[pantsd] Address pantsd-runner hang on Linux and re-enable integration test. by kwlzn · Pull Request #4407 · pantsbuild/pants

kwlzn · 2017-03-31T02:08:23Z

Problem

Currently, on Linux the first thin client call to the daemon can deadlock just after the pantsd->fork->pantsd-runner workflow. Connecting to the process with gdb reveals a deadlock in the following stack in the post_fork drop of the CpuPool:

#0  0x00007f63f04c31bd in __lll_lock_wait () from /lib64/libpthread.so.0
No symbol table info available.
#1  0x00007f63f04c0ded in pthread_cond_signal@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
No symbol table info available.
#2  0x00007f63d3cfa438 in notify_one ()
    at /buildslave/rust-buildbot/slave/stable-dist-rustc-linux/build/src/libstd/sys/unix/condvar.rs:52
No locals.
#3  notify_one () at /buildslave/rust-buildbot/slave/stable-dist-rustc-linux/build/src/libstd/sys_common/condvar.rs:39
No locals.
#4  notify_one () at /buildslave/rust-buildbot/slave/stable-dist-rustc-linux/build/src/libstd/sync/condvar.rs:208
No locals.
#5  std::thread::{{impl}}::unpark () at /buildslave/rust-buildbot/slave/stable-dist-rustc-linux/build/src/libstd/thread/mod.rs:633
No locals.
#6  0x00007f63d3c583d1 in crossbeam::sync::ms_queue::{{impl}}::push<futures_cpupool::Message> (self=<optimized out>, t=...)
    at /home/kwilson/.cache/pants/rust-toolchain/registry/src/github.com-1ecc6299db9ec823/crossbeam-0.2.10/src/sync/ms_queue.rs:178
        guard = <optimized out>
        self = <optimized out>
#7  0x00007f63d3c588ed in futures_cpupool::{{impl}}::drop (self=<optimized out>)
    at /home/kwilson/.cache/pants/rust-toolchain/git/checkouts/futures-rs-a4f11d094efefb0a/f7e6bc8/futures-cpupool/src/lib.rs:236
        self = 0x37547a0
#8  0x00007f63d3be871c in engine::fs::{{impl}}::post_fork (self=0x3754778) at /home/kwilson/dev/pants/src/rust/engine/src/fs.rs:355
        self = 0x3754778
#9  0x00007f63d3be10e4 in engine::context::{{impl}}::post_fork (self=0x37545b0)
    at /home/kwilson/dev/pants/src/rust/engine/src/context.rs:93
        self = 0x37545b0
#10 0x00007f63d3c0de5a in {{closure}} (scheduler=<optimized out>) at /home/kwilson/dev/pants/src/rust/engine/src/lib.rs:275
        scheduler = 0x3740580
#11 with_scheduler<closure,()> (scheduler_ptr=<optimized out>, f=...) at /home/kwilson/dev/pants/src/rust/engine/src/lib.rs:584
        scheduler = 0x3740580
        scheduler_ptr = 0x3740580
#12 engine::scheduler_post_fork (scheduler_ptr=0x3740580) at /home/kwilson/dev/pants/src/rust/engine/src/lib.rs:274
        scheduler_ptr = 0x3740580
#13 0x00007f63d3c1be8c in _cffi_f_scheduler_post_fork (self=<optimized out>, arg0=0x35798f0) at src/cffi/native_engine.c:2234
        _save = 0x34a65a0
        x0 = 0x3740580
        datasize = <optimized out>
#14 0x00007f63f07b5a62 in PyEval_EvalFrameEx () from /lib64/libpython2.7.so.1.0

This presents as a hang in the thin client, because the pailgun socket is left open in the pantsd-runner.

Solution

Add pre-fork hooks and tear down the CpuPool instances prior to forking and rebuilding them.

Result

Can no longer reproduce the hang.

stuhood · 2017-03-31T02:12:17Z

src/rust/engine/src/graph.rs

        match self.node.clone() {
          EntryKey::Valid(n) => {
-            let pool = context_factory.pool();
+            let pool_opt = context_factory.pool();


One line for this is probably fine.

not according to the compiler.

error: borrowed value does not live long enough --> src/rust/engine/src/graph.rs:115:88 | 115 | let pool = context_factory.pool().as_ref().expect("Uninitialized CpuPool!"); | ---------------------- temporary value created here ^ temporary value dropped here while still borrowed ... 122 | }, | - temporary value needs to live until here | = note: consider using a `let` binding to increase its lifetime

Ah. Yea, makes sense. Thanks.

stuhood

Thanks!

stuhood · 2017-03-31T02:14:26Z

src/python/pants/bin/daemon_pants_runner.py

    """Fork, daemonize and invoke self.post_fork_child() (via ProcessManager)."""
    self.daemonize(write_pid=False)

+  def pre_fork(self):


IIRC, post_fork is called in two places... does this need to be as well?

…n test. (pantsbuild#4407) ### Problem Currently, on Linux the first thin client call to the daemon can deadlock just after the pantsd->fork->pantsd-runner workflow. Connecting to the process with `gdb` reveals a deadlock in the following stack in the `post_fork` `drop` of the `CpuPool`: ``` #0 0x00007f63f04c31bd in __lll_lock_wait () from /lib64/libpthread.so.0 No symbol table info available. pantsbuild#1 0x00007f63f04c0ded in pthread_cond_signal@@GLIBC_2.3.2 () from /lib64/libpthread.so.0 No symbol table info available. pantsbuild#2 0x00007f63d3cfa438 in notify_one () at /buildslave/rust-buildbot/slave/stable-dist-rustc-linux/build/src/libstd/sys/unix/condvar.rs:52 No locals. pantsbuild#3 notify_one () at /buildslave/rust-buildbot/slave/stable-dist-rustc-linux/build/src/libstd/sys_common/condvar.rs:39 No locals. pantsbuild#4 notify_one () at /buildslave/rust-buildbot/slave/stable-dist-rustc-linux/build/src/libstd/sync/condvar.rs:208 No locals. pantsbuild#5 std::thread::{{impl}}::unpark () at /buildslave/rust-buildbot/slave/stable-dist-rustc-linux/build/src/libstd/thread/mod.rs:633 No locals. pantsbuild#6 0x00007f63d3c583d1 in crossbeam::sync::ms_queue::{{impl}}::push<futures_cpupool::Message> (self=<optimized out>, t=...) at /home/kwilson/.cache/pants/rust-toolchain/registry/src/github.com-1ecc6299db9ec823/crossbeam-0.2.10/src/sync/ms_queue.rs:178 guard = <optimized out> self = <optimized out> pantsbuild#7 0x00007f63d3c588ed in futures_cpupool::{{impl}}::drop (self=<optimized out>) at /home/kwilson/.cache/pants/rust-toolchain/git/checkouts/futures-rs-a4f11d094efefb0a/f7e6bc8/futures-cpupool/src/lib.rs:236 self = 0x37547a0 pantsbuild#8 0x00007f63d3be871c in engine::fs::{{impl}}::post_fork (self=0x3754778) at /home/kwilson/dev/pants/src/rust/engine/src/fs.rs:355 self = 0x3754778 pantsbuild#9 0x00007f63d3be10e4 in engine::context::{{impl}}::post_fork (self=0x37545b0) at /home/kwilson/dev/pants/src/rust/engine/src/context.rs:93 self = 0x37545b0 pantsbuild#10 0x00007f63d3c0de5a in {{closure}} (scheduler=<optimized out>) at /home/kwilson/dev/pants/src/rust/engine/src/lib.rs:275 scheduler = 0x3740580 pantsbuild#11 with_scheduler<closure,()> (scheduler_ptr=<optimized out>, f=...) at /home/kwilson/dev/pants/src/rust/engine/src/lib.rs:584 scheduler = 0x3740580 scheduler_ptr = 0x3740580 pantsbuild#12 engine::scheduler_post_fork (scheduler_ptr=0x3740580) at /home/kwilson/dev/pants/src/rust/engine/src/lib.rs:274 scheduler_ptr = 0x3740580 pantsbuild#13 0x00007f63d3c1be8c in _cffi_f_scheduler_post_fork (self=<optimized out>, arg0=0x35798f0) at src/cffi/native_engine.c:2234 _save = 0x34a65a0 x0 = 0x3740580 datasize = <optimized out> pantsbuild#14 0x00007f63f07b5a62 in PyEval_EvalFrameEx () from /lib64/libpython2.7.so.1.0 ``` This presents as a hang in the thin client, because the pailgun socket is left open in the pantsd-runner. ### Solution Add pre-fork hooks and tear down the `CpuPool` instances prior to forking and rebuilding them. ### Result Can no longer reproduce the hang.

https://python-poetry.org/history/#118---2021-08-19 Fixed an error with repository prioritization when specifying secondary repositories. (pantsbuild#4241) Fixed the detection of the system environment when the setting virtualenvs.create is deactivated. (pantsbuild#4330, pantsbuild#4407) Fixed the evaluation of relative path dependencies. (pantsbuild#4246) Fixed environment detection for Python 3.10 environments. (pantsbuild#4387) Fixed an error in the evaluation of in/not in markers (python-poetry/poetry-core#189)

https://python-poetry.org/history/#118---2021-08-19 Fixed an error with repository prioritization when specifying secondary repositories. (#4241) Fixed the detection of the system environment when the setting virtualenvs.create is deactivated. (#4330, #4407) Fixed the evaluation of relative path dependencies. (#4246) Fixed environment detection for Python 3.10 environments. (#4387) Fixed an error in the evaluation of in/not in markers (python-poetry/poetry-core#189)

Uncomment disabled test.

3748d7a

stuhood reviewed Mar 31, 2017

View reviewed changes

stuhood approved these changes Mar 31, 2017

View reviewed changes

Add pre-fork handling for Core::pool.

65e5697

kwlzn force-pushed the kwlzn/linux_pantsd_hang branch from 207f8d6 to 65e5697 Compare March 31, 2017 02:25

kwlzn added 4 commits March 30, 2017 21:22

And the pre-fork handling for SchedulerService.

9936f36

And the fs.rs treatment.

61ea558

Condense.

22f61a1

Unwrap to expect.

78e2845

kwlzn merged commit 177deb9 into pantsbuild:master Mar 31, 2017

jsirois mentioned this pull request Apr 20, 2017

[testing] Unskip pantsd+watchman test. #4363

Closed

asherf mentioned this pull request Sep 11, 2021

Upgrade to poetry 1.1.8 #12856

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Comments

[pantsd] Address pantsd-runner hang on Linux and re-enable integration test.#4407

[pantsd] Address pantsd-runner hang on Linux and re-enable integration test.#4407
kwlzn merged 6 commits intopantsbuild:masterfrom
twitter:kwlzn/linux_pantsd_hang

kwlzn commented Mar 31, 2017 •

edited

Loading

Uh oh!

stuhood Mar 31, 2017

Uh oh!

kwlzn Mar 31, 2017 •

edited

Loading

Uh oh!

stuhood Mar 31, 2017

Uh oh!

stuhood left a comment

Uh oh!

stuhood Mar 31, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Comments

Conversation

kwlzn commented Mar 31, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Solution

Result

Uh oh!

stuhood Mar 31, 2017

Choose a reason for hiding this comment

Uh oh!

kwlzn Mar 31, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

stuhood Mar 31, 2017

Choose a reason for hiding this comment

Uh oh!

stuhood left a comment

Choose a reason for hiding this comment

Uh oh!

stuhood Mar 31, 2017

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

kwlzn commented Mar 31, 2017 •

edited

Loading

kwlzn Mar 31, 2017 •

edited

Loading