drivers/lightnvm/rrpc.c

   1 /*
   2  * Copyright (C) 2015 IT University of Copenhagen
   3  * Initial release: Matias Bjorling <m@bjorling.me>
   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 rrpc_block *rblk = a->rblk;
  32         unsigned int pg_offset;
  33
  34         lockdep_assert_held(&rrpc->rev_lock);
  35
  36         if (a->addr == ADDR_EMPTY || !rblk)
  37                 return;
  38
  39         spin_lock(&rblk->lock);
  40
  41         div_u64_rem(a->addr, rrpc->dev->sec_per_blk, &pg_offset);
  42         WARN_ON(test_and_set_bit(pg_offset, rblk->invalid_pages));
  43         rblk->nr_invalid_pages++;
  44
  45         spin_unlock(&rblk->lock);
  46
  47         rrpc->rev_trans_map[a->addr - rrpc->poffset].addr = ADDR_EMPTY;
  48 }
  49
  50 static void rrpc_invalidate_range(struct rrpc *rrpc, sector_t slba,
  51                                                         unsigned int len)
  52 {
  53         sector_t i;
  54
  55         spin_lock(&rrpc->rev_lock);
  56         for (i = slba; i < slba + len; i++) {
  57                 struct rrpc_addr *gp = &rrpc->trans_map[i];
  58
  59                 rrpc_page_invalidate(rrpc, gp);
  60                 gp->rblk = NULL;
  61         }
  62         spin_unlock(&rrpc->rev_lock);
  63 }
  64
  65 static struct nvm_rq *rrpc_inflight_laddr_acquire(struct rrpc *rrpc,
  66                                         sector_t laddr, unsigned int pages)
  67 {
  68         struct nvm_rq *rqd;
  69         struct rrpc_inflight_rq *inf;
  70
  71         rqd = mempool_alloc(rrpc->rq_pool, GFP_ATOMIC);
  72         if (!rqd)
  73                 return ERR_PTR(-ENOMEM);
  74
  75         inf = rrpc_get_inflight_rq(rqd);
  76         if (rrpc_lock_laddr(rrpc, laddr, pages, inf)) {
  77                 mempool_free(rqd, rrpc->rq_pool);
  78                 return NULL;
  79         }
  80
  81         return rqd;
  82 }
  83
  84 static void rrpc_inflight_laddr_release(struct rrpc *rrpc, struct nvm_rq *rqd)
  85 {
  86         struct rrpc_inflight_rq *inf = rrpc_get_inflight_rq(rqd);
  87
  88         rrpc_unlock_laddr(rrpc, inf);
  89
  90         mempool_free(rqd, rrpc->rq_pool);
  91 }
  92
  93 static void rrpc_discard(struct rrpc *rrpc, struct bio *bio)
  94 {
  95         sector_t slba = bio->bi_iter.bi_sector / NR_PHY_IN_LOG;
  96         sector_t len = bio->bi_iter.bi_size / RRPC_EXPOSED_PAGE_SIZE;
  97         struct nvm_rq *rqd;
  98
  99         while (1) {
 100                 rqd = rrpc_inflight_laddr_acquire(rrpc, slba, len);
 101                 if (rqd)
 102                         break;
 103
 104                 schedule();
 105         }
 106
 107         if (IS_ERR(rqd)) {
 108                 pr_err("rrpc: unable to acquire inflight IO\n");
 109                 bio_io_error(bio);
 110                 return;
 111         }
 112
 113         rrpc_invalidate_range(rrpc, slba, len);
 114         rrpc_inflight_laddr_release(rrpc, rqd);
 115 }
 116
 117 static int block_is_full(struct rrpc *rrpc, struct rrpc_block *rblk)
 118 {
 119         return (rblk->next_page == rrpc->dev->sec_per_blk);
 120 }
 121
 122 /* Calculate relative addr for the given block, considering instantiated LUNs */
 123 static u64 block_to_rel_addr(struct rrpc *rrpc, struct rrpc_block *rblk)
 124 {
 125         struct nvm_block *blk = rblk->parent;
 126         int lun_blk = blk->id % (rrpc->dev->blks_per_lun * rrpc->nr_luns);
 127
 128         return lun_blk * rrpc->dev->sec_per_blk;
 129 }
 130
 131 /* Calculate global addr for the given block */
 132 static u64 block_to_addr(struct rrpc *rrpc, struct rrpc_block *rblk)
 133 {
 134         struct nvm_block *blk = rblk->parent;
 135
 136         return blk->id * rrpc->dev->sec_per_blk;
 137 }
 138
 139 static struct ppa_addr linear_to_generic_addr(struct nvm_dev *dev,
 140                                                         struct ppa_addr r)
 141 {
 142         struct ppa_addr l;
 143         int secs, pgs, blks, luns;
 144         sector_t ppa = r.ppa;
 145
 146         l.ppa = 0;
 147
 148         div_u64_rem(ppa, dev->sec_per_pg, &secs);
 149         l.g.sec = secs;
 150
 151         sector_div(ppa, dev->sec_per_pg);
 152         div_u64_rem(ppa, dev->pgs_per_blk, &pgs);
 153         l.g.pg = pgs;
 154
 155         sector_div(ppa, dev->pgs_per_blk);
 156         div_u64_rem(ppa, dev->blks_per_lun, &blks);
 157         l.g.blk = blks;
 158
 159         sector_div(ppa, dev->blks_per_lun);
 160         div_u64_rem(ppa, dev->luns_per_chnl, &luns);
 161         l.g.lun = luns;
 162
 163         sector_div(ppa, dev->luns_per_chnl);
 164         l.g.ch = ppa;
 165
 166         return l;
 167 }
 168
 169 static struct ppa_addr rrpc_ppa_to_gaddr(struct nvm_dev *dev, u64 addr)
 170 {
 171         struct ppa_addr paddr;
 172
 173         paddr.ppa = addr;
 174         return linear_to_generic_addr(dev, paddr);
 175 }
 176
 177 /* requires lun->lock taken */
 178 static void rrpc_set_lun_cur(struct rrpc_lun *rlun, struct rrpc_block *new_rblk,
 179                                                 struct rrpc_block **cur_rblk)
 180 {
 181         struct rrpc *rrpc = rlun->rrpc;
 182
 183         if (*cur_rblk) {
 184                 spin_lock(&(*cur_rblk)->lock);
 185                 WARN_ON(!block_is_full(rrpc, *cur_rblk));
 186                 spin_unlock(&(*cur_rblk)->lock);
 187         }
 188         *cur_rblk = new_rblk;
 189 }
 190
 191 static struct rrpc_block *rrpc_get_blk(struct rrpc *rrpc, struct rrpc_lun *rlun,
 192                                                         unsigned long flags)
 193 {
 194         struct nvm_block *blk;
 195         struct rrpc_block *rblk;
 196
 197         blk = nvm_get_blk(rrpc->dev, rlun->parent, flags);
 198         if (!blk) {
 199                 pr_err("nvm: rrpc: cannot get new block from media manager\n");
 200                 return NULL;
 201         }
 202
 203         rblk = rrpc_get_rblk(rlun, blk->id);
 204         blk->priv = rblk;
 205         bitmap_zero(rblk->invalid_pages, rrpc->dev->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         nvm_put_blk(rrpc->dev, rblk->parent);
 216 }
 217
 218 static void rrpc_put_blks(struct rrpc *rrpc)
 219 {
 220         struct rrpc_lun *rlun;
 221         int i;
 222
 223         for (i = 0; i < rrpc->nr_luns; i++) {
 224                 rlun = &rrpc->luns[i];
 225                 if (rlun->cur)
 226                         rrpc_put_blk(rrpc, rlun->cur);
 227                 if (rlun->gc_cur)
 228                         rrpc_put_blk(rrpc, rlun->gc_cur);
 229         }
 230 }
 231
 232 static struct rrpc_lun *get_next_lun(struct rrpc *rrpc)
 233 {
 234         int next = atomic_inc_return(&rrpc->next_lun);
 235
 236         return &rrpc->luns[next % rrpc->nr_luns];
 237 }
 238
 239 static void rrpc_gc_kick(struct rrpc *rrpc)
 240 {
 241         struct rrpc_lun *rlun;
 242         unsigned int i;
 243
 244         for (i = 0; i < rrpc->nr_luns; i++) {
 245                 rlun = &rrpc->luns[i];
 246                 queue_work(rrpc->krqd_wq, &rlun->ws_gc);
 247         }
 248 }
 249
 250 /*
 251  * timed GC every interval.
 252  */
 253 static void rrpc_gc_timer(unsigned long data)
 254 {
 255         struct rrpc *rrpc = (struct rrpc *)data;
 256
 257         rrpc_gc_kick(rrpc);
 258         mod_timer(&rrpc->gc_timer, jiffies + msecs_to_jiffies(10));
 259 }
 260
 261 static void rrpc_end_sync_bio(struct bio *bio)
 262 {
 263         struct completion *waiting = bio->bi_private;
 264
 265         if (bio->bi_error)
 266                 pr_err("nvm: gc request failed (%u).\n", bio->bi_error);
 267
 268         complete(waiting);
 269 }
 270
 271 /*
 272  * rrpc_move_valid_pages -- migrate live data off the block
 273  * @rrpc: the 'rrpc' structure
 274  * @block: the block from which to migrate live pages
 275  *
 276  * Description:
 277  *   GC algorithms may call this function to migrate remaining live
 278  *   pages off the block prior to erasing it. This function blocks
 279  *   further execution until the operation is complete.
 280  */
 281 static int rrpc_move_valid_pages(struct rrpc *rrpc, struct rrpc_block *rblk)
 282 {
 283         struct request_queue *q = rrpc->dev->q;
 284         struct rrpc_rev_addr *rev;
 285         struct nvm_rq *rqd;
 286         struct bio *bio;
 287         struct page *page;
 288         int slot;
 289         int nr_sec_per_blk = rrpc->dev->sec_per_blk;
 290         u64 phys_addr;
 291         DECLARE_COMPLETION_ONSTACK(wait);
 292
 293         if (bitmap_full(rblk->invalid_pages, nr_sec_per_blk))
 294                 return 0;
 295
 296         bio = bio_alloc(GFP_NOIO, 1);
 297         if (!bio) {
 298                 pr_err("nvm: could not alloc bio to gc\n");
 299                 return -ENOMEM;
 300         }
 301
 302         page = mempool_alloc(rrpc->page_pool, GFP_NOIO);
 303         if (!page) {
 304                 bio_put(bio);
 305                 return -ENOMEM;
 306         }
 307
 308         while ((slot = find_first_zero_bit(rblk->invalid_pages,
 309                                             nr_sec_per_blk)) < nr_sec_per_blk) {
 310
 311                 /* Lock laddr */
 312                 phys_addr = rblk->parent->id * nr_sec_per_blk + slot;
 313
 314 try:
 315                 spin_lock(&rrpc->rev_lock);
 316                 /* Get logical address from physical to logical table */
 317                 rev = &rrpc->rev_trans_map[phys_addr - rrpc->poffset];
 318                 /* already updated by previous regular write */
 319                 if (rev->addr == ADDR_EMPTY) {
 320                         spin_unlock(&rrpc->rev_lock);
 321                         continue;
 322                 }
 323
 324                 rqd = rrpc_inflight_laddr_acquire(rrpc, rev->addr, 1);
 325                 if (IS_ERR_OR_NULL(rqd)) {
 326                         spin_unlock(&rrpc->rev_lock);
 327                         schedule();
 328                         goto try;
 329                 }
 330
 331                 spin_unlock(&rrpc->rev_lock);
 332
 333                 /* Perform read to do GC */
 334                 bio->bi_iter.bi_sector = rrpc_get_sector(rev->addr);
 335                 bio_set_op_attrs(bio,  REQ_OP_READ, 0);
 336                 bio->bi_private = &wait;
 337                 bio->bi_end_io = rrpc_end_sync_bio;
 338
 339                 /* TODO: may fail when EXP_PG_SIZE > PAGE_SIZE */
 340                 bio_add_pc_page(q, bio, page, RRPC_EXPOSED_PAGE_SIZE, 0);
 341
 342                 if (rrpc_submit_io(rrpc, bio, rqd, NVM_IOTYPE_GC)) {
 343                         pr_err("rrpc: gc read failed.\n");
 344                         rrpc_inflight_laddr_release(rrpc, rqd);
 345                         goto finished;
 346                 }
 347                 wait_for_completion_io(&wait);
 348                 if (bio->bi_error) {
 349                         rrpc_inflight_laddr_release(rrpc, rqd);
 350                         goto finished;
 351                 }
 352
 353                 bio_reset(bio);
 354                 reinit_completion(&wait);
 355
 356                 bio->bi_iter.bi_sector = rrpc_get_sector(rev->addr);
 357                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 358                 bio->bi_private = &wait;
 359                 bio->bi_end_io = rrpc_end_sync_bio;
 360
 361                 bio_add_pc_page(q, bio, page, RRPC_EXPOSED_PAGE_SIZE, 0);
 362
 363                 /* turn the command around and write the data back to a new
 364                  * address
 365                  */
 366                 if (rrpc_submit_io(rrpc, bio, rqd, NVM_IOTYPE_GC)) {
 367                         pr_err("rrpc: gc write failed.\n");
 368                         rrpc_inflight_laddr_release(rrpc, rqd);
 369                         goto finished;
 370                 }
 371                 wait_for_completion_io(&wait);
 372
 373                 rrpc_inflight_laddr_release(rrpc, rqd);
 374                 if (bio->bi_error)
 375                         goto finished;
 376
 377                 bio_reset(bio);
 378         }
 379
 380 finished:
 381         mempool_free(page, rrpc->page_pool);
 382         bio_put(bio);
 383
 384         if (!bitmap_full(rblk->invalid_pages, nr_sec_per_blk)) {
 385                 pr_err("nvm: failed to garbage collect block\n");
 386                 return -EIO;
 387         }
 388
 389         return 0;
 390 }
 391
 392 static void rrpc_block_gc(struct work_struct *work)
 393 {
 394         struct rrpc_block_gc *gcb = container_of(work, struct rrpc_block_gc,
 395                                                                         ws_gc);
 396         struct rrpc *rrpc = gcb->rrpc;
 397         struct rrpc_block *rblk = gcb->rblk;
 398         struct rrpc_lun *rlun = rblk->rlun;
 399         struct nvm_dev *dev = rrpc->dev;
 400
 401         mempool_free(gcb, rrpc->gcb_pool);
 402         pr_debug("nvm: block '%lu' being reclaimed\n", rblk->parent->id);
 403
 404         if (rrpc_move_valid_pages(rrpc, rblk))
 405                 goto put_back;
 406
 407         if (nvm_erase_blk(dev, rblk->parent, 0))
 408                 goto put_back;
 409
 410         rrpc_put_blk(rrpc, rblk);
 411
 412         return;
 413
 414 put_back:
 415         spin_lock(&rlun->lock);
 416         list_add_tail(&rblk->prio, &rlun->prio_list);
 417         spin_unlock(&rlun->lock);
 418 }
 419
 420 /* the block with highest number of invalid pages, will be in the beginning
 421  * of the list
 422  */
 423 static struct rrpc_block *rblock_max_invalid(struct rrpc_block *ra,
 424                                                         struct rrpc_block *rb)
 425 {
 426         if (ra->nr_invalid_pages == rb->nr_invalid_pages)
 427                 return ra;
 428
 429         return (ra->nr_invalid_pages < rb->nr_invalid_pages) ? rb : ra;
 430 }
 431
 432 /* linearly find the block with highest number of invalid pages
 433  * requires lun->lock
 434  */
 435 static struct rrpc_block *block_prio_find_max(struct rrpc_lun *rlun)
 436 {
 437         struct list_head *prio_list = &rlun->prio_list;
 438         struct rrpc_block *rblock, *max;
 439
 440         BUG_ON(list_empty(prio_list));
 441
 442         max = list_first_entry(prio_list, struct rrpc_block, prio);
 443         list_for_each_entry(rblock, prio_list, prio)
 444                 max = rblock_max_invalid(max, rblock);
 445
 446         return max;
 447 }
 448
 449 static void rrpc_lun_gc(struct work_struct *work)
 450 {
 451         struct rrpc_lun *rlun = container_of(work, struct rrpc_lun, ws_gc);
 452         struct rrpc *rrpc = rlun->rrpc;
 453         struct nvm_lun *lun = rlun->parent;
 454         struct rrpc_block_gc *gcb;
 455         unsigned int nr_blocks_need;
 456
 457         nr_blocks_need = rrpc->dev->blks_per_lun / GC_LIMIT_INVERSE;
 458
 459         if (nr_blocks_need < rrpc->nr_luns)
 460                 nr_blocks_need = rrpc->nr_luns;
 461
 462         spin_lock(&rlun->lock);
 463         while (nr_blocks_need > lun->nr_free_blocks &&
 464                                         !list_empty(&rlun->prio_list)) {
 465                 struct rrpc_block *rblock = block_prio_find_max(rlun);
 466                 struct nvm_block *block = rblock->parent;
 467
 468                 if (!rblock->nr_invalid_pages)
 469                         break;
 470
 471                 gcb = mempool_alloc(rrpc->gcb_pool, GFP_ATOMIC);
 472                 if (!gcb)
 473                         break;
 474
 475                 list_del_init(&rblock->prio);
 476
 477                 BUG_ON(!block_is_full(rrpc, rblock));
 478
 479                 pr_debug("rrpc: selected block '%lu' for GC\n", block->id);
 480
 481                 gcb->rrpc = rrpc;
 482                 gcb->rblk = rblock;
 483                 INIT_WORK(&gcb->ws_gc, rrpc_block_gc);
 484
 485                 queue_work(rrpc->kgc_wq, &gcb->ws_gc);
 486
 487                 nr_blocks_need--;
 488         }
 489         spin_unlock(&rlun->lock);
 490
 491         /* TODO: Hint that request queue can be started again */
 492 }
 493
 494 static void rrpc_gc_queue(struct work_struct *work)
 495 {
 496         struct rrpc_block_gc *gcb = container_of(work, struct rrpc_block_gc,
 497                                                                         ws_gc);
 498         struct rrpc *rrpc = gcb->rrpc;
 499         struct rrpc_block *rblk = gcb->rblk;
 500         struct rrpc_lun *rlun = rblk->rlun;
 501
 502         spin_lock(&rlun->lock);
 503         list_add_tail(&rblk->prio, &rlun->prio_list);
 504         spin_unlock(&rlun->lock);
 505
 506         mempool_free(gcb, rrpc->gcb_pool);
 507         pr_debug("nvm: block '%lu' is full, allow GC (sched)\n",
 508                                                         rblk->parent->id);
 509 }
 510
 511 static const struct block_device_operations rrpc_fops = {
 512         .owner          = THIS_MODULE,
 513 };
 514
 515 static struct rrpc_lun *rrpc_get_lun_rr(struct rrpc *rrpc, int is_gc)
 516 {
 517         unsigned int i;
 518         struct rrpc_lun *rlun, *max_free;
 519
 520         if (!is_gc)
 521                 return get_next_lun(rrpc);
 522
 523         /* during GC, we don't care about RR, instead we want to make
 524          * sure that we maintain evenness between the block luns.
 525          */
 526         max_free = &rrpc->luns[0];
 527         /* prevent GC-ing lun from devouring pages of a lun with
 528          * little free blocks. We don't take the lock as we only need an
 529          * estimate.
 530          */
 531         rrpc_for_each_lun(rrpc, rlun, i) {
 532                 if (rlun->parent->nr_free_blocks >
 533                                         max_free->parent->nr_free_blocks)
 534                         max_free = rlun;
 535         }
 536
 537         return max_free;
 538 }
 539
 540 static struct rrpc_addr *rrpc_update_map(struct rrpc *rrpc, sector_t laddr,
 541                                         struct rrpc_block *rblk, u64 paddr)
 542 {
 543         struct rrpc_addr *gp;
 544         struct rrpc_rev_addr *rev;
 545
 546         BUG_ON(laddr >= rrpc->nr_sects);
 547
 548         gp = &rrpc->trans_map[laddr];
 549         spin_lock(&rrpc->rev_lock);
 550         if (gp->rblk)
 551                 rrpc_page_invalidate(rrpc, gp);
 552
 553         gp->addr = paddr;
 554         gp->rblk = rblk;
 555
 556         rev = &rrpc->rev_trans_map[gp->addr - rrpc->poffset];
 557         rev->addr = laddr;
 558         spin_unlock(&rrpc->rev_lock);
 559
 560         return gp;
 561 }
 562
 563 static u64 rrpc_alloc_addr(struct rrpc *rrpc, struct rrpc_block *rblk)
 564 {
 565         u64 addr = ADDR_EMPTY;
 566
 567         spin_lock(&rblk->lock);
 568         if (block_is_full(rrpc, rblk))
 569                 goto out;
 570
 571         addr = block_to_addr(rrpc, rblk) + rblk->next_page;
 572
 573         rblk->next_page++;
 574 out:
 575         spin_unlock(&rblk->lock);
 576         return addr;
 577 }
 578
 579 /* Map logical address to a physical page. The mapping implements a round robin
 580  * approach and allocates a page from the next lun available.
 581  *
 582  * Returns rrpc_addr with the physical address and block. Returns NULL if no
 583  * blocks in the next rlun are available.
 584  */
 585 static struct rrpc_addr *rrpc_map_page(struct rrpc *rrpc, sector_t laddr,
 586                                                                 int is_gc)
 587 {
 588         struct rrpc_lun *rlun;
 589         struct rrpc_block *rblk, **cur_rblk;
 590         struct nvm_lun *lun;
 591         u64 paddr;
 592         int gc_force = 0;
 593
 594         rlun = rrpc_get_lun_rr(rrpc, is_gc);
 595         lun = rlun->parent;
 596
 597         if (!is_gc && lun->nr_free_blocks < rrpc->nr_luns * 4)
 598                 return NULL;
 599
 600         /*
 601          * page allocation steps:
 602          * 1. Try to allocate new page from current rblk
 603          * 2a. If succeed, proceed to map it in and return
 604          * 2b. If fail, first try to allocate a new block from media manger,
 605          *     and then retry step 1. Retry until the normal block pool is
 606          *     exhausted.
 607          * 3. If exhausted, and garbage collector is requesting the block,
 608          *    go to the reserved block and retry step 1.
 609          *    In the case that this fails as well, or it is not GC
 610          *    requesting, report not able to retrieve a block and let the
 611          *    caller handle further processing.
 612          */
 613
 614         spin_lock(&rlun->lock);
 615         cur_rblk = &rlun->cur;
 616         rblk = rlun->cur;
 617 retry:
 618         paddr = rrpc_alloc_addr(rrpc, rblk);
 619
 620         if (paddr != ADDR_EMPTY)
 621                 goto done;
 622
 623         if (!list_empty(&rlun->wblk_list)) {
 624 new_blk:
 625                 rblk = list_first_entry(&rlun->wblk_list, struct rrpc_block,
 626                                                                         prio);
 627                 rrpc_set_lun_cur(rlun, rblk, cur_rblk);
 628                 list_del(&rblk->prio);
 629                 goto retry;
 630         }
 631         spin_unlock(&rlun->lock);
 632
 633         rblk = rrpc_get_blk(rrpc, rlun, gc_force);
 634         if (rblk) {
 635                 spin_lock(&rlun->lock);
 636                 list_add_tail(&rblk->prio, &rlun->wblk_list);
 637                 /*
 638                  * another thread might already have added a new block,
 639                  * Therefore, make sure that one is used, instead of the
 640                  * one just added.
 641                  */
 642                 goto new_blk;
 643         }
 644
 645         if (unlikely(is_gc) && !gc_force) {
 646                 /* retry from emergency gc block */
 647                 cur_rblk = &rlun->gc_cur;
 648                 rblk = rlun->gc_cur;
 649                 gc_force = 1;
 650                 spin_lock(&rlun->lock);
 651                 goto retry;
 652         }
 653
 654         pr_err("rrpc: failed to allocate new block\n");
 655         return NULL;
 656 done:
 657         spin_unlock(&rlun->lock);
 658         return rrpc_update_map(rrpc, laddr, rblk, paddr);
 659 }
 660
 661 static void rrpc_run_gc(struct rrpc *rrpc, struct rrpc_block *rblk)
 662 {
 663         struct rrpc_block_gc *gcb;
 664
 665         gcb = mempool_alloc(rrpc->gcb_pool, GFP_ATOMIC);
 666         if (!gcb) {
 667                 pr_err("rrpc: unable to queue block for gc.");
 668                 return;
 669         }
 670
 671         gcb->rrpc = rrpc;
 672         gcb->rblk = rblk;
 673
 674         INIT_WORK(&gcb->ws_gc, rrpc_gc_queue);
 675         queue_work(rrpc->kgc_wq, &gcb->ws_gc);
 676 }
 677
 678 static void __rrpc_mark_bad_block(struct nvm_dev *dev, struct ppa_addr *ppa)
 679 {
 680                 nvm_mark_blk(dev, *ppa, NVM_BLK_ST_BAD);
 681                 nvm_set_bb_tbl(dev, ppa, 1, NVM_BLK_T_GRWN_BAD);
 682 }
 683
 684 static void rrpc_mark_bad_block(struct rrpc *rrpc, struct nvm_rq *rqd)
 685 {
 686         struct nvm_dev *dev = rrpc->dev;
 687         void *comp_bits = &rqd->ppa_status;
 688         struct ppa_addr ppa, prev_ppa;
 689         int nr_ppas = rqd->nr_ppas;
 690         int bit;
 691
 692         if (rqd->nr_ppas == 1)
 693                 __rrpc_mark_bad_block(dev, &rqd->ppa_addr);
 694
 695         ppa_set_empty(&prev_ppa);
 696         bit = -1;
 697         while ((bit = find_next_bit(comp_bits, nr_ppas, bit + 1)) < nr_ppas) {
 698                 ppa = rqd->ppa_list[bit];
 699                 if (ppa_cmp_blk(ppa, prev_ppa))
 700                         continue;
 701
 702                 __rrpc_mark_bad_block(dev, &ppa);
 703         }
 704 }
 705
 706 static void rrpc_end_io_write(struct rrpc *rrpc, struct rrpc_rq *rrqd,
 707                                                 sector_t laddr, uint8_t npages)
 708 {
 709         struct rrpc_addr *p;
 710         struct rrpc_block *rblk;
 711         struct nvm_lun *lun;
 712         int cmnt_size, i;
 713
 714         for (i = 0; i < npages; i++) {
 715                 p = &rrpc->trans_map[laddr + i];
 716                 rblk = p->rblk;
 717                 lun = rblk->parent->lun;
 718
 719                 cmnt_size = atomic_inc_return(&rblk->data_cmnt_size);
 720                 if (unlikely(cmnt_size == rrpc->dev->sec_per_blk))
 721                         rrpc_run_gc(rrpc, rblk);
 722         }
 723 }
 724
 725 static void rrpc_end_io(struct nvm_rq *rqd)
 726 {
 727         struct rrpc *rrpc = container_of(rqd->ins, struct rrpc, instance);
 728         struct rrpc_rq *rrqd = nvm_rq_to_pdu(rqd);
 729         uint8_t npages = rqd->nr_ppas;
 730         sector_t laddr = rrpc_get_laddr(rqd->bio) - npages;
 731
 732         if (bio_data_dir(rqd->bio) == WRITE) {
 733                 if (rqd->error == NVM_RSP_ERR_FAILWRITE)
 734                         rrpc_mark_bad_block(rrpc, rqd);
 735
 736                 rrpc_end_io_write(rrpc, rrqd, laddr, npages);
 737         }
 738
 739         bio_put(rqd->bio);
 740
 741         if (rrqd->flags & NVM_IOTYPE_GC)
 742                 return;
 743
 744         rrpc_unlock_rq(rrpc, rqd);
 745
 746         if (npages > 1)
 747                 nvm_dev_dma_free(rrpc->dev, rqd->ppa_list, rqd->dma_ppa_list);
 748
 749         mempool_free(rqd, rrpc->rq_pool);
 750 }
 751
 752 static int rrpc_read_ppalist_rq(struct rrpc *rrpc, struct bio *bio,
 753                         struct nvm_rq *rqd, unsigned long flags, int npages)
 754 {
 755         struct rrpc_inflight_rq *r = rrpc_get_inflight_rq(rqd);
 756         struct rrpc_addr *gp;
 757         sector_t laddr = rrpc_get_laddr(bio);
 758         int is_gc = flags & NVM_IOTYPE_GC;
 759         int i;
 760
 761         if (!is_gc && rrpc_lock_rq(rrpc, bio, rqd)) {
 762                 nvm_dev_dma_free(rrpc->dev, rqd->ppa_list, rqd->dma_ppa_list);
 763                 return NVM_IO_REQUEUE;
 764         }
 765
 766         for (i = 0; i < npages; i++) {
 767                 /* We assume that mapping occurs at 4KB granularity */
 768                 BUG_ON(!(laddr + i >= 0 && laddr + i < rrpc->nr_sects));
 769                 gp = &rrpc->trans_map[laddr + i];
 770
 771                 if (gp->rblk) {
 772                         rqd->ppa_list[i] = rrpc_ppa_to_gaddr(rrpc->dev,
 773                                                                 gp->addr);
 774                 } else {
 775                         BUG_ON(is_gc);
 776                         rrpc_unlock_laddr(rrpc, r);
 777                         nvm_dev_dma_free(rrpc->dev, rqd->ppa_list,
 778                                                         rqd->dma_ppa_list);
 779                         return NVM_IO_DONE;
 780                 }
 781         }
 782
 783         rqd->opcode = NVM_OP_HBREAD;
 784
 785         return NVM_IO_OK;
 786 }
 787
 788 static int rrpc_read_rq(struct rrpc *rrpc, struct bio *bio, struct nvm_rq *rqd,
 789                                                         unsigned long flags)
 790 {
 791         struct rrpc_rq *rrqd = nvm_rq_to_pdu(rqd);
 792         int is_gc = flags & NVM_IOTYPE_GC;
 793         sector_t laddr = rrpc_get_laddr(bio);
 794         struct rrpc_addr *gp;
 795
 796         if (!is_gc && rrpc_lock_rq(rrpc, bio, rqd))
 797                 return NVM_IO_REQUEUE;
 798
 799         BUG_ON(!(laddr >= 0 && laddr < rrpc->nr_sects));
 800         gp = &rrpc->trans_map[laddr];
 801
 802         if (gp->rblk) {
 803                 rqd->ppa_addr = rrpc_ppa_to_gaddr(rrpc->dev, gp->addr);
 804         } else {
 805                 BUG_ON(is_gc);
 806                 rrpc_unlock_rq(rrpc, rqd);
 807                 return NVM_IO_DONE;
 808         }
 809
 810         rqd->opcode = NVM_OP_HBREAD;
 811         rrqd->addr = gp;
 812
 813         return NVM_IO_OK;
 814 }
 815
 816 static int rrpc_write_ppalist_rq(struct rrpc *rrpc, struct bio *bio,
 817                         struct nvm_rq *rqd, unsigned long flags, int npages)
 818 {
 819         struct rrpc_inflight_rq *r = rrpc_get_inflight_rq(rqd);
 820         struct rrpc_addr *p;
 821         sector_t laddr = rrpc_get_laddr(bio);
 822         int is_gc = flags & NVM_IOTYPE_GC;
 823         int i;
 824
 825         if (!is_gc && rrpc_lock_rq(rrpc, bio, rqd)) {
 826                 nvm_dev_dma_free(rrpc->dev, rqd->ppa_list, rqd->dma_ppa_list);
 827                 return NVM_IO_REQUEUE;
 828         }
 829
 830         for (i = 0; i < npages; i++) {
 831                 /* We assume that mapping occurs at 4KB granularity */
 832                 p = rrpc_map_page(rrpc, laddr + i, is_gc);
 833                 if (!p) {
 834                         BUG_ON(is_gc);
 835                         rrpc_unlock_laddr(rrpc, r);
 836                         nvm_dev_dma_free(rrpc->dev, rqd->ppa_list,
 837                                                         rqd->dma_ppa_list);
 838                         rrpc_gc_kick(rrpc);
 839                         return NVM_IO_REQUEUE;
 840                 }
 841
 842                 rqd->ppa_list[i] = rrpc_ppa_to_gaddr(rrpc->dev,
 843                                                                 p->addr);
 844         }
 845
 846         rqd->opcode = NVM_OP_HBWRITE;
 847
 848         return NVM_IO_OK;
 849 }
 850
 851 static int rrpc_write_rq(struct rrpc *rrpc, struct bio *bio,
 852                                 struct nvm_rq *rqd, unsigned long flags)
 853 {
 854         struct rrpc_rq *rrqd = nvm_rq_to_pdu(rqd);
 855         struct rrpc_addr *p;
 856         int is_gc = flags & NVM_IOTYPE_GC;
 857         sector_t laddr = rrpc_get_laddr(bio);
 858
 859         if (!is_gc && rrpc_lock_rq(rrpc, bio, rqd))
 860                 return NVM_IO_REQUEUE;
 861
 862         p = rrpc_map_page(rrpc, laddr, is_gc);
 863         if (!p) {
 864                 BUG_ON(is_gc);
 865                 rrpc_unlock_rq(rrpc, rqd);
 866                 rrpc_gc_kick(rrpc);
 867                 return NVM_IO_REQUEUE;
 868         }
 869
 870         rqd->ppa_addr = rrpc_ppa_to_gaddr(rrpc->dev, p->addr);
 871         rqd->opcode = NVM_OP_HBWRITE;
 872         rrqd->addr = p;
 873
 874         return NVM_IO_OK;
 875 }
 876
 877 static int rrpc_setup_rq(struct rrpc *rrpc, struct bio *bio,
 878                         struct nvm_rq *rqd, unsigned long flags, uint8_t npages)
 879 {
 880         if (npages > 1) {
 881                 rqd->ppa_list = nvm_dev_dma_alloc(rrpc->dev, GFP_KERNEL,
 882                                                         &rqd->dma_ppa_list);
 883                 if (!rqd->ppa_list) {
 884                         pr_err("rrpc: not able to allocate ppa list\n");
 885                         return NVM_IO_ERR;
 886                 }
 887
 888                 if (bio_op(bio) == REQ_OP_WRITE)
 889                         return rrpc_write_ppalist_rq(rrpc, bio, rqd, flags,
 890                                                                         npages);
 891
 892                 return rrpc_read_ppalist_rq(rrpc, bio, rqd, flags, npages);
 893         }
 894
 895         if (bio_op(bio) == REQ_OP_WRITE)
 896                 return rrpc_write_rq(rrpc, bio, rqd, flags);
 897
 898         return rrpc_read_rq(rrpc, bio, rqd, flags);
 899 }
 900
 901 static int rrpc_submit_io(struct rrpc *rrpc, struct bio *bio,
 902                                 struct nvm_rq *rqd, unsigned long flags)
 903 {
 904         int err;
 905         struct rrpc_rq *rrq = nvm_rq_to_pdu(rqd);
 906         uint8_t nr_pages = rrpc_get_pages(bio);
 907         int bio_size = bio_sectors(bio) << 9;
 908
 909         if (bio_size < rrpc->dev->sec_size)
 910                 return NVM_IO_ERR;
 911         else if (bio_size > rrpc->dev->max_rq_size)
 912                 return NVM_IO_ERR;
 913
 914         err = rrpc_setup_rq(rrpc, bio, rqd, flags, nr_pages);
 915         if (err)
 916                 return err;
 917
 918         bio_get(bio);
 919         rqd->bio = bio;
 920         rqd->ins = &rrpc->instance;
 921         rqd->nr_ppas = nr_pages;
 922         rrq->flags = flags;
 923
 924         err = nvm_submit_io(rrpc->dev, rqd);
 925         if (err) {
 926                 pr_err("rrpc: I/O submission failed: %d\n", err);
 927                 bio_put(bio);
 928                 if (!(flags & NVM_IOTYPE_GC)) {
 929                         rrpc_unlock_rq(rrpc, rqd);
 930                         if (rqd->nr_ppas > 1)
 931                                 nvm_dev_dma_free(rrpc->dev,
 932                         rqd->ppa_list, rqd->dma_ppa_list);
 933                 }
 934                 return NVM_IO_ERR;
 935         }
 936
 937         return NVM_IO_OK;
 938 }
 939
 940 static blk_qc_t rrpc_make_rq(struct request_queue *q, struct bio *bio)
 941 {
 942         struct rrpc *rrpc = q->queuedata;
 943         struct nvm_rq *rqd;
 944         int err;
 945
 946         if (bio_op(bio) == REQ_OP_DISCARD) {
 947                 rrpc_discard(rrpc, bio);
 948                 return BLK_QC_T_NONE;
 949         }
 950
 951         rqd = mempool_alloc(rrpc->rq_pool, GFP_KERNEL);
 952         if (!rqd) {
 953                 pr_err_ratelimited("rrpc: not able to queue bio.");
 954                 bio_io_error(bio);
 955                 return BLK_QC_T_NONE;
 956         }
 957         memset(rqd, 0, sizeof(struct nvm_rq));
 958
 959         err = rrpc_submit_io(rrpc, bio, rqd, NVM_IOTYPE_NONE);
 960         switch (err) {
 961         case NVM_IO_OK:
 962                 return BLK_QC_T_NONE;
 963         case NVM_IO_ERR:
 964                 bio_io_error(bio);
 965                 break;
 966         case NVM_IO_DONE:
 967                 bio_endio(bio);
 968                 break;
 969         case NVM_IO_REQUEUE:
 970                 spin_lock(&rrpc->bio_lock);
 971                 bio_list_add(&rrpc->requeue_bios, bio);
 972                 spin_unlock(&rrpc->bio_lock);
 973                 queue_work(rrpc->kgc_wq, &rrpc->ws_requeue);
 974                 break;
 975         }
 976
 977         mempool_free(rqd, rrpc->rq_pool);
 978         return BLK_QC_T_NONE;
 979 }
 980
 981 static void rrpc_requeue(struct work_struct *work)
 982 {
 983         struct rrpc *rrpc = container_of(work, struct rrpc, ws_requeue);
 984         struct bio_list bios;
 985         struct bio *bio;
 986
 987         bio_list_init(&bios);
 988
 989         spin_lock(&rrpc->bio_lock);
 990         bio_list_merge(&bios, &rrpc->requeue_bios);
 991         bio_list_init(&rrpc->requeue_bios);
 992         spin_unlock(&rrpc->bio_lock);
 993
 994         while ((bio = bio_list_pop(&bios)))
 995                 rrpc_make_rq(rrpc->disk->queue, bio);
 996 }
 997
 998 static void rrpc_gc_free(struct rrpc *rrpc)
 999 {
1000         if (rrpc->krqd_wq)
1001                 destroy_workqueue(rrpc->krqd_wq);
1002
1003         if (rrpc->kgc_wq)
1004                 destroy_workqueue(rrpc->kgc_wq);
1005 }
1006
1007 static int rrpc_gc_init(struct rrpc *rrpc)
1008 {
1009         rrpc->krqd_wq = alloc_workqueue("rrpc-lun", WQ_MEM_RECLAIM|WQ_UNBOUND,
1010                                                                 rrpc->nr_luns);
1011         if (!rrpc->krqd_wq)
1012                 return -ENOMEM;
1013
1014         rrpc->kgc_wq = alloc_workqueue("rrpc-bg", WQ_MEM_RECLAIM, 1);
1015         if (!rrpc->kgc_wq)
1016                 return -ENOMEM;
1017
1018         setup_timer(&rrpc->gc_timer, rrpc_gc_timer, (unsigned long)rrpc);
1019
1020         return 0;
1021 }
1022
1023 static void rrpc_map_free(struct rrpc *rrpc)
1024 {
1025         vfree(rrpc->rev_trans_map);
1026         vfree(rrpc->trans_map);
1027 }
1028
1029 static int rrpc_l2p_update(u64 slba, u32 nlb, __le64 *entries, void *private)
1030 {
1031         struct rrpc *rrpc = (struct rrpc *)private;
1032         struct nvm_dev *dev = rrpc->dev;
1033         struct rrpc_addr *addr = rrpc->trans_map + slba;
1034         struct rrpc_rev_addr *raddr = rrpc->rev_trans_map;
1035         u64 elba = slba + nlb;
1036         u64 i;
1037
1038         if (unlikely(elba > dev->total_secs)) {
1039                 pr_err("nvm: L2P data from device is out of bounds!\n");
1040                 return -EINVAL;
1041         }
1042
1043         for (i = 0; i < nlb; i++) {
1044                 u64 pba = le64_to_cpu(entries[i]);
1045                 unsigned int mod;
1046                 /* LNVM treats address-spaces as silos, LBA and PBA are
1047                  * equally large and zero-indexed.
1048                  */
1049                 if (unlikely(pba >= dev->total_secs && pba != U64_MAX)) {
1050                         pr_err("nvm: L2P data entry is out of bounds!\n");
1051                         return -EINVAL;
1052                 }
1053
1054                 /* Address zero is a special one. The first page on a disk is
1055                  * protected. As it often holds internal device boot
1056                  * information.
1057                  */
1058                 if (!pba)
1059                         continue;
1060
1061                 div_u64_rem(pba, rrpc->nr_sects, &mod);
1062
1063                 addr[i].addr = pba;
1064                 raddr[mod].addr = slba + i;
1065         }
1066
1067         return 0;
1068 }
1069
1070 static int rrpc_map_init(struct rrpc *rrpc)
1071 {
1072         struct nvm_dev *dev = rrpc->dev;
1073         sector_t i;
1074         int ret;
1075
1076         rrpc->trans_map = vzalloc(sizeof(struct rrpc_addr) * rrpc->nr_sects);
1077         if (!rrpc->trans_map)
1078                 return -ENOMEM;
1079
1080         rrpc->rev_trans_map = vmalloc(sizeof(struct rrpc_rev_addr)
1081                                                         * rrpc->nr_sects);
1082         if (!rrpc->rev_trans_map)
1083                 return -ENOMEM;
1084
1085         for (i = 0; i < rrpc->nr_sects; i++) {
1086                 struct rrpc_addr *p = &rrpc->trans_map[i];
1087                 struct rrpc_rev_addr *r = &rrpc->rev_trans_map[i];
1088
1089                 p->addr = ADDR_EMPTY;
1090                 r->addr = ADDR_EMPTY;
1091         }
1092
1093         if (!dev->ops->get_l2p_tbl)
1094                 return 0;
1095
1096         /* Bring up the mapping table from device */
1097         ret = dev->ops->get_l2p_tbl(dev, rrpc->soffset, rrpc->nr_sects,
1098                                         rrpc_l2p_update, rrpc);
1099         if (ret) {
1100                 pr_err("nvm: rrpc: could not read L2P table.\n");
1101                 return -EINVAL;
1102         }
1103
1104         return 0;
1105 }
1106
1107 /* Minimum pages needed within a lun */
1108 #define PAGE_POOL_SIZE 16
1109 #define ADDR_POOL_SIZE 64
1110
1111 static int rrpc_core_init(struct rrpc *rrpc)
1112 {
1113         down_write(&rrpc_lock);
1114         if (!rrpc_gcb_cache) {
1115                 rrpc_gcb_cache = kmem_cache_create("rrpc_gcb",
1116                                 sizeof(struct rrpc_block_gc), 0, 0, NULL);
1117                 if (!rrpc_gcb_cache) {
1118                         up_write(&rrpc_lock);
1119                         return -ENOMEM;
1120                 }
1121
1122                 rrpc_rq_cache = kmem_cache_create("rrpc_rq",
1123                                 sizeof(struct nvm_rq) + sizeof(struct rrpc_rq),
1124                                 0, 0, NULL);
1125                 if (!rrpc_rq_cache) {
1126                         kmem_cache_destroy(rrpc_gcb_cache);
1127                         up_write(&rrpc_lock);
1128                         return -ENOMEM;
1129                 }
1130         }
1131         up_write(&rrpc_lock);
1132
1133         rrpc->page_pool = mempool_create_page_pool(PAGE_POOL_SIZE, 0);
1134         if (!rrpc->page_pool)
1135                 return -ENOMEM;
1136
1137         rrpc->gcb_pool = mempool_create_slab_pool(rrpc->dev->nr_luns,
1138                                                                 rrpc_gcb_cache);
1139         if (!rrpc->gcb_pool)
1140                 return -ENOMEM;
1141
1142         rrpc->rq_pool = mempool_create_slab_pool(64, rrpc_rq_cache);
1143         if (!rrpc->rq_pool)
1144                 return -ENOMEM;
1145
1146         spin_lock_init(&rrpc->inflights.lock);
1147         INIT_LIST_HEAD(&rrpc->inflights.reqs);
1148
1149         return 0;
1150 }
1151
1152 static void rrpc_core_free(struct rrpc *rrpc)
1153 {
1154         mempool_destroy(rrpc->page_pool);
1155         mempool_destroy(rrpc->gcb_pool);
1156         mempool_destroy(rrpc->rq_pool);
1157 }
1158
1159 static void rrpc_luns_free(struct rrpc *rrpc)
1160 {
1161         struct nvm_dev *dev = rrpc->dev;
1162         struct nvm_lun *lun;
1163         struct rrpc_lun *rlun;
1164         int i;
1165
1166         if (!rrpc->luns)
1167                 return;
1168
1169         for (i = 0; i < rrpc->nr_luns; i++) {
1170                 rlun = &rrpc->luns[i];
1171                 lun = rlun->parent;
1172                 if (!lun)
1173                         break;
1174                 dev->mt->release_lun(dev, lun->id);
1175                 vfree(rlun->blocks);
1176         }
1177
1178         kfree(rrpc->luns);
1179 }
1180
1181 static int rrpc_luns_init(struct rrpc *rrpc, int lun_begin, int lun_end)
1182 {
1183         struct nvm_dev *dev = rrpc->dev;
1184         struct rrpc_lun *rlun;
1185         int i, j, ret = -EINVAL;
1186
1187         if (dev->sec_per_blk > MAX_INVALID_PAGES_STORAGE * BITS_PER_LONG) {
1188                 pr_err("rrpc: number of pages per block too high.");
1189                 return -EINVAL;
1190         }
1191
1192         spin_lock_init(&rrpc->rev_lock);
1193
1194         rrpc->luns = kcalloc(rrpc->nr_luns, sizeof(struct rrpc_lun),
1195                                                                 GFP_KERNEL);
1196         if (!rrpc->luns)
1197                 return -ENOMEM;
1198
1199         /* 1:1 mapping */
1200         for (i = 0; i < rrpc->nr_luns; i++) {
1201                 int lunid = lun_begin + i;
1202                 struct nvm_lun *lun;
1203
1204                 if (dev->mt->reserve_lun(dev, lunid)) {
1205                         pr_err("rrpc: lun %u is already allocated\n", lunid);
1206                         goto err;
1207                 }
1208
1209                 lun = dev->mt->get_lun(dev, lunid);
1210                 if (!lun)
1211                         goto err;
1212
1213                 rlun = &rrpc->luns[i];
1214                 rlun->parent = lun;
1215                 rlun->blocks = vzalloc(sizeof(struct rrpc_block) *
1216                                                 rrpc->dev->blks_per_lun);
1217                 if (!rlun->blocks) {
1218                         ret = -ENOMEM;
1219                         goto err;
1220                 }
1221
1222                 for (j = 0; j < rrpc->dev->blks_per_lun; j++) {
1223                         struct rrpc_block *rblk = &rlun->blocks[j];
1224                         struct nvm_block *blk = &lun->blocks[j];
1225
1226                         rblk->parent = blk;
1227                         rblk->rlun = rlun;
1228                         INIT_LIST_HEAD(&rblk->prio);
1229                         spin_lock_init(&rblk->lock);
1230                 }
1231
1232                 rlun->rrpc = rrpc;
1233                 INIT_LIST_HEAD(&rlun->prio_list);
1234                 INIT_LIST_HEAD(&rlun->wblk_list);
1235
1236                 INIT_WORK(&rlun->ws_gc, rrpc_lun_gc);
1237                 spin_lock_init(&rlun->lock);
1238         }
1239
1240         return 0;
1241 err:
1242         return ret;
1243 }
1244
1245 /* returns 0 on success and stores the beginning address in *begin */
1246 static int rrpc_area_init(struct rrpc *rrpc, sector_t *begin)
1247 {
1248         struct nvm_dev *dev = rrpc->dev;
1249         struct nvmm_type *mt = dev->mt;
1250         sector_t size = rrpc->nr_sects * dev->sec_size;
1251         int ret;
1252
1253         size >>= 9;
1254
1255         ret = mt->get_area(dev, begin, size);
1256         if (!ret)
1257                 *begin >>= (ilog2(dev->sec_size) - 9);
1258
1259         return ret;
1260 }
1261
1262 static void rrpc_area_free(struct rrpc *rrpc)
1263 {
1264         struct nvm_dev *dev = rrpc->dev;
1265         struct nvmm_type *mt = dev->mt;
1266         sector_t begin = rrpc->soffset << (ilog2(dev->sec_size) - 9);
1267
1268         mt->put_area(dev, begin);
1269 }
1270
1271 static void rrpc_free(struct rrpc *rrpc)
1272 {
1273         rrpc_gc_free(rrpc);
1274         rrpc_map_free(rrpc);
1275         rrpc_core_free(rrpc);
1276         rrpc_luns_free(rrpc);
1277         rrpc_area_free(rrpc);
1278
1279         kfree(rrpc);
1280 }
1281
1282 static void rrpc_exit(void *private)
1283 {
1284         struct rrpc *rrpc = private;
1285
1286         del_timer(&rrpc->gc_timer);
1287
1288         flush_workqueue(rrpc->krqd_wq);
1289         flush_workqueue(rrpc->kgc_wq);
1290
1291         rrpc_free(rrpc);
1292 }
1293
1294 static sector_t rrpc_capacity(void *private)
1295 {
1296         struct rrpc *rrpc = private;
1297         struct nvm_dev *dev = rrpc->dev;
1298         sector_t reserved, provisioned;
1299
1300         /* cur, gc, and two emergency blocks for each lun */
1301         reserved = rrpc->nr_luns * dev->sec_per_blk * 4;
1302         provisioned = rrpc->nr_sects - reserved;
1303
1304         if (reserved > rrpc->nr_sects) {
1305                 pr_err("rrpc: not enough space available to expose storage.\n");
1306                 return 0;
1307         }
1308
1309         sector_div(provisioned, 10);
1310         return provisioned * 9 * NR_PHY_IN_LOG;
1311 }
1312
1313 /*
1314  * Looks up the logical address from reverse trans map and check if its valid by
1315  * comparing the logical to physical address with the physical address.
1316  * Returns 0 on free, otherwise 1 if in use
1317  */
1318 static void rrpc_block_map_update(struct rrpc *rrpc, struct rrpc_block *rblk)
1319 {
1320         struct nvm_dev *dev = rrpc->dev;
1321         int offset;
1322         struct rrpc_addr *laddr;
1323         u64 bpaddr, paddr, pladdr;
1324
1325         bpaddr = block_to_rel_addr(rrpc, rblk);
1326         for (offset = 0; offset < dev->sec_per_blk; offset++) {
1327                 paddr = bpaddr + offset;
1328
1329                 pladdr = rrpc->rev_trans_map[paddr].addr;
1330                 if (pladdr == ADDR_EMPTY)
1331                         continue;
1332
1333                 laddr = &rrpc->trans_map[pladdr];
1334
1335                 if (paddr == laddr->addr) {
1336                         laddr->rblk = rblk;
1337                 } else {
1338                         set_bit(offset, rblk->invalid_pages);
1339                         rblk->nr_invalid_pages++;
1340                 }
1341         }
1342 }
1343
1344 static int rrpc_blocks_init(struct rrpc *rrpc)
1345 {
1346         struct rrpc_lun *rlun;
1347         struct rrpc_block *rblk;
1348         int lun_iter, blk_iter;
1349
1350         for (lun_iter = 0; lun_iter < rrpc->nr_luns; lun_iter++) {
1351                 rlun = &rrpc->luns[lun_iter];
1352
1353                 for (blk_iter = 0; blk_iter < rrpc->dev->blks_per_lun;
1354                                                                 blk_iter++) {
1355                         rblk = &rlun->blocks[blk_iter];
1356                         rrpc_block_map_update(rrpc, rblk);
1357                 }
1358         }
1359
1360         return 0;
1361 }
1362
1363 static int rrpc_luns_configure(struct rrpc *rrpc)
1364 {
1365         struct rrpc_lun *rlun;
1366         struct rrpc_block *rblk;
1367         int i;
1368
1369         for (i = 0; i < rrpc->nr_luns; i++) {
1370                 rlun = &rrpc->luns[i];
1371
1372                 rblk = rrpc_get_blk(rrpc, rlun, 0);
1373                 if (!rblk)
1374                         goto err;
1375                 rrpc_set_lun_cur(rlun, rblk, &rlun->cur);
1376
1377                 /* Emergency gc block */
1378                 rblk = rrpc_get_blk(rrpc, rlun, 1);
1379                 if (!rblk)
1380                         goto err;
1381                 rrpc_set_lun_cur(rlun, rblk, &rlun->gc_cur);
1382         }
1383
1384         return 0;
1385 err:
1386         rrpc_put_blks(rrpc);
1387         return -EINVAL;
1388 }
1389
1390 static struct nvm_tgt_type tt_rrpc;
1391
1392 static void *rrpc_init(struct nvm_dev *dev, struct gendisk *tdisk,
1393                                                 int lun_begin, int lun_end)
1394 {
1395         struct request_queue *bqueue = dev->q;
1396         struct request_queue *tqueue = tdisk->queue;
1397         struct rrpc *rrpc;
1398         sector_t soffset;
1399         int ret;
1400
1401         if (!(dev->identity.dom & NVM_RSP_L2P)) {
1402                 pr_err("nvm: rrpc: device does not support l2p (%x)\n",
1403                                                         dev->identity.dom);
1404                 return ERR_PTR(-EINVAL);
1405         }
1406
1407         rrpc = kzalloc(sizeof(struct rrpc), GFP_KERNEL);
1408         if (!rrpc)
1409                 return ERR_PTR(-ENOMEM);
1410
1411         rrpc->instance.tt = &tt_rrpc;
1412         rrpc->dev = dev;
1413         rrpc->disk = tdisk;
1414
1415         bio_list_init(&rrpc->requeue_bios);
1416         spin_lock_init(&rrpc->bio_lock);
1417         INIT_WORK(&rrpc->ws_requeue, rrpc_requeue);
1418
1419         rrpc->nr_luns = lun_end - lun_begin + 1;
1420         rrpc->total_blocks = (unsigned long)dev->blks_per_lun * rrpc->nr_luns;
1421         rrpc->nr_sects = (unsigned long long)dev->sec_per_lun * rrpc->nr_luns;
1422
1423         /* simple round-robin strategy */
1424         atomic_set(&rrpc->next_lun, -1);
1425
1426         ret = rrpc_area_init(rrpc, &soffset);
1427         if (ret < 0) {
1428                 pr_err("nvm: rrpc: could not initialize area\n");
1429                 return ERR_PTR(ret);
1430         }
1431         rrpc->soffset = soffset;
1432
1433         ret = rrpc_luns_init(rrpc, lun_begin, lun_end);
1434         if (ret) {
1435                 pr_err("nvm: rrpc: could not initialize luns\n");
1436                 goto err;
1437         }
1438
1439         rrpc->poffset = dev->sec_per_lun * lun_begin;
1440         rrpc->lun_offset = lun_begin;
1441
1442         ret = rrpc_core_init(rrpc);
1443         if (ret) {
1444                 pr_err("nvm: rrpc: could not initialize core\n");
1445                 goto err;
1446         }
1447
1448         ret = rrpc_map_init(rrpc);
1449         if (ret) {
1450                 pr_err("nvm: rrpc: could not initialize maps\n");
1451                 goto err;
1452         }
1453
1454         ret = rrpc_blocks_init(rrpc);
1455         if (ret) {
1456                 pr_err("nvm: rrpc: could not initialize state for blocks\n");
1457                 goto err;
1458         }
1459
1460         ret = rrpc_luns_configure(rrpc);
1461         if (ret) {
1462                 pr_err("nvm: rrpc: not enough blocks available in LUNs.\n");
1463                 goto err;
1464         }
1465
1466         ret = rrpc_gc_init(rrpc);
1467         if (ret) {
1468                 pr_err("nvm: rrpc: could not initialize gc\n");
1469                 goto err;
1470         }
1471
1472         /* inherit the size from the underlying device */
1473         blk_queue_logical_block_size(tqueue, queue_physical_block_size(bqueue));
1474         blk_queue_max_hw_sectors(tqueue, queue_max_hw_sectors(bqueue));
1475
1476         pr_info("nvm: rrpc initialized with %u luns and %llu pages.\n",
1477                         rrpc->nr_luns, (unsigned long long)rrpc->nr_sects);
1478
1479         mod_timer(&rrpc->gc_timer, jiffies + msecs_to_jiffies(10));
1480
1481         return rrpc;
1482 err:
1483         rrpc_free(rrpc);
1484         return ERR_PTR(ret);
1485 }
1486
1487 /* round robin, page-based FTL, and cost-based GC */
1488 static struct nvm_tgt_type tt_rrpc = {
1489         .name           = "rrpc",
1490         .version        = {1, 0, 0},
1491
1492         .make_rq        = rrpc_make_rq,
1493         .capacity       = rrpc_capacity,
1494         .end_io         = rrpc_end_io,
1495
1496         .init           = rrpc_init,
1497         .exit           = rrpc_exit,
1498 };
1499
1500 static int __init rrpc_module_init(void)
1501 {
1502         return nvm_register_tgt_type(&tt_rrpc);
1503 }
1504
1505 static void rrpc_module_exit(void)
1506 {
1507         nvm_unregister_tgt_type(&tt_rrpc);
1508 }
1509
1510 module_init(rrpc_module_init);
1511 module_exit(rrpc_module_exit);
1512 MODULE_LICENSE("GPL v2");
1513 MODULE_DESCRIPTION("Block-Device Target for Open-Channel SSDs");