Skip to contentSkip to conversation

Drive-by MST fixes for amdgpu

PATCH New Lyude Paul 1 participant 4 messages Jun 2, 2022

Subsystems (3)
lore
Share & tools
lore

Lyude Paul

author
#0 lore raw

cover Drive-by MST fixes for amdgpu

Now that I'm finishing up my work to remove the legacy MST code from the tree, I've come across a couple of various issues that I wrote up patches for along the way. These are some of those patches for amdgpu.

Lyude Paul (3):
drm/amdgpu/dm/mst: Stop grabbing mst_mgr->lock in
compute_mst_dsc_configs_for_state()
drm/amdgpu/dm/mst: Stop grabbing mst_mgr->lock in
pre_compute_mst_dsc_configs_for_state()
drm/amdgpu/dm: Drop != NULL check in dm_mst_get_pbn_divider()

 .../drm/amd/display/amdgpu_dm/amdgpu_dm_mst_types.c  | 12 +-----------
 1 file changed, 1 insertion(+), 11 deletions(-)
Signature · 1 line
2.35.3
author
#1 lore raw

PATCH 1/3 drm/amdgpu/dm/mst: Stop grabbing mst_mgr->lock in compute_mst_dsc_configs_for_state()

Noticed this while trying to update amdgpu for the non-atomic MST removal changes - for some reason we appear to grab mst_mgr->lock before computing mst DSC configs. This is wrong though - mst_mgr->lock is only needed while traversing the actual MST topology state - which is not typically something that DRM drivers should be doing themselves anyway.

---
 drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_mst_types.c | 3 ---
 1 file changed, 3 deletions(-)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_mst_types.c +0 -3
index 9221b6690a4a..cb3b0e08acc4 100644
@@ -1056,13 +1056,10 @@ bool compute_mst_dsc_configs_for_state(struct drm_atomic_state *state,
 		if (!is_dsc_need_re_compute(state, dc_state, stream->link))
 			continue;
 
-		mutex_lock(&aconnector->mst_mgr.lock);
 		if (!compute_mst_dsc_configs_for_link(state, dc_state, stream->link,
 			vars, &link_vars_start_index)) {
-			mutex_unlock(&aconnector->mst_mgr.lock);
 			return false;
 		}
-		mutex_unlock(&aconnector->mst_mgr.lock);
 
 		for (j = 0; j < dc_state->stream_count; j++) {
 			if (dc_state->streams[j]->link == stream->link)
Signature · 1 line
2.35.3
author
#2 lore raw

PATCH 2/3 drm/amdgpu/dm/mst: Stop grabbing mst_mgr->lock in pre_compute_mst_dsc_configs_for_state()

In reply to Lyude Paul

This lock is only needed if you're iterating through the in-memory topology (e.g. drm_dp_mst_branch->ports, drm_dp_mst_port->mstb, etc.). This doesn't actually seem to be what's going on here though, so we can just drop this lock.

---
 drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_mst_types.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_mst_types.c +1 -5
index cb3b0e08acc4..1259f2f7a8f9 100644
@@ -1112,16 +1112,12 @@ static bool
 		if (!is_dsc_need_re_compute(state, dc_state, stream->link))
 			continue;
 
-		mutex_lock(&aconnector->mst_mgr.lock);
 		if (!compute_mst_dsc_configs_for_link(state,
 						      dc_state,
 						      stream->link,
 						      vars,
-						      &link_vars_start_index)) {
-			mutex_unlock(&aconnector->mst_mgr.lock);
+						      &link_vars_start_index))
 			return false;
-		}
-		mutex_unlock(&aconnector->mst_mgr.lock);
 
 		for (j = 0; j < dc_state->stream_count; j++) {
 			if (dc_state->streams[j]->link == stream->link)
Signature · 1 line
2.35.3
author
#3 lore raw

PATCH 3/3 drm/amdgpu/dm: Drop != NULL check in dm_mst_get_pbn_divider()

In reply to Lyude Paul

A lot of code in amdgpu seems to sprinkle in

if (foo != NULL)
…

Checks pretty much all over the place, many times in locations where it's clear foo (whatever foo may be) should never be NULL unless we've run into a programming error. This is definitely one of those places, as dm_mst_get_pbn_divider() should never be getting called with a NULL dc_link pointer.

The problem with this code pattern is that many times the places I've seen it used in amdgpu have no real error handling. This is actually quite bad, if we try to avoid the NULL pointer and instead simply skip any code that was expecting a valid pointer - we're already in undefined territory. Subsequent code we execute may have expected sideaffects from the code we skipped that are no longer present, which leads to even more unpredictable behavior then a simple segfault. This could be silent errors or even just another segfault somewhere else.

If we simply segfault though, that's not good either. But unlike the former solution, no subsequent code in the kernel thread will execute - and we will likely even get a clear backtrace from the invalid memory access. Of course, the preferred approach is to simply handle the possibility of both NULL and non-NULL pointers with nice error handling code. However, that's not always desirable or even possible, and in those cases it's likely just better to fail predictably rather than unpredictably.

This code is a nice example of that - if link is NULL, you'll return a PBN divisor of 0. And thus, you've simply traded in your potential segfault for a potential divide by 0 error. This was something I actually managed to hit while working on the legacy MST removal work.

---
 drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_mst_types.c | 3 ---
 1 file changed, 3 deletions(-)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_mst_types.c +0 -3
index 1259f2f7a8f9..35c7def8f2bd 100644
@@ -537,9 +537,6 @@ void amdgpu_dm_initialize_dp_connector(struct amdgpu_display_manager *dm,
 
 int dm_mst_get_pbn_divider(struct dc_link *link)
 {
-	if (!link)
-		return 0;
-
 	return dc_link_bandwidth_kbps(link,
 			dc_link_get_link_cap(link)) / (8 * 1000 * 54);
 }
Signature · 1 line
2.35.3