[RFC PATCH 0/2] iommu: Avoid unnecessary PRI queue flushes
jean-philippe at linaro.org
Sat Oct 24 00:34:23 AEDT 2020
On Mon, Oct 19, 2020 at 02:16:08PM -0700, Raj, Ashok wrote:
> Hi Jean
> On Mon, Oct 19, 2020 at 04:08:24PM +0200, Jean-Philippe Brucker wrote:
> > On Sat, Oct 17, 2020 at 04:25:25AM -0700, Raj, Ashok wrote:
> > > > For devices that *don't* use a stop marker, the PCIe spec says (10.4.1.2):
> > > >
> > > > To stop [using a PASID] without using a Stop Marker Message, the
> > > > function shall:
> > > > 1. Stop queueing new Page Request Messages for this PASID.
> > >
> > > The device driver would need to tell stop sending any new PR's.
> > >
> > > > 2. Finish transmitting any multi-page Page Request Messages for this
> > > > PASID (i.e. send the Page Request Message with the L bit Set).
> > > > 3. Wait for PRG Response Messages associated any outstanding Page
> > > > Request Messages for the PASID.
> > > >
> > > > So they have to flush their PR themselves. And since the device driver
> > > > completes this sequence before calling unbind(), then there shouldn't be
> > > > any oustanding PR for the PASID, and unbind() doesn't need to flush,
> > > > right?
> > >
> > > I can see how the device can complete #2,3 above. But the device driver
> > > isn't the one managing page-responses right. So in order for the device to
> > > know the above sequence is complete, it would need to get some assist from
> > > IOMMU driver?
> > No the device driver just waits for the device to indicate that it has
> > completed the sequence. That's what the magic stop-PASID mechanism
> > described by PCIe does. In 6.20.1 "Managing PASID TLP Prefix Usage" it
> > says:
> The goal is we do this when the device is in a messup up state. So we can't
> take for granted the device is properly operating which is why we are going
> to wack the device with a flr().
> The only thing that's supposed to work without a brain freeze is the
> invalidation logic. Spec requires devices to respond to invalidations even when
> they are in the process of flr().
> So when IOMMU does an invalidation wait with a Page-Drain, IOMMU waits till
> the response for that arrives before completing the descriptor. Due to
> the posted semantics it will ensure any PR's issued and in the fabric are flushed
> out to memory.
> I suppose you can wait for the device to vouch for all responses, but that
> is assuming the device is still functioning properly. Given that we use it
> in two places,
> * Reclaiming a PASID - only during a tear down sequence, skipping it
> doesn't really buy us much.
Yes I was only wondering about normal PASID reclaim operations, in
unbind(). Agreed that for FLR we need to properly clean the queue, though
I do need to do more thinking about this.
Anyway, having a full priq drain in unbind() isn't harmful, just
unnecessary delay in my opinion. I'll drop these patches for now but
thanks for the discussion.
> * During FLR this can't be skipped anyway due to the above sequence
> > "A Function must have a mechanism to request that it gracefully stop using
> > a specific PASID. This mechanism is device specific but must satisfy the
> > following rules:
> > [...]
> > * When the stop request mechanism indicates completion, the Function has:
> > [...]
> > * Complied with additional rules described in Address Translation
> > Services (Chapter 10 [10.4.1.2 quoted above]) if Address Translations
> > or Page Requests were issued on the behalf of this PASID."
> > So after the device driver initiates this mechanism in the device, the
> > device must be able to indicate completion of the mechanism, which
> > includes completing all in-flight Page Requests. At that point the device
> > driver can call unbind() knowing there is no pending PR for this PASID.
More information about the Linux-accelerators