Re: mremap vs sysctl_max_map_count

["Kirill A. Shutemov" <kirill@shutemov.name>]

On Mon, Feb 18, 2019 at 10:57:18AM +0100, Vlastimil Babka wrote:
> On 2/18/19 9:33 AM, Oscar Salvador wrote:
> > 
> > Hi all,
> > 
> > I would like to bring up a topic that comes from an issue a customer of ours
> > is facing with the mremap syscall + hitting the max_map_count threshold:
> > 
> > When passing the MREMAP_FIXED flag, mremap() calls mremap_to() which does the
> > following:
> > 
> > 1) it unmaps the region where we want to put the new map:
> >    (new_addr, new_addr + new_len] [1]
> > 2) IFF old_len > new_len, it unmaps the region:
> >    (old_addr + new_len, (old_addr + new_len) + (old_len - new_len)] [2]
> > 
> > Now, having gone through steps 1) and 2), we eventually call move_vma() to do
> > the actual move.
> > 
> > move_vma() checks if we are at least 4 maps below max_map_count, otherwise
> > it bails out with -ENOMEM [3].
> > The problem is that we might have already unmapped the vma's in steps 1) and 2),
> > so it is not possible for userspace to figure out the state of the vma's after
> > it gets -ENOMEM.
> > 
> > - Did new_addr got unmaped?
> > - Did part of the old_addr got unmaped?
> > 
> > Because of that, it gets tricky for userspace to clean up properly on error
> > path.
> > 
> > While it is true that we can return -ENOMEM for more reasons
> > (e.g: see vma_to_resize()->may_expand_vm()), I think that we might be able to
> > pre-compute the number of maps that we are going add/release during the first
> > two do_munmaps(), and check whether we are 4 maps below the threshold
> > (as move_vma() does).
> > Should not be the case, we can bail out early before we unmap anything, so we
> > make sure the vma's are left untouched in case we are going to be short of maps.
> > 
> > I am not sure if that is realistically doable, or there are limitations
> > I overlooked, or we simply do not want to do that.
> 
> IMHO it makes sense to do all such resource limit checks upfront. It
> should all be protected by mmap_sem and thus stable, right? Even if it
> was racy, I'd think it's better to breach the limit a bit due to a race
> than bail out in the middle of operation. Being also resilient against
> "real" ENOMEM's due to e.g. failure to alocate a vma would be much
> harder perhaps (but maybe it's already mostly covered by the
> too-small-to-fail in page allocator), but I'd try with the artificial
> limits at least.

There's slight chance of false-postive -ENOMEM with upfront approach:
unmapping can reduce number of VMAs so in some cases upfront check would
fail something that could succeed otherwise.

We could check also what number of VMA unmap would free (if any). But it
complicates the picture and I don't think worth it in the end.

-- 
 Kirill A. Shutemov