···
1
-
commit 827b86ad1dd21feed4c0b99faf6059f245f7dadb
2
-
Author: Tejun Heo <tj@kernel.org>
3
-
Date: Fri Mar 11 07:31:23 2016 -0500
5
-
sched: Misc preps for cgroup unified hierarchy interface
7
-
Make the following changes in preparation for the cpu controller
8
-
interface implementation for the unified hierarchy. This patch
9
-
doesn't cause any functional differences.
11
-
* s/cpu_stats_show()/cpu_cfs_stats_show()/
13
-
* s/cpu_files/cpu_legacy_files/
15
-
* Separate out cpuacct_stats_read() from cpuacct_stats_show(). While
16
-
at it, make the @val array u64 for consistency.
18
-
Signed-off-by: Tejun Heo <tj@kernel.org>
19
-
Cc: Ingo Molnar <mingo@redhat.com>
20
-
Cc: Peter Zijlstra <peterz@infradead.org>
21
-
Cc: Li Zefan <lizefan@huawei.com>
22
-
Cc: Johannes Weiner <hannes@cmpxchg.org>
24
-
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
25
-
index 3b31fc05a0f1..a1b95e83fa87 100644
26
-
--- a/kernel/sched/core.c
27
-
+++ b/kernel/sched/core.c
28
-
@@ -7174,7 +7174,7 @@ static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
32
-
-static int cpu_stats_show(struct seq_file *sf, void *v)
33
-
+static int cpu_cfs_stats_show(struct seq_file *sf, void *v)
35
-
struct task_group *tg = css_tg(seq_css(sf));
36
-
struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
37
-
@@ -7214,7 +7214,7 @@ static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
39
-
#endif /* CONFIG_RT_GROUP_SCHED */
41
-
-static struct cftype cpu_files[] = {
42
-
+static struct cftype cpu_legacy_files[] = {
43
-
#ifdef CONFIG_FAIR_GROUP_SCHED
46
-
@@ -7235,7 +7235,7 @@ static struct cftype cpu_files[] = {
50
-
- .seq_show = cpu_stats_show,
51
-
+ .seq_show = cpu_cfs_stats_show,
54
-
#ifdef CONFIG_RT_GROUP_SCHED
55
-
@@ -7261,7 +7261,7 @@ struct cgroup_subsys cpu_cgrp_subsys = {
56
-
.fork = cpu_cgroup_fork,
57
-
.can_attach = cpu_cgroup_can_attach,
58
-
.attach = cpu_cgroup_attach,
59
-
- .legacy_cftypes = cpu_files,
60
-
+ .legacy_cftypes = cpu_legacy_files,
64
-
diff --git a/kernel/sched/cpuacct.c b/kernel/sched/cpuacct.c
65
-
index f95ab29a45d0..6151c23f722f 100644
66
-
--- a/kernel/sched/cpuacct.c
67
-
+++ b/kernel/sched/cpuacct.c
68
-
@@ -276,26 +276,33 @@ static int cpuacct_all_seq_show(struct seq_file *m, void *V)
72
-
-static int cpuacct_stats_show(struct seq_file *sf, void *v)
73
-
+static void cpuacct_stats_read(struct cpuacct *ca,
74
-
+ u64 (*val)[CPUACCT_STAT_NSTATS])
76
-
- struct cpuacct *ca = css_ca(seq_css(sf));
77
-
- s64 val[CPUACCT_STAT_NSTATS];
81
-
- memset(val, 0, sizeof(val));
82
-
+ memset(val, 0, sizeof(*val));
84
-
for_each_possible_cpu(cpu) {
85
-
u64 *cpustat = per_cpu_ptr(ca->cpustat, cpu)->cpustat;
87
-
- val[CPUACCT_STAT_USER] += cpustat[CPUTIME_USER];
88
-
- val[CPUACCT_STAT_USER] += cpustat[CPUTIME_NICE];
89
-
- val[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_SYSTEM];
90
-
- val[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_IRQ];
91
-
- val[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_SOFTIRQ];
92
-
+ (*val)[CPUACCT_STAT_USER] += cpustat[CPUTIME_USER];
93
-
+ (*val)[CPUACCT_STAT_USER] += cpustat[CPUTIME_NICE];
94
-
+ (*val)[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_SYSTEM];
95
-
+ (*val)[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_IRQ];
96
-
+ (*val)[CPUACCT_STAT_SYSTEM] += cpustat[CPUTIME_SOFTIRQ];
100
-
+static int cpuacct_stats_show(struct seq_file *sf, void *v)
102
-
+ u64 val[CPUACCT_STAT_NSTATS];
105
-
+ cpuacct_stats_read(css_ca(seq_css(sf)), &val);
107
-
for (stat = 0; stat < CPUACCT_STAT_NSTATS; stat++) {
108
-
- seq_printf(sf, "%s %lld\n",
109
-
+ seq_printf(sf, "%s %llu\n",
110
-
cpuacct_stat_desc[stat],
111
-
(long long)nsec_to_clock_t(val[stat]));
114
-
commit fdb64d002b3a223ce4bb11aa4448a42050470052
115
-
Author: Tejun Heo <tj@kernel.org>
116
-
Date: Fri Mar 11 07:31:23 2016 -0500
118
-
sched: Implement interface for cgroup unified hierarchy
120
-
While the cpu controller doesn't have any functional problems, there
121
-
are a couple interface issues which can be addressed in the v2
124
-
* cpuacct being a separate controller. This separation is artificial
125
-
and rather pointless as demonstrated by most use cases co-mounting
126
-
the two controllers. It also forces certain information to be
129
-
* Use of different time units. Writable control knobs use
130
-
microseconds, some stat fields use nanoseconds while other cpuacct
131
-
stat fields use centiseconds.
133
-
* Control knobs which can't be used in the root cgroup still show up
136
-
* Control knob names and semantics aren't consistent with other
139
-
This patchset implements cpu controller's interface on the unified
140
-
hierarchy which adheres to the controller file conventions described
141
-
in Documentation/cgroups/unified-hierarchy.txt. Overall, the
142
-
following changes are made.
144
-
* cpuacct is implictly enabled and disabled by cpu and its information
145
-
is reported through "cpu.stat" which now uses microseconds for all
146
-
time durations. All time duration fields now have "_usec" appended
147
-
to them for clarity. While this doesn't solve the double accounting
148
-
immediately, once majority of users switch to v2, cpu can directly
149
-
account and report the relevant stats and cpuacct can be disabled on
150
-
the unified hierarchy.
152
-
Note that cpuacct.usage_percpu is currently not included in
153
-
"cpu.stat". If this information is actually called for, it can be
156
-
* "cpu.shares" is replaced with "cpu.weight" and operates on the
157
-
standard scale defined by CGROUP_WEIGHT_MIN/DFL/MAX (1, 100, 10000).
158
-
The weight is scaled to scheduler weight so that 100 maps to 1024
159
-
and the ratio relationship is preserved - if weight is W and its
160
-
scaled value is S, W / 100 == S / 1024. While the mapped range is a
161
-
bit smaller than the orignal scheduler weight range, the dead zones
162
-
on both sides are relatively small and covers wider range than the
163
-
nice value mappings. This file doesn't make sense in the root
164
-
cgroup and isn't create on root.
166
-
* "cpu.cfs_quota_us" and "cpu.cfs_period_us" are replaced by "cpu.max"
167
-
which contains both quota and period.
169
-
* "cpu.rt_runtime_us" and "cpu.rt_period_us" are replaced by
170
-
"cpu.rt.max" which contains both runtime and period.
172
-
v2: cpu_stats_show() was incorrectly using CONFIG_FAIR_GROUP_SCHED for
173
-
CFS bandwidth stats and also using raw division for u64. Use
174
-
CONFIG_CFS_BANDWITH and do_div() instead.
176
-
The semantics of "cpu.rt.max" is not fully decided yet. Dropped
179
-
Signed-off-by: Tejun Heo <tj@kernel.org>
180
-
Cc: Ingo Molnar <mingo@redhat.com>
181
-
Cc: Peter Zijlstra <peterz@infradead.org>
182
-
Cc: Li Zefan <lizefan@huawei.com>
183
-
Cc: Johannes Weiner <hannes@cmpxchg.org>
185
-
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
186
-
index a1b95e83fa87..f01d56e58a1b 100644
187
-
--- a/kernel/sched/core.c
188
-
+++ b/kernel/sched/core.c
189
-
@@ -7253,6 +7253,139 @@ static struct cftype cpu_legacy_files[] = {
190
-
{ } /* Terminate */
193
-
+static int cpu_stats_show(struct seq_file *sf, void *v)
195
-
+ cpuacct_cpu_stats_show(sf);
197
-
+#ifdef CONFIG_CFS_BANDWIDTH
199
-
+ struct task_group *tg = css_tg(seq_css(sf));
200
-
+ struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
201
-
+ u64 throttled_usec;
203
-
+ throttled_usec = cfs_b->throttled_time;
204
-
+ do_div(throttled_usec, NSEC_PER_USEC);
206
-
+ seq_printf(sf, "nr_periods %d\n"
207
-
+ "nr_throttled %d\n"
208
-
+ "throttled_usec %llu\n",
209
-
+ cfs_b->nr_periods, cfs_b->nr_throttled,
216
-
+#ifdef CONFIG_FAIR_GROUP_SCHED
217
-
+static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
218
-
+ struct cftype *cft)
220
-
+ struct task_group *tg = css_tg(css);
221
-
+ u64 weight = scale_load_down(tg->shares);
223
-
+ return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
226
-
+static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
227
-
+ struct cftype *cftype, u64 weight)
230
-
+ * cgroup weight knobs should use the common MIN, DFL and MAX
231
-
+ * values which are 1, 100 and 10000 respectively. While it loses
232
-
+ * a bit of range on both ends, it maps pretty well onto the shares
233
-
+ * value used by scheduler and the round-trip conversions preserve
234
-
+ * the original value over the entire range.
236
-
+ if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
239
-
+ weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
241
-
+ return sched_group_set_shares(css_tg(css), scale_load(weight));
245
-
+static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
246
-
+ long period, long quota)
249
-
+ seq_puts(sf, "max");
251
-
+ seq_printf(sf, "%ld", quota);
253
-
+ seq_printf(sf, " %ld\n", period);
256
-
+/* caller should put the current value in *@periodp before calling */
257
-
+static int __maybe_unused cpu_period_quota_parse(char *buf,
258
-
+ u64 *periodp, u64 *quotap)
260
-
+ char tok[21]; /* U64_MAX */
262
-
+ if (!sscanf(buf, "%s %llu", tok, periodp))
265
-
+ *periodp *= NSEC_PER_USEC;
267
-
+ if (sscanf(tok, "%llu", quotap))
268
-
+ *quotap *= NSEC_PER_USEC;
269
-
+ else if (!strcmp(tok, "max"))
270
-
+ *quotap = RUNTIME_INF;
277
-
+#ifdef CONFIG_CFS_BANDWIDTH
278
-
+static int cpu_max_show(struct seq_file *sf, void *v)
280
-
+ struct task_group *tg = css_tg(seq_css(sf));
282
-
+ cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
286
-
+static ssize_t cpu_max_write(struct kernfs_open_file *of,
287
-
+ char *buf, size_t nbytes, loff_t off)
289
-
+ struct task_group *tg = css_tg(of_css(of));
290
-
+ u64 period = tg_get_cfs_period(tg);
294
-
+ ret = cpu_period_quota_parse(buf, &period, "a);
296
-
+ ret = tg_set_cfs_bandwidth(tg, period, quota);
297
-
+ return ret ?: nbytes;
301
-
+static struct cftype cpu_files[] = {
304
-
+ .flags = CFTYPE_NOT_ON_ROOT,
305
-
+ .seq_show = cpu_stats_show,
307
-
+#ifdef CONFIG_FAIR_GROUP_SCHED
309
-
+ .name = "weight",
310
-
+ .flags = CFTYPE_NOT_ON_ROOT,
311
-
+ .read_u64 = cpu_weight_read_u64,
312
-
+ .write_u64 = cpu_weight_write_u64,
315
-
+#ifdef CONFIG_CFS_BANDWIDTH
318
-
+ .flags = CFTYPE_NOT_ON_ROOT,
319
-
+ .seq_show = cpu_max_show,
320
-
+ .write = cpu_max_write,
323
-
+ { } /* terminate */
326
-
struct cgroup_subsys cpu_cgrp_subsys = {
327
-
.css_alloc = cpu_cgroup_css_alloc,
328
-
.css_online = cpu_cgroup_css_online,
329
-
@@ -7262,7 +7395,15 @@ struct cgroup_subsys cpu_cgrp_subsys = {
330
-
.can_attach = cpu_cgroup_can_attach,
331
-
.attach = cpu_cgroup_attach,
332
-
.legacy_cftypes = cpu_legacy_files,
333
-
+ .dfl_cftypes = cpu_files,
334
-
.early_init = true,
335
-
+#ifdef CONFIG_CGROUP_CPUACCT
337
-
+ * cpuacct is enabled together with cpu on the unified hierarchy
338
-
+ * and its stats are reported through "cpu.stat".
340
-
+ .depends_on = 1 << cpuacct_cgrp_id,
344
-
#endif /* CONFIG_CGROUP_SCHED */
345
-
diff --git a/kernel/sched/cpuacct.c b/kernel/sched/cpuacct.c
346
-
index 6151c23f722f..fc1cf13c3af1 100644
347
-
--- a/kernel/sched/cpuacct.c
348
-
+++ b/kernel/sched/cpuacct.c
349
-
@@ -347,6 +347,31 @@ static struct cftype files[] = {
350
-
{ } /* terminate */
353
-
+/* used to print cpuacct stats in cpu.stat on the unified hierarchy */
354
-
+void cpuacct_cpu_stats_show(struct seq_file *sf)
356
-
+ struct cgroup_subsys_state *css;
357
-
+ u64 usage, val[CPUACCT_STAT_NSTATS];
359
-
+ css = cgroup_get_e_css(seq_css(sf)->cgroup, &cpuacct_cgrp_subsys);
361
-
+ usage = cpuusage_read(css, seq_cft(sf));
362
-
+ cpuacct_stats_read(css_ca(css), &val);
364
-
+ val[CPUACCT_STAT_USER] *= TICK_NSEC;
365
-
+ val[CPUACCT_STAT_SYSTEM] *= TICK_NSEC;
366
-
+ do_div(usage, NSEC_PER_USEC);
367
-
+ do_div(val[CPUACCT_STAT_USER], NSEC_PER_USEC);
368
-
+ do_div(val[CPUACCT_STAT_SYSTEM], NSEC_PER_USEC);
370
-
+ seq_printf(sf, "usage_usec %llu\n"
371
-
+ "user_usec %llu\n"
372
-
+ "system_usec %llu\n",
373
-
+ usage, val[CPUACCT_STAT_USER], val[CPUACCT_STAT_SYSTEM]);
379
-
* charge this task's execution time to its accounting group.
381
-
diff --git a/kernel/sched/cpuacct.h b/kernel/sched/cpuacct.h
382
-
index ba72807c73d4..ddf7af466d35 100644
383
-
--- a/kernel/sched/cpuacct.h
384
-
+++ b/kernel/sched/cpuacct.h
387
-
extern void cpuacct_charge(struct task_struct *tsk, u64 cputime);
388
-
extern void cpuacct_account_field(struct task_struct *tsk, int index, u64 val);
389
-
+extern void cpuacct_cpu_stats_show(struct seq_file *sf);
393
-
@@ -14,4 +15,8 @@ cpuacct_account_field(struct task_struct *tsk, int index, u64 val)
397
-
+static inline void cpuacct_cpu_stats_show(struct seq_file *sf)
403
-
commit 8dde150866b8c433216105c50b7e889d5242d583
404
-
Author: Tejun Heo <tj@kernel.org>
405
-
Date: Fri Aug 5 12:41:01 2016 -0400
407
-
cgroup: add documentation regarding CPU controller cgroup v2 support
409
-
Signed-off-by: Tejun Heo <tj@kernel.org>
411
-
diff --git a/Documentation/cgroup-v2-cpu.txt b/Documentation/cgroup-v2-cpu.txt
412
-
new file mode 100644
413
-
index 000000000000..1ed7032d4472
415
-
+++ b/Documentation/cgroup-v2-cpu.txt
419
-
+CPU Controller on Control Group v2
421
-
+August, 2016 Tejun Heo <tj@kernel.org>
424
-
+While most controllers have support for cgroup v2 now, the CPU
425
-
+controller support is not upstream yet due to objections from the
426
-
+scheduler maintainers on the basic designs of cgroup v2. This
427
-
+document explains the current situation as well as an interim
428
-
+solution, and details the disagreements and arguments. The latest
429
-
+version of this document can be found at the following URL.
431
-
+ https://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup.git/tree/Documentation/cgroup-v2-cpu.txt?h=cgroup-v2-cpu
433
-
+This document was posted to the linux-kernel and cgroup mailing lists.
434
-
+Unfortunately, no consensus was reached as of Oct, 2016. The thread
435
-
+can be found at the following URL.
437
-
+ http://lkml.kernel.org/r/20160805170752.GK2542@mtj.duckdns.org
442
-
+1. Current Situation and Interim Solution
443
-
+2. Disagreements and Arguments
444
-
+ 2-1. Contentious Restrictions
445
-
+ 2-1-1. Process Granularity
446
-
+ 2-1-2. No Internal Process Constraint
447
-
+ 2-2. Impact on CPU Controller
448
-
+ 2-2-1. Impact of Process Granularity
449
-
+ 2-2-2. Impact of No Internal Process Constraint
450
-
+ 2-3. Arguments for cgroup v2
455
-
+1. Current Situation and Interim Solution
457
-
+All objections from the scheduler maintainers apply to cgroup v2 core
458
-
+design, and there are no known objections to the specifics of the CPU
459
-
+controller cgroup v2 interface. The only blocked part is changes to
460
-
+expose the CPU controller interface on cgroup v2, which comprises the
461
-
+following two patches:
463
-
+ [1] sched: Misc preps for cgroup unified hierarchy interface
464
-
+ [2] sched: Implement interface for cgroup unified hierarchy
466
-
+The necessary changes are superficial and implement the interface
467
-
+files on cgroup v2. The combined diffstat is as follows.
469
-
+ kernel/sched/core.c | 149 +++++++++++++++++++++++++++++++++++++++++++++++--
470
-
+ kernel/sched/cpuacct.c | 57 ++++++++++++------
471
-
+ kernel/sched/cpuacct.h | 5 +
472
-
+ 3 files changed, 189 insertions(+), 22 deletions(-)
474
-
+The patches are easy to apply and forward-port. The following git
475
-
+branch will always carry the two patches on top of the latest release
476
-
+of the upstream kernel.
478
-
+ git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup.git/cgroup-v2-cpu
480
-
+There also are versioned branches going back to v4.4.
482
-
+ git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup.git/cgroup-v2-cpu-$KERNEL_VER
484
-
+While it's difficult to tell whether the CPU controller support will
485
-
+be merged, there are crucial resource control features in cgroup v2
486
-
+that are only possible due to the design choices that are being
487
-
+objected to, and every effort will be made to ease enabling the CPU
488
-
+controller cgroup v2 support out-of-tree for parties which choose to.
491
-
+2. Disagreements and Arguments
493
-
+There have been several lengthy discussion threads [3][4] on LKML
494
-
+around the structural constraints of cgroup v2. The two that affect
495
-
+the CPU controller are process granularity and no internal process
496
-
+constraint. Both arise primarily from the need for common resource
497
-
+domain definition across different resources.
499
-
+The common resource domain is a powerful concept in cgroup v2 that
500
-
+allows controllers to make basic assumptions about the structural
501
-
+organization of processes and controllers inside the cgroup hierarchy,
502
-
+and thus solve problems spanning multiple types of resources. The
503
-
+prime example for this is page cache writeback: dirty page cache is
504
-
+regulated through throttling buffered writers based on memory
505
-
+availability, and initiating batched write outs to the disk based on
506
-
+IO capacity. Tracking and controlling writeback inside a cgroup thus
507
-
+requires the direct cooperation of the memory and the IO controller.
509
-
+This easily extends to other areas, such as CPU cycles consumed while
510
-
+performing memory reclaim or IO encryption.
513
-
+2-1. Contentious Restrictions
515
-
+For controllers of different resources to work together, they must
516
-
+agree on a common organization. This uniform model across controllers
517
-
+imposes two contentious restrictions on the CPU controller: process
518
-
+granularity and the no-internal-process constraint.
521
-
+ 2-1-1. Process Granularity
523
-
+ For memory, because an address space is shared between all threads
524
-
+ of a process, the terminal consumer is a process, not a thread.
525
-
+ Separating the threads of a single process into different memory
526
-
+ control domains doesn't make semantical sense. cgroup v2 ensures
527
-
+ that all controller can agree on the same organization by requiring
528
-
+ that threads of the same process belong to the same cgroup.
530
-
+ There are other reasons to enforce process granularity. One
531
-
+ important one is isolating system-level management operations from
532
-
+ in-process application operations. The cgroup interface, being a
533
-
+ virtual filesystem, is very unfit for multiple independent
534
-
+ operations taking place at the same time as most operations have to
535
-
+ be multi-step and there is no way to synchronize multiple accessors.
536
-
+ See also [5] Documentation/cgroup-v2.txt, "R-2. Thread Granularity"
539
-
+ 2-1-2. No Internal Process Constraint
541
-
+ cgroup v2 does not allow processes to belong to any cgroup which has
542
-
+ child cgroups when resource controllers are enabled on it (the
543
-
+ notable exception being the root cgroup itself). This is because,
544
-
+ for some resources, a resource domain (cgroup) is not directly
545
-
+ comparable to the terminal consumer (process/task) of said resource,
546
-
+ and so putting the two into a sibling relationship isn't meaningful.
548
-
+ - Differing Control Parameters and Capabilities
550
-
+ A cgroup controller has different resource control parameters and
551
-
+ capabilities from a terminal consumer, be that a task or process.
552
-
+ There are a couple cases where a cgroup control knob can be mapped
553
-
+ to a per-task or per-process API but they are exceptions and the
554
-
+ mappings aren't obvious even in those cases.
556
-
+ For example, task priorities (also known as nice values) set
557
-
+ through setpriority(2) are mapped to the CPU controller
558
-
+ "cpu.shares" values. However, how exactly the two ranges map and
559
-
+ even the fact that they map to each other at all are not obvious.
561
-
+ The situation gets further muddled when considering other resource
562
-
+ types and control knobs. IO priorities set through ioprio_set(2)
563
-
+ cannot be mapped to IO controller weights and most cgroup resource
564
-
+ control knobs including the bandwidth control knobs of the CPU
565
-
+ controller don't have counterparts in the terminal consumers.
567
-
+ - Anonymous Resource Consumption
569
-
+ For CPU, every time slice consumed from inside a cgroup, which
570
-
+ comprises most but not all of consumed CPU time for the cgroup,
571
-
+ can be clearly attributed to a specific task or process. Because
572
-
+ these two types of entities are directly comparable as consumers
573
-
+ of CPU time, it's theoretically possible to mix tasks and cgroups
574
-
+ on the same tree levels and let them directly compete for the time
575
-
+ quota available to their common ancestor.
577
-
+ However, the same can't be said for resource types like memory or
578
-
+ IO: the memory consumed by the page cache, for example, can be
579
-
+ tracked on a per-cgroup level, but due to mismatches in lifetimes
580
-
+ of involved objects (page cache can persist long after processes
581
-
+ are gone), shared usages and the implementation overhead of
582
-
+ tracking persistent state, it can no longer be attributed to
583
-
+ individual processes after instantiation. Consequently, any IO
584
-
+ incurred by page cache writeback can be attributed to a cgroup,
585
-
+ but not to the individual consumers inside the cgroup.
587
-
+ For memory and IO, this makes a resource domain (cgroup) an object
588
-
+ of a fundamentally different type than a terminal consumer
589
-
+ (process). A process can't be a first class object in the resource
590
-
+ distribution graph as its total resource consumption can't be
591
-
+ described without the containing resource domain.
593
-
+ Disallowing processes in internal cgroups avoids competition between
594
-
+ cgroups and processes which cannot be meaningfully defined for these
595
-
+ resources. All resource control takes place among cgroups and a
596
-
+ terminal consumer interacts with the containing cgroup the same way
597
-
+ it would with the system without cgroup.
599
-
+ Root cgroup is exempt from this constraint, which is in line with
600
-
+ how root cgroup is handled in general - it's excluded from cgroup
601
-
+ resource accounting and control.
604
-
+Enforcing process granularity and no internal process constraint
605
-
+allows all controllers to be on the same footing in terms of resource
606
-
+distribution hierarchy.
609
-
+2-2. Impact on CPU Controller
611
-
+As indicated earlier, the CPU controller's resource distribution graph
612
-
+is the simplest. Every schedulable resource consumption can be
613
-
+attributed to a specific task. In addition, for weight based control,
614
-
+the per-task priority set through setpriority(2) can be translated to
615
-
+and from a per-cgroup weight. As such, the CPU controller can treat a
616
-
+task and a cgroup symmetrically, allowing support for any tree layout
617
-
+of cgroups and tasks. Both process granularity and the no internal
618
-
+process constraint restrict how the CPU controller can be used.
621
-
+ 2-2-1. Impact of Process Granularity
623
-
+ Process granularity prevents tasks belonging to the same process to
624
-
+ be assigned to different cgroups. It was pointed out [6] that this
625
-
+ excludes the valid use case of hierarchical CPU distribution within
628
-
+ To address this issue, the rgroup (resource group) [7][8][9]
629
-
+ interface, an extension of the existing setpriority(2) API, was
630
-
+ proposed, which is in line with other programmable priority
631
-
+ mechanisms and eliminates the risk of in-application configuration
632
-
+ and system configuration stepping on each other's toes.
633
-
+ Unfortunately, the proposal quickly turned into discussions around
634
-
+ cgroup v2 design decisions [4] and no consensus could be reached.
637
-
+ 2-2-2. Impact of No Internal Process Constraint
639
-
+ The no internal process constraint disallows tasks from competing
640
-
+ directly against cgroups. Here is an excerpt from Peter Zijlstra
641
-
+ pointing out the issue [10] - R, L and A are cgroups; t1, t2, t3 and
652
-
+ Is fundamentally different from:
662
-
+ Because if in the first hierarchy you add a task (t5) to R, all of
663
-
+ its A will run at 1/4th of total bandwidth where before it had
664
-
+ 1/3rd, whereas with the second example, if you add our t5 to L, A
665
-
+ doesn't get any less bandwidth.
668
-
+ It is true that the trees are semantically different from each other
669
-
+ and the symmetric handling of tasks and cgroups is aesthetically
670
-
+ pleasing. However, it isn't clear what the practical usefulness of
671
-
+ a layout with direct competition between tasks and cgroups would be,
672
-
+ considering that number and behavior of tasks are controlled by each
673
-
+ application, and cgroups primarily deal with system level resource
674
-
+ distribution; changes in the number of active threads would directly
675
-
+ impact resource distribution. Real world use cases of such layouts
676
-
+ could not be established during the discussions.
679
-
+2-3. Arguments for cgroup v2
681
-
+There are strong demands for comprehensive hierarchical resource
682
-
+control across all major resources, and establishing a common resource
683
-
+hierarchy is an essential step. As with most engineering decisions,
684
-
+common resource hierarchy definition comes with its trade-offs. With
685
-
+cgroup v2, the trade-offs are in the form of structural constraints
686
-
+which, among others, restrict the CPU controller's space of possible
689
-
+However, even with the restrictions, cgroup v2, in combination with
690
-
+rgroup, covers most of identified real world use cases while enabling
691
-
+new important use cases of resource control across multiple resource
692
-
+types that were fundamentally broken previously.
694
-
+Furthermore, for resource control, treating resource domains as
695
-
+objects of a different type from terminal consumers has important
696
-
+advantages - it can account for resource consumptions which are not
697
-
+tied to any specific terminal consumer, be that a task or process, and
698
-
+allows decoupling resource distribution controls from in-application
699
-
+APIs. Even the CPU controller may benefit from it as the kernel can
700
-
+consume significant amount of CPU cycles in interrupt context or tasks
701
-
+shared across multiple resource domains (e.g. softirq).
703
-
+Finally, it's important to note that enabling cgroup v2 support for
704
-
+the CPU controller doesn't block use cases which require the features
705
-
+which are not available on cgroup v2. Unlikely, but should anybody
706
-
+actually rely on the CPU controller's symmetric handling of tasks and
707
-
+cgroups, backward compatibility is and will be maintained by being
708
-
+able to disconnect the controller from the cgroup v2 hierarchy and use
709
-
+it standalone. This also holds for cpuset which is often used in
710
-
+highly customized configurations which might be a poor fit for common
713
-
+The required changes are minimal, the benefits for the target use
714
-
+cases are critical and obvious, and use cases which have to use v1 can
715
-
+continue to do so.
720
-
+cgroup v2 primarily aims to solve the problem of comprehensive
721
-
+hierarchical resource control across all major computing resources,
722
-
+which is one of the core problems of modern server infrastructure
723
-
+engineering. The trade-offs that cgroup v2 took are results of
724
-
+pursuing that goal and gaining a better understanding of the nature of
725
-
+resource control in the process.
727
-
+I believe that real world usages will prove cgroup v2's model right,
728
-
+considering the crucial pieces of comprehensive resource control that
729
-
+cannot be implemented without common resource domains. This is not to
730
-
+say that cgroup v2 is fixed in stone and can't be updated; if there is
731
-
+an approach which better serves both comprehensive resource control
732
-
+and the CPU controller's flexibility, we will surely move towards
733
-
+that. It goes without saying that discussions around such approach
734
-
+should consider practical aspects of resource control as a whole
735
-
+rather than absolutely focusing on a particular controller.
737
-
+Until such consensus can be reached, the CPU controller cgroup v2
738
-
+support will be maintained out of the mainline kernel in an easily
739
-
+accessible form. If there is anything cgroup developers can do to
740
-
+ease the pain, please feel free to contact us on the cgroup mailing
741
-
+list at cgroups@vger.kernel.org.
746
-
+[1] http://lkml.kernel.org/r/20160105164834.GE5995@mtj.duckdns.org
747
-
+ [PATCH 1/2] sched: Misc preps for cgroup unified hierarchy interface
748
-
+ Tejun Heo <tj@kernel.org>
750
-
+[2] http://lkml.kernel.org/r/20160105164852.GF5995@mtj.duckdns.org
751
-
+ [PATCH 2/2] sched: Implement interface for cgroup unified hierarchy
752
-
+ Tejun Heo <tj@kernel.org>
754
-
+[3] http://lkml.kernel.org/r/1438641689-14655-4-git-send-email-tj@kernel.org
755
-
+ [PATCH 3/3] sched: Implement interface for cgroup unified hierarchy
756
-
+ Tejun Heo <tj@kernel.org>
758
-
+[4] http://lkml.kernel.org/r/20160407064549.GH3430@twins.programming.kicks-ass.net
759
-
+ Re: [PATCHSET RFC cgroup/for-4.6] cgroup, sched: implement resource group and PRIO_RGRP
760
-
+ Peter Zijlstra <peterz@infradead.org>
762
-
+[5] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/cgroup-v2.txt
764
-
+ Tejun Heo <tj@kernel.org>
766
-
+[6] http://lkml.kernel.org/r/CAPM31RJNy3jgG=DYe6GO=wyL4BPPxwUm1f2S6YXacQmo7viFZA@mail.gmail.com
767
-
+ Re: [PATCH 3/3] sched: Implement interface for cgroup unified hierarchy
768
-
+ Paul Turner <pjt@google.com>
770
-
+[7] http://lkml.kernel.org/r/20160105154503.GC5995@mtj.duckdns.org
771
-
+ [RFD] cgroup: thread granularity support for cpu controller
772
-
+ Tejun Heo <tj@kernel.org>
774
-
+[8] http://lkml.kernel.org/r/1457710888-31182-1-git-send-email-tj@kernel.org
775
-
+ [PATCHSET RFC cgroup/for-4.6] cgroup, sched: implement resource group and PRIO_RGRP
776
-
+ Tejun Heo <tj@kernel.org>
778
-
+[9] http://lkml.kernel.org/r/20160311160522.GA24046@htj.duckdns.org
779
-
+ Example program for PRIO_RGRP
780
-
+ Tejun Heo <tj@kernel.org>
782
-
+[10] http://lkml.kernel.org/r/20160407082810.GN3430@twins.programming.kicks-ass.net
783
-
+ Re: [PATCHSET RFC cgroup/for-4.6] cgroup, sched: implement resource
784
-
+ Peter Zijlstra <peterz@infradead.org>
pyrox.dev