1 /******************************************************************************
2 * Xen selfballoon driver (and optional frontswap self-shrinking driver)
4 * Copyright (c) 2009-2011, Dan Magenheimer, Oracle Corp.
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
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.
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.
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.
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.
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).
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.
66 * Selfballooning is disallowed in domain0 and force-disabled.
70 #include <linux/kernel.h>
71 #include <linux/bootmem.h>
72 #include <linux/swap.h>
74 #include <linux/mman.h>
75 #include <linux/module.h>
76 #include <linux/workqueue.h>
77 #include <xen/balloon.h>
81 /* Enable/disable with sysfs. */
82 static int xen_selfballooning_enabled __read_mostly;
85 * Controls rate at which memory target (this iteration) approaches
86 * ultimate goal when memory need is increasing (up-hysteresis) or
87 * decreasing (down-hysteresis). Higher values of hysteresis cause
88 * slower increases/decreases. The default values for the various
89 * parameters were deemed reasonable by experimentation, may be
90 * workload-dependent, and can all be adjusted via sysfs.
92 static unsigned int selfballoon_downhysteresis __read_mostly = 8;
93 static unsigned int selfballoon_uphysteresis __read_mostly = 1;
95 /* In HZ, controls frequency of worker invocation. */
96 static unsigned int selfballoon_interval __read_mostly = 5;
99 * Minimum usable RAM in MB for selfballooning target for balloon.
100 * If non-zero, it is added to totalreserve_pages and self-ballooning
101 * will not balloon below the sum. If zero, a piecewise linear function
102 * is calculated as a minimum and added to totalreserve_pages. Note that
103 * setting this value indiscriminately may cause OOMs and crashes.
105 static unsigned int selfballoon_min_usable_mb;
107 static void selfballoon_process(struct work_struct *work);
108 static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
110 #ifdef CONFIG_FRONTSWAP
111 #include <linux/frontswap.h>
113 /* Enable/disable with sysfs. */
114 static bool frontswap_selfshrinking __read_mostly;
116 /* Enable/disable with kernel boot option. */
117 static bool use_frontswap_selfshrink __initdata = true;
120 * The default values for the following parameters were deemed reasonable
121 * by experimentation, may be workload-dependent, and can all be
122 * adjusted via sysfs.
125 /* Control rate for frontswap shrinking. Higher hysteresis is slower. */
126 static unsigned int frontswap_hysteresis __read_mostly = 20;
129 * Number of selfballoon worker invocations to wait before observing that
130 * frontswap selfshrinking should commence. Note that selfshrinking does
131 * not use a separate worker thread.
133 static unsigned int frontswap_inertia __read_mostly = 3;
135 /* Countdown to next invocation of frontswap_shrink() */
136 static unsigned long frontswap_inertia_counter;
139 * Invoked by the selfballoon worker thread, uses current number of pages
140 * in frontswap (frontswap_curr_pages()), previous status, and control
141 * values (hysteresis and inertia) to determine if frontswap should be
142 * shrunk and what the new frontswap size should be. Note that
143 * frontswap_shrink is essentially a partial swapoff that immediately
144 * transfers pages from the "swap device" (frontswap) back into kernel
145 * RAM; despite the name, frontswap "shrinking" is very different from
146 * the "shrinker" interface used by the kernel MM subsystem to reclaim
149 static void frontswap_selfshrink(void)
151 static unsigned long cur_frontswap_pages;
152 static unsigned long last_frontswap_pages;
153 static unsigned long tgt_frontswap_pages;
155 last_frontswap_pages = cur_frontswap_pages;
156 cur_frontswap_pages = frontswap_curr_pages();
157 if (!cur_frontswap_pages ||
158 (cur_frontswap_pages > last_frontswap_pages)) {
159 frontswap_inertia_counter = frontswap_inertia;
162 if (frontswap_inertia_counter && --frontswap_inertia_counter)
164 if (cur_frontswap_pages <= frontswap_hysteresis)
165 tgt_frontswap_pages = 0;
167 tgt_frontswap_pages = cur_frontswap_pages -
168 (cur_frontswap_pages / frontswap_hysteresis);
169 frontswap_shrink(tgt_frontswap_pages);
172 static int __init xen_nofrontswap_selfshrink_setup(char *s)
174 use_frontswap_selfshrink = false;
178 __setup("noselfshrink", xen_nofrontswap_selfshrink_setup);
180 /* Disable with kernel boot option. */
181 static bool use_selfballooning __initdata = true;
183 static int __init xen_noselfballooning_setup(char *s)
185 use_selfballooning = false;
189 __setup("noselfballooning", xen_noselfballooning_setup);
190 #else /* !CONFIG_FRONTSWAP */
191 /* Enable with kernel boot option. */
192 static bool use_selfballooning __initdata = false;
194 static int __init xen_selfballooning_setup(char *s)
196 use_selfballooning = true;
200 __setup("selfballooning", xen_selfballooning_setup);
201 #endif /* CONFIG_FRONTSWAP */
203 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
206 * Use current balloon size, the goal (vm_committed_as), and hysteresis
207 * parameters to set a new target balloon size
209 static void selfballoon_process(struct work_struct *work)
211 unsigned long cur_pages, goal_pages, tgt_pages, floor_pages;
212 unsigned long useful_pages;
213 bool reset_timer = false;
215 if (xen_selfballooning_enabled) {
216 cur_pages = totalram_pages;
217 tgt_pages = cur_pages; /* default is no change */
218 goal_pages = percpu_counter_read_positive(&vm_committed_as) +
220 #ifdef CONFIG_FRONTSWAP
221 /* allow space for frontswap pages to be repatriated */
222 if (frontswap_selfshrinking && frontswap_enabled)
223 goal_pages += frontswap_curr_pages();
225 if (cur_pages > goal_pages)
226 tgt_pages = cur_pages -
227 ((cur_pages - goal_pages) /
228 selfballoon_downhysteresis);
229 else if (cur_pages < goal_pages)
230 tgt_pages = cur_pages +
231 ((goal_pages - cur_pages) /
232 selfballoon_uphysteresis);
233 /* else if cur_pages == goal_pages, no change */
234 useful_pages = max_pfn - totalreserve_pages;
235 if (selfballoon_min_usable_mb != 0)
236 floor_pages = totalreserve_pages +
237 MB2PAGES(selfballoon_min_usable_mb);
238 /* piecewise linear function ending in ~3% slope */
239 else if (useful_pages < MB2PAGES(16))
240 floor_pages = max_pfn; /* not worth ballooning */
241 else if (useful_pages < MB2PAGES(64))
242 floor_pages = totalreserve_pages + MB2PAGES(16) +
243 ((useful_pages - MB2PAGES(16)) >> 1);
244 else if (useful_pages < MB2PAGES(512))
245 floor_pages = totalreserve_pages + MB2PAGES(40) +
246 ((useful_pages - MB2PAGES(40)) >> 3);
247 else /* useful_pages >= MB2PAGES(512) */
248 floor_pages = totalreserve_pages + MB2PAGES(99) +
249 ((useful_pages - MB2PAGES(99)) >> 5);
250 if (tgt_pages < floor_pages)
251 tgt_pages = floor_pages;
252 balloon_set_new_target(tgt_pages +
253 balloon_stats.current_pages - totalram_pages);
256 #ifdef CONFIG_FRONTSWAP
257 if (frontswap_selfshrinking && frontswap_enabled) {
258 frontswap_selfshrink();
263 schedule_delayed_work(&selfballoon_worker,
264 selfballoon_interval * HZ);
269 #include <linux/capability.h>
271 #define SELFBALLOON_SHOW(name, format, args...) \
272 static ssize_t show_##name(struct device *dev, \
273 struct device_attribute *attr, \
276 return sprintf(buf, format, ##args); \
279 SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
281 static ssize_t store_selfballooning(struct device *dev,
282 struct device_attribute *attr,
286 bool was_enabled = xen_selfballooning_enabled;
290 if (!capable(CAP_SYS_ADMIN))
293 err = strict_strtoul(buf, 10, &tmp);
294 if (err || ((tmp != 0) && (tmp != 1)))
297 xen_selfballooning_enabled = !!tmp;
298 if (!was_enabled && xen_selfballooning_enabled)
299 schedule_delayed_work(&selfballoon_worker,
300 selfballoon_interval * HZ);
305 static DEVICE_ATTR(selfballooning, S_IRUGO | S_IWUSR,
306 show_selfballooning, store_selfballooning);
308 SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
310 static ssize_t store_selfballoon_interval(struct device *dev,
311 struct device_attribute *attr,
318 if (!capable(CAP_SYS_ADMIN))
320 err = strict_strtoul(buf, 10, &val);
323 selfballoon_interval = val;
327 static DEVICE_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
328 show_selfballoon_interval, store_selfballoon_interval);
330 SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
332 static ssize_t store_selfballoon_downhys(struct device *dev,
333 struct device_attribute *attr,
340 if (!capable(CAP_SYS_ADMIN))
342 err = strict_strtoul(buf, 10, &val);
345 selfballoon_downhysteresis = val;
349 static DEVICE_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
350 show_selfballoon_downhys, store_selfballoon_downhys);
353 SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
355 static ssize_t store_selfballoon_uphys(struct device *dev,
356 struct device_attribute *attr,
363 if (!capable(CAP_SYS_ADMIN))
365 err = strict_strtoul(buf, 10, &val);
368 selfballoon_uphysteresis = val;
372 static DEVICE_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
373 show_selfballoon_uphys, store_selfballoon_uphys);
375 SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
376 selfballoon_min_usable_mb);
378 static ssize_t store_selfballoon_min_usable_mb(struct device *dev,
379 struct device_attribute *attr,
386 if (!capable(CAP_SYS_ADMIN))
388 err = strict_strtoul(buf, 10, &val);
391 selfballoon_min_usable_mb = val;
395 static DEVICE_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
396 show_selfballoon_min_usable_mb,
397 store_selfballoon_min_usable_mb);
400 #ifdef CONFIG_FRONTSWAP
401 SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
403 static ssize_t store_frontswap_selfshrinking(struct device *dev,
404 struct device_attribute *attr,
408 bool was_enabled = frontswap_selfshrinking;
412 if (!capable(CAP_SYS_ADMIN))
414 err = strict_strtoul(buf, 10, &tmp);
415 if (err || ((tmp != 0) && (tmp != 1)))
417 frontswap_selfshrinking = !!tmp;
418 if (!was_enabled && !xen_selfballooning_enabled &&
419 frontswap_selfshrinking)
420 schedule_delayed_work(&selfballoon_worker,
421 selfballoon_interval * HZ);
426 static DEVICE_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
427 show_frontswap_selfshrinking, store_frontswap_selfshrinking);
429 SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
431 static ssize_t store_frontswap_inertia(struct device *dev,
432 struct device_attribute *attr,
439 if (!capable(CAP_SYS_ADMIN))
441 err = strict_strtoul(buf, 10, &val);
444 frontswap_inertia = val;
445 frontswap_inertia_counter = val;
449 static DEVICE_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
450 show_frontswap_inertia, store_frontswap_inertia);
452 SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
454 static ssize_t store_frontswap_hysteresis(struct device *dev,
455 struct device_attribute *attr,
462 if (!capable(CAP_SYS_ADMIN))
464 err = strict_strtoul(buf, 10, &val);
467 frontswap_hysteresis = val;
471 static DEVICE_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
472 show_frontswap_hysteresis, store_frontswap_hysteresis);
474 #endif /* CONFIG_FRONTSWAP */
476 static struct attribute *selfballoon_attrs[] = {
477 &dev_attr_selfballooning.attr,
478 &dev_attr_selfballoon_interval.attr,
479 &dev_attr_selfballoon_downhysteresis.attr,
480 &dev_attr_selfballoon_uphysteresis.attr,
481 &dev_attr_selfballoon_min_usable_mb.attr,
482 #ifdef CONFIG_FRONTSWAP
483 &dev_attr_frontswap_selfshrinking.attr,
484 &dev_attr_frontswap_hysteresis.attr,
485 &dev_attr_frontswap_inertia.attr,
490 static struct attribute_group selfballoon_group = {
491 .name = "selfballoon",
492 .attrs = selfballoon_attrs
496 int register_xen_selfballooning(struct device *dev)
501 error = sysfs_create_group(&dev->kobj, &selfballoon_group);
505 EXPORT_SYMBOL(register_xen_selfballooning);
507 static int __init xen_selfballoon_init(void)
514 if (xen_initial_domain()) {
515 pr_info("xen/balloon: Xen selfballooning driver "
516 "disabled for domain0.\n");
520 xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
521 if (xen_selfballooning_enabled) {
522 pr_info("xen/balloon: Initializing Xen "
523 "selfballooning driver.\n");
526 #ifdef CONFIG_FRONTSWAP
527 frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
528 if (frontswap_selfshrinking) {
529 pr_info("xen/balloon: Initializing frontswap "
530 "selfshrinking driver.\n");
537 schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
542 subsys_initcall(xen_selfballoon_init);
544 MODULE_LICENSE("GPL");