Re: [PATCH v8 03/10] mm/lru: replace pgdat lru_lock with lruvec lock

[Alex Shi]

> In a previous review, I pointed out the following race condition
> between page charging and compaction:
> 
> compaction:				generic_file_buffered_read:
> 
> 					page_cache_alloc()
> 
> !PageBuddy()
> 
> lock_page_lruvec(page)
>   lruvec = mem_cgroup_page_lruvec()
>   spin_lock(&lruvec->lru_lock)
>   if lruvec != mem_cgroup_page_lruvec()
>     goto again
> 
> 					add_to_page_cache_lru()
> 					  mem_cgroup_commit_charge()
> 					    page->mem_cgroup = foo
> 					  lru_cache_add()
> 					    __pagevec_lru_add()
> 					      SetPageLRU()
> 
> if PageLRU(page):
>   __isolate_lru_page()
> 
> As far as I can see, you have not addressed this. You have added
> lock_page_memcg(), but that prevents charged pages from moving between
> cgroups, it does not prevent newly allocated pages from being charged.
> 

yes, it's my fault to oversee this problem.

...

> 
> So here is a crazy idea that may be worth exploring:
> 
> Right now, pgdat->lru_lock protects both PageLRU *and* the lruvec's
> linked list.
> 
> Can we make PageLRU atomic and use it to stabilize the lru_lock
> instead, and then use the lru_lock only serialize list operations?
> 

Sounds a good idea. I will try this.

Thanks
Alex

> I.e. in compaction, you'd do
> 
> 	if (!TestClearPageLRU(page))
> 		goto isolate_fail;
> 	/*
> 	 * We isolated the page's LRU state and thereby locked out all
> 	 * other isolators, including cgroup page moving, page reclaim,
> 	 * page freeing etc. That means page->mem_cgroup is now stable
> 	 * and we can safely look up the correct lruvec and take the
> 	 * page off its physical LRU list.
> 	 */
> 	lruvec = mem_cgroup_page_lruvec(page);
> 	spin_lock_irq(&lruvec->lru_lock);
> 	del_page_from_lru_list(page, lruvec, page_lru(page));
> 
> Putback would mostly remain the same (although you could take the
> PageLRU setting out of the list update locked section, as long as it's
> set after the page is physically linked):
> 
> 	/* LRU isolation pins page->mem_cgroup */
> 	lruvec = mem_cgroup_page_lruvec(page)
> 	spin_lock_irq(&lruvec->lru_lock);
> 	add_page_to_lru_list(...);
> 	spin_unlock_irq(&lruvec->lru_lock);
> 
> 	SetPageLRU(page);
> 
> And you'd have to carefully review and rework other sites that rely on
> PageLRU: reclaim, __page_cache_release(), __activate_page() etc.
> 
> Especially things like activate_page(), which used to only check
> PageLRU to shuffle the page on the LRU list would now have to briefly
> clear PageLRU and then set it again afterwards.
> 
> However, aside from a bit more churn in those cases, and the
> unfortunate additional atomic operations, I currently can't think of a
> fundamental reason why this wouldn't work.
> 
> Hugh, what do you think?
>