Re: tty update speed regression (was: 2.6.14-rc2-mm1)

[Alexey Dobriyan]

On Fri, Sep 23, 2005 at 12:07:49PM -0700, Nishanth Aravamudan wrote:
> On 23.09.2005 [22:42:16 +0400], Alexey Dobriyan wrote:
> > poll([{fd=0, events=POLLIN}], 1, 0) = 0

> > I can send full strace log if needed.
> 
> Nope, that helped tremendously! I think I know what the issue is (and
> why it's HZ dependent).
> 
> In the current code, (2.6.13.2, e.g) we allow 0 timeout poll-requests to
> be resolved as 0 jiffy requests. But in my patch, those requests become
> 1 jiffy (which of course depends on HZ and gets quite long if HZ=100)!
> 
> Care to try the following patch?

It works! Now, even with HZ=100, gameplay is smooth.

Andrew, please, apply.

> Description: Modifying sys_poll() to handle large timeouts correctly
> resulted in 0 being treated just like any other millisecond request,
> while the current code treats it as an optimized case. Do the same in
> the new code. Most of the code change is tabbing due to the inserted if.

> diff -urpN 2.6.14-rc2-mm1/fs/select.c 2.6.14-rc2-mm1-dev/fs/select.c
> --- 2.6.14-rc2-mm1/fs/select.c	2005-09-23 11:52:36.000000000 -0700
> +++ 2.6.14-rc2-mm1-dev/fs/select.c	2005-09-23 12:04:03.000000000 -0700
> @@ -485,26 +485,35 @@ asmlinkage long sys_poll(struct pollfd _
>  	if (nfds > max_fdset && nfds > OPEN_MAX)
>  		return -EINVAL;
>  
> -	/*
> -	 * We compare HZ with 1000 to work out which side of the
> -	 * expression needs conversion.  Because we want to avoid
> -	 * converting any value to a numerically higher value, which
> -	 * could overflow.
> -	 */
> +	if (timeout_msecs) {
> +		/*
> +		 * We compare HZ with 1000 to work out which side of the
> +		 * expression needs conversion.  Because we want to
> +		 * avoid converting any value to a numerically higher
> +		 * value, which could overflow.
> +		 */
>  #if HZ > 1000
> -	overflow = timeout_msecs >= jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT);
> +		overflow = timeout_msecs >=
> +				jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT);
>  #else
> -	overflow = msecs_to_jiffies(timeout_msecs) >= MAX_SCHEDULE_TIMEOUT;
> +		overflow = msecs_to_jiffies(timeout_msecs) >=
> +				MAX_SCHEDULE_TIMEOUT;
>  #endif
>  
> -	/*
> -	 * If we would overflow in the conversion or a negative timeout
> -	 * is requested, sleep indefinitely.
> -	 */
> -	if (overflow || timeout_msecs < 0)
> -		timeout_jiffies = MAX_SCHEDULE_TIMEOUT;
> -	else
> -		timeout_jiffies = msecs_to_jiffies(timeout_msecs) + 1;
> +		/*
> +		 * If we would overflow in the conversion or a negative
> +		 * timeout is requested, sleep indefinitely.
> +		 */
> +		if (overflow || timeout_msecs < 0)
> +			timeout_jiffies = MAX_SCHEDULE_TIMEOUT;
> +		else
> +			timeout_jiffies = msecs_to_jiffies(timeout_msecs) + 1;
> +	} else {
> +		/*
> +		 * 0 millisecond requests become 0 jiffy requests
> +		 */
> +		timeout_jiffies = 0;
> +	}
>  
>  	poll_initwait(&table);
>  

-
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/