This patchset tries to address some arm64 perf issues.
While the 1st patch is a generic one and just removes a dead/left-over
declaration:
[PATCH 1/2]: hw_breakpoint: Remove unused '__register_perf_hw_breakpoint' declaration
the 2nd patch addresses 'perf stat -e' like use-cases on arm64 which
allow per-task address execution h/w breakpoints:
[PATCH 2/2]: perf/arm64: Allow per-task kernel breakpoints
Cc: Mark Rutland <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Catalin Marinas <[email protected]>
Cc: Ard Biesheuvel <[email protected]>
Bhupesh Sharma (2):
hw_breakpoint: Remove unused '__register_perf_hw_breakpoint'
declaration
perf/arm64: Allow per-task kernel breakpoints
arch/arm64/kernel/hw_breakpoint.c | 7 -------
include/linux/hw_breakpoint.h | 3 ---
2 files changed, 10 deletions(-)
--
2.7.4
commit b326e9560a28 ("hw-breakpoints: Use overflow handler instead of
the event callback") removed '__register_perf_hw_breakpoint' function
usage and replaced it with 'register_perf_hw_breakpoint' function.
Remove the left-over unused '__register_perf_hw_breakpoint' declaration
from 'hw_breakpoint.h' as well.
Cc: Mark Rutland <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Catalin Marinas <[email protected]>
Cc: Ard Biesheuvel <[email protected]>
Signed-off-by: Bhupesh Sharma <[email protected]>
---
include/linux/hw_breakpoint.h | 3 ---
1 file changed, 3 deletions(-)
diff --git a/include/linux/hw_breakpoint.h b/include/linux/hw_breakpoint.h
index 6058c3844a76..fe1302da8e0f 100644
--- a/include/linux/hw_breakpoint.h
+++ b/include/linux/hw_breakpoint.h
@@ -72,7 +72,6 @@ register_wide_hw_breakpoint(struct perf_event_attr *attr,
void *context);
extern int register_perf_hw_breakpoint(struct perf_event *bp);
-extern int __register_perf_hw_breakpoint(struct perf_event *bp);
extern void unregister_hw_breakpoint(struct perf_event *bp);
extern void unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events);
@@ -115,8 +114,6 @@ register_wide_hw_breakpoint(struct perf_event_attr *attr,
void *context) { return NULL; }
static inline int
register_perf_hw_breakpoint(struct perf_event *bp) { return -ENOSYS; }
-static inline int
-__register_perf_hw_breakpoint(struct perf_event *bp) { return -ENOSYS; }
static inline void unregister_hw_breakpoint(struct perf_event *bp) { }
static inline void
unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events) { }
--
2.7.4
commit 478fcb2cdb23 ("arm64: Debugging support") disallowed per-task
kernel breakpoints on arm64 since these would have potentially
complicated the stepping code.
However, we now have several use-cases (for e.g. perf) which require
per-task address execution h/w breakpoint to be exercised/set on arm64:
For e.g. we can set address execution h/w breakpoints, using the
format prescribed by 'perf-list' command:
mem:<addr>[/len][:access] [Hardware breakpoint]
Without this patch, currently 'perf stat -e' reports that per-task
address execution h/w breakpoints are 'not supported' on arm64. See
below:
$ TEST_FUNC="vfs_read"
$ ADDR=0x`cat /proc/kallsyms | grep -P "\\s$TEST_FUNC\$" | cut -f1 -d' '`
$ perf stat -e mem:$ADDR:x -x';' -- cat /proc/cpuinfo > /dev/null
<not supported>;;mem:0xffff00001031dd68:x;0;100.00;;
After this patch, this use-case can be supported:
$ TEST_FUNC="vfs_read"
$ ADDR=0x`cat /proc/kallsyms | grep -P "\\s$TEST_FUNC\$" | cut -f1 -d' '`
$ perf stat -e mem:$ADDR:x -x';' -- cat /proc/cpuinfo > /dev/null
5;;mem:0xfffffe0010361d20:x;912200;100.00;;
Cc: Mark Rutland <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Catalin Marinas <[email protected]>
Cc: Ard Biesheuvel <[email protected]>
Signed-off-by: Bhupesh Sharma <[email protected]>
---
arch/arm64/kernel/hw_breakpoint.c | 7 -------
1 file changed, 7 deletions(-)
diff --git a/arch/arm64/kernel/hw_breakpoint.c b/arch/arm64/kernel/hw_breakpoint.c
index 0b727edf4104..c28f04e02845 100644
--- a/arch/arm64/kernel/hw_breakpoint.c
+++ b/arch/arm64/kernel/hw_breakpoint.c
@@ -562,13 +562,6 @@ int hw_breakpoint_arch_parse(struct perf_event *bp,
hw->address &= ~alignment_mask;
hw->ctrl.len <<= offset;
- /*
- * Disallow per-task kernel breakpoints since these would
- * complicate the stepping code.
- */
- if (hw->ctrl.privilege == AARCH64_BREAKPOINT_EL1 && bp->hw.target)
- return -EINVAL;
-
return 0;
}
--
2.7.4
On Tue, Feb 04, 2020 at 03:56:24AM +0530, Bhupesh Sharma wrote:
> commit 478fcb2cdb23 ("arm64: Debugging support") disallowed per-task
> kernel breakpoints on arm64 since these would have potentially
> complicated the stepping code.
>
> However, we now have several use-cases (for e.g. perf) which require
> per-task address execution h/w breakpoint to be exercised/set on arm64:
To be honest, the perf interface to hw_breakpoint is an abomination and
I think we should remove it entirely for arm64. It's flakey, complicated,
adds code to context-switch and reduces the capabilities available to
ptrace.
> For e.g. we can set address execution h/w breakpoints, using the
> format prescribed by 'perf-list' command:
> mem:<addr>[/len][:access] [Hardware breakpoint]
>
> Without this patch, currently 'perf stat -e' reports that per-task
> address execution h/w breakpoints are 'not supported' on arm64. See
> below:
>
> $ TEST_FUNC="vfs_read"
> $ ADDR=0x`cat /proc/kallsyms | grep -P "\\s$TEST_FUNC\$" | cut -f1 -d' '`
> $ perf stat -e mem:$ADDR:x -x';' -- cat /proc/cpuinfo > /dev/null
>
> <not supported>;;mem:0xffff00001031dd68:x;0;100.00;;
>
> After this patch, this use-case can be supported:
>
> $ TEST_FUNC="vfs_read"
> $ ADDR=0x`cat /proc/kallsyms | grep -P "\\s$TEST_FUNC\$" | cut -f1 -d' '`
> $ perf stat -e mem:$ADDR:x -x';' -- cat /proc/cpuinfo > /dev/null
>
> 5;;mem:0xfffffe0010361d20:x;912200;100.00;;
>
> Cc: Mark Rutland <[email protected]>
> Cc: Will Deacon <[email protected]>
> Cc: Catalin Marinas <[email protected]>
> Cc: Ard Biesheuvel <[email protected]>
> Signed-off-by: Bhupesh Sharma <[email protected]>
> ---
> arch/arm64/kernel/hw_breakpoint.c | 7 -------
> 1 file changed, 7 deletions(-)
>
> diff --git a/arch/arm64/kernel/hw_breakpoint.c b/arch/arm64/kernel/hw_breakpoint.c
> index 0b727edf4104..c28f04e02845 100644
> --- a/arch/arm64/kernel/hw_breakpoint.c
> +++ b/arch/arm64/kernel/hw_breakpoint.c
> @@ -562,13 +562,6 @@ int hw_breakpoint_arch_parse(struct perf_event *bp,
> hw->address &= ~alignment_mask;
> hw->ctrl.len <<= offset;
>
> - /*
> - * Disallow per-task kernel breakpoints since these would
> - * complicate the stepping code.
> - */
> - if (hw->ctrl.privilege == AARCH64_BREAKPOINT_EL1 && bp->hw.target)
> - return -EINVAL;
> -
Sorry, but this is broken; the check is there for a reason, not just for
fun!
Look at how the step handler toggles the bp registers.
Will
Hi Will,
Thanks for your inputs.
On Thu, Feb 6, 2020 at 4:09 PM Will Deacon <[email protected]> wrote:
>
> On Tue, Feb 04, 2020 at 03:56:24AM +0530, Bhupesh Sharma wrote:
> > commit 478fcb2cdb23 ("arm64: Debugging support") disallowed per-task
> > kernel breakpoints on arm64 since these would have potentially
> > complicated the stepping code.
> >
> > However, we now have several use-cases (for e.g. perf) which require
> > per-task address execution h/w breakpoint to be exercised/set on arm64:
>
> To be honest, the perf interface to hw_breakpoint is an abomination and
> I think we should remove it entirely for arm64. It's flakey, complicated,
> adds code to context-switch and reduces the capabilities available to
> ptrace.
Sure, I agree.
> > For e.g. we can set address execution h/w breakpoints, using the
> > format prescribed by 'perf-list' command:
> > mem:<addr>[/len][:access] [Hardware breakpoint]
> >
> > Without this patch, currently 'perf stat -e' reports that per-task
> > address execution h/w breakpoints are 'not supported' on arm64. See
> > below:
> >
> > $ TEST_FUNC="vfs_read"
> > $ ADDR=0x`cat /proc/kallsyms | grep -P "\\s$TEST_FUNC\$" | cut -f1 -d' '`
> > $ perf stat -e mem:$ADDR:x -x';' -- cat /proc/cpuinfo > /dev/null
> >
> > <not supported>;;mem:0xffff00001031dd68:x;0;100.00;;
> >
> > After this patch, this use-case can be supported:
> >
> > $ TEST_FUNC="vfs_read"
> > $ ADDR=0x`cat /proc/kallsyms | grep -P "\\s$TEST_FUNC\$" | cut -f1 -d' '`
> > $ perf stat -e mem:$ADDR:x -x';' -- cat /proc/cpuinfo > /dev/null
> >
> > 5;;mem:0xfffffe0010361d20:x;912200;100.00;;
> >
> > Cc: Mark Rutland <[email protected]>
> > Cc: Will Deacon <[email protected]>
> > Cc: Catalin Marinas <[email protected]>
> > Cc: Ard Biesheuvel <[email protected]>
> > Signed-off-by: Bhupesh Sharma <[email protected]>
> > ---
> > arch/arm64/kernel/hw_breakpoint.c | 7 -------
> > 1 file changed, 7 deletions(-)
> >
> > diff --git a/arch/arm64/kernel/hw_breakpoint.c b/arch/arm64/kernel/hw_breakpoint.c
> > index 0b727edf4104..c28f04e02845 100644
> > --- a/arch/arm64/kernel/hw_breakpoint.c
> > +++ b/arch/arm64/kernel/hw_breakpoint.c
> > @@ -562,13 +562,6 @@ int hw_breakpoint_arch_parse(struct perf_event *bp,
> > hw->address &= ~alignment_mask;
> > hw->ctrl.len <<= offset;
> >
> > - /*
> > - * Disallow per-task kernel breakpoints since these would
> > - * complicate the stepping code.
> > - */
> > - if (hw->ctrl.privilege == AARCH64_BREAKPOINT_EL1 && bp->hw.target)
> > - return -EINVAL;
> > -
>
> Sorry, but this is broken; the check is there for a reason, not just for
> fun!
>
> Look at how the step handler toggles the bp registers.
Not sure I follow. Can you please give me some pointers. All the perf
tests I have from test-suite run fine with this chunk removed.
Thanks for your help.
Regards,
Bhupesh