QTI EAS学习之find_energy_efficient_cpu
Energy Awareness Scheduler是由ARM和Linaro开发的新的linux kernel调度器。
原先CFS调度器是基于policy进行调度，并有不同的吞吐量。例如，有一个新的task创建，同时也有一个idle cpu时，CFS始终会把新的task放到这个idle cpu上运行。但是，这样对节省功耗来说，并不是一个最好的决定。而EAS就是为了解决这样的问题。在不影响性能的前提下，EAS会在调度时实现节省功耗。
从SDM845开始，QTI在EAS基础上进行了一些修改，以满足移动市场的需要。所以QTI在EAS基础上添加了一些feature，来获得更好的性能和功耗。
Energy model
在dts中，针对不同的cpu平台，已定义好不同的energy model。模型主要是由【频率，能量】的数组构成，对应了CPU和cluster不同的OOP（Operating Performance Point）；同时也提供了不同idle state的能量消耗：idle cost。
CPU0: cpu@0 { device_type = "cpu"; compatible = "arm,armv8"; reg = <0x0 0x0>; enable-method = "psci"; efficiency = <1024>; cache-size = <0x8000>; cpu-release-addr = <0x0 0x90000000>; qcom,lmh-dcvs = <&lmh_dcvs0>; #cooling-cells = <2>; next-level-cache = <&L2_0>; sched-energy-costs = <&CPU_COST_0 &CLUSTER_COST_0>; //小核都用CPU_COST_0 CLUSTER_COST_0 。。。。。。 CPU4: cpu@400 { device_type = "cpu"; compatible = "arm,armv8"; reg = <0x0 0x400>; enable-method = "psci"; efficiency = <1740>; cache-size = <0x20000>; cpu-release-addr = <0x0 0x90000000>; qcom,lmh-dcvs = <&lmh_dcvs1>; #cooling-cells = <2>; next-level-cache = <&L2_400>; sched-energy-costs = <&CPU_COST_1 &CLUSTER_COST_1>; //大核都用CPU_COST_1 CLUSTER_COST_1
。。。。。。。
对应的数组如下，
energy_costs: energy-costs { compatible = "sched-energy"; CPU_COST_0: core-cost0 { busy-cost-data = < 300000 31 422400 38 499200 42 576000 46 652800 51 748800 58 825600 64 902400 70 979200 76 1056000 83 1132800 90 1209600 97 1286400 105 1363200 114 1440000 124 1516800 136 1593600 152 1651200 167 /* speedbin 0,1 */ 1670400 173 /* speedbin 2 */ 1708800 186 /* speedbin 0,1 */ 1747200 201 /* speedbin 2 */ >; idle-cost-data = < 22 18 14 12 >; }; CPU_COST_1: core-cost1 { busy-cost-data = < 300000 258 422400 260 499200 261 576000 263 652800 267 729600 272 806400 280 883200 291 960000 305 1036800 324 1113600 348 1190400 378 1267200 415 1344000 460 1420800 513 1497600 576 1574400 649 1651200 732 1728000 824 1804800 923 1881600 1027 1958400 1131 2035000 1228 /* speedbin 1,2 */ 2092000 1290 /* speedbin 1 */ 2112000 1308 /* speedbin 2 */ 2208000 1363 /* speedbin 2 */ >; idle-cost-data = < 100 80 60 40 >; }; CLUSTER_COST_0: cluster-cost0 { busy-cost-data = < 300000 3 422400 4 499200 4 576000 4 652800 5 748800 5 825600 6 902400 7 979200 7 1056000 8 1132800 9 1209600 9 1286400 10 1363200 11 1440000 12 1516800 13 1593600 15 1651200 17 /* speedbin 0,1 */ 1670400 19 /* speedbin 2 */ 1708800 21 /* speedbin 0,1 */ 1747200 23 /* speedbin 2 */ >; idle-cost-data = < 4 3 2 1 >; }; CLUSTER_COST_1: cluster-cost1 { busy-cost-data = < 300000 24 422400 24 499200 25 576000 25 652800 26 729600 27 806400 28 883200 29 960000 30 1036800 32 1113600 34 1190400 37 1267200 40 1344000 45 1420800 50 1497600 57 1574400 64 1651200 74 1728000 84 1804800 96 1881600 106 1958400 113 2035000 120 /* speedbin 1,2 */ 2092000 125 /* speedbin 1 */ 2112000 127 /* speedbin 2 */ 2208000 130 /* speedbin 2 */ >; idle-cost-data = < 4 3 2 1 >; }; }; /* energy-costs */
在代码kernel/sched/energy.c中遍历所有cpu，并读取dts中的数据
for_each_possible_cpu(cpu) { cn = of_get_cpu_node(cpu, NULL); if (!cn) { pr_warn("CPU device node missing for CPU %d\n", cpu); return; } if (!of_find_property(cn, "sched-energy-costs", NULL)) { pr_warn("CPU device node has no sched-energy-costs\n"); return; } for_each_possible_sd_level(sd_level) { cp = of_parse_phandle(cn, "sched-energy-costs", sd_level); if (!cp) break; prop = of_find_property(cp, "busy-cost-data", NULL); if (!prop || !prop->value) { pr_warn("No busy-cost data, skipping sched_energy init\n"); goto out; } sge = kcalloc(1, sizeof(struct sched_group_energy), GFP_NOWAIT); if (!sge) goto out; nstates = (prop->length / sizeof(u32)) / 2; cap_states = kcalloc(nstates, sizeof(struct capacity_state), GFP_NOWAIT); if (!cap_states) { kfree(sge); goto out; } for (i = 0, val = prop->value; i < nstates; i++) { //将读取的[freq,energy]数组存放起来 cap_states[i].cap = SCHED_CAPACITY_SCALE; cap_states[i].frequency = be32_to_cpup(val++); cap_states[i].power = be32_to_cpup(val++); } sge->nr_cap_states = nstates; //state为[freq,energy]组合个数，就是支持多少个状态：将所有数据flatten之后，再处以2 sge->cap_states = cap_states; prop = of_find_property(cp, "idle-cost-data", NULL); if (!prop || !prop->value) { pr_warn("No idle-cost data, skipping sched_energy init\n"); kfree(sge); kfree(cap_states); goto out; } nstates = (prop->length / sizeof(u32)); idle_states = kcalloc(nstates, sizeof(struct idle_state), GFP_NOWAIT); if (!idle_states) { kfree(sge); kfree(cap_states); goto out; } for (i = 0, val = prop->value; i < nstates; i++) idle_states[i].power = be32_to_cpup(val++); //将读取的idle cost data存放起来 sge->nr_idle_states = nstates; //idle state的个数，就是idle cost data的长度 sge->idle_states = idle_states; sge_array[cpu][sd_level] = sge; //将当前cpu获取的energy模型存放再sge_array[cpu][sd_level]中。其中cpu就是对应哪个cpu，sd_level则对应是哪个sched_domain，也就是是cpu level还是cluster level } }
Load Tracking
QTI EAS使用的负载计算是WALT，是基于时间窗口的load统计方法，具体参考之前文章：https://www.cnblogs.com/lingjiajun/p/12317090.html
其中会跟踪计算出2个比较关键的数据，就是task_util和cpu_util
当执行wakeup task placement，scheduler就会使用task utilization和CPU utilization
可以理解为将load的情况转化为Utilization，并且将其标准化为1024的值。
task_util = demand *1024 / window_size
= (delta / window_size) * (cur_freq / max_freq) * cpu_max_capacity
-----delta是task在一个window中运行的真实时间；window_size默认是20ms；
cur_freq为cpu当前频率；max_freq为cpu最大频率；
Task utilization boosted = Task utilization + (1024-task_util) x boost_percent -----boost percent是使用schedtune boost时，所需要乘上的百分比
CPU utilization = 1024 x (累计的runnable均值 / window size)--------累计的runnable均值，个人理解就是rq上所有task util的总和
Task placement的主要概念：
EAS是Task placement 是EAS影响调度的主要模块。 其主要keypoint如下：
1、EAS依靠energy model来进行精确地进行选择CPU运行
2、使用energy model估算：把一个任务安排在一个CPU上，或者将任务从一个CPU迁移到另一个CPU上，所发生的能量变化
3、EAS会在不影响performance情况下（比如满足满足最低的latency），趋向于选择消耗能量最小的CPU，去运行当前的task
4、EAS仅发生在system没有overutilized的情况下
5、EAS的概念与QTI EAS的一样
6、一旦系统处于overutilized，QTI EAS仍然在wake up的path下进行energy aware。不会考虑系统overutilized的情形。
补充：
overutilization，一个cpu_util大于cpu capacity的95%（sched_capacity_margin_up[cpu]），那么就认为这个cpu处于overutilization。并且整个系统也被认为overutilizaion。
EAS核心调度算法
不同版本的EAS在不同版本下的主要task placement实现函数（针对CFS task）：
Zone scheduler： select_best_cpu() QTI EAS r1.2： energy_aware_wake_cpu() QTI EAS r1.5： find_energy_efficienct_cpu()
task placement调用路径：
QTI EAS r1.5 (Kernel 4.14) Task wake-up： try_to_wake_up() →select_task_rq_fair() →invokes find_energy_efficient_cpu() Scheduler tick occurs： scheduler_tick() →check_for_migration() →invokes find_energy_efficient_cpu() New task arrives： do_fork() →wake_up_new_task() →select_task_rq_fair() →invokes find_energy_efficient_cpu()
EAS的task placement代码流程，主要目标是找到一个合适的cpu来运行当前这个task p。
主要代码就是find_energy_efficient_cpu（）这个函数里面，如下：
1 /* 2 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the 3 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum 4 * spare capacity in each performance domain and uses it as a potential 5 * candidate to execute the task. Then, it uses the Energy Model to figure 6 * out which of the CPU candidates is the most energy-efficient. 7 * 8 * The rationale for this heuristic is as follows. In a performance domain, 9 * all the most energy efficient CPU candidates (according to the Energy 10 * Model) are those for which we'll request a low frequency. When there are 11 * several CPUs for which the frequency request will be the same, we don't 12 * have enough data to break the tie between them, because the Energy Model 13 * only includes active power costs. With this model, if we assume that 14 * frequency requests follow utilization (e.g. using schedutil), the CPU with 15 * the maximum spare capacity in a performance domain is guaranteed to be among 16 * the best candidates of the performance domain. 17 * 18 * In practice, it could be preferable from an energy standpoint to pack 19 * small tasks on a CPU in order to let other CPUs go in deeper idle states, 20 * but that could also hurt our chances to go cluster idle, and we have no 21 * ways to tell with the current Energy Model if this is actually a good 22 * idea or not. So, find_energy_efficient_cpu() basically favors 23 * cluster-packing, and spreading inside a cluster. That should at least be 24 * a good thing for latency, and this is consistent with the idea that most 25 * of the energy savings of EAS come from the asymmetry of the system, and 26 * not so much from breaking the tie between identical CPUs. That's also the 27 * reason why EAS is enabled in the topology code only for systems where 28 * SD_ASYM_CPUCAPACITY is set. 29 * 30 * NOTE: Forkees are not accepted in the energy-aware wake-up path because 31 * they don't have any useful utilization data yet and it's not possible to 32 * forecast their impact on energy consumption. Consequently, they will be 33 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out 34 * to be energy-inefficient in some use-cases. The alternative would be to 35 * bias new tasks towards specific types of CPUs first, or to try to infer 36 * their util_avg from the parent task, but those heuristics could hurt 37 * other use-cases too. So, until someone finds a better way to solve this, 38 * let's keep things simple by re-using the existing slow path. 39 */ 40 41 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu, 42 int sync, int sibling_count_hint) 43 { 44 unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX; 45 struct root_domain *rd = cpu_rq(smp_processor_id())->rd; 46 int weight, cpu = smp_processor_id(), best_energy_cpu = prev_cpu; //cpu：当前执行的cpu 47 unsigned long cur_energy; 48 struct perf_domain *pd; 49 struct sched_domain *sd; 50 cpumask_t *candidates; 51 bool is_rtg; 52 struct find_best_target_env fbt_env; 53 bool need_idle = wake_to_idle(p); //是否set flag PF_WAKE_UP_IDLE 54 int placement_boost = task_boost_policy(p); //获取task sched boost policy：none/on_big/on_all 与sched_boost、schedtune设置也有关 55 u64 start_t = 0; 56 int delta = 0; 57 int task_boost = per_task_boost(p); //仅网络有打开该boost，这里可以认为没有boost 58 int boosted = (schedtune_task_boost(p) > 0) || (task_boost > 0); //查看task的schedtune有没有打开boost 59 int start_cpu = get_start_cpu(p); //获取从哪个cpu core开始，尝试作为target cpu 60 61 if (start_cpu < 0) 62 goto eas_not_ready; 63 64 is_rtg = task_in_related_thread_group(p); //判断task是否在一个group内 65 66 fbt_env.fastpath = 0; 67 68 if (trace_sched_task_util_enabled()) 69 start_t = sched_clock(); //trace log 70 71 /* Pre-select a set of candidate CPUs. */ 72 candidates = this_cpu_ptr(&energy_cpus); 73 cpumask_clear(candidates); 74 75 if (need_idle) 76 sync = 0; 77 78 if (sysctl_sched_sync_hint_enable && sync && 79 bias_to_this_cpu(p, cpu, start_cpu)) { //满足3个调节：sync hint enable/flag：sync=1/bias to当前cpu 80 best_energy_cpu = cpu; //当前执行的cpu 81 fbt_env.fastpath = SYNC_WAKEUP; 82 goto done; 83 } 84 85 if (is_many_wakeup(sibling_count_hint) && prev_cpu != cpu && //sibling_count_hint代表有多少个thread在当前event中唤醒 86 bias_to_this_cpu(p, prev_cpu, start_cpu)) { 87 best_energy_cpu = prev_cpu; //选择prev cpu 88 fbt_env.fastpath = MANY_WAKEUP; 89 goto done; 90 } 91 92 rcu_read_lock(); 93 pd = rcu_dereference(rd->pd); 94 if (!pd) 95 goto fail; 96 97 /* 98 * Energy-aware wake-up happens on the lowest sched_domain starting 99 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. 100 */ 101 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); 102 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) 103 sd = sd->parent; 104 if (!sd) 105 goto fail; 106 107 sync_entity_load_avg(&p->se); //更新task所在sched_entity的PELT load 108 if (!task_util_est(p)) 109 goto unlock; 110 111 if (sched_feat(FIND_BEST_TARGET)) { //检查FIND_BEST_TARGET这个调度特性是否打开：目前是打开的 112 fbt_env.is_rtg = is_rtg; 113 fbt_env.placement_boost = placement_boost; 114 fbt_env.need_idle = need_idle; 115 fbt_env.start_cpu = start_cpu; 116 fbt_env.boosted = boosted; 117 fbt_env.strict_max = is_rtg && 118 (task_boost == TASK_BOOST_STRICT_MAX); 119 fbt_env.skip_cpu = is_many_wakeup(sibling_count_hint) ? 120 cpu : -1; 121 122 find_best_target(NULL, candidates, p, &fbt_env); //（1）核心函数，最终是将找到的target_cpu和backup_cpu都存放进了candidates中 123 } else { 124 select_cpu_candidates(sd, candidates, pd, p, prev_cpu); 125 } 126 127 /* Bail out if no candidate was found. */ 128 weight = cpumask_weight(candidates); //判断如果没有找到target cpu和backup cpu时，直接goto unlock 129 if (!weight) 130 goto unlock; 131 132 /* If there is only one sensible candidate, select it now. */ 133 cpu = cpumask_first(candidates); 134 if (weight == 1 && ((schedtune_prefer_idle(p) && idle_cpu(cpu)) || //如果只找到了1个cpu，task是prefer_idle并且这个cpu也是idle的；或者cpu就是prev_cpu 135 (cpu == prev_cpu))) { 136 best_energy_cpu = cpu; //那么就选这个cpu为【best_energy_cpu】 137 goto unlock; 138 } 139 140 #ifdef CONFIG_SCHED_WALT 141 if (p->state == TASK_WAKING) //如果是新唤醒的task，获取task_util 142 delta = task_util(p); 143 #endif 144 if (task_placement_boost_enabled(p) || need_idle || boosted || //满足一下条件之一，那么第一个candidate cpu就作为【best_energy_cpu】不再考虑计算energy 145 is_rtg || __cpu_overutilized(prev_cpu, delta) || //打开了sched_boost、need_idle（PF_WAKE_UP_IDLE）、开了schedtune boost、related_thread_group限制使用小核、prev_cpu+delta没有overutil、 146 !task_fits_max(p, prev_cpu) || cpu_isolated(prev_cpu)) { //p放在prev_cpu上会misfit、prev_cpu处于isolated 147 best_energy_cpu = cpu; 148 goto unlock; 149 } 150 151 if (cpumask_test_cpu(prev_cpu, &p->cpus_allowed)) //根据prev_cpu是否在task p的cpuset范围内 152 prev_energy = best_energy = compute_energy(p, prev_cpu, pd); //（2）在范围内，则计算p在prev_cpu上的energy 153 else 154 prev_energy = best_energy = ULONG_MAX; //不在范围内，energy就设为最大，说明prev_cpu不考虑作为best_energy_cpu了 155 156 /* Select the best candidate energy-wise. */ //通过比较energy，挑选出best_energy_cpu、best_energy 157 for_each_cpu(cpu, candidates) { 158 if (cpu == prev_cpu) //过滤prev_cpu 159 continue; 160 cur_energy = compute_energy(p, cpu, pd); //计算p迁移到candidate cpu上的energy 161 trace_sched_compute_energy(p, cpu, cur_energy, prev_energy, 162 best_energy, best_energy_cpu); 163 if (cur_energy < best_energy) { 164 best_energy = cur_energy; 165 best_energy_cpu = cpu; 166 } else if (cur_energy == best_energy) { 167 if (select_cpu_same_energy(cpu, best_energy_cpu, //当candidate cpu的energy与best_cpu一样的话，怎么选 168 prev_cpu)) { 169 best_energy = cur_energy; 170 best_energy_cpu = cpu; 171 } 172 } 173 } 174 unlock: 175 rcu_read_unlock(); 176 177 /* 178 * Pick the prev CPU, if best energy CPU can't saves at least 6% of 179 * the energy used by prev_cpu. 180 */ 181 if ((prev_energy != ULONG_MAX) && (best_energy_cpu != prev_cpu) && //找到了非prev_cpu的best_energy_cpu、且省电下来的energy要大于在prev_energy上的6%，那么best_energy_cpu则满足条件；否则仍然使用prev_cpu 182 ((prev_energy - best_energy) <= prev_energy >> 4)) //这里巧妙地使用了位移：右移1位代表÷2，所以prev_energy/2/2/2/2 = prev_energy*6% 183 best_energy_cpu = prev_cpu; 184 185 done: 186 187 trace_sched_task_util(p, cpumask_bits(candidates)[0], best_energy_cpu, 188 sync, need_idle, fbt_env.fastpath, placement_boost, 189 start_t, boosted, is_rtg, get_rtg_status(p), start_cpu); 190 191 return best_energy_cpu; 192 193 fail: 194 rcu_read_unlock(); 195 eas_not_ready: 196 return -1; 197 }
（1）find_best_target（）
1 static void find_best_target(struct sched_domain *sd, cpumask_t *cpus, 2 struct task_struct *p, 3 struct find_best_target_env *fbt_env) 4 { 5 unsigned long min_util = boosted_task_util(p); //获取p的boosted_task_util 6 unsigned long target_capacity = ULONG_MAX; 7 unsigned long min_wake_util = ULONG_MAX; 8 unsigned long target_max_spare_cap = 0; 9 unsigned long best_active_util = ULONG_MAX; 10 unsigned long best_active_cuml_util = ULONG_MAX; 11 unsigned long best_idle_cuml_util = ULONG_MAX; 12 bool prefer_idle = schedtune_prefer_idle(p); //获取task prefer_idle配置 13 bool boosted = fbt_env->boosted; 14 /* Initialise with deepest possible cstate (INT_MAX) */ 15 int shallowest_idle_cstate = INT_MAX; 16 struct sched_domain *start_sd; 17 struct sched_group *sg; 18 int best_active_cpu = -1; 19 int best_idle_cpu = -1; 20 int target_cpu = -1; 21 int backup_cpu = -1; 22 int i, start_cpu; 23 long spare_wake_cap, most_spare_wake_cap = 0; 24 int most_spare_cap_cpu = -1; 25 int prev_cpu = task_cpu(p); 26 bool next_group_higher_cap = false; 27 int isolated_candidate = -1; 28 29 /* 30 * In most cases, target_capacity tracks capacity_orig of the most 31 * energy efficient CPU candidate, thus requiring to minimise 32 * target_capacity. For these cases target_capacity is already 33 * initialized to ULONG_MAX. 34 * However, for prefer_idle and boosted tasks we look for a high 35 * performance CPU, thus requiring to maximise target_capacity. In this 36 * case we initialise target_capacity to 0. 37 */ 38 if (prefer_idle && boosted) 39 target_capacity = 0; 40 41 if (fbt_env->strict_max) 42 most_spare_wake_cap = LONG_MIN; 43 44 /* Find start CPU based on boost value */ 45 start_cpu = fbt_env->start_cpu; 46 /* Find SD for the start CPU */ 47 start_sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, start_cpu)); //找到start cpu所在的sched domain，sd_asym_cpucapacity表示是非对称cpu capacity级别，应该就是DIE level，所以domain是cpu0-7 48 if (!start_sd) 49 goto out; 50 51 /* fast path for prev_cpu */ 52 if (((capacity_orig_of(prev_cpu) == capacity_orig_of(start_cpu)) || //prev cpu和start cpu的当前max_policy_freq下的capacity相等 53 asym_cap_siblings(prev_cpu, start_cpu)) && 54 !cpu_isolated(prev_cpu) && cpu_online(prev_cpu) && 55 idle_cpu(prev_cpu)) { 56 57 if (idle_get_state_idx(cpu_rq(prev_cpu)) <= 1) { //prev cpu idle state的index <1，说明休眠不深 58 target_cpu = prev_cpu; 59 60 fbt_env->fastpath = PREV_CPU_FASTPATH; 61 goto target; 62 } 63 } 64 65 /* Scan CPUs in all SDs */ 66 sg = start_sd->groups; 67 do { //do-while循环，针对start cpu的调度域中的所有调度组进行遍历，由于domain是cpu0-7，那么调度组就是2个大小cluster：cpu0-3，cpu4-7 68 for_each_cpu_and(i, &p->cpus_allowed, sched_group_span(sg)) { //寻找task允许的cpuset和调度组可用cpu范围内 69 unsigned long capacity_curr = capacity_curr_of(i); //当前freq的cpu_capacity 70 unsigned long capacity_orig = capacity_orig_of(i); //当前max_policy_freq的cpu_capacity, >=capacity_curr 71 unsigned long wake_util, new_util, new_util_cuml; 72 long spare_cap; 73 int idle_idx = INT_MAX; 74 75 trace_sched_cpu_util(i); 76 77 if (!cpu_online(i) || cpu_isolated(i)) //cpu处于非online，或者isolate状态，则直接不考虑 78 continue; 79 80 if (isolated_candidate == -1) 81 isolated_candidate = i; 82 83 /* 84 * This CPU is the target of an active migration that's 85 * yet to complete. Avoid placing another task on it. 86 * See check_for_migration() 87 */ 88 if (is_reserved(i)) //已经有task要迁移到上面，但是还没有迁移完成。所以这样的cpu不考虑 89 continue; 90 91 if (sched_cpu_high_irqload(i)) //高irq load的cpu不考虑。irq load可以参考之前WALT文章：https://www.cnblogs.com/lingjiajun/p/12317090.html 92 continue; 93 94 if (fbt_env->skip_cpu == i) //当前活动的cpu是否有很多event一起wakeup，如果有，那么也不考虑该cpu 95 continue; 96 97 /* 98 * p's blocked utilization is still accounted for on prev_cpu 99 * so prev_cpu will receive a negative bias due to the double 100 * accounting. However, the blocked utilization may be zero. 101 */ 102 wake_util = cpu_util_without(i, p); //计算没有除了p以外的cpu_util（p不在该cpu rq的情况下，实际就是当前cpu_util） 103 new_util = wake_util + task_util_est(p); //计算cpu_util + p的task_util（p的task_util就是walt统计的demand_scaled） 104 spare_wake_cap = capacity_orig - wake_util; //剩余的capacity = capacity_orig - p以外的cpu_util 105 106 if (spare_wake_cap > most_spare_wake_cap) { 107 most_spare_wake_cap = spare_wake_cap; //在循环中，找到有剩余capacity最多（最空闲）的cpu = i，并保存剩余的capacity 108 most_spare_cap_cpu = i; 109 } 110 111 if (per_task_boost(cpu_rq(i)->curr) == //cpu【i】当前running_task的task_boost == TASK_BOOST_STRICT_MAX，那么不适合作为tager_cpu 112 TASK_BOOST_STRICT_MAX) 113 continue; 114 /* 115 * Cumulative demand may already be accounting for the 116 * task. If so, add just the boost-utilization to 117 * the cumulative demand of the cpu. 118 */ 119 if (task_in_cum_window_demand(cpu_rq(i), p)) //计算新的cpu【i】的cpu_util_cum = cpu_util_cum + p的boosted_task_util 120 new_util_cuml = cpu_util_cum(i, 0) + //特别地，如果p已经在cpu【i】的rq中，或者p的部分demand被统计在了walt中。那么防止统计2次，所以要减去p的task_util（denamd_scaled） 121 min_util - task_util(p); 122 else 123 new_util_cuml = cpu_util_cum(i, 0) + min_util; 124 125 /* 126 * Ensure minimum capacity to grant the required boost. 127 * The target CPU can be already at a capacity level higher 128 * than the one required to boost the task. 129 */ 130 new_util = max(min_util, new_util); //取 p的booted_task_util、加入p之后的cpu_util，之间的较大值 131 if (new_util > capacity_orig) //与capacity_orig比较，大于capacity_orig的情况下，不适合作为target_cpu 132 continue; 133 134 /* 135 * Pre-compute the maximum possible capacity we expect 136 * to have available on this CPU once the task is 137 * enqueued here. 138 */ 139 spare_cap = capacity_orig - new_util; //预计算当p迁移到cpu【i】上后，剩余的可能最大capacity 140 141 if (idle_cpu(i)) //判断当前cpu【i】是否处于idle，并获取idle index（idle的深度） 142 idle_idx = idle_get_state_idx(cpu_rq(i)); 143 144 145 /* 146 * Case A) Latency sensitive tasks 147 * 148 * Unconditionally favoring tasks that prefer idle CPU to 149 * improve latency. 150 * 151 * Looking for: 152 * - an idle CPU, whatever its idle_state is, since 153 * the first CPUs we explore are more likely to be 154 * reserved for latency sensitive tasks. 155 * - a non idle CPU where the task fits in its current 156 * capacity and has the maximum spare capacity. 157 * - a non idle CPU with lower contention from other 158 * tasks and running at the lowest possible OPP. 159 * 160 * The last two goals tries to favor a non idle CPU 161 * where the task can run as if it is "almost alone". 162 * A maximum spare capacity CPU is favoured since 163 * the task already fits into that CPU's capacity 164 * without waiting for an OPP chance. 165 * 166 * The following code path is the only one in the CPUs 167 * exploration loop which is always used by 168 * prefer_idle tasks. It exits the loop with wither a 169 * best_active_cpu or a target_cpu which should 170 * represent an optimal choice for latency sensitive 171 * tasks. 172 */ 173 if (prefer_idle) { //对lantency有要求的task 174 /* 175 * Case A.1: IDLE CPU 176 * Return the best IDLE CPU we find: 177 * - for boosted tasks: the CPU with the highest 178 * performance (i.e. biggest capacity_orig) 179 * - for !boosted tasks: the most energy 180 * efficient CPU (i.e. smallest capacity_orig) 181 */ 182 if (idle_cpu(i)) { //如果cpu【i】是idle的 183 if (boosted && 184 capacity_orig < target_capacity) //对于boosted task，cpu需要选择最大capacity_orig，不满足要continue 185 continue; 186 if (!boosted && 187 capacity_orig > target_capacity) //对于非boosted task，cpu选择最小capacity_orig，不满足要continue 188 continue; 189 /* 190 * Minimise value of idle state: skip 191 * deeper idle states and pick the 192 * shallowest. 193 */ 194 if (capacity_orig == target_capacity && 195 sysctl_sched_cstate_aware && 196 idle_idx >= shallowest_idle_cstate) //包括下面的continue，都是为了挑选出处于idle最浅的cpu 197 continue; 198 199 target_capacity = capacity_orig; 200 shallowest_idle_cstate = idle_idx; 201 best_idle_cpu = i; //选出【prefer_idle】best_idle_cpu 202 continue; 203 } 204 if (best_idle_cpu != -1) //过滤上面已经找到best_idle_cpu的情况，不需要走下面流程了 205 continue; 206 207 /* 208 * Case A.2: Target ACTIVE CPU 209 * Favor CPUs with max spare capacity. 210 */ 211 if (capacity_curr > new_util && 212 spare_cap > target_max_spare_cap) { //找到capacity_curr满足包含进程p的cpu_util，并且找到空闲capacity最多的那个cpu 213 target_max_spare_cap = spare_cap; 214 target_cpu = i; //选出【prefer_idle】target_cpu 215 continue; 216 } 217 if (target_cpu != -1) //如果cpu条件不满足，则continue，继续找target_cpu 218 continue; 219 220 221 /* 222 * Case A.3: Backup ACTIVE CPU 223 * Favor CPUs with: 224 * - lower utilization due to other tasks 225 * - lower utilization with the task in 226 */ 227 if (wake_util > min_wake_util) //找出除了p以外的cpu_util最小的cpu 228 continue; 229 230 /* 231 * If utilization is the same between CPUs, 232 * break the ties with WALT's cumulative 233 * demand 234 */ 235 if (new_util == best_active_util && 236 new_util_cuml > best_active_cuml_util) //如果包含p的cpu_util相等，那么就挑选cpu_util_cum + p的boosted_task_util最小的那个cpu 237 continue; 238 min_wake_util = wake_util; 239 best_active_util = new_util; 240 best_active_cuml_util = new_util_cuml; 241 best_active_cpu = i; //选出【prefer_idle】best_active_cpu 242 continue; 243 } 244 245 /* 246 * Skip processing placement further if we are visiting 247 * cpus with lower capacity than start cpu 248 */ 249 if (capacity_orig < capacity_orig_of(start_cpu)) //cpu【i】capacity_orig < 【start_cpu】capacity_orig的不考虑 250 continue; 251 252 /* 253 * Case B) Non latency sensitive tasks on IDLE CPUs. 254 * 255 * Find an optimal backup IDLE CPU for non latency 256 * sensitive tasks. 257 * 258 * Looking for: 259 * - minimizing the capacity_orig, 260 * i.e. preferring LITTLE CPUs 261 * - favoring shallowest idle states 262 * i.e. avoid to wakeup deep-idle CPUs 263 * 264 * The following code path is used by non latency 265 * sensitive tasks if IDLE CPUs are available. If at 266 * least one of such CPUs are available it sets the 267 * best_idle_cpu to the most suitable idle CPU to be 268 * selected. 269 * 270 * If idle CPUs are available, favour these CPUs to 271 * improve performances by spreading tasks. 272 * Indeed, the energy_diff() computed by the caller67jkkk 273 * will take care to ensure the minimization of energy 274 * consumptions without affecting performance. 275 */ //对latency要求不高的task，并要求idle cpu作为target的情况 276 if (idle_cpu(i)) { //判断cpu【i】是否idle 277 /* 278 * Prefer shallowest over deeper idle state cpu, 279 * of same capacity cpus. 280 */ 281 if (capacity_orig == target_capacity && //选出capacity相同情况下，idle最浅的cpu 282 sysctl_sched_cstate_aware && 283 idle_idx > shallowest_idle_cstate) 284 continue; 285 286 if (shallowest_idle_cstate == idle_idx && 287 target_capacity == capacity_orig && 288 (best_idle_cpu == prev_cpu || 289 (i != prev_cpu && 290 new_util_cuml > best_idle_cuml_util))) //best_idle_cpu非prev_cpu，并且挑选cpu_util_cum + p的boosted_task_util最小的 291 continue; 292 293 target_capacity = capacity_orig; 294 shallowest_idle_cstate = idle_idx; 295 best_idle_cuml_util = new_util_cuml; 296 best_idle_cpu = i; //选出【normal-idle】best_idle_cpu 297 continue; 298 } 299 300 /* 301 * Consider only idle CPUs for active migration. 302 */ 303 if (p->state == TASK_RUNNING) //task p正在运行说明是misfit task，只考虑idle cpu作为target，不进行下面流程 304 continue; 305 306 /* 307 * Case C) Non latency sensitive tasks on ACTIVE CPUs. 308 * 309 * Pack tasks in the most energy efficient capacities. 310 * 311 * This task packing strategy prefers more energy 312 * efficient CPUs (i.e. pack on smaller maximum 313 * capacity CPUs) while also trying to spread tasks to 314 * run them all at the lower OPP. 315 * 316 * This assumes for example that it's more energy 317 * efficient to run two tasks on two CPUs at a lower 318 * OPP than packing both on a single CPU but running 319 * that CPU at an higher OPP. 320 * 321 * Thus, this case keep track of the CPU with the 322 * smallest maximum capacity and highest spare maximum 323 * capacity. 324 */ //对latency要求不高，并需要ACTIVE cpu作为target的情况 325 326 /* Favor CPUs with maximum spare capacity */ 327 if (spare_cap < target_max_spare_cap) //找到迁移p之后，剩余capacity最多的cpu 328 continue; 329 330 target_max_spare_cap = spare_cap; 331 target_capacity = capacity_orig; 332 target_cpu = i; //找出【normal-ACTIVe】的target_cpu 333 } //到此就是一个调度组（cluster）内cpu的循环查找 334 335 next_group_higher_cap = (capacity_orig_of(group_first_cpu(sg)) < 336 capacity_orig_of(group_first_cpu(sg->next))); //尝试查找下一个capacity更大的big cluster 337 338 /* 339 * If we've found a cpu, but the boost is ON_ALL we continue 340 * visiting other clusters. If the boost is ON_BIG we visit 341 * next cluster if they are higher in capacity. If we are 342 * not in any kind of boost, we break. 343 * 344 * And always visit higher capacity group, if solo cpu group 345 * is not in idle. 346 */ 347 if (!prefer_idle && !boosted && //上面找到cpu但是boost=ON_ALL，那么还要查找其他cluster 348 ((target_cpu != -1 && (sg->group_weight > 1 || //上面找到cpu但是boost=ON_BIG，那么还要在capacity更大的cluster中查找 349 !next_group_higher_cap)) || //上面找到了cpu，并且不在任何boost。那么break 350 best_idle_cpu != -1) && //如果上面group中，没有cpu是idle，那么always在capacity更大的cluster中查找 351 (fbt_env->placement_boost == SCHED_BOOST_NONE || 352 !is_full_throttle_boost() || 353 (fbt_env->placement_boost == SCHED_BOOST_ON_BIG && 354 !next_group_higher_cap))) 355 break; 356 357 /* 358 * if we are in prefer_idle and have found an idle cpu, 359 * break from searching more groups based on the stune.boost and 360 * group cpu capacity. For !prefer_idle && boosted case, don't 361 * iterate lower capacity CPUs unless the task can't be 362 * accommodated in the higher capacity CPUs. 363 */ 364 if ((prefer_idle && best_idle_cpu != -1) || //如果设置了prefer_idle，并且找到了一个idle cpu；根据schedtune是否打开boost和是否有更大capacity的cluster进行判断是否break 365 (boosted && (best_idle_cpu != -1 || target_cpu != -1 || //没有prefer_idle，但是打开boost的情况，除非high capacity的cpu不能接受task，否则不用再遍历low capacity的cpu 366 (fbt_env->strict_max && most_spare_cap_cpu != -1)))) { 367 if (boosted) { 368 if (!next_group_higher_cap) 369 break; 370 } else { 371 if (next_group_higher_cap) 372 break; 373 } 374 } 375 376 } while (sg = sg->next, sg != start_sd->groups); 377 378 adjust_cpus_for_packing(p, &target_cpu, &best_idle_cpu, //计算将task放在target_cpu时，在考虑20%的余量，和sched_load_boost之后，看capacity是否满足target_cpu当前freq的capacity 379 shallowest_idle_cstate, //另外检查rtg，看是否不考虑idle cpu 380 fbt_env, boosted); 381 382 /* 383 * For non latency sensitive tasks, cases B and C in the previous loop, 384 * we pick the best IDLE CPU only if we was not able to find a target 385 * ACTIVE CPU. //latency要求不高的task选择cpu优先级：ACTIVE cpu > idle cpu；没有ACITVE，则选idle cpu 386 * 387 * Policies priorities: 388 * 389 * - prefer_idle tasks: //prefer_idle的task选择cpu优先级：idle cpu > ACTIVE cpu(包含task之后又更多spare capacity) > ACTIVE cpu(更小cpu_util+boosted_task_util) 390 * 391 * a) IDLE CPU available: best_idle_cpu 392 * b) ACTIVE CPU where task fits and has the bigger maximum spare 393 * capacity (i.e. target_cpu) 394 * c) ACTIVE CPU with less contention due to other tasks 395 * (i.e. best_active_cpu) 396 * 397 * - NON prefer_idle tasks: //非prefer_idle的task选择cpu优先级：ACTIVE cpu > idle cpu 398 * 399 * a) ACTIVE CPU: target_cpu 400 * b) IDLE CPU: best_idle_cpu 401 */ 402 403 if (prefer_idle && (best_idle_cpu != -1)) { //prefer_idle的task，直接选择best_idle_cpu作为target 404 target_cpu = best_idle_cpu; 405 goto target; 406 } 407 408 if (target_cpu == -1) //假如target没有找到，那么重新找target： 409 target_cpu = prefer_idle 410 ? best_active_cpu //1、prefer_idle的task选择best_active_cpu; 411 : best_idle_cpu; //2、而非prefer_idle的task选择best_idle_cpu 412 else 413 backup_cpu = prefer_idle //假如找到了target，那么再选backup_cpu： 414 ? best_active_cpu //1、prefer_idle的task选择 best_active_cpu 415 : best_idle_cpu; //2、非prefer_idle的task选择 best_idle_cpu 416 417 if (target_cpu == -1 && most_spare_cap_cpu != -1 && 418 /* ensure we use active cpu for active migration */ //active migration（misfit task迁移）情况只选择active cpu 419 !(p->state == TASK_RUNNING && !idle_cpu(most_spare_cap_cpu))) 420 target_cpu = most_spare_cap_cpu; 421 422 if (target_cpu == -1 && isolated_candidate != -1 && //假如没有找到target_cpu，prev_cpu又处于isolated，而task允许的所有cpu中有online并且unisolated的 423 cpu_isolated(prev_cpu)) 424 target_cpu = isolated_candidate; //那么就选择最后一个online并unisolated的cpu作为target 425 426 if (backup_cpu >= 0) 427 cpumask_set_cpu(backup_cpu, cpus); //将backup_cpu存放进cpus中 428 if (target_cpu >= 0) { 429 target: 430 cpumask_set_cpu(target_cpu, cpus); //将找出的target cpu存放进cpus中 431 } 432 433 out: 434 trace_sched_find_best_target(p, prefer_idle, min_util, start_cpu, 435 best_idle_cpu, best_active_cpu, 436 most_spare_cap_cpu, 437 target_cpu, backup_cpu); 438 }
（2）计算energy
/* * compute_energy(): Estimates the energy that would be consumed if @p was * migrated to @dst_cpu. compute_energy() predicts what will be the utilization * landscape of the * CPUs after the task migration, and uses the Energy Model * to compute what would be the energy if we decided to actually migrate that * task. */ static long compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) { long util, max_util, sum_util, energy = 0; int cpu; for (; pd; pd = pd->next) { max_util = sum_util = 0; /* * The capacity state of CPUs of the current rd can be driven by * CPUs of another rd if they belong to the same performance * domain. So, account for the utilization of these CPUs too * by masking pd with cpu_online_mask instead of the rd span. * * If an entire performance domain is outside of the current rd, * it will not appear in its pd list and will not be accounted * by compute_energy(). */ for_each_cpu_and(cpu, perf_domain_span(pd), cpu_online_mask) { //在perf domain的cpu中找出online的 #ifdef CONFIG_SCHED_WALT util = cpu_util_next_walt(cpu, p, dst_cpu); //计算迁移task p之后，每个cpu的util情况 #else util = cpu_util_next(cpu, p, dst_cpu); util += cpu_util_rt(cpu_rq(cpu)); util = schedutil_energy_util(cpu, util); #endif max_util = max(util, max_util); //找到perf domain中cpu util最大的值（同perf domain，即cluster，最大的util决定了freq的设定） sum_util += util; //统计迁移之后，perf domain内的总util } energy += em_pd_energy(pd->em_pd, max_util, sum_util); //计算perf domain的energy，并累计大小cluster的energy，就是整个系统energy } return energy; }
获取perf domain内的energy，在其中有2个重要的结构体：
/** * em_cap_state - Capacity state of a performance domain * @frequency: The CPU frequency in KHz, for consistency with CPUFreq * @power: The power consumed by 1 CPU at this level, in milli-watts * @cost: The cost coefficient associated with this level, used during * energy calculation. Equal to: power * max_frequency / frequency */ struct em_cap_state { unsigned long frequency; unsigned long power; unsigned long cost; }; /** * em_perf_domain - Performance domain * @table: List of capacity states, in ascending order * @nr_cap_states: Number of capacity states * @cpus: Cpumask covering the CPUs of the domain * * A "performance domain" represents a group of CPUs whose performance is * scaled together. All CPUs of a performance domain must have the same * micro-architecture. Performance domains often have a 1-to-1 mapping with * CPUFreq policies. */ struct em_perf_domain { struct em_cap_state *table; int nr_cap_states; unsigned long cpus[0]; };
em_pd_energy函数可以得到perf domain的energy。
/** * em_pd_energy() - Estimates the energy consumed by the CPUs of a perf. domain * @pd : performance domain for which energy has to be estimated * @max_util : highest utilization among CPUs of the domain * @sum_util : sum of the utilization of all CPUs in the domain * * Return: the sum of the energy consumed by the CPUs of the domain assuming * a capacity state satisfying the max utilization of the domain. */ static inline unsigned long em_pd_energy(struct em_perf_domain *pd, unsigned long max_util, unsigned long sum_util) { unsigned long freq, scale_cpu; struct em_cap_state *cs; int i, cpu; if (!sum_util) return 0; /* * In order to predict the capacity state, map the utilization of the * most utilized CPU of the performance domain to a requested frequency, * like schedutil. */ cpu = cpumask_first(to_cpumask(pd->cpus)); scale_cpu = arch_scale_cpu_capacity(NULL, cpu); //获取cpu的max_capacity cs = &pd->table[pd->nr_cap_states - 1]; //获取capacity state，是为了获取最大频点（因为cs的table是升序排列的，所以最后一个配置就是最大的频点） freq = map_util_freq(max_util, cs->frequency, scale_cpu); //利用上面获取的最大频点、max_capacity，根据当前的cpu util映射到当前的cpu freq /* * Find the lowest capacity state of the Energy Model above the * requested frequency. */ for (i = 0; i < pd->nr_cap_states; i++) { //通过循环找到能满足当前cpu freq的最小的频点，及其对应的capacity state cs = &pd->table[i]; //同样因为cs的table是升序排列的，所以递增找到第一个满足的，就是满足条件的最小频点 if (cs->frequency >= freq) break; } /* * The capacity of a CPU in the domain at that capacity state (cs) * can be computed as: * * cs->freq * scale_cpu * cs->cap = -------------------- (1) * cpu_max_freq * * So, ignoring the costs of idle states (which are not available in * the EM), the energy consumed by this CPU at that capacity state is * estimated as: * * cs->power * cpu_util * cpu_nrg = -------------------- (2) * cs->cap * * since 'cpu_util / cs->cap' represents its percentage of busy time. * * NOTE: Although the result of this computation actually is in * units of power, it can be manipulated as an energy value * over a scheduling period, since it is assumed to be * constant during that interval. * * By injecting (1) in (2), 'cpu_nrg' can be re-expressed as a product * of two terms: * * cs->power * cpu_max_freq cpu_util * cpu_nrg = ------------------------ * --------- (3) * cs->freq scale_cpu * * The first term is static, and is stored in the em_cap_state struct * as 'cs->cost'. * * Since all CPUs of the domain have the same micro-architecture, they * share the same 'cs->cost', and the same CPU capacity. Hence, the * total energy of the domain (which is the simple sum of the energy of * all of its CPUs) can be factorized as: * * cs->cost * \Sum cpu_util * pd_nrg = ------------------------ (4) * scale_cpu */ return cs->cost * sum_util / scale_cpu; //通过上面的注释以及公式，推导出energy计算公式，并计算出perf doamin的总energy }
总结
1、find_best_target()函数主要是根据当前情况，找到task迁移的candidate cpu（target_cpu、backup cpu、prev_cpu）
具体逻辑：
prefer_idle：
best_idle_cpu:必须选择idle状态的cpu
---【task打开boost，选大核cpu】 && 【idle state最浅】
---【task没有打开boost，选小核cpu】 && 【idle state最浅】
target_cpu:必选选择ACTIVE状态的cpu
---【当前freq的cpu_capacity > 迁移task后的cpu_util】 && 【迁移task之后，剩余capacity最多的cpu】
best_active_cpu:必选选择ACTIVE状态的cpu
---【当前cpu_util更小的cpu】 && 【迁移task之后的cpu_util相等的话，选择cpu_util_cum + boosted_task_util】
normal：
该cpu的capacity_orig > start_cpu的capacity_orig（假如设置boost on big，只会往更大的cluster中寻找）
best_idle_cpu:必须选择idle状态的cpu
---【capacity相同情况下，idle最浅的cpu】 && 【选择cpu_util_cum + boosted_task_util中最小的】 && 【不能是prev_cpu】
非misfit task迁移的情况下，还要选出target_cpu
target_cpu:必选选择ACTIVE状态的cpu
---【迁移task之后，剩余capacity最多的cpu】】
2、在find_energy_efficient_cpu()后半段，计算task迁移到每个candidate cpu后的系统总energy。计算出的最小energy假如比prev_cpu的总energy少6%以上，那么这个cpu就是best_energy_cpu。
后续：在energy model与energy计算，目前还未弄清楚如何联系起来，后续需要找到如何联系。
还在成长。。。
