Re: [RFC PATCH 00/16] mm/madvise: batch tlb flushes for MADV_DONTNEED and MADV_FREE

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

On Wed, Mar 05, 2025 at 08:35:36PM +0100, David Hildenbrand wrote:
> On 05.03.25 20:26, Lorenzo Stoakes wrote:
> > On Wed, Mar 05, 2025 at 08:19:41PM +0100, David Hildenbrand wrote:
> > > On 05.03.25 19:56, Matthew Wilcox wrote:
> > > > On Wed, Mar 05, 2025 at 10:15:55AM -0800, SeongJae Park wrote:
> > > > > For MADV_DONTNEED[_LOCKED] or MADV_FREE madvise requests, tlb flushes
> > > > > can happen for each vma of the given address ranges.  Because such tlb
> > > > > flushes are for address ranges of same process, doing those in a batch
> > > > > is more efficient while still being safe.  Modify madvise() and
> > > > > process_madvise() entry level code path to do such batched tlb flushes,
> > > > > while the internal unmap logics do only gathering of the tlb entries to
> > > > > flush.
> > > >
> > > > Do real applications actually do madvise requests that span multiple
> > > > VMAs?  It just seems weird to me.  Like, each vma comes from a separate
> > > > call to mmap [1], so why would it make sense for an application to
> > > > call madvise() across a VMA boundary?
> > >
> > > I had the same question. If this happens in an app, I would assume that a
> > > single MADV_DONTNEED call would usually not span multiples VMAs, and if it
> > > does, not that many (and that often) that we would really care about it.
> > >
> > > OTOH, optimizing tlb flushing when using a vectored MADV_DONTNEED version
> > > would make more sense to me. I don't recall if process_madvise() allows for
> > > that already, and if it does, is this series primarily tackling optimizing
> > > that?
> >
> > Yeah it's weird, but people can get caught out by unexpected failures to merge
> > if they do fun stuff with mremap().
> >
> > Then again mremap() itself _mandates_ that you only span a single VMA (or part
> > of one) :)
>
> Maybe some garbage collection use cases that shuffle individual pages, and
> later free larger chunks using MADV_DONTNEED. Doesn't sound unlikely.
>
> >
> > Can we talk about the _true_ horror show that - you can span multiple VMAs _with
> > gaps_ and it'll allow you, only it'll return -ENOMEM at the end?
> >
> > In madvise_walk_vmas():
> >
> > 	for (;;) {
> > 		...
> >
> > 		if (start < vma->vm_start) {
> > 			unmapped_error = -ENOMEM;
> > 			start = vma->vm_start;
> > 			...
> > 		}
> >
> > 		...
> >
> > 		error = visit(vma, &prev, start, tmp, arg);
> > 		if (error)
> > 			return error;
> >
> > 		...
> > 	}
> >
> > 	return unmapped_error;
> >
> > So, you have no idea if that -ENOMEM is due to a gap, or do to the
> > operation returning an -ENOMEM?
> > > I mean can we just drop this? Does anybody in their right mind rely on
> > this? Or is it intentional to deal with somehow a racing unmap?
> > > But, no, we hold an mmap lock so that's not it.
>
> Races could still happen if user space would do it from separate threads.
> But then, who would prevent user space from doing another mmap() and getting
> pages zapped ... so that sounds unlikely.

Ah yeah, I mean if you got unlucky on timing, munmap()/mmap() or mremap() with
MAP_FIXED quickly unmaps on another thread before you grab the mmap lock and fun
times.

I mean for that to happen you'd have to be doing something... very odd or insane
so. But still.

>
> >
> > Yeah OK so can we drop this madness? :) or am I missing some very important
> > detail about why we allow this?
>
> I stumbled over that myself a while ago. It's well documented behavior in
> the man page :(

Haha ok I guess we have to live with it then.

>
> At that point I stopped caring, because apparently somebody else cared
> enough to document that clearly in the man page :)
>
> >
> > I guess spanning multiple VMAs we _have_ to leave in because plausibly
> > there are users of that out there?
>
> Spanning multiple VMAs can probably easily happen. At least in QEMU I know
> some sane ways to trigger it on guest memory. But, all corner cases, so
> nothing relevant for performance.
>
>
> --
> Cheers,
>
> David / dhildenb
>