Skip to contentSkip to conversation

scsi: support packing multi-segment in UNMAP command

PATCH Superseded Chao Yu 3 participants 16 messages Mar 10, 2023 – Oct 7, 2023

Subsystems (2)

v5 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Chao Yu

author
#0 lore raw

PATCH v5 1/1 scsi: support packing multi-segment in UNMAP command

As SCSI SBC4 specification section 5.30.2 describes that it can support unmapping one or more LBA range in single UNMAP command.

However, previously we only pack one LBA range in UNMAP command by default no matter device gives the block limits that says it can support unmapping multiple LBA ranges with a single UNMAP command.

This patch sets max_discard_segments config according to block limits of device, and supports unmapping multiple LBA ranges with a single UNMAP command.

---
v5:
- fix to assign cmd->cmnd[7] and cmd->cmnd[8] w/ data_len correctly.
- fix to let ufshpb recognize and handle discard request which contains multiple ranges correctly.

 drivers/scsi/sd.c         | 36 +++++++++++++++++++++++++-----------
 drivers/scsi/sd.h         |  1 +
 drivers/ufs/core/ufshpb.c | 21 +++++++++++++++++++--
 3 files changed, 45 insertions(+), 13 deletions(-)
drivers/scsi/sd.c +25 -11
index 4bb87043e6db..1908b31c7342 100644
@@ -792,6 +792,8 @@ static void sd_config_discard(struct scsi_disk *sdkp, unsigned int mode)
 	q->limits.discard_granularity =
 		max(sdkp->physical_block_size,
 		    sdkp->unmap_granularity * logical_block_size);
+	blk_queue_max_discard_segments(q, min_t(u32, U16_MAX,
+				sdkp->max_unmap_block_desc_count));
 	sdkp->provisioning_mode = mode;
 
 	switch (mode) {
@@ -851,9 +853,10 @@ static blk_status_t sd_setup_unmap_cmnd(struct scsi_cmnd *cmd)
 	struct scsi_device *sdp = cmd->device;
 	struct request *rq = scsi_cmd_to_rq(cmd);
 	struct scsi_disk *sdkp = scsi_disk(rq->q->disk);
-	u64 lba = sectors_to_logical(sdp, blk_rq_pos(rq));
-	u32 nr_blocks = sectors_to_logical(sdp, blk_rq_sectors(rq));
-	unsigned int data_len = 24;
+	unsigned short segments = blk_rq_nr_discard_segments(rq);
+	unsigned int data_len = 8 + 16 * segments;
+	unsigned int descriptor_offset = 8;
+	struct bio *bio;
 	char *buf;
 
 	buf = sd_set_special_bvec(rq, data_len);
@@ -862,12 +865,20 @@ static blk_status_t sd_setup_unmap_cmnd(struct scsi_cmnd *cmd)
 
 	cmd->cmd_len = 10;
 	cmd->cmnd[0] = UNMAP;
-	cmd->cmnd[8] = 24;
+	cmd->cmnd[7] = data_len >> 8;
+	cmd->cmnd[8] = data_len & 0xff;
+
+	put_unaligned_be16(6 + 16 * segments, &buf[0]);
+	put_unaligned_be16(16 * segments, &buf[2]);
 
-	put_unaligned_be16(6 + 16, &buf[0]);
-	put_unaligned_be16(16, &buf[2]);
-	put_unaligned_be64(lba, &buf[8]);
-	put_unaligned_be32(nr_blocks, &buf[16]);
+	__rq_for_each_bio(bio, rq) {
+		u64 lba = sectors_to_logical(sdp, bio->bi_iter.bi_sector);
+		u32 nr_blocks = sectors_to_logical(sdp, bio_sectors(bio));
+
+		put_unaligned_be64(lba, &buf[descriptor_offset]);
+		put_unaligned_be32(nr_blocks, &buf[descriptor_offset + 8]);
+		descriptor_offset += 16;
+	}
 
 	cmd->allowed = sdkp->max_retries;
 	cmd->transfersize = data_len;
@@ -2917,7 +2928,7 @@ static void sd_read_block_limits(struct scsi_disk *sdkp)
 	sdkp->opt_xfer_blocks = get_unaligned_be32(&vpd->data[12]);
 
 	if (vpd->len >= 64) {
-		unsigned int lba_count, desc_count;
+		unsigned int lba_count;
 
 		sdkp->max_ws_blocks = (u32)get_unaligned_be64(&vpd->data[36]);
 
@@ -2925,9 +2936,12 @@ static void sd_read_block_limits(struct scsi_disk *sdkp)
 			goto out;
 
 		lba_count = get_unaligned_be32(&vpd->data[20]);
-		desc_count = get_unaligned_be32(&vpd->data[24]);
 
-		if (lba_count && desc_count)
+		/* Extract the MAXIMUM UNMAP BLOCK DESCRIPTOR COUNT. */
+		sdkp->max_unmap_block_desc_count =
+					get_unaligned_be32(&vpd->data[24]);
+
+		if (lba_count && sdkp->max_unmap_block_desc_count)
 			sdkp->max_unmap_blocks = lba_count;
 
 		sdkp->unmap_granularity = get_unaligned_be32(&vpd->data[28]);
drivers/scsi/sd.h +1 -0
index 5eea762f84d1..e7c51d23395b 100644
@@ -119,6 +119,7 @@ struct scsi_disk {
 	u32		opt_xfer_blocks;
 	u32		max_ws_blocks;
 	u32		max_unmap_blocks;
+	u32		max_unmap_block_desc_count;
 	u32		unmap_granularity;
 	u32		unmap_alignment;
 	u32		index;
drivers/ufs/core/ufshpb.c +19 -2
index a46a7666c891..327b29cf506f 100644
@@ -383,13 +383,30 @@ int ufshpb_prep(struct ufs_hba *hba, struct ufshcd_lrb *lrbp)
 	rgn = hpb->rgn_tbl + rgn_idx;
 	srgn = rgn->srgn_tbl + srgn_idx;
 
-	/* If command type is WRITE or DISCARD, set bitmap as dirty */
-	if (ufshpb_is_write_or_discard(cmd)) {
+	/* If command type is WRITE, set bitmap as dirty */
+	if (op_is_write(req_op(scsi_cmd_to_rq(cmd)))) {
 		ufshpb_iterate_rgn(hpb, rgn_idx, srgn_idx, srgn_offset,
 				   transfer_len, true);
 		return 0;
 	}
 
+	/* If command type is DISCARD, set bitmap as dirty */
+	if (op_is_discard(req_op(scsi_cmd_to_rq(cmd)))) {
+		struct bio *bio;
+
+		__rq_for_each_bio(bio, scsi_cmd_to_rq(cmd)) {
+			lpn = sectors_to_logical(cmd->device,
+						bio->bi_iter.bi_sector);
+			transfer_len = sectors_to_logical(cmd->device,
+							bio_sectors(bio));
+			ufshpb_get_pos_from_lpn(hpb, lpn, &rgn_idx,
+						&srgn_idx, &srgn_offset);
+			ufshpb_iterate_rgn(hpb, rgn_idx, srgn_idx,
+					srgn_offset, transfer_len, true);
+		}
+		return 0;
+	}
+
 	if (!ufshpb_is_supported_chunk(hpb, transfer_len))
 		return 0;
 
Signature · 1 line
2.25.1
#1 lore raw

drivers/ufs/core/ufshpb.c int ufshpb_prep(struct ufs_hba *hba, struct ufshcd_lrb *lrbp) · line 387

Quoted 4 lines
drivers/ufs/core/ufshpb.c +2 -2
-	/* If command type is WRITE or DISCARD, set bitmap as dirty */
-	if (ufshpb_is_write_or_discard(cmd)) {
+	/* If command type is WRITE, set bitmap as dirty */
+	if (op_is_write(req_op(scsi_cmd_to_rq(cmd)))) {

Umm, a driver has absolutely no business poking into the UNMAP
payload. Someone needs to fix the UFS driver first to not do this.

author
#2 lore raw
Quoted 7 lines · Christoph Hellwig

drivers/ufs/core/ufshpb.c int ufshpb_prep(struct ufs_hba *hba, struct ufshcd_lrb *lrbp) · line 387

Quoted 4 lines
drivers/ufs/core/ufshpb.c +2 -2
-	/* If command type is WRITE or DISCARD, set bitmap as dirty */
-	if (ufshpb_is_write_or_discard(cmd)) {
+	/* If command type is WRITE, set bitmap as dirty */
+	if (op_is_write(req_op(scsi_cmd_to_rq(cmd)))) {

Umm, a driver has absolutely no business poking into the UNMAP
payload. Someone needs to fix the UFS driver first to not do this.

IIUC,originally, HPB driver tries to lookup LBA range{,s} from WRITE/DISCARD request, and will dirty mapped HPB regions based on LBA range{,s}, do you mean HPB driver should not parse DISCARD request?

#3 lore raw
Quoted 3 lines · Chao Yu

IIUC,originally, HPB driver tries to lookup LBA range{,s} from WRITE/DISCARD request, and will dirty mapped HPB regions based on LBA range{,s}, do you mean HPB driver should not parse DISCARD request?

Eww.. This just means we really need to drop this bogus HPB code. It will otherwise break every single time when a new SCSI command get added that modifies LBA content.

author
#4 lore raw
In reply to Chao Yu

SCSI maintainers,

Any comments on this patch?

To Christoph, should I just drop HPB part change?

Quoted 149 lines · Chao Yu

As SCSI SBC4 specification section 5.30.2 describes that it can support unmapping one or more LBA range in single UNMAP command.

However, previously we only pack one LBA range in UNMAP command by default no matter device gives the block limits that says it can support unmapping multiple LBA ranges with a single UNMAP command.

This patch sets max_discard_segments config according to block limits of device, and supports unmapping multiple LBA ranges with a single UNMAP command.

---
v5:
- fix to assign cmd->cmnd[7] and cmd->cmnd[8] w/ data_len correctly.
- fix to let ufshpb recognize and handle discard request which contains multiple ranges correctly.

  drivers/scsi/sd.c         | 36 +++++++++++++++++++++++++-----------
  drivers/scsi/sd.h         |  1 +
  drivers/ufs/core/ufshpb.c | 21 +++++++++++++++++++--
  3 files changed, 45 insertions(+), 13 deletions(-)
drivers/scsi/sd.c +25 -11
index 4bb87043e6db..1908b31c7342 100644
@@ -792,6 +792,8 @@ static void sd_config_discard(struct scsi_disk *sdkp, unsigned int mode)
  	q->limits.discard_granularity =
  		max(sdkp->physical_block_size,
  		    sdkp->unmap_granularity * logical_block_size);
+	blk_queue_max_discard_segments(q, min_t(u32, U16_MAX,
+				sdkp->max_unmap_block_desc_count));
  	sdkp->provisioning_mode = mode;
 
  	switch (mode) {
@@ -851,9 +853,10 @@ static blk_status_t sd_setup_unmap_cmnd(struct scsi_cmnd *cmd)
  	struct scsi_device *sdp = cmd->device;
  	struct request *rq = scsi_cmd_to_rq(cmd);
  	struct scsi_disk *sdkp = scsi_disk(rq->q->disk);
-	u64 lba = sectors_to_logical(sdp, blk_rq_pos(rq));
-	u32 nr_blocks = sectors_to_logical(sdp, blk_rq_sectors(rq));
-	unsigned int data_len = 24;
+	unsigned short segments = blk_rq_nr_discard_segments(rq);
+	unsigned int data_len = 8 + 16 * segments;
+	unsigned int descriptor_offset = 8;
+	struct bio *bio;
  	char *buf;
 
  	buf = sd_set_special_bvec(rq, data_len);
@@ -862,12 +865,20 @@ static blk_status_t sd_setup_unmap_cmnd(struct scsi_cmnd *cmd)
 
  	cmd->cmd_len = 10;
  	cmd->cmnd[0] = UNMAP;
-	cmd->cmnd[8] = 24;
+	cmd->cmnd[7] = data_len >> 8;
+	cmd->cmnd[8] = data_len & 0xff;
+
+	put_unaligned_be16(6 + 16 * segments, &buf[0]);
+	put_unaligned_be16(16 * segments, &buf[2]);
 
-	put_unaligned_be16(6 + 16, &buf[0]);
-	put_unaligned_be16(16, &buf[2]);
-	put_unaligned_be64(lba, &buf[8]);
-	put_unaligned_be32(nr_blocks, &buf[16]);
+	__rq_for_each_bio(bio, rq) {
+		u64 lba = sectors_to_logical(sdp, bio->bi_iter.bi_sector);
+		u32 nr_blocks = sectors_to_logical(sdp, bio_sectors(bio));
+
+		put_unaligned_be64(lba, &buf[descriptor_offset]);
+		put_unaligned_be32(nr_blocks, &buf[descriptor_offset + 8]);
+		descriptor_offset += 16;
+	}
 
  	cmd->allowed = sdkp->max_retries;
  	cmd->transfersize = data_len;
@@ -2917,7 +2928,7 @@ static void sd_read_block_limits(struct scsi_disk *sdkp)
  	sdkp->opt_xfer_blocks = get_unaligned_be32(&vpd->data[12]);
 
  	if (vpd->len >= 64) {
-		unsigned int lba_count, desc_count;
+		unsigned int lba_count;
 
  		sdkp->max_ws_blocks = (u32)get_unaligned_be64(&vpd->data[36]);
 
@@ -2925,9 +2936,12 @@ static void sd_read_block_limits(struct scsi_disk *sdkp)
  			goto out;
 
  		lba_count = get_unaligned_be32(&vpd->data[20]);
-		desc_count = get_unaligned_be32(&vpd->data[24]);
 
-		if (lba_count && desc_count)
+		/* Extract the MAXIMUM UNMAP BLOCK DESCRIPTOR COUNT. */
+		sdkp->max_unmap_block_desc_count =
+					get_unaligned_be32(&vpd->data[24]);
+
+		if (lba_count && sdkp->max_unmap_block_desc_count)
  			sdkp->max_unmap_blocks = lba_count;
 
  		sdkp->unmap_granularity = get_unaligned_be32(&vpd->data[28]);
drivers/scsi/sd.h +1 -0
index 5eea762f84d1..e7c51d23395b 100644
@@ -119,6 +119,7 @@ struct scsi_disk {
  	u32		opt_xfer_blocks;
  	u32		max_ws_blocks;
  	u32		max_unmap_blocks;
+	u32		max_unmap_block_desc_count;
  	u32		unmap_granularity;
  	u32		unmap_alignment;
  	u32		index;
drivers/ufs/core/ufshpb.c +19 -2
index a46a7666c891..327b29cf506f 100644
@@ -383,13 +383,30 @@ int ufshpb_prep(struct ufs_hba *hba, struct ufshcd_lrb *lrbp)
  	rgn = hpb->rgn_tbl + rgn_idx;
  	srgn = rgn->srgn_tbl + srgn_idx;
 
-	/* If command type is WRITE or DISCARD, set bitmap as dirty */
-	if (ufshpb_is_write_or_discard(cmd)) {
+	/* If command type is WRITE, set bitmap as dirty */
+	if (op_is_write(req_op(scsi_cmd_to_rq(cmd)))) {
  		ufshpb_iterate_rgn(hpb, rgn_idx, srgn_idx, srgn_offset,
  				   transfer_len, true);
  		return 0;
  	}
 
+	/* If command type is DISCARD, set bitmap as dirty */
+	if (op_is_discard(req_op(scsi_cmd_to_rq(cmd)))) {
+		struct bio *bio;
+
+		__rq_for_each_bio(bio, scsi_cmd_to_rq(cmd)) {
+			lpn = sectors_to_logical(cmd->device,
+						bio->bi_iter.bi_sector);
+			transfer_len = sectors_to_logical(cmd->device,
+							bio_sectors(bio));
+			ufshpb_get_pos_from_lpn(hpb, lpn, &rgn_idx,
+						&srgn_idx, &srgn_offset);
+			ufshpb_iterate_rgn(hpb, rgn_idx, srgn_idx,
+					srgn_offset, transfer_len, true);
+		}
+		return 0;
+	}
+
  	if (!ufshpb_is_supported_chunk(hpb, transfer_len))
  		return 0;
 
#5 lore raw

Chao,

Quoted 3 lines

SCSI maintainers,

Any comments on this patch?

I have an updated version of your original multi-segment UNMAP support queued in a discard topic branch that I intend to submit shortly. Without the UFS HPB bits.

Signature · 1 line
Martin K. Petersen	Oracle Linux Engineering
#7 lore raw

Chao,

Quoted 1 line

Any progress on this patch?

I'll resubmit this series for 6.6.

Signature · 1 line
Martin K. Petersen	Oracle Linux Engineering
author
#8 lore raw
Quoted 6 lines · Martin K. Petersen

Chao,

Quoted 1 line

Any progress on this patch?

I'll resubmit this series for 6.6.

Martin,

Thank you!

Thanks,

Quoted 1 line
#11 lore raw

Chao,

Quoted 2 lines
Quoted 2 lines
Quoted 1 line

Any progress on this patch?

I'll resubmit this series for 6.6.

Been working on this series to address the reported regressions. Spent quite a bit of time on it last week.

Signature · 1 line
Martin K. Petersen	Oracle Linux Engineering
author
#12 lore raw
Quoted 2 lines · Martin K. Petersen

Been working on this series to address the reported regressions. Spent quite a bit of time on it last week.

Martin, thanks for the help. :)

Could you please share the report? Maybe I can join to check it.

Thanks,

author
#13 lore raw
Quoted 7 lines · Chao Yu
Quoted 2 lines · Martin K. Petersen

Been working on this series to address the reported regressions. Spent quite a bit of time on it last week.

Martin, thanks for the help. :)

Could you please share the report? Maybe I can join to check it.

Oh, I may misunderstand it, the reported regression wasn't caused by this patch, right?

Thanks,

Quoted 2 lines

Thanks,

author
#14 lore raw
In reply to Chao Yu
Quoted 7 lines · Chao Yu
Quoted 2 lines · Martin K. Petersen

Been working on this series to address the reported regressions. Spent quite a bit of time on it last week.

Martin, thanks for the help. :)

Could you please share the report? Maybe I can join to check it.

It looks we will miss another merge window again... :(

Not sure I understand it correctly...
If this patch caused regression, could you please share the report w/ me?

Thanks,

Quoted 2 lines

Thanks,

author
#15 lore raw
In reply to Martin K. Petersen

Hi Martin,

Is it possible to commit this patch only to catch up 6.7-rc1, w/o other changes in 5.20/discovery? I don't feel there is any strong correlation between them, or am I missing something?

Thanks,

Quoted 9 lines · Martin K. Petersen

Chao,

Quoted 2 lines
Quoted 2 lines
Quoted 1 line

Any progress on this patch?

I'll resubmit this series for 6.6.

Been working on this series to address the reported regressions. Spent quite a bit of time on it last week.