On 15/08/19 11:25 AM, Leo Yan wrote:
> Arm and arm64 architecture reserve some memory regions prior to the
> symbol '_stext' and these memory regions later will be used by device
> module and BPF jit. The current code misses to consider these memory
> regions thus any address in the regions will be taken as user space
> mode, but perf cannot find the corresponding dso with the wrong CPU
> mode so we misses to generate samples for device module and BPF
> related trace data.
>
> This patch parse the link scripts to get the memory size prior to start
> address and reduce this size from 'machine>->kernel_start', then can
> get a fixed up kernel start address which contain memory regions for
> device module and BPF. Finally, machine__get_kernel_start() can reflect
> more complete kernel memory regions and perf can successfully generate
> samples.
>
> The reason for parsing the link scripts is Arm architecture changes text
> offset dependent on different platforms, which define multiple text
> offsets in $kernel/arch/arm/Makefile. This offset is decided when build
> kernel and the final value is extended in the link script, so we can
> extract the used value from the link script. We use the same way to
> parse arm64 link script as well. If fail to find the link script, the
> pre start memory size is assumed as zero, in this case it has no any
> change caused with this patch.
>
> Below is detailed info for testing this patch:
>
> - Install or build LLVM/Clang;
>
> - Configure perf with ~/.perfconfig:
>
> root@debian:~# cat ~/.perfconfig
> # this file is auto-generated.
> [llvm]
> clang-path = /mnt/build/llvm-build/build/install/bin/clang
> kbuild-dir = /mnt/linux-kernel/linux-cs-dev/
> clang-opt = "-g"
> dump-obj = true
>
> [trace]
> show_zeros = yes
> show_duration = no
> no_inherit = yes
> show_timestamp = no
> show_arg_names = no
> args_alignment = 40
> show_prefix = yes
>
> - Run 'perf trace' command with eBPF event:
>
> root@debian:~# perf trace -e string \
> -e $kernel/tools/perf/examples/bpf/augmented_raw_syscalls.c
>
> - Read eBPF program memory mapping in kernel:
>
> root@debian:~# echo 1 > /proc/sys/net/core/bpf_jit_kallsyms
> root@debian:~# cat /proc/kallsyms | grep -E "bpf_prog_.+_sys_[enter|exit]"
> ffff00000008a0d0 t bpf_prog_e470211b846088d5_sys_enter [bpf]
> ffff00000008c6a4 t bpf_prog_29c7ae234d79bd5c_sys_exit [bpf]
>
> - Launch any program which accesses file system frequently so can hit
> the system calls trace flow with eBPF event;
>
> - Capture CoreSight trace data with filtering eBPF program:
>
> root@debian:~# perf record -e cs_etm/@tmc_etr0/ \
> --filter 'filter 0xffff00000008a0d0/0x800' -a sleep 5s
>
> - Decode the eBPF program symbol 'bpf_prog_f173133dc38ccf87_sys_enter':
>
> root@debian:~# perf script -F,ip,sym
> Frame deformatter: Found 4 FSYNCS
> 0 [unknown]
> ffff00000008a1ac bpf_prog_e470211b846088d5_sys_enter
> ffff00000008a250 bpf_prog_e470211b846088d5_sys_enter
> 0 [unknown]
> ffff00000008a124 bpf_prog_e470211b846088d5_sys_enter
> 0 [unknown]
> ffff00000008a14c bpf_prog_e470211b846088d5_sys_enter
> ffff00000008a13c bpf_prog_e470211b846088d5_sys_enter
> ffff00000008a14c bpf_prog_e470211b846088d5_sys_enter
> 0 [unknown]
> ffff00000008a180 bpf_prog_e470211b846088d5_sys_enter
> 0 [unknown]
> ffff00000008a1ac bpf_prog_e470211b846088d5_sys_enter
> ffff00000008a190 bpf_prog_e470211b846088d5_sys_enter
> ffff00000008a1ac bpf_prog_e470211b846088d5_sys_enter
> ffff00000008a250 bpf_prog_e470211b846088d5_sys_enter
> 0 [unknown]
> ffff00000008a124 bpf_prog_e470211b846088d5_sys_enter
> 0 [unknown]
> ffff00000008a14c bpf_prog_e470211b846088d5_sys_enter
> 0 [unknown]
> ffff00000008a180 bpf_prog_e470211b846088d5_sys_enter
> [...]
>
> Cc: Mathieu Poirier <mathieu.poir...@linaro.org>
> Cc: Alexander Shishkin <alexander.shish...@linux.intel.com>
> Cc: Jiri Olsa <jo...@redhat.com>
> Cc: Namhyung Kim <namhy...@kernel.org>
> Cc: Peter Zijlstra <pet...@infradead.org>
> Cc: Suzuki Poulouse <suzuki.poul...@arm.com>
> Cc: Adrian Hunter <adrian.hun...@intel.com>
> Cc: coresi...@lists.linaro.org
> Cc: linux-arm-ker...@lists.infradead.org
> Signed-off-by: Leo Yan <leo....@linaro.org>
> ---
> tools/perf/Makefile.config | 22 ++++++++++++++++++++++
> tools/perf/util/machine.c | 15 ++++++++++++++-
> 2 files changed, 36 insertions(+), 1 deletion(-)
>
> 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?
How come you cannot use kallsyms to get the information?
> 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;
> }
>
