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>

drivers/lightnvm/pblk-gc.c

@@ -156,7 +156,8 @@ static void pblk_gc_line_ws(struct work_struct *work)
 	struct pblk_line_mgmt *l_mg = &pblk->l_mg;
 	struct pblk_line *line = line_ws->line;
 	struct pblk_line_meta *lm = &pblk->lm;
-	__le64 *lba_list = line_ws->priv;
+	struct line_emeta *emeta_buf = line_ws->priv;
+	__le64 *lba_list;
 	u64 *gc_list;
 	int sec_left;
 	int nr_ppas, bit;
@@ -164,8 +165,18 @@ static void pblk_gc_line_ws(struct work_struct *work)
 
 	pr_debug("pblk: line '%d' being reclaimed for GC\n", line->id);
 
+	/* If this read fails, it means that emeta is corrupted. For now, leave
+	 * the line untouched. TODO: Implement a recovery routine that scans and
+	 * moves all sectors on the line.
+	 */
+	lba_list = pblk_recov_get_lba_list(pblk, emeta_buf);
+	if (!lba_list) {
+		pr_err("pblk: could not interpret emeta (line %d)\n", line->id);
+		goto out;
+	}
+
 	spin_lock(&line->lock);
-	sec_left = line->vsc;
+	sec_left = le32_to_cpu(*line->vsc);
 	if (!sec_left) {
 		/* Lines are erased before being used (l_mg->data_/log_next) */
 		spin_unlock(&line->lock);
@@ -206,7 +217,7 @@ static void pblk_gc_line_ws(struct work_struct *work)
 
 	if (pblk_gc_move_valid_secs(pblk, line, gc_list, nr_ppas)) {
 		pr_err("pblk: could not GC all sectors: line:%d (%d/%d/%d)\n",
-						line->id, line->vsc,
+						line->id, *line->vsc,
 						nr_ppas, nr_ppas);
 		put_line = 0;
 		pblk_put_line_back(pblk, line);
@@ -218,7 +229,7 @@ static void pblk_gc_line_ws(struct work_struct *work)
 		goto next_rq;
 
 out:
-	pblk_mfree(line->emeta, l_mg->emeta_alloc_type);
+	pblk_mfree(emeta_buf, l_mg->emeta_alloc_type);
 	mempool_free(line_ws, pblk->line_ws_pool);
 	atomic_dec(&pblk->gc.inflight_gc);
 	if (put_line)
@@ -229,37 +240,27 @@ static int pblk_gc_line(struct pblk *pblk, struct pblk_line *line)
 {
 	struct pblk_line_mgmt *l_mg = &pblk->l_mg;
 	struct pblk_line_meta *lm = &pblk->lm;
+	struct line_emeta *emeta_buf;
 	struct pblk_line_ws *line_ws;
-	__le64 *lba_list;
 	int ret;
 
 	line_ws = mempool_alloc(pblk->line_ws_pool, GFP_KERNEL);
-	line->emeta = pblk_malloc(lm->emeta_len, l_mg->emeta_alloc_type,
+	emeta_buf = pblk_malloc(lm->emeta_len[0], l_mg->emeta_alloc_type,
 								GFP_KERNEL);
-	if (!line->emeta) {
+	if (!emeta_buf) {
 		pr_err("pblk: cannot use GC emeta\n");
 		goto fail_free_ws;
 	}
 
-	ret = pblk_line_read_emeta(pblk, line);
+	ret = pblk_line_read_emeta(pblk, line, emeta_buf);
 	if (ret) {
 		pr_err("pblk: line %d read emeta failed (%d)\n", line->id, ret);
 		goto fail_free_emeta;
 	}
 
-	/* If this read fails, it means that emeta is corrupted. For now, leave
-	 * the line untouched. TODO: Implement a recovery routine that scans and
-	 * moves all sectors on the line.
-	 */
-	lba_list = pblk_recov_get_lba_list(pblk, line->emeta);
-	if (!lba_list) {
-		pr_err("pblk: could not interpret emeta (line %d)\n", line->id);
-		goto fail_free_emeta;
-	}
-
 	line_ws->pblk = pblk;
 	line_ws->line = line;
-	line_ws->priv = lba_list;
+	line_ws->priv = emeta_buf;
 
 	INIT_WORK(&line_ws->ws, pblk_gc_line_ws);
 	queue_work(pblk->gc.gc_reader_wq, &line_ws->ws);
@@ -267,7 +268,7 @@ static int pblk_gc_line(struct pblk *pblk, struct pblk_line *line)
 	return 0;
 
 fail_free_emeta:
-	pblk_mfree(line->emeta, l_mg->emeta_alloc_type);
+	pblk_mfree(emeta_buf, l_mg->emeta_alloc_type);
 fail_free_ws:
 	mempool_free(line_ws, pblk->line_ws_pool);
 	pblk_put_line_back(pblk, line);

drivers/lightnvm/pblk-map.c

@@ -25,9 +25,9 @@ static void pblk_map_page_data(struct pblk *pblk, unsigned int sentry,
 			       unsigned int valid_secs)
 {
 	struct pblk_line *line = pblk_line_get_data(pblk);
-	struct line_emeta *emeta = line->emeta;
+	struct pblk_emeta *emeta = line->emeta;
 	struct pblk_w_ctx *w_ctx;
-	__le64 *lba_list = pblk_line_emeta_to_lbas(emeta);
+	__le64 *lba_list = emeta_to_lbas(pblk, emeta->buf);
 	u64 paddr;
 	int nr_secs = pblk->min_write_pgs;
 	int i;
@@ -51,7 +51,7 @@ static void pblk_map_page_data(struct pblk *pblk, unsigned int sentry,
 			w_ctx->ppa = ppa_list[i];
 			meta_list[i].lba = cpu_to_le64(w_ctx->lba);
 			lba_list[paddr] = cpu_to_le64(w_ctx->lba);
-			le64_add_cpu(&line->emeta->nr_valid_lbas, 1);
+			line->nr_valid_lbas++;
 		} else {
 			u64 addr_empty = cpu_to_le64(ADDR_EMPTY);
 
@@ -61,9 +61,11 @@ static void pblk_map_page_data(struct pblk *pblk, unsigned int sentry,
 	}
 
 	if (pblk_line_is_full(line)) {
+		struct pblk_line *prev_line = line;
 		line = pblk_line_replace_data(pblk);
 		if (!line)
 			return;
+		pblk_line_close_meta(pblk, prev_line);
 	}
 
 	pblk_down_rq(pblk, ppa_list, nr_secs, lun_bitmap);
@@ -104,11 +106,10 @@ void pblk_map_erase_rq(struct pblk *pblk, struct nvm_rq *rqd,
 		pblk_map_page_data(pblk, sentry + i, &rqd->ppa_list[i],
 					lun_bitmap, &meta_list[i], map_secs);
 
-		erase_lun = rqd->ppa_list[i].g.lun * geo->nr_chnls +
-							rqd->ppa_list[i].g.ch;
-
 		/* line can change after page map */
 		e_line = pblk_line_get_erase(pblk);
+		erase_lun = pblk_ppa_to_pos(geo, rqd->ppa_list[i]);
+
 		spin_lock(&e_line->lock);
 		if (!test_bit(erase_lun, e_line->erase_bitmap)) {
 			set_bit(erase_lun, e_line->erase_bitmap);

drivers/lightnvm/pblk-recovery.c

@@ -120,18 +120,18 @@ int pblk_recov_setup_rq(struct pblk *pblk, struct pblk_c_ctx *c_ctx,
 	return 0;
 }
 
-__le64 *pblk_recov_get_lba_list(struct pblk *pblk, struct line_emeta *emeta)
+__le64 *pblk_recov_get_lba_list(struct pblk *pblk, struct line_emeta *emeta_buf)
 {
 	u32 crc;
 
-	crc = pblk_calc_emeta_crc(pblk, emeta);
-	if (le32_to_cpu(emeta->crc) != crc)
+	crc = pblk_calc_emeta_crc(pblk, emeta_buf);
+	if (le32_to_cpu(emeta_buf->crc) != crc)
 		return NULL;
 
-	if (le32_to_cpu(emeta->header.identifier) != PBLK_MAGIC)
+	if (le32_to_cpu(emeta_buf->header.identifier) != PBLK_MAGIC)
 		return NULL;
 
-	return pblk_line_emeta_to_lbas(emeta);
+	return emeta_to_lbas(pblk, emeta_buf);
 }
 
 static int pblk_recov_l2p_from_emeta(struct pblk *pblk, struct pblk_line *line)
@@ -139,19 +139,20 @@ static int pblk_recov_l2p_from_emeta(struct pblk *pblk, struct pblk_line *line)
 	struct nvm_tgt_dev *dev = pblk->dev;
 	struct nvm_geo *geo = &dev->geo;
 	struct pblk_line_meta *lm = &pblk->lm;
-	struct line_emeta *emeta = line->emeta;
+	struct pblk_emeta *emeta = line->emeta;
+	struct line_emeta *emeta_buf = emeta->buf;
 	__le64 *lba_list;
 	int data_start;
 	int nr_data_lbas, nr_valid_lbas, nr_lbas = 0;
 	int i;
 
-	lba_list = pblk_recov_get_lba_list(pblk, emeta);
+	lba_list = pblk_recov_get_lba_list(pblk, emeta_buf);
 	if (!lba_list)
 		return 1;
 
 	data_start = pblk_line_smeta_start(pblk, line) + lm->smeta_sec;
-	nr_data_lbas = lm->sec_per_line - lm->emeta_sec;
-	nr_valid_lbas = le64_to_cpu(emeta->nr_valid_lbas);
+	nr_data_lbas = lm->sec_per_line - lm->emeta_sec[0];
+	nr_valid_lbas = le64_to_cpu(emeta_buf->nr_valid_lbas);
 
 	for (i = data_start; i < nr_data_lbas && nr_lbas < nr_valid_lbas; i++) {
 		struct ppa_addr ppa;
@@ -169,7 +170,7 @@ static int pblk_recov_l2p_from_emeta(struct pblk *pblk, struct pblk_line *line)
 			if (test_and_set_bit(i, line->invalid_bitmap))
 				WARN_ONCE(1, "pblk: rec. double invalidate:\n");
 			else
-				line->vsc--;
+				le32_add_cpu(line->vsc, -1);
 			spin_unlock(&line->lock);
 
 			continue;
@@ -181,7 +182,7 @@ static int pblk_recov_l2p_from_emeta(struct pblk *pblk, struct pblk_line *line)
 
 	if (nr_valid_lbas != nr_lbas)
 		pr_err("pblk: line %d - inconsistent lba list(%llu/%d)\n",
-				line->id, line->emeta->nr_valid_lbas, nr_lbas);
+				line->id, emeta_buf->nr_valid_lbas, nr_lbas);
 
 	line->left_msecs = 0;
 
@@ -195,7 +196,7 @@ static int pblk_calc_sec_in_line(struct pblk *pblk, struct pblk_line *line)
 	struct pblk_line_meta *lm = &pblk->lm;
 	int nr_bb = bitmap_weight(line->blk_bitmap, lm->blk_per_line);
 
-	return lm->sec_per_line - lm->smeta_sec - lm->emeta_sec -
+	return lm->sec_per_line - lm->smeta_sec - lm->emeta_sec[0] -
 				nr_bb * geo->sec_per_blk;
 }
 
@@ -333,7 +334,7 @@ static int pblk_recov_pad_oob(struct pblk *pblk, struct pblk_line *line,
 	struct bio *bio;
 	void *data;
 	dma_addr_t dma_ppa_list, dma_meta_list;
-	__le64 *lba_list = pblk_line_emeta_to_lbas(line->emeta);
+	__le64 *lba_list = emeta_to_lbas(pblk, line->emeta->buf);
 	u64 w_ptr = line->cur_sec;
 	int left_line_ppas = line->left_msecs;
 	int rq_ppas, rq_len;
@@ -770,8 +771,9 @@ struct pblk_line *pblk_recov_l2p(struct pblk *pblk)
 	struct pblk_line_meta *lm = &pblk->lm;
 	struct pblk_line_mgmt *l_mg = &pblk->l_mg;
 	struct pblk_line *line, *tline, *data_line = NULL;
-	struct line_smeta *smeta;
-	struct line_emeta *emeta;
+	struct pblk_smeta *smeta;
+	struct pblk_emeta *emeta;
+	struct line_smeta *smeta_buf;
 	int found_lines = 0, recovered_lines = 0, open_lines = 0;
 	int is_next = 0;
 	int meta_line;
@@ -784,8 +786,9 @@ struct pblk_line *pblk_recov_l2p(struct pblk *pblk)
 	spin_lock(&l_mg->free_lock);
 	meta_line = find_first_zero_bit(&l_mg->meta_bitmap, PBLK_DATA_LINES);
 	set_bit(meta_line, &l_mg->meta_bitmap);
-	smeta = l_mg->sline_meta[meta_line].meta;
-	emeta = l_mg->eline_meta[meta_line].meta;
+	smeta = l_mg->sline_meta[meta_line];
+	emeta = l_mg->eline_meta[meta_line];
+	smeta_buf = smeta->buf;
 	spin_unlock(&l_mg->free_lock);
 
 	/* Order data lines using their sequence number */
@@ -796,33 +799,33 @@ struct pblk_line *pblk_recov_l2p(struct pblk *pblk)
 
 		memset(smeta, 0, lm->smeta_len);
 		line->smeta = smeta;
-		line->lun_bitmap = ((void *)(smeta)) +
+		line->lun_bitmap = ((void *)(smeta_buf)) +
 						sizeof(struct line_smeta);
 
 		/* Lines that cannot be read are assumed as not written here */
 		if (pblk_line_read_smeta(pblk, line))
 			continue;
 
-		crc = pblk_calc_smeta_crc(pblk, smeta);
-		if (le32_to_cpu(smeta->crc) != crc)
+		crc = pblk_calc_smeta_crc(pblk, smeta_buf);
+		if (le32_to_cpu(smeta_buf->crc) != crc)
 			continue;
 
-		if (le32_to_cpu(smeta->header.identifier) != PBLK_MAGIC)
+		if (le32_to_cpu(smeta_buf->header.identifier) != PBLK_MAGIC)
 			continue;
 
-		if (le16_to_cpu(smeta->header.version) != 1) {
+		if (le16_to_cpu(smeta_buf->header.version) != 1) {
 			pr_err("pblk: found incompatible line version %u\n",
-					smeta->header.version);
+					smeta_buf->header.version);
 			return ERR_PTR(-EINVAL);
 		}
 
 		/* The first valid instance uuid is used for initialization */
 		if (!valid_uuid) {
-			memcpy(pblk->instance_uuid, smeta->header.uuid, 16);
+			memcpy(pblk->instance_uuid, smeta_buf->header.uuid, 16);
 			valid_uuid = 1;
 		}
 
-		if (memcmp(pblk->instance_uuid, smeta->header.uuid, 16)) {
+		if (memcmp(pblk->instance_uuid, smeta_buf->header.uuid, 16)) {
 			pr_debug("pblk: ignore line %u due to uuid mismatch\n",
 					i);
 			continue;
@@ -830,9 +833,9 @@ struct pblk_line *pblk_recov_l2p(struct pblk *pblk)
 
 		/* Update line metadata */
 		spin_lock(&line->lock);
-		line->id = le32_to_cpu(line->smeta->header.id);
-		line->type = le16_to_cpu(line->smeta->header.type);
-		line->seq_nr = le64_to_cpu(line->smeta->seq_nr);
+		line->id = le32_to_cpu(smeta_buf->header.id);
+		line->type = le16_to_cpu(smeta_buf->header.type);
+		line->seq_nr = le64_to_cpu(smeta_buf->seq_nr);
 		spin_unlock(&line->lock);
 
 		/* Update general metadata */
@@ -848,7 +851,7 @@ struct pblk_line *pblk_recov_l2p(struct pblk *pblk)
 		pblk_recov_line_add_ordered(&recov_list, line);
 		found_lines++;
 		pr_debug("pblk: recovering data line %d, seq:%llu\n",
-						line->id, smeta->seq_nr);
+						line->id, smeta_buf->seq_nr);
 	}
 
 	if (!found_lines) {
@@ -868,15 +871,15 @@ struct pblk_line *pblk_recov_l2p(struct pblk *pblk)
 
 		recovered_lines++;
 		/* Calculate where emeta starts based on the line bb */
-		off = lm->sec_per_line - lm->emeta_sec;
+		off = lm->sec_per_line - lm->emeta_sec[0];
 		nr_bb = bitmap_weight(line->blk_bitmap, lm->blk_per_line);
 		off -= nr_bb * geo->sec_per_pl;
 
-		memset(emeta, 0, lm->emeta_len);
+		memset(&emeta->buf, 0, lm->emeta_len[0]);
 		line->emeta = emeta;
 		line->emeta_ssec = off;
 
-		if (pblk_line_read_emeta(pblk, line)) {
+		if (pblk_line_read_emeta(pblk, line, line->emeta->buf)) {
 			pblk_recov_l2p_from_oob(pblk, line);
 			goto next;
 		}

drivers/lightnvm/pblk-sysfs.c

@@ -150,7 +150,7 @@ static ssize_t pblk_sysfs_lines(struct pblk *pblk, char *page)
 	ssize_t sz = 0;
 	int nr_free_lines;
 	int cur_data, cur_log;
-	int free_line_cnt = 0, closed_line_cnt = 0;
+	int free_line_cnt = 0, closed_line_cnt = 0, emeta_line_cnt = 0;
 	int d_line_cnt = 0, l_line_cnt = 0;
 	int gc_full = 0, gc_high = 0, gc_mid = 0, gc_low = 0, gc_empty = 0;
 	int free = 0, bad = 0, cor = 0;
@@ -166,6 +166,11 @@ static ssize_t pblk_sysfs_lines(struct pblk *pblk, char *page)
 		free_line_cnt++;
 	spin_unlock(&l_mg->free_lock);
 
+	spin_lock(&l_mg->close_lock);
+	list_for_each_entry(line, &l_mg->emeta_list, list)
+		emeta_line_cnt++;
+	spin_unlock(&l_mg->close_lock);
+
 	spin_lock(&l_mg->gc_lock);
 	list_for_each_entry(line, &l_mg->gc_full_list, list) {
 		if (line->type == PBLK_LINETYPE_DATA)
@@ -225,7 +230,7 @@ static ssize_t pblk_sysfs_lines(struct pblk *pblk, char *page)
 		cur_sec = l_mg->data_line->cur_sec;
 		msecs = l_mg->data_line->left_msecs;
 		ssecs = l_mg->data_line->left_ssecs;
-		vsc = l_mg->data_line->vsc;
+		vsc = le32_to_cpu(*l_mg->data_line->vsc);
 		sec_in_line = l_mg->data_line->sec_in_line;
 		meta_weight = bitmap_weight(&l_mg->meta_bitmap,
 							PBLK_DATA_LINES);
@@ -242,10 +247,11 @@ static ssize_t pblk_sysfs_lines(struct pblk *pblk, char *page)
 		geo->nr_luns, lm->blk_per_line, lm->sec_per_line);
 
 	sz += snprintf(page + sz, PAGE_SIZE - sz,
-		"lines:d:%d,l:%d-f:%d(%d),b:%d,co:%d,c:%d(d:%d,l:%d)t:%d\n",
+		"lines:d:%d,l:%d-f:%d(%d),m:%d,c:%d,b:%d,co:%d(d:%d,l:%d)t:%d\n",
 					cur_data, cur_log,
-					free, nr_free_lines, bad, cor,
+					free, nr_free_lines, emeta_line_cnt,
 					closed_line_cnt,
+					bad, cor,
 					d_line_cnt, l_line_cnt,
 					l_mg->nr_lines);
 
@@ -274,7 +280,7 @@ static ssize_t pblk_sysfs_lines_info(struct pblk *pblk, char *page)
 					lm->smeta_len, lm->smeta_sec);
 	sz += snprintf(page + sz, PAGE_SIZE - sz,
 				"emeta - len:%d, sec:%d, bb_start:%d\n",
-					lm->emeta_len, lm->emeta_sec,
+					lm->emeta_len[0], lm->emeta_sec[0],
 					lm->emeta_bb);
 	sz += snprintf(page + sz, PAGE_SIZE - sz,
 				"bitmap lengths: sec:%d, blk:%d, lun:%d\n",

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);
@@ -308,25 +498,40 @@ static int pblk_submit_io_set(struct pblk *pblk, struct nvm_rq *rqd)
 		return NVM_IO_ERR;
 	}
 
-	/* Submit write for current data line */
+	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 data write for current data line */
 		err = pblk_submit_io(pblk, rqd);
 		if (err) {
-		pr_err("pblk: I/O submission failed: %d\n", 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 (unlikely(!ppa_empty(erase_ppa)) &&
-				pblk_blk_erase_async(pblk, erase_ppa)) {
+		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 = erase_ppa.g.lun * geo->nr_chnls + erase_ppa.g.ch;
+			bit = pblk_ppa_to_pos(geo, erase_ppa);
 			WARN_ON(!test_and_clear_bit(bit, e_line->erase_bitmap));
 		}
+	}
 
 	return NVM_IO_OK;
 }

drivers/lightnvm/pblk.h

@@ -258,8 +258,6 @@ struct pblk_rl {
 	atomic_t free_blocks;
 };
 
-#define PBLK_LINE_NR_LUN_BITMAP 2
-#define PBLK_LINE_NR_SEC_BITMAP 2
 #define PBLK_LINE_EMPTY (~0U)
 
 enum {
@@ -310,16 +308,19 @@ struct line_smeta {
 	__le32 window_wr_lun;	/* Number of parallel LUNs to write */
 
 	__le32 rsvd[2];
+
+	__le64 lun_bitmap[];
 };
 
 /*
- * Metadata Layout:
- *	1. struct pblk_emeta
- *	2. nr_lbas u64 forming lba list
- *	3. nr_lines (all) u32 valid sector count (vsc) (~0U: non-alloc line)
- *	4. nr_luns bits (u64 format) forming line bad block bitmap
- *
- *	3. and 4. will be part of FTL log
+ * Metadata layout in media:
+ *	First sector:
+ *		1. struct line_emeta
+ *		2. bad block bitmap (u64 * window_wr_lun)
+ *	Mid sectors (start at lbas_sector):
+ *		3. nr_lbas (u64) forming lba list
+ *	Last sectors (start at vsc_sector):
+ *		4. u32 valid sector count (vsc) for all lines (~0U: free line)
  */
 struct line_emeta {
 	struct line_header header;
@@ -339,6 +340,23 @@ struct line_emeta {
 	__le32 next_id;		/* Line id for next line */
 	__le64 nr_lbas;		/* Number of lbas mapped in line */
 	__le64 nr_valid_lbas;	/* Number of valid lbas mapped in line */
+	__le64 bb_bitmap[];	/* Updated bad block bitmap for line */
+};
+
+struct pblk_emeta {
+	struct line_emeta *buf;		/* emeta buffer in media format */
+	int mem;			/* Write offset - points to next
+					 * writable entry in memory
+					 */
+	atomic_t sync;			/* Synced - backpointer that signals the
+					 * last entry that has been successfully
+					 * persisted to media
+					 */
+	unsigned int nr_entries;	/* Number of emeta entries */
+};
+
+struct pblk_smeta {
+	struct line_smeta *buf;		/* smeta buffer in persistent format */
 };
 
 struct pblk_line {
@@ -355,9 +373,12 @@ struct pblk_line {
 
 	unsigned long *lun_bitmap;	/* Bitmap for LUNs mapped in line */
 
-	struct line_smeta *smeta;	/* Start metadata */
-	struct line_emeta *emeta;	/* End metadata */
+	struct pblk_smeta *smeta;	/* Start metadata */
+	struct pblk_emeta *emeta;	/* End medatada */
+
 	int meta_line;			/* Metadata line id */
+	int meta_distance;		/* Distance between data and metadata */
+
 	u64 smeta_ssec;			/* Sector where smeta starts */
 	u64 emeta_ssec;			/* Sector where emeta starts */
 
@@ -376,7 +397,9 @@ struct pblk_line {
 	int left_msecs;			/* Sectors left for mapping */
 	int left_ssecs;			/* Sectors left to sync */
 	unsigned int cur_sec;		/* Sector map pointer */
-	unsigned int vsc;		/* Valid sector count in line */
+	unsigned int nr_valid_lbas;	/* Number of valid lbas in line */
+
+	__le32 *vsc;			/* Valid sector count in line */
 
 	struct kref ref;		/* Write buffer L2P references */
 
@@ -385,13 +408,15 @@ struct pblk_line {
 
 #define PBLK_DATA_LINES 4
 
-enum{
+enum {
 	PBLK_KMALLOC_META = 1,
 	PBLK_VMALLOC_META = 2,
 };
 
-struct pblk_line_metadata {
-	void *meta;
+enum {
+	PBLK_EMETA_TYPE_HEADER = 1,	/* struct line_emeta first sector */
+	PBLK_EMETA_TYPE_LLBA = 2,	/* lba list - type: __le64 */
+	PBLK_EMETA_TYPE_VSC = 3,	/* vsc list - type: __le32 */
 };
 
 struct pblk_line_mgmt {
@@ -417,13 +442,17 @@ struct pblk_line_mgmt {
 	struct pblk_line *log_next;	/* Next FTL log line */
 	struct pblk_line *data_next;	/* Next data line */
 
+	struct list_head emeta_list;	/* Lines queued to schedule emeta */
+
+	__le32 *vsc_list;		/* Valid sector counts for all lines */
+
 	/* Metadata allocation type: VMALLOC | KMALLOC */
 	int smeta_alloc_type;
 	int emeta_alloc_type;
 
 	/* Pre-allocated metadata for data lines */
-	struct pblk_line_metadata sline_meta[PBLK_DATA_LINES];
-	struct pblk_line_metadata eline_meta[PBLK_DATA_LINES];
+	struct pblk_smeta *sline_meta[PBLK_DATA_LINES];
+	struct pblk_emeta *eline_meta[PBLK_DATA_LINES];
 	unsigned long meta_bitmap;
 
 	/* Helpers for fast bitmap calculations */
@@ -434,25 +463,40 @@ struct pblk_line_mgmt {
 	unsigned long l_seq_nr;		/* Log line unique sequence number */
 
 	spinlock_t free_lock;
+	spinlock_t close_lock;
 	spinlock_t gc_lock;
 };
 
 struct pblk_line_meta {
 	unsigned int smeta_len;		/* Total length for smeta */
-	unsigned int smeta_sec;		/* Sectors needed for smeta*/
-	unsigned int emeta_len;		/* Total length for emeta */
-	unsigned int emeta_sec;		/* Sectors needed for emeta*/
+	unsigned int smeta_sec;		/* Sectors needed for smeta */
+
+	unsigned int emeta_len[4];	/* Lengths for emeta:
+					 *  [0]: Total length
+					 *  [1]: struct line_emeta length
+					 *  [2]: L2P portion length
+					 *  [3]: vsc list length
+					 */
+	unsigned int emeta_sec[4];	/* Sectors needed for emeta. Same layout
+					 * as emeta_len
+					 */
+
 	unsigned int emeta_bb;		/* Boundary for bb that affects emeta */
+
+	unsigned int vsc_list_len;	/* Length for vsc list */
 	unsigned int sec_bitmap_len;	/* Length for sector bitmap in line */
 	unsigned int blk_bitmap_len;	/* Length for block bitmap in line */
 	unsigned int lun_bitmap_len;	/* Length for lun bitmap in line */
 
 	unsigned int blk_per_line;	/* Number of blocks in a full line */
 	unsigned int sec_per_line;	/* Number of sectors in a line */
+	unsigned int dsec_per_line;	/* Number of data sectors in a line */
 	unsigned int min_blk_line;	/* Min. number of good blocks in line */
 
 	unsigned int mid_thrs;		/* Threshold for GC mid list */
 	unsigned int high_thrs;		/* Threshold for GC high list */
+
+	unsigned int meta_distance;	/* Distance between data and metadata */
 };
 
 struct pblk_addr_format {
@@ -621,6 +665,7 @@ void pblk_discard(struct pblk *pblk, struct bio *bio);
 void pblk_log_write_err(struct pblk *pblk, struct nvm_rq *rqd);
 void pblk_log_read_err(struct pblk *pblk, struct nvm_rq *rqd);
 int pblk_submit_io(struct pblk *pblk, struct nvm_rq *rqd);
+int pblk_submit_meta_io(struct pblk *pblk, struct pblk_line *meta_line);
 struct bio *pblk_bio_map_addr(struct pblk *pblk, void *data,
 			      unsigned int nr_secs, unsigned int len,
 			      gfp_t gfp_mask);
@@ -634,18 +679,23 @@ struct pblk_line *pblk_line_get_erase(struct pblk *pblk);
 int pblk_line_erase(struct pblk *pblk, struct pblk_line *line);
 int pblk_line_is_full(struct pblk_line *line);
 void pblk_line_free(struct pblk *pblk, struct pblk_line *line);
-void pblk_line_close_ws(struct work_struct *work);
+void pblk_line_close_meta(struct pblk *pblk, struct pblk_line *line);
 void pblk_line_close(struct pblk *pblk, struct pblk_line *line);
+void pblk_line_close_ws(struct work_struct *work);
 void pblk_line_mark_bb(struct work_struct *work);
 void pblk_line_run_ws(struct pblk *pblk, struct pblk_line *line, void *priv,
 		      void (*work)(struct work_struct *));
 u64 pblk_line_smeta_start(struct pblk *pblk, struct pblk_line *line);
 int pblk_line_read_smeta(struct pblk *pblk, struct pblk_line *line);
-int pblk_line_read_emeta(struct pblk *pblk, struct pblk_line *line);
+int pblk_line_read_emeta(struct pblk *pblk, struct pblk_line *line,
+			 void *emeta_buf);
 int pblk_blk_erase_async(struct pblk *pblk, struct ppa_addr erase_ppa);
 void pblk_line_put(struct kref *ref);
 struct list_head *pblk_line_gc_list(struct pblk *pblk, struct pblk_line *line);
+u64 pblk_lookup_page(struct pblk *pblk, struct pblk_line *line);
+void pblk_dealloc_page(struct pblk *pblk, struct pblk_line *line, int nr_secs);
 u64 pblk_alloc_page(struct pblk *pblk, struct pblk_line *line, int nr_secs);
+u64 __pblk_alloc_page(struct pblk *pblk, struct pblk_line *line, int nr_secs);
 int pblk_calc_secs(struct pblk *pblk, unsigned long secs_avail,
 		   unsigned long secs_to_flush);
 void pblk_down_rq(struct pblk *pblk, struct ppa_addr *ppa_list, int nr_ppas,
@@ -775,9 +825,19 @@ static inline struct nvm_rq *nvm_rq_from_c_ctx(void *c_ctx)
 	return c_ctx - sizeof(struct nvm_rq);
 }
 
-static inline void *pblk_line_emeta_to_lbas(struct line_emeta *emeta)
+static inline void *emeta_to_bb(struct line_emeta *emeta)
 {
-	return (emeta) + 1;
+	return emeta->bb_bitmap;
+}
+
+static inline void *emeta_to_lbas(struct pblk *pblk, struct line_emeta *emeta)
+{
+	return ((void *)emeta + pblk->lm.emeta_len[1]);
+}
+
+static inline void *emeta_to_vsc(struct pblk *pblk, struct line_emeta *emeta)
+{
+	return (emeta_to_lbas(pblk, emeta) + pblk->lm.emeta_len[2]);
 }
 
 #define NVM_MEM_PAGE_WRITE (8)
@@ -965,11 +1025,11 @@ static inline struct ppa_addr addr_to_pblk_ppa(struct pblk *pblk, u64 paddr,
 }
 
 static inline u32 pblk_calc_meta_header_crc(struct pblk *pblk,
-					    struct line_smeta *smeta)
+					    struct line_header *header)
 {
 	u32 crc = ~(u32)0;
 
-	crc = crc32_le(crc, (unsigned char *)smeta + sizeof(crc),
+	crc = crc32_le(crc, (unsigned char *)header + sizeof(crc),
 				sizeof(struct line_header) - sizeof(crc));
 
 	return crc;
@@ -997,7 +1057,7 @@ static inline u32 pblk_calc_emeta_crc(struct pblk *pblk,
 
 	crc = crc32_le(crc, (unsigned char *)emeta +
 				sizeof(struct line_header) + sizeof(crc),
-				lm->emeta_len -
+				lm->emeta_len[0] -
 				sizeof(struct line_header) - sizeof(crc));
 
 	return crc;