Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2016 CNEX Labs
4 * Initial release: Javier Gonzalez <javier@cnexlabs.com>
5 * Matias Bjorling <matias@cnexlabs.com>
6 *
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.
10 *
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.
15 *
16 * pblk-write.c - pblk's write path from write buffer to media
17 */
18
19#include "pblk.h"
20#include "pblk-trace.h"
21
22static unsigned long pblk_end_w_bio(struct pblk *pblk, struct nvm_rq *rqd,
23 struct pblk_c_ctx *c_ctx)
24{
25 struct bio *original_bio;
26 struct pblk_rb *rwb = &pblk->rwb;
27 unsigned long ret;
28 int i;
29
30 for (i = 0; i < c_ctx->nr_valid; i++) {
31 struct pblk_w_ctx *w_ctx;
32 int pos = c_ctx->sentry + i;
33 int flags;
34
35 w_ctx = pblk_rb_w_ctx(rwb, pos);
36 flags = READ_ONCE(w_ctx->flags);
37
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);
42
43#ifdef CONFIG_NVM_PBLK_DEBUG
44 atomic_dec(&rwb->inflight_flush_point);
45#endif
46 }
47
48 while ((original_bio = bio_list_pop(&w_ctx->bios)))
49 bio_endio(original_bio);
50 }
51
52 if (c_ctx->nr_padded)
53 pblk_bio_free_pages(pblk, rqd->bio, c_ctx->nr_valid,
54 c_ctx->nr_padded);
55
56#ifdef CONFIG_NVM_PBLK_DEBUG
57 atomic_long_add(rqd->nr_ppas, &pblk->sync_writes);
58#endif
59
60 ret = pblk_rb_sync_advance(&pblk->rwb, c_ctx->nr_valid);
61
62 bio_put(rqd->bio);
63 pblk_free_rqd(pblk, rqd, PBLK_WRITE);
64
65 return ret;
66}
67
68static unsigned long pblk_end_queued_w_bio(struct pblk *pblk,
69 struct nvm_rq *rqd,
70 struct pblk_c_ctx *c_ctx)
71{
72 list_del(&c_ctx->list);
73 return pblk_end_w_bio(pblk, rqd, c_ctx);
74}
75
76static void pblk_complete_write(struct pblk *pblk, struct nvm_rq *rqd,
77 struct pblk_c_ctx *c_ctx)
78{
79 struct pblk_c_ctx *c, *r;
80 unsigned long flags;
81 unsigned long pos;
82
83#ifdef CONFIG_NVM_PBLK_DEBUG
84 atomic_long_sub(c_ctx->nr_valid, &pblk->inflight_writes);
85#endif
86 pblk_up_rq(pblk, c_ctx->lun_bitmap);
87
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);
91
92retry:
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);
97 goto retry;
98 }
99 }
100 } else {
101 WARN_ON(nvm_rq_from_c_ctx(c_ctx) != rqd);
102 list_add_tail(&c_ctx->list, &pblk->compl_list);
103 }
104 pblk_rb_sync_end(&pblk->rwb, &flags);
105}
106
107/* Map remaining sectors in chunk, starting from ppa */
108static void pblk_map_remaining(struct pblk *pblk, struct ppa_addr *ppa,
109 int rqd_ppas)
110{
111 struct pblk_line *line;
112 struct ppa_addr map_ppa = *ppa;
113 __le64 addr_empty = cpu_to_le64(ADDR_EMPTY);
114 __le64 *lba_list;
115 u64 paddr;
116 int done = 0;
117 int n = 0;
118
119 line = pblk_ppa_to_line(pblk, *ppa);
120 lba_list = emeta_to_lbas(pblk, line->emeta->buf);
121
122 spin_lock(&line->lock);
123
124 while (!done) {
125 paddr = pblk_dev_ppa_to_line_addr(pblk, map_ppa);
126
127 if (!test_and_set_bit(paddr, line->map_bitmap))
128 line->left_msecs--;
129
130 if (n < rqd_ppas && lba_list[paddr] != addr_empty)
131 line->nr_valid_lbas--;
132
133 lba_list[paddr] = addr_empty;
134
135 if (!test_and_set_bit(paddr, line->invalid_bitmap))
136 le32_add_cpu(line->vsc, -1);
137
138 done = nvm_next_ppa_in_chk(pblk->dev, &map_ppa);
139
140 n++;
141 }
142
143 line->w_err_gc->has_write_err = 1;
144 spin_unlock(&line->lock);
145}
146
147static void pblk_prepare_resubmit(struct pblk *pblk, unsigned int sentry,
148 unsigned int nr_entries)
149{
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;
155 int flags;
156 unsigned int i;
157
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;
162
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;
168 }
169
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);
175
176 /* Decrease the reference count to the line as we will
177 * re-map these entries
178 */
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);
182 }
183 spin_unlock(&pblk->trans_lock);
184}
185
186static void pblk_queue_resubmit(struct pblk *pblk, struct pblk_c_ctx *c_ctx)
187{
188 struct pblk_c_ctx *r_ctx;
189
190 r_ctx = kzalloc(sizeof(struct pblk_c_ctx), GFP_KERNEL);
191 if (!r_ctx)
192 return;
193
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;
198
199 spin_lock(&pblk->resubmit_lock);
200 list_add_tail(&r_ctx->list, &pblk->resubmit_list);
201 spin_unlock(&pblk->resubmit_lock);
202
203#ifdef CONFIG_NVM_PBLK_DEBUG
204 atomic_long_add(c_ctx->nr_valid, &pblk->recov_writes);
205#endif
206}
207
208static void pblk_submit_rec(struct work_struct *work)
209{
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);
216
217 pblk_log_write_err(pblk, rqd);
218
219 pblk_map_remaining(pblk, ppa_list, rqd->nr_ppas);
220 pblk_queue_resubmit(pblk, c_ctx);
221
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,
225 c_ctx->nr_padded);
226 bio_put(rqd->bio);
227 pblk_free_rqd(pblk, rqd, PBLK_WRITE);
228 mempool_free(recovery, &pblk->rec_pool);
229
230 atomic_dec(&pblk->inflight_io);
231 pblk_write_kick(pblk);
232}
233
234
235static void pblk_end_w_fail(struct pblk *pblk, struct nvm_rq *rqd)
236{
237 struct pblk_rec_ctx *recovery;
238
239 recovery = mempool_alloc(&pblk->rec_pool, GFP_ATOMIC);
240 if (!recovery) {
241 pblk_err(pblk, "could not allocate recovery work\n");
242 return;
243 }
244
245 recovery->pblk = pblk;
246 recovery->rqd = rqd;
247
248 INIT_WORK(&recovery->ws_rec, pblk_submit_rec);
249 queue_work(pblk->close_wq, &recovery->ws_rec);
250}
251
252static void pblk_end_io_write(struct nvm_rq *rqd)
253{
254 struct pblk *pblk = rqd->private;
255 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
256
257 if (rqd->error) {
258 pblk_end_w_fail(pblk, rqd);
259 return;
260 } else {
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");
265#endif
266 }
267
268 pblk_complete_write(pblk, rqd, c_ctx);
269 atomic_dec(&pblk->inflight_io);
270}
271
272static void pblk_end_io_write_meta(struct nvm_rq *rqd)
273{
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);
279 int sync;
280
281 pblk_up_chunk(pblk, ppa_list[0]);
282
283 if (rqd->error) {
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;
287 } else {
288 if (trace_pblk_chunk_state_enabled())
289 pblk_check_chunk_state_update(pblk, rqd);
290 }
291
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);
296
297 pblk_free_rqd(pblk, rqd, PBLK_WRITE_INT);
298
299 atomic_dec(&pblk->inflight_io);
300}
301
302static int pblk_alloc_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
303 unsigned int nr_secs, nvm_end_io_fn(*end_io))
304{
305 /* Setup write request */
306 rqd->opcode = NVM_OP_PWRITE;
307 rqd->nr_ppas = nr_secs;
308 rqd->is_seq = 1;
309 rqd->private = pblk;
310 rqd->end_io = end_io;
311
312 return pblk_alloc_rqd_meta(pblk, rqd);
313}
314
315static int pblk_setup_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
316 struct ppa_addr *erase_ppa)
317{
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;
325 int ret;
326
327 lun_bitmap = kzalloc(lm->lun_bitmap_len, GFP_KERNEL);
328 if (!lun_bitmap)
329 return -ENOMEM;
330 c_ctx->lun_bitmap = lun_bitmap;
331
332 ret = pblk_alloc_w_rq(pblk, rqd, nr_secs, pblk_end_io_write);
333 if (ret) {
334 kfree(lun_bitmap);
335 return ret;
336 }
337
338 if (likely(!e_line || !atomic_read(&e_line->left_eblks)))
339 ret = pblk_map_rq(pblk, rqd, c_ctx->sentry, lun_bitmap,
340 valid, 0);
341 else
342 ret = pblk_map_erase_rq(pblk, rqd, c_ctx->sentry, lun_bitmap,
343 valid, erase_ppa);
344
345 return ret;
346}
347
348static int pblk_calc_secs_to_sync(struct pblk *pblk, unsigned int secs_avail,
349 unsigned int secs_to_flush)
350{
351 int secs_to_sync;
352
353 secs_to_sync = pblk_calc_secs(pblk, secs_avail, secs_to_flush, true);
354
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);
361 }
362#endif
363
364 return secs_to_sync;
365}
366
367int pblk_submit_meta_io(struct pblk *pblk, struct pblk_line *meta_line)
368{
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;
376 struct nvm_rq *rqd;
377 void *data;
378 u64 paddr;
379 int rq_ppas = pblk->min_write_pgs;
380 int id = meta_line->id;
381 int rq_len;
382 int i, j;
383 int ret;
384
385 rqd = pblk_alloc_rqd(pblk, PBLK_WRITE_INT);
386
387 m_ctx = nvm_rq_to_pdu(rqd);
388 m_ctx->private = meta_line;
389
390 rq_len = rq_ppas * geo->csecs;
391 data = ((void *)emeta->buf) + emeta->mem;
392
393 ret = pblk_alloc_w_rq(pblk, rqd, rq_ppas, pblk_end_io_write_meta);
394 if (ret)
395 goto fail_free_rqd;
396
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);
404 }
405
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);
411
412 pblk_down_chunk(pblk, ppa_list[0]);
413
414 ret = pblk_submit_io(pblk, rqd, data);
415 if (ret) {
416 pblk_err(pblk, "emeta I/O submission failed: %d\n", ret);
417 goto fail_rollback;
418 }
419
420 return NVM_IO_OK;
421
422fail_rollback:
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);
428fail_free_rqd:
429 pblk_free_rqd(pblk, rqd, PBLK_WRITE_INT);
430 return ret;
431}
432
433static inline bool pblk_valid_meta_ppa(struct pblk *pblk,
434 struct pblk_line *meta_line,
435 struct nvm_rq *data_rqd)
436{
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;
442 u64 paddr;
443 int pos_opt;
444
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.
448 *
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.
453 */
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);
458
459 if (test_bit(pos_opt, data_c_ctx->lun_bitmap) ||
460 test_bit(pos_opt, data_line->blk_bitmap))
461 return true;
462
463 if (unlikely(pblk_ppa_comp(ppa_opt, ppa)))
464 data_line->meta_distance--;
465
466 return false;
467}
468
469static struct pblk_line *pblk_should_submit_meta_io(struct pblk *pblk,
470 struct nvm_rq *data_rqd)
471{
472 struct pblk_line_meta *lm = &pblk->lm;
473 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
474 struct pblk_line *meta_line;
475
476 spin_lock(&l_mg->close_lock);
477 if (list_empty(&l_mg->emeta_list)) {
478 spin_unlock(&l_mg->close_lock);
479 return NULL;
480 }
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);
484 return NULL;
485 }
486 spin_unlock(&l_mg->close_lock);
487
488 if (!pblk_valid_meta_ppa(pblk, meta_line, data_rqd))
489 return NULL;
490
491 return meta_line;
492}
493
494static int pblk_submit_io_set(struct pblk *pblk, struct nvm_rq *rqd)
495{
496 struct ppa_addr erase_ppa;
497 struct pblk_line *meta_line;
498 int err;
499
500 pblk_ppa_set_empty(&erase_ppa);
501
502 /* Assign lbas to ppas and populate request structure */
503 err = pblk_setup_w_rq(pblk, rqd, &erase_ppa);
504 if (err) {
505 pblk_err(pblk, "could not setup write request: %d\n", err);
506 return NVM_IO_ERR;
507 }
508
509 meta_line = pblk_should_submit_meta_io(pblk, rqd);
510
511 /* Submit data write for current data line */
512 err = pblk_submit_io(pblk, rqd, NULL);
513 if (err) {
514 pblk_err(pblk, "data I/O submission failed: %d\n", err);
515 return NVM_IO_ERR;
516 }
517
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;
524 int bit;
525
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));
529 }
530 }
531
532 if (meta_line) {
533 /* Submit metadata write for previous data line */
534 err = pblk_submit_meta_io(pblk, meta_line);
535 if (err) {
536 pblk_err(pblk, "metadata I/O submission failed: %d",
537 err);
538 return NVM_IO_ERR;
539 }
540 }
541
542 return NVM_IO_OK;
543}
544
545static void pblk_free_write_rqd(struct pblk *pblk, struct nvm_rq *rqd)
546{
547 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
548 struct bio *bio = rqd->bio;
549
550 if (c_ctx->nr_padded)
551 pblk_bio_free_pages(pblk, bio, c_ctx->nr_valid,
552 c_ctx->nr_padded);
553}
554
555static int pblk_submit_write(struct pblk *pblk, int *secs_left)
556{
557 struct bio *bio;
558 struct nvm_rq *rqd;
559 unsigned int secs_avail, secs_to_sync, secs_to_com;
560 unsigned int secs_to_flush, packed_meta_pgs;
561 unsigned long pos;
562 unsigned int resubmit;
563
564 *secs_left = 0;
565
566 spin_lock(&pblk->resubmit_lock);
567 resubmit = !list_empty(&pblk->resubmit_list);
568 spin_unlock(&pblk->resubmit_lock);
569
570 /* Resubmit failed writes first */
571 if (resubmit) {
572 struct pblk_c_ctx *r_ctx;
573
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);
579
580 secs_avail = r_ctx->nr_valid;
581 pos = r_ctx->sentry;
582
583 pblk_prepare_resubmit(pblk, pos, secs_avail);
584 secs_to_sync = pblk_calc_secs_to_sync(pblk, secs_avail,
585 secs_avail);
586
587 kfree(r_ctx);
588 } else {
589 /* If there are no sectors in the cache,
590 * flushes (bios without data) will be cleared on
591 * the cache threads
592 */
593 secs_avail = pblk_rb_read_count(&pblk->rwb);
594 if (!secs_avail)
595 return 0;
596
597 secs_to_flush = pblk_rb_flush_point_count(&pblk->rwb);
598 if (!secs_to_flush && secs_avail < pblk->min_write_pgs_data)
599 return 0;
600
601 secs_to_sync = pblk_calc_secs_to_sync(pblk, secs_avail,
602 secs_to_flush);
603 if (secs_to_sync > pblk->max_write_pgs) {
604 pblk_err(pblk, "bad buffer sync calculation\n");
605 return 0;
606 }
607
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);
611 }
612
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);
615
616 bio->bi_iter.bi_sector = 0; /* internal bio */
617 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
618
619 rqd = pblk_alloc_rqd(pblk, PBLK_WRITE);
620 rqd->bio = bio;
621
622 if (pblk_rb_read_to_bio(&pblk->rwb, rqd, pos, secs_to_sync,
623 secs_avail)) {
624 pblk_err(pblk, "corrupted write bio\n");
625 goto fail_put_bio;
626 }
627
628 if (pblk_submit_io_set(pblk, rqd))
629 goto fail_free_bio;
630
631#ifdef CONFIG_NVM_PBLK_DEBUG
632 atomic_long_add(secs_to_sync, &pblk->sub_writes);
633#endif
634
635 *secs_left = 1;
636 return 0;
637
638fail_free_bio:
639 pblk_free_write_rqd(pblk, rqd);
640fail_put_bio:
641 bio_put(bio);
642 pblk_free_rqd(pblk, rqd, PBLK_WRITE);
643
644 return -EINTR;
645}
646
647int pblk_write_ts(void *data)
648{
649 struct pblk *pblk = data;
650 int secs_left;
651 int write_failure = 0;
652
653 while (!kthread_should_stop()) {
654 if (!write_failure) {
655 write_failure = pblk_submit_write(pblk, &secs_left);
656
657 if (secs_left)
658 continue;
659 }
660 set_current_state(TASK_INTERRUPTIBLE);
661 io_schedule();
662 }
663
664 return 0;
665}
1/*
2 * Copyright (C) 2016 CNEX Labs
3 * Initial release: Javier Gonzalez <javier@cnexlabs.com>
4 * Matias Bjorling <matias@cnexlabs.com>
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License version
8 * 2 as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License for more details.
14 *
15 * pblk-write.c - pblk's write path from write buffer to media
16 */
17
18#include "pblk.h"
19
20static unsigned long pblk_end_w_bio(struct pblk *pblk, struct nvm_rq *rqd,
21 struct pblk_c_ctx *c_ctx)
22{
23 struct bio *original_bio;
24 struct pblk_rb *rwb = &pblk->rwb;
25 unsigned long ret;
26 int i;
27
28 for (i = 0; i < c_ctx->nr_valid; i++) {
29 struct pblk_w_ctx *w_ctx;
30 int pos = c_ctx->sentry + i;
31 int flags;
32
33 w_ctx = pblk_rb_w_ctx(rwb, pos);
34 flags = READ_ONCE(w_ctx->flags);
35
36 if (flags & PBLK_FLUSH_ENTRY) {
37 flags &= ~PBLK_FLUSH_ENTRY;
38 /* Release flags on context. Protect from writes */
39 smp_store_release(&w_ctx->flags, flags);
40
41#ifdef CONFIG_NVM_DEBUG
42 atomic_dec(&rwb->inflight_flush_point);
43#endif
44 }
45
46 while ((original_bio = bio_list_pop(&w_ctx->bios)))
47 bio_endio(original_bio);
48 }
49
50 if (c_ctx->nr_padded)
51 pblk_bio_free_pages(pblk, rqd->bio, c_ctx->nr_valid,
52 c_ctx->nr_padded);
53
54#ifdef CONFIG_NVM_DEBUG
55 atomic_long_add(rqd->nr_ppas, &pblk->sync_writes);
56#endif
57
58 ret = pblk_rb_sync_advance(&pblk->rwb, c_ctx->nr_valid);
59
60 bio_put(rqd->bio);
61 pblk_free_rqd(pblk, rqd, PBLK_WRITE);
62
63 return ret;
64}
65
66static unsigned long pblk_end_queued_w_bio(struct pblk *pblk,
67 struct nvm_rq *rqd,
68 struct pblk_c_ctx *c_ctx)
69{
70 list_del(&c_ctx->list);
71 return pblk_end_w_bio(pblk, rqd, c_ctx);
72}
73
74static void pblk_complete_write(struct pblk *pblk, struct nvm_rq *rqd,
75 struct pblk_c_ctx *c_ctx)
76{
77 struct pblk_c_ctx *c, *r;
78 unsigned long flags;
79 unsigned long pos;
80
81#ifdef CONFIG_NVM_DEBUG
82 atomic_long_sub(c_ctx->nr_valid, &pblk->inflight_writes);
83#endif
84
85 pblk_up_rq(pblk, rqd->ppa_list, rqd->nr_ppas, c_ctx->lun_bitmap);
86
87 pos = pblk_rb_sync_init(&pblk->rwb, &flags);
88 if (pos == c_ctx->sentry) {
89 pos = pblk_end_w_bio(pblk, rqd, c_ctx);
90
91retry:
92 list_for_each_entry_safe(c, r, &pblk->compl_list, list) {
93 rqd = nvm_rq_from_c_ctx(c);
94 if (c->sentry == pos) {
95 pos = pblk_end_queued_w_bio(pblk, rqd, c);
96 goto retry;
97 }
98 }
99 } else {
100 WARN_ON(nvm_rq_from_c_ctx(c_ctx) != rqd);
101 list_add_tail(&c_ctx->list, &pblk->compl_list);
102 }
103 pblk_rb_sync_end(&pblk->rwb, &flags);
104}
105
106/* When a write fails, we are not sure whether the block has grown bad or a page
107 * range is more susceptible to write errors. If a high number of pages fail, we
108 * assume that the block is bad and we mark it accordingly. In all cases, we
109 * remap and resubmit the failed entries as fast as possible; if a flush is
110 * waiting on a completion, the whole stack would stall otherwise.
111 */
112static void pblk_end_w_fail(struct pblk *pblk, struct nvm_rq *rqd)
113{
114 void *comp_bits = &rqd->ppa_status;
115 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
116 struct pblk_rec_ctx *recovery;
117 struct ppa_addr *ppa_list = rqd->ppa_list;
118 int nr_ppas = rqd->nr_ppas;
119 unsigned int c_entries;
120 int bit, ret;
121
122 if (unlikely(nr_ppas == 1))
123 ppa_list = &rqd->ppa_addr;
124
125 recovery = mempool_alloc(pblk->rec_pool, GFP_ATOMIC);
126
127 INIT_LIST_HEAD(&recovery->failed);
128
129 bit = -1;
130 while ((bit = find_next_bit(comp_bits, nr_ppas, bit + 1)) < nr_ppas) {
131 struct pblk_rb_entry *entry;
132 struct ppa_addr ppa;
133
134 /* Logic error */
135 if (bit > c_ctx->nr_valid) {
136 WARN_ONCE(1, "pblk: corrupted write request\n");
137 mempool_free(recovery, pblk->rec_pool);
138 goto out;
139 }
140
141 ppa = ppa_list[bit];
142 entry = pblk_rb_sync_scan_entry(&pblk->rwb, &ppa);
143 if (!entry) {
144 pr_err("pblk: could not scan entry on write failure\n");
145 mempool_free(recovery, pblk->rec_pool);
146 goto out;
147 }
148
149 /* The list is filled first and emptied afterwards. No need for
150 * protecting it with a lock
151 */
152 list_add_tail(&entry->index, &recovery->failed);
153 }
154
155 c_entries = find_first_bit(comp_bits, nr_ppas);
156 ret = pblk_recov_setup_rq(pblk, c_ctx, recovery, comp_bits, c_entries);
157 if (ret) {
158 pr_err("pblk: could not recover from write failure\n");
159 mempool_free(recovery, pblk->rec_pool);
160 goto out;
161 }
162
163 INIT_WORK(&recovery->ws_rec, pblk_submit_rec);
164 queue_work(pblk->close_wq, &recovery->ws_rec);
165
166out:
167 pblk_complete_write(pblk, rqd, c_ctx);
168}
169
170static void pblk_end_io_write(struct nvm_rq *rqd)
171{
172 struct pblk *pblk = rqd->private;
173 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
174
175 if (rqd->error) {
176 pblk_log_write_err(pblk, rqd);
177 return pblk_end_w_fail(pblk, rqd);
178 }
179#ifdef CONFIG_NVM_DEBUG
180 else
181 WARN_ONCE(rqd->bio->bi_status, "pblk: corrupted write error\n");
182#endif
183
184 pblk_complete_write(pblk, rqd, c_ctx);
185 atomic_dec(&pblk->inflight_io);
186}
187
188static void pblk_end_io_write_meta(struct nvm_rq *rqd)
189{
190 struct pblk *pblk = rqd->private;
191 struct pblk_g_ctx *m_ctx = nvm_rq_to_pdu(rqd);
192 struct pblk_line *line = m_ctx->private;
193 struct pblk_emeta *emeta = line->emeta;
194 int sync;
195
196 pblk_up_page(pblk, rqd->ppa_list, rqd->nr_ppas);
197
198 if (rqd->error) {
199 pblk_log_write_err(pblk, rqd);
200 pr_err("pblk: metadata I/O failed. Line %d\n", line->id);
201 }
202
203 sync = atomic_add_return(rqd->nr_ppas, &emeta->sync);
204 if (sync == emeta->nr_entries)
205 pblk_gen_run_ws(pblk, line, NULL, pblk_line_close_ws,
206 GFP_ATOMIC, pblk->close_wq);
207
208 pblk_free_rqd(pblk, rqd, PBLK_WRITE_INT);
209
210 atomic_dec(&pblk->inflight_io);
211}
212
213static int pblk_alloc_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
214 unsigned int nr_secs,
215 nvm_end_io_fn(*end_io))
216{
217 struct nvm_tgt_dev *dev = pblk->dev;
218
219 /* Setup write request */
220 rqd->opcode = NVM_OP_PWRITE;
221 rqd->nr_ppas = nr_secs;
222 rqd->flags = pblk_set_progr_mode(pblk, PBLK_WRITE);
223 rqd->private = pblk;
224 rqd->end_io = end_io;
225
226 rqd->meta_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
227 &rqd->dma_meta_list);
228 if (!rqd->meta_list)
229 return -ENOMEM;
230
231 rqd->ppa_list = rqd->meta_list + pblk_dma_meta_size;
232 rqd->dma_ppa_list = rqd->dma_meta_list + pblk_dma_meta_size;
233
234 return 0;
235}
236
237static int pblk_setup_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
238 struct ppa_addr *erase_ppa)
239{
240 struct pblk_line_meta *lm = &pblk->lm;
241 struct pblk_line *e_line = pblk_line_get_erase(pblk);
242 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
243 unsigned int valid = c_ctx->nr_valid;
244 unsigned int padded = c_ctx->nr_padded;
245 unsigned int nr_secs = valid + padded;
246 unsigned long *lun_bitmap;
247 int ret;
248
249 lun_bitmap = kzalloc(lm->lun_bitmap_len, GFP_KERNEL);
250 if (!lun_bitmap)
251 return -ENOMEM;
252 c_ctx->lun_bitmap = lun_bitmap;
253
254 ret = pblk_alloc_w_rq(pblk, rqd, nr_secs, pblk_end_io_write);
255 if (ret) {
256 kfree(lun_bitmap);
257 return ret;
258 }
259
260 if (likely(!e_line || !atomic_read(&e_line->left_eblks)))
261 pblk_map_rq(pblk, rqd, c_ctx->sentry, lun_bitmap, valid, 0);
262 else
263 pblk_map_erase_rq(pblk, rqd, c_ctx->sentry, lun_bitmap,
264 valid, erase_ppa);
265
266 return 0;
267}
268
269int pblk_setup_w_rec_rq(struct pblk *pblk, struct nvm_rq *rqd,
270 struct pblk_c_ctx *c_ctx)
271{
272 struct pblk_line_meta *lm = &pblk->lm;
273 unsigned long *lun_bitmap;
274 int ret;
275
276 lun_bitmap = kzalloc(lm->lun_bitmap_len, GFP_KERNEL);
277 if (!lun_bitmap)
278 return -ENOMEM;
279
280 c_ctx->lun_bitmap = lun_bitmap;
281
282 ret = pblk_alloc_w_rq(pblk, rqd, rqd->nr_ppas, pblk_end_io_write);
283 if (ret)
284 return ret;
285
286 pblk_map_rq(pblk, rqd, c_ctx->sentry, lun_bitmap, c_ctx->nr_valid, 0);
287
288 rqd->ppa_status = (u64)0;
289 rqd->flags = pblk_set_progr_mode(pblk, PBLK_WRITE);
290
291 return ret;
292}
293
294static int pblk_calc_secs_to_sync(struct pblk *pblk, unsigned int secs_avail,
295 unsigned int secs_to_flush)
296{
297 int secs_to_sync;
298
299 secs_to_sync = pblk_calc_secs(pblk, secs_avail, secs_to_flush);
300
301#ifdef CONFIG_NVM_DEBUG
302 if ((!secs_to_sync && secs_to_flush)
303 || (secs_to_sync < 0)
304 || (secs_to_sync > secs_avail && !secs_to_flush)) {
305 pr_err("pblk: bad sector calculation (a:%d,s:%d,f:%d)\n",
306 secs_avail, secs_to_sync, secs_to_flush);
307 }
308#endif
309
310 return secs_to_sync;
311}
312
313int pblk_submit_meta_io(struct pblk *pblk, struct pblk_line *meta_line)
314{
315 struct nvm_tgt_dev *dev = pblk->dev;
316 struct nvm_geo *geo = &dev->geo;
317 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
318 struct pblk_line_meta *lm = &pblk->lm;
319 struct pblk_emeta *emeta = meta_line->emeta;
320 struct pblk_g_ctx *m_ctx;
321 struct bio *bio;
322 struct nvm_rq *rqd;
323 void *data;
324 u64 paddr;
325 int rq_ppas = pblk->min_write_pgs;
326 int id = meta_line->id;
327 int rq_len;
328 int i, j;
329 int ret;
330
331 rqd = pblk_alloc_rqd(pblk, PBLK_WRITE_INT);
332
333 m_ctx = nvm_rq_to_pdu(rqd);
334 m_ctx->private = meta_line;
335
336 rq_len = rq_ppas * geo->csecs;
337 data = ((void *)emeta->buf) + emeta->mem;
338
339 bio = pblk_bio_map_addr(pblk, data, rq_ppas, rq_len,
340 l_mg->emeta_alloc_type, GFP_KERNEL);
341 if (IS_ERR(bio)) {
342 ret = PTR_ERR(bio);
343 goto fail_free_rqd;
344 }
345 bio->bi_iter.bi_sector = 0; /* internal bio */
346 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
347 rqd->bio = bio;
348
349 ret = pblk_alloc_w_rq(pblk, rqd, rq_ppas, pblk_end_io_write_meta);
350 if (ret)
351 goto fail_free_bio;
352
353 for (i = 0; i < rqd->nr_ppas; ) {
354 spin_lock(&meta_line->lock);
355 paddr = __pblk_alloc_page(pblk, meta_line, rq_ppas);
356 spin_unlock(&meta_line->lock);
357 for (j = 0; j < rq_ppas; j++, i++, paddr++)
358 rqd->ppa_list[i] = addr_to_gen_ppa(pblk, paddr, id);
359 }
360
361 emeta->mem += rq_len;
362 if (emeta->mem >= lm->emeta_len[0]) {
363 spin_lock(&l_mg->close_lock);
364 list_del(&meta_line->list);
365 spin_unlock(&l_mg->close_lock);
366 }
367
368 pblk_down_page(pblk, rqd->ppa_list, rqd->nr_ppas);
369
370 ret = pblk_submit_io(pblk, rqd);
371 if (ret) {
372 pr_err("pblk: emeta I/O submission failed: %d\n", ret);
373 goto fail_rollback;
374 }
375
376 return NVM_IO_OK;
377
378fail_rollback:
379 pblk_up_page(pblk, rqd->ppa_list, rqd->nr_ppas);
380 spin_lock(&l_mg->close_lock);
381 pblk_dealloc_page(pblk, meta_line, rq_ppas);
382 list_add(&meta_line->list, &meta_line->list);
383 spin_unlock(&l_mg->close_lock);
384fail_free_bio:
385 bio_put(bio);
386fail_free_rqd:
387 pblk_free_rqd(pblk, rqd, PBLK_WRITE_INT);
388 return ret;
389}
390
391static inline bool pblk_valid_meta_ppa(struct pblk *pblk,
392 struct pblk_line *meta_line,
393 struct nvm_rq *data_rqd)
394{
395 struct nvm_tgt_dev *dev = pblk->dev;
396 struct nvm_geo *geo = &dev->geo;
397 struct pblk_c_ctx *data_c_ctx = nvm_rq_to_pdu(data_rqd);
398 struct pblk_line *data_line = pblk_line_get_data(pblk);
399 struct ppa_addr ppa, ppa_opt;
400 u64 paddr;
401 int pos_opt;
402
403 /* Schedule a metadata I/O that is half the distance from the data I/O
404 * with regards to the number of LUNs forming the pblk instance. This
405 * balances LUN conflicts across every I/O.
406 *
407 * When the LUN configuration changes (e.g., due to GC), this distance
408 * can align, which would result on metadata and data I/Os colliding. In
409 * this case, modify the distance to not be optimal, but move the
410 * optimal in the right direction.
411 */
412 paddr = pblk_lookup_page(pblk, meta_line);
413 ppa = addr_to_gen_ppa(pblk, paddr, 0);
414 ppa_opt = addr_to_gen_ppa(pblk, paddr + data_line->meta_distance, 0);
415 pos_opt = pblk_ppa_to_pos(geo, ppa_opt);
416
417 if (test_bit(pos_opt, data_c_ctx->lun_bitmap) ||
418 test_bit(pos_opt, data_line->blk_bitmap))
419 return true;
420
421 if (unlikely(pblk_ppa_comp(ppa_opt, ppa)))
422 data_line->meta_distance--;
423
424 return false;
425}
426
427static struct pblk_line *pblk_should_submit_meta_io(struct pblk *pblk,
428 struct nvm_rq *data_rqd)
429{
430 struct pblk_line_meta *lm = &pblk->lm;
431 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
432 struct pblk_line *meta_line;
433
434 spin_lock(&l_mg->close_lock);
435retry:
436 if (list_empty(&l_mg->emeta_list)) {
437 spin_unlock(&l_mg->close_lock);
438 return NULL;
439 }
440 meta_line = list_first_entry(&l_mg->emeta_list, struct pblk_line, list);
441 if (meta_line->emeta->mem >= lm->emeta_len[0])
442 goto retry;
443 spin_unlock(&l_mg->close_lock);
444
445 if (!pblk_valid_meta_ppa(pblk, meta_line, data_rqd))
446 return NULL;
447
448 return meta_line;
449}
450
451static int pblk_submit_io_set(struct pblk *pblk, struct nvm_rq *rqd)
452{
453 struct ppa_addr erase_ppa;
454 struct pblk_line *meta_line;
455 int err;
456
457 pblk_ppa_set_empty(&erase_ppa);
458
459 /* Assign lbas to ppas and populate request structure */
460 err = pblk_setup_w_rq(pblk, rqd, &erase_ppa);
461 if (err) {
462 pr_err("pblk: could not setup write request: %d\n", err);
463 return NVM_IO_ERR;
464 }
465
466 meta_line = pblk_should_submit_meta_io(pblk, rqd);
467
468 /* Submit data write for current data line */
469 err = pblk_submit_io(pblk, rqd);
470 if (err) {
471 pr_err("pblk: data I/O submission failed: %d\n", err);
472 return NVM_IO_ERR;
473 }
474
475 if (!pblk_ppa_empty(erase_ppa)) {
476 /* Submit erase for next data line */
477 if (pblk_blk_erase_async(pblk, erase_ppa)) {
478 struct pblk_line *e_line = pblk_line_get_erase(pblk);
479 struct nvm_tgt_dev *dev = pblk->dev;
480 struct nvm_geo *geo = &dev->geo;
481 int bit;
482
483 atomic_inc(&e_line->left_eblks);
484 bit = pblk_ppa_to_pos(geo, erase_ppa);
485 WARN_ON(!test_and_clear_bit(bit, e_line->erase_bitmap));
486 }
487 }
488
489 if (meta_line) {
490 /* Submit metadata write for previous data line */
491 err = pblk_submit_meta_io(pblk, meta_line);
492 if (err) {
493 pr_err("pblk: metadata I/O submission failed: %d", err);
494 return NVM_IO_ERR;
495 }
496 }
497
498 return NVM_IO_OK;
499}
500
501static void pblk_free_write_rqd(struct pblk *pblk, struct nvm_rq *rqd)
502{
503 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
504 struct bio *bio = rqd->bio;
505
506 if (c_ctx->nr_padded)
507 pblk_bio_free_pages(pblk, bio, c_ctx->nr_valid,
508 c_ctx->nr_padded);
509}
510
511static int pblk_submit_write(struct pblk *pblk)
512{
513 struct bio *bio;
514 struct nvm_rq *rqd;
515 unsigned int secs_avail, secs_to_sync, secs_to_com;
516 unsigned int secs_to_flush;
517 unsigned long pos;
518
519 /* If there are no sectors in the cache, flushes (bios without data)
520 * will be cleared on the cache threads
521 */
522 secs_avail = pblk_rb_read_count(&pblk->rwb);
523 if (!secs_avail)
524 return 1;
525
526 secs_to_flush = pblk_rb_flush_point_count(&pblk->rwb);
527 if (!secs_to_flush && secs_avail < pblk->min_write_pgs)
528 return 1;
529
530 secs_to_sync = pblk_calc_secs_to_sync(pblk, secs_avail, secs_to_flush);
531 if (secs_to_sync > pblk->max_write_pgs) {
532 pr_err("pblk: bad buffer sync calculation\n");
533 return 1;
534 }
535
536 secs_to_com = (secs_to_sync > secs_avail) ? secs_avail : secs_to_sync;
537 pos = pblk_rb_read_commit(&pblk->rwb, secs_to_com);
538
539 bio = bio_alloc(GFP_KERNEL, secs_to_sync);
540
541 bio->bi_iter.bi_sector = 0; /* internal bio */
542 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
543
544 rqd = pblk_alloc_rqd(pblk, PBLK_WRITE);
545 rqd->bio = bio;
546
547 if (pblk_rb_read_to_bio(&pblk->rwb, rqd, pos, secs_to_sync,
548 secs_avail)) {
549 pr_err("pblk: corrupted write bio\n");
550 goto fail_put_bio;
551 }
552
553 if (pblk_submit_io_set(pblk, rqd))
554 goto fail_free_bio;
555
556#ifdef CONFIG_NVM_DEBUG
557 atomic_long_add(secs_to_sync, &pblk->sub_writes);
558#endif
559
560 return 0;
561
562fail_free_bio:
563 pblk_free_write_rqd(pblk, rqd);
564fail_put_bio:
565 bio_put(bio);
566 pblk_free_rqd(pblk, rqd, PBLK_WRITE);
567
568 return 1;
569}
570
571int pblk_write_ts(void *data)
572{
573 struct pblk *pblk = data;
574
575 while (!kthread_should_stop()) {
576 if (!pblk_submit_write(pblk))
577 continue;
578 set_current_state(TASK_INTERRUPTIBLE);
579 io_schedule();
580 }
581
582 return 0;
583}