Re: [PATCH v2 1/2] tools/perf: Add bitfield_swap to handle branch_stack endian issue
From: Madhavan Srinivasan
Date: Mon Oct 25 2021 - 07:33:08 EST
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@xxxxxxxxxxxxx>
---
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