On Mon, Sep 13, 2010 at 06:35:10AM +0200, Bret S. Lambert wrote: > On Mon, Sep 13, 2010 at 10:12:44AM +0600, Anton Maksimenkov wrote: > > 2010/9/13 Henning Brauer <lists-open...@bsws.de>: > > >> hangs. 1-2 sec after start. > > >> --- interrupt --- > > >> pool_do_get(d0a10b60,0,0,0,60) at pool_do_get+0x2c2 > > >> pool_get(d0a10b60,0,80000000,0,0) at pool_get+0x54 > > >> m_gethdr(1,1,80000000,369e99,0) at m_gethdr+0x39 > > > too me that simply looks like you are running out of memory in mbpl > > > and the pool_get is a M_WAITOK one > > > > But it not unfreezed even after minute. SSH connections dropped, com > > console didn't response (but it can be dropped into ddb, of course). > > yes, because you've soaked up all the memory that's available for > handling incoming/outgoing network traffic; you've got a bunch of > processes that try to grab a limited number of resources, fail to > get all they need, and sleep while holding already-allocated mbufs, > meaning that nobody else can get them, and none of your processes > can advance. > > That said, the pool_get that's failing in the re driver is set as > non-blocking, so it should fail. However, it's hard to see how > you're tickling this without seeing the source that you're running, > since we don't know how you're cornholing the network stack. >
Oh, that pool_get succeds since mbufs don't have a limit but the allocation of the cluster fails so the driver will reuse the old buffer on the queue. By breaking into ddb it just happend to be in there instead of somewhere else. When running with that many sockets a prominent warning about increasing kern.maxclusters shows up. This is not just dmesg spam, running out of mbuf clusters will stop your network stack. Most systems today come up with 6144 clusters. A TCP connection needs at least 4 clusters plus the network stack itself needs a few clusters for things like defragmenting long mbuf chains. So to be able to run with 10000 sockets you should set kern.maxclusters to around 100000. This will allow the network stack to allocate around 200MB of memory in kva. -- :wq Claudio
