On Mon 25-06-18 16:04:04, peter enderborg wrote: > On 06/25/2018 03:07 PM, Michal Hocko wrote: > > > On Mon 25-06-18 15:03:40, peter enderborg wrote: > >> On 06/20/2018 01:55 PM, Michal Hocko wrote: > >>> On Wed 20-06-18 20:20:38, Tetsuo Handa wrote: > >>>> Sleeping with oom_lock held can cause AB-BA lockup bug because > >>>> __alloc_pages_may_oom() does not wait for oom_lock. Since > >>>> blocking_notifier_call_chain() in out_of_memory() might sleep, sleeping > >>>> with oom_lock held is currently an unavoidable problem. > >>> Could you be more specific about the potential deadlock? Sleeping while > >>> holding oom lock is certainly not nice but I do not see how that would > >>> result in a deadlock assuming that the sleeping context doesn't sleep on > >>> the memory allocation obviously. > >> It is a mutex you are supposed to be able to sleep. It's even exported. > > What do you mean? oom_lock is certainly not exported for general use. It > > is not local to oom_killer.c just because it is needed in other _mm_ > > code. > > > > It is in the oom.h file include/linux/oom.h, if it that sensitive it should > be in mm/ and a documented note about the special rules. It is only used > in drivers/tty/sysrq.c and that be replaced by a help function in mm that > do the oom stuff. Well, there are many things defined in kernel header files and not meant for wider use. Using random locks is generally discouraged I would say unless you are sure you know what you are doing. We could do some more work to hide internals for sure, though. > >>>> As a preparation for not to sleep with oom_lock held, this patch brings > >>>> OOM notifier callbacks to outside of OOM killer, with two small behavior > >>>> changes explained below. > >>> Can we just eliminate this ugliness and remove it altogether? We do not > >>> have that many notifiers. Is there anything fundamental that would > >>> prevent us from moving them to shrinkers instead? > >> @Hocko Do you remember the lowmemorykiller from android? Some things > >> might not be the right thing for shrinkers. > > Just that lmk did it wrong doesn't mean others have to follow. > > > If all you have is a hammer, everything looks like a nail. (I don’t argument that it was right) > But if you don’t have a way to interact with the memory system we will get attempts like lmk. > Oom notifiers and vmpressure is for this task better than shrinkers. A lack of feature should be a trigger for a discussion rather than a quick hack that seems to work for a particular usecase and live out of tree, then get to staging and hope it will fix itself. Seriously, the kernel development is not a nail hammering. -- Michal Hocko SUSE Labs
