scx_layered: Prefer preempting the CPU the task was previously on

Currently, when preempting, searching for the candidate CPU always starts
from the RR preemption cursor. Let's first try the previous CPU the
preempting task was on as that may have some locality benefits.
This commit is contained in:
Tejun Heo 2024-06-07 11:49:38 -10:00
parent 3db3257911
commit 7c48814ed0

View File

@ -463,11 +463,54 @@ s32 BPF_STRUCT_OPS(layered_select_cpu, struct task_struct *p, s32 prev_cpu, u64
}
}
static __attribute__((always_inline))
bool try_preempt(s32 cand, struct task_struct *p, struct cpu_ctx *cctx,
struct task_ctx *tctx, struct layer *layer)
{
struct cpu_ctx *cand_cctx, *sib_cctx = NULL;
s32 sib;
if (!bpf_cpumask_test_cpu(cand, p->cpus_ptr))
return false;
if (!(cand_cctx = lookup_cpu_ctx(cand)) || cand_cctx->current_preempt)
return false;
/*
* If exclusive, we want to make sure the sibling CPU, if there's
* one, is idle. However, if the sibling CPU is already running a
* preempt task, we shouldn't kick it out.
*/
if (layer->exclusive && (sib = sibling_cpu(cand)) >= 0 &&
(!(sib_cctx = lookup_cpu_ctx(sib)) || sib_cctx->current_preempt)) {
lstat_inc(LSTAT_EXCL_COLLISION, layer, cctx);
return false;
}
scx_bpf_kick_cpu(cand, SCX_KICK_PREEMPT);
/*
* $sib_cctx is set iff @p is an exclusive task, a sibling CPU
* exists which is not running a preempt task. Let's preempt the
* sibling CPU so that it can become idle. The ->maybe_idle test is
* inaccurate and racy but should be good enough for best-effort
* optimization.
*/
if (sib_cctx && !sib_cctx->maybe_idle) {
lstat_inc(LSTAT_EXCL_PREEMPT, layer, cctx);
scx_bpf_kick_cpu(sib, SCX_KICK_PREEMPT);
}
lstat_inc(LSTAT_PREEMPT, layer, cctx);
return true;
}
void BPF_STRUCT_OPS(layered_enqueue, struct task_struct *p, u64 enq_flags)
{
struct cpu_ctx *cctx;
struct task_ctx *tctx;
struct layer *layer;
s32 task_cpu = scx_bpf_task_cpu(p);
u64 vtime = p->scx.dsq_vtime;
u32 idx;
@ -514,7 +557,7 @@ void BPF_STRUCT_OPS(layered_enqueue, struct task_struct *p, u64 enq_flags)
if (!layer->open &&
(layer_cpumask = lookup_layer_cpumask(tctx->layer)) &&
!bpf_cpumask_test_cpu(scx_bpf_task_cpu(p), layer_cpumask))
!bpf_cpumask_test_cpu(task_cpu, layer_cpumask))
lstat_inc(LSTAT_AFFN_VIOL, layer, cctx);
scx_bpf_dispatch(p, HI_FALLBACK_DSQ, slice_ns, enq_flags);
@ -544,7 +587,7 @@ find_cpu:
if (!(enq_flags & SCX_ENQ_WAKEUP)) {
s32 cpu;
cpu = pick_idle_cpu(p, scx_bpf_task_cpu(p), cctx, tctx, layer);
cpu = pick_idle_cpu(p, task_cpu, cctx, tctx, layer);
if (cpu >= 0) {
lstat_inc(LSTAT_KICK, layer, cctx);
scx_bpf_kick_cpu(cpu, __COMPAT_SCX_KICK_IDLE);
@ -555,51 +598,20 @@ find_cpu:
if (!layer->preempt)
return;
bpf_for(idx, 0, nr_possible_cpus) {
struct cpu_ctx *cand_cctx, *sib_cctx = NULL;
u32 cpu = (preempt_cursor + idx) % nr_possible_cpus;
s32 sib;
if (!bpf_cpumask_test_cpu(cpu, p->cpus_ptr))
continue;
if (!(cand_cctx = lookup_cpu_ctx(cpu)) || cand_cctx->current_preempt)
continue;
/*
* If exclusive, we want to make sure the sibling CPU, if
* there's one, is idle. However, if the sibling CPU is already
* running a preempt task, we shouldn't kick it out.
*/
if (layer->exclusive && (sib = sibling_cpu(cpu)) >= 0 &&
(!(sib_cctx = lookup_cpu_ctx(sib)) ||
sib_cctx->current_preempt)) {
lstat_inc(LSTAT_EXCL_COLLISION, layer, cctx);
continue;
}
scx_bpf_kick_cpu(cpu, SCX_KICK_PREEMPT);
/*
* $sib_cctx is set iff @p is an exclusive task, a sibling CPU
* exists which is not running a preempt task. Let's preempt the
* sibling CPU so that it can become idle. The ->maybe_idle test
* is inaccurate and racy but should be good enough for
* best-effort optimization.
*/
if (sib_cctx && !sib_cctx->maybe_idle) {
lstat_inc(LSTAT_EXCL_PREEMPT, layer, cctx);
scx_bpf_kick_cpu(sib, SCX_KICK_PREEMPT);
}
/*
* Round-robining doesn't have to be strict. Let's not bother
* with atomic ops on $preempt_cursor.
*/
preempt_cursor = (cpu + 1) % nr_possible_cpus;
lstat_inc(LSTAT_PREEMPT, layer, cctx);
if (try_preempt(task_cpu, p, cctx, tctx, layer))
return;
bpf_for(idx, 0, nr_possible_cpus) {
s32 cand = (preempt_cursor + idx) % nr_possible_cpus;
if (try_preempt(cand, p, cctx, tctx, layer)) {
/*
* Round-robining doesn't have to be strict. Let's
* not bother with atomic ops on $preempt_cursor.
*/
preempt_cursor = (cand + 1) % nr_possible_cpus;
return;
}
}
lstat_inc(LSTAT_PREEMPT_FAIL, layer, cctx);