lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <Zmy5I-XNK5bffSe3@google.com>
Date: Fri, 14 Jun 2024 21:41:55 +0000
From: Igor Pylypiv <ipylypiv@...gle.com>
To: Niklas Cassel <cassel@...nel.org>
Cc: TJ Adams <tadamsjr@...gle.com>, Jack Wang <jinpu.wang@...ud.ionos.com>,
	"James E . J . Bottomley" <James.Bottomley@...senpartnership.com>,
	"Martin K . Petersen" <martin.petersen@...cle.com>,
	linux-scsi@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 2/3] scsi: pm80xx: Do not issue hard reset before NCQ EH

On Tue, Jun 11, 2024 at 10:40:48AM +0200, Niklas Cassel wrote:
> Hello Igor, TJ,
> 
Hi Niklas,

Thank you for the feedback!

> On Fri, Jun 07, 2024 at 05:57:42PM +0000, TJ Adams wrote:
> > From: Igor Pylypiv <ipylypiv@...gle.com>
> > 
> > v6.2 commit 811be570a9a8 ("scsi: pm8001: Use sas_ata_device_link_abort()
> 
> Do not specify kernel version (it is irrelevant), SHA1 is enough.
> 
Noted.

> 
> > to handle NCQ errors") removed duplicate NCQ EH from the pm80xx driver
> > and started relying on libata to handle the NCQ errors. The PM8006
> > controller has a special EH sequence that was added in v4.15 commit
> > 869ddbdcae3b ("scsi: pm80xx: corrected SATA abort handling sequence.").
> 
> Do not specify kernel version (it is irrelevant), SHA1 is enough.
> 
> Since the code added in 869ddbdcae3b still exists in the pm80xx driver,
> I think that you should mention the commits in chronological order.
> (Right now you mention the oldest still existing code last, which seems
> a bit backwards.)
>
Noted. I wanted to emphasise that newer commit 811be570a9a8 broke the NCQ EH
for pm8006 so I put it first. I should have added a Fixes tag to make it
clear. 

> 
> > The special EH sequence issues a hard reset to a drive before libata EH
> > has a chance to read the NCQ log page. Libata EH gets confused by empty
> > NCQ log page which results in HSM violation. The failed command gets
> > retried a few times and each time fails with the same HSM violation.
> > Finally, libata decides to disable NCQ due to subsequent HSM vioaltions.
> 
> s/vioaltions/violations/
> 
> I'm not an expert in libsas EH, but I think that your commit message fails
> to explain why this change actually fixes anything. You do not mention the
> relationship between the code that you add pm8001_work_fn() and the
> existing code in pm8001_abort_task(), and the order in which the functions
> get executed.
> 
Noted, will update with more details.

> Does calling sas_execute_internal_abort_dev() from pm8001_work_fn() ensure
> that the libsas EH is never invoked? Or does it cancel the hard reset that
> is part of the "special EH sequence" in pm8001_abort_task() ?
> 
It ensures that all I/Os are aborted before libsas EH kicks in. Since all
I/Os are aborted by the controller the pm8001_abort_task() will not be called.
Going to add that to commit message as well.

> Wouldn't it be better if this was fixed in pm8001_abort_task() or similar
> instead? It appears that the code you add to pm8001_work_fn() (that has a
> very ugly if (pm8006)) is only there to undo or avoid the hard reset that
> is done in pm8001_abort_task() (which also has a very ugly if (pm8006)).
>

It would definetely be better to fix this in pm8001_abort_task(), if possible.
One way would be to add a flag that will be set when NCQ error happens (when
IO_XFER_ERROR_ABORTED_NCQ_MODE event is received) and then check that flag
in pm8001_abort_task() to skip hard reset. This approach adds another type
of ugliness to the code and I'm not sure which of these ugly approaches is
less ugly.

> Now we have this ugly if (pm8006) in two different functions... which
> makes my "this could be solved in a nicer way" detector go off.
> 

I would be very happy if we can drop those ugly if (pm8006) checks and have
a generic nice solution.

> If this patch (as is) is really the way to go, then I think there should
> be a more detailed reasoning why this change is the most sensible one.
> 

Let me investigate this issue more to see if there is a way to drop 
the ugly pm8006 checks.

Any ideas/suggestions on how to fix this nicely are very welcomed.

> 
> Kind regards,
> Niklas

Thank you,
Igor

Powered by blists - more mailing lists

Powered by Openwall GNU/*/Linux Powered by OpenVZ