Re: [PATCH] raceless request_region() fix (was Re: Linux 2.6.0-test4)

From: insecure <insecure@mail.od.ua>
To: Jeff Garzik <jgarzik@pobox.com>
Cc: Kernel Mailing List <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] raceless request_region() fix (was Re: Linux 2.6.0-test4)
Date: Tue, 26 Aug 2003 08:20:13 +0300	[thread overview]
Message-ID: <200308260820.13205.insecure@mail.od.ua> (raw)
In-Reply-To: <3F4A86B8.9080404@pobox.com>

[de-cc'ing Linus]

On Tuesday 26 August 2003 00:59, Jeff Garzik wrote:
> Is it a race if noone cares?  :)
>
> The code does
>
> 	if (!request_region(...))
> 		fail
> 	touch hardware
> 	release_region
> 	if (!request_region(...))
> 		fail
>
> If the HIGHLY UNLIKELY event of another ISA driver claiming this region
> occurs, the system continues working just fine.

Yes, this particular driver is fine. My patch just eliminate back-to-back
 	release_region(...)
 	if (!request_region(...)) fail;
pair.

I just have a feeling that if I am going to replace check_region with
request_region, I can close race window altogether.

BTW, it all started when I tried to get rid of check_region in de4x5.c.
That's a bit worse. For example:
        barrier();
        request_region(iobase, (lp->bus == PCI ? DE4X5_PCI_TOTAL_SIZE :
                                DE4X5_EISA_TOTAL_SIZE),
                       lp->adapter_name);
        lp->rxRingSize = NUM_RX_DESC;
-- 
vda