Date: Tue, 17 Dec 2002 16:05:23 -0500 (EST) From: Andrew Gallatin <gallatin@cs.duke.edu> To: Kyunghwan Kim <redjade@ada.snu.ac.kr> Cc: current@FreeBSD.ORG Subject: Re: INTR_MPSAFE to network device drivers Message-ID: <15871.37267.31358.280147@grasshopper.cs.duke.edu> In-Reply-To: <20021217202402.GA57385@ada.snu.ac.kr> References: <20021217191841.GA57094@ada.snu.ac.kr> <15871.31635.962334.855790@grasshopper.cs.duke.edu> <20021217195300.GB57094@ada.snu.ac.kr> <20021217202402.GA57385@ada.snu.ac.kr>
next in thread | previous in thread | raw e-mail | index | archive | help
Kyunghwan Kim writes: > On Wed, Dec 18, 2002 at 04:53:00AM +0900, Kyunghwan Kim wrote: > > On Tue, Dec 17, 2002 at 02:31:31PM -0500, Andrew Gallatin wrote: > > > > mbuf and bpf routines are all mp-safe, so it seems that > > > > it is safe to make network device drivers out of Giant lock. > > > > Or is there any unresolved related issues? > > > > > > Yes, the mbuf allocator must occasionally call kmem_malloc(), which > > > requires Giant. This means no net driver can be made INTR_MPSAFE, > > > or it will eventually panic when kmem_malloc is called. > > > > I found and read the thread that you and Alan had discussed about this > > problem just before. Then what about making updated version of mb_pop_cont() > > that accepts occasionally acquiring Giant? > > Oh, sorry. Conclusion of the thread was preallocation. > But it doesn't seem that preallocation is the correct way. Well, the "right" way is bringing kmem_malloc() out from under Giant. Preallocation would just be an interim measure until kmem_malloc() was mp-safe. The VM maintainers were only recently made aware of this issue. I have no idea what the scope of the changes to make kmem_malloc() mpsafe would be, so I have no idea when kmem_malloc() will become mpsafe. Drew To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?15871.37267.31358.280147>