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 "alloc_foreground.h"
   9 #include "btree_iter.h"
  10 #include "btree_update.h"
  11 #include "buckets.h"
  12 #include "clock.h"
  13 #include "disk_groups.h"
  14 #include "extents.h"
  15 #include "eytzinger.h"
  16 #include "io.h"
  17 #include "keylist.h"
  18 #include "move.h"
  19 #include "movinggc.h"
  20 #include "super-io.h"
  21
  22 #include <trace/events/bcachefs.h>
  23 #include <linux/freezer.h>
  24 #include <linux/kthread.h>
  25 #include <linux/math64.h>
  26 #include <linux/sched/task.h>
  27 #include <linux/sort.h>
  28 #include <linux/wait.h>
  29
  30 /*
  31  * We can't use the entire copygc reserve in one iteration of copygc: we may
  32  * need the buckets we're freeing up to go back into the copygc reserve to make
  33  * forward progress, but if the copygc reserve is full they'll be available for
  34  * any allocation - and it's possible that in a given iteration, we free up most
  35  * of the buckets we're going to free before we allocate most of the buckets
  36  * we're going to allocate.
  37  *
  38  * If we only use half of the reserve per iteration, then in steady state we'll
  39  * always have room in the reserve for the buckets we're going to need in the
  40  * next iteration:
  41  */
  42 #define COPYGC_BUCKETS_PER_ITER(ca)                                     \
  43         ((ca)->free[RESERVE_MOVINGGC].size / 2)
  44
  45 /*
  46  * Max sectors to move per iteration: Have to take into account internal
  47  * fragmentation from the multiple write points for each generation:
  48  */
  49 #define COPYGC_SECTORS_PER_ITER(ca)                                     \
  50         ((ca)->mi.bucket_size * COPYGC_BUCKETS_PER_ITER(ca))
  51
  52 static inline int sectors_used_cmp(copygc_heap *heap,
  53                                    struct copygc_heap_entry l,
  54                                    struct copygc_heap_entry r)
  55 {
  56         return (l.sectors > r.sectors) - (l.sectors < r.sectors);
  57 }
  58
  59 static int bucket_offset_cmp(const void *_l, const void *_r, size_t size)
  60 {
  61         const struct copygc_heap_entry *l = _l;
  62         const struct copygc_heap_entry *r = _r;
  63
  64         return (l->offset > r->offset) - (l->offset < r->offset);
  65 }
  66
  67 static bool __copygc_pred(struct bch_dev *ca,
  68                           struct bkey_s_c k)
  69 {
  70         copygc_heap *h = &ca->copygc_heap;
  71
  72         switch (k.k->type) {
  73         case KEY_TYPE_extent: {
  74                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
  75                 const struct bch_extent_ptr *ptr =
  76                         bch2_extent_has_device(e, ca->dev_idx);
  77
  78                 if (ptr) {
  79                         struct copygc_heap_entry search = { .offset = ptr->offset };
  80
  81                         ssize_t i = eytzinger0_find_le(h->data, h->used,
  82                                                        sizeof(h->data[0]),
  83                                                        bucket_offset_cmp, &search);
  84
  85                         return (i >= 0 &&
  86                                 ptr->offset < h->data[i].offset + ca->mi.bucket_size &&
  87                                 ptr->gen == h->data[i].gen);
  88                 }
  89                 break;
  90         }
  91         }
  92
  93         return false;
  94 }
  95
  96 static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
  97                                  struct bkey_s_c k,
  98                                  struct bch_io_opts *io_opts,
  99                                  struct data_opts *data_opts)
 100 {
 101         struct bch_dev *ca = arg;
 102
 103         if (!__copygc_pred(ca, k))
 104                 return DATA_SKIP;
 105
 106         data_opts->target               = dev_to_target(ca->dev_idx);
 107         data_opts->btree_insert_flags   = BTREE_INSERT_USE_RESERVE;
 108         data_opts->rewrite_dev          = ca->dev_idx;
 109         return DATA_REWRITE;
 110 }
 111
 112 static bool have_copygc_reserve(struct bch_dev *ca)
 113 {
 114         bool ret;
 115
 116         spin_lock(&ca->freelist_lock);
 117         ret = fifo_full(&ca->free[RESERVE_MOVINGGC]) ||
 118                 ca->allocator_blocked;
 119         spin_unlock(&ca->freelist_lock);
 120
 121         return ret;
 122 }
 123
 124 static void bch2_copygc(struct bch_fs *c, struct bch_dev *ca)
 125 {
 126         copygc_heap *h = &ca->copygc_heap;
 127         struct copygc_heap_entry e, *i;
 128         struct bucket_array *buckets;
 129         struct bch_move_stats move_stats;
 130         u64 sectors_to_move = 0, sectors_not_moved = 0;
 131         u64 buckets_to_move, buckets_not_moved = 0;
 132         size_t b;
 133         int ret;
 134
 135         memset(&move_stats, 0, sizeof(move_stats));
 136         closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
 137
 138         /*
 139          * Find buckets with lowest sector counts, skipping completely
 140          * empty buckets, by building a maxheap sorted by sector count,
 141          * and repeatedly replacing the maximum element until all
 142          * buckets have been visited.
 143          */
 144         h->used = 0;
 145
 146         /*
 147          * We need bucket marks to be up to date - gc can't be recalculating
 148          * them:
 149          */
 150         down_read(&c->gc_lock);
 151         down_read(&ca->bucket_lock);
 152         buckets = bucket_array(ca);
 153
 154         for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
 155                 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
 156                 struct copygc_heap_entry e;
 157
 158                 if (m.owned_by_allocator ||
 159                     m.data_type != BCH_DATA_USER ||
 160                     !bucket_sectors_used(m) ||
 161                     bucket_sectors_used(m) >= ca->mi.bucket_size)
 162                         continue;
 163
 164                 e = (struct copygc_heap_entry) {
 165                         .gen            = m.gen,
 166                         .sectors        = bucket_sectors_used(m),
 167                         .offset         = bucket_to_sector(ca, b),
 168                 };
 169                 heap_add_or_replace(h, e, -sectors_used_cmp, NULL);
 170         }
 171         up_read(&ca->bucket_lock);
 172         up_read(&c->gc_lock);
 173
 174         for (i = h->data; i < h->data + h->used; i++)
 175                 sectors_to_move += i->sectors;
 176
 177         while (sectors_to_move > COPYGC_SECTORS_PER_ITER(ca)) {
 178                 BUG_ON(!heap_pop(h, e, -sectors_used_cmp, NULL));
 179                 sectors_to_move -= e.sectors;
 180         }
 181
 182         buckets_to_move = h->used;
 183
 184         if (!buckets_to_move)
 185                 return;
 186
 187         eytzinger0_sort(h->data, h->used,
 188                         sizeof(h->data[0]),
 189                         bucket_offset_cmp, NULL);
 190
 191         ret = bch2_move_data(c, &ca->copygc_pd.rate,
 192                              writepoint_ptr(&ca->copygc_write_point),
 193                              POS_MIN, POS_MAX,
 194                              copygc_pred, ca,
 195                              &move_stats);
 196
 197         down_read(&ca->bucket_lock);
 198         buckets = bucket_array(ca);
 199         for (i = h->data; i < h->data + h->used; i++) {
 200                 size_t b = sector_to_bucket(ca, i->offset);
 201                 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
 202
 203                 if (i->gen == m.gen && bucket_sectors_used(m)) {
 204                         sectors_not_moved += bucket_sectors_used(m);
 205                         buckets_not_moved++;
 206                 }
 207         }
 208         up_read(&ca->bucket_lock);
 209
 210         if (sectors_not_moved && !ret)
 211                 bch_warn(c, "copygc finished but %llu/%llu sectors, %llu/%llu buckets not moved",
 212                          sectors_not_moved, sectors_to_move,
 213                          buckets_not_moved, buckets_to_move);
 214
 215         trace_copygc(ca,
 216                      atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
 217                      buckets_to_move, buckets_not_moved);
 218 }
 219
 220 static int bch2_copygc_thread(void *arg)
 221 {
 222         struct bch_dev *ca = arg;
 223         struct bch_fs *c = ca->fs;
 224         struct io_clock *clock = &c->io_clock[WRITE];
 225         struct bch_dev_usage usage;
 226         unsigned long last;
 227         u64 available, fragmented, reserve, next;
 228
 229         set_freezable();
 230
 231         while (!kthread_should_stop()) {
 232                 if (kthread_wait_freezable(c->copy_gc_enabled))
 233                         break;
 234
 235                 last = atomic_long_read(&clock->now);
 236
 237                 reserve = ca->copygc_threshold;
 238
 239                 usage = bch2_dev_usage_read(c, ca);
 240
 241                 available = __dev_buckets_available(ca, usage) *
 242                         ca->mi.bucket_size;
 243                 if (available > reserve) {
 244                         next = last + available - reserve;
 245                         bch2_kthread_io_clock_wait(clock, next,
 246                                         MAX_SCHEDULE_TIMEOUT);
 247                         continue;
 248                 }
 249
 250                 /*
 251                  * don't start copygc until there's more than half the copygc
 252                  * reserve of fragmented space:
 253                  */
 254                 fragmented = usage.sectors_fragmented;
 255                 if (fragmented < reserve) {
 256                         next = last + reserve - fragmented;
 257                         bch2_kthread_io_clock_wait(clock, next,
 258                                         MAX_SCHEDULE_TIMEOUT);
 259                         continue;
 260                 }
 261
 262                 bch2_copygc(c, ca);
 263         }
 264
 265         return 0;
 266 }
 267
 268 void bch2_copygc_stop(struct bch_dev *ca)
 269 {
 270         ca->copygc_pd.rate.rate = UINT_MAX;
 271         bch2_ratelimit_reset(&ca->copygc_pd.rate);
 272
 273         if (ca->copygc_thread) {
 274                 kthread_stop(ca->copygc_thread);
 275                 put_task_struct(ca->copygc_thread);
 276         }
 277         ca->copygc_thread = NULL;
 278 }
 279
 280 int bch2_copygc_start(struct bch_fs *c, struct bch_dev *ca)
 281 {
 282         struct task_struct *t;
 283
 284         BUG_ON(ca->copygc_thread);
 285
 286         if (c->opts.nochanges)
 287                 return 0;
 288
 289         if (bch2_fs_init_fault("copygc_start"))
 290                 return -ENOMEM;
 291
 292         t = kthread_create(bch2_copygc_thread, ca,
 293                            "bch_copygc[%s]", ca->name);
 294         if (IS_ERR(t))
 295                 return PTR_ERR(t);
 296
 297         get_task_struct(t);
 298
 299         ca->copygc_thread = t;
 300         wake_up_process(ca->copygc_thread);
 301
 302         return 0;
 303 }
 304
 305 void bch2_dev_copygc_init(struct bch_dev *ca)
 306 {
 307         bch2_pd_controller_init(&ca->copygc_pd);
 308         ca->copygc_pd.d_term = 0;
 309 }