JakeHillion/scx
1
0
Fork 0
mirror of https://github.com/JakeHillion/scx.git synced 2024-11-30 21:10:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #92 from sched-ext/nest_callbacks

Browse Source 
scx_nest: Set timer callback after cancelling
This commit is contained in:
Tejun Heo 2024-01-17 09:27:22 -10:00 committed by GitHub
commit 9089cc09bb
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 45 additions and 53 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

92
scheds/c/scx_nest.bpf.c
View File

@ -107,9 +107,6 @@ struct pcpu_ctx {
/* Whether the current core has been scheduled for compaction. */
bool scheduled_compaction;
/* Number of times a primary core has been scheduled for compaction. */
u32 num_schedulings;
};
struct {
@ -197,6 +194,39 @@ static void update_attached(struct task_ctx *tctx, s32 prev_cpu, s32 new_cpu)
tctx->prev_cpu = prev_cpu;
}
static int compact_primary_core(void *map, int *key, struct bpf_timer *timer)
{
struct bpf_cpumask *primary, *reserve;
s32 cpu = bpf_get_smp_processor_id();
struct pcpu_ctx *pcpu_ctx;
stat_inc(NEST_STAT(CALLBACK_COMPACTED));
/*
* If we made it to this callback, it means that the timer callback was
* never cancelled, and so the core needs to be demoted from the
* primary nest.
*/
pcpu_ctx = bpf_map_lookup_elem(&pcpu_ctxs, &cpu);
if (!pcpu_ctx) {
scx_bpf_error("Couldn't lookup pcpu ctx");
return 0;
}
bpf_rcu_read_lock();
primary = primary_cpumask;
reserve = reserve_cpumask;
if (!primary || !reserve) {
scx_bpf_error("Couldn't find primary or reserve");
bpf_rcu_read_unlock();
return 0;
}
bpf_cpumask_clear_cpu(cpu, primary);
try_make_core_reserved(cpu, reserve, false);
bpf_rcu_read_unlock();
pcpu_ctx->scheduled_compaction = false;
return 0;
}
s32 BPF_STRUCT_OPS(nest_select_cpu, struct task_struct *p, s32 prev_cpu,
u64 wake_flags)
{
@ -344,6 +374,8 @@ migrate_primary:
if (pcpu_ctx->scheduled_compaction) {
if (bpf_timer_cancel(&pcpu_ctx->timer) < 0)
scx_bpf_error("Failed to cancel pcpu timer");
if (bpf_timer_set_callback(&pcpu_ctx->timer, compact_primary_core))
scx_bpf_error("Failed to re-arm pcpu timer");
pcpu_ctx->scheduled_compaction = false;
stat_inc(NEST_STAT(CANCELLED_COMPACTION));
}
@ -425,10 +457,8 @@ void BPF_STRUCT_OPS(nest_dispatch, s32 cpu, struct task_struct *prev)
stat_inc(NEST_STAT(NOT_CONSUMED));
if (in_primary) {
/*
* Immediately demote a primary core if:
* - It's been scheduled for compaction at least
* r_depth times without actually being compacted.
* - The previous task on it is dying
* Immediately demote a primary core if the previous
* task on it is dying
*
* Note that we elect to not compact the "first" CPU in
* the mask so as to encourage at least one core to
@ -437,12 +467,11 @@ void BPF_STRUCT_OPS(nest_dispatch, s32 cpu, struct task_struct *prev)
* nest, but BPF doesn't yet have a kfunc for querying
* cpumask weight.
*/
if ((prev && prev->__state == TASK_DEAD) ||
(cpu != bpf_cpumask_first(cast_mask(primary)) && pcpu_ctx->num_schedulings >= r_depth)) {
if ((prev && prev->__state == TASK_DEAD) &&
(cpu != bpf_cpumask_first(cast_mask(primary)))) {
stat_inc(NEST_STAT(EAGERLY_COMPACTED));
bpf_cpumask_clear_cpu(cpu, primary);
try_make_core_reserved(cpu, reserve, false);
pcpu_ctx->num_schedulings = 0;
} else {
pcpu_ctx->scheduled_compaction = true;
/*
@ -452,14 +481,12 @@ void BPF_STRUCT_OPS(nest_dispatch, s32 cpu, struct task_struct *prev)
*/
bpf_timer_start(&pcpu_ctx->timer, p_remove_ns,
BPF_F_TIMER_CPU_PIN);
pcpu_ctx->num_schedulings++;
stat_inc(NEST_STAT(SCHEDULED_COMPACTION));
}
}
return;
}
stat_inc(NEST_STAT(CONSUMED));
pcpu_ctx->num_schedulings = 0;
}
void BPF_STRUCT_OPS(nest_running, struct task_struct *p)
@ -514,40 +541,6 @@ void BPF_STRUCT_OPS(nest_enable, struct task_struct *p)
p->scx.dsq_vtime = vtime_now;
}
static int compact_primary_core(void *map, int *key, struct bpf_timer *timer)
{
struct bpf_cpumask *primary, *reserve;
s32 cpu = bpf_get_smp_processor_id();
struct pcpu_ctx *pcpu_ctx;
stat_inc(NEST_STAT(CALLBACK_COMPACTED));
/*
* If we made it to this callback, it means that the timer callback was
* never cancelled, and so the core needs to be demoted from the
* primary nest.
*/
pcpu_ctx = bpf_map_lookup_elem(&pcpu_ctxs, &cpu);
if (!pcpu_ctx) {
scx_bpf_error("Couldn't lookup pcpu ctx");
return 0;
}
bpf_rcu_read_lock();
primary = primary_cpumask;
reserve = reserve_cpumask;
if (!primary || !reserve) {
scx_bpf_error("Couldn't find primary or reserve");
bpf_rcu_read_unlock();
return 0;
}
bpf_cpumask_clear_cpu(cpu, primary);
try_make_core_reserved(cpu, reserve, false);
bpf_rcu_read_unlock();
pcpu_ctx->num_schedulings = 0;
pcpu_ctx->scheduled_compaction = false;
return 0;
}
static int stats_timerfn(void *map, int *key, struct bpf_timer *timer)
{
s32 cpu;
@ -636,8 +629,11 @@ s32 BPF_STRUCT_OPS_SLEEPABLE(nest_init)
scx_bpf_error("Failed to initialize pcpu timer");
return -EINVAL;
}
ctx->num_schedulings = 0;
bpf_timer_set_callback(&ctx->timer, compact_primary_core);
err = bpf_timer_set_callback(&ctx->timer, compact_primary_core);
if (err) {
scx_bpf_error("Failed to set pcpu timer callback");
return -EINVAL;
}
}
timer = bpf_map_lookup_elem(&stats_timer, &key);

6
scheds/c/scx_nest.c
View File

@ -28,7 +28,6 @@ const char help_fmt[] =
" -m R_MAX Maximum number of cores in the reserve nest (default 5)\n"
" -i ITERS Number of successive placement failures tolerated before trying to aggressively expand primary nest (default 2), or 0 to disable\n"
" -s SLICE_US Override slice duration in us (default 20000us / 20ms)\n"
" -D R_SCHED Override the number of times that a core may be scheduled for compaction before having compaction happen immediately (default 5)\n"
" -I First try to find a fully idle core, and then any idle core, when searching nests. Default behavior is to ignore hypertwins and check for any idle core.\n"
" -h Display this help and exit\n";
@ -165,14 +164,11 @@ int main(int argc, char **argv)
skel->rodata->nr_cpus = libbpf_num_possible_cpus();
skel->rodata->sampling_cadence_ns = SAMPLING_CADENCE_S * 1000 * 1000 * 1000;
while ((opt = getopt(argc, argv, "hId:D:m:i:s:")) != -1) {
while ((opt = getopt(argc, argv, "hId:m:i:s:")) != -1) {
switch (opt) {
case 'd':
skel->rodata->p_remove_ns = strtoull(optarg, NULL, 0) * 1000;
break;
case 'D':
skel->rodata->r_depth = strtoull(optarg, NULL, 0);
break;
case 'm':
skel->rodata->r_max = strtoull(optarg, NULL, 0);
break;