drivers/lightnvm/pblk-write.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-write.c - pblk's write path from write buffer to media
  16  */
  17
  18 #include "pblk.h"
  19
  20 static unsigned long pblk_end_w_bio(struct pblk *pblk, struct nvm_rq *rqd,
  21                                     struct pblk_c_ctx *c_ctx)
  22 {
  23         struct bio *original_bio;
  24         struct pblk_rb *rwb = &pblk->rwb;
  25         unsigned long ret;
  26         int i;
  27
  28         for (i = 0; i < c_ctx->nr_valid; i++) {
  29                 struct pblk_w_ctx *w_ctx;
  30                 int pos = c_ctx->sentry + i;
  31                 int flags;
  32
  33                 w_ctx = pblk_rb_w_ctx(rwb, pos);
  34                 flags = READ_ONCE(w_ctx->flags);
  35
  36                 if (flags & PBLK_FLUSH_ENTRY) {
  37                         flags &= ~PBLK_FLUSH_ENTRY;
  38                         /* Release flags on context. Protect from writes */
  39                         smp_store_release(&w_ctx->flags, flags);
  40
  41 #ifdef CONFIG_NVM_DEBUG
  42                         atomic_dec(&rwb->inflight_flush_point);
  43 #endif
  44                 }
  45
  46                 while ((original_bio = bio_list_pop(&w_ctx->bios)))
  47                         bio_endio(original_bio);
  48         }
  49
  50         if (c_ctx->nr_padded)
  51                 pblk_bio_free_pages(pblk, rqd->bio, c_ctx->nr_valid,
  52                                                         c_ctx->nr_padded);
  53
  54 #ifdef CONFIG_NVM_DEBUG
  55         atomic_long_add(rqd->nr_ppas, &pblk->sync_writes);
  56 #endif
  57
  58         ret = pblk_rb_sync_advance(&pblk->rwb, c_ctx->nr_valid);
  59
  60         bio_put(rqd->bio);
  61         pblk_free_rqd(pblk, rqd, PBLK_WRITE);
  62
  63         return ret;
  64 }
  65
  66 static unsigned long pblk_end_queued_w_bio(struct pblk *pblk,
  67                                            struct nvm_rq *rqd,
  68                                            struct pblk_c_ctx *c_ctx)
  69 {
  70         list_del(&c_ctx->list);
  71         return pblk_end_w_bio(pblk, rqd, c_ctx);
  72 }
  73
  74 static void pblk_complete_write(struct pblk *pblk, struct nvm_rq *rqd,
  75                                 struct pblk_c_ctx *c_ctx)
  76 {
  77         struct pblk_c_ctx *c, *r;
  78         unsigned long flags;
  79         unsigned long pos;
  80
  81 #ifdef CONFIG_NVM_DEBUG
  82         atomic_long_sub(c_ctx->nr_valid, &pblk->inflight_writes);
  83 #endif
  84
  85         pblk_up_rq(pblk, rqd->ppa_list, rqd->nr_ppas, c_ctx->lun_bitmap);
  86
  87         pos = pblk_rb_sync_init(&pblk->rwb, &flags);
  88         if (pos == c_ctx->sentry) {
  89                 pos = pblk_end_w_bio(pblk, rqd, c_ctx);
  90
  91 retry:
  92                 list_for_each_entry_safe(c, r, &pblk->compl_list, list) {
  93                         rqd = nvm_rq_from_c_ctx(c);
  94                         if (c->sentry == pos) {
  95                                 pos = pblk_end_queued_w_bio(pblk, rqd, c);
  96                                 goto retry;
  97                         }
  98                 }
  99         } else {
 100                 WARN_ON(nvm_rq_from_c_ctx(c_ctx) != rqd);
 101                 list_add_tail(&c_ctx->list, &pblk->compl_list);
 102         }
 103         pblk_rb_sync_end(&pblk->rwb, &flags);
 104 }
 105
 106 /* When a write fails, we are not sure whether the block has grown bad or a page
 107  * range is more susceptible to write errors. If a high number of pages fail, we
 108  * assume that the block is bad and we mark it accordingly. In all cases, we
 109  * remap and resubmit the failed entries as fast as possible; if a flush is
 110  * waiting on a completion, the whole stack would stall otherwise.
 111  */
 112 static void pblk_end_w_fail(struct pblk *pblk, struct nvm_rq *rqd)
 113 {
 114         void *comp_bits = &rqd->ppa_status;
 115         struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
 116         struct pblk_rec_ctx *recovery;
 117         struct ppa_addr *ppa_list = rqd->ppa_list;
 118         int nr_ppas = rqd->nr_ppas;
 119         unsigned int c_entries;
 120         int bit, ret;
 121
 122         if (unlikely(nr_ppas == 1))
 123                 ppa_list = &rqd->ppa_addr;
 124
 125         recovery = mempool_alloc(pblk->rec_pool, GFP_ATOMIC);
 126
 127         INIT_LIST_HEAD(&recovery->failed);
 128
 129         bit = -1;
 130         while ((bit = find_next_bit(comp_bits, nr_ppas, bit + 1)) < nr_ppas) {
 131                 struct pblk_rb_entry *entry;
 132                 struct ppa_addr ppa;
 133
 134                 /* Logic error */
 135                 if (bit > c_ctx->nr_valid) {
 136                         WARN_ONCE(1, "pblk: corrupted write request\n");
 137                         mempool_free(recovery, pblk->rec_pool);
 138                         goto out;
 139                 }
 140
 141                 ppa = ppa_list[bit];
 142                 entry = pblk_rb_sync_scan_entry(&pblk->rwb, &ppa);
 143                 if (!entry) {
 144                         pr_err("pblk: could not scan entry on write failure\n");
 145                         mempool_free(recovery, pblk->rec_pool);
 146                         goto out;
 147                 }
 148
 149                 /* The list is filled first and emptied afterwards. No need for
 150                  * protecting it with a lock
 151                  */
 152                 list_add_tail(&entry->index, &recovery->failed);
 153         }
 154
 155         c_entries = find_first_bit(comp_bits, nr_ppas);
 156         ret = pblk_recov_setup_rq(pblk, c_ctx, recovery, comp_bits, c_entries);
 157         if (ret) {
 158                 pr_err("pblk: could not recover from write failure\n");
 159                 mempool_free(recovery, pblk->rec_pool);
 160                 goto out;
 161         }
 162
 163         INIT_WORK(&recovery->ws_rec, pblk_submit_rec);
 164         queue_work(pblk->close_wq, &recovery->ws_rec);
 165
 166 out:
 167         pblk_complete_write(pblk, rqd, c_ctx);
 168 }
 169
 170 static void pblk_end_io_write(struct nvm_rq *rqd)
 171 {
 172         struct pblk *pblk = rqd->private;
 173         struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
 174
 175         if (rqd->error) {
 176                 pblk_log_write_err(pblk, rqd);
 177                 return pblk_end_w_fail(pblk, rqd);
 178         }
 179 #ifdef CONFIG_NVM_DEBUG
 180         else
 181                 WARN_ONCE(rqd->bio->bi_status, "pblk: corrupted write error\n");
 182 #endif
 183
 184         pblk_complete_write(pblk, rqd, c_ctx);
 185         atomic_dec(&pblk->inflight_io);
 186 }
 187
 188 static void pblk_end_io_write_meta(struct nvm_rq *rqd)
 189 {
 190         struct pblk *pblk = rqd->private;
 191         struct pblk_g_ctx *m_ctx = nvm_rq_to_pdu(rqd);
 192         struct pblk_line *line = m_ctx->private;
 193         struct pblk_emeta *emeta = line->emeta;
 194         int sync;
 195
 196         pblk_up_page(pblk, rqd->ppa_list, rqd->nr_ppas);
 197
 198         if (rqd->error) {
 199                 pblk_log_write_err(pblk, rqd);
 200                 pr_err("pblk: metadata I/O failed. Line %d\n", line->id);
 201         }
 202
 203         sync = atomic_add_return(rqd->nr_ppas, &emeta->sync);
 204         if (sync == emeta->nr_entries)
 205                 pblk_gen_run_ws(pblk, line, NULL, pblk_line_close_ws,
 206                                                 GFP_ATOMIC, pblk->close_wq);
 207
 208         pblk_free_rqd(pblk, rqd, PBLK_WRITE_INT);
 209
 210         atomic_dec(&pblk->inflight_io);
 211 }
 212
 213 static int pblk_alloc_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
 214                            unsigned int nr_secs,
 215                            nvm_end_io_fn(*end_io))
 216 {
 217         struct nvm_tgt_dev *dev = pblk->dev;
 218
 219         /* Setup write request */
 220         rqd->opcode = NVM_OP_PWRITE;
 221         rqd->nr_ppas = nr_secs;
 222         rqd->flags = pblk_set_progr_mode(pblk, PBLK_WRITE);
 223         rqd->private = pblk;
 224         rqd->end_io = end_io;
 225
 226         rqd->meta_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
 227                                                         &rqd->dma_meta_list);
 228         if (!rqd->meta_list)
 229                 return -ENOMEM;
 230
 231         rqd->ppa_list = rqd->meta_list + pblk_dma_meta_size;
 232         rqd->dma_ppa_list = rqd->dma_meta_list + pblk_dma_meta_size;
 233
 234         return 0;
 235 }
 236
 237 static int pblk_setup_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
 238                            struct ppa_addr *erase_ppa)
 239 {
 240         struct pblk_line_meta *lm = &pblk->lm;
 241         struct pblk_line *e_line = pblk_line_get_erase(pblk);
 242         struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
 243         unsigned int valid = c_ctx->nr_valid;
 244         unsigned int padded = c_ctx->nr_padded;
 245         unsigned int nr_secs = valid + padded;
 246         unsigned long *lun_bitmap;
 247         int ret;
 248
 249         lun_bitmap = kzalloc(lm->lun_bitmap_len, GFP_KERNEL);
 250         if (!lun_bitmap)
 251                 return -ENOMEM;
 252         c_ctx->lun_bitmap = lun_bitmap;
 253
 254         ret = pblk_alloc_w_rq(pblk, rqd, nr_secs, pblk_end_io_write);
 255         if (ret) {
 256                 kfree(lun_bitmap);
 257                 return ret;
 258         }
 259
 260         if (likely(!e_line || !atomic_read(&e_line->left_eblks)))
 261                 pblk_map_rq(pblk, rqd, c_ctx->sentry, lun_bitmap, valid, 0);
 262         else
 263                 pblk_map_erase_rq(pblk, rqd, c_ctx->sentry, lun_bitmap,
 264                                                         valid, erase_ppa);
 265
 266         return 0;
 267 }
 268
 269 int pblk_setup_w_rec_rq(struct pblk *pblk, struct nvm_rq *rqd,
 270                         struct pblk_c_ctx *c_ctx)
 271 {
 272         struct pblk_line_meta *lm = &pblk->lm;
 273         unsigned long *lun_bitmap;
 274         int ret;
 275
 276         lun_bitmap = kzalloc(lm->lun_bitmap_len, GFP_KERNEL);
 277         if (!lun_bitmap)
 278                 return -ENOMEM;
 279
 280         c_ctx->lun_bitmap = lun_bitmap;
 281
 282         ret = pblk_alloc_w_rq(pblk, rqd, rqd->nr_ppas, pblk_end_io_write);
 283         if (ret)
 284                 return ret;
 285
 286         pblk_map_rq(pblk, rqd, c_ctx->sentry, lun_bitmap, c_ctx->nr_valid, 0);
 287
 288         rqd->ppa_status = (u64)0;
 289         rqd->flags = pblk_set_progr_mode(pblk, PBLK_WRITE);
 290
 291         return ret;
 292 }
 293
 294 static int pblk_calc_secs_to_sync(struct pblk *pblk, unsigned int secs_avail,
 295                                   unsigned int secs_to_flush)
 296 {
 297         int secs_to_sync;
 298
 299         secs_to_sync = pblk_calc_secs(pblk, secs_avail, secs_to_flush);
 300
 301 #ifdef CONFIG_NVM_DEBUG
 302         if ((!secs_to_sync && secs_to_flush)
 303                         || (secs_to_sync < 0)
 304                         || (secs_to_sync > secs_avail && !secs_to_flush)) {
 305                 pr_err("pblk: bad sector calculation (a:%d,s:%d,f:%d)\n",
 306                                 secs_avail, secs_to_sync, secs_to_flush);
 307         }
 308 #endif
 309
 310         return secs_to_sync;
 311 }
 312
 313 int pblk_submit_meta_io(struct pblk *pblk, struct pblk_line *meta_line)
 314 {
 315         struct nvm_tgt_dev *dev = pblk->dev;
 316         struct nvm_geo *geo = &dev->geo;
 317         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
 318         struct pblk_line_meta *lm = &pblk->lm;
 319         struct pblk_emeta *emeta = meta_line->emeta;
 320         struct pblk_g_ctx *m_ctx;
 321         struct bio *bio;
 322         struct nvm_rq *rqd;
 323         void *data;
 324         u64 paddr;
 325         int rq_ppas = pblk->min_write_pgs;
 326         int id = meta_line->id;
 327         int rq_len;
 328         int i, j;
 329         int ret;
 330
 331         rqd = pblk_alloc_rqd(pblk, PBLK_WRITE_INT);
 332
 333         m_ctx = nvm_rq_to_pdu(rqd);
 334         m_ctx->private = meta_line;
 335
 336         rq_len = rq_ppas * geo->sec_size;
 337         data = ((void *)emeta->buf) + emeta->mem;
 338
 339         bio = pblk_bio_map_addr(pblk, data, rq_ppas, rq_len,
 340                                         l_mg->emeta_alloc_type, GFP_KERNEL);
 341         if (IS_ERR(bio)) {
 342                 ret = PTR_ERR(bio);
 343                 goto fail_free_rqd;
 344         }
 345         bio->bi_iter.bi_sector = 0; /* internal bio */
 346         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 347         rqd->bio = bio;
 348
 349         ret = pblk_alloc_w_rq(pblk, rqd, rq_ppas, pblk_end_io_write_meta);
 350         if (ret)
 351                 goto fail_free_bio;
 352
 353         for (i = 0; i < rqd->nr_ppas; ) {
 354                 spin_lock(&meta_line->lock);
 355                 paddr = __pblk_alloc_page(pblk, meta_line, rq_ppas);
 356                 spin_unlock(&meta_line->lock);
 357                 for (j = 0; j < rq_ppas; j++, i++, paddr++)
 358                         rqd->ppa_list[i] = addr_to_gen_ppa(pblk, paddr, id);
 359         }
 360
 361         emeta->mem += rq_len;
 362         if (emeta->mem >= lm->emeta_len[0]) {
 363                 spin_lock(&l_mg->close_lock);
 364                 list_del(&meta_line->list);
 365                 spin_unlock(&l_mg->close_lock);
 366         }
 367
 368         pblk_down_page(pblk, rqd->ppa_list, rqd->nr_ppas);
 369
 370         ret = pblk_submit_io(pblk, rqd);
 371         if (ret) {
 372                 pr_err("pblk: emeta I/O submission failed: %d\n", ret);
 373                 goto fail_rollback;
 374         }
 375
 376         return NVM_IO_OK;
 377
 378 fail_rollback:
 379         pblk_up_page(pblk, rqd->ppa_list, rqd->nr_ppas);
 380         spin_lock(&l_mg->close_lock);
 381         pblk_dealloc_page(pblk, meta_line, rq_ppas);
 382         list_add(&meta_line->list, &meta_line->list);
 383         spin_unlock(&l_mg->close_lock);
 384 fail_free_bio:
 385         bio_put(bio);
 386 fail_free_rqd:
 387         pblk_free_rqd(pblk, rqd, PBLK_WRITE_INT);
 388         return ret;
 389 }
 390
 391 static inline bool pblk_valid_meta_ppa(struct pblk *pblk,
 392                                        struct pblk_line *meta_line,
 393                                        struct nvm_rq *data_rqd)
 394 {
 395         struct nvm_tgt_dev *dev = pblk->dev;
 396         struct nvm_geo *geo = &dev->geo;
 397         struct pblk_c_ctx *data_c_ctx = nvm_rq_to_pdu(data_rqd);
 398         struct pblk_line *data_line = pblk_line_get_data(pblk);
 399         struct ppa_addr ppa, ppa_opt;
 400         u64 paddr;
 401         int pos_opt;
 402
 403         /* Schedule a metadata I/O that is half the distance from the data I/O
 404          * with regards to the number of LUNs forming the pblk instance. This
 405          * balances LUN conflicts across every I/O.
 406          *
 407          * When the LUN configuration changes (e.g., due to GC), this distance
 408          * can align, which would result on metadata and data I/Os colliding. In
 409          * this case, modify the distance to not be optimal, but move the
 410          * optimal in the right direction.
 411          */
 412         paddr = pblk_lookup_page(pblk, meta_line);
 413         ppa = addr_to_gen_ppa(pblk, paddr, 0);
 414         ppa_opt = addr_to_gen_ppa(pblk, paddr + data_line->meta_distance, 0);
 415         pos_opt = pblk_ppa_to_pos(geo, ppa_opt);
 416
 417         if (test_bit(pos_opt, data_c_ctx->lun_bitmap) ||
 418                                 test_bit(pos_opt, data_line->blk_bitmap))
 419                 return true;
 420
 421         if (unlikely(pblk_ppa_comp(ppa_opt, ppa)))
 422                 data_line->meta_distance--;
 423
 424         return false;
 425 }
 426
 427 static struct pblk_line *pblk_should_submit_meta_io(struct pblk *pblk,
 428                                                     struct nvm_rq *data_rqd)
 429 {
 430         struct pblk_line_meta *lm = &pblk->lm;
 431         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
 432         struct pblk_line *meta_line;
 433
 434         spin_lock(&l_mg->close_lock);
 435 retry:
 436         if (list_empty(&l_mg->emeta_list)) {
 437                 spin_unlock(&l_mg->close_lock);
 438                 return NULL;
 439         }
 440         meta_line = list_first_entry(&l_mg->emeta_list, struct pblk_line, list);
 441         if (meta_line->emeta->mem >= lm->emeta_len[0])
 442                 goto retry;
 443         spin_unlock(&l_mg->close_lock);
 444
 445         if (!pblk_valid_meta_ppa(pblk, meta_line, data_rqd))
 446                 return NULL;
 447
 448         return meta_line;
 449 }
 450
 451 static int pblk_submit_io_set(struct pblk *pblk, struct nvm_rq *rqd)
 452 {
 453         struct ppa_addr erase_ppa;
 454         struct pblk_line *meta_line;
 455         int err;
 456
 457         pblk_ppa_set_empty(&erase_ppa);
 458
 459         /* Assign lbas to ppas and populate request structure */
 460         err = pblk_setup_w_rq(pblk, rqd, &erase_ppa);
 461         if (err) {
 462                 pr_err("pblk: could not setup write request: %d\n", err);
 463                 return NVM_IO_ERR;
 464         }
 465
 466         meta_line = pblk_should_submit_meta_io(pblk, rqd);
 467
 468         /* Submit data write for current data line */
 469         err = pblk_submit_io(pblk, rqd);
 470         if (err) {
 471                 pr_err("pblk: data I/O submission failed: %d\n", err);
 472                 return NVM_IO_ERR;
 473         }
 474
 475         if (!pblk_ppa_empty(erase_ppa)) {
 476                 /* Submit erase for next data line */
 477                 if (pblk_blk_erase_async(pblk, erase_ppa)) {
 478                         struct pblk_line *e_line = pblk_line_get_erase(pblk);
 479                         struct nvm_tgt_dev *dev = pblk->dev;
 480                         struct nvm_geo *geo = &dev->geo;
 481                         int bit;
 482
 483                         atomic_inc(&e_line->left_eblks);
 484                         bit = pblk_ppa_to_pos(geo, erase_ppa);
 485                         WARN_ON(!test_and_clear_bit(bit, e_line->erase_bitmap));
 486                 }
 487         }
 488
 489         if (meta_line) {
 490                 /* Submit metadata write for previous data line */
 491                 err = pblk_submit_meta_io(pblk, meta_line);
 492                 if (err) {
 493                         pr_err("pblk: metadata I/O submission failed: %d", err);
 494                         return NVM_IO_ERR;
 495                 }
 496         }
 497
 498         return NVM_IO_OK;
 499 }
 500
 501 static void pblk_free_write_rqd(struct pblk *pblk, struct nvm_rq *rqd)
 502 {
 503         struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
 504         struct bio *bio = rqd->bio;
 505
 506         if (c_ctx->nr_padded)
 507                 pblk_bio_free_pages(pblk, bio, c_ctx->nr_valid,
 508                                                         c_ctx->nr_padded);
 509 }
 510
 511 static int pblk_submit_write(struct pblk *pblk)
 512 {
 513         struct bio *bio;
 514         struct nvm_rq *rqd;
 515         unsigned int secs_avail, secs_to_sync, secs_to_com;
 516         unsigned int secs_to_flush;
 517         unsigned long pos;
 518
 519         /* If there are no sectors in the cache, flushes (bios without data)
 520          * will be cleared on the cache threads
 521          */
 522         secs_avail = pblk_rb_read_count(&pblk->rwb);
 523         if (!secs_avail)
 524                 return 1;
 525
 526         secs_to_flush = pblk_rb_flush_point_count(&pblk->rwb);
 527         if (!secs_to_flush && secs_avail < pblk->min_write_pgs)
 528                 return 1;
 529
 530         secs_to_sync = pblk_calc_secs_to_sync(pblk, secs_avail, secs_to_flush);
 531         if (secs_to_sync > pblk->max_write_pgs) {
 532                 pr_err("pblk: bad buffer sync calculation\n");
 533                 return 1;
 534         }
 535
 536         secs_to_com = (secs_to_sync > secs_avail) ? secs_avail : secs_to_sync;
 537         pos = pblk_rb_read_commit(&pblk->rwb, secs_to_com);
 538
 539         bio = bio_alloc(GFP_KERNEL, secs_to_sync);
 540
 541         bio->bi_iter.bi_sector = 0; /* internal bio */
 542         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 543
 544         rqd = pblk_alloc_rqd(pblk, PBLK_WRITE);
 545         rqd->bio = bio;
 546
 547         if (pblk_rb_read_to_bio(&pblk->rwb, rqd, pos, secs_to_sync,
 548                                                                 secs_avail)) {
 549                 pr_err("pblk: corrupted write bio\n");
 550                 goto fail_put_bio;
 551         }
 552
 553         if (pblk_submit_io_set(pblk, rqd))
 554                 goto fail_free_bio;
 555
 556 #ifdef CONFIG_NVM_DEBUG
 557         atomic_long_add(secs_to_sync, &pblk->sub_writes);
 558 #endif
 559
 560         return 0;
 561
 562 fail_free_bio:
 563         pblk_free_write_rqd(pblk, rqd);
 564 fail_put_bio:
 565         bio_put(bio);
 566         pblk_free_rqd(pblk, rqd, PBLK_WRITE);
 567
 568         return 1;
 569 }
 570
 571 int pblk_write_ts(void *data)
 572 {
 573         struct pblk *pblk = data;
 574
 575         while (!kthread_should_stop()) {
 576                 if (!pblk_submit_write(pblk))
 577                         continue;
 578                 set_current_state(TASK_INTERRUPTIBLE);
 579                 io_schedule();
 580         }
 581
 582         return 0;
 583 }