On Wed 04-09-19 15:41:44, Sergey Senozhatsky wrote: > On (09/04/19 08:15), Michal Hocko wrote: > > > If you look at the original report, the failed allocation dump_stack() is, > > > > > > <IRQ> > > > warn_alloc.cold.43+0x8a/0x148 > > > __alloc_pages_nodemask+0x1a5c/0x1bb0 > > > alloc_pages_current+0x9c/0x110 > > > allocate_slab+0x34a/0x11f0 > > > new_slab+0x46/0x70 > > > ___slab_alloc+0x604/0x950 > > > __slab_alloc+0x12/0x20 > > > kmem_cache_alloc+0x32a/0x400 > > > __build_skb+0x23/0x60 > > > build_skb+0x1a/0xb0 > > > igb_clean_rx_irq+0xafc/0x1010 [igb] > > > igb_poll+0x4bb/0xe30 [igb] > > > net_rx_action+0x244/0x7a0 > > > __do_softirq+0x1a0/0x60a > > > irq_exit+0xb5/0xd0 > > > do_IRQ+0x81/0x170 > > > common_interrupt+0xf/0xf > > > </IRQ> > > > > > > Since it has no __GFP_NOWARN to begin with, it will call, > > I think that DEFAULT_RATELIMIT_INTERVAL and DEFAULT_RATELIMIT_BURST > are good when we ratelimit just a single printk() call, so the ratelimit > is "max 10 kernel log lines in 5 seconds". I am sorry, I could have been more explicit when CCing you. Sure the ratelimit is part of the problem. But I was more interested in the potential livelock (infinite loop) mentioned by Qian Cai. It is not important whether we generate one or more lines of output from the softirq context as long as the printk generates more irq processing which might end up doing the same. Is this really possible? -- Michal Hocko SUSE Labs
