mpool opt

git-svn: trunk@4354
2026-05-07 15:15:40 -04:00 · 2008-11-08 05:34:25 +00:00
parent 08402afa4c
commit b26d0ef3fd
3 changed files with 270 additions and 237 deletions
--- a/5
+++ b/5
@@ -1,3 +1,8 @@
+Sat Nov  8 05:27:43 CET 2008 (acab)
+-----------------------------------
+ * libclamav/mpool.c: slightly optimized (we're now faster and more compact
+		      than malloc() - it's just because we cheat more :)
+
 Fri Nov  7 19:37:38 EET 2008 (edwin)
 ------------------------------------
 * clamd/, libclamav/, libclamunrar/, unit_tests/: fix more compiler
--- a/contrib/mpoolparse/mpoolparse.pl
+++ b/contrib/mpoolparse/mpoolparse.pl
@@ -0,0 +1,17 @@
+#!/usr/bin/perl
+
+use strict;
+use warnings;
+
+my %frags = ();
+
+while (<>) {
+	chomp;
+	next unless /^LibClamAV Warning: [mc]alloc .* size (\d+) .*$/;
+	$frags{$1}++;
+}
+
+foreach (sort {$a<=>$b} (keys(%frags))) {
+	print "$_, /* ($frags{$_}) */\n";
+}
+
--- a/libclamav/mpool.c
+++ b/libclamav/mpool.c
@@ -44,9 +44,8 @@
 #include <assert.h>

 #define MPOOLMAGIC 0x5adeada5
-/*#define DEBUGMPOOL /\* DO NOT define *\/ */
+/* #define DEBUGMPOOL /\* DO NOT define *\/ */
 #ifdef DEBUGMPOOL
-FILE *lfd = NULL;
 #define spam(...) cli_warnmsg( __VA_ARGS__)
 #else
 #define spam
@@ -68,85 +67,94 @@ FILE *lfd = NULL;

 #if SIZEOF_VOID_P==8
 static const unsigned int fragsz[] = {
-24, /* (33067) */
-32, /* (93970) */
-40, /* (731473) */
-48, /* (35342) */
-56, /* (6967) */
-64, /* (442053) */
-72, /* (2706) */
-80, /* (2281) */
-88, /* (3658) */
-96, /* (2620) */
-104, /* (3705) */
-112, /* (3722) */
-120, /* (4012) */
-128, /* (5161) */
-136, /* (22458) */
-144, /* (4339) */
-152, /* (4315) */
-160, /* (12195) */
-168, /* (18042) */
-/* 176, /\* (341) *\/ */
-/* 184, /\* (329) *\/ */
-192, /* (202) */
-/* 200, /\* (238) *\/ */
-/* 208, /\* (166) *\/ */
-/* 216, /\* (179) *\/ */
-224, /* (216) */
-/* 232, /\* (189) *\/ */
-/* 240, /\* (157) *\/ */
-/* 248, /\* (178) *\/ */
-256, /* (133) */
-/* 264, /\* (157) *\/ */
-/* 272, /\* (140) *\/ */
-/* 280, /\* (161) *\/ */
-/* 288, /\* (125) *\/ */
-/* 296, /\* (141) *\/ */
-304, /* (100) */
-/* 312, /\* (114) *\/ */
-/* 320, /\* (70) *\/ */
-/* 328, /\* (13) *\/ */
-/* 336, /\* (21) *\/ */
-/* 344, /\* (21) *\/ */
-352, /* (13) */
-/* 360, /\* (26) *\/ */
-/* 368, /\* (8) *\/ */
-/* 376, /\* (14) *\/ */
-/* 384, /\* (5) *\/ */
-/* 392, /\* (6) *\/ */
-/* 400, /\* (4) *\/ */
-/* 408, /\* (2) *\/ */
-/* 416, /\* (6) *\/ */
-424, /* (20) */
+16, /* (37189) */
+24, /* (94629) */
+32, /* (756188) */
+40, /* (35658) */
+48, /* (6989) */
+56, /* (454597) */
+64, /* (2736) */
+72, /* (2298) */
+80, /* (3664) */
+88, /* (2636) */
+96, /* (3715) */
+104, /* (3732) */
+112, /* (4021) */
+120, /* (5165) */
+128, /* (22788) */
+136, /* (4343) */
+144, /* (4321) */
+152, /* (12196) */
+160, /* (18044) */
+/* 168, /\* (344) *\/ */
+/* 176, /\* (335) *\/ */
+/* 184, /\* (204) *\/ */
+192, /* (226) */
+/* 200, /\* (185) *\/ */
+/* 208, /\* (181) *\/ */
+/* 216, /\* (217) *\/ */
+224, /* (172) */
+/* 232, /\* (175) *\/ */
+/* 240, /\* (178) *\/ */
+/* 248, /\* (135) *\/ */
+256, /* (140) */
+/* 264, /\* (158) *\/ */
+/* 272, /\* (161) *\/ */
+/* 280, /\* (126) *\/ */
+/* 288, /\* (141) *\/ */
+296, /* (101) */
+/* 304, /\* (97) *\/ */
+/* 312, /\* (85) *\/ */
+/* 320, /\* (14) *\/ */
+/* 328, /\* (21) *\/ */
+336, /* (21) */
+/* 344, /\* (12) *\/ */
+/* 352, /\* (10) *\/ */
+/* 360, /\* (7) *\/ */
+/* 368, /\* (14) *\/ */
+376, /* (21) */
+/* 384, /\* (6) *\/ */
+/* 392, /\* (4) *\/ */
+/* 400, /\* (3) *\/ */
+/* 408, /\* (6) *\/ */
+/* 416, /\* (5) *\/ */
+/* 424, /\* (5) *\/ */
 /* 432, /\* (4) *\/ */
-/* 440, /\* (4) *\/ */
-/* 448, /\* (4) *\/ */
-/* 464, /\* (2) *\/ */
-/* 472, /\* (2) *\/ */
-/* 480, /\* (1) *\/ */
-/* 496, /\* (1) *\/ */
-/* 512, /\* (2) *\/ */
-528, /* (15) */
-/* 536, /\* (1) *\/ */
-/* 544, /\* (2) *\/ */
-/* 552, /\* (1) *\/ */
-/* 584, /\* (3) *\/ */
-/* 600, /\* (1) *\/ */
-/* 624, /\* (1) *\/ */
+440, /* (19) */
+/* 456, /\* (2) *\/ */
+/* 464, /\* (8) *\/ */
+/* 472, /\* (3) *\/ */
+/* 488, /\* (1) *\/ */
+/* 496, /\* (4) *\/ */
+/* 504, /\* (3) *\/ */
+/* 512, /\* (1) *\/ */
+/* 520, /\* (2) *\/ */
+/* 528, /\* (6) *\/ */
+/* 536, /\* (3) *\/ */
+544, /* (16) */
+/* 560, /\* (4) *\/ */
+/* 576, /\* (3) *\/ */
+/* 592, /\* (8) *\/ */
+/* 616, /\* (2) *\/ */
+/* 624, /\* (5) *\/ */
+/* 648, /\* (1) *\/ */
 /* 656, /\* (1) *\/ */
-/* 784, /\* (2) *\/ */
-1040, /* (15) */
-/* 2064, /\* (7456) *\/ */
-2072, /* (14) */
-4112, /* (14) */
-8208, /* (9) */
-16400, /* (6) */
-32784, /* (4) */
-63512, /* (7) */
-134408, /* (2) */
-507984, /* (7) */
-1051040, /* (1) */
+/* 680, /\* (1) *\/ */
+/* 704, /\* (1) *\/ */
+/* 720, /\* (1) *\/ */
+/* 776, /\* (2) *\/ */
+1056, /* (15) */
+2056, /* (7585) */
+/* 2064, /\* (14) *\/ */
+2088, /* (14) */
+4128, /* (14) */
+8224, /* (9) */
+16416, /* (6) */
+32800, /* (4) */
+63504, /* (7) */
+136352, /* (1) */
+507976, /* (7) */
+1051032, /* (1) */
 2097152
 /* ^^ This shouldn't be reached but it's a good fall back
 * MAX_ALLOCATION is 184549376 but that's really not need here */
@@ -155,150 +163,149 @@ static const unsigned int fragsz[] = {
 #else

 static const unsigned int fragsz[] = {
-12, /* (2297) */
-16, /* (30785) */
-20, /* (41460) */
-24, /* (69214) */
-28, /* (639488) */
-32, /* (107920) */
-36, /* (454213) */
-40, /* (11497) */
-44, /* (1688) */
-48, /* (5294) */
-52, /* (1496) */
-56, /* (3738) */
-60, /* (1719) */
-64, /* (918) */
-68, /* (956) */
-72, /* (1324) */
-76, /* (1905) */
-80, /* (1745) */
-84, /* (1053) */
-88, /* (1566) */
-92, /* (2081) */
-96, /* (20851) */
-100, /* (1882) */
-104, /* (1848) */
-108, /* (1931) */
-112, /* (2079) */
-116, /* (1736) */
-120, /* (3425) */
-124, /* (2115) */
-128, /* (1298) */
-132, /* (2307) */
-136, /* (2033) */
-140, /* (2837) */
-144, /* (1479) */
-148, /* (1607) */
-152, /* (10587) */
-156, /* (2719) */
-160, /* (15311) */
-164, /* (196) */
-168, /* (145) */
-172, /* (211) */
-176, /* (140) */
-180, /* (116) */
-/* 184, /\* (86) *\/ */
-188, /* (119) */
-192, /* (104) */
-/* 196, /\* (99) *\/ */
-/* 200, /\* (84) *\/ */
-/* 204, /\* (94) *\/ */
-/* 208, /\* (86) *\/ */
-212, /* (136) */
-/* 216, /\* (80) *\/ */
-/* 220, /\* (75) *\/ */
-/* 224, /\* (97) *\/ */
-/* 228, /\* (99) *\/ */
-/* 232, /\* (74) *\/ */
-236, /* (114) */
-/* 240, /\* (64) *\/ */
-/* 244, /\* (73) *\/ */
-/* 248, /\* (62) *\/ */
-/* 252, /\* (71) *\/ */
-/* 256, /\* (69) *\/ */
-/* 260, /\* (85) *\/ */
-/* 264, /\* (71) *\/ */
-268, /* (92) */
-/* 272, /\* (69) *\/ */
-/* 276, /\* (56) *\/ */
-/* 280, /\* (69) *\/ */
-/* 284, /\* (71) *\/ */
-/* 288, /\* (70) *\/ */
-/* 292, /\* (62) *\/ */
-/* 296, /\* (39) *\/ */
-/* 300, /\* (54) *\/ */
-/* 304, /\* (43) *\/ */
-/* 308, /\* (54) *\/ */
-312, /* (30) */
-/* 316, /\* (8) *\/ */
-/* 320, /\* (5) *\/ */
-/* 324, /\* (7) *\/ */
-/* 328, /\* (14) *\/ */
-/* 332, /\* (13) *\/ */
-/* 336, /\* (8) *\/ */
-/* 340, /\* (7) *\/ */
-/* 344, /\* (6) *\/ */
-/* 348, /\* (2) *\/ */
-/* 352, /\* (7) *\/ */
-/* 356, /\* (18) *\/ */
-/* 360, /\* (5) *\/ */
-364, /* (12) */
-/* 368, /\* (2) *\/ */
-/* 372, /\* (4) *\/ */
-/* 376, /\* (2) *\/ */
-/* 380, /\* (5) *\/ */
-/* 384, /\* (1) *\/ */
-/* 392, /\* (4) *\/ */
-/* 396, /\* (3) *\/ */
-/* 404, /\* (4) *\/ */
-/* 408, /\* (2) *\/ */
-/* 412, /\* (3) *\/ */
-/* 416, /\* (2) *\/ */
-/* 420, /\* (3) *\/ */
-/* 424, /\* (2) *\/ */
-428, /* (16) */
-/* 432, /\* (4) *\/ */
-/* 436, /\* (1) *\/ */
-/* 440, /\* (3) *\/ */
-/* 452, /\* (1) *\/ */
-/* 456, /\* (2) *\/ */
-/* 460, /\* (8) *\/ */
-/* 468, /\* (1) *\/ */
-/* 472, /\* (2) *\/ */
-/* 484, /\* (1) *\/ */
-/* 492, /\* (4) *\/ */
-/* 500, /\* (1) *\/ */
-/* 504, /\* (2) *\/ */
-/* 508, /\* (1) *\/ */
-/* 516, /\* (2) *\/ */
-/* 524, /\* (5) *\/ */
-532, /* (15) */
-/* 536, /\* (3) *\/ */
-/* 540, /\* (1) *\/ */
-/* 556, /\* (4) *\/ */
-/* 576, /\* (3) *\/ */
-/* 588, /\* (8) *\/ */
+8, /* (6381) */
+12, /* (30903) */
+16, /* (41616) */
+20, /* (69507) */
+24, /* (646497) */
+28, /* (108677) */
+32, /* (458074) */
+36, /* (11537) */
+40, /* (1690) */
+44, /* (5326) */
+48, /* (1505) */
+52, /* (3777) */
+56, /* (1730) */
+60, /* (943) */
+64, /* (963) */
+68, /* (1345) */
+72, /* (1909) */
+76, /* (1765) */
+80, /* (1060) */
+84, /* (1586) */
+88, /* (2082) */
+92, /* (21009) */
+96, /* (1886) */
+100, /* (1869) */
+104, /* (1936) */
+108, /* (2097) */
+112, /* (1736) */
+116, /* (3442) */
+120, /* (2117) */
+124, /* (1317) */
+128, /* (2307) */
+132, /* (2051) */
+136, /* (2839) */
+140, /* (1497) */
+144, /* (1607) */
+148, /* (10604) */
+152, /* (2719) */
+156, /* (15328) */
+160, /* (197) */
+/* 164, /\* (161) *\/ */
+/* 168, /\* (195) *\/ */
+172, /* (156) */
+/* 176, /\* (132) *\/ */
+/* 180, /\* (86) *\/ */
+/* 184, /\* (120) *\/ */
+188, /* (122) */
+/* 192, /\* (84) *\/ */
+/* 196, /\* (84) *\/ */
+/* 200, /\* (95) *\/ */
+/* 204, /\* (86) *\/ */
+208, /* (137) */
+/* 212, /\* (95) *\/ */
+/* 216, /\* (75) *\/ */
+/* 220, /\* (97) *\/ */
+/* 224, /\* (84) *\/ */
+/* 228, /\* (74) *\/ */
+232, /* (114) */
+/* 236, /\* (79) *\/ */
+/* 240, /\* (73) *\/ */
+/* 244, /\* (62) *\/ */
+/* 248, /\* (71) *\/ */
+/* 252, /\* (69) *\/ */
+256, /* (70) */
+/* 260, /\* (86) *\/ */
+/* 264, /\* (92) *\/ */
+/* 268, /\* (69) *\/ */
+/* 272, /\* (56) *\/ */
+/* 276, /\* (70) *\/ */
+280, /* (71) */
+/* 284, /\* (70) *\/ */
+/* 288, /\* (62) *\/ */
+/* 292, /\* (54) *\/ */
+/* 296, /\* (54) *\/ */
+/* 300, /\* (43) *\/ */
+/* 304, /\* (39) *\/ */
+308, /* (30) */
+/* 312, /\* (9) *\/ */
+/* 316, /\* (5) *\/ */
+/* 320, /\* (7) *\/ */
+/* 324, /\* (14) *\/ */
+/* 328, /\* (13) *\/ */
+332, /* (23) */
+/* 336, /\* (6) *\/ */
+/* 340, /\* (6) *\/ */
+/* 344, /\* (3) *\/ */
+/* 348, /\* (7) *\/ */
+/* 352, /\* (3) *\/ */
+/* 356, /\* (4) *\/ */
+/* 360, /\* (12) *\/ */
+/* 364, /\* (2) *\/ */
+/* 368, /\* (3) *\/ */
+372, /* (17) */
+/* 376, /\* (5) *\/ */
+/* 380, /\* (1) *\/ */
+/* 388, /\* (4) *\/ */
+/* 392, /\* (3) *\/ */
+/* 400, /\* (4) *\/ */
+/* 404, /\* (2) *\/ */
+/* 408, /\* (3) *\/ */
+/* 412, /\* (2) *\/ */
+/* 416, /\* (3) *\/ */
+/* 420, /\* (2) *\/ */
+/* 428, /\* (4) *\/ */
+/* 432, /\* (1) *\/ */
+436, /* (18) */
+/* 452, /\* (2) *\/ */
+/* 456, /\* (8) *\/ */
+/* 464, /\* (1) *\/ */
+/* 468, /\* (2) *\/ */
+/* 480, /\* (1) *\/ */
+/* 488, /\* (4) *\/ */
+/* 496, /\* (1) *\/ */
+/* 500, /\* (2) *\/ */
+/* 504, /\* (1) *\/ */
+/* 512, /\* (2) *\/ */
+/* 520, /\* (6) *\/ */
+/* 532, /\* (3) *\/ */
+/* 536, /\* (1) *\/ */
+540, /* (15) */
+/* 552, /\* (4) *\/ */
+/* 572, /\* (3) *\/ */
+/* 584, /\* (7) *\/ */
+/* 588, /\* (1) *\/ */
+/* 608, /\* (1) *\/ */
 /* 612, /\* (1) *\/ */
-/* 616, /\* (1) *\/ */
-/* 620, /\* (5) *\/ */
+/* 616, /\* (5) *\/ */
+/* 644, /\* (1) *\/ */
 /* 648, /\* (1) *\/ */
-/* 652, /\* (1) *\/ */
-/* 680, /\* (1) *\/ */
-/* 704, /\* (1) *\/ */
-/* 716, /\* (1) *\/ */
+/* 676, /\* (1) *\/ */
+/* 700, /\* (1) *\/ */
+/* 712, /\* (1) *\/ */
+/* 768, /\* (1) *\/ */
 /* 772, /\* (1) *\/ */
-/* 776, /\* (1) *\/ */
-1032, /* (7549) */
-/* 1044, /\* (14) *\/ */
-2076, /* (14) */
-4116, /* (9) */
-8212, /* (6) */
-16404, /* (4) */
-63504, /* (7) */
-135636, /* (2) */
-253992, /* (7) */
-1050864, /* (1) */
+1028, /* (7585) */
+1032, /* (14) */
+2084, /* (14) */
+4124, /* (9) */
+8220, /* (6) */
+16412, /* (4) */
+63500, /* (7) */
+136348, /* (1) */
+253988, /* (7) */
+1050860, /* (1) */
 2097152
 };
 #endif
@@ -317,8 +324,10 @@ struct MP {
 };

 struct FRAG {
-  struct FRAG *next;
-  unsigned int sbits;
+  union {
+    struct FRAG *next;
+    unsigned int sbits;
+  } u;
 #ifdef CL_DEBUG
  unsigned int magic;
 #endif
@@ -349,9 +358,6 @@ struct MP *mp_create() {
  unsigned int sz;
  memset(&mp, 0, sizeof(mp));
  mp.psize = getpagesize();
-#ifdef DEBUGMPOOL
-  lfd = fopen("mmpool_log", "w");
-#endif
  sz = align_to_pagesize(&mp, MIN_FRAGSIZE);
  mp.mpm.usize = align_to_voidptr(sizeof(struct MPMAP));
  mp.mpm.size = sz - align_to_voidptr(sizeof(mp));
@@ -425,7 +431,8 @@ void *mp_malloc(struct MP *mp, size_t size) {
  /* Case 1: We have a free'd frag */
  if((f = mp->avail[sbits])) {
    spam("malloc %p size %u (freed)\n", f, mp_roundup(size));
-    mp->avail[sbits] = f->next;
+    mp->avail[sbits] = f->u.next;
+    f->u.sbits = sbits;
 #ifdef CL_DEBUG
      f->magic = MPOOLMAGIC;
 #endif
@@ -443,7 +450,7 @@ void *mp_malloc(struct MP *mp, size_t size) {
      f = (struct FRAG *)((void *)mpm + mpm->usize);
      spam("malloc %p size %u (hole)\n", f, mp_roundup(size));
      mpm->usize += needed;
-      f->sbits = sbits;
+      f->u.sbits = sbits;
 #ifdef CL_DEBUG
      f->magic = MPOOLMAGIC;
 #endif
@@ -469,7 +476,7 @@ void *mp_malloc(struct MP *mp, size_t size) {
  mp->mpm.next = mpm;
  f = (struct FRAG *)((void *)mpm + align_to_voidptr(sizeof(*mpm)));
  spam("malloc %p size %u (new map)\n", f, mp_roundup(size));
-  f->sbits = sbits;
+  f->u.sbits = sbits;
 #ifdef CL_DEBUG
      f->magic = MPOOLMAGIC;
 #endif
@@ -478,14 +485,16 @@ void *mp_malloc(struct MP *mp, size_t size) {

 void mp_free(struct MP *mp, void *ptr) {
  struct FRAG *f = (struct FRAG *)(ptr - FRAG_OVERHEAD);
+  unsigned int sbits;
  if (!ptr) return;

 #ifdef CL_DEBUG
  assert(f->magic == MPOOLMAGIC && "Attempt to mp_free a pointer we did not allocate!");
 #endif

-  f->next = mp->avail[f->sbits];
-  mp->avail[f->sbits] = f;
+  sbits = f->u.sbits;
+  f->u.next = mp->avail[sbits];
+  mp->avail[sbits] = f;
  spam("free @ %p\n", f);
 }

@@ -501,17 +510,18 @@ void *mp_calloc(struct MP *mp, size_t nmemb, size_t size) {

 void *mp_realloc(struct MP *mp, void *ptr, size_t size) {
  struct FRAG *f = (struct FRAG *)(ptr - FRAG_OVERHEAD);
-  unsigned int csize;
+  unsigned int csize, sbits;
  void *new_ptr;
  if (!ptr) return mp_malloc(mp, size);

-  spam("realloc @ %p (size %u -> %u))\n", f, from_bits(f->sbits), size);
-  if(!size || !(csize = from_bits(f->sbits))) {
+  spam("realloc @ %p (size %u -> %u))\n", f, from_bits(f->u.sbits), size);
+  if(!size || !(csize = from_bits(f->u.sbits))) {
    cli_errmsg("mp_realloc(): Attempt to allocate %lu bytes. Please report to http://bugs.clamav.net\n", (unsigned long int) size);
    return NULL;
  }
  csize -= FRAG_OVERHEAD;
-  if (csize >= size) return ptr;
+  if (csize >= size && (!f->u.sbits || from_bits(f->u.sbits-1)-FRAG_OVERHEAD < size))
+    return ptr;
  if (!(new_ptr = mp_malloc(mp, size)))
    return NULL;
  memcpy(new_ptr, ptr, csize);
@@ -525,14 +535,15 @@ void *mp_realloc2(struct MP *mp, void *ptr, size_t size) {
  void *new_ptr;
  if (!ptr) return mp_malloc(mp, size);

-  spam("realloc @ %p (size %u -> %u))\n", f, from_bits(f->sbits), size);
-  if(!size || !(csize = from_bits(f->sbits))) {
+  spam("realloc @ %p (size %u -> %u))\n", f, from_bits(f->u.sbits), size);
+  if(!size || !(csize = from_bits(f->u.sbits))) {
    cli_errmsg("mp_realloc2(): Attempt to allocate %lu bytes. Please report to http://bugs.clamav.net\n", (unsigned long int) size);
    mp_free(mp, ptr);
    return NULL;
  }
  csize -= FRAG_OVERHEAD;
-  if (csize >= size) return ptr;
+  if (csize >= size && (!f->u.sbits || from_bits(f->u.sbits-1)-FRAG_OVERHEAD < size))
+    return ptr;
  if ((new_ptr = mp_malloc(mp, size)))
    memcpy(new_ptr, ptr, csize);
  mp_free(mp, ptr);