lightnvm: pblk: sched. metadata on write thread

At the moment, line metadata is persisted on a separate work queue, that
is kicked each time that a line is closed. The assumption when designing
this was that freeing the write thread from creating a new write request
was better than the potential impact of writes colliding on the media
(user I/O and metadata I/O). Experimentation has proven that this
assumption is wrong; collision can cause up to 25% of bandwidth and
introduce long tail latencies on the write thread, which potentially
cause user write threads to spend more time spinning to get a free entry
on the write buffer.

This patch moves the metadata logic to the write thread. When a line is
closed, remaining metadata is written in memory and is placed on a
metadata queue. The write thread then takes the metadata corresponding
to the previous line, creates the write request and schedules it to
minimize collisions on the media. Using this approach, we see that we
can saturate the media's bandwidth, which helps reducing both write
latencies and the spinning time for user writer threads.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

1 files changed, 228 insertions, 23 deletions

diff --git a/drivers/lightnvm/pblk-write.c b/drivers/lightnvm/pblk-write.c
index c745a22057f8..1739c970692e 100644
--- a/drivers/lightnvm/pblk-write.c
+++ b/drivers/lightnvm/pblk-write.c
@@ -25,8 +25,6 @@ static void pblk_sync_line(struct pblk *pblk, struct pblk_line *line)
 
 	/* Counter protected by rb sync lock */
 	line->left_ssecs--;
-	if (!line->left_ssecs)
-		pblk_line_run_ws(pblk, line, NULL, pblk_line_close_ws);
 }
 
 static unsigned long pblk_end_w_bio(struct pblk *pblk, struct nvm_rq *rqd,
@@ -192,8 +190,40 @@ static void pblk_end_io_write(struct nvm_rq *rqd)
 	pblk_complete_write(pblk, rqd, c_ctx);
 }
 
+static void pblk_end_io_write_meta(struct nvm_rq *rqd)
+{
+	struct pblk *pblk = rqd->private;
+	struct nvm_tgt_dev *dev = pblk->dev;
+	struct nvm_geo *geo = &dev->geo;
+	struct pblk_g_ctx *m_ctx = nvm_rq_to_pdu(rqd);
+	struct pblk_line *line = m_ctx->private;
+	struct pblk_emeta *emeta = line->emeta;
+	int pos = pblk_ppa_to_pos(geo, rqd->ppa_list[0]);
+	struct pblk_lun *rlun = &pblk->luns[pos];
+	int sync;
+
+	up(&rlun->wr_sem);
+
+	if (rqd->error) {
+		pblk_log_write_err(pblk, rqd);
+		pr_err("pblk: metadata I/O failed\n");
+	}
+#ifdef CONFIG_NVM_DEBUG
+	else
+		WARN_ONCE(rqd->bio->bi_status, "pblk: corrupted write error\n");
+#endif
+
+	sync = atomic_add_return(rqd->nr_ppas, &emeta->sync);
+	if (sync == emeta->nr_entries)
+		pblk_line_run_ws(pblk, line, NULL, pblk_line_close_ws);
+
+	bio_put(rqd->bio);
+	pblk_free_rqd(pblk, rqd, READ);
+}
+
 static int pblk_alloc_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
-			   unsigned int nr_secs)
+			   unsigned int nr_secs,
+			   nvm_end_io_fn(*end_io))
 {
 	struct nvm_tgt_dev *dev = pblk->dev;
 
@@ -202,7 +232,7 @@ static int pblk_alloc_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
 	rqd->nr_ppas = nr_secs;
 	rqd->flags = pblk_set_progr_mode(pblk, WRITE);
 	rqd->private = pblk;
-	rqd->end_io = pblk_end_io_write;
+	rqd->end_io = end_io;
 
 	rqd->meta_list = nvm_dev_dma_alloc(dev->parent, GFP_KERNEL,
 							&rqd->dma_meta_list);
@@ -234,7 +264,7 @@ static int pblk_setup_w_rq(struct pblk *pblk, struct nvm_rq *rqd,
 		return -ENOMEM;
 	c_ctx->lun_bitmap = lun_bitmap;
 
-	ret = pblk_alloc_w_rq(pblk, rqd, nr_secs);
+	ret = pblk_alloc_w_rq(pblk, rqd, nr_secs, pblk_end_io_write);
 	if (ret) {
 		kfree(lun_bitmap);
 		return ret;
@@ -262,7 +292,7 @@ int pblk_setup_w_rec_rq(struct pblk *pblk, struct nvm_rq *rqd,
 
 	c_ctx->lun_bitmap = lun_bitmap;
 
-	ret = pblk_alloc_w_rq(pblk, rqd, rqd->nr_ppas);
+	ret = pblk_alloc_w_rq(pblk, rqd, rqd->nr_ppas, pblk_end_io_write);
 	if (ret)
 		return ret;
 
@@ -293,6 +323,166 @@ static int pblk_calc_secs_to_sync(struct pblk *pblk, unsigned int secs_avail,
 	return secs_to_sync;
 }
 
+static inline int pblk_valid_meta_ppa(struct pblk *pblk,
+				      struct pblk_line *meta_line,
+				      struct ppa_addr *ppa_list, int nr_ppas)
+{
+	struct nvm_tgt_dev *dev = pblk->dev;
+	struct nvm_geo *geo = &dev->geo;
+	struct pblk_line *data_line;
+	struct ppa_addr ppa, ppa_opt;
+	u64 paddr;
+	int i;
+
+	data_line = &pblk->lines[pblk_dev_ppa_to_line(ppa_list[0])];
+	paddr = pblk_lookup_page(pblk, meta_line);
+	ppa = addr_to_gen_ppa(pblk, paddr, 0);
+
+	if (test_bit(pblk_ppa_to_pos(geo, ppa), data_line->blk_bitmap))
+		return 1;
+
+	/* Schedule a metadata I/O that is half the distance from the data I/O
+	 * with regards to the number of LUNs forming the pblk instance. This
+	 * balances LUN conflicts across every I/O.
+	 *
+	 * When the LUN configuration changes (e.g., due to GC), this distance
+	 * can align, which would result on a LUN deadlock. In this case, modify
+	 * the distance to not be optimal, but allow metadata I/Os to succeed.
+	 */
+	ppa_opt = addr_to_gen_ppa(pblk, paddr + data_line->meta_distance, 0);
+	if (unlikely(ppa_opt.ppa == ppa.ppa)) {
+		data_line->meta_distance--;
+		return 0;
+	}
+
+	for (i = 0; i < nr_ppas; i += pblk->min_write_pgs)
+		if (ppa_list[i].g.ch == ppa_opt.g.ch &&
+					ppa_list[i].g.lun == ppa_opt.g.lun)
+			return 1;
+
+	if (test_bit(pblk_ppa_to_pos(geo, ppa_opt), data_line->blk_bitmap)) {
+		for (i = 0; i < nr_ppas; i += pblk->min_write_pgs)
+			if (ppa_list[i].g.ch == ppa.g.ch &&
+						ppa_list[i].g.lun == ppa.g.lun)
+				return 0;
+
+		return 1;
+	}
+
+	return 0;
+}
+
+int pblk_submit_meta_io(struct pblk *pblk, struct pblk_line *meta_line)
+{
+	struct nvm_tgt_dev *dev = pblk->dev;
+	struct nvm_geo *geo = &dev->geo;
+	struct pblk_line_mgmt *l_mg = &pblk->l_mg;
+	struct pblk_line_meta *lm = &pblk->lm;
+	struct pblk_emeta *emeta = meta_line->emeta;
+	struct pblk_g_ctx *m_ctx;
+	struct pblk_lun *rlun;
+	struct bio *bio;
+	struct nvm_rq *rqd;
+	void *data;
+	u64 paddr;
+	int rq_ppas = pblk->min_write_pgs;
+	int id = meta_line->id;
+	int rq_len;
+	int i, j;
+	int ret;
+
+	rqd = pblk_alloc_rqd(pblk, READ);
+	if (IS_ERR(rqd)) {
+		pr_err("pblk: cannot allocate write req.\n");
+		return PTR_ERR(rqd);
+	}
+	m_ctx = nvm_rq_to_pdu(rqd);
+	m_ctx->private = meta_line;
+
+	rq_len = rq_ppas * geo->sec_size;
+	data = ((void *)emeta->buf) + emeta->mem;
+
+	bio = pblk_bio_map_addr(pblk, data, rq_ppas, rq_len, GFP_KERNEL);
+	if (IS_ERR(bio)) {
+		ret = PTR_ERR(bio);
+		goto fail_free_rqd;
+	}
+	bio->bi_iter.bi_sector = 0; /* internal bio */
+	bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
+	rqd->bio = bio;
+
+	ret = pblk_alloc_w_rq(pblk, rqd, rq_ppas, pblk_end_io_write_meta);
+	if (ret)
+		goto fail_free_bio;
+
+	for (i = 0; i < rqd->nr_ppas; ) {
+		spin_lock(&meta_line->lock);
+		paddr = __pblk_alloc_page(pblk, meta_line, rq_ppas);
+		spin_unlock(&meta_line->lock);
+		for (j = 0; j < rq_ppas; j++, i++, paddr++)
+			rqd->ppa_list[i] = addr_to_gen_ppa(pblk, paddr, id);
+	}
+
+	rlun = &pblk->luns[pblk_ppa_to_pos(geo, rqd->ppa_list[0])];
+	ret = down_timeout(&rlun->wr_sem, msecs_to_jiffies(5000));
+	if (ret) {
+		pr_err("pblk: lun semaphore timed out (%d)\n", ret);
+		goto fail_free_bio;
+	}
+
+	emeta->mem += rq_len;
+	if (emeta->mem >= lm->emeta_len[0]) {
+		spin_lock(&l_mg->close_lock);
+		list_del(&meta_line->list);
+		WARN(!bitmap_full(meta_line->map_bitmap, lm->sec_per_line),
+				"pblk: corrupt meta line %d\n", meta_line->id);
+		spin_unlock(&l_mg->close_lock);
+	}
+
+	ret = pblk_submit_io(pblk, rqd);
+	if (ret) {
+		pr_err("pblk: emeta I/O submission failed: %d\n", ret);
+		goto fail_rollback;
+	}
+
+	return NVM_IO_OK;
+
+fail_rollback:
+	spin_lock(&l_mg->close_lock);
+	pblk_dealloc_page(pblk, meta_line, rq_ppas);
+	list_add(&meta_line->list, &meta_line->list);
+	spin_unlock(&l_mg->close_lock);
+fail_free_bio:
+	bio_put(bio);
+fail_free_rqd:
+	pblk_free_rqd(pblk, rqd, READ);
+	return ret;
+}
+
+static int pblk_sched_meta_io(struct pblk *pblk, struct ppa_addr *prev_list,
+			       int prev_n)
+{
+	struct pblk_line_meta *lm = &pblk->lm;
+	struct pblk_line_mgmt *l_mg = &pblk->l_mg;
+	struct pblk_line *meta_line;
+
+	spin_lock(&l_mg->close_lock);
+retry:
+	if (list_empty(&l_mg->emeta_list)) {
+		spin_unlock(&l_mg->close_lock);
+		return 0;
+	}
+	meta_line = list_first_entry(&l_mg->emeta_list, struct pblk_line, list);
+	if (bitmap_full(meta_line->map_bitmap, lm->sec_per_line))
+		goto retry;
+	spin_unlock(&l_mg->close_lock);
+
+	if (!pblk_valid_meta_ppa(pblk, meta_line, prev_list, prev_n))
+		return 0;
+
+	return pblk_submit_meta_io(pblk, meta_line);
+}
+
 static int pblk_submit_io_set(struct pblk *pblk, struct nvm_rq *rqd)
 {
 	struct pblk_c_ctx *c_ctx = nvm_rq_to_pdu(rqd);
@@ -308,24 +498,39 @@ static int pblk_submit_io_set(struct pblk *pblk, struct nvm_rq *rqd)
 		return NVM_IO_ERR;
 	}
 
-	/* Submit write for current data line */
-	err = pblk_submit_io(pblk, rqd);
-	if (err) {
-		pr_err("pblk: I/O submission failed: %d\n", err);
-		return NVM_IO_ERR;
-	}
+	if (likely(ppa_empty(erase_ppa))) {
+		/* Submit metadata write for previous data line */
+		err = pblk_sched_meta_io(pblk, rqd->ppa_list, rqd->nr_ppas);
+		if (err) {
+			pr_err("pblk: metadata I/O submission failed: %d", err);
+			return NVM_IO_ERR;
+		}
 
-	/* Submit available erase for next data line */
-	if (unlikely(!ppa_empty(erase_ppa)) &&
-				pblk_blk_erase_async(pblk, erase_ppa)) {
-		struct pblk_line *e_line = pblk_line_get_erase(pblk);
-		struct nvm_tgt_dev *dev = pblk->dev;
-		struct nvm_geo *geo = &dev->geo;
-		int bit;
-
-		atomic_inc(&e_line->left_eblks);
-		bit = erase_ppa.g.lun * geo->nr_chnls + erase_ppa.g.ch;
-		WARN_ON(!test_and_clear_bit(bit, e_line->erase_bitmap));
+		/* Submit data write for current data line */
+		err = pblk_submit_io(pblk, rqd);
+		if (err) {
+			pr_err("pblk: data I/O submission failed: %d\n", err);
+			return NVM_IO_ERR;
+		}
+	} else {
+		/* Submit data write for current data line */
+		err = pblk_submit_io(pblk, rqd);
+		if (err) {
+			pr_err("pblk: data I/O submission failed: %d\n", err);
+			return NVM_IO_ERR;
+		}
+
+		/* Submit available erase for next data line */
+		if (pblk_blk_erase_async(pblk, erase_ppa)) {
+			struct pblk_line *e_line = pblk_line_get_erase(pblk);
+			struct nvm_tgt_dev *dev = pblk->dev;
+			struct nvm_geo *geo = &dev->geo;
+			int bit;
+
+			atomic_inc(&e_line->left_eblks);
+			bit = pblk_ppa_to_pos(geo, erase_ppa);
+			WARN_ON(!test_and_clear_bit(bit, e_line->erase_bitmap));
+		}
 	}
 
 	return NVM_IO_OK;