drivers/lightnvm/pblk-read.c

   1 /*
   2  * Copyright (C) 2016 CNEX Labs
   3  * Initial release: Javier Gonzalez <javier@cnexlabs.com>
   4  *                  Matias Bjorling <matias@cnexlabs.com>
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License version
   8  * 2 as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License for more details.
  14  *
  15  * pblk-read.c - pblk's read path
  16  */
  17
  18 #include "pblk.h"
  19
  20 /*
  21  * There is no guarantee that the value read from cache has not been updated and
  22  * resides at another location in the cache. We guarantee though that if the
  23  * value is read from the cache, it belongs to the mapped lba. In order to
  24  * guarantee and order between writes and reads are ordered, a flush must be
  25  * issued.
  26  */
  27 static int pblk_read_from_cache(struct pblk *pblk, struct bio *bio,
  28                                 sector_t lba, struct ppa_addr ppa,
  29                                 int bio_iter)
  30 {
  31 #ifdef CONFIG_NVM_DEBUG
  32         /* Callers must ensure that the ppa points to a cache address */
  33         BUG_ON(pblk_ppa_empty(ppa));
  34         BUG_ON(!pblk_addr_in_cache(ppa));
  35 #endif
  36
  37         return pblk_rb_copy_to_bio(&pblk->rwb, bio, lba,
  38                                         pblk_addr_to_cacheline(ppa), bio_iter);
  39 }
  40
  41 static void pblk_read_ppalist_rq(struct pblk *pblk, struct nvm_rq *rqd,
  42                                  unsigned long *read_bitmap)
  43 {
  44         struct bio *bio = rqd->bio;
  45         struct ppa_addr ppas[PBLK_MAX_REQ_ADDRS];
  46         sector_t blba = pblk_get_lba(bio);
  47         int nr_secs = rqd->nr_ppas;
  48         int advanced_bio = 0;
  49         int i, j = 0;
  50
  51         /* logic error: lba out-of-bounds. Ignore read request */
  52         if (blba + nr_secs >= pblk->rl.nr_secs) {
  53                 WARN(1, "pblk: read lbas out of bounds\n");
  54                 return;
  55         }
  56
  57         pblk_lookup_l2p_seq(pblk, ppas, blba, nr_secs);
  58
  59         for (i = 0; i < nr_secs; i++) {
  60                 struct ppa_addr p = ppas[i];
  61                 sector_t lba = blba + i;
  62
  63 retry:
  64                 if (pblk_ppa_empty(p)) {
  65                         WARN_ON(test_and_set_bit(i, read_bitmap));
  66                         continue;
  67                 }
  68
  69                 /* Try to read from write buffer. The address is later checked
  70                  * on the write buffer to prevent retrieving overwritten data.
  71                  */
  72                 if (pblk_addr_in_cache(p)) {
  73                         if (!pblk_read_from_cache(pblk, bio, lba, p, i)) {
  74                                 pblk_lookup_l2p_seq(pblk, &p, lba, 1);
  75                                 goto retry;
  76                         }
  77                         WARN_ON(test_and_set_bit(i, read_bitmap));
  78                         advanced_bio = 1;
  79 #ifdef CONFIG_NVM_DEBUG
  80                         atomic_long_inc(&pblk->cache_reads);
  81 #endif
  82                 } else {
  83                         /* Read from media non-cached sectors */
  84                         rqd->ppa_list[j++] = p;
  85                 }
  86
  87                 if (advanced_bio)
  88                         bio_advance(bio, PBLK_EXPOSED_PAGE_SIZE);
  89         }
  90
  91 #ifdef CONFIG_NVM_DEBUG
  92         atomic_long_add(nr_secs, &pblk->inflight_reads);
  93 #endif
  94 }
  95
  96 static int pblk_submit_read_io(struct pblk *pblk, struct nvm_rq *rqd)
  97 {
  98         int err;
  99
 100         rqd->flags = pblk_set_read_mode(pblk);
 101
 102         err = pblk_submit_io(pblk, rqd);
 103         if (err)
 104                 return NVM_IO_ERR;
 105
 106         return NVM_IO_OK;
 107 }
 108
 109 static void pblk_end_io_read(struct nvm_rq *rqd)
 110 {
 111         struct pblk *pblk = rqd->private;
 112         struct nvm_tgt_dev *dev = pblk->dev;
 113         struct pblk_g_ctx *r_ctx = nvm_rq_to_pdu(rqd);
 114         struct bio *bio = rqd->bio;
 115
 116         if (rqd->error)
 117                 pblk_log_read_err(pblk, rqd);
 118 #ifdef CONFIG_NVM_DEBUG
 119         else
 120                 WARN_ONCE(bio->bi_status, "pblk: corrupted read error\n");
 121 #endif
 122
 123         if (rqd->nr_ppas > 1)
 124                 nvm_dev_dma_free(dev->parent, rqd->ppa_list, rqd->dma_ppa_list);
 125
 126         bio_put(bio);
 127         if (r_ctx->private) {
 128                 struct bio *orig_bio = r_ctx->private;
 129
 130 #ifdef CONFIG_NVM_DEBUG
 131                 WARN_ONCE(orig_bio->bi_status, "pblk: corrupted read bio\n");
 132 #endif
 133                 bio_endio(orig_bio);
 134                 bio_put(orig_bio);
 135         }
 136
 137 #ifdef CONFIG_NVM_DEBUG
 138         atomic_long_add(rqd->nr_ppas, &pblk->sync_reads);
 139         atomic_long_sub(rqd->nr_ppas, &pblk->inflight_reads);
 140 #endif
 141
 142         pblk_free_rqd(pblk, rqd, READ);
 143 }
 144
 145 static int pblk_fill_partial_read_bio(struct pblk *pblk, struct nvm_rq *rqd,
 146                                       unsigned int bio_init_idx,
 147                                       unsigned long *read_bitmap)
 148 {
 149         struct bio *new_bio, *bio = rqd->bio;
 150         struct bio_vec src_bv, dst_bv;
 151         void *ppa_ptr = NULL;
 152         void *src_p, *dst_p;
 153         dma_addr_t dma_ppa_list = 0;
 154         int nr_secs = rqd->nr_ppas;
 155         int nr_holes = nr_secs - bitmap_weight(read_bitmap, nr_secs);
 156         int i, ret, hole;
 157         DECLARE_COMPLETION_ONSTACK(wait);
 158
 159         new_bio = bio_alloc(GFP_KERNEL, nr_holes);
 160         if (!new_bio) {
 161                 pr_err("pblk: could not alloc read bio\n");
 162                 return NVM_IO_ERR;
 163         }
 164
 165         if (pblk_bio_add_pages(pblk, new_bio, GFP_KERNEL, nr_holes))
 166                 goto err;
 167
 168         if (nr_holes != new_bio->bi_vcnt) {
 169                 pr_err("pblk: malformed bio\n");
 170                 goto err;
 171         }
 172
 173         new_bio->bi_iter.bi_sector = 0; /* internal bio */
 174         bio_set_op_attrs(new_bio, REQ_OP_READ, 0);
 175         new_bio->bi_private = &wait;
 176         new_bio->bi_end_io = pblk_end_bio_sync;
 177
 178         rqd->bio = new_bio;
 179         rqd->nr_ppas = nr_holes;
 180         rqd->end_io = NULL;
 181
 182         if (unlikely(nr_secs > 1 && nr_holes == 1)) {
 183                 ppa_ptr = rqd->ppa_list;
 184                 dma_ppa_list = rqd->dma_ppa_list;
 185                 rqd->ppa_addr = rqd->ppa_list[0];
 186         }
 187
 188         ret = pblk_submit_read_io(pblk, rqd);
 189         if (ret) {
 190                 bio_put(rqd->bio);
 191                 pr_err("pblk: read IO submission failed\n");
 192                 goto err;
 193         }
 194
 195         if (!wait_for_completion_io_timeout(&wait,
 196                                 msecs_to_jiffies(PBLK_COMMAND_TIMEOUT_MS))) {
 197                 pr_err("pblk: partial read I/O timed out\n");
 198         }
 199
 200         if (rqd->error) {
 201                 atomic_long_inc(&pblk->read_failed);
 202 #ifdef CONFIG_NVM_DEBUG
 203                 pblk_print_failed_rqd(pblk, rqd, rqd->error);
 204 #endif
 205         }
 206
 207         if (unlikely(nr_secs > 1 && nr_holes == 1)) {
 208                 rqd->ppa_list = ppa_ptr;
 209                 rqd->dma_ppa_list = dma_ppa_list;
 210         }
 211
 212         /* Fill the holes in the original bio */
 213         i = 0;
 214         hole = find_first_zero_bit(read_bitmap, nr_secs);
 215         do {
 216                 src_bv = new_bio->bi_io_vec[i++];
 217                 dst_bv = bio->bi_io_vec[bio_init_idx + hole];
 218
 219                 src_p = kmap_atomic(src_bv.bv_page);
 220                 dst_p = kmap_atomic(dst_bv.bv_page);
 221
 222                 memcpy(dst_p + dst_bv.bv_offset,
 223                         src_p + src_bv.bv_offset,
 224                         PBLK_EXPOSED_PAGE_SIZE);
 225
 226                 kunmap_atomic(src_p);
 227                 kunmap_atomic(dst_p);
 228
 229                 mempool_free(src_bv.bv_page, pblk->page_pool);
 230
 231                 hole = find_next_zero_bit(read_bitmap, nr_secs, hole + 1);
 232         } while (hole < nr_secs);
 233
 234         bio_put(new_bio);
 235
 236         /* Complete the original bio and associated request */
 237         rqd->bio = bio;
 238         rqd->nr_ppas = nr_secs;
 239         rqd->private = pblk;
 240
 241         bio_endio(bio);
 242         pblk_end_io_read(rqd);
 243         return NVM_IO_OK;
 244
 245 err:
 246         /* Free allocated pages in new bio */
 247         pblk_bio_free_pages(pblk, bio, 0, new_bio->bi_vcnt);
 248         rqd->private = pblk;
 249         pblk_end_io_read(rqd);
 250         return NVM_IO_ERR;
 251 }
 252
 253 static void pblk_read_rq(struct pblk *pblk, struct nvm_rq *rqd,
 254                          unsigned long *read_bitmap)
 255 {
 256         struct bio *bio = rqd->bio;
 257         struct ppa_addr ppa;
 258         sector_t lba = pblk_get_lba(bio);
 259
 260         /* logic error: lba out-of-bounds. Ignore read request */
 261         if (lba >= pblk->rl.nr_secs) {
 262                 WARN(1, "pblk: read lba out of bounds\n");
 263                 return;
 264         }
 265
 266         pblk_lookup_l2p_seq(pblk, &ppa, lba, 1);
 267
 268 #ifdef CONFIG_NVM_DEBUG
 269         atomic_long_inc(&pblk->inflight_reads);
 270 #endif
 271
 272 retry:
 273         if (pblk_ppa_empty(ppa)) {
 274                 WARN_ON(test_and_set_bit(0, read_bitmap));
 275                 return;
 276         }
 277
 278         /* Try to read from write buffer. The address is later checked on the
 279          * write buffer to prevent retrieving overwritten data.
 280          */
 281         if (pblk_addr_in_cache(ppa)) {
 282                 if (!pblk_read_from_cache(pblk, bio, lba, ppa, 0)) {
 283                         pblk_lookup_l2p_seq(pblk, &ppa, lba, 1);
 284                         goto retry;
 285                 }
 286                 WARN_ON(test_and_set_bit(0, read_bitmap));
 287 #ifdef CONFIG_NVM_DEBUG
 288                         atomic_long_inc(&pblk->cache_reads);
 289 #endif
 290         } else {
 291                 rqd->ppa_addr = ppa;
 292         }
 293 }
 294
 295 int pblk_submit_read(struct pblk *pblk, struct bio *bio)
 296 {
 297         struct nvm_tgt_dev *dev = pblk->dev;
 298         unsigned int nr_secs = pblk_get_secs(bio);
 299         struct nvm_rq *rqd;
 300         unsigned long read_bitmap; /* Max 64 ppas per request */
 301         unsigned int bio_init_idx;
 302         int ret = NVM_IO_ERR;
 303
 304         if (nr_secs > PBLK_MAX_REQ_ADDRS)
 305                 return NVM_IO_ERR;
 306
 307         bitmap_zero(&read_bitmap, nr_secs);
 308
 309         rqd = pblk_alloc_rqd(pblk, READ);
 310         if (IS_ERR(rqd)) {
 311                 pr_err_ratelimited("pblk: not able to alloc rqd");
 312                 return NVM_IO_ERR;
 313         }
 314
 315         rqd->opcode = NVM_OP_PREAD;
 316         rqd->bio = bio;
 317         rqd->nr_ppas = nr_secs;
 318         rqd->private = pblk;
 319         rqd->end_io = pblk_end_io_read;
 320
 321         /* Save the index for this bio's start. This is needed in case
 322          * we need to fill a partial read.
 323          */
 324         bio_init_idx = pblk_get_bi_idx(bio);
 325
 326         if (nr_secs > 1) {
 327                 rqd->ppa_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
 328                                                 &rqd->dma_ppa_list);
 329                 if (!rqd->ppa_list) {
 330                         pr_err("pblk: not able to allocate ppa list\n");
 331                         goto fail_rqd_free;
 332                 }
 333
 334                 pblk_read_ppalist_rq(pblk, rqd, &read_bitmap);
 335         } else {
 336                 pblk_read_rq(pblk, rqd, &read_bitmap);
 337         }
 338
 339         bio_get(bio);
 340         if (bitmap_full(&read_bitmap, nr_secs)) {
 341                 bio_endio(bio);
 342                 pblk_end_io_read(rqd);
 343                 return NVM_IO_OK;
 344         }
 345
 346         /* All sectors are to be read from the device */
 347         if (bitmap_empty(&read_bitmap, rqd->nr_ppas)) {
 348                 struct bio *int_bio = NULL;
 349                 struct pblk_g_ctx *r_ctx = nvm_rq_to_pdu(rqd);
 350
 351                 /* Clone read bio to deal with read errors internally */
 352                 int_bio = bio_clone_fast(bio, GFP_KERNEL, pblk_bio_set);
 353                 if (!int_bio) {
 354                         pr_err("pblk: could not clone read bio\n");
 355                         return NVM_IO_ERR;
 356                 }
 357
 358                 rqd->bio = int_bio;
 359                 r_ctx->private = bio;
 360
 361                 ret = pblk_submit_read_io(pblk, rqd);
 362                 if (ret) {
 363                         pr_err("pblk: read IO submission failed\n");
 364                         if (int_bio)
 365                                 bio_put(int_bio);
 366                         return ret;
 367                 }
 368
 369                 return NVM_IO_OK;
 370         }
 371
 372         /* The read bio request could be partially filled by the write buffer,
 373          * but there are some holes that need to be read from the drive.
 374          */
 375         ret = pblk_fill_partial_read_bio(pblk, rqd, bio_init_idx, &read_bitmap);
 376         if (ret) {
 377                 pr_err("pblk: failed to perform partial read\n");
 378                 return ret;
 379         }
 380
 381         return NVM_IO_OK;
 382
 383 fail_rqd_free:
 384         pblk_free_rqd(pblk, rqd, READ);
 385         return ret;
 386 }
 387
 388 static int read_ppalist_rq_gc(struct pblk *pblk, struct nvm_rq *rqd,
 389                               struct pblk_line *line, u64 *lba_list,
 390                               unsigned int nr_secs)
 391 {
 392         struct ppa_addr ppas[PBLK_MAX_REQ_ADDRS];
 393         int valid_secs = 0;
 394         int i;
 395
 396         pblk_lookup_l2p_rand(pblk, ppas, lba_list, nr_secs);
 397
 398         for (i = 0; i < nr_secs; i++) {
 399                 if (pblk_addr_in_cache(ppas[i]) || ppas[i].g.blk != line->id ||
 400                                                 pblk_ppa_empty(ppas[i])) {
 401                         lba_list[i] = ADDR_EMPTY;
 402                         continue;
 403                 }
 404
 405                 rqd->ppa_list[valid_secs++] = ppas[i];
 406         }
 407
 408 #ifdef CONFIG_NVM_DEBUG
 409         atomic_long_add(valid_secs, &pblk->inflight_reads);
 410 #endif
 411         return valid_secs;
 412 }
 413
 414 static int read_rq_gc(struct pblk *pblk, struct nvm_rq *rqd,
 415                       struct pblk_line *line, sector_t lba)
 416 {
 417         struct ppa_addr ppa;
 418         int valid_secs = 0;
 419
 420         if (lba == ADDR_EMPTY)
 421                 goto out;
 422
 423         /* logic error: lba out-of-bounds */
 424         if (lba >= pblk->rl.nr_secs) {
 425                 WARN(1, "pblk: read lba out of bounds\n");
 426                 goto out;
 427         }
 428
 429         spin_lock(&pblk->trans_lock);
 430         ppa = pblk_trans_map_get(pblk, lba);
 431         spin_unlock(&pblk->trans_lock);
 432
 433         /* Ignore updated values until the moment */
 434         if (pblk_addr_in_cache(ppa) || ppa.g.blk != line->id ||
 435                                                         pblk_ppa_empty(ppa))
 436                 goto out;
 437
 438         rqd->ppa_addr = ppa;
 439         valid_secs = 1;
 440
 441 #ifdef CONFIG_NVM_DEBUG
 442         atomic_long_inc(&pblk->inflight_reads);
 443 #endif
 444
 445 out:
 446         return valid_secs;
 447 }
 448
 449 int pblk_submit_read_gc(struct pblk *pblk, u64 *lba_list, void *data,
 450                         unsigned int nr_secs, unsigned int *secs_to_gc,
 451                         struct pblk_line *line)
 452 {
 453         struct nvm_tgt_dev *dev = pblk->dev;
 454         struct nvm_geo *geo = &dev->geo;
 455         struct request_queue *q = dev->q;
 456         struct bio *bio;
 457         struct nvm_rq rqd;
 458         int ret, data_len;
 459         DECLARE_COMPLETION_ONSTACK(wait);
 460
 461         memset(&rqd, 0, sizeof(struct nvm_rq));
 462
 463         if (nr_secs > 1) {
 464                 rqd.ppa_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
 465                                                         &rqd.dma_ppa_list);
 466                 if (!rqd.ppa_list)
 467                         return NVM_IO_ERR;
 468
 469                 *secs_to_gc = read_ppalist_rq_gc(pblk, &rqd, line, lba_list,
 470                                                                 nr_secs);
 471                 if (*secs_to_gc == 1) {
 472                         struct ppa_addr ppa;
 473
 474                         ppa = rqd.ppa_list[0];
 475                         nvm_dev_dma_free(dev->parent, rqd.ppa_list,
 476                                                         rqd.dma_ppa_list);
 477                         rqd.ppa_addr = ppa;
 478                 }
 479         } else {
 480                 *secs_to_gc = read_rq_gc(pblk, &rqd, line, lba_list[0]);
 481         }
 482
 483         if (!(*secs_to_gc))
 484                 goto out;
 485
 486         data_len = (*secs_to_gc) * geo->sec_size;
 487         bio = bio_map_kern(q, data, data_len, GFP_KERNEL);
 488         if (IS_ERR(bio)) {
 489                 pr_err("pblk: could not allocate GC bio (%lu)\n", PTR_ERR(bio));
 490                 goto err_free_dma;
 491         }
 492
 493         bio->bi_iter.bi_sector = 0; /* internal bio */
 494         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 495
 496         rqd.opcode = NVM_OP_PREAD;
 497         rqd.end_io = pblk_end_io_sync;
 498         rqd.private = &wait;
 499         rqd.nr_ppas = *secs_to_gc;
 500         rqd.bio = bio;
 501
 502         ret = pblk_submit_read_io(pblk, &rqd);
 503         if (ret) {
 504                 bio_endio(bio);
 505                 pr_err("pblk: GC read request failed\n");
 506                 goto err_free_dma;
 507         }
 508
 509         if (!wait_for_completion_io_timeout(&wait,
 510                                 msecs_to_jiffies(PBLK_COMMAND_TIMEOUT_MS))) {
 511                 pr_err("pblk: GC read I/O timed out\n");
 512         }
 513
 514         if (rqd.error) {
 515                 atomic_long_inc(&pblk->read_failed_gc);
 516 #ifdef CONFIG_NVM_DEBUG
 517                 pblk_print_failed_rqd(pblk, &rqd, rqd.error);
 518 #endif
 519         }
 520
 521 #ifdef CONFIG_NVM_DEBUG
 522         atomic_long_add(*secs_to_gc, &pblk->sync_reads);
 523         atomic_long_add(*secs_to_gc, &pblk->recov_gc_reads);
 524         atomic_long_sub(*secs_to_gc, &pblk->inflight_reads);
 525 #endif
 526
 527 out:
 528         if (rqd.nr_ppas > 1)
 529                 nvm_dev_dma_free(dev->parent, rqd.ppa_list, rqd.dma_ppa_list);
 530         return NVM_IO_OK;
 531
 532 err_free_dma:
 533         if (rqd.nr_ppas > 1)
 534                 nvm_dev_dma_free(dev->parent, rqd.ppa_list, rqd.dma_ppa_list);
 535         return NVM_IO_ERR;
 536 }