Add user_reserve_kbytes knob.

Limit the growth of the memory reserved for other user processes to
min(3% current process size, user_reserve_pages).  Only about 8MB is
necessary to enable recovery in the default mode, and only a few hundred
MB are required even when overcommit is disabled.

user_reserve_pages defaults to min(3% free pages, 128MB)

I arrived at 128MB by taking the max VSZ of sshd, login, bash, and top ...
then adding the RSS of each.

This only affects OVERCOMMIT_NEVER mode.

Background

1. user reserve

__vm_enough_memory reserves a hardcoded 3% of the current process size for
other applications when overcommit is disabled.  This was done so that a
user could recover if they launched a memory hogging process.  Without the
reserve, a user would easily run into a message such as:

bash: fork: Cannot allocate memory

2. admin reserve

Additionally, a hardcoded 3% of free memory is reserved for root in both
overcommit 'guess' and 'never' modes.  This was intended to prevent a
scenario where root-cant-log-in and perform recovery operations.

Note that this reserve shrinks, and doesn't guarantee a useful reserve.

Motivation

The two hardcoded memory reserves should be updated to account for current
memory sizes.

Also, the admin reserve would be more useful if it didn't shrink too much.

When the current code was originally written, 1GB was considered
"enterprise".  Now the 3% reserve can grow to multiple GB on large memory
systems, and it only needs to be a few hundred MB at most to enable a user
or admin to recover a system with an unwanted memory hogging process.

I've found that reducing these reserves is especially beneficial for a
specific type of application load:

 * single application system
 * one or few processes (e.g. one per core)
 * allocating all available memory
 * not initializing every page immediately
 * long running

I've run scientific clusters with this sort of load.  A long running job
sometimes failed many hours (weeks of CPU time) into a calculation.  They
weren't initializing all of their memory immediately, and they weren't
using calloc, so I put systems into overcommit 'never' mode.  These
clusters run diskless and have no swap.

However, with the current reserves, a user wishing to allocate as much
memory as possible to one process may be prevented from using, for
example, almost 2GB out of 32GB.

The effect is less, but still significant when a user starts a job with
one process per core.  I have repeatedly seen a set of processes
requesting the same amount of memory fail because one of them could not
allocate the amount of memory a user would expect to be able to allocate.
For example, Message Passing Interfce (MPI) processes, one per core.  And
it is similar for other parallel programming frameworks.

Changing this reserve code will make the overcommit never mode more useful
by allowing applications to allocate nearly all of the available memory.

Also, the new admin_reserve_kbytes will be safer than the current behavior
since the hardcoded 3% of available memory reserve can shrink to something
useless in the case where applications have grabbed all available memory.

Risks

* "bash: fork: Cannot allocate memory"

  The downside of the first patch-- which creates a tunable user reserve
  that is only used in overcommit 'never' mode--is that an admin can set
  it so low that a user may not be able to kill their process, even if
  they already have a shell prompt.

  Of course, a user can get in the same predicament with the current 3%
  reserve--they just have to launch processes until 3% becomes negligible.

* root-cant-log-in problem

  The second patch, adding the tunable rootuser_reserve_pages, allows
  the admin to shoot themselves in the foot by setting it too small.  They
  can easily get the system into a state where root-can't-log-in.

  However, the new admin_reserve_kbytes will be safer than the current
  behavior since the hardcoded 3% of available memory reserve can shrink
  to something useless in the case where applications have grabbed all
  available memory.

Alternatives

 * Memory cgroups provide a more flexible way to limit application memory.

   Not everyone wants to set up cgroups or deal with their overhead.

 * We could create a fourth overcommit mode which provides smaller reserves.

   The size of useful reserves may be drastically different depending
   on the whether the system is embedded or enterprise.

 * Force users to initialize all of their memory or use calloc.

   Some users don't want/expect the system to overcommit when they malloc.
   Overcommit 'never' mode is for this scenario, and it should work well.

The new user and admin reserve tunables are simple to use, with low
overhead compared to cgroups.  The patches preserve current behavior where
3% of memory is less than 128MB, except that the admin reserve doesn't
shrink to an unusable size under pressure.  The code allows admins to tune
for embedded and enterprise usage.

FAQ

 * How is the root-cant-login problem addressed?
   What happens if admin_reserve_pages is set to 0?

   Root is free to shoot themselves in the foot by setting
   admin_reserve_kbytes too low.

   On x86_64, the minimum useful reserve is:
     8MB for overcommit 'guess'
   128MB for overcommit 'never'

   admin_reserve_pages defaults to min(3% free memory, 8MB)

   So, anyone switching to 'never' mode needs to adjust
   admin_reserve_pages.

 * How do you calculate a minimum useful reserve?

   A user or the admin needs enough memory to login and perform
   recovery operations, which includes, at a minimum:

   sshd or login + bash (or some other shell) + top (or ps, kill, etc.)

   For overcommit 'guess', we can sum resident set sizes (RSS)
   because we only need enough memory to handle what the recovery
   programs will typically use. On x86_64 this is about 8MB.

   For overcommit 'never', we can take the max of their virtual sizes (VSZ)
   and add the sum of their RSS. We use VSZ instead of RSS because mode
   forces us to ensure we can fulfill all of the requested memory allocations--
   even if the programs only use a fraction of what they ask for.
   On x86_64 this is about 128MB.

   When swap is enabled, reserves are useful even when they are as
   small as 10MB, regardless of overcommit mode.

   When both swap and overcommit are disabled, then the admin should
   tune the reserves higher to be absolutley safe. Over 230MB each
   was safest in my testing.

 * What happens if user_reserve_pages is set to 0?

   Note, this only affects overcomitt 'never' mode.

   Then a user will be able to allocate all available memory minus
   admin_reserve_kbytes.

   However, they will easily see a message such as:

   "bash: fork: Cannot allocate memory"

   And they won't be able to recover/kill their application.
   The admin should be able to recover the system if
   admin_reserve_kbytes is set appropriately.

 * What's the difference between overcommit 'guess' and 'never'?

   "Guess" allows an allocation if there are enough free + reclaimable
   pages. It has a hardcoded 3% of free pages reserved for root.

   "Never" allows an allocation if there is enough swap + a configurable
   percentage (default is 50) of physical RAM. It has a hardcoded 3% of
   free pages reserved for root, like "Guess" mode. It also has a
   hardcoded 3% of the current process size reserved for additional
   applications.

 * Why is overcommit 'guess' not suitable even when an app eventually
   writes to every page? It takes free pages, file pages, available
   swap pages, reclaimable slab pages into consideration. In other words,
   these are all pages available, then why isn't overcommit suitable?

   Because it only looks at the present state of the system. It
   does not take into account the memory that other applications have
   malloced, but haven't initialized yet. It overcommits the system.

Test Summary

There was little change in behavior in the default overcommit 'guess'
mode with swap enabled before and after the patch. This was expected.

Systems run most predictably (i.e. no oom kills) in overcommit 'never'
mode with swap enabled. This also allowed the most memory to be allocated
to a user application.

Overcommit 'guess' mode without swap is a bad idea. It is easy to
crash the system. None of the other tested combinations crashed.
This matches my experience on the Roadrunner supercomputer.

Without the tunable user reserve, a system in overcommit 'never' mode
and without swap does not allow the admin to recover, although the
admin can.

With the new tunable reserves, a system in overcommit 'never' mode
and without swap can be configured to:

1. maximize user-allocatable memory, running close to the edge of
recoverability

2. maximize recoverability, sacrificing allocatable memory to
ensure that a user cannot take down a system

Test Description

Fedora 18 VM - 4 x86_64 cores, 5725MB RAM, 4GB Swap

System is booted into multiuser console mode, with unnecessary services
turned off. Caches were dropped before each test.

Hogs are user memtester processes that attempt to allocate all free memory
as reported by /proc/meminfo

In overcommit 'never' mode, memory_ratio=100

Test Results

3.9.0-rc1-mm1

Overcommit | Swap | Hogs | MB Got/Wanted | OOMs | User Recovery | Admin Recovery
----------   ----   ----   -------------   ----   -------------   --------------
guess        yes    1      5432/5432       no     yes             yes
guess        yes    4      5444/5444       1      yes             yes
guess        no     1      5302/5449       no     yes             yes
guess        no     4      -               crash  no              no

never        yes    1      5460/5460       1      yes             yes
never        yes    4      5460/5460       1      yes             yes
never        no     1      5218/5432       no     no              yes
never        no     4      5203/5448       no     no              yes

3.9.0-rc1-mm1-tunablereserves

User and Admin Recovery show their respective reserves, if applicable.

Overcommit | Swap | Hogs | MB Got/Wanted | OOMs | User Recovery | Admin Recovery
----------   ----   ----   -------------   ----   -------------   --------------
guess        yes    1      5419/5419       no     - yes           8MB yes
guess        yes    4      5436/5436       1      - yes           8MB yes
guess        no     1      5440/5440       *      - yes           8MB yes
guess        no     4      -               crash  - no            8MB no

* process would successfully mlock, then the oom killer would pick it

never        yes    1      5446/5446       no     10MB yes        20MB yes
never        yes    4      5456/5456       no     10MB yes        20MB yes
never        no     1      5387/5429       no     128MB no        8MB barely
never        no     1      5323/5428       no     226MB barely    8MB barely
never        no     1      5323/5428       no     226MB barely    8MB barely

never        no     1      5359/5448       no     10MB no         10MB barely

never        no     1      5323/5428       no     0MB no          10MB barely
never        no     1      5332/5428       no     0MB no          50MB yes
never        no     1      5293/5429       no     0MB no          90MB yes

never        no     1      5001/5427       no     230MB yes       338MB yes
never        no     4*     4998/5424       no     230MB yes       338MB yes

* more memtesters were launched, able to allocate approximately another 100MB

Future Work

 - Test larger memory systems.

 - Test an embedded image.

 - Test other architectures.

 - Time malloc microbenchmarks.

 - Would it be useful to be able to set overcommit policy for
   each memory cgroup?

 - Some lines are slightly above 80 chars.
   Perhaps define a macro to convert between pages and kb?
   Other places in the kernel do this.

[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: make init_user_reserve() static]
Signed-off-by: Andrew Shewmaker <agshew@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>