xen-selfballoon.c 18 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577
  1. /******************************************************************************
  2. * Xen selfballoon driver (and optional frontswap self-shrinking driver)
  3. *
  4. * Copyright (c) 2009-2011, Dan Magenheimer, Oracle Corp.
  5. *
  6. * This code complements the cleancache and frontswap patchsets to optimize
  7. * support for Xen Transcendent Memory ("tmem"). The policy it implements
  8. * is rudimentary and will likely improve over time, but it does work well
  9. * enough today.
  10. *
  11. * Two functionalities are implemented here which both use "control theory"
  12. * (feedback) to optimize memory utilization. In a virtualized environment
  13. * such as Xen, RAM is often a scarce resource and we would like to ensure
  14. * that each of a possibly large number of virtual machines is using RAM
  15. * efficiently, i.e. using as little as possible when under light load
  16. * and obtaining as much as possible when memory demands are high.
  17. * Since RAM needs vary highly dynamically and sometimes dramatically,
  18. * "hysteresis" is used, that is, memory target is determined not just
  19. * on current data but also on past data stored in the system.
  20. *
  21. * "Selfballooning" creates memory pressure by managing the Xen balloon
  22. * driver to decrease and increase available kernel memory, driven
  23. * largely by the target value of "Committed_AS" (see /proc/meminfo).
  24. * Since Committed_AS does not account for clean mapped pages (i.e. pages
  25. * in RAM that are identical to pages on disk), selfballooning has the
  26. * affect of pushing less frequently used clean pagecache pages out of
  27. * kernel RAM and, presumably using cleancache, into Xen tmem where
  28. * Xen can more efficiently optimize RAM utilization for such pages.
  29. *
  30. * When kernel memory demand unexpectedly increases faster than Xen, via
  31. * the selfballoon driver, is able to (or chooses to) provide usable RAM,
  32. * the kernel may invoke swapping. In most cases, frontswap is able
  33. * to absorb this swapping into Xen tmem. However, due to the fact
  34. * that the kernel swap subsystem assumes swapping occurs to a disk,
  35. * swapped pages may sit on the disk for a very long time; even if
  36. * the kernel knows the page will never be used again. This is because
  37. * the disk space costs very little and can be overwritten when
  38. * necessary. When such stale pages are in frontswap, however, they
  39. * are taking up valuable real estate. "Frontswap selfshrinking" works
  40. * to resolve this: When frontswap activity is otherwise stable
  41. * and the guest kernel is not under memory pressure, the "frontswap
  42. * selfshrinking" accounts for this by providing pressure to remove some
  43. * pages from frontswap and return them to kernel memory.
  44. *
  45. * For both "selfballooning" and "frontswap-selfshrinking", a worker
  46. * thread is used and sysfs tunables are provided to adjust the frequency
  47. * and rate of adjustments to achieve the goal, as well as to disable one
  48. * or both functions independently.
  49. *
  50. * While some argue that this functionality can and should be implemented
  51. * in userspace, it has been observed that bad things happen (e.g. OOMs).
  52. *
  53. * System configuration note: Selfballooning should not be enabled on
  54. * systems without a sufficiently large swap device configured; for best
  55. * results, it is recommended that total swap be increased by the size
  56. * of the guest memory. Also, while technically not required to be
  57. * configured, it is highly recommended that frontswap also be configured
  58. * and enabled when selfballooning is running. So, selfballooning
  59. * is disabled by default if frontswap is not configured and can only
  60. * be enabled with the "selfballooning" kernel boot option; similarly
  61. * selfballooning is enabled by default if frontswap is configured and
  62. * can be disabled with the "noselfballooning" kernel boot option. Finally,
  63. * when frontswap is configured, frontswap-selfshrinking can be disabled
  64. * with the "noselfshrink" kernel boot option.
  65. *
  66. * Selfballooning is disallowed in domain0 and force-disabled.
  67. *
  68. */
  69. #include <linux/kernel.h>
  70. #include <linux/bootmem.h>
  71. #include <linux/swap.h>
  72. #include <linux/mm.h>
  73. #include <linux/mman.h>
  74. #include <linux/module.h>
  75. #include <linux/workqueue.h>
  76. #include <linux/device.h>
  77. #include <xen/balloon.h>
  78. #include <xen/tmem.h>
  79. #include <xen/xen.h>
  80. /* Enable/disable with sysfs. */
  81. static int xen_selfballooning_enabled __read_mostly;
  82. /*
  83. * Controls rate at which memory target (this iteration) approaches
  84. * ultimate goal when memory need is increasing (up-hysteresis) or
  85. * decreasing (down-hysteresis). Higher values of hysteresis cause
  86. * slower increases/decreases. The default values for the various
  87. * parameters were deemed reasonable by experimentation, may be
  88. * workload-dependent, and can all be adjusted via sysfs.
  89. */
  90. static unsigned int selfballoon_downhysteresis __read_mostly = 8;
  91. static unsigned int selfballoon_uphysteresis __read_mostly = 1;
  92. /* In HZ, controls frequency of worker invocation. */
  93. static unsigned int selfballoon_interval __read_mostly = 5;
  94. /*
  95. * Minimum usable RAM in MB for selfballooning target for balloon.
  96. * If non-zero, it is added to totalreserve_pages and self-ballooning
  97. * will not balloon below the sum. If zero, a piecewise linear function
  98. * is calculated as a minimum and added to totalreserve_pages. Note that
  99. * setting this value indiscriminately may cause OOMs and crashes.
  100. */
  101. static unsigned int selfballoon_min_usable_mb;
  102. /*
  103. * Amount of RAM in MB to add to the target number of pages.
  104. * Can be used to reserve some more room for caches and the like.
  105. */
  106. static unsigned int selfballoon_reserved_mb;
  107. static void selfballoon_process(struct work_struct *work);
  108. static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
  109. #ifdef CONFIG_FRONTSWAP
  110. #include <linux/frontswap.h>
  111. /* Enable/disable with sysfs. */
  112. static bool frontswap_selfshrinking __read_mostly;
  113. /* Enable/disable with kernel boot option. */
  114. static bool use_frontswap_selfshrink __initdata = true;
  115. /*
  116. * The default values for the following parameters were deemed reasonable
  117. * by experimentation, may be workload-dependent, and can all be
  118. * adjusted via sysfs.
  119. */
  120. /* Control rate for frontswap shrinking. Higher hysteresis is slower. */
  121. static unsigned int frontswap_hysteresis __read_mostly = 20;
  122. /*
  123. * Number of selfballoon worker invocations to wait before observing that
  124. * frontswap selfshrinking should commence. Note that selfshrinking does
  125. * not use a separate worker thread.
  126. */
  127. static unsigned int frontswap_inertia __read_mostly = 3;
  128. /* Countdown to next invocation of frontswap_shrink() */
  129. static unsigned long frontswap_inertia_counter;
  130. /*
  131. * Invoked by the selfballoon worker thread, uses current number of pages
  132. * in frontswap (frontswap_curr_pages()), previous status, and control
  133. * values (hysteresis and inertia) to determine if frontswap should be
  134. * shrunk and what the new frontswap size should be. Note that
  135. * frontswap_shrink is essentially a partial swapoff that immediately
  136. * transfers pages from the "swap device" (frontswap) back into kernel
  137. * RAM; despite the name, frontswap "shrinking" is very different from
  138. * the "shrinker" interface used by the kernel MM subsystem to reclaim
  139. * memory.
  140. */
  141. static void frontswap_selfshrink(void)
  142. {
  143. static unsigned long cur_frontswap_pages;
  144. static unsigned long last_frontswap_pages;
  145. static unsigned long tgt_frontswap_pages;
  146. last_frontswap_pages = cur_frontswap_pages;
  147. cur_frontswap_pages = frontswap_curr_pages();
  148. if (!cur_frontswap_pages ||
  149. (cur_frontswap_pages > last_frontswap_pages)) {
  150. frontswap_inertia_counter = frontswap_inertia;
  151. return;
  152. }
  153. if (frontswap_inertia_counter && --frontswap_inertia_counter)
  154. return;
  155. if (cur_frontswap_pages <= frontswap_hysteresis)
  156. tgt_frontswap_pages = 0;
  157. else
  158. tgt_frontswap_pages = cur_frontswap_pages -
  159. (cur_frontswap_pages / frontswap_hysteresis);
  160. frontswap_shrink(tgt_frontswap_pages);
  161. }
  162. static int __init xen_nofrontswap_selfshrink_setup(char *s)
  163. {
  164. use_frontswap_selfshrink = false;
  165. return 1;
  166. }
  167. __setup("noselfshrink", xen_nofrontswap_selfshrink_setup);
  168. /* Disable with kernel boot option. */
  169. static bool use_selfballooning __initdata = true;
  170. static int __init xen_noselfballooning_setup(char *s)
  171. {
  172. use_selfballooning = false;
  173. return 1;
  174. }
  175. __setup("noselfballooning", xen_noselfballooning_setup);
  176. #else /* !CONFIG_FRONTSWAP */
  177. /* Enable with kernel boot option. */
  178. static bool use_selfballooning __initdata = false;
  179. static int __init xen_selfballooning_setup(char *s)
  180. {
  181. use_selfballooning = true;
  182. return 1;
  183. }
  184. __setup("selfballooning", xen_selfballooning_setup);
  185. #endif /* CONFIG_FRONTSWAP */
  186. #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
  187. /*
  188. * Use current balloon size, the goal (vm_committed_as), and hysteresis
  189. * parameters to set a new target balloon size
  190. */
  191. static void selfballoon_process(struct work_struct *work)
  192. {
  193. unsigned long cur_pages, goal_pages, tgt_pages, floor_pages;
  194. unsigned long useful_pages;
  195. bool reset_timer = false;
  196. if (xen_selfballooning_enabled) {
  197. cur_pages = totalram_pages;
  198. tgt_pages = cur_pages; /* default is no change */
  199. goal_pages = vm_memory_committed() +
  200. totalreserve_pages +
  201. MB2PAGES(selfballoon_reserved_mb);
  202. #ifdef CONFIG_FRONTSWAP
  203. /* allow space for frontswap pages to be repatriated */
  204. if (frontswap_selfshrinking && frontswap_enabled)
  205. goal_pages += frontswap_curr_pages();
  206. #endif
  207. if (cur_pages > goal_pages)
  208. tgt_pages = cur_pages -
  209. ((cur_pages - goal_pages) /
  210. selfballoon_downhysteresis);
  211. else if (cur_pages < goal_pages)
  212. tgt_pages = cur_pages +
  213. ((goal_pages - cur_pages) /
  214. selfballoon_uphysteresis);
  215. /* else if cur_pages == goal_pages, no change */
  216. useful_pages = max_pfn - totalreserve_pages;
  217. if (selfballoon_min_usable_mb != 0)
  218. floor_pages = totalreserve_pages +
  219. MB2PAGES(selfballoon_min_usable_mb);
  220. /* piecewise linear function ending in ~3% slope */
  221. else if (useful_pages < MB2PAGES(16))
  222. floor_pages = max_pfn; /* not worth ballooning */
  223. else if (useful_pages < MB2PAGES(64))
  224. floor_pages = totalreserve_pages + MB2PAGES(16) +
  225. ((useful_pages - MB2PAGES(16)) >> 1);
  226. else if (useful_pages < MB2PAGES(512))
  227. floor_pages = totalreserve_pages + MB2PAGES(40) +
  228. ((useful_pages - MB2PAGES(40)) >> 3);
  229. else /* useful_pages >= MB2PAGES(512) */
  230. floor_pages = totalreserve_pages + MB2PAGES(99) +
  231. ((useful_pages - MB2PAGES(99)) >> 5);
  232. if (tgt_pages < floor_pages)
  233. tgt_pages = floor_pages;
  234. balloon_set_new_target(tgt_pages +
  235. balloon_stats.current_pages - totalram_pages);
  236. reset_timer = true;
  237. }
  238. #ifdef CONFIG_FRONTSWAP
  239. if (frontswap_selfshrinking && frontswap_enabled) {
  240. frontswap_selfshrink();
  241. reset_timer = true;
  242. }
  243. #endif
  244. if (reset_timer)
  245. schedule_delayed_work(&selfballoon_worker,
  246. selfballoon_interval * HZ);
  247. }
  248. #ifdef CONFIG_SYSFS
  249. #include <linux/capability.h>
  250. #define SELFBALLOON_SHOW(name, format, args...) \
  251. static ssize_t show_##name(struct device *dev, \
  252. struct device_attribute *attr, \
  253. char *buf) \
  254. { \
  255. return sprintf(buf, format, ##args); \
  256. }
  257. SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
  258. static ssize_t store_selfballooning(struct device *dev,
  259. struct device_attribute *attr,
  260. const char *buf,
  261. size_t count)
  262. {
  263. bool was_enabled = xen_selfballooning_enabled;
  264. unsigned long tmp;
  265. int err;
  266. if (!capable(CAP_SYS_ADMIN))
  267. return -EPERM;
  268. err = strict_strtoul(buf, 10, &tmp);
  269. if (err || ((tmp != 0) && (tmp != 1)))
  270. return -EINVAL;
  271. xen_selfballooning_enabled = !!tmp;
  272. if (!was_enabled && xen_selfballooning_enabled)
  273. schedule_delayed_work(&selfballoon_worker,
  274. selfballoon_interval * HZ);
  275. return count;
  276. }
  277. static DEVICE_ATTR(selfballooning, S_IRUGO | S_IWUSR,
  278. show_selfballooning, store_selfballooning);
  279. SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
  280. static ssize_t store_selfballoon_interval(struct device *dev,
  281. struct device_attribute *attr,
  282. const char *buf,
  283. size_t count)
  284. {
  285. unsigned long val;
  286. int err;
  287. if (!capable(CAP_SYS_ADMIN))
  288. return -EPERM;
  289. err = strict_strtoul(buf, 10, &val);
  290. if (err || val == 0)
  291. return -EINVAL;
  292. selfballoon_interval = val;
  293. return count;
  294. }
  295. static DEVICE_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
  296. show_selfballoon_interval, store_selfballoon_interval);
  297. SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
  298. static ssize_t store_selfballoon_downhys(struct device *dev,
  299. struct device_attribute *attr,
  300. const char *buf,
  301. size_t count)
  302. {
  303. unsigned long val;
  304. int err;
  305. if (!capable(CAP_SYS_ADMIN))
  306. return -EPERM;
  307. err = strict_strtoul(buf, 10, &val);
  308. if (err || val == 0)
  309. return -EINVAL;
  310. selfballoon_downhysteresis = val;
  311. return count;
  312. }
  313. static DEVICE_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
  314. show_selfballoon_downhys, store_selfballoon_downhys);
  315. SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
  316. static ssize_t store_selfballoon_uphys(struct device *dev,
  317. struct device_attribute *attr,
  318. const char *buf,
  319. size_t count)
  320. {
  321. unsigned long val;
  322. int err;
  323. if (!capable(CAP_SYS_ADMIN))
  324. return -EPERM;
  325. err = strict_strtoul(buf, 10, &val);
  326. if (err || val == 0)
  327. return -EINVAL;
  328. selfballoon_uphysteresis = val;
  329. return count;
  330. }
  331. static DEVICE_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
  332. show_selfballoon_uphys, store_selfballoon_uphys);
  333. SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
  334. selfballoon_min_usable_mb);
  335. static ssize_t store_selfballoon_min_usable_mb(struct device *dev,
  336. struct device_attribute *attr,
  337. const char *buf,
  338. size_t count)
  339. {
  340. unsigned long val;
  341. int err;
  342. if (!capable(CAP_SYS_ADMIN))
  343. return -EPERM;
  344. err = strict_strtoul(buf, 10, &val);
  345. if (err || val == 0)
  346. return -EINVAL;
  347. selfballoon_min_usable_mb = val;
  348. return count;
  349. }
  350. static DEVICE_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
  351. show_selfballoon_min_usable_mb,
  352. store_selfballoon_min_usable_mb);
  353. SELFBALLOON_SHOW(selfballoon_reserved_mb, "%d\n",
  354. selfballoon_reserved_mb);
  355. static ssize_t store_selfballoon_reserved_mb(struct device *dev,
  356. struct device_attribute *attr,
  357. const char *buf,
  358. size_t count)
  359. {
  360. unsigned long val;
  361. int err;
  362. if (!capable(CAP_SYS_ADMIN))
  363. return -EPERM;
  364. err = strict_strtoul(buf, 10, &val);
  365. if (err || val == 0)
  366. return -EINVAL;
  367. selfballoon_reserved_mb = val;
  368. return count;
  369. }
  370. static DEVICE_ATTR(selfballoon_reserved_mb, S_IRUGO | S_IWUSR,
  371. show_selfballoon_reserved_mb,
  372. store_selfballoon_reserved_mb);
  373. #ifdef CONFIG_FRONTSWAP
  374. SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
  375. static ssize_t store_frontswap_selfshrinking(struct device *dev,
  376. struct device_attribute *attr,
  377. const char *buf,
  378. size_t count)
  379. {
  380. bool was_enabled = frontswap_selfshrinking;
  381. unsigned long tmp;
  382. int err;
  383. if (!capable(CAP_SYS_ADMIN))
  384. return -EPERM;
  385. err = strict_strtoul(buf, 10, &tmp);
  386. if (err || ((tmp != 0) && (tmp != 1)))
  387. return -EINVAL;
  388. frontswap_selfshrinking = !!tmp;
  389. if (!was_enabled && !xen_selfballooning_enabled &&
  390. frontswap_selfshrinking)
  391. schedule_delayed_work(&selfballoon_worker,
  392. selfballoon_interval * HZ);
  393. return count;
  394. }
  395. static DEVICE_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
  396. show_frontswap_selfshrinking, store_frontswap_selfshrinking);
  397. SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
  398. static ssize_t store_frontswap_inertia(struct device *dev,
  399. struct device_attribute *attr,
  400. const char *buf,
  401. size_t count)
  402. {
  403. unsigned long val;
  404. int err;
  405. if (!capable(CAP_SYS_ADMIN))
  406. return -EPERM;
  407. err = strict_strtoul(buf, 10, &val);
  408. if (err || val == 0)
  409. return -EINVAL;
  410. frontswap_inertia = val;
  411. frontswap_inertia_counter = val;
  412. return count;
  413. }
  414. static DEVICE_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
  415. show_frontswap_inertia, store_frontswap_inertia);
  416. SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
  417. static ssize_t store_frontswap_hysteresis(struct device *dev,
  418. struct device_attribute *attr,
  419. const char *buf,
  420. size_t count)
  421. {
  422. unsigned long val;
  423. int err;
  424. if (!capable(CAP_SYS_ADMIN))
  425. return -EPERM;
  426. err = strict_strtoul(buf, 10, &val);
  427. if (err || val == 0)
  428. return -EINVAL;
  429. frontswap_hysteresis = val;
  430. return count;
  431. }
  432. static DEVICE_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
  433. show_frontswap_hysteresis, store_frontswap_hysteresis);
  434. #endif /* CONFIG_FRONTSWAP */
  435. static struct attribute *selfballoon_attrs[] = {
  436. &dev_attr_selfballooning.attr,
  437. &dev_attr_selfballoon_interval.attr,
  438. &dev_attr_selfballoon_downhysteresis.attr,
  439. &dev_attr_selfballoon_uphysteresis.attr,
  440. &dev_attr_selfballoon_min_usable_mb.attr,
  441. &dev_attr_selfballoon_reserved_mb.attr,
  442. #ifdef CONFIG_FRONTSWAP
  443. &dev_attr_frontswap_selfshrinking.attr,
  444. &dev_attr_frontswap_hysteresis.attr,
  445. &dev_attr_frontswap_inertia.attr,
  446. #endif
  447. NULL
  448. };
  449. static const struct attribute_group selfballoon_group = {
  450. .name = "selfballoon",
  451. .attrs = selfballoon_attrs
  452. };
  453. #endif
  454. int register_xen_selfballooning(struct device *dev)
  455. {
  456. int error = -1;
  457. #ifdef CONFIG_SYSFS
  458. error = sysfs_create_group(&dev->kobj, &selfballoon_group);
  459. #endif
  460. return error;
  461. }
  462. EXPORT_SYMBOL(register_xen_selfballooning);
  463. static int __init xen_selfballoon_init(void)
  464. {
  465. bool enable = false;
  466. if (!xen_domain())
  467. return -ENODEV;
  468. if (xen_initial_domain()) {
  469. pr_info("xen/balloon: Xen selfballooning driver "
  470. "disabled for domain0.\n");
  471. return -ENODEV;
  472. }
  473. xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
  474. if (xen_selfballooning_enabled) {
  475. pr_info("xen/balloon: Initializing Xen "
  476. "selfballooning driver.\n");
  477. enable = true;
  478. }
  479. #ifdef CONFIG_FRONTSWAP
  480. frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
  481. if (frontswap_selfshrinking) {
  482. pr_info("xen/balloon: Initializing frontswap "
  483. "selfshrinking driver.\n");
  484. enable = true;
  485. }
  486. #endif
  487. if (!enable)
  488. return -ENODEV;
  489. schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
  490. return 0;
  491. }
  492. subsys_initcall(xen_selfballoon_init);
  493. MODULE_LICENSE("GPL");