Re: [PATCH 2/8] writeback: move dirty inodes from super_block to backing_dev_info

[Jens Axboe <jens.axboe@oracle.com>]

On Tue, Sep 08 2009, Christoph Hellwig wrote:
> On Sun, Sep 06, 2009 at 08:43:59PM +0200, Jens Axboe wrote:
> > On Fri, Sep 04 2009, Christoph Hellwig wrote:
> > > On Fri, Sep 04, 2009 at 08:53:57AM +0200, Jens Axboe wrote:
> > > > > +	if (wbc->sync_mode == WB_SYNC_ALL)
> > > > > +		bdi_wait_on_work_clear(&work);
> > > > >  }
> > > > 
> > > > That doesn't work, you have to wait for on-stack work. So either we just
> > > > punt and not do anything for WB_SYNC_NONE if the allocation fails, or we
> > > > punt to stack and do the wait. Since it's a cleaning action and
> > > > allocation fails, falling back to the stack and waiting seems like the
> > > > most appropriate choice.
> > > 
> > > True, the wait needs to be unconditional.  Updated version below.
> > 
> > (did you forget that patch? it's not there).
> 
> Here we go, sorry:

I have applied this to a postmerge writeback branch. I made one change,
though:

> +	if (wbc->sync_mode != WB_SYNC_ALL) {
> +		struct bdi_work *w = bdi_alloc_work(wbc);
> +		if (w) {
> +			bdi_queue_work(wbc->bdi, w);
> +			return;
> +		}

We should make that bdi_queue_work() unconditional, if you want to make
sure that we current thread wakes up and actually flushes some old data
when allocation fails.

void bdi_start_writeback(struct writeback_control *wbc)
{
        /*
         * WB_SYNC_NONE is opportunistic writeback. If this allocation fails,
         * bdi_queue_work() will wake up the thread and flush old data. This
         * should ensure some amount of progress in freeing memory.
         */
        if (wbc->sync_mode != WB_SYNC_ALL) {
                struct bdi_work *w = bdi_alloc_work(wbc);

                bdi_queue_work(wbc->bdi, w);
        } else {
                struct bdi_work work;

                bdi_work_init(&work, wbc);
                work.state |= WS_ONSTACK;

                bdi_queue_work(wbc->bdi, &work);
                bdi_wait_on_work_clear(&work);
        }
}

-- 
Jens Axboe