Hi, Jean and Baolu,
On Fri, Apr 29, 2022 at 03:34:36PM +0100, Jean-Philippe Brucker wrote:
On Fri, Apr 29, 2022 at 06:51:17AM -0700, Fenghua Yu wrote:
On Fri, Apr 29, 2022 at 03:53:57PM +0800, Baolu Lu wrote:
On 2022/4/28 16:39, Jean-Philippe Brucker wrote:
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.
But I didn't find mmgrab()/mmdrop() get called in both arm and intel
$ git grep mmgrab drivers/iommu/
Do we need to add these in a separated fix patch, or I missed anything
On both ARM and X86, sva_bind() calls mmu_notifier_register()->mmgrab() and
sva_unbind() calls mmu_notifier_unregister()/mmu_notifier_put()->mmdrop().
Yes, although for Arm I realized the mmu_notifier grab wasn't sufficient
so I sent a separate fix that should go in 5.18 as well
The Arm driver still touches the arch mm context after mmu_notifier_put().
I don't think X86 has that problem.
I think so too. On X86, the mm is not used after mmu_notifier_unregister().
In case of supervisor mode SVM (i.e. svm is bound to init_mm), the code
is right too because init_mm and its PASID cannot be dropped and
mmu_notifier_register()/mmu_notifier_unregister() are not called.
So I think no extra fix patch is needed on X86.