On 16.10.2024 8.47, Michał Pecio wrote:
With some experimentation I found that the bug is a variant of the
old "stop after restart" issue - the doorbell ring is internally
reordered after the subsequent command. By busy-waiting I confirmed
that EP state which is initially seen as Stopped becomes Running
some time later.
Seems host controllers aren't designed to stop, move dequeue, and
restart an endpoint in quick succession.
As it was you who added the Running case handling, do you know hardware
other than NEC which triggers this? Or could it be just a single vendor
who screwed up once 15 years ago and caused all the chaos?
NEC sometimes triggers the Running case too and it is obvious why. I'm
not sure how I missed it back in January and assumed it's some sort of
random failure for no reason.
BTW, the NEC problem appears to be limited to periodic endpoints. I am
unable to reproduce it on bulk. I thought that I reproduced it on bulk
back then, but on second thought it may have been interrupt, which that
device also has. Unfortunatel I wasn't printing endpoint numbers then.
Regards,
Michal
Sorry about the reply delay.
I don't think this is a NEC only issue.
I was originally fixing halted endpoints at stop endpoint command completion,
did some stress testing, and was able to hit that running case on Intel
xHC controllers
See:
9ebf30007858 xhci: Fix halted endpoint at stop endpoint command completion
1174d44906d5 xhci: handle stop endpoint command completion with endpoint in running state.
I also just got a report off-list about an exactly similar case as yours, endpoint
stopped with ctx error, endpoint state was still stopped even if doorbell was
already rung.
This caused Set TR Deq command to fail with context error as endpoint was running
by the time this command was processed.
This was on a Intel host, se we need a generic solution to this.
Thanks
-Mathias
