Re: ksoftirqd uses 99% CPU triggered by network traffic (maybe RLT-8139 related)

[Robert Olsson <Robert.Olsson@data.slu.se>]

Pasi Sjoholm writes:

 > Pid: 2, comm:          ksoftirqd/0
 > EIP: 0060:[<e0871224>] CPU: 0
 > EIP is at rtl8139_poll+0xb4/0x100 [8139too]
 >  EFLAGS: 00000247    Not tainted  (2.6.7-mm7)
 > EAX: ffffe000 EBX: 00000040 ECX: df4824f8 EDX: c0441978
 > ESI: df482400 EDI: e0868000 EBP: dff85f80 DS: 007b ES: 007b
 > CR0: 8005003b CR2: b7c5a000 CR3: 1fafd000 CR4: 000006d0
 >  [<c0119580>] ksoftirqd+0x0/0xc0
 >  [<c02c5f3a>] net_rx_action+0x6a/0x110
 >  [<c01191a9>] __do_softirq+0xa9/0xb0
 >  [<c01191d7>] do_softirq+0x27/0x30
 >  [<c01195e8>] ksoftirqd+0x68/0xc0
 >  [<c01277e5>] kthread+0xa5/0xb0
 >  [<c0127740>] kthread+0x0/0xb0
 >  [<c0102111>] kernel_thread_helper+0x5/0x14
 > --
 > 
 > I'm not a kernel expert but it would seem that ksoftirqd is in some sort a 
 > loop because I didn't get any "printk("%s wakes ksoftirqd\n", 
 > current->comm);"-lines.

 Hello!
 This looks very much like the problem we see when doing route DoS testing
 with Alexey.

 In summary: High softirq loads can totally kill userland. The reason is that 
 do_softirq() is run from many places hard interrupts, local_bh_enable etc 
 and bypasses the ksoftirqd protection. It just been discussed at OLS with
 Andrea and Dipankar and others. Current RCU suffers from this problem as well.

 I've experimented some code to defer softirq's to ksoftirqd after a time as 
 well as deferring all softirq's to ksoftirqd. Andrea had some ideas as well 
 as Ingo.

 Cheers.
						--ro