[PATCH v13 57/70] mm/mlock: use vma iterator and maple state instead of vma linked list

[Catalin Marinas]

On Thu, Aug 25, 2022 at 01:21:01PM +0000, Liam Howlett wrote:
> * Mark Brown <broonie at kernel.org> [220824 20:34]:
> > On Mon, Aug 22, 2022 at 03:06:30PM +0000, Liam Howlett wrote:
> > > From: "Matthew Wilcox (Oracle)" <willy at infradead.org>
> > > 
> > > Handle overflow checking in count_mm_mlocked_page_nr() differently.
> > 
> > Our QA team found that since next-20220823 we're seeing a couple of test
> > failures in the check_mmap_options kselftest on arm64 platforms with MTE
> > that aren't present in mainline:
> > 
> > # # FAIL: mprotect not ignoring clear PROT_MTE property
> > # not ok 21 Check clear PROT_MTE flags with private mapping, sync error mode and mmap memory
> > # # FAIL: mprotect not ignoring clear PROT_MTE property
> > # not ok 22 Check clear PROT_MTE flags with private mapping and sync error mode and mmap/mprotect memory
> 
> Thanks.
> 
> > I bisected this using qemu[1] which landed on 4ceb4bca479d41a
> > ("mm/mprotect: use maple tree navigation instead of vma linked list"),
> > though I'm not 100% sure I trust the specific identification of the
> > commit I'm pretty confident it's at the very least in this series.  I've
> > not done any analysis of the failure beyond getting this bisect result.
> > 
> > [1] qemu -smp cpus=4 -cpu max -machine virt,gic-version=3,mte=on
> 
> This helps a lot.  I think your bisect is accurate:
> 
> ...
>         struct mmu_gather tlb;
> +       MA_STATE(mas, &current->mm->mm_mt, start, start);
>  
>         start = untagged_addr(start);
> ...
> 
> It looks like I search against the tagged address.  I should initialize
> the state to 0 and mas_set(&mas, start) after untagging the address.
> 
> I'll send out a patch once I have recreated and verified this is the
> issue.

Thanks. I did a quick test and untagging start seems to fix the issue (I
was wondering why mprotect() returned -ENOMEM when failing).

-- 
Catalin