[PATCH v5 01/12] sched/isolation: Enforce nohz_full as a subset of isolcpus=domain at boot

[Qiliang Yuan]

HK_TYPE_KERNEL_NOISE and HK_TYPE_DOMAIN are configured independently
when nohz_full=/isolcpus=nohz and isolcpus=domain are passed as
separate boot parameters, or when nohz_full=/isolcpus=nohz is passed
without isolcpus=domain at all.  Nothing stops a CPU from ending up
tick-suppressed (nohz_full) while still scheduled as part of the
normal, non-isolated sched domain, which defeats the point of
isolating it from kernel noise in the first place.

A CPU with the tick stopped must always be excluded from the normal
scheduler domain: HK_TYPE_KERNEL_NOISE's housekeeping set has to stay
a superset of HK_TYPE_DOMAIN's.  Checking this while __setup()
parameters are still being parsed is order-dependent: rejecting
nohz_full= the moment it is seen, before a later isolcpus=domain on
the same command line has been parsed yet, would discard a
perfectly valid combination just because of argument order.

Check the invariant once in housekeeping_init() instead, which runs
after every __setup() cmdline parameter has been parsed regardless of
order.  Disable HK_TYPE_KERNEL_NOISE rather than panicking or leaving
the inconsistency in place: this gives the same end state as not
having passed nohz_full=/isolcpus=nohz at all.

housekeeping_cpumask() only dereferences the real per-type masks once
housekeeping_overridden is enabled; before that every call returns
cpu_possible_mask regardless of housekeeping.flags.  Enable the static
key before the subset check runs, not after: checking first would
compare cpu_possible_mask against itself for both sides and never
reject anything.

Signed-off-by: Qiliang Yuan <odys.yuan@xxxxxxxxx>
---
 Documentation/admin-guide/kernel-parameters.txt | 12 ++++++++++++
 kernel/sched/isolation.c                        | 24 ++++++++++++++++++++++++
 2 files changed, 36 insertions(+)

diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
index e75344f4e0cde..cd4fb6d752a6c 100644
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -2810,6 +2810,11 @@ Kernel parameters
 			  so to protect individual CPUs the 'cpumask' file has to
 			  be configured manually after bootup.
 
+			  When 'domain' is configured through a separate
+			  isolcpus= or nohz_full= invocation, every 'nohz'
+			  CPU must also be a 'domain' CPU; a combination that
+			  violates this is rejected at boot with a warning.
+
 			domain
 			  Isolate from the general SMP balancing and scheduling
 			  algorithms. Note that performing domain isolation this way
@@ -4564,6 +4569,13 @@ Kernel parameters
 			Note that this argument takes precedence over
 			the CONFIG_RCU_NOCB_CPU_DEFAULT_ALL option.
 
+			When isolcpus=domain is also given, every CPU in
+			this list (or in isolcpus=nohz) must also be in
+			the isolcpus=domain list: a tick-suppressed CPU
+			must always be excluded from the normal scheduler
+			domain.  A combination that violates this is
+			rejected at boot with a warning.
+
 	noinitrd	[Deprecated,RAM] Tells the kernel not to load any configured
 			initial RAM disk. Currently this parameter applies to
 			initrd only, not to initramfs. But it applies to both
diff --git a/kernel/sched/isolation.c b/kernel/sched/isolation.c
index 156025ef81b75..c6a41095a9002 100644
--- a/kernel/sched/isolation.c
+++ b/kernel/sched/isolation.c
@@ -171,8 +171,32 @@ void __init housekeeping_init(void)
 	if (!housekeeping.flags)
 		return;
 
+	/*
+	 * housekeeping_cpumask() only dereferences the real per-type masks
+	 * once housekeeping_overridden is live; before that it always
+	 * returns cpu_possible_mask regardless of housekeeping.flags, which
+	 * would make the subset check below vacuously pass for every type.
+	 * Enable it first so the check actually sees the parsed masks.
+	 */
 	static_branch_enable(&housekeeping_overridden);
 
+	/*
+	 * A nohz_full CPU must always run on an isolated sched domain:
+	 * reject nohz_full=/isolcpus=nohz unless isolcpus=domain covers
+	 * at least the same CPUs.  Checked here, after all __setup()
+	 * cmdline parsing has run, so the outcome does not depend on
+	 * which of the two parameters came first on the command line.
+	 */
+	if ((housekeeping.flags & HK_FLAG_KERNEL_NOISE) &&
+	    !cpumask_subset(housekeeping_cpumask(HK_TYPE_DOMAIN),
+			    housekeeping_cpumask(HK_TYPE_KERNEL_NOISE))) {
+		pr_warn("Housekeeping: nohz_full=/isolcpus=nohz must be a subset "
+			"of isolcpus=domain, disabling nohz_full\n");
+		housekeeping.flags &= ~HK_FLAG_KERNEL_NOISE;
+		if (!housekeeping.flags)
+			return;
+	}
+
 	if (housekeeping.flags & HK_FLAG_KERNEL_NOISE)
 		sched_tick_offload_init();
 	/*

-- 
2.43.0