[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <3a4cdc4f-63e9-48b1-80b0-54c5d1eb3e7f@kernel.dk>
Date: Thu, 17 Oct 2024 19:10:49 -0600
From: Jens Axboe <axboe@...nel.dk>
To: Josh Poimboeuf <jpoimboe@...nel.org>
Cc: Phillip Potter <phil@...lpotter.co.uk>, linux-kernel@...r.kernel.org,
Jordy Zomer <jordyzomer@...gle.com>
Subject: Re: [PATCH] cdrom: Avoid barrier_nospec() in
cdrom_ioctl_media_changed()
On 10/17/24 6:52 PM, Josh Poimboeuf wrote:
> On Thu, Oct 17, 2024 at 06:33:24PM -0600, Jens Axboe wrote:
>> On 10/17/24 4:09 PM, Josh Poimboeuf wrote:
>>> The barrier_nospec() after the array bounds check is overkill and
>>> painfully slow for arches which implement it.
>>>
>>> Furthermore, most arches don't implement it, so they remain exposed to
>>> Spectre v1 (which can affect pretty much any CPU with branch
>>> prediction).
>>>
>>> Instead, clamp the user pointer to a valid range so it's guaranteed to
>>> be a valid array index even when the bounds check mispredicts.
>>
>> It's a cdrom, and media change detection to be more specific. I really
>> don't think anyone would care about performance here, it's not even
>> a hot path for a cdrom driver. That said, I don't disagree with
>> the change, just don't think it'll make one iota of difference
>> in the real world.
>
> Fair, though it's also about hardening as barrier_nospec() is only
> implemented by x86 and powerpc (see 2nd paragraph). Most/all arches are
> affected by Spectre v1.
Yep agree, if we don't have full arch coverage, then that's a better
reason for getting it included rather than performance reasons.
--
Jens Axboe
Powered by blists - more mailing lists