[RFC] tracing: aggregate ring-buffer memory statistics

[Xiang Gao <gxxa03070307@gmail.com>]

Hi,

This is a resend of the RFC for exposing tracing buffer memory
statistics under tracefs.

The motivation is Android lost-RAM attribution: tracing buffer memory
is not represented as a separate category, so it can be counted as lost
RAM. The memory can be spread across the global trace array, dynamically
created instances, and snapshot buffers, and userspace currently has to
discover and sum every instance.

The patch adds:

  /sys/kernel/tracing/trace_stats/total_memory_kb

reporting:

  total:
  buffers_mem:
  snapshot_buffers_mem:

The values cover the global trace array and all instances across all
CPUs. A per-CPU view is provided under:

  /sys/kernel/tracing/trace_stats/per_cpu/cpuN/total_memory_kb

These *_mem values report the full pages backing the data sub-buffers
and reader page, unlike buffer_size_kb / buffer_total_size_kb which
report usable data capacity. The new trace_stats directory is distinct
from the existing trace_stat directory used for function and branch
counters.

For example, with 4 CPUs and 2 sub-buffers per CPU on a 4 KiB page
system:

  buffer_size_kb        = (4096 - 16) * 2 / 1024 = 7 kB     (one CPU)
  buffer_total_size_kb  = 7 * 4 = 28 kB                    (one array)
  buffers_mem           = (2 + 1) * 4096 * 4 / 1024 = 48 kB

The extra memory in buffers_mem comes from the per-page header and the
reader page.

Slab-allocated ring-buffer metadata is not included.

One open question is whether reporting page memory only is acceptable
for the initial version.

Thanks,
Xiang