Occasionally there are LINK_ADDRESS sideband messages timing out with the Lenovo MST dock + Dell MST monitor(w/ in-built branch) setup I have. These failures lead to the display not coming up on boot. Power cycling the port corresponding to the MST monitor's branch device and resending the message fixes the issue. I am not entirely sure if this is specific to my setup. However, as the power state is toggled conditionally on LINK_ADDRESS timeouts, this should not affect the working cases.
Cc: Lyude lyude@redhat.com Cc: Dave Airlie airlied@redhat.com Cc: Jani Nikula jani.nikula@intel.com Signed-off-by: Dhinakaran Pandiyan dhinakaran.pandiyan@intel.com --- drivers/gpu/drm/drm_dp_mst_topology.c | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/drm_dp_mst_topology.c b/drivers/gpu/drm/drm_dp_mst_topology.c index 70dcfa58d3c2..e06defcdcf18 100644 --- a/drivers/gpu/drm/drm_dp_mst_topology.c +++ b/drivers/gpu/drm/drm_dp_mst_topology.c @@ -1596,8 +1596,9 @@ static void drm_dp_send_link_address(struct drm_dp_mst_topology_mgr *mgr, int len; struct drm_dp_sideband_msg_tx *txmsg; int ret; + int attempts = 5;
- txmsg = kzalloc(sizeof(*txmsg), GFP_KERNEL); +retry: txmsg = kzalloc(sizeof(*txmsg), GFP_KERNEL); if (!txmsg) return;
@@ -1635,9 +1636,17 @@ static void drm_dp_send_link_address(struct drm_dp_mst_topology_mgr *mgr, } (*mgr->cbs->hotplug)(mgr); } + } else if (attempts--) { + kfree(txmsg); + drm_dp_send_power_updown_phy(mstb->mgr, mstb->port_parent, + false); + drm_dp_send_power_updown_phy(mstb->mgr, mstb->port_parent, + true); + DRM_DEBUG_KMS("link address failed %d, retrying\n", ret); + goto retry; } else { mstb->link_address_sent = false; - DRM_DEBUG_KMS("link address failed %d\n", ret); + DRM_DEBUG_KMS("link address failed %d, giving up\n", ret); }
kfree(txmsg);
On Wed, 20 Dec 2017, Dhinakaran Pandiyan dhinakaran.pandiyan@intel.com wrote:
With stuff like this, I always wonder if catering for a failing setup blocks us from improving working setups, because once we add this, we can't regress it. For example, is there a valid scenario where we'd want to fail fast here, instead of retrying?
Some nits below.
You'll end up doing (attempts + 1) attempts, including the first one.
How about memset(txmsg, 0, sizoef(*txmsg)); here and move the goto label down to avoid repeated allocations?
Maybe do the debug message before you power down/up?
BR, Jani.
On Thu, 2017-12-21 at 08:53 +0200, Jani Nikula wrote:
Link address failure would result in not probing a branch device that already has been detected. I guess the fail fast case will be applicable if the said device was not really present but the parent branch told us otherwise.
Yeah, that is what I intended to do :) I renamed it from 'retry' to 'attempt' at the last moment, which made it a bit confusing I suppose.
I am stressing testing my setup more to see how well this recovery works and update this patch.
Absolutely.
Ok.
On Fri, 2017-12-22 at 00:48 +0000, Pandiyan, Dhinakaran wrote:
The other option is we could check the device ID of the dock and implement this as a quirk.
Btw I found some relevant information in the spec.
"The Message Transaction originator must perform the reply timeout check. If an error to a request causes the system to be in an invalid state (e.g., all nodes failed to delete a virtual channel, it is the responsibility of the Message Transaction originator to return the system to a valid state). The Message Transaction originator is responsible for any retries."
and
"SET_POWER_STATE Must be programmed to 001 (binary) upon power-on reset or an upstream device disconnect. 001 = Set local Sink device and all downstream Sink devices to D0 (normal operation mode)."
I wonder if the dock is missing this step because the monitor seems to work well with a discrete MST hub.
Here's what I got with 266 rounds of reboots. Correct link address at 1st try 180 Retry 1 45 Retry 2 32 Retry 3 0 Retry 4 0 Retry 5 2 Giving up 3 Incorrect link address 4
The retries help about 30% of the cases.
On Wed, Dec 20, 2017 at 10:36:24PM -0800, Dhinakaran Pandiyan wrote:
Does the spec say this should be retried 5 times or is this just an experimental number? We have such magical numbers for retries all over the DP code and that makes debugging harder later, so atleast a comment of why its 5 would help.
This should be --attempts if you want the total attempts to be 5
Manasi
On Thu, 2017-12-21 at 10:52 -0800, Manasi Navare wrote:
The spec. does not say how many times to retry, but it does say the source is responsible for retrying.
We have such magical numbers for retries all over the DP code and that makes debugging harder later, so atleast a comment of why its 5 would help.
Takes about 22 seconds from boot to complete 5 retries on my SKL, I think that's enough trying from the kernel side before pulling out the DP cable makes sense :)
I don't.
On Thu, Dec 21, 2017 at 05:06:22PM -0800, Pandiyan, Dhinakaran wrote:
It still appears to be a magical number so better to comment it properly. Helps in the debug
Yes the variable attempts is misleading in that case. Probably call it "tries"
Manasi
Sorry for the late reply, I've been having very similar issues on my own MST hub and I wanted to make sure that they were the same issue, although it seems like they aren't.
So; I've been doing a lot of MST debugging this week and last and something I've discovered needs to be taken into account sometimes with MST hubs is the actual state that they're in at the point that the DRM driver detects them. I've managed to on multiple occasions, get my hub into a weird state by:
- Plugging in MST displays into the hub - Turning on the machine - Unplugging MST displays from the hub (while still in the BIOS) - Booting into linux - Plugging MST displays into the hub - Everything times out, the world explodes, the economy collapses, etc.
I think maybe, especially since this should be perfectly valid behavior and not break well or poor behaving hubs, we should do a power cycle with the display like this when the DP port initially detects an MST hub and before we start doing any serious communication with it. Could you see if this fixes your issue instead of the patch you've got here?
As well, mind attaching your full dmesg with drm.debug=0x6?
On Wed, 2017-12-20 at 22:36 -0800, Dhinakaran Pandiyan wrote:
On Thu, 2018-01-04 at 18:21 -0500, Lyude Paul wrote:
Well, my reasoning to power cycle after a failure was to not affect hubs that work. Besides that I also saw an odd cycle of timeouts and late replies when I did this.
1. Detect hub 2. Toggle power state and send LINK_ADDRESS req. 3. LINK_ADDRESS req times out. 4. Toggle power state and send LINK_ADDRESS req. 5. Get late response for the first (and expired) LINK_ADDRESS req. 6. Go to step 3
It seems like toggling the power states flushes out some message buffers in the MST hub.
But, in the retry approach, power cycling resulted in getting the response for the current LINK_ADDRESS request along with the previous one that timed out.
I could not come up with an explanation for all the behavior. So, I decided to get back to this later.
As well, mind attaching your full dmesg with drm.debug=0x6?
I don't have the old logs anymore, will try to get something for you.
-DK
On Thu, 2018-01-04 at 23:46 +0000, Pandiyan, Dhinakaran wrote:
Here you go
Setup: Laptop -> ThinkPad dock -> Dell MST monitor -> Dell monitor Hotplugged display to dock[passed] - https://pastebin.com/CemRGsCb Connected boot[failed] - https://pastebin.com/jjXP6HzB
dri-devel@lists.freedesktop.org
-
Dhinakaran Pandiyan -
Jani Nikula -
Lyude Paul -
Manasi Navare -
Pandiyan, Dhinakaran