drivers/lightnvm/rrpc.c

   1 /*
   2  * Copyright (C) 2015 IT University of Copenhagen
   3  * Initial release: Matias Bjorling <[email protected]>
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License version
   7  * 2 as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful, but
  10  * WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  12  * General Public License for more details.
  13  *
  14  * Implementation of a Round-robin page-based Hybrid FTL for Open-channel SSDs.
  15  */
  16
  17 #include "rrpc.h"
  18
  19 static struct kmem_cache *rrpc_gcb_cache, *rrpc_rq_cache;
  20 static DECLARE_RWSEM(rrpc_lock);
  21
  22 static int rrpc_submit_io(struct rrpc *rrpc, struct bio *bio,
  23                                 struct nvm_rq *rqd, unsigned long flags);
  24
  25 #define rrpc_for_each_lun(rrpc, rlun, i) \
  26                 for ((i) = 0, rlun = &(rrpc)->luns[0]; \
  27                         (i) < (rrpc)->nr_luns; (i)++, rlun = &(rrpc)->luns[(i)])
  28
  29 static void rrpc_page_invalidate(struct rrpc *rrpc, struct rrpc_addr *a)
  30 {
  31         struct nvm_tgt_dev *dev = rrpc->dev;
  32         struct rrpc_block *rblk = a->rblk;
  33         unsigned int pg_offset;
  34
  35         lockdep_assert_held(&rrpc->rev_lock);
  36
  37         if (a->addr == ADDR_EMPTY || !rblk)
  38                 return;
  39
  40         spin_lock(&rblk->lock);
  41
  42         div_u64_rem(a->addr, dev->geo.sec_per_blk, &pg_offset);
  43         WARN_ON(test_and_set_bit(pg_offset, rblk->invalid_pages));
  44         rblk->nr_invalid_pages++;
  45
  46         spin_unlock(&rblk->lock);
  47
  48         rrpc->rev_trans_map[a->addr].addr = ADDR_EMPTY;
  49 }
  50
  51 static void rrpc_invalidate_range(struct rrpc *rrpc, sector_t slba,
  52                                                         unsigned int len)
  53 {
  54         sector_t i;
  55
  56         spin_lock(&rrpc->rev_lock);
  57         for (i = slba; i < slba + len; i++) {
  58                 struct rrpc_addr *gp = &rrpc->trans_map[i];
  59
  60                 rrpc_page_invalidate(rrpc, gp);
  61                 gp->rblk = NULL;
  62         }
  63         spin_unlock(&rrpc->rev_lock);
  64 }
  65
  66 static struct nvm_rq *rrpc_inflight_laddr_acquire(struct rrpc *rrpc,
  67                                         sector_t laddr, unsigned int pages)
  68 {
  69         struct nvm_rq *rqd;
  70         struct rrpc_inflight_rq *inf;
  71
  72         rqd = mempool_alloc(rrpc->rq_pool, GFP_ATOMIC);
  73         if (!rqd)
  74                 return ERR_PTR(-ENOMEM);
  75
  76         inf = rrpc_get_inflight_rq(rqd);
  77         if (rrpc_lock_laddr(rrpc, laddr, pages, inf)) {
  78                 mempool_free(rqd, rrpc->rq_pool);
  79                 return NULL;
  80         }
  81
  82         return rqd;
  83 }
  84
  85 static void rrpc_inflight_laddr_release(struct rrpc *rrpc, struct nvm_rq *rqd)
  86 {
  87         struct rrpc_inflight_rq *inf = rrpc_get_inflight_rq(rqd);
  88
  89         rrpc_unlock_laddr(rrpc, inf);
  90
  91         mempool_free(rqd, rrpc->rq_pool);
  92 }
  93
  94 static void rrpc_discard(struct rrpc *rrpc, struct bio *bio)
  95 {
  96         sector_t slba = bio->bi_iter.bi_sector / NR_PHY_IN_LOG;
  97         sector_t len = bio->bi_iter.bi_size / RRPC_EXPOSED_PAGE_SIZE;
  98         struct nvm_rq *rqd;
  99
 100         while (1) {
 101                 rqd = rrpc_inflight_laddr_acquire(rrpc, slba, len);
 102                 if (rqd)
 103                         break;
 104
 105                 schedule();
 106         }
 107
 108         if (IS_ERR(rqd)) {
 109                 pr_err("rrpc: unable to acquire inflight IO\n");
 110                 bio_io_error(bio);
 111                 return;
 112         }
 113
 114         rrpc_invalidate_range(rrpc, slba, len);
 115         rrpc_inflight_laddr_release(rrpc, rqd);
 116 }
 117
 118 static int block_is_full(struct rrpc *rrpc, struct rrpc_block *rblk)
 119 {
 120         struct nvm_tgt_dev *dev = rrpc->dev;
 121
 122         return (rblk->next_page == dev->geo.sec_per_blk);
 123 }
 124
 125 /* Calculate relative addr for the given block, considering instantiated LUNs */
 126 static u64 block_to_rel_addr(struct rrpc *rrpc, struct rrpc_block *rblk)
 127 {
 128         struct nvm_tgt_dev *dev = rrpc->dev;
 129         struct rrpc_lun *rlun = rblk->rlun;
 130
 131         return rlun->id * dev->geo.sec_per_blk;
 132 }
 133
 134 static struct ppa_addr rrpc_ppa_to_gaddr(struct nvm_tgt_dev *dev,
 135                                          struct rrpc_addr *gp)
 136 {
 137         struct rrpc_block *rblk = gp->rblk;
 138         struct rrpc_lun *rlun = rblk->rlun;
 139         u64 addr = gp->addr;
 140         struct ppa_addr paddr;
 141
 142         paddr.ppa = addr;
 143         paddr = rrpc_linear_to_generic_addr(&dev->geo, paddr);
 144         paddr.g.ch = rlun->bppa.g.ch;
 145         paddr.g.lun = rlun->bppa.g.lun;
 146         paddr.g.blk = rblk->id;
 147
 148         return paddr;
 149 }
 150
 151 /* requires lun->lock taken */
 152 static void rrpc_set_lun_cur(struct rrpc_lun *rlun, struct rrpc_block *new_rblk,
 153                                                 struct rrpc_block **cur_rblk)
 154 {
 155         struct rrpc *rrpc = rlun->rrpc;
 156
 157         if (*cur_rblk) {
 158                 spin_lock(&(*cur_rblk)->lock);
 159                 WARN_ON(!block_is_full(rrpc, *cur_rblk));
 160                 spin_unlock(&(*cur_rblk)->lock);
 161         }
 162         *cur_rblk = new_rblk;
 163 }
 164
 165 static struct rrpc_block *__rrpc_get_blk(struct rrpc *rrpc,
 166                                                         struct rrpc_lun *rlun)
 167 {
 168         struct rrpc_block *rblk = NULL;
 169
 170         if (list_empty(&rlun->free_list))
 171                 goto out;
 172
 173         rblk = list_first_entry(&rlun->free_list, struct rrpc_block, list);
 174
 175         list_move_tail(&rblk->list, &rlun->used_list);
 176         rblk->state = NVM_BLK_ST_TGT;
 177         rlun->nr_free_blocks--;
 178
 179 out:
 180         return rblk;
 181 }
 182
 183 static struct rrpc_block *rrpc_get_blk(struct rrpc *rrpc, struct rrpc_lun *rlun,
 184                                                         unsigned long flags)
 185 {
 186         struct nvm_tgt_dev *dev = rrpc->dev;
 187         struct rrpc_block *rblk;
 188         int is_gc = flags & NVM_IOTYPE_GC;
 189
 190         spin_lock(&rlun->lock);
 191         if (!is_gc && rlun->nr_free_blocks < rlun->reserved_blocks) {
 192                 pr_err("nvm: rrpc: cannot give block to non GC request\n");
 193                 spin_unlock(&rlun->lock);
 194                 return NULL;
 195         }
 196
 197         rblk = __rrpc_get_blk(rrpc, rlun);
 198         if (!rblk) {
 199                 pr_err("nvm: rrpc: cannot get new block\n");
 200                 spin_unlock(&rlun->lock);
 201                 return NULL;
 202         }
 203         spin_unlock(&rlun->lock);
 204
 205         bitmap_zero(rblk->invalid_pages, dev->geo.sec_per_blk);
 206         rblk->next_page = 0;
 207         rblk->nr_invalid_pages = 0;
 208         atomic_set(&rblk->data_cmnt_size, 0);
 209
 210         return rblk;
 211 }
 212
 213 static void rrpc_put_blk(struct rrpc *rrpc, struct rrpc_block *rblk)
 214 {
 215         struct rrpc_lun *rlun = rblk->rlun;
 216
 217         spin_lock(&rlun->lock);
 218         if (rblk->state & NVM_BLK_ST_TGT) {
 219                 list_move_tail(&rblk->list, &rlun->free_list);
 220                 rlun->nr_free_blocks++;
 221                 rblk->state = NVM_BLK_ST_FREE;
 222         } else if (rblk->state & NVM_BLK_ST_BAD) {
 223                 list_move_tail(&rblk->list, &rlun->bb_list);
 224                 rblk->state = NVM_BLK_ST_BAD;
 225         } else {
 226                 WARN_ON_ONCE(1);
 227                 pr_err("rrpc: erroneous type (ch:%d,lun:%d,blk%d-> %u)\n",
 228                                         rlun->bppa.g.ch, rlun->bppa.g.lun,
 229                                         rblk->id, rblk->state);
 230                 list_move_tail(&rblk->list, &rlun->bb_list);
 231         }
 232         spin_unlock(&rlun->lock);
 233 }
 234
 235 static void rrpc_put_blks(struct rrpc *rrpc)
 236 {
 237         struct rrpc_lun *rlun;
 238         int i;
 239
 240         for (i = 0; i < rrpc->nr_luns; i++) {
 241                 rlun = &rrpc->luns[i];
 242                 if (rlun->cur)
 243                         rrpc_put_blk(rrpc, rlun->cur);
 244                 if (rlun->gc_cur)
 245                         rrpc_put_blk(rrpc, rlun->gc_cur);
 246         }
 247 }
 248
 249 static struct rrpc_lun *get_next_lun(struct rrpc *rrpc)
 250 {
 251         int next = atomic_inc_return(&rrpc->next_lun);
 252
 253         return &rrpc->luns[next % rrpc->nr_luns];
 254 }
 255
 256 static void rrpc_gc_kick(struct rrpc *rrpc)
 257 {
 258         struct rrpc_lun *rlun;
 259         unsigned int i;
 260
 261         for (i = 0; i < rrpc->nr_luns; i++) {
 262                 rlun = &rrpc->luns[i];
 263                 queue_work(rrpc->krqd_wq, &rlun->ws_gc);
 264         }
 265 }
 266
 267 /*
 268  * timed GC every interval.
 269  */
 270 static void rrpc_gc_timer(unsigned long data)
 271 {
 272         struct rrpc *rrpc = (struct rrpc *)data;
 273
 274         rrpc_gc_kick(rrpc);
 275         mod_timer(&rrpc->gc_timer, jiffies + msecs_to_jiffies(10));
 276 }
 277
 278 static void rrpc_end_sync_bio(struct bio *bio)
 279 {
 280         struct completion *waiting = bio->bi_private;
 281
 282         if (bio->bi_error)
 283                 pr_err("nvm: gc request failed (%u).\n", bio->bi_error);
 284
 285         complete(waiting);
 286 }
 287
 288 /*
 289  * rrpc_move_valid_pages -- migrate live data off the block
 290  * @rrpc: the 'rrpc' structure
 291  * @block: the block from which to migrate live pages
 292  *
 293  * Description:
 294  *   GC algorithms may call this function to migrate remaining live
 295  *   pages off the block prior to erasing it. This function blocks
 296  *   further execution until the operation is complete.
 297  */
 298 static int rrpc_move_valid_pages(struct rrpc *rrpc, struct rrpc_block *rblk)
 299 {
 300         struct nvm_tgt_dev *dev = rrpc->dev;
 301         struct request_queue *q = dev->q;
 302         struct rrpc_rev_addr *rev;
 303         struct nvm_rq *rqd;
 304         struct bio *bio;
 305         struct page *page;
 306         int slot;
 307         int nr_sec_per_blk = dev->geo.sec_per_blk;
 308         u64 phys_addr;
 309         DECLARE_COMPLETION_ONSTACK(wait);
 310
 311         if (bitmap_full(rblk->invalid_pages, nr_sec_per_blk))
 312                 return 0;
 313
 314         bio = bio_alloc(GFP_NOIO, 1);
 315         if (!bio) {
 316                 pr_err("nvm: could not alloc bio to gc\n");
 317                 return -ENOMEM;
 318         }
 319
 320         page = mempool_alloc(rrpc->page_pool, GFP_NOIO);
 321         if (!page) {
 322                 bio_put(bio);
 323                 return -ENOMEM;
 324         }
 325
 326         while ((slot = find_first_zero_bit(rblk->invalid_pages,
 327                                             nr_sec_per_blk)) < nr_sec_per_blk) {
 328
 329                 /* Lock laddr */
 330                 phys_addr = rrpc_blk_to_ppa(rrpc, rblk) + slot;
 331
 332 try:
 333                 spin_lock(&rrpc->rev_lock);
 334                 /* Get logical address from physical to logical table */
 335                 rev = &rrpc->rev_trans_map[phys_addr];
 336                 /* already updated by previous regular write */
 337                 if (rev->addr == ADDR_EMPTY) {
 338                         spin_unlock(&rrpc->rev_lock);
 339                         continue;
 340                 }
 341
 342                 rqd = rrpc_inflight_laddr_acquire(rrpc, rev->addr, 1);
 343                 if (IS_ERR_OR_NULL(rqd)) {
 344                         spin_unlock(&rrpc->rev_lock);
 345                         schedule();
 346                         goto try;
 347                 }
 348
 349                 spin_unlock(&rrpc->rev_lock);
 350
 351                 /* Perform read to do GC */
 352                 bio->bi_iter.bi_sector = rrpc_get_sector(rev->addr);
 353                 bio_set_op_attrs(bio,  REQ_OP_READ, 0);
 354                 bio->bi_private = &wait;
 355                 bio->bi_end_io = rrpc_end_sync_bio;
 356
 357                 /* TODO: may fail when EXP_PG_SIZE > PAGE_SIZE */
 358                 bio_add_pc_page(q, bio, page, RRPC_EXPOSED_PAGE_SIZE, 0);
 359
 360                 if (rrpc_submit_io(rrpc, bio, rqd, NVM_IOTYPE_GC)) {
 361                         pr_err("rrpc: gc read failed.\n");
 362                         rrpc_inflight_laddr_release(rrpc, rqd);
 363                         goto finished;
 364                 }
 365                 wait_for_completion_io(&wait);
 366                 if (bio->bi_error) {
 367                         rrpc_inflight_laddr_release(rrpc, rqd);
 368                         goto finished;
 369                 }
 370
 371                 bio_reset(bio);
 372                 reinit_completion(&wait);
 373
 374                 bio->bi_iter.bi_sector = rrpc_get_sector(rev->addr);
 375                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 376                 bio->bi_private = &wait;
 377                 bio->bi_end_io = rrpc_end_sync_bio;
 378
 379                 bio_add_pc_page(q, bio, page, RRPC_EXPOSED_PAGE_SIZE, 0);
 380
 381                 /* turn the command around and write the data back to a new
 382                  * address
 383                  */
 384                 if (rrpc_submit_io(rrpc, bio, rqd, NVM_IOTYPE_GC)) {
 385                         pr_err("rrpc: gc write failed.\n");
 386                         rrpc_inflight_laddr_release(rrpc, rqd);
 387                         goto finished;
 388                 }
 389                 wait_for_completion_io(&wait);
 390
 391                 rrpc_inflight_laddr_release(rrpc, rqd);
 392                 if (bio->bi_error)
 393                         goto finished;
 394
 395                 bio_reset(bio);
 396         }
 397
 398 finished:
 399         mempool_free(page, rrpc->page_pool);
 400         bio_put(bio);
 401
 402         if (!bitmap_full(rblk->invalid_pages, nr_sec_per_blk)) {
 403                 pr_err("nvm: failed to garbage collect block\n");
 404                 return -EIO;
 405         }
 406
 407         return 0;
 408 }
 409
 410 static void rrpc_block_gc(struct work_struct *work)
 411 {
 412         struct rrpc_block_gc *gcb = container_of(work, struct rrpc_block_gc,
 413                                                                         ws_gc);
 414         struct rrpc *rrpc = gcb->rrpc;
 415         struct rrpc_block *rblk = gcb->rblk;
 416         struct rrpc_lun *rlun = rblk->rlun;
 417         struct nvm_tgt_dev *dev = rrpc->dev;
 418         struct ppa_addr ppa;
 419
 420         mempool_free(gcb, rrpc->gcb_pool);
 421         pr_debug("nvm: block 'ch:%d,lun:%d,blk:%d' being reclaimed\n",
 422                         rlun->bppa.g.ch, rlun->bppa.g.lun,
 423                         rblk->id);
 424
 425         if (rrpc_move_valid_pages(rrpc, rblk))
 426                 goto put_back;
 427
 428         ppa.ppa = 0;
 429         ppa.g.ch = rlun->bppa.g.ch;
 430         ppa.g.lun = rlun->bppa.g.lun;
 431         ppa.g.blk = rblk->id;
 432
 433         if (nvm_erase_blk(dev, &ppa, 0))
 434                 goto put_back;
 435
 436         rrpc_put_blk(rrpc, rblk);
 437
 438         return;
 439
 440 put_back:
 441         spin_lock(&rlun->lock);
 442         list_add_tail(&rblk->prio, &rlun->prio_list);
 443         spin_unlock(&rlun->lock);
 444 }
 445
 446 /* the block with highest number of invalid pages, will be in the beginning
 447  * of the list
 448  */
 449 static struct rrpc_block *rblk_max_invalid(struct rrpc_block *ra,
 450                                                         struct rrpc_block *rb)
 451 {
 452         if (ra->nr_invalid_pages == rb->nr_invalid_pages)
 453                 return ra;
 454
 455         return (ra->nr_invalid_pages < rb->nr_invalid_pages) ? rb : ra;
 456 }
 457
 458 /* linearly find the block with highest number of invalid pages
 459  * requires lun->lock
 460  */
 461 static struct rrpc_block *block_prio_find_max(struct rrpc_lun *rlun)
 462 {
 463         struct list_head *prio_list = &rlun->prio_list;
 464         struct rrpc_block *rblk, *max;
 465
 466         BUG_ON(list_empty(prio_list));
 467
 468         max = list_first_entry(prio_list, struct rrpc_block, prio);
 469         list_for_each_entry(rblk, prio_list, prio)
 470                 max = rblk_max_invalid(max, rblk);
 471
 472         return max;
 473 }
 474
 475 static void rrpc_lun_gc(struct work_struct *work)
 476 {
 477         struct rrpc_lun *rlun = container_of(work, struct rrpc_lun, ws_gc);
 478         struct rrpc *rrpc = rlun->rrpc;
 479         struct nvm_tgt_dev *dev = rrpc->dev;
 480         struct rrpc_block_gc *gcb;
 481         unsigned int nr_blocks_need;
 482
 483         nr_blocks_need = dev->geo.blks_per_lun / GC_LIMIT_INVERSE;
 484
 485         if (nr_blocks_need < rrpc->nr_luns)
 486                 nr_blocks_need = rrpc->nr_luns;
 487
 488         spin_lock(&rlun->lock);
 489         while (nr_blocks_need > rlun->nr_free_blocks &&
 490                                         !list_empty(&rlun->prio_list)) {
 491                 struct rrpc_block *rblk = block_prio_find_max(rlun);
 492
 493                 if (!rblk->nr_invalid_pages)
 494                         break;
 495
 496                 gcb = mempool_alloc(rrpc->gcb_pool, GFP_ATOMIC);
 497                 if (!gcb)
 498                         break;
 499
 500                 list_del_init(&rblk->prio);
 501
 502                 WARN_ON(!block_is_full(rrpc, rblk));
 503
 504                 pr_debug("rrpc: selected block 'ch:%d,lun:%d,blk:%d' for GC\n",
 505                                         rlun->bppa.g.ch, rlun->bppa.g.lun,
 506                                         rblk->id);
 507
 508                 gcb->rrpc = rrpc;
 509                 gcb->rblk = rblk;
 510                 INIT_WORK(&gcb->ws_gc, rrpc_block_gc);
 511
 512                 queue_work(rrpc->kgc_wq, &gcb->ws_gc);
 513
 514                 nr_blocks_need--;
 515         }
 516         spin_unlock(&rlun->lock);
 517
 518         /* TODO: Hint that request queue can be started again */
 519 }
 520
 521 static void rrpc_gc_queue(struct work_struct *work)
 522 {
 523         struct rrpc_block_gc *gcb = container_of(work, struct rrpc_block_gc,
 524                                                                         ws_gc);
 525         struct rrpc *rrpc = gcb->rrpc;
 526         struct rrpc_block *rblk = gcb->rblk;
 527         struct rrpc_lun *rlun = rblk->rlun;
 528
 529         spin_lock(&rlun->lock);
 530         list_add_tail(&rblk->prio, &rlun->prio_list);
 531         spin_unlock(&rlun->lock);
 532
 533         mempool_free(gcb, rrpc->gcb_pool);
 534         pr_debug("nvm: block 'ch:%d,lun:%d,blk:%d' full, allow GC (sched)\n",
 535                                         rlun->bppa.g.ch, rlun->bppa.g.lun,
 536                                         rblk->id);
 537 }
 538
 539 static const struct block_device_operations rrpc_fops = {
 540         .owner          = THIS_MODULE,
 541 };
 542
 543 static struct rrpc_lun *rrpc_get_lun_rr(struct rrpc *rrpc, int is_gc)
 544 {
 545         unsigned int i;
 546         struct rrpc_lun *rlun, *max_free;
 547
 548         if (!is_gc)
 549                 return get_next_lun(rrpc);
 550
 551         /* during GC, we don't care about RR, instead we want to make
 552          * sure that we maintain evenness between the block luns.
 553          */
 554         max_free = &rrpc->luns[0];
 555         /* prevent GC-ing lun from devouring pages of a lun with
 556          * little free blocks. We don't take the lock as we only need an
 557          * estimate.
 558          */
 559         rrpc_for_each_lun(rrpc, rlun, i) {
 560                 if (rlun->nr_free_blocks > max_free->nr_free_blocks)
 561                         max_free = rlun;
 562         }
 563
 564         return max_free;
 565 }
 566
 567 static struct rrpc_addr *rrpc_update_map(struct rrpc *rrpc, sector_t laddr,
 568                                         struct rrpc_block *rblk, u64 paddr)
 569 {
 570         struct rrpc_addr *gp;
 571         struct rrpc_rev_addr *rev;
 572
 573         BUG_ON(laddr >= rrpc->nr_sects);
 574
 575         gp = &rrpc->trans_map[laddr];
 576         spin_lock(&rrpc->rev_lock);
 577         if (gp->rblk)
 578                 rrpc_page_invalidate(rrpc, gp);
 579
 580         gp->addr = paddr;
 581         gp->rblk = rblk;
 582
 583         rev = &rrpc->rev_trans_map[gp->addr];
 584         rev->addr = laddr;
 585         spin_unlock(&rrpc->rev_lock);
 586
 587         return gp;
 588 }
 589
 590 static u64 rrpc_alloc_addr(struct rrpc *rrpc, struct rrpc_block *rblk)
 591 {
 592         u64 addr = ADDR_EMPTY;
 593
 594         spin_lock(&rblk->lock);
 595         if (block_is_full(rrpc, rblk))
 596                 goto out;
 597
 598         addr = rblk->next_page;
 599
 600         rblk->next_page++;
 601 out:
 602         spin_unlock(&rblk->lock);
 603         return addr;
 604 }
 605
 606 /* Map logical address to a physical page. The mapping implements a round robin
 607  * approach and allocates a page from the next lun available.
 608  *
 609  * Returns rrpc_addr with the physical address and block. Returns NULL if no
 610  * blocks in the next rlun are available.
 611  */
 612 static struct ppa_addr rrpc_map_page(struct rrpc *rrpc, sector_t laddr,
 613                                                                 int is_gc)
 614 {
 615         struct nvm_tgt_dev *tgt_dev = rrpc->dev;
 616         struct rrpc_lun *rlun;
 617         struct rrpc_block *rblk, **cur_rblk;
 618         struct rrpc_addr *p;
 619         struct ppa_addr ppa;
 620         u64 paddr;
 621         int gc_force = 0;
 622
 623         ppa.ppa = ADDR_EMPTY;
 624         rlun = rrpc_get_lun_rr(rrpc, is_gc);
 625
 626         if (!is_gc && rlun->nr_free_blocks < rrpc->nr_luns * 4)
 627                 return ppa;
 628
 629         /*
 630          * page allocation steps:
 631          * 1. Try to allocate new page from current rblk
 632          * 2a. If succeed, proceed to map it in and return
 633          * 2b. If fail, first try to allocate a new block from media manger,
 634          *     and then retry step 1. Retry until the normal block pool is
 635          *     exhausted.
 636          * 3. If exhausted, and garbage collector is requesting the block,
 637          *    go to the reserved block and retry step 1.
 638          *    In the case that this fails as well, or it is not GC
 639          *    requesting, report not able to retrieve a block and let the
 640          *    caller handle further processing.
 641          */
 642
 643         spin_lock(&rlun->lock);
 644         cur_rblk = &rlun->cur;
 645         rblk = rlun->cur;
 646 retry:
 647         paddr = rrpc_alloc_addr(rrpc, rblk);
 648
 649         if (paddr != ADDR_EMPTY)
 650                 goto done;
 651
 652         if (!list_empty(&rlun->wblk_list)) {
 653 new_blk:
 654                 rblk = list_first_entry(&rlun->wblk_list, struct rrpc_block,
 655                                                                         prio);
 656                 rrpc_set_lun_cur(rlun, rblk, cur_rblk);
 657                 list_del(&rblk->prio);
 658                 goto retry;
 659         }
 660         spin_unlock(&rlun->lock);
 661
 662         rblk = rrpc_get_blk(rrpc, rlun, gc_force);
 663         if (rblk) {
 664                 spin_lock(&rlun->lock);
 665                 list_add_tail(&rblk->prio, &rlun->wblk_list);
 666                 /*
 667                  * another thread might already have added a new block,
 668                  * Therefore, make sure that one is used, instead of the
 669                  * one just added.
 670                  */
 671                 goto new_blk;
 672         }
 673
 674         if (unlikely(is_gc) && !gc_force) {
 675                 /* retry from emergency gc block */
 676                 cur_rblk = &rlun->gc_cur;
 677                 rblk = rlun->gc_cur;
 678                 gc_force = 1;
 679                 spin_lock(&rlun->lock);
 680                 goto retry;
 681         }
 682
 683         pr_err("rrpc: failed to allocate new block\n");
 684         return ppa;
 685 done:
 686         spin_unlock(&rlun->lock);
 687         p = rrpc_update_map(rrpc, laddr, rblk, paddr);
 688         if (!p)
 689                 return ppa;
 690
 691         /* return global address */
 692         return rrpc_ppa_to_gaddr(tgt_dev, p);
 693 }
 694
 695 static void rrpc_run_gc(struct rrpc *rrpc, struct rrpc_block *rblk)
 696 {
 697         struct rrpc_block_gc *gcb;
 698
 699         gcb = mempool_alloc(rrpc->gcb_pool, GFP_ATOMIC);
 700         if (!gcb) {
 701                 pr_err("rrpc: unable to queue block for gc.");
 702                 return;
 703         }
 704
 705         gcb->rrpc = rrpc;
 706         gcb->rblk = rblk;
 707
 708         INIT_WORK(&gcb->ws_gc, rrpc_gc_queue);
 709         queue_work(rrpc->kgc_wq, &gcb->ws_gc);
 710 }
 711
 712 static struct rrpc_lun *rrpc_ppa_to_lun(struct rrpc *rrpc, struct ppa_addr p)
 713 {
 714         struct rrpc_lun *rlun = NULL;
 715         int i;
 716
 717         for (i = 0; i < rrpc->nr_luns; i++) {
 718                 if (rrpc->luns[i].bppa.g.ch == p.g.ch &&
 719                                 rrpc->luns[i].bppa.g.lun == p.g.lun) {
 720                         rlun = &rrpc->luns[i];
 721                         break;
 722                 }
 723         }
 724
 725         return rlun;
 726 }
 727
 728 static void __rrpc_mark_bad_block(struct rrpc *rrpc, struct ppa_addr ppa)
 729 {
 730         struct nvm_tgt_dev *dev = rrpc->dev;
 731         struct rrpc_lun *rlun;
 732         struct rrpc_block *rblk;
 733
 734         rlun = rrpc_ppa_to_lun(rrpc, ppa);
 735         rblk = &rlun->blocks[ppa.g.blk];
 736         rblk->state = NVM_BLK_ST_BAD;
 737
 738         nvm_set_tgt_bb_tbl(dev, &ppa, 1, NVM_BLK_T_GRWN_BAD);
 739 }
 740
 741 static void rrpc_mark_bad_block(struct rrpc *rrpc, struct nvm_rq *rqd)
 742 {
 743         void *comp_bits = &rqd->ppa_status;
 744         struct ppa_addr ppa, prev_ppa;
 745         int nr_ppas = rqd->nr_ppas;
 746         int bit;
 747
 748         if (rqd->nr_ppas == 1)
 749                 __rrpc_mark_bad_block(rrpc, rqd->ppa_addr);
 750
 751         ppa_set_empty(&prev_ppa);
 752         bit = -1;
 753         while ((bit = find_next_bit(comp_bits, nr_ppas, bit + 1)) < nr_ppas) {
 754                 ppa = rqd->ppa_list[bit];
 755                 if (ppa_cmp_blk(ppa, prev_ppa))
 756                         continue;
 757
 758                 __rrpc_mark_bad_block(rrpc, ppa);
 759         }
 760 }
 761
 762 static void rrpc_end_io_write(struct rrpc *rrpc, struct rrpc_rq *rrqd,
 763                                                 sector_t laddr, uint8_t npages)
 764 {
 765         struct nvm_tgt_dev *dev = rrpc->dev;
 766         struct rrpc_addr *p;
 767         struct rrpc_block *rblk;
 768         int cmnt_size, i;
 769
 770         for (i = 0; i < npages; i++) {
 771                 p = &rrpc->trans_map[laddr + i];
 772                 rblk = p->rblk;
 773
 774                 cmnt_size = atomic_inc_return(&rblk->data_cmnt_size);
 775                 if (unlikely(cmnt_size == dev->geo.sec_per_blk))
 776                         rrpc_run_gc(rrpc, rblk);
 777         }
 778 }
 779
 780 static void rrpc_end_io(struct nvm_rq *rqd)
 781 {
 782         struct rrpc *rrpc = container_of(rqd->ins, struct rrpc, instance);
 783         struct nvm_tgt_dev *dev = rrpc->dev;
 784         struct rrpc_rq *rrqd = nvm_rq_to_pdu(rqd);
 785         uint8_t npages = rqd->nr_ppas;
 786         sector_t laddr = rrpc_get_laddr(rqd->bio) - npages;
 787
 788         if (bio_data_dir(rqd->bio) == WRITE) {
 789                 if (rqd->error == NVM_RSP_ERR_FAILWRITE)
 790                         rrpc_mark_bad_block(rrpc, rqd);
 791
 792                 rrpc_end_io_write(rrpc, rrqd, laddr, npages);
 793         }
 794
 795         bio_put(rqd->bio);
 796
 797         if (rrqd->flags & NVM_IOTYPE_GC)
 798                 return;
 799
 800         rrpc_unlock_rq(rrpc, rqd);
 801
 802         if (npages > 1)
 803                 nvm_dev_dma_free(dev->parent, rqd->ppa_list, rqd->dma_ppa_list);
 804
 805         mempool_free(rqd, rrpc->rq_pool);
 806 }
 807
 808 static int rrpc_read_ppalist_rq(struct rrpc *rrpc, struct bio *bio,
 809                         struct nvm_rq *rqd, unsigned long flags, int npages)
 810 {
 811         struct nvm_tgt_dev *dev = rrpc->dev;
 812         struct rrpc_inflight_rq *r = rrpc_get_inflight_rq(rqd);
 813         struct rrpc_addr *gp;
 814         sector_t laddr = rrpc_get_laddr(bio);
 815         int is_gc = flags & NVM_IOTYPE_GC;
 816         int i;
 817
 818         if (!is_gc && rrpc_lock_rq(rrpc, bio, rqd)) {
 819                 nvm_dev_dma_free(dev->parent, rqd->ppa_list, rqd->dma_ppa_list);
 820                 return NVM_IO_REQUEUE;
 821         }
 822
 823         for (i = 0; i < npages; i++) {
 824                 /* We assume that mapping occurs at 4KB granularity */
 825                 BUG_ON(!(laddr + i >= 0 && laddr + i < rrpc->nr_sects));
 826                 gp = &rrpc->trans_map[laddr + i];
 827
 828                 if (gp->rblk) {
 829                         rqd->ppa_list[i] = rrpc_ppa_to_gaddr(dev, gp);
 830                 } else {
 831                         BUG_ON(is_gc);
 832                         rrpc_unlock_laddr(rrpc, r);
 833                         nvm_dev_dma_free(dev->parent, rqd->ppa_list,
 834                                                         rqd->dma_ppa_list);
 835                         return NVM_IO_DONE;
 836                 }
 837         }
 838
 839         rqd->opcode = NVM_OP_HBREAD;
 840
 841         return NVM_IO_OK;
 842 }
 843
 844 static int rrpc_read_rq(struct rrpc *rrpc, struct bio *bio, struct nvm_rq *rqd,
 845                                                         unsigned long flags)
 846 {
 847         int is_gc = flags & NVM_IOTYPE_GC;
 848         sector_t laddr = rrpc_get_laddr(bio);
 849         struct rrpc_addr *gp;
 850
 851         if (!is_gc && rrpc_lock_rq(rrpc, bio, rqd))
 852                 return NVM_IO_REQUEUE;
 853
 854         BUG_ON(!(laddr >= 0 && laddr < rrpc->nr_sects));
 855         gp = &rrpc->trans_map[laddr];
 856
 857         if (gp->rblk) {
 858                 rqd->ppa_addr = rrpc_ppa_to_gaddr(rrpc->dev, gp);
 859         } else {
 860                 BUG_ON(is_gc);
 861                 rrpc_unlock_rq(rrpc, rqd);
 862                 return NVM_IO_DONE;
 863         }
 864
 865         rqd->opcode = NVM_OP_HBREAD;
 866
 867         return NVM_IO_OK;
 868 }
 869
 870 static int rrpc_write_ppalist_rq(struct rrpc *rrpc, struct bio *bio,
 871                         struct nvm_rq *rqd, unsigned long flags, int npages)
 872 {
 873         struct nvm_tgt_dev *dev = rrpc->dev;
 874         struct rrpc_inflight_rq *r = rrpc_get_inflight_rq(rqd);
 875         struct ppa_addr p;
 876         sector_t laddr = rrpc_get_laddr(bio);
 877         int is_gc = flags & NVM_IOTYPE_GC;
 878         int i;
 879
 880         if (!is_gc && rrpc_lock_rq(rrpc, bio, rqd)) {
 881                 nvm_dev_dma_free(dev->parent, rqd->ppa_list, rqd->dma_ppa_list);
 882                 return NVM_IO_REQUEUE;
 883         }
 884
 885         for (i = 0; i < npages; i++) {
 886                 /* We assume that mapping occurs at 4KB granularity */
 887                 p = rrpc_map_page(rrpc, laddr + i, is_gc);
 888                 if (p.ppa == ADDR_EMPTY) {
 889                         BUG_ON(is_gc);
 890                         rrpc_unlock_laddr(rrpc, r);
 891                         nvm_dev_dma_free(dev->parent, rqd->ppa_list,
 892                                                         rqd->dma_ppa_list);
 893                         rrpc_gc_kick(rrpc);
 894                         return NVM_IO_REQUEUE;
 895                 }
 896
 897                 rqd->ppa_list[i] = p;
 898         }
 899
 900         rqd->opcode = NVM_OP_HBWRITE;
 901
 902         return NVM_IO_OK;
 903 }
 904
 905 static int rrpc_write_rq(struct rrpc *rrpc, struct bio *bio,
 906                                 struct nvm_rq *rqd, unsigned long flags)
 907 {
 908         struct ppa_addr p;
 909         int is_gc = flags & NVM_IOTYPE_GC;
 910         sector_t laddr = rrpc_get_laddr(bio);
 911
 912         if (!is_gc && rrpc_lock_rq(rrpc, bio, rqd))
 913                 return NVM_IO_REQUEUE;
 914
 915         p = rrpc_map_page(rrpc, laddr, is_gc);
 916         if (p.ppa == ADDR_EMPTY) {
 917                 BUG_ON(is_gc);
 918                 rrpc_unlock_rq(rrpc, rqd);
 919                 rrpc_gc_kick(rrpc);
 920                 return NVM_IO_REQUEUE;
 921         }
 922
 923         rqd->ppa_addr = p;
 924         rqd->opcode = NVM_OP_HBWRITE;
 925
 926         return NVM_IO_OK;
 927 }
 928
 929 static int rrpc_setup_rq(struct rrpc *rrpc, struct bio *bio,
 930                         struct nvm_rq *rqd, unsigned long flags, uint8_t npages)
 931 {
 932         struct nvm_tgt_dev *dev = rrpc->dev;
 933
 934         if (npages > 1) {
 935                 rqd->ppa_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
 936                                                         &rqd->dma_ppa_list);
 937                 if (!rqd->ppa_list) {
 938                         pr_err("rrpc: not able to allocate ppa list\n");
 939                         return NVM_IO_ERR;
 940                 }
 941
 942                 if (bio_op(bio) == REQ_OP_WRITE)
 943                         return rrpc_write_ppalist_rq(rrpc, bio, rqd, flags,
 944                                                                         npages);
 945
 946                 return rrpc_read_ppalist_rq(rrpc, bio, rqd, flags, npages);
 947         }
 948
 949         if (bio_op(bio) == REQ_OP_WRITE)
 950                 return rrpc_write_rq(rrpc, bio, rqd, flags);
 951
 952         return rrpc_read_rq(rrpc, bio, rqd, flags);
 953 }
 954
 955 static int rrpc_submit_io(struct rrpc *rrpc, struct bio *bio,
 956                                 struct nvm_rq *rqd, unsigned long flags)
 957 {
 958         struct nvm_tgt_dev *dev = rrpc->dev;
 959         struct rrpc_rq *rrq = nvm_rq_to_pdu(rqd);
 960         uint8_t nr_pages = rrpc_get_pages(bio);
 961         int bio_size = bio_sectors(bio) << 9;
 962         int err;
 963
 964         if (bio_size < dev->geo.sec_size)
 965                 return NVM_IO_ERR;
 966         else if (bio_size > dev->geo.max_rq_size)
 967                 return NVM_IO_ERR;
 968
 969         err = rrpc_setup_rq(rrpc, bio, rqd, flags, nr_pages);
 970         if (err)
 971                 return err;
 972
 973         bio_get(bio);
 974         rqd->bio = bio;
 975         rqd->ins = &rrpc->instance;
 976         rqd->nr_ppas = nr_pages;
 977         rrq->flags = flags;
 978
 979         err = nvm_submit_io(dev, rqd);
 980         if (err) {
 981                 pr_err("rrpc: I/O submission failed: %d\n", err);
 982                 bio_put(bio);
 983                 if (!(flags & NVM_IOTYPE_GC)) {
 984                         rrpc_unlock_rq(rrpc, rqd);
 985                         if (rqd->nr_ppas > 1)
 986                                 nvm_dev_dma_free(dev->parent, rqd->ppa_list,
 987                                                         rqd->dma_ppa_list);
 988                 }
 989                 return NVM_IO_ERR;
 990         }
 991
 992         return NVM_IO_OK;
 993 }
 994
 995 static blk_qc_t rrpc_make_rq(struct request_queue *q, struct bio *bio)
 996 {
 997         struct rrpc *rrpc = q->queuedata;
 998         struct nvm_rq *rqd;
 999         int err;
1000
1001         blk_queue_split(q, &bio, q->bio_split);
1002
1003         if (bio_op(bio) == REQ_OP_DISCARD) {
1004                 rrpc_discard(rrpc, bio);
1005                 return BLK_QC_T_NONE;
1006         }
1007
1008         rqd = mempool_alloc(rrpc->rq_pool, GFP_KERNEL);
1009         if (!rqd) {
1010                 pr_err_ratelimited("rrpc: not able to queue bio.");
1011                 bio_io_error(bio);
1012                 return BLK_QC_T_NONE;
1013         }
1014         memset(rqd, 0, sizeof(struct nvm_rq));
1015
1016         err = rrpc_submit_io(rrpc, bio, rqd, NVM_IOTYPE_NONE);
1017         switch (err) {
1018         case NVM_IO_OK:
1019                 return BLK_QC_T_NONE;
1020         case NVM_IO_ERR:
1021                 bio_io_error(bio);
1022                 break;
1023         case NVM_IO_DONE:
1024                 bio_endio(bio);
1025                 break;
1026         case NVM_IO_REQUEUE:
1027                 spin_lock(&rrpc->bio_lock);
1028                 bio_list_add(&rrpc->requeue_bios, bio);
1029                 spin_unlock(&rrpc->bio_lock);
1030                 queue_work(rrpc->kgc_wq, &rrpc->ws_requeue);
1031                 break;
1032         }
1033
1034         mempool_free(rqd, rrpc->rq_pool);
1035         return BLK_QC_T_NONE;
1036 }
1037
1038 static void rrpc_requeue(struct work_struct *work)
1039 {
1040         struct rrpc *rrpc = container_of(work, struct rrpc, ws_requeue);
1041         struct bio_list bios;
1042         struct bio *bio;
1043
1044         bio_list_init(&bios);
1045
1046         spin_lock(&rrpc->bio_lock);
1047         bio_list_merge(&bios, &rrpc->requeue_bios);
1048         bio_list_init(&rrpc->requeue_bios);
1049         spin_unlock(&rrpc->bio_lock);
1050
1051         while ((bio = bio_list_pop(&bios)))
1052                 rrpc_make_rq(rrpc->disk->queue, bio);
1053 }
1054
1055 static void rrpc_gc_free(struct rrpc *rrpc)
1056 {
1057         if (rrpc->krqd_wq)
1058                 destroy_workqueue(rrpc->krqd_wq);
1059
1060         if (rrpc->kgc_wq)
1061                 destroy_workqueue(rrpc->kgc_wq);
1062 }
1063
1064 static int rrpc_gc_init(struct rrpc *rrpc)
1065 {
1066         rrpc->krqd_wq = alloc_workqueue("rrpc-lun", WQ_MEM_RECLAIM|WQ_UNBOUND,
1067                                                                 rrpc->nr_luns);
1068         if (!rrpc->krqd_wq)
1069                 return -ENOMEM;
1070
1071         rrpc->kgc_wq = alloc_workqueue("rrpc-bg", WQ_MEM_RECLAIM, 1);
1072         if (!rrpc->kgc_wq)
1073                 return -ENOMEM;
1074
1075         setup_timer(&rrpc->gc_timer, rrpc_gc_timer, (unsigned long)rrpc);
1076
1077         return 0;
1078 }
1079
1080 static void rrpc_map_free(struct rrpc *rrpc)
1081 {
1082         vfree(rrpc->rev_trans_map);
1083         vfree(rrpc->trans_map);
1084 }
1085
1086 static int rrpc_l2p_update(u64 slba, u32 nlb, __le64 *entries, void *private)
1087 {
1088         struct rrpc *rrpc = (struct rrpc *)private;
1089         struct nvm_tgt_dev *dev = rrpc->dev;
1090         struct rrpc_addr *addr = rrpc->trans_map + slba;
1091         struct rrpc_rev_addr *raddr = rrpc->rev_trans_map;
1092         struct rrpc_lun *rlun;
1093         struct rrpc_block *rblk;
1094         u64 i;
1095
1096         for (i = 0; i < nlb; i++) {
1097                 struct ppa_addr gaddr;
1098                 u64 pba = le64_to_cpu(entries[i]);
1099                 unsigned int mod;
1100
1101                 /* LNVM treats address-spaces as silos, LBA and PBA are
1102                  * equally large and zero-indexed.
1103                  */
1104                 if (unlikely(pba >= dev->total_secs && pba != U64_MAX)) {
1105                         pr_err("nvm: L2P data entry is out of bounds!\n");
1106                         pr_err("nvm: Maybe loaded an old target L2P\n");
1107                         return -EINVAL;
1108                 }
1109
1110                 /* Address zero is a special one. The first page on a disk is
1111                  * protected. As it often holds internal device boot
1112                  * information.
1113                  */
1114                 if (!pba)
1115                         continue;
1116
1117                 div_u64_rem(pba, rrpc->nr_sects, &mod);
1118
1119                 gaddr = rrpc_recov_addr(dev, pba);
1120                 rlun = rrpc_ppa_to_lun(rrpc, gaddr);
1121                 if (!rlun) {
1122                         pr_err("rrpc: l2p corruption on lba %llu\n",
1123                                                         slba + i);
1124                         return -EINVAL;
1125                 }
1126
1127                 rblk = &rlun->blocks[gaddr.g.blk];
1128                 if (!rblk->state) {
1129                         /* at this point, we don't know anything about the
1130                          * block. It's up to the FTL on top to re-etablish the
1131                          * block state. The block is assumed to be open.
1132                          */
1133                         list_move_tail(&rblk->list, &rlun->used_list);
1134                         rblk->state = NVM_BLK_ST_TGT;
1135                         rlun->nr_free_blocks--;
1136                 }
1137
1138                 addr[i].addr = pba;
1139                 addr[i].rblk = rblk;
1140                 raddr[mod].addr = slba + i;
1141         }
1142
1143         return 0;
1144 }
1145
1146 static int rrpc_map_init(struct rrpc *rrpc)
1147 {
1148         struct nvm_tgt_dev *dev = rrpc->dev;
1149         sector_t i;
1150         int ret;
1151
1152         rrpc->trans_map = vzalloc(sizeof(struct rrpc_addr) * rrpc->nr_sects);
1153         if (!rrpc->trans_map)
1154                 return -ENOMEM;
1155
1156         rrpc->rev_trans_map = vmalloc(sizeof(struct rrpc_rev_addr)
1157                                                         * rrpc->nr_sects);
1158         if (!rrpc->rev_trans_map)
1159                 return -ENOMEM;
1160
1161         for (i = 0; i < rrpc->nr_sects; i++) {
1162                 struct rrpc_addr *p = &rrpc->trans_map[i];
1163                 struct rrpc_rev_addr *r = &rrpc->rev_trans_map[i];
1164
1165                 p->addr = ADDR_EMPTY;
1166                 r->addr = ADDR_EMPTY;
1167         }
1168
1169         /* Bring up the mapping table from device */
1170         ret = nvm_get_l2p_tbl(dev, rrpc->soffset, rrpc->nr_sects,
1171                                                         rrpc_l2p_update, rrpc);
1172         if (ret) {
1173                 pr_err("nvm: rrpc: could not read L2P table.\n");
1174                 return -EINVAL;
1175         }
1176
1177         return 0;
1178 }
1179
1180 /* Minimum pages needed within a lun */
1181 #define PAGE_POOL_SIZE 16
1182 #define ADDR_POOL_SIZE 64
1183
1184 static int rrpc_core_init(struct rrpc *rrpc)
1185 {
1186         down_write(&rrpc_lock);
1187         if (!rrpc_gcb_cache) {
1188                 rrpc_gcb_cache = kmem_cache_create("rrpc_gcb",
1189                                 sizeof(struct rrpc_block_gc), 0, 0, NULL);
1190                 if (!rrpc_gcb_cache) {
1191                         up_write(&rrpc_lock);
1192                         return -ENOMEM;
1193                 }
1194
1195                 rrpc_rq_cache = kmem_cache_create("rrpc_rq",
1196                                 sizeof(struct nvm_rq) + sizeof(struct rrpc_rq),
1197                                 0, 0, NULL);
1198                 if (!rrpc_rq_cache) {
1199                         kmem_cache_destroy(rrpc_gcb_cache);
1200                         up_write(&rrpc_lock);
1201                         return -ENOMEM;
1202                 }
1203         }
1204         up_write(&rrpc_lock);
1205
1206         rrpc->page_pool = mempool_create_page_pool(PAGE_POOL_SIZE, 0);
1207         if (!rrpc->page_pool)
1208                 return -ENOMEM;
1209
1210         rrpc->gcb_pool = mempool_create_slab_pool(rrpc->dev->geo.nr_luns,
1211                                                                 rrpc_gcb_cache);
1212         if (!rrpc->gcb_pool)
1213                 return -ENOMEM;
1214
1215         rrpc->rq_pool = mempool_create_slab_pool(64, rrpc_rq_cache);
1216         if (!rrpc->rq_pool)
1217                 return -ENOMEM;
1218
1219         spin_lock_init(&rrpc->inflights.lock);
1220         INIT_LIST_HEAD(&rrpc->inflights.reqs);
1221
1222         return 0;
1223 }
1224
1225 static void rrpc_core_free(struct rrpc *rrpc)
1226 {
1227         mempool_destroy(rrpc->page_pool);
1228         mempool_destroy(rrpc->gcb_pool);
1229         mempool_destroy(rrpc->rq_pool);
1230 }
1231
1232 static void rrpc_luns_free(struct rrpc *rrpc)
1233 {
1234         struct rrpc_lun *rlun;
1235         int i;
1236
1237         if (!rrpc->luns)
1238                 return;
1239
1240         for (i = 0; i < rrpc->nr_luns; i++) {
1241                 rlun = &rrpc->luns[i];
1242                 vfree(rlun->blocks);
1243         }
1244
1245         kfree(rrpc->luns);
1246 }
1247
1248 static int rrpc_bb_discovery(struct nvm_tgt_dev *dev, struct rrpc_lun *rlun)
1249 {
1250         struct nvm_geo *geo = &dev->geo;
1251         struct rrpc_block *rblk;
1252         struct ppa_addr ppa;
1253         u8 *blks;
1254         int nr_blks;
1255         int i;
1256         int ret;
1257
1258         if (!dev->parent->ops->get_bb_tbl)
1259                 return 0;
1260
1261         nr_blks = geo->blks_per_lun * geo->plane_mode;
1262         blks = kmalloc(nr_blks, GFP_KERNEL);
1263         if (!blks)
1264                 return -ENOMEM;
1265
1266         ppa.ppa = 0;
1267         ppa.g.ch = rlun->bppa.g.ch;
1268         ppa.g.lun = rlun->bppa.g.lun;
1269
1270         ret = nvm_get_tgt_bb_tbl(dev, ppa, blks);
1271         if (ret) {
1272                 pr_err("rrpc: could not get BB table\n");
1273                 goto out;
1274         }
1275
1276         nr_blks = nvm_bb_tbl_fold(dev->parent, blks, nr_blks);
1277         if (nr_blks < 0)
1278                 return nr_blks;
1279
1280         for (i = 0; i < nr_blks; i++) {
1281                 if (blks[i] == NVM_BLK_T_FREE)
1282                         continue;
1283
1284                 rblk = &rlun->blocks[i];
1285                 list_move_tail(&rblk->list, &rlun->bb_list);
1286                 rblk->state = NVM_BLK_ST_BAD;
1287                 rlun->nr_free_blocks--;
1288         }
1289
1290 out:
1291         kfree(blks);
1292         return ret;
1293 }
1294
1295 static void rrpc_set_lun_ppa(struct rrpc_lun *rlun, struct ppa_addr ppa)
1296 {
1297         rlun->bppa.ppa = 0;
1298         rlun->bppa.g.ch = ppa.g.ch;
1299         rlun->bppa.g.lun = ppa.g.lun;
1300 }
1301
1302 static int rrpc_luns_init(struct rrpc *rrpc, struct ppa_addr *luns)
1303 {
1304         struct nvm_tgt_dev *dev = rrpc->dev;
1305         struct nvm_geo *geo = &dev->geo;
1306         struct rrpc_lun *rlun;
1307         int i, j, ret = -EINVAL;
1308
1309         if (geo->sec_per_blk > MAX_INVALID_PAGES_STORAGE * BITS_PER_LONG) {
1310                 pr_err("rrpc: number of pages per block too high.");
1311                 return -EINVAL;
1312         }
1313
1314         spin_lock_init(&rrpc->rev_lock);
1315
1316         rrpc->luns = kcalloc(rrpc->nr_luns, sizeof(struct rrpc_lun),
1317                                                                 GFP_KERNEL);
1318         if (!rrpc->luns)
1319                 return -ENOMEM;
1320
1321         /* 1:1 mapping */
1322         for (i = 0; i < rrpc->nr_luns; i++) {
1323                 rlun = &rrpc->luns[i];
1324                 rlun->id = i;
1325                 rrpc_set_lun_ppa(rlun, luns[i]);
1326                 rlun->blocks = vzalloc(sizeof(struct rrpc_block) *
1327                                                         geo->blks_per_lun);
1328                 if (!rlun->blocks) {
1329                         ret = -ENOMEM;
1330                         goto err;
1331                 }
1332
1333                 INIT_LIST_HEAD(&rlun->free_list);
1334                 INIT_LIST_HEAD(&rlun->used_list);
1335                 INIT_LIST_HEAD(&rlun->bb_list);
1336
1337                 for (j = 0; j < geo->blks_per_lun; j++) {
1338                         struct rrpc_block *rblk = &rlun->blocks[j];
1339
1340                         rblk->id = j;
1341                         rblk->rlun = rlun;
1342                         rblk->state = NVM_BLK_T_FREE;
1343                         INIT_LIST_HEAD(&rblk->prio);
1344                         INIT_LIST_HEAD(&rblk->list);
1345                         spin_lock_init(&rblk->lock);
1346
1347                         list_add_tail(&rblk->list, &rlun->free_list);
1348                 }
1349
1350                 rlun->rrpc = rrpc;
1351                 rlun->nr_free_blocks = geo->blks_per_lun;
1352                 rlun->reserved_blocks = 2; /* for GC only */
1353
1354                 INIT_LIST_HEAD(&rlun->prio_list);
1355                 INIT_LIST_HEAD(&rlun->wblk_list);
1356
1357                 INIT_WORK(&rlun->ws_gc, rrpc_lun_gc);
1358                 spin_lock_init(&rlun->lock);
1359
1360                 if (rrpc_bb_discovery(dev, rlun))
1361                         goto err;
1362
1363         }
1364
1365         return 0;
1366 err:
1367         return ret;
1368 }
1369
1370 /* returns 0 on success and stores the beginning address in *begin */
1371 static int rrpc_area_init(struct rrpc *rrpc, sector_t *begin)
1372 {
1373         struct nvm_tgt_dev *dev = rrpc->dev;
1374         sector_t size = rrpc->nr_sects * dev->geo.sec_size;
1375         int ret;
1376
1377         size >>= 9;
1378
1379         ret = nvm_get_area(dev, begin, size);
1380         if (!ret)
1381                 *begin >>= (ilog2(dev->geo.sec_size) - 9);
1382
1383         return ret;
1384 }
1385
1386 static void rrpc_area_free(struct rrpc *rrpc)
1387 {
1388         struct nvm_tgt_dev *dev = rrpc->dev;
1389         sector_t begin = rrpc->soffset << (ilog2(dev->geo.sec_size) - 9);
1390
1391         nvm_put_area(dev, begin);
1392 }
1393
1394 static void rrpc_free(struct rrpc *rrpc)
1395 {
1396         rrpc_gc_free(rrpc);
1397         rrpc_map_free(rrpc);
1398         rrpc_core_free(rrpc);
1399         rrpc_luns_free(rrpc);
1400         rrpc_area_free(rrpc);
1401
1402         kfree(rrpc);
1403 }
1404
1405 static void rrpc_exit(void *private)
1406 {
1407         struct rrpc *rrpc = private;
1408
1409         del_timer(&rrpc->gc_timer);
1410
1411         flush_workqueue(rrpc->krqd_wq);
1412         flush_workqueue(rrpc->kgc_wq);
1413
1414         rrpc_free(rrpc);
1415 }
1416
1417 static sector_t rrpc_capacity(void *private)
1418 {
1419         struct rrpc *rrpc = private;
1420         struct nvm_tgt_dev *dev = rrpc->dev;
1421         sector_t reserved, provisioned;
1422
1423         /* cur, gc, and two emergency blocks for each lun */
1424         reserved = rrpc->nr_luns * dev->geo.sec_per_blk * 4;
1425         provisioned = rrpc->nr_sects - reserved;
1426
1427         if (reserved > rrpc->nr_sects) {
1428                 pr_err("rrpc: not enough space available to expose storage.\n");
1429                 return 0;
1430         }
1431
1432         sector_div(provisioned, 10);
1433         return provisioned * 9 * NR_PHY_IN_LOG;
1434 }
1435
1436 /*
1437  * Looks up the logical address from reverse trans map and check if its valid by
1438  * comparing the logical to physical address with the physical address.
1439  * Returns 0 on free, otherwise 1 if in use
1440  */
1441 static void rrpc_block_map_update(struct rrpc *rrpc, struct rrpc_block *rblk)
1442 {
1443         struct nvm_tgt_dev *dev = rrpc->dev;
1444         int offset;
1445         struct rrpc_addr *laddr;
1446         u64 bpaddr, paddr, pladdr;
1447
1448         bpaddr = block_to_rel_addr(rrpc, rblk);
1449         for (offset = 0; offset < dev->geo.sec_per_blk; offset++) {
1450                 paddr = bpaddr + offset;
1451
1452                 pladdr = rrpc->rev_trans_map[paddr].addr;
1453                 if (pladdr == ADDR_EMPTY)
1454                         continue;
1455
1456                 laddr = &rrpc->trans_map[pladdr];
1457
1458                 if (paddr == laddr->addr) {
1459                         laddr->rblk = rblk;
1460                 } else {
1461                         set_bit(offset, rblk->invalid_pages);
1462                         rblk->nr_invalid_pages++;
1463                 }
1464         }
1465 }
1466
1467 static int rrpc_blocks_init(struct rrpc *rrpc)
1468 {
1469         struct nvm_tgt_dev *dev = rrpc->dev;
1470         struct rrpc_lun *rlun;
1471         struct rrpc_block *rblk;
1472         int lun_iter, blk_iter;
1473
1474         for (lun_iter = 0; lun_iter < rrpc->nr_luns; lun_iter++) {
1475                 rlun = &rrpc->luns[lun_iter];
1476
1477                 for (blk_iter = 0; blk_iter < dev->geo.blks_per_lun;
1478                                                                 blk_iter++) {
1479                         rblk = &rlun->blocks[blk_iter];
1480                         rrpc_block_map_update(rrpc, rblk);
1481                 }
1482         }
1483
1484         return 0;
1485 }
1486
1487 static int rrpc_luns_configure(struct rrpc *rrpc)
1488 {
1489         struct rrpc_lun *rlun;
1490         struct rrpc_block *rblk;
1491         int i;
1492
1493         for (i = 0; i < rrpc->nr_luns; i++) {
1494                 rlun = &rrpc->luns[i];
1495
1496                 rblk = rrpc_get_blk(rrpc, rlun, 0);
1497                 if (!rblk)
1498                         goto err;
1499                 rrpc_set_lun_cur(rlun, rblk, &rlun->cur);
1500
1501                 /* Emergency gc block */
1502                 rblk = rrpc_get_blk(rrpc, rlun, 1);
1503                 if (!rblk)
1504                         goto err;
1505                 rrpc_set_lun_cur(rlun, rblk, &rlun->gc_cur);
1506         }
1507
1508         return 0;
1509 err:
1510         rrpc_put_blks(rrpc);
1511         return -EINVAL;
1512 }
1513
1514 static struct nvm_tgt_type tt_rrpc;
1515
1516 static void *rrpc_init(struct nvm_tgt_dev *dev, struct gendisk *tdisk)
1517 {
1518         struct request_queue *bqueue = dev->q;
1519         struct request_queue *tqueue = tdisk->queue;
1520         struct nvm_geo *geo = &dev->geo;
1521         struct rrpc *rrpc;
1522         sector_t soffset;
1523         int ret;
1524
1525         if (!(dev->identity.dom & NVM_RSP_L2P)) {
1526                 pr_err("nvm: rrpc: device does not support l2p (%x)\n",
1527                                                         dev->identity.dom);
1528                 return ERR_PTR(-EINVAL);
1529         }
1530
1531         rrpc = kzalloc(sizeof(struct rrpc), GFP_KERNEL);
1532         if (!rrpc)
1533                 return ERR_PTR(-ENOMEM);
1534
1535         rrpc->instance.tt = &tt_rrpc;
1536         rrpc->dev = dev;
1537         rrpc->disk = tdisk;
1538
1539         bio_list_init(&rrpc->requeue_bios);
1540         spin_lock_init(&rrpc->bio_lock);
1541         INIT_WORK(&rrpc->ws_requeue, rrpc_requeue);
1542
1543         rrpc->nr_luns = geo->nr_luns;
1544         rrpc->nr_sects = (unsigned long long)geo->sec_per_lun * rrpc->nr_luns;
1545
1546         /* simple round-robin strategy */
1547         atomic_set(&rrpc->next_lun, -1);
1548
1549         ret = rrpc_area_init(rrpc, &soffset);
1550         if (ret < 0) {
1551                 pr_err("nvm: rrpc: could not initialize area\n");
1552                 return ERR_PTR(ret);
1553         }
1554         rrpc->soffset = soffset;
1555
1556         ret = rrpc_luns_init(rrpc, dev->luns);
1557         if (ret) {
1558                 pr_err("nvm: rrpc: could not initialize luns\n");
1559                 goto err;
1560         }
1561
1562         ret = rrpc_core_init(rrpc);
1563         if (ret) {
1564                 pr_err("nvm: rrpc: could not initialize core\n");
1565                 goto err;
1566         }
1567
1568         ret = rrpc_map_init(rrpc);
1569         if (ret) {
1570                 pr_err("nvm: rrpc: could not initialize maps\n");
1571                 goto err;
1572         }
1573
1574         ret = rrpc_blocks_init(rrpc);
1575         if (ret) {
1576                 pr_err("nvm: rrpc: could not initialize state for blocks\n");
1577                 goto err;
1578         }
1579
1580         ret = rrpc_luns_configure(rrpc);
1581         if (ret) {
1582                 pr_err("nvm: rrpc: not enough blocks available in LUNs.\n");
1583                 goto err;
1584         }
1585
1586         ret = rrpc_gc_init(rrpc);
1587         if (ret) {
1588                 pr_err("nvm: rrpc: could not initialize gc\n");
1589                 goto err;
1590         }
1591
1592         /* inherit the size from the underlying device */
1593         blk_queue_logical_block_size(tqueue, queue_physical_block_size(bqueue));
1594         blk_queue_max_hw_sectors(tqueue, queue_max_hw_sectors(bqueue));
1595
1596         pr_info("nvm: rrpc initialized with %u luns and %llu pages.\n",
1597                         rrpc->nr_luns, (unsigned long long)rrpc->nr_sects);
1598
1599         mod_timer(&rrpc->gc_timer, jiffies + msecs_to_jiffies(10));
1600
1601         return rrpc;
1602 err:
1603         rrpc_free(rrpc);
1604         return ERR_PTR(ret);
1605 }
1606
1607 /* round robin, page-based FTL, and cost-based GC */
1608 static struct nvm_tgt_type tt_rrpc = {
1609         .name           = "rrpc",
1610         .version        = {1, 0, 0},
1611
1612         .make_rq        = rrpc_make_rq,
1613         .capacity       = rrpc_capacity,
1614         .end_io         = rrpc_end_io,
1615
1616         .init           = rrpc_init,
1617         .exit           = rrpc_exit,
1618 };
1619
1620 static int __init rrpc_module_init(void)
1621 {
1622         return nvm_register_tgt_type(&tt_rrpc);
1623 }
1624
1625 static void rrpc_module_exit(void)
1626 {
1627         nvm_unregister_tgt_type(&tt_rrpc);
1628 }
1629
1630 module_init(rrpc_module_init);
1631 module_exit(rrpc_module_exit);
1632 MODULE_LICENSE("GPL v2");
1633 MODULE_DESCRIPTION("Block-Device Target for Open-Channel SSDs");