Documentation/arch/arm64/booting.rst | 1 + arch/arm64/include/asm/el2_setup.h | 9 ++++++++- 2 files changed, 9 insertions(+), 1 deletion(-)
__init_el2_fgt2() writes one mask to both HDFGRTR2_EL2 and HDFGWTR2_EL2.
PMZR_EL0 is write-only, so its trap bit, nPMZR_EL0, exists only in
HDFGWTR2_EL2 and is therefore never set: a PMZR_EL0 write from the host
traps to EL2, where the nVHE hypervisor has no handler and BUG()s. The
kernel never writes PMZR_EL0, but kernel.perf_user_access=1 has the PMU
driver set PMUSERENR_EL0.UEN for a task with a user-read event, so a
write from EL0 reaches the trap and takes the host down without a panic
message.
Accumulate the HDFGWTR2_EL2 bits separately, as __init_el2_fgt() already
does for HDFGWTR_EL2, and set nPMZR_EL0 with the other FEAT_PMUv3p9
bits.
Fixes: 858c7bfcb35e1 ("arm64/boot: Enable EL2 requirements for FEAT_PMUv3p9")
Cc: stable@vger.kernel.org
Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
---
Based on Linux 7.3-rc4 (93f51579e7df2).
Documentation/arch/arm64/booting.rst | 1 +
arch/arm64/include/asm/el2_setup.h | 9 ++++++++-
2 files changed, 9 insertions(+), 1 deletion(-)
diff --git a/Documentation/arch/arm64/booting.rst b/Documentation/arch/arm64/booting.rst
index 13ef311dace83..3fea4b14ef7c2 100644
--- a/Documentation/arch/arm64/booting.rst
+++ b/Documentation/arch/arm64/booting.rst
@@ -465,6 +465,7 @@ Before jumping into the kernel, the following conditions must be met:
- HDFGWTR2_EL2.nPMICNTR_EL0 (bit 2) must be initialised to 0b1.
- HDFGWTR2_EL2.nPMICFILTR_EL0 (bit 3) must be initialised to 0b1.
- HDFGWTR2_EL2.nPMUACR_EL1 (bit 4) must be initialised to 0b1.
+ - HDFGWTR2_EL2.nPMZR_EL0 (bit 21) must be initialised to 0b1.
For CPUs with SPE data source filtering (FEAT_SPE_FDS):
diff --git a/arch/arm64/include/asm/el2_setup.h b/arch/arm64/include/asm/el2_setup.h
index aa8ec9df80243..87560d8b254e6 100644
--- a/arch/arm64/include/asm/el2_setup.h
+++ b/arch/arm64/include/asm/el2_setup.h
@@ -418,6 +418,7 @@
b.lt .Lskip_fgt2_\@
mov x0, xzr
+ mov x2, xzr
mrs x1, id_aa64dfr0_el1
ubfx x1, x1, #ID_AA64DFR0_EL1_PMUVer_SHIFT, #4
cmp x1, #ID_AA64DFR0_EL1_PMUVer_V3P9
@@ -426,6 +427,11 @@
orr x0, x0, #HDFGRTR2_EL2_nPMICNTR_EL0
orr x0, x0, #HDFGRTR2_EL2_nPMICFILTR_EL0
orr x0, x0, #HDFGRTR2_EL2_nPMUACR_EL1
+ orr x2, x2, #HDFGWTR2_EL2_nPMICNTR_EL0
+ orr x2, x2, #HDFGWTR2_EL2_nPMICFILTR_EL0
+ orr x2, x2, #HDFGWTR2_EL2_nPMUACR_EL1
+ /* PMZR_EL0 is write-only, so it has no read trap to disable */
+ orr x2, x2, #HDFGWTR2_EL2_nPMZR_EL0
.Lskip_pmuv3p9_\@:
/* If SPE is implemented, */
__spe_vers_imp .Lskip_spefds_\@, ID_AA64DFR0_EL1_PMSVer_IMP, x1
@@ -436,10 +442,11 @@
cbz x1, .Lskip_spefds_\@
/* disable traps of PMSDSFR to EL2. */
orr x0, x0, #HDFGRTR2_EL2_nPMSDSFR_EL1
+ orr x2, x2, #HDFGWTR2_EL2_nPMSDSFR_EL1
.Lskip_spefds_\@:
msr_s SYS_HDFGRTR2_EL2, x0
- msr_s SYS_HDFGWTR2_EL2, x0
+ msr_s SYS_HDFGWTR2_EL2, x2
msr_s SYS_HFGRTR2_EL2, xzr
msr_s SYS_HFGWTR2_EL2, xzr
msr_s SYS_HFGITR2_EL2, xzr
base-commit: 93f51579e7df248780214094418f205253383cc5
--
2.39.5
On Tue, 22 Sep 2026 19:14:30 +0100, Fuad Tabba wrote:
> __init_el2_fgt2() writes one mask to both HDFGRTR2_EL2 and HDFGWTR2_EL2.
> PMZR_EL0 is write-only, so its trap bit, nPMZR_EL0, exists only in
> HDFGWTR2_EL2 and is therefore never set: a PMZR_EL0 write from the host
> traps to EL2, where the nVHE hypervisor has no handler and BUG()s. The
> kernel never writes PMZR_EL0, but kernel.perf_user_access=1 has the PMU
> driver set PMUSERENR_EL0.UEN for a task with a user-read event, so a
> write from EL0 reaches the trap and takes the host down without a panic
> message.
>
> [...]
Applied to arm64 (for-next/fixes), thanks!
[1/1] arm64/boot: Disable trapping of PMZR_EL0 writes to EL2
https://git.kernel.org/arm64/c/2bc6b218717b
Cheers,
--
Will
https://fixes.arm64.dev
https://next.arm64.dev
https://will.arm64.dev
On Tue, Sep 22, 2026 at 07:14:30PM +0100, Fuad Tabba wrote:
> __init_el2_fgt2() writes one mask to both HDFGRTR2_EL2 and HDFGWTR2_EL2.
> PMZR_EL0 is write-only, so its trap bit, nPMZR_EL0, exists only in
> HDFGWTR2_EL2 and is therefore never set: a PMZR_EL0 write from the host
> traps to EL2, where the nVHE hypervisor has no handler and BUG()s. The
> kernel never writes PMZR_EL0, but kernel.perf_user_access=1 has the PMU
> driver set PMUSERENR_EL0.UEN for a task with a user-read event, so a
> write from EL0 reaches the trap and takes the host down without a panic
> message.
>
> Accumulate the HDFGWTR2_EL2 bits separately, as __init_el2_fgt() already
> does for HDFGWTR_EL2, and set nPMZR_EL0 with the other FEAT_PMUv3p9
> bits.
>
> Fixes: 858c7bfcb35e1 ("arm64/boot: Enable EL2 requirements for FEAT_PMUv3p9")
> Cc: stable@vger.kernel.org
> Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
Reviewed-by: Anshuman Khandual <anshuman.khandual@arm.com>
> ---
> Based on Linux 7.3-rc4 (93f51579e7df2).
>
> Documentation/arch/arm64/booting.rst | 1 +
> arch/arm64/include/asm/el2_setup.h | 9 ++++++++-
> 2 files changed, 9 insertions(+), 1 deletion(-)
>
> diff --git a/Documentation/arch/arm64/booting.rst b/Documentation/arch/arm64/booting.rst
> index 13ef311dace83..3fea4b14ef7c2 100644
> --- a/Documentation/arch/arm64/booting.rst
> +++ b/Documentation/arch/arm64/booting.rst
> @@ -465,6 +465,7 @@ Before jumping into the kernel, the following conditions must be met:
> - HDFGWTR2_EL2.nPMICNTR_EL0 (bit 2) must be initialised to 0b1.
> - HDFGWTR2_EL2.nPMICFILTR_EL0 (bit 3) must be initialised to 0b1.
> - HDFGWTR2_EL2.nPMUACR_EL1 (bit 4) must be initialised to 0b1.
> + - HDFGWTR2_EL2.nPMZR_EL0 (bit 21) must be initialised to 0b1.
>
> For CPUs with SPE data source filtering (FEAT_SPE_FDS):
>
> diff --git a/arch/arm64/include/asm/el2_setup.h b/arch/arm64/include/asm/el2_setup.h
> index aa8ec9df80243..87560d8b254e6 100644
> --- a/arch/arm64/include/asm/el2_setup.h
> +++ b/arch/arm64/include/asm/el2_setup.h
> @@ -418,6 +418,7 @@
> b.lt .Lskip_fgt2_\@
>
> mov x0, xzr
> + mov x2, xzr
> mrs x1, id_aa64dfr0_el1
> ubfx x1, x1, #ID_AA64DFR0_EL1_PMUVer_SHIFT, #4
> cmp x1, #ID_AA64DFR0_EL1_PMUVer_V3P9
> @@ -426,6 +427,11 @@
> orr x0, x0, #HDFGRTR2_EL2_nPMICNTR_EL0
> orr x0, x0, #HDFGRTR2_EL2_nPMICFILTR_EL0
> orr x0, x0, #HDFGRTR2_EL2_nPMUACR_EL1
> + orr x2, x2, #HDFGWTR2_EL2_nPMICNTR_EL0
> + orr x2, x2, #HDFGWTR2_EL2_nPMICFILTR_EL0
> + orr x2, x2, #HDFGWTR2_EL2_nPMUACR_EL1
> + /* PMZR_EL0 is write-only, so it has no read trap to disable */
> + orr x2, x2, #HDFGWTR2_EL2_nPMZR_EL0
> .Lskip_pmuv3p9_\@:
> /* If SPE is implemented, */
> __spe_vers_imp .Lskip_spefds_\@, ID_AA64DFR0_EL1_PMSVer_IMP, x1
> @@ -436,10 +442,11 @@
> cbz x1, .Lskip_spefds_\@
> /* disable traps of PMSDSFR to EL2. */
> orr x0, x0, #HDFGRTR2_EL2_nPMSDSFR_EL1
> + orr x2, x2, #HDFGWTR2_EL2_nPMSDSFR_EL1
>
> .Lskip_spefds_\@:
> msr_s SYS_HDFGRTR2_EL2, x0
> - msr_s SYS_HDFGWTR2_EL2, x0
> + msr_s SYS_HDFGWTR2_EL2, x2
> msr_s SYS_HFGRTR2_EL2, xzr
> msr_s SYS_HFGWTR2_EL2, xzr
> msr_s SYS_HFGITR2_EL2, xzr
>
> base-commit: 93f51579e7df248780214094418f205253383cc5
> --
> 2.39.5
>
Hi Fuad, On Tue, Sep 22, 2026 at 07:14:30PM +0100, Fuad Tabba wrote: > @@ -426,6 +427,11 @@ > orr x0, x0, #HDFGRTR2_EL2_nPMICNTR_EL0 > orr x0, x0, #HDFGRTR2_EL2_nPMICFILTR_EL0 > orr x0, x0, #HDFGRTR2_EL2_nPMUACR_EL1 > + orr x2, x2, #HDFGWTR2_EL2_nPMICNTR_EL0 > + orr x2, x2, #HDFGWTR2_EL2_nPMICFILTR_EL0 > + orr x2, x2, #HDFGWTR2_EL2_nPMUACR_EL1 > + /* PMZR_EL0 is write-only, so it has no read trap to disable */ > + orr x2, x2, #HDFGWTR2_EL2_nPMZR_EL0 We can trim 3 instructions from this by using the read-trap bits as the starting point for the write-trap bits, like below. It also makes clear the point that most of our trap bits are symmetrical for read and write. Thanks, Oliver diff --git a/arch/arm64/include/asm/el2_setup.h b/arch/arm64/include/asm/el2_setup.h index aa8ec9df8024..6c5d2cc4e6fa 100644 --- a/arch/arm64/include/asm/el2_setup.h +++ b/arch/arm64/include/asm/el2_setup.h @@ -418,6 +418,7 @@ b.lt .Lskip_fgt2_\@ mov x0, xzr + mov x2, xzr mrs x1, id_aa64dfr0_el1 ubfx x1, x1, #ID_AA64DFR0_EL1_PMUVer_SHIFT, #4 cmp x1, #ID_AA64DFR0_EL1_PMUVer_V3P9 @@ -426,6 +427,9 @@ orr x0, x0, #HDFGRTR2_EL2_nPMICNTR_EL0 orr x0, x0, #HDFGRTR2_EL2_nPMICFILTR_EL0 orr x0, x0, #HDFGRTR2_EL2_nPMUACR_EL1 + + /* PMZR_EL0 is write-only, so it has no read trap to disable */ + orr x2, x0, #HDFGWTR2_EL2_nPMZR_EL0 .Lskip_pmuv3p9_\@: /* If SPE is implemented, */ __spe_vers_imp .Lskip_spefds_\@, ID_AA64DFR0_EL1_PMSVer_IMP, x1 @@ -439,7 +443,7 @@ .Lskip_spefds_\@: msr_s SYS_HDFGRTR2_EL2, x0 - msr_s SYS_HDFGWTR2_EL2, x0 + msr_s SYS_HDFGWTR2_EL2, x2 msr_s SYS_HFGRTR2_EL2, xzr msr_s SYS_HFGWTR2_EL2, xzr msr_s SYS_HFGITR2_EL2, xzr
Hi Oliver, On Tue, 22 Sep 2026 20:47:58 +0100, Oliver Upton <oupton@kernel.org> wrote: [...] > We can trim 3 instructions from this by using the read-trap bits as the > starting point for the write-trap bits, like below. It also makes clear > the point that most of our trap bits are symmetrical for read and write. I did consider that, and went the other way for two reasons. The read side has three bits with no write-side counterpart (nPMSSDATA, nSPMDEVAFF_EL1, nSPMID), so copying x0 into x2 writes ones into three RES0 bits of HDFGWTR2_EL2. Harmless today, but it's this bug in reverse: the first revision that defines one of those positions in the write register turns the copy into a silent behaviour change. Building each mask from its own register's names can't do that. It also matches __init_el2_fgt, which already keeps x0 and x2 apart for HDFGRTR_EL2 and HDFGWTR_EL2. And even in that form the copy would have to sit just before the msr_s, after the SPE block, or nPMSDSFR_EL1 only lands in x0. Cheers, /fuad
On Tue, Sep 22, 2026 at 10:09:37PM +0100, Fuad Tabba wrote: > Hi Oliver, > > On Tue, 22 Sep 2026 20:47:58 +0100, Oliver Upton <oupton@kernel.org> wrote: > [...] > > We can trim 3 instructions from this by using the read-trap bits as the > > starting point for the write-trap bits, like below. It also makes clear > > the point that most of our trap bits are symmetrical for read and write. > > I did consider that, and went the other way for two reasons. > > The read side has three bits with no write-side counterpart > (nPMSSDATA, nSPMDEVAFF_EL1, nSPMID), so copying x0 into x2 writes ones > into three RES0 bits of HDFGWTR2_EL2. Harmless today, but it's this > bug in reverse: the first revision that defines one of those positions > in the write register turns the copy into a silent behaviour change. > Building each mask from its own register's names can't do that. My thinking was that we could have a third block, but... > It also matches __init_el2_fgt, which already keeps x0 and x2 apart > for HDFGRTR_EL2 and HDFGWTR_EL2. And even in that form the copy would > have to sit just before the msr_s, after the SPE block, or > nPMSDSFR_EL1 only lands in x0. new feature-specific initialization blocks will make this an absolute mess. So I agree, this is the right fix. Reviewed-by: Oliver Upton <oupton@kernel.org> Best, Oliver
© 2016 - 2026 Red Hat, Inc.