perf: arm64: Kernel support for Dwarf unwinding through SVE functions
Subsystems (3)
v2 cover perf: arm64: Kernel support for Dwarf unwinding through SVE functions
Changes since v1:
* Add Mark's review tag
* Clarify in docs that it's the SVE register length
* Split patchset into kernel side and Perf tool changes
When SVE registers are pushed onto the stack the VG register is required to unwind because the stack offsets would vary by the SVE register width at the time when the sample was taken.
These first two patches add support for sampling the VG register to the kernel and the docs. There is another patchset to add support to userspace perf.
A small change is also required to libunwind or libdw depending on which unwinder is used, and these will be published later. Without these changes Perf continues to work with both libraries, although the VG register is still not used for unwinding.
Thanks
James
James Clark (2):
perf: arm64: Add SVE vector granule register to user regs arm64/sve: Add Perf extensions documentation
Documentation/arm64/sve.rst | 20 +++++++++++++++++ arch/arm64/include/uapi/asm/perf_regs.h | 7 +++++- arch/arm64/kernel/perf_regs.c | 30 +++++++++++++++++++++++-- drivers/perf/arm_pmu.c | 2 +- 4 files changed, 55 insertions(+), 4 deletions(-)
Signature · 7 lines
2.28.0 _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
PATCH v2 1/2 perf: arm64: Add SVE vector granule register to user regs
Dwarf based unwinding in a function that pushes SVE registers onto the stack requires the unwinder to know the length of the SVE register to calculate the stack offsets correctly. This was added to the Arm specific Dwarf spec as the VG pseudo register[1].
Add the vector length at position 46 if it's requested by userspace and SVE is supported. If it's not supported then fail to open the event.
The vector length must be on each sample because it can be changed at runtime via a prctl or ptrace call. Also by adding it as a register rather than a separate attribute, minimal changes will be required in an unwinder that already indexes into the register list.
[1]: https://github.com/ARM-software/abi-aa/blob/main/aadwarf64/aadwarf64.rst
--- arch/arm64/include/uapi/asm/perf_regs.h | 7 +++++- arch/arm64/kernel/perf_regs.c | 30 +++++++++++++++++++++++-- drivers/perf/arm_pmu.c | 2 +- 3 files changed, 35 insertions(+), 4 deletions(-)
arch/arm64/include/uapi/asm/perf_regs.h +6 -1
@@ -36,6 +36,11 @@ enum perf_event_arm_regs {PERF_REG_ARM64_LR ,PERF_REG_ARM64_SP ,PERF_REG_ARM64_PC , - PERF_REG_ARM64_MAX, + +/* Extended/pseudo registers */ +PERF_REG_ARM64_VG =46 ,// SVE Vector Granule + +PERF_REG_ARM64_MAX =PERF_REG_ARM64_PC +1 , +PERF_REG_ARM64_EXTENDED_MAX =PERF_REG_ARM64_VG +1 };#endif /* _ASM_ARM64_PERF_REGS_H */
arch/arm64/kernel/perf_regs.c +28 -2
@@ -9,9 +9,27 @@#include <asm/perf_regs.h> #include <asm/ptrace.h> +static u64 perf_ext_regs_value (int idx) +{ +switch (idx) { +case PERF_REG_ARM64_VG : +if (WARN_ON_ONCE (!system_supports_sve ())) +return 0 ; + +/* +* Vector granule is current length in bits of SVE registers +* divided by 64. +*/ +return (task_get_sve_vl (current) *8 ) /64 ; +default : +WARN_ON_ONCE (true); +return 0 ; + } +} +u64 perf_reg_value (struct pt_regs *regs,int idx) { -if (WARN_ON_ONCE ((u32 )idx >=PERF_REG_ARM64_MAX )) +if (WARN_ON_ONCE ((u32 )idx >=PERF_REG_ARM64_EXTENDED_MAX ))return 0 ;/*
@@ -51,6 +69,9 @@ u64 perf_reg_value(struct pt_regs *regs, int idx)if ((u32 )idx ==PERF_REG_ARM64_PC )return regs->pc; +if ((u32 )idx >=PERF_REG_ARM64_MAX ) +return perf_ext_regs_value (idx); +return regs->regs[idx]; }
@@ -58,7 +79,12 @@ u64 perf_reg_value(struct pt_regs *regs, int idx)int perf_reg_validate (u64 mask) { -if (!mask || mask ®_RESERVED ) +u64 reserved_mask =REG_RESERVED ; + +if (system_supports_sve ()) + reserved_mask &= ~(1ULL <<PERF_REG_ARM64_VG ); + +if (!mask || mask & reserved_mask)return -EINVAL ;return 0 ;
drivers/perf/arm_pmu.c +1 -1
@@ -894,7 +894,7 @@ static struct arm_pmu *__armpmu_alloc(gfp_t flags)* pmu::filter_match callback and pmu::event_init group * validation). */ - .capabilities = PERF_PMU_CAP_HETEROGENEOUS_CPUS, + .capabilities = PERF_PMU_CAP_HETEROGENEOUS_CPUS | PERF_PMU_CAP_EXTENDED_REGS, }; pmu->attr_groups[ARMPMU_ATTR_GROUP_COMMON] =
Signature · 7 lines
2.28.0 _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Will Deacon
maintainerRe: [PATCH v2 1/2] perf: arm64: Add SVE vector granule register to user regs
Quoted 28 lines · James Clark
Dwarf based unwinding in a function that pushes SVE registers onto the stack requires the unwinder to know the length of the SVE register to calculate the stack offsets correctly. This was added to the Arm specific Dwarf spec as the VG pseudo register[1].
Add the vector length at position 46 if it's requested by userspace and SVE is supported. If it's not supported then fail to open the event.
The vector length must be on each sample because it can be changed at runtime via a prctl or ptrace call. Also by adding it as a register rather than a separate attribute, minimal changes will be required in an unwinder that already indexes into the register list.
[1]: https://github.com/ARM-software/abi-aa/blob/main/aadwarf64/aadwarf64.rst
Reviewed-by Mark Brown <[email protected]>Signed-off-by James Clark <[email protected]>--- arch/arm64/include/uapi/asm/perf_regs.h | 7 +++++- arch/arm64/kernel/perf_regs.c | 30 +++++++++++++++++++++++-- drivers/perf/arm_pmu.c | 2 +- 3 files changed, 35 insertions(+), 4 deletions(-)arch/arm64/include/uapi/asm/perf_regs.h +0
-0
arch/arm64/include/uapi/asm/perf_regs.h enum perf_event_arm_regs { · line 44
Quoted 10 lines
arch/arm64/include/uapi/asm/perf_regs.h +6
-1@@ -36,6 +36,11 @@ enum perf_event_arm_regs {PERF_REG_ARM64_LR ,PERF_REG_ARM64_SP ,PERF_REG_ARM64_PC , - PERF_REG_ARM64_MAX, + +/* Extended/pseudo registers */ +PERF_REG_ARM64_VG =46 ,// SVE Vector Granule + +PERF_REG_ARM64_MAX =PERF_REG_ARM64_PC +1 , +PERF_REG_ARM64_EXTENDED_MAX =PERF_REG_ARM64_VG +1
I think you can leave PERF_REG_ARM64_MAX alone and just add:
PERF_REG_ARM64_VG = 46, PERF_REG_ARM64_EXTENDED_MAX,
no?
Quoted 12 lines
arch/arm64/include/uapi/asm/perf_regs.h +0
-0};#endif /* _ASM_ARM64_PERF_REGS_H */ arch/arm64/kernel/perf_regs.c +2
-0@@ -9,9 +9,27 @@#include <asm/perf_regs.h> #include <asm/ptrace.h> +static u64 perf_ext_regs_value (int idx) +{
arch/arm64/kernel/perf_regs.c line 23
Quoted 10 lines
arch/arm64/kernel/perf_regs.c +10
-0+switch (idx) { +case PERF_REG_ARM64_VG : +if (WARN_ON_ONCE (!system_supports_sve ())) +return 0 ; + +/* +* Vector granule is current length in bits of SVE registers +* divided by 64. +*/ +return (task_get_sve_vl (current) *8 ) /64 ;
Is 'current' the right thing to use here? We pass the regs everywhere else, so I'd prefer to stick with that if possible.
Quoted 42 lines
arch/arm64/kernel/perf_regs.c +16
-2+default : +WARN_ON_ONCE (true); +return 0 ; + } +} +u64 perf_reg_value (struct pt_regs *regs,int idx) { -if (WARN_ON_ONCE ((u32 )idx >=PERF_REG_ARM64_MAX )) +if (WARN_ON_ONCE ((u32 )idx >=PERF_REG_ARM64_EXTENDED_MAX ))return 0 ;/* @@ -51,6 +69,9 @@ u64 perf_reg_value(struct pt_regs *regs, int idx)if ((u32 )idx ==PERF_REG_ARM64_PC )return regs->pc; +if ((u32 )idx >=PERF_REG_ARM64_MAX ) +return perf_ext_regs_value (idx); +return regs->regs[idx]; }@@ -58,7 +79,12 @@ u64 perf_reg_value(struct pt_regs *regs, int idx)int perf_reg_validate (u64 mask) { -if (!mask || mask ®_RESERVED ) +u64 reserved_mask =REG_RESERVED ; + +if (system_supports_sve ()) + reserved_mask &= ~(1ULL <<PERF_REG_ARM64_VG ); + +if (!mask || mask & reserved_mask)return -EINVAL ;return 0 ;drivers/perf/arm_pmu.c +0
-0
drivers/perf/arm_pmu.c static struct arm_pmu *__armpmu_alloc(gfp_t flags) · line 897
Quoted 5 lines
drivers/perf/arm_pmu.c +1
-1@@ -894,7 +894,7 @@ static struct arm_pmu *__armpmu_alloc(gfp_t flags)* pmu::filter_match callback and pmu::event_init group * validation). */ - .capabilities = PERF_PMU_CAP_HETEROGENEOUS_CPUS, + .capabilities = PERF_PMU_CAP_HETEROGENEOUS_CPUS | PERF_PMU_CAP_EXTENDED_REGS,
How does userspace know this capability is available? Should we advertise the set of extended registers that we support, rather than make this a one-trick pony for the vector length?
Also, you don't appear to #define PERF_REG_EXTENDED_MASK so I don't understand how userspace is supposed to interact with this. Won't has_extended_regs() always return false?
Will
Re: [PATCH v2 1/2] perf: arm64: Add SVE vector granule register to user regs
Quoted 24 lines · Will Deacon
Quoted 23 lines · James Clark
Dwarf based unwinding in a function that pushes SVE registers onto the stack requires the unwinder to know the length of the SVE register to calculate the stack offsets correctly. This was added to the Arm specific Dwarf spec as the VG pseudo register[1].
Add the vector length at position 46 if it's requested by userspace and SVE is supported. If it's not supported then fail to open the event.
The vector length must be on each sample because it can be changed at runtime via a prctl or ptrace call. Also by adding it as a register rather than a separate attribute, minimal changes will be required in an unwinder that already indexes into the register list.
[1]: https://github.com/ARM-software/abi-aa/blob/main/aadwarf64/aadwarf64.rst
Reviewed-by Mark Brown <[email protected]>Signed-off-by James Clark <[email protected]>--- arch/arm64/include/uapi/asm/perf_regs.h | 7 +++++- arch/arm64/kernel/perf_regs.c | 30 +++++++++++++++++++++++-- drivers/perf/arm_pmu.c | 2 +- 3 files changed, 35 insertions(+), 4 deletions(-)
Quoted 15 lines
Quoted 5 lines
arch/arm64/include/uapi/asm/perf_regs.h +0
-0arch/arm64/include/uapi/asm/perf_regs.h enum perf_event_arm_regs { · line 44
Quoted 10 lines
arch/arm64/include/uapi/asm/perf_regs.h +6
-1@@ -36,6 +36,11 @@ enum perf_event_arm_regs {PERF_REG_ARM64_LR ,PERF_REG_ARM64_SP ,PERF_REG_ARM64_PC , - PERF_REG_ARM64_MAX, + +/* Extended/pseudo registers */ +PERF_REG_ARM64_VG =46 ,// SVE Vector Granule + +PERF_REG_ARM64_MAX =PERF_REG_ARM64_PC +1 , +PERF_REG_ARM64_EXTENDED_MAX =PERF_REG_ARM64_VG +1
Hi Will,
Apologies for the late reply, I was on a long break.
Quoted 7 lines
I think you can leave
PERF_REG_ARM64_MAXalone and just add:PERF_REG_ARM64_VG = 46, PERF_REG_ARM64_EXTENDED_MAX,no?
Yes that would be the same. I was going for consistency with arch/powerpc/include/uapi/asm/perf_regs.h which did it this way, possibly so that ..._MAX could always be found at the end.
But I'm happy to simplify it. Although this change has already made it onto Perf's copy of this header for v5.19-rc5. It can still be changed but I'm not sure if that fact changes your preference for this?
Quoted 11 lines
Quoted 11 lines
arch/arm64/include/uapi/asm/perf_regs.h +0
-0};#endif /* _ASM_ARM64_PERF_REGS_H */ arch/arm64/kernel/perf_regs.c +0
-0@@ -9,9 +9,27 @@#include <asm/perf_regs.h> #include <asm/ptrace.h>
Quoted 15 lines
arch/arm64/kernel/perf_regs.c line 23
Quoted 12 lines
arch/arm64/kernel/perf_regs.c +12
-0+static u64 perf_ext_regs_value (int idx) +{ +switch (idx) { +case PERF_REG_ARM64_VG : +if (WARN_ON_ONCE (!system_supports_sve ())) +return 0 ; + +/* +* Vector granule is current length in bits of SVE registers +* divided by 64. +*/ +return (task_get_sve_vl (current) *8 ) /64 ;Is 'current' the right thing to use here? We pass the regs everywhere else, so I'd prefer to stick with that if possible.
I did try that to start with but it required either a change to 'struct pt_regs' which didn't feel right, or adding an additional struct to be passed around in the core perf code.
Ultimately for userspace samples, 'current' is used anyway:
regs_user->regs = task_pt_regs(current); regs_user->abi = perf_reg_abi(current);
The reason prepare_sample() and then output_sample() are called separately is to get the total size of the sample ahead of outputting it, but this is already taken into account with registers without storing them because the number of set bits are counted, not the number of saved registers.
I also referenced get_ext_regs_value() in arch/powerpc/perf/perf_regs.c which doesn't change any function signatures or structs and just grabs whichever current extra register was needed.
So TLDR: yes it would be possible to save the register somewhere and pass it around, but it would be the saved version of 'current' anyway, and it would require much bigger changes.
I can make the change if you like if it's easier to judge by seeing what it would look like?
Quoted 37 lines
Quoted 37 lines
arch/arm64/kernel/perf_regs.c +16
-2+default : +WARN_ON_ONCE (true); +return 0 ; + } +} +u64 perf_reg_value (struct pt_regs *regs,int idx) { -if (WARN_ON_ONCE ((u32 )idx >=PERF_REG_ARM64_MAX )) +if (WARN_ON_ONCE ((u32 )idx >=PERF_REG_ARM64_EXTENDED_MAX ))return 0 ;/* @@ -51,6 +69,9 @@ u64 perf_reg_value(struct pt_regs *regs, int idx)if ((u32 )idx ==PERF_REG_ARM64_PC )return regs->pc; +if ((u32 )idx >=PERF_REG_ARM64_MAX ) +return perf_ext_regs_value (idx); +return regs->regs[idx]; }@@ -58,7 +79,12 @@ u64 perf_reg_value(struct pt_regs *regs, int idx)int perf_reg_validate (u64 mask) { -if (!mask || mask ®_RESERVED ) +u64 reserved_mask =REG_RESERVED ; + +if (system_supports_sve ()) + reserved_mask &= ~(1ULL <<PERF_REG_ARM64_VG ); + +if (!mask || mask & reserved_mask)return -EINVAL ;
Quoted 13 lines
drivers/perf/arm_pmu.c static struct arm_pmu *__armpmu_alloc(gfp_t flags) · line 897
Quoted 11 lines
arch/arm64/kernel/perf_regs.c +0
-0return 0 ;drivers/perf/arm_pmu.c +1
-1@@ -894,7 +894,7 @@ static struct arm_pmu *__armpmu_alloc(gfp_t flags)* pmu::filter_match callback and pmu::event_init group * validation). */ - .capabilities = PERF_PMU_CAP_HETEROGENEOUS_CPUS, + .capabilities = PERF_PMU_CAP_HETEROGENEOUS_CPUS | PERF_PMU_CAP_EXTENDED_REGS,How does userspace know this capability is available?
It doesn't, this change is only needed kernel side because if a register matching PERF_REG_EXTENDED_MASK is requested, the PMU has to have this capability otherwise the perf_event_open is rejected.
Quoted 3 lines
Should we advertise
the set of extended registers that we support, rather than make this a one-trick pony for the vector length?
Maybe, this is another copy of what was done for extended registers on Power and x86 where the event is attempted to be opened and if it works it's assumed that it's supported:
/*
* check if the pmu supports perf extended regs, before
* returning the register mask to sample.
*/
fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
if (fd != -1) {
close(fd);
mask |= extended_mask;
}I did something similar in user space for this change:
https://www.spinics.net/lists/linux-perf-users/msg19569.html
Actually I also check HWCAP_SVE, which is not strictly necessary. I suppose we could report the bitmask of which extended registers are supported. But for me just trying to open the event works just as well.
Quoted 4 lines
Also, you don't appear to #define
PERF_REG_EXTENDED_MASKso I don't understand how userspace is supposed to interact with this. Won'thas_extended_regs()always return false?
Oops, nice catch. Yes I think I accidentally dropped this part but didn't notice because it still works without it. If I add this, then the event doesn't open unless PERF_PMU_CAP_EXTENDED_REGS is added, which is what I wanted:
#define PERF_REG_EXTENDED_MASK (~((1ULL << PERF_REG_ARM64_VG) - 1))
So the only reason to use PERF_PMU_CAP_EXTENDED_REGS is for consistency and to avoid changing PERF_REG_ARM64_MAX which could potentially break userspace programs depending on it.
James
Quoted 2 lines
Will
PATCH v2 2/2 arm64/sve: Add Perf extensions documentation
In reply to James ClarkDocument that the VG register is available in Perf samples
--- Documentation/arm64/sve.rst | 20 ++++++++++++++++++++ 1 file changed, 20 insertions(+)
Documentation/arm64/sve.rst +20 -0
@@ -402,6 +402,24 @@ The regset data starts with struct user_sve_header, containing: * Modifying the system default vector length does not affect the vector length of any existing process or thread that does not make an execve() call. +10. Perf extensions +-------------------------------- + +* The arm64 specific DWARF standard [5] added the VG (Vector Granule) register + at index 46. This register is used for DWARF unwinding when variable length + SVE registers are pushed onto the stack. + +* Its value is equivalent to the current SVE vector length (VL) in bits divided + by 64. + +* The value is included in Perf samples in the regs[46] field if + PERF_SAMPLE_REGS_USER is set and the sample_regs_user mask has bit 46 set. + +* The value is the current value at the time the sample was taken, and it can + change over time. + +* If the system doesn't support SVE when perf_event_open is called with these + settings, the event will fail to open. Appendix A. SVE programmer's model (informative) =================================================
@@ -543,3 +561,5 @@ References http://infocenter.arm.com/help/topic/com.arm.doc.ihi0055c/IHI0055C_beta_aapcs64.pdf http://infocenter.arm.com/help/topic/com.arm.doc.subset.swdev.abi/index.html Procedure Call Standard for the ARM 64-bit Architecture (AArch64) + +[5] https://github.com/ARM-software/abi-aa/blob/main/aadwarf64/aadwarf64.rst
Signature · 7 lines
2.28.0 _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH v2 2/2] arm64/sve: Add Perf extensions documentation
Quoted 3 lines · James Clark
Document that the VG register is available in Perf samples
Signed-off-by James Clark <[email protected]>
Quoted 6 lines · James Clark
Changes since v1:
* Add Mark's review tag
* Clarify in docs that it's the SVE register length
* Split patchset into kernel side and Perf tool changes
Quoted 6 lines
When SVE registers are pushed onto the stack the VG register is required to unwind because the stack offsets would vary by the SVE register width at the time when the sample was taken.
These first two patches add support for sampling the VG register to the kernel and the docs. There is another patchset to add support to userspace perf.
Hi Catalin, Will,
Since James is on vacation, just want to ping if you could pick up this two patches? Mark.B has given review tags for this patch set.
I did this is because there has another patch set in perf tool to enable SVE registsers [1], which is dependent on this patch set's merging.
Thanks,
Leo
Quoted 11 lines · Leo Yan
Quoted 10 lines · James Clark
Changes since v1:
* Add Mark's review tag
* Clarify in docs that it's the SVE register length
* Split patchset into kernel side and Perf tool changesWhen SVE registers are pushed onto the stack the VG register is required to unwind because the stack offsets would vary by the SVE register width at the time when the sample was taken.
Quoted 7 lines
Quoted 2 lines
These first two patches add support for sampling the VG register to the kernel and the docs. There is another patchset to add support to userspace perf.
Hi Catalin, Will,
Since James is on vacation, just want to ping if you could pick up this two patches? Mark.B has given review tags for this patch set.
Sorry, or the kernel patch should be via arm tree and the document patch via Jonathan's tree?
Quoted 8 lines
I did this is because there has another patch set in perf tool to enable SVE registsers [1], which is dependent on this patch set's merging.
Thanks,
Leo