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 struct rrpc_lun
*get_next_lun(struct rrpc
*rrpc
)
207 int next
= atomic_inc_return(&rrpc
->next_lun
);
209 return &rrpc
->luns
[next
% rrpc
->nr_luns
];
212 static void rrpc_gc_kick(struct rrpc
*rrpc
)
214 struct rrpc_lun
*rlun
;
217 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
218 rlun
= &rrpc
->luns
[i
];
219 queue_work(rrpc
->krqd_wq
, &rlun
->ws_gc
);
224 * timed GC every interval.
226 static void rrpc_gc_timer(unsigned long data
)
228 struct rrpc
*rrpc
= (struct rrpc
*)data
;
231 mod_timer(&rrpc
->gc_timer
, jiffies
+ msecs_to_jiffies(10));
234 static void rrpc_end_sync_bio(struct bio
*bio
)
236 struct completion
*waiting
= bio
->bi_private
;
239 pr_err("nvm: gc request failed (%u).\n", bio
->bi_error
);
245 * rrpc_move_valid_pages -- migrate live data off the block
246 * @rrpc: the 'rrpc' structure
247 * @block: the block from which to migrate live pages
250 * GC algorithms may call this function to migrate remaining live
251 * pages off the block prior to erasing it. This function blocks
252 * further execution until the operation is complete.
254 static int rrpc_move_valid_pages(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
256 struct request_queue
*q
= rrpc
->dev
->q
;
257 struct rrpc_rev_addr
*rev
;
262 int nr_pgs_per_blk
= rrpc
->dev
->pgs_per_blk
;
264 DECLARE_COMPLETION_ONSTACK(wait
);
266 if (bitmap_full(rblk
->invalid_pages
, nr_pgs_per_blk
))
269 bio
= bio_alloc(GFP_NOIO
, 1);
271 pr_err("nvm: could not alloc bio to gc\n");
275 page
= mempool_alloc(rrpc
->page_pool
, GFP_NOIO
);
277 while ((slot
= find_first_zero_bit(rblk
->invalid_pages
,
278 nr_pgs_per_blk
)) < nr_pgs_per_blk
) {
281 phys_addr
= (rblk
->parent
->id
* nr_pgs_per_blk
) + slot
;
284 spin_lock(&rrpc
->rev_lock
);
285 /* Get logical address from physical to logical table */
286 rev
= &rrpc
->rev_trans_map
[phys_addr
- rrpc
->poffset
];
287 /* already updated by previous regular write */
288 if (rev
->addr
== ADDR_EMPTY
) {
289 spin_unlock(&rrpc
->rev_lock
);
293 rqd
= rrpc_inflight_laddr_acquire(rrpc
, rev
->addr
, 1);
294 if (IS_ERR_OR_NULL(rqd
)) {
295 spin_unlock(&rrpc
->rev_lock
);
300 spin_unlock(&rrpc
->rev_lock
);
302 /* Perform read to do GC */
303 bio
->bi_iter
.bi_sector
= rrpc_get_sector(rev
->addr
);
305 bio
->bi_private
= &wait
;
306 bio
->bi_end_io
= rrpc_end_sync_bio
;
308 /* TODO: may fail when EXP_PG_SIZE > PAGE_SIZE */
309 bio_add_pc_page(q
, bio
, page
, RRPC_EXPOSED_PAGE_SIZE
, 0);
311 if (rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_GC
)) {
312 pr_err("rrpc: gc read failed.\n");
313 rrpc_inflight_laddr_release(rrpc
, rqd
);
316 wait_for_completion_io(&wait
);
319 reinit_completion(&wait
);
321 bio
->bi_iter
.bi_sector
= rrpc_get_sector(rev
->addr
);
323 bio
->bi_private
= &wait
;
324 bio
->bi_end_io
= rrpc_end_sync_bio
;
326 bio_add_pc_page(q
, bio
, page
, RRPC_EXPOSED_PAGE_SIZE
, 0);
328 /* turn the command around and write the data back to a new
331 if (rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_GC
)) {
332 pr_err("rrpc: gc write failed.\n");
333 rrpc_inflight_laddr_release(rrpc
, rqd
);
336 wait_for_completion_io(&wait
);
338 rrpc_inflight_laddr_release(rrpc
, rqd
);
344 mempool_free(page
, rrpc
->page_pool
);
347 if (!bitmap_full(rblk
->invalid_pages
, nr_pgs_per_blk
)) {
348 pr_err("nvm: failed to garbage collect block\n");
355 static void rrpc_block_gc(struct work_struct
*work
)
357 struct rrpc_block_gc
*gcb
= container_of(work
, struct rrpc_block_gc
,
359 struct rrpc
*rrpc
= gcb
->rrpc
;
360 struct rrpc_block
*rblk
= gcb
->rblk
;
361 struct nvm_dev
*dev
= rrpc
->dev
;
363 pr_debug("nvm: block '%lu' being reclaimed\n", rblk
->parent
->id
);
365 if (rrpc_move_valid_pages(rrpc
, rblk
))
368 nvm_erase_blk(dev
, rblk
->parent
);
369 rrpc_put_blk(rrpc
, rblk
);
371 mempool_free(gcb
, rrpc
->gcb_pool
);
374 /* the block with highest number of invalid pages, will be in the beginning
377 static struct rrpc_block
*rblock_max_invalid(struct rrpc_block
*ra
,
378 struct rrpc_block
*rb
)
380 if (ra
->nr_invalid_pages
== rb
->nr_invalid_pages
)
383 return (ra
->nr_invalid_pages
< rb
->nr_invalid_pages
) ? rb
: ra
;
386 /* linearly find the block with highest number of invalid pages
389 static struct rrpc_block
*block_prio_find_max(struct rrpc_lun
*rlun
)
391 struct list_head
*prio_list
= &rlun
->prio_list
;
392 struct rrpc_block
*rblock
, *max
;
394 BUG_ON(list_empty(prio_list
));
396 max
= list_first_entry(prio_list
, struct rrpc_block
, prio
);
397 list_for_each_entry(rblock
, prio_list
, prio
)
398 max
= rblock_max_invalid(max
, rblock
);
403 static void rrpc_lun_gc(struct work_struct
*work
)
405 struct rrpc_lun
*rlun
= container_of(work
, struct rrpc_lun
, ws_gc
);
406 struct rrpc
*rrpc
= rlun
->rrpc
;
407 struct nvm_lun
*lun
= rlun
->parent
;
408 struct rrpc_block_gc
*gcb
;
409 unsigned int nr_blocks_need
;
411 nr_blocks_need
= rrpc
->dev
->blks_per_lun
/ GC_LIMIT_INVERSE
;
413 if (nr_blocks_need
< rrpc
->nr_luns
)
414 nr_blocks_need
= rrpc
->nr_luns
;
416 spin_lock(&lun
->lock
);
417 while (nr_blocks_need
> lun
->nr_free_blocks
&&
418 !list_empty(&rlun
->prio_list
)) {
419 struct rrpc_block
*rblock
= block_prio_find_max(rlun
);
420 struct nvm_block
*block
= rblock
->parent
;
422 if (!rblock
->nr_invalid_pages
)
425 list_del_init(&rblock
->prio
);
427 BUG_ON(!block_is_full(rrpc
, rblock
));
429 pr_debug("rrpc: selected block '%lu' for GC\n", block
->id
);
431 gcb
= mempool_alloc(rrpc
->gcb_pool
, GFP_ATOMIC
);
437 INIT_WORK(&gcb
->ws_gc
, rrpc_block_gc
);
439 queue_work(rrpc
->kgc_wq
, &gcb
->ws_gc
);
443 spin_unlock(&lun
->lock
);
445 /* TODO: Hint that request queue can be started again */
448 static void rrpc_gc_queue(struct work_struct
*work
)
450 struct rrpc_block_gc
*gcb
= container_of(work
, struct rrpc_block_gc
,
452 struct rrpc
*rrpc
= gcb
->rrpc
;
453 struct rrpc_block
*rblk
= gcb
->rblk
;
454 struct nvm_lun
*lun
= rblk
->parent
->lun
;
455 struct rrpc_lun
*rlun
= &rrpc
->luns
[lun
->id
- rrpc
->lun_offset
];
457 spin_lock(&rlun
->lock
);
458 list_add_tail(&rblk
->prio
, &rlun
->prio_list
);
459 spin_unlock(&rlun
->lock
);
461 mempool_free(gcb
, rrpc
->gcb_pool
);
462 pr_debug("nvm: block '%lu' is full, allow GC (sched)\n",
466 static const struct block_device_operations rrpc_fops
= {
467 .owner
= THIS_MODULE
,
470 static struct rrpc_lun
*rrpc_get_lun_rr(struct rrpc
*rrpc
, int is_gc
)
473 struct rrpc_lun
*rlun
, *max_free
;
476 return get_next_lun(rrpc
);
478 /* during GC, we don't care about RR, instead we want to make
479 * sure that we maintain evenness between the block luns.
481 max_free
= &rrpc
->luns
[0];
482 /* prevent GC-ing lun from devouring pages of a lun with
483 * little free blocks. We don't take the lock as we only need an
486 rrpc_for_each_lun(rrpc
, rlun
, i
) {
487 if (rlun
->parent
->nr_free_blocks
>
488 max_free
->parent
->nr_free_blocks
)
495 static struct rrpc_addr
*rrpc_update_map(struct rrpc
*rrpc
, sector_t laddr
,
496 struct rrpc_block
*rblk
, u64 paddr
)
498 struct rrpc_addr
*gp
;
499 struct rrpc_rev_addr
*rev
;
501 BUG_ON(laddr
>= rrpc
->nr_pages
);
503 gp
= &rrpc
->trans_map
[laddr
];
504 spin_lock(&rrpc
->rev_lock
);
506 rrpc_page_invalidate(rrpc
, gp
);
511 rev
= &rrpc
->rev_trans_map
[gp
->addr
- rrpc
->poffset
];
513 spin_unlock(&rrpc
->rev_lock
);
518 static u64
rrpc_alloc_addr(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
520 u64 addr
= ADDR_EMPTY
;
522 spin_lock(&rblk
->lock
);
523 if (block_is_full(rrpc
, rblk
))
526 addr
= block_to_addr(rrpc
, rblk
) + rblk
->next_page
;
530 spin_unlock(&rblk
->lock
);
534 /* Simple round-robin Logical to physical address translation.
536 * Retrieve the mapping using the active append point. Then update the ap for
537 * the next write to the disk.
539 * Returns rrpc_addr with the physical address and block. Remember to return to
540 * rrpc->addr_cache when request is finished.
542 static struct rrpc_addr
*rrpc_map_page(struct rrpc
*rrpc
, sector_t laddr
,
545 struct rrpc_lun
*rlun
;
546 struct rrpc_block
*rblk
;
550 rlun
= rrpc_get_lun_rr(rrpc
, is_gc
);
553 if (!is_gc
&& lun
->nr_free_blocks
< rrpc
->nr_luns
* 4)
556 spin_lock(&rlun
->lock
);
560 paddr
= rrpc_alloc_addr(rrpc
, rblk
);
562 if (paddr
== ADDR_EMPTY
) {
563 rblk
= rrpc_get_blk(rrpc
, rlun
, 0);
565 rrpc_set_lun_cur(rlun
, rblk
);
570 /* retry from emergency gc block */
571 paddr
= rrpc_alloc_addr(rrpc
, rlun
->gc_cur
);
572 if (paddr
== ADDR_EMPTY
) {
573 rblk
= rrpc_get_blk(rrpc
, rlun
, 1);
575 pr_err("rrpc: no more blocks");
580 paddr
= rrpc_alloc_addr(rrpc
, rlun
->gc_cur
);
586 spin_unlock(&rlun
->lock
);
587 return rrpc_update_map(rrpc
, laddr
, rblk
, paddr
);
589 spin_unlock(&rlun
->lock
);
593 static void rrpc_run_gc(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
595 struct rrpc_block_gc
*gcb
;
597 gcb
= mempool_alloc(rrpc
->gcb_pool
, GFP_ATOMIC
);
599 pr_err("rrpc: unable to queue block for gc.");
606 INIT_WORK(&gcb
->ws_gc
, rrpc_gc_queue
);
607 queue_work(rrpc
->kgc_wq
, &gcb
->ws_gc
);
610 static void rrpc_end_io_write(struct rrpc
*rrpc
, struct rrpc_rq
*rrqd
,
611 sector_t laddr
, uint8_t npages
)
614 struct rrpc_block
*rblk
;
618 for (i
= 0; i
< npages
; i
++) {
619 p
= &rrpc
->trans_map
[laddr
+ i
];
621 lun
= rblk
->parent
->lun
;
623 cmnt_size
= atomic_inc_return(&rblk
->data_cmnt_size
);
624 if (unlikely(cmnt_size
== rrpc
->dev
->pgs_per_blk
))
625 rrpc_run_gc(rrpc
, rblk
);
629 static int rrpc_end_io(struct nvm_rq
*rqd
, int error
)
631 struct rrpc
*rrpc
= container_of(rqd
->ins
, struct rrpc
, instance
);
632 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
633 uint8_t npages
= rqd
->nr_pages
;
634 sector_t laddr
= rrpc_get_laddr(rqd
->bio
) - npages
;
636 if (bio_data_dir(rqd
->bio
) == WRITE
)
637 rrpc_end_io_write(rrpc
, rrqd
, laddr
, npages
);
639 if (rrqd
->flags
& NVM_IOTYPE_GC
)
642 rrpc_unlock_rq(rrpc
, rqd
);
646 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
648 nvm_dev_dma_free(rrpc
->dev
, rqd
->metadata
, rqd
->dma_metadata
);
650 mempool_free(rqd
, rrpc
->rq_pool
);
655 static int rrpc_read_ppalist_rq(struct rrpc
*rrpc
, struct bio
*bio
,
656 struct nvm_rq
*rqd
, unsigned long flags
, int npages
)
658 struct rrpc_inflight_rq
*r
= rrpc_get_inflight_rq(rqd
);
659 struct rrpc_addr
*gp
;
660 sector_t laddr
= rrpc_get_laddr(bio
);
661 int is_gc
= flags
& NVM_IOTYPE_GC
;
664 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
)) {
665 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
666 return NVM_IO_REQUEUE
;
669 for (i
= 0; i
< npages
; i
++) {
670 /* We assume that mapping occurs at 4KB granularity */
671 BUG_ON(!(laddr
+ i
>= 0 && laddr
+ i
< rrpc
->nr_pages
));
672 gp
= &rrpc
->trans_map
[laddr
+ i
];
675 rqd
->ppa_list
[i
] = rrpc_ppa_to_gaddr(rrpc
->dev
,
679 rrpc_unlock_laddr(rrpc
, r
);
680 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
,
686 rqd
->opcode
= NVM_OP_HBREAD
;
691 static int rrpc_read_rq(struct rrpc
*rrpc
, struct bio
*bio
, struct nvm_rq
*rqd
,
694 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
695 int is_gc
= flags
& NVM_IOTYPE_GC
;
696 sector_t laddr
= rrpc_get_laddr(bio
);
697 struct rrpc_addr
*gp
;
699 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
))
700 return NVM_IO_REQUEUE
;
702 BUG_ON(!(laddr
>= 0 && laddr
< rrpc
->nr_pages
));
703 gp
= &rrpc
->trans_map
[laddr
];
706 rqd
->ppa_addr
= rrpc_ppa_to_gaddr(rrpc
->dev
, gp
->addr
);
709 rrpc_unlock_rq(rrpc
, rqd
);
713 rqd
->opcode
= NVM_OP_HBREAD
;
719 static int rrpc_write_ppalist_rq(struct rrpc
*rrpc
, struct bio
*bio
,
720 struct nvm_rq
*rqd
, unsigned long flags
, int npages
)
722 struct rrpc_inflight_rq
*r
= rrpc_get_inflight_rq(rqd
);
724 sector_t laddr
= rrpc_get_laddr(bio
);
725 int is_gc
= flags
& NVM_IOTYPE_GC
;
728 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
)) {
729 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
, rqd
->dma_ppa_list
);
730 return NVM_IO_REQUEUE
;
733 for (i
= 0; i
< npages
; i
++) {
734 /* We assume that mapping occurs at 4KB granularity */
735 p
= rrpc_map_page(rrpc
, laddr
+ i
, is_gc
);
738 rrpc_unlock_laddr(rrpc
, r
);
739 nvm_dev_dma_free(rrpc
->dev
, rqd
->ppa_list
,
742 return NVM_IO_REQUEUE
;
745 rqd
->ppa_list
[i
] = rrpc_ppa_to_gaddr(rrpc
->dev
,
749 rqd
->opcode
= NVM_OP_HBWRITE
;
754 static int rrpc_write_rq(struct rrpc
*rrpc
, struct bio
*bio
,
755 struct nvm_rq
*rqd
, unsigned long flags
)
757 struct rrpc_rq
*rrqd
= nvm_rq_to_pdu(rqd
);
759 int is_gc
= flags
& NVM_IOTYPE_GC
;
760 sector_t laddr
= rrpc_get_laddr(bio
);
762 if (!is_gc
&& rrpc_lock_rq(rrpc
, bio
, rqd
))
763 return NVM_IO_REQUEUE
;
765 p
= rrpc_map_page(rrpc
, laddr
, is_gc
);
768 rrpc_unlock_rq(rrpc
, rqd
);
770 return NVM_IO_REQUEUE
;
773 rqd
->ppa_addr
= rrpc_ppa_to_gaddr(rrpc
->dev
, p
->addr
);
774 rqd
->opcode
= NVM_OP_HBWRITE
;
780 static int rrpc_setup_rq(struct rrpc
*rrpc
, struct bio
*bio
,
781 struct nvm_rq
*rqd
, unsigned long flags
, uint8_t npages
)
784 rqd
->ppa_list
= nvm_dev_dma_alloc(rrpc
->dev
, GFP_KERNEL
,
786 if (!rqd
->ppa_list
) {
787 pr_err("rrpc: not able to allocate ppa list\n");
791 if (bio_rw(bio
) == WRITE
)
792 return rrpc_write_ppalist_rq(rrpc
, bio
, rqd
, flags
,
795 return rrpc_read_ppalist_rq(rrpc
, bio
, rqd
, flags
, npages
);
798 if (bio_rw(bio
) == WRITE
)
799 return rrpc_write_rq(rrpc
, bio
, rqd
, flags
);
801 return rrpc_read_rq(rrpc
, bio
, rqd
, flags
);
804 static int rrpc_submit_io(struct rrpc
*rrpc
, struct bio
*bio
,
805 struct nvm_rq
*rqd
, unsigned long flags
)
808 struct rrpc_rq
*rrq
= nvm_rq_to_pdu(rqd
);
809 uint8_t nr_pages
= rrpc_get_pages(bio
);
810 int bio_size
= bio_sectors(bio
) << 9;
812 if (bio_size
< rrpc
->dev
->sec_size
)
814 else if (bio_size
> rrpc
->dev
->max_rq_size
)
817 err
= rrpc_setup_rq(rrpc
, bio
, rqd
, flags
, nr_pages
);
823 rqd
->ins
= &rrpc
->instance
;
824 rqd
->nr_pages
= nr_pages
;
827 err
= nvm_submit_io(rrpc
->dev
, rqd
);
829 pr_err("rrpc: I/O submission failed: %d\n", err
);
836 static blk_qc_t
rrpc_make_rq(struct request_queue
*q
, struct bio
*bio
)
838 struct rrpc
*rrpc
= q
->queuedata
;
842 if (bio
->bi_rw
& REQ_DISCARD
) {
843 rrpc_discard(rrpc
, bio
);
844 return BLK_QC_T_NONE
;
847 rqd
= mempool_alloc(rrpc
->rq_pool
, GFP_KERNEL
);
849 pr_err_ratelimited("rrpc: not able to queue bio.");
851 return BLK_QC_T_NONE
;
853 memset(rqd
, 0, sizeof(struct nvm_rq
));
855 err
= rrpc_submit_io(rrpc
, bio
, rqd
, NVM_IOTYPE_NONE
);
858 return BLK_QC_T_NONE
;
866 spin_lock(&rrpc
->bio_lock
);
867 bio_list_add(&rrpc
->requeue_bios
, bio
);
868 spin_unlock(&rrpc
->bio_lock
);
869 queue_work(rrpc
->kgc_wq
, &rrpc
->ws_requeue
);
873 mempool_free(rqd
, rrpc
->rq_pool
);
874 return BLK_QC_T_NONE
;
877 static void rrpc_requeue(struct work_struct
*work
)
879 struct rrpc
*rrpc
= container_of(work
, struct rrpc
, ws_requeue
);
880 struct bio_list bios
;
883 bio_list_init(&bios
);
885 spin_lock(&rrpc
->bio_lock
);
886 bio_list_merge(&bios
, &rrpc
->requeue_bios
);
887 bio_list_init(&rrpc
->requeue_bios
);
888 spin_unlock(&rrpc
->bio_lock
);
890 while ((bio
= bio_list_pop(&bios
)))
891 rrpc_make_rq(rrpc
->disk
->queue
, bio
);
894 static void rrpc_gc_free(struct rrpc
*rrpc
)
896 struct rrpc_lun
*rlun
;
900 destroy_workqueue(rrpc
->krqd_wq
);
903 destroy_workqueue(rrpc
->kgc_wq
);
908 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
909 rlun
= &rrpc
->luns
[i
];
917 static int rrpc_gc_init(struct rrpc
*rrpc
)
919 rrpc
->krqd_wq
= alloc_workqueue("rrpc-lun", WQ_MEM_RECLAIM
|WQ_UNBOUND
,
924 rrpc
->kgc_wq
= alloc_workqueue("rrpc-bg", WQ_MEM_RECLAIM
, 1);
928 setup_timer(&rrpc
->gc_timer
, rrpc_gc_timer
, (unsigned long)rrpc
);
933 static void rrpc_map_free(struct rrpc
*rrpc
)
935 vfree(rrpc
->rev_trans_map
);
936 vfree(rrpc
->trans_map
);
939 static int rrpc_l2p_update(u64 slba
, u32 nlb
, __le64
*entries
, void *private)
941 struct rrpc
*rrpc
= (struct rrpc
*)private;
942 struct nvm_dev
*dev
= rrpc
->dev
;
943 struct rrpc_addr
*addr
= rrpc
->trans_map
+ slba
;
944 struct rrpc_rev_addr
*raddr
= rrpc
->rev_trans_map
;
945 sector_t max_pages
= dev
->total_pages
* (dev
->sec_size
>> 9);
946 u64 elba
= slba
+ nlb
;
949 if (unlikely(elba
> dev
->total_pages
)) {
950 pr_err("nvm: L2P data from device is out of bounds!\n");
954 for (i
= 0; i
< nlb
; i
++) {
955 u64 pba
= le64_to_cpu(entries
[i
]);
956 /* LNVM treats address-spaces as silos, LBA and PBA are
957 * equally large and zero-indexed.
959 if (unlikely(pba
>= max_pages
&& pba
!= U64_MAX
)) {
960 pr_err("nvm: L2P data entry is out of bounds!\n");
964 /* Address zero is a special one. The first page on a disk is
965 * protected. As it often holds internal device boot
972 raddr
[pba
].addr
= slba
+ i
;
978 static int rrpc_map_init(struct rrpc
*rrpc
)
980 struct nvm_dev
*dev
= rrpc
->dev
;
984 rrpc
->trans_map
= vzalloc(sizeof(struct rrpc_addr
) * rrpc
->nr_pages
);
985 if (!rrpc
->trans_map
)
988 rrpc
->rev_trans_map
= vmalloc(sizeof(struct rrpc_rev_addr
)
990 if (!rrpc
->rev_trans_map
)
993 for (i
= 0; i
< rrpc
->nr_pages
; i
++) {
994 struct rrpc_addr
*p
= &rrpc
->trans_map
[i
];
995 struct rrpc_rev_addr
*r
= &rrpc
->rev_trans_map
[i
];
997 p
->addr
= ADDR_EMPTY
;
998 r
->addr
= ADDR_EMPTY
;
1001 if (!dev
->ops
->get_l2p_tbl
)
1004 /* Bring up the mapping table from device */
1005 ret
= dev
->ops
->get_l2p_tbl(dev
->q
, 0, dev
->total_pages
,
1006 rrpc_l2p_update
, rrpc
);
1008 pr_err("nvm: rrpc: could not read L2P table.\n");
1016 /* Minimum pages needed within a lun */
1017 #define PAGE_POOL_SIZE 16
1018 #define ADDR_POOL_SIZE 64
1020 static int rrpc_core_init(struct rrpc
*rrpc
)
1022 down_write(&rrpc_lock
);
1023 if (!rrpc_gcb_cache
) {
1024 rrpc_gcb_cache
= kmem_cache_create("rrpc_gcb",
1025 sizeof(struct rrpc_block_gc
), 0, 0, NULL
);
1026 if (!rrpc_gcb_cache
) {
1027 up_write(&rrpc_lock
);
1031 rrpc_rq_cache
= kmem_cache_create("rrpc_rq",
1032 sizeof(struct nvm_rq
) + sizeof(struct rrpc_rq
),
1034 if (!rrpc_rq_cache
) {
1035 kmem_cache_destroy(rrpc_gcb_cache
);
1036 up_write(&rrpc_lock
);
1040 up_write(&rrpc_lock
);
1042 rrpc
->page_pool
= mempool_create_page_pool(PAGE_POOL_SIZE
, 0);
1043 if (!rrpc
->page_pool
)
1046 rrpc
->gcb_pool
= mempool_create_slab_pool(rrpc
->dev
->nr_luns
,
1048 if (!rrpc
->gcb_pool
)
1051 rrpc
->rq_pool
= mempool_create_slab_pool(64, rrpc_rq_cache
);
1055 spin_lock_init(&rrpc
->inflights
.lock
);
1056 INIT_LIST_HEAD(&rrpc
->inflights
.reqs
);
1061 static void rrpc_core_free(struct rrpc
*rrpc
)
1063 mempool_destroy(rrpc
->page_pool
);
1064 mempool_destroy(rrpc
->gcb_pool
);
1065 mempool_destroy(rrpc
->rq_pool
);
1068 static void rrpc_luns_free(struct rrpc
*rrpc
)
1073 static int rrpc_luns_init(struct rrpc
*rrpc
, int lun_begin
, int lun_end
)
1075 struct nvm_dev
*dev
= rrpc
->dev
;
1076 struct rrpc_lun
*rlun
;
1079 spin_lock_init(&rrpc
->rev_lock
);
1081 rrpc
->luns
= kcalloc(rrpc
->nr_luns
, sizeof(struct rrpc_lun
),
1087 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
1088 struct nvm_lun
*lun
= dev
->mt
->get_lun(dev
, lun_begin
+ i
);
1090 if (dev
->pgs_per_blk
>
1091 MAX_INVALID_PAGES_STORAGE
* BITS_PER_LONG
) {
1092 pr_err("rrpc: number of pages per block too high.");
1096 rlun
= &rrpc
->luns
[i
];
1099 INIT_LIST_HEAD(&rlun
->prio_list
);
1100 INIT_WORK(&rlun
->ws_gc
, rrpc_lun_gc
);
1101 spin_lock_init(&rlun
->lock
);
1103 rrpc
->total_blocks
+= dev
->blks_per_lun
;
1104 rrpc
->nr_pages
+= dev
->sec_per_lun
;
1106 rlun
->blocks
= vzalloc(sizeof(struct rrpc_block
) *
1107 rrpc
->dev
->blks_per_lun
);
1111 for (j
= 0; j
< rrpc
->dev
->blks_per_lun
; j
++) {
1112 struct rrpc_block
*rblk
= &rlun
->blocks
[j
];
1113 struct nvm_block
*blk
= &lun
->blocks
[j
];
1116 INIT_LIST_HEAD(&rblk
->prio
);
1117 spin_lock_init(&rblk
->lock
);
1126 static void rrpc_free(struct rrpc
*rrpc
)
1129 rrpc_map_free(rrpc
);
1130 rrpc_core_free(rrpc
);
1131 rrpc_luns_free(rrpc
);
1136 static void rrpc_exit(void *private)
1138 struct rrpc
*rrpc
= private;
1140 del_timer(&rrpc
->gc_timer
);
1142 flush_workqueue(rrpc
->krqd_wq
);
1143 flush_workqueue(rrpc
->kgc_wq
);
1148 static sector_t
rrpc_capacity(void *private)
1150 struct rrpc
*rrpc
= private;
1151 struct nvm_dev
*dev
= rrpc
->dev
;
1152 sector_t reserved
, provisioned
;
1154 /* cur, gc, and two emergency blocks for each lun */
1155 reserved
= rrpc
->nr_luns
* dev
->max_pages_per_blk
* 4;
1156 provisioned
= rrpc
->nr_pages
- reserved
;
1158 if (reserved
> rrpc
->nr_pages
) {
1159 pr_err("rrpc: not enough space available to expose storage.\n");
1163 sector_div(provisioned
, 10);
1164 return provisioned
* 9 * NR_PHY_IN_LOG
;
1168 * Looks up the logical address from reverse trans map and check if its valid by
1169 * comparing the logical to physical address with the physical address.
1170 * Returns 0 on free, otherwise 1 if in use
1172 static void rrpc_block_map_update(struct rrpc
*rrpc
, struct rrpc_block
*rblk
)
1174 struct nvm_dev
*dev
= rrpc
->dev
;
1176 struct rrpc_addr
*laddr
;
1179 for (offset
= 0; offset
< dev
->pgs_per_blk
; offset
++) {
1180 paddr
= block_to_addr(rrpc
, rblk
) + offset
;
1182 pladdr
= rrpc
->rev_trans_map
[paddr
].addr
;
1183 if (pladdr
== ADDR_EMPTY
)
1186 laddr
= &rrpc
->trans_map
[pladdr
];
1188 if (paddr
== laddr
->addr
) {
1191 set_bit(offset
, rblk
->invalid_pages
);
1192 rblk
->nr_invalid_pages
++;
1197 static int rrpc_blocks_init(struct rrpc
*rrpc
)
1199 struct rrpc_lun
*rlun
;
1200 struct rrpc_block
*rblk
;
1201 int lun_iter
, blk_iter
;
1203 for (lun_iter
= 0; lun_iter
< rrpc
->nr_luns
; lun_iter
++) {
1204 rlun
= &rrpc
->luns
[lun_iter
];
1206 for (blk_iter
= 0; blk_iter
< rrpc
->dev
->blks_per_lun
;
1208 rblk
= &rlun
->blocks
[blk_iter
];
1209 rrpc_block_map_update(rrpc
, rblk
);
1216 static int rrpc_luns_configure(struct rrpc
*rrpc
)
1218 struct rrpc_lun
*rlun
;
1219 struct rrpc_block
*rblk
;
1222 for (i
= 0; i
< rrpc
->nr_luns
; i
++) {
1223 rlun
= &rrpc
->luns
[i
];
1225 rblk
= rrpc_get_blk(rrpc
, rlun
, 0);
1229 rrpc_set_lun_cur(rlun
, rblk
);
1231 /* Emergency gc block */
1232 rblk
= rrpc_get_blk(rrpc
, rlun
, 1);
1235 rlun
->gc_cur
= rblk
;
1241 static struct nvm_tgt_type tt_rrpc
;
1243 static void *rrpc_init(struct nvm_dev
*dev
, struct gendisk
*tdisk
,
1244 int lun_begin
, int lun_end
)
1246 struct request_queue
*bqueue
= dev
->q
;
1247 struct request_queue
*tqueue
= tdisk
->queue
;
1251 if (!(dev
->identity
.dom
& NVM_RSP_L2P
)) {
1252 pr_err("nvm: rrpc: device does not support l2p (%x)\n",
1254 return ERR_PTR(-EINVAL
);
1257 rrpc
= kzalloc(sizeof(struct rrpc
), GFP_KERNEL
);
1259 return ERR_PTR(-ENOMEM
);
1261 rrpc
->instance
.tt
= &tt_rrpc
;
1265 bio_list_init(&rrpc
->requeue_bios
);
1266 spin_lock_init(&rrpc
->bio_lock
);
1267 INIT_WORK(&rrpc
->ws_requeue
, rrpc_requeue
);
1269 rrpc
->nr_luns
= lun_end
- lun_begin
+ 1;
1271 /* simple round-robin strategy */
1272 atomic_set(&rrpc
->next_lun
, -1);
1274 ret
= rrpc_luns_init(rrpc
, lun_begin
, lun_end
);
1276 pr_err("nvm: rrpc: could not initialize luns\n");
1280 rrpc
->poffset
= dev
->sec_per_lun
* lun_begin
;
1281 rrpc
->lun_offset
= lun_begin
;
1283 ret
= rrpc_core_init(rrpc
);
1285 pr_err("nvm: rrpc: could not initialize core\n");
1289 ret
= rrpc_map_init(rrpc
);
1291 pr_err("nvm: rrpc: could not initialize maps\n");
1295 ret
= rrpc_blocks_init(rrpc
);
1297 pr_err("nvm: rrpc: could not initialize state for blocks\n");
1301 ret
= rrpc_luns_configure(rrpc
);
1303 pr_err("nvm: rrpc: not enough blocks available in LUNs.\n");
1307 ret
= rrpc_gc_init(rrpc
);
1309 pr_err("nvm: rrpc: could not initialize gc\n");
1313 /* inherit the size from the underlying device */
1314 blk_queue_logical_block_size(tqueue
, queue_physical_block_size(bqueue
));
1315 blk_queue_max_hw_sectors(tqueue
, queue_max_hw_sectors(bqueue
));
1317 pr_info("nvm: rrpc initialized with %u luns and %llu pages.\n",
1318 rrpc
->nr_luns
, (unsigned long long)rrpc
->nr_pages
);
1320 mod_timer(&rrpc
->gc_timer
, jiffies
+ msecs_to_jiffies(10));
1325 return ERR_PTR(ret
);
1328 /* round robin, page-based FTL, and cost-based GC */
1329 static struct nvm_tgt_type tt_rrpc
= {
1331 .version
= {1, 0, 0},
1333 .make_rq
= rrpc_make_rq
,
1334 .capacity
= rrpc_capacity
,
1335 .end_io
= rrpc_end_io
,
1341 static int __init
rrpc_module_init(void)
1343 return nvm_register_target(&tt_rrpc
);
1346 static void rrpc_module_exit(void)
1348 nvm_unregister_target(&tt_rrpc
);
1351 module_init(rrpc_module_init
);
1352 module_exit(rrpc_module_exit
);
1353 MODULE_LICENSE("GPL v2");
1354 MODULE_DESCRIPTION("Block-Device Target for Open-Channel SSDs");