We discussed if that is really the right approach for quite a while now, but digging deeper into a bug report on arm turned out that this is actually horrible broken right now.
The reason for this is that vmf_insert_mixed_prot() always tries to grab a reference to the underlaying page on architectures without ARCH_HAS_PTE_SPECIAL and as far as I can see also enabled GUP.
So nuke using VM_MIXEDMAP here and use VM_PFNMAP instead.
Also set VM_SHARED, not 100% sure if that is needed with VM_PFNMAP, but better save than sorry.
Signed-off-by: Christian König christian.koenig@amd.com Bugs: https://gitlab.freedesktop.org/drm/amd/-/issues/1606#note_936174 --- drivers/gpu/drm/ttm/ttm_bo_vm.c | 29 +++++++---------------------- 1 file changed, 7 insertions(+), 22 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c index 9bd15cb39145..bf86ae849340 100644 --- a/drivers/gpu/drm/ttm/ttm_bo_vm.c +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c @@ -359,12 +359,7 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, * at arbitrary times while the data is mmap'ed. * See vmf_insert_mixed_prot() for a discussion. */ - if (vma->vm_flags & VM_MIXEDMAP) - ret = vmf_insert_mixed_prot(vma, address, - __pfn_to_pfn_t(pfn, PFN_DEV), - prot); - else - ret = vmf_insert_pfn_prot(vma, address, pfn, prot); + ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
/* Never error on prefaulted PTEs */ if (unlikely((ret & VM_FAULT_ERROR))) { @@ -411,15 +406,9 @@ vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot) pfn = page_to_pfn(page);
/* Prefault the entire VMA range right away to avoid further faults */ - for (address = vma->vm_start; address < vma->vm_end; address += PAGE_SIZE) { - - if (vma->vm_flags & VM_MIXEDMAP) - ret = vmf_insert_mixed_prot(vma, address, - __pfn_to_pfn_t(pfn, PFN_DEV), - prot); - else - ret = vmf_insert_pfn_prot(vma, address, pfn, prot); - } + for (address = vma->vm_start; address < vma->vm_end; + address += PAGE_SIZE) + ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
return ret; } @@ -576,14 +565,10 @@ static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object *bo, struct vm_area_s
vma->vm_private_data = bo;
- /* - * We'd like to use VM_PFNMAP on shared mappings, where - * (vma->vm_flags & VM_SHARED) != 0, for performance reasons, - * but for some reason VM_PFNMAP + x86 PAT + write-combine is very - * bad for performance. Until that has been sorted out, use - * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719 + /* Enforce VM_SHARED here since no driver backend actually supports COW + * on TTM buffer object mappings. */ - vma->vm_flags |= VM_MIXEDMAP; + vma->vm_flags |= VM_PFNMAP | VM_SHARED; vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP; }
On 6/2/21 10:30 AM, Christian König wrote:
I think vmwgfx still uses MIXEDMAP. (Which is ofc same bug and should be changed).
I think by default all TTM drivers support COW mappings in the sense that written data never makes it to the bo but stays in anonymous pages, although I can't find a single usecase. So comment should be changed to state that they are useless for us and that we can't support COW mappings with VM_PFNMAP.
Hmm, shouldn't we refuse COW mappings instead, like my old patch on this subject did? In theory someone could be setting up what she thinks is a private mapping to a shared buffer object, and write sensitive data to it, which will immediately leak. It's a simple check, could open-code if necessary.
vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP; }
/Thomas
Am 02.06.21 um 11:07 schrieb Thomas Hellström (Intel):
Mhm, the only thing I could find is that it is clearing VM_MIXEDMAP and adding VM_PFNMAP instead.
But going to clean that up as well.
Well the problem I see with that is that it only works as long as the BO is in system memory. When it then suddenly migrates to VRAM everybody sees the same content again and the COW pages are dropped. That is really inconsistent and I can't see why we would want to do that.
Additionally to that when you allow COW mappings you need to make sure your COWed pages have the right caching attribute and that the reference count is initialized and taken into account properly. Not driver actually gets that right at the moment.
Yeah, though about that as well. Rejecting things would mean we potentially break userspace which just happened to work by coincident previously. Not totally evil, but not nice either.
How about we do a WARN_ON_ONCE(!(vma->vm_flags & VM_SHARED)); instead?
Thanks, Christian.
Hi,
On 6/2/21 12:03 PM, Christian König wrote:
Hmm, yes, that's actually a bug in drm_vma_manager().
I was under the impression that COW'ed pages were handled transparently by the vm, you'd always get cached properly refcounted COW'ed pages but anyway since we're going to ditch support for them, doesn't really matter.
Umm, yes but that wouldn't notify the user, and would be triggerable from user-space. But you can also set up legal non-COW mappings without the VM_SHARED flag, IIRC, see is_cow_mapping(). I think when this was up for discussion last time we arrived in a vma_is_cow_mapping() utility...
/Thomas
Am 02.06.21 um 13:24 schrieb Thomas Hellström (Intel):
Hui? How is that related to drm_vma_manager() ?
Yeah, but I would have expected that the new COWed page should have the same caching attributes as the old one and that is not really the case.
Well userspace could trigger that only once, so no spamming of the log can be expected here. And extra warnings in the logs are usually reported by people rather quickly.
Christian.
/Thomas
On Wed, Jun 02, 2021 at 02:21:17PM +0200, Thomas Hellström wrote:
VM_SHARED is already cleared in vma_set_page_prot() due to the VM_PFNMAP check in vma_wants_writenotify.
I'm honestly not sure whether userspace then even can notice this or anything, so might be worth a quick testcase.
Even if I'm wrong here we shouldn't allow cow mappings of gem_bo, that just seems too nasty with all the side-effects. -Daniel
On 6/2/21 8:36 PM, Daniel Vetter wrote:
Yes, but that's only on a local variable to get a write-protected prot. vmwgfx does the same for its dirty-tracking. Here we're debating setting VM_SHARED on a private mapping.
I'm honestly not sure whether userspace then even can notice this or anything, so might be worth a quick testcase.
The net result is that in the very unlikely case the user requested a private GPU mapping to write secret data into, That secret data is no longer secret. And, for example in the case of AMD's SEV encryption that data would have been encrypted in an anonymous page with COW mappings, but not so if we add VM_SHARED, then it will be unencrypted in in GPU pages. Then I think it's better to refuse COW mappings in mmap:
if (is_cow_mapping(vma->vm_flags)) return -EINVAL;
This will still allow private read-only mappings which is OK. And if someone was actually relying on private COW'd GPU mappings, we'd only break the code slightly more...
/Thomas
Even if I'm wrong here we shouldn't allow cow mappings of gem_bo, that just seems too nasty with all the side-effects.
Completely agree.
On Wed, Jun 02, 2021 at 10:30:13AM +0200, Christian König wrote:
I thought we still have the same issue open for ttm_bo_vm_insert_huge()? Or at least a potentially pretty big bug, because our current huge entries don't stop gup (because there's no pmd_mkspecial right now in the kernel).
So I think if you want to close this for good then we also need to (temporarily at least) disable the huge entry code? -Daniel
Am 02.06.21 um 20:38 schrieb Daniel Vetter:
That's already done (at least for ~vmwgfx) because we ran into problems we couldn't explain.
Going to add something which explicitly disable it with a comment.
What's the conclusion on VM_SHARED? Should I enforce this, warn about it or just ignore it because it doesn't matter for VM_PFNMAP?
Thanks, Christian.
dri-devel@lists.freedesktop.org
-
Christian König -
Daniel Vetter -
Thomas Hellström -
Thomas Hellström (Intel)