Re: Root filesystem read access for firmware load during hibernation image writing

["Darrick J. Wong" <djwong@kernel.org>]

On Fri, Oct 11, 2024 at 04:11:49PM -0700, Luis Chamberlain wrote:
> On Sat, Oct 05, 2024 at 03:16:27PM +0200, Maciej S. Szmigiero wrote:
> > The issue below still happens on kernel version 6.11.1.
> > 
> > Created a kernel bugzilla entry for it:
> > https://bugzilla.kernel.org/show_bug.cgi?id=219353
> > 
> > It would be nice to at least know whether the filesystem read access supposed is
> > to be working normally at PMSG_THAW hibernation stage to assign the issue accordingly.
> 
> No, there are races possible if you trigger IO to a fs before a suspend
> is going on. If you have *more* IO ongoing, then you are likely to stall
> suspend and not be able to recover.
> 
> > CC: request_firmware() maintainers
> 
> If IO is ongoing suspend is broken today because of the kthread freezer
> which puts kthreads which should sleep idle, but if IO is ongoing we
> can stall. This is work which we've been working to remove for years and
> after its removal we can gracefully freeze filesystems [0] [1]
> properly on suspend.
> 
> Other than that, obviously upon resume we want firmware to be present,
> and to prevent races on resume we have a firmware cache. So on suspend
> we cache used firmware so its available in cache on resume. See
> device_cache_fw_images().

Reads are generally supposed to succeed on a suspended system but if the
read causes a write (e.g. atime update) they can block just like a write
would.  So in the end you probably have to cache the firmware blobs in a
tmpfs or something like that.

> So.. we just now gotta respin the latest effort. I had stopped because
> I know Darrick had some changes which he needed to get in sooner but

Do you remember what changes those were?  I don't. :(

--D

> since this is low hanging fruit I never got around to it. So someone
> just needs to respin the series.
> 
> [0] https://lwn.net/Articles/752588/
> [1] https://lwn.net/Articles/935602/
> 
>   Luis