arm64: reject prctl(PR_PAC_RESET_KEYS) on compat tasks
PATCH 1/1 arm64: reject prctl(PR_PAC_RESET_KEYS) on compat tasks
It doesn't make sense to issue prctl(PR_PAC_RESET_KEYS) on a compat task because the 32-bit instruction set does not offer PAuth instructions. For consistency with other 64-bit only prctls such as {SET,GET}_TAGGED_ADDR_CTRL, reject the prctl on compat tasks.
Although this is a userspace-visible change, maybe it isn't too late to make this change given that the hardware isn't available yet and it's very unlikely that anyone has 32-bit software that actually depends on this succeeding.
--- arch/arm64/kernel/pointer_auth.c | 4 ++++ 1 file changed, 4 insertions(+)
arch/arm64/kernel/pointer_auth.c +4 -0
@@ -1,5 +1,6 @@// SPDX-License-Identifier: GPL-2.0 +#include <linux/compat.h> #include <linux/errno.h> #include <linux/prctl.h> #include <linux/random.h>
@@ -17,6 +18,9 @@ int ptrauth_prctl_reset_keys(struct task_struct *tsk, unsigned long arg)if (!system_supports_address_auth () && !system_supports_generic_auth ())return -EINVAL ; +if (is_compat_task ()) +return -EINVAL ; +if (!arg) {ptrauth_keys_init_user (keys);return 0 ;
Signature · 7 lines
2.28.0.1011.ga647a8990f-goog _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 5 lines · Peter Collingbourne
It doesn't make sense to issue
prctl(PR_PAC_RESET_KEYS)on a compat task because the 32-bit instruction set does not offer PAuth instructions. For consistency with other 64-bit only prctls such as {SET,GET}_TAGGED_ADDR_CTRL, reject the prctl on compat tasks.
Quoted 7 lines
Although this is a userspace-visible change, maybe it isn't too late to make this change given that the hardware isn't available yet and it's very unlikely that anyone has 32-bit software that actually depends on this succeeding.
This does seem an anomaly, but it's not an isolated case. I suspect that some other prctls are also missing a compat check -- PR_SVE_SET_VL doesn't have it, for example.
So, I'm not sure it's worth fixing this one case in isolation. Fixing all affected cases may have greater risk, and it won't stay fixed, since the compat check will likely often get forgotten when a new prctl is added.
So, is this anomaly in any way harmful?
Can the code be refactored in such a way as to make it hard to forget the check in future?
Cheers
---Dave
Quoted 28 lines
--- arch/arm64/kernel/pointer_auth.c | 4 ++++ 1 file changed, 4 insertions(+)arch/arm64/kernel/pointer_auth.c +4
-0@@ -1,5 +1,6 @@// SPDX-License-Identifier: GPL-2.0 +#include <linux/compat.h> #include <linux/errno.h> #include <linux/prctl.h> #include <linux/random.h> @@ -17,6 +18,9 @@ int ptrauth_prctl_reset_keys(struct task_struct *tsk, unsigned long arg)if (!system_supports_address_auth () && !system_supports_generic_auth ())return -EINVAL ; +if (is_compat_task ()) +return -EINVAL ; +if (!arg) {ptrauth_keys_init_user (keys);return 0 ;Signature · 1 line
2.28.0.1011.ga647a8990f-goog
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 15 lines · Dave Martin
Quoted 5 lines · Peter Collingbourne
It doesn't make sense to issue
prctl(PR_PAC_RESET_KEYS)on a compat task because the 32-bit instruction set does not offer PAuth instructions. For consistency with other 64-bit only prctls such as {SET,GET}_TAGGED_ADDR_CTRL, reject the prctl on compat tasks.Quoted 7 lines
Although this is a userspace-visible change, maybe it isn't too late to make this change given that the hardware isn't available yet and it's very unlikely that anyone has 32-bit software that actually depends on this succeeding.
Quoted 8 lines
This does seem an anomaly, but it's not an isolated case. I suspect that some other prctls are also missing a compat check --
PR_SVE_SET_VLdoesn't have it, for example.So, I'm not sure it's worth fixing this one case in isolation. Fixing all affected cases may have greater risk, and it won't stay fixed, since the compat check will likely often get forgotten when a new prctl is added.
The only other affected cases involve SVE and that doesn't have hardware available yet either, right? I'm going by the binutils CPU list, which is the closest thing that I'm aware of to an official list of all microarchitectures and their supported ISA features:
(and I know that Neoverse V1/N2 isn't available yet)
Quoted 1 line
So, is this anomaly in any way harmful?
Not as far as I can tell, at least for this specific prctl.
Quoted 2 lines
Can the code be refactored in such a way as to make it hard to forget the check in future?
I've never been a fan of the arch-specific prctls being listed in kernel/sys.c. It seems better to me to have that handling be moved into a new arch hook and that should let us remove some boilerplate as well. We can make the default case in the prctl syscall handler look like this:
default:
return arch_handle_prctrl(option, arg2, arg3, arg4, arg5);
And move the arch-specific prctls into a switch in arch_handle_prctl(). Now, since (as far as I can tell) all of the arm64-specific prctls do not make sense on compat tasks, we can add an:
if (is_compat_task())
return -EINVAL;
to the top of arch_handle_prctl() and any new arm64-specific prctls will get the compat check by default. Of course, if we add a new compat-compatible prctl in the future, we may add it to a new switch before the if statement.
Peter
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 25 lines · Peter Collingbourne
Quoted 15 lines · Dave Martin
Quoted 5 lines · Peter Collingbourne
It doesn't make sense to issue
prctl(PR_PAC_RESET_KEYS)on a compat task because the 32-bit instruction set does not offer PAuth instructions. For consistency with other 64-bit only prctls such as {SET,GET}_TAGGED_ADDR_CTRL, reject the prctl on compat tasks.Quoted 7 lines
Although this is a userspace-visible change, maybe it isn't too late to make this change given that the hardware isn't available yet and it's very unlikely that anyone has 32-bit software that actually depends on this succeeding.
Quoted 8 lines
This does seem an anomaly, but it's not an isolated case. I suspect that some other prctls are also missing a compat check --
PR_SVE_SET_VLdoesn't have it, for example.So, I'm not sure it's worth fixing this one case in isolation. Fixing all affected cases may have greater risk, and it won't stay fixed, since the compat check will likely often get forgotten when a new prctl is added.
Quoted 8 lines
The only other affected cases involve SVE and that doesn't have hardware available yet either, right? I'm going by the binutils CPU list, which is the closest thing that I'm aware of to an official list of all microarchitectures and their supported ISA features:
(and I know that Neoverse V1/N2 isn't available yet)
Yes, that's probably a reasonable list.
SVE is supported by Fujitsu's A64FX part, but that's in few pockets. In any case, it doesn't feel legitimate for a compat process to be using this. The chance of extra compat checks there breaking something seems pretty low.
It's not like we ever promised it would work on compat.
Quoted 3 lines
Quoted 1 line
So, is this anomaly in any way harmful?
Not as far as I can tell, at least for this specific prctl.
Ack
Quoted 20 lines
Quoted 2 lines
Can the code be refactored in such a way as to make it hard to forget the check in future?
I've never been a fan of the arch-specific prctls being listed in
kernel/sys.c. It seems better to me to have that handling be moved into a new arch hook and that should let us remove some boilerplate as well. We can make the default case in the prctl syscall handler look like this:default:
returnarch_handle_prctrl(option, arg2, arg3, arg4, arg5);And move the arch-specific prctls into a switch in
arch_handle_prctl(). Now, since (as far as I can tell) all of the arm64-specific prctls do not make sense on compat tasks, we can add an:if (
is_compat_task())
return-EINVAL;
Quoted 6 lines
to the top of
arch_handle_prctl()and any new arm64-specific prctls will get the compat check by default. Of course, if we add a new compat-compatible prctl in the future, we may add it to a new switch before the if statement.Peter
Something like that could work.
I did have a try at cleaning some of this up some time ago [1], but other stuff happened and I never finished it -- not sure how relevant it still is, but might be worth a look.
Cheers
---Dave
[1] [RFC PATCH 00/11] prctl: Modernise wiring for optional prctl() calls
https://lore.kernel.org/lkml/[email protected]/
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Will Deacon
maintainerQuoted 10 lines · Peter Collingbourne
It doesn't make sense to issue
prctl(PR_PAC_RESET_KEYS)on a compat task because the 32-bit instruction set does not offer PAuth instructions. For consistency with other 64-bit only prctls such as {SET,GET}_TAGGED_ADDR_CTRL, reject the prctl on compat tasks.Although this is a userspace-visible change, maybe it isn't too late to make this change given that the hardware isn't available yet and it's very unlikely that anyone has 32-bit software that actually depends on this succeeding.
Quoted 5 lines
Signed-off-by Peter Collingbourne <[email protected]>--- arch/arm64/kernel/pointer_auth.c | 4 ++++ 1 file changed, 4 insertions(+)
Cheers. Although I agree with the discussion here that it would be better to catch this in one place, for now I'll merge this and add similar checks to the SVE prctl()s too. Patch below (I also tweaked it to use the 'tsk' parameter for ptrauth, for consistency).
Will
--->8
arch/arm64/kernel/fpsimd.c +2 -2
@@ -678,7 +678,7 @@ int sve_set_current_vl(unsigned long arg) vl = arg &PR_SVE_VL_LEN_MASK ; flags = arg & ~vl; -if (!system_supports_sve ()) +if (!system_supports_sve () ||is_compat_task ())return -EINVAL ; ret =sve_set_vector_length (current, vl, flags);
@@ -691,7 +691,7 @@ int sve_set_current_vl(unsigned long arg)/* PR_SVE_GET_VL */ int sve_get_current_vl (void ) { -if (!system_supports_sve ()) +if (!system_supports_sve () ||is_compat_task ())return -EINVAL ;return sve_prctl_status (0 );
arch/arm64/kernel/pointer_auth.c +4 -0
@@ -1,5 +1,6 @@// SPDX-License-Identifier: GPL-2.0 +#include <linux/compat.h> #include <linux/errno.h> #include <linux/prctl.h> #include <linux/random.h>
@@ -17,6 +18,9 @@ int ptrauth_prctl_reset_keys(struct task_struct *tsk, unsigned long arg)if (!system_supports_address_auth () && !system_supports_generic_auth ())return -EINVAL ; +if (is_compat_thread (task_thread_info (tsk))) +return -EINVAL ; +if (!arg) {ptrauth_keys_init_user (keys);return 0 ;
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 9 lines · Peter Collingbourne
It doesn't make sense to issue
prctl(PR_PAC_RESET_KEYS)on a compat task because the 32-bit instruction set does not offer PAuth instructions. For consistency with other 64-bit only prctls such as {SET,GET}_TAGGED_ADDR_CTRL, reject the prctl on compat tasks.Although this is a userspace-visible change, maybe it isn't too late to make this change given that the hardware isn't available yet and it's very unlikely that anyone has 32-bit software that actually depends on this succeeding.
Applied to arm64 (for-next/core), thanks!
[1/1] arm64: reject prctl(PR_PAC_RESET_KEYS) on compat tasks
https://git.kernel.org/arm64/c/de7ae1c57273
Please check my changes if you get a chance.
Cheers,
Signature · 10 lines
Will https://fixes.arm64.dev https://next.arm64.dev https://will.arm64.dev _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel