Date: Sun, 27 Nov 2005 21:45:45 +0200 From: Ruslan Ermilov <ru@FreeBSD.org> To: Gleb Smirnoff <glebius@FreeBSD.org> Cc: Vsevolod Lobko <seva@ip.net.ua>, rwatson@FreeBSD.org, net@FreeBSD.org Subject: Re: parallelizing ipfw table Message-ID: <20051127194545.GA76200@ip.net.ua> In-Reply-To: <20051127135529.GF25711@cell.sick.ru> References: <20051127005943.GR25711@cell.sick.ru> <20051127135529.GF25711@cell.sick.ru>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sun, Nov 27, 2005 at 04:55:29PM +0300, Gleb Smirnoff wrote: > On Sun, Nov 27, 2005 at 03:59:43AM +0300, Gleb Smirnoff wrote: > T> A patch displaying the idea is attached. Not tested yet, read > T> below. The patch moves the tables array into the ip_fw_chain > T> structure. This is not necessary now, but in future we can > T> have multiple independent chains in ipfw, that's why I try > T> to avoid usage of &layer3_chain in the functions that are > T> deeper in the call graph. I try to supply chain pointer > T> from the caller. > T> > T> The only problem is the caching in table lookup. This "hack" > T> makes the lookup function modify the table structure. We need > T> to remove caching to make the lookup_table() function fully > T> lockless and reenterable at the same time. The attached patch > T> doesn't removes caching, since it only displays the original > T> idea. > > Okay, I have made a working patch, that is now undergoing testing > on SMP. I have axed all the caching from ipfw tables, to make > lookup_table() lockless and reenterable. This axing simplified > things much. I believe that the caching gives a benefit only > when we serve a small number of clients, and is only additional > workload when we are routing hundreds and thousands of simultaneous > IP flows. > > The patch attached. I'm going to put it into production testing as > soon as I can reboot the prod box. > Nope, I need this caching. It's for looking up the same table several times in a row but with various values. For example, we use ipfw tables to route the traffic to the correct dummynet pipe, where value is the bandwidth, and this caching helps a lot. Cheers, -- Ruslan Ermilov ru@FreeBSD.org FreeBSD committer
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20051127194545.GA76200>