diff options
author | 2018-06-01 16:41:05 +0200 | |
---|---|---|
committer | 2018-06-01 09:02:53 -0600 | |
commit | 6a3abf5beef6ae46381c1fb6976e6f313c40f0c1 (patch) | |
tree | 0cb1da46fdb14e1a0de58d656331b0cc6c0c9082 /drivers/lightnvm/pblk-write.c | |
parent | lightnvm: pblk: remove dead function (diff) | |
download | linux-6a3abf5beef6ae46381c1fb6976e6f313c40f0c1.tar.gz linux-6a3abf5beef6ae46381c1fb6976e6f313c40f0c1.tar.bz2 linux-6a3abf5beef6ae46381c1fb6976e6f313c40f0c1.zip |
lightnvm: pblk: rework write error recovery path
The write error recovery path is incomplete, so rework
the write error recovery handling to do resubmits directly
from the write buffer.
When a write error occurs, the remaining sectors in the chunk are
mapped out and invalidated and the request inserted in a resubmit list.
The writer thread checks if there are any requests to resubmit,
scans and invalidates any lbas that have been overwritten by later
writes and resubmits the failed entries.
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Reviewed-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Diffstat (limited to 'drivers/lightnvm/pblk-write.c')
-rw-r--r-- | drivers/lightnvm/pblk-write.c | 267 |
1 files changed, 176 insertions, 91 deletions
diff --git a/drivers/lightnvm/pblk-write.c b/drivers/lightnvm/pblk-write.c index aef7fa2d401d..7f9491bc64b5 100644 --- a/drivers/lightnvm/pblk-write.c +++ b/drivers/lightnvm/pblk-write.c @@ -103,68 +103,149 @@ retry: pblk_rb_sync_end(&pblk->rwb, &flags); } -/* When a write fails, we are not sure whether the block has grown bad or a page - * range is more susceptible to write errors. If a high number of pages fail, we - * assume that the block is bad and we mark it accordingly. In all cases, we - * remap and resubmit the failed entries as fast as possible; if a flush is - * waiting on a completion, the whole stack would stall otherwise. - */ -static void pblk_end_w_fail(struct pblk *pblk, struct nvm_rq *rqd) +/* Map remaining sectors in chunk, starting from ppa */ +static void pblk_map_remaining(struct pblk *pblk, struct ppa_addr *ppa) { - void *comp_bits = &rqd->ppa_status; - struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd); - struct pblk_rec_ctx *recovery; - struct ppa_addr *ppa_list = rqd->ppa_list; - int nr_ppas = rqd->nr_ppas; - unsigned int c_entries; - int bit, ret; + struct nvm_tgt_dev *dev = pblk->dev; + struct nvm_geo *geo = &dev->geo; + struct pblk_line *line; + struct ppa_addr map_ppa = *ppa; + u64 paddr; + int done = 0; - if (unlikely(nr_ppas == 1)) - ppa_list = &rqd->ppa_addr; + line = &pblk->lines[pblk_ppa_to_line(*ppa)]; + spin_lock(&line->lock); - recovery = mempool_alloc(&pblk->rec_pool, GFP_ATOMIC); + while (!done) { + paddr = pblk_dev_ppa_to_line_addr(pblk, map_ppa); - INIT_LIST_HEAD(&recovery->failed); + if (!test_and_set_bit(paddr, line->map_bitmap)) + line->left_msecs--; - bit = -1; - while ((bit = find_next_bit(comp_bits, nr_ppas, bit + 1)) < nr_ppas) { - struct pblk_rb_entry *entry; - struct ppa_addr ppa; + if (!test_and_set_bit(paddr, line->invalid_bitmap)) + le32_add_cpu(line->vsc, -1); - /* Logic error */ - if (bit > c_ctx->nr_valid) { - WARN_ONCE(1, "pblk: corrupted write request\n"); - mempool_free(recovery, &pblk->rec_pool); - goto out; + if (geo->version == NVM_OCSSD_SPEC_12) { + map_ppa.ppa++; + if (map_ppa.g.pg == geo->num_pg) + done = 1; + } else { + map_ppa.m.sec++; + if (map_ppa.m.sec == geo->clba) + done = 1; } + } - ppa = ppa_list[bit]; - entry = pblk_rb_sync_scan_entry(&pblk->rwb, &ppa); - if (!entry) { - pr_err("pblk: could not scan entry on write failure\n"); - mempool_free(recovery, &pblk->rec_pool); - goto out; - } + spin_unlock(&line->lock); +} + +static void pblk_prepare_resubmit(struct pblk *pblk, unsigned int sentry, + unsigned int nr_entries) +{ + struct pblk_rb *rb = &pblk->rwb; + struct pblk_rb_entry *entry; + struct pblk_line *line; + struct pblk_w_ctx *w_ctx; + struct ppa_addr ppa_l2p; + int flags; + unsigned int pos, i; + + spin_lock(&pblk->trans_lock); + pos = sentry; + for (i = 0; i < nr_entries; i++) { + entry = &rb->entries[pos]; + w_ctx = &entry->w_ctx; + + /* Check if the lba has been overwritten */ + ppa_l2p = pblk_trans_map_get(pblk, w_ctx->lba); + if (!pblk_ppa_comp(ppa_l2p, entry->cacheline)) + w_ctx->lba = ADDR_EMPTY; + + /* Mark up the entry as submittable again */ + flags = READ_ONCE(w_ctx->flags); + flags |= PBLK_WRITTEN_DATA; + /* Release flags on write context. Protect from writes */ + smp_store_release(&w_ctx->flags, flags); - /* The list is filled first and emptied afterwards. No need for - * protecting it with a lock + /* Decrese the reference count to the line as we will + * re-map these entries */ - list_add_tail(&entry->index, &recovery->failed); + line = &pblk->lines[pblk_ppa_to_line(w_ctx->ppa)]; + kref_put(&line->ref, pblk_line_put); + + pos = (pos + 1) & (rb->nr_entries - 1); } + spin_unlock(&pblk->trans_lock); +} - c_entries = find_first_bit(comp_bits, nr_ppas); - ret = pblk_recov_setup_rq(pblk, c_ctx, recovery, comp_bits, c_entries); - if (ret) { - pr_err("pblk: could not recover from write failure\n"); - mempool_free(recovery, &pblk->rec_pool); - goto out; +static void pblk_queue_resubmit(struct pblk *pblk, struct pblk_c_ctx *c_ctx) +{ + struct pblk_c_ctx *r_ctx; + + r_ctx = kzalloc(sizeof(struct pblk_c_ctx), GFP_KERNEL); + if (!r_ctx) + return; + + r_ctx->lun_bitmap = NULL; + r_ctx->sentry = c_ctx->sentry; + r_ctx->nr_valid = c_ctx->nr_valid; + r_ctx->nr_padded = c_ctx->nr_padded; + + spin_lock(&pblk->resubmit_lock); + list_add_tail(&r_ctx->list, &pblk->resubmit_list); + spin_unlock(&pblk->resubmit_lock); + +#ifdef CONFIG_NVM_DEBUG + atomic_long_add(c_ctx->nr_valid, &pblk->recov_writes); +#endif +} + +static void pblk_submit_rec(struct work_struct *work) +{ + struct pblk_rec_ctx *recovery = + container_of(work, struct pblk_rec_ctx, ws_rec); + struct pblk *pblk = recovery->pblk; + struct nvm_rq *rqd = recovery->rqd; + struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd); + struct ppa_addr *ppa_list; + + pblk_log_write_err(pblk, rqd); + + if (rqd->nr_ppas == 1) + ppa_list = &rqd->ppa_addr; + else + ppa_list = rqd->ppa_list; + + pblk_map_remaining(pblk, ppa_list); + pblk_queue_resubmit(pblk, c_ctx); + + pblk_up_rq(pblk, rqd->ppa_list, rqd->nr_ppas, c_ctx->lun_bitmap); + if (c_ctx->nr_padded) + pblk_bio_free_pages(pblk, rqd->bio, c_ctx->nr_valid, + c_ctx->nr_padded); + bio_put(rqd->bio); + pblk_free_rqd(pblk, rqd, PBLK_WRITE); + mempool_free(recovery, &pblk->rec_pool); + + atomic_dec(&pblk->inflight_io); +} + + +static void pblk_end_w_fail(struct pblk *pblk, struct nvm_rq *rqd) +{ + struct pblk_rec_ctx *recovery; + + recovery = mempool_alloc(&pblk->rec_pool, GFP_ATOMIC); + if (!recovery) { + pr_err("pblk: could not allocate recovery work\n"); + return; } + recovery->pblk = pblk; + recovery->rqd = rqd; + INIT_WORK(&recovery->ws_rec, pblk_submit_rec); queue_work(pblk->close_wq, &recovery->ws_rec); - -out: - pblk_complete_write(pblk, rqd, c_ctx); } static void pblk_end_io_write(struct nvm_rq *rqd) @@ -173,8 +254,8 @@ static void pblk_end_io_write(struct nvm_rq *rqd) struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd); if (rqd->error) { - pblk_log_write_err(pblk, rqd); - return pblk_end_w_fail(pblk, rqd); + pblk_end_w_fail(pblk, rqd); + return; } #ifdef CONFIG_NVM_DEBUG else @@ -266,31 +347,6 @@ static int pblk_setup_w_rq(struct pblk *pblk, struct nvm_rq *rqd, return 0; } -int pblk_setup_w_rec_rq(struct pblk *pblk, struct nvm_rq *rqd, - struct pblk_c_ctx *c_ctx) -{ - struct pblk_line_meta *lm = &pblk->lm; - unsigned long *lun_bitmap; - int ret; - - lun_bitmap = kzalloc(lm->lun_bitmap_len, GFP_KERNEL); - if (!lun_bitmap) - return -ENOMEM; - - c_ctx->lun_bitmap = lun_bitmap; - - ret = pblk_alloc_w_rq(pblk, rqd, rqd->nr_ppas, pblk_end_io_write); - if (ret) - return ret; - - pblk_map_rq(pblk, rqd, c_ctx->sentry, lun_bitmap, c_ctx->nr_valid, 0); - - rqd->ppa_status = (u64)0; - rqd->flags = pblk_set_progr_mode(pblk, PBLK_WRITE); - - return ret; -} - static int pblk_calc_secs_to_sync(struct pblk *pblk, unsigned int secs_avail, unsigned int secs_to_flush) { @@ -339,6 +395,7 @@ int pblk_submit_meta_io(struct pblk *pblk, struct pblk_line *meta_line) bio = pblk_bio_map_addr(pblk, data, rq_ppas, rq_len, l_mg->emeta_alloc_type, GFP_KERNEL); if (IS_ERR(bio)) { + pr_err("pblk: failed to map emeta io"); ret = PTR_ERR(bio); goto fail_free_rqd; } @@ -515,26 +572,54 @@ static int pblk_submit_write(struct pblk *pblk) unsigned int secs_avail, secs_to_sync, secs_to_com; unsigned int secs_to_flush; unsigned long pos; + unsigned int resubmit; - /* If there are no sectors in the cache, flushes (bios without data) - * will be cleared on the cache threads - */ - secs_avail = pblk_rb_read_count(&pblk->rwb); - if (!secs_avail) - return 1; - - secs_to_flush = pblk_rb_flush_point_count(&pblk->rwb); - if (!secs_to_flush && secs_avail < pblk->min_write_pgs) - return 1; - - secs_to_sync = pblk_calc_secs_to_sync(pblk, secs_avail, secs_to_flush); - if (secs_to_sync > pblk->max_write_pgs) { - pr_err("pblk: bad buffer sync calculation\n"); - return 1; - } + spin_lock(&pblk->resubmit_lock); + resubmit = !list_empty(&pblk->resubmit_list); + spin_unlock(&pblk->resubmit_lock); + + /* Resubmit failed writes first */ + if (resubmit) { + struct pblk_c_ctx *r_ctx; + + spin_lock(&pblk->resubmit_lock); + r_ctx = list_first_entry(&pblk->resubmit_list, + struct pblk_c_ctx, list); + list_del(&r_ctx->list); + spin_unlock(&pblk->resubmit_lock); + + secs_avail = r_ctx->nr_valid; + pos = r_ctx->sentry; + + pblk_prepare_resubmit(pblk, pos, secs_avail); + secs_to_sync = pblk_calc_secs_to_sync(pblk, secs_avail, + secs_avail); - secs_to_com = (secs_to_sync > secs_avail) ? secs_avail : secs_to_sync; - pos = pblk_rb_read_commit(&pblk->rwb, secs_to_com); + kfree(r_ctx); + } else { + /* If there are no sectors in the cache, + * flushes (bios without data) will be cleared on + * the cache threads + */ + secs_avail = pblk_rb_read_count(&pblk->rwb); + if (!secs_avail) + return 1; + + secs_to_flush = pblk_rb_flush_point_count(&pblk->rwb); + if (!secs_to_flush && secs_avail < pblk->min_write_pgs) + return 1; + + secs_to_sync = pblk_calc_secs_to_sync(pblk, secs_avail, + secs_to_flush); + if (secs_to_sync > pblk->max_write_pgs) { + pr_err("pblk: bad buffer sync calculation\n"); + return 1; + } + + secs_to_com = (secs_to_sync > secs_avail) ? + secs_avail : secs_to_sync; + pos = pblk_rb_read_commit(&pblk->rwb, secs_to_com); + } bio = bio_alloc(GFP_KERNEL, secs_to_sync); |