Bug#1036755: linux: 6.1.26 <= x < 6.1.30 breaks applications using mmap(MAP_32BIT)

[Salvatore Bonaccorso <carnil@debian.org>]

On Thu, May 25, 2023 at 07:12:47PM +0200, Salvatore Bonaccorso wrote:
> Control: tags -1 + confirmed pending
> 
> Hi Apollon,
> 
> On Thu, May 25, 2023 at 03:26:50PM +0300, Apollon Oikonomopoulos wrote:
> > Source: linux
> > Version: 6.1.27-1
> > Severity: critical
> > Tags: upstream
> > Affects: ganeti
> > Justification: breaks unrelated software
> > 
> > Dear Kernel Maintainers,
> > 
> > A regression was introduced with upstream commit 58c5d0d6d5,  
> > subsequently backported to 6.1 as 7e6631f7 and released with 6.1.26.  
> > These commits break MAP_32BIT and affect applications relying on it. In 
> > particular, we are seeing failures in ganeti's autopkgtests[1], where a 
> > daemon written in Haskell (the GHC linker uses MAP_32BIT) segfaults with 
> > a null pointer dereference. From the build artifacts:
> > 
> >  ganeti-luxid: mmap 4096 bytes at (nil): Cannot allocate memory
> >  ganeti-luxid: Try specifying an address with +RTS -xm<addr> -RTS
> > 
> > This has already been discussed on GHC's discourse[2] and has been fixed 
> > upstream, with the fix queued for 6.1.30[3]. Please consider upgrading 
> > to 6.1.30 or cherry-picking the fix for Bookworm.
> 
> I'm not yet certain, that critical severity is right, and making it
> RC. But will look into the issue today.
> 
> Defitnively the issue will be fixed. Just at this stage for the
> bookworm release I will pick it for the first upload after the release
> in a point release update or a security update (whatever will come
> first likely).
> 
> In any case I'm already preparing the 6.1.30 import for bookworm.

For the record,
https://lore.kernel.org/linux-mm/cb8dc31a-fef2-1d09-f133-e9f7b9f9e77a@sony.com/
gives a reproducer/verifier for the issue.

Regards,
Salvatore