This adds the ability to run Android reftests from a taskcluster interactive loaner using
mach.
MozReview-Commit-ID: 5HMfm64wmK2
--HG--
extra : rebase_source : c17763e6916ef5dbaab25c6de6691aeb7541c3d9
When using an interactive loaner, we compute an explicit path to the adb binary. Passing it in
directly to mozdevice seems a bit cleaner than modifying the $PATH. This also does a minor
refactor around how the DM instances are instantiated.
MozReview-Commit-ID: 5HMfm64wmK2
--HG--
extra : rebase_source : a42afcc49f2b07892c2da083c5ce3589276cc765
Reftests hackily create a second argument parser and set defaults on it to get around the fact
that the mach command sends the harness a dict rather than a Namespace object. This is bad because:
1. It's much less hacky to just create a Namespace object directly (rather than making a second parser)
2. Most other mach commands actually *do* return a Namespace object, reftest is the odd one out here.
So this patch makes the reftest mach commands convert to the Namespace object, which allows us to get
rid of the hacks in the reftest harness. This also does some light refactoring of the entry points to
the reftest harness so make it more consistent with mochitest and xpcshell.
MozReview-Commit-ID: 5HMfm64wmK2
--HG--
extra : rebase_source : 7e9c0fc64e3aeced7662a9df13f94b1b5d751bc8
Android reftests still use os.chdir and imp to import reftest files. But there's a
_setup_objdir method (used by b2g) that does a similar thing instead. I decided to
try getting Android to use this as well and it seems to work. This is just a cleanup.
MozReview-Commit-ID: 5HMfm64wmK2
--HG--
extra : rebase_source : b9f52c01e0d1311a0c4511ffbc0a60e7d149ffe8
I noticed that if you request a loaner for a non-e10s job then run mach, it will be run with e10s
enabled. The mach command should accurately reflect the type of job that got requested. This patch
grabs the 'e10s' argument from the mozharness localconfig.json and uses that.
MozReview-Commit-ID: 4lsKGpizfH7
--HG--
extra : rebase_source : 6f7b45146bb7f3431c242d92b4206d41f7217c03
You can set attributes on a mach context by using the 'key' argument to the context_handler. Basically,
whatever gets returned by the handler when <key> is passed in, will get set (i.e cached) on the context
object for fast retrieval next time. This is a way to lazy load these attributes.
Previously I was setting functions like 'find_firefox()' on the context object, and then having the
mach_commands call that directly. But this way is much cleaner.
Now, the loaded 'mozharness_config' can be stored as an attribute on the context. Also 'find_firefox()'
is now an attribute called 'firefox_bin'.
MozReview-Commit-ID: 4lsKGpizfH7
--HG--
extra : rebase_source : af5e32e05a29b8e91d4cd0005689baa8079ec137
Ensure that the first line of the log for failing tests is identical to
one that would be produced by the tbplformatter from mozlog. This means
that treeherder will be able to sucessfully cross-match error lines.
Also make the data structures used for storing screenshots identical
between reftests and web-platform-tests so that we can start to move
toward removing the reftest-specific codepath here.
MozReview-Commit-ID: FZQXLjj9Ejv
--HG--
extra : rebase_source : 3e9bf70ff7915817b8c02032c8da0d0a659e4661
This fixes a regression from bug 1288827. It happened because I moved the logic that finds
the application path a little later on in the test harness. But there was an instance where
it was being used in the android mach command before that point.
As it turned out, we don't really *need* that value there. This patch grabs the same value
from build_obj.substs which is already an argument to the function.
MozReview-Commit-ID: 3IsI4VzEIIF
--HG--
extra : rebase_source : ceb718b265c7886b831e0378436baebf2b4e7598
This removes the unnecessary setting of c-basic-offset from all
python-mode files.
This was automatically generated using
perl -pi -e 's/; *c-basic-offset: *[0-9]+//'
... on the affected files.
The bulk of these files are moz.build files but there a few others as
well.
MozReview-Commit-ID: 2pPf3DEiZqx
--HG--
extra : rebase_source : 0a7dcac80b924174a2c429b093791148ea6ac204
This adds reftest support to the test package mach environment. It will allow
developers to easily run reftests after checking out an interactive worker.
MozReview-Commit-ID: fBAbfuG5XQ
--HG--
extra : rebase_source : 84b4a9fff7f2f27a325ffad4af1de7726bad296e
The idea of mozlog ids being tuples unfortunately didn't work so well
with external systems that found the tuple|string datatype difficult to
work with. Convert reftests to use simple ids of the form "<test url>
<comparison> <refurl>" e.g. "about:blank == data:text/html," instead of
tuples.
MozReview-Commit-ID: 18jufbssn4A
--HG--
extra : rebase_source : 542888186fecf3671847d1fca307770c7a989ad3
In bug 1271035 gps identified disk I/O as a major cause for reftest slowness.
He was able to fix most of it, but the new highest source of I/O in reftest is
reftest.log.
To note, we were only saving reftest.log when running via mach, so this won't
impact automation in any way:
https://hg.mozilla.org/mozilla-central/file/043082cb7bd8/layout/tools/reftest/mach_commands.py#l204
I don't know why we are doing this given that the same output goes to stdout.
And as dholbert pointed out, since bug 1034290 landed, that log contains raw
structured logs, which are not at all useful for debugging. Given that it is
no longer useful and causes slowness, we should stop saving it.
If anyone wishes to keep saving to a log, they can use:
./mach reftest --log-tbpl reftest.log
If they wish this to be the default behaviour they can make a machrc (or
.machrc) in either topsrcdir, ~/.mozbuild or $MACHRC. Then add:
[alias]
reftest = reftest --log-tbpl reftest.log
MozReview-Commit-ID: A3e2X7qF90H
--HG--
extra : rebase_source : 9286e3061c068c6fef55b503017b3f4a9588be46
It was a cold Friday night in San Francisco. Earlier in the day, I
informed Chris AtLee that I was going to start focusing on improving
the efficiency of Firefox automation and asked him where the biggest
capacity issues were. He said "we're hurting most on Windows tests."
As I was casually drinking a barleywine (note to reader: barleywines
are serious beers - there's nothing casual about them), I found myself
tediously clicking through Treeherder looking at logs for Windows jobs,
looking for patterns and other oddities. As I was clicking through,
something stood out to me: the sheer number of reftest jobs. I
recalled a random project I started a few years ago. Its aim was to
analyze buildbot job metadata so we could better understand where time
was spent in automation. I had mostly written off the side project as
a failure and a near complete waste of my time. Not even a stray
random thought of this project had entered my mind in the past year.
But clicking through Treeherder after a few glasses of barleywine
somehow reminded me of one of the few useful findings of that project:
reftest jobs consumed a seemingly disproportiate amount of machine time,
something like 35 or 40% IIRC of the time spent on all jobs.
Now, this finding was several years ago and almost certainly no longer
relevant. But, again, I had a few glasses of barleywine in me and was
bothered by the amount of reftest jobs and their duration, so I thought
"hey, why don't I run reftests and see why they take so long." So I
built Firefox on Windows - the platform Chris AtLee said we're "hurting
most on."
I decided to start my very casual profiling session by recording a
`mach reftest` run using Sysinternals Process Monitor. To my surprise,
it yielded a very obvious and unexpected result: the Places SQLite
database was incurring a lot of I/O. On my i7-6700K Windows 10 desktop
with a high performance SSD, `mach reftest` yielded the following:
File Time #Events #Reads #Writes RBytes WBytes Path
198s 980,872 243,270 669,231 7,971,471,360 20,667,084,080 places.sqlite-wal
165s 645,853 222,407 367,775 7,287,701,820 14.071,529,472 places.sqlite
2s 377,121 1 0 32,768 0 places.sqlite-shm
The Places SQLite database accounts for 2,003,846 of the total of
3,547,527 system calls (56.49%) recorded by procmon during `mach
reftest` execution. This represents a staggering 49,997,786,732 of the
50,307,660,589 (99.38%) bytes of I/O recorded! Yes, that's 50 GB.
I reckon the reason the Places database accumulates so much I/O load
during reftests is because the reftest suite essentially loads thousands
of pages as quickly as possible. This effectively performs a stress
test against the Places history recording service.
As effective as reftests are at stress-testing Places, it adds no value
to reftests because reftests are testing the layout features, not the
performance of history recording. So these 2M system calls and 50 GB
of I/O are overhead.
This commit disables Places when executing reftests and prevents
the overhead.
After this commit, `mach reftest` has significantly reduced interaction
with the Places SQLite database:
File Time #Events #Reads #Writes RBytes WBytes Path
0.09s 502 138 302 4,521,984 8,961,528 places.sqlite-wal
0.07s 254 20 140 524,604 8,126,464 places.sqlite
0.01s 3,289 1 0 32,768 0 places.sqlite-shm
Of the 948,033 system calls recorded with this change (26.7% of
original), 691,322 were related to I/O. The Places SQLite database
only consumed ~22MB of I/O, <0.01% of original. It's worth noting that
~half of the remaining I/O system calls are related to reftest.log,
which now accounts for the largest percentage of write I/O (only
~53 MB, however). It's worth noting that reftest.log appears to be
using unbuffered writes and is requiring an excessive amount of
system calls for writing. But that's for another bug and commit.
In terms of wall time, the drastic I/O reduction during `mach reftest`
appears to have minimal impact on my machine: maybe 30s shaved from a
~900s execution, or ~3%. But my machine with its modern SSD doesn't
struggle with I/O.
In automation, it is a different story.
I pushed this change to Try along with the base revision and triggered
4 runs of most reftest jobs. The runtime improvements in automation
are impressive. Here are the fastest reported times for various jobs:
Job Before After Delta
Linux Opt R1 31m 34m +3m
Linux Opt R2 43m 35m -8m
Linux Opt Ru1 40m 34m -6m
Linux Opt Ru2 43m 37m -6m
Linux Opt R E10s 89m 72m -17m
Linux Debug R1 52m 40m -12m
Linux Debug R2 49m 42m -7m
Linux Debug R3 60m 51m -9m
Linux Debug R4 42m 37m -5m
Linux Debug R1 E10s 84m 72m -12m
Linux Debug R2 E10s 97m 85m -12m
Linux64 Opt R1 35m 24m -11m
Linux64 Opt R2 37m 26m -11m
Linux64 Opt Ru1 32m 29m -3m
Linux64 Opt Ru2 37m 26m -12m
Linux64 Opt TC R1 12m 10m -2m
Linux64 Opt TC R2 10m 7m -3m
Linux64 Opt TC R3 11m 9m -2m
Linux64 Opt TC R4 11m 9m -2m
Linux64 Opt TC R5 13m 11m -2m
Linux64 Opt TC R6 11m 9m -2m
Linux64 Opt TC R7 9m 8m -1m
Linux64 Opt TC R8 11m 10m -1m
Linux64 Opt TC Ru1 30m 25m -5m
Linux64 Opt TC Ru2 36m 27m -11m
OS X 10.10 Opt 31m 27m -4m
OS X 10.10 Opt E10s 26m 25m -1m
OS X 10.10 Debug 68m 55m -13m
Win7 Opt R 30m 28m -2m
Win7 Opt Ru 28m 26m -2m
Win7 Opt R E10S 29m 27m -2m
Win7 Debug R 85m 76m -9m
Win7 Debug R E10S 75m 65m -10m
Win8 x64 Opt R 29m 26m -3m
Win8 x64 Opt Ru 27m 25m -2m
Win8 x64 Debug R 90m 71m -19m
Android 4.3 API15 Opt R1 89m 71m -18m
Android 4.3 API15 Opt R2 78m 64m -14m
Android 4.3 API15 Opt R3 75m 64m -11m
Android 4.3 API15 Opt R4 74m 68m -6m
Android 4.3 API15 Opt R5 75m 69m -6m
Android 4.3 API15 Opt R6 91m 86m -5m
Android 4.3 API15 Opt R7 87m 66m -21m
Android 4.3 API15 Opt R8 87m 82m -5m
Android 4.3 API15 Opt R9 80m 66m -14m
Android 4.3 API15 Opt R10 80m 67m -13m
Android 4.3 API15 Opt R11 73m 66m -7m
Android 4.3 API15 Opt R12 105m 91m -14m
Android 4.3 API15 Opt R13 72m 59m -13m
Android 4.3 API15 Opt R14 82m 61m -21m
Android 4.3 API15 Opt R15 73m 62m -11m
Android 4.3 API15 Opt R16 79m 78m -1m
The savings total 6+ *hours* or ~15% when running all reftests. I'd
say this isn't bad for a one-line code change!
MozReview-Commit-ID: H1LkACgSpVn
--HG--
extra : rebase_source : 891a5ce8e1f6c3d70fc646f116c2f49f897ad735
As of bug 1245092, reftest depends on marionette. Normally marionette client has an internal timeout of around
60 seconds where it waits for the server to become available. But when using debuggers (and especially valgrind),
it can take much much longer than a minute for Firefox to start. This patch adds a couple hidden command line
args to reftest to tweak the marionette timeouts.
MozReview-Commit-ID: 3xF0InBJNEf
--HG--
extra : rebase_source : f18092bc90a7d8aab34b527a15bd3b1dc6afc997
This extra window was initially left open because closing it was causing memory leaks
on debug e10s crashtests. There was pressure to get the regressing patch landed due
to addon signing, so it got landed with this extra window hanging around (as it didn't
impact test results).
But it is a UX wart for several reasons. Upon testing it again recently, the leaks all
seem to have vanished. I'm not sure why, possibly it was a bug fixed in e10s.
MozReview-Commit-ID: CEI2enKAOyv
--HG--
extra : rebase_source : da0e8889f67d57aa61670105c14c2235607d79a7
This is a internal-only syntax for guarding rules from a boolean
preference. Nothing causes @supports rules to be re-evaluated except
html.css registered in Part 2.
This is needed for rendering the disclosure triangle of the
summary element by using "display: list-item".
Usage example:
@supports -moz-bool-pref("dom.details_element.enabled") {
/* css rules */
}
MozReview-Commit-ID: HDCa8zHxYTA
--HG--
extra : rebase_source : b7a72a48166edf1d486014ff37363ed8be9127d9
Andrew Halberstadt
Geoff Brown
Terrence Cole
Phil Ringnalda
Xidorn Quan
Ryan VanderMeulen
Andrew McCreight
Mason Chang
James Graham
Seth Fowler
Tom Tromey
Benjamin Smedberg
Matt Woodrow
Brad Werth
Krishna
Thomas Nguyen
Jonathan Hao
Gijs Kruitbosch
Chris Peterson
Andreas Tolfsen
Gregory Szorc
Mike Shal
Daniel Holbert
Hiroyuki Ikezoe
Sebastian Hengst
Joel Maher
Ting-Yu Lin
Tobias Schneider
Francois Marier
Xidorn Quan
Stanley Mwangi