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
);
334 rrpc_inflight_laddr_release(rrpc
, rqd
);
339 reinit_completion(&wait
);
341 bio
->bi_iter
.bi_sector
= rrpc_get_sector(rev
->addr
);
343 bio
->bi_private
= &wait
;
344 bio
->bi_end_io
= rrpc_end_sync_bio
;
346 bio_add_pc_page(q
, bio
, page
, RRPC_EXPOSED_PAGE_SIZE
, 0);
348 /* turn the command around and write the data back to a new
351 if (rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_GC
)) {
352 pr_err("rrpc: gc write failed.\n");
353 rrpc_inflight_laddr_release(rrpc
, rqd
);
356 wait_for_completion_io(&wait
);
358 rrpc_inflight_laddr_release(rrpc
, rqd
);
366 mempool_free(page
, rrpc
->page_pool
);
369 if (!bitmap_full(rblk
->invalid_pages
, nr_pgs_per_blk
)) {
370 pr_err("nvm: failed to garbage collect block\n");
377 static void rrpc_block_gc(struct work_struct
*work
)
379 struct rrpc_block_gc
*gcb
= container_of(work
, struct rrpc_block_gc
,
381 struct rrpc
*rrpc
= gcb
->rrpc
;
382 struct rrpc_block
*rblk
= gcb
->rblk
;
383 struct nvm_dev
*dev
= rrpc
->dev
;
384 struct nvm_lun
*lun
= rblk
->parent
->lun
;
385 struct rrpc_lun
*rlun
= &rrpc
->luns
[lun
->id
- rrpc
->lun_offset
];
387 mempool_free(gcb
, rrpc
->gcb_pool
);
388 pr_debug("nvm: block '%lu' being reclaimed\n", rblk
->parent
->id
);
390 if (rrpc_move_valid_pages(rrpc
, rblk
))
393 if (nvm_erase_blk(dev
, rblk
->parent
))
396 rrpc_put_blk(rrpc
, rblk
);
401 spin_lock(&rlun
->lock
);
402 list_add_tail(&rblk
->prio
, &rlun
->prio_list
);
403 spin_unlock(&rlun
->lock
);
406 /* the block with highest number of invalid pages, will be in the beginning
409 static struct rrpc_block
*rblock_max_invalid(struct rrpc_block
*ra
,
410 struct rrpc_block
*rb
)
412 if (ra
->nr_invalid_pages
== rb
->nr_invalid_pages
)
415 return (ra
->nr_invalid_pages
< rb
->nr_invalid_pages
) ? rb
: ra
;
418 /* linearly find the block with highest number of invalid pages
421 static struct rrpc_block
*block_prio_find_max(struct rrpc_lun
*rlun
)
423 struct list_head
*prio_list
= &rlun
->prio_list
;
424 struct rrpc_block
*rblock
, *max
;
426 BUG_ON(list_empty(prio_list
));
428 max
= list_first_entry(prio_list
, struct rrpc_block
, prio
);
429 list_for_each_entry(rblock
, prio_list
, prio
)
430 max
= rblock_max_invalid(max
, rblock
);
435 static void rrpc_lun_gc(struct work_struct
*work
)
437 struct rrpc_lun
*rlun
= container_of(work
, struct rrpc_lun
, ws_gc
);
438 struct rrpc
*rrpc
= rlun
->rrpc
;
439 struct nvm_lun
*lun
= rlun
->parent
;
440 struct rrpc_block_gc
*gcb
;
441 unsigned int nr_blocks_need
;
443 nr_blocks_need
= rrpc
->dev
->blks_per_lun
/ GC_LIMIT_INVERSE
;
445 if (nr_blocks_need
< rrpc
->nr_luns
)
446 nr_blocks_need
= rrpc
->nr_luns
;
448 spin_lock(&lun
->lock
);
449 while (nr_blocks_need
> lun
->nr_free_blocks
&&
450 !list_empty(&rlun
->prio_list
)) {
451 struct rrpc_block
*rblock
= block_prio_find_max(rlun
);
452 struct nvm_block
*block
= rblock
->parent
;
454 if (!rblock
->nr_invalid_pages
)
457 list_del_init(&rblock
->prio
);
459 BUG_ON(!block_is_full(rrpc
, rblock
));
461 pr_debug("rrpc: selected block '%lu' for GC\n", block
->id
);
463 gcb
= mempool_alloc(rrpc
->gcb_pool
, GFP_ATOMIC
);
469 INIT_WORK(&gcb
->ws_gc
, rrpc_block_gc
);
471 queue_work(rrpc
->kgc_wq
, &gcb
->ws_gc
);
475 spin_unlock(&lun
->lock
);
477 /* TODO: Hint that request queue can be started again */
480 static void rrpc_gc_queue(struct work_struct
*work
)
482 struct rrpc_block_gc
*gcb
= container_of(work
, struct rrpc_block_gc
,
484 struct rrpc
*rrpc
= gcb
->rrpc
;
485 struct rrpc_block
*rblk
= gcb
->rblk
;
486 struct nvm_lun
*lun
= rblk
->parent
->lun
;
487 struct rrpc_lun
*rlun
= &rrpc
->luns
[lun
->id
- rrpc
->lun_offset
];
489 spin_lock(&rlun
->lock
);
490 list_add_tail(&rblk
->prio
, &rlun
->prio_list
);
491 spin_unlock(&rlun
->lock
);
493 mempool_free(gcb
, rrpc
->gcb_pool
);
494 pr_debug("nvm: block '%lu' is full, allow GC (sched)\n",
498 static const struct block_device_operations rrpc_fops
= {
499 .owner
= THIS_MODULE
,
502 static struct rrpc_lun
*rrpc_get_lun_rr(struct rrpc
*rrpc
, int is_gc
)
505 struct rrpc_lun
*rlun
, *max_free
;
508 return get_next_lun(rrpc
);
510 /* during GC, we don't care about RR, instead we want to make
511 * sure that we maintain evenness between the block luns.
513 max_free
= &rrpc
->luns
[0];
514 /* prevent GC-ing lun from devouring pages of a lun with
515 * little free blocks. We don't take the lock as we only need an
518 rrpc_for_each_lun(rrpc
, rlun
, i
) {
519 if (rlun
->parent
->nr_free_blocks
>
520 max_free
->parent
->nr_free_blocks
)
527 static struct rrpc_addr
*rrpc_update_map(struct rrpc
*rrpc
, sector_t laddr
,
528 struct rrpc_block
*rblk
, u64 paddr
)
530 struct rrpc_addr
*gp
;
531 struct rrpc_rev_addr
*rev
;
533 BUG_ON(laddr
>= rrpc
->nr_pages
);
535 gp
= &rrpc
->trans_map
[laddr
];
536 spin_lock(&rrpc
->rev_lock
);
538 rrpc_page_invalidate(rrpc
, gp
);
543 rev
= &rrpc
->rev_trans_map
[gp
->addr
- rrpc
->poffset
];
545 spin_unlock(&rrpc
->rev_lock
);
550 static u64
rrpc_alloc_addr(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
552 u64 addr
= ADDR_EMPTY
;
554 spin_lock(&rblk
->lock
);
555 if (block_is_full(rrpc
, rblk
))
558 addr
= block_to_addr(rrpc
, rblk
) + rblk
->next_page
;
562 spin_unlock(&rblk
->lock
);
566 /* Simple round-robin Logical to physical address translation.
568 * Retrieve the mapping using the active append point. Then update the ap for
569 * the next write to the disk.
571 * Returns rrpc_addr with the physical address and block. Remember to return to
572 * rrpc->addr_cache when request is finished.
574 static struct rrpc_addr
*rrpc_map_page(struct rrpc
*rrpc
, sector_t laddr
,
577 struct rrpc_lun
*rlun
;
578 struct rrpc_block
*rblk
;
582 rlun
= rrpc_get_lun_rr(rrpc
, is_gc
);
585 if (!is_gc
&& lun
->nr_free_blocks
< rrpc
->nr_luns
* 4)
588 spin_lock(&rlun
->lock
);
592 paddr
= rrpc_alloc_addr(rrpc
, rblk
);
594 if (paddr
== ADDR_EMPTY
) {
595 rblk
= rrpc_get_blk(rrpc
, rlun
, 0);
597 rrpc_set_lun_cur(rlun
, rblk
);
602 /* retry from emergency gc block */
603 paddr
= rrpc_alloc_addr(rrpc
, rlun
->gc_cur
);
604 if (paddr
== ADDR_EMPTY
) {
605 rblk
= rrpc_get_blk(rrpc
, rlun
, 1);
607 pr_err("rrpc: no more blocks");
612 paddr
= rrpc_alloc_addr(rrpc
, rlun
->gc_cur
);
618 spin_unlock(&rlun
->lock
);
619 return rrpc_update_map(rrpc
, laddr
, rblk
, paddr
);
621 spin_unlock(&rlun
->lock
);
625 static void rrpc_run_gc(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
627 struct rrpc_block_gc
*gcb
;
629 gcb
= mempool_alloc(rrpc
->gcb_pool
, GFP_ATOMIC
);
631 pr_err("rrpc: unable to queue block for gc.");
638 INIT_WORK(&gcb
->ws_gc
, rrpc_gc_queue
);
639 queue_work(rrpc
->kgc_wq
, &gcb
->ws_gc
);
642 static void rrpc_end_io_write(struct rrpc
*rrpc
, struct rrpc_rq
*rrqd
,
643 sector_t laddr
, uint8_t npages
)
646 struct rrpc_block
*rblk
;
650 for (i
= 0; i
< npages
; i
++) {
651 p
= &rrpc
->trans_map
[laddr
+ i
];
653 lun
= rblk
->parent
->lun
;
655 cmnt_size
= atomic_inc_return(&rblk
->data_cmnt_size
);
656 if (unlikely(cmnt_size
== rrpc
->dev
->pgs_per_blk
))
657 rrpc_run_gc(rrpc
, rblk
);
661 static void rrpc_end_io(struct nvm_rq
*rqd
, int error
)
663 struct rrpc
*rrpc
= container_of(rqd
->ins
, struct rrpc
, instance
);
664 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
665 uint8_t npages
= rqd
->nr_pages
;
666 sector_t laddr
= rrpc_get_laddr(rqd
->bio
) - npages
;
668 if (bio_data_dir(rqd
->bio
) == WRITE
)
669 rrpc_end_io_write(rrpc
, rrqd
, laddr
, npages
);
673 if (rrqd
->flags
& NVM_IOTYPE_GC
)
676 rrpc_unlock_rq(rrpc
, rqd
);
679 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
681 nvm_dev_dma_free(rrpc
->dev
, rqd
->metadata
, rqd
->dma_metadata
);
683 mempool_free(rqd
, rrpc
->rq_pool
);
686 static int rrpc_read_ppalist_rq(struct rrpc
*rrpc
, struct bio
*bio
,
687 struct nvm_rq
*rqd
, unsigned long flags
, int npages
)
689 struct rrpc_inflight_rq
*r
= rrpc_get_inflight_rq(rqd
);
690 struct rrpc_addr
*gp
;
691 sector_t laddr
= rrpc_get_laddr(bio
);
692 int is_gc
= flags
& NVM_IOTYPE_GC
;
695 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
)) {
696 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
697 return NVM_IO_REQUEUE
;
700 for (i
= 0; i
< npages
; i
++) {
701 /* We assume that mapping occurs at 4KB granularity */
702 BUG_ON(!(laddr
+ i
>= 0 && laddr
+ i
< rrpc
->nr_pages
));
703 gp
= &rrpc
->trans_map
[laddr
+ i
];
706 rqd
->ppa_list
[i
] = rrpc_ppa_to_gaddr(rrpc
->dev
,
710 rrpc_unlock_laddr(rrpc
, r
);
711 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
,
717 rqd
->opcode
= NVM_OP_HBREAD
;
722 static int rrpc_read_rq(struct rrpc
*rrpc
, struct bio
*bio
, struct nvm_rq
*rqd
,
725 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
726 int is_gc
= flags
& NVM_IOTYPE_GC
;
727 sector_t laddr
= rrpc_get_laddr(bio
);
728 struct rrpc_addr
*gp
;
730 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
))
731 return NVM_IO_REQUEUE
;
733 BUG_ON(!(laddr
>= 0 && laddr
< rrpc
->nr_pages
));
734 gp
= &rrpc
->trans_map
[laddr
];
737 rqd
->ppa_addr
= rrpc_ppa_to_gaddr(rrpc
->dev
, gp
->addr
);
740 rrpc_unlock_rq(rrpc
, rqd
);
744 rqd
->opcode
= NVM_OP_HBREAD
;
750 static int rrpc_write_ppalist_rq(struct rrpc
*rrpc
, struct bio
*bio
,
751 struct nvm_rq
*rqd
, unsigned long flags
, int npages
)
753 struct rrpc_inflight_rq
*r
= rrpc_get_inflight_rq(rqd
);
755 sector_t laddr
= rrpc_get_laddr(bio
);
756 int is_gc
= flags
& NVM_IOTYPE_GC
;
759 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
)) {
760 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
761 return NVM_IO_REQUEUE
;
764 for (i
= 0; i
< npages
; i
++) {
765 /* We assume that mapping occurs at 4KB granularity */
766 p
= rrpc_map_page(rrpc
, laddr
+ i
, is_gc
);
769 rrpc_unlock_laddr(rrpc
, r
);
770 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
,
773 return NVM_IO_REQUEUE
;
776 rqd
->ppa_list
[i
] = rrpc_ppa_to_gaddr(rrpc
->dev
,
780 rqd
->opcode
= NVM_OP_HBWRITE
;
785 static int rrpc_write_rq(struct rrpc
*rrpc
, struct bio
*bio
,
786 struct nvm_rq
*rqd
, unsigned long flags
)
788 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
790 int is_gc
= flags
& NVM_IOTYPE_GC
;
791 sector_t laddr
= rrpc_get_laddr(bio
);
793 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
))
794 return NVM_IO_REQUEUE
;
796 p
= rrpc_map_page(rrpc
, laddr
, is_gc
);
799 rrpc_unlock_rq(rrpc
, rqd
);
801 return NVM_IO_REQUEUE
;
804 rqd
->ppa_addr
= rrpc_ppa_to_gaddr(rrpc
->dev
, p
->addr
);
805 rqd
->opcode
= NVM_OP_HBWRITE
;
811 static int rrpc_setup_rq(struct rrpc
*rrpc
, struct bio
*bio
,
812 struct nvm_rq
*rqd
, unsigned long flags
, uint8_t npages
)
815 rqd
->ppa_list
= nvm_dev_dma_alloc(rrpc
->dev
, GFP_KERNEL
,
817 if (!rqd
->ppa_list
) {
818 pr_err("rrpc: not able to allocate ppa list\n");
822 if (bio_rw(bio
) == WRITE
)
823 return rrpc_write_ppalist_rq(rrpc
, bio
, rqd
, flags
,
826 return rrpc_read_ppalist_rq(rrpc
, bio
, rqd
, flags
, npages
);
829 if (bio_rw(bio
) == WRITE
)
830 return rrpc_write_rq(rrpc
, bio
, rqd
, flags
);
832 return rrpc_read_rq(rrpc
, bio
, rqd
, flags
);
835 static int rrpc_submit_io(struct rrpc
*rrpc
, struct bio
*bio
,
836 struct nvm_rq
*rqd
, unsigned long flags
)
839 struct rrpc_rq
*rrq
= nvm_rq_to_pdu(rqd
);
840 uint8_t nr_pages
= rrpc_get_pages(bio
);
841 int bio_size
= bio_sectors(bio
) << 9;
843 if (bio_size
< rrpc
->dev
->sec_size
)
845 else if (bio_size
> rrpc
->dev
->max_rq_size
)
848 err
= rrpc_setup_rq(rrpc
, bio
, rqd
, flags
, nr_pages
);
854 rqd
->ins
= &rrpc
->instance
;
855 rqd
->nr_pages
= nr_pages
;
858 err
= nvm_submit_io(rrpc
->dev
, rqd
);
860 pr_err("rrpc: I/O submission failed: %d\n", err
);
862 if (!(flags
& NVM_IOTYPE_GC
)) {
863 rrpc_unlock_rq(rrpc
, rqd
);
864 if (rqd
->nr_pages
> 1)
865 nvm_dev_dma_free(rrpc
->dev
,
866 rqd
->ppa_list
, rqd
->dma_ppa_list
);
874 static blk_qc_t
rrpc_make_rq(struct request_queue
*q
, struct bio
*bio
)
876 struct rrpc
*rrpc
= q
->queuedata
;
880 if (bio
->bi_rw
& REQ_DISCARD
) {
881 rrpc_discard(rrpc
, bio
);
882 return BLK_QC_T_NONE
;
885 rqd
= mempool_alloc(rrpc
->rq_pool
, GFP_KERNEL
);
887 pr_err_ratelimited("rrpc: not able to queue bio.");
889 return BLK_QC_T_NONE
;
891 memset(rqd
, 0, sizeof(struct nvm_rq
));
893 err
= rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_NONE
);
896 return BLK_QC_T_NONE
;
904 spin_lock(&rrpc
->bio_lock
);
905 bio_list_add(&rrpc
->requeue_bios
, bio
);
906 spin_unlock(&rrpc
->bio_lock
);
907 queue_work(rrpc
->kgc_wq
, &rrpc
->ws_requeue
);
911 mempool_free(rqd
, rrpc
->rq_pool
);
912 return BLK_QC_T_NONE
;
915 static void rrpc_requeue(struct work_struct
*work
)
917 struct rrpc
*rrpc
= container_of(work
, struct rrpc
, ws_requeue
);
918 struct bio_list bios
;
921 bio_list_init(&bios
);
923 spin_lock(&rrpc
->bio_lock
);
924 bio_list_merge(&bios
, &rrpc
->requeue_bios
);
925 bio_list_init(&rrpc
->requeue_bios
);
926 spin_unlock(&rrpc
->bio_lock
);
928 while ((bio
= bio_list_pop(&bios
)))
929 rrpc_make_rq(rrpc
->disk
->queue
, bio
);
932 static void rrpc_gc_free(struct rrpc
*rrpc
)
934 struct rrpc_lun
*rlun
;
938 destroy_workqueue(rrpc
->krqd_wq
);
941 destroy_workqueue(rrpc
->kgc_wq
);
946 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
947 rlun
= &rrpc
->luns
[i
];
955 static int rrpc_gc_init(struct rrpc
*rrpc
)
957 rrpc
->krqd_wq
= alloc_workqueue("rrpc-lun", WQ_MEM_RECLAIM
|WQ_UNBOUND
,
962 rrpc
->kgc_wq
= alloc_workqueue("rrpc-bg", WQ_MEM_RECLAIM
, 1);
966 setup_timer(&rrpc
->gc_timer
, rrpc_gc_timer
, (unsigned long)rrpc
);
971 static void rrpc_map_free(struct rrpc
*rrpc
)
973 vfree(rrpc
->rev_trans_map
);
974 vfree(rrpc
->trans_map
);
977 static int rrpc_l2p_update(u64 slba
, u32 nlb
, __le64
*entries
, void *private)
979 struct rrpc
*rrpc
= (struct rrpc
*)private;
980 struct nvm_dev
*dev
= rrpc
->dev
;
981 struct rrpc_addr
*addr
= rrpc
->trans_map
+ slba
;
982 struct rrpc_rev_addr
*raddr
= rrpc
->rev_trans_map
;
983 sector_t max_pages
= dev
->total_pages
* (dev
->sec_size
>> 9);
984 u64 elba
= slba
+ nlb
;
987 if (unlikely(elba
> dev
->total_pages
)) {
988 pr_err("nvm: L2P data from device is out of bounds!\n");
992 for (i
= 0; i
< nlb
; i
++) {
993 u64 pba
= le64_to_cpu(entries
[i
]);
994 /* LNVM treats address-spaces as silos, LBA and PBA are
995 * equally large and zero-indexed.
997 if (unlikely(pba
>= max_pages
&& pba
!= U64_MAX
)) {
998 pr_err("nvm: L2P data entry is out of bounds!\n");
1002 /* Address zero is a special one. The first page on a disk is
1003 * protected. As it often holds internal device boot
1010 raddr
[pba
].addr
= slba
+ i
;
1016 static int rrpc_map_init(struct rrpc
*rrpc
)
1018 struct nvm_dev
*dev
= rrpc
->dev
;
1022 rrpc
->trans_map
= vzalloc(sizeof(struct rrpc_addr
) * rrpc
->nr_pages
);
1023 if (!rrpc
->trans_map
)
1026 rrpc
->rev_trans_map
= vmalloc(sizeof(struct rrpc_rev_addr
)
1028 if (!rrpc
->rev_trans_map
)
1031 for (i
= 0; i
< rrpc
->nr_pages
; i
++) {
1032 struct rrpc_addr
*p
= &rrpc
->trans_map
[i
];
1033 struct rrpc_rev_addr
*r
= &rrpc
->rev_trans_map
[i
];
1035 p
->addr
= ADDR_EMPTY
;
1036 r
->addr
= ADDR_EMPTY
;
1039 if (!dev
->ops
->get_l2p_tbl
)
1042 /* Bring up the mapping table from device */
1043 ret
= dev
->ops
->get_l2p_tbl(dev
, 0, dev
->total_pages
,
1044 rrpc_l2p_update
, rrpc
);
1046 pr_err("nvm: rrpc: could not read L2P table.\n");
1054 /* Minimum pages needed within a lun */
1055 #define PAGE_POOL_SIZE 16
1056 #define ADDR_POOL_SIZE 64
1058 static int rrpc_core_init(struct rrpc
*rrpc
)
1060 down_write(&rrpc_lock
);
1061 if (!rrpc_gcb_cache
) {
1062 rrpc_gcb_cache
= kmem_cache_create("rrpc_gcb",
1063 sizeof(struct rrpc_block_gc
), 0, 0, NULL
);
1064 if (!rrpc_gcb_cache
) {
1065 up_write(&rrpc_lock
);
1069 rrpc_rq_cache
= kmem_cache_create("rrpc_rq",
1070 sizeof(struct nvm_rq
) + sizeof(struct rrpc_rq
),
1072 if (!rrpc_rq_cache
) {
1073 kmem_cache_destroy(rrpc_gcb_cache
);
1074 up_write(&rrpc_lock
);
1078 up_write(&rrpc_lock
);
1080 rrpc
->page_pool
= mempool_create_page_pool(PAGE_POOL_SIZE
, 0);
1081 if (!rrpc
->page_pool
)
1084 rrpc
->gcb_pool
= mempool_create_slab_pool(rrpc
->dev
->nr_luns
,
1086 if (!rrpc
->gcb_pool
)
1089 rrpc
->rq_pool
= mempool_create_slab_pool(64, rrpc_rq_cache
);
1093 spin_lock_init(&rrpc
->inflights
.lock
);
1094 INIT_LIST_HEAD(&rrpc
->inflights
.reqs
);
1099 static void rrpc_core_free(struct rrpc
*rrpc
)
1101 mempool_destroy(rrpc
->page_pool
);
1102 mempool_destroy(rrpc
->gcb_pool
);
1103 mempool_destroy(rrpc
->rq_pool
);
1106 static void rrpc_luns_free(struct rrpc
*rrpc
)
1111 static int rrpc_luns_init(struct rrpc
*rrpc
, int lun_begin
, int lun_end
)
1113 struct nvm_dev
*dev
= rrpc
->dev
;
1114 struct rrpc_lun
*rlun
;
1117 spin_lock_init(&rrpc
->rev_lock
);
1119 rrpc
->luns
= kcalloc(rrpc
->nr_luns
, sizeof(struct rrpc_lun
),
1125 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
1126 struct nvm_lun
*lun
= dev
->mt
->get_lun(dev
, lun_begin
+ i
);
1128 if (dev
->pgs_per_blk
>
1129 MAX_INVALID_PAGES_STORAGE
* BITS_PER_LONG
) {
1130 pr_err("rrpc: number of pages per block too high.");
1134 rlun
= &rrpc
->luns
[i
];
1137 INIT_LIST_HEAD(&rlun
->prio_list
);
1138 INIT_WORK(&rlun
->ws_gc
, rrpc_lun_gc
);
1139 spin_lock_init(&rlun
->lock
);
1141 rrpc
->total_blocks
+= dev
->blks_per_lun
;
1142 rrpc
->nr_pages
+= dev
->sec_per_lun
;
1144 rlun
->blocks
= vzalloc(sizeof(struct rrpc_block
) *
1145 rrpc
->dev
->blks_per_lun
);
1149 for (j
= 0; j
< rrpc
->dev
->blks_per_lun
; j
++) {
1150 struct rrpc_block
*rblk
= &rlun
->blocks
[j
];
1151 struct nvm_block
*blk
= &lun
->blocks
[j
];
1154 INIT_LIST_HEAD(&rblk
->prio
);
1155 spin_lock_init(&rblk
->lock
);
1164 static void rrpc_free(struct rrpc
*rrpc
)
1167 rrpc_map_free(rrpc
);
1168 rrpc_core_free(rrpc
);
1169 rrpc_luns_free(rrpc
);
1174 static void rrpc_exit(void *private)
1176 struct rrpc
*rrpc
= private;
1178 del_timer(&rrpc
->gc_timer
);
1180 flush_workqueue(rrpc
->krqd_wq
);
1181 flush_workqueue(rrpc
->kgc_wq
);
1186 static sector_t
rrpc_capacity(void *private)
1188 struct rrpc
*rrpc
= private;
1189 struct nvm_dev
*dev
= rrpc
->dev
;
1190 sector_t reserved
, provisioned
;
1192 /* cur, gc, and two emergency blocks for each lun */
1193 reserved
= rrpc
->nr_luns
* dev
->max_pages_per_blk
* 4;
1194 provisioned
= rrpc
->nr_pages
- reserved
;
1196 if (reserved
> rrpc
->nr_pages
) {
1197 pr_err("rrpc: not enough space available to expose storage.\n");
1201 sector_div(provisioned
, 10);
1202 return provisioned
* 9 * NR_PHY_IN_LOG
;
1206 * Looks up the logical address from reverse trans map and check if its valid by
1207 * comparing the logical to physical address with the physical address.
1208 * Returns 0 on free, otherwise 1 if in use
1210 static void rrpc_block_map_update(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
1212 struct nvm_dev
*dev
= rrpc
->dev
;
1214 struct rrpc_addr
*laddr
;
1217 for (offset
= 0; offset
< dev
->pgs_per_blk
; offset
++) {
1218 paddr
= block_to_addr(rrpc
, rblk
) + offset
;
1220 pladdr
= rrpc
->rev_trans_map
[paddr
].addr
;
1221 if (pladdr
== ADDR_EMPTY
)
1224 laddr
= &rrpc
->trans_map
[pladdr
];
1226 if (paddr
== laddr
->addr
) {
1229 set_bit(offset
, rblk
->invalid_pages
);
1230 rblk
->nr_invalid_pages
++;
1235 static int rrpc_blocks_init(struct rrpc
*rrpc
)
1237 struct rrpc_lun
*rlun
;
1238 struct rrpc_block
*rblk
;
1239 int lun_iter
, blk_iter
;
1241 for (lun_iter
= 0; lun_iter
< rrpc
->nr_luns
; lun_iter
++) {
1242 rlun
= &rrpc
->luns
[lun_iter
];
1244 for (blk_iter
= 0; blk_iter
< rrpc
->dev
->blks_per_lun
;
1246 rblk
= &rlun
->blocks
[blk_iter
];
1247 rrpc_block_map_update(rrpc
, rblk
);
1254 static int rrpc_luns_configure(struct rrpc
*rrpc
)
1256 struct rrpc_lun
*rlun
;
1257 struct rrpc_block
*rblk
;
1260 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
1261 rlun
= &rrpc
->luns
[i
];
1263 rblk
= rrpc_get_blk(rrpc
, rlun
, 0);
1267 rrpc_set_lun_cur(rlun
, rblk
);
1269 /* Emergency gc block */
1270 rblk
= rrpc_get_blk(rrpc
, rlun
, 1);
1273 rlun
->gc_cur
= rblk
;
1278 rrpc_put_blks(rrpc
);
1282 static struct nvm_tgt_type tt_rrpc
;
1284 static void *rrpc_init(struct nvm_dev
*dev
, struct gendisk
*tdisk
,
1285 int lun_begin
, int lun_end
)
1287 struct request_queue
*bqueue
= dev
->q
;
1288 struct request_queue
*tqueue
= tdisk
->queue
;
1292 if (!(dev
->identity
.dom
& NVM_RSP_L2P
)) {
1293 pr_err("nvm: rrpc: device does not support l2p (%x)\n",
1295 return ERR_PTR(-EINVAL
);
1298 rrpc
= kzalloc(sizeof(struct rrpc
), GFP_KERNEL
);
1300 return ERR_PTR(-ENOMEM
);
1302 rrpc
->instance
.tt
= &tt_rrpc
;
1306 bio_list_init(&rrpc
->requeue_bios
);
1307 spin_lock_init(&rrpc
->bio_lock
);
1308 INIT_WORK(&rrpc
->ws_requeue
, rrpc_requeue
);
1310 rrpc
->nr_luns
= lun_end
- lun_begin
+ 1;
1312 /* simple round-robin strategy */
1313 atomic_set(&rrpc
->next_lun
, -1);
1315 ret
= rrpc_luns_init(rrpc
, lun_begin
, lun_end
);
1317 pr_err("nvm: rrpc: could not initialize luns\n");
1321 rrpc
->poffset
= dev
->sec_per_lun
* lun_begin
;
1322 rrpc
->lun_offset
= lun_begin
;
1324 ret
= rrpc_core_init(rrpc
);
1326 pr_err("nvm: rrpc: could not initialize core\n");
1330 ret
= rrpc_map_init(rrpc
);
1332 pr_err("nvm: rrpc: could not initialize maps\n");
1336 ret
= rrpc_blocks_init(rrpc
);
1338 pr_err("nvm: rrpc: could not initialize state for blocks\n");
1342 ret
= rrpc_luns_configure(rrpc
);
1344 pr_err("nvm: rrpc: not enough blocks available in LUNs.\n");
1348 ret
= rrpc_gc_init(rrpc
);
1350 pr_err("nvm: rrpc: could not initialize gc\n");
1354 /* inherit the size from the underlying device */
1355 blk_queue_logical_block_size(tqueue
, queue_physical_block_size(bqueue
));
1356 blk_queue_max_hw_sectors(tqueue
, queue_max_hw_sectors(bqueue
));
1358 pr_info("nvm: rrpc initialized with %u luns and %llu pages.\n",
1359 rrpc
->nr_luns
, (unsigned long long)rrpc
->nr_pages
);
1361 mod_timer(&rrpc
->gc_timer
, jiffies
+ msecs_to_jiffies(10));
1366 return ERR_PTR(ret
);
1369 /* round robin, page-based FTL, and cost-based GC */
1370 static struct nvm_tgt_type tt_rrpc
= {
1372 .version
= {1, 0, 0},
1374 .make_rq
= rrpc_make_rq
,
1375 .capacity
= rrpc_capacity
,
1376 .end_io
= rrpc_end_io
,
1382 static int __init
rrpc_module_init(void)
1384 return nvm_register_target(&tt_rrpc
);
1387 static void rrpc_module_exit(void)
1389 nvm_unregister_target(&tt_rrpc
);
1392 module_init(rrpc_module_init
);
1393 module_exit(rrpc_module_exit
);
1394 MODULE_LICENSE("GPL v2");
1395 MODULE_DESCRIPTION("Block-Device Target for Open-Channel SSDs");