1The Kernel Address Sanitizer (KASAN)
2====================================
3
4Overview
5--------
6
7KernelAddressSANitizer (KASAN) is a dynamic memory error detector designed to
8find out-of-bound and use-after-free bugs. KASAN has two modes: generic KASAN
9(similar to userspace ASan) and software tag-based KASAN (similar to userspace
10HWASan).
11
12KASAN uses compile-time instrumentation to insert validity checks before every
13memory access, and therefore requires a compiler version that supports that.
14
15Generic KASAN is supported in both GCC and Clang. With GCC it requires version
164.9.2 or later for basic support and version 5.0 or later for detection of
17out-of-bounds accesses for stack and global variables and for inline
18instrumentation mode (see the Usage section). With Clang it requires version
197.0.0 or later and it doesn't support detection of out-of-bounds accesses for
20global variables yet.
21
22Tag-based KASAN is only supported in Clang and requires version 7.0.0 or later.
23
24Currently generic KASAN is supported for the x86_64, arm64, xtensa, s390 and
25riscv architectures, and tag-based KASAN is supported only for arm64.
26
27Usage
28-----
29
30To enable KASAN configure kernel with::
31
32	  CONFIG_KASAN = y
33
34and choose between CONFIG_KASAN_GENERIC (to enable generic KASAN) and
35CONFIG_KASAN_SW_TAGS (to enable software tag-based KASAN).
36
37You also need to choose between CONFIG_KASAN_OUTLINE and CONFIG_KASAN_INLINE.
38Outline and inline are compiler instrumentation types. The former produces
39smaller binary while the latter is 1.1 - 2 times faster.
40
41Both KASAN modes work with both SLUB and SLAB memory allocators.
42For better bug detection and nicer reporting, enable CONFIG_STACKTRACE.
43
44To augment reports with last allocation and freeing stack of the physical page,
45it is recommended to enable also CONFIG_PAGE_OWNER and boot with page_owner=on.
46
47To disable instrumentation for specific files or directories, add a line
48similar to the following to the respective kernel Makefile:
49
50- For a single file (e.g. main.o)::
51
52    KASAN_SANITIZE_main.o := n
53
54- For all files in one directory::
55
56    KASAN_SANITIZE := n
57
58Error reports
59~~~~~~~~~~~~~
60
61A typical out-of-bounds access generic KASAN report looks like this::
62
63    ==================================================================
64    BUG: KASAN: slab-out-of-bounds in kmalloc_oob_right+0xa8/0xbc [test_kasan]
65    Write of size 1 at addr ffff8801f44ec37b by task insmod/2760
66
67    CPU: 1 PID: 2760 Comm: insmod Not tainted 4.19.0-rc3+ #698
68    Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
69    Call Trace:
70     dump_stack+0x94/0xd8
71     print_address_description+0x73/0x280
72     kasan_report+0x144/0x187
73     __asan_report_store1_noabort+0x17/0x20
74     kmalloc_oob_right+0xa8/0xbc [test_kasan]
75     kmalloc_tests_init+0x16/0x700 [test_kasan]
76     do_one_initcall+0xa5/0x3ae
77     do_init_module+0x1b6/0x547
78     load_module+0x75df/0x8070
79     __do_sys_init_module+0x1c6/0x200
80     __x64_sys_init_module+0x6e/0xb0
81     do_syscall_64+0x9f/0x2c0
82     entry_SYSCALL_64_after_hwframe+0x44/0xa9
83    RIP: 0033:0x7f96443109da
84    RSP: 002b:00007ffcf0b51b08 EFLAGS: 00000202 ORIG_RAX: 00000000000000af
85    RAX: ffffffffffffffda RBX: 000055dc3ee521a0 RCX: 00007f96443109da
86    RDX: 00007f96445cff88 RSI: 0000000000057a50 RDI: 00007f9644992000
87    RBP: 000055dc3ee510b0 R08: 0000000000000003 R09: 0000000000000000
88    R10: 00007f964430cd0a R11: 0000000000000202 R12: 00007f96445cff88
89    R13: 000055dc3ee51090 R14: 0000000000000000 R15: 0000000000000000
90
91    Allocated by task 2760:
92     save_stack+0x43/0xd0
93     kasan_kmalloc+0xa7/0xd0
94     kmem_cache_alloc_trace+0xe1/0x1b0
95     kmalloc_oob_right+0x56/0xbc [test_kasan]
96     kmalloc_tests_init+0x16/0x700 [test_kasan]
97     do_one_initcall+0xa5/0x3ae
98     do_init_module+0x1b6/0x547
99     load_module+0x75df/0x8070
100     __do_sys_init_module+0x1c6/0x200
101     __x64_sys_init_module+0x6e/0xb0
102     do_syscall_64+0x9f/0x2c0
103     entry_SYSCALL_64_after_hwframe+0x44/0xa9
104
105    Freed by task 815:
106     save_stack+0x43/0xd0
107     __kasan_slab_free+0x135/0x190
108     kasan_slab_free+0xe/0x10
109     kfree+0x93/0x1a0
110     umh_complete+0x6a/0xa0
111     call_usermodehelper_exec_async+0x4c3/0x640
112     ret_from_fork+0x35/0x40
113
114    The buggy address belongs to the object at ffff8801f44ec300
115     which belongs to the cache kmalloc-128 of size 128
116    The buggy address is located 123 bytes inside of
117     128-byte region [ffff8801f44ec300, ffff8801f44ec380)
118    The buggy address belongs to the page:
119    page:ffffea0007d13b00 count:1 mapcount:0 mapping:ffff8801f7001640 index:0x0
120    flags: 0x200000000000100(slab)
121    raw: 0200000000000100 ffffea0007d11dc0 0000001a0000001a ffff8801f7001640
122    raw: 0000000000000000 0000000080150015 00000001ffffffff 0000000000000000
123    page dumped because: kasan: bad access detected
124
125    Memory state around the buggy address:
126     ffff8801f44ec200: fc fc fc fc fc fc fc fc fb fb fb fb fb fb fb fb
127     ffff8801f44ec280: fb fb fb fb fb fb fb fb fc fc fc fc fc fc fc fc
128    >ffff8801f44ec300: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 03
129                                                                    ^
130     ffff8801f44ec380: fc fc fc fc fc fc fc fc fb fb fb fb fb fb fb fb
131     ffff8801f44ec400: fb fb fb fb fb fb fb fb fc fc fc fc fc fc fc fc
132    ==================================================================
133
134The header of the report provides a short summary of what kind of bug happened
135and what kind of access caused it. It's followed by a stack trace of the bad
136access, a stack trace of where the accessed memory was allocated (in case bad
137access happens on a slab object), and a stack trace of where the object was
138freed (in case of a use-after-free bug report). Next comes a description of
139the accessed slab object and information about the accessed memory page.
140
141In the last section the report shows memory state around the accessed address.
142Reading this part requires some understanding of how KASAN works.
143
144The state of each 8 aligned bytes of memory is encoded in one shadow byte.
145Those 8 bytes can be accessible, partially accessible, freed or be a redzone.
146We use the following encoding for each shadow byte: 0 means that all 8 bytes
147of the corresponding memory region are accessible; number N (1 <= N <= 7) means
148that the first N bytes are accessible, and other (8 - N) bytes are not;
149any negative value indicates that the entire 8-byte word is inaccessible.
150We use different negative values to distinguish between different kinds of
151inaccessible memory like redzones or freed memory (see mm/kasan/kasan.h).
152
153In the report above the arrows point to the shadow byte 03, which means that
154the accessed address is partially accessible.
155
156For tag-based KASAN this last report section shows the memory tags around the
157accessed address (see Implementation details section).
158
159
160Implementation details
161----------------------
162
163Generic KASAN
164~~~~~~~~~~~~~
165
166From a high level, our approach to memory error detection is similar to that
167of kmemcheck: use shadow memory to record whether each byte of memory is safe
168to access, and use compile-time instrumentation to insert checks of shadow
169memory on each memory access.
170
171Generic KASAN dedicates 1/8th of kernel memory to its shadow memory (e.g. 16TB
172to cover 128TB on x86_64) and uses direct mapping with a scale and offset to
173translate a memory address to its corresponding shadow address.
174
175Here is the function which translates an address to its corresponding shadow
176address::
177
178    static inline void *kasan_mem_to_shadow(const void *addr)
179    {
180	return ((unsigned long)addr >> KASAN_SHADOW_SCALE_SHIFT)
181		+ KASAN_SHADOW_OFFSET;
182    }
183
184where ``KASAN_SHADOW_SCALE_SHIFT = 3``.
185
186Compile-time instrumentation is used to insert memory access checks. Compiler
187inserts function calls (__asan_load*(addr), __asan_store*(addr)) before each
188memory access of size 1, 2, 4, 8 or 16. These functions check whether memory
189access is valid or not by checking corresponding shadow memory.
190
191GCC 5.0 has possibility to perform inline instrumentation. Instead of making
192function calls GCC directly inserts the code to check the shadow memory.
193This option significantly enlarges kernel but it gives x1.1-x2 performance
194boost over outline instrumented kernel.
195
196Software tag-based KASAN
197~~~~~~~~~~~~~~~~~~~~~~~~
198
199Tag-based KASAN uses the Top Byte Ignore (TBI) feature of modern arm64 CPUs to
200store a pointer tag in the top byte of kernel pointers. Like generic KASAN it
201uses shadow memory to store memory tags associated with each 16-byte memory
202cell (therefore it dedicates 1/16th of the kernel memory for shadow memory).
203
204On each memory allocation tag-based KASAN generates a random tag, tags the
205allocated memory with this tag, and embeds this tag into the returned pointer.
206Software tag-based KASAN uses compile-time instrumentation to insert checks
207before each memory access. These checks make sure that tag of the memory that
208is being accessed is equal to tag of the pointer that is used to access this
209memory. In case of a tag mismatch tag-based KASAN prints a bug report.
210
211Software tag-based KASAN also has two instrumentation modes (outline, that
212emits callbacks to check memory accesses; and inline, that performs the shadow
213memory checks inline). With outline instrumentation mode, a bug report is
214simply printed from the function that performs the access check. With inline
215instrumentation a brk instruction is emitted by the compiler, and a dedicated
216brk handler is used to print bug reports.
217
218A potential expansion of this mode is a hardware tag-based mode, which would
219use hardware memory tagging support instead of compiler instrumentation and
220manual shadow memory manipulation.
221
222What memory accesses are sanitised by KASAN?
223--------------------------------------------
224
225The kernel maps memory in a number of different parts of the address
226space. This poses something of a problem for KASAN, which requires
227that all addresses accessed by instrumented code have a valid shadow
228region.
229
230The range of kernel virtual addresses is large: there is not enough
231real memory to support a real shadow region for every address that
232could be accessed by the kernel.
233
234By default
235~~~~~~~~~~
236
237By default, architectures only map real memory over the shadow region
238for the linear mapping (and potentially other small areas). For all
239other areas - such as vmalloc and vmemmap space - a single read-only
240page is mapped over the shadow area. This read-only shadow page
241declares all memory accesses as permitted.
242
243This presents a problem for modules: they do not live in the linear
244mapping, but in a dedicated module space. By hooking in to the module
245allocator, KASAN can temporarily map real shadow memory to cover
246them. This allows detection of invalid accesses to module globals, for
247example.
248
249This also creates an incompatibility with ``VMAP_STACK``: if the stack
250lives in vmalloc space, it will be shadowed by the read-only page, and
251the kernel will fault when trying to set up the shadow data for stack
252variables.
253
254CONFIG_KASAN_VMALLOC
255~~~~~~~~~~~~~~~~~~~~
256
257With ``CONFIG_KASAN_VMALLOC``, KASAN can cover vmalloc space at the
258cost of greater memory usage. Currently this is only supported on x86.
259
260This works by hooking into vmalloc and vmap, and dynamically
261allocating real shadow memory to back the mappings.
262
263Most mappings in vmalloc space are small, requiring less than a full
264page of shadow space. Allocating a full shadow page per mapping would
265therefore be wasteful. Furthermore, to ensure that different mappings
266use different shadow pages, mappings would have to be aligned to
267``KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE``.
268
269Instead, we share backing space across multiple mappings. We allocate
270a backing page when a mapping in vmalloc space uses a particular page
271of the shadow region. This page can be shared by other vmalloc
272mappings later on.
273
274We hook in to the vmap infrastructure to lazily clean up unused shadow
275memory.
276
277To avoid the difficulties around swapping mappings around, we expect
278that the part of the shadow region that covers the vmalloc space will
279not be covered by the early shadow page, but will be left
280unmapped. This will require changes in arch-specific code.
281
282This allows ``VMAP_STACK`` support on x86, and can simplify support of
283architectures that do not have a fixed module region.
284