Linux Audio

Check our new training course

Loading...
v6.2
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Linux VM pressure
  4 *
  5 * Copyright 2012 Linaro Ltd.
  6 *		  Anton Vorontsov <anton.vorontsov@linaro.org>
  7 *
  8 * Based on ideas from Andrew Morton, David Rientjes, KOSAKI Motohiro,
  9 * Leonid Moiseichuk, Mel Gorman, Minchan Kim and Pekka Enberg.
 
 
 
 
 10 */
 11
 12#include <linux/cgroup.h>
 13#include <linux/fs.h>
 14#include <linux/log2.h>
 15#include <linux/sched.h>
 16#include <linux/mm.h>
 17#include <linux/vmstat.h>
 18#include <linux/eventfd.h>
 19#include <linux/slab.h>
 20#include <linux/swap.h>
 21#include <linux/printk.h>
 22#include <linux/vmpressure.h>
 23
 24/*
 25 * The window size (vmpressure_win) is the number of scanned pages before
 26 * we try to analyze scanned/reclaimed ratio. So the window is used as a
 27 * rate-limit tunable for the "low" level notification, and also for
 28 * averaging the ratio for medium/critical levels. Using small window
 29 * sizes can cause lot of false positives, but too big window size will
 30 * delay the notifications.
 31 *
 32 * As the vmscan reclaimer logic works with chunks which are multiple of
 33 * SWAP_CLUSTER_MAX, it makes sense to use it for the window size as well.
 34 *
 35 * TODO: Make the window size depend on machine size, as we do for vmstat
 36 * thresholds. Currently we set it to 512 pages (2MB for 4KB pages).
 37 */
 38static const unsigned long vmpressure_win = SWAP_CLUSTER_MAX * 16;
 39
 40/*
 41 * These thresholds are used when we account memory pressure through
 42 * scanned/reclaimed ratio. The current values were chosen empirically. In
 43 * essence, they are percents: the higher the value, the more number
 44 * unsuccessful reclaims there were.
 45 */
 46static const unsigned int vmpressure_level_med = 60;
 47static const unsigned int vmpressure_level_critical = 95;
 48
 49/*
 50 * When there are too little pages left to scan, vmpressure() may miss the
 51 * critical pressure as number of pages will be less than "window size".
 52 * However, in that case the vmscan priority will raise fast as the
 53 * reclaimer will try to scan LRUs more deeply.
 54 *
 55 * The vmscan logic considers these special priorities:
 56 *
 57 * prio == DEF_PRIORITY (12): reclaimer starts with that value
 58 * prio <= DEF_PRIORITY - 2 : kswapd becomes somewhat overwhelmed
 59 * prio == 0                : close to OOM, kernel scans every page in an lru
 60 *
 61 * Any value in this range is acceptable for this tunable (i.e. from 12 to
 62 * 0). Current value for the vmpressure_level_critical_prio is chosen
 63 * empirically, but the number, in essence, means that we consider
 64 * critical level when scanning depth is ~10% of the lru size (vmscan
 65 * scans 'lru_size >> prio' pages, so it is actually 12.5%, or one
 66 * eights).
 67 */
 68static const unsigned int vmpressure_level_critical_prio = ilog2(100 / 10);
 69
 70static struct vmpressure *work_to_vmpressure(struct work_struct *work)
 71{
 72	return container_of(work, struct vmpressure, work);
 73}
 74
 75static struct vmpressure *vmpressure_parent(struct vmpressure *vmpr)
 76{
 77	struct mem_cgroup *memcg = vmpressure_to_memcg(vmpr);
 
 78
 79	memcg = parent_mem_cgroup(memcg);
 80	if (!memcg)
 81		return NULL;
 82	return memcg_to_vmpressure(memcg);
 83}
 84
 85enum vmpressure_levels {
 86	VMPRESSURE_LOW = 0,
 87	VMPRESSURE_MEDIUM,
 88	VMPRESSURE_CRITICAL,
 89	VMPRESSURE_NUM_LEVELS,
 90};
 91
 92enum vmpressure_modes {
 93	VMPRESSURE_NO_PASSTHROUGH = 0,
 94	VMPRESSURE_HIERARCHY,
 95	VMPRESSURE_LOCAL,
 96	VMPRESSURE_NUM_MODES,
 97};
 98
 99static const char * const vmpressure_str_levels[] = {
100	[VMPRESSURE_LOW] = "low",
101	[VMPRESSURE_MEDIUM] = "medium",
102	[VMPRESSURE_CRITICAL] = "critical",
103};
104
105static const char * const vmpressure_str_modes[] = {
106	[VMPRESSURE_NO_PASSTHROUGH] = "default",
107	[VMPRESSURE_HIERARCHY] = "hierarchy",
108	[VMPRESSURE_LOCAL] = "local",
109};
110
111static enum vmpressure_levels vmpressure_level(unsigned long pressure)
112{
113	if (pressure >= vmpressure_level_critical)
114		return VMPRESSURE_CRITICAL;
115	else if (pressure >= vmpressure_level_med)
116		return VMPRESSURE_MEDIUM;
117	return VMPRESSURE_LOW;
118}
119
120static enum vmpressure_levels vmpressure_calc_level(unsigned long scanned,
121						    unsigned long reclaimed)
122{
123	unsigned long scale = scanned + reclaimed;
124	unsigned long pressure = 0;
125
126	/*
127	 * reclaimed can be greater than scanned for things such as reclaimed
128	 * slab pages. shrink_node() just adds reclaimed pages without a
129	 * related increment to scanned pages.
130	 */
131	if (reclaimed >= scanned)
132		goto out;
133	/*
134	 * We calculate the ratio (in percents) of how many pages were
135	 * scanned vs. reclaimed in a given time frame (window). Note that
136	 * time is in VM reclaimer's "ticks", i.e. number of pages
137	 * scanned. This makes it possible to set desired reaction time
138	 * and serves as a ratelimit.
139	 */
140	pressure = scale - (reclaimed * scale / scanned);
141	pressure = pressure * 100 / scale;
142
143out:
144	pr_debug("%s: %3lu  (s: %lu  r: %lu)\n", __func__, pressure,
145		 scanned, reclaimed);
146
147	return vmpressure_level(pressure);
148}
149
150struct vmpressure_event {
151	struct eventfd_ctx *efd;
152	enum vmpressure_levels level;
153	enum vmpressure_modes mode;
154	struct list_head node;
155};
156
157static bool vmpressure_event(struct vmpressure *vmpr,
158			     const enum vmpressure_levels level,
159			     bool ancestor, bool signalled)
160{
161	struct vmpressure_event *ev;
162	bool ret = false;
163
164	mutex_lock(&vmpr->events_lock);
 
165	list_for_each_entry(ev, &vmpr->events, node) {
166		if (ancestor && ev->mode == VMPRESSURE_LOCAL)
167			continue;
168		if (signalled && ev->mode == VMPRESSURE_NO_PASSTHROUGH)
169			continue;
170		if (level < ev->level)
171			continue;
172		eventfd_signal(ev->efd, 1);
173		ret = true;
174	}
 
175	mutex_unlock(&vmpr->events_lock);
176
177	return ret;
178}
179
180static void vmpressure_work_fn(struct work_struct *work)
181{
182	struct vmpressure *vmpr = work_to_vmpressure(work);
183	unsigned long scanned;
184	unsigned long reclaimed;
185	enum vmpressure_levels level;
186	bool ancestor = false;
187	bool signalled = false;
188
189	spin_lock(&vmpr->sr_lock);
190	/*
191	 * Several contexts might be calling vmpressure(), so it is
192	 * possible that the work was rescheduled again before the old
193	 * work context cleared the counters. In that case we will run
194	 * just after the old work returns, but then scanned might be zero
195	 * here. No need for any locks here since we don't care if
196	 * vmpr->reclaimed is in sync.
197	 */
198	scanned = vmpr->tree_scanned;
199	if (!scanned) {
200		spin_unlock(&vmpr->sr_lock);
201		return;
202	}
203
204	reclaimed = vmpr->tree_reclaimed;
205	vmpr->tree_scanned = 0;
206	vmpr->tree_reclaimed = 0;
207	spin_unlock(&vmpr->sr_lock);
208
209	level = vmpressure_calc_level(scanned, reclaimed);
210
211	do {
212		if (vmpressure_event(vmpr, level, ancestor, signalled))
213			signalled = true;
214		ancestor = true;
 
 
 
215	} while ((vmpr = vmpressure_parent(vmpr)));
216}
217
218/**
219 * vmpressure() - Account memory pressure through scanned/reclaimed ratio
220 * @gfp:	reclaimer's gfp mask
221 * @memcg:	cgroup memory controller handle
222 * @tree:	legacy subtree mode
223 * @scanned:	number of pages scanned
224 * @reclaimed:	number of pages reclaimed
225 *
226 * This function should be called from the vmscan reclaim path to account
227 * "instantaneous" memory pressure (scanned/reclaimed ratio). The raw
228 * pressure index is then further refined and averaged over time.
229 *
230 * If @tree is set, vmpressure is in traditional userspace reporting
231 * mode: @memcg is considered the pressure root and userspace is
232 * notified of the entire subtree's reclaim efficiency.
233 *
234 * If @tree is not set, reclaim efficiency is recorded for @memcg, and
235 * only in-kernel users are notified.
236 *
237 * This function does not return any value.
238 */
239void vmpressure(gfp_t gfp, struct mem_cgroup *memcg, bool tree,
240		unsigned long scanned, unsigned long reclaimed)
241{
242	struct vmpressure *vmpr;
243
244	if (mem_cgroup_disabled())
245		return;
246
247	vmpr = memcg_to_vmpressure(memcg);
248
249	/*
250	 * Here we only want to account pressure that userland is able to
251	 * help us with. For example, suppose that DMA zone is under
252	 * pressure; if we notify userland about that kind of pressure,
253	 * then it will be mostly a waste as it will trigger unnecessary
254	 * freeing of memory by userland (since userland is more likely to
255	 * have HIGHMEM/MOVABLE pages instead of the DMA fallback). That
256	 * is why we include only movable, highmem and FS/IO pages.
257	 * Indirect reclaim (kswapd) sets sc->gfp_mask to GFP_KERNEL, so
258	 * we account it too.
259	 */
260	if (!(gfp & (__GFP_HIGHMEM | __GFP_MOVABLE | __GFP_IO | __GFP_FS)))
261		return;
262
263	/*
264	 * If we got here with no pages scanned, then that is an indicator
265	 * that reclaimer was unable to find any shrinkable LRUs at the
266	 * current scanning depth. But it does not mean that we should
267	 * report the critical pressure, yet. If the scanning priority
268	 * (scanning depth) goes too high (deep), we will be notified
269	 * through vmpressure_prio(). But so far, keep calm.
270	 */
271	if (!scanned)
272		return;
273
274	if (tree) {
275		spin_lock(&vmpr->sr_lock);
276		scanned = vmpr->tree_scanned += scanned;
277		vmpr->tree_reclaimed += reclaimed;
278		spin_unlock(&vmpr->sr_lock);
279
280		if (scanned < vmpressure_win)
281			return;
282		schedule_work(&vmpr->work);
283	} else {
284		enum vmpressure_levels level;
285
286		/* For now, no users for root-level efficiency */
287		if (!memcg || mem_cgroup_is_root(memcg))
288			return;
289
290		spin_lock(&vmpr->sr_lock);
291		scanned = vmpr->scanned += scanned;
292		reclaimed = vmpr->reclaimed += reclaimed;
293		if (scanned < vmpressure_win) {
294			spin_unlock(&vmpr->sr_lock);
295			return;
296		}
297		vmpr->scanned = vmpr->reclaimed = 0;
298		spin_unlock(&vmpr->sr_lock);
299
300		level = vmpressure_calc_level(scanned, reclaimed);
301
302		if (level > VMPRESSURE_LOW) {
303			/*
304			 * Let the socket buffer allocator know that
305			 * we are having trouble reclaiming LRU pages.
306			 *
307			 * For hysteresis keep the pressure state
308			 * asserted for a second in which subsequent
309			 * pressure events can occur.
310			 */
311			WRITE_ONCE(memcg->socket_pressure, jiffies + HZ);
312		}
313	}
314}
315
316/**
317 * vmpressure_prio() - Account memory pressure through reclaimer priority level
318 * @gfp:	reclaimer's gfp mask
319 * @memcg:	cgroup memory controller handle
320 * @prio:	reclaimer's priority
321 *
322 * This function should be called from the reclaim path every time when
323 * the vmscan's reclaiming priority (scanning depth) changes.
324 *
325 * This function does not return any value.
326 */
327void vmpressure_prio(gfp_t gfp, struct mem_cgroup *memcg, int prio)
328{
329	/*
330	 * We only use prio for accounting critical level. For more info
331	 * see comment for vmpressure_level_critical_prio variable above.
332	 */
333	if (prio > vmpressure_level_critical_prio)
334		return;
335
336	/*
337	 * OK, the prio is below the threshold, updating vmpressure
338	 * information before shrinker dives into long shrinking of long
339	 * range vmscan. Passing scanned = vmpressure_win, reclaimed = 0
340	 * to the vmpressure() basically means that we signal 'critical'
341	 * level.
342	 */
343	vmpressure(gfp, memcg, true, vmpressure_win, 0);
344}
345
346#define MAX_VMPRESSURE_ARGS_LEN	(strlen("critical") + strlen("hierarchy") + 2)
347
348/**
349 * vmpressure_register_event() - Bind vmpressure notifications to an eventfd
350 * @memcg:	memcg that is interested in vmpressure notifications
351 * @eventfd:	eventfd context to link notifications with
352 * @args:	event arguments (pressure level threshold, optional mode)
353 *
354 * This function associates eventfd context with the vmpressure
355 * infrastructure, so that the notifications will be delivered to the
356 * @eventfd. The @args parameter is a comma-delimited string that denotes a
357 * pressure level threshold (one of vmpressure_str_levels, i.e. "low", "medium",
358 * or "critical") and an optional mode (one of vmpressure_str_modes, i.e.
359 * "hierarchy" or "local").
360 *
361 * To be used as memcg event method.
362 *
363 * Return: 0 on success, -ENOMEM on memory failure or -EINVAL if @args could
364 * not be parsed.
365 */
366int vmpressure_register_event(struct mem_cgroup *memcg,
367			      struct eventfd_ctx *eventfd, const char *args)
368{
369	struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
370	struct vmpressure_event *ev;
371	enum vmpressure_modes mode = VMPRESSURE_NO_PASSTHROUGH;
372	enum vmpressure_levels level;
373	char *spec, *spec_orig;
374	char *token;
375	int ret = 0;
376
377	spec_orig = spec = kstrndup(args, MAX_VMPRESSURE_ARGS_LEN, GFP_KERNEL);
378	if (!spec)
379		return -ENOMEM;
380
381	/* Find required level */
382	token = strsep(&spec, ",");
383	ret = match_string(vmpressure_str_levels, VMPRESSURE_NUM_LEVELS, token);
384	if (ret < 0)
385		goto out;
386	level = ret;
387
388	/* Find optional mode */
389	token = strsep(&spec, ",");
390	if (token) {
391		ret = match_string(vmpressure_str_modes, VMPRESSURE_NUM_MODES, token);
392		if (ret < 0)
393			goto out;
394		mode = ret;
395	}
396
 
 
 
397	ev = kzalloc(sizeof(*ev), GFP_KERNEL);
398	if (!ev) {
399		ret = -ENOMEM;
400		goto out;
401	}
402
403	ev->efd = eventfd;
404	ev->level = level;
405	ev->mode = mode;
406
407	mutex_lock(&vmpr->events_lock);
408	list_add(&ev->node, &vmpr->events);
409	mutex_unlock(&vmpr->events_lock);
410	ret = 0;
411out:
412	kfree(spec_orig);
413	return ret;
414}
415
416/**
417 * vmpressure_unregister_event() - Unbind eventfd from vmpressure
418 * @memcg:	memcg handle
419 * @eventfd:	eventfd context that was used to link vmpressure with the @cg
420 *
421 * This function does internal manipulations to detach the @eventfd from
422 * the vmpressure notifications, and then frees internal resources
423 * associated with the @eventfd (but the @eventfd itself is not freed).
424 *
425 * To be used as memcg event method.
426 */
427void vmpressure_unregister_event(struct mem_cgroup *memcg,
428				 struct eventfd_ctx *eventfd)
429{
430	struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
431	struct vmpressure_event *ev;
432
433	mutex_lock(&vmpr->events_lock);
434	list_for_each_entry(ev, &vmpr->events, node) {
435		if (ev->efd != eventfd)
436			continue;
437		list_del(&ev->node);
438		kfree(ev);
439		break;
440	}
441	mutex_unlock(&vmpr->events_lock);
442}
443
444/**
445 * vmpressure_init() - Initialize vmpressure control structure
446 * @vmpr:	Structure to be initialized
447 *
448 * This function should be called on every allocated vmpressure structure
449 * before any usage.
450 */
451void vmpressure_init(struct vmpressure *vmpr)
452{
453	spin_lock_init(&vmpr->sr_lock);
454	mutex_init(&vmpr->events_lock);
455	INIT_LIST_HEAD(&vmpr->events);
456	INIT_WORK(&vmpr->work, vmpressure_work_fn);
457}
458
459/**
460 * vmpressure_cleanup() - shuts down vmpressure control structure
461 * @vmpr:	Structure to be cleaned up
462 *
463 * This function should be called before the structure in which it is
464 * embedded is cleaned up.
465 */
466void vmpressure_cleanup(struct vmpressure *vmpr)
467{
468	/*
469	 * Make sure there is no pending work before eventfd infrastructure
470	 * goes away.
471	 */
472	flush_work(&vmpr->work);
473}
v4.10.11
 
  1/*
  2 * Linux VM pressure
  3 *
  4 * Copyright 2012 Linaro Ltd.
  5 *		  Anton Vorontsov <anton.vorontsov@linaro.org>
  6 *
  7 * Based on ideas from Andrew Morton, David Rientjes, KOSAKI Motohiro,
  8 * Leonid Moiseichuk, Mel Gorman, Minchan Kim and Pekka Enberg.
  9 *
 10 * This program is free software; you can redistribute it and/or modify it
 11 * under the terms of the GNU General Public License version 2 as published
 12 * by the Free Software Foundation.
 13 */
 14
 15#include <linux/cgroup.h>
 16#include <linux/fs.h>
 17#include <linux/log2.h>
 18#include <linux/sched.h>
 19#include <linux/mm.h>
 20#include <linux/vmstat.h>
 21#include <linux/eventfd.h>
 22#include <linux/slab.h>
 23#include <linux/swap.h>
 24#include <linux/printk.h>
 25#include <linux/vmpressure.h>
 26
 27/*
 28 * The window size (vmpressure_win) is the number of scanned pages before
 29 * we try to analyze scanned/reclaimed ratio. So the window is used as a
 30 * rate-limit tunable for the "low" level notification, and also for
 31 * averaging the ratio for medium/critical levels. Using small window
 32 * sizes can cause lot of false positives, but too big window size will
 33 * delay the notifications.
 34 *
 35 * As the vmscan reclaimer logic works with chunks which are multiple of
 36 * SWAP_CLUSTER_MAX, it makes sense to use it for the window size as well.
 37 *
 38 * TODO: Make the window size depend on machine size, as we do for vmstat
 39 * thresholds. Currently we set it to 512 pages (2MB for 4KB pages).
 40 */
 41static const unsigned long vmpressure_win = SWAP_CLUSTER_MAX * 16;
 42
 43/*
 44 * These thresholds are used when we account memory pressure through
 45 * scanned/reclaimed ratio. The current values were chosen empirically. In
 46 * essence, they are percents: the higher the value, the more number
 47 * unsuccessful reclaims there were.
 48 */
 49static const unsigned int vmpressure_level_med = 60;
 50static const unsigned int vmpressure_level_critical = 95;
 51
 52/*
 53 * When there are too little pages left to scan, vmpressure() may miss the
 54 * critical pressure as number of pages will be less than "window size".
 55 * However, in that case the vmscan priority will raise fast as the
 56 * reclaimer will try to scan LRUs more deeply.
 57 *
 58 * The vmscan logic considers these special priorities:
 59 *
 60 * prio == DEF_PRIORITY (12): reclaimer starts with that value
 61 * prio <= DEF_PRIORITY - 2 : kswapd becomes somewhat overwhelmed
 62 * prio == 0                : close to OOM, kernel scans every page in an lru
 63 *
 64 * Any value in this range is acceptable for this tunable (i.e. from 12 to
 65 * 0). Current value for the vmpressure_level_critical_prio is chosen
 66 * empirically, but the number, in essence, means that we consider
 67 * critical level when scanning depth is ~10% of the lru size (vmscan
 68 * scans 'lru_size >> prio' pages, so it is actually 12.5%, or one
 69 * eights).
 70 */
 71static const unsigned int vmpressure_level_critical_prio = ilog2(100 / 10);
 72
 73static struct vmpressure *work_to_vmpressure(struct work_struct *work)
 74{
 75	return container_of(work, struct vmpressure, work);
 76}
 77
 78static struct vmpressure *vmpressure_parent(struct vmpressure *vmpr)
 79{
 80	struct cgroup_subsys_state *css = vmpressure_to_css(vmpr);
 81	struct mem_cgroup *memcg = mem_cgroup_from_css(css);
 82
 83	memcg = parent_mem_cgroup(memcg);
 84	if (!memcg)
 85		return NULL;
 86	return memcg_to_vmpressure(memcg);
 87}
 88
 89enum vmpressure_levels {
 90	VMPRESSURE_LOW = 0,
 91	VMPRESSURE_MEDIUM,
 92	VMPRESSURE_CRITICAL,
 93	VMPRESSURE_NUM_LEVELS,
 94};
 95
 
 
 
 
 
 
 
 96static const char * const vmpressure_str_levels[] = {
 97	[VMPRESSURE_LOW] = "low",
 98	[VMPRESSURE_MEDIUM] = "medium",
 99	[VMPRESSURE_CRITICAL] = "critical",
100};
101
 
 
 
 
 
 
102static enum vmpressure_levels vmpressure_level(unsigned long pressure)
103{
104	if (pressure >= vmpressure_level_critical)
105		return VMPRESSURE_CRITICAL;
106	else if (pressure >= vmpressure_level_med)
107		return VMPRESSURE_MEDIUM;
108	return VMPRESSURE_LOW;
109}
110
111static enum vmpressure_levels vmpressure_calc_level(unsigned long scanned,
112						    unsigned long reclaimed)
113{
114	unsigned long scale = scanned + reclaimed;
115	unsigned long pressure = 0;
116
117	/*
118	 * reclaimed can be greater than scanned in cases
119	 * like THP, where the scanned is 1 and reclaimed
120	 * could be 512
121	 */
122	if (reclaimed >= scanned)
123		goto out;
124	/*
125	 * We calculate the ratio (in percents) of how many pages were
126	 * scanned vs. reclaimed in a given time frame (window). Note that
127	 * time is in VM reclaimer's "ticks", i.e. number of pages
128	 * scanned. This makes it possible to set desired reaction time
129	 * and serves as a ratelimit.
130	 */
131	pressure = scale - (reclaimed * scale / scanned);
132	pressure = pressure * 100 / scale;
133
134out:
135	pr_debug("%s: %3lu  (s: %lu  r: %lu)\n", __func__, pressure,
136		 scanned, reclaimed);
137
138	return vmpressure_level(pressure);
139}
140
141struct vmpressure_event {
142	struct eventfd_ctx *efd;
143	enum vmpressure_levels level;
 
144	struct list_head node;
145};
146
147static bool vmpressure_event(struct vmpressure *vmpr,
148			     enum vmpressure_levels level)
 
149{
150	struct vmpressure_event *ev;
151	bool signalled = false;
152
153	mutex_lock(&vmpr->events_lock);
154
155	list_for_each_entry(ev, &vmpr->events, node) {
156		if (level >= ev->level) {
157			eventfd_signal(ev->efd, 1);
158			signalled = true;
159		}
 
 
 
 
160	}
161
162	mutex_unlock(&vmpr->events_lock);
163
164	return signalled;
165}
166
167static void vmpressure_work_fn(struct work_struct *work)
168{
169	struct vmpressure *vmpr = work_to_vmpressure(work);
170	unsigned long scanned;
171	unsigned long reclaimed;
172	enum vmpressure_levels level;
 
 
173
174	spin_lock(&vmpr->sr_lock);
175	/*
176	 * Several contexts might be calling vmpressure(), so it is
177	 * possible that the work was rescheduled again before the old
178	 * work context cleared the counters. In that case we will run
179	 * just after the old work returns, but then scanned might be zero
180	 * here. No need for any locks here since we don't care if
181	 * vmpr->reclaimed is in sync.
182	 */
183	scanned = vmpr->tree_scanned;
184	if (!scanned) {
185		spin_unlock(&vmpr->sr_lock);
186		return;
187	}
188
189	reclaimed = vmpr->tree_reclaimed;
190	vmpr->tree_scanned = 0;
191	vmpr->tree_reclaimed = 0;
192	spin_unlock(&vmpr->sr_lock);
193
194	level = vmpressure_calc_level(scanned, reclaimed);
195
196	do {
197		if (vmpressure_event(vmpr, level))
198			break;
199		/*
200		 * If not handled, propagate the event upward into the
201		 * hierarchy.
202		 */
203	} while ((vmpr = vmpressure_parent(vmpr)));
204}
205
206/**
207 * vmpressure() - Account memory pressure through scanned/reclaimed ratio
208 * @gfp:	reclaimer's gfp mask
209 * @memcg:	cgroup memory controller handle
210 * @tree:	legacy subtree mode
211 * @scanned:	number of pages scanned
212 * @reclaimed:	number of pages reclaimed
213 *
214 * This function should be called from the vmscan reclaim path to account
215 * "instantaneous" memory pressure (scanned/reclaimed ratio). The raw
216 * pressure index is then further refined and averaged over time.
217 *
218 * If @tree is set, vmpressure is in traditional userspace reporting
219 * mode: @memcg is considered the pressure root and userspace is
220 * notified of the entire subtree's reclaim efficiency.
221 *
222 * If @tree is not set, reclaim efficiency is recorded for @memcg, and
223 * only in-kernel users are notified.
224 *
225 * This function does not return any value.
226 */
227void vmpressure(gfp_t gfp, struct mem_cgroup *memcg, bool tree,
228		unsigned long scanned, unsigned long reclaimed)
229{
230	struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
 
 
 
 
 
231
232	/*
233	 * Here we only want to account pressure that userland is able to
234	 * help us with. For example, suppose that DMA zone is under
235	 * pressure; if we notify userland about that kind of pressure,
236	 * then it will be mostly a waste as it will trigger unnecessary
237	 * freeing of memory by userland (since userland is more likely to
238	 * have HIGHMEM/MOVABLE pages instead of the DMA fallback). That
239	 * is why we include only movable, highmem and FS/IO pages.
240	 * Indirect reclaim (kswapd) sets sc->gfp_mask to GFP_KERNEL, so
241	 * we account it too.
242	 */
243	if (!(gfp & (__GFP_HIGHMEM | __GFP_MOVABLE | __GFP_IO | __GFP_FS)))
244		return;
245
246	/*
247	 * If we got here with no pages scanned, then that is an indicator
248	 * that reclaimer was unable to find any shrinkable LRUs at the
249	 * current scanning depth. But it does not mean that we should
250	 * report the critical pressure, yet. If the scanning priority
251	 * (scanning depth) goes too high (deep), we will be notified
252	 * through vmpressure_prio(). But so far, keep calm.
253	 */
254	if (!scanned)
255		return;
256
257	if (tree) {
258		spin_lock(&vmpr->sr_lock);
259		scanned = vmpr->tree_scanned += scanned;
260		vmpr->tree_reclaimed += reclaimed;
261		spin_unlock(&vmpr->sr_lock);
262
263		if (scanned < vmpressure_win)
264			return;
265		schedule_work(&vmpr->work);
266	} else {
267		enum vmpressure_levels level;
268
269		/* For now, no users for root-level efficiency */
270		if (!memcg || memcg == root_mem_cgroup)
271			return;
272
273		spin_lock(&vmpr->sr_lock);
274		scanned = vmpr->scanned += scanned;
275		reclaimed = vmpr->reclaimed += reclaimed;
276		if (scanned < vmpressure_win) {
277			spin_unlock(&vmpr->sr_lock);
278			return;
279		}
280		vmpr->scanned = vmpr->reclaimed = 0;
281		spin_unlock(&vmpr->sr_lock);
282
283		level = vmpressure_calc_level(scanned, reclaimed);
284
285		if (level > VMPRESSURE_LOW) {
286			/*
287			 * Let the socket buffer allocator know that
288			 * we are having trouble reclaiming LRU pages.
289			 *
290			 * For hysteresis keep the pressure state
291			 * asserted for a second in which subsequent
292			 * pressure events can occur.
293			 */
294			memcg->socket_pressure = jiffies + HZ;
295		}
296	}
297}
298
299/**
300 * vmpressure_prio() - Account memory pressure through reclaimer priority level
301 * @gfp:	reclaimer's gfp mask
302 * @memcg:	cgroup memory controller handle
303 * @prio:	reclaimer's priority
304 *
305 * This function should be called from the reclaim path every time when
306 * the vmscan's reclaiming priority (scanning depth) changes.
307 *
308 * This function does not return any value.
309 */
310void vmpressure_prio(gfp_t gfp, struct mem_cgroup *memcg, int prio)
311{
312	/*
313	 * We only use prio for accounting critical level. For more info
314	 * see comment for vmpressure_level_critical_prio variable above.
315	 */
316	if (prio > vmpressure_level_critical_prio)
317		return;
318
319	/*
320	 * OK, the prio is below the threshold, updating vmpressure
321	 * information before shrinker dives into long shrinking of long
322	 * range vmscan. Passing scanned = vmpressure_win, reclaimed = 0
323	 * to the vmpressure() basically means that we signal 'critical'
324	 * level.
325	 */
326	vmpressure(gfp, memcg, true, vmpressure_win, 0);
327}
328
 
 
329/**
330 * vmpressure_register_event() - Bind vmpressure notifications to an eventfd
331 * @memcg:	memcg that is interested in vmpressure notifications
332 * @eventfd:	eventfd context to link notifications with
333 * @args:	event arguments (used to set up a pressure level threshold)
334 *
335 * This function associates eventfd context with the vmpressure
336 * infrastructure, so that the notifications will be delivered to the
337 * @eventfd. The @args parameter is a string that denotes pressure level
338 * threshold (one of vmpressure_str_levels, i.e. "low", "medium", or
339 * "critical").
 
340 *
341 * To be used as memcg event method.
 
 
 
342 */
343int vmpressure_register_event(struct mem_cgroup *memcg,
344			      struct eventfd_ctx *eventfd, const char *args)
345{
346	struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
347	struct vmpressure_event *ev;
348	int level;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
349
350	for (level = 0; level < VMPRESSURE_NUM_LEVELS; level++) {
351		if (!strcmp(vmpressure_str_levels[level], args))
352			break;
 
 
 
 
353	}
354
355	if (level >= VMPRESSURE_NUM_LEVELS)
356		return -EINVAL;
357
358	ev = kzalloc(sizeof(*ev), GFP_KERNEL);
359	if (!ev)
360		return -ENOMEM;
 
 
361
362	ev->efd = eventfd;
363	ev->level = level;
 
364
365	mutex_lock(&vmpr->events_lock);
366	list_add(&ev->node, &vmpr->events);
367	mutex_unlock(&vmpr->events_lock);
368
369	return 0;
 
 
370}
371
372/**
373 * vmpressure_unregister_event() - Unbind eventfd from vmpressure
374 * @memcg:	memcg handle
375 * @eventfd:	eventfd context that was used to link vmpressure with the @cg
376 *
377 * This function does internal manipulations to detach the @eventfd from
378 * the vmpressure notifications, and then frees internal resources
379 * associated with the @eventfd (but the @eventfd itself is not freed).
380 *
381 * To be used as memcg event method.
382 */
383void vmpressure_unregister_event(struct mem_cgroup *memcg,
384				 struct eventfd_ctx *eventfd)
385{
386	struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
387	struct vmpressure_event *ev;
388
389	mutex_lock(&vmpr->events_lock);
390	list_for_each_entry(ev, &vmpr->events, node) {
391		if (ev->efd != eventfd)
392			continue;
393		list_del(&ev->node);
394		kfree(ev);
395		break;
396	}
397	mutex_unlock(&vmpr->events_lock);
398}
399
400/**
401 * vmpressure_init() - Initialize vmpressure control structure
402 * @vmpr:	Structure to be initialized
403 *
404 * This function should be called on every allocated vmpressure structure
405 * before any usage.
406 */
407void vmpressure_init(struct vmpressure *vmpr)
408{
409	spin_lock_init(&vmpr->sr_lock);
410	mutex_init(&vmpr->events_lock);
411	INIT_LIST_HEAD(&vmpr->events);
412	INIT_WORK(&vmpr->work, vmpressure_work_fn);
413}
414
415/**
416 * vmpressure_cleanup() - shuts down vmpressure control structure
417 * @vmpr:	Structure to be cleaned up
418 *
419 * This function should be called before the structure in which it is
420 * embedded is cleaned up.
421 */
422void vmpressure_cleanup(struct vmpressure *vmpr)
423{
424	/*
425	 * Make sure there is no pending work before eventfd infrastructure
426	 * goes away.
427	 */
428	flush_work(&vmpr->work);
429}