Re: Fwd: Control page reclaim granularity

From: Konstantin Khlebnikov
Date: Mon Mar 12 2012 - 10:18:30 EST


Minchan Kim wrote:
On Mon, Mar 12, 2012 at 04:14:14PM +0800, Zheng Liu wrote:
On 03/12/2012 02:20 PM, Konstantin Khlebnikov wrote:
Minchan Kim wrote:
On Mon, Mar 12, 2012 at 10:06:09AM +0800, Zheng Liu wrote:
On Mon, Mar 12, 2012 at 09:29:34AM +0900, Minchan Kim wrote:
I forgot to Ccing you.
Sorry.

---------- Forwarded message ----------
From: Minchan Kim<minchan@xxxxxxxxxx>
Date: Mon, Mar 12, 2012 at 9:28 AM
Subject: Re: Control page reclaim granularity
To: Minchan Kim<minchan@xxxxxxxxxx>, linux-mm<linux-mm@xxxxxxxxx>,
linux-kernel<linux-kernel@xxxxxxxxxxxxxxx>, Konstantin Khlebnikov<
khlebnikov@xxxxxxxxxx>, riel@xxxxxxxxxx, kosaki.motohiro@xxxxxxxxxxxxxx


On Fri, Mar 09, 2012 at 12:54:03AM +0800, Zheng Liu wrote:
Hi Minchan,

Sorry, I forgot to say that I don't subscribe linux-mm and
linux-kernel
mailing list. So please Cc me.

IMHO, maybe we should re-think about how does user use mmap(2). I
describe the cases I known in our product system. They can be
categorized into two cases. One is mmaped all data files into memory
and sometime it uses write(2) to append some data, and another uses
mmap(2)/munmap(2) and read(2)/write(2) to manipulate the files. In
the
second case, the application wants to keep mmaped page into memory
and
let file pages to be reclaimed firstly. So, IMO, when application
uses
mmap(2) to manipulate files, it is possible to imply that it wants
keep
these mmaped pages into memory and do not be reclaimed. At least
these
pages do not be reclaimed early than file pages. I think that
maybe we
can recover that routine and provide a sysctl parameter to let the
user
to set this ratio between mmaped pages and file pages.

I am not convinced why we should handle mapped page specially.
Sometimem, someone may use mmap by reducing buffer copy compared to
read
system call.
So I think we can't make sure mmaped pages are always win.

My suggestion is that it would be better to declare by user explicitly.
I think we can implement it by madvise and fadvise's WILLNEED option.
Current implementation is just readahead if there isn't a page in
memory
but I think
we can promote from inactive to active if there is already a page in
memory.

It's more clear and it couldn't be affected by kernel page reclaim
algorithm change
like this.

Thank you for your advice. But I still have question about this
solution. If we improve the madvise(2) and fadvise(2)'s WILLNEED
option, it will cause an inconsistently status for pages that be
manipulated by madvise(2) and/or fadvise(2). For example, when I call
madvise with WILLNEED flag, some pages will be moved into active list if
they already have been in memory, and other pages will be read into
memory and be saved in inactive list if they don't be in memory. Then
pages that are in inactive list are possible to be reclaim. So from the
view of users, it is inconsistent because some pages are in memory and
some pages are reclaimed. But actually the user hopes that all of pages
can be kept in memory. IMHO, this inconsistency is weird and makes
users
puzzled.

Now problem is that

1. User want to keep pages which are used once in a while in memory.
2. Kernel want to reclaim them because they are surely reclaim target
pages in point of view by LRU.

The most desriable approach is that user should use mlock to guarantee
them in memory. But mlock is too big overhead and user doesn't want to
keep
memory all pages all at once.(Ie, he want demand paging when he need
the page)
Right?

madvise, it's a just hint for kernel and kernel doesn't need to make
sure madvise's behavior.
In point of view, such inconsistency might not be a big problem.

Big problem I think now is that user should use madvise(WILLNEED)
periodically because such
activation happens once when user calls madvise. If user doesn't use
page frequently after
user calls it, it ends up moving into inactive list and even could be
reclaimed.
It's not good. :-(

Okay. How about adding new VM_WORKINGSET?
And reclaimer would give one more round trip in active/inactive list
erwhen reclaim happens
if the page is referenced.

Sigh. We have no room for new VM_FLAG in 32 bit.
p
It would be nice to mark struct address_space with this flag and export
AS_UNEVICTABLE somehow.
Maybe we can reuse file-locking engine for managing these bits =)

Make sense to me. We can mark this flag in struct address_space and check
it in page_refereneced_file(). If this flag is set, it will be cleard and

Disadvantage is that we could set reclaim granularity as per-inode.
I want to set it as per-vma, not per-inode.

But with per-inode flag we can tune all files, not only memory-mapped.
See, attached patch. Currently I thinking about managing code,
file-locking engine really fits perfectly =)


the function returns referenced> 1. Then this page can be promoted into
activate list. But I prefer to set/clear this flag in madvise.

Hmm, My idea is following as,
If we can set new VM flag into VMA or something, reclaimer can check it when shrink_[in]active_list
and he can prevent to deactivate/reclaim if he takes a look the page is in VMA which
are set by new VM flag and the page is referenced recently at least once.
It means it gives one more round trip in his list(ie, active/inactive list)
rather than activation so that the page would become less reclaimable.


PS, I have subscribed linux-mm mailing list. :-)

Congratulations! :)


Regards,
Zheng

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email:<a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx</a>

mm: introduce mapping AS_WORKINGSET flag

From: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxx>

This patch introduces new flag AS_WORKINGSET in mapping->flags.
If it set reclaimer will activates all pages for this inode after first usage.

Signed-off-by: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxx>
---
include/linux/pagemap.h | 16 ++++++++++++++++
mm/vmscan.c | 15 ++++++++++++---
2 files changed, 28 insertions(+), 3 deletions(-)

diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
index cfaaa69..c15fc17 100644
--- a/include/linux/pagemap.h
+++ b/include/linux/pagemap.h
@@ -24,6 +24,7 @@ enum mapping_flags {
AS_ENOSPC = __GFP_BITS_SHIFT + 1, /* ENOSPC on async write */
AS_MM_ALL_LOCKS = __GFP_BITS_SHIFT + 2, /* under mm_take_all_locks() */
AS_UNEVICTABLE = __GFP_BITS_SHIFT + 3, /* e.g., ramdisk, SHM_LOCK */
+ AS_WORKINGSET = __GFP_BITS_SHIFT + 4, /* promote pages activation */
};

static inline void mapping_set_error(struct address_space *mapping, int error)
@@ -53,6 +54,21 @@ static inline int mapping_unevictable(struct address_space *mapping)
return !!mapping;
}

+static inline void mapping_set_workingset(struct address_space *mapping)
+{
+ set_bit(AS_WORKINGSET, &mapping->flags);
+}
+
+static inline void mapping_clear_workingset(struct address_space *mapping)
+{
+ clear_bit(AS_WORKINGSET, &mapping->flags);
+}
+
+static inline int mapping_test_workingset(struct address_space *mapping)
+{
+ return mapping && test_bit(AS_WORKINGSET, &mapping->flags);
+}
+
static inline gfp_t mapping_gfp_mask(struct address_space * mapping)
{
return (__force gfp_t)mapping->flags & __GFP_BITS_MASK;
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 57b9658..5ccbe8c 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -701,6 +701,7 @@ enum page_references {
};

static enum page_references page_check_references(struct page *page,
+ struct address_space *mapping,
struct mem_cgroup_zone *mz,
struct scan_control *sc)
{
@@ -721,6 +722,13 @@ static enum page_references page_check_references(struct page *page,
if (vm_flags & VM_LOCKED)
return PAGEREF_RECLAIM;

+ /*
+ * Activate workingset page if referenced at least once.
+ */
+ if (mapping_test_workingset(mapping) &&
+ (referenced_ptes || referenced_page))
+ return PAGEREF_ACTIVATE;
+
if (referenced_ptes) {
if (PageAnon(page))
return PAGEREF_ACTIVATE;
@@ -828,7 +836,9 @@ static unsigned long shrink_page_list(struct list_head *page_list,
}
}

- references = page_check_references(page, mz, sc);
+ mapping = page_mapping(page);
+
+ references = page_check_references(page, mapping, mz, sc);
switch (references) {
case PAGEREF_ACTIVATE:
goto activate_locked;
@@ -848,11 +858,10 @@ static unsigned long shrink_page_list(struct list_head *page_list,
goto keep_locked;
if (!add_to_swap(page))
goto activate_locked;
+ mapping = &swapper_space;
may_enter_fs = 1;
}

- mapping = page_mapping(page);
-
/*
* The page is mapped into the page tables of one or more
* processes. Try to unmap it here.