• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //--------------------------------------------------------------------*/
2 //--- Massif: a heap profiling tool.                     ms_main.c ---*/
3 //--------------------------------------------------------------------*/
4 
5 /*
6    This file is part of Massif, a Valgrind tool for profiling memory
7    usage of programs.
8 
9    Copyright (C) 2003-2013 Nicholas Nethercote
10       njn@valgrind.org
11 
12    This program is free software; you can redistribute it and/or
13    modify it under the terms of the GNU General Public License as
14    published by the Free Software Foundation; either version 2 of the
15    License, or (at your option) any later version.
16 
17    This program is distributed in the hope that it will be useful, but
18    WITHOUT ANY WARRANTY; without even the implied warranty of
19    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
20    General Public License for more details.
21 
22    You should have received a copy of the GNU General Public License
23    along with this program; if not, write to the Free Software
24    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
25    02111-1307, USA.
26 
27    The GNU General Public License is contained in the file COPYING.
28 */
29 
30 //---------------------------------------------------------------------------
31 // XXX:
32 //---------------------------------------------------------------------------
33 // Todo -- nice, but less critical:
34 // - do a graph-drawing test
35 // - make file format more generic.  Obstacles:
36 //   - unit prefixes are not generic
37 //   - preset column widths for stats are not generic
38 //   - preset column headers are not generic
39 //   - "Massif arguments:" line is not generic
40 // - do snapshots on client requests
41 //   - (Michael Meeks): have an interactive way to request a dump
42 //     (callgrind_control-style)
43 //     - "profile now"
44 //     - "show me the extra allocations since the last snapshot"
45 //     - "start/stop logging" (eg. quickly skip boring bits)
46 // - Add ability to draw multiple graphs, eg. heap-only, stack-only, total.
47 //   Give each graph a title.  (try to do it generically!)
48 // - allow truncation of long fnnames if the exact line number is
49 //   identified?  [hmm, could make getting the name of alloc-fns more
50 //   difficult] [could dump full names to file, truncate in ms_print]
51 // - make --show-below-main=no work
52 // - Options like --alloc-fn='operator new(unsigned, std::nothrow_t const&)'
53 //   don't work in a .valgrindrc file or in $VALGRIND_OPTS.
54 //   m_commandline.c:add_args_from_string() needs to respect single quotes.
55 // - With --stack=yes, want to add a stack trace for detailed snapshots so
56 //   it's clear where/why the peak is occurring. (Mattieu Castet)  Also,
57 //   possibly useful even with --stack=no? (Andi Yin)
58 //
59 // Performance:
60 // - To run the benchmarks:
61 //
62 //     perl perf/vg_perf --tools=massif --reps=3 perf/{heap,tinycc} massif
63 //     time valgrind --tool=massif --depth=100 konqueror
64 //
65 //   The other benchmarks don't do much allocation, and so give similar speeds
66 //   to Nulgrind.
67 //
68 //   Timing results on 'nevermore' (njn's machine) as of r7013:
69 //
70 //     heap      0.53s  ma:12.4s (23.5x, -----)
71 //     tinycc    0.46s  ma: 4.9s (10.7x, -----)
72 //     many-xpts 0.08s  ma: 2.0s (25.0x, -----)
73 //     konqueror 29.6s real  0:21.0s user
74 //
75 //   [Introduction of --time-unit=i as the default slowed things down by
76 //   roughly 0--20%.]
77 //
78 // - get_XCon accounts for about 9% of konqueror startup time.  Try
79 //   keeping XPt children sorted by 'ip' and use binary search in get_XCon.
80 //   Requires factoring out binary search code from various places into a
81 //   VG_(bsearch) function.
82 //
83 // Todo -- low priority:
84 // - In each XPt, record both bytes and the number of allocations, and
85 //   possibly the global number of allocations.
86 // - (Andy Lin) Give a stack trace on detailed snapshots?
87 // - (Artur Wisz) add a feature to Massif to ignore any heap blocks larger
88 //   than a certain size!  Because: "linux's malloc allows to set a
89 //   MMAP_THRESHOLD value, so we set it to 4096 - all blocks above that will
90 //   be handled directly by the kernel, and are guaranteed to be returned to
91 //   the system when freed. So we needed to profile only blocks below this
92 //   limit."
93 //
94 // File format working notes:
95 
96 #if 0
97 desc: --heap-admin=foo
98 cmd: date
99 time_unit: ms
100 #-----------
101 snapshot=0
102 #-----------
103 time=0
104 mem_heap_B=0
105 mem_heap_admin_B=0
106 mem_stacks_B=0
107 heap_tree=empty
108 #-----------
109 snapshot=1
110 #-----------
111 time=353
112 mem_heap_B=5
113 mem_heap_admin_B=0
114 mem_stacks_B=0
115 heap_tree=detailed
116 n1: 5 (heap allocation functions) malloc/new/new[], --alloc-fns, etc.
117  n1: 5 0x27F6E0: _nl_normalize_codeset (in /lib/libc-2.3.5.so)
118   n1: 5 0x279DE6: _nl_load_locale_from_archive (in /lib/libc-2.3.5.so)
119    n1: 5 0x278E97: _nl_find_locale (in /lib/libc-2.3.5.so)
120     n1: 5 0x278871: setlocale (in /lib/libc-2.3.5.so)
121      n1: 5 0x8049821: (within /bin/date)
122       n0: 5 0x26ED5E: (below main) (in /lib/libc-2.3.5.so)
123 
124 
125 n_events: n  time(ms)  total(B)    useful-heap(B)  admin-heap(B)  stacks(B)
126 t_events: B
127 n 0 0 0 0 0
128 n 0 0 0 0 0
129 t1: 5 <string...>
130  t1: 6 <string...>
131 
132 Ideas:
133 - each snapshot specifies an x-axis value and one or more y-axis values.
134 - can display the y-axis values separately if you like
135 - can completely separate connection between snapshots and trees.
136 
137 Challenges:
138 - how to specify and scale/abbreviate units on axes?
139 - how to combine multiple values into the y-axis?
140 
141 --------------------------------------------------------------------------------Command:            date
142 Massif arguments:   --heap-admin=foo
143 ms_print arguments: massif.out
144 --------------------------------------------------------------------------------
145     KB
146 6.472^                                                       :#
147      |                                                       :#  ::  .    .
148      ...
149      |                                     ::@  :@    :@ :@:::#  ::  :    ::::
150    0 +-----------------------------------@---@---@-----@--@---#-------------->ms     0                                                                     713
151 
152 Number of snapshots: 50
153  Detailed snapshots: [2, 11, 13, 19, 25, 32 (peak)]
154 --------------------------------------------------------------------------------  n       time(ms)         total(B)   useful-heap(B) admin-heap(B)    stacks(B)
155 --------------------------------------------------------------------------------  0              0                0                0             0            0
156   1            345                5                5             0            0
157   2            353                5                5             0            0
158 100.00% (5B) (heap allocation functions) malloc/new/new[], --alloc-fns, etc.
159 ->100.00% (5B) 0x27F6E0: _nl_normalize_codeset (in /lib/libc-2.3.5.so)
160 #endif
161 
162 //---------------------------------------------------------------------------
163 
164 #include "pub_tool_basics.h"
165 #include "pub_tool_vki.h"
166 #include "pub_tool_aspacemgr.h"
167 #include "pub_tool_debuginfo.h"
168 #include "pub_tool_hashtable.h"
169 #include "pub_tool_libcbase.h"
170 #include "pub_tool_libcassert.h"
171 #include "pub_tool_libcfile.h"
172 #include "pub_tool_libcprint.h"
173 #include "pub_tool_libcproc.h"
174 #include "pub_tool_machine.h"
175 #include "pub_tool_mallocfree.h"
176 #include "pub_tool_options.h"
177 #include "pub_tool_replacemalloc.h"
178 #include "pub_tool_stacktrace.h"
179 #include "pub_tool_threadstate.h"
180 #include "pub_tool_tooliface.h"
181 #include "pub_tool_xarray.h"
182 #include "pub_tool_clientstate.h"
183 #include "pub_tool_gdbserver.h"
184 
185 #include "pub_tool_clreq.h"           // For {MALLOC,FREE}LIKE_BLOCK
186 
187 //------------------------------------------------------------*/
188 //--- Overview of operation                                ---*/
189 //------------------------------------------------------------*/
190 
191 // The size of the stacks and heap is tracked.  The heap is tracked in a lot
192 // of detail, enough to tell how many bytes each line of code is responsible
193 // for, more or less.  The main data structure is a tree representing the
194 // call tree beneath all the allocation functions like malloc().
195 // (Alternatively, if --pages-as-heap=yes is specified, memory is tracked at
196 // the page level, and each page is treated much like a heap block.  We use
197 // "heap" throughout below to cover this case because the concepts are all the
198 // same.)
199 //
200 // "Snapshots" are recordings of the memory usage.  There are two basic
201 // kinds:
202 // - Normal:  these record the current time, total memory size, total heap
203 //   size, heap admin size and stack size.
204 // - Detailed: these record those things in a normal snapshot, plus a very
205 //   detailed XTree (see below) indicating how the heap is structured.
206 //
207 // Snapshots are taken every so often.  There are two storage classes of
208 // snapshots:
209 // - Temporary:  Massif does a temporary snapshot every so often.  The idea
210 //   is to always have a certain number of temporary snapshots around.  So
211 //   we take them frequently to begin with, but decreasingly often as the
212 //   program continues to run.  Also, we remove some old ones after a while.
213 //   Overall it's a kind of exponential decay thing.  Most of these are
214 //   normal snapshots, a small fraction are detailed snapshots.
215 // - Permanent:  Massif takes a permanent (detailed) snapshot in some
216 //   circumstances.  They are:
217 //   - Peak snapshot:  When the memory usage peak is reached, it takes a
218 //     snapshot.  It keeps this, unless the peak is subsequently exceeded,
219 //     in which case it will overwrite the peak snapshot.
220 //   - User-requested snapshots:  These are done in response to client
221 //     requests.  They are always kept.
222 
223 // Used for printing things when clo_verbosity > 1.
224 #define VERB(verb, format, args...) \
225    if (VG_(clo_verbosity) > verb) { \
226       VG_(dmsg)("Massif: " format, ##args); \
227    }
228 
229 //------------------------------------------------------------//
230 //--- Statistics                                           ---//
231 //------------------------------------------------------------//
232 
233 // Konqueror startup, to give an idea of the numbers involved with a biggish
234 // program, with default depth:
235 //
236 //  depth=3                   depth=40
237 //  - 310,000 allocations
238 //  - 300,000 frees
239 //  -  15,000 XPts            800,000 XPts
240 //  -   1,800 top-XPts
241 
242 static UInt n_heap_allocs           = 0;
243 static UInt n_heap_reallocs         = 0;
244 static UInt n_heap_frees            = 0;
245 static UInt n_ignored_heap_allocs   = 0;
246 static UInt n_ignored_heap_frees    = 0;
247 static UInt n_ignored_heap_reallocs = 0;
248 static UInt n_stack_allocs          = 0;
249 static UInt n_stack_frees           = 0;
250 static UInt n_xpts                  = 0;
251 static UInt n_xpt_init_expansions   = 0;
252 static UInt n_xpt_later_expansions  = 0;
253 static UInt n_sxpt_allocs           = 0;
254 static UInt n_sxpt_frees            = 0;
255 static UInt n_skipped_snapshots     = 0;
256 static UInt n_real_snapshots        = 0;
257 static UInt n_detailed_snapshots    = 0;
258 static UInt n_peak_snapshots        = 0;
259 static UInt n_cullings              = 0;
260 static UInt n_XCon_redos            = 0;
261 
262 //------------------------------------------------------------//
263 //--- Globals                                              ---//
264 //------------------------------------------------------------//
265 
266 // Number of guest instructions executed so far.  Only used with
267 // --time-unit=i.
268 static Long guest_instrs_executed = 0;
269 
270 static SizeT heap_szB       = 0; // Live heap size
271 static SizeT heap_extra_szB = 0; // Live heap extra size -- slop + admin bytes
272 static SizeT stacks_szB     = 0; // Live stacks size
273 
274 // This is the total size from the current peak snapshot, or 0 if no peak
275 // snapshot has been taken yet.
276 static SizeT peak_snapshot_total_szB = 0;
277 
278 // Incremented every time memory is allocated/deallocated, by the
279 // allocated/deallocated amount;  includes heap, heap-admin and stack
280 // memory.  An alternative to milliseconds as a unit of program "time".
281 static ULong total_allocs_deallocs_szB = 0;
282 
283 // When running with --heap=yes --pages-as-heap=no, we don't start taking
284 // snapshots until the first basic block is executed, rather than doing it in
285 // ms_post_clo_init (which is the obvious spot), for two reasons.
286 // - It lets us ignore stack events prior to that, because they're not
287 //   really proper ones and just would screw things up.
288 // - Because there's still some core initialisation to do, and so there
289 //   would be an artificial time gap between the first and second snapshots.
290 //
291 // When running with --heap=yes --pages-as-heap=yes, snapshots start much
292 // earlier due to new_mem_startup so this isn't relevant.
293 //
294 static Bool have_started_executing_code = False;
295 
296 //------------------------------------------------------------//
297 //--- Alloc fns                                            ---//
298 //------------------------------------------------------------//
299 
300 static XArray* alloc_fns;
301 static XArray* ignore_fns;
302 
init_alloc_fns(void)303 static void init_alloc_fns(void)
304 {
305    // Create the list, and add the default elements.
306    alloc_fns = VG_(newXA)(VG_(malloc), "ms.main.iaf.1",
307                                        VG_(free), sizeof(HChar*));
308    #define DO(x)  { const HChar* s = x; VG_(addToXA)(alloc_fns, &s); }
309 
310    // Ordered roughly according to (presumed) frequency.
311    // Nb: The C++ "operator new*" ones are overloadable.  We include them
312    // always anyway, because even if they're overloaded, it would be a
313    // prodigiously stupid overloading that caused them to not allocate
314    // memory.
315    //
316    // XXX: because we don't look at the first stack entry (unless it's a
317    // custom allocation) there's not much point to having all these alloc
318    // functions here -- they should never appear anywhere (I think?) other
319    // than the top stack entry.  The only exceptions are those that in
320    // vg_replace_malloc.c are partly or fully implemented in terms of another
321    // alloc function: realloc (which uses malloc);  valloc,
322    // malloc_zone_valloc, posix_memalign and memalign_common (which use
323    // memalign).
324    //
325    DO("malloc"                                              );
326    DO("__builtin_new"                                       );
327    DO("operator new(unsigned)"                              );
328    DO("operator new(unsigned long)"                         );
329    DO("__builtin_vec_new"                                   );
330    DO("operator new[](unsigned)"                            );
331    DO("operator new[](unsigned long)"                       );
332    DO("calloc"                                              );
333    DO("realloc"                                             );
334    DO("memalign"                                            );
335    DO("posix_memalign"                                      );
336    DO("valloc"                                              );
337    DO("operator new(unsigned, std::nothrow_t const&)"       );
338    DO("operator new[](unsigned, std::nothrow_t const&)"     );
339    DO("operator new(unsigned long, std::nothrow_t const&)"  );
340    DO("operator new[](unsigned long, std::nothrow_t const&)");
341 #if defined(VGO_darwin)
342    DO("malloc_zone_malloc"                                  );
343    DO("malloc_zone_calloc"                                  );
344    DO("malloc_zone_realloc"                                 );
345    DO("malloc_zone_memalign"                                );
346    DO("malloc_zone_valloc"                                  );
347 #endif
348 }
349 
init_ignore_fns(void)350 static void init_ignore_fns(void)
351 {
352    // Create the (empty) list.
353    ignore_fns = VG_(newXA)(VG_(malloc), "ms.main.iif.1",
354                                         VG_(free), sizeof(HChar*));
355 }
356 
357 // Determines if the named function is a member of the XArray.
is_member_fn(XArray * fns,const HChar * fnname)358 static Bool is_member_fn(XArray* fns, const HChar* fnname)
359 {
360    HChar** fn_ptr;
361    Int i;
362 
363    // Nb: It's a linear search through the list, because we're comparing
364    // strings rather than pointers to strings.
365    // Nb: This gets called a lot.  It was an OSet, but they're quite slow to
366    // iterate through so it wasn't a good choice.
367    for (i = 0; i < VG_(sizeXA)(fns); i++) {
368       fn_ptr = VG_(indexXA)(fns, i);
369       if (VG_STREQ(fnname, *fn_ptr))
370          return True;
371    }
372    return False;
373 }
374 
375 
376 //------------------------------------------------------------//
377 //--- Command line args                                    ---//
378 //------------------------------------------------------------//
379 
380 #define MAX_DEPTH       200
381 
382 typedef enum { TimeI, TimeMS, TimeB } TimeUnit;
383 
TimeUnit_to_string(TimeUnit time_unit)384 static const HChar* TimeUnit_to_string(TimeUnit time_unit)
385 {
386    switch (time_unit) {
387    case TimeI:  return "i";
388    case TimeMS: return "ms";
389    case TimeB:  return "B";
390    default:     tl_assert2(0, "TimeUnit_to_string: unrecognised TimeUnit");
391    }
392 }
393 
394 static Bool   clo_heap            = True;
395    // clo_heap_admin is deliberately a word-sized type.  At one point it was
396    // a UInt, but this caused problems on 64-bit machines when it was
397    // multiplied by a small negative number and then promoted to a
398    // word-sized type -- it ended up with a value of 4.2 billion.  Sigh.
399 static SSizeT clo_heap_admin      = 8;
400 static Bool   clo_pages_as_heap   = False;
401 static Bool   clo_stacks          = False;
402 static Int    clo_depth           = 30;
403 static double clo_threshold       = 1.0;  // percentage
404 static double clo_peak_inaccuracy = 1.0;  // percentage
405 static Int    clo_time_unit       = TimeI;
406 static Int    clo_detailed_freq   = 10;
407 static Int    clo_max_snapshots   = 100;
408 static const HChar* clo_massif_out_file = "massif.out.%p";
409 
410 static XArray* args_for_massif;
411 
ms_process_cmd_line_option(const HChar * arg)412 static Bool ms_process_cmd_line_option(const HChar* arg)
413 {
414    const HChar* tmp_str;
415 
416    // Remember the arg for later use.
417    VG_(addToXA)(args_for_massif, &arg);
418 
419         if VG_BOOL_CLO(arg, "--heap",           clo_heap)   {}
420    else if VG_BINT_CLO(arg, "--heap-admin",     clo_heap_admin, 0, 1024) {}
421 
422    else if VG_BOOL_CLO(arg, "--stacks",         clo_stacks) {}
423 
424    else if VG_BOOL_CLO(arg, "--pages-as-heap",  clo_pages_as_heap) {}
425 
426    else if VG_BINT_CLO(arg, "--depth",          clo_depth, 1, MAX_DEPTH) {}
427 
428    else if VG_STR_CLO(arg, "--alloc-fn",        tmp_str) {
429       VG_(addToXA)(alloc_fns, &tmp_str);
430    }
431    else if VG_STR_CLO(arg, "--ignore-fn",       tmp_str) {
432       VG_(addToXA)(ignore_fns, &tmp_str);
433    }
434 
435    else if VG_DBL_CLO(arg, "--threshold",  clo_threshold) {
436       if (clo_threshold < 0 || clo_threshold > 100) {
437          VG_(fmsg_bad_option)(arg,
438             "--threshold must be between 0.0 and 100.0\n");
439       }
440    }
441 
442    else if VG_DBL_CLO(arg, "--peak-inaccuracy", clo_peak_inaccuracy) {}
443 
444    else if VG_XACT_CLO(arg, "--time-unit=i",    clo_time_unit, TimeI)  {}
445    else if VG_XACT_CLO(arg, "--time-unit=ms",   clo_time_unit, TimeMS) {}
446    else if VG_XACT_CLO(arg, "--time-unit=B",    clo_time_unit, TimeB)  {}
447 
448    else if VG_BINT_CLO(arg, "--detailed-freq",  clo_detailed_freq, 1, 1000000) {}
449 
450    else if VG_BINT_CLO(arg, "--max-snapshots",  clo_max_snapshots, 10, 1000) {}
451 
452    else if VG_STR_CLO(arg, "--massif-out-file", clo_massif_out_file) {}
453 
454    else
455       return VG_(replacement_malloc_process_cmd_line_option)(arg);
456 
457    return True;
458 }
459 
ms_print_usage(void)460 static void ms_print_usage(void)
461 {
462    VG_(printf)(
463 "    --heap=no|yes             profile heap blocks [yes]\n"
464 "    --heap-admin=<size>       average admin bytes per heap block;\n"
465 "                               ignored if --heap=no [8]\n"
466 "    --stacks=no|yes           profile stack(s) [no]\n"
467 "    --pages-as-heap=no|yes    profile memory at the page level [no]\n"
468 "    --depth=<number>          depth of contexts [30]\n"
469 "    --alloc-fn=<name>         specify <name> as an alloc function [empty]\n"
470 "    --ignore-fn=<name>        ignore heap allocations within <name> [empty]\n"
471 "    --threshold=<m.n>         significance threshold, as a percentage [1.0]\n"
472 "    --peak-inaccuracy=<m.n>   maximum peak inaccuracy, as a percentage [1.0]\n"
473 "    --time-unit=i|ms|B        time unit: instructions executed, milliseconds\n"
474 "                              or heap bytes alloc'd/dealloc'd [i]\n"
475 "    --detailed-freq=<N>       every Nth snapshot should be detailed [10]\n"
476 "    --max-snapshots=<N>       maximum number of snapshots recorded [100]\n"
477 "    --massif-out-file=<file>  output file name [massif.out.%%p]\n"
478    );
479 }
480 
ms_print_debug_usage(void)481 static void ms_print_debug_usage(void)
482 {
483    VG_(printf)(
484 "    (none)\n"
485    );
486 }
487 
488 
489 //------------------------------------------------------------//
490 //--- XPts, XTrees and XCons                               ---//
491 //------------------------------------------------------------//
492 
493 // An XPt represents an "execution point", ie. a code address.  Each XPt is
494 // part of a tree of XPts (an "execution tree", or "XTree").  The details of
495 // the heap are represented by a single XTree.
496 //
497 // The root of the tree is 'alloc_xpt', which represents all allocation
498 // functions, eg:
499 // - malloc/calloc/realloc/memalign/new/new[];
500 // - user-specified allocation functions (using --alloc-fn);
501 // - custom allocation (MALLOCLIKE) points
502 // It's a bit of a fake XPt (ie. its 'ip' is zero), and is only used because
503 // it makes the code simpler.
504 //
505 // Any child of 'alloc_xpt' is called a "top-XPt".  The XPts at the bottom
506 // of an XTree (leaf nodes) are "bottom-XPTs".
507 //
508 // Each path from a top-XPt to a bottom-XPt through an XTree gives an
509 // execution context ("XCon"), ie. a stack trace.  (And sub-paths represent
510 // stack sub-traces.)  The number of XCons in an XTree is equal to the
511 // number of bottom-XPTs in that XTree.
512 //
513 //      alloc_xpt       XTrees are bi-directional.
514 //        | ^
515 //        v |
516 //     > parent <       Example: if child1() calls parent() and child2()
517 //    /    |     \      also calls parent(), and parent() calls malloc(),
518 //   |    / \     |     the XTree will look like this.
519 //   |   v   v    |
520 //  child1   child2
521 //
522 // (Note that malformed stack traces can lead to difficulties.  See the
523 // comment at the bottom of get_XCon.)
524 //
525 // XTrees and XPts are mirrored by SXTrees and SXPts, where the 'S' is short
526 // for "saved".  When the XTree is duplicated for a snapshot, we duplicate
527 // it as an SXTree, which is similar but omits some things it does not need,
528 // and aggregates up insignificant nodes.  This is important as an SXTree is
529 // typically much smaller than an XTree.
530 
531 // XXX: make XPt and SXPt extensible arrays, to avoid having to do two
532 // allocations per Pt.
533 
534 typedef struct _XPt XPt;
535 struct _XPt {
536    Addr  ip;              // code address
537 
538    // Bottom-XPts: space for the precise context.
539    // Other XPts:  space of all the descendent bottom-XPts.
540    // Nb: this value goes up and down as the program executes.
541    SizeT szB;
542 
543    XPt*  parent;           // pointer to parent XPt
544 
545    // Children.
546    // n_children and max_children are 32-bit integers.  16-bit integers
547    // are too small -- a very big program might have more than 65536
548    // allocation points (ie. top-XPts) -- Konqueror starting up has 1800.
549    UInt  n_children;       // number of children
550    UInt  max_children;     // capacity of children array
551    XPt** children;         // pointers to children XPts
552 };
553 
554 typedef
555    enum {
556       SigSXPt,
557       InsigSXPt
558    }
559    SXPtTag;
560 
561 typedef struct _SXPt SXPt;
562 struct _SXPt {
563    SXPtTag tag;
564    SizeT szB;              // memory size for the node, be it Sig or Insig
565    union {
566       // An SXPt representing a single significant code location.  Much like
567       // an XPt, minus the fields that aren't necessary.
568       struct {
569          Addr   ip;
570          UInt   n_children;
571          SXPt** children;
572       }
573       Sig;
574 
575       // An SXPt representing one or more code locations, all below the
576       // significance threshold.
577       struct {
578          Int   n_xpts;     // number of aggregated XPts
579       }
580       Insig;
581    };
582 };
583 
584 // Fake XPt representing all allocation functions like malloc().  Acts as
585 // parent node to all top-XPts.
586 static XPt* alloc_xpt;
587 
new_XPt(Addr ip,XPt * parent)588 static XPt* new_XPt(Addr ip, XPt* parent)
589 {
590    // XPts are never freed, so we can use VG_(perm_malloc) to allocate them.
591    // Note that we cannot use VG_(perm_malloc) for the 'children' array, because
592    // that needs to be resizable.
593    XPt* xpt    = VG_(perm_malloc)(sizeof(XPt), vg_alignof(XPt));
594    xpt->ip     = ip;
595    xpt->szB    = 0;
596    xpt->parent = parent;
597 
598    // We don't initially allocate any space for children.  We let that
599    // happen on demand.  Many XPts (ie. all the bottom-XPts) don't have any
600    // children anyway.
601    xpt->n_children   = 0;
602    xpt->max_children = 0;
603    xpt->children     = NULL;
604 
605    // Update statistics
606    n_xpts++;
607 
608    return xpt;
609 }
610 
add_child_xpt(XPt * parent,XPt * child)611 static void add_child_xpt(XPt* parent, XPt* child)
612 {
613    // Expand 'children' if necessary.
614    tl_assert(parent->n_children <= parent->max_children);
615    if (parent->n_children == parent->max_children) {
616       if (0 == parent->max_children) {
617          parent->max_children = 4;
618          parent->children = VG_(malloc)( "ms.main.acx.1",
619                                          parent->max_children * sizeof(XPt*) );
620          n_xpt_init_expansions++;
621       } else {
622          parent->max_children *= 2;    // Double size
623          parent->children = VG_(realloc)( "ms.main.acx.2",
624                                           parent->children,
625                                           parent->max_children * sizeof(XPt*) );
626          n_xpt_later_expansions++;
627       }
628    }
629 
630    // Insert new child XPt in parent's children list.
631    parent->children[ parent->n_children++ ] = child;
632 }
633 
634 // Reverse comparison for a reverse sort -- biggest to smallest.
SXPt_revcmp_szB(const void * n1,const void * n2)635 static Int SXPt_revcmp_szB(const void* n1, const void* n2)
636 {
637    const SXPt* sxpt1 = *(const SXPt *const *)n1;
638    const SXPt* sxpt2 = *(const SXPt *const *)n2;
639    return ( sxpt1->szB < sxpt2->szB ?  1
640           : sxpt1->szB > sxpt2->szB ? -1
641           :                            0);
642 }
643 
644 //------------------------------------------------------------//
645 //--- XTree Operations                                     ---//
646 //------------------------------------------------------------//
647 
648 // Duplicates an XTree as an SXTree.
dup_XTree(XPt * xpt,SizeT total_szB)649 static SXPt* dup_XTree(XPt* xpt, SizeT total_szB)
650 {
651    Int  i, n_sig_children, n_insig_children, n_child_sxpts;
652    SizeT sig_child_threshold_szB;
653    SXPt* sxpt;
654 
655    // Number of XPt children  Action for SXPT
656    // ------------------      ---------------
657    // 0 sig, 0 insig          alloc 0 children
658    // N sig, 0 insig          alloc N children, dup all
659    // N sig, M insig          alloc N+1, dup first N, aggregate remaining M
660    // 0 sig, M insig          alloc 1, aggregate M
661 
662    // Work out how big a child must be to be significant.  If the current
663    // total_szB is zero, then we set it to 1, which means everything will be
664    // judged insignificant -- this is sensible, as there's no point showing
665    // any detail for this case.  Unless they used --threshold=0, in which
666    // case we show them everything because that's what they asked for.
667    //
668    // Nb: We do this once now, rather than once per child, because if we do
669    // that the cost of all the divisions adds up to something significant.
670    if (0 == total_szB && 0 != clo_threshold) {
671       sig_child_threshold_szB = 1;
672    } else {
673       sig_child_threshold_szB = (SizeT)((total_szB * clo_threshold) / 100);
674    }
675 
676    // How many children are significant?  And do we need an aggregate SXPt?
677    n_sig_children = 0;
678    for (i = 0; i < xpt->n_children; i++) {
679       if (xpt->children[i]->szB >= sig_child_threshold_szB) {
680          n_sig_children++;
681       }
682    }
683    n_insig_children = xpt->n_children - n_sig_children;
684    n_child_sxpts = n_sig_children + ( n_insig_children > 0 ? 1 : 0 );
685 
686    // Duplicate the XPt.
687    sxpt                 = VG_(malloc)("ms.main.dX.1", sizeof(SXPt));
688    n_sxpt_allocs++;
689    sxpt->tag            = SigSXPt;
690    sxpt->szB            = xpt->szB;
691    sxpt->Sig.ip         = xpt->ip;
692    sxpt->Sig.n_children = n_child_sxpts;
693 
694    // Create the SXPt's children.
695    if (n_child_sxpts > 0) {
696       Int j;
697       SizeT sig_children_szB = 0, insig_children_szB = 0;
698       sxpt->Sig.children = VG_(malloc)("ms.main.dX.2",
699                                        n_child_sxpts * sizeof(SXPt*));
700 
701       // Duplicate the significant children.  (Nb: sig_children_szB +
702       // insig_children_szB doesn't necessarily equal xpt->szB.)
703       j = 0;
704       for (i = 0; i < xpt->n_children; i++) {
705          if (xpt->children[i]->szB >= sig_child_threshold_szB) {
706             sxpt->Sig.children[j++] = dup_XTree(xpt->children[i], total_szB);
707             sig_children_szB   += xpt->children[i]->szB;
708          } else {
709             insig_children_szB += xpt->children[i]->szB;
710          }
711       }
712 
713       // Create the SXPt for the insignificant children, if any, and put it
714       // in the last child entry.
715       if (n_insig_children > 0) {
716          // Nb: We 'n_sxpt_allocs' here because creating an Insig SXPt
717          // doesn't involve a call to dup_XTree().
718          SXPt* insig_sxpt = VG_(malloc)("ms.main.dX.3", sizeof(SXPt));
719          n_sxpt_allocs++;
720          insig_sxpt->tag = InsigSXPt;
721          insig_sxpt->szB = insig_children_szB;
722          insig_sxpt->Insig.n_xpts = n_insig_children;
723          sxpt->Sig.children[n_sig_children] = insig_sxpt;
724       }
725    } else {
726       sxpt->Sig.children = NULL;
727    }
728 
729    return sxpt;
730 }
731 
free_SXTree(SXPt * sxpt)732 static void free_SXTree(SXPt* sxpt)
733 {
734    Int  i;
735    tl_assert(sxpt != NULL);
736 
737    switch (sxpt->tag) {
738     case SigSXPt:
739       // Free all children SXPts, then the children array.
740       for (i = 0; i < sxpt->Sig.n_children; i++) {
741          free_SXTree(sxpt->Sig.children[i]);
742          sxpt->Sig.children[i] = NULL;
743       }
744       VG_(free)(sxpt->Sig.children);  sxpt->Sig.children = NULL;
745       break;
746 
747     case InsigSXPt:
748       break;
749 
750     default: tl_assert2(0, "free_SXTree: unknown SXPt tag");
751    }
752 
753    // Free the SXPt itself.
754    VG_(free)(sxpt);     sxpt = NULL;
755    n_sxpt_frees++;
756 }
757 
758 // Sanity checking:  we periodically check the heap XTree with
759 // ms_expensive_sanity_check.
sanity_check_XTree(XPt * xpt,XPt * parent)760 static void sanity_check_XTree(XPt* xpt, XPt* parent)
761 {
762    tl_assert(xpt != NULL);
763 
764    // Check back-pointer.
765    tl_assert2(xpt->parent == parent,
766       "xpt->parent = %p, parent = %p\n", xpt->parent, parent);
767 
768    // Check children counts look sane.
769    tl_assert(xpt->n_children <= xpt->max_children);
770 
771    // Unfortunately, xpt's size is not necessarily equal to the sum of xpt's
772    // children's sizes.  See comment at the bottom of get_XCon.
773 }
774 
775 // Sanity checking:  we check SXTrees (which are in snapshots) after
776 // snapshots are created, before they are deleted, and before they are
777 // printed.
sanity_check_SXTree(SXPt * sxpt)778 static void sanity_check_SXTree(SXPt* sxpt)
779 {
780    Int i;
781 
782    tl_assert(sxpt != NULL);
783 
784    // Check the sum of any children szBs equals the SXPt's szB.  Check the
785    // children at the same time.
786    switch (sxpt->tag) {
787     case SigSXPt: {
788       if (sxpt->Sig.n_children > 0) {
789          for (i = 0; i < sxpt->Sig.n_children; i++) {
790             sanity_check_SXTree(sxpt->Sig.children[i]);
791          }
792       }
793       break;
794     }
795     case InsigSXPt:
796       break;         // do nothing
797 
798     default: tl_assert2(0, "sanity_check_SXTree: unknown SXPt tag");
799    }
800 }
801 
802 
803 //------------------------------------------------------------//
804 //--- XCon Operations                                      ---//
805 //------------------------------------------------------------//
806 
807 // This is the limit on the number of removed alloc-fns that can be in a
808 // single XCon.
809 #define MAX_OVERESTIMATE   50
810 #define MAX_IPS            (MAX_DEPTH + MAX_OVERESTIMATE)
811 
812 // This is used for various buffers which can hold function names/IP
813 // description.  Some C++ names can get really long so 1024 isn't big
814 // enough.
815 #define BUF_LEN   2048
816 
817 // Determine if the given IP belongs to a function that should be ignored.
fn_should_be_ignored(Addr ip)818 static Bool fn_should_be_ignored(Addr ip)
819 {
820    static HChar buf[BUF_LEN];
821    return
822       ( VG_(get_fnname)(ip, buf, BUF_LEN) && is_member_fn(ignore_fns, buf)
823       ? True : False );
824 }
825 
826 // Get the stack trace for an XCon, filtering out uninteresting entries:
827 // alloc-fns and entries above alloc-fns, and entries below main-or-below-main.
828 //   Eg:       alloc-fn1 / alloc-fn2 / a / b / main / (below main) / c
829 //   becomes:  a / b / main
830 // Nb: it's possible to end up with an empty trace, eg. if 'main' is marked
831 // as an alloc-fn.  This is ok.
832 static
get_IPs(ThreadId tid,Bool exclude_first_entry,Addr ips[])833 Int get_IPs( ThreadId tid, Bool exclude_first_entry, Addr ips[])
834 {
835    static HChar buf[BUF_LEN];
836    Int n_ips, i, n_alloc_fns_removed;
837    Int overestimate;
838    Bool redo;
839 
840    // We ask for a few more IPs than clo_depth suggests we need.  Then we
841    // remove every entry that is an alloc-fn.  Depending on the
842    // circumstances, we may need to redo it all, asking for more IPs.
843    // Details:
844    // - If the original stack trace is smaller than asked-for, redo=False
845    // - Else if after filtering we have >= clo_depth IPs,      redo=False
846    // - Else redo=True
847    // In other words, to redo, we'd have to get a stack trace as big as we
848    // asked for and remove more than 'overestimate' alloc-fns.
849 
850    // Main loop.
851    redo = True;      // Assume this to begin with.
852    for (overestimate = 3; redo; overestimate += 6) {
853       // This should never happen -- would require MAX_OVERESTIMATE
854       // alloc-fns to be removed from the stack trace.
855       if (overestimate > MAX_OVERESTIMATE)
856          VG_(tool_panic)("get_IPs: ips[] too small, inc. MAX_OVERESTIMATE?");
857 
858       // Ask for more IPs than clo_depth suggests we need.
859       n_ips = VG_(get_StackTrace)( tid, ips, clo_depth + overestimate,
860                                    NULL/*array to dump SP values in*/,
861                                    NULL/*array to dump FP values in*/,
862                                    0/*first_ip_delta*/ );
863       tl_assert(n_ips > 0);
864 
865       // If the original stack trace is smaller than asked-for, redo=False.
866       if (n_ips < clo_depth + overestimate) { redo = False; }
867 
868       // Filter out alloc fns.  If requested, we automatically remove the
869       // first entry (which presumably will be something like malloc or
870       // __builtin_new that we're sure to filter out) without looking at it,
871       // because VG_(get_fnname) is expensive.
872       n_alloc_fns_removed = ( exclude_first_entry ? 1 : 0 );
873       for (i = n_alloc_fns_removed; i < n_ips; i++) {
874          if (VG_(get_fnname)(ips[i], buf, BUF_LEN)) {
875             if (is_member_fn(alloc_fns, buf)) {
876                n_alloc_fns_removed++;
877             } else {
878                break;
879             }
880          }
881       }
882       // Remove the alloc fns by shuffling the rest down over them.
883       n_ips -= n_alloc_fns_removed;
884       for (i = 0; i < n_ips; i++) {
885          ips[i] = ips[i + n_alloc_fns_removed];
886       }
887 
888       // If after filtering we have >= clo_depth IPs, redo=False
889       if (n_ips >= clo_depth) {
890          redo = False;
891          n_ips = clo_depth;      // Ignore any IPs below --depth.
892       }
893 
894       if (redo) {
895          n_XCon_redos++;
896       }
897    }
898    return n_ips;
899 }
900 
901 // Gets an XCon and puts it in the tree.  Returns the XCon's bottom-XPt.
902 // Unless the allocation should be ignored, in which case we return NULL.
get_XCon(ThreadId tid,Bool exclude_first_entry)903 static XPt* get_XCon( ThreadId tid, Bool exclude_first_entry )
904 {
905    static Addr ips[MAX_IPS];
906    Int i;
907    XPt* xpt = alloc_xpt;
908 
909    // After this call, the IPs we want are in ips[0]..ips[n_ips-1].
910    Int n_ips = get_IPs(tid, exclude_first_entry, ips);
911 
912    // Should we ignore this allocation?  (Nb: n_ips can be zero, eg. if
913    // 'main' is marked as an alloc-fn.)
914    if (n_ips > 0 && fn_should_be_ignored(ips[0])) {
915       return NULL;
916    }
917 
918    // Now do the search/insertion of the XCon.
919    for (i = 0; i < n_ips; i++) {
920       Addr ip = ips[i];
921       Int ch;
922       // Look for IP in xpt's children.
923       // Linear search, ugh -- about 10% of time for konqueror startup tried
924       // caching last result, only hit about 4% for konqueror.
925       // Nb:  this search hits about 98% of the time for konqueror
926       for (ch = 0; True; ch++) {
927          if (ch == xpt->n_children) {
928             // IP not found in the children.
929             // Create and add new child XPt, then stop.
930             XPt* new_child_xpt = new_XPt(ip, xpt);
931             add_child_xpt(xpt, new_child_xpt);
932             xpt = new_child_xpt;
933             break;
934 
935          } else if (ip == xpt->children[ch]->ip) {
936             // Found the IP in the children, stop.
937             xpt = xpt->children[ch];
938             break;
939          }
940       }
941    }
942 
943    // [Note: several comments refer to this comment.  Do not delete it
944    // without updating them.]
945    //
946    // A complication... If all stack traces were well-formed, then the
947    // returned xpt would always be a bottom-XPt.  As a consequence, an XPt's
948    // size would always be equal to the sum of its children's sizes, which
949    // is an excellent sanity check.
950    //
951    // Unfortunately, stack traces occasionally are malformed, ie. truncated.
952    // This allows a stack trace to be a sub-trace of another, eg. a/b/c is a
953    // sub-trace of a/b/c/d.  So we can't assume this xpt is a bottom-XPt;
954    // nor can we do sanity check an XPt's size against its children's sizes.
955    // This is annoying, but must be dealt with.  (Older versions of Massif
956    // had this assertion in, and it was reported to fail by real users a
957    // couple of times.)  Even more annoyingly, I can't come up with a simple
958    // test case that exhibit such a malformed stack trace, so I can't
959    // regression test it.  Sigh.
960    //
961    // However, we can print a warning, so that if it happens (unexpectedly)
962    // in existing regression tests we'll know.  Also, it warns users that
963    // the output snapshots may not add up the way they might expect.
964    //
965    //tl_assert(0 == xpt->n_children); // Must be bottom-XPt
966    if (0 != xpt->n_children) {
967       static Int n_moans = 0;
968       if (n_moans < 3) {
969          VG_(umsg)(
970             "Warning: Malformed stack trace detected.  In Massif's output,\n");
971          VG_(umsg)(
972             "         the size of an entry's child entries may not sum up\n");
973          VG_(umsg)(
974             "         to the entry's size as they normally do.\n");
975          n_moans++;
976          if (3 == n_moans)
977             VG_(umsg)(
978             "         (And Massif now won't warn about this again.)\n");
979       }
980    }
981    return xpt;
982 }
983 
984 // Update 'szB' of every XPt in the XCon, by percolating upwards.
update_XCon(XPt * xpt,SSizeT space_delta)985 static void update_XCon(XPt* xpt, SSizeT space_delta)
986 {
987    tl_assert(clo_heap);
988    tl_assert(NULL != xpt);
989 
990    if (0 == space_delta)
991       return;
992 
993    while (xpt != alloc_xpt) {
994       if (space_delta < 0) tl_assert(xpt->szB >= -space_delta);
995       xpt->szB += space_delta;
996       xpt = xpt->parent;
997    }
998    if (space_delta < 0) tl_assert(alloc_xpt->szB >= -space_delta);
999    alloc_xpt->szB += space_delta;
1000 }
1001 
1002 
1003 //------------------------------------------------------------//
1004 //--- Snapshots                                            ---//
1005 //------------------------------------------------------------//
1006 
1007 // Snapshots are done in a way so that we always have a reasonable number of
1008 // them.  We start by taking them quickly.  Once we hit our limit, we cull
1009 // some (eg. half), and start taking them more slowly.  Once we hit the
1010 // limit again, we again cull and then take them even more slowly, and so
1011 // on.
1012 
1013 // Time is measured either in i or ms or bytes, depending on the --time-unit
1014 // option.  It's a Long because it can exceed 32-bits reasonably easily, and
1015 // because we need to allow negative values to represent unset times.
1016 typedef Long Time;
1017 
1018 #define UNUSED_SNAPSHOT_TIME  -333  // A conspicuous negative number.
1019 
1020 typedef
1021    enum {
1022       Normal = 77,
1023       Peak,
1024       Unused
1025    }
1026    SnapshotKind;
1027 
1028 typedef
1029    struct {
1030       SnapshotKind kind;
1031       Time  time;
1032       SizeT heap_szB;
1033       SizeT heap_extra_szB;// Heap slop + admin bytes.
1034       SizeT stacks_szB;
1035       SXPt* alloc_sxpt;    // Heap XTree root, if a detailed snapshot,
1036    }                       // otherwise NULL.
1037    Snapshot;
1038 
1039 static UInt      next_snapshot_i = 0;  // Index of where next snapshot will go.
1040 static Snapshot* snapshots;            // Array of snapshots.
1041 
is_snapshot_in_use(Snapshot * snapshot)1042 static Bool is_snapshot_in_use(Snapshot* snapshot)
1043 {
1044    if (Unused == snapshot->kind) {
1045       // If snapshot is unused, check all the fields are unset.
1046       tl_assert(snapshot->time           == UNUSED_SNAPSHOT_TIME);
1047       tl_assert(snapshot->heap_extra_szB == 0);
1048       tl_assert(snapshot->heap_szB       == 0);
1049       tl_assert(snapshot->stacks_szB     == 0);
1050       tl_assert(snapshot->alloc_sxpt     == NULL);
1051       return False;
1052    } else {
1053       tl_assert(snapshot->time           != UNUSED_SNAPSHOT_TIME);
1054       return True;
1055    }
1056 }
1057 
is_detailed_snapshot(Snapshot * snapshot)1058 static Bool is_detailed_snapshot(Snapshot* snapshot)
1059 {
1060    return (snapshot->alloc_sxpt ? True : False);
1061 }
1062 
is_uncullable_snapshot(Snapshot * snapshot)1063 static Bool is_uncullable_snapshot(Snapshot* snapshot)
1064 {
1065    return &snapshots[0] == snapshot                   // First snapshot
1066        || &snapshots[next_snapshot_i-1] == snapshot   // Last snapshot
1067        || snapshot->kind == Peak;                     // Peak snapshot
1068 }
1069 
sanity_check_snapshot(Snapshot * snapshot)1070 static void sanity_check_snapshot(Snapshot* snapshot)
1071 {
1072    if (snapshot->alloc_sxpt) {
1073       sanity_check_SXTree(snapshot->alloc_sxpt);
1074    }
1075 }
1076 
1077 // All the used entries should look used, all the unused ones should be clear.
sanity_check_snapshots_array(void)1078 static void sanity_check_snapshots_array(void)
1079 {
1080    Int i;
1081    for (i = 0; i < next_snapshot_i; i++) {
1082       tl_assert( is_snapshot_in_use( & snapshots[i] ));
1083    }
1084    for (    ; i < clo_max_snapshots; i++) {
1085       tl_assert(!is_snapshot_in_use( & snapshots[i] ));
1086    }
1087 }
1088 
1089 // This zeroes all the fields in the snapshot, but does not free the heap
1090 // XTree if present.  It also does a sanity check unless asked not to;  we
1091 // can't sanity check at startup when clearing the initial snapshots because
1092 // they're full of junk.
clear_snapshot(Snapshot * snapshot,Bool do_sanity_check)1093 static void clear_snapshot(Snapshot* snapshot, Bool do_sanity_check)
1094 {
1095    if (do_sanity_check) sanity_check_snapshot(snapshot);
1096    snapshot->kind           = Unused;
1097    snapshot->time           = UNUSED_SNAPSHOT_TIME;
1098    snapshot->heap_extra_szB = 0;
1099    snapshot->heap_szB       = 0;
1100    snapshot->stacks_szB     = 0;
1101    snapshot->alloc_sxpt     = NULL;
1102 }
1103 
1104 // This zeroes all the fields in the snapshot, and frees the heap XTree if
1105 // present.
delete_snapshot(Snapshot * snapshot)1106 static void delete_snapshot(Snapshot* snapshot)
1107 {
1108    // Nb: if there's an XTree, we free it after calling clear_snapshot,
1109    // because clear_snapshot does a sanity check which includes checking the
1110    // XTree.
1111    SXPt* tmp_sxpt = snapshot->alloc_sxpt;
1112    clear_snapshot(snapshot, /*do_sanity_check*/True);
1113    if (tmp_sxpt) {
1114       free_SXTree(tmp_sxpt);
1115    }
1116 }
1117 
VERB_snapshot(Int verbosity,const HChar * prefix,Int i)1118 static void VERB_snapshot(Int verbosity, const HChar* prefix, Int i)
1119 {
1120    Snapshot* snapshot = &snapshots[i];
1121    const HChar* suffix;
1122    switch (snapshot->kind) {
1123    case Peak:   suffix = "p";                                            break;
1124    case Normal: suffix = ( is_detailed_snapshot(snapshot) ? "d" : "." ); break;
1125    case Unused: suffix = "u";                                            break;
1126    default:
1127       tl_assert2(0, "VERB_snapshot: unknown snapshot kind: %d", snapshot->kind);
1128    }
1129    VERB(verbosity, "%s S%s%3d (t:%lld, hp:%ld, ex:%ld, st:%ld)\n",
1130       prefix, suffix, i,
1131       snapshot->time,
1132       snapshot->heap_szB,
1133       snapshot->heap_extra_szB,
1134       snapshot->stacks_szB
1135    );
1136 }
1137 
1138 // Cull half the snapshots;  we choose those that represent the smallest
1139 // time-spans, because that gives us the most even distribution of snapshots
1140 // over time.  (It's possible to lose interesting spikes, however.)
1141 //
1142 // Algorithm for N snapshots:  We find the snapshot representing the smallest
1143 // timeframe, and remove it.  We repeat this until (N/2) snapshots are gone.
1144 // We have to do this one snapshot at a time, rather than finding the (N/2)
1145 // smallest snapshots in one hit, because when a snapshot is removed, its
1146 // neighbours immediately cover greater timespans.  So it's O(N^2), but N is
1147 // small, and it's not done very often.
1148 //
1149 // Once we're done, we return the new smallest interval between snapshots.
1150 // That becomes our minimum time interval.
cull_snapshots(void)1151 static UInt cull_snapshots(void)
1152 {
1153    Int  i, jp, j, jn, min_timespan_i;
1154    Int  n_deleted = 0;
1155    Time min_timespan;
1156 
1157    n_cullings++;
1158 
1159    // Sets j to the index of the first not-yet-removed snapshot at or after i
1160    #define FIND_SNAPSHOT(i, j) \
1161       for (j = i; \
1162            j < clo_max_snapshots && !is_snapshot_in_use(&snapshots[j]); \
1163            j++) { }
1164 
1165    VERB(2, "Culling...\n");
1166 
1167    // First we remove enough snapshots by clearing them in-place.  Once
1168    // that's done, we can slide the remaining ones down.
1169    for (i = 0; i < clo_max_snapshots/2; i++) {
1170       // Find the snapshot representing the smallest timespan.  The timespan
1171       // for snapshot n = d(N-1,N)+d(N,N+1), where d(A,B) is the time between
1172       // snapshot A and B.  We don't consider the first and last snapshots for
1173       // removal.
1174       Snapshot* min_snapshot;
1175       Int min_j;
1176 
1177       // Initial triple: (prev, curr, next) == (jp, j, jn)
1178       // Initial min_timespan is the first one.
1179       jp = 0;
1180       FIND_SNAPSHOT(1,   j);
1181       FIND_SNAPSHOT(j+1, jn);
1182       min_timespan = 0x7fffffffffffffffLL;
1183       min_j        = -1;
1184       while (jn < clo_max_snapshots) {
1185          Time timespan = snapshots[jn].time - snapshots[jp].time;
1186          tl_assert(timespan >= 0);
1187          // Nb: We never cull the peak snapshot.
1188          if (Peak != snapshots[j].kind && timespan < min_timespan) {
1189             min_timespan = timespan;
1190             min_j        = j;
1191          }
1192          // Move on to next triple
1193          jp = j;
1194          j  = jn;
1195          FIND_SNAPSHOT(jn+1, jn);
1196       }
1197       // We've found the least important snapshot, now delete it.  First
1198       // print it if necessary.
1199       tl_assert(-1 != min_j);    // Check we found a minimum.
1200       min_snapshot = & snapshots[ min_j ];
1201       if (VG_(clo_verbosity) > 1) {
1202          HChar buf[64];
1203          VG_(snprintf)(buf, 64, " %3d (t-span = %lld)", i, min_timespan);
1204          VERB_snapshot(2, buf, min_j);
1205       }
1206       delete_snapshot(min_snapshot);
1207       n_deleted++;
1208    }
1209 
1210    // Slide down the remaining snapshots over the removed ones.  First set i
1211    // to point to the first empty slot, and j to the first full slot after
1212    // i.  Then slide everything down.
1213    for (i = 0;  is_snapshot_in_use( &snapshots[i] ); i++) { }
1214    for (j = i; !is_snapshot_in_use( &snapshots[j] ); j++) { }
1215    for (  ; j < clo_max_snapshots; j++) {
1216       if (is_snapshot_in_use( &snapshots[j] )) {
1217          snapshots[i++] = snapshots[j];
1218          clear_snapshot(&snapshots[j], /*do_sanity_check*/True);
1219       }
1220    }
1221    next_snapshot_i = i;
1222 
1223    // Check snapshots array looks ok after changes.
1224    sanity_check_snapshots_array();
1225 
1226    // Find the minimum timespan remaining;  that will be our new minimum
1227    // time interval.  Note that above we were finding timespans by measuring
1228    // two intervals around a snapshot that was under consideration for
1229    // deletion.  Here we only measure single intervals because all the
1230    // deletions have occurred.
1231    //
1232    // But we have to be careful -- some snapshots (eg. snapshot 0, and the
1233    // peak snapshot) are uncullable.  If two uncullable snapshots end up
1234    // next to each other, they'll never be culled (assuming the peak doesn't
1235    // change), and the time gap between them will not change.  However, the
1236    // time between the remaining cullable snapshots will grow ever larger.
1237    // This means that the min_timespan found will always be that between the
1238    // two uncullable snapshots, and it will be much smaller than it should
1239    // be.  To avoid this problem, when computing the minimum timespan, we
1240    // ignore any timespans between two uncullable snapshots.
1241    tl_assert(next_snapshot_i > 1);
1242    min_timespan = 0x7fffffffffffffffLL;
1243    min_timespan_i = -1;
1244    for (i = 1; i < next_snapshot_i; i++) {
1245       if (is_uncullable_snapshot(&snapshots[i]) &&
1246           is_uncullable_snapshot(&snapshots[i-1]))
1247       {
1248          VERB(2, "(Ignoring interval %d--%d when computing minimum)\n", i-1, i);
1249       } else {
1250          Time timespan = snapshots[i].time - snapshots[i-1].time;
1251          tl_assert(timespan >= 0);
1252          if (timespan < min_timespan) {
1253             min_timespan = timespan;
1254             min_timespan_i = i;
1255          }
1256       }
1257    }
1258    tl_assert(-1 != min_timespan_i);    // Check we found a minimum.
1259 
1260    // Print remaining snapshots, if necessary.
1261    if (VG_(clo_verbosity) > 1) {
1262       VERB(2, "Finished culling (%3d of %3d deleted)\n",
1263          n_deleted, clo_max_snapshots);
1264       for (i = 0; i < next_snapshot_i; i++) {
1265          VERB_snapshot(2, "  post-cull", i);
1266       }
1267       VERB(2, "New time interval = %lld (between snapshots %d and %d)\n",
1268          min_timespan, min_timespan_i-1, min_timespan_i);
1269    }
1270 
1271    return min_timespan;
1272 }
1273 
get_time(void)1274 static Time get_time(void)
1275 {
1276    // Get current time, in whatever time unit we're using.
1277    if (clo_time_unit == TimeI) {
1278       return guest_instrs_executed;
1279    } else if (clo_time_unit == TimeMS) {
1280       // Some stuff happens between the millisecond timer being initialised
1281       // to zero and us taking our first snapshot.  We determine that time
1282       // gap so we can subtract it from all subsequent times so that our
1283       // first snapshot is considered to be at t = 0ms.  Unfortunately, a
1284       // bunch of symbols get read after the first snapshot is taken but
1285       // before the second one (which is triggered by the first allocation),
1286       // so when the time-unit is 'ms' we always have a big gap between the
1287       // first two snapshots.  But at least users won't have to wonder why
1288       // the first snapshot isn't at t=0.
1289       static Bool is_first_get_time = True;
1290       static Time start_time_ms;
1291       if (is_first_get_time) {
1292          start_time_ms = VG_(read_millisecond_timer)();
1293          is_first_get_time = False;
1294          return 0;
1295       } else {
1296          return VG_(read_millisecond_timer)() - start_time_ms;
1297       }
1298    } else if (clo_time_unit == TimeB) {
1299       return total_allocs_deallocs_szB;
1300    } else {
1301       tl_assert2(0, "bad --time-unit value");
1302    }
1303 }
1304 
1305 // Take a snapshot, and only that -- decisions on whether to take a
1306 // snapshot, or what kind of snapshot, are made elsewhere.
1307 // Nb: we call the arg "my_time" because "time" shadows a global declaration
1308 // in /usr/include/time.h on Darwin.
1309 static void
take_snapshot(Snapshot * snapshot,SnapshotKind kind,Time my_time,Bool is_detailed)1310 take_snapshot(Snapshot* snapshot, SnapshotKind kind, Time my_time,
1311               Bool is_detailed)
1312 {
1313    tl_assert(!is_snapshot_in_use(snapshot));
1314    if (!clo_pages_as_heap) {
1315       tl_assert(have_started_executing_code);
1316    }
1317 
1318    // Heap and heap admin.
1319    if (clo_heap) {
1320       snapshot->heap_szB = heap_szB;
1321       if (is_detailed) {
1322          SizeT total_szB = heap_szB + heap_extra_szB + stacks_szB;
1323          snapshot->alloc_sxpt = dup_XTree(alloc_xpt, total_szB);
1324          tl_assert(           alloc_xpt->szB == heap_szB);
1325          tl_assert(snapshot->alloc_sxpt->szB == heap_szB);
1326       }
1327       snapshot->heap_extra_szB = heap_extra_szB;
1328    }
1329 
1330    // Stack(s).
1331    if (clo_stacks) {
1332       snapshot->stacks_szB = stacks_szB;
1333    }
1334 
1335    // Rest of snapshot.
1336    snapshot->kind = kind;
1337    snapshot->time = my_time;
1338    sanity_check_snapshot(snapshot);
1339 
1340    // Update stats.
1341    if (Peak == kind) n_peak_snapshots++;
1342    if (is_detailed)  n_detailed_snapshots++;
1343    n_real_snapshots++;
1344 }
1345 
1346 
1347 // Take a snapshot, if it's time, or if we've hit a peak.
1348 static void
maybe_take_snapshot(SnapshotKind kind,const HChar * what)1349 maybe_take_snapshot(SnapshotKind kind, const HChar* what)
1350 {
1351    // 'min_time_interval' is the minimum time interval between snapshots.
1352    // If we try to take a snapshot and less than this much time has passed,
1353    // we don't take it.  It gets larger as the program runs longer.  It's
1354    // initialised to zero so that we begin by taking snapshots as quickly as
1355    // possible.
1356    static Time min_time_interval = 0;
1357    // Zero allows startup snapshot.
1358    static Time earliest_possible_time_of_next_snapshot = 0;
1359    static Int  n_snapshots_since_last_detailed         = 0;
1360    static Int  n_skipped_snapshots_since_last_snapshot = 0;
1361 
1362    Snapshot* snapshot;
1363    Bool      is_detailed;
1364    // Nb: we call this variable "my_time" because "time" shadows a global
1365    // declaration in /usr/include/time.h on Darwin.
1366    Time      my_time = get_time();
1367 
1368    switch (kind) {
1369     case Normal:
1370       // Only do a snapshot if it's time.
1371       if (my_time < earliest_possible_time_of_next_snapshot) {
1372          n_skipped_snapshots++;
1373          n_skipped_snapshots_since_last_snapshot++;
1374          return;
1375       }
1376       is_detailed = (clo_detailed_freq-1 == n_snapshots_since_last_detailed);
1377       break;
1378 
1379     case Peak: {
1380       // Because we're about to do a deallocation, we're coming down from a
1381       // local peak.  If it is (a) actually a global peak, and (b) a certain
1382       // amount bigger than the previous peak, then we take a peak snapshot.
1383       // By not taking a snapshot for every peak, we save a lot of effort --
1384       // because many peaks remain peak only for a short time.
1385       SizeT total_szB = heap_szB + heap_extra_szB + stacks_szB;
1386       SizeT excess_szB_for_new_peak =
1387          (SizeT)((peak_snapshot_total_szB * clo_peak_inaccuracy) / 100);
1388       if (total_szB <= peak_snapshot_total_szB + excess_szB_for_new_peak) {
1389          return;
1390       }
1391       is_detailed = True;
1392       break;
1393     }
1394 
1395     default:
1396       tl_assert2(0, "maybe_take_snapshot: unrecognised snapshot kind");
1397    }
1398 
1399    // Take the snapshot.
1400    snapshot = & snapshots[next_snapshot_i];
1401    take_snapshot(snapshot, kind, my_time, is_detailed);
1402 
1403    // Record if it was detailed.
1404    if (is_detailed) {
1405       n_snapshots_since_last_detailed = 0;
1406    } else {
1407       n_snapshots_since_last_detailed++;
1408    }
1409 
1410    // Update peak data, if it's a Peak snapshot.
1411    if (Peak == kind) {
1412       Int i, number_of_peaks_snapshots_found = 0;
1413 
1414       // Sanity check the size, then update our recorded peak.
1415       SizeT snapshot_total_szB =
1416          snapshot->heap_szB + snapshot->heap_extra_szB + snapshot->stacks_szB;
1417       tl_assert2(snapshot_total_szB > peak_snapshot_total_szB,
1418          "%ld, %ld\n", snapshot_total_szB, peak_snapshot_total_szB);
1419       peak_snapshot_total_szB = snapshot_total_szB;
1420 
1421       // Find the old peak snapshot, if it exists, and mark it as normal.
1422       for (i = 0; i < next_snapshot_i; i++) {
1423          if (Peak == snapshots[i].kind) {
1424             snapshots[i].kind = Normal;
1425             number_of_peaks_snapshots_found++;
1426          }
1427       }
1428       tl_assert(number_of_peaks_snapshots_found <= 1);
1429    }
1430 
1431    // Finish up verbosity and stats stuff.
1432    if (n_skipped_snapshots_since_last_snapshot > 0) {
1433       VERB(2, "  (skipped %d snapshot%s)\n",
1434          n_skipped_snapshots_since_last_snapshot,
1435          ( 1 == n_skipped_snapshots_since_last_snapshot ? "" : "s") );
1436    }
1437    VERB_snapshot(2, what, next_snapshot_i);
1438    n_skipped_snapshots_since_last_snapshot = 0;
1439 
1440    // Cull the entries, if our snapshot table is full.
1441    next_snapshot_i++;
1442    if (clo_max_snapshots == next_snapshot_i) {
1443       min_time_interval = cull_snapshots();
1444    }
1445 
1446    // Work out the earliest time when the next snapshot can happen.
1447    earliest_possible_time_of_next_snapshot = my_time + min_time_interval;
1448 }
1449 
1450 
1451 //------------------------------------------------------------//
1452 //--- Sanity checking                                      ---//
1453 //------------------------------------------------------------//
1454 
ms_cheap_sanity_check(void)1455 static Bool ms_cheap_sanity_check ( void )
1456 {
1457    return True;   // Nothing useful we can cheaply check.
1458 }
1459 
ms_expensive_sanity_check(void)1460 static Bool ms_expensive_sanity_check ( void )
1461 {
1462    sanity_check_XTree(alloc_xpt, /*parent*/NULL);
1463    sanity_check_snapshots_array();
1464    return True;
1465 }
1466 
1467 
1468 //------------------------------------------------------------//
1469 //--- Heap management                                      ---//
1470 //------------------------------------------------------------//
1471 
1472 // Metadata for heap blocks.  Each one contains a pointer to a bottom-XPt,
1473 // which is a foothold into the XCon at which it was allocated.  From
1474 // HP_Chunks, XPt 'space' fields are incremented (at allocation) and
1475 // decremented (at deallocation).
1476 //
1477 // Nb: first two fields must match core's VgHashNode.
1478 typedef
1479    struct _HP_Chunk {
1480       struct _HP_Chunk* next;
1481       Addr              data;       // Ptr to actual block
1482       SizeT             req_szB;    // Size requested
1483       SizeT             slop_szB;   // Extra bytes given above those requested
1484       XPt*              where;      // Where allocated; bottom-XPt
1485    }
1486    HP_Chunk;
1487 
1488 static VgHashTable malloc_list  = NULL;   // HP_Chunks
1489 
update_alloc_stats(SSizeT szB_delta)1490 static void update_alloc_stats(SSizeT szB_delta)
1491 {
1492    // Update total_allocs_deallocs_szB.
1493    if (szB_delta < 0) szB_delta = -szB_delta;
1494    total_allocs_deallocs_szB += szB_delta;
1495 }
1496 
update_heap_stats(SSizeT heap_szB_delta,Int heap_extra_szB_delta)1497 static void update_heap_stats(SSizeT heap_szB_delta, Int heap_extra_szB_delta)
1498 {
1499    if (heap_szB_delta < 0)
1500       tl_assert(heap_szB >= -heap_szB_delta);
1501    if (heap_extra_szB_delta < 0)
1502       tl_assert(heap_extra_szB >= -heap_extra_szB_delta);
1503 
1504    heap_extra_szB += heap_extra_szB_delta;
1505    heap_szB       += heap_szB_delta;
1506 
1507    update_alloc_stats(heap_szB_delta + heap_extra_szB_delta);
1508 }
1509 
1510 static
record_block(ThreadId tid,void * p,SizeT req_szB,SizeT slop_szB,Bool exclude_first_entry,Bool maybe_snapshot)1511 void* record_block( ThreadId tid, void* p, SizeT req_szB, SizeT slop_szB,
1512                     Bool exclude_first_entry, Bool maybe_snapshot )
1513 {
1514    // Make new HP_Chunk node, add to malloc_list
1515    HP_Chunk* hc = VG_(malloc)("ms.main.rb.1", sizeof(HP_Chunk));
1516    hc->req_szB  = req_szB;
1517    hc->slop_szB = slop_szB;
1518    hc->data     = (Addr)p;
1519    hc->where    = NULL;
1520    VG_(HT_add_node)(malloc_list, hc);
1521 
1522    if (clo_heap) {
1523       VERB(3, "<<< record_block (%lu, %lu)\n", req_szB, slop_szB);
1524 
1525       hc->where = get_XCon( tid, exclude_first_entry );
1526 
1527       if (hc->where) {
1528          // Update statistics.
1529          n_heap_allocs++;
1530 
1531          // Update heap stats.
1532          update_heap_stats(req_szB, clo_heap_admin + slop_szB);
1533 
1534          // Update XTree.
1535          update_XCon(hc->where, req_szB);
1536 
1537          // Maybe take a snapshot.
1538          if (maybe_snapshot) {
1539             maybe_take_snapshot(Normal, "  alloc");
1540          }
1541 
1542       } else {
1543          // Ignored allocation.
1544          n_ignored_heap_allocs++;
1545 
1546          VERB(3, "(ignored)\n");
1547       }
1548 
1549       VERB(3, ">>>\n");
1550    }
1551 
1552    return p;
1553 }
1554 
1555 static __inline__
alloc_and_record_block(ThreadId tid,SizeT req_szB,SizeT req_alignB,Bool is_zeroed)1556 void* alloc_and_record_block ( ThreadId tid, SizeT req_szB, SizeT req_alignB,
1557                                Bool is_zeroed )
1558 {
1559    SizeT actual_szB, slop_szB;
1560    void* p;
1561 
1562    if ((SSizeT)req_szB < 0) return NULL;
1563 
1564    // Allocate and zero if necessary.
1565    p = VG_(cli_malloc)( req_alignB, req_szB );
1566    if (!p) {
1567       return NULL;
1568    }
1569    if (is_zeroed) VG_(memset)(p, 0, req_szB);
1570    actual_szB = VG_(malloc_usable_size)(p);
1571    tl_assert(actual_szB >= req_szB);
1572    slop_szB = actual_szB - req_szB;
1573 
1574    // Record block.
1575    record_block(tid, p, req_szB, slop_szB, /*exclude_first_entry*/True,
1576                 /*maybe_snapshot*/True);
1577 
1578    return p;
1579 }
1580 
1581 static __inline__
unrecord_block(void * p,Bool maybe_snapshot)1582 void unrecord_block ( void* p, Bool maybe_snapshot )
1583 {
1584    // Remove HP_Chunk from malloc_list
1585    HP_Chunk* hc = VG_(HT_remove)(malloc_list, (UWord)p);
1586    if (NULL == hc) {
1587       return;   // must have been a bogus free()
1588    }
1589 
1590    if (clo_heap) {
1591       VERB(3, "<<< unrecord_block\n");
1592 
1593       if (hc->where) {
1594          // Update statistics.
1595          n_heap_frees++;
1596 
1597          // Maybe take a peak snapshot, since it's a deallocation.
1598          if (maybe_snapshot) {
1599             maybe_take_snapshot(Peak, "de-PEAK");
1600          }
1601 
1602          // Update heap stats.
1603          update_heap_stats(-hc->req_szB, -clo_heap_admin - hc->slop_szB);
1604 
1605          // Update XTree.
1606          update_XCon(hc->where, -hc->req_szB);
1607 
1608          // Maybe take a snapshot.
1609          if (maybe_snapshot) {
1610             maybe_take_snapshot(Normal, "dealloc");
1611          }
1612 
1613       } else {
1614          n_ignored_heap_frees++;
1615 
1616          VERB(3, "(ignored)\n");
1617       }
1618 
1619       VERB(3, ">>> (-%lu, -%lu)\n", hc->req_szB, hc->slop_szB);
1620    }
1621 
1622    // Actually free the chunk, and the heap block (if necessary)
1623    VG_(free)( hc );  hc = NULL;
1624 }
1625 
1626 // Nb: --ignore-fn is tricky for realloc.  If the block's original alloc was
1627 // ignored, but the realloc is not requested to be ignored, and we are
1628 // shrinking the block, then we have to ignore the realloc -- otherwise we
1629 // could end up with negative heap sizes.  This isn't a danger if we are
1630 // growing such a block, but for consistency (it also simplifies things) we
1631 // ignore such reallocs as well.
1632 static __inline__
realloc_block(ThreadId tid,void * p_old,SizeT new_req_szB)1633 void* realloc_block ( ThreadId tid, void* p_old, SizeT new_req_szB )
1634 {
1635    HP_Chunk* hc;
1636    void*     p_new;
1637    SizeT     old_req_szB, old_slop_szB, new_slop_szB, new_actual_szB;
1638    XPt      *old_where, *new_where;
1639    Bool      is_ignored = False;
1640 
1641    // Remove the old block
1642    hc = VG_(HT_remove)(malloc_list, (UWord)p_old);
1643    if (hc == NULL) {
1644       return NULL;   // must have been a bogus realloc()
1645    }
1646 
1647    old_req_szB  = hc->req_szB;
1648    old_slop_szB = hc->slop_szB;
1649 
1650    tl_assert(!clo_pages_as_heap);  // Shouldn't be here if --pages-as-heap=yes.
1651    if (clo_heap) {
1652       VERB(3, "<<< realloc_block (%lu)\n", new_req_szB);
1653 
1654       if (hc->where) {
1655          // Update statistics.
1656          n_heap_reallocs++;
1657 
1658          // Maybe take a peak snapshot, if it's (effectively) a deallocation.
1659          if (new_req_szB < old_req_szB) {
1660             maybe_take_snapshot(Peak, "re-PEAK");
1661          }
1662       } else {
1663          // The original malloc was ignored, so we have to ignore the
1664          // realloc as well.
1665          is_ignored = True;
1666       }
1667    }
1668 
1669    // Actually do the allocation, if necessary.
1670    if (new_req_szB <= old_req_szB + old_slop_szB) {
1671       // New size is smaller or same;  block not moved.
1672       p_new = p_old;
1673       new_slop_szB = old_slop_szB + (old_req_szB - new_req_szB);
1674 
1675    } else {
1676       // New size is bigger;  make new block, copy shared contents, free old.
1677       p_new = VG_(cli_malloc)(VG_(clo_alignment), new_req_szB);
1678       if (!p_new) {
1679          // Nb: if realloc fails, NULL is returned but the old block is not
1680          // touched.  What an awful function.
1681          return NULL;
1682       }
1683       VG_(memcpy)(p_new, p_old, old_req_szB + old_slop_szB);
1684       VG_(cli_free)(p_old);
1685       new_actual_szB = VG_(malloc_usable_size)(p_new);
1686       tl_assert(new_actual_szB >= new_req_szB);
1687       new_slop_szB = new_actual_szB - new_req_szB;
1688    }
1689 
1690    if (p_new) {
1691       // Update HP_Chunk.
1692       hc->data     = (Addr)p_new;
1693       hc->req_szB  = new_req_szB;
1694       hc->slop_szB = new_slop_szB;
1695       old_where    = hc->where;
1696       hc->where    = NULL;
1697 
1698       // Update XTree.
1699       if (clo_heap) {
1700          new_where = get_XCon( tid, /*exclude_first_entry*/True);
1701          if (!is_ignored && new_where) {
1702             hc->where = new_where;
1703             update_XCon(old_where, -old_req_szB);
1704             update_XCon(new_where,  new_req_szB);
1705          } else {
1706             // The realloc itself is ignored.
1707             is_ignored = True;
1708 
1709             // Update statistics.
1710             n_ignored_heap_reallocs++;
1711          }
1712       }
1713    }
1714 
1715    // Now insert the new hc (with a possibly new 'data' field) into
1716    // malloc_list.  If this realloc() did not increase the memory size, we
1717    // will have removed and then re-added hc unnecessarily.  But that's ok
1718    // because shrinking a block with realloc() is (presumably) much rarer
1719    // than growing it, and this way simplifies the growing case.
1720    VG_(HT_add_node)(malloc_list, hc);
1721 
1722    if (clo_heap) {
1723       if (!is_ignored) {
1724          // Update heap stats.
1725          update_heap_stats(new_req_szB - old_req_szB,
1726                           new_slop_szB - old_slop_szB);
1727 
1728          // Maybe take a snapshot.
1729          maybe_take_snapshot(Normal, "realloc");
1730       } else {
1731 
1732          VERB(3, "(ignored)\n");
1733       }
1734 
1735       VERB(3, ">>> (%ld, %ld)\n",
1736          new_req_szB - old_req_szB, new_slop_szB - old_slop_szB);
1737    }
1738 
1739    return p_new;
1740 }
1741 
1742 
1743 //------------------------------------------------------------//
1744 //--- malloc() et al replacement wrappers                  ---//
1745 //------------------------------------------------------------//
1746 
ms_malloc(ThreadId tid,SizeT szB)1747 static void* ms_malloc ( ThreadId tid, SizeT szB )
1748 {
1749    return alloc_and_record_block( tid, szB, VG_(clo_alignment), /*is_zeroed*/False );
1750 }
1751 
ms___builtin_new(ThreadId tid,SizeT szB)1752 static void* ms___builtin_new ( ThreadId tid, SizeT szB )
1753 {
1754    return alloc_and_record_block( tid, szB, VG_(clo_alignment), /*is_zeroed*/False );
1755 }
1756 
ms___builtin_vec_new(ThreadId tid,SizeT szB)1757 static void* ms___builtin_vec_new ( ThreadId tid, SizeT szB )
1758 {
1759    return alloc_and_record_block( tid, szB, VG_(clo_alignment), /*is_zeroed*/False );
1760 }
1761 
ms_calloc(ThreadId tid,SizeT m,SizeT szB)1762 static void* ms_calloc ( ThreadId tid, SizeT m, SizeT szB )
1763 {
1764    return alloc_and_record_block( tid, m*szB, VG_(clo_alignment), /*is_zeroed*/True );
1765 }
1766 
ms_memalign(ThreadId tid,SizeT alignB,SizeT szB)1767 static void *ms_memalign ( ThreadId tid, SizeT alignB, SizeT szB )
1768 {
1769    return alloc_and_record_block( tid, szB, alignB, False );
1770 }
1771 
ms_free(ThreadId tid,void * p)1772 static void ms_free ( ThreadId tid __attribute__((unused)), void* p )
1773 {
1774    unrecord_block(p, /*maybe_snapshot*/True);
1775    VG_(cli_free)(p);
1776 }
1777 
ms___builtin_delete(ThreadId tid,void * p)1778 static void ms___builtin_delete ( ThreadId tid, void* p )
1779 {
1780    unrecord_block(p, /*maybe_snapshot*/True);
1781    VG_(cli_free)(p);
1782 }
1783 
ms___builtin_vec_delete(ThreadId tid,void * p)1784 static void ms___builtin_vec_delete ( ThreadId tid, void* p )
1785 {
1786    unrecord_block(p, /*maybe_snapshot*/True);
1787    VG_(cli_free)(p);
1788 }
1789 
ms_realloc(ThreadId tid,void * p_old,SizeT new_szB)1790 static void* ms_realloc ( ThreadId tid, void* p_old, SizeT new_szB )
1791 {
1792    return realloc_block(tid, p_old, new_szB);
1793 }
1794 
ms_malloc_usable_size(ThreadId tid,void * p)1795 static SizeT ms_malloc_usable_size ( ThreadId tid, void* p )
1796 {
1797    HP_Chunk* hc = VG_(HT_lookup)( malloc_list, (UWord)p );
1798 
1799    return ( hc ? hc->req_szB + hc->slop_szB : 0 );
1800 }
1801 
1802 //------------------------------------------------------------//
1803 //--- Page handling                                        ---//
1804 //------------------------------------------------------------//
1805 
1806 static
ms_record_page_mem(Addr a,SizeT len)1807 void ms_record_page_mem ( Addr a, SizeT len )
1808 {
1809    ThreadId tid = VG_(get_running_tid)();
1810    Addr end;
1811    tl_assert(VG_IS_PAGE_ALIGNED(len));
1812    tl_assert(len >= VKI_PAGE_SIZE);
1813    // Record the first N-1 pages as blocks, but don't do any snapshots.
1814    for (end = a + len - VKI_PAGE_SIZE; a < end; a += VKI_PAGE_SIZE) {
1815       record_block( tid, (void*)a, VKI_PAGE_SIZE, /*slop_szB*/0,
1816                     /*exclude_first_entry*/False, /*maybe_snapshot*/False );
1817    }
1818    // Record the last page as a block, and maybe do a snapshot afterwards.
1819    record_block( tid, (void*)a, VKI_PAGE_SIZE, /*slop_szB*/0,
1820                  /*exclude_first_entry*/False, /*maybe_snapshot*/True );
1821 }
1822 
1823 static
ms_unrecord_page_mem(Addr a,SizeT len)1824 void ms_unrecord_page_mem( Addr a, SizeT len )
1825 {
1826    Addr end;
1827    tl_assert(VG_IS_PAGE_ALIGNED(len));
1828    tl_assert(len >= VKI_PAGE_SIZE);
1829    for (end = a + len - VKI_PAGE_SIZE; a < end; a += VKI_PAGE_SIZE) {
1830       unrecord_block((void*)a, /*maybe_snapshot*/False);
1831    }
1832    unrecord_block((void*)a, /*maybe_snapshot*/True);
1833 }
1834 
1835 //------------------------------------------------------------//
1836 
1837 static
ms_new_mem_mmap(Addr a,SizeT len,Bool rr,Bool ww,Bool xx,ULong di_handle)1838 void ms_new_mem_mmap ( Addr a, SizeT len,
1839                        Bool rr, Bool ww, Bool xx, ULong di_handle )
1840 {
1841    tl_assert(VG_IS_PAGE_ALIGNED(len));
1842    ms_record_page_mem(a, len);
1843 }
1844 
1845 static
ms_new_mem_startup(Addr a,SizeT len,Bool rr,Bool ww,Bool xx,ULong di_handle)1846 void ms_new_mem_startup( Addr a, SizeT len,
1847                          Bool rr, Bool ww, Bool xx, ULong di_handle )
1848 {
1849    // startup maps are always be page-sized, except the trampoline page is
1850    // marked by the core as only being the size of the trampoline itself,
1851    // which is something like 57 bytes.  Round it up to page size.
1852    len = VG_PGROUNDUP(len);
1853    ms_record_page_mem(a, len);
1854 }
1855 
1856 static
ms_new_mem_brk(Addr a,SizeT len,ThreadId tid)1857 void ms_new_mem_brk ( Addr a, SizeT len, ThreadId tid )
1858 {
1859    // brk limit is not necessarily aligned on a page boundary.
1860    // If new memory being brk-ed implies to allocate a new page,
1861    // then call ms_record_page_mem with page aligned parameters
1862    // otherwise just ignore.
1863    Addr old_bottom_page = VG_PGROUNDDN(a - 1);
1864    Addr new_top_page = VG_PGROUNDDN(a + len - 1);
1865    if (old_bottom_page != new_top_page)
1866       ms_record_page_mem(VG_PGROUNDDN(a),
1867                          (new_top_page - old_bottom_page));
1868 }
1869 
1870 static
ms_copy_mem_remap(Addr from,Addr to,SizeT len)1871 void ms_copy_mem_remap( Addr from, Addr to, SizeT len)
1872 {
1873    tl_assert(VG_IS_PAGE_ALIGNED(len));
1874    ms_unrecord_page_mem(from, len);
1875    ms_record_page_mem(to, len);
1876 }
1877 
1878 static
ms_die_mem_munmap(Addr a,SizeT len)1879 void ms_die_mem_munmap( Addr a, SizeT len )
1880 {
1881    tl_assert(VG_IS_PAGE_ALIGNED(len));
1882    ms_unrecord_page_mem(a, len);
1883 }
1884 
1885 static
ms_die_mem_brk(Addr a,SizeT len)1886 void ms_die_mem_brk( Addr a, SizeT len )
1887 {
1888    // Call ms_unrecord_page_mem only if one or more pages are de-allocated.
1889    // See ms_new_mem_brk for more details.
1890    Addr new_bottom_page = VG_PGROUNDDN(a - 1);
1891    Addr old_top_page = VG_PGROUNDDN(a + len - 1);
1892    if (old_top_page != new_bottom_page)
1893       ms_unrecord_page_mem(VG_PGROUNDDN(a),
1894                            (old_top_page - new_bottom_page));
1895 
1896 }
1897 
1898 //------------------------------------------------------------//
1899 //--- Stacks                                               ---//
1900 //------------------------------------------------------------//
1901 
1902 // We really want the inlining to occur...
1903 #define INLINE    inline __attribute__((always_inline))
1904 
update_stack_stats(SSizeT stack_szB_delta)1905 static void update_stack_stats(SSizeT stack_szB_delta)
1906 {
1907    if (stack_szB_delta < 0) tl_assert(stacks_szB >= -stack_szB_delta);
1908    stacks_szB += stack_szB_delta;
1909 
1910    update_alloc_stats(stack_szB_delta);
1911 }
1912 
new_mem_stack_2(SizeT len,const HChar * what)1913 static INLINE void new_mem_stack_2(SizeT len, const HChar* what)
1914 {
1915    if (have_started_executing_code) {
1916       VERB(3, "<<< new_mem_stack (%ld)\n", len);
1917       n_stack_allocs++;
1918       update_stack_stats(len);
1919       maybe_take_snapshot(Normal, what);
1920       VERB(3, ">>>\n");
1921    }
1922 }
1923 
die_mem_stack_2(SizeT len,const HChar * what)1924 static INLINE void die_mem_stack_2(SizeT len, const HChar* what)
1925 {
1926    if (have_started_executing_code) {
1927       VERB(3, "<<< die_mem_stack (%ld)\n", -len);
1928       n_stack_frees++;
1929       maybe_take_snapshot(Peak,   "stkPEAK");
1930       update_stack_stats(-len);
1931       maybe_take_snapshot(Normal, what);
1932       VERB(3, ">>>\n");
1933    }
1934 }
1935 
new_mem_stack(Addr a,SizeT len)1936 static void new_mem_stack(Addr a, SizeT len)
1937 {
1938    new_mem_stack_2(len, "stk-new");
1939 }
1940 
die_mem_stack(Addr a,SizeT len)1941 static void die_mem_stack(Addr a, SizeT len)
1942 {
1943    die_mem_stack_2(len, "stk-die");
1944 }
1945 
new_mem_stack_signal(Addr a,SizeT len,ThreadId tid)1946 static void new_mem_stack_signal(Addr a, SizeT len, ThreadId tid)
1947 {
1948    new_mem_stack_2(len, "sig-new");
1949 }
1950 
die_mem_stack_signal(Addr a,SizeT len)1951 static void die_mem_stack_signal(Addr a, SizeT len)
1952 {
1953    die_mem_stack_2(len, "sig-die");
1954 }
1955 
1956 
1957 //------------------------------------------------------------//
1958 //--- Client Requests                                      ---//
1959 //------------------------------------------------------------//
1960 
print_monitor_help(void)1961 static void print_monitor_help ( void )
1962 {
1963    VG_(gdb_printf) ("\n");
1964    VG_(gdb_printf) ("massif monitor commands:\n");
1965    VG_(gdb_printf) ("  snapshot [<filename>]\n");
1966    VG_(gdb_printf) ("  detailed_snapshot [<filename>]\n");
1967    VG_(gdb_printf) ("       takes a snapshot (or a detailed snapshot)\n");
1968    VG_(gdb_printf) ("       and saves it in <filename>\n");
1969    VG_(gdb_printf) ("             default <filename> is massif.vgdb.out\n");
1970    VG_(gdb_printf) ("\n");
1971 }
1972 
1973 
1974 /* Forward declaration.
1975    return True if request recognised, False otherwise */
1976 static Bool handle_gdb_monitor_command (ThreadId tid, HChar *req);
ms_handle_client_request(ThreadId tid,UWord * argv,UWord * ret)1977 static Bool ms_handle_client_request ( ThreadId tid, UWord* argv, UWord* ret )
1978 {
1979    switch (argv[0]) {
1980    case VG_USERREQ__MALLOCLIKE_BLOCK: {
1981       void* p   = (void*)argv[1];
1982       SizeT szB =        argv[2];
1983       record_block( tid, p, szB, /*slop_szB*/0, /*exclude_first_entry*/False,
1984                     /*maybe_snapshot*/True );
1985       *ret = 0;
1986       return True;
1987    }
1988    case VG_USERREQ__RESIZEINPLACE_BLOCK: {
1989       void* p        = (void*)argv[1];
1990       SizeT newSizeB =       argv[3];
1991 
1992       unrecord_block(p, /*maybe_snapshot*/True);
1993       record_block(tid, p, newSizeB, /*slop_szB*/0,
1994                    /*exclude_first_entry*/False, /*maybe_snapshot*/True);
1995       return True;
1996    }
1997    case VG_USERREQ__FREELIKE_BLOCK: {
1998       void* p = (void*)argv[1];
1999       unrecord_block(p, /*maybe_snapshot*/True);
2000       *ret = 0;
2001       return True;
2002    }
2003    case VG_USERREQ__GDB_MONITOR_COMMAND: {
2004      Bool handled = handle_gdb_monitor_command (tid, (HChar*)argv[1]);
2005      if (handled)
2006        *ret = 1;
2007      else
2008        *ret = 0;
2009      return handled;
2010    }
2011 
2012    default:
2013       *ret = 0;
2014       return False;
2015    }
2016 }
2017 
2018 //------------------------------------------------------------//
2019 //--- Instrumentation                                      ---//
2020 //------------------------------------------------------------//
2021 
add_counter_update(IRSB * sbOut,Int n)2022 static void add_counter_update(IRSB* sbOut, Int n)
2023 {
2024    #if defined(VG_BIGENDIAN)
2025    # define END Iend_BE
2026    #elif defined(VG_LITTLEENDIAN)
2027    # define END Iend_LE
2028    #else
2029    # error "Unknown endianness"
2030    #endif
2031    // Add code to increment 'guest_instrs_executed' by 'n', like this:
2032    //   WrTmp(t1, Load64(&guest_instrs_executed))
2033    //   WrTmp(t2, Add64(RdTmp(t1), Const(n)))
2034    //   Store(&guest_instrs_executed, t2)
2035    IRTemp t1 = newIRTemp(sbOut->tyenv, Ity_I64);
2036    IRTemp t2 = newIRTemp(sbOut->tyenv, Ity_I64);
2037    IRExpr* counter_addr = mkIRExpr_HWord( (HWord)&guest_instrs_executed );
2038 
2039    IRStmt* st1 = IRStmt_WrTmp(t1, IRExpr_Load(END, Ity_I64, counter_addr));
2040    IRStmt* st2 =
2041       IRStmt_WrTmp(t2,
2042                    IRExpr_Binop(Iop_Add64, IRExpr_RdTmp(t1),
2043                                            IRExpr_Const(IRConst_U64(n))));
2044    IRStmt* st3 = IRStmt_Store(END, counter_addr, IRExpr_RdTmp(t2));
2045 
2046    addStmtToIRSB( sbOut, st1 );
2047    addStmtToIRSB( sbOut, st2 );
2048    addStmtToIRSB( sbOut, st3 );
2049 }
2050 
ms_instrument2(IRSB * sbIn)2051 static IRSB* ms_instrument2( IRSB* sbIn )
2052 {
2053    Int   i, n = 0;
2054    IRSB* sbOut;
2055 
2056    // We increment the instruction count in two places:
2057    // - just before any Ist_Exit statements;
2058    // - just before the IRSB's end.
2059    // In the former case, we zero 'n' and then continue instrumenting.
2060 
2061    sbOut = deepCopyIRSBExceptStmts(sbIn);
2062 
2063    for (i = 0; i < sbIn->stmts_used; i++) {
2064       IRStmt* st = sbIn->stmts[i];
2065 
2066       if (!st || st->tag == Ist_NoOp) continue;
2067 
2068       if (st->tag == Ist_IMark) {
2069          n++;
2070       } else if (st->tag == Ist_Exit) {
2071          if (n > 0) {
2072             // Add an increment before the Exit statement, then reset 'n'.
2073             add_counter_update(sbOut, n);
2074             n = 0;
2075          }
2076       }
2077       addStmtToIRSB( sbOut, st );
2078    }
2079 
2080    if (n > 0) {
2081       // Add an increment before the SB end.
2082       add_counter_update(sbOut, n);
2083    }
2084    return sbOut;
2085 }
2086 
2087 static
ms_instrument(VgCallbackClosure * closure,IRSB * sbIn,VexGuestLayout * layout,VexGuestExtents * vge,VexArchInfo * archinfo_host,IRType gWordTy,IRType hWordTy)2088 IRSB* ms_instrument ( VgCallbackClosure* closure,
2089                       IRSB* sbIn,
2090                       VexGuestLayout* layout,
2091                       VexGuestExtents* vge,
2092                       VexArchInfo* archinfo_host,
2093                       IRType gWordTy, IRType hWordTy )
2094 {
2095    if (! have_started_executing_code) {
2096       // Do an initial sample to guarantee that we have at least one.
2097       // We use 'maybe_take_snapshot' instead of 'take_snapshot' to ensure
2098       // 'maybe_take_snapshot's internal static variables are initialised.
2099       have_started_executing_code = True;
2100       maybe_take_snapshot(Normal, "startup");
2101    }
2102 
2103    if      (clo_time_unit == TimeI)  { return ms_instrument2(sbIn); }
2104    else if (clo_time_unit == TimeMS) { return sbIn; }
2105    else if (clo_time_unit == TimeB)  { return sbIn; }
2106    else                              { tl_assert2(0, "bad --time-unit value"); }
2107 }
2108 
2109 
2110 //------------------------------------------------------------//
2111 //--- Writing snapshots                                    ---//
2112 //------------------------------------------------------------//
2113 
2114 HChar FP_buf[BUF_LEN];
2115 
2116 // XXX: implement f{,n}printf in m_libcprint.c eventually, and use it here.
2117 // Then change Cachegrind to use it too.
2118 #define FP(format, args...) ({ \
2119    VG_(snprintf)(FP_buf, BUF_LEN, format, ##args); \
2120    FP_buf[BUF_LEN-1] = '\0';  /* Make sure the string is terminated. */ \
2121    VG_(write)(fd, (void*)FP_buf, VG_(strlen)(FP_buf)); \
2122 })
2123 
2124 // Nb: uses a static buffer, each call trashes the last string returned.
make_perc(double x)2125 static const HChar* make_perc(double x)
2126 {
2127    static HChar mbuf[32];
2128 
2129    VG_(percentify)((ULong)(x * 100), 10000, 2, 6, mbuf);
2130    // XXX: this is bogus if the denominator was zero -- resulting string is
2131    // something like "0 --%")
2132    if (' ' == mbuf[0]) mbuf[0] = '0';
2133    return mbuf;
2134 }
2135 
pp_snapshot_SXPt(Int fd,SXPt * sxpt,Int depth,HChar * depth_str,Int depth_str_len,SizeT snapshot_heap_szB,SizeT snapshot_total_szB)2136 static void pp_snapshot_SXPt(Int fd, SXPt* sxpt, Int depth, HChar* depth_str,
2137                             Int depth_str_len,
2138                             SizeT snapshot_heap_szB, SizeT snapshot_total_szB)
2139 {
2140    Int   i, j, n_insig_children_sxpts;
2141    SXPt* child = NULL;
2142 
2143    // Used for printing function names.  Is made static to keep it out
2144    // of the stack frame -- this function is recursive.  Obviously this
2145    // now means its contents are trashed across the recursive call.
2146    static HChar ip_desc_array[BUF_LEN];
2147    const HChar* ip_desc = ip_desc_array;
2148 
2149    switch (sxpt->tag) {
2150     case SigSXPt:
2151       // Print the SXPt itself.
2152       if (0 == depth) {
2153          if (clo_heap) {
2154             ip_desc =
2155                ( clo_pages_as_heap
2156                ? "(page allocation syscalls) mmap/mremap/brk, --alloc-fns, etc."
2157                : "(heap allocation functions) malloc/new/new[], --alloc-fns, etc."
2158                );
2159          } else {
2160             // XXX: --alloc-fns?
2161 
2162             // Nick thinks this case cannot happen. ip_desc_array would be
2163             // conceptually uninitialised here. Therefore:
2164             tl_assert2(0, "pp_snapshot_SXPt: unexpected");
2165          }
2166       } else {
2167          // If it's main-or-below-main, we (if appropriate) ignore everything
2168          // below it by pretending it has no children.
2169          if ( ! VG_(clo_show_below_main) ) {
2170             Vg_FnNameKind kind = VG_(get_fnname_kind_from_IP)(sxpt->Sig.ip);
2171             if (Vg_FnNameMain == kind || Vg_FnNameBelowMain == kind) {
2172                sxpt->Sig.n_children = 0;
2173             }
2174          }
2175 
2176          // We need the -1 to get the line number right, But I'm not sure why.
2177          ip_desc = VG_(describe_IP)(sxpt->Sig.ip-1, ip_desc_array, BUF_LEN);
2178       }
2179 
2180       // Do the non-ip_desc part first...
2181       FP("%sn%d: %lu ", depth_str, sxpt->Sig.n_children, sxpt->szB);
2182 
2183       // For ip_descs beginning with "0xABCD...:" addresses, we first
2184       // measure the length of the "0xabcd: " address at the start of the
2185       // ip_desc.
2186       j = 0;
2187       if ('0' == ip_desc[0] && 'x' == ip_desc[1]) {
2188          j = 2;
2189          while (True) {
2190             if (ip_desc[j]) {
2191                if (':' == ip_desc[j]) break;
2192                j++;
2193             } else {
2194                tl_assert2(0, "ip_desc has unexpected form: %s\n", ip_desc);
2195             }
2196          }
2197       }
2198       // Nb: We treat this specially (ie. we don't use FP) so that if the
2199       // ip_desc is too long (eg. due to a long C++ function name), it'll
2200       // get truncated, but the '\n' is still there so its a valid file.
2201       // (At one point we were truncating without adding the '\n', which
2202       // caused bug #155929.)
2203       //
2204       // Also, we account for the length of the address in ip_desc when
2205       // truncating.  (The longest address we could have is 18 chars:  "0x"
2206       // plus 16 address digits.)  This ensures that the truncated function
2207       // name always has the same length, which makes truncation
2208       // deterministic and thus makes testing easier.
2209       tl_assert(j <= 18);
2210       VG_(snprintf)(FP_buf, BUF_LEN, "%s\n", ip_desc);
2211       FP_buf[BUF_LEN-18+j-5] = '.';    // "..." at the end make the
2212       FP_buf[BUF_LEN-18+j-4] = '.';    //   truncation more obvious.
2213       FP_buf[BUF_LEN-18+j-3] = '.';
2214       FP_buf[BUF_LEN-18+j-2] = '\n';   // The last char is '\n'.
2215       FP_buf[BUF_LEN-18+j-1] = '\0';   // The string is terminated.
2216       VG_(write)(fd, (void*)FP_buf, VG_(strlen)(FP_buf));
2217 
2218       // Indent.
2219       tl_assert(depth+1 < depth_str_len-1);    // -1 for end NUL char
2220       depth_str[depth+0] = ' ';
2221       depth_str[depth+1] = '\0';
2222 
2223       // Sort SXPt's children by szB (reverse order:  biggest to smallest).
2224       // Nb: we sort them here, rather than earlier (eg. in dup_XTree), for
2225       // two reasons.  First, if we do it during dup_XTree, it can get
2226       // expensive (eg. 15% of execution time for konqueror
2227       // startup/shutdown).  Second, this way we get the Insig SXPt (if one
2228       // is present) in its sorted position, not at the end.
2229       VG_(ssort)(sxpt->Sig.children, sxpt->Sig.n_children, sizeof(SXPt*),
2230                  SXPt_revcmp_szB);
2231 
2232       // Print the SXPt's children.  They should already be in sorted order.
2233       n_insig_children_sxpts = 0;
2234       for (i = 0; i < sxpt->Sig.n_children; i++) {
2235          child = sxpt->Sig.children[i];
2236 
2237          if (InsigSXPt == child->tag)
2238             n_insig_children_sxpts++;
2239 
2240          // Ok, print the child.  NB: contents of ip_desc_array will be
2241          // trashed by this recursive call.  Doesn't matter currently,
2242          // but worth noting.
2243          pp_snapshot_SXPt(fd, child, depth+1, depth_str, depth_str_len,
2244             snapshot_heap_szB, snapshot_total_szB);
2245       }
2246 
2247       // Unindent.
2248       depth_str[depth+0] = '\0';
2249       depth_str[depth+1] = '\0';
2250 
2251       // There should be 0 or 1 Insig children SXPts.
2252       tl_assert(n_insig_children_sxpts <= 1);
2253       break;
2254 
2255     case InsigSXPt: {
2256       const HChar* s = ( 1 == sxpt->Insig.n_xpts ? "," : "s, all" );
2257       FP("%sn0: %lu in %d place%s below massif's threshold (%s)\n",
2258          depth_str, sxpt->szB, sxpt->Insig.n_xpts, s,
2259          make_perc(clo_threshold));
2260       break;
2261     }
2262 
2263     default:
2264       tl_assert2(0, "pp_snapshot_SXPt: unrecognised SXPt tag");
2265    }
2266 }
2267 
pp_snapshot(Int fd,Snapshot * snapshot,Int snapshot_n)2268 static void pp_snapshot(Int fd, Snapshot* snapshot, Int snapshot_n)
2269 {
2270    sanity_check_snapshot(snapshot);
2271 
2272    FP("#-----------\n");
2273    FP("snapshot=%d\n", snapshot_n);
2274    FP("#-----------\n");
2275    FP("time=%lld\n",            snapshot->time);
2276    FP("mem_heap_B=%lu\n",       snapshot->heap_szB);
2277    FP("mem_heap_extra_B=%lu\n", snapshot->heap_extra_szB);
2278    FP("mem_stacks_B=%lu\n",     snapshot->stacks_szB);
2279 
2280    if (is_detailed_snapshot(snapshot)) {
2281       // Detailed snapshot -- print heap tree.
2282       Int   depth_str_len = clo_depth + 3;
2283       HChar* depth_str = VG_(malloc)("ms.main.pps.1",
2284                                      sizeof(HChar) * depth_str_len);
2285       SizeT snapshot_total_szB =
2286          snapshot->heap_szB + snapshot->heap_extra_szB + snapshot->stacks_szB;
2287       depth_str[0] = '\0';   // Initialise depth_str to "".
2288 
2289       FP("heap_tree=%s\n", ( Peak == snapshot->kind ? "peak" : "detailed" ));
2290       pp_snapshot_SXPt(fd, snapshot->alloc_sxpt, 0, depth_str,
2291                        depth_str_len, snapshot->heap_szB,
2292                        snapshot_total_szB);
2293 
2294       VG_(free)(depth_str);
2295 
2296    } else {
2297       FP("heap_tree=empty\n");
2298    }
2299 }
2300 
write_snapshots_to_file(const HChar * massif_out_file,Snapshot snapshots_array[],Int nr_elements)2301 static void write_snapshots_to_file(const HChar* massif_out_file,
2302                                     Snapshot snapshots_array[],
2303                                     Int nr_elements)
2304 {
2305    Int i, fd;
2306    SysRes sres;
2307 
2308    sres = VG_(open)(massif_out_file, VKI_O_CREAT|VKI_O_TRUNC|VKI_O_WRONLY,
2309                                      VKI_S_IRUSR|VKI_S_IWUSR);
2310    if (sr_isError(sres)) {
2311       // If the file can't be opened for whatever reason (conflict
2312       // between multiple cachegrinded processes?), give up now.
2313       VG_(umsg)("error: can't open output file '%s'\n", massif_out_file );
2314       VG_(umsg)("       ... so profiling results will be missing.\n");
2315       return;
2316    } else {
2317       fd = sr_Res(sres);
2318    }
2319 
2320    // Print massif-specific options that were used.
2321    // XXX: is it worth having a "desc:" line?  Could just call it "options:"
2322    // -- this file format isn't as generic as Cachegrind's, so the
2323    // implied genericity of "desc:" is bogus.
2324    FP("desc:");
2325    for (i = 0; i < VG_(sizeXA)(args_for_massif); i++) {
2326       HChar* arg = *(HChar**)VG_(indexXA)(args_for_massif, i);
2327       FP(" %s", arg);
2328    }
2329    if (0 == i) FP(" (none)");
2330    FP("\n");
2331 
2332    // Print "cmd:" line.
2333    FP("cmd: ");
2334    if (VG_(args_the_exename)) {
2335       FP("%s", VG_(args_the_exename));
2336       for (i = 0; i < VG_(sizeXA)( VG_(args_for_client) ); i++) {
2337          HChar* arg = * (HChar**) VG_(indexXA)( VG_(args_for_client), i );
2338          if (arg)
2339             FP(" %s", arg);
2340       }
2341    } else {
2342       FP(" ???");
2343    }
2344    FP("\n");
2345 
2346    FP("time_unit: %s\n", TimeUnit_to_string(clo_time_unit));
2347 
2348    for (i = 0; i < nr_elements; i++) {
2349       Snapshot* snapshot = & snapshots_array[i];
2350       pp_snapshot(fd, snapshot, i);     // Detailed snapshot!
2351    }
2352    VG_(close) (fd);
2353 }
2354 
write_snapshots_array_to_file(void)2355 static void write_snapshots_array_to_file(void)
2356 {
2357    // Setup output filename.  Nb: it's important to do this now, ie. as late
2358    // as possible.  If we do it at start-up and the program forks and the
2359    // output file format string contains a %p (pid) specifier, both the
2360    // parent and child will incorrectly write to the same file;  this
2361    // happened in 3.3.0.
2362    HChar* massif_out_file =
2363       VG_(expand_file_name)("--massif-out-file", clo_massif_out_file);
2364    write_snapshots_to_file (massif_out_file, snapshots, next_snapshot_i);
2365    VG_(free)(massif_out_file);
2366 }
2367 
handle_snapshot_monitor_command(const HChar * filename,Bool detailed)2368 static void handle_snapshot_monitor_command (const HChar *filename,
2369                                              Bool detailed)
2370 {
2371    Snapshot snapshot;
2372 
2373    if (!clo_pages_as_heap && !have_started_executing_code) {
2374       // See comments of variable have_started_executing_code.
2375       VG_(gdb_printf)
2376          ("error: cannot take snapshot before execution has started\n");
2377       return;
2378    }
2379 
2380    clear_snapshot(&snapshot, /* do_sanity_check */ False);
2381    take_snapshot(&snapshot, Normal, get_time(), detailed);
2382    write_snapshots_to_file ((filename == NULL) ?
2383                             "massif.vgdb.out" : filename,
2384                             &snapshot,
2385                             1);
2386    delete_snapshot(&snapshot);
2387 }
2388 
handle_gdb_monitor_command(ThreadId tid,HChar * req)2389 static Bool handle_gdb_monitor_command (ThreadId tid, HChar *req)
2390 {
2391    HChar* wcmd;
2392    HChar s[VG_(strlen(req)) + 1]; /* copy for strtok_r */
2393    HChar *ssaveptr;
2394 
2395    VG_(strcpy) (s, req);
2396 
2397    wcmd = VG_(strtok_r) (s, " ", &ssaveptr);
2398    switch (VG_(keyword_id) ("help snapshot detailed_snapshot",
2399                             wcmd, kwd_report_duplicated_matches)) {
2400    case -2: /* multiple matches */
2401       return True;
2402    case -1: /* not found */
2403       return False;
2404    case  0: /* help */
2405       print_monitor_help();
2406       return True;
2407    case  1: { /* snapshot */
2408       HChar* filename;
2409       filename = VG_(strtok_r) (NULL, " ", &ssaveptr);
2410       handle_snapshot_monitor_command (filename, False /* detailed */);
2411       return True;
2412    }
2413    case  2: { /* detailed_snapshot */
2414       HChar* filename;
2415       filename = VG_(strtok_r) (NULL, " ", &ssaveptr);
2416       handle_snapshot_monitor_command (filename, True /* detailed */);
2417       return True;
2418    }
2419    default:
2420       tl_assert(0);
2421       return False;
2422    }
2423 }
2424 
ms_print_stats(void)2425 static void ms_print_stats (void)
2426 {
2427 #define STATS(format, args...) \
2428       VG_(dmsg)("Massif: " format, ##args)
2429 
2430    STATS("heap allocs:           %u\n", n_heap_allocs);
2431    STATS("heap reallocs:         %u\n", n_heap_reallocs);
2432    STATS("heap frees:            %u\n", n_heap_frees);
2433    STATS("ignored heap allocs:   %u\n", n_ignored_heap_allocs);
2434    STATS("ignored heap frees:    %u\n", n_ignored_heap_frees);
2435    STATS("ignored heap reallocs: %u\n", n_ignored_heap_reallocs);
2436    STATS("stack allocs:          %u\n", n_stack_allocs);
2437    STATS("stack frees:           %u\n", n_stack_frees);
2438    STATS("XPts:                  %u\n", n_xpts);
2439    STATS("top-XPts:              %u (%d%%)\n",
2440       alloc_xpt->n_children,
2441       ( n_xpts ? alloc_xpt->n_children * 100 / n_xpts : 0));
2442    STATS("XPt init expansions:   %u\n", n_xpt_init_expansions);
2443    STATS("XPt later expansions:  %u\n", n_xpt_later_expansions);
2444    STATS("SXPt allocs:           %u\n", n_sxpt_allocs);
2445    STATS("SXPt frees:            %u\n", n_sxpt_frees);
2446    STATS("skipped snapshots:     %u\n", n_skipped_snapshots);
2447    STATS("real snapshots:        %u\n", n_real_snapshots);
2448    STATS("detailed snapshots:    %u\n", n_detailed_snapshots);
2449    STATS("peak snapshots:        %u\n", n_peak_snapshots);
2450    STATS("cullings:              %u\n", n_cullings);
2451    STATS("XCon redos:            %u\n", n_XCon_redos);
2452 #undef STATS
2453 }
2454 
2455 //------------------------------------------------------------//
2456 //--- Finalisation                                         ---//
2457 //------------------------------------------------------------//
2458 
ms_fini(Int exit_status)2459 static void ms_fini(Int exit_status)
2460 {
2461    // Output.
2462    write_snapshots_array_to_file();
2463 
2464    // Stats
2465    tl_assert(n_xpts > 0);  // always have alloc_xpt
2466 
2467    if (VG_(clo_stats))
2468       ms_print_stats();
2469 }
2470 
2471 
2472 //------------------------------------------------------------//
2473 //--- Initialisation                                       ---//
2474 //------------------------------------------------------------//
2475 
ms_post_clo_init(void)2476 static void ms_post_clo_init(void)
2477 {
2478    Int i;
2479    HChar* LD_PRELOAD_val;
2480    HChar* s;
2481    HChar* s2;
2482 
2483    // Check options.
2484    if (clo_pages_as_heap) {
2485       if (clo_stacks) {
2486          VG_(fmsg_bad_option)(
2487             "--pages-as-heap=yes together with --stacks=yes", "");
2488       }
2489    }
2490    if (!clo_heap) {
2491       clo_pages_as_heap = False;
2492    }
2493 
2494    // If --pages-as-heap=yes we don't want malloc replacement to occur.  So we
2495    // disable vgpreload_massif-$PLATFORM.so by removing it from LD_PRELOAD (or
2496    // platform-equivalent).  We replace it entirely with spaces because then
2497    // the linker doesn't complain (it does complain if we just change the name
2498    // to a bogus file).  This is a bit of a hack, but LD_PRELOAD is setup well
2499    // before tool initialisation, so this seems the best way to do it.
2500    if (clo_pages_as_heap) {
2501       clo_heap_admin = 0;     // No heap admin on pages.
2502 
2503       LD_PRELOAD_val = VG_(getenv)( VG_(LD_PRELOAD_var_name) );
2504       tl_assert(LD_PRELOAD_val);
2505 
2506       // Make sure the vgpreload_core-$PLATFORM entry is there, for sanity.
2507       s2 = VG_(strstr)(LD_PRELOAD_val, "vgpreload_core");
2508       tl_assert(s2);
2509 
2510       // Now find the vgpreload_massif-$PLATFORM entry.
2511       s2 = VG_(strstr)(LD_PRELOAD_val, "vgpreload_massif");
2512       tl_assert(s2);
2513 
2514       // Blank out everything to the previous ':', which must be there because
2515       // of the preceding vgpreload_core-$PLATFORM entry.
2516       for (s = s2; *s != ':'; s--) {
2517          *s = ' ';
2518       }
2519 
2520       // Blank out everything to the end of the entry, which will be '\0' if
2521       // LD_PRELOAD was empty before Valgrind started, or ':' otherwise.
2522       for (s = s2; *s != ':' && *s != '\0'; s++) {
2523          *s = ' ';
2524       }
2525    }
2526 
2527    // Print alloc-fns and ignore-fns, if necessary.
2528    if (VG_(clo_verbosity) > 1) {
2529       VERB(1, "alloc-fns:\n");
2530       for (i = 0; i < VG_(sizeXA)(alloc_fns); i++) {
2531          HChar** fn_ptr = VG_(indexXA)(alloc_fns, i);
2532          VERB(1, "  %s\n", *fn_ptr);
2533       }
2534 
2535       VERB(1, "ignore-fns:\n");
2536       if (0 == VG_(sizeXA)(ignore_fns)) {
2537          VERB(1, "  <empty>\n");
2538       }
2539       for (i = 0; i < VG_(sizeXA)(ignore_fns); i++) {
2540          HChar** fn_ptr = VG_(indexXA)(ignore_fns, i);
2541          VERB(1, "  %d: %s\n", i, *fn_ptr);
2542       }
2543    }
2544 
2545    // Events to track.
2546    if (clo_stacks) {
2547       VG_(track_new_mem_stack)        ( new_mem_stack        );
2548       VG_(track_die_mem_stack)        ( die_mem_stack        );
2549       VG_(track_new_mem_stack_signal) ( new_mem_stack_signal );
2550       VG_(track_die_mem_stack_signal) ( die_mem_stack_signal );
2551    }
2552 
2553    if (clo_pages_as_heap) {
2554       VG_(track_new_mem_startup) ( ms_new_mem_startup );
2555       VG_(track_new_mem_brk)     ( ms_new_mem_brk     );
2556       VG_(track_new_mem_mmap)    ( ms_new_mem_mmap    );
2557 
2558       VG_(track_copy_mem_remap)  ( ms_copy_mem_remap  );
2559 
2560       VG_(track_die_mem_brk)     ( ms_die_mem_brk     );
2561       VG_(track_die_mem_munmap)  ( ms_die_mem_munmap  );
2562    }
2563 
2564    // Initialise snapshot array, and sanity-check it.
2565    snapshots = VG_(malloc)("ms.main.mpoci.1",
2566                            sizeof(Snapshot) * clo_max_snapshots);
2567    // We don't want to do snapshot sanity checks here, because they're
2568    // currently uninitialised.
2569    for (i = 0; i < clo_max_snapshots; i++) {
2570       clear_snapshot( & snapshots[i], /*do_sanity_check*/False );
2571    }
2572    sanity_check_snapshots_array();
2573 }
2574 
ms_pre_clo_init(void)2575 static void ms_pre_clo_init(void)
2576 {
2577    VG_(details_name)            ("Massif");
2578    VG_(details_version)         (NULL);
2579    VG_(details_description)     ("a heap profiler");
2580    VG_(details_copyright_author)(
2581       "Copyright (C) 2003-2013, and GNU GPL'd, by Nicholas Nethercote");
2582    VG_(details_bug_reports_to)  (VG_BUGS_TO);
2583 
2584    VG_(details_avg_translation_sizeB) ( 330 );
2585 
2586    VG_(clo_vex_control).iropt_register_updates
2587       = VexRegUpdSpAtMemAccess; // overridable by the user.
2588 
2589    // Basic functions.
2590    VG_(basic_tool_funcs)          (ms_post_clo_init,
2591                                    ms_instrument,
2592                                    ms_fini);
2593 
2594    // Needs.
2595    VG_(needs_libc_freeres)();
2596    VG_(needs_command_line_options)(ms_process_cmd_line_option,
2597                                    ms_print_usage,
2598                                    ms_print_debug_usage);
2599    VG_(needs_client_requests)     (ms_handle_client_request);
2600    VG_(needs_sanity_checks)       (ms_cheap_sanity_check,
2601                                    ms_expensive_sanity_check);
2602    VG_(needs_print_stats)         (ms_print_stats);
2603    VG_(needs_malloc_replacement)  (ms_malloc,
2604                                    ms___builtin_new,
2605                                    ms___builtin_vec_new,
2606                                    ms_memalign,
2607                                    ms_calloc,
2608                                    ms_free,
2609                                    ms___builtin_delete,
2610                                    ms___builtin_vec_delete,
2611                                    ms_realloc,
2612                                    ms_malloc_usable_size,
2613                                    0 );
2614 
2615    // HP_Chunks.
2616    malloc_list = VG_(HT_construct)( "Massif's malloc list" );
2617 
2618    // Dummy node at top of the context structure.
2619    alloc_xpt = new_XPt(/*ip*/0, /*parent*/NULL);
2620 
2621    // Initialise alloc_fns and ignore_fns.
2622    init_alloc_fns();
2623    init_ignore_fns();
2624 
2625    // Initialise args_for_massif.
2626    args_for_massif = VG_(newXA)(VG_(malloc), "ms.main.mprci.1",
2627                                 VG_(free), sizeof(HChar*));
2628 }
2629 
2630 VG_DETERMINE_INTERFACE_VERSION(ms_pre_clo_init)
2631 
2632 //--------------------------------------------------------------------//
2633 //--- end                                                          ---//
2634 //--------------------------------------------------------------------//
2635