Home | History | Annotate | Download | only in x86_64
      1 /* libunwind - a platform-independent unwind library
      2    Copyright (C) 2010, 2011 by FERMI NATIONAL ACCELERATOR LABORATORY
      3 
      4 This file is part of libunwind.
      5 
      6 Permission is hereby granted, free of charge, to any person obtaining
      7 a copy of this software and associated documentation files (the
      8 "Software"), to deal in the Software without restriction, including
      9 without limitation the rights to use, copy, modify, merge, publish,
     10 distribute, sublicense, and/or sell copies of the Software, and to
     11 permit persons to whom the Software is furnished to do so, subject to
     12 the following conditions:
     13 
     14 The above copyright notice and this permission notice shall be
     15 included in all copies or substantial portions of the Software.
     16 
     17 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
     18 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
     19 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
     20 NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
     21 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
     22 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
     23 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
     24 
     25 #include "unwind_i.h"
     26 #include "ucontext_i.h"
     27 #include <signal.h>
     28 #include <limits.h>
     29 
     30 #pragma weak pthread_once
     31 #pragma weak pthread_key_create
     32 #pragma weak pthread_getspecific
     33 #pragma weak pthread_setspecific
     34 
     35 /* Initial hash table size. Table expands by 2 bits (times four). */
     36 #define HASH_MIN_BITS 14
     37 
     38 typedef struct
     39 {
     40   unw_tdep_frame_t *frames;
     41   size_t log_size;
     42   size_t used;
     43   size_t dtor_count;  /* Counts how many times our destructor has already
     44 			 been called. */
     45 } unw_trace_cache_t;
     46 
     47 static const unw_tdep_frame_t empty_frame = { 0, UNW_X86_64_FRAME_OTHER, -1, -1, 0, -1, -1 };
     48 static define_lock (trace_init_lock);
     49 static pthread_once_t trace_cache_once = PTHREAD_ONCE_INIT;
     50 static sig_atomic_t trace_cache_once_happen;
     51 static pthread_key_t trace_cache_key;
     52 static struct mempool trace_cache_pool;
     53 static __thread  unw_trace_cache_t *tls_cache;
     54 static __thread  int tls_cache_destroyed;
     55 
     56 /* Free memory for a thread's trace cache. */
     57 static void
     58 trace_cache_free (void *arg)
     59 {
     60   unw_trace_cache_t *cache = arg;
     61   if (++cache->dtor_count < PTHREAD_DESTRUCTOR_ITERATIONS)
     62   {
     63     /* Not yet our turn to get destroyed. Re-install ourselves into the key. */
     64     pthread_setspecific(trace_cache_key, cache);
     65     Debug(5, "delayed freeing cache %p (%zx to go)\n", cache,
     66 	  PTHREAD_DESTRUCTOR_ITERATIONS - cache->dtor_count);
     67     return;
     68   }
     69   tls_cache_destroyed = 1;
     70   tls_cache = NULL;
     71   munmap (cache->frames, (1u << cache->log_size) * sizeof(unw_tdep_frame_t));
     72   mempool_free (&trace_cache_pool, cache);
     73   Debug(5, "freed cache %p\n", cache);
     74 }
     75 
     76 /* Initialise frame tracing for threaded use. */
     77 static void
     78 trace_cache_init_once (void)
     79 {
     80   pthread_key_create (&trace_cache_key, &trace_cache_free);
     81   mempool_init (&trace_cache_pool, sizeof (unw_trace_cache_t), 0);
     82   trace_cache_once_happen = 1;
     83 }
     84 
     85 static unw_tdep_frame_t *
     86 trace_cache_buckets (size_t n)
     87 {
     88   unw_tdep_frame_t *frames;
     89   size_t i;
     90 
     91   GET_MEMORY(frames, n * sizeof (unw_tdep_frame_t));
     92   if (likely(frames != NULL))
     93     for (i = 0; i < n; ++i)
     94       frames[i] = empty_frame;
     95 
     96   return frames;
     97 }
     98 
     99 /* Allocate and initialise hash table for frame cache lookups.
    100    Returns the cache initialised with (1u << HASH_LOW_BITS) hash
    101    buckets, or NULL if there was a memory allocation problem. */
    102 static unw_trace_cache_t *
    103 trace_cache_create (void)
    104 {
    105   unw_trace_cache_t *cache;
    106 
    107   if (tls_cache_destroyed)
    108   {
    109     /* The current thread is in the process of exiting. Don't recreate
    110        cache, as we wouldn't have another chance to free it. */
    111     Debug(5, "refusing to reallocate cache: "
    112 	     "thread-locals are being deallocated\n");
    113     return NULL;
    114   }
    115 
    116   if (! (cache = mempool_alloc(&trace_cache_pool)))
    117   {
    118     Debug(5, "failed to allocate cache\n");
    119     return NULL;
    120   }
    121 
    122   if (! (cache->frames = trace_cache_buckets(1u << HASH_MIN_BITS)))
    123   {
    124     Debug(5, "failed to allocate buckets\n");
    125     mempool_free(&trace_cache_pool, cache);
    126     return NULL;
    127   }
    128 
    129   cache->log_size = HASH_MIN_BITS;
    130   cache->used = 0;
    131   cache->dtor_count = 0;
    132   tls_cache_destroyed = 0;  /* Paranoia: should already be 0. */
    133   Debug(5, "allocated cache %p\n", cache);
    134   return cache;
    135 }
    136 
    137 /* Expand the hash table in the frame cache if possible. This always
    138    quadruples the hash size, and clears all previous frame entries. */
    139 static int
    140 trace_cache_expand (unw_trace_cache_t *cache)
    141 {
    142   size_t old_size = (1u << cache->log_size);
    143   size_t new_log_size = cache->log_size + 2;
    144   unw_tdep_frame_t *new_frames = trace_cache_buckets (1u << new_log_size);
    145 
    146   if (unlikely(! new_frames))
    147   {
    148     Debug(5, "failed to expand cache to 2^%lu buckets\n", new_log_size);
    149     return -UNW_ENOMEM;
    150   }
    151 
    152   Debug(5, "expanded cache from 2^%lu to 2^%lu buckets\n", cache->log_size, new_log_size);
    153   munmap(cache->frames, old_size * sizeof(unw_tdep_frame_t));
    154   cache->frames = new_frames;
    155   cache->log_size = new_log_size;
    156   cache->used = 0;
    157   return 0;
    158 }
    159 
    160 static unw_trace_cache_t *
    161 trace_cache_get_unthreaded (void)
    162 {
    163   unw_trace_cache_t *cache;
    164   intrmask_t saved_mask;
    165   static unw_trace_cache_t *global_cache = NULL;
    166   lock_acquire (&trace_init_lock, saved_mask);
    167   if (! global_cache)
    168   {
    169     mempool_init (&trace_cache_pool, sizeof (unw_trace_cache_t), 0);
    170     global_cache = trace_cache_create ();
    171   }
    172   cache = global_cache;
    173   lock_release (&trace_init_lock, saved_mask);
    174   Debug(5, "using cache %p\n", cache);
    175   return cache;
    176 }
    177 
    178 /* Get the frame cache for the current thread. Create it if there is none. */
    179 static unw_trace_cache_t *
    180 trace_cache_get (void)
    181 {
    182   unw_trace_cache_t *cache;
    183   if (likely (pthread_once != NULL))
    184   {
    185     pthread_once(&trace_cache_once, &trace_cache_init_once);
    186     if (!trace_cache_once_happen)
    187     {
    188       return trace_cache_get_unthreaded();
    189     }
    190     if (! (cache = tls_cache))
    191     {
    192       cache = trace_cache_create();
    193       pthread_setspecific(trace_cache_key, cache);
    194       tls_cache = cache;
    195     }
    196     Debug(5, "using cache %p\n", cache);
    197     return cache;
    198   }
    199   else
    200   {
    201     return trace_cache_get_unthreaded();
    202   }
    203 }
    204 
    205 /* Initialise frame properties for address cache slot F at address
    206    RIP using current CFA, RBP and RSP values.  Modifies CURSOR to
    207    that location, performs one unw_step(), and fills F with what
    208    was discovered about the location.  Returns F.
    209 
    210    FIXME: This probably should tell DWARF handling to never evaluate
    211    or use registers other than RBP, RSP and RIP in case there is
    212    highly unusual unwind info which uses these creatively. */
    213 static unw_tdep_frame_t *
    214 trace_init_addr (unw_tdep_frame_t *f,
    215 		 unw_cursor_t *cursor,
    216 		 unw_word_t cfa,
    217 		 unw_word_t rip,
    218 		 unw_word_t rbp,
    219 		 unw_word_t rsp)
    220 {
    221   struct cursor *c = (struct cursor *) cursor;
    222   struct dwarf_cursor *d = &c->dwarf;
    223   int ret = -UNW_EINVAL;
    224 
    225   /* Initialise frame properties: unknown, not last. */
    226   f->virtual_address = rip;
    227   f->frame_type = UNW_X86_64_FRAME_OTHER;
    228   f->last_frame = 0;
    229   f->cfa_reg_rsp = -1;
    230   f->cfa_reg_offset = 0;
    231   f->rbp_cfa_offset = -1;
    232   f->rsp_cfa_offset = -1;
    233 
    234   /* Reinitialise cursor to this instruction - but undo next/prev RIP
    235      adjustment because unw_step will redo it - and force RIP, RBP
    236      RSP into register locations (=~ ucontext we keep), then set
    237      their desired values. Then perform the step. */
    238   d->ip = rip + d->use_prev_instr;
    239   d->cfa = cfa;
    240   d->loc[UNW_X86_64_RIP] = DWARF_REG_LOC (d, UNW_X86_64_RIP);
    241   d->loc[UNW_X86_64_RBP] = DWARF_REG_LOC (d, UNW_X86_64_RBP);
    242   d->loc[UNW_X86_64_RSP] = DWARF_REG_LOC (d, UNW_X86_64_RSP);
    243   c->frame_info = *f;
    244 
    245   if (likely(dwarf_put (d, d->loc[UNW_X86_64_RIP], rip) >= 0)
    246       && likely(dwarf_put (d, d->loc[UNW_X86_64_RBP], rbp) >= 0)
    247       && likely(dwarf_put (d, d->loc[UNW_X86_64_RSP], rsp) >= 0)
    248       && likely((ret = unw_step (cursor)) >= 0))
    249     *f = c->frame_info;
    250 
    251   /* If unw_step() stopped voluntarily, remember that, even if it
    252      otherwise could not determine anything useful.  This avoids
    253      failing trace if we hit frames without unwind info, which is
    254      common for the outermost frame (CRT stuff) on many systems.
    255      This avoids failing trace in very common circumstances; failing
    256      to unw_step() loop wouldn't produce any better result. */
    257   if (ret == 0)
    258     f->last_frame = -1;
    259 
    260   Debug (3, "frame va %lx type %d last %d cfa %s+%d rbp @ cfa%+d rsp @ cfa%+d\n",
    261 	 f->virtual_address, f->frame_type, f->last_frame,
    262 	 f->cfa_reg_rsp ? "rsp" : "rbp", f->cfa_reg_offset,
    263 	 f->rbp_cfa_offset, f->rsp_cfa_offset);
    264 
    265   return f;
    266 }
    267 
    268 /* Look up and if necessary fill in frame attributes for address RIP
    269    in CACHE using current CFA, RBP and RSP values.  Uses CURSOR to
    270    perform any unwind steps necessary to fill the cache.  Returns the
    271    frame cache slot which describes RIP. */
    272 static unw_tdep_frame_t *
    273 trace_lookup (unw_cursor_t *cursor,
    274 	      unw_trace_cache_t *cache,
    275 	      unw_word_t cfa,
    276 	      unw_word_t rip,
    277 	      unw_word_t rbp,
    278 	      unw_word_t rsp)
    279 {
    280   /* First look up for previously cached information using cache as
    281      linear probing hash table with probe step of 1.  Majority of
    282      lookups should be completed within few steps, but it is very
    283      important the hash table does not fill up, or performance falls
    284      off the cliff. */
    285   uint64_t i, addr;
    286   uint64_t cache_size = 1u << cache->log_size;
    287   uint64_t slot = ((rip * 0x9e3779b97f4a7c16) >> 43) & (cache_size-1);
    288   unw_tdep_frame_t *frame;
    289 
    290   for (i = 0; i < 16; ++i)
    291   {
    292     frame = &cache->frames[slot];
    293     addr = frame->virtual_address;
    294 
    295     /* Return if we found the address. */
    296     if (likely(addr == rip))
    297     {
    298       Debug (4, "found address after %ld steps\n", i);
    299       return frame;
    300     }
    301 
    302     /* If slot is empty, reuse it. */
    303     if (likely(! addr))
    304       break;
    305 
    306     /* Linear probe to next slot candidate, step = 1. */
    307     if (++slot >= cache_size)
    308       slot -= cache_size;
    309   }
    310 
    311   /* If we collided after 16 steps, or if the hash is more than half
    312      full, force the hash to expand. Fill the selected slot, whether
    313      it's free or collides. Note that hash expansion drops previous
    314      contents; further lookups will refill the hash. */
    315   Debug (4, "updating slot %lu after %ld steps, replacing 0x%lx\n", slot, i, addr);
    316   if (unlikely(addr || cache->used >= cache_size / 2))
    317   {
    318     if (unlikely(trace_cache_expand (cache) < 0))
    319       return NULL;
    320 
    321     cache_size = 1u << cache->log_size;
    322     slot = ((rip * 0x9e3779b97f4a7c16) >> 43) & (cache_size-1);
    323     frame = &cache->frames[slot];
    324     addr = frame->virtual_address;
    325   }
    326 
    327   if (! addr)
    328     ++cache->used;
    329 
    330   return trace_init_addr (frame, cursor, cfa, rip, rbp, rsp);
    331 }
    332 
    333 /* Fast stack backtrace for x86-64.
    334 
    335    This is used by backtrace() implementation to accelerate frequent
    336    queries for current stack, without any desire to unwind. It fills
    337    BUFFER with the call tree from CURSOR upwards for at most SIZE
    338    stack levels. The first frame, backtrace itself, is omitted. When
    339    called, SIZE should give the maximum number of entries that can be
    340    stored into BUFFER. Uses an internal thread-specific cache to
    341    accelerate queries.
    342 
    343    The caller should fall back to a unw_step() loop if this function
    344    fails by returning -UNW_ESTOPUNWIND, meaning the routine hit a
    345    stack frame that is too complex to be traced in the fast path.
    346 
    347    This function is tuned for clients which only need to walk the
    348    stack to get the call tree as fast as possible but without any
    349    other details, for example profilers sampling the stack thousands
    350    to millions of times per second.  The routine handles the most
    351    common x86-64 ABI stack layouts: CFA is RBP or RSP plus/minus
    352    constant offset, return address is at CFA-8, and RBP and RSP are
    353    either unchanged or saved on stack at constant offset from the CFA;
    354    the signal return frame; and frames without unwind info provided
    355    they are at the outermost (final) frame or can conservatively be
    356    assumed to be frame-pointer based.
    357 
    358    Any other stack layout will cause the routine to give up. There
    359    are only a handful of relatively rarely used functions which do
    360    not have a stack in the standard form: vfork, longjmp, setcontext
    361    and _dl_runtime_profile on common linux systems for example.
    362 
    363    On success BUFFER and *SIZE reflect the trace progress up to *SIZE
    364    stack levels or the outermost frame, which ever is less.  It may
    365    stop short of outermost frame if unw_step() loop would also do so,
    366    e.g. if there is no more unwind information; this is not reported
    367    as an error.
    368 
    369    The function returns a negative value for errors, -UNW_ESTOPUNWIND
    370    if tracing stopped because of an unusual frame unwind info.  The
    371    BUFFER and *SIZE reflect tracing progress up to the error frame.
    372 
    373    Callers of this function would normally look like this:
    374 
    375      unw_cursor_t     cur;
    376      unw_context_t    ctx;
    377      void             addrs[128];
    378      int              depth = 128;
    379      int              ret;
    380 
    381      unw_getcontext(&ctx);
    382      unw_init_local(&cur, &ctx);
    383      if ((ret = unw_tdep_trace(&cur, addrs, &depth)) < 0)
    384      {
    385        depth = 0;
    386        unw_getcontext(&ctx);
    387        unw_init_local(&cur, &ctx);
    388        while ((ret = unw_step(&cur)) > 0 && depth < 128)
    389        {
    390          unw_word_t ip;
    391          unw_get_reg(&cur, UNW_REG_IP, &ip);
    392          addresses[depth++] = (void *) ip;
    393        }
    394      }
    395 */
    396 HIDDEN int
    397 tdep_trace (unw_cursor_t *cursor, void **buffer, int *size)
    398 {
    399   struct cursor *c = (struct cursor *) cursor;
    400   struct dwarf_cursor *d = &c->dwarf;
    401   unw_trace_cache_t *cache;
    402   unw_word_t rbp, rsp, rip, cfa;
    403   int maxdepth = 0;
    404   int depth = 0;
    405   int ret;
    406 
    407   /* Check input parametres. */
    408   if (unlikely(! cursor || ! buffer || ! size || (maxdepth = *size) <= 0))
    409     return -UNW_EINVAL;
    410 
    411   Debug (1, "begin ip 0x%lx cfa 0x%lx\n", d->ip, d->cfa);
    412 
    413   /* Tell core dwarf routines to call back to us. */
    414   d->stash_frames = 1;
    415 
    416   /* Determine initial register values. These are direct access safe
    417      because we know they come from the initial machine context. */
    418   rip = d->ip;
    419   rsp = cfa = d->cfa;
    420   ACCESS_MEM_FAST(ret, 0, d, DWARF_GET_LOC(d->loc[UNW_X86_64_RBP]), rbp);
    421   assert(ret == 0);
    422 
    423   /* Get frame cache. */
    424   if (unlikely(! (cache = trace_cache_get())))
    425   {
    426     Debug (1, "returning %d, cannot get trace cache\n", -UNW_ENOMEM);
    427     *size = 0;
    428     d->stash_frames = 0;
    429     return -UNW_ENOMEM;
    430   }
    431 
    432   /* Trace the stack upwards, starting from current RIP.  Adjust
    433      the RIP address for previous/next instruction as the main
    434      unwinding logic would also do.  We undo this before calling
    435      back into unw_step(). */
    436   while (depth < maxdepth)
    437   {
    438     rip -= d->use_prev_instr;
    439     Debug (2, "depth %d cfa 0x%lx rip 0x%lx rsp 0x%lx rbp 0x%lx\n",
    440 	   depth, cfa, rip, rsp, rbp);
    441 
    442     /* See if we have this address cached.  If not, evaluate enough of
    443        the dwarf unwind information to fill the cache line data, or to
    444        decide this frame cannot be handled in fast trace mode.  We
    445        cache negative results too to prevent unnecessary dwarf parsing
    446        for common failures. */
    447     unw_tdep_frame_t *f = trace_lookup (cursor, cache, cfa, rip, rbp, rsp);
    448 
    449     /* If we don't have information for this frame, give up. */
    450     if (unlikely(! f))
    451     {
    452       ret = -UNW_ENOINFO;
    453       break;
    454     }
    455 
    456     Debug (3, "frame va %lx type %d last %d cfa %s+%d rbp @ cfa%+d rsp @ cfa%+d\n",
    457            f->virtual_address, f->frame_type, f->last_frame,
    458            f->cfa_reg_rsp ? "rsp" : "rbp", f->cfa_reg_offset,
    459            f->rbp_cfa_offset, f->rsp_cfa_offset);
    460 
    461     assert (f->virtual_address == rip);
    462 
    463     /* Stop if this was the last frame.  In particular don't evaluate
    464        new register values as it may not be safe - we don't normally
    465        run with full validation on, and do not want to - and there's
    466        enough bad unwind info floating around that we need to trust
    467        what unw_step() previously said, in potentially bogus frames. */
    468     if (f->last_frame)
    469       break;
    470 
    471     /* Evaluate CFA and registers for the next frame. */
    472     switch (f->frame_type)
    473     {
    474     case UNW_X86_64_FRAME_GUESSED:
    475       /* Fall thru to standard processing after forcing validation. */
    476       c->validate = 1;
    477 
    478     case UNW_X86_64_FRAME_STANDARD:
    479       /* Advance standard traceable frame. */
    480       cfa = (f->cfa_reg_rsp ? rsp : rbp) + f->cfa_reg_offset;
    481       ACCESS_MEM_FAST(ret, c->validate, d, cfa - 8, rip);
    482       if (likely(ret >= 0) && likely(f->rbp_cfa_offset != -1))
    483 	ACCESS_MEM_FAST(ret, c->validate, d, cfa + f->rbp_cfa_offset, rbp);
    484 
    485       /* Don't bother reading RSP from DWARF, CFA becomes new RSP. */
    486       rsp = cfa;
    487 
    488       /* Next frame needs to back up for unwind info lookup. */
    489       d->use_prev_instr = 1;
    490       break;
    491 
    492     case UNW_X86_64_FRAME_SIGRETURN:
    493       cfa = cfa + f->cfa_reg_offset; /* cfa now points to ucontext_t.  */
    494 
    495       ACCESS_MEM_FAST(ret, c->validate, d, cfa + UC_MCONTEXT_GREGS_RIP, rip);
    496       if (likely(ret >= 0))
    497         ACCESS_MEM_FAST(ret, c->validate, d, cfa + UC_MCONTEXT_GREGS_RBP, rbp);
    498       if (likely(ret >= 0))
    499         ACCESS_MEM_FAST(ret, c->validate, d, cfa + UC_MCONTEXT_GREGS_RSP, rsp);
    500 
    501       /* Resume stack at signal restoration point. The stack is not
    502          necessarily continuous here, especially with sigaltstack(). */
    503       cfa = rsp;
    504 
    505       /* Next frame should not back up. */
    506       d->use_prev_instr = 0;
    507       break;
    508 
    509     default:
    510       /* We cannot trace through this frame, give up and tell the
    511 	 caller we had to stop.  Data collected so far may still be
    512 	 useful to the caller, so let it know how far we got.  */
    513       ret = -UNW_ESTOPUNWIND;
    514       break;
    515     }
    516 
    517     Debug (4, "new cfa 0x%lx rip 0x%lx rsp 0x%lx rbp 0x%lx\n",
    518 	   cfa, rip, rsp, rbp);
    519 
    520     /* If we failed or ended up somewhere bogus, stop. */
    521     if (unlikely(ret < 0 || rip < 0x4000))
    522       break;
    523 
    524     /* Record this address in stack trace. We skipped the first address. */
    525     buffer[depth++] = (void *) (rip - d->use_prev_instr);
    526   }
    527 
    528 #if UNW_DEBUG
    529   Debug (1, "returning %d, depth %d\n", ret, depth);
    530 #endif
    531   *size = depth;
    532   return ret;
    533 }
    534