Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Thu, Sep 17, 2020 at 08:39:34AM -0700, Alexander Duyck wrote: > On Thu, Sep 17, 2020 at 7:26 AM Daniel Jordan > wrote: > > > > On Thu, Sep 17, 2020 at 10:37:45AM +0800, Alex Shi wrote: > > > 在 2020/9/16 上午12:58, Daniel Jordan 写道: > > > > On Tue, Sep 15, 2020 at 01:21:56AM -0700, Hugh Dickins wrote: > > > >> On Sun, 13 Sep 2020, Alex Shi wrote: > > > >>> Uh, I updated the testing with some new results here: > > > >>> https://lkml.org/lkml/2020/8/26/212 > > > >> Right, I missed that, that's better, thanks. Any other test results? > > > > Alex, you were doing some will-it-scale runs earlier. Are you planning > > > > to do > > > > more of those? Otherwise I can add them in. > > > > > > Hi Daniel, > > > > > > Does compaction perf scalable, like thpscale, I except they could get > > > some benefit. > > > > Yep, I plan to stress compaction. Reclaim as well. > > > > I should have said which Alex I meant. I was asking Alex Duyck since he'd > > done > > some will-it-scale runs. > > I probably won't be able to do any will-it-scale runs any time soon. > If I recall I ran them for this latest v18 patch set and didn't see > any regressions like I did with the previous set. However the system I > was using is tied up for other purposes and it may be awhile before I > can free it up to look into this again. Ok, sure. I hadn't seen the regressions were taken case of, that's good to hear. Might still add them to my testing for v19 and beyond, we'll see.
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Thu, Sep 17, 2020 at 7:26 AM Daniel Jordan wrote: > > On Thu, Sep 17, 2020 at 10:37:45AM +0800, Alex Shi wrote: > > 在 2020/9/16 上午12:58, Daniel Jordan 写道: > > > On Tue, Sep 15, 2020 at 01:21:56AM -0700, Hugh Dickins wrote: > > >> On Sun, 13 Sep 2020, Alex Shi wrote: > > >>> Uh, I updated the testing with some new results here: > > >>> https://lkml.org/lkml/2020/8/26/212 > > >> Right, I missed that, that's better, thanks. Any other test results? > > > Alex, you were doing some will-it-scale runs earlier. Are you planning > > > to do > > > more of those? Otherwise I can add them in. > > > > Hi Daniel, > > > > Does compaction perf scalable, like thpscale, I except they could get some > > benefit. > > Yep, I plan to stress compaction. Reclaim as well. > > I should have said which Alex I meant. I was asking Alex Duyck since he'd > done > some will-it-scale runs. I probably won't be able to do any will-it-scale runs any time soon. If I recall I ran them for this latest v18 patch set and didn't see any regressions like I did with the previous set. However the system I was using is tied up for other purposes and it may be awhile before I can free it up to look into this again. Thanks. - Alex
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Thu, Sep 17, 2020 at 10:37:45AM +0800, Alex Shi wrote: > 在 2020/9/16 上午12:58, Daniel Jordan 写道: > > On Tue, Sep 15, 2020 at 01:21:56AM -0700, Hugh Dickins wrote: > >> On Sun, 13 Sep 2020, Alex Shi wrote: > >>> Uh, I updated the testing with some new results here: > >>> https://lkml.org/lkml/2020/8/26/212 > >> Right, I missed that, that's better, thanks. Any other test results? > > Alex, you were doing some will-it-scale runs earlier. Are you planning to > > do > > more of those? Otherwise I can add them in. > > Hi Daniel, > > Does compaction perf scalable, like thpscale, I except they could get some > benefit. Yep, I plan to stress compaction. Reclaim as well. I should have said which Alex I meant. I was asking Alex Duyck since he'd done some will-it-scale runs.
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
在 2020/9/16 上午12:58, Daniel Jordan 写道: > On Tue, Sep 15, 2020 at 01:21:56AM -0700, Hugh Dickins wrote: >> On Sun, 13 Sep 2020, Alex Shi wrote: >>> Uh, I updated the testing with some new results here: >>> https://lkml.org/lkml/2020/8/26/212 >> Right, I missed that, that's better, thanks. Any other test results? > Alex, you were doing some will-it-scale runs earlier. Are you planning to do > more of those? Otherwise I can add them in. Hi Daniel, Does compaction perf scalable, like thpscale, I except they could get some benefit. Thanks Alex
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Tue, Sep 15, 2020 at 01:21:56AM -0700, Hugh Dickins wrote: > On Sun, 13 Sep 2020, Alex Shi wrote: > > Uh, I updated the testing with some new results here: > > https://lkml.org/lkml/2020/8/26/212 > > Right, I missed that, that's better, thanks. Any other test results? Alex, you were doing some will-it-scale runs earlier. Are you planning to do more of those? Otherwise I can add them in. This is what I have so far. sysbench oltp read-only --- The goal was to run a real world benchmark, at least more so than something like vm-scalability, with the memory controller enabled but unused to check for regressions. I chose sysbench because it was relatively straightforward to run, but I'm open to ideas for other high level benchmarks that might be more sensitive to this series. CoeffVar shows the test was pretty noisy overall. It's nice to see there's no significant difference between the kernels for low thread counts (1-12), but I'm not sure what to make of the 18 and 20 thread cases. At 20 threads, the CPUs of the node that the test was confined to were saturated and the variance is especially high. I'm tempted to write the 18 and 20 thread cases off as noise. - 2-socket * 10-core * 2-hyperthread broadwell server - test bound to node 1 to lower variance - 251G memory, divided evenly between the nodes (memory size of test shrunk to accommodate confining to one node) - 12 iterations per thread count per kernel - THP enabled export OLTP_CACHESIZE=$(($MEMTOTAL_BYTES/4)) export OLTP_SHAREDBUFFERS=$((MEMTOTAL_BYTES/8)) export OLTP_PAGESIZES="default" export SYSBENCH_DRIVER=postgres export SYSBENCH_MAX_TRANSACTIONS=auto export SYSBENCH_READONLY=yes export SYSBENCH_MAX_THREADS=$((NUMCPUS / 2)) export SYSBENCH_ITERATIONS=12 export SYSBENCH_WORKLOAD_SIZE=$((MEMTOTAL_BYTES*3/8)) export SYSBENCH_CACHE_COLD=no export DATABASE_INIT_ONCE=yes export MMTESTS_NUMA_POLICY=fullbind_single_instance_node numactl --cpunodebind=1 --membind=1 sysbench Transactions per second 5.9-rc25.9-rc2-lru-v18 Min 1 593.23 ( 0.00%) 583.37 ( -1.66%) Min 4 1897.34 ( 0.00%) 1871.77 ( -1.35%) Min 7 2471.14 ( 0.00%) 2449.77 ( -0.86%) Min 12 2680.00 ( 0.00%) 2853.25 ( 6.46%) Min 18 2183.82 ( 0.00%) 1191.43 ( -45.44%) Min 20 924.96 ( 0.00%) 526.66 ( -43.06%) Hmean 1 912.08 ( 0.00%) 904.24 ( -0.86%) Hmean 4 2057.11 ( 0.00%) 2044.69 ( -0.60%) Hmean 7 2817.59 ( 0.00%) 2812.80 ( -0.17%) Hmean 12 3201.05 ( 0.00%) 3171.09 ( -0.94%) Hmean 18 2529.10 ( 0.00%) 2009.99 * -20.53%* Hmean 20 1742.29 ( 0.00%) 1127.77 * -35.27%* Stddev1 219.21 ( 0.00%) 220.92 ( -0.78%) Stddev494.94 ( 0.00%) 84.34 ( 11.17%) Stddev7 189.42 ( 0.00%) 167.58 ( 11.53%) Stddev12 372.13 ( 0.00%) 199.40 ( 46.42%) Stddev18 248.42 ( 0.00%) 574.66 (-131.32%) Stddev20 757.69 ( 0.00%) 666.87 ( 11.99%) CoeffVar 122.54 ( 0.00%) 22.86 ( -1.42%) CoeffVar 4 4.61 ( 0.00%)4.12 ( 10.60%) CoeffVar 7 6.69 ( 0.00%)5.94 ( 11.30%) CoeffVar 12 11.49 ( 0.00%)6.27 ( 45.46%) CoeffVar 189.74 ( 0.00%) 26.22 (-169.23%) CoeffVar 20 36.32 ( 0.00%) 47.18 ( -29.89%) Max 1 1117.45 ( 0.00%) 1107.33 ( -0.91%) Max 4 2184.92 ( 0.00%) 2136.65 ( -2.21%) Max 7 3086.81 ( 0.00%) 3049.52 ( -1.21%) Max 12 4020.07 ( 0.00%) 3580.95 ( -10.92%) Max 18 3032.30 ( 0.00%) 2810.85 ( -7.30%) Max 20 2891.27 ( 0.00%) 2675.80 ( -7.45%) BHmean-50 1 1098.77 ( 0.00%) 1093.58 ( -0.47%) BHmean-50 4 2139.76 ( 0.00%) 2107.13 ( -1.52%) BHmean-50 7 2972.18 ( 0.00%) 2953.94 ( -0.61%) BHmean-50 12 3494.73 ( 0.00%) 3311.33 ( -5.25%) BHmean-50 18 2729.70 ( 0.00%) 2606.32 ( -4.52%) BHmean-50 20 2668.72 ( 0.00%) 1779.87 ( -33.31%) BHmean-95 1 958.94 ( 0.00%) 951.84 ( -0.74%) BHmean-95 4 2072.98 ( 0.00%) 2062.01 ( -0.53%) BHmean-95 7 2853.96 ( 0.00%) 2851.21 ( -0.10%) BHmean-95 12 3258.65 ( 0.00%) 3203.53 ( -1.69%) BHmean-95 18 2565.99 ( 0.00%) 2143.90 ( -16.45%) BHmean-95 20 1894.47 ( 0.00%) 1258.34 ( -33.58%) BHmean-99 1 958.94 ( 0.00%) 951.84 ( -0.74%) BHmean-99 4 2072.98 ( 0.00%) 2062.01 ( -0.53%) BHmean-99 7 2853.96 ( 0.00%) 2851.21 ( -0.10%) BHmean-99 12 3258.65 ( 0.00%) 3203.53 ( -1.69%) BHmean-99 18 2565.99 ( 0.00%) 2143.90 ( -16.45%) BHmean-99 20 1894.47 ( 0.00%) 1258.34 ( -33.58%) sysbench Time 5.9-rc25.9-rc2-lru Min
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Sun, 13 Sep 2020, Alex Shi wrote: > > IIRC, all of comments are accepted and push to > https://github.com/alexshi/linux.git lruv19 I just had to relax for the weekend, so no progress from me. I'll take a look at your tree tomorrow, er, later today. > If you don't minder, could you change everything and send out a new version > for further review? Sorry, no. Tiresome though it is for both of us, I'll continue to send you comments, and leave all the posting to you. > Uh, I updated the testing with some new results here: > https://lkml.org/lkml/2020/8/26/212 Right, I missed that, that's better, thanks. Any other test results? Hugh
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
在 2020/9/12 上午10:13, Hugh Dickins 写道: > On Fri, 11 Sep 2020, Alex Shi wrote: >> 在 2020/9/10 上午7:16, Hugh Dickins 写道: >>> On Wed, 9 Sep 2020, Alex Shi wrote: 在 2020/9/9 上午7:41, Hugh Dickins 写道: > > [PATCH v18 05/32] mm/thp: remove code path which never got into > This is a good simplification, but I see no sign that you understand > why it's valid: it relies on lru_add_page_tail() being called while > head refcount is frozen to 0: we would not get this far if someone > else holds a reference to the THP - which they must hold if they have > isolated the page from its lru (and that's true before or after your > per-memcg changes - but even truer after those changes, since PageLRU > can then be flipped without lru_lock at any instant): please explain > something of this in the commit message. Is the following commit log better? split_huge_page() will never call on a page which isn't on lru list, so this code never got a chance to run, and should not be run, to add tail pages on a lru list which head page isn't there. Hugh Dickins' mentioned: The path should never be called since lru_add_page_tail() being called while head refcount is frozen to 0: we would not get this far if someone else holds a reference to the THP - which they must hold if they have isolated the page from its lru. Although the bug was never triggered, it'better be removed for code correctness, and add a warn for unexpected calling. >>> >>> Not much better, no. split_huge_page() can easily be called for a page >>> which is not on the lru list at the time, >> >> Hi Hugh, >> >> Thanks for comments! >> >> There are some discussion on this point a couple of weeks ago, >> https://lkml.org/lkml/2020/7/9/760 >> >> Matthew Wilcox and Kirill have the following comments, >>> I don't understand how we get to split_huge_page() with a page that's >>> not on an LRU list. Both anonymous and page cache pages should be on >>> an LRU list. What am I missing? >> >> Right, and it's never got removed from LRU during the split. The tail >> pages have to be added to LRU because they now separate from the tail >> page. >> >> -- >> Kirill A. Shutemov > > Yes, those were among the mails that I read through before getting > down to review. I was surprised by their not understanding, but > it was a bit late to reply to that thread. > > Perhaps everybody had been focused on pages which have been and > naturally belong on an LRU list, rather than pages which are on > the LRU list at the instant that split_huge_page() is called. > > There are a number of places where PageLRU gets cleared, and a > number of places where we del_page_from_lru_list(), I think you'll > agree: your patches touch all or most of them. Let's think of a > common one, isolate_lru_pages() used by page reclaim, but the same > would apply to most of the others. > > Then there a number of places where split_huge_page() is called: > I am having difficulty finding any of those which cannot race with > page reclaim, but shall we choose anon THP's deferred_split_scan(), > or shmem THP's shmem_punch_compound()? > > What prevents either of those from calling split_huge_page() at > a time when isolate_lru_pages() has removed the page from LRU? > > But there's no problem in this race, because anyone isolating the > page from LRU must hold their own reference to the page (to prevent > it from being freed independently), and the can_split_huge_page() or > page_ref_freeze() in split_huge_page_to_list() will detect that and > fail the split with -EBUSY (or else succeed and prevent new references > from being acquired). So this case never reaches lru_add_page_tail(). Hi Hugh, Thanks for comments! We are the same page here, we all know split_huge_page_to_list could block them go futher and the code is functionality right. If the comments 'Split start from PageLRU(head), and ...' doesn't make things clear as it's should be, I am glad to see you rewrite and improve them. > >> >>> and I don't know what was the >>> bug which was never triggered. >> >> So the only path to the removed part should be a bug, like sth here, >> https://lkml.org/lkml/2020/7/10/118 >> or >> https://lkml.org/lkml/2020/7/10/972 > > Oh, the use of split_huge_page() in __iommu_dma_alloc_pages() is just > nonsense, I thought it had already been removed - perhaps some debate > over __GFP_COMP held it up. Not something you need worry about in > this patchset. > >> >>> Stick with whatever text you end up with >>> for the combination of 05/32 and 18/32, and I'll rewrite it after. >> >> I am not object to merge them into one, I just don't know how to say >> clear about 2 patches in commit log. As patch 18, TestClearPageLRU >> add the incorrect posibility of remove lru bit during split, that's >> the reason of code path rewrite and a WARN there. > > I did not
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
在 2020/9/12 下午4:38, Hugh Dickins 写道:
> On Wed, 9 Sep 2020, Alex Shi wrote:
>> 在 2020/9/9 上午7:41, Hugh Dickins 写道:
>>>
>>> The use of lock_page_memcg() in __munlock_pagevec() in 20/32,
>>> introduced in patchset v17, looks good but it isn't: I was lucky that
>>> systemd at reboot did some munlocking that exposed the problem to lockdep.
>>> The first time into the loop, lock_page_memcg() is done before lru_lock
>>> (as 06/32 has allowed); but the second time around the loop, it is done
>>> while still holding lru_lock.
>>
>> I don't know the details of lockdep show. Just wondering could it possible
>> to solid the move_lock/lru_lock sequence?
>> or try other blocking way which mentioned in commit_charge()?
>>
>>>
>>> lock_page_memcg() really needs to be absorbed into (a variant of)
>>> relock_page_lruvec(), and I do have that (it's awkward because of
>>> the different ways in which the IRQ flags are handled). And out of
>>> curiosity, I've also tried using that in mm/swap.c too, instead of the
>>> TestClearPageLRU technique: lockdep is happy, but an update_lru_size()
>>> warning showed that it cannot safely be mixed with the TestClearPageLRU
>>> technique (that I'd left in isolate_lru_page()). So I'll stash away
>>> that relock_page_lruvec(), and consider what's best for mm/mlock.c:
>>> now that I've posted these comments so far, that's my priority, then
>>> to get the result under testing again, before resuming these comments.
>>
>> No idea of your solution, but looking forward for your good news! :)
>
> Yes, it is good news, and simpler than anything suggested above.
Awesome!
>
> The main difficulties will probably be to look good in the 80 columns
> (I know that limit has been lifted recently, but some of us use xterms
> side by side), and to explain it.
>
> mm/mlock.c has not been kept up-to-date very well: and in particular,
> you have taken too seriously that "Serialize with any parallel
> __split_huge_page_refcount()..." comment that you updated to two
> comments "Serialize split tail pages in __split_huge_page_tail()...".
>
> Delete them! The original comment was by Vlastimil for v3.14 in 2014.
> But Kirill redesigned THP refcounting for v4.5 in 2016: that's when
> __split_huge_page_refcount() went away. And with the new refcounting,
> the THP splitting races that lru_lock protected munlock_vma_page()
> and __munlock_pagevec() from: those races have become impossible.
>
> Or maybe there never was such a race in __munlock_pagevec(): you
> have added the comment there, assuming lru_lock was for that purpose,
> but that was probably just the convenient place to take it,
> to cover all the del_page_from_lru()s.
>
> Observe how split_huge_page_to_list() uses unmap_page() to remove
> all pmds and all ptes for the huge page being split, and remap_page()
> only replaces the migration entries (used for anon but not for shmem
> or file) after doing all of the __split_huge_page_tail()s, before
> unlocking any of the pages. Recall that munlock_vma_page() and
> __munlock_pagevec() are being applied to pages found mapped
> into userspace, by ptes or pmd: there are none of those while
> __split_huge_page_tail() is being used, so no race to protect from.
>
> (Could a newly detached tail be freshly faulted into userspace just
> before __split_huge_page() has reached the head? Not quite, the
> fault has to wait to get the tail's page lock. But even if it
> could, how would that be a problem for __munlock_pagevec()?)
>
> There's lots more that could be said: for example, PageMlocked will
> always be clear on the THP head during __split_huge_page_tail(),
> because the last unmap of a PageMlocked page does clear_page_mlock().
> But that's not required to prove the case, it's just another argument
> against the "Serialize" comment you have in __munlock_pagevec().
>
> So, no need for the problematic lock_page_memcg(page) there in
> __munlock_pagevec(), nor to lock (or relock) lruvec just below it.
> __munlock_pagevec() still needs lru_lock to del_page_from_lru_list(),
> of course, but that must be done after your TestClearPageMlocked has
> stabilized page->memcg. Use relock_page_lruvec_irq() here? I suppose
> that will be easiest, but notice how __munlock_pagevec_fill() has
> already made sure that all the pages in the pagevec are from the same
> zone (and it cannot do the same for memcg without locking page memcg);
> so some of relock's work will be redundant.
It sounds reasonable for me.
>
> Otherwise, I'm much happier with your mm/mlock.c since looking at it
> in more detail: a couple of nits though - drop the clear_page_mlock()
> hunk from 25/32 - kernel style says do it the way you are undoing by
> - if (!isolate_lru_page(page)) {
> + if (!isolate_lru_page(page))
> putback_lru_page(page);
> - } else {
> + else {
> I don't always follow that over-braced style when making changes,
> but you should not touch otherwise untouched code just to make it
> go against the
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Wed, 9 Sep 2020, Alex Shi wrote:
> 在 2020/9/9 上午7:41, Hugh Dickins 写道:
> >
> > The use of lock_page_memcg() in __munlock_pagevec() in 20/32,
> > introduced in patchset v17, looks good but it isn't: I was lucky that
> > systemd at reboot did some munlocking that exposed the problem to lockdep.
> > The first time into the loop, lock_page_memcg() is done before lru_lock
> > (as 06/32 has allowed); but the second time around the loop, it is done
> > while still holding lru_lock.
>
> I don't know the details of lockdep show. Just wondering could it possible
> to solid the move_lock/lru_lock sequence?
> or try other blocking way which mentioned in commit_charge()?
>
> >
> > lock_page_memcg() really needs to be absorbed into (a variant of)
> > relock_page_lruvec(), and I do have that (it's awkward because of
> > the different ways in which the IRQ flags are handled). And out of
> > curiosity, I've also tried using that in mm/swap.c too, instead of the
> > TestClearPageLRU technique: lockdep is happy, but an update_lru_size()
> > warning showed that it cannot safely be mixed with the TestClearPageLRU
> > technique (that I'd left in isolate_lru_page()). So I'll stash away
> > that relock_page_lruvec(), and consider what's best for mm/mlock.c:
> > now that I've posted these comments so far, that's my priority, then
> > to get the result under testing again, before resuming these comments.
>
> No idea of your solution, but looking forward for your good news! :)
Yes, it is good news, and simpler than anything suggested above.
The main difficulties will probably be to look good in the 80 columns
(I know that limit has been lifted recently, but some of us use xterms
side by side), and to explain it.
mm/mlock.c has not been kept up-to-date very well: and in particular,
you have taken too seriously that "Serialize with any parallel
__split_huge_page_refcount()..." comment that you updated to two
comments "Serialize split tail pages in __split_huge_page_tail()...".
Delete them! The original comment was by Vlastimil for v3.14 in 2014.
But Kirill redesigned THP refcounting for v4.5 in 2016: that's when
__split_huge_page_refcount() went away. And with the new refcounting,
the THP splitting races that lru_lock protected munlock_vma_page()
and __munlock_pagevec() from: those races have become impossible.
Or maybe there never was such a race in __munlock_pagevec(): you
have added the comment there, assuming lru_lock was for that purpose,
but that was probably just the convenient place to take it,
to cover all the del_page_from_lru()s.
Observe how split_huge_page_to_list() uses unmap_page() to remove
all pmds and all ptes for the huge page being split, and remap_page()
only replaces the migration entries (used for anon but not for shmem
or file) after doing all of the __split_huge_page_tail()s, before
unlocking any of the pages. Recall that munlock_vma_page() and
__munlock_pagevec() are being applied to pages found mapped
into userspace, by ptes or pmd: there are none of those while
__split_huge_page_tail() is being used, so no race to protect from.
(Could a newly detached tail be freshly faulted into userspace just
before __split_huge_page() has reached the head? Not quite, the
fault has to wait to get the tail's page lock. But even if it
could, how would that be a problem for __munlock_pagevec()?)
There's lots more that could be said: for example, PageMlocked will
always be clear on the THP head during __split_huge_page_tail(),
because the last unmap of a PageMlocked page does clear_page_mlock().
But that's not required to prove the case, it's just another argument
against the "Serialize" comment you have in __munlock_pagevec().
So, no need for the problematic lock_page_memcg(page) there in
__munlock_pagevec(), nor to lock (or relock) lruvec just below it.
__munlock_pagevec() still needs lru_lock to del_page_from_lru_list(),
of course, but that must be done after your TestClearPageMlocked has
stabilized page->memcg. Use relock_page_lruvec_irq() here? I suppose
that will be easiest, but notice how __munlock_pagevec_fill() has
already made sure that all the pages in the pagevec are from the same
zone (and it cannot do the same for memcg without locking page memcg);
so some of relock's work will be redundant.
Otherwise, I'm much happier with your mm/mlock.c since looking at it
in more detail: a couple of nits though - drop the clear_page_mlock()
hunk from 25/32 - kernel style says do it the way you are undoing by
- if (!isolate_lru_page(page)) {
+ if (!isolate_lru_page(page))
putback_lru_page(page);
- } else {
+ else {
I don't always follow that over-braced style when making changes,
but you should not touch otherwise untouched code just to make it
go against the approved style. And in munlock_vma_page(),
- if (!TestClearPageMlocked(page)) {
+ if (!TestClearPageMlocked(page))
/* Potentially, PTE-mapped THP: do not skip the rest PTEs */
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Thu, 10 Sep 2020, Alexander Duyck wrote: > On Wed, Sep 9, 2020 at 5:32 PM Hugh Dickins wrote: > > On Wed, 9 Sep 2020, Alexander Duyck wrote: > > > On Tue, Sep 8, 2020 at 4:41 PM Hugh Dickins wrote: > > > > [PATCH v18 28/32] mm/compaction: Drop locked from > > > > isolate_migratepages_block > > > > Most of this consists of replacing "locked" by "lruvec", which is good: > > > > but please fold those changes back into 20/32 (or would it be 17/32? > > > > I've not yet looked into the relationship between those two), so we > > > > can then see more clearly what change this 28/32 (will need renaming!) > > > > actually makes, to use lruvec_holds_page_lru_lock(). That may be a > > > > good change, but it's mixed up with the "locked"->"lruvec" at present, > > > > and I think you could have just used lruvec for locked all along > > > > (but of course there's a place where you'll need new_lruvec too). > > > > > > I am good with my patch being folded in. No need to keep it separate. > > > > Thanks. Though it was only the "locked"->"lruvec" changes I was > > suggesting to fold back, to minimize the diff, so that we could > > see your use of lruvec_holds_page_lru_lock() more clearly - you > > had not introduced that function at the stage of the earlier patches. > > > > But now that I stare at it again, using lruvec_holds_page_lru_lock() > > there doesn't look like an advantage to me: when it decides no, the > > same calculation is made all over again in mem_cgroup_page_lruvec(), > > whereas the code before only had to calculate it once. > > > > So, the code before looks better to me: I wonder, do you think that > > rcu_read_lock() is more expensive than I think it? There can be > > debug instrumentation that makes it heavier, but by itself it is > > very cheap (by design) - not worth branching around. > > Actually what I was more concerned with was the pointer chase that > required the RCU lock. With this function we are able to compare a > pair of pointers from the page and the lruvec and avoid the need for > the RCU lock. The way the old code was working we had to crawl through > the memcg to get to the lruvec before we could compare it to the one > we currently hold. The general idea is to use the data we have instead > of having to pull in some additional cache lines to perform the test. When you say "With this function...", I think you are referring to lruvec_holds_page_lru_lock(). Yes, I appreciate what you're doing there, making calculations from known-stable data, and taking it no further than the required comparison; and I think (I don't yet claim to have reviewed 21/32) what you do with it in relock_page_lruvec*() is an improvement over what we had there before. But here I'm talking about using it in isolate_migratepages_block() in 28/32: in this case, the code before evaluated the new lruvec, compared against the old, and immediately used the new lruvec if different; whereas using lruvec_holds_page_lru_lock() makes an almost (I agree not entirely, and I haven't counted cachelines) equivalent evaluation, but its results have to thrown away when it's false, then the new lruvec actually calculated and used. The same "results thrown away" criticism can be made of relock_page_lruvec*(), but what was done there before your rewrite in v18 was no better: they both resort to lock_page_lruvec*(page), working it all out again from page. And I'm not suggesting that be changed, not at this point anyway; but 28/32 looks to me like a regression from what was done there before 28/32. > > > > > > > > [PATCH v18 29/32] mm: Identify compound pages sooner in > > > > isolate_migratepages_block > > > > NAK. I agree that isolate_migratepages_block() looks nicer this way, but > > > > take a look at prep_new_page() in mm/page_alloc.c: post_alloc_hook() is > > > > where set_page_refcounted() changes page->_refcount from 0 to 1, > > > > allowing > > > > a racing get_page_unless_zero() to succeed; then later > > > > prep_compound_page() > > > > is where PageHead and PageTails get set. So there's a small race window > > > > in > > > > which this patch could deliver a compound page when it should not. > > > > > > So the main motivation for the patch was to avoid the case where we > > > are having to reset the LRU flag. > > > > That would be satisfying. Not necessary, but I agree satisfying. > > Maybe depends also on your "skip" change, which I've not looked at yet? > > My concern is that we have scenarios where isolate_migratepages_block > could possibly prevent another page from being able to isolate a page. > I'm mostly concerned with us potentially creating something like an > isolation leak if multiple threads are doing something like clearing > and then resetting the LRU flag. In my mind if we clear the LRU flag > we should be certain we are going to remove the page as otherwise > another thread would have done it if it would have been allowed > access. I agree it's nicer not to TestClearPageLRU unnecessarily;
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Fri, 11 Sep 2020, Alex Shi wrote: > 在 2020/9/10 上午7:16, Hugh Dickins 写道: > > On Wed, 9 Sep 2020, Alex Shi wrote: > >> 在 2020/9/9 上午7:41, Hugh Dickins 写道: > >>> > >>> [PATCH v18 05/32] mm/thp: remove code path which never got into > >>> This is a good simplification, but I see no sign that you understand > >>> why it's valid: it relies on lru_add_page_tail() being called while > >>> head refcount is frozen to 0: we would not get this far if someone > >>> else holds a reference to the THP - which they must hold if they have > >>> isolated the page from its lru (and that's true before or after your > >>> per-memcg changes - but even truer after those changes, since PageLRU > >>> can then be flipped without lru_lock at any instant): please explain > >>> something of this in the commit message. > >> > >> Is the following commit log better? > >> > >> split_huge_page() will never call on a page which isn't on lru list, so > >> this code never got a chance to run, and should not be run, to add tail > >> pages on a lru list which head page isn't there. > >> > >> Hugh Dickins' mentioned: > >> The path should never be called since lru_add_page_tail() being called > >> while head refcount is frozen to 0: we would not get this far if > >> someone > >> else holds a reference to the THP - which they must hold if they have > >> isolated the page from its lru. > >> > >> Although the bug was never triggered, it'better be removed for code > >> correctness, and add a warn for unexpected calling. > > > > Not much better, no. split_huge_page() can easily be called for a page > > which is not on the lru list at the time, > > Hi Hugh, > > Thanks for comments! > > There are some discussion on this point a couple of weeks ago, > https://lkml.org/lkml/2020/7/9/760 > > Matthew Wilcox and Kirill have the following comments, > > I don't understand how we get to split_huge_page() with a page that's > > not on an LRU list. Both anonymous and page cache pages should be on > > an LRU list. What am I missing? > > Right, and it's never got removed from LRU during the split. The tail > pages have to be added to LRU because they now separate from the tail > page. > > -- > Kirill A. Shutemov Yes, those were among the mails that I read through before getting down to review. I was surprised by their not understanding, but it was a bit late to reply to that thread. Perhaps everybody had been focused on pages which have been and naturally belong on an LRU list, rather than pages which are on the LRU list at the instant that split_huge_page() is called. There are a number of places where PageLRU gets cleared, and a number of places where we del_page_from_lru_list(), I think you'll agree: your patches touch all or most of them. Let's think of a common one, isolate_lru_pages() used by page reclaim, but the same would apply to most of the others. Then there a number of places where split_huge_page() is called: I am having difficulty finding any of those which cannot race with page reclaim, but shall we choose anon THP's deferred_split_scan(), or shmem THP's shmem_punch_compound()? What prevents either of those from calling split_huge_page() at a time when isolate_lru_pages() has removed the page from LRU? But there's no problem in this race, because anyone isolating the page from LRU must hold their own reference to the page (to prevent it from being freed independently), and the can_split_huge_page() or page_ref_freeze() in split_huge_page_to_list() will detect that and fail the split with -EBUSY (or else succeed and prevent new references from being acquired). So this case never reaches lru_add_page_tail(). > > > and I don't know what was the > > bug which was never triggered. > > So the only path to the removed part should be a bug, like sth here, > https://lkml.org/lkml/2020/7/10/118 > or > https://lkml.org/lkml/2020/7/10/972 Oh, the use of split_huge_page() in __iommu_dma_alloc_pages() is just nonsense, I thought it had already been removed - perhaps some debate over __GFP_COMP held it up. Not something you need worry about in this patchset. > > > Stick with whatever text you end up with > > for the combination of 05/32 and 18/32, and I'll rewrite it after. > > I am not object to merge them into one, I just don't know how to say > clear about 2 patches in commit log. As patch 18, TestClearPageLRU > add the incorrect posibility of remove lru bit during split, that's > the reason of code path rewrite and a WARN there. I did not know that was why you were putting 18/32 in at that point, it does not mention TestClearPageLRU at all. But the fact remains that it's a nice cleanup, contains a reassuring WARN if we got it wrong (and I've suggested a WARN on the other branch too), it was valid before your changes, and it's valid after your changes. Please merge it back into the uglier 05/32, and again I'll rewrite whatever comment you come up with if necessary. > > > >>>
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
在 2020/9/10 上午7:16, Hugh Dickins 写道: > On Wed, 9 Sep 2020, Alex Shi wrote: >> 在 2020/9/9 上午7:41, Hugh Dickins 写道: >>> >>> [PATCH v18 05/32] mm/thp: remove code path which never got into >>> This is a good simplification, but I see no sign that you understand >>> why it's valid: it relies on lru_add_page_tail() being called while >>> head refcount is frozen to 0: we would not get this far if someone >>> else holds a reference to the THP - which they must hold if they have >>> isolated the page from its lru (and that's true before or after your >>> per-memcg changes - but even truer after those changes, since PageLRU >>> can then be flipped without lru_lock at any instant): please explain >>> something of this in the commit message. >> >> Is the following commit log better? >> >> split_huge_page() will never call on a page which isn't on lru list, so >> this code never got a chance to run, and should not be run, to add tail >> pages on a lru list which head page isn't there. >> >> Hugh Dickins' mentioned: >> The path should never be called since lru_add_page_tail() being called >> while head refcount is frozen to 0: we would not get this far if someone >> else holds a reference to the THP - which they must hold if they have >> isolated the page from its lru. >> >> Although the bug was never triggered, it'better be removed for code >> correctness, and add a warn for unexpected calling. > > Not much better, no. split_huge_page() can easily be called for a page > which is not on the lru list at the time, Hi Hugh, Thanks for comments! There are some discussion on this point a couple of weeks ago, https://lkml.org/lkml/2020/7/9/760 Matthew Wilcox and Kirill have the following comments, > I don't understand how we get to split_huge_page() with a page that's > not on an LRU list. Both anonymous and page cache pages should be on > an LRU list. What am I missing? Right, and it's never got removed from LRU during the split. The tail pages have to be added to LRU because they now separate from the tail page. -- Kirill A. Shutemov > and I don't know what was the > bug which was never triggered. So the only path to the removed part should be a bug, like sth here, https://lkml.org/lkml/2020/7/10/118 or https://lkml.org/lkml/2020/7/10/972 > Stick with whatever text you end up with > for the combination of 05/32 and 18/32, and I'll rewrite it after. I am not object to merge them into one, I just don't know how to say clear about 2 patches in commit log. As patch 18, TestClearPageLRU add the incorrect posibility of remove lru bit during split, that's the reason of code path rewrite and a WARN there. Thanks Alex > >>> [PATCH v18 06/32] mm/thp: narrow lru locking >>> Why? What part does this play in the series? "narrow lru locking" can >>> also be described as "widen page cache locking": >> >> Uh, the page cache locking isn't widen, it's still on the old place. > > I'm not sure if you're joking there. Perhaps just a misunderstanding. > > Yes, patch 06/32 does not touch the xa_lock(>i_pages) and > xa_lock(_cache->i_pages) lines (odd how we've arrived at two of > those, but please do not get into cleaning it up now); but it removes > the spin_lock_irqsave(>lru_lock, flags) which used to come > before them, and inserts a spin_lock(>lru_lock) after them. > > You call that narrowing the lru locking, okay, but I see it as also > pushing the page cache locking outwards: before this patch, page cache > lock was taken inside lru_lock; after this patch, page cache lock is > taken outside lru_lock. If you cannot see that, then I think you > should not have touched this code at all; but it's what we have > been testing, and I think we should go forward with it. > >>> But I wish you could give some reason for it in the commit message! >> >> It's a head scratch task. Would you like to tell me what's detailed info >> should be there? Thanks! > > So, you don't know why you did it either: then it will be hard to > justify. I guess I'll have to write something for it later. I'm > strongly tempted just to drop the patch, but expect it will become > useful later, for using lock_page_memcg() before getting lru_lock. > I thought the xa_lock and lru_lock relationship was described clear in the commit log, and still no idea of the move_lock in the chain. Please refill them for what I overlooked. Thanks! >>> Signed-off-by: Wei Yang >>> Is that correct? Or Wei Yang suggested some part of it perhaps? >> >> Yes, we talked a lot to confirm the locking change is safe. > > Okay, but the patch was written by you, and sent by you to Andrew: > that is not a case for "Signed-off-by: Someone Else". > Ok. let's remove his signed-off. >>> [PATCH v18 27/32] mm/swap.c: optimizing __pagevec_lru_add lru_lock >>> Could we please drop this one for the moment? And come back to it later >>> when the basic series is safely in. It's a good idea to try sorting >>> together those pages which
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Wed, Sep 9, 2020 at 5:32 PM Hugh Dickins wrote: > > On Wed, 9 Sep 2020, Alexander Duyck wrote: > > On Tue, Sep 8, 2020 at 4:41 PM Hugh Dickins wrote: > > > [PATCH v18 28/32] mm/compaction: Drop locked from > > > isolate_migratepages_block > > > Most of this consists of replacing "locked" by "lruvec", which is good: > > > but please fold those changes back into 20/32 (or would it be 17/32? > > > I've not yet looked into the relationship between those two), so we > > > can then see more clearly what change this 28/32 (will need renaming!) > > > actually makes, to use lruvec_holds_page_lru_lock(). That may be a > > > good change, but it's mixed up with the "locked"->"lruvec" at present, > > > and I think you could have just used lruvec for locked all along > > > (but of course there's a place where you'll need new_lruvec too). > > > > I am good with my patch being folded in. No need to keep it separate. > > Thanks. Though it was only the "locked"->"lruvec" changes I was > suggesting to fold back, to minimize the diff, so that we could > see your use of lruvec_holds_page_lru_lock() more clearly - you > had not introduced that function at the stage of the earlier patches. > > But now that I stare at it again, using lruvec_holds_page_lru_lock() > there doesn't look like an advantage to me: when it decides no, the > same calculation is made all over again in mem_cgroup_page_lruvec(), > whereas the code before only had to calculate it once. > > So, the code before looks better to me: I wonder, do you think that > rcu_read_lock() is more expensive than I think it? There can be > debug instrumentation that makes it heavier, but by itself it is > very cheap (by design) - not worth branching around. Actually what I was more concerned with was the pointer chase that required the RCU lock. With this function we are able to compare a pair of pointers from the page and the lruvec and avoid the need for the RCU lock. The way the old code was working we had to crawl through the memcg to get to the lruvec before we could compare it to the one we currently hold. The general idea is to use the data we have instead of having to pull in some additional cache lines to perform the test. > > > > > [PATCH v18 29/32] mm: Identify compound pages sooner in > > > isolate_migratepages_block > > > NAK. I agree that isolate_migratepages_block() looks nicer this way, but > > > take a look at prep_new_page() in mm/page_alloc.c: post_alloc_hook() is > > > where set_page_refcounted() changes page->_refcount from 0 to 1, allowing > > > a racing get_page_unless_zero() to succeed; then later > > > prep_compound_page() > > > is where PageHead and PageTails get set. So there's a small race window in > > > which this patch could deliver a compound page when it should not. > > > > So the main motivation for the patch was to avoid the case where we > > are having to reset the LRU flag. > > That would be satisfying. Not necessary, but I agree satisfying. > Maybe depends also on your "skip" change, which I've not looked at yet? My concern is that we have scenarios where isolate_migratepages_block could possibly prevent another page from being able to isolate a page. I'm mostly concerned with us potentially creating something like an isolation leak if multiple threads are doing something like clearing and then resetting the LRU flag. In my mind if we clear the LRU flag we should be certain we are going to remove the page as otherwise another thread would have done it if it would have been allowed access. > > One question I would have is what if > > we swapped the code block with the __isolate_lru_page_prepare section? > > WIth that we would be taking a reference on the page, then verifying > > the LRU flag is set, and then testing for compound page flag bit. > > Would doing that close the race window since the LRU flag being set > > should indicate that the allocation has already been completed has it > > not? > > Yes, I think that would be safe, and would look better. But I am > very hesitant to give snap assurances here (I've twice missed out > a vital PageLRU check from this sequence myself): it is very easy > to deceive myself and only see it later. I'm not looking for assurances, just sanity checks to make sure I am not missing something obvious. > If you can see a bug in what's there before these patches, certainly > we need to fix it. But adding non-essential patches to the already > overlong series risks delaying it. My concern ends up being that if we are clearing the bit and restoring it while holding the LRU lock we can effectively cause pages to become pseudo-pinned on the LRU. In my mind I would want us to avoid clearing the LRU flag until we know we are going to be pulling the page from the list once we take the lruvec lock. I interpret clearing of the flag to indicate the page has already been pulled, it just hasn't left the list yet. With us resetting the bit we are violating that which I worry will lead to issues.
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Wed, 9 Sep 2020, Alex Shi wrote: > 在 2020/9/9 上午7:41, Hugh Dickins 写道: > > > > [PATCH v18 05/32] mm/thp: remove code path which never got into > > This is a good simplification, but I see no sign that you understand > > why it's valid: it relies on lru_add_page_tail() being called while > > head refcount is frozen to 0: we would not get this far if someone > > else holds a reference to the THP - which they must hold if they have > > isolated the page from its lru (and that's true before or after your > > per-memcg changes - but even truer after those changes, since PageLRU > > can then be flipped without lru_lock at any instant): please explain > > something of this in the commit message. > > Is the following commit log better? > > split_huge_page() will never call on a page which isn't on lru list, so > this code never got a chance to run, and should not be run, to add tail > pages on a lru list which head page isn't there. > > Hugh Dickins' mentioned: > The path should never be called since lru_add_page_tail() being called > while head refcount is frozen to 0: we would not get this far if someone > else holds a reference to the THP - which they must hold if they have > isolated the page from its lru. > > Although the bug was never triggered, it'better be removed for code > correctness, and add a warn for unexpected calling. Not much better, no. split_huge_page() can easily be called for a page which is not on the lru list at the time, and I don't know what was the bug which was never triggered. Stick with whatever text you end up with for the combination of 05/32 and 18/32, and I'll rewrite it after. > > [PATCH v18 06/32] mm/thp: narrow lru locking > > Why? What part does this play in the series? "narrow lru locking" can > > also be described as "widen page cache locking": > > Uh, the page cache locking isn't widen, it's still on the old place. I'm not sure if you're joking there. Perhaps just a misunderstanding. Yes, patch 06/32 does not touch the xa_lock(>i_pages) and xa_lock(_cache->i_pages) lines (odd how we've arrived at two of those, but please do not get into cleaning it up now); but it removes the spin_lock_irqsave(>lru_lock, flags) which used to come before them, and inserts a spin_lock(>lru_lock) after them. You call that narrowing the lru locking, okay, but I see it as also pushing the page cache locking outwards: before this patch, page cache lock was taken inside lru_lock; after this patch, page cache lock is taken outside lru_lock. If you cannot see that, then I think you should not have touched this code at all; but it's what we have been testing, and I think we should go forward with it. > > But I wish you could give some reason for it in the commit message! > > It's a head scratch task. Would you like to tell me what's detailed info > should be there? Thanks! So, you don't know why you did it either: then it will be hard to justify. I guess I'll have to write something for it later. I'm strongly tempted just to drop the patch, but expect it will become useful later, for using lock_page_memcg() before getting lru_lock. > > Signed-off-by: Wei Yang > > Is that correct? Or Wei Yang suggested some part of it perhaps? > > Yes, we talked a lot to confirm the locking change is safe. Okay, but the patch was written by you, and sent by you to Andrew: that is not a case for "Signed-off-by: Someone Else". > > [PATCH v18 27/32] mm/swap.c: optimizing __pagevec_lru_add lru_lock > > Could we please drop this one for the moment? And come back to it later > > when the basic series is safely in. It's a good idea to try sorting > > together those pages which come under the same lock (though my guess is > > that they naturally gather themselves together quite well already); but > > I'm not happy adding 360 bytes to the kernel stack here (and that in > > addition to 192 bytes of horrid pseudo-vma in the shmem swapin case), > > though that could be avoided by making it per-cpu. But I hope there's > > a simpler way of doing it, as efficient, but also useful for the other > > pagevec operations here: perhaps scanning the pagevec for same page-> > > mem_cgroup (and flags node bits), NULLing entries as they are done. > > Another, easily fixed, minor defect in this patch: if I'm reading it > > right, it reverses the order in which the pages are put on the lru? > > this patch could give about 10+% performance gain on my multiple memcg > readtwice testing. fairness locking cost the performance much. Good to know, should have been mentioned. s/fairness/Repeated/ But what was the gain or loss on your multiple memcg readtwice testing without this patch, compared against node-only lru_lock? The 80% gain mentioned before, I presume. So this further optimization can wait until the rest is solid. > > I also tried per cpu solution but that cause much trouble of per cpu func > things, and looks no benefit except a bit struct size of stack, so if >
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Wed, 9 Sep 2020, Alexander Duyck wrote: > On Tue, Sep 8, 2020 at 4:41 PM Hugh Dickins wrote: > > [PATCH v18 28/32] mm/compaction: Drop locked from isolate_migratepages_block > > Most of this consists of replacing "locked" by "lruvec", which is good: > > but please fold those changes back into 20/32 (or would it be 17/32? > > I've not yet looked into the relationship between those two), so we > > can then see more clearly what change this 28/32 (will need renaming!) > > actually makes, to use lruvec_holds_page_lru_lock(). That may be a > > good change, but it's mixed up with the "locked"->"lruvec" at present, > > and I think you could have just used lruvec for locked all along > > (but of course there's a place where you'll need new_lruvec too). > > I am good with my patch being folded in. No need to keep it separate. Thanks. Though it was only the "locked"->"lruvec" changes I was suggesting to fold back, to minimize the diff, so that we could see your use of lruvec_holds_page_lru_lock() more clearly - you had not introduced that function at the stage of the earlier patches. But now that I stare at it again, using lruvec_holds_page_lru_lock() there doesn't look like an advantage to me: when it decides no, the same calculation is made all over again in mem_cgroup_page_lruvec(), whereas the code before only had to calculate it once. So, the code before looks better to me: I wonder, do you think that rcu_read_lock() is more expensive than I think it? There can be debug instrumentation that makes it heavier, but by itself it is very cheap (by design) - not worth branching around. > > > [PATCH v18 29/32] mm: Identify compound pages sooner in > > isolate_migratepages_block > > NAK. I agree that isolate_migratepages_block() looks nicer this way, but > > take a look at prep_new_page() in mm/page_alloc.c: post_alloc_hook() is > > where set_page_refcounted() changes page->_refcount from 0 to 1, allowing > > a racing get_page_unless_zero() to succeed; then later prep_compound_page() > > is where PageHead and PageTails get set. So there's a small race window in > > which this patch could deliver a compound page when it should not. > > So the main motivation for the patch was to avoid the case where we > are having to reset the LRU flag. That would be satisfying. Not necessary, but I agree satisfying. Maybe depends also on your "skip" change, which I've not looked at yet? > One question I would have is what if > we swapped the code block with the __isolate_lru_page_prepare section? > WIth that we would be taking a reference on the page, then verifying > the LRU flag is set, and then testing for compound page flag bit. > Would doing that close the race window since the LRU flag being set > should indicate that the allocation has already been completed has it > not? Yes, I think that would be safe, and would look better. But I am very hesitant to give snap assurances here (I've twice missed out a vital PageLRU check from this sequence myself): it is very easy to deceive myself and only see it later. If you can see a bug in what's there before these patches, certainly we need to fix it. But adding non-essential patches to the already overlong series risks delaying it. Hugh
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
Hi Hugh,
Thanks a lot for so rich review and comments!
在 2020/9/9 上午7:41, Hugh Dickins 写道:
> Miscellaneous Acks and NAKs and other comments on the beginning and
> the end of the series, but not much yet on the all-important middle.
> I'm hoping to be spared sending ~20 email replies to ~20 patches.
>
> [PATCH v18 01/32] mm/memcg: warning on !memcg after readahead page charged
> Acked-by: Hugh Dickins
> if you make these changes:
>
> Please add "Add VM_WARN_ON_ONCE_PAGE() macro." or something like that to
> the commit message: that's a good addition that we shall find useful in
> other places, so please advertise it.
Accepted!
>
> Delete the four comment lines
> /* Readahead page is charged too, to see if other page uncharged */
> which make no sense on their own.
>
Accepted!
> [PATCH v18 02/32] mm/memcg: bail out early from swap accounting when memcg is
> disabled
> Acked-by: Hugh Dickins
>
> [PATCH v18 03/32] mm/thp: move lru_add_page_tail func to huge_memory.c
> Acked-by: Hugh Dickins
>
> [PATCH v18 04/32] mm/thp: clean up lru_add_page_tail
> Acked-by: Hugh Dickins
>
> Though I'd prefer "mm/thp: use head for head page in lru_add_page_tail"
> to the unnecessarily vague "clean up". But you're right to keep this
> renaming separate from the code movement in the previous commit, and
> perhaps right to keep it from the more interesting cleanup next.
>
> [PATCH v18 05/32] mm/thp: remove code path which never got into
> This is a good simplification, but I see no sign that you understand
> why it's valid: it relies on lru_add_page_tail() being called while
> head refcount is frozen to 0: we would not get this far if someone
> else holds a reference to the THP - which they must hold if they have
> isolated the page from its lru (and that's true before or after your
> per-memcg changes - but even truer after those changes, since PageLRU
> can then be flipped without lru_lock at any instant): please explain
> something of this in the commit message.
Is the following commit log better?
split_huge_page() will never call on a page which isn't on lru list, so
this code never got a chance to run, and should not be run, to add tail
pages on a lru list which head page isn't there.
Hugh Dickins' mentioned:
The path should never be called since lru_add_page_tail() being called
while head refcount is frozen to 0: we would not get this far if someone
else holds a reference to the THP - which they must hold if they have
isolated the page from its lru.
Although the bug was never triggered, it'better be removed for code
correctness, and add a warn for unexpected calling.
>
> You revisit this same code in 18/32, and I much prefer the way it looks
> after that (if (list) {} else {}) - this 05/32 is a bit weird, it would
> be easier to understand if it just did VM_WARN_ON(1). Please pull the
> 18/32 mods back into this one, maybe adding a VM_WARN_ON(PageLRU) into
> the "if (list)" block too.
Accepted.
>
> [PATCH v18 18/32] mm/thp: add tail pages into lru anyway in split_huge_page()
> Please merge into 05/32.
> But what do "Split_huge_page() must start with
> PageLRU(head)" and "Split start from PageLRU(head)" mean? Perhaps you mean
> that if list is NULL, then if the head was not on the LRU, then it cannot
> have got through page_ref_freeze(), because isolator would hold page ref?
No, what I mean is only PageLRU(head) could be called and get here. Would you
like to give a suggestion to replace old one?
> That is subtle, and deserves mention in the commit comment, but is not
> what you have said at all. s/unexpected/unexpectedly/.
Thanks!
>
> [PATCH v18 06/32] mm/thp: narrow lru locking
> Why? What part does this play in the series? "narrow lru locking" can
> also be described as "widen page cache locking":
Uh, the page cache locking isn't widen, it's still on the old place.
> you are changing the
> lock ordering, and not giving any reason to do so. This may be an
> excellent change, or it may be a terrible change: I find that usually
> lock ordering is forced upon us, and it's rare to meet an instance like
> this that could go either way, and I don't know myself how to judge it.
>
> I do want this commit to go in, partly because it has been present in
> all the testing we have done, and partly because I *can at last* see a
> logical advantage to it - it also nests lru_lock inside memcg->move_lock,
I must overlook sth on the lock nest. Would you like to reveal it for me?
Thanks!
> allowing lock_page_memcg() to be used to stabilize page->mem_cgroup when
> getting per-memcg lru_lock - though only in one place, starting in v17,
> do you actually use that (and, warning: it's not used correctly there).
>
> I'm not very bothered by how the local_irq_disable() looks to RT: THP
> seems a very bad idea in an RT kernel. Earlier I asked you to run this
> past Kirill and Matthew and Johannes: you did so, thank you, and Kirill
> has blessed it, and no one has
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Tue, Sep 8, 2020 at 4:41 PM Hugh Dickins wrote: > > [PATCH v18 28/32] mm/compaction: Drop locked from isolate_migratepages_block > Most of this consists of replacing "locked" by "lruvec", which is good: > but please fold those changes back into 20/32 (or would it be 17/32? > I've not yet looked into the relationship between those two), so we > can then see more clearly what change this 28/32 (will need renaming!) > actually makes, to use lruvec_holds_page_lru_lock(). That may be a > good change, but it's mixed up with the "locked"->"lruvec" at present, > and I think you could have just used lruvec for locked all along > (but of course there's a place where you'll need new_lruvec too). I am good with my patch being folded in. No need to keep it separate. > [PATCH v18 29/32] mm: Identify compound pages sooner in > isolate_migratepages_block > NAK. I agree that isolate_migratepages_block() looks nicer this way, but > take a look at prep_new_page() in mm/page_alloc.c: post_alloc_hook() is > where set_page_refcounted() changes page->_refcount from 0 to 1, allowing > a racing get_page_unless_zero() to succeed; then later prep_compound_page() > is where PageHead and PageTails get set. So there's a small race window in > which this patch could deliver a compound page when it should not. So the main motivation for the patch was to avoid the case where we are having to reset the LRU flag. One question I would have is what if we swapped the code block with the __isolate_lru_page_prepare section? WIth that we would be taking a reference on the page, then verifying the LRU flag is set, and then testing for compound page flag bit. Would doing that close the race window since the LRU flag being set should indicate that the allocation has already been completed has it not? > [PATCH v18 30/32] mm: Drop use of test_and_set_skip in favor of just setting > skip > I haven't looked at this yet (but recall that per-memcg lru_lock can > change the point at which compaction should skip a contended lock: IIRC > the current kernel needs nothing extra, whereas some earlier kernels did > need extra; but when I look at 30/32, may find these remarks irrelevant). > > [PATCH v18 31/32] mm: Add explicit page decrement in exception path for > isolate_lru_pages > The title of this patch is definitely wrong: there was an explicit page > decrement there before (put_page), now it's wrapping it up inside a > WARN_ON(). We usually prefer to avoid doing functional operations > inside WARN/BUGs, but I think I'll overlook that - anyone else worried? > The comment is certainly better than what was there before: yes, this > warning reflects the difficulty we have in thinking about the > TestClearPageLRU protocol: which I'm still not sold on, but > agree we should proceed with. With a change in title, perhaps > "mm: add warning where TestClearPageLRU failed on freeable page"? > Acked-by: Hugh Dickins I can update that and resubmit it if needed. I know there were also some suggestions from Matthew.
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
On Tue, Sep 08, 2020 at 04:41:00PM -0700, Hugh Dickins wrote: [...] >[PATCH v18 06/32] mm/thp: narrow lru locking >Why? What part does this play in the series? "narrow lru locking" can >also be described as "widen page cache locking": you are changing the >lock ordering, and not giving any reason to do so. This may be an >excellent change, or it may be a terrible change: I find that usually >lock ordering is forced upon us, and it's rare to meet an instance like >this that could go either way, and I don't know myself how to judge it. > >I do want this commit to go in, partly because it has been present in >all the testing we have done, and partly because I *can at last* see a >logical advantage to it - it also nests lru_lock inside memcg->move_lock, >allowing lock_page_memcg() to be used to stabilize page->mem_cgroup when >getting per-memcg lru_lock - though only in one place, starting in v17, >do you actually use that (and, warning: it's not used correctly there). > >I'm not very bothered by how the local_irq_disable() looks to RT: THP >seems a very bad idea in an RT kernel. Earlier I asked you to run this >past Kirill and Matthew and Johannes: you did so, thank you, and Kirill >has blessed it, and no one has nacked it, and I have not noticed any >disadvantage from this change in lock ordering (documented in 23/32), >so I'm now going to say > >Acked-by: Hugh Dickins > >But I wish you could give some reason for it in the commit message! > >Signed-off-by: Wei Yang >Is that correct? Or Wei Yang suggested some part of it perhaps? > If my memory is correct, we had some offline discussion about this change. -- Wei Yang Help you, Help me
Re: [PATCH v18 00/32] per memcg lru_lock: reviews
Miscellaneous Acks and NAKs and other comments on the beginning and
the end of the series, but not much yet on the all-important middle.
I'm hoping to be spared sending ~20 email replies to ~20 patches.
[PATCH v18 01/32] mm/memcg: warning on !memcg after readahead page charged
Acked-by: Hugh Dickins
if you make these changes:
Please add "Add VM_WARN_ON_ONCE_PAGE() macro." or something like that to
the commit message: that's a good addition that we shall find useful in
other places, so please advertise it.
Delete the four comment lines
/* Readahead page is charged too, to see if other page uncharged */
which make no sense on their own.
[PATCH v18 02/32] mm/memcg: bail out early from swap accounting when memcg is
disabled
Acked-by: Hugh Dickins
[PATCH v18 03/32] mm/thp: move lru_add_page_tail func to huge_memory.c
Acked-by: Hugh Dickins
[PATCH v18 04/32] mm/thp: clean up lru_add_page_tail
Acked-by: Hugh Dickins
Though I'd prefer "mm/thp: use head for head page in lru_add_page_tail"
to the unnecessarily vague "clean up". But you're right to keep this
renaming separate from the code movement in the previous commit, and
perhaps right to keep it from the more interesting cleanup next.
[PATCH v18 05/32] mm/thp: remove code path which never got into
This is a good simplification, but I see no sign that you understand
why it's valid: it relies on lru_add_page_tail() being called while
head refcount is frozen to 0: we would not get this far if someone
else holds a reference to the THP - which they must hold if they have
isolated the page from its lru (and that's true before or after your
per-memcg changes - but even truer after those changes, since PageLRU
can then be flipped without lru_lock at any instant): please explain
something of this in the commit message.
You revisit this same code in 18/32, and I much prefer the way it looks
after that (if (list) {} else {}) - this 05/32 is a bit weird, it would
be easier to understand if it just did VM_WARN_ON(1). Please pull the
18/32 mods back into this one, maybe adding a VM_WARN_ON(PageLRU) into
the "if (list)" block too.
[PATCH v18 18/32] mm/thp: add tail pages into lru anyway in split_huge_page()
Please merge into 05/32. But what do "Split_huge_page() must start with
PageLRU(head)" and "Split start from PageLRU(head)" mean? Perhaps you mean
that if list is NULL, then if the head was not on the LRU, then it cannot
have got through page_ref_freeze(), because isolator would hold page ref?
That is subtle, and deserves mention in the commit comment, but is not
what you have said at all. s/unexpected/unexpectedly/.
[PATCH v18 06/32] mm/thp: narrow lru locking
Why? What part does this play in the series? "narrow lru locking" can
also be described as "widen page cache locking": you are changing the
lock ordering, and not giving any reason to do so. This may be an
excellent change, or it may be a terrible change: I find that usually
lock ordering is forced upon us, and it's rare to meet an instance like
this that could go either way, and I don't know myself how to judge it.
I do want this commit to go in, partly because it has been present in
all the testing we have done, and partly because I *can at last* see a
logical advantage to it - it also nests lru_lock inside memcg->move_lock,
allowing lock_page_memcg() to be used to stabilize page->mem_cgroup when
getting per-memcg lru_lock - though only in one place, starting in v17,
do you actually use that (and, warning: it's not used correctly there).
I'm not very bothered by how the local_irq_disable() looks to RT: THP
seems a very bad idea in an RT kernel. Earlier I asked you to run this
past Kirill and Matthew and Johannes: you did so, thank you, and Kirill
has blessed it, and no one has nacked it, and I have not noticed any
disadvantage from this change in lock ordering (documented in 23/32),
so I'm now going to say
Acked-by: Hugh Dickins
But I wish you could give some reason for it in the commit message!
Signed-off-by: Wei Yang
Is that correct? Or Wei Yang suggested some part of it perhaps?
[PATCH v18 07/32] mm/swap.c: stop deactivate_file_page if page not on lru
Perhaps; or perhaps by the time the pagevec is full, the page has been
drained to the lru, and it should be deactivated? I'm indifferent.
Is this important for per-memcg lru_lock?
[PATCH v18 08/32] mm/vmscan: remove unnecessary lruvec adding
You are optimizing for a case which you then mark unlikely(), and I
don't agree that it makes the flow clearer; but you've added a useful
comment on the race there, so please s/intergrity/integrity/ in commit
message and in code comment, then
Acked-by: Hugh Dickins
[PATCH v18 09/32] mm/page_idle: no unlikely double check for idle page counting
I strongly approve of removing the abuse of lru_lock here, but the
patch is wrong: you are mistaken in thinking the PageLRU check after
get_page_unless_zero() is an unnecessary duplicaton of the one before.
No, the one before is an optimization, and the one
