Merge pull request #1099 from hodgesds/layered-namespace-aware

scx_layered: Make layered work in pid namespaces

scheds/include/scx/namespace.bpf.h

@@ -5,11 +5,16 @@
 #ifndef __SCHED_EXT_NAMESPACE_BPF_H
 #define __SCHED_EXT_NAMESPACE_BPF_H
 
+#ifdef LSP
+#define __bpf__
+#include "../vmlinux.h"
+#else
 #include "vmlinux.h"
+#endif
 
-struct pid_namespace* get_task_pid_ns(const struct task_struct* task);
+struct pid_namespace* get_task_pid_ns(const struct task_struct* task, enum pid_type);
 struct pid* get_task_pid_ptr(const struct task_struct* task, enum pid_type type);
-pid_t get_task_ns_pid(const struct task_struct* task, enum pid_type type);
+pid_t get_task_ns_pid(const struct task_struct* task);
 
 pid_t get_pid_nr_ns(struct pid* pid, struct pid_namespace* ns);
 pid_t get_ns_pid(void);

scheds/rust/scx_layered/src/bpf/main.bpf.c

@@ -25,7 +25,6 @@ char _license[] SEC("license") = "GPL";
 extern unsigned CONFIG_HZ __kconfig;
 
 const volatile u32 debug;
-const volatile s32 layered_tgid;
 const volatile u64 slice_ns;
 const volatile u64 max_exec_ns;
 const volatile u32 nr_cpu_ids = 1;
@@ -59,6 +58,7 @@ private(big_cpumask) struct bpf_cpumask __kptr *big_cpumask;
 struct layer layers[MAX_LAYERS];
 u32 fallback_cpu;
 static u32 preempt_cursor;
+u32 layered_root_tgid = 0;
 
 u32 empty_layer_ids[MAX_LAYERS];
 u32 nr_empty_layer_ids;
@@ -143,6 +143,11 @@ static inline bool is_fb_dsq(u64 dsq_id)
 	return dsq_id & (HI_FB_DSQ_BASE | LO_FB_DSQ_BASE);
 }
 
+static __always_inline bool is_scheduler_task(struct task_struct *p)
+{
+	return (u32)p->tgid == layered_root_tgid;
+}
+
 struct {
 	__uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
 	__type(key, u32);
@@ -485,6 +490,23 @@ int BPF_PROG(tp_task_rename, struct task_struct *p, const char *buf)
 	return 0;
 }
 
+/*
+ * Initializes the scheduler to support running in a pid namespace.
+ */
+SEC("syscall")
+int BPF_PROG(initialize_pid_namespace)
+{
+	struct task_struct *p;
+
+	if (!(p = (struct task_struct*)bpf_get_current_task_btf()))
+		return -ENOENT;
+
+	layered_root_tgid = BPF_PROBE_READ(p, tgid);
+	trace("CFG layered running with tgid: %d", layered_root_tgid);
+
+	return 0;
+}
+
 static bool should_refresh_cached_cpus(struct cached_cpus *ccpus, s64 id, u64 cpus_seq)
 {
 	return ccpus->id != id || ccpus->seq != cpus_seq;
@@ -1081,7 +1103,7 @@ void BPF_STRUCT_OPS(layered_enqueue, struct task_struct *p, u64 enq_flags)
 	 * are usually important for system performance and responsiveness.
 	 */
 	if (((p->flags & PF_KTHREAD) && p->nr_cpus_allowed < nr_possible_cpus) ||
-	    p->tgid == layered_tgid) {
+	    is_scheduler_task(p)) {
 		struct cpumask *layer_cpumask;
 
 		if (layer->kind == LAYER_KIND_CONFINED &&

scheds/rust/scx_layered/src/main.rs

@@ -1568,9 +1568,6 @@ impl<'a> Scheduler<'a> {
         skel.struct_ops.layered_mut().exit_dump_len = opts.exit_dump_len;
 
         skel.maps.rodata_data.debug = opts.verbose as u32;
-        // Running scx_layered inside a PID namespace would break the
-        // following.
-        skel.maps.rodata_data.layered_tgid = std::process::id() as i32;
         skel.maps.rodata_data.slice_ns = opts.slice_us * 1000;
         skel.maps.rodata_data.max_exec_ns = if opts.max_exec_us > 0 {
             opts.max_exec_us * 1000
@@ -1637,6 +1634,13 @@ impl<'a> Scheduler<'a> {
         // Other stuff.
         let proc_reader = procfs::ProcReader::new();
 
+        // Handle setup if layered is running in a pid namespace.
+        let input = ProgramInput {
+            ..Default::default()
+        };
+        let prog = &mut skel.progs.initialize_pid_namespace;
+        let _ = prog.test_run(input);
+
         // XXX If we try to refresh the cpumasks here before attaching, we
         // sometimes (non-deterministically) don't see the updated values in
         // BPF. It would be better to update the cpumasks here before we