Add facilities to produce pprof profiles

[javierhonduco]

This (rather large) commit adds pprof profile generation to lightswitch. They can be dumped to disk and the read with pprof tooling (see test plan) as well as sent to a remote server with a new collector. By default profiles are written to dis as flamegraphs in svg format.

Test Plan / Demo

build with: nix develop --command 'cargo build --release'

$ sudo target/release/lightswitch --duration=5

produces a flamegraph in svg format

$ sudo target/release/lightswitch --duration=5 --profile-format=pprof

produces a pprof file (uncompressed) that can be read with standard tooling

$ go tool pprof profile.pb
For tag pid used unit task-tgid, also encountered unit(s) task-id
File: ruby
Build ID: gnu-bae4d9744266cf1f3e27a1ba7512c1c97b5e4a86
lightswitch
Type: cpu
Time: Jul 23, 2024 at 10:03am (BST)
Duration: 5s, Total samples = 1.70s (34.07%)
Entering interactive mode (type "help" for commands, "o" for options)
(pprof) top
For tag pid used unit task-tgid, also encountered unit(s) task-id
Showing nodes accounting for 703.70ms, 41.30% of 1703.70ms total
Showing top 10 nodes out of 140
      flat  flat%   sum%        cum   cum%
  111.11ms  6.52%  6.52%   148.15ms  8.70%  BSD_vfprintf
  111.11ms  6.52% 13.04%   259.26ms 15.22%  lightswitch::unwind_info::UnwindInfoBuilder::process
  111.11ms  6.52% 19.57%   222.22ms 13.04%  n_tty_inherit_ops
   74.07ms  4.35% 23.91%   111.11ms  6.52%  __memcpy_avx_unaligned_erms
   74.07ms  4.35% 28.26%   407.41ms 23.91%  do_syscall_64
   74.07ms  4.35% 32.61%    74.07ms  4.35%  queue_work_on
   37.04ms  2.17% 34.78%    37.04ms  2.17%  __GI___pthread_enable_asynccancel
   37.04ms  2.17% 36.96%   444.44ms 26.09%  __GI___writev
   37.04ms  2.17% 39.13%    37.04ms  2.17%  __malloc_usable_size
   37.04ms  2.17% 41.30%    37.04ms  2.17%  __probestub_page_fault_kernel

$ sudo target/release/lightswitch --duration=5 --sender=remote

Performs an HTTP POST with the pprof as the body to a hardcoded server URL (this will change in the future).

[javierhonduco]

There must be a better way to do this... 🤔