On Mon, Sep 09, 2013 at 01:08:47PM +0200, Michal Hocko wrote: > On Fri 06-09-13 22:59:16, Hugh Dickins wrote: > > Hit divide-by-0 in vmpressure_work_fn(): checking vmpr->scanned before > > taking the lock is not enough, we must check scanned afterwards too. > > As vmpressure_work_fn seems the be the only place where we set scanned > to 0 (except for the rare occasion when scanned overflows which > would be really surprising) then the only possible way would be two > vmpressure_work_fn racing over the same work item. system_wq is > !WQ_NON_REENTRANT so one work item might be processed by multiple > workers on different CPUs. This means that the vmpr->scanned check in > the beginning of vmpressure_work_fn is inherently racy. > > Hugh's patch fixes the issue obviously but doesn't it make more sense to > move the initial vmpr->scanned check under the lock instead? > > Anton, what was the initial motivation for the out of the lock > check? Does it really optimize anything? Thanks a lot for the explanation. Answering your question: the idea was to minimize the lock section, but the section is quite small anyway so I doubt that it makes any difference (during development I could not measure any effect of vmpressure() calls in my system, though the system itself was quite small). I am happy with moving the check under the lock or moving the work into its own WQ_NON_REENTRANT queue. Anton -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>