThreadBlockInVM() CTOR and DTOR may grab Threads_lock.
6090 ThreadBlockInVM tbivm(jt);
6091
6092 // Don't wait if cannot get lock since interference arises from
6093 // unblocking. Also. check interrupt before trying wait
6094 if (Thread::is_interrupted(thread, false) || pthread_mutex_trylock(_mutex) != 0) {
6095 return;
6096 }
6097
6098 int status ;
6099 if (_counter > 0) { // no wait needed
6100 _counter = 0;
6101 status = pthread_mutex_unlock(_mutex);
6102 assert (status == 0, "invariant") ;
6103 // Paranoia to ensure our locked and lock-free paths interact
6104 // correctly with each other and Java-level accesses.
6105 OrderAccess::fence();
6106 return;
6107 }
6108
6109 #ifdef ASSERT
6110 // Don't catch signals while blocked; let the running threads have the signals.
6111 // (This allows a debugger to break into the running thread.)
6112 sigset_t oldsigs;
6113 sigset_t* allowdebug_blocked = os::Linux::allowdebug_blocked_signals();
6114 pthread_sigmask(SIG_BLOCK, allowdebug_blocked, &oldsigs);
6115 #endif
6116
6117 OSThreadWaitState osts(thread->osthread(), false /* not Object.wait() */);
6118 jt->set_suspend_equivalent();
6119 // cleared by handle_special_suspend_equivalent_condition() or java_suspend_self()
6120
6121 assert(_cur_index == -1, "invariant");
6122 if (time == 0) {
6123 _cur_index = REL_INDEX; // arbitrary choice when not timed
6124 status = pthread_cond_wait (&_cond[_cur_index], _mutex) ;
6125 } else {
6126 _cur_index = isAbsolute ? ABS_INDEX : REL_INDEX;
6127 status = os::Linux::safe_cond_timedwait (&_cond[_cur_index], _mutex, &absTime) ;
6128 if (status != 0 && WorkAroundNPTLTimedWaitHang) {
6129 pthread_cond_destroy (&_cond[_cur_index]) ;
6130 pthread_cond_init (&_cond[_cur_index], isAbsolute ? NULL : os::Linux::condAttr());
6131 }
6132 }
6133 _cur_index = -1;
6134 assert_status(status == 0 || status == EINTR ||
6135 status == ETIME || status == ETIMEDOUT,
6136 status, "cond_timedwait");
6137
6138 #ifdef ASSERT
6139 pthread_sigmask(SIG_SETMASK, &oldsigs, NULL);
6140 #endif
6141
6142 _counter = 0 ;
6143 status = pthread_mutex_unlock(_mutex) ;
6144 assert_status(status == 0, status, "invariant") ;
6145 // Paranoia to ensure our locked and lock-free paths interact
6146 // correctly with each other and Java-level accesses.
6147 OrderAccess::fence();
6148
6149 // If externally suspended while waiting, re-suspend
6150 if (jt->handle_special_suspend_equivalent_condition()) {
6151 jt->java_suspend_self();
6152 }
6153 }
6154
6155 void Parker::unpark() {
6156 int s, status ;
6157 status = pthread_mutex_lock(_mutex);
6158 assert (status == 0, "invariant") ;
6159 s = _counter;
6160 _counter = 1;
6161 if (s < 1) {
6162 // thread might be parked
6163 if (_cur_index != -1) {
6164 // thread is definitely parked
6165 if (WorkAroundNPTLTimedWaitHang) {
6166 status = pthread_cond_signal (&_cond[_cur_index]);
6167 assert (status == 0, "invariant");
6168 status = pthread_mutex_unlock(_mutex);
6169 assert (status == 0, "invariant");
6170 } else {
6171 // must capture correct index before unlocking
6172 int index = _cur_index;
6173 status = pthread_mutex_unlock(_mutex);
6174 assert (status == 0, "invariant");
6175 status = pthread_cond_signal (&_cond[index]);
6176 assert (status == 0, "invariant");
6177 }
6178 } else {
6179 pthread_mutex_unlock(_mutex);
6180 assert (status == 0, "invariant") ;
6181 }
6182 } else {
6183 pthread_mutex_unlock(_mutex);
6184 assert (status == 0, "invariant") ;
6185 }
6186 }
6187
6188
6189 extern char** environ;
6190
6191 // Run the specified command in a separate process. Return its exit value,
6192 // or -1 on failure (e.g. can't fork a new process).
6193 // Unlike system(), this function can be called from signal handler. It
6194 // doesn't block SIGINT et al.
6195 int os::fork_and_exec(char* cmd) {
6196 const char * argv[4] = {"sh", "-c", cmd, NULL};
6197
6198 pid_t pid = fork();
6199
6200 if (pid < 0) {
6201 // fork failed
6202 return -1;
6203
6204 } else if (pid == 0) {
6205 // child process
6206
6207 execve("/bin/sh", (char* const*)argv, environ);
6208
6209 // execve failed
6210 _exit(-1);
6211
6212 } else {
6213 // copied from J2SE ..._waitForProcessExit() in UNIXProcess_md.c; we don't
6214 // care about the actual exit code, for now.
6215
6216 int status;
6217
6218 // Wait for the child process to exit. This returns immediately if
6219 // the child has already exited. */
6220 while (waitpid(pid, &status, 0) < 0) {
6221 switch (errno) {
6222 case ECHILD: return 0;
6223 case EINTR: break;
6224 default: return -1;
6225 }
6226 }
6227
6228 if (WIFEXITED(status)) {
6229 // The child exited normally; get its exit code.
6230 return WEXITSTATUS(status);
6231 } else if (WIFSIGNALED(status)) {
6232 // The child exited because of a signal
6233 // The best value to return is 0x80 + signal number,
6234 // because that is what all Unix shells do, and because
6235 // it allows callers to distinguish between process exit and
6236 // process death by signal.
6237 return 0x80 + WTERMSIG(status);
6238 } else {
6239 // Unknown exit code; pass it through
6240 return status;
6241 }
6242 }
6243 }
6244
6245 // is_headless_jre()
6246 //
6247 // Test for the existence of xawt/libmawt.so or libawt_xawt.so
6248 // in order to report if we are running in a headless jre
6249 //
6250 // Since JDK8 xawt/libmawt.so was moved into the same directory
6251 // as libawt.so, and renamed libawt_xawt.so
6252 //
6253 bool os::is_headless_jre() {
6254 struct stat statbuf;
6255 char buf[MAXPATHLEN];
6256 char libmawtpath[MAXPATHLEN];
6257 const char *xawtstr = "/xawt/libmawt.so";
6258 const char *new_xawtstr = "/libawt_xawt.so";
6259 char *p;
6260
6261 // Get path to libjvm.so
6262 os::jvm_path(buf, sizeof(buf));
6263
6264 // Get rid of libjvm.so
6265 p = strrchr(buf, '/');
6266 if (p == NULL) return false;
6267 else *p = '\0';
6268
6269 // Get rid of client or server
6270 p = strrchr(buf, '/');
6271 if (p == NULL) return false;
6272 else *p = '\0';
6273
6274 // check xawt/libmawt.so
6275 strcpy(libmawtpath, buf);
6276 strcat(libmawtpath, xawtstr);
6277 if (::stat(libmawtpath, &statbuf) == 0) return false;
6278
6279 // check libawt_xawt.so
6280 strcpy(libmawtpath, buf);
6281 strcat(libmawtpath, new_xawtstr);
6282 if (::stat(libmawtpath, &statbuf) == 0) return false;
6283
6284 return true;
6285 }
6286
6287 // Get the default path to the core file
6288 // Returns the length of the string
6289 int os::get_core_path(char* buffer, size_t bufferSize) {
6290 const char* p = get_current_directory(buffer, bufferSize);
6291
6292 if (p == NULL) {
6293 assert(p != NULL, "failed to get current directory");
6294 return 0;
6295 }
6296
6297 return strlen(buffer);
6298 }
6299
6300 /////////////// Unit tests ///////////////
6301
6302 #ifndef PRODUCT
6303
6304 #define test_log(...) \
6305 do {\
6306 if (VerboseInternalVMTests) { \
6307 tty->print_cr(__VA_ARGS__); \
6308 tty->flush(); \
6309 }\
6310 } while (false)
6311
6312 class TestReserveMemorySpecial : AllStatic {
6313 public:
6314 static void small_page_write(void* addr, size_t size) {
6315 size_t page_size = os::vm_page_size();
6316
6317 char* end = (char*)addr + size;
6318 for (char* p = (char*)addr; p < end; p += page_size) {
6319 *p = 1;
6320 }
6321 }
6322
6323 static void test_reserve_memory_special_huge_tlbfs_only(size_t size) {
6324 if (!UseHugeTLBFS) {
6325 return;
6326 }
6327
6328 test_log("test_reserve_memory_special_huge_tlbfs_only(" SIZE_FORMAT ")", size);
6329
6330 char* addr = os::Linux::reserve_memory_special_huge_tlbfs_only(size, NULL, false);
6331
6332 if (addr != NULL) {
6333 small_page_write(addr, size);
6334
6335 os::Linux::release_memory_special_huge_tlbfs(addr, size);
6336 }
6337 }
6338
6339 static void test_reserve_memory_special_huge_tlbfs_only() {
6340 if (!UseHugeTLBFS) {
6341 return;
6342 }
6343
6344 size_t lp = os::large_page_size();
6345
6346 for (size_t size = lp; size <= lp * 10; size += lp) {
6347 test_reserve_memory_special_huge_tlbfs_only(size);
6348 }
6349 }
6350
6351 static void test_reserve_memory_special_huge_tlbfs_mixed() {
6352 size_t lp = os::large_page_size();
6353 size_t ag = os::vm_allocation_granularity();
6354
6355 // sizes to test
6356 const size_t sizes[] = {
6357 lp, lp + ag, lp + lp / 2, lp * 2,
6358 lp * 2 + ag, lp * 2 - ag, lp * 2 + lp / 2,
6359 lp * 10, lp * 10 + lp / 2
6360 };
6361 const int num_sizes = sizeof(sizes) / sizeof(size_t);
6362
6363 // For each size/alignment combination, we test three scenarios:
6364 // 1) with req_addr == NULL
6365 // 2) with a non-null req_addr at which we expect to successfully allocate
6366 // 3) with a non-null req_addr which contains a pre-existing mapping, at which we
6367 // expect the allocation to either fail or to ignore req_addr
6368
6369 // Pre-allocate two areas; they shall be as large as the largest allocation
6370 // and aligned to the largest alignment we will be testing.
6371 const size_t mapping_size = sizes[num_sizes - 1] * 2;
6372 char* const mapping1 = (char*) ::mmap(NULL, mapping_size,
6373 PROT_NONE, MAP_PRIVATE|MAP_ANONYMOUS|MAP_NORESERVE,
6374 -1, 0);
6375 assert(mapping1 != MAP_FAILED, "should work");
6376
6377 char* const mapping2 = (char*) ::mmap(NULL, mapping_size,
6378 PROT_NONE, MAP_PRIVATE|MAP_ANONYMOUS|MAP_NORESERVE,
6379 -1, 0);
6380 assert(mapping2 != MAP_FAILED, "should work");
6381
6382 // Unmap the first mapping, but leave the second mapping intact: the first
6383 // mapping will serve as a value for a "good" req_addr (case 2). The second
6384 // mapping, still intact, as "bad" req_addr (case 3).
6385 ::munmap(mapping1, mapping_size);
6386
6387 // Case 1
6388 test_log("%s, req_addr NULL:", __FUNCTION__);
6389 test_log("size align result");
6390
6391 for (int i = 0; i < num_sizes; i++) {
6392 const size_t size = sizes[i];
6393 for (size_t alignment = ag; is_size_aligned(size, alignment); alignment *= 2) {
6394 char* p = os::Linux::reserve_memory_special_huge_tlbfs_mixed(size, alignment, NULL, false);
6395 test_log(SIZE_FORMAT_HEX " " SIZE_FORMAT_HEX " -> " PTR_FORMAT " %s",
6396 size, alignment, p, (p != NULL ? "" : "(failed)"));
6397 if (p != NULL) {
6398 assert(is_ptr_aligned(p, alignment), "must be");
6399 small_page_write(p, size);
6400 os::Linux::release_memory_special_huge_tlbfs(p, size);
6401 }
6402 }
6403 }
6404
6405 // Case 2
6406 test_log("%s, req_addr non-NULL:", __FUNCTION__);
6407 test_log("size align req_addr result");
6408
6409 for (int i = 0; i < num_sizes; i++) {
6410 const size_t size = sizes[i];
6411 for (size_t alignment = ag; is_size_aligned(size, alignment); alignment *= 2) {
6412 char* const req_addr = (char*) align_ptr_up(mapping1, alignment);
6413 char* p = os::Linux::reserve_memory_special_huge_tlbfs_mixed(size, alignment, req_addr, false);
6414 test_log(SIZE_FORMAT_HEX " " SIZE_FORMAT_HEX " " PTR_FORMAT " -> " PTR_FORMAT " %s",
6415 size, alignment, req_addr, p,
6416 ((p != NULL ? (p == req_addr ? "(exact match)" : "") : "(failed)")));
6417 if (p != NULL) {
6418 assert(p == req_addr, "must be");
6419 small_page_write(p, size);
6420 os::Linux::release_memory_special_huge_tlbfs(p, size);
6421 }
6422 }
6423 }
6424
6425 // Case 3
6426 test_log("%s, req_addr non-NULL with preexisting mapping:", __FUNCTION__);
6427 test_log("size align req_addr result");
6428
6429 for (int i = 0; i < num_sizes; i++) {
6430 const size_t size = sizes[i];
6431 for (size_t alignment = ag; is_size_aligned(size, alignment); alignment *= 2) {
6432 char* const req_addr = (char*) align_ptr_up(mapping2, alignment);
6433 char* p = os::Linux::reserve_memory_special_huge_tlbfs_mixed(size, alignment, req_addr, false);
6434 test_log(SIZE_FORMAT_HEX " " SIZE_FORMAT_HEX " " PTR_FORMAT " -> " PTR_FORMAT " %s",
6435 size, alignment, req_addr, p,
6436 ((p != NULL ? "" : "(failed)")));
6437 // as the area around req_addr contains already existing mappings, the API should always
6438 // return NULL (as per contract, it cannot return another address)
6439 assert(p == NULL, "must be");
6440 }
6441 }
6442
6443 ::munmap(mapping2, mapping_size);
6444
6445 }
6446
6447 static void test_reserve_memory_special_huge_tlbfs() {
6448 if (!UseHugeTLBFS) {
6449 return;
6450 }
6451
6452 test_reserve_memory_special_huge_tlbfs_only();
6453 test_reserve_memory_special_huge_tlbfs_mixed();
6454 }
6455
6456 static void test_reserve_memory_special_shm(size_t size, size_t alignment) {
6457 if (!UseSHM) {
6458 return;
6459 }
6460
6461 test_log("test_reserve_memory_special_shm(" SIZE_FORMAT ", " SIZE_FORMAT ")", size, alignment);
6462
6463 char* addr = os::Linux::reserve_memory_special_shm(size, alignment, NULL, false);
6464
6465 if (addr != NULL) {
6466 assert(is_ptr_aligned(addr, alignment), "Check");
6467 assert(is_ptr_aligned(addr, os::large_page_size()), "Check");
6468
6469 small_page_write(addr, size);
6470
6471 os::Linux::release_memory_special_shm(addr, size);
6472 }
6473 }
6474
6475 static void test_reserve_memory_special_shm() {
6476 size_t lp = os::large_page_size();
6477 size_t ag = os::vm_allocation_granularity();
6478
6479 for (size_t size = ag; size < lp * 3; size += ag) {
6480 for (size_t alignment = ag; is_size_aligned(size, alignment); alignment *= 2) {
6481 test_reserve_memory_special_shm(size, alignment);
6482 }
6483 }
6484 }
6485
6486 static void test() {
6487 test_reserve_memory_special_huge_tlbfs();
6488 test_reserve_memory_special_shm();
6489 }
6490 };
6491
6492 void TestReserveMemorySpecial_test() {
6493 TestReserveMemorySpecial::test();
6494 }
6495
6496 #endif