Re: [rfc/patch] mm/slub: restore/expand unfreeze_partials() local exclusion scope

[Vlastimil Babka <vbabka@suse.cz>]

On 7/25/21 6:27 PM, Vlastimil Babka wrote:
>>> Hope fixing that helps then?
>>
>> Yeah, though RT should perhaps be pinned across release/re-acquire?
> 
> Probably not necessary, this_cpu_cmpxchg() will effectively recognize
> being moved to a different CPU.
> Might also move __unfreeze_partials() out of the whole loop to avoid the
> relock. Yeah that should be better.
> 
>> Actually, local locks should rediscover the recursion handling skills
>> they long had so such RT specific hole poking isn't necessary.  There
>> previously would have been no ifdef+typo there for eyeballs to miss and
>> miss and miss.
> 
> Hm, now I'm realizing that local_lock() on !RT is just
> preempt_disable(), i.e. equivalent to get_cpu_ptr(), so some of the
> ifdeffery could go away?

How much will this explode? Thanks.

----8<----
From 99808b198bdf867951131bb9d1ca1bd1cd12b8c4 Mon Sep 17 00:00:00 2001
From: Vlastimil Babka <vbabka@suse.cz>
Date: Fri, 23 Jul 2021 23:17:18 +0200
Subject: [PATCH] PREEMPT_RT+SLUB_CPU_PARTIAL fix attempt

---
 mm/slub.c | 34 +++++++++++++++++++++++++++++++---
 1 file changed, 31 insertions(+), 3 deletions(-)

diff --git a/mm/slub.c b/mm/slub.c
index 581004a5aca9..d12a50b5ee6f 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -2437,13 +2437,19 @@ static void unfreeze_partials(struct kmem_cache *s)
 {
 	struct page *partial_page;
 
+#ifndef CONFIG_PREEMPT_RT
 	do {
 		partial_page = this_cpu_read(s->cpu_slab->partial);
 
 	} while (partial_page &&
 		 this_cpu_cmpxchg(s->cpu_slab->partial, partial_page, NULL)
 				  != partial_page);
-
+#else
+	local_lock(&s->cpu_slab->lock);
+	partial_page = this_cpu_read(s->cpu_slab->partial);
+	this_cpu_write(s->cpu_slab->partial, NULL);
+	local_unlock(&s->cpu_slab->lock);
+#endif
 	if (partial_page)
 		__unfreeze_partials(s, partial_page);
 }
@@ -2479,10 +2485,15 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
 {
 #ifdef CONFIG_SLUB_CPU_PARTIAL
 	struct page *oldpage;
+	struct page *page_to_unfreeze = NULL;
 	int pages;
 	int pobjects;
 
-	slub_get_cpu_ptr(s->cpu_slab);
+	/*
+	 * On !RT we just want to disable preemption, on RT we need the lock
+	 * for real. This happens to match local_lock() semantics.
+	 */
+	local_lock(&s->cpu_slab->lock);
 	do {
 		pages = 0;
 		pobjects = 0;
@@ -2496,7 +2507,16 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
 				 * partial array is full. Move the existing
 				 * set to the per node partial list.
 				 */
+#ifndef CONFIG_PREEMPT_RT
 				unfreeze_partials(s);
+#else
+				/*
+				 * Postpone unfreezing until we drop the local
+				 * lock to avoid relocking.
+				 */
+				page_to_unfreeze = oldpage;
+#endif
+
 				oldpage = NULL;
 				pobjects = 0;
 				pages = 0;
@@ -2511,9 +2531,17 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
 		page->pobjects = pobjects;
 		page->next = oldpage;
 
+#ifndef CONFIG_PREEMPT_RT
 	} while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
 								!= oldpage);
-	slub_put_cpu_ptr(s->cpu_slab);
+#else
+		this_cpu_write(s->cpu_slab->partial, page);
+	} while (false);
+#endif
+
+	local_unlock(&s->cpu_slab->lock);
+	if (page_to_unfreeze)
+		__unfreeze_partials(s, page_to_unfreeze);
 #endif	/* CONFIG_SLUB_CPU_PARTIAL */
 }
 
-- 
2.32.0