2 * Copyright (C) 2016 CNEX Labs
3 * Initial release: Javier Gonzalez <javier@cnexlabs.com>
4 * Matias Bjorling <matias@cnexlabs.com>
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.
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.
15 * pblk-write.c - pblk's write path from write buffer to media
20 static unsigned long pblk_end_w_bio(struct pblk *pblk, struct nvm_rq *rqd,
21 struct pblk_c_ctx *c_ctx)
23 struct nvm_tgt_dev *dev = pblk->dev;
24 struct bio *original_bio;
28 for (i = 0; i < c_ctx->nr_valid; i++) {
29 struct pblk_w_ctx *w_ctx;
31 w_ctx = pblk_rb_w_ctx(&pblk->rwb, c_ctx->sentry + i);
32 while ((original_bio = bio_list_pop(&w_ctx->bios)))
33 bio_endio(original_bio);
36 #ifdef CONFIG_NVM_DEBUG
37 atomic_long_add(c_ctx->nr_valid, &pblk->sync_writes);
40 ret = pblk_rb_sync_advance(&pblk->rwb, c_ctx->nr_valid);
42 nvm_dev_dma_free(dev->parent, rqd->meta_list, rqd->dma_meta_list);
45 pblk_free_rqd(pblk, rqd, WRITE);
50 static unsigned long pblk_end_queued_w_bio(struct pblk *pblk,
52 struct pblk_c_ctx *c_ctx)
54 list_del(&c_ctx->list);
55 return pblk_end_w_bio(pblk, rqd, c_ctx);
58 static void pblk_complete_write(struct pblk *pblk, struct nvm_rq *rqd,
59 struct pblk_c_ctx *c_ctx)
61 struct pblk_c_ctx *c, *r;
65 #ifdef CONFIG_NVM_DEBUG
66 atomic_long_sub(c_ctx->nr_valid, &pblk->inflight_writes);
69 pblk_up_rq(pblk, rqd->ppa_list, rqd->nr_ppas, c_ctx->lun_bitmap);
71 pos = pblk_rb_sync_init(&pblk->rwb, &flags);
72 if (pos == c_ctx->sentry) {
73 pos = pblk_end_w_bio(pblk, rqd, c_ctx);
76 list_for_each_entry_safe(c, r, &pblk->compl_list, list) {
77 rqd = nvm_rq_from_c_ctx(c);
78 if (c->sentry == pos) {
79 pos = pblk_end_queued_w_bio(pblk, rqd, c);
84 WARN_ON(nvm_rq_from_c_ctx(c_ctx) != rqd);
85 list_add_tail(&c_ctx->list, &pblk->compl_list);
87 pblk_rb_sync_end(&pblk->rwb, &flags);
90 /* When a write fails, we are not sure whether the block has grown bad or a page
91 * range is more susceptible to write errors. If a high number of pages fail, we
92 * assume that the block is bad and we mark it accordingly. In all cases, we
93 * remap and resubmit the failed entries as fast as possible; if a flush is
94 * waiting on a completion, the whole stack would stall otherwise.
96 static void pblk_end_w_fail(struct pblk *pblk, struct nvm_rq *rqd)
98 void *comp_bits = &rqd->ppa_status;
99 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
100 struct pblk_rec_ctx *recovery;
101 struct ppa_addr *ppa_list = rqd->ppa_list;
102 int nr_ppas = rqd->nr_ppas;
103 unsigned int c_entries;
106 if (unlikely(nr_ppas == 1))
107 ppa_list = &rqd->ppa_addr;
109 recovery = mempool_alloc(pblk->rec_pool, GFP_ATOMIC);
111 pr_err("pblk: could not allocate recovery context\n");
114 INIT_LIST_HEAD(&recovery->failed);
117 while ((bit = find_next_bit(comp_bits, nr_ppas, bit + 1)) < nr_ppas) {
118 struct pblk_rb_entry *entry;
122 if (bit > c_ctx->nr_valid) {
123 WARN_ONCE(1, "pblk: corrupted write request\n");
124 mempool_free(recovery, pblk->rec_pool);
129 entry = pblk_rb_sync_scan_entry(&pblk->rwb, &ppa);
131 pr_err("pblk: could not scan entry on write failure\n");
132 mempool_free(recovery, pblk->rec_pool);
136 /* The list is filled first and emptied afterwards. No need for
137 * protecting it with a lock
139 list_add_tail(&entry->index, &recovery->failed);
142 c_entries = find_first_bit(comp_bits, nr_ppas);
143 ret = pblk_recov_setup_rq(pblk, c_ctx, recovery, comp_bits, c_entries);
145 pr_err("pblk: could not recover from write failure\n");
146 mempool_free(recovery, pblk->rec_pool);
150 INIT_WORK(&recovery->ws_rec, pblk_submit_rec);
151 queue_work(pblk->close_wq, &recovery->ws_rec);
154 pblk_complete_write(pblk, rqd, c_ctx);
157 static void pblk_end_io_write(struct nvm_rq *rqd)
159 struct pblk *pblk = rqd->private;
160 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
163 pblk_log_write_err(pblk, rqd);
164 return pblk_end_w_fail(pblk, rqd);
166 #ifdef CONFIG_NVM_DEBUG
168 WARN_ONCE(rqd->bio->bi_status, "pblk: corrupted write error\n");
171 pblk_complete_write(pblk, rqd, c_ctx);
172 atomic_dec(&pblk->inflight_io);
175 static void pblk_end_io_write_meta(struct nvm_rq *rqd)
177 struct pblk *pblk = rqd->private;
178 struct nvm_tgt_dev *dev = pblk->dev;
179 struct pblk_g_ctx *m_ctx = nvm_rq_to_pdu(rqd);
180 struct pblk_line *line = m_ctx->private;
181 struct pblk_emeta *emeta = line->emeta;
184 pblk_up_page(pblk, rqd->ppa_list, rqd->nr_ppas);
187 pblk_log_write_err(pblk, rqd);
188 pr_err("pblk: metadata I/O failed. Line %d\n", line->id);
190 #ifdef CONFIG_NVM_DEBUG
192 WARN_ONCE(rqd->bio->bi_status, "pblk: corrupted write error\n");
195 sync = atomic_add_return(rqd->nr_ppas, &emeta->sync);
196 if (sync == emeta->nr_entries)
197 pblk_line_run_ws(pblk, line, NULL, pblk_line_close_ws,
201 nvm_dev_dma_free(dev->parent, rqd->meta_list, rqd->dma_meta_list);
202 pblk_free_rqd(pblk, rqd, READ);
204 atomic_dec(&pblk->inflight_io);
207 static int pblk_alloc_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
208 unsigned int nr_secs,
209 nvm_end_io_fn(*end_io))
211 struct nvm_tgt_dev *dev = pblk->dev;
213 /* Setup write request */
214 rqd->opcode = NVM_OP_PWRITE;
215 rqd->nr_ppas = nr_secs;
216 rqd->flags = pblk_set_progr_mode(pblk, WRITE);
218 rqd->end_io = end_io;
220 rqd->meta_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
221 &rqd->dma_meta_list);
225 rqd->ppa_list = rqd->meta_list + pblk_dma_meta_size;
226 rqd->dma_ppa_list = rqd->dma_meta_list + pblk_dma_meta_size;
231 static int pblk_setup_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
232 struct pblk_c_ctx *c_ctx, struct ppa_addr *erase_ppa)
234 struct pblk_line_meta *lm = &pblk->lm;
235 struct pblk_line *e_line = pblk_line_get_erase(pblk);
236 unsigned int valid = c_ctx->nr_valid;
237 unsigned int padded = c_ctx->nr_padded;
238 unsigned int nr_secs = valid + padded;
239 unsigned long *lun_bitmap;
242 lun_bitmap = kzalloc(lm->lun_bitmap_len, GFP_KERNEL);
245 c_ctx->lun_bitmap = lun_bitmap;
247 ret = pblk_alloc_w_rq(pblk, rqd, nr_secs, pblk_end_io_write);
253 if (likely(!e_line || !atomic_read(&e_line->left_eblks)))
254 pblk_map_rq(pblk, rqd, c_ctx->sentry, lun_bitmap, valid, 0);
256 pblk_map_erase_rq(pblk, rqd, c_ctx->sentry, lun_bitmap,
262 int pblk_setup_w_rec_rq(struct pblk *pblk, struct nvm_rq *rqd,
263 struct pblk_c_ctx *c_ctx)
265 struct pblk_line_meta *lm = &pblk->lm;
266 unsigned long *lun_bitmap;
269 lun_bitmap = kzalloc(lm->lun_bitmap_len, GFP_KERNEL);
273 c_ctx->lun_bitmap = lun_bitmap;
275 ret = pblk_alloc_w_rq(pblk, rqd, rqd->nr_ppas, pblk_end_io_write);
279 pblk_map_rq(pblk, rqd, c_ctx->sentry, lun_bitmap, c_ctx->nr_valid, 0);
281 rqd->ppa_status = (u64)0;
282 rqd->flags = pblk_set_progr_mode(pblk, WRITE);
287 static int pblk_calc_secs_to_sync(struct pblk *pblk, unsigned int secs_avail,
288 unsigned int secs_to_flush)
292 secs_to_sync = pblk_calc_secs(pblk, secs_avail, secs_to_flush);
294 #ifdef CONFIG_NVM_DEBUG
295 if ((!secs_to_sync && secs_to_flush)
296 || (secs_to_sync < 0)
297 || (secs_to_sync > secs_avail && !secs_to_flush)) {
298 pr_err("pblk: bad sector calculation (a:%d,s:%d,f:%d)\n",
299 secs_avail, secs_to_sync, secs_to_flush);
306 static inline int pblk_valid_meta_ppa(struct pblk *pblk,
307 struct pblk_line *meta_line,
308 struct ppa_addr *ppa_list, int nr_ppas)
310 struct nvm_tgt_dev *dev = pblk->dev;
311 struct nvm_geo *geo = &dev->geo;
312 struct pblk_line *data_line;
313 struct ppa_addr ppa, ppa_opt;
317 data_line = &pblk->lines[pblk_dev_ppa_to_line(ppa_list[0])];
318 paddr = pblk_lookup_page(pblk, meta_line);
319 ppa = addr_to_gen_ppa(pblk, paddr, 0);
321 if (test_bit(pblk_ppa_to_pos(geo, ppa), data_line->blk_bitmap))
324 /* Schedule a metadata I/O that is half the distance from the data I/O
325 * with regards to the number of LUNs forming the pblk instance. This
326 * balances LUN conflicts across every I/O.
328 * When the LUN configuration changes (e.g., due to GC), this distance
329 * can align, which would result on a LUN deadlock. In this case, modify
330 * the distance to not be optimal, but allow metadata I/Os to succeed.
332 ppa_opt = addr_to_gen_ppa(pblk, paddr + data_line->meta_distance, 0);
333 if (unlikely(ppa_opt.ppa == ppa.ppa)) {
334 data_line->meta_distance--;
338 for (i = 0; i < nr_ppas; i += pblk->min_write_pgs)
339 if (ppa_list[i].g.ch == ppa_opt.g.ch &&
340 ppa_list[i].g.lun == ppa_opt.g.lun)
343 if (test_bit(pblk_ppa_to_pos(geo, ppa_opt), data_line->blk_bitmap)) {
344 for (i = 0; i < nr_ppas; i += pblk->min_write_pgs)
345 if (ppa_list[i].g.ch == ppa.g.ch &&
346 ppa_list[i].g.lun == ppa.g.lun)
355 int pblk_submit_meta_io(struct pblk *pblk, struct pblk_line *meta_line)
357 struct nvm_tgt_dev *dev = pblk->dev;
358 struct nvm_geo *geo = &dev->geo;
359 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
360 struct pblk_line_meta *lm = &pblk->lm;
361 struct pblk_emeta *emeta = meta_line->emeta;
362 struct pblk_g_ctx *m_ctx;
367 int rq_ppas = pblk->min_write_pgs;
368 int id = meta_line->id;
373 rqd = pblk_alloc_rqd(pblk, READ);
375 pr_err("pblk: cannot allocate write req.\n");
378 m_ctx = nvm_rq_to_pdu(rqd);
379 m_ctx->private = meta_line;
381 rq_len = rq_ppas * geo->sec_size;
382 data = ((void *)emeta->buf) + emeta->mem;
384 bio = pblk_bio_map_addr(pblk, data, rq_ppas, rq_len,
385 l_mg->emeta_alloc_type, GFP_KERNEL);
390 bio->bi_iter.bi_sector = 0; /* internal bio */
391 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
394 ret = pblk_alloc_w_rq(pblk, rqd, rq_ppas, pblk_end_io_write_meta);
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 rqd->ppa_list[i] = addr_to_gen_ppa(pblk, paddr, id);
406 emeta->mem += rq_len;
407 if (emeta->mem >= lm->emeta_len[0]) {
408 spin_lock(&l_mg->close_lock);
409 list_del(&meta_line->list);
410 WARN(!bitmap_full(meta_line->map_bitmap, lm->sec_per_line),
411 "pblk: corrupt meta line %d\n", meta_line->id);
412 spin_unlock(&l_mg->close_lock);
415 pblk_down_page(pblk, rqd->ppa_list, rqd->nr_ppas);
417 ret = pblk_submit_io(pblk, rqd);
419 pr_err("pblk: emeta I/O submission failed: %d\n", ret);
426 pblk_up_page(pblk, rqd->ppa_list, rqd->nr_ppas);
427 spin_lock(&l_mg->close_lock);
428 pblk_dealloc_page(pblk, meta_line, rq_ppas);
429 list_add(&meta_line->list, &meta_line->list);
430 spin_unlock(&l_mg->close_lock);
432 nvm_dev_dma_free(dev->parent, rqd->meta_list, rqd->dma_meta_list);
434 if (likely(l_mg->emeta_alloc_type == PBLK_VMALLOC_META))
437 pblk_free_rqd(pblk, rqd, READ);
441 static int pblk_sched_meta_io(struct pblk *pblk, struct ppa_addr *prev_list,
444 struct pblk_line_meta *lm = &pblk->lm;
445 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
446 struct pblk_line *meta_line;
448 spin_lock(&l_mg->close_lock);
450 if (list_empty(&l_mg->emeta_list)) {
451 spin_unlock(&l_mg->close_lock);
454 meta_line = list_first_entry(&l_mg->emeta_list, struct pblk_line, list);
455 if (bitmap_full(meta_line->map_bitmap, lm->sec_per_line))
457 spin_unlock(&l_mg->close_lock);
459 if (!pblk_valid_meta_ppa(pblk, meta_line, prev_list, prev_n))
462 return pblk_submit_meta_io(pblk, meta_line);
465 static int pblk_submit_io_set(struct pblk *pblk, struct nvm_rq *rqd)
467 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
468 struct ppa_addr erase_ppa;
471 ppa_set_empty(&erase_ppa);
473 /* Assign lbas to ppas and populate request structure */
474 err = pblk_setup_w_rq(pblk, rqd, c_ctx, &erase_ppa);
476 pr_err("pblk: could not setup write request: %d\n", err);
480 if (likely(ppa_empty(erase_ppa))) {
481 /* Submit metadata write for previous data line */
482 err = pblk_sched_meta_io(pblk, rqd->ppa_list, rqd->nr_ppas);
484 pr_err("pblk: metadata I/O submission failed: %d", err);
488 /* Submit data write for current data line */
489 err = pblk_submit_io(pblk, rqd);
491 pr_err("pblk: data I/O submission failed: %d\n", err);
495 /* Submit data write for current data line */
496 err = pblk_submit_io(pblk, rqd);
498 pr_err("pblk: data I/O submission failed: %d\n", err);
502 /* Submit available erase for next data line */
503 if (pblk_blk_erase_async(pblk, erase_ppa)) {
504 struct pblk_line *e_line = pblk_line_get_erase(pblk);
505 struct nvm_tgt_dev *dev = pblk->dev;
506 struct nvm_geo *geo = &dev->geo;
509 atomic_inc(&e_line->left_eblks);
510 bit = pblk_ppa_to_pos(geo, erase_ppa);
511 WARN_ON(!test_and_clear_bit(bit, e_line->erase_bitmap));
518 static void pblk_free_write_rqd(struct pblk *pblk, struct nvm_rq *rqd)
520 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
521 struct bio *bio = rqd->bio;
523 if (c_ctx->nr_padded)
524 pblk_bio_free_pages(pblk, bio, rqd->nr_ppas, c_ctx->nr_padded);
527 static int pblk_submit_write(struct pblk *pblk)
531 unsigned int secs_avail, secs_to_sync, secs_to_com;
532 unsigned int secs_to_flush;
535 /* If there are no sectors in the cache, flushes (bios without data)
536 * will be cleared on the cache threads
538 secs_avail = pblk_rb_read_count(&pblk->rwb);
542 secs_to_flush = pblk_rb_sync_point_count(&pblk->rwb);
543 if (!secs_to_flush && secs_avail < pblk->min_write_pgs)
546 rqd = pblk_alloc_rqd(pblk, WRITE);
548 pr_err("pblk: cannot allocate write req.\n");
552 bio = bio_alloc(GFP_KERNEL, pblk->max_write_pgs);
554 pr_err("pblk: cannot allocate write bio\n");
557 bio->bi_iter.bi_sector = 0; /* internal bio */
558 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
561 secs_to_sync = pblk_calc_secs_to_sync(pblk, secs_avail, secs_to_flush);
562 if (secs_to_sync > pblk->max_write_pgs) {
563 pr_err("pblk: bad buffer sync calculation\n");
567 secs_to_com = (secs_to_sync > secs_avail) ? secs_avail : secs_to_sync;
568 pos = pblk_rb_read_commit(&pblk->rwb, secs_to_com);
570 if (pblk_rb_read_to_bio(&pblk->rwb, rqd, bio, pos, secs_to_sync,
572 pr_err("pblk: corrupted write bio\n");
576 if (pblk_submit_io_set(pblk, rqd))
579 #ifdef CONFIG_NVM_DEBUG
580 atomic_long_add(secs_to_sync, &pblk->sub_writes);
586 pblk_free_write_rqd(pblk, rqd);
590 pblk_free_rqd(pblk, rqd, WRITE);
595 int pblk_write_ts(void *data)
597 struct pblk *pblk = data;
599 while (!kthread_should_stop()) {
600 if (!pblk_submit_write(pblk))
602 set_current_state(TASK_INTERRUPTIBLE);