On 26 January 2016 at 13:04, Felix Fietkau [off-list ref] wrote:

On 2016-01-26 12:56, Michal Kazior wrote:

quoted

On 26 January 2016 at 11:45, Felix Fietkau [off-list ref] wrote:

quoted

On 2016-01-21 14:23, Michal Kazior wrote:

quoted

This will allow drivers to make more educated
decisions whether to defer transmission or not.

Relying on wake_tx_queue() call count implicitly
was not possible because it could be called
without queued frame count actually changing on
software tx aggregation start/stop code paths.

It was also not possible to know how long
byte-wise queue was without dequeueing.

Signed-off-by: Michal Kazior <redacted>

Instead of exposing these in the struct to the driver directly, please
make a function to get them. Since the number of frames is already
tracked in txqi->queue, you can avoid counter duplication that way.

Hmm, so you suggest to have something like:

[...]

quoted

quoted

Also, that way you can fix a race condition between accessing the number
of frames counter and the bytes counter.

I don't see a point in maintaining coherency between the two counters
with regard to each other alone. Do you have a use-case that would
actually make use of that property?

I'd like to avoid any unnecessary spinlocks.

OK. I guess we can leave them out for now. How frequently are you going
to call this function?

Depends, but with new data path in ath10k[1][2] it'll be at least once
for each wake_tx_queue() + once for each txq in each fetch-ind event.


Michał

[1]: https://patchwork.kernel.org/patch/8081301/
[2]: https://patchwork.kernel.org/patch/8081331/