Re: [PATCH v3 26/35] mm: fall back to mmap_lock if vma->anon_vma is not yet set

[Matthew Wilcox]

On Fri, Feb 17, 2023 at 08:10:35AM -0800, Suren Baghdasaryan wrote:
> On Fri, Feb 17, 2023 at 8:05 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >
> > On Thu, Feb 16, 2023 at 06:14:59PM -0800, Suren Baghdasaryan wrote:
> > > On Thu, Feb 16, 2023 at 11:43 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> > > >
> > > > On Thu, Feb 16, 2023 at 7:44 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Wed, Feb 15, 2023 at 09:17:41PM -0800, Suren Baghdasaryan wrote:
> > > > > > When vma->anon_vma is not set, page fault handler will set it by either
> > > > > > reusing anon_vma of an adjacent VMA if VMAs are compatible or by
> > > > > > allocating a new one. find_mergeable_anon_vma() walks VMA tree to find
> > > > > > a compatible adjacent VMA and that requires not only the faulting VMA
> > > > > > to be stable but also the tree structure and other VMAs inside that tree.
> > > > > > Therefore locking just the faulting VMA is not enough for this search.
> > > > > > Fall back to taking mmap_lock when vma->anon_vma is not set. This
> > > > > > situation happens only on the first page fault and should not affect
> > > > > > overall performance.
> > > > >
> > > > > I think I asked this before, but don't remember getting an aswer.
> > > > > Why do we defer setting anon_vma to the first fault?  Why don't we
> > > > > set it up at mmap time?
> > > >
> > > > Yeah, I remember that conversation Matthew and I could not find the
> > > > definitive answer at the time. I'll look into that again or maybe
> > > > someone can answer it here.
> > >
> > > After looking into it again I'm still under the impression that
> > > vma->anon_vma is populated lazily (during the first page fault rather
> > > than at mmap time) to avoid doing extra work for areas which are never
> > > faulted. Though I might be missing some important detail here.
> >
> > How often does userspace call mmap() and then _never_ fault on it?
> > I appreciate that userspace might mmap() gigabytes of address space and
> > then only end up using a small amount of it, so populating it lazily
> > makes sense.  But creating a region and never faulting on it?  The only
> > use-case I can think of is loading shared libraries:
> >
> > openat(AT_FDCWD, "/lib/x86_64-linux-gnu/libc.so.6", O_RDONLY|O_CLOEXEC) = 3
> > (...)
> > mmap(NULL, 1970000, PROT_READ, MAP_PRIVATE|MAP_DENYWRITE, 3, 0) = 0x7f0ce612e000
> > mmap(0x7f0ce6154000, 1396736, PROT_READ|PROT_EXEC, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x26000) = 0x7f0ce6154000
> > mmap(0x7f0ce62a9000, 339968, PROT_READ, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x17b000) = 0x7f0ce62a9000
> > mmap(0x7f0ce62fc000, 24576, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x1ce000) = 0x7f0ce62fc000
> > mmap(0x7f0ce6302000, 53072, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x7f0ce6302000
> >
> > but that's a file-backed VMA, not an anon VMA.
> 
> Might the case of dup_mmap() while forking be the reason why a VMA in
> the child process might be never used while parent uses it (or visa
> versa)? Again, I'm not sure this is the reason but I can find no other
> good explanation.

I found an explanation!  Well, a partial one.  If we MAP_PRIVATE a file
mapping (like, er those ones up there) and only take read faults on it,
we can postpone allocation of the anon_vma indefinitely.  But once we
take a write fault in that VMA, we need to allocate an anon_vma for it
so that we can track the anonymous pages that have been allocated to
satisfy the copy-on-write (see do_cow_fault()).

However, I think in that caase, we could probably skip the
find_mergeable_anon_vma() step.  We don't today; we check whether
a->vm_file == b->vm_file in anon_vma_compatible, but I wonder if that
triggers often.