Skip to contentSkip to conversation

Six UFS patches

PATCH In review Bart Van Assche 5 participants 28 messages Dec 24, 2019 – May 29, 2020 SCSI

lore
Share & tools
lore
#0 lore raw

cover Six UFS patches

Hi Martin,

This patch series includes patches that clean up the UFS driver code somewhat. Please consider these patches for kernel v5.6.

Thanks,

Bart.

Bart Van Assche (6):
ufs: Fix indentation in ufshcd_query_attr_retry()
ufs: Make ufshcd_add_command_trace() easier to read
ufs: Make ufshcd_prepare_utp_scsi_cmd_upiu() easier to read
ufs: Fix a race condition in the tracing code
ufs: Remove superfluous memory barriers
ufs: Remove the SCSI timeout handler

 drivers/scsi/ufs/ufshcd.c | 64 ++++++++-------------------------------
 1 file changed, 12 insertions(+), 52 deletions(-)
#1 lore raw

PATCH 1/6 ufs: Fix indentation in ufshcd_query_attr_retry()

Remove a space that occurs after a tab.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +1 -1
index d6d0f83c9044..48f2f94d51bc 100644
@@ -2869,7 +2869,7 @@ static int ufshcd_query_attr_retry(struct ufs_hba *hba,
 	int ret = 0;
 	u32 retries;
 
-	 for (retries = QUERY_REQ_RETRIES; retries > 0; retries--) {
+	for (retries = QUERY_REQ_RETRIES; retries > 0; retries--) {
 		ret = ufshcd_query_attr(hba, opcode, idn, index,
 						selector, attr_val);
 		if (ret)
Reviewed-by on v1 1/6
#2 lore raw

Re: [PATCH 1/6] ufs: Fix indentation in ufshcd_query_attr_retry()

Hi Bart,

Quoted 17 lines · Bart Van Assche

Remove a space that occurs after a tab.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +0 -0
index d6d0f83c9044..48f2f94d51bc 100644

drivers/scsi/ufs/ufshcd.c static int ufshcd_query_attr_retry(struct ufs_hba *hba, · line 2875

Quoted 8 lines
drivers/scsi/ufs/ufshcd.c +1 -1
@@ -2869,7 +2869,7 @@ static int ufshcd_query_attr_retry(struct ufs_hba *hba,
 	int ret = 0;
 	u32 retries;
 
-	 for (retries = QUERY_REQ_RETRIES; retries > 0; retries--) {
+	for (retries = QUERY_REQ_RETRIES; retries > 0; retries--) {
 		ret = ufshcd_query_attr(hba, opcode, idn, index,
 						selector, attr_val);
 		if (ret)
Reviewed-by on v1 1/6
#3 lore raw

Re: [PATCH 1/6] ufs: Fix indentation in ufshcd_query_attr_retry()

In reply to Bart Van Assche
Quoted 8 lines · Bart Van Assche

Remove a space that occurs after a tab.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
Quoted 18 lines
---
 drivers/scsi/ufs/ufshcd.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +1 -1
index d6d0f83c9044..48f2f94d51bc 100644
@@ -2869,7 +2869,7 @@ static int ufshcd_query_attr_retry(struct ufs_hba 
 *hba,
 	int ret = 0;
 	u32 retries;
 
-	 for (retries = QUERY_REQ_RETRIES; retries > 0; retries--) {
+	for (retries = QUERY_REQ_RETRIES; retries > 0; retries--) {
 		ret = ufshcd_query_attr(hba, opcode, idn, index,
 						selector, attr_val);

if (ret)

Reviewed-by Alim Akhar on v1 1/6
#4 lore raw

Re: [PATCH 1/6] ufs: Fix indentation in ufshcd_query_attr_retry()

In reply to Bart Van Assche

Hi Bart

Quoted 10 lines · Bart Van Assche

Remove a space that occurs after a tab.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>

---

Quoted 16 lines
 drivers/scsi/ufs/ufshcd.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +1 -1
index d6d0f83c9044..48f2f94d51bc 100644
@@ -2869,7 +2869,7 @@ static int ufshcd_query_attr_retry(struct ufs_hba *hba,
        int ret = 0;
        u32 retries;
 
-        for (retries = QUERY_REQ_RETRIES; retries > 0; retries--) {
+       for (retries = QUERY_REQ_RETRIES; retries > 0; retries--) {
                ret = ufshcd_query_attr(hba, opcode, idn, index,
                                                selector, attr_val);
                if (ret)
Signature · 2 lines
Regards,
Alim
#5 lore raw

PATCH 2/6 ufs: Make ufshcd_add_command_trace() easier to read

In reply to Bart Van Assche

Since the lrbp->cmd expression occurs multiple times, introduce a new local variable to hold that pointer. This patch does not change any functionality.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)
drivers/scsi/ufs/ufshcd.c +6 -6
index 48f2f94d51bc..acc84e964e8f 100644
@@ -327,27 +327,27 @@ static void ufshcd_add_command_trace(struct ufs_hba *hba,
 	u8 opcode = 0;
 	u32 intr, doorbell;
 	struct ufshcd_lrb *lrbp = &hba->lrb[tag];
+	struct scsi_cmnd *cmd = lrbp->cmd;
 	int transfer_len = -1;
 
 	if (!trace_ufshcd_command_enabled()) {
 		/* trace UPIU W/O tracing command */
-		if (lrbp->cmd)
+		if (cmd)
 			ufshcd_add_cmd_upiu_trace(hba, tag, str);
 		return;
 	}
 
-	if (lrbp->cmd) { /* data phase exists */
+	if (cmd) { /* data phase exists */
 		/* trace UPIU also */
 		ufshcd_add_cmd_upiu_trace(hba, tag, str);
-		opcode = (u8)(*lrbp->cmd->cmnd);
+		opcode = cmd->cmnd[0];
 		if ((opcode == READ_10) || (opcode == WRITE_10)) {
 			/*
 			 * Currently we only fully trace read(10) and write(10)
 			 * commands
 			 */
-			if (lrbp->cmd->request && lrbp->cmd->request->bio)
-				lba =
-				  lrbp->cmd->request->bio->bi_iter.bi_sector;
+			if (cmd->request && cmd->request->bio)
+				lba = cmd->request->bio->bi_iter.bi_sector;
 			transfer_len = be32_to_cpu(
 				lrbp->ucd_req_ptr->sc.exp_data_transfer_len);
 		}
Reviewed-by on v1 2/6
#6 lore raw

Re: [PATCH 2/6] ufs: Make ufshcd_add_command_trace() easier to read

Hi Bart,

Quoted 11 lines · Bart Van Assche

Since the lrbp->cmd expression occurs multiple times, introduce a new local variable to hold that pointer. This patch does not change any functionality.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>

---

Reviewed-by on v1 2/6
#7 lore raw

Re: [PATCH 2/6] ufs: Make ufshcd_add_command_trace() easier to read

In reply to Bart Van Assche
Quoted 10 lines · Bart Van Assche

Since the lrbp->cmd expression occurs multiple times, introduce a new local variable to hold that pointer. This patch does not change any functionality.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
Quoted 43 lines
---
 drivers/scsi/ufs/ufshcd.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)
drivers/scsi/ufs/ufshcd.c +6 -6
index 48f2f94d51bc..acc84e964e8f 100644
@@ -327,27 +327,27 @@ static void ufshcd_add_command_trace(struct 
 ufs_hba *hba,
 	u8 opcode = 0;
 	u32 intr, doorbell;
 	struct ufshcd_lrb *lrbp = &hba->lrb[tag];
+	struct scsi_cmnd *cmd = lrbp->cmd;
 	int transfer_len = -1;
 
 	if (!trace_ufshcd_command_enabled()) {
 		/* trace UPIU W/O tracing command */
-		if (lrbp->cmd)
+		if (cmd)
 			ufshcd_add_cmd_upiu_trace(hba, tag, str);
 		return;
 	}
 
-	if (lrbp->cmd) { /* data phase exists */
+	if (cmd) { /* data phase exists */
 		/* trace UPIU also */
 		ufshcd_add_cmd_upiu_trace(hba, tag, str);
-		opcode = (u8)(*lrbp->cmd->cmnd);
+		opcode = cmd->cmnd[0];
 		if ((opcode == READ_10) || (opcode == WRITE_10)) {
 			/*
 			 * Currently we only fully trace read(10) and write(10)
 			 * commands
 			 */
-			if (lrbp->cmd->request && lrbp->cmd->request->bio)
-				lba =
-				  lrbp->cmd->request->bio->bi_iter.bi_sector;
+			if (cmd->request && cmd->request->bio)
+				lba = cmd->request->bio->bi_iter.bi_sector;
 			transfer_len = be32_to_cpu(
 				lrbp->ucd_req_ptr->sc.exp_data_transfer_len);

}

Reviewed-by on v1 2/6
#8 lore raw

Re: [PATCH 2/6] ufs: Make ufshcd_add_command_trace() easier to read

In reply to Bart Van Assche
Quoted 5 lines · Bart Van Assche

Since the lrbp->cmd expression occurs multiple times, introduce a new local variable to hold that pointer. This patch does not change any functionality.

FYMI, Any Advantage of doing this? or it just we don't want to fetch cmd from lrbp every time? I mean in terms of execution speed.

Quoted 7 lines
Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>

---

Quoted 41 lines
 drivers/scsi/ufs/ufshcd.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)
drivers/scsi/ufs/ufshcd.c +6 -6
index 48f2f94d51bc..acc84e964e8f 100644
@@ -327,27 +327,27 @@ static void ufshcd_add_command_trace(struct ufs_hba *hba,
        u8 opcode = 0;
        u32 intr, doorbell;
        struct ufshcd_lrb *lrbp = &hba->lrb[tag];
+       struct scsi_cmnd *cmd = lrbp->cmd;
        int transfer_len = -1;
 
        if (!trace_ufshcd_command_enabled()) {
                /* trace UPIU W/O tracing command */
-               if (lrbp->cmd)
+               if (cmd)
                        ufshcd_add_cmd_upiu_trace(hba, tag, str);
                return;
        }
 
-       if (lrbp->cmd) { /* data phase exists */
+       if (cmd) { /* data phase exists */
                /* trace UPIU also */
                ufshcd_add_cmd_upiu_trace(hba, tag, str);
-               opcode = (u8)(*lrbp->cmd->cmnd);
+               opcode = cmd->cmnd[0];
                if ((opcode == READ_10) || (opcode == WRITE_10)) {
                        /*
                         * Currently we only fully trace read(10) and write(10)
                         * commands
                         */
-                       if (lrbp->cmd->request && lrbp->cmd->request->bio)
-                               lba =
-                                 lrbp->cmd->request->bio->bi_iter.bi_sector;
+                       if (cmd->request && cmd->request->bio)
+                               lba = cmd->request->bio->bi_iter.bi_sector;
                        transfer_len = be32_to_cpu(
                                lrbp->ucd_req_ptr->sc.exp_data_transfer_len);
                }
Signature · 2 lines
Regards,
Alim
#9 lore raw

PATCH 3/6 ufs: Make ufshcd_prepare_utp_scsi_cmd_upiu() easier to read

In reply to Bart Van Assche

Since the lrbp->cmd expression occurs multiple times, introduce a new local variable to hold that pointer. This patch does not change any functionality.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/scsi/ufs/ufshcd.c +4 -4
index acc84e964e8f..80b028ba39e9 100644
@@ -2233,6 +2233,7 @@ static void ufshcd_prepare_req_desc_hdr(struct ufshcd_lrb *lrbp,
 static
 void ufshcd_prepare_utp_scsi_cmd_upiu(struct ufshcd_lrb *lrbp, u32 upiu_flags)
 {
+	struct scsi_cmnd *cmd = lrbp->cmd;
 	struct utp_upiu_req *ucd_req_ptr = lrbp->ucd_req_ptr;
 	unsigned short cdb_len;
 
@@ -2246,12 +2247,11 @@ void ufshcd_prepare_utp_scsi_cmd_upiu(struct ufshcd_lrb *lrbp, u32 upiu_flags)
 	/* Total EHS length and Data segment length will be zero */
 	ucd_req_ptr->header.dword_2 = 0;
 
-	ucd_req_ptr->sc.exp_data_transfer_len =
-		cpu_to_be32(lrbp->cmd->sdb.length);
+	ucd_req_ptr->sc.exp_data_transfer_len = cpu_to_be32(cmd->sdb.length);
 
-	cdb_len = min_t(unsigned short, lrbp->cmd->cmd_len, UFS_CDB_SIZE);
+	cdb_len = min_t(unsigned short, cmd->cmd_len, UFS_CDB_SIZE);
 	memset(ucd_req_ptr->sc.cdb, 0, UFS_CDB_SIZE);
-	memcpy(ucd_req_ptr->sc.cdb, lrbp->cmd->cmnd, cdb_len);
+	memcpy(ucd_req_ptr->sc.cdb, cmd->cmnd, cdb_len);
 
 	memset(lrbp->ucd_rsp_ptr, 0, sizeof(struct utp_upiu_rsp));
 }
Reviewed-by on v1 3/6
#10 lore raw

Re: [PATCH 3/6] ufs: Make ufshcd_prepare_utp_scsi_cmd_upiu() easier to read

Hi Bart,

Quoted 10 lines · Bart Van Assche

Since the lrbp->cmd expression occurs multiple times, introduce a new local variable to hold that pointer. This patch does not change any functionality.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
Reviewed-by on v1 3/6
#11 lore raw

Re: [PATCH 3/6] ufs: Make ufshcd_prepare_utp_scsi_cmd_upiu() easier to read

In reply to Bart Van Assche
Quoted 10 lines · Bart Van Assche

Since the lrbp->cmd expression occurs multiple times, introduce a new local variable to hold that pointer. This patch does not change any functionality.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
Quoted 35 lines
---
 drivers/scsi/ufs/ufshcd.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/scsi/ufs/ufshcd.c +1 -0
index acc84e964e8f..80b028ba39e9 100644
@@ -2233,6 +2233,7 @@ static void ufshcd_prepare_req_desc_hdr(struct
 ufshcd_lrb *lrbp,
 static
 void ufshcd_prepare_utp_scsi_cmd_upiu(struct ufshcd_lrb *lrbp, u32 
 upiu_flags)
 {
+	struct scsi_cmnd *cmd = lrbp->cmd;
 	struct utp_upiu_req *ucd_req_ptr = lrbp->ucd_req_ptr;
drivers/scsi/ufs/ufshcd.c +3 -4
 	unsigned short cdb_len;
@@ -2246,12 +2247,11 @@ void ufshcd_prepare_utp_scsi_cmd_upiu(struct
 ufshcd_lrb *lrbp, u32 upiu_flags)
 	/* Total EHS length and Data segment length will be zero */
 	ucd_req_ptr->header.dword_2 = 0;
 
-	ucd_req_ptr->sc.exp_data_transfer_len =
-		cpu_to_be32(lrbp->cmd->sdb.length);
+	ucd_req_ptr->sc.exp_data_transfer_len = cpu_to_be32(cmd->sdb.length);
 
-	cdb_len = min_t(unsigned short, lrbp->cmd->cmd_len, UFS_CDB_SIZE);
+	cdb_len = min_t(unsigned short, cmd->cmd_len, UFS_CDB_SIZE);
 	memset(ucd_req_ptr->sc.cdb, 0, UFS_CDB_SIZE);
-	memcpy(ucd_req_ptr->sc.cdb, lrbp->cmd->cmnd, cdb_len);
+	memcpy(ucd_req_ptr->sc.cdb, cmd->cmnd, cdb_len);
 
 	memset(lrbp->ucd_rsp_ptr, 0, sizeof(struct utp_upiu_rsp));

}

#12 lore raw

PATCH 4/6 ufs: Fix a race condition in the tracing code

In reply to Bart Van Assche

Starting execution of a command before tracing a command may cause the completion handler to free data while it is being traced. Fix this race by tracing a command before it is submitted.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +1 -1
index 80b028ba39e9..4d9bb1932b39 100644
@@ -1875,12 +1875,12 @@ void ufshcd_send_command(struct ufs_hba *hba, unsigned int task_tag)
 {
 	hba->lrb[task_tag].issue_time_stamp = ktime_get();
 	hba->lrb[task_tag].compl_time_stamp = ktime_set(0, 0);
+	ufshcd_add_command_trace(hba, task_tag, "send");
 	ufshcd_clk_scaling_start_busy(hba);
 	__set_bit(task_tag, &hba->outstanding_reqs);
 	ufshcd_writel(hba, 1 << task_tag, REG_UTP_TRANSFER_REQ_DOOR_BELL);
 	/* Make sure that doorbell is committed immediately */
 	wmb();
-	ufshcd_add_command_trace(hba, task_tag, "send");
 }
 
 /**
#13 lore raw

Re: [PATCH 4/6] ufs: Fix a race condition in the tracing code

Quoted 19 lines · Bart Van Assche

Starting execution of a command before tracing a command may cause the completion handler to free data while it is being traced. Fix this race by tracing a command before it is submitted.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +0 -0
index 80b028ba39e9..4d9bb1932b39 100644

drivers/scsi/ufs/ufshcd.c void ufshcd_send_command(struct ufs_hba *hba, · line 1879

Quoted 5 lines
drivers/scsi/ufs/ufshcd.c +1 -0
@@ -1875,12 +1875,12 @@ void ufshcd_send_command(struct ufs_hba *hba,
 unsigned int task_tag)
 {
 	hba->lrb[task_tag].issue_time_stamp = ktime_get();
 	hba->lrb[task_tag].compl_time_stamp = ktime_set(0, 0);
+	ufshcd_add_command_trace(hba, task_tag, "send");

How about moving it after __set_bit(task_tag, &hba->outstanding_reqs);?

Thanks,

Can Guo.

Quoted 9 lines
drivers/scsi/ufs/ufshcd.c +0 -1
 	ufshcd_clk_scaling_start_busy(hba);
 	__set_bit(task_tag, &hba->outstanding_reqs);
 	ufshcd_writel(hba, 1 << task_tag, REG_UTP_TRANSFER_REQ_DOOR_BELL);
 	/* Make sure that doorbell is committed immediately */
 	wmb();
-	ufshcd_add_command_trace(hba, task_tag, "send");
 }
 

/**

#14 lore raw

Re: [PATCH 4/6] ufs: Fix a race condition in the tracing code

Quoted 28 lines · Can Guo
Quoted 19 lines · Bart Van Assche

Starting execution of a command before tracing a command may cause the completion handler to free data while it is being traced. Fix this race by tracing a command before it is submitted.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>

---
 drivers/scsi/ufs/ufshcd.c | 2 +-

 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +0 -0
index 80b028ba39e9..4d9bb1932b39 100644

drivers/scsi/ufs/ufshcd.c void ufshcd_send_command(struct ufs_hba *hba, · line 1879

Quoted 5 lines
drivers/scsi/ufs/ufshcd.c +1 -0
@@ -1875,12 +1875,12 @@ void ufshcd_send_command(struct ufs_hba *hba,
 unsigned int task_tag)
  {
      hba->lrb[task_tag].issue_time_stamp = ktime_get();
      hba->lrb[task_tag].compl_time_stamp = ktime_set(0, 0);
+    ufshcd_add_command_trace(hba, task_tag, "send");

How about moving it after __set_bit(task_tag, &hba->outstanding_reqs);?

Quoted 11 lines

Thanks,

Can Guo.

Quoted 7 lines
drivers/scsi/ufs/ufshcd.c +0 -1
      ufshcd_clk_scaling_start_busy(hba);
      __set_bit(task_tag, &hba->outstanding_reqs);
      ufshcd_writel(hba, 1 << task_tag, REG_UTP_TRANSFER_REQ_DOOR_BELL);
      /* Make sure that doorbell is committed immediately */
      wmb();
-    ufshcd_add_command_trace(hba, task_tag, "send");
  }

Hi Can,

As far as I can see ufshcd_add_command_trace() does not read hba->outstanding_reqs so calling ufshcd_add_command_trace() before changing outstanding_reqs should be fine.

The order chosen in the posted patch should minimize energy usage of the UFS controller, namely by calling the tracing function before starting clock scaling.

Thanks,

Bart.

#15 lore raw

Re: [PATCH 4/6] ufs: Fix a race condition in the tracing code

Quoted 47 lines · Bart Van Assche
Quoted 28 lines · Can Guo
Quoted 19 lines · Bart Van Assche

Starting execution of a command before tracing a command may cause the completion handler to free data while it is being traced. Fix this race by tracing a command before it is submitted.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>

---
 drivers/scsi/ufs/ufshcd.c | 2 +-

 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +0 -0
index 80b028ba39e9..4d9bb1932b39 100644

drivers/scsi/ufs/ufshcd.c void ufshcd_send_command(struct ufs_hba *hba, · line 1879

Quoted 5 lines
drivers/scsi/ufs/ufshcd.c +1 -0
@@ -1875,12 +1875,12 @@ void ufshcd_send_command(struct ufs_hba *hba,
 unsigned int task_tag)
  {
      hba->lrb[task_tag].issue_time_stamp = ktime_get();
      hba->lrb[task_tag].compl_time_stamp = ktime_set(0, 0);
+    ufshcd_add_command_trace(hba, task_tag, "send");

How about moving it after __set_bit(task_tag, &hba->outstanding_reqs);?

Quoted 11 lines

Thanks,

Can Guo.

Quoted 7 lines
drivers/scsi/ufs/ufshcd.c +0 -1
      ufshcd_clk_scaling_start_busy(hba);
      __set_bit(task_tag, &hba->outstanding_reqs);
      ufshcd_writel(hba, 1 << task_tag, REG_UTP_TRANSFER_REQ_DOOR_BELL);
      /* Make sure that doorbell is committed immediately */
      wmb();
-    ufshcd_add_command_trace(hba, task_tag, "send");
  }

Hi Can,

As far as I can see ufshcd_add_command_trace() does not read hba->outstanding_reqs so calling ufshcd_add_command_trace() before changing outstanding_reqs should be fine.

Quoted 7 lines

The order chosen in the posted patch should minimize energy usage of the UFS controller, namely by calling the tracing function before starting clock scaling.

Thanks,

Bart.

True.

One concern regards this change is that since we move the ufshcd_add_command_trace() before ringing doorbell, in tracing log, we will lose the info/status of this specific tag in doorbell register.

Usually, with the old code, after we collect the tracing log, we can tell whether the doorbell bit was set/rung correctly in doorbell register or not for an specific tag/task, because it is an important sign of whether HW had actually accepted the task on that tag. It is really helpful sometime if things go wrong.

Any good ways to keep the doorbell info as before?

Thanks,

Can Guo.

Reviewed-by on v1 4/6
#16 lore raw

Re: [PATCH 4/6] ufs: Fix a race condition in the tracing code

In reply to Bart Van Assche
Quoted 5 lines · Bart Van Assche

Starting execution of a command before tracing a command may cause the completion handler to free data while it is being traced. Fix this race by tracing a command before it is submitted.

Quoted 7 lines
Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>

---

Quoted 21 lines
 drivers/scsi/ufs/ufshcd.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/scsi/ufs/ufshcd.c +1 -1
index 80b028ba39e9..4d9bb1932b39 100644
@@ -1875,12 +1875,12 @@ void ufshcd_send_command(struct ufs_hba *hba, unsigned int task_tag)
 {
        hba->lrb[task_tag].issue_time_stamp = ktime_get();
        hba->lrb[task_tag].compl_time_stamp = ktime_set(0, 0);
+       ufshcd_add_command_trace(hba, task_tag, "send");
        ufshcd_clk_scaling_start_busy(hba);
        __set_bit(task_tag, &hba->outstanding_reqs);
        ufshcd_writel(hba, 1 << task_tag, REG_UTP_TRANSFER_REQ_DOOR_BELL);
        /* Make sure that doorbell is committed immediately */
        wmb();
-       ufshcd_add_command_trace(hba, task_tag, "send");
 }
 
 /**
Signature · 2 lines
Regards,
Alim
#17 lore raw

PATCH 5/6 ufs: Remove superfluous memory barriers

In reply to Bart Van Assche

Calling wmb() after having written to a doorbell slows down code and does not help to commit the doorbell write faster. Hence remove such wmb() calls. Note: detailed information about the semantics of writel() is available in Documentation/driver-api/device-io.rst.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 4 ----
 1 file changed, 4 deletions(-)
drivers/scsi/ufs/ufshcd.c +0 -4
index 4d9bb1932b39..edcc137c436b 100644
@@ -1879,8 +1879,6 @@ void ufshcd_send_command(struct ufs_hba *hba, unsigned int task_tag)
 	ufshcd_clk_scaling_start_busy(hba);
 	__set_bit(task_tag, &hba->outstanding_reqs);
 	ufshcd_writel(hba, 1 << task_tag, REG_UTP_TRANSFER_REQ_DOOR_BELL);
-	/* Make sure that doorbell is committed immediately */
-	wmb();
 }
 
 /**
@@ -5766,8 +5764,6 @@ static int __ufshcd_issue_tm_cmd(struct ufs_hba *hba,
 	wmb();
 
 	ufshcd_writel(hba, 1 << free_slot, REG_UTP_TASK_REQ_DOOR_BELL);
-	/* Make sure that doorbell is committed immediately */
-	wmb();
 
 	spin_unlock_irqrestore(host->host_lock, flags);
 
#18 lore raw

Re: [PATCH 5/6] ufs: Remove superfluous memory barriers

Quoted 30 lines · Bart Van Assche

Calling wmb() after having written to a doorbell slows down code and does not help to commit the doorbell write faster. Hence remove such wmb() calls. Note: detailed information about the semantics of writel() is available in Documentation/driver-api/device-io.rst.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 4 ----
 1 file changed, 4 deletions(-)
drivers/scsi/ufs/ufshcd.c +0 -2
index 4d9bb1932b39..edcc137c436b 100644
@@ -1879,8 +1879,6 @@ void ufshcd_send_command(struct ufs_hba *hba,
 unsigned int task_tag)
 	ufshcd_clk_scaling_start_busy(hba);
 	__set_bit(task_tag, &hba->outstanding_reqs);
 	ufshcd_writel(hba, 1 << task_tag, REG_UTP_TRANSFER_REQ_DOOR_BELL);
-	/* Make sure that doorbell is committed immediately */
-	wmb();
 }
 

/**

drivers/scsi/ufs/ufshcd.c static int __ufshcd_issue_tm_cmd(struct ufs_hba · line 5769

Quoted 8 lines
drivers/scsi/ufs/ufshcd.c +0 -2
@@ -5766,8 +5764,6 @@ static int __ufshcd_issue_tm_cmd(struct ufs_hba 
 *hba,
 	wmb();
 
 	ufshcd_writel(hba, 1 << free_slot, REG_UTP_TASK_REQ_DOOR_BELL);
-	/* Make sure that doorbell is committed immediately */
-	wmb();
 
 	spin_unlock_irqrestore(host->host_lock, flags);

Hi Bart,

Three wmb()s were added in commit ad1a1b9cd because we did see instances on which OCS=3(MISMATCH_DATA_BUFFER_SIZE) error were observed in large scale test. Commit ad1a1b9cd fixed the error and we had confirmed it through large amount of tests. I am not sure removing the 2 wmb()s here would cause regression or not.

Thanks,

Can Guo.

#19 lore raw

Re: [PATCH 5/6] ufs: Remove superfluous memory barriers

Quoted 42 lines · Can Guo
Quoted 30 lines · Bart Van Assche

Calling wmb() after having written to a doorbell slows down code and does not help to commit the doorbell write faster. Hence remove such wmb() calls. Note: detailed information about the semantics of writel() is available in Documentation/driver-api/device-io.rst.

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>

---
 drivers/scsi/ufs/ufshcd.c | 4 ----

 1 file changed, 4 deletions(-)
drivers/scsi/ufs/ufshcd.c +0 -2
index 4d9bb1932b39..edcc137c436b 100644
@@ -1879,8 +1879,6 @@ void ufshcd_send_command(struct ufs_hba *hba,
 unsigned int task_tag)
      ufshcd_clk_scaling_start_busy(hba);
      __set_bit(task_tag, &hba->outstanding_reqs);
      ufshcd_writel(hba, 1 << task_tag, REG_UTP_TRANSFER_REQ_DOOR_BELL);
-    /* Make sure that doorbell is committed immediately */
-    wmb();
  }
 

 /**

drivers/scsi/ufs/ufshcd.c static int __ufshcd_issue_tm_cmd(struct ufs_hba · line 5769

Quoted 8 lines
drivers/scsi/ufs/ufshcd.c +0 -2
@@ -5766,8 +5764,6 @@ static int __ufshcd_issue_tm_cmd(struct ufs_hba 
 *hba,
      wmb();
 
      ufshcd_writel(hba, 1 << free_slot, REG_UTP_TASK_REQ_DOOR_BELL);
-    /* Make sure that doorbell is committed immediately */
-    wmb();
 
      spin_unlock_irqrestore(host->host_lock, flags);

Hi Bart,

Quoted 5 lines

Three wmb()s were added in commit ad1a1b9cd because we did see instances on which OCS=3(MISMATCH_DATA_BUFFER_SIZE) error were observed in large scale test. Commit ad1a1b9cd fixed the error and we had confirmed it through large amount of tests. I am not sure removing the 2 wmb()s here would cause regression or not.

Hi Can,

Thank you for having reported this. I will drop this patch.

Bart.

#20 lore raw

PATCH 6/6 ufs: Remove the SCSI timeout handler

In reply to Bart Van Assche

The UFS SCSI timeout handler was needed to compensate that ufshcd_queuecommand() could return SCSI_MLQUEUE_HOST_BUSY for a long time. Commit a276c19e3e98 ("scsi: ufs: Avoid busy-waiting by eliminating tag conflicts") fixed this so the timeout handler is no longer necessary.

See also commit f550c65b543b ("scsi: ufs: implement scsi host timeout handler").

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 36 ------------------------------------
 1 file changed, 36 deletions(-)
drivers/scsi/ufs/ufshcd.c +0 -36
index edcc137c436b..763e41286a4b 100644
@@ -7092,41 +7092,6 @@ static void ufshcd_async_scan(void *data, async_cookie_t cookie)
 	ufshcd_probe_hba(hba);
 }
 
-static enum blk_eh_timer_return ufshcd_eh_timed_out(struct scsi_cmnd *scmd)
-{
-	unsigned long flags;
-	struct Scsi_Host *host;
-	struct ufs_hba *hba;
-	int index;
-	bool found = false;
-
-	if (!scmd || !scmd->device || !scmd->device->host)
-		return BLK_EH_DONE;
-
-	host = scmd->device->host;
-	hba = shost_priv(host);
-	if (!hba)
-		return BLK_EH_DONE;
-
-	spin_lock_irqsave(host->host_lock, flags);
-
-	for_each_set_bit(index, &hba->outstanding_reqs, hba->nutrs) {
-		if (hba->lrb[index].cmd == scmd) {
-			found = true;
-			break;
-		}
-	}
-
-	spin_unlock_irqrestore(host->host_lock, flags);
-
-	/*
-	 * Bypass SCSI error handling and reset the block layer timer if this
-	 * SCSI command was not actually dispatched to UFS driver, otherwise
-	 * let SCSI layer handle the error as usual.
-	 */
-	return found ? BLK_EH_DONE : BLK_EH_RESET_TIMER;
-}
-
 static const struct attribute_group *ufshcd_driver_groups[] = {
 	&ufs_sysfs_unit_descriptor_group,
 	&ufs_sysfs_lun_attributes_group,
@@ -7145,7 +7110,6 @@ static struct scsi_host_template ufshcd_driver_template = {
 	.eh_abort_handler	= ufshcd_abort,
 	.eh_device_reset_handler = ufshcd_eh_device_reset_handler,
 	.eh_host_reset_handler   = ufshcd_eh_host_reset_handler,
-	.eh_timed_out		= ufshcd_eh_timed_out,
 	.this_id		= -1,
 	.sg_tablesize		= SG_ALL,
 	.cmd_per_lun		= UFSHCD_CMD_PER_LUN,
Reviewed-by on v1 6/6
#21 lore raw

Re: [PATCH 6/6] ufs: Remove the SCSI timeout handler

Quoted 73 lines · Bart Van Assche

The UFS SCSI timeout handler was needed to compensate that ufshcd_queuecommand() could return SCSI_MLQUEUE_HOST_BUSY for a long time. Commit a276c19e3e98 ("scsi: ufs: Avoid busy-waiting by eliminating tag conflicts") fixed this so the timeout handler is no longer necessary.

See also commit f550c65b543b ("scsi: ufs: implement scsi host timeout handler").

Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 36 ------------------------------------
 1 file changed, 36 deletions(-)
drivers/scsi/ufs/ufshcd.c +0 -35
index edcc137c436b..763e41286a4b 100644
@@ -7092,41 +7092,6 @@ static void ufshcd_async_scan(void *data,
 async_cookie_t cookie)
 	ufshcd_probe_hba(hba);
 }
 
-static enum blk_eh_timer_return ufshcd_eh_timed_out(struct scsi_cmnd 
 *scmd)
-{
-	unsigned long flags;
-	struct Scsi_Host *host;
-	struct ufs_hba *hba;
-	int index;
-	bool found = false;
-
-	if (!scmd || !scmd->device || !scmd->device->host)
-		return BLK_EH_DONE;
-
-	host = scmd->device->host;
-	hba = shost_priv(host);
-	if (!hba)
-		return BLK_EH_DONE;
-
-	spin_lock_irqsave(host->host_lock, flags);
-
-	for_each_set_bit(index, &hba->outstanding_reqs, hba->nutrs) {
-		if (hba->lrb[index].cmd == scmd) {
-			found = true;
-			break;
-		}
-	}
-
-	spin_unlock_irqrestore(host->host_lock, flags);
-
-	/*
-	 * Bypass SCSI error handling and reset the block layer timer if this
-	 * SCSI command was not actually dispatched to UFS driver, otherwise
-	 * let SCSI layer handle the error as usual.
-	 */
-	return found ? BLK_EH_DONE : BLK_EH_RESET_TIMER;
-}
-
 static const struct attribute_group *ufshcd_driver_groups[] = {
drivers/scsi/ufs/ufshcd.c +0 -1
 	&ufs_sysfs_unit_descriptor_group,
 	&ufs_sysfs_lun_attributes_group,
@@ -7145,7 +7110,6 @@ static struct scsi_host_template
 ufshcd_driver_template = {
 	.eh_abort_handler	= ufshcd_abort,
 	.eh_device_reset_handler = ufshcd_eh_device_reset_handler,
 	.eh_host_reset_handler   = ufshcd_eh_host_reset_handler,
-	.eh_timed_out		= ufshcd_eh_timed_out,
 	.this_id		= -1,
 	.sg_tablesize		= SG_ALL,
Quoted 1 line

.cmd_per_lun = UFSHCD_CMD_PER_LUN,

Reviewed-by on v1 6/6
#22 lore raw

Re: [PATCH 6/6] ufs: Remove the SCSI timeout handler

In reply to Bart Van Assche
Quoted 8 lines · Bart Van Assche

The UFS SCSI timeout handler was needed to compensate that ufshcd_queuecommand() could return SCSI_MLQUEUE_HOST_BUSY for a long time. Commit a276c19e3e98 ("scsi: ufs: Avoid busy-waiting by eliminating tag conflicts") fixed this so the timeout handler is no longer necessary.

See also commit f550c65b543b ("scsi: ufs: implement scsi host timeout handler").

Quoted 7 lines
Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>

---

Quoted 57 lines
 drivers/scsi/ufs/ufshcd.c | 36 ------------------------------------
 1 file changed, 36 deletions(-)
drivers/scsi/ufs/ufshcd.c +0 -36
index edcc137c436b..763e41286a4b 100644
@@ -7092,41 +7092,6 @@ static void ufshcd_async_scan(void *data, async_cookie_t cookie)
        ufshcd_probe_hba(hba);
 }
 
-static enum blk_eh_timer_return ufshcd_eh_timed_out(struct scsi_cmnd *scmd)
-{
-       unsigned long flags;
-       struct Scsi_Host *host;
-       struct ufs_hba *hba;
-       int index;
-       bool found = false;
-
-       if (!scmd || !scmd->device || !scmd->device->host)
-               return BLK_EH_DONE;
-
-       host = scmd->device->host;
-       hba = shost_priv(host);
-       if (!hba)
-               return BLK_EH_DONE;
-
-       spin_lock_irqsave(host->host_lock, flags);
-
-       for_each_set_bit(index, &hba->outstanding_reqs, hba->nutrs) {
-               if (hba->lrb[index].cmd == scmd) {
-                       found = true;
-                       break;
-               }
-       }
-
-       spin_unlock_irqrestore(host->host_lock, flags);
-
-       /*
-        * Bypass SCSI error handling and reset the block layer timer if this
-        * SCSI command was not actually dispatched to UFS driver, otherwise
-        * let SCSI layer handle the error as usual.
-        */
-       return found ? BLK_EH_DONE : BLK_EH_RESET_TIMER;
-}
-
 static const struct attribute_group *ufshcd_driver_groups[] = {
        &ufs_sysfs_unit_descriptor_group,
        &ufs_sysfs_lun_attributes_group,
@@ -7145,7 +7110,6 @@ static struct scsi_host_template ufshcd_driver_template = {
        .eh_abort_handler       = ufshcd_abort,
        .eh_device_reset_handler = ufshcd_eh_device_reset_handler,
        .eh_host_reset_handler   = ufshcd_eh_host_reset_handler,
-       .eh_timed_out           = ufshcd_eh_timed_out,
        .this_id                = -1,
        .sg_tablesize           = SG_ALL,
        .cmd_per_lun            = UFSHCD_CMD_PER_LUN,
Signature · 2 lines
Regards,
Alim
#23 lore raw

Re: [PATCH 6/6] ufs: Remove the SCSI timeout handler

In reply to Bart Van Assche

Hi Bart,

Quoted 7 lines · Bart Van Assche

The UFS SCSI timeout handler was needed to compensate that ufshcd_queuecommand() could return SCSI_MLQUEUE_HOST_BUSY for a long time. Commit a276c19e3e98 ("scsi: ufs: Avoid busy-waiting by eliminating tag conflicts") fixed this so the timeout handler is no longer necessary.

See also commit f550c65b543b ("scsi: ufs: implement scsi host timeout handler").

Sorry for bugging you on this old change. I am afraid we may need to add this timeout handler back. Because there is till chances that a request gets stuck somewhere in ufshcd_queuecommand() path before ufshcd_send_command() gets called. e.g.

ufshcd_queuecommand()
->ufshcd_map_sg()
-->scsi_dma_map()
--->dma_map_sg()
---->dev->ops->map_sg()

map_sg() ops may get stuck. map_sg() method can vary on different platforms based on actual IOMMU engines. We cannot gaurantee map_sg() ops must return immediately as we don't know what is actually inside map_sg() ops.

And if it gets stuck there for a long time till the request times out, without the UFS timeout handler, scsi layer will try to abort this request from UFS driver by calling ufshcd_abort() eventually. ufshcd_abort() will think this request has been completed due to its tag is not in hba->outstanding_reqs or UFS host's door bell reg. However, actually, this request is still in ufshcd_queuecommand() path. I don't need to continue on the subsequent impact to UFS driver if ufshcd_abort() happens in this case. This is a corner case, but it is still possible (I did see map_sg() ops hangs on real devices).

Having the UFS timeout handler back will prevent this situation as UFS timeout handler checks if the tag is in hba->outstanding_reqs (for our case, it is not in there), if no, it returns BLK_EH_RESET_TIMER so that scsi/block layer will keep waiting.

What do you think? Please let me know your ideas on this, thanks!

<--code snippet-->
static int ufshcd_abort(struct scsi_cmnd *cmd)
{
...

        reg = ufshcd_readl(hba, REG_UTP_TRANSFER_REQ_DOOR_BELL);
        /* If command is already aborted/completed, return SUCCESS */
        if (!(test_bit(tag, &hba->outstanding_reqs))) {
                dev_err(hba->dev,
                        "%s: cmd at tag %d already completed, 
outstanding=0x%lx, doorbell=0x%x\n",
                        __func__, tag, hba->outstanding_reqs, reg);
                goto out;
        }

...
}
<--code snippet-->

Thanks,
Can Guo.

Quoted 67 lines
Cc Bean Huo <[email protected]>
Cc Can Guo <[email protected]>
Cc Avri Altman <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Tomas Winkler <[email protected]>
---
 drivers/scsi/ufs/ufshcd.c | 36 ------------------------------------
 1 file changed, 36 deletions(-)
drivers/scsi/ufs/ufshcd.c +0 -35
index edcc137c436b..763e41286a4b 100644
@@ -7092,41 +7092,6 @@ static void ufshcd_async_scan(void *data,
 async_cookie_t cookie)
 	ufshcd_probe_hba(hba);
 }
 
-static enum blk_eh_timer_return ufshcd_eh_timed_out(struct scsi_cmnd 
 *scmd)
-{
-	unsigned long flags;
-	struct Scsi_Host *host;
-	struct ufs_hba *hba;
-	int index;
-	bool found = false;
-
-	if (!scmd || !scmd->device || !scmd->device->host)
-		return BLK_EH_DONE;
-
-	host = scmd->device->host;
-	hba = shost_priv(host);
-	if (!hba)
-		return BLK_EH_DONE;
-
-	spin_lock_irqsave(host->host_lock, flags);
-
-	for_each_set_bit(index, &hba->outstanding_reqs, hba->nutrs) {
-		if (hba->lrb[index].cmd == scmd) {
-			found = true;
-			break;
-		}
-	}
-
-	spin_unlock_irqrestore(host->host_lock, flags);
-
-	/*
-	 * Bypass SCSI error handling and reset the block layer timer if this
-	 * SCSI command was not actually dispatched to UFS driver, otherwise
-	 * let SCSI layer handle the error as usual.
-	 */
-	return found ? BLK_EH_DONE : BLK_EH_RESET_TIMER;
-}
-
 static const struct attribute_group *ufshcd_driver_groups[] = {
drivers/scsi/ufs/ufshcd.c +0 -1
 	&ufs_sysfs_unit_descriptor_group,
 	&ufs_sysfs_lun_attributes_group,
@@ -7145,7 +7110,6 @@ static struct scsi_host_template
 ufshcd_driver_template = {
 	.eh_abort_handler	= ufshcd_abort,
 	.eh_device_reset_handler = ufshcd_eh_device_reset_handler,
 	.eh_host_reset_handler   = ufshcd_eh_host_reset_handler,
-	.eh_timed_out		= ufshcd_eh_timed_out,
 	.this_id		= -1,
 	.sg_tablesize		= SG_ALL,

.cmd_per_lun = UFSHCD_CMD_PER_LUN,

#24 lore raw

Re: [PATCH 6/6] ufs: Remove the SCSI timeout handler

Quoted 47 lines · Can Guo

Hi Bart,

Quoted 8 lines · Bart Van Assche

The UFS SCSI timeout handler was needed to compensate that ufshcd_queuecommand() could return SCSI_MLQUEUE_HOST_BUSY for a long time. Commit a276c19e3e98 ("scsi: ufs: Avoid busy-waiting by eliminating tag conflicts") fixed this so the timeout handler is no longer necessary.

See also commit f550c65b543b ("scsi: ufs: implement scsi host timeout handler").

Sorry for bugging you on this old change. I am afraid we may need to add this timeout handler back. Because there is till chances that a request gets stuck somewhere in ufshcd_queuecommand() path before ufshcd_send_command() gets called. e.g.

ufshcd_queuecommand()
->ufshcd_map_sg()
-->scsi_dma_map()
--->dma_map_sg()
---->dev->ops->map_sg()

map_sg() ops may get stuck. map_sg() method can vary on different platforms based on actual IOMMU engines. We cannot gaurantee map_sg() ops must return immediately as we don't know what is actually inside map_sg() ops.

And if it gets stuck there for a long time till the request times out, without
the UFS timeout handler, scsi layer will try to abort this request from UFS driver by calling ufshcd_abort() eventually. ufshcd_abort() will think this request has been completed due to its tag is not in hba->outstanding_reqs or UFS host's door bell reg. However, actually, this request is still in ufshcd_queuecommand() path. I don't need to continue on the subsequent impact
to UFS driver if ufshcd_abort() happens in this case. This is a corner case,
but it is still possible (I did see map_sg() ops hangs on real devices).

Having the UFS timeout handler back will prevent this situation as UFS timeout
handler checks if the tag is in hba->outstanding_reqs (for our case, it is not
in there), if no, it returns BLK_EH_RESET_TIMER so that scsi/block layer will
keep waiting.

Quoted 1 line

What do you think? Please let me know your ideas on this, thanks!

Hi Can,

I see the following issues with the above proposal:
- Although I haven't been able to find explicit documentation of this, I think that dma_map_sg() must not sleep. If it would sleep that would break most block and SCSI drivers because many of these drivers call dma_map_sg() from their .queue_rq() or .queuecommand() implementation and if BLK_MQ_F_BLOCKING has not been set these functions must not sleep.
- A timeout handler must not be invoked while .queuecommand() is still in progress. The SCSI core calls blk_mq_start_request() before it calls ufshcd_queuecommand(). The blk_mq_start_request() activates the block layer timeout mechanism. ufshcd_queuecommand() must have finished before the block layer timeout handler is activated.

Please fix the root cause, namely the map_sg implementation that may get stuck.

Thanks,

Bart.

#25 lore raw

Re: [PATCH 6/6] ufs: Remove the SCSI timeout handler

Quoted 48 lines · Bart Van Assche
Quoted 47 lines · Can Guo

Hi Bart,

Quoted 8 lines · Bart Van Assche

The UFS SCSI timeout handler was needed to compensate that ufshcd_queuecommand() could return SCSI_MLQUEUE_HOST_BUSY for a long time. Commit a276c19e3e98 ("scsi: ufs: Avoid busy-waiting by eliminating tag conflicts") fixed this so the timeout handler is no longer necessary.

See also commit f550c65b543b ("scsi: ufs: implement scsi host timeout handler").

Sorry for bugging you on this old change. I am afraid we may need to add this timeout handler back. Because there is till chances that a request gets stuck somewhere in ufshcd_queuecommand() path before ufshcd_send_command() gets called. e.g.

ufshcd_queuecommand()
->ufshcd_map_sg()
-->scsi_dma_map()
--->dma_map_sg()
---->dev->ops->map_sg()

map_sg() ops may get stuck. map_sg() method can vary on different platforms based on actual IOMMU engines. We cannot gaurantee map_sg() ops must return immediately as we don't know what is actually inside map_sg() ops.

And if it gets stuck there for a long time till the request times out, without
the UFS timeout handler, scsi layer will try to abort this request from UFS driver by calling ufshcd_abort() eventually. ufshcd_abort() will think this request has been completed due to its tag is not in hba->outstanding_reqs or UFS host's door bell reg. However, actually, this request is still in ufshcd_queuecommand() path. I don't need to continue on the subsequent impact
to UFS driver if ufshcd_abort() happens in this case. This is a corner case,
but it is still possible (I did see map_sg() ops hangs on real devices).

Having the UFS timeout handler back will prevent this situation as UFS timeout
handler checks if the tag is in hba->outstanding_reqs (for our case, it is not
in there), if no, it returns BLK_EH_RESET_TIMER so that scsi/block layer will
keep waiting.

Quoted 1 line
Quoted 1 line

What do you think? Please let me know your ideas on this, thanks!

Hi Bart,

Quoted 17 lines

Hi Can,

I see the following issues with the above proposal:
- Although I haven't been able to find explicit documentation of this, I
think that dma_map_sg() must not sleep. If it would sleep that would break most block and SCSI drivers because many of these drivers call dma_map_sg() from their .queue_rq() or .queuecommand() implementation and if BLK_MQ_F_BLOCKING has not been set these functions must not sleep.
- A timeout handler must not be invoked while .queuecommand() is still in progress. The SCSI core calls blk_mq_start_request() before it calls ufshcd_queuecommand(). The blk_mq_start_request() activates the block layer timeout mechanism. ufshcd_queuecommand() must have finished before the block layer timeout handler is activated.

Quoted 6 lines

Please fix the root cause, namely the map_sg implementation that may get stuck.

Thanks,

Bart.

queuecommand path should not sleep - that is right, due to queuecommand can be invoked from contexts where preempt is disabled, e.g. softirq.

I don't know why map_sg() ops can take that long, but apparently it does not sleep, otherwise we should have seen schedule while atomic error long time ago.

Quoted 2 lines

ufshcd_queuecommand() must have
finished before the block layer timeout handler is activated.

This is the ideal/expected situation, but we are seeing the corner case.

Fixing the root cause of that is one thing, but having the timeout handler back can prevent UFS driver from messing up the subsequent requests further in such case, causing possible data corruption. Is there any drawbacks if we have it back?

Thanks,
Can Guo.

#26 lore raw

Re: [PATCH 6/6] ufs: Remove the SCSI timeout handler

Quoted 10 lines · Can Guo
Quoted 2 lines · Bart Van Assche

ufshcd_queuecommand() must have
finished before the block layer timeout handler is activated.

This is the ideal/expected situation, but we are seeing the corner case.

Fixing the root cause of that is one thing, but having the timeout handler back can prevent UFS driver from messing up the subsequent requests further in such case, causing possible data corruption. Is there any drawbacks if we have it back?

Hi Can,

My conclusion from your emails is that ufshcd_queuecommand() can spend more time than the SCSI timeout (30 seconds) in dma_map_sg(). A dma_map_sg() call that keeps the CPU busy during more than 30 seconds is not only weird but it is also a disaster from the point of view of energy consumption. Please fix the root cause.

Thanks,

Bart.

#27 lore raw
In reply to Bart Van Assche

Bart,

Quoted 2 lines

This patch series includes patches that clean up the UFS driver code somewhat. Please consider these patches for kernel v5.6.

Applied 1-4,6 to 5.6/scsi-queue. Thank you!

Signature · 1 line
Martin K. Petersen	Oracle Linux Engineering