On Thu 05-05-16 10:14:52, Jan Kara wrote: > When system is under memory pressure memory management frequently calls > wakeup_flusher_threads() to writeback pages to that they can be freed. > This was observed to exhaust reserves for atomic allocations since > wakeup_flusher_threads() allocates one writeback work for each device > with dirty data with GFP_ATOMIC. > > However it is pointless to allocate new work items when requested work > is identical. Instead, we can merge the new work with the pending work > items and thus save memory allocation. Makes sense. See one question below: > Reported-by: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx> > Signed-off-by: Jan Kara <jack@xxxxxxx> > --- > fs/fs-writeback.c | 37 +++++++++++++++++++++++++++++++++++++ > include/trace/events/writeback.h | 1 + > 2 files changed, 38 insertions(+) > > This is a patch which should (and in my basic testing does) address the issues > with many atomic allocations Tetsuo reported. What do people think? > > diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c > index fee81e8768c9..bb6725f5b1ba 100644 > --- a/fs/fs-writeback.c > +++ b/fs/fs-writeback.c > @@ -189,6 +189,35 @@ out_unlock: > spin_unlock_bh(&wb->work_lock); > } > > +/* > + * Check whether the request to writeback some pages can be merged with some > + * other request which is already pending. If yes, merge it and return true. > + * If no, return false. > + */ > +static bool wb_merge_request(struct bdi_writeback *wb, long nr_pages, > + struct super_block *sb, bool range_cyclic, > + enum wb_reason reason) > +{ > + struct wb_writeback_work *work; > + bool merged = false; > + > + spin_lock_bh(&wb->work_lock); > + list_for_each_entry(work, &wb->work_list, list) { Is the lenght of the list bounded somehow? In other words is it possible that the spinlock would be held for too long to traverse the whole list? > + if (work->reason == reason && > + work->range_cyclic == range_cyclic && > + work->auto_free == 1 && work->sb == sb && > + work->for_sync == 0) { > + work->nr_pages += nr_pages; > + merged = true; > + trace_writeback_merged(wb, work); > + break; > + } > + } > + spin_unlock_bh(&wb->work_lock); > + > + return merged; > +} -- Michal Hocko SUSE Labs -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html