I stumbled upon "isolcpus=managed_irq" which is the last piece which
can only be handled by isolcpus= and has no runtime knob. I knew roughly
what managed interrupts should do but I lacked some details how it is
used and what the managed_irq sub parameter means in practise.

This documents what we have as of today and how it works. I added some
examples how the parameter affects the configuration. Did I miss
something?

Given that the spreading as computed group_cpus_evenly() does not take
the mask of isolated CPUs into account I'm not sure how relevant the
managed_irq argument is. The virtio_scsi driver has no way to limit the
interrupts and I don't see this for the nvme. Even if the number of
queues can be reduced to two (as in the example) it is still spread
evenly in the system instead and the isolated CPUs are not taken into
account.
To make this worse, you can even argue further whether or not the
application on the isolated CPU wants to receive the interrupt directly
or would prefer not to.

Given all this, I am not sure if it makes sense to add 'io_queue' to the
mix or if it could be incorporated into 'managed_irq'.

One more point: Given that isolcpus= is marked deprecated as of commit
   b0d40d2b22fe4 ("sched/isolation: Document isolcpus= boot parameter flags, mark it deprecated")

and the 'managed_irq' is evaluated at device's probe time it would
require additional callbacks to re-evaluate the situation. Probably for
'io_queue', too. Does is make sense or should we simply drop the
"deprecation" notice and allowing using it long term?
Dynamic partitions work with cpusets, there this (managed_irq)
limitation but is it really? And if static partition is the use case why
bother.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 Documentation/core-api/irq/index.rst       |   1 +
 Documentation/core-api/irq/managed_irq.rst | 116 +++++++++++++++++++++
 2 files changed, 117 insertions(+)
 create mode 100644 Documentation/core-api/irq/managed_irq.rst

diff --git a/Documentation/core-api/irq/index.rst b/Documentation/core-api/irq/index.rst
index 0d65d11e54200..13bd24dd2b1cc 100644
--- a/Documentation/core-api/irq/index.rst
+++ b/Documentation/core-api/irq/index.rst
@@ -9,3 +9,4 @@ IRQs
    irq-affinity
    irq-domain
    irqflags-tracing
+   managed_irq
diff --git a/Documentation/core-api/irq/managed_irq.rst b/Documentation/core-api/irq/managed_irq.rst
new file mode 100644
index 0000000000000..05e295f3c289d
--- /dev/null
+++ b/Documentation/core-api/irq/managed_irq.rst
@@ -0,0 +1,116 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+===========================
+Affinity managed interrupts
+===========================
+
+The IRQ core provides support for managing interrupts according to a specified
+CPU affinity. Under normal operation, an interrupt is associated with a
+particular CPU. If that CPU is taken offline, the interrupt is migrated to
+another online CPU.
+
+Devices with large numbers of interrupt vectors can stress the available vector
+space. For example, an NVMe device with 128 I/O queues typically requests one
+interrupt per queue on systems with at least 128 CPUs. Two such devices
+therefore request 256 interrupts. On x86, the interrupt vector space is
+notoriously low, providing only 256 vectors per CPU, and the kernel reserves a
+subset of these, further reducing the number available for device interrupts.
+In practice this is not an issue because the interrupts are distributed across
+many CPUs, so each CPU only receives a small number of vectors.
+
+During system suspend, however, all secondary CPUs are taken offline and all
+interrupts are migrated to the single CPU that remains online. This can exhaust
+the available interrupt vectors on that CPU and cause the suspend operation to
+fail.
+
+Affinity‑managed interrupts address this limitation. Each interrupt is assigned
+a CPU affinity mask that specifies the set of CPUs on which the interrupt may
+be targeted. When a CPU in the mask goes offline, the interrupt is moved to the
+next CPU in the mask. If the last CPU in the mask goes offline, the interrupt
+is shut down. Drivers using affinity‑managed interrupts must ensure that the
+associated queue is quiesced before the interrupt is disabled so that no
+further interrupts are generated. When a CPU in the affinity mask comes back
+online, the interrupt is re‑enabled.
+
+Implementation
+--------------
+
+Devices must provide per‑instance interrupts, such as per‑I/O‑queue interrupts
+for storage devices like NVMe. The driver allocates interrupt vectors with the
+required affinity settings using struct irq_affinity. For MSI‑X devices, this
+is done via pci_alloc_irq_vectors_affinity() with the PCI_IRQ_AFFINITY flag
+set.
+
+Based on the provided affinity information, the IRQ core attempts to spread the
+interrupts evenly across the system. The affinity masks are computed during
+this allocation step, but the final IRQ assignment is performed when
+request_irq() is invoked.
+
+Isolated CPUs
+-------------
+
+The affinity of managed interrupts is handled entirely in the kernel and cannot
+be modified from user space through the /proc interfaces. The managed_irq
+sub‑parameter of the isolcpus boot option specifies a CPU mask that managed
+interrupts should attempt to avoid. This isolation is best‑effort and only
+applies if the automatically assigned interrupt mask also contains online CPUs
+outside the avoided mask. If the requested mask contains only isolated CPUs,
+the setting has no effect.
+
+CPUs listed in the avoided mask remain part of the interrupt’s affinity mask.
+This means that if all non‑isolated CPUs go offline while isolated CPUs remain
+online, the interrupt will be assigned to one of the isolated CPUs.
+
+The following examples assume a system with 8 CPUs.
+
+- A QEMU instance is booted with "-device virtio-scsi-pci".
+  The MSI‑X device exposes 11 interrupts: 3 "management" interrupts and 8
+  "queue" interrupts. The driver requests the 8 queue interrupts, each of which
+  is affine to exactly one CPU. If that CPU goes offline, the interrupt is shut
+  down.
+
+  Assuming interrupt 48 is one of the queue interrupts, the following appears::
+
+    /proc/irq/48/effective_affinity_list:7
+    /proc/irq/48/smp_affinity_list:7
+
+  This indicates that the interrupt is served only by CPU7. Shutting down CPU7
+  does not migrate the interrupt to another CPU::
+
+    /proc/irq/48/effective_affinity_list:0
+    /proc/irq/48/smp_affinity_list:7
+
+  This can be verified via the debugfs interface
+  (/sys/kernel/debug/irq/irqs/48). The dstate field will include
+  IRQD_IRQ_DISABLED, IRQD_IRQ_MASKED and IRQD_MANAGED_SHUTDOWN.
+
+- A QEMU instance is booted with "-device virtio-scsi-pci,num_queues=2"
+  and the kernel command line includes:
+  "irqaffinity=0,1 isolcpus=domain,2-7 isolcpus=managed_irq,1-3,5-7".
+  The MSI‑X device exposes 5 interrupts: 3 management interrupts and 2 queue
+  interrupts. The management interrupts follow the irqaffinity= setting. The
+  queue interrupts are spread across available CPUs::
+
+    /proc/irq/47/effective_affinity_list:0
+    /proc/irq/47/smp_affinity_list:0-3
+    /proc/irq/48/effective_affinity_list:4
+    /proc/irq/48/smp_affinity_list:4-7
+
+  The two queue interrupts are evenly distributed. Interrupt 48 is placed on CPU4
+  because the managed_irq mask avoids CPUs 5–7 when possible.
+
+  Replacing the managed_irq argument with "isolcpus=managed_irq,1-3,4-5,7"
+  results in::
+
+    /proc/irq/48/effective_affinity_list:6
+    /proc/irq/48/smp_affinity_list:4-7
+
+  Interrupt 48 is now served on CPU6 because the system avoids CPUs 4, 5 and
+  7. If CPU6 is taken offline, the interrupt migrates to one of the "isolated"
+  CPUs::
+
+    /proc/irq/48/effective_affinity_list:7
+    /proc/irq/48/smp_affinity_list:4-7
+
+  The interrupt is shut down once all CPUs listed in its smp_affinity mask are
+  offline.
-- 
2.53.0