Re: [PATCH v13 4/5] mm: support reporting free page blocks

From: Michal Hocko
Date: Thu Aug 03 2017 - 09:50:59 EST


On Thu 03-08-17 21:17:25, Wei Wang wrote:
> On 08/03/2017 08:41 PM, Michal Hocko wrote:
> >On Thu 03-08-17 20:11:58, Wei Wang wrote:
> >>On 08/03/2017 07:28 PM, Michal Hocko wrote:
> >>>On Thu 03-08-17 19:27:19, Wei Wang wrote:
> >>>>On 08/03/2017 06:44 PM, Michal Hocko wrote:
> >>>>>On Thu 03-08-17 18:42:15, Wei Wang wrote:
> >>>>>>On 08/03/2017 05:11 PM, Michal Hocko wrote:
> >>>>>>>On Thu 03-08-17 14:38:18, Wei Wang wrote:
> >>>>>[...]
> >>>>>>>>+static int report_free_page_block(struct zone *zone, unsigned int order,
> >>>>>>>>+ unsigned int migratetype, struct page **page)
> >>>>>>>This is just too ugly and wrong actually. Never provide struct page
> >>>>>>>pointers outside of the zone->lock. What I've had in mind was to simply
> >>>>>>>walk free lists of the suitable order and call the callback for each one.
> >>>>>>>Something as simple as
> >>>>>>>
> >>>>>>> for (i = 0; i < MAX_NR_ZONES; i++) {
> >>>>>>> struct zone *zone = &pgdat->node_zones[i];
> >>>>>>>
> >>>>>>> if (!populated_zone(zone))
> >>>>>>> continue;
> >>>>>>> spin_lock_irqsave(&zone->lock, flags);
> >>>>>>> for (order = min_order; order < MAX_ORDER; ++order) {
> >>>>>>> struct free_area *free_area = &zone->free_area[order];
> >>>>>>> enum migratetype mt;
> >>>>>>> struct page *page;
> >>>>>>>
> >>>>>>> if (!free_area->nr_pages)
> >>>>>>> continue;
> >>>>>>>
> >>>>>>> for_each_migratetype_order(order, mt) {
> >>>>>>> list_for_each_entry(page,
> >>>>>>> &free_area->free_list[mt], lru) {
> >>>>>>>
> >>>>>>> pfn = page_to_pfn(page);
> >>>>>>> visit(opaque2, prn, 1<<order);
> >>>>>>> }
> >>>>>>> }
> >>>>>>> }
> >>>>>>>
> >>>>>>> spin_unlock_irqrestore(&zone->lock, flags);
> >>>>>>> }
> >>>>>>>
> >>>>>>>[...]
> >>>>>>I think the above would take the lock for too long time. That's why we
> >>>>>>prefer to take one free page block each time, and taking it one by one
> >>>>>>also doesn't make a difference, in terms of the performance that we
> >>>>>>need.
> >>>>>I think you should start with simple approach and impove incrementally
> >>>>>if this turns out to be not optimal. I really detest taking struct pages
> >>>>>outside of the lock. You never know what might happen after the lock is
> >>>>>dropped. E.g. can you race with the memory hotremove?
> >>>>The caller won't use pages returned from the function, so I think there
> >>>>shouldn't be an issue or race if the returned pages are used (i.e. not free
> >>>>anymore) or simply gone due to hotremove.
> >>>No, this is just too error prone. Consider that struct page pointer
> >>>itself could get invalid in the meantime. Please always keep robustness
> >>>in mind first. Optimizations are nice but it is even not clear whether
> >>>the simple variant will cause any problems.
> >>
> >>how about this:
> >>
> >>for_each_populated_zone(zone) {
> >> for_each_migratetype_order_decend(min_order, order, type) {
> >> do {
> >> => spin_lock_irqsave(&zone->lock, flags);
> >> ret = report_free_page_block(zone, order, type,
> >> &page)) {
> >> pfn = page_to_pfn(page);
> >> nr_pages = 1 << order;
> >> visit(opaque1, pfn, nr_pages);
> >> }
> >> => spin_unlock_irqrestore(&zone->lock, flags);
> >> } while (!ret)
> >>}
> >>
> >>In this way, we can still keep the lock granularity at one free page block
> >>while having the struct page operated under the lock.
> >How can you continue iteration of free_list after the lock has been
> >dropped?
>
> report_free_page_block() has handled all the possible cases after the lock
> is
> dropped. For example, if the previous reported page has not been on the free
> list, then the first node from the list of this order will be given. This is
> because
> page allocation takes page blocks from the head to end, for example:
>
> 1,2,3,4,5,6
> if the previous reported free block is 2, when we give 2 to the report
> function
> to get the next page block, and find 1,2,3 have all gone, it will report 4,
> which
> is the head of the free list.

As I've said earlier. Start simple optimize incrementally with some
numbers to justify a more subtle code.
--
Michal Hocko
SUSE Labs