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);
37 pblk_bio_free_pages(pblk, rqd->bio, c_ctx->nr_valid,
40 #ifdef CONFIG_NVM_DEBUG
41 atomic_long_add(c_ctx->nr_valid, &pblk->sync_writes);
44 ret = pblk_rb_sync_advance(&pblk->rwb, c_ctx->nr_valid);
46 nvm_dev_dma_free(dev->parent, rqd->meta_list, rqd->dma_meta_list);
49 pblk_free_rqd(pblk, rqd, WRITE);
54 static unsigned long pblk_end_queued_w_bio(struct pblk *pblk,
56 struct pblk_c_ctx *c_ctx)
58 list_del(&c_ctx->list);
59 return pblk_end_w_bio(pblk, rqd, c_ctx);
62 static void pblk_complete_write(struct pblk *pblk, struct nvm_rq *rqd,
63 struct pblk_c_ctx *c_ctx)
65 struct pblk_c_ctx *c, *r;
69 #ifdef CONFIG_NVM_DEBUG
70 atomic_long_sub(c_ctx->nr_valid, &pblk->inflight_writes);
73 pblk_up_rq(pblk, rqd->ppa_list, rqd->nr_ppas, c_ctx->lun_bitmap);
75 pos = pblk_rb_sync_init(&pblk->rwb, &flags);
76 if (pos == c_ctx->sentry) {
77 pos = pblk_end_w_bio(pblk, rqd, c_ctx);
80 list_for_each_entry_safe(c, r, &pblk->compl_list, list) {
81 rqd = nvm_rq_from_c_ctx(c);
82 if (c->sentry == pos) {
83 pos = pblk_end_queued_w_bio(pblk, rqd, c);
88 WARN_ON(nvm_rq_from_c_ctx(c_ctx) != rqd);
89 list_add_tail(&c_ctx->list, &pblk->compl_list);
91 pblk_rb_sync_end(&pblk->rwb, &flags);
94 /* When a write fails, we are not sure whether the block has grown bad or a page
95 * range is more susceptible to write errors. If a high number of pages fail, we
96 * assume that the block is bad and we mark it accordingly. In all cases, we
97 * remap and resubmit the failed entries as fast as possible; if a flush is
98 * waiting on a completion, the whole stack would stall otherwise.
100 static void pblk_end_w_fail(struct pblk *pblk, struct nvm_rq *rqd)
102 void *comp_bits = &rqd->ppa_status;
103 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
104 struct pblk_rec_ctx *recovery;
105 struct ppa_addr *ppa_list = rqd->ppa_list;
106 int nr_ppas = rqd->nr_ppas;
107 unsigned int c_entries;
110 if (unlikely(nr_ppas == 1))
111 ppa_list = &rqd->ppa_addr;
113 recovery = mempool_alloc(pblk->rec_pool, GFP_ATOMIC);
115 pr_err("pblk: could not allocate recovery context\n");
118 INIT_LIST_HEAD(&recovery->failed);
121 while ((bit = find_next_bit(comp_bits, nr_ppas, bit + 1)) < nr_ppas) {
122 struct pblk_rb_entry *entry;
126 if (bit > c_ctx->nr_valid) {
127 WARN_ONCE(1, "pblk: corrupted write request\n");
128 mempool_free(recovery, pblk->rec_pool);
133 entry = pblk_rb_sync_scan_entry(&pblk->rwb, &ppa);
135 pr_err("pblk: could not scan entry on write failure\n");
136 mempool_free(recovery, pblk->rec_pool);
140 /* The list is filled first and emptied afterwards. No need for
141 * protecting it with a lock
143 list_add_tail(&entry->index, &recovery->failed);
146 c_entries = find_first_bit(comp_bits, nr_ppas);
147 ret = pblk_recov_setup_rq(pblk, c_ctx, recovery, comp_bits, c_entries);
149 pr_err("pblk: could not recover from write failure\n");
150 mempool_free(recovery, pblk->rec_pool);
154 INIT_WORK(&recovery->ws_rec, pblk_submit_rec);
155 queue_work(pblk->close_wq, &recovery->ws_rec);
158 pblk_complete_write(pblk, rqd, c_ctx);
161 static void pblk_end_io_write(struct nvm_rq *rqd)
163 struct pblk *pblk = rqd->private;
164 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
167 pblk_log_write_err(pblk, rqd);
168 return pblk_end_w_fail(pblk, rqd);
170 #ifdef CONFIG_NVM_DEBUG
172 WARN_ONCE(rqd->bio->bi_status, "pblk: corrupted write error\n");
175 pblk_complete_write(pblk, rqd, c_ctx);
176 atomic_dec(&pblk->inflight_io);
179 static void pblk_end_io_write_meta(struct nvm_rq *rqd)
181 struct pblk *pblk = rqd->private;
182 struct nvm_tgt_dev *dev = pblk->dev;
183 struct pblk_g_ctx *m_ctx = nvm_rq_to_pdu(rqd);
184 struct pblk_line *line = m_ctx->private;
185 struct pblk_emeta *emeta = line->emeta;
188 pblk_up_page(pblk, rqd->ppa_list, rqd->nr_ppas);
191 pblk_log_write_err(pblk, rqd);
192 pr_err("pblk: metadata I/O failed. Line %d\n", line->id);
194 #ifdef CONFIG_NVM_DEBUG
196 WARN_ONCE(rqd->bio->bi_status, "pblk: corrupted write error\n");
199 sync = atomic_add_return(rqd->nr_ppas, &emeta->sync);
200 if (sync == emeta->nr_entries)
201 pblk_line_run_ws(pblk, line, NULL, pblk_line_close_ws,
205 nvm_dev_dma_free(dev->parent, rqd->meta_list, rqd->dma_meta_list);
206 pblk_free_rqd(pblk, rqd, READ);
208 atomic_dec(&pblk->inflight_io);
211 static int pblk_alloc_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
212 unsigned int nr_secs,
213 nvm_end_io_fn(*end_io))
215 struct nvm_tgt_dev *dev = pblk->dev;
217 /* Setup write request */
218 rqd->opcode = NVM_OP_PWRITE;
219 rqd->nr_ppas = nr_secs;
220 rqd->flags = pblk_set_progr_mode(pblk, WRITE);
222 rqd->end_io = end_io;
224 rqd->meta_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
225 &rqd->dma_meta_list);
229 rqd->ppa_list = rqd->meta_list + pblk_dma_meta_size;
230 rqd->dma_ppa_list = rqd->dma_meta_list + pblk_dma_meta_size;
235 static int pblk_setup_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
236 struct pblk_c_ctx *c_ctx, struct ppa_addr *erase_ppa)
238 struct pblk_line_meta *lm = &pblk->lm;
239 struct pblk_line *e_line = pblk_line_get_erase(pblk);
240 unsigned int valid = c_ctx->nr_valid;
241 unsigned int padded = c_ctx->nr_padded;
242 unsigned int nr_secs = valid + padded;
243 unsigned long *lun_bitmap;
246 lun_bitmap = kzalloc(lm->lun_bitmap_len, GFP_KERNEL);
249 c_ctx->lun_bitmap = lun_bitmap;
251 ret = pblk_alloc_w_rq(pblk, rqd, nr_secs, pblk_end_io_write);
257 if (likely(!e_line || !atomic_read(&e_line->left_eblks)))
258 pblk_map_rq(pblk, rqd, c_ctx->sentry, lun_bitmap, valid, 0);
260 pblk_map_erase_rq(pblk, rqd, c_ctx->sentry, lun_bitmap,
266 int pblk_setup_w_rec_rq(struct pblk *pblk, struct nvm_rq *rqd,
267 struct pblk_c_ctx *c_ctx)
269 struct pblk_line_meta *lm = &pblk->lm;
270 unsigned long *lun_bitmap;
273 lun_bitmap = kzalloc(lm->lun_bitmap_len, GFP_KERNEL);
277 c_ctx->lun_bitmap = lun_bitmap;
279 ret = pblk_alloc_w_rq(pblk, rqd, rqd->nr_ppas, pblk_end_io_write);
283 pblk_map_rq(pblk, rqd, c_ctx->sentry, lun_bitmap, c_ctx->nr_valid, 0);
285 rqd->ppa_status = (u64)0;
286 rqd->flags = pblk_set_progr_mode(pblk, WRITE);
291 static int pblk_calc_secs_to_sync(struct pblk *pblk, unsigned int secs_avail,
292 unsigned int secs_to_flush)
296 secs_to_sync = pblk_calc_secs(pblk, secs_avail, secs_to_flush);
298 #ifdef CONFIG_NVM_DEBUG
299 if ((!secs_to_sync && secs_to_flush)
300 || (secs_to_sync < 0)
301 || (secs_to_sync > secs_avail && !secs_to_flush)) {
302 pr_err("pblk: bad sector calculation (a:%d,s:%d,f:%d)\n",
303 secs_avail, secs_to_sync, secs_to_flush);
310 static inline int pblk_valid_meta_ppa(struct pblk *pblk,
311 struct pblk_line *meta_line,
312 struct ppa_addr *ppa_list, int nr_ppas)
314 struct nvm_tgt_dev *dev = pblk->dev;
315 struct nvm_geo *geo = &dev->geo;
316 struct pblk_line *data_line;
317 struct ppa_addr ppa, ppa_opt;
321 data_line = &pblk->lines[pblk_dev_ppa_to_line(ppa_list[0])];
322 paddr = pblk_lookup_page(pblk, meta_line);
323 ppa = addr_to_gen_ppa(pblk, paddr, 0);
325 if (test_bit(pblk_ppa_to_pos(geo, ppa), data_line->blk_bitmap))
328 /* Schedule a metadata I/O that is half the distance from the data I/O
329 * with regards to the number of LUNs forming the pblk instance. This
330 * balances LUN conflicts across every I/O.
332 * When the LUN configuration changes (e.g., due to GC), this distance
333 * can align, which would result on a LUN deadlock. In this case, modify
334 * the distance to not be optimal, but allow metadata I/Os to succeed.
336 ppa_opt = addr_to_gen_ppa(pblk, paddr + data_line->meta_distance, 0);
337 if (unlikely(ppa_opt.ppa == ppa.ppa)) {
338 data_line->meta_distance--;
342 for (i = 0; i < nr_ppas; i += pblk->min_write_pgs)
343 if (ppa_list[i].g.ch == ppa_opt.g.ch &&
344 ppa_list[i].g.lun == ppa_opt.g.lun)
347 if (test_bit(pblk_ppa_to_pos(geo, ppa_opt), data_line->blk_bitmap)) {
348 for (i = 0; i < nr_ppas; i += pblk->min_write_pgs)
349 if (ppa_list[i].g.ch == ppa.g.ch &&
350 ppa_list[i].g.lun == ppa.g.lun)
359 int pblk_submit_meta_io(struct pblk *pblk, struct pblk_line *meta_line)
361 struct nvm_tgt_dev *dev = pblk->dev;
362 struct nvm_geo *geo = &dev->geo;
363 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
364 struct pblk_line_meta *lm = &pblk->lm;
365 struct pblk_emeta *emeta = meta_line->emeta;
366 struct pblk_g_ctx *m_ctx;
371 int rq_ppas = pblk->min_write_pgs;
372 int id = meta_line->id;
377 rqd = pblk_alloc_rqd(pblk, READ);
379 pr_err("pblk: cannot allocate write req.\n");
382 m_ctx = nvm_rq_to_pdu(rqd);
383 m_ctx->private = meta_line;
385 rq_len = rq_ppas * geo->sec_size;
386 data = ((void *)emeta->buf) + emeta->mem;
388 bio = pblk_bio_map_addr(pblk, data, rq_ppas, rq_len,
389 l_mg->emeta_alloc_type, GFP_KERNEL);
394 bio->bi_iter.bi_sector = 0; /* internal bio */
395 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
398 ret = pblk_alloc_w_rq(pblk, rqd, rq_ppas, pblk_end_io_write_meta);
402 for (i = 0; i < rqd->nr_ppas; ) {
403 spin_lock(&meta_line->lock);
404 paddr = __pblk_alloc_page(pblk, meta_line, rq_ppas);
405 spin_unlock(&meta_line->lock);
406 for (j = 0; j < rq_ppas; j++, i++, paddr++)
407 rqd->ppa_list[i] = addr_to_gen_ppa(pblk, paddr, id);
410 emeta->mem += rq_len;
411 if (emeta->mem >= lm->emeta_len[0]) {
412 spin_lock(&l_mg->close_lock);
413 list_del(&meta_line->list);
414 WARN(!bitmap_full(meta_line->map_bitmap, lm->sec_per_line),
415 "pblk: corrupt meta line %d\n", meta_line->id);
416 spin_unlock(&l_mg->close_lock);
419 pblk_down_page(pblk, rqd->ppa_list, rqd->nr_ppas);
421 ret = pblk_submit_io(pblk, rqd);
423 pr_err("pblk: emeta I/O submission failed: %d\n", ret);
430 pblk_up_page(pblk, rqd->ppa_list, rqd->nr_ppas);
431 spin_lock(&l_mg->close_lock);
432 pblk_dealloc_page(pblk, meta_line, rq_ppas);
433 list_add(&meta_line->list, &meta_line->list);
434 spin_unlock(&l_mg->close_lock);
436 nvm_dev_dma_free(dev->parent, rqd->meta_list, rqd->dma_meta_list);
438 if (likely(l_mg->emeta_alloc_type == PBLK_VMALLOC_META))
441 pblk_free_rqd(pblk, rqd, READ);
445 static int pblk_sched_meta_io(struct pblk *pblk, struct ppa_addr *prev_list,
448 struct pblk_line_meta *lm = &pblk->lm;
449 struct pblk_line_mgmt *l_mg = &pblk->l_mg;
450 struct pblk_line *meta_line;
452 spin_lock(&l_mg->close_lock);
454 if (list_empty(&l_mg->emeta_list)) {
455 spin_unlock(&l_mg->close_lock);
458 meta_line = list_first_entry(&l_mg->emeta_list, struct pblk_line, list);
459 if (bitmap_full(meta_line->map_bitmap, lm->sec_per_line))
461 spin_unlock(&l_mg->close_lock);
463 if (!pblk_valid_meta_ppa(pblk, meta_line, prev_list, prev_n))
466 return pblk_submit_meta_io(pblk, meta_line);
469 static int pblk_submit_io_set(struct pblk *pblk, struct nvm_rq *rqd)
471 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
472 struct ppa_addr erase_ppa;
475 ppa_set_empty(&erase_ppa);
477 /* Assign lbas to ppas and populate request structure */
478 err = pblk_setup_w_rq(pblk, rqd, c_ctx, &erase_ppa);
480 pr_err("pblk: could not setup write request: %d\n", err);
484 if (likely(ppa_empty(erase_ppa))) {
485 /* Submit metadata write for previous data line */
486 err = pblk_sched_meta_io(pblk, rqd->ppa_list, rqd->nr_ppas);
488 pr_err("pblk: metadata I/O submission failed: %d", err);
492 /* Submit data write for current data line */
493 err = pblk_submit_io(pblk, rqd);
495 pr_err("pblk: data I/O submission failed: %d\n", err);
499 /* Submit data write for current data line */
500 err = pblk_submit_io(pblk, rqd);
502 pr_err("pblk: data I/O submission failed: %d\n", err);
506 /* Submit available erase for next data line */
507 if (pblk_blk_erase_async(pblk, erase_ppa)) {
508 struct pblk_line *e_line = pblk_line_get_erase(pblk);
509 struct nvm_tgt_dev *dev = pblk->dev;
510 struct nvm_geo *geo = &dev->geo;
513 atomic_inc(&e_line->left_eblks);
514 bit = pblk_ppa_to_pos(geo, erase_ppa);
515 WARN_ON(!test_and_clear_bit(bit, e_line->erase_bitmap));
522 static void pblk_free_write_rqd(struct pblk *pblk, struct nvm_rq *rqd)
524 struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
525 struct bio *bio = rqd->bio;
527 if (c_ctx->nr_padded)
528 pblk_bio_free_pages(pblk, bio, c_ctx->nr_valid,
532 static int pblk_submit_write(struct pblk *pblk)
536 unsigned int secs_avail, secs_to_sync, secs_to_com;
537 unsigned int secs_to_flush;
540 /* If there are no sectors in the cache, flushes (bios without data)
541 * will be cleared on the cache threads
543 secs_avail = pblk_rb_read_count(&pblk->rwb);
547 secs_to_flush = pblk_rb_sync_point_count(&pblk->rwb);
548 if (!secs_to_flush && secs_avail < pblk->min_write_pgs)
551 rqd = pblk_alloc_rqd(pblk, WRITE);
553 pr_err("pblk: cannot allocate write req.\n");
557 bio = bio_alloc(GFP_KERNEL, pblk->max_write_pgs);
559 pr_err("pblk: cannot allocate write bio\n");
562 bio->bi_iter.bi_sector = 0; /* internal bio */
563 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
566 secs_to_sync = pblk_calc_secs_to_sync(pblk, secs_avail, secs_to_flush);
567 if (secs_to_sync > pblk->max_write_pgs) {
568 pr_err("pblk: bad buffer sync calculation\n");
572 secs_to_com = (secs_to_sync > secs_avail) ? secs_avail : secs_to_sync;
573 pos = pblk_rb_read_commit(&pblk->rwb, secs_to_com);
575 if (pblk_rb_read_to_bio(&pblk->rwb, rqd, bio, pos, secs_to_sync,
577 pr_err("pblk: corrupted write bio\n");
581 if (pblk_submit_io_set(pblk, rqd))
584 #ifdef CONFIG_NVM_DEBUG
585 atomic_long_add(secs_to_sync, &pblk->sub_writes);
591 pblk_free_write_rqd(pblk, rqd);
595 pblk_free_rqd(pblk, rqd, WRITE);
600 int pblk_write_ts(void *data)
602 struct pblk *pblk = data;
604 while (!kthread_should_stop()) {
605 if (!pblk_submit_write(pblk))
607 set_current_state(TASK_INTERRUPTIBLE);