Re: [PATCH v5] perf machine: arm/arm64: Improve completeness for kernel address space

From: Adrian Hunter
Date: Thu Aug 15 2019 - 07:47:13 EST


On 15/08/19 2:32 PM, Leo Yan wrote:
> Hi Adrian,
>
> On Thu, Aug 15, 2019 at 11:54:54AM +0300, Adrian Hunter wrote:
>
> [...]
>
>>> diff --git a/tools/perf/Makefile.config b/tools/perf/Makefile.config
>>> index e4988f49ea79..d7ff839d8b20 100644
>>> --- a/tools/perf/Makefile.config
>>> +++ b/tools/perf/Makefile.config
>>> @@ -48,9 +48,20 @@ ifeq ($(SRCARCH),x86)
>>> NO_PERF_REGS := 0
>>> endif
>>>
>>> +ARM_PRE_START_SIZE := 0
>>> +
>>> ifeq ($(SRCARCH),arm)
>>> NO_PERF_REGS := 0
>>> LIBUNWIND_LIBS = -lunwind -lunwind-arm
>>> + ifneq ($(wildcard $(srctree)/arch/$(SRCARCH)/kernel/vmlinux.lds),)
>>> + # Extract info from lds:
>>> + # . = ((0xC0000000)) + 0x00208000;
>>> + # ARM_PRE_START_SIZE := 0x00208000
>>> + ARM_PRE_START_SIZE := $(shell egrep ' \. \= \({2}0x[0-9a-fA-F]+\){2}' \
>>> + $(srctree)/arch/$(SRCARCH)/kernel/vmlinux.lds | \
>>> + sed -e 's/[(|)|.|=|+|<|;|-]//g' -e 's/ \+/ /g' -e 's/^[ \t]*//' | \
>>> + awk -F' ' '{printf "0x%x", $$2}' 2>/dev/null)
>>> + endif
>>> endif
>>>
>>> ifeq ($(SRCARCH),arm64)
>>> @@ -58,8 +69,19 @@ ifeq ($(SRCARCH),arm64)
>>> NO_SYSCALL_TABLE := 0
>>> CFLAGS += -I$(OUTPUT)arch/arm64/include/generated
>>> LIBUNWIND_LIBS = -lunwind -lunwind-aarch64
>>> + ifneq ($(wildcard $(srctree)/arch/$(SRCARCH)/kernel/vmlinux.lds),)
>>> + # Extract info from lds:
>>> + # . = ((((((((0xffffffffffffffff)) - (((1)) << (48)) + 1) + (0)) + (0x08000000))) + (0x08000000))) + 0x00080000;
>>> + # ARM_PRE_START_SIZE := (0x08000000 + 0x08000000 + 0x00080000) = 0x10080000
>>> + ARM_PRE_START_SIZE := $(shell egrep ' \. \= \({8}0x[0-9a-fA-F]+\){2}' \
>>> + $(srctree)/arch/$(SRCARCH)/kernel/vmlinux.lds | \
>>> + sed -e 's/[(|)|.|=|+|<|;|-]//g' -e 's/ \+/ /g' -e 's/^[ \t]*//' | \
>>> + awk -F' ' '{printf "0x%x", $$6+$$7+$$8}' 2>/dev/null)
>>> + endif
>>
>> So, that is not going to work if you take a perf.data file to a non-arm machine?
>
> Yeah, this patch will only allow perf to work correctly when perf
> run natively on arm/arm64, so it can resolve partial of the issue.
>
>> How come you cannot use kallsyms to get the information?
>
> Thanks for pointing out this. Sorry I skipped your comment "I don't
> know how you intend to calculate ARM_PRE_START_SIZE" when you reviewed
> the patch v3, I should use that chance to elaborate the detailed idea
> and so can get more feedback/guidance before procceed.
>
> Actually, I have considered to use kallsyms when worked on the previous
> patch set.
>
> As mentioned in patch set v4's cover letter, I tried to implement
> machine__create_extra_kernel_maps() for arm/arm64, the purpose is to
> parse kallsyms so can find more kernel maps and thus also can fixup
> the kernel start address. But I found the 'perf script' tool directly
> calls machine__get_kernel_start() instead of running into the flow for
> machine__create_extra_kernel_maps();

Doesn't it just need to loop through each kernel map to find the lowest
start address?

> so I finally gave up to use
> machine__create_extra_kernel_maps() for tweaking kernel start address
> and went back to use this patch's approach by parsing lds files.
>
> So for next step, I want to get some guidances:
>
> - One method is to add a new weak function, e.g.
> arch__fix_kernel_text_start(), then every arch can implement its own
> function to fixup the kernel start address;
>
> For arm/arm64, can use kallsyms to find the symbols with least
> address and fixup for kernel start address.
>
> - Another method is to directly parse kallsyms in the function
> machine__get_kernel_start(), thus the change can be used for all
> archs;
>
> Seems to me the second method is to address this issue as a common
> issue crossing all archs. But not sure if this is the requirement for
> all archs or just this is only required for arm/arm64. Please let me
> know what's your preference or other thoughts. Thanks a lot!
>
> Leo.
>
>>> endif
>>>
>>> +CFLAGS += -DARM_PRE_START_SIZE=$(ARM_PRE_START_SIZE)
>>> +
>>> ifeq ($(SRCARCH),csky)
>>> NO_PERF_REGS := 0
>>> endif
>>> diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
>>> index f6ee7fbad3e4..e993f891bb82 100644
>>> --- a/tools/perf/util/machine.c
>>> +++ b/tools/perf/util/machine.c
>>> @@ -2687,13 +2687,26 @@ int machine__get_kernel_start(struct machine *machine)
>>> machine->kernel_start = 1ULL << 63;
>>> if (map) {
>>> err = map__load(map);
>>> + if (err)
>>> + return err;
>>> +
>>> /*
>>> * On x86_64, PTI entry trampolines are less than the
>>> * start of kernel text, but still above 2^63. So leave
>>> * kernel_start = 1ULL << 63 for x86_64.
>>> */
>>> - if (!err && !machine__is(machine, "x86_64"))
>>> + if (!machine__is(machine, "x86_64"))
>>> machine->kernel_start = map->start;
>>> +
>>> + /*
>>> + * On arm/arm64, the kernel uses some memory regions which are
>>> + * prior to '_stext' symbol; to reflect the complete kernel
>>> + * address space, compensate these pre-defined regions for
>>> + * kernel start address.
>>> + */
>>> + if (!strcmp(perf_env__arch(machine->env), "arm") ||
>>> + !strcmp(perf_env__arch(machine->env), "arm64"))
>>> + machine->kernel_start -= ARM_PRE_START_SIZE;
>>> }
>>> return err;
>>> }
>>>
>>
>