Re: ->poll() instances shouldn't be indefinitely blocking

From: Linus Torvalds
Date: Fri Nov 27 2015 - 12:49:30 EST


On Fri, Nov 27, 2015 at 7:18 AM, Mauro Carvalho Chehab
<mchehab@xxxxxxxxxxxxxxx> wrote:
> Al Viro <viro@xxxxxxxxxxxxxxxxxx> escreveu:
>
>> Take a look at this:
>> static unsigned int gsc_m2m_poll(struct file *file,
>> struct poll_table_struct *wait)
>> {
>> struct gsc_ctx *ctx = fh_to_ctx(file->private_data);
>> struct gsc_dev *gsc = ctx->gsc_dev;
>> int ret;
>>
>> if (mutex_lock_interruptible(&gsc->lock))
>> return -ERESTARTSYS;
>>
>> ret = v4l2_m2m_poll(file, ctx->m2m_ctx, wait);
>> mutex_unlock(&gsc->lock);
>>
>> return ret;
>> }
>>
>> a) ->poll() should not return -E...; callers expect just a bitmap of
>> POLL... values.
>
> Yeah. We fixed issues like that on other drivers along the time. I guess
> this is a some bad code that people just cut-and-paste from legacy drivers
> without looking into it.

Actually, while returning -ERESTARTSYS is bogus, returning _zero_
would not be. The top-level poll() code will happily notice the
signal, and return -EINTR like poll should (unless something else is
pending, in which case it will return zero and the bits set for that
something else).

So having a driver with a ->poll() function that does that kind of
conditional locking is not wrong per se. It's just he return value
that is crap.

I also do wonder if we might not make the generic code a bit more
robust wrt things like this. The bitmask we use is only about the low
bits, so we *could* certainly allow the driver poll() functions to
return errors - possibly just ignoring them. Or perhaps have a
WARN_ON_OCNE() to find them.

Al, what do you think? The whole "generic code should be robust wrt
drivers making silly mistakes" just sounds like a good idea. Finding
these things through code inspection is all well and good, but having
a nice warning report from users might be even better.

Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/