When amdgpu_dm_atomic_commit_tail is running in the workqueue, drm_atomic_state_put will get called while amdgpu_dm_atomic_commit_tail is running, causing a race condition where state (and then dm_state) is sometimes freed while amdgpu_dm_atomic_commit_tail is running. This bug has occurred since 5.7-rc1 and is well documented among polaris11 users [1].
Prior to 5.7, this was not a noticeable issue since the freelist pointer was stored at the beginning of dm_state (base), which was unused. After changing the freelist pointer to be stored in the middle of the struct, the freelist pointer overwrote the context, causing dc_state to become garbage data and made the call to dm_enable_per_frame_crtc_master_sync dereference a freelist pointer.
This patch fixes the aforementioned issue by calling drm_atomic_state_get in amdgpu_dm_atomic_commit before drm_atomic_helper_commit is called and drm_atomic_state_put after amdgpu_dm_atomic_commit_tail is complete.
According to my testing on 5.8.0-rc6, this should fix bug 207383 on Bugzilla [1].
[1] https://bugzilla.kernel.org/show_bug.cgi?id=207383
Fixes: 3202fa62f ("slub: relocate freelist pointer to middle of object") Reported-by: Duncan 1i5t5.duncan@cox.net Signed-off-by: Mazin Rezk mnrzk@protonmail.com --- drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c index 86ffa0c2880f..86d6652872f2 100644 --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c @@ -7303,6 +7303,7 @@ static int amdgpu_dm_atomic_commit(struct drm_device *dev, * unset legacy_cursor_update */
+ drm_atomic_state_get(state); return drm_atomic_helper_commit(dev, state, nonblock);
/*TODO Handle EINTR, reenable IRQ*/ @@ -7628,6 +7629,8 @@ static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_state *state)
if (dc_state_temp) dc_release_state(dc_state_temp); + + drm_atomic_state_put(state); }
-- 2.27.0
On 2020-07-23 5:10 p.m., Mazin Rezk wrote:
When amdgpu_dm_atomic_commit_tail is running in the workqueue, drm_atomic_state_put will get called while amdgpu_dm_atomic_commit_tail is running, causing a race condition where state (and then dm_state) is sometimes freed while amdgpu_dm_atomic_commit_tail is running. This bug has occurred since 5.7-rc1 and is well documented among polaris11 users [1].
Prior to 5.7, this was not a noticeable issue since the freelist pointer was stored at the beginning of dm_state (base), which was unused. After changing the freelist pointer to be stored in the middle of the struct, the freelist pointer overwrote the context, causing dc_state to become garbage data and made the call to dm_enable_per_frame_crtc_master_sync dereference a freelist pointer.
This patch fixes the aforementioned issue by calling drm_atomic_state_get in amdgpu_dm_atomic_commit before drm_atomic_helper_commit is called and drm_atomic_state_put after amdgpu_dm_atomic_commit_tail is complete.
According to my testing on 5.8.0-rc6, this should fix bug 207383 on Bugzilla [1].
[1] https://bugzilla.kernel.org/show_bug.cgi?id=207383
Fixes: 3202fa62f ("slub: relocate freelist pointer to middle of object") Reported-by: Duncan 1i5t5.duncan@cox.net Signed-off-by: Mazin Rezk mnrzk@protonmail.com
Thanks for the investigation and your patch. I appreciate the help in trying to narrow down the root cause as this issue has been difficult to reproduce on my setups.
Though I'm not sure this really resolves the issue - we make use of the drm_atomic_helper_commit helper function from DRM which internally does what you're doing with this patch:
drm_atomic_state_get(state); if (nonblock) queue_work(system_unbound_wq, &state->commit_work); else commit_tail(state);
So even when it gets queued off to the unbound workqueue we still have a reference on the state.
That reference gets dropped as part of commit tail helper in DRM as well:
if (funcs && funcs->atomic_commit_tail) funcs->atomic_commit_tail(old_state); else drm_atomic_helper_commit_tail(old_state);
commit_time_ms = ktime_ms_delta(ktime_get(), start); if (commit_time_ms > 0) drm_self_refresh_helper_update_avg_times(old_state, (unsigned long)commit_time_ms, new_self_refresh_mask);
drm_atomic_helper_commit_cleanup_done(old_state);
drm_atomic_state_put(old_state);
So instead of a use after free happening when we access the state we get a double-free happening later at the end of commit tail in DRM.
What I think would be the right next step here is to actually determine what sequence of IOCTLs and atomic commits are happening under your setup with a very verbose dmesg log. You can set a debug level for DRM in your kernel parameters with something like:
drm.debug=0x54
I don't see anything in amdgpu_dm.c that looks like it would be freeing the state so I suspect something in the core is this doing this.
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c index 86ffa0c2880f..86d6652872f2 100644 --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c @@ -7303,6 +7303,7 @@ static int amdgpu_dm_atomic_commit(struct drm_device *dev, * unset legacy_cursor_update */
- drm_atomic_state_get(state);
Also note that if the drm_atomic_helper_commit() call fails here then we're going to never free this structure. So we should really be checking the return code here below before trying to do this, if at all.
Regards, Nicholas Kazlauskas
return drm_atomic_helper_commit(dev, state, nonblock);
/*TODO Handle EINTR, reenable IRQ*/ @@ -7628,6 +7629,8 @@ static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_state *state)
if (dc_state_temp) dc_release_state(dc_state_temp);
- drm_atomic_state_put(state); }
-- 2.27.0
It seems that I spoke too soon. I ran the system for another hour after submitting the patch and the bug just occurred. :/
Sadly, that means the bug isn't really fixed and that I have to go investigate further.
At the very least, this patch seems to delay the occurrence of the bug significantly which may help in further discovering the cause.
On Thursday, July 23, 2020 6:16 PM, Kazlauskas, Nicholas nicholas.kazlauskas@amd.com wrote:
On 2020-07-23 5:10 p.m., Mazin Rezk wrote:
When amdgpu_dm_atomic_commit_tail is running in the workqueue, drm_atomic_state_put will get called while amdgpu_dm_atomic_commit_tail is running, causing a race condition where state (and then dm_state) is sometimes freed while amdgpu_dm_atomic_commit_tail is running. This bug has occurred since 5.7-rc1 and is well documented among polaris11 users [1]. Prior to 5.7, this was not a noticeable issue since the freelist pointer was stored at the beginning of dm_state (base), which was unused. After changing the freelist pointer to be stored in the middle of the struct, the freelist pointer overwrote the context, causing dc_state to become garbage data and made the call to dm_enable_per_frame_crtc_master_sync dereference a freelist pointer. This patch fixes the aforementioned issue by calling drm_atomic_state_get in amdgpu_dm_atomic_commit before drm_atomic_helper_commit is called and drm_atomic_state_put after amdgpu_dm_atomic_commit_tail is complete. According to my testing on 5.8.0-rc6, this should fix bug 207383 on Bugzilla [1]. [1] https://bugzilla.kernel.org/show_bug.cgi?id=207383 Fixes: 3202fa62f ("slub: relocate freelist pointer to middle of object") Reported-by: Duncan 1i5t5.duncan@cox.net Signed-off-by: Mazin Rezk mnrzk@protonmail.com
Thanks for the investigation and your patch. I appreciate the help in trying to narrow down the root cause as this issue has been difficult to reproduce on my setups.
Though I'm not sure this really resolves the issue - we make use of the drm_atomic_helper_commit helper function from DRM which internally does what you're doing with this patch:
drm_atomic_state_get(state); if (nonblock) queue_work(system_unbound_wq, &state->commit_work);
else commit_tail(state);So even when it gets queued off to the unbound workqueue we still have a reference on the state.
That reference gets dropped as part of commit tail helper in DRM as well:
if (funcs && funcs->atomic_commit_tail)
funcs->atomic_commit_tail(old_state); else drm_atomic_helper_commit_tail(old_state);commit_time_ms = ktime_ms_delta(ktime_get(), start); if (commit_time_ms > 0)
drm_self_refresh_helper_update_avg_times(old_state, (unsigned long)commit_time_ms, new_self_refresh_mask);drm_atomic_helper_commit_cleanup_done(old_state);
drm_atomic_state_put(old_state);
I initially noticed that right after I wrote this patch so I was expecting the patch to fail. However, after several hours of testing, the crash just didn't occur so I believed the bug was fixed.
So instead of a use after free happening when we access the state we get a double-free happening later at the end of commit tail in DRM.
What I think would be the right next step here is to actually determine what sequence of IOCTLs and atomic commits are happening under your setup with a very verbose dmesg log. You can set a debug level for DRM in your kernel parameters with something like:
drm.debug=0x54
I don't see anything in amdgpu_dm.c that looks like it would be freeing the state so I suspect something in the core is this doing this.
Going through the KASAN use-after-free bug report in the Bugzilla attachments, it appears that the state is being freed at the end of commit_tail. Perhaps amdgpu_dm_atomic_commit_tail is being called on the the same old state twice? I can't quite think of any other possible explanation as to why that happens.
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 +++ 1 file changed, 3 insertions(+) diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c index 86ffa0c2880f..86d6652872f2 100644 --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c @@ -7303,6 +7303,7 @@ static int amdgpu_dm_atomic_commit(struct drm_device *dev,
- unset legacy_cursor_update
*/
- drm_atomic_state_get(state);
Also note that if the drm_atomic_helper_commit() call fails here then we're going to never free this structure. So we should really be checking the return code here below before trying to do this, if at all.
Oh right, that's true. I looked at amdgpu_dm_atomic_commit_tail and didn't see any return statements in there, so I thought it was safe.
Regards, Nicholas Kazlauskas
return drm_atomic_helper_commit(dev, state, nonblock); /*TODO Handle EINTR, reenable IRQ*/@@ -7628,6 +7629,8 @@ static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_state *state)
if (dc_state_temp) dc_release_state(dc_state_temp);
- drm_atomic_state_put(state); }
-- 2.27.0
Thanks, Mazin Rezk
On Thursday, July 23, 2020 6:57 PM, Mazin Rezk mnrzk@protonmail.com wrote:
It seems that I spoke too soon. I ran the system for another hour after submitting the patch and the bug just occurred. :/
Sadly, that means the bug isn't really fixed and that I have to go investigate further.
At the very least, this patch seems to delay the occurrence of the bug significantly which may help in further discovering the cause.
On Thursday, July 23, 2020 6:16 PM, Kazlauskas, Nicholas nicholas.kazlauskas@amd.com wrote:
On 2020-07-23 5:10 p.m., Mazin Rezk wrote:
When amdgpu_dm_atomic_commit_tail is running in the workqueue, drm_atomic_state_put will get called while amdgpu_dm_atomic_commit_tail is running, causing a race condition where state (and then dm_state) is sometimes freed while amdgpu_dm_atomic_commit_tail is running. This bug has occurred since 5.7-rc1 and is well documented among polaris11 users [1]. Prior to 5.7, this was not a noticeable issue since the freelist pointer was stored at the beginning of dm_state (base), which was unused. After changing the freelist pointer to be stored in the middle of the struct, the freelist pointer overwrote the context, causing dc_state to become garbage data and made the call to dm_enable_per_frame_crtc_master_sync dereference a freelist pointer. This patch fixes the aforementioned issue by calling drm_atomic_state_get in amdgpu_dm_atomic_commit before drm_atomic_helper_commit is called and drm_atomic_state_put after amdgpu_dm_atomic_commit_tail is complete. According to my testing on 5.8.0-rc6, this should fix bug 207383 on Bugzilla [1]. [1] https://bugzilla.kernel.org/show_bug.cgi?id=207383 Fixes: 3202fa62f ("slub: relocate freelist pointer to middle of object") Reported-by: Duncan 1i5t5.duncan@cox.net Signed-off-by: Mazin Rezk mnrzk@protonmail.com
Thanks for the investigation and your patch. I appreciate the help in trying to narrow down the root cause as this issue has been difficult to reproduce on my setups. Though I'm not sure this really resolves the issue - we make use of the drm_atomic_helper_commit helper function from DRM which internally does what you're doing with this patch: drm_atomic_state_get(state); if (nonblock) queue_work(system_unbound_wq, &state->commit_work);
else commit_tail(state);So even when it gets queued off to the unbound workqueue we still have a reference on the state. That reference gets dropped as part of commit tail helper in DRM as well: if (funcs && funcs->atomic_commit_tail)
funcs->atomic_commit_tail(old_state); else drm_atomic_helper_commit_tail(old_state);commit_time_ms = ktime_ms_delta(ktime_get(), start); if (commit_time_ms > 0)
drm_self_refresh_helper_update_avg_times(old_state, (unsigned long)commit_time_ms, new_self_refresh_mask);drm_atomic_helper_commit_cleanup_done(old_state); drm_atomic_state_put(old_state);
I initially noticed that right after I wrote this patch so I was expecting the patch to fail. However, after several hours of testing, the crash just didn't occur so I believed the bug was fixed.
So instead of a use after free happening when we access the state we get a double-free happening later at the end of commit tail in DRM. What I think would be the right next step here is to actually determine what sequence of IOCTLs and atomic commits are happening under your setup with a very verbose dmesg log. You can set a debug level for DRM in your kernel parameters with something like: drm.debug=0x54 I don't see anything in amdgpu_dm.c that looks like it would be freeing the state so I suspect something in the core is this doing this.
Going through the KASAN use-after-free bug report in the Bugzilla attachments, it appears that the state is being freed at the end of commit_tail. Perhaps amdgpu_dm_atomic_commit_tail is being called on the the same old state twice? I can't quite think of any other possible explanation as to why that happens.
I think I've more or less confirmed that this is the case.
I created two padding variables, one to store debug magic numbers and one to store the freelist pointer. I had magic numbers for initialised, preuse, and used states. When the dm_atomic_state is initialised, the padding is set to the init magic number. Right before commit_tail is called, the padding is set to the preuse magic number. During dm_atomic_get_new_state checks the magic number to confirm that it was in the preuse state and then set it to used. If it failed that check and it was already in a used state, there was a breakpoint set so I could gather further information.
At one point (presumably where the crash would have occurred), the debug padding variable was set to the used state during the call to commit_tail which I believe confirms my guess that amdgpu_dm_atomic_commit_tail is being called on the same state twice.
What's weird, however, is that dmesg (w/ drm.debug=0x54) says this right before amdgpu_dm_atomic_commit_tail is called:
[ 3277.580205] [drm:drm_atomic_state_init [drm]] Allocated atomic state 00000000a06f4024 [ 3277.580262] [drm:drm_atomic_get_crtc_state [drm]] Added [CRTC:49:crtc-1] 000000003b9da5c1 state to 00000000a06f4024 [ 3277.580316] [drm:drm_atomic_get_plane_state [drm]] Added [PLANE:44:plane-4] 000000003488c027 state to 00000000a06f4024 [ 3277.580366] [drm:drm_atomic_set_fb_for_plane [drm]] Set [FB:103] for [PLANE:44:plane-4] state 000000003488c027 [ 3277.580417] [drm:drm_atomic_check_only [drm]] checking 00000000a06f4024 [ 3277.580519] [drm:drm_atomic_get_private_obj_state [drm]] Added new private object 0000000002a633ab state 00000000695dff15 to 00000000a06f4024 [ 3277.580579] [drm:drm_atomic_nonblocking_commit [drm]] committing 00000000a06f4024 nonblocking [ 3277.582325] [drm:drm_atomic_state_default_clear [drm]] Clearing atomic state 00000000a06f4024 [ 3277.582393] [drm:__drm_atomic_state_free [drm]] Freeing atomic state 00000000a06f4024
From the log, I'm noticing that drm_atomic_nonblocking_commit is only
called once and that whatever is calling the second non-blocking commit_tail on the same state doesn't seem to be using drm_atomic_nonblocking_commit.
Perhaps someone with more knowledge of the code can give a possible explanation as to why that's happening.
Thanks, Mazin Rezk
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 +++ 1 file changed, 3 insertions(+) diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c index 86ffa0c2880f..86d6652872f2 100644 --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c @@ -7303,6 +7303,7 @@ static int amdgpu_dm_atomic_commit(struct drm_device *dev,
unset legacy_cursor_update */
drm_atomic_state_get(state);
Also note that if the drm_atomic_helper_commit() call fails here then we're going to never free this structure. So we should really be checking the return code here below before trying to do this, if at all.
Oh right, that's true. I looked at amdgpu_dm_atomic_commit_tail and didn't see any return statements in there, so I thought it was safe.
Regards, Nicholas Kazlauskas
return drm_atomic_helper_commit(dev, state, nonblock); /*TODO Handle EINTR, reenable IRQ*/@@ -7628,6 +7629,8 @@ static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_state *state)
if (dc_state_temp) dc_release_state(dc_state_temp);
- drm_atomic_state_put(state); }
-- 2.27.0
Thanks, Mazin Rezk
On Thu, Jul 23, 2020 at 09:10:15PM +0000, Mazin Rezk wrote:
When amdgpu_dm_atomic_commit_tail is running in the workqueue, drm_atomic_state_put will get called while amdgpu_dm_atomic_commit_tail is running, causing a race condition where state (and then dm_state) is sometimes freed while amdgpu_dm_atomic_commit_tail is running. This bug has occurred since 5.7-rc1 and is well documented among polaris11 users [1].
Prior to 5.7, this was not a noticeable issue since the freelist pointer was stored at the beginning of dm_state (base), which was unused. After changing the freelist pointer to be stored in the middle of the struct, the freelist pointer overwrote the context, causing dc_state to become garbage data and made the call to dm_enable_per_frame_crtc_master_sync dereference a freelist pointer.
This patch fixes the aforementioned issue by calling drm_atomic_state_get in amdgpu_dm_atomic_commit before drm_atomic_helper_commit is called and drm_atomic_state_put after amdgpu_dm_atomic_commit_tail is complete.
According to my testing on 5.8.0-rc6, this should fix bug 207383 on Bugzilla [1].
Nice work tracking this down!
Fixes: 3202fa62f ("slub: relocate freelist pointer to middle of object")
I do, however, object to this Fixes tag. :) The flaw appears to have been with amdgpu_dm's reference tracking of "state" in the nonblocking case. (How this reference counting is supposed to work correctly, though, I'm not sure.) If I look at where the drm helper was split from being the default callback, it looks like this was what introduced the bug:
da5c47f682ab ("drm/amd/display: Remove acrtc->stream")
? 3202fa62f certainly exposed it much more quickly, but there was a race even without 3202fa62f where something could have realloced the memory and written over it.
On Thursday, July 23, 2020 6:32 PM, Kees Cook keescook@chromium.org wrote:
On Thu, Jul 23, 2020 at 09:10:15PM +0000, Mazin Rezk wrote:
When amdgpu_dm_atomic_commit_tail is running in the workqueue, drm_atomic_state_put will get called while amdgpu_dm_atomic_commit_tail is running, causing a race condition where state (and then dm_state) is sometimes freed while amdgpu_dm_atomic_commit_tail is running. This bug has occurred since 5.7-rc1 and is well documented among polaris11 users [1]. Prior to 5.7, this was not a noticeable issue since the freelist pointer was stored at the beginning of dm_state (base), which was unused. After changing the freelist pointer to be stored in the middle of the struct, the freelist pointer overwrote the context, causing dc_state to become garbage data and made the call to dm_enable_per_frame_crtc_master_sync dereference a freelist pointer. This patch fixes the aforementioned issue by calling drm_atomic_state_get in amdgpu_dm_atomic_commit before drm_atomic_helper_commit is called and drm_atomic_state_put after amdgpu_dm_atomic_commit_tail is complete. According to my testing on 5.8.0-rc6, this should fix bug 207383 on Bugzilla [1]. [1] https://bugzilla.kernel.org/show_bug.cgi?id=207383
Nice work tracking this down!
Fixes: 3202fa62f ("slub: relocate freelist pointer to middle of object")
I do, however, object to this Fixes tag. :) The flaw appears to have been with amdgpu_dm's reference tracking of "state" in the nonblocking case. (How this reference counting is supposed to work correctly, though, I'm not sure.) If I look at where the drm helper was split from being the default callback, it looks like this was what introduced the bug:
da5c47f682ab ("drm/amd/display: Remove acrtc->stream")
? 3202fa62f certainly exposed it much more quickly, but there was a race even without 3202fa62f where something could have realloced the memory and written over it.
Kees Cook
Thanks, I'll be sure to avoid using 3202fa62f as the cause next time. I just thought to do that because it was what made the use-after-free cause a noticeable bug.
Also, by the way, I just realised the patch didn't completely solve the bug. Sorry about that, making an LKML thread on this was hasty on my part. Should I get further confirmation from the Bugzilla thread before submitting a patch for this bug in the future?
Thanks, Mazin Rezk
Am 24.07.20 um 00:58 schrieb Mazin Rezk:
On Thursday, July 23, 2020 6:32 PM, Kees Cook keescook@chromium.org wrote:
On Thu, Jul 23, 2020 at 09:10:15PM +0000, Mazin Rezk wrote:
When amdgpu_dm_atomic_commit_tail is running in the workqueue, drm_atomic_state_put will get called while amdgpu_dm_atomic_commit_tail is running, causing a race condition where state (and then dm_state) is sometimes freed while amdgpu_dm_atomic_commit_tail is running. This bug has occurred since 5.7-rc1 and is well documented among polaris11 users [1]. Prior to 5.7, this was not a noticeable issue since the freelist pointer was stored at the beginning of dm_state (base), which was unused. After changing the freelist pointer to be stored in the middle of the struct, the freelist pointer overwrote the context, causing dc_state to become garbage data and made the call to dm_enable_per_frame_crtc_master_sync dereference a freelist pointer. This patch fixes the aforementioned issue by calling drm_atomic_state_get in amdgpu_dm_atomic_commit before drm_atomic_helper_commit is called and drm_atomic_state_put after amdgpu_dm_atomic_commit_tail is complete. According to my testing on 5.8.0-rc6, this should fix bug 207383 on Bugzilla [1]. [1] https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fbugzilla.k...
Nice work tracking this down!
Fixes: 3202fa62f ("slub: relocate freelist pointer to middle of object")
I do, however, object to this Fixes tag. :) The flaw appears to have been with amdgpu_dm's reference tracking of "state" in the nonblocking case. (How this reference counting is supposed to work correctly, though, I'm not sure.) If I look at where the drm helper was split from being the default callback, it looks like this was what introduced the bug:
da5c47f682ab ("drm/amd/display: Remove acrtc->stream")
? 3202fa62f certainly exposed it much more quickly, but there was a race even without 3202fa62f where something could have realloced the memory and written over it.
Kees Cook
Thanks, I'll be sure to avoid using 3202fa62f as the cause next time. I just thought to do that because it was what made the use-after-free cause a noticeable bug.
Also, by the way, I just realised the patch didn't completely solve the bug. Sorry about that, making an LKML thread on this was hasty on my part. Should I get further confirmation from the Bugzilla thread before submitting a patch for this bug in the future?
Submitting stuff as early as possible is mostly a good idea. Just if the code is utterly broken or completely unreadable you should probably expect a harsh response :)
Maybe ask for more testing in the commit message if you are not 100% sure if that really fixes a bug or not.
Regards, Christian.
Thanks, Mazin Rezk
Dear Kees,
Am 24.07.20 um 00:32 schrieb Kees Cook:
On Thu, Jul 23, 2020 at 09:10:15PM +0000, Mazin Rezk wrote:
When amdgpu_dm_atomic_commit_tail is running in the workqueue, drm_atomic_state_put will get called while amdgpu_dm_atomic_commit_tail is running, causing a race condition where state (and then dm_state) is sometimes freed while amdgpu_dm_atomic_commit_tail is running. This bug has occurred since 5.7-rc1 and is well documented among polaris11 users [1].
Prior to 5.7, this was not a noticeable issue since the freelist pointer was stored at the beginning of dm_state (base), which was unused. After changing the freelist pointer to be stored in the middle of the struct, the freelist pointer overwrote the context, causing dc_state to become garbage data and made the call to dm_enable_per_frame_crtc_master_sync dereference a freelist pointer.
This patch fixes the aforementioned issue by calling drm_atomic_state_get in amdgpu_dm_atomic_commit before drm_atomic_helper_commit is called and drm_atomic_state_put after amdgpu_dm_atomic_commit_tail is complete.
According to my testing on 5.8.0-rc6, this should fix bug 207383 on Bugzilla [1].
Nice work tracking this down!
Fixes: 3202fa62f ("slub: relocate freelist pointer to middle of object")
I do, however, object to this Fixes tag. :) The flaw appears to have been with amdgpu_dm's reference tracking of "state" in the nonblocking case. (How this reference counting is supposed to work correctly, though, I'm not sure.) If I look at where the drm helper was split from being the default callback, it looks like this was what introduced the bug:
da5c47f682ab ("drm/amd/display: Remove acrtc->stream")
? 3202fa62f certainly exposed it much more quickly, but there was a race even without 3202fa62f where something could have realloced the memory and written over it.
I understand the Fixes tag mainly a help when backporting commits.
As Linux 5.8-rc7 is going to be released this Sunday, I wonder, if commit 3202fa62f ("slub: relocate freelist pointer to middle of object") should be reverted for now to fix the regression for the users according to Linux’ no regression policy. Once the AMDGPU/DRM driver issue is fixed, it can be reapplied. I know it’s not optimal, but as some testing is going to be involved for the fix, I’d argue it’s the best option for the users.
Kind regards,
Paul
On Fri, Jul 24, 2020 at 09:45:18AM +0200, Paul Menzel wrote:
Am 24.07.20 um 00:32 schrieb Kees Cook:
On Thu, Jul 23, 2020 at 09:10:15PM +0000, Mazin Rezk wrote:
As Linux 5.8-rc7 is going to be released this Sunday, I wonder, if commit 3202fa62f ("slub: relocate freelist pointer to middle of object") should be reverted for now to fix the regression for the users according to Linux’ no regression policy. Once the AMDGPU/DRM driver issue is fixed, it can be reapplied. I know it’s not optimal, but as some testing is going to be involved for the fix, I’d argue it’s the best option for the users.
Well, the SLUB defense was already released in v5.7, so I'm not sure it really helps for amdgpu_dm users seeing it there too. There was a fix to disable the async path for this driver that worked around the bug too, yes? That seems like a safer and more focused change that doesn't revert the SLUB defense for all users, and would actually provide a complete, I think, workaround whereas reverting the SLUB change means the race still exists. For example, it would be hit with slab poisoning, etc.
Dear Kees,
Am 24.07.20 um 19:33 schrieb Kees Cook:
On Fri, Jul 24, 2020 at 09:45:18AM +0200, Paul Menzel wrote:
Am 24.07.20 um 00:32 schrieb Kees Cook:
On Thu, Jul 23, 2020 at 09:10:15PM +0000, Mazin Rezk wrote:
As Linux 5.8-rc7 is going to be released this Sunday, I wonder, if commit 3202fa62f ("slub: relocate freelist pointer to middle of object") should be reverted for now to fix the regression for the users according to Linux’ no regression policy. Once the AMDGPU/DRM driver issue is fixed, it can be reapplied. I know it’s not optimal, but as some testing is going to be involved for the fix, I’d argue it’s the best option for the users.
Well, the SLUB defense was already released in v5.7, so I'm not sure it really helps for amdgpu_dm users seeing it there too.
In my opinion, it would help, as the stable release could pick up the revert, ones it’s in Linus’ master branch.
There was a fix to disable the async path for this driver that worked around the bug too, yes? That seems like a safer and more focused change that doesn't revert the SLUB defense for all users, and would actually provide a complete, I think, workaround whereas reverting the SLUB change means the race still exists. For example, it would be hit with slab poisoning, etc.
I do not know. If there is such a fix, that would be great. But if you do not know, how should a normal user? ;-)
Kind regards,
Paul
Kind regards,
Paul
On Friday, July 24, 2020 5:19 PM, Paul Menzel pmenzel@molgen.mpg.de wrote:
Dear Kees,
Am 24.07.20 um 19:33 schrieb Kees Cook:
On Fri, Jul 24, 2020 at 09:45:18AM +0200, Paul Menzel wrote:
Am 24.07.20 um 00:32 schrieb Kees Cook:
On Thu, Jul 23, 2020 at 09:10:15PM +0000, Mazin Rezk wrote: As Linux 5.8-rc7 is going to be released this Sunday, I wonder, if commit 3202fa62f ("slub: relocate freelist pointer to middle of object") should be reverted for now to fix the regression for the users according to Linux’ no regression policy. Once the AMDGPU/DRM driver issue is fixed, it can be reapplied. I know it’s not optimal, but as some testing is going to be involved for the fix, I’d argue it’s the best option for the users.
Well, the SLUB defense was already released in v5.7, so I'm not sure it really helps for amdgpu_dm users seeing it there too.
In my opinion, it would help, as the stable release could pick up the revert, ones it’s in Linus’ master branch.
There was a fix to disable the async path for this driver that worked around the bug too, yes? That seems like a safer and more focused change that doesn't revert the SLUB defense for all users, and would actually provide a complete, I think, workaround whereas reverting the SLUB change means the race still exists. For example, it would be hit with slab poisoning, etc.
I do not know. If there is such a fix, that would be great. But if you do not know, how should a normal user? ;-)
Kind regards,
Paul
Kind regards,
Paul
If we're talking about workarounds now, I suggest simply swapping the base and context variables in struct dm_atomic_state. By that way, we won't need to change non-amdgpu parts of the code (e.g. by reverting the SLUB patch).
Prior to 3202fa62f, the freelist pointer was stored in dm_state->base which was never dereferenced and therefore caused no noticeable issue. After 3202fa62f, the freelist pointer is stored in the middle of the struct (i.e. dm_state->context).
Swapping the position of the base and context variables in dm_atomic_state should, in theory, revert this code back to it's pre-5.7 state since the code would be back to overwriting base instead.
If we decide to use this workaround, I can write the patch and do more extended tests to confirm it works around the issues.
That said, I haven't seen the async disabling patch. If you could link to it, I'd be glad to test it out and perhaps we can use that instead.
Thanks, Mazin Rezk
On Sat, 25 Jul 2020 03:03:52 +0000 Mazin Rezk mnrzk@protonmail.com wrote:
Am 24.07.20 um 19:33 schrieb Kees Cook:
There was a fix to disable the async path for this driver that worked around the bug too, yes? That seems like a safer and more focused change that doesn't revert the SLUB defense for all users, and would actually provide a complete, I think, workaround
That said, I haven't seen the async disabling patch. If you could link to it, I'd be glad to test it out and perhaps we can use that instead.
I'm confused. Not to put words in Kees' mouth; /I/ am confused (which admittedly could well be just because I make no claims to be a coder and am simply reading the bug and thread, but I'd appreciate some "unconfusing" anyway).
My interpretation of the "async disabling" reference was that it was to comment #30 on the bug:
https://bugzilla.kernel.org/show_bug.cgi?id=207383#c30
... which (if I'm not confused on this point too) appears to be yours. There it was stated...
I've also found that this bug exclusively occurs when commit_work is on the workqueue. After forcing drm_atomic_helper_commit to run all of the commits without adding to the workqueue and running the OS, the issue seems to have disappeared. <<<<
Would not forcing all commits to run directly, without placing them on the workqueue, be "async disabling"? That's what I /thought/ he was referencing.
OTOH your base/context swap idea sounds like a possibly "less disturbance" workaround, if it works, and given the point in the commit cycle... (But if it's out Sunday it's likely too late to test and get it in now anyway; if it's another week, tho...)
On Saturday, July 25, 2020 12:59 AM, Duncan 1i5t5.duncan@cox.net wrote:
On Sat, 25 Jul 2020 03:03:52 +0000 Mazin Rezk mnrzk@protonmail.com wrote:
Am 24.07.20 um 19:33 schrieb Kees Cook:
There was a fix to disable the async path for this driver that worked around the bug too, yes? That seems like a safer and more focused change that doesn't revert the SLUB defense for all users, and would actually provide a complete, I think, workaround
That said, I haven't seen the async disabling patch. If you could link to it, I'd be glad to test it out and perhaps we can use that instead.
I'm confused. Not to put words in Kees' mouth; /I/ am confused (which admittedly could well be just because I make no claims to be a coder and am simply reading the bug and thread, but I'd appreciate some "unconfusing" anyway).
My interpretation of the "async disabling" reference was that it was to comment #30 on the bug:
https://bugzilla.kernel.org/show_bug.cgi?id=207383#c30
... which (if I'm not confused on this point too) appears to be yours. There it was stated...
I've also found that this bug exclusively occurs when commit_work is on the workqueue. After forcing drm_atomic_helper_commit to run all of the commits without adding to the workqueue and running the OS, the issue seems to have disappeared. <<<<
Would not forcing all commits to run directly, without placing them on the workqueue, be "async disabling"? That's what I /thought/ he was referencing.
Oh, I thought he was referring to a different patch. Kees, could I get your confirmation on this?
The change I made actually affected all of the DRM code, although this could easily be changed to be specific to amdgpu. (By forcing blocking on amdgpu_dm's non-blocking commit code)
That said, I'd still need to test further because I only did test it for a couple of hours then. Although it should work in theory.
OTOH your base/context swap idea sounds like a possibly "less disturbance" workaround, if it works, and given the point in the commit cycle... (But if it's out Sunday it's likely too late to test and get it in now anyway; if it's another week, tho...)
The base/context swap idea should make the use-after-free behave how it did in 5.6. Since the bug doesn't cause an issue in 5.6, it's less of a "less disturbance" workaround and more of a "no disturbance" workaround.
Thanks, Mazin Rezk
Duncan - No HTML messages please; they are filtered as spam. "Every nonfree program has a lord, a master -- and if you use the program, he is your master." Richard Stallman
Dear Linux folks,
Am 25.07.20 um 07:20 schrieb Mazin Rezk:
On Saturday, July 25, 2020 12:59 AM, Duncan wrote:
On Sat, 25 Jul 2020 03:03:52 +0000 Mazin Rezk wrote:
Am 24.07.20 um 19:33 schrieb Kees Cook:
There was a fix to disable the async path for this driver that worked around the bug too, yes? That seems like a safer and more focused change that doesn't revert the SLUB defense for all users, and would actually provide a complete, I think, workaround
That said, I haven't seen the async disabling patch. If you could link to it, I'd be glad to test it out and perhaps we can use that instead.
I'm confused. Not to put words in Kees' mouth; /I/ am confused (which admittedly could well be just because I make no claims to be a coder and am simply reading the bug and thread, but I'd appreciate some "unconfusing" anyway).
My interpretation of the "async disabling" reference was that it was to comment #30 on the bug:
https://bugzilla.kernel.org/show_bug.cgi?id=207383#c30
... which (if I'm not confused on this point too) appears to be yours. There it was stated...
I've also found that this bug exclusively occurs when commit_work is on the workqueue. After forcing drm_atomic_helper_commit to run all of the commits without adding to the workqueue and running the OS, the issue seems to have disappeared. <<<<
Would not forcing all commits to run directly, without placing them on the workqueue, be "async disabling"? That's what I /thought/ he was referencing.
Oh, I thought he was referring to a different patch. Kees, could I get your confirmation on this?
The change I made actually affected all of the DRM code, although this could easily be changed to be specific to amdgpu. (By forcing blocking on amdgpu_dm's non-blocking commit code)
That said, I'd still need to test further because I only did test it for a couple of hours then. Although it should work in theory.
OTOH your base/context swap idea sounds like a possibly "less disturbance" workaround, if it works, and given the point in the commit cycle... (But if it's out Sunday it's likely too late to test and get it in now anyway; if it's another week, tho...)
The base/context swap idea should make the use-after-free behave how it did in 5.6. Since the bug doesn't cause an issue in 5.6, it's less of a "less disturbance" workaround and more of a "no disturbance" workaround.
Sorry for bothering, but is there now a solution, besides reverting the commits, to avoid freezes/crashes *without* performance regressions?
Kind regards,
Paul
On 2020-07-28 5:22 a.m., Paul Menzel wrote:
Dear Linux folks,
Am 25.07.20 um 07:20 schrieb Mazin Rezk:
On Saturday, July 25, 2020 12:59 AM, Duncan wrote:
On Sat, 25 Jul 2020 03:03:52 +0000 Mazin Rezk wrote:
Am 24.07.20 um 19:33 schrieb Kees Cook:
There was a fix to disable the async path for this driver that worked around the bug too, yes? That seems like a safer and more focused change that doesn't revert the SLUB defense for all users, and would actually provide a complete, I think, workaround
That said, I haven't seen the async disabling patch. If you could link to it, I'd be glad to test it out and perhaps we can use that instead.
I'm confused. Not to put words in Kees' mouth; /I/ am confused (which admittedly could well be just because I make no claims to be a coder and am simply reading the bug and thread, but I'd appreciate some "unconfusing" anyway).
My interpretation of the "async disabling" reference was that it was to comment #30 on the bug:
https://bugzilla.kernel.org/show_bug.cgi?id=207383#c30
... which (if I'm not confused on this point too) appears to be yours. There it was stated...
I've also found that this bug exclusively occurs when commit_work is on the workqueue. After forcing drm_atomic_helper_commit to run all of the commits without adding to the workqueue and running the OS, the issue seems to have disappeared. <<<<
Would not forcing all commits to run directly, without placing them on the workqueue, be "async disabling"? That's what I /thought/ he was referencing.
Oh, I thought he was referring to a different patch. Kees, could I get your confirmation on this?
The change I made actually affected all of the DRM code, although this could easily be changed to be specific to amdgpu. (By forcing blocking on amdgpu_dm's non-blocking commit code)
That said, I'd still need to test further because I only did test it for a couple of hours then. Although it should work in theory.
OTOH your base/context swap idea sounds like a possibly "less disturbance" workaround, if it works, and given the point in the commit cycle... (But if it's out Sunday it's likely too late to test and get it in now anyway; if it's another week, tho...)
The base/context swap idea should make the use-after-free behave how it did in 5.6. Since the bug doesn't cause an issue in 5.6, it's less of a "less disturbance" workaround and more of a "no disturbance" workaround.
Sorry for bothering, but is there now a solution, besides reverting the commits, to avoid freezes/crashes *without* performance regressions?
Kind regards,
Paul
Mazin's "drm/amd/display: Clear dm_state for fast updates" change accomplishes this, at least as a temporary hack.
I've started work on a more large scale fix that we could get in in after.
Regards, Nicholas Kazlauskas
On Tue, Jul 28, 2020 at 01:07:13PM -0400, Kazlauskas, Nicholas wrote:
On 2020-07-28 5:22 a.m., Paul Menzel wrote:
Dear Linux folks,
Am 25.07.20 um 07:20 schrieb Mazin Rezk:
On Saturday, July 25, 2020 12:59 AM, Duncan wrote:
On Sat, 25 Jul 2020 03:03:52 +0000 Mazin Rezk wrote:
Am 24.07.20 um 19:33 schrieb Kees Cook:
> There was a fix to disable the async path for this driver that > worked around the bug too, yes? That seems like a safer and more > focused change that doesn't revert the SLUB defense for all > users, and would actually provide a complete, I think, workaround
That said, I haven't seen the async disabling patch. If you could link to it, I'd be glad to test it out and perhaps we can use that instead.
I'm confused. Not to put words in Kees' mouth; /I/ am confused (which admittedly could well be just because I make no claims to be a coder and am simply reading the bug and thread, but I'd appreciate some "unconfusing" anyway).
My interpretation of the "async disabling" reference was that it was to comment #30 on the bug:
https://bugzilla.kernel.org/show_bug.cgi?id=207383#c30
... which (if I'm not confused on this point too) appears to be yours. There it was stated...
I've also found that this bug exclusively occurs when commit_work is on the workqueue. After forcing drm_atomic_helper_commit to run all of the commits without adding to the workqueue and running the OS, the issue seems to have disappeared. <<<<
Would not forcing all commits to run directly, without placing them on the workqueue, be "async disabling"? That's what I /thought/ he was referencing.
Oh, I thought he was referring to a different patch. Kees, could I get your confirmation on this?
The change I made actually affected all of the DRM code, although this could easily be changed to be specific to amdgpu. (By forcing blocking on amdgpu_dm's non-blocking commit code)
That said, I'd still need to test further because I only did test it for a couple of hours then. Although it should work in theory.
OTOH your base/context swap idea sounds like a possibly "less disturbance" workaround, if it works, and given the point in the commit cycle... (But if it's out Sunday it's likely too late to test and get it in now anyway; if it's another week, tho...)
The base/context swap idea should make the use-after-free behave how it did in 5.6. Since the bug doesn't cause an issue in 5.6, it's less of a "less disturbance" workaround and more of a "no disturbance" workaround.
Sorry for bothering, but is there now a solution, besides reverting the commits, to avoid freezes/crashes *without* performance regressions?
Kind regards,
Paul
Mazin's "drm/amd/display: Clear dm_state for fast updates" change accomplishes this, at least as a temporary hack.
Yeah I gets it's horrible, but better than nothing. Reverting the old amdgpu change to a private state object is probably a lot more invasive.
I've started work on a more large scale fix that we could get in in after.
Does that include a fix for the "stuff needed by irq handler"? Either way pls cc dri-devel, I think this is something worth of a bit wider discussion. Feels like unsolved homework from the entire "make DC integrate into linux" saga ... -Daniel
dri-devel@lists.freedesktop.org
-
Christian König -
daniel＠ffwll.ch -
Duncan -
Kazlauskas, Nicholas -
Kees Cook -
Mazin Rezk -
Paul Menzel