- 01 Aug, 2024 14 commits
-
-
Namhyung Kim authored
Now it can run the BPF filtering test with normal user if the BPF objects are pinned by 'sudo perf record --setup-filter pin'. Let's update the test case to verify the behavior. It'll skip the test if the filter check is failed from a normal user, but it shows a message how to set up the filters. First, run the test as a normal user and it fails. $ perf test -vv filtering 95: perf record sample filtering (by BPF) tests: --- start --- test child forked, pid 425677 Checking BPF-filter privilege try 'sudo perf record --setup-filter pin' first. <<<--- here bpf-filter test [Skipped permission] ---- end(-2) ---- 95: perf record sample filtering (by BPF) tests : Skip According to the message, run the perf record command to pin the BPF objects. $ sudo perf record --setup-filter pin And re-run the test as a normal user. $ perf test -vv filtering 95: perf record sample filtering (by BPF) tests: --- start --- test child forked, pid 424486 Checking BPF-filter privilege Basic bpf-filter test Basic bpf-filter test [Success] Failing bpf-filter test Error: task-clock event does not have PERF_SAMPLE_CPU Failing bpf-filter test [Success] Group bpf-filter test Error: task-clock event does not have PERF_SAMPLE_CPU Error: task-clock event does not have PERF_SAMPLE_CODE_PAGE_SIZE Group bpf-filter test [Success] ---- end(0) ---- 95: perf record sample filtering (by BPF) tests : Ok Signed-off-by:
Namhyung Kim <namhyung@kernel.org> Tested-by:
Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Ian Rogers <irogers@google.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Jiri Olsa <jolsa@kernel.org> Cc: KP Singh <kpsingh@kernel.org> Cc: Kan Liang <kan.liang@linux.intel.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Song Liu <song@kernel.org> Cc: Stephane Eranian <eranian@google.com> Cc: bpf@vger.kernel.org Link: https://lore.kernel.org/r/20240703223035.2024586-9-namhyung@kernel.orgSigned-off-by:
-
Namhyung Kim authored
To allow BPF filters for unprivileged users it needs to pin the BPF objects to BPF-fs first. Let's add a new option to pin and unpin the objects easily. I'm not sure 'perf record' is a right place to do this but I don't have a better idea right now. $ sudo perf record --setup-filter pin The above command would pin BPF program and maps for the filter when the system has BPF-fs (usually at /sys/fs/bpf/). To unpin the objects, users can run the following command (as root). $ sudo perf record --setup-filter unpin Committer testing: root@number:~# perf record --setup-filter pin root@number:~# ls -la /sys/fs/bpf/perf_filter/ total 0 drwxr-xr-x. 2 root root 0 Jul 31 10:43 . drwxr-xr-t. 3 root root 0 Jul 31 10:43 .. -rw-rw-rw-. 1 root root 0 Jul 31 10:43 dropped -rw-rw-rw-. 1 root root 0 Jul 31 10:43 filters -rwxrwxrwx. 1 root root 0 Jul 31 10:43 perf_sample_filter -rw-rw-rw-. 1 root root 0 Jul 31 10:43 pid_hash -rw-------. 1 root root 0 Jul 31 10:43 sample_f_rodata root@number:~# ls -la /sys/fs/bpf/perf_filter/perf_sample_filter -rwxrwxrwx. 1 root root 0 Jul 31 10:43 /sys/fs/bpf/perf_filter/perf_sample_filter root@number:~# Signed-off-by:
-
Namhyung Kim authored
The evlist is allocated at the beginning of cmd_record(). Also free-ing thread masks should be paired with record__init_thread_masks() which is called right before __cmd_record(). Let's change the order of these functions to release the resources correctly in case of errors. This is maybe fine as the process exits, but it might be a problem if it manages some system-wide resources that live longer than the process. Signed-off-by:
-
Namhyung Kim authored
As the BPF filter is shared between other processes, it should have its own counter for each invocation. Add a new array map (lost_count) to save the count using the same index as the filter. It should clear the count before running the filter. Signed-off-by:
-
Namhyung Kim authored
And use the pinned objects for unprivileged users to profile their own tasks. The BPF objects need to be pinned in the BPF-fs by root first and it'll be handled in the later patch. Signed-off-by:
-
Namhyung Kim authored
If the target is a list of tasks, it can use a shared hash map for filter expressions. The key of the filter map is an integer index like in an array. A separate pid_hash map is added to get the index for the filter map using the tgid. For system-wide mode including per-cpu or per-user targets are handled by the single entry map like before. Signed-off-by:
-
Namhyung Kim authored
This is needed to prepare target-specific actions in the later patch. We want to reuse the pinned BPF program and map for regular users to profile their own processes. Signed-off-by:
-
Namhyung Kim authored
And the value is now an array. This is to support multiple filter entries in the map later. No functional changes intended. Signed-off-by:
-
Ian Rogers authored
-1000 was used as a special value added in Commit 3d504549 ("perf pmu-events: Add pmu_events_table__find_event()") to show that 1 table lacked a PMU/event but that didn't terminate the search in other tables. Add a new constant PMU_EVENTS__NOT_FOUND for this value and use it. Reviewed-by:
John Garry <john.g.garry@oracle.com> Signed-off-by:
Ian Rogers <irogers@google.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jing Zhang <renyu.zj@linux.alibaba.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Kan Liang <kan.liang@linux.intel.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Oliver Sang <oliver.sang@intel.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Philip Li <philip.li@intel.com> Cc: Sandipan Das <sandipan.das@amd.com> Cc: Weilin Wang <weilin.wang@intel.com> Cc: Xu Yang <xu.yang_2@nxp.com> Link: https://lore.kernel.org/r/20240730191744.3097329-2-irogers@google.comSigned-off-by:
-
Tiezhu Yang authored
When executing the command "perf list", I met "Error: failed to open tracing events directory" twice, the first reason is that there is no "/sys/kernel/tracing/events" directory due to it does not enable the kernel tracing infrastructure with CONFIG_FTRACE, the second reason is that there is no root privileges. Add the error string to tell the users what happened and what should to do, and also call put_tracing_file() to free events_path a little later to avoid messy code in the error message. At the same time, just remove the redundant "/" of the file path in the function get_tracing_file(), otherwise it shows something like "/sys/kernel/tracing//events". Before: $ ./perf list Error: failed to open tracing events directory After: (1) Without CONFIG_FTRACE $ ./perf list Error: failed to open tracing events directory /sys/kernel/tracing/events: No such file or directory (2) With CONFIG_FTRACE but no root privileges $ ./perf list Error: failed to open tracing events directory /sys/kernel/tracing/events: Permission denied Committer testing: Redirect stdout to null to quickly test the patch: Before: $ perf list > /dev/null Error: failed to open tracing events directory $ After: $ perf list > /dev/null Error: failed to open tracing events directory /sys/kernel/tracing/events: Permission denied $ Signed-off-by:
Tiezhu Yang <yangtiezhu@loongson.cn> Tested-by:
-
Tiezhu Yang authored
There is only one job when running "make JOBS=1", it should print "sequential build" rather than "parallel build". Before: $ cd tools/perf && make JOBS=1 BUILD: Doing 'make -j1' parallel build After: $ cd tools/perf && make JOBS=1 BUILD: Doing 'make -j1' sequential build Signed-off-by:
-
Charlie Jenkins authored
Ignore files that are generated by libperf and libperf tests. Signed-off-by:
Charlie Jenkins <charlie@rivosinc.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Ian Rogers <irogers@google.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: https://lore.kernel.org/lkml/20240729-libperf_gitignore-v1-1-1c70dd98edf9@rivosinc.comSigned-off-by:
-
Leo Yan authored
As the flag 'auxtrace' has been set for Arm SPE events, now it is ready to use evsel__is_aux_event() to check if an event is AUX trace event or not. Use this function to replace the old checking for only the first Arm SPE event. Signed-off-by:
Leo Yan <leo.yan@arm.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Suzuki K Poulose <suzuki.poulose@arm.com> Cc: Ian Rogers <irogers@google.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Namhyung Kim <namhyung@kernel.org> Cc: James Clark <james.clark@linaro.org> Cc: Mike Leach <mike.leach@linaro.org> Cc: <coresight@lists.linaro.org> Cc: John Garry <john.g.garry@oracle.com> Cc: <linux-perf-users@vger.kernel.org> Signed-off-by:
-
Leo Yan authored
The evsel for Arm SPE PMU needs to be set up. Extract the setting up into a function arm_spe_setup_evsel(). Signed-off-by:
-
- 31 Jul, 2024 26 commits
-
-
Weilin Wang authored
perf test: make metric validation test return early when there is no metric supported on the test system Add a check to return the metric validation test early when perf list metric does not output any metric. This would happen when NO_JEVENTS=1 is set or in a system that there is no metric supported. Signed-off-by:
Weilin Wang <weilin.wang@intel.com> Tested-by:
-
Namhyung Kim authored
The -s/--sort option is to sort the output by given column. $ sudo perf ftrace profile -s max sync | head # Total (us) Avg (us) Max (us) Count Function 6301.811 6301.811 6301.811 1 __do_sys_sync 6301.328 6301.328 6301.328 1 ksys_sync 5320.300 1773.433 2858.819 3 iterate_supers 2755.875 17.012 2610.633 162 sync_fs_one_sb 2728.351 682.088 2610.413 4 ext4_sync_fs [ext4] 2603.654 2603.654 2603.654 1 jbd2_log_wait_commit [jbd2] 4750.615 593.827 2597.427 8 schedule 2164.986 26.728 2115.673 81 sync_inodes_one_sb 2143.842 26.467 2115.438 81 sync_inodes_sb Reviewed-by: -
Namhyung Kim authored
The 'perf ftrace profile' command is to get function execution profiles using function-graph tracer so that users can see the total, average, max execution time as well as the number of invocations easily. The following is a profile for the perf_event_open syscall. $ sudo perf ftrace profile -G __x64_sys_perf_event_open -- \ perf stat -e cycles -C1 true 2> /dev/null | head # Total (us) Avg (us) Max (us) Count Function 65.611 65.611 65.611 1 __x64_sys_perf_event_open 30.527 30.527 30.527 1 anon_inode_getfile 30.260 30.260 30.260 1 __anon_inode_getfile 29.700 29.700 29.700 1 alloc_file_pseudo 17.578 17.578 17.578 1 d_alloc_pseudo 17.382 17.382 17.382 1 __d_alloc 16.738 16.738 16.738 1 kmem_cache_alloc_lru 15.686 15.686 15.686 1 perf_event_alloc 14.012 7.006 11.264 2 obj_cgroup_charge # Reviewed-by: -
Namhyung Kim authored
The check is a common part of the ftrace commands, let's move it out. Reviewed-by:
-
Namhyung Kim authored
The 'graph-tail' option is to print function name as a comment at the end. This is useful when a large function is mixed with other functions (possibly from different CPUs). For example, $ sudo perf ftrace -- perf stat true ... 1) | get_unused_fd_flags() { 1) | alloc_fd() { 1) 0.178 us | _raw_spin_lock(); 1) 0.187 us | expand_files(); 1) 0.169 us | _raw_spin_unlock(); 1) 1.211 us | } 1) 1.503 us | } $ sudo perf ftrace --graph-opts tail -- perf stat true ... 1) | get_unused_fd_flags() { 1) | alloc_fd() { 1) 0.099 us | _raw_spin_lock(); 1) 0.083 us | expand_files(); 1) 0.081 us | _raw_spin_unlock(); 1) 0.601 us | } /* alloc_fd */ 1) 0.751 us | } /* get_unused_fd_flags */ Reviewed-by: -
Dr. David Alan Gilbert authored
Commit aa1551f2 ("perf test pmu: Refactor format test and exposed test APIs") added the 'test_pmus' list, but didn't use it. (It seems to put them on the other_pmus list?) Remove it. Fixes: aa1551f2 ("perf test pmu: Refactor format test and exposed test APIs") Reviewed-by:
Dr. David Alan Gilbert <linux@treblig.org> Cc: Ian Rogers <irogers@google.com> Cc: Kan Liang <kan.liang@linux.intel.com> Link: https://lore.kernel.org/lkml/20240727175919.1041468-1-linux@treblig.orgSigned-off-by:
-
Adrian Hunter authored
evsel__is_aux_event() identifies AUX area tracing selected events. S390_CPUMSF uses a raw event type (PERF_TYPE_RAW - refer s390_cpumsf_evsel_is_auxtrace()) not a PMU type value that could be checked in evsel__is_aux_event(). However it sets needs_auxtrace_mmap (refer auxtrace_record__init()), so check that first. Currently, the features that use evsel__is_aux_event() are used only by Intel PT, but that may change in the future. Reviewed-by:
Andi Kleen <ak@linux.intel.com> Reviewed-by:
Adrian Hunter <adrian.hunter@intel.com> Acked-by:
-
Adrian Hunter authored
Set pmu->auxtrace on ARM/ARM64 AUX area PMUs. evsel__is_aux_event() needs the setting to identify AUX area tracing selected events. Currently, the features that use evsel__is_aux_event() are used only by Intel PT, but that may change in the future. Reviewed-by:
-
James Clark authored
The linked commit moved the early return on the first sample to before the verbose log, so move the log earlier too. Now the first sample is also logged and not skipped. Fixes: 2d98dbb4 ("perf scripts python arm-cs-trace-disasm.py: Do not ignore disam first sample") Reviewed-by:
James Clark <james.clark@linaro.org> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Benjamin Gray <bgray@linux.ibm.com> Cc: coresight@lists.linaro.org Cc: gankulkarni@os.amperecomputing.com Cc: Ian Rogers <irogers@google.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mike Leach <mike.leach@linaro.org> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ruidong Tian <tianruidong@linux.alibaba.com> Cc: Suzuki Poulouse <suzuki.poulose@arm.com> Link: https://lore.kernel.org/r/20240723132858.12747-1-james.clark@linaro.orgSigned-off-by:
-
James Clark authored
Normally exception packets don't directly output a branch sample, but if they're the last record in a buffer then they will. Because they don't have addresses set we'll see the placeholder value CS_ETM_INVAL_ADDR (0xdeadbeef) in the output. Since commit 6035b680 ("perf cs-etm: Support dummy address value for CS_ETM_TRACE_ON packet") we've used 0 as an externally visible "not set" address value. For consistency reasons and to not make exceptions look like an error, change them to use 0 too. This is particularly visible when doing userspace only tracing because trace is disabled when jumping to the kernel, causing the flush and then forcing the last exception packet to be emitted as a branch. With kernel trace included, there is no flush so exception packets don't generate samples until the next range packet and they'll pick up the correct address. Before: $ perf record -e cs_etm//u -- stress -i 1 -t 1 $ perf script -F comm,ip,addr,flags stress syscall ffffb7eedbc0 => deadbeefdeadbeef stress syscall ffffb7f14a14 => deadbeefdeadbeef stress syscall ffffb7eedbc0 => deadbeefdeadbeef After: stress syscall ffffb7eedbc0 => 0 stress syscall ffffb7f14a14 => 0 stress syscall ffffb7eedbc0 => 0 Reviewed-by:
Mike Leach <mike.leach@linaro.org> Signed-off-by:
-
Chen Ni authored
Replace a comma between expression statements by a semicolon. Reviewed-by:
Chen Ni <nichen@iscas.ac.cn> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Kan Liang <kan.liang@linux.intel.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Link: https://lore.kernel.org/r/20240716075347.969041-1-nichen@iscas.ac.cnSigned-off-by:
-
Chen Ni authored
Replace a comma between expression statements by a semicolon. Reviewed-by:
-
Chen Ni authored
Replace a comma between expression statements by a semicolon. Reviewed-by:
-
Kajol Jain authored
Update JSON/events for power10 platform with additional events. Also move PM_VECTOR_LD_CMPL event from others.json to frontend.json file. Reviewed-by:
Kajol Jain <kjain@linux.ibm.com> Tested-by:
Disha Goel <disgoel@linux.ibm.com> Cc: Akanksha J N <akanksha@linux.ibm.com> Cc: Athira Rajeev <atrajeev@linux.vnet.ibm.com> Cc: Disha Goel <disgoel@linux.vnet.ibm.com> Cc: Madhavan Srinivasan <maddy@linux.ibm.com> Cc: Namhyung Kim <namhyung@kernel.org> Cc: hbathini@linux.ibm.com Cc: linuxppc-dev@lists.ozlabs.org Link: https://lore.kernel.org/r/20240723052154.96202-1-kjain@linux.ibm.com [ Remove alternative to ' char that made the build break in some distros with a unicode parsing python error ] Signed-off-by:
-
Athira Rajeev authored
Since the "ins.name" is not set while using raw instruction, 'perf annotate' with insn-stat gives wrong data: Result from "./perf annotate --data-type --insn-stat": Annotate Instruction stats total 615, ok 419 (68.1%), bad 196 (31.9%) Name : Good Bad ----------------------------------------------------------- : 419 196 This patch sets "dl->ins.name" in arch specific function "check_ppc_insn" while initialising "struct disasm_line". Also update "ins_find" function to pass "struct disasm_line" as a parameter so as to set its name field in arch specific call. With the patch changes: Annotate Instruction stats total 609, ok 446 (73.2%), bad 163 (26.8%) Name/opcode : Good Bad ----------------------------------------------------------- 58 : 323 80 32 : 49 43 34 : 33 11 OP_31_XOP_LDX : 8 20 40 : 23 0 OP_31_XOP_LWARX : 5 1 OP_31_XOP_LWZX : 2 3 OP_31_XOP_LDARX : 3 0 33 : 0 2 OP_31_XOP_LBZX : 0 1 OP_31_XOP_LWAX : 0 1 OP_31_XOP_LHZX : 0 1 Reviewed-by:Athira Rajeev <atrajeev@linux.vnet.ibm.com> Tested-by:
-
Athira Rajeev authored
Now perf uses the capstone library to disassemble the instructions in x86. capstone is used (if available) for perf annotate to speed up. Currently it only supports x86 architecture. This patch includes changes to enable this in powerpc. For now, only for data type sort keys, this method is used and only binary code (raw instruction) is read. This is because powerpc approach to understand instructions and reg fields uses raw instruction. The "cs_disasm" is currently not enabled. While attempting to do cs_disasm, observation is that some of the instructions were not identified (ex: extswsli, maddld) and it had to fallback to use objdump. Hence enabling "cs_disasm" is added in comment section as a TODO for powerpc. Reviewed-by:
-
Athira Rajeev authored
capstone_init is made availbale for all archs to use and updated to enable support for CS_ARCH_PPC as well. Patch removes open_capstone_handle and uses capstone_init in all the places. Committer notes: Avoid including capstone/capstone.h from print_insn.h to not break the build in builtin-script.c due to the namespace clash with libbpf: /usr/include/capstone/bpf.h:94:14: error: 'bpf_insn' defined as wrong kind of tag Reviewed-by:
-
Athira Rajeev authored
symbol__disassemble_capstone in util/disasm.c calls function open_capstone_handle to open/init the capstone. We already have a capstone_init function in "util/print_insn.c". But capstone_init is defined as a static function in util/print_insn.c. Change this and also add the function in print_insn.h The open_capstone_handle checks the disassembler_style option from annotation_options to decide whether to set CS_OPT_SYNTAX_ATT. Add that logic in capstone_init also and by default set it to true. Reviewed-by:
-
Athira Rajeev authored
Add instruction tracking function "update_insn_state_powerpc" for powerpc. Example sequence in powerpc: ld r10,264(r3) mr r31,r3 <<after some sequence> ld r9,312(r31) Consider ithe sample is pointing to: "ld r9,312(r31)". Here the memory reference is hit at "312(r31)" where 312 is the offset and r31 is the source register. Previous instruction sequence shows that register state of r3 is moved to r31. So to identify the data type for r31 access, the previous instruction ("mr") needs to be tracked and the state type entry has to be updated. Current instruction tracking support in perf tools infrastructure is specific to x86. Patch adds this support for powerpc as well. Reviewed-by: -
Athira Rajeev authored
Add few more instructions and use opcode as search key to find if it is supported by the architecture. The added ones are: addi, addic, addic., addis, subfic and mulli Reviewed-by:
-
Athira Rajeev authored
Data-type profiling has the concept of instruction tracking. Example sequence in powerpc: ld r10,264(r3) mr r31,r3 <<after some sequence> ld r9,312(r31) or differently lwz r10,264(r3) add r31, r3, RB lwz r9, 0(r31) If a sample is hit at "lwz r9, 0(r31)", data type of r31 depends on previous instruction sequence here. So to track the previous instructions, patch adds changes to identify some of the arithmetic instructions which are having opcode as 31. Since memory instructions also has cases with opcode 31, use the bits 22:30 to filter the arithmetic instructions here. Also there are instructions with just two operands like "addme", "addze". This patch adds new instructions ops "arithmetic_ops" to handle this Reviewed-by:
-
Athira Rajeev authored
There are memory instructions in powerpc with opcode as 31. Example: "ldx RT,RA,RB" , Its X form is as below: ______________________________________ | 31 | RT | RA | RB | 21 |/| -------------------------------------- 0 6 11 16 21 30 31 The opcode for "ldx" is 31. There are other instructions also with opcode 31 which are memory insn like ldux, stbx, lwzx, lhaux But all instructions with opcode 31 are not memory. Example is add instruction: "add RT,RA,RB" The value in bit 21-30 [ 21 for ldx ] is different for these instructions. Patch uses this value to assign instruction ops for these cases. The naming convention and value to identify these are picked from defines in "arch/powerpc/include/asm/ppc-opcode.h" Reviewed-by:
-
Athira Rajeev authored
Use the raw instruction code and macros to identify memory instructions, extract register fields and also offset. The implementation addresses the D-form, X-form, DS-form instructions. Two main functions are added. New parse function "load_store__parse" as instruction ops parser for memory instructions. Unlike other parsers (like mov__parse), this one fills in the "multi_regs" field for source/target and new added "mem_ref" field. No other fields are set because, here there is no need to parse the disassembled code and arch specific macros will take care of extracting offset and regs which is easier and will be precise. In powerpc, all instructions with a primary opcode from 32 to 63 are memory instructions. Update "ins__find" function to have "raw_insn" also as a parameter. Reviewed-by:
-
Athira Rajeev authored
Use the raw instruction code and macros to identify memory instructions, extract register fields and also offset. The implementation addresses the D-form, X-form, DS-form instructions. Adds "mem_ref" field to check whether source/target has memory reference. Add function "get_powerpc_regs" which will set these fields: reg1, reg2, offset depending of where it is source or target ops. Update "parse" callback for "struct ins_ops" to also pass "struct disasm_line" as argument. This is needed in parse functions where opcode is used to determine whether to set multi_regs and other fields Reviewed-by:
-
Athira Rajeev authored
perf annotate: Add support to capture and parse raw instruction in powerpc using dso__data_read_offset utility Add support to capture and parse raw instruction in powerpc. Currently, the perf tool infrastructure uses two ways to disassemble and understand the instruction. One is objdump and other option is via libcapstone. Currently, the perf tool infrastructure uses "--no-show-raw-insn" option with "objdump" while disassemble. Example from powerpc with this option for an instruction address is: Snippet from: objdump --start-address=<address> --stop-address=<address> -d --no-show-raw-insn -C <vmlinux> c0000000010224b4: lwz r10,0(r9) This line "lwz r10,0(r9)" is parsed to extract instruction name, registers names and offset. Also to find whether there is a memory reference in the operands, "memory_ref_char" field of objdump is used. For x86, "(" is used as memory_ref_char to tackle instructions of the form "mov (%rax), %rcx". In case of powerpc, not all instructions using "(" are the only memory instructions. Example, above instruction can also be of extended form (X form) "lwzx r10,0,r19". Inorder to easy identify the instruction category and extract the source/target registers, patch adds support to use raw instruction for powerpc. Approach used is to read the raw instruction directly from the DSO file using "dso__data_read_offset" utility which is already implemented in perf infrastructure in "util/dso.c". Example: 38 01 81 e8 ld r4,312(r1) Here "38 01 81 e8" is the raw instruction representation. In powerpc, this translates to instruction form: "ld RT,DS(RA)" and binary code as: | 58 | RT | RA | DS | | ------------------------------------- 0 6 11 16 30 31 Function "symbol__disassemble_dso" is updated to read raw instruction directly from DSO using dso__data_read_offset utility. In case of above example, this captures: line: 38 01 81 e8 The above works well when 'perf report' is invoked with only sort keys for data type ie type and typeoff. Because there is no instruction level annotation needed if only data type information is requested for. For annotating sample, along with type and typeoff sort key, "sym" sort key is also needed. And by default invoking just "perf report" uses sort key "sym" that displays the symbol information. With approach changes in powerpc which first reads DSO for raw instruction, "perf annotate" and "perf report" + a key breaks since it doesn't do the instruction level disassembly. Snippet of result from 'perf report': Samples: 1K of event 'mem-loads', 4000 Hz, Event count (approx.): 937238 do_work /usr/bin/pmlogger [Percent: local period] Percent│ ea230010 │ 3a550010 │ 3a600000 │ 38f60001 │ 39490008 │ 42400438 51.44 │ 81290008 │ 7d485378 Here, raw instruction is displayed in the output instead of human readable annotated form. One way to get the appropriate data is to specify "--objdump path", by which code annotation will be done. But the default behaviour will be changed. To fix this breakage, check if "sym" sort key is set. If so fallback and use the libcapstone/objdump way of disassmbling the sample. With the changes and "perf report" Samples: 1K of event 'mem-loads', 4000 Hz, Event count (approx.): 937238 do_work /usr/bin/pmlogger [Percent: local period] Percent│ ld r17,16(r3) │ addi r18,r21,16 │ li r19,0 │ 8b0: rldicl r10,r10,63,33 │ addi r10,r10,1 │ mtctr r10 │ ↓ b 8e4 │ 8c0: addi r7,r22,1 │ addi r10,r9,8 │ ↓ bdz d00 51.44 │ lwz r9,8(r9) │ mr r8,r10 │ cmpw r20,r9 Committer notes: Just add the extern for 'sort_order' in disasm.c so that we don't end up breaking the build due to this type colision with capstone and libbpf: In file included from /usr/include/capstone/capstone.h:325, from /git/perf-6.10.0/tools/perf/util/print_insn.h:23, from builtin-script.c:38: /usr/include/capstone/bpf.h:94:14: error: 'bpf_insn' defined as wrong kind of tag 94 | typedef enum bpf_insn { I reported this to the bpf mailing list, see one of the links below. Reviewed-by: -
Athira Rajeev authored
Currently, the perf tool infrastructure uses the disasm_line__parse function to parse disassembled line. Example snippet from objdump: objdump --start-address=<address> --stop-address=<address> -d --no-show-raw-insn -C <vmlinux> c0000000010224b4: lwz r10,0(r9) This line "lwz r10,0(r9)" is parsed to extract instruction name, registers names and offset. In powerpc, the approach for data type profiling uses raw instruction instead of result from objdump to identify the instruction category and extract the source/target registers. Example: 38 01 81 e8 ld r4,312(r1) Here "38 01 81 e8" is the raw instruction representation. Add function "disasm_line__parse_powerpc" to handle parsing of raw instruction. Also update "struct disasm_line" to save the binary code/ With the change, function captures: line -> "38 01 81 e8 ld r4,312(r1)" raw instruction "38 01 81 e8" Raw instruction is used later to extract the reg/offset fields. Macros are added to extract opcode and register fields. "struct disasm_line" is updated to carry union of "bytes" and "raw_insn" of 32 bit to carry raw code (raw). Function "disasm_line__parse_powerpc fills the raw instruction hex value and can use macros to get opcode. There is no changes in existing code paths, which parses the disassembled code. The size of raw instruction depends on architecture. In case of powerpc, the parsing the disasm line needs to handle cases for reading binary code directly from DSO as well as parsing the objdump result. Hence adding the logic into separate function instead of updating "disasm_line__parse". The architecture using the instruction name and present approach is not altered. Since this approach targets powerpc, the macro implementation is added for powerpc as of now. Since the disasm_line__parse is used in other cases (perf annotate) and not only data tye profiling, the powerpc callback includes changes to work with binary code as well as mnemonic representation. Also in case if the DSO read fails and libcapstone is not supported, the approach fallback to use objdump as option. Hence as option, patch has changes to ensure objdump option also works well. Reviewed-by:
-
