On 23.05.2022 16:48:32, David Jander wrote:
> Btw, I just discovered that there was indeed another often unnecessary context
> switch happening in spi_sync(). When spi_finalize_current_message() is called,
> it will always queue work, even if we just managed to do everything in the
> calling context:
> 
> https://elixir.bootlin.com/linux/v5.18-rc7/source/drivers/spi/spi.c#L1909
> 
> This is needed to have the SPI worker thread unprepare transfer
> hardware and free the dummy buffers if required. My question is why
> this needs to be done from the thread. Putting the relevant code after
> the call to ctlr->transfer_one_message() in __spi_pump_messages(),
> saves this extra bounce to the worker thread if no more messages are
> queued, saving ca 10-12us extra time between consecutive spi_sync
> messages.

Martin Sperl tried to do a delayed teardown, see:

| 412e60373245 spi: core: avoid waking pump thread from spi_sync instead run teardown delayed

But that turned out be not working properly:

| https://lore.kernel.org/all/f86eaebb-0359-13be-f4a2-4f2b8832252e@nvidia.com/

Marc

-- 
Pengutronix e.K.                 | Marc Kleine-Budde           |
Embedded Linux                   | https://www.pengutronix.de  |
Vertretung West/Dortmund         | Phone: +49-231-2826-924     |
Amtsgericht Hildesheim, HRA 2686 | Fax:   +49-5121-206917-5555 |