sample/hw_breakpoint: avoid sample hw_breakpoint recursion for arm/arm64
PATCH 1/1 sample/hw_breakpoint: avoid sample hw_breakpoint recursion for arm/arm64
From: Wang Xu <[email protected]>
For x86/ppc, hw_breakpoint is triggered after the instruction is executed.
For arm/arm64, which is triggered before the instruction executed. Arm/arm64 skips the instruction by using single step. But it only supports default overflow_handler.
This patch provides a chance to avoid sample hw_breakpoint recursion for arm/arm64 by adding 'struct perf_event_attr.bp_step'.
--- include/linux/perf_event.h | 3 +++ include/uapi/linux/perf_event.h | 3 ++- samples/hw_breakpoint/data_breakpoint.c | 1 + 3 files changed, 6 insertions(+), 1 deletion(-)
include/linux/perf_event.h +3 -0
@@ -1024,6 +1024,9 @@ extern int perf_event_output(struct perf_event *event,return true;if (unlikely (event->overflow_handler == perf_event_output_backward))return true; +/* avoid sample hw_breakpoint recursion */ +if (unlikely (event->attr.bp_step)) +return true;return false; }
include/uapi/linux/perf_event.h +2 -1
@@ -375,7 +375,8 @@ struct perf_event_attr { ksymbol : 1, /* include ksymbol events */ bpf_event : 1, /* include bpf events */ aux_output : 1, /* generate AUX records instead of events */ - __reserved_1 : 32; + bp_step : 1, + __reserved_1 : 31;union {__u32 wakeup_events;/* wakeup every n events */
samples/hw_breakpoint/data_breakpoint.c +1 -0
@@ -46,6 +46,7 @@ static int __init hw_break_module_init(void) attr.bp_addr =kallsyms_lookup_name (ksym_name); attr.bp_len =HW_BREAKPOINT_LEN_4 ; attr.bp_type =HW_BREAKPOINT_W |HW_BREAKPOINT_R ; + attr.bp_step = true; sample_hbp =register_wide_hw_breakpoint (&attr, sample_hbp_handler,NULL );if (IS_ERR ((void __force *)sample_hbp)) {
Signature · 1 line
1.8.5.6
Peter Zijlstra
maintainerQuoted 8 lines · wangxu
From: Wang Xu <[email protected]>
For x86/ppc,
hw_breakpointis triggered after the instruction is executed.For arm/arm64, which is triggered before the instruction executed. Arm/arm64 skips the instruction by using single step. But it only supports default
overflow_handler.
Where is the recusion.. ?
Quoted 2 lines
This patch provides a chance to avoid sample
hw_breakpointrecursion for arm/arm64 by adding 'struct perf_event_attr.bp_step'.
This patch also lacks justification for why this needs to come with ABI changes. There is also a distinct lack of comments.
Quoted 12 lines
Signed-off-by Wang Xu <[email protected]>--- include/linux/perf_event.h | 3 +++ include/uapi/linux/perf_event.h | 3 ++- samples/hw_breakpoint/data_breakpoint.c | 1 + 3 files changed, 6 insertions(+), 1 deletion(-)include/linux/perf_event.h +0
-0
include/linux/perf_event.h extern int perf_event_output(struct perf_event *event, · line 1029
Quoted 6 lines
include/linux/perf_event.h +3
-0@@ -1024,6 +1024,9 @@ extern int perf_event_output(struct perf_event *event,return true;if (unlikely (event->overflow_handler == perf_event_output_backward))return true; +/* avoid sample hw_breakpoint recursion */ +if (unlikely (event->attr.bp_step)) +return true;
This is just _wrong_.. it says that every event with bp_step set always is a 'default overflow handler', irrespective of what the overflow handler actually is.
Quoted 3 lines
include/linux/perf_event.h +0
-0return false; }
-----Original Message-----
From: Peter Zijlstra [mailto:[email protected]]
Sent: Thursday, September 26, 2019 5:14 PM
To: wangxu (AE) <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH] sample/hw_breakpoint: avoid sample hw_breakpoint recursion for arm/arm64
Quoted 8 lines · wangxu
From: Wang Xu <[email protected]>
For x86/ppc,
hw_breakpointis triggered after the instruction is executed.For arm/arm64, which is triggered before the instruction executed. Arm/arm64 skips the instruction by using single step. But it only supports default
overflow_handler.
Where is the recusion.. ?
For arm/arm64, hw_breakpoint is triggered before the instruction executed.
When instruction_A is triggered, watchpoint_handler() will deal with this exception, and after return instruction_A will be triggerd ...
One using samples/hw_breakpoint/data_breakpoint.c in arm/arm64 will meet this problem.
Quoted 2 lines
This patch provides a chance to avoid sample
hw_breakpointrecursion for arm/arm64 by adding 'struct perf_event_attr.bp_step'.
This patch also lacks justification for why this needs to come with ABI changes. There is also a distinct lack of comments.
I agree too. but have no better idea...
This problem is really a big pit, especially for one not familiar with implementation differences in hw breakpoint for different architectures.
Quoted 12 lines
Signed-off-by Wang Xu <[email protected]>--- include/linux/perf_event.h | 3 +++ include/uapi/linux/perf_event.h | 3 ++- samples/hw_breakpoint/data_breakpoint.c | 1 + 3 files changed, 6 insertions(+), 1 deletion(-)include/linux/perf_event.h +0
-0
include/linux/perf_event.h extern int perf_event_output(struct perf_event *event, · line 1029
Quoted 6 lines
include/linux/perf_event.h +3
-0@@ -1024,6 +1024,9 @@ extern int perf_event_output(struct perf_event *event,return true;if (unlikely (event->overflow_handler == perf_event_output_backward))return true; +/* avoid sample hw_breakpoint recursion */ +if (unlikely (event->attr.bp_step)) +return true;
This is just _wrong_.. it says that every event with bp_step set always is a 'default overflow handler', irrespective of what the overflow handler actually is.
Thanks for comments.
Function is_default_overflow_handler() was introduced in 1879445dfa7bbd6fe21b09c5cc72f4934798afed , which is only be called in arch/arm[64]/kernel/hw_breakpoint.c, and will never be used in other arch/ (I think).
But keeping is_default_overflow_handler() unchanged, changing ' if (is_default_overflow_handler(bp)) ' to ' if (is_default_overflow_handler(bp) || unlikely(event->attr.bp_step) ) ' will be better in arch/arm[64]/kernel/hw_breakpoint.c.
Quoted 3 lines
include/linux/perf_event.h +0
-0return false; }
Quoted 11 lines · wangxu
From: Wang Xu <[email protected]>
For x86/ppc,
hw_breakpointis triggered after the instruction is executed.For arm/arm64, which is triggered before the instruction executed. Arm/arm64 skips the instruction by using single step. But it only supports default
overflow_handler.This patch provides a chance to avoid sample
hw_breakpointrecursion for arm/arm64 by adding 'struct perf_event_attr.bp_step'.
Issues like this come up every so often [1], [2], [3] but I'm still of the opinion that we should rip out the perf interface to hw_breakpoint on arm64 and implement something better directly for ptrace, which is what GDB cares about. The current "let's convert to perf and back again" is a wreck, mainly because we've not been able to abstract the debug trap behaviour across different architectures. GDB just wants to poke registers, and this all gets in the way of that.
Will
[1] https://lkml.org/lkml/2018/11/15/205
[2] https://lore.kernel.org/lkml/[email protected]/
[3] https://lkml.org/lkml/2016/3/21/504