drivers/md/bcache/movinggc.c

   1 /*
   2  * Moving/copying garbage collector
   3  *
   4  * Copyright 2012 Google, Inc.
   5  */
   6
   7 #include "bcache.h"
   8 #include "btree.h"
   9 #include "debug.h"
  10 #include "request.h"
  11
  12 #include <trace/events/bcache.h>
  13
  14 struct moving_io {
  15         struct keybuf_key       *w;
  16         struct search           s;
  17         struct bbio             bio;
  18 };
  19
  20 static bool moving_pred(struct keybuf *buf, struct bkey *k)
  21 {
  22         struct cache_set *c = container_of(buf, struct cache_set,
  23                                            moving_gc_keys);
  24         unsigned i;
  25
  26         for (i = 0; i < KEY_PTRS(k); i++) {
  27                 struct cache *ca = PTR_CACHE(c, k, i);
  28                 struct bucket *g = PTR_BUCKET(c, k, i);
  29
  30                 if (GC_SECTORS_USED(g) < ca->gc_move_threshold)
  31                         return true;
  32         }
  33
  34         return false;
  35 }
  36
  37 /* Moving GC - IO loop */
  38
  39 static void moving_io_destructor(struct closure *cl)
  40 {
  41         struct moving_io *io = container_of(cl, struct moving_io, s.cl);
  42         kfree(io);
  43 }
  44
  45 static void write_moving_finish(struct closure *cl)
  46 {
  47         struct moving_io *io = container_of(cl, struct moving_io, s.cl);
  48         struct bio *bio = &io->bio.bio;
  49         struct bio_vec *bv;
  50         int i;
  51
  52         bio_for_each_segment_all(bv, bio, i)
  53                 __free_page(bv->bv_page);
  54
  55         if (io->s.op.insert_collision)
  56                 trace_bcache_gc_copy_collision(&io->w->key);
  57
  58         bch_keybuf_del(&io->s.op.c->moving_gc_keys, io->w);
  59
  60         up(&io->s.op.c->moving_in_flight);
  61
  62         closure_return_with_destructor(cl, moving_io_destructor);
  63 }
  64
  65 static void read_moving_endio(struct bio *bio, int error)
  66 {
  67         struct moving_io *io = container_of(bio->bi_private,
  68                                             struct moving_io, s.cl);
  69
  70         if (error)
  71                 io->s.error = error;
  72
  73         bch_bbio_endio(io->s.op.c, bio, error, "reading data to move");
  74 }
  75
  76 static void moving_init(struct moving_io *io)
  77 {
  78         struct bio *bio = &io->bio.bio;
  79
  80         bio_init(bio);
  81         bio_get(bio);
  82         bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
  83
  84         bio->bi_size            = KEY_SIZE(&io->w->key) << 9;
  85         bio->bi_max_vecs        = DIV_ROUND_UP(KEY_SIZE(&io->w->key),
  86                                                PAGE_SECTORS);
  87         bio->bi_private         = &io->s.cl;
  88         bio->bi_io_vec          = bio->bi_inline_vecs;
  89         bch_bio_map(bio, NULL);
  90 }
  91
  92 static void write_moving(struct closure *cl)
  93 {
  94         struct search *s = container_of(cl, struct search, cl);
  95         struct moving_io *io = container_of(s, struct moving_io, s);
  96
  97         if (!s->error) {
  98                 moving_init(io);
  99
 100                 io->bio.bio.bi_sector   = KEY_START(&io->w->key);
 101                 s->op.lock              = -1;
 102                 s->op.write_prio        = 1;
 103                 s->op.cache_bio         = &io->bio.bio;
 104
 105                 s->writeback            = KEY_DIRTY(&io->w->key);
 106                 s->op.csum              = KEY_CSUM(&io->w->key);
 107
 108                 s->op.type = BTREE_REPLACE;
 109                 bkey_copy(&s->op.replace, &io->w->key);
 110
 111                 closure_init(&s->op.cl, cl);
 112                 bch_data_insert(&s->op.cl);
 113         }
 114
 115         continue_at(cl, write_moving_finish, system_wq);
 116 }
 117
 118 static void read_moving_submit(struct closure *cl)
 119 {
 120         struct search *s = container_of(cl, struct search, cl);
 121         struct moving_io *io = container_of(s, struct moving_io, s);
 122         struct bio *bio = &io->bio.bio;
 123
 124         bch_submit_bbio(bio, s->op.c, &io->w->key, 0);
 125
 126         continue_at(cl, write_moving, system_wq);
 127 }
 128
 129 static void read_moving(struct cache_set *c)
 130 {
 131         struct keybuf_key *w;
 132         struct moving_io *io;
 133         struct bio *bio;
 134         struct closure cl;
 135
 136         closure_init_stack(&cl);
 137
 138         /* XXX: if we error, background writeback could stall indefinitely */
 139
 140         while (!test_bit(CACHE_SET_STOPPING, &c->flags)) {
 141                 w = bch_keybuf_next_rescan(c, &c->moving_gc_keys,
 142                                            &MAX_KEY, moving_pred);
 143                 if (!w)
 144                         break;
 145
 146                 io = kzalloc(sizeof(struct moving_io) + sizeof(struct bio_vec)
 147                              * DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
 148                              GFP_KERNEL);
 149                 if (!io)
 150                         goto err;
 151
 152                 w->private      = io;
 153                 io->w           = w;
 154                 io->s.op.inode  = KEY_INODE(&w->key);
 155                 io->s.op.c      = c;
 156
 157                 moving_init(io);
 158                 bio = &io->bio.bio;
 159
 160                 bio->bi_rw      = READ;
 161                 bio->bi_end_io  = read_moving_endio;
 162
 163                 if (bio_alloc_pages(bio, GFP_KERNEL))
 164                         goto err;
 165
 166                 trace_bcache_gc_copy(&w->key);
 167
 168                 down(&c->moving_in_flight);
 169                 closure_call(&io->s.cl, read_moving_submit, NULL, &cl);
 170         }
 171
 172         if (0) {
 173 err:            if (!IS_ERR_OR_NULL(w->private))
 174                         kfree(w->private);
 175
 176                 bch_keybuf_del(&c->moving_gc_keys, w);
 177         }
 178
 179         closure_sync(&cl);
 180 }
 181
 182 static bool bucket_cmp(struct bucket *l, struct bucket *r)
 183 {
 184         return GC_SECTORS_USED(l) < GC_SECTORS_USED(r);
 185 }
 186
 187 static unsigned bucket_heap_top(struct cache *ca)
 188 {
 189         return GC_SECTORS_USED(heap_peek(&ca->heap));
 190 }
 191
 192 void bch_moving_gc(struct cache_set *c)
 193 {
 194         struct cache *ca;
 195         struct bucket *b;
 196         unsigned i;
 197
 198         if (!c->copy_gc_enabled)
 199                 return;
 200
 201         mutex_lock(&c->bucket_lock);
 202
 203         for_each_cache(ca, c, i) {
 204                 unsigned sectors_to_move = 0;
 205                 unsigned reserve_sectors = ca->sb.bucket_size *
 206                         min(fifo_used(&ca->free), ca->free.size / 2);
 207
 208                 ca->heap.used = 0;
 209
 210                 for_each_bucket(b, ca) {
 211                         if (!GC_SECTORS_USED(b))
 212                                 continue;
 213
 214                         if (!heap_full(&ca->heap)) {
 215                                 sectors_to_move += GC_SECTORS_USED(b);
 216                                 heap_add(&ca->heap, b, bucket_cmp);
 217                         } else if (bucket_cmp(b, heap_peek(&ca->heap))) {
 218                                 sectors_to_move -= bucket_heap_top(ca);
 219                                 sectors_to_move += GC_SECTORS_USED(b);
 220
 221                                 ca->heap.data[0] = b;
 222                                 heap_sift(&ca->heap, 0, bucket_cmp);
 223                         }
 224                 }
 225
 226                 while (sectors_to_move > reserve_sectors) {
 227                         heap_pop(&ca->heap, b, bucket_cmp);
 228                         sectors_to_move -= GC_SECTORS_USED(b);
 229                 }
 230
 231                 ca->gc_move_threshold = bucket_heap_top(ca);
 232
 233                 pr_debug("threshold %u", ca->gc_move_threshold);
 234         }
 235
 236         mutex_unlock(&c->bucket_lock);
 237
 238         c->moving_gc_keys.last_scanned = ZERO_KEY;
 239
 240         read_moving(c);
 241 }
 242
 243 void bch_moving_init_cache_set(struct cache_set *c)
 244 {
 245         bch_keybuf_init(&c->moving_gc_keys);
 246         sema_init(&c->moving_in_flight, 64);
 247 }