(I am going to try responding to this tomorrow btw. I haven't been super busy this week, but this has been a surprisingly difficult email to respond to because I need to actually need to do a deep dive some of the MST helpers tomorrow to figure out more of the specifics on why I realized we couldn't just hot add/remove port->connector here).
On Wed, 2021-08-25 at 03:35 +0000, Lin, Wayne wrote:
[Public]
-----Original Message----- From: Lyude Paul lyude@redhat.com Sent: Tuesday, August 24, 2021 5:18 AM To: Lin, Wayne Wayne.Lin@amd.com; dri-devel@lists.freedesktop.org Cc: Kazlauskas, Nicholas Nicholas.Kazlauskas@amd.com; Wentland, Harry Harry.Wentland@amd.com; Zuo, Jerry Jerry.Zuo@amd.com; Wu, Hersen hersenxs.wu@amd.com; Juston Li juston.li@intel.com; Imre Deak imre.deak@intel.com; Ville Syrjälä ville.syrjala@linux.intel.com; Daniel Vetter daniel.vetter@ffwll.ch; Sean Paul sean@poorly.run; Maarten Lankhorst maarten.lankhorst@linux.intel.com; Maxime Ripard mripard@kernel.org; Thomas Zimmermann tzimmermann@suse.de; David Airlie airlied@linux.ie; Daniel Vetter daniel@ffwll.ch; Deucher, Alexander Alexander.Deucher@amd.com; Siqueira, Rodrigo Rodrigo.Siqueira@amd.com; Pillai, Aurabindo Aurabindo.Pillai@amd.com; Bas Nieuwenhuizen bas@basnieuwenhuizen.nl; Cornij, Nikola Nikola.Cornij@amd.com; Jani Nikula jani.nikula@intel.com; Manasi Navare manasi.d.navare@intel.com; Ankit Nautiyal ankit.k.nautiyal@intel.com; José Roberto de Souza jose.souza@intel.com; Sean Paul seanpaul@chromium.org; Ben Skeggs bskeggs@redhat.com; stable@vger.kernel.org Subject: Re: [PATCH 2/4] drm/dp_mst: Only create connector for connected end device
[snip]
I think I might still be misunderstanding something, some comments below
On Mon, 2021-08-23 at 06:33 +0000, Lin, Wayne wrote:
Hi Lyude,
Really thankful for willing to explain in such details. Really appreciate.
I'm trying to fix some problems that observed after these 2 patches
- 09b974e8983 drm/amd/amdgpu_dm/mst: Remove ->destroy_connector()
callback
- 72dc0f51591 drm/dp_mst: Remove
drm_dp_mst_topology_cbs.destroy_connector
With above patches, we now change to remove dc_sink when connector is about to be destroyed. However, we found out that connectors won't get destroyed after hotplugs. Thus, after few times hotplugs, we won't create any new dc_sink since number of sink is exceeding our limitation. As the result of that, I'm trying to figure out why the refcount of connectors won't get zero.
Based on my analysis, I found out that if we connect a sst monitor to a mst hub then connect the hub to the system, and then unplug the sst monitor from the hub. E.g. src - mst hub - sst monitor => src - mst hub (unplug) sst monitor
Within this case, we won't try to put refcount of the sst monitor. Which is what I tried to resolve by [PATCH 3/4]. But here comes a problem which is confusing me that if I can destroy connector in this case. By comparing to another case, if now mst hub is connected with a mst monitor like this: src - mst hub - mst monitor => src - mst hub (unplug) mst monitor
We will put the topology refcount of mst monitor's branching unit in and drm_dp_port_set_pdt() and eventually call drm_dp_delayed_destroy_port() to unregister the connector of the logical port. So following the same rule, I think to dynamically unregister a mst connector is what we want and should be reasonable to also destroy sst connectors in my case. But this conflicts the idea what we have here. We want to create connectors for all output ports. So if dynamically creating/destroying connectors is what we want, when is the appropriate time for us to create one is what I'm considering.
Take the StartTech hub DP 1to4 DP output ports for instance. This hub, internally, is constructed by 3 1-to-2 mst branch chips. 2 output ports of 1st chip are hardwired to another 2 chips. It's how it makes it to support 1-to-4 mst branching. So within this case, the internal 2 output ports of 1st chip is not connecting to a stream sink and will never get connected to one. Thus, I'm thinking maybe the best timing to attach a connector to a port is when the port is connected, and the connected PDT is determined as a stream sink.
Sorry if I misunderstand anything here and really thanks for your time to shed light on this : ) Thanks Lyude.
It's no problem, it is my job after all! Sorry for how long my responses have been taking, but my plate seems to be finally clearing up for the foreseeable future.
That being said - it sounds like with this we still aren't actually clear on where the topology refcount leak is happening - only when it's happening, which says to me that's the issue we really need to be figuring out the cause of as opposed to trying to workaround it.
Actually - refcount leaks is an issue I've ran into a number of times before in the past, so a while back I actually added some nice debugging features to assist with debugging leaks. If you enable the following options in your kernel config:
CONFIG_EXPERT=y # This must be set first before the next option CONFIG_DRM_DEBUG_DP_MST_TOPOLOGY_REFS=y
Unfortunately, I'm suddenly realizing after typing this that apparently I never bothered adding a way for us to debug the
refcounts of ports/mstbs that haven't been released yet - only the ones for ones that have. This shouldn't be difficult at all for me to add, so I'll send you a patch either today or at the start of next week to try debugging with using this, and then we can figure out where this leak is really coming from.
Thanks Lyude!
Sorry to bother you, but I would like to clarify this again. So it sounds
It's no problem! It's my job and I'm happy to help :).
Thanks! I would like to learn more from you as below : p
like you also agree that we should destroy associated connector
Not quite. I think a better way of explaining this might be to point out that the lifetime of an MST port and its connector isn't supposed to be determined by whether or not it has something plugged into it - its lifetime is supposed to depend on whether there's a valid path from us down the MST topology to the port we're trying to reach. So an MSTB with ports that is unplugged would destroy all of its ports - but an unplugged port should just be the same as a disconnected DRM connector - even if the port itself is just hosting a branching device.
This is the part a bit difficult to me. I treat DRM connector as the place where we associate with a stream sink. So if the statement is "All DP mst output ports are places we connect with stream sink", I would say false to this since I can find the negative example when output port is connected with mst branch device. Thus, looks like we could only determine whether to create a connector for an output port when the peer device type is known?
Additionally - we don't want to try "delaying" connector creation either. In the modern world hotplugging is almost always reliable in normal situations, but even so there's still use cases for wanting force probing for analog devices on DP converters and just in general as it's a feature commonly used by developers or users working around monitors with problematic HPD issues or EDID issues.
I think I understand that why we want to create connectors for all output ports here. But under these mentioned use cases, aren't we still capable to force connector to enable stream? MST hub with muti-functon capability, it will enumerate connected virtual DP peer device. For problematic HPD issues or EDID issues, their connection status is also connected.
My understanding of output port is it is an internal node to help construct an end-to-end virtual channel between a stream source device and a stream sink device. Creating connectors for internal nodes within a virtual channel is a bit hard for me to get the idea. Please correct me if I misunderstand anything here. Thanks Lyude!
when we unplug sst monitor from a mst hub in the case that I described? In the case I described (unplug sst monitor), we only receive CSN from the hub that notifying us the connection status of one of its downstream output ports is changed to disconnected. There is no topology refcount needed to be decreased on this disconnected port but the malloc refcount. Since the output port is still declared by
Apologies - I misunderstood your original mail as implying that topology refcounts were being leaked - but it sounds like it's actually malloc refcounts being leaked instead? In any case - that means we're still tracing down a leak, just a malloc ref leak.
But, this still doesn't totally make sense to me. Malloc refs only keep the actual drm_dp_mst_port/drm_dp_mst_branch struct alive in memory. Nothing else is kept around, meaning the DRM connector (and I assume by proxy, the dc_sink) should both be getting dropped still and the only thing that should be leaked is a memory allocation. These things should instead be dropped once there's no longer any topology references around. So, are we _sure_ that the problem here is a missing drm_dp_mst_port_put_malloc() or drm_dp_mst_mstb_put_malloc()?
Just my two cents, I don't think it's leak of malloc ref neither. As you said, malloc ref is dealing with the last step to free port/mstb. If there is still topology refcount on port/mstb in my case, we won't free port/mstb.
If we are unfortunately we don't have equivalent tools for malloc() tracing. I'm totally fine with trying to add some if we have trouble figuring out this issue, but I'm a bit suspicious of the commits you mentioned that introduced this problem. If the problem doesn't happen until those two commits, then it's something in the code changes there that are causing this problem.
I think we probably also have the problem before these commits, but we didn't notice this before. Just when we change to clean up all things in dm_dp_mst_connector_destroy(), I start to try to figure out all these things out.
The main thing I'm suspicious of just from looking at changes in 09b974e8983a4b163d4a406b46d50bf869da3073 is that the call to amdgpu_dm_update_freesync_caps() that was previously in dm_dp_destroy_mst_connector() appears to be dropped and not re-added in (oh dear, this is a /very/ confusingly similar function
Lol. I also have hard time on this..
name!!!) dm_dp_mst_connector_destroy(). I don't remember if this was intentional on my part, but does adding a call back to amdgpu_dm_update_freesync_caps() into dm_dp_destroy_mst_connector() right before the dc_link_remove_remote_sink() call fix anything?
As well, I'm far less suspicious of this one but does re-adding this hunk:
aconnector->dc_sink = NULL; aconnector->dc_link->cur_link_settings.lane_count = 0;
After dc_sink_release() fix anything either?
So the main problem is we don't have chance to call dc_link_remove_remote_sink() in the unplugging SST case. We only have chance to remove the remote sink of a link when unplug a mstb.
the mst hub, I think we shouldn't destroy the port. Actually, no ports nor mst branch devices should get destroyed in this case I think. The result of LINK_ADDRESS is still the same before/after removing the sst monitor except the DisplayPort_Device_Plug_Status/ Legacy_Device_Plug_Status.
Hence, if you agree that we should put refcount of the connector of the disconnected port within the unplugging sst monitor case to release the allocated resource, it means we don't want to create connectors for those disconnected ports. Which conflicts current flow to create connectors for all declared output ports.
Thanks again for your time Lyude!
-- Cheers, Lyude Paul (she/her) Software Engineer at Red Hat
-- Regards, Wayne