首页 > 代码库 > Linux内核情景分析的alloc_pages

Linux内核情景分析的alloc_pages

NUMA结构的alloc_pages
  1. ==================== mm/numa.c 43 43 ====================
  2. 43 #ifdef CONFIG_DISCONTIGMEM
  3. ==================== mm/numa.c 91 128 ====================
  4. 91 /*
  5. 92 * This can be refined. Currently, tries to do round robin, instead
  6. 93 * should do concentratic circle search, starting from current node.
  7. 94 */
  8. //分配策略, 所需物理块的大小,2的order次方
  9. 95 struct page * alloc_pages(int gfp_mask, unsigned long order)
  10. 96 {
  11. 97 struct page *ret = 0;
  12. 98 pg_data_t *start, *temp;
  13. 99 #ifndef CONFIG_NUMA
  14. 100 unsigned long flags;
  15. 101 static pg_data_t *next = 0;
  16. 102 #endif
  17. 103
  18. 104 if (order >= MAX_ORDER)
  19. 105 return NULL;
  20. 106 #ifdef CONFIG_NUMA//NUMA结构
  21. 107 temp = NODE_DATA(numa_node_id());//可以通过宏操作找到cpu的节数据结构队列
  22. 108 #else
  23. 109 spin_lock_irqsave(&node_lock, flags);
  24. 110 if (!next) next = pgdat_list;
  25. 111 temp = next;
  26. 112 next = next->node_next;
  27. 113 spin_unlock_irqrestore(&node_lock, flags);
  28. 114 #endif
  29. /*
  30. 函数主要操作2个循环,一个从temp到队列末尾,一个从队头到temp,扫描所有节,直到某节点内存分配成功
  31. */
  32. 115 start = temp;
  33. 116 while (temp) {
  34. 117 if ((ret = alloc_pages_pgdat(temp, gfp_mask, order)))//接下来解析此函数
  35. 118 return(ret);
  36. 119 temp = temp->node_next;
  37. 120 }
  38. 121 temp = pgdat_list;
  39. 122 while (temp != start) {
  40. 123 if ((ret = alloc_pages_pgdat(temp, gfp_mask, order)))
  41. 124 return(ret);
  42. 125 temp = temp->node_next;
  43. 126 }
  44. 127 return(0);
  45. 128 }
alloc_pages_pgdat试图分配所需页面,是__alloc_pages的封装
  1. 85 static struct page * alloc_pages_pgdat(pg_data_t *pgdat, int gfp_mask,
  2. 86 unsigned long order)
  3. 87 { //node_zonelist决定分配策略数组
  4. 88     return __alloc_pages(pgdat->node_zonelists + gfp_mask, order);
  5. 89 }
与UMA的alloc_pages()相比较,UMA只有一个节点,contig_page_data.UMA与NUMA共同使用__alloc_pages
  1. ==================== include/linux/mm.h 343 352 ====================
  2. 343 #ifndef CONFIG_DISCONTIGMEM//只有这个无定义,才使用uma的__alloc_pages
  3. 344 static inline struct page * alloc_pages(int gfp_mask, unsigned long order)
  4. 83
  5. 345 {
  6. 346 /*
  7. 347 * Gets optimized away by the compiler.
  8. 348 */
  9. 349 if (order >= MAX_ORDER)
  10. 350 return NULL;
  11. 351 return __alloc_pages(contig_page_data.node_zonelists+(gfp_mask), order);
  12. 352 }
查看__alloc_page,
  1. 如果只分配一个页面,而且要等待完成分配,又不适用于管理的目的
  2. 把direct_reclaim设置为1,表示可以从相应的管理区的不活跃干净页面缓冲队列中回收
  3. 84发现空闲页面短缺,唤醒以下2个进程,试图腾出一些页面出来
  1. [alloc_pages()>__alloc_pages()]
  2. 270 /*
  3. 271 * This is the ‘heart‘ of the zoned buddy allocator:
  4. 272 */
  5. 273 struct page * __alloc_pages(zonelist_t *zonelist, unsigned long order)
  6. 274 {
  7. 275 zone_t **zone;
  8. 276 int direct_reclaim = 0;
  9. 277 unsigned int gfp_mask = zonelist->gfp_mask;//获取具体的分配策略
  10. 278 struct page * page;
  11. 279
  12. 280 /*
  13. 281 * Allocations put pressure on the VM subsystem.
  14. 282 */
  15. 283 memory_pressure++;//表示内存管理所承受的压力,分配++,归还--
  16. 284
  17. 285 /*
  18. 286 * (If anyone calls gfp from interrupts nonatomically then it
  19. 287 * will sooner or later tripped up by a schedule().)
  20. 288 *
  21. 289 * We are falling back to lower-level zones if allocation
  22. 290 * in a higher zone fails.
  23. 291 */
  24. 292
  25. 293 /*
  26. 294 如果只分配一个页面,而且要等待完成分配,又不适用于管理的目的
  27. 把direct_reclaim设置为1,表示可以从相应的管理区的不活跃干净页面缓冲队列中回收
  28. 296 */
  29. 297 if (order == 0 && (gfp_mask & __GFP_WAIT) &&
  30. 298 !(current->flags & PF_MEMALLOC))
  31. 299 direct_reclaim = 1;
  32. 300
  33. 301 /*
  34. 302 * If we are about to get low on free pages and we also have
  35. 303 * an inactive page shortage, wake up kswapd.
  36. 84发现空闲页面短缺,唤醒以下2个进程,试图腾出一些页面出来
  37. 304 */
  38. 305 if (inactive_shortage() > inactive_target / 2 && free_shortage())
  39. 306 wakeup_kswapd(0);
  40. 307 /*
  41. 308 * If we are about to get low on free pages and cleaning
  42. 309 * the inactive_dirty pages would fix the situation,
  43. 310 * wake up bdflush.
  44. 311 */
  45. 312 else if (free_shortage() && nr_inactive_dirty_pages > free_shortage()
  46. 313 && nr_inactive_dirty_pages >= freepages.high)
  47. 314 wakeup_bdflush(0);
  48. 315
继续看__alloc_page代码
//如果管理区的空闲页面大于其最低标准,分配成功直接返回
  1. //否则有进程(内核线程kreclaimd)在等待队列睡眠,把它唤醒,用于回收一些页面,备用
  1. ==================== mm/page_alloc.c 316 340 ====================
  2. [alloc_pages()>__alloc_pages()]
  3. 316 try_again:
  4. 317 /*
  5. 318 * First, see if we have any zones with lots of free memory.
  6. 319 *
  7. 320 * We allocate free memory first because it doesn‘t contain
  8. 321 * any data ... DUH!
  9. 322 */
  10. 323 zone = zonelist->zones;//获取管理区指针
  11. 324 for (;;) {
  12. 325 zone_t *z = *(zone++);//管理区
  13. 326 if (!z)
  14. 327 break;
  15. 328 if (!z->size)
  16. 329 BUG();
  17. 330//如果管理区的空闲页面大于其最低标准
  18. 331 if (z->free_pages >= z->pages_low) {
  19. 332 page = rmqueue(z, order);//分配内存,接下来分析此函数
  20. 333 if (page)
  21. 334 return page;
  22. 335 } 
  23. //否则有进程(内核线程kreclaimd)在等待队列睡眠,把它唤醒,用于回收一些页面,备用
  24. else if (z->free_pages < z->pages_min &&
  25. 336 waitqueue_active(&kreclaimd_wait)) {
  26. 85
  27. 337 wake_up_interruptible(&kreclaimd_wait);
  28. 338 }
  29. 339 }
  30. 340

  1. [alloc_pages()>__alloc_pages()>rmqueue()]
  2. 172 static struct page * rmqueue(zone_t *zone, unsigned long order)
  3. 173 {
  4. 174 free_area_t * area = zone->free_area + order;//获取其数组对应的元素
  5. 175 unsigned long curr_order = order;
  6. 176 struct list_head *head, *curr;
  7. 177 unsigned long flags;
  8. 178 struct page *page;
  9. 179
  10. 180 spin_lock_irqsave(&zone->lock, flags);//相应管理区加锁
  11. 181 do {
  12. 182 head = &area->free_list;//头
  13. 183 curr = memlist_next(head);//头的下一个节点
  14. 184
  15. 185 if (curr != head) {//不等于空,说明有物理页块
  16. 186 unsigned int index;
  17. 187//从非空队列中取出第一个结构page元素
  18. 188 page = memlist_entry(curr, struct page, list);
  19. 189 if (BAD_RANGE(zone,page))
  20. 190 BUG();
  21. 191 memlist_del(curr);//删除队列中的元素
  22. 192 index = (page - mem_map) - zone->offset;//偏移
  23. 193 MARK_USED(index, curr_order, area);//将相应位图设置为1
  24. 194 zone->free_pages -= 1 << order;
  25. 195//分配成功,把大块剩余的部分分解为小块,链入相应的队列
  26. 196 page = expand(zone, page, index, order, curr_order, area);
  27. 197 spin_unlock_irqrestore(&zone->lock, flags);
  28. 198
  29. 199 set_page_count(page, 1);
  30. 200 if (BAD_RANGE(zone,page))
  31. 201 BUG();
  32. 202 DEBUG_ADD_PAGE
  33. 203 return page;
  34. 204 }
  35. 205 curr_order++;
  36. 206 area++;
  37. 86
  38. 207 } while (curr_order < MAX_ORDER);
  39. 208 spin_unlock_irqrestore(&zone->lock, flags);
  40. 209
  41. 210 return NULL;
  42. 211 }


  1. [alloc_pages()>__alloc_pages()>rmqueue()>expand()]
  2. /*
  3.     
        low表示所需块大小,high表示实际大小
  4.     */
  5. 150 static inline struct page * expand (zone_t *zone, struct page *page,
  6. 151 unsigned long index, int low, int high, free_area_t * area)
  7. 152 {
  8. 153 unsigned long size = 1 << high;
  9. 154
  10. 155 while (high > low) {
  11. 156 if (BAD_RANGE(zone,page))
  12. 157 BUG();
  13. 158 area--;
  14. 159 high--;
  15. 160 size >>= 1;//每次减少2的n次方
  16. 161 memlist_add_head(&(page)->list, &(area)->free_list);
  17. 162 MARK_USED(index, high, area);//标记位图
  18. //处理更低一档的空闲块队列
  19. 163 index += size;
  20. 164 page += size;
  21. 165 }
  22. 166 if (BAD_RANGE(zone,page))
  23. 167 BUG();
  24. 168 return page;
  25. 169 }
就这样rmqueue队列一直往上扫描,直到分配成功或者失败,如果失败,则__alloc_pages通过for循环
指向下一个管理区(按照分配策略),直到成功.
要是给定的分配策略中的所有页面管理区都失败,那就只能加大力度再试试.要么降低对页面的水位要求
要么把缓冲在管理区的不活跃干净页面也给考虑进去
  1. [alloc_pages()>__alloc_pages()]
  2. 341 /*
  3. 342 * Try to allocate a page from a zone with a HIGH
  4. 343 * amount of free + inactive_clean pages.
  5. 344 *
  6. 345 * If there is a lot of activity, inactive_target
  7. 346 * will be high and we‘ll have a good chance of
  8. 347 * finding a page using the HIGH limit.
  9. 348 */
  10. //先用page_high,如果不行再用page_low
  11. 349 page = __alloc_pages_limit(zonelist, order, PAGES_HIGH, direct_reclaim);
  12. 350 if (page)
  13. 351 return page;
  14. 352
  15. 353 /*
  16. 354 * Then try to allocate a page from a zone with more
  17. 355 * than zone->pages_low free + inactive_clean pages.
  18. 356 *
  19. 357 * When the working set is very large and VM activity
  20. 358 * is low, we‘re most likely to have our allocation
  21. 359 * succeed here.
  22. 360 */
  23. 361 page = __alloc_pages_limit(zonelist, order, PAGES_LOW, direct_reclaim);
  24. 362 if (page)
  25. 363 return page;
  26. 364


  1. [alloc_pages()>__alloc_pages()>__alloc_pages_limit()]
  2. 213 #define PAGES_MIN 0
  3. 214 #define PAGES_LOW 1
  4. 215 #define PAGES_HIGH 2
  5. 88
  6. 216
  7. 217 /*
  8. 218 * This function does the dirty work for __alloc_pages
  9. 219 * and is separated out to keep the code size smaller.
  10. 220 * (suggested by Davem at 1:30 AM, typed by Rik at 6 AM)
  11. 221 */
  12. 222 static struct page * __alloc_pages_limit(zonelist_t *zonelist,
  13. 223 unsigned long order, int limit, int direct_reclaim)
  14. 224 {
  15. 225 zone_t **zone = zonelist->zones;
  16. 226
  17. 227 for (;;) {
  18. 228 zone_t *z = *(zone++);
  19. 229 unsigned long water_mark;
  20. 230
  21. 231 if (!z)
  22. 232 break;
  23. 233 if (!z->size)
  24. 234 BUG();
  25. 235
  26. 236 /*
  27. 237 * We allocate if the number of free + inactive_clean
  28. 238 * pages is above the watermark.
  29. 239 */
  30. 240 switch (limit) {
  31. 241 default:
  32. 242 case PAGES_MIN://通过分配策略,改变水位
  33. 243 water_mark = z->pages_min;
  34. 244 break;
  35. 245 case PAGES_LOW:
  36. 246 water_mark = z->pages_low;
  37. 247 break;
  38. 248 case PAGES_HIGH:
  39. 249 water_mark = z->pages_high;
  40. 250 }
  41. 251//如果空闲页面+干净回收页面大于最低水位
  42. 252 if (z->free_pages + z->inactive_clean_pages > water_mark) {
  43. 253 struct page *page = NULL;
  44. 254 /* 如果空闲页面小于最低水位+8,那就回收. */
  45. 255 if (direct_reclaim && z->free_pages < z->pages_min + 8)
  46. 256 page = reclaim_page(z);//把inactive_clean_list队列回收页面
  47. 257 /* If that fails, fall back to rmqueue. */
  48. 258 if (!page)
  49. 259 page = rmqueue(z, order);
  50. 260 if (page)
  51. 261 return page;
  52. 262 }
  53. 263 }
  54. 264
  55. 89
  56. 265 /* Found nothing. */
  57. 266 return NULL;
  58. 267 }
如果还是不行,那就说明管理区的页面很短缺了
  1. [alloc_pages()>__alloc_pages()]
  2. 365 /*
  3. 366 * OK, none of the zones on our zonelist has lots
  4. 367 * of pages free.
  5. 368 *
  6. 369 * We wake up kswapd, in the hope that kswapd will
  7. 370 * resolve this situation before memory gets tight.
  8. 371 *
  9. 372 * We also yield the CPU, because that:
  10. 373 * - gives kswapd a chance to do something
  11. 374 * - slows down allocations, in particular the
  12. 375 * allocations from the fast allocator that‘s
  13. 376 * causing the problems ...
  14. 377 * - ... which minimises the impact the "bad guys"
  15. 378 * have on the rest of the system
  16. 379 * - if we don‘t have __GFP_IO set, kswapd may be
  17. 380 * able to free some memory we can‘t free ourselves
  18. 381 */
  19. 382 wakeup_kswapd(0);//唤醒内核线程,想办法换出一些页面
  20. 383 if (gfp_mask & __GFP_WAIT) {//要求必须获取页面,分配不到时等待,那就让系统再调用一次(目的为了调度kswapd线程)
  21. //以此获取一些页面
  22. 384 __set_current_state(TASK_RUNNING);
  23. 385 current->policy |= SCHED_YIELD;
  24. 386 schedule();
  25. 387 }
  26. 388
  27. 389 /*
  28. 390 * After waking up kswapd, we try to allocate a page
  29. 391 * from any zone which isn‘t critical yet.
  30. 392 *
  31. 393 * Kswapd should, in most situations, bring the situation
  32. 394 * back to normal in no time.
  33. 395 */
  34. /*
  35. 如果不允许等待,那就用pages_min再调用一次__alloc_pages_limit
  36. */
  37. 396 page = __alloc_pages_limit(zonelist, order, PAGES_MIN, direct_reclaim);
  38. 397 if (page)
  39. 398 return page;
  40. 399
要是再失败,那就看谁在要求分配内存页面.如果是kswaped,本身就是内存分配工作者
是要更好的分配页面,比一般进程更重要,那就PF_memalloc标志位为1,不过我们先看一般进程
即pe_memalloc标志位为0的策略.

  1. ==================== mm/page_alloc.c 400 477 ====================
  2. [alloc_pages()>__alloc_pages()]
  3. 400 /*
  4. 401 * Damn, we didn‘t succeed.
  5. 402 *
  6. 403 * This can be due to 2 reasons:
  7. 404 * - we‘re doing a higher-order allocation
  8. 405 * --> move pages to the free list until we succeed
  9. 406 * - we‘re /really/ tight on memory
  10. 407 * --> wait on the kswapd waitqueue until memory is freed
  11. 408 */
  12. 409 if (!(current->flags & PF_MEMALLOC)) {
  13. 410 /*
  14. 411 * Are we dealing with a higher order allocation?
  15. 412 *
  16. 413 * Move pages from the inactive_clean to the free list
  17. 414 * in the hope of creating a large, physically contiguous
  18. 415 * piece of free memory.
  19. 416 */
  20. 417 if (order > 0 && (gfp_mask & __GFP_WAIT)) {
  21. 418 zone = zonelist->zones;
  22. 419 /* First, clean some dirty pages. */
  23. 420 current->flags |= PF_MEMALLOC;
  24. 421 page_launder(gfp_mask, 1);//把脏页洗干净(页面的定期换出)
  25. 422 current->flags &= ~PF_MEMALLOC;
  26. 423 for (;;) {
  27. 424 zone_t *z = *(zone++);//通过一个for循环把干净页面等待队列的页面回收
  28. 425 if (!z)
  29. 426 break;
  30. 427 if (!z->size)
  31. 428 continue;
  32. //是否有干净页面
  33. 429 while (z->inactive_clean_pages) {
  34. 430 struct page * page;
  35. 431 /* Move one page to the free list. */
  36. 432 page = reclaim_page(z);//回收干净页面等待队列
  37. 433 if (!page)
  38. 434 break;
  39. 91
  40. 435 __free_page(page);//通过__free_page释放页面的同时,把空闲页面拼接成大的页面块
  41. 436 /* Try if the allocation succeeds. */
  42. 437 page = rmqueue(z, order);//试图再次请求成功
  43. 438 if (page)
  44. 439 return page;
  45. 440 }
  46. 441 }
  47. 442 }
  48. 443 /*
  49. 444 * When we arrive here, we are really tight on memory.
  50. 445 *
  51. 446 * We wake up kswapd and sleep until kswapd wakes us
  52. 447 * up again. After that we loop back to the start.
  53. 448 *
  54. 449 * We have to do this because something else might eat
  55. 450 * the memory kswapd frees for us and we need to be
  56. 451 * reliable. Note that we don‘t loop back for higher
  57. 452 * order allocations since it is possible that kswapd
  58. 453 * simply cannot free a large enough contiguous area
  59. 454 * of memory *ever*.
  60. 455 */
  61. /*
  62. 如果依旧失败,而且必须要求分配到页面,那就等待,进程睡眠
  63. */
  64. 456 if ((gfp_mask & (__GFP_WAIT|__GFP_IO)) == (__GFP_WAIT|__GFP_IO)) {
  65. 457 wakeup_kswapd(1);//唤醒kswaped,要求分配页面进程睡眠,等待kswapd完成一轮运行再唤醒需要页面的进程
  66. 458 memory_pressure++;
  67. 459 if (!order)//如果要求分配的是1个页面,跳到try_again
  68. 460 goto try_again;
  69. 461 /*
  70. 462 * If __GFP_IO isn‘t set, we can‘t wait on kswapd because
  71. 463 * kswapd just might need some IO locks /we/ are holding ...
  72. 464 *
  73. 465 * SUBTLE: The scheduling point above makes sure that
  74. 466 * kswapd does get the chance to free memory we can‘t
  75. 467 * free ourselves...
  76. 468 */
  77. 469 } else if (gfp_mask & __GFP_WAIT) {
  78. 470 try_to_free_pages(gfp_mask);//另外一种方案...直接调用此函数获取页面(本来就是kswaped函数调用的)
  79. 471 memory_pressure++;
  80. 472 if (!order)
  81. 473 goto try_again;
  82. 474 }
  83. 475
  84. 476 }
  85. 477
技术分享最后的办法了
 
  1. [alloc_pages()>__alloc_pages()]
  2. 478 /*
  3. 479 * Final phase: allocate anything we can!
  4. 480 *
  5. 481 * Higher order allocations, GFP_ATOMIC allocations and
  6. 482 * recursive allocations (PF_MEMALLOC) end up here.
  7. 483 *
  8. 484 * Only recursive allocations can use the very last pages
  9. 485 * in the system, otherwise it would be just too easy to
  10. 486 * deadlock the system...
  11. 487 */
  12. 488 zone = zonelist->zones;
  13. 489 for (;;) {
  14. 490 zone_t *z = *(zone++);
  15. 491 struct page * page = NULL;
  16. 492 if (!z)
  17. 493 break;
  18. 494 if (!z->size)
  19. 495 BUG();
  20. 496
  21. 497 /*
  22. 498 * SUBTLE: direct_reclaim is only possible if the task
  23. 499 * becomes PF_MEMALLOC while looping above. This will
  24. 500 * happen when the OOM killer selects this task for
  25. 501 * instant execution...
  26. 93
  27. 502 */
  28. 503 if (direct_reclaim) {
  29. 504 page = reclaim_page(z);
  30. 505 if (page)
  31. 506 return page;
  32. 507 }
  33. 508
  34. 509 /* XXX: is pages_min/4 a good amount to reserve for this? */
  35. 510 if (z->free_pages < z->pages_min / 4 &&
  36. 511 !(current->flags & PF_MEMALLOC))
  37. 512 continue;
  38. 513 page = rmqueue(z, order);
  39. 514 if (page)
  40. 515 return page;
  41. 516 }
  42. 517
  43. 518 /* No luck.. */
  44. 519 printk(KERN_ERR "__alloc_pages: %lu-order allocation failed.\n", order);
  45. 520 return NULL;
  46. 521 }
如果这都失败,那就系统一定出现了问题了
节点->页面短缺->调用线程,试图腾出页面->
开始遍历每个管理区->一旦管理区的空闲页面大于最低水位,那就调用rmqueue进行分配,否则把kcreclaimd线程唤醒,回收页面
rmqueue分析->如果失败,换一个管理区(按照分配策略),如果全部失败->降低页面的水位要求,把不活跃干净的页面考虑进来
->调用__alloc_pages_limit->如果空闲页面小于最低水位+8,那就回收干净页面队列(换出,腾出空间)->失败,唤醒内核线程,获取页面
->依旧失败,把脏页面洗干净,换出.获取页面->依旧失败,再次调用线程换取页面,依旧失败->把水位降低到1/4看能否满足分配->
依旧不能,系统出了问题
技术分享
 







来自为知笔记(Wiz)


Linux内核情景分析的alloc_pages