gen/stack_logging.c

   1 /*
   2  * Copyright (c) 1999 Apple Computer, Inc. All rights reserved.
   3  *
   4  * @APPLE_LICENSE_HEADER_START@
   5  *
   6  * Copyright (c) 1999-2003 Apple Computer, Inc.  All Rights Reserved.
   7  *
   8  * This file contains Original Code and/or Modifications of Original Code
   9  * as defined in and that are subject to the Apple Public Source License
  10  * Version 2.0 (the 'License'). You may not use this file except in
  11  * compliance with the License. Please obtain a copy of the License at
  12  * http://www.opensource.apple.com/apsl/ and read it before using this
  13  * file.
  14  *
  15  * The Original Code and all software distributed under the License are
  16  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  17  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  18  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
  20  * Please see the License for the specific language governing rights and
  21  * limitations under the License.
  22  *
  23  * @APPLE_LICENSE_HEADER_END@
  24  */
  25
  26 /*      Bertrand from vmutils -> CF -> System */
  27
  28 #import "stack_logging.h"
  29
  30 #import <libc.h>
  31 #import <pthread.h>
  32 #import <mach/mach.h>
  33 #include <mach/vm_statistics.h>
  34
  35 extern void spin_lock(int *);
  36
  37 static inline void *allocate_pages(unsigned bytes) {
  38     void *address;
  39     if (vm_allocate(mach_task_self(), (vm_address_t *)&address, bytes,
  40                     VM_MAKE_TAG(VM_MEMORY_ANALYSIS_TOOL)| TRUE)) {
  41         address = 0;
  42     }
  43     return (void *)address;
  44 }
  45
  46 static inline void deallocate_pages(void *ptr, unsigned bytes) {
  47     vm_deallocate(mach_task_self(), (vm_address_t)ptr, bytes);
  48 }
  49
  50 static inline void copy_pages(const void *source, void *dest, unsigned bytes) {
  51     if (vm_copy(mach_task_self(), (vm_address_t)source, bytes, (vm_address_t)dest)) memmove(dest, source, bytes);
  52 }
  53
  54 /***************        Recording stack         ***********/
  55
  56 static void *first_frame_address(void) {
  57 #if 0
  58     return __builtin_frame_address(1);
  59 #elif defined(__ppc__)
  60     void *addr;
  61 #warning __builtin_frame_address IS BROKEN IN BEAKER: RADAR #2340421
  62     __asm__ volatile("mr %0, r1" : "=r" (addr));
  63     return addr;
  64 #else
  65 #warning first_frame_address WILL NOT BE FUNCTIONAL ON THIS ARCHITECTURE
  66     return NULL;
  67 #endif
  68 }
  69
  70 static void *next_frame_address(void *addr) {
  71     void *ret;
  72 #if defined(__MACH__) && defined(__i386__)
  73     __asm__ volatile("movl (%1),%0" : "=r" (ret) : "r" (addr));
  74 #elif defined(__MACH__) && defined(__ppc__)
  75     __asm__ volatile("lwz %0,0x0(%1)" : "=r" (ret) : "b" (addr));
  76 #elif defined(__hpux__)
  77     __asm__ volatile("ldw 0x0(%1),%0" : "=r" (ret) : "r" (addr));
  78 #elif defined(__svr4__)
  79     __asm__ volatile("ta 0x3");
  80     __asm__ volatile("ld [%1 + 56],%0" : "=r" (ret) : "r" (addr));
  81 #else
  82 #error Unknown architecture
  83 #endif
  84     return ret;
  85 }
  86
  87 #if defined(__i386__) || defined (__m68k__)
  88 #define FP_LINK_OFFSET 1
  89 #elif defined(__ppc__)
  90 #define FP_LINK_OFFSET 2
  91 #elif defined(__hppa__)
  92 #define FP_LINK_OFFSET -5
  93 #elif defined(__sparc__)
  94 #define FP_LINK_OFFSET 14
  95 #else
  96 #error  ********** Unimplemented architecture
  97 #endif
  98
  99 void thread_stack_pcs(vm_address_t *buffer, unsigned max, unsigned *nb) {
 100     void *addr;
 101     addr = first_frame_address();
 102     *nb = 0;
 103     while ((addr >= (void *)0x800) && (max--)) {
 104         vm_address_t    fp_link = (vm_address_t)(((unsigned *)addr)+FP_LINK_OFFSET);
 105         void    *addr2;
 106         buffer[*nb] = *((vm_address_t *)fp_link);
 107         (*nb)++;
 108         addr2 = next_frame_address(addr);
 109 #if defined(__ppc__)
 110         if ((unsigned)addr2 <= (unsigned)addr) break; // catch bozo frames
 111 #endif
 112         addr = addr2;
 113     }
 114 }
 115
 116 /***************        Uniquing stack          ***********/
 117
 118 #define MAX_COLLIDE     8
 119
 120 #define MAX_NUM_PC      512
 121
 122 static int enter_pair_in_table(unsigned *table, unsigned numPages, unsigned *uniquedParent, unsigned thisPC) {
 123     // uniquedParent is in-out; return 1 is collisions max not exceeded
 124     unsigned    base = numPages * vm_page_size / (sizeof(int)*2*2);
 125     unsigned    hash = base + (((*uniquedParent) << 4) ^ (thisPC >> 2)) % (base - 1); // modulo odd number for hashing
 126     unsigned    collisions = MAX_COLLIDE;
 127     while (collisions--) {
 128         unsigned        *head = table + hash*2;
 129         if (! head[0] && !head[1]) {
 130             /* end of chain; store this entry! */
 131             /* Note that we need to test for both head[0] and head[1] as (0, -1) is a valid entry */
 132             head[0] = thisPC;
 133             head[1] = *uniquedParent;
 134             *uniquedParent = hash;
 135             return 1;
 136         }
 137         if ((head[0] == thisPC) && (head[1] == *uniquedParent)) {
 138             /* we found the proper entry, the value for the pair is the entry offset */
 139             *uniquedParent = hash;
 140             return 1;
 141         }
 142         hash++;
 143         if (hash == base*2) hash = base;
 144     }
 145     return 0;
 146 }
 147
 148 unsigned stack_logging_get_unique_stack(unsigned **table, unsigned *table_num_pages, unsigned *stack_entries, unsigned count, unsigned num_hot_to_skip) {
 149     unsigned    uniquedParent = (unsigned)-1;
 150     // we skip the warmest entries that are an artefact of the code
 151     while (num_hot_to_skip--) {
 152         if (count > 0) { stack_entries++; count--; }
 153     }
 154     while (count--) {
 155         unsigned        thisPC = stack_entries[count];
 156         while (!enter_pair_in_table(*table, *table_num_pages, &uniquedParent, thisPC)) {
 157             unsigned    *newTable;
 158             unsigned    oldBytes = (*table_num_pages) * vm_page_size;
 159             newTable = allocate_pages(oldBytes*2);
 160             copy_pages(*table, newTable, oldBytes);
 161             deallocate_pages(*table, oldBytes);
 162             *table_num_pages *= 2;
 163             *table = newTable;
 164         }
 165     }
 166     return uniquedParent;
 167 }
 168
 169 /***************        Logging stack and arguments             ***********/
 170
 171 stack_logging_record_list_t *stack_logging_the_record_list = NULL;
 172
 173 int stack_logging_enable_logging = 0;
 174
 175 int stack_logging_dontcompact = 0;
 176
 177 static stack_logging_record_list_t *GrowLogRecords(stack_logging_record_list_t *records, unsigned desiredNumRecords) {
 178     stack_logging_record_list_t *new_records;
 179     unsigned    old_size = records->overall_num_bytes;
 180     if (desiredNumRecords*sizeof(stack_logging_record_t)+sizeof(stack_logging_record_list_t) < records->overall_num_bytes) return records;
 181     records->overall_num_bytes += records->overall_num_bytes + vm_page_size; // in order to always get an even number of pages
 182     new_records = allocate_pages(records->overall_num_bytes);
 183     copy_pages(records, new_records, old_size);
 184     deallocate_pages(records, old_size);
 185     return new_records;
 186 }
 187
 188 static void prepare_to_log_stack(void) {
 189     if (!stack_logging_the_record_list) {
 190         unsigned        totalSize = 4 * vm_page_size;
 191         stack_logging_the_record_list = allocate_pages(totalSize);
 192         memset(stack_logging_the_record_list, 0, sizeof(stack_logging_record_list_t));
 193         stack_logging_the_record_list->overall_num_bytes = totalSize;
 194         stack_logging_the_record_list->uniquing_table_num_pages = 128;
 195         stack_logging_the_record_list->uniquing_table = allocate_pages(stack_logging_the_record_list->uniquing_table_num_pages * vm_page_size);
 196     }
 197 }
 198
 199 void stack_logging_log_stack(unsigned type, unsigned arg1, unsigned arg2, unsigned arg3, unsigned result, unsigned num_hot_to_skip) {
 200     stack_logging_record_t      *rec;
 201     if (!stack_logging_enable_logging) return;
 202     // printf("stack_logging_log_stack 0x%x 0x%x 0x%x 0x%x -> 0x%x\n", type, arg1, arg2, arg3, result);
 203     if (type & stack_logging_flag_zone) {
 204         // just process it now and be done with it!
 205         arg1 = arg2; arg2 = arg3; arg3 = 0; type &= ~stack_logging_flag_zone;
 206     }
 207     if (type & stack_logging_flag_calloc) {
 208         // just process it now and be done with it!
 209         arg1 *= arg2; arg2 = arg3; arg3 = 0; type &= ~stack_logging_flag_calloc;
 210     }
 211     if (type & stack_logging_flag_object) {
 212         unsigned        *class = (unsigned *)arg1;
 213         arg1 = arg2 + class[5]; // corresponds to the instance_size field
 214         arg2 = 0; arg3 = 0; type = stack_logging_type_alloc;
 215     }
 216     if (type & stack_logging_flag_cleared) {
 217         type &= ~stack_logging_flag_cleared;
 218     }
 219     if (type & stack_logging_flag_handle) {
 220         if (stack_logging_type_alloc) {
 221             if (!result) return;
 222             stack_logging_log_stack(stack_logging_type_alloc, 0, 0, 0, result, num_hot_to_skip+1);
 223             stack_logging_log_stack(stack_logging_type_alloc, arg1, 0, 0, *((int *)result), num_hot_to_skip+1);
 224             return;
 225         }
 226         if (stack_logging_type_dealloc) {
 227             if (!arg1) return;
 228             stack_logging_log_stack(stack_logging_type_dealloc, *((int *)arg1), 0, 0, 0, num_hot_to_skip+1);
 229             stack_logging_log_stack(stack_logging_type_dealloc, arg1, 0, 0, 0, num_hot_to_skip+1);
 230             return;
 231         }
 232         printf("*** Unknown logging type: 0x%x\n", type);
 233     }
 234     if (type == stack_logging_flag_set_handle_size) {
 235         if (!arg1) return;
 236         // Thanks to a horrible hack, arg3 contains the prvious handle value
 237         if (arg3 == *((int *)arg1)) return;
 238         stack_logging_log_stack(stack_logging_type_dealloc, arg3, 0, 0, 0, num_hot_to_skip+1);
 239         stack_logging_log_stack(stack_logging_type_alloc, arg2, 0, 0, *((int *)arg1), num_hot_to_skip+1);
 240         return;
 241     }
 242     if (type == (stack_logging_type_dealloc|stack_logging_type_alloc)) {
 243         if (arg1 == result) return; // realloc had no effect, skipping
 244         if (!arg1) {
 245             // realloc(NULL, size) same as malloc(size)
 246             type = stack_logging_type_alloc; arg1 = arg2; arg2 = arg3; arg3 = 0;
 247         } else {
 248             // realloc(arg1, arg2) -> result is same as free(arg1); malloc(arg2) -> result
 249             stack_logging_log_stack(stack_logging_type_dealloc, arg1, 0, 0, 0, num_hot_to_skip+1);
 250             stack_logging_log_stack(stack_logging_type_alloc, arg2, 0, 0, result, num_hot_to_skip+1);
 251             return;
 252         }
 253     }
 254     if (type == stack_logging_type_dealloc) {
 255         // simple free
 256         if (!arg1) return; // free(nil)
 257     }
 258     prepare_to_log_stack();
 259     spin_lock(&stack_logging_the_record_list->lock);
 260     stack_logging_enable_logging = 0;
 261     stack_logging_the_record_list = GrowLogRecords(stack_logging_the_record_list, stack_logging_the_record_list->num_records + 1);
 262     rec = stack_logging_the_record_list->records + stack_logging_the_record_list->num_records;
 263     // We take care of the common case of alloc-dealloc
 264     if (!stack_logging_dontcompact && stack_logging_the_record_list->num_records && (type == stack_logging_type_dealloc) && arg1 && ((rec-1)->type == stack_logging_type_alloc) && (arg1 == STACK_LOGGING_DISGUISE((rec-1)->address))) {
 265         stack_logging_the_record_list->num_records--;
 266         // printf("Erased previous record in alloc-dealloc sequence\n");
 267     } else {
 268         unsigned        stack_entries[MAX_NUM_PC];
 269         unsigned        count = 0;
 270         rec->type = type;
 271         if (type == stack_logging_type_dealloc) {
 272             rec->argument = 0;
 273             rec->address = STACK_LOGGING_DISGUISE(arg1); // we disguise the address
 274         } else if (type == stack_logging_type_alloc) {
 275             rec->argument = arg1;
 276             rec->address = STACK_LOGGING_DISGUISE(result); // we disguise the address
 277         } else {
 278             rec->argument = arg2;
 279             rec->address = STACK_LOGGING_DISGUISE(arg1); // we disguise the address
 280         }
 281         // printf("Before getting samples  0x%x 0x%x 0x%x 0x%x -> 0x%x\n", type, arg1, arg2, arg3, result);
 282         thread_stack_pcs(stack_entries, MAX_NUM_PC - 1, &count);
 283         // We put at the bottom of the stack a marker that denotes the thread (+1 for good measure...)
 284         stack_entries[count++] = (int)pthread_self() + 1;
 285         /* now let's unique the sample */
 286         // printf("Uniquing 0x%x 0x%x 0x%x 0x%x -> 0x%x\n", type, arg1, arg2, arg3, result);
 287         rec->uniqued_stack = stack_logging_get_unique_stack(&stack_logging_the_record_list->uniquing_table, &stack_logging_the_record_list->uniquing_table_num_pages, stack_entries, count, num_hot_to_skip+2); // we additionally skip the warmest 2 entries that are an artefact of the code
 288         stack_logging_the_record_list->num_records++;
 289     }
 290     stack_logging_enable_logging = 1;
 291     stack_logging_the_record_list->lock = 0;
 292 }
 293
 294 static kern_return_t default_reader(task_t task, vm_address_t address, vm_size_t size, void **ptr) {
 295     *ptr = (void *)address;
 296     return 0;
 297 }
 298
 299 static kern_return_t get_remote_records(task_t task, memory_reader_t reader, stack_logging_record_list_t **records) {
 300     // sets records
 301     vm_address_t        *remote_records_address_ref;
 302     kern_return_t       err;
 303     *records = NULL;
 304     err = reader(task, (vm_address_t)&stack_logging_the_record_list, sizeof(vm_address_t), (void **)&remote_records_address_ref);
 305     if (err) return err;
 306     if (!*remote_records_address_ref) {
 307         // printf("stack_logging: no stack record\n");
 308         return 0;
 309     }
 310     // printf("stack_logging: stack records at %p\n", (void *)(*remote_records_address_ref));
 311     // printf("stack_logging: reading %d bytes\n", sizeof(stack_logging_record_list_t));
 312     err = reader(task, *remote_records_address_ref, sizeof(stack_logging_record_list_t), (void **)records); // get the list head
 313     if (err) return err;
 314     // printf("stack_logging: overall num bytes = %d\n", records->overall_num_bytes);
 315     return reader(task, *remote_records_address_ref, (*records)->overall_num_bytes, (void **)records);
 316 }
 317
 318 kern_return_t stack_logging_get_frames(task_t task, memory_reader_t reader, vm_address_t address, vm_address_t *stack_frames_buffer, unsigned max_stack_frames, unsigned *num_frames) {
 319     stack_logging_record_list_t *records;
 320     kern_return_t       err;
 321     unsigned            index;
 322     unsigned            disguised = STACK_LOGGING_DISGUISE(address);
 323     if (!reader) reader = default_reader;
 324     *num_frames = 0;
 325     err = get_remote_records(task, reader, &records);
 326     if (err || !records) return err;
 327     // printf("stack_logging: %d records\n", records->num_records);
 328     index = 0;
 329     while (index < records->num_records) {
 330         stack_logging_record_t  *record = records->records + index;
 331         if (record->address == disguised) {
 332             return stack_logging_frames_for_uniqued_stack(task, reader, record->uniqued_stack, stack_frames_buffer, max_stack_frames, num_frames);
 333         }
 334         index++;
 335     }
 336     fprintf(stderr, "*** stack_logging: no record found for 0x%x\n", address);
 337     return 0;
 338 }
 339
 340 kern_return_t stack_logging_enumerate_records(task_t task, memory_reader_t reader, vm_address_t address, void enumerator(stack_logging_record_t, void *), void *context) {
 341     stack_logging_record_list_t *records;
 342     kern_return_t       err;
 343     unsigned            index;
 344     unsigned            disguised = STACK_LOGGING_DISGUISE(address);
 345     if (!reader) reader = default_reader;
 346     err = get_remote_records(task, reader, &records);
 347     if (err || !records) return err;
 348     // printf("stack_logging: %d records\n", records->num_records);
 349     index = 0;
 350     while (index < records->num_records) {
 351         stack_logging_record_t  *record = records->records + index;
 352         if (!address || (record->address == disguised)) enumerator(*record, context);
 353         index++;
 354     }
 355     return 0;
 356 }
 357
 358 kern_return_t stack_logging_frames_for_uniqued_stack(task_t task, memory_reader_t reader, unsigned uniqued_stack, vm_address_t *stack_frames_buffer, unsigned max_stack_frames, unsigned *num_frames) {
 359     stack_logging_record_list_t *records;
 360     unsigned            *uniquing_table;
 361     kern_return_t       err;
 362     if (!reader) reader = default_reader;
 363     *num_frames = 0;
 364     err = get_remote_records(task, reader, &records);
 365     if (err || !records) return err;
 366     err = reader(task, (vm_address_t)records->uniquing_table, records->uniquing_table_num_pages * vm_page_size, (void **)&uniquing_table);
 367     if (err) return err;
 368     while (max_stack_frames && (uniqued_stack != -1)) {
 369         unsigned        thisPC;
 370         if ((uniqued_stack * 2 + 1) * sizeof(unsigned) >= records->uniquing_table_num_pages * vm_page_size) {
 371             fprintf(stderr, "*** stack_logging: Invalid uniqued stack 0x%x", uniqued_stack);
 372             break;
 373         }
 374         thisPC = uniquing_table[uniqued_stack * 2];
 375         uniqued_stack = uniquing_table[uniqued_stack * 2 + 1];
 376         if (!thisPC && !uniqued_stack) {
 377             // Invalid entry
 378             fprintf(stderr, "*** stack_logging: Invalid entry 0x%x", thisPC);
 379             break;
 380         }
 381         stack_frames_buffer[0] = thisPC;
 382         stack_frames_buffer++;
 383         (*num_frames)++;
 384         max_stack_frames--;
 385     }
 386     return 0;
 387 }