| 1 | /* Scheduler load balancing policy */ |
| 2 | #include <mem/numa.h> |
| 3 | #include <smp/domain.h> |
| 4 | |
| 5 | #include "internal.h" |
| 6 | |
| 7 | #define SCHEDULER_REMOTE_NODE_SCALE_NUMERATOR 1 |
| 8 | #define SCHEDULER_REMOTE_NODE_SCALE_DENOMINATOR 5 |
| 9 | #define IDLE_LONG_ENOUGH 10 /* if the other core is idle for 10ms */ |
| 10 | #define IDLE_MIN_MIGRATABLE 3 /* and we have 3 migratable */ |
| 11 | |
| 12 | static inline bool can_push_thread(size_t caller, struct thread *t) { |
| 13 | return scheduler_can_take_thread(core: caller, target: t) && t != thread_get_current(); |
| 14 | } |
| 15 | |
| 16 | /* fraction = remote_scale * (1 / (1 + dist)) |
| 17 | * |
| 18 | * using integer math: |
| 19 | * |
| 20 | * to_migrate = (count * remote_scale_num) / ((1 + dist) * remote_scale_den); |
| 21 | */ |
| 22 | |
| 23 | /* it's OK if this races, we are just counting threads */ |
| 24 | static size_t migratable_in_tree(size_t caller, struct rbt *rbt) { |
| 25 | struct rbt_node *rb; |
| 26 | size_t agg = 0; |
| 27 | rbt_for_each(rb, rbt) { |
| 28 | struct thread *t = thread_from_rq_rbt_node(rb); |
| 29 | if (can_push_thread(caller, t)) |
| 30 | agg++; |
| 31 | } |
| 32 | return agg; |
| 33 | } |
| 34 | |
| 35 | static size_t migratable_in_list(size_t caller, struct list_head *tq) { |
| 36 | struct list_head *ln; |
| 37 | size_t agg = 0; |
| 38 | list_for_each(ln, tq) { |
| 39 | struct thread *t = thread_from_rq_list_node(ln); |
| 40 | if (can_push_thread(caller, t)) |
| 41 | agg++; |
| 42 | } |
| 43 | return agg; |
| 44 | } |
| 45 | |
| 46 | void scheduler_count_migratable_threads(struct scheduler *caller, |
| 47 | struct scheduler *s, |
| 48 | size_t agg[THREAD_PRIO_CLASS_COUNT]) { |
| 49 | SPINLOCK_ASSERT_HELD(&caller->lock); |
| 50 | SPINLOCK_ASSERT_HELD(&s->lock); |
| 51 | /* count urgent, count realtime, count timesharing, count background */ |
| 52 | size_t c = caller->core_id; |
| 53 | agg[THREAD_PRIO_CLASS_URGENT] = migratable_in_list(caller: c, tq: &s->urgent_threads); |
| 54 | agg[THREAD_PRIO_CLASS_RT] = migratable_in_list(caller: c, tq: &s->rt_threads); |
| 55 | agg[THREAD_PRIO_CLASS_BACKGROUND] = migratable_in_list(caller: c, tq: &s->bg_threads); |
| 56 | agg[THREAD_PRIO_CLASS_TIMESHARE] = migratable_in_tree(caller: c, rbt: &s->completed_rbt); |
| 57 | agg[THREAD_PRIO_CLASS_TIMESHARE] += migratable_in_tree(caller: c, rbt: &s->thread_rbt); |
| 58 | } |
| 59 | |
| 60 | /* this is a fun trick... if there is no NUMA, the `associated_node` will be |
| 61 | * NULL, and this will always return true. if there is NUMA, then this will |
| 62 | * actually do the comparison, so we can use this function the same |
| 63 | * way in either setting! :) */ |
| 64 | static inline bool cores_in_same_numa_node(struct core *a, struct core *b) { |
| 65 | return a->domain->associated_node == b->domain->associated_node; |
| 66 | } |
| 67 | |
| 68 | static void move_ts_thread_raw(struct scheduler *dest, struct scheduler *source, |
| 69 | struct rbt *tree, struct thread *thread) { |
| 70 | rbt_delete(tree, z: &thread->rq_tree_node); |
| 71 | scheduler_decrement_thread_count(sched: source, t: thread); |
| 72 | |
| 73 | rbt_insert(tree: &dest->thread_rbt, new_node: &thread->rq_tree_node); |
| 74 | scheduler_increment_thread_count(sched: dest, t: thread); |
| 75 | thread_set_runqueue(t: thread, s: dest); |
| 76 | thread_post_migrate(t: thread, old_cpu: source->core_id, new_cpu: dest->core_id); |
| 77 | } |
| 78 | |
| 79 | static size_t migrate_from_tree(struct scheduler *to, |
| 80 | struct scheduler *from_sched, struct rbt *from, |
| 81 | size_t target) { |
| 82 | size_t migrated = 0; |
| 83 | struct rbt_node *rb; |
| 84 | |
| 85 | /* the idea is that we should ideally check "every other thread" to migrate |
| 86 | * them over. however, we will sometimes encounter threads that cannot be |
| 87 | * migrated. if we detect during iteration that there is a next thread and |
| 88 | * it cannot be migrated, we will opt to migrate the current thread instead, |
| 89 | * even if it slightly breaks our "every other thread" rule. */ |
| 90 | bool prev_migrated = false; |
| 91 | rbt_for_each(rb, from) { |
| 92 | if (migrated >= target) |
| 93 | break; |
| 94 | |
| 95 | struct thread *t = thread_from_rq_rbt_node(rb); |
| 96 | |
| 97 | /* we are on a thread we will give priority to migrating */ |
| 98 | if (!prev_migrated) { |
| 99 | if (can_push_thread(caller: to->core_id, t)) { |
| 100 | move_ts_thread_raw(dest: to, source: from_sched, tree: from, thread: t); |
| 101 | prev_migrated = true; |
| 102 | migrated++; |
| 103 | } |
| 104 | |
| 105 | } else { |
| 106 | prev_migrated = false; |
| 107 | } |
| 108 | } |
| 109 | |
| 110 | return migrated; |
| 111 | } |
| 112 | |
| 113 | /* migrate `num_threads` threads, and return how many threads we migrated. In |
| 114 | * the event that we are migrating from the timesharing threads, every OTHER |
| 115 | * thread will be migrated until `num_threads` is reached. we do this to |
| 116 | * prevent stealing either all the high prio threads, all the low prio threads, |
| 117 | * or all the middlers. */ |
| 118 | static size_t migrate_from_prio_class(struct scheduler *to, |
| 119 | struct scheduler *from, |
| 120 | enum thread_prio_class class, |
| 121 | size_t nthreads) { |
| 122 | if (!nthreads) |
| 123 | return 0; |
| 124 | |
| 125 | size_t migrated = 0; |
| 126 | if (class != THREAD_PRIO_CLASS_TIMESHARE) { |
| 127 | struct list_head *from_queue = |
| 128 | scheduler_get_this_thread_queue(sched: from, prio: class); |
| 129 | struct list_head *to_queue = scheduler_get_this_thread_queue(sched: to, prio: class); |
| 130 | |
| 131 | struct list_head *ln, *tmp; |
| 132 | list_for_each_safe(ln, tmp, from_queue) { |
| 133 | if (migrated >= nthreads) |
| 134 | break; |
| 135 | |
| 136 | struct thread *t = thread_from_rq_list_node(ln); |
| 137 | if (can_push_thread(caller: to->core_id, t)) { |
| 138 | |
| 139 | list_del_init(entry: ln); |
| 140 | scheduler_decrement_thread_count(sched: from, t); |
| 141 | |
| 142 | list_add_tail(new: ln, head: to_queue); |
| 143 | scheduler_increment_thread_count(sched: to, t); |
| 144 | |
| 145 | migrated++; |
| 146 | thread_set_runqueue(t, s: to); |
| 147 | thread_post_migrate(t, old_cpu: from->core_id, new_cpu: to->core_id); |
| 148 | } |
| 149 | } |
| 150 | } else { |
| 151 | /* migrating timesharing threads. first try to migrate threads that have |
| 152 | * not ran this period, skipping every other thread, and then try and |
| 153 | * migrate the completed threads */ |
| 154 | size_t took = migrate_from_tree(to, from_sched: from, from: &from->thread_rbt, target: nthreads); |
| 155 | migrated += took; |
| 156 | nthreads -= took; |
| 157 | |
| 158 | migrated += migrate_from_tree(to, from_sched: from, from: &from->completed_rbt, target: nthreads); |
| 159 | } |
| 160 | |
| 161 | return migrated; |
| 162 | } |
| 163 | |
| 164 | size_t scheduler_try_push_to_idle_core(struct scheduler *sched) { |
| 165 | if (!global.scheduler_domains[TOPOLOGY_LEVEL_MACHINE]) |
| 166 | return 0; |
| 167 | |
| 168 | int32_t other_cpu_id = scheduler_push_target(from: global.cores[sched->core_id]); |
| 169 | if (other_cpu_id < 0) |
| 170 | return 0; |
| 171 | |
| 172 | struct core *this_core = global.cores[sched->core_id]; |
| 173 | struct core *other_core = global.cores[other_cpu_id]; |
| 174 | |
| 175 | struct scheduler *other = global.schedulers[other_cpu_id]; |
| 176 | |
| 177 | /* we trylock to avoid spinning as this is called from the hotpath and can |
| 178 | * also have a circular dependency deadlock/starvation chance... */ |
| 179 | if (!spin_trylock_raw(lock: &other->lock)) |
| 180 | return 0; |
| 181 | |
| 182 | /* from the OTHER core's perspective, how many threads can we migrate? */ |
| 183 | size_t migratable[THREAD_PRIO_CLASS_COUNT]; |
| 184 | scheduler_count_migratable_threads(caller: other, s: sched, agg: migratable); |
| 185 | |
| 186 | size_t total_migratable = 0; |
| 187 | for (size_t i = 0; i < THREAD_PRIO_CLASS_COUNT; i++) |
| 188 | total_migratable += migratable[i]; |
| 189 | |
| 190 | size_t migrated = 0; |
| 191 | if (!total_migratable) |
| 192 | goto out; |
| 193 | |
| 194 | /* if we are in the same NUMA node, migrate half of our threads |
| 195 | * from each priority class so we are near identical */ |
| 196 | if (cores_in_same_numa_node(a: this_core, b: other_core)) { |
| 197 | for (size_t i = 0; i < THREAD_PRIO_CLASS_COUNT; i++) { |
| 198 | size_t to_migrate = migratable[i] / 2; |
| 199 | migrated += migrate_from_prio_class(to: other, from: sched, class: i, nthreads: to_migrate); |
| 200 | } |
| 201 | } else { |
| 202 | /* remote node */ |
| 203 | size_t numa_id = other_core->domain->id; |
| 204 | size_t dist = this_core->domain->associated_node->rel_dists[numa_id]; |
| 205 | |
| 206 | if (dist == 0) |
| 207 | dist = 1; /* shouldn't happen here */ |
| 208 | |
| 209 | /* TODO: tune */ |
| 210 | const size_t remote_scale_num = 1; /* numerator of scale, tuneable */ |
| 211 | const size_t remote_scale_den = 5; /* denominator of scale, tuneable */ |
| 212 | |
| 213 | size_t dist_factor = (1 + dist) * remote_scale_den; |
| 214 | |
| 215 | for (size_t i = 0; i < THREAD_PRIO_CLASS_COUNT; i++) { |
| 216 | size_t count = migratable[i]; |
| 217 | if (count == 0) |
| 218 | continue; |
| 219 | |
| 220 | size_t to_migrate = (count * remote_scale_num) / dist_factor; |
| 221 | |
| 222 | /* Enforce minimum if remote move is allowed */ |
| 223 | if (to_migrate == 0 && count > 0) |
| 224 | to_migrate = 1; |
| 225 | |
| 226 | time_t idle_entry = other->idle_thread_data.last_entry_ms; |
| 227 | time_t idle_for = time_get_ms() - idle_entry; |
| 228 | if (to_migrate == 0 && count >= IDLE_MIN_MIGRATABLE && |
| 229 | idle_for >= IDLE_LONG_ENOUGH) { |
| 230 | to_migrate = 1; |
| 231 | } |
| 232 | |
| 233 | migrated += migrate_from_prio_class(to: other, from: sched, class: i, nthreads: to_migrate); |
| 234 | } |
| 235 | } |
| 236 | |
| 237 | /* if the other core is in the same node as us, we push half of our threads |
| 238 | * over there. otherwise, we push (1 / distance * scale) of our threads */ |
| 239 | |
| 240 | out: |
| 241 | spin_unlock_raw(lock: &other->lock); |
| 242 | |
| 243 | if (migrated) |
| 244 | ipi_send(apic_id: other->core_id, IRQ_SCHEDULER); |
| 245 | |
| 246 | return migrated; |
| 247 | } |
| 248 | |