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 
70 #include <linux/kernel.h>
71 #include <linux/bootmem.h>
72 #include <linux/swap.h>
73 #include <linux/mm.h>
74 #include <linux/mman.h>
75 #include <linux/module.h>
76 #include <linux/workqueue.h>
77 #include <linux/device.h>
78 #include <xen/balloon.h>
79 #include <xen/tmem.h>
80 #include <xen/xen.h>
81 
82 /* Enable/disable with sysfs. */
83 static int xen_selfballooning_enabled __read_mostly;
84 
85 /*
86  * Controls rate at which memory target (this iteration) approaches
87  * ultimate goal when memory need is increasing (up-hysteresis) or
88  * decreasing (down-hysteresis). Higher values of hysteresis cause
89  * slower increases/decreases. The default values for the various
90  * parameters were deemed reasonable by experimentation, may be
91  * workload-dependent, and can all be adjusted via sysfs.
92  */
93 static unsigned int selfballoon_downhysteresis __read_mostly = 8;
94 static unsigned int selfballoon_uphysteresis __read_mostly = 1;
95 
96 /* In HZ, controls frequency of worker invocation. */
97 static unsigned int selfballoon_interval __read_mostly = 5;
98 
99 /*
100  * Minimum usable RAM in MB for selfballooning target for balloon.
101  * If non-zero, it is added to totalreserve_pages and self-ballooning
102  * will not balloon below the sum.  If zero, a piecewise linear function
103  * is calculated as a minimum and added to totalreserve_pages.  Note that
104  * setting this value indiscriminately may cause OOMs and crashes.
105  */
106 static unsigned int selfballoon_min_usable_mb;
107 
108 static void selfballoon_process(struct work_struct *work);
109 static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
110 
111 #ifdef CONFIG_FRONTSWAP
112 #include <linux/frontswap.h>
113 
114 /* Enable/disable with sysfs. */
115 static bool frontswap_selfshrinking __read_mostly;
116 
117 /* Enable/disable with kernel boot option. */
118 static bool use_frontswap_selfshrink __initdata = true;
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 
xen_nofrontswap_selfshrink_setup(char * s)173 static int __init xen_nofrontswap_selfshrink_setup(char *s)
174 {
175 	use_frontswap_selfshrink = false;
176 	return 1;
177 }
178 
179 __setup("noselfshrink", xen_nofrontswap_selfshrink_setup);
180 
181 /* Disable with kernel boot option. */
182 static bool use_selfballooning __initdata = true;
183 
xen_noselfballooning_setup(char * s)184 static int __init xen_noselfballooning_setup(char *s)
185 {
186 	use_selfballooning = false;
187 	return 1;
188 }
189 
190 __setup("noselfballooning", xen_noselfballooning_setup);
191 #else /* !CONFIG_FRONTSWAP */
192 /* Enable with kernel boot option. */
193 static bool use_selfballooning __initdata = false;
194 
xen_selfballooning_setup(char * s)195 static int __init xen_selfballooning_setup(char *s)
196 {
197 	use_selfballooning = true;
198 	return 1;
199 }
200 
201 __setup("selfballooning", xen_selfballooning_setup);
202 #endif /* CONFIG_FRONTSWAP */
203 
204 #define MB2PAGES(mb)	((mb) << (20 - PAGE_SHIFT))
205 
206 /*
207  * Use current balloon size, the goal (vm_committed_as), and hysteresis
208  * parameters to set a new target balloon size
209  */
selfballoon_process(struct work_struct * work)210 static void selfballoon_process(struct work_struct *work)
211 {
212 	unsigned long cur_pages, goal_pages, tgt_pages, floor_pages;
213 	unsigned long useful_pages;
214 	bool reset_timer = false;
215 
216 	if (xen_selfballooning_enabled) {
217 		cur_pages = totalram_pages;
218 		tgt_pages = cur_pages; /* default is no change */
219 		goal_pages = percpu_counter_read_positive(&vm_committed_as) +
220 				totalreserve_pages;
221 #ifdef CONFIG_FRONTSWAP
222 		/* allow space for frontswap pages to be repatriated */
223 		if (frontswap_selfshrinking && frontswap_enabled)
224 			goal_pages += frontswap_curr_pages();
225 #endif
226 		if (cur_pages > goal_pages)
227 			tgt_pages = cur_pages -
228 				((cur_pages - goal_pages) /
229 				  selfballoon_downhysteresis);
230 		else if (cur_pages < goal_pages)
231 			tgt_pages = cur_pages +
232 				((goal_pages - cur_pages) /
233 				  selfballoon_uphysteresis);
234 		/* else if cur_pages == goal_pages, no change */
235 		useful_pages = max_pfn - totalreserve_pages;
236 		if (selfballoon_min_usable_mb != 0)
237 			floor_pages = totalreserve_pages +
238 					MB2PAGES(selfballoon_min_usable_mb);
239 		/* piecewise linear function ending in ~3% slope */
240 		else if (useful_pages < MB2PAGES(16))
241 			floor_pages = max_pfn; /* not worth ballooning */
242 		else if (useful_pages < MB2PAGES(64))
243 			floor_pages = totalreserve_pages + MB2PAGES(16) +
244 					((useful_pages - MB2PAGES(16)) >> 1);
245 		else if (useful_pages < MB2PAGES(512))
246 			floor_pages = totalreserve_pages + MB2PAGES(40) +
247 					((useful_pages - MB2PAGES(40)) >> 3);
248 		else /* useful_pages >= MB2PAGES(512) */
249 			floor_pages = totalreserve_pages + MB2PAGES(99) +
250 					((useful_pages - MB2PAGES(99)) >> 5);
251 		if (tgt_pages < floor_pages)
252 			tgt_pages = floor_pages;
253 		balloon_set_new_target(tgt_pages +
254 			balloon_stats.current_pages - totalram_pages);
255 		reset_timer = true;
256 	}
257 #ifdef CONFIG_FRONTSWAP
258 	if (frontswap_selfshrinking && frontswap_enabled) {
259 		frontswap_selfshrink();
260 		reset_timer = true;
261 	}
262 #endif
263 	if (reset_timer)
264 		schedule_delayed_work(&selfballoon_worker,
265 			selfballoon_interval * HZ);
266 }
267 
268 #ifdef CONFIG_SYSFS
269 
270 #include <linux/capability.h>
271 
272 #define SELFBALLOON_SHOW(name, format, args...)				\
273 	static ssize_t show_##name(struct device *dev,	\
274 					  struct device_attribute *attr, \
275 					  char *buf) \
276 	{ \
277 		return sprintf(buf, format, ##args); \
278 	}
279 
280 SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
281 
store_selfballooning(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)282 static ssize_t store_selfballooning(struct device *dev,
283 			    struct device_attribute *attr,
284 			    const char *buf,
285 			    size_t count)
286 {
287 	bool was_enabled = xen_selfballooning_enabled;
288 	unsigned long tmp;
289 	int err;
290 
291 	if (!capable(CAP_SYS_ADMIN))
292 		return -EPERM;
293 
294 	err = strict_strtoul(buf, 10, &tmp);
295 	if (err || ((tmp != 0) && (tmp != 1)))
296 		return -EINVAL;
297 
298 	xen_selfballooning_enabled = !!tmp;
299 	if (!was_enabled && xen_selfballooning_enabled)
300 		schedule_delayed_work(&selfballoon_worker,
301 			selfballoon_interval * HZ);
302 
303 	return count;
304 }
305 
306 static DEVICE_ATTR(selfballooning, S_IRUGO | S_IWUSR,
307 		   show_selfballooning, store_selfballooning);
308 
309 SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
310 
store_selfballoon_interval(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)311 static ssize_t store_selfballoon_interval(struct device *dev,
312 					  struct device_attribute *attr,
313 					  const char *buf,
314 					  size_t count)
315 {
316 	unsigned long val;
317 	int err;
318 
319 	if (!capable(CAP_SYS_ADMIN))
320 		return -EPERM;
321 	err = strict_strtoul(buf, 10, &val);
322 	if (err || val == 0)
323 		return -EINVAL;
324 	selfballoon_interval = val;
325 	return count;
326 }
327 
328 static DEVICE_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
329 		   show_selfballoon_interval, store_selfballoon_interval);
330 
331 SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
332 
store_selfballoon_downhys(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)333 static ssize_t store_selfballoon_downhys(struct device *dev,
334 					 struct device_attribute *attr,
335 					 const char *buf,
336 					 size_t count)
337 {
338 	unsigned long val;
339 	int err;
340 
341 	if (!capable(CAP_SYS_ADMIN))
342 		return -EPERM;
343 	err = strict_strtoul(buf, 10, &val);
344 	if (err || val == 0)
345 		return -EINVAL;
346 	selfballoon_downhysteresis = val;
347 	return count;
348 }
349 
350 static DEVICE_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
351 		   show_selfballoon_downhys, store_selfballoon_downhys);
352 
353 
354 SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
355 
store_selfballoon_uphys(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)356 static ssize_t store_selfballoon_uphys(struct device *dev,
357 				       struct device_attribute *attr,
358 				       const char *buf,
359 				       size_t count)
360 {
361 	unsigned long val;
362 	int err;
363 
364 	if (!capable(CAP_SYS_ADMIN))
365 		return -EPERM;
366 	err = strict_strtoul(buf, 10, &val);
367 	if (err || val == 0)
368 		return -EINVAL;
369 	selfballoon_uphysteresis = val;
370 	return count;
371 }
372 
373 static DEVICE_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
374 		   show_selfballoon_uphys, store_selfballoon_uphys);
375 
376 SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
377 				selfballoon_min_usable_mb);
378 
store_selfballoon_min_usable_mb(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)379 static ssize_t store_selfballoon_min_usable_mb(struct device *dev,
380 					       struct device_attribute *attr,
381 					       const char *buf,
382 					       size_t count)
383 {
384 	unsigned long val;
385 	int err;
386 
387 	if (!capable(CAP_SYS_ADMIN))
388 		return -EPERM;
389 	err = strict_strtoul(buf, 10, &val);
390 	if (err || val == 0)
391 		return -EINVAL;
392 	selfballoon_min_usable_mb = val;
393 	return count;
394 }
395 
396 static DEVICE_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
397 		   show_selfballoon_min_usable_mb,
398 		   store_selfballoon_min_usable_mb);
399 
400 
401 #ifdef CONFIG_FRONTSWAP
402 SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
403 
store_frontswap_selfshrinking(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)404 static ssize_t store_frontswap_selfshrinking(struct device *dev,
405 					     struct device_attribute *attr,
406 					     const char *buf,
407 					     size_t count)
408 {
409 	bool was_enabled = frontswap_selfshrinking;
410 	unsigned long tmp;
411 	int err;
412 
413 	if (!capable(CAP_SYS_ADMIN))
414 		return -EPERM;
415 	err = strict_strtoul(buf, 10, &tmp);
416 	if (err || ((tmp != 0) && (tmp != 1)))
417 		return -EINVAL;
418 	frontswap_selfshrinking = !!tmp;
419 	if (!was_enabled && !xen_selfballooning_enabled &&
420 	     frontswap_selfshrinking)
421 		schedule_delayed_work(&selfballoon_worker,
422 			selfballoon_interval * HZ);
423 
424 	return count;
425 }
426 
427 static DEVICE_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
428 		   show_frontswap_selfshrinking, store_frontswap_selfshrinking);
429 
430 SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
431 
store_frontswap_inertia(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)432 static ssize_t store_frontswap_inertia(struct device *dev,
433 				       struct device_attribute *attr,
434 				       const char *buf,
435 				       size_t count)
436 {
437 	unsigned long val;
438 	int err;
439 
440 	if (!capable(CAP_SYS_ADMIN))
441 		return -EPERM;
442 	err = strict_strtoul(buf, 10, &val);
443 	if (err || val == 0)
444 		return -EINVAL;
445 	frontswap_inertia = val;
446 	frontswap_inertia_counter = val;
447 	return count;
448 }
449 
450 static DEVICE_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
451 		   show_frontswap_inertia, store_frontswap_inertia);
452 
453 SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
454 
store_frontswap_hysteresis(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)455 static ssize_t store_frontswap_hysteresis(struct device *dev,
456 					  struct device_attribute *attr,
457 					  const char *buf,
458 					  size_t count)
459 {
460 	unsigned long val;
461 	int err;
462 
463 	if (!capable(CAP_SYS_ADMIN))
464 		return -EPERM;
465 	err = strict_strtoul(buf, 10, &val);
466 	if (err || val == 0)
467 		return -EINVAL;
468 	frontswap_hysteresis = val;
469 	return count;
470 }
471 
472 static DEVICE_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
473 		   show_frontswap_hysteresis, store_frontswap_hysteresis);
474 
475 #endif /* CONFIG_FRONTSWAP */
476 
477 static struct attribute *selfballoon_attrs[] = {
478 	&dev_attr_selfballooning.attr,
479 	&dev_attr_selfballoon_interval.attr,
480 	&dev_attr_selfballoon_downhysteresis.attr,
481 	&dev_attr_selfballoon_uphysteresis.attr,
482 	&dev_attr_selfballoon_min_usable_mb.attr,
483 #ifdef CONFIG_FRONTSWAP
484 	&dev_attr_frontswap_selfshrinking.attr,
485 	&dev_attr_frontswap_hysteresis.attr,
486 	&dev_attr_frontswap_inertia.attr,
487 #endif
488 	NULL
489 };
490 
491 static struct attribute_group selfballoon_group = {
492 	.name = "selfballoon",
493 	.attrs = selfballoon_attrs
494 };
495 #endif
496 
register_xen_selfballooning(struct device * dev)497 int register_xen_selfballooning(struct device *dev)
498 {
499 	int error = -1;
500 
501 #ifdef CONFIG_SYSFS
502 	error = sysfs_create_group(&dev->kobj, &selfballoon_group);
503 #endif
504 	return error;
505 }
506 EXPORT_SYMBOL(register_xen_selfballooning);
507 
xen_selfballoon_init(void)508 static int __init xen_selfballoon_init(void)
509 {
510 	bool enable = false;
511 
512 	if (!xen_domain())
513 		return -ENODEV;
514 
515 	if (xen_initial_domain()) {
516 		pr_info("xen/balloon: Xen selfballooning driver "
517 				"disabled for domain0.\n");
518 		return -ENODEV;
519 	}
520 
521 	xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
522 	if (xen_selfballooning_enabled) {
523 		pr_info("xen/balloon: Initializing Xen "
524 					"selfballooning driver.\n");
525 		enable = true;
526 	}
527 #ifdef CONFIG_FRONTSWAP
528 	frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
529 	if (frontswap_selfshrinking) {
530 		pr_info("xen/balloon: Initializing frontswap "
531 					"selfshrinking driver.\n");
532 		enable = true;
533 	}
534 #endif
535 	if (!enable)
536 		return -ENODEV;
537 
538 	schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
539 
540 	return 0;
541 }
542 
543 subsys_initcall(xen_selfballoon_init);
544 
545 MODULE_LICENSE("GPL");
546