It is possible that bus_reset_cleanup() or .eh_abort_handler could
be invoked during NCR5380_queuecommand(). If that takes place before
the new command is enqueued and after the ST-DMA "lock" has been
acquired, the ST-DMA "lock" will be released again. This will result
in a lost DMA interrupt and a command timeout. Fix this by excluding
EH and interrupt handlers while the new command is enqueued.
Signed-off-by: Finn Thain <[email protected]>
---
Michael, would you please send your Acked-by or Reviewed-and-tested-by?
These two patches taken together should be equivalent to the one you tested
recently. I've split it into two as that seemed to make more sense.
---
drivers/scsi/NCR5380.c | 9 ++++++---
drivers/scsi/atari_scsi.c | 10 +++-------
2 files changed, 9 insertions(+), 10 deletions(-)
diff --git a/drivers/scsi/NCR5380.c b/drivers/scsi/NCR5380.c
index d654a6cc4162..ea4b5749e7da 100644
--- a/drivers/scsi/NCR5380.c
+++ b/drivers/scsi/NCR5380.c
@@ -580,11 +580,14 @@ static int NCR5380_queue_command(struct Scsi_Host *instance,
cmd->result = 0;
- if (!NCR5380_acquire_dma_irq(instance))
- return SCSI_MLQUEUE_HOST_BUSY;
-
spin_lock_irqsave(&hostdata->lock, flags);
+ if (!NCR5380_acquire_dma_irq(instance)) {
+ spin_unlock_irqrestore(&hostdata->lock, flags);
+
+ return SCSI_MLQUEUE_HOST_BUSY;
+ }
+
/*
* Insert the cmd into the issue queue. Note that REQUEST SENSE
* commands are added to the head of the queue since any command will
diff --git a/drivers/scsi/atari_scsi.c b/drivers/scsi/atari_scsi.c
index a82b63a66635..95d7a3586083 100644
--- a/drivers/scsi/atari_scsi.c
+++ b/drivers/scsi/atari_scsi.c
@@ -376,15 +376,11 @@ static int falcon_get_lock(struct Scsi_Host *instance)
if (IS_A_TT())
return 1;
- if (stdma_is_locked_by(scsi_falcon_intr) &&
- instance->hostt->can_queue > 1)
+ if (stdma_is_locked_by(scsi_falcon_intr))
return 1;
- if (in_interrupt())
- return stdma_try_lock(scsi_falcon_intr, instance);
-
- stdma_lock(scsi_falcon_intr, instance);
- return 1;
+ /* stdma_lock() may sleep which means it can't be used here */
+ return stdma_try_lock(scsi_falcon_intr, instance);
}
#ifndef MODULE
--
2.26.2
Hi Finn,
thanks for your patch!
Tested on Atari Falcon (with falconide, and pata_falcon modules).
Reviewed-by: Michael Schmitz <[email protected]>
Tested-by: Michael Schmitz <[email protected]>
Am 20.11.2020 um 17:39 schrieb Finn Thain:
> It is possible that bus_reset_cleanup() or .eh_abort_handler could
> be invoked during NCR5380_queuecommand(). If that takes place before
> the new command is enqueued and after the ST-DMA "lock" has been
> acquired, the ST-DMA "lock" will be released again. This will result
> in a lost DMA interrupt and a command timeout. Fix this by excluding
> EH and interrupt handlers while the new command is enqueued.
>
> Signed-off-by: Finn Thain <[email protected]>
> ---
> Michael, would you please send your Acked-by or Reviewed-and-tested-by?
> These two patches taken together should be equivalent to the one you tested
> recently. I've split it into two as that seemed to make more sense.
> ---
> drivers/scsi/NCR5380.c | 9 ++++++---
> drivers/scsi/atari_scsi.c | 10 +++-------
> 2 files changed, 9 insertions(+), 10 deletions(-)
>
> diff --git a/drivers/scsi/NCR5380.c b/drivers/scsi/NCR5380.c
> index d654a6cc4162..ea4b5749e7da 100644
> --- a/drivers/scsi/NCR5380.c
> +++ b/drivers/scsi/NCR5380.c
> @@ -580,11 +580,14 @@ static int NCR5380_queue_command(struct Scsi_Host *instance,
>
> cmd->result = 0;
>
> - if (!NCR5380_acquire_dma_irq(instance))
> - return SCSI_MLQUEUE_HOST_BUSY;
> -
> spin_lock_irqsave(&hostdata->lock, flags);
>
> + if (!NCR5380_acquire_dma_irq(instance)) {
> + spin_unlock_irqrestore(&hostdata->lock, flags);
> +
> + return SCSI_MLQUEUE_HOST_BUSY;
> + }
> +
> /*
> * Insert the cmd into the issue queue. Note that REQUEST SENSE
> * commands are added to the head of the queue since any command will
> diff --git a/drivers/scsi/atari_scsi.c b/drivers/scsi/atari_scsi.c
> index a82b63a66635..95d7a3586083 100644
> --- a/drivers/scsi/atari_scsi.c
> +++ b/drivers/scsi/atari_scsi.c
> @@ -376,15 +376,11 @@ static int falcon_get_lock(struct Scsi_Host *instance)
> if (IS_A_TT())
> return 1;
>
> - if (stdma_is_locked_by(scsi_falcon_intr) &&
> - instance->hostt->can_queue > 1)
> + if (stdma_is_locked_by(scsi_falcon_intr))
> return 1;
>
> - if (in_interrupt())
> - return stdma_try_lock(scsi_falcon_intr, instance);
> -
> - stdma_lock(scsi_falcon_intr, instance);
> - return 1;
> + /* stdma_lock() may sleep which means it can't be used here */
> + return stdma_try_lock(scsi_falcon_intr, instance);
> }
>
> #ifndef MODULE
>
Finn,
> It is possible that bus_reset_cleanup() or .eh_abort_handler could be
> invoked during NCR5380_queuecommand(). If that takes place before the
> new command is enqueued and after the ST-DMA "lock" has been acquired,
> the ST-DMA "lock" will be released again. This will result in a lost
> DMA interrupt and a command timeout. Fix this by excluding EH and
> interrupt handlers while the new command is enqueued.
Applied to 5.11/scsi-staging, thanks!
--
Martin K. Petersen Oracle Linux Engineering
On Fri, 20 Nov 2020 15:39:56 +1100, Finn Thain wrote:
> It is possible that bus_reset_cleanup() or .eh_abort_handler could
> be invoked during NCR5380_queuecommand(). If that takes place before
> the new command is enqueued and after the ST-DMA "lock" has been
> acquired, the ST-DMA "lock" will be released again. This will result
> in a lost DMA interrupt and a command timeout. Fix this by excluding
> EH and interrupt handlers while the new command is enqueued.
Applied to 5.11/scsi-queue, thanks!
[1/1] scsi: atari_scsi: Fix race condition between .queuecommand and EH
https://git.kernel.org/mkp/scsi/c/03fe6a640a05
--
Martin K. Petersen Oracle Linux Engineering