Re: [PATCH RFC v3 0/7] epoll: Introduce new syscalls, epoll_ctl_batch and epoll_pwait1

[Fam Zheng]

Hi Seymour,

On Mon, 02/16 07:25, Seymour, Shane M wrote:
> I found the manual pages really confusing so I had a go at rewriting
> them - there were places in the manual page that didn't match the
> functionality provided by your code as well as I could tell).

Could you point which places don't match the code?

> 
> My apologies for a few formatting issues though. I still don't like
> parts of epoll_pwait1 but it's less confusing than it was.

Any other than the timespec question don't you like?

> 
> You are free to take some or all or none of the changes.
> 
> I did have a question I marked with **** below about what you
> describe and what your code does.
> 

<snip>

>        The timeout member specifies the minimum time that epoll_wait(2) will
>        block. The time spent waiting will be rounded up to the clock
>        granularity. Kernel scheduling delays mean that the blocking
>        interval may overrun by a small amount. Specifying a -1 for either
>        tv_sec or tv_nsec member of the struct timespec timeout will cause
>        causes epoll_pwait1(2) to block indefinitely. Specifying a timeout
>        equal to zero (both tv_sec or tv_nsec member of the struct timespec
>        timeout are zero) causes epoll_wait(2) to return immediately, even
>        if no events are available.
> 
> **** Are you really really sure about this for the -1 stuff? your code copies
> in the timespec and just passes it to timespec_to_ktime:
> 
> +	if (copy_from_user(&p, params, sizeof(p)))
> +		return -EFAULT;
> ...
> +	kt = timespec_to_ktime(p.timeout);
> 
> Compare that to something like the futex syscall which does this:
> 
> 		if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
> 			return -EFAULT;
> 		if (!timespec_valid(&ts))
> 			return -EINVAL;
> 
> 		t = timespec_to_ktime(ts);
> 
> If the timespec is not valid it returns -EINVAL back to user space. With your
> settings of tv_sec and/or tv_usec to -1 are you relying on a side effect of
> the conversion that could break your code in the future if in the unlikely
> event someone changes timespec_to_ktime() and should it be:
> 
> +	if (copy_from_user(&p, params, sizeof(p)))
> +		return -EFAULT;
> +       if ((p.timeout.tv_sec == -1) || (p.timeout.tv_nsec == -1)) {
> +  /* this is off the top of my head no idea if it will compile */
> +		p.timeout.tv_sec = KTIME_SEC_MAX;
> +		p.timeout.tv_nsec = 0;
> +	}
> +       if (!timespec_valid(&p.timeout))
> +       	return -EINVAL;
> ...
> +	kt = timespec_to_ktime(p.timeout);

OK. timespec_valid() is clear about this: negative tv_sec is invalid, so I
don't think accepting -1 from user is the right thing to do.

We cannot do pointer check as ppoll already because the structure is embedded
in epoll_wait_params.

Maybe it's best to use a flags bit (#define EPOLL_PWAIT1_BLOCK 1).  What do you
think?

Fam

<snip>
--
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/