Re: [PATCH v2 ath-next 2/2] wifi: ath11k: fix HTC rx insufficient length
From: Miaoqing Pan
Date: Tue Mar 11 2025 - 21:12:06 EST
On 3/11/2025 11:20 PM, Jeff Johnson wrote:
On 3/11/2025 1:29 AM, Miaoqing Pan wrote:
On 3/10/2025 6:09 PM, Johan Hovold wrote:
I'm still waiting for feedback from one user that can reproduce the
ring-buffer corruption very easily, but another user mentioned seeing
multiple zero-length descriptor warnings over the weekend when running
with this patch:
ath11k_pci 0006:01:00.0: rxed invalid length (nbytes 0, max 2048)
Are there ever any valid reasons for seeing a zero-length descriptor
(i.e. unrelated to the race at hand)? IIUC the warning would only be
printed when processing such descriptors a second time (i.e. when
is_desc_len0 is set).
That's exactly the logic, only can see the warning in a second time. For
the first time, ath12k_ce_completed_recv_next() returns '-EIO'.
That didn't answer Johan's first question:
Are there ever any valid reasons for seeing a zero-length descriptor?
The events currently observed are all firmware logs. The discarded
packets will not affect normal operation. I will adjust the logging to
debug level.
We have an issue that there is a race condition where hardware updates the
pointer before it has filled all the data. The current solution is just to
read the data a second time. But if that second read also occurs before
hardware has updated the data, then the issue isn't fixed.
Thanks for the addition.
So should there be some forced delay before we read a second time?
Or should we attempt to read more times?
The initial fix was to keep waiting for the data to be ready. The
observed phenomenon is that when the second read fails, subsequent reads
will continue to fail until the firmware's CE2 ring is full and triggers
an assert after timeout. However, this situation is relatively rare, and
in most cases, the second read will succeed. Therefore, adding a delay
or multiple read attempts is not useful.