blob: 55a5b7bb1fe010cb847da97cf335b566f91e81d1 [file] [log] [blame]
njn43c799e2003-04-08 00:08:52 +00001
2/*--------------------------------------------------------------------*/
njn1d0825f2006-03-27 11:37:07 +00003/*--- The leak checker. mc_leakcheck.c ---*/
njn43c799e2003-04-08 00:08:52 +00004/*--------------------------------------------------------------------*/
5
6/*
nethercote137bc552003-11-14 17:47:54 +00007 This file is part of MemCheck, a heavyweight Valgrind tool for
njn1d0825f2006-03-27 11:37:07 +00008 detecting memory errors.
njn43c799e2003-04-08 00:08:52 +00009
sewardjec062e82011-10-23 07:32:08 +000010 Copyright (C) 2000-2011 Julian Seward
njn43c799e2003-04-08 00:08:52 +000011 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
njnc7561b92005-06-19 01:24:32 +000031#include "pub_tool_basics.h"
sewardj4cfea4f2006-10-14 19:26:10 +000032#include "pub_tool_vki.h"
njnac1e0332009-05-08 00:39:31 +000033#include "pub_tool_aspacehl.h"
njn4802b382005-06-11 04:58:29 +000034#include "pub_tool_aspacemgr.h"
njn1d0825f2006-03-27 11:37:07 +000035#include "pub_tool_execontext.h"
36#include "pub_tool_hashtable.h"
njn97405b22005-06-02 03:39:33 +000037#include "pub_tool_libcbase.h"
njn132bfcc2005-06-04 19:16:06 +000038#include "pub_tool_libcassert.h"
njn36a20fa2005-06-03 03:08:39 +000039#include "pub_tool_libcprint.h"
njnde62cbf2005-06-10 22:08:14 +000040#include "pub_tool_libcsignal.h"
njn6ace3ea2005-06-17 03:06:27 +000041#include "pub_tool_machine.h"
njnc7561b92005-06-19 01:24:32 +000042#include "pub_tool_mallocfree.h"
43#include "pub_tool_options.h"
njn29a5c012009-05-06 06:15:55 +000044#include "pub_tool_oset.h"
philippe6643e962012-01-17 21:16:30 +000045#include "pub_tool_poolalloc.h"
46#include "pub_tool_signals.h" // Needed for mc_include.h
sewardj6c591e12011-04-11 16:17:51 +000047#include "pub_tool_libcsetjmp.h" // setjmp facilities
njn1d0825f2006-03-27 11:37:07 +000048#include "pub_tool_tooliface.h" // Needed for mc_include.h
njn43c799e2003-04-08 00:08:52 +000049
njn1d0825f2006-03-27 11:37:07 +000050#include "mc_include.h"
njnc7561b92005-06-19 01:24:32 +000051
njn8225cc02009-03-09 22:52:24 +000052/*------------------------------------------------------------*/
53/*--- An overview of leak checking. ---*/
54/*------------------------------------------------------------*/
njnc7561b92005-06-19 01:24:32 +000055
njn8225cc02009-03-09 22:52:24 +000056// 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?
njnc2f8b1b2009-08-10 06:47:00 +0000116// [bug 197930]
njn8225cc02009-03-09 22:52:24 +0000117//
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.
sewardjb5f6f512005-03-10 23:59:00 +0000238#define VG_DEBUG_LEAKCHECK 0
njn8225cc02009-03-09 22:52:24 +0000239#define VG_DEBUG_CLIQUE 0
240
sewardjb5f6f512005-03-10 23:59:00 +0000241
njn43c799e2003-04-08 00:08:52 +0000242/*------------------------------------------------------------*/
njn8225cc02009-03-09 22:52:24 +0000243/*--- Getting the initial chunks, and searching them. ---*/
njn43c799e2003-04-08 00:08:52 +0000244/*------------------------------------------------------------*/
245
njn8225cc02009-03-09 22:52:24 +0000246// Compare the MC_Chunks by 'data' (i.e. the address of the block).
247static Int compare_MC_Chunks(void* n1, void* n2)
njn43c799e2003-04-08 00:08:52 +0000248{
njn8225cc02009-03-09 22:52:24 +0000249 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;
njn43c799e2003-04-08 00:08:52 +0000254}
255
njn8225cc02009-03-09 22:52:24 +0000256#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.
tom1d0f3f62010-10-04 20:55:21 +0000428 UInt pending:1; // Scan pending.
429 SizeT indirect_szB : (sizeof(SizeT)*8)-3; // If Unreached, how many bytes
njn8225cc02009-03-09 22:52:24 +0000430 // are unreachable from here.
431 }
432 LC_Extra;
433
434// An array holding pointers to every chunk we're checking. Sorted by address.
435static MC_Chunk** lc_chunks;
436// How many chunks we're dealing with.
437static Int lc_n_chunks;
sewardjc8bd1df2011-06-26 12:41:33 +0000438// chunks will be converted and merged in loss record, maintained in lr_table
439// lr_table elements are kept from one leak_search to another to implement
440// the "print new/changed leaks" client request
441static OSet* lr_table;
442
443// DeltaMode used the last time we called detect_memory_leaks.
444// The recorded leak errors must be output using a logic based on this delta_mode.
445// The below avoids replicating the delta_mode in each LossRecord.
446LeakCheckDeltaMode MC_(detect_memory_leaks_last_delta_mode);
447
njn8225cc02009-03-09 22:52:24 +0000448
449// This has the same number of entries as lc_chunks, and each entry
450// in lc_chunks corresponds with the entry here (ie. lc_chunks[i] and
451// lc_extras[i] describe the same block).
452static LC_Extra* lc_extras;
453
454// Records chunks that are currently being processed. Each element in the
455// stack is an index into lc_chunks and lc_extras. Its size is
456// 'lc_n_chunks' because in the worst case that's how many chunks could be
457// pushed onto it (actually I think the maximum is lc_n_chunks-1 but let's
458// be conservative).
459static Int* lc_markstack;
460// The index of the top element of the stack; -1 if the stack is empty, 0 if
461// the stack has one element, 1 if it has two, etc.
462static Int lc_markstack_top;
463
464// Keeps track of how many bytes of memory we've scanned, for printing.
465// (Nb: We don't keep track of how many register bytes we've scanned.)
466static SizeT lc_scanned_szB;
467
468
469SizeT MC_(bytes_leaked) = 0;
470SizeT MC_(bytes_indirect) = 0;
471SizeT MC_(bytes_dubious) = 0;
472SizeT MC_(bytes_reachable) = 0;
473SizeT MC_(bytes_suppressed) = 0;
474
475SizeT MC_(blocks_leaked) = 0;
476SizeT MC_(blocks_indirect) = 0;
477SizeT MC_(blocks_dubious) = 0;
478SizeT MC_(blocks_reachable) = 0;
479SizeT MC_(blocks_suppressed) = 0;
480
sewardj45f4e7c2005-09-27 19:20:21 +0000481
njn8225cc02009-03-09 22:52:24 +0000482// Determines if a pointer is to a chunk. Returns the chunk number et al
483// via call-by-reference.
484static Bool
485lc_is_a_chunk_ptr(Addr ptr, Int* pch_no, MC_Chunk** pch, LC_Extra** pex)
njn43c799e2003-04-08 00:08:52 +0000486{
njn8225cc02009-03-09 22:52:24 +0000487 Int ch_no;
488 MC_Chunk* ch;
489 LC_Extra* ex;
njn43c799e2003-04-08 00:08:52 +0000490
njn8225cc02009-03-09 22:52:24 +0000491 // Quick filter.
492 if (!VG_(am_is_valid_for_client)(ptr, 1, VKI_PROT_READ)) {
493 return False;
sewardjb5f6f512005-03-10 23:59:00 +0000494 } else {
njn8225cc02009-03-09 22:52:24 +0000495 ch_no = find_chunk_for(ptr, lc_chunks, lc_n_chunks);
496 tl_assert(ch_no >= -1 && ch_no < lc_n_chunks);
497
498 if (ch_no == -1) {
499 return False;
500 } else {
501 // Ok, we've found a pointer to a chunk. Get the MC_Chunk and its
502 // LC_Extra.
503 ch = lc_chunks[ch_no];
504 ex = &(lc_extras[ch_no]);
505
506 tl_assert(ptr >= ch->data);
507 tl_assert(ptr < ch->data + ch->szB + (ch->szB==0 ? 1 : 0));
508
509 if (VG_DEBUG_LEAKCHECK)
510 VG_(printf)("ptr=%#lx -> block %d\n", ptr, ch_no);
511
512 *pch_no = ch_no;
513 *pch = ch;
514 *pex = ex;
515
516 return True;
517 }
sewardjb5f6f512005-03-10 23:59:00 +0000518 }
519}
520
njn8225cc02009-03-09 22:52:24 +0000521// Push a chunk (well, just its index) onto the mark stack.
522static void lc_push(Int ch_no, MC_Chunk* ch)
sewardjb5f6f512005-03-10 23:59:00 +0000523{
tom1d0f3f62010-10-04 20:55:21 +0000524 if (!lc_extras[ch_no].pending) {
525 if (0) {
526 VG_(printf)("pushing %#lx-%#lx\n", ch->data, ch->data + ch->szB);
527 }
528 lc_markstack_top++;
529 tl_assert(lc_markstack_top < lc_n_chunks);
530 lc_markstack[lc_markstack_top] = ch_no;
531 tl_assert(!lc_extras[ch_no].pending);
532 lc_extras[ch_no].pending = True;
njn8225cc02009-03-09 22:52:24 +0000533 }
sewardjb5f6f512005-03-10 23:59:00 +0000534}
535
njn8225cc02009-03-09 22:52:24 +0000536// Return the index of the chunk on the top of the mark stack, or -1 if
537// there isn't one.
538static Bool lc_pop(Int* ret)
sewardjb5f6f512005-03-10 23:59:00 +0000539{
njn8225cc02009-03-09 22:52:24 +0000540 if (-1 == lc_markstack_top) {
541 return False;
542 } else {
543 tl_assert(0 <= lc_markstack_top && lc_markstack_top < lc_n_chunks);
544 *ret = lc_markstack[lc_markstack_top];
545 lc_markstack_top--;
tom1d0f3f62010-10-04 20:55:21 +0000546 tl_assert(lc_extras[*ret].pending);
547 lc_extras[*ret].pending = False;
njn8225cc02009-03-09 22:52:24 +0000548 return True;
549 }
550}
sewardjb5f6f512005-03-10 23:59:00 +0000551
njn8225cc02009-03-09 22:52:24 +0000552
553// If 'ptr' is pointing to a heap-allocated block which hasn't been seen
554// before, push it onto the mark stack.
555static void
556lc_push_without_clique_if_a_chunk_ptr(Addr ptr, Bool is_prior_definite)
557{
558 Int ch_no;
559 MC_Chunk* ch;
560 LC_Extra* ex;
561
562 if ( ! lc_is_a_chunk_ptr(ptr, &ch_no, &ch, &ex) )
563 return;
tom1d0f3f62010-10-04 20:55:21 +0000564
njn8225cc02009-03-09 22:52:24 +0000565 // Possibly upgrade the state, ie. one of:
566 // - Unreached --> Possible
567 // - Unreached --> Reachable
568 // - Possible --> Reachable
tom1d0f3f62010-10-04 20:55:21 +0000569 if (ptr == ch->data && is_prior_definite && ex->state != Reachable) {
njn8225cc02009-03-09 22:52:24 +0000570 // 'ptr' points to the start of the block, and the prior node is
571 // definite, which means that this block is definitely reachable.
572 ex->state = Reachable;
573
tom1d0f3f62010-10-04 20:55:21 +0000574 // State has changed to Reachable so (re)scan the block to make
575 // sure any blocks it points to are correctly marked.
576 lc_push(ch_no, ch);
577
njn8225cc02009-03-09 22:52:24 +0000578 } else if (ex->state == Unreached) {
579 // Either 'ptr' is a interior-pointer, or the prior node isn't definite,
580 // which means that we can only mark this block as possibly reachable.
581 ex->state = Possible;
tom1d0f3f62010-10-04 20:55:21 +0000582
583 // State has changed to Possible so (re)scan the block to make
584 // sure any blocks it points to are correctly marked.
585 lc_push(ch_no, ch);
njn8225cc02009-03-09 22:52:24 +0000586 }
587}
588
589static void
590lc_push_if_a_chunk_ptr_register(Addr ptr)
591{
592 lc_push_without_clique_if_a_chunk_ptr(ptr, /*is_prior_definite*/True);
593}
594
595// If ptr is pointing to a heap-allocated block which hasn't been seen
596// before, push it onto the mark stack. Clique is the index of the
597// clique leader.
598static void
599lc_push_with_clique_if_a_chunk_ptr(Addr ptr, Int clique)
600{
601 Int ch_no;
602 MC_Chunk* ch;
603 LC_Extra* ex;
604
605 tl_assert(0 <= clique && clique < lc_n_chunks);
606
607 if ( ! lc_is_a_chunk_ptr(ptr, &ch_no, &ch, &ex) )
608 return;
609
610 // If it's not Unreached, it's already been handled so ignore it.
611 // If ch_no==clique, it's the clique leader, which means this is a cyclic
612 // structure; again ignore it because it's already been handled.
613 if (ex->state == Unreached && ch_no != clique) {
614 // Note that, unlike reachable blocks, we currently don't distinguish
615 // between start-pointers and interior-pointers here. We probably
616 // should, though.
617 ex->state = IndirectLeak;
618 lc_push(ch_no, ch);
619
620 // Add the block to the clique, and add its size to the
621 // clique-leader's indirect size. Also, if the new block was
622 // itself a clique leader, it isn't any more, so add its
623 // indirect_szB to the new clique leader.
624 if (VG_DEBUG_CLIQUE) {
625 if (ex->indirect_szB > 0)
626 VG_(printf)(" clique %d joining clique %d adding %lu+%lu\n",
bartdc429d12011-07-29 14:24:07 +0000627 ch_no, clique, (unsigned long)ch->szB,
628 (unsigned long)ex->indirect_szB);
njn8225cc02009-03-09 22:52:24 +0000629 else
630 VG_(printf)(" block %d joining clique %d adding %lu\n",
bartdc429d12011-07-29 14:24:07 +0000631 ch_no, clique, (unsigned long)ch->szB);
njn8225cc02009-03-09 22:52:24 +0000632 }
633
634 lc_extras[clique].indirect_szB += ch->szB;
635 lc_extras[clique].indirect_szB += ex->indirect_szB;
636 ex->indirect_szB = 0; // Shouldn't matter.
637 }
638}
639
640static void
641lc_push_if_a_chunk_ptr(Addr ptr, Int clique, Bool is_prior_definite)
642{
643 if (-1 == clique)
644 lc_push_without_clique_if_a_chunk_ptr(ptr, is_prior_definite);
645 else
646 lc_push_with_clique_if_a_chunk_ptr(ptr, clique);
sewardjb5f6f512005-03-10 23:59:00 +0000647}
648
sewardj45d94cc2005-04-20 14:44:11 +0000649
sewardj97d3ebb2011-04-11 18:36:34 +0000650static VG_MINIMAL_JMP_BUF(memscan_jmpbuf);
sewardjb5f6f512005-03-10 23:59:00 +0000651
njn8225cc02009-03-09 22:52:24 +0000652static
653void scan_all_valid_memory_catcher ( Int sigNo, Addr addr )
sewardjb5f6f512005-03-10 23:59:00 +0000654{
njn8225cc02009-03-09 22:52:24 +0000655 if (0)
656 VG_(printf)("OUCH! sig=%d addr=%#lx\n", sigNo, addr);
657 if (sigNo == VKI_SIGSEGV || sigNo == VKI_SIGBUS)
sewardj6c591e12011-04-11 16:17:51 +0000658 VG_MINIMAL_LONGJMP(memscan_jmpbuf);
njn8225cc02009-03-09 22:52:24 +0000659}
660
661// Scan a block of memory between [start, start+len). This range may
662// be bogus, inaccessable, or otherwise strange; we deal with it. For each
663// valid aligned word we assume it's a pointer to a chunk a push the chunk
664// onto the mark stack if so.
665static void
666lc_scan_memory(Addr start, SizeT len, Bool is_prior_definite, Int clique)
667{
668 Addr ptr = VG_ROUNDUP(start, sizeof(Addr));
njn13bfd852005-06-02 03:52:53 +0000669 Addr end = VG_ROUNDDN(start+len, sizeof(Addr));
sewardjb5f6f512005-03-10 23:59:00 +0000670 vki_sigset_t sigmask;
671
672 if (VG_DEBUG_LEAKCHECK)
njn8225cc02009-03-09 22:52:24 +0000673 VG_(printf)("scan %#lx-%#lx (%lu)\n", start, end, len);
674
sewardjb5f6f512005-03-10 23:59:00 +0000675 VG_(sigprocmask)(VKI_SIG_SETMASK, NULL, &sigmask);
njn695c16e2005-03-27 03:40:28 +0000676 VG_(set_fault_catcher)(scan_all_valid_memory_catcher);
sewardjb5f6f512005-03-10 23:59:00 +0000677
njn8225cc02009-03-09 22:52:24 +0000678 // We might be in the middle of a page. Do a cheap check to see if
679 // it's valid; if not, skip onto the next page.
sewardj45f4e7c2005-09-27 19:20:21 +0000680 if (!VG_(am_is_valid_for_client)(ptr, sizeof(Addr), VKI_PROT_READ))
njn8225cc02009-03-09 22:52:24 +0000681 ptr = VG_PGROUNDUP(ptr+1); // First page is bad - skip it.
sewardjb5f6f512005-03-10 23:59:00 +0000682
sewardj05fe85e2005-04-27 22:46:36 +0000683 while (ptr < end) {
sewardjb5f6f512005-03-10 23:59:00 +0000684 Addr addr;
685
njn8225cc02009-03-09 22:52:24 +0000686 // Skip invalid chunks.
687 if ( ! MC_(is_within_valid_secondary)(ptr) ) {
688 ptr = VG_ROUNDUP(ptr+1, SM_SIZE);
689 continue;
sewardjb5f6f512005-03-10 23:59:00 +0000690 }
691
njn8225cc02009-03-09 22:52:24 +0000692 // Look to see if this page seems reasonable.
sewardjb5f6f512005-03-10 23:59:00 +0000693 if ((ptr % VKI_PAGE_SIZE) == 0) {
njn8225cc02009-03-09 22:52:24 +0000694 if (!VG_(am_is_valid_for_client)(ptr, sizeof(Addr), VKI_PROT_READ)) {
695 ptr += VKI_PAGE_SIZE; // Bad page - skip it.
696 continue;
697 }
sewardjb5f6f512005-03-10 23:59:00 +0000698 }
699
sewardj6c591e12011-04-11 16:17:51 +0000700 if (VG_MINIMAL_SETJMP(memscan_jmpbuf) == 0) {
njn8225cc02009-03-09 22:52:24 +0000701 if ( MC_(is_valid_aligned_word)(ptr) ) {
702 lc_scanned_szB += sizeof(Addr);
703 addr = *(Addr *)ptr;
704 // If we get here, the scanned word is in valid memory. Now
705 // let's see if its contents point to a chunk.
706 lc_push_if_a_chunk_ptr(addr, clique, is_prior_definite);
707 } else if (0 && VG_DEBUG_LEAKCHECK) {
708 VG_(printf)("%#lx not valid\n", ptr);
709 }
710 ptr += sizeof(Addr);
sewardjb5f6f512005-03-10 23:59:00 +0000711 } else {
njn8225cc02009-03-09 22:52:24 +0000712 // We need to restore the signal mask, because we were
713 // longjmped out of a signal handler.
714 VG_(sigprocmask)(VKI_SIG_SETMASK, &sigmask, NULL);
sewardjb5f6f512005-03-10 23:59:00 +0000715
njn8225cc02009-03-09 22:52:24 +0000716 ptr = VG_PGROUNDUP(ptr+1); // Bad page - skip it.
sewardjb5f6f512005-03-10 23:59:00 +0000717 }
718 }
719
720 VG_(sigprocmask)(VKI_SIG_SETMASK, &sigmask, NULL);
721 VG_(set_fault_catcher)(NULL);
722}
723
sewardj45d94cc2005-04-20 14:44:11 +0000724
njn8225cc02009-03-09 22:52:24 +0000725// Process the mark stack until empty.
726static void lc_process_markstack(Int clique)
sewardjb5f6f512005-03-10 23:59:00 +0000727{
njne3675d62009-05-19 02:08:25 +0000728 Int top = -1; // shut gcc up
njn8225cc02009-03-09 22:52:24 +0000729 Bool is_prior_definite;
sewardjb5f6f512005-03-10 23:59:00 +0000730
njn8225cc02009-03-09 22:52:24 +0000731 while (lc_pop(&top)) {
tom1d0f3f62010-10-04 20:55:21 +0000732 tl_assert(top >= 0 && top < lc_n_chunks);
sewardjb5f6f512005-03-10 23:59:00 +0000733
njn8225cc02009-03-09 22:52:24 +0000734 // See comment about 'is_prior_definite' at the top to understand this.
735 is_prior_definite = ( Possible != lc_extras[top].state );
sewardjb5f6f512005-03-10 23:59:00 +0000736
njn8225cc02009-03-09 22:52:24 +0000737 lc_scan_memory(lc_chunks[top]->data, lc_chunks[top]->szB,
738 is_prior_definite, clique);
sewardjb5f6f512005-03-10 23:59:00 +0000739 }
740}
741
njn29a5c012009-05-06 06:15:55 +0000742static Word cmp_LossRecordKey_LossRecord(const void* key, const void* elem)
743{
744 LossRecordKey* a = (LossRecordKey*)key;
745 LossRecordKey* b = &(((LossRecord*)elem)->key);
746
747 // Compare on states first because that's fast.
748 if (a->state < b->state) return -1;
749 if (a->state > b->state) return 1;
750 // Ok, the states are equal. Now compare the locations, which is slower.
751 if (VG_(eq_ExeContext)(
752 MC_(clo_leak_resolution), a->allocated_at, b->allocated_at))
753 return 0;
754 // Different locations. Ordering is arbitrary, just use the ec pointer.
755 if (a->allocated_at < b->allocated_at) return -1;
756 if (a->allocated_at > b->allocated_at) return 1;
757 VG_(tool_panic)("bad LossRecord comparison");
758}
759
760static Int cmp_LossRecords(void* va, void* vb)
761{
762 LossRecord* lr_a = *(LossRecord**)va;
763 LossRecord* lr_b = *(LossRecord**)vb;
764 SizeT total_szB_a = lr_a->szB + lr_a->indirect_szB;
765 SizeT total_szB_b = lr_b->szB + lr_b->indirect_szB;
766
767 // First compare by sizes.
768 if (total_szB_a < total_szB_b) return -1;
769 if (total_szB_a > total_szB_b) return 1;
770 // If size are equal, compare by states.
771 if (lr_a->key.state < lr_b->key.state) return -1;
772 if (lr_a->key.state > lr_b->key.state) return 1;
njne10c7f82009-05-06 06:52:47 +0000773 // If they're still equal here, it doesn't matter that much, but we keep
774 // comparing other things so that regtests are as deterministic as
775 // possible. So: compare num_blocks.
776 if (lr_a->num_blocks < lr_b->num_blocks) return -1;
777 if (lr_a->num_blocks > lr_b->num_blocks) return 1;
778 // Finally, compare ExeContext addresses... older ones are likely to have
779 // lower addresses.
780 if (lr_a->key.allocated_at < lr_b->key.allocated_at) return -1;
781 if (lr_a->key.allocated_at > lr_b->key.allocated_at) return 1;
njn29a5c012009-05-06 06:15:55 +0000782 return 0;
783}
784
philippe84234902012-01-14 13:53:13 +0000785
786static void get_printing_rules(LeakCheckParams* lcp,
787 LossRecord* lr,
788 Bool* count_as_error,
789 Bool* print_record)
sewardjb5f6f512005-03-10 23:59:00 +0000790{
philippe84234902012-01-14 13:53:13 +0000791 // Rules for printing:
792 // - We don't show suppressed loss records ever (and that's controlled
793 // within the error manager).
794 // - We show non-suppressed loss records that are not "reachable" if
795 // --leak-check=yes.
796 // - We show all non-suppressed loss records if --leak-check=yes and
797 // --show-reachable=yes.
798 //
799 // Nb: here "reachable" means Reachable *or* IndirectLeak; note that
800 // this is different to "still reachable" used elsewhere because it
801 // includes indirectly lost blocks!
802
803 Bool delta_considered;
804
805 switch (lcp->deltamode) {
806 case LCD_Any:
807 delta_considered = lr->num_blocks > 0;
808 break;
809 case LCD_Increased:
810 delta_considered
811 = lr->szB > lr->old_szB
812 || lr->indirect_szB > lr->old_indirect_szB
813 || lr->num_blocks > lr->old_num_blocks;
814 break;
815 case LCD_Changed:
816 delta_considered = lr->szB != lr->old_szB
817 || lr->indirect_szB != lr->old_indirect_szB
818 || lr->num_blocks != lr->old_num_blocks;
819 break;
820 default:
821 tl_assert(0);
822 }
823
824 *print_record = lcp->mode == LC_Full && delta_considered &&
825 ( lcp->show_reachable ||
826 Unreached == lr->key.state ||
827 ( lcp->show_possibly_lost &&
828 Possible == lr->key.state ) );
829 // We don't count a leaks as errors with lcp->mode==LC_Summary.
830 // Otherwise you can get high error counts with few or no error
831 // messages, which can be confusing. Also, you could argue that
832 // indirect leaks should be counted as errors, but it seems better to
833 // make the counting criteria similar to the printing criteria. So we
834 // don't count them.
835 *count_as_error = lcp->mode == LC_Full && delta_considered &&
836 ( Unreached == lr->key.state ||
837 Possible == lr->key.state );
838}
839
840static void print_results(ThreadId tid, LeakCheckParams* lcp)
841{
842 Int i, n_lossrecords, start_lr_output_scan;
njn29a5c012009-05-06 06:15:55 +0000843 LossRecord** lr_array;
844 LossRecord* lr;
845 Bool is_suppressed;
sewardjc8bd1df2011-06-26 12:41:33 +0000846 SizeT old_bytes_leaked = MC_(bytes_leaked); /* to report delta in summary */
847 SizeT old_bytes_indirect = MC_(bytes_indirect);
848 SizeT old_bytes_dubious = MC_(bytes_dubious);
849 SizeT old_bytes_reachable = MC_(bytes_reachable);
850 SizeT old_bytes_suppressed = MC_(bytes_suppressed);
851 SizeT old_blocks_leaked = MC_(blocks_leaked);
852 SizeT old_blocks_indirect = MC_(blocks_indirect);
853 SizeT old_blocks_dubious = MC_(blocks_dubious);
854 SizeT old_blocks_reachable = MC_(blocks_reachable);
855 SizeT old_blocks_suppressed = MC_(blocks_suppressed);
sewardjb5f6f512005-03-10 23:59:00 +0000856
sewardjc8bd1df2011-06-26 12:41:33 +0000857 if (lr_table == NULL)
858 // Create the lr_table, which holds the loss records.
859 // If the lr_table already exists, it means it contains
860 // loss_records from the previous leak search. The old_*
861 // values in these records are used to implement the
862 // leak check delta mode
863 lr_table =
864 VG_(OSetGen_Create)(offsetof(LossRecord, key),
865 cmp_LossRecordKey_LossRecord,
866 VG_(malloc), "mc.pr.1",
867 VG_(free));
868
njn29a5c012009-05-06 06:15:55 +0000869
870 // Convert the chunks into loss records, merging them where appropriate.
njn8225cc02009-03-09 22:52:24 +0000871 for (i = 0; i < lc_n_chunks; i++) {
njn29a5c012009-05-06 06:15:55 +0000872 MC_Chunk* ch = lc_chunks[i];
873 LC_Extra* ex = &(lc_extras)[i];
874 LossRecord* old_lr;
875 LossRecordKey lrkey;
876 lrkey.state = ex->state;
877 lrkey.allocated_at = ch->where;
sewardjb5f6f512005-03-10 23:59:00 +0000878
njn29a5c012009-05-06 06:15:55 +0000879 old_lr = VG_(OSetGen_Lookup)(lr_table, &lrkey);
880 if (old_lr) {
881 // We found an existing loss record matching this chunk. Update the
882 // loss record's details in-situ. This is safe because we don't
883 // change the elements used as the OSet key.
884 old_lr->szB += ch->szB;
885 old_lr->indirect_szB += ex->indirect_szB;
886 old_lr->num_blocks++;
sewardjb5f6f512005-03-10 23:59:00 +0000887 } else {
njn29a5c012009-05-06 06:15:55 +0000888 // No existing loss record matches this chunk. Create a new loss
889 // record, initialise it from the chunk, and insert it into lr_table.
890 lr = VG_(OSetGen_AllocNode)(lr_table, sizeof(LossRecord));
891 lr->key = lrkey;
892 lr->szB = ch->szB;
893 lr->indirect_szB = ex->indirect_szB;
894 lr->num_blocks = 1;
sewardjc8bd1df2011-06-26 12:41:33 +0000895 lr->old_szB = 0;
896 lr->old_indirect_szB = 0;
897 lr->old_num_blocks = 0;
njn29a5c012009-05-06 06:15:55 +0000898 VG_(OSetGen_Insert)(lr_table, lr);
sewardjb5f6f512005-03-10 23:59:00 +0000899 }
900 }
njn29a5c012009-05-06 06:15:55 +0000901 n_lossrecords = VG_(OSetGen_Size)(lr_table);
sewardjb5f6f512005-03-10 23:59:00 +0000902
njn29a5c012009-05-06 06:15:55 +0000903 // Create an array of pointers to the loss records.
904 lr_array = VG_(malloc)("mc.pr.2", n_lossrecords * sizeof(LossRecord*));
905 i = 0;
906 VG_(OSetGen_ResetIter)(lr_table);
907 while ( (lr = VG_(OSetGen_Next)(lr_table)) ) {
908 lr_array[i++] = lr;
909 }
910 tl_assert(i == n_lossrecords);
911
912 // Sort the array by loss record sizes.
913 VG_(ssort)(lr_array, n_lossrecords, sizeof(LossRecord*),
914 cmp_LossRecords);
915
916 // Zero totals.
njn8225cc02009-03-09 22:52:24 +0000917 MC_(blocks_leaked) = MC_(bytes_leaked) = 0;
918 MC_(blocks_indirect) = MC_(bytes_indirect) = 0;
919 MC_(blocks_dubious) = MC_(bytes_dubious) = 0;
920 MC_(blocks_reachable) = MC_(bytes_reachable) = 0;
921 MC_(blocks_suppressed) = MC_(bytes_suppressed) = 0;
922
philippe84234902012-01-14 13:53:13 +0000923 // If there is a maximum nr of loss records we can output, then first
924 // compute from where the output scan has to start.
925 // By default, start from the first loss record. Compute a higher
926 // value if there is a maximum to respect. We need to print the last
927 // records, as the one with the biggest sizes are more interesting.
928 start_lr_output_scan = 0;
929 if (lcp->mode == LC_Full && lcp->max_loss_records_output < n_lossrecords) {
930 Int nr_printable_records = 0;
931 for (i = n_lossrecords - 1; i >= 0 && start_lr_output_scan == 0; i--) {
932 Bool count_as_error, print_record;
933 lr = lr_array[i];
934 get_printing_rules (lcp, lr, &count_as_error, &print_record);
935 // Do not use get_printing_rules results for is_suppressed, as we
936 // only want to check if the record would be suppressed.
937 is_suppressed =
938 MC_(record_leak_error) ( tid, i+1, n_lossrecords, lr,
939 False /* print_record */,
940 False /* count_as_error */);
941 if (print_record && !is_suppressed) {
942 nr_printable_records++;
943 if (nr_printable_records == lcp->max_loss_records_output)
944 start_lr_output_scan = i;
945 }
sewardjc8bd1df2011-06-26 12:41:33 +0000946 }
philippe84234902012-01-14 13:53:13 +0000947 }
sewardjc8bd1df2011-06-26 12:41:33 +0000948
philippe84234902012-01-14 13:53:13 +0000949 // Print the loss records (in size order) and collect summary stats.
950 for (i = start_lr_output_scan; i < n_lossrecords; i++) {
951 Bool count_as_error, print_record;
952 lr = lr_array[i];
953 get_printing_rules(lcp, lr, &count_as_error, &print_record);
sewardjb5f6f512005-03-10 23:59:00 +0000954 is_suppressed =
njn18afe5d2009-08-10 08:25:39 +0000955 MC_(record_leak_error) ( tid, i+1, n_lossrecords, lr, print_record,
956 count_as_error );
sewardjb5f6f512005-03-10 23:59:00 +0000957
958 if (is_suppressed) {
njn29a5c012009-05-06 06:15:55 +0000959 MC_(blocks_suppressed) += lr->num_blocks;
960 MC_(bytes_suppressed) += lr->szB;
sewardjb5f6f512005-03-10 23:59:00 +0000961
njn29a5c012009-05-06 06:15:55 +0000962 } else if (Unreached == lr->key.state) {
963 MC_(blocks_leaked) += lr->num_blocks;
964 MC_(bytes_leaked) += lr->szB;
sewardjb5f6f512005-03-10 23:59:00 +0000965
njn29a5c012009-05-06 06:15:55 +0000966 } else if (IndirectLeak == lr->key.state) {
967 MC_(blocks_indirect) += lr->num_blocks;
968 MC_(bytes_indirect) += lr->szB;
sewardjb5f6f512005-03-10 23:59:00 +0000969
njn29a5c012009-05-06 06:15:55 +0000970 } else if (Possible == lr->key.state) {
971 MC_(blocks_dubious) += lr->num_blocks;
972 MC_(bytes_dubious) += lr->szB;
sewardjb5f6f512005-03-10 23:59:00 +0000973
njn29a5c012009-05-06 06:15:55 +0000974 } else if (Reachable == lr->key.state) {
975 MC_(blocks_reachable) += lr->num_blocks;
976 MC_(bytes_reachable) += lr->szB;
sewardjb5f6f512005-03-10 23:59:00 +0000977
978 } else {
njn8225cc02009-03-09 22:52:24 +0000979 VG_(tool_panic)("unknown loss mode");
sewardjb5f6f512005-03-10 23:59:00 +0000980 }
sewardjb5f6f512005-03-10 23:59:00 +0000981 }
sewardjb5f6f512005-03-10 23:59:00 +0000982
sewardjc8bd1df2011-06-26 12:41:33 +0000983 for (i = 0; i < n_lossrecords; i++)
984 {
985 if (lr->num_blocks == 0)
986 // remove from lr_table the old loss_records with 0 bytes found
987 VG_(OSetGen_Remove) (lr_table, &lr_array[i]->key);
988 else
989 {
990 // move the leak sizes to old_* and zero the current sizes
991 // for next leak search
992 lr_array[i]->old_szB = lr_array[i]->szB;
993 lr_array[i]->old_indirect_szB = lr_array[i]->indirect_szB;
994 lr_array[i]->old_num_blocks = lr_array[i]->num_blocks;
995 lr_array[i]->szB = 0;
996 lr_array[i]->indirect_szB = 0;
997 lr_array[i]->num_blocks = 0;
998 }
999 }
1000 VG_(free)(lr_array);
1001
njn8225cc02009-03-09 22:52:24 +00001002 if (VG_(clo_verbosity) > 0 && !VG_(clo_xml)) {
sewardjc8bd1df2011-06-26 12:41:33 +00001003 char d_bytes[20];
1004 char d_blocks[20];
1005
sewardj6b523cd2009-07-15 14:49:40 +00001006 VG_(umsg)("LEAK SUMMARY:\n");
sewardjc8bd1df2011-06-26 12:41:33 +00001007 VG_(umsg)(" definitely lost: %'lu%s bytes in %'lu%s blocks\n",
1008 MC_(bytes_leaked),
philippe84234902012-01-14 13:53:13 +00001009 MC_(snprintf_delta) (d_bytes, 20, MC_(bytes_leaked), old_bytes_leaked, lcp->deltamode),
sewardjc8bd1df2011-06-26 12:41:33 +00001010 MC_(blocks_leaked),
philippe84234902012-01-14 13:53:13 +00001011 MC_(snprintf_delta) (d_blocks, 20, MC_(blocks_leaked), old_blocks_leaked, lcp->deltamode));
sewardjc8bd1df2011-06-26 12:41:33 +00001012 VG_(umsg)(" indirectly lost: %'lu%s bytes in %'lu%s blocks\n",
1013 MC_(bytes_indirect),
philippe84234902012-01-14 13:53:13 +00001014 MC_(snprintf_delta) (d_bytes, 20, MC_(bytes_indirect), old_bytes_indirect, lcp->deltamode),
sewardjc8bd1df2011-06-26 12:41:33 +00001015 MC_(blocks_indirect),
philippe84234902012-01-14 13:53:13 +00001016 MC_(snprintf_delta) (d_blocks, 20, MC_(blocks_indirect), old_blocks_indirect, lcp->deltamode) );
sewardjc8bd1df2011-06-26 12:41:33 +00001017 VG_(umsg)(" possibly lost: %'lu%s bytes in %'lu%s blocks\n",
1018 MC_(bytes_dubious),
philippe84234902012-01-14 13:53:13 +00001019 MC_(snprintf_delta) (d_bytes, 20, MC_(bytes_dubious), old_bytes_dubious, lcp->deltamode),
sewardjc8bd1df2011-06-26 12:41:33 +00001020 MC_(blocks_dubious),
philippe84234902012-01-14 13:53:13 +00001021 MC_(snprintf_delta) (d_blocks, 20, MC_(blocks_dubious), old_blocks_dubious, lcp->deltamode) );
sewardjc8bd1df2011-06-26 12:41:33 +00001022 VG_(umsg)(" still reachable: %'lu%s bytes in %'lu%s blocks\n",
1023 MC_(bytes_reachable),
philippe84234902012-01-14 13:53:13 +00001024 MC_(snprintf_delta) (d_bytes, 20, MC_(bytes_reachable), old_bytes_reachable, lcp->deltamode),
sewardjc8bd1df2011-06-26 12:41:33 +00001025 MC_(blocks_reachable),
philippe84234902012-01-14 13:53:13 +00001026 MC_(snprintf_delta) (d_blocks, 20, MC_(blocks_reachable), old_blocks_reachable, lcp->deltamode) );
sewardjc8bd1df2011-06-26 12:41:33 +00001027 VG_(umsg)(" suppressed: %'lu%s bytes in %'lu%s blocks\n",
1028 MC_(bytes_suppressed),
philippe84234902012-01-14 13:53:13 +00001029 MC_(snprintf_delta) (d_bytes, 20, MC_(bytes_suppressed), old_bytes_suppressed, lcp->deltamode),
sewardjc8bd1df2011-06-26 12:41:33 +00001030 MC_(blocks_suppressed),
philippe84234902012-01-14 13:53:13 +00001031 MC_(snprintf_delta) (d_blocks, 20, MC_(blocks_suppressed), old_blocks_suppressed, lcp->deltamode) );
1032 if (lcp->mode != LC_Full &&
njn8225cc02009-03-09 22:52:24 +00001033 (MC_(blocks_leaked) + MC_(blocks_indirect) +
1034 MC_(blocks_dubious) + MC_(blocks_reachable)) > 0) {
philippe84234902012-01-14 13:53:13 +00001035 if (lcp->requested_by_monitor_command)
sewardj30b3eca2011-06-28 08:20:39 +00001036 VG_(umsg)("To see details of leaked memory, give 'full' arg to leak_check\n");
sewardjc8bd1df2011-06-26 12:41:33 +00001037 else
1038 VG_(umsg)("Rerun with --leak-check=full to see details "
1039 "of leaked memory\n");
njn8225cc02009-03-09 22:52:24 +00001040 }
philippe84234902012-01-14 13:53:13 +00001041 if (lcp->mode == LC_Full &&
1042 MC_(blocks_reachable) > 0 && !lcp->show_reachable)
njn8225cc02009-03-09 22:52:24 +00001043 {
sewardj6b523cd2009-07-15 14:49:40 +00001044 VG_(umsg)("Reachable blocks (those to which a pointer "
1045 "was found) are not shown.\n");
philippe84234902012-01-14 13:53:13 +00001046 if (lcp->requested_by_monitor_command)
sewardj30b3eca2011-06-28 08:20:39 +00001047 VG_(umsg)("To see them, add 'reachable any' args to leak_check\n");
sewardjc8bd1df2011-06-26 12:41:33 +00001048 else
1049 VG_(umsg)("To see them, rerun with: --leak-check=full "
1050 "--show-reachable=yes\n");
sewardjb5f6f512005-03-10 23:59:00 +00001051 }
njnb6267bd2009-08-12 00:14:16 +00001052 VG_(umsg)("\n");
sewardjb5f6f512005-03-10 23:59:00 +00001053 }
1054}
1055
njn8225cc02009-03-09 22:52:24 +00001056/*------------------------------------------------------------*/
1057/*--- Top-level entry point. ---*/
1058/*------------------------------------------------------------*/
sewardj3cf26a52006-07-27 23:48:53 +00001059
philippe84234902012-01-14 13:53:13 +00001060void MC_(detect_memory_leaks) ( ThreadId tid, LeakCheckParams* lcp)
njn43c799e2003-04-08 00:08:52 +00001061{
njnb965efb2009-08-10 07:36:54 +00001062 Int i, j;
njn43c799e2003-04-08 00:08:52 +00001063
philippe84234902012-01-14 13:53:13 +00001064 tl_assert(lcp->mode != LC_Off);
sewardjc8bd1df2011-06-26 12:41:33 +00001065
philippe84234902012-01-14 13:53:13 +00001066 MC_(detect_memory_leaks_last_delta_mode) = lcp->deltamode;
njn43c799e2003-04-08 00:08:52 +00001067
njn8225cc02009-03-09 22:52:24 +00001068 // Get the chunks, stop if there were none.
1069 lc_chunks = find_active_chunks(&lc_n_chunks);
1070 if (lc_n_chunks == 0) {
1071 tl_assert(lc_chunks == NULL);
sewardjc8bd1df2011-06-26 12:41:33 +00001072 if (lr_table != NULL) {
1073 // forget the previous recorded LossRecords as next leak search will in any case
1074 // just create new leaks.
1075 // Maybe it would be better to rather call print_result ?
1076 // (at least when leak decrease are requested)
1077 // This will then output all LossRecords with a size decreasing to 0
1078 VG_(OSetGen_Destroy) (lr_table);
1079 }
sewardj71bc3cb2005-05-19 00:25:45 +00001080 if (VG_(clo_verbosity) >= 1 && !VG_(clo_xml)) {
njnb6267bd2009-08-12 00:14:16 +00001081 VG_(umsg)("All heap blocks were freed -- no leaks are possible\n");
sewardj2d9e8742009-08-07 15:46:56 +00001082 VG_(umsg)("\n");
sewardj37d06f22003-09-17 21:48:26 +00001083 }
njn43c799e2003-04-08 00:08:52 +00001084 return;
1085 }
1086
njn8225cc02009-03-09 22:52:24 +00001087 // Sort the array so blocks are in ascending order in memory.
1088 VG_(ssort)(lc_chunks, lc_n_chunks, sizeof(VgHashNode*), compare_MC_Chunks);
njn43c799e2003-04-08 00:08:52 +00001089
njn8225cc02009-03-09 22:52:24 +00001090 // Sanity check -- make sure they're in order.
1091 for (i = 0; i < lc_n_chunks-1; i++) {
1092 tl_assert( lc_chunks[i]->data <= lc_chunks[i+1]->data);
1093 }
njn43c799e2003-04-08 00:08:52 +00001094
njnb965efb2009-08-10 07:36:54 +00001095 // Sanity check -- make sure they don't overlap. The one exception is that
1096 // we allow a MALLOCLIKE block to sit entirely within a malloc() block.
1097 // This is for bug 100628. If this occurs, we ignore the malloc() block
1098 // for leak-checking purposes. This is a hack and probably should be done
1099 // better, but at least it's consistent with mempools (which are treated
1100 // like this in find_active_chunks). Mempools have a separate VgHashTable
1101 // for mempool chunks, but if custom-allocated blocks are put in a separate
1102 // table from normal heap blocks it makes free-mismatch checking more
1103 // difficult.
1104 //
1105 // If this check fails, it probably means that the application
njn8225cc02009-03-09 22:52:24 +00001106 // has done something stupid with VALGRIND_MALLOCLIKE_BLOCK client
njnb965efb2009-08-10 07:36:54 +00001107 // requests, eg. has made overlapping requests (which are
1108 // nonsensical), or used VALGRIND_MALLOCLIKE_BLOCK for stack locations;
1109 // again nonsensical.
1110 //
njn8225cc02009-03-09 22:52:24 +00001111 for (i = 0; i < lc_n_chunks-1; i++) {
1112 MC_Chunk* ch1 = lc_chunks[i];
1113 MC_Chunk* ch2 = lc_chunks[i+1];
njnb965efb2009-08-10 07:36:54 +00001114
1115 Addr start1 = ch1->data;
1116 Addr start2 = ch2->data;
1117 Addr end1 = ch1->data + ch1->szB - 1;
1118 Addr end2 = ch2->data + ch2->szB - 1;
1119 Bool isCustom1 = ch1->allockind == MC_AllocCustom;
1120 Bool isCustom2 = ch2->allockind == MC_AllocCustom;
1121
1122 if (end1 < start2) {
1123 // Normal case - no overlap.
1124
1125 // We used to allow exact duplicates, I'm not sure why. --njn
1126 //} else if (start1 == start2 && end1 == end2) {
1127 // Degenerate case: exact duplicates.
1128
1129 } else if (start1 >= start2 && end1 <= end2 && isCustom1 && !isCustom2) {
1130 // Block i is MALLOCLIKE and entirely within block i+1.
1131 // Remove block i+1.
1132 for (j = i+1; j < lc_n_chunks-1; j++) {
1133 lc_chunks[j] = lc_chunks[j+1];
1134 }
1135 lc_n_chunks--;
1136
1137 } else if (start2 >= start1 && end2 <= end1 && isCustom2 && !isCustom1) {
1138 // Block i+1 is MALLOCLIKE and entirely within block i.
1139 // Remove block i.
1140 for (j = i; j < lc_n_chunks-1; j++) {
1141 lc_chunks[j] = lc_chunks[j+1];
1142 }
1143 lc_n_chunks--;
1144
1145 } else {
1146 VG_(umsg)("Block 0x%lx..0x%lx overlaps with block 0x%lx..0x%lx",
bart3c4fa9f2011-05-09 10:46:55 +00001147 start1, end1, start2, end2);
njnb965efb2009-08-10 07:36:54 +00001148 VG_(umsg)("This is usually caused by using VALGRIND_MALLOCLIKE_BLOCK");
1149 VG_(umsg)("in an inappropriate way.");
1150 tl_assert (0);
njn8225cc02009-03-09 22:52:24 +00001151 }
njn8225cc02009-03-09 22:52:24 +00001152 }
1153
1154 // Initialise lc_extras.
1155 lc_extras = VG_(malloc)( "mc.dml.2", lc_n_chunks * sizeof(LC_Extra) );
1156 for (i = 0; i < lc_n_chunks; i++) {
1157 lc_extras[i].state = Unreached;
tom1d0f3f62010-10-04 20:55:21 +00001158 lc_extras[i].pending = False;
njn8225cc02009-03-09 22:52:24 +00001159 lc_extras[i].indirect_szB = 0;
1160 }
1161
1162 // Initialise lc_markstack.
1163 lc_markstack = VG_(malloc)( "mc.dml.2", lc_n_chunks * sizeof(Int) );
1164 for (i = 0; i < lc_n_chunks; i++) {
1165 lc_markstack[i] = -1;
sewardjb5f6f512005-03-10 23:59:00 +00001166 }
1167 lc_markstack_top = -1;
njn43c799e2003-04-08 00:08:52 +00001168
njn8225cc02009-03-09 22:52:24 +00001169 // Verbosity.
sewardj2d9e8742009-08-07 15:46:56 +00001170 if (VG_(clo_verbosity) > 1 && !VG_(clo_xml)) {
njnb6267bd2009-08-12 00:14:16 +00001171 VG_(umsg)( "Searching for pointers to %'d not-freed blocks\n",
sewardj6b523cd2009-07-15 14:49:40 +00001172 lc_n_chunks );
sewardj2d9e8742009-08-07 15:46:56 +00001173 }
sewardjb5f6f512005-03-10 23:59:00 +00001174
njn8225cc02009-03-09 22:52:24 +00001175 // Scan the memory root-set, pushing onto the mark stack any blocks
1176 // pointed to.
1177 {
1178 Int n_seg_starts;
njnac1e0332009-05-08 00:39:31 +00001179 Addr* seg_starts = VG_(get_segment_starts)( &n_seg_starts );
sewardjb5f6f512005-03-10 23:59:00 +00001180
njn8225cc02009-03-09 22:52:24 +00001181 tl_assert(seg_starts && n_seg_starts > 0);
sewardjde3ad732006-07-27 23:12:17 +00001182
njn8225cc02009-03-09 22:52:24 +00001183 lc_scanned_szB = 0;
sewardjde3ad732006-07-27 23:12:17 +00001184
njn8225cc02009-03-09 22:52:24 +00001185 // VG_(am_show_nsegments)( 0, "leakcheck");
1186 for (i = 0; i < n_seg_starts; i++) {
1187 SizeT seg_size;
1188 NSegment const* seg = VG_(am_find_nsegment)( seg_starts[i] );
1189 tl_assert(seg);
1190
1191 if (seg->kind != SkFileC && seg->kind != SkAnonC) continue;
1192 if (!(seg->hasR && seg->hasW)) continue;
1193 if (seg->isCH) continue;
1194
1195 // Don't poke around in device segments as this may cause
1196 // hangs. Exclude /dev/zero just in case someone allocated
1197 // memory by explicitly mapping /dev/zero.
1198 if (seg->kind == SkFileC
1199 && (VKI_S_ISCHR(seg->mode) || VKI_S_ISBLK(seg->mode))) {
1200 HChar* dev_name = VG_(am_get_filename)( (NSegment*)seg );
1201 if (dev_name && 0 == VG_(strcmp)(dev_name, "/dev/zero")) {
1202 // Don't skip /dev/zero.
1203 } else {
1204 // Skip this device mapping.
1205 continue;
1206 }
1207 }
1208
1209 if (0)
1210 VG_(printf)("ACCEPT %2d %#lx %#lx\n", i, seg->start, seg->end);
1211
1212 // Scan the segment. We use -1 for the clique number, because this
1213 // is a root-set.
1214 seg_size = seg->end - seg->start + 1;
1215 if (VG_(clo_verbosity) > 2) {
1216 VG_(message)(Vg_DebugMsg,
sewardj6b523cd2009-07-15 14:49:40 +00001217 " Scanning root segment: %#lx..%#lx (%lu)\n",
njn8225cc02009-03-09 22:52:24 +00001218 seg->start, seg->end, seg_size);
1219 }
1220 lc_scan_memory(seg->start, seg_size, /*is_prior_definite*/True, -1);
1221 }
sewardj45f4e7c2005-09-27 19:20:21 +00001222 }
sewardjb5f6f512005-03-10 23:59:00 +00001223
njn8225cc02009-03-09 22:52:24 +00001224 // Scan GP registers for chunk pointers.
1225 VG_(apply_to_GP_regs)(lc_push_if_a_chunk_ptr_register);
sewardjb5f6f512005-03-10 23:59:00 +00001226
njn8225cc02009-03-09 22:52:24 +00001227 // Process the pushed blocks. After this, every block that is reachable
1228 // from the root-set has been traced.
1229 lc_process_markstack(/*clique*/-1);
njn43c799e2003-04-08 00:08:52 +00001230
njnb6267bd2009-08-12 00:14:16 +00001231 if (VG_(clo_verbosity) > 1 && !VG_(clo_xml)) {
1232 VG_(umsg)("Checked %'lu bytes\n", lc_scanned_szB);
1233 VG_(umsg)( "\n" );
1234 }
njn43c799e2003-04-08 00:08:52 +00001235
njn8225cc02009-03-09 22:52:24 +00001236 // Trace all the leaked blocks to determine which are directly leaked and
1237 // which are indirectly leaked. For each Unreached block, push it onto
1238 // the mark stack, and find all the as-yet-Unreached blocks reachable
1239 // from it. These form a clique and are marked IndirectLeak, and their
1240 // size is added to the clique leader's indirect size. If one of the
1241 // found blocks was itself a clique leader (from a previous clique), then
1242 // the cliques are merged.
1243 for (i = 0; i < lc_n_chunks; i++) {
1244 MC_Chunk* ch = lc_chunks[i];
1245 LC_Extra* ex = &(lc_extras[i]);
njn43c799e2003-04-08 00:08:52 +00001246
njn8225cc02009-03-09 22:52:24 +00001247 if (VG_DEBUG_CLIQUE)
1248 VG_(printf)("cliques: %d at %#lx -> Loss state %d\n",
1249 i, ch->data, ex->state);
njn43c799e2003-04-08 00:08:52 +00001250
njn8225cc02009-03-09 22:52:24 +00001251 tl_assert(lc_markstack_top == -1);
1252
1253 if (ex->state == Unreached) {
1254 if (VG_DEBUG_CLIQUE)
1255 VG_(printf)("%d: gathering clique %#lx\n", i, ch->data);
1256
1257 // Push this Unreached block onto the stack and process it.
1258 lc_push(i, ch);
1259 lc_process_markstack(i);
1260
1261 tl_assert(lc_markstack_top == -1);
1262 tl_assert(ex->state == Unreached);
nethercote0f19bce2003-12-02 10:17:44 +00001263 }
njn43c799e2003-04-08 00:08:52 +00001264 }
njn8225cc02009-03-09 22:52:24 +00001265
sewardjc8bd1df2011-06-26 12:41:33 +00001266 print_results( tid, lcp);
njn43c799e2003-04-08 00:08:52 +00001267
njn8225cc02009-03-09 22:52:24 +00001268 VG_(free) ( lc_chunks );
1269 VG_(free) ( lc_extras );
sewardjb5f6f512005-03-10 23:59:00 +00001270 VG_(free) ( lc_markstack );
njn43c799e2003-04-08 00:08:52 +00001271}
1272
1273/*--------------------------------------------------------------------*/
njn1d0825f2006-03-27 11:37:07 +00001274/*--- end ---*/
njn43c799e2003-04-08 00:08:52 +00001275/*--------------------------------------------------------------------*/
1276