Date:   Thu, 8 Apr 2021 01:37:34 -0700
From:   Michel Lespinasse <michel@...pinasse.org>
To:     Matthew Wilcox <willy@...radead.org>
Cc:     Peter Zijlstra <peterz@...radead.org>,
        Michel Lespinasse <michel@...pinasse.org>,
        Linux-MM <linux-mm@...ck.org>,
        Laurent Dufour <ldufour@...ux.ibm.com>,
        Michal Hocko <mhocko@...e.com>,
        Rik van Riel <riel@...riel.com>,
        Paul McKenney <paulmck@...nel.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Suren Baghdasaryan <surenb@...gle.com>,
        Joel Fernandes <joelaf@...gle.com>,
        Rom Lemarchand <romlem@...gle.com>,
        Linux-Kernel <linux-kernel@...r.kernel.org>
Subject: Re: [RFC PATCH 24/37] mm: implement speculative handling in
 __do_fault()

On Thu, Apr 08, 2021 at 08:13:43AM +0100, Matthew Wilcox wrote:
> On Thu, Apr 08, 2021 at 09:00:26AM +0200, Peter Zijlstra wrote:
> > On Wed, Apr 07, 2021 at 10:27:12PM +0100, Matthew Wilcox wrote:
> > > Doing I/O without any lock held already works; it just uses the file
> > > refcount.  It would be better to use a vma refcount, as I already said.
> > 
> > The original workload that I developed SPF for (waaaay back when) was
> > prefaulting a single huge vma. Using a vma refcount was a total loss
> > because it resulted in the same cacheline contention that down_read()
> > was having.
> > 
> > As such, I'm always incredibly sad to see mention of vma refcounts.
> > They're fundamentally not solving the problem :/
> 
> OK, let me outline my locking scheme because I think it's rather better
> than Michel's.  The vma refcount is the slow path.
> 
> 1. take the RCU read lock
> 2. walk the pgd/p4d/pud/pmd
> 3. allocate page tables if necessary.  *handwave GFP flags*.
> 4. walk the vma tree
> 5. call ->map_pages
> 6. take ptlock
> 7. insert page(s)
> 8. drop ptlock
> if this all worked out, we're done, drop the RCU read lock and return.
> 9. increment vma refcount
> 10. drop RCU read lock
> 11. call ->fault
> 12. decrement vma refcount

Note that most of your proposed steps seem similar in principle to mine.
Looking at the fast path (steps 1-8):
- step 2 sounds like the speculative part of __handle_mm_fault()
- (step 3 not included in my proposal)
- step 4 is basically the lookup I currently have in the arch fault handler
- step 6 sounds like the speculative part of map_pte_lock()

I have working implementations for each step, while your proposal
summarizes each as a point item. It's not clear to me what to make of it;
presumably you would be "filling in the blanks" in a different way
than I have but you are not explaining how. Are you suggesting that
the precautions taken in each step to avoid races with mmap writers
would not be necessary in your proposal ? if that is the case, what is
the alternative mechanism would you use to handle such races ?

Going back to the source of this, you suggested not copying the VMA,
what is your proposed alternative ? Do you suggest that fault handlers
should deal with the vma potentially mutating under them ? Or should
mmap writers consider vmas as immutable and copy them whenever they
want to change them ? or are you implying a locking mechanism that would
prevent mmap writers from executing while the fault is running ?

> Compared to today, where we bump the refcount on the file underlying the
> vma, this is _better_ scalability -- different mappings of the same file
> will not contend on the file's refcount.
> 
> I suspect your huge VMA was anon, and that wouldn't need a vma refcount
> as faulting in new pages doesn't need to do I/O, just drop the RCU
> lock, allocate and retry.

Powered by blists - more mailing lists