BusyBox symbol profile · librelane-playground

P84 told us the shell workload was slow, but all low userspace PCs were raw addresses. P85 resolves the command-window samples against BusyBox’s unstripped ELF, so the profile can name the code the shell is actually running.

The filtered window is:

marker	cycle
`shell_start`	159,733,543
`file_ok`	239,533,859

That leaves boot, initramfs unpacking, and early console setup out of the BusyBox profile.

Hot Functions

hot functions label P85 BusyBox shell symbols samples 77,930 period every 1,024 cycles

printf_core busybox

4.8% of samples (3,718 samples)

4.8% 3,718
memset kernel

4.3% of samples (3,347 samples)

4.3% 3,347
vruntime_eligible kernel

3.8% of samples (2,954 samples)

3.8% 2,954
memcpy busybox

3.1% of samples (2,404 samples)

3.1% 2,404
n_tty_write kernel

2.6% of samples (2,028 samples)

2.6% 2,028
memcpy kernel

2.4% of samples (1,837 samples)

2.4% 1,837
blake2s_compress_generic kernel

2.3% of samples (1,809 samples)

2.3% 1,809
__fwritex busybox

2.3% of samples (1,763 samples)

2.3% 1,763
avg_vruntime kernel

1.8% of samples (1,436 samples)

1.8% 1,436
handle_exception kernel

1.5% of samples (1,160 samples)

1.5% 1,160
unmap_page_range kernel

1.5% of samples (1,154 samples)

1.5% 1,154
update_curr kernel

1.2% of samples (912 samples)

1.2% 912
memset busybox

1.1% of samples (845 samples)

1.1% 845
__enqueue_entity.constprop.0 kernel

1.1% of samples (820 samples)

1.1% 820
ret_from_exception kernel

1% of samples (789 samples)

1% 789
(remaining) remaining

56.9% of samples (44,330 samples)

56.9% 44,330

The top entries are no longer mysterious raw addresses:

bucket	function	pct
BusyBox	`printf_core`	4.77%
kernel	`memset`	4.29%
kernel	`vruntime_eligible`	3.79%
BusyBox	`memcpy`	3.08%
kernel	`n_tty_write`	2.60%
BusyBox	`__fwritex`	2.26%

This is a shell and terminal workload, not an integer benchmark. The shape says that output formatting, stdio writes, terminal writes, exception entry/return, and scheduler accounting are visible costs.

What This Means For CPU Work

The next CPU-side perf rounds should be judged against P84/P85, not just against userspace hello:

Larger or split TLB: P84 saw 2.26M fetch walks and 2.94M load/store walks.
Better memory path: P84 saw 39.6M memory handshakes and 91.8M memory stall cycles.
Console batching: printf_core, __fwritex, and n_tty_write are now plainly hot.
Trap/syscall cleanup: handle_exception, ret_from_exception, and do_trap_ecall_u are visible in the shell window.

Honest Status

check	status
P84 shell sample data reused	PASS
BusyBox unstripped ELF found and symbolized	PASS
Shell-window cycle filter used	PASS
Site hot-function JSON staged	PASS
Folded flamegraph input staged	PASS
True stack unwinding	NOT RUN
Per-PID userspace symbolization	PARTIAL
LibreLane hardening	NOT RUN

The remaining caveat is process identity. The sample stream has PC and cycle, not PID or current address-space identity. Filtering to the shell window makes BusyBox the right practical interpretation, but the harness does not yet prove it per sample.