Re: [PATCH 6/6] fs-writeback: only allow one inflight and pending !nr_pages flush

From: Johannes Weiner
Date: Tue Sep 19 2017 - 16:18:51 EST


On Tue, Sep 19, 2017 at 01:53:07PM -0600, Jens Axboe wrote:
> A few callers pass in nr_pages == 0 when they wakeup the flusher
> threads, which means that the flusher should just flush everything
> that was currently dirty. If we are tight on memory, we can get
> tons of these queued from kswapd/vmscan. This causes (at least)
> two problems:
>
> 1) We consume a ton of memory just allocating writeback work items.
> 2) We spend so much time processing these work items, that we
> introduce a softlockup in writeback processing.
>
> Fix this by adding a 'zero_pages' bit to the writeback structure,
> and set that when someone queues a nr_pages==0 flusher thread
> wakeup. The bit is cleared when we start writeback on that work
> item. If the bit is already set when we attempt to queue !nr_pages
> writeback, then we simply ignore it.
>
> This provides us one of full flush in flight, with one pending as
> well, and makes for more efficient handling of this type of
> writeback.
>
> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>

Acked-by: Johannes Weiner <hannes@xxxxxxxxxxx>

Just a nitpick:

> @@ -948,15 +949,25 @@ static void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
> bool range_cyclic, enum wb_reason reason)
> {
> struct wb_writeback_work *work;
> + bool zero_pages = false;
>
> if (!wb_has_dirty_io(wb))
> return;
>
> /*
> - * If someone asked for zero pages, we write out the WORLD
> + * If someone asked for zero pages, we write out the WORLD.
> + * Places like vmscan and laptop mode want to queue a wakeup to
> + * the flusher threads to clean out everything. To avoid potentially
> + * having tons of these pending, ensure that we only allow one of
> + * them pending and inflight at the time
> */
> - if (!nr_pages)
> + if (!nr_pages) {
> + if (test_bit(WB_zero_pages, &wb->state))
> + return;
> + set_bit(WB_zero_pages, &wb->state);
> nr_pages = get_nr_dirty_pages();

We could rely on the work->older_than_this and pass LONG_MAX here
instead to write out the world as it was at the time wb commences.

get_nr_dirty_pages() is somewhat clearer on intent, but on the other
hand it returns global state and is used here in a split-bdi context,
and we can end up in sum requesting the system-wide dirty pages
several times over. It'll work fine, relying on work->older_than_this
to contain it also, it just seems a little ugly and subtle.