Skip to content

add perfmap generation - #8163

Open
csjh wants to merge 2 commits into
mainfrom
cormick/add-perfmap-generation
Open

csjh wants to merge 2 commits into
mainfrom
cormick/add-perfmap-generation

Conversation

@csjh

@csjh csjh commented Nov 2, 2026

Copy link
Copy Markdown
Contributor

enhances our existing code gen event tracking to record a corresponding perfmap file

not merging there's some alignment with the perf team on how to integrate into continuous profiling, but approach will be the same either way (I think only concern is cardinality, so only change would be changing name writing in PerfMapWriter::add)

@csjh
csjh requested review from a team as code owners November 2, 2026 21:24

@clanky-the-clanker clanky-the-clanker Bot left a comment •

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded by the review of b2fff3b3.

This review, of bffe36da

🛑 Significant concerns — One BLOCKING issue: perf-map.c++ is built into the jsg library on all platforms but uses <unistd.h>, fcntl(F_GETFL) and write() without a Windows guard, which should break the Windows build and release. The Linux-only test is a WARNING.

Findings: 1 blocking · 1 warning · 1 info · 1 suggestion (4 are inline comments on the diff)

The decision is advisory: this review is a comment and never approves, blocks or requests changes on the pull request.

🔧 Fix with your agent

Paste the following into your coding agent to investigate and fix these issues.

Fix the following code review issues found in pull request #8163.

Before making changes, review each issue and present a brief summary table:
- For each finding, state whether you agree, disagree, or need clarification
- If you disagree (e.g., the fix would require disproportionate changes
  for minimal ROI, or the finding is factually incorrect), explain why
- If you need to clarify something before deciding, ask those questions
- Then share your plan for which issues to tackle and in what order

## BLOCKING (1)

### [Code Quality] perf-map.c++ is compiled into the jsg library on all platforms but uses <unistd.h>, fcntl(F_GETFL), O_APPEND and write() with no _WIN32 guard, so it should break the Windows build (.github/workflows/test.yml:48, release.yml:102-105).
- **File:** src/workerd/jsg/perf-map.c++:9

## WARNING (1)

### [Code Quality] perf-map-test.c++ uses the Linux-only memfd_create/MFD_CLOEXEC with no platform guard, but the *-test.c++ glob builds it on macOS and Windows too, so those CI test builds should fail.
- **File:** src/workerd/jsg/perf-map-test.c++:17

## INFO (1)

### [Performance] Each JIT code event results in its own synchronous write() syscall on the isolate thread.
- **File:** src/workerd/jsg/perf-map.c++:38

## SUGGESTION (1)

### [Code Quality] The KJ_REQUIRE line in V8System::enablePerfMap is over the 100-column limit and is not clang-formatted.
- **File:** src/workerd/jsg/setup.c++:227

📋 Code Quality: 🛑 1 blocking · 1 warning · 1 suggestion

The new perf-map source isn't guarded for Windows, the test uses a Linux-only API, and one line is not formatted.

Recommendation: 🔴 Serious Concerns

  • [BLOCKING] perf-map.c++ is compiled into the jsg library on all platforms but uses <unistd.h>, fcntl(F_GETFL), O_APPEND and write() with no _WIN32 guard, so it should break the Windows build (.github/workflows/test.yml:48, release.yml:102-105). (src/workerd/jsg/perf-map.c++:9)
  • [WARNING] perf-map-test.c++ uses the Linux-only memfd_create/MFD_CLOEXEC with no platform guard, but the *-test.c++ glob builds it on macOS and Windows too, so those CI test builds should fail. (src/workerd/jsg/perf-map-test.c++:17)
  • [SUGGESTION] The KJ_REQUIRE line in V8System::enablePerfMap is over the 100-column limit and is not clang-formatted. (src/workerd/jsg/setup.c++:227)
View full review
  • Windows build (src/workerd/jsg/perf-map.c++:7-9): perf-map.c++ is added to the main jsg wd_cc_library with no platform restriction. It includes <unistd.h> and calls fcntl(F_GETFL), O_APPEND, write() and ssize_t, none of which the Windows toolchain provides. Other files in the tree put <unistd.h> behind #if !_WIN32, for example server/factory/bootstrap.c++:37 and io/worker.c++:74. Windows is built and tested in .github/workflows/test.yml:48 and released in .github/workflows/release.yml:102-105. setup.h now includes perf-map.h, so the header must also stay portable.
  • Linux-only test (src/workerd/jsg/perf-map-test.c++:9,17,58): the *-test.c++ glob builds this test on every platform, but it uses memfd_create, MFD_CLOEXEC and <sys/mman.h>. Guard it with #ifdef __linux__ the way fast-api-test.c++ does, or exclude the file from the glob on non-Linux targets.
  • Formatting (src/workerd/jsg/setup.c++:227): the KJ_REQUIRE line is over 100 columns and is not clang-formatted.

⚡ Performance: 💬 1 info

Each JIT code event costs one synchronous write() on the isolate thread.

Recommendation: 💬 Optional Nits

  • [INFO] Each JIT code event results in its own synchronous write() syscall on the isolate thread. (src/workerd/jsg/perf-map.c++:38)
View full review

PerfMapWriter::add makes one unbuffered write() per ADD_CODE_LINE_POS and CODE_MOVED event, inline in the isolate's V8 callback. The feature is opt-in, so the cost is acceptable, but keep it in mind while the cardinality question is still open.


✅ All other checks passed

🔒 Security: ✅ LGTM

The writer receives a descriptor that is already open; this pull request does not open any path itself. Newlines and carriage returns in JS-derived names are replaced, so a script cannot inject extra map records. The line buffer is bounded at 1024 bytes.


📖 AGENTS.md: ✅ LGTM

The change follows existing JSG and Bazel patterns.


📝 Documentation: ✅ LGTM

No documentation issues.


Reviewed commit: bffe36da

Comment thread src/workerd/jsg/perf-map.c++ Resolved
Comment thread src/workerd/jsg/perf-map-test.c++ Resolved

ssize_t n;
do {
n = write(fd.get(), line, pos - line);

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ℹ️ INFO · Performance

Each JIT code event results in its own synchronous write() syscall on the isolate thread.

This runs inline in IsolateBase::jitCodeEvent on the isolate thread, one syscall for every code event. A cold start can produce thousands of these, and a slow or full filesystem would stall the isolate. The feature is opt-in, so this is acceptable, but if you change how names are written to address cardinality, consider whether per-event writes are still the right choice.

Comment thread src/workerd/jsg/setup.c++ Resolved

@clanky-the-clanker clanky-the-clanker Bot left a comment •

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded by the review of d6cb3401.

This review, of b2fff3b3

💬 Approved with comments — The earlier Windows build BLOCKING and the Linux-only test WARNING are both fixed. The only remaining note is an INFO about one write() call per JIT event.

Findings: 1 info

The decision is advisory: this review is a comment and never approves, blocks or requests changes on the pull request.


⚡ Performance: 💬 1 info

One write() syscall per JIT code event on the isolate thread

Recommendation: 💬 Optional Nits

  • [INFO] Each ADD_CODE_LINE_POS and CODE_MOVED event makes its own synchronous write() syscall on the isolate thread. (src/workerd/jsg/perf-map.c++:49)
View full review

PerfMapWriter::add runs inline from V8's JIT code event callback. Each call does one unbuffered write(), and warm-up or GC compaction can produce many of these events. The feature is opt-in, so this is acceptable. Revisit it if the cardinality question turns out to mean high event volume.


✅ All other checks passed

🔒 Security: ✅ LGTM

No security issues. The writer takes a descriptor that is already open, and it replaces newlines in names so a script cannot inject extra map records. The line buffer is bounded at 1024 bytes.


📖 AGENTS.md: ✅ LGTM

The change follows the existing JSG and Bazel patterns.


📋 Code Quality: ✅ LGTM

All three earlier findings are fixed:

  • Windows build: perf-map.c++:9-25 puts the POSIX code behind #if !_WIN32.
  • Linux-only test: the test uses mkstemp + unlink, and BUILD.bazel:373-381 builds it as a separate kj_test that is incompatible with Windows.
  • Line length: the KJ_REQUIRE line in setup.c++ is now wrapped.

📝 Documentation: ✅ LGTM

No documentation issues.


Reviewed commit: b2fff3b3

csjh and others added 2 commits November 16, 2026 19:36
perf-map.c++ used <unistd.h>, fcntl() and write() unguarded, which
breaks the Windows build. On Windows the writer now throws
KJ_UNIMPLEMENTED, as getJsStackTrace() already does.

The test used memfd_create(), which is Linux-only, so it also failed
to build on macOS. It now uses an unlinked mkstemp() file, and is
marked incompatible with Windows.

Also explain in perf-map.h why each line is its own write(), and
wrap an over-long KJ_REQUIRE.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@csjh
csjh force-pushed the cormick/add-perfmap-generation branch from b2fff3b to d6cb340 Compare November 17, 2026 00:45

@clanky-the-clanker clanky-the-clanker Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ Approved — No findings at this commit. The earlier Windows-build and Linux-only test problems are still fixed, and the setup.h and BUILD.bazel changes are correct.

The decision is advisory: this review is a comment and never approves, blocks or requests changes on the pull request.


✅ All checks passed

🔒 Security: ✅ LGTM

The writer takes a descriptor that is already open, so no path comes from untrusted input. It replaces newlines in function names, so a script cannot inject extra map records. The line buffer is bounded at 1024 bytes.


📖 AGENTS.md: ✅ LGTM

The change follows the existing JSG and Bazel conventions.


📋 Code Quality: ✅ LGTM

No new problems in the code changed since the last review. setup.c++ skips bytecode events in both the CODE_ADDED and CODE_MOVED paths. perf-map.c++ keeps its POSIX code behind #if !_WIN32. The test target is marked incompatible with Windows. enablePerfMap documents that it requires JitCodeEventTracking::YES and must be called before any isolate is created.


⚡ Performance: ✅ LGTM

No new performance concerns. The earlier note still applies: each JIT code event makes one synchronous write() call. The feature is opt-in, so this is acceptable for now.


📝 Documentation: ✅ LGTM

No documentation issues.


Reviewed commit: d6cb3401

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant