Date: Tue, 25 May 2004 14:17:40 -0700 (PDT) From: Julian Elischer <julian@elischer.org> To: Matthew Dillon <dillon@apollo.backplane.com> Cc: Alan Cox <alc@cs.rice.edu> Subject: Re: Network Stack Locking Message-ID: <Pine.BSF.4.21.0405251410580.43418-100000@InterJet.elischer.org> In-Reply-To: <200405251829.i4PITjbp093921@apollo.backplane.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, 25 May 2004, Matthew Dillon wrote: > > :>:Sounds a lot like a lot of the Mach IPC optimizations, including their use > :>:of continuations during IPC to avoid a full context switch. > :>: > :>:Robert N M Watson FreeBSD Core Team, TrustedBSD Projects > :>:robert@fledge.watson.org Senior Research Scientist, McAfee Research > :> > :> Well, I like the performance aspects of a continuation mechanism, but > :> I really dislike the memory overhead. Even a minimal stack is > :> expensive when you multiply it by potentially hundreds of thousands > :> of 'blocking' entities such as PCBs.. say, a TCP output stream. > :> Because of this the overhead and cache pollution generated by the > :> continuation mechanism increases as system load increases rather > :> then decreases. > : > :When the explicit continuation mechanism was used, the thread's stack > :was freed when the thread blocked and a new stack allocated when the > :thread was restarted. Here is a URL: > :http://citeseer.ist.psu.edu/draves91using.html. Notice the mention of > :space reduction in the abstract. It's worth reading. > : > :Alan > > Ah, now I understand. The continuation is an exit from the procedure. > > This sounds very similar to an interrupt mechanism I designed about > a decade ago. Instead of saving and restoring the interrupt > context for the interrupt thread, the thread switch was special-cased > to basically create a context (by pushing a procedure call on the stack) > on switch-in and to throw it away on switch-out (which was always at > the end of the interrupt routine). I extended the same mechanism down > into 'userland' by creating a 'waitforever()' system call which basically > did nothing but wait for and dispatch signal vectors (most of the programs > were event oriented and had no main loop). The nice thing about this was > that no context had to be saved while the program was sitting in > waitforever(), or restored when the program returned from a signal handler. > This more then doubled scheduler performance (which on a 10MHz 68000 was > important). > > In many ways, the continuation mechanism and the message queue mechanism > appear to be nearly identical. If an explicit exit from a procedure > is required to optimize the stack with the continuation mechanism, then > that isn't much different then moving the message to an event queue > and returning to the message processing loop. Neither case allows > you to save stack context or to save the current procedural stacking > level, and both mechanisms allow you to reuse your current stack to > handle multiple messages/continuations. This is how netgraph works. (particularly in 5.x) messages are queued for a node.. they can have 3 types 1/ mbuf data 2/ control message 3/ procedure execution request. the 3rd type is for such things as: timeouts.. the timeout request simply adds a request to the queue that the node execute it's timeout function. The queue entry specifies whether it needs to have a reader or writer lock on the node. By default (but not written in stone), mbuf data looks for a reader, and the others request a writer lock, so data could pass through the same node at the same time on different cpus but a control message woudl not be able to be done at the same time as anything else. If the node lock can be had at the time that the item is to be queued, and there is nothing else in the queue, then the context of the caller can be used to process the item imediatly. Otherwise whoever has the lock should process all waiting items when they finish. > > -Matt > Matthew Dillon > <dillon@backplane.com> > _______________________________________________ > freebsd-arch@freebsd.org mailing list > http://lists.freebsd.org/mailman/listinfo/freebsd-arch > To unsubscribe, send any mail to "freebsd-arch-unsubscribe@freebsd.org" >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.BSF.4.21.0405251410580.43418-100000>