[RFCv2 PATCH 0/7] A General Accelerator Framework, WarpDrive

Jerome Glisse jglisse at redhat.com
Mon Sep 17 11:42:44 AEST 2018

So i want to summarize issues i have as this threads have dig deep into
details. For this i would like to differentiate two cases first the easy
one when relying on SVA/SVM. Then the second one when there is no SVA/SVM.
In both cases your objectives as i understand them:

[R1]- expose a common user space API that make it easy to share boiler
      plate code accross many devices (discovering devices, opening
      device, creating context, creating command queue ...).
[R2]- try to share the device as much as possible up to device limits
      (number of independant queues the device has)
[R3]- minimize syscall by allowing user space to directly schedule on the
      device queue without a round trip to the kernel

I don't think i missed any.

(1) Device with SVA/SVM

For that case it is easy, you do not need to be in VFIO or part of any
thing specific in the kernel. There is no security risk (modulo bug in
the SVA/SVM silicon). Fork/exec is properly handle and binding a process
to a device is just couple dozen lines of code.

(2) Device does not have SVA/SVM (or it is disabled)

You want to still allow device to be part of your framework. However
here i see fundamentals securities issues and you move the burden of
being careful to user space which i think is a bad idea. We should
never trus the userspace from kernel space.

To keep the same API for the user space code you want a 1:1 mapping
between device physical address and process virtual address (ie if
device access device physical address A it is accessing the same
memory as what is backing the virtual address A in the process.

Security issues are on two things:
[I1]- fork/exec, a process who opened any such device and created an
      active queue can transfer without its knowledge control of its
      commands queue through COW. The parent map some anonymous region
      to the device as a command queue buffer but because of COW the
      parent can be the first to copy on write and thus the child can
      inherit the original pages that are mapped to the hardware.
      Here parent lose control and child gain it.

[I2]- Because of [R3] you want to allow userspace to schedule commands
      on the device without doing an ioctl and thus here user space
      can schedule any commands to the device with any address. What
      happens if that address have not been mapped by the user space
      is undefined and in fact can not be defined as what each IOMMU
      does on invalid address access is different from IOMMU to IOMMU.

      In case of a bad IOMMU, or simply an IOMMU improperly setup by
      the kernel, this can potentialy allow user space to DMA anywhere.

[I3]- By relying on GUP in VFIO you are not abiding by the implicit
      contract (at least i hope it is implicit) that you should not
      try to map to the device any file backed vma (private or share).

      The VFIO code never check the vma controlling the addresses that
      are provided to VFIO_IOMMU_MAP_DMA ioctl. Which means that the
      user space can provide file backed range.

      I am guessing that the VFIO code never had any issues because its
      number one user is QEMU and QEMU never does that (and that's good
      as no one should ever do that).

      So if process does that you are opening your self to serious file
      system corruption (depending on file system this can lead to total
      data loss for the filesystem).

      Issue is that once you GUP you never abide to file system flushing
      which write protect the page before writing to the disk. So
      because the page is still map with write permission to the device
      (assuming VFIO_IOMMU_MAP_DMA was a write map) then the device can
      write to the page while it is in the middle of being written back
      to disk. Consult your nearest file system specialist to ask him
      how bad that can be.

[I4]- Design issue, mdev design As Far As I Understand It is about
      sharing a single device to multiple clients (most obvious case
      here is again QEMU guest). But you are going against that model,
      in fact AFAIUI you are doing the exect opposite. When there is
      no SVA/SVM you want only one mdev device that can not be share.

      So this is counter intuitive to the mdev existing design. It is
      not about sharing device among multiple users but about giving
      exclusive access to the device to one user.

All the reasons above is why i believe a different model would serve
you and your user better. Below is a design that avoids all of the
above issues and still delivers all of your objectives with the
exceptions of the third one [R3] when there is no SVA/SVM.

Create a subsystem (very much boiler plate code) which allow device to
register themself against (very much like what you do in your current
patchset but outside of VFIO).

That subsystem will create a device file for each registered system and
expose a common API (ie set of ioctl) for each of those device files.

When user space create a queue (through an ioctl after opening the device
file) the kernel can return -EBUSY if all the device queue are in use,
or create a device queue and return a flag like SYNC_ONLY for device that
do not have SVA/SVM.

For device with SVA/SVM at the time the process create a queue you bind
the process PASID to the device queue. From there on the userspace can
schedule commands and use the device without going to kernel space.

For device without SVA/SVM you create a fake queue that is just pure
memory is not related to the device. From there on the userspace must
call an ioctl every time it wants the device to consume its queue
(hence why the SYNC_ONLY flag for synchronous operation only). The
kernel portion read the fake queue expose to user space and copy
commands into the real hardware queue but first it properly map any
of the process memory needed for those commands to the device and
adjust the device physical address with the one it gets from dma_map

With that model it is "easy" to listen to mmu_notifier and to abide by
them to avoid issues [I1], [I3] and [I4]. You obviously avoid the [I2]
issue by only mapping a fake device queue to userspace.

So yes with that models it means that every device that wish to support
the non SVA/SVM case will have to do extra work (ie emulate its command
queue in software in the kernel). But by doing so, you support an
unlimited number of process on your device (ie all the process can share
one single hardware command queues or multiple hardware queues).

The big advantages i see here is that the process do not have to worry
about doing something wrong. You are protecting yourself and your user
from stupid mistakes.

I hope this is useful to you.


More information about the Linux-accelerators mailing list