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

[Alex Shi]

在 2020/4/14 上午2:07, Johannes Weiner 写道:
> On Mon, Apr 13, 2020 at 06:48:22PM +0800, Alex Shi wrote:
>>> 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.
>>>
>>> It doesn't matter how many times you check the lruvec before and after
>>> locking - if you're looking at a free page, it might get allocated,
>>> charged and put on a new lruvec after you're done checking, and then
>>> you isolate a page from an unlocked lruvec.
>>>
>>> You simply cannot serialize on page->mem_cgroup->lruvec when
>>> page->mem_cgroup isn't stable. You need to serialize on the page
>>> itself, one way or another, to make this work.
>>>
>>>
>>> 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?
>>>
>>> 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?
>>>
>>
>> Hi Johannes
>>
>> As to the idea of TestClearPageLRU, we except the following scenario
>>     compaction                       commit_charge
>>                                      if (TestClearPageLRU)
>>         !TestClearPageLRU                 lock_page_lruvec
>>             goto isolate_fail;            del_from_lru_list
>>                                           unlock_page_lruvec
>>
>> But there is a difficult situation to handle:
>>
>>    compaction                        commit_charge
>>         TestClearPageLRU
>>                                     !TestClearPageLRU
>>
>>                                     page possible state:
>>                                     a, reclaiming, b, moving between lru list, c, migrating, like in compaction
>>                                     d, mlocking,   e, split_huge_page,
>>
>> If the page lru bit was cleared in commit_charge with lrucare,
>> we still have no idea if the page was isolated by the reason from a~e
>> or the page is never on LRU, to deal with different reasons is high cost.
>>
>> So as to the above issue you mentioned, Maybe the better idea is to
>> set lrucare when do mem_cgroup_commit_charge(), since the charge action
>> is not often. What's your idea of this solution?
> 
> Hm, yes, the lrucare scenario is a real problem. If it can isolate the
> page, fine, but if not, it changes page->mem_cgroup on a page that
> somebody else has isolated, having indeed no idea who they are and how
> they are going to access page->mem_cgroup.
> 
> Right now it's safe because of secondary protection on top of
> isolation: split_huge_page keeps the lru_lock held throughout;
> reclaim, cgroup migration, page migration, compaction etc. hold the
> page lock which locks out swapcache charging.
> 
> But it complicates the serialization model immensely and makes it
> subtle and error prone.
> 
> I'm not sure how unconditionally taking the lru_lock when charging
> would help. Can you lay out what you have in mind in prototype code,
> like I'm using below, for isolation, putback, charging, compaction?

The situation would back to relock scheme, the lru_lock will compete with 
the some root_memcg->lru_lock in practical. So no needs to distinguish 
putback, compaction etc. 

But I don't know how much impact on this alloc path...

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()
					    spin_lock_irq(page->memcg->lruvec->lru_lock)
 					    page->mem_cgroup = foo
					    spin_unlock_irq(page->memcg->lruvec->lru_lock)
 					  lru_cache_add()
 					    __pagevec_lru_add()
 					      SetPageLRU()

 if PageLRU(page):
   __isolate_lru_page()

> 
> That said, charging actually is a hotpath. I'm reluctant to
> unconditionally take the LRU lock there. But if you can make things a
> lot simpler this way, it could be worth exploring.
> 
> In the PageLRU locking scheme, I can see a way to make putback safe
> wrt lrucare charging, but I'm not sure about isolation:
> 
> putback:
> lruvec = page->mem_cgroup->lruvecs[pgdat]
> spin_lock(lruvec->lru_lock)
> if lruvec != page->mem_cgroup->lruvecs[pgdat]:
>   /*
>    * commit_charge(lrucare=true) can charge an uncharged swapcache
>    * page while we had it isolated. This changes page->mem_cgroup,
>    * but it can only happen once. Look up the new cgroup.
>    */
>   spin_unlock(lruvec->lru_lock)
>   lruvec = page->mem_cgroup->lruvecs[pgdat]
>   spin_lock(lruvec->lru_lock)
> add_page_to_lru_list(page, lruvec, ...)
> SetPageLRU(page);
> spin_unlock(lruvec->lru_lock)
> 
> commit_charge:
> if (lrucare)
>   spin_lock(root_memcg->lru_lock)
>   /*
>    * If we can isolate the page, we'll move it to the new
>    * cgroup's LRU list. If somebody else has the page
>    * isolated, we need their putback to move it to the
>    * new cgroup. If they see the old cgroup - the root -
>    * they will spin until we're done and recheck.
>    */
>   if ((lru = TestClearPageLRU(page)))
>     del_page_from_lru_list()
> page->mem_cgroup = new;
> if (lrucare)
>   spin_unlock(root_memcg->lru_lock)
>   if (lru)
>     spin_lock(new->lru_lock)
>     add_page_to_lru_list()
>     spin_unlock(new->lru_lock);
>     SetPageLRU(page)
> 
> putback would need to 1) recheck once after acquiring the lock and 2)
> SetPageLRU while holding the lru_lock after all. But it works because
> we know the old cgroup: if the putback sees the old cgroup, we know
> it's the root cgroup, and we have that locked until we're done with
> the update. And if putback manages to lock the old cgroup before us,
> we will spin until the isolator is done, and then either be able to
> isolate it ourselves or, if racing with yet another isolator, hold the
> lock and delay putback until we're done.
> 
> But isolation actually needs to lock out charging, or it would operate
> on the wrong list:
> 
> isolation:                                     commit_charge:
> if (TestClearPageLRU(page))
>                                                page->mem_cgroup = new
>   // page is still physically on
>   // the root_mem_cgroup's LRU. We're
>   // updating the wrong list:
>   memcg = page->mem_cgroup
>   spin_lock(memcg->lru_lock)
>   del_page_from_lru_list(page, memcg)
>   spin_unlock(memcg->lru_lock)
> 

Yes, this is the problem I encountered now for mem_cgroup_lru_size incorrect.

 
> lrucare really is a mess. Even before this patch series, it makes
> things tricky and subtle and error prone.
> 
> The only reason we're doing it is for when there is swapping without
> swap tracking, in which case swap reahadead needs to put pages on the
> LRU but cannot charge them until we have a faulting vma later.
> 
> But it's not clear how practical such a configuration is. Both memory
> and swap are shared resources, and isolation isn't really effective
> when you restrict access to memory but then let workloads swap freely.

Yes, we didn't figure a good usage on MEMCG_SWAP too. And if swaping happens
often, the different memcg's memory were swaped to same disk and mixed together
which cause readahead useless.

> 
> Plus, the overhead of tracking is tiny - 512k per G of swap (0.04%).
> 
> Maybe we should just delete MEMCG_SWAP and unconditionally track swap
> entry ownership when the memory controller is enabled. I don't see a
> good reason not to, and it would simplify the entire swapin path, the
> LRU locking, and the page->mem_cgroup stabilization rules.
> 

Sorry for not follow you up, did you mean just remove the MEMCG_SWAP configuration
and keep the feature in default memcg? 
That does can remove lrucare, but PageLRU lock scheme still fails since
we can not isolate the page during commit_charge, is that right?

Thanks a lot!
Alex