3 =======================
4 Memory Allocation Guide
5 =======================
7 Linux provides a variety of APIs for memory allocation. You can
8 allocate small chunks using `kmalloc` or `kmem_cache_alloc` families,
9 large virtually contiguous areas using `vmalloc` and its derivatives,
10 or you can directly request pages from the page allocator with
11 `alloc_pages`. It is also possible to use more specialized allocators,
12 for instance `cma_alloc` or `zs_malloc`.
14 Most of the memory allocation APIs use GFP flags to express how that
15 memory should be allocated. The GFP acronym stands for "get free
16 pages", the underlying memory allocation function.
18 Diversity of the allocation APIs combined with the numerous GFP flags
19 makes the question "How should I allocate memory?" not that easy to
20 answer, although very likely you should use
24 kzalloc(<size>, GFP_KERNEL);
26 Of course there are cases when other allocation APIs and different GFP
32 The GFP flags control the allocators behavior. They tell what memory
33 zones can be used, how hard the allocator should try to find free
34 memory, whether the memory can be accessed by the userspace etc. The
35 :ref:`Documentation/core-api/mm-api.rst <mm-api-gfp-flags>` provides
36 reference documentation for the GFP flags and their combinations and
37 here we briefly outline their recommended usage:
39 * Most of the time ``GFP_KERNEL`` is what you need. Memory for the
40 kernel data structures, DMAable memory, inode cache, all these and
41 many other allocations types can use ``GFP_KERNEL``. Note, that
42 using ``GFP_KERNEL`` implies ``GFP_RECLAIM``, which means that
43 direct reclaim may be triggered under memory pressure; the calling
44 context must be allowed to sleep.
45 * If the allocation is performed from an atomic context, e.g interrupt
46 handler, use ``GFP_NOWAIT``. This flag prevents direct reclaim and
47 IO or filesystem operations. Consequently, under memory pressure
48 ``GFP_NOWAIT`` allocation is likely to fail. Users of this flag need
49 to provide a suitable fallback to cope with such failures where
51 * If you think that accessing memory reserves is justified and the kernel
52 will be stressed unless allocation succeeds, you may use ``GFP_ATOMIC``.
53 * Untrusted allocations triggered from userspace should be a subject
54 of kmem accounting and must have ``__GFP_ACCOUNT`` bit set. There
55 is the handy ``GFP_KERNEL_ACCOUNT`` shortcut for ``GFP_KERNEL``
56 allocations that should be accounted.
57 * Userspace allocations should use either of the ``GFP_USER``,
58 ``GFP_HIGHUSER`` or ``GFP_HIGHUSER_MOVABLE`` flags. The longer
59 the flag name the less restrictive it is.
61 ``GFP_HIGHUSER_MOVABLE`` does not require that allocated memory
62 will be directly accessible by the kernel and implies that the
65 ``GFP_HIGHUSER`` means that the allocated memory is not movable,
66 but it is not required to be directly accessible by the kernel. An
67 example may be a hardware allocation that maps data directly into
68 userspace but has no addressing limitations.
70 ``GFP_USER`` means that the allocated memory is not movable and it
71 must be directly accessible by the kernel.
73 You may notice that quite a few allocations in the existing code
74 specify ``GFP_NOIO`` or ``GFP_NOFS``. Historically, they were used to
75 prevent recursion deadlocks caused by direct memory reclaim calling
76 back into the FS or IO paths and blocking on already held
77 resources. Since 4.12 the preferred way to address this issue is to
78 use new scope APIs described in
79 :ref:`Documentation/core-api/gfp_mask-from-fs-io.rst <gfp_mask_from_fs_io>`.
81 Other legacy GFP flags are ``GFP_DMA`` and ``GFP_DMA32``. They are
82 used to ensure that the allocated memory is accessible by hardware
83 with limited addressing capabilities. So unless you are writing a
84 driver for a device with such restrictions, avoid using these flags.
85 And even with hardware with restrictions it is preferable to use
88 GFP flags and reclaim behavior
89 ------------------------------
90 Memory allocations may trigger direct or background reclaim and it is
91 useful to understand how hard the page allocator will try to satisfy that
94 * ``GFP_KERNEL & ~__GFP_RECLAIM`` - optimistic allocation without _any_
95 attempt to free memory at all. The most light weight mode which even
96 doesn't kick the background reclaim. Should be used carefully because it
97 might deplete the memory and the next user might hit the more aggressive
100 * ``GFP_KERNEL & ~__GFP_DIRECT_RECLAIM`` (or ``GFP_NOWAIT``)- optimistic
101 allocation without any attempt to free memory from the current
102 context but can wake kswapd to reclaim memory if the zone is below
103 the low watermark. Can be used from either atomic contexts or when
104 the request is a performance optimization and there is another
105 fallback for a slow path.
107 * ``(GFP_KERNEL|__GFP_HIGH) & ~__GFP_DIRECT_RECLAIM`` (aka ``GFP_ATOMIC``) -
108 non sleeping allocation with an expensive fallback so it can access
109 some portion of memory reserves. Usually used from interrupt/bottom-half
110 context with an expensive slow path fallback.
112 * ``GFP_KERNEL`` - both background and direct reclaim are allowed and the
113 **default** page allocator behavior is used. That means that not costly
114 allocation requests are basically no-fail but there is no guarantee of
115 that behavior so failures have to be checked properly by callers
116 (e.g. OOM killer victim is allowed to fail currently).
118 * ``GFP_KERNEL | __GFP_NORETRY`` - overrides the default allocator behavior
119 and all allocation requests fail early rather than cause disruptive
120 reclaim (one round of reclaim in this implementation). The OOM killer
123 * ``GFP_KERNEL | __GFP_RETRY_MAYFAIL`` - overrides the default allocator
124 behavior and all allocation requests try really hard. The request
125 will fail if the reclaim cannot make any progress. The OOM killer
128 * ``GFP_KERNEL | __GFP_NOFAIL`` - overrides the default allocator behavior
129 and all allocation requests will loop endlessly until they succeed.
130 This might be really dangerous especially for larger orders.
132 Selecting memory allocator
133 ==========================
135 The most straightforward way to allocate memory is to use a function
136 from the kmalloc() family. And, to be on the safe side it's best to use
137 routines that set memory to zero, like kzalloc(). If you need to
138 allocate memory for an array, there are kmalloc_array() and kcalloc()
139 helpers. The helpers struct_size(), array_size() and array3_size() can
140 be used to safely calculate object sizes without overflowing.
142 The maximal size of a chunk that can be allocated with `kmalloc` is
143 limited. The actual limit depends on the hardware and the kernel
144 configuration, but it is a good practice to use `kmalloc` for objects
145 smaller than page size.
147 The address of a chunk allocated with `kmalloc` is aligned to at least
148 ARCH_KMALLOC_MINALIGN bytes. For sizes which are a power of two, the
149 alignment is also guaranteed to be at least the respective size. For other
150 sizes, the alignment is guaranteed to be at least the largest power-of-two
153 Chunks allocated with kmalloc() can be resized with krealloc(). Similarly
154 to kmalloc_array(): a helper for resizing arrays is provided in the form of
157 For large allocations you can use vmalloc() and vzalloc(), or directly
158 request pages from the page allocator. The memory allocated by `vmalloc`
159 and related functions is not physically contiguous.
161 If you are not sure whether the allocation size is too large for
162 `kmalloc`, it is possible to use kvmalloc() and its derivatives. It will
163 try to allocate memory with `kmalloc` and if the allocation fails it
164 will be retried with `vmalloc`. There are restrictions on which GFP
165 flags can be used with `kvmalloc`; please see kvmalloc_node() reference
166 documentation. Note that `kvmalloc` may return memory that is not
167 physically contiguous.
169 If you need to allocate many identical objects you can use the slab
170 cache allocator. The cache should be set up with kmem_cache_create() or
171 kmem_cache_create_usercopy() before it can be used. The second function
172 should be used if a part of the cache might be copied to the userspace.
173 After the cache is created kmem_cache_alloc() and its convenience
174 wrappers can allocate memory from that cache.
176 When the allocated memory is no longer needed it must be freed.
178 Objects allocated by `kmalloc` can be freed by `kfree` or `kvfree`. Objects
179 allocated by `kmem_cache_alloc` can be freed with `kmem_cache_free`, `kfree`
180 or `kvfree`, where the latter two might be more convenient thanks to not
181 needing the kmem_cache pointer.
183 The same rules apply to _bulk and _rcu flavors of freeing functions.
185 Memory allocated by `vmalloc` can be freed with `vfree` or `kvfree`.
186 Memory allocated by `kvmalloc` can be freed with `kvfree`.
187 Caches created by `kmem_cache_create` should be freed with
188 `kmem_cache_destroy` only after freeing all the allocated objects first.