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