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#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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
81/* Enable/disable with sysfs. */
82static int xen_selfballooning_enabled __read_mostly;
83
84/*
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.
91 */
92static unsigned int selfballoon_downhysteresis __read_mostly = 8;
93static unsigned int selfballoon_uphysteresis __read_mostly = 1;
94
95/* In HZ, controls frequency of worker invocation. */
96static unsigned int selfballoon_interval __read_mostly = 5;
97
98/*
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.
104 */
105static unsigned int selfballoon_min_usable_mb;
106
107/*
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.
110 */
111static unsigned int selfballoon_reserved_mb;
112
113static void selfballoon_process(struct work_struct *work);
114static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
115
116#ifdef CONFIG_FRONTSWAP
117#include <linux/frontswap.h>
118
119/* Enable/disable with sysfs. */
120static bool frontswap_selfshrinking __read_mostly;
121
122/*
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.
126 */
127
128/* Control rate for frontswap shrinking. Higher hysteresis is slower. */
129static unsigned int frontswap_hysteresis __read_mostly = 20;
130
131/*
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.
135 */
136static unsigned int frontswap_inertia __read_mostly = 3;
137
138/* Countdown to next invocation of frontswap_shrink() */
139static unsigned long frontswap_inertia_counter;
140
141/*
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
150 * memory.
151 */
152static void frontswap_selfshrink(void)
153{
154	static unsigned long cur_frontswap_pages;
155	static unsigned long last_frontswap_pages;
156	static unsigned long tgt_frontswap_pages;
157
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;
163		return;
164	}
165	if (frontswap_inertia_counter && --frontswap_inertia_counter)
166		return;
167	if (cur_frontswap_pages <= frontswap_hysteresis)
168		tgt_frontswap_pages = 0;
169	else
170		tgt_frontswap_pages = cur_frontswap_pages -
171			(cur_frontswap_pages / frontswap_hysteresis);
172	frontswap_shrink(tgt_frontswap_pages);
173	frontswap_inertia_counter = frontswap_inertia;
174}
175
176#endif /* CONFIG_FRONTSWAP */
177
178#define MB2PAGES(mb)	((mb) << (20 - PAGE_SHIFT))
179#define PAGES2MB(pages) ((pages) >> (20 - PAGE_SHIFT))
180
181/*
182 * Use current balloon size, the goal (vm_committed_as), and hysteresis
183 * parameters to set a new target balloon size
184 */
185static void selfballoon_process(struct work_struct *work)
186{
187	unsigned long cur_pages, goal_pages, tgt_pages, floor_pages;
188	unsigned long useful_pages;
189	bool reset_timer = false;
190
191	if (xen_selfballooning_enabled) {
192		cur_pages = totalram_pages;
193		tgt_pages = cur_pages; /* default is no change */
194		goal_pages = vm_memory_committed() +
195				totalreserve_pages +
196				MB2PAGES(selfballoon_reserved_mb);
197#ifdef CONFIG_FRONTSWAP
198		/* allow space for frontswap pages to be repatriated */
199		if (frontswap_selfshrinking && frontswap_enabled)
200			goal_pages += frontswap_curr_pages();
201#endif
202		if (cur_pages > goal_pages)
203			tgt_pages = cur_pages -
204				((cur_pages - goal_pages) /
205				  selfballoon_downhysteresis);
206		else if (cur_pages < goal_pages)
207			tgt_pages = cur_pages +
208				((goal_pages - cur_pages) /
209				  selfballoon_uphysteresis);
210		/* else if cur_pages == goal_pages, no change */
211		useful_pages = max_pfn - totalreserve_pages;
212		if (selfballoon_min_usable_mb != 0)
213			floor_pages = totalreserve_pages +
214					MB2PAGES(selfballoon_min_usable_mb);
215		/* piecewise linear function ending in ~3% slope */
216		else if (useful_pages < MB2PAGES(16))
217			floor_pages = max_pfn; /* not worth ballooning */
218		else if (useful_pages < MB2PAGES(64))
219			floor_pages = totalreserve_pages + MB2PAGES(16) +
220					((useful_pages - MB2PAGES(16)) >> 1);
221		else if (useful_pages < MB2PAGES(512))
222			floor_pages = totalreserve_pages + MB2PAGES(40) +
223					((useful_pages - MB2PAGES(40)) >> 3);
224		else /* useful_pages >= MB2PAGES(512) */
225			floor_pages = totalreserve_pages + MB2PAGES(99) +
226					((useful_pages - MB2PAGES(99)) >> 5);
227		if (tgt_pages < floor_pages)
228			tgt_pages = floor_pages;
229		balloon_set_new_target(tgt_pages +
230			balloon_stats.current_pages - totalram_pages);
231		reset_timer = true;
232	}
233#ifdef CONFIG_FRONTSWAP
234	if (frontswap_selfshrinking && frontswap_enabled) {
235		frontswap_selfshrink();
236		reset_timer = true;
237	}
238#endif
239	if (reset_timer)
240		schedule_delayed_work(&selfballoon_worker,
241			selfballoon_interval * HZ);
242}
243
244#ifdef CONFIG_SYSFS
245
246#include <linux/capability.h>
247
248#define SELFBALLOON_SHOW(name, format, args...)				\
249	static ssize_t show_##name(struct device *dev,	\
250					  struct device_attribute *attr, \
251					  char *buf) \
252	{ \
253		return sprintf(buf, format, ##args); \
254	}
255
256SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
257
258static 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
267	if (!capable(CAP_SYS_ADMIN))
268		return -EPERM;
269
270	err = kstrtoul(buf, 10, &tmp);
271	if (err)
272		return err;
273	if ((tmp != 0) && (tmp != 1))
274		return -EINVAL;
275
276	xen_selfballooning_enabled = !!tmp;
277	if (!was_enabled && xen_selfballooning_enabled)
278		schedule_delayed_work(&selfballoon_worker,
279			selfballoon_interval * HZ);
280
281	return count;
282}
283
284static DEVICE_ATTR(selfballooning, S_IRUGO | S_IWUSR,
285		   show_selfballooning, store_selfballooning);
286
287SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
288
289static ssize_t store_selfballoon_interval(struct device *dev,
290					  struct device_attribute *attr,
291					  const char *buf,
292					  size_t count)
293{
294	unsigned long val;
295	int err;
296
297	if (!capable(CAP_SYS_ADMIN))
298		return -EPERM;
299	err = kstrtoul(buf, 10, &val);
300	if (err)
301		return err;
302	if (val == 0)
303		return -EINVAL;
304	selfballoon_interval = val;
305	return count;
306}
307
308static DEVICE_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
309		   show_selfballoon_interval, store_selfballoon_interval);
310
311SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
312
313static ssize_t store_selfballoon_downhys(struct device *dev,
314					 struct device_attribute *attr,
315					 const char *buf,
316					 size_t count)
317{
318	unsigned long val;
319	int err;
320
321	if (!capable(CAP_SYS_ADMIN))
322		return -EPERM;
323	err = kstrtoul(buf, 10, &val);
324	if (err)
325		return err;
326	if (val == 0)
327		return -EINVAL;
328	selfballoon_downhysteresis = val;
329	return count;
330}
331
332static DEVICE_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
333		   show_selfballoon_downhys, store_selfballoon_downhys);
334
335
336SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
337
338static ssize_t store_selfballoon_uphys(struct device *dev,
339				       struct device_attribute *attr,
340				       const char *buf,
341				       size_t count)
342{
343	unsigned long val;
344	int err;
345
346	if (!capable(CAP_SYS_ADMIN))
347		return -EPERM;
348	err = kstrtoul(buf, 10, &val);
349	if (err)
350		return err;
351	if (val == 0)
352		return -EINVAL;
353	selfballoon_uphysteresis = val;
354	return count;
355}
356
357static DEVICE_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
358		   show_selfballoon_uphys, store_selfballoon_uphys);
359
360SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
361				selfballoon_min_usable_mb);
362
363static ssize_t store_selfballoon_min_usable_mb(struct device *dev,
364					       struct device_attribute *attr,
365					       const char *buf,
366					       size_t count)
367{
368	unsigned long val;
369	int err;
370
371	if (!capable(CAP_SYS_ADMIN))
372		return -EPERM;
373	err = kstrtoul(buf, 10, &val);
374	if (err)
375		return err;
376	if (val == 0)
377		return -EINVAL;
378	selfballoon_min_usable_mb = val;
379	return count;
380}
381
382static DEVICE_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
383		   show_selfballoon_min_usable_mb,
384		   store_selfballoon_min_usable_mb);
385
386SELFBALLOON_SHOW(selfballoon_reserved_mb, "%d\n",
387				selfballoon_reserved_mb);
388
389static ssize_t store_selfballoon_reserved_mb(struct device *dev,
390					     struct device_attribute *attr,
391					     const char *buf,
392					     size_t count)
393{
394	unsigned long val;
395	int err;
396
397	if (!capable(CAP_SYS_ADMIN))
398		return -EPERM;
399	err = kstrtoul(buf, 10, &val);
400	if (err)
401		return err;
402	if (val == 0)
403		return -EINVAL;
404	selfballoon_reserved_mb = val;
405	return count;
406}
407
408static DEVICE_ATTR(selfballoon_reserved_mb, S_IRUGO | S_IWUSR,
409		   show_selfballoon_reserved_mb,
410		   store_selfballoon_reserved_mb);
411
412
413#ifdef CONFIG_FRONTSWAP
414SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
415
416static ssize_t store_frontswap_selfshrinking(struct device *dev,
417					     struct device_attribute *attr,
418					     const char *buf,
419					     size_t count)
420{
421	bool was_enabled = frontswap_selfshrinking;
422	unsigned long tmp;
423	int err;
424
425	if (!capable(CAP_SYS_ADMIN))
426		return -EPERM;
427	err = kstrtoul(buf, 10, &tmp);
428	if (err)
429		return err;
430	if ((tmp != 0) && (tmp != 1))
431		return -EINVAL;
432	frontswap_selfshrinking = !!tmp;
433	if (!was_enabled && !xen_selfballooning_enabled &&
434	     frontswap_selfshrinking)
435		schedule_delayed_work(&selfballoon_worker,
436			selfballoon_interval * HZ);
437
438	return count;
439}
440
441static DEVICE_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
442		   show_frontswap_selfshrinking, store_frontswap_selfshrinking);
443
444SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
445
446static ssize_t store_frontswap_inertia(struct device *dev,
447				       struct device_attribute *attr,
448				       const char *buf,
449				       size_t count)
450{
451	unsigned long val;
452	int err;
453
454	if (!capable(CAP_SYS_ADMIN))
455		return -EPERM;
456	err = kstrtoul(buf, 10, &val);
457	if (err)
458		return err;
459	if (val == 0)
460		return -EINVAL;
461	frontswap_inertia = val;
462	frontswap_inertia_counter = val;
463	return count;
464}
465
466static DEVICE_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
467		   show_frontswap_inertia, store_frontswap_inertia);
468
469SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
470
471static ssize_t store_frontswap_hysteresis(struct device *dev,
472					  struct device_attribute *attr,
473					  const char *buf,
474					  size_t count)
475{
476	unsigned long val;
477	int err;
478
479	if (!capable(CAP_SYS_ADMIN))
480		return -EPERM;
481	err = kstrtoul(buf, 10, &val);
482	if (err)
483		return err;
484	if (val == 0)
485		return -EINVAL;
486	frontswap_hysteresis = val;
487	return count;
488}
489
490static DEVICE_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
491		   show_frontswap_hysteresis, store_frontswap_hysteresis);
492
493#endif /* CONFIG_FRONTSWAP */
494
495static struct attribute *selfballoon_attrs[] = {
496	&dev_attr_selfballooning.attr,
497	&dev_attr_selfballoon_interval.attr,
498	&dev_attr_selfballoon_downhysteresis.attr,
499	&dev_attr_selfballoon_uphysteresis.attr,
500	&dev_attr_selfballoon_min_usable_mb.attr,
501	&dev_attr_selfballoon_reserved_mb.attr,
502#ifdef CONFIG_FRONTSWAP
503	&dev_attr_frontswap_selfshrinking.attr,
504	&dev_attr_frontswap_hysteresis.attr,
505	&dev_attr_frontswap_inertia.attr,
506#endif
507	NULL
508};
509
510static const struct attribute_group selfballoon_group = {
511	.name = "selfballoon",
512	.attrs = selfballoon_attrs
513};
514#endif
515
516int register_xen_selfballooning(struct device *dev)
517{
518	int error = -1;
519
520#ifdef CONFIG_SYSFS
521	error = sysfs_create_group(&dev->kobj, &selfballoon_group);
522#endif
523	return error;
524}
525EXPORT_SYMBOL(register_xen_selfballooning);
526
527int xen_selfballoon_init(bool use_selfballooning, bool use_frontswap_selfshrink)
528{
529	bool enable = false;
530	unsigned long reserve_pages;
531
532	if (!xen_domain())
533		return -ENODEV;
534
535	if (xen_initial_domain()) {
536		pr_info("Xen selfballooning driver disabled for domain0\n");
537		return -ENODEV;
538	}
539
540	xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
541	if (xen_selfballooning_enabled) {
542		pr_info("Initializing Xen selfballooning driver\n");
543		enable = true;
544	}
545#ifdef CONFIG_FRONTSWAP
546	frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
547	if (frontswap_selfshrinking) {
548		pr_info("Initializing frontswap selfshrinking driver\n");
549		enable = true;
550	}
551#endif
552	if (!enable)
553		return -ENODEV;
554
555	/*
556	 * Give selfballoon_reserved_mb a default value(10% of total ram pages)
557	 * to make selfballoon not so aggressive.
558	 *
559	 * There are mainly two reasons:
560	 * 1) The original goal_page didn't consider some pages used by kernel
561	 *    space, like slab pages and memory used by device drivers.
562	 *
563	 * 2) The balloon driver may not give back memory to guest OS fast
564	 *    enough when the workload suddenly aquries a lot of physical memory.
565	 *
566	 * In both cases, the guest OS will suffer from memory pressure and
567	 * OOM killer may be triggered.
568	 * By reserving extra 10% of total ram pages, we can keep the system
569	 * much more reliably and response faster in some cases.
570	 */
571	if (!selfballoon_reserved_mb) {
572		reserve_pages = totalram_pages / 10;
573		selfballoon_reserved_mb = PAGES2MB(reserve_pages);
574	}
575	schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
576
577	return 0;
578}
579EXPORT_SYMBOL(xen_selfballoon_init);
580