On (09/29/16 15:25), Petr Mladek wrote: > On Tue 2016-09-27 23:22:30, Sergey Senozhatsky wrote: > > Hello, > > > > RFC > > > > This patch set extends a lock-less NMI per-cpu buffers idea to > > handle recursive printk() calls. The basic mechanism is pretty much the > > same -- at the beginning of a deadlock-prone section we switch to lock-less > > printk callback, and return back to a default printk implementation at the > > end; the messages are getting flushed to a logbuf buffer from a safer > > context. > > I was skeptical but I really like this way now. > > The switching of the buffers is a bit hairy in this version but I > think that we could make it much better. > > Other than that it looks like a big win. It kills a lot of > printk-related pain points. And it will not be that complicated > after all.
many thanks for looking at this train wreck. so, like I said, it addresses printk()-recursion in *ideally* quite a minimalistic way -- just several alt_printk_enter/exit calls in printk.c, without ever touching any other parts of the kernel. gunning down printk deadlocks in general, however, requires much more effort; or even a completely different approach. a) a lock-less printk() by default um, `#define printk alt_printk'. but this will break printk() from irq. and the ordering of messages from per-cpu buffers may be far from correct. b) combining a DEFERRED_WARN + alt_printk DEFERRED_WARN potentially is a never ending thing. we can add some lockdep annotations, perhaps, and hope that error handling branches that may contain WARN_ONs/printk-s will be executed with prove_locking enabled on someone's machine. c) ... -ss