Skip to content

workqueue: don't migrate pending works from the dead CPU

Oleg Nesterov lkml 1 version New

Get series with b4
b4 am [email protected]

Versions newest first

  1. v1 Latest0 of 1 patch 0 repliesNew

Patches in v1 No applies recorded yet

The patch the cover letter announces is not in this archive.

Reviews trailers per patch

No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.

Read the v1 thread

Cover letter [PATCH 0/1] workqueue: don't migrate pending works from the dead CPU

Andrew, please drop the workqueue-rework-threads-hotplug-management.patch, and please take this one (feel free to just kill me instead).

We have workqueue callbacks (slab for example) which assume that CPU is stable during the execution since cwq->thread is pinned to a processor. With or without that patch this is not true with CONFIG_HOTPLUG_CPU, and the problem is not connected to workqueue.c.

However, my patch makes the things worse, because CPU_UP_PREPARE creates cwq->thread which is not pinned to CPU until CPU_ONLINE. This is solvable, but needs a serious complication, so this patch can't be considered as a cleanup.

The new patch removes take_over_work() and friends (as the old one did), but retains an unfortunate property of the current implementation: the workqueue_mutex is held from CPU_LOCK_ACQUIRE to CPU_LOCK_RELEASE, and CPU_DEAD stops cwq->thread synchronously.

This means that cpu-hotplug callbacks can't create/destroy workqueues (Gautham needs this?), and work->func() can't share a lock with cpu hotplug. May be we can change this later, but this needs a separate patch.

Srivatsa, do you have any objections?

The new patch is very similar (in fact, a subset), I don't send incremental update because this will break changelogs.

Oleg.