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
->pgs_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
);
105 pr_err("rrpc: unable to acquire inflight IO\n");
110 rrpc_invalidate_range(rrpc
, slba
, len
);
111 rrpc_inflight_laddr_release(rrpc
, rqd
);
114 static int block_is_full(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
116 return (rblk
->next_page
== rrpc
->dev
->pgs_per_blk
);
119 static u64
block_to_addr(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
121 struct nvm_block
*blk
= rblk
->parent
;
123 return blk
->id
* rrpc
->dev
->pgs_per_blk
;
126 static struct ppa_addr
linear_to_generic_addr(struct nvm_dev
*dev
,
130 int secs
, pgs
, blks
, luns
;
131 sector_t ppa
= r
.ppa
;
135 div_u64_rem(ppa
, dev
->sec_per_pg
, &secs
);
138 sector_div(ppa
, dev
->sec_per_pg
);
139 div_u64_rem(ppa
, dev
->sec_per_blk
, &pgs
);
142 sector_div(ppa
, dev
->pgs_per_blk
);
143 div_u64_rem(ppa
, dev
->blks_per_lun
, &blks
);
146 sector_div(ppa
, dev
->blks_per_lun
);
147 div_u64_rem(ppa
, dev
->luns_per_chnl
, &luns
);
150 sector_div(ppa
, dev
->luns_per_chnl
);
156 static struct ppa_addr
rrpc_ppa_to_gaddr(struct nvm_dev
*dev
, u64 addr
)
158 struct ppa_addr paddr
;
161 return linear_to_generic_addr(dev
, paddr
);
164 /* requires lun->lock taken */
165 static void rrpc_set_lun_cur(struct rrpc_lun
*rlun
, struct rrpc_block
*rblk
)
167 struct rrpc
*rrpc
= rlun
->rrpc
;
172 spin_lock(&rlun
->cur
->lock
);
173 WARN_ON(!block_is_full(rrpc
, rlun
->cur
));
174 spin_unlock(&rlun
->cur
->lock
);
179 static struct rrpc_block
*rrpc_get_blk(struct rrpc
*rrpc
, struct rrpc_lun
*rlun
,
182 struct nvm_lun
*lun
= rlun
->parent
;
183 struct nvm_block
*blk
;
184 struct rrpc_block
*rblk
;
186 spin_lock(&lun
->lock
);
187 blk
= nvm_get_blk_unlocked(rrpc
->dev
, rlun
->parent
, flags
);
189 pr_err("nvm: rrpc: cannot get new block from media manager\n");
190 spin_unlock(&lun
->lock
);
194 rblk
= &rlun
->blocks
[blk
->id
];
195 list_add_tail(&rblk
->list
, &rlun
->open_list
);
196 spin_unlock(&lun
->lock
);
199 bitmap_zero(rblk
->invalid_pages
, rrpc
->dev
->pgs_per_blk
);
201 rblk
->nr_invalid_pages
= 0;
202 atomic_set(&rblk
->data_cmnt_size
, 0);
207 static void rrpc_put_blk(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
209 struct rrpc_lun
*rlun
= rblk
->rlun
;
210 struct nvm_lun
*lun
= rlun
->parent
;
212 spin_lock(&lun
->lock
);
213 nvm_put_blk_unlocked(rrpc
->dev
, rblk
->parent
);
214 list_del(&rblk
->list
);
215 spin_unlock(&lun
->lock
);
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_pgs_per_blk
= rrpc
->dev
->pgs_per_blk
;
291 DECLARE_COMPLETION_ONSTACK(wait
);
293 if (bitmap_full(rblk
->invalid_pages
, nr_pgs_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_pgs_per_blk
)) < nr_pgs_per_blk
) {
312 phys_addr
= (rblk
->parent
->id
* nr_pgs_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
);
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
);
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_pgs_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 nvm_dev
*dev
= rrpc
->dev
;
399 struct nvm_lun
*lun
= rblk
->parent
->lun
;
400 struct rrpc_lun
*rlun
= &rrpc
->luns
[lun
->id
- rrpc
->lun_offset
];
402 mempool_free(gcb
, rrpc
->gcb_pool
);
403 pr_debug("nvm: block '%lu' being reclaimed\n", rblk
->parent
->id
);
405 if (rrpc_move_valid_pages(rrpc
, rblk
))
408 if (nvm_erase_blk(dev
, rblk
->parent
))
411 rrpc_put_blk(rrpc
, rblk
);
416 spin_lock(&rlun
->lock
);
417 list_add_tail(&rblk
->prio
, &rlun
->prio_list
);
418 spin_unlock(&rlun
->lock
);
421 /* the block with highest number of invalid pages, will be in the beginning
424 static struct rrpc_block
*rblock_max_invalid(struct rrpc_block
*ra
,
425 struct rrpc_block
*rb
)
427 if (ra
->nr_invalid_pages
== rb
->nr_invalid_pages
)
430 return (ra
->nr_invalid_pages
< rb
->nr_invalid_pages
) ? rb
: ra
;
433 /* linearly find the block with highest number of invalid pages
436 static struct rrpc_block
*block_prio_find_max(struct rrpc_lun
*rlun
)
438 struct list_head
*prio_list
= &rlun
->prio_list
;
439 struct rrpc_block
*rblock
, *max
;
441 BUG_ON(list_empty(prio_list
));
443 max
= list_first_entry(prio_list
, struct rrpc_block
, prio
);
444 list_for_each_entry(rblock
, prio_list
, prio
)
445 max
= rblock_max_invalid(max
, rblock
);
450 static void rrpc_lun_gc(struct work_struct
*work
)
452 struct rrpc_lun
*rlun
= container_of(work
, struct rrpc_lun
, ws_gc
);
453 struct rrpc
*rrpc
= rlun
->rrpc
;
454 struct nvm_lun
*lun
= rlun
->parent
;
455 struct rrpc_block_gc
*gcb
;
456 unsigned int nr_blocks_need
;
458 nr_blocks_need
= rrpc
->dev
->blks_per_lun
/ GC_LIMIT_INVERSE
;
460 if (nr_blocks_need
< rrpc
->nr_luns
)
461 nr_blocks_need
= rrpc
->nr_luns
;
463 spin_lock(&rlun
->lock
);
464 while (nr_blocks_need
> lun
->nr_free_blocks
&&
465 !list_empty(&rlun
->prio_list
)) {
466 struct rrpc_block
*rblock
= block_prio_find_max(rlun
);
467 struct nvm_block
*block
= rblock
->parent
;
469 if (!rblock
->nr_invalid_pages
)
472 gcb
= mempool_alloc(rrpc
->gcb_pool
, GFP_ATOMIC
);
476 list_del_init(&rblock
->prio
);
478 BUG_ON(!block_is_full(rrpc
, rblock
));
480 pr_debug("rrpc: selected block '%lu' for GC\n", block
->id
);
484 INIT_WORK(&gcb
->ws_gc
, rrpc_block_gc
);
486 queue_work(rrpc
->kgc_wq
, &gcb
->ws_gc
);
490 spin_unlock(&rlun
->lock
);
492 /* TODO: Hint that request queue can be started again */
495 static void rrpc_gc_queue(struct work_struct
*work
)
497 struct rrpc_block_gc
*gcb
= container_of(work
, struct rrpc_block_gc
,
499 struct rrpc
*rrpc
= gcb
->rrpc
;
500 struct rrpc_block
*rblk
= gcb
->rblk
;
501 struct nvm_lun
*lun
= rblk
->parent
->lun
;
502 struct rrpc_lun
*rlun
= &rrpc
->luns
[lun
->id
- rrpc
->lun_offset
];
504 spin_lock(&rlun
->lock
);
505 list_add_tail(&rblk
->prio
, &rlun
->prio_list
);
506 spin_unlock(&rlun
->lock
);
508 mempool_free(gcb
, rrpc
->gcb_pool
);
509 pr_debug("nvm: block '%lu' is full, allow GC (sched)\n",
513 static const struct block_device_operations rrpc_fops
= {
514 .owner
= THIS_MODULE
,
517 static struct rrpc_lun
*rrpc_get_lun_rr(struct rrpc
*rrpc
, int is_gc
)
520 struct rrpc_lun
*rlun
, *max_free
;
523 return get_next_lun(rrpc
);
525 /* during GC, we don't care about RR, instead we want to make
526 * sure that we maintain evenness between the block luns.
528 max_free
= &rrpc
->luns
[0];
529 /* prevent GC-ing lun from devouring pages of a lun with
530 * little free blocks. We don't take the lock as we only need an
533 rrpc_for_each_lun(rrpc
, rlun
, i
) {
534 if (rlun
->parent
->nr_free_blocks
>
535 max_free
->parent
->nr_free_blocks
)
542 static struct rrpc_addr
*rrpc_update_map(struct rrpc
*rrpc
, sector_t laddr
,
543 struct rrpc_block
*rblk
, u64 paddr
)
545 struct rrpc_addr
*gp
;
546 struct rrpc_rev_addr
*rev
;
548 BUG_ON(laddr
>= rrpc
->nr_pages
);
550 gp
= &rrpc
->trans_map
[laddr
];
551 spin_lock(&rrpc
->rev_lock
);
553 rrpc_page_invalidate(rrpc
, gp
);
558 rev
= &rrpc
->rev_trans_map
[gp
->addr
- rrpc
->poffset
];
560 spin_unlock(&rrpc
->rev_lock
);
565 static u64
rrpc_alloc_addr(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
567 u64 addr
= ADDR_EMPTY
;
569 spin_lock(&rblk
->lock
);
570 if (block_is_full(rrpc
, rblk
))
573 addr
= block_to_addr(rrpc
, rblk
) + rblk
->next_page
;
577 spin_unlock(&rblk
->lock
);
581 /* Simple round-robin Logical to physical address translation.
583 * Retrieve the mapping using the active append point. Then update the ap for
584 * the next write to the disk.
586 * Returns rrpc_addr with the physical address and block. Remember to return to
587 * rrpc->addr_cache when request is finished.
589 static struct rrpc_addr
*rrpc_map_page(struct rrpc
*rrpc
, sector_t laddr
,
592 struct rrpc_lun
*rlun
;
593 struct rrpc_block
*rblk
;
597 rlun
= rrpc_get_lun_rr(rrpc
, is_gc
);
600 if (!is_gc
&& lun
->nr_free_blocks
< rrpc
->nr_luns
* 4)
603 spin_lock(&rlun
->lock
);
607 paddr
= rrpc_alloc_addr(rrpc
, rblk
);
609 if (paddr
== ADDR_EMPTY
) {
610 rblk
= rrpc_get_blk(rrpc
, rlun
, 0);
612 rrpc_set_lun_cur(rlun
, rblk
);
617 /* retry from emergency gc block */
618 paddr
= rrpc_alloc_addr(rrpc
, rlun
->gc_cur
);
619 if (paddr
== ADDR_EMPTY
) {
620 rblk
= rrpc_get_blk(rrpc
, rlun
, 1);
622 pr_err("rrpc: no more blocks");
627 paddr
= rrpc_alloc_addr(rrpc
, rlun
->gc_cur
);
633 spin_unlock(&rlun
->lock
);
634 return rrpc_update_map(rrpc
, laddr
, rblk
, paddr
);
636 spin_unlock(&rlun
->lock
);
640 static void rrpc_run_gc(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
642 struct rrpc_block_gc
*gcb
;
644 gcb
= mempool_alloc(rrpc
->gcb_pool
, GFP_ATOMIC
);
646 pr_err("rrpc: unable to queue block for gc.");
653 INIT_WORK(&gcb
->ws_gc
, rrpc_gc_queue
);
654 queue_work(rrpc
->kgc_wq
, &gcb
->ws_gc
);
657 static void rrpc_end_io_write(struct rrpc
*rrpc
, struct rrpc_rq
*rrqd
,
658 sector_t laddr
, uint8_t npages
)
661 struct rrpc_block
*rblk
;
665 for (i
= 0; i
< npages
; i
++) {
666 p
= &rrpc
->trans_map
[laddr
+ i
];
668 lun
= rblk
->parent
->lun
;
670 cmnt_size
= atomic_inc_return(&rblk
->data_cmnt_size
);
671 if (unlikely(cmnt_size
== rrpc
->dev
->pgs_per_blk
)) {
672 struct nvm_block
*blk
= rblk
->parent
;
673 struct rrpc_lun
*rlun
= rblk
->rlun
;
675 spin_lock(&lun
->lock
);
676 lun
->nr_open_blocks
--;
677 lun
->nr_closed_blocks
++;
678 blk
->state
&= ~NVM_BLK_ST_OPEN
;
679 blk
->state
|= NVM_BLK_ST_CLOSED
;
680 list_move_tail(&rblk
->list
, &rlun
->closed_list
);
681 spin_unlock(&lun
->lock
);
683 rrpc_run_gc(rrpc
, rblk
);
688 static void rrpc_end_io(struct nvm_rq
*rqd
)
690 struct rrpc
*rrpc
= container_of(rqd
->ins
, struct rrpc
, instance
);
691 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
692 uint8_t npages
= rqd
->nr_pages
;
693 sector_t laddr
= rrpc_get_laddr(rqd
->bio
) - npages
;
695 if (bio_data_dir(rqd
->bio
) == WRITE
)
696 rrpc_end_io_write(rrpc
, rrqd
, laddr
, npages
);
700 if (rrqd
->flags
& NVM_IOTYPE_GC
)
703 rrpc_unlock_rq(rrpc
, rqd
);
706 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
708 nvm_dev_dma_free(rrpc
->dev
, rqd
->metadata
, rqd
->dma_metadata
);
710 mempool_free(rqd
, rrpc
->rq_pool
);
713 static int rrpc_read_ppalist_rq(struct rrpc
*rrpc
, struct bio
*bio
,
714 struct nvm_rq
*rqd
, unsigned long flags
, int npages
)
716 struct rrpc_inflight_rq
*r
= rrpc_get_inflight_rq(rqd
);
717 struct rrpc_addr
*gp
;
718 sector_t laddr
= rrpc_get_laddr(bio
);
719 int is_gc
= flags
& NVM_IOTYPE_GC
;
722 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
)) {
723 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
724 return NVM_IO_REQUEUE
;
727 for (i
= 0; i
< npages
; i
++) {
728 /* We assume that mapping occurs at 4KB granularity */
729 BUG_ON(!(laddr
+ i
>= 0 && laddr
+ i
< rrpc
->nr_pages
));
730 gp
= &rrpc
->trans_map
[laddr
+ i
];
733 rqd
->ppa_list
[i
] = rrpc_ppa_to_gaddr(rrpc
->dev
,
737 rrpc_unlock_laddr(rrpc
, r
);
738 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
,
744 rqd
->opcode
= NVM_OP_HBREAD
;
749 static int rrpc_read_rq(struct rrpc
*rrpc
, struct bio
*bio
, struct nvm_rq
*rqd
,
752 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
753 int is_gc
= flags
& NVM_IOTYPE_GC
;
754 sector_t laddr
= rrpc_get_laddr(bio
);
755 struct rrpc_addr
*gp
;
757 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
))
758 return NVM_IO_REQUEUE
;
760 BUG_ON(!(laddr
>= 0 && laddr
< rrpc
->nr_pages
));
761 gp
= &rrpc
->trans_map
[laddr
];
764 rqd
->ppa_addr
= rrpc_ppa_to_gaddr(rrpc
->dev
, gp
->addr
);
767 rrpc_unlock_rq(rrpc
, rqd
);
771 rqd
->opcode
= NVM_OP_HBREAD
;
777 static int rrpc_write_ppalist_rq(struct rrpc
*rrpc
, struct bio
*bio
,
778 struct nvm_rq
*rqd
, unsigned long flags
, int npages
)
780 struct rrpc_inflight_rq
*r
= rrpc_get_inflight_rq(rqd
);
782 sector_t laddr
= rrpc_get_laddr(bio
);
783 int is_gc
= flags
& NVM_IOTYPE_GC
;
786 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
)) {
787 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
788 return NVM_IO_REQUEUE
;
791 for (i
= 0; i
< npages
; i
++) {
792 /* We assume that mapping occurs at 4KB granularity */
793 p
= rrpc_map_page(rrpc
, laddr
+ i
, is_gc
);
796 rrpc_unlock_laddr(rrpc
, r
);
797 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
,
800 return NVM_IO_REQUEUE
;
803 rqd
->ppa_list
[i
] = rrpc_ppa_to_gaddr(rrpc
->dev
,
807 rqd
->opcode
= NVM_OP_HBWRITE
;
812 static int rrpc_write_rq(struct rrpc
*rrpc
, struct bio
*bio
,
813 struct nvm_rq
*rqd
, unsigned long flags
)
815 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
817 int is_gc
= flags
& NVM_IOTYPE_GC
;
818 sector_t laddr
= rrpc_get_laddr(bio
);
820 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
))
821 return NVM_IO_REQUEUE
;
823 p
= rrpc_map_page(rrpc
, laddr
, is_gc
);
826 rrpc_unlock_rq(rrpc
, rqd
);
828 return NVM_IO_REQUEUE
;
831 rqd
->ppa_addr
= rrpc_ppa_to_gaddr(rrpc
->dev
, p
->addr
);
832 rqd
->opcode
= NVM_OP_HBWRITE
;
838 static int rrpc_setup_rq(struct rrpc
*rrpc
, struct bio
*bio
,
839 struct nvm_rq
*rqd
, unsigned long flags
, uint8_t npages
)
842 rqd
->ppa_list
= nvm_dev_dma_alloc(rrpc
->dev
, GFP_KERNEL
,
844 if (!rqd
->ppa_list
) {
845 pr_err("rrpc: not able to allocate ppa list\n");
849 if (bio_rw(bio
) == WRITE
)
850 return rrpc_write_ppalist_rq(rrpc
, bio
, rqd
, flags
,
853 return rrpc_read_ppalist_rq(rrpc
, bio
, rqd
, flags
, npages
);
856 if (bio_rw(bio
) == WRITE
)
857 return rrpc_write_rq(rrpc
, bio
, rqd
, flags
);
859 return rrpc_read_rq(rrpc
, bio
, rqd
, flags
);
862 static int rrpc_submit_io(struct rrpc
*rrpc
, struct bio
*bio
,
863 struct nvm_rq
*rqd
, unsigned long flags
)
866 struct rrpc_rq
*rrq
= nvm_rq_to_pdu(rqd
);
867 uint8_t nr_pages
= rrpc_get_pages(bio
);
868 int bio_size
= bio_sectors(bio
) << 9;
870 if (bio_size
< rrpc
->dev
->sec_size
)
872 else if (bio_size
> rrpc
->dev
->max_rq_size
)
875 err
= rrpc_setup_rq(rrpc
, bio
, rqd
, flags
, nr_pages
);
881 rqd
->ins
= &rrpc
->instance
;
882 rqd
->nr_pages
= nr_pages
;
885 err
= nvm_submit_io(rrpc
->dev
, rqd
);
887 pr_err("rrpc: I/O submission failed: %d\n", err
);
889 if (!(flags
& NVM_IOTYPE_GC
)) {
890 rrpc_unlock_rq(rrpc
, rqd
);
891 if (rqd
->nr_pages
> 1)
892 nvm_dev_dma_free(rrpc
->dev
,
893 rqd
->ppa_list
, rqd
->dma_ppa_list
);
901 static blk_qc_t
rrpc_make_rq(struct request_queue
*q
, struct bio
*bio
)
903 struct rrpc
*rrpc
= q
->queuedata
;
907 if (bio
->bi_rw
& REQ_DISCARD
) {
908 rrpc_discard(rrpc
, bio
);
909 return BLK_QC_T_NONE
;
912 rqd
= mempool_alloc(rrpc
->rq_pool
, GFP_KERNEL
);
914 pr_err_ratelimited("rrpc: not able to queue bio.");
916 return BLK_QC_T_NONE
;
918 memset(rqd
, 0, sizeof(struct nvm_rq
));
920 err
= rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_NONE
);
923 return BLK_QC_T_NONE
;
931 spin_lock(&rrpc
->bio_lock
);
932 bio_list_add(&rrpc
->requeue_bios
, bio
);
933 spin_unlock(&rrpc
->bio_lock
);
934 queue_work(rrpc
->kgc_wq
, &rrpc
->ws_requeue
);
938 mempool_free(rqd
, rrpc
->rq_pool
);
939 return BLK_QC_T_NONE
;
942 static void rrpc_requeue(struct work_struct
*work
)
944 struct rrpc
*rrpc
= container_of(work
, struct rrpc
, ws_requeue
);
945 struct bio_list bios
;
948 bio_list_init(&bios
);
950 spin_lock(&rrpc
->bio_lock
);
951 bio_list_merge(&bios
, &rrpc
->requeue_bios
);
952 bio_list_init(&rrpc
->requeue_bios
);
953 spin_unlock(&rrpc
->bio_lock
);
955 while ((bio
= bio_list_pop(&bios
)))
956 rrpc_make_rq(rrpc
->disk
->queue
, bio
);
959 static void rrpc_gc_free(struct rrpc
*rrpc
)
961 struct rrpc_lun
*rlun
;
965 destroy_workqueue(rrpc
->krqd_wq
);
968 destroy_workqueue(rrpc
->kgc_wq
);
973 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
974 rlun
= &rrpc
->luns
[i
];
982 static int rrpc_gc_init(struct rrpc
*rrpc
)
984 rrpc
->krqd_wq
= alloc_workqueue("rrpc-lun", WQ_MEM_RECLAIM
|WQ_UNBOUND
,
989 rrpc
->kgc_wq
= alloc_workqueue("rrpc-bg", WQ_MEM_RECLAIM
, 1);
993 setup_timer(&rrpc
->gc_timer
, rrpc_gc_timer
, (unsigned long)rrpc
);
998 static void rrpc_map_free(struct rrpc
*rrpc
)
1000 vfree(rrpc
->rev_trans_map
);
1001 vfree(rrpc
->trans_map
);
1004 static int rrpc_l2p_update(u64 slba
, u32 nlb
, __le64
*entries
, void *private)
1006 struct rrpc
*rrpc
= (struct rrpc
*)private;
1007 struct nvm_dev
*dev
= rrpc
->dev
;
1008 struct rrpc_addr
*addr
= rrpc
->trans_map
+ slba
;
1009 struct rrpc_rev_addr
*raddr
= rrpc
->rev_trans_map
;
1010 sector_t max_pages
= dev
->total_pages
* (dev
->sec_size
>> 9);
1011 u64 elba
= slba
+ nlb
;
1014 if (unlikely(elba
> dev
->total_pages
)) {
1015 pr_err("nvm: L2P data from device is out of bounds!\n");
1019 for (i
= 0; i
< nlb
; i
++) {
1020 u64 pba
= le64_to_cpu(entries
[i
]);
1021 /* LNVM treats address-spaces as silos, LBA and PBA are
1022 * equally large and zero-indexed.
1024 if (unlikely(pba
>= max_pages
&& pba
!= U64_MAX
)) {
1025 pr_err("nvm: L2P data entry is out of bounds!\n");
1029 /* Address zero is a special one. The first page on a disk is
1030 * protected. As it often holds internal device boot
1037 raddr
[pba
].addr
= slba
+ i
;
1043 static int rrpc_map_init(struct rrpc
*rrpc
)
1045 struct nvm_dev
*dev
= rrpc
->dev
;
1049 rrpc
->trans_map
= vzalloc(sizeof(struct rrpc_addr
) * rrpc
->nr_pages
);
1050 if (!rrpc
->trans_map
)
1053 rrpc
->rev_trans_map
= vmalloc(sizeof(struct rrpc_rev_addr
)
1055 if (!rrpc
->rev_trans_map
)
1058 for (i
= 0; i
< rrpc
->nr_pages
; i
++) {
1059 struct rrpc_addr
*p
= &rrpc
->trans_map
[i
];
1060 struct rrpc_rev_addr
*r
= &rrpc
->rev_trans_map
[i
];
1062 p
->addr
= ADDR_EMPTY
;
1063 r
->addr
= ADDR_EMPTY
;
1066 if (!dev
->ops
->get_l2p_tbl
)
1069 /* Bring up the mapping table from device */
1070 ret
= dev
->ops
->get_l2p_tbl(dev
, 0, dev
->total_pages
,
1071 rrpc_l2p_update
, rrpc
);
1073 pr_err("nvm: rrpc: could not read L2P table.\n");
1081 /* Minimum pages needed within a lun */
1082 #define PAGE_POOL_SIZE 16
1083 #define ADDR_POOL_SIZE 64
1085 static int rrpc_core_init(struct rrpc
*rrpc
)
1087 down_write(&rrpc_lock
);
1088 if (!rrpc_gcb_cache
) {
1089 rrpc_gcb_cache
= kmem_cache_create("rrpc_gcb",
1090 sizeof(struct rrpc_block_gc
), 0, 0, NULL
);
1091 if (!rrpc_gcb_cache
) {
1092 up_write(&rrpc_lock
);
1096 rrpc_rq_cache
= kmem_cache_create("rrpc_rq",
1097 sizeof(struct nvm_rq
) + sizeof(struct rrpc_rq
),
1099 if (!rrpc_rq_cache
) {
1100 kmem_cache_destroy(rrpc_gcb_cache
);
1101 up_write(&rrpc_lock
);
1105 up_write(&rrpc_lock
);
1107 rrpc
->page_pool
= mempool_create_page_pool(PAGE_POOL_SIZE
, 0);
1108 if (!rrpc
->page_pool
)
1111 rrpc
->gcb_pool
= mempool_create_slab_pool(rrpc
->dev
->nr_luns
,
1113 if (!rrpc
->gcb_pool
)
1116 rrpc
->rq_pool
= mempool_create_slab_pool(64, rrpc_rq_cache
);
1120 spin_lock_init(&rrpc
->inflights
.lock
);
1121 INIT_LIST_HEAD(&rrpc
->inflights
.reqs
);
1126 static void rrpc_core_free(struct rrpc
*rrpc
)
1128 mempool_destroy(rrpc
->page_pool
);
1129 mempool_destroy(rrpc
->gcb_pool
);
1130 mempool_destroy(rrpc
->rq_pool
);
1133 static void rrpc_luns_free(struct rrpc
*rrpc
)
1138 static int rrpc_luns_init(struct rrpc
*rrpc
, int lun_begin
, int lun_end
)
1140 struct nvm_dev
*dev
= rrpc
->dev
;
1141 struct rrpc_lun
*rlun
;
1144 if (dev
->pgs_per_blk
> MAX_INVALID_PAGES_STORAGE
* BITS_PER_LONG
) {
1145 pr_err("rrpc: number of pages per block too high.");
1149 spin_lock_init(&rrpc
->rev_lock
);
1151 rrpc
->luns
= kcalloc(rrpc
->nr_luns
, sizeof(struct rrpc_lun
),
1157 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
1158 struct nvm_lun
*lun
= dev
->mt
->get_lun(dev
, lun_begin
+ i
);
1160 rlun
= &rrpc
->luns
[i
];
1163 INIT_LIST_HEAD(&rlun
->prio_list
);
1164 INIT_LIST_HEAD(&rlun
->open_list
);
1165 INIT_LIST_HEAD(&rlun
->closed_list
);
1167 INIT_WORK(&rlun
->ws_gc
, rrpc_lun_gc
);
1168 spin_lock_init(&rlun
->lock
);
1170 rrpc
->total_blocks
+= dev
->blks_per_lun
;
1171 rrpc
->nr_pages
+= dev
->sec_per_lun
;
1173 rlun
->blocks
= vzalloc(sizeof(struct rrpc_block
) *
1174 rrpc
->dev
->blks_per_lun
);
1178 for (j
= 0; j
< rrpc
->dev
->blks_per_lun
; j
++) {
1179 struct rrpc_block
*rblk
= &rlun
->blocks
[j
];
1180 struct nvm_block
*blk
= &lun
->blocks
[j
];
1184 INIT_LIST_HEAD(&rblk
->prio
);
1185 spin_lock_init(&rblk
->lock
);
1194 static void rrpc_free(struct rrpc
*rrpc
)
1197 rrpc_map_free(rrpc
);
1198 rrpc_core_free(rrpc
);
1199 rrpc_luns_free(rrpc
);
1204 static void rrpc_exit(void *private)
1206 struct rrpc
*rrpc
= private;
1208 del_timer(&rrpc
->gc_timer
);
1210 flush_workqueue(rrpc
->krqd_wq
);
1211 flush_workqueue(rrpc
->kgc_wq
);
1216 static sector_t
rrpc_capacity(void *private)
1218 struct rrpc
*rrpc
= private;
1219 struct nvm_dev
*dev
= rrpc
->dev
;
1220 sector_t reserved
, provisioned
;
1222 /* cur, gc, and two emergency blocks for each lun */
1223 reserved
= rrpc
->nr_luns
* dev
->max_pages_per_blk
* 4;
1224 provisioned
= rrpc
->nr_pages
- reserved
;
1226 if (reserved
> rrpc
->nr_pages
) {
1227 pr_err("rrpc: not enough space available to expose storage.\n");
1231 sector_div(provisioned
, 10);
1232 return provisioned
* 9 * NR_PHY_IN_LOG
;
1236 * Looks up the logical address from reverse trans map and check if its valid by
1237 * comparing the logical to physical address with the physical address.
1238 * Returns 0 on free, otherwise 1 if in use
1240 static void rrpc_block_map_update(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
1242 struct nvm_dev
*dev
= rrpc
->dev
;
1244 struct rrpc_addr
*laddr
;
1247 for (offset
= 0; offset
< dev
->pgs_per_blk
; offset
++) {
1248 paddr
= block_to_addr(rrpc
, rblk
) + offset
;
1250 pladdr
= rrpc
->rev_trans_map
[paddr
].addr
;
1251 if (pladdr
== ADDR_EMPTY
)
1254 laddr
= &rrpc
->trans_map
[pladdr
];
1256 if (paddr
== laddr
->addr
) {
1259 set_bit(offset
, rblk
->invalid_pages
);
1260 rblk
->nr_invalid_pages
++;
1265 static int rrpc_blocks_init(struct rrpc
*rrpc
)
1267 struct rrpc_lun
*rlun
;
1268 struct rrpc_block
*rblk
;
1269 int lun_iter
, blk_iter
;
1271 for (lun_iter
= 0; lun_iter
< rrpc
->nr_luns
; lun_iter
++) {
1272 rlun
= &rrpc
->luns
[lun_iter
];
1274 for (blk_iter
= 0; blk_iter
< rrpc
->dev
->blks_per_lun
;
1276 rblk
= &rlun
->blocks
[blk_iter
];
1277 rrpc_block_map_update(rrpc
, rblk
);
1284 static int rrpc_luns_configure(struct rrpc
*rrpc
)
1286 struct rrpc_lun
*rlun
;
1287 struct rrpc_block
*rblk
;
1290 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
1291 rlun
= &rrpc
->luns
[i
];
1293 rblk
= rrpc_get_blk(rrpc
, rlun
, 0);
1297 rrpc_set_lun_cur(rlun
, rblk
);
1299 /* Emergency gc block */
1300 rblk
= rrpc_get_blk(rrpc
, rlun
, 1);
1303 rlun
->gc_cur
= rblk
;
1308 rrpc_put_blks(rrpc
);
1312 static struct nvm_tgt_type tt_rrpc
;
1314 static void *rrpc_init(struct nvm_dev
*dev
, struct gendisk
*tdisk
,
1315 int lun_begin
, int lun_end
)
1317 struct request_queue
*bqueue
= dev
->q
;
1318 struct request_queue
*tqueue
= tdisk
->queue
;
1322 if (!(dev
->identity
.dom
& NVM_RSP_L2P
)) {
1323 pr_err("nvm: rrpc: device does not support l2p (%x)\n",
1325 return ERR_PTR(-EINVAL
);
1328 rrpc
= kzalloc(sizeof(struct rrpc
), GFP_KERNEL
);
1330 return ERR_PTR(-ENOMEM
);
1332 rrpc
->instance
.tt
= &tt_rrpc
;
1336 bio_list_init(&rrpc
->requeue_bios
);
1337 spin_lock_init(&rrpc
->bio_lock
);
1338 INIT_WORK(&rrpc
->ws_requeue
, rrpc_requeue
);
1340 rrpc
->nr_luns
= lun_end
- lun_begin
+ 1;
1342 /* simple round-robin strategy */
1343 atomic_set(&rrpc
->next_lun
, -1);
1345 ret
= rrpc_luns_init(rrpc
, lun_begin
, lun_end
);
1347 pr_err("nvm: rrpc: could not initialize luns\n");
1351 rrpc
->poffset
= dev
->sec_per_lun
* lun_begin
;
1352 rrpc
->lun_offset
= lun_begin
;
1354 ret
= rrpc_core_init(rrpc
);
1356 pr_err("nvm: rrpc: could not initialize core\n");
1360 ret
= rrpc_map_init(rrpc
);
1362 pr_err("nvm: rrpc: could not initialize maps\n");
1366 ret
= rrpc_blocks_init(rrpc
);
1368 pr_err("nvm: rrpc: could not initialize state for blocks\n");
1372 ret
= rrpc_luns_configure(rrpc
);
1374 pr_err("nvm: rrpc: not enough blocks available in LUNs.\n");
1378 ret
= rrpc_gc_init(rrpc
);
1380 pr_err("nvm: rrpc: could not initialize gc\n");
1384 /* inherit the size from the underlying device */
1385 blk_queue_logical_block_size(tqueue
, queue_physical_block_size(bqueue
));
1386 blk_queue_max_hw_sectors(tqueue
, queue_max_hw_sectors(bqueue
));
1388 pr_info("nvm: rrpc initialized with %u luns and %llu pages.\n",
1389 rrpc
->nr_luns
, (unsigned long long)rrpc
->nr_pages
);
1391 mod_timer(&rrpc
->gc_timer
, jiffies
+ msecs_to_jiffies(10));
1396 return ERR_PTR(ret
);
1399 /* round robin, page-based FTL, and cost-based GC */
1400 static struct nvm_tgt_type tt_rrpc
= {
1402 .version
= {1, 0, 0},
1404 .make_rq
= rrpc_make_rq
,
1405 .capacity
= rrpc_capacity
,
1406 .end_io
= rrpc_end_io
,
1412 static int __init
rrpc_module_init(void)
1414 return nvm_register_target(&tt_rrpc
);
1417 static void rrpc_module_exit(void)
1419 nvm_unregister_target(&tt_rrpc
);
1422 module_init(rrpc_module_init
);
1423 module_exit(rrpc_module_exit
);
1424 MODULE_LICENSE("GPL v2");
1425 MODULE_DESCRIPTION("Block-Device Target for Open-Channel SSDs");