linux/Documentation/dev-tools/kasan.rst

487 lines
21 KiB
ReStructuredText
Raw Normal View History

The Kernel Address Sanitizer (KASAN)
====================================
Overview
--------
KernelAddressSANitizer (KASAN) is a dynamic memory safety error detector
designed to find out-of-bound and use-after-free bugs. KASAN has three modes:
1. generic KASAN (similar to userspace ASan),
2. software tag-based KASAN (similar to userspace HWASan),
3. hardware tag-based KASAN (based on hardware memory tagging).
Generic KASAN is mainly used for debugging due to a large memory overhead.
Software tag-based KASAN can be used for dogfood testing as it has a lower
memory overhead that allows using it with real workloads. Hardware tag-based
KASAN comes with low memory and performance overheads and, therefore, can be
used in production. Either as an in-field memory bug detector or as a security
mitigation.
Software KASAN modes (#1 and #2) use compile-time instrumentation to insert
validity checks before every memory access and, therefore, require a compiler
version that supports that.
Generic KASAN is supported in GCC and Clang. With GCC, it requires version
8.3.0 or later. Any supported Clang version is compatible, but detection of
out-of-bounds accesses for global variables is only supported since Clang 11.
Software tag-based KASAN mode is only supported in Clang.
The hardware KASAN mode (#3) relies on hardware to perform the checks but
still requires a compiler version that supports memory tagging instructions.
This mode is supported in GCC 10+ and Clang 11+.
Both software KASAN modes work with SLUB and SLAB memory allocators,
while the hardware tag-based KASAN currently only supports SLUB.
Currently, generic KASAN is supported for the x86_64, arm, arm64, xtensa, s390,
and riscv architectures, and tag-based KASAN modes are supported only for arm64.
Usage
-----
To enable KASAN, configure the kernel with::
CONFIG_KASAN=y
and choose between ``CONFIG_KASAN_GENERIC`` (to enable generic KASAN),
``CONFIG_KASAN_SW_TAGS`` (to enable software tag-based KASAN), and
``CONFIG_KASAN_HW_TAGS`` (to enable hardware tag-based KASAN).
For software modes, also choose between ``CONFIG_KASAN_OUTLINE`` and
``CONFIG_KASAN_INLINE``. Outline and inline are compiler instrumentation types.
The former produces a smaller binary while the latter is 1.1-2 times faster.
To include alloc and free stack traces of affected slab objects into reports,
enable ``CONFIG_STACKTRACE``. To include alloc and free stack traces of affected
physical pages, enable ``CONFIG_PAGE_OWNER`` and boot with ``page_owner=on``.
mm, page_owner: decouple freeing stack trace from debug_pagealloc Commit 8974558f49a6 ("mm, page_owner, debug_pagealloc: save and dump freeing stack trace") enhanced page_owner to also store freeing stack trace, when debug_pagealloc is also enabled. KASAN would also like to do this [1] to improve error reports to debug e.g. UAF issues. Kirill has suggested that the freeing stack trace saving should be also possible to be enabled separately from KASAN or debug_pagealloc, i.e. with an extra boot option. Qian argued that we have enough options already, and avoiding the extra overhead is not worth the complications in the case of a debugging option. Kirill noted that the extra stack handle in struct page_owner requires 0.1% of memory. This patch therefore enables free stack saving whenever page_owner is enabled, regardless of whether debug_pagealloc or KASAN is also enabled. KASAN kernels booted with page_owner=on will thus benefit from the improved error reports. [1] https://bugzilla.kernel.org/show_bug.cgi?id=203967 [vbabka@suse.cz: v3] Link: http://lkml.kernel.org/r/20191007091808.7096-3-vbabka@suse.cz Link: http://lkml.kernel.org/r/20190930122916.14969-3-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Reviewed-by: Qian Cai <cai@lca.pw> Suggested-by: Dmitry Vyukov <dvyukov@google.com> Suggested-by: Walter Wu <walter-zh.wu@mediatek.com> Suggested-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Suggested-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Suggested-by: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-15 05:11:44 +08:00
Error reports
~~~~~~~~~~~~~
A typical KASAN report looks like this::
==================================================================
BUG: KASAN: slab-out-of-bounds in kmalloc_oob_right+0xa8/0xbc [test_kasan]
Write of size 1 at addr ffff8801f44ec37b by task insmod/2760
CPU: 1 PID: 2760 Comm: insmod Not tainted 4.19.0-rc3+ #698
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
Call Trace:
dump_stack+0x94/0xd8
print_address_description+0x73/0x280
kasan_report+0x144/0x187
__asan_report_store1_noabort+0x17/0x20
kmalloc_oob_right+0xa8/0xbc [test_kasan]
kmalloc_tests_init+0x16/0x700 [test_kasan]
do_one_initcall+0xa5/0x3ae
do_init_module+0x1b6/0x547
load_module+0x75df/0x8070
__do_sys_init_module+0x1c6/0x200
__x64_sys_init_module+0x6e/0xb0
do_syscall_64+0x9f/0x2c0
entry_SYSCALL_64_after_hwframe+0x44/0xa9
RIP: 0033:0x7f96443109da
RSP: 002b:00007ffcf0b51b08 EFLAGS: 00000202 ORIG_RAX: 00000000000000af
RAX: ffffffffffffffda RBX: 000055dc3ee521a0 RCX: 00007f96443109da
RDX: 00007f96445cff88 RSI: 0000000000057a50 RDI: 00007f9644992000
RBP: 000055dc3ee510b0 R08: 0000000000000003 R09: 0000000000000000
R10: 00007f964430cd0a R11: 0000000000000202 R12: 00007f96445cff88
R13: 000055dc3ee51090 R14: 0000000000000000 R15: 0000000000000000
Allocated by task 2760:
save_stack+0x43/0xd0
kasan_kmalloc+0xa7/0xd0
kmem_cache_alloc_trace+0xe1/0x1b0
kmalloc_oob_right+0x56/0xbc [test_kasan]
kmalloc_tests_init+0x16/0x700 [test_kasan]
do_one_initcall+0xa5/0x3ae
do_init_module+0x1b6/0x547
load_module+0x75df/0x8070
__do_sys_init_module+0x1c6/0x200
__x64_sys_init_module+0x6e/0xb0
do_syscall_64+0x9f/0x2c0
entry_SYSCALL_64_after_hwframe+0x44/0xa9
Freed by task 815:
save_stack+0x43/0xd0
__kasan_slab_free+0x135/0x190
kasan_slab_free+0xe/0x10
kfree+0x93/0x1a0
umh_complete+0x6a/0xa0
call_usermodehelper_exec_async+0x4c3/0x640
ret_from_fork+0x35/0x40
The buggy address belongs to the object at ffff8801f44ec300
which belongs to the cache kmalloc-128 of size 128
The buggy address is located 123 bytes inside of
128-byte region [ffff8801f44ec300, ffff8801f44ec380)
The buggy address belongs to the page:
page:ffffea0007d13b00 count:1 mapcount:0 mapping:ffff8801f7001640 index:0x0
flags: 0x200000000000100(slab)
raw: 0200000000000100 ffffea0007d11dc0 0000001a0000001a ffff8801f7001640
raw: 0000000000000000 0000000080150015 00000001ffffffff 0000000000000000
page dumped because: kasan: bad access detected
Memory state around the buggy address:
ffff8801f44ec200: fc fc fc fc fc fc fc fc fb fb fb fb fb fb fb fb
ffff8801f44ec280: fb fb fb fb fb fb fb fb fc fc fc fc fc fc fc fc
>ffff8801f44ec300: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 03
^
ffff8801f44ec380: fc fc fc fc fc fc fc fc fb fb fb fb fb fb fb fb
ffff8801f44ec400: fb fb fb fb fb fb fb fb fc fc fc fc fc fc fc fc
==================================================================
The report header summarizes what kind of bug happened and what kind of access
caused it. It is followed by a stack trace of the bad access, a stack trace of
where the accessed memory was allocated (in case a slab object was accessed),
and a stack trace of where the object was freed (in case of a use-after-free
bug report). Next comes a description of the accessed slab object and the
information about the accessed memory page.
In the end, the report shows the memory state around the accessed address.
Internally, KASAN tracks memory state separately for each memory granule, which
is either 8 or 16 aligned bytes depending on KASAN mode. Each number in the
memory state section of the report shows the state of one of the memory
granules that surround the accessed address.
For generic KASAN, the size of each memory granule is 8. The state of each
granule is encoded in one shadow byte. Those 8 bytes can be accessible,
partially accessible, freed, or be a part of a redzone. KASAN uses the following
encoding for each shadow byte: 00 means that all 8 bytes of the corresponding
memory region are accessible; number N (1 <= N <= 7) means that the first N
bytes are accessible, and other (8 - N) bytes are not; any negative value
indicates that the entire 8-byte word is inaccessible. KASAN uses different
negative values to distinguish between different kinds of inaccessible memory
like redzones or freed memory (see mm/kasan/kasan.h).
In the report above, the arrow points to the shadow byte ``03``, which means
that the accessed address is partially accessible.
For tag-based KASAN modes, this last report section shows the memory tags around
the accessed address (see the `Implementation details`_ section).
Note that KASAN bug titles (like ``slab-out-of-bounds`` or ``use-after-free``)
are best-effort: KASAN prints the most probable bug type based on the limited
information it has. The actual type of the bug might be different.
Generic KASAN also reports up to two auxiliary call stack traces. These stack
traces point to places in code that interacted with the object but that are not
directly present in the bad access stack trace. Currently, this includes
call_rcu() and workqueue queuing.
Boot parameters
~~~~~~~~~~~~~~~
KASAN is affected by the generic ``panic_on_warn`` command line parameter.
When it is enabled, KASAN panics the kernel after printing a bug report.
By default, KASAN prints a bug report only for the first invalid memory access.
With ``kasan_multi_shot``, KASAN prints a report on every invalid access. This
effectively disables ``panic_on_warn`` for KASAN reports.
Alternatively, independent of ``panic_on_warn`` the ``kasan.fault=`` boot
parameter can be used to control panic and reporting behaviour:
- ``kasan.fault=report`` or ``=panic`` controls whether to only print a KASAN
report or also panic the kernel (default: ``report``). The panic happens even
if ``kasan_multi_shot`` is enabled.
Hardware tag-based KASAN mode (see the section about various modes below) is
intended for use in production as a security mitigation. Therefore, it supports
additional boot parameters that allow disabling KASAN or controlling features:
kasan: fix HW_TAGS boot parameters The initially proposed KASAN command line parameters are redundant. This change drops the complex "kasan.mode=off/prod/full" parameter and adds a simpler kill switch "kasan=off/on" instead. The new parameter together with the already existing ones provides a cleaner way to express the same set of features. The full set of parameters with this change: kasan=off/on - whether KASAN is enabled kasan.fault=report/panic - whether to only print a report or also panic kasan.stacktrace=off/on - whether to collect alloc/free stack traces Default values: kasan=on kasan.fault=report kasan.stacktrace=on (if CONFIG_DEBUG_KERNEL=y) kasan.stacktrace=off (otherwise) Link: https://linux-review.googlesource.com/id/Ib3694ed90b1e8ccac6cf77dfd301847af4aba7b8 Link: https://lkml.kernel.org/r/4e9c4a4bdcadc168317deb2419144582a9be6e61.1610736745.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Peter Collingbourne <pcc@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-01-24 13:01:34 +08:00
- ``kasan=off`` or ``=on`` controls whether KASAN is enabled (default: ``on``).
kasan: Add KASAN mode kernel parameter Architectures supported by KASAN_HW_TAGS can provide a sync or async mode of execution. On an MTE enabled arm64 hw for example this can be identified with the synchronous or asynchronous tagging mode of execution. In synchronous mode, an exception is triggered if a tag check fault occurs. In asynchronous mode, if a tag check fault occurs, the TFSR_EL1 register is updated asynchronously. The kernel checks the corresponding bits periodically. KASAN requires a specific kernel command line parameter to make use of this hw features. Add KASAN HW execution mode kernel command line parameter. Note: This patch adds the kasan.mode kernel parameter and the sync/async kernel command line options to enable the described features. [ Add a new var instead of exposing kasan_arg_mode to be consistent with flags for other command line arguments. ] Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Andrey Konovalov <andreyknvl@google.com> Acked-by: Catalin Marinas <catalin.marinas@arm.com> Acked-by: Andrey Konovalov <andreyknvl@google.com> Tested-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Link: https://lore.kernel.org/r/20210315132019.33202-3-vincenzo.frascino@arm.com Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
2021-03-15 21:20:12 +08:00
- ``kasan.mode=sync`` or ``=async`` controls whether KASAN is configured in
synchronous or asynchronous mode of execution (default: ``sync``).
Synchronous mode: a bad access is detected immediately when a tag
check fault occurs.
Asynchronous mode: a bad access detection is delayed. When a tag check
fault occurs, the information is stored in hardware (in the TFSR_EL1
register for arm64). The kernel periodically checks the hardware and
only reports tag faults during these checks.
kasan: fix HW_TAGS boot parameters The initially proposed KASAN command line parameters are redundant. This change drops the complex "kasan.mode=off/prod/full" parameter and adds a simpler kill switch "kasan=off/on" instead. The new parameter together with the already existing ones provides a cleaner way to express the same set of features. The full set of parameters with this change: kasan=off/on - whether KASAN is enabled kasan.fault=report/panic - whether to only print a report or also panic kasan.stacktrace=off/on - whether to collect alloc/free stack traces Default values: kasan=on kasan.fault=report kasan.stacktrace=on (if CONFIG_DEBUG_KERNEL=y) kasan.stacktrace=off (otherwise) Link: https://linux-review.googlesource.com/id/Ib3694ed90b1e8ccac6cf77dfd301847af4aba7b8 Link: https://lkml.kernel.org/r/4e9c4a4bdcadc168317deb2419144582a9be6e61.1610736745.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Peter Collingbourne <pcc@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-01-24 13:01:34 +08:00
- ``kasan.stacktrace=off`` or ``=on`` disables or enables alloc and free stack
kasan: fix stack traces dependency for HW_TAGS Currently, whether the alloc/free stack traces collection is enabled by default for hardware tag-based KASAN depends on CONFIG_DEBUG_KERNEL. The intention for this dependency was to only enable collection on slow debug kernels due to a significant perf and memory impact. As it turns out, CONFIG_DEBUG_KERNEL is not considered a debug option and is enabled on many productions kernels including Android and Ubuntu. As the result, this dependency is pointless and only complicates the code and documentation. Having stack traces collection disabled by default would make the hardware mode work differently to to the software ones, which is confusing. This change removes the dependency and enables stack traces collection by default. Looking into the future, this default might makes sense for production kernels, assuming we implement a fast stack trace collection approach. Link: https://lkml.kernel.org/r/6678d77ceffb71f1cff2cf61560e2ffe7bb6bfe9.1612808820.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Peter Collingbourne <pcc@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-10 05:42:03 +08:00
traces collection (default: ``on``).
Implementation details
----------------------
Generic KASAN
~~~~~~~~~~~~~
Software KASAN modes use shadow memory to record whether each byte of memory is
safe to access and use compile-time instrumentation to insert shadow memory
checks before each memory access.
Generic KASAN dedicates 1/8th of kernel memory to its shadow memory (16TB
to cover 128TB on x86_64) and uses direct mapping with a scale and offset to
translate a memory address to its corresponding shadow address.
Here is the function which translates an address to its corresponding shadow
address::
static inline void *kasan_mem_to_shadow(const void *addr)
{
return (void *)((unsigned long)addr >> KASAN_SHADOW_SCALE_SHIFT)
+ KASAN_SHADOW_OFFSET;
}
where ``KASAN_SHADOW_SCALE_SHIFT = 3``.
Compile-time instrumentation is used to insert memory access checks. Compiler
inserts function calls (``__asan_load*(addr)``, ``__asan_store*(addr)``) before
each memory access of size 1, 2, 4, 8, or 16. These functions check whether
memory accesses are valid or not by checking corresponding shadow memory.
With inline instrumentation, instead of making function calls, the compiler
directly inserts the code to check shadow memory. This option significantly
enlarges the kernel, but it gives an x1.1-x2 performance boost over the
outline-instrumented kernel.
Generic KASAN is the only mode that delays the reuse of freed objects via
quarantine (see mm/kasan/quarantine.c for implementation).
Software tag-based KASAN
~~~~~~~~~~~~~~~~~~~~~~~~
Software tag-based KASAN uses a software memory tagging approach to checking
access validity. It is currently only implemented for the arm64 architecture.
Software tag-based KASAN uses the Top Byte Ignore (TBI) feature of arm64 CPUs
to store a pointer tag in the top byte of kernel pointers. It uses shadow memory
to store memory tags associated with each 16-byte memory cell (therefore, it
dedicates 1/16th of the kernel memory for shadow memory).
On each memory allocation, software tag-based KASAN generates a random tag, tags
the allocated memory with this tag, and embeds the same tag into the returned
pointer.
Software tag-based KASAN uses compile-time instrumentation to insert checks
before each memory access. These checks make sure that the tag of the memory
that is being accessed is equal to the tag of the pointer that is used to access
this memory. In case of a tag mismatch, software tag-based KASAN prints a bug
report.
Software tag-based KASAN also has two instrumentation modes (outline, which
emits callbacks to check memory accesses; and inline, which performs the shadow
memory checks inline). With outline instrumentation mode, a bug report is
printed from the function that performs the access check. With inline
instrumentation, a ``brk`` instruction is emitted by the compiler, and a
dedicated ``brk`` handler is used to print bug reports.
Software tag-based KASAN uses 0xFF as a match-all pointer tag (accesses through
pointers with the 0xFF pointer tag are not checked). The value 0xFE is currently
reserved to tag freed memory regions.
Software tag-based KASAN currently only supports tagging of slab and page_alloc
memory.
Hardware tag-based KASAN
~~~~~~~~~~~~~~~~~~~~~~~~
Hardware tag-based KASAN is similar to the software mode in concept but uses
hardware memory tagging support instead of compiler instrumentation and
shadow memory.
Hardware tag-based KASAN is currently only implemented for arm64 architecture
and based on both arm64 Memory Tagging Extension (MTE) introduced in ARMv8.5
Instruction Set Architecture and Top Byte Ignore (TBI).
Special arm64 instructions are used to assign memory tags for each allocation.
Same tags are assigned to pointers to those allocations. On every memory
access, hardware makes sure that the tag of the memory that is being accessed is
equal to the tag of the pointer that is used to access this memory. In case of a
tag mismatch, a fault is generated, and a report is printed.
Hardware tag-based KASAN uses 0xFF as a match-all pointer tag (accesses through
pointers with the 0xFF pointer tag are not checked). The value 0xFE is currently
reserved to tag freed memory regions.
Hardware tag-based KASAN currently only supports tagging of slab and page_alloc
memory.
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
If the hardware does not support MTE (pre ARMv8.5), hardware tag-based KASAN
will not be enabled. In this case, all KASAN boot parameters are ignored.
Note that enabling CONFIG_KASAN_HW_TAGS always results in in-kernel TBI being
enabled. Even when ``kasan.mode=off`` is provided or when the hardware does not
support MTE (but supports TBI).
Hardware tag-based KASAN only reports the first found bug. After that, MTE tag
checking gets disabled.
Shadow memory
-------------
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
The kernel maps memory in several different parts of the address space.
The range of kernel virtual addresses is large: there is not enough real
memory to support a real shadow region for every address that could be
accessed by the kernel. Therefore, KASAN only maps real shadow for certain
parts of the address space.
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
Default behaviour
~~~~~~~~~~~~~~~~~
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
By default, architectures only map real memory over the shadow region
for the linear mapping (and potentially other small areas). For all
other areas - such as vmalloc and vmemmap space - a single read-only
page is mapped over the shadow area. This read-only shadow page
declares all memory accesses as permitted.
This presents a problem for modules: they do not live in the linear
mapping but in a dedicated module space. By hooking into the module
allocator, KASAN temporarily maps real shadow memory to cover them.
This allows detection of invalid accesses to module globals, for example.
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
This also creates an incompatibility with ``VMAP_STACK``: if the stack
lives in vmalloc space, it will be shadowed by the read-only page, and
the kernel will fault when trying to set up the shadow data for stack
variables.
CONFIG_KASAN_VMALLOC
~~~~~~~~~~~~~~~~~~~~
With ``CONFIG_KASAN_VMALLOC``, KASAN can cover vmalloc space at the
cost of greater memory usage. Currently, this is supported on x86,
riscv, s390, and powerpc.
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
This works by hooking into vmalloc and vmap and dynamically
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
allocating real shadow memory to back the mappings.
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
2020-12-23 04:00:24 +08:00
``KASAN_GRANULE_SIZE * PAGE_SIZE``.
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
Instead, KASAN shares backing space across multiple mappings. It allocates
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
a backing page when a mapping in vmalloc space uses a particular page
of the shadow region. This page can be shared by other vmalloc
mappings later on.
KASAN hooks into the vmap infrastructure to lazily clean up unused shadow
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
memory.
To avoid the difficulties around swapping mappings around, KASAN expects
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
that the part of the shadow region that covers the vmalloc space will
not be covered by the early shadow page but will be left unmapped.
This will require changes in arch-specific code.
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
This allows ``VMAP_STACK`` support on x86 and can simplify support of
kasan: support backing vmalloc space with real shadow memory Patch series "kasan: support backing vmalloc space with real shadow memory", v11. Currently, vmalloc space is backed by the early shadow page. This means that kasan is incompatible with VMAP_STACK. This series provides a mechanism to back vmalloc space with real, dynamically allocated memory. I have only wired up x86, because that's the only currently supported arch I can work with easily, but it's very easy to wire up other architectures, and it appears that there is some work-in-progress code to do this on arm64 and s390. This has been discussed before in the context of VMAP_STACK: - https://bugzilla.kernel.org/show_bug.cgi?id=202009 - https://lkml.org/lkml/2018/7/22/198 - https://lkml.org/lkml/2019/7/19/822 In terms of implementation details: Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=1) This is unfortunate but given that this is a debug feature only, not the end of the world. The benchmarks are also a stress-test for the vmalloc subsystem: they're not indicative of an overall 2x slowdown! This patch (of 4): Hook into vmalloc and vmap, and dynamically allocate real shadow memory to back the mappings. Most mappings in vmalloc space are small, requiring less than a full page of shadow space. Allocating a full shadow page per mapping would therefore be wasteful. Furthermore, to ensure that different mappings use different shadow pages, mappings would have to be aligned to KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE. Instead, share backing space across multiple mappings. Allocate a backing page when a mapping in vmalloc space uses a particular page of the shadow region. This page can be shared by other vmalloc mappings later on. We hook in to the vmap infrastructure to lazily clean up unused shadow memory. To avoid the difficulties around swapping mappings around, this code expects that the part of the shadow region that covers the vmalloc space will not be covered by the early shadow page, but will be left unmapped. This will require changes in arch-specific code. This allows KASAN with VMAP_STACK, and may be helpful for architectures that do not have a separate module space (e.g. powerpc64, which I am currently working on). It also allows relaxing the module alignment back to PAGE_SIZE. Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that: - Turning on KASAN, inline instrumentation, without vmalloc, introuduces a 4.1x-4.2x slowdown in vmalloc operations. - Turning this on introduces the following slowdowns over KASAN: * ~1.76x slower single-threaded (test_vmalloc.sh performance) * ~2.18x slower when both cpus are performing operations simultaneously (test_vmalloc.sh sequential_test_order=3D1) This is unfortunate but given that this is a debug feature only, not the end of the world. The full benchmark results are: Performance No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68 full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10 long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89 random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04 fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05 random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75 align_shift_alloc_test 147 830 5.65 5692 38.72 6.86 pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12 Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82 Sequential, 2 cpus No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94 full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02 long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05 random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58 fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50 random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16 align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08 pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43 Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11 fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94 full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03 long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06 random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58 fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49 random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15 align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57 pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10 Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11 [dja@axtens.net: fixups] Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009 Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework] Signed-off-by: Daniel Axtens <dja@axtens.net> Co-developed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Vasily Gorbik <gor@linux.ibm.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Christophe Leroy <christophe.leroy@c-s.fr> Cc: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
architectures that do not have a fixed module region.
For developers
--------------
Ignoring accesses
~~~~~~~~~~~~~~~~~
Software KASAN modes use compiler instrumentation to insert validity checks.
Such instrumentation might be incompatible with some parts of the kernel, and
therefore needs to be disabled.
Other parts of the kernel might access metadata for allocated objects.
Normally, KASAN detects and reports such accesses, but in some cases (e.g.,
in memory allocators), these accesses are valid.
For software KASAN modes, to disable instrumentation for a specific file or
directory, add a ``KASAN_SANITIZE`` annotation to the respective kernel
Makefile:
- For a single file (e.g., main.o)::
KASAN_SANITIZE_main.o := n
- For all files in one directory::
KASAN_SANITIZE := n
For software KASAN modes, to disable instrumentation on a per-function basis,
use the KASAN-specific ``__no_sanitize_address`` function attribute or the
generic ``noinstr`` one.
Note that disabling compiler instrumentation (either on a per-file or a
per-function basis) makes KASAN ignore the accesses that happen directly in
that code for software KASAN modes. It does not help when the accesses happen
indirectly (through calls to instrumented functions) or with the hardware
tag-based mode that does not use compiler instrumentation.
For software KASAN modes, to disable KASAN reports in a part of the kernel code
for the current task, annotate this part of the code with a
``kasan_disable_current()``/``kasan_enable_current()`` section. This also
disables the reports for indirect accesses that happen through function calls.
For tag-based KASAN modes (include the hardware one), to disable access
checking, use ``kasan_reset_tag()`` or ``page_kasan_tag_reset()``. Note that
temporarily disabling access checking via ``page_kasan_tag_reset()`` requires
saving and restoring the per-page KASAN tag via
``page_kasan_tag``/``page_kasan_tag_set``.
Tests
~~~~~
There are KASAN tests that allow verifying that KASAN works and can detect
certain types of memory corruptions. The tests consist of two parts:
1. Tests that are integrated with the KUnit Test Framework. Enabled with
``CONFIG_KASAN_KUNIT_TEST``. These tests can be run and partially verified
automatically in a few different ways; see the instructions below.
2. Tests that are currently incompatible with KUnit. Enabled with
``CONFIG_KASAN_MODULE_TEST`` and can only be run as a module. These tests can
only be verified manually by loading the kernel module and inspecting the
kernel log for KASAN reports.
Each KUnit-compatible KASAN test prints one of multiple KASAN reports if an
error is detected. Then the test prints its number and status.
When a test passes::
ok 28 - kmalloc_double_kzfree
When a test fails due to a failed ``kmalloc``::
# kmalloc_large_oob_right: ASSERTION FAILED at lib/test_kasan.c:163
Expected ptr is not null, but is
not ok 4 - kmalloc_large_oob_right
When a test fails due to a missing KASAN report::
kasan: test: improve failure message in KUNIT_EXPECT_KASAN_FAIL() The KUNIT_EXPECT_KASAN_FAIL() macro currently uses KUNIT_EXPECT_EQ() to compare fail_data.report_expected and fail_data.report_found. This always gave a somewhat useless error message on failure, but the addition of extra compile-time checking with READ_ONCE() has caused it to get much longer, and be truncated before anything useful is displayed. Instead, just check fail_data.report_found by hand (we've just set report_expected to 'true'), and print a better failure message with KUNIT_FAIL(). Because of this, report_expected is no longer used anywhere, and can be removed. Beforehand, a failure in: KUNIT_EXPECT_KASAN_FAIL(test, ((volatile char *)area)[3100]); would have looked like: [22:00:34] [FAILED] vmalloc_oob [22:00:34] # vmalloc_oob: EXPECTATION FAILED at lib/test_kasan.c:991 [22:00:34] Expected ({ do { extern void __compiletime_assert_705(void) __attribute__((__error__("Unsupported access size for {READ,WRITE}_ONCE()."))); if (!((sizeof(fail_data.report_expected) == sizeof(char) || sizeof(fail_data.repp [22:00:34] not ok 45 - vmalloc_oob With this change, it instead looks like: [22:04:04] [FAILED] vmalloc_oob [22:04:04] # vmalloc_oob: EXPECTATION FAILED at lib/test_kasan.c:993 [22:04:04] KASAN failure expected in "((volatile char *)area)[3100]", but none occurred [22:04:04] not ok 45 - vmalloc_oob Also update the example failure in the documentation to reflect this. Link: https://lkml.kernel.org/r/20210606005531.165954-1-davidgow@google.com Signed-off-by: David Gow <davidgow@google.com> Reviewed-by: Andrey Konovalov <andreyknvl@gmail.com> Reviewed-by: Marco Elver <elver@google.com> Acked-by: Brendan Higgins <brendanhiggins@google.com> Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Daniel Axtens <dja@axtens.net> Cc: David Gow <davidgow@google.com> Cc: Jonathan Corbet <corbet@lwn.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 10:40:36 +08:00
# kmalloc_double_kzfree: EXPECTATION FAILED at lib/test_kasan.c:974
KASAN failure expected in "kfree_sensitive(ptr)", but none occurred
not ok 44 - kmalloc_double_kzfree
At the end the cumulative status of all KASAN tests is printed. On success::
ok 1 - kasan
Or, if one of the tests failed::
not ok 1 - kasan
There are a few ways to run KUnit-compatible KASAN tests.
1. Loadable module
With ``CONFIG_KUNIT`` enabled, KASAN-KUnit tests can be built as a loadable
module and run by loading ``test_kasan.ko`` with ``insmod`` or ``modprobe``.
2. Built-In
With ``CONFIG_KUNIT`` built-in, KASAN-KUnit tests can be built-in as well.
In this case, the tests will run at boot as a late-init call.
3. Using kunit_tool
With ``CONFIG_KUNIT`` and ``CONFIG_KASAN_KUNIT_TEST`` built-in, it is also
possible to use ``kunit_tool`` to see the results of KUnit tests in a more
readable way. This will not print the KASAN reports of the tests that passed.
See `KUnit documentation <https://www.kernel.org/doc/html/latest/dev-tools/kunit/index.html>`_
for more up-to-date information on ``kunit_tool``.
.. _KUnit: https://www.kernel.org/doc/html/latest/dev-tools/kunit/index.html