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_block
*blk
;
183 struct rrpc_block
*rblk
;
185 blk
= nvm_get_blk(rrpc
->dev
, rlun
->parent
, flags
);
189 rblk
= &rlun
->blocks
[blk
->id
];
192 bitmap_zero(rblk
->invalid_pages
, rrpc
->dev
->pgs_per_blk
);
194 rblk
->nr_invalid_pages
= 0;
195 atomic_set(&rblk
->data_cmnt_size
, 0);
200 static void rrpc_put_blk(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
202 nvm_put_blk(rrpc
->dev
, rblk
->parent
);
205 static void rrpc_put_blks(struct rrpc
*rrpc
)
207 struct rrpc_lun
*rlun
;
210 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
211 rlun
= &rrpc
->luns
[i
];
213 rrpc_put_blk(rrpc
, rlun
->cur
);
215 rrpc_put_blk(rrpc
, rlun
->gc_cur
);
219 static struct rrpc_lun
*get_next_lun(struct rrpc
*rrpc
)
221 int next
= atomic_inc_return(&rrpc
->next_lun
);
223 return &rrpc
->luns
[next
% rrpc
->nr_luns
];
226 static void rrpc_gc_kick(struct rrpc
*rrpc
)
228 struct rrpc_lun
*rlun
;
231 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
232 rlun
= &rrpc
->luns
[i
];
233 queue_work(rrpc
->krqd_wq
, &rlun
->ws_gc
);
238 * timed GC every interval.
240 static void rrpc_gc_timer(unsigned long data
)
242 struct rrpc
*rrpc
= (struct rrpc
*)data
;
245 mod_timer(&rrpc
->gc_timer
, jiffies
+ msecs_to_jiffies(10));
248 static void rrpc_end_sync_bio(struct bio
*bio
)
250 struct completion
*waiting
= bio
->bi_private
;
253 pr_err("nvm: gc request failed (%u).\n", bio
->bi_error
);
259 * rrpc_move_valid_pages -- migrate live data off the block
260 * @rrpc: the 'rrpc' structure
261 * @block: the block from which to migrate live pages
264 * GC algorithms may call this function to migrate remaining live
265 * pages off the block prior to erasing it. This function blocks
266 * further execution until the operation is complete.
268 static int rrpc_move_valid_pages(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
270 struct request_queue
*q
= rrpc
->dev
->q
;
271 struct rrpc_rev_addr
*rev
;
276 int nr_pgs_per_blk
= rrpc
->dev
->pgs_per_blk
;
278 DECLARE_COMPLETION_ONSTACK(wait
);
280 if (bitmap_full(rblk
->invalid_pages
, nr_pgs_per_blk
))
283 bio
= bio_alloc(GFP_NOIO
, 1);
285 pr_err("nvm: could not alloc bio to gc\n");
289 page
= mempool_alloc(rrpc
->page_pool
, GFP_NOIO
);
293 while ((slot
= find_first_zero_bit(rblk
->invalid_pages
,
294 nr_pgs_per_blk
)) < nr_pgs_per_blk
) {
297 phys_addr
= (rblk
->parent
->id
* nr_pgs_per_blk
) + slot
;
300 spin_lock(&rrpc
->rev_lock
);
301 /* Get logical address from physical to logical table */
302 rev
= &rrpc
->rev_trans_map
[phys_addr
- rrpc
->poffset
];
303 /* already updated by previous regular write */
304 if (rev
->addr
== ADDR_EMPTY
) {
305 spin_unlock(&rrpc
->rev_lock
);
309 rqd
= rrpc_inflight_laddr_acquire(rrpc
, rev
->addr
, 1);
310 if (IS_ERR_OR_NULL(rqd
)) {
311 spin_unlock(&rrpc
->rev_lock
);
316 spin_unlock(&rrpc
->rev_lock
);
318 /* Perform read to do GC */
319 bio
->bi_iter
.bi_sector
= rrpc_get_sector(rev
->addr
);
321 bio
->bi_private
= &wait
;
322 bio
->bi_end_io
= rrpc_end_sync_bio
;
324 /* TODO: may fail when EXP_PG_SIZE > PAGE_SIZE */
325 bio_add_pc_page(q
, bio
, page
, RRPC_EXPOSED_PAGE_SIZE
, 0);
327 if (rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_GC
)) {
328 pr_err("rrpc: gc read failed.\n");
329 rrpc_inflight_laddr_release(rrpc
, rqd
);
332 wait_for_completion_io(&wait
);
335 reinit_completion(&wait
);
337 bio
->bi_iter
.bi_sector
= rrpc_get_sector(rev
->addr
);
339 bio
->bi_private
= &wait
;
340 bio
->bi_end_io
= rrpc_end_sync_bio
;
342 bio_add_pc_page(q
, bio
, page
, RRPC_EXPOSED_PAGE_SIZE
, 0);
344 /* turn the command around and write the data back to a new
347 if (rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_GC
)) {
348 pr_err("rrpc: gc write failed.\n");
349 rrpc_inflight_laddr_release(rrpc
, rqd
);
352 wait_for_completion_io(&wait
);
354 rrpc_inflight_laddr_release(rrpc
, rqd
);
360 mempool_free(page
, rrpc
->page_pool
);
363 if (!bitmap_full(rblk
->invalid_pages
, nr_pgs_per_blk
)) {
364 pr_err("nvm: failed to garbage collect block\n");
371 static void rrpc_block_gc(struct work_struct
*work
)
373 struct rrpc_block_gc
*gcb
= container_of(work
, struct rrpc_block_gc
,
375 struct rrpc
*rrpc
= gcb
->rrpc
;
376 struct rrpc_block
*rblk
= gcb
->rblk
;
377 struct nvm_dev
*dev
= rrpc
->dev
;
379 pr_debug("nvm: block '%lu' being reclaimed\n", rblk
->parent
->id
);
381 if (rrpc_move_valid_pages(rrpc
, rblk
))
384 nvm_erase_blk(dev
, rblk
->parent
);
385 rrpc_put_blk(rrpc
, rblk
);
387 mempool_free(gcb
, rrpc
->gcb_pool
);
390 /* the block with highest number of invalid pages, will be in the beginning
393 static struct rrpc_block
*rblock_max_invalid(struct rrpc_block
*ra
,
394 struct rrpc_block
*rb
)
396 if (ra
->nr_invalid_pages
== rb
->nr_invalid_pages
)
399 return (ra
->nr_invalid_pages
< rb
->nr_invalid_pages
) ? rb
: ra
;
402 /* linearly find the block with highest number of invalid pages
405 static struct rrpc_block
*block_prio_find_max(struct rrpc_lun
*rlun
)
407 struct list_head
*prio_list
= &rlun
->prio_list
;
408 struct rrpc_block
*rblock
, *max
;
410 BUG_ON(list_empty(prio_list
));
412 max
= list_first_entry(prio_list
, struct rrpc_block
, prio
);
413 list_for_each_entry(rblock
, prio_list
, prio
)
414 max
= rblock_max_invalid(max
, rblock
);
419 static void rrpc_lun_gc(struct work_struct
*work
)
421 struct rrpc_lun
*rlun
= container_of(work
, struct rrpc_lun
, ws_gc
);
422 struct rrpc
*rrpc
= rlun
->rrpc
;
423 struct nvm_lun
*lun
= rlun
->parent
;
424 struct rrpc_block_gc
*gcb
;
425 unsigned int nr_blocks_need
;
427 nr_blocks_need
= rrpc
->dev
->blks_per_lun
/ GC_LIMIT_INVERSE
;
429 if (nr_blocks_need
< rrpc
->nr_luns
)
430 nr_blocks_need
= rrpc
->nr_luns
;
432 spin_lock(&lun
->lock
);
433 while (nr_blocks_need
> lun
->nr_free_blocks
&&
434 !list_empty(&rlun
->prio_list
)) {
435 struct rrpc_block
*rblock
= block_prio_find_max(rlun
);
436 struct nvm_block
*block
= rblock
->parent
;
438 if (!rblock
->nr_invalid_pages
)
441 list_del_init(&rblock
->prio
);
443 BUG_ON(!block_is_full(rrpc
, rblock
));
445 pr_debug("rrpc: selected block '%lu' for GC\n", block
->id
);
447 gcb
= mempool_alloc(rrpc
->gcb_pool
, GFP_ATOMIC
);
453 INIT_WORK(&gcb
->ws_gc
, rrpc_block_gc
);
455 queue_work(rrpc
->kgc_wq
, &gcb
->ws_gc
);
459 spin_unlock(&lun
->lock
);
461 /* TODO: Hint that request queue can be started again */
464 static void rrpc_gc_queue(struct work_struct
*work
)
466 struct rrpc_block_gc
*gcb
= container_of(work
, struct rrpc_block_gc
,
468 struct rrpc
*rrpc
= gcb
->rrpc
;
469 struct rrpc_block
*rblk
= gcb
->rblk
;
470 struct nvm_lun
*lun
= rblk
->parent
->lun
;
471 struct rrpc_lun
*rlun
= &rrpc
->luns
[lun
->id
- rrpc
->lun_offset
];
473 spin_lock(&rlun
->lock
);
474 list_add_tail(&rblk
->prio
, &rlun
->prio_list
);
475 spin_unlock(&rlun
->lock
);
477 mempool_free(gcb
, rrpc
->gcb_pool
);
478 pr_debug("nvm: block '%lu' is full, allow GC (sched)\n",
482 static const struct block_device_operations rrpc_fops
= {
483 .owner
= THIS_MODULE
,
486 static struct rrpc_lun
*rrpc_get_lun_rr(struct rrpc
*rrpc
, int is_gc
)
489 struct rrpc_lun
*rlun
, *max_free
;
492 return get_next_lun(rrpc
);
494 /* during GC, we don't care about RR, instead we want to make
495 * sure that we maintain evenness between the block luns.
497 max_free
= &rrpc
->luns
[0];
498 /* prevent GC-ing lun from devouring pages of a lun with
499 * little free blocks. We don't take the lock as we only need an
502 rrpc_for_each_lun(rrpc
, rlun
, i
) {
503 if (rlun
->parent
->nr_free_blocks
>
504 max_free
->parent
->nr_free_blocks
)
511 static struct rrpc_addr
*rrpc_update_map(struct rrpc
*rrpc
, sector_t laddr
,
512 struct rrpc_block
*rblk
, u64 paddr
)
514 struct rrpc_addr
*gp
;
515 struct rrpc_rev_addr
*rev
;
517 BUG_ON(laddr
>= rrpc
->nr_pages
);
519 gp
= &rrpc
->trans_map
[laddr
];
520 spin_lock(&rrpc
->rev_lock
);
522 rrpc_page_invalidate(rrpc
, gp
);
527 rev
= &rrpc
->rev_trans_map
[gp
->addr
- rrpc
->poffset
];
529 spin_unlock(&rrpc
->rev_lock
);
534 static u64
rrpc_alloc_addr(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
536 u64 addr
= ADDR_EMPTY
;
538 spin_lock(&rblk
->lock
);
539 if (block_is_full(rrpc
, rblk
))
542 addr
= block_to_addr(rrpc
, rblk
) + rblk
->next_page
;
546 spin_unlock(&rblk
->lock
);
550 /* Simple round-robin Logical to physical address translation.
552 * Retrieve the mapping using the active append point. Then update the ap for
553 * the next write to the disk.
555 * Returns rrpc_addr with the physical address and block. Remember to return to
556 * rrpc->addr_cache when request is finished.
558 static struct rrpc_addr
*rrpc_map_page(struct rrpc
*rrpc
, sector_t laddr
,
561 struct rrpc_lun
*rlun
;
562 struct rrpc_block
*rblk
;
566 rlun
= rrpc_get_lun_rr(rrpc
, is_gc
);
569 if (!is_gc
&& lun
->nr_free_blocks
< rrpc
->nr_luns
* 4)
572 spin_lock(&rlun
->lock
);
576 paddr
= rrpc_alloc_addr(rrpc
, rblk
);
578 if (paddr
== ADDR_EMPTY
) {
579 rblk
= rrpc_get_blk(rrpc
, rlun
, 0);
581 rrpc_set_lun_cur(rlun
, rblk
);
586 /* retry from emergency gc block */
587 paddr
= rrpc_alloc_addr(rrpc
, rlun
->gc_cur
);
588 if (paddr
== ADDR_EMPTY
) {
589 rblk
= rrpc_get_blk(rrpc
, rlun
, 1);
591 pr_err("rrpc: no more blocks");
596 paddr
= rrpc_alloc_addr(rrpc
, rlun
->gc_cur
);
602 spin_unlock(&rlun
->lock
);
603 return rrpc_update_map(rrpc
, laddr
, rblk
, paddr
);
605 spin_unlock(&rlun
->lock
);
609 static void rrpc_run_gc(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
611 struct rrpc_block_gc
*gcb
;
613 gcb
= mempool_alloc(rrpc
->gcb_pool
, GFP_ATOMIC
);
615 pr_err("rrpc: unable to queue block for gc.");
622 INIT_WORK(&gcb
->ws_gc
, rrpc_gc_queue
);
623 queue_work(rrpc
->kgc_wq
, &gcb
->ws_gc
);
626 static void rrpc_end_io_write(struct rrpc
*rrpc
, struct rrpc_rq
*rrqd
,
627 sector_t laddr
, uint8_t npages
)
630 struct rrpc_block
*rblk
;
634 for (i
= 0; i
< npages
; i
++) {
635 p
= &rrpc
->trans_map
[laddr
+ i
];
637 lun
= rblk
->parent
->lun
;
639 cmnt_size
= atomic_inc_return(&rblk
->data_cmnt_size
);
640 if (unlikely(cmnt_size
== rrpc
->dev
->pgs_per_blk
))
641 rrpc_run_gc(rrpc
, rblk
);
645 static int rrpc_end_io(struct nvm_rq
*rqd
, int error
)
647 struct rrpc
*rrpc
= container_of(rqd
->ins
, struct rrpc
, instance
);
648 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
649 uint8_t npages
= rqd
->nr_pages
;
650 sector_t laddr
= rrpc_get_laddr(rqd
->bio
) - npages
;
652 if (bio_data_dir(rqd
->bio
) == WRITE
)
653 rrpc_end_io_write(rrpc
, rrqd
, laddr
, npages
);
657 if (rrqd
->flags
& NVM_IOTYPE_GC
)
660 rrpc_unlock_rq(rrpc
, rqd
);
663 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
665 nvm_dev_dma_free(rrpc
->dev
, rqd
->metadata
, rqd
->dma_metadata
);
667 mempool_free(rqd
, rrpc
->rq_pool
);
672 static int rrpc_read_ppalist_rq(struct rrpc
*rrpc
, struct bio
*bio
,
673 struct nvm_rq
*rqd
, unsigned long flags
, int npages
)
675 struct rrpc_inflight_rq
*r
= rrpc_get_inflight_rq(rqd
);
676 struct rrpc_addr
*gp
;
677 sector_t laddr
= rrpc_get_laddr(bio
);
678 int is_gc
= flags
& NVM_IOTYPE_GC
;
681 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
)) {
682 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
683 return NVM_IO_REQUEUE
;
686 for (i
= 0; i
< npages
; i
++) {
687 /* We assume that mapping occurs at 4KB granularity */
688 BUG_ON(!(laddr
+ i
>= 0 && laddr
+ i
< rrpc
->nr_pages
));
689 gp
= &rrpc
->trans_map
[laddr
+ i
];
692 rqd
->ppa_list
[i
] = rrpc_ppa_to_gaddr(rrpc
->dev
,
696 rrpc_unlock_laddr(rrpc
, r
);
697 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
,
703 rqd
->opcode
= NVM_OP_HBREAD
;
708 static int rrpc_read_rq(struct rrpc
*rrpc
, struct bio
*bio
, struct nvm_rq
*rqd
,
711 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
712 int is_gc
= flags
& NVM_IOTYPE_GC
;
713 sector_t laddr
= rrpc_get_laddr(bio
);
714 struct rrpc_addr
*gp
;
716 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
))
717 return NVM_IO_REQUEUE
;
719 BUG_ON(!(laddr
>= 0 && laddr
< rrpc
->nr_pages
));
720 gp
= &rrpc
->trans_map
[laddr
];
723 rqd
->ppa_addr
= rrpc_ppa_to_gaddr(rrpc
->dev
, gp
->addr
);
726 rrpc_unlock_rq(rrpc
, rqd
);
730 rqd
->opcode
= NVM_OP_HBREAD
;
736 static int rrpc_write_ppalist_rq(struct rrpc
*rrpc
, struct bio
*bio
,
737 struct nvm_rq
*rqd
, unsigned long flags
, int npages
)
739 struct rrpc_inflight_rq
*r
= rrpc_get_inflight_rq(rqd
);
741 sector_t laddr
= rrpc_get_laddr(bio
);
742 int is_gc
= flags
& NVM_IOTYPE_GC
;
745 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
)) {
746 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
747 return NVM_IO_REQUEUE
;
750 for (i
= 0; i
< npages
; i
++) {
751 /* We assume that mapping occurs at 4KB granularity */
752 p
= rrpc_map_page(rrpc
, laddr
+ i
, is_gc
);
755 rrpc_unlock_laddr(rrpc
, r
);
756 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
,
759 return NVM_IO_REQUEUE
;
762 rqd
->ppa_list
[i
] = rrpc_ppa_to_gaddr(rrpc
->dev
,
766 rqd
->opcode
= NVM_OP_HBWRITE
;
771 static int rrpc_write_rq(struct rrpc
*rrpc
, struct bio
*bio
,
772 struct nvm_rq
*rqd
, unsigned long flags
)
774 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
776 int is_gc
= flags
& NVM_IOTYPE_GC
;
777 sector_t laddr
= rrpc_get_laddr(bio
);
779 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
))
780 return NVM_IO_REQUEUE
;
782 p
= rrpc_map_page(rrpc
, laddr
, is_gc
);
785 rrpc_unlock_rq(rrpc
, rqd
);
787 return NVM_IO_REQUEUE
;
790 rqd
->ppa_addr
= rrpc_ppa_to_gaddr(rrpc
->dev
, p
->addr
);
791 rqd
->opcode
= NVM_OP_HBWRITE
;
797 static int rrpc_setup_rq(struct rrpc
*rrpc
, struct bio
*bio
,
798 struct nvm_rq
*rqd
, unsigned long flags
, uint8_t npages
)
801 rqd
->ppa_list
= nvm_dev_dma_alloc(rrpc
->dev
, GFP_KERNEL
,
803 if (!rqd
->ppa_list
) {
804 pr_err("rrpc: not able to allocate ppa list\n");
808 if (bio_rw(bio
) == WRITE
)
809 return rrpc_write_ppalist_rq(rrpc
, bio
, rqd
, flags
,
812 return rrpc_read_ppalist_rq(rrpc
, bio
, rqd
, flags
, npages
);
815 if (bio_rw(bio
) == WRITE
)
816 return rrpc_write_rq(rrpc
, bio
, rqd
, flags
);
818 return rrpc_read_rq(rrpc
, bio
, rqd
, flags
);
821 static int rrpc_submit_io(struct rrpc
*rrpc
, struct bio
*bio
,
822 struct nvm_rq
*rqd
, unsigned long flags
)
825 struct rrpc_rq
*rrq
= nvm_rq_to_pdu(rqd
);
826 uint8_t nr_pages
= rrpc_get_pages(bio
);
827 int bio_size
= bio_sectors(bio
) << 9;
829 if (bio_size
< rrpc
->dev
->sec_size
)
831 else if (bio_size
> rrpc
->dev
->max_rq_size
)
834 err
= rrpc_setup_rq(rrpc
, bio
, rqd
, flags
, nr_pages
);
840 rqd
->ins
= &rrpc
->instance
;
841 rqd
->nr_pages
= nr_pages
;
844 err
= nvm_submit_io(rrpc
->dev
, rqd
);
846 pr_err("rrpc: I/O submission failed: %d\n", err
);
848 if (!(flags
& NVM_IOTYPE_GC
)) {
849 rrpc_unlock_rq(rrpc
, rqd
);
850 if (rqd
->nr_pages
> 1)
851 nvm_dev_dma_free(rrpc
->dev
,
852 rqd
->ppa_list
, rqd
->dma_ppa_list
);
860 static blk_qc_t
rrpc_make_rq(struct request_queue
*q
, struct bio
*bio
)
862 struct rrpc
*rrpc
= q
->queuedata
;
866 if (bio
->bi_rw
& REQ_DISCARD
) {
867 rrpc_discard(rrpc
, bio
);
868 return BLK_QC_T_NONE
;
871 rqd
= mempool_alloc(rrpc
->rq_pool
, GFP_KERNEL
);
873 pr_err_ratelimited("rrpc: not able to queue bio.");
875 return BLK_QC_T_NONE
;
877 memset(rqd
, 0, sizeof(struct nvm_rq
));
879 err
= rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_NONE
);
882 return BLK_QC_T_NONE
;
890 spin_lock(&rrpc
->bio_lock
);
891 bio_list_add(&rrpc
->requeue_bios
, bio
);
892 spin_unlock(&rrpc
->bio_lock
);
893 queue_work(rrpc
->kgc_wq
, &rrpc
->ws_requeue
);
897 mempool_free(rqd
, rrpc
->rq_pool
);
898 return BLK_QC_T_NONE
;
901 static void rrpc_requeue(struct work_struct
*work
)
903 struct rrpc
*rrpc
= container_of(work
, struct rrpc
, ws_requeue
);
904 struct bio_list bios
;
907 bio_list_init(&bios
);
909 spin_lock(&rrpc
->bio_lock
);
910 bio_list_merge(&bios
, &rrpc
->requeue_bios
);
911 bio_list_init(&rrpc
->requeue_bios
);
912 spin_unlock(&rrpc
->bio_lock
);
914 while ((bio
= bio_list_pop(&bios
)))
915 rrpc_make_rq(rrpc
->disk
->queue
, bio
);
918 static void rrpc_gc_free(struct rrpc
*rrpc
)
920 struct rrpc_lun
*rlun
;
924 destroy_workqueue(rrpc
->krqd_wq
);
927 destroy_workqueue(rrpc
->kgc_wq
);
932 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
933 rlun
= &rrpc
->luns
[i
];
941 static int rrpc_gc_init(struct rrpc
*rrpc
)
943 rrpc
->krqd_wq
= alloc_workqueue("rrpc-lun", WQ_MEM_RECLAIM
|WQ_UNBOUND
,
948 rrpc
->kgc_wq
= alloc_workqueue("rrpc-bg", WQ_MEM_RECLAIM
, 1);
952 setup_timer(&rrpc
->gc_timer
, rrpc_gc_timer
, (unsigned long)rrpc
);
957 static void rrpc_map_free(struct rrpc
*rrpc
)
959 vfree(rrpc
->rev_trans_map
);
960 vfree(rrpc
->trans_map
);
963 static int rrpc_l2p_update(u64 slba
, u32 nlb
, __le64
*entries
, void *private)
965 struct rrpc
*rrpc
= (struct rrpc
*)private;
966 struct nvm_dev
*dev
= rrpc
->dev
;
967 struct rrpc_addr
*addr
= rrpc
->trans_map
+ slba
;
968 struct rrpc_rev_addr
*raddr
= rrpc
->rev_trans_map
;
969 sector_t max_pages
= dev
->total_pages
* (dev
->sec_size
>> 9);
970 u64 elba
= slba
+ nlb
;
973 if (unlikely(elba
> dev
->total_pages
)) {
974 pr_err("nvm: L2P data from device is out of bounds!\n");
978 for (i
= 0; i
< nlb
; i
++) {
979 u64 pba
= le64_to_cpu(entries
[i
]);
980 /* LNVM treats address-spaces as silos, LBA and PBA are
981 * equally large and zero-indexed.
983 if (unlikely(pba
>= max_pages
&& pba
!= U64_MAX
)) {
984 pr_err("nvm: L2P data entry is out of bounds!\n");
988 /* Address zero is a special one. The first page on a disk is
989 * protected. As it often holds internal device boot
996 raddr
[pba
].addr
= slba
+ i
;
1002 static int rrpc_map_init(struct rrpc
*rrpc
)
1004 struct nvm_dev
*dev
= rrpc
->dev
;
1008 rrpc
->trans_map
= vzalloc(sizeof(struct rrpc_addr
) * rrpc
->nr_pages
);
1009 if (!rrpc
->trans_map
)
1012 rrpc
->rev_trans_map
= vmalloc(sizeof(struct rrpc_rev_addr
)
1014 if (!rrpc
->rev_trans_map
)
1017 for (i
= 0; i
< rrpc
->nr_pages
; i
++) {
1018 struct rrpc_addr
*p
= &rrpc
->trans_map
[i
];
1019 struct rrpc_rev_addr
*r
= &rrpc
->rev_trans_map
[i
];
1021 p
->addr
= ADDR_EMPTY
;
1022 r
->addr
= ADDR_EMPTY
;
1025 if (!dev
->ops
->get_l2p_tbl
)
1028 /* Bring up the mapping table from device */
1029 ret
= dev
->ops
->get_l2p_tbl(dev
, 0, dev
->total_pages
,
1030 rrpc_l2p_update
, rrpc
);
1032 pr_err("nvm: rrpc: could not read L2P table.\n");
1040 /* Minimum pages needed within a lun */
1041 #define PAGE_POOL_SIZE 16
1042 #define ADDR_POOL_SIZE 64
1044 static int rrpc_core_init(struct rrpc
*rrpc
)
1046 down_write(&rrpc_lock
);
1047 if (!rrpc_gcb_cache
) {
1048 rrpc_gcb_cache
= kmem_cache_create("rrpc_gcb",
1049 sizeof(struct rrpc_block_gc
), 0, 0, NULL
);
1050 if (!rrpc_gcb_cache
) {
1051 up_write(&rrpc_lock
);
1055 rrpc_rq_cache
= kmem_cache_create("rrpc_rq",
1056 sizeof(struct nvm_rq
) + sizeof(struct rrpc_rq
),
1058 if (!rrpc_rq_cache
) {
1059 kmem_cache_destroy(rrpc_gcb_cache
);
1060 up_write(&rrpc_lock
);
1064 up_write(&rrpc_lock
);
1066 rrpc
->page_pool
= mempool_create_page_pool(PAGE_POOL_SIZE
, 0);
1067 if (!rrpc
->page_pool
)
1070 rrpc
->gcb_pool
= mempool_create_slab_pool(rrpc
->dev
->nr_luns
,
1072 if (!rrpc
->gcb_pool
)
1075 rrpc
->rq_pool
= mempool_create_slab_pool(64, rrpc_rq_cache
);
1079 spin_lock_init(&rrpc
->inflights
.lock
);
1080 INIT_LIST_HEAD(&rrpc
->inflights
.reqs
);
1085 static void rrpc_core_free(struct rrpc
*rrpc
)
1087 mempool_destroy(rrpc
->page_pool
);
1088 mempool_destroy(rrpc
->gcb_pool
);
1089 mempool_destroy(rrpc
->rq_pool
);
1092 static void rrpc_luns_free(struct rrpc
*rrpc
)
1097 static int rrpc_luns_init(struct rrpc
*rrpc
, int lun_begin
, int lun_end
)
1099 struct nvm_dev
*dev
= rrpc
->dev
;
1100 struct rrpc_lun
*rlun
;
1103 spin_lock_init(&rrpc
->rev_lock
);
1105 rrpc
->luns
= kcalloc(rrpc
->nr_luns
, sizeof(struct rrpc_lun
),
1111 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
1112 struct nvm_lun
*lun
= dev
->mt
->get_lun(dev
, lun_begin
+ i
);
1114 if (dev
->pgs_per_blk
>
1115 MAX_INVALID_PAGES_STORAGE
* BITS_PER_LONG
) {
1116 pr_err("rrpc: number of pages per block too high.");
1120 rlun
= &rrpc
->luns
[i
];
1123 INIT_LIST_HEAD(&rlun
->prio_list
);
1124 INIT_WORK(&rlun
->ws_gc
, rrpc_lun_gc
);
1125 spin_lock_init(&rlun
->lock
);
1127 rrpc
->total_blocks
+= dev
->blks_per_lun
;
1128 rrpc
->nr_pages
+= dev
->sec_per_lun
;
1130 rlun
->blocks
= vzalloc(sizeof(struct rrpc_block
) *
1131 rrpc
->dev
->blks_per_lun
);
1135 for (j
= 0; j
< rrpc
->dev
->blks_per_lun
; j
++) {
1136 struct rrpc_block
*rblk
= &rlun
->blocks
[j
];
1137 struct nvm_block
*blk
= &lun
->blocks
[j
];
1140 INIT_LIST_HEAD(&rblk
->prio
);
1141 spin_lock_init(&rblk
->lock
);
1150 static void rrpc_free(struct rrpc
*rrpc
)
1153 rrpc_map_free(rrpc
);
1154 rrpc_core_free(rrpc
);
1155 rrpc_luns_free(rrpc
);
1160 static void rrpc_exit(void *private)
1162 struct rrpc
*rrpc
= private;
1164 del_timer(&rrpc
->gc_timer
);
1166 flush_workqueue(rrpc
->krqd_wq
);
1167 flush_workqueue(rrpc
->kgc_wq
);
1172 static sector_t
rrpc_capacity(void *private)
1174 struct rrpc
*rrpc
= private;
1175 struct nvm_dev
*dev
= rrpc
->dev
;
1176 sector_t reserved
, provisioned
;
1178 /* cur, gc, and two emergency blocks for each lun */
1179 reserved
= rrpc
->nr_luns
* dev
->max_pages_per_blk
* 4;
1180 provisioned
= rrpc
->nr_pages
- reserved
;
1182 if (reserved
> rrpc
->nr_pages
) {
1183 pr_err("rrpc: not enough space available to expose storage.\n");
1187 sector_div(provisioned
, 10);
1188 return provisioned
* 9 * NR_PHY_IN_LOG
;
1192 * Looks up the logical address from reverse trans map and check if its valid by
1193 * comparing the logical to physical address with the physical address.
1194 * Returns 0 on free, otherwise 1 if in use
1196 static void rrpc_block_map_update(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
1198 struct nvm_dev
*dev
= rrpc
->dev
;
1200 struct rrpc_addr
*laddr
;
1203 for (offset
= 0; offset
< dev
->pgs_per_blk
; offset
++) {
1204 paddr
= block_to_addr(rrpc
, rblk
) + offset
;
1206 pladdr
= rrpc
->rev_trans_map
[paddr
].addr
;
1207 if (pladdr
== ADDR_EMPTY
)
1210 laddr
= &rrpc
->trans_map
[pladdr
];
1212 if (paddr
== laddr
->addr
) {
1215 set_bit(offset
, rblk
->invalid_pages
);
1216 rblk
->nr_invalid_pages
++;
1221 static int rrpc_blocks_init(struct rrpc
*rrpc
)
1223 struct rrpc_lun
*rlun
;
1224 struct rrpc_block
*rblk
;
1225 int lun_iter
, blk_iter
;
1227 for (lun_iter
= 0; lun_iter
< rrpc
->nr_luns
; lun_iter
++) {
1228 rlun
= &rrpc
->luns
[lun_iter
];
1230 for (blk_iter
= 0; blk_iter
< rrpc
->dev
->blks_per_lun
;
1232 rblk
= &rlun
->blocks
[blk_iter
];
1233 rrpc_block_map_update(rrpc
, rblk
);
1240 static int rrpc_luns_configure(struct rrpc
*rrpc
)
1242 struct rrpc_lun
*rlun
;
1243 struct rrpc_block
*rblk
;
1246 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
1247 rlun
= &rrpc
->luns
[i
];
1249 rblk
= rrpc_get_blk(rrpc
, rlun
, 0);
1253 rrpc_set_lun_cur(rlun
, rblk
);
1255 /* Emergency gc block */
1256 rblk
= rrpc_get_blk(rrpc
, rlun
, 1);
1259 rlun
->gc_cur
= rblk
;
1264 rrpc_put_blks(rrpc
);
1268 static struct nvm_tgt_type tt_rrpc
;
1270 static void *rrpc_init(struct nvm_dev
*dev
, struct gendisk
*tdisk
,
1271 int lun_begin
, int lun_end
)
1273 struct request_queue
*bqueue
= dev
->q
;
1274 struct request_queue
*tqueue
= tdisk
->queue
;
1278 if (!(dev
->identity
.dom
& NVM_RSP_L2P
)) {
1279 pr_err("nvm: rrpc: device does not support l2p (%x)\n",
1281 return ERR_PTR(-EINVAL
);
1284 rrpc
= kzalloc(sizeof(struct rrpc
), GFP_KERNEL
);
1286 return ERR_PTR(-ENOMEM
);
1288 rrpc
->instance
.tt
= &tt_rrpc
;
1292 bio_list_init(&rrpc
->requeue_bios
);
1293 spin_lock_init(&rrpc
->bio_lock
);
1294 INIT_WORK(&rrpc
->ws_requeue
, rrpc_requeue
);
1296 rrpc
->nr_luns
= lun_end
- lun_begin
+ 1;
1298 /* simple round-robin strategy */
1299 atomic_set(&rrpc
->next_lun
, -1);
1301 ret
= rrpc_luns_init(rrpc
, lun_begin
, lun_end
);
1303 pr_err("nvm: rrpc: could not initialize luns\n");
1307 rrpc
->poffset
= dev
->sec_per_lun
* lun_begin
;
1308 rrpc
->lun_offset
= lun_begin
;
1310 ret
= rrpc_core_init(rrpc
);
1312 pr_err("nvm: rrpc: could not initialize core\n");
1316 ret
= rrpc_map_init(rrpc
);
1318 pr_err("nvm: rrpc: could not initialize maps\n");
1322 ret
= rrpc_blocks_init(rrpc
);
1324 pr_err("nvm: rrpc: could not initialize state for blocks\n");
1328 ret
= rrpc_luns_configure(rrpc
);
1330 pr_err("nvm: rrpc: not enough blocks available in LUNs.\n");
1334 ret
= rrpc_gc_init(rrpc
);
1336 pr_err("nvm: rrpc: could not initialize gc\n");
1340 /* inherit the size from the underlying device */
1341 blk_queue_logical_block_size(tqueue
, queue_physical_block_size(bqueue
));
1342 blk_queue_max_hw_sectors(tqueue
, queue_max_hw_sectors(bqueue
));
1344 pr_info("nvm: rrpc initialized with %u luns and %llu pages.\n",
1345 rrpc
->nr_luns
, (unsigned long long)rrpc
->nr_pages
);
1347 mod_timer(&rrpc
->gc_timer
, jiffies
+ msecs_to_jiffies(10));
1352 return ERR_PTR(ret
);
1355 /* round robin, page-based FTL, and cost-based GC */
1356 static struct nvm_tgt_type tt_rrpc
= {
1358 .version
= {1, 0, 0},
1360 .make_rq
= rrpc_make_rq
,
1361 .capacity
= rrpc_capacity
,
1362 .end_io
= rrpc_end_io
,
1368 static int __init
rrpc_module_init(void)
1370 return nvm_register_target(&tt_rrpc
);
1373 static void rrpc_module_exit(void)
1375 nvm_unregister_target(&tt_rrpc
);
1378 module_init(rrpc_module_init
);
1379 module_exit(rrpc_module_exit
);
1380 MODULE_LICENSE("GPL v2");
1381 MODULE_DESCRIPTION("Block-Device Target for Open-Channel SSDs");