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