On 29.9.2015. 1:00, Hrvoje Popovski wrote: > On 28.9.2015. 23:53, Mark Kettenis wrote: >> This diff makes the tx completion path run without the kernel lock >> held. With this change, the interrupt handler will not grab the >> kernel lock under normal circumstances. The diff follows the same >> approach as dlg@ took with vmx(4). >> >> The diff removes the code that tries to reclaim tx descriptors in >> if_start() if we have only EM_TX_CLEANUP_THRESHOLD available >> descriptors left. Running tx_eof() in both the normal transmit path >> and from the interrupt handler becomes a locking nightmare. >> >> Seems to survive some serious stress-testing with tcpbench and ping -f >> on my x220: >> >> em0 at pci0 dev 25 function 0 "Intel 82579LM" rev 0x04: msi, address >> f0:de:f1:70:87:8a >> >> Could use some further testing of course. > > > > Hi, > i'm getting 720kpps with 128 byte packets through bridge with two i350. > Everything seems stable and openbsd box is responsive. Will leave it > over night ....
with routed setup i'm getting stable 1Mpps (64b packets) and box is happy and responsive ... both tests are with pf disabled ... sysctl kern.pool_debug=0 net.inet.ip.forwarding=1 net.inet.ip.ifq.maxlen=8192 ddb.console=1
