[v3] scsi: libsas: abort all inflight requests when device is gone

Commit Message

Jason Yan March 30, 2023, 11:09 a.m. UTC

When a disk is removed with inflight IO, the application need to wait
for 30 senconds(depends on the timeout configuration) to get back from
the kernel. Xingui tried to fix this issue by aborting the ATA link for
SATA devices[1]. However this approach left the SAS devices unresolved.

This patch try to fix this issue by aborting all inflight requests while
the device is gone. This is implemented by itering the tagset.

[1] https://lore.kernel.org/lkml/[e-mail]/T/

Cc: Xingui Yang <[e-mail]>
Cc: John Garry <[e-mail]>
Cc: Damien Le Moal <[e-mail]>
Cc: Hannes Reinecke <[e-mail]>
Signed-off-by: Jason Yan <[e-mail]>
Reviewed-by: John Garry <[e-mail]>
---
 v2->v3:
   1. Remove 'userspace' before application and fix typo in commit message.
   2. Add John's reviewed-by tag.
 v1->v2:
   1. Rename sas_abort_domain_cmds() to sas_abort_device_scsi_cmds().
   2. Don't do the aborting for expanders and for devices not completely initialinzed.
   3. Add a comment to explain why we need to abort these commands.

 drivers/scsi/libsas/sas_discover.c | 29 +++++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)

Comments

Martin K. Petersen April 3, 2023, 1:19 a.m. UTC | #1

Jason,

> When a disk is removed with inflight IO, the application need to wait
> for 30 senconds(depends on the timeout configuration) to get back from
> the kernel. Xingui tried to fix this issue by aborting the ATA link
> for SATA devices[1]. However this approach left the SAS devices
> unresolved.

Applied to 6.4/scsi-staging, thanks!

Martin K. Petersen April 12, 2023, 2:04 a.m. UTC | #2

On Thu, 30 Mar 2023 19:09:30 +0800, Jason Yan wrote:

> When a disk is removed with inflight IO, the application need to wait
> for 30 senconds(depends on the timeout configuration) to get back from
> the kernel. Xingui tried to fix this issue by aborting the ATA link for
> SATA devices[1]. However this approach left the SAS devices unresolved.
> 
> This patch try to fix this issue by aborting all inflight requests while
> the device is gone. This is implemented by itering the tagset.
> 
> [...]

Applied to 6.4/scsi-queue, thanks!

[1/1] scsi: libsas: abort all inflight requests when device is gone
      https://git.kernel.org/mkp/scsi/c/0e4b1791d9b1

Patch

diff --git a/drivers/scsi/libsas/sas_discover.c b/drivers/scsi/libsas/sas_discover.c
index 72fdb2e5d047..8c6afe724944 100644
--- a/drivers/scsi/libsas/sas_discover.c
+++ b/drivers/scsi/libsas/sas_discover.c
@@ -360,6 +360,33 @@  static void sas_destruct_ports(struct asd_sas_port *port)
 	}
 }
 
+static bool sas_abort_cmd(struct request *req, void *data)
+{
+	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
+	struct domain_device *dev = data;
+
+	if (dev == cmd_to_domain_dev(cmd))
+		blk_abort_request(req);
+	return true;
+}
+
+static void sas_abort_device_scsi_cmds(struct domain_device *dev)
+{
+	struct sas_ha_struct *sas_ha = dev->port->ha;
+	struct Scsi_Host *shost = sas_ha->core.shost;
+
+	if (dev_is_expander(dev->dev_type))
+		return;
+
+	/*
+	 * For removed device with active IOs, the user space applications have
+	 * to spend very long time waiting for the timeout. This is not
+	 * necessary because a removed device will not return the IOs.
+	 * Abort the inflight IOs here so that EH can be quickly kicked in.
+	 */
+	blk_mq_tagset_busy_iter(&shost->tag_set, sas_abort_cmd, dev);
+}
+
 void sas_unregister_dev(struct asd_sas_port *port, struct domain_device *dev)
 {
 	if (!test_bit(SAS_DEV_DESTROY, &dev->state) &&
@@ -372,6 +399,8 @@  void sas_unregister_dev(struct asd_sas_port *port, struct domain_device *dev)
 	}
 
 	if (!test_and_set_bit(SAS_DEV_DESTROY, &dev->state)) {
+		if (test_bit(SAS_DEV_GONE, &dev->state))
+			sas_abort_device_scsi_cmds(dev);
 		sas_rphy_unlink(dev->rphy);
 		list_move_tail(&dev->disco_list_node, &port->destroy_list);
 	}