Re: [PATCH v2 3/8] perf/core: Implement BUILD_ID_OFFSET sample type

[sashiko-bot@kernel.org]

> Implement PERF_SAMPLE_BUILD_ID_OFFSET and
> PERF_SAMPLE_CALLCHAIN_BUILD_ID_OFFSET payload writing in the perf event
> core by reusing BPF's stack_map_get_build_id_offset().
> 
> When attr.defer_callchain is set, user-space IPs and callchains request
> a deferred unwind via unwind_deferred_request() and emit a
> PERF_CONTEXT_USER_DEFERRED marker and cookie in the sample. In the
> faultable task_work callback (perf_unwind_deferred_callback), the
> unwound user IPs are resolved to build IDs and file offsets using
> stack_map_get_build_id_offset(..., may_fault=true) and emitted in
> PERF_RECORD_CALLCHAIN_DEFERRED with PERF_RECORD_MISC_MMAP_BUILD_ID set.
> 
> When attr.defer_callchain is not set (or if deferred unwinding is
> unavailable), user-space IPs and callchains are resolved in-place during
> perf_output_sample() using stack_map_get_build_id_offset(...,
> may_fault=false). If the mmap_lock cannot be trylocked or the ELF
> header/note pages are not present in the page cache (nor mlocked into
> memory), resolution falls back to emitting a zero-sized build ID
> (size = 0) with the virtual address as the offset. Using the deferred
> approach (attr.defer_callchain and attr.defer_output) is therefore
> preferred as it can fault in pages and wait on locks, resolving more
> build IDs without requiring ELF header pages to be mlocked in the page
> cache.
> 
> Kernel-space IPs and callchain entries are emitted with a zero-sized
> build ID (size = 0) and the virtual address as the offset.
> 
> Signed-off-by: Ian Rogers <irogers@google.com>

Sashiko has reviewed this patch and found no issues. It looks great!

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261002173848.3228217-1-irogers@google.com?part=3