mc_leakcheck.c revision 18afe5d1c3dc9e4e16d3e10617d6afa4bf9fcd40
1
2/*--------------------------------------------------------------------*/
3/*--- The leak checker.                             mc_leakcheck.c ---*/
4/*--------------------------------------------------------------------*/
5
6/*
7   This file is part of MemCheck, a heavyweight Valgrind tool for
8   detecting memory errors.
9
10   Copyright (C) 2000-2009 Julian Seward
11      jseward@acm.org
12
13   This program is free software; you can redistribute it and/or
14   modify it under the terms of the GNU General Public License as
15   published by the Free Software Foundation; either version 2 of the
16   License, or (at your option) any later version.
17
18   This program is distributed in the hope that it will be useful, but
19   WITHOUT ANY WARRANTY; without even the implied warranty of
20   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
21   General Public License for more details.
22
23   You should have received a copy of the GNU General Public License
24   along with this program; if not, write to the Free Software
25   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
26   02111-1307, USA.
27
28   The GNU General Public License is contained in the file COPYING.
29*/
30
31#include "pub_tool_basics.h"
32#include "pub_tool_vki.h"
33#include "pub_tool_aspacehl.h"
34#include "pub_tool_aspacemgr.h"
35#include "pub_tool_execontext.h"
36#include "pub_tool_hashtable.h"
37#include "pub_tool_libcbase.h"
38#include "pub_tool_libcassert.h"
39#include "pub_tool_libcprint.h"
40#include "pub_tool_libcsignal.h"
41#include "pub_tool_machine.h"
42#include "pub_tool_mallocfree.h"
43#include "pub_tool_options.h"
44#include "pub_tool_oset.h"
45#include "pub_tool_signals.h"
46#include "pub_tool_tooliface.h"     // Needed for mc_include.h
47
48#include "mc_include.h"
49
50#include <setjmp.h>                 // For jmp_buf
51
52/*------------------------------------------------------------*/
53/*--- An overview of leak checking.                        ---*/
54/*------------------------------------------------------------*/
55
56// Leak-checking is a directed-graph traversal problem.  The graph has
57// two kinds of nodes:
58// - root-set nodes:
59//   - GP registers of all threads;
60//   - valid, aligned, pointer-sized data words in valid client memory,
61//     including stacks, but excluding words within client heap-allocated
62//     blocks (they are excluded so that later on we can differentiate
63//     between heap blocks that are indirectly leaked vs. directly leaked).
64// - heap-allocated blocks.  A block is a mempool chunk or a malloc chunk
65//   that doesn't contain a mempool chunk.  Nb: the terms "blocks" and
66//   "chunks" are used interchangeably below.
67//
68// There are two kinds of edges:
69// - start-pointers, i.e. pointers to the start of a block;
70// - interior-pointers, i.e. pointers to the interior of a block.
71//
72// We use "pointers" rather than "edges" below.
73//
74// Root set nodes only point to blocks.  Blocks only point to blocks;
75// a block can point to itself.
76//
77// The aim is to traverse the graph and determine the status of each block.
78//
79// There are 9 distinct cases.  See memcheck/docs/mc-manual.xml for details.
80// Presenting all nine categories to the user is probably too much.
81// Currently we do this:
82// - definitely lost:  case 3
83// - indirectly lost:  case 4, 9
84// - possibly lost:    cases 5..8
85// - still reachable:  cases 1, 2
86//
87// It's far from clear that this is the best possible categorisation;  it's
88// accreted over time without any central guiding principle.
89
90/*------------------------------------------------------------*/
91/*--- XXX: Thoughts for improvement.                       ---*/
92/*------------------------------------------------------------*/
93
94// From the user's point of view:
95// - If they aren't using interior-pointers, they just have to fix the
96//   directly lost blocks, and the indirectly lost ones will be fixed as
97//   part of that.  Any possibly lost blocks will just be due to random
98//   pointer garbage and can be ignored.
99//
100// - If they are using interior-pointers, the fact that they currently are not
101//   being told which ones might be directly lost vs. indirectly lost makes
102//   it hard to know where to begin.
103//
104// All this makes me wonder if new option is warranted:
105// --follow-interior-pointers.  By default it would be off, the leak checker
106// wouldn't follow interior-pointers and there would only be 3 categories:
107// R, DL, IL.
108//
109// If turned on, then it would show 7 categories (R, DL, IL, DR/DL, IR/IL,
110// IR/IL/DL, IL/DL).  That output is harder to understand but it's your own
111// damn fault for using interior-pointers...
112//
113// ----
114//
115// Also, why are two blank lines printed between each loss record?
116// [bug 197930]
117//
118// ----
119//
120// Also, --show-reachable is a bad name because it also turns on the showing
121// of indirectly leaked blocks(!)  It would be better named --show-all or
122// --show-all-heap-blocks, because that's the end result.
123//
124// ----
125//
126// Also, the VALGRIND_LEAK_CHECK and VALGRIND_QUICK_LEAK_CHECK aren't great
127// names.  VALGRIND_FULL_LEAK_CHECK and VALGRIND_SUMMARY_LEAK_CHECK would be
128// better.
129//
130// ----
131//
132// Also, VALGRIND_COUNT_LEAKS and VALGRIND_COUNT_LEAK_BLOCKS aren't great as
133// they combine direct leaks and indirect leaks into one.  New, more precise
134// ones (they'll need new names) would be good.  If more categories are
135// used, as per the --follow-interior-pointers option, they should be
136// updated accordingly.  And they should use a struct to return the values.
137//
138// ----
139//
140// Also, for this case:
141//
142//  (4)  p4      BBB ---> AAA
143//
144// BBB is definitely directly lost.  AAA is definitely indirectly lost.
145// Here's the relevant loss records printed for a full check (each block is
146// 16 bytes):
147//
148// ==20397== 16 bytes in 1 blocks are indirectly lost in loss record 9 of 15
149// ==20397==    at 0x4C2694E: malloc (vg_replace_malloc.c:177)
150// ==20397==    by 0x400521: mk (leak-cases.c:49)
151// ==20397==    by 0x400578: main (leak-cases.c:72)
152//
153// ==20397== 32 (16 direct, 16 indirect) bytes in 1 blocks are definitely
154// lost in loss record 14 of 15
155// ==20397==    at 0x4C2694E: malloc (vg_replace_malloc.c:177)
156// ==20397==    by 0x400521: mk (leak-cases.c:49)
157// ==20397==    by 0x400580: main (leak-cases.c:72)
158//
159// The first one is fine -- it describes AAA.
160//
161// The second one is for BBB.  It's correct in that 16 bytes in 1 block are
162// directly lost. It's also correct that 16 are indirectly lost as a result,
163// but it means that AAA is being counted twice in the loss records.  (It's
164// not, thankfully, counted twice in the summary counts).  Argh.
165//
166// This would be less confusing for the second one:
167//
168// ==20397== 16 bytes in 1 blocks are definitely lost in loss record 14
169// of 15 (and 16 bytes in 1 block are indirectly lost as a result;  they
170// are mentioned elsewhere (if --show-reachable=yes is given!))
171// ==20397==    at 0x4C2694E: malloc (vg_replace_malloc.c:177)
172// ==20397==    by 0x400521: mk (leak-cases.c:49)
173// ==20397==    by 0x400580: main (leak-cases.c:72)
174//
175// But ideally we'd present the loss record for the directly lost block and
176// then the resultant indirectly lost blocks and make it clear the
177// dependence.  Double argh.
178
179/*------------------------------------------------------------*/
180/*--- The actual algorithm.                                ---*/
181/*------------------------------------------------------------*/
182
183// - Find all the blocks (a.k.a. chunks) to check.  Mempool chunks require
184//   some special treatment because they can be within malloc'd blocks.
185// - Scan every word in the root set (GP registers and valid
186//   non-heap memory words).
187//   - First, we skip if it doesn't point to valid memory.
188//   - Then, we see if it points to the start or interior of a block.  If
189//     so, we push the block onto the mark stack and mark it as having been
190//     reached.
191// - Then, we process the mark stack, repeating the scanning for each block;
192//   this can push more blocks onto the mark stack.  We repeat until the
193//   mark stack is empty.  Each block is marked as definitely or possibly
194//   reachable, depending on whether interior-pointers were required to
195//   reach it.
196// - At this point we know for every block if it's reachable or not.
197// - We then push each unreached block onto the mark stack, using the block
198//   number as the "clique" number.
199// - We process the mark stack again, this time grouping blocks into cliques
200//   in order to facilitate the directly/indirectly lost categorisation.
201// - We group blocks by their ExeContexts and categorisation, and print them
202//   if --leak-check=full.  We also print summary numbers.
203//
204// A note on "cliques":
205// - A directly lost block is one with no pointers to it.  An indirectly
206//   lost block is one that is pointed to by a directly or indirectly lost
207//   block.
208// - Each directly lost block has zero or more indirectly lost blocks
209//   hanging off it.  All these blocks together form a "clique".  The
210//   directly lost block is called the "clique leader".  The clique number
211//   is the number (in lc_chunks[]) of the clique leader.
212// - Actually, a directly lost block may be pointed to if it's part of a
213//   cycle.  In that case, there may be more than one choice for the clique
214//   leader, and the choice is arbitrary.  Eg. if you have A-->B and B-->A
215//   either A or B could be the clique leader.
216// - Cliques cannot overlap, and will be truncated to avoid this.  Eg. if we
217//   have A-->C and B-->C, the two cliques will be {A,C} and {B}, or {A} and
218//   {B,C} (again the choice is arbitrary).  This is because we don't want
219//   to count a block as indirectly lost more than once.
220//
221// A note on 'is_prior_definite':
222// - This is a boolean used in various places that indicates if the chain
223//   up to the prior node (prior to the one being considered) is definite.
224// - In the clique == -1 case:
225//   - if True it means that the prior node is a root-set node, or that the
226//     prior node is a block which is reachable from the root-set via
227//     start-pointers.
228//   - if False it means that the prior node is a block that is only
229//     reachable from the root-set via a path including at least one
230//     interior-pointer.
231// - In the clique != -1 case, currently it's always True because we treat
232//   start-pointers and interior-pointers the same for direct/indirect leak
233//   checking.  If we added a PossibleIndirectLeak state then this would
234//   change.
235
236
237// Define to debug the memory-leak-detector.
238#define VG_DEBUG_LEAKCHECK 0
239#define VG_DEBUG_CLIQUE    0
240
241
242/*------------------------------------------------------------*/
243/*--- Getting the initial chunks, and searching them.      ---*/
244/*------------------------------------------------------------*/
245
246// Compare the MC_Chunks by 'data' (i.e. the address of the block).
247static Int compare_MC_Chunks(void* n1, void* n2)
248{
249   MC_Chunk* mc1 = *(MC_Chunk**)n1;
250   MC_Chunk* mc2 = *(MC_Chunk**)n2;
251   if (mc1->data < mc2->data) return -1;
252   if (mc1->data > mc2->data) return  1;
253   return 0;
254}
255
256#if VG_DEBUG_LEAKCHECK
257// Used to sanity-check the fast binary-search mechanism.
258static
259Int find_chunk_for_OLD ( Addr       ptr,
260                         MC_Chunk** chunks,
261                         Int        n_chunks )
262
263{
264   Int  i;
265   Addr a_lo, a_hi;
266   PROF_EVENT(70, "find_chunk_for_OLD");
267   for (i = 0; i < n_chunks; i++) {
268      PROF_EVENT(71, "find_chunk_for_OLD(loop)");
269      a_lo = chunks[i]->data;
270      a_hi = ((Addr)chunks[i]->data) + chunks[i]->szB;
271      if (a_lo <= ptr && ptr < a_hi)
272         return i;
273   }
274   return -1;
275}
276#endif
277
278// Find the i such that ptr points at or inside the block described by
279// chunks[i].  Return -1 if none found.  This assumes that chunks[]
280// has been sorted on the 'data' field.
281static
282Int find_chunk_for ( Addr       ptr,
283                     MC_Chunk** chunks,
284                     Int        n_chunks )
285{
286   Addr a_mid_lo, a_mid_hi;
287   Int lo, mid, hi, retVal;
288   // VG_(printf)("find chunk for %p = ", ptr);
289   retVal = -1;
290   lo = 0;
291   hi = n_chunks-1;
292   while (True) {
293      // Invariant: current unsearched space is from lo to hi, inclusive.
294      if (lo > hi) break; // not found
295
296      mid      = (lo + hi) / 2;
297      a_mid_lo = chunks[mid]->data;
298      a_mid_hi = chunks[mid]->data + chunks[mid]->szB;
299      // Extent of block 'mid' is [a_mid_lo .. a_mid_hi).
300      // Special-case zero-sized blocks - treat them as if they had
301      // size 1.  Not doing so causes them to not cover any address
302      // range at all and so will never be identified as the target of
303      // any pointer, which causes them to be incorrectly reported as
304      // definitely leaked.
305      if (chunks[mid]->szB == 0)
306         a_mid_hi++;
307
308      if (ptr < a_mid_lo) {
309         hi = mid-1;
310         continue;
311      }
312      if (ptr >= a_mid_hi) {
313         lo = mid+1;
314         continue;
315      }
316      tl_assert(ptr >= a_mid_lo && ptr < a_mid_hi);
317      retVal = mid;
318      break;
319   }
320
321#  if VG_DEBUG_LEAKCHECK
322   tl_assert(retVal == find_chunk_for_OLD ( ptr, chunks, n_chunks ));
323#  endif
324   // VG_(printf)("%d\n", retVal);
325   return retVal;
326}
327
328
329static MC_Chunk**
330find_active_chunks(UInt* pn_chunks)
331{
332   // Our goal is to construct a set of chunks that includes every
333   // mempool chunk, and every malloc region that *doesn't* contain a
334   // mempool chunk.
335   MC_Mempool *mp;
336   MC_Chunk **mallocs, **chunks, *mc;
337   UInt n_mallocs, n_chunks, m, s;
338   Bool *malloc_chunk_holds_a_pool_chunk;
339
340   // First we collect all the malloc chunks into an array and sort it.
341   // We do this because we want to query the chunks by interior
342   // pointers, requiring binary search.
343   mallocs = (MC_Chunk**) VG_(HT_to_array)( MC_(malloc_list), &n_mallocs );
344   if (n_mallocs == 0) {
345      tl_assert(mallocs == NULL);
346      *pn_chunks = 0;
347      return NULL;
348   }
349   VG_(ssort)(mallocs, n_mallocs, sizeof(VgHashNode*), compare_MC_Chunks);
350
351   // Then we build an array containing a Bool for each malloc chunk,
352   // indicating whether it contains any mempools.
353   malloc_chunk_holds_a_pool_chunk = VG_(calloc)( "mc.fas.1",
354                                                  n_mallocs, sizeof(Bool) );
355   n_chunks = n_mallocs;
356
357   // Then we loop over the mempool tables. For each chunk in each
358   // pool, we set the entry in the Bool array corresponding to the
359   // malloc chunk containing the mempool chunk.
360   VG_(HT_ResetIter)(MC_(mempool_list));
361   while ( (mp = VG_(HT_Next)(MC_(mempool_list))) ) {
362      VG_(HT_ResetIter)(mp->chunks);
363      while ( (mc = VG_(HT_Next)(mp->chunks)) ) {
364
365         // We'll need to record this chunk.
366         n_chunks++;
367
368         // Possibly invalidate the malloc holding the beginning of this chunk.
369         m = find_chunk_for(mc->data, mallocs, n_mallocs);
370         if (m != -1 && malloc_chunk_holds_a_pool_chunk[m] == False) {
371            tl_assert(n_chunks > 0);
372            n_chunks--;
373            malloc_chunk_holds_a_pool_chunk[m] = True;
374         }
375
376         // Possibly invalidate the malloc holding the end of this chunk.
377         if (mc->szB > 1) {
378            m = find_chunk_for(mc->data + (mc->szB - 1), mallocs, n_mallocs);
379            if (m != -1 && malloc_chunk_holds_a_pool_chunk[m] == False) {
380               tl_assert(n_chunks > 0);
381               n_chunks--;
382               malloc_chunk_holds_a_pool_chunk[m] = True;
383            }
384         }
385      }
386   }
387   tl_assert(n_chunks > 0);
388
389   // Create final chunk array.
390   chunks = VG_(malloc)("mc.fas.2", sizeof(VgHashNode*) * (n_chunks));
391   s = 0;
392
393   // Copy the mempool chunks and the non-marked malloc chunks into a
394   // combined array of chunks.
395   VG_(HT_ResetIter)(MC_(mempool_list));
396   while ( (mp = VG_(HT_Next)(MC_(mempool_list))) ) {
397      VG_(HT_ResetIter)(mp->chunks);
398      while ( (mc = VG_(HT_Next)(mp->chunks)) ) {
399         tl_assert(s < n_chunks);
400         chunks[s++] = mc;
401      }
402   }
403   for (m = 0; m < n_mallocs; ++m) {
404      if (!malloc_chunk_holds_a_pool_chunk[m]) {
405         tl_assert(s < n_chunks);
406         chunks[s++] = mallocs[m];
407      }
408   }
409   tl_assert(s == n_chunks);
410
411   // Free temporaries.
412   VG_(free)(mallocs);
413   VG_(free)(malloc_chunk_holds_a_pool_chunk);
414
415   *pn_chunks = n_chunks;
416
417   return chunks;
418}
419
420/*------------------------------------------------------------*/
421/*--- The leak detector proper.                            ---*/
422/*------------------------------------------------------------*/
423
424// Holds extra info about each block during leak checking.
425typedef
426   struct {
427      UInt  state:2;    // Reachedness.
428      SizeT indirect_szB : (sizeof(SizeT)*8)-2; // If Unreached, how many bytes
429                                                //   are unreachable from here.
430   }
431   LC_Extra;
432
433// An array holding pointers to every chunk we're checking.  Sorted by address.
434static MC_Chunk** lc_chunks;
435// How many chunks we're dealing with.
436static Int        lc_n_chunks;
437
438// This has the same number of entries as lc_chunks, and each entry
439// in lc_chunks corresponds with the entry here (ie. lc_chunks[i] and
440// lc_extras[i] describe the same block).
441static LC_Extra* lc_extras;
442
443// Records chunks that are currently being processed.  Each element in the
444// stack is an index into lc_chunks and lc_extras.  Its size is
445// 'lc_n_chunks' because in the worst case that's how many chunks could be
446// pushed onto it (actually I think the maximum is lc_n_chunks-1 but let's
447// be conservative).
448static Int* lc_markstack;
449// The index of the top element of the stack; -1 if the stack is empty, 0 if
450// the stack has one element, 1 if it has two, etc.
451static Int  lc_markstack_top;
452
453// Keeps track of how many bytes of memory we've scanned, for printing.
454// (Nb: We don't keep track of how many register bytes we've scanned.)
455static SizeT lc_scanned_szB;
456
457
458SizeT MC_(bytes_leaked)     = 0;
459SizeT MC_(bytes_indirect)   = 0;
460SizeT MC_(bytes_dubious)    = 0;
461SizeT MC_(bytes_reachable)  = 0;
462SizeT MC_(bytes_suppressed) = 0;
463
464SizeT MC_(blocks_leaked)     = 0;
465SizeT MC_(blocks_indirect)   = 0;
466SizeT MC_(blocks_dubious)    = 0;
467SizeT MC_(blocks_reachable)  = 0;
468SizeT MC_(blocks_suppressed) = 0;
469
470
471// Determines if a pointer is to a chunk.  Returns the chunk number et al
472// via call-by-reference.
473static Bool
474lc_is_a_chunk_ptr(Addr ptr, Int* pch_no, MC_Chunk** pch, LC_Extra** pex)
475{
476   Int ch_no;
477   MC_Chunk* ch;
478   LC_Extra* ex;
479
480   // Quick filter.
481   if (!VG_(am_is_valid_for_client)(ptr, 1, VKI_PROT_READ)) {
482      return False;
483   } else {
484      ch_no = find_chunk_for(ptr, lc_chunks, lc_n_chunks);
485      tl_assert(ch_no >= -1 && ch_no < lc_n_chunks);
486
487      if (ch_no == -1) {
488         return False;
489      } else {
490         // Ok, we've found a pointer to a chunk.  Get the MC_Chunk and its
491         // LC_Extra.
492         ch = lc_chunks[ch_no];
493         ex = &(lc_extras[ch_no]);
494
495         tl_assert(ptr >= ch->data);
496         tl_assert(ptr < ch->data + ch->szB + (ch->szB==0  ? 1  : 0));
497
498         if (VG_DEBUG_LEAKCHECK)
499            VG_(printf)("ptr=%#lx -> block %d\n", ptr, ch_no);
500
501         *pch_no = ch_no;
502         *pch    = ch;
503         *pex    = ex;
504
505         return True;
506      }
507   }
508}
509
510// Push a chunk (well, just its index) onto the mark stack.
511static void lc_push(Int ch_no, MC_Chunk* ch)
512{
513   if (0) {
514      VG_(printf)("pushing %#lx-%#lx\n", ch->data, ch->data + ch->szB);
515   }
516   lc_markstack_top++;
517   tl_assert(lc_markstack_top < lc_n_chunks);
518   lc_markstack[lc_markstack_top] = ch_no;
519}
520
521// Return the index of the chunk on the top of the mark stack, or -1 if
522// there isn't one.
523static Bool lc_pop(Int* ret)
524{
525   if (-1 == lc_markstack_top) {
526      return False;
527   } else {
528      tl_assert(0 <= lc_markstack_top && lc_markstack_top < lc_n_chunks);
529      *ret = lc_markstack[lc_markstack_top];
530      lc_markstack_top--;
531      return True;
532   }
533}
534
535
536// If 'ptr' is pointing to a heap-allocated block which hasn't been seen
537// before, push it onto the mark stack.
538static void
539lc_push_without_clique_if_a_chunk_ptr(Addr ptr, Bool is_prior_definite)
540{
541   Int ch_no;
542   MC_Chunk* ch;
543   LC_Extra* ex;
544
545   if ( ! lc_is_a_chunk_ptr(ptr, &ch_no, &ch, &ex) )
546      return;
547
548   // Only push it if it hasn't been seen previously.
549   if (ex->state == Unreached) {
550      lc_push(ch_no, ch);
551   }
552
553   // Possibly upgrade the state, ie. one of:
554   // - Unreached --> Possible
555   // - Unreached --> Reachable
556   // - Possible  --> Reachable
557   if (ptr == ch->data && is_prior_definite) {
558      // 'ptr' points to the start of the block, and the prior node is
559      // definite, which means that this block is definitely reachable.
560      ex->state = Reachable;
561
562   } else if (ex->state == Unreached) {
563      // Either 'ptr' is a interior-pointer, or the prior node isn't definite,
564      // which means that we can only mark this block as possibly reachable.
565      ex->state = Possible;
566   }
567}
568
569static void
570lc_push_if_a_chunk_ptr_register(Addr ptr)
571{
572   lc_push_without_clique_if_a_chunk_ptr(ptr, /*is_prior_definite*/True);
573}
574
575// If ptr is pointing to a heap-allocated block which hasn't been seen
576// before, push it onto the mark stack.  Clique is the index of the
577// clique leader.
578static void
579lc_push_with_clique_if_a_chunk_ptr(Addr ptr, Int clique)
580{
581   Int ch_no;
582   MC_Chunk* ch;
583   LC_Extra* ex;
584
585   tl_assert(0 <= clique && clique < lc_n_chunks);
586
587   if ( ! lc_is_a_chunk_ptr(ptr, &ch_no, &ch, &ex) )
588      return;
589
590   // If it's not Unreached, it's already been handled so ignore it.
591   // If ch_no==clique, it's the clique leader, which means this is a cyclic
592   // structure;  again ignore it because it's already been handled.
593   if (ex->state == Unreached && ch_no != clique) {
594      // Note that, unlike reachable blocks, we currently don't distinguish
595      // between start-pointers and interior-pointers here.  We probably
596      // should, though.
597      ex->state = IndirectLeak;
598      lc_push(ch_no, ch);
599
600      // Add the block to the clique, and add its size to the
601      // clique-leader's indirect size.  Also, if the new block was
602      // itself a clique leader, it isn't any more, so add its
603      // indirect_szB to the new clique leader.
604      if (VG_DEBUG_CLIQUE) {
605         if (ex->indirect_szB > 0)
606            VG_(printf)("  clique %d joining clique %d adding %lu+%lu\n",
607                        ch_no, clique, (SizeT)ch->szB, (SizeT)ex->indirect_szB);
608         else
609            VG_(printf)("  block %d joining clique %d adding %lu\n",
610                        ch_no, clique, (SizeT)ch->szB);
611      }
612
613      lc_extras[clique].indirect_szB += ch->szB;
614      lc_extras[clique].indirect_szB += ex->indirect_szB;
615      ex->indirect_szB = 0;    // Shouldn't matter.
616   }
617}
618
619static void
620lc_push_if_a_chunk_ptr(Addr ptr, Int clique, Bool is_prior_definite)
621{
622   if (-1 == clique)
623      lc_push_without_clique_if_a_chunk_ptr(ptr, is_prior_definite);
624   else
625      lc_push_with_clique_if_a_chunk_ptr(ptr, clique);
626}
627
628
629static jmp_buf memscan_jmpbuf;
630
631static
632void scan_all_valid_memory_catcher ( Int sigNo, Addr addr )
633{
634   if (0)
635      VG_(printf)("OUCH! sig=%d addr=%#lx\n", sigNo, addr);
636   if (sigNo == VKI_SIGSEGV || sigNo == VKI_SIGBUS)
637      __builtin_longjmp(memscan_jmpbuf, 1);
638}
639
640// Scan a block of memory between [start, start+len).  This range may
641// be bogus, inaccessable, or otherwise strange; we deal with it.  For each
642// valid aligned word we assume it's a pointer to a chunk a push the chunk
643// onto the mark stack if so.
644static void
645lc_scan_memory(Addr start, SizeT len, Bool is_prior_definite, Int clique)
646{
647   Addr ptr = VG_ROUNDUP(start,     sizeof(Addr));
648   Addr end = VG_ROUNDDN(start+len, sizeof(Addr));
649   vki_sigset_t sigmask;
650
651   if (VG_DEBUG_LEAKCHECK)
652      VG_(printf)("scan %#lx-%#lx (%lu)\n", start, end, len);
653
654   VG_(sigprocmask)(VKI_SIG_SETMASK, NULL, &sigmask);
655   VG_(set_fault_catcher)(scan_all_valid_memory_catcher);
656
657   // We might be in the middle of a page.  Do a cheap check to see if
658   // it's valid;  if not, skip onto the next page.
659   if (!VG_(am_is_valid_for_client)(ptr, sizeof(Addr), VKI_PROT_READ))
660      ptr = VG_PGROUNDUP(ptr+1);        // First page is bad - skip it.
661
662   while (ptr < end) {
663      Addr addr;
664
665      // Skip invalid chunks.
666      if ( ! MC_(is_within_valid_secondary)(ptr) ) {
667         ptr = VG_ROUNDUP(ptr+1, SM_SIZE);
668         continue;
669      }
670
671      // Look to see if this page seems reasonable.
672      if ((ptr % VKI_PAGE_SIZE) == 0) {
673         if (!VG_(am_is_valid_for_client)(ptr, sizeof(Addr), VKI_PROT_READ)) {
674            ptr += VKI_PAGE_SIZE;      // Bad page - skip it.
675            continue;
676         }
677      }
678
679      if (__builtin_setjmp(memscan_jmpbuf) == 0) {
680         if ( MC_(is_valid_aligned_word)(ptr) ) {
681            lc_scanned_szB += sizeof(Addr);
682            addr = *(Addr *)ptr;
683            // If we get here, the scanned word is in valid memory.  Now
684            // let's see if its contents point to a chunk.
685            lc_push_if_a_chunk_ptr(addr, clique, is_prior_definite);
686         } else if (0 && VG_DEBUG_LEAKCHECK) {
687            VG_(printf)("%#lx not valid\n", ptr);
688         }
689         ptr += sizeof(Addr);
690      } else {
691         // We need to restore the signal mask, because we were
692         // longjmped out of a signal handler.
693         VG_(sigprocmask)(VKI_SIG_SETMASK, &sigmask, NULL);
694
695         ptr = VG_PGROUNDUP(ptr+1);     // Bad page - skip it.
696      }
697   }
698
699   VG_(sigprocmask)(VKI_SIG_SETMASK, &sigmask, NULL);
700   VG_(set_fault_catcher)(NULL);
701}
702
703
704// Process the mark stack until empty.
705static void lc_process_markstack(Int clique)
706{
707   Int  top = -1;    // shut gcc up
708   Bool is_prior_definite;
709
710   while (lc_pop(&top)) {
711      tl_assert(top >= 0 && top < lc_n_chunks);
712
713      // See comment about 'is_prior_definite' at the top to understand this.
714      is_prior_definite = ( Possible != lc_extras[top].state );
715
716      lc_scan_memory(lc_chunks[top]->data, lc_chunks[top]->szB,
717                     is_prior_definite, clique);
718   }
719}
720
721static Word cmp_LossRecordKey_LossRecord(const void* key, const void* elem)
722{
723   LossRecordKey* a = (LossRecordKey*)key;
724   LossRecordKey* b = &(((LossRecord*)elem)->key);
725
726   // Compare on states first because that's fast.
727   if (a->state < b->state) return -1;
728   if (a->state > b->state) return  1;
729   // Ok, the states are equal.  Now compare the locations, which is slower.
730   if (VG_(eq_ExeContext)(
731            MC_(clo_leak_resolution), a->allocated_at, b->allocated_at))
732      return 0;
733   // Different locations.  Ordering is arbitrary, just use the ec pointer.
734   if (a->allocated_at < b->allocated_at) return -1;
735   if (a->allocated_at > b->allocated_at) return  1;
736   VG_(tool_panic)("bad LossRecord comparison");
737}
738
739static Int cmp_LossRecords(void* va, void* vb)
740{
741   LossRecord* lr_a = *(LossRecord**)va;
742   LossRecord* lr_b = *(LossRecord**)vb;
743   SizeT total_szB_a = lr_a->szB + lr_a->indirect_szB;
744   SizeT total_szB_b = lr_b->szB + lr_b->indirect_szB;
745
746   // First compare by sizes.
747   if (total_szB_a < total_szB_b) return -1;
748   if (total_szB_a > total_szB_b) return  1;
749   // If size are equal, compare by states.
750   if (lr_a->key.state < lr_b->key.state) return -1;
751   if (lr_a->key.state > lr_b->key.state) return  1;
752   // If they're still equal here, it doesn't matter that much, but we keep
753   // comparing other things so that regtests are as deterministic as
754   // possible.  So:  compare num_blocks.
755   if (lr_a->num_blocks < lr_b->num_blocks) return -1;
756   if (lr_a->num_blocks > lr_b->num_blocks) return  1;
757   // Finally, compare ExeContext addresses... older ones are likely to have
758   // lower addresses.
759   if (lr_a->key.allocated_at < lr_b->key.allocated_at) return -1;
760   if (lr_a->key.allocated_at > lr_b->key.allocated_at) return  1;
761   return 0;
762}
763
764static void print_results(ThreadId tid, Bool is_full_check)
765{
766   Int          i, n_lossrecords;
767   OSet*        lr_table;
768   LossRecord** lr_array;
769   LossRecord*  lr;
770   Bool         is_suppressed;
771
772   // Create the lr_table, which holds the loss records.
773   lr_table =
774      VG_(OSetGen_Create)(offsetof(LossRecord, key),
775                          cmp_LossRecordKey_LossRecord,
776                          VG_(malloc), "mc.pr.1",
777                          VG_(free));
778
779   // Convert the chunks into loss records, merging them where appropriate.
780   for (i = 0; i < lc_n_chunks; i++) {
781      MC_Chunk*     ch = lc_chunks[i];
782      LC_Extra*     ex = &(lc_extras)[i];
783      LossRecord*   old_lr;
784      LossRecordKey lrkey;
785      lrkey.state        = ex->state;
786      lrkey.allocated_at = ch->where;
787
788      old_lr = VG_(OSetGen_Lookup)(lr_table, &lrkey);
789      if (old_lr) {
790         // We found an existing loss record matching this chunk.  Update the
791         // loss record's details in-situ.  This is safe because we don't
792         // change the elements used as the OSet key.
793         old_lr->szB          += ch->szB;
794         old_lr->indirect_szB += ex->indirect_szB;
795         old_lr->num_blocks++;
796      } else {
797         // No existing loss record matches this chunk.  Create a new loss
798         // record, initialise it from the chunk, and insert it into lr_table.
799         lr = VG_(OSetGen_AllocNode)(lr_table, sizeof(LossRecord));
800         lr->key              = lrkey;
801         lr->szB              = ch->szB;
802         lr->indirect_szB     = ex->indirect_szB;
803         lr->num_blocks       = 1;
804         VG_(OSetGen_Insert)(lr_table, lr);
805      }
806   }
807   n_lossrecords = VG_(OSetGen_Size)(lr_table);
808
809   // Create an array of pointers to the loss records.
810   lr_array = VG_(malloc)("mc.pr.2", n_lossrecords * sizeof(LossRecord*));
811   i = 0;
812   VG_(OSetGen_ResetIter)(lr_table);
813   while ( (lr = VG_(OSetGen_Next)(lr_table)) ) {
814      lr_array[i++] = lr;
815   }
816   tl_assert(i == n_lossrecords);
817
818   // Sort the array by loss record sizes.
819   VG_(ssort)(lr_array, n_lossrecords, sizeof(LossRecord*),
820              cmp_LossRecords);
821
822   // Zero totals.
823   MC_(blocks_leaked)     = MC_(bytes_leaked)     = 0;
824   MC_(blocks_indirect)   = MC_(bytes_indirect)   = 0;
825   MC_(blocks_dubious)    = MC_(bytes_dubious)    = 0;
826   MC_(blocks_reachable)  = MC_(bytes_reachable)  = 0;
827   MC_(blocks_suppressed) = MC_(bytes_suppressed) = 0;
828
829   // Print the loss records (in size order) and collect summary stats.
830   for (i = 0; i < n_lossrecords; i++) {
831      Bool count_as_error, print_record;
832      // Rules for printing:
833      // - We don't show suppressed loss records ever (and that's controlled
834      //   within the error manager).
835      // - We show non-suppressed loss records that are not "reachable" if
836      //   --leak-check=yes.
837      // - We show all non-suppressed loss records if --leak-check=yes and
838      //   --show-reachable=yes.
839      //
840      // Nb: here "reachable" means Reachable *or* IndirectLeak;  note that
841      // this is different to "still reachable" used elsewhere because it
842      // includes indirectly lost blocks!
843      //
844      lr = lr_array[i];
845      count_as_error = Unreached == lr->key.state ||
846                       Possible  == lr->key.state;
847      print_record = is_full_check &&
848                     ( MC_(clo_show_reachable) || count_as_error );
849      is_suppressed =
850         MC_(record_leak_error) ( tid, i+1, n_lossrecords, lr, print_record,
851                                  count_as_error );
852
853      if (is_suppressed) {
854         MC_(blocks_suppressed) += lr->num_blocks;
855         MC_(bytes_suppressed)  += lr->szB;
856
857      } else if (Unreached == lr->key.state) {
858         MC_(blocks_leaked)     += lr->num_blocks;
859         MC_(bytes_leaked)      += lr->szB;
860
861      } else if (IndirectLeak == lr->key.state) {
862         MC_(blocks_indirect)   += lr->num_blocks;
863         MC_(bytes_indirect)    += lr->szB;
864
865      } else if (Possible == lr->key.state) {
866         MC_(blocks_dubious)    += lr->num_blocks;
867         MC_(bytes_dubious)     += lr->szB;
868
869      } else if (Reachable == lr->key.state) {
870         MC_(blocks_reachable)  += lr->num_blocks;
871         MC_(bytes_reachable)   += lr->szB;
872
873      } else {
874         VG_(tool_panic)("unknown loss mode");
875      }
876   }
877
878   if (VG_(clo_verbosity) > 0 && !VG_(clo_xml)) {
879      VG_(umsg)("\n");
880      VG_(umsg)("LEAK SUMMARY:\n");
881      VG_(umsg)("   definitely lost: %'lu bytes in %'lu blocks.\n",
882                MC_(bytes_leaked), MC_(blocks_leaked) );
883      VG_(umsg)("   indirectly lost: %'lu bytes in %'lu blocks.\n",
884                MC_(bytes_indirect), MC_(blocks_indirect) );
885      VG_(umsg)("     possibly lost: %'lu bytes in %'lu blocks.\n",
886                MC_(bytes_dubious), MC_(blocks_dubious) );
887      VG_(umsg)("   still reachable: %'lu bytes in %'lu blocks.\n",
888                MC_(bytes_reachable), MC_(blocks_reachable) );
889      VG_(umsg)("        suppressed: %'lu bytes in %'lu blocks.\n",
890                MC_(bytes_suppressed), MC_(blocks_suppressed) );
891      if (!is_full_check &&
892          (MC_(blocks_leaked) + MC_(blocks_indirect) +
893           MC_(blocks_dubious) + MC_(blocks_reachable)) > 0) {
894         VG_(umsg)("Rerun with --leak-check=full to see details "
895                   "of leaked memory.\n");
896      }
897      if (is_full_check &&
898          MC_(blocks_reachable) > 0 && !MC_(clo_show_reachable))
899      {
900         VG_(umsg)("Reachable blocks (those to which a pointer "
901                   "was found) are not shown.\n");
902         VG_(umsg)("To see them, rerun with: --leak-check=full "
903                   "--show-reachable=yes\n");
904      }
905   }
906}
907
908/*------------------------------------------------------------*/
909/*--- Top-level entry point.                               ---*/
910/*------------------------------------------------------------*/
911
912void MC_(detect_memory_leaks) ( ThreadId tid, LeakCheckMode mode )
913{
914   Int i, j;
915
916   tl_assert(mode != LC_Off);
917
918   // Get the chunks, stop if there were none.
919   lc_chunks = find_active_chunks(&lc_n_chunks);
920   if (lc_n_chunks == 0) {
921      tl_assert(lc_chunks == NULL);
922      if (VG_(clo_verbosity) >= 1 && !VG_(clo_xml)) {
923         VG_(umsg)("\n");
924         VG_(umsg)("All heap blocks were freed -- no leaks are possible.\n");
925      }
926      return;
927   }
928
929   // Sort the array so blocks are in ascending order in memory.
930   VG_(ssort)(lc_chunks, lc_n_chunks, sizeof(VgHashNode*), compare_MC_Chunks);
931
932   // Sanity check -- make sure they're in order.
933   for (i = 0; i < lc_n_chunks-1; i++) {
934      tl_assert( lc_chunks[i]->data <= lc_chunks[i+1]->data);
935   }
936
937   // Sanity check -- make sure they don't overlap.  The one exception is that
938   // we allow a MALLOCLIKE block to sit entirely within a malloc() block.
939   // This is for bug 100628.  If this occurs, we ignore the malloc() block
940   // for leak-checking purposes.  This is a hack and probably should be done
941   // better, but at least it's consistent with mempools (which are treated
942   // like this in find_active_chunks).  Mempools have a separate VgHashTable
943   // for mempool chunks, but if custom-allocated blocks are put in a separate
944   // table from normal heap blocks it makes free-mismatch checking more
945   // difficult.
946   //
947   // If this check fails, it probably means that the application
948   // has done something stupid with VALGRIND_MALLOCLIKE_BLOCK client
949   // requests, eg. has made overlapping requests (which are
950   // nonsensical), or used VALGRIND_MALLOCLIKE_BLOCK for stack locations;
951   // again nonsensical.
952   //
953   for (i = 0; i < lc_n_chunks-1; i++) {
954      MC_Chunk* ch1 = lc_chunks[i];
955      MC_Chunk* ch2 = lc_chunks[i+1];
956
957      Addr start1    = ch1->data;
958      Addr start2    = ch2->data;
959      Addr end1      = ch1->data + ch1->szB - 1;
960      Addr end2      = ch2->data + ch2->szB - 1;
961      Bool isCustom1 = ch1->allockind == MC_AllocCustom;
962      Bool isCustom2 = ch2->allockind == MC_AllocCustom;
963
964      if (end1 < start2) {
965         // Normal case - no overlap.
966
967      // We used to allow exact duplicates, I'm not sure why.  --njn
968      //} else if (start1 == start2 && end1 == end2) {
969         // Degenerate case: exact duplicates.
970
971      } else if (start1 >= start2 && end1 <= end2 && isCustom1 && !isCustom2) {
972         // Block i is MALLOCLIKE and entirely within block i+1.
973         // Remove block i+1.
974         for (j = i+1; j < lc_n_chunks-1; j++) {
975            lc_chunks[j] = lc_chunks[j+1];
976         }
977         lc_n_chunks--;
978
979      } else if (start2 >= start1 && end2 <= end1 && isCustom2 && !isCustom1) {
980         // Block i+1 is MALLOCLIKE and entirely within block i.
981         // Remove block i.
982         for (j = i; j < lc_n_chunks-1; j++) {
983            lc_chunks[j] = lc_chunks[j+1];
984         }
985         lc_n_chunks--;
986
987      } else {
988         VG_(umsg)("Block 0x%lx..0x%lx overlaps with block 0x%lx..0x%lx",
989                   start1, end1, start1, end2);
990         VG_(umsg)("This is usually caused by using VALGRIND_MALLOCLIKE_BLOCK");
991         VG_(umsg)("in an inappropriate way.");
992         tl_assert (0);
993      }
994   }
995
996   // Initialise lc_extras.
997   lc_extras = VG_(malloc)( "mc.dml.2", lc_n_chunks * sizeof(LC_Extra) );
998   for (i = 0; i < lc_n_chunks; i++) {
999      lc_extras[i].state        = Unreached;
1000      lc_extras[i].indirect_szB = 0;
1001   }
1002
1003   // Initialise lc_markstack.
1004   lc_markstack = VG_(malloc)( "mc.dml.2", lc_n_chunks * sizeof(Int) );
1005   for (i = 0; i < lc_n_chunks; i++) {
1006      lc_markstack[i] = -1;
1007   }
1008   lc_markstack_top = -1;
1009
1010   // Verbosity.
1011   if (VG_(clo_verbosity) > 1 && !VG_(clo_xml)) {
1012      VG_(umsg)( "\n" );
1013      VG_(umsg)( "Searching for pointers to %'d not-freed blocks.\n",
1014                 lc_n_chunks );
1015   }
1016
1017   // Scan the memory root-set, pushing onto the mark stack any blocks
1018   // pointed to.
1019   {
1020      Int   n_seg_starts;
1021      Addr* seg_starts = VG_(get_segment_starts)( &n_seg_starts );
1022
1023      tl_assert(seg_starts && n_seg_starts > 0);
1024
1025      lc_scanned_szB = 0;
1026
1027      // VG_(am_show_nsegments)( 0, "leakcheck");
1028      for (i = 0; i < n_seg_starts; i++) {
1029         SizeT seg_size;
1030         NSegment const* seg = VG_(am_find_nsegment)( seg_starts[i] );
1031         tl_assert(seg);
1032
1033         if (seg->kind != SkFileC && seg->kind != SkAnonC) continue;
1034         if (!(seg->hasR && seg->hasW))                    continue;
1035         if (seg->isCH)                                    continue;
1036
1037         // Don't poke around in device segments as this may cause
1038         // hangs.  Exclude /dev/zero just in case someone allocated
1039         // memory by explicitly mapping /dev/zero.
1040         if (seg->kind == SkFileC
1041             && (VKI_S_ISCHR(seg->mode) || VKI_S_ISBLK(seg->mode))) {
1042            HChar* dev_name = VG_(am_get_filename)( (NSegment*)seg );
1043            if (dev_name && 0 == VG_(strcmp)(dev_name, "/dev/zero")) {
1044               // Don't skip /dev/zero.
1045            } else {
1046               // Skip this device mapping.
1047               continue;
1048            }
1049         }
1050
1051         if (0)
1052            VG_(printf)("ACCEPT %2d  %#lx %#lx\n", i, seg->start, seg->end);
1053
1054         // Scan the segment.  We use -1 for the clique number, because this
1055         // is a root-set.
1056         seg_size = seg->end - seg->start + 1;
1057         if (VG_(clo_verbosity) > 2) {
1058            VG_(message)(Vg_DebugMsg,
1059                         "  Scanning root segment: %#lx..%#lx (%lu)\n",
1060                         seg->start, seg->end, seg_size);
1061         }
1062         lc_scan_memory(seg->start, seg_size, /*is_prior_definite*/True, -1);
1063      }
1064   }
1065
1066   // Scan GP registers for chunk pointers.
1067   VG_(apply_to_GP_regs)(lc_push_if_a_chunk_ptr_register);
1068
1069   // Process the pushed blocks.  After this, every block that is reachable
1070   // from the root-set has been traced.
1071   lc_process_markstack(/*clique*/-1);
1072
1073   if (VG_(clo_verbosity) > 1 && !VG_(clo_xml))
1074      VG_(umsg)("Checked %'lu bytes.\n", lc_scanned_szB);
1075
1076   // Trace all the leaked blocks to determine which are directly leaked and
1077   // which are indirectly leaked.  For each Unreached block, push it onto
1078   // the mark stack, and find all the as-yet-Unreached blocks reachable
1079   // from it.  These form a clique and are marked IndirectLeak, and their
1080   // size is added to the clique leader's indirect size.  If one of the
1081   // found blocks was itself a clique leader (from a previous clique), then
1082   // the cliques are merged.
1083   for (i = 0; i < lc_n_chunks; i++) {
1084      MC_Chunk* ch = lc_chunks[i];
1085      LC_Extra* ex = &(lc_extras[i]);
1086
1087      if (VG_DEBUG_CLIQUE)
1088         VG_(printf)("cliques: %d at %#lx -> Loss state %d\n",
1089                     i, ch->data, ex->state);
1090
1091      tl_assert(lc_markstack_top == -1);
1092
1093      if (ex->state == Unreached) {
1094         if (VG_DEBUG_CLIQUE)
1095            VG_(printf)("%d: gathering clique %#lx\n", i, ch->data);
1096
1097         // Push this Unreached block onto the stack and process it.
1098         lc_push(i, ch);
1099         lc_process_markstack(i);
1100
1101         tl_assert(lc_markstack_top == -1);
1102         tl_assert(ex->state == Unreached);
1103      }
1104   }
1105
1106   print_results( tid, ( mode == LC_Full ? True : False ) );
1107
1108   VG_(free) ( lc_chunks );
1109   VG_(free) ( lc_extras );
1110   VG_(free) ( lc_markstack );
1111}
1112
1113/*--------------------------------------------------------------------*/
1114/*--- end                                                          ---*/
1115/*--------------------------------------------------------------------*/
1116
1117