From ee08d1284ea9235b29bd2d9b7493b4b4cf3da09c Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Wed, 13 Jun 2012 15:24:45 +0200 Subject: [PATCH 01/10] sched/x86: Remove broken power estimation The x86 sched power implementation has been broken forever and gets in the way of other stuff, remove it. [ For archaeological interest, fixing this code would require dealing with the cross-cpu calling of these functions and more importantly, we need to filter idle time out of the a/m-perf stuff because the ratio will go down to 0 when idle, giving a 0 capacity which is not what we'd want. ] Signed-off-by: Peter Zijlstra Cc: Linus Torvalds Cc: Andrew Morton Cc: Arjan van de Ven Link: http://lkml.kernel.org/r/1339594110.8980.38.camel@twins Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/Makefile | 2 +- arch/x86/kernel/cpu/sched.c | 55 ------------------------------------ 2 files changed, 1 insertion(+), 56 deletions(-) delete mode 100644 arch/x86/kernel/cpu/sched.c diff --git a/arch/x86/kernel/cpu/Makefile b/arch/x86/kernel/cpu/Makefile index 6ab6aa2fdfdd..c5981267a60c 100644 --- a/arch/x86/kernel/cpu/Makefile +++ b/arch/x86/kernel/cpu/Makefile @@ -14,7 +14,7 @@ CFLAGS_common.o := $(nostackp) obj-y := intel_cacheinfo.o scattered.o topology.o obj-y += proc.o capflags.o powerflags.o common.o -obj-y += vmware.o hypervisor.o sched.o mshyperv.o +obj-y += vmware.o hypervisor.o mshyperv.o obj-y += rdrand.o obj-y += match.o diff --git a/arch/x86/kernel/cpu/sched.c b/arch/x86/kernel/cpu/sched.c deleted file mode 100644 index a640ae5ad201..000000000000 --- a/arch/x86/kernel/cpu/sched.c +++ /dev/null @@ -1,55 +0,0 @@ -#include -#include -#include -#include - -#include -#include - -#ifdef CONFIG_SMP - -static DEFINE_PER_CPU(struct aperfmperf, old_perf_sched); - -static unsigned long scale_aperfmperf(void) -{ - struct aperfmperf val, *old = &__get_cpu_var(old_perf_sched); - unsigned long ratio, flags; - - local_irq_save(flags); - get_aperfmperf(&val); - local_irq_restore(flags); - - ratio = calc_aperfmperf_ratio(old, &val); - *old = val; - - return ratio; -} - -unsigned long arch_scale_freq_power(struct sched_domain *sd, int cpu) -{ - /* - * do aperf/mperf on the cpu level because it includes things - * like turbo mode, which are relevant to full cores. - */ - if (boot_cpu_has(X86_FEATURE_APERFMPERF)) - return scale_aperfmperf(); - - /* - * maybe have something cpufreq here - */ - - return default_scale_freq_power(sd, cpu); -} - -unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu) -{ - /* - * aperf/mperf already includes the smt gain - */ - if (boot_cpu_has(X86_FEATURE_APERFMPERF)) - return SCHED_LOAD_SCALE; - - return default_scale_smt_power(sd, cpu); -} - -#endif From d35be8bab9b0ce44bed4b9453f86ebf64062721e Mon Sep 17 00:00:00 2001 From: "Srivatsa S. Bhat" Date: Thu, 24 May 2012 19:46:26 +0530 Subject: [PATCH 02/10] CPU hotplug, cpusets, suspend: Don't modify cpusets during suspend/resume In the event of CPU hotplug, the kernel modifies the cpusets' cpus_allowed masks as and when necessary to ensure that the tasks belonging to the cpusets have some place (online CPUs) to run on. And regular CPU hotplug is destructive in the sense that the kernel doesn't remember the original cpuset configurations set by the user, across hotplug operations. However, suspend/resume (which uses CPU hotplug) is a special case in which the kernel has the responsibility to restore the system (during resume), to exactly the same state it was in before suspend. In order to achieve that, do the following: 1. Don't modify cpusets during suspend/resume. At all. In particular, don't move the tasks from one cpuset to another, and don't modify any cpuset's cpus_allowed mask. So, simply ignore cpusets during the CPU hotplug operations that are carried out in the suspend/resume path. 2. However, cpusets and sched domains are related. We just want to avoid altering cpusets alone. So, to keep the sched domains updated, build a single sched domain (containing all active cpus) during each of the CPU hotplug operations carried out in s/r path, effectively ignoring the cpusets' cpus_allowed masks. (Since userspace is frozen while doing all this, it will go unnoticed.) 3. During the last CPU online operation during resume, build the sched domains by looking up the (unaltered) cpusets' cpus_allowed masks. That will bring back the system to the same original state as it was in before suspend. Ultimately, this will not only solve the cpuset problem related to suspend resume (ie., restores the cpusets to exactly what it was before suspend, by not touching it at all) but also speeds up suspend/resume because we avoid running cpuset update code for every CPU being offlined/onlined. Signed-off-by: Srivatsa S. Bhat Signed-off-by: Peter Zijlstra Cc: Linus Torvalds Cc: Andrew Morton Cc: Thomas Gleixner Link: http://lkml.kernel.org/r/20120524141611.3692.20155.stgit@srivatsabhat.in.ibm.com Signed-off-by: Ingo Molnar --- kernel/cpuset.c | 3 +++ kernel/sched/core.c | 40 ++++++++++++++++++++++++++++++++++++---- 2 files changed, 39 insertions(+), 4 deletions(-) diff --git a/kernel/cpuset.c b/kernel/cpuset.c index 8c8bd652dd12..746d1eeb5dbe 100644 --- a/kernel/cpuset.c +++ b/kernel/cpuset.c @@ -2054,6 +2054,9 @@ static void scan_for_empty_cpusets(struct cpuset *root) * (of no affect) on systems that are actively using CPU hotplug * but making no active use of cpusets. * + * The only exception to this is suspend/resume, where we don't + * modify cpusets at all. + * * This routine ensures that top_cpuset.cpus_allowed tracks * cpu_active_mask on each CPU hotplug (cpuhp) event. * diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 468bdd44c1ba..4c1d80c6b318 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -7097,34 +7097,66 @@ void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], mutex_unlock(&sched_domains_mutex); } +static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ + /* * Update cpusets according to cpu_active mask. If cpusets are * disabled, cpuset_update_active_cpus() becomes a simple wrapper * around partition_sched_domains(). + * + * If we come here as part of a suspend/resume, don't touch cpusets because we + * want to restore it back to its original state upon resume anyway. */ static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, void *hcpu) { - switch (action & ~CPU_TASKS_FROZEN) { + switch (action) { + case CPU_ONLINE_FROZEN: + case CPU_DOWN_FAILED_FROZEN: + + /* + * num_cpus_frozen tracks how many CPUs are involved in suspend + * resume sequence. As long as this is not the last online + * operation in the resume sequence, just build a single sched + * domain, ignoring cpusets. + */ + num_cpus_frozen--; + if (likely(num_cpus_frozen)) { + partition_sched_domains(1, NULL, NULL); + break; + } + + /* + * This is the last CPU online operation. So fall through and + * restore the original sched domains by considering the + * cpuset configurations. + */ + case CPU_ONLINE: case CPU_DOWN_FAILED: cpuset_update_active_cpus(); - return NOTIFY_OK; + break; default: return NOTIFY_DONE; } + return NOTIFY_OK; } static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, void *hcpu) { - switch (action & ~CPU_TASKS_FROZEN) { + switch (action) { case CPU_DOWN_PREPARE: cpuset_update_active_cpus(); - return NOTIFY_OK; + break; + case CPU_DOWN_PREPARE_FROZEN: + num_cpus_frozen++; + partition_sched_domains(1, NULL, NULL); + break; default: return NOTIFY_DONE; } + return NOTIFY_OK; } void __init sched_init_smp(void) From 80d1fa6463d934969b7aebf04107fc133463f0f6 Mon Sep 17 00:00:00 2001 From: "Srivatsa S. Bhat" Date: Thu, 24 May 2012 19:46:41 +0530 Subject: [PATCH 03/10] cpusets, hotplug: Implement cpuset tree traversal in a helper function At present, the functions that deal with cpusets during CPU/Mem hotplug are quite messy, since a lot of the functionality is mixed up without clear separation. And this takes a toll on optimization as well. For example, the function cpuset_update_active_cpus() is called on both CPU offline and CPU online events; and it invokes scan_for_empty_cpusets(), which makes sense only for CPU offline events. And hence, the current code ends up unnecessarily traversing the cpuset tree during CPU online also. As a first step towards cleaning up those functions, encapsulate the cpuset tree traversal in a helper function, so as to facilitate upcoming changes. Signed-off-by: Srivatsa S. Bhat Signed-off-by: Peter Zijlstra Cc: Linus Torvalds Cc: Andrew Morton Cc: Thomas Gleixner Link: http://lkml.kernel.org/r/20120524141635.3692.893.stgit@srivatsabhat.in.ibm.com Signed-off-by: Ingo Molnar --- kernel/cpuset.c | 36 +++++++++++++++++++++++++++--------- 1 file changed, 27 insertions(+), 9 deletions(-) diff --git a/kernel/cpuset.c b/kernel/cpuset.c index 746d1eeb5dbe..ba96349aa522 100644 --- a/kernel/cpuset.c +++ b/kernel/cpuset.c @@ -1989,6 +1989,32 @@ static void remove_tasks_in_empty_cpuset(struct cpuset *cs) move_member_tasks_to_cpuset(cs, parent); } +/* + * Helper function to traverse cpusets. + * It can be used to walk the cpuset tree from top to bottom, completing + * one layer before dropping down to the next (thus always processing a + * node before any of its children). + */ +static struct cpuset *cpuset_next(struct list_head *queue) +{ + struct cpuset *cp; + struct cpuset *child; /* scans child cpusets of cp */ + struct cgroup *cont; + + if (list_empty(queue)) + return NULL; + + cp = list_first_entry(queue, struct cpuset, stack_list); + list_del(queue->next); + list_for_each_entry(cont, &cp->css.cgroup->children, sibling) { + child = cgroup_cs(cont); + list_add_tail(&child->stack_list, queue); + } + + return cp; +} + + /* * Walk the specified cpuset subtree and look for empty cpusets. * The tasks of such cpuset must be moved to a parent cpuset. @@ -2008,19 +2034,11 @@ static void scan_for_empty_cpusets(struct cpuset *root) { LIST_HEAD(queue); struct cpuset *cp; /* scans cpusets being updated */ - struct cpuset *child; /* scans child cpusets of cp */ - struct cgroup *cont; static nodemask_t oldmems; /* protected by cgroup_mutex */ list_add_tail((struct list_head *)&root->stack_list, &queue); - while (!list_empty(&queue)) { - cp = list_first_entry(&queue, struct cpuset, stack_list); - list_del(queue.next); - list_for_each_entry(cont, &cp->css.cgroup->children, sibling) { - child = cgroup_cs(cont); - list_add_tail(&child->stack_list, &queue); - } + while ((cp = cpuset_next(&queue)) != NULL) { /* Continue past cpusets with all cpus, mems online */ if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) && From 7ddf96b02fe8dd441f452deef879040def5f7b34 Mon Sep 17 00:00:00 2001 From: "Srivatsa S. Bhat" Date: Thu, 24 May 2012 19:46:55 +0530 Subject: [PATCH 04/10] cpusets, hotplug: Restructure functions that are invoked during hotplug Separate out the cpuset related handling for CPU/Memory online/offline. This also helps us exploit the most obvious and basic level of optimization that any notification mechanism (CPU/Mem online/offline) has to offer us: "We *know* why we have been invoked. So stop pretending that we are lost, and do only the necessary amount of processing!". And while at it, rename scan_for_empty_cpusets() to scan_cpusets_upon_hotplug(), which is more appropriate considering how it is restructured. Signed-off-by: Srivatsa S. Bhat Signed-off-by: Peter Zijlstra Cc: Linus Torvalds Cc: Andrew Morton Cc: Thomas Gleixner Link: http://lkml.kernel.org/r/20120524141650.3692.48637.stgit@srivatsabhat.in.ibm.com Signed-off-by: Ingo Molnar --- include/linux/cpuset.h | 4 +- kernel/cpuset.c | 88 +++++++++++++++++++++++++++++------------- kernel/sched/core.c | 4 +- 3 files changed, 65 insertions(+), 31 deletions(-) diff --git a/include/linux/cpuset.h b/include/linux/cpuset.h index 668f66baac7b..838320fc3d1d 100644 --- a/include/linux/cpuset.h +++ b/include/linux/cpuset.h @@ -20,7 +20,7 @@ extern int number_of_cpusets; /* How many cpusets are defined in system? */ extern int cpuset_init(void); extern void cpuset_init_smp(void); -extern void cpuset_update_active_cpus(void); +extern void cpuset_update_active_cpus(bool cpu_online); extern void cpuset_cpus_allowed(struct task_struct *p, struct cpumask *mask); extern void cpuset_cpus_allowed_fallback(struct task_struct *p); extern nodemask_t cpuset_mems_allowed(struct task_struct *p); @@ -124,7 +124,7 @@ static inline void set_mems_allowed(nodemask_t nodemask) static inline int cpuset_init(void) { return 0; } static inline void cpuset_init_smp(void) {} -static inline void cpuset_update_active_cpus(void) +static inline void cpuset_update_active_cpus(bool cpu_online) { partition_sched_domains(1, NULL, NULL); } diff --git a/kernel/cpuset.c b/kernel/cpuset.c index ba96349aa522..ba0a4d74d262 100644 --- a/kernel/cpuset.c +++ b/kernel/cpuset.c @@ -147,6 +147,12 @@ typedef enum { CS_SPREAD_SLAB, } cpuset_flagbits_t; +/* the type of hotplug event */ +enum hotplug_event { + CPUSET_CPU_OFFLINE, + CPUSET_MEM_OFFLINE, +}; + /* convenient tests for these bits */ static inline int is_cpu_exclusive(const struct cpuset *cs) { @@ -2016,8 +2022,10 @@ static struct cpuset *cpuset_next(struct list_head *queue) /* - * Walk the specified cpuset subtree and look for empty cpusets. - * The tasks of such cpuset must be moved to a parent cpuset. + * Walk the specified cpuset subtree upon a hotplug operation (CPU/Memory + * online/offline) and update the cpusets accordingly. + * For regular CPU/Mem hotplug, look for empty cpusets; the tasks of such + * cpuset must be moved to a parent cpuset. * * Called with cgroup_mutex held. We take callback_mutex to modify * cpus_allowed and mems_allowed. @@ -2030,38 +2038,58 @@ static struct cpuset *cpuset_next(struct list_head *queue) * that has tasks along with an empty 'mems'. But if we did see such * a cpuset, we'd handle it just like we do if its 'cpus' was empty. */ -static void scan_for_empty_cpusets(struct cpuset *root) +static void +scan_cpusets_upon_hotplug(struct cpuset *root, enum hotplug_event event) { LIST_HEAD(queue); - struct cpuset *cp; /* scans cpusets being updated */ + struct cpuset *cp; /* scans cpusets being updated */ static nodemask_t oldmems; /* protected by cgroup_mutex */ list_add_tail((struct list_head *)&root->stack_list, &queue); - while ((cp = cpuset_next(&queue)) != NULL) { + switch (event) { + case CPUSET_CPU_OFFLINE: + while ((cp = cpuset_next(&queue)) != NULL) { - /* Continue past cpusets with all cpus, mems online */ - if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) && - nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY])) - continue; + /* Continue past cpusets with all cpus online */ + if (cpumask_subset(cp->cpus_allowed, cpu_active_mask)) + continue; - oldmems = cp->mems_allowed; + /* Remove offline cpus from this cpuset. */ + mutex_lock(&callback_mutex); + cpumask_and(cp->cpus_allowed, cp->cpus_allowed, + cpu_active_mask); + mutex_unlock(&callback_mutex); - /* Remove offline cpus and mems from this cpuset. */ - mutex_lock(&callback_mutex); - cpumask_and(cp->cpus_allowed, cp->cpus_allowed, - cpu_active_mask); - nodes_and(cp->mems_allowed, cp->mems_allowed, + /* Move tasks from the empty cpuset to a parent */ + if (cpumask_empty(cp->cpus_allowed)) + remove_tasks_in_empty_cpuset(cp); + else + update_tasks_cpumask(cp, NULL); + } + break; + + case CPUSET_MEM_OFFLINE: + while ((cp = cpuset_next(&queue)) != NULL) { + + /* Continue past cpusets with all mems online */ + if (nodes_subset(cp->mems_allowed, + node_states[N_HIGH_MEMORY])) + continue; + + oldmems = cp->mems_allowed; + + /* Remove offline mems from this cpuset. */ + mutex_lock(&callback_mutex); + nodes_and(cp->mems_allowed, cp->mems_allowed, node_states[N_HIGH_MEMORY]); - mutex_unlock(&callback_mutex); + mutex_unlock(&callback_mutex); - /* Move tasks from the empty cpuset to a parent */ - if (cpumask_empty(cp->cpus_allowed) || - nodes_empty(cp->mems_allowed)) - remove_tasks_in_empty_cpuset(cp); - else { - update_tasks_cpumask(cp, NULL); - update_tasks_nodemask(cp, &oldmems, NULL); + /* Move tasks from the empty cpuset to a parent */ + if (nodes_empty(cp->mems_allowed)) + remove_tasks_in_empty_cpuset(cp); + else + update_tasks_nodemask(cp, &oldmems, NULL); } } } @@ -2080,8 +2108,11 @@ static void scan_for_empty_cpusets(struct cpuset *root) * * Called within get_online_cpus(). Needs to call cgroup_lock() * before calling generate_sched_domains(). + * + * @cpu_online: Indicates whether this is a CPU online event (true) or + * a CPU offline event (false). */ -void cpuset_update_active_cpus(void) +void cpuset_update_active_cpus(bool cpu_online) { struct sched_domain_attr *attr; cpumask_var_t *doms; @@ -2091,7 +2122,10 @@ void cpuset_update_active_cpus(void) mutex_lock(&callback_mutex); cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask); mutex_unlock(&callback_mutex); - scan_for_empty_cpusets(&top_cpuset); + + if (!cpu_online) + scan_cpusets_upon_hotplug(&top_cpuset, CPUSET_CPU_OFFLINE); + ndoms = generate_sched_domains(&doms, &attr); cgroup_unlock(); @@ -2122,9 +2156,9 @@ static int cpuset_track_online_nodes(struct notifier_block *self, case MEM_OFFLINE: /* * needn't update top_cpuset.mems_allowed explicitly because - * scan_for_empty_cpusets() will update it. + * scan_cpusets_upon_hotplug() will update it. */ - scan_for_empty_cpusets(&top_cpuset); + scan_cpusets_upon_hotplug(&top_cpuset, CPUSET_MEM_OFFLINE); break; default: break; diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 4c1d80c6b318..4b4a63d34396 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -7134,7 +7134,7 @@ static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, case CPU_ONLINE: case CPU_DOWN_FAILED: - cpuset_update_active_cpus(); + cpuset_update_active_cpus(true); break; default: return NOTIFY_DONE; @@ -7147,7 +7147,7 @@ static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, { switch (action) { case CPU_DOWN_PREPARE: - cpuset_update_active_cpus(); + cpuset_update_active_cpus(false); break; case CPU_DOWN_PREPARE_FROZEN: num_cpus_frozen++; From a1cd2b13f754b2c56fb87b8c4912c015f8f57c0c Mon Sep 17 00:00:00 2001 From: "Srivatsa S. Bhat" Date: Thu, 24 May 2012 19:47:03 +0530 Subject: [PATCH 05/10] cpusets: Remove/update outdated comments cpuset_track_online_cpus() is no longer present. So remove the outdated comment and replace it with reference to cpuset_update_active_cpus() which is its equivalent. Also, we don't lack memory hot-unplug anymore. And David Rientjes pointed out how it is dealt with. So update that comment as well. Signed-off-by: Srivatsa S. Bhat Signed-off-by: Peter Zijlstra Cc: Linus Torvalds Cc: Andrew Morton Cc: Thomas Gleixner Link: http://lkml.kernel.org/r/20120524141700.3692.98192.stgit@srivatsabhat.in.ibm.com Signed-off-by: Ingo Molnar --- kernel/cpuset.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/kernel/cpuset.c b/kernel/cpuset.c index ba0a4d74d262..f33c7153b6d7 100644 --- a/kernel/cpuset.c +++ b/kernel/cpuset.c @@ -2034,9 +2034,8 @@ static struct cpuset *cpuset_next(struct list_head *queue) * before dropping down to the next. It always processes a node before * any of its children. * - * For now, since we lack memory hot unplug, we'll never see a cpuset - * that has tasks along with an empty 'mems'. But if we did see such - * a cpuset, we'd handle it just like we do if its 'cpus' was empty. + * In the case of memory hot-unplug, it will remove nodes from N_HIGH_MEMORY + * if all present pages from a node are offlined. */ static void scan_cpusets_upon_hotplug(struct cpuset *root, enum hotplug_event event) @@ -2137,7 +2136,7 @@ void cpuset_update_active_cpus(bool cpu_online) /* * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY]. * Call this routine anytime after node_states[N_HIGH_MEMORY] changes. - * See also the previous routine cpuset_track_online_cpus(). + * See cpuset_update_active_cpus() for CPU hotplug handling. */ static int cpuset_track_online_nodes(struct notifier_block *self, unsigned long action, void *arg) From 970e178985cadbca660feb02f4d2ee3a09f7fdda Mon Sep 17 00:00:00 2001 From: Mike Galbraith Date: Tue, 12 Jun 2012 05:18:32 +0200 Subject: [PATCH 06/10] sched: Improve scalability via 'CPU buddies', which withstand random perturbations Traversing an entire package is not only expensive, it also leads to tasks bouncing all over a partially idle and possible quite large package. Fix that up by assigning a 'buddy' CPU to try to motivate. Each buddy may try to motivate that one other CPU, if it's busy, tough, it may then try its SMT sibling, but that's all this optimization is allowed to cost. Sibling cache buddies are cross-wired to prevent bouncing. 4 socket 40 core + SMT Westmere box, single 30 sec tbench runs, higher is better: clients 1 2 4 8 16 32 64 128 .......................................................................... pre 30 41 118 645 3769 6214 12233 14312 post 299 603 1211 2418 4697 6847 11606 14557 A nice increase in performance. Signed-off-by: Mike Galbraith Signed-off-by: Peter Zijlstra Cc: Linus Torvalds Cc: Andrew Morton Cc: Thomas Gleixner Link: http://lkml.kernel.org/r/1339471112.7352.32.camel@marge.simpson.net Signed-off-by: Ingo Molnar --- include/linux/sched.h | 1 + kernel/sched/core.c | 39 ++++++++++++++++++++++++++++++++++++++- kernel/sched/fair.c | 28 +++++++--------------------- 3 files changed, 46 insertions(+), 22 deletions(-) diff --git a/include/linux/sched.h b/include/linux/sched.h index 4a1f493e0fef..bc9952991710 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -949,6 +949,7 @@ struct sched_domain { unsigned int smt_gain; int flags; /* See SD_* */ int level; + int idle_buddy; /* cpu assigned to select_idle_sibling() */ /* Runtime fields. */ unsigned long last_balance; /* init to jiffies. units in jiffies */ diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 4b4a63d34396..536b213f0ce5 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -6024,6 +6024,11 @@ static void destroy_sched_domains(struct sched_domain *sd, int cpu) * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this * allows us to avoid some pointer chasing select_idle_sibling(). * + * Iterate domains and sched_groups downward, assigning CPUs to be + * select_idle_sibling() hw buddy. Cross-wiring hw makes bouncing + * due to random perturbation self canceling, ie sw buddies pull + * their counterpart to their CPU's hw counterpart. + * * Also keep a unique ID per domain (we use the first cpu number in * the cpumask of the domain), this allows us to quickly tell if * two cpus are in the same cache domain, see cpus_share_cache(). @@ -6037,8 +6042,40 @@ static void update_top_cache_domain(int cpu) int id = cpu; sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); - if (sd) + if (sd) { + struct sched_domain *tmp = sd; + struct sched_group *sg, *prev; + bool right; + + /* + * Traverse to first CPU in group, and count hops + * to cpu from there, switching direction on each + * hop, never ever pointing the last CPU rightward. + */ + do { + id = cpumask_first(sched_domain_span(tmp)); + prev = sg = tmp->groups; + right = 1; + + while (cpumask_first(sched_group_cpus(sg)) != id) + sg = sg->next; + + while (!cpumask_test_cpu(cpu, sched_group_cpus(sg))) { + prev = sg; + sg = sg->next; + right = !right; + } + + /* A CPU went down, never point back to domain start. */ + if (right && cpumask_first(sched_group_cpus(sg->next)) == id) + right = false; + + sg = right ? sg->next : prev; + tmp->idle_buddy = cpumask_first(sched_group_cpus(sg)); + } while ((tmp = tmp->child)); + id = cpumask_first(sched_domain_span(sd)); + } rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); per_cpu(sd_llc_id, cpu) = id; diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index c099cc6eebe3..dd00aaf44fda 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -2637,8 +2637,6 @@ static int select_idle_sibling(struct task_struct *p, int target) int cpu = smp_processor_id(); int prev_cpu = task_cpu(p); struct sched_domain *sd; - struct sched_group *sg; - int i; /* * If the task is going to be woken-up on this cpu and if it is @@ -2655,29 +2653,17 @@ static int select_idle_sibling(struct task_struct *p, int target) return prev_cpu; /* - * Otherwise, iterate the domains and find an elegible idle cpu. + * Otherwise, check assigned siblings to find an elegible idle cpu. */ sd = rcu_dereference(per_cpu(sd_llc, target)); + for_each_lower_domain(sd) { - sg = sd->groups; - do { - if (!cpumask_intersects(sched_group_cpus(sg), - tsk_cpus_allowed(p))) - goto next; - - for_each_cpu(i, sched_group_cpus(sg)) { - if (!idle_cpu(i)) - goto next; - } - - target = cpumask_first_and(sched_group_cpus(sg), - tsk_cpus_allowed(p)); - goto done; -next: - sg = sg->next; - } while (sg != sd->groups); + if (!cpumask_test_cpu(sd->idle_buddy, tsk_cpus_allowed(p))) + continue; + if (idle_cpu(sd->idle_buddy)) + return sd->idle_buddy; } -done: + return target; } From 85c1e7dae165acd004429f81fe52bfbf55b57a98 Mon Sep 17 00:00:00 2001 From: Prashanth Nageshappa Date: Tue, 19 Jun 2012 17:47:34 +0530 Subject: [PATCH 07/10] sched: Reorder 'struct lb_env' members to reduce its size Members of 'struct lb_env' are not in appropriate order to reuse compiler added padding on 64bit architectures. In this patch we reorder those struct members and help reduce the size of the structure from 96 bytes to 80 bytes on 64 bit architectures. Suggested-by: Srivatsa Vaddagiri Signed-off-by: Prashanth Nageshappa Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/r/4FE06DDE.7000403@linux.vnet.ibm.com Signed-off-by: Ingo Molnar --- kernel/sched/fair.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index dd00aaf44fda..9361669d4242 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -3058,8 +3058,8 @@ static unsigned long __read_mostly max_load_balance_interval = HZ/10; struct lb_env { struct sched_domain *sd; - int src_cpu; struct rq *src_rq; + int src_cpu; int dst_cpu; struct rq *dst_rq; From bbf18b19495942cc730e8ff11fc3ffadf20cbfe1 Mon Sep 17 00:00:00 2001 From: Prashanth Nageshappa Date: Tue, 19 Jun 2012 17:52:07 +0530 Subject: [PATCH 08/10] sched: Reset loop counters if all tasks are pinned and we need to redo load balance While load balancing, if all tasks on the source runqueue are pinned, we retry after excluding the corresponding source cpu. However, loop counters env.loop and env.loop_break are not reset before retrying, which can lead to failure in moving the tasks. In this patch we reset env.loop and env.loop_break to their inital values before we retry. Signed-off-by: Prashanth Nageshappa Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/r/4FE06EEF.2090709@linux.vnet.ibm.com Signed-off-by: Ingo Molnar --- kernel/sched/fair.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 9361669d4242..f9f9aa0edf3c 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -4288,8 +4288,11 @@ static int load_balance(int this_cpu, struct rq *this_rq, /* All tasks on this runqueue were pinned by CPU affinity */ if (unlikely(env.flags & LBF_ALL_PINNED)) { cpumask_clear_cpu(cpu_of(busiest), cpus); - if (!cpumask_empty(cpus)) + if (!cpumask_empty(cpus)) { + env.loop = 0; + env.loop_break = sched_nr_migrate_break; goto redo; + } goto out_balanced; } } From 88b8dac0a14c511ff41486b83a8c3d688936eec0 Mon Sep 17 00:00:00 2001 From: Srivatsa Vaddagiri Date: Tue, 19 Jun 2012 17:43:15 +0530 Subject: [PATCH 09/10] sched: Improve balance_cpu() to consider other cpus in its group as target of (pinned) task Current load balance scheme requires only one cpu in a sched_group (balance_cpu) to look at other peer sched_groups for imbalance and pull tasks towards itself from a busy cpu. Tasks thus pulled by balance_cpu could later get picked up by cpus that are in the same sched_group as that of balance_cpu. This scheme however fails to pull tasks that are not allowed to run on balance_cpu (but are allowed to run on other cpus in its sched_group). That can affect fairness and in some worst case scenarios cause starvation. Consider a two core (2 threads/core) system running tasks as below: Core0 Core1 / \ / \ C0 C1 C2 C3 | | | | v v v v F0 T1 F1 [idle] T2 F0 = SCHED_FIFO task (pinned to C0) F1 = SCHED_FIFO task (pinned to C2) T1 = SCHED_OTHER task (pinned to C1) T2 = SCHED_OTHER task (pinned to C1 and C2) F1 could become a cpu hog, which will starve T2 unless C1 pulls it. Between C0 and C1 however, C0 is required to look for imbalance between cores, which will fail to pull T2 towards Core0. T2 will starve eternally in this case. The same scenario can arise in presence of non-rt tasks as well (say we replace F1 with high irq load). We tackle this problem by having balance_cpu move pinned tasks to one of its sibling cpus (where they can run). We first check if load balance goal can be met by ignoring pinned tasks, failing which we retry move_tasks() with a new env->dst_cpu. This patch modifies load balance semantics on who can move load towards a given cpu in a given sched_domain. Before this patch, a given_cpu or a ilb_cpu acting on behalf of an idle given_cpu is responsible for moving load to given_cpu. With this patch applied, balance_cpu can in addition decide on moving some load to a given_cpu. There is a remote possibility that excess load could get moved as a result of this (balance_cpu and given_cpu/ilb_cpu deciding *independently* and at *same* time to move some load to a given_cpu). However we should see less of such conflicting decisions in practice and moreover subsequent load balance cycles should correct the excess load moved to given_cpu. Signed-off-by: Srivatsa Vaddagiri Signed-off-by: Prashanth Nageshappa Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/r/4FE06CDB.2060605@linux.vnet.ibm.com [ minor edits ] Signed-off-by: Ingo Molnar --- kernel/sched/fair.c | 78 ++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 74 insertions(+), 4 deletions(-) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index f9f9aa0edf3c..22321db64952 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -3054,6 +3054,7 @@ static unsigned long __read_mostly max_load_balance_interval = HZ/10; #define LBF_ALL_PINNED 0x01 #define LBF_NEED_BREAK 0x02 +#define LBF_SOME_PINNED 0x04 struct lb_env { struct sched_domain *sd; @@ -3064,6 +3065,8 @@ struct lb_env { int dst_cpu; struct rq *dst_rq; + struct cpumask *dst_grpmask; + int new_dst_cpu; enum cpu_idle_type idle; long imbalance; unsigned int flags; @@ -3131,9 +3134,31 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env) * 3) are cache-hot on their current CPU. */ if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { + int new_dst_cpu; + schedstat_inc(p, se.statistics.nr_failed_migrations_affine); + + /* + * Remember if this task can be migrated to any other cpu in + * our sched_group. We may want to revisit it if we couldn't + * meet load balance goals by pulling other tasks on src_cpu. + * + * Also avoid computing new_dst_cpu if we have already computed + * one in current iteration. + */ + if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED)) + return 0; + + new_dst_cpu = cpumask_first_and(env->dst_grpmask, + tsk_cpus_allowed(p)); + if (new_dst_cpu < nr_cpu_ids) { + env->flags |= LBF_SOME_PINNED; + env->new_dst_cpu = new_dst_cpu; + } return 0; } + + /* Record that we found atleast one task that could run on dst_cpu */ env->flags &= ~LBF_ALL_PINNED; if (task_running(env->src_rq, p)) { @@ -4213,7 +4238,8 @@ static int load_balance(int this_cpu, struct rq *this_rq, struct sched_domain *sd, enum cpu_idle_type idle, int *balance) { - int ld_moved, active_balance = 0; + int ld_moved, cur_ld_moved, active_balance = 0; + int lb_iterations, max_lb_iterations; struct sched_group *group; struct rq *busiest; unsigned long flags; @@ -4223,11 +4249,13 @@ static int load_balance(int this_cpu, struct rq *this_rq, .sd = sd, .dst_cpu = this_cpu, .dst_rq = this_rq, + .dst_grpmask = sched_group_cpus(sd->groups), .idle = idle, .loop_break = sched_nr_migrate_break, }; cpumask_copy(cpus, cpu_active_mask); + max_lb_iterations = cpumask_weight(env.dst_grpmask); schedstat_inc(sd, lb_count[idle]); @@ -4253,6 +4281,7 @@ static int load_balance(int this_cpu, struct rq *this_rq, schedstat_add(sd, lb_imbalance[idle], env.imbalance); ld_moved = 0; + lb_iterations = 1; if (busiest->nr_running > 1) { /* * Attempt to move tasks. If find_busiest_group has found @@ -4270,7 +4299,13 @@ static int load_balance(int this_cpu, struct rq *this_rq, double_rq_lock(this_rq, busiest); if (!env.loop) update_h_load(env.src_cpu); - ld_moved += move_tasks(&env); + + /* + * cur_ld_moved - load moved in current iteration + * ld_moved - cumulative load moved across iterations + */ + cur_ld_moved = move_tasks(&env); + ld_moved += cur_ld_moved; double_rq_unlock(this_rq, busiest); local_irq_restore(flags); @@ -4282,8 +4317,43 @@ static int load_balance(int this_cpu, struct rq *this_rq, /* * some other cpu did the load balance for us. */ - if (ld_moved && this_cpu != smp_processor_id()) - resched_cpu(this_cpu); + if (cur_ld_moved && env.dst_cpu != smp_processor_id()) + resched_cpu(env.dst_cpu); + + /* + * Revisit (affine) tasks on src_cpu that couldn't be moved to + * us and move them to an alternate dst_cpu in our sched_group + * where they can run. The upper limit on how many times we + * iterate on same src_cpu is dependent on number of cpus in our + * sched_group. + * + * This changes load balance semantics a bit on who can move + * load to a given_cpu. In addition to the given_cpu itself + * (or a ilb_cpu acting on its behalf where given_cpu is + * nohz-idle), we now have balance_cpu in a position to move + * load to given_cpu. In rare situations, this may cause + * conflicts (balance_cpu and given_cpu/ilb_cpu deciding + * _independently_ and at _same_ time to move some load to + * given_cpu) causing exceess load to be moved to given_cpu. + * This however should not happen so much in practice and + * moreover subsequent load balance cycles should correct the + * excess load moved. + */ + if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 && + lb_iterations++ < max_lb_iterations) { + + this_rq = cpu_rq(env.new_dst_cpu); + env.dst_rq = this_rq; + env.dst_cpu = env.new_dst_cpu; + env.flags &= ~LBF_SOME_PINNED; + env.loop = 0; + env.loop_break = sched_nr_migrate_break; + /* + * Go back to "more_balance" rather than "redo" since we + * need to continue with same src_cpu. + */ + goto more_balance; + } /* All tasks on this runqueue were pinned by CPU affinity */ if (unlikely(env.flags & LBF_ALL_PINNED)) { From 8323f26ce3425460769605a6aece7a174edaa7d1 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Fri, 22 Jun 2012 13:36:05 +0200 Subject: [PATCH 10/10] sched: Fix race in task_group() Stefan reported a crash on a kernel before a3e5d1091c1 ("sched: Don't call task_group() too many times in set_task_rq()"), he found the reason to be that the multiple task_group() invocations in set_task_rq() returned different values. Looking at all that I found a lack of serialization and plain wrong comments. The below tries to fix it using an extra pointer which is updated under the appropriate scheduler locks. Its not pretty, but I can't really see another way given how all the cgroup stuff works. Reported-and-tested-by: Stefan Bader Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/r/1340364965.18025.71.camel@twins Signed-off-by: Ingo Molnar --- include/linux/init_task.h | 12 +++++++++++- include/linux/sched.h | 5 ++++- kernel/sched/core.c | 9 ++++++++- kernel/sched/sched.h | 23 ++++++++++------------- 4 files changed, 33 insertions(+), 16 deletions(-) diff --git a/include/linux/init_task.h b/include/linux/init_task.h index 9e65eff6af3b..b806b821e735 100644 --- a/include/linux/init_task.h +++ b/include/linux/init_task.h @@ -123,8 +123,17 @@ extern struct group_info init_groups; extern struct cred init_cred; +extern struct task_group root_task_group; + +#ifdef CONFIG_CGROUP_SCHED +# define INIT_CGROUP_SCHED(tsk) \ + .sched_task_group = &root_task_group, +#else +# define INIT_CGROUP_SCHED(tsk) +#endif + #ifdef CONFIG_PERF_EVENTS -# define INIT_PERF_EVENTS(tsk) \ +# define INIT_PERF_EVENTS(tsk) \ .perf_event_mutex = \ __MUTEX_INITIALIZER(tsk.perf_event_mutex), \ .perf_event_list = LIST_HEAD_INIT(tsk.perf_event_list), @@ -161,6 +170,7 @@ extern struct cred init_cred; }, \ .tasks = LIST_HEAD_INIT(tsk.tasks), \ INIT_PUSHABLE_TASKS(tsk) \ + INIT_CGROUP_SCHED(tsk) \ .ptraced = LIST_HEAD_INIT(tsk.ptraced), \ .ptrace_entry = LIST_HEAD_INIT(tsk.ptrace_entry), \ .real_parent = &tsk, \ diff --git a/include/linux/sched.h b/include/linux/sched.h index bc9952991710..fd9436a3a545 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -1245,6 +1245,9 @@ struct task_struct { const struct sched_class *sched_class; struct sched_entity se; struct sched_rt_entity rt; +#ifdef CONFIG_CGROUP_SCHED + struct task_group *sched_task_group; +#endif #ifdef CONFIG_PREEMPT_NOTIFIERS /* list of struct preempt_notifier: */ @@ -2724,7 +2727,7 @@ extern int sched_group_set_rt_period(struct task_group *tg, extern long sched_group_rt_period(struct task_group *tg); extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk); #endif -#endif +#endif /* CONFIG_CGROUP_SCHED */ extern int task_can_switch_user(struct user_struct *up, struct task_struct *tsk); diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 536b213f0ce5..5d011ef4c0df 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -1096,7 +1096,7 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu) * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. * * sched_move_task() holds both and thus holding either pins the cgroup, - * see set_task_rq(). + * see task_group(). * * Furthermore, all task_rq users should acquire both locks, see * task_rq_lock(). @@ -7658,6 +7658,7 @@ void sched_destroy_group(struct task_group *tg) */ void sched_move_task(struct task_struct *tsk) { + struct task_group *tg; int on_rq, running; unsigned long flags; struct rq *rq; @@ -7672,6 +7673,12 @@ void sched_move_task(struct task_struct *tsk) if (unlikely(running)) tsk->sched_class->put_prev_task(rq, tsk); + tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, + lockdep_is_held(&tsk->sighand->siglock)), + struct task_group, css); + tg = autogroup_task_group(tsk, tg); + tsk->sched_task_group = tg; + #ifdef CONFIG_FAIR_GROUP_SCHED if (tsk->sched_class->task_move_group) tsk->sched_class->task_move_group(tsk, on_rq); diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h index 55844f24435a..c35a1a7dd4d6 100644 --- a/kernel/sched/sched.h +++ b/kernel/sched/sched.h @@ -538,22 +538,19 @@ extern int group_balance_cpu(struct sched_group *sg); /* * Return the group to which this tasks belongs. * - * We use task_subsys_state_check() and extend the RCU verification with - * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each - * task it moves into the cgroup. Therefore by holding either of those locks, - * we pin the task to the current cgroup. + * We cannot use task_subsys_state() and friends because the cgroup + * subsystem changes that value before the cgroup_subsys::attach() method + * is called, therefore we cannot pin it and might observe the wrong value. + * + * The same is true for autogroup's p->signal->autogroup->tg, the autogroup + * core changes this before calling sched_move_task(). + * + * Instead we use a 'copy' which is updated from sched_move_task() while + * holding both task_struct::pi_lock and rq::lock. */ static inline struct task_group *task_group(struct task_struct *p) { - struct task_group *tg; - struct cgroup_subsys_state *css; - - css = task_subsys_state_check(p, cpu_cgroup_subsys_id, - lockdep_is_held(&p->pi_lock) || - lockdep_is_held(&task_rq(p)->lock)); - tg = container_of(css, struct task_group, css); - - return autogroup_task_group(p, tg); + return p->sched_task_group; } /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */