/*****************************************************************************\ * DHCMP - Compare DH.EXE outputs. * * Copyright (c) 1995-2000 Microsoft Corporation. All rights reserved. * * DHCMP is a character-mode tool which processes DH output file(s) into forms * which may be more useful in investigate memory leaks etc. * * DH is a useful tool which displays heap allocations in a properly enabled * system, but the output is sometimes hard to analyze and interpret. * The output is a list of allocation backtraces: each backtrace contains up to * MAX_BT call-sites, and is accompanied by the number of bytes allocated. * * 02-01-95 IanJa bugfixes and handle BackTraceNNNNN identifiers from dh.exe * 03/22/95 IanJa modify to cope with current DH output format. * 07/27/98 t-mattba added -v switch \*****************************************************************************/ char *pszHow = " DHCMP has two modes:\n" "\n" " 1) DHCMP [-d] dh_dump1.txt dh_dump2.txt\n" " This compares two DH dumps, useful for finding leaks.\n" " dh_dump1.txt & dh_dump2.txt are obtained before and after some test\n" " scenario. DHCMP matches the backtraces from each file and calculates\n" " the increase in bytes allocated for each backtrace. These are then\n" " displayed in descending order of size of leak\n" " The first line of each backtrace output shows the size of the leak in\n" " bytes, followed by the (last-first) difference in parentheses.\n" " Leaks of size 0 are not shown.\n" "\n" " 2) DHCMP [-d] dh_dump.txt\n" " For each allocation backtrace, the number of bytes allocated will be\n" " attributed to each callsite (each line of the backtrace). The number\n" " of bytes allocated per callsite are summed and the callsites are then\n" " displayed in descending order of bytes allocated. This is useful for\n" " finding a leak that is reached via many different codepaths.\n" " ntdll!RtlAllocateHeap@12 will appear first when analyzing DH dumps of\n" " csrss.exe, since all allocation will have gone through that routine.\n" " Similarly, ProcessApiRequest will be very prominent too, since that\n" " appears in most allocation backtraces. Hence the useful thing to do\n" " with mode 2 output is to use dhcmp to comapre two of them:\n" " dhcmp dh_dump1.txt > tmp1.txt\n" " dhcmp dh_dump2.txt > tmp2.txt\n" " dhcmp tmp1.txt tmp2.txt\n" " the output will show the differences.\n" "\n" " Flags:\n" " -d Output in decimal (default is hexadecimal)\n" // " -t Find Totals (NOT YET IMPLEMENTED)\n" " -v Verbose output: include the actual backtraces as well as summary information\n" " (Verbose output is only interesting in mode 1 above.)\n" " -? This help\n"; #include #include #include #define NHASH 47 #define TRUE 1 #define FALSE 0 typedef int BOOL; #define TYPE_WHOLESTACK 0 #define TYPE_FUNCTIONS 1 #define MAXLINELENGTH 4096 #define MAXFUNCNAMELENGTH 1024 #define MAX_BT 48 /* max length of back trace stack */ void AddToName(char *fnname, unsigned __int64 nb, int sign); void SetAllocs(char *fnname, unsigned __int64 nb, int sign); void Process(char *fnam, int sign, int type); void SortAll(); void AddToStackTrace(char *fnname, char *line); void ResetStackTrace(char *fnname); /* * Hashing */ int MakeHash(char *pName); void InitHashTab(); #define DUMPF_FIRST (1) #define DUMPF_SECOND (2) #define DUMPF_RESULT (4) #define DUMPF_ALL (DUMPF_FIRST | DUMPF_SECOND | DUMPF_RESULT) void DumpNodes(int Flags); #define F_DECIMAL 0x0001 #define F_TOTAL 0x0002 #define F_VERBOSE 0x0004 // // Globals // int gFlags = 0; int cdecl main(int argc, char *argv[]) { int n, DumpType; InitHashTab(); for (n = 1; n < argc; n++) { if ((argv[n][0] == '-') || (argv[n][0] == '/')) { /* * Flags */ switch (argv[n][1]) { case 'd': gFlags |= F_DECIMAL; break; //NOT YET IMPLEMENTED //case 't': // gFlags |= F_TOTAL; // break; case 'v': gFlags |= F_VERBOSE; break; case '?': default: printf("%s\n", pszHow); return 1; } } else { /* * No more flags */ break; } } if ((argc - n) == 2) { DumpType = DUMPF_ALL; Process(argv[n], -1, TYPE_WHOLESTACK); Process(argv[n+1], +1, TYPE_WHOLESTACK); } else if ((argc - n) == 1) { // // F_VERBOSE is not meaningful when groveling only one dump. // gFlags &= ~F_VERBOSE; DumpType = DUMPF_RESULT; Process(argv[n], +1, TYPE_FUNCTIONS); } else { printf("%s\n", pszHow); return 1; } // printf("==================== BEFORE SORTING ====================\n"); // DumpNodes(DUMPF_ALL); SortAll(); // printf("==================== AFTER SORTING ====================\n"); DumpNodes(DumpType); return 0; } void Process(char *fname, int sign, int type) { FILE *stream; char linebuff[MAXLINELENGTH]; char fnnamebuff[MAXFUNCNAMELENGTH]; // // BogdanA - 02/19/2002: we don't really need that much stack... // MAXLINELENGTH should be OK, we get it by copying from a line // char BackTraceBuff[MAXLINELENGTH] = {0}; char *p; int lineno = 0; BOOL skip = TRUE; // start out skipping lines int iT; unsigned __int64 ulT = 0L; unsigned __int64 nBytes = 0L; unsigned __int64 ulConsumed; unsigned __int64 lAllocs; // printf("PROCESS %s %d %d\n", fname, sign, type); stream = fopen(fname, "r"); if (stream == NULL) { fprintf(stderr, "Can't open %s for reading\n", fname); exit (2); } nBytes = 0; while (fgets(linebuff, sizeof(linebuff), stream) != NULL) { lineno++; //fprintf(stderr, "Line #%d\r", lineno); if (linebuff[0] == '*') { // // If we find a "hogs" line, stack traces follow, any other line // started by "*" should cause us to go back to searching for a // hogs block. // if (strstr(linebuff, "Hogs")) { skip = FALSE; } else { skip = TRUE; } continue; } if (skip) { // // Skip is enabled, skip this line, it is data about the heap // between 'heap information' and 'heap hogs' lines. // continue; } if (linebuff[0] != ' ') { // // Scan for byte count and find out how many characters have // been consumed by this action. // ulConsumed = 0; iT = sscanf(linebuff, "%I64x bytes in %I64x", &ulT, &lAllocs); if (iT > 0) { nBytes = ulT; p = strstr(linebuff, "BackTrace"); if (!p) { // // What's this ? // continue; } strncpy(BackTraceBuff, p, sizeof(BackTraceBuff) - 1); p = strchr(BackTraceBuff, '\n'); if (p) { *p = '\0'; } if (type == TYPE_FUNCTIONS) { // // BackTraceBuff is now saved for use with the rest of the // trace. // continue; } AddToName(BackTraceBuff, nBytes, sign); if(iT == 1) { lAllocs = 1; } SetAllocs(BackTraceBuff, lAllocs, sign); ResetStackTrace(BackTraceBuff); } } else if (nBytes != 0) { /* * If TYPE_WHOLESTACK, then add the count to each line of the * stack backtrace. */ if (sscanf(linebuff, " %[^+]+0x", fnnamebuff) == 1) { if (type == TYPE_FUNCTIONS) { AddToName(fnnamebuff, nBytes, sign); } if ((gFlags & F_VERBOSE) == F_VERBOSE) { AddToStackTrace(BackTraceBuff, linebuff); } continue; } else { nBytes = 0; } } } fclose(stream); /* * make sure to account for the final one. */ if (type == TYPE_WHOLESTACK) { AddToName(BackTraceBuff, nBytes, sign); } } /* * Hashing */ typedef struct tagNODE { char *pName; __int64 lValue; __int64 lFirst; __int64 lSecond; char BackTrace[MAX_BT][MAXFUNCNAMELENGTH]; long lPosition; __int64 lAllocsFirst; __int64 lAllocsSecond; struct tagNODE *pNext; } NODE, *PNODE; void DumpStackTrace(PNODE pNode); PNODE HashTab[NHASH]; void InitHashTab() { int i; for (i = 0; i < NHASH; i++) { HashTab[i] = NULL; } } int MakeHash(char *pName) { int hash = 0; while (*pName) { hash += *pName; pName++; } return hash % NHASH; } void DumpNodes(int Flags) { PNODE pNode; int i; unsigned __int64 ulTotal = 0; char *fmt1; char *fmt2; char *fmt3; char *fmt4; char *fmt5; char *fmt6; char *fmt7; if ((gFlags & F_VERBOSE) == F_VERBOSE) { if (gFlags & F_DECIMAL) { fmt1 = "% 8I64d %s\n"; fmt2 = "% 8I64d bytes by: %s\n"; fmt3 = "+% 8I64d ( %6I64d - %6I64d) %6I64d allocs\t%s\n"; fmt4 = "-% 8I64d ( %6I64d - %6I64d) %6I64d allocs\t%s\n"; fmt5 = "\nTotal increase == %I64d\n"; fmt6 = "+% 8I64d ( %6I64d - %6I64d)\t%s\tallocations\n"; fmt7 = "-% 8I64d ( %6I64d - %6I64d)\t%s\tallocations\n"; } else { fmt1 = "%08I64x %s\n"; fmt2 = "%08I64x bytes by: %s\n"; fmt3 = "+% 8I64x ( %5I64x - %5I64x) %6I64x allocs\t%s\n"; fmt4 = "-% 8I64x ( %5I64x - %5I64x) %6I64x allocs\t%s\n"; fmt5 = "\nTotal increase == %I64x\n"; fmt6 = "+% 8I64x ( %5I64x - %5I64x)\t%s\tallocations\n"; fmt7 = "-% 8I64x ( %5I64x - %5I64x)\t%s\tallocations\n"; } } else { if (gFlags & F_DECIMAL) { fmt1 = "% 8I64d %s\n"; fmt2 = "% 8I64d bytes by: %s\n"; fmt3 = "+% 8I64d ( %6I64d - %6I64d) %6I64d allocs\t%s\n"; fmt4 = "\n-% 8I64d ( %6I64d - %6I64d) %6I64d allocs\t%s"; fmt5 = "\nTotal increase == %I64d\n"; } else { fmt1 = "%08I64x %s\n"; fmt2 = "%08I64x bytes by: %s\n"; fmt3 = "+% 8I64x ( %5I64x - %5I64x) %6I64x allocs\t%s\n"; fmt4 = "\n-% 8I64x ( %5I64x - %5I64x) %6I64x allocs\t%s"; fmt5 = "\nTotal increase == %I64x\n"; } } for (i = 0; i < NHASH; i++) { // printf("========= HASH %d ==========\n", i); for (pNode = HashTab[i]; pNode != NULL; pNode = pNode->pNext) { switch (Flags) { case DUMPF_FIRST: printf(fmt1, pNode->lFirst, pNode->pName); break; case DUMPF_SECOND: printf(fmt1, pNode->lSecond, pNode->pName); break; case DUMPF_RESULT: printf(fmt2, pNode->lValue, pNode->pName); break; case DUMPF_ALL: if (pNode->lValue > 0) { printf(fmt3, pNode->lValue, pNode->lSecond, pNode->lFirst, (pNode->lAllocsSecond), pNode->pName); } else if (pNode->lValue < 0) { printf(fmt4, -pNode->lValue, pNode->lSecond, pNode->lFirst, (pNode->lAllocsSecond), pNode->pName); } if((gFlags & F_VERBOSE) == F_VERBOSE) { if(pNode->lAllocsSecond-pNode->lAllocsFirst > 0) { printf(fmt6, pNode->lAllocsSecond-pNode->lAllocsFirst, pNode->lAllocsSecond, pNode->lAllocsFirst, pNode->pName); } else if(pNode->lAllocsSecond-pNode->lAllocsFirst < 0) { printf(fmt7, -(pNode->lAllocsSecond-pNode->lAllocsFirst), pNode->lAllocsSecond, pNode->lAllocsFirst, pNode->pName); } } break; } ulTotal += pNode->lValue; if(((gFlags & F_VERBOSE) == F_VERBOSE) && (pNode->lValue != 0)) { DumpStackTrace(pNode); } } } if (Flags == DUMPF_ALL) { printf(fmt5, ulTotal); } } PNODE FindNode(char *pName) { int i; PNODE pNode; i = MakeHash(pName); pNode = HashTab[i]; while (pNode) { if (strcmp(pName, pNode->pName) == 0) { return pNode; } pNode = pNode->pNext; } // Not found // fprintf(stderr, "NEW %s\n", pName); pNode = malloc(sizeof(NODE)); if (!pNode) { fprintf(stderr, "malloc failed in FindNode\n"); exit(2); } pNode->pName = _strdup(pName); if (!pNode->pName) { fprintf(stderr, "strdup failed in FindNode\n"); exit(2); } pNode->pNext = HashTab[i]; HashTab[i] = pNode; pNode->lValue = 0L; pNode->lFirst = 0L; pNode->lSecond = 0L; pNode->lPosition = 0L; pNode->lAllocsFirst = 0L; pNode->lAllocsSecond = 0L; return pNode; } void AddToName(char *fnname, unsigned __int64 nb, int sign) { PNODE pNode; // fprintf(stderr, "%s += %lx\n", fnname, nb); pNode = FindNode(fnname); pNode->lValue += nb * sign; if (sign == -1) { pNode->lFirst += nb; } else { pNode->lSecond += nb; } // fprintf(stderr, "%s == %lx\n", fnname, pNode->lValue); } void SetAllocs(char *fnname, unsigned __int64 nb, int sign) { PNODE pNode; // fprintf(stderr, "%s += %lx\n", fnname, nb); pNode = FindNode(fnname); if (sign == -1) { pNode->lAllocsFirst = nb; } else { pNode->lAllocsSecond = nb; } // fprintf(stderr, "%s == %lx\n", fnname, pNode->lValue); } void ResetStackTrace(char *fnname) { PNODE pNode; pNode = FindNode(fnname); pNode->lPosition = 0L; } void AddToStackTrace(char *fnname, char *line) { PNODE pNode; pNode = FindNode(fnname); // // Make sure we don't write too much data in the BackTrace field. // if (pNode -> lPosition >= MAX_BT) { // // MAX_BT should be the number of entries in a stack trace that // DH/UMDH captures. If we trigger this we have tried to attach // more than MAX_BT entries in this stack. // fprintf(stderr, "More than %d entries in this stack trace, " "did the max change ?\n", MAX_BT); exit(EXIT_FAILURE); } // // BogdanA - 02/19/2002 - do counted copy and // don't forget to increment lPosition // strncpy(pNode->BackTrace[pNode->lPosition], line, sizeof(pNode->BackTrace[pNode->lPosition]) - 1); pNode->lPosition += 1; } /* * Insert pNode into the list at ppNodeHead. * Sort in ascending order. * Insert pNode BEFORE the first item >= pNode. */ void Reinsert(PNODE pNode, PNODE *ppNodeHead) { PNODE *ppT; ppT = ppNodeHead; while (*ppT && (pNode->lValue < (*ppT)->lValue)) { ppT = &((*ppT)->pNext); } /* * Insert pNode before *ppT */ pNode->pNext = *ppT; *ppT = pNode; } void SortList(PNODE *ppNodeHead) { PNODE pNode; PNODE pNext; pNode = *ppNodeHead; if (pNode == NULL) { return; } pNext = pNode->pNext; if (pNext == NULL) { return; } while (TRUE) { while (pNext != NULL) { if (pNode->lValue < pNext->lValue) { /* * cut the unordered node from the list */ pNode->pNext = pNext->pNext; Reinsert(pNext, ppNodeHead); break; } pNode = pNext; pNext = pNode->pNext; } if (pNext == NULL) { return; } pNode = *ppNodeHead; pNext = pNode->pNext; } } /* * Merge ordered list 1 into ordered list 2 * Leaves list 1 empty; list 2 ordered */ void MergeLists(PNODE *ppNode1, PNODE *ppNode2) { PNODE *pp1; PNODE *pp2; PNODE p1; PNODE p2; pp1 = ppNode1; pp2 = ppNode2; while (TRUE) { p1 = *pp1; p2 = *pp2; if (p1 == NULL) { return; } if (p2 == NULL) { *pp2 = *pp1; *pp1 = NULL; return; } if (p1->lValue > p2->lValue) { *pp1 = p1->pNext; p1->pNext = p2; *pp2 = p1; pp2 = &(p1->pNext); } else { pp2 = &(p2->pNext); } } } void SortAll() { int i; for (i = 0; i < NHASH; i++) { SortList(&HashTab[i]); } // printf(" ======================== SORTED ========================\n"); // DumpNodes(DUMPF_ALL); for (i = 0; i < NHASH-1; i++) { // printf(" ======================== MERGING %d and %d ======================== \n", i, i+1); MergeLists(&HashTab[i], &HashTab[i+1]); // DumpNodes(DUMPF_ALL); } } void DumpStackTrace(PNODE pNode) { int n; for(n = 0; n < pNode->lPosition; n++) { printf("%s", pNode->BackTrace[n]); } }