drm/amd/display: Fix writeback completion timing
Subsystems (3)
PATCH 1/1 drm/amd/display: Fix writeback completion timing
[WHY]
The out fence was signalled on the first vblank after arming, before the DMA finished copying, and the old code worked around this with an mdelay() in the IRQ handler.
[HOW]
Hold a vblank reference while writeback is pending and signal the out fence on the second vblank instead of using mdelay(). Add amdgpu_dm_crtc_complete_writeback() to finish and clean up writeback from both the IRQ and teardown paths.
This can be verified by running IGT's kms_writeback 20 times without timeout errors.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h | 1 + .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 2 ++ .../drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c | 33 ++++++++++--------- 3 files changed, 21 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h +1 -0
@@ -509,6 +509,7 @@ struct amdgpu_crtc {struct drm_pending_vblank_event *event;bool wb_pending; +bool wb_frame_done;bool wb_enabled;struct drm_writeback_connector *wb_conn; };
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +2 -0
@@ -4521,6 +4521,7 @@ bool amdgpu_dm_crtc_complete_writeback(struct amdgpu_crtc *acrtc)spin_lock_irqsave (&acrtc->wb_conn->job_lock, flags); pending = acrtc->wb_pending; acrtc->wb_pending = false; + acrtc->wb_frame_done = false;spin_unlock_irqrestore (&acrtc->wb_conn->job_lock, flags);if (!pending)
@@ -4988,6 +4989,7 @@ static void dm_set_writeback(struct amdgpu_display_manager *dm,* cannot run its matching vblank_put before this get. */ WARN_ON (drm_crtc_vblank_get (&acrtc->base)); + acrtc->wb_frame_done = false; acrtc->wb_pending = true; }
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c +18 -15
@@ -1974,23 +1974,26 @@ static void dm_crtc_high_irq(void *interrupt_params)return ;if (acrtc->wb_conn && acrtc->wb_pending) { -struct dc_stream_state *stream = acrtc->dm_irq_params.stream; -unsigned int v_total, refresh_hz; - - v_total = stream->adjust.v_total_max ? - stream->adjust.v_total_max : stream->timing.v_total; - refresh_hz =div_u64 ((uint64_t ) stream->timing.pix_clk_100hz * -100LL , (v_total * stream->timing.h_total)); -mdelay (1000 / refresh_hz); - -/* -* Completion (signalling the out fence and releasing the vblank -* reference taken in dm_set_writeback()) is handled by the shared -* helper, which is also used by the teardown path. -*/ -if (amdgpu_dm_crtc_complete_writeback (acrtc)) +if (acrtc->wb_frame_done) { +/* +* Second vblank: the DMA for the captured frame has +* had a full frame period to flush to memory. Signal +* the out fence now. +*/ +amdgpu_dm_crtc_complete_writeback (acrtc); + }else { +/* +* First vblank after arming: the frame has been +* scanned out and the DMA is finishing. Disable +* writeback immediately to prevent the hardware from +* starting a new capture that would overwrite the +* buffer. Signal completion on the next vblank to +* ensure the DMA is fully flushed to memory. +*/ dc_stream_fc_disable_writeback (adev->dm.dc, acrtc->dm_irq_params.stream,0 ); + acrtc->wb_frame_done = true; + } } vrr_active =amdgpu_dm_crtc_vrr_active_irq (acrtc);
--
2.43.0
[email protected]
AI reviewThank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
New issues:
- [High] Lockless state machine corruption on overlapping writeback jobs stalls fences and leaks vblank references.
Pre-existing issues:
- [High] Unconditional memory leak of dc_writeback_info structure in dm_set_writeback. --
--- Patch [1]: [PATCH] drm/amd/display: Fix writeback completion timing --- commit d7178e82119f1d066c5082ca676728119bef3a13
Author: Alex Hung <[email protected]>
drm/amd/display: Fix writeback completion timing
Hold a vblank reference while writeback is pending and signal the out fence on the second vblank instead of using mdelay(). Add amdgpu_dm_crtc_complete_writeback() to finish and clean up writeback from both the IRQ and teardown paths.
Quoted 4 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +0
-0
[ ... ]
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c static void dm_set_writeback(struct amdgpu_display_manager *dm, · line 4993
Quoted 6 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +1
-0@@ -4988,6 +4989,7 @@ static void dm_set_writeback(struct amdgpu_display_manager *dm,* cannot run its matching vblank_put before this get. */ WARN_ON (drm_crtc_vblank_get (&acrtc->base)); + acrtc->wb_frame_done = false; acrtc->wb_pending = true;
[Severity: High]
Does this lockless state update cause issues if userspace submits another writeback job immediately?
Since page flip events can signal on the first vblank, userspace could submit a new atomic commit before the second vblank occurs. If a new job is submitted in this window, dm_set_writeback() will overwrite acrtc->wb_frame_done = false and acrtc->wb_pending = true without holding a lock to protect the active job's state.
[Severity: High]
This is a pre-existing issue, but does this function leak the memory allocated for wb_info?
Earlier in dm_set_writeback(), wb_info is allocated:
wb_info = kzalloc(sizeof(*wb_info), GFP_KERNEL);
Then it is passed to dc_stream_add_writeback():
dc_stream_add_writeback(dm->dc, crtc_state->stream, wb_info);
Since dc_stream_add_writeback() copies the structure by value into the stream's array and does not take ownership of the pointer, it seems the originally allocated wb_info pointer is never freed before returning. Could this lead to continuous kernel memory leakage when userspace streams writeback?
Quoted 16 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +0
-0}drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c +0
-6@@ -1974,23 +1974,26 @@ static void dm_crtc_high_irq(void *interrupt_params)return ;if (acrtc->wb_conn && acrtc->wb_pending) { -struct dc_stream_state *stream = acrtc->dm_irq_params.stream; -unsigned int v_total, refresh_hz; - - v_total = stream->adjust.v_total_max ? - stream->adjust.v_total_max : stream->timing.v_total; - refresh_hz = div_u64((uint64_t) stream->timing.pix_clk_100hz *
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c static void dm_crtc_high_irq(void *interrupt_params) · line 1977
Quoted 10 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c +1
-9- 100LL, (v_total * stream->timing.h_total)); -mdelay (1000 / refresh_hz); - -/* -* Completion (signalling the out fence and releasing the vblank -* reference taken in dm_set_writeback()) is handled by the shared -* helper, which is also used by the teardown path. -*/ -if (amdgpu_dm_crtc_complete_writeback (acrtc)) +if (acrtc->wb_frame_done) {
[Severity: High]
If the state was overwritten by a new commit as mentioned above, will this read wb_frame_done == false on the second vblank?
If so, it seems it would take the else branch, set wb_frame_done = true again, and completely skip calling amdgpu_dm_crtc_complete_writeback() for the original job, potentially stalling the out-fence permanently and leaking a vblank reference.
Quoted 11 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c +11
-0+/* +* Second vblank: the DMA for the captured frame has +* had a full frame period to flush to memory. Signal +* the out fence now. +*/ +amdgpu_dm_crtc_complete_writeback (acrtc); + }else { +/* +* First vblank after arming: the frame has been +* scanned out and the DMA is finishing. Disable +* writeback immediately to prevent the hardware from
[ ... ]
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Quoted 11 lines · Alex Hung
[WHY]
The out fence was signalled on the first vblank after arming, before the DMA finished copying, and the old code worked around this with anmdelay()in the IRQ handler.[HOW]
Hold a vblank reference while writeback is pending and signal the out fence on the second vblank instead of usingmdelay(). Addamdgpu_dm_crtc_complete_writeback()to finish and clean up writeback from both the IRQ and teardown paths.
Quoted 5 lines
This can be verified by running IGT's
kms_writeback20 times without timeout errors.Assisted-by Copilot:Claude-Opus-4.8Signed-off-by Alex Hung <[email protected]>
Quoted 87 lines
--- drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h | 1 + .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 2 ++ .../drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c | 33 ++++++++++--------- 3 files changed, 21 insertions(+), 15 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h +1
-0@@ -509,6 +509,7 @@ struct amdgpu_crtc {struct drm_pending_vblank_event *event;bool wb_pending; +bool wb_frame_done;bool wb_enabled;struct drm_writeback_connector *wb_conn; };drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +2
-0@@ -4521,6 +4521,7 @@ bool amdgpu_dm_crtc_complete_writeback(struct amdgpu_crtc *acrtc)spin_lock_irqsave (&acrtc->wb_conn->job_lock, flags); pending = acrtc->wb_pending; acrtc->wb_pending = false; + acrtc->wb_frame_done = false;spin_unlock_irqrestore (&acrtc->wb_conn->job_lock, flags);if (!pending)@@ -4988,6 +4989,7 @@ static void dm_set_writeback(struct amdgpu_display_manager *dm,* cannot run its matching vblank_put before this get. */ WARN_ON (drm_crtc_vblank_get (&acrtc->base)); + acrtc->wb_frame_done = false; acrtc->wb_pending = true; }drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c +18
-16@@ -1974,23 +1974,26 @@ static void dm_crtc_high_irq(void *interrupt_params)return ;if (acrtc->wb_conn && acrtc->wb_pending) { -struct dc_stream_state *stream = acrtc->dm_irq_params.stream; -unsigned int v_total, refresh_hz; - - v_total = stream->adjust.v_total_max ? - stream->adjust.v_total_max : stream->timing.v_total; - refresh_hz =div_u64 ((uint64_t ) stream->timing.pix_clk_100hz * -100LL , (v_total * stream->timing.h_total)); -mdelay (1000 / refresh_hz); - -/* -* Completion (signalling the out fence and releasing the vblank -* reference taken in dm_set_writeback()) is handled by the shared -* helper, which is also used by the teardown path. -*/ -if (amdgpu_dm_crtc_complete_writeback (acrtc)) +if (acrtc->wb_frame_done) { +/* +* Second vblank: the DMA for the captured frame has +* had a full frame period to flush to memory. Signal +* the out fence now. +*/ +amdgpu_dm_crtc_complete_writeback (acrtc); + }else { +/* +* First vblank after arming: the frame has been +* scanned out and the DMA is finishing. Disable +* writeback immediately to prevent the hardware from +* starting a new capture that would overwrite the +* buffer. Signal completion on the next vblank to +* ensure the DMA is fully flushed to memory. +*/ dc_stream_fc_disable_writeback (adev->dm.dc, acrtc->dm_irq_params.stream,0 ); + acrtc->wb_frame_done = true; + } } vrr_active =amdgpu_dm_crtc_vrr_active_irq (acrtc); --2.43.0
Quoted 11 lines · Alex Hung
[WHY]
The out fence was signalled on the first vblank after arming, before the DMA finished copying, and the old code worked around this with anmdelay()in the IRQ handler.[HOW]
Hold a vblank reference while writeback is pending and signal the out fence on the second vblank instead of usingmdelay(). Addamdgpu_dm_crtc_complete_writeback()to finish and clean up writeback from both the IRQ and teardown paths.
Quoted 5 lines
This can be verified by running IGT's
kms_writeback20 times without timeout errors.Assisted-by Copilot:Claude-Opus-4.8Signed-off-by Alex Hung <[email protected]>
Harry
Quoted 87 lines
--- drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h | 1 + .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 2 ++ .../drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c | 33 ++++++++++--------- 3 files changed, 21 insertions(+), 15 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h +1
-0@@ -509,6 +509,7 @@ struct amdgpu_crtc {struct drm_pending_vblank_event *event;bool wb_pending; +bool wb_frame_done;bool wb_enabled;struct drm_writeback_connector *wb_conn; };drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +2
-0@@ -4521,6 +4521,7 @@ bool amdgpu_dm_crtc_complete_writeback(struct amdgpu_crtc *acrtc)spin_lock_irqsave (&acrtc->wb_conn->job_lock, flags); pending = acrtc->wb_pending; acrtc->wb_pending = false; + acrtc->wb_frame_done = false;spin_unlock_irqrestore (&acrtc->wb_conn->job_lock, flags);if (!pending)@@ -4988,6 +4989,7 @@ static void dm_set_writeback(struct amdgpu_display_manager *dm,* cannot run its matching vblank_put before this get. */ WARN_ON (drm_crtc_vblank_get (&acrtc->base)); + acrtc->wb_frame_done = false; acrtc->wb_pending = true; }drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c +18
-16@@ -1974,23 +1974,26 @@ static void dm_crtc_high_irq(void *interrupt_params)return ;if (acrtc->wb_conn && acrtc->wb_pending) { -struct dc_stream_state *stream = acrtc->dm_irq_params.stream; -unsigned int v_total, refresh_hz; - - v_total = stream->adjust.v_total_max ? - stream->adjust.v_total_max : stream->timing.v_total; - refresh_hz =div_u64 ((uint64_t ) stream->timing.pix_clk_100hz * -100LL , (v_total * stream->timing.h_total)); -mdelay (1000 / refresh_hz); - -/* -* Completion (signalling the out fence and releasing the vblank -* reference taken in dm_set_writeback()) is handled by the shared -* helper, which is also used by the teardown path. -*/ -if (amdgpu_dm_crtc_complete_writeback (acrtc)) +if (acrtc->wb_frame_done) { +/* +* Second vblank: the DMA for the captured frame has +* had a full frame period to flush to memory. Signal +* the out fence now. +*/ +amdgpu_dm_crtc_complete_writeback (acrtc); + }else { +/* +* First vblank after arming: the frame has been +* scanned out and the DMA is finishing. Disable +* writeback immediately to prevent the hardware from +* starting a new capture that would overwrite the +* buffer. Signal completion on the next vblank to +* ensure the DMA is fully flushed to memory. +*/ dc_stream_fc_disable_writeback (adev->dm.dc, acrtc->dm_irq_params.stream,0 ); + acrtc->wb_frame_done = true; + } } vrr_active =amdgpu_dm_crtc_vrr_active_irq (acrtc); --2.43.0
Christian König
maintainerQuoted 4 lines · Harry Wentland
Quoted 3 lines · Alex Hung
[WHY]
The out fence was signalled on the first vblank after arming, before the DMA finished copying,
Uff, that's a clear violation of dma_fence rules and can easily be used to cause massive security problems by writing back to freed up memory.
Quoted 2 lines
Quoted 2 lines
and the old code worked around this with an
mdelay()in the IRQ handler.
And that is a really ugly workaround to the problem. Not sure if that is 100% reliable.
Quoted 7 lines
Quoted 7 lines
[HOW]
Hold a vblank reference while writeback is pending and signal the out fence on the second vblank instead of usingmdelay(). Addamdgpu_dm_crtc_complete_writeback()to finish and clean up writeback from both the IRQ and teardown paths.
Quoted 7 lines
Quoted 5 lines
This can be verified by running IGT's
kms_writeback20 times without timeout errors.Assisted-by Copilot:Claude-Opus-4.8Signed-off-by Alex Hung <[email protected]>Reviewed-by Harry Wentland <[email protected]>
Can we backport this patch to stable kernels?
Regards,
Christian.
Quoted 91 lines
Harry
Quoted 87 lines
--- drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h | 1 + .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 2 ++ .../drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c | 33 ++++++++++--------- 3 files changed, 21 insertions(+), 15 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h +1
-0@@ -509,6 +509,7 @@ struct amdgpu_crtc {struct drm_pending_vblank_event *event;bool wb_pending; +bool wb_frame_done;bool wb_enabled;struct drm_writeback_connector *wb_conn; };drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +2
-0@@ -4521,6 +4521,7 @@ bool amdgpu_dm_crtc_complete_writeback(struct amdgpu_crtc *acrtc)spin_lock_irqsave (&acrtc->wb_conn->job_lock, flags); pending = acrtc->wb_pending; acrtc->wb_pending = false; + acrtc->wb_frame_done = false;spin_unlock_irqrestore (&acrtc->wb_conn->job_lock, flags);if (!pending)@@ -4988,6 +4989,7 @@ static void dm_set_writeback(struct amdgpu_display_manager *dm,* cannot run its matching vblank_put before this get. */ WARN_ON (drm_crtc_vblank_get (&acrtc->base)); + acrtc->wb_frame_done = false; acrtc->wb_pending = true; }drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c +18
-16@@ -1974,23 +1974,26 @@ static void dm_crtc_high_irq(void *interrupt_params)return ;if (acrtc->wb_conn && acrtc->wb_pending) { -struct dc_stream_state *stream = acrtc->dm_irq_params.stream; -unsigned int v_total, refresh_hz; - - v_total = stream->adjust.v_total_max ? - stream->adjust.v_total_max : stream->timing.v_total; - refresh_hz =div_u64 ((uint64_t ) stream->timing.pix_clk_100hz * -100LL , (v_total * stream->timing.h_total)); -mdelay (1000 / refresh_hz); - -/* -* Completion (signalling the out fence and releasing the vblank -* reference taken in dm_set_writeback()) is handled by the shared -* helper, which is also used by the teardown path. -*/ -if (amdgpu_dm_crtc_complete_writeback (acrtc)) +if (acrtc->wb_frame_done) { +/* +* Second vblank: the DMA for the captured frame has +* had a full frame period to flush to memory. Signal +* the out fence now. +*/ +amdgpu_dm_crtc_complete_writeback (acrtc); + }else { +/* +* First vblank after arming: the frame has been +* scanned out and the DMA is finishing. Disable +* writeback immediately to prevent the hardware from +* starting a new capture that would overwrite the +* buffer. Signal completion on the next vblank to +* ensure the DMA is fully flushed to memory. +*/ dc_stream_fc_disable_writeback (adev->dm.dc, acrtc->dm_irq_params.stream,0 ); + acrtc->wb_frame_done = true; + } } vrr_active =amdgpu_dm_crtc_vrr_active_irq (acrtc); --2.43.0