git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/movinggc.c

   1 /*
   2  * Moving/copying garbage collector
   3  *
   4  * Copyright 2012 Google, Inc.
   5  */
   6
   7 #include "bcachefs.h"
   8 #include "btree_iter.h"
   9 #include "btree_update.h"
  10 #include "buckets.h"
  11 #include "clock.h"
  12 #include "extents.h"
  13 #include "eytzinger.h"
  14 #include "io.h"
  15 #include "keylist.h"
  16 #include "move.h"
  17 #include "movinggc.h"
  18 #include "super-io.h"
  19
  20 #include <trace/events/bcachefs.h>
  21 #include <linux/freezer.h>
  22 #include <linux/kthread.h>
  23 #include <linux/math64.h>
  24 #include <linux/sort.h>
  25 #include <linux/wait.h>
  26
  27 /*
  28  * We can't use the entire copygc reserve in one iteration of copygc: we may
  29  * need the buckets we're freeing up to go back into the copygc reserve to make
  30  * forward progress, but if the copygc reserve is full they'll be available for
  31  * any allocation - and it's possible that in a given iteration, we free up most
  32  * of the buckets we're going to free before we allocate most of the buckets
  33  * we're going to allocate.
  34  *
  35  * If we only use half of the reserve per iteration, then in steady state we'll
  36  * always have room in the reserve for the buckets we're going to need in the
  37  * next iteration:
  38  */
  39 #define COPYGC_BUCKETS_PER_ITER(ca)                                     \
  40         ((ca)->free[RESERVE_MOVINGGC].size / 2)
  41
  42 /*
  43  * Max sectors to move per iteration: Have to take into account internal
  44  * fragmentation from the multiple write points for each generation:
  45  */
  46 #define COPYGC_SECTORS_PER_ITER(ca)                                     \
  47         ((ca)->mi.bucket_size * COPYGC_BUCKETS_PER_ITER(ca))
  48
  49 static inline int sectors_used_cmp(copygc_heap *heap,
  50                                    struct copygc_heap_entry l,
  51                                    struct copygc_heap_entry r)
  52 {
  53         return bucket_sectors_used(l.mark) - bucket_sectors_used(r.mark);
  54 }
  55
  56 static int bucket_offset_cmp(const void *_l, const void *_r, size_t size)
  57 {
  58         const struct copygc_heap_entry *l = _l;
  59         const struct copygc_heap_entry *r = _r;
  60
  61         return (l->offset > r->offset) - (l->offset < r->offset);
  62 }
  63
  64 static bool __copygc_pred(struct bch_dev *ca,
  65                           struct bkey_s_c_extent e)
  66 {
  67         copygc_heap *h = &ca->copygc_heap;
  68         const struct bch_extent_ptr *ptr =
  69                 bch2_extent_has_device(e, ca->dev_idx);
  70
  71         if (ptr) {
  72                 struct copygc_heap_entry search = { .offset = ptr->offset };
  73
  74                 size_t i = eytzinger0_find_le(h->data, h->used,
  75                                               sizeof(h->data[0]),
  76                                               bucket_offset_cmp, &search);
  77
  78                 return (i >= 0 &&
  79                         ptr->offset < h->data[i].offset + ca->mi.bucket_size &&
  80                         ptr->gen == h->data[i].mark.gen);
  81         }
  82
  83         return false;
  84 }
  85
  86 static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
  87                                  enum bkey_type type,
  88                                  struct bkey_s_c_extent e,
  89                                  struct bch_io_opts *io_opts,
  90                                  struct data_opts *data_opts)
  91 {
  92         struct bch_dev *ca = arg;
  93
  94         if (!__copygc_pred(ca, e))
  95                 return DATA_SKIP;
  96
  97         data_opts->btree_insert_flags   = BTREE_INSERT_USE_RESERVE,
  98         data_opts->rewrite_dev          = ca->dev_idx;
  99         return DATA_REWRITE;
 100 }
 101
 102 static bool have_copygc_reserve(struct bch_dev *ca)
 103 {
 104         bool ret;
 105
 106         spin_lock(&ca->freelist_lock);
 107         ret = fifo_used(&ca->free[RESERVE_MOVINGGC]) >=
 108                 COPYGC_BUCKETS_PER_ITER(ca);
 109         spin_unlock(&ca->freelist_lock);
 110
 111         return ret;
 112 }
 113
 114 static void bch2_copygc(struct bch_fs *c, struct bch_dev *ca)
 115 {
 116         copygc_heap *h = &ca->copygc_heap;
 117         struct copygc_heap_entry e, *i;
 118         struct bucket_array *buckets;
 119         struct bch_move_stats move_stats;
 120         u64 sectors_to_move = 0, sectors_not_moved = 0;
 121         u64 buckets_to_move, buckets_not_moved = 0;
 122         size_t b;
 123         int ret;
 124
 125         memset(&move_stats, 0, sizeof(move_stats));
 126         closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
 127
 128         /*
 129          * Find buckets with lowest sector counts, skipping completely
 130          * empty buckets, by building a maxheap sorted by sector count,
 131          * and repeatedly replacing the maximum element until all
 132          * buckets have been visited.
 133          */
 134         h->used = 0;
 135
 136         /*
 137          * We need bucket marks to be up to date - gc can't be recalculating
 138          * them:
 139          */
 140         down_read(&c->gc_lock);
 141         down_read(&ca->bucket_lock);
 142         buckets = bucket_array(ca);
 143
 144         for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
 145                 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
 146                 struct copygc_heap_entry e;
 147
 148                 if (m.owned_by_allocator ||
 149                     m.data_type != BCH_DATA_USER ||
 150                     !bucket_sectors_used(m) ||
 151                     bucket_sectors_used(m) >= ca->mi.bucket_size)
 152                         continue;
 153
 154                 e = (struct copygc_heap_entry) {
 155                         .offset = bucket_to_sector(ca, b),
 156                         .mark   = m
 157                 };
 158                 heap_add_or_replace(h, e, -sectors_used_cmp);
 159         }
 160         up_read(&ca->bucket_lock);
 161         up_read(&c->gc_lock);
 162
 163         for (i = h->data; i < h->data + h->used; i++)
 164                 sectors_to_move += bucket_sectors_used(i->mark);
 165
 166         while (sectors_to_move > COPYGC_SECTORS_PER_ITER(ca)) {
 167                 BUG_ON(!heap_pop(h, e, -sectors_used_cmp));
 168                 sectors_to_move -= bucket_sectors_used(e.mark);
 169         }
 170
 171         buckets_to_move = h->used;
 172
 173         if (!buckets_to_move)
 174                 return;
 175
 176         eytzinger0_sort(h->data, h->used,
 177                         sizeof(h->data[0]),
 178                         bucket_offset_cmp, NULL);
 179
 180         ret = bch2_move_data(c, &ca->copygc_pd.rate,
 181                              SECTORS_IN_FLIGHT_PER_DEVICE,
 182                              &ca->self,
 183                              writepoint_ptr(&ca->copygc_write_point),
 184                              POS_MIN, POS_MAX,
 185                              copygc_pred, ca,
 186                              &move_stats);
 187
 188         down_read(&ca->bucket_lock);
 189         buckets = bucket_array(ca);
 190         for (i = h->data; i < h->data + h->used; i++) {
 191                 size_t b = sector_to_bucket(ca, i->offset);
 192                 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
 193
 194                 if (i->mark.gen == m.gen && bucket_sectors_used(m)) {
 195                         sectors_not_moved += bucket_sectors_used(m);
 196                         buckets_not_moved++;
 197                 }
 198         }
 199         up_read(&ca->bucket_lock);
 200
 201         if (sectors_not_moved && !ret)
 202                 bch_warn(c, "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved",
 203                          sectors_not_moved, sectors_to_move,
 204                          buckets_not_moved, buckets_to_move);
 205
 206         trace_copygc(ca,
 207                      atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
 208                      buckets_to_move, buckets_not_moved);
 209 }
 210
 211 static int bch2_copygc_thread(void *arg)
 212 {
 213         struct bch_dev *ca = arg;
 214         struct bch_fs *c = ca->fs;
 215         struct io_clock *clock = &c->io_clock[WRITE];
 216         unsigned long last;
 217         u64 available, want, next;
 218
 219         set_freezable();
 220
 221         while (!kthread_should_stop()) {
 222                 if (kthread_wait_freezable(c->copy_gc_enabled))
 223                         break;
 224
 225                 last = atomic_long_read(&clock->now);
 226                 /*
 227                  * don't start copygc until less than half the gc reserve is
 228                  * available:
 229                  */
 230                 available = dev_buckets_available(c, ca);
 231                 want = div64_u64((ca->mi.nbuckets - ca->mi.first_bucket) *
 232                                  c->opts.gc_reserve_percent, 200);
 233                 if (available > want) {
 234                         next = last + (available - want) *
 235                                 ca->mi.bucket_size;
 236                         bch2_kthread_io_clock_wait(clock, next);
 237                         continue;
 238                 }
 239
 240                 bch2_copygc(c, ca);
 241         }
 242
 243         return 0;
 244 }
 245
 246 void bch2_copygc_stop(struct bch_dev *ca)
 247 {
 248         ca->copygc_pd.rate.rate = UINT_MAX;
 249         bch2_ratelimit_reset(&ca->copygc_pd.rate);
 250
 251         if (ca->copygc_thread)
 252                 kthread_stop(ca->copygc_thread);
 253         ca->copygc_thread = NULL;
 254 }
 255
 256 int bch2_copygc_start(struct bch_fs *c, struct bch_dev *ca)
 257 {
 258         struct task_struct *t;
 259
 260         BUG_ON(ca->copygc_thread);
 261
 262         if (c->opts.nochanges)
 263                 return 0;
 264
 265         if (bch2_fs_init_fault("copygc_start"))
 266                 return -ENOMEM;
 267
 268         t = kthread_create(bch2_copygc_thread, ca, "bch_copygc");
 269         if (IS_ERR(t))
 270                 return PTR_ERR(t);
 271
 272         ca->copygc_thread = t;
 273         wake_up_process(ca->copygc_thread);
 274
 275         return 0;
 276 }
 277
 278 void bch2_dev_copygc_init(struct bch_dev *ca)
 279 {
 280         bch2_pd_controller_init(&ca->copygc_pd);
 281         ca->copygc_pd.d_term = 0;
 282 }