mm/page_counter.c

   1 /*
   2  * Lockless hierarchical page accounting & limiting
   3  *
   4  * Copyright (C) 2014 Red Hat, Inc., Johannes Weiner
   5  */
   6
   7 #include <linux/page_counter.h>
   8 #include <linux/atomic.h>
   9 #include <linux/kernel.h>
  10 #include <linux/string.h>
  11 #include <linux/sched.h>
  12 #include <linux/bug.h>
  13 #include <asm/page.h>
  14
  15 /**
  16  * page_counter_cancel - take pages out of the local counter
  17  * @counter: counter
  18  * @nr_pages: number of pages to cancel
  19  */
  20 void page_counter_cancel(struct page_counter *counter, unsigned long nr_pages)
  21 {
  22         long new;
  23
  24         new = atomic_long_sub_return(nr_pages, &counter->count);
  25         /* More uncharges than charges? */
  26         WARN_ON_ONCE(new < 0);
  27 }
  28
  29 /**
  30  * page_counter_charge - hierarchically charge pages
  31  * @counter: counter
  32  * @nr_pages: number of pages to charge
  33  *
  34  * NOTE: This does not consider any configured counter limits.
  35  */
  36 void page_counter_charge(struct page_counter *counter, unsigned long nr_pages)
  37 {
  38         struct page_counter *c;
  39
  40         for (c = counter; c; c = c->parent) {
  41                 long new;
  42
  43                 new = atomic_long_add_return(nr_pages, &c->count);
  44                 /*
  45                  * This is indeed racy, but we can live with some
  46                  * inaccuracy in the watermark.
  47                  */
  48                 if (new > c->watermark)
  49                         c->watermark = new;
  50         }
  51 }
  52
  53 /**
  54  * page_counter_try_charge - try to hierarchically charge pages
  55  * @counter: counter
  56  * @nr_pages: number of pages to charge
  57  * @fail: points first counter to hit its limit, if any
  58  *
  59  * Returns 0 on success, or -ENOMEM and @fail if the counter or one of
  60  * its ancestors has hit its configured limit.
  61  */
  62 int page_counter_try_charge(struct page_counter *counter,
  63                             unsigned long nr_pages,
  64                             struct page_counter **fail)
  65 {
  66         struct page_counter *c;
  67
  68         for (c = counter; c; c = c->parent) {
  69                 long new;
  70                 /*
  71                  * Charge speculatively to avoid an expensive CAS.  If
  72                  * a bigger charge fails, it might falsely lock out a
  73                  * racing smaller charge and send it into reclaim
  74                  * early, but the error is limited to the difference
  75                  * between the two sizes, which is less than 2M/4M in
  76                  * case of a THP locking out a regular page charge.
  77                  *
  78                  * The atomic_long_add_return() implies a full memory
  79                  * barrier between incrementing the count and reading
  80                  * the limit.  When racing with page_counter_limit(),
  81                  * we either see the new limit or the setter sees the
  82                  * counter has changed and retries.
  83                  */
  84                 new = atomic_long_add_return(nr_pages, &c->count);
  85                 if (new > c->limit) {
  86                         atomic_long_sub(nr_pages, &c->count);
  87                         /*
  88                          * This is racy, but we can live with some
  89                          * inaccuracy in the failcnt.
  90                          */
  91                         c->failcnt++;
  92                         *fail = c;
  93                         goto failed;
  94                 }
  95                 /*
  96                  * Just like with failcnt, we can live with some
  97                  * inaccuracy in the watermark.
  98                  */
  99                 if (new > c->watermark)
 100                         c->watermark = new;
 101         }
 102         return 0;
 103
 104 failed:
 105         for (c = counter; c != *fail; c = c->parent)
 106                 page_counter_cancel(c, nr_pages);
 107
 108         return -ENOMEM;
 109 }
 110
 111 /**
 112  * page_counter_uncharge - hierarchically uncharge pages
 113  * @counter: counter
 114  * @nr_pages: number of pages to uncharge
 115  */
 116 void page_counter_uncharge(struct page_counter *counter, unsigned long nr_pages)
 117 {
 118         struct page_counter *c;
 119
 120         for (c = counter; c; c = c->parent)
 121                 page_counter_cancel(c, nr_pages);
 122 }
 123
 124 /**
 125  * page_counter_limit - limit the number of pages allowed
 126  * @counter: counter
 127  * @limit: limit to set
 128  *
 129  * Returns 0 on success, -EBUSY if the current number of pages on the
 130  * counter already exceeds the specified limit.
 131  *
 132  * The caller must serialize invocations on the same counter.
 133  */
 134 int page_counter_limit(struct page_counter *counter, unsigned long limit)
 135 {
 136         for (;;) {
 137                 unsigned long old;
 138                 long count;
 139
 140                 /*
 141                  * Update the limit while making sure that it's not
 142                  * below the concurrently-changing counter value.
 143                  *
 144                  * The xchg implies two full memory barriers before
 145                  * and after, so the read-swap-read is ordered and
 146                  * ensures coherency with page_counter_try_charge():
 147                  * that function modifies the count before checking
 148                  * the limit, so if it sees the old limit, we see the
 149                  * modified counter and retry.
 150                  */
 151                 count = atomic_long_read(&counter->count);
 152
 153                 if (count > limit)
 154                         return -EBUSY;
 155
 156                 old = xchg(&counter->limit, limit);
 157
 158                 if (atomic_long_read(&counter->count) <= count)
 159                         return 0;
 160
 161                 counter->limit = old;
 162                 cond_resched();
 163         }
 164 }
 165
 166 /**
 167  * page_counter_memparse - memparse() for page counter limits
 168  * @buf: string to parse
 169  * @max: string meaning maximum possible value
 170  * @nr_pages: returns the result in number of pages
 171  *
 172  * Returns -EINVAL, or 0 and @nr_pages on success.  @nr_pages will be
 173  * limited to %PAGE_COUNTER_MAX.
 174  */
 175 int page_counter_memparse(const char *buf, const char *max,
 176                           unsigned long *nr_pages)
 177 {
 178         char *end;
 179         u64 bytes;
 180
 181         if (!strcmp(buf, max)) {
 182                 *nr_pages = PAGE_COUNTER_MAX;
 183                 return 0;
 184         }
 185
 186         bytes = memparse(buf, &end);
 187         if (*end != '\0')
 188                 return -EINVAL;
 189
 190         *nr_pages = min(bytes / PAGE_SIZE, (u64)PAGE_COUNTER_MAX);
 191
 192         return 0;
 193 }