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. 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.
62 *
63 * Selfballooning is disallowed in domain0 and force-disabled.
64 *
65 */
66
67 #include <linux/kernel.h>
68 #include <linux/bootmem.h>
69 #include <linux/swap.h>
70 #include <linux/mm.h>
71 #include <linux/mman.h>
72 #include <linux/module.h>
73 #include <linux/workqueue.h>
74 #include <linux/device.h>
75 #include <xen/balloon.h>
76 #include <xen/tmem.h>
77 #include <xen/xen.h>
78
79 /* Enable/disable with sysfs. */
80 static int xen_selfballooning_enabled __read_mostly;
81
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
93 /* In HZ, controls frequency of worker invocation. */
94 static unsigned int selfballoon_interval __read_mostly = 5;
95
96 /*
97 * Minimum usable RAM in MB for selfballooning target for balloon.
98 * If non-zero, it is added to totalreserve_pages and self-ballooning
99 * will not balloon below the sum. If zero, a piecewise linear function
100 * is calculated as a minimum and added to totalreserve_pages. Note that
101 * setting this value indiscriminately may cause OOMs and crashes.
102 */
103 static unsigned int selfballoon_min_usable_mb;
104
105 /*
106 * Amount of RAM in MB to add to the target number of pages.
107 * Can be used to reserve some more room for caches and the like.
108 */
109 static unsigned int selfballoon_reserved_mb;
110
111 static void selfballoon_process(struct work_struct *work);
112 static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
113
114 #ifdef CONFIG_FRONTSWAP
115 #include <linux/frontswap.h>
116
117 /* Enable/disable with sysfs. */
118 static bool frontswap_selfshrinking __read_mostly;
119
120 /*
121 * The default values for the following parameters were deemed reasonable
122 * by experimentation, may be workload-dependent, and can all be
123 * adjusted via sysfs.
124 */
125
126 /* Control rate for frontswap shrinking. Higher hysteresis is slower. */
127 static unsigned int frontswap_hysteresis __read_mostly = 20;
128
129 /*
130 * Number of selfballoon worker invocations to wait before observing that
131 * frontswap selfshrinking should commence. Note that selfshrinking does
132 * not use a separate worker thread.
133 */
134 static unsigned int frontswap_inertia __read_mostly = 3;
135
136 /* Countdown to next invocation of frontswap_shrink() */
137 static unsigned long frontswap_inertia_counter;
138
139 /*
140 * Invoked by the selfballoon worker thread, uses current number of pages
141 * in frontswap (frontswap_curr_pages()), previous status, and control
142 * values (hysteresis and inertia) to determine if frontswap should be
143 * shrunk and what the new frontswap size should be. Note that
144 * frontswap_shrink is essentially a partial swapoff that immediately
145 * transfers pages from the "swap device" (frontswap) back into kernel
146 * RAM; despite the name, frontswap "shrinking" is very different from
147 * the "shrinker" interface used by the kernel MM subsystem to reclaim
148 * memory.
149 */
frontswap_selfshrink(void)150 static void frontswap_selfshrink(void)
151 {
152 static unsigned long cur_frontswap_pages;
153 static unsigned long last_frontswap_pages;
154 static unsigned long tgt_frontswap_pages;
155
156 last_frontswap_pages = cur_frontswap_pages;
157 cur_frontswap_pages = frontswap_curr_pages();
158 if (!cur_frontswap_pages ||
159 (cur_frontswap_pages > last_frontswap_pages)) {
160 frontswap_inertia_counter = frontswap_inertia;
161 return;
162 }
163 if (frontswap_inertia_counter && --frontswap_inertia_counter)
164 return;
165 if (cur_frontswap_pages <= frontswap_hysteresis)
166 tgt_frontswap_pages = 0;
167 else
168 tgt_frontswap_pages = cur_frontswap_pages -
169 (cur_frontswap_pages / frontswap_hysteresis);
170 frontswap_shrink(tgt_frontswap_pages);
171 }
172
173 #endif /* CONFIG_FRONTSWAP */
174
175 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
176
177 /*
178 * Use current balloon size, the goal (vm_committed_as), and hysteresis
179 * parameters to set a new target balloon size
180 */
selfballoon_process(struct work_struct * work)181 static void selfballoon_process(struct work_struct *work)
182 {
183 unsigned long cur_pages, goal_pages, tgt_pages, floor_pages;
184 unsigned long useful_pages;
185 bool reset_timer = false;
186
187 if (xen_selfballooning_enabled) {
188 cur_pages = totalram_pages;
189 tgt_pages = cur_pages; /* default is no change */
190 goal_pages = vm_memory_committed() +
191 totalreserve_pages +
192 MB2PAGES(selfballoon_reserved_mb);
193 #ifdef CONFIG_FRONTSWAP
194 /* allow space for frontswap pages to be repatriated */
195 if (frontswap_selfshrinking && frontswap_enabled)
196 goal_pages += frontswap_curr_pages();
197 #endif
198 if (cur_pages > goal_pages)
199 tgt_pages = cur_pages -
200 ((cur_pages - goal_pages) /
201 selfballoon_downhysteresis);
202 else if (cur_pages < goal_pages)
203 tgt_pages = cur_pages +
204 ((goal_pages - cur_pages) /
205 selfballoon_uphysteresis);
206 /* else if cur_pages == goal_pages, no change */
207 useful_pages = max_pfn - totalreserve_pages;
208 if (selfballoon_min_usable_mb != 0)
209 floor_pages = totalreserve_pages +
210 MB2PAGES(selfballoon_min_usable_mb);
211 /* piecewise linear function ending in ~3% slope */
212 else if (useful_pages < MB2PAGES(16))
213 floor_pages = max_pfn; /* not worth ballooning */
214 else if (useful_pages < MB2PAGES(64))
215 floor_pages = totalreserve_pages + MB2PAGES(16) +
216 ((useful_pages - MB2PAGES(16)) >> 1);
217 else if (useful_pages < MB2PAGES(512))
218 floor_pages = totalreserve_pages + MB2PAGES(40) +
219 ((useful_pages - MB2PAGES(40)) >> 3);
220 else /* useful_pages >= MB2PAGES(512) */
221 floor_pages = totalreserve_pages + MB2PAGES(99) +
222 ((useful_pages - MB2PAGES(99)) >> 5);
223 if (tgt_pages < floor_pages)
224 tgt_pages = floor_pages;
225 balloon_set_new_target(tgt_pages +
226 balloon_stats.current_pages - totalram_pages);
227 reset_timer = true;
228 }
229 #ifdef CONFIG_FRONTSWAP
230 if (frontswap_selfshrinking && frontswap_enabled) {
231 frontswap_selfshrink();
232 reset_timer = true;
233 }
234 #endif
235 if (reset_timer)
236 schedule_delayed_work(&selfballoon_worker,
237 selfballoon_interval * HZ);
238 }
239
240 #ifdef CONFIG_SYSFS
241
242 #include <linux/capability.h>
243
244 #define SELFBALLOON_SHOW(name, format, args...) \
245 static ssize_t show_##name(struct device *dev, \
246 struct device_attribute *attr, \
247 char *buf) \
248 { \
249 return sprintf(buf, format, ##args); \
250 }
251
252 SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
253
store_selfballooning(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)254 static ssize_t store_selfballooning(struct device *dev,
255 struct device_attribute *attr,
256 const char *buf,
257 size_t count)
258 {
259 bool was_enabled = xen_selfballooning_enabled;
260 unsigned long tmp;
261 int err;
262
263 if (!capable(CAP_SYS_ADMIN))
264 return -EPERM;
265
266 err = strict_strtoul(buf, 10, &tmp);
267 if (err || ((tmp != 0) && (tmp != 1)))
268 return -EINVAL;
269
270 xen_selfballooning_enabled = !!tmp;
271 if (!was_enabled && xen_selfballooning_enabled)
272 schedule_delayed_work(&selfballoon_worker,
273 selfballoon_interval * HZ);
274
275 return count;
276 }
277
278 static DEVICE_ATTR(selfballooning, S_IRUGO | S_IWUSR,
279 show_selfballooning, store_selfballooning);
280
281 SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
282
store_selfballoon_interval(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)283 static ssize_t store_selfballoon_interval(struct device *dev,
284 struct device_attribute *attr,
285 const char *buf,
286 size_t count)
287 {
288 unsigned long val;
289 int err;
290
291 if (!capable(CAP_SYS_ADMIN))
292 return -EPERM;
293 err = strict_strtoul(buf, 10, &val);
294 if (err || val == 0)
295 return -EINVAL;
296 selfballoon_interval = val;
297 return count;
298 }
299
300 static DEVICE_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
301 show_selfballoon_interval, store_selfballoon_interval);
302
303 SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
304
store_selfballoon_downhys(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)305 static ssize_t store_selfballoon_downhys(struct device *dev,
306 struct device_attribute *attr,
307 const char *buf,
308 size_t count)
309 {
310 unsigned long val;
311 int err;
312
313 if (!capable(CAP_SYS_ADMIN))
314 return -EPERM;
315 err = strict_strtoul(buf, 10, &val);
316 if (err || val == 0)
317 return -EINVAL;
318 selfballoon_downhysteresis = val;
319 return count;
320 }
321
322 static DEVICE_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
323 show_selfballoon_downhys, store_selfballoon_downhys);
324
325
326 SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
327
store_selfballoon_uphys(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)328 static ssize_t store_selfballoon_uphys(struct device *dev,
329 struct device_attribute *attr,
330 const char *buf,
331 size_t count)
332 {
333 unsigned long val;
334 int err;
335
336 if (!capable(CAP_SYS_ADMIN))
337 return -EPERM;
338 err = strict_strtoul(buf, 10, &val);
339 if (err || val == 0)
340 return -EINVAL;
341 selfballoon_uphysteresis = val;
342 return count;
343 }
344
345 static DEVICE_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
346 show_selfballoon_uphys, store_selfballoon_uphys);
347
348 SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
349 selfballoon_min_usable_mb);
350
store_selfballoon_min_usable_mb(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)351 static ssize_t store_selfballoon_min_usable_mb(struct device *dev,
352 struct device_attribute *attr,
353 const char *buf,
354 size_t count)
355 {
356 unsigned long val;
357 int err;
358
359 if (!capable(CAP_SYS_ADMIN))
360 return -EPERM;
361 err = strict_strtoul(buf, 10, &val);
362 if (err || val == 0)
363 return -EINVAL;
364 selfballoon_min_usable_mb = val;
365 return count;
366 }
367
368 static DEVICE_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
369 show_selfballoon_min_usable_mb,
370 store_selfballoon_min_usable_mb);
371
372 SELFBALLOON_SHOW(selfballoon_reserved_mb, "%d\n",
373 selfballoon_reserved_mb);
374
store_selfballoon_reserved_mb(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)375 static ssize_t store_selfballoon_reserved_mb(struct device *dev,
376 struct device_attribute *attr,
377 const char *buf,
378 size_t count)
379 {
380 unsigned long val;
381 int err;
382
383 if (!capable(CAP_SYS_ADMIN))
384 return -EPERM;
385 err = strict_strtoul(buf, 10, &val);
386 if (err || val == 0)
387 return -EINVAL;
388 selfballoon_reserved_mb = val;
389 return count;
390 }
391
392 static DEVICE_ATTR(selfballoon_reserved_mb, S_IRUGO | S_IWUSR,
393 show_selfballoon_reserved_mb,
394 store_selfballoon_reserved_mb);
395
396
397 #ifdef CONFIG_FRONTSWAP
398 SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
399
store_frontswap_selfshrinking(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)400 static ssize_t store_frontswap_selfshrinking(struct device *dev,
401 struct device_attribute *attr,
402 const char *buf,
403 size_t count)
404 {
405 bool was_enabled = frontswap_selfshrinking;
406 unsigned long tmp;
407 int err;
408
409 if (!capable(CAP_SYS_ADMIN))
410 return -EPERM;
411 err = strict_strtoul(buf, 10, &tmp);
412 if (err || ((tmp != 0) && (tmp != 1)))
413 return -EINVAL;
414 frontswap_selfshrinking = !!tmp;
415 if (!was_enabled && !xen_selfballooning_enabled &&
416 frontswap_selfshrinking)
417 schedule_delayed_work(&selfballoon_worker,
418 selfballoon_interval * HZ);
419
420 return count;
421 }
422
423 static DEVICE_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
424 show_frontswap_selfshrinking, store_frontswap_selfshrinking);
425
426 SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
427
store_frontswap_inertia(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)428 static ssize_t store_frontswap_inertia(struct device *dev,
429 struct device_attribute *attr,
430 const char *buf,
431 size_t count)
432 {
433 unsigned long val;
434 int err;
435
436 if (!capable(CAP_SYS_ADMIN))
437 return -EPERM;
438 err = strict_strtoul(buf, 10, &val);
439 if (err || val == 0)
440 return -EINVAL;
441 frontswap_inertia = val;
442 frontswap_inertia_counter = val;
443 return count;
444 }
445
446 static DEVICE_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
447 show_frontswap_inertia, store_frontswap_inertia);
448
449 SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
450
store_frontswap_hysteresis(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)451 static ssize_t store_frontswap_hysteresis(struct device *dev,
452 struct device_attribute *attr,
453 const char *buf,
454 size_t count)
455 {
456 unsigned long val;
457 int err;
458
459 if (!capable(CAP_SYS_ADMIN))
460 return -EPERM;
461 err = strict_strtoul(buf, 10, &val);
462 if (err || val == 0)
463 return -EINVAL;
464 frontswap_hysteresis = val;
465 return count;
466 }
467
468 static DEVICE_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
469 show_frontswap_hysteresis, store_frontswap_hysteresis);
470
471 #endif /* CONFIG_FRONTSWAP */
472
473 static struct attribute *selfballoon_attrs[] = {
474 &dev_attr_selfballooning.attr,
475 &dev_attr_selfballoon_interval.attr,
476 &dev_attr_selfballoon_downhysteresis.attr,
477 &dev_attr_selfballoon_uphysteresis.attr,
478 &dev_attr_selfballoon_min_usable_mb.attr,
479 &dev_attr_selfballoon_reserved_mb.attr,
480 #ifdef CONFIG_FRONTSWAP
481 &dev_attr_frontswap_selfshrinking.attr,
482 &dev_attr_frontswap_hysteresis.attr,
483 &dev_attr_frontswap_inertia.attr,
484 #endif
485 NULL
486 };
487
488 static const struct attribute_group selfballoon_group = {
489 .name = "selfballoon",
490 .attrs = selfballoon_attrs
491 };
492 #endif
493
register_xen_selfballooning(struct device * dev)494 int register_xen_selfballooning(struct device *dev)
495 {
496 int error = -1;
497
498 #ifdef CONFIG_SYSFS
499 error = sysfs_create_group(&dev->kobj, &selfballoon_group);
500 #endif
501 return error;
502 }
503 EXPORT_SYMBOL(register_xen_selfballooning);
504
xen_selfballoon_init(bool use_selfballooning,bool use_frontswap_selfshrink)505 int xen_selfballoon_init(bool use_selfballooning, bool use_frontswap_selfshrink)
506 {
507 bool enable = false;
508
509 if (!xen_domain())
510 return -ENODEV;
511
512 if (xen_initial_domain()) {
513 pr_info("xen/balloon: Xen selfballooning driver "
514 "disabled for domain0.\n");
515 return -ENODEV;
516 }
517
518 xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
519 if (xen_selfballooning_enabled) {
520 pr_info("xen/balloon: Initializing Xen "
521 "selfballooning driver.\n");
522 enable = true;
523 }
524 #ifdef CONFIG_FRONTSWAP
525 frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
526 if (frontswap_selfshrinking) {
527 pr_info("xen/balloon: Initializing frontswap "
528 "selfshrinking driver.\n");
529 enable = true;
530 }
531 #endif
532 if (!enable)
533 return -ENODEV;
534
535 schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
536
537 return 0;
538 }
539 EXPORT_SYMBOL(xen_selfballoon_init);
540