On Tue, Apr 26, 2022 at 04:31:57PM -0700, Dave Hansen wrote:
On 4/26/22 09:48, Jean-Philippe Brucker wrote:
On Tue, Apr 26, 2022 at 08:27:00AM -0700, Dave Hansen wrote:
On 4/25/22 09:40, Jean-Philippe Brucker wrote:The latter I think, this one was about releasing pages as fast as possible
The problem is that we'd have to request the device driver to stop DMAIs the concern truly about blocking mmput() itself? Or, is it about
before we can destroy the context and free the PASID. We did consider
doing this in the release() MMU notifier, but there were concerns about
blocking mmput() for too long (for example
though I think there was a more recent discussion). We also need to drain
the PRI and fault queues to get rid of all references to that PASID.
releasing the resources associated with the mm?
if the process is picked by the OOM killer.
We're tying the PASID to the life of the mm itself, not the mm's address
space. That means the PASID should be tied to
The address space is what the OOM killer is after. That gets refcounted
with mmget()/mmput()/mm->mm_users. The OOM killer is satiated by the
page freeing done in __mmput()->exit_mmap().
Also, all the VMAs should be gone after exit_mmap(). So, even if
vma->vm_file was holding a reference to a device driver, that reference
should be gone by the time __mmdrop() is actually freeing the PASID.
I agree with all that. The concern was about tearing down the PASID in the
IOMMU and device from the release() MMU notifier, which would happen in
exit_mmap(). But doing the teardown at or before __mmdrop() is fine. And
since the IOMMU drivers need to hold mm->mm_count anyway between bind()
and unbind(), I think Fenghua's fix works.