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                               int alloc_type, gfp_t gfp_mask)
 429 {
 430         struct nvm_tgt_dev *dev = pblk->dev;
 431         void *kaddr = data;
 432         struct page *page;
 433         struct bio *bio;
 434         int i, ret;
 435
 436         if (alloc_type == PBLK_KMALLOC_META)
 437                 return bio_map_kern(dev->q, kaddr, len, gfp_mask);
 438
 439         bio = bio_kmalloc(gfp_mask, nr_secs);
 440         if (!bio)
 441                 return ERR_PTR(-ENOMEM);
 442
 443         for (i = 0; i < nr_secs; i++) {
 444                 page = vmalloc_to_page(kaddr);
 445                 if (!page) {
 446                         pr_err("pblk: could not map vmalloc bio\n");
 447                         bio_put(bio);
 448                         bio = ERR_PTR(-ENOMEM);
 449                         goto out;
 450                 }
 451
 452                 ret = bio_add_pc_page(dev->q, bio, page, PAGE_SIZE, 0);
 453                 if (ret != PAGE_SIZE) {
 454                         pr_err("pblk: could not add page to bio\n");
 455                         bio_put(bio);
 456                         bio = ERR_PTR(-ENOMEM);
 457                         goto out;
 458                 }
 459
 460                 kaddr += PAGE_SIZE;
 461         }
 462 out:
 463         return bio;
 464 }
 465
 466 int pblk_calc_secs(struct pblk *pblk, unsigned long secs_avail,
 467                    unsigned long secs_to_flush)
 468 {
 469         int max = pblk->sec_per_write;
 470         int min = pblk->min_write_pgs;
 471         int secs_to_sync = 0;
 472
 473         if (secs_avail >= max)
 474                 secs_to_sync = max;
 475         else if (secs_avail >= min)
 476                 secs_to_sync = min * (secs_avail / min);
 477         else if (secs_to_flush)
 478                 secs_to_sync = min;
 479
 480         return secs_to_sync;
 481 }
 482
 483 void pblk_dealloc_page(struct pblk *pblk, struct pblk_line *line, int nr_secs)
 484 {
 485         u64 addr;
 486         int i;
 487
 488         addr = find_next_zero_bit(line->map_bitmap,
 489                                         pblk->lm.sec_per_line, line->cur_sec);
 490         line->cur_sec = addr - nr_secs;
 491
 492         for (i = 0; i < nr_secs; i++, line->cur_sec--)
 493                 WARN_ON(!test_and_clear_bit(line->cur_sec, line->map_bitmap));
 494 }
 495
 496 u64 __pblk_alloc_page(struct pblk *pblk, struct pblk_line *line, int nr_secs)
 497 {
 498         u64 addr;
 499         int i;
 500
 501         lockdep_assert_held(&line->lock);
 502
 503         /* logic error: ppa out-of-bounds. Prevent generating bad address */
 504         if (line->cur_sec + nr_secs > pblk->lm.sec_per_line) {
 505                 WARN(1, "pblk: page allocation out of bounds\n");
 506                 nr_secs = pblk->lm.sec_per_line - line->cur_sec;
 507         }
 508
 509         line->cur_sec = addr = find_next_zero_bit(line->map_bitmap,
 510                                         pblk->lm.sec_per_line, line->cur_sec);
 511         for (i = 0; i < nr_secs; i++, line->cur_sec++)
 512                 WARN_ON(test_and_set_bit(line->cur_sec, line->map_bitmap));
 513
 514         return addr;
 515 }
 516
 517 u64 pblk_alloc_page(struct pblk *pblk, struct pblk_line *line, int nr_secs)
 518 {
 519         u64 addr;
 520
 521         /* Lock needed in case a write fails and a recovery needs to remap
 522          * failed write buffer entries
 523          */
 524         spin_lock(&line->lock);
 525         addr = __pblk_alloc_page(pblk, line, nr_secs);
 526         line->left_msecs -= nr_secs;
 527         WARN(line->left_msecs < 0, "pblk: page allocation out of bounds\n");
 528         spin_unlock(&line->lock);
 529
 530         return addr;
 531 }
 532
 533 u64 pblk_lookup_page(struct pblk *pblk, struct pblk_line *line)
 534 {
 535         u64 paddr;
 536
 537         spin_lock(&line->lock);
 538         paddr = find_next_zero_bit(line->map_bitmap,
 539                                         pblk->lm.sec_per_line, line->cur_sec);
 540         spin_unlock(&line->lock);
 541
 542         return paddr;
 543 }
 544
 545 /*
 546  * Submit emeta to one LUN in the raid line at the time to avoid a deadlock when
 547  * taking the per LUN semaphore.
 548  */
 549 static int pblk_line_submit_emeta_io(struct pblk *pblk, struct pblk_line *line,
 550                                      void *emeta_buf, u64 paddr, int dir)
 551 {
 552         struct nvm_tgt_dev *dev = pblk->dev;
 553         struct nvm_geo *geo = &dev->geo;
 554         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
 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,
 593                                         l_mg->emeta_alloc_type, GFP_KERNEL);
 594         if (IS_ERR(bio)) {
 595                 ret = PTR_ERR(bio);
 596                 goto free_rqd_dma;
 597         }
 598
 599         bio->bi_iter.bi_sector = 0; /* internal bio */
 600         bio_set_op_attrs(bio, bio_op, 0);
 601
 602         rqd.bio = bio;
 603         rqd.meta_list = meta_list;
 604         rqd.ppa_list = ppa_list;
 605         rqd.dma_meta_list = dma_meta_list;
 606         rqd.dma_ppa_list = dma_ppa_list;
 607         rqd.opcode = cmd_op;
 608         rqd.nr_ppas = rq_ppas;
 609         rqd.end_io = pblk_end_io_sync;
 610         rqd.private = &wait;
 611
 612         if (dir == WRITE) {
 613                 struct pblk_sec_meta *meta_list = rqd.meta_list;
 614
 615                 rqd.flags = pblk_set_progr_mode(pblk, WRITE);
 616                 for (i = 0; i < rqd.nr_ppas; ) {
 617                         spin_lock(&line->lock);
 618                         paddr = __pblk_alloc_page(pblk, line, min);
 619                         spin_unlock(&line->lock);
 620                         for (j = 0; j < min; j++, i++, paddr++) {
 621                                 meta_list[i].lba = cpu_to_le64(ADDR_EMPTY);
 622                                 rqd.ppa_list[i] =
 623                                         addr_to_gen_ppa(pblk, paddr, id);
 624                         }
 625                 }
 626         } else {
 627                 for (i = 0; i < rqd.nr_ppas; ) {
 628                         struct ppa_addr ppa = addr_to_gen_ppa(pblk, paddr, id);
 629                         int pos = pblk_dev_ppa_to_pos(geo, ppa);
 630                         int read_type = PBLK_READ_RANDOM;
 631
 632                         if (pblk_io_aligned(pblk, rq_ppas))
 633                                 read_type = PBLK_READ_SEQUENTIAL;
 634                         rqd.flags = pblk_set_read_mode(pblk, read_type);
 635
 636                         while (test_bit(pos, line->blk_bitmap)) {
 637                                 paddr += min;
 638                                 if (pblk_boundary_paddr_checks(pblk, paddr)) {
 639                                         pr_err("pblk: corrupt emeta line:%d\n",
 640                                                                 line->id);
 641                                         bio_put(bio);
 642                                         ret = -EINTR;
 643                                         goto free_rqd_dma;
 644                                 }
 645
 646                                 ppa = addr_to_gen_ppa(pblk, paddr, id);
 647                                 pos = pblk_dev_ppa_to_pos(geo, ppa);
 648                         }
 649
 650                         if (pblk_boundary_paddr_checks(pblk, paddr + min)) {
 651                                 pr_err("pblk: corrupt emeta line:%d\n",
 652                                                                 line->id);
 653                                 bio_put(bio);
 654                                 ret = -EINTR;
 655                                 goto free_rqd_dma;
 656                         }
 657
 658                         for (j = 0; j < min; j++, i++, paddr++)
 659                                 rqd.ppa_list[i] =
 660                                         addr_to_gen_ppa(pblk, paddr, line->id);
 661                 }
 662         }
 663
 664         ret = pblk_submit_io(pblk, &rqd);
 665         if (ret) {
 666                 pr_err("pblk: emeta I/O submission failed: %d\n", ret);
 667                 bio_put(bio);
 668                 goto free_rqd_dma;
 669         }
 670
 671         if (!wait_for_completion_io_timeout(&wait,
 672                                 msecs_to_jiffies(PBLK_COMMAND_TIMEOUT_MS))) {
 673                 pr_err("pblk: emeta I/O timed out\n");
 674         }
 675         atomic_dec(&pblk->inflight_io);
 676         reinit_completion(&wait);
 677
 678         if (likely(pblk->l_mg.emeta_alloc_type == PBLK_VMALLOC_META))
 679                 bio_put(bio);
 680
 681         if (rqd.error) {
 682                 if (dir == WRITE)
 683                         pblk_log_write_err(pblk, &rqd);
 684                 else
 685                         pblk_log_read_err(pblk, &rqd);
 686         }
 687
 688         emeta_buf += rq_len;
 689         left_ppas -= rq_ppas;
 690         if (left_ppas)
 691                 goto next_rq;
 692 free_rqd_dma:
 693         nvm_dev_dma_free(dev->parent, rqd.meta_list, rqd.dma_meta_list);
 694         return ret;
 695 }
 696
 697 u64 pblk_line_smeta_start(struct pblk *pblk, struct pblk_line *line)
 698 {
 699         struct nvm_tgt_dev *dev = pblk->dev;
 700         struct nvm_geo *geo = &dev->geo;
 701         struct pblk_line_meta *lm = &pblk->lm;
 702         int bit;
 703
 704         /* This usually only happens on bad lines */
 705         bit = find_first_zero_bit(line->blk_bitmap, lm->blk_per_line);
 706         if (bit >= lm->blk_per_line)
 707                 return -1;
 708
 709         return bit * geo->sec_per_pl;
 710 }
 711
 712 static int pblk_line_submit_smeta_io(struct pblk *pblk, struct pblk_line *line,
 713                                      u64 paddr, int dir)
 714 {
 715         struct nvm_tgt_dev *dev = pblk->dev;
 716         struct pblk_line_meta *lm = &pblk->lm;
 717         struct bio *bio;
 718         struct nvm_rq rqd;
 719         __le64 *lba_list = NULL;
 720         int i, ret;
 721         int cmd_op, bio_op;
 722         int flags;
 723         DECLARE_COMPLETION_ONSTACK(wait);
 724
 725         if (dir == WRITE) {
 726                 bio_op = REQ_OP_WRITE;
 727                 cmd_op = NVM_OP_PWRITE;
 728                 flags = pblk_set_progr_mode(pblk, WRITE);
 729                 lba_list = emeta_to_lbas(pblk, line->emeta->buf);
 730         } else if (dir == READ) {
 731                 bio_op = REQ_OP_READ;
 732                 cmd_op = NVM_OP_PREAD;
 733                 flags = pblk_set_read_mode(pblk, PBLK_READ_SEQUENTIAL);
 734         } else
 735                 return -EINVAL;
 736
 737         memset(&rqd, 0, sizeof(struct nvm_rq));
 738
 739         rqd.meta_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
 740                                                         &rqd.dma_meta_list);
 741         if (!rqd.meta_list)
 742                 return -ENOMEM;
 743
 744         rqd.ppa_list = rqd.meta_list + pblk_dma_meta_size;
 745         rqd.dma_ppa_list = rqd.dma_meta_list + pblk_dma_meta_size;
 746
 747         bio = bio_map_kern(dev->q, line->smeta, lm->smeta_len, GFP_KERNEL);
 748         if (IS_ERR(bio)) {
 749                 ret = PTR_ERR(bio);
 750                 goto free_ppa_list;
 751         }
 752
 753         bio->bi_iter.bi_sector = 0; /* internal bio */
 754         bio_set_op_attrs(bio, bio_op, 0);
 755
 756         rqd.bio = bio;
 757         rqd.opcode = cmd_op;
 758         rqd.flags = flags;
 759         rqd.nr_ppas = lm->smeta_sec;
 760         rqd.end_io = pblk_end_io_sync;
 761         rqd.private = &wait;
 762
 763         for (i = 0; i < lm->smeta_sec; i++, paddr++) {
 764                 struct pblk_sec_meta *meta_list = rqd.meta_list;
 765
 766                 rqd.ppa_list[i] = addr_to_gen_ppa(pblk, paddr, line->id);
 767
 768                 if (dir == WRITE) {
 769                         __le64 addr_empty = cpu_to_le64(ADDR_EMPTY);
 770
 771                         meta_list[i].lba = lba_list[paddr] = addr_empty;
 772                 }
 773         }
 774
 775         /*
 776          * This I/O is sent by the write thread when a line is replace. Since
 777          * the write thread is the only one sending write and erase commands,
 778          * there is no need to take the LUN semaphore.
 779          */
 780         ret = pblk_submit_io(pblk, &rqd);
 781         if (ret) {
 782                 pr_err("pblk: smeta I/O submission failed: %d\n", ret);
 783                 bio_put(bio);
 784                 goto free_ppa_list;
 785         }
 786
 787         if (!wait_for_completion_io_timeout(&wait,
 788                                 msecs_to_jiffies(PBLK_COMMAND_TIMEOUT_MS))) {
 789                 pr_err("pblk: smeta I/O timed out\n");
 790         }
 791         atomic_dec(&pblk->inflight_io);
 792
 793         if (rqd.error) {
 794                 if (dir == WRITE)
 795                         pblk_log_write_err(pblk, &rqd);
 796                 else
 797                         pblk_log_read_err(pblk, &rqd);
 798         }
 799
 800 free_ppa_list:
 801         nvm_dev_dma_free(dev->parent, rqd.meta_list, rqd.dma_meta_list);
 802
 803         return ret;
 804 }
 805
 806 int pblk_line_read_smeta(struct pblk *pblk, struct pblk_line *line)
 807 {
 808         u64 bpaddr = pblk_line_smeta_start(pblk, line);
 809
 810         return pblk_line_submit_smeta_io(pblk, line, bpaddr, READ);
 811 }
 812
 813 int pblk_line_read_emeta(struct pblk *pblk, struct pblk_line *line,
 814                          void *emeta_buf)
 815 {
 816         return pblk_line_submit_emeta_io(pblk, line, emeta_buf,
 817                                                 line->emeta_ssec, READ);
 818 }
 819
 820 static void pblk_setup_e_rq(struct pblk *pblk, struct nvm_rq *rqd,
 821                             struct ppa_addr ppa)
 822 {
 823         rqd->opcode = NVM_OP_ERASE;
 824         rqd->ppa_addr = ppa;
 825         rqd->nr_ppas = 1;
 826         rqd->flags = pblk_set_progr_mode(pblk, ERASE);
 827         rqd->bio = NULL;
 828 }
 829
 830 static int pblk_blk_erase_sync(struct pblk *pblk, struct ppa_addr ppa)
 831 {
 832         struct nvm_rq rqd;
 833         int ret = 0;
 834         DECLARE_COMPLETION_ONSTACK(wait);
 835
 836         memset(&rqd, 0, sizeof(struct nvm_rq));
 837
 838         pblk_setup_e_rq(pblk, &rqd, ppa);
 839
 840         rqd.end_io = pblk_end_io_sync;
 841         rqd.private = &wait;
 842
 843         /* The write thread schedules erases so that it minimizes disturbances
 844          * with writes. Thus, there is no need to take the LUN semaphore.
 845          */
 846         ret = pblk_submit_io(pblk, &rqd);
 847         if (ret) {
 848                 struct nvm_tgt_dev *dev = pblk->dev;
 849                 struct nvm_geo *geo = &dev->geo;
 850
 851                 pr_err("pblk: could not sync erase line:%d,blk:%d\n",
 852                                         pblk_dev_ppa_to_line(ppa),
 853                                         pblk_dev_ppa_to_pos(geo, ppa));
 854
 855                 rqd.error = ret;
 856                 goto out;
 857         }
 858
 859         if (!wait_for_completion_io_timeout(&wait,
 860                                 msecs_to_jiffies(PBLK_COMMAND_TIMEOUT_MS))) {
 861                 pr_err("pblk: sync erase timed out\n");
 862         }
 863
 864 out:
 865         rqd.private = pblk;
 866         __pblk_end_io_erase(pblk, &rqd);
 867
 868         return ret;
 869 }
 870
 871 int pblk_line_erase(struct pblk *pblk, struct pblk_line *line)
 872 {
 873         struct pblk_line_meta *lm = &pblk->lm;
 874         struct ppa_addr ppa;
 875         int ret, bit = -1;
 876
 877         /* Erase only good blocks, one at a time */
 878         do {
 879                 spin_lock(&line->lock);
 880                 bit = find_next_zero_bit(line->erase_bitmap, lm->blk_per_line,
 881                                                                 bit + 1);
 882                 if (bit >= lm->blk_per_line) {
 883                         spin_unlock(&line->lock);
 884                         break;
 885                 }
 886
 887                 ppa = pblk->luns[bit].bppa; /* set ch and lun */
 888                 ppa.g.blk = line->id;
 889
 890                 atomic_dec(&line->left_eblks);
 891                 WARN_ON(test_and_set_bit(bit, line->erase_bitmap));
 892                 spin_unlock(&line->lock);
 893
 894                 ret = pblk_blk_erase_sync(pblk, ppa);
 895                 if (ret) {
 896                         pr_err("pblk: failed to erase line %d\n", line->id);
 897                         return ret;
 898                 }
 899         } while (1);
 900
 901         return 0;
 902 }
 903
 904 static void pblk_line_setup_metadata(struct pblk_line *line,
 905                                      struct pblk_line_mgmt *l_mg,
 906                                      struct pblk_line_meta *lm)
 907 {
 908         int meta_line;
 909
 910         lockdep_assert_held(&l_mg->free_lock);
 911
 912 retry_meta:
 913         meta_line = find_first_zero_bit(&l_mg->meta_bitmap, PBLK_DATA_LINES);
 914         if (meta_line == PBLK_DATA_LINES) {
 915                 spin_unlock(&l_mg->free_lock);
 916                 io_schedule();
 917                 spin_lock(&l_mg->free_lock);
 918                 goto retry_meta;
 919         }
 920
 921         set_bit(meta_line, &l_mg->meta_bitmap);
 922         line->meta_line = meta_line;
 923
 924         line->smeta = l_mg->sline_meta[meta_line];
 925         line->emeta = l_mg->eline_meta[meta_line];
 926
 927         memset(line->smeta, 0, lm->smeta_len);
 928         memset(line->emeta->buf, 0, lm->emeta_len[0]);
 929
 930         line->emeta->mem = 0;
 931         atomic_set(&line->emeta->sync, 0);
 932 }
 933
 934 /* For now lines are always assumed full lines. Thus, smeta former and current
 935  * lun bitmaps are omitted.
 936  */
 937 static int pblk_line_init_metadata(struct pblk *pblk, struct pblk_line *line,
 938                                   struct pblk_line *cur)
 939 {
 940         struct nvm_tgt_dev *dev = pblk->dev;
 941         struct nvm_geo *geo = &dev->geo;
 942         struct pblk_line_meta *lm = &pblk->lm;
 943         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
 944         struct pblk_emeta *emeta = line->emeta;
 945         struct line_emeta *emeta_buf = emeta->buf;
 946         struct line_smeta *smeta_buf = (struct line_smeta *)line->smeta;
 947         int nr_blk_line;
 948
 949         /* After erasing the line, new bad blocks might appear and we risk
 950          * having an invalid line
 951          */
 952         nr_blk_line = lm->blk_per_line -
 953                         bitmap_weight(line->blk_bitmap, lm->blk_per_line);
 954         if (nr_blk_line < lm->min_blk_line) {
 955                 spin_lock(&l_mg->free_lock);
 956                 spin_lock(&line->lock);
 957                 line->state = PBLK_LINESTATE_BAD;
 958                 spin_unlock(&line->lock);
 959
 960                 list_add_tail(&line->list, &l_mg->bad_list);
 961                 spin_unlock(&l_mg->free_lock);
 962
 963                 pr_debug("pblk: line %d is bad\n", line->id);
 964
 965                 return 0;
 966         }
 967
 968         /* Run-time metadata */
 969         line->lun_bitmap = ((void *)(smeta_buf)) + sizeof(struct line_smeta);
 970
 971         /* Mark LUNs allocated in this line (all for now) */
 972         bitmap_set(line->lun_bitmap, 0, lm->lun_bitmap_len);
 973
 974         smeta_buf->header.identifier = cpu_to_le32(PBLK_MAGIC);
 975         memcpy(smeta_buf->header.uuid, pblk->instance_uuid, 16);
 976         smeta_buf->header.id = cpu_to_le32(line->id);
 977         smeta_buf->header.type = cpu_to_le16(line->type);
 978         smeta_buf->header.version = cpu_to_le16(1);
 979
 980         /* Start metadata */
 981         smeta_buf->seq_nr = cpu_to_le64(line->seq_nr);
 982         smeta_buf->window_wr_lun = cpu_to_le32(geo->nr_luns);
 983
 984         /* Fill metadata among lines */
 985         if (cur) {
 986                 memcpy(line->lun_bitmap, cur->lun_bitmap, lm->lun_bitmap_len);
 987                 smeta_buf->prev_id = cpu_to_le32(cur->id);
 988                 cur->emeta->buf->next_id = cpu_to_le32(line->id);
 989         } else {
 990                 smeta_buf->prev_id = cpu_to_le32(PBLK_LINE_EMPTY);
 991         }
 992
 993         /* All smeta must be set at this point */
 994         smeta_buf->header.crc = cpu_to_le32(
 995                         pblk_calc_meta_header_crc(pblk, &smeta_buf->header));
 996         smeta_buf->crc = cpu_to_le32(pblk_calc_smeta_crc(pblk, smeta_buf));
 997
 998         /* End metadata */
 999         memcpy(&emeta_buf->header, &smeta_buf->header,
1000                                                 sizeof(struct line_header));
1001         emeta_buf->seq_nr = cpu_to_le64(line->seq_nr);
1002         emeta_buf->nr_lbas = cpu_to_le64(line->sec_in_line);
1003         emeta_buf->nr_valid_lbas = cpu_to_le64(0);
1004         emeta_buf->next_id = cpu_to_le32(PBLK_LINE_EMPTY);
1005         emeta_buf->crc = cpu_to_le32(0);
1006         emeta_buf->prev_id = smeta_buf->prev_id;
1007
1008         return 1;
1009 }
1010
1011 /* For now lines are always assumed full lines. Thus, smeta former and current
1012  * lun bitmaps are omitted.
1013  */
1014 static int pblk_line_init_bb(struct pblk *pblk, struct pblk_line *line,
1015                              int init)
1016 {
1017         struct nvm_tgt_dev *dev = pblk->dev;
1018         struct nvm_geo *geo = &dev->geo;
1019         struct pblk_line_meta *lm = &pblk->lm;
1020         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1021         int nr_bb = 0;
1022         u64 off;
1023         int bit = -1;
1024
1025         line->sec_in_line = lm->sec_per_line;
1026
1027         /* Capture bad block information on line mapping bitmaps */
1028         while ((bit = find_next_bit(line->blk_bitmap, lm->blk_per_line,
1029                                         bit + 1)) < lm->blk_per_line) {
1030                 off = bit * geo->sec_per_pl;
1031                 bitmap_shift_left(l_mg->bb_aux, l_mg->bb_template, off,
1032                                                         lm->sec_per_line);
1033                 bitmap_or(line->map_bitmap, line->map_bitmap, l_mg->bb_aux,
1034                                                         lm->sec_per_line);
1035                 line->sec_in_line -= geo->sec_per_blk;
1036                 if (bit >= lm->emeta_bb)
1037                         nr_bb++;
1038         }
1039
1040         /* Mark smeta metadata sectors as bad sectors */
1041         bit = find_first_zero_bit(line->blk_bitmap, lm->blk_per_line);
1042         off = bit * geo->sec_per_pl;
1043         bitmap_set(line->map_bitmap, off, lm->smeta_sec);
1044         line->sec_in_line -= lm->smeta_sec;
1045         line->smeta_ssec = off;
1046         line->cur_sec = off + lm->smeta_sec;
1047
1048         if (init && pblk_line_submit_smeta_io(pblk, line, off, WRITE)) {
1049                 pr_debug("pblk: line smeta I/O failed. Retry\n");
1050                 return 1;
1051         }
1052
1053         bitmap_copy(line->invalid_bitmap, line->map_bitmap, lm->sec_per_line);
1054
1055         /* Mark emeta metadata sectors as bad sectors. We need to consider bad
1056          * blocks to make sure that there are enough sectors to store emeta
1057          */
1058         bit = lm->sec_per_line;
1059         off = lm->sec_per_line - lm->emeta_sec[0];
1060         bitmap_set(line->invalid_bitmap, off, lm->emeta_sec[0]);
1061         while (nr_bb) {
1062                 off -= geo->sec_per_pl;
1063                 if (!test_bit(off, line->invalid_bitmap)) {
1064                         bitmap_set(line->invalid_bitmap, off, geo->sec_per_pl);
1065                         nr_bb--;
1066                 }
1067         }
1068
1069         line->sec_in_line -= lm->emeta_sec[0];
1070         line->emeta_ssec = off;
1071         line->nr_valid_lbas = 0;
1072         line->left_msecs = line->sec_in_line;
1073         *line->vsc = cpu_to_le32(line->sec_in_line);
1074
1075         if (lm->sec_per_line - line->sec_in_line !=
1076                 bitmap_weight(line->invalid_bitmap, lm->sec_per_line)) {
1077                 spin_lock(&line->lock);
1078                 line->state = PBLK_LINESTATE_BAD;
1079                 spin_unlock(&line->lock);
1080
1081                 list_add_tail(&line->list, &l_mg->bad_list);
1082                 pr_err("pblk: unexpected line %d is bad\n", line->id);
1083
1084                 return 0;
1085         }
1086
1087         return 1;
1088 }
1089
1090 static int pblk_line_prepare(struct pblk *pblk, struct pblk_line *line)
1091 {
1092         struct pblk_line_meta *lm = &pblk->lm;
1093         int blk_in_line = atomic_read(&line->blk_in_line);
1094
1095         line->map_bitmap = mempool_alloc(pblk->line_meta_pool, GFP_ATOMIC);
1096         if (!line->map_bitmap)
1097                 return -ENOMEM;
1098         memset(line->map_bitmap, 0, lm->sec_bitmap_len);
1099
1100         /* invalid_bitmap is special since it is used when line is closed. No
1101          * need to zeroized; it will be initialized using bb info form
1102          * map_bitmap
1103          */
1104         line->invalid_bitmap = mempool_alloc(pblk->line_meta_pool, GFP_ATOMIC);
1105         if (!line->invalid_bitmap) {
1106                 mempool_free(line->map_bitmap, pblk->line_meta_pool);
1107                 return -ENOMEM;
1108         }
1109
1110         spin_lock(&line->lock);
1111         if (line->state != PBLK_LINESTATE_FREE) {
1112                 mempool_free(line->invalid_bitmap, pblk->line_meta_pool);
1113                 mempool_free(line->map_bitmap, pblk->line_meta_pool);
1114                 spin_unlock(&line->lock);
1115                 WARN(1, "pblk: corrupted line %d, state %d\n",
1116                                                         line->id, line->state);
1117                 return -EAGAIN;
1118         }
1119
1120         line->state = PBLK_LINESTATE_OPEN;
1121
1122         atomic_set(&line->left_eblks, blk_in_line);
1123         atomic_set(&line->left_seblks, blk_in_line);
1124
1125         line->meta_distance = lm->meta_distance;
1126         spin_unlock(&line->lock);
1127
1128         /* Bad blocks do not need to be erased */
1129         bitmap_copy(line->erase_bitmap, line->blk_bitmap, lm->blk_per_line);
1130
1131         kref_init(&line->ref);
1132
1133         return 0;
1134 }
1135
1136 int pblk_line_recov_alloc(struct pblk *pblk, struct pblk_line *line)
1137 {
1138         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1139         int ret;
1140
1141         spin_lock(&l_mg->free_lock);
1142         l_mg->data_line = line;
1143         list_del(&line->list);
1144
1145         ret = pblk_line_prepare(pblk, line);
1146         if (ret) {
1147                 list_add(&line->list, &l_mg->free_list);
1148                 spin_unlock(&l_mg->free_lock);
1149                 return ret;
1150         }
1151         spin_unlock(&l_mg->free_lock);
1152
1153         pblk_rl_free_lines_dec(&pblk->rl, line);
1154
1155         if (!pblk_line_init_bb(pblk, line, 0)) {
1156                 list_add(&line->list, &l_mg->free_list);
1157                 return -EINTR;
1158         }
1159
1160         return 0;
1161 }
1162
1163 void pblk_line_recov_close(struct pblk *pblk, struct pblk_line *line)
1164 {
1165         mempool_free(line->map_bitmap, pblk->line_meta_pool);
1166         line->map_bitmap = NULL;
1167         line->smeta = NULL;
1168         line->emeta = NULL;
1169 }
1170
1171 struct pblk_line *pblk_line_get(struct pblk *pblk)
1172 {
1173         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1174         struct pblk_line_meta *lm = &pblk->lm;
1175         struct pblk_line *line;
1176         int ret, bit;
1177
1178         lockdep_assert_held(&l_mg->free_lock);
1179
1180 retry:
1181         if (list_empty(&l_mg->free_list)) {
1182                 pr_err("pblk: no free lines\n");
1183                 return NULL;
1184         }
1185
1186         line = list_first_entry(&l_mg->free_list, struct pblk_line, list);
1187         list_del(&line->list);
1188         l_mg->nr_free_lines--;
1189
1190         bit = find_first_zero_bit(line->blk_bitmap, lm->blk_per_line);
1191         if (unlikely(bit >= lm->blk_per_line)) {
1192                 spin_lock(&line->lock);
1193                 line->state = PBLK_LINESTATE_BAD;
1194                 spin_unlock(&line->lock);
1195
1196                 list_add_tail(&line->list, &l_mg->bad_list);
1197
1198                 pr_debug("pblk: line %d is bad\n", line->id);
1199                 goto retry;
1200         }
1201
1202         ret = pblk_line_prepare(pblk, line);
1203         if (ret) {
1204                 if (ret == -EAGAIN) {
1205                         list_add(&line->list, &l_mg->corrupt_list);
1206                         goto retry;
1207                 } else {
1208                         pr_err("pblk: failed to prepare line %d\n", line->id);
1209                         list_add(&line->list, &l_mg->free_list);
1210                         l_mg->nr_free_lines++;
1211                         return NULL;
1212                 }
1213         }
1214
1215         return line;
1216 }
1217
1218 static struct pblk_line *pblk_line_retry(struct pblk *pblk,
1219                                          struct pblk_line *line)
1220 {
1221         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1222         struct pblk_line *retry_line;
1223
1224 retry:
1225         spin_lock(&l_mg->free_lock);
1226         retry_line = pblk_line_get(pblk);
1227         if (!retry_line) {
1228                 l_mg->data_line = NULL;
1229                 spin_unlock(&l_mg->free_lock);
1230                 return NULL;
1231         }
1232
1233         retry_line->smeta = line->smeta;
1234         retry_line->emeta = line->emeta;
1235         retry_line->meta_line = line->meta_line;
1236
1237         pblk_line_free(pblk, line);
1238         l_mg->data_line = retry_line;
1239         spin_unlock(&l_mg->free_lock);
1240
1241         pblk_rl_free_lines_dec(&pblk->rl, retry_line);
1242
1243         if (pblk_line_erase(pblk, retry_line))
1244                 goto retry;
1245
1246         return retry_line;
1247 }
1248
1249 static void pblk_set_space_limit(struct pblk *pblk)
1250 {
1251         struct pblk_rl *rl = &pblk->rl;
1252
1253         atomic_set(&rl->rb_space, 0);
1254 }
1255
1256 struct pblk_line *pblk_line_get_first_data(struct pblk *pblk)
1257 {
1258         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1259         struct pblk_line *line;
1260         int is_next = 0;
1261
1262         spin_lock(&l_mg->free_lock);
1263         line = pblk_line_get(pblk);
1264         if (!line) {
1265                 spin_unlock(&l_mg->free_lock);
1266                 return NULL;
1267         }
1268
1269         line->seq_nr = l_mg->d_seq_nr++;
1270         line->type = PBLK_LINETYPE_DATA;
1271         l_mg->data_line = line;
1272
1273         pblk_line_setup_metadata(line, l_mg, &pblk->lm);
1274
1275         /* Allocate next line for preparation */
1276         l_mg->data_next = pblk_line_get(pblk);
1277         if (!l_mg->data_next) {
1278                 /* If we cannot get a new line, we need to stop the pipeline.
1279                  * Only allow as many writes in as we can store safely and then
1280                  * fail gracefully
1281                  */
1282                 pblk_set_space_limit(pblk);
1283
1284                 l_mg->data_next = NULL;
1285         } else {
1286                 l_mg->data_next->seq_nr = l_mg->d_seq_nr++;
1287                 l_mg->data_next->type = PBLK_LINETYPE_DATA;
1288                 is_next = 1;
1289         }
1290         spin_unlock(&l_mg->free_lock);
1291
1292         if (pblk_line_erase(pblk, line)) {
1293                 line = pblk_line_retry(pblk, line);
1294                 if (!line)
1295                         return NULL;
1296         }
1297
1298         pblk_rl_free_lines_dec(&pblk->rl, line);
1299         if (is_next)
1300                 pblk_rl_free_lines_dec(&pblk->rl, l_mg->data_next);
1301
1302 retry_setup:
1303         if (!pblk_line_init_metadata(pblk, line, NULL)) {
1304                 line = pblk_line_retry(pblk, line);
1305                 if (!line)
1306                         return NULL;
1307
1308                 goto retry_setup;
1309         }
1310
1311         if (!pblk_line_init_bb(pblk, line, 1)) {
1312                 line = pblk_line_retry(pblk, line);
1313                 if (!line)
1314                         return NULL;
1315
1316                 goto retry_setup;
1317         }
1318
1319         return line;
1320 }
1321
1322 static void pblk_stop_writes(struct pblk *pblk, struct pblk_line *line)
1323 {
1324         lockdep_assert_held(&pblk->l_mg.free_lock);
1325
1326         pblk_set_space_limit(pblk);
1327         pblk->state = PBLK_STATE_STOPPING;
1328 }
1329
1330 void pblk_pipeline_stop(struct pblk *pblk)
1331 {
1332         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1333         int ret;
1334
1335         spin_lock(&l_mg->free_lock);
1336         if (pblk->state == PBLK_STATE_RECOVERING ||
1337                                         pblk->state == PBLK_STATE_STOPPED) {
1338                 spin_unlock(&l_mg->free_lock);
1339                 return;
1340         }
1341         pblk->state = PBLK_STATE_RECOVERING;
1342         spin_unlock(&l_mg->free_lock);
1343
1344         pblk_flush_writer(pblk);
1345         pblk_wait_for_meta(pblk);
1346
1347         ret = pblk_recov_pad(pblk);
1348         if (ret) {
1349                 pr_err("pblk: could not close data on teardown(%d)\n", ret);
1350                 return;
1351         }
1352
1353         pblk_line_close_meta_sync(pblk);
1354
1355         spin_lock(&l_mg->free_lock);
1356         pblk->state = PBLK_STATE_STOPPED;
1357         l_mg->data_line = NULL;
1358         l_mg->data_next = NULL;
1359         spin_unlock(&l_mg->free_lock);
1360 }
1361
1362 void pblk_line_replace_data(struct pblk *pblk)
1363 {
1364         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1365         struct pblk_line *cur, *new;
1366         unsigned int left_seblks;
1367         int is_next = 0;
1368
1369         cur = l_mg->data_line;
1370         new = l_mg->data_next;
1371         if (!new)
1372                 return;
1373         l_mg->data_line = new;
1374
1375         spin_lock(&l_mg->free_lock);
1376         if (pblk->state != PBLK_STATE_RUNNING) {
1377                 l_mg->data_line = NULL;
1378                 l_mg->data_next = NULL;
1379                 spin_unlock(&l_mg->free_lock);
1380                 return;
1381         }
1382
1383         pblk_line_setup_metadata(new, l_mg, &pblk->lm);
1384         spin_unlock(&l_mg->free_lock);
1385
1386 retry_erase:
1387         left_seblks = atomic_read(&new->left_seblks);
1388         if (left_seblks) {
1389                 /* If line is not fully erased, erase it */
1390                 if (atomic_read(&new->left_eblks)) {
1391                         if (pblk_line_erase(pblk, new))
1392                                 return;
1393                 } else {
1394                         io_schedule();
1395                 }
1396                 goto retry_erase;
1397         }
1398
1399 retry_setup:
1400         if (!pblk_line_init_metadata(pblk, new, cur)) {
1401                 new = pblk_line_retry(pblk, new);
1402                 if (!new)
1403                         return;
1404
1405                 goto retry_setup;
1406         }
1407
1408         if (!pblk_line_init_bb(pblk, new, 1)) {
1409                 new = pblk_line_retry(pblk, new);
1410                 if (!new)
1411                         return;
1412
1413                 goto retry_setup;
1414         }
1415
1416         /* Allocate next line for preparation */
1417         spin_lock(&l_mg->free_lock);
1418         l_mg->data_next = pblk_line_get(pblk);
1419         if (!l_mg->data_next) {
1420                 /* If we cannot get a new line, we need to stop the pipeline.
1421                  * Only allow as many writes in as we can store safely and then
1422                  * fail gracefully
1423                  */
1424                 pblk_stop_writes(pblk, new);
1425                 l_mg->data_next = NULL;
1426         } else {
1427                 l_mg->data_next->seq_nr = l_mg->d_seq_nr++;
1428                 l_mg->data_next->type = PBLK_LINETYPE_DATA;
1429                 is_next = 1;
1430         }
1431         spin_unlock(&l_mg->free_lock);
1432
1433         if (is_next)
1434                 pblk_rl_free_lines_dec(&pblk->rl, l_mg->data_next);
1435 }
1436
1437 void pblk_line_free(struct pblk *pblk, struct pblk_line *line)
1438 {
1439         if (line->map_bitmap)
1440                 mempool_free(line->map_bitmap, pblk->line_meta_pool);
1441         if (line->invalid_bitmap)
1442                 mempool_free(line->invalid_bitmap, pblk->line_meta_pool);
1443
1444         *line->vsc = cpu_to_le32(EMPTY_ENTRY);
1445
1446         line->map_bitmap = NULL;
1447         line->invalid_bitmap = NULL;
1448         line->smeta = NULL;
1449         line->emeta = NULL;
1450 }
1451
1452 void pblk_line_put(struct kref *ref)
1453 {
1454         struct pblk_line *line = container_of(ref, struct pblk_line, ref);
1455         struct pblk *pblk = line->pblk;
1456         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1457
1458         spin_lock(&line->lock);
1459         WARN_ON(line->state != PBLK_LINESTATE_GC);
1460         line->state = PBLK_LINESTATE_FREE;
1461         line->gc_group = PBLK_LINEGC_NONE;
1462         pblk_line_free(pblk, line);
1463         spin_unlock(&line->lock);
1464
1465         spin_lock(&l_mg->free_lock);
1466         list_add_tail(&line->list, &l_mg->free_list);
1467         l_mg->nr_free_lines++;
1468         spin_unlock(&l_mg->free_lock);
1469
1470         pblk_rl_free_lines_inc(&pblk->rl, line);
1471 }
1472
1473 int pblk_blk_erase_async(struct pblk *pblk, struct ppa_addr ppa)
1474 {
1475         struct nvm_rq *rqd;
1476         int err;
1477
1478         rqd = mempool_alloc(pblk->g_rq_pool, GFP_KERNEL);
1479         memset(rqd, 0, pblk_g_rq_size);
1480
1481         pblk_setup_e_rq(pblk, rqd, ppa);
1482
1483         rqd->end_io = pblk_end_io_erase;
1484         rqd->private = pblk;
1485
1486         /* The write thread schedules erases so that it minimizes disturbances
1487          * with writes. Thus, there is no need to take the LUN semaphore.
1488          */
1489         err = pblk_submit_io(pblk, rqd);
1490         if (err) {
1491                 struct nvm_tgt_dev *dev = pblk->dev;
1492                 struct nvm_geo *geo = &dev->geo;
1493
1494                 pr_err("pblk: could not async erase line:%d,blk:%d\n",
1495                                         pblk_dev_ppa_to_line(ppa),
1496                                         pblk_dev_ppa_to_pos(geo, ppa));
1497         }
1498
1499         return err;
1500 }
1501
1502 struct pblk_line *pblk_line_get_data(struct pblk *pblk)
1503 {
1504         return pblk->l_mg.data_line;
1505 }
1506
1507 /* For now, always erase next line */
1508 struct pblk_line *pblk_line_get_erase(struct pblk *pblk)
1509 {
1510         return pblk->l_mg.data_next;
1511 }
1512
1513 int pblk_line_is_full(struct pblk_line *line)
1514 {
1515         return (line->left_msecs == 0);
1516 }
1517
1518 void pblk_line_close_meta_sync(struct pblk *pblk)
1519 {
1520         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1521         struct pblk_line_meta *lm = &pblk->lm;
1522         struct pblk_line *line, *tline;
1523         LIST_HEAD(list);
1524
1525         spin_lock(&l_mg->close_lock);
1526         if (list_empty(&l_mg->emeta_list)) {
1527                 spin_unlock(&l_mg->close_lock);
1528                 return;
1529         }
1530
1531         list_cut_position(&list, &l_mg->emeta_list, l_mg->emeta_list.prev);
1532         spin_unlock(&l_mg->close_lock);
1533
1534         list_for_each_entry_safe(line, tline, &list, list) {
1535                 struct pblk_emeta *emeta = line->emeta;
1536
1537                 while (emeta->mem < lm->emeta_len[0]) {
1538                         int ret;
1539
1540                         ret = pblk_submit_meta_io(pblk, line);
1541                         if (ret) {
1542                                 pr_err("pblk: sync meta line %d failed (%d)\n",
1543                                                         line->id, ret);
1544                                 return;
1545                         }
1546                 }
1547         }
1548
1549         pblk_wait_for_meta(pblk);
1550 }
1551
1552 static void pblk_line_should_sync_meta(struct pblk *pblk)
1553 {
1554         if (pblk_rl_is_limit(&pblk->rl))
1555                 pblk_line_close_meta_sync(pblk);
1556 }
1557
1558 void pblk_line_close(struct pblk *pblk, struct pblk_line *line)
1559 {
1560         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1561         struct pblk_line_meta *lm = &pblk->lm;
1562         struct list_head *move_list;
1563
1564         WARN(!bitmap_full(line->map_bitmap, lm->sec_per_line),
1565                                 "pblk: corrupt closed line %d\n", line->id);
1566
1567         spin_lock(&l_mg->free_lock);
1568         WARN_ON(!test_and_clear_bit(line->meta_line, &l_mg->meta_bitmap));
1569         spin_unlock(&l_mg->free_lock);
1570
1571         spin_lock(&l_mg->gc_lock);
1572         spin_lock(&line->lock);
1573         WARN_ON(line->state != PBLK_LINESTATE_OPEN);
1574         line->state = PBLK_LINESTATE_CLOSED;
1575         move_list = pblk_line_gc_list(pblk, line);
1576
1577         list_add_tail(&line->list, move_list);
1578
1579         mempool_free(line->map_bitmap, pblk->line_meta_pool);
1580         line->map_bitmap = NULL;
1581         line->smeta = NULL;
1582         line->emeta = NULL;
1583
1584         spin_unlock(&line->lock);
1585         spin_unlock(&l_mg->gc_lock);
1586
1587         pblk_gc_should_kick(pblk);
1588 }
1589
1590 void pblk_line_close_meta(struct pblk *pblk, struct pblk_line *line)
1591 {
1592         struct pblk_line_mgmt *l_mg = &pblk->l_mg;
1593         struct pblk_line_meta *lm = &pblk->lm;
1594         struct pblk_emeta *emeta = line->emeta;
1595         struct line_emeta *emeta_buf = emeta->buf;
1596
1597         /* No need for exact vsc value; avoid a big line lock and take aprox. */
1598         memcpy(emeta_to_vsc(pblk, emeta_buf), l_mg->vsc_list, lm->vsc_list_len);
1599         memcpy(emeta_to_bb(emeta_buf), line->blk_bitmap, lm->blk_bitmap_len);
1600
1601         emeta_buf->nr_valid_lbas = cpu_to_le64(line->nr_valid_lbas);
1602         emeta_buf->crc = cpu_to_le32(pblk_calc_emeta_crc(pblk, emeta_buf));
1603
1604         spin_lock(&l_mg->close_lock);
1605         spin_lock(&line->lock);
1606         list_add_tail(&line->list, &l_mg->emeta_list);
1607         spin_unlock(&line->lock);
1608         spin_unlock(&l_mg->close_lock);
1609
1610         pblk_line_should_sync_meta(pblk);
1611 }
1612
1613 void pblk_line_close_ws(struct work_struct *work)
1614 {
1615         struct pblk_line_ws *line_ws = container_of(work, struct pblk_line_ws,
1616                                                                         ws);
1617         struct pblk *pblk = line_ws->pblk;
1618         struct pblk_line *line = line_ws->line;
1619
1620         pblk_line_close(pblk, line);
1621         mempool_free(line_ws, pblk->line_ws_pool);
1622 }
1623
1624 void pblk_line_mark_bb(struct work_struct *work)
1625 {
1626         struct pblk_line_ws *line_ws = container_of(work, struct pblk_line_ws,
1627                                                                         ws);
1628         struct pblk *pblk = line_ws->pblk;
1629         struct nvm_tgt_dev *dev = pblk->dev;
1630         struct ppa_addr *ppa = line_ws->priv;
1631         int ret;
1632
1633         ret = nvm_set_tgt_bb_tbl(dev, ppa, 1, NVM_BLK_T_GRWN_BAD);
1634         if (ret) {
1635                 struct pblk_line *line;
1636                 int pos;
1637
1638                 line = &pblk->lines[pblk_dev_ppa_to_line(*ppa)];
1639                 pos = pblk_dev_ppa_to_pos(&dev->geo, *ppa);
1640
1641                 pr_err("pblk: failed to mark bb, line:%d, pos:%d\n",
1642                                 line->id, pos);
1643         }
1644
1645         kfree(ppa);
1646         mempool_free(line_ws, pblk->line_ws_pool);
1647 }
1648
1649 void pblk_line_run_ws(struct pblk *pblk, struct pblk_line *line, void *priv,
1650                       void (*work)(struct work_struct *),
1651                       struct workqueue_struct *wq)
1652 {
1653         struct pblk_line_ws *line_ws;
1654
1655         line_ws = mempool_alloc(pblk->line_ws_pool, GFP_ATOMIC);
1656         if (!line_ws)
1657                 return;
1658
1659         line_ws->pblk = pblk;
1660         line_ws->line = line;
1661         line_ws->priv = priv;
1662
1663         INIT_WORK(&line_ws->ws, work);
1664         queue_work(wq, &line_ws->ws);
1665 }
1666
1667 void pblk_down_rq(struct pblk *pblk, struct ppa_addr *ppa_list, int nr_ppas,
1668                   unsigned long *lun_bitmap)
1669 {
1670         struct nvm_tgt_dev *dev = pblk->dev;
1671         struct nvm_geo *geo = &dev->geo;
1672         struct pblk_lun *rlun;
1673         int pos = pblk_ppa_to_pos(geo, ppa_list[0]);
1674         int ret;
1675
1676         /*
1677          * Only send one inflight I/O per LUN. Since we map at a page
1678          * granurality, all ppas in the I/O will map to the same LUN
1679          */
1680 #ifdef CONFIG_NVM_DEBUG
1681         int i;
1682
1683         for (i = 1; i < nr_ppas; i++)
1684                 WARN_ON(ppa_list[0].g.lun != ppa_list[i].g.lun ||
1685                                 ppa_list[0].g.ch != ppa_list[i].g.ch);
1686 #endif
1687         /* If the LUN has been locked for this same request, do no attempt to
1688          * lock it again
1689          */
1690         if (test_and_set_bit(pos, lun_bitmap))
1691                 return;
1692
1693         rlun = &pblk->luns[pos];
1694         ret = down_timeout(&rlun->wr_sem, msecs_to_jiffies(5000));
1695         if (ret) {
1696                 switch (ret) {
1697                 case -ETIME:
1698                         pr_err("pblk: lun semaphore timed out\n");
1699                         break;
1700                 case -EINTR:
1701                         pr_err("pblk: lun semaphore timed out\n");
1702                         break;
1703                 }
1704         }
1705 }
1706
1707 void pblk_up_rq(struct pblk *pblk, struct ppa_addr *ppa_list, int nr_ppas,
1708                 unsigned long *lun_bitmap)
1709 {
1710         struct nvm_tgt_dev *dev = pblk->dev;
1711         struct nvm_geo *geo = &dev->geo;
1712         struct pblk_lun *rlun;
1713         int nr_luns = geo->nr_luns;
1714         int bit = -1;
1715
1716         while ((bit = find_next_bit(lun_bitmap, nr_luns, bit + 1)) < nr_luns) {
1717                 rlun = &pblk->luns[bit];
1718                 up(&rlun->wr_sem);
1719         }
1720
1721         kfree(lun_bitmap);
1722 }
1723
1724 void pblk_update_map(struct pblk *pblk, sector_t lba, struct ppa_addr ppa)
1725 {
1726         struct ppa_addr l2p_ppa;
1727
1728         /* logic error: lba out-of-bounds. Ignore update */
1729         if (!(lba < pblk->rl.nr_secs)) {
1730                 WARN(1, "pblk: corrupted L2P map request\n");
1731                 return;
1732         }
1733
1734         spin_lock(&pblk->trans_lock);
1735         l2p_ppa = pblk_trans_map_get(pblk, lba);
1736
1737         if (!pblk_addr_in_cache(l2p_ppa) && !pblk_ppa_empty(l2p_ppa))
1738                 pblk_map_invalidate(pblk, l2p_ppa);
1739
1740         pblk_trans_map_set(pblk, lba, ppa);
1741         spin_unlock(&pblk->trans_lock);
1742 }
1743
1744 void pblk_update_map_cache(struct pblk *pblk, sector_t lba, struct ppa_addr ppa)
1745 {
1746 #ifdef CONFIG_NVM_DEBUG
1747         /* Callers must ensure that the ppa points to a cache address */
1748         BUG_ON(!pblk_addr_in_cache(ppa));
1749         BUG_ON(pblk_rb_pos_oob(&pblk->rwb, pblk_addr_to_cacheline(ppa)));
1750 #endif
1751
1752         pblk_update_map(pblk, lba, ppa);
1753 }
1754
1755 int pblk_update_map_gc(struct pblk *pblk, sector_t lba, struct ppa_addr ppa,
1756                        struct pblk_line *gc_line)
1757 {
1758         struct ppa_addr l2p_ppa;
1759         int ret = 1;
1760
1761 #ifdef CONFIG_NVM_DEBUG
1762         /* Callers must ensure that the ppa points to a cache address */
1763         BUG_ON(!pblk_addr_in_cache(ppa));
1764         BUG_ON(pblk_rb_pos_oob(&pblk->rwb, pblk_addr_to_cacheline(ppa)));
1765 #endif
1766
1767         /* logic error: lba out-of-bounds. Ignore update */
1768         if (!(lba < pblk->rl.nr_secs)) {
1769                 WARN(1, "pblk: corrupted L2P map request\n");
1770                 return 0;
1771         }
1772
1773         spin_lock(&pblk->trans_lock);
1774         l2p_ppa = pblk_trans_map_get(pblk, lba);
1775
1776         /* Prevent updated entries to be overwritten by GC */
1777         if (pblk_addr_in_cache(l2p_ppa) || pblk_ppa_empty(l2p_ppa) ||
1778                                 pblk_tgt_ppa_to_line(l2p_ppa) != gc_line->id) {
1779                 ret = 0;
1780                 goto out;
1781         }
1782
1783         pblk_trans_map_set(pblk, lba, ppa);
1784 out:
1785         spin_unlock(&pblk->trans_lock);
1786         return ret;
1787 }
1788
1789 void pblk_update_map_dev(struct pblk *pblk, sector_t lba, struct ppa_addr ppa,
1790                          struct ppa_addr entry_line)
1791 {
1792         struct ppa_addr l2p_line;
1793
1794 #ifdef CONFIG_NVM_DEBUG
1795         /* Callers must ensure that the ppa points to a device address */
1796         BUG_ON(pblk_addr_in_cache(ppa));
1797 #endif
1798         /* Invalidate and discard padded entries */
1799         if (lba == ADDR_EMPTY) {
1800 #ifdef CONFIG_NVM_DEBUG
1801                 atomic_long_inc(&pblk->padded_wb);
1802 #endif
1803                 pblk_map_invalidate(pblk, ppa);
1804                 return;
1805         }
1806
1807         /* logic error: lba out-of-bounds. Ignore update */
1808         if (!(lba < pblk->rl.nr_secs)) {
1809                 WARN(1, "pblk: corrupted L2P map request\n");
1810                 return;
1811         }
1812
1813         spin_lock(&pblk->trans_lock);
1814         l2p_line = pblk_trans_map_get(pblk, lba);
1815
1816         /* Do not update L2P if the cacheline has been updated. In this case,
1817          * the mapped ppa must be invalidated
1818          */
1819         if (l2p_line.ppa != entry_line.ppa) {
1820                 if (!pblk_ppa_empty(ppa))
1821                         pblk_map_invalidate(pblk, ppa);
1822                 goto out;
1823         }
1824
1825 #ifdef CONFIG_NVM_DEBUG
1826         WARN_ON(!pblk_addr_in_cache(l2p_line) && !pblk_ppa_empty(l2p_line));
1827 #endif
1828
1829         pblk_trans_map_set(pblk, lba, ppa);
1830 out:
1831         spin_unlock(&pblk->trans_lock);
1832 }
1833
1834 void pblk_lookup_l2p_seq(struct pblk *pblk, struct ppa_addr *ppas,
1835                          sector_t blba, int nr_secs)
1836 {
1837         int i;
1838
1839         spin_lock(&pblk->trans_lock);
1840         for (i = 0; i < nr_secs; i++)
1841                 ppas[i] = pblk_trans_map_get(pblk, blba + i);
1842         spin_unlock(&pblk->trans_lock);
1843 }
1844
1845 void pblk_lookup_l2p_rand(struct pblk *pblk, struct ppa_addr *ppas,
1846                           u64 *lba_list, int nr_secs)
1847 {
1848         sector_t lba;
1849         int i;
1850
1851         spin_lock(&pblk->trans_lock);
1852         for (i = 0; i < nr_secs; i++) {
1853                 lba = lba_list[i];
1854                 if (lba == ADDR_EMPTY) {
1855                         ppas[i].ppa = ADDR_EMPTY;
1856                 } else {
1857                         /* logic error: lba out-of-bounds. Ignore update */
1858                         if (!(lba < pblk->rl.nr_secs)) {
1859                                 WARN(1, "pblk: corrupted L2P map request\n");
1860                                 continue;
1861                         }
1862                         ppas[i] = pblk_trans_map_get(pblk, lba);
1863                 }
1864         }
1865         spin_unlock(&pblk->trans_lock);
1866 }