Re: [PATCH v4] lib/group_cpus: Snapshot cluster masks to keep grouping hotplug invariant

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 26 Aug 2026 22:45:37 +0530 Vishal Badole <Vishal.Badole@amd.com> wrote:

> group_cpus_evenly() builds the managed-IRQ affinity spread used by
> multi-queue devices such as NVMe. That spread is meant to be a property
> of the static CPU topology: it walks cpu_present_mask and then
> cpu_possible_mask so every hardware queue owns a fixed set of CPUs,
> including CPUs that are offline at the time. A driver depends on that
> partition staying stable across re-computation - the CPUs a queue is
> given at probe must still describe the same queue after the device is
> later reset and its affinity recomputed.
> 
> On an AMD system that stability breaks across an s2idle cycle. With CPUs
> 3-11 offlined and only CPUs 0-2 left online, the machine is suspended to
> s2idle and resumed. The NVMe controller uses the simple-suspend quirk, so
> resume fully re-initialises it and recomputes the affinity spread. The
> system then hangs for roughly two minutes and stays sluggish afterwards,
> the controller only making progress through its command-timeout poll:
> 
>   nvme nvme0: I/O tag 898 (3382) QID 9 timeout, completion polled
>   nvme nvme0: I/O tag 398 (618e) QID 11 timeout, completion polled
> 
> QID 9 and QID 11 are the queues whose CPUs were offline when the spread
> was recomputed. "completion polled" means the commands did finish in
> hardware, but their interrupts were never delivered to a CPU that was
> watching the queue, so nothing reaped them until the timeout fired.
> 
> It happens because commit 89802ca36c96 ("lib/group_cpus: make group CPU
> cluster aware") derives the cluster groups from topology_cluster_cpumask(),
> which lists only the cluster siblings that are online when it is called.
> The resulting partition therefore depends on the transient online mask
> rather than on the topology alone. Recomputed on resume while the non-boot
> CPUs are still offline, it no longer matches the boot-time partition, and a
> queue is left with an affinity that does not cover the CPU it is meant to
> serve once that CPU comes back online. The dependence is on the online
> mask, not on any AMD-specific behaviour, so the same stall is reproducible
> on Intel platforms as well.
> 
> Make the cluster grouping depend on the complete cluster topology rather
> than on whichever CPUs happen to be online. Snapshot the cluster masks
> once while every present CPU is online and reuse that view for every
> later spread. Every spread then groups from the same masks, so the
> partition computed when the controller is reset matches the one computed
> at probe and each queue's IRQ still covers the CPUs it serves. If the
> snapshot was never taken, the cluster path is skipped and the plain
> present/possible spread is used.

I fed this into my evolving AI reviewbot and it had a few things to
say:


The snapshot helps when the first affinity spread is computed with all
present CPUs online, but I don't think it guarantees the invariant
claimed in the changelog.

If the first spread happens while some present CPUs are offline,
capture_cluster_snapshot() declines to run and we use plain grouping.
When those CPUs come online a later spread captures the clusters and
changes the grouping algorithm. For clusters {0,2} and {1,3}, two IRQ
groups can change from {0,1}/{2,3} to {0,2}/{1,3}.

Also, even after the snapshot, grp_spread_init_one() still consults
live topology_sibling_cpumask(). x86 CPU offlining changes these masks.
On an L2 cluster containing multiple SMT cores, assigning multiple
IRQ groups inside that cluster can therefore change across hotplug.

There's also a new failure mode: if the first snapshot allocation or
cpus_read_trylock() fails, we use plain grouping, then switch to
cluster grouping after a subsequent successful capture. I can produce
different masks this way with all CPUs remaining online.

I've verified the mask differences in reduced GCC and Clang models,
not in a kernel/NVMe reproducer. Can we make the entire assignment
independent of online sibling masks and avoid switching algorithms
after an affinity spread has already been handed to a device?