[Nut-upsdev] A lot of EAGAIN (Resource temporarily unavailable) using nutdrv_qx

[Charles Lepple]

On May 26, 2016, at 1:57 PM, Antonio Pérez <aperez at skarcha.com> wrote:
> 
> I downloaded the code and patched and compiled myself. Using some
> "sleeps" it goes better, but not so fine I would like:
> 
> %<------------%<------------%<------------%<------------
> write(2, "send: QS\n", 9)               = 9
> nanosleep({0, 100000000}, NULL)         = 0
> ioctl(4, USBDEVFS_SUBMITURB, 0x7ffc0d2c66e0) = 0
> ioctl(4, USBDEVFS_REAPURBNDELAY, 0x7ffc0d2c66a8) = -1 EAGAIN (Resource
> temporarily unavailable)
> select(5, NULL, [4], NULL, {0, 1000})   = 1 (out [4], left {0, 564})

How would you improve it?

There will definitely be a tradeoff between responsiveness and CPU usage.

I CC'd the author of nutdrv_qx - I think it could work to add a "usleep" corresponding to the approximate size of the sent and expected receive packet, but this would be different for 1.5 Mb/s and 12 Mb/s USB devices.

You might also want to add timing numbers to the strace output (either "-tt" or "-T").

> Is this the right behaviour? If not, could it goes better using
> libusb-1.0? Are the port to libusb-1.0 in the roadmap?

So far, there have not been any good arguments for libusb-1.0 that would offset the amount of testing needed. Feel free to experiment, though. For timing purposes, it should be sufficient to write a small loop that sends the command for your UPS, and reads the response via libusb-1.0 (rather than trying to port the entire NUT driver).

-- 
Charles Lepple
clepple at gmail