[PATCH RFC v2] mpt2/mpt3sas lock reduction for scsi-mq

[PATCH RFC v2] mpt2/mpt3sas lock reduction for scsi-mq

[Jens Axboe]

Hi,

This is v2 of the series. Generally addressed Christoph's review comments,
but more specifically, changes since v1:

- Kill the MPI2_FUNCTION_SCSI_TASK_MGMT support, which in turn
  enabled me to drop the SCSI/blk iter patches and the smid work-around
  for scsi-mq in mpt2/3 for bumping to hpr for NULL command.
- Made init/exit_command() work (sort of) on non scsi-mq. Drivers
  will have to set request_index related information separately
  still for non scsi-mq, since we don't have a reliable one
- Use scsi_cmd_priv() instead of adding a new helper.
- Make scsi_find_tag() only return in-progress IO for scsi-mq. This is
  simiar to the ->special check for non-mq.
- Ensure that mpt3sas is in sync with the mpt2sas changes.

 drivers/scsi/mpt2sas/mpt2sas_base.c  |  187 ++++++++++++++++++++++++-----------
 drivers/scsi/mpt2sas/mpt2sas_base.h  |    3 
 drivers/scsi/mpt2sas/mpt2sas_ctl.c   |  126 +----------------------
 drivers/scsi/mpt2sas/mpt2sas_scsih.c |   87 ++++++++++++----
 drivers/scsi/mpt3sas/mpt3sas_base.c  |  173 +++++++++++++++++++++++---------
 drivers/scsi/mpt3sas/mpt3sas_base.h  |    2 
 drivers/scsi/mpt3sas/mpt3sas_ctl.c   |  130 +-----------------------
 drivers/scsi/mpt3sas/mpt3sas_scsih.c |   69 ++++++++++--
 drivers/scsi/scsi.c                  |    9 +
 drivers/scsi/scsi_lib.c              |   16 ++
 include/scsi/scsi_host.h             |    6 -
 include/scsi/scsi_tcq.h              |   11 +-
 12 files changed, 448 insertions(+), 371 deletions(-)

[PATCH RFC v2] mpt2/mpt3sas lock reduction for scsi-mq

[Jens Axboe]

Hi,

This is v2 of the series. Generally addressed Christoph's review comments,
but more specifically, changes since v1:

- Kill the MPI2_FUNCTION_SCSI_TASK_MGMT support, which in turn
  enabled me to drop the SCSI/blk iter patches and the smid work-around
  for scsi-mq in mpt2/3 for bumping to hpr for NULL command.
- Made init/exit_command() work (sort of) on non scsi-mq. Drivers
  will have to set request_index related information separately
  still for non scsi-mq, since we don't have a reliable one
- Use scsi_cmd_priv() instead of adding a new helper.
- Make scsi_find_tag() only return in-progress IO for scsi-mq. This is
  simiar to the ->special check for non-mq.
- Ensure that mpt3sas is in sync with the mpt2sas changes.

 drivers/scsi/mpt2sas/mpt2sas_base.c  |  187 ++++++++++++++++++++++++-----------
 drivers/scsi/mpt2sas/mpt2sas_base.h  |    3 
 drivers/scsi/mpt2sas/mpt2sas_ctl.c   |  126 +----------------------
 drivers/scsi/mpt2sas/mpt2sas_scsih.c |   87 ++++++++++++----
 drivers/scsi/mpt3sas/mpt3sas_base.c  |  173 +++++++++++++++++++++++---------
 drivers/scsi/mpt3sas/mpt3sas_base.h  |    2 
 drivers/scsi/mpt3sas/mpt3sas_ctl.c   |  130 +-----------------------
 drivers/scsi/mpt3sas/mpt3sas_scsih.c |   69 ++++++++++--
 drivers/scsi/scsi.c                  |    9 +
 drivers/scsi/scsi_lib.c              |   16 ++
 include/scsi/scsi_host.h             |    6 -
 include/scsi/scsi_tcq.h              |   11 +-
 12 files changed, 448 insertions(+), 371 deletions(-)

[PATCH 1/6] scsi: add host template init/exit_command hooks

[Jens Axboe]

If a LLD has hardware commands in the request pdu, then we need some
helper hooks to help the driver initialize state at load time, and
potentially to tear down state at rmmod time. Add a host template
->init_command() and ->exit_command() hook to help with that.

Note that for the non scsi-mq case, we can't pass in a reliable
request index, since we don't have one...

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 drivers/scsi/scsi.c      |  9 +++++++++
 drivers/scsi/scsi_lib.c  | 16 ++++++++++++++++
 include/scsi/scsi_host.h |  6 +++++-
 3 files changed, 30 insertions(+), 1 deletion(-)

diff --git a/drivers/scsi/scsi.c b/drivers/scsi/scsi.c
index c9c3b579eece..53e935c52300 100644
--- a/drivers/scsi/scsi.c
+++ b/drivers/scsi/scsi.c
@@ -182,6 +182,8 @@ scsi_host_free_command(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
 {
 	struct scsi_host_cmd_pool *pool = shost->cmd_pool;
 
+	if (shost->hostt->exit_command)
+		shost->hostt->exit_command(shost, cmd);
 	if (cmd->prot_sdb)
 		kmem_cache_free(scsi_sdb_cache, cmd->prot_sdb);
 	kmem_cache_free(pool->sense_slab, cmd->sense_buffer);
@@ -217,8 +219,15 @@ scsi_host_alloc_command(struct Scsi_Host *shost, gfp_t gfp_mask)
 			goto fail_free_sense;
 	}
 
+	if (shost->hostt->init_command) {
+		if (shost->hostt->init_command(shost, cmd, 0))
+			goto fail_init;
+	}
+
 	return cmd;
 
+fail_init:
+	kmem_cache_free(scsi_sdb_cache, cmd->prot_sdb);
 fail_free_sense:
 	kmem_cache_free(pool->sense_slab, cmd->sense_buffer);
 fail_free_cmd:
diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index 54d7a6cbb98a..df46946aa708 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -2056,11 +2056,23 @@ static int scsi_init_request(void *data, struct request *rq,
 		unsigned int numa_node)
 {
 	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
+	struct Scsi_Host *shost = data;
 
 	cmd->sense_buffer = kzalloc_node(SCSI_SENSE_BUFFERSIZE, GFP_KERNEL,
 			numa_node);
 	if (!cmd->sense_buffer)
 		return -ENOMEM;
+
+	if (shost->hostt->init_command) {
+		int ret;
+
+		ret = shost->hostt->init_command(shost, cmd, request_idx);
+		if (ret) {
+			kfree(cmd->sense_buffer);
+			return ret;
+		}
+	}
+
 	return 0;
 }
 
@@ -2068,8 +2080,12 @@ static void scsi_exit_request(void *data, struct request *rq,
 		unsigned int hctx_idx, unsigned int request_idx)
 {
 	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
+	struct Scsi_Host *shost = data;
 
 	kfree(cmd->sense_buffer);
+
+	if (shost->hostt->exit_command)
+		shost->hostt->exit_command(shost, cmd);
 }
 
 static u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
diff --git a/include/scsi/scsi_host.h b/include/scsi/scsi_host.h
index e113c757d555..baaa7a2fc07d 100644
--- a/include/scsi/scsi_host.h
+++ b/include/scsi/scsi_host.h
@@ -496,10 +496,14 @@ struct scsi_host_template {
 	u64 vendor_id;
 
 	/*
-	 * Additional per-command data allocated for the driver.
+	 * Additional per-command data allocated for the driver, along
+	 * with init/exit helper hooks.
 	 */
 	unsigned int cmd_size;
 	struct scsi_host_cmd_pool *cmd_pool;
+	int (*init_command)(struct Scsi_Host *, struct scsi_cmnd *,
+				unsigned int);
+	void (*exit_command)(struct Scsi_Host *, struct scsi_cmnd *);
 
 	/* temporary flag to disable blk-mq I/O path */
 	bool disable_blk_mq;
-- 
1.9.1

[PATCH 1/6] scsi: add host template init/exit_command hooks

[Jens Axboe]

If a LLD has hardware commands in the request pdu, then we need some
helper hooks to help the driver initialize state at load time, and
potentially to tear down state at rmmod time. Add a host template
->init_command() and ->exit_command() hook to help with that.

Note that for the non scsi-mq case, we can't pass in a reliable
request index, since we don't have one...

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 drivers/scsi/scsi.c      |  9 +++++++++
 drivers/scsi/scsi_lib.c  | 16 ++++++++++++++++
 include/scsi/scsi_host.h |  6 +++++-
 3 files changed, 30 insertions(+), 1 deletion(-)

diff --git a/drivers/scsi/scsi.c b/drivers/scsi/scsi.c
index c9c3b579eece..53e935c52300 100644
--- a/drivers/scsi/scsi.c
+++ b/drivers/scsi/scsi.c
@@ -182,6 +182,8 @@ scsi_host_free_command(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
 {
 	struct scsi_host_cmd_pool *pool = shost->cmd_pool;
 
+	if (shost->hostt->exit_command)
+		shost->hostt->exit_command(shost, cmd);
 	if (cmd->prot_sdb)
 		kmem_cache_free(scsi_sdb_cache, cmd->prot_sdb);
 	kmem_cache_free(pool->sense_slab, cmd->sense_buffer);
@@ -217,8 +219,15 @@ scsi_host_alloc_command(struct Scsi_Host *shost, gfp_t gfp_mask)
 			goto fail_free_sense;
 	}
 
+	if (shost->hostt->init_command) {
+		if (shost->hostt->init_command(shost, cmd, 0))
+			goto fail_init;
+	}
+
 	return cmd;
 
+fail_init:
+	kmem_cache_free(scsi_sdb_cache, cmd->prot_sdb);
 fail_free_sense:
 	kmem_cache_free(pool->sense_slab, cmd->sense_buffer);
 fail_free_cmd:
diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index 54d7a6cbb98a..df46946aa708 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -2056,11 +2056,23 @@ static int scsi_init_request(void *data, struct request *rq,
 		unsigned int numa_node)
 {
 	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
+	struct Scsi_Host *shost = data;
 
 	cmd->sense_buffer = kzalloc_node(SCSI_SENSE_BUFFERSIZE, GFP_KERNEL,
 			numa_node);
 	if (!cmd->sense_buffer)
 		return -ENOMEM;
+
+	if (shost->hostt->init_command) {
+		int ret;
+
+		ret = shost->hostt->init_command(shost, cmd, request_idx);
+		if (ret) {
+			kfree(cmd->sense_buffer);
+			return ret;
+		}
+	}
+
 	return 0;
 }
 
@@ -2068,8 +2080,12 @@ static void scsi_exit_request(void *data, struct request *rq,
 		unsigned int hctx_idx, unsigned int request_idx)
 {
 	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
+	struct Scsi_Host *shost = data;
 
 	kfree(cmd->sense_buffer);
+
+	if (shost->hostt->exit_command)
+		shost->hostt->exit_command(shost, cmd);
 }
 
 static u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
diff --git a/include/scsi/scsi_host.h b/include/scsi/scsi_host.h
index e113c757d555..baaa7a2fc07d 100644
--- a/include/scsi/scsi_host.h
+++ b/include/scsi/scsi_host.h
@@ -496,10 +496,14 @@ struct scsi_host_template {
 	u64 vendor_id;
 
 	/*
-	 * Additional per-command data allocated for the driver.
+	 * Additional per-command data allocated for the driver, along
+	 * with init/exit helper hooks.
 	 */
 	unsigned int cmd_size;
 	struct scsi_host_cmd_pool *cmd_pool;
+	int (*init_command)(struct Scsi_Host *, struct scsi_cmnd *,
+				unsigned int);
+	void (*exit_command)(struct Scsi_Host *, struct scsi_cmnd *);
 
 	/* temporary flag to disable blk-mq I/O path */
 	bool disable_blk_mq;
-- 
1.9.1

[PATCH 2/6] scsi: ensure that a request is valid when returning it from scsi_find_tag()

[Jens Axboe]

For the scsi-mq case, ensure that the request has been started before
returning it from scsi_find_tag(). This is akin to the ->special
check for the non-mq case, it ensures that the request has been setup
and issued.

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 include/scsi/scsi_tcq.h | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

diff --git a/include/scsi/scsi_tcq.h b/include/scsi/scsi_tcq.h
index b27977e8aaed..8634cab61aeb 100644
--- a/include/scsi/scsi_tcq.h
+++ b/include/scsi/scsi_tcq.h
@@ -35,8 +35,15 @@ static inline struct scsi_cmnd *scsi_find_tag(struct scsi_device *sdev, int tag)
         struct request *req;
 
         if (tag != SCSI_NO_TAG) {
-		if (shost_use_blk_mq(sdev->host))
-			return scsi_mq_find_tag(sdev->host, tag);
+		if (shost_use_blk_mq(sdev->host)) {
+			struct scsi_cmnd *scmd;
+
+			scmd = scsi_mq_find_tag(sdev->host, tag);
+			if (blk_mq_request_started(scmd->request))
+				return scmd;
+
+			return NULL;
+		}
 
 		req = blk_queue_find_tag(sdev->request_queue, tag);
 	        return req ? (struct scsi_cmnd *)req->special : NULL;
-- 
1.9.1

[PATCH 2/6] scsi: ensure that a request is valid when returning it from scsi_find_tag()

[Jens Axboe]

For the scsi-mq case, ensure that the request has been started before
returning it from scsi_find_tag(). This is akin to the ->special
check for the non-mq case, it ensures that the request has been setup
and issued.

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 include/scsi/scsi_tcq.h | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

diff --git a/include/scsi/scsi_tcq.h b/include/scsi/scsi_tcq.h
index b27977e8aaed..8634cab61aeb 100644
--- a/include/scsi/scsi_tcq.h
+++ b/include/scsi/scsi_tcq.h
@@ -35,8 +35,15 @@ static inline struct scsi_cmnd *scsi_find_tag(struct scsi_device *sdev, int tag)
         struct request *req;
 
         if (tag != SCSI_NO_TAG) {
-		if (shost_use_blk_mq(sdev->host))
-			return scsi_mq_find_tag(sdev->host, tag);
+		if (shost_use_blk_mq(sdev->host)) {
+			struct scsi_cmnd *scmd;
+
+			scmd = scsi_mq_find_tag(sdev->host, tag);
+			if (blk_mq_request_started(scmd->request))
+				return scmd;
+
+			return NULL;
+		}
 
 		req = blk_queue_find_tag(sdev->request_queue, tag);
 	        return req ? (struct scsi_cmnd *)req->special : NULL;
-- 
1.9.1

Re: [PATCH 2/6] scsi: ensure that a request is valid when returning it from scsi_find_tag()

[Christoph Hellwig]

On Tue, Apr 07, 2015 at 11:18:36AM -0600, Jens Axboe wrote:
> For the scsi-mq case, ensure that the request has been started before
> returning it from scsi_find_tag(). This is akin to the ->special
> check for the non-mq case, it ensures that the request has been setup
> and issued.
> 
> Signed-off-by: Jens Axboe <axboe@fb.com>

Shouldn't this be done in blk_mq_tag_to_rq?  In general when a driver
looks for an active tag it wants an active tag, so working around
this in the driver seems like a bad idea.

[PATCH 3/6] mpt2sas: get rid of MPI2_FUNCTION_SCSI_TASK_MGMT

[Jens Axboe]

This involves a lot of nasty lookup of a specific running command,
across the host. If anybody is using this code, it should be moved
to proper error handling instead.

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 drivers/scsi/mpt2sas/mpt2sas_ctl.c | 126 +++----------------------------------
 1 file changed, 10 insertions(+), 116 deletions(-)

diff --git a/drivers/scsi/mpt2sas/mpt2sas_ctl.c b/drivers/scsi/mpt2sas/mpt2sas_ctl.c
index 4e509604b571..dd777e53298d 100644
--- a/drivers/scsi/mpt2sas/mpt2sas_ctl.c
+++ b/drivers/scsi/mpt2sas/mpt2sas_ctl.c
@@ -530,84 +530,6 @@ _ctl_poll(struct file *filep, poll_table *wait)
 }
 
 /**
- * _ctl_set_task_mid - assign an active smid to tm request
- * @ioc: per adapter object
- * @karg - (struct mpt2_ioctl_command)
- * @tm_request - pointer to mf from user space
- *
- * Returns 0 when an smid if found, else fail.
- * during failure, the reply frame is filled.
- */
-static int
-_ctl_set_task_mid(struct MPT2SAS_ADAPTER *ioc, struct mpt2_ioctl_command *karg,
-    Mpi2SCSITaskManagementRequest_t *tm_request)
-{
-	u8 found = 0;
-	u16 i;
-	u16 handle;
-	struct scsi_cmnd *scmd;
-	struct MPT2SAS_DEVICE *priv_data;
-	unsigned long flags;
-	Mpi2SCSITaskManagementReply_t *tm_reply;
-	u32 sz;
-	u32 lun;
-	char *desc = NULL;
-
-	if (tm_request->TaskType == MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK)
-		desc = "abort_task";
-	else if (tm_request->TaskType == MPI2_SCSITASKMGMT_TASKTYPE_QUERY_TASK)
-		desc = "query_task";
-	else
-		return 0;
-
-	lun = scsilun_to_int((struct scsi_lun *)tm_request->LUN);
-
-	handle = le16_to_cpu(tm_request->DevHandle);
-	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
-	for (i = ioc->scsiio_depth; i && !found; i--) {
-		scmd = ioc->scsi_lookup[i - 1].scmd;
-		if (scmd == NULL || scmd->device == NULL ||
-		    scmd->device->hostdata == NULL)
-			continue;
-		if (lun != scmd->device->lun)
-			continue;
-		priv_data = scmd->device->hostdata;
-		if (priv_data->sas_target == NULL)
-			continue;
-		if (priv_data->sas_target->handle != handle)
-			continue;
-		tm_request->TaskMID = cpu_to_le16(ioc->scsi_lookup[i - 1].smid);
-		found = 1;
-	}
-	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
-
-	if (!found) {
-		dctlprintk(ioc, printk(MPT2SAS_INFO_FMT "%s: "
-		    "handle(0x%04x), lun(%d), no active mid!!\n", ioc->name,
-		    desc, le16_to_cpu(tm_request->DevHandle), lun));
-		tm_reply = ioc->ctl_cmds.reply;
-		tm_reply->DevHandle = tm_request->DevHandle;
-		tm_reply->Function = MPI2_FUNCTION_SCSI_TASK_MGMT;
-		tm_reply->TaskType = tm_request->TaskType;
-		tm_reply->MsgLength = sizeof(Mpi2SCSITaskManagementReply_t)/4;
-		tm_reply->VP_ID = tm_request->VP_ID;
-		tm_reply->VF_ID = tm_request->VF_ID;
-		sz = min_t(u32, karg->max_reply_bytes, ioc->reply_sz);
-		if (copy_to_user(karg->reply_frame_buf_ptr, ioc->ctl_cmds.reply,
-		    sz))
-			printk(KERN_ERR "failure at %s:%d/%s()!\n", __FILE__,
-			    __LINE__, __func__);
-		return 1;
-	}
-
-	dctlprintk(ioc, printk(MPT2SAS_INFO_FMT "%s: "
-	    "handle(0x%04x), lun(%d), task_mid(%d)\n", ioc->name,
-	    desc, le16_to_cpu(tm_request->DevHandle), lun,
-	     le16_to_cpu(tm_request->TaskMID)));
-	return 0;
-}
-
-/**
  * _ctl_do_mpt_command - main handler for MPT2COMMAND opcode
  * @ioc: per adapter object
  * @karg - (struct mpt2_ioctl_command)
@@ -689,22 +611,18 @@ _ctl_do_mpt_command(struct MPT2SAS_ADAPTER *ioc, struct mpt2_ioctl_command karg,
 	}
 
 	if (mpi_request->Function == MPI2_FUNCTION_SCSI_TASK_MGMT) {
-		smid = mpt2sas_base_get_smid_hpr(ioc, ioc->ctl_cb_idx);
-		if (!smid) {
-			printk(MPT2SAS_ERR_FMT "%s: failed obtaining a smid\n",
-			    ioc->name, __func__);
-			ret = -EAGAIN;
-			goto out;
-		}
-	} else {
+		printk(MPT2SAS_ERR_FMT "%s: ignoring "
+			"MPI2_FUNCTION_SCSI_TASK_MGMT\n", ioc->name, __func__);
+		ret = -EINVAL;
+		goto out;
+	}
 
-		smid = mpt2sas_base_get_smid_scsiio(ioc, ioc->ctl_cb_idx, NULL);
-		if (!smid) {
-			printk(MPT2SAS_ERR_FMT "%s: failed obtaining a smid\n",
+	smid = mpt2sas_base_get_smid_hpr(ioc, ioc->ctl_cb_idx);
+	if (!smid) {
+		printk(MPT2SAS_ERR_FMT "%s: failed obtaining a smid\n",
 			    ioc->name, __func__);
-			ret = -EAGAIN;
-			goto out;
-		}
+		ret = -EAGAIN;
+		goto out;
 	}
 
 	ret = 0;
@@ -822,30 +740,6 @@ _ctl_do_mpt_command(struct MPT2SAS_ADAPTER *ioc, struct mpt2_ioctl_command karg,
 			mpt2sas_base_put_smid_default(ioc, smid);
 		break;
 	}
-	case MPI2_FUNCTION_SCSI_TASK_MGMT:
-	{
-		Mpi2SCSITaskManagementRequest_t *tm_request =
-		    (Mpi2SCSITaskManagementRequest_t *)request;
-
-		dtmprintk(ioc, printk(MPT2SAS_INFO_FMT "TASK_MGMT: "
-		    "handle(0x%04x), task_type(0x%02x)\n", ioc->name,
-		    le16_to_cpu(tm_request->DevHandle), tm_request->TaskType));
-
-		if (tm_request->TaskType ==
-		    MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK ||
-		    tm_request->TaskType ==
-		    MPI2_SCSITASKMGMT_TASKTYPE_QUERY_TASK) {
-			if (_ctl_set_task_mid(ioc, &karg, tm_request)) {
-				mpt2sas_base_free_smid(ioc, smid);
-				goto out;
-			}
-		}
-
-		mpt2sas_scsih_set_tm_flag(ioc, le16_to_cpu(
-		    tm_request->DevHandle));
-		mpt2sas_base_put_smid_hi_priority(ioc, smid);
-		break;
-	}
 	case MPI2_FUNCTION_SMP_PASSTHROUGH:
 	{
 		Mpi2SmpPassthroughRequest_t *smp_request =
-- 
1.9.1

[PATCH 3/6] mpt2sas: get rid of MPI2_FUNCTION_SCSI_TASK_MGMT

[Jens Axboe]

This involves a lot of nasty lookup of a specific running command,
across the host. If anybody is using this code, it should be moved
to proper error handling instead.

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 drivers/scsi/mpt2sas/mpt2sas_ctl.c | 126 +++----------------------------------
 1 file changed, 10 insertions(+), 116 deletions(-)

diff --git a/drivers/scsi/mpt2sas/mpt2sas_ctl.c b/drivers/scsi/mpt2sas/mpt2sas_ctl.c
index 4e509604b571..dd777e53298d 100644
--- a/drivers/scsi/mpt2sas/mpt2sas_ctl.c
+++ b/drivers/scsi/mpt2sas/mpt2sas_ctl.c
@@ -530,84 +530,6 @@ _ctl_poll(struct file *filep, poll_table *wait)
 }
 
 /**
- * _ctl_set_task_mid - assign an active smid to tm request
- * @ioc: per adapter object
- * @karg - (struct mpt2_ioctl_command)
- * @tm_request - pointer to mf from user space
- *
- * Returns 0 when an smid if found, else fail.
- * during failure, the reply frame is filled.
- */
-static int
-_ctl_set_task_mid(struct MPT2SAS_ADAPTER *ioc, struct mpt2_ioctl_command *karg,
-    Mpi2SCSITaskManagementRequest_t *tm_request)
-{
-	u8 found = 0;
-	u16 i;
-	u16 handle;
-	struct scsi_cmnd *scmd;
-	struct MPT2SAS_DEVICE *priv_data;
-	unsigned long flags;
-	Mpi2SCSITaskManagementReply_t *tm_reply;
-	u32 sz;
-	u32 lun;
-	char *desc = NULL;
-
-	if (tm_request->TaskType == MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK)
-		desc = "abort_task";
-	else if (tm_request->TaskType == MPI2_SCSITASKMGMT_TASKTYPE_QUERY_TASK)
-		desc = "query_task";
-	else
-		return 0;
-
-	lun = scsilun_to_int((struct scsi_lun *)tm_request->LUN);
-
-	handle = le16_to_cpu(tm_request->DevHandle);
-	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
-	for (i = ioc->scsiio_depth; i && !found; i--) {
-		scmd = ioc->scsi_lookup[i - 1].scmd;
-		if (scmd == NULL || scmd->device == NULL ||
-		    scmd->device->hostdata == NULL)
-			continue;
-		if (lun != scmd->device->lun)
-			continue;
-		priv_data = scmd->device->hostdata;
-		if (priv_data->sas_target == NULL)
-			continue;
-		if (priv_data->sas_target->handle != handle)
-			continue;
-		tm_request->TaskMID = cpu_to_le16(ioc->scsi_lookup[i - 1].smid);
-		found = 1;
-	}
-	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
-
-	if (!found) {
-		dctlprintk(ioc, printk(MPT2SAS_INFO_FMT "%s: "
-		    "handle(0x%04x), lun(%d), no active mid!!\n", ioc->name,
-		    desc, le16_to_cpu(tm_request->DevHandle), lun));
-		tm_reply = ioc->ctl_cmds.reply;
-		tm_reply->DevHandle = tm_request->DevHandle;
-		tm_reply->Function = MPI2_FUNCTION_SCSI_TASK_MGMT;
-		tm_reply->TaskType = tm_request->TaskType;
-		tm_reply->MsgLength = sizeof(Mpi2SCSITaskManagementReply_t)/4;
-		tm_reply->VP_ID = tm_request->VP_ID;
-		tm_reply->VF_ID = tm_request->VF_ID;
-		sz = min_t(u32, karg->max_reply_bytes, ioc->reply_sz);
-		if (copy_to_user(karg->reply_frame_buf_ptr, ioc->ctl_cmds.reply,
-		    sz))
-			printk(KERN_ERR "failure at %s:%d/%s()!\n", __FILE__,
-			    __LINE__, __func__);
-		return 1;
-	}
-
-	dctlprintk(ioc, printk(MPT2SAS_INFO_FMT "%s: "
-	    "handle(0x%04x), lun(%d), task_mid(%d)\n", ioc->name,
-	    desc, le16_to_cpu(tm_request->DevHandle), lun,
-	     le16_to_cpu(tm_request->TaskMID)));
-	return 0;
-}
-
-/**
  * _ctl_do_mpt_command - main handler for MPT2COMMAND opcode
  * @ioc: per adapter object
  * @karg - (struct mpt2_ioctl_command)
@@ -689,22 +611,18 @@ _ctl_do_mpt_command(struct MPT2SAS_ADAPTER *ioc, struct mpt2_ioctl_command karg,
 	}
 
 	if (mpi_request->Function == MPI2_FUNCTION_SCSI_TASK_MGMT) {
-		smid = mpt2sas_base_get_smid_hpr(ioc, ioc->ctl_cb_idx);
-		if (!smid) {
-			printk(MPT2SAS_ERR_FMT "%s: failed obtaining a smid\n",
-			    ioc->name, __func__);
-			ret = -EAGAIN;
-			goto out;
-		}
-	} else {
+		printk(MPT2SAS_ERR_FMT "%s: ignoring "
+			"MPI2_FUNCTION_SCSI_TASK_MGMT\n", ioc->name, __func__);
+		ret = -EINVAL;
+		goto out;
+	}
 
-		smid = mpt2sas_base_get_smid_scsiio(ioc, ioc->ctl_cb_idx, NULL);
-		if (!smid) {
-			printk(MPT2SAS_ERR_FMT "%s: failed obtaining a smid\n",
+	smid = mpt2sas_base_get_smid_hpr(ioc, ioc->ctl_cb_idx);
+	if (!smid) {
+		printk(MPT2SAS_ERR_FMT "%s: failed obtaining a smid\n",
 			    ioc->name, __func__);
-			ret = -EAGAIN;
-			goto out;
-		}
+		ret = -EAGAIN;
+		goto out;
 	}
 
 	ret = 0;
@@ -822,30 +740,6 @@ _ctl_do_mpt_command(struct MPT2SAS_ADAPTER *ioc, struct mpt2_ioctl_command karg,
 			mpt2sas_base_put_smid_default(ioc, smid);
 		break;
 	}
-	case MPI2_FUNCTION_SCSI_TASK_MGMT:
-	{
-		Mpi2SCSITaskManagementRequest_t *tm_request =
-		    (Mpi2SCSITaskManagementRequest_t *)request;
-
-		dtmprintk(ioc, printk(MPT2SAS_INFO_FMT "TASK_MGMT: "
-		    "handle(0x%04x), task_type(0x%02x)\n", ioc->name,
-		    le16_to_cpu(tm_request->DevHandle), tm_request->TaskType));
-
-		if (tm_request->TaskType ==
-		    MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK ||
-		    tm_request->TaskType ==
-		    MPI2_SCSITASKMGMT_TASKTYPE_QUERY_TASK) {
-			if (_ctl_set_task_mid(ioc, &karg, tm_request)) {
-				mpt2sas_base_free_smid(ioc, smid);
-				goto out;
-			}
-		}
-
-		mpt2sas_scsih_set_tm_flag(ioc, le16_to_cpu(
-		    tm_request->DevHandle));
-		mpt2sas_base_put_smid_hi_priority(ioc, smid);
-		break;
-	}
 	case MPI2_FUNCTION_SMP_PASSTHROUGH:
 	{
 		Mpi2SmpPassthroughRequest_t *smp_request =
-- 
1.9.1

Re: [PATCH 3/6] mpt2sas: get rid of MPI2_FUNCTION_SCSI_TASK_MGMT

[Christoph Hellwig]

This should also mention that you now allocate the hdr for the normal
pass through from a different pool.  Actually that change probably should
really be a separate patch..

[PATCH 4/6] mpt2sas: store scsi io tracker data in the scsi request

[Jens Axboe]

Instead of storing the IO tracker structure in a separate list
that we need to pop/push to on every submit and complete (and
lock), store it in the pdu associated with a request. This is
possible on scsi-mq only, and further cuts the spinlock associated
time for higher IOPS IO workloads. At 100K IOPS, this effectively
cuts the locking time in half.

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 drivers/scsi/mpt2sas/mpt2sas_base.c  | 187 ++++++++++++++++++++++++-----------
 drivers/scsi/mpt2sas/mpt2sas_base.h  |   3 +
 drivers/scsi/mpt2sas/mpt2sas_scsih.c |  87 ++++++++++++----
 3 files changed, 204 insertions(+), 73 deletions(-)

diff --git a/drivers/scsi/mpt2sas/mpt2sas_base.c b/drivers/scsi/mpt2sas/mpt2sas_base.c
index 11248de92b3b..5adb8469ede4 100644
--- a/drivers/scsi/mpt2sas/mpt2sas_base.c
+++ b/drivers/scsi/mpt2sas/mpt2sas_base.c
@@ -828,6 +828,20 @@ _base_async_event(struct MPT2SAS_ADAPTER *ioc, u8 msix_index, u32 reply)
 	return;
 }
 
+struct scsiio_tracker *
+mpt2sas_get_st_from_smid(struct MPT2SAS_ADAPTER *ioc, u16 smid)
+{
+	if (shost_use_blk_mq(ioc->shost)) {
+		struct scsi_cmnd *scmd;
+
+		scmd = scsi_mq_find_tag(ioc->shost, smid - 1);
+		if (!scmd)
+			return NULL;
+		return scsi_cmd_priv(scmd);
+	} else
+		return &ioc->scsi_lookup[smid - 1];
+}
+
 /**
  * _base_get_cb_idx - obtain the callback index
  * @ioc: per adapter object
@@ -842,8 +856,10 @@ _base_get_cb_idx(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 	u8 cb_idx;
 
 	if (smid < ioc->hi_priority_smid) {
-		i = smid - 1;
-		cb_idx = ioc->scsi_lookup[i].cb_idx;
+		struct scsiio_tracker *st;
+
+		st = mpt2sas_get_st_from_smid(ioc, smid);
+		cb_idx = st->cb_idx;
 	} else if (smid < ioc->internal_smid) {
 		i = smid - ioc->hi_priority_smid;
 		cb_idx = ioc->hpr_lookup[i].cb_idx;
@@ -962,18 +978,17 @@ _base_interrupt(int irq, void *bus_id)
 			goto next;
 		if (smid) {
 			cb_idx = _base_get_cb_idx(ioc, smid);
-		if ((likely(cb_idx < MPT_MAX_CALLBACKS))
+			if ((likely(cb_idx < MPT_MAX_CALLBACKS))
 			    && (likely(mpt_callbacks[cb_idx] != NULL))) {
 				rc = mpt_callbacks[cb_idx](ioc, smid,
 				    msix_index, reply);
-			if (reply)
-				_base_display_reply_info(ioc, smid,
-				    msix_index, reply);
-			if (rc)
-				mpt2sas_base_free_smid(ioc, smid);
+				if (reply)
+					_base_display_reply_info(ioc, smid,
+							msix_index, reply);
+				if (rc)
+					mpt2sas_base_free_smid(ioc, smid);
 			}
-		}
-		if (!smid)
+		} else
 			_base_async_event(ioc, msix_index, reply);
 
 		/* reply free queue handling */
@@ -1724,6 +1739,11 @@ mpt2sas_base_get_smid_scsiio(struct MPT2SAS_ADAPTER *ioc, u8 cb_idx,
 	struct scsiio_tracker *request;
 	u16 smid;
 
+	if (shost_use_blk_mq(ioc->shost)) {
+		request = scsi_cmd_priv(scmd);
+		return request->smid;
+	}
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	if (list_empty(&ioc->free_list)) {
 		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
@@ -1771,6 +1791,31 @@ mpt2sas_base_get_smid_hpr(struct MPT2SAS_ADAPTER *ioc, u8 cb_idx)
 	return smid;
 }
 
+static void
+_base_recovery_check(struct MPT2SAS_ADAPTER *ioc)
+{
+	/*
+	 * See _wait_for_commands_to_complete() call with regards to this code.
+	 */
+	if (ioc->shost_recovery && ioc->pending_io_count) {
+		if (ioc->pending_io_count == 1)
+			wake_up(&ioc->reset_wq);
+		ioc->pending_io_count = 0;
+	}
+}
+
+static void
+_dechain_st(struct MPT2SAS_ADAPTER *ioc, struct scsiio_tracker *st)
+{
+	struct chain_tracker *chain_req;
+
+	while (!list_empty(&st->chain_list)) {
+		chain_req = list_first_entry(&st->chain_list,
+						struct chain_tracker,
+						tracker_list);
+		list_move(&chain_req->tracker_list, &ioc->free_chain_list);
+	}
+}
 
 /**
  * mpt2sas_base_free_smid - put smid back on free_list
@@ -1784,20 +1829,32 @@ mpt2sas_base_free_smid(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 {
 	unsigned long flags;
 	int i;
-	struct chain_tracker *chain_req, *next;
+
+	if (shost_use_blk_mq(ioc->shost) && smid < ioc->hi_priority_smid) {
+		struct scsiio_tracker *st;
+
+		st = mpt2sas_get_st_from_smid(ioc, smid);
+		if (!st)
+			return;
+
+		st->direct_io = 0;
+
+		if (!list_empty(&st->chain_list)) {
+			spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
+			_dechain_st(ioc, st);
+			spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
+		}
+
+		_base_recovery_check(ioc);
+		return;
+	}
 
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	if (smid < ioc->hi_priority_smid) {
 		/* scsiio queue */
 		i = smid - 1;
-		if (!list_empty(&ioc->scsi_lookup[i].chain_list)) {
-			list_for_each_entry_safe(chain_req, next,
-			    &ioc->scsi_lookup[i].chain_list, tracker_list) {
-				list_del_init(&chain_req->tracker_list);
-				list_add(&chain_req->tracker_list,
-				    &ioc->free_chain_list);
-			}
-		}
+		if (!list_empty(&ioc->scsi_lookup[i].chain_list))
+			_dechain_st(ioc, &ioc->scsi_lookup[i]);
 		ioc->scsi_lookup[i].cb_idx = 0xFF;
 		ioc->scsi_lookup[i].scmd = NULL;
 		ioc->scsi_lookup[i].direct_io = 0;
@@ -1805,15 +1862,7 @@ mpt2sas_base_free_smid(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 		    &ioc->free_list);
 		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 
-		/*
-		 * See _wait_for_commands_to_complete() call with regards
-		 * to this code.
-		 */
-		if (ioc->shost_recovery && ioc->pending_io_count) {
-			if (ioc->pending_io_count == 1)
-				wake_up(&ioc->reset_wq);
-			ioc->pending_io_count--;
-		}
+		_base_recovery_check(ioc);
 		return;
 	} else if (smid < ioc->internal_smid) {
 		/* hi-priority */
@@ -2723,14 +2772,23 @@ _base_allocate_memory_pools(struct MPT2SAS_ADAPTER *ioc,  int sleep_flag)
 	    ioc->name, (unsigned long long) ioc->request_dma));
 	total_sz += sz;
 
-	sz = ioc->scsiio_depth * sizeof(struct scsiio_tracker);
-	ioc->scsi_lookup_pages = get_order(sz);
-	ioc->scsi_lookup = (struct scsiio_tracker *)__get_free_pages(
-	    GFP_KERNEL, ioc->scsi_lookup_pages);
-	if (!ioc->scsi_lookup) {
-		printk(MPT2SAS_ERR_FMT "scsi_lookup: get_free_pages failed, "
-		    "sz(%d)\n", ioc->name, (int)sz);
-		goto out;
+	/*
+	 * Don't need to allocate memory for scsiio_tracker array if we
+	 * are using scsi-mq, we embed it in the scsi_cmnd for that case.
+	 */
+	if (!shost_use_blk_mq(ioc->shost)) {
+		sz = ioc->scsiio_depth * sizeof(struct scsiio_tracker);
+		ioc->scsi_lookup_pages = get_order(sz);
+		ioc->scsi_lookup = (struct scsiio_tracker *)__get_free_pages(
+				    GFP_KERNEL, ioc->scsi_lookup_pages);
+		if (!ioc->scsi_lookup) {
+			printk(MPT2SAS_ERR_FMT "scsi_lookup: get_free_pages "
+				"failed, sz(%d)\n", ioc->name, (int)sz);
+			goto out;
+		}
+	} else {
+		ioc->scsi_lookup_pages = 0;
+		ioc->scsi_lookup = NULL;
 	}
 
 	dinitprintk(ioc, printk(MPT2SAS_INFO_FMT "scsiio(0x%p): "
@@ -4299,15 +4357,17 @@ _base_make_ioc_operational(struct MPT2SAS_ADAPTER *ioc, int sleep_flag)
 	/* initialize the scsi lookup free list */
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	INIT_LIST_HEAD(&ioc->free_list);
-	smid = 1;
-	for (i = 0; i < ioc->scsiio_depth; i++, smid++) {
-		INIT_LIST_HEAD(&ioc->scsi_lookup[i].chain_list);
-		ioc->scsi_lookup[i].cb_idx = 0xFF;
-		ioc->scsi_lookup[i].smid = smid;
-		ioc->scsi_lookup[i].scmd = NULL;
-		ioc->scsi_lookup[i].direct_io = 0;
-		list_add_tail(&ioc->scsi_lookup[i].tracker_list,
-		    &ioc->free_list);
+	if (!shost_use_blk_mq(ioc->shost)) {
+		smid = 1;
+		for (i = 0; i < ioc->scsiio_depth; i++, smid++) {
+			INIT_LIST_HEAD(&ioc->scsi_lookup[i].chain_list);
+			ioc->scsi_lookup[i].cb_idx = 0xFF;
+			ioc->scsi_lookup[i].smid = smid;
+			ioc->scsi_lookup[i].scmd = NULL;
+			ioc->scsi_lookup[i].direct_io = 0;
+			list_add_tail(&ioc->scsi_lookup[i].tracker_list,
+					    &ioc->free_list);
+		}
 	}
 
 	/* hi-priority queue */
@@ -4772,7 +4832,7 @@ _wait_for_commands_to_complete(struct MPT2SAS_ADAPTER *ioc, int sleep_flag)
 {
 	u32 ioc_state;
 	unsigned long flags;
-	u16 i;
+	u16 i, pending, loops;
 
 	ioc->pending_io_count = 0;
 	if (sleep_flag != CAN_SLEEP)
@@ -4783,17 +4843,34 @@ _wait_for_commands_to_complete(struct MPT2SAS_ADAPTER *ioc, int sleep_flag)
 		return;
 
 	/* pending command count */
-	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
-	for (i = 0; i < ioc->scsiio_depth; i++)
-		if (ioc->scsi_lookup[i].cb_idx != 0xFF)
-			ioc->pending_io_count++;
-	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
+	loops = 0;
+	do {
+		pending = 0;
+		spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
+		for (i = 0; i < ioc->scsiio_depth; i++) {
+			struct scsiio_tracker *st;
+			struct scsi_cmnd *scmd;
+
+			if (shost_use_blk_mq(ioc->shost)) {
+				scmd = scsi_mq_find_tag(ioc->shost,  i);
+				if (scmd)
+					pending++;
+			} else {
+				st = mpt2sas_get_st_from_smid(ioc, i + 1);
+				if (st->cb_idx != 0xFF)
+					pending++;
+			}
+		}
+		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 
-	if (!ioc->pending_io_count)
-		return;
+		if (!pending)
+			break;
+
+		ioc->pending_io_count = 1;
 
-	/* wait for pending commands to complete */
-	wait_event_timeout(ioc->reset_wq, ioc->pending_io_count == 0, 10 * HZ);
+		/* wait for pending commands to complete */
+		wait_event_timeout(ioc->reset_wq, ioc->pending_io_count == 0, HZ);
+	} while (++loops <= 10);
 }
 
 /**
diff --git a/drivers/scsi/mpt2sas/mpt2sas_base.h b/drivers/scsi/mpt2sas/mpt2sas_base.h
index caff8d10cca4..cadb392126e0 100644
--- a/drivers/scsi/mpt2sas/mpt2sas_base.h
+++ b/drivers/scsi/mpt2sas/mpt2sas_base.h
@@ -1045,6 +1045,9 @@ u16 mpt2sas_base_get_smid_hpr(struct MPT2SAS_ADAPTER *ioc, u8 cb_idx);
 u16 mpt2sas_base_get_smid_scsiio(struct MPT2SAS_ADAPTER *ioc, u8 cb_idx,
     struct scsi_cmnd *scmd);
 
+
+struct scsiio_tracker *mpt2sas_get_st_from_smid(struct MPT2SAS_ADAPTER *ioc,
+    u16 smid);
 u16 mpt2sas_base_get_smid(struct MPT2SAS_ADAPTER *ioc, u8 cb_idx);
 void mpt2sas_base_free_smid(struct MPT2SAS_ADAPTER *ioc, u16 smid);
 void mpt2sas_base_put_smid_scsi_io(struct MPT2SAS_ADAPTER *ioc, u16 smid,
diff --git a/drivers/scsi/mpt2sas/mpt2sas_scsih.c b/drivers/scsi/mpt2sas/mpt2sas_scsih.c
index 3f26147bbc64..7543eb3a4a8d 100644
--- a/drivers/scsi/mpt2sas/mpt2sas_scsih.c
+++ b/drivers/scsi/mpt2sas/mpt2sas_scsih.c
@@ -884,7 +884,10 @@ _scsih_is_end_device(u32 device_info)
 static struct scsi_cmnd *
 _scsih_scsi_lookup_get(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 {
-	return ioc->scsi_lookup[smid - 1].scmd;
+	if (shost_use_blk_mq(ioc->shost))
+		return scsi_mq_find_tag(ioc->shost, smid - 1);
+	else
+		return ioc->scsi_lookup[smid - 1].scmd;
 }
 
 /**
@@ -901,6 +904,8 @@ _scsih_scsi_lookup_get_clear(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 	unsigned long flags;
 	struct scsi_cmnd *scmd;
 
+	BUG_ON(shost_use_blk_mq(ioc->shost));
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	scmd = ioc->scsi_lookup[smid - 1].scmd;
 	ioc->scsi_lookup[smid - 1].scmd = NULL;
@@ -927,6 +932,13 @@ _scsih_scsi_lookup_find_by_scmd(struct MPT2SAS_ADAPTER *ioc, struct scsi_cmnd
 	unsigned long	flags;
 	int i;
 
+	if (shost_use_blk_mq(ioc->shost)) {
+		struct scsiio_tracker *st;
+
+		st = scsi_cmd_priv(scmd);
+		return st->smid;
+	}
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	smid = 0;
 	for (i = 0; i < ioc->scsiio_depth; i++) {
@@ -961,9 +973,14 @@ _scsih_scsi_lookup_find_by_target(struct MPT2SAS_ADAPTER *ioc, int id,
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	found = 0;
 	for (i = 0 ; i < ioc->scsiio_depth; i++) {
-		if (ioc->scsi_lookup[i].scmd &&
-		    (ioc->scsi_lookup[i].scmd->device->id == id &&
-		    ioc->scsi_lookup[i].scmd->device->channel == channel)) {
+		struct scsiio_tracker *st;
+
+		st = mpt2sas_get_st_from_smid(ioc, i + 1);
+		if (!st)
+			continue;
+		if (st->scmd &&
+		    (st->scmd->device->id == id &&
+		    st->scmd->device->channel == channel)) {
 			found = 1;
 			goto out;
 		}
@@ -995,10 +1012,15 @@ _scsih_scsi_lookup_find_by_lun(struct MPT2SAS_ADAPTER *ioc, int id,
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	found = 0;
 	for (i = 0 ; i < ioc->scsiio_depth; i++) {
-		if (ioc->scsi_lookup[i].scmd &&
-		    (ioc->scsi_lookup[i].scmd->device->id == id &&
-		    ioc->scsi_lookup[i].scmd->device->channel == channel &&
-		    ioc->scsi_lookup[i].scmd->device->lun == lun)) {
+		struct scsiio_tracker *st;
+
+		st = mpt2sas_get_st_from_smid(ioc, i + 1);
+		if (!st)
+			continue;
+		if (st->scmd &&
+		    (st->scmd->device->id == id &&
+		    st->scmd->device->channel == channel &&
+		    st->scmd->device->lun == lun)) {
 			found = 1;
 			goto out;
 		}
@@ -1019,6 +1041,7 @@ static struct chain_tracker *
 _scsih_get_chain_buffer_tracker(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 {
 	struct chain_tracker *chain_req;
+	struct scsiio_tracker *st;
 	unsigned long flags;
 
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
@@ -1031,8 +1054,8 @@ _scsih_get_chain_buffer_tracker(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 	chain_req = list_entry(ioc->free_chain_list.next,
 	    struct chain_tracker, tracker_list);
 	list_del_init(&chain_req->tracker_list);
-	list_add_tail(&chain_req->tracker_list,
-	    &ioc->scsi_lookup[smid - 1].chain_list);
+	st = mpt2sas_get_st_from_smid(ioc, smid);
+	list_add_tail(&chain_req->tracker_list, &st->chain_list);
 	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 	return chain_req;
 }
@@ -2387,7 +2410,7 @@ mpt2sas_scsih_issue_tm(struct MPT2SAS_ADAPTER *ioc, u16 handle, uint channel,
 	}
 
 	if (type == MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK)
-		scsi_lookup = &ioc->scsi_lookup[smid_task - 1];
+		scsi_lookup = mpt2sas_get_st_from_smid(ioc, smid_task);
 
 	dtmprintk(ioc, printk(MPT2SAS_INFO_FMT "sending tm: handle(0x%04x),"
 	    " task_type(0x%02x), smid(%d)\n", ioc->name, handle, type,
@@ -3698,7 +3721,11 @@ _scsih_flush_running_cmds(struct MPT2SAS_ADAPTER *ioc)
 	u16 count = 0;
 
 	for (smid = 1; smid <= ioc->scsiio_depth; smid++) {
-		scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+		if (shost_use_blk_mq(ioc->shost))
+			scmd = _scsih_scsi_lookup_get(ioc, smid);
+		else
+			scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+
 		if (!scmd)
 			continue;
 		count++;
@@ -3809,7 +3836,7 @@ _scsih_eedp_error_handling(struct scsi_cmnd *scmd, u16 ioc_status)
 static inline u8
 _scsih_scsi_direct_io_get(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 {
-	return ioc->scsi_lookup[smid - 1].direct_io;
+	return mpt2sas_get_st_from_smid(ioc, smid)->direct_io;
 }
 
 /**
@@ -3823,7 +3850,7 @@ _scsih_scsi_direct_io_get(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 static inline void
 _scsih_scsi_direct_io_set(struct MPT2SAS_ADAPTER *ioc, u16 smid, u8 direct_io)
 {
-	ioc->scsi_lookup[smid - 1].direct_io = direct_io;
+	mpt2sas_get_st_from_smid(ioc, smid)->direct_io = direct_io;
 }
 
 
@@ -4443,7 +4470,11 @@ _scsih_io_done(struct MPT2SAS_ADAPTER *ioc, u16 smid, u8 msix_index, u32 reply)
 	unsigned long flags;
 
 	mpi_reply = mpt2sas_base_get_reply_virt_addr(ioc, reply);
-	scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+	if (shost_use_blk_mq(ioc->shost))
+		scmd = scsi_mq_find_tag(ioc->shost, smid - 1);
+	else
+		scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+
 	if (scmd == NULL)
 		return 1;
 
@@ -4468,10 +4499,12 @@ _scsih_io_done(struct MPT2SAS_ADAPTER *ioc, u16 smid, u8 msix_index, u32 reply)
 	if (_scsih_scsi_direct_io_get(ioc, smid) &&
 	    ((ioc_status & MPI2_IOCSTATUS_MASK)
 	    != MPI2_IOCSTATUS_SCSI_TASK_TERMINATED)) {
-		spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
-		ioc->scsi_lookup[smid - 1].scmd = scmd;
+		if (ioc->scsi_lookup) {
+			spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
+			ioc->scsi_lookup[smid - 1].scmd = scmd;
+			spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
+		}
 		_scsih_scsi_direct_io_set(ioc, smid, 0);
-		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 		memcpy(mpi_request->CDB.CDB32, scmd->cmnd, scmd->cmd_len);
 		mpi_request->DevHandle =
 		    cpu_to_le16(sas_device_priv_data->sas_target->handle);
@@ -7623,6 +7656,22 @@ mpt2sas_scsih_event_callback(struct MPT2SAS_ADAPTER *ioc, u8 msix_index,
 	return;
 }
 
+static int
+_scsih_init_command(struct Scsi_Host *shost, struct scsi_cmnd *cmd,
+			unsigned int request_idx)
+{
+	struct MPT2SAS_ADAPTER *ioc = shost_priv(shost);
+	struct scsiio_tracker *st;
+
+	st = (void *) cmd + sizeof(*cmd);
+	INIT_LIST_HEAD(&st->chain_list);
+	st->scmd = cmd;
+	st->cb_idx = ioc->scsi_io_cb_idx;
+	st->smid = request_idx + 1;
+	st->direct_io = 0;
+	return 0;
+}
+
 /* shost template */
 static struct scsi_host_template scsih_driver_template = {
 	.module				= THIS_MODULE,
@@ -7651,6 +7700,8 @@ static struct scsi_host_template scsih_driver_template = {
 	.shost_attrs			= mpt2sas_host_attrs,
 	.sdev_attrs			= mpt2sas_dev_attrs,
 	.track_queue_depth		= 1,
+	.cmd_size			= sizeof(struct scsiio_tracker),
+	.init_command			= _scsih_init_command,
 };
 
 /**
-- 
1.9.1

[PATCH 4/6] mpt2sas: store scsi io tracker data in the scsi request

[Jens Axboe]

Instead of storing the IO tracker structure in a separate list
that we need to pop/push to on every submit and complete (and
lock), store it in the pdu associated with a request. This is
possible on scsi-mq only, and further cuts the spinlock associated
time for higher IOPS IO workloads. At 100K IOPS, this effectively
cuts the locking time in half.

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 drivers/scsi/mpt2sas/mpt2sas_base.c  | 187 ++++++++++++++++++++++++-----------
 drivers/scsi/mpt2sas/mpt2sas_base.h  |   3 +
 drivers/scsi/mpt2sas/mpt2sas_scsih.c |  87 ++++++++++++----
 3 files changed, 204 insertions(+), 73 deletions(-)

diff --git a/drivers/scsi/mpt2sas/mpt2sas_base.c b/drivers/scsi/mpt2sas/mpt2sas_base.c
index 11248de92b3b..5adb8469ede4 100644
--- a/drivers/scsi/mpt2sas/mpt2sas_base.c
+++ b/drivers/scsi/mpt2sas/mpt2sas_base.c
@@ -828,6 +828,20 @@ _base_async_event(struct MPT2SAS_ADAPTER *ioc, u8 msix_index, u32 reply)
 	return;
 }
 
+struct scsiio_tracker *
+mpt2sas_get_st_from_smid(struct MPT2SAS_ADAPTER *ioc, u16 smid)
+{
+	if (shost_use_blk_mq(ioc->shost)) {
+		struct scsi_cmnd *scmd;
+
+		scmd = scsi_mq_find_tag(ioc->shost, smid - 1);
+		if (!scmd)
+			return NULL;
+		return scsi_cmd_priv(scmd);
+	} else
+		return &ioc->scsi_lookup[smid - 1];
+}
+
 /**
  * _base_get_cb_idx - obtain the callback index
  * @ioc: per adapter object
@@ -842,8 +856,10 @@ _base_get_cb_idx(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 	u8 cb_idx;
 
 	if (smid < ioc->hi_priority_smid) {
-		i = smid - 1;
-		cb_idx = ioc->scsi_lookup[i].cb_idx;
+		struct scsiio_tracker *st;
+
+		st = mpt2sas_get_st_from_smid(ioc, smid);
+		cb_idx = st->cb_idx;
 	} else if (smid < ioc->internal_smid) {
 		i = smid - ioc->hi_priority_smid;
 		cb_idx = ioc->hpr_lookup[i].cb_idx;
@@ -962,18 +978,17 @@ _base_interrupt(int irq, void *bus_id)
 			goto next;
 		if (smid) {
 			cb_idx = _base_get_cb_idx(ioc, smid);
-		if ((likely(cb_idx < MPT_MAX_CALLBACKS))
+			if ((likely(cb_idx < MPT_MAX_CALLBACKS))
 			    && (likely(mpt_callbacks[cb_idx] != NULL))) {
 				rc = mpt_callbacks[cb_idx](ioc, smid,
 				    msix_index, reply);
-			if (reply)
-				_base_display_reply_info(ioc, smid,
-				    msix_index, reply);
-			if (rc)
-				mpt2sas_base_free_smid(ioc, smid);
+				if (reply)
+					_base_display_reply_info(ioc, smid,
+							msix_index, reply);
+				if (rc)
+					mpt2sas_base_free_smid(ioc, smid);
 			}
-		}
-		if (!smid)
+		} else
 			_base_async_event(ioc, msix_index, reply);
 
 		/* reply free queue handling */
@@ -1724,6 +1739,11 @@ mpt2sas_base_get_smid_scsiio(struct MPT2SAS_ADAPTER *ioc, u8 cb_idx,
 	struct scsiio_tracker *request;
 	u16 smid;
 
+	if (shost_use_blk_mq(ioc->shost)) {
+		request = scsi_cmd_priv(scmd);
+		return request->smid;
+	}
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	if (list_empty(&ioc->free_list)) {
 		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
@@ -1771,6 +1791,31 @@ mpt2sas_base_get_smid_hpr(struct MPT2SAS_ADAPTER *ioc, u8 cb_idx)
 	return smid;
 }
 
+static void
+_base_recovery_check(struct MPT2SAS_ADAPTER *ioc)
+{
+	/*
+	 * See _wait_for_commands_to_complete() call with regards to this code.
+	 */
+	if (ioc->shost_recovery && ioc->pending_io_count) {
+		if (ioc->pending_io_count == 1)
+			wake_up(&ioc->reset_wq);
+		ioc->pending_io_count = 0;
+	}
+}
+
+static void
+_dechain_st(struct MPT2SAS_ADAPTER *ioc, struct scsiio_tracker *st)
+{
+	struct chain_tracker *chain_req;
+
+	while (!list_empty(&st->chain_list)) {
+		chain_req = list_first_entry(&st->chain_list,
+						struct chain_tracker,
+						tracker_list);
+		list_move(&chain_req->tracker_list, &ioc->free_chain_list);
+	}
+}
 
 /**
  * mpt2sas_base_free_smid - put smid back on free_list
@@ -1784,20 +1829,32 @@ mpt2sas_base_free_smid(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 {
 	unsigned long flags;
 	int i;
-	struct chain_tracker *chain_req, *next;
+
+	if (shost_use_blk_mq(ioc->shost) && smid < ioc->hi_priority_smid) {
+		struct scsiio_tracker *st;
+
+		st = mpt2sas_get_st_from_smid(ioc, smid);
+		if (!st)
+			return;
+
+		st->direct_io = 0;
+
+		if (!list_empty(&st->chain_list)) {
+			spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
+			_dechain_st(ioc, st);
+			spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
+		}
+
+		_base_recovery_check(ioc);
+		return;
+	}
 
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	if (smid < ioc->hi_priority_smid) {
 		/* scsiio queue */
 		i = smid - 1;
-		if (!list_empty(&ioc->scsi_lookup[i].chain_list)) {
-			list_for_each_entry_safe(chain_req, next,
-			    &ioc->scsi_lookup[i].chain_list, tracker_list) {
-				list_del_init(&chain_req->tracker_list);
-				list_add(&chain_req->tracker_list,
-				    &ioc->free_chain_list);
-			}
-		}
+		if (!list_empty(&ioc->scsi_lookup[i].chain_list))
+			_dechain_st(ioc, &ioc->scsi_lookup[i]);
 		ioc->scsi_lookup[i].cb_idx = 0xFF;
 		ioc->scsi_lookup[i].scmd = NULL;
 		ioc->scsi_lookup[i].direct_io = 0;
@@ -1805,15 +1862,7 @@ mpt2sas_base_free_smid(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 		    &ioc->free_list);
 		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 
-		/*
-		 * See _wait_for_commands_to_complete() call with regards
-		 * to this code.
-		 */
-		if (ioc->shost_recovery && ioc->pending_io_count) {
-			if (ioc->pending_io_count == 1)
-				wake_up(&ioc->reset_wq);
-			ioc->pending_io_count--;
-		}
+		_base_recovery_check(ioc);
 		return;
 	} else if (smid < ioc->internal_smid) {
 		/* hi-priority */
@@ -2723,14 +2772,23 @@ _base_allocate_memory_pools(struct MPT2SAS_ADAPTER *ioc,  int sleep_flag)
 	    ioc->name, (unsigned long long) ioc->request_dma));
 	total_sz += sz;
 
-	sz = ioc->scsiio_depth * sizeof(struct scsiio_tracker);
-	ioc->scsi_lookup_pages = get_order(sz);
-	ioc->scsi_lookup = (struct scsiio_tracker *)__get_free_pages(
-	    GFP_KERNEL, ioc->scsi_lookup_pages);
-	if (!ioc->scsi_lookup) {
-		printk(MPT2SAS_ERR_FMT "scsi_lookup: get_free_pages failed, "
-		    "sz(%d)\n", ioc->name, (int)sz);
-		goto out;
+	/*
+	 * Don't need to allocate memory for scsiio_tracker array if we
+	 * are using scsi-mq, we embed it in the scsi_cmnd for that case.
+	 */
+	if (!shost_use_blk_mq(ioc->shost)) {
+		sz = ioc->scsiio_depth * sizeof(struct scsiio_tracker);
+		ioc->scsi_lookup_pages = get_order(sz);
+		ioc->scsi_lookup = (struct scsiio_tracker *)__get_free_pages(
+				    GFP_KERNEL, ioc->scsi_lookup_pages);
+		if (!ioc->scsi_lookup) {
+			printk(MPT2SAS_ERR_FMT "scsi_lookup: get_free_pages "
+				"failed, sz(%d)\n", ioc->name, (int)sz);
+			goto out;
+		}
+	} else {
+		ioc->scsi_lookup_pages = 0;
+		ioc->scsi_lookup = NULL;
 	}
 
 	dinitprintk(ioc, printk(MPT2SAS_INFO_FMT "scsiio(0x%p): "
@@ -4299,15 +4357,17 @@ _base_make_ioc_operational(struct MPT2SAS_ADAPTER *ioc, int sleep_flag)
 	/* initialize the scsi lookup free list */
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	INIT_LIST_HEAD(&ioc->free_list);
-	smid = 1;
-	for (i = 0; i < ioc->scsiio_depth; i++, smid++) {
-		INIT_LIST_HEAD(&ioc->scsi_lookup[i].chain_list);
-		ioc->scsi_lookup[i].cb_idx = 0xFF;
-		ioc->scsi_lookup[i].smid = smid;
-		ioc->scsi_lookup[i].scmd = NULL;
-		ioc->scsi_lookup[i].direct_io = 0;
-		list_add_tail(&ioc->scsi_lookup[i].tracker_list,
-		    &ioc->free_list);
+	if (!shost_use_blk_mq(ioc->shost)) {
+		smid = 1;
+		for (i = 0; i < ioc->scsiio_depth; i++, smid++) {
+			INIT_LIST_HEAD(&ioc->scsi_lookup[i].chain_list);
+			ioc->scsi_lookup[i].cb_idx = 0xFF;
+			ioc->scsi_lookup[i].smid = smid;
+			ioc->scsi_lookup[i].scmd = NULL;
+			ioc->scsi_lookup[i].direct_io = 0;
+			list_add_tail(&ioc->scsi_lookup[i].tracker_list,
+					    &ioc->free_list);
+		}
 	}
 
 	/* hi-priority queue */
@@ -4772,7 +4832,7 @@ _wait_for_commands_to_complete(struct MPT2SAS_ADAPTER *ioc, int sleep_flag)
 {
 	u32 ioc_state;
 	unsigned long flags;
-	u16 i;
+	u16 i, pending, loops;
 
 	ioc->pending_io_count = 0;
 	if (sleep_flag != CAN_SLEEP)
@@ -4783,17 +4843,34 @@ _wait_for_commands_to_complete(struct MPT2SAS_ADAPTER *ioc, int sleep_flag)
 		return;
 
 	/* pending command count */
-	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
-	for (i = 0; i < ioc->scsiio_depth; i++)
-		if (ioc->scsi_lookup[i].cb_idx != 0xFF)
-			ioc->pending_io_count++;
-	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
+	loops = 0;
+	do {
+		pending = 0;
+		spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
+		for (i = 0; i < ioc->scsiio_depth; i++) {
+			struct scsiio_tracker *st;
+			struct scsi_cmnd *scmd;
+
+			if (shost_use_blk_mq(ioc->shost)) {
+				scmd = scsi_mq_find_tag(ioc->shost,  i);
+				if (scmd)
+					pending++;
+			} else {
+				st = mpt2sas_get_st_from_smid(ioc, i + 1);
+				if (st->cb_idx != 0xFF)
+					pending++;
+			}
+		}
+		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 
-	if (!ioc->pending_io_count)
-		return;
+		if (!pending)
+			break;
+
+		ioc->pending_io_count = 1;
 
-	/* wait for pending commands to complete */
-	wait_event_timeout(ioc->reset_wq, ioc->pending_io_count == 0, 10 * HZ);
+		/* wait for pending commands to complete */
+		wait_event_timeout(ioc->reset_wq, ioc->pending_io_count == 0, HZ);
+	} while (++loops <= 10);
 }
 
 /**
diff --git a/drivers/scsi/mpt2sas/mpt2sas_base.h b/drivers/scsi/mpt2sas/mpt2sas_base.h
index caff8d10cca4..cadb392126e0 100644
--- a/drivers/scsi/mpt2sas/mpt2sas_base.h
+++ b/drivers/scsi/mpt2sas/mpt2sas_base.h
@@ -1045,6 +1045,9 @@ u16 mpt2sas_base_get_smid_hpr(struct MPT2SAS_ADAPTER *ioc, u8 cb_idx);
 u16 mpt2sas_base_get_smid_scsiio(struct MPT2SAS_ADAPTER *ioc, u8 cb_idx,
     struct scsi_cmnd *scmd);
 
+
+struct scsiio_tracker *mpt2sas_get_st_from_smid(struct MPT2SAS_ADAPTER *ioc,
+    u16 smid);
 u16 mpt2sas_base_get_smid(struct MPT2SAS_ADAPTER *ioc, u8 cb_idx);
 void mpt2sas_base_free_smid(struct MPT2SAS_ADAPTER *ioc, u16 smid);
 void mpt2sas_base_put_smid_scsi_io(struct MPT2SAS_ADAPTER *ioc, u16 smid,
diff --git a/drivers/scsi/mpt2sas/mpt2sas_scsih.c b/drivers/scsi/mpt2sas/mpt2sas_scsih.c
index 3f26147bbc64..7543eb3a4a8d 100644
--- a/drivers/scsi/mpt2sas/mpt2sas_scsih.c
+++ b/drivers/scsi/mpt2sas/mpt2sas_scsih.c
@@ -884,7 +884,10 @@ _scsih_is_end_device(u32 device_info)
 static struct scsi_cmnd *
 _scsih_scsi_lookup_get(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 {
-	return ioc->scsi_lookup[smid - 1].scmd;
+	if (shost_use_blk_mq(ioc->shost))
+		return scsi_mq_find_tag(ioc->shost, smid - 1);
+	else
+		return ioc->scsi_lookup[smid - 1].scmd;
 }
 
 /**
@@ -901,6 +904,8 @@ _scsih_scsi_lookup_get_clear(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 	unsigned long flags;
 	struct scsi_cmnd *scmd;
 
+	BUG_ON(shost_use_blk_mq(ioc->shost));
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	scmd = ioc->scsi_lookup[smid - 1].scmd;
 	ioc->scsi_lookup[smid - 1].scmd = NULL;
@@ -927,6 +932,13 @@ _scsih_scsi_lookup_find_by_scmd(struct MPT2SAS_ADAPTER *ioc, struct scsi_cmnd
 	unsigned long	flags;
 	int i;
 
+	if (shost_use_blk_mq(ioc->shost)) {
+		struct scsiio_tracker *st;
+
+		st = scsi_cmd_priv(scmd);
+		return st->smid;
+	}
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	smid = 0;
 	for (i = 0; i < ioc->scsiio_depth; i++) {
@@ -961,9 +973,14 @@ _scsih_scsi_lookup_find_by_target(struct MPT2SAS_ADAPTER *ioc, int id,
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	found = 0;
 	for (i = 0 ; i < ioc->scsiio_depth; i++) {
-		if (ioc->scsi_lookup[i].scmd &&
-		    (ioc->scsi_lookup[i].scmd->device->id == id &&
-		    ioc->scsi_lookup[i].scmd->device->channel == channel)) {
+		struct scsiio_tracker *st;
+
+		st = mpt2sas_get_st_from_smid(ioc, i + 1);
+		if (!st)
+			continue;
+		if (st->scmd &&
+		    (st->scmd->device->id == id &&
+		    st->scmd->device->channel == channel)) {
 			found = 1;
 			goto out;
 		}
@@ -995,10 +1012,15 @@ _scsih_scsi_lookup_find_by_lun(struct MPT2SAS_ADAPTER *ioc, int id,
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	found = 0;
 	for (i = 0 ; i < ioc->scsiio_depth; i++) {
-		if (ioc->scsi_lookup[i].scmd &&
-		    (ioc->scsi_lookup[i].scmd->device->id == id &&
-		    ioc->scsi_lookup[i].scmd->device->channel == channel &&
-		    ioc->scsi_lookup[i].scmd->device->lun == lun)) {
+		struct scsiio_tracker *st;
+
+		st = mpt2sas_get_st_from_smid(ioc, i + 1);
+		if (!st)
+			continue;
+		if (st->scmd &&
+		    (st->scmd->device->id == id &&
+		    st->scmd->device->channel == channel &&
+		    st->scmd->device->lun == lun)) {
 			found = 1;
 			goto out;
 		}
@@ -1019,6 +1041,7 @@ static struct chain_tracker *
 _scsih_get_chain_buffer_tracker(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 {
 	struct chain_tracker *chain_req;
+	struct scsiio_tracker *st;
 	unsigned long flags;
 
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
@@ -1031,8 +1054,8 @@ _scsih_get_chain_buffer_tracker(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 	chain_req = list_entry(ioc->free_chain_list.next,
 	    struct chain_tracker, tracker_list);
 	list_del_init(&chain_req->tracker_list);
-	list_add_tail(&chain_req->tracker_list,
-	    &ioc->scsi_lookup[smid - 1].chain_list);
+	st = mpt2sas_get_st_from_smid(ioc, smid);
+	list_add_tail(&chain_req->tracker_list, &st->chain_list);
 	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 	return chain_req;
 }
@@ -2387,7 +2410,7 @@ mpt2sas_scsih_issue_tm(struct MPT2SAS_ADAPTER *ioc, u16 handle, uint channel,
 	}
 
 	if (type == MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK)
-		scsi_lookup = &ioc->scsi_lookup[smid_task - 1];
+		scsi_lookup = mpt2sas_get_st_from_smid(ioc, smid_task);
 
 	dtmprintk(ioc, printk(MPT2SAS_INFO_FMT "sending tm: handle(0x%04x),"
 	    " task_type(0x%02x), smid(%d)\n", ioc->name, handle, type,
@@ -3698,7 +3721,11 @@ _scsih_flush_running_cmds(struct MPT2SAS_ADAPTER *ioc)
 	u16 count = 0;
 
 	for (smid = 1; smid <= ioc->scsiio_depth; smid++) {
-		scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+		if (shost_use_blk_mq(ioc->shost))
+			scmd = _scsih_scsi_lookup_get(ioc, smid);
+		else
+			scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+
 		if (!scmd)
 			continue;
 		count++;
@@ -3809,7 +3836,7 @@ _scsih_eedp_error_handling(struct scsi_cmnd *scmd, u16 ioc_status)
 static inline u8
 _scsih_scsi_direct_io_get(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 {
-	return ioc->scsi_lookup[smid - 1].direct_io;
+	return mpt2sas_get_st_from_smid(ioc, smid)->direct_io;
 }
 
 /**
@@ -3823,7 +3850,7 @@ _scsih_scsi_direct_io_get(struct MPT2SAS_ADAPTER *ioc, u16 smid)
 static inline void
 _scsih_scsi_direct_io_set(struct MPT2SAS_ADAPTER *ioc, u16 smid, u8 direct_io)
 {
-	ioc->scsi_lookup[smid - 1].direct_io = direct_io;
+	mpt2sas_get_st_from_smid(ioc, smid)->direct_io = direct_io;
 }
 
 
@@ -4443,7 +4470,11 @@ _scsih_io_done(struct MPT2SAS_ADAPTER *ioc, u16 smid, u8 msix_index, u32 reply)
 	unsigned long flags;
 
 	mpi_reply = mpt2sas_base_get_reply_virt_addr(ioc, reply);
-	scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+	if (shost_use_blk_mq(ioc->shost))
+		scmd = scsi_mq_find_tag(ioc->shost, smid - 1);
+	else
+		scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+
 	if (scmd == NULL)
 		return 1;
 
@@ -4468,10 +4499,12 @@ _scsih_io_done(struct MPT2SAS_ADAPTER *ioc, u16 smid, u8 msix_index, u32 reply)
 	if (_scsih_scsi_direct_io_get(ioc, smid) &&
 	    ((ioc_status & MPI2_IOCSTATUS_MASK)
 	    != MPI2_IOCSTATUS_SCSI_TASK_TERMINATED)) {
-		spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
-		ioc->scsi_lookup[smid - 1].scmd = scmd;
+		if (ioc->scsi_lookup) {
+			spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
+			ioc->scsi_lookup[smid - 1].scmd = scmd;
+			spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
+		}
 		_scsih_scsi_direct_io_set(ioc, smid, 0);
-		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 		memcpy(mpi_request->CDB.CDB32, scmd->cmnd, scmd->cmd_len);
 		mpi_request->DevHandle =
 		    cpu_to_le16(sas_device_priv_data->sas_target->handle);
@@ -7623,6 +7656,22 @@ mpt2sas_scsih_event_callback(struct MPT2SAS_ADAPTER *ioc, u8 msix_index,
 	return;
 }
 
+static int
+_scsih_init_command(struct Scsi_Host *shost, struct scsi_cmnd *cmd,
+			unsigned int request_idx)
+{
+	struct MPT2SAS_ADAPTER *ioc = shost_priv(shost);
+	struct scsiio_tracker *st;
+
+	st = (void *) cmd + sizeof(*cmd);
+	INIT_LIST_HEAD(&st->chain_list);
+	st->scmd = cmd;
+	st->cb_idx = ioc->scsi_io_cb_idx;
+	st->smid = request_idx + 1;
+	st->direct_io = 0;
+	return 0;
+}
+
 /* shost template */
 static struct scsi_host_template scsih_driver_template = {
 	.module				= THIS_MODULE,
@@ -7651,6 +7700,8 @@ static struct scsi_host_template scsih_driver_template = {
 	.shost_attrs			= mpt2sas_host_attrs,
 	.sdev_attrs			= mpt2sas_dev_attrs,
 	.track_queue_depth		= 1,
+	.cmd_size			= sizeof(struct scsiio_tracker),
+	.init_command			= _scsih_init_command,
 };
 
 /**
-- 
1.9.1

Re: [PATCH 4/6] mpt2sas: store scsi io tracker data in the scsi request

[Christoph Hellwig]

On Tue, Apr 07, 2015 at 11:18:38AM -0600, Jens Axboe wrote:
> Instead of storing the IO tracker structure in a separate list
> that we need to pop/push to on every submit and complete (and
> lock), store it in the pdu associated with a request. This is
> possible on scsi-mq only, and further cuts the spinlock associated
> time for higher IOPS IO workloads. At 100K IOPS, this effectively
> cuts the locking time in half.

As mentioned before please call scsi_init_shared_tag_map and use your
new code path for the !mq case as well.

[PATCH 5/6] mpt3sas: get rid of MPI2_FUNCTION_SCSI_TASK_MGMT

[Jens Axboe]

This involves a lot of nasty lookup of a specific running command,
across the host. If anybody is using this code, it should be moved
to proper error handling instead.

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 drivers/scsi/mpt3sas/mpt3sas_ctl.c | 130 +++----------------------------------
 1 file changed, 10 insertions(+), 120 deletions(-)

diff --git a/drivers/scsi/mpt3sas/mpt3sas_ctl.c b/drivers/scsi/mpt3sas/mpt3sas_ctl.c
index 080c8a76d23d..2cdcff321d7b 100644
--- a/drivers/scsi/mpt3sas/mpt3sas_ctl.c
+++ b/drivers/scsi/mpt3sas/mpt3sas_ctl.c
@@ -524,85 +524,6 @@ _ctl_poll(struct file *filep, poll_table *wait)
 }
 
 /**
- * _ctl_set_task_mid - assign an active smid to tm request
- * @ioc: per adapter object
- * @karg - (struct mpt3_ioctl_command)
- * @tm_request - pointer to mf from user space
- *
- * Returns 0 when an smid if found, else fail.
- * during failure, the reply frame is filled.
- */
-static int
-_ctl_set_task_mid(struct MPT3SAS_ADAPTER *ioc, struct mpt3_ioctl_command *karg,
-	Mpi2SCSITaskManagementRequest_t *tm_request)
-{
-	u8 found = 0;
-	u16 i;
-	u16 handle;
-	struct scsi_cmnd *scmd;
-	struct MPT3SAS_DEVICE *priv_data;
-	unsigned long flags;
-	Mpi2SCSITaskManagementReply_t *tm_reply;
-	u32 sz;
-	u32 lun;
-	char *desc = NULL;
-
-	if (tm_request->TaskType == MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK)
-		desc = "abort_task";
-	else if (tm_request->TaskType == MPI2_SCSITASKMGMT_TASKTYPE_QUERY_TASK)
-		desc = "query_task";
-	else
-		return 0;
-
-	lun = scsilun_to_int((struct scsi_lun *)tm_request->LUN);
-
-	handle = le16_to_cpu(tm_request->DevHandle);
-	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
-	for (i = ioc->scsiio_depth; i && !found; i--) {
-		scmd = ioc->scsi_lookup[i - 1].scmd;
-		if (scmd == NULL || scmd->device == NULL ||
-		    scmd->device->hostdata == NULL)
-			continue;
-		if (lun != scmd->device->lun)
-			continue;
-		priv_data = scmd->device->hostdata;
-		if (priv_data->sas_target == NULL)
-			continue;
-		if (priv_data->sas_target->handle != handle)
-			continue;
-		tm_request->TaskMID = cpu_to_le16(ioc->scsi_lookup[i - 1].smid);
-		found = 1;
-	}
-	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
-
-	if (!found) {
-		dctlprintk(ioc, pr_info(MPT3SAS_FMT
-			"%s: handle(0x%04x), lun(%d), no active mid!!\n",
-			ioc->name,
-		    desc, le16_to_cpu(tm_request->DevHandle), lun));
-		tm_reply = ioc->ctl_cmds.reply;
-		tm_reply->DevHandle = tm_request->DevHandle;
-		tm_reply->Function = MPI2_FUNCTION_SCSI_TASK_MGMT;
-		tm_reply->TaskType = tm_request->TaskType;
-		tm_reply->MsgLength = sizeof(Mpi2SCSITaskManagementReply_t)/4;
-		tm_reply->VP_ID = tm_request->VP_ID;
-		tm_reply->VF_ID = tm_request->VF_ID;
-		sz = min_t(u32, karg->max_reply_bytes, ioc->reply_sz);
-		if (copy_to_user(karg->reply_frame_buf_ptr, ioc->ctl_cmds.reply,
-		    sz))
-			pr_err("failure at %s:%d/%s()!\n", __FILE__,
-			    __LINE__, __func__);
-		return 1;
-	}
-
-	dctlprintk(ioc, pr_info(MPT3SAS_FMT
-		"%s: handle(0x%04x), lun(%d), task_mid(%d)\n", ioc->name,
-	    desc, le16_to_cpu(tm_request->DevHandle), lun,
-	     le16_to_cpu(tm_request->TaskMID)));
-	return 0;
-}
-
-/**
  * _ctl_do_mpt_command - main handler for MPT3COMMAND opcode
  * @ioc: per adapter object
  * @karg - (struct mpt3_ioctl_command)
@@ -685,22 +606,18 @@ _ctl_do_mpt_command(struct MPT3SAS_ADAPTER *ioc, struct mpt3_ioctl_command karg,
 	}
 
 	if (mpi_request->Function == MPI2_FUNCTION_SCSI_TASK_MGMT) {
-		smid = mpt3sas_base_get_smid_hpr(ioc, ioc->ctl_cb_idx);
-		if (!smid) {
-			pr_err(MPT3SAS_FMT "%s: failed obtaining a smid\n",
-			    ioc->name, __func__);
-			ret = -EAGAIN;
-			goto out;
-		}
-	} else {
+		printk(MPT3SAS_FMT "%s: ignoring "
+			"MPI2_FUNCTION_SCSI_TASK_MGMT\n", ioc->name, __func__);
+		ret = -EINVAL;
+		goto out;
+	}
 
-		smid = mpt3sas_base_get_smid_scsiio(ioc, ioc->ctl_cb_idx, NULL);
-		if (!smid) {
-			pr_err(MPT3SAS_FMT "%s: failed obtaining a smid\n",
+	smid = mpt3sas_base_get_smid_hpr(ioc, ioc->ctl_cb_idx);
+	if (!smid) {
+		pr_err(MPT3SAS_FMT "%s: failed obtaining a smid\n",
 			    ioc->name, __func__);
-			ret = -EAGAIN;
-			goto out;
-		}
+		ret = -EAGAIN;
+		goto out;
 	}
 
 	ret = 0;
@@ -784,33 +701,6 @@ _ctl_do_mpt_command(struct MPT3SAS_ADAPTER *ioc, struct mpt3_ioctl_command karg,
 			mpt3sas_base_put_smid_default(ioc, smid);
 		break;
 	}
-	case MPI2_FUNCTION_SCSI_TASK_MGMT:
-	{
-		Mpi2SCSITaskManagementRequest_t *tm_request =
-		    (Mpi2SCSITaskManagementRequest_t *)request;
-
-		dtmprintk(ioc, pr_info(MPT3SAS_FMT
-			"TASK_MGMT: handle(0x%04x), task_type(0x%02x)\n",
-			ioc->name,
-		    le16_to_cpu(tm_request->DevHandle), tm_request->TaskType));
-
-		if (tm_request->TaskType ==
-		    MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK ||
-		    tm_request->TaskType ==
-		    MPI2_SCSITASKMGMT_TASKTYPE_QUERY_TASK) {
-			if (_ctl_set_task_mid(ioc, &karg, tm_request)) {
-				mpt3sas_base_free_smid(ioc, smid);
-				goto out;
-			}
-		}
-
-		mpt3sas_scsih_set_tm_flag(ioc, le16_to_cpu(
-		    tm_request->DevHandle));
-		ioc->build_sg_mpi(ioc, psge, data_out_dma, data_out_sz,
-		    data_in_dma, data_in_sz);
-		mpt3sas_base_put_smid_hi_priority(ioc, smid);
-		break;
-	}
 	case MPI2_FUNCTION_SMP_PASSTHROUGH:
 	{
 		Mpi2SmpPassthroughRequest_t *smp_request =
-- 
1.9.1

[PATCH 5/6] mpt3sas: get rid of MPI2_FUNCTION_SCSI_TASK_MGMT

[Jens Axboe]

This involves a lot of nasty lookup of a specific running command,
across the host. If anybody is using this code, it should be moved
to proper error handling instead.

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 drivers/scsi/mpt3sas/mpt3sas_ctl.c | 130 +++----------------------------------
 1 file changed, 10 insertions(+), 120 deletions(-)

diff --git a/drivers/scsi/mpt3sas/mpt3sas_ctl.c b/drivers/scsi/mpt3sas/mpt3sas_ctl.c
index 080c8a76d23d..2cdcff321d7b 100644
--- a/drivers/scsi/mpt3sas/mpt3sas_ctl.c
+++ b/drivers/scsi/mpt3sas/mpt3sas_ctl.c
@@ -524,85 +524,6 @@ _ctl_poll(struct file *filep, poll_table *wait)
 }
 
 /**
- * _ctl_set_task_mid - assign an active smid to tm request
- * @ioc: per adapter object
- * @karg - (struct mpt3_ioctl_command)
- * @tm_request - pointer to mf from user space
- *
- * Returns 0 when an smid if found, else fail.
- * during failure, the reply frame is filled.
- */
-static int
-_ctl_set_task_mid(struct MPT3SAS_ADAPTER *ioc, struct mpt3_ioctl_command *karg,
-	Mpi2SCSITaskManagementRequest_t *tm_request)
-{
-	u8 found = 0;
-	u16 i;
-	u16 handle;
-	struct scsi_cmnd *scmd;
-	struct MPT3SAS_DEVICE *priv_data;
-	unsigned long flags;
-	Mpi2SCSITaskManagementReply_t *tm_reply;
-	u32 sz;
-	u32 lun;
-	char *desc = NULL;
-
-	if (tm_request->TaskType == MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK)
-		desc = "abort_task";
-	else if (tm_request->TaskType == MPI2_SCSITASKMGMT_TASKTYPE_QUERY_TASK)
-		desc = "query_task";
-	else
-		return 0;
-
-	lun = scsilun_to_int((struct scsi_lun *)tm_request->LUN);
-
-	handle = le16_to_cpu(tm_request->DevHandle);
-	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
-	for (i = ioc->scsiio_depth; i && !found; i--) {
-		scmd = ioc->scsi_lookup[i - 1].scmd;
-		if (scmd == NULL || scmd->device == NULL ||
-		    scmd->device->hostdata == NULL)
-			continue;
-		if (lun != scmd->device->lun)
-			continue;
-		priv_data = scmd->device->hostdata;
-		if (priv_data->sas_target == NULL)
-			continue;
-		if (priv_data->sas_target->handle != handle)
-			continue;
-		tm_request->TaskMID = cpu_to_le16(ioc->scsi_lookup[i - 1].smid);
-		found = 1;
-	}
-	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
-
-	if (!found) {
-		dctlprintk(ioc, pr_info(MPT3SAS_FMT
-			"%s: handle(0x%04x), lun(%d), no active mid!!\n",
-			ioc->name,
-		    desc, le16_to_cpu(tm_request->DevHandle), lun));
-		tm_reply = ioc->ctl_cmds.reply;
-		tm_reply->DevHandle = tm_request->DevHandle;
-		tm_reply->Function = MPI2_FUNCTION_SCSI_TASK_MGMT;
-		tm_reply->TaskType = tm_request->TaskType;
-		tm_reply->MsgLength = sizeof(Mpi2SCSITaskManagementReply_t)/4;
-		tm_reply->VP_ID = tm_request->VP_ID;
-		tm_reply->VF_ID = tm_request->VF_ID;
-		sz = min_t(u32, karg->max_reply_bytes, ioc->reply_sz);
-		if (copy_to_user(karg->reply_frame_buf_ptr, ioc->ctl_cmds.reply,
-		    sz))
-			pr_err("failure at %s:%d/%s()!\n", __FILE__,
-			    __LINE__, __func__);
-		return 1;
-	}
-
-	dctlprintk(ioc, pr_info(MPT3SAS_FMT
-		"%s: handle(0x%04x), lun(%d), task_mid(%d)\n", ioc->name,
-	    desc, le16_to_cpu(tm_request->DevHandle), lun,
-	     le16_to_cpu(tm_request->TaskMID)));
-	return 0;
-}
-
-/**
  * _ctl_do_mpt_command - main handler for MPT3COMMAND opcode
  * @ioc: per adapter object
  * @karg - (struct mpt3_ioctl_command)
@@ -685,22 +606,18 @@ _ctl_do_mpt_command(struct MPT3SAS_ADAPTER *ioc, struct mpt3_ioctl_command karg,
 	}
 
 	if (mpi_request->Function == MPI2_FUNCTION_SCSI_TASK_MGMT) {
-		smid = mpt3sas_base_get_smid_hpr(ioc, ioc->ctl_cb_idx);
-		if (!smid) {
-			pr_err(MPT3SAS_FMT "%s: failed obtaining a smid\n",
-			    ioc->name, __func__);
-			ret = -EAGAIN;
-			goto out;
-		}
-	} else {
+		printk(MPT3SAS_FMT "%s: ignoring "
+			"MPI2_FUNCTION_SCSI_TASK_MGMT\n", ioc->name, __func__);
+		ret = -EINVAL;
+		goto out;
+	}
 
-		smid = mpt3sas_base_get_smid_scsiio(ioc, ioc->ctl_cb_idx, NULL);
-		if (!smid) {
-			pr_err(MPT3SAS_FMT "%s: failed obtaining a smid\n",
+	smid = mpt3sas_base_get_smid_hpr(ioc, ioc->ctl_cb_idx);
+	if (!smid) {
+		pr_err(MPT3SAS_FMT "%s: failed obtaining a smid\n",
 			    ioc->name, __func__);
-			ret = -EAGAIN;
-			goto out;
-		}
+		ret = -EAGAIN;
+		goto out;
 	}
 
 	ret = 0;
@@ -784,33 +701,6 @@ _ctl_do_mpt_command(struct MPT3SAS_ADAPTER *ioc, struct mpt3_ioctl_command karg,
 			mpt3sas_base_put_smid_default(ioc, smid);
 		break;
 	}
-	case MPI2_FUNCTION_SCSI_TASK_MGMT:
-	{
-		Mpi2SCSITaskManagementRequest_t *tm_request =
-		    (Mpi2SCSITaskManagementRequest_t *)request;
-
-		dtmprintk(ioc, pr_info(MPT3SAS_FMT
-			"TASK_MGMT: handle(0x%04x), task_type(0x%02x)\n",
-			ioc->name,
-		    le16_to_cpu(tm_request->DevHandle), tm_request->TaskType));
-
-		if (tm_request->TaskType ==
-		    MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK ||
-		    tm_request->TaskType ==
-		    MPI2_SCSITASKMGMT_TASKTYPE_QUERY_TASK) {
-			if (_ctl_set_task_mid(ioc, &karg, tm_request)) {
-				mpt3sas_base_free_smid(ioc, smid);
-				goto out;
-			}
-		}
-
-		mpt3sas_scsih_set_tm_flag(ioc, le16_to_cpu(
-		    tm_request->DevHandle));
-		ioc->build_sg_mpi(ioc, psge, data_out_dma, data_out_sz,
-		    data_in_dma, data_in_sz);
-		mpt3sas_base_put_smid_hi_priority(ioc, smid);
-		break;
-	}
 	case MPI2_FUNCTION_SMP_PASSTHROUGH:
 	{
 		Mpi2SmpPassthroughRequest_t *smp_request =
-- 
1.9.1

[PATCH 6/6] mpt3sas: store scsi io tracker data in the scsi request

[Jens Axboe]

Instead of storing the IO tracker structure in a separate list
that we need to pop/push to on every submit and complete (and
lock), store it in the pdu associated with a request. This is
possible on scsi-mq only, and further cuts the spinlock associated
time for higher IOPS IO workloads. At 100K IOPS, this effectively
cuts the locking time in half.

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 drivers/scsi/mpt3sas/mpt3sas_base.c  | 173 +++++++++++++++++++++++++----------
 drivers/scsi/mpt3sas/mpt3sas_base.h  |   2 +
 drivers/scsi/mpt3sas/mpt3sas_scsih.c |  69 +++++++++++---
 3 files changed, 185 insertions(+), 59 deletions(-)

diff --git a/drivers/scsi/mpt3sas/mpt3sas_base.c b/drivers/scsi/mpt3sas/mpt3sas_base.c
index 14a781b6b88d..4c67b685e1f5 100644
--- a/drivers/scsi/mpt3sas/mpt3sas_base.c
+++ b/drivers/scsi/mpt3sas/mpt3sas_base.c
@@ -815,6 +815,20 @@ _base_async_event(struct MPT3SAS_ADAPTER *ioc, u8 msix_index, u32 reply)
 	return 1;
 }
 
+struct scsiio_tracker *
+mpt3sas_get_st_from_smid(struct MPT3SAS_ADAPTER *ioc, u16 smid)
+{
+	if (shost_use_blk_mq(ioc->shost)) {
+		struct scsi_cmnd *scmd;
+
+		scmd = scsi_mq_find_tag(ioc->shost, smid - 1);
+		if (!scmd)
+			return NULL;
+		return scsi_cmd_priv(scmd);
+	} else
+		return &ioc->scsi_lookup[smid - 1];
+}
+
 /**
  * _base_get_cb_idx - obtain the callback index
  * @ioc: per adapter object
@@ -829,8 +843,10 @@ _base_get_cb_idx(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 	u8 cb_idx;
 
 	if (smid < ioc->hi_priority_smid) {
-		i = smid - 1;
-		cb_idx = ioc->scsi_lookup[i].cb_idx;
+		struct scsiio_tracker *st;
+
+		st = mpt3sas_get_st_from_smid(ioc, smid);
+		cb_idx = st->cb_idx;
 	} else if (smid < ioc->internal_smid) {
 		i = smid - ioc->hi_priority_smid;
 		cb_idx = ioc->hpr_lookup[i].cb_idx;
@@ -1176,6 +1192,7 @@ static struct chain_tracker *
 _base_get_chain_buffer_tracker(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 {
 	struct chain_tracker *chain_req;
+	struct scsiio_tracker *st;
 	unsigned long flags;
 
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
@@ -1188,8 +1205,8 @@ _base_get_chain_buffer_tracker(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 	chain_req = list_entry(ioc->free_chain_list.next,
 	    struct chain_tracker, tracker_list);
 	list_del_init(&chain_req->tracker_list);
-	list_add_tail(&chain_req->tracker_list,
-	    &ioc->scsi_lookup[smid - 1].chain_list);
+	st = mpt3sas_get_st_from_smid(ioc, smid);
+	list_add_tail(&chain_req->tracker_list, &st->chain_list);
 	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 	return chain_req;
 }
@@ -2006,6 +2023,11 @@ mpt3sas_base_get_smid_scsiio(struct MPT3SAS_ADAPTER *ioc, u8 cb_idx,
 	struct scsiio_tracker *request;
 	u16 smid;
 
+	if (shost_use_blk_mq(ioc->shost)) {
+		request = scsi_cmd_priv(scmd);
+		return request->smid;
+	}
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	if (list_empty(&ioc->free_list)) {
 		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
@@ -2053,6 +2075,32 @@ mpt3sas_base_get_smid_hpr(struct MPT3SAS_ADAPTER *ioc, u8 cb_idx)
 	return smid;
 }
 
+static void
+_base_recovery_check(struct MPT3SAS_ADAPTER *ioc)
+{
+	/*
+	 * See _wait_for_commands_to_complete() call with regards to this code.
+	 */
+	if (ioc->shost_recovery && ioc->pending_io_count) {
+		if (ioc->pending_io_count == 1)
+			wake_up(&ioc->reset_wq);
+		ioc->pending_io_count = 0;
+	}
+}
+
+static void
+_dechain_st(struct MPT3SAS_ADAPTER *ioc, struct scsiio_tracker *st)
+{
+	struct chain_tracker *chain_req;
+
+	while (!list_empty(&st->chain_list)) {
+		chain_req = list_first_entry(&st->chain_list,
+						struct chain_tracker,
+						tracker_list);
+		list_move(&chain_req->tracker_list, &ioc->free_chain_list);
+	}
+}
+
 /**
  * mpt3sas_base_free_smid - put smid back on free_list
  * @ioc: per adapter object
@@ -2065,34 +2113,36 @@ mpt3sas_base_free_smid(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 {
 	unsigned long flags;
 	int i;
-	struct chain_tracker *chain_req, *next;
+
+	if (shost_use_blk_mq(ioc->shost) && smid < ioc->hi_priority_smid) {
+		struct scsiio_tracker *st;
+
+		st = mpt3sas_get_st_from_smid(ioc, smid);
+		if (!st)
+			return;
+
+		if (!list_empty(&st->chain_list)) {
+			spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
+			_dechain_st(ioc, st);
+			spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
+		}
+
+		_base_recovery_check(ioc);
+		return;
+	}
 
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	if (smid < ioc->hi_priority_smid) {
 		/* scsiio queue */
 		i = smid - 1;
-		if (!list_empty(&ioc->scsi_lookup[i].chain_list)) {
-			list_for_each_entry_safe(chain_req, next,
-			    &ioc->scsi_lookup[i].chain_list, tracker_list) {
-				list_del_init(&chain_req->tracker_list);
-				list_add(&chain_req->tracker_list,
-				    &ioc->free_chain_list);
-			}
-		}
+		if (!list_empty(&ioc->scsi_lookup[i].chain_list))
+			_dechain_st(ioc, &ioc->scsi_lookup[i]);
 		ioc->scsi_lookup[i].cb_idx = 0xFF;
 		ioc->scsi_lookup[i].scmd = NULL;
 		list_add(&ioc->scsi_lookup[i].tracker_list, &ioc->free_list);
 		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 
-		/*
-		 * See _wait_for_commands_to_complete() call with regards
-		 * to this code.
-		 */
-		if (ioc->shost_recovery && ioc->pending_io_count) {
-			if (ioc->pending_io_count == 1)
-				wake_up(&ioc->reset_wq);
-			ioc->pending_io_count--;
-		}
+		_base_recovery_check(ioc);
 		return;
 	} else if (smid < ioc->internal_smid) {
 		/* hi-priority */
@@ -2896,14 +2946,23 @@ _base_allocate_memory_pools(struct MPT3SAS_ADAPTER *ioc,  int sleep_flag)
 	    ioc->name, (unsigned long long) ioc->request_dma));
 	total_sz += sz;
 
-	sz = ioc->scsiio_depth * sizeof(struct scsiio_tracker);
-	ioc->scsi_lookup_pages = get_order(sz);
-	ioc->scsi_lookup = (struct scsiio_tracker *)__get_free_pages(
-	    GFP_KERNEL, ioc->scsi_lookup_pages);
-	if (!ioc->scsi_lookup) {
-		pr_err(MPT3SAS_FMT "scsi_lookup: get_free_pages failed, sz(%d)\n",
-			ioc->name, (int)sz);
-		goto out;
+	/*
+	 * Don't need to allocate memory for scsiio_tracker array if we
+	 * are using scsi-mq, we embed it in the scsi_cmnd for that case.
+	 */
+	if (!shost_use_blk_mq(ioc->shost)) {
+		sz = ioc->scsiio_depth * sizeof(struct scsiio_tracker);
+		ioc->scsi_lookup_pages = get_order(sz);
+		ioc->scsi_lookup = (struct scsiio_tracker *)__get_free_pages(
+				    GFP_KERNEL, ioc->scsi_lookup_pages);
+		if (!ioc->scsi_lookup) {
+			pr_err(MPT3SAS_FMT "scsi_lookup: get_free_pages "
+					"failed, sz(%d)\n", ioc->name, (int)sz);
+			goto out;
+		}
+	} else {
+		ioc->scsi_lookup_pages = 0;
+		ioc->scsi_lookup = NULL;
 	}
 
 	dinitprintk(ioc, pr_info(MPT3SAS_FMT "scsiio(0x%p): depth(%d)\n",
@@ -4439,14 +4498,17 @@ _base_make_ioc_operational(struct MPT3SAS_ADAPTER *ioc, int sleep_flag)
 	/* initialize the scsi lookup free list */
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	INIT_LIST_HEAD(&ioc->free_list);
-	smid = 1;
-	for (i = 0; i < ioc->scsiio_depth; i++, smid++) {
-		INIT_LIST_HEAD(&ioc->scsi_lookup[i].chain_list);
-		ioc->scsi_lookup[i].cb_idx = 0xFF;
-		ioc->scsi_lookup[i].smid = smid;
-		ioc->scsi_lookup[i].scmd = NULL;
-		list_add_tail(&ioc->scsi_lookup[i].tracker_list,
-		    &ioc->free_list);
+
+	if (!shost_use_blk_mq(ioc->shost)) {
+		smid = 1;
+		for (i = 0; i < ioc->scsiio_depth; i++, smid++) {
+			INIT_LIST_HEAD(&ioc->scsi_lookup[i].chain_list);
+			ioc->scsi_lookup[i].cb_idx = 0xFF;
+			ioc->scsi_lookup[i].smid = smid;
+			ioc->scsi_lookup[i].scmd = NULL;
+			list_add_tail(&ioc->scsi_lookup[i].tracker_list,
+					    &ioc->free_list);
+		}
 	}
 
 	/* hi-priority queue */
@@ -4895,7 +4957,7 @@ _wait_for_commands_to_complete(struct MPT3SAS_ADAPTER *ioc, int sleep_flag)
 {
 	u32 ioc_state;
 	unsigned long flags;
-	u16 i;
+	u16 i, pending, loops;
 
 	ioc->pending_io_count = 0;
 	if (sleep_flag != CAN_SLEEP)
@@ -4906,17 +4968,32 @@ _wait_for_commands_to_complete(struct MPT3SAS_ADAPTER *ioc, int sleep_flag)
 		return;
 
 	/* pending command count */
-	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
-	for (i = 0; i < ioc->scsiio_depth; i++)
-		if (ioc->scsi_lookup[i].cb_idx != 0xFF)
-			ioc->pending_io_count++;
-	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
+	loops = 0;
+	do {
+		pending = 0;
+		spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
+		for (i = 0; i < ioc->scsiio_depth; i++) {
+			struct scsiio_tracker *st;
+			struct scsi_cmnd *scmd;
+
+			if (shost_use_blk_mq(ioc->shost)) {
+				scmd = scsi_mq_find_tag(ioc->shost, i);
+				if (scmd)
+					pending++;
+			} else {
+				st = mpt3sas_get_st_from_smid(ioc, i + 1);
+				if (st->cb_idx != 0xFF)
+					pending++;
+			}
+		}
+		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 
-	if (!ioc->pending_io_count)
-		return;
+		if (!pending)
+			break;
 
-	/* wait for pending commands to complete */
-	wait_event_timeout(ioc->reset_wq, ioc->pending_io_count == 0, 10 * HZ);
+		/* wait for pending commands to complete */
+		wait_event_timeout(ioc->reset_wq, ioc->pending_io_count == 0, HZ);
+	} while (++loops <= 10);
 }
 
 /**
diff --git a/drivers/scsi/mpt3sas/mpt3sas_base.h b/drivers/scsi/mpt3sas/mpt3sas_base.h
index afa881682bef..1a94791dbfe8 100644
--- a/drivers/scsi/mpt3sas/mpt3sas_base.h
+++ b/drivers/scsi/mpt3sas/mpt3sas_base.h
@@ -986,6 +986,8 @@ u16 mpt3sas_base_get_smid_hpr(struct MPT3SAS_ADAPTER *ioc, u8 cb_idx);
 u16 mpt3sas_base_get_smid_scsiio(struct MPT3SAS_ADAPTER *ioc, u8 cb_idx,
 	struct scsi_cmnd *scmd);
 
+struct scsiio_tracker *mpt3sas_get_st_from_smid(struct MPT3SAS_ADAPTER *ioc,
+    u16 smid);
 u16 mpt3sas_base_get_smid(struct MPT3SAS_ADAPTER *ioc, u8 cb_idx);
 void mpt3sas_base_free_smid(struct MPT3SAS_ADAPTER *ioc, u16 smid);
 void mpt3sas_base_put_smid_scsi_io(struct MPT3SAS_ADAPTER *ioc, u16 smid,
diff --git a/drivers/scsi/mpt3sas/mpt3sas_scsih.c b/drivers/scsi/mpt3sas/mpt3sas_scsih.c
index 5a97e3286719..6b478ae0b680 100644
--- a/drivers/scsi/mpt3sas/mpt3sas_scsih.c
+++ b/drivers/scsi/mpt3sas/mpt3sas_scsih.c
@@ -930,7 +930,10 @@ _scsih_is_end_device(u32 device_info)
 static struct scsi_cmnd *
 _scsih_scsi_lookup_get(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 {
-	return ioc->scsi_lookup[smid - 1].scmd;
+	if (shost_use_blk_mq(ioc->shost))
+		return scsi_mq_find_tag(ioc->shost, smid - 1);
+	else
+		return ioc->scsi_lookup[smid - 1].scmd;
 }
 
 /**
@@ -947,6 +950,8 @@ _scsih_scsi_lookup_get_clear(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 	unsigned long flags;
 	struct scsi_cmnd *scmd;
 
+	BUG_ON(shost_use_blk_mq(ioc->shost));
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	scmd = ioc->scsi_lookup[smid - 1].scmd;
 	ioc->scsi_lookup[smid - 1].scmd = NULL;
@@ -973,6 +978,13 @@ _scsih_scsi_lookup_find_by_scmd(struct MPT3SAS_ADAPTER *ioc, struct scsi_cmnd
 	unsigned long	flags;
 	int i;
 
+	if (shost_use_blk_mq(ioc->shost)) {
+		struct scsiio_tracker *st;
+
+		st = scsi_cmd_priv(scmd);
+		return st->smid;
+	}
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	smid = 0;
 	for (i = 0; i < ioc->scsiio_depth; i++) {
@@ -1007,9 +1019,14 @@ _scsih_scsi_lookup_find_by_target(struct MPT3SAS_ADAPTER *ioc, int id,
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	found = 0;
 	for (i = 0 ; i < ioc->scsiio_depth; i++) {
-		if (ioc->scsi_lookup[i].scmd &&
-		    (ioc->scsi_lookup[i].scmd->device->id == id &&
-		    ioc->scsi_lookup[i].scmd->device->channel == channel)) {
+		struct scsiio_tracker *st;
+
+		st = mpt3sas_get_st_from_smid(ioc, i + 1);
+		if (!st)
+			continue;
+		if (st->scmd &&
+		    (st->scmd->device->id == id &&
+		    st->scmd->device->channel == channel)) {
 			found = 1;
 			goto out;
 		}
@@ -1041,10 +1058,15 @@ _scsih_scsi_lookup_find_by_lun(struct MPT3SAS_ADAPTER *ioc, int id,
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	found = 0;
 	for (i = 0 ; i < ioc->scsiio_depth; i++) {
-		if (ioc->scsi_lookup[i].scmd &&
-		    (ioc->scsi_lookup[i].scmd->device->id == id &&
-		    ioc->scsi_lookup[i].scmd->device->channel == channel &&
-		    ioc->scsi_lookup[i].scmd->device->lun == lun)) {
+		struct scsiio_tracker *st;
+
+		st = mpt3sas_get_st_from_smid(ioc, i + 1);
+		if (!st)
+			continue;
+		if (st->scmd &&
+		    (st->scmd->device->id == id &&
+		    st->scmd->device->channel == channel &&
+		    st->scmd->device->lun == lun)) {
 			found = 1;
 			goto out;
 		}
@@ -2053,7 +2075,7 @@ mpt3sas_scsih_issue_tm(struct MPT3SAS_ADAPTER *ioc, u16 handle, uint channel,
 	}
 
 	if (type == MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK)
-		scsi_lookup = &ioc->scsi_lookup[smid_task - 1];
+		scsi_lookup = mpt3sas_get_st_from_smid(ioc, smid_task);
 
 	dtmprintk(ioc, pr_info(MPT3SAS_FMT
 		"sending tm: handle(0x%04x), task_type(0x%02x), smid(%d)\n",
@@ -3392,7 +3414,11 @@ _scsih_flush_running_cmds(struct MPT3SAS_ADAPTER *ioc)
 	u16 count = 0;
 
 	for (smid = 1; smid <= ioc->scsiio_depth; smid++) {
-		scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+		if (shost_use_blk_mq(ioc->shost))
+			scmd = _scsih_scsi_lookup_get(ioc, smid);
+		else
+			scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+
 		if (!scmd)
 			continue;
 		count++;
@@ -4058,7 +4084,11 @@ _scsih_io_done(struct MPT3SAS_ADAPTER *ioc, u16 smid, u8 msix_index, u32 reply)
 	u32 response_code = 0;
 
 	mpi_reply = mpt3sas_base_get_reply_virt_addr(ioc, reply);
-	scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+	if (shost_use_blk_mq(ioc->shost))
+		scmd = scsi_mq_find_tag(ioc->shost, smid - 1);
+	else
+		scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+
 	if (scmd == NULL)
 		return 1;
 
@@ -7255,6 +7285,21 @@ mpt3sas_scsih_event_callback(struct MPT3SAS_ADAPTER *ioc, u8 msix_index,
 	return 1;
 }
 
+static int
+_scsih_init_command(struct Scsi_Host *shost, struct scsi_cmnd *cmd,
+			unsigned int request_idx)
+{
+	struct MPT3SAS_ADAPTER *ioc = shost_priv(shost);
+	struct scsiio_tracker *st;
+
+	st = (void *) cmd + sizeof(*cmd);
+	INIT_LIST_HEAD(&st->chain_list);
+	st->scmd = cmd;
+	st->cb_idx = ioc->scsi_io_cb_idx;
+	st->smid = request_idx + 1;
+	return 0;
+}
+
 /* shost template */
 static struct scsi_host_template scsih_driver_template = {
 	.module				= THIS_MODULE,
@@ -7283,6 +7328,8 @@ static struct scsi_host_template scsih_driver_template = {
 	.shost_attrs			= mpt3sas_host_attrs,
 	.sdev_attrs			= mpt3sas_dev_attrs,
 	.track_queue_depth		= 1,
+	.cmd_size			= sizeof(struct scsiio_tracker),
+	.init_command			= _scsih_init_command,
 };
 
 /**
-- 
1.9.1

[PATCH 6/6] mpt3sas: store scsi io tracker data in the scsi request

[Jens Axboe]

Instead of storing the IO tracker structure in a separate list
that we need to pop/push to on every submit and complete (and
lock), store it in the pdu associated with a request. This is
possible on scsi-mq only, and further cuts the spinlock associated
time for higher IOPS IO workloads. At 100K IOPS, this effectively
cuts the locking time in half.

Signed-off-by: Jens Axboe <axboe@fb.com>
---
 drivers/scsi/mpt3sas/mpt3sas_base.c  | 173 +++++++++++++++++++++++++----------
 drivers/scsi/mpt3sas/mpt3sas_base.h  |   2 +
 drivers/scsi/mpt3sas/mpt3sas_scsih.c |  69 +++++++++++---
 3 files changed, 185 insertions(+), 59 deletions(-)

diff --git a/drivers/scsi/mpt3sas/mpt3sas_base.c b/drivers/scsi/mpt3sas/mpt3sas_base.c
index 14a781b6b88d..4c67b685e1f5 100644
--- a/drivers/scsi/mpt3sas/mpt3sas_base.c
+++ b/drivers/scsi/mpt3sas/mpt3sas_base.c
@@ -815,6 +815,20 @@ _base_async_event(struct MPT3SAS_ADAPTER *ioc, u8 msix_index, u32 reply)
 	return 1;
 }
 
+struct scsiio_tracker *
+mpt3sas_get_st_from_smid(struct MPT3SAS_ADAPTER *ioc, u16 smid)
+{
+	if (shost_use_blk_mq(ioc->shost)) {
+		struct scsi_cmnd *scmd;
+
+		scmd = scsi_mq_find_tag(ioc->shost, smid - 1);
+		if (!scmd)
+			return NULL;
+		return scsi_cmd_priv(scmd);
+	} else
+		return &ioc->scsi_lookup[smid - 1];
+}
+
 /**
  * _base_get_cb_idx - obtain the callback index
  * @ioc: per adapter object
@@ -829,8 +843,10 @@ _base_get_cb_idx(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 	u8 cb_idx;
 
 	if (smid < ioc->hi_priority_smid) {
-		i = smid - 1;
-		cb_idx = ioc->scsi_lookup[i].cb_idx;
+		struct scsiio_tracker *st;
+
+		st = mpt3sas_get_st_from_smid(ioc, smid);
+		cb_idx = st->cb_idx;
 	} else if (smid < ioc->internal_smid) {
 		i = smid - ioc->hi_priority_smid;
 		cb_idx = ioc->hpr_lookup[i].cb_idx;
@@ -1176,6 +1192,7 @@ static struct chain_tracker *
 _base_get_chain_buffer_tracker(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 {
 	struct chain_tracker *chain_req;
+	struct scsiio_tracker *st;
 	unsigned long flags;
 
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
@@ -1188,8 +1205,8 @@ _base_get_chain_buffer_tracker(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 	chain_req = list_entry(ioc->free_chain_list.next,
 	    struct chain_tracker, tracker_list);
 	list_del_init(&chain_req->tracker_list);
-	list_add_tail(&chain_req->tracker_list,
-	    &ioc->scsi_lookup[smid - 1].chain_list);
+	st = mpt3sas_get_st_from_smid(ioc, smid);
+	list_add_tail(&chain_req->tracker_list, &st->chain_list);
 	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 	return chain_req;
 }
@@ -2006,6 +2023,11 @@ mpt3sas_base_get_smid_scsiio(struct MPT3SAS_ADAPTER *ioc, u8 cb_idx,
 	struct scsiio_tracker *request;
 	u16 smid;
 
+	if (shost_use_blk_mq(ioc->shost)) {
+		request = scsi_cmd_priv(scmd);
+		return request->smid;
+	}
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	if (list_empty(&ioc->free_list)) {
 		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
@@ -2053,6 +2075,32 @@ mpt3sas_base_get_smid_hpr(struct MPT3SAS_ADAPTER *ioc, u8 cb_idx)
 	return smid;
 }
 
+static void
+_base_recovery_check(struct MPT3SAS_ADAPTER *ioc)
+{
+	/*
+	 * See _wait_for_commands_to_complete() call with regards to this code.
+	 */
+	if (ioc->shost_recovery && ioc->pending_io_count) {
+		if (ioc->pending_io_count == 1)
+			wake_up(&ioc->reset_wq);
+		ioc->pending_io_count = 0;
+	}
+}
+
+static void
+_dechain_st(struct MPT3SAS_ADAPTER *ioc, struct scsiio_tracker *st)
+{
+	struct chain_tracker *chain_req;
+
+	while (!list_empty(&st->chain_list)) {
+		chain_req = list_first_entry(&st->chain_list,
+						struct chain_tracker,
+						tracker_list);
+		list_move(&chain_req->tracker_list, &ioc->free_chain_list);
+	}
+}
+
 /**
  * mpt3sas_base_free_smid - put smid back on free_list
  * @ioc: per adapter object
@@ -2065,34 +2113,36 @@ mpt3sas_base_free_smid(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 {
 	unsigned long flags;
 	int i;
-	struct chain_tracker *chain_req, *next;
+
+	if (shost_use_blk_mq(ioc->shost) && smid < ioc->hi_priority_smid) {
+		struct scsiio_tracker *st;
+
+		st = mpt3sas_get_st_from_smid(ioc, smid);
+		if (!st)
+			return;
+
+		if (!list_empty(&st->chain_list)) {
+			spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
+			_dechain_st(ioc, st);
+			spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
+		}
+
+		_base_recovery_check(ioc);
+		return;
+	}
 
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	if (smid < ioc->hi_priority_smid) {
 		/* scsiio queue */
 		i = smid - 1;
-		if (!list_empty(&ioc->scsi_lookup[i].chain_list)) {
-			list_for_each_entry_safe(chain_req, next,
-			    &ioc->scsi_lookup[i].chain_list, tracker_list) {
-				list_del_init(&chain_req->tracker_list);
-				list_add(&chain_req->tracker_list,
-				    &ioc->free_chain_list);
-			}
-		}
+		if (!list_empty(&ioc->scsi_lookup[i].chain_list))
+			_dechain_st(ioc, &ioc->scsi_lookup[i]);
 		ioc->scsi_lookup[i].cb_idx = 0xFF;
 		ioc->scsi_lookup[i].scmd = NULL;
 		list_add(&ioc->scsi_lookup[i].tracker_list, &ioc->free_list);
 		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 
-		/*
-		 * See _wait_for_commands_to_complete() call with regards
-		 * to this code.
-		 */
-		if (ioc->shost_recovery && ioc->pending_io_count) {
-			if (ioc->pending_io_count == 1)
-				wake_up(&ioc->reset_wq);
-			ioc->pending_io_count--;
-		}
+		_base_recovery_check(ioc);
 		return;
 	} else if (smid < ioc->internal_smid) {
 		/* hi-priority */
@@ -2896,14 +2946,23 @@ _base_allocate_memory_pools(struct MPT3SAS_ADAPTER *ioc,  int sleep_flag)
 	    ioc->name, (unsigned long long) ioc->request_dma));
 	total_sz += sz;
 
-	sz = ioc->scsiio_depth * sizeof(struct scsiio_tracker);
-	ioc->scsi_lookup_pages = get_order(sz);
-	ioc->scsi_lookup = (struct scsiio_tracker *)__get_free_pages(
-	    GFP_KERNEL, ioc->scsi_lookup_pages);
-	if (!ioc->scsi_lookup) {
-		pr_err(MPT3SAS_FMT "scsi_lookup: get_free_pages failed, sz(%d)\n",
-			ioc->name, (int)sz);
-		goto out;
+	/*
+	 * Don't need to allocate memory for scsiio_tracker array if we
+	 * are using scsi-mq, we embed it in the scsi_cmnd for that case.
+	 */
+	if (!shost_use_blk_mq(ioc->shost)) {
+		sz = ioc->scsiio_depth * sizeof(struct scsiio_tracker);
+		ioc->scsi_lookup_pages = get_order(sz);
+		ioc->scsi_lookup = (struct scsiio_tracker *)__get_free_pages(
+				    GFP_KERNEL, ioc->scsi_lookup_pages);
+		if (!ioc->scsi_lookup) {
+			pr_err(MPT3SAS_FMT "scsi_lookup: get_free_pages "
+					"failed, sz(%d)\n", ioc->name, (int)sz);
+			goto out;
+		}
+	} else {
+		ioc->scsi_lookup_pages = 0;
+		ioc->scsi_lookup = NULL;
 	}
 
 	dinitprintk(ioc, pr_info(MPT3SAS_FMT "scsiio(0x%p): depth(%d)\n",
@@ -4439,14 +4498,17 @@ _base_make_ioc_operational(struct MPT3SAS_ADAPTER *ioc, int sleep_flag)
 	/* initialize the scsi lookup free list */
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	INIT_LIST_HEAD(&ioc->free_list);
-	smid = 1;
-	for (i = 0; i < ioc->scsiio_depth; i++, smid++) {
-		INIT_LIST_HEAD(&ioc->scsi_lookup[i].chain_list);
-		ioc->scsi_lookup[i].cb_idx = 0xFF;
-		ioc->scsi_lookup[i].smid = smid;
-		ioc->scsi_lookup[i].scmd = NULL;
-		list_add_tail(&ioc->scsi_lookup[i].tracker_list,
-		    &ioc->free_list);
+
+	if (!shost_use_blk_mq(ioc->shost)) {
+		smid = 1;
+		for (i = 0; i < ioc->scsiio_depth; i++, smid++) {
+			INIT_LIST_HEAD(&ioc->scsi_lookup[i].chain_list);
+			ioc->scsi_lookup[i].cb_idx = 0xFF;
+			ioc->scsi_lookup[i].smid = smid;
+			ioc->scsi_lookup[i].scmd = NULL;
+			list_add_tail(&ioc->scsi_lookup[i].tracker_list,
+					    &ioc->free_list);
+		}
 	}
 
 	/* hi-priority queue */
@@ -4895,7 +4957,7 @@ _wait_for_commands_to_complete(struct MPT3SAS_ADAPTER *ioc, int sleep_flag)
 {
 	u32 ioc_state;
 	unsigned long flags;
-	u16 i;
+	u16 i, pending, loops;
 
 	ioc->pending_io_count = 0;
 	if (sleep_flag != CAN_SLEEP)
@@ -4906,17 +4968,32 @@ _wait_for_commands_to_complete(struct MPT3SAS_ADAPTER *ioc, int sleep_flag)
 		return;
 
 	/* pending command count */
-	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
-	for (i = 0; i < ioc->scsiio_depth; i++)
-		if (ioc->scsi_lookup[i].cb_idx != 0xFF)
-			ioc->pending_io_count++;
-	spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
+	loops = 0;
+	do {
+		pending = 0;
+		spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
+		for (i = 0; i < ioc->scsiio_depth; i++) {
+			struct scsiio_tracker *st;
+			struct scsi_cmnd *scmd;
+
+			if (shost_use_blk_mq(ioc->shost)) {
+				scmd = scsi_mq_find_tag(ioc->shost, i);
+				if (scmd)
+					pending++;
+			} else {
+				st = mpt3sas_get_st_from_smid(ioc, i + 1);
+				if (st->cb_idx != 0xFF)
+					pending++;
+			}
+		}
+		spin_unlock_irqrestore(&ioc->scsi_lookup_lock, flags);
 
-	if (!ioc->pending_io_count)
-		return;
+		if (!pending)
+			break;
 
-	/* wait for pending commands to complete */
-	wait_event_timeout(ioc->reset_wq, ioc->pending_io_count == 0, 10 * HZ);
+		/* wait for pending commands to complete */
+		wait_event_timeout(ioc->reset_wq, ioc->pending_io_count == 0, HZ);
+	} while (++loops <= 10);
 }
 
 /**
diff --git a/drivers/scsi/mpt3sas/mpt3sas_base.h b/drivers/scsi/mpt3sas/mpt3sas_base.h
index afa881682bef..1a94791dbfe8 100644
--- a/drivers/scsi/mpt3sas/mpt3sas_base.h
+++ b/drivers/scsi/mpt3sas/mpt3sas_base.h
@@ -986,6 +986,8 @@ u16 mpt3sas_base_get_smid_hpr(struct MPT3SAS_ADAPTER *ioc, u8 cb_idx);
 u16 mpt3sas_base_get_smid_scsiio(struct MPT3SAS_ADAPTER *ioc, u8 cb_idx,
 	struct scsi_cmnd *scmd);
 
+struct scsiio_tracker *mpt3sas_get_st_from_smid(struct MPT3SAS_ADAPTER *ioc,
+    u16 smid);
 u16 mpt3sas_base_get_smid(struct MPT3SAS_ADAPTER *ioc, u8 cb_idx);
 void mpt3sas_base_free_smid(struct MPT3SAS_ADAPTER *ioc, u16 smid);
 void mpt3sas_base_put_smid_scsi_io(struct MPT3SAS_ADAPTER *ioc, u16 smid,
diff --git a/drivers/scsi/mpt3sas/mpt3sas_scsih.c b/drivers/scsi/mpt3sas/mpt3sas_scsih.c
index 5a97e3286719..6b478ae0b680 100644
--- a/drivers/scsi/mpt3sas/mpt3sas_scsih.c
+++ b/drivers/scsi/mpt3sas/mpt3sas_scsih.c
@@ -930,7 +930,10 @@ _scsih_is_end_device(u32 device_info)
 static struct scsi_cmnd *
 _scsih_scsi_lookup_get(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 {
-	return ioc->scsi_lookup[smid - 1].scmd;
+	if (shost_use_blk_mq(ioc->shost))
+		return scsi_mq_find_tag(ioc->shost, smid - 1);
+	else
+		return ioc->scsi_lookup[smid - 1].scmd;
 }
 
 /**
@@ -947,6 +950,8 @@ _scsih_scsi_lookup_get_clear(struct MPT3SAS_ADAPTER *ioc, u16 smid)
 	unsigned long flags;
 	struct scsi_cmnd *scmd;
 
+	BUG_ON(shost_use_blk_mq(ioc->shost));
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	scmd = ioc->scsi_lookup[smid - 1].scmd;
 	ioc->scsi_lookup[smid - 1].scmd = NULL;
@@ -973,6 +978,13 @@ _scsih_scsi_lookup_find_by_scmd(struct MPT3SAS_ADAPTER *ioc, struct scsi_cmnd
 	unsigned long	flags;
 	int i;
 
+	if (shost_use_blk_mq(ioc->shost)) {
+		struct scsiio_tracker *st;
+
+		st = scsi_cmd_priv(scmd);
+		return st->smid;
+	}
+
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	smid = 0;
 	for (i = 0; i < ioc->scsiio_depth; i++) {
@@ -1007,9 +1019,14 @@ _scsih_scsi_lookup_find_by_target(struct MPT3SAS_ADAPTER *ioc, int id,
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	found = 0;
 	for (i = 0 ; i < ioc->scsiio_depth; i++) {
-		if (ioc->scsi_lookup[i].scmd &&
-		    (ioc->scsi_lookup[i].scmd->device->id == id &&
-		    ioc->scsi_lookup[i].scmd->device->channel == channel)) {
+		struct scsiio_tracker *st;
+
+		st = mpt3sas_get_st_from_smid(ioc, i + 1);
+		if (!st)
+			continue;
+		if (st->scmd &&
+		    (st->scmd->device->id == id &&
+		    st->scmd->device->channel == channel)) {
 			found = 1;
 			goto out;
 		}
@@ -1041,10 +1058,15 @@ _scsih_scsi_lookup_find_by_lun(struct MPT3SAS_ADAPTER *ioc, int id,
 	spin_lock_irqsave(&ioc->scsi_lookup_lock, flags);
 	found = 0;
 	for (i = 0 ; i < ioc->scsiio_depth; i++) {
-		if (ioc->scsi_lookup[i].scmd &&
-		    (ioc->scsi_lookup[i].scmd->device->id == id &&
-		    ioc->scsi_lookup[i].scmd->device->channel == channel &&
-		    ioc->scsi_lookup[i].scmd->device->lun == lun)) {
+		struct scsiio_tracker *st;
+
+		st = mpt3sas_get_st_from_smid(ioc, i + 1);
+		if (!st)
+			continue;
+		if (st->scmd &&
+		    (st->scmd->device->id == id &&
+		    st->scmd->device->channel == channel &&
+		    st->scmd->device->lun == lun)) {
 			found = 1;
 			goto out;
 		}
@@ -2053,7 +2075,7 @@ mpt3sas_scsih_issue_tm(struct MPT3SAS_ADAPTER *ioc, u16 handle, uint channel,
 	}
 
 	if (type == MPI2_SCSITASKMGMT_TASKTYPE_ABORT_TASK)
-		scsi_lookup = &ioc->scsi_lookup[smid_task - 1];
+		scsi_lookup = mpt3sas_get_st_from_smid(ioc, smid_task);
 
 	dtmprintk(ioc, pr_info(MPT3SAS_FMT
 		"sending tm: handle(0x%04x), task_type(0x%02x), smid(%d)\n",
@@ -3392,7 +3414,11 @@ _scsih_flush_running_cmds(struct MPT3SAS_ADAPTER *ioc)
 	u16 count = 0;
 
 	for (smid = 1; smid <= ioc->scsiio_depth; smid++) {
-		scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+		if (shost_use_blk_mq(ioc->shost))
+			scmd = _scsih_scsi_lookup_get(ioc, smid);
+		else
+			scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+
 		if (!scmd)
 			continue;
 		count++;
@@ -4058,7 +4084,11 @@ _scsih_io_done(struct MPT3SAS_ADAPTER *ioc, u16 smid, u8 msix_index, u32 reply)
 	u32 response_code = 0;
 
 	mpi_reply = mpt3sas_base_get_reply_virt_addr(ioc, reply);
-	scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+	if (shost_use_blk_mq(ioc->shost))
+		scmd = scsi_mq_find_tag(ioc->shost, smid - 1);
+	else
+		scmd = _scsih_scsi_lookup_get_clear(ioc, smid);
+
 	if (scmd == NULL)
 		return 1;
 
@@ -7255,6 +7285,21 @@ mpt3sas_scsih_event_callback(struct MPT3SAS_ADAPTER *ioc, u8 msix_index,
 	return 1;
 }
 
+static int
+_scsih_init_command(struct Scsi_Host *shost, struct scsi_cmnd *cmd,
+			unsigned int request_idx)
+{
+	struct MPT3SAS_ADAPTER *ioc = shost_priv(shost);
+	struct scsiio_tracker *st;
+
+	st = (void *) cmd + sizeof(*cmd);
+	INIT_LIST_HEAD(&st->chain_list);
+	st->scmd = cmd;
+	st->cb_idx = ioc->scsi_io_cb_idx;
+	st->smid = request_idx + 1;
+	return 0;
+}
+
 /* shost template */
 static struct scsi_host_template scsih_driver_template = {
 	.module				= THIS_MODULE,
@@ -7283,6 +7328,8 @@ static struct scsi_host_template scsih_driver_template = {
 	.shost_attrs			= mpt3sas_host_attrs,
 	.sdev_attrs			= mpt3sas_dev_attrs,
 	.track_queue_depth		= 1,
+	.cmd_size			= sizeof(struct scsiio_tracker),
+	.init_command			= _scsih_init_command,
 };
 
 /**
-- 
1.9.1

Re: [PATCH RFC v2] mpt2/mpt3sas lock reduction for scsi-mq

[Christoph Hellwig]

This looks good modulo my minor comments and using and testing the
new code path for the !mq case as well.  Please also Cc the Avago
people on your next submission.