2 * Copyright (C) 2015 IT University of Copenhagen
3 * Initial release: Matias Bjorling <m@bjorling.me>
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.
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.
14 * Implementation of a Round-robin page-based Hybrid FTL for Open-channel SSDs.
19 static struct kmem_cache *rrpc_gcb_cache, *rrpc_rq_cache;
20 static DECLARE_RWSEM(rrpc_lock);
22 static int rrpc_submit_io(struct rrpc *rrpc, struct bio *bio,
23 struct nvm_rq *rqd, unsigned long flags);
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)])
29 static void rrpc_page_invalidate(struct rrpc *rrpc, struct rrpc_addr *a)
31 struct rrpc_block *rblk = a->rblk;
32 unsigned int pg_offset;
34 lockdep_assert_held(&rrpc->rev_lock);
36 if (a->addr == ADDR_EMPTY || !rblk)
39 spin_lock(&rblk->lock);
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++;
45 spin_unlock(&rblk->lock);
47 rrpc->rev_trans_map[a->addr - rrpc->poffset].addr = ADDR_EMPTY;
50 static void rrpc_invalidate_range(struct rrpc *rrpc, sector_t slba,
55 spin_lock(&rrpc->rev_lock);
56 for (i = slba; i < slba + len; i++) {
57 struct rrpc_addr *gp = &rrpc->trans_map[i];
59 rrpc_page_invalidate(rrpc, gp);
62 spin_unlock(&rrpc->rev_lock);
65 static struct nvm_rq *rrpc_inflight_laddr_acquire(struct rrpc *rrpc,
66 sector_t laddr, unsigned int pages)
69 struct rrpc_inflight_rq *inf;
71 rqd = mempool_alloc(rrpc->rq_pool, GFP_ATOMIC);
73 return ERR_PTR(-ENOMEM);
75 inf = rrpc_get_inflight_rq(rqd);
76 if (rrpc_lock_laddr(rrpc, laddr, pages, inf)) {
77 mempool_free(rqd, rrpc->rq_pool);
84 static void rrpc_inflight_laddr_release(struct rrpc *rrpc, struct nvm_rq *rqd)
86 struct rrpc_inflight_rq *inf = rrpc_get_inflight_rq(rqd);
88 rrpc_unlock_laddr(rrpc, inf);
90 mempool_free(rqd, rrpc->rq_pool);
93 static void rrpc_discard(struct rrpc *rrpc, struct bio *bio)
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;
100 rqd = rrpc_inflight_laddr_acquire(rrpc, slba, len);
108 pr_err("rrpc: unable to acquire inflight IO\n");
113 rrpc_invalidate_range(rrpc, slba, len);
114 rrpc_inflight_laddr_release(rrpc, rqd);
117 static int block_is_full(struct rrpc *rrpc, struct rrpc_block *rblk)
119 return (rblk->next_page == rrpc->dev->sec_per_blk);
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)
125 struct nvm_block *blk = rblk->parent;
126 int lun_blk = blk->id % (rrpc->dev->blks_per_lun * rrpc->nr_luns);
128 return lun_blk * rrpc->dev->sec_per_blk;
131 /* Calculate global addr for the given block */
132 static u64 block_to_addr(struct rrpc *rrpc, struct rrpc_block *rblk)
134 struct nvm_block *blk = rblk->parent;
136 return blk->id * rrpc->dev->sec_per_blk;
139 static struct ppa_addr linear_to_generic_addr(struct nvm_dev *dev,
143 int secs, pgs, blks, luns;
144 sector_t ppa = r.ppa;
148 div_u64_rem(ppa, dev->sec_per_pg, &secs);
151 sector_div(ppa, dev->sec_per_pg);
152 div_u64_rem(ppa, dev->pgs_per_blk, &pgs);
155 sector_div(ppa, dev->pgs_per_blk);
156 div_u64_rem(ppa, dev->blks_per_lun, &blks);
159 sector_div(ppa, dev->blks_per_lun);
160 div_u64_rem(ppa, dev->luns_per_chnl, &luns);
163 sector_div(ppa, dev->luns_per_chnl);
169 static struct ppa_addr rrpc_ppa_to_gaddr(struct nvm_dev *dev, u64 addr)
171 struct ppa_addr paddr;
174 return linear_to_generic_addr(dev, paddr);
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)
181 struct rrpc *rrpc = rlun->rrpc;
184 spin_lock(&(*cur_rblk)->lock);
185 WARN_ON(!block_is_full(rrpc, *cur_rblk));
186 spin_unlock(&(*cur_rblk)->lock);
188 *cur_rblk = new_rblk;
191 static struct rrpc_block *rrpc_get_blk(struct rrpc *rrpc, struct rrpc_lun *rlun,
194 struct nvm_block *blk;
195 struct rrpc_block *rblk;
197 blk = nvm_get_blk(rrpc->dev, rlun->parent, flags);
199 pr_err("nvm: rrpc: cannot get new block from media manager\n");
203 rblk = rrpc_get_rblk(rlun, blk->id);
205 bitmap_zero(rblk->invalid_pages, rrpc->dev->sec_per_blk);
207 rblk->nr_invalid_pages = 0;
208 atomic_set(&rblk->data_cmnt_size, 0);
213 static void rrpc_put_blk(struct rrpc *rrpc, struct rrpc_block *rblk)
215 nvm_put_blk(rrpc->dev, rblk->parent);
218 static void rrpc_put_blks(struct rrpc *rrpc)
220 struct rrpc_lun *rlun;
223 for (i = 0; i < rrpc->nr_luns; i++) {
224 rlun = &rrpc->luns[i];
226 rrpc_put_blk(rrpc, rlun->cur);
228 rrpc_put_blk(rrpc, rlun->gc_cur);
232 static struct rrpc_lun *get_next_lun(struct rrpc *rrpc)
234 int next = atomic_inc_return(&rrpc->next_lun);
236 return &rrpc->luns[next % rrpc->nr_luns];
239 static void rrpc_gc_kick(struct rrpc *rrpc)
241 struct rrpc_lun *rlun;
244 for (i = 0; i < rrpc->nr_luns; i++) {
245 rlun = &rrpc->luns[i];
246 queue_work(rrpc->krqd_wq, &rlun->ws_gc);
251 * timed GC every interval.
253 static void rrpc_gc_timer(unsigned long data)
255 struct rrpc *rrpc = (struct rrpc *)data;
258 mod_timer(&rrpc->gc_timer, jiffies + msecs_to_jiffies(10));
261 static void rrpc_end_sync_bio(struct bio *bio)
263 struct completion *waiting = bio->bi_private;
266 pr_err("nvm: gc request failed (%u).\n", bio->bi_error);
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
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.
281 static int rrpc_move_valid_pages(struct rrpc *rrpc, struct rrpc_block *rblk)
283 struct request_queue *q = rrpc->dev->q;
284 struct rrpc_rev_addr *rev;
289 int nr_sec_per_blk = rrpc->dev->sec_per_blk;
291 DECLARE_COMPLETION_ONSTACK(wait);
293 if (bitmap_full(rblk->invalid_pages, nr_sec_per_blk))
296 bio = bio_alloc(GFP_NOIO, 1);
298 pr_err("nvm: could not alloc bio to gc\n");
302 page = mempool_alloc(rrpc->page_pool, GFP_NOIO);
308 while ((slot = find_first_zero_bit(rblk->invalid_pages,
309 nr_sec_per_blk)) < nr_sec_per_blk) {
312 phys_addr = rblk->parent->id * nr_sec_per_blk + slot;
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);
324 rqd = rrpc_inflight_laddr_acquire(rrpc, rev->addr, 1);
325 if (IS_ERR_OR_NULL(rqd)) {
326 spin_unlock(&rrpc->rev_lock);
331 spin_unlock(&rrpc->rev_lock);
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;
339 /* TODO: may fail when EXP_PG_SIZE > PAGE_SIZE */
340 bio_add_pc_page(q, bio, page, RRPC_EXPOSED_PAGE_SIZE, 0);
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);
347 wait_for_completion_io(&wait);
349 rrpc_inflight_laddr_release(rrpc, rqd);
354 reinit_completion(&wait);
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;
361 bio_add_pc_page(q, bio, page, RRPC_EXPOSED_PAGE_SIZE, 0);
363 /* turn the command around and write the data back to a new
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);
371 wait_for_completion_io(&wait);
373 rrpc_inflight_laddr_release(rrpc, rqd);
381 mempool_free(page, rrpc->page_pool);
384 if (!bitmap_full(rblk->invalid_pages, nr_sec_per_blk)) {
385 pr_err("nvm: failed to garbage collect block\n");
392 static void rrpc_block_gc(struct work_struct *work)
394 struct rrpc_block_gc *gcb = container_of(work, struct rrpc_block_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;
401 mempool_free(gcb, rrpc->gcb_pool);
402 pr_debug("nvm: block '%lu' being reclaimed\n", rblk->parent->id);
404 if (rrpc_move_valid_pages(rrpc, rblk))
407 if (nvm_erase_blk(dev, rblk->parent, 0))
410 rrpc_put_blk(rrpc, rblk);
415 spin_lock(&rlun->lock);
416 list_add_tail(&rblk->prio, &rlun->prio_list);
417 spin_unlock(&rlun->lock);
420 /* the block with highest number of invalid pages, will be in the beginning
423 static struct rrpc_block *rblock_max_invalid(struct rrpc_block *ra,
424 struct rrpc_block *rb)
426 if (ra->nr_invalid_pages == rb->nr_invalid_pages)
429 return (ra->nr_invalid_pages < rb->nr_invalid_pages) ? rb : ra;
432 /* linearly find the block with highest number of invalid pages
435 static struct rrpc_block *block_prio_find_max(struct rrpc_lun *rlun)
437 struct list_head *prio_list = &rlun->prio_list;
438 struct rrpc_block *rblock, *max;
440 BUG_ON(list_empty(prio_list));
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);
449 static void rrpc_lun_gc(struct work_struct *work)
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;
457 nr_blocks_need = rrpc->dev->blks_per_lun / GC_LIMIT_INVERSE;
459 if (nr_blocks_need < rrpc->nr_luns)
460 nr_blocks_need = rrpc->nr_luns;
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;
468 if (!rblock->nr_invalid_pages)
471 gcb = mempool_alloc(rrpc->gcb_pool, GFP_ATOMIC);
475 list_del_init(&rblock->prio);
477 BUG_ON(!block_is_full(rrpc, rblock));
479 pr_debug("rrpc: selected block '%lu' for GC\n", block->id);
483 INIT_WORK(&gcb->ws_gc, rrpc_block_gc);
485 queue_work(rrpc->kgc_wq, &gcb->ws_gc);
489 spin_unlock(&rlun->lock);
491 /* TODO: Hint that request queue can be started again */
494 static void rrpc_gc_queue(struct work_struct *work)
496 struct rrpc_block_gc *gcb = container_of(work, struct rrpc_block_gc,
498 struct rrpc *rrpc = gcb->rrpc;
499 struct rrpc_block *rblk = gcb->rblk;
500 struct rrpc_lun *rlun = rblk->rlun;
502 spin_lock(&rlun->lock);
503 list_add_tail(&rblk->prio, &rlun->prio_list);
504 spin_unlock(&rlun->lock);
506 mempool_free(gcb, rrpc->gcb_pool);
507 pr_debug("nvm: block '%lu' is full, allow GC (sched)\n",
511 static const struct block_device_operations rrpc_fops = {
512 .owner = THIS_MODULE,
515 static struct rrpc_lun *rrpc_get_lun_rr(struct rrpc *rrpc, int is_gc)
518 struct rrpc_lun *rlun, *max_free;
521 return get_next_lun(rrpc);
523 /* during GC, we don't care about RR, instead we want to make
524 * sure that we maintain evenness between the block luns.
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
531 rrpc_for_each_lun(rrpc, rlun, i) {
532 if (rlun->parent->nr_free_blocks >
533 max_free->parent->nr_free_blocks)
540 static struct rrpc_addr *rrpc_update_map(struct rrpc *rrpc, sector_t laddr,
541 struct rrpc_block *rblk, u64 paddr)
543 struct rrpc_addr *gp;
544 struct rrpc_rev_addr *rev;
546 BUG_ON(laddr >= rrpc->nr_sects);
548 gp = &rrpc->trans_map[laddr];
549 spin_lock(&rrpc->rev_lock);
551 rrpc_page_invalidate(rrpc, gp);
556 rev = &rrpc->rev_trans_map[gp->addr - rrpc->poffset];
558 spin_unlock(&rrpc->rev_lock);
563 static u64 rrpc_alloc_addr(struct rrpc *rrpc, struct rrpc_block *rblk)
565 u64 addr = ADDR_EMPTY;
567 spin_lock(&rblk->lock);
568 if (block_is_full(rrpc, rblk))
571 addr = block_to_addr(rrpc, rblk) + rblk->next_page;
575 spin_unlock(&rblk->lock);
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.
582 * Returns rrpc_addr with the physical address and block. Returns NULL if no
583 * blocks in the next rlun are available.
585 static struct rrpc_addr *rrpc_map_page(struct rrpc *rrpc, sector_t laddr,
588 struct rrpc_lun *rlun;
589 struct rrpc_block *rblk, **cur_rblk;
594 rlun = rrpc_get_lun_rr(rrpc, is_gc);
597 if (!is_gc && lun->nr_free_blocks < rrpc->nr_luns * 4)
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
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.
614 spin_lock(&rlun->lock);
615 cur_rblk = &rlun->cur;
618 paddr = rrpc_alloc_addr(rrpc, rblk);
620 if (paddr != ADDR_EMPTY)
623 if (!list_empty(&rlun->wblk_list)) {
625 rblk = list_first_entry(&rlun->wblk_list, struct rrpc_block,
627 rrpc_set_lun_cur(rlun, rblk, cur_rblk);
628 list_del(&rblk->prio);
631 spin_unlock(&rlun->lock);
633 rblk = rrpc_get_blk(rrpc, rlun, gc_force);
635 spin_lock(&rlun->lock);
636 list_add_tail(&rblk->prio, &rlun->wblk_list);
638 * another thread might already have added a new block,
639 * Therefore, make sure that one is used, instead of the
645 if (unlikely(is_gc) && !gc_force) {
646 /* retry from emergency gc block */
647 cur_rblk = &rlun->gc_cur;
650 spin_lock(&rlun->lock);
654 pr_err("rrpc: failed to allocate new block\n");
657 spin_unlock(&rlun->lock);
658 return rrpc_update_map(rrpc, laddr, rblk, paddr);
661 static void rrpc_run_gc(struct rrpc *rrpc, struct rrpc_block *rblk)
663 struct rrpc_block_gc *gcb;
665 gcb = mempool_alloc(rrpc->gcb_pool, GFP_ATOMIC);
667 pr_err("rrpc: unable to queue block for gc.");
674 INIT_WORK(&gcb->ws_gc, rrpc_gc_queue);
675 queue_work(rrpc->kgc_wq, &gcb->ws_gc);
678 static void __rrpc_mark_bad_block(struct nvm_dev *dev, struct ppa_addr *ppa)
680 nvm_mark_blk(dev, *ppa, NVM_BLK_ST_BAD);
681 nvm_set_bb_tbl(dev, ppa, 1, NVM_BLK_T_GRWN_BAD);
684 static void rrpc_mark_bad_block(struct rrpc *rrpc, struct nvm_rq *rqd)
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;
692 if (rqd->nr_ppas == 1)
693 __rrpc_mark_bad_block(dev, &rqd->ppa_addr);
695 ppa_set_empty(&prev_ppa);
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))
702 __rrpc_mark_bad_block(dev, &ppa);
706 static void rrpc_end_io_write(struct rrpc *rrpc, struct rrpc_rq *rrqd,
707 sector_t laddr, uint8_t npages)
710 struct rrpc_block *rblk;
714 for (i = 0; i < npages; i++) {
715 p = &rrpc->trans_map[laddr + i];
717 lun = rblk->parent->lun;
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);
725 static void rrpc_end_io(struct nvm_rq *rqd)
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;
732 if (bio_data_dir(rqd->bio) == WRITE) {
733 if (rqd->error == NVM_RSP_ERR_FAILWRITE)
734 rrpc_mark_bad_block(rrpc, rqd);
736 rrpc_end_io_write(rrpc, rrqd, laddr, npages);
741 if (rrqd->flags & NVM_IOTYPE_GC)
744 rrpc_unlock_rq(rrpc, rqd);
747 nvm_dev_dma_free(rrpc->dev, rqd->ppa_list, rqd->dma_ppa_list);
749 mempool_free(rqd, rrpc->rq_pool);
752 static int rrpc_read_ppalist_rq(struct rrpc *rrpc, struct bio *bio,
753 struct nvm_rq *rqd, unsigned long flags, int npages)
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;
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;
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];
772 rqd->ppa_list[i] = rrpc_ppa_to_gaddr(rrpc->dev,
776 rrpc_unlock_laddr(rrpc, r);
777 nvm_dev_dma_free(rrpc->dev, rqd->ppa_list,
783 rqd->opcode = NVM_OP_HBREAD;
788 static int rrpc_read_rq(struct rrpc *rrpc, struct bio *bio, struct nvm_rq *rqd,
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;
796 if (!is_gc && rrpc_lock_rq(rrpc, bio, rqd))
797 return NVM_IO_REQUEUE;
799 BUG_ON(!(laddr >= 0 && laddr < rrpc->nr_sects));
800 gp = &rrpc->trans_map[laddr];
803 rqd->ppa_addr = rrpc_ppa_to_gaddr(rrpc->dev, gp->addr);
806 rrpc_unlock_rq(rrpc, rqd);
810 rqd->opcode = NVM_OP_HBREAD;
816 static int rrpc_write_ppalist_rq(struct rrpc *rrpc, struct bio *bio,
817 struct nvm_rq *rqd, unsigned long flags, int npages)
819 struct rrpc_inflight_rq *r = rrpc_get_inflight_rq(rqd);
821 sector_t laddr = rrpc_get_laddr(bio);
822 int is_gc = flags & NVM_IOTYPE_GC;
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;
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);
835 rrpc_unlock_laddr(rrpc, r);
836 nvm_dev_dma_free(rrpc->dev, rqd->ppa_list,
839 return NVM_IO_REQUEUE;
842 rqd->ppa_list[i] = rrpc_ppa_to_gaddr(rrpc->dev,
846 rqd->opcode = NVM_OP_HBWRITE;
851 static int rrpc_write_rq(struct rrpc *rrpc, struct bio *bio,
852 struct nvm_rq *rqd, unsigned long flags)
854 struct rrpc_rq *rrqd = nvm_rq_to_pdu(rqd);
856 int is_gc = flags & NVM_IOTYPE_GC;
857 sector_t laddr = rrpc_get_laddr(bio);
859 if (!is_gc && rrpc_lock_rq(rrpc, bio, rqd))
860 return NVM_IO_REQUEUE;
862 p = rrpc_map_page(rrpc, laddr, is_gc);
865 rrpc_unlock_rq(rrpc, rqd);
867 return NVM_IO_REQUEUE;
870 rqd->ppa_addr = rrpc_ppa_to_gaddr(rrpc->dev, p->addr);
871 rqd->opcode = NVM_OP_HBWRITE;
877 static int rrpc_setup_rq(struct rrpc *rrpc, struct bio *bio,
878 struct nvm_rq *rqd, unsigned long flags, uint8_t npages)
881 rqd->ppa_list = nvm_dev_dma_alloc(rrpc->dev, GFP_KERNEL,
883 if (!rqd->ppa_list) {
884 pr_err("rrpc: not able to allocate ppa list\n");
888 if (bio_op(bio) == REQ_OP_WRITE)
889 return rrpc_write_ppalist_rq(rrpc, bio, rqd, flags,
892 return rrpc_read_ppalist_rq(rrpc, bio, rqd, flags, npages);
895 if (bio_op(bio) == REQ_OP_WRITE)
896 return rrpc_write_rq(rrpc, bio, rqd, flags);
898 return rrpc_read_rq(rrpc, bio, rqd, flags);
901 static int rrpc_submit_io(struct rrpc *rrpc, struct bio *bio,
902 struct nvm_rq *rqd, unsigned long flags)
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;
909 if (bio_size < rrpc->dev->sec_size)
911 else if (bio_size > rrpc->dev->max_rq_size)
914 err = rrpc_setup_rq(rrpc, bio, rqd, flags, nr_pages);
920 rqd->ins = &rrpc->instance;
921 rqd->nr_ppas = nr_pages;
924 err = nvm_submit_io(rrpc->dev, rqd);
926 pr_err("rrpc: I/O submission failed: %d\n", err);
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);
940 static blk_qc_t rrpc_make_rq(struct request_queue *q, struct bio *bio)
942 struct rrpc *rrpc = q->queuedata;
946 blk_queue_split(q, &bio, q->bio_split);
948 if (bio_op(bio) == REQ_OP_DISCARD) {
949 rrpc_discard(rrpc, bio);
950 return BLK_QC_T_NONE;
953 rqd = mempool_alloc(rrpc->rq_pool, GFP_KERNEL);
955 pr_err_ratelimited("rrpc: not able to queue bio.");
957 return BLK_QC_T_NONE;
959 memset(rqd, 0, sizeof(struct nvm_rq));
961 err = rrpc_submit_io(rrpc, bio, rqd, NVM_IOTYPE_NONE);
964 return BLK_QC_T_NONE;
972 spin_lock(&rrpc->bio_lock);
973 bio_list_add(&rrpc->requeue_bios, bio);
974 spin_unlock(&rrpc->bio_lock);
975 queue_work(rrpc->kgc_wq, &rrpc->ws_requeue);
979 mempool_free(rqd, rrpc->rq_pool);
980 return BLK_QC_T_NONE;
983 static void rrpc_requeue(struct work_struct *work)
985 struct rrpc *rrpc = container_of(work, struct rrpc, ws_requeue);
986 struct bio_list bios;
989 bio_list_init(&bios);
991 spin_lock(&rrpc->bio_lock);
992 bio_list_merge(&bios, &rrpc->requeue_bios);
993 bio_list_init(&rrpc->requeue_bios);
994 spin_unlock(&rrpc->bio_lock);
996 while ((bio = bio_list_pop(&bios)))
997 rrpc_make_rq(rrpc->disk->queue, bio);
1000 static void rrpc_gc_free(struct rrpc *rrpc)
1003 destroy_workqueue(rrpc->krqd_wq);
1006 destroy_workqueue(rrpc->kgc_wq);
1009 static int rrpc_gc_init(struct rrpc *rrpc)
1011 rrpc->krqd_wq = alloc_workqueue("rrpc-lun", WQ_MEM_RECLAIM|WQ_UNBOUND,
1016 rrpc->kgc_wq = alloc_workqueue("rrpc-bg", WQ_MEM_RECLAIM, 1);
1020 setup_timer(&rrpc->gc_timer, rrpc_gc_timer, (unsigned long)rrpc);
1025 static void rrpc_map_free(struct rrpc *rrpc)
1027 vfree(rrpc->rev_trans_map);
1028 vfree(rrpc->trans_map);
1031 static int rrpc_l2p_update(u64 slba, u32 nlb, __le64 *entries, void *private)
1033 struct rrpc *rrpc = (struct rrpc *)private;
1034 struct nvm_dev *dev = rrpc->dev;
1035 struct rrpc_addr *addr = rrpc->trans_map + slba;
1036 struct rrpc_rev_addr *raddr = rrpc->rev_trans_map;
1037 u64 elba = slba + nlb;
1040 if (unlikely(elba > dev->total_secs)) {
1041 pr_err("nvm: L2P data from device is out of bounds!\n");
1045 for (i = 0; i < nlb; i++) {
1046 u64 pba = le64_to_cpu(entries[i]);
1048 /* LNVM treats address-spaces as silos, LBA and PBA are
1049 * equally large and zero-indexed.
1051 if (unlikely(pba >= dev->total_secs && pba != U64_MAX)) {
1052 pr_err("nvm: L2P data entry is out of bounds!\n");
1056 /* Address zero is a special one. The first page on a disk is
1057 * protected. As it often holds internal device boot
1063 div_u64_rem(pba, rrpc->nr_sects, &mod);
1066 raddr[mod].addr = slba + i;
1072 static int rrpc_map_init(struct rrpc *rrpc)
1074 struct nvm_dev *dev = rrpc->dev;
1078 rrpc->trans_map = vzalloc(sizeof(struct rrpc_addr) * rrpc->nr_sects);
1079 if (!rrpc->trans_map)
1082 rrpc->rev_trans_map = vmalloc(sizeof(struct rrpc_rev_addr)
1084 if (!rrpc->rev_trans_map)
1087 for (i = 0; i < rrpc->nr_sects; i++) {
1088 struct rrpc_addr *p = &rrpc->trans_map[i];
1089 struct rrpc_rev_addr *r = &rrpc->rev_trans_map[i];
1091 p->addr = ADDR_EMPTY;
1092 r->addr = ADDR_EMPTY;
1095 if (!dev->ops->get_l2p_tbl)
1098 /* Bring up the mapping table from device */
1099 ret = dev->ops->get_l2p_tbl(dev, rrpc->soffset, rrpc->nr_sects,
1100 rrpc_l2p_update, rrpc);
1102 pr_err("nvm: rrpc: could not read L2P table.\n");
1109 /* Minimum pages needed within a lun */
1110 #define PAGE_POOL_SIZE 16
1111 #define ADDR_POOL_SIZE 64
1113 static int rrpc_core_init(struct rrpc *rrpc)
1115 down_write(&rrpc_lock);
1116 if (!rrpc_gcb_cache) {
1117 rrpc_gcb_cache = kmem_cache_create("rrpc_gcb",
1118 sizeof(struct rrpc_block_gc), 0, 0, NULL);
1119 if (!rrpc_gcb_cache) {
1120 up_write(&rrpc_lock);
1124 rrpc_rq_cache = kmem_cache_create("rrpc_rq",
1125 sizeof(struct nvm_rq) + sizeof(struct rrpc_rq),
1127 if (!rrpc_rq_cache) {
1128 kmem_cache_destroy(rrpc_gcb_cache);
1129 up_write(&rrpc_lock);
1133 up_write(&rrpc_lock);
1135 rrpc->page_pool = mempool_create_page_pool(PAGE_POOL_SIZE, 0);
1136 if (!rrpc->page_pool)
1139 rrpc->gcb_pool = mempool_create_slab_pool(rrpc->dev->nr_luns,
1141 if (!rrpc->gcb_pool)
1144 rrpc->rq_pool = mempool_create_slab_pool(64, rrpc_rq_cache);
1148 spin_lock_init(&rrpc->inflights.lock);
1149 INIT_LIST_HEAD(&rrpc->inflights.reqs);
1154 static void rrpc_core_free(struct rrpc *rrpc)
1156 mempool_destroy(rrpc->page_pool);
1157 mempool_destroy(rrpc->gcb_pool);
1158 mempool_destroy(rrpc->rq_pool);
1161 static void rrpc_luns_free(struct rrpc *rrpc)
1163 struct nvm_dev *dev = rrpc->dev;
1164 struct nvm_lun *lun;
1165 struct rrpc_lun *rlun;
1171 for (i = 0; i < rrpc->nr_luns; i++) {
1172 rlun = &rrpc->luns[i];
1176 dev->mt->release_lun(dev, lun->id);
1177 vfree(rlun->blocks);
1183 static int rrpc_luns_init(struct rrpc *rrpc, int lun_begin, int lun_end)
1185 struct nvm_dev *dev = rrpc->dev;
1186 struct rrpc_lun *rlun;
1187 int i, j, ret = -EINVAL;
1189 if (dev->sec_per_blk > MAX_INVALID_PAGES_STORAGE * BITS_PER_LONG) {
1190 pr_err("rrpc: number of pages per block too high.");
1194 spin_lock_init(&rrpc->rev_lock);
1196 rrpc->luns = kcalloc(rrpc->nr_luns, sizeof(struct rrpc_lun),
1202 for (i = 0; i < rrpc->nr_luns; i++) {
1203 int lunid = lun_begin + i;
1204 struct nvm_lun *lun;
1206 if (dev->mt->reserve_lun(dev, lunid)) {
1207 pr_err("rrpc: lun %u is already allocated\n", lunid);
1211 lun = dev->mt->get_lun(dev, lunid);
1215 rlun = &rrpc->luns[i];
1217 rlun->blocks = vzalloc(sizeof(struct rrpc_block) *
1218 rrpc->dev->blks_per_lun);
1219 if (!rlun->blocks) {
1224 for (j = 0; j < rrpc->dev->blks_per_lun; j++) {
1225 struct rrpc_block *rblk = &rlun->blocks[j];
1226 struct nvm_block *blk = &lun->blocks[j];
1230 INIT_LIST_HEAD(&rblk->prio);
1231 spin_lock_init(&rblk->lock);
1235 INIT_LIST_HEAD(&rlun->prio_list);
1236 INIT_LIST_HEAD(&rlun->wblk_list);
1238 INIT_WORK(&rlun->ws_gc, rrpc_lun_gc);
1239 spin_lock_init(&rlun->lock);
1247 /* returns 0 on success and stores the beginning address in *begin */
1248 static int rrpc_area_init(struct rrpc *rrpc, sector_t *begin)
1250 struct nvm_dev *dev = rrpc->dev;
1251 struct nvmm_type *mt = dev->mt;
1252 sector_t size = rrpc->nr_sects * dev->sec_size;
1257 ret = mt->get_area(dev, begin, size);
1259 *begin >>= (ilog2(dev->sec_size) - 9);
1264 static void rrpc_area_free(struct rrpc *rrpc)
1266 struct nvm_dev *dev = rrpc->dev;
1267 struct nvmm_type *mt = dev->mt;
1268 sector_t begin = rrpc->soffset << (ilog2(dev->sec_size) - 9);
1270 mt->put_area(dev, begin);
1273 static void rrpc_free(struct rrpc *rrpc)
1276 rrpc_map_free(rrpc);
1277 rrpc_core_free(rrpc);
1278 rrpc_luns_free(rrpc);
1279 rrpc_area_free(rrpc);
1284 static void rrpc_exit(void *private)
1286 struct rrpc *rrpc = private;
1288 del_timer(&rrpc->gc_timer);
1290 flush_workqueue(rrpc->krqd_wq);
1291 flush_workqueue(rrpc->kgc_wq);
1296 static sector_t rrpc_capacity(void *private)
1298 struct rrpc *rrpc = private;
1299 struct nvm_dev *dev = rrpc->dev;
1300 sector_t reserved, provisioned;
1302 /* cur, gc, and two emergency blocks for each lun */
1303 reserved = rrpc->nr_luns * dev->sec_per_blk * 4;
1304 provisioned = rrpc->nr_sects - reserved;
1306 if (reserved > rrpc->nr_sects) {
1307 pr_err("rrpc: not enough space available to expose storage.\n");
1311 sector_div(provisioned, 10);
1312 return provisioned * 9 * NR_PHY_IN_LOG;
1316 * Looks up the logical address from reverse trans map and check if its valid by
1317 * comparing the logical to physical address with the physical address.
1318 * Returns 0 on free, otherwise 1 if in use
1320 static void rrpc_block_map_update(struct rrpc *rrpc, struct rrpc_block *rblk)
1322 struct nvm_dev *dev = rrpc->dev;
1324 struct rrpc_addr *laddr;
1325 u64 bpaddr, paddr, pladdr;
1327 bpaddr = block_to_rel_addr(rrpc, rblk);
1328 for (offset = 0; offset < dev->sec_per_blk; offset++) {
1329 paddr = bpaddr + offset;
1331 pladdr = rrpc->rev_trans_map[paddr].addr;
1332 if (pladdr == ADDR_EMPTY)
1335 laddr = &rrpc->trans_map[pladdr];
1337 if (paddr == laddr->addr) {
1340 set_bit(offset, rblk->invalid_pages);
1341 rblk->nr_invalid_pages++;
1346 static int rrpc_blocks_init(struct rrpc *rrpc)
1348 struct rrpc_lun *rlun;
1349 struct rrpc_block *rblk;
1350 int lun_iter, blk_iter;
1352 for (lun_iter = 0; lun_iter < rrpc->nr_luns; lun_iter++) {
1353 rlun = &rrpc->luns[lun_iter];
1355 for (blk_iter = 0; blk_iter < rrpc->dev->blks_per_lun;
1357 rblk = &rlun->blocks[blk_iter];
1358 rrpc_block_map_update(rrpc, rblk);
1365 static int rrpc_luns_configure(struct rrpc *rrpc)
1367 struct rrpc_lun *rlun;
1368 struct rrpc_block *rblk;
1371 for (i = 0; i < rrpc->nr_luns; i++) {
1372 rlun = &rrpc->luns[i];
1374 rblk = rrpc_get_blk(rrpc, rlun, 0);
1377 rrpc_set_lun_cur(rlun, rblk, &rlun->cur);
1379 /* Emergency gc block */
1380 rblk = rrpc_get_blk(rrpc, rlun, 1);
1383 rrpc_set_lun_cur(rlun, rblk, &rlun->gc_cur);
1388 rrpc_put_blks(rrpc);
1392 static struct nvm_tgt_type tt_rrpc;
1394 static void *rrpc_init(struct nvm_dev *dev, struct gendisk *tdisk,
1395 int lun_begin, int lun_end)
1397 struct request_queue *bqueue = dev->q;
1398 struct request_queue *tqueue = tdisk->queue;
1403 if (!(dev->identity.dom & NVM_RSP_L2P)) {
1404 pr_err("nvm: rrpc: device does not support l2p (%x)\n",
1406 return ERR_PTR(-EINVAL);
1409 rrpc = kzalloc(sizeof(struct rrpc), GFP_KERNEL);
1411 return ERR_PTR(-ENOMEM);
1413 rrpc->instance.tt = &tt_rrpc;
1417 bio_list_init(&rrpc->requeue_bios);
1418 spin_lock_init(&rrpc->bio_lock);
1419 INIT_WORK(&rrpc->ws_requeue, rrpc_requeue);
1421 rrpc->nr_luns = lun_end - lun_begin + 1;
1422 rrpc->total_blocks = (unsigned long)dev->blks_per_lun * rrpc->nr_luns;
1423 rrpc->nr_sects = (unsigned long long)dev->sec_per_lun * rrpc->nr_luns;
1425 /* simple round-robin strategy */
1426 atomic_set(&rrpc->next_lun, -1);
1428 ret = rrpc_area_init(rrpc, &soffset);
1430 pr_err("nvm: rrpc: could not initialize area\n");
1431 return ERR_PTR(ret);
1433 rrpc->soffset = soffset;
1435 ret = rrpc_luns_init(rrpc, lun_begin, lun_end);
1437 pr_err("nvm: rrpc: could not initialize luns\n");
1441 rrpc->poffset = dev->sec_per_lun * lun_begin;
1442 rrpc->lun_offset = lun_begin;
1444 ret = rrpc_core_init(rrpc);
1446 pr_err("nvm: rrpc: could not initialize core\n");
1450 ret = rrpc_map_init(rrpc);
1452 pr_err("nvm: rrpc: could not initialize maps\n");
1456 ret = rrpc_blocks_init(rrpc);
1458 pr_err("nvm: rrpc: could not initialize state for blocks\n");
1462 ret = rrpc_luns_configure(rrpc);
1464 pr_err("nvm: rrpc: not enough blocks available in LUNs.\n");
1468 ret = rrpc_gc_init(rrpc);
1470 pr_err("nvm: rrpc: could not initialize gc\n");
1474 /* inherit the size from the underlying device */
1475 blk_queue_logical_block_size(tqueue, queue_physical_block_size(bqueue));
1476 blk_queue_max_hw_sectors(tqueue, queue_max_hw_sectors(bqueue));
1478 pr_info("nvm: rrpc initialized with %u luns and %llu pages.\n",
1479 rrpc->nr_luns, (unsigned long long)rrpc->nr_sects);
1481 mod_timer(&rrpc->gc_timer, jiffies + msecs_to_jiffies(10));
1486 return ERR_PTR(ret);
1489 /* round robin, page-based FTL, and cost-based GC */
1490 static struct nvm_tgt_type tt_rrpc = {
1492 .version = {1, 0, 0},
1494 .make_rq = rrpc_make_rq,
1495 .capacity = rrpc_capacity,
1496 .end_io = rrpc_end_io,
1502 static int __init rrpc_module_init(void)
1504 return nvm_register_tgt_type(&tt_rrpc);
1507 static void rrpc_module_exit(void)
1509 nvm_unregister_tgt_type(&tt_rrpc);
1512 module_init(rrpc_module_init);
1513 module_exit(rrpc_module_exit);
1514 MODULE_LICENSE("GPL v2");
1515 MODULE_DESCRIPTION("Block-Device Target for Open-Channel SSDs");