From 10717dcde10d09f9fcee53a12a4236af1a82b484 Mon Sep 17 00:00:00 2001 From: Alex Shi Date: Wed, 6 Jun 2012 14:52:51 +0800 Subject: sched/numa: Load balance between remote nodes Commit cb83b629b ("sched/numa: Rewrite the CONFIG_NUMA sched domain support") removed the NODE sched domain and started checking if the node distance in SLIT table is farther than REMOTE_DISTANCE, if so, it will lose the load balance chance at exec/fork/wake_affine points. But actually, even the node distance is farther than REMOTE_DISTANCE. Modern CPUs also has QPI like connections, which ensures that memory access is not too slow between nodes. So the above change in behavior on NUMA machine causes a performance regression on various benchmarks: hackbench, tbench, netperf, oltp, etc. This patch will recover the scheduler behavior to old mode on all my Intel platforms: NHM EP/EX, WSM EP, SNB EP/EP4S, and thus fixes the perfromance regressions. (all of them just have 2 kinds distance, 10, 21) Signed-off-by: Alex Shi Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/r/1338965571-9812-1-git-send-email-alex.shi@intel.com Signed-off-by: Ingo Molnar --- kernel/sched/core.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'kernel') diff --git a/kernel/sched/core.c b/kernel/sched/core.c index c46958e2612..6546083af3e 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -6321,7 +6321,7 @@ static int sched_domains_curr_level; static inline int sd_local_flags(int level) { - if (sched_domains_numa_distance[level] > REMOTE_DISTANCE) + if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) return 0; return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; -- cgit v1.2.3 From 7f1b43936f0ecad14770634c021cf4a929aec74d Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Thu, 17 May 2012 21:19:46 +0200 Subject: sched/rt: Fix lockdep annotation within find_lock_lowest_rq() Roland Dreier reported spurious, hard to trigger lockdep warnings within the scheduler - without any real lockup. This bit gives us the right clue: > [89945.640512] [] double_lock_balance+0x5a/0x90 > [89945.640568] [] push_rt_task+0xc6/0x290 if you look at that code you'll find the double_lock_balance() in question is the one in find_lock_lowest_rq() [yay for inlining]. Now find_lock_lowest_rq() has a bug.. it fails to use double_unlock_balance() in one exit path, if this results in a retry in push_rt_task() we'll call double_lock_balance() again, at which point we'll run into said lockdep confusion. Reported-by: Roland Dreier Acked-by: Steven Rostedt Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/r/1337282386.4281.77.camel@twins Signed-off-by: Ingo Molnar --- kernel/sched/rt.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'kernel') diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c index 2a4e8dffbd6..573e1ca0110 100644 --- a/kernel/sched/rt.c +++ b/kernel/sched/rt.c @@ -1562,7 +1562,7 @@ static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq) task_running(rq, task) || !task->on_rq)) { - raw_spin_unlock(&lowest_rq->lock); + double_unlock_balance(rq, lowest_rq); lowest_rq = NULL; break; } -- cgit v1.2.3 From c1174876874dcf8986806e4dad3d7d07af20b439 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Thu, 31 May 2012 14:47:33 +0200 Subject: sched: Fix domain iteration Weird topologies can lead to asymmetric domain setups. This needs further consideration since these setups are typically non-minimal too. For now, make it work by adding an extra mask selecting which CPUs are allowed to iterate up. The topology that triggered it is the one from David Rientjes: 10 20 20 30 20 10 20 20 20 20 10 20 30 20 20 10 resulting in boxes that wouldn't even boot. Reported-by: David Rientjes Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/n/tip-3p86l9cuaqnxz7uxsojmz5rm@git.kernel.org Signed-off-by: Ingo Molnar --- kernel/sched/core.c | 64 +++++++++++++++++++++++++++++++++++++++++++++------- kernel/sched/fair.c | 5 ++-- kernel/sched/sched.h | 2 ++ 3 files changed, 61 insertions(+), 10 deletions(-) (limited to 'kernel') diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 6546083af3e..781acb91a50 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -5994,6 +5994,44 @@ struct sched_domain_topology_level { struct sd_data data; }; +/* + * Build an iteration mask that can exclude certain CPUs from the upwards + * domain traversal. + * + * Asymmetric node setups can result in situations where the domain tree is of + * unequal depth, make sure to skip domains that already cover the entire + * range. + * + * In that case build_sched_domains() will have terminated the iteration early + * and our sibling sd spans will be empty. Domains should always include the + * cpu they're built on, so check that. + * + */ +static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) +{ + const struct cpumask *span = sched_domain_span(sd); + struct sd_data *sdd = sd->private; + struct sched_domain *sibling; + int i; + + for_each_cpu(i, span) { + sibling = *per_cpu_ptr(sdd->sd, i); + if (!cpumask_test_cpu(i, sched_domain_span(sibling))) + continue; + + cpumask_set_cpu(i, sched_group_mask(sg)); + } +} + +/* + * Return the canonical balance cpu for this group, this is the first cpu + * of this group that's also in the iteration mask. + */ +int group_balance_cpu(struct sched_group *sg) +{ + return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); +} + static int build_overlap_sched_groups(struct sched_domain *sd, int cpu) { @@ -6012,6 +6050,12 @@ build_overlap_sched_groups(struct sched_domain *sd, int cpu) if (cpumask_test_cpu(i, covered)) continue; + child = *per_cpu_ptr(sdd->sd, i); + + /* See the comment near build_group_mask(). */ + if (!cpumask_test_cpu(i, sched_domain_span(child))) + continue; + sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), GFP_KERNEL, cpu_to_node(cpu)); @@ -6019,8 +6063,6 @@ build_overlap_sched_groups(struct sched_domain *sd, int cpu) goto fail; sg_span = sched_group_cpus(sg); - - child = *per_cpu_ptr(sdd->sd, i); if (child->child) { child = child->child; cpumask_copy(sg_span, sched_domain_span(child)); @@ -6030,13 +6072,18 @@ build_overlap_sched_groups(struct sched_domain *sd, int cpu) cpumask_or(covered, covered, sg_span); sg->sgp = *per_cpu_ptr(sdd->sgp, i); - atomic_inc(&sg->sgp->ref); + if (atomic_inc_return(&sg->sgp->ref) == 1) + build_group_mask(sd, sg); + + /* + * Make sure the first group of this domain contains the + * canonical balance cpu. Otherwise the sched_domain iteration + * breaks. See update_sg_lb_stats(). + */ if ((!groups && cpumask_test_cpu(cpu, sg_span)) || - cpumask_first(sg_span) == cpu) { - WARN_ON_ONCE(!cpumask_test_cpu(cpu, sg_span)); + group_balance_cpu(sg) == cpu) groups = sg; - } if (!first) first = sg; @@ -6109,6 +6156,7 @@ build_sched_groups(struct sched_domain *sd, int cpu) cpumask_clear(sched_group_cpus(sg)); sg->sgp->power = 0; + cpumask_setall(sched_group_mask(sg)); for_each_cpu(j, span) { if (get_group(j, sdd, NULL) != group) @@ -6150,7 +6198,7 @@ static void init_sched_groups_power(int cpu, struct sched_domain *sd) sg = sg->next; } while (sg != sd->groups); - if (cpu != group_first_cpu(sg)) + if (cpu != group_balance_cpu(sg)) return; update_group_power(sd, cpu); @@ -6525,7 +6573,7 @@ static int __sdt_alloc(const struct cpumask *cpu_map) *per_cpu_ptr(sdd->sg, j) = sg; - sgp = kzalloc_node(sizeof(struct sched_group_power), + sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), GFP_KERNEL, cpu_to_node(j)); if (!sgp) return -ENOMEM; diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index b2a2d236f27..54cbaa4e7b3 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -3652,7 +3652,7 @@ static inline void update_sg_lb_stats(struct lb_env *env, int i; if (local_group) - balance_cpu = group_first_cpu(group); + balance_cpu = group_balance_cpu(group); /* Tally up the load of all CPUs in the group */ max_cpu_load = 0; @@ -3667,7 +3667,8 @@ static inline void update_sg_lb_stats(struct lb_env *env, /* Bias balancing toward cpus of our domain */ if (local_group) { - if (idle_cpu(i) && !first_idle_cpu) { + if (idle_cpu(i) && !first_idle_cpu && + cpumask_test_cpu(i, sched_group_mask(group))) { first_idle_cpu = 1; balance_cpu = i; } diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h index ba9dccfd24c..6d52cea7f33 100644 --- a/kernel/sched/sched.h +++ b/kernel/sched/sched.h @@ -526,6 +526,8 @@ static inline struct sched_domain *highest_flag_domain(int cpu, int flag) DECLARE_PER_CPU(struct sched_domain *, sd_llc); DECLARE_PER_CPU(int, sd_llc_id); +extern int group_balance_cpu(struct sched_group *sg); + #endif /* CONFIG_SMP */ #include "stats.h" -- cgit v1.2.3 From c3decf0dfbc95736b7c0ab68fa4e5854c4734da9 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Thu, 31 May 2012 12:05:32 +0200 Subject: sched: Always initialize cpu-power Often when we run into mis-shapen topologies the balance iteration fails to update the cpu power properly and we'll end up in /0 traps. Always initialize the cpu-power to a semi-sane value so that we can at least boot the machine, even if the load-balancer might not function correctly. Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/n/tip-3lbhyj25sr169ha7z3qht5na@git.kernel.org Signed-off-by: Ingo Molnar --- kernel/sched/core.c | 13 ++++++++++++- kernel/sched/fair.c | 2 +- 2 files changed, 13 insertions(+), 2 deletions(-) (limited to 'kernel') diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 781acb91a50..725ee7c1c8c 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -5604,7 +5604,12 @@ static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, break; } - if (!group->sgp->power) { + /* + * Even though we initialize ->power to something semi-sane, + * we leave power_orig unset. This allows us to detect if + * domain iteration is still funny without causing /0 traps. + */ + if (!group->sgp->power_orig) { printk(KERN_CONT "\n"); printk(KERN_ERR "ERROR: domain->cpu_power not " "set\n"); @@ -6075,6 +6080,12 @@ build_overlap_sched_groups(struct sched_domain *sd, int cpu) if (atomic_inc_return(&sg->sgp->ref) == 1) build_group_mask(sd, sg); + /* + * Initialize sgp->power such that even if we mess up the + * domains and no possible iteration will get us here, we won't + * die on a /0 trap. + */ + sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); /* * Make sure the first group of this domain contains the diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 54cbaa4e7b3..c9fd6d673d0 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -3602,7 +3602,7 @@ void update_group_power(struct sched_domain *sd, int cpu) } while (group != child->groups); } - sdg->sgp->power = power; + sdg->sgp->power_orig = sdg->sgp->power = power; } /* -- cgit v1.2.3 From d039ac60800fe8ed8522ec3b9ca796aaf748c18b Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Thu, 31 May 2012 21:20:16 +0200 Subject: sched: Validate assumptions in sched_init_numa() Add some code to validate assumptions we're making and output warnings if they are not. If this trigger we want to know about it. Signed-off-by: Peter Zijlstra Cc: Alex Shi Link: http://lkml.kernel.org/n/tip-6uc3wk5s9udxtdl9cnku0vtt@git.kernel.org Signed-off-by: Ingo Molnar --- kernel/sched/core.c | 99 +++++++++++++++++++++++++++++++++++++++++++---------- 1 file changed, 80 insertions(+), 19 deletions(-) (limited to 'kernel') diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 725ee7c1c8c..2bdd1761643 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -5556,15 +5556,20 @@ static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ #ifdef CONFIG_SCHED_DEBUG -static __read_mostly int sched_domain_debug_enabled; +static __read_mostly int sched_debug_enabled; -static int __init sched_domain_debug_setup(char *str) +static int __init sched_debug_setup(char *str) { - sched_domain_debug_enabled = 1; + sched_debug_enabled = 1; return 0; } -early_param("sched_debug", sched_domain_debug_setup); +early_param("sched_debug", sched_debug_setup); + +static inline bool sched_debug(void) +{ + return sched_debug_enabled; +} static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, struct cpumask *groupmask) @@ -5657,7 +5662,7 @@ static void sched_domain_debug(struct sched_domain *sd, int cpu) { int level = 0; - if (!sched_domain_debug_enabled) + if (!sched_debug_enabled) return; if (!sd) { @@ -5678,6 +5683,10 @@ static void sched_domain_debug(struct sched_domain *sd, int cpu) } #else /* !CONFIG_SCHED_DEBUG */ # define sched_domain_debug(sd, cpu) do { } while (0) +static inline bool sched_debug(void) +{ + return false; +} #endif /* CONFIG_SCHED_DEBUG */ static int sd_degenerate(struct sched_domain *sd) @@ -6373,7 +6382,6 @@ static struct sched_domain_topology_level *sched_domain_topology = default_topol #ifdef CONFIG_NUMA static int sched_domains_numa_levels; -static int sched_domains_numa_scale; static int *sched_domains_numa_distance; static struct cpumask ***sched_domains_numa_masks; static int sched_domains_curr_level; @@ -6438,6 +6446,42 @@ static const struct cpumask *sd_numa_mask(int cpu) return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; } +static void sched_numa_warn(const char *str) +{ + static int done = false; + int i,j; + + if (done) + return; + + done = true; + + printk(KERN_WARNING "ERROR: %s\n\n", str); + + for (i = 0; i < nr_node_ids; i++) { + printk(KERN_WARNING " "); + for (j = 0; j < nr_node_ids; j++) + printk(KERN_CONT "%02d ", node_distance(i,j)); + printk(KERN_CONT "\n"); + } + printk(KERN_WARNING "\n"); +} + +static bool find_numa_distance(int distance) +{ + int i; + + if (distance == node_distance(0, 0)) + return true; + + for (i = 0; i < sched_domains_numa_levels; i++) { + if (sched_domains_numa_distance[i] == distance) + return true; + } + + return false; +} + static void sched_init_numa(void) { int next_distance, curr_distance = node_distance(0, 0); @@ -6445,7 +6489,6 @@ static void sched_init_numa(void) int level = 0; int i, j, k; - sched_domains_numa_scale = curr_distance; sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); if (!sched_domains_numa_distance) return; @@ -6456,23 +6499,41 @@ static void sched_init_numa(void) * * Assumes node_distance(0,j) includes all distances in * node_distance(i,j) in order to avoid cubic time. - * - * XXX: could be optimized to O(n log n) by using sort() */ next_distance = curr_distance; for (i = 0; i < nr_node_ids; i++) { for (j = 0; j < nr_node_ids; j++) { - int distance = node_distance(0, j); - if (distance > curr_distance && - (distance < next_distance || - next_distance == curr_distance)) - next_distance = distance; + for (k = 0; k < nr_node_ids; k++) { + int distance = node_distance(i, k); + + if (distance > curr_distance && + (distance < next_distance || + next_distance == curr_distance)) + next_distance = distance; + + /* + * While not a strong assumption it would be nice to know + * about cases where if node A is connected to B, B is not + * equally connected to A. + */ + if (sched_debug() && node_distance(k, i) != distance) + sched_numa_warn("Node-distance not symmetric"); + + if (sched_debug() && i && !find_numa_distance(distance)) + sched_numa_warn("Node-0 not representative"); + } + if (next_distance != curr_distance) { + sched_domains_numa_distance[level++] = next_distance; + sched_domains_numa_levels = level; + curr_distance = next_distance; + } else break; } - if (next_distance != curr_distance) { - sched_domains_numa_distance[level++] = next_distance; - sched_domains_numa_levels = level; - curr_distance = next_distance; - } else break; + + /* + * In case of sched_debug() we verify the above assumption. + */ + if (!sched_debug()) + break; } /* * 'level' contains the number of unique distances, excluding the -- cgit v1.2.3 From a841f8cef4bb124f0f5563314d0beaf2e1249d72 Mon Sep 17 00:00:00 2001 From: Dimitri Sivanich Date: Tue, 5 Jun 2012 13:44:36 -0500 Subject: sched: Fix the relax_domain_level boot parameter It does not get processed because sched_domain_level_max is 0 at the time that setup_relax_domain_level() is run. Simply accept the value as it is, as we don't know the value of sched_domain_level_max until sched domain construction is completed. Fix sched_relax_domain_level in cpuset. The build_sched_domain() routine calls the set_domain_attribute() routine prior to setting the sd->level, however, the set_domain_attribute() routine relies on the sd->level to decide whether idle load balancing will be off/on. Signed-off-by: Dimitri Sivanich Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/r/20120605184436.GA15668@sgi.com Signed-off-by: Ingo Molnar --- kernel/sched/core.c | 9 +++------ 1 file changed, 3 insertions(+), 6 deletions(-) (limited to 'kernel') diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 2bdd1761643..d5594a4268d 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -6268,11 +6268,8 @@ int sched_domain_level_max; static int __init setup_relax_domain_level(char *str) { - unsigned long val; - - val = simple_strtoul(str, NULL, 0); - if (val < sched_domain_level_max) - default_relax_domain_level = val; + if (kstrtoint(str, 0, &default_relax_domain_level)) + pr_warn("Unable to set relax_domain_level\n"); return 1; } @@ -6698,7 +6695,6 @@ struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, if (!sd) return child; - set_domain_attribute(sd, attr); cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); if (child) { sd->level = child->level + 1; @@ -6706,6 +6702,7 @@ struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, child->parent = sd; } sd->child = child; + set_domain_attribute(sd, attr); return sd; } -- cgit v1.2.3