Re: [RFC PATCHES 00/17] IOMMUFD: Deliver IO page faults to user space

Re: [RFC PATCHES 00/17] IOMMUFD: Deliver IO page faults to user space

[Jason Gunthorpe]

On Tue, May 30, 2023 at 01:37:07PM +0800, Lu Baolu wrote:
> Hi folks,
> 
> This series implements the functionality of delivering IO page faults to
> user space through the IOMMUFD framework. The use case is nested
> translation, where modern IOMMU hardware supports two-stage translation
> tables. The second-stage translation table is managed by the host VMM
> while the first-stage translation table is owned by the user space.
> Hence, any IO page fault that occurs on the first-stage page table
> should be delivered to the user space and handled there. The user space
> should respond the page fault handling result to the device top-down
> through the IOMMUFD response uAPI.
> 
> User space indicates its capablity of handling IO page faults by setting
> a user HWPT allocation flag IOMMU_HWPT_ALLOC_FLAGS_IOPF_CAPABLE. IOMMUFD
> will then setup its infrastructure for page fault delivery. Together
> with the iopf-capable flag, user space should also provide an eventfd
> where it will listen on any down-top page fault messages.
> 
> On a successful return of the allocation of iopf-capable HWPT, a fault
> fd will be returned. User space can open and read fault messages from it
> once the eventfd is signaled.

This is a performance path so we really need to think about this more,
polling on an eventfd and then reading a different fd is not a good
design.

What I would like is to have a design from the start that fits into
io_uring, so we can have pre-posted 'recvs' in io_uring that just get
completed at high speed when PRIs come in.

This suggests that the PRI should be delivered via read() on a single
FD and pollability on the single FD without any eventfd.

> Besides the overall design, I'd like to hear comments about below
> designs:
> 
> - The IOMMUFD fault message format. It is very similar to that in
>   uapi/linux/iommu which has been discussed before and partially used by
>   the IOMMU SVA implementation. I'd like to get more comments on the
>   format when it comes to IOMMUFD.

We have to have the same discussion as always, does a generic fault
message format make any sense here?

PRI seems more likely that it would but it needs a big carefull cross
vendor check out.

Jason
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [RFC PATCHES 00/17] IOMMUFD: Deliver IO page faults to user space

[Jason Gunthorpe]

On Fri, Jun 23, 2023 at 02:18:38PM +0800, Baolu Lu wrote:

> 	struct io_uring ring;
> 
> 	io_uring_setup(IOPF_ENTRIES, &ring);
> 
> 	while (1) {
> 		struct io_uring_prep_read read;
> 		struct io_uring_cqe *cqe;
> 
> 		read.fd = iopf_fd;
> 		read.buf = malloc(IOPF_SIZE);
> 		read.len = IOPF_SIZE;
> 		read.flags = 0;
> 
> 		io_uring_prep_read(&ring, &read);
> 		io_uring_submit(&ring);
> 
> 		// Wait for the read to complete
> 		while ((cqe = io_uring_get_cqe(&ring)) != NULL) {
> 			// Check if the read completed
> 			if (cqe->res < 0)
> 				break;
> 
> 			if (page_fault_read_completion(cqe)) {
> 				// Get the fault data
> 				void *data = cqe->buf;
> 				size_t size = cqe->res;
> 
> 				// Handle the page fault
> 				handle_page_fault(data);
> 
> 				// Respond the fault
> 				struct io_uring_prep_write write;
> 				write.fd = iopf_fd;
> 				write.buf = malloc(IOPF_RESPONSE_SIZE);
> 				write.len = IOPF_RESPONSE_SIZE;
> 				write.flags = 0;
> 
> 				io_uring_prep_write(&ring, &write);
>             			io_uring_submit(&ring);
> 			}
> 
> 			// Reap the cqe
> 			io_uring_cqe_free(&ring, cqe);
> 		}
> 	}
> 
> Did I understand you correctly?

Yes, basically this is the right idea. There are more complex ways to
use the iouring that would be faster still.

And the kernel side can have support to speed it up as well.

I'm wondering if we should be pushing invalidations on io_uring as
well?

Jason
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [RFC PATCHES 00/17] IOMMUFD: Deliver IO page faults to user space

[Jean-Philippe Brucker]

Hi Baolu,

On Tue, May 30, 2023 at 01:37:07PM +0800, Lu Baolu wrote:
> - The timeout value for the pending page fault messages. Ideally we
>   should determine the timeout value from the device configuration, but
>   I failed to find any statement in the PCI specification (version 6.x).
>   A default 100 milliseconds is selected in the implementation, but it
>   leave the room for grow the code for per-device setting.

If it helps we had some discussions about this timeout [1]. It's useful to
print out a warning for debugging, but I don't think completing the fault
on timeout is correct, we should leave the fault pending. Given that the
PCI spec does not indicate a timeout, the guest can wait as long as it
wants to complete the fault (and 100ms may even be reasonable on an
emulator, who knows how many layers and context switches the fault has to
go through).


Another outstanding issue was what to do for PASID stop. When the guest
device driver stops using a PASID it issues a PASID stop request to the
device (a device-specific mechanism). If the device is not using PRI stop
markers it waits for pending PRs to complete and we're fine. Otherwise it
sends a stop marker which is flushed to the PRI queue, but does not wait
for pending PRs.

Handling stop markers is annoying. If the device issues one, then the PRI
queue contains stale faults, a stop marker, followed by valid faults for
the next address space bound to this PASID. The next address space will
get all the spurious faults because the fault handler doesn't know that
there is a stop marker coming. Linux is probably alright with spurious
faults, though maybe not in all cases, and other guests may not support
them at all.

We might need to revisit supporting stop markers: request that each device
driver declares whether their device uses stop markers on unbind() ("This
mechanism must indicate that a Stop Marker Message will be generated."
says the spec, but doesn't say if the function always uses one or the
other mechanism so it's per-unbind). Then we still have to synchronize
unbind() with the fault handler to deal with the pending stop marker,
which might have already gone through or be generated later.

Currently we ignore all that and just flush the PRI queue, followed by the
IOPF queue, to get rid of any stale fault before reassigning the PASID. A
guest however would also need to first flush the HW PRI queue, but doesn't
have a direct way to do that. If we want to support guests that don't deal
with stop markers, the host needs to flush the PRI queue when a PASID is
detached. I guess on Intel detaching the PASID goes through the host which
can flush the host queue. On Arm we'll probably need to flush the queue
when receiving a PASID cache invalidation, which the guest issues after
clearing a PASID table entry.

Thanks,
Jean

[1] https://lore.kernel.org/linux-iommu/20180423153622.GC38106@ostrya.localdomain/
    Also unregistration, not sure if relevant here
    https://lore.kernel.org/linux-iommu/20190605154553.0d00ad8d@jacob-builder/
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [RFC PATCHES 00/17] IOMMUFD: Deliver IO page faults to user space

[Jason Gunthorpe]

On Fri, Jun 16, 2023 at 12:32:32PM +0100, Jean-Philippe Brucker wrote:

> We might need to revisit supporting stop markers: request that each device
> driver declares whether their device uses stop markers on unbind() ("This
> mechanism must indicate that a Stop Marker Message will be generated."
> says the spec, but doesn't say if the function always uses one or the
> other mechanism so it's per-unbind). Then we still have to synchronize
> unbind() with the fault handler to deal with the pending stop marker,
> which might have already gone through or be generated later.

An explicit API to wait for the stop marker makes sense, with the
expectation that well behaved devices will generate it and well
behaved drivers will wait for it.

Things like VFIO should have a way to barrier/drain the PRI queue
after issuing FLR. ie the VMM processing FLR should also barrier the
real HW queues and flush them to VM visibility.

> with stop markers, the host needs to flush the PRI queue when a PASID is
> detached. I guess on Intel detaching the PASID goes through the host which
> can flush the host queue. On Arm we'll probably need to flush the queue
> when receiving a PASID cache invalidation, which the guest issues after
> clearing a PASID table entry.

We are trying to get ARM to a point where invalidations don't need to
be trapped. It would be good to not rely on that anyplace.

Jason
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [RFC PATCHES 00/17] IOMMUFD: Deliver IO page faults to user space

[Jean-Philippe Brucker]

On Mon, Jun 19, 2023 at 11:35:50AM +0800, Baolu Lu wrote:
> > Another outstanding issue was what to do for PASID stop. When the guest
> > device driver stops using a PASID it issues a PASID stop request to the
> > device (a device-specific mechanism). If the device is not using PRI stop
> > markers it waits for pending PRs to complete and we're fine. Otherwise it
> > sends a stop marker which is flushed to the PRI queue, but does not wait
> > for pending PRs.
> > 
> > Handling stop markers is annoying. If the device issues one, then the PRI
> > queue contains stale faults, a stop marker, followed by valid faults for
> > the next address space bound to this PASID. The next address space will
> > get all the spurious faults because the fault handler doesn't know that
> > there is a stop marker coming. Linux is probably alright with spurious
> > faults, though maybe not in all cases, and other guests may not support
> > them at all.
> > 
> > We might need to revisit supporting stop markers: request that each device
> > driver declares whether their device uses stop markers on unbind() ("This
> > mechanism must indicate that a Stop Marker Message will be generated."
> > says the spec, but doesn't say if the function always uses one or the
> > other mechanism so it's per-unbind). Then we still have to synchronize
> > unbind() with the fault handler to deal with the pending stop marker,
> > which might have already gone through or be generated later.
> 
> I don't quite follow here. Once a PASID is unbound from the device, the
> device driver should be free to release the PASID. The PASID could then
> be used for any other purpose. The device driver has no idea when the
> pending page requests are flushed after unbind(), so it cannot decide
> how long should the PASID be delayed for reuse. Therefore, I understand
> that a successful return from the unbind() function denotes that all
> pending page requests have been flushed and the PASID is viable for
> other use.

Yes that's the contract for unbind() at the moment

> 
> > 
> > Currently we ignore all that and just flush the PRI queue, followed by the
> > IOPF queue, to get rid of any stale fault before reassigning the PASID. A
> > guest however would also need to first flush the HW PRI queue, but doesn't
> > have a direct way to do that. If we want to support guests that don't deal
> > with stop markers, the host needs to flush the PRI queue when a PASID is
> > detached. I guess on Intel detaching the PASID goes through the host which
> > can flush the host queue. On Arm we'll probably need to flush the queue
> > when receiving a PASID cache invalidation, which the guest issues after
> > clearing a PASID table entry.
> 
> The Intel VT-d driver follows below steps to drain pending page requests
> when a PASID is unbound from a device.
> 
> - Tear down the device's pasid table entry for the stopped pasid.
>   This ensures that ATS/PRI will stop putting more page requests for the
>   pasid in VT-d PRQ.

Oh that's interesting, I didn't know about the implicit TLB invalidations
on page requests for VT-d.

For Arm SMMU, clearing the PASID table entry does cause ATS Translation
Requests to return with Completer Abort, but does not affect PRI. The SMMU
pushes page requests directly into the PRI queue without reading any table
(unless the queue overflows).

We're counting on the device driver to perform the PASID stop request
before calling unbind(), described in PCIe 6.20.1 (Managing PASID Usage)
and 10.4.1.2 (Managing PASID Usage on PRG Requests). This ensures that
when unbind() is called, no more page request for the PASID is pushed into
the PRI queue. But some may still be in the queue if the device uses stop
markers.

> - Sync with the PRQ handling thread until all related page requests in
>   PRQ have been delivered.

This is what I'm concerned about. For VT-d this happens in the host which
is in charge of modifying the PASID table. For SMMU, the guest writes the
PASID table. It flushes its virtual PRI queue, but not the physical queue
that is managed by the host.

One synchronization point where the host could flush the physical PRI
queue is the PASID config invalidation (CMD_CFGI_CD). As Jason pointed out
the host may not be able to observe those if a command queue is assigned
directly to the guest (a theoretical SMMU extension), though in that case
the guest may also have direct access to a PRI queue (like the AMD vIOMMU
extension) and be able to flush the queue directly.

But we can just wait for PRI implementations and see what the drivers
need. Maybe no device will implement stop markers.

Thanks,
Jean

> - Flush the iopf queue with iopf_queue_flush_dev().
> - Follow the steps defined in VT-d spec section 7.10 to drain all page
>   requests and responses between VT-d and the endpoint device.
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [RFC PATCHES 00/17] IOMMUFD: Deliver IO page faults to user space

[Jason Gunthorpe]

On Sun, Jun 25, 2023 at 02:30:46PM +0800, Baolu Lu wrote:

> Agreed. We should avoid workqueue in sva iopf framework. Perhaps we
> could go ahead with below code? It will be registered to device with
> iommu_register_device_fault_handler() in IOMMU_DEV_FEAT_IOPF enabling
> path. Un-registering in the disable path of cause.

This maze needs to be undone as well.

It makes no sense that all the drivers are calling 

 iommu_register_device_fault_handler(dev, iommu_queue_iopf, dev);

The driver should RX a PRI fault and deliver it to some core code
function, this looks like a good start:

> static int io_pgfault_handler(struct iommu_fault *fault, void *cookie)
> {
>         ioasid_t pasid = fault->prm.pasid;
>         struct device *dev = cookie;
>         struct iommu_domain *domain;
> 
>         if (fault->type != IOMMU_FAULT_PAGE_REQ)
>                 return -EOPNOTSUPP;
> 
>         if (fault->prm.flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID)
>                 domain = iommu_get_domain_for_dev_pasid(dev, pasid, 0);
>         else
>                 domain = iommu_get_domain_for_dev(dev);
> 
>         if (!domain || !domain->iopf_handler)
>                 return -ENODEV;
> 
>         if (domain->type == IOMMU_DOMAIN_SVA)
>                 return iommu_queue_iopf(fault, cookie);
> 
>         return domain->iopf_handler(fault, dev, domain->fault_data);

Then we find the domain that owns the translation and invoke its
domain->ops->iopf_handler()

If the driver created a SVA domain then the op should point to some
generic 'handle sva fault' function. There shouldn't be weird SVA
stuff in the core code.

The weird SVA stuff is really just a generic per-device workqueue
dispatcher, so if we think that is valuable then it should be
integrated into the iommu_domain (domain->ops->use_iopf_workqueue =
true for instance). Then it could route the fault through the
workqueue and still invoke domain->ops->iopf_handler.

The word "SVA" should not appear in any of this.

Not sure what iommu_register_device_fault_handler() has to do with all
of this.. Setting up the dev_iommu stuff to allow for the workqueue
should happen dynamically during domain attach, ideally in the core
code before calling to the driver.

Also, I can understand there is a need to turn on PRI support really
early, and it can make sense to have some IOMMU_DEV_FEAT_IOPF/SVA to
ask to turn it on.. But that should really only be needed if the HW
cannot turn it on dynamically during domain attach of a PRI enabled
domain.

It needs cleaning up..

Jason
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [RFC PATCHES 00/17] IOMMUFD: Deliver IO page faults to user space

[Jason Gunthorpe]

On Wed, Jun 28, 2023 at 10:00:56AM +0800, Baolu Lu wrote:
> > If the driver created a SVA domain then the op should point to some
> > generic 'handle sva fault' function. There shouldn't be weird SVA
> > stuff in the core code.
> > 
> > The weird SVA stuff is really just a generic per-device workqueue
> > dispatcher, so if we think that is valuable then it should be
> > integrated into the iommu_domain (domain->ops->use_iopf_workqueue =
> > true for instance). Then it could route the fault through the
> > workqueue and still invoke domain->ops->iopf_handler.
> > 
> > The word "SVA" should not appear in any of this.
> 
> Yes. We should make it generic. The domain->use_iopf_workqueue flag
> denotes that the page faults of a fault group should be put together and
> then be handled and responded in a workqueue. Otherwise, the page fault
> is dispatched to domain->iopf_handler directly.

It might be better to have iopf_handler and
iopf_handler_work function pointers to distinguish to two cases.

> > Not sure what iommu_register_device_fault_handler() has to do with all
> > of this.. Setting up the dev_iommu stuff to allow for the workqueue
> > should happen dynamically during domain attach, ideally in the core
> > code before calling to the driver.
> 
> There are two pointers under struct dev_iommu for fault handling.
> 
> /**
>  * struct dev_iommu - Collection of per-device IOMMU data
>  *
>  * @fault_param: IOMMU detected device fault reporting data
>  * @iopf_param:  I/O Page Fault queue and data
> 
> [...]
> 
> struct dev_iommu {
>         struct mutex lock;
>         struct iommu_fault_param        *fault_param;
>         struct iopf_device_param        *iopf_param;
> 
> My understanding is that @fault_param is a place holder for generic
> things, while @iopf_param is workqueue specific.

Well, lets look

struct iommu_fault_param {
	iommu_dev_fault_handler_t handler;
	void *data;

These two make no sense now. handler is always iommu_queue_iopf. Given
our domain centric design we want the function pointer in the domain,
not in the device. So delete it.

	struct list_head faults;
	struct mutex lock;

Queue of unhandled/unacked faults? Seems sort of reasonable

> @iopf_param could be allocated on demand. (perhaps renaming it to a more
> meaningful one?) It happens before a domain with use_iopf_workqueue flag
> set attaches to a device. iopf_param keeps alive until device_release.

Yes

Do this for the iommu_fault_param as well, in fact, probably just put
the two things together in one allocation and allocate if we attach a
PRI using domain. I don't think we need to micro optimze further..
 
Jason
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization