On 6/10/21 12:29 AM, Cyrill Gorcunov wrote:

On Wed, Jun 09, 2021 at 01:38:58PM +0200, Vlastimil Babka wrote:

quoted

+static DEFINE_MUTEX(flush_lock);
+static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
+
 static void flush_all(struct kmem_cache *s)
 {
-	on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
+	struct slub_flush_work *sfw;
+	unsigned int cpu;
+
+	cpus_read_lock();
+	mutex_lock(&flush_lock);
+

Hi, Vlastimil! Could you please point why do you lock cpus first and
mutex only after? Why not mutex_lock + cpus_read_lock instead?

Good question! I must admit I didn't think about it much and just followed the
order that was in the original Sebastian's patch [1]
But there was a good reason for this order as some paths via
__kmem_cache_shutdown() and __kmem_cache_shrink() were alreadu called under
cpus_read_lock. Meanwhile mainline (me, actually) removed those, so now it
doesn't seem to be a need to keep this order anymore and we could switch it.

Thanks,
Vlastimil

[1]
https://git.kernel.org/pub/scm/linux/kernel/git/rt/linux-rt-devel.git/tree/patches/0005-mm-slub-Move-flush_cpu_slab-invocations-__free_slab-.patch?h=linux-5.12.y-rt-patches