On So 24-12-05 10:11:13, Alan Stern wrote: > On Fri, 23 Dec 2005, Greg KH wrote: > > > On Fri, Dec 23, 2005 at 04:22:44PM -0500, Alan Stern wrote: > > > On Fri, 23 Dec 2005, Pavel Machek wrote: > > > > > > > Well, if you can find some elegant solution in the core, I think thats > > > > the best way. > > > > > > > > You could set system_state to "suspending" or something like that, and > > > > just if() out notifications in that case. > > > > > > How about simply adding a call to try_to_freeze() somewhere inside > > > kernel/kmod.c:____call_usermodehelper()? That ought to do pretty much > > > what I want, in theory. The hotplug processes would get frozen before > > > /sbin/hotplug is exec'ed. > > > > On modern distros, /sbin/hotplug is set to NULL, so this isn't an issue. > > We use netlink to send the data out, so this might not even be a problem > > anymore... > > Indeed it might not. On my older FC3 system I end up with unwanted > processes, but under FC4 that doesn't happen. (Note however that even on > the FC4 system, /sbin/hotplug is not empty and /proc/sys/kernel/hotplug > does point to /sbin/hotplug.) > > The ideal situation would be to have PF_FREEZE automatically set for every > new process created while userspace was frozen. Kernel threads could > remove the flag and set PF_NOFREEZE if they wanted, while user processes > would automatically get frozen before returning to userspace. This > approach would eliminate all the loopholes at once. It would also prevent some clever stuff with u-swsusp. We'll want to keep u-swsusp program controlling suspend. I don't see how it needs to exec() today, but eventually it may want to run some properly-audited helpers... ...okay, better not, because that would mean memory allocation while suspended => bad. Anyway, doing check in exec() or something like that would slow down hot path. Just take fix userland exec's during suspend one-by-one. We are doing something wrong if we attempt that, anyway. Pavel -- Thanks, Sharp!
