xref: /openbmc/linux/mm/vmpressure.c (revision be709d48)
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  */
41 static 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  */
49 static const unsigned int vmpressure_level_med = 60;
50 static 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  */
71 static const unsigned int vmpressure_level_critical_prio = ilog2(100 / 10);
72 
73 static struct vmpressure *work_to_vmpressure(struct work_struct *work)
74 {
75 	return container_of(work, struct vmpressure, work);
76 }
77 
78 static 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 
89 enum vmpressure_levels {
90 	VMPRESSURE_LOW = 0,
91 	VMPRESSURE_MEDIUM,
92 	VMPRESSURE_CRITICAL,
93 	VMPRESSURE_NUM_LEVELS,
94 };
95 
96 enum vmpressure_modes {
97 	VMPRESSURE_NO_PASSTHROUGH = 0,
98 	VMPRESSURE_HIERARCHY,
99 	VMPRESSURE_LOCAL,
100 	VMPRESSURE_NUM_MODES,
101 };
102 
103 static const char * const vmpressure_str_levels[] = {
104 	[VMPRESSURE_LOW] = "low",
105 	[VMPRESSURE_MEDIUM] = "medium",
106 	[VMPRESSURE_CRITICAL] = "critical",
107 };
108 
109 static const char * const vmpressure_str_modes[] = {
110 	[VMPRESSURE_NO_PASSTHROUGH] = "default",
111 	[VMPRESSURE_HIERARCHY] = "hierarchy",
112 	[VMPRESSURE_LOCAL] = "local",
113 };
114 
115 static enum vmpressure_levels vmpressure_level(unsigned long pressure)
116 {
117 	if (pressure >= vmpressure_level_critical)
118 		return VMPRESSURE_CRITICAL;
119 	else if (pressure >= vmpressure_level_med)
120 		return VMPRESSURE_MEDIUM;
121 	return VMPRESSURE_LOW;
122 }
123 
124 static enum vmpressure_levels vmpressure_calc_level(unsigned long scanned,
125 						    unsigned long reclaimed)
126 {
127 	unsigned long scale = scanned + reclaimed;
128 	unsigned long pressure = 0;
129 
130 	/*
131 	 * reclaimed can be greater than scanned for things such as reclaimed
132 	 * slab pages. shrink_node() just adds reclaimed pages without a
133 	 * related increment to scanned pages.
134 	 */
135 	if (reclaimed >= scanned)
136 		goto out;
137 	/*
138 	 * We calculate the ratio (in percents) of how many pages were
139 	 * scanned vs. reclaimed in a given time frame (window). Note that
140 	 * time is in VM reclaimer's "ticks", i.e. number of pages
141 	 * scanned. This makes it possible to set desired reaction time
142 	 * and serves as a ratelimit.
143 	 */
144 	pressure = scale - (reclaimed * scale / scanned);
145 	pressure = pressure * 100 / scale;
146 
147 out:
148 	pr_debug("%s: %3lu  (s: %lu  r: %lu)\n", __func__, pressure,
149 		 scanned, reclaimed);
150 
151 	return vmpressure_level(pressure);
152 }
153 
154 struct vmpressure_event {
155 	struct eventfd_ctx *efd;
156 	enum vmpressure_levels level;
157 	enum vmpressure_modes mode;
158 	struct list_head node;
159 };
160 
161 static bool vmpressure_event(struct vmpressure *vmpr,
162 			     const enum vmpressure_levels level,
163 			     bool ancestor, bool signalled)
164 {
165 	struct vmpressure_event *ev;
166 	bool ret = false;
167 
168 	mutex_lock(&vmpr->events_lock);
169 	list_for_each_entry(ev, &vmpr->events, node) {
170 		if (ancestor && ev->mode == VMPRESSURE_LOCAL)
171 			continue;
172 		if (signalled && ev->mode == VMPRESSURE_NO_PASSTHROUGH)
173 			continue;
174 		if (level < ev->level)
175 			continue;
176 		eventfd_signal(ev->efd, 1);
177 		ret = true;
178 	}
179 	mutex_unlock(&vmpr->events_lock);
180 
181 	return ret;
182 }
183 
184 static void vmpressure_work_fn(struct work_struct *work)
185 {
186 	struct vmpressure *vmpr = work_to_vmpressure(work);
187 	unsigned long scanned;
188 	unsigned long reclaimed;
189 	enum vmpressure_levels level;
190 	bool ancestor = false;
191 	bool signalled = false;
192 
193 	spin_lock(&vmpr->sr_lock);
194 	/*
195 	 * Several contexts might be calling vmpressure(), so it is
196 	 * possible that the work was rescheduled again before the old
197 	 * work context cleared the counters. In that case we will run
198 	 * just after the old work returns, but then scanned might be zero
199 	 * here. No need for any locks here since we don't care if
200 	 * vmpr->reclaimed is in sync.
201 	 */
202 	scanned = vmpr->tree_scanned;
203 	if (!scanned) {
204 		spin_unlock(&vmpr->sr_lock);
205 		return;
206 	}
207 
208 	reclaimed = vmpr->tree_reclaimed;
209 	vmpr->tree_scanned = 0;
210 	vmpr->tree_reclaimed = 0;
211 	spin_unlock(&vmpr->sr_lock);
212 
213 	level = vmpressure_calc_level(scanned, reclaimed);
214 
215 	do {
216 		if (vmpressure_event(vmpr, level, ancestor, signalled))
217 			signalled = true;
218 		ancestor = true;
219 	} while ((vmpr = vmpressure_parent(vmpr)));
220 }
221 
222 /**
223  * vmpressure() - Account memory pressure through scanned/reclaimed ratio
224  * @gfp:	reclaimer's gfp mask
225  * @memcg:	cgroup memory controller handle
226  * @tree:	legacy subtree mode
227  * @scanned:	number of pages scanned
228  * @reclaimed:	number of pages reclaimed
229  *
230  * This function should be called from the vmscan reclaim path to account
231  * "instantaneous" memory pressure (scanned/reclaimed ratio). The raw
232  * pressure index is then further refined and averaged over time.
233  *
234  * If @tree is set, vmpressure is in traditional userspace reporting
235  * mode: @memcg is considered the pressure root and userspace is
236  * notified of the entire subtree's reclaim efficiency.
237  *
238  * If @tree is not set, reclaim efficiency is recorded for @memcg, and
239  * only in-kernel users are notified.
240  *
241  * This function does not return any value.
242  */
243 void vmpressure(gfp_t gfp, struct mem_cgroup *memcg, bool tree,
244 		unsigned long scanned, unsigned long reclaimed)
245 {
246 	struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
247 
248 	/*
249 	 * Here we only want to account pressure that userland is able to
250 	 * help us with. For example, suppose that DMA zone is under
251 	 * pressure; if we notify userland about that kind of pressure,
252 	 * then it will be mostly a waste as it will trigger unnecessary
253 	 * freeing of memory by userland (since userland is more likely to
254 	 * have HIGHMEM/MOVABLE pages instead of the DMA fallback). That
255 	 * is why we include only movable, highmem and FS/IO pages.
256 	 * Indirect reclaim (kswapd) sets sc->gfp_mask to GFP_KERNEL, so
257 	 * we account it too.
258 	 */
259 	if (!(gfp & (__GFP_HIGHMEM | __GFP_MOVABLE | __GFP_IO | __GFP_FS)))
260 		return;
261 
262 	/*
263 	 * If we got here with no pages scanned, then that is an indicator
264 	 * that reclaimer was unable to find any shrinkable LRUs at the
265 	 * current scanning depth. But it does not mean that we should
266 	 * report the critical pressure, yet. If the scanning priority
267 	 * (scanning depth) goes too high (deep), we will be notified
268 	 * through vmpressure_prio(). But so far, keep calm.
269 	 */
270 	if (!scanned)
271 		return;
272 
273 	if (tree) {
274 		spin_lock(&vmpr->sr_lock);
275 		scanned = vmpr->tree_scanned += scanned;
276 		vmpr->tree_reclaimed += reclaimed;
277 		spin_unlock(&vmpr->sr_lock);
278 
279 		if (scanned < vmpressure_win)
280 			return;
281 		schedule_work(&vmpr->work);
282 	} else {
283 		enum vmpressure_levels level;
284 
285 		/* For now, no users for root-level efficiency */
286 		if (!memcg || memcg == root_mem_cgroup)
287 			return;
288 
289 		spin_lock(&vmpr->sr_lock);
290 		scanned = vmpr->scanned += scanned;
291 		reclaimed = vmpr->reclaimed += reclaimed;
292 		if (scanned < vmpressure_win) {
293 			spin_unlock(&vmpr->sr_lock);
294 			return;
295 		}
296 		vmpr->scanned = vmpr->reclaimed = 0;
297 		spin_unlock(&vmpr->sr_lock);
298 
299 		level = vmpressure_calc_level(scanned, reclaimed);
300 
301 		if (level > VMPRESSURE_LOW) {
302 			/*
303 			 * Let the socket buffer allocator know that
304 			 * we are having trouble reclaiming LRU pages.
305 			 *
306 			 * For hysteresis keep the pressure state
307 			 * asserted for a second in which subsequent
308 			 * pressure events can occur.
309 			 */
310 			memcg->socket_pressure = jiffies + HZ;
311 		}
312 	}
313 }
314 
315 /**
316  * vmpressure_prio() - Account memory pressure through reclaimer priority level
317  * @gfp:	reclaimer's gfp mask
318  * @memcg:	cgroup memory controller handle
319  * @prio:	reclaimer's priority
320  *
321  * This function should be called from the reclaim path every time when
322  * the vmscan's reclaiming priority (scanning depth) changes.
323  *
324  * This function does not return any value.
325  */
326 void vmpressure_prio(gfp_t gfp, struct mem_cgroup *memcg, int prio)
327 {
328 	/*
329 	 * We only use prio for accounting critical level. For more info
330 	 * see comment for vmpressure_level_critical_prio variable above.
331 	 */
332 	if (prio > vmpressure_level_critical_prio)
333 		return;
334 
335 	/*
336 	 * OK, the prio is below the threshold, updating vmpressure
337 	 * information before shrinker dives into long shrinking of long
338 	 * range vmscan. Passing scanned = vmpressure_win, reclaimed = 0
339 	 * to the vmpressure() basically means that we signal 'critical'
340 	 * level.
341 	 */
342 	vmpressure(gfp, memcg, true, vmpressure_win, 0);
343 }
344 
345 #define MAX_VMPRESSURE_ARGS_LEN	(strlen("critical") + strlen("hierarchy") + 2)
346 
347 /**
348  * vmpressure_register_event() - Bind vmpressure notifications to an eventfd
349  * @memcg:	memcg that is interested in vmpressure notifications
350  * @eventfd:	eventfd context to link notifications with
351  * @args:	event arguments (pressure level threshold, optional mode)
352  *
353  * This function associates eventfd context with the vmpressure
354  * infrastructure, so that the notifications will be delivered to the
355  * @eventfd. The @args parameter is a comma-delimited string that denotes a
356  * pressure level threshold (one of vmpressure_str_levels, i.e. "low", "medium",
357  * or "critical") and an optional mode (one of vmpressure_str_modes, i.e.
358  * "hierarchy" or "local").
359  *
360  * To be used as memcg event method.
361  */
362 int vmpressure_register_event(struct mem_cgroup *memcg,
363 			      struct eventfd_ctx *eventfd, const char *args)
364 {
365 	struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
366 	struct vmpressure_event *ev;
367 	enum vmpressure_modes mode = VMPRESSURE_NO_PASSTHROUGH;
368 	enum vmpressure_levels level = -1;
369 	char *spec, *spec_orig;
370 	char *token;
371 	int ret = 0;
372 
373 	spec_orig = spec = kstrndup(args, MAX_VMPRESSURE_ARGS_LEN, GFP_KERNEL);
374 	if (!spec) {
375 		ret = -ENOMEM;
376 		goto out;
377 	}
378 
379 	/* Find required level */
380 	token = strsep(&spec, ",");
381 	level = match_string(vmpressure_str_levels, VMPRESSURE_NUM_LEVELS, token);
382 	if (level < 0) {
383 		ret = level;
384 		goto out;
385 	}
386 
387 	/* Find optional mode */
388 	token = strsep(&spec, ",");
389 	if (token) {
390 		mode = match_string(vmpressure_str_modes, VMPRESSURE_NUM_MODES, token);
391 		if (mode < 0) {
392 			ret = mode;
393 			goto out;
394 		}
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 out:
411 	kfree(spec_orig);
412 	return ret;
413 }
414 
415 /**
416  * vmpressure_unregister_event() - Unbind eventfd from vmpressure
417  * @memcg:	memcg handle
418  * @eventfd:	eventfd context that was used to link vmpressure with the @cg
419  *
420  * This function does internal manipulations to detach the @eventfd from
421  * the vmpressure notifications, and then frees internal resources
422  * associated with the @eventfd (but the @eventfd itself is not freed).
423  *
424  * To be used as memcg event method.
425  */
426 void vmpressure_unregister_event(struct mem_cgroup *memcg,
427 				 struct eventfd_ctx *eventfd)
428 {
429 	struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
430 	struct vmpressure_event *ev;
431 
432 	mutex_lock(&vmpr->events_lock);
433 	list_for_each_entry(ev, &vmpr->events, node) {
434 		if (ev->efd != eventfd)
435 			continue;
436 		list_del(&ev->node);
437 		kfree(ev);
438 		break;
439 	}
440 	mutex_unlock(&vmpr->events_lock);
441 }
442 
443 /**
444  * vmpressure_init() - Initialize vmpressure control structure
445  * @vmpr:	Structure to be initialized
446  *
447  * This function should be called on every allocated vmpressure structure
448  * before any usage.
449  */
450 void vmpressure_init(struct vmpressure *vmpr)
451 {
452 	spin_lock_init(&vmpr->sr_lock);
453 	mutex_init(&vmpr->events_lock);
454 	INIT_LIST_HEAD(&vmpr->events);
455 	INIT_WORK(&vmpr->work, vmpressure_work_fn);
456 }
457 
458 /**
459  * vmpressure_cleanup() - shuts down vmpressure control structure
460  * @vmpr:	Structure to be cleaned up
461  *
462  * This function should be called before the structure in which it is
463  * embedded is cleaned up.
464  */
465 void vmpressure_cleanup(struct vmpressure *vmpr)
466 {
467 	/*
468 	 * Make sure there is no pending work before eventfd infrastructure
469 	 * goes away.
470 	 */
471 	flush_work(&vmpr->work);
472 }
473