On Wed, Sep 11, 2024 at 2:35 PM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Mon, Sep 9, 2024 at 7:09 PM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> >
> > On Mon, Sep 9, 2024 at 5:35 AM Jann Horn <jannh@xxxxxxxxxx> wrote:
> > >
> > > On Fri, Sep 6, 2024 at 7:12 AM Andrii Nakryiko <andrii@xxxxxxxxxx> wrote:
> > > > +static inline bool mmap_lock_speculation_end(struct mm_struct *mm, int seq)
> > > > +{
> > > > + /* Pairs with RELEASE semantics in inc_mm_lock_seq(). */
> > > > + return seq == smp_load_acquire(&mm->mm_lock_seq);
> > > > +}
> > >
> > > A load-acquire can't provide "end of locked section" semantics - a
> > > load-acquire is a one-way barrier, you can basically use it for
> > > "acquire lock" semantics but not for "release lock" semantics, because
> > > the CPU will prevent reordering the load with *later* loads but not
> > > with *earlier* loads. So if you do:
> > >
> > > mmap_lock_speculation_start()
> > > [locked reads go here]
> > > mmap_lock_speculation_end()
> > >
> > > then the CPU is allowed to reorder your instructions like this:
> > >
> > > mmap_lock_speculation_start()
> > > mmap_lock_speculation_end()
> > > [locked reads go here]
> > >
> > > so the lock is broken.
> >
> > Hi Jann,
> > Thanks for the review!
> > Yeah, you are right, we do need an smp_rmb() before we compare
> > mm->mm_lock_seq with the stored seq.
> >
> > Otherwise reads might get reordered this way:
> >
> > CPU1 CPU2
> > mmap_lock_speculation_start() // seq = mm->mm_lock_seq
> > reloaded_seq = mm->mm_lock_seq; // reordered read
> > mmap_write_lock() // inc_mm_lock_seq(mm)
> > vma->vm_file = ...;
> > mmap_write_unlock() // inc_mm_lock_seq(mm)
> > <speculate>
> > mmap_lock_speculation_end() // return (reloaded_seq == seq)
> >
> > >
> > > > static inline void mmap_write_lock(struct mm_struct *mm)
> > > > {
> > > > __mmap_lock_trace_start_locking(mm, true);
> > > > down_write(&mm->mmap_lock);
> > > > + inc_mm_lock_seq(mm);
> > > > __mmap_lock_trace_acquire_returned(mm, true, true);
> > > > }
> > >
> > > Similarly, inc_mm_lock_seq(), which does a store-release, can only
> > > provide "release lock" semantics, not "take lock" semantics, because
> > > the CPU can reorder it with later stores; for example, this code:
> > >
> > > inc_mm_lock_seq()
> > > [locked stuff goes here]
> > > inc_mm_lock_seq()
> > >
> > > can be reordered into this:
> > >
> > > [locked stuff goes here]
> > > inc_mm_lock_seq()
> > > inc_mm_lock_seq()
> > >
> > > so the lock is broken.
> >
> > Ugh, yes. We do need smp_wmb() AFTER the inc_mm_lock_seq(). Whenever
>
> Suren, can you share with me an updated patch for mm_lock_seq with the
> right memory barriers? Do you think this might have a noticeable
> impact on performance? What sort of benchmark do mm folks use to
> quantify changes like that?
Yes, I think I can get it to you before leaving for LPC.
It might end up affecting paths where we take mmap_lock for write
(mmap/munmap/mprotect/etc) but these are not considered fast paths.
I'll think about possible tests we can run to evaluate it.
>
> > we use inc_mm_lock_seq() for "take lock" semantics, it's preceded by a
> > down_write(&mm->mmap_lock) with implied ACQUIRE ordering. So I thought
> > we can use it but I realize now that this reordering is still
> > possible:
> > CPU1 CPU2
> > mmap_write_lock()
> > down_write(&mm->mmap_lock);
> > vma->vm_file = ...;
> >
> > mmap_lock_speculation_start() // seq = mm->mm_lock_seq
> > <speculate>
> > mmap_lock_speculation_end() // return (mm->mm_lock_seq == seq)
> >
> > inc_mm_lock_seq(mm);
> > mmap_write_unlock() // inc_mm_lock_seq(mm)
> >
> > Is that what you were describing?
> > Thanks,
> > Suren.
> >
> > >
> > > For "taking a lock" with a memory store, or "dropping a lock" with a
> > > memory load, you need heavier memory barriers, see
> > > Documentation/memory-barriers.txt.
