drivers/lightnvm/pblk-core.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-core.c - pblk's core functionality
  16  *
  17  */
  18
  19 #include "pblk.h"
  20
  21 static void pblk_mark_bb(struct pblk *pblk, struct pblk_line *line,
  22                          struct ppa_addr *ppa)
  23 {
  24         struct nvm_tgt_dev *dev = pblk->dev;
  25         struct nvm_geo *geo = &dev->geo;
  26         int pos = pblk_dev_ppa_to_pos(geo, *ppa);
  27
  28         pr_debug("pblk: erase failed: line:%d, pos:%d\n", line->id, pos);
  29         atomic_long_inc(&pblk->erase_failed);
  30
  31         atomic_dec(&line->blk_in_line);
  32         if (test_and_set_bit(pos, line->blk_bitmap))
  33                 pr_err("pblk: attempted to erase bb: line:%d, pos:%d\n",
  34                                                         line->id, pos);
  35
  36         pblk_line_run_ws(pblk, NULL, ppa, pblk_line_mark_bb, pblk->bb_wq);
  37 }
  38
  39 static void __pblk_end_io_erase(struct pblk *pblk, struct nvm_rq *rqd)
  40 {
  41         struct pblk_line *line;
  42
  43         line = &pblk->lines[pblk_dev_ppa_to_line(rqd->ppa_addr)];
  44         atomic_dec(&line->left_seblks);
  45
  46         if (rqd->error) {
  47                 struct ppa_addr *ppa;
  48
  49                 ppa = kmalloc(sizeof(struct ppa_addr), GFP_ATOMIC);
  50                 if (!ppa)
  51                         return;
  52
  53                 *ppa = rqd->ppa_addr;
  54                 pblk_mark_bb(pblk, line, ppa);
  55         }
  56
  57         atomic_dec(&pblk->inflight_io);
  58 }
  59
  60 /* Erase completion assumes that only one block is erased at the time */
  61 static void pblk_end_io_erase(struct nvm_rq *rqd)
  62 {
  63         struct pblk *pblk = rqd->private;
  64
  65         __pblk_end_io_erase(pblk, rqd);
  66         mempool_free(rqd, pblk->g_rq_pool);
  67 }
  68
  69 void __pblk_map_invalidate(struct pblk *pblk, struct pblk_line *line,
  70                            u64 paddr)
  71 {
  72         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
  73         struct list_head *move_list = NULL;
  74
  75         /* Lines being reclaimed (GC'ed) cannot be invalidated. Before the L2P
  76          * table is modified with reclaimed sectors, a check is done to endure
  77          * that newer updates are not overwritten.
  78          */
  79         spin_lock(&line->lock);
  80         if (line->state == PBLK_LINESTATE_GC ||
  81                                         line->state == PBLK_LINESTATE_FREE) {
  82                 spin_unlock(&line->lock);
  83                 return;
  84         }
  85
  86         if (test_and_set_bit(paddr, line->invalid_bitmap)) {
  87                 WARN_ONCE(1, "pblk: double invalidate\n");
  88                 spin_unlock(&line->lock);
  89                 return;
  90         }
  91         le32_add_cpu(line->vsc, -1);
  92
  93         if (line->state == PBLK_LINESTATE_CLOSED)
  94                 move_list = pblk_line_gc_list(pblk, line);
  95         spin_unlock(&line->lock);
  96
  97         if (move_list) {
  98                 spin_lock(&l_mg->gc_lock);
  99                 spin_lock(&line->lock);
 100                 /* Prevent moving a line that has just been chosen for GC */
 101                 if (line->state == PBLK_LINESTATE_GC ||
 102                                         line->state == PBLK_LINESTATE_FREE) {
 103                         spin_unlock(&line->lock);
 104                         spin_unlock(&l_mg->gc_lock);
 105                         return;
 106                 }
 107                 spin_unlock(&line->lock);
 108
 109                 list_move_tail(&line->list, move_list);
 110                 spin_unlock(&l_mg->gc_lock);
 111         }
 112 }
 113
 114 void pblk_map_invalidate(struct pblk *pblk, struct ppa_addr ppa)
 115 {
 116         struct pblk_line *line;
 117         u64 paddr;
 118         int line_id;
 119
 120 #ifdef CONFIG_NVM_DEBUG
 121         /* Callers must ensure that the ppa points to a device address */
 122         BUG_ON(pblk_addr_in_cache(ppa));
 123         BUG_ON(pblk_ppa_empty(ppa));
 124 #endif
 125
 126         line_id = pblk_tgt_ppa_to_line(ppa);
 127         line = &pblk->lines[line_id];
 128         paddr = pblk_dev_ppa_to_line_addr(pblk, ppa);
 129
 130         __pblk_map_invalidate(pblk, line, paddr);
 131 }
 132
 133 static void pblk_invalidate_range(struct pblk *pblk, sector_t slba,
 134                                   unsigned int nr_secs)
 135 {
 136         sector_t lba;
 137
 138         spin_lock(&pblk->trans_lock);
 139         for (lba = slba; lba < slba + nr_secs; lba++) {
 140                 struct ppa_addr ppa;
 141
 142                 ppa = pblk_trans_map_get(pblk, lba);
 143
 144                 if (!pblk_addr_in_cache(ppa) && !pblk_ppa_empty(ppa))
 145                         pblk_map_invalidate(pblk, ppa);
 146
 147                 pblk_ppa_set_empty(&ppa);
 148                 pblk_trans_map_set(pblk, lba, ppa);
 149         }
 150         spin_unlock(&pblk->trans_lock);
 151 }
 152
 153 struct nvm_rq *pblk_alloc_rqd(struct pblk *pblk, int rw)
 154 {
 155         mempool_t *pool;
 156         struct nvm_rq *rqd;
 157         int rq_size;
 158
 159         if (rw == WRITE) {
 160                 pool = pblk->w_rq_pool;
 161                 rq_size = pblk_w_rq_size;
 162         } else {
 163                 pool = pblk->g_rq_pool;
 164                 rq_size = pblk_g_rq_size;
 165         }
 166
 167         rqd = mempool_alloc(pool, GFP_KERNEL);
 168         memset(rqd, 0, rq_size);
 169
 170         return rqd;
 171 }
 172
 173 void pblk_free_rqd(struct pblk *pblk, struct nvm_rq *rqd, int rw)
 174 {
 175         mempool_t *pool;
 176
 177         if (rw == WRITE)
 178                 pool = pblk->w_rq_pool;
 179         else
 180                 pool = pblk->g_rq_pool;
 181
 182         mempool_free(rqd, pool);
 183 }
 184
 185 void pblk_bio_free_pages(struct pblk *pblk, struct bio *bio, int off,
 186                          int nr_pages)
 187 {
 188         struct bio_vec bv;
 189         int i;
 190
 191         WARN_ON(off + nr_pages != bio->bi_vcnt);
 192
 193         bio_advance(bio, off * PBLK_EXPOSED_PAGE_SIZE);
 194         for (i = off; i < nr_pages + off; i++) {
 195                 bv = bio->bi_io_vec[i];
 196                 mempool_free(bv.bv_page, pblk->page_pool);
 197         }
 198 }
 199
 200 int pblk_bio_add_pages(struct pblk *pblk, struct bio *bio, gfp_t flags,
 201                        int nr_pages)
 202 {
 203         struct request_queue *q = pblk->dev->q;
 204         struct page *page;
 205         int i, ret;
 206
 207         for (i = 0; i < nr_pages; i++) {
 208                 page = mempool_alloc(pblk->page_pool, flags);
 209                 if (!page)
 210                         goto err;
 211
 212                 ret = bio_add_pc_page(q, bio, page, PBLK_EXPOSED_PAGE_SIZE, 0);
 213                 if (ret != PBLK_EXPOSED_PAGE_SIZE) {
 214                         pr_err("pblk: could not add page to bio\n");
 215                         mempool_free(page, pblk->page_pool);
 216                         goto err;
 217                 }
 218         }
 219
 220         return 0;
 221 err:
 222         pblk_bio_free_pages(pblk, bio, 0, i - 1);
 223         return -1;
 224 }
 225
 226 static void pblk_write_kick(struct pblk *pblk)
 227 {
 228         wake_up_process(pblk->writer_ts);
 229         mod_timer(&pblk->wtimer, jiffies + msecs_to_jiffies(1000));
 230 }
 231
 232 void pblk_write_timer_fn(unsigned long data)
 233 {
 234         struct pblk *pblk = (struct pblk *)data;
 235
 236         /* kick the write thread every tick to flush outstanding data */
 237         pblk_write_kick(pblk);
 238 }
 239
 240 void pblk_write_should_kick(struct pblk *pblk)
 241 {
 242         unsigned int secs_avail = pblk_rb_read_count(&pblk->rwb);
 243
 244         if (secs_avail >= pblk->min_write_pgs)
 245                 pblk_write_kick(pblk);
 246 }
 247
 248 void pblk_end_bio_sync(struct bio *bio)
 249 {
 250         struct completion *waiting = bio->bi_private;
 251
 252         complete(waiting);
 253 }
 254
 255 void pblk_end_io_sync(struct nvm_rq *rqd)
 256 {
 257         struct completion *waiting = rqd->private;
 258
 259         complete(waiting);
 260 }
 261
 262 void pblk_wait_for_meta(struct pblk *pblk)
 263 {
 264         do {
 265                 if (!atomic_read(&pblk->inflight_io))
 266                         break;
 267
 268                 schedule();
 269         } while (1);
 270 }
 271
 272 static void pblk_flush_writer(struct pblk *pblk)
 273 {
 274         pblk_rb_flush(&pblk->rwb);
 275         do {
 276                 if (!pblk_rb_read_count(&pblk->rwb))
 277                         break;
 278
 279                 schedule();
 280         } while (1);
 281 }
 282
 283 struct list_head *pblk_line_gc_list(struct pblk *pblk, struct pblk_line *line)
 284 {
 285         struct pblk_line_meta *lm = &pblk->lm;
 286         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
 287         struct list_head *move_list = NULL;
 288         int vsc = le32_to_cpu(*line->vsc);
 289
 290         lockdep_assert_held(&line->lock);
 291
 292         if (!vsc) {
 293                 if (line->gc_group != PBLK_LINEGC_FULL) {
 294                         line->gc_group = PBLK_LINEGC_FULL;
 295                         move_list = &l_mg->gc_full_list;
 296                 }
 297         } else if (vsc < lm->high_thrs) {
 298                 if (line->gc_group != PBLK_LINEGC_HIGH) {
 299                         line->gc_group = PBLK_LINEGC_HIGH;
 300                         move_list = &l_mg->gc_high_list;
 301                 }
 302         } else if (vsc < lm->mid_thrs) {
 303                 if (line->gc_group != PBLK_LINEGC_MID) {
 304                         line->gc_group = PBLK_LINEGC_MID;
 305                         move_list = &l_mg->gc_mid_list;
 306                 }
 307         } else if (vsc < line->sec_in_line) {
 308                 if (line->gc_group != PBLK_LINEGC_LOW) {
 309                         line->gc_group = PBLK_LINEGC_LOW;
 310                         move_list = &l_mg->gc_low_list;
 311                 }
 312         } else if (vsc == line->sec_in_line) {
 313                 if (line->gc_group != PBLK_LINEGC_EMPTY) {
 314                         line->gc_group = PBLK_LINEGC_EMPTY;
 315                         move_list = &l_mg->gc_empty_list;
 316                 }
 317         } else {
 318                 line->state = PBLK_LINESTATE_CORRUPT;
 319                 line->gc_group = PBLK_LINEGC_NONE;
 320                 move_list =  &l_mg->corrupt_list;
 321                 pr_err("pblk: corrupted vsc for line %d, vsc:%d (%d/%d/%d)\n",
 322                                                 line->id, vsc,
 323                                                 line->sec_in_line,
 324                                                 lm->high_thrs, lm->mid_thrs);
 325         }
 326
 327         return move_list;
 328 }
 329
 330 void pblk_discard(struct pblk *pblk, struct bio *bio)
 331 {
 332         sector_t slba = pblk_get_lba(bio);
 333         sector_t nr_secs = pblk_get_secs(bio);
 334
 335         pblk_invalidate_range(pblk, slba, nr_secs);
 336 }
 337
 338 struct ppa_addr pblk_get_lba_map(struct pblk *pblk, sector_t lba)
 339 {
 340         struct ppa_addr ppa;
 341
 342         spin_lock(&pblk->trans_lock);
 343         ppa = pblk_trans_map_get(pblk, lba);
 344         spin_unlock(&pblk->trans_lock);
 345
 346         return ppa;
 347 }
 348
 349 void pblk_log_write_err(struct pblk *pblk, struct nvm_rq *rqd)
 350 {
 351         atomic_long_inc(&pblk->write_failed);
 352 #ifdef CONFIG_NVM_DEBUG
 353         pblk_print_failed_rqd(pblk, rqd, rqd->error);
 354 #endif
 355 }
 356
 357 void pblk_log_read_err(struct pblk *pblk, struct nvm_rq *rqd)
 358 {
 359         /* Empty page read is not necessarily an error (e.g., L2P recovery) */
 360         if (rqd->error == NVM_RSP_ERR_EMPTYPAGE) {
 361                 atomic_long_inc(&pblk->read_empty);
 362                 return;
 363         }
 364
 365         switch (rqd->error) {
 366         case NVM_RSP_WARN_HIGHECC:
 367                 atomic_long_inc(&pblk->read_high_ecc);
 368                 break;
 369         case NVM_RSP_ERR_FAILECC:
 370         case NVM_RSP_ERR_FAILCRC:
 371                 atomic_long_inc(&pblk->read_failed);
 372                 break;
 373         default:
 374                 pr_err("pblk: unknown read error:%d\n", rqd->error);
 375         }
 376 #ifdef CONFIG_NVM_DEBUG
 377         pblk_print_failed_rqd(pblk, rqd, rqd->error);
 378 #endif
 379 }
 380
 381 void pblk_set_sec_per_write(struct pblk *pblk, int sec_per_write)
 382 {
 383         pblk->sec_per_write = sec_per_write;
 384 }
 385
 386 int pblk_submit_io(struct pblk *pblk, struct nvm_rq *rqd)
 387 {
 388         struct nvm_tgt_dev *dev = pblk->dev;
 389
 390 #ifdef CONFIG_NVM_DEBUG
 391         struct ppa_addr *ppa_list;
 392
 393         ppa_list = (rqd->nr_ppas > 1) ? rqd->ppa_list : &rqd->ppa_addr;
 394         if (pblk_boundary_ppa_checks(dev, ppa_list, rqd->nr_ppas)) {
 395                 WARN_ON(1);
 396                 return -EINVAL;
 397         }
 398
 399         if (rqd->opcode == NVM_OP_PWRITE) {
 400                 struct pblk_line *line;
 401                 struct ppa_addr ppa;
 402                 int i;
 403
 404                 for (i = 0; i < rqd->nr_ppas; i++) {
 405                         ppa = ppa_list[i];
 406                         line = &pblk->lines[pblk_dev_ppa_to_line(ppa)];
 407
 408                         spin_lock(&line->lock);
 409                         if (line->state != PBLK_LINESTATE_OPEN) {
 410                                 pr_err("pblk: bad ppa: line:%d,state:%d\n",
 411                                                         line->id, line->state);
 412                                 WARN_ON(1);
 413                                 spin_unlock(&line->lock);
 414                                 return -EINVAL;
 415                         }
 416                         spin_unlock(&line->lock);
 417                 }
 418         }
 419 #endif
 420
 421         atomic_inc(&pblk->inflight_io);
 422
 423         return nvm_submit_io(dev, rqd);
 424 }
 425
 426 struct bio *pblk_bio_map_addr(struct pblk *pblk, void *data,
 427                               unsigned int nr_secs, unsigned int len,
 428                               gfp_t gfp_mask)
 429 {
 430         struct nvm_tgt_dev *dev = pblk->dev;
 431         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
 432         void *kaddr = data;
 433         struct page *page;
 434         struct bio *bio;
 435         int i, ret;
 436
 437         if (l_mg->emeta_alloc_type == PBLK_KMALLOC_META)
 438                 return bio_map_kern(dev->q, kaddr, len, gfp_mask);
 439
 440         bio = bio_kmalloc(gfp_mask, nr_secs);
 441         if (!bio)
 442                 return ERR_PTR(-ENOMEM);
 443
 444         for (i = 0; i < nr_secs; i++) {
 445                 page = vmalloc_to_page(kaddr);
 446                 if (!page) {
 447                         pr_err("pblk: could not map vmalloc bio\n");
 448                         bio_put(bio);
 449                         bio = ERR_PTR(-ENOMEM);
 450                         goto out;
 451                 }
 452
 453                 ret = bio_add_pc_page(dev->q, bio, page, PAGE_SIZE, 0);
 454                 if (ret != PAGE_SIZE) {
 455                         pr_err("pblk: could not add page to bio\n");
 456                         bio_put(bio);
 457                         bio = ERR_PTR(-ENOMEM);
 458                         goto out;
 459                 }
 460
 461                 kaddr += PAGE_SIZE;
 462         }
 463 out:
 464         return bio;
 465 }
 466
 467 int pblk_calc_secs(struct pblk *pblk, unsigned long secs_avail,
 468                    unsigned long secs_to_flush)
 469 {
 470         int max = pblk->sec_per_write;
 471         int min = pblk->min_write_pgs;
 472         int secs_to_sync = 0;
 473
 474         if (secs_avail >= max)
 475                 secs_to_sync = max;
 476         else if (secs_avail >= min)
 477                 secs_to_sync = min * (secs_avail / min);
 478         else if (secs_to_flush)
 479                 secs_to_sync = min;
 480
 481         return secs_to_sync;
 482 }
 483
 484 void pblk_dealloc_page(struct pblk *pblk, struct pblk_line *line, int nr_secs)
 485 {
 486         u64 addr;
 487         int i;
 488
 489         addr = find_next_zero_bit(line->map_bitmap,
 490                                         pblk->lm.sec_per_line, line->cur_sec);
 491         line->cur_sec = addr - nr_secs;
 492
 493         for (i = 0; i < nr_secs; i++, line->cur_sec--)
 494                 WARN_ON(!test_and_clear_bit(line->cur_sec, line->map_bitmap));
 495 }
 496
 497 u64 __pblk_alloc_page(struct pblk *pblk, struct pblk_line *line, int nr_secs)
 498 {
 499         u64 addr;
 500         int i;
 501
 502         lockdep_assert_held(&line->lock);
 503
 504         /* logic error: ppa out-of-bounds. Prevent generating bad address */
 505         if (line->cur_sec + nr_secs > pblk->lm.sec_per_line) {
 506                 WARN(1, "pblk: page allocation out of bounds\n");
 507                 nr_secs = pblk->lm.sec_per_line - line->cur_sec;
 508         }
 509
 510         line->cur_sec = addr = find_next_zero_bit(line->map_bitmap,
 511                                         pblk->lm.sec_per_line, line->cur_sec);
 512         for (i = 0; i < nr_secs; i++, line->cur_sec++)
 513                 WARN_ON(test_and_set_bit(line->cur_sec, line->map_bitmap));
 514
 515         return addr;
 516 }
 517
 518 u64 pblk_alloc_page(struct pblk *pblk, struct pblk_line *line, int nr_secs)
 519 {
 520         u64 addr;
 521
 522         /* Lock needed in case a write fails and a recovery needs to remap
 523          * failed write buffer entries
 524          */
 525         spin_lock(&line->lock);
 526         addr = __pblk_alloc_page(pblk, line, nr_secs);
 527         line->left_msecs -= nr_secs;
 528         WARN(line->left_msecs < 0, "pblk: page allocation out of bounds\n");
 529         spin_unlock(&line->lock);
 530
 531         return addr;
 532 }
 533
 534 u64 pblk_lookup_page(struct pblk *pblk, struct pblk_line *line)
 535 {
 536         u64 paddr;
 537
 538         spin_lock(&line->lock);
 539         paddr = find_next_zero_bit(line->map_bitmap,
 540                                         pblk->lm.sec_per_line, line->cur_sec);
 541         spin_unlock(&line->lock);
 542
 543         return paddr;
 544 }
 545
 546 /*
 547  * Submit emeta to one LUN in the raid line at the time to avoid a deadlock when
 548  * taking the per LUN semaphore.
 549  */
 550 static int pblk_line_submit_emeta_io(struct pblk *pblk, struct pblk_line *line,
 551                                      void *emeta_buf, u64 paddr, int dir)
 552 {
 553         struct nvm_tgt_dev *dev = pblk->dev;
 554         struct nvm_geo *geo = &dev->geo;
 555         struct pblk_line_meta *lm = &pblk->lm;
 556         void *ppa_list, *meta_list;
 557         struct bio *bio;
 558         struct nvm_rq rqd;
 559         dma_addr_t dma_ppa_list, dma_meta_list;
 560         int min = pblk->min_write_pgs;
 561         int left_ppas = lm->emeta_sec[0];
 562         int id = line->id;
 563         int rq_ppas, rq_len;
 564         int cmd_op, bio_op;
 565         int i, j;
 566         int ret;
 567         DECLARE_COMPLETION_ONSTACK(wait);
 568
 569         if (dir == WRITE) {
 570                 bio_op = REQ_OP_WRITE;
 571                 cmd_op = NVM_OP_PWRITE;
 572         } else if (dir == READ) {
 573                 bio_op = REQ_OP_READ;
 574                 cmd_op = NVM_OP_PREAD;
 575         } else
 576                 return -EINVAL;
 577
 578         meta_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
 579                                                         &dma_meta_list);
 580         if (!meta_list)
 581                 return -ENOMEM;
 582
 583         ppa_list = meta_list + pblk_dma_meta_size;
 584         dma_ppa_list = dma_meta_list + pblk_dma_meta_size;
 585
 586 next_rq:
 587         memset(&rqd, 0, sizeof(struct nvm_rq));
 588
 589         rq_ppas = pblk_calc_secs(pblk, left_ppas, 0);
 590         rq_len = rq_ppas * geo->sec_size;
 591
 592         bio = pblk_bio_map_addr(pblk, emeta_buf, rq_ppas, rq_len, GFP_KERNEL);
 593         if (IS_ERR(bio)) {
 594                 ret = PTR_ERR(bio);
 595                 goto free_rqd_dma;
 596         }
 597
 598         bio->bi_iter.bi_sector = 0; /* internal bio */
 599         bio_set_op_attrs(bio, bio_op, 0);
 600
 601         rqd.bio = bio;
 602         rqd.meta_list = meta_list;
 603         rqd.ppa_list = ppa_list;
 604         rqd.dma_meta_list = dma_meta_list;
 605         rqd.dma_ppa_list = dma_ppa_list;
 606         rqd.opcode = cmd_op;
 607         rqd.nr_ppas = rq_ppas;
 608         rqd.end_io = pblk_end_io_sync;
 609         rqd.private = &wait;
 610
 611         if (dir == WRITE) {
 612                 struct pblk_sec_meta *meta_list = rqd.meta_list;
 613
 614                 rqd.flags = pblk_set_progr_mode(pblk, WRITE);
 615                 for (i = 0; i < rqd.nr_ppas; ) {
 616                         spin_lock(&line->lock);
 617                         paddr = __pblk_alloc_page(pblk, line, min);
 618                         spin_unlock(&line->lock);
 619                         for (j = 0; j < min; j++, i++, paddr++) {
 620                                 meta_list[i].lba = cpu_to_le64(ADDR_EMPTY);
 621                                 rqd.ppa_list[i] =
 622                                         addr_to_gen_ppa(pblk, paddr, id);
 623                         }
 624                 }
 625         } else {
 626                 for (i = 0; i < rqd.nr_ppas; ) {
 627                         struct ppa_addr ppa = addr_to_gen_ppa(pblk, paddr, id);
 628                         int pos = pblk_dev_ppa_to_pos(geo, ppa);
 629                         int read_type = PBLK_READ_RANDOM;
 630
 631                         if (pblk_io_aligned(pblk, rq_ppas))
 632                                 read_type = PBLK_READ_SEQUENTIAL;
 633                         rqd.flags = pblk_set_read_mode(pblk, read_type);
 634
 635                         while (test_bit(pos, line->blk_bitmap)) {
 636                                 paddr += min;
 637                                 if (pblk_boundary_paddr_checks(pblk, paddr)) {
 638                                         pr_err("pblk: corrupt emeta line:%d\n",
 639                                                                 line->id);
 640                                         bio_put(bio);
 641                                         ret = -EINTR;
 642                                         goto free_rqd_dma;
 643                                 }
 644
 645                                 ppa = addr_to_gen_ppa(pblk, paddr, id);
 646                                 pos = pblk_dev_ppa_to_pos(geo, ppa);
 647                         }
 648
 649                         if (pblk_boundary_paddr_checks(pblk, paddr + min)) {
 650                                 pr_err("pblk: corrupt emeta line:%d\n",
 651                                                                 line->id);
 652                                 bio_put(bio);
 653                                 ret = -EINTR;
 654                                 goto free_rqd_dma;
 655                         }
 656
 657                         for (j = 0; j < min; j++, i++, paddr++)
 658                                 rqd.ppa_list[i] =
 659                                         addr_to_gen_ppa(pblk, paddr, line->id);
 660                 }
 661         }
 662
 663         ret = pblk_submit_io(pblk, &rqd);
 664         if (ret) {
 665                 pr_err("pblk: emeta I/O submission failed: %d\n", ret);
 666                 bio_put(bio);
 667                 goto free_rqd_dma;
 668         }
 669
 670         if (!wait_for_completion_io_timeout(&wait,
 671                                 msecs_to_jiffies(PBLK_COMMAND_TIMEOUT_MS))) {
 672                 pr_err("pblk: emeta I/O timed out\n");
 673         }
 674         atomic_dec(&pblk->inflight_io);
 675         reinit_completion(&wait);
 676
 677         if (likely(pblk->l_mg.emeta_alloc_type == PBLK_VMALLOC_META))
 678                 bio_put(bio);
 679
 680         if (rqd.error) {
 681                 if (dir == WRITE)
 682                         pblk_log_write_err(pblk, &rqd);
 683                 else
 684                         pblk_log_read_err(pblk, &rqd);
 685         }
 686
 687         emeta_buf += rq_len;
 688         left_ppas -= rq_ppas;
 689         if (left_ppas)
 690                 goto next_rq;
 691 free_rqd_dma:
 692         nvm_dev_dma_free(dev->parent, rqd.meta_list, rqd.dma_meta_list);
 693         return ret;
 694 }
 695
 696 u64 pblk_line_smeta_start(struct pblk *pblk, struct pblk_line *line)
 697 {
 698         struct nvm_tgt_dev *dev = pblk->dev;
 699         struct nvm_geo *geo = &dev->geo;
 700         struct pblk_line_meta *lm = &pblk->lm;
 701         int bit;
 702
 703         /* This usually only happens on bad lines */
 704         bit = find_first_zero_bit(line->blk_bitmap, lm->blk_per_line);
 705         if (bit >= lm->blk_per_line)
 706                 return -1;
 707
 708         return bit * geo->sec_per_pl;
 709 }
 710
 711 static int pblk_line_submit_smeta_io(struct pblk *pblk, struct pblk_line *line,
 712                                      u64 paddr, int dir)
 713 {
 714         struct nvm_tgt_dev *dev = pblk->dev;
 715         struct pblk_line_meta *lm = &pblk->lm;
 716         struct bio *bio;
 717         struct nvm_rq rqd;
 718         __le64 *lba_list = NULL;
 719         int i, ret;
 720         int cmd_op, bio_op;
 721         int flags;
 722         DECLARE_COMPLETION_ONSTACK(wait);
 723
 724         if (dir == WRITE) {
 725                 bio_op = REQ_OP_WRITE;
 726                 cmd_op = NVM_OP_PWRITE;
 727                 flags = pblk_set_progr_mode(pblk, WRITE);
 728                 lba_list = emeta_to_lbas(pblk, line->emeta->buf);
 729         } else if (dir == READ) {
 730                 bio_op = REQ_OP_READ;
 731                 cmd_op = NVM_OP_PREAD;
 732                 flags = pblk_set_read_mode(pblk, PBLK_READ_SEQUENTIAL);
 733         } else
 734                 return -EINVAL;
 735
 736         memset(&rqd, 0, sizeof(struct nvm_rq));
 737
 738         rqd.meta_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
 739                                                         &rqd.dma_meta_list);
 740         if (!rqd.meta_list)
 741                 return -ENOMEM;
 742
 743         rqd.ppa_list = rqd.meta_list + pblk_dma_meta_size;
 744         rqd.dma_ppa_list = rqd.dma_meta_list + pblk_dma_meta_size;
 745
 746         bio = bio_map_kern(dev->q, line->smeta, lm->smeta_len, GFP_KERNEL);
 747         if (IS_ERR(bio)) {
 748                 ret = PTR_ERR(bio);
 749                 goto free_ppa_list;
 750         }
 751
 752         bio->bi_iter.bi_sector = 0; /* internal bio */
 753         bio_set_op_attrs(bio, bio_op, 0);
 754
 755         rqd.bio = bio;
 756         rqd.opcode = cmd_op;
 757         rqd.flags = flags;
 758         rqd.nr_ppas = lm->smeta_sec;
 759         rqd.end_io = pblk_end_io_sync;
 760         rqd.private = &wait;
 761
 762         for (i = 0; i < lm->smeta_sec; i++, paddr++) {
 763                 struct pblk_sec_meta *meta_list = rqd.meta_list;
 764
 765                 rqd.ppa_list[i] = addr_to_gen_ppa(pblk, paddr, line->id);
 766
 767                 if (dir == WRITE) {
 768                         __le64 addr_empty = cpu_to_le64(ADDR_EMPTY);
 769
 770                         meta_list[i].lba = lba_list[paddr] = addr_empty;
 771                 }
 772         }
 773
 774         /*
 775          * This I/O is sent by the write thread when a line is replace. Since
 776          * the write thread is the only one sending write and erase commands,
 777          * there is no need to take the LUN semaphore.
 778          */
 779         ret = pblk_submit_io(pblk, &rqd);
 780         if (ret) {
 781                 pr_err("pblk: smeta I/O submission failed: %d\n", ret);
 782                 bio_put(bio);
 783                 goto free_ppa_list;
 784         }
 785
 786         if (!wait_for_completion_io_timeout(&wait,
 787                                 msecs_to_jiffies(PBLK_COMMAND_TIMEOUT_MS))) {
 788                 pr_err("pblk: smeta I/O timed out\n");
 789         }
 790         atomic_dec(&pblk->inflight_io);
 791
 792         if (rqd.error) {
 793                 if (dir == WRITE)
 794                         pblk_log_write_err(pblk, &rqd);
 795                 else
 796                         pblk_log_read_err(pblk, &rqd);
 797         }
 798
 799 free_ppa_list:
 800         nvm_dev_dma_free(dev->parent, rqd.meta_list, rqd.dma_meta_list);
 801
 802         return ret;
 803 }
 804
 805 int pblk_line_read_smeta(struct pblk *pblk, struct pblk_line *line)
 806 {
 807         u64 bpaddr = pblk_line_smeta_start(pblk, line);
 808
 809         return pblk_line_submit_smeta_io(pblk, line, bpaddr, READ);
 810 }
 811
 812 int pblk_line_read_emeta(struct pblk *pblk, struct pblk_line *line,
 813                          void *emeta_buf)
 814 {
 815         return pblk_line_submit_emeta_io(pblk, line, emeta_buf,
 816                                                 line->emeta_ssec, READ);
 817 }
 818
 819 static void pblk_setup_e_rq(struct pblk *pblk, struct nvm_rq *rqd,
 820                             struct ppa_addr ppa)
 821 {
 822         rqd->opcode = NVM_OP_ERASE;
 823         rqd->ppa_addr = ppa;
 824         rqd->nr_ppas = 1;
 825         rqd->flags = pblk_set_progr_mode(pblk, ERASE);
 826         rqd->bio = NULL;
 827 }
 828
 829 static int pblk_blk_erase_sync(struct pblk *pblk, struct ppa_addr ppa)
 830 {
 831         struct nvm_rq rqd;
 832         int ret = 0;
 833         DECLARE_COMPLETION_ONSTACK(wait);
 834
 835         memset(&rqd, 0, sizeof(struct nvm_rq));
 836
 837         pblk_setup_e_rq(pblk, &rqd, ppa);
 838
 839         rqd.end_io = pblk_end_io_sync;
 840         rqd.private = &wait;
 841
 842         /* The write thread schedules erases so that it minimizes disturbances
 843          * with writes. Thus, there is no need to take the LUN semaphore.
 844          */
 845         ret = pblk_submit_io(pblk, &rqd);
 846         if (ret) {
 847                 struct nvm_tgt_dev *dev = pblk->dev;
 848                 struct nvm_geo *geo = &dev->geo;
 849
 850                 pr_err("pblk: could not sync erase line:%d,blk:%d\n",
 851                                         pblk_dev_ppa_to_line(ppa),
 852                                         pblk_dev_ppa_to_pos(geo, ppa));
 853
 854                 rqd.error = ret;
 855                 goto out;
 856         }
 857
 858         if (!wait_for_completion_io_timeout(&wait,
 859                                 msecs_to_jiffies(PBLK_COMMAND_TIMEOUT_MS))) {
 860                 pr_err("pblk: sync erase timed out\n");
 861         }
 862
 863 out:
 864         rqd.private = pblk;
 865         __pblk_end_io_erase(pblk, &rqd);
 866
 867         return ret;
 868 }
 869
 870 int pblk_line_erase(struct pblk *pblk, struct pblk_line *line)
 871 {
 872         struct pblk_line_meta *lm = &pblk->lm;
 873         struct ppa_addr ppa;
 874         int ret, bit = -1;
 875
 876         /* Erase only good blocks, one at a time */
 877         do {
 878                 spin_lock(&line->lock);
 879                 bit = find_next_zero_bit(line->erase_bitmap, lm->blk_per_line,
 880                                                                 bit + 1);
 881                 if (bit >= lm->blk_per_line) {
 882                         spin_unlock(&line->lock);
 883                         break;
 884                 }
 885
 886                 ppa = pblk->luns[bit].bppa; /* set ch and lun */
 887                 ppa.g.blk = line->id;
 888
 889                 atomic_dec(&line->left_eblks);
 890                 WARN_ON(test_and_set_bit(bit, line->erase_bitmap));
 891                 spin_unlock(&line->lock);
 892
 893                 ret = pblk_blk_erase_sync(pblk, ppa);
 894                 if (ret) {
 895                         pr_err("pblk: failed to erase line %d\n", line->id);
 896                         return ret;
 897                 }
 898         } while (1);
 899
 900         return 0;
 901 }
 902
 903 static void pblk_line_setup_metadata(struct pblk_line *line,
 904                                      struct pblk_line_mgmt *l_mg,
 905                                      struct pblk_line_meta *lm)
 906 {
 907         int meta_line;
 908
 909         lockdep_assert_held(&l_mg->free_lock);
 910
 911 retry_meta:
 912         meta_line = find_first_zero_bit(&l_mg->meta_bitmap, PBLK_DATA_LINES);
 913         if (meta_line == PBLK_DATA_LINES) {
 914                 spin_unlock(&l_mg->free_lock);
 915                 io_schedule();
 916                 spin_lock(&l_mg->free_lock);
 917                 goto retry_meta;
 918         }
 919
 920         set_bit(meta_line, &l_mg->meta_bitmap);
 921         line->meta_line = meta_line;
 922
 923         line->smeta = l_mg->sline_meta[meta_line];
 924         line->emeta = l_mg->eline_meta[meta_line];
 925
 926         memset(line->smeta, 0, lm->smeta_len);
 927         memset(line->emeta->buf, 0, lm->emeta_len[0]);
 928
 929         line->emeta->mem = 0;
 930         atomic_set(&line->emeta->sync, 0);
 931 }
 932
 933 /* For now lines are always assumed full lines. Thus, smeta former and current
 934  * lun bitmaps are omitted.
 935  */
 936 static int pblk_line_init_metadata(struct pblk *pblk, struct pblk_line *line,
 937                                   struct pblk_line *cur)
 938 {
 939         struct nvm_tgt_dev *dev = pblk->dev;
 940         struct nvm_geo *geo = &dev->geo;
 941         struct pblk_line_meta *lm = &pblk->lm;
 942         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
 943         struct pblk_emeta *emeta = line->emeta;
 944         struct line_emeta *emeta_buf = emeta->buf;
 945         struct line_smeta *smeta_buf = (struct line_smeta *)line->smeta;
 946         int nr_blk_line;
 947
 948         /* After erasing the line, new bad blocks might appear and we risk
 949          * having an invalid line
 950          */
 951         nr_blk_line = lm->blk_per_line -
 952                         bitmap_weight(line->blk_bitmap, lm->blk_per_line);
 953         if (nr_blk_line < lm->min_blk_line) {
 954                 spin_lock(&l_mg->free_lock);
 955                 spin_lock(&line->lock);
 956                 line->state = PBLK_LINESTATE_BAD;
 957                 spin_unlock(&line->lock);
 958
 959                 list_add_tail(&line->list, &l_mg->bad_list);
 960                 spin_unlock(&l_mg->free_lock);
 961
 962                 pr_debug("pblk: line %d is bad\n", line->id);
 963
 964                 return 0;
 965         }
 966
 967         /* Run-time metadata */
 968         line->lun_bitmap = ((void *)(smeta_buf)) + sizeof(struct line_smeta);
 969
 970         /* Mark LUNs allocated in this line (all for now) */
 971         bitmap_set(line->lun_bitmap, 0, lm->lun_bitmap_len);
 972
 973         smeta_buf->header.identifier = cpu_to_le32(PBLK_MAGIC);
 974         memcpy(smeta_buf->header.uuid, pblk->instance_uuid, 16);
 975         smeta_buf->header.id = cpu_to_le32(line->id);
 976         smeta_buf->header.type = cpu_to_le16(line->type);
 977         smeta_buf->header.version = cpu_to_le16(1);
 978
 979         /* Start metadata */
 980         smeta_buf->seq_nr = cpu_to_le64(line->seq_nr);
 981         smeta_buf->window_wr_lun = cpu_to_le32(geo->nr_luns);
 982
 983         /* Fill metadata among lines */
 984         if (cur) {
 985                 memcpy(line->lun_bitmap, cur->lun_bitmap, lm->lun_bitmap_len);
 986                 smeta_buf->prev_id = cpu_to_le32(cur->id);
 987                 cur->emeta->buf->next_id = cpu_to_le32(line->id);
 988         } else {
 989                 smeta_buf->prev_id = cpu_to_le32(PBLK_LINE_EMPTY);
 990         }
 991
 992         /* All smeta must be set at this point */
 993         smeta_buf->header.crc = cpu_to_le32(
 994                         pblk_calc_meta_header_crc(pblk, &smeta_buf->header));
 995         smeta_buf->crc = cpu_to_le32(pblk_calc_smeta_crc(pblk, smeta_buf));
 996
 997         /* End metadata */
 998         memcpy(&emeta_buf->header, &smeta_buf->header,
 999                                                 sizeof(struct line_header));
1000         emeta_buf->seq_nr = cpu_to_le64(line->seq_nr);
1001         emeta_buf->nr_lbas = cpu_to_le64(line->sec_in_line);
1002         emeta_buf->nr_valid_lbas = cpu_to_le64(0);
1003         emeta_buf->next_id = cpu_to_le32(PBLK_LINE_EMPTY);
1004         emeta_buf->crc = cpu_to_le32(0);
1005         emeta_buf->prev_id = smeta_buf->prev_id;
1006
1007         return 1;
1008 }
1009
1010 /* For now lines are always assumed full lines. Thus, smeta former and current
1011  * lun bitmaps are omitted.
1012  */
1013 static int pblk_line_init_bb(struct pblk *pblk, struct pblk_line *line,
1014                              int init)
1015 {
1016         struct nvm_tgt_dev *dev = pblk->dev;
1017         struct nvm_geo *geo = &dev->geo;
1018         struct pblk_line_meta *lm = &pblk->lm;
1019         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1020         int nr_bb = 0;
1021         u64 off;
1022         int bit = -1;
1023
1024         line->sec_in_line = lm->sec_per_line;
1025
1026         /* Capture bad block information on line mapping bitmaps */
1027         while ((bit = find_next_bit(line->blk_bitmap, lm->blk_per_line,
1028                                         bit + 1)) < lm->blk_per_line) {
1029                 off = bit * geo->sec_per_pl;
1030                 bitmap_shift_left(l_mg->bb_aux, l_mg->bb_template, off,
1031                                                         lm->sec_per_line);
1032                 bitmap_or(line->map_bitmap, line->map_bitmap, l_mg->bb_aux,
1033                                                         lm->sec_per_line);
1034                 line->sec_in_line -= geo->sec_per_blk;
1035                 if (bit >= lm->emeta_bb)
1036                         nr_bb++;
1037         }
1038
1039         /* Mark smeta metadata sectors as bad sectors */
1040         bit = find_first_zero_bit(line->blk_bitmap, lm->blk_per_line);
1041         off = bit * geo->sec_per_pl;
1042         bitmap_set(line->map_bitmap, off, lm->smeta_sec);
1043         line->sec_in_line -= lm->smeta_sec;
1044         line->smeta_ssec = off;
1045         line->cur_sec = off + lm->smeta_sec;
1046
1047         if (init && pblk_line_submit_smeta_io(pblk, line, off, WRITE)) {
1048                 pr_debug("pblk: line smeta I/O failed. Retry\n");
1049                 return 1;
1050         }
1051
1052         bitmap_copy(line->invalid_bitmap, line->map_bitmap, lm->sec_per_line);
1053
1054         /* Mark emeta metadata sectors as bad sectors. We need to consider bad
1055          * blocks to make sure that there are enough sectors to store emeta
1056          */
1057         bit = lm->sec_per_line;
1058         off = lm->sec_per_line - lm->emeta_sec[0];
1059         bitmap_set(line->invalid_bitmap, off, lm->emeta_sec[0]);
1060         while (nr_bb) {
1061                 off -= geo->sec_per_pl;
1062                 if (!test_bit(off, line->invalid_bitmap)) {
1063                         bitmap_set(line->invalid_bitmap, off, geo->sec_per_pl);
1064                         nr_bb--;
1065                 }
1066         }
1067
1068         line->sec_in_line -= lm->emeta_sec[0];
1069         line->emeta_ssec = off;
1070         line->nr_valid_lbas = 0;
1071         line->left_msecs = line->sec_in_line;
1072         *line->vsc = cpu_to_le32(line->sec_in_line);
1073
1074         if (lm->sec_per_line - line->sec_in_line !=
1075                 bitmap_weight(line->invalid_bitmap, lm->sec_per_line)) {
1076                 spin_lock(&line->lock);
1077                 line->state = PBLK_LINESTATE_BAD;
1078                 spin_unlock(&line->lock);
1079
1080                 list_add_tail(&line->list, &l_mg->bad_list);
1081                 pr_err("pblk: unexpected line %d is bad\n", line->id);
1082
1083                 return 0;
1084         }
1085
1086         return 1;
1087 }
1088
1089 static int pblk_line_prepare(struct pblk *pblk, struct pblk_line *line)
1090 {
1091         struct pblk_line_meta *lm = &pblk->lm;
1092         int blk_in_line = atomic_read(&line->blk_in_line);
1093
1094         line->map_bitmap = mempool_alloc(pblk->line_meta_pool, GFP_ATOMIC);
1095         if (!line->map_bitmap)
1096                 return -ENOMEM;
1097         memset(line->map_bitmap, 0, lm->sec_bitmap_len);
1098
1099         /* invalid_bitmap is special since it is used when line is closed. No
1100          * need to zeroized; it will be initialized using bb info form
1101          * map_bitmap
1102          */
1103         line->invalid_bitmap = mempool_alloc(pblk->line_meta_pool, GFP_ATOMIC);
1104         if (!line->invalid_bitmap) {
1105                 mempool_free(line->map_bitmap, pblk->line_meta_pool);
1106                 return -ENOMEM;
1107         }
1108
1109         spin_lock(&line->lock);
1110         if (line->state != PBLK_LINESTATE_FREE) {
1111                 mempool_free(line->invalid_bitmap, pblk->line_meta_pool);
1112                 mempool_free(line->map_bitmap, pblk->line_meta_pool);
1113                 spin_unlock(&line->lock);
1114                 WARN(1, "pblk: corrupted line %d, state %d\n",
1115                                                         line->id, line->state);
1116                 return -EAGAIN;
1117         }
1118
1119         line->state = PBLK_LINESTATE_OPEN;
1120
1121         atomic_set(&line->left_eblks, blk_in_line);
1122         atomic_set(&line->left_seblks, blk_in_line);
1123
1124         line->meta_distance = lm->meta_distance;
1125         spin_unlock(&line->lock);
1126
1127         /* Bad blocks do not need to be erased */
1128         bitmap_copy(line->erase_bitmap, line->blk_bitmap, lm->blk_per_line);
1129
1130         kref_init(&line->ref);
1131
1132         return 0;
1133 }
1134
1135 int pblk_line_recov_alloc(struct pblk *pblk, struct pblk_line *line)
1136 {
1137         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1138         int ret;
1139
1140         spin_lock(&l_mg->free_lock);
1141         l_mg->data_line = line;
1142         list_del(&line->list);
1143
1144         ret = pblk_line_prepare(pblk, line);
1145         if (ret) {
1146                 list_add(&line->list, &l_mg->free_list);
1147                 spin_unlock(&l_mg->free_lock);
1148                 return ret;
1149         }
1150         spin_unlock(&l_mg->free_lock);
1151
1152         pblk_rl_free_lines_dec(&pblk->rl, line);
1153
1154         if (!pblk_line_init_bb(pblk, line, 0)) {
1155                 list_add(&line->list, &l_mg->free_list);
1156                 return -EINTR;
1157         }
1158
1159         return 0;
1160 }
1161
1162 void pblk_line_recov_close(struct pblk *pblk, struct pblk_line *line)
1163 {
1164         mempool_free(line->map_bitmap, pblk->line_meta_pool);
1165         line->map_bitmap = NULL;
1166         line->smeta = NULL;
1167         line->emeta = NULL;
1168 }
1169
1170 struct pblk_line *pblk_line_get(struct pblk *pblk)
1171 {
1172         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1173         struct pblk_line_meta *lm = &pblk->lm;
1174         struct pblk_line *line;
1175         int ret, bit;
1176
1177         lockdep_assert_held(&l_mg->free_lock);
1178
1179 retry:
1180         if (list_empty(&l_mg->free_list)) {
1181                 pr_err("pblk: no free lines\n");
1182                 return NULL;
1183         }
1184
1185         line = list_first_entry(&l_mg->free_list, struct pblk_line, list);
1186         list_del(&line->list);
1187         l_mg->nr_free_lines--;
1188
1189         bit = find_first_zero_bit(line->blk_bitmap, lm->blk_per_line);
1190         if (unlikely(bit >= lm->blk_per_line)) {
1191                 spin_lock(&line->lock);
1192                 line->state = PBLK_LINESTATE_BAD;
1193                 spin_unlock(&line->lock);
1194
1195                 list_add_tail(&line->list, &l_mg->bad_list);
1196
1197                 pr_debug("pblk: line %d is bad\n", line->id);
1198                 goto retry;
1199         }
1200
1201         ret = pblk_line_prepare(pblk, line);
1202         if (ret) {
1203                 if (ret == -EAGAIN) {
1204                         list_add(&line->list, &l_mg->corrupt_list);
1205                         goto retry;
1206                 } else {
1207                         pr_err("pblk: failed to prepare line %d\n", line->id);
1208                         list_add(&line->list, &l_mg->free_list);
1209                         l_mg->nr_free_lines++;
1210                         return NULL;
1211                 }
1212         }
1213
1214         return line;
1215 }
1216
1217 static struct pblk_line *pblk_line_retry(struct pblk *pblk,
1218                                          struct pblk_line *line)
1219 {
1220         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1221         struct pblk_line *retry_line;
1222
1223 retry:
1224         spin_lock(&l_mg->free_lock);
1225         retry_line = pblk_line_get(pblk);
1226         if (!retry_line) {
1227                 l_mg->data_line = NULL;
1228                 spin_unlock(&l_mg->free_lock);
1229                 return NULL;
1230         }
1231
1232         retry_line->smeta = line->smeta;
1233         retry_line->emeta = line->emeta;
1234         retry_line->meta_line = line->meta_line;
1235
1236         pblk_line_free(pblk, line);
1237         l_mg->data_line = retry_line;
1238         spin_unlock(&l_mg->free_lock);
1239
1240         pblk_rl_free_lines_dec(&pblk->rl, retry_line);
1241
1242         if (pblk_line_erase(pblk, retry_line))
1243                 goto retry;
1244
1245         return retry_line;
1246 }
1247
1248 static void pblk_set_space_limit(struct pblk *pblk)
1249 {
1250         struct pblk_rl *rl = &pblk->rl;
1251
1252         atomic_set(&rl->rb_space, 0);
1253 }
1254
1255 struct pblk_line *pblk_line_get_first_data(struct pblk *pblk)
1256 {
1257         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1258         struct pblk_line *line;
1259         int is_next = 0;
1260
1261         spin_lock(&l_mg->free_lock);
1262         line = pblk_line_get(pblk);
1263         if (!line) {
1264                 spin_unlock(&l_mg->free_lock);
1265                 return NULL;
1266         }
1267
1268         line->seq_nr = l_mg->d_seq_nr++;
1269         line->type = PBLK_LINETYPE_DATA;
1270         l_mg->data_line = line;
1271
1272         pblk_line_setup_metadata(line, l_mg, &pblk->lm);
1273
1274         /* Allocate next line for preparation */
1275         l_mg->data_next = pblk_line_get(pblk);
1276         if (!l_mg->data_next) {
1277                 /* If we cannot get a new line, we need to stop the pipeline.
1278                  * Only allow as many writes in as we can store safely and then
1279                  * fail gracefully
1280                  */
1281                 pblk_set_space_limit(pblk);
1282
1283                 l_mg->data_next = NULL;
1284         } else {
1285                 l_mg->data_next->seq_nr = l_mg->d_seq_nr++;
1286                 l_mg->data_next->type = PBLK_LINETYPE_DATA;
1287                 is_next = 1;
1288         }
1289         spin_unlock(&l_mg->free_lock);
1290
1291         if (pblk_line_erase(pblk, line)) {
1292                 line = pblk_line_retry(pblk, line);
1293                 if (!line)
1294                         return NULL;
1295         }
1296
1297         pblk_rl_free_lines_dec(&pblk->rl, line);
1298         if (is_next)
1299                 pblk_rl_free_lines_dec(&pblk->rl, l_mg->data_next);
1300
1301 retry_setup:
1302         if (!pblk_line_init_metadata(pblk, line, NULL)) {
1303                 line = pblk_line_retry(pblk, line);
1304                 if (!line)
1305                         return NULL;
1306
1307                 goto retry_setup;
1308         }
1309
1310         if (!pblk_line_init_bb(pblk, line, 1)) {
1311                 line = pblk_line_retry(pblk, line);
1312                 if (!line)
1313                         return NULL;
1314
1315                 goto retry_setup;
1316         }
1317
1318         return line;
1319 }
1320
1321 static void pblk_stop_writes(struct pblk *pblk, struct pblk_line *line)
1322 {
1323         lockdep_assert_held(&pblk->l_mg.free_lock);
1324
1325         pblk_set_space_limit(pblk);
1326         pblk->state = PBLK_STATE_STOPPING;
1327 }
1328
1329 void pblk_pipeline_stop(struct pblk *pblk)
1330 {
1331         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1332         int ret;
1333
1334         spin_lock(&l_mg->free_lock);
1335         if (pblk->state == PBLK_STATE_RECOVERING ||
1336                                         pblk->state == PBLK_STATE_STOPPED) {
1337                 spin_unlock(&l_mg->free_lock);
1338                 return;
1339         }
1340         pblk->state = PBLK_STATE_RECOVERING;
1341         spin_unlock(&l_mg->free_lock);
1342
1343         pblk_flush_writer(pblk);
1344         pblk_wait_for_meta(pblk);
1345
1346         ret = pblk_recov_pad(pblk);
1347         if (ret) {
1348                 pr_err("pblk: could not close data on teardown(%d)\n", ret);
1349                 return;
1350         }
1351
1352         pblk_line_close_meta_sync(pblk);
1353
1354         spin_lock(&l_mg->free_lock);
1355         pblk->state = PBLK_STATE_STOPPED;
1356         l_mg->data_line = NULL;
1357         l_mg->data_next = NULL;
1358         spin_unlock(&l_mg->free_lock);
1359 }
1360
1361 void pblk_line_replace_data(struct pblk *pblk)
1362 {
1363         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1364         struct pblk_line *cur, *new;
1365         unsigned int left_seblks;
1366         int is_next = 0;
1367
1368         cur = l_mg->data_line;
1369         new = l_mg->data_next;
1370         if (!new)
1371                 return;
1372         l_mg->data_line = new;
1373
1374         spin_lock(&l_mg->free_lock);
1375         if (pblk->state != PBLK_STATE_RUNNING) {
1376                 l_mg->data_line = NULL;
1377                 l_mg->data_next = NULL;
1378                 spin_unlock(&l_mg->free_lock);
1379                 return;
1380         }
1381
1382         pblk_line_setup_metadata(new, l_mg, &pblk->lm);
1383         spin_unlock(&l_mg->free_lock);
1384
1385 retry_erase:
1386         left_seblks = atomic_read(&new->left_seblks);
1387         if (left_seblks) {
1388                 /* If line is not fully erased, erase it */
1389                 if (atomic_read(&new->left_eblks)) {
1390                         if (pblk_line_erase(pblk, new))
1391                                 return;
1392                 } else {
1393                         io_schedule();
1394                 }
1395                 goto retry_erase;
1396         }
1397
1398 retry_setup:
1399         if (!pblk_line_init_metadata(pblk, new, cur)) {
1400                 new = pblk_line_retry(pblk, new);
1401                 if (!new)
1402                         return;
1403
1404                 goto retry_setup;
1405         }
1406
1407         if (!pblk_line_init_bb(pblk, new, 1)) {
1408                 new = pblk_line_retry(pblk, new);
1409                 if (!new)
1410                         return;
1411
1412                 goto retry_setup;
1413         }
1414
1415         /* Allocate next line for preparation */
1416         spin_lock(&l_mg->free_lock);
1417         l_mg->data_next = pblk_line_get(pblk);
1418         if (!l_mg->data_next) {
1419                 /* If we cannot get a new line, we need to stop the pipeline.
1420                  * Only allow as many writes in as we can store safely and then
1421                  * fail gracefully
1422                  */
1423                 pblk_stop_writes(pblk, new);
1424                 l_mg->data_next = NULL;
1425         } else {
1426                 l_mg->data_next->seq_nr = l_mg->d_seq_nr++;
1427                 l_mg->data_next->type = PBLK_LINETYPE_DATA;
1428                 is_next = 1;
1429         }
1430         spin_unlock(&l_mg->free_lock);
1431
1432         if (is_next)
1433                 pblk_rl_free_lines_dec(&pblk->rl, l_mg->data_next);
1434 }
1435
1436 void pblk_line_free(struct pblk *pblk, struct pblk_line *line)
1437 {
1438         if (line->map_bitmap)
1439                 mempool_free(line->map_bitmap, pblk->line_meta_pool);
1440         if (line->invalid_bitmap)
1441                 mempool_free(line->invalid_bitmap, pblk->line_meta_pool);
1442
1443         *line->vsc = cpu_to_le32(EMPTY_ENTRY);
1444
1445         line->map_bitmap = NULL;
1446         line->invalid_bitmap = NULL;
1447         line->smeta = NULL;
1448         line->emeta = NULL;
1449 }
1450
1451 void pblk_line_put(struct kref *ref)
1452 {
1453         struct pblk_line *line = container_of(ref, struct pblk_line, ref);
1454         struct pblk *pblk = line->pblk;
1455         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1456
1457         spin_lock(&line->lock);
1458         WARN_ON(line->state != PBLK_LINESTATE_GC);
1459         line->state = PBLK_LINESTATE_FREE;
1460         line->gc_group = PBLK_LINEGC_NONE;
1461         pblk_line_free(pblk, line);
1462         spin_unlock(&line->lock);
1463
1464         spin_lock(&l_mg->free_lock);
1465         list_add_tail(&line->list, &l_mg->free_list);
1466         l_mg->nr_free_lines++;
1467         spin_unlock(&l_mg->free_lock);
1468
1469         pblk_rl_free_lines_inc(&pblk->rl, line);
1470 }
1471
1472 int pblk_blk_erase_async(struct pblk *pblk, struct ppa_addr ppa)
1473 {
1474         struct nvm_rq *rqd;
1475         int err;
1476
1477         rqd = mempool_alloc(pblk->g_rq_pool, GFP_KERNEL);
1478         memset(rqd, 0, pblk_g_rq_size);
1479
1480         pblk_setup_e_rq(pblk, rqd, ppa);
1481
1482         rqd->end_io = pblk_end_io_erase;
1483         rqd->private = pblk;
1484
1485         /* The write thread schedules erases so that it minimizes disturbances
1486          * with writes. Thus, there is no need to take the LUN semaphore.
1487          */
1488         err = pblk_submit_io(pblk, rqd);
1489         if (err) {
1490                 struct nvm_tgt_dev *dev = pblk->dev;
1491                 struct nvm_geo *geo = &dev->geo;
1492
1493                 pr_err("pblk: could not async erase line:%d,blk:%d\n",
1494                                         pblk_dev_ppa_to_line(ppa),
1495                                         pblk_dev_ppa_to_pos(geo, ppa));
1496         }
1497
1498         return err;
1499 }
1500
1501 struct pblk_line *pblk_line_get_data(struct pblk *pblk)
1502 {
1503         return pblk->l_mg.data_line;
1504 }
1505
1506 /* For now, always erase next line */
1507 struct pblk_line *pblk_line_get_erase(struct pblk *pblk)
1508 {
1509         return pblk->l_mg.data_next;
1510 }
1511
1512 int pblk_line_is_full(struct pblk_line *line)
1513 {
1514         return (line->left_msecs == 0);
1515 }
1516
1517 void pblk_line_close_meta_sync(struct pblk *pblk)
1518 {
1519         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1520         struct pblk_line_meta *lm = &pblk->lm;
1521         struct pblk_line *line, *tline;
1522         LIST_HEAD(list);
1523
1524         spin_lock(&l_mg->close_lock);
1525         if (list_empty(&l_mg->emeta_list)) {
1526                 spin_unlock(&l_mg->close_lock);
1527                 return;
1528         }
1529
1530         list_cut_position(&list, &l_mg->emeta_list, l_mg->emeta_list.prev);
1531         spin_unlock(&l_mg->close_lock);
1532
1533         list_for_each_entry_safe(line, tline, &list, list) {
1534                 struct pblk_emeta *emeta = line->emeta;
1535
1536                 while (emeta->mem < lm->emeta_len[0]) {
1537                         int ret;
1538
1539                         ret = pblk_submit_meta_io(pblk, line);
1540                         if (ret) {
1541                                 pr_err("pblk: sync meta line %d failed (%d)\n",
1542                                                         line->id, ret);
1543                                 return;
1544                         }
1545                 }
1546         }
1547
1548         pblk_wait_for_meta(pblk);
1549 }
1550
1551 static void pblk_line_should_sync_meta(struct pblk *pblk)
1552 {
1553         if (pblk_rl_is_limit(&pblk->rl))
1554                 pblk_line_close_meta_sync(pblk);
1555 }
1556
1557 void pblk_line_close(struct pblk *pblk, struct pblk_line *line)
1558 {
1559         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1560         struct pblk_line_meta *lm = &pblk->lm;
1561         struct list_head *move_list;
1562
1563         WARN(!bitmap_full(line->map_bitmap, lm->sec_per_line),
1564                                 "pblk: corrupt closed line %d\n", line->id);
1565
1566         spin_lock(&l_mg->free_lock);
1567         WARN_ON(!test_and_clear_bit(line->meta_line, &l_mg->meta_bitmap));
1568         spin_unlock(&l_mg->free_lock);
1569
1570         spin_lock(&l_mg->gc_lock);
1571         spin_lock(&line->lock);
1572         WARN_ON(line->state != PBLK_LINESTATE_OPEN);
1573         line->state = PBLK_LINESTATE_CLOSED;
1574         move_list = pblk_line_gc_list(pblk, line);
1575
1576         list_add_tail(&line->list, move_list);
1577
1578         mempool_free(line->map_bitmap, pblk->line_meta_pool);
1579         line->map_bitmap = NULL;
1580         line->smeta = NULL;
1581         line->emeta = NULL;
1582
1583         spin_unlock(&line->lock);
1584         spin_unlock(&l_mg->gc_lock);
1585
1586         pblk_gc_should_kick(pblk);
1587 }
1588
1589 void pblk_line_close_meta(struct pblk *pblk, struct pblk_line *line)
1590 {
1591         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1592         struct pblk_line_meta *lm = &pblk->lm;
1593         struct pblk_emeta *emeta = line->emeta;
1594         struct line_emeta *emeta_buf = emeta->buf;
1595
1596         /* No need for exact vsc value; avoid a big line lock and take aprox. */
1597         memcpy(emeta_to_vsc(pblk, emeta_buf), l_mg->vsc_list, lm->vsc_list_len);
1598         memcpy(emeta_to_bb(emeta_buf), line->blk_bitmap, lm->blk_bitmap_len);
1599
1600         emeta_buf->nr_valid_lbas = cpu_to_le64(line->nr_valid_lbas);
1601         emeta_buf->crc = cpu_to_le32(pblk_calc_emeta_crc(pblk, emeta_buf));
1602
1603         spin_lock(&l_mg->close_lock);
1604         spin_lock(&line->lock);
1605         list_add_tail(&line->list, &l_mg->emeta_list);
1606         spin_unlock(&line->lock);
1607         spin_unlock(&l_mg->close_lock);
1608
1609         pblk_line_should_sync_meta(pblk);
1610 }
1611
1612 void pblk_line_close_ws(struct work_struct *work)
1613 {
1614         struct pblk_line_ws *line_ws = container_of(work, struct pblk_line_ws,
1615                                                                         ws);
1616         struct pblk *pblk = line_ws->pblk;
1617         struct pblk_line *line = line_ws->line;
1618
1619         pblk_line_close(pblk, line);
1620         mempool_free(line_ws, pblk->line_ws_pool);
1621 }
1622
1623 void pblk_line_mark_bb(struct work_struct *work)
1624 {
1625         struct pblk_line_ws *line_ws = container_of(work, struct pblk_line_ws,
1626                                                                         ws);
1627         struct pblk *pblk = line_ws->pblk;
1628         struct nvm_tgt_dev *dev = pblk->dev;
1629         struct ppa_addr *ppa = line_ws->priv;
1630         int ret;
1631
1632         ret = nvm_set_tgt_bb_tbl(dev, ppa, 1, NVM_BLK_T_GRWN_BAD);
1633         if (ret) {
1634                 struct pblk_line *line;
1635                 int pos;
1636
1637                 line = &pblk->lines[pblk_dev_ppa_to_line(*ppa)];
1638                 pos = pblk_dev_ppa_to_pos(&dev->geo, *ppa);
1639
1640                 pr_err("pblk: failed to mark bb, line:%d, pos:%d\n",
1641                                 line->id, pos);
1642         }
1643
1644         kfree(ppa);
1645         mempool_free(line_ws, pblk->line_ws_pool);
1646 }
1647
1648 void pblk_line_run_ws(struct pblk *pblk, struct pblk_line *line, void *priv,
1649                       void (*work)(struct work_struct *),
1650                       struct workqueue_struct *wq)
1651 {
1652         struct pblk_line_ws *line_ws;
1653
1654         line_ws = mempool_alloc(pblk->line_ws_pool, GFP_ATOMIC);
1655         if (!line_ws)
1656                 return;
1657
1658         line_ws->pblk = pblk;
1659         line_ws->line = line;
1660         line_ws->priv = priv;
1661
1662         INIT_WORK(&line_ws->ws, work);
1663         queue_work(wq, &line_ws->ws);
1664 }
1665
1666 void pblk_down_rq(struct pblk *pblk, struct ppa_addr *ppa_list, int nr_ppas,
1667                   unsigned long *lun_bitmap)
1668 {
1669         struct nvm_tgt_dev *dev = pblk->dev;
1670         struct nvm_geo *geo = &dev->geo;
1671         struct pblk_lun *rlun;
1672         int pos = pblk_ppa_to_pos(geo, ppa_list[0]);
1673         int ret;
1674
1675         /*
1676          * Only send one inflight I/O per LUN. Since we map at a page
1677          * granurality, all ppas in the I/O will map to the same LUN
1678          */
1679 #ifdef CONFIG_NVM_DEBUG
1680         int i;
1681
1682         for (i = 1; i < nr_ppas; i++)
1683                 WARN_ON(ppa_list[0].g.lun != ppa_list[i].g.lun ||
1684                                 ppa_list[0].g.ch != ppa_list[i].g.ch);
1685 #endif
1686         /* If the LUN has been locked for this same request, do no attempt to
1687          * lock it again
1688          */
1689         if (test_and_set_bit(pos, lun_bitmap))
1690                 return;
1691
1692         rlun = &pblk->luns[pos];
1693         ret = down_timeout(&rlun->wr_sem, msecs_to_jiffies(5000));
1694         if (ret) {
1695                 switch (ret) {
1696                 case -ETIME:
1697                         pr_err("pblk: lun semaphore timed out\n");
1698                         break;
1699                 case -EINTR:
1700                         pr_err("pblk: lun semaphore timed out\n");
1701                         break;
1702                 }
1703         }
1704 }
1705
1706 void pblk_up_rq(struct pblk *pblk, struct ppa_addr *ppa_list, int nr_ppas,
1707                 unsigned long *lun_bitmap)
1708 {
1709         struct nvm_tgt_dev *dev = pblk->dev;
1710         struct nvm_geo *geo = &dev->geo;
1711         struct pblk_lun *rlun;
1712         int nr_luns = geo->nr_luns;
1713         int bit = -1;
1714
1715         while ((bit = find_next_bit(lun_bitmap, nr_luns, bit + 1)) < nr_luns) {
1716                 rlun = &pblk->luns[bit];
1717                 up(&rlun->wr_sem);
1718         }
1719
1720         kfree(lun_bitmap);
1721 }
1722
1723 void pblk_update_map(struct pblk *pblk, sector_t lba, struct ppa_addr ppa)
1724 {
1725         struct ppa_addr l2p_ppa;
1726
1727         /* logic error: lba out-of-bounds. Ignore update */
1728         if (!(lba < pblk->rl.nr_secs)) {
1729                 WARN(1, "pblk: corrupted L2P map request\n");
1730                 return;
1731         }
1732
1733         spin_lock(&pblk->trans_lock);
1734         l2p_ppa = pblk_trans_map_get(pblk, lba);
1735
1736         if (!pblk_addr_in_cache(l2p_ppa) && !pblk_ppa_empty(l2p_ppa))
1737                 pblk_map_invalidate(pblk, l2p_ppa);
1738
1739         pblk_trans_map_set(pblk, lba, ppa);
1740         spin_unlock(&pblk->trans_lock);
1741 }
1742
1743 void pblk_update_map_cache(struct pblk *pblk, sector_t lba, struct ppa_addr ppa)
1744 {
1745 #ifdef CONFIG_NVM_DEBUG
1746         /* Callers must ensure that the ppa points to a cache address */
1747         BUG_ON(!pblk_addr_in_cache(ppa));
1748         BUG_ON(pblk_rb_pos_oob(&pblk->rwb, pblk_addr_to_cacheline(ppa)));
1749 #endif
1750
1751         pblk_update_map(pblk, lba, ppa);
1752 }
1753
1754 int pblk_update_map_gc(struct pblk *pblk, sector_t lba, struct ppa_addr ppa,
1755                        struct pblk_line *gc_line)
1756 {
1757         struct ppa_addr l2p_ppa;
1758         int ret = 1;
1759
1760 #ifdef CONFIG_NVM_DEBUG
1761         /* Callers must ensure that the ppa points to a cache address */
1762         BUG_ON(!pblk_addr_in_cache(ppa));
1763         BUG_ON(pblk_rb_pos_oob(&pblk->rwb, pblk_addr_to_cacheline(ppa)));
1764 #endif
1765
1766         /* logic error: lba out-of-bounds. Ignore update */
1767         if (!(lba < pblk->rl.nr_secs)) {
1768                 WARN(1, "pblk: corrupted L2P map request\n");
1769                 return 0;
1770         }
1771
1772         spin_lock(&pblk->trans_lock);
1773         l2p_ppa = pblk_trans_map_get(pblk, lba);
1774
1775         /* Prevent updated entries to be overwritten by GC */
1776         if (pblk_addr_in_cache(l2p_ppa) || pblk_ppa_empty(l2p_ppa) ||
1777                                 pblk_tgt_ppa_to_line(l2p_ppa) != gc_line->id) {
1778                 ret = 0;
1779                 goto out;
1780         }
1781
1782         pblk_trans_map_set(pblk, lba, ppa);
1783 out:
1784         spin_unlock(&pblk->trans_lock);
1785         return ret;
1786 }
1787
1788 void pblk_update_map_dev(struct pblk *pblk, sector_t lba, struct ppa_addr ppa,
1789                          struct ppa_addr entry_line)
1790 {
1791         struct ppa_addr l2p_line;
1792
1793 #ifdef CONFIG_NVM_DEBUG
1794         /* Callers must ensure that the ppa points to a device address */
1795         BUG_ON(pblk_addr_in_cache(ppa));
1796 #endif
1797         /* Invalidate and discard padded entries */
1798         if (lba == ADDR_EMPTY) {
1799 #ifdef CONFIG_NVM_DEBUG
1800                 atomic_long_inc(&pblk->padded_wb);
1801 #endif
1802                 pblk_map_invalidate(pblk, ppa);
1803                 return;
1804         }
1805
1806         /* logic error: lba out-of-bounds. Ignore update */
1807         if (!(lba < pblk->rl.nr_secs)) {
1808                 WARN(1, "pblk: corrupted L2P map request\n");
1809                 return;
1810         }
1811
1812         spin_lock(&pblk->trans_lock);
1813         l2p_line = pblk_trans_map_get(pblk, lba);
1814
1815         /* Do not update L2P if the cacheline has been updated. In this case,
1816          * the mapped ppa must be invalidated
1817          */
1818         if (l2p_line.ppa != entry_line.ppa) {
1819                 if (!pblk_ppa_empty(ppa))
1820                         pblk_map_invalidate(pblk, ppa);
1821                 goto out;
1822         }
1823
1824 #ifdef CONFIG_NVM_DEBUG
1825         WARN_ON(!pblk_addr_in_cache(l2p_line) && !pblk_ppa_empty(l2p_line));
1826 #endif
1827
1828         pblk_trans_map_set(pblk, lba, ppa);
1829 out:
1830         spin_unlock(&pblk->trans_lock);
1831 }
1832
1833 void pblk_lookup_l2p_seq(struct pblk *pblk, struct ppa_addr *ppas,
1834                          sector_t blba, int nr_secs)
1835 {
1836         int i;
1837
1838         spin_lock(&pblk->trans_lock);
1839         for (i = 0; i < nr_secs; i++)
1840                 ppas[i] = pblk_trans_map_get(pblk, blba + i);
1841         spin_unlock(&pblk->trans_lock);
1842 }
1843
1844 void pblk_lookup_l2p_rand(struct pblk *pblk, struct ppa_addr *ppas,
1845                           u64 *lba_list, int nr_secs)
1846 {
1847         sector_t lba;
1848         int i;
1849
1850         spin_lock(&pblk->trans_lock);
1851         for (i = 0; i < nr_secs; i++) {
1852                 lba = lba_list[i];
1853                 if (lba == ADDR_EMPTY) {
1854                         ppas[i].ppa = ADDR_EMPTY;
1855                 } else {
1856                         /* logic error: lba out-of-bounds. Ignore update */
1857                         if (!(lba < pblk->rl.nr_secs)) {
1858                                 WARN(1, "pblk: corrupted L2P map request\n");
1859                                 continue;
1860                         }
1861                         ppas[i] = pblk_trans_map_get(pblk, lba);
1862                 }
1863         }
1864         spin_unlock(&pblk->trans_lock);
1865 }