1The Kernel Address Sanitizer (KASAN)
2====================================
3
4Overview
5--------
6
7Kernel Address Sanitizer (KASAN) is a dynamic memory safety error detector
8designed to find out-of-bounds and use-after-free bugs.
9
10KASAN has three modes:
11
121. Generic KASAN
132. Software Tag-Based KASAN
143. Hardware Tag-Based KASAN
15
16Generic KASAN, enabled with CONFIG_KASAN_GENERIC, is the mode intended for
17debugging, similar to userspace ASan. This mode is supported on many CPU
18architectures, but it has significant performance and memory overheads.
19
20Software Tag-Based KASAN or SW_TAGS KASAN, enabled with CONFIG_KASAN_SW_TAGS,
21can be used for both debugging and dogfood testing, similar to userspace HWASan.
22This mode is only supported for arm64, but its moderate memory overhead allows
23using it for testing on memory-restricted devices with real workloads.
24
25Hardware Tag-Based KASAN or HW_TAGS KASAN, enabled with CONFIG_KASAN_HW_TAGS,
26is the mode intended to be used as an in-field memory bug detector or as a
27security mitigation. This mode only works on arm64 CPUs that support MTE
28(Memory Tagging Extension), but it has low memory and performance overheads and
29thus can be used in production.
30
31For details about the memory and performance impact of each KASAN mode, see the
32descriptions of the corresponding Kconfig options.
33
34The Generic and the Software Tag-Based modes are commonly referred to as the
35software modes. The Software Tag-Based and the Hardware Tag-Based modes are
36referred to as the tag-based modes.
37
38Support
39-------
40
41Architectures
42~~~~~~~~~~~~~
43
44Generic KASAN is supported on x86_64, arm, arm64, powerpc, riscv, s390, xtensa,
45and loongarch, and the tag-based KASAN modes are supported only on arm64.
46
47Compilers
48~~~~~~~~~
49
50Software KASAN modes use compile-time instrumentation to insert validity checks
51before every memory access and thus require a compiler version that provides
52support for that. The Hardware Tag-Based mode relies on hardware to perform
53these checks but still requires a compiler version that supports the memory
54tagging instructions.
55
56Generic KASAN requires GCC version 8.3.0 or later
57or any Clang version supported by the kernel.
58
59Software Tag-Based KASAN requires GCC 11+
60or any Clang version supported by the kernel.
61
62Hardware Tag-Based KASAN requires GCC 10+ or Clang 12+.
63
64Memory types
65~~~~~~~~~~~~
66
67Generic KASAN supports finding bugs in all of slab, page_alloc, vmap, vmalloc,
68stack, and global memory.
69
70Software Tag-Based KASAN supports slab, page_alloc, vmalloc, and stack memory.
71
72Hardware Tag-Based KASAN supports slab, page_alloc, and non-executable vmalloc
73memory.
74
75For slab, both software KASAN modes support SLUB and SLAB allocators, while
76Hardware Tag-Based KASAN only supports SLUB.
77
78Usage
79-----
80
81To enable KASAN, configure the kernel with::
82
83	  CONFIG_KASAN=y
84
85and choose between ``CONFIG_KASAN_GENERIC`` (to enable Generic KASAN),
86``CONFIG_KASAN_SW_TAGS`` (to enable Software Tag-Based KASAN), and
87``CONFIG_KASAN_HW_TAGS`` (to enable Hardware Tag-Based KASAN).
88
89For the software modes, also choose between ``CONFIG_KASAN_OUTLINE`` and
90``CONFIG_KASAN_INLINE``. Outline and inline are compiler instrumentation types.
91The former produces a smaller binary while the latter is up to 2 times faster.
92
93To include alloc and free stack traces of affected slab objects into reports,
94enable ``CONFIG_STACKTRACE``. To include alloc and free stack traces of affected
95physical pages, enable ``CONFIG_PAGE_OWNER`` and boot with ``page_owner=on``.
96
97Boot parameters
98~~~~~~~~~~~~~~~
99
100KASAN is affected by the generic ``panic_on_warn`` command line parameter.
101When it is enabled, KASAN panics the kernel after printing a bug report.
102
103By default, KASAN prints a bug report only for the first invalid memory access.
104With ``kasan_multi_shot``, KASAN prints a report on every invalid access. This
105effectively disables ``panic_on_warn`` for KASAN reports.
106
107Alternatively, independent of ``panic_on_warn``, the ``kasan.fault=`` boot
108parameter can be used to control panic and reporting behaviour:
109
110- ``kasan.fault=report``, ``=panic``, or ``=panic_on_write`` controls whether
111  to only print a KASAN report, panic the kernel, or panic the kernel on
112  invalid writes only (default: ``report``). The panic happens even if
113  ``kasan_multi_shot`` is enabled. Note that when using asynchronous mode of
114  Hardware Tag-Based KASAN, ``kasan.fault=panic_on_write`` always panics on
115  asynchronously checked accesses (including reads).
116
117Software and Hardware Tag-Based KASAN modes (see the section about various
118modes below) support altering stack trace collection behavior:
119
120- ``kasan.stacktrace=off`` or ``=on`` disables or enables alloc and free stack
121  traces collection (default: ``on``).
122- ``kasan.stack_ring_size=<number of entries>`` specifies the number of entries
123  in the stack ring (default: ``32768``).
124
125Hardware Tag-Based KASAN mode is intended for use in production as a security
126mitigation. Therefore, it supports additional boot parameters that allow
127disabling KASAN altogether or controlling its features:
128
129- ``kasan=off`` or ``=on`` controls whether KASAN is enabled (default: ``on``).
130
131- ``kasan.mode=sync``, ``=async`` or ``=asymm`` controls whether KASAN
132  is configured in synchronous, asynchronous or asymmetric mode of
133  execution (default: ``sync``).
134  Synchronous mode: a bad access is detected immediately when a tag
135  check fault occurs.
136  Asynchronous mode: a bad access detection is delayed. When a tag check
137  fault occurs, the information is stored in hardware (in the TFSR_EL1
138  register for arm64). The kernel periodically checks the hardware and
139  only reports tag faults during these checks.
140  Asymmetric mode: a bad access is detected synchronously on reads and
141  asynchronously on writes.
142
143- ``kasan.vmalloc=off`` or ``=on`` disables or enables tagging of vmalloc
144  allocations (default: ``on``).
145
146- ``kasan.page_alloc.sample=<sampling interval>`` makes KASAN tag only every
147  Nth page_alloc allocation with the order equal or greater than
148  ``kasan.page_alloc.sample.order``, where N is the value of the ``sample``
149  parameter (default: ``1``, or tag every such allocation).
150  This parameter is intended to mitigate the performance overhead introduced
151  by KASAN.
152  Note that enabling this parameter makes Hardware Tag-Based KASAN skip checks
153  of allocations chosen by sampling and thus miss bad accesses to these
154  allocations. Use the default value for accurate bug detection.
155
156- ``kasan.page_alloc.sample.order=<minimum page order>`` specifies the minimum
157  order of allocations that are affected by sampling (default: ``3``).
158  Only applies when ``kasan.page_alloc.sample`` is set to a value greater
159  than ``1``.
160  This parameter is intended to allow sampling only large page_alloc
161  allocations, which is the biggest source of the performance overhead.
162
163Error reports
164~~~~~~~~~~~~~
165
166A typical KASAN report looks like this::
167
168    ==================================================================
169    BUG: KASAN: slab-out-of-bounds in kmalloc_oob_right+0xa8/0xbc [test_kasan]
170    Write of size 1 at addr ffff8801f44ec37b by task insmod/2760
171
172    CPU: 1 PID: 2760 Comm: insmod Not tainted 4.19.0-rc3+ #698
173    Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
174    Call Trace:
175     dump_stack+0x94/0xd8
176     print_address_description+0x73/0x280
177     kasan_report+0x144/0x187
178     __asan_report_store1_noabort+0x17/0x20
179     kmalloc_oob_right+0xa8/0xbc [test_kasan]
180     kmalloc_tests_init+0x16/0x700 [test_kasan]
181     do_one_initcall+0xa5/0x3ae
182     do_init_module+0x1b6/0x547
183     load_module+0x75df/0x8070
184     __do_sys_init_module+0x1c6/0x200
185     __x64_sys_init_module+0x6e/0xb0
186     do_syscall_64+0x9f/0x2c0
187     entry_SYSCALL_64_after_hwframe+0x44/0xa9
188    RIP: 0033:0x7f96443109da
189    RSP: 002b:00007ffcf0b51b08 EFLAGS: 00000202 ORIG_RAX: 00000000000000af
190    RAX: ffffffffffffffda RBX: 000055dc3ee521a0 RCX: 00007f96443109da
191    RDX: 00007f96445cff88 RSI: 0000000000057a50 RDI: 00007f9644992000
192    RBP: 000055dc3ee510b0 R08: 0000000000000003 R09: 0000000000000000
193    R10: 00007f964430cd0a R11: 0000000000000202 R12: 00007f96445cff88
194    R13: 000055dc3ee51090 R14: 0000000000000000 R15: 0000000000000000
195
196    Allocated by task 2760:
197     save_stack+0x43/0xd0
198     kasan_kmalloc+0xa7/0xd0
199     kmem_cache_alloc_trace+0xe1/0x1b0
200     kmalloc_oob_right+0x56/0xbc [test_kasan]
201     kmalloc_tests_init+0x16/0x700 [test_kasan]
202     do_one_initcall+0xa5/0x3ae
203     do_init_module+0x1b6/0x547
204     load_module+0x75df/0x8070
205     __do_sys_init_module+0x1c6/0x200
206     __x64_sys_init_module+0x6e/0xb0
207     do_syscall_64+0x9f/0x2c0
208     entry_SYSCALL_64_after_hwframe+0x44/0xa9
209
210    Freed by task 815:
211     save_stack+0x43/0xd0
212     __kasan_slab_free+0x135/0x190
213     kasan_slab_free+0xe/0x10
214     kfree+0x93/0x1a0
215     umh_complete+0x6a/0xa0
216     call_usermodehelper_exec_async+0x4c3/0x640
217     ret_from_fork+0x35/0x40
218
219    The buggy address belongs to the object at ffff8801f44ec300
220     which belongs to the cache kmalloc-128 of size 128
221    The buggy address is located 123 bytes inside of
222     128-byte region [ffff8801f44ec300, ffff8801f44ec380)
223    The buggy address belongs to the page:
224    page:ffffea0007d13b00 count:1 mapcount:0 mapping:ffff8801f7001640 index:0x0
225    flags: 0x200000000000100(slab)
226    raw: 0200000000000100 ffffea0007d11dc0 0000001a0000001a ffff8801f7001640
227    raw: 0000000000000000 0000000080150015 00000001ffffffff 0000000000000000
228    page dumped because: kasan: bad access detected
229
230    Memory state around the buggy address:
231     ffff8801f44ec200: fc fc fc fc fc fc fc fc fb fb fb fb fb fb fb fb
232     ffff8801f44ec280: fb fb fb fb fb fb fb fb fc fc fc fc fc fc fc fc
233    >ffff8801f44ec300: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 03
234                                                                    ^
235     ffff8801f44ec380: fc fc fc fc fc fc fc fc fb fb fb fb fb fb fb fb
236     ffff8801f44ec400: fb fb fb fb fb fb fb fb fc fc fc fc fc fc fc fc
237    ==================================================================
238
239The report header summarizes what kind of bug happened and what kind of access
240caused it. It is followed by a stack trace of the bad access, a stack trace of
241where the accessed memory was allocated (in case a slab object was accessed),
242and a stack trace of where the object was freed (in case of a use-after-free
243bug report). Next comes a description of the accessed slab object and the
244information about the accessed memory page.
245
246In the end, the report shows the memory state around the accessed address.
247Internally, KASAN tracks memory state separately for each memory granule, which
248is either 8 or 16 aligned bytes depending on KASAN mode. Each number in the
249memory state section of the report shows the state of one of the memory
250granules that surround the accessed address.
251
252For Generic KASAN, the size of each memory granule is 8. The state of each
253granule is encoded in one shadow byte. Those 8 bytes can be accessible,
254partially accessible, freed, or be a part of a redzone. KASAN uses the following
255encoding for each shadow byte: 00 means that all 8 bytes of the corresponding
256memory region are accessible; number N (1 <= N <= 7) means that the first N
257bytes are accessible, and other (8 - N) bytes are not; any negative value
258indicates that the entire 8-byte word is inaccessible. KASAN uses different
259negative values to distinguish between different kinds of inaccessible memory
260like redzones or freed memory (see mm/kasan/kasan.h).
261
262In the report above, the arrow points to the shadow byte ``03``, which means
263that the accessed address is partially accessible.
264
265For tag-based KASAN modes, this last report section shows the memory tags around
266the accessed address (see the `Implementation details`_ section).
267
268Note that KASAN bug titles (like ``slab-out-of-bounds`` or ``use-after-free``)
269are best-effort: KASAN prints the most probable bug type based on the limited
270information it has. The actual type of the bug might be different.
271
272Generic KASAN also reports up to two auxiliary call stack traces. These stack
273traces point to places in code that interacted with the object but that are not
274directly present in the bad access stack trace. Currently, this includes
275call_rcu() and workqueue queuing.
276
277Implementation details
278----------------------
279
280Generic KASAN
281~~~~~~~~~~~~~
282
283Software KASAN modes use shadow memory to record whether each byte of memory is
284safe to access and use compile-time instrumentation to insert shadow memory
285checks before each memory access.
286
287Generic KASAN dedicates 1/8th of kernel memory to its shadow memory (16TB
288to cover 128TB on x86_64) and uses direct mapping with a scale and offset to
289translate a memory address to its corresponding shadow address.
290
291Here is the function which translates an address to its corresponding shadow
292address::
293
294    static inline void *kasan_mem_to_shadow(const void *addr)
295    {
296	return (void *)((unsigned long)addr >> KASAN_SHADOW_SCALE_SHIFT)
297		+ KASAN_SHADOW_OFFSET;
298    }
299
300where ``KASAN_SHADOW_SCALE_SHIFT = 3``.
301
302Compile-time instrumentation is used to insert memory access checks. Compiler
303inserts function calls (``__asan_load*(addr)``, ``__asan_store*(addr)``) before
304each memory access of size 1, 2, 4, 8, or 16. These functions check whether
305memory accesses are valid or not by checking corresponding shadow memory.
306
307With inline instrumentation, instead of making function calls, the compiler
308directly inserts the code to check shadow memory. This option significantly
309enlarges the kernel, but it gives an x1.1-x2 performance boost over the
310outline-instrumented kernel.
311
312Generic KASAN is the only mode that delays the reuse of freed objects via
313quarantine (see mm/kasan/quarantine.c for implementation).
314
315Software Tag-Based KASAN
316~~~~~~~~~~~~~~~~~~~~~~~~
317
318Software Tag-Based KASAN uses a software memory tagging approach to checking
319access validity. It is currently only implemented for the arm64 architecture.
320
321Software Tag-Based KASAN uses the Top Byte Ignore (TBI) feature of arm64 CPUs
322to store a pointer tag in the top byte of kernel pointers. It uses shadow memory
323to store memory tags associated with each 16-byte memory cell (therefore, it
324dedicates 1/16th of the kernel memory for shadow memory).
325
326On each memory allocation, Software Tag-Based KASAN generates a random tag, tags
327the allocated memory with this tag, and embeds the same tag into the returned
328pointer.
329
330Software Tag-Based KASAN uses compile-time instrumentation to insert checks
331before each memory access. These checks make sure that the tag of the memory
332that is being accessed is equal to the tag of the pointer that is used to access
333this memory. In case of a tag mismatch, Software Tag-Based KASAN prints a bug
334report.
335
336Software Tag-Based KASAN also has two instrumentation modes (outline, which
337emits callbacks to check memory accesses; and inline, which performs the shadow
338memory checks inline). With outline instrumentation mode, a bug report is
339printed from the function that performs the access check. With inline
340instrumentation, a ``brk`` instruction is emitted by the compiler, and a
341dedicated ``brk`` handler is used to print bug reports.
342
343Software Tag-Based KASAN uses 0xFF as a match-all pointer tag (accesses through
344pointers with the 0xFF pointer tag are not checked). The value 0xFE is currently
345reserved to tag freed memory regions.
346
347Hardware Tag-Based KASAN
348~~~~~~~~~~~~~~~~~~~~~~~~
349
350Hardware Tag-Based KASAN is similar to the software mode in concept but uses
351hardware memory tagging support instead of compiler instrumentation and
352shadow memory.
353
354Hardware Tag-Based KASAN is currently only implemented for arm64 architecture
355and based on both arm64 Memory Tagging Extension (MTE) introduced in ARMv8.5
356Instruction Set Architecture and Top Byte Ignore (TBI).
357
358Special arm64 instructions are used to assign memory tags for each allocation.
359Same tags are assigned to pointers to those allocations. On every memory
360access, hardware makes sure that the tag of the memory that is being accessed is
361equal to the tag of the pointer that is used to access this memory. In case of a
362tag mismatch, a fault is generated, and a report is printed.
363
364Hardware Tag-Based KASAN uses 0xFF as a match-all pointer tag (accesses through
365pointers with the 0xFF pointer tag are not checked). The value 0xFE is currently
366reserved to tag freed memory regions.
367
368If the hardware does not support MTE (pre ARMv8.5), Hardware Tag-Based KASAN
369will not be enabled. In this case, all KASAN boot parameters are ignored.
370
371Note that enabling CONFIG_KASAN_HW_TAGS always results in in-kernel TBI being
372enabled. Even when ``kasan.mode=off`` is provided or when the hardware does not
373support MTE (but supports TBI).
374
375Hardware Tag-Based KASAN only reports the first found bug. After that, MTE tag
376checking gets disabled.
377
378Shadow memory
379-------------
380
381The contents of this section are only applicable to software KASAN modes.
382
383The kernel maps memory in several different parts of the address space.
384The range of kernel virtual addresses is large: there is not enough real
385memory to support a real shadow region for every address that could be
386accessed by the kernel. Therefore, KASAN only maps real shadow for certain
387parts of the address space.
388
389Default behaviour
390~~~~~~~~~~~~~~~~~
391
392By default, architectures only map real memory over the shadow region
393for the linear mapping (and potentially other small areas). For all
394other areas - such as vmalloc and vmemmap space - a single read-only
395page is mapped over the shadow area. This read-only shadow page
396declares all memory accesses as permitted.
397
398This presents a problem for modules: they do not live in the linear
399mapping but in a dedicated module space. By hooking into the module
400allocator, KASAN temporarily maps real shadow memory to cover them.
401This allows detection of invalid accesses to module globals, for example.
402
403This also creates an incompatibility with ``VMAP_STACK``: if the stack
404lives in vmalloc space, it will be shadowed by the read-only page, and
405the kernel will fault when trying to set up the shadow data for stack
406variables.
407
408CONFIG_KASAN_VMALLOC
409~~~~~~~~~~~~~~~~~~~~
410
411With ``CONFIG_KASAN_VMALLOC``, KASAN can cover vmalloc space at the
412cost of greater memory usage. Currently, this is supported on x86,
413arm64, riscv, s390, and powerpc.
414
415This works by hooking into vmalloc and vmap and dynamically
416allocating real shadow memory to back the mappings.
417
418Most mappings in vmalloc space are small, requiring less than a full
419page of shadow space. Allocating a full shadow page per mapping would
420therefore be wasteful. Furthermore, to ensure that different mappings
421use different shadow pages, mappings would have to be aligned to
422``KASAN_GRANULE_SIZE * PAGE_SIZE``.
423
424Instead, KASAN shares backing space across multiple mappings. It allocates
425a backing page when a mapping in vmalloc space uses a particular page
426of the shadow region. This page can be shared by other vmalloc
427mappings later on.
428
429KASAN hooks into the vmap infrastructure to lazily clean up unused shadow
430memory.
431
432To avoid the difficulties around swapping mappings around, KASAN expects
433that the part of the shadow region that covers the vmalloc space will
434not be covered by the early shadow page but will be left unmapped.
435This will require changes in arch-specific code.
436
437This allows ``VMAP_STACK`` support on x86 and can simplify support of
438architectures that do not have a fixed module region.
439
440For developers
441--------------
442
443Ignoring accesses
444~~~~~~~~~~~~~~~~~
445
446Software KASAN modes use compiler instrumentation to insert validity checks.
447Such instrumentation might be incompatible with some parts of the kernel, and
448therefore needs to be disabled.
449
450Other parts of the kernel might access metadata for allocated objects.
451Normally, KASAN detects and reports such accesses, but in some cases (e.g.,
452in memory allocators), these accesses are valid.
453
454For software KASAN modes, to disable instrumentation for a specific file or
455directory, add a ``KASAN_SANITIZE`` annotation to the respective kernel
456Makefile:
457
458- For a single file (e.g., main.o)::
459
460    KASAN_SANITIZE_main.o := n
461
462- For all files in one directory::
463
464    KASAN_SANITIZE := n
465
466For software KASAN modes, to disable instrumentation on a per-function basis,
467use the KASAN-specific ``__no_sanitize_address`` function attribute or the
468generic ``noinstr`` one.
469
470Note that disabling compiler instrumentation (either on a per-file or a
471per-function basis) makes KASAN ignore the accesses that happen directly in
472that code for software KASAN modes. It does not help when the accesses happen
473indirectly (through calls to instrumented functions) or with Hardware
474Tag-Based KASAN, which does not use compiler instrumentation.
475
476For software KASAN modes, to disable KASAN reports in a part of the kernel code
477for the current task, annotate this part of the code with a
478``kasan_disable_current()``/``kasan_enable_current()`` section. This also
479disables the reports for indirect accesses that happen through function calls.
480
481For tag-based KASAN modes, to disable access checking, use
482``kasan_reset_tag()`` or ``page_kasan_tag_reset()``. Note that temporarily
483disabling access checking via ``page_kasan_tag_reset()`` requires saving and
484restoring the per-page KASAN tag via ``page_kasan_tag``/``page_kasan_tag_set``.
485
486Tests
487~~~~~
488
489There are KASAN tests that allow verifying that KASAN works and can detect
490certain types of memory corruptions. The tests consist of two parts:
491
4921. Tests that are integrated with the KUnit Test Framework. Enabled with
493``CONFIG_KASAN_KUNIT_TEST``. These tests can be run and partially verified
494automatically in a few different ways; see the instructions below.
495
4962. Tests that are currently incompatible with KUnit. Enabled with
497``CONFIG_KASAN_MODULE_TEST`` and can only be run as a module. These tests can
498only be verified manually by loading the kernel module and inspecting the
499kernel log for KASAN reports.
500
501Each KUnit-compatible KASAN test prints one of multiple KASAN reports if an
502error is detected. Then the test prints its number and status.
503
504When a test passes::
505
506        ok 28 - kmalloc_double_kzfree
507
508When a test fails due to a failed ``kmalloc``::
509
510        # kmalloc_large_oob_right: ASSERTION FAILED at lib/test_kasan.c:163
511        Expected ptr is not null, but is
512        not ok 4 - kmalloc_large_oob_right
513
514When a test fails due to a missing KASAN report::
515
516        # kmalloc_double_kzfree: EXPECTATION FAILED at lib/test_kasan.c:974
517        KASAN failure expected in "kfree_sensitive(ptr)", but none occurred
518        not ok 44 - kmalloc_double_kzfree
519
520
521At the end the cumulative status of all KASAN tests is printed. On success::
522
523        ok 1 - kasan
524
525Or, if one of the tests failed::
526
527        not ok 1 - kasan
528
529There are a few ways to run KUnit-compatible KASAN tests.
530
5311. Loadable module
532
533   With ``CONFIG_KUNIT`` enabled, KASAN-KUnit tests can be built as a loadable
534   module and run by loading ``test_kasan.ko`` with ``insmod`` or ``modprobe``.
535
5362. Built-In
537
538   With ``CONFIG_KUNIT`` built-in, KASAN-KUnit tests can be built-in as well.
539   In this case, the tests will run at boot as a late-init call.
540
5413. Using kunit_tool
542
543   With ``CONFIG_KUNIT`` and ``CONFIG_KASAN_KUNIT_TEST`` built-in, it is also
544   possible to use ``kunit_tool`` to see the results of KUnit tests in a more
545   readable way. This will not print the KASAN reports of the tests that passed.
546   See `KUnit documentation <https://www.kernel.org/doc/html/latest/dev-tools/kunit/index.html>`_
547   for more up-to-date information on ``kunit_tool``.
548
549.. _KUnit: https://www.kernel.org/doc/html/latest/dev-tools/kunit/index.html
550