This is based on Linus' idea of creating cpu_active_map that prevents
scheduler load balancer from migrating tasks to the cpu that is going
down.
It allows us to simplify domain management code and avoid unecessary
domain rebuilds during cpu hotplug event handling.
Please ignore the cpusets part for now. It needs some more work in order
to avoid crazy lock nesting. Although I did simplfy and unify domain
reinitialization logic. We now simply call partition_sched_domains() in
all the cases. This means that we're using exact same code paths as in
cpusets case and hence the test below cover cpusets too.
Cpuset changes to make rebuild_sched_domains() callable from various
contexts are in the separate patch (right next after this one).
This not only boots but also easily handles
while true; do make clean; make -j 8; done
and
while true; do on-off-cpu 1; done
at the same time.
(on-off-cpu 1 simple does echo 0/1 > /sys/.../cpu1/online thing).
Suprisingly the box (dual-core Core2) is quite usable. In fact I'm typing
this on right now in gnome-terminal and things are moving just fine.
Also this is running with most of the debug features enabled (lockdep,
mutex, etc) no BUG_ONs or lockdep complaints so far.
I believe I addressed all of the Dmitry's comments for original Linus'
version. I changed both fair and rt balancer to mask out non-active cpus.
And replaced cpu_is_offline() with !cpu_active() in the main scheduler
code where it made sense (to me).
Signed-off-by: Max Krasnyanskiy <maxk@qualcomm.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Gregory Haskins <ghaskins@novell.com>
Cc: dmitry.adamushko@gmail.com
Cc: pj@sgi.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The MPOL_BIND policy creates a zonelist that is used for allocations
controlled by that mempolicy. As the per-node zonelist is already being
filtered based on a zone id, this patch adds a version of __alloc_pages() that
takes a nodemask for further filtering. This eliminates the need for
MPOL_BIND to create a custom zonelist.
A positive benefit of this is that allocations using MPOL_BIND now use the
local node's distance-ordered zonelist instead of a custom node-id-ordered
zonelist. I.e., pages will be allocated from the closest allowed node with
available memory.
[Lee.Schermerhorn@hp.com: Mempolicy: update stale documentation and comments]
[Lee.Schermerhorn@hp.com: Mempolicy: make dequeue_huge_page_vma() obey MPOL_BIND nodemask]
[Lee.Schermerhorn@hp.com: Mempolicy: make dequeue_huge_page_vma() obey MPOL_BIND nodemask rework]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* Modify cpuset_cpus_allowed to return the currently allowed cpuset
via a pointer argument instead of as the function return value.
* Use new set_cpus_allowed_ptr function.
* Cleanup CPU_MASK_ALL and NODE_MASK_ALL uses.
Depends on:
[sched-devel]: sched: add new set_cpus_allowed_ptr function
Signed-off-by: Mike Travis <travis@sgi.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Kosaki Motohito noted that "numactl --interleave=all ..." failed in the
presence of memoryless nodes. This patch attempts to fix that problem.
Some background:
numactl --interleave=all calls set_mempolicy(2) with a fully populated
[out to MAXNUMNODES] nodemask. set_mempolicy() [in do_set_mempolicy()]
calls contextualize_policy() which requires that the nodemask be a
subset of the current task's mems_allowed; else EINVAL will be returned.
A task's mems_allowed will always be a subset of node_states[N_HIGH_MEMORY]
i.e., nodes with memory. So, a fully populated nodemask will be
declared invalid if it includes memoryless nodes.
NOTE: the same thing will occur when running in a cpuset
with restricted mem_allowed--for the same reason:
node mask contains dis-allowed nodes.
mbind(2), on the other hand, just masks off any nodes in the nodemask
that are not included in the caller's mems_allowed.
In each case [mbind() and set_mempolicy()], mpol_check_policy() will
complain [again, resulting in EINVAL] if the nodemask contains any
memoryless nodes. This is somewhat redundant as mpol_new() will remove
memoryless nodes for interleave policy, as will bind_zonelist()--called
by mpol_new() for BIND policy.
Proposed fix:
1) modify contextualize_policy logic to:
a) remember whether the incoming node mask is empty.
b) if not, restrict the nodemask to allowed nodes, as is
currently done in-line for mbind(). This guarantees
that the resulting mask includes only nodes with memory.
NOTE: this is a [benign, IMO] change in behavior for
set_mempolicy(). Dis-allowed nodes will be
silently ignored, rather than returning an error.
c) fold this code into mpol_check_policy(), replace 2 calls to
contextualize_policy() to call mpol_check_policy() directly
and remove contextualize_policy().
2) In existing mpol_check_policy() logic, after "contextualization":
a) MPOL_DEFAULT: require that in coming mask "was_empty"
b) MPOL_{BIND|INTERLEAVE}: require that contextualized nodemask
contains at least one node.
c) add a case for MPOL_PREFERRED: if in coming was not empty
and resulting mask IS empty, user specified invalid nodes.
Return EINVAL.
c) remove the now redundant check for memoryless nodes
3) remove the now redundant masking of policy nodes for interleave
policy from mpol_new().
4) Now that mpol_check_policy() contextualizes the nodemask, remove
the in-line nodes_and() from sys_mbind(). I believe that this
restores mbind() to the behavior before the memoryless-nodes
patch series. E.g., we'll no longer treat an invalid nodemask
with MPOL_PREFERRED as local allocation.
[ Patch history:
v1 -> v2:
- Communicate whether or not incoming node mask was empty to
mpol_check_policy() for better error checking.
- As suggested by David Rientjes, remove the now unused
cpuset_nodes_subset_current_mems_allowed() from cpuset.h
v2 -> v3:
- As suggested by Kosaki Motohito, fold the "contextualization"
of policy nodemask into mpol_check_policy(). Looks a little
cleaner. ]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently we possibly lookup the pid in the wrong pid namespace. So
seq_file convert proc_pid_status which ensures the proper pid namespaces is
passed in.
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: another build fix]
[akpm@linux-foundation.org: s390 build fix]
[akpm@linux-foundation.org: fix task_name() output]
[akpm@linux-foundation.org: fix nommu build]
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
Cc: Andrew Morgan <morgan@kernel.org>
Cc: Serge Hallyn <serue@us.ibm.com>
Cc: Cedric Le Goater <clg@fr.ibm.com>
Cc: Pavel Emelyanov <xemul@openvz.org>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Paul Menage <menage@google.com>
Cc: Paul Jackson <pj@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When a cpu is disabled, move_task_off_dead_cpu() is called for tasks that have
been running on that cpu.
Currently, such a task is migrated:
1) to any cpu on the same node as the disabled cpu, which is both online
and among that task's cpus_allowed
2) to any cpu which is both online and among that task's cpus_allowed
It is typical of a multithreaded application running on a large NUMA system to
have its tasks confined to a cpuset so as to cluster them near the memory that
they share. Furthermore, it is typical to explicitly place such a task on a
specific cpu in that cpuset. And in that case the task's cpus_allowed
includes only a single cpu.
This patch would insert a preference to migrate such a task to some cpu within
its cpuset (and set its cpus_allowed to its entire cpuset).
With this patch, migrate the task to:
1) to any cpu on the same node as the disabled cpu, which is both online
and among that task's cpus_allowed
2) to any online cpu within the task's cpuset
3) to any cpu which is both online and among that task's cpus_allowed
In order to do this, move_task_off_dead_cpu() must make a call to
cpuset_cpus_allowed_locked(), a new subset of cpuset_cpus_allowed(), that will
not block. (name change - per Oleg's suggestion)
Calls are made to cpuset_lock() and cpuset_unlock() in migration_call() to set
the cpuset mutex during the whole migrate_live_tasks() and
migrate_dead_tasks() procedure.
[akpm@linux-foundation.org: build fix]
[pj@sgi.com: Fix indentation and spacing]
Signed-off-by: Cliff Wickman <cpw@sgi.com>
Cc: Oleg Nesterov <oleg@tv-sign.ru>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Paul Jackson <pj@sgi.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Remove the filesystem support logic from the cpusets system and makes cpusets
a cgroup subsystem
The "cpuset" filesystem becomes a dummy filesystem; attempts to mount it get
passed through to the cgroup filesystem with the appropriate options to
emulate the old cpuset filesystem behaviour.
Signed-off-by: Paul Menage <menage@google.com>
Cc: Serge E. Hallyn <serue@us.ibm.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Dave Hansen <haveblue@us.ibm.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Paul Jackson <pj@sgi.com>
Cc: Kirill Korotaev <dev@openvz.org>
Cc: Herbert Poetzl <herbert@13thfloor.at>
Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com>
Cc: Cedric Le Goater <clg@fr.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Instead of testing for overlap in the memory nodes of the the nearest
exclusive ancestor of both current and the candidate task, it is better to
simply test for intersection between the task's mems_allowed in their task
descriptors. This does not require taking callback_mutex since it is only
used as a hint in the badness scoring.
Tasks that do not have an intersection in their mems_allowed with the current
task are not explicitly restricted from being OOM killed because it is quite
possible that the candidate task has allocated memory there before and has
since changed its mems_allowed.
Cc: Andrea Arcangeli <andrea@suse.de>
Acked-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
cpusets try to ensure that any node added to a cpuset's mems_allowed is
on-line and contains memory. The assumption was that online nodes contained
memory. Thus, it is possible to add memoryless nodes to a cpuset and then add
tasks to this cpuset. This results in continuous series of oom-kill and
apparent system hang.
Change cpusets to use node_states[N_HIGH_MEMORY] [a.k.a. node_memory_map] in
place of node_online_map when vetting memories. Return error if admin
attempts to write a non-empty mems_allowed node mask containing only
memoryless-nodes.
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@skynet.ie>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Many struct file_operations in the kernel can be "const". Marking them const
moves these to the .rodata section, which avoids false sharing with potential
dirty data. In addition it'll catch accidental writes at compile time to
these shared resources.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/proc/base.c:1869: warning: initialization discards qualifiers from pointer target type
fs/proc/base.c:2150: warning: initialization discards qualifiers from pointer target type
Cc: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Elaborate the API for calling cpuset_zone_allowed(), so that users have to
explicitly choose between the two variants:
cpuset_zone_allowed_hardwall()
cpuset_zone_allowed_softwall()
Until now, whether or not you got the hardwall flavor depended solely on
whether or not you or'd in the __GFP_HARDWALL gfp flag to the gfp_mask
argument.
If you didn't specify __GFP_HARDWALL, you implicitly got the softwall
version.
Unfortunately, this meant that users would end up with the softwall version
without thinking about it. Since only the softwall version might sleep,
this led to bugs with possible sleeping in interrupt context on more than
one occassion.
The hardwall version requires that the current tasks mems_allowed allows
the node of the specified zone (or that you're in interrupt or that
__GFP_THISNODE is set or that you're on a one cpuset system.)
The softwall version, depending on the gfp_mask, might allow a node if it
was allowed in the nearest enclusing cpuset marked mem_exclusive (which
requires taking the cpuset lock 'callback_mutex' to evaluate.)
This patch removes the cpuset_zone_allowed() call, and forces the caller to
explicitly choose between the hardwall and the softwall case.
If the caller wants the gfp_mask to determine this choice, they should (1)
be sure they can sleep or that __GFP_HARDWALL is set, and (2) invoke the
cpuset_zone_allowed_softwall() routine.
This adds another 100 or 200 bytes to the kernel text space, due to the few
lines of nearly duplicate code at the top of both cpuset_zone_allowed_*
routines. It should save a few instructions executed for the calls that
turned into calls of cpuset_zone_allowed_hardwall, thanks to not having to
set (before the call) then check (within the call) the __GFP_HARDWALL flag.
For the most critical call, from get_page_from_freelist(), the same
instructions are executed as before -- the old cpuset_zone_allowed()
routine it used to call is the same code as the
cpuset_zone_allowed_softwall() routine that it calls now.
Not a perfect win, but seems worth it, to reduce this chance of hitting a
sleeping with irq off complaint again.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
- move some file_operations structs into the .rodata section
- move static strings from policy_types[] array into the .rodata section
- fix generic seq_operations usages, so that those structs may be defined
as "const" as well
[akpm@osdl.org: couple of fixes]
Signed-off-by: Helge Deller <deller@gmx.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Optimize the critical zonelist scanning for free pages in the kernel memory
allocator by caching the zones that were found to be full recently, and
skipping them.
Remembers the zones in a zonelist that were short of free memory in the
last second. And it stashes a zone-to-node table in the zonelist struct,
to optimize that conversion (minimize its cache footprint.)
Recent changes:
This differs in a significant way from a similar patch that I
posted a week ago. Now, instead of having a nodemask_t of
recently full nodes, I have a bitmask of recently full zones.
This solves a problem that last weeks patch had, which on
systems with multiple zones per node (such as DMA zone) would
take seeing any of these zones full as meaning that all zones
on that node were full.
Also I changed names - from "zonelist faster" to "zonelist cache",
as that seemed to better convey what we're doing here - caching
some of the key zonelist state (for faster access.)
See below for some performance benchmark results. After all that
discussion with David on why I didn't need them, I went and got
some ;). I wanted to verify that I had not hurt the normal case
of memory allocation noticeably. At least for my one little
microbenchmark, I found (1) the normal case wasn't affected, and
(2) workloads that forced scanning across multiple nodes for
memory improved up to 10% fewer System CPU cycles and lower
elapsed clock time ('sys' and 'real'). Good. See details, below.
I didn't have the logic in get_page_from_freelist() for various
full nodes and zone reclaim failures correct. That should be
fixed up now - notice the new goto labels zonelist_scan,
this_zone_full, and try_next_zone, in get_page_from_freelist().
There are two reasons I persued this alternative, over some earlier
proposals that would have focused on optimizing the fake numa
emulation case by caching the last useful zone:
1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems)
have seen real customer loads where the cost to scan the zonelist
was a problem, due to many nodes being full of memory before
we got to a node we could use. Or at least, I think we have.
This was related to me by another engineer, based on experiences
from some time past. So this is not guaranteed. Most likely, though.
The following approach should help such real numa systems just as
much as it helps fake numa systems, or any combination thereof.
2) The effort to distinguish fake from real numa, using node_distance,
so that we could cache a fake numa node and optimize choosing
it over equivalent distance fake nodes, while continuing to
properly scan all real nodes in distance order, was going to
require a nasty blob of zonelist and node distance munging.
The following approach has no new dependency on node distances or
zone sorting.
See comment in the patch below for a description of what it actually does.
Technical details of note (or controversy):
- See the use of "zlc_active" and "did_zlc_setup" below, to delay
adding any work for this new mechanism until we've looked at the
first zone in zonelist. I figured the odds of the first zone
having the memory we needed were high enough that we should just
look there, first, then get fancy only if we need to keep looking.
- Some odd hackery was needed to add items to struct zonelist, while
not tripping up the custom zonelists built by the mm/mempolicy.c
code for MPOL_BIND. My usual wordy comments below explain this.
Search for "MPOL_BIND".
- Some per-node data in the struct zonelist is now modified frequently,
with no locking. Multiple CPU cores on a node could hit and mangle
this data. The theory is that this is just performance hint data,
and the memory allocator will work just fine despite any such mangling.
The fields at risk are the struct 'zonelist_cache' fields 'fullzones'
(a bitmask) and 'last_full_zap' (unsigned long jiffies). It should
all be self correcting after at most a one second delay.
- This still does a linear scan of the same lengths as before. All
I've optimized is making the scan faster, not algorithmically
shorter. It is now able to scan a compact array of 'unsigned
short' in the case of many full nodes, so one cache line should
cover quite a few nodes, rather than each node hitting another
one or two new and distinct cache lines.
- If both Andi and Nick don't find this too complicated, I will be
(pleasantly) flabbergasted.
- I removed the comment claiming we only use one cachline's worth of
zonelist. We seem, at least in the fake numa case, to have put the
lie to that claim.
- I pay no attention to the various watermarks and such in this performance
hint. A node could be marked full for one watermark, and then skipped
over when searching for a page using a different watermark. I think
that's actually quite ok, as it will tend to slightly increase the
spreading of memory over other nodes, away from a memory stressed node.
===============
Performance - some benchmark results and analysis:
This benchmark runs a memory hog program that uses multiple
threads to touch alot of memory as quickly as it can.
Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of
the total 96 GBytes on the system, and using 1, 19, 37, or 55
threads (on a 56 CPU system.) System, user and real (elapsed)
timings were recorded for each run, shown in units of seconds,
in the table below.
Two kernels were tested - 2.6.18-mm3 and the same kernel with
this zonelist caching patch added. The table also shows the
percentage improvement the zonelist caching sys time is over
(lower than) the stock *-mm kernel.
number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent
GBs N ------------ -------------- ---------------- systime
mem threads sys user real sys user real sys user real better
12 1 153 24 177 151 24 176 -2 0 -1 1%
12 19 99 22 8 99 22 8 0 0 0 0%
12 37 111 25 6 112 25 6 1 0 0 -0%
12 55 115 25 5 110 23 5 -5 -2 0 4%
38 1 502 74 576 497 73 570 -5 -1 -6 0%
38 19 426 78 48 373 76 39 -53 -2 -9 12%
38 37 544 83 36 547 82 36 3 -1 0 -0%
38 55 501 77 23 511 80 24 10 3 1 -1%
64 1 917 125 1042 890 124 1014 -27 -1 -28 2%
64 19 1118 138 119 965 141 103 -153 3 -16 13%
64 37 1202 151 94 1136 150 81 -66 -1 -13 5%
64 55 1118 141 61 1072 140 58 -46 -1 -3 4%
90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4%
90 19 2392 199 192 2116 189 176 -276 -10 -16 11%
90 37 3313 238 175 2972 225 145 -341 -13 -30 10%
90 55 1948 210 104 1843 213 100 -105 3 -4 5%
Notes:
1) This test ran a memory hog program that started a specified number N of
threads, and had each thread allocate and touch 1/N'th of
the total memory to be used in the test run in a single loop,
writing a constant word to memory, one store every 4096 bytes.
Watching this test during some earlier trial runs, I would see
each of these threads sit down on one CPU and stay there, for
the remainder of the pass, a different CPU for each thread.
2) The 'real' column is not comparable to the 'sys' or 'user' columns.
The 'real' column is seconds wall clock time elapsed, from beginning
to end of that test pass. The 'sys' and 'user' columns are total
CPU seconds spent on that test pass. For a 19 thread test run,
for example, the sum of 'sys' and 'user' could be up to 19 times the
number of 'real' elapsed wall clock seconds.
3) Tests were run on a fresh, single-user boot, to minimize the amount
of memory already in use at the start of the test, and to minimize
the amount of background activity that might interfere.
4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM.
5) Notice that the 'real' time gets large for the single thread runs, even
though the measured 'sys' and 'user' times are modest. I'm not sure what
that means - probably something to do with it being slow for one thread to
be accessing memory along ways away. Perhaps the fake numa system, running
ostensibly the same workload, would not show this substantial degradation
of 'real' time for one thread on many nodes -- lets hope not.
6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs)
ran quite efficiently, as one might expect. Each pair of threads needed
to allocate and touch the memory on the node the two threads shared, a
pleasantly parallizable workload.
7) The intermediate thread count passes, when asking for alot of memory forcing
them to go to a few neighboring nodes, improved the most with this zonelist
caching patch.
Conclusions:
* This zonelist cache patch probably makes little difference one way or the
other for most workloads on real numa hardware, if those workloads avoid
heavy off node allocations.
* For memory intensive workloads requiring substantial off-node allocations
on real numa hardware, this patch improves both kernel and elapsed timings
up to ten per-cent.
* For fake numa systems, I'm optimistic, but will have to leave that up to
Rohit Seth to actually test (once I get him a 2.6.18 backport.)
Signed-off-by: Paul Jackson <pj@sgi.com>
Cc: Rohit Seth <rohitseth@google.com>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Cc: David Rientjes <rientjes@cs.washington.edu>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Change the list of memory nodes allowed to tasks in the top (root) nodeset
to dynamically track what cpus are online, using a call to a cpuset hook
from the memory hotplug code. Make this top cpus file read-only.
On systems that have cpusets configured in their kernel, but that aren't
actively using cpusets (for some distros, this covers the majority of
systems) all tasks end up in the top cpuset.
If that system does support memory hotplug, then these tasks cannot make
use of memory nodes that are added after system boot, because the memory
nodes are not allowed in the top cpuset. This is a surprising regression
over earlier kernels that didn't have cpusets enabled.
One key motivation for this change is to remain consistent with the
behaviour for the top_cpuset's 'cpus', which is also read-only, and which
automatically tracks the cpu_online_map.
This change also has the minor benefit that it fixes a long standing,
little noticed, minor bug in cpusets. The cpuset performance tweak to
short circuit the cpuset_zone_allowed() check on systems with just a single
cpuset (see 'number_of_cpusets', in linux/cpuset.h) meant that simply
changing the 'mems' of the top_cpuset had no affect, even though the change
(the write system call) appeared to succeed. With the following change,
that write to the 'mems' file fails -EACCES, and the 'mems' file stubbornly
refuses to be changed via user space writes. Thus no one should be mislead
into thinking they've changed the top_cpusets's 'mems' when in affect they
haven't.
In order to keep the behaviour of cpusets consistent between systems
actively making use of them and systems not using them, this patch changes
the behaviour of the 'mems' file in the top (root) cpuset, making it read
only, and making it automatically track the value of node_online_map. Thus
tasks in the top cpuset will have automatic use of hot plugged memory nodes
allowed by their cpuset.
[akpm@osdl.org: build fix]
[bunk@stusta.de: build fix]
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch provides the implementation and cpuset interface for an alternative
memory allocation policy that can be applied to certain kinds of memory
allocations, such as the page cache (file system buffers) and some slab caches
(such as inode caches).
The policy is called "memory spreading." If enabled, it spreads out these
kinds of memory allocations over all the nodes allowed to a task, instead of
preferring to place them on the node where the task is executing.
All other kinds of allocations, including anonymous pages for a tasks stack
and data regions, are not affected by this policy choice, and continue to be
allocated preferring the node local to execution, as modified by the NUMA
mempolicy.
There are two boolean flag files per cpuset that control where the kernel
allocates pages for the file system buffers and related in kernel data
structures. They are called 'memory_spread_page' and 'memory_spread_slab'.
If the per-cpuset boolean flag file 'memory_spread_page' is set, then the
kernel will spread the file system buffers (page cache) evenly over all the
nodes that the faulting task is allowed to use, instead of preferring to put
those pages on the node where the task is running.
If the per-cpuset boolean flag file 'memory_spread_slab' is set, then the
kernel will spread some file system related slab caches, such as for inodes
and dentries evenly over all the nodes that the faulting task is allowed to
use, instead of preferring to put those pages on the node where the task is
running.
The implementation is simple. Setting the cpuset flags 'memory_spread_page'
or 'memory_spread_cache' turns on the per-process flags PF_SPREAD_PAGE or
PF_SPREAD_SLAB, respectively, for each task that is in the cpuset or
subsequently joins that cpuset. In subsequent patches, the page allocation
calls for the affected page cache and slab caches are modified to perform an
inline check for these flags, and if set, a call to a new routine
cpuset_mem_spread_node() returns the node to prefer for the allocation.
The cpuset_mem_spread_node() routine is also simple. It uses the value of a
per-task rotor cpuset_mem_spread_rotor to select the next node in the current
tasks mems_allowed to prefer for the allocation.
This policy can provide substantial improvements for jobs that need to place
thread local data on the corresponding node, but that need to access large
file system data sets that need to be spread across the several nodes in the
jobs cpuset in order to fit. Without this patch, especially for jobs that
might have one thread reading in the data set, the memory allocation across
the nodes in the jobs cpuset can become very uneven.
A couple of Copyright year ranges are updated as well. And a couple of email
addresses that can be found in the MAINTAINERS file are removed.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus 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: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Remove a couple of more lines of code from the cpuset hooks in the page
allocation code path.
There was a check for a NULL cpuset pointer in the routine
cpuset_update_task_memory_state() that was only needed during system boot,
after the memory subsystem was initialized, before the cpuset subsystem was
initialized, to catch a NULL task->cpuset pointer.
Add a cpuset_init_early() routine, just before the mem_init() call in
init/main.c, that sets up just enough of the init tasks cpuset structure to
render cpuset_update_task_memory_state() calls harmless.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Easy little optimization hack to avoid actually having to call
cpuset_zone_allowed() and check mems_allowed, in the main page allocation
routine, __alloc_pages(). This saves several CPU cycles per page allocation
on systems not using cpusets.
A counter is updated each time a cpuset is created or removed, and whenever
there is only one cpuset in the system, it must be the root cpuset, which
contains all CPUs and all Memory Nodes. In that case, when the counter is
one, all allocations are allowed.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Provide a cpuset_mems_allowed() method, which the sys_migrate_pages() code
needed, to obtain the mems_allowed vector of a cpuset, and replaced the
workaround in sys_migrate_pages() to call this new method.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The important code paths through alloc_pages_current() and alloc_page_vma(),
by which most kernel page allocations go, both called
cpuset_update_current_mems_allowed(), which in turn called refresh_mems().
-Both- of these latter two routines did a tasklock, got the tasks cpuset
pointer, and checked for out of date cpuset->mems_generation.
That was a silly duplication of code and waste of CPU cycles on an important
code path.
Consolidated those two routines into a single routine, called
cpuset_update_task_memory_state(), since it updates more than just
mems_allowed.
Changed all callers of either routine to call the new consolidated routine.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Provide a simple per-cpuset metric of memory pressure, tracking the -rate-
that the tasks in a cpuset call try_to_free_pages(), the synchronous
(direct) memory reclaim code.
This enables batch managers monitoring jobs running in dedicated cpusets to
efficiently detect what level of memory pressure that job is causing.
This is useful both on tightly managed systems running a wide mix of
submitted jobs, which may choose to terminate or reprioritize jobs that are
trying to use more memory than allowed on the nodes assigned them, and with
tightly coupled, long running, massively parallel scientific computing jobs
that will dramatically fail to meet required performance goals if they
start to use more memory than allowed to them.
This patch just provides a very economical way for the batch manager to
monitor a cpuset for signs of memory pressure. It's up to the batch
manager or other user code to decide what to do about it and take action.
==> Unless this feature is enabled by writing "1" to the special file
/dev/cpuset/memory_pressure_enabled, the hook in the rebalance
code of __alloc_pages() for this metric reduces to simply noticing
that the cpuset_memory_pressure_enabled flag is zero. So only
systems that enable this feature will compute the metric.
Why a per-cpuset, running average:
Because this meter is per-cpuset, rather than per-task or mm, the
system load imposed by a batch scheduler monitoring this metric is
sharply reduced on large systems, because a scan of the tasklist can be
avoided on each set of queries.
Because this meter is a running average, instead of an accumulating
counter, a batch scheduler can detect memory pressure with a single
read, instead of having to read and accumulate results for a period of
time.
Because this meter is per-cpuset rather than per-task or mm, the
batch scheduler can obtain the key information, memory pressure in a
cpuset, with a single read, rather than having to query and accumulate
results over all the (dynamically changing) set of tasks in the cpuset.
A per-cpuset simple digital filter (requires a spinlock and 3 words of data
per-cpuset) is kept, and updated by any task attached to that cpuset, if it
enters the synchronous (direct) page reclaim code.
A per-cpuset file provides an integer number representing the recent
(half-life of 10 seconds) rate of direct page reclaims caused by the tasks
in the cpuset, in units of reclaims attempted per second, times 1000.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Finish converting mm/mempolicy.c from bitmaps to nodemasks. The previous
conversion had left one routine using bitmaps, since it involved a
corresponding change to kernel/cpuset.c
Fix that interface by replacing with a simple macro that calls nodes_subset(),
or if !CONFIG_CPUSET, returns (1).
Signed-off-by: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <christoph@lameter.com>
Cc: Andi Kleen <ak@muc.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus 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: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus 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: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch makes use of the previously underutilized cpuset flag
'mem_exclusive' to provide what amounts to another layer of memory placement
resolution. With this patch, there are now the following four layers of
memory placement available:
1) The whole system (interrupt and GFP_ATOMIC allocations can use this),
2) The nearest enclosing mem_exclusive cpuset (GFP_KERNEL allocations can use),
3) The current tasks cpuset (GFP_USER allocations constrained to here), and
4) Specific node placement, using mbind and set_mempolicy.
These nest - each layer is a subset (same or within) of the previous.
Layer (2) above is new, with this patch. The call used to check whether a
zone (its node, actually) is in a cpuset (in its mems_allowed, actually) is
extended to take a gfp_mask argument, and its logic is extended, in the case
that __GFP_HARDWALL is not set in the flag bits, to look up the cpuset
hierarchy for the nearest enclosing mem_exclusive cpuset, to determine if
placement is allowed. The definition of GFP_USER, which used to be identical
to GFP_KERNEL, is changed to also set the __GFP_HARDWALL bit, in the previous
cpuset_gfp_hardwall_flag patch.
GFP_ATOMIC and GFP_KERNEL allocations will stay within the current tasks
cpuset, so long as any node therein is not too tight on memory, but will
escape to the larger layer, if need be.
The intended use is to allow something like a batch manager to handle several
jobs, each job in its own cpuset, but using common kernel memory for caches
and such. Swapper and oom_kill activity is also constrained to Layer (2). A
task in or below one mem_exclusive cpuset should not cause swapping on nodes
in another non-overlapping mem_exclusive cpuset, nor provoke oom_killing of a
task in another such cpuset. Heavy use of kernel memory for i/o caching and
such by one job should not impact the memory available to jobs in other
non-overlapping mem_exclusive cpusets.
This patch enables providing hardwall, inescapable cpusets for memory
allocations of each job, while sharing kernel memory allocations between
several jobs, in an enclosing mem_exclusive cpuset.
Like Dinakar's patch earlier to enable administering sched domains using the
cpu_exclusive flag, this patch also provides a useful meaning to a cpuset flag
that had previously done nothing much useful other than restrict what cpuset
configurations were allowed.
Signed-off-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
gcc-4 warns with
include/linux/cpuset.h:21: warning: type qualifiers ignored on function
return type
cpuset_cpus_allowed is declared with const
extern const cpumask_t cpuset_cpus_allowed(const struct task_struct *p);
First const should be __attribute__((const)), but the gcc manual
explains that:
"Note that a function that has pointer arguments and examines the data
pointed to must not be declared const. Likewise, a function that calls a
non-const function usually must not be const. It does not make sense for
a const function to return void."
The following patch remove const from the function declaration.
Signed-off-by: Benoit Boissinot <benoit.boissinot@ens-lyon.org>
Acked-by: Paul Jackson <pj@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus 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!
Max Krasnyansky