mm/vmpressure.c

   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/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  */
  40 static 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  */
  48 static const unsigned int vmpressure_level_med = 60;
  49 static 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  */
  70 static const unsigned int vmpressure_level_critical_prio = ilog2(100 / 10);
  71
  72 static struct vmpressure *work_to_vmpressure(struct work_struct *work)
  73 {
  74         return container_of(work, struct vmpressure, work);
  75 }
  76
  77 static struct vmpressure *cg_to_vmpressure(struct cgroup *cg)
  78 {
  79         return css_to_vmpressure(cgroup_subsys_state(cg, mem_cgroup_subsys_id));
  80 }
  81
  82 static struct vmpressure *vmpressure_parent(struct vmpressure *vmpr)
  83 {
  84         struct cgroup *cg = vmpressure_to_css(vmpr)->cgroup;
  85         struct mem_cgroup *memcg = mem_cgroup_from_cont(cg);
  86
  87         memcg = parent_mem_cgroup(memcg);
  88         if (!memcg)
  89                 return NULL;
  90         return memcg_to_vmpressure(memcg);
  91 }
  92
  93 enum vmpressure_levels {
  94         VMPRESSURE_LOW = 0,
  95         VMPRESSURE_MEDIUM,
  96         VMPRESSURE_CRITICAL,
  97         VMPRESSURE_NUM_LEVELS,
  98 };
  99
 100 static const char * const vmpressure_str_levels[] = {
 101         [VMPRESSURE_LOW] = "low",
 102         [VMPRESSURE_MEDIUM] = "medium",
 103         [VMPRESSURE_CRITICAL] = "critical",
 104 };
 105
 106 static enum vmpressure_levels vmpressure_level(unsigned long pressure)
 107 {
 108         if (pressure >= vmpressure_level_critical)
 109                 return VMPRESSURE_CRITICAL;
 110         else if (pressure >= vmpressure_level_med)
 111                 return VMPRESSURE_MEDIUM;
 112         return VMPRESSURE_LOW;
 113 }
 114
 115 static enum vmpressure_levels vmpressure_calc_level(unsigned long scanned,
 116                                                     unsigned long reclaimed)
 117 {
 118         unsigned long scale = scanned + reclaimed;
 119         unsigned long pressure;
 120
 121         /*
 122          * We calculate the ratio (in percents) of how many pages were
 123          * scanned vs. reclaimed in a given time frame (window). Note that
 124          * time is in VM reclaimer's "ticks", i.e. number of pages
 125          * scanned. This makes it possible to set desired reaction time
 126          * and serves as a ratelimit.
 127          */
 128         pressure = scale - (reclaimed * scale / scanned);
 129         pressure = pressure * 100 / scale;
 130
 131         pr_debug("%s: %3lu  (s: %lu  r: %lu)\n", __func__, pressure,
 132                  scanned, reclaimed);
 133
 134         return vmpressure_level(pressure);
 135 }
 136
 137 struct vmpressure_event {
 138         struct eventfd_ctx *efd;
 139         enum vmpressure_levels level;
 140         struct list_head node;
 141 };
 142
 143 static bool vmpressure_event(struct vmpressure *vmpr,
 144                              unsigned long scanned, unsigned long reclaimed)
 145 {
 146         struct vmpressure_event *ev;
 147         enum vmpressure_levels level;
 148         bool signalled = false;
 149
 150         level = vmpressure_calc_level(scanned, reclaimed);
 151
 152         mutex_lock(&vmpr->events_lock);
 153
 154         list_for_each_entry(ev, &vmpr->events, node) {
 155                 if (level >= ev->level) {
 156                         eventfd_signal(ev->efd, 1);
 157                         signalled = true;
 158                 }
 159         }
 160
 161         mutex_unlock(&vmpr->events_lock);
 162
 163         return signalled;
 164 }
 165
 166 static void vmpressure_work_fn(struct work_struct *work)
 167 {
 168         struct vmpressure *vmpr = work_to_vmpressure(work);
 169         unsigned long scanned;
 170         unsigned long reclaimed;
 171
 172         /*
 173          * Several contexts might be calling vmpressure(), so it is
 174          * possible that the work was rescheduled again before the old
 175          * work context cleared the counters. In that case we will run
 176          * just after the old work returns, but then scanned might be zero
 177          * here. No need for any locks here since we don't care if
 178          * vmpr->reclaimed is in sync.
 179          */
 180         if (!vmpr->scanned)
 181                 return;
 182
 183         spin_lock(&vmpr->sr_lock);
 184         scanned = vmpr->scanned;
 185         reclaimed = vmpr->reclaimed;
 186         vmpr->scanned = 0;
 187         vmpr->reclaimed = 0;
 188         spin_unlock(&vmpr->sr_lock);
 189
 190         do {
 191                 if (vmpressure_event(vmpr, scanned, reclaimed))
 192                         break;
 193                 /*
 194                  * If not handled, propagate the event upward into the
 195                  * hierarchy.
 196                  */
 197         } while ((vmpr = vmpressure_parent(vmpr)));
 198 }
 199
 200 /**
 201  * vmpressure() - Account memory pressure through scanned/reclaimed ratio
 202  * @gfp:        reclaimer's gfp mask
 203  * @memcg:      cgroup memory controller handle
 204  * @scanned:    number of pages scanned
 205  * @reclaimed:  number of pages reclaimed
 206  *
 207  * This function should be called from the vmscan reclaim path to account
 208  * "instantaneous" memory pressure (scanned/reclaimed ratio). The raw
 209  * pressure index is then further refined and averaged over time.
 210  *
 211  * This function does not return any value.
 212  */
 213 void vmpressure(gfp_t gfp, struct mem_cgroup *memcg,
 214                 unsigned long scanned, unsigned long reclaimed)
 215 {
 216         struct vmpressure *vmpr = memcg_to_vmpressure(memcg);
 217
 218         /*
 219          * Here we only want to account pressure that userland is able to
 220          * help us with. For example, suppose that DMA zone is under
 221          * pressure; if we notify userland about that kind of pressure,
 222          * then it will be mostly a waste as it will trigger unnecessary
 223          * freeing of memory by userland (since userland is more likely to
 224          * have HIGHMEM/MOVABLE pages instead of the DMA fallback). That
 225          * is why we include only movable, highmem and FS/IO pages.
 226          * Indirect reclaim (kswapd) sets sc->gfp_mask to GFP_KERNEL, so
 227          * we account it too.
 228          */
 229         if (!(gfp & (__GFP_HIGHMEM | __GFP_MOVABLE | __GFP_IO | __GFP_FS)))
 230                 return;
 231
 232         /*
 233          * If we got here with no pages scanned, then that is an indicator
 234          * that reclaimer was unable to find any shrinkable LRUs at the
 235          * current scanning depth. But it does not mean that we should
 236          * report the critical pressure, yet. If the scanning priority
 237          * (scanning depth) goes too high (deep), we will be notified
 238          * through vmpressure_prio(). But so far, keep calm.
 239          */
 240         if (!scanned)
 241                 return;
 242
 243         spin_lock(&vmpr->sr_lock);
 244         vmpr->scanned += scanned;
 245         vmpr->reclaimed += reclaimed;
 246         scanned = vmpr->scanned;
 247         spin_unlock(&vmpr->sr_lock);
 248
 249         if (scanned < vmpressure_win)
 250                 return;
 251         schedule_work(&vmpr->work);
 252 }
 253
 254 /**
 255  * vmpressure_prio() - Account memory pressure through reclaimer priority level
 256  * @gfp:        reclaimer's gfp mask
 257  * @memcg:      cgroup memory controller handle
 258  * @prio:       reclaimer's priority
 259  *
 260  * This function should be called from the reclaim path every time when
 261  * the vmscan's reclaiming priority (scanning depth) changes.
 262  *
 263  * This function does not return any value.
 264  */
 265 void vmpressure_prio(gfp_t gfp, struct mem_cgroup *memcg, int prio)
 266 {
 267         /*
 268          * We only use prio for accounting critical level. For more info
 269          * see comment for vmpressure_level_critical_prio variable above.
 270          */
 271         if (prio > vmpressure_level_critical_prio)
 272                 return;
 273
 274         /*
 275          * OK, the prio is below the threshold, updating vmpressure
 276          * information before shrinker dives into long shrinking of long
 277          * range vmscan. Passing scanned = vmpressure_win, reclaimed = 0
 278          * to the vmpressure() basically means that we signal 'critical'
 279          * level.
 280          */
 281         vmpressure(gfp, memcg, vmpressure_win, 0);
 282 }
 283
 284 /**
 285  * vmpressure_register_event() - Bind vmpressure notifications to an eventfd
 286  * @cg:         cgroup that is interested in vmpressure notifications
 287  * @cft:        cgroup control files handle
 288  * @eventfd:    eventfd context to link notifications with
 289  * @args:       event arguments (used to set up a pressure level threshold)
 290  *
 291  * This function associates eventfd context with the vmpressure
 292  * infrastructure, so that the notifications will be delivered to the
 293  * @eventfd. The @args parameter is a string that denotes pressure level
 294  * threshold (one of vmpressure_str_levels, i.e. "low", "medium", or
 295  * "critical").
 296  *
 297  * This function should not be used directly, just pass it to (struct
 298  * cftype).register_event, and then cgroup core will handle everything by
 299  * itself.
 300  */
 301 int vmpressure_register_event(struct cgroup *cg, struct cftype *cft,
 302                               struct eventfd_ctx *eventfd, const char *args)
 303 {
 304         struct vmpressure *vmpr = cg_to_vmpressure(cg);
 305         struct vmpressure_event *ev;
 306         int level;
 307
 308         for (level = 0; level < VMPRESSURE_NUM_LEVELS; level++) {
 309                 if (!strcmp(vmpressure_str_levels[level], args))
 310                         break;
 311         }
 312
 313         if (level >= VMPRESSURE_NUM_LEVELS)
 314                 return -EINVAL;
 315
 316         ev = kzalloc(sizeof(*ev), GFP_KERNEL);
 317         if (!ev)
 318                 return -ENOMEM;
 319
 320         ev->efd = eventfd;
 321         ev->level = level;
 322
 323         mutex_lock(&vmpr->events_lock);
 324         list_add(&ev->node, &vmpr->events);
 325         mutex_unlock(&vmpr->events_lock);
 326
 327         return 0;
 328 }
 329
 330 /**
 331  * vmpressure_unregister_event() - Unbind eventfd from vmpressure
 332  * @cg:         cgroup handle
 333  * @cft:        cgroup control files handle
 334  * @eventfd:    eventfd context that was used to link vmpressure with the @cg
 335  *
 336  * This function does internal manipulations to detach the @eventfd from
 337  * the vmpressure notifications, and then frees internal resources
 338  * associated with the @eventfd (but the @eventfd itself is not freed).
 339  *
 340  * This function should not be used directly, just pass it to (struct
 341  * cftype).unregister_event, and then cgroup core will handle everything
 342  * by itself.
 343  */
 344 void vmpressure_unregister_event(struct cgroup *cg, struct cftype *cft,
 345                                  struct eventfd_ctx *eventfd)
 346 {
 347         struct vmpressure *vmpr = cg_to_vmpressure(cg);
 348         struct vmpressure_event *ev;
 349
 350         mutex_lock(&vmpr->events_lock);
 351         list_for_each_entry(ev, &vmpr->events, node) {
 352                 if (ev->efd != eventfd)
 353                         continue;
 354                 list_del(&ev->node);
 355                 kfree(ev);
 356                 break;
 357         }
 358         mutex_unlock(&vmpr->events_lock);
 359 }
 360
 361 /**
 362  * vmpressure_init() - Initialize vmpressure control structure
 363  * @vmpr:       Structure to be initialized
 364  *
 365  * This function should be called on every allocated vmpressure structure
 366  * before any usage.
 367  */
 368 void vmpressure_init(struct vmpressure *vmpr)
 369 {
 370         spin_lock_init(&vmpr->sr_lock);
 371         mutex_init(&vmpr->events_lock);
 372         INIT_LIST_HEAD(&vmpr->events);
 373         INIT_WORK(&vmpr->work, vmpressure_work_fn);
 374 }
 375
 376 /**
 377  * vmpressure_cleanup() - shuts down vmpressure control structure
 378  * @vmpr:       Structure to be cleaned up
 379  *
 380  * This function should be called before the structure in which it is
 381  * embedded is cleaned up.
 382  */
 383 void vmpressure_cleanup(struct vmpressure *vmpr)
 384 {
 385         /*
 386          * Make sure there is no pending work before eventfd infrastructure
 387          * goes away.
 388          */
 389         flush_work(&vmpr->work);
 390 }