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(void *arg, struct bkey_s_c_extent e)
  65 {
  66         struct bch_dev *ca = arg;
  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 bool have_copygc_reserve(struct bch_dev *ca)
  87 {
  88         bool ret;
  89
  90         spin_lock(&ca->freelist_lock);
  91         ret = fifo_used(&ca->free[RESERVE_MOVINGGC]) >=
  92                 COPYGC_BUCKETS_PER_ITER(ca);
  93         spin_unlock(&ca->freelist_lock);
  94
  95         return ret;
  96 }
  97
  98 static void bch2_copygc(struct bch_fs *c, struct bch_dev *ca)
  99 {
 100         copygc_heap *h = &ca->copygc_heap;
 101         struct copygc_heap_entry e, *i;
 102         struct bucket *g;
 103         u64 keys_moved, sectors_moved;
 104         u64 sectors_to_move = 0, sectors_not_moved = 0;
 105         u64 buckets_to_move, buckets_not_moved = 0;
 106         int ret;
 107
 108         closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
 109
 110         /*
 111          * Find buckets with lowest sector counts, skipping completely
 112          * empty buckets, by building a maxheap sorted by sector count,
 113          * and repeatedly replacing the maximum element until all
 114          * buckets have been visited.
 115          */
 116
 117         /*
 118          * We need bucket marks to be up to date - gc can't be recalculating
 119          * them:
 120          */
 121         down_read(&c->gc_lock);
 122         h->used = 0;
 123         for_each_bucket(g, ca) {
 124                 struct bucket_mark m = READ_ONCE(g->mark);
 125                 struct copygc_heap_entry e;
 126
 127                 if (m.owned_by_allocator ||
 128                     m.data_type != BCH_DATA_USER ||
 129                     !bucket_sectors_used(m) ||
 130                     bucket_sectors_used(m) >= ca->mi.bucket_size)
 131                         continue;
 132
 133                 e = (struct copygc_heap_entry) {
 134                         .offset = bucket_to_sector(ca, g - ca->buckets),
 135                         .mark   = m
 136                 };
 137                 heap_add_or_replace(h, e, -sectors_used_cmp);
 138         }
 139         up_read(&c->gc_lock);
 140
 141         for (i = h->data; i < h->data + h->used; i++)
 142                 sectors_to_move += bucket_sectors_used(i->mark);
 143
 144         while (sectors_to_move > COPYGC_SECTORS_PER_ITER(ca)) {
 145                 BUG_ON(!heap_pop(h, e, -sectors_used_cmp));
 146                 sectors_to_move -= bucket_sectors_used(e.mark);
 147         }
 148
 149         buckets_to_move = h->used;
 150
 151         if (!buckets_to_move)
 152                 return;
 153
 154         eytzinger0_sort(h->data, h->used,
 155                         sizeof(h->data[0]),
 156                         bucket_offset_cmp, NULL);
 157
 158         ret = bch2_move_data(c, &ca->copygc_pd.rate,
 159                              SECTORS_IN_FLIGHT_PER_DEVICE,
 160                              &ca->self,
 161                              writepoint_ptr(&ca->copygc_write_point),
 162                              BTREE_INSERT_USE_RESERVE,
 163                              ca->dev_idx,
 164                              copygc_pred, ca,
 165                              &keys_moved,
 166                              &sectors_moved);
 167
 168         for (i = h->data; i < h->data + h->used; i++) {
 169                 size_t bucket = sector_to_bucket(ca, i->offset);
 170                 struct bucket_mark m = READ_ONCE(ca->buckets[bucket].mark);
 171
 172                 if (i->mark.gen == m.gen && bucket_sectors_used(m)) {
 173                         sectors_not_moved += bucket_sectors_used(m);
 174                         buckets_not_moved++;
 175                 }
 176         }
 177
 178         if (sectors_not_moved && !ret)
 179                 bch_warn(c, "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved",
 180                          sectors_not_moved, sectors_to_move,
 181                          buckets_not_moved, buckets_to_move);
 182
 183         trace_copygc(ca,
 184                      sectors_moved, sectors_not_moved,
 185                      buckets_to_move, buckets_not_moved);
 186 }
 187
 188 static int bch2_copygc_thread(void *arg)
 189 {
 190         struct bch_dev *ca = arg;
 191         struct bch_fs *c = ca->fs;
 192         struct io_clock *clock = &c->io_clock[WRITE];
 193         unsigned long last;
 194         u64 available, want, next;
 195
 196         set_freezable();
 197
 198         while (!kthread_should_stop()) {
 199                 if (kthread_wait_freezable(c->copy_gc_enabled))
 200                         break;
 201
 202                 last = atomic_long_read(&clock->now);
 203                 /*
 204                  * don't start copygc until less than half the gc reserve is
 205                  * available:
 206                  */
 207                 available = dev_buckets_available(c, ca);
 208                 want = div64_u64((ca->mi.nbuckets - ca->mi.first_bucket) *
 209                                  c->opts.gc_reserve_percent, 200);
 210                 if (available > want) {
 211                         next = last + (available - want) *
 212                                 ca->mi.bucket_size;
 213                         bch2_kthread_io_clock_wait(clock, next);
 214                         continue;
 215                 }
 216
 217                 bch2_copygc(c, ca);
 218         }
 219
 220         return 0;
 221 }
 222
 223 void bch2_copygc_stop(struct bch_dev *ca)
 224 {
 225         ca->copygc_pd.rate.rate = UINT_MAX;
 226         bch2_ratelimit_reset(&ca->copygc_pd.rate);
 227
 228         if (ca->copygc_thread)
 229                 kthread_stop(ca->copygc_thread);
 230         ca->copygc_thread = NULL;
 231 }
 232
 233 int bch2_copygc_start(struct bch_fs *c, struct bch_dev *ca)
 234 {
 235         struct task_struct *t;
 236
 237         BUG_ON(ca->copygc_thread);
 238
 239         if (c->opts.nochanges)
 240                 return 0;
 241
 242         if (bch2_fs_init_fault("copygc_start"))
 243                 return -ENOMEM;
 244
 245         t = kthread_create(bch2_copygc_thread, ca, "bch_copygc");
 246         if (IS_ERR(t))
 247                 return PTR_ERR(t);
 248
 249         ca->copygc_thread = t;
 250         wake_up_process(ca->copygc_thread);
 251
 252         return 0;
 253 }
 254
 255 void bch2_dev_copygc_init(struct bch_dev *ca)
 256 {
 257         bch2_pd_controller_init(&ca->copygc_pd);
 258         ca->copygc_pd.d_term = 0;
 259 }