Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame^] | 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 | static void bch_bi_idx_hack_endio(struct bio *bio, int error) |
| 13 | { |
| 14 | struct bio *p = bio->bi_private; |
| 15 | |
| 16 | bio_endio(p, error); |
| 17 | bio_put(bio); |
| 18 | } |
| 19 | |
| 20 | static void bch_generic_make_request_hack(struct bio *bio) |
| 21 | { |
| 22 | if (bio->bi_idx) { |
| 23 | struct bio *clone = bio_alloc(GFP_NOIO, bio_segments(bio)); |
| 24 | |
| 25 | memcpy(clone->bi_io_vec, |
| 26 | bio_iovec(bio), |
| 27 | bio_segments(bio) * sizeof(struct bio_vec)); |
| 28 | |
| 29 | clone->bi_sector = bio->bi_sector; |
| 30 | clone->bi_bdev = bio->bi_bdev; |
| 31 | clone->bi_rw = bio->bi_rw; |
| 32 | clone->bi_vcnt = bio_segments(bio); |
| 33 | clone->bi_size = bio->bi_size; |
| 34 | |
| 35 | clone->bi_private = bio; |
| 36 | clone->bi_end_io = bch_bi_idx_hack_endio; |
| 37 | |
| 38 | bio = clone; |
| 39 | } |
| 40 | |
| 41 | generic_make_request(bio); |
| 42 | } |
| 43 | |
| 44 | /** |
| 45 | * bch_bio_split - split a bio |
| 46 | * @bio: bio to split |
| 47 | * @sectors: number of sectors to split from the front of @bio |
| 48 | * @gfp: gfp mask |
| 49 | * @bs: bio set to allocate from |
| 50 | * |
| 51 | * Allocates and returns a new bio which represents @sectors from the start of |
| 52 | * @bio, and updates @bio to represent the remaining sectors. |
| 53 | * |
| 54 | * If bio_sectors(@bio) was less than or equal to @sectors, returns @bio |
| 55 | * unchanged. |
| 56 | * |
| 57 | * The newly allocated bio will point to @bio's bi_io_vec, if the split was on a |
| 58 | * bvec boundry; it is the caller's responsibility to ensure that @bio is not |
| 59 | * freed before the split. |
| 60 | * |
| 61 | * If bch_bio_split() is running under generic_make_request(), it's not safe to |
| 62 | * allocate more than one bio from the same bio set. Therefore, if it is running |
| 63 | * under generic_make_request() it masks out __GFP_WAIT when doing the |
| 64 | * allocation. The caller must check for failure if there's any possibility of |
| 65 | * it being called from under generic_make_request(); it is then the caller's |
| 66 | * responsibility to retry from a safe context (by e.g. punting to workqueue). |
| 67 | */ |
| 68 | struct bio *bch_bio_split(struct bio *bio, int sectors, |
| 69 | gfp_t gfp, struct bio_set *bs) |
| 70 | { |
| 71 | unsigned idx = bio->bi_idx, vcnt = 0, nbytes = sectors << 9; |
| 72 | struct bio_vec *bv; |
| 73 | struct bio *ret = NULL; |
| 74 | |
| 75 | BUG_ON(sectors <= 0); |
| 76 | |
| 77 | /* |
| 78 | * If we're being called from underneath generic_make_request() and we |
| 79 | * already allocated any bios from this bio set, we risk deadlock if we |
| 80 | * use the mempool. So instead, we possibly fail and let the caller punt |
| 81 | * to workqueue or somesuch and retry in a safe context. |
| 82 | */ |
| 83 | if (current->bio_list) |
| 84 | gfp &= ~__GFP_WAIT; |
| 85 | |
| 86 | if (sectors >= bio_sectors(bio)) |
| 87 | return bio; |
| 88 | |
| 89 | if (bio->bi_rw & REQ_DISCARD) { |
| 90 | ret = bio_alloc_bioset(gfp, 1, bs); |
| 91 | idx = 0; |
| 92 | goto out; |
| 93 | } |
| 94 | |
| 95 | bio_for_each_segment(bv, bio, idx) { |
| 96 | vcnt = idx - bio->bi_idx; |
| 97 | |
| 98 | if (!nbytes) { |
| 99 | ret = bio_alloc_bioset(gfp, vcnt, bs); |
| 100 | if (!ret) |
| 101 | return NULL; |
| 102 | |
| 103 | memcpy(ret->bi_io_vec, bio_iovec(bio), |
| 104 | sizeof(struct bio_vec) * vcnt); |
| 105 | |
| 106 | break; |
| 107 | } else if (nbytes < bv->bv_len) { |
| 108 | ret = bio_alloc_bioset(gfp, ++vcnt, bs); |
| 109 | if (!ret) |
| 110 | return NULL; |
| 111 | |
| 112 | memcpy(ret->bi_io_vec, bio_iovec(bio), |
| 113 | sizeof(struct bio_vec) * vcnt); |
| 114 | |
| 115 | ret->bi_io_vec[vcnt - 1].bv_len = nbytes; |
| 116 | bv->bv_offset += nbytes; |
| 117 | bv->bv_len -= nbytes; |
| 118 | break; |
| 119 | } |
| 120 | |
| 121 | nbytes -= bv->bv_len; |
| 122 | } |
| 123 | out: |
| 124 | ret->bi_bdev = bio->bi_bdev; |
| 125 | ret->bi_sector = bio->bi_sector; |
| 126 | ret->bi_size = sectors << 9; |
| 127 | ret->bi_rw = bio->bi_rw; |
| 128 | ret->bi_vcnt = vcnt; |
| 129 | ret->bi_max_vecs = vcnt; |
| 130 | |
| 131 | bio->bi_sector += sectors; |
| 132 | bio->bi_size -= sectors << 9; |
| 133 | bio->bi_idx = idx; |
| 134 | |
| 135 | if (bio_integrity(bio)) { |
| 136 | if (bio_integrity_clone(ret, bio, gfp)) { |
| 137 | bio_put(ret); |
| 138 | return NULL; |
| 139 | } |
| 140 | |
| 141 | bio_integrity_trim(ret, 0, bio_sectors(ret)); |
| 142 | bio_integrity_trim(bio, bio_sectors(ret), bio_sectors(bio)); |
| 143 | } |
| 144 | |
| 145 | return ret; |
| 146 | } |
| 147 | |
| 148 | static unsigned bch_bio_max_sectors(struct bio *bio) |
| 149 | { |
| 150 | unsigned ret = bio_sectors(bio); |
| 151 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
| 152 | struct bio_vec *bv, *end = bio_iovec(bio) + |
| 153 | min_t(int, bio_segments(bio), queue_max_segments(q)); |
| 154 | |
| 155 | struct bvec_merge_data bvm = { |
| 156 | .bi_bdev = bio->bi_bdev, |
| 157 | .bi_sector = bio->bi_sector, |
| 158 | .bi_size = 0, |
| 159 | .bi_rw = bio->bi_rw, |
| 160 | }; |
| 161 | |
| 162 | if (bio->bi_rw & REQ_DISCARD) |
| 163 | return min(ret, q->limits.max_discard_sectors); |
| 164 | |
| 165 | if (bio_segments(bio) > queue_max_segments(q) || |
| 166 | q->merge_bvec_fn) { |
| 167 | ret = 0; |
| 168 | |
| 169 | for (bv = bio_iovec(bio); bv < end; bv++) { |
| 170 | if (q->merge_bvec_fn && |
| 171 | q->merge_bvec_fn(q, &bvm, bv) < (int) bv->bv_len) |
| 172 | break; |
| 173 | |
| 174 | ret += bv->bv_len >> 9; |
| 175 | bvm.bi_size += bv->bv_len; |
| 176 | } |
| 177 | |
| 178 | if (ret >= (BIO_MAX_PAGES * PAGE_SIZE) >> 9) |
| 179 | return (BIO_MAX_PAGES * PAGE_SIZE) >> 9; |
| 180 | } |
| 181 | |
| 182 | ret = min(ret, queue_max_sectors(q)); |
| 183 | |
| 184 | WARN_ON(!ret); |
| 185 | ret = max_t(int, ret, bio_iovec(bio)->bv_len >> 9); |
| 186 | |
| 187 | return ret; |
| 188 | } |
| 189 | |
| 190 | static void bch_bio_submit_split_done(struct closure *cl) |
| 191 | { |
| 192 | struct bio_split_hook *s = container_of(cl, struct bio_split_hook, cl); |
| 193 | |
| 194 | s->bio->bi_end_io = s->bi_end_io; |
| 195 | s->bio->bi_private = s->bi_private; |
| 196 | bio_endio(s->bio, 0); |
| 197 | |
| 198 | closure_debug_destroy(&s->cl); |
| 199 | mempool_free(s, s->p->bio_split_hook); |
| 200 | } |
| 201 | |
| 202 | static void bch_bio_submit_split_endio(struct bio *bio, int error) |
| 203 | { |
| 204 | struct closure *cl = bio->bi_private; |
| 205 | struct bio_split_hook *s = container_of(cl, struct bio_split_hook, cl); |
| 206 | |
| 207 | if (error) |
| 208 | clear_bit(BIO_UPTODATE, &s->bio->bi_flags); |
| 209 | |
| 210 | bio_put(bio); |
| 211 | closure_put(cl); |
| 212 | } |
| 213 | |
| 214 | static void __bch_bio_submit_split(struct closure *cl) |
| 215 | { |
| 216 | struct bio_split_hook *s = container_of(cl, struct bio_split_hook, cl); |
| 217 | struct bio *bio = s->bio, *n; |
| 218 | |
| 219 | do { |
| 220 | n = bch_bio_split(bio, bch_bio_max_sectors(bio), |
| 221 | GFP_NOIO, s->p->bio_split); |
| 222 | if (!n) |
| 223 | continue_at(cl, __bch_bio_submit_split, system_wq); |
| 224 | |
| 225 | n->bi_end_io = bch_bio_submit_split_endio; |
| 226 | n->bi_private = cl; |
| 227 | |
| 228 | closure_get(cl); |
| 229 | bch_generic_make_request_hack(n); |
| 230 | } while (n != bio); |
| 231 | |
| 232 | continue_at(cl, bch_bio_submit_split_done, NULL); |
| 233 | } |
| 234 | |
| 235 | void bch_generic_make_request(struct bio *bio, struct bio_split_pool *p) |
| 236 | { |
| 237 | struct bio_split_hook *s; |
| 238 | |
| 239 | if (!bio_has_data(bio) && !(bio->bi_rw & REQ_DISCARD)) |
| 240 | goto submit; |
| 241 | |
| 242 | if (bio_sectors(bio) <= bch_bio_max_sectors(bio)) |
| 243 | goto submit; |
| 244 | |
| 245 | s = mempool_alloc(p->bio_split_hook, GFP_NOIO); |
| 246 | |
| 247 | s->bio = bio; |
| 248 | s->p = p; |
| 249 | s->bi_end_io = bio->bi_end_io; |
| 250 | s->bi_private = bio->bi_private; |
| 251 | bio_get(bio); |
| 252 | |
| 253 | closure_call(&s->cl, __bch_bio_submit_split, NULL, NULL); |
| 254 | return; |
| 255 | submit: |
| 256 | bch_generic_make_request_hack(bio); |
| 257 | } |
| 258 | |
| 259 | /* Bios with headers */ |
| 260 | |
| 261 | void bch_bbio_free(struct bio *bio, struct cache_set *c) |
| 262 | { |
| 263 | struct bbio *b = container_of(bio, struct bbio, bio); |
| 264 | mempool_free(b, c->bio_meta); |
| 265 | } |
| 266 | |
| 267 | struct bio *bch_bbio_alloc(struct cache_set *c) |
| 268 | { |
| 269 | struct bbio *b = mempool_alloc(c->bio_meta, GFP_NOIO); |
| 270 | struct bio *bio = &b->bio; |
| 271 | |
| 272 | bio_init(bio); |
| 273 | bio->bi_flags |= BIO_POOL_NONE << BIO_POOL_OFFSET; |
| 274 | bio->bi_max_vecs = bucket_pages(c); |
| 275 | bio->bi_io_vec = bio->bi_inline_vecs; |
| 276 | |
| 277 | return bio; |
| 278 | } |
| 279 | |
| 280 | void __bch_submit_bbio(struct bio *bio, struct cache_set *c) |
| 281 | { |
| 282 | struct bbio *b = container_of(bio, struct bbio, bio); |
| 283 | |
| 284 | bio->bi_sector = PTR_OFFSET(&b->key, 0); |
| 285 | bio->bi_bdev = PTR_CACHE(c, &b->key, 0)->bdev; |
| 286 | |
| 287 | b->submit_time_us = local_clock_us(); |
| 288 | closure_bio_submit(bio, bio->bi_private, PTR_CACHE(c, &b->key, 0)); |
| 289 | } |
| 290 | |
| 291 | void bch_submit_bbio(struct bio *bio, struct cache_set *c, |
| 292 | struct bkey *k, unsigned ptr) |
| 293 | { |
| 294 | struct bbio *b = container_of(bio, struct bbio, bio); |
| 295 | bch_bkey_copy_single_ptr(&b->key, k, ptr); |
| 296 | __bch_submit_bbio(bio, c); |
| 297 | } |
| 298 | |
| 299 | /* IO errors */ |
| 300 | |
| 301 | void bch_count_io_errors(struct cache *ca, int error, const char *m) |
| 302 | { |
| 303 | /* |
| 304 | * The halflife of an error is: |
| 305 | * log2(1/2)/log2(127/128) * refresh ~= 88 * refresh |
| 306 | */ |
| 307 | |
| 308 | if (ca->set->error_decay) { |
| 309 | unsigned count = atomic_inc_return(&ca->io_count); |
| 310 | |
| 311 | while (count > ca->set->error_decay) { |
| 312 | unsigned errors; |
| 313 | unsigned old = count; |
| 314 | unsigned new = count - ca->set->error_decay; |
| 315 | |
| 316 | /* |
| 317 | * First we subtract refresh from count; each time we |
| 318 | * succesfully do so, we rescale the errors once: |
| 319 | */ |
| 320 | |
| 321 | count = atomic_cmpxchg(&ca->io_count, old, new); |
| 322 | |
| 323 | if (count == old) { |
| 324 | count = new; |
| 325 | |
| 326 | errors = atomic_read(&ca->io_errors); |
| 327 | do { |
| 328 | old = errors; |
| 329 | new = ((uint64_t) errors * 127) / 128; |
| 330 | errors = atomic_cmpxchg(&ca->io_errors, |
| 331 | old, new); |
| 332 | } while (old != errors); |
| 333 | } |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | if (error) { |
| 338 | char buf[BDEVNAME_SIZE]; |
| 339 | unsigned errors = atomic_add_return(1 << IO_ERROR_SHIFT, |
| 340 | &ca->io_errors); |
| 341 | errors >>= IO_ERROR_SHIFT; |
| 342 | |
| 343 | if (errors < ca->set->error_limit) |
| 344 | pr_err("%s: IO error on %s, recovering", |
| 345 | bdevname(ca->bdev, buf), m); |
| 346 | else |
| 347 | bch_cache_set_error(ca->set, |
| 348 | "%s: too many IO errors %s", |
| 349 | bdevname(ca->bdev, buf), m); |
| 350 | } |
| 351 | } |
| 352 | |
| 353 | void bch_bbio_count_io_errors(struct cache_set *c, struct bio *bio, |
| 354 | int error, const char *m) |
| 355 | { |
| 356 | struct bbio *b = container_of(bio, struct bbio, bio); |
| 357 | struct cache *ca = PTR_CACHE(c, &b->key, 0); |
| 358 | |
| 359 | unsigned threshold = bio->bi_rw & REQ_WRITE |
| 360 | ? c->congested_write_threshold_us |
| 361 | : c->congested_read_threshold_us; |
| 362 | |
| 363 | if (threshold) { |
| 364 | unsigned t = local_clock_us(); |
| 365 | |
| 366 | int us = t - b->submit_time_us; |
| 367 | int congested = atomic_read(&c->congested); |
| 368 | |
| 369 | if (us > (int) threshold) { |
| 370 | int ms = us / 1024; |
| 371 | c->congested_last_us = t; |
| 372 | |
| 373 | ms = min(ms, CONGESTED_MAX + congested); |
| 374 | atomic_sub(ms, &c->congested); |
| 375 | } else if (congested < 0) |
| 376 | atomic_inc(&c->congested); |
| 377 | } |
| 378 | |
| 379 | bch_count_io_errors(ca, error, m); |
| 380 | } |
| 381 | |
| 382 | void bch_bbio_endio(struct cache_set *c, struct bio *bio, |
| 383 | int error, const char *m) |
| 384 | { |
| 385 | struct closure *cl = bio->bi_private; |
| 386 | |
| 387 | bch_bbio_count_io_errors(c, bio, error, m); |
| 388 | bio_put(bio); |
| 389 | closure_put(cl); |
| 390 | } |