src/runtime/linux-os.c

   1 /*
   2  * the Linux incarnation of OS-dependent routines.  See also
   3  * $(sbcl_arch)-linux-os.c
   4  *
   5  * This file (along with os.h) exports an OS-independent interface to
   6  * the operating system VM facilities. Surprise surprise, this
   7  * interface looks a lot like the Mach interface (but simpler in some
   8  * places). For some operating systems, a subset of these functions
   9  * will have to be emulated.
  10  */
  11
  12 /*
  13  * This software is part of the SBCL system. See the README file for
  14  * more information.
  15  *
  16  * This software is derived from the CMU CL system, which was
  17  * written at Carnegie Mellon University and released into the
  18  * public domain. The software is in the public domain and is
  19  * provided with absolutely no warranty. See the COPYING and CREDITS
  20  * files for more information.
  21  */
  22
  23 #include <stdio.h>
  24 #include <sys/param.h>
  25 #include <sys/file.h>
  26 #include "sbcl.h"
  27 #include "./signal.h"
  28 #include "os.h"
  29 #include "arch.h"
  30 #include "globals.h"
  31 #include "interrupt.h"
  32 #include "interr.h"
  33 #include "lispregs.h"
  34 #include "runtime.h"
  35 #include "genesis/static-symbols.h"
  36 #include "genesis/fdefn.h"
  37
  38 #include <sys/socket.h>
  39 #include <sys/utsname.h>
  40 #include <errno.h>
  41
  42 #include <sys/types.h>
  43 #include <signal.h>
  44 /* #include <sys/sysinfo.h> */
  45 #include <sys/time.h>
  46 #include <sys/stat.h>
  47 #include <unistd.h>
  48 #include <linux/version.h>
  49
  50 #include "validate.h"
  51 #include "thread.h"
  52 #include "gc.h"
  53 #if defined LISP_FEATURE_GENCGC
  54 #include "gencgc-internal.h"
  55 #else
  56 #include "cheneygc-internal.h"
  57 #endif
  58
  59 #ifdef LISP_FEATURE_X86
  60 /* Prototype for personality(2). Done inline here since the header file
  61  * for this isn't available on old versions of glibc. */
  62 int personality (unsigned long);
  63 #endif
  64
  65 size_t os_vm_page_size;
  66
  67 #if defined(LISP_FEATURE_SB_THREAD) && !defined(LISP_FEATURE_SB_LUTEX) && !defined(LISP_FEATURE_SB_PTHREAD_FUTEX)
  68 #include <sys/syscall.h>
  69 #include <unistd.h>
  70 #include <errno.h>
  71
  72 /* values taken from the kernel's linux/futex.h.  This header file
  73    doesn't exist in userspace, which is our excuse for not grovelling
  74    them automatically */
  75 #define FUTEX_WAIT (0)
  76 #define FUTEX_WAKE (1)
  77 #define FUTEX_FD (2)
  78 #define FUTEX_REQUEUE (3)
  79
  80 #define sys_futex sbcl_sys_futex
  81 static inline int sys_futex (void *futex, int op, int val, struct timespec *rel)
  82 {
  83     return syscall (SYS_futex, futex, op, val, rel);
  84 }
  85
  86 int
  87 futex_wait(int *lock_word, int oldval, long sec, unsigned long usec)
  88 {
  89   struct timespec timeout;
  90   int t;
  91
  92  again:
  93   if (sec<0) {
  94     t = sys_futex(lock_word,FUTEX_WAIT,oldval, 0);
  95   }
  96   else {
  97     timeout.tv_sec = sec;
  98     timeout.tv_nsec = usec * 1000;
  99     t = sys_futex(lock_word,FUTEX_WAIT,oldval, &timeout);
 100   }
 101   if (t==0)
 102       return 0;
 103   else if (errno==ETIMEDOUT)
 104       return 1;
 105   else if (errno==EINTR)
 106       /* spurious wakeup from interrupt */
 107       goto again;
 108   else
 109       /* EWOULDBLOCK and others, need to check the lock */
 110       return -1;
 111 }
 112
 113 int
 114 futex_wake(int *lock_word, int n)
 115 {
 116     return sys_futex(lock_word,FUTEX_WAKE,n,0);
 117 }
 118 #endif
 119
 120 \f
 121 int linux_sparc_siginfo_bug = 0;
 122
 123 /* This variable was in real use for a few months, basically for
 124  * storing autodetected information about whether the Linux
 125  * installation was recent enough to support SBCL threads, and make
 126  * some run-time decisions based on that. But this turned out to be
 127  * unstable, so now we just flat-out refuse to start on the old installations
 128  * when thread support has been compiled in.
 129  *
 130  * Unfortunately, in the meanwhile Slime started depending on this
 131  * variable for deciding which communication style to use. So even
 132  * though this variable looks unused, it shouldn't be deleted until
 133  * it's no longer used in the versions of Slime that people are
 134  * likely to download first. -- JES, 2006-06-07
 135  */
 136 int linux_no_threads_p = 0;
 137
 138 #ifdef LISP_FEATURE_SB_THREAD
 139 int
 140 isnptl (void)
 141 {
 142   size_t n = confstr (_CS_GNU_LIBPTHREAD_VERSION, NULL, 0);
 143   if (n > 0) {
 144       char *buf = alloca (n);
 145       confstr (_CS_GNU_LIBPTHREAD_VERSION, buf, n);
 146       if (strstr (buf, "NPTL")) {
 147           return 1;
 148       }
 149   }
 150   return 0;
 151 }
 152 #endif
 153
 154 void
 155 os_init(char *argv[], char *envp[])
 156 {
 157     /* Conduct various version checks: do we have enough mmap(), is
 158      * this a sparc running 2.2, can we do threads? */
 159 #if defined(LISP_FEATURE_SB_THREAD) && !defined(LISP_FEATURE_SB_LUTEX) && !defined(LISP_FEATURE_SB_PTHREAD_FUTEX)
 160     int *futex=0;
 161 #endif
 162     struct utsname name;
 163     int major_version;
 164     int minor_version;
 165     int patch_version;
 166     char *p;
 167     uname(&name);
 168     p=name.release;
 169     major_version = atoi(p);
 170     p=strchr(p,'.')+1;
 171     minor_version = atoi(p);
 172     p=strchr(p,'.')+1;
 173     patch_version = atoi(p);
 174     if (major_version<2) {
 175         lose("linux kernel version too old: major version=%d (can't run in version < 2.0.0)\n",
 176              major_version);
 177     }
 178     if (!(major_version>2 || minor_version >= 4)) {
 179 #ifdef LISP_FEATURE_SPARC
 180         FSHOW((stderr,"linux kernel %d.%d predates 2.4;\n enabling workarounds for SPARC kernel bugs in signal handling.\n", major_version,minor_version));
 181         linux_sparc_siginfo_bug = 1;
 182 #endif
 183     }
 184 #ifdef LISP_FEATURE_SB_THREAD
 185 #if !defined(LISP_FEATURE_SB_LUTEX) && !defined(LISP_FEATURE_SB_PTHREAD_FUTEX)
 186     futex_wait(futex,-1,-1,0);
 187     if(errno==ENOSYS) {
 188        lose("This version of SBCL is compiled with threading support, but your kernel\n"
 189             "is too old to support this. Please use a more recent kernel or\n"
 190             "a version of SBCL without threading support.\n");
 191     }
 192 #endif
 193     if(! isnptl()) {
 194        lose("This version of SBCL only works correctly with the NPTL threading\n"
 195             "library. Please use a newer glibc, use an older SBCL, or stop using\n"
 196             "LD_ASSUME_KERNEL\n");
 197     }
 198 #endif
 199
 200     /* Don't use getpagesize(), since it's not constant across Linux
 201      * kernel versions on some architectures (for example PPC). FIXME:
 202      * possibly the same should be done on other architectures too.
 203      */
 204     os_vm_page_size = BACKEND_PAGE_BYTES;
 205
 206     /* KLUDGE: Disable memory randomization on new Linux kernels
 207      * by setting a personality flag and re-executing. (We need
 208      * to re-execute, since the memory maps that can conflict with
 209      * the SBCL spaces have already been done at this point).
 210      *
 211      * Since randomization is currently implemented only on x86 kernels,
 212      * don't do this trick on other platforms.
 213      */
 214 #ifdef LISP_FEATURE_X86
 215     if ((major_version == 2
 216          /* Some old kernels will apparently lose unsupported personality flags
 217           * on exec() */
 218          && ((minor_version == 6 && patch_version >= 11)
 219              || (minor_version > 6)
 220              /* This is what RHEL 3 reports */
 221              || (minor_version == 4 && patch_version > 20)))
 222         || major_version >= 3)
 223     {
 224         int pers = personality(0xffffffffUL);
 225         /* 0x40000 aka. ADDR_NO_RANDOMIZE */
 226         if (!(pers & 0x40000)) {
 227             int retval = personality(pers | 0x40000);
 228             /* Allegedly some Linux kernels (the reported case was
 229              * "hardened Linux 2.6.7") won't set the new personality,
 230              * but nor will they return -1 for an error. So as a
 231              * workaround query the new personality...
 232              */
 233             int newpers = personality(0xffffffffUL);
 234             /* ... and don't re-execute if either the setting resulted
 235              * in an error or if the value didn't change. Otherwise
 236              * this might result in an infinite loop.
 237              */
 238             if (retval != -1 && newpers != pers) {
 239                 /* Use /proc/self/exe instead of trying to figure out
 240                  * the executable path from PATH and argv[0], since
 241                  * that's unreliable. We follow the symlink instead of
 242                  * executing the file directly in order to prevent top
 243                  * from displaying the name of the process as "exe". */
 244                 char runtime[PATH_MAX+1];
 245                 int i = readlink("/proc/self/exe", runtime, PATH_MAX);
 246                 if (i != -1) {
 247                     runtime[i] = '\0';
 248                     execve(runtime, argv, envp);
 249                 }
 250             }
 251             /* Either changing the personality or execve() failed. Either
 252              * way we might as well continue, and hope that the random
 253              * memory maps are ok this time around.
 254              */
 255             fprintf(stderr, "WARNING: Couldn't re-execute SBCL with the proper personality flags (maybe /proc isn't mounted?). Trying to continue anyway.\n");
 256         }
 257     }
 258     /* Use SSE detector.  Recent versions of Linux enable SSE support
 259      * on SSE capable CPUs.  */
 260     /* FIXME: Are there any old versions that does not support SSE?  */
 261     fast_bzero_pointer = fast_bzero_detect;
 262 #endif
 263 }
 264
 265
 266 #ifdef LISP_FEATURE_ALPHA
 267 /* The Alpha is a 64 bit CPU.  SBCL is a 32 bit application.  Due to all
 268  * the places that assume we can get a pointer into a fixnum with no
 269  * information loss, we have to make sure it allocates all its ram in the
 270  * 0-2Gb region.  */
 271
 272 static void * under_2gb_free_pointer=DYNAMIC_1_SPACE_END;
 273 #endif
 274
 275 os_vm_address_t
 276 os_validate(os_vm_address_t addr, os_vm_size_t len)
 277 {
 278     int flags =  MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE;
 279     os_vm_address_t actual;
 280
 281 #ifdef LISP_FEATURE_ALPHA
 282     if (!addr) {
 283         addr=under_2gb_free_pointer;
 284     }
 285 #endif
 286     actual = mmap(addr, len, OS_VM_PROT_ALL, flags, -1, 0);
 287     if (actual == MAP_FAILED) {
 288         perror("mmap");
 289         return 0;               /* caller should check this */
 290     }
 291
 292     if (addr && (addr!=actual)) {
 293         fprintf(stderr, "mmap: wanted %lu bytes at %p, actually mapped at %p\n",
 294                 (unsigned long) len, addr, actual);
 295         return 0;
 296     }
 297
 298 #ifdef LISP_FEATURE_ALPHA
 299
 300     len=(len+(os_vm_page_size-1))&(~(os_vm_page_size-1));
 301     under_2gb_free_pointer+=len;
 302 #endif
 303
 304     return actual;
 305 }
 306
 307 void
 308 os_invalidate(os_vm_address_t addr, os_vm_size_t len)
 309 {
 310     if (munmap(addr,len) == -1) {
 311         perror("munmap");
 312     }
 313 }
 314
 315 os_vm_address_t
 316 os_map(int fd, int offset, os_vm_address_t addr, os_vm_size_t len)
 317 {
 318     os_vm_address_t actual;
 319
 320     actual = mmap(addr, len, OS_VM_PROT_ALL, MAP_PRIVATE | MAP_FIXED,
 321                   fd, (off_t) offset);
 322     if (actual == MAP_FAILED || (addr && (addr != actual))) {
 323         perror("mmap");
 324         lose("unexpected mmap(..) failure\n");
 325     }
 326
 327     return actual;
 328 }
 329
 330 void
 331 os_protect(os_vm_address_t address, os_vm_size_t length, os_vm_prot_t prot)
 332 {
 333     if (mprotect(address, length, prot) == -1) {
 334         if (errno == ENOMEM) {
 335             lose("An mprotect call failed with ENOMEM. This probably means that the maximum amount\n"
 336                  "of separate memory mappings was exceeded. To fix the problem, either increase\n"
 337                  "the maximum with e.g. 'echo 262144 > /proc/sys/vm/max_map_count' or recompile\n"
 338                  "SBCL with a larger value for GENCGC-PAGE-SIZE in 'src/target/parms.lisp'.");
 339         } else {
 340             perror("mprotect");
 341         }
 342     }
 343 }
 344 \f
 345 boolean
 346 is_valid_lisp_addr(os_vm_address_t addr)
 347 {
 348     struct thread *th;
 349     size_t ad = (size_t) addr;
 350
 351     if ((READ_ONLY_SPACE_START <= ad && ad < READ_ONLY_SPACE_END)
 352         || (STATIC_SPACE_START <= ad && ad < STATIC_SPACE_END)
 353 #if defined LISP_FEATURE_GENCGC
 354         || (DYNAMIC_SPACE_START <= ad && ad < DYNAMIC_SPACE_END)
 355 #else
 356         || (DYNAMIC_0_SPACE_START <= ad && ad < DYNAMIC_0_SPACE_END)
 357         || (DYNAMIC_1_SPACE_START <= ad && ad < DYNAMIC_1_SPACE_END)
 358 #endif
 359         )
 360         return 1;
 361     for_each_thread(th) {
 362         if((size_t)(th->control_stack_start) <= ad
 363            && ad < (size_t)(th->control_stack_end))
 364             return 1;
 365         if((size_t)(th->binding_stack_start) <= ad
 366            && ad < (size_t)(th->binding_stack_start + BINDING_STACK_SIZE))
 367             return 1;
 368     }
 369     return 0;
 370 }
 371 \f
 372 /*
 373  * any OS-dependent special low-level handling for signals
 374  */
 375
 376 /*
 377  * The GC needs to be hooked into whatever signal is raised for
 378  * page fault on this OS.
 379  */
 380 static void
 381 sigsegv_handler(int signal, siginfo_t *info, void* void_context)
 382 {
 383     os_context_t *context = arch_os_get_context(&void_context);
 384     os_vm_address_t addr = arch_get_bad_addr(signal, info, context);
 385
 386 #ifdef LISP_FEATURE_ALPHA
 387     /* Alpha stuff: This is the end of a pseudo-atomic section during
 388        which a signal was received.  We must deal with the pending
 389        interrupt (see also interrupt.c, ../code/interrupt.lisp)
 390
 391        (how we got here: when interrupting, we set bit 63 in reg_ALLOC.
 392        At the end of the atomic section we tried to write to reg_ALLOC,
 393        got a SIGSEGV (there's nothing mapped there) so ended up here. */
 394     if (addr != NULL &&
 395         *os_context_register_addr(context, reg_ALLOC) & (1L<<63)) {
 396         *os_context_register_addr(context, reg_ALLOC) -= (1L<<63);
 397         interrupt_handle_pending(context);
 398         return;
 399     }
 400 #endif
 401
 402 #ifdef LISP_FEATURE_GENCGC
 403     if (!gencgc_handle_wp_violation(addr))
 404 #else
 405     if (!cheneygc_handle_wp_violation(context, addr))
 406 #endif
 407         if (!handle_guard_page_triggered(context, addr))
 408 #ifdef LISP_FEATURE_C_STACK_IS_CONTROL_STACK
 409             lisp_memory_fault_error(context, addr);
 410 #else
 411             interrupt_handle_now(signal, info, context);
 412 #endif
 413 }
 414
 415 void
 416 os_install_interrupt_handlers(void)
 417 {
 418     undoably_install_low_level_interrupt_handler(SIG_MEMORY_FAULT,
 419                                                  sigsegv_handler);
 420 #ifdef LISP_FEATURE_SB_THREAD
 421     undoably_install_low_level_interrupt_handler(SIG_INTERRUPT_THREAD,
 422                                                  interrupt_thread_handler);
 423     undoably_install_low_level_interrupt_handler(SIG_STOP_FOR_GC,
 424                                                  sig_stop_for_gc_handler);
 425 #endif
 426 }
 427
 428 char *
 429 os_get_runtime_executable_path()
 430 {
 431     char path[PATH_MAX + 1];
 432     int size;
 433
 434     size = readlink("/proc/self/exe", path, sizeof(path)-1);
 435     if (size < 0)
 436         return NULL;
 437     else
 438         path[size] = '\0';
 439
 440     return copied_string(path);
 441 }