On Sat, Jan 11, 2025 at 12:14 PM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Sat, Jan 11, 2025 at 3:24 AM Mateusz Guzik <mjguzik@xxxxxxxxx> wrote:
> >
> > On Fri, Jan 10, 2025 at 08:25:58PM -0800, Suren Baghdasaryan wrote:
> >
> > So there were quite a few iterations of the patch and I have not been
> > reading majority of the feedback, so it may be I missed something,
> > apologies upfront. :)
> >
> > > /*
> > > * Try to read-lock a vma. The function is allowed to occasionally yield false
> > > * locked result to avoid performance overhead, in which case we fall back to
> > > @@ -710,6 +742,8 @@ static inline void vma_lock_init(struct vm_area_struct *vma)
> > > */
> > > static inline bool vma_start_read(struct vm_area_struct *vma)
> > > {
> > > + int oldcnt;
> > > +
> > > /*
> > > * Check before locking. A race might cause false locked result.
> > > * We can use READ_ONCE() for the mm_lock_seq here, and don't need
> > > @@ -720,13 +754,19 @@ static inline bool vma_start_read(struct vm_area_struct *vma)
> > > if (READ_ONCE(vma->vm_lock_seq) == READ_ONCE(vma->vm_mm->mm_lock_seq.sequence))
> > > return false;
> > >
> > > - if (unlikely(down_read_trylock(&vma->vm_lock.lock) == 0))
> > > + /*
> > > + * If VMA_LOCK_OFFSET is set, __refcount_inc_not_zero_limited() will fail
> > > + * because VMA_REF_LIMIT is less than VMA_LOCK_OFFSET.
> > > + */
> > > + if (unlikely(!__refcount_inc_not_zero_limited(&vma->vm_refcnt, &oldcnt,
> > > + VMA_REF_LIMIT)))
> > > return false;
> > >
> >
> > Replacing down_read_trylock() with the new routine loses an acquire
> > fence. That alone is not a problem, but see below.
>
> Hmm. I think this acquire fence is actually necessary. We don't want
> the later vm_lock_seq check to be reordered and happen before we take
> the refcount. Otherwise this might happen:
>
> reader writer
> if (vm_lock_seq == mm_lock_seq) // check got reordered
> return false;
> vm_refcnt += VMA_LOCK_OFFSET
> vm_lock_seq == mm_lock_seq
s/vm_lock_seq == mm_lock_seq/vm_lock_seq = mm_lock_seq
> vm_refcnt -= VMA_LOCK_OFFSET
> if (!__refcount_inc_not_zero_limited())
> return false;
>
> Both reader's checks will pass and the reader would read-lock a vma
> that was write-locked.
>
> >
> > > + rwsem_acquire_read(&vma->vmlock_dep_map, 0, 1, _RET_IP_);
> > > /*
> > > - * Overflow might produce false locked result.
> > > + * Overflow of vm_lock_seq/mm_lock_seq might produce false locked result.
> > > * False unlocked result is impossible because we modify and check
> > > - * vma->vm_lock_seq under vma->vm_lock protection and mm->mm_lock_seq
> > > + * vma->vm_lock_seq under vma->vm_refcnt protection and mm->mm_lock_seq
> > > * modification invalidates all existing locks.
> > > *
> > > * We must use ACQUIRE semantics for the mm_lock_seq so that if we are
> > > @@ -735,9 +775,10 @@ static inline bool vma_start_read(struct vm_area_struct *vma)
> > > * This pairs with RELEASE semantics in vma_end_write_all().
> > > */
> > > if (unlikely(vma->vm_lock_seq == raw_read_seqcount(&vma->vm_mm->mm_lock_seq))) {
> >
> > The previous modification of this spot to raw_read_seqcount loses the
> > acquire fence, making the above comment not line up with the code.
>
> Is it? From reading the seqcount code
> (https://elixir.bootlin.com/linux/v6.13-rc3/source/include/linux/seqlock.h#L211):
>
> raw_read_seqcount()
> seqprop_sequence()
> __seqprop(s, sequence)
> __seqprop_sequence()
> smp_load_acquire()
>
> smp_load_acquire() still provides the acquire fence. Am I missing something?
>
> >
> > I don't know if the stock code (with down_read_trylock()) is correct as
> > is -- looks fine for cursory reading fwiw. However, if it indeed works,
> > the acquire fence stemming from the lock routine is a mandatory part of
> > it afaics.
> >
> > I think the best way forward is to add a new refcount routine which
> > ships with an acquire fence.
>
> I plan on replacing refcount_t usage here with an atomic since, as
> Hillf noted, refcount is not designed to be used for locking. And will
> make sure the down_read_trylock() replacement will provide an acquire
> fence.
>
> >
> > Otherwise I would suggest:
> > 1. a comment above __refcount_inc_not_zero_limited saying there is an
> > acq fence issued later
> > 2. smp_rmb() slapped between that and seq accesses
> >
> > If the now removed fence is somehow not needed, I think a comment
> > explaining it is necessary.
> >
> > > @@ -813,36 +856,33 @@ static inline void vma_assert_write_locked(struct vm_area_struct *vma)
> > >
> > > static inline void vma_assert_locked(struct vm_area_struct *vma)
> > > {
> > > - if (!rwsem_is_locked(&vma->vm_lock.lock))
> > > + if (refcount_read(&vma->vm_refcnt) <= 1)
> > > vma_assert_write_locked(vma);
> > > }
> > >
> >
> > This now forces the compiler to emit a load from vm_refcnt even if
> > vma_assert_write_locked expands to nothing. iow this wants to hide
> > behind the same stuff as vma_assert_write_locked.
>
> True. I guess I'll have to avoid using vma_assert_write_locked() like this:
>
> static inline void vma_assert_locked(struct vm_area_struct *vma)
> {
> unsigned int mm_lock_seq;
>
> VM_BUG_ON_VMA(refcount_read(&vma->vm_refcnt) <= 1 &&
> !__is_vma_write_locked(vma,
> &mm_lock_seq), vma);
> }
>
> Will make the change.
>
> Thanks for the feedback!
