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. Note, that selfballooning should be disabled by default
57 * if frontswap is not configured. Similarly selfballooning should be enabled
58 * by default if frontswap is configured and can be disabled with the
59 * "tmem.selfballooning=0" kernel boot option. Finally, when frontswap is
60 * configured, frontswap-selfshrinking can be disabled with the
61 * "tmem.selfshrink=0" kernel boot option.
63 * Selfballooning is disallowed in domain0 and force-disabled.
67 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
69 #include <linux/kernel.h>
70 #include <linux/bootmem.h>
71 #include <linux/swap.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>
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
;
108 * Amount of RAM in MB to add to the target number of pages.
109 * Can be used to reserve some more room for caches and the like.
111 static unsigned int selfballoon_reserved_mb
;
113 static void selfballoon_process(struct work_struct
*work
);
114 static DECLARE_DELAYED_WORK(selfballoon_worker
, selfballoon_process
);
116 #ifdef CONFIG_FRONTSWAP
117 #include <linux/frontswap.h>
119 /* Enable/disable with sysfs. */
120 static bool frontswap_selfshrinking __read_mostly
;
123 * The default values for the following parameters were deemed reasonable
124 * by experimentation, may be workload-dependent, and can all be
125 * adjusted via sysfs.
128 /* Control rate for frontswap shrinking. Higher hysteresis is slower. */
129 static unsigned int frontswap_hysteresis __read_mostly
= 20;
132 * Number of selfballoon worker invocations to wait before observing that
133 * frontswap selfshrinking should commence. Note that selfshrinking does
134 * not use a separate worker thread.
136 static unsigned int frontswap_inertia __read_mostly
= 3;
138 /* Countdown to next invocation of frontswap_shrink() */
139 static unsigned long frontswap_inertia_counter
;
142 * Invoked by the selfballoon worker thread, uses current number of pages
143 * in frontswap (frontswap_curr_pages()), previous status, and control
144 * values (hysteresis and inertia) to determine if frontswap should be
145 * shrunk and what the new frontswap size should be. Note that
146 * frontswap_shrink is essentially a partial swapoff that immediately
147 * transfers pages from the "swap device" (frontswap) back into kernel
148 * RAM; despite the name, frontswap "shrinking" is very different from
149 * the "shrinker" interface used by the kernel MM subsystem to reclaim
152 static void frontswap_selfshrink(void)
154 static unsigned long cur_frontswap_pages
;
155 static unsigned long last_frontswap_pages
;
156 static unsigned long tgt_frontswap_pages
;
158 last_frontswap_pages
= cur_frontswap_pages
;
159 cur_frontswap_pages
= frontswap_curr_pages();
160 if (!cur_frontswap_pages
||
161 (cur_frontswap_pages
> last_frontswap_pages
)) {
162 frontswap_inertia_counter
= frontswap_inertia
;
165 if (frontswap_inertia_counter
&& --frontswap_inertia_counter
)
167 if (cur_frontswap_pages
<= frontswap_hysteresis
)
168 tgt_frontswap_pages
= 0;
170 tgt_frontswap_pages
= cur_frontswap_pages
-
171 (cur_frontswap_pages
/ frontswap_hysteresis
);
172 frontswap_shrink(tgt_frontswap_pages
);
175 #endif /* CONFIG_FRONTSWAP */
177 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
178 #define PAGES2MB(pages) ((pages) >> (20 - PAGE_SHIFT))
181 * Use current balloon size, the goal (vm_committed_as), and hysteresis
182 * parameters to set a new target balloon size
184 static void selfballoon_process(struct work_struct
*work
)
186 unsigned long cur_pages
, goal_pages
, tgt_pages
, floor_pages
;
187 unsigned long useful_pages
;
188 bool reset_timer
= false;
190 if (xen_selfballooning_enabled
) {
191 cur_pages
= totalram_pages
;
192 tgt_pages
= cur_pages
; /* default is no change */
193 goal_pages
= vm_memory_committed() +
195 MB2PAGES(selfballoon_reserved_mb
);
196 #ifdef CONFIG_FRONTSWAP
197 /* allow space for frontswap pages to be repatriated */
198 if (frontswap_selfshrinking
&& frontswap_enabled
)
199 goal_pages
+= frontswap_curr_pages();
201 if (cur_pages
> goal_pages
)
202 tgt_pages
= cur_pages
-
203 ((cur_pages
- goal_pages
) /
204 selfballoon_downhysteresis
);
205 else if (cur_pages
< goal_pages
)
206 tgt_pages
= cur_pages
+
207 ((goal_pages
- cur_pages
) /
208 selfballoon_uphysteresis
);
209 /* else if cur_pages == goal_pages, no change */
210 useful_pages
= max_pfn
- totalreserve_pages
;
211 if (selfballoon_min_usable_mb
!= 0)
212 floor_pages
= totalreserve_pages
+
213 MB2PAGES(selfballoon_min_usable_mb
);
214 /* piecewise linear function ending in ~3% slope */
215 else if (useful_pages
< MB2PAGES(16))
216 floor_pages
= max_pfn
; /* not worth ballooning */
217 else if (useful_pages
< MB2PAGES(64))
218 floor_pages
= totalreserve_pages
+ MB2PAGES(16) +
219 ((useful_pages
- MB2PAGES(16)) >> 1);
220 else if (useful_pages
< MB2PAGES(512))
221 floor_pages
= totalreserve_pages
+ MB2PAGES(40) +
222 ((useful_pages
- MB2PAGES(40)) >> 3);
223 else /* useful_pages >= MB2PAGES(512) */
224 floor_pages
= totalreserve_pages
+ MB2PAGES(99) +
225 ((useful_pages
- MB2PAGES(99)) >> 5);
226 if (tgt_pages
< floor_pages
)
227 tgt_pages
= floor_pages
;
228 balloon_set_new_target(tgt_pages
+
229 balloon_stats
.current_pages
- totalram_pages
);
232 #ifdef CONFIG_FRONTSWAP
233 if (frontswap_selfshrinking
&& frontswap_enabled
) {
234 frontswap_selfshrink();
239 schedule_delayed_work(&selfballoon_worker
,
240 selfballoon_interval
* HZ
);
245 #include <linux/capability.h>
247 #define SELFBALLOON_SHOW(name, format, args...) \
248 static ssize_t show_##name(struct device *dev, \
249 struct device_attribute *attr, \
252 return sprintf(buf, format, ##args); \
255 SELFBALLOON_SHOW(selfballooning
, "%d\n", xen_selfballooning_enabled
);
257 static ssize_t
store_selfballooning(struct device
*dev
,
258 struct device_attribute
*attr
,
262 bool was_enabled
= xen_selfballooning_enabled
;
266 if (!capable(CAP_SYS_ADMIN
))
269 err
= kstrtoul(buf
, 10, &tmp
);
272 if ((tmp
!= 0) && (tmp
!= 1))
275 xen_selfballooning_enabled
= !!tmp
;
276 if (!was_enabled
&& xen_selfballooning_enabled
)
277 schedule_delayed_work(&selfballoon_worker
,
278 selfballoon_interval
* HZ
);
283 static DEVICE_ATTR(selfballooning
, S_IRUGO
| S_IWUSR
,
284 show_selfballooning
, store_selfballooning
);
286 SELFBALLOON_SHOW(selfballoon_interval
, "%d\n", selfballoon_interval
);
288 static ssize_t
store_selfballoon_interval(struct device
*dev
,
289 struct device_attribute
*attr
,
296 if (!capable(CAP_SYS_ADMIN
))
298 err
= kstrtoul(buf
, 10, &val
);
303 selfballoon_interval
= val
;
307 static DEVICE_ATTR(selfballoon_interval
, S_IRUGO
| S_IWUSR
,
308 show_selfballoon_interval
, store_selfballoon_interval
);
310 SELFBALLOON_SHOW(selfballoon_downhys
, "%d\n", selfballoon_downhysteresis
);
312 static ssize_t
store_selfballoon_downhys(struct device
*dev
,
313 struct device_attribute
*attr
,
320 if (!capable(CAP_SYS_ADMIN
))
322 err
= kstrtoul(buf
, 10, &val
);
327 selfballoon_downhysteresis
= val
;
331 static DEVICE_ATTR(selfballoon_downhysteresis
, S_IRUGO
| S_IWUSR
,
332 show_selfballoon_downhys
, store_selfballoon_downhys
);
335 SELFBALLOON_SHOW(selfballoon_uphys
, "%d\n", selfballoon_uphysteresis
);
337 static ssize_t
store_selfballoon_uphys(struct device
*dev
,
338 struct device_attribute
*attr
,
345 if (!capable(CAP_SYS_ADMIN
))
347 err
= kstrtoul(buf
, 10, &val
);
352 selfballoon_uphysteresis
= val
;
356 static DEVICE_ATTR(selfballoon_uphysteresis
, S_IRUGO
| S_IWUSR
,
357 show_selfballoon_uphys
, store_selfballoon_uphys
);
359 SELFBALLOON_SHOW(selfballoon_min_usable_mb
, "%d\n",
360 selfballoon_min_usable_mb
);
362 static ssize_t
store_selfballoon_min_usable_mb(struct device
*dev
,
363 struct device_attribute
*attr
,
370 if (!capable(CAP_SYS_ADMIN
))
372 err
= kstrtoul(buf
, 10, &val
);
377 selfballoon_min_usable_mb
= val
;
381 static DEVICE_ATTR(selfballoon_min_usable_mb
, S_IRUGO
| S_IWUSR
,
382 show_selfballoon_min_usable_mb
,
383 store_selfballoon_min_usable_mb
);
385 SELFBALLOON_SHOW(selfballoon_reserved_mb
, "%d\n",
386 selfballoon_reserved_mb
);
388 static ssize_t
store_selfballoon_reserved_mb(struct device
*dev
,
389 struct device_attribute
*attr
,
396 if (!capable(CAP_SYS_ADMIN
))
398 err
= kstrtoul(buf
, 10, &val
);
403 selfballoon_reserved_mb
= val
;
407 static DEVICE_ATTR(selfballoon_reserved_mb
, S_IRUGO
| S_IWUSR
,
408 show_selfballoon_reserved_mb
,
409 store_selfballoon_reserved_mb
);
412 #ifdef CONFIG_FRONTSWAP
413 SELFBALLOON_SHOW(frontswap_selfshrinking
, "%d\n", frontswap_selfshrinking
);
415 static ssize_t
store_frontswap_selfshrinking(struct device
*dev
,
416 struct device_attribute
*attr
,
420 bool was_enabled
= frontswap_selfshrinking
;
424 if (!capable(CAP_SYS_ADMIN
))
426 err
= kstrtoul(buf
, 10, &tmp
);
429 if ((tmp
!= 0) && (tmp
!= 1))
431 frontswap_selfshrinking
= !!tmp
;
432 if (!was_enabled
&& !xen_selfballooning_enabled
&&
433 frontswap_selfshrinking
)
434 schedule_delayed_work(&selfballoon_worker
,
435 selfballoon_interval
* HZ
);
440 static DEVICE_ATTR(frontswap_selfshrinking
, S_IRUGO
| S_IWUSR
,
441 show_frontswap_selfshrinking
, store_frontswap_selfshrinking
);
443 SELFBALLOON_SHOW(frontswap_inertia
, "%d\n", frontswap_inertia
);
445 static ssize_t
store_frontswap_inertia(struct device
*dev
,
446 struct device_attribute
*attr
,
453 if (!capable(CAP_SYS_ADMIN
))
455 err
= kstrtoul(buf
, 10, &val
);
460 frontswap_inertia
= val
;
461 frontswap_inertia_counter
= val
;
465 static DEVICE_ATTR(frontswap_inertia
, S_IRUGO
| S_IWUSR
,
466 show_frontswap_inertia
, store_frontswap_inertia
);
468 SELFBALLOON_SHOW(frontswap_hysteresis
, "%d\n", frontswap_hysteresis
);
470 static ssize_t
store_frontswap_hysteresis(struct device
*dev
,
471 struct device_attribute
*attr
,
478 if (!capable(CAP_SYS_ADMIN
))
480 err
= kstrtoul(buf
, 10, &val
);
485 frontswap_hysteresis
= val
;
489 static DEVICE_ATTR(frontswap_hysteresis
, S_IRUGO
| S_IWUSR
,
490 show_frontswap_hysteresis
, store_frontswap_hysteresis
);
492 #endif /* CONFIG_FRONTSWAP */
494 static struct attribute
*selfballoon_attrs
[] = {
495 &dev_attr_selfballooning
.attr
,
496 &dev_attr_selfballoon_interval
.attr
,
497 &dev_attr_selfballoon_downhysteresis
.attr
,
498 &dev_attr_selfballoon_uphysteresis
.attr
,
499 &dev_attr_selfballoon_min_usable_mb
.attr
,
500 &dev_attr_selfballoon_reserved_mb
.attr
,
501 #ifdef CONFIG_FRONTSWAP
502 &dev_attr_frontswap_selfshrinking
.attr
,
503 &dev_attr_frontswap_hysteresis
.attr
,
504 &dev_attr_frontswap_inertia
.attr
,
509 static const struct attribute_group selfballoon_group
= {
510 .name
= "selfballoon",
511 .attrs
= selfballoon_attrs
515 int register_xen_selfballooning(struct device
*dev
)
520 error
= sysfs_create_group(&dev
->kobj
, &selfballoon_group
);
524 EXPORT_SYMBOL(register_xen_selfballooning
);
526 int xen_selfballoon_init(bool use_selfballooning
, bool use_frontswap_selfshrink
)
529 unsigned long reserve_pages
;
534 if (xen_initial_domain()) {
535 pr_info("Xen selfballooning driver disabled for domain0\n");
539 xen_selfballooning_enabled
= tmem_enabled
&& use_selfballooning
;
540 if (xen_selfballooning_enabled
) {
541 pr_info("Initializing Xen selfballooning driver\n");
544 #ifdef CONFIG_FRONTSWAP
545 frontswap_selfshrinking
= tmem_enabled
&& use_frontswap_selfshrink
;
546 if (frontswap_selfshrinking
) {
547 pr_info("Initializing frontswap selfshrinking driver\n");
555 * Give selfballoon_reserved_mb a default value(10% of total ram pages)
556 * to make selfballoon not so aggressive.
558 * There are mainly two reasons:
559 * 1) The original goal_page didn't consider some pages used by kernel
560 * space, like slab pages and memory used by device drivers.
562 * 2) The balloon driver may not give back memory to guest OS fast
563 * enough when the workload suddenly aquries a lot of physical memory.
565 * In both cases, the guest OS will suffer from memory pressure and
566 * OOM killer may be triggered.
567 * By reserving extra 10% of total ram pages, we can keep the system
568 * much more reliably and response faster in some cases.
570 if (!selfballoon_reserved_mb
) {
571 reserve_pages
= totalram_pages
/ 10;
572 selfballoon_reserved_mb
= PAGES2MB(reserve_pages
);
574 schedule_delayed_work(&selfballoon_worker
, selfballoon_interval
* HZ
);
578 EXPORT_SYMBOL(xen_selfballoon_init
);