drivers/md/bcache/io.c

   1 /*
   2  * Some low level IO code, and hacks for various block layer limitations
   3  *
   4  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   5  * Copyright 2012 Google, Inc.
   6  */
   7
   8 #include "bcache.h"
   9 #include "bset.h"
  10 #include "debug.h"
  11
  12 #include <linux/blkdev.h>
  13
  14 /**
  15  * bch_bio_split - split a bio
  16  * @bio:        bio to split
  17  * @sectors:    number of sectors to split from the front of @bio
  18  * @gfp:        gfp mask
  19  * @bs:         bio set to allocate from
  20  *
  21  * Allocates and returns a new bio which represents @sectors from the start of
  22  * @bio, and updates @bio to represent the remaining sectors.
  23  *
  24  * If bio_sectors(@bio) was less than or equal to @sectors, returns @bio
  25  * unchanged.
  26  *
  27  * The newly allocated bio will point to @bio's bi_io_vec, if the split was on a
  28  * bvec boundry; it is the caller's responsibility to ensure that @bio is not
  29  * freed before the split.
  30  */
  31 struct bio *bch_bio_split(struct bio *bio, int sectors,
  32                           gfp_t gfp, struct bio_set *bs)
  33 {
  34         unsigned vcnt = 0, nbytes = sectors << 9;
  35         struct bio_vec bv;
  36         struct bvec_iter iter;
  37         struct bio *ret = NULL;
  38
  39         BUG_ON(sectors <= 0);
  40
  41         if (sectors >= bio_sectors(bio))
  42                 return bio;
  43
  44         if (bio->bi_rw & REQ_DISCARD) {
  45                 ret = bio_alloc_bioset(gfp, 1, bs);
  46                 if (!ret)
  47                         return NULL;
  48                 goto out;
  49         }
  50
  51         bio_for_each_segment(bv, bio, iter) {
  52                 vcnt++;
  53
  54                 if (nbytes <= bv.bv_len)
  55                         break;
  56
  57                 nbytes -= bv.bv_len;
  58         }
  59
  60         ret = bio_alloc_bioset(gfp, vcnt, bs);
  61         if (!ret)
  62                 return NULL;
  63
  64         bio_for_each_segment(bv, bio, iter) {
  65                 ret->bi_io_vec[ret->bi_vcnt++] = bv;
  66
  67                 if (ret->bi_vcnt == vcnt)
  68                         break;
  69         }
  70
  71         ret->bi_io_vec[ret->bi_vcnt - 1].bv_len = nbytes;
  72 out:
  73         ret->bi_bdev    = bio->bi_bdev;
  74         ret->bi_iter.bi_sector  = bio->bi_iter.bi_sector;
  75         ret->bi_iter.bi_size    = sectors << 9;
  76         ret->bi_rw      = bio->bi_rw;
  77
  78         if (bio_integrity(bio)) {
  79                 if (bio_integrity_clone(ret, bio, gfp)) {
  80                         bio_put(ret);
  81                         return NULL;
  82                 }
  83
  84                 bio_integrity_trim(ret, 0, bio_sectors(ret));
  85         }
  86
  87         bio_advance(bio, ret->bi_iter.bi_size);
  88
  89         return ret;
  90 }
  91
  92 static unsigned bch_bio_max_sectors(struct bio *bio)
  93 {
  94         struct request_queue *q = bdev_get_queue(bio->bi_bdev);
  95         struct bio_vec bv;
  96         struct bvec_iter iter;
  97         unsigned ret = 0, seg = 0;
  98
  99         if (bio->bi_rw & REQ_DISCARD)
 100                 return min(bio_sectors(bio), q->limits.max_discard_sectors);
 101
 102         bio_for_each_segment(bv, bio, iter) {
 103                 struct bvec_merge_data bvm = {
 104                         .bi_bdev        = bio->bi_bdev,
 105                         .bi_sector      = bio->bi_iter.bi_sector,
 106                         .bi_size        = ret << 9,
 107                         .bi_rw          = bio->bi_rw,
 108                 };
 109
 110                 if (seg == min_t(unsigned, BIO_MAX_PAGES,
 111                                  queue_max_segments(q)))
 112                         break;
 113
 114                 if (q->merge_bvec_fn &&
 115                     q->merge_bvec_fn(q, &bvm, &bv) < (int) bv.bv_len)
 116                         break;
 117
 118                 seg++;
 119                 ret += bv.bv_len >> 9;
 120         }
 121
 122         ret = min(ret, queue_max_sectors(q));
 123
 124         WARN_ON(!ret);
 125         ret = max_t(int, ret, bio_iovec(bio).bv_len >> 9);
 126
 127         return ret;
 128 }
 129
 130 static void bch_bio_submit_split_done(struct closure *cl)
 131 {
 132         struct bio_split_hook *s = container_of(cl, struct bio_split_hook, cl);
 133
 134         s->bio->bi_end_io = s->bi_end_io;
 135         s->bio->bi_private = s->bi_private;
 136         bio_endio_nodec(s->bio, 0);
 137
 138         closure_debug_destroy(&s->cl);
 139         mempool_free(s, s->p->bio_split_hook);
 140 }
 141
 142 static void bch_bio_submit_split_endio(struct bio *bio, int error)
 143 {
 144         struct closure *cl = bio->bi_private;
 145         struct bio_split_hook *s = container_of(cl, struct bio_split_hook, cl);
 146
 147         if (error)
 148                 clear_bit(BIO_UPTODATE, &s->bio->bi_flags);
 149
 150         bio_put(bio);
 151         closure_put(cl);
 152 }
 153
 154 void bch_generic_make_request(struct bio *bio, struct bio_split_pool *p)
 155 {
 156         struct bio_split_hook *s;
 157         struct bio *n;
 158
 159         if (!bio_has_data(bio) && !(bio->bi_rw & REQ_DISCARD))
 160                 goto submit;
 161
 162         if (bio_sectors(bio) <= bch_bio_max_sectors(bio))
 163                 goto submit;
 164
 165         s = mempool_alloc(p->bio_split_hook, GFP_NOIO);
 166         closure_init(&s->cl, NULL);
 167
 168         s->bio          = bio;
 169         s->p            = p;
 170         s->bi_end_io    = bio->bi_end_io;
 171         s->bi_private   = bio->bi_private;
 172         bio_get(bio);
 173
 174         do {
 175                 n = bch_bio_split(bio, bch_bio_max_sectors(bio),
 176                                   GFP_NOIO, s->p->bio_split);
 177
 178                 n->bi_end_io    = bch_bio_submit_split_endio;
 179                 n->bi_private   = &s->cl;
 180
 181                 closure_get(&s->cl);
 182                 generic_make_request(n);
 183         } while (n != bio);
 184
 185         continue_at(&s->cl, bch_bio_submit_split_done, NULL);
 186 submit:
 187         generic_make_request(bio);
 188 }
 189
 190 /* Bios with headers */
 191
 192 void bch_bbio_free(struct bio *bio, struct cache_set *c)
 193 {
 194         struct bbio *b = container_of(bio, struct bbio, bio);
 195         mempool_free(b, c->bio_meta);
 196 }
 197
 198 struct bio *bch_bbio_alloc(struct cache_set *c)
 199 {
 200         struct bbio *b = mempool_alloc(c->bio_meta, GFP_NOIO);
 201         struct bio *bio = &b->bio;
 202
 203         bio_init(bio);
 204         bio->bi_flags           |= BIO_POOL_NONE << BIO_POOL_OFFSET;
 205         bio->bi_max_vecs         = bucket_pages(c);
 206         bio->bi_io_vec           = bio->bi_inline_vecs;
 207
 208         return bio;
 209 }
 210
 211 void __bch_submit_bbio(struct bio *bio, struct cache_set *c)
 212 {
 213         struct bbio *b = container_of(bio, struct bbio, bio);
 214
 215         bio->bi_iter.bi_sector  = PTR_OFFSET(&b->key, 0);
 216         bio->bi_bdev            = PTR_CACHE(c, &b->key, 0)->bdev;
 217
 218         b->submit_time_us = local_clock_us();
 219         closure_bio_submit(bio, bio->bi_private, PTR_CACHE(c, &b->key, 0));
 220 }
 221
 222 void bch_submit_bbio(struct bio *bio, struct cache_set *c,
 223                      struct bkey *k, unsigned ptr)
 224 {
 225         struct bbio *b = container_of(bio, struct bbio, bio);
 226         bch_bkey_copy_single_ptr(&b->key, k, ptr);
 227         __bch_submit_bbio(bio, c);
 228 }
 229
 230 /* IO errors */
 231
 232 void bch_count_io_errors(struct cache *ca, int error, const char *m)
 233 {
 234         /*
 235          * The halflife of an error is:
 236          * log2(1/2)/log2(127/128) * refresh ~= 88 * refresh
 237          */
 238
 239         if (ca->set->error_decay) {
 240                 unsigned count = atomic_inc_return(&ca->io_count);
 241
 242                 while (count > ca->set->error_decay) {
 243                         unsigned errors;
 244                         unsigned old = count;
 245                         unsigned new = count - ca->set->error_decay;
 246
 247                         /*
 248                          * First we subtract refresh from count; each time we
 249                          * succesfully do so, we rescale the errors once:
 250                          */
 251
 252                         count = atomic_cmpxchg(&ca->io_count, old, new);
 253
 254                         if (count == old) {
 255                                 count = new;
 256
 257                                 errors = atomic_read(&ca->io_errors);
 258                                 do {
 259                                         old = errors;
 260                                         new = ((uint64_t) errors * 127) / 128;
 261                                         errors = atomic_cmpxchg(&ca->io_errors,
 262                                                                 old, new);
 263                                 } while (old != errors);
 264                         }
 265                 }
 266         }
 267
 268         if (error) {
 269                 char buf[BDEVNAME_SIZE];
 270                 unsigned errors = atomic_add_return(1 << IO_ERROR_SHIFT,
 271                                                     &ca->io_errors);
 272                 errors >>= IO_ERROR_SHIFT;
 273
 274                 if (errors < ca->set->error_limit)
 275                         pr_err("%s: IO error on %s, recovering",
 276                                bdevname(ca->bdev, buf), m);
 277                 else
 278                         bch_cache_set_error(ca->set,
 279                                             "%s: too many IO errors %s",
 280                                             bdevname(ca->bdev, buf), m);
 281         }
 282 }
 283
 284 void bch_bbio_count_io_errors(struct cache_set *c, struct bio *bio,
 285                               int error, const char *m)
 286 {
 287         struct bbio *b = container_of(bio, struct bbio, bio);
 288         struct cache *ca = PTR_CACHE(c, &b->key, 0);
 289
 290         unsigned threshold = bio->bi_rw & REQ_WRITE
 291                 ? c->congested_write_threshold_us
 292                 : c->congested_read_threshold_us;
 293
 294         if (threshold) {
 295                 unsigned t = local_clock_us();
 296
 297                 int us = t - b->submit_time_us;
 298                 int congested = atomic_read(&c->congested);
 299
 300                 if (us > (int) threshold) {
 301                         int ms = us / 1024;
 302                         c->congested_last_us = t;
 303
 304                         ms = min(ms, CONGESTED_MAX + congested);
 305                         atomic_sub(ms, &c->congested);
 306                 } else if (congested < 0)
 307                         atomic_inc(&c->congested);
 308         }
 309
 310         bch_count_io_errors(ca, error, m);
 311 }
 312
 313 void bch_bbio_endio(struct cache_set *c, struct bio *bio,
 314                     int error, const char *m)
 315 {
 316         struct closure *cl = bio->bi_private;
 317
 318         bch_bbio_count_io_errors(c, bio, error, m);
 319         bio_put(bio);
 320         closure_put(cl);
 321 }