On Thu, 2012-05-10 at 14:08 -0500, Robert Jennings wrote:

* Benjamin Herrenschmidt (benh@kernel.crashing.org) wrote:

quoted

Is this meant to be called in atomic context ? If not, maybe it should
at the very least do a cond_resched() ?

Else, what about ceding the processor ? Or at the very least reducing
the thread priority for a bit ?

Shouldn't we also enforce to always have a timeout ? IE. Something like
30s or so if nothing specified to avoid having the kernel just hard
lock...

In general I don't like that sort of synchronous code, I'd rather return
the busy status up the chain which gives a chance to the caller to take
more appropriate measures depending on what it's doing, but that really
depends what you use that synchronous call for. I suppose if it's for
configuration type operations, it's ok...

This function is called in atomic context, it is used by PFO-type device
drivers to perform operations with the nest accelerator unit (like
crypto acceleration).

Having the timeout and retries in this function is the wrong thing to do.
We'll resubmit this without the loop and the caller will be responsible for
retrying the operations.

I would rather have the caller cede the processor or alter thread
priority where appropriate than doing that in this function.  I don't
think this should be done in this crypto driver.

That sounds right indeed... as long as the upper crypto layer has a
concept of "try again later"... if it doesn't it will result in random
funny failures :-)

Cheers,
Ben.