PCI_P2PDMA_MAP_BUS_ADDR and PCIe ATS?
Hi!
I was looking at the way pci_p2pdma_distance() returns PCI_P2PDMA_MAP_BUS_ADDR if devices sit beneath a local switch with ACS disabled or not present.
The client then programs the bus address directly to perform p2p dma, as handed out by the dma layer.
However if the device (p2pdma client) has IOMMU turned on and ATS enabled, the device may or may not send a translation request to the IOMMU. Wouldn't that fail returnint an IOMMU fault if the device is programmed with the bus address rather than the IOVA?
So if ATS is enabled, shouldn't the device always be handed the IOVA to avoid this? Granted, untranslated transactions will then take the round-trip to the host bridge, but pre-translated transactions will still benefit from the direct routing through the switch, although that is not directly visible to the kernel p2pdma layer?
Does pci_p2pdma_distance() need to be client ATS-enabled aware?
Any input appreciated,
Thanks,
Thomas
Quoted 6 lines · Thomas Hellström
Hi!
I was looking at the way
pci_p2pdma_distance()returnsPCI_P2PDMA_MAP_BUS_ADDRif devices sit beneath a local switch with ACS disabled or not present.
Quoted 7 lines
The client then programs the bus address directly to perform p2p dma, as handed out by the dma layer.
However if the device (p2pdma client) has IOMMU turned on and ATS enabled, the device may or may not send a translation request to the IOMMU. Wouldn't that fail returnint an IOMMU fault if the device is programmed with the bus address rather than the IOVA?
There are many goofy configurations you can setup on PCIe, something like this would be one of them :)
Linux only (barely) works with a few mainstream ones.
Quoted 2 lines
So if ATS is enabled, shouldn't the device always be handed the IOVA to avoid this?
No, Linux has no idea if the device will use ATS or not for a certain request. ATS is not an all or nothing feature in PCI.
If you want this kind of configuration then set the ACS flags to achieve it, or disable ATS in the device.
IMHO such a PCIe fabric is misconfigured. I can't see a reason why anyone would want to have ACS flags allowing fast path and then Linux setup to never supply an IOVA that can use it.
Jason
Quoted 30 lines · Jason Gunthorpe
Quoted 11 lines · Thomas Hellström
Hi!
I was looking at the way
pci_p2pdma_distance()returnsPCI_P2PDMA_MAP_BUS_ADDRif devices sit beneath a local switch with ACS
disabled or not present.The client then programs the bus address directly to perform p2p dma,
as handed out by the dma layer.Quoted 5 lines
However if the device (p2pdma client) has IOMMU turned on and ATS enabled, the device may or may not send a translation request to the
IOMMU. Wouldn't that fail returnint an IOMMU fault if the device is programmed with the bus address rather than the IOVA?There are many goofy configurations you can setup on PCIe, something like this would be one of them :)
Linux only (barely) works with a few mainstream ones.
Quoted 3 lines
So if ATS is enabled, shouldn't the device always be handed the IOVA to
avoid this?No, Linux has no idea if the device will use ATS or not for a certain request. ATS is not an all or nothing feature in PCI.
Quoted 6 lines
If you want this kind of configuration then set the ACS flags to achieve it, or disable ATS in the device.
IMHO such a PCIe fabric is misconfigured. I can't see a reason why anyone would want to have ACS flags allowing fast path and then Linux setup to never supply an IOVA that can use it.
From my understanding, ATS translated requests would use the fast-path as soon as a translation is established in the device, without the BUS address being explicitly programmed. ACS redirect would block that.
Admittedly untranslated requests would take the detour around the host bridge, but a GPU device that would otherwise enable ATS by default to offload the IOMMU TLB would outright fail with the current code with little clue what was going on.
So wouldn't it be better to be safe than sorry and return a host bridge map and make something usable and potentially fast out of something currently unusable?
The only way I can see one could support the fast-path for both untranslated and translated request would be to set up a 1:1 IOVA for the particular mapping.
/Thomas
Quoted 2 lines
Jason
Quoted 3 lines · Thomas Hellström
From my understanding, ATS translated requests would use the fast-path as soon as a translation is established in the device, without the BUS address being explicitly programmed. ACS redirect would block that.
Oh it's super complex. There are seperate ACS bits for translated and untranslated requests. You need a PCI expert to design your fabric and devices to get a working system unfortunately.
If you use ATS and don't set the ACS to route all translations to host then your IOVA space is no longer uniform, you have different results depending on which fabric route you take. That's a pretty unreasonable fabric design, and can't work securely with virtualization.
Quoted 3 lines
So wouldn't it be better to be safe than sorry and return a host bridge map and make something usable and potentially fast out of something currently unusable?
What we have today works for lots of devices and if you change it like this then you'll break existing users.
Probably what you want to do is add a flag that the device is going to use ATS for the P2P DMA. Then we can force setup an IOVA and check the fabric path for translated request routing to make sure it will work.
And yes, there is an existing issue with DMABUF not being able to negotiate these little pathing details with the importer when the exporter forms the scatterlist..
Jason
Quoted 20 lines · Jason Gunthorpe
Quoted 5 lines · Thomas Hellström
From my understanding, ATS translated requests would use the fast- path
as soon as a translation is established in the device, without the BUS
address being explicitly programmed. ACS redirect would block that.Oh it's super complex. There are seperate ACS bits for translated and untranslated requests. You need a PCI expert to design your fabric and
devices to get a working system unfortunately.If you use ATS and don't set the ACS to route all translations to host
then your IOVA space is no longer uniform, you have different results depending on which fabric route you take. That's a pretty unreasonable
fabric design, and can't work securely with virtualization.
Quoted 8 lines
Quoted 4 lines
So wouldn't it be better to be safe than sorry and return a host bridge
map and make something usable and potentially fast out of something currently unusable?What we have today works for lots of devices and if you change it like
this then you'll break existing users.
I guess if there are setups with devices with PCIe ATS enabled but only ever sending untranslated requests, indeed they would see a performance regression with this approach.
Quoted 6 lines
Probably what you want to do is add a flag that the device is going to
use ATS for the P2P DMA. Then we can force setup an IOVA and check the
fabric path for translated request routing to make sure it will work.
Yes, sounds like a way forward. To be honest, we don't have an existing use-case yet, but taking the union of how we want to configure our GPUs moving forward and the support requests we see from p2pdma users, this is probably something we are going to hit sooner than later.
Quoted 4 lines
And yes, there is an existing issue with DMABUF not being able to negotiate these little pathing details with the importer when the exporter forms the scatterlist..
Yes, it was when looking into that I actually stumbled upon the ATS issue as well. I recall you and Christian have touched upon this in the past. What is an acceptable way forward here? A dma_map_resource_p2p() interface that takes into account the mapping type?
Thanks,
Thomas
Quoted 2 lines
Jason
Quoted 10 lines · Thomas Hellström
Quoted 5 lines
Probably what you want to do is add a flag that the device is going to
use ATS for the P2P DMA. Then we can force setup an IOVA and check the
fabric path for translated request routing to make sure it will work.Yes, sounds like a way forward. To be honest, we don't have an existing use-case yet, but taking the union of how we want to configure our GPUs moving forward and the support requests we see from p2pdma users, this is probably something we are going to hit sooner than later.
I will say we've had to deploy alot of experts to design our systems and it is really hard to get something that actually works :|
Quoted 8 lines
Quoted 3 lines
And yes, there is an existing issue with DMABUF not being able to negotiate these little pathing details with the importer when the exporter forms the scatterlist..
Yes, it was when looking into that I actually stumbled upon the ATS issue as well. I recall you and Christian have touched upon this in the past. What is an acceptable way forward here? A
dma_map_resource_p2p()interface that takes into account the mapping type?
IDK, it is painful. There is alot of stuff coming that the importer will need to customize when performing the mapping. Locking the DMA mapping process in the exporter is not helpful. For example your device might need to say 'I will always use ATS', while mlx5 would prefer 'Should I disable ATS for this mapping'.
Christian seems unrelenting on this point so I don't really have a good solution beyond some endless list of single user flags.
I wanted to rebase and revise the mapping type series but I've been just buried in security things, patch reviews and travel lately. Maybe I will get there after this merge window
Jason