Re: ata timeout exceptions

[Niklas Cassel <cassel@kernel.org>]

Hello Eyal,

On Mon, Nov 03, 2025 at 03:13:34PM +1100, Eyal Lebedinsky wrote:
> 
> I want to understand what is going on:
> 
> 1) Why do I not see an I/O error and the writes to the disk (rsync) seem to complete?
>    Which layer absorbs the errors, hiding them from the application?

SCSI layer.

For a timed out command, libata will set DID_TIME_OUT:
https://github.com/torvalds/linux/blob/v6.18-rc4/drivers/ata/libata-eh.c#L652-L654

For most commands SCSI layer, SCSI will set cmd->allowed to sdkp->max_retries:
https://github.com/torvalds/linux/blob/v6.18-rc4/drivers/scsi/sd.c#L1411

which by default is 5:
https://github.com/torvalds/linux/blob/v6.18-rc4/drivers/scsi/sd.c#L3962

Thus, most commands will be retried up to 5 times:
https://github.com/torvalds/linux/blob/v6.18-rc4/drivers/scsi/scsi_error.c#L2225

Thus, the user will only see the I/O as an error if the command failed
6 times.

(Note that if the command returns sense data instead of timeout, depending on
the sense data returned, we might report an I/O error to the user immediately.


> 
> 2) Why do I get only one command_timeout counted (originally, with ncq active) and none when ncq is disabled?

You are right that even if it is only a single command that times out,
the whole queue will be drained and retried.
(Because we always do a hard reset after a command timeout.)

command_timeout is most likely increased only by one because it was
only a single command that timed out. (The other commands might have
been queued but were never executed/finished.)

I have no idea why a command timeout, when NCQ has been disabled,
does not increase the command_timeout counter. My expectation would
have been for the counter to still be increased by one.


Kind regards,
Niklas