In the badness() calculation, there's currently this piece of code:

        /*
         * Processes which fork a lot of child processes are likely
         * a good choice. We add the vmsize of the children if they
         * have an own mm. This prevents forking servers to flood the
         * machine with an endless amount of children
         */
        list_for_each(tsk, &p->children) {
                struct task_struct *chld;
                chld = list_entry(tsk, struct task_struct, sibling);
                if (chld->mm = p->mm && chld->mm)
                        points += chld->mm->total_vm;
        }

The intention is clear: If some server (apache) keeps spawning new children
and we run OOM, we want to kill the father rather than picking a child.

This -- to some degree -- also helps a bit with getting fork bombs under
control, though I'd consider this a desirable side-effect rather than a
feature.

There's one problem with this: No matter how many or few children there are,
if just one of them misbehaves, and all others (including the father) do
everything right, we still always kill the whole family.  This hits in real
life; whether it's javascript in konqueror resulting in kdeinit (and thus the
whole KDE session) being hit or just a classical server that spawns children.

Sidenote: The killer does kill all direct children as well, not only the
selected father, see oom_kill_process().

The idea in attached patch is that we do want to account the memory
consumption of the (direct) children to the father -- however not fully.
This maintains the property that fathers with too many children will still
very likely be picked, whereas a single misbehaving child has the chance to
be picked by the OOM killer.

In the patch I account only half (rounded up) of the children's vm_size to
the parent.  This means that if one child eats more mem than the rest of
the family, it will be picked, otherwise it's still the father and thus the
whole family that gets selected.

This is heuristics -- we could debate whether accounting for a fourth would
be better than for half of it.  Or -- if people would consider it worth the
trouble -- make it a sysctl.  For now I sticked to accounting for half,
which should IMHO be a significant improvement.

The patch does one more thing: As users tend to be irritated by the choice
of killed processes (mainly because the children are killed first, despite
some of them having a very low OOM score), I added some more output: The
selected (father) process will be reported first and it's oom_score printed
to syslog.

Description:

Only account for half of children's vm size in oom score calculation

This should still give the parent enough point in case of fork bombs.  If
any child however has more than 50% of the vm size of all children
together, it'll get a higher score and be elected.

This patch also makes the kernel display the oom_score.
Signed-off-by: NKurt Garloff <garloff@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Sometimes it's nice to know who's calling.
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

The problem, reported in:

  http://bugzilla.kernel.org/show_bug.cgi?id=5859

and by various other email messages and lkml posts is that the cpuset hook
in the oom (out of memory) code can try to take a cpuset semaphore while
holding the tasklist_lock (a spinlock).

One must not sleep while holding a spinlock.

The fix seems easy enough - move the cpuset semaphore region outside the
tasklist_lock region.

This required a few lines of mechanism to implement.  The oom code where
the locking needs to be changed does not have access to the cpuset locks,
which are internal to kernel/cpuset.c only.  So I provided a couple more
cpuset interface routines, available to the rest of the kernel, which
simple take and drop the lock needed here (cpusets callback_sem).
Signed-off-by: NPaul Jackson <pj@sgi.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

When oom_killer kills current there's no need to call
schedule_timeout_interruptible() since task must die ASAP.
Signed-Off-By: NPavel Emelianov <xemul@sw.ru>
Signed-Off-By: NKirill Korotaev <dev@openvz.org>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

 - added typedef unsigned int __nocast gfp_t;

 - replaced __nocast uses for gfp flags with gfp_t - it gives exactly
   the same warnings as far as sparse is concerned, doesn't change
   generated code (from gcc point of view we replaced unsigned int with
   typedef) and documents what's going on far better.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Use schedule_timeout_{,un}interruptible() instead of
set_current_state()/schedule_timeout() to reduce kernel size.
Signed-off-by: NNishanth Aravamudan <nacc@us.ibm.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Now the real motivation for this cpuset mem_exclusive patch series seems
trivial.

This patch keeps a task in or under one mem_exclusive cpuset from provoking an
oom kill of a task under a non-overlapping mem_exclusive cpuset.  Since only
interrupt and GFP_ATOMIC allocations are allowed to escape mem_exclusive
containment, there is little to gain from oom killing a task under a
non-overlapping mem_exclusive cpuset, as almost all kernel and user memory
allocation must come from disjoint memory nodes.

This patch enables configuring a system so that a runaway job under one
mem_exclusive cpuset cannot cause the killing of a job in another such cpuset
that might be using very high compute and memory resources for a prolonged
time.
Signed-off-by: NPaul Jackson <pj@sgi.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

This patch series extends the use of the cpuset attribute 'mem_exclusive'
to support cpuset configurations that:
 1) allow GFP_KERNEL allocations to come from a potentially larger
    set of memory nodes than GFP_USER allocations, and
 2) can constrain the oom killer to tasks running in cpusets in
    a specified subtree of the cpuset hierarchy.

Here's an example usage scenario.  For a few hours or more, a large NUMA
system at a University is to be divided in two halves, with a bunch of student
jobs running in half the system under some form of batch manager, and with a
big research project running in the other half.  Each of the student jobs is
placed in a small cpuset, but should share the classic Unix time share
facilities, such as buffered pages of files in /bin and /usr/lib.  The big
research project wants no interference whatsoever from the student jobs, and
has highly tuned, unusual memory and i/o patterns that intend to make full use
of all the main memory on the nodes available to it.

In this example, we have two big sibling cpusets, one of which is further
divided into a more dynamic set of child cpusets.

We want kernel memory allocations constrained by the two big cpusets, and user
allocations constrained by the smaller child cpusets where present.  And we
require that the oom killer not operate across the two halves of this system,
or else the first time a student job runs amuck, the big research project will
likely be first inline to get shot.

Tweaking /proc/<pid>/oom_adj is not ideal -- if the big research project
really does run amuck allocating memory, it should be shot, not some other
task outside the research projects mem_exclusive cpuset.

I propose to extend the use of the 'mem_exclusive' flag of cpusets to manage
such scenarios.  Let memory allocations for user space (GFP_USER) be
constrained by a tasks current cpuset, but memory allocations for kernel space
(GFP_KERNEL) by constrained by the nearest mem_exclusive ancestor of the
current cpuset, even though kernel space allocations will still _prefer_ to
remain within the current tasks cpuset, if memory is easily available.

Let the oom killer be constrained to consider only tasks that are in
overlapping mem_exclusive cpusets (it won't help much to kill a task that
normally cannot allocate memory on any of the same nodes as the ones on which
the current task can allocate.)

The current constraints imposed on setting mem_exclusive are unchanged.  A
cpuset may only be mem_exclusive if its parent is also mem_exclusive, and a
mem_exclusive cpuset may not overlap any of its siblings memory nodes.

This patch was presented on linux-mm in early July 2005, though did not
generate much feedback at that time.  It has been built for a variety of
arch's using cross tools, and built, booted and tested for function on SN2
(ia64).

There are 4 patches in this set:
  1) Some minor cleanup, and some improvements to the code layout
     of one routine to make subsequent patches cleaner.
  2) Add another GFP flag - __GFP_HARDWALL.  It marks memory
     requests for USER space, which are tightly confined by the
     current tasks cpuset.
  3) Now memory requests (such as KERNEL) that not marked HARDWALL can
     if short on memory, look in the potentially larger pool of memory
     defined by the nearest mem_exclusive ancestor cpuset of the current
     tasks cpuset.
  4) Finally, modify the oom killer to skip any task whose mem_exclusive
     cpuset doesn't overlap ours.

Patch (1), the one time I looked on an SN2 (ia64) build, actually saved 32
bytes of kernel text space.  Patch (2) has no affect on the size of kernel
text space (it just adds a preprocessor flag).  Patches (3) and (4) added
about 600 bytes each of kernel text space, mostly in kernel/cpuset.c, which
matters only if CONFIG_CPUSET is enabled.

This patch:

This patch applies a few comment and code cleanups to mm/oom_kill.c prior to
applying a few small patches to improve cpuset management of memory placement.

The comment changed in oom_kill.c was seriously misleading.  The code layout
change in select_bad_process() makes room for adding another condition on
which a process can be spared the oom killer (see the subsequent
cpuset_nodes_overlap patch for this addition).

Also a couple typos and spellos that bugged me, while I was here.

This patch should have no material affect.
Signed-off-by: NPaul Jackson <pj@sgi.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

We now print statistics when invoking the OOM killer, however this
information is not rate limited and you can get into situations where the
console is continually spammed.

For example, when a task is exiting the OOM killer will simply return
(waiting for that task to exit and clear up memory).  If the VM continually
calls back into the OOM killer we get thousands of copies of show_mem() on
the console.

Use printk_ratelimit() to quieten it.
Signed-off-by: NAnton Blanchard <anton@samba.org>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Dump the current allocation order when OOM killing.
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

This patch provides more debug info when the system is OOM.  It displays
memory stats (basically sysrq-m info) from __alloc_pages() when page
allocation fails and during OOM kill.

Thanks to Dave Jones for coming up with the idea.
Signed-off-by: NJanet Morgan <janetmor@us.ibm.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

iscsi/lvm2/multipath needs guaranteed protection from the oom-killer, so
make the magical value of -17 in /proc/<pid>/oom_adj defeat the oom-killer
altogether.

(akpm: we still need to document oom_adj and friends in
Documentation/filesystems/proc.txt!)
Signed-off-by: NAndrea Arcangeli <andrea@suse.de>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.

Let it rip!