Messages in this thread Patch in this message | | | Subject | Re: [PATCH v2 1/2] tools/perf: Add bitfield_swap to handle branch_stack endian issue | From | Madhavan Srinivasan <> | Date | Mon, 25 Oct 2021 17:02:47 +0530 |
| |
On 10/21/21 6:55 PM, Jiri Olsa wrote: > On Sat, Oct 16, 2021 at 06:20:58PM +0530, Madhavan Srinivasan wrote: >> branch_stack struct has bit field definition which >> produces different bit ordering for big/little endian. >> Because of this, when branch_stack sample is collected >> in a BE system and viewed/reported in a LE system, bit >> fields of the branch stack are not presented properly. >> To address this issue, a evsel__bitfield_swap_branch_stack() >> is defined and introduced in evsel__parse_sample. >> >> Signed-off-by: Madhavan Srinivasan <maddy@linux.ibm.com> >> --- >> Changelog v1: >> - Renamed function and macro >> - Added comments in code >> >> tools/perf/util/evsel.c | 74 +++++++++++++++++++++++++++++++++++++++-- >> tools/perf/util/evsel.h | 13 ++++++++ >> 2 files changed, 85 insertions(+), 2 deletions(-) >> >> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c >> index dbfeceb2546c..746e642d4d32 100644 >> --- a/tools/perf/util/evsel.c >> +++ b/tools/perf/util/evsel.c >> @@ -2221,6 +2221,51 @@ void __weak arch_perf_parse_sample_weight(struct perf_sample *data, >> data->weight = *array; >> } >> >> +u64 evsel__bitfield_swap_branch_flags(u64 value) >> +{ >> + u64 new_val = 0; >> + >> + /* >> + * branch_flags >> + * union { >> + * u64 values; >> + * struct { >> + * mispred:1 //target mispredicted >> + * predicted:1 //target predicted >> + * in_tx:1 //in transaction >> + * abort:1 //transaction abort >> + * cycles:16 //cycle count to last branch >> + * type:4 //branch type >> + * reserved:40 >> + * } >> + * } >> + * >> + * Avoid bswap64() the entire branch_flag.value, >> + * as it has variable bit-field sizes. Instead the >> + * macro takes the bit-field position/size, >> + * swaps it based on the host endianness. >> + */ >> + if (bigendian()) { >> + new_val = bitfield_swap(value, 0, 1); >> + new_val |= bitfield_swap(value, 1, 1); >> + new_val |= bitfield_swap(value, 2, 1); >> + new_val |= bitfield_swap(value, 3, 1); >> + new_val |= bitfield_swap(value, 4, 16); >> + new_val |= bitfield_swap(value, 20, 4); >> + new_val |= bitfield_swap(value, 24, 40); >> + } else { >> + new_val = bitfield_swap(value, 63, 1); >> + new_val |= bitfield_swap(value, 62, 1); >> + new_val |= bitfield_swap(value, 61, 1); >> + new_val |= bitfield_swap(value, 60, 1); >> + new_val |= bitfield_swap(value, 44, 16); >> + new_val |= bitfield_swap(value, 40, 4); >> + new_val |= bitfield_swap(value, 0, 40); >> + } >> + >> + return new_val; >> +} >> + >> int evsel__parse_sample(struct evsel *evsel, union perf_event *event, >> struct perf_sample *data) >> { >> @@ -2408,6 +2453,8 @@ int evsel__parse_sample(struct evsel *evsel, union perf_event *event, >> if (type & PERF_SAMPLE_BRANCH_STACK) { >> const u64 max_branch_nr = UINT64_MAX / >> sizeof(struct branch_entry); >> + struct branch_entry *e; >> + unsigned int i; >> >> OVERFLOW_CHECK_u64(array); >> data->branch_stack = (struct branch_stack *)array++; >> @@ -2416,10 +2463,33 @@ int evsel__parse_sample(struct evsel *evsel, union perf_event *event, >> return -EFAULT; >> >> sz = data->branch_stack->nr * sizeof(struct branch_entry); >> - if (evsel__has_branch_hw_idx(evsel)) >> + if (evsel__has_branch_hw_idx(evsel)) { >> sz += sizeof(u64); >> - else >> + e = &data->branch_stack->entries[0]; >> + } else { >> data->no_hw_idx = true; >> + /* >> + * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied, >> + * only nr and entries[] will be output by kernel. >> + */ >> + e = (struct branch_entry *)&data->branch_stack->hw_idx; >> + } >> + >> + if (swapped) { > thanks for all the comments, make it easy to review > >> + /* >> + * struct branch_flag does not have endian >> + * specific bit field definition. And bswap >> + * will not resolve the issue, since these >> + * are bit fields. >> + * >> + * evsel__bitfield_swap_branch_flags() uses a >> + * bitfield_swap macro to swap the bit position >> + * based on the host endians. >> + */ >> + for (i = 0; i < data->branch_stack->nr; i++, e++) >> + e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value); > one last thing that confuses me.. the sample is already swapped at this > point, right? with perf_event__all64_swap function > > should you swap it back first and then do the proper bitfield swap? > like we do in PERF_SAMPLE_RAW to get proper u32 values
No, in case of PERF_SAMPLE_RAW type, first word contains the size.
Let me explain with an example (hope will not confuse further :) ), say we have branch_flag.value as 0x4000000000000000 generated in a BE system, which has predict set, now if we undo the swap done by perf_event__all64_swap, value comes out as 0x40 (in a LE system) and if this value fed to bitswap macro, we end up 0x0000000040000000, which is wrong. Instead, by not swapping back, we feed the value in the source-endian to the new macro.
Maddy
> thanks, > jirka > >> + } >> + >> OVERFLOW_CHECK(array, sz, max_size); >> array = (void *)array + sz; >> } >> diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h >> index 1f7edfa8568a..2e82cdbe2c08 100644 >> --- a/tools/perf/util/evsel.h >> +++ b/tools/perf/util/evsel.h >> @@ -482,4 +482,17 @@ struct evsel *evsel__leader(struct evsel *evsel); >> bool evsel__has_leader(struct evsel *evsel, struct evsel *leader); >> bool evsel__is_leader(struct evsel *evsel); >> void evsel__set_leader(struct evsel *evsel, struct evsel *leader); >> + >> +/* >> + * Macro to swap the bit-field postition and size. >> + * Used when, >> + * - dont need to swap the entire u64 && >> + * - when u64 has variable bit-field sizes && >> + * - when presented in a host endian which is different >> + * than the source endian of the perf.data file >> + */ >> +#define bitfield_swap(src, pos, size) \ >> + ((((src) >> (pos)) & ((1ull << (size)) - 1)) << (63 - ((pos) + (size) - 1))) >> + >> +u64 evsel__bitfield_swap_branch_flags(u64 value); >> #endif /* __PERF_EVSEL_H */ >> -- >> 2.31.1 >>
| |