1291954 - Make SM(tsan) a tier 1 build

Status: VERIFIED FIXED

(Core :: JavaScript Engine, defect, P3)

Description

[Steve Fink [:sfink] [:s:]]

See also 1288596.

Comment 1

[Steve Fink [:sfink] [:s:]]

Attachment: Report added tsan failures

I don't even remember what I was waiting on to land tsan as a tier 1 job, but I know I felt it was at least somewhat important to get somewhat decent reporting of added failures.

This commit also bumps up the number of expected failures from 14 to 20 unique locations, which is not great.

Comment 2

[Steve Fink [:sfink] [:s:]]

Attachment: Turn tsan into a canary

Ah! *This* is the important piece. Without this, a tsan error will make the test fail, which will turn the job red even if it is one of the expected errors.

Comment 3

[Steve Fink [:sfink] [:s:]]

Two remaining problems.

A bunch of tests are timing out when run under tsan. Unfortunately, they're jit-tests, which means we don't have a nice mechanism for logging all of the tests that take longer than n seconds so we can skip them. Hopefully the log will have enough info, or I can add in that option there too. (It's really annoying having 2 very similar test harnesses.)

A bunch of new races have been added, probably from the new thread work bhackett has been doing. See bug 1342473. Currently, it looks like it's probably just that the JS test shell is using things in a way not yet supported. (The setup is for cooperative threads, and it looks like shell workers run concurrently on the same runtime.)

Comment 4

[Julian Seward [:jseward]]

Comment on attachment 8840728 [details] [diff] [review]
Report added tsan failures

Review of attachment 8840728 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/devtools/automation/variants/tsan
@@ +7,5 @@
>          "JITTEST_EXTRA_ARGS": "--jitflags=debug --ignore-timeouts={DIR}/cgc-jittest-timeouts.txt",
>          "JSTESTS_EXTRA_ARGS": "--exclude-file={DIR}/cgc-jstests-slow.txt",
>          "TSAN_OPTIONS": "log_path={OUTDIR}/sanitize_log"
>      },
> +    "expect-errors": [

It would be nice to add a comment here explaining why some
of these entries are duplicated.  If not here, maybe in the
Python part?

Comment 5

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/dccd52fb0a8d
Turn tsan into a canary, r=jseward
https://hg.mozilla.org/integration/mozilla-inbound/rev/5806128b981e
Report added tsan failures, r=jseward

Comment 6

[Carsten Book [:Tomcat]]

https://hg.mozilla.org/mozilla-central/rev/dccd52fb0a8d
https://hg.mozilla.org/mozilla-central/rev/5806128b981e

Comment 7

[Steve Fink [:sfink] [:s:]]

Attachment: Avoid problems with timeouts when running under tsan

Message in the commit message: Several jit-tests expect to timeout, and are annotated with an expected status code. Currently, we have to force tsan to report a zero status if it finds an error, since otherwise it will cause lots of tests to fail (due to hitting a tsan-detectable problem.) But those zero exit statuses cause the test to fail. Add --unusable-error-status to treat those as passing.

Now, it turns out that this was *not* the reason for those tests to be failing. They didn't have any tsan errors, so did not trigger the forced-exitstatus setting. But this patch still seems worthwhile if we *do* run into that situation (with any expected error test, not just timeouts.)

Comment 8

[Steve Fink [:sfink] [:s:]]

Attachment: Skip wasm/asm.js for tsan test

After the next patch I'm going to post, the main purpose of this patch is to tune the set of flags that tsan is running with.

Comment 9

[Steve Fink [:sfink] [:s:]]

Attachment: Skip tests that depend on the wasm signal handler, since tsan interferes with it

It turns out that it's some sort of subtle interaction with how tsan manages signal handlers and the signal handler we use for wasm. Just skip these tests for now.

Comment 10

[Jon Coppeard (:jonco) (PTO until 3rd April)]

Comment on attachment 8865618 [details] [diff] [review]
Avoid problems with timeouts when running under tsan

Review of attachment 8865618 [details] [diff] [review]:
-----------------------------------------------------------------

Yes, seems sensible.

Comment 11

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/3b3067e901d4
Avoid problems with timeouts when running under tsan, r=jonco
https://hg.mozilla.org/integration/mozilla-inbound/rev/e03be5e59295
Skip wasm/asm.js for tsan test, r=jonco
https://hg.mozilla.org/integration/mozilla-inbound/rev/3a43c13ce065
Skip tests that depend on the wasm signal handler, since tsan interferes with it, r=jonco

Comment 12

[Steve Fink [:sfink] [:s:]]

Attachment: Update tsan failures to current set and move SM(tsan) to tier 1

jonco, nothing much here to review. This moves SM(tsan) to tier 1. I can't think of any remaining reason not to now that it's green. I'll try to fix some of the failures in fast followups.

Comment 13

[Jon Coppeard (:jonco) (PTO until 3rd April)]

Comment on attachment 8867031 [details] [diff] [review]
Update tsan failures to current set and move SM(tsan) to tier 1

Review of attachment 8867031 [details] [diff] [review]:
-----------------------------------------------------------------

\o/

Comment 14

[Carsten Book [:Tomcat]]

https://hg.mozilla.org/mozilla-central/rev/3b3067e901d4
https://hg.mozilla.org/mozilla-central/rev/e03be5e59295
https://hg.mozilla.org/mozilla-central/rev/3a43c13ce065

Comment 15

[Steve Fink [:sfink] [:s:]]

I don't think this verification approach is going to work. tsan is a dynamic checker, so the error count is random. Also, the reports of "expected error XXX not seen" are noise.

I'm thinking a workable approach would be to run a bunch of times, collect all failures seen, and add them to the expected set. Then make the job fail whenever any new error appears, regardless of the error count. We will just have to accept that low-probability errors will appear occasionally, and we'll need to add them to the expected set. Distinguishing these from legitimate new errors will be the tough part.

Also, this raises the question of when an error can be removed from the expected set. I think for now, we'll have to rely on people to remove errors when they believe they are fixing them, and allowing some number of stale errors to pile up over time. We can periodically purge.

A better approach would be to report the results of every run to a common place, and notice when an error hasn't appeared in N runs. That smells like future work.

Unfortunately, all of this does mean that new errors can hide on a per-file basis. Since line numbers are unstable, the expectation is a mapping of filenames to error counts. So if you introduce a new error in foo.cpp that already expects 3 errors, and you don't happen to hit one of the known errors, then the job will pass and some random later patch will be blamed when it reports 4 errors in that file. Hopefully this shouldn't come up often, and the best fix would be getting down to zero errors anyway.

Comment 16

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/2db29775e18f
Update tsan failures to current set and move SM(tsan) to tier 1, r=jonco

Comment 17

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/64e6fe94b7b0
followup, revert the part of changeset 2db29775e18f which prematurely declared SM(tsan) tier-1

Comment 18

[Phil Ringnalda (:philor)]

No, absolutely not.

https://wiki.mozilla.org/Sheriffing/Job_Visibility_Policy#Usable_job_logs - a TinderboxPrint which appears in the never-seen Job Details panel is not enough, we never should have let you get away with the way hazards are displayed, but at least they are better than this. It's not even tier-2 until it displays failure messages in a treeherder-parsable way in the Failure Summary panel.

https://wiki.mozilla.org/Sheriffing/Job_Visibility_Policy#Runs_on_mozilla-central_and_all_trees_that_merge_into_it - I believe that it's the "run-on-projects: []", which my vague memory says is how you say that you only want a job to run on Try and only when someone explicitly asks for it, but it doesn't run on any trees other than Try when you ask for it, making it the weirdest tier-1 ever.

https://wiki.mozilla.org/Sheriffing/Job_Visibility_Policy#Low_intermittent_failure_rate - five runs to declare victory was far too few, so I triggered another 45 on https://treeherder.mozilla.org/#/jobs?repo=try&revision=4c58a67710e6ffc72c8ebe049ab9425fe62c4f9a&group_state=expanded&noautoclassify and that's a 22% failure rate, not the barely tolerable for a new suite 5%.

Comment 19

[Jan de Mooij [:jandem]]

This seems fine now:

https://treeherder.mozilla.org/#/jobs?repo=mozilla-inbound&revision=1f6d4b0df2226e001496060c4f4ba439e2b1f143&filter-tier=1&filter-tier=2&filter-tier=3&filter-searchStr=tsan&group_state=expanded

We also dump errors to stderr now.

Comment 20

[Steve Fink [:sfink] [:s:]]

sheriffs - anything else you'd want us to do before making this tier 1? The known issues have been miraculously fixed, and all of the fancy output processing I had has been ripped out so that I believe this should report in the same way as eg asan builds.

(Also, re: comment 18: I hadn't realized the hazard error reporting was a problem, so when philor complained I immediately jumped up, paused for a year, then filed bug 1459378. Time is slippery.)

Comment 21

[Sebastian Hengst [:aryx] (needinfo me if it's about an intermittent or backout)]

Looks good to me.

Comment 22

[Phil Ringnalda (:philor)]

Dunno, I guess maybe? There have been zero failures since last Friday, which is nice but makes it impossible to assess whether or not it actually displays its errors reasonably.

Comment 23

[Jan de Mooij [:jandem]]

NI to make sure we don't forget about this (I'm worried it will regress quickly if we don't make it tier 1).

Comment 24

[GitHub Bugzilla PR Linker]

Attachment: Link to GitHub pull-request: https://github.com/mozilla/treeherder/pull/3553

Comment 25

[Steve Fink [:sfink] [:s:]]

I wanted to check out what failures look like, so I introduced a bunch of races (it didn't complain about my first attempt!). I can't say it looks all that great:

https://treeherder.mozilla.org/#/jobs?repo=try&author=[e-mail]&fromchange=02a45e75c05f98df89f87e1171e47b64f6309020&selectedJob=178122029&filter-tier=1&filter-tier=2&filter-tier=3

The number of errors is deceptive because this is failing on many, many tests. I scanned through try for a recent tsan failure. Here's one:

https://treeherder.mozilla.org/#/jobs?repo=try&revision=89fdae9f4f5e6dbcf0abb820fd04b6cdaf3c953f&filter-searchStr=tsan&filter-tier=1&filter-tier=2&filter-tier=3&selectedJob=178305107

Its Failure Summary, in its entirety, is:

TEST-UNEXPECTED-FAIL | js/src/jit-test/tests/gc/bug-1136597.js | ================== (code 66, args "") [1.1 s]
make[1]: *** [check-jit-test] Error 2
make: *** [check-jit-test] Error 2

Yuck. One thing that would make this much better, good enough for landing, is to just do what AddressSanitizer does and highlight the SUMMARY: line on failures. I've opened the PR https://github.com/mozilla/treeherder/pull/3553 for that.

So I think I'll go ahead and land this tier-1 now.

Comment 26

[Steve Fink [:sfink] [:s:]]

Attachment: Promote SM(tsan) to tier 1

Ok, maybe I'll get sheriff review first, just to cover my argyle socks.

Comment 27

[Treeherder GitHub Bugbot]

Commit pushed to master at https://github.com/mozilla/treeherder

https://github.com/mozilla/treeherder/commit/ffdb06fc37b414cf77e03b532b7bcfe46f3d0ef1
Bug 1291954 - Highlight tsan errors (#3553)

Comment 28

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/a4257fbc42c8
Promote SM(tsan) to tier 1, r=Aryx

Comment 29

[Pulsebot]

Backout by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/88d835e400ea
Backed out changeset a4257fbc42c8 because tsan is permafail right now

Comment 30

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/4481e5d95c62
Promote SM(tsan) to tier 1, r=Aryx

Comment 31

[Jan de Mooij [:jandem]]

(In reply to Pulsebot from comment #30)
> Pushed by [e-mail]:
> https://hg.mozilla.org/integration/mozilla-inbound/rev/4481e5d95c62
> Promote SM(tsan) to tier 1, r=Aryx

\o/

Comment 32

[Dorel Luca [:dluca]]

https://hg.mozilla.org/mozilla-central/rev/4481e5d95c62

Comment 33

[Jon Coppeard (:jonco) (PTO until 3rd April)]

Nice!

Comment 34

[Ryan VanderMeulen [:RyanVM]]

Confirmed running and visible on Treeherder :)

Comment 35

[Sebastian Hengst [:aryx] (needinfo me if it's about an intermittent or backout)]

Please add documentation for this test suite to https://developer.mozilla.org/en-US/docs/Mozilla/QA/Automated_testing and also explain what kind of failures shall be reported as restricted security bugs and which ones not. Thank you.

Comment 36

[Steve Fink [:sfink] [:s:]]

(In reply to Sebastian Hengst [:aryx][:archaeopteryx] (needinfo on intermittent or backout) from comment #35)
> Please add documentation for this test suite to
> https://developer.mozilla.org/en-US/docs/Mozilla/QA/Automated_testing and

It is already covered under the SM(...) row, which links to <https://wiki.mozilla.org/Javascript:Automation_Builds>. I have replaced the contents of the tsan section on that page to remove the outdated whitelist info. Is that adequate? I could add a row specifically for this job, but it would turn out to be a complete duplicate of the SM(...) row.

> also explain what kind of failures shall be reported as restricted security
> bugs and which ones not. Thank you.

I put in some weaselly language for that "ThreadSanitizer failures are generally not security-critical, though they can be." But my confidence in that assessment is not high. Mostly, if this build fails on the initial push of a problem, then it's somewhat likely that it could be a security issue -- but it doesn't really matter with this tier 1, since it'll immediately get backed out anyway. The intermittents are more problematic, and while I could imagine them indeed being security sensitive, I'm not sure if they are any more so than other intermittents?

I dunno, other people may feel wildly differently.

Comment 37

[Steve Fink [:sfink] [:s:]]

Ok, that seemed a little too glib, so I updated to "ThreadSanitizer failures are generally not security sensitive, though they can be. The usual rules apply -- eg if it's in GC code, it should be assumed to be sensitive until investigated."

Comment 38

[Sebastian Hengst [:aryx] (needinfo me if it's about an intermittent or backout)]

Thank you, that works.

Comment 39

[Nicholas Nethercote [inactive]]

Sounds reasonable to me too. I would say that TSAN errors aren't security sensitive the way ASAN errors are, for example.