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
);
306 while ((slot
= find_first_zero_bit(rblk
->invalid_pages
,
307 nr_pgs_per_blk
)) < nr_pgs_per_blk
) {
310 phys_addr
= (rblk
->parent
->id
* nr_pgs_per_blk
) + slot
;
313 spin_lock(&rrpc
->rev_lock
);
314 /* Get logical address from physical to logical table */
315 rev
= &rrpc
->rev_trans_map
[phys_addr
- rrpc
->poffset
];
316 /* already updated by previous regular write */
317 if (rev
->addr
== ADDR_EMPTY
) {
318 spin_unlock(&rrpc
->rev_lock
);
322 rqd
= rrpc_inflight_laddr_acquire(rrpc
, rev
->addr
, 1);
323 if (IS_ERR_OR_NULL(rqd
)) {
324 spin_unlock(&rrpc
->rev_lock
);
329 spin_unlock(&rrpc
->rev_lock
);
331 /* Perform read to do GC */
332 bio
->bi_iter
.bi_sector
= rrpc_get_sector(rev
->addr
);
334 bio
->bi_private
= &wait
;
335 bio
->bi_end_io
= rrpc_end_sync_bio
;
337 /* TODO: may fail when EXP_PG_SIZE > PAGE_SIZE */
338 bio_add_pc_page(q
, bio
, page
, RRPC_EXPOSED_PAGE_SIZE
, 0);
340 if (rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_GC
)) {
341 pr_err("rrpc: gc read failed.\n");
342 rrpc_inflight_laddr_release(rrpc
, rqd
);
345 wait_for_completion_io(&wait
);
347 rrpc_inflight_laddr_release(rrpc
, rqd
);
352 reinit_completion(&wait
);
354 bio
->bi_iter
.bi_sector
= rrpc_get_sector(rev
->addr
);
356 bio
->bi_private
= &wait
;
357 bio
->bi_end_io
= rrpc_end_sync_bio
;
359 bio_add_pc_page(q
, bio
, page
, RRPC_EXPOSED_PAGE_SIZE
, 0);
361 /* turn the command around and write the data back to a new
364 if (rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_GC
)) {
365 pr_err("rrpc: gc write failed.\n");
366 rrpc_inflight_laddr_release(rrpc
, rqd
);
369 wait_for_completion_io(&wait
);
371 rrpc_inflight_laddr_release(rrpc
, rqd
);
379 mempool_free(page
, rrpc
->page_pool
);
382 if (!bitmap_full(rblk
->invalid_pages
, nr_pgs_per_blk
)) {
383 pr_err("nvm: failed to garbage collect block\n");
390 static void rrpc_block_gc(struct work_struct
*work
)
392 struct rrpc_block_gc
*gcb
= container_of(work
, struct rrpc_block_gc
,
394 struct rrpc
*rrpc
= gcb
->rrpc
;
395 struct rrpc_block
*rblk
= gcb
->rblk
;
396 struct nvm_dev
*dev
= rrpc
->dev
;
397 struct nvm_lun
*lun
= rblk
->parent
->lun
;
398 struct rrpc_lun
*rlun
= &rrpc
->luns
[lun
->id
- rrpc
->lun_offset
];
400 mempool_free(gcb
, rrpc
->gcb_pool
);
401 pr_debug("nvm: block '%lu' being reclaimed\n", rblk
->parent
->id
);
403 if (rrpc_move_valid_pages(rrpc
, rblk
))
406 if (nvm_erase_blk(dev
, rblk
->parent
))
409 rrpc_put_blk(rrpc
, rblk
);
414 spin_lock(&rlun
->lock
);
415 list_add_tail(&rblk
->prio
, &rlun
->prio_list
);
416 spin_unlock(&rlun
->lock
);
419 /* the block with highest number of invalid pages, will be in the beginning
422 static struct rrpc_block
*rblock_max_invalid(struct rrpc_block
*ra
,
423 struct rrpc_block
*rb
)
425 if (ra
->nr_invalid_pages
== rb
->nr_invalid_pages
)
428 return (ra
->nr_invalid_pages
< rb
->nr_invalid_pages
) ? rb
: ra
;
431 /* linearly find the block with highest number of invalid pages
434 static struct rrpc_block
*block_prio_find_max(struct rrpc_lun
*rlun
)
436 struct list_head
*prio_list
= &rlun
->prio_list
;
437 struct rrpc_block
*rblock
, *max
;
439 BUG_ON(list_empty(prio_list
));
441 max
= list_first_entry(prio_list
, struct rrpc_block
, prio
);
442 list_for_each_entry(rblock
, prio_list
, prio
)
443 max
= rblock_max_invalid(max
, rblock
);
448 static void rrpc_lun_gc(struct work_struct
*work
)
450 struct rrpc_lun
*rlun
= container_of(work
, struct rrpc_lun
, ws_gc
);
451 struct rrpc
*rrpc
= rlun
->rrpc
;
452 struct nvm_lun
*lun
= rlun
->parent
;
453 struct rrpc_block_gc
*gcb
;
454 unsigned int nr_blocks_need
;
456 nr_blocks_need
= rrpc
->dev
->blks_per_lun
/ GC_LIMIT_INVERSE
;
458 if (nr_blocks_need
< rrpc
->nr_luns
)
459 nr_blocks_need
= rrpc
->nr_luns
;
461 spin_lock(&rlun
->lock
);
462 while (nr_blocks_need
> lun
->nr_free_blocks
&&
463 !list_empty(&rlun
->prio_list
)) {
464 struct rrpc_block
*rblock
= block_prio_find_max(rlun
);
465 struct nvm_block
*block
= rblock
->parent
;
467 if (!rblock
->nr_invalid_pages
)
470 gcb
= mempool_alloc(rrpc
->gcb_pool
, GFP_ATOMIC
);
474 list_del_init(&rblock
->prio
);
476 BUG_ON(!block_is_full(rrpc
, rblock
));
478 pr_debug("rrpc: selected block '%lu' for GC\n", block
->id
);
482 INIT_WORK(&gcb
->ws_gc
, rrpc_block_gc
);
484 queue_work(rrpc
->kgc_wq
, &gcb
->ws_gc
);
488 spin_unlock(&rlun
->lock
);
490 /* TODO: Hint that request queue can be started again */
493 static void rrpc_gc_queue(struct work_struct
*work
)
495 struct rrpc_block_gc
*gcb
= container_of(work
, struct rrpc_block_gc
,
497 struct rrpc
*rrpc
= gcb
->rrpc
;
498 struct rrpc_block
*rblk
= gcb
->rblk
;
499 struct nvm_lun
*lun
= rblk
->parent
->lun
;
500 struct rrpc_lun
*rlun
= &rrpc
->luns
[lun
->id
- rrpc
->lun_offset
];
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_pages
);
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 /* Simple round-robin Logical to physical address translation.
581 * Retrieve the mapping using the active append point. Then update the ap for
582 * the next write to the disk.
584 * Returns rrpc_addr with the physical address and block. Remember to return to
585 * rrpc->addr_cache when request is finished.
587 static struct rrpc_addr
*rrpc_map_page(struct rrpc
*rrpc
, sector_t laddr
,
590 struct rrpc_lun
*rlun
;
591 struct rrpc_block
*rblk
;
595 rlun
= rrpc_get_lun_rr(rrpc
, is_gc
);
598 if (!is_gc
&& lun
->nr_free_blocks
< rrpc
->nr_luns
* 4)
601 spin_lock(&rlun
->lock
);
605 paddr
= rrpc_alloc_addr(rrpc
, rblk
);
607 if (paddr
== ADDR_EMPTY
) {
608 rblk
= rrpc_get_blk(rrpc
, rlun
, 0);
610 rrpc_set_lun_cur(rlun
, rblk
);
615 /* retry from emergency gc block */
616 paddr
= rrpc_alloc_addr(rrpc
, rlun
->gc_cur
);
617 if (paddr
== ADDR_EMPTY
) {
618 rblk
= rrpc_get_blk(rrpc
, rlun
, 1);
620 pr_err("rrpc: no more blocks");
625 paddr
= rrpc_alloc_addr(rrpc
, rlun
->gc_cur
);
631 spin_unlock(&rlun
->lock
);
632 return rrpc_update_map(rrpc
, laddr
, rblk
, paddr
);
634 spin_unlock(&rlun
->lock
);
638 static void rrpc_run_gc(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
640 struct rrpc_block_gc
*gcb
;
642 gcb
= mempool_alloc(rrpc
->gcb_pool
, GFP_ATOMIC
);
644 pr_err("rrpc: unable to queue block for gc.");
651 INIT_WORK(&gcb
->ws_gc
, rrpc_gc_queue
);
652 queue_work(rrpc
->kgc_wq
, &gcb
->ws_gc
);
655 static void rrpc_end_io_write(struct rrpc
*rrpc
, struct rrpc_rq
*rrqd
,
656 sector_t laddr
, uint8_t npages
)
659 struct rrpc_block
*rblk
;
663 for (i
= 0; i
< npages
; i
++) {
664 p
= &rrpc
->trans_map
[laddr
+ i
];
666 lun
= rblk
->parent
->lun
;
668 cmnt_size
= atomic_inc_return(&rblk
->data_cmnt_size
);
669 if (unlikely(cmnt_size
== rrpc
->dev
->pgs_per_blk
)) {
670 struct nvm_block
*blk
= rblk
->parent
;
671 struct rrpc_lun
*rlun
= rblk
->rlun
;
673 spin_lock(&lun
->lock
);
674 lun
->nr_open_blocks
--;
675 lun
->nr_closed_blocks
++;
676 blk
->state
&= ~NVM_BLK_ST_OPEN
;
677 blk
->state
|= NVM_BLK_ST_CLOSED
;
678 list_move_tail(&rblk
->list
, &rlun
->closed_list
);
679 spin_unlock(&lun
->lock
);
681 rrpc_run_gc(rrpc
, rblk
);
686 static void rrpc_end_io(struct nvm_rq
*rqd
)
688 struct rrpc
*rrpc
= container_of(rqd
->ins
, struct rrpc
, instance
);
689 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
690 uint8_t npages
= rqd
->nr_pages
;
691 sector_t laddr
= rrpc_get_laddr(rqd
->bio
) - npages
;
693 if (bio_data_dir(rqd
->bio
) == WRITE
)
694 rrpc_end_io_write(rrpc
, rrqd
, laddr
, npages
);
698 if (rrqd
->flags
& NVM_IOTYPE_GC
)
701 rrpc_unlock_rq(rrpc
, rqd
);
704 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
706 nvm_dev_dma_free(rrpc
->dev
, rqd
->metadata
, rqd
->dma_metadata
);
708 mempool_free(rqd
, rrpc
->rq_pool
);
711 static int rrpc_read_ppalist_rq(struct rrpc
*rrpc
, struct bio
*bio
,
712 struct nvm_rq
*rqd
, unsigned long flags
, int npages
)
714 struct rrpc_inflight_rq
*r
= rrpc_get_inflight_rq(rqd
);
715 struct rrpc_addr
*gp
;
716 sector_t laddr
= rrpc_get_laddr(bio
);
717 int is_gc
= flags
& NVM_IOTYPE_GC
;
720 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
)) {
721 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
722 return NVM_IO_REQUEUE
;
725 for (i
= 0; i
< npages
; i
++) {
726 /* We assume that mapping occurs at 4KB granularity */
727 BUG_ON(!(laddr
+ i
>= 0 && laddr
+ i
< rrpc
->nr_pages
));
728 gp
= &rrpc
->trans_map
[laddr
+ i
];
731 rqd
->ppa_list
[i
] = rrpc_ppa_to_gaddr(rrpc
->dev
,
735 rrpc_unlock_laddr(rrpc
, r
);
736 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
,
742 rqd
->opcode
= NVM_OP_HBREAD
;
747 static int rrpc_read_rq(struct rrpc
*rrpc
, struct bio
*bio
, struct nvm_rq
*rqd
,
750 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
751 int is_gc
= flags
& NVM_IOTYPE_GC
;
752 sector_t laddr
= rrpc_get_laddr(bio
);
753 struct rrpc_addr
*gp
;
755 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
))
756 return NVM_IO_REQUEUE
;
758 BUG_ON(!(laddr
>= 0 && laddr
< rrpc
->nr_pages
));
759 gp
= &rrpc
->trans_map
[laddr
];
762 rqd
->ppa_addr
= rrpc_ppa_to_gaddr(rrpc
->dev
, gp
->addr
);
765 rrpc_unlock_rq(rrpc
, rqd
);
769 rqd
->opcode
= NVM_OP_HBREAD
;
775 static int rrpc_write_ppalist_rq(struct rrpc
*rrpc
, struct bio
*bio
,
776 struct nvm_rq
*rqd
, unsigned long flags
, int npages
)
778 struct rrpc_inflight_rq
*r
= rrpc_get_inflight_rq(rqd
);
780 sector_t laddr
= rrpc_get_laddr(bio
);
781 int is_gc
= flags
& NVM_IOTYPE_GC
;
784 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
)) {
785 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
786 return NVM_IO_REQUEUE
;
789 for (i
= 0; i
< npages
; i
++) {
790 /* We assume that mapping occurs at 4KB granularity */
791 p
= rrpc_map_page(rrpc
, laddr
+ i
, is_gc
);
794 rrpc_unlock_laddr(rrpc
, r
);
795 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
,
798 return NVM_IO_REQUEUE
;
801 rqd
->ppa_list
[i
] = rrpc_ppa_to_gaddr(rrpc
->dev
,
805 rqd
->opcode
= NVM_OP_HBWRITE
;
810 static int rrpc_write_rq(struct rrpc
*rrpc
, struct bio
*bio
,
811 struct nvm_rq
*rqd
, unsigned long flags
)
813 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
815 int is_gc
= flags
& NVM_IOTYPE_GC
;
816 sector_t laddr
= rrpc_get_laddr(bio
);
818 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
))
819 return NVM_IO_REQUEUE
;
821 p
= rrpc_map_page(rrpc
, laddr
, is_gc
);
824 rrpc_unlock_rq(rrpc
, rqd
);
826 return NVM_IO_REQUEUE
;
829 rqd
->ppa_addr
= rrpc_ppa_to_gaddr(rrpc
->dev
, p
->addr
);
830 rqd
->opcode
= NVM_OP_HBWRITE
;
836 static int rrpc_setup_rq(struct rrpc
*rrpc
, struct bio
*bio
,
837 struct nvm_rq
*rqd
, unsigned long flags
, uint8_t npages
)
840 rqd
->ppa_list
= nvm_dev_dma_alloc(rrpc
->dev
, GFP_KERNEL
,
842 if (!rqd
->ppa_list
) {
843 pr_err("rrpc: not able to allocate ppa list\n");
847 if (bio_rw(bio
) == WRITE
)
848 return rrpc_write_ppalist_rq(rrpc
, bio
, rqd
, flags
,
851 return rrpc_read_ppalist_rq(rrpc
, bio
, rqd
, flags
, npages
);
854 if (bio_rw(bio
) == WRITE
)
855 return rrpc_write_rq(rrpc
, bio
, rqd
, flags
);
857 return rrpc_read_rq(rrpc
, bio
, rqd
, flags
);
860 static int rrpc_submit_io(struct rrpc
*rrpc
, struct bio
*bio
,
861 struct nvm_rq
*rqd
, unsigned long flags
)
864 struct rrpc_rq
*rrq
= nvm_rq_to_pdu(rqd
);
865 uint8_t nr_pages
= rrpc_get_pages(bio
);
866 int bio_size
= bio_sectors(bio
) << 9;
868 if (bio_size
< rrpc
->dev
->sec_size
)
870 else if (bio_size
> rrpc
->dev
->max_rq_size
)
873 err
= rrpc_setup_rq(rrpc
, bio
, rqd
, flags
, nr_pages
);
879 rqd
->ins
= &rrpc
->instance
;
880 rqd
->nr_pages
= nr_pages
;
883 err
= nvm_submit_io(rrpc
->dev
, rqd
);
885 pr_err("rrpc: I/O submission failed: %d\n", err
);
887 if (!(flags
& NVM_IOTYPE_GC
)) {
888 rrpc_unlock_rq(rrpc
, rqd
);
889 if (rqd
->nr_pages
> 1)
890 nvm_dev_dma_free(rrpc
->dev
,
891 rqd
->ppa_list
, rqd
->dma_ppa_list
);
899 static blk_qc_t
rrpc_make_rq(struct request_queue
*q
, struct bio
*bio
)
901 struct rrpc
*rrpc
= q
->queuedata
;
905 if (bio
->bi_rw
& REQ_DISCARD
) {
906 rrpc_discard(rrpc
, bio
);
907 return BLK_QC_T_NONE
;
910 rqd
= mempool_alloc(rrpc
->rq_pool
, GFP_KERNEL
);
912 pr_err_ratelimited("rrpc: not able to queue bio.");
914 return BLK_QC_T_NONE
;
916 memset(rqd
, 0, sizeof(struct nvm_rq
));
918 err
= rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_NONE
);
921 return BLK_QC_T_NONE
;
929 spin_lock(&rrpc
->bio_lock
);
930 bio_list_add(&rrpc
->requeue_bios
, bio
);
931 spin_unlock(&rrpc
->bio_lock
);
932 queue_work(rrpc
->kgc_wq
, &rrpc
->ws_requeue
);
936 mempool_free(rqd
, rrpc
->rq_pool
);
937 return BLK_QC_T_NONE
;
940 static void rrpc_requeue(struct work_struct
*work
)
942 struct rrpc
*rrpc
= container_of(work
, struct rrpc
, ws_requeue
);
943 struct bio_list bios
;
946 bio_list_init(&bios
);
948 spin_lock(&rrpc
->bio_lock
);
949 bio_list_merge(&bios
, &rrpc
->requeue_bios
);
950 bio_list_init(&rrpc
->requeue_bios
);
951 spin_unlock(&rrpc
->bio_lock
);
953 while ((bio
= bio_list_pop(&bios
)))
954 rrpc_make_rq(rrpc
->disk
->queue
, bio
);
957 static void rrpc_gc_free(struct rrpc
*rrpc
)
959 struct rrpc_lun
*rlun
;
963 destroy_workqueue(rrpc
->krqd_wq
);
966 destroy_workqueue(rrpc
->kgc_wq
);
971 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
972 rlun
= &rrpc
->luns
[i
];
980 static int rrpc_gc_init(struct rrpc
*rrpc
)
982 rrpc
->krqd_wq
= alloc_workqueue("rrpc-lun", WQ_MEM_RECLAIM
|WQ_UNBOUND
,
987 rrpc
->kgc_wq
= alloc_workqueue("rrpc-bg", WQ_MEM_RECLAIM
, 1);
991 setup_timer(&rrpc
->gc_timer
, rrpc_gc_timer
, (unsigned long)rrpc
);
996 static void rrpc_map_free(struct rrpc
*rrpc
)
998 vfree(rrpc
->rev_trans_map
);
999 vfree(rrpc
->trans_map
);
1002 static int rrpc_l2p_update(u64 slba
, u32 nlb
, __le64
*entries
, void *private)
1004 struct rrpc
*rrpc
= (struct rrpc
*)private;
1005 struct nvm_dev
*dev
= rrpc
->dev
;
1006 struct rrpc_addr
*addr
= rrpc
->trans_map
+ slba
;
1007 struct rrpc_rev_addr
*raddr
= rrpc
->rev_trans_map
;
1008 sector_t max_pages
= dev
->total_pages
* (dev
->sec_size
>> 9);
1009 u64 elba
= slba
+ nlb
;
1012 if (unlikely(elba
> dev
->total_pages
)) {
1013 pr_err("nvm: L2P data from device is out of bounds!\n");
1017 for (i
= 0; i
< nlb
; i
++) {
1018 u64 pba
= le64_to_cpu(entries
[i
]);
1019 /* LNVM treats address-spaces as silos, LBA and PBA are
1020 * equally large and zero-indexed.
1022 if (unlikely(pba
>= max_pages
&& pba
!= U64_MAX
)) {
1023 pr_err("nvm: L2P data entry is out of bounds!\n");
1027 /* Address zero is a special one. The first page on a disk is
1028 * protected. As it often holds internal device boot
1035 raddr
[pba
].addr
= slba
+ i
;
1041 static int rrpc_map_init(struct rrpc
*rrpc
)
1043 struct nvm_dev
*dev
= rrpc
->dev
;
1047 rrpc
->trans_map
= vzalloc(sizeof(struct rrpc_addr
) * rrpc
->nr_pages
);
1048 if (!rrpc
->trans_map
)
1051 rrpc
->rev_trans_map
= vmalloc(sizeof(struct rrpc_rev_addr
)
1053 if (!rrpc
->rev_trans_map
)
1056 for (i
= 0; i
< rrpc
->nr_pages
; i
++) {
1057 struct rrpc_addr
*p
= &rrpc
->trans_map
[i
];
1058 struct rrpc_rev_addr
*r
= &rrpc
->rev_trans_map
[i
];
1060 p
->addr
= ADDR_EMPTY
;
1061 r
->addr
= ADDR_EMPTY
;
1064 if (!dev
->ops
->get_l2p_tbl
)
1067 /* Bring up the mapping table from device */
1068 ret
= dev
->ops
->get_l2p_tbl(dev
, 0, dev
->total_pages
,
1069 rrpc_l2p_update
, rrpc
);
1071 pr_err("nvm: rrpc: could not read L2P table.\n");
1079 /* Minimum pages needed within a lun */
1080 #define PAGE_POOL_SIZE 16
1081 #define ADDR_POOL_SIZE 64
1083 static int rrpc_core_init(struct rrpc
*rrpc
)
1085 down_write(&rrpc_lock
);
1086 if (!rrpc_gcb_cache
) {
1087 rrpc_gcb_cache
= kmem_cache_create("rrpc_gcb",
1088 sizeof(struct rrpc_block_gc
), 0, 0, NULL
);
1089 if (!rrpc_gcb_cache
) {
1090 up_write(&rrpc_lock
);
1094 rrpc_rq_cache
= kmem_cache_create("rrpc_rq",
1095 sizeof(struct nvm_rq
) + sizeof(struct rrpc_rq
),
1097 if (!rrpc_rq_cache
) {
1098 kmem_cache_destroy(rrpc_gcb_cache
);
1099 up_write(&rrpc_lock
);
1103 up_write(&rrpc_lock
);
1105 rrpc
->page_pool
= mempool_create_page_pool(PAGE_POOL_SIZE
, 0);
1106 if (!rrpc
->page_pool
)
1109 rrpc
->gcb_pool
= mempool_create_slab_pool(rrpc
->dev
->nr_luns
,
1111 if (!rrpc
->gcb_pool
)
1114 rrpc
->rq_pool
= mempool_create_slab_pool(64, rrpc_rq_cache
);
1118 spin_lock_init(&rrpc
->inflights
.lock
);
1119 INIT_LIST_HEAD(&rrpc
->inflights
.reqs
);
1124 static void rrpc_core_free(struct rrpc
*rrpc
)
1126 mempool_destroy(rrpc
->page_pool
);
1127 mempool_destroy(rrpc
->gcb_pool
);
1128 mempool_destroy(rrpc
->rq_pool
);
1131 static void rrpc_luns_free(struct rrpc
*rrpc
)
1136 static int rrpc_luns_init(struct rrpc
*rrpc
, int lun_begin
, int lun_end
)
1138 struct nvm_dev
*dev
= rrpc
->dev
;
1139 struct rrpc_lun
*rlun
;
1142 if (dev
->pgs_per_blk
> MAX_INVALID_PAGES_STORAGE
* BITS_PER_LONG
) {
1143 pr_err("rrpc: number of pages per block too high.");
1147 spin_lock_init(&rrpc
->rev_lock
);
1149 rrpc
->luns
= kcalloc(rrpc
->nr_luns
, sizeof(struct rrpc_lun
),
1155 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
1156 struct nvm_lun
*lun
= dev
->mt
->get_lun(dev
, lun_begin
+ i
);
1158 rlun
= &rrpc
->luns
[i
];
1161 INIT_LIST_HEAD(&rlun
->prio_list
);
1162 INIT_LIST_HEAD(&rlun
->open_list
);
1163 INIT_LIST_HEAD(&rlun
->closed_list
);
1165 INIT_WORK(&rlun
->ws_gc
, rrpc_lun_gc
);
1166 spin_lock_init(&rlun
->lock
);
1168 rrpc
->total_blocks
+= dev
->blks_per_lun
;
1169 rrpc
->nr_pages
+= dev
->sec_per_lun
;
1171 rlun
->blocks
= vzalloc(sizeof(struct rrpc_block
) *
1172 rrpc
->dev
->blks_per_lun
);
1176 for (j
= 0; j
< rrpc
->dev
->blks_per_lun
; j
++) {
1177 struct rrpc_block
*rblk
= &rlun
->blocks
[j
];
1178 struct nvm_block
*blk
= &lun
->blocks
[j
];
1182 INIT_LIST_HEAD(&rblk
->prio
);
1183 spin_lock_init(&rblk
->lock
);
1192 static void rrpc_free(struct rrpc
*rrpc
)
1195 rrpc_map_free(rrpc
);
1196 rrpc_core_free(rrpc
);
1197 rrpc_luns_free(rrpc
);
1202 static void rrpc_exit(void *private)
1204 struct rrpc
*rrpc
= private;
1206 del_timer(&rrpc
->gc_timer
);
1208 flush_workqueue(rrpc
->krqd_wq
);
1209 flush_workqueue(rrpc
->kgc_wq
);
1214 static sector_t
rrpc_capacity(void *private)
1216 struct rrpc
*rrpc
= private;
1217 struct nvm_dev
*dev
= rrpc
->dev
;
1218 sector_t reserved
, provisioned
;
1220 /* cur, gc, and two emergency blocks for each lun */
1221 reserved
= rrpc
->nr_luns
* dev
->max_pages_per_blk
* 4;
1222 provisioned
= rrpc
->nr_pages
- reserved
;
1224 if (reserved
> rrpc
->nr_pages
) {
1225 pr_err("rrpc: not enough space available to expose storage.\n");
1229 sector_div(provisioned
, 10);
1230 return provisioned
* 9 * NR_PHY_IN_LOG
;
1234 * Looks up the logical address from reverse trans map and check if its valid by
1235 * comparing the logical to physical address with the physical address.
1236 * Returns 0 on free, otherwise 1 if in use
1238 static void rrpc_block_map_update(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
1240 struct nvm_dev
*dev
= rrpc
->dev
;
1242 struct rrpc_addr
*laddr
;
1245 for (offset
= 0; offset
< dev
->pgs_per_blk
; offset
++) {
1246 paddr
= block_to_addr(rrpc
, rblk
) + offset
;
1248 pladdr
= rrpc
->rev_trans_map
[paddr
].addr
;
1249 if (pladdr
== ADDR_EMPTY
)
1252 laddr
= &rrpc
->trans_map
[pladdr
];
1254 if (paddr
== laddr
->addr
) {
1257 set_bit(offset
, rblk
->invalid_pages
);
1258 rblk
->nr_invalid_pages
++;
1263 static int rrpc_blocks_init(struct rrpc
*rrpc
)
1265 struct rrpc_lun
*rlun
;
1266 struct rrpc_block
*rblk
;
1267 int lun_iter
, blk_iter
;
1269 for (lun_iter
= 0; lun_iter
< rrpc
->nr_luns
; lun_iter
++) {
1270 rlun
= &rrpc
->luns
[lun_iter
];
1272 for (blk_iter
= 0; blk_iter
< rrpc
->dev
->blks_per_lun
;
1274 rblk
= &rlun
->blocks
[blk_iter
];
1275 rrpc_block_map_update(rrpc
, rblk
);
1282 static int rrpc_luns_configure(struct rrpc
*rrpc
)
1284 struct rrpc_lun
*rlun
;
1285 struct rrpc_block
*rblk
;
1288 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
1289 rlun
= &rrpc
->luns
[i
];
1291 rblk
= rrpc_get_blk(rrpc
, rlun
, 0);
1295 rrpc_set_lun_cur(rlun
, rblk
);
1297 /* Emergency gc block */
1298 rblk
= rrpc_get_blk(rrpc
, rlun
, 1);
1301 rlun
->gc_cur
= rblk
;
1306 rrpc_put_blks(rrpc
);
1310 static struct nvm_tgt_type tt_rrpc
;
1312 static void *rrpc_init(struct nvm_dev
*dev
, struct gendisk
*tdisk
,
1313 int lun_begin
, int lun_end
)
1315 struct request_queue
*bqueue
= dev
->q
;
1316 struct request_queue
*tqueue
= tdisk
->queue
;
1320 if (!(dev
->identity
.dom
& NVM_RSP_L2P
)) {
1321 pr_err("nvm: rrpc: device does not support l2p (%x)\n",
1323 return ERR_PTR(-EINVAL
);
1326 rrpc
= kzalloc(sizeof(struct rrpc
), GFP_KERNEL
);
1328 return ERR_PTR(-ENOMEM
);
1330 rrpc
->instance
.tt
= &tt_rrpc
;
1334 bio_list_init(&rrpc
->requeue_bios
);
1335 spin_lock_init(&rrpc
->bio_lock
);
1336 INIT_WORK(&rrpc
->ws_requeue
, rrpc_requeue
);
1338 rrpc
->nr_luns
= lun_end
- lun_begin
+ 1;
1340 /* simple round-robin strategy */
1341 atomic_set(&rrpc
->next_lun
, -1);
1343 ret
= rrpc_luns_init(rrpc
, lun_begin
, lun_end
);
1345 pr_err("nvm: rrpc: could not initialize luns\n");
1349 rrpc
->poffset
= dev
->sec_per_lun
* lun_begin
;
1350 rrpc
->lun_offset
= lun_begin
;
1352 ret
= rrpc_core_init(rrpc
);
1354 pr_err("nvm: rrpc: could not initialize core\n");
1358 ret
= rrpc_map_init(rrpc
);
1360 pr_err("nvm: rrpc: could not initialize maps\n");
1364 ret
= rrpc_blocks_init(rrpc
);
1366 pr_err("nvm: rrpc: could not initialize state for blocks\n");
1370 ret
= rrpc_luns_configure(rrpc
);
1372 pr_err("nvm: rrpc: not enough blocks available in LUNs.\n");
1376 ret
= rrpc_gc_init(rrpc
);
1378 pr_err("nvm: rrpc: could not initialize gc\n");
1382 /* inherit the size from the underlying device */
1383 blk_queue_logical_block_size(tqueue
, queue_physical_block_size(bqueue
));
1384 blk_queue_max_hw_sectors(tqueue
, queue_max_hw_sectors(bqueue
));
1386 pr_info("nvm: rrpc initialized with %u luns and %llu pages.\n",
1387 rrpc
->nr_luns
, (unsigned long long)rrpc
->nr_pages
);
1389 mod_timer(&rrpc
->gc_timer
, jiffies
+ msecs_to_jiffies(10));
1394 return ERR_PTR(ret
);
1397 /* round robin, page-based FTL, and cost-based GC */
1398 static struct nvm_tgt_type tt_rrpc
= {
1400 .version
= {1, 0, 0},
1402 .make_rq
= rrpc_make_rq
,
1403 .capacity
= rrpc_capacity
,
1404 .end_io
= rrpc_end_io
,
1410 static int __init
rrpc_module_init(void)
1412 return nvm_register_target(&tt_rrpc
);
1415 static void rrpc_module_exit(void)
1417 nvm_unregister_target(&tt_rrpc
);
1420 module_init(rrpc_module_init
);
1421 module_exit(rrpc_module_exit
);
1422 MODULE_LICENSE("GPL v2");
1423 MODULE_DESCRIPTION("Block-Device Target for Open-Channel SSDs");