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

[Mauro Carvalho Chehab <mchehab@osg.samsung.com>]

Em Fri, 27 Nov 2015 05:00:26 +0000
Al Viro <viro@ZenIV.linux.org.uk> 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.

The same kind of crap were found (and fixed) on other drivers, like
the fix on this changeset: 45053edc05 ('[media] saa7164: fix poll bugs').

> b) sure, it's nice that if this thing hangs, we'll be able to kill it.
> However, if one's ->poll() can hang indefinitely, it means bad things
> for poll(2), select(2), etc. semantics.  What the hell had been intended
> there?

I guess there was no special intent. It is just a bad driver code that
was replicated from other drivers.

> c) a bunch of v4l2_m2m_poll() callers are also taking some kind of
> mutex; AFAICS, all of those appear bogus (the rest of them do not
> play wiht ERESTARTSYS, just plain mutex_lock() for those).
> 
> What's going on there?