Re: mm, slab/slub: Ensure kmem_cache_alloc_bulk() is available early

[Vlastimil Babka <vbabka@suse.cz>]

On 2/7/23 15:45, Vlastimil Babka wrote:
> On 2/7/23 15:16, Thomas Gleixner wrote:
>> The memory allocators are available during early boot even in the phase
>> where interrupts are disabled and scheduling is not yet possible.
>> 
>> The setup is so that GFP_KERNEL allocations work in this phase without
>> causing might_alloc() splats to be emitted because the system state is
>> SYSTEM_BOOTING at that point which prevents the warnings to trigger.
>> 
>> Most allocation/free functions use local_irq_save()/restore() or a lock
>> variant of that. But kmem_cache_alloc_bulk() and kmem_cache_free_bulk() use
>> local_[lock]_irq_disable()/enable(), which leads to a lockdep warning when
>> interrupts are enabled during the early boot phase.
>> 
>> This went unnoticed so far as there are no early users of these
>> interfaces. The upcoming conversion of the interrupt descriptor store from
>> radix_tree to maple_tree triggered this warning as maple_tree uses the bulk
>> interface.
>> 
>> Cure this by moving the kmem_cache_alloc/free() bulk variants of SLUB and
>> SLAB to local[_lock]_irq_save()/restore().
>> 
>> There is obviously no reclaim possible and required at this point so there
>> is no need to expand this coverage further.
>> 
>> No functional change.
>> 
>> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
> 
> +Cc rest of slab folks
> 
> Thanks, added to slab/for-6.3/fixes

After your patch, I think it also makes sense to do the following:
----8<----
From 340d7c7b99f3e67780f6dec480ed1d27e6f325eb Mon Sep 17 00:00:00 2001
From: Vlastimil Babka <vbabka@suse.cz>
Date: Tue, 7 Feb 2023 15:34:53 +0100
Subject: [PATCH] mm, slab/slub: remove notes that bulk alloc/free needs
 interrupts enabled

The slab functions kmem_cache_[alloc|free]_bulk() have been documented
as requiring interrupts to be enabled, since their addition in 2015.
It's unclear whether that was a fundamental restriction, or an attempt
to save some cpu cycles by not having to save and restore the irq flags.

However, it appears that most of the code involved was/became safe to be
called with interrupts disabled, and the remaining bits were fixed by
commit f244b0182b8e ("mm, slab/slub: Ensure kmem_cache_alloc_bulk() is
available early"). While the commit was aimed at early boot scenario, we
can now also remove the documented restrictions for any interrupt
disabled scenarios.

Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
---
 include/linux/slab.h | 2 --
 mm/slub.c            | 2 --
 2 files changed, 4 deletions(-)

diff --git a/include/linux/slab.h b/include/linux/slab.h
index 45af70315a94..ea439b4e2b34 100644
--- a/include/linux/slab.h
+++ b/include/linux/slab.h
@@ -481,8 +481,6 @@ void kmem_cache_free(struct kmem_cache *s, void *objp);
  * Bulk allocation and freeing operations. These are accelerated in an
  * allocator specific way to avoid taking locks repeatedly or building
  * metadata structures unnecessarily.
- *
- * Note that interrupts must be enabled when calling these functions.
  */
 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p);
 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, void **p);
diff --git a/mm/slub.c b/mm/slub.c
index c16d78698e3f..23b3fb86045d 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -3889,7 +3889,6 @@ int build_detached_freelist(struct kmem_cache *s, size_t size,
 	return same;
 }
 
-/* Note that interrupts must be enabled when calling this function. */
 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
 {
 	if (!size)
@@ -4009,7 +4008,6 @@ static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
 }
 #endif /* CONFIG_SLUB_TINY */
 
-/* Note that interrupts must be enabled when calling this function. */
 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
 			  void **p)
 {
-- 
2.39.1