Re: [PATCH v3] perf cs-etm: Improve completeness for kernel address space

From: Leo Yan
Date: Tue Jul 09 2019 - 01:12:12 EST


Hi Mathieu,

On Mon, Jul 08, 2019 at 11:33:59AM -0600, Mathieu Poirier wrote:
> On Wed, 19 Jun 2019 at 21:45, Leo Yan <leo.yan@xxxxxxxxxx> 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 'etmq->etm->kernel_start', then can
> > get a fixed up kernel start address which contain memory regions for
> > device module and BPF. Finally, cs_etm__cpu_mode() can return right
> > mode for these 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:
> >
> > - Build LLVM/Clang 8.0 or later version;
> >
> > - 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]"
> > ffff000000086a84 t bpf_prog_f173133dc38ccf87_sys_enter [bpf]
> > ffff000000088618 t bpf_prog_c1bd85c092d6e4aa_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/@20070000.etr/ \
> > --filter 'filter 0xffff000000086a84/0x800' -a sleep 5s
> >
> > - Annotate for symbol 'bpf_prog_f173133dc38ccf87_sys_enter':
> >
> > root@debian:~# perf report
> > Then select 'branches' samples and press 'a' to annotate symbol
> > 'bpf_prog_f173133dc38ccf87_sys_enter', press 'P' to print to the
> > bpf_prog_f173133dc38ccf87_sys_enter.annotation file:
> >
> > root@debian:~# cat bpf_prog_f173133dc38ccf87_sys_enter.annotation
> >
> > bpf_prog_f173133dc38ccf87_sys_enter() bpf_prog_f173133dc38ccf87_sys_enter
> > Event: branches
> >
> > Percent int sys_enter(struct syscall_enter_args *args)
> > stp x29, x30, [sp, #-16]!
> >
> > int key = 0;
> > mov x29, sp
> >
> > augmented_args = bpf_map_lookup_elem(&augmented_filename_map, &key);
> > stp x19, x20, [sp, #-16]!
> >
> > augmented_args = bpf_map_lookup_elem(&augmented_filename_map, &key);
> > stp x21, x22, [sp, #-16]!
> >
> > stp x25, x26, [sp, #-16]!
> >
> > return bpf_get_current_pid_tgid();
> > mov x25, sp
> >
> > return bpf_get_current_pid_tgid();
> > mov x26, #0x0 // #0
> >
> > sub sp, sp, #0x10
> >
> > return bpf_map_lookup_elem(pids, &pid) != NULL;
> > add x19, x0, #0x0
> >
> > mov x0, #0x0 // #0
> >
> > mov x10, #0xfffffffffffffff8 // #-8
> >
> > if (pid_filter__has(&pids_filtered, getpid()))
> > str w0, [x25, x10]
> >
> > probe_read(&augmented_args->args, sizeof(augmented_args->args), args);
> > add x1, x25, #0x0
> >
> > probe_read(&augmented_args->args, sizeof(augmented_args->args), args);
> > mov x10, #0xfffffffffffffff8 // #-8
> >
> > syscall = bpf_map_lookup_elem(&syscalls, &augmented_args->args.syscall_nr);
> > add x1, x1, x10
> >
> > syscall = bpf_map_lookup_elem(&syscalls, &augmented_args->args.syscall_nr);
> > mov x0, #0xffff8009ffffffff // #-140694538682369
> >
> > movk x0, #0x6698, lsl #16
> >
> > movk x0, #0x3e00
> >
> > mov x10, #0xffffffffffff1040 // #-61376
> >
> > if (syscall == NULL || !syscall->enabled)
> > movk x10, #0x1023, lsl #16
> >
> > if (syscall == NULL || !syscall->enabled)
> > movk x10, #0x0, lsl #32
> >
> > loop_iter_first()
> > 3.69 â blr bpf_prog_f173133dc38ccf87_sys_enter
> > loop_iter_first()
> > add x7, x0, #0x0
> >
> > loop_iter_first()
> > add x20, x7, #0x0
> >
> > int size = probe_read_str(&augmented_filename->value, filename_len, filename_arg);
> > mov x0, #0x1 // #1
>
> I'm not sure all this information about annotation should be in the
> changelog. This patch is about being able to decode traces that
> executed outside the current kernel addresse range and as such simply
> using "perf report" or "perf script" successfully is enough to test
> this set. Any information that goes beyond that muddies the water.

Agree. Will remove this in the new patch.

> > [...]
> >
> > Cc: Mathieu Poirier <mathieu.poirier@xxxxxxxxxx>
> > Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
> > Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
> > Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
> > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > Cc: Suzuki Poulouse <suzuki.poulose@xxxxxxx>
> > Cc: coresight@xxxxxxxxxxxxxxxx
> > Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
> > Signed-off-by: Leo Yan <leo.yan@xxxxxxxxxx>
> > ---
> > tools/perf/Makefile.config | 22 ++++++++++++++++++++++
> > tools/perf/util/cs-etm.c | 19 ++++++++++++++++++-
> > 2 files changed, 40 insertions(+), 1 deletion(-)
> >
> > diff --git a/tools/perf/Makefile.config b/tools/perf/Makefile.config
> > index 51dd00f65709..a58cd5a43a98 100644
> > --- a/tools/perf/Makefile.config
> > +++ b/tools/perf/Makefile.config
> > @@ -418,6 +418,28 @@ ifdef CORESIGHT
> > endif
> > LDFLAGS += $(LIBOPENCSD_LDFLAGS)
> > EXTLIBS += $(OPENCSDLIBS)
> > + PRE_START_SIZE := 0
> > + ifneq ($(wildcard $(srctree)/arch/$(SRCARCH)/kernel/vmlinux.lds),)
> > + ifeq ($(SRCARCH),arm64)
> > + # Extract info from lds:
> > + # . = ((((((((0xffffffffffffffff)) - (((1)) << (48)) + 1) + (0)) + (0x08000000))) + (0x08000000))) + 0x00080000;
> > + # PRE_START_SIZE := (0x08000000 + 0x08000000 + 0x00080000) = 0x10080000
> > + 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
> > + ifeq ($(SRCARCH),arm)
> > + # Extract info from lds:
> > + # . = ((0xC0000000)) + 0x00208000;
> > + # PRE_START_SIZE := 0x00208000
> > + 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
> > + CFLAGS += -DARM_PRE_START_SIZE=$(PRE_START_SIZE)
>
> It might be useful to do this for arm and arm64 regardless of
> CoreSight but I'll let Arnaldo decide on this.

Ah, good point! On Arm/arm64 platforms, if we want to parse kernel
address space with considering eBPF/module, the kernel start address
needs to be compensated in the function machine__get_kernel_start(),
so this can let PMU or other events to work correctly.

Will wait a bit if Arnaldo could give out guidance.

> > $(call detected,CONFIG_LIBOPENCSD)
> > ifdef CSTRACE_RAW
> > CFLAGS += -DCS_DEBUG_RAW
> > diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c
> > index 0c7776b51045..5fa0be3a3904 100644
> > --- a/tools/perf/util/cs-etm.c
> > +++ b/tools/perf/util/cs-etm.c
> > @@ -613,10 +613,27 @@ static void cs_etm__free(struct perf_session *session)
> > static u8 cs_etm__cpu_mode(struct cs_etm_queue *etmq, u64 address)
> > {
> > struct machine *machine;
> > + u64 fixup_kernel_start = 0;
> >
> > machine = etmq->etm->machine;
> >
> > - if (address >= etmq->etm->kernel_start) {
> > + /*
> > + * Since arm and arm64 specify some memory regions prior to
> > + * 'kernel_start', kernel addresses can be less than 'kernel_start'.
> > + *
> > + * For arm architecture, the 16MB virtual memory space prior to
> > + * 'kernel_start' is allocated to device modules, a PMD table if
> > + * CONFIG_HIGHMEM is enabled and a PGD table.
> > + *
> > + * For arm64 architecture, the root PGD table, device module memory
> > + * region and BPF jit region are prior to 'kernel_start'.
> > + *
> > + * To reflect the complete kernel address space, compensate these
> > + * pre-defined regions for kernel start address.
> > + */
> > + fixup_kernel_start = etmq->etm->kernel_start - ARM_PRE_START_SIZE;
> > +
> > + if (address >= fixup_kernel_start) {
> > if (machine__is_host(machine))
> > return PERF_RECORD_MISC_KERNEL;
> > else
>
> Tested-by: Mathieu Poirier <mathieu.poirier@xxxxxxxxxx>

Thanks a lot for the reviewing & testing; and very appreciate your much
consumed time :)

Thanks,
Leo Yan