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 struct moving_io {
  13         struct keybuf_key       *w;
  14         struct search           s;
  15         struct bbio             bio;
  16 };
  17
  18 static bool moving_pred(struct keybuf *buf, struct bkey *k)
  19 {
  20         struct cache_set *c = container_of(buf, struct cache_set,
  21                                            moving_gc_keys);
  22         unsigned i;
  23
  24         for (i = 0; i < KEY_PTRS(k); i++) {
  25                 struct cache *ca = PTR_CACHE(c, k, i);
  26                 struct bucket *g = PTR_BUCKET(c, k, i);
  27
  28                 if (GC_SECTORS_USED(g) < ca->gc_move_threshold)
  29                         return true;
  30         }
  31
  32         return false;
  33 }
  34
  35 /* Moving GC - IO loop */
  36
  37 static void moving_io_destructor(struct closure *cl)
  38 {
  39         struct moving_io *io = container_of(cl, struct moving_io, s.cl);
  40         kfree(io);
  41 }
  42
  43 static void write_moving_finish(struct closure *cl)
  44 {
  45         struct moving_io *io = container_of(cl, struct moving_io, s.cl);
  46         struct bio *bio = &io->bio.bio;
  47         struct bio_vec *bv = bio_iovec_idx(bio, bio->bi_vcnt);
  48
  49         while (bv-- != bio->bi_io_vec)
  50                 __free_page(bv->bv_page);
  51
  52         pr_debug("%s %s", io->s.op.insert_collision
  53                  ? "collision moving" : "moved",
  54                  pkey(&io->w->key));
  55
  56         bch_keybuf_del(&io->s.op.c->moving_gc_keys, io->w);
  57
  58         atomic_dec_bug(&io->s.op.c->in_flight);
  59         closure_wake_up(&io->s.op.c->moving_gc_wait);
  60
  61         closure_return_with_destructor(cl, moving_io_destructor);
  62 }
  63
  64 static void read_moving_endio(struct bio *bio, int error,
  65                               struct batch_complete *batch)
  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                 trace_bcache_write_moving(&io->bio.bio);
  99
 100                 moving_init(io);
 101
 102                 io->bio.bio.bi_sector   = KEY_START(&io->w->key);
 103                 s->op.lock              = -1;
 104                 s->op.write_prio        = 1;
 105                 s->op.cache_bio         = &io->bio.bio;
 106
 107                 s->writeback            = KEY_DIRTY(&io->w->key);
 108                 s->op.csum              = KEY_CSUM(&io->w->key);
 109
 110                 s->op.type = BTREE_REPLACE;
 111                 bkey_copy(&s->op.replace, &io->w->key);
 112
 113                 closure_init(&s->op.cl, cl);
 114                 bch_insert_data(&s->op.cl);
 115         }
 116
 117         continue_at(cl, write_moving_finish, NULL);
 118 }
 119
 120 static void read_moving_submit(struct closure *cl)
 121 {
 122         struct search *s = container_of(cl, struct search, cl);
 123         struct moving_io *io = container_of(s, struct moving_io, s);
 124         struct bio *bio = &io->bio.bio;
 125
 126         trace_bcache_read_moving(bio);
 127         bch_submit_bbio(bio, s->op.c, &io->w->key, 0);
 128
 129         continue_at(cl, write_moving, bch_gc_wq);
 130 }
 131
 132 static void read_moving(struct closure *cl)
 133 {
 134         struct cache_set *c = container_of(cl, struct cache_set, moving_gc);
 135         struct keybuf_key *w;
 136         struct moving_io *io;
 137         struct bio *bio;
 138
 139         /* XXX: if we error, background writeback could stall indefinitely */
 140
 141         while (!test_bit(CACHE_SET_STOPPING, &c->flags)) {
 142                 w = bch_keybuf_next_rescan(c, &c->moving_gc_keys, &MAX_KEY);
 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 (bch_bio_alloc_pages(bio, GFP_KERNEL))
 164                         goto err;
 165
 166                 pr_debug("%s", pkey(&w->key));
 167
 168                 closure_call(&io->s.cl, read_moving_submit, NULL, &c->gc.cl);
 169
 170                 if (atomic_inc_return(&c->in_flight) >= 64) {
 171                         closure_wait_event(&c->moving_gc_wait, cl,
 172                                            atomic_read(&c->in_flight) < 64);
 173                         continue_at(cl, read_moving, bch_gc_wq);
 174                 }
 175         }
 176
 177         if (0) {
 178 err:            if (!IS_ERR_OR_NULL(w->private))
 179                         kfree(w->private);
 180
 181                 bch_keybuf_del(&c->moving_gc_keys, w);
 182         }
 183
 184         closure_return(cl);
 185 }
 186
 187 static bool bucket_cmp(struct bucket *l, struct bucket *r)
 188 {
 189         return GC_SECTORS_USED(l) < GC_SECTORS_USED(r);
 190 }
 191
 192 static unsigned bucket_heap_top(struct cache *ca)
 193 {
 194         return GC_SECTORS_USED(heap_peek(&ca->heap));
 195 }
 196
 197 void bch_moving_gc(struct closure *cl)
 198 {
 199         struct cache_set *c = container_of(cl, struct cache_set, gc.cl);
 200         struct cache *ca;
 201         struct bucket *b;
 202         unsigned i;
 203
 204         if (!c->copy_gc_enabled)
 205                 closure_return(cl);
 206
 207         mutex_lock(&c->bucket_lock);
 208
 209         for_each_cache(ca, c, i) {
 210                 unsigned sectors_to_move = 0;
 211                 unsigned reserve_sectors = ca->sb.bucket_size *
 212                         min(fifo_used(&ca->free), ca->free.size / 2);
 213
 214                 ca->heap.used = 0;
 215
 216                 for_each_bucket(b, ca) {
 217                         if (!GC_SECTORS_USED(b))
 218                                 continue;
 219
 220                         if (!heap_full(&ca->heap)) {
 221                                 sectors_to_move += GC_SECTORS_USED(b);
 222                                 heap_add(&ca->heap, b, bucket_cmp);
 223                         } else if (bucket_cmp(b, heap_peek(&ca->heap))) {
 224                                 sectors_to_move -= bucket_heap_top(ca);
 225                                 sectors_to_move += GC_SECTORS_USED(b);
 226
 227                                 ca->heap.data[0] = b;
 228                                 heap_sift(&ca->heap, 0, bucket_cmp);
 229                         }
 230                 }
 231
 232                 while (sectors_to_move > reserve_sectors) {
 233                         heap_pop(&ca->heap, b, bucket_cmp);
 234                         sectors_to_move -= GC_SECTORS_USED(b);
 235                 }
 236
 237                 ca->gc_move_threshold = bucket_heap_top(ca);
 238
 239                 pr_debug("threshold %u", ca->gc_move_threshold);
 240         }
 241
 242         mutex_unlock(&c->bucket_lock);
 243
 244         c->moving_gc_keys.last_scanned = ZERO_KEY;
 245
 246         closure_init(&c->moving_gc, cl);
 247         read_moving(&c->moving_gc);
 248
 249         closure_return(cl);
 250 }
 251
 252 void bch_moving_init_cache_set(struct cache_set *c)
 253 {
 254         bch_keybuf_init(&c->moving_gc_keys, moving_pred);
 255 }