]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
IB/rdmavt: WQ_PERCPU added to alloc_workqueue users
authorMarco Crivellari <marco.crivellari@suse.com>
Sat, 1 Nov 2025 16:31:15 +0000 (17:31 +0100)
committerLeon Romanovsky <leon@kernel.org>
Thu, 6 Nov 2025 07:23:23 +0000 (02:23 -0500)
Currently if a user enqueue a work item using schedule_delayed_work() the
used wq is "system_wq" (per-cpu wq) while queue_delayed_work() use
WORK_CPU_UNBOUND (used when a cpu is not specified). The same applies to
schedule_work() that is using system_wq and queue_work(), that makes use
again of WORK_CPU_UNBOUND.
This lack of consistentcy cannot be addressed without refactoring the API.

alloc_workqueue() treats all queues as per-CPU by default, while unbound
workqueues must opt-in via WQ_UNBOUND.

This default is suboptimal: most workloads benefit from unbound queues,
allowing the scheduler to place worker threads where they’re needed and
reducing noise when CPUs are isolated.

This change adds a new WQ_PERCPU flag to explicitly request
alloc_workqueue() to be per-cpu when WQ_UNBOUND has not been specified.

With the introduction of the WQ_PERCPU flag (equivalent to !WQ_UNBOUND),
any alloc_workqueue() caller that doesn’t explicitly specify WQ_UNBOUND
must now use WQ_PERCPU.

Once migration is complete, WQ_UNBOUND can be removed and unbound will
become the implicit default.

CC: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Marco Crivellari <marco.crivellari@suse.com>
Link: https://patch.msgid.link/20251101163121.78400-6-marco.crivellari@suse.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
drivers/infiniband/sw/rdmavt/cq.c

index 0ca2743f107524b84a501f3e69d06f4528c4ef98..e7835ca70e2b21493080334f4544242161ef68d0 100644 (file)
@@ -518,7 +518,8 @@ int rvt_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
  */
 int rvt_driver_cq_init(void)
 {
-       comp_vector_wq = alloc_workqueue("%s", WQ_HIGHPRI | WQ_CPU_INTENSIVE,
+       comp_vector_wq = alloc_workqueue("%s",
+                                        WQ_HIGHPRI | WQ_CPU_INTENSIVE | WQ_PERCPU,
                                         0, "rdmavt_cq");
        if (!comp_vector_wq)
                return -ENOMEM;