OpenCloudOS-Kernel/include/linux/fortify-string.h

478 lines
16 KiB
C
Raw Normal View History

/* SPDX-License-Identifier: GPL-2.0 */
#ifndef _LINUX_FORTIFY_STRING_H_
#define _LINUX_FORTIFY_STRING_H_
#include <linux/const.h>
#define __FORTIFY_INLINE extern __always_inline __gnu_inline __overloadable
#define __RENAME(x) __asm__(#x)
void fortify_panic(const char *name) __noreturn __cold;
void __read_overflow(void) __compiletime_error("detected read beyond size of object (1st parameter)");
void __read_overflow2(void) __compiletime_error("detected read beyond size of object (2nd parameter)");
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
void __read_overflow2_field(size_t avail, size_t wanted) __compiletime_warning("detected read beyond size of field (2nd parameter); maybe use struct_group()?");
void __write_overflow(void) __compiletime_error("detected write beyond size of object (1st parameter)");
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
void __write_overflow_field(size_t avail, size_t wanted) __compiletime_warning("detected write beyond size of field (1st parameter); maybe use struct_group()?");
#define __compiletime_strlen(p) \
({ \
unsigned char *__p = (unsigned char *)(p); \
size_t __ret = (size_t)-1; \
size_t __p_size = __builtin_object_size(p, 1); \
if (__p_size != (size_t)-1) { \
size_t __p_len = __p_size - 1; \
if (__builtin_constant_p(__p[__p_len]) && \
__p[__p_len] == '\0') \
__ret = __builtin_strlen(__p); \
} \
__ret; \
})
#if defined(CONFIG_KASAN_GENERIC) || defined(CONFIG_KASAN_SW_TAGS)
extern void *__underlying_memchr(const void *p, int c, __kernel_size_t size) __RENAME(memchr);
extern int __underlying_memcmp(const void *p, const void *q, __kernel_size_t size) __RENAME(memcmp);
extern void *__underlying_memcpy(void *p, const void *q, __kernel_size_t size) __RENAME(memcpy);
extern void *__underlying_memmove(void *p, const void *q, __kernel_size_t size) __RENAME(memmove);
extern void *__underlying_memset(void *p, int c, __kernel_size_t size) __RENAME(memset);
extern char *__underlying_strcat(char *p, const char *q) __RENAME(strcat);
extern char *__underlying_strcpy(char *p, const char *q) __RENAME(strcpy);
extern __kernel_size_t __underlying_strlen(const char *p) __RENAME(strlen);
extern char *__underlying_strncat(char *p, const char *q, __kernel_size_t count) __RENAME(strncat);
extern char *__underlying_strncpy(char *p, const char *q, __kernel_size_t size) __RENAME(strncpy);
#else
#define __underlying_memchr __builtin_memchr
#define __underlying_memcmp __builtin_memcmp
#define __underlying_memcpy __builtin_memcpy
#define __underlying_memmove __builtin_memmove
#define __underlying_memset __builtin_memset
#define __underlying_strcat __builtin_strcat
#define __underlying_strcpy __builtin_strcpy
#define __underlying_strlen __builtin_strlen
#define __underlying_strncat __builtin_strncat
#define __underlying_strncpy __builtin_strncpy
#endif
/*
* Clang's use of __builtin_object_size() within inlines needs hinting via
* __pass_object_size(). The preference is to only ever use type 1 (member
* size, rather than struct size), but there remain some stragglers using
* type 0 that will be converted in the future.
*/
#define POS __pass_object_size(1)
#define POS0 __pass_object_size(0)
__FORTIFY_INLINE __diagnose_as(__builtin_strncpy, 1, 2, 3)
char *strncpy(char * const POS p, const char *q, __kernel_size_t size)
{
size_t p_size = __builtin_object_size(p, 1);
if (__builtin_constant_p(size) && p_size < size)
__write_overflow();
if (p_size < size)
fortify_panic(__func__);
return __underlying_strncpy(p, q, size);
}
__FORTIFY_INLINE __diagnose_as(__builtin_strcat, 1, 2)
char *strcat(char * const POS p, const char *q)
{
size_t p_size = __builtin_object_size(p, 1);
if (p_size == (size_t)-1)
return __underlying_strcat(p, q);
if (strlcat(p, q, p_size) >= p_size)
fortify_panic(__func__);
return p;
}
extern __kernel_size_t __real_strnlen(const char *, __kernel_size_t) __RENAME(strnlen);
__FORTIFY_INLINE __kernel_size_t strnlen(const char * const POS p, __kernel_size_t maxlen)
{
size_t p_size = __builtin_object_size(p, 1);
size_t p_len = __compiletime_strlen(p);
size_t ret;
/* We can take compile-time actions when maxlen is const. */
if (__builtin_constant_p(maxlen) && p_len != (size_t)-1) {
/* If p is const, we can use its compile-time-known len. */
if (maxlen >= p_size)
return p_len;
}
/* Do not check characters beyond the end of p. */
ret = __real_strnlen(p, maxlen < p_size ? maxlen : p_size);
if (p_size <= ret && maxlen != ret)
fortify_panic(__func__);
return ret;
}
/*
* Defined after fortified strnlen to reuse it. However, it must still be
* possible for strlen() to be used on compile-time strings for use in
* static initializers (i.e. as a constant expression).
*/
#define strlen(p) \
__builtin_choose_expr(__is_constexpr(__builtin_strlen(p)), \
__builtin_strlen(p), __fortify_strlen(p))
__FORTIFY_INLINE __diagnose_as(__builtin_strlen, 1)
__kernel_size_t __fortify_strlen(const char * const POS p)
{
__kernel_size_t ret;
size_t p_size = __builtin_object_size(p, 1);
/* Give up if we don't know how large p is. */
if (p_size == (size_t)-1)
return __underlying_strlen(p);
ret = strnlen(p, p_size);
if (p_size <= ret)
fortify_panic(__func__);
return ret;
}
/* defined after fortified strlen to reuse it */
extern size_t __real_strlcpy(char *, const char *, size_t) __RENAME(strlcpy);
__FORTIFY_INLINE size_t strlcpy(char * const POS p, const char * const POS q, size_t size)
{
size_t p_size = __builtin_object_size(p, 1);
size_t q_size = __builtin_object_size(q, 1);
size_t q_len; /* Full count of source string length. */
size_t len; /* Count of characters going into destination. */
if (p_size == (size_t)-1 && q_size == (size_t)-1)
return __real_strlcpy(p, q, size);
q_len = strlen(q);
len = (q_len >= size) ? size - 1 : q_len;
if (__builtin_constant_p(size) && __builtin_constant_p(q_len) && size) {
/* Write size is always larger than destination. */
if (len >= p_size)
__write_overflow();
}
if (size) {
if (len >= p_size)
fortify_panic(__func__);
__underlying_memcpy(p, q, len);
p[len] = '\0';
}
return q_len;
}
/* defined after fortified strnlen to reuse it */
extern ssize_t __real_strscpy(char *, const char *, size_t) __RENAME(strscpy);
__FORTIFY_INLINE ssize_t strscpy(char * const POS p, const char * const POS q, size_t size)
{
size_t len;
/* Use string size rather than possible enclosing struct size. */
size_t p_size = __builtin_object_size(p, 1);
size_t q_size = __builtin_object_size(q, 1);
/* If we cannot get size of p and q default to call strscpy. */
if (p_size == (size_t) -1 && q_size == (size_t) -1)
return __real_strscpy(p, q, size);
/*
* If size can be known at compile time and is greater than
* p_size, generate a compile time write overflow error.
*/
if (__builtin_constant_p(size) && size > p_size)
__write_overflow();
/*
* This call protects from read overflow, because len will default to q
* length if it smaller than size.
*/
len = strnlen(q, size);
/*
* If len equals size, we will copy only size bytes which leads to
* -E2BIG being returned.
* Otherwise we will copy len + 1 because of the final '\O'.
*/
len = len == size ? size : len + 1;
/*
* Generate a runtime write overflow error if len is greater than
* p_size.
*/
if (len > p_size)
fortify_panic(__func__);
/*
* We can now safely call vanilla strscpy because we are protected from:
* 1. Read overflow thanks to call to strnlen().
* 2. Write overflow thanks to above ifs.
*/
return __real_strscpy(p, q, len);
}
/* defined after fortified strlen and strnlen to reuse them */
__FORTIFY_INLINE __diagnose_as(__builtin_strncat, 1, 2, 3)
char *strncat(char * const POS p, const char * const POS q, __kernel_size_t count)
{
size_t p_len, copy_len;
size_t p_size = __builtin_object_size(p, 1);
size_t q_size = __builtin_object_size(q, 1);
if (p_size == (size_t)-1 && q_size == (size_t)-1)
return __underlying_strncat(p, q, count);
p_len = strlen(p);
copy_len = strnlen(q, count);
if (p_size < p_len + copy_len + 1)
fortify_panic(__func__);
__underlying_memcpy(p + p_len, q, copy_len);
p[p_len + copy_len] = '\0';
return p;
}
__FORTIFY_INLINE void fortify_memset_chk(__kernel_size_t size,
const size_t p_size,
const size_t p_size_field)
{
if (__builtin_constant_p(size)) {
/*
* Length argument is a constant expression, so we
* can perform compile-time bounds checking where
* buffer sizes are known.
*/
/* Error when size is larger than enclosing struct. */
if (p_size > p_size_field && p_size < size)
__write_overflow();
/* Warn when write size is larger than dest field. */
if (p_size_field < size)
__write_overflow_field(p_size_field, size);
}
/*
* At this point, length argument may not be a constant expression,
* so run-time bounds checking can be done where buffer sizes are
* known. (This is not an "else" because the above checks may only
* be compile-time warnings, and we want to still warn for run-time
* overflows.)
*/
/*
* Always stop accesses beyond the struct that contains the
* field, when the buffer's remaining size is known.
* (The -1 test is to optimize away checks where the buffer
* lengths are unknown.)
*/
if (p_size != (size_t)(-1) && p_size < size)
fortify_panic("memset");
}
#define __fortify_memset_chk(p, c, size, p_size, p_size_field) ({ \
size_t __fortify_size = (size_t)(size); \
fortify_memset_chk(__fortify_size, p_size, p_size_field), \
__underlying_memset(p, c, __fortify_size); \
})
/*
* __builtin_object_size() must be captured here to avoid evaluating argument
* side-effects further into the macro layers.
*/
#define memset(p, c, s) __fortify_memset_chk(p, c, s, \
__builtin_object_size(p, 0), __builtin_object_size(p, 1))
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
/*
* To make sure the compiler can enforce protection against buffer overflows,
* memcpy(), memmove(), and memset() must not be used beyond individual
* struct members. If you need to copy across multiple members, please use
* struct_group() to create a named mirror of an anonymous struct union.
* (e.g. see struct sk_buff.) Read overflow checking is currently only
* done when a write overflow is also present, or when building with W=1.
*
* Mitigation coverage matrix
* Bounds checking at:
* +-------+-------+-------+-------+
* | Compile time | Run time |
* memcpy() argument sizes: | write | read | write | read |
* dest source length +-------+-------+-------+-------+
* memcpy(known, known, constant) | y | y | n/a | n/a |
* memcpy(known, unknown, constant) | y | n | n/a | V |
* memcpy(known, known, dynamic) | n | n | B | B |
* memcpy(known, unknown, dynamic) | n | n | B | V |
* memcpy(unknown, known, constant) | n | y | V | n/a |
* memcpy(unknown, unknown, constant) | n | n | V | V |
* memcpy(unknown, known, dynamic) | n | n | V | B |
* memcpy(unknown, unknown, dynamic) | n | n | V | V |
* +-------+-------+-------+-------+
*
* y = perform deterministic compile-time bounds checking
* n = cannot perform deterministic compile-time bounds checking
* n/a = no run-time bounds checking needed since compile-time deterministic
* B = can perform run-time bounds checking (currently unimplemented)
* V = vulnerable to run-time overflow (will need refactoring to solve)
*
*/
__FORTIFY_INLINE void fortify_memcpy_chk(__kernel_size_t size,
const size_t p_size,
const size_t q_size,
const size_t p_size_field,
const size_t q_size_field,
const char *func)
{
if (__builtin_constant_p(size)) {
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
/*
* Length argument is a constant expression, so we
* can perform compile-time bounds checking where
* buffer sizes are known.
*/
/* Error when size is larger than enclosing struct. */
if (p_size > p_size_field && p_size < size)
__write_overflow();
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
if (q_size > q_size_field && q_size < size)
__read_overflow2();
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
/* Warn when write size argument larger than dest field. */
if (p_size_field < size)
__write_overflow_field(p_size_field, size);
/*
* Warn for source field over-read when building with W=1
* or when an over-write happened, so both can be fixed at
* the same time.
*/
if ((IS_ENABLED(KBUILD_EXTRA_WARN1) || p_size_field < size) &&
q_size_field < size)
__read_overflow2_field(q_size_field, size);
}
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
/*
* At this point, length argument may not be a constant expression,
* so run-time bounds checking can be done where buffer sizes are
* known. (This is not an "else" because the above checks may only
* be compile-time warnings, and we want to still warn for run-time
* overflows.)
*/
/*
* Always stop accesses beyond the struct that contains the
* field, when the buffer's remaining size is known.
* (The -1 test is to optimize away checks where the buffer
* lengths are unknown.)
*/
if ((p_size != (size_t)(-1) && p_size < size) ||
(q_size != (size_t)(-1) && q_size < size))
fortify_panic(func);
}
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
#define __fortify_memcpy_chk(p, q, size, p_size, q_size, \
p_size_field, q_size_field, op) ({ \
size_t __fortify_size = (size_t)(size); \
fortify_memcpy_chk(__fortify_size, p_size, q_size, \
p_size_field, q_size_field, #op); \
__underlying_##op(p, q, __fortify_size); \
})
/*
* __builtin_object_size() must be captured here to avoid evaluating argument
* side-effects further into the macro layers.
*/
#define memcpy(p, q, s) __fortify_memcpy_chk(p, q, s, \
__builtin_object_size(p, 0), __builtin_object_size(q, 0), \
__builtin_object_size(p, 1), __builtin_object_size(q, 1), \
memcpy)
#define memmove(p, q, s) __fortify_memcpy_chk(p, q, s, \
__builtin_object_size(p, 0), __builtin_object_size(q, 0), \
__builtin_object_size(p, 1), __builtin_object_size(q, 1), \
memmove)
extern void *__real_memscan(void *, int, __kernel_size_t) __RENAME(memscan);
__FORTIFY_INLINE void *memscan(void * const POS0 p, int c, __kernel_size_t size)
{
size_t p_size = __builtin_object_size(p, 0);
if (__builtin_constant_p(size) && p_size < size)
__read_overflow();
if (p_size < size)
fortify_panic(__func__);
return __real_memscan(p, c, size);
}
__FORTIFY_INLINE __diagnose_as(__builtin_memcmp, 1, 2, 3)
int memcmp(const void * const POS0 p, const void * const POS0 q, __kernel_size_t size)
{
size_t p_size = __builtin_object_size(p, 0);
size_t q_size = __builtin_object_size(q, 0);
if (__builtin_constant_p(size)) {
if (p_size < size)
__read_overflow();
if (q_size < size)
__read_overflow2();
}
if (p_size < size || q_size < size)
fortify_panic(__func__);
return __underlying_memcmp(p, q, size);
}
__FORTIFY_INLINE __diagnose_as(__builtin_memchr, 1, 2, 3)
void *memchr(const void * const POS0 p, int c, __kernel_size_t size)
{
size_t p_size = __builtin_object_size(p, 0);
if (__builtin_constant_p(size) && p_size < size)
__read_overflow();
if (p_size < size)
fortify_panic(__func__);
return __underlying_memchr(p, c, size);
}
void *__real_memchr_inv(const void *s, int c, size_t n) __RENAME(memchr_inv);
__FORTIFY_INLINE void *memchr_inv(const void * const POS0 p, int c, size_t size)
{
size_t p_size = __builtin_object_size(p, 0);
if (__builtin_constant_p(size) && p_size < size)
__read_overflow();
if (p_size < size)
fortify_panic(__func__);
return __real_memchr_inv(p, c, size);
}
extern void *__real_kmemdup(const void *src, size_t len, gfp_t gfp) __RENAME(kmemdup);
__FORTIFY_INLINE void *kmemdup(const void * const POS0 p, size_t size, gfp_t gfp)
{
size_t p_size = __builtin_object_size(p, 0);
if (__builtin_constant_p(size) && p_size < size)
__read_overflow();
if (p_size < size)
fortify_panic(__func__);
return __real_kmemdup(p, size, gfp);
}
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
/* Defined after fortified strlen to reuse it. */
__FORTIFY_INLINE __diagnose_as(__builtin_strcpy, 1, 2)
char *strcpy(char * const POS p, const char * const POS q)
{
size_t p_size = __builtin_object_size(p, 1);
size_t q_size = __builtin_object_size(q, 1);
size_t size;
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
/* If neither buffer size is known, immediately give up. */
if (p_size == (size_t)-1 && q_size == (size_t)-1)
return __underlying_strcpy(p, q);
size = strlen(q) + 1;
/* Compile-time check for const size overflow. */
if (__builtin_constant_p(size) && p_size < size)
__write_overflow();
/* Run-time check for dynamic size overflow. */
if (p_size < size)
fortify_panic(__func__);
fortify: Detect struct member overflows in memcpy() at compile-time memcpy() is dead; long live memcpy() tl;dr: In order to eliminate a large class of common buffer overflow flaws that continue to persist in the kernel, have memcpy() (under CONFIG_FORTIFY_SOURCE) perform bounds checking of the destination struct member when they have a known size. This would have caught all of the memcpy()-related buffer write overflow flaws identified in at least the last three years. Background and analysis: While stack-based buffer overflow flaws are largely mitigated by stack canaries (and similar) features, heap-based buffer overflow flaws continue to regularly appear in the kernel. Many classes of heap buffer overflows are mitigated by FORTIFY_SOURCE when using the strcpy() family of functions, but a significant number remain exposed through the memcpy() family of functions. At its core, FORTIFY_SOURCE uses the compiler's __builtin_object_size() internal[0] to determine the available size at a target address based on the compile-time known structure layout details. It operates in two modes: outer bounds (0) and inner bounds (1). In mode 0, the size of the enclosing structure is used. In mode 1, the size of the specific field is used. For example: struct object { u16 scalar1; /* 2 bytes */ char array[6]; /* 6 bytes */ u64 scalar2; /* 8 bytes */ u32 scalar3; /* 4 bytes */ u32 scalar4; /* 4 bytes */ } instance; __builtin_object_size(instance.array, 0) == 22, since the remaining size of the enclosing structure starting from "array" is 22 bytes (6 + 8 + 4 + 4). __builtin_object_size(instance.array, 1) == 6, since the remaining size of the specific field "array" is 6 bytes. The initial implementation of FORTIFY_SOURCE used mode 0 because there were many cases of both strcpy() and memcpy() functions being used to write (or read) across multiple fields in a structure. For example, it would catch this, which is writing 2 bytes beyond the end of "instance": memcpy(&instance.array, data, 25); While this didn't protect against overwriting adjacent fields in a given structure, it would at least stop overflows from reaching beyond the end of the structure into neighboring memory, and provided a meaningful mitigation of a subset of buffer overflow flaws. However, many desirable targets remain within the enclosing structure (for example function pointers). As it happened, there were very few cases of strcpy() family functions intentionally writing beyond the end of a string buffer. Once all known cases were removed from the kernel, the strcpy() family was tightened[1] to use mode 1, providing greater mitigation coverage. What remains is switching memcpy() to mode 1 as well, but making the switch is much more difficult because of how frustrating it can be to find existing "normal" uses of memcpy() that expect to write (or read) across multiple fields. The root cause of the problem is that the C language lacks a common pattern to indicate the intent of an author's use of memcpy(), and is further complicated by the available compile-time and run-time mitigation behaviors. The FORTIFY_SOURCE mitigation comes in two halves: the compile-time half, when both the buffer size _and_ the length of the copy is known, and the run-time half, when only the buffer size is known. If neither size is known, there is no bounds checking possible. At compile-time when the compiler sees that a length will always exceed a known buffer size, a warning can be deterministically emitted. For the run-time half, the length is tested against the known size of the buffer, and the overflowing operation is detected. (The performance overhead for these tests is virtually zero.) It is relatively easy to find compile-time false-positives since a warning is always generated. Fixing the false positives, however, can be very time-consuming as there are hundreds of instances. While it's possible some over-read conditions could lead to kernel memory exposures, the bulk of the risk comes from the run-time flaws where the length of a write may end up being attacker-controlled and lead to an overflow. Many of the compile-time false-positives take a form similar to this: memcpy(&instance.scalar2, data, sizeof(instance.scalar2) + sizeof(instance.scalar3)); and the run-time ones are similar, but lack a constant expression for the size of the copy: memcpy(instance.array, data, length); The former is meant to cover multiple fields (though its style has been frowned upon more recently), but has been technically legal. Both lack any expressivity in the C language about the author's _intent_ in a way that a compiler can check when the length isn't known at compile time. A comment doesn't work well because what's needed is something a compiler can directly reason about. Is a given memcpy() call expected to overflow into neighbors? Is it not? By using the new struct_group() macro, this intent can be much more easily encoded. It is not as easy to find the run-time false-positives since the code path to exercise a seemingly out-of-bounds condition that is actually expected may not be trivially reachable. Tightening the restrictions to block an operation for a false positive will either potentially create a greater flaw (if a copy is truncated by the mitigation), or destabilize the kernel (e.g. with a BUG()), making things completely useless for the end user. As a result, tightening the memcpy() restriction (when there is a reasonable level of uncertainty of the number of false positives), needs to first WARN() with no truncation. (Though any sufficiently paranoid end-user can always opt to set the panic_on_warn=1 sysctl.) Once enough development time has passed, the mitigation can be further intensified. (Note that this patch is only the compile-time checking step, which is a prerequisite to doing run-time checking, which will come in future patches.) Given the potential frustrations of weeding out all the false positives when tightening the run-time checks, it is reasonable to wonder if these changes would actually add meaningful protection. Looking at just the last three years, there are 23 identified flaws with a CVE that mention "buffer overflow", and 11 are memcpy()-related buffer overflows. (For the remaining 12: 7 are array index overflows that would be mitigated by systems built with CONFIG_UBSAN_BOUNDS=y: CVE-2019-0145, CVE-2019-14835, CVE-2019-14896, CVE-2019-14897, CVE-2019-14901, CVE-2019-17666, CVE-2021-28952. 2 are miscalculated allocation sizes which could be mitigated with memory tagging: CVE-2019-16746, CVE-2019-2181. 1 is an iovec buffer bug maybe mitigated by memory tagging: CVE-2020-10742. 1 is a type confusion bug mitigated by stack canaries: CVE-2020-10942. 1 is a string handling logic bug with no mitigation I'm aware of: CVE-2021-28972.) At my last count on an x86_64 allmodconfig build, there are 35,294 calls to memcpy(). With callers instrumented to report all places where the buffer size is known but the length remains unknown (i.e. a run-time bounds check is added), we can count how many new run-time bounds checks are added when the destination and source arguments of memcpy() are changed to use "mode 1" bounds checking: 1,276. This means for the future run-time checking, there is a worst-case upper bounds of 3.6% false positives to fix. In addition, there were around 150 new compile-time warnings to evaluate and fix (which have now been fixed). With this instrumentation it's also possible to compare the places where the known 11 memcpy() flaw overflows manifested against the resulting list of potential new run-time bounds checks, as a measure of potential efficacy of the tightened mitigation. Much to my surprise, horror, and delight, all 11 flaws would have been detected by the newly added run-time bounds checks, making this a distinctly clear mitigation improvement: 100% coverage for known memcpy() flaws, with a possible 2 orders of magnitude gain in coverage over existing but undiscovered run-time dynamic length flaws (i.e. 1265 newly covered sites in addition to the 11 known), against only <4% of all memcpy() callers maybe gaining a false positive run-time check, with only about 150 new compile-time instances needing evaluation. Specifically these would have been mitigated: CVE-2020-24490 https://git.kernel.org/linus/a2ec905d1e160a33b2e210e45ad30445ef26ce0e CVE-2020-12654 https://git.kernel.org/linus/3a9b153c5591548612c3955c9600a98150c81875 CVE-2020-12653 https://git.kernel.org/linus/b70261a288ea4d2f4ac7cd04be08a9f0f2de4f4d CVE-2019-14895 https://git.kernel.org/linus/3d94a4a8373bf5f45cf5f939e88b8354dbf2311b CVE-2019-14816 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14815 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-14814 https://git.kernel.org/linus/7caac62ed598a196d6ddf8d9c121e12e082cac3a CVE-2019-10126 https://git.kernel.org/linus/69ae4f6aac1578575126319d3f55550e7e440449 CVE-2019-9500 https://git.kernel.org/linus/1b5e2423164b3670e8bc9174e4762d297990deff no-CVE-yet https://git.kernel.org/linus/130f634da1af649205f4a3dd86cbe5c126b57914 no-CVE-yet https://git.kernel.org/linus/d10a87a3535cce2b890897914f5d0d83df669c63 To accelerate the review of potential run-time false positives, it's also worth noting that it is possible to partially automate checking by examining the memcpy() buffer argument to check for the destination struct member having a neighboring array member. It is reasonable to expect that the vast majority of run-time false positives would look like the already evaluated and fixed compile-time false positives, where the most common pattern is neighboring arrays. (And, FWIW, many of the compile-time fixes were actual bugs, so it is reasonable to assume we'll have similar cases of actual bugs getting fixed for run-time checks.) Implementation: Tighten the memcpy() destination buffer size checking to use the actual ("mode 1") target buffer size as the bounds check instead of their enclosing structure's ("mode 0") size. Use a common inline for memcpy() (and memmove() in a following patch), since all the tests are the same. All new cross-field memcpy() uses must use the struct_group() macro or similar to target a specific range of fields, so that FORTIFY_SOURCE can reason about the size and safety of the copy. For now, cross-member "mode 1" _read_ detection at compile-time will be limited to W=1 builds, since it is, unfortunately, very common. As the priority is solving write overflows, read overflows will be part of a future phase (and can be fixed in parallel, for anyone wanting to look at W=1 build output). For run-time, the "mode 0" size checking and mitigation is left unchanged, with "mode 1" to be added in stages. In this patch, no new run-time checks are added. Future patches will first bounds-check writes, and only perform a WARN() for now. This way any missed run-time false positives can be flushed out over the coming several development cycles, but system builders who have tested their workloads to be WARN()-free can enable the panic_on_warn=1 sysctl to immediately gain a mitigation against this class of buffer overflows. Once that is under way, run-time bounds-checking of reads can be similarly enabled. Related classes of flaws that will remain unmitigated: - memcpy() with flexible array structures, as the compiler does not currently have visibility into the size of the trailing flexible array. These can be fixed in the future by refactoring such cases to use a new set of flexible array structure helpers to perform the common serialization/deserialization code patterns doing allocation and/or copying. - memcpy() with raw pointers (e.g. void *, char *, etc), or otherwise having their buffer size unknown at compile time, have no good mitigation beyond memory tagging (and even that would only protect against inter-object overflow, not intra-object neighboring field overflows), or refactoring. Some kind of "fat pointer" solution is likely needed to gain proper size-of-buffer awareness. (e.g. see struct membuf) - type confusion where a higher level type's allocation size does not match the resulting cast type eventually passed to a deeper memcpy() call where the compiler cannot see the true type. In theory, greater static analysis could catch these, and the use of -Warray-bounds will help find some of these. [0] https://gcc.gnu.org/onlinedocs/gcc/Object-Size-Checking.html [1] https://git.kernel.org/linus/6a39e62abbafd1d58d1722f40c7d26ef379c6a2f Signed-off-by: Kees Cook <keescook@chromium.org>
2021-04-21 14:22:52 +08:00
__underlying_memcpy(p, q, size);
return p;
}
/* Don't use these outside the FORITFY_SOURCE implementation */
#undef __underlying_memchr
#undef __underlying_memcmp
#undef __underlying_strcat
#undef __underlying_strcpy
#undef __underlying_strlen
#undef __underlying_strncat
#undef __underlying_strncpy
#undef POS
#undef POS0
#endif /* _LINUX_FORTIFY_STRING_H_ */