1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2016 CNEX Labs
4 * Initial release: Javier Gonzalez <javier@cnexlabs.com>
5 * Matias Bjorling <matias@cnexlabs.com>
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License version
9 * 2 as published by the Free Software Foundation.
11 * This program is distributed in the hope that it will be useful, but
12 * WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * General Public License for more details.
16 * pblk-write.c - pblk's write path from write buffer to media
20 #include "pblk-trace.h"
22 static unsigned long pblk_end_w_bio(struct pblk
*pblk
, struct nvm_rq
*rqd
,
23 struct pblk_c_ctx
*c_ctx
)
25 struct bio
*original_bio
;
26 struct pblk_rb
*rwb
= &pblk
->rwb
;
30 for (i
= 0; i
< c_ctx
->nr_valid
; i
++) {
31 struct pblk_w_ctx
*w_ctx
;
32 int pos
= c_ctx
->sentry
+ i
;
35 w_ctx
= pblk_rb_w_ctx(rwb
, pos
);
36 flags
= READ_ONCE(w_ctx
->flags
);
38 if (flags
& PBLK_FLUSH_ENTRY
) {
39 flags
&= ~PBLK_FLUSH_ENTRY
;
40 /* Release flags on context. Protect from writes */
41 smp_store_release(&w_ctx
->flags
, flags
);
43 #ifdef CONFIG_NVM_PBLK_DEBUG
44 atomic_dec(&rwb
->inflight_flush_point
);
48 while ((original_bio
= bio_list_pop(&w_ctx
->bios
)))
49 bio_endio(original_bio
);
53 pblk_bio_free_pages(pblk
, rqd
->bio
, c_ctx
->nr_valid
,
56 #ifdef CONFIG_NVM_PBLK_DEBUG
57 atomic_long_add(rqd
->nr_ppas
, &pblk
->sync_writes
);
60 ret
= pblk_rb_sync_advance(&pblk
->rwb
, c_ctx
->nr_valid
);
63 pblk_free_rqd(pblk
, rqd
, PBLK_WRITE
);
68 static unsigned long pblk_end_queued_w_bio(struct pblk
*pblk
,
70 struct pblk_c_ctx
*c_ctx
)
72 list_del(&c_ctx
->list
);
73 return pblk_end_w_bio(pblk
, rqd
, c_ctx
);
76 static void pblk_complete_write(struct pblk
*pblk
, struct nvm_rq
*rqd
,
77 struct pblk_c_ctx
*c_ctx
)
79 struct pblk_c_ctx
*c
, *r
;
83 #ifdef CONFIG_NVM_PBLK_DEBUG
84 atomic_long_sub(c_ctx
->nr_valid
, &pblk
->inflight_writes
);
86 pblk_up_rq(pblk
, c_ctx
->lun_bitmap
);
88 pos
= pblk_rb_sync_init(&pblk
->rwb
, &flags
);
89 if (pos
== c_ctx
->sentry
) {
90 pos
= pblk_end_w_bio(pblk
, rqd
, c_ctx
);
93 list_for_each_entry_safe(c
, r
, &pblk
->compl_list
, list
) {
94 rqd
= nvm_rq_from_c_ctx(c
);
95 if (c
->sentry
== pos
) {
96 pos
= pblk_end_queued_w_bio(pblk
, rqd
, c
);
101 WARN_ON(nvm_rq_from_c_ctx(c_ctx
) != rqd
);
102 list_add_tail(&c_ctx
->list
, &pblk
->compl_list
);
104 pblk_rb_sync_end(&pblk
->rwb
, &flags
);
107 /* Map remaining sectors in chunk, starting from ppa */
108 static void pblk_map_remaining(struct pblk
*pblk
, struct ppa_addr
*ppa
,
111 struct pblk_line
*line
;
112 struct ppa_addr map_ppa
= *ppa
;
113 __le64 addr_empty
= cpu_to_le64(ADDR_EMPTY
);
119 line
= pblk_ppa_to_line(pblk
, *ppa
);
120 lba_list
= emeta_to_lbas(pblk
, line
->emeta
->buf
);
122 spin_lock(&line
->lock
);
125 paddr
= pblk_dev_ppa_to_line_addr(pblk
, map_ppa
);
127 if (!test_and_set_bit(paddr
, line
->map_bitmap
))
130 if (n
< rqd_ppas
&& lba_list
[paddr
] != addr_empty
)
131 line
->nr_valid_lbas
--;
133 lba_list
[paddr
] = addr_empty
;
135 if (!test_and_set_bit(paddr
, line
->invalid_bitmap
))
136 le32_add_cpu(line
->vsc
, -1);
138 done
= nvm_next_ppa_in_chk(pblk
->dev
, &map_ppa
);
143 line
->w_err_gc
->has_write_err
= 1;
144 spin_unlock(&line
->lock
);
147 static void pblk_prepare_resubmit(struct pblk
*pblk
, unsigned int sentry
,
148 unsigned int nr_entries
)
150 struct pblk_rb
*rb
= &pblk
->rwb
;
151 struct pblk_rb_entry
*entry
;
152 struct pblk_line
*line
;
153 struct pblk_w_ctx
*w_ctx
;
154 struct ppa_addr ppa_l2p
;
158 spin_lock(&pblk
->trans_lock
);
159 for (i
= 0; i
< nr_entries
; i
++) {
160 entry
= &rb
->entries
[pblk_rb_ptr_wrap(rb
, sentry
, i
)];
161 w_ctx
= &entry
->w_ctx
;
163 /* Check if the lba has been overwritten */
164 if (w_ctx
->lba
!= ADDR_EMPTY
) {
165 ppa_l2p
= pblk_trans_map_get(pblk
, w_ctx
->lba
);
166 if (!pblk_ppa_comp(ppa_l2p
, entry
->cacheline
))
167 w_ctx
->lba
= ADDR_EMPTY
;
170 /* Mark up the entry as submittable again */
171 flags
= READ_ONCE(w_ctx
->flags
);
172 flags
|= PBLK_WRITTEN_DATA
;
173 /* Release flags on write context. Protect from writes */
174 smp_store_release(&w_ctx
->flags
, flags
);
176 /* Decrease the reference count to the line as we will
177 * re-map these entries
179 line
= pblk_ppa_to_line(pblk
, w_ctx
->ppa
);
180 atomic_dec(&line
->sec_to_update
);
181 kref_put(&line
->ref
, pblk_line_put
);
183 spin_unlock(&pblk
->trans_lock
);
186 static void pblk_queue_resubmit(struct pblk
*pblk
, struct pblk_c_ctx
*c_ctx
)
188 struct pblk_c_ctx
*r_ctx
;
190 r_ctx
= kzalloc(sizeof(struct pblk_c_ctx
), GFP_KERNEL
);
194 r_ctx
->lun_bitmap
= NULL
;
195 r_ctx
->sentry
= c_ctx
->sentry
;
196 r_ctx
->nr_valid
= c_ctx
->nr_valid
;
197 r_ctx
->nr_padded
= c_ctx
->nr_padded
;
199 spin_lock(&pblk
->resubmit_lock
);
200 list_add_tail(&r_ctx
->list
, &pblk
->resubmit_list
);
201 spin_unlock(&pblk
->resubmit_lock
);
203 #ifdef CONFIG_NVM_PBLK_DEBUG
204 atomic_long_add(c_ctx
->nr_valid
, &pblk
->recov_writes
);
208 static void pblk_submit_rec(struct work_struct
*work
)
210 struct pblk_rec_ctx
*recovery
=
211 container_of(work
, struct pblk_rec_ctx
, ws_rec
);
212 struct pblk
*pblk
= recovery
->pblk
;
213 struct nvm_rq
*rqd
= recovery
->rqd
;
214 struct pblk_c_ctx
*c_ctx
= nvm_rq_to_pdu(rqd
);
215 struct ppa_addr
*ppa_list
= nvm_rq_to_ppa_list(rqd
);
217 pblk_log_write_err(pblk
, rqd
);
219 pblk_map_remaining(pblk
, ppa_list
, rqd
->nr_ppas
);
220 pblk_queue_resubmit(pblk
, c_ctx
);
222 pblk_up_rq(pblk
, c_ctx
->lun_bitmap
);
223 if (c_ctx
->nr_padded
)
224 pblk_bio_free_pages(pblk
, rqd
->bio
, c_ctx
->nr_valid
,
227 pblk_free_rqd(pblk
, rqd
, PBLK_WRITE
);
228 mempool_free(recovery
, &pblk
->rec_pool
);
230 atomic_dec(&pblk
->inflight_io
);
231 pblk_write_kick(pblk
);
235 static void pblk_end_w_fail(struct pblk
*pblk
, struct nvm_rq
*rqd
)
237 struct pblk_rec_ctx
*recovery
;
239 recovery
= mempool_alloc(&pblk
->rec_pool
, GFP_ATOMIC
);
241 pblk_err(pblk
, "could not allocate recovery work\n");
245 recovery
->pblk
= pblk
;
248 INIT_WORK(&recovery
->ws_rec
, pblk_submit_rec
);
249 queue_work(pblk
->close_wq
, &recovery
->ws_rec
);
252 static void pblk_end_io_write(struct nvm_rq
*rqd
)
254 struct pblk
*pblk
= rqd
->private;
255 struct pblk_c_ctx
*c_ctx
= nvm_rq_to_pdu(rqd
);
258 pblk_end_w_fail(pblk
, rqd
);
261 if (trace_pblk_chunk_state_enabled())
262 pblk_check_chunk_state_update(pblk
, rqd
);
263 #ifdef CONFIG_NVM_PBLK_DEBUG
264 WARN_ONCE(rqd
->bio
->bi_status
, "pblk: corrupted write error\n");
268 pblk_complete_write(pblk
, rqd
, c_ctx
);
269 atomic_dec(&pblk
->inflight_io
);
272 static void pblk_end_io_write_meta(struct nvm_rq
*rqd
)
274 struct pblk
*pblk
= rqd
->private;
275 struct pblk_g_ctx
*m_ctx
= nvm_rq_to_pdu(rqd
);
276 struct pblk_line
*line
= m_ctx
->private;
277 struct pblk_emeta
*emeta
= line
->emeta
;
278 struct ppa_addr
*ppa_list
= nvm_rq_to_ppa_list(rqd
);
281 pblk_up_chunk(pblk
, ppa_list
[0]);
284 pblk_log_write_err(pblk
, rqd
);
285 pblk_err(pblk
, "metadata I/O failed. Line %d\n", line
->id
);
286 line
->w_err_gc
->has_write_err
= 1;
288 if (trace_pblk_chunk_state_enabled())
289 pblk_check_chunk_state_update(pblk
, rqd
);
292 sync
= atomic_add_return(rqd
->nr_ppas
, &emeta
->sync
);
293 if (sync
== emeta
->nr_entries
)
294 pblk_gen_run_ws(pblk
, line
, NULL
, pblk_line_close_ws
,
295 GFP_ATOMIC
, pblk
->close_wq
);
297 pblk_free_rqd(pblk
, rqd
, PBLK_WRITE_INT
);
299 atomic_dec(&pblk
->inflight_io
);
302 static int pblk_alloc_w_rq(struct pblk
*pblk
, struct nvm_rq
*rqd
,
303 unsigned int nr_secs
, nvm_end_io_fn(*end_io
))
305 /* Setup write request */
306 rqd
->opcode
= NVM_OP_PWRITE
;
307 rqd
->nr_ppas
= nr_secs
;
310 rqd
->end_io
= end_io
;
312 return pblk_alloc_rqd_meta(pblk
, rqd
);
315 static int pblk_setup_w_rq(struct pblk
*pblk
, struct nvm_rq
*rqd
,
316 struct ppa_addr
*erase_ppa
)
318 struct pblk_line_meta
*lm
= &pblk
->lm
;
319 struct pblk_line
*e_line
= pblk_line_get_erase(pblk
);
320 struct pblk_c_ctx
*c_ctx
= nvm_rq_to_pdu(rqd
);
321 unsigned int valid
= c_ctx
->nr_valid
;
322 unsigned int padded
= c_ctx
->nr_padded
;
323 unsigned int nr_secs
= valid
+ padded
;
324 unsigned long *lun_bitmap
;
327 lun_bitmap
= kzalloc(lm
->lun_bitmap_len
, GFP_KERNEL
);
330 c_ctx
->lun_bitmap
= lun_bitmap
;
332 ret
= pblk_alloc_w_rq(pblk
, rqd
, nr_secs
, pblk_end_io_write
);
338 if (likely(!e_line
|| !atomic_read(&e_line
->left_eblks
)))
339 ret
= pblk_map_rq(pblk
, rqd
, c_ctx
->sentry
, lun_bitmap
,
342 ret
= pblk_map_erase_rq(pblk
, rqd
, c_ctx
->sentry
, lun_bitmap
,
348 static int pblk_calc_secs_to_sync(struct pblk
*pblk
, unsigned int secs_avail
,
349 unsigned int secs_to_flush
)
353 secs_to_sync
= pblk_calc_secs(pblk
, secs_avail
, secs_to_flush
, true);
355 #ifdef CONFIG_NVM_PBLK_DEBUG
356 if ((!secs_to_sync
&& secs_to_flush
)
357 || (secs_to_sync
< 0)
358 || (secs_to_sync
> secs_avail
&& !secs_to_flush
)) {
359 pblk_err(pblk
, "bad sector calculation (a:%d,s:%d,f:%d)\n",
360 secs_avail
, secs_to_sync
, secs_to_flush
);
367 int pblk_submit_meta_io(struct pblk
*pblk
, struct pblk_line
*meta_line
)
369 struct nvm_tgt_dev
*dev
= pblk
->dev
;
370 struct nvm_geo
*geo
= &dev
->geo
;
371 struct pblk_line_mgmt
*l_mg
= &pblk
->l_mg
;
372 struct pblk_line_meta
*lm
= &pblk
->lm
;
373 struct pblk_emeta
*emeta
= meta_line
->emeta
;
374 struct ppa_addr
*ppa_list
;
375 struct pblk_g_ctx
*m_ctx
;
379 int rq_ppas
= pblk
->min_write_pgs
;
380 int id
= meta_line
->id
;
385 rqd
= pblk_alloc_rqd(pblk
, PBLK_WRITE_INT
);
387 m_ctx
= nvm_rq_to_pdu(rqd
);
388 m_ctx
->private = meta_line
;
390 rq_len
= rq_ppas
* geo
->csecs
;
391 data
= ((void *)emeta
->buf
) + emeta
->mem
;
393 ret
= pblk_alloc_w_rq(pblk
, rqd
, rq_ppas
, pblk_end_io_write_meta
);
397 ppa_list
= nvm_rq_to_ppa_list(rqd
);
398 for (i
= 0; i
< rqd
->nr_ppas
; ) {
399 spin_lock(&meta_line
->lock
);
400 paddr
= __pblk_alloc_page(pblk
, meta_line
, rq_ppas
);
401 spin_unlock(&meta_line
->lock
);
402 for (j
= 0; j
< rq_ppas
; j
++, i
++, paddr
++)
403 ppa_list
[i
] = addr_to_gen_ppa(pblk
, paddr
, id
);
406 spin_lock(&l_mg
->close_lock
);
407 emeta
->mem
+= rq_len
;
408 if (emeta
->mem
>= lm
->emeta_len
[0])
409 list_del(&meta_line
->list
);
410 spin_unlock(&l_mg
->close_lock
);
412 pblk_down_chunk(pblk
, ppa_list
[0]);
414 ret
= pblk_submit_io(pblk
, rqd
, data
);
416 pblk_err(pblk
, "emeta I/O submission failed: %d\n", ret
);
423 pblk_up_chunk(pblk
, ppa_list
[0]);
424 spin_lock(&l_mg
->close_lock
);
425 pblk_dealloc_page(pblk
, meta_line
, rq_ppas
);
426 list_add(&meta_line
->list
, &meta_line
->list
);
427 spin_unlock(&l_mg
->close_lock
);
429 pblk_free_rqd(pblk
, rqd
, PBLK_WRITE_INT
);
433 static inline bool pblk_valid_meta_ppa(struct pblk
*pblk
,
434 struct pblk_line
*meta_line
,
435 struct nvm_rq
*data_rqd
)
437 struct nvm_tgt_dev
*dev
= pblk
->dev
;
438 struct nvm_geo
*geo
= &dev
->geo
;
439 struct pblk_c_ctx
*data_c_ctx
= nvm_rq_to_pdu(data_rqd
);
440 struct pblk_line
*data_line
= pblk_line_get_data(pblk
);
441 struct ppa_addr ppa
, ppa_opt
;
445 /* Schedule a metadata I/O that is half the distance from the data I/O
446 * with regards to the number of LUNs forming the pblk instance. This
447 * balances LUN conflicts across every I/O.
449 * When the LUN configuration changes (e.g., due to GC), this distance
450 * can align, which would result on metadata and data I/Os colliding. In
451 * this case, modify the distance to not be optimal, but move the
452 * optimal in the right direction.
454 paddr
= pblk_lookup_page(pblk
, meta_line
);
455 ppa
= addr_to_gen_ppa(pblk
, paddr
, 0);
456 ppa_opt
= addr_to_gen_ppa(pblk
, paddr
+ data_line
->meta_distance
, 0);
457 pos_opt
= pblk_ppa_to_pos(geo
, ppa_opt
);
459 if (test_bit(pos_opt
, data_c_ctx
->lun_bitmap
) ||
460 test_bit(pos_opt
, data_line
->blk_bitmap
))
463 if (unlikely(pblk_ppa_comp(ppa_opt
, ppa
)))
464 data_line
->meta_distance
--;
469 static struct pblk_line
*pblk_should_submit_meta_io(struct pblk
*pblk
,
470 struct nvm_rq
*data_rqd
)
472 struct pblk_line_meta
*lm
= &pblk
->lm
;
473 struct pblk_line_mgmt
*l_mg
= &pblk
->l_mg
;
474 struct pblk_line
*meta_line
;
476 spin_lock(&l_mg
->close_lock
);
477 if (list_empty(&l_mg
->emeta_list
)) {
478 spin_unlock(&l_mg
->close_lock
);
481 meta_line
= list_first_entry(&l_mg
->emeta_list
, struct pblk_line
, list
);
482 if (meta_line
->emeta
->mem
>= lm
->emeta_len
[0]) {
483 spin_unlock(&l_mg
->close_lock
);
486 spin_unlock(&l_mg
->close_lock
);
488 if (!pblk_valid_meta_ppa(pblk
, meta_line
, data_rqd
))
494 static int pblk_submit_io_set(struct pblk
*pblk
, struct nvm_rq
*rqd
)
496 struct ppa_addr erase_ppa
;
497 struct pblk_line
*meta_line
;
500 pblk_ppa_set_empty(&erase_ppa
);
502 /* Assign lbas to ppas and populate request structure */
503 err
= pblk_setup_w_rq(pblk
, rqd
, &erase_ppa
);
505 pblk_err(pblk
, "could not setup write request: %d\n", err
);
509 meta_line
= pblk_should_submit_meta_io(pblk
, rqd
);
511 /* Submit data write for current data line */
512 err
= pblk_submit_io(pblk
, rqd
, NULL
);
514 pblk_err(pblk
, "data I/O submission failed: %d\n", err
);
518 if (!pblk_ppa_empty(erase_ppa
)) {
519 /* Submit erase for next data line */
520 if (pblk_blk_erase_async(pblk
, erase_ppa
)) {
521 struct pblk_line
*e_line
= pblk_line_get_erase(pblk
);
522 struct nvm_tgt_dev
*dev
= pblk
->dev
;
523 struct nvm_geo
*geo
= &dev
->geo
;
526 atomic_inc(&e_line
->left_eblks
);
527 bit
= pblk_ppa_to_pos(geo
, erase_ppa
);
528 WARN_ON(!test_and_clear_bit(bit
, e_line
->erase_bitmap
));
533 /* Submit metadata write for previous data line */
534 err
= pblk_submit_meta_io(pblk
, meta_line
);
536 pblk_err(pblk
, "metadata I/O submission failed: %d",
545 static void pblk_free_write_rqd(struct pblk
*pblk
, struct nvm_rq
*rqd
)
547 struct pblk_c_ctx
*c_ctx
= nvm_rq_to_pdu(rqd
);
548 struct bio
*bio
= rqd
->bio
;
550 if (c_ctx
->nr_padded
)
551 pblk_bio_free_pages(pblk
, bio
, c_ctx
->nr_valid
,
555 static int pblk_submit_write(struct pblk
*pblk
, int *secs_left
)
559 unsigned int secs_avail
, secs_to_sync
, secs_to_com
;
560 unsigned int secs_to_flush
, packed_meta_pgs
;
562 unsigned int resubmit
;
566 spin_lock(&pblk
->resubmit_lock
);
567 resubmit
= !list_empty(&pblk
->resubmit_list
);
568 spin_unlock(&pblk
->resubmit_lock
);
570 /* Resubmit failed writes first */
572 struct pblk_c_ctx
*r_ctx
;
574 spin_lock(&pblk
->resubmit_lock
);
575 r_ctx
= list_first_entry(&pblk
->resubmit_list
,
576 struct pblk_c_ctx
, list
);
577 list_del(&r_ctx
->list
);
578 spin_unlock(&pblk
->resubmit_lock
);
580 secs_avail
= r_ctx
->nr_valid
;
583 pblk_prepare_resubmit(pblk
, pos
, secs_avail
);
584 secs_to_sync
= pblk_calc_secs_to_sync(pblk
, secs_avail
,
589 /* If there are no sectors in the cache,
590 * flushes (bios without data) will be cleared on
593 secs_avail
= pblk_rb_read_count(&pblk
->rwb
);
597 secs_to_flush
= pblk_rb_flush_point_count(&pblk
->rwb
);
598 if (!secs_to_flush
&& secs_avail
< pblk
->min_write_pgs_data
)
601 secs_to_sync
= pblk_calc_secs_to_sync(pblk
, secs_avail
,
603 if (secs_to_sync
> pblk
->max_write_pgs
) {
604 pblk_err(pblk
, "bad buffer sync calculation\n");
608 secs_to_com
= (secs_to_sync
> secs_avail
) ?
609 secs_avail
: secs_to_sync
;
610 pos
= pblk_rb_read_commit(&pblk
->rwb
, secs_to_com
);
613 packed_meta_pgs
= (pblk
->min_write_pgs
- pblk
->min_write_pgs_data
);
614 bio
= bio_alloc(GFP_KERNEL
, secs_to_sync
+ packed_meta_pgs
);
616 bio
->bi_iter
.bi_sector
= 0; /* internal bio */
617 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
619 rqd
= pblk_alloc_rqd(pblk
, PBLK_WRITE
);
622 if (pblk_rb_read_to_bio(&pblk
->rwb
, rqd
, pos
, secs_to_sync
,
624 pblk_err(pblk
, "corrupted write bio\n");
628 if (pblk_submit_io_set(pblk
, rqd
))
631 #ifdef CONFIG_NVM_PBLK_DEBUG
632 atomic_long_add(secs_to_sync
, &pblk
->sub_writes
);
639 pblk_free_write_rqd(pblk
, rqd
);
642 pblk_free_rqd(pblk
, rqd
, PBLK_WRITE
);
647 int pblk_write_ts(void *data
)
649 struct pblk
*pblk
= data
;
651 int write_failure
= 0;
653 while (!kthread_should_stop()) {
654 if (!write_failure
) {
655 write_failure
= pblk_submit_write(pblk
, &secs_left
);
660 set_current_state(TASK_INTERRUPTIBLE
);