Am 23.11.22 um 13:53 schrieb Jason Gunthorpe:
On Wed, Nov 23, 2022 at 01:49:41PM +0100, Christian König wrote:
Am 23.11.22 um 13:46 schrieb Jason Gunthorpe:
On Wed, Nov 23, 2022 at 11:06:55AM +0100, Daniel Vetter wrote:
Maybe a GFP flag to set the page reference count to zero or something like this?
Hm yeah that might work. I'm not sure what it will all break though? And we'd need to make sure that underflowing the page refcount dies in a backtrace.
Mucking with the refcount like this to protect against crazy out of tree drives seems horrible..
Well not only out of tree drivers. The intree KVM got that horrible wrong as well, those where the latest guys complaining about it.
kvm was taking refs on special PTEs? That seems really unlikely?
Well then look at this code here:
commit add6a0cd1c5ba51b201e1361b05a5df817083618 Author: Paolo Bonzini pbonzini@redhat.com Date: Tue Jun 7 17:51:18 2016 +0200
KVM: MMU: try to fix up page faults before giving up
The vGPU folks would like to trap the first access to a BAR by setting vm_ops on the VMAs produced by mmap-ing a VFIO device. The fault handler then can use remap_pfn_range to place some non-reserved pages in the VMA.
This kind of VM_PFNMAP mapping is not handled by KVM, but follow_pfn and fixup_user_fault together help supporting it. The patch also supports VM_MIXEDMAP vmas where the pfns are not reserved and thus subject to reference counting.
Cc: Xiao Guangrong guangrong.xiao@linux.intel.com Cc: Andrea Arcangeli aarcange@redhat.com Cc: Radim Krčmář rkrcmar@redhat.com Tested-by: Neo Jia cjia@nvidia.com Reported-by: Kirti Wankhede kwankhede@nvidia.com Signed-off-by: Paolo Bonzini pbonzini@redhat.com
And see also the discussion here: https://patchwork.freedesktop.org/patch/414123/
as well as here: https://patchwork.freedesktop.org/patch/499190/
I can't count how often I have pointed out that this is absolutely illegal and KVM can't touch pages in VMAs with VM_PFNMAP.
The WARN_ON(pag_count(p) != 1) seems like a reasonable thing to do though, though you must combine this with the special PTE flag..
That's not sufficient. The pages are released much later than things actually go wrong. In most cases this WARN_ON here won't hit.
How so? As long as the page is mapped into the PTE there is no issue with corruption. If dmabuf checks the refcount after it does the unmap mapping range it should catch any bogus pin that might be confused about address coherency.
Yeah, that would work. The problem is this WARN_ON() comes much later.
The device drivers usually keep the page around for a while even after it is unmapped. IIRC the cleanup worker only runs every 10ms or so.
Christian.
Jason