Message-Id: <20251023-sheaves-for-all-v1-18-6ffa2c9941c0@suse.cz>
Date: Thu, 23 Oct 2025 15:52:40 +0200
From: Vlastimil Babka <vbabka@...e.cz>
To: Andrew Morton <akpm@...ux-foundation.org>, 
 Christoph Lameter <cl@...two.org>, David Rientjes <rientjes@...gle.com>, 
 Roman Gushchin <roman.gushchin@...ux.dev>, Harry Yoo <harry.yoo@...cle.com>
Cc: Uladzislau Rezki <urezki@...il.com>, 
 "Liam R. Howlett" <Liam.Howlett@...cle.com>, 
 Suren Baghdasaryan <surenb@...gle.com>, 
 Sebastian Andrzej Siewior <bigeasy@...utronix.de>, 
 Alexei Starovoitov <ast@...nel.org>, linux-mm@...ck.org, 
 linux-kernel@...r.kernel.org, linux-rt-devel@...ts.linux.dev, 
 bpf@...r.kernel.org, kasan-dev@...glegroups.com, 
 Vlastimil Babka <vbabka@...e.cz>
Subject: [PATCH RFC 18/19] slab: update overview comments

The changes related to sheaves made the description of locking and other
details outdated. Update it to reflect current state.

Also add a new copyright line due to major changes.

Signed-off-by: Vlastimil Babka <vbabka@...e.cz>
---
 mm/slub.c | 141 +++++++++++++++++++++++++++++---------------------------------
 1 file changed, 67 insertions(+), 74 deletions(-)

diff --git a/mm/slub.c b/mm/slub.c
index 4e003493ba60..515a2b59cb52 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -1,13 +1,15 @@
 // SPDX-License-Identifier: GPL-2.0
 /*
- * SLUB: A slab allocator that limits cache line use instead of queuing
- * objects in per cpu and per node lists.
+ * SLUB: A slab allocator with low overhead percpu array caches and mostly
+ * lockless freeing of objects to slabs in the slowpath.
  *
- * The allocator synchronizes using per slab locks or atomic operations
- * and only uses a centralized lock to manage a pool of partial slabs.
+ * The allocator synchronizes using spin_trylock for percpu arrays in the
+ * fastpath, and cmpxchg_double (or bit spinlock) for slowpath freeing.
+ * Uses a centralized lock to manage a pool of partial slabs.
  *
  * (C) 2007 SGI, Christoph Lameter
  * (C) 2011 Linux Foundation, Christoph Lameter
+ * (C) 2025 SUSE, Vlastimil Babka
  */
 
 #include <linux/mm.h>
@@ -53,11 +55,13 @@
 
 /*
  * Lock order:
- *   1. slab_mutex (Global Mutex)
- *   2. node->list_lock (Spinlock)
- *   3. kmem_cache->cpu_slab->lock (Local lock)
- *   4. slab_lock(slab) (Only on some arches)
- *   5. object_map_lock (Only for debugging)
+ *   0.  cpu_hotplug_lock
+ *   1.  slab_mutex (Global Mutex)
+ *   2a. kmem_cache->cpu_sheaves->lock (Local trylock)
+ *   2b. node->barn->lock (Spinlock)
+ *   2c. node->list_lock (Spinlock)
+ *   3.  slab_lock(slab) (Only on some arches)
+ *   4.  object_map_lock (Only for debugging)
  *
  *   slab_mutex
  *
@@ -78,31 +82,38 @@
  *	C. slab->objects	-> Number of objects in slab
  *	D. slab->frozen		-> frozen state
  *
- *   Frozen slabs
+ *   SL_partial slabs
+ *
+ *   Slabs on node partial list have at least one free object. A limited number
+ *   of slabs on the list can be fully free (slab->inuse == 0), until we start
+ *   discarding them. These slabs are marked with SL_partial, and the flag is
+ *   cleared while removing them, usually to grab their freelist afterwards.
+ *   This clearing also exempts them from list management. Please see
+ *   __slab_free() for more details.
  *
- *   If a slab is frozen then it is exempt from list management. It is
- *   the cpu slab which is actively allocated from by the processor that
- *   froze it and it is not on any list. The processor that froze the
- *   slab is the one who can perform list operations on the slab. Other
- *   processors may put objects onto the freelist but the processor that
- *   froze the slab is the only one that can retrieve the objects from the
- *   slab's freelist.
+ *   Full slabs
  *
- *   CPU partial slabs
+ *   For caches without debugging enabled, full slabs (slab->inuse ==
+ *   slab->objects and slab->freelist == NULL) are not placed on any list.
+ *   The __slab_free() freeing the first object from such a slab will place
+ *   it on the partial list. Caches with debugging enabled place such slab
+ *   on the full list and use different allocation and freeing paths.
+ *
+ *   Frozen slabs
  *
- *   The partially empty slabs cached on the CPU partial list are used
- *   for performance reasons, which speeds up the allocation process.
- *   These slabs are not frozen, but are also exempt from list management,
- *   by clearing the SL_partial flag when moving out of the node
- *   partial list. Please see __slab_free() for more details.
+ *   If a slab is frozen then it is exempt from list management. It is used to
+ *   indicate a slab that has failed consistency checks and thus cannot be
+ *   allocated from anymore - it is also marked as full. Any previously
+ *   allocated objects will be simply leaked upon freeing instead of attempting
+ *   to modify the potentially corrupted freelist and metadata.
  *
  *   To sum up, the current scheme is:
- *   - node partial slab: SL_partial && !frozen
- *   - cpu partial slab: !SL_partial && !frozen
- *   - cpu slab: !SL_partial && frozen
- *   - full slab: !SL_partial && !frozen
+ *   - node partial slab:            SL_partial && !full && !frozen
+ *   - taken off partial list:      !SL_partial && !full && !frozen
+ *   - full slab, not on any list:  !SL_partial &&  full && !frozen
+ *   - frozen due to inconsistency: !SL_partial &&  full &&  frozen
  *
- *   list_lock
+ *   node->list_lock (spinlock)
  *
  *   The list_lock protects the partial and full list on each node and
  *   the partial slab counter. If taken then no new slabs may be added or
@@ -112,47 +123,46 @@
  *
  *   The list_lock is a centralized lock and thus we avoid taking it as
  *   much as possible. As long as SLUB does not have to handle partial
- *   slabs, operations can continue without any centralized lock. F.e.
- *   allocating a long series of objects that fill up slabs does not require
- *   the list lock.
+ *   slabs, operations can continue without any centralized lock.
  *
  *   For debug caches, all allocations are forced to go through a list_lock
  *   protected region to serialize against concurrent validation.
  *
- *   cpu_slab->lock local lock
+ *   cpu_sheaves->lock (local_trylock)
  *
- *   This locks protect slowpath manipulation of all kmem_cache_cpu fields
- *   except the stat counters. This is a percpu structure manipulated only by
- *   the local cpu, so the lock protects against being preempted or interrupted
- *   by an irq. Fast path operations rely on lockless operations instead.
+ *   This lock protects fastpath operations on the percpu sheaves. On !RT it
+ *   only disables preemption and does no atomic operations. As long as the main
+ *   or spare sheaf can handle the allocation or free, there is no other
+ *   overhead.
  *
- *   On PREEMPT_RT, the local lock neither disables interrupts nor preemption
- *   which means the lockless fastpath cannot be used as it might interfere with
- *   an in-progress slow path operations. In this case the local lock is always
- *   taken but it still utilizes the freelist for the common operations.
+ *   node->barn->lock (spinlock)
  *
- *   lockless fastpaths
+ *   This lock protects the operations on per-NUMA-node barn. It can quickly
+ *   serve an empty or full sheaf if available, and avoid more expensive refill
+ *   or flush operation.
  *
- *   The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
- *   are fully lockless when satisfied from the percpu slab (and when
- *   cmpxchg_double is possible to use, otherwise slab_lock is taken).
- *   They also don't disable preemption or migration or irqs. They rely on
- *   the transaction id (tid) field to detect being preempted or moved to
- *   another cpu.
+ *   Lockless freeing
+ *
+ *   Objects may have to be freed to their slabs when they are from a remote
+ *   node (where we want to avoid filling local sheaves with remote objects)
+ *   or when there are too many full sheaves. On architectures supporting
+ *   cmpxchg_double this is done by a lockless update of slab's freelist and
+ *   counters, otherwise slab_lock is taken. This only needs to take the
+ *   list_lock if it's a first free to a full slab, or when there are too many
+ *   fully free slabs and some need to be discarded.
  *
  *   irq, preemption, migration considerations
  *
- *   Interrupts are disabled as part of list_lock or local_lock operations, or
+ *   Interrupts are disabled as part of list_lock or barn lock operations, or
  *   around the slab_lock operation, in order to make the slab allocator safe
  *   to use in the context of an irq.
+ *   Preemption is disabled as part of local_trylock operations.
+ *   kmalloc_nolock() and kfree_nolock() are safe in NMI context but see
+ *   their limitations.
  *
- *   In addition, preemption (or migration on PREEMPT_RT) is disabled in the
- *   allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
- *   local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
- *   doesn't have to be revalidated in each section protected by the local lock.
- *
- * SLUB assigns one slab for allocation to each processor.
- * Allocations only occur from these slabs called cpu slabs.
+ * SLUB assigns two object arrays called sheaves for caching allocation and
+ * frees on each cpu, with a NUMA node shared barn for balancing between cpus.
+ * Allocations and frees are primarily served from these sheaves.
  *
  * Slabs with free elements are kept on a partial list and during regular
  * operations no list for full slabs is used. If an object in a full slab is
@@ -160,25 +170,8 @@
  * We track full slabs for debugging purposes though because otherwise we
  * cannot scan all objects.
  *
- * Slabs are freed when they become empty. Teardown and setup is
- * minimal so we rely on the page allocators per cpu caches for
- * fast frees and allocs.
- *
- * slab->frozen		The slab is frozen and exempt from list processing.
- * 			This means that the slab is dedicated to a purpose
- * 			such as satisfying allocations for a specific
- * 			processor. Objects may be freed in the slab while
- * 			it is frozen but slab_free will then skip the usual
- * 			list operations. It is up to the processor holding
- * 			the slab to integrate the slab into the slab lists
- * 			when the slab is no longer needed.
- *
- * 			One use of this flag is to mark slabs that are
- * 			used for allocations. Then such a slab becomes a cpu
- * 			slab. The cpu slab may be equipped with an additional
- * 			freelist that allows lockless access to
- * 			free objects in addition to the regular freelist
- * 			that requires the slab lock.
+ * Slabs are freed when they become empty. Teardown and setup is minimal so we
+ * rely on the page allocators per cpu caches for fast frees and allocs.
  *
  * SLAB_DEBUG_FLAGS	Slab requires special handling due to debug
  * 			options set. This moves	slab handling out of

-- 
2.51.1

Powered by blists - more mailing lists