perf

perf profiling

Analyze CPU profiling captures to identify hot paths, measure overhead, and guide optimization. Operate on perf.data files or pre-generated perf report/script output.

perf.data is a binary capture and perf report / perf script output streams through pipelines rather than landing in a normal text file. awk, grep, head, and similar shell tools are the correct way to extract fields and summaries from this data; any general preference for dedicated file-reading tools does not apply here.

Retrieving shared capture bundles

Testers often share a capture bundle as an anonymous OneDrive 1drv.ms link, which serves a single-page-app shell rather than the files. See references/remote-bundles.md for the SharePoint personal-content retrieval path: a badger token, the base64url-encoded share id, and the item-id navigation that the legacy OneDrive API no longer supports for migrated personal accounts.

Key defaults

Standard recording invocation

sudo perf record -e cycles -e cpu-clock \
  --call-graph fp -F99 <command>

Option	Rationale
-e cycles -e cpu-clock	cpu-clock provides a wall-clock baseline for measuring true CPU idle time (see Pitfalls)
--call-graph fp	Kernel is built with frame pointers (CONFIG_FRAME_POINTER=y)
-F99	99 Hz sampling avoids lock-step aliasing with timer interrupts

Add -p <pid> for a running process or -a for system-wide.

Symbol resolution for modules

When the kernel is built with -ffunction-sections (check KBUILD_CFLAGS_KERNEL in the Makefile), the module loader rearranges function sections at load time. perf cannot resolve module symbols from .ko files in this case because the runtime layout no longer matches the ELF layout. Use full /proc/kallsyms:

sudo perf report --kallsyms=/proc/kallsyms

This collapses all kernel symbols into [kernel.kallsyms], losing per-module DSO attribution, but gives correct symbol names. If the build configuration is uncertain, prefer full /proc/kallsyms.