Date:	Mon, 20 Jul 2015 09:28:10 +0100
From:	Mel Gorman <mgorman@...e.de>
To:	Pintu Kumar <pintu.k@...sung.com>
Cc:	akpm@...ux-foundation.org, corbet@....net, vbabka@...e.cz,
	gorcunov@...nvz.org, mhocko@...e.cz, emunson@...mai.com,
	kirill.shutemov@...ux.intel.com, standby24x7@...il.com,
	hannes@...xchg.org, vdavydov@...allels.com, hughd@...gle.com,
	minchan@...nel.org, tj@...nel.org, rientjes@...gle.com,
	xypron.glpk@....de, dzickus@...hat.com, prarit@...hat.com,
	ebiederm@...ssion.com, rostedt@...dmis.org, uobergfe@...hat.com,
	paulmck@...ux.vnet.ibm.com, iamjoonsoo.kim@....com,
	ddstreet@...e.org, sasha.levin@...cle.com, koct9i@...il.com,
	cj@...ux.com, opensource.ganesh@...il.com, vinmenon@...eaurora.org,
	linux-doc@...r.kernel.org, linux-kernel@...r.kernel.org,
	linux-mm@...ck.org, linux-pm@...r.kernel.org, qiuxishi@...wei.com,
	Valdis.Kletnieks@...edu, cpgs@...sung.com, pintu_agarwal@...oo.com,
	vishnu.ps@...sung.com, rohit.kr@...sung.com, iqbal.ams@...sung.com,
	pintu.ping@...il.com, pintu.k@...look.com
Subject: Re: [PATCH v3 1/1] kernel/sysctl.c: Add /proc/sys/vm/shrink_memory
 feature

On Mon, Jul 20, 2015 at 09:59:04AM +0530, Pintu Kumar wrote:
> This patch provides 2 things:
> 1. Add new control called shrink_memory in /proc/sys/vm/.
> This control can be used to aggressively reclaim memory system-wide
> in one shot from the user space. A value of 1 will instruct the
> kernel to reclaim as much as totalram_pages in the system.
> Example: echo 1 > /proc/sys/vm/shrink_memory
> 
> If any other value than 1 is written to shrink_memory an error EINVAL
> occurs.
> 
> 2. Enable shrink_all_memory API in kernel with new CONFIG_SHRINK_MEMORY.
> Currently, shrink_all_memory function is used only during hibernation.
> With the new config we can make use of this API for non-hibernation case
> also without disturbing the hibernation case.
> 
> The detailed paper was presented in Embedded Linux Conference, Mar-2015
> http://events.linuxfoundation.org/sites/events/files/slides/
> %5BELC-2015%5D-System-wide-Memory-Defragmenter.pdf
> 

Johannes has already reviewed this series and explained why it's a bad
idea. This is just a note to say that I agree the points he made and also
think that adding an additional knob to reclaim data from user space is a bad
idea. Even drop_caches is only intended as a debugging tool to illustrate
cases where normal reclaim is broken. Similarly compact_node exists as a
debugging tool to check if direct compaction is not behaving as expected.

If this is invoked when high-order allocations start failing and memory is
fragmented with unreclaimable memory then it'll potentially keep thrashing
depending on the userspace monitor implementation.  If the latency of
a high order allocation is important then reclaim/compaction should be
examined and improved. If the reliability of high-order allocations are
important then you either need to reserve the memory in advance. If that
is undesirable due to a constrained memory environment then one approach
is to modify how pages are grouped by mobility as described in the leader
of the series "Remove zonelist cache and high-order watermark checking".
There are two suggestions there for out-of-tree patches that would make
high-order allocations more reliable that are not suitable for mainline.

Yes, I read your presentation but lets go through the use cases you
list again;

> Various other use cases where this can be used:
> ----------------------------------------------------------------------------
> 1) Just after system boot-up is finished, using the sysctl configuration from
>    bootup script.

Almost no benefit. Any page cache that is active and now cold would be
trivially reclaimed later.

> 2) During system suspend state, after suspend_freeze_processes()
>    [kernel/power/suspend.c]
>    Based on certain condition about fragmentation or free memory state.

No gain.

> 3) From Android ION system heap driver, when order-4 allocation starts failing.
>    By calling shrink_all_memory, in a separate worker thread, based on certain
>    condition.

If order-4 allocations fail when shrink_all_memory works and the order-4
allocation is required to work then the aggressiveness of reclaim/compaction
needs to be fixed to reclaim all system memory if necessary. Right now
it can bail because generally it is expected that no subsystem depends on
high order allocations succeeding for functional correctness.

> 4) It can be combined with compact_memory to achieve better results on memory
>    fragmentation.

Only by reclaiming the world. In 3.0 the system behaved like this. High order
stress tests could take hours to complete as the system was continually
thrashed. Today the same test would complete in about 15 minutes albeit
with lower allocation success rates. We ran into multiple issues where
high order allocation requests caused the system to thrash and triggering
such thrashing from userspace is not an improvement.

> 5) It can be helpful in debugging and tuning various vm parameters.

No more than drop_caches is.

> 6) It can be helpful to identify how much of maximum memory could be
>    reclaimable at any point of time.

Only by reclaiming the world. A less destructive means is using
MemAvailable from /proc/meminfo

>    And how much higher-order pages could be formed with this amount of
>    reclaimable memory.

Only by reclaiming the world

>    Thus it can be helpful in accordingly tuning the reserved memory needs
>    of a system.

By which time it's too late as a reboot will be necessary to set the
reserve.

> 7) It can be helpful in properly tuning the SWAP size in the system.

Only for a single point in time as it's workload dependant. The same
data can be inferred from smaps.

>    In shrink_all_memory, we enable may_swap = 1, that means all unused pages
>    will be swapped out.
>    Thus, running shrink_memory on a heavy loaded system, we can check how much
>    swap is getting full.
>    That can be the maximum swap size with a 10% delta.
>    Also if ZRAM is used, it helps us in compressing and storing the pages for
>    later use.
> 8) It can be helpful to allow more new applications to be launched, without
>    killing the older once.

Reclaim would achieve the same effect over time.

>    And moving the least recently used pages to the SWAP area.
>    Thus user data can be retained.
> 9) Can be part of a system utility to quickly defragment entire system
>    memory.

Any memory that is not on the LRU or indirectly pinned by pages on the
LRU are unaffected.

If high-order allocation latency or reliability is important then you
really need a different solution because unless this thing runs
continually to keep memory unused then it'll eventually fail hard and
the system will perform poorly in the meantime.

-- 
Mel Gorman
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists