linux-kernel - Re: [PATCHv2 1/7] block: check for valid bio while splitting

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 for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <b6860c56-e91d-45c8-8d4c-05bcae97a2bb@kernel.org>
Date: Thu, 14 Aug 2025 10:42:35 +0900
From: Damien Le Moal <dlemoal@...nel.org>
To: Keith Busch <kbusch@...nel.org>, Bart Van Assche <bvanassche@....org>
Cc: Keith Busch <kbusch@...a.com>, linux-block@...r.kernel.org,
 linux-fsdevel@...r.kernel.org, linux-kernel@...r.kernel.org,
 snitzer@...nel.org, axboe@...nel.dk, dw@...idwei.uk, brauner@...nel.org,
 Hannes Reinecke <hare@...e.de>
Subject: Re: [PATCHv2 1/7] block: check for valid bio while splitting

On 8/14/25 6:01 AM, Keith Busch wrote:
> On Wed, Aug 13, 2025 at 01:41:49PM -0700, Bart Van Assche wrote:
>> On 8/13/25 1:06 PM, Keith Busch wrote:
>>> But I can't make that change because many scsi devices don't set the dma
>>> alignment and get the default 511 value. This is fine for the memory
>>> address offset, but the lengths sent for various inquriy commands are
>>> much smaller, like 4 and 32 byte lengths. That length wouldn't pass the
>>> dma alignment granularity, so I think the default value is far too
>>> conservative. Does the address start size need to be a different limit
>>> than minimum length? I feel like they should be the same, but maybe
>>> that's just an nvme thing.
>>
>> Hi Keith,
>>
>> Maybe I misunderstood your question. It seems to me that the SCSI core
>> sets the DMA alignment by default to four bytes. From
>> drivers/scsi/hosts.c:
> 
> Thanks, I think you got my meaning. 
> 
> I'm using the AHCI driver. It looks like ata_scsi_dev_config() overrides
> the dma_alignment to sector_size - 1, and that pattern goes way back,
> almost 20 years ago, so maybe I can't change it.

That is probably buggy now in the sense that the scsi layer should be able to
send any command with a size not aligned to the LBA size or ATA sector (512 B)
and libata-scsi SAT should do the translation using an internal 512B aligned
command size.

What makes a mess here is that SCSI allows having a media-access command
specifying a transfer size that is not aligned on the LBA size. The transfer
will be "short" in that case, which is perfectly fine with SCSI. But ATA does
not allow that. It is all or nothing and the command size thus must always be
aligned to the LBA size.

I think that dma_alignment was abused to check that. But I think it should not
be too hard to check the alignment in libata-scsi when translating the command.
SAS HBAs should be doing something similar too. Have never exactly tested that
though, and I am afraid how many SAS HBAs will not like unaligned command to
ATA devices...

We also have the different alignment for management commands (most of which use
512B sector size) and media access commands which use the actual device LBA
size alignment.

So it is a mess :)

-- 
Damien Le Moal
Western Digital Research