On Thu, May 21, 2015 at 09:38:23AM +0800, long.wanglong wrote: > On 2015/5/20 21:22, Petr Mladek wrote: > > On Tue 2015-05-19 14:57:46, Petr Mladek wrote: > >> On Tue 2015-05-19 09:08:45, Wang Long wrote: > >>> This is my backport patch series to Fix the problem(backport to 3.10): > >>> " > >>> When trigger_all_cpu_backtrace() is called on x86, it will trigger an > >>> NMI on each CPU and call show_regs(). But this can lead to a hard lock > >>> up if the NMI comes in on another printk(). > >>> " > >>> The solution is described in commit "a9edc88093287183ac934be44f295f183b2c62dd": > >>> when the NMI triggers, it switches the printk routine for that CPU to call > >>> a NMI safe printk function that records the printk in a per_cpu seq_buf > >>> descriptor. After all NMIs have finished recording its data, the trace_ > >>> seqs are printed in a safe context. > >>> > >>> The solution use "switch printk routine" and "seq_buf" infrastructures, but the > >>> 3.10 stable have no both of them. > >>> > >>> The patch 1-13 backport the "seq_buf" infrastructures. in detail, patch 1, 2 > >>> and 6 only backport "seq_buf" related code. > >>> > >>> The patch 14-15 backport the "switch printk routine". > >>> > >>> The patch 16-17 is the patch to print all cpu stacks from NMI safely > >>> > >>> as discussed in https://lkml.org/lkml/2015/5/13/497, in 3.10 stable, this is > >>> the only way to solve the problem and the backport code is a bit more. > >>> > >>> v1 -> v2: > >>> * fix the indent error. > >>> * rebase on 3.10.79 > >>> > >>> Any thoughts? > >> > >> Please, wait with the integration. I am testing it with a storm of > >> sysrq requests: > >> > >> $> while true ; do echo l >/proc/sysrq-trigger ; done > >> > >> with iptables enabled: > >> > >> $> iptables -A INPUT -j LOG --log-prefix "incomming packet:" > >> > >> and storm of pings from other machine: > >> > >> $> ping -f <patched-host> > >> > >> > >> The machine somehow freezes. It does not make sense. I am trying to investigate. > > > > OK, it seems that the machine freezes because there are still few > > messages printed in the NMI context, e.g.: > > > > [ 3080.286277] Uhhuh. NMI received for unknown reason 3d on CPU 12. > > [ 3637.939276] Uhhuh. NMI received for unknown reason 2d on CPU 13. > > > > I am not exactly sure why I get them on the test machine. But I get > > such messages from time to time when hammering it by the pings and > > sysrq-l requests. > > > > I modified vprintk_emit() to do raw_spin_trylock(&logbuf_lock) > > and do not try to lock console in NMI context. The trylock fails > > from time to time but it does not longer freeze. > > > > I am going to clean up the vprintk_emit() modification and send it for > > review. > > > > Anyway, this patch set seems to work as expected. It heavily reduces > > the risk of NMI/printk-related deadlocks => it is worth having. > > > > Feel free to use the following for the whole patchset (backport): > > > > Reviewed-by: Petr Mladek <pmladek@xxxxxxx> > > Tested-by: Petr Mladek <pmladek@xxxxxxx> > > Hi Greg, > > This patch set is the only way to solve the NMI/printk-related deadlock problems. > Could you please include them to 3.10 stable? > > Although the code a bit more, most of the code is "seq_buf" infrastructures and > it does not affect other parts of the kernel. Yeah, but this is way too much for a -stable kernel. I suggest that if a user has this problem, please move to 3.14 or newer kernels, which has this fixed. There's too many changes here for me to be confortable accepting to a -stable kernel, sorry. greg k-h -- To unsubscribe from this list: send the line "unsubscribe stable" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html