Skip to contentSkip to conversation

aacraid: reply queue mapping to CPUs based of IRQ affinity

PATCH Applied Sagar Biradar 6 participants 7 messages May 19, 2023 – Nov 24, 2023

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v4 1/1 aacraid: reply queue mapping to CPUs based of IRQ affinity

Fix the IO hang that arises because of MSIx vector not having a mapped online CPU upon receiving completion.

The SCSI cmds take the blk_mq route, which is setup during the init. The reserved cmds fetch the vector_no from mq_map after the init is complete and before the init, they use 0 - as per the norm.

---
 drivers/scsi/aacraid/aacraid.h  |  1 +
 drivers/scsi/aacraid/comminit.c |  1 -
 drivers/scsi/aacraid/commsup.c  |  6 +++++-
 drivers/scsi/aacraid/linit.c    | 14 ++++++++++++++
 drivers/scsi/aacraid/src.c      | 25 +++++++++++++++++++++++--
 5 files changed, 43 insertions(+), 4 deletions(-)
drivers/scsi/aacraid/aacraid.h +1 -0
index 5e115e8b2ba4..7c6efde75da6 100644
@@ -1678,6 +1678,7 @@ struct aac_dev
 	u32			handle_pci_error;
 	bool			init_reset;
 	u8			soft_reset_support;
+	u8			use_map_queue;
 };
 
 #define aac_adapter_interrupt(dev) \
drivers/scsi/aacraid/comminit.c +0 -1
index bd99c5492b7d..a5483e7e283a 100644
@@ -657,4 +657,3 @@ struct aac_dev *aac_init_adapter(struct aac_dev *dev)
 
 	return dev;
 }
-
drivers/scsi/aacraid/commsup.c +5 -1
index deb32c9f4b3e..3f062e4013ab 100644
@@ -223,8 +223,12 @@ int aac_fib_setup(struct aac_dev * dev)
 struct fib *aac_fib_alloc_tag(struct aac_dev *dev, struct scsi_cmnd *scmd)
 {
 	struct fib *fibptr;
+	u32 blk_tag;
+	int i;
 
-	fibptr = &dev->fibs[scsi_cmd_to_rq(scmd)->tag];
+	blk_tag = blk_mq_unique_tag(scsi_cmd_to_rq(scmd));
+	i = blk_mq_unique_tag_to_tag(blk_tag);
+	fibptr = &dev->fibs[i];
 	/*
 	 *	Null out fields that depend on being zero at the start of
 	 *	each I/O
drivers/scsi/aacraid/linit.c +14 -0
index 5ba5c18b77b4..9caf8c314ce1 100644
@@ -19,6 +19,7 @@
 
 #include <linux/compat.h>
 #include <linux/blkdev.h>
+#include <linux/blk-mq-pci.h>
 #include <linux/completion.h>
 #include <linux/init.h>
 #include <linux/interrupt.h>
@@ -505,6 +506,15 @@ static int aac_slave_configure(struct scsi_device *sdev)
 	return 0;
 }
 
+static void aac_map_queues(struct Scsi_Host *shost)
+{
+	struct aac_dev *aac = (struct aac_dev *)shost->hostdata;
+
+	blk_mq_pci_map_queues(&shost->tag_set.map[HCTX_TYPE_DEFAULT],
+				aac->pdev, 0);
+	aac->use_map_queue = true;
+}
+
 /**
  *	aac_change_queue_depth		-	alter queue depths
  *	@sdev:	SCSI device we are considering
@@ -1489,6 +1499,7 @@ static struct scsi_host_template aac_driver_template = {
 	.bios_param			= aac_biosparm,
 	.shost_groups			= aac_host_groups,
 	.slave_configure		= aac_slave_configure,
+	.map_queues			= aac_map_queues,
 	.change_queue_depth		= aac_change_queue_depth,
 	.sdev_groups			= aac_dev_groups,
 	.eh_abort_handler		= aac_eh_abort,
@@ -1776,6 +1787,8 @@ static int aac_probe_one(struct pci_dev *pdev, const struct pci_device_id *id)
 	shost->max_lun = AAC_MAX_LUN;
 
 	pci_set_drvdata(pdev, shost);
+	shost->nr_hw_queues = aac->max_msix;
+	shost->host_tagset = 1;
 
 	error = scsi_add_host(shost, &pdev->dev);
 	if (error)
@@ -1908,6 +1921,7 @@ static void aac_remove_one(struct pci_dev *pdev)
 	struct aac_dev *aac = (struct aac_dev *)shost->hostdata;
 
 	aac_cancel_rescan_worker(aac);
+	aac->use_map_queue = false;
 	scsi_remove_host(shost);
 
 	__aac_shutdown(aac);
drivers/scsi/aacraid/src.c +23 -2
index 11ef58204e96..61949f374188 100644
@@ -493,6 +493,10 @@ static int aac_src_deliver_message(struct fib *fib)
 #endif
 
 	u16 vector_no;
+	struct scsi_cmnd *scmd;
+	u32 blk_tag;
+	struct Scsi_Host *shost = dev->scsi_host_ptr;
+	struct blk_mq_queue_map *qmap;
 
 	atomic_inc(&q->numpending);
 
@@ -505,8 +509,25 @@ static int aac_src_deliver_message(struct fib *fib)
 		if ((dev->comm_interface == AAC_COMM_MESSAGE_TYPE3)
 			&& dev->sa_firmware)
 			vector_no = aac_get_vector(dev);
-		else
-			vector_no = fib->vector_no;
+		else {
+			if (!fib->vector_no || !fib->callback_data) {
+				if (shost && dev->use_map_queue) {
+					qmap = &shost->tag_set.map[HCTX_TYPE_DEFAULT];
+					vector_no = qmap->mq_map[raw_smp_processor_id()];
+				}
+				/*
+				 *	We hardcode the vector_no for
+				 *	reserved commands as a valid shost is
+				 *	absent during the init
+				 */
+				else
+					vector_no = 0;
+			} else {
+				scmd = (struct scsi_cmnd *)fib->callback_data;
+				blk_tag = blk_mq_unique_tag(scsi_cmd_to_rq(scmd));
+				vector_no = blk_mq_unique_tag_to_hwq(blk_tag);
+			}
+		}
 
 		if (native_hba) {
 			if (fib->flags & FIB_CONTEXT_FLAG_NATIVE_HBA_TMF) {
Signature · 1 line
2.29.0
#1 lore raw
FYI: kernel test robot notices the stable kernel rule is not satisfied.

Hi,

Thanks for your patch.

FYI: kernel test robot notices the stable kernel rule is not satisfied.

Rule: 'Cc: [email protected]' or 'commit <sha1> upstream.'
Subject: [PATCH v4] aacraid: reply queue mapping to CPUs based of IRQ affinity

The check is based on https://www.kernel.org/doc/html/latest/process/stable-kernel-rules.html

Signature · 2 lines
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests
author Reviewed-by Gilbert Wu on v4 1/1
#2 lore raw
In reply to Sagar Biradar

Hi James et al.
Since there were no activity/comments - I pinged to check if this patch was in line to be accepted/merged. I have addressed the comments on the earlier versions of the patch.

Thanks
-----Original Message-----
From: Sagar Biradar <[email protected]>
Sent: Friday, May 19, 2023 4:09 PM
To: Don Brace - C33706 <[email protected]>; Sagar Biradar - C34249 <[email protected]>; Gilbert Wu - C33504 <[email protected]>; [email protected]; Martin Petersen <[email protected]>; James Bottomley <[email protected]>; Brian King <[email protected]>; [email protected]; Tom White - C33503 <[email protected]>
Subject: [PATCH v4] aacraid: reply queue mapping to CPUs based of IRQ affinity

Fix the IO hang that arises because of MSIx vector not having a mapped online CPU upon receiving completion.

The SCSI cmds take the blk_mq route, which is setup during the init.
The reserved cmds fetch the vector_no from mq_map after the init is complete and before the init, they use 0 - as per the norm.

---
 drivers/scsi/aacraid/aacraid.h  |  1 +
 drivers/scsi/aacraid/comminit.c |  1 -
 drivers/scsi/aacraid/commsup.c  |  6 +++++-
 drivers/scsi/aacraid/linit.c    | 14 ++++++++++++++
 drivers/scsi/aacraid/src.c      | 25 +++++++++++++++++++++++--
 5 files changed, 43 insertions(+), 4 deletions(-)
drivers/scsi/aacraid/aacraid.h +1 -0
@@ -1678,6 +1678,7 @@ struct aac_dev
 	u32			handle_pci_error;
 	bool			init_reset;
 	u8			soft_reset_support;
+	u8			use_map_queue;
 };
 
 #define aac_adapter_interrupt(dev) \
drivers/scsi/aacraid/comminit.c +0 -1
@@ -657,4 +657,3 @@ struct aac_dev *aac_init_adapter(struct aac_dev *dev)
 
 	return dev;
 }
-
drivers/scsi/aacraid/commsup.c +5 -1
@@ -223,8 +223,12 @@ int aac_fib_setup(struct aac_dev * dev)  struct fib *aac_fib_alloc_tag(struct aac_dev *dev, struct scsi_cmnd *scmd)  {
 	struct fib *fibptr;
+	u32 blk_tag;
+	int i;
 
-	fibptr = &dev->fibs[scsi_cmd_to_rq(scmd)->tag];
+	blk_tag = blk_mq_unique_tag(scsi_cmd_to_rq(scmd));
+	i = blk_mq_unique_tag_to_tag(blk_tag);
+	fibptr = &dev->fibs[i];
 	/*
 	 *	Null out fields that depend on being zero at the start of
 	 *	each I/O
drivers/scsi/aacraid/linit.c +13 -0
@@ -19,6 +19,7 @@
 
 #include <linux/compat.h>
 #include <linux/blkdev.h>
+#include <linux/blk-mq-pci.h>
 #include <linux/completion.h>
 #include <linux/init.h>
 #include <linux/interrupt.h>
@@ -505,6 +506,15 @@ static int aac_slave_configure(struct scsi_device *sdev)
 	return 0;
 }
 
+static void aac_map_queues(struct Scsi_Host *shost) {
+	struct aac_dev *aac = (struct aac_dev *)shost->hostdata;
+
+	blk_mq_pci_map_queues(&shost->tag_set.map[HCTX_TYPE_DEFAULT],
+				aac->pdev, 0);
+	aac->use_map_queue = true;
+}
+
 /**
  *	aac_change_queue_depth		-	alter queue depths
  *	@sdev:	SCSI device we are considering
@@ -1489,6 +1499,7 @@ static struct scsi_host_template aac_driver_template = {
 	.bios_param			= aac_biosparm,
 	.shost_groups			= aac_host_groups,
 	.slave_configure		= aac_slave_configure,
+	.map_queues			= aac_map_queues,
 	.change_queue_depth		= aac_change_queue_depth,
 	.sdev_groups			= aac_dev_groups,
 	.eh_abort_handler		= aac_eh_abort,
@@ -1776,6 +1787,8 @@ static int aac_probe_one(struct pci_dev *pdev, const struct pci_device_id *id)
 	shost->max_lun = AAC_MAX_LUN;
 
 	pci_set_drvdata(pdev, shost);
+	shost->nr_hw_queues = aac->max_msix;
+	shost->host_tagset = 1;
 
 	error = scsi_add_host(shost, &pdev->dev);
 	if (error)
@@ -1908,6 +1921,7 @@ static void aac_remove_one(struct pci_dev *pdev)
 	struct aac_dev *aac = (struct aac_dev *)shost->hostdata;
 
 	aac_cancel_rescan_worker(aac);
+	aac->use_map_queue = false;
 	scsi_remove_host(shost);
 
 	__aac_shutdown(aac);
drivers/scsi/aacraid/src.c +23 -2
@@ -493,6 +493,10 @@ static int aac_src_deliver_message(struct fib *fib)  #endif
 
 	u16 vector_no;
+	struct scsi_cmnd *scmd;
+	u32 blk_tag;
+	struct Scsi_Host *shost = dev->scsi_host_ptr;
+	struct blk_mq_queue_map *qmap;
 
 	atomic_inc(&q->numpending);
 
@@ -505,8 +509,25 @@ static int aac_src_deliver_message(struct fib *fib)
 		if ((dev->comm_interface == AAC_COMM_MESSAGE_TYPE3)
 			&& dev->sa_firmware)
 			vector_no = aac_get_vector(dev);
-		else
-			vector_no = fib->vector_no;
+		else {
+			if (!fib->vector_no || !fib->callback_data) {
+				if (shost && dev->use_map_queue) {
+					qmap = &shost->tag_set.map[HCTX_TYPE_DEFAULT];
+					vector_no = qmap->mq_map[raw_smp_processor_id()];
+				}
+				/*
+				 *	We hardcode the vector_no for
+				 *	reserved commands as a valid shost is
+				 *	absent during the init
+				 */
+				else
+					vector_no = 0;
+			} else {
+				scmd = (struct scsi_cmnd *)fib->callback_data;
+				blk_tag = blk_mq_unique_tag(scsi_cmd_to_rq(scmd));
+				vector_no = blk_mq_unique_tag_to_hwq(blk_tag);
+			}
+		}
 
 		if (native_hba) {
 			if (fib->flags & FIB_CONTEXT_FLAG_NATIVE_HBA_TMF) {

--
2.29.0

#3 lore raw
In reply to Sagar Biradar
Quoted 8 lines · Sagar Biradar

Fix the IO hang that arises because of MSIx vector not having a mapped online CPU upon receiving completion.

The SCSI cmds take the blk_mq route, which is setup during the init. The reserved cmds fetch the vector_no from mq_map after the init is complete and before the init, they use 0 - as per the norm.

[...]

Applied to 6.4/scsi-fixes, thanks!

[1/1] aacraid: reply queue mapping to CPUs based of IRQ affinity

      https://git.kernel.org/mkp/scsi/c/9dc704dcc09e
Signature · 1 line
Martin K. Petersen	Oracle Linux Engineering
#4 lore raw
In reply to Sagar Biradar
Quoted 9 lines · Sagar Biradar

Fix the IO hang that arises because of MSIx vector not having a mapped online CPU upon receiving completion.

The SCSI cmds take the blk_mq route, which is setup during the init. The reserved cmds fetch the vector_no from mq_map after the init is complete and before the init, they use 0 - as per the norm.

This the patch which seems to be causing the issue in https://bugzilla.kernel.org/show_bug.cgi?id=217599

I will comment here since I got no response there...

Quoted 128 lines
---
  drivers/scsi/aacraid/aacraid.h  |  1 +
  drivers/scsi/aacraid/comminit.c |  1 -
  drivers/scsi/aacraid/commsup.c  |  6 +++++-
  drivers/scsi/aacraid/linit.c    | 14 ++++++++++++++
  drivers/scsi/aacraid/src.c      | 25 +++++++++++++++++++++++--
  5 files changed, 43 insertions(+), 4 deletions(-)
drivers/scsi/aacraid/aacraid.h +1 -0
index 5e115e8b2ba4..7c6efde75da6 100644
@@ -1678,6 +1678,7 @@ struct aac_dev
  	u32			handle_pci_error;
  	bool			init_reset;
  	u8			soft_reset_support;
+	u8			use_map_queue;
  };
 
  #define aac_adapter_interrupt(dev) \
drivers/scsi/aacraid/comminit.c +0 -1
index bd99c5492b7d..a5483e7e283a 100644
@@ -657,4 +657,3 @@ struct aac_dev *aac_init_adapter(struct aac_dev *dev)
 
  	return dev;
  }
-
drivers/scsi/aacraid/commsup.c +5 -1
index deb32c9f4b3e..3f062e4013ab 100644
@@ -223,8 +223,12 @@ int aac_fib_setup(struct aac_dev * dev)
  struct fib *aac_fib_alloc_tag(struct aac_dev *dev, struct scsi_cmnd *scmd)
  {
  	struct fib *fibptr;
+	u32 blk_tag;
+	int i;
 
-	fibptr = &dev->fibs[scsi_cmd_to_rq(scmd)->tag];
+	blk_tag = blk_mq_unique_tag(scsi_cmd_to_rq(scmd));
+	i = blk_mq_unique_tag_to_tag(blk_tag);
+	fibptr = &dev->fibs[i];
  	/*
  	 *	Null out fields that depend on being zero at the start of
  	 *	each I/O
drivers/scsi/aacraid/linit.c +14 -0
index 5ba5c18b77b4..9caf8c314ce1 100644
@@ -19,6 +19,7 @@
 
  #include <linux/compat.h>
  #include <linux/blkdev.h>
+#include <linux/blk-mq-pci.h>
  #include <linux/completion.h>
  #include <linux/init.h>
  #include <linux/interrupt.h>
@@ -505,6 +506,15 @@ static int aac_slave_configure(struct scsi_device *sdev)
  	return 0;
  }
 
+static void aac_map_queues(struct Scsi_Host *shost)
+{
+	struct aac_dev *aac = (struct aac_dev *)shost->hostdata;
+
+	blk_mq_pci_map_queues(&shost->tag_set.map[HCTX_TYPE_DEFAULT],
+				aac->pdev, 0);
+	aac->use_map_queue = true;
+}
+
  /**
   *	aac_change_queue_depth		-	alter queue depths
   *	@sdev:	SCSI device we are considering
@@ -1489,6 +1499,7 @@ static struct scsi_host_template aac_driver_template = {
  	.bios_param			= aac_biosparm,
  	.shost_groups			= aac_host_groups,
  	.slave_configure		= aac_slave_configure,
+	.map_queues			= aac_map_queues,
  	.change_queue_depth		= aac_change_queue_depth,
  	.sdev_groups			= aac_dev_groups,
  	.eh_abort_handler		= aac_eh_abort,
@@ -1776,6 +1787,8 @@ static int aac_probe_one(struct pci_dev *pdev, const struct pci_device_id *id)
  	shost->max_lun = AAC_MAX_LUN;
 
  	pci_set_drvdata(pdev, shost);
+	shost->nr_hw_queues = aac->max_msix;
+	shost->host_tagset = 1;
 
  	error = scsi_add_host(shost, &pdev->dev);
  	if (error)
@@ -1908,6 +1921,7 @@ static void aac_remove_one(struct pci_dev *pdev)
  	struct aac_dev *aac = (struct aac_dev *)shost->hostdata;
 
  	aac_cancel_rescan_worker(aac);
+	aac->use_map_queue = false;
  	scsi_remove_host(shost);
 
  	__aac_shutdown(aac);
drivers/scsi/aacraid/src.c +11 -2
index 11ef58204e96..61949f374188 100644
@@ -493,6 +493,10 @@ static int aac_src_deliver_message(struct fib *fib)
  #endif
 
  	u16 vector_no;
+	struct scsi_cmnd *scmd;
+	u32 blk_tag;
+	struct Scsi_Host *shost = dev->scsi_host_ptr;
+	struct blk_mq_queue_map *qmap;
 
  	atomic_inc(&q->numpending);
 
@@ -505,8 +509,25 @@ static int aac_src_deliver_message(struct fib *fib)
  		if ((dev->comm_interface == AAC_COMM_MESSAGE_TYPE3)
  			&& dev->sa_firmware)
  			vector_no = aac_get_vector(dev);
-		else
-			vector_no = fib->vector_no;
+		else {
+			if (!fib->vector_no || !fib->callback_data) {
+				if (shost && dev->use_map_queue) {
+					qmap = &shost->tag_set.map[HCTX_TYPE_DEFAULT];
+					vector_no = qmap->mq_map[raw_smp_processor_id()];
+				}
+				/*

drivers/scsi/aacraid/src.c static int aac_src_deliver_message(struct fib *fib) · line 528

Quoted 10 lines
drivers/scsi/aacraid/src.c +10 -0
+				 *	We hardcode the vector_no for
+				 *	reserved commands as a valid shost is
+				 *	absent during the init
+				 */
+				else
+					vector_no = 0;
+			} else {
+				scmd = (struct scsi_cmnd *)fib->callback_data;
+				blk_tag = blk_mq_unique_tag(scsi_cmd_to_rq(scmd));
+				vector_no = blk_mq_unique_tag_to_hwq(blk_tag);

Hannes' patch in the bugzilla was to revert to using hw queue #0 always for internal commands, and it didn't help.

Could there be any issue in using hw queue #0 for regular SCSI commands?

AFAICS, that's a significant change. Previously we would use fib->vector_no to decide the queue, which was in range (1, dev->max_msix).

BTW, is there any code which relies on a command being sent/received on the HW queue same as fib->vector_no?

Thanks,
John

Quoted 5 lines
drivers/scsi/aacraid/src.c +2 -0
+			
+		}
 
  		if (native_hba) {
  			if (fib->flags & FIB_CONTEXT_FLAG_NATIVE_HBA_TMF) {
#5 lore raw
Quoted 8 lines · John Garry
Quoted 7 lines · Sagar Biradar

Fix the IO hang that arises because of MSIx vector not having a mapped online CPU upon receiving completion.

The SCSI cmds take the blk_mq route, which is setup during the init. The reserved cmds fetch the vector_no from mq_map after the init is complete and before the init, they use 0 - as per the norm.

Quoted 7 lines
Quoted 2 lines

This the patch which seems to be causing the issue in
https://bugzilla.kernel.org/show_bug.cgi?id=217599

I will comment here since I got no response there...

We can still do a clean revert of this commit if no other solution is found before the end of the 6.7 rc cycle.

Regards,

James

#6 lore raw
In reply to John Garry
Quoted 149 lines · John Garry
Quoted 9 lines · Sagar Biradar

Fix the IO hang that arises because of MSIx vector not having a mapped online CPU upon receiving completion.

The SCSI cmds take the blk_mq route, which is setup during the init. The reserved cmds fetch the vector_no from mq_map after the init is complete and before the init, they use 0 - as per the norm.

This the patch which seems to be causing the issue in https://bugzilla.kernel.org/show_bug.cgi?id=217599

I will comment here since I got no response there...

Quoted 119 lines

---
  drivers/scsi/aacraid/aacraid.h  |  1 + drivers/scsi/aacraid/comminit.c |  1 - drivers/scsi/aacraid/commsup.c  |  6 +++++- drivers/scsi/aacraid/linit.c    | 14 ++++++++++++++ drivers/scsi/aacraid/src.c      | 25 +++++++++++++++++++++++--

  5 files changed, 43 insertions(+), 4 deletions(-)
drivers/scsi/aacraid/aacraid.h +1 -0
index 5e115e8b2ba4..7c6efde75da6 100644
@@ -1678,6 +1678,7 @@ struct aac_dev
       u32            handle_pci_error;
       bool            init_reset;
       u8            soft_reset_support;
+    u8            use_map_queue;
   };
   #define aac_adapter_interrupt(dev) \
drivers/scsi/aacraid/comminit.c +0 -1
index bd99c5492b7d..a5483e7e283a 100644
@@ -657,4 +657,3 @@ struct aac_dev *aac_init_adapter(struct aac_dev *dev)
       return dev;
   }
-
drivers/scsi/aacraid/commsup.c +5 -1
index deb32c9f4b3e..3f062e4013ab 100644
@@ -223,8 +223,12 @@ int aac_fib_setup(struct aac_dev * dev)
   struct fib *aac_fib_alloc_tag(struct aac_dev *dev, struct scsi_cmnd *scmd)
   {
       struct fib *fibptr;
+    u32 blk_tag;
+    int i;
-    fibptr = &dev->fibs[scsi_cmd_to_rq(scmd)->tag];
+    blk_tag = blk_mq_unique_tag(scsi_cmd_to_rq(scmd));
+    i = blk_mq_unique_tag_to_tag(blk_tag);
+    fibptr = &dev->fibs[i];
       /*
        *    Null out fields that depend on being zero at the start of
        *    each I/O
drivers/scsi/aacraid/linit.c +11 -0
index 5ba5c18b77b4..9caf8c314ce1 100644
@@ -19,6 +19,7 @@
   #include <linux/compat.h>
   #include <linux/blkdev.h>
+#include <linux/blk-mq-pci.h>
   #include <linux/completion.h>
   #include <linux/init.h>
   #include <linux/interrupt.h>
@@ -505,6 +506,15 @@ static int aac_slave_configure(struct scsi_device 
 *sdev)
       return 0;
   }
+static void aac_map_queues(struct Scsi_Host *shost)
+{
+    struct aac_dev *aac = (struct aac_dev *)shost->hostdata;
+
+    blk_mq_pci_map_queues(&shost->tag_set.map[HCTX_TYPE_DEFAULT],
+                aac->pdev, 0);
+    aac->use_map_queue = true;
+}
+
   /**
    *    aac_change_queue_depth        -    alter queue depths
    *    @sdev:    SCSI device we are considering
@@ -1489,6 +1499,7 @@ static struct scsi_host_template 
 aac_driver_template = {
       .bios_param            = aac_biosparm,
       .shost_groups            = aac_host_groups,
       .slave_configure        = aac_slave_configure,
+    .map_queues            = aac_map_queues,
       .change_queue_depth        = aac_change_queue_depth,
       .sdev_groups            = aac_dev_groups,

      .eh_abort_handler        = aac_eh_abort,

drivers/scsi/aacraid/linit.c +3 -0
@@ -1776,6 +1787,8 @@ static int aac_probe_one(struct pci_dev *pdev, 
 const struct pci_device_id *id)
       shost->max_lun = AAC_MAX_LUN;
       pci_set_drvdata(pdev, shost);
+    shost->nr_hw_queues = aac->max_msix;
+    shost->host_tagset = 1;
       error = scsi_add_host(shost, &pdev->dev);
       if (error)
@@ -1908,6 +1921,7 @@ static void aac_remove_one(struct pci_dev *pdev)
       struct aac_dev *aac = (struct aac_dev *)shost->hostdata;
       aac_cancel_rescan_worker(aac);
+    aac->use_map_queue = false;
       scsi_remove_host(shost);
       __aac_shutdown(aac);
drivers/scsi/aacraid/src.c +11 -2
index 11ef58204e96..61949f374188 100644
@@ -493,6 +493,10 @@ static int aac_src_deliver_message(struct fib *fib)
   #endif
       u16 vector_no;
+    struct scsi_cmnd *scmd;
+    u32 blk_tag;
+    struct Scsi_Host *shost = dev->scsi_host_ptr;
+    struct blk_mq_queue_map *qmap;
       atomic_inc(&q->numpending);
@@ -505,8 +509,25 @@ static int aac_src_deliver_message(struct fib *fib)
           if ((dev->comm_interface == AAC_COMM_MESSAGE_TYPE3)
               && dev->sa_firmware)
               vector_no = aac_get_vector(dev);
-        else
-            vector_no = fib->vector_no;
+        else {
+            if (!fib->vector_no || !fib->callback_data) {
+                if (shost && dev->use_map_queue) {
+                    qmap = &shost->tag_set.map[HCTX_TYPE_DEFAULT];
+                    vector_no = qmap->mq_map[raw_smp_processor_id()];
+                }
+                /*

drivers/scsi/aacraid/src.c static int aac_src_deliver_message(struct fib *fib) · line 528

Quoted 10 lines
drivers/scsi/aacraid/src.c +10 -0
+                 *    We hardcode the vector_no for
+                 *    reserved commands as a valid shost is
+                 *    absent during the init
+                 */
+                else
+                    vector_no = 0;
+            } else {
+                scmd = (struct scsi_cmnd *)fib->callback_data;
+                blk_tag = blk_mq_unique_tag(scsi_cmd_to_rq(scmd));
+                vector_no = blk_mq_unique_tag_to_hwq(blk_tag);

Hannes' patch in the bugzilla was to revert to using hw queue #0 always for internal commands, and it didn't help.

Quoted 6 lines

Could there be any issue in using hw queue #0 for regular SCSI commands?

AFAICS, that's a significant change. Previously we would use fib->vector_no to decide the queue, which was in range (1, dev->max_msix).

BTW, is there any code which relies on a command being sent/received on the HW queue same as fib->vector_no?

Yeah, and that's the clincher.
The vector/MSIx interrupt to use for the completion is encoded in the command (cf drivers/scsi/aacraid/src.c:aac_src_deliver_message()):

fib->hw_fib_va)->reply_qid = vector_no;

_and_ the command index:

fib->hw_fib_va)->request_id += (vector_no << 16);

so if we miscalculate the vector number here the command completion will lookup the wrong command, and we get this issue.

But maybe the fix is relatively simple; can you try this:

drivers/scsi/aacraid/src.c +1 -1
index 61949f374188..698a206a2b43 100644
@@ -510,7 +510,7 @@ static int aac_src_deliver_message(struct fib *fib)
                        && dev->sa_firmware)
                        vector_no = aac_get_vector(dev);
                else {
-                       if (!fib->vector_no || !fib->callback_data) {
+                       if (!fib->callback_data) {
                                if (shost && dev->use_map_queue) {
                                        qmap = 

&shost->tag_set.map[HCTX_TYPE_DEFAULT];
vector_no =
qmap->mq_map[raw_smp_processor_id()];

Hmm?

Cheers,

Hannes

Signature · 5 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), GF: Ivo Totev, Andrew McDonald,
Werner Knoblich