| Age | Commit message (Collapse) | Author |
|---|
|
Previously different architectures were using random sources of
differing strength and cost to decide the random kstack offset. A number
of architectures (loongarch, powerpc, s390, x86) were using their
timestamp counter, at whatever the frequency happened to be. Other
arches (arm64, riscv) were using entropy from the crng via
get_random_u16().
There have been concerns that in some cases the timestamp counters may
be too weak, because they can be easily guessed or influenced by user
space. And get_random_u16() has been shown to be too costly for the
level of protection kstack offset randomization provides.
So let's use a common, architecture-agnostic source of entropy; a
per-cpu prng, seeded at boot-time from the crng. This has a few
benefits:
- We can remove choose_random_kstack_offset(); That was only there to
try to make the timestamp counter value a bit harder to influence
from user space [*].
- The architecture code is simplified. All it has to do now is call
add_random_kstack_offset() in the syscall path.
- The strength of the randomness can be reasoned about independently
of the architecture.
- Arches previously using get_random_u16() now have much faster
syscall paths, see below results.
[*] Additionally, this gets rid of some redundant work on s390 and x86.
Before this patch, those architectures called
choose_random_kstack_offset() under arch_exit_to_user_mode_prepare(),
which is also called for exception returns to userspace which were *not*
syscalls (e.g. regular interrupts). Getting rid of
choose_random_kstack_offset() avoids a small amount of redundant work
for the non-syscall cases.
In some configurations, add_random_kstack_offset() will now call
instrumentable code, so for a couple of arches, I have moved the call a
bit later to the first point where instrumentation is allowed. This
doesn't impact the efficacy of the mechanism.
There have been some claims that a prng may be less strong than the
timestamp counter if not regularly reseeded. But the prng has a period
of about 2^113. So as long as the prng state remains secret, it should
not be possible to guess. If the prng state can be accessed, we have
bigger problems.
Additionally, we are only consuming 6 bits to randomize the stack, so
there are only 64 possible random offsets. I assert that it would be
trivial for an attacker to brute force by repeating their attack and
waiting for the random stack offset to be the desired one. The prng
approach seems entirely proportional to this level of protection.
Performance data are provided below. The baseline is v6.18 with rndstack
on for each respective arch. (I)/(R) indicate statistically significant
improvement/regression. arm64 platform is AWS Graviton3 (m7g.metal).
x86_64 platform is AWS Sapphire Rapids (m7i.24xlarge):
+-----------------+--------------+---------------+---------------+
| Benchmark | Result Class | per-cpu-prng | per-cpu-prng |
| | | arm64 (metal) | x86_64 (VM) |
+=================+==============+===============+===============+
| syscall/getpid | mean (ns) | (I) -9.50% | (I) -17.65% |
| | p99 (ns) | (I) -59.24% | (I) -24.41% |
| | p99.9 (ns) | (I) -59.52% | (I) -28.52% |
+-----------------+--------------+---------------+---------------+
| syscall/getppid | mean (ns) | (I) -9.52% | (I) -19.24% |
| | p99 (ns) | (I) -59.25% | (I) -25.03% |
| | p99.9 (ns) | (I) -59.50% | (I) -28.17% |
+-----------------+--------------+---------------+---------------+
| syscall/invalid | mean (ns) | (I) -10.31% | (I) -18.56% |
| | p99 (ns) | (I) -60.79% | (I) -20.06% |
| | p99.9 (ns) | (I) -61.04% | (I) -25.04% |
+-----------------+--------------+---------------+---------------+
I tested an earlier version of this change on x86 bare metal and it
showed a smaller but still significant improvement. The bare metal
system wasn't available this time around so testing was done in a VM
instance. I'm guessing the cost of rdtsc is higher for VMs.
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Link: https://patch.msgid.link/20260303150840.3789438-3-ryan.roberts@arm.com
Signed-off-by: Kees Cook <kees@kernel.org>
|
|
kstack_offset was previously maintained per-cpu, but this caused a
couple of issues. So let's instead make it per-task.
Issue 1: add_random_kstack_offset() and choose_random_kstack_offset()
expected and required to be called with interrupts and preemption
disabled so that it could manipulate per-cpu state. But arm64, loongarch
and risc-v are calling them with interrupts and preemption enabled. I
don't _think_ this causes any functional issues, but it's certainly
unexpected and could lead to manipulating the wrong cpu's state, which
could cause a minor performance degradation due to bouncing the cache
lines. By maintaining the state per-task those functions can safely be
called in preemptible context.
Issue 2: add_random_kstack_offset() is called before executing the
syscall and expands the stack using a previously chosen random offset.
choose_random_kstack_offset() is called after executing the syscall and
chooses and stores a new random offset for the next syscall. With
per-cpu storage for this offset, an attacker could force cpu migration
during the execution of the syscall and prevent the offset from being
updated for the original cpu such that it is predictable for the next
syscall on that cpu. By maintaining the state per-task, this problem
goes away because the per-task random offset is updated after the
syscall regardless of which cpu it is executing on.
Fixes: 39218ff4c625 ("stack: Optionally randomize kernel stack offset each syscall")
Closes: https://lore.kernel.org/all/dd8c37bc-795f-4c7a-9086-69e584d8ab24@arm.com/
Cc: stable@vger.kernel.org
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
Link: https://patch.msgid.link/20260303150840.3789438-2-ryan.roberts@arm.com
Signed-off-by: Kees Cook <kees@kernel.org>
|
|
The codgen for adding architecture-specific stack alignment to the
effective alloca() usage is somewhat inefficient and allows a bit to get
carried beyond the desired entropy range. This isn't really a problem,
but it's unexpected and the codegen is kind of bad.
Quoting Mark[1], the disassembly for arm64's invoke_syscall() looks like:
// offset = raw_cpu_read(kstack_offset)
mov x4, sp
adrp x0, kstack_offset
mrs x5, tpidr_el1
add x0, x0, #:lo12:kstack_offset
ldr w0, [x0, x5]
// offset = KSTACK_OFFSET_MAX(offset)
and x0, x0, #0x3ff
// alloca(offset)
add x0, x0, #0xf
and x0, x0, #0x7f0
sub sp, x4, x0
... which in C would be:
offset = raw_cpu_read(kstack_offset)
offset &= 0x3ff; // [0x0, 0x3ff]
offset += 0xf; // [0xf, 0x40e]
offset &= 0x7f0; // [0x0,
... so when *all* bits [3:0] are 0, they'll have no impact, and when
*any* of bits [3:0] are 1 they'll trigger a carry into bit 4, which
could ripple all the way up and spill into bit 10.
Switch the masking in KSTACK_OFFSET_MAX() to explicitly clear the bottom
bits to avoid the rounding by using 0b1111110000 instead of 0b1111111111:
// offset = raw_cpu_read(kstack_offset)
mov x4, sp
adrp x0, 0 <kstack_offset>
mrs x5, tpidr_el1
add x0, x0, #:lo12:kstack_offset
ldr w0, [x0, x5]
// offset = KSTACK_OFFSET_MAX(offset)
and x0, x0, #0x3f0
// alloca(offset)
sub sp, x4, x0
Suggested-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lore.kernel.org/lkml/ZnVfOnIuFl2kNWkT@J2N7QTR9R3/ [1]
Link: https://lore.kernel.org/r/20240702211612.work.576-kees@kernel.org
Signed-off-by: Kees Cook <kees@kernel.org>
|
|
The kstack_offset variable was really only ever using the low bits for
kernel stack offset entropy. Add a ror32() to increase bit diffusion.
Suggested-by: Arnd Bergmann <arnd@arndb.de>
Fixes: 39218ff4c625 ("stack: Optionally randomize kernel stack offset each syscall")
Link: https://lore.kernel.org/r/20240309202445.work.165-kees@kernel.org
Signed-off-by: Kees Cook <keescook@chromium.org>
|
|
There were some recent questions about where and why to use the
random_kstack routines when applying them to new architectures[1].
Update the header comments to reflect the design choices for the
routines.
[1] https://lore.kernel.org/lkml/1652173338.7bltwybi0c.astroid@bobo.none
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Xiu Jianfeng <xiujianfeng@huawei.com>
Signed-off-by: Kees Cook <keescook@chromium.org>
|
|
All supported versions of Clang perform auto-init of __builtin_alloca()
when stack auto-init is on (CONFIG_INIT_STACK_ALL_{ZERO,PATTERN}).
add_random_kstack_offset() uses __builtin_alloca() to add a stack
offset. This means, when CONFIG_INIT_STACK_ALL_{ZERO,PATTERN} is
enabled, add_random_kstack_offset() will auto-init that unused portion
of the stack used to add an offset.
There are several problems with this:
1. These offsets can be as large as 1023 bytes. Performing
memset() on them isn't exactly cheap, and this is done on
every syscall entry.
2. Architectures adding add_random_kstack_offset() to syscall
entry implemented in C require them to be 'noinstr' (e.g. see
x86 and s390). The potential problem here is that a call to
memset may occur, which is not noinstr.
A x86_64 defconfig kernel with Clang 11 and CONFIG_VMLINUX_VALIDATION shows:
| vmlinux.o: warning: objtool: do_syscall_64()+0x9d: call to memset() leaves .noinstr.text section
| vmlinux.o: warning: objtool: do_int80_syscall_32()+0xab: call to memset() leaves .noinstr.text section
| vmlinux.o: warning: objtool: __do_fast_syscall_32()+0xe2: call to memset() leaves .noinstr.text section
| vmlinux.o: warning: objtool: fixup_bad_iret()+0x2f: call to memset() leaves .noinstr.text section
Clang 14 (unreleased) will introduce a way to skip alloca initialization
via __builtin_alloca_uninitialized() (https://reviews.llvm.org/D115440).
Constrain RANDOMIZE_KSTACK_OFFSET to only be enabled if no stack
auto-init is enabled, the compiler is GCC, or Clang is version 14+. Use
__builtin_alloca_uninitialized() if the compiler provides it, as is done
by Clang 14.
Link: https://lkml.kernel.org/r/YbHTKUjEejZCLyhX@elver.google.com
Fixes: 39218ff4c625 ("stack: Optionally randomize kernel stack offset each syscall")
Signed-off-by: Marco Elver <elver@google.com>
Reviewed-by: Nathan Chancellor <nathan@kernel.org>
Acked-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: https://lore.kernel.org/r/20220131090521.1947110-2-elver@google.com
|
|
The randomize_kstack_offset feature is unconditionally compiled in when
the architecture supports it.
To add constraints on compiler versions, we require a dedicated Kconfig
variable. Therefore, introduce RANDOMIZE_KSTACK_OFFSET.
Furthermore, this option is now also configurable by EXPERT kernels:
while the feature is supposed to have zero performance overhead when
disabled, due to its use of static branches, there are few cases where
giving a distribution the option to disable the feature entirely makes
sense. For example, in very resource constrained environments, which
would never enable the feature to begin with, in which case the
additional kernel code size increase would be redundant.
Signed-off-by: Marco Elver <elver@google.com>
Reviewed-by: Nathan Chancellor <nathan@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: https://lore.kernel.org/r/20220131090521.1947110-1-elver@google.com
|
|
"o" isn't a common asm() constraint to use; it triggers an assertion in
assert-enabled builds of LLVM that it's not recognized when targeting
aarch64 (though it appears to fall back to "m"). It's fixed in LLVM 13 now,
but there isn't really a good reason to use "o" in particular here. To
avoid causing build issues for those using assert-enabled builds of earlier
LLVM versions, the constraint needs changing.
Instead, if the point is to retain the __builtin_alloca(), make ptr appear
to "escape" via being an input to an empty inline asm block. This is
preferable anyways, since otherwise this looks like a dead store.
While the use of "r" was considered in
https://lore.kernel.org/lkml/202104011447.2E7F543@keescook/
it was only tested as an output (which looks like a dead store, and wasn't
sufficient).
Use "r" as an input constraint instead, which behaves correctly across
compilers and architectures.
Fixes: 39218ff4c625 ("stack: Optionally randomize kernel stack offset each syscall")
Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>
Signed-off-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Kees Cook <keescook@chromium.org>
Tested-by: Nathan Chancellor <nathan@kernel.org>
Reviewed-by: Nathan Chancellor <nathan@kernel.org>
Link: https://reviews.llvm.org/D100412
Link: https://bugs.llvm.org/show_bug.cgi?id=49956
Link: https://lore.kernel.org/r/20210419231741.4084415-1-keescook@chromium.org
|
|
This provides the ability for architectures to enable kernel stack base
address offset randomization. This feature is controlled by the boot
param "randomize_kstack_offset=on/off", with its default value set by
CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT.
This feature is based on the original idea from the last public release
of PaX's RANDKSTACK feature: https://pax.grsecurity.net/docs/randkstack.txt
All the credit for the original idea goes to the PaX team. Note that
the design and implementation of this upstream randomize_kstack_offset
feature differs greatly from the RANDKSTACK feature (see below).
Reasoning for the feature:
This feature aims to make harder the various stack-based attacks that
rely on deterministic stack structure. We have had many such attacks in
past (just to name few):
https://jon.oberheide.org/files/infiltrate12-thestackisback.pdf
https://jon.oberheide.org/files/stackjacking-infiltrate11.pdf
https://googleprojectzero.blogspot.com/2016/06/exploiting-recursion-in-linux-kernel_20.html
As Linux kernel stack protections have been constantly improving
(vmap-based stack allocation with guard pages, removal of thread_info,
STACKLEAK), attackers have had to find new ways for their exploits
to work. They have done so, continuing to rely on the kernel's stack
determinism, in situations where VMAP_STACK and THREAD_INFO_IN_TASK_STRUCT
were not relevant. For example, the following recent attacks would have
been hampered if the stack offset was non-deterministic between syscalls:
https://repositorio-aberto.up.pt/bitstream/10216/125357/2/374717.pdf
(page 70: targeting the pt_regs copy with linear stack overflow)
https://a13xp0p0v.github.io/2020/02/15/CVE-2019-18683.html
(leaked stack address from one syscall as a target during next syscall)
The main idea is that since the stack offset is randomized on each system
call, it is harder for an attack to reliably land in any particular place
on the thread stack, even with address exposures, as the stack base will
change on the next syscall. Also, since randomization is performed after
placing pt_regs, the ptrace-based approach[1] to discover the randomized
offset during a long-running syscall should not be possible.
Design description:
During most of the kernel's execution, it runs on the "thread stack",
which is pretty deterministic in its structure: it is fixed in size,
and on every entry from userspace to kernel on a syscall the thread
stack starts construction from an address fetched from the per-cpu
cpu_current_top_of_stack variable. The first element to be pushed to the
thread stack is the pt_regs struct that stores all required CPU registers
and syscall parameters. Finally the specific syscall function is called,
with the stack being used as the kernel executes the resulting request.
The goal of randomize_kstack_offset feature is to add a random offset
after the pt_regs has been pushed to the stack and before the rest of the
thread stack is used during the syscall processing, and to change it every
time a process issues a syscall. The source of randomness is currently
architecture-defined (but x86 is using the low byte of rdtsc()). Future
improvements for different entropy sources is possible, but out of scope
for this patch. Further more, to add more unpredictability, new offsets
are chosen at the end of syscalls (the timing of which should be less
easy to measure from userspace than at syscall entry time), and stored
in a per-CPU variable, so that the life of the value does not stay
explicitly tied to a single task.
As suggested by Andy Lutomirski, the offset is added using alloca()
and an empty asm() statement with an output constraint, since it avoids
changes to assembly syscall entry code, to the unwinder, and provides
correct stack alignment as defined by the compiler.
In order to make this available by default with zero performance impact
for those that don't want it, it is boot-time selectable with static
branches. This way, if the overhead is not wanted, it can just be
left turned off with no performance impact.
The generated assembly for x86_64 with GCC looks like this:
...
ffffffff81003977: 65 8b 05 02 ea 00 7f mov %gs:0x7f00ea02(%rip),%eax
# 12380 <kstack_offset>
ffffffff8100397e: 25 ff 03 00 00 and $0x3ff,%eax
ffffffff81003983: 48 83 c0 0f add $0xf,%rax
ffffffff81003987: 25 f8 07 00 00 and $0x7f8,%eax
ffffffff8100398c: 48 29 c4 sub %rax,%rsp
ffffffff8100398f: 48 8d 44 24 0f lea 0xf(%rsp),%rax
ffffffff81003994: 48 83 e0 f0 and $0xfffffffffffffff0,%rax
...
As a result of the above stack alignment, this patch introduces about
5 bits of randomness after pt_regs is spilled to the thread stack on
x86_64, and 6 bits on x86_32 (since its has 1 fewer bit required for
stack alignment). The amount of entropy could be adjusted based on how
much of the stack space we wish to trade for security.
My measure of syscall performance overhead (on x86_64):
lmbench: /usr/lib/lmbench/bin/x86_64-linux-gnu/lat_syscall -N 10000 null
randomize_kstack_offset=y Simple syscall: 0.7082 microseconds
randomize_kstack_offset=n Simple syscall: 0.7016 microseconds
So, roughly 0.9% overhead growth for a no-op syscall, which is very
manageable. And for people that don't want this, it's off by default.
There are two gotchas with using the alloca() trick. First,
compilers that have Stack Clash protection (-fstack-clash-protection)
enabled by default (e.g. Ubuntu[3]) add pagesize stack probes to
any dynamic stack allocations. While the randomization offset is
always less than a page, the resulting assembly would still contain
(unreachable!) probing routines, bloating the resulting assembly. To
avoid this, -fno-stack-clash-protection is unconditionally added to
the kernel Makefile since this is the only dynamic stack allocation in
the kernel (now that VLAs have been removed) and it is provably safe
from Stack Clash style attacks.
The second gotcha with alloca() is a negative interaction with
-fstack-protector*, in that it sees the alloca() as an array allocation,
which triggers the unconditional addition of the stack canary function
pre/post-amble which slows down syscalls regardless of the static
branch. In order to avoid adding this unneeded check and its associated
performance impact, architectures need to carefully remove uses of
-fstack-protector-strong (or -fstack-protector) in the compilation units
that use the add_random_kstack() macro and to audit the resulting stack
mitigation coverage (to make sure no desired coverage disappears). No
change is visible for this on x86 because the stack protector is already
unconditionally disabled for the compilation unit, but the change is
required on arm64. There is, unfortunately, no attribute that can be
used to disable stack protector for specific functions.
Comparison to PaX RANDKSTACK feature:
The RANDKSTACK feature randomizes the location of the stack start
(cpu_current_top_of_stack), i.e. including the location of pt_regs
structure itself on the stack. Initially this patch followed the same
approach, but during the recent discussions[2], it has been determined
to be of a little value since, if ptrace functionality is available for
an attacker, they can use PTRACE_PEEKUSR/PTRACE_POKEUSR to read/write
different offsets in the pt_regs struct, observe the cache behavior of
the pt_regs accesses, and figure out the random stack offset. Another
difference is that the random offset is stored in a per-cpu variable,
rather than having it be per-thread. As a result, these implementations
differ a fair bit in their implementation details and results, though
obviously the intent is similar.
[1] https://lore.kernel.org/kernel-hardening/2236FBA76BA1254E88B949DDB74E612BA4BC57C1@IRSMSX102.ger.corp.intel.com/
[2] https://lore.kernel.org/kernel-hardening/20190329081358.30497-1-elena.reshetova@intel.com/
[3] https://lists.ubuntu.com/archives/ubuntu-devel/2019-June/040741.html
Co-developed-by: Elena Reshetova <elena.reshetova@intel.com>
Signed-off-by: Elena Reshetova <elena.reshetova@intel.com>
Signed-off-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/r/20210401232347.2791257-4-keescook@chromium.org
|
