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