How to ensure perf buffer memory alignment?

[yzhao1012]

Hi All,

We have a struct used in C code, some thing like:
struct event {
u64 data;
} attribute((packed, aligned(8));

BPF_PERF_OUTPUT(events);
events.perf_submit(..., sizeof(event));

In the user space C++ code, we open the perf buffer with 8 pages:
BPF bpf;
bpf.open_perf_buffer(events, output_fn, loss_fn, cookie, 8);

When we read the data from perf buffer, we sometimes get unaligned pointers.

Because the submitted memory is 8 bytes, that means that the perf buffer's own memory is not aligned on 8 bytes.

Is there anyway we can ensure perf buffer's own memory alignment?

[yzhao1012]

Note:
IIRC, open_perf_buffer() eventually mmap() the pages, and that memory region is always aligned:
https://stackoverflow.com/questions/42259495/does-mmap-return-aligned-pointer-values

If that's the case, then submitting aligned memory chunk should always result into read at aligned address, right?

But we do saw unaligned pointer. :(

[palmtenor]

A little bit more details: actual struct got written to the Perf Buffer is a PERF_RECORD_SAMPLE (see perf_event_open.2 man page). In BCC case we don't set any of the PERF_SAMPLE_* bits, so it would be just the 8-byte struct perf_event_header followed by size and the raw data. The man page says:

    size, data[size]
        If PERF_SAMPLE_RAW is enabled, then a 32-bit value
        indicating size is included followed by an array of
        8-bit values of length size.  The values are padded
        with 0 to have 64-bit alignment.

[yonghong-song]

From @palmtenor explanation in the above, the data alignment in the memory will be 4, but not 8. Maybe you want to design your struct event to be aligned at 4 (not 8) and your interested fields aligned at 8.

[yzhao1012]

Thanks All!

1 quick clarification:
"size, data[size]" "The values are padded with 0 to have 64-bit alignment"
This sounds like #1 perf buffer's memory region is 64bit/8bytes aligned, but when read, the data pointer points 'data', which is 4 bytes after size, therefore it becomes 4 bytes aligned.