Page MenuHomeFreeBSD

pci: Add ioctls to perform I/O to I/O port BARs
ClosedPublic

Authored by markj on Jul 26 2021, 4:45 PM.
Tags
None
Referenced Files
Unknown Object (File)
Thu, Oct 31, 8:34 PM
Unknown Object (File)
Wed, Oct 23, 2:51 AM
Unknown Object (File)
Sun, Oct 20, 6:24 AM
Unknown Object (File)
Sun, Oct 20, 6:23 AM
Unknown Object (File)
Sun, Oct 20, 6:23 AM
Unknown Object (File)
Sun, Oct 20, 6:23 AM
Unknown Object (File)
Sun, Oct 20, 6:23 AM
Unknown Object (File)
Sun, Oct 20, 6:23 AM
Subscribers

Details

Summary

This is useful for bhyve, which otherwise has to use /dev/io to handle
accesses to I/O port BARs when PCI passthrough is in use. In
particular, /dev/io cannot be used in jails even if devfs access is
granted.

Test Plan

See D31308.

Diff Detail

Repository
rG FreeBSD src repository
Lint
Lint Not Applicable
Unit
Tests Not Applicable

Event Timeline

markj requested review of this revision.Jul 26 2021, 4:45 PM
markj added reviewers: grehan, jhb, kib.

My feel is that it would be more useful to have single ioctl, but encode read/write operation in struct pci_ioport_io. My reasoning is that just reads and writes are not enough, eventually for this API to be more useful, there will be requests for operations like and/or/nor. I do not propose to implement them outright, but it is better to leave the flexibility to introduce them when needed.

sys/dev/pci/pci_user.c
943

I believe that this feature should not be limited to the io bars. Or rather, I see no reason why it is useful to limit it.

956

This is an overflow check, am I right? Isn't it better to check that offset+width does not overflow, and then leave only second test for rman_get_size?

1056

I believe ports (or MMIO) reads should be treated same as writes from the security PoV. It is very common for any access to some registers to have side effect.

markj marked 3 inline comments as done.

Address feedback:

  • Add a single ioctl and encode the requested op in the parameters
  • Fix FWRITE check
  • Permit I/O to any BAR
  • Correct the overflow check
sys/dev/pci/pci_user.c
1056

Ugh, thanks, I misread the check and inverted it in my head.

sys/dev/pci/pci_user.c
946

So you still require port io space? The comment for diff stated that you allowed any bar.

949

I am completely not sure, but wouldn't RF_SHAREABLE useful (or even required) there, if we already have a driver attached?

markj marked an inline comment as done.

Actually remove the BAR type check.

sys/dev/pci/pci_user.c
949

It's not required at least for PCI passthrough with ppt. Is there anything preventing a mmap of a memory-mapped BAR if a driver has attached and mapped it already? If not then it probably makes sense to add RF_SHAREABLE.

sys/dev/pci/pci_user.c
949

As I said, I am not sure. My belief is that it is needed on both sides to get the resource shared between two activations, but I do not know for sure.

This revision is now accepted and ready to land.Jul 26 2021, 7:40 PM
sys/dev/pci/pci_user.c
947

You need to pick between SYS_RES_IOPORT and SYS_RES_MEMORY here to support memory BARs. I don't think we have a helper for that yet, but you can use this:

if (PCI_BAR_MEM(pm->pm_value))
    type = SYS_RES_MEMORY;
else
    type = SYS_RES_IOPORT;
949

As I said, I am not sure. My belief is that it is needed on both sides to get the resource shared between two activations, but I do not know for sure.

So it is true that all parties have to use RF_SHAREABLE. It may also be a feature that you can't do I/O to BARs mapped by drivers (I think ppt doesn't alloc resources for its BARs which is why Mark's use case works). I'm not sure how I feel about what we should do here. Not passing the flag is the most conservative approach, and I think most drivers don't allocate their BARs shareable anyway. Actually, the PCI bus only allows for one allocation of a BAR at a time, though a few places like vgapci allow multiple allocs via a ref count.

950–951

Ugh, new code shouldn't use this.

956

You could move these checks earlier and use pm->pm_size as the BAR size and validate the offset and width before the call to bus_alloc_resource_any.

967

Here and the other places below.

markj marked 4 inline comments as done.
  • Address feedback.
  • Rename the ioctl and request structure to indicate that they can be used for both I/O port and memory-mapped BARs.
This revision now requires review to proceed.Jul 30 2021, 3:22 PM
This revision is now accepted and ready to land.Jul 30 2021, 3:45 PM

Generally this is looking good. I agree with the RF_SHAREABLE bit. But to support that protocol correctly, you need to activate/deactivate while holding the allocation, which suggests a more elaborate protocol would be needed than the stateless one we have here. So it's best to omit it entirely.

share/man/man4/pci.4
454

I'd be tempted to add 'in bytes' here to make it clear the units. And why aren't 8-byte I/Os supported?

sys/dev/pci/pci_user.c
937

How do we know it's a type 0 device here? Do you allow these operations on PCI bridges? If not, perhaps you should filter them out. If so, you need to test for TYPE 1 here. Type 2 is CardBus bridges and likely can be ignored, or at least filtered, given their rarity these days.

946

I think you want an invariant check to make sure this is a valid width.
Or at least have an invariants message in the case below.

970

Why not case 8: here?

markj marked 3 inline comments as done.

Address feedback from Warner:

  • man page clarification
  • permit 8-byte accesses
  • remove the BAR index check and rely on pci_find_bar() instead

Also add a comment noting that the ioctl could be modified to work if a driver
has already allocated the resource.

This revision now requires review to proceed.Aug 13 2021, 8:53 PM

I'm not seeing anything new, and you addressed all my prior feedback (though the large number of comments in spots makes it a little hard to read)

This revision is now accepted and ready to land.Aug 13 2021, 10:34 PM
This revision was automatically updated to reflect the committed changes.