Re: [PATCH v17 4/9] mm: Introduce Reported pages

[Mel Gorman]

On Fri, Feb 21, 2020 at 11:25:49AM -0800, Alexander Duyck wrote:
> On Thu, 2020-02-20 at 22:35 +0000, Mel Gorman wrote:
> > On Thu, Feb 20, 2020 at 10:44:21AM -0800, Alexander Duyck wrote:
> > > > > +static int
> > > > > +page_reporting_cycle(struct page_reporting_dev_info *prdev, struct zone *zone,
> > > > > +		     unsigned int order, unsigned int mt,
> > > > > +		     struct scatterlist *sgl, unsigned int *offset)
> > > > > +{
> > > > > +	struct free_area *area = &zone->free_area[order];
> > > > > +	struct list_head *list = &area->free_list[mt];
> > > > > +	unsigned int page_len = PAGE_SIZE << order;
> > > > > +	struct page *page, *next;
> > > > > +	int err = 0;
> > > > > +
> > > > > +	/*
> > > > > +	 * Perform early check, if free area is empty there is
> > > > > +	 * nothing to process so we can skip this free_list.
> > > > > +	 */
> > > > > +	if (list_empty(list))
> > > > > +		return err;
> > > > > +
> > > > > +	spin_lock_irq(&zone->lock);
> > > > > +
> > > > > +	/* loop through free list adding unreported pages to sg list */
> > > > > +	list_for_each_entry_safe(page, next, list, lru) {
> > > > > +		/* We are going to skip over the reported pages. */
> > > > > +		if (PageReported(page))
> > > > > +			continue;
> > > > > +
> > > > > +		/* Attempt to pull page from list */
> > > > > +		if (!__isolate_free_page(page, order))
> > > > > +			break;
> > > > > +
> > > > 
> > > > Might want to note that you are breaking because the only reason to fail
> > > > the isolation is that watermarks are not met and we are likely under
> > > > memory pressure. It's not a big issue.
> > > > 
> > > > However, while I think this is correct, it's hard to follow. This loop can
> > > > be broken out of with pages still on the scatter gather list. The current
> > > > flow guarantees that err will not be set at this point so the caller
> > > > cleans it up so we always drain the list either here or in the caller.
> > > 
> > > I can probably submit a follow-up patch to update the comments. The reason
> > > for not returning an error is because I didn't consider it an error that
> > > we encountered the watermark and were not able to pull any more pages.
> > > Instead I considered that the "stop" point for this pass and have it just
> > > exit out of the loop and flush the data.
> > > 
> > 
> > I don't consider it an error and I don't think you should return an
> > error. The comment just needs to explain that the draining happens in
> > the caller in this case. That should be enough of a warning to a future
> > developer to double check the flow after any changes to make sure the
> > drain is reached.
> 
> The comment I can do, that shouldn't be an issue. The point I was getting
> at is that a separate drain call is expected for this any time the
> function is not returning an error, and the only way it can return an
> error is if there was a reporting issue.
> 

I'm not suggesting you return an error. I'm suggesting you put a warn in
before you break due to watermarks *if* there is an error. It should
*never* trigger unless someone modifies the flow and breaks it in which
case the warning will not kill the system but give a strong hint to the
developer that they need to think a bit more.

It's ok to leave it out because at this point, it's a distraction and I
do not see a problem with the current code.

-- 
Mel Gorman
SUSE Labs