ata: libata-scsi: Avoid deadlock on rescan after device resume
authorDamien Le Moal <dlemoal@kernel.org>
Thu, 15 Jun 2023 08:18:53 +0000 (17:18 +0900)
committerDamien Le Moal <dlemoal@kernel.org>
Sun, 18 Jun 2023 03:00:49 +0000 (12:00 +0900)
When an ATA port is resumed from sleep, the port is reset and a power
management request issued to libata EH to reset the port and rescanning
the device(s) attached to the port. Device rescanning is done by
scheduling an ata_scsi_dev_rescan() work, which will execute
scsi_rescan_device().

However, scsi_rescan_device() takes the generic device lock, which is
also taken by dpm_resume() when the SCSI device is resumed as well. If
a device rescan execution starts before the completion of the SCSI
device resume, the rcu locking used to refresh the cached VPD pages of
the device, combined with the generic device locking from
scsi_rescan_device() and from dpm_resume() can cause a deadlock.

Avoid this situation by changing struct ata_port scsi_rescan_task to be
a delayed work instead of a simple work_struct. ata_scsi_dev_rescan() is
modified to check if the SCSI device associated with the ATA device that
must be rescanned is not suspended. If the SCSI device is still
suspended, ata_scsi_dev_rescan() returns early and reschedule itself for
execution after an arbitrary delay of 5ms.

Reported-by: Kai-Heng Feng <kai.heng.feng@canonical.com>
Reported-by: Joe Breuer <linux-kernel@jmbreuer.net>
Closes: https://bugzilla.kernel.org/show_bug.cgi?id=217530
Fixes: a19a93e4c6a9 ("scsi: core: pm: Rely on the device driver core for async power management")
Signed-off-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Tested-by: Kai-Heng Feng <kai.heng.feng@canonical.com>
Tested-by: Joe Breuer <linux-kernel@jmbreuer.net>
drivers/ata/libata-core.c
drivers/ata/libata-eh.c
drivers/ata/libata-scsi.c
include/linux/libata.h

index 8bf612bdd61a5a1d8d04d70fd8579249ede7c728..b4f246f0cac7f05a6942d41526af51d6f65eb766 100644 (file)
@@ -5348,7 +5348,7 @@ struct ata_port *ata_port_alloc(struct ata_host *host)
 
        mutex_init(&ap->scsi_scan_mutex);
        INIT_DELAYED_WORK(&ap->hotplug_task, ata_scsi_hotplug);
-       INIT_WORK(&ap->scsi_rescan_task, ata_scsi_dev_rescan);
+       INIT_DELAYED_WORK(&ap->scsi_rescan_task, ata_scsi_dev_rescan);
        INIT_LIST_HEAD(&ap->eh_done_q);
        init_waitqueue_head(&ap->eh_wait_q);
        init_completion(&ap->park_req_pending);
@@ -5954,6 +5954,7 @@ static void ata_port_detach(struct ata_port *ap)
        WARN_ON(!(ap->pflags & ATA_PFLAG_UNLOADED));
 
        cancel_delayed_work_sync(&ap->hotplug_task);
+       cancel_delayed_work_sync(&ap->scsi_rescan_task);
 
  skip_eh:
        /* clean up zpodd on port removal */
index a6c90181180271bd0c435a658b4dc71a57f3968c..6f8d141915939441dafd1a9852e2172da128cadf 100644 (file)
@@ -2984,7 +2984,7 @@ static int ata_eh_revalidate_and_attach(struct ata_link *link,
                        ehc->i.flags |= ATA_EHI_SETMODE;
 
                        /* schedule the scsi_rescan_device() here */
-                       schedule_work(&(ap->scsi_rescan_task));
+                       schedule_delayed_work(&ap->scsi_rescan_task, 0);
                } else if (dev->class == ATA_DEV_UNKNOWN &&
                           ehc->tries[dev->devno] &&
                           ata_class_enabled(ehc->classes[dev->devno])) {
index 8ce90284eb344b6a49e8d40ad5be930e8d7485f7..551077cea4e4e023e20e913e37ee7b939488083a 100644 (file)
@@ -4597,10 +4597,11 @@ int ata_scsi_user_scan(struct Scsi_Host *shost, unsigned int channel,
 void ata_scsi_dev_rescan(struct work_struct *work)
 {
        struct ata_port *ap =
-               container_of(work, struct ata_port, scsi_rescan_task);
+               container_of(work, struct ata_port, scsi_rescan_task.work);
        struct ata_link *link;
        struct ata_device *dev;
        unsigned long flags;
+       bool delay_rescan = false;
 
        mutex_lock(&ap->scsi_scan_mutex);
        spin_lock_irqsave(ap->lock, flags);
@@ -4614,6 +4615,21 @@ void ata_scsi_dev_rescan(struct work_struct *work)
                        if (scsi_device_get(sdev))
                                continue;
 
+                       /*
+                        * If the rescan work was scheduled because of a resume
+                        * event, the port is already fully resumed, but the
+                        * SCSI device may not yet be fully resumed. In such
+                        * case, executing scsi_rescan_device() may cause a
+                        * deadlock with the PM code on device_lock(). Prevent
+                        * this by giving up and retrying rescan after a short
+                        * delay.
+                        */
+                       delay_rescan = sdev->sdev_gendev.power.is_suspended;
+                       if (delay_rescan) {
+                               scsi_device_put(sdev);
+                               break;
+                       }
+
                        spin_unlock_irqrestore(ap->lock, flags);
                        scsi_rescan_device(&(sdev->sdev_gendev));
                        scsi_device_put(sdev);
@@ -4623,4 +4639,8 @@ void ata_scsi_dev_rescan(struct work_struct *work)
 
        spin_unlock_irqrestore(ap->lock, flags);
        mutex_unlock(&ap->scsi_scan_mutex);
+
+       if (delay_rescan)
+               schedule_delayed_work(&ap->scsi_rescan_task,
+                                     msecs_to_jiffies(5));
 }
index 311cd93377c75e20a80a3dbe48f60d3453b940a3..dd5797fb6305cb29b3edff708215b0459a7516b2 100644 (file)
@@ -836,7 +836,7 @@ struct ata_port {
 
        struct mutex            scsi_scan_mutex;
        struct delayed_work     hotplug_task;
-       struct work_struct      scsi_rescan_task;
+       struct delayed_work     scsi_rescan_task;
 
        unsigned int            hsm_task_state;