Re: linux-next 20141216 BUG: sleeping function called from invalid context at mm/slab.c:2849

On 14/12/18, Eric Paris wrote: > On Thu, 2014-12-18 at 12:46 -0500, Richard Guy Briggs wrote: > > On 14/12/18, Eric Paris wrote: > > > On Thu, 2014-12-18 at 11:45 -0500, Valdis.Kletnieks(a)vt.edu wrote: > > > > On Tue, 16 Dec 2014 20:09:54 -0500, Valdis Kletnieks said: > > > > > Spotted these two while booting single-user on 20141216. 20141208 > > > > > doesn't throw these, so it's something in the last week or so.. > > > > > > > > Gaah! Turns out that 20141208 *is* susceptible - it had been booting > > > > just fine for several days, but it went around the bend, apparently due > > > > to a userspace or initrd change. > > > > > > $5 says you updated systemd? > > > > > > Richard? > > > > Ok, so if you are correct, then either we justify dropping the lock (I > > assume the one commone to both BUG reports [sig->cred_guard_mutex] ), > > or we make yet another queue were were hoping to avoid... > > > > It would also be good to narrow it down to a rule that triggers this. > > I thought the first message was enough to find the problem, but: > > static void kauditd_send_multicast_skb(struct sk_buff *skb) > { > ... > nlmsg_multicast(sock, copy, 0, AUDIT_NLGRP_READLOG, GFP_KERNEL); > ... > } > > Since kauditd_send_multicast_skb() gets called in audit_log_end(), which > can come from any context (aka even a sleeping context) you can't use > GFP_KERNEL. The audit_buffer know what context it should use. So pass > that down and use that. Ok, that looks more obvious now... We just need to change the internal interface to kauditd_send_multicast_skb() to accept an audit_buffer instead of just the skb and use the gfp_mask value from there instead of using our own... Thanks, Eric.