Skip to contentSkip to conversation

drm/vma-manager: Don't unmap COW'd pages when zapping bo ptes

PATCH In review Thomas Hellstrom 3 participants 8 messages Nov 20, 2013 – Nov 21, 2013

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/vma-manager: Don't unmap COW'd pages when zapping bo ptes

Not sure if there are any user-space users of private bo mappings, but if there are, or will be, zapping the COW'd pages when, for example, moving a bo would confuse the user immensely since the net effect for the user would be that pages written to would lose their contents.

---
 include/drm/drm_vma_manager.h |    6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
include/drm/drm_vma_manager.h +3 -3
index c18a593..347077d 100644
@@ -231,9 +231,9 @@ static inline void drm_vma_node_unmap(struct drm_vma_offset_node *node,
 				      struct address_space *file_mapping)
 {
 	if (file_mapping && drm_vma_node_has_offset(node))
-		unmap_mapping_range(file_mapping,
-				    drm_vma_node_offset_addr(node),
-				    drm_vma_node_size(node) << PAGE_SHIFT, 1);
+		unmap_shared_mapping_range
+			(file_mapping, drm_vma_node_offset_addr(node),
+			 drm_vma_node_size(node) << PAGE_SHIFT);
 }
 
 /**
Signature · 1 line
1.7.10.4
#1 lore raw

Hi

Quoted 4 lines · Thomas Hellstrom

Not sure if there are any user-space users of private bo mappings, but if there are, or will be, zapping the COW'd pages when, for example, moving a bo would confuse the user immensely since the net effect for the user would be that pages written to would lose their contents.

You're only talking about moving bos, but what happens when you evict a bo? You *must* clear COW mappings, too. Otherwise, they might still get read-access to the evicted bo.

Note that we currently cause SIGBUS on access to mmap'd bos if they are evicted and cannot be restored (think: user-space mapped the VESA-FB but a real driver took over). So maybe we want two independent helpers here?

Thanks
David

Quoted 24 lines
---
 include/drm/drm_vma_manager.h |    6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
include/drm/drm_vma_manager.h +3 -4
index c18a593..347077d 100644
@@ -231,9 +231,9 @@ static inline void drm_vma_node_unmap(struct drm_vma_offset_node *node,
                                      struct address_space *file_mapping)
 {
        if (file_mapping && drm_vma_node_has_offset(node))
-               unmap_mapping_range(file_mapping,
-                                   drm_vma_node_offset_addr(node),
-                                   drm_vma_node_size(node) << PAGE_SHIFT, 1);
+               unmap_shared_mapping_range
+                       (file_mapping, drm_vma_node_offset_addr(node),
+                        drm_vma_node_size(node) << PAGE_SHIFT);
 }
 
 /**
--

1.7.10.4

#2 lore raw
Quoted 11 lines · David Herrmann

Hi

Quoted 4 lines · Thomas Hellstrom

Not sure if there are any user-space users of private bo mappings, but if there are, or will be, zapping the COW'd pages when, for example, moving a bo would confuse the user immensely since the net effect for the user would be that pages written to would lose their contents.

You're only talking about moving bos, but what happens when you evict a bo? You *must* clear COW mappings, too. Otherwise, they might still get read-access to the evicted bo.

Hmm. When I talk about COW'd pages, I mean individual anonymous pages.
If there is a COW mapping left in a VMA, it points to an anonymous page and will never access the evicted bo. All (read-only-enabled) ptes pointing to the evicted bo would still be zapped, so I'm afraid I don't understand your comment?

Quoted 5 lines

Note that we currently cause SIGBUS on access to mmap'd bos if they are evicted and cannot be restored (think: user-space mapped the VESA-FB but a real driver took over). So maybe we want two independent helpers here?

It shouldn't be needed.

Thanks,
Thomas

Quoted 28 lines

Thanks
David

Quoted 24 lines
---
  include/drm/drm_vma_manager.h |    6 +++---
  1 file changed, 3 insertions(+), 3 deletions(-)
include/drm/drm_vma_manager.h +3 -4
index c18a593..347077d 100644
@@ -231,9 +231,9 @@ static inline void drm_vma_node_unmap(struct drm_vma_offset_node *node,
                                       struct address_space *file_mapping)
  {
         if (file_mapping && drm_vma_node_has_offset(node))
-               unmap_mapping_range(file_mapping,
-                                   drm_vma_node_offset_addr(node),
-                                   drm_vma_node_size(node) << PAGE_SHIFT, 1);
+               unmap_shared_mapping_range
+                       (file_mapping, drm_vma_node_offset_addr(node),
+                        drm_vma_node_size(node) << PAGE_SHIFT);
  }
 
  /**
--

1.7.10.4

Reviewed-by on v1 1/1
#3 lore raw
In reply to Thomas Hellstrom
Quoted 6 lines · Thomas Hellstrom

Not sure if there are any user-space users of private bo mappings, but if there are, or will be, zapping the COW'd pages when, for example, moving a bo would confuse the user immensely since the net effect for the user would be that pages written to would lose their contents.

Presuming I'm not horribly confused about that all the vm slang in the kerneldoc means this changes is

Now I still hold that userspace creating anynomous bo mappings is rather crazy, but meh ;-) I guess the real question is whether we have anyone relying on this out there (or planing to), in which case we either need to funnel this through stable kernels or whack a drm feature flag onto drivers/kernels with this fixed. But I seriously hope the answer is no.

Cheers, Daniel

Quoted 27 lines
---
 include/drm/drm_vma_manager.h |    6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
include/drm/drm_vma_manager.h +3 -3
index c18a593..347077d 100644
@@ -231,9 +231,9 @@ static inline void drm_vma_node_unmap(struct drm_vma_offset_node *node,
 				      struct address_space *file_mapping)
 {
 	if (file_mapping && drm_vma_node_has_offset(node))
-		unmap_mapping_range(file_mapping,
-				    drm_vma_node_offset_addr(node),
-				    drm_vma_node_size(node) << PAGE_SHIFT, 1);
+		unmap_shared_mapping_range
+			(file_mapping, drm_vma_node_offset_addr(node),
+			 drm_vma_node_size(node) << PAGE_SHIFT);
 }
 
 /**
Signature · 5 lines
1.7.10.4
_______________________________________________
dri-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/dri-devel
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#4 lore raw
Quoted 12 lines · Daniel Vetter
Quoted 6 lines · Thomas Hellstrom

Not sure if there are any user-space users of private bo mappings, but if there are, or will be, zapping the COW'd pages when, for example, moving a bo would confuse the user immensely since the net effect for the user would be that pages written to would lose their contents.

Presuming I'm not horribly confused about that all the vm slang in the kerneldoc means this changes is

Quoted 8 lines

Now I still hold that userspace creating anynomous bo mappings is rather crazy, but meh ;-) I guess the real question is whether we have anyone relying on this out there (or planing to), in which case we either need to funnel this through stable kernels or whack a drm feature flag onto drivers/kernels with this fixed. But I seriously hope the answer is no.

Cheers, Daniel

Thanks for reviewing. I don't think there's a need to take this through stable, since I don't know of anyone using private VMAs,

Actually I'll need to hold off on this for a while since on some archs this may cause
ptes of shared pages to not be zapped. If the arch doesn't have PTE_SPECIAL, shared pages on MIXEDMAP vmas will come through as vm_normal_page, and since page->mapping is usually (un)set to NULL by our drivers, this will result in false positives for COW'ed pages.

So that test is buggy, or us not setting page->mapping to the mapping we use is buggy. It's too late in the day to decide which.

/Thomas

#5 lore raw
Quoted 32 lines · Thomas Hellstrom
Quoted 12 lines · Daniel Vetter
Quoted 6 lines · Thomas Hellstrom

Not sure if there are any user-space users of private bo mappings, but if there are, or will be, zapping the COW'd pages when, for example, moving a bo would confuse the user immensely since the net effect for the user would be that pages written to would lose their contents.

Presuming I'm not horribly confused about that all the vm slang in the kerneldoc means this changes is

Quoted 8 lines

Now I still hold that userspace creating anynomous bo mappings is rather crazy, but meh ;-) I guess the real question is whether we have anyone relying on this out there (or planing to), in which case we either need to funnel this through stable kernels or whack a drm feature flag onto drivers/kernels with this fixed. But I seriously hope the answer is no.

Cheers, Daniel

Thanks for reviewing. I don't think there's a need to take this through stable, since I don't know of anyone using private VMAs,

Actually I'll need to hold off on this for a while since on some archs this may cause
ptes of shared pages to not be zapped. If the arch doesn't have PTE_SPECIAL, shared pages on MIXEDMAP vmas will come through as vm_normal_page, and since page->mapping is usually (un)set to NULL by our drivers, this will result in false positives for COW'ed pages.

Quoted 2 lines

So that test is buggy, or us not setting page->mapping to the mapping we use is buggy. It's too late in the day to decide which.

On second thought we also use this helper when zapping a bo when deleting it or purging it's backing storage. And there I'd say it would be consistent with normal semantics to also zap cow pages. So I'm no longer sure this is the right approach in general. Maybe we should instead wire up the switch, but not sure whether that's indeed worth the effort. -Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#6 lore raw
Quoted 33 lines · Daniel Vetter
Quoted 32 lines · Thomas Hellstrom
Quoted 12 lines · Daniel Vetter
Quoted 6 lines · Thomas Hellstrom

Not sure if there are any user-space users of private bo mappings, but if there are, or will be, zapping the COW'd pages when, for example, moving a bo would confuse the user immensely since the net effect for the user would be that pages written to would lose their contents.

Presuming I'm not horribly confused about that all the vm slang in the kerneldoc means this changes is

Quoted 8 lines

Now I still hold that userspace creating anynomous bo mappings is rather crazy, but meh ;-) I guess the real question is whether we have anyone relying on this out there (or planing to), in which case we either need to funnel this through stable kernels or whack a drm feature flag onto drivers/kernels with this fixed. But I seriously hope the answer is no.

Cheers, Daniel

Thanks for reviewing. I don't think there's a need to take this through stable, since I don't know of anyone using private VMAs,

Actually I'll need to hold off on this for a while since on some archs this may cause
ptes of shared pages to not be zapped. If the arch doesn't have PTE_SPECIAL, shared pages on MIXEDMAP vmas will come through as vm_normal_page, and since page->mapping is usually (un)set to NULL by our drivers, this will result in false positives for COW'ed pages.

Quoted 5 lines
Quoted 2 lines

So that test is buggy, or us not setting page->mapping to the mapping we use is buggy. It's too late in the day to decide which.

On second thought we also use this helper when zapping a bo when deleting it or purging it's backing storage. And there I'd say it would be consistent with normal semantics to also zap cow pages.

Agreed. That's consistent with truncation. Perhaps this was what David meant. Out of interest, why do you delete a bo or purge backing store before all vmas are gone?

Quoted 3 lines

So I'm no longer
sure this is the right approach in general. Maybe we should instead wire up the switch, but not sure whether that's indeed worth the effort.

I'll leave it as is. If I happen to sort out the non-cow zapping, I'll post a new patch.

/Thomas

Quoted 1 line

-Daniel

#7 lore raw
Quoted 36 lines · Thomas Hellstrom
Quoted 35 lines · Daniel Vetter
Quoted 34 lines · Thomas Hellstrom
Quoted 12 lines · Daniel Vetter
Quoted 6 lines · Thomas Hellstrom

Not sure if there are any user-space users of private bo mappings, but if there are, or will be, zapping the COW'd pages when, for example, moving a bo would confuse the user immensely since the net effect for the user would be that pages written to would lose their contents.

Presuming I'm not horribly confused about that all the vm slang in the kerneldoc means this changes is

Quoted 8 lines

Now I still hold that userspace creating anynomous bo mappings is rather crazy, but meh ;-) I guess the real question is whether we have anyone relying on this out there (or planing to), in which case we either need to funnel this through stable kernels or whack a drm feature flag onto drivers/kernels with this fixed. But I seriously hope the answer is no.

Cheers, Daniel

Thanks for reviewing. I don't think there's a need to take this through stable, since I don't know of anyone using private VMAs,

Actually I'll need to hold off on this for a while since on some archs this may cause
ptes of shared pages to not be zapped. If the arch doesn't have PTE_SPECIAL, shared pages on MIXEDMAP vmas will come through as vm_normal_page, and since page->mapping is usually (un)set to NULL by our drivers, this will result in false positives for COW'ed pages.

So that test is buggy, or us not setting page->mapping to the mapping we use is buggy. It's too late in the day to decide which.

Quoted 8 lines
Quoted 3 lines

On second thought we also use this helper when zapping a bo when deleting it or purging it's backing storage. And there I'd say it would be consistent with normal semantics to also zap cow pages.

Agreed. That's consistent with truncation. Perhaps this was what David meant.
Out of interest, why do you delete a bo or purge backing store before all vmas are gone?

I don't think we actually delete it before all the vmas are gone, but purging can happen. Userspace (libdrm) aggressively caches both bos and their memory mappings, but sets them to purgeable. Hence when the kernel is low on memory and we preferentially remove those objects we can end up purging backing storage for which userspace still has a cached mmaping.

Of course userspace first needs to adjust the madvise flag to willneed before it can use that mmap region again. And if the kernel tells it that the object is unfortunately gone userspace can only do an munmap and close its gem handle.
-Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch