KVM: arm64: Add missing BTI instructions
PATCH v2 1/1 KVM: arm64: Add missing BTI instructions
Some bti instructions were missing from commit b53d4a272349 ("KVM: arm64: Use BTI for nvhe")
1) kvm_host_psci_cpu_entrykvm_host_psci_cpu_entry is called from __kvm_hyp_init_cpu through "br" instruction as __kvm_hyp_init_cpu resides in idmap section while kvm_host_psci_cpu_entry is in hyp .text so the offset is larger than 128MB range covered by "b".
Which means that this function should start with "bti j" instruction.
LLVM which is the only compiler supporting BTI for Linux, adds "bti j" for jump tables or by when taking the address of the block [1]. Same behaviour is observed with GCC.
As kvm_host_psci_cpu_entry is a C function, this must be done in assembly.
Another solution is to use X16/X17 with "br", as according to ARM ARM DDI0487I.a RLJHCL/IGMGRS, PACIASP has an implicit branch target identification instruction that is compatible with PSTATE.BTYPE 0b01 which includes "br X16/X17"
And the kvm_host_psci_cpu_entry has PACIASP as it is an external function.
Although, using explicit "bti" makes it more clear than relying on which register is used.
A third solution is to clear SCTLR_EL2.BT, which would make PACIASP compatible PSTATE.BTYPE 0b11 ("br" to other registers). However this deviates from the kernel behaviour (in bti_enable()).
2) Spectre vector table
"br" instructions are generated at runtime for the vector table (__bp_harden_hyp_vecs).
These branches would land on vectors in __kvm_hyp_vector at offset 8. As all the macros are defined with valid_vect/invalid_vect, it is sufficient to add "bti j" at the correct offset.
[1] https://reviews.llvm.org/D52867
---
v1 -> v2:
- Add another missing bti in spectre vector table.
v1: https://lore.kernel.org/all/[email protected]/
--- arch/arm64/kvm/hyp/hyp-entry.S | 8 ++++++++ arch/arm64/kvm/hyp/nvhe/host.S | 10 ++++++++++ arch/arm64/kvm/hyp/nvhe/psci-relay.c | 2 +- 3 files changed, 19 insertions(+), 1 deletion(-)
arch/arm64/kvm/hyp/hyp-entry.S +8 -0
@@ -154,6 +154,12 @@ SYM_CODE_END(\label) esb stp x0, x1, [sp, #-16]! 662: + /* + * spectre vectors __bp_harden_hyp_vecs generate br instructions at runtime + * that jump at offset 8 at __kvm_hyp_vector. + * As hyp .text is guarded section, it needs bti j. + */ + bti j b \target check_preamble_length 661b, 662b
@@ -165,6 +171,8 @@ check_preamble_length 661b, 662b nop stp x0, x1, [sp, #-16]! 662: + /* Check valid_vect */ + bti j b \target check_preamble_length 661b, 662b
arch/arm64/kvm/hyp/nvhe/host.S +10 -0
@@ -297,3 +297,13 @@ SYM_CODE_START(__kvm_hyp_host_forward_smc) ret SYM_CODE_END(__kvm_hyp_host_forward_smc) + +/* + * kvm_host_psci_cpu_entry is called through br instruction, which requires + * bti j instruction as compilers (gcc and llvm) doesn't insert bti j for external + * functions, but bti c instead. + */ +SYM_CODE_START(kvm_host_psci_cpu_entry) + bti j + b __kvm_host_psci_cpu_entry +SYM_CODE_END(kvm_host_psci_cpu_entry)
arch/arm64/kvm/hyp/nvhe/psci-relay.c +1 -1
@@ -200,7 +200,7 @@ static int psci_system_suspend(u64 func_id, struct kvm_cpu_context *host_ctxt) __hyp_pa(init_params), 0); } -asmlinkage void __noreturn kvm_host_psci_cpu_entry(bool is_cpu_on) +asmlinkage void __noreturn __kvm_host_psci_cpu_entry(bool is_cpu_on) {struct psci_boot_args *boot_args;struct kvm_cpu_context *host_ctxt;
Signature · 1 line
2.41.0.255.g8b1d071c50-goog
Quoted 37 lines · Mostafa Saleh
Some bti instructions were missing from commit
b53d4a272349("KVM: arm64: Use BTI for nvhe")1)
kvm_host_psci_cpu_entrykvm_host_psci_cpu_entryis called from__kvm_hyp_init_cputhrough "br" instruction as__kvm_hyp_init_cpuresides in idmap section whilekvm_host_psci_cpu_entryis in hyp .text so the offset is larger than 128MB range covered by "b".
Which means that this function should start with "bti j" instruction.LLVM which is the only compiler supporting BTI for Linux, adds "bti j" for jump tables or by when taking the address of the block [1]. Same behaviour is observed with GCC.
As
kvm_host_psci_cpu_entryis a C function, this must be done in assembly.Another solution is to use X16/X17 with "br", as according to ARM ARM DDI0487I.a RLJHCL/IGMGRS, PACIASP has an implicit branch target identification instruction that is compatible with PSTATE.BTYPE 0b01 which includes "br X16/X17"
And thekvm_host_psci_cpu_entryhas PACIASP as it is an external function.
Although, using explicit "bti" makes it more clear than relying on which register is used.A third solution is to clear SCTLR_EL2.BT, which would make PACIASP compatible PSTATE.BTYPE 0b11 ("br" to other registers). However this deviates from the kernel behaviour (in
bti_enable()).2) Spectre vector table
"br" instructions are generated at runtime for the vector table (__bp_harden_hyp_vecs).
These branches would land on vectors in__kvm_hyp_vectorat offset 8. As all the macros are defined with valid_vect/invalid_vect, it is sufficient to add "bti j" at the correct offset.
Quoted 5 lines
[1] https://reviews.llvm.org/D52867
Fixes b53d4a272349 ("KVM: arm64: Use BTI for nvhe")Signed-off-by Mostafa Saleh <[email protected]>Reported-by Sudeep Holla <[email protected]>
Nothing change w.r.t cpu suspend-resume path in v2 anyways, but I assure I tested this again just be absolutely sure and it still fixes the issue I reported 😄, so
Signature · 2 lines
Regards, Sudeep
Quoted 44 lines · Sudeep Holla
Quoted 37 lines · Mostafa Saleh
Some bti instructions were missing from commit
b53d4a272349("KVM: arm64: Use BTI for nvhe")1)
kvm_host_psci_cpu_entrykvm_host_psci_cpu_entryis called from__kvm_hyp_init_cputhrough "br" instruction as__kvm_hyp_init_cpuresides in idmap section whilekvm_host_psci_cpu_entryis in hyp .text so the offset is larger than 128MB range covered by "b".
Which means that this function should start with "bti j" instruction.LLVM which is the only compiler supporting BTI for Linux, adds "bti j" for jump tables or by when taking the address of the block [1]. Same behaviour is observed with GCC.
As
kvm_host_psci_cpu_entryis a C function, this must be done in assembly.Another solution is to use X16/X17 with "br", as according to ARM ARM DDI0487I.a RLJHCL/IGMGRS, PACIASP has an implicit branch target identification instruction that is compatible with PSTATE.BTYPE 0b01 which includes "br X16/X17"
And thekvm_host_psci_cpu_entryhas PACIASP as it is an external function.
Although, using explicit "bti" makes it more clear than relying on which register is used.A third solution is to clear SCTLR_EL2.BT, which would make PACIASP compatible PSTATE.BTYPE 0b11 ("br" to other registers). However this deviates from the kernel behaviour (in
bti_enable()).2) Spectre vector table
"br" instructions are generated at runtime for the vector table (__bp_harden_hyp_vecs).
These branches would land on vectors in__kvm_hyp_vectorat offset 8. As all the macros are defined with valid_vect/invalid_vect, it is sufficient to add "bti j" at the correct offset.Quoted 5 lines
[1] https://reviews.llvm.org/D52867
Fixes b53d4a272349 ("KVM: arm64: Use BTI for nvhe")Signed-off-by Mostafa Saleh <[email protected]>Reported-by Sudeep Holla <[email protected]>
Quoted 5 lines
Nothing change w.r.t cpu suspend-resume path in v2 anyways, but I assure I tested this again just be absolutely sure and it still fixes the issue I reported 😄, so
Tested-by Sudeep Holla <[email protected]>
Thanks for testing the patch again!
Quoted 40 lines · Mostafa Saleh
Some bti instructions were missing from commit
b53d4a272349("KVM: arm64: Use BTI for nvhe")1)
kvm_host_psci_cpu_entrykvm_host_psci_cpu_entryis called from__kvm_hyp_init_cputhrough "br" instruction as__kvm_hyp_init_cpuresides in idmap section whilekvm_host_psci_cpu_entryis in hyp .text so the offset is larger than 128MB range covered by "b".
Which means that this function should start with "bti j" instruction.LLVM which is the only compiler supporting BTI for Linux, adds "bti j" for jump tables or by when taking the address of the block [1]. Same behaviour is observed with GCC.
As
kvm_host_psci_cpu_entryis a C function, this must be done in assembly.Another solution is to use X16/X17 with "br", as according to ARM ARM DDI0487I.a RLJHCL/IGMGRS, PACIASP has an implicit branch target identification instruction that is compatible with PSTATE.BTYPE 0b01 which includes "br X16/X17"
And thekvm_host_psci_cpu_entryhas PACIASP as it is an external function.
Although, using explicit "bti" makes it more clear than relying on which register is used.A third solution is to clear SCTLR_EL2.BT, which would make PACIASP compatible PSTATE.BTYPE 0b11 ("br" to other registers). However this deviates from the kernel behaviour (in
bti_enable()).2) Spectre vector table
"br" instructions are generated at runtime for the vector table (__bp_harden_hyp_vecs).
These branches would land on vectors in__kvm_hyp_vectorat offset 8. As all the macros are defined with valid_vect/invalid_vect, it is sufficient to add "bti j" at the correct offset.
Quoted 7 lines
Fixes b53d4a272349 ("KVM: arm64: Use BTI for nvhe")Signed-off-by Mostafa Saleh <[email protected]>Reported-by Sudeep Holla <[email protected]>---
v1 -> v2:
- Add another missing bti in spectre vector table.
v1: https://lore.kernel.org/all/[email protected]/
M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
Oliver Upton
AppliedQuoted 11 lines · Mostafa Saleh
Some bti instructions were missing from commit
b53d4a272349("KVM: arm64: Use BTI for nvhe")1)
kvm_host_psci_cpu_entrykvm_host_psci_cpu_entryis called from__kvm_hyp_init_cputhrough "br" instruction as__kvm_hyp_init_cpuresides in idmap section whilekvm_host_psci_cpu_entryis in hyp .text so the offset is larger than 128MB range covered by "b".
Which means that this function should start with "bti j" instruction.[...]
Applied to kvmarm/fixes. Sorry this took longer than expected, Mostafa. I was hitting an unrelated issue with pKVM initialization on the only system I have that supports BTI.
[1/1] KVM: arm64: Add missing BTI instructions
https://git.kernel.org/kvmarm/kvmarm/c/dcf89d111199
--
Best,
Oliver
Quoted 11 lines · Oliver Upton
Quoted 10 lines · Mostafa Saleh
Some bti instructions were missing from commit
b53d4a272349("KVM: arm64: Use BTI for nvhe")1)
kvm_host_psci_cpu_entrykvm_host_psci_cpu_entryis called from__kvm_hyp_init_cputhrough "br" instruction as__kvm_hyp_init_cpuresides in idmap section whilekvm_host_psci_cpu_entryis in hyp .text so the offset is larger than 128MB range covered by "b".
Which means that this function should start with "bti j" instruction.
Quoted 9 lines
Quoted 1 line
[...]
Applied to kvmarm/fixes. Sorry this took longer than expected, Mostafa. I was hitting an unrelated issue with pKVM initialization on the only system I have that supports BTI.
[1/1] KVM: arm64: Add missing BTI instructions
https://git.kernel.org/kvmarm/kvmarm/c/dcf89d111199
Thanks Oliver, hopefully everything is sorted out now.