Hugh Dickins' found a memcg change bug on original version:
If we want to change the pgdat->lru_lock to memcg's lruvec lock, we have
to serialize mem_cgroup_move_account during pagevec_lru_move_fn. The
possible bad scenario would like:
cpu 0 cpu 1
lruvec = mem_cgroup_page_lruvec()
if (!isolate_lru_page())
mem_cgroup_move_account
spin_lock_irqsave(&lruvec->lru_lock <== wrong lock.
So we need the ClearPageLRU to block isolate_lru_page(), then serialize
the memcg change here.
Reported-by: Hugh Dickins <[email protected]>
Signed-off-by: Alex Shi <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: [email protected]
Cc: [email protected]
---
mm/swap.c | 31 +++++++++++++++++++------------
1 file changed, 19 insertions(+), 12 deletions(-)
diff --git a/mm/swap.c b/mm/swap.c
index b24d5f69b93a..55eb2c2eed03 100644
--- a/mm/swap.c
+++ b/mm/swap.c
@@ -203,7 +203,7 @@ int get_kernel_page(unsigned long start, int write, struct page **pages)
EXPORT_SYMBOL_GPL(get_kernel_page);
static void pagevec_lru_move_fn(struct pagevec *pvec,
- void (*move_fn)(struct page *page, struct lruvec *lruvec))
+ void (*move_fn)(struct page *page, struct lruvec *lruvec), bool add)
{
int i;
struct pglist_data *pgdat = NULL;
@@ -221,8 +221,15 @@ static void pagevec_lru_move_fn(struct pagevec *pvec,
spin_lock_irqsave(&pgdat->lru_lock, flags);
}
+ /* new page add to lru or page moving between lru */
+ if (!add && !TestClearPageLRU(page))
+ continue;
+
lruvec = mem_cgroup_page_lruvec(page, pgdat);
(*move_fn)(page, lruvec);
+
+ if (!add)
+ SetPageLRU(page);
}
if (pgdat)
spin_unlock_irqrestore(&pgdat->lru_lock, flags);
@@ -259,7 +266,7 @@ void rotate_reclaimable_page(struct page *page)
local_lock_irqsave(&lru_rotate.lock, flags);
pvec = this_cpu_ptr(&lru_rotate.pvec);
if (!pagevec_add(pvec, page) || PageCompound(page))
- pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
+ pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, false);
local_unlock_irqrestore(&lru_rotate.lock, flags);
}
}
@@ -328,7 +335,7 @@ static void activate_page_drain(int cpu)
struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
if (pagevec_count(pvec))
- pagevec_lru_move_fn(pvec, __activate_page);
+ pagevec_lru_move_fn(pvec, __activate_page, false);
}
static bool need_activate_page_drain(int cpu)
@@ -346,7 +353,7 @@ void activate_page(struct page *page)
pvec = this_cpu_ptr(&lru_pvecs.activate_page);
get_page(page);
if (!pagevec_add(pvec, page) || PageCompound(page))
- pagevec_lru_move_fn(pvec, __activate_page);
+ pagevec_lru_move_fn(pvec, __activate_page, false);
local_unlock(&lru_pvecs.lock);
}
}
@@ -621,21 +628,21 @@ void lru_add_drain_cpu(int cpu)
/* No harm done if a racing interrupt already did this */
local_lock_irqsave(&lru_rotate.lock, flags);
- pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
+ pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, false);
local_unlock_irqrestore(&lru_rotate.lock, flags);
}
pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
if (pagevec_count(pvec))
- pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
+ pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, false);
pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
if (pagevec_count(pvec))
- pagevec_lru_move_fn(pvec, lru_deactivate_fn);
+ pagevec_lru_move_fn(pvec, lru_deactivate_fn, false);
pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
if (pagevec_count(pvec))
- pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
+ pagevec_lru_move_fn(pvec, lru_lazyfree_fn, false);
activate_page_drain(cpu);
}
@@ -664,7 +671,7 @@ void deactivate_file_page(struct page *page)
pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
if (!pagevec_add(pvec, page) || PageCompound(page))
- pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
+ pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, false);
local_unlock(&lru_pvecs.lock);
}
}
@@ -686,7 +693,7 @@ void deactivate_page(struct page *page)
pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
get_page(page);
if (!pagevec_add(pvec, page) || PageCompound(page))
- pagevec_lru_move_fn(pvec, lru_deactivate_fn);
+ pagevec_lru_move_fn(pvec, lru_deactivate_fn, false);
local_unlock(&lru_pvecs.lock);
}
}
@@ -708,7 +715,7 @@ void mark_page_lazyfree(struct page *page)
pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
get_page(page);
if (!pagevec_add(pvec, page) || PageCompound(page))
- pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
+ pagevec_lru_move_fn(pvec, lru_lazyfree_fn, false);
local_unlock(&lru_pvecs.lock);
}
}
@@ -976,7 +983,7 @@ static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec)
*/
void __pagevec_lru_add(struct pagevec *pvec)
{
- pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn);
+ pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, true);
}
/**
--
1.8.3.1
On Fri, Jul 3, 2020 at 8:09 AM Alex Shi <[email protected]> wrote:
>
> Hugh Dickins' found a memcg change bug on original version:
> If we want to change the pgdat->lru_lock to memcg's lruvec lock, we have
> to serialize mem_cgroup_move_account during pagevec_lru_move_fn. The
> possible bad scenario would like:
>
> cpu 0 cpu 1
> lruvec = mem_cgroup_page_lruvec()
> if (!isolate_lru_page())
> mem_cgroup_move_account
>
> spin_lock_irqsave(&lruvec->lru_lock <== wrong lock.
>
> So we need the ClearPageLRU to block isolate_lru_page(), then serialize
> the memcg change here.
>
> Reported-by: Hugh Dickins <[email protected]>
> Signed-off-by: Alex Shi <[email protected]>
> Cc: Andrew Morton <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
> ---
> mm/swap.c | 31 +++++++++++++++++++------------
> 1 file changed, 19 insertions(+), 12 deletions(-)
>
> diff --git a/mm/swap.c b/mm/swap.c
> index b24d5f69b93a..55eb2c2eed03 100644
> --- a/mm/swap.c
> +++ b/mm/swap.c
> @@ -203,7 +203,7 @@ int get_kernel_page(unsigned long start, int write, struct page **pages)
> EXPORT_SYMBOL_GPL(get_kernel_page);
>
> static void pagevec_lru_move_fn(struct pagevec *pvec,
> - void (*move_fn)(struct page *page, struct lruvec *lruvec))
> + void (*move_fn)(struct page *page, struct lruvec *lruvec), bool add)
> {
> int i;
> struct pglist_data *pgdat = NULL;
> @@ -221,8 +221,15 @@ static void pagevec_lru_move_fn(struct pagevec *pvec,
> spin_lock_irqsave(&pgdat->lru_lock, flags);
> }
>
> + /* new page add to lru or page moving between lru */
> + if (!add && !TestClearPageLRU(page))
> + continue;
> +
> lruvec = mem_cgroup_page_lruvec(page, pgdat);
> (*move_fn)(page, lruvec);
> +
> + if (!add)
> + SetPageLRU(page);
> }
> if (pgdat)
> spin_unlock_irqrestore(&pgdat->lru_lock, flags);
> @@ -259,7 +266,7 @@ void rotate_reclaimable_page(struct page *page)
> local_lock_irqsave(&lru_rotate.lock, flags);
> pvec = this_cpu_ptr(&lru_rotate.pvec);
> if (!pagevec_add(pvec, page) || PageCompound(page))
> - pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
> + pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, false);
> local_unlock_irqrestore(&lru_rotate.lock, flags);
> }
> }
> @@ -328,7 +335,7 @@ static void activate_page_drain(int cpu)
> struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
>
> if (pagevec_count(pvec))
> - pagevec_lru_move_fn(pvec, __activate_page);
> + pagevec_lru_move_fn(pvec, __activate_page, false);
> }
>
> static bool need_activate_page_drain(int cpu)
> @@ -346,7 +353,7 @@ void activate_page(struct page *page)
> pvec = this_cpu_ptr(&lru_pvecs.activate_page);
> get_page(page);
> if (!pagevec_add(pvec, page) || PageCompound(page))
> - pagevec_lru_move_fn(pvec, __activate_page);
> + pagevec_lru_move_fn(pvec, __activate_page, false);
> local_unlock(&lru_pvecs.lock);
> }
> }
> @@ -621,21 +628,21 @@ void lru_add_drain_cpu(int cpu)
>
> /* No harm done if a racing interrupt already did this */
> local_lock_irqsave(&lru_rotate.lock, flags);
> - pagevec_lru_move_fn(pvec, pagevec_move_tail_fn);
> + pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, false);
> local_unlock_irqrestore(&lru_rotate.lock, flags);
> }
>
> pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
> if (pagevec_count(pvec))
> - pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
> + pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, false);
>
> pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
> if (pagevec_count(pvec))
> - pagevec_lru_move_fn(pvec, lru_deactivate_fn);
> + pagevec_lru_move_fn(pvec, lru_deactivate_fn, false);
>
> pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
> if (pagevec_count(pvec))
> - pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
> + pagevec_lru_move_fn(pvec, lru_lazyfree_fn, false);
>
> activate_page_drain(cpu);
> }
> @@ -664,7 +671,7 @@ void deactivate_file_page(struct page *page)
> pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
>
> if (!pagevec_add(pvec, page) || PageCompound(page))
> - pagevec_lru_move_fn(pvec, lru_deactivate_file_fn);
> + pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, false);
> local_unlock(&lru_pvecs.lock);
> }
> }
> @@ -686,7 +693,7 @@ void deactivate_page(struct page *page)
> pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
> get_page(page);
> if (!pagevec_add(pvec, page) || PageCompound(page))
> - pagevec_lru_move_fn(pvec, lru_deactivate_fn);
> + pagevec_lru_move_fn(pvec, lru_deactivate_fn, false);
> local_unlock(&lru_pvecs.lock);
> }
> }
> @@ -708,7 +715,7 @@ void mark_page_lazyfree(struct page *page)
> pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
> get_page(page);
> if (!pagevec_add(pvec, page) || PageCompound(page))
> - pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
> + pagevec_lru_move_fn(pvec, lru_lazyfree_fn, false);
> local_unlock(&lru_pvecs.lock);
> }
> }
> @@ -976,7 +983,7 @@ static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec)
> */
> void __pagevec_lru_add(struct pagevec *pvec)
> {
> - pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn);
> + pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, true);
> }
It seems better to open code version in lru_add than adding a bool
argument which is true just for one user.
Also with this new lru protection logic lru_add could be optimized:
It could prepare a list of pages and under lru_lock do only list
splice and bumping counter.
Since PageLRU isn't set yet nobody could touch these pages in lru.
After that lru_add could iterate pages from first to last without
lru_lock to set PageLRU and drop reference.
So, lru_add will do O(1) operations under lru_lock regardless of the
count of pages it added.
Actually per-cpu vector for adding could be replaced with per-cpu
lists and\or per-lruvec atomic slist.
Thus incommig pages will be already in list structure rather than page vector.
This allows to accumulate more pages and offload adding to kswapd or
direct reclaim.
>
> /**
> --
> 1.8.3.1
>
>
在 2020/7/3 下午5:13, Konstantin Khlebnikov 写道:
>> @@ -976,7 +983,7 @@ static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec)
>> */
>> void __pagevec_lru_add(struct pagevec *pvec)
>> {
>> - pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn);
>> + pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, true);
>> }
> It seems better to open code version in lru_add than adding a bool
> argument which is true just for one user.
Right, I will rewrite this part as your suggestion. Thanks!
>
> Also with this new lru protection logic lru_add could be optimized:
> It could prepare a list of pages and under lru_lock do only list
> splice and bumping counter.
> Since PageLRU isn't set yet nobody could touch these pages in lru.
> After that lru_add could iterate pages from first to last without
> lru_lock to set PageLRU and drop reference.
>
> So, lru_add will do O(1) operations under lru_lock regardless of the
> count of pages it added.
>
> Actually per-cpu vector for adding could be replaced with per-cpu
> lists and\or per-lruvec atomic slist.
> Thus incommig pages will be already in list structure rather than page vector.
> This allows to accumulate more pages and offload adding to kswapd or
> direct reclaim.
>
That's a great idea! Guess what the new struct we need would be like this?
I like to try this. :)
diff --git a/include/linux/pagevec.h b/include/linux/pagevec.h
index 081d934eda64..d62778c8c184 100644
--- a/include/linux/pagevec.h
+++ b/include/linux/pagevec.h
@@ -20,7 +20,7 @@
struct pagevec {
unsigned char nr;
bool percpu_pvec_drained;
- struct page *pages[PAGEVEC_SIZE];
+ struct list_head veclist;
};
On Sat, Jul 04, 2020 at 07:34:59PM +0800, Alex Shi wrote:
> That's a great idea! Guess what the new struct we need would be like this?
> I like to try this. :)
>
>
> diff --git a/include/linux/pagevec.h b/include/linux/pagevec.h
> index 081d934eda64..d62778c8c184 100644
> --- a/include/linux/pagevec.h
> +++ b/include/linux/pagevec.h
> @@ -20,7 +20,7 @@
> struct pagevec {
> unsigned char nr;
> bool percpu_pvec_drained;
> - struct page *pages[PAGEVEC_SIZE];
> + struct list_head veclist;
> };
pagevecs are used not just for LRU. If you want to use a list_head for
LRU then define a new structure.
?? 2020/7/4 ????7:39, Matthew Wilcox д??:
> On Sat, Jul 04, 2020 at 07:34:59PM +0800, Alex Shi wrote:
>> That's a great idea! Guess what the new struct we need would be like this?
>> I like to try this. :)
>>
>>
>> diff --git a/include/linux/pagevec.h b/include/linux/pagevec.h
>> index 081d934eda64..d62778c8c184 100644
>> --- a/include/linux/pagevec.h
>> +++ b/include/linux/pagevec.h
>> @@ -20,7 +20,7 @@
>> struct pagevec {
>> unsigned char nr;
>> bool percpu_pvec_drained;
>> - struct page *pages[PAGEVEC_SIZE];
>> + struct list_head veclist;
>> };
>
> pagevecs are used not just for LRU. If you want to use a list_head for
> LRU then define a new structure.
>
yes, there are much page don't use page->lru, like slab etc. we need a new struct.
Thanks!
Alex
On Sat, Jul 04, 2020 at 09:12:46PM +0800, Alex Shi wrote:
> 在 2020/7/4 下午7:39, Matthew Wilcox 写道:
> > On Sat, Jul 04, 2020 at 07:34:59PM +0800, Alex Shi wrote:
> >> That's a great idea! Guess what the new struct we need would be like this?
> >> I like to try this. :)
> >>
> >>
> >> diff --git a/include/linux/pagevec.h b/include/linux/pagevec.h
> >> index 081d934eda64..d62778c8c184 100644
> >> --- a/include/linux/pagevec.h
> >> +++ b/include/linux/pagevec.h
> >> @@ -20,7 +20,7 @@
> >> struct pagevec {
> >> unsigned char nr;
> >> bool percpu_pvec_drained;
> >> - struct page *pages[PAGEVEC_SIZE];
> >> + struct list_head veclist;
> >> };
> >
> > pagevecs are used not just for LRU. If you want to use a list_head for
> > LRU then define a new structure.
>
> yes, there are much page don't use page->lru, like slab etc. we need a new struct.
That's not what I mean. Slab pages aren't on the LRU anyway.
Consider the callers of page_cache_delete_batch(). These use a pagevec
for a non-LRU purpose, and they will be much slower with a list_head than
with an array.
在 2020/7/4 下午9:33, Matthew Wilcox 写道:
> On Sat, Jul 04, 2020 at 09:12:46PM +0800, Alex Shi wrote:
>> 在 2020/7/4 下午7:39, Matthew Wilcox 写道:
>>> On Sat, Jul 04, 2020 at 07:34:59PM +0800, Alex Shi wrote:
>>>> That's a great idea! Guess what the new struct we need would be like this?
>>>> I like to try this. :)
>>>>
>>>>
>>>> diff --git a/include/linux/pagevec.h b/include/linux/pagevec.h
>>>> index 081d934eda64..d62778c8c184 100644
>>>> --- a/include/linux/pagevec.h
>>>> +++ b/include/linux/pagevec.h
>>>> @@ -20,7 +20,7 @@
>>>> struct pagevec {
>>>> unsigned char nr;
>>>> bool percpu_pvec_drained;
>>>> - struct page *pages[PAGEVEC_SIZE];
>>>> + struct list_head veclist;
>>>> };
>>>
>>> pagevecs are used not just for LRU. If you want to use a list_head for
>>> LRU then define a new structure.
>>
>> yes, there are much page don't use page->lru, like slab etc. we need a new struct>
> That's not what I mean. Slab pages aren't on the LRU anyway.
Right. I mean, that's reason for a new struct if we change to list.
>
> Consider the callers of page_cache_delete_batch(). These use a pagevec
> for a non-LRU purpose, and they will be much slower with a list_head than
> with an array.
>
Thanks for the info.
If the list is slower than pagevec, maybe it's not worth to do the change.
Since pagevec could handle any kind of pages, anon/file, non-active/active, but one
list only fit for just one kind of list. 5 kinds of list adding would increase the
complexity. Consider this, I am wondering if it's worth?
Thanks
Alex