blob: e0f62837c3f4873ea2d54c14845c30cb888a386b [file] [log] [blame]
Anton Vorontsov70ddf632013-04-29 15:08:31 -07001/*
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/swap.h>
23#include <linux/printk.h>
24#include <linux/vmpressure.h>
25
26/*
27 * The window size (vmpressure_win) is the number of scanned pages before
28 * we try to analyze scanned/reclaimed ratio. So the window is used as a
29 * rate-limit tunable for the "low" level notification, and also for
30 * averaging the ratio for medium/critical levels. Using small window
31 * sizes can cause lot of false positives, but too big window size will
32 * delay the notifications.
33 *
34 * As the vmscan reclaimer logic works with chunks which are multiple of
35 * SWAP_CLUSTER_MAX, it makes sense to use it for the window size as well.
36 *
37 * TODO: Make the window size depend on machine size, as we do for vmstat
38 * thresholds. Currently we set it to 512 pages (2MB for 4KB pages).
39 */
40static const unsigned long vmpressure_win = SWAP_CLUSTER_MAX * 16;
41
42/*
43 * These thresholds are used when we account memory pressure through
44 * scanned/reclaimed ratio. The current values were chosen empirically. In
45 * essence, they are percents: the higher the value, the more number
46 * unsuccessful reclaims there were.
47 */
48static const unsigned int vmpressure_level_med = 60;
49static const unsigned int vmpressure_level_critical = 95;
50
51/*
52 * When there are too little pages left to scan, vmpressure() may miss the
53 * critical pressure as number of pages will be less than "window size".
54 * However, in that case the vmscan priority will raise fast as the
55 * reclaimer will try to scan LRUs more deeply.
56 *
57 * The vmscan logic considers these special priorities:
58 *
59 * prio == DEF_PRIORITY (12): reclaimer starts with that value
60 * prio <= DEF_PRIORITY - 2 : kswapd becomes somewhat overwhelmed
61 * prio == 0 : close to OOM, kernel scans every page in an lru
62 *
63 * Any value in this range is acceptable for this tunable (i.e. from 12 to
64 * 0). Current value for the vmpressure_level_critical_prio is chosen
65 * empirically, but the number, in essence, means that we consider
66 * critical level when scanning depth is ~10% of the lru size (vmscan
67 * scans 'lru_size >> prio' pages, so it is actually 12.5%, or one
68 * eights).
69 */
70static const unsigned int vmpressure_level_critical_prio = ilog2(100 / 10);
71
72static struct vmpressure *work_to_vmpressure(struct work_struct *work)
73{
74 return container_of(work, struct vmpressure, work);
75}
76
Anton Vorontsov70ddf632013-04-29 15:08:31 -070077static struct vmpressure *vmpressure_parent(struct vmpressure *vmpr)
78{
Tejun Heo182446d2013-08-08 20:11:24 -040079 struct cgroup_subsys_state *css = vmpressure_to_css(vmpr);
80 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
Anton Vorontsov70ddf632013-04-29 15:08:31 -070081
82 memcg = parent_mem_cgroup(memcg);
83 if (!memcg)
84 return NULL;
85 return memcg_to_vmpressure(memcg);
86}
87
88enum vmpressure_levels {
89 VMPRESSURE_LOW = 0,
90 VMPRESSURE_MEDIUM,
91 VMPRESSURE_CRITICAL,
92 VMPRESSURE_NUM_LEVELS,
93};
94
95static const char * const vmpressure_str_levels[] = {
96 [VMPRESSURE_LOW] = "low",
97 [VMPRESSURE_MEDIUM] = "medium",
98 [VMPRESSURE_CRITICAL] = "critical",
99};
100
101static enum vmpressure_levels vmpressure_level(unsigned long pressure)
102{
103 if (pressure >= vmpressure_level_critical)
104 return VMPRESSURE_CRITICAL;
105 else if (pressure >= vmpressure_level_med)
106 return VMPRESSURE_MEDIUM;
107 return VMPRESSURE_LOW;
108}
109
110static enum vmpressure_levels vmpressure_calc_level(unsigned long scanned,
111 unsigned long reclaimed)
112{
113 unsigned long scale = scanned + reclaimed;
114 unsigned long pressure;
115
116 /*
117 * We calculate the ratio (in percents) of how many pages were
118 * scanned vs. reclaimed in a given time frame (window). Note that
119 * time is in VM reclaimer's "ticks", i.e. number of pages
120 * scanned. This makes it possible to set desired reaction time
121 * and serves as a ratelimit.
122 */
123 pressure = scale - (reclaimed * scale / scanned);
124 pressure = pressure * 100 / scale;
125
126 pr_debug("%s: %3lu (s: %lu r: %lu)\n", __func__, pressure,
127 scanned, reclaimed);
128
129 return vmpressure_level(pressure);
130}
131
132struct vmpressure_event {
133 struct eventfd_ctx *efd;
134 enum vmpressure_levels level;
135 struct list_head node;
136};
137
138static bool vmpressure_event(struct vmpressure *vmpr,
139 unsigned long scanned, unsigned long reclaimed)
140{
141 struct vmpressure_event *ev;
142 enum vmpressure_levels level;
143 bool signalled = false;
144
145 level = vmpressure_calc_level(scanned, reclaimed);
146
147 mutex_lock(&vmpr->events_lock);
148
149 list_for_each_entry(ev, &vmpr->events, node) {
150 if (level >= ev->level) {
151 eventfd_signal(ev->efd, 1);
152 signalled = true;
153 }
154 }
155
156 mutex_unlock(&vmpr->events_lock);
157
158 return signalled;
159}
160
161static void vmpressure_work_fn(struct work_struct *work)
162{
163 struct vmpressure *vmpr = work_to_vmpressure(work);
164 unsigned long scanned;
165 unsigned long reclaimed;
166
167 /*
168 * Several contexts might be calling vmpressure(), so it is
169 * possible that the work was rescheduled again before the old
170 * work context cleared the counters. In that case we will run
171 * just after the old work returns, but then scanned might be zero
172 * here. No need for any locks here since we don't care if
173 * vmpr->reclaimed is in sync.
174 */
175 if (!vmpr->scanned)
176 return;
177
Michal Hocko22f20202013-07-31 13:53:48 -0700178 spin_lock(&vmpr->sr_lock);
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700179 scanned = vmpr->scanned;
180 reclaimed = vmpr->reclaimed;
181 vmpr->scanned = 0;
182 vmpr->reclaimed = 0;
Michal Hocko22f20202013-07-31 13:53:48 -0700183 spin_unlock(&vmpr->sr_lock);
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700184
185 do {
186 if (vmpressure_event(vmpr, scanned, reclaimed))
187 break;
188 /*
189 * If not handled, propagate the event upward into the
190 * hierarchy.
191 */
192 } while ((vmpr = vmpressure_parent(vmpr)));
193}
194
195/**
196 * vmpressure() - Account memory pressure through scanned/reclaimed ratio
197 * @gfp: reclaimer's gfp mask
198 * @memcg: cgroup memory controller handle
199 * @scanned: number of pages scanned
200 * @reclaimed: number of pages reclaimed
201 *
202 * This function should be called from the vmscan reclaim path to account
203 * "instantaneous" memory pressure (scanned/reclaimed ratio). The raw
204 * pressure index is then further refined and averaged over time.
205 *
206 * This function does not return any value.
207 */
208void vmpressure(gfp_t gfp, struct mem_cgroup *memcg,
209 unsigned long scanned, unsigned long reclaimed)
210{
211 struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
212
213 /*
214 * Here we only want to account pressure that userland is able to
215 * help us with. For example, suppose that DMA zone is under
216 * pressure; if we notify userland about that kind of pressure,
217 * then it will be mostly a waste as it will trigger unnecessary
218 * freeing of memory by userland (since userland is more likely to
219 * have HIGHMEM/MOVABLE pages instead of the DMA fallback). That
220 * is why we include only movable, highmem and FS/IO pages.
221 * Indirect reclaim (kswapd) sets sc->gfp_mask to GFP_KERNEL, so
222 * we account it too.
223 */
224 if (!(gfp & (__GFP_HIGHMEM | __GFP_MOVABLE | __GFP_IO | __GFP_FS)))
225 return;
226
227 /*
228 * If we got here with no pages scanned, then that is an indicator
229 * that reclaimer was unable to find any shrinkable LRUs at the
230 * current scanning depth. But it does not mean that we should
231 * report the critical pressure, yet. If the scanning priority
232 * (scanning depth) goes too high (deep), we will be notified
233 * through vmpressure_prio(). But so far, keep calm.
234 */
235 if (!scanned)
236 return;
237
Michal Hocko22f20202013-07-31 13:53:48 -0700238 spin_lock(&vmpr->sr_lock);
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700239 vmpr->scanned += scanned;
240 vmpr->reclaimed += reclaimed;
241 scanned = vmpr->scanned;
Michal Hocko22f20202013-07-31 13:53:48 -0700242 spin_unlock(&vmpr->sr_lock);
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700243
Michal Hocko8e0ed442013-07-31 13:53:50 -0700244 if (scanned < vmpressure_win)
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700245 return;
246 schedule_work(&vmpr->work);
247}
248
249/**
250 * vmpressure_prio() - Account memory pressure through reclaimer priority level
251 * @gfp: reclaimer's gfp mask
252 * @memcg: cgroup memory controller handle
253 * @prio: reclaimer's priority
254 *
255 * This function should be called from the reclaim path every time when
256 * the vmscan's reclaiming priority (scanning depth) changes.
257 *
258 * This function does not return any value.
259 */
260void vmpressure_prio(gfp_t gfp, struct mem_cgroup *memcg, int prio)
261{
262 /*
263 * We only use prio for accounting critical level. For more info
264 * see comment for vmpressure_level_critical_prio variable above.
265 */
266 if (prio > vmpressure_level_critical_prio)
267 return;
268
269 /*
270 * OK, the prio is below the threshold, updating vmpressure
271 * information before shrinker dives into long shrinking of long
272 * range vmscan. Passing scanned = vmpressure_win, reclaimed = 0
273 * to the vmpressure() basically means that we signal 'critical'
274 * level.
275 */
276 vmpressure(gfp, memcg, vmpressure_win, 0);
277}
278
279/**
280 * vmpressure_register_event() - Bind vmpressure notifications to an eventfd
Tejun Heo81eeaf02013-08-08 20:11:26 -0400281 * @css: css that is interested in vmpressure notifications
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700282 * @cft: cgroup control files handle
283 * @eventfd: eventfd context to link notifications with
284 * @args: event arguments (used to set up a pressure level threshold)
285 *
286 * This function associates eventfd context with the vmpressure
287 * infrastructure, so that the notifications will be delivered to the
288 * @eventfd. The @args parameter is a string that denotes pressure level
289 * threshold (one of vmpressure_str_levels, i.e. "low", "medium", or
290 * "critical").
291 *
292 * This function should not be used directly, just pass it to (struct
293 * cftype).register_event, and then cgroup core will handle everything by
294 * itself.
295 */
Tejun Heo81eeaf02013-08-08 20:11:26 -0400296int vmpressure_register_event(struct cgroup_subsys_state *css,
297 struct cftype *cft, struct eventfd_ctx *eventfd,
298 const char *args)
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700299{
Tejun Heo81eeaf02013-08-08 20:11:26 -0400300 struct vmpressure *vmpr = css_to_vmpressure(css);
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700301 struct vmpressure_event *ev;
302 int level;
303
304 for (level = 0; level < VMPRESSURE_NUM_LEVELS; level++) {
305 if (!strcmp(vmpressure_str_levels[level], args))
306 break;
307 }
308
309 if (level >= VMPRESSURE_NUM_LEVELS)
310 return -EINVAL;
311
312 ev = kzalloc(sizeof(*ev), GFP_KERNEL);
313 if (!ev)
314 return -ENOMEM;
315
316 ev->efd = eventfd;
317 ev->level = level;
318
319 mutex_lock(&vmpr->events_lock);
320 list_add(&ev->node, &vmpr->events);
321 mutex_unlock(&vmpr->events_lock);
322
323 return 0;
324}
325
326/**
327 * vmpressure_unregister_event() - Unbind eventfd from vmpressure
Tejun Heo81eeaf02013-08-08 20:11:26 -0400328 * @css: css handle
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700329 * @cft: cgroup control files handle
330 * @eventfd: eventfd context that was used to link vmpressure with the @cg
331 *
332 * This function does internal manipulations to detach the @eventfd from
333 * the vmpressure notifications, and then frees internal resources
334 * associated with the @eventfd (but the @eventfd itself is not freed).
335 *
336 * This function should not be used directly, just pass it to (struct
337 * cftype).unregister_event, and then cgroup core will handle everything
338 * by itself.
339 */
Tejun Heo81eeaf02013-08-08 20:11:26 -0400340void vmpressure_unregister_event(struct cgroup_subsys_state *css,
341 struct cftype *cft,
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700342 struct eventfd_ctx *eventfd)
343{
Tejun Heo81eeaf02013-08-08 20:11:26 -0400344 struct vmpressure *vmpr = css_to_vmpressure(css);
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700345 struct vmpressure_event *ev;
346
347 mutex_lock(&vmpr->events_lock);
348 list_for_each_entry(ev, &vmpr->events, node) {
349 if (ev->efd != eventfd)
350 continue;
351 list_del(&ev->node);
352 kfree(ev);
353 break;
354 }
355 mutex_unlock(&vmpr->events_lock);
356}
357
358/**
359 * vmpressure_init() - Initialize vmpressure control structure
360 * @vmpr: Structure to be initialized
361 *
362 * This function should be called on every allocated vmpressure structure
363 * before any usage.
364 */
365void vmpressure_init(struct vmpressure *vmpr)
366{
Michal Hocko22f20202013-07-31 13:53:48 -0700367 spin_lock_init(&vmpr->sr_lock);
Anton Vorontsov70ddf632013-04-29 15:08:31 -0700368 mutex_init(&vmpr->events_lock);
369 INIT_LIST_HEAD(&vmpr->events);
370 INIT_WORK(&vmpr->work, vmpressure_work_fn);
371}
Michal Hocko33cb8762013-07-31 13:53:51 -0700372
373/**
374 * vmpressure_cleanup() - shuts down vmpressure control structure
375 * @vmpr: Structure to be cleaned up
376 *
377 * This function should be called before the structure in which it is
378 * embedded is cleaned up.
379 */
380void vmpressure_cleanup(struct vmpressure *vmpr)
381{
382 /*
383 * Make sure there is no pending work before eventfd infrastructure
384 * goes away.
385 */
386 flush_work(&vmpr->work);
387}