On Fri, Jul 18, 2014 at 10:35:10AM +0200, Sebastian Andrzej Siewior wrote:

On 07/17/2014 06:18 PM, Felipe Balbi wrote:

quoted

quoted

No, this is okay. If you look, it checks for "up->ier & 
UART_IER_THRI". On the second invocation it will see that this
bit is already set and therefore won't call get_sync() for the
second time. That bit is removed in the _stop_tx() path.

oh, right. But that's actually unnecessary. Calling
pm_runtime_get() multiple times will just increment the usage
counter multiple times, which means you can call __stop_tx()
multiple times too and everything gets balanced, right ?

No. start_tx() will be called multiple times but only the first
invocation invoke pm_runtime_get(). Now I noticed that I forgot to

right, but that's unnecessary. You can pm_runtime_get() every time
start_tx() is called. Just make sure to put everytime stop_tx() is
called too.

remove pm_runtime_put_autosuspend() at the bottom of it. But you get
the idea right?
pm_get() on the while the UART_IER_THRI is not yet set. pm_put() once
the fifo is completely empty.

quoted

quoted

Do you have other ideas? It doesn't look like this is exported at
all. If we call _stop_tx() right away, then we have 64 bytes in
the TX fifo in the worst case. They should be gone "soon" but the
HW-flow control may delay it (in theory for a long time)).

this can be problematic, specially for OMAP which can go into OFF
while idle. Whatever is in the FIFO would get lost. It seems like
omap-serial solved this within transmit_chars().

No, it didn't.

quoted

See how transmit_chars() is called from within IRQ handler with
clocks enabled then it conditionally calls serial_omap_stop_tx()
which will pm_runtime_get_sync() -> do_the_harlem_shake() -> 
pm_runtime_put_autosuspend(). That leaves one unbalanced 
pm_runtime_get() which is balanced when we're exitting the IRQ
handler.

omap-serial and the 8250 do the following on tx path:
- start_tx()
  -> sets UART_IER_THRI. This will generate an interrupt once the FIFO
     is empty.
- interrupt, notices the empty fifo, invokes serial8250_start_tx()/
  transmit_chars().
  Both have a while loop that fills the FIFO. This loop is left once
  the tty-buffer is empty (uart_circ_empty() is true) or the FIFO full.

Lets say you filled 64 bytes into the FIFO and then left because your
FIFO is full and tty-buffer is empty. That means you will invoke
serial_omap_stop_tx() and remove UART_IER_THRI bit.
This is okay because you are not interested in further FIFO empty
interrupts because you don't have any TX-bytes to be sent. However,
once you leave the transmit_chars() you leave serial_omap_irq() which
does the last pm_put(). That means you have data in the TX FIFO that is
about to be sent and the device is in auto-suspend.
This is "fine" as long as the timeout is greater then the time required
for the data be sent (plus assuming HW-float control does not stall it
for too long) so nobody notices a thing.

the time is set to -1 by default. I guess this only works because nobody
has ever tested long transfers with slow baud rates :-p

For that reason I added the hack / #if0 block in the 8250 driver. To
ensure we do not disable the TX-FIFO-empty interrupt even if there is
nothing to send. Instead we enter serial8250_tx_chars() once again with
empty FIFO and empty tty-buffer and will invoke _stop_tx() which also
finally does the pm_put().
That is the plan. The problem I have is how to figure out that the
device is using auto-suspend. If I don't then I would have to remove
the #if0 block and that would mean for everybody an extra interrupt
(which I wanted to avoid).

looks like the closest you have is:

if (pm_runtime_autosuspend_expiration(dev) > 0)
	foo();

Another possibility would be to implement the ->runtime_idle() callback
and only return 0 if fifo is empty, otherwise return -EAGAIN ? then, if
the autosuspend timer expires, ->runtime_idle gets called and you can do
the right thing depending on fifo empty or not.

Take a look at
drivers/usb/core/driver.c::usb_runtime_{idle,resume,suspend} for
examples. That seems to work pretty well.

quoted

This seems work fine and dandy without DMA, but for DMA work, I
think we need to make sure this IP stays powered until we get DMA
completion callback. But that's future, I guess.

Yes, probably. That means one get at dma start, one put at dma complete
callback. And I assume we get that callbacks once the DMA transfer is
complete, not when the FIFO is empty :) So lets leave it to the future
for now?

k

-- 
balbi
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: Digital signature
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20140718/6299a783/attachment-0001.sig>