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