On Mon, Dec 23, 2024 at 08:05:53PM -0800, Yury Norov wrote:
> On Fri, Dec 20, 2024 at 04:11:39PM +0100, Andrea Righi wrote:
> > Using a single global idle mask can lead to inefficiencies and a lot of
> > stress on the cache coherency protocol on large systems with multiple
> > NUMA nodes, since all the CPUs can create a really intense read/write
> > activity on the single global cpumask.
> >
> > Therefore, split the global cpumask into multiple per-NUMA node cpumasks
> > to improve scalability and performance on large systems.
> >
> > The concept is that each cpumask will track only the idle CPUs within
> > its corresponding NUMA node, treating CPUs in other NUMA nodes as busy.
> > In this way concurrent access to the idle cpumask will be restricted
> > within each NUMA node.
> >
> > NOTE: if a scheduler enables the per-node idle cpumasks (via
> > SCX_OPS_BUILTIN_IDLE_PER_NODE), scx_bpf_get_idle_cpu/smtmask() will
> > trigger an scx error, since there are no system-wide cpumasks.
> >
> > By default (when SCX_OPS_BUILTIN_IDLE_PER_NODE is not enabled), only the
> > cpumask of node 0 is used as a single global flat CPU mask, maintaining
> > the previous behavior.
> >
> > Signed-off-by: Andrea Righi <arighi@xxxxxxxxxx>
>
> This is a rather big patch... Can you split it somehow? Maybe
> introduce new functions in a separate patch, and use them in the
> following patch(es)?
Yes, it's quite big unfortunately. I've tried to split it more, but it's
not trivial to have self-consistent changes that don't break things or
introduce unused functions...
>
> > ---
> > kernel/sched/ext.c | 7 +-
> > kernel/sched/ext_idle.c | 258 +++++++++++++++++++++++++++++++---------
> > 2 files changed, 208 insertions(+), 57 deletions(-)
> >
> > diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
> > index 148ec04d4a0a..143938e935f1 100644
> > --- a/kernel/sched/ext.c
> > +++ b/kernel/sched/ext.c
> > @@ -3228,7 +3228,7 @@ static void handle_hotplug(struct rq *rq, bool online)
> > atomic_long_inc(&scx_hotplug_seq);
> >
> > if (scx_enabled())
> > - update_selcpu_topology();
> > + update_selcpu_topology(&scx_ops);
> >
> > if (online && SCX_HAS_OP(cpu_online))
> > SCX_CALL_OP(SCX_KF_UNLOCKED, cpu_online, cpu);
> > @@ -5107,7 +5107,7 @@ static int scx_ops_enable(struct sched_ext_ops *ops, struct bpf_link *link)
> >
> > check_hotplug_seq(ops);
> > #ifdef CONFIG_SMP
> > - update_selcpu_topology();
> > + update_selcpu_topology(ops);
> > #endif
> > cpus_read_unlock();
> >
> > @@ -5800,8 +5800,7 @@ void __init init_sched_ext_class(void)
> >
> > BUG_ON(rhashtable_init(&dsq_hash, &dsq_hash_params));
> > #ifdef CONFIG_SMP
> > - BUG_ON(!alloc_cpumask_var(&idle_masks.cpu, GFP_KERNEL));
> > - BUG_ON(!alloc_cpumask_var(&idle_masks.smt, GFP_KERNEL));
> > + idle_masks_init();
> > #endif
> > scx_kick_cpus_pnt_seqs =
> > __alloc_percpu(sizeof(scx_kick_cpus_pnt_seqs[0]) * nr_cpu_ids,
> > diff --git a/kernel/sched/ext_idle.c b/kernel/sched/ext_idle.c
> > index 4952e2793304..444f2a15f1d4 100644
> > --- a/kernel/sched/ext_idle.c
> > +++ b/kernel/sched/ext_idle.c
> > @@ -10,7 +10,14 @@
> > * Copyright (c) 2024 Andrea Righi <arighi@xxxxxxxxxx>
> > */
> >
> > +/*
> > + * If NUMA awareness is disabled consider only node 0 as a single global
> > + * NUMA node.
> > + */
> > +#define NUMA_FLAT_NODE 0
>
> If it's a global idle node maybe
>
> #define GLOBAL_IDLE_NODE 0
>
> This actually bypasses NUMA, so it's weird to mention NUMA here.
Ok.
>
> > +
> > static DEFINE_STATIC_KEY_FALSE(scx_builtin_idle_enabled);
> > +static DEFINE_STATIC_KEY_FALSE(scx_builtin_idle_per_node);
> >
> > static bool check_builtin_idle_enabled(void)
> > {
> > @@ -22,22 +29,82 @@ static bool check_builtin_idle_enabled(void)
> > }
> >
> > #ifdef CONFIG_SMP
> > -#ifdef CONFIG_CPUMASK_OFFSTACK
> > -#define CL_ALIGNED_IF_ONSTACK
> > -#else
> > -#define CL_ALIGNED_IF_ONSTACK __cacheline_aligned_in_smp
> > -#endif
> > -
> > -static struct {
> > +struct idle_cpumask {
> > cpumask_var_t cpu;
> > cpumask_var_t smt;
> > -} idle_masks CL_ALIGNED_IF_ONSTACK;
> > +};
>
> We already have struct cpumask, and this struct idle_cpumask may
> mislead. Maybe struct idle_cpus or something?
Ok, I like struct idle_cpus.
>
> > +
> > +/*
> > + * cpumasks to track idle CPUs within each NUMA node.
> > + *
> > + * If SCX_OPS_BUILTIN_IDLE_PER_NODE is not specified, a single flat cpumask
> > + * from node 0 is used to track all idle CPUs system-wide.
> > + */
> > +static struct idle_cpumask **scx_idle_masks;
> > +
> > +static struct idle_cpumask *get_idle_mask(int node)
>
> Didn't we agree to drop this 'get' thing?
Hm... no? :)
The analogy you pointed out was with get_parity8() which implements a pure
function, so we should just use parity8() as "get" is implicit.
This case is a bit different IMHO, because it's getting a reference to the
object (so not a pure function) and we may even have a put_idle_mask()
potentially.
Personally I like to have the "get" here, but if you think it's confusing
or it makes the code less readable I'm ok to drop it.
>
> > +{
> > + if (node == NUMA_NO_NODE)
> > + node = numa_node_id();
> > + else if (WARN_ON_ONCE(node < 0 || node >= nr_node_ids))
> > + return NULL;
>
> Kernel users always provide correct parameters. I don't even think you
> need to check for NO_NODE, because if I as user of your API need to
> provide current node, I can use numa_node_id() just as well.
>
> If you drop all that sanity bloating, your function will be a
> one-liner, and the question is: do you need it at all?
>
> We usually need such wrappers to apply 'const' qualifier or do some
> housekeeping before dereferencing. But in this case you just return
> a pointer, and I don't understand why local users can't do it
> themself.
>
> The following idle_mask_init() happily ignores just added accessor...
Ok, makes sense. I'll drop the sanity checks here.
>
> > + return scx_idle_masks[node];
> > +}
>
> > +
> > +static struct cpumask *get_idle_cpumask(int node)
> > +{
> > + struct idle_cpumask *mask = get_idle_mask(node);
> > +
> > + return mask ? mask->cpu : cpu_none_mask;
> > +}
> > +
> > +static struct cpumask *get_idle_smtmask(int node)
> > +{
> > + struct idle_cpumask *mask = get_idle_mask(node);
> > +
> > + return mask ? mask->smt : cpu_none_mask;
> > +}
>
> For those two guys... I think you agreed with Tejun that you don't
> need them. To me the following is more verbose:
>
> idle_cpus(node)->smt;
Ok.
>
> > +
> > +static void idle_masks_init(void)
> > +{
> > + int node;
> > +
> > + scx_idle_masks = kcalloc(num_possible_nodes(), sizeof(*scx_idle_masks), GFP_KERNEL);
> > + BUG_ON(!scx_idle_masks);
> > +
> > + for_each_node_state(node, N_POSSIBLE) {
> > + scx_idle_masks[node] = kzalloc_node(sizeof(**scx_idle_masks), GFP_KERNEL, node);
> > + BUG_ON(!scx_idle_masks[node]);
> > +
> > + BUG_ON(!alloc_cpumask_var_node(&scx_idle_masks[node]->cpu, GFP_KERNEL, node));
> > + BUG_ON(!alloc_cpumask_var_node(&scx_idle_masks[node]->smt, GFP_KERNEL, node));
> > + }
> > +}
> >
> > static DEFINE_STATIC_KEY_FALSE(scx_selcpu_topo_llc);
> > static DEFINE_STATIC_KEY_FALSE(scx_selcpu_topo_numa);
> >
> > +/*
> > + * Return the node id associated to a target idle CPU (used to determine
> > + * the proper idle cpumask).
> > + */
> > +static int idle_cpu_to_node(int cpu)
> > +{
> > + int node;
> > +
> > + if (static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node))
> > + node = cpu_to_node(cpu);
>
> Nit: can you just return cpu_to_node(cpu). This will save 3 LOCs
Ok.
>
> > + else
> > + node = NUMA_FLAT_NODE;
> > +
> > + return node;
> > +}
> > +
> > static bool test_and_clear_cpu_idle(int cpu)
> > {
> > + int node = idle_cpu_to_node(cpu);
> > + struct cpumask *idle_cpus = get_idle_cpumask(node);
> > +
> > #ifdef CONFIG_SCHED_SMT
> > /*
> > * SMT mask should be cleared whether we can claim @cpu or not. The SMT
> > @@ -46,33 +113,37 @@ static bool test_and_clear_cpu_idle(int cpu)
> > */
> > if (sched_smt_active()) {
> > const struct cpumask *smt = cpu_smt_mask(cpu);
> > + struct cpumask *idle_smts = get_idle_smtmask(node);
> >
> > /*
> > * If offline, @cpu is not its own sibling and
> > * scx_pick_idle_cpu() can get caught in an infinite loop as
> > - * @cpu is never cleared from idle_masks.smt. Ensure that @cpu
> > - * is eventually cleared.
> > + * @cpu is never cleared from the idle SMT mask. Ensure that
> > + * @cpu is eventually cleared.
> > *
> > * NOTE: Use cpumask_intersects() and cpumask_test_cpu() to
> > * reduce memory writes, which may help alleviate cache
> > * coherence pressure.
> > */
> > - if (cpumask_intersects(smt, idle_masks.smt))
> > - cpumask_andnot(idle_masks.smt, idle_masks.smt, smt);
> > - else if (cpumask_test_cpu(cpu, idle_masks.smt))
> > - __cpumask_clear_cpu(cpu, idle_masks.smt);
> > + if (cpumask_intersects(smt, idle_smts))
> > + cpumask_andnot(idle_smts, idle_smts, smt);
> > + else if (cpumask_test_cpu(cpu, idle_smts))
> > + __cpumask_clear_cpu(cpu, idle_smts);
> > }
> > #endif
> > - return cpumask_test_and_clear_cpu(cpu, idle_masks.cpu);
> > + return cpumask_test_and_clear_cpu(cpu, idle_cpus);
> > }
> >
> > -static s32 scx_pick_idle_cpu(const struct cpumask *cpus_allowed, u64 flags)
> > +/*
> > + * Pick an idle CPU in a specific NUMA node.
> > + */
> > +static s32 pick_idle_cpu_from_node(const struct cpumask *cpus_allowed, int node, u64 flags)
> > {
> > int cpu;
> >
> > retry:
> > if (sched_smt_active()) {
> > - cpu = cpumask_any_and_distribute(idle_masks.smt, cpus_allowed);
> > + cpu = cpumask_any_and_distribute(get_idle_smtmask(node), cpus_allowed);
> > if (cpu < nr_cpu_ids)
> > goto found;
> >
> > @@ -80,15 +151,57 @@ static s32 scx_pick_idle_cpu(const struct cpumask *cpus_allowed, u64 flags)
> > return -EBUSY;
> > }
> >
> > - cpu = cpumask_any_and_distribute(idle_masks.cpu, cpus_allowed);
> > + cpu = cpumask_any_and_distribute(get_idle_cpumask(node), cpus_allowed);
> > if (cpu >= nr_cpu_ids)
> > return -EBUSY;
> >
> > found:
> > if (test_and_clear_cpu_idle(cpu))
> > return cpu;
> > - else
> > - goto retry;
> > + goto retry;
> > +}
>
> Yes, I see this too. But to me minimizing your patch and preserving as
> much history as you can is more important.
>
> After all, newcomers should have a room to practice :)
checkpatch.pl is bugging me. :)
But ok, makes sense, I'll ignore that for now.
>
> > +
> > +/*
> > + * Find the best idle CPU in the system, relative to @node.
> > + *
> > + * If @node is NUMA_NO_NODE, start from the current node.
> > + */
>
> And if you don't invent this rule for kernel users, you don't need to
> explain it everywhere.
I think we mentioned treating NUMA_NO_NODE as current node, but I might
have misunderstood. In an earlier patch set I was just ignoring
NUMA_NO_NODE. Should we return an error instead?
>
> > +static s32 scx_pick_idle_cpu(const struct cpumask *cpus_allowed, int node, u64 flags)
> > +{
> > + nodemask_t hop_nodes = NODE_MASK_NONE;
> > + s32 cpu = -EBUSY;
> > +
> > + if (!static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node))
> > + return pick_idle_cpu_from_node(cpus_allowed, NUMA_FLAT_NODE, flags);
> > +
> > + /*
> > + * If a NUMA node was not specified, start with the current one.
> > + */
> > + if (node == NUMA_NO_NODE)
> > + node = numa_node_id();
>
> And enforce too...
>
> > +
> > + /*
> > + * Traverse all nodes in order of increasing distance, starting
> > + * from prev_cpu's node.
> > + *
> > + * This loop is O(N^2), with N being the amount of NUMA nodes,
> > + * which might be quite expensive in large NUMA systems. However,
> > + * this complexity comes into play only when a scheduler enables
> > + * SCX_OPS_BUILTIN_IDLE_PER_NODE and it's requesting an idle CPU
> > + * without specifying a target NUMA node, so it shouldn't be a
> > + * bottleneck is most cases.
> > + *
> > + * As a future optimization we may want to cache the list of hop
> > + * nodes in a per-node array, instead of actually traversing them
> > + * every time.
> > + */
> > + for_each_numa_hop_node(n, node, hop_nodes, N_POSSIBLE) {
> > + cpu = pick_idle_cpu_from_node(cpus_allowed, n, flags);
> > + if (cpu >= 0)
> > + break;
> > + }
> > +
> > + return cpu;
> > }
> >
> > /*
> > @@ -208,7 +321,7 @@ static bool llc_numa_mismatch(void)
> > * CPU belongs to a single LLC domain, and that each LLC domain is entirely
> > * contained within a single NUMA node.
> > */
> > -static void update_selcpu_topology(void)
> > +static void update_selcpu_topology(struct sched_ext_ops *ops)
> > {
> > bool enable_llc = false, enable_numa = false;
> > unsigned int nr_cpus;
> > @@ -298,6 +411,7 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> > {
> > const struct cpumask *llc_cpus = NULL;
> > const struct cpumask *numa_cpus = NULL;
> > + int node = idle_cpu_to_node(prev_cpu);
> > s32 cpu;
> >
> > *found = false;
> > @@ -355,9 +469,9 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> > * piled up on it even if there is an idle core elsewhere on
> > * the system.
> > */
> > - if (!cpumask_empty(idle_masks.cpu) &&
> > - !(current->flags & PF_EXITING) &&
> > - cpu_rq(cpu)->scx.local_dsq.nr == 0) {
> > + if (!(current->flags & PF_EXITING) &&
> > + cpu_rq(cpu)->scx.local_dsq.nr == 0 &&
> > + !cpumask_empty(get_idle_cpumask(idle_cpu_to_node(cpu)))) {
> > if (cpumask_test_cpu(cpu, p->cpus_ptr))
> > goto cpu_found;
> > }
> > @@ -371,7 +485,7 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> > /*
> > * Keep using @prev_cpu if it's part of a fully idle core.
> > */
> > - if (cpumask_test_cpu(prev_cpu, idle_masks.smt) &&
> > + if (cpumask_test_cpu(prev_cpu, get_idle_smtmask(node)) &&
> > test_and_clear_cpu_idle(prev_cpu)) {
> > cpu = prev_cpu;
> > goto cpu_found;
> > @@ -381,7 +495,7 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> > * Search for any fully idle core in the same LLC domain.
> > */
> > if (llc_cpus) {
> > - cpu = scx_pick_idle_cpu(llc_cpus, SCX_PICK_IDLE_CORE);
> > + cpu = pick_idle_cpu_from_node(llc_cpus, node, SCX_PICK_IDLE_CORE);
> > if (cpu >= 0)
> > goto cpu_found;
> > }
> > @@ -390,15 +504,19 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> > * Search for any fully idle core in the same NUMA node.
> > */
> > if (numa_cpus) {
> > - cpu = scx_pick_idle_cpu(numa_cpus, SCX_PICK_IDLE_CORE);
> > + cpu = scx_pick_idle_cpu(numa_cpus, node, SCX_PICK_IDLE_CORE);
> > if (cpu >= 0)
> > goto cpu_found;
> > }
> >
> > /*
> > * Search for any full idle core usable by the task.
> > + *
> > + * If NUMA aware idle selection is enabled, the search will
> > + * begin in prev_cpu's node and proceed to other nodes in
> > + * order of increasing distance.
> > */
> > - cpu = scx_pick_idle_cpu(p->cpus_ptr, SCX_PICK_IDLE_CORE);
> > + cpu = scx_pick_idle_cpu(p->cpus_ptr, node, SCX_PICK_IDLE_CORE);
> > if (cpu >= 0)
> > goto cpu_found;
> > }
> > @@ -415,7 +533,7 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> > * Search for any idle CPU in the same LLC domain.
> > */
> > if (llc_cpus) {
> > - cpu = scx_pick_idle_cpu(llc_cpus, 0);
> > + cpu = pick_idle_cpu_from_node(llc_cpus, node, 0);
> > if (cpu >= 0)
> > goto cpu_found;
> > }
> > @@ -424,7 +542,7 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> > * Search for any idle CPU in the same NUMA node.
> > */
> > if (numa_cpus) {
> > - cpu = scx_pick_idle_cpu(numa_cpus, 0);
> > + cpu = pick_idle_cpu_from_node(numa_cpus, node, 0);
> > if (cpu >= 0)
> > goto cpu_found;
> > }
> > @@ -432,7 +550,7 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> > /*
> > * Search for any idle CPU usable by the task.
> > */
> > - cpu = scx_pick_idle_cpu(p->cpus_ptr, 0);
> > + cpu = scx_pick_idle_cpu(p->cpus_ptr, node, 0);
> > if (cpu >= 0)
> > goto cpu_found;
> >
> > @@ -448,17 +566,33 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> >
> > static void reset_idle_masks(void)
> > {
> > + int node;
> > +
> > + if (!static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) {
> > + cpumask_copy(get_idle_cpumask(NUMA_FLAT_NODE), cpu_online_mask);
> > + cpumask_copy(get_idle_smtmask(NUMA_FLAT_NODE), cpu_online_mask);
> > + return;
> > + }
> > +
> > /*
> > * Consider all online cpus idle. Should converge to the actual state
> > * quickly.
> > */
> > - cpumask_copy(idle_masks.cpu, cpu_online_mask);
> > - cpumask_copy(idle_masks.smt, cpu_online_mask);
> > + for_each_node_state(node, N_POSSIBLE) {
> > + const struct cpumask *node_mask = cpumask_of_node(node);
> > + struct cpumask *idle_cpu = get_idle_cpumask(node);
> > + struct cpumask *idle_smt = get_idle_smtmask(node);
> > +
> > + cpumask_and(idle_cpu, cpu_online_mask, node_mask);
> > + cpumask_copy(idle_smt, idle_cpu);
>
> Tejun asked you to use cpumask_and() in both cases, didn't he?
Ah! My bad, I missed this change during the last rebase. Will fix it.
-Andrea
