Re: [PATCH] swiotlb: fix the check whether a device has used software IO TLB

From: Petr Tesařík
Date: Fri Sep 22 2023 - 14:21:09 EST


On Fri, 22 Sep 2023 19:12:13 +0200
Petr Tesařík <petr@xxxxxxxxxxx> wrote:

> Hi Catalin,
>
> On Fri, 22 Sep 2023 15:31:29 +0200
> Petr Tesařík <petr@xxxxxxxxxxx> wrote:
>
> >[...]
> > On Mon, 18 Sep 2023 16:45:34 +0100
> > Catalin Marinas <catalin.marinas@xxxxxxx> wrote:
> >
> > > On Sun, Sep 17, 2023 at 11:47:41AM +0200, Petr Tesařík wrote:
> >[...]
> > > > Ah... You may have a point after all if this sequence of events is
> > > > possible:
> > > >
> > > > - CPU 0 writes new value to mem->pools->next in swiotlb_dyn_alloc().
> > > >
> > > > - CPU 1 observes the new value in swiotlb_find_slots(), even though it
> > > > is not guaranteed by any barrier, allocates a slot and sets the
> > > > dev->dma_uses_io_tlb flag.
> > > >
> > > > - CPU 1 (driver code) writes the returned buffer address into its
> > > > private struct. This write is ordered after dev->dma_uses_io_tlb
> > > > thanks to the smp_wmb() in swiotlb_find_slots().
> > > >
> > > > - CPU 2 (driver code) reads the buffer address, and DMA core passes it
> > > > to is_swiotlb_buffer(), which contains smp_rmb().
> > > >
> > > > - IIUC CPU 2 is guaranteed to observe the new value of
> > > > dev->dma_uses_io_tlb, but it may still use the old value of
> > > > mem->pools->next, because the write on CPU 0 was not ordered
> > > > against anything. The fact that the new value was observed by CPU 1
> > > > does not mean that it is also observed by CPU 2.
> > >
> > > Yes, that's possible. On CPU 1 there is a control dependency between the
> > > read of mem->pools->next and the write of dev->dma_uses_io_tlb but I
> > > don't think this is sufficient to claim multi-copy atomicity (if CPU 1
> > > sees mem->pools->next write by CPU 0, CPU 2 must see it as well), at
> > > least not on all architectures supported by Linux. memory-barriers.txt
> > > says that a full barrier on CPU 1 is needed between the read and write,
> > > i.e. smp_mb() before WRITE_ONCE(dev->dma_uses_io_tlb). You could add it
> > > just before "goto found" in swiotlb_find_slots() since it's only needed
> > > on this path.
> >
> > Let me check my understanding. This smp_mb() is not needed to make sure
> > that the write to dev->dma_uses_io_tlb cannot be visible before the
> > read of mem->pools->next. Since stores are not speculated, that
> > ordering is provided by the control dependency alone.
> >
> > But a general barrier ensures that a third CPU will observe the write to
> > mem->pools->next after the read of mem->pools->next. Makes sense.
>
> Now that I'm writing the patch, I get your idea to replace WRITE_ONCE()
> with smp_store_release(). Since a full memory barrier is required for
> multicopy atomicity, it is not "more than I need". Instead, the
> ordering contraints may be possibly restricted to "CPUs participating
> in a release-acquire chain" if I also replace READ_ONCE() in
> is_swiotlb_buffer() with smp_read_acquire().
>
> I believe it does not matter that the CPU which writes a new value to
> mem->pools->next in swiotlb_dyn_alloc() does not participate in the
> chain, because the value must have been visible to the CPU which
> executes swiotlb_find_slots() and which does participate in the chain.
>
> Let me double-check this thinking with a litmus test.

Hm. I didn't have much luck with smp_store_release(), because I need
to ensure ordering of the SUBSEQUENT store (by a device driver).

However, inserting smp_mb() _after_ WRITE_ONCE(dev->dma_uses_io_tlb)
seems to be enough to ensure proper ordering. I could even remove the
write memory barrier in swiotlb_dyn_alloc().

This is my first time using herd7, so I can only hope I got everything
right. FWIW this is my litmus test:

C swiotlb-new-pool

(*
* Result: Never
*
* Check that a newly allocated pool is always visible when the corresponding
* swiotlb buffer is visible.
*)

{}

P0(int *pool)
{
/* swiotlb_dyn_alloc() */
WRITE_ONCE(*pool, 999);
}

P1(int *pool, int *flag, int *buf)
{
/* swiotlb_find_slots() */
int r0 = READ_ONCE(*pool);
if (r0) {
WRITE_ONCE(*flag, 1);
smp_mb();
}

/* device driver (presumed) */
WRITE_ONCE(*buf, r0);
}

P2(int *pool, int *flag, int *buf)
{
/* device driver (presumed) */
int r1 = READ_ONCE(*buf);

/* is_swiotlb_buffer() */
int r2;
int r3;

smp_rmb();
r2 = READ_ONCE(*flag);
if (r2) {
r3 = READ_ONCE(*pool);
}
}

exists (2:r1<>0 /\ 2:r3=0) (* Not flagged or pool not found. *)

Petr T