drivers/lightnvm/pblk-read.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2016 CNEX Labs
   4  * Initial release: Javier Gonzalez <javier@cnexlabs.com>
   5  *                  Matias Bjorling <matias@cnexlabs.com>
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License version
   9  * 2 as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope that it will be useful, but
  12  * WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License for more details.
  15  *
  16  * pblk-read.c - pblk's read path
  17  */
  18
  19 #include "pblk.h"
  20
  21 /*
  22  * There is no guarantee that the value read from cache has not been updated and
  23  * resides at another location in the cache. We guarantee though that if the
  24  * value is read from the cache, it belongs to the mapped lba. In order to
  25  * guarantee and order between writes and reads are ordered, a flush must be
  26  * issued.
  27  */
  28 static int pblk_read_from_cache(struct pblk *pblk, struct bio *bio,
  29                                 sector_t lba, struct ppa_addr ppa)
  30 {
  31 #ifdef CONFIG_NVM_PBLK_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, ppa);
  38 }
  39
  40 static int pblk_read_ppalist_rq(struct pblk *pblk, struct nvm_rq *rqd,
  41                                  struct bio *bio, sector_t blba,
  42                                  bool *from_cache)
  43 {
  44         void *meta_list = rqd->meta_list;
  45         int nr_secs, i;
  46
  47 retry:
  48         nr_secs = pblk_lookup_l2p_seq(pblk, rqd->ppa_list, blba, rqd->nr_ppas,
  49                                         from_cache);
  50
  51         if (!*from_cache)
  52                 goto end;
  53
  54         for (i = 0; i < nr_secs; i++) {
  55                 struct pblk_sec_meta *meta = pblk_get_meta(pblk, meta_list, i);
  56                 sector_t lba = blba + i;
  57
  58                 if (pblk_ppa_empty(rqd->ppa_list[i])) {
  59                         __le64 addr_empty = cpu_to_le64(ADDR_EMPTY);
  60
  61                         meta->lba = addr_empty;
  62                 } else if (pblk_addr_in_cache(rqd->ppa_list[i])) {
  63                         /*
  64                          * Try to read from write buffer. The address is later
  65                          * checked on the write buffer to prevent retrieving
  66                          * overwritten data.
  67                          */
  68                         if (!pblk_read_from_cache(pblk, bio, lba,
  69                                                         rqd->ppa_list[i])) {
  70                                 if (i == 0) {
  71                                         /*
  72                                          * We didn't call with bio_advance()
  73                                          * yet, so we can just retry.
  74                                          */
  75                                         goto retry;
  76                                 } else {
  77                                         /*
  78                                          * We already call bio_advance()
  79                                          * so we cannot retry and we need
  80                                          * to quit that function in order
  81                                          * to allow caller to handle the bio
  82                                          * splitting in the current sector
  83                                          * position.
  84                                          */
  85                                         nr_secs = i;
  86                                         goto end;
  87                                 }
  88                         }
  89                         meta->lba = cpu_to_le64(lba);
  90 #ifdef CONFIG_NVM_PBLK_DEBUG
  91                         atomic_long_inc(&pblk->cache_reads);
  92 #endif
  93                 }
  94                 bio_advance(bio, PBLK_EXPOSED_PAGE_SIZE);
  95         }
  96
  97 end:
  98         if (pblk_io_aligned(pblk, nr_secs))
  99                 rqd->is_seq = 1;
 100
 101 #ifdef CONFIG_NVM_PBLK_DEBUG
 102         atomic_long_add(nr_secs, &pblk->inflight_reads);
 103 #endif
 104
 105         return nr_secs;
 106 }
 107
 108
 109 static void pblk_read_check_seq(struct pblk *pblk, struct nvm_rq *rqd,
 110                                 sector_t blba)
 111 {
 112         void *meta_list = rqd->meta_list;
 113         int nr_lbas = rqd->nr_ppas;
 114         int i;
 115
 116         if (!pblk_is_oob_meta_supported(pblk))
 117                 return;
 118
 119         for (i = 0; i < nr_lbas; i++) {
 120                 struct pblk_sec_meta *meta = pblk_get_meta(pblk, meta_list, i);
 121                 u64 lba = le64_to_cpu(meta->lba);
 122
 123                 if (lba == ADDR_EMPTY)
 124                         continue;
 125
 126                 if (lba != blba + i) {
 127 #ifdef CONFIG_NVM_PBLK_DEBUG
 128                         struct ppa_addr *ppa_list = nvm_rq_to_ppa_list(rqd);
 129
 130                         print_ppa(pblk, &ppa_list[i], "seq", i);
 131 #endif
 132                         pblk_err(pblk, "corrupted read LBA (%llu/%llu)\n",
 133                                                         lba, (u64)blba + i);
 134                         WARN_ON(1);
 135                 }
 136         }
 137 }
 138
 139 /*
 140  * There can be holes in the lba list.
 141  */
 142 static void pblk_read_check_rand(struct pblk *pblk, struct nvm_rq *rqd,
 143                                  u64 *lba_list, int nr_lbas)
 144 {
 145         void *meta_lba_list = rqd->meta_list;
 146         int i, j;
 147
 148         if (!pblk_is_oob_meta_supported(pblk))
 149                 return;
 150
 151         for (i = 0, j = 0; i < nr_lbas; i++) {
 152                 struct pblk_sec_meta *meta = pblk_get_meta(pblk,
 153                                                            meta_lba_list, j);
 154                 u64 lba = lba_list[i];
 155                 u64 meta_lba;
 156
 157                 if (lba == ADDR_EMPTY)
 158                         continue;
 159
 160                 meta_lba = le64_to_cpu(meta->lba);
 161
 162                 if (lba != meta_lba) {
 163 #ifdef CONFIG_NVM_PBLK_DEBUG
 164                         struct ppa_addr *ppa_list = nvm_rq_to_ppa_list(rqd);
 165
 166                         print_ppa(pblk, &ppa_list[j], "rnd", j);
 167 #endif
 168                         pblk_err(pblk, "corrupted read LBA (%llu/%llu)\n",
 169                                                         meta_lba, lba);
 170                         WARN_ON(1);
 171                 }
 172
 173                 j++;
 174         }
 175
 176         WARN_ONCE(j != rqd->nr_ppas, "pblk: corrupted random request\n");
 177 }
 178
 179 static void pblk_end_user_read(struct bio *bio, int error)
 180 {
 181         if (error && error != NVM_RSP_WARN_HIGHECC)
 182                 bio_io_error(bio);
 183         else
 184                 bio_endio(bio);
 185 }
 186
 187 static void __pblk_end_io_read(struct pblk *pblk, struct nvm_rq *rqd,
 188                                bool put_line)
 189 {
 190         struct pblk_g_ctx *r_ctx = nvm_rq_to_pdu(rqd);
 191         struct bio *int_bio = rqd->bio;
 192         unsigned long start_time = r_ctx->start_time;
 193
 194         bio_end_io_acct(int_bio, start_time);
 195
 196         if (rqd->error)
 197                 pblk_log_read_err(pblk, rqd);
 198
 199         pblk_read_check_seq(pblk, rqd, r_ctx->lba);
 200         bio_put(int_bio);
 201
 202         if (put_line)
 203                 pblk_rq_to_line_put(pblk, rqd);
 204
 205 #ifdef CONFIG_NVM_PBLK_DEBUG
 206         atomic_long_add(rqd->nr_ppas, &pblk->sync_reads);
 207         atomic_long_sub(rqd->nr_ppas, &pblk->inflight_reads);
 208 #endif
 209
 210         pblk_free_rqd(pblk, rqd, PBLK_READ);
 211         atomic_dec(&pblk->inflight_io);
 212 }
 213
 214 static void pblk_end_io_read(struct nvm_rq *rqd)
 215 {
 216         struct pblk *pblk = rqd->private;
 217         struct pblk_g_ctx *r_ctx = nvm_rq_to_pdu(rqd);
 218         struct bio *bio = (struct bio *)r_ctx->private;
 219
 220         pblk_end_user_read(bio, rqd->error);
 221         __pblk_end_io_read(pblk, rqd, true);
 222 }
 223
 224 static void pblk_read_rq(struct pblk *pblk, struct nvm_rq *rqd, struct bio *bio,
 225                          sector_t lba, bool *from_cache)
 226 {
 227         struct pblk_sec_meta *meta = pblk_get_meta(pblk, rqd->meta_list, 0);
 228         struct ppa_addr ppa;
 229
 230         pblk_lookup_l2p_seq(pblk, &ppa, lba, 1, from_cache);
 231
 232 #ifdef CONFIG_NVM_PBLK_DEBUG
 233         atomic_long_inc(&pblk->inflight_reads);
 234 #endif
 235
 236 retry:
 237         if (pblk_ppa_empty(ppa)) {
 238                 __le64 addr_empty = cpu_to_le64(ADDR_EMPTY);
 239
 240                 meta->lba = addr_empty;
 241                 return;
 242         }
 243
 244         /* Try to read from write buffer. The address is later checked on the
 245          * write buffer to prevent retrieving overwritten data.
 246          */
 247         if (pblk_addr_in_cache(ppa)) {
 248                 if (!pblk_read_from_cache(pblk, bio, lba, ppa)) {
 249                         pblk_lookup_l2p_seq(pblk, &ppa, lba, 1, from_cache);
 250                         goto retry;
 251                 }
 252
 253                 meta->lba = cpu_to_le64(lba);
 254
 255 #ifdef CONFIG_NVM_PBLK_DEBUG
 256                 atomic_long_inc(&pblk->cache_reads);
 257 #endif
 258         } else {
 259                 rqd->ppa_addr = ppa;
 260         }
 261 }
 262
 263 void pblk_submit_read(struct pblk *pblk, struct bio *bio)
 264 {
 265         sector_t blba = pblk_get_lba(bio);
 266         unsigned int nr_secs = pblk_get_secs(bio);
 267         bool from_cache;
 268         struct pblk_g_ctx *r_ctx;
 269         struct nvm_rq *rqd;
 270         struct bio *int_bio, *split_bio;
 271         unsigned long start_time;
 272
 273         start_time = bio_start_io_acct(bio);
 274
 275         rqd = pblk_alloc_rqd(pblk, PBLK_READ);
 276
 277         rqd->opcode = NVM_OP_PREAD;
 278         rqd->nr_ppas = nr_secs;
 279         rqd->private = pblk;
 280         rqd->end_io = pblk_end_io_read;
 281
 282         r_ctx = nvm_rq_to_pdu(rqd);
 283         r_ctx->start_time = start_time;
 284         r_ctx->lba = blba;
 285
 286         if (pblk_alloc_rqd_meta(pblk, rqd)) {
 287                 bio_io_error(bio);
 288                 pblk_free_rqd(pblk, rqd, PBLK_READ);
 289                 return;
 290         }
 291
 292         /* Clone read bio to deal internally with:
 293          * -read errors when reading from drive
 294          * -bio_advance() calls during cache reads
 295          */
 296         int_bio = bio_clone_fast(bio, GFP_KERNEL, &pblk_bio_set);
 297
 298         if (nr_secs > 1)
 299                 nr_secs = pblk_read_ppalist_rq(pblk, rqd, int_bio, blba,
 300                                                 &from_cache);
 301         else
 302                 pblk_read_rq(pblk, rqd, int_bio, blba, &from_cache);
 303
 304 split_retry:
 305         r_ctx->private = bio; /* original bio */
 306         rqd->bio = int_bio; /* internal bio */
 307
 308         if (from_cache && nr_secs == rqd->nr_ppas) {
 309                 /* All data was read from cache, we can complete the IO. */
 310                 pblk_end_user_read(bio, 0);
 311                 atomic_inc(&pblk->inflight_io);
 312                 __pblk_end_io_read(pblk, rqd, false);
 313         } else if (nr_secs != rqd->nr_ppas) {
 314                 /* The read bio request could be partially filled by the write
 315                  * buffer, but there are some holes that need to be read from
 316                  * the drive. In order to handle this, we will use block layer
 317                  * mechanism to split this request in to smaller ones and make
 318                  * a chain of it.
 319                  */
 320                 split_bio = bio_split(bio, nr_secs * NR_PHY_IN_LOG, GFP_KERNEL,
 321                                         &pblk_bio_set);
 322                 bio_chain(split_bio, bio);
 323                 submit_bio_noacct(bio);
 324
 325                 /* New bio contains first N sectors of the previous one, so
 326                  * we can continue to use existing rqd, but we need to shrink
 327                  * the number of PPAs in it. New bio is also guaranteed that
 328                  * it contains only either data from cache or from drive, newer
 329                  * mix of them.
 330                  */
 331                 bio = split_bio;
 332                 rqd->nr_ppas = nr_secs;
 333                 if (rqd->nr_ppas == 1)
 334                         rqd->ppa_addr = rqd->ppa_list[0];
 335
 336                 /* Recreate int_bio - existing might have some needed internal
 337                  * fields modified already.
 338                  */
 339                 bio_put(int_bio);
 340                 int_bio = bio_clone_fast(bio, GFP_KERNEL, &pblk_bio_set);
 341                 goto split_retry;
 342         } else if (pblk_submit_io(pblk, rqd, NULL)) {
 343                 /* Submitting IO to drive failed, let's report an error */
 344                 rqd->error = -ENODEV;
 345                 pblk_end_io_read(rqd);
 346         }
 347 }
 348
 349 static int read_ppalist_rq_gc(struct pblk *pblk, struct nvm_rq *rqd,
 350                               struct pblk_line *line, u64 *lba_list,
 351                               u64 *paddr_list_gc, unsigned int nr_secs)
 352 {
 353         struct ppa_addr ppa_list_l2p[NVM_MAX_VLBA];
 354         struct ppa_addr ppa_gc;
 355         int valid_secs = 0;
 356         int i;
 357
 358         pblk_lookup_l2p_rand(pblk, ppa_list_l2p, lba_list, nr_secs);
 359
 360         for (i = 0; i < nr_secs; i++) {
 361                 if (lba_list[i] == ADDR_EMPTY)
 362                         continue;
 363
 364                 ppa_gc = addr_to_gen_ppa(pblk, paddr_list_gc[i], line->id);
 365                 if (!pblk_ppa_comp(ppa_list_l2p[i], ppa_gc)) {
 366                         paddr_list_gc[i] = lba_list[i] = ADDR_EMPTY;
 367                         continue;
 368                 }
 369
 370                 rqd->ppa_list[valid_secs++] = ppa_list_l2p[i];
 371         }
 372
 373 #ifdef CONFIG_NVM_PBLK_DEBUG
 374         atomic_long_add(valid_secs, &pblk->inflight_reads);
 375 #endif
 376
 377         return valid_secs;
 378 }
 379
 380 static int read_rq_gc(struct pblk *pblk, struct nvm_rq *rqd,
 381                       struct pblk_line *line, sector_t lba,
 382                       u64 paddr_gc)
 383 {
 384         struct ppa_addr ppa_l2p, ppa_gc;
 385         int valid_secs = 0;
 386
 387         if (lba == ADDR_EMPTY)
 388                 goto out;
 389
 390         /* logic error: lba out-of-bounds */
 391         if (lba >= pblk->capacity) {
 392                 WARN(1, "pblk: read lba out of bounds\n");
 393                 goto out;
 394         }
 395
 396         spin_lock(&pblk->trans_lock);
 397         ppa_l2p = pblk_trans_map_get(pblk, lba);
 398         spin_unlock(&pblk->trans_lock);
 399
 400         ppa_gc = addr_to_gen_ppa(pblk, paddr_gc, line->id);
 401         if (!pblk_ppa_comp(ppa_l2p, ppa_gc))
 402                 goto out;
 403
 404         rqd->ppa_addr = ppa_l2p;
 405         valid_secs = 1;
 406
 407 #ifdef CONFIG_NVM_PBLK_DEBUG
 408         atomic_long_inc(&pblk->inflight_reads);
 409 #endif
 410
 411 out:
 412         return valid_secs;
 413 }
 414
 415 int pblk_submit_read_gc(struct pblk *pblk, struct pblk_gc_rq *gc_rq)
 416 {
 417         struct nvm_rq rqd;
 418         int ret = NVM_IO_OK;
 419
 420         memset(&rqd, 0, sizeof(struct nvm_rq));
 421
 422         ret = pblk_alloc_rqd_meta(pblk, &rqd);
 423         if (ret)
 424                 return ret;
 425
 426         if (gc_rq->nr_secs > 1) {
 427                 gc_rq->secs_to_gc = read_ppalist_rq_gc(pblk, &rqd, gc_rq->line,
 428                                                         gc_rq->lba_list,
 429                                                         gc_rq->paddr_list,
 430                                                         gc_rq->nr_secs);
 431                 if (gc_rq->secs_to_gc == 1)
 432                         rqd.ppa_addr = rqd.ppa_list[0];
 433         } else {
 434                 gc_rq->secs_to_gc = read_rq_gc(pblk, &rqd, gc_rq->line,
 435                                                         gc_rq->lba_list[0],
 436                                                         gc_rq->paddr_list[0]);
 437         }
 438
 439         if (!(gc_rq->secs_to_gc))
 440                 goto out;
 441
 442         rqd.opcode = NVM_OP_PREAD;
 443         rqd.nr_ppas = gc_rq->secs_to_gc;
 444
 445         if (pblk_submit_io_sync(pblk, &rqd, gc_rq->data)) {
 446                 ret = -EIO;
 447                 goto err_free_dma;
 448         }
 449
 450         pblk_read_check_rand(pblk, &rqd, gc_rq->lba_list, gc_rq->nr_secs);
 451
 452         atomic_dec(&pblk->inflight_io);
 453
 454         if (rqd.error) {
 455                 atomic_long_inc(&pblk->read_failed_gc);
 456 #ifdef CONFIG_NVM_PBLK_DEBUG
 457                 pblk_print_failed_rqd(pblk, &rqd, rqd.error);
 458 #endif
 459         }
 460
 461 #ifdef CONFIG_NVM_PBLK_DEBUG
 462         atomic_long_add(gc_rq->secs_to_gc, &pblk->sync_reads);
 463         atomic_long_add(gc_rq->secs_to_gc, &pblk->recov_gc_reads);
 464         atomic_long_sub(gc_rq->secs_to_gc, &pblk->inflight_reads);
 465 #endif
 466
 467 out:
 468         pblk_free_rqd_meta(pblk, &rqd);
 469         return ret;
 470
 471 err_free_dma:
 472         pblk_free_rqd_meta(pblk, &rqd);
 473         return ret;
 474 }