Re: [REGRESSION][BISECTED][STABLE] hdparm errors since 28ab9769117c

[Damien Le Moal <dlemoal@kernel.org>]

On 2024/08/09 7:30, Niklas Cassel wrote:
> Hello Damien,
> 
> If we want to no longer respect the D_SENSE bit for successful ATA-passthrough
> commands, e.g. by replacing the ata_scsi_set_sense() call with a
> scsi_build_sense() call in the else clause:
> https://github.com/torvalds/linux/blob/v6.11-rc2/drivers/ata/libata-scsi.c#L955
> 
> ...then I think that we should also replace the ata_scsi_set_sense() call with
> a scsi_build_sense() call for failed ATA-passthrough commands too
> (in the non-else clause):
> https://github.com/torvalds/linux/blob/v6.11-rc2/drivers/ata/libata-scsi.c#L952
> 
> ..however, that does not sound like a very nice solution IMO.
> 
> 
> Another option, if there are a lot of user space programs that incorrectly
> assume that the sense data (for both successful and failed commands) is in
> descriptor format, without bothering to check the sense data type, one option
> might be to change the default value of D_SENSE in the control mode page to 1
> in libata's SATL, i.e. set ATA_DFLAG_D_SENSE in dev->flags by default.
> 
> That way, we will still respect D_SENSE while generating the sense data
> (in case the user issues a mode select to modify the bit), and the default
> will be to generate the sense data in descriptor format, as it has been
> since 11093cb1ef56 ("libata-scsi: generate correct ATA pass-through sense").

That indeed should be acceptable. And we should also patch hdparm to properly
look at the sense format and not assume descriptor format by default.


-- 
Damien Le Moal
Western Digital Research