Commit da4d897e ("core: add cgroup memory controller support on the unified
hierarchy (#3315)") changed the code in src/core/cgroup.c to always write
the real numeric value from the cgroup parameters to the
"memory.limit_in_bytes" attribute file.
For parameters set to CGROUP_LIMIT_MAX, this results in the string
"18446744073709551615" being written into that file, which is UINT64_MAX.
Before that commit, CGROUP_LIMIT_MAX was special-cased to the string "-1".
This causes a regression on CentOS 7, which is based on kernel 3.10, as the
value is interpreted as *signed* 64 bit, and clamped to 0:
[root@n54 ~]# echo 18446744073709551615 >/sys/fs/cgroup/memory/user.slice/memory.limit_in_bytes
[root@n54 ~]# cat /sys/fs/cgroup/memory/user.slice/memory.limit_in_bytes
0
[root@n54 ~]# echo -1 >/sys/fs/cgroup/memory/user.slice/memory.limit_in_bytes
[root@n54 ~]# cat /sys/fs/cgroup/memory/user.slice/memory.limit_in_bytes
9223372036854775807
Hence, all units that are subject to the limits enforced by the memory
controller will crash immediately, even though they have no actual limit
set. This happens to for the user.slice, for instance:
[ 453.577153] Hardware name: SeaMicro SM15000-64-CC-AA-1Ox1/AMD Server CRB, BIOS Estoc.3.72.19.0018 08/19/2014
[ 453.587024] ffff880810c56780 00000000aae9501f ffff880813d7fcd0 ffffffff816360fc
[ 453.594544] ffff880813d7fd60 ffffffff8163109c ffff88080ffc5000 ffff880813d7fd28
[ 453.602120] ffffffff00000202 fffeefff00000000 0000000000000001 ffff880810c56c03
[ 453.609680] Call Trace:
[ 453.612156] [<ffffffff816360fc>] dump_stack+0x19/0x1b
[ 453.617324] [<ffffffff8163109c>] dump_header+0x8e/0x214
[ 453.622671] [<ffffffff8116d20e>] oom_kill_process+0x24e/0x3b0
[ 453.628559] [<ffffffff81088dae>] ? has_capability_noaudit+0x1e/0x30
[ 453.634969] [<ffffffff811d4155>] mem_cgroup_oom_synchronize+0x575/0x5a0
[ 453.641721] [<ffffffff811d3520>] ? mem_cgroup_charge_common+0xc0/0xc0
[ 453.648299] [<ffffffff8116da84>] pagefault_out_of_memory+0x14/0x90
[ 453.654621] [<ffffffff8162f4cc>] mm_fault_error+0x68/0x12b
[ 453.660233] [<ffffffff81642012>] __do_page_fault+0x3e2/0x450
[ 453.666017] [<ffffffff816420a3>] do_page_fault+0x23/0x80
[ 453.671467] [<ffffffff8163e308>] page_fault+0x28/0x30
[ 453.676656] Task in /user.slice/user-0.slice/user@0.service killed as a result of limit of /user.slice/user-0.slice/user@0.service
[ 453.688477] memory: usage 0kB, limit 0kB, failcnt 7
[ 453.693391] memory+swap: usage 0kB, limit 9007199254740991kB, failcnt 0
[ 453.700039] kmem: usage 0kB, limit 9007199254740991kB, failcnt 0
[ 453.706076] Memory cgroup stats for /user.slice/user-0.slice/user@0.service: cache:0KB rss:0KB rss_huge:0KB mapped_file:0KB swap:0KB inactive_anon:0KB active_anon:0KB inactive_file:0KB active_file:0KB unevictable:0KB
[ 453.725702] [ pid ] uid tgid total_vm rss nr_ptes swapents oom_score_adj name
[ 453.733614] [ 2837] 0 2837 11950 899 23 0 0 (systemd)
[ 453.741919] Memory cgroup out of memory: Kill process 2837 ((systemd)) score 1 or sacrifice child
[ 453.750831] Killed process 2837 ((systemd)) total-vm:47800kB, anon-rss:3188kB, file-rss:408kB
Fix this issue by special-casing the UINT64_MAX case again.
To accommodate changes in kernel interface, cgroup unified hierarchy support
added several configuration items which overlap with the existing resource
control settings and there is simple config translation between the overlapping
settings to ease the transition. As why certain cgroup knobs are being
configured can become confusing, this patch adds a master warning message which
is printed once when such translation is first used and logs each translation
with a debug message.
v2:
- Switched to log_unit*().
cgroup_context_apply() and friends take CGroupContext and cgroup path as input
and has no way of getting back to the associated Unit and thus uses raw cgroup
path for logging. This makes the log messages difficult to track down.
There's no reason to avoid passing in Unit into these functions. Pass in Unit
and use log_unit*() instead.
While at it, make cgroup_context_apply(), which has no outside users, static.
Also, drop cgroup path from log messages where the path itself isn't too
interesting and can be easily obtained from the unit.
On the unified hierarchy, memory controller implements three control knobs -
low, high and max which enables more useable and versatile control over memory
usage. This patch implements support for the three control knobs.
* MemoryLow, MemoryHigh and MemoryMax are added for memory.low, memory.high and
memory.max, respectively.
* As all absolute limits on the unified hierarchy use "max" for no limit, make
memory limit parse functions accept "max" in addition to "infinity" and
document "max" for the new knobs.
* Implement compatibility translation between MemoryMax and MemoryLimit.
v2:
- Fixed missing else's in config_parse_memory_limit().
- Fixed missing newline when writing out drop-ins.
- Coding style updates to use "val > 0" instead of "val".
- Minor updates to documentation.
Due to the substantial interface changes in cgroup unified hierarchy, new IO
settings are introduced. Currently, IO settings apply only to unified
hierarchy and BlockIO to legacy. While the transition is necessary, it's
painful for users to have to provide configs for both. This patch implements
translation from one config set to another for configs which make sense.
* The translation takes place during application of the configs. Users won't
see IO or BlockIO settings appearing without being explicitly created.
* The translation takes place only if there is no config for the matching
cgroup hierarchy type at all.
While this doesn't provide comprehensive compatibility, it should considerably
ease transition to the new IO settings which are a superset of BlockIO
settings.
v2:
- Update test-cgroup-mask.c so that it accounts for the fact that
CGROUP_MASK_IO and CGROUP_MASK_BLKIO move together. Also, test/parent.slice
now sets IOWeight instead of BlockIOWeight.
Factor out the following functions out of cgroup_context_apply()
* cgroup_context_[blk]io_weight()
* cgroup_apply_[blk]io_device_weight()
* cgroup_apply_[blk]io_device_limit()
This is pure refactoring and shouldn't cause any functional differences.
CGroupBlockIODeviceBandwith is used to keep track of IO bandwidth limits for
legacy cgroup hierarchies. Unlike the unified hierarchy counterpart
CGroupIODeviceLimit, a CGroupBlockIODeviceBandwiddth records either a read or
write limit and has a couple issues.
* There's no way to clear specific config entry.
* When configs are cleared for an IO direction of a unit, the kernel settings
aren't cleared accordingly creating discrepancies.
This patch updates CGroupBlockIODeviceBandwidth so that it behaves similarly to
CGroupIODeviceLimit - each entry records both rbps and wbps limits and is
cleared if both are at default values after kernel settings are updated.
cgroup IO controller supports maximum limits for both bandwidth and IOPS but
systemd resource control currently only supports bandwidth limits. This patch
adds support for IOReadIOPSMax and IOWriteIOPSMax when unified cgroup hierarchy
is in use.
It isn't difficult to also add BlockIOReadIOPS and BlockIOWriteIOPS for legacy
hierarchies but IO control on legacy hierarchies is half-broken anyway, so
let's leave it alone for now.
Currently, there are two cgroup IO limits, bandwidth max for read and write,
and they are hard-coded in various places. This is fine for two limits but IO
is expected to grow more limits - low, high and max limits for bandwidth and
IOPS - and hard-coding each limit won't make sense.
This patch replaces hard-coded limits with an array indexed by
CGroupIOLimitType and accompanying string and default value tables so that new
limits can be added trivially.
On the unified hierarchy, blkio controller is renamed to io and the interface
is changed significantly.
* blkio.weight and blkio.weight_device are consolidated into io.weight which
uses the standardized weight range [1, 10000] with 100 as the default value.
* blkio.throttle.{read|write}_{bps|iops}_device are consolidated into io.max.
Expansion of throttling features is being worked on to support
work-conserving absolute limits (io.low and io.high).
* All stats are consolidated into io.stats.
This patchset adds support for the new interface. As the interface has been
revamped and new features are expected to be added, it seems best to treat it
as a separate controller rather than trying to expand the blkio settings
although we might add automatic translation if only blkio settings are
specified.
* io.weight handling is mostly identical to blkio.weight[_device] handling
except that the weight range is different.
* Both read and write bandwidth settings are consolidated into
CGroupIODeviceLimit which describes all limits applicable to the device.
This makes it less painful to add new limits.
* "max" can be used to specify the maximum limit which is equivalent to no
config for max limits and treated as such. If a given CGroupIODeviceLimit
doesn't contain any non-default configs, the config struct is discarded once
the no limit config is applied to cgroup.
* lookup_blkio_device() is renamed to lookup_block_device().
Signed-off-by: Tejun Heo <htejun@fb.com>
dbus-daemon currently uses a backlog of 30 on its D-bus system bus socket. On
overloaded systems this means that only 30 connections may be queued without
dbus-daemon processing them before further connection attempts fail. Our
cgroups-agent binary so far used D-Bus for its messaging, and hitting this
limit hence may result in us losing cgroup empty messages.
This patch adds a seperate cgroup agent socket of type AF_UNIX/SOCK_DGRAM.
Since sockets of these types need no connection set up, no listen() backlog
applies. Our cgroup-agent binary will hence simply block as long as it can't
enqueue its datagram message, so that we won't lose cgroup empty messages as
likely anymore.
This also rearranges the ordering of the processing of SIGCHLD signals, service
notification messages (sd_notify()...) and the two types of cgroup
notifications (inotify for the unified hierarchy support, and agent for the
classic hierarchy support). We now always process events for these in the
following order:
1. service notification messages (SD_EVENT_PRIORITY_NORMAL-7)
2. SIGCHLD signals (SD_EVENT_PRIORITY_NORMAL-6)
3. cgroup inotify and cgroup agent (SD_EVENT_PRIORITY_NORMAL-5)
This is because when receiving SIGCHLD we invalidate PID information, which we
need to process the service notification messages which are bound to PIDs.
Hence the order between the first two items. And we want to process SIGCHLD
metadata to detect whether a service is gone, before using cgroup
notifications, to decide when a service is gone, since the former carries more
useful metadata.
Related to this:
https://bugs.freedesktop.org/show_bug.cgi?id=95264https://github.com/systemd/systemd/issues/1961
unit_has_mask_realized() determines whether the specified unit has its cgroups
set up properly given the desired target_mask; however, on the unified
hierarchy, controllers need to be enabled explicitly for children and the mask
of enabled controllers can deviate from target_mask. Only considering
target_mask in unit_has_mask_realized() can lead to false positives and
skipping enabling the requested controllers.
This patch adds unit->cgroup_enabled_mask to track which controllers are
enabled and updates unit_has_mask_realized() to also consider enable_mask.
Signed-off-by: Tejun Heo <htejun@fb.com>
Previously, we had two enums ManagerRunningAs and UnitFileScope, that were
mostly identical and converted from one to the other all the time. The latter
had one more value UNIT_FILE_GLOBAL however.
Let's simplify things, and remove ManagerRunningAs and replace it by
UnitFileScope everywhere, thus making the translation unnecessary. Introduce
two new macros MANAGER_IS_SYSTEM() and MANAGER_IS_USER() to simplify checking
if we are running in one or the user context.
Earlier during the development of unified hierarchy, the populated event was
reported through by the dedicated "cgroup.populated" file; however, the
interface was updated so that it's reported through the "populated" field of
"cgroup.events" file. Update populated event handling logic accordingly.
Support for net_cls.class_id through the NetClass= configuration directive
has been added in v227 in preparation for a per-unit packet filter mechanism.
However, it turns out the kernel people have decided to deprecate the net_cls
and net_prio controllers in v2. Tejun provides a comprehensive justification
for this in his commit, which has landed during the merge window for kernel
v4.5:
https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=bd1060a1d671
As we're aiming for full support for the v2 cgroup hierarchy, we can no
longer support this feature. Userspace tool such as nftables are moving over
to setting rules that are specific to the full cgroup path of a task, which
obsoletes these controllers anyway.
This commit removes support for tweaking details in the net_cls controller,
but keeps the NetClass= directive around for legacy compatibility reasons.
There are more than enough calls doing string manipulations to deserve
its own files, hence do something about it.
This patch also sorts the #include blocks of all files that needed to be
updated, according to the sorting suggestions from CODING_STYLE. Since
pretty much every file needs our string manipulation functions this
effectively means that most files have sorted #include blocks now.
Also touches a few unrelated include files.
Add a new config directive called NetClass= to CGroup enabled units.
Allowed values are positive numbers for fix assignments and "auto" for
picking a free value automatically, for which we need to keep track of
dynamically assigned net class IDs of units. Introduce a hash table for
this, and also record the last ID that was given out, so the allocator
can start its search for the next 'hole' from there. This could
eventually be optimized with something like an irb.
The class IDs up to 65536 are considered reserved and won't be
assigned automatically by systemd. This barrier can be made a config
directive in the future.
Values set in unit files are stored in the CGroupContext of the
unit and considered read-only. The actually assigned number (which
may have been chosen dynamically) is stored in the unit itself and
is guaranteed to remain stable as long as the unit is active.
In the CGroup controller, set the configured CGroup net class to
net_cls.classid. Multiple unit may share the same net class ID,
and those which do are linked together.
Let's make sure that we follow the same codepaths when adjusting a
cgroup property via the dbus SetProperty() call, and when we execute the
StartupCPUShares= effect.
Let's stop using the "unsigned long" type for weights/shares, and let's
just use uint64_t for this, as that's what we expose on the bus.
Unify parsers, and always validate the range for these fields.
Correct the default blockio weight to 500, since that's what the kernel
actually uses.
When parsing the weight/shares settings from unit files accept the empty
string as a way to reset the weight/shares value. When getting it via
the bus, uniformly map (uint64_t) -1 to unset.
Open up StartupCPUShares= and StartupBlockIOWeight= to transient units.
This adds support for the new "pids" cgroup controller of 4.3 kernels.
It allows accounting the number of tasks in a cgroup and enforcing
limits on it.
This adds two new setting TasksAccounting= and TasksMax= to each unit,
as well as a gloabl option DefaultTasksAccounting=.
This also updated "cgtop" to optionally make use of the new
kernel-provided accounting.
systemctl has been updated to show the number of tasks for each service
if it is available.
This patch also adds correct support for undoing memory limits for units
using a MemoryLimit=infinity syntax. We do the same for TasksMax= now
and hence keep things in sync here.
Delegation to unpriviliged processes is safe in the unified hierarchy,
hence allow it. This has the benefit of permitting "systemd --user"
instances to further partition their resources between user services.
Let's move the actual cgroup part of it into a new separate function
manager_get_unit_by_pid_cgroup(), and then make
manager_get_unit_by_pid() just a wrapper that also checks the two pid
hashmaps.
Then, let's make sure the various calls that want to deliver events to
the owners of a PID check both hashmaps and the cgroup and deliver the
event to *each* of them. OTOH make sure bus calls like GetUnitByPID()
continue to check the PID hashmaps first and the cgroup only as
fallback.
This adds a new PID_TO_PTR() macro, plus PTR_TO_PID() and makes use of
it wherever we maintain processes in a hash table. Previously we
sometimes used LONG_TO_PTR() and other times ULONG_TO_PTR() for that,
hence let's make this more explicit and clean up things.
This patch set adds full support the new unified cgroup hierarchy logic
of modern kernels.
A new kernel command line option "systemd.unified_cgroup_hierarchy=1" is
added. If specified the unified hierarchy is mounted to /sys/fs/cgroup
instead of a tmpfs. No further hierarchies are mounted. The kernel
command line option defaults to off. We can turn it on by default as
soon as the kernel's APIs regarding this are stabilized (but even then
downstream distros might want to turn this off, as this will break any
tools that access cgroupfs directly).
It is possibly to choose for each boot individually whether the unified
or the legacy hierarchy is used. nspawn will by default provide the
legacy hierarchy to containers if the host is using it, and the unified
otherwise. However it is possible to run containers with the unified
hierarchy on a legacy host and vice versa, by setting the
$UNIFIED_CGROUP_HIERARCHY environment variable for nspawn to 1 or 0,
respectively.
The unified hierarchy provides reliable cgroup empty notifications for
the first time, via inotify. To make use of this we maintain one
manager-wide inotify fd, and each cgroup to it.
This patch also removes cg_delete() which is unused now.
On kernel 4.2 only the "memory" controller is compatible with the
unified hierarchy, hence that's the only controller systemd exposes when
booted in unified heirarchy mode.
This introduces a new enum for enumerating supported controllers, plus a
related enum for the mask bits mapping to it. The core is changed to
make use of this everywhere.
This moves PID 1 into a new "init.scope" implicit scope unit in the root
slice. This is necessary since on the unified hierarchy cgroups may
either contain subgroups or processes but not both. PID 1 hence has to
move out of the root cgroup (strictly speaking the root cgroup is the
only one where processes and subgroups are still allowed, but in order
to support containers nicey, we move PID 1 into the new scope in all
cases.) This new unit is also used on legacy hierarchy setups. It's
actually pretty useful on all systems, as it can then be used to filter
journal messages coming from PID 1, and so on.
The root slice ("-.slice") is now implicitly created and started (and
does not require a unit file on disk anymore), since
that's where "init.scope" is located and the slice needs to be started
before the scope can.
To check whether we are in unified or legacy hierarchy mode we use
statfs() on /sys/fs/cgroup. If the .f_type field reports tmpfs we are in
legacy mode, if it reports cgroupfs we are in unified mode.
This patch set carefuly makes sure that cgls and cgtop continue to work
as desired.
When invoking nspawn as a service it will implicitly create two
subcgroups in the cgroup it is using, one to move the nspawn process
into, the other to move the actual container processes into. This is
done because of the requirement that cgroups may either contain
processes or other subgroups.
It's cheaper that going to cgroupfs, and also usually the better choice
since it's not racy and can map PIDs even if they were moved to a
different unit.
In all cases where the function (or cg_is_empty_recursive()) ignoring
the calling process is actually wrong, as a process keeps a cgroup busy
regardless if its the current one or another. Hence, let's simplify
things and drop the "ignore_self" parameter.
The legacy cgroup hierarchy does not support reliable empty
notifications in containers and if there are left-over subgroups in a
cgroup. This makes it hard to correctly wait for them running empty, and
thus we previously disabled this logic entirely.
With this change we explicitly check for the container case, and whether
the unit is a "delegation" unit (i.e. one where programs may create
their own subgroups). If we are neither in a container, nor operating on
a delegation unit cgroup empty notifications become reliable and thus we
start waiting for the empty notifications again.
This doesn't really fix the general problem around cgroup notifications
but reduces the effect around it.
(This also reorders #include lines by their focus, as suggsted in
CODING_STYLE. We have to add "virt.h", so let's do that at the right
place.)
Also see #317.
It's primarily just a property of the Manager object after all, and we
try to refer to PID 1 as "manager" instead of "systemd", hence let's to
stick to this here too.
Torstein Husebø