linux-kernel - Re: [PATCH] [v2] nvme-pci: disable write zeros support on specific SSDs

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>] [day] [month] [year] [list]

Message-ID: <YjNrkknvwFYhW3ME@C02CK6Q3MD6M>
Date:   Thu, 17 Mar 2022 11:10:42 -0600
From:   Keith Busch <kbusch@...nel.org>
To:     Mark Hsieh/WHQ/Wistron <Mark_Hsieh@...tron.com>
Cc:     Christoph Hellwig <hch@....de>,
        Mark Hsieh <mark_hsieh@...tron.corp-partner.google.com>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "linux-nvme@...ts.infradead.org" <linux-nvme@...ts.infradead.org>,
        "sagi@...mberg.me" <sagi@...mberg.me>,
        "axboe@...com" <axboe@...com>,
        "alsa-devel@...a-project.org" <alsa-devel@...a-project.org>
Subject: Re: [PATCH] [v2] nvme-pci: disable write zeros support on specific
 SSDs

On Thu, Mar 17, 2022 at 07:03:52AM +0000, Mark Hsieh/WHQ/Wistron wrote:
> Sorry late reply as it take a while to confirm with Micron engineering team.
 
It would help keep the thread readable if you send mailing list messages in
plain text. 
 
> But more importantly something seems really broken here if the device claims to support Write Zeroes but then doesn't claims it is not supported.
> 
>   *   We support write zero command indeed, but if host send too much write zero commands in due to the architecture design, we need spend lots of time to handle it, so we want to add 2450 in the google check list to resolve this issue.
>   *   2450 doesn't violate the write zero command spec, but total write zero command handle time will be too long for host point of view.
> 
> Can you see if this only happens with some I/O size?  With our without the nounmap flag?  Does the device report non-MDTS command limits?
> 
>   *   If the LBA range of each write zero command is quite small, it will let drive spend more time to handle it.
>   *   With noun-map flag (Deallocate = 0)
>   *   We report MDTS as 256KB, but write zero command doesn't take it as a reference.

The driver already limits Write Zeroes to MDTS if it doesn't report non-MDTS
limits. The spec doesn't require a driver do that, but we do out of an
abundance of caution.

So you are saying that sending many small Write Zeroes is taking more time on
the drive side compared to a single larger Write Zeroes? Would it be successful
if we weren't splitting these?