Re: [PATCH hotfix 6.12] maple_tree: correct tree corruption on spanning store

[Lorenzo Stoakes <lorenzo.stoakes@oracle.com>]

On Sat, Oct 05, 2024 at 03:24:39PM +0200, Bert Karwatzki wrote:
> Am Samstag, dem 05.10.2024 um 12:17 +0100 schrieb Lorenzo Stoakes:
> > On Sat, Oct 05, 2024 at 07:41:14AM +0100, Lorenzo Stoakes wrote:
> > > Writing a data range into a maple tree may involve overwriting a number of
> > > existing entries that span across more than one node. Doing so invokes a
> > > 'spanning' store.
> > >
> >
> > [snip]
> >
> > Andrew - just to note that I have intentionally left stable off this, in
> > order that wre can allow this to stabilise in the 6.12 release candidates.
> >
> > Up until 6.12 this bug seemed much harder to hit, and as far as I'm aware
> > we've never had a bug report for it prior to this.
>
> I still suspect that this could have been the same error:
> https://lkml.org/lkml/2024/8/28/1558
> When compiling the kernel without CONFIG_DEBUG_VM maple tree bug results in an
> unkillable task, and when trying to kill it first produced the rwsem warning
> (and soon after took down the whole system).
> But I couldn't reproduce it with the given reproducer, either.
>
> Bert Karwatzki
>

Thanks for reminding me of that one!

Yeah unfortunately that thread was very unproductive in that the reporter
gave no feedback or further information. They spammed the list with a bunch
of such reports many looking suspect...

So it is possible, and I suspect that this bug may have caused some other
'weird' crashes that were non-repro in the past.

The difference here may be that we (or rather specifically - you! :)
finally found a way to reliable repro this to the point where we could
diagnose it.

As far as I can tell this could happen even with vma_iter_clear_gfp(), so
old unmap/MAP_FIXED behaviour could have hit it.

BUT a difference now is that we essentially combine the MAP_FIXED with a
merge and overwrite everything in between, so this addition of up to 2
extra entries probably pushed it over the edge to make this event
statistically likely enough for you to have hit it.

Note in your case it took unmapping 6 (!) entries and merging another 2 for
a total of an overwrite spanning 8 entries.

Anyway, assuming we so no issues stabilising this in the rc's I will ping
stable to get this backported and fix this everywhere.