copies: cache the ancestor checking call when tracing copy
A good share of the time spent in this function is spent doing ancestors
checking. To avoid spending time in duplicated call, we cache the result of
calls.
In the slower case, this provide a quite significant performance boost. Below
are the result for a set of selected pairs (many of them pathological):
(And further down is another table that summarize the current state of filelog
based vs changeset base copy tracing)
The benchmark have been configured to be killed after 6 minutes of runtime,
which mean that any detect slower than 2 minutes will be marked as "killed".
This drop some useful information about how much slower these case are… but also
prevent 99% of the benchmark time to be spent on case that can be labelled "very
slow" anyway.
Repo Case Source-Rev Dest-Rev Old-Time New-Time Difference Factor
------------------------------------------------------------------------------------------------------------------------------------
mercurial x_revs_x_added_0_copies
ad6b123de1c7 39cfcef4f463 : 0.000044 s, 0.000044 s, +0.000000 s, × 1.0000
mercurial x_revs_x_added_x_copies
2b1c78674230 0c1d10351869 : 0.000138 s, 0.000138 s, +0.000000 s, × 1.0000
mercurial x000_revs_x000_added_x_copies
81f8ff2a9bf2 dd3267698d84 : 0.005067 s, 0.005052 s, -0.000015 s, × 0.9970
pypy x_revs_x_added_0_copies
aed021ee8ae8 099ed31b181b : 0.000218 s, 0.000219 s, +0.000001 s, × 1.0046
pypy x_revs_x000_added_0_copies
4aa4e1f8e19a 359343b9ac0e : 0.000053 s, 0.000055 s, +0.000002 s, × 1.0377
pypy x_revs_x_added_x_copies
ac52eb7bbbb0 72e022663155 : 0.000125 s, 0.000128 s, +0.000003 s, × 1.0240
pypy x_revs_x00_added_x_copies
c3b14617fbd7 ace7255d9a26 : 0.001098 s, 0.001089 s, -0.000009 s, × 0.9918
pypy x_revs_x000_added_x000_copies
df6f7a526b60 a83dc6a2d56f : 0.017546 s, 0.017407 s, -0.000139 s, × 0.9921
pypy x000_revs_xx00_added_0_copies
89a76aede314 2f22446ff07e : 0.096723 s, 0.094175 s, -0.002548 s, × 0.9737
pypy x000_revs_x000_added_x_copies
8a3b5bfd266e 2c68e87c3efe : 0.271796 s, 0.238009 s, -0.033787 s, × 0.8757
pypy x000_revs_x000_added_x000_copies
89a76aede314 7b3dda341c84 : 0.128602 s, 0.125876 s, -0.002726 s, × 0.9788
pypy x0000_revs_x_added_0_copies
d1defd0dc478 c9cb1334cc78 : 7.086742 s, 3.581556 s, -3.505186 s, × 0.5054
pypy x0000_revs_xx000_added_0_copies
bf2c629d0071 4ffed77c095c : 0.016634 s, 0.016721 s, +0.000087 s, × 1.0052
pypy x0000_revs_xx000_added_x000_copies
08ea3258278e d9fa043f30c0 : 0.254225 s, 0.242367 s, -0.011858 s, × 0.9534
netbeans x_revs_x_added_0_copies
fb0955ffcbcd a01e9239f9e7 : 0.000166 s, 0.000165 s, -0.000001 s, × 0.9940
netbeans x_revs_x000_added_0_copies
6f360122949f 20eb231cc7d0 : 0.000118 s, 0.000114 s, -0.000004 s, × 0.9661
netbeans x_revs_x_added_x_copies
1ada3faf6fb6 5a39d12eecf4 : 0.000296 s, 0.000296 s, +0.000000 s, × 1.0000
netbeans x_revs_x00_added_x_copies
35be93ba1e2c 9eec5e90c05f : 0.001137 s, 0.001124 s, -0.000013 s, × 0.9886
netbeans x000_revs_xx00_added_0_copies
eac3045b4fdd 51d4ae7f1290 : 0.014133 s, 0.013060 s, -0.001073 s, × 0.9241
netbeans x000_revs_x000_added_x_copies
e2063d266acd 6081d72689dc : 0.016988 s, 0.017112 s, +0.000124 s, × 1.0073
netbeans x000_revs_x000_added_x000_copies
ff453e9fee32 411350406ec2 : 0.676361 s, 0.660350 s, -0.016011 s, × 0.9763
netbeans x0000_revs_xx000_added_x000_copies
588c2d1ced70 1aad62e59ddd : 12.515149 s, 10.032499 s, -2.482650 s, × 0.8016
mozilla-central x_revs_x_added_0_copies
3697f962bb7b 7015fcdd43a2 : 0.000186 s, 0.000189 s, +0.000003 s, × 1.0161
mozilla-central x_revs_x000_added_0_copies
dd390860c6c9 40d0c5bed75d : 0.000459 s, 0.000462 s, +0.000003 s, × 1.0065
mozilla-central x_revs_x_added_x_copies
8d198483ae3b 14207ffc2b2f : 0.000273 s, 0.000270 s, -0.000003 s, × 0.9890
mozilla-central x_revs_x00_added_x_copies
98cbc58cc6bc 446a150332c3 : 0.001503 s, 0.001474 s, -0.000029 s, × 0.9807
mozilla-central x_revs_x000_added_x000_copies
3c684b4b8f68 0a5e72d1b479 : 0.004862 s, 0.004806 s, -0.000056 s, × 0.9885
mozilla-central x_revs_x0000_added_x0000_copies
effb563bb7e5 c07a39dc4e80 : 0.088291 s, 0.085150 s, -0.003141 s, × 0.9644
mozilla-central x000_revs_xx00_added_0_copies
6100d773079a 04a55431795e : 0.007113 s, 0.007064 s, -0.000049 s, × 0.9931
mozilla-central x000_revs_x000_added_x_copies
9f17a6fc04f9 2d37b966abed : 0.004687 s, 0.004741 s, +0.000054 s, × 1.0115
mozilla-central x000_revs_x000_added_x000_copies
7c97034feb78 4407bd0c6330 : 0.198710 s, 0.190133 s, -0.008577 s, × 0.9568
mozilla-central x0000_revs_xx000_added_0_copies
9eec5917337d 67118cc6dcad : 0.036068 s, 0.035651 s, -0.000417 s, × 0.9884
mozilla-central x0000_revs_xx000_added_x000_copies
f78c615a656c 96a38b690156 : 0.465362 s, 0.440694 s, -0.024668 s, × 0.9470
mozilla-central x00000_revs_x0000_added_x0000_copies
6832ae71433c 4c222a1d9a00 : 24.519684 s, 18.454163 s, -6.065521 s, × 0.7526
mozilla-central x00000_revs_x00000_added_x000_copies
76caed42cf7c 1daa622bbe42 : 42.711897 s, 31.562719 s, -11.149178 s, × 0.7390
mozilla-try x_revs_x_added_0_copies
aaf6dde0deb8 9790f499805a : 0.001201 s, 0.001189 s, -0.000012 s, × 0.9900
mozilla-try x_revs_x000_added_0_copies
d8d0222927b4 5bb8ce8c7450 : 0.001216 s, 0.001204 s, -0.000012 s, × 0.9901
mozilla-try x_revs_x_added_x_copies
092fcca11bdb 936255a0384a : 0.000595 s, 0.000586 s, -0.000009 s, × 0.9849
mozilla-try x_revs_x00_added_x_copies
b53d2fadbdb5 017afae788ec : 0.001856 s, 0.001845 s, -0.000011 s, × 0.9941
mozilla-try x_revs_x000_added_x000_copies
20408ad61ce5 6f0ee96e21ad : 0.064936 s, 0.063822 s, -0.001114 s, × 0.9828
mozilla-try x_revs_x0000_added_x0000_copies
effb563bb7e5 c07a39dc4e80 : 0.090601 s, 0.088038 s, -0.002563 s, × 0.9717
mozilla-try x000_revs_xx00_added_0_copies
6100d773079a 04a55431795e : 0.007510 s, 0.007389 s, -0.000121 s, × 0.9839
mozilla-try x000_revs_x000_added_x_copies
9f17a6fc04f9 2d37b966abed : 0.004911 s, 0.004868 s, -0.000043 s, × 0.9912
mozilla-try x000_revs_x000_added_x000_copies
1346fd0130e4 4c65cbdabc1f : 0.233231 s, 0.222450 s, -0.010781 s, × 0.9538
mozilla-try x0000_revs_x_added_0_copies
63519bfd42ee a36a2a865d92 : 0.419989 s, 0.370675 s, -0.049314 s, × 0.8826
mozilla-try x0000_revs_x_added_x_copies
9fe69ff0762d bcabf2a78927 : 0.401521 s, 0.358020 s, -0.043501 s, × 0.8917
mozilla-try x0000_revs_xx000_added_x_copies
156f6e2674f2 4d0f2c178e66 : 0.179555 s, 0.145235 s, -0.034320 s, × 0.8089
mozilla-try x0000_revs_xx000_added_0_copies
9eec5917337d 67118cc6dcad : 0.038004 s, 0.037606 s, -0.000398 s, × 0.9895
mozilla-try x0000_revs_xx000_added_x000_copies
89294cd501d9 7ccb2fc7ccb5 : 52.838482 s, 7.382439 s, -45.456043 s, × 0.1397
mozilla-try x0000_revs_x0000_added_x0000_copies
e928c65095ed e951f4ad123a : 8.705874 s, 7.273506 s, -1.432368 s, × 0.8355
mozilla-try x00000_revs_x00000_added_0_copies
dc8a3ca7010e d16fde900c9c : 1.126708 s, 1.074593 s, -0.052115 s, × 0.9537
mozilla-try x00000_revs_x0000_added_x0000_copies
8d3fafa80d4b eb884023b810 : 83.854020 s, 27.746195 s, -56.107825 s, × 0.3309
Below is a table comparing the runtime of the current "filelog centric"
algorithm, with the "changeset centric" one, we just modified.
The changeset centric algorithm is a significant win in many scenario, but they
are still various cases where it is quite slower. When many revision has to be
considered the cost of retrieving the copy information, creating new
dictionaries, merging dictionaries and checking if revision are ancestors of
each other can slow things down.
The rest of this series, will introduce a rust version of the copy tracing code
to deal with most of theses issues.
Repo Case Source-Rev Dest-Rev filelog sidedata Difference Factor
---------------------------------------------------------------------------------------------------------------------------------------
mercurial x_revs_x_added_0_copies
ad6b123de1c7 39cfcef4f463 : 0.000914 s, 0.000044 s, - 0.000870 s, × 0.048140
mercurial x_revs_x_added_x_copies
2b1c78674230 0c1d10351869 : 0.001812 s, 0.000138 s, - 0.001674 s, × 0.076159
mercurial x000_revs_x000_added_x_copies
81f8ff2a9bf2 dd3267698d84 : 0.017954 s, 0.005052 s, - 0.012902 s, × 0.281386
pypy x_revs_x_added_0_copies
aed021ee8ae8 099ed31b181b : 0.001509 s, 0.000219 s, - 0.001290 s, × 0.145129
pypy x_revs_x000_added_0_copies
4aa4e1f8e19a 359343b9ac0e : 0.206881 s, 0.000055 s, - 0.206826 s, × 0.000266
pypy x_revs_x_added_x_copies
ac52eb7bbbb0 72e022663155 : 0.016951 s, 0.000128 s, - 0.016823 s, × 0.007551
pypy x_revs_x00_added_x_copies
c3b14617fbd7 ace7255d9a26 : 0.019096 s, 0.001089 s, - 0.018007 s, × 0.057028
pypy x_revs_x000_added_x000_copies
df6f7a526b60 a83dc6a2d56f : 0.762506 s, 0.017407 s, - 0.745099 s, × 0.022829
pypy x000_revs_xx00_added_0_copies
89a76aede314 2f22446ff07e : 1.179211 s, 0.094175 s, - 1.085036 s, × 0.079863
pypy x000_revs_x000_added_x_copies
8a3b5bfd266e 2c68e87c3efe : 1.249058 s, 0.238009 s, - 1.011049 s, × 0.190551
pypy x000_revs_x000_added_x000_copies
89a76aede314 7b3dda341c84 : 1.614107 s, 0.125876 s, - 1.488231 s, × 0.077985
pypy x0000_revs_x_added_0_copies
d1defd0dc478 c9cb1334cc78 : 0.001064 s, 3.581556 s, + 3.580492 s, × 3366.124060
pypy x0000_revs_xx000_added_0_copies
bf2c629d0071 4ffed77c095c : 1.061275 s, 0.016721 s, - 1.044554 s, × 0.015756
pypy x0000_revs_xx000_added_x000_copies
08ea3258278e d9fa043f30c0 : 1.341119 s, 0.242367 s, - 1.098752 s, × 0.180720
netbeans x_revs_x_added_0_copies
fb0955ffcbcd a01e9239f9e7 : 0.027803 s, 0.000165 s, - 0.027638 s, × 0.005935
netbeans x_revs_x000_added_0_copies
6f360122949f 20eb231cc7d0 : 0.130014 s, 0.000114 s, - 0.129900 s, × 0.000877
netbeans x_revs_x_added_x_copies
1ada3faf6fb6 5a39d12eecf4 : 0.024990 s, 0.000296 s, - 0.024694 s, × 0.011845
netbeans x_revs_x00_added_x_copies
35be93ba1e2c 9eec5e90c05f : 0.052201 s, 0.001124 s, - 0.051077 s, × 0.021532
netbeans x000_revs_xx00_added_0_copies
eac3045b4fdd 51d4ae7f1290 : 0.037642 s, 0.013060 s, - 0.024582 s, × 0.346953
netbeans x000_revs_x000_added_x_copies
e2063d266acd 6081d72689dc : 0.197086 s, 0.017112 s, - 0.179974 s, × 0.086825
netbeans x000_revs_x000_added_x000_copies
ff453e9fee32 411350406ec2 : 0.935148 s, 0.660350 s, - 0.274798 s, × 0.706145
netbeans x0000_revs_xx000_added_x000_copies
588c2d1ced70 1aad62e59ddd : 3.920674 s, 10.032499 s, + 6.111825 s, × 2.558871
mozilla-central x_revs_x_added_0_copies
3697f962bb7b 7015fcdd43a2 : 0.024232 s, 0.000189 s, - 0.024043 s, × 0.007800
mozilla-central x_revs_x000_added_0_copies
dd390860c6c9 40d0c5bed75d : 0.141483 s, 0.000462 s, - 0.141021 s, × 0.003265
mozilla-central x_revs_x_added_x_copies
8d198483ae3b 14207ffc2b2f : 0.025775 s, 0.000270 s, - 0.025505 s, × 0.010475
mozilla-central x_revs_x00_added_x_copies
98cbc58cc6bc 446a150332c3 : 0.084922 s, 0.001474 s, - 0.083448 s, × 0.017357
mozilla-central x_revs_x000_added_x000_copies
3c684b4b8f68 0a5e72d1b479 : 0.194784 s, 0.004806 s, - 0.189978 s, × 0.024673
mozilla-central x_revs_x0000_added_x0000_copies
effb563bb7e5 c07a39dc4e80 : 2.161103 s, 0.085150 s, - 2.075953 s, × 0.039401
mozilla-central x000_revs_xx00_added_0_copies
6100d773079a 04a55431795e : 0.089347 s, 0.007064 s, - 0.082283 s, × 0.079063
mozilla-central x000_revs_x000_added_x_copies
9f17a6fc04f9 2d37b966abed : 0.732171 s, 0.004741 s, - 0.727430 s, × 0.006475
mozilla-central x000_revs_x000_added_x000_copies
7c97034feb78 4407bd0c6330 : 1.157287 s, 0.190133 s, - 0.967154 s, × 0.164292
mozilla-central x0000_revs_xx000_added_0_copies
9eec5917337d 67118cc6dcad : 6.726568 s, 0.035651 s, - 6.690917 s, × 0.005300
mozilla-central x0000_revs_xx000_added_x000_copies
f78c615a656c 96a38b690156 : 3.266229 s, 0.440694 s, - 2.825535 s, × 0.134924
mozilla-central x00000_revs_x0000_added_x0000_copies
6832ae71433c 4c222a1d9a00 : 15.860534 s, 18.454163 s, + 2.593629 s, × 1.163527
mozilla-central x00000_revs_x00000_added_x000_copies
76caed42cf7c 1daa622bbe42 : 20.450475 s, 31.562719 s, +11.112244 s, × 1.543373
mozilla-try x_revs_x_added_0_copies
aaf6dde0deb8 9790f499805a : 0.080442 s, 0.001189 s, - 0.079253 s, × 0.014781
mozilla-try x_revs_x000_added_0_copies
d8d0222927b4 5bb8ce8c7450 : 0.497672 s, 0.001204 s, - 0.496468 s, × 0.002419
mozilla-try x_revs_x_added_x_copies
092fcca11bdb 936255a0384a : 0.021183 s, 0.000586 s, - 0.020597 s, × 0.027664
mozilla-try x_revs_x00_added_x_copies
b53d2fadbdb5 017afae788ec : 0.230991 s, 0.001845 s, - 0.229146 s, × 0.007987
mozilla-try x_revs_x000_added_x000_copies
20408ad61ce5 6f0ee96e21ad : 1.118461 s, 0.063822 s, - 1.054639 s, × 0.057062
mozilla-try x_revs_x0000_added_x0000_copies
effb563bb7e5 c07a39dc4e80 : 2.206083 s, 0.088038 s, - 2.118045 s, × 0.039907
mozilla-try x000_revs_xx00_added_0_copies
6100d773079a 04a55431795e : 0.089404 s, 0.007389 s, - 0.082015 s, × 0.082647
mozilla-try x000_revs_x000_added_x_copies
9f17a6fc04f9 2d37b966abed : 0.733043 s, 0.004868 s, - 0.728175 s, × 0.006641
mozilla-try x000_revs_x000_added_x000_copies
1346fd0130e4 4c65cbdabc1f : 1.163367 s, 0.222450 s, - 0.940917 s, × 0.191212
mozilla-try x0000_revs_x_added_0_copies
63519bfd42ee a36a2a865d92 : 0.085456 s, 0.370675 s, + 0.285219 s, × 4.337612
mozilla-try x0000_revs_x_added_x_copies
9fe69ff0762d bcabf2a78927 : 0.083601 s, 0.358020 s, + 0.274419 s, × 4.282485
mozilla-try x0000_revs_xx000_added_x_copies
156f6e2674f2 4d0f2c178e66 : 7.366614 s, 0.145235 s, - 7.221379 s, × 0.019715
mozilla-try x0000_revs_xx000_added_0_copies
9eec5917337d 67118cc6dcad : 6.664464 s, 0.037606 s, - 6.626858 s, × 0.005643
mozilla-try x0000_revs_xx000_added_x000_copies
89294cd501d9 7ccb2fc7ccb5 : 7.467836 s, 7.382439 s, - 0.085397 s, × 0.988565
mozilla-try x0000_revs_x0000_added_x0000_copies
e928c65095ed e951f4ad123a : 9.801294 s, 7.273506 s, - 2.527788 s, × 0.742097
mozilla-try x00000_revs_x_added_0_copies
6a320851d377 1ebb79acd503 : 0.091886 s, killed
mozilla-try x00000_revs_x00000_added_0_copies
dc8a3ca7010e d16fde900c9c : 26.491140 s, 1.074593 s, -25.416547 s, × 0.040564
mozilla-try x00000_revs_x_added_x_copies
5173c4b6f97c 95d83ee7242d : 0.092863 s, killed
mozilla-try x00000_revs_x000_added_x_copies
9126823d0e9c ca82787bb23c : 0.226823 s, killed
mozilla-try x00000_revs_x0000_added_x0000_copies
8d3fafa80d4b eb884023b810 : 18.914630 s, 27.746195 s, + 8.831565 s, × 1.466917
mozilla-try x00000_revs_x00000_added_x0000_copies
1b661134e2ca 1ae03d022d6d : 21.198903 s, killed
mozilla-try x00000_revs_x00000_added_x000_copies
9b2a99adc05e 8e29777b48e6 : 24.952268 s, killed
Differential Revision: https://phab.mercurial-scm.org/D9296
# __init__.py - fsmonitor initialization and overrides
#
# Copyright 2013-2016 Facebook, Inc.
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.
'''Faster status operations with the Watchman file monitor (EXPERIMENTAL)
Integrates the file-watching program Watchman with Mercurial to produce faster
status results.
On a particular Linux system, for a real-world repository with over 400,000
files hosted on ext4, vanilla `hg status` takes 1.3 seconds. On the same
system, with fsmonitor it takes about 0.3 seconds.
fsmonitor requires no configuration -- it will tell Watchman about your
repository as necessary. You'll need to install Watchman from
https://facebook.github.io/watchman/ and make sure it is in your PATH.
fsmonitor is incompatible with the largefiles and eol extensions, and
will disable itself if any of those are active.
The following configuration options exist:
::
[fsmonitor]
mode = {off, on, paranoid}
When `mode = off`, fsmonitor will disable itself (similar to not loading the
extension at all). When `mode = on`, fsmonitor will be enabled (the default).
When `mode = paranoid`, fsmonitor will query both Watchman and the filesystem,
and ensure that the results are consistent.
::
[fsmonitor]
timeout = (float)
A value, in seconds, that determines how long fsmonitor will wait for Watchman
to return results. Defaults to `2.0`.
::
[fsmonitor]
blacklistusers = (list of userids)
A list of usernames for which fsmonitor will disable itself altogether.
::
[fsmonitor]
walk_on_invalidate = (boolean)
Whether or not to walk the whole repo ourselves when our cached state has been
invalidated, for example when Watchman has been restarted or .hgignore rules
have been changed. Walking the repo in that case can result in competing for
I/O with Watchman. For large repos it is recommended to set this value to
false. You may wish to set this to true if you have a very fast filesystem
that can outpace the IPC overhead of getting the result data for the full repo
from Watchman. Defaults to false.
::
[fsmonitor]
warn_when_unused = (boolean)
Whether to print a warning during certain operations when fsmonitor would be
beneficial to performance but isn't enabled.
::
[fsmonitor]
warn_update_file_count = (integer)
# or when mercurial is built with rust support
warn_update_file_count_rust = (integer)
If ``warn_when_unused`` is set and fsmonitor isn't enabled, a warning will
be printed during working directory updates if this many files will be
created.
'''
# Platforms Supported
# ===================
#
# **Linux:** *Stable*. Watchman and fsmonitor are both known to work reliably,
# even under severe loads.
#
# **Mac OS X:** *Stable*. The Mercurial test suite passes with fsmonitor
# turned on, on case-insensitive HFS+. There has been a reasonable amount of
# user testing under normal loads.
#
# **Solaris, BSD:** *Alpha*. watchman and fsmonitor are believed to work, but
# very little testing has been done.
#
# **Windows:** *Alpha*. Not in a release version of watchman or fsmonitor yet.
#
# Known Issues
# ============
#
# * fsmonitor will disable itself if any of the following extensions are
# enabled: largefiles, inotify, eol; or if the repository has subrepos.
# * fsmonitor will produce incorrect results if nested repos that are not
# subrepos exist. *Workaround*: add nested repo paths to your `.hgignore`.
#
# The issues related to nested repos and subrepos are probably not fundamental
# ones. Patches to fix them are welcome.
from __future__ import absolute_import
import codecs
import os
import stat
import sys
import tempfile
import weakref
from mercurial.i18n import _
from mercurial.node import hex
from mercurial.pycompat import open
from mercurial import (
context,
encoding,
error,
extensions,
localrepo,
merge,
pathutil,
pycompat,
registrar,
scmutil,
util,
)
from mercurial import match as matchmod
from mercurial.utils import (
hashutil,
stringutil,
)
from . import (
pywatchman,
state,
watchmanclient,
)
# Note for extension authors: ONLY specify testedwith = 'ships-with-hg-core' for
# extensions which SHIP WITH MERCURIAL. Non-mainline extensions should
# be specifying the version(s) of Mercurial they are tested with, or
# leave the attribute unspecified.
testedwith = b'ships-with-hg-core'
configtable = {}
configitem = registrar.configitem(configtable)
configitem(
b'fsmonitor', b'mode', default=b'on',
)
configitem(
b'fsmonitor', b'walk_on_invalidate', default=False,
)
configitem(
b'fsmonitor', b'timeout', default=b'2',
)
configitem(
b'fsmonitor', b'blacklistusers', default=list,
)
configitem(
b'fsmonitor', b'watchman_exe', default=b'watchman',
)
configitem(
b'fsmonitor', b'verbose', default=True, experimental=True,
)
configitem(
b'experimental', b'fsmonitor.transaction_notify', default=False,
)
# This extension is incompatible with the following blacklisted extensions
# and will disable itself when encountering one of these:
_blacklist = [b'largefiles', b'eol']
def debuginstall(ui, fm):
fm.write(
b"fsmonitor-watchman",
_(b"fsmonitor checking for watchman binary... (%s)\n"),
ui.configpath(b"fsmonitor", b"watchman_exe"),
)
root = tempfile.mkdtemp()
c = watchmanclient.client(ui, root)
err = None
try:
v = c.command(b"version")
fm.write(
b"fsmonitor-watchman-version",
_(b" watchman binary version %s\n"),
pycompat.bytestr(v["version"]),
)
except watchmanclient.Unavailable as e:
err = stringutil.forcebytestr(e)
fm.condwrite(
err,
b"fsmonitor-watchman-error",
_(b" watchman binary missing or broken: %s\n"),
err,
)
return 1 if err else 0
def _handleunavailable(ui, state, ex):
"""Exception handler for Watchman interaction exceptions"""
if isinstance(ex, watchmanclient.Unavailable):
# experimental config: fsmonitor.verbose
if ex.warn and ui.configbool(b'fsmonitor', b'verbose'):
if b'illegal_fstypes' not in stringutil.forcebytestr(ex):
ui.warn(stringutil.forcebytestr(ex) + b'\n')
if ex.invalidate:
state.invalidate()
# experimental config: fsmonitor.verbose
if ui.configbool(b'fsmonitor', b'verbose'):
ui.log(
b'fsmonitor',
b'Watchman unavailable: %s\n',
stringutil.forcebytestr(ex.msg),
)
else:
ui.log(
b'fsmonitor',
b'Watchman exception: %s\n',
stringutil.forcebytestr(ex),
)
def _hashignore(ignore):
"""Calculate hash for ignore patterns and filenames
If this information changes between Mercurial invocations, we can't
rely on Watchman information anymore and have to re-scan the working
copy.
"""
sha1 = hashutil.sha1()
sha1.update(pycompat.byterepr(ignore))
return pycompat.sysbytes(sha1.hexdigest())
_watchmanencoding = pywatchman.encoding.get_local_encoding()
_fsencoding = sys.getfilesystemencoding() or sys.getdefaultencoding()
_fixencoding = codecs.lookup(_watchmanencoding) != codecs.lookup(_fsencoding)
def _watchmantofsencoding(path):
"""Fix path to match watchman and local filesystem encoding
watchman's paths encoding can differ from filesystem encoding. For example,
on Windows, it's always utf-8.
"""
try:
decoded = path.decode(_watchmanencoding)
except UnicodeDecodeError as e:
raise error.Abort(
stringutil.forcebytestr(e), hint=b'watchman encoding error'
)
try:
encoded = decoded.encode(_fsencoding, 'strict')
except UnicodeEncodeError as e:
raise error.Abort(stringutil.forcebytestr(e))
return encoded
def overridewalk(orig, self, match, subrepos, unknown, ignored, full=True):
'''Replacement for dirstate.walk, hooking into Watchman.
Whenever full is False, ignored is False, and the Watchman client is
available, use Watchman combined with saved state to possibly return only a
subset of files.'''
def bail(reason):
self._ui.debug(b'fsmonitor: fallback to core status, %s\n' % reason)
return orig(match, subrepos, unknown, ignored, full=True)
if full:
return bail(b'full rewalk requested')
if ignored:
return bail(b'listing ignored files')
if not self._watchmanclient.available():
return bail(b'client unavailable')
state = self._fsmonitorstate
clock, ignorehash, notefiles = state.get()
if not clock:
if state.walk_on_invalidate:
return bail(b'no clock')
# Initial NULL clock value, see
# https://facebook.github.io/watchman/docs/clockspec.html
clock = b'c:0:0'
notefiles = []
ignore = self._ignore
dirignore = self._dirignore
if unknown:
if _hashignore(ignore) != ignorehash and clock != b'c:0:0':
# ignore list changed -- can't rely on Watchman state any more
if state.walk_on_invalidate:
return bail(b'ignore rules changed')
notefiles = []
clock = b'c:0:0'
else:
# always ignore
ignore = util.always
dirignore = util.always
matchfn = match.matchfn
matchalways = match.always()
dmap = self._map
if util.safehasattr(dmap, b'_map'):
# for better performance, directly access the inner dirstate map if the
# standard dirstate implementation is in use.
dmap = dmap._map
nonnormalset = self._map.nonnormalset
copymap = self._map.copymap
getkind = stat.S_IFMT
dirkind = stat.S_IFDIR
regkind = stat.S_IFREG
lnkkind = stat.S_IFLNK
join = self._join
normcase = util.normcase
fresh_instance = False
exact = skipstep3 = False
if match.isexact(): # match.exact
exact = True
dirignore = util.always # skip step 2
elif match.prefix(): # match.match, no patterns
skipstep3 = True
if not exact and self._checkcase:
# note that even though we could receive directory entries, we're only
# interested in checking if a file with the same name exists. So only
# normalize files if possible.
normalize = self._normalizefile
skipstep3 = False
else:
normalize = None
# step 1: find all explicit files
results, work, dirsnotfound = self._walkexplicit(match, subrepos)
skipstep3 = skipstep3 and not (work or dirsnotfound)
work = [d for d in work if not dirignore(d[0])]
if not work and (exact or skipstep3):
for s in subrepos:
del results[s]
del results[b'.hg']
return results
# step 2: query Watchman
try:
# Use the user-configured timeout for the query.
# Add a little slack over the top of the user query to allow for
# overheads while transferring the data
self._watchmanclient.settimeout(state.timeout + 0.1)
result = self._watchmanclient.command(
b'query',
{
b'fields': [b'mode', b'mtime', b'size', b'exists', b'name'],
b'since': clock,
b'expression': [
b'not',
[
b'anyof',
[b'dirname', b'.hg'],
[b'name', b'.hg', b'wholename'],
],
],
b'sync_timeout': int(state.timeout * 1000),
b'empty_on_fresh_instance': state.walk_on_invalidate,
},
)
except Exception as ex:
_handleunavailable(self._ui, state, ex)
self._watchmanclient.clearconnection()
return bail(b'exception during run')
else:
# We need to propagate the last observed clock up so that we
# can use it for our next query
state.setlastclock(pycompat.sysbytes(result[b'clock']))
if result[b'is_fresh_instance']:
if state.walk_on_invalidate:
state.invalidate()
return bail(b'fresh instance')
fresh_instance = True
# Ignore any prior noteable files from the state info
notefiles = []
# for file paths which require normalization and we encounter a case
# collision, we store our own foldmap
if normalize:
foldmap = {normcase(k): k for k in results}
switch_slashes = pycompat.ossep == b'\\'
# The order of the results is, strictly speaking, undefined.
# For case changes on a case insensitive filesystem we may receive
# two entries, one with exists=True and another with exists=False.
# The exists=True entries in the same response should be interpreted
# as being happens-after the exists=False entries due to the way that
# Watchman tracks files. We use this property to reconcile deletes
# for name case changes.
for entry in result[b'files']:
fname = entry[b'name']
# Watchman always give us a str. Normalize to bytes on Python 3
# using Watchman's encoding, if needed.
if not isinstance(fname, bytes):
fname = fname.encode(_watchmanencoding)
if _fixencoding:
fname = _watchmantofsencoding(fname)
if switch_slashes:
fname = fname.replace(b'\\', b'/')
if normalize:
normed = normcase(fname)
fname = normalize(fname, True, True)
foldmap[normed] = fname
fmode = entry[b'mode']
fexists = entry[b'exists']
kind = getkind(fmode)
if b'/.hg/' in fname or fname.endswith(b'/.hg'):
return bail(b'nested-repo-detected')
if not fexists:
# if marked as deleted and we don't already have a change
# record, mark it as deleted. If we already have an entry
# for fname then it was either part of walkexplicit or was
# an earlier result that was a case change
if (
fname not in results
and fname in dmap
and (matchalways or matchfn(fname))
):
results[fname] = None
elif kind == dirkind:
if fname in dmap and (matchalways or matchfn(fname)):
results[fname] = None
elif kind == regkind or kind == lnkkind:
if fname in dmap:
if matchalways or matchfn(fname):
results[fname] = entry
elif (matchalways or matchfn(fname)) and not ignore(fname):
results[fname] = entry
elif fname in dmap and (matchalways or matchfn(fname)):
results[fname] = None
# step 3: query notable files we don't already know about
# XXX try not to iterate over the entire dmap
if normalize:
# any notable files that have changed case will already be handled
# above, so just check membership in the foldmap
notefiles = {
normalize(f, True, True)
for f in notefiles
if normcase(f) not in foldmap
}
visit = {
f
for f in notefiles
if (f not in results and matchfn(f) and (f in dmap or not ignore(f)))
}
if not fresh_instance:
if matchalways:
visit.update(f for f in nonnormalset if f not in results)
visit.update(f for f in copymap if f not in results)
else:
visit.update(
f for f in nonnormalset if f not in results and matchfn(f)
)
visit.update(f for f in copymap if f not in results and matchfn(f))
else:
if matchalways:
visit.update(
f for f, st in pycompat.iteritems(dmap) if f not in results
)
visit.update(f for f in copymap if f not in results)
else:
visit.update(
f
for f, st in pycompat.iteritems(dmap)
if f not in results and matchfn(f)
)
visit.update(f for f in copymap if f not in results and matchfn(f))
audit = pathutil.pathauditor(self._root, cached=True).check
auditpass = [f for f in visit if audit(f)]
auditpass.sort()
auditfail = visit.difference(auditpass)
for f in auditfail:
results[f] = None
nf = iter(auditpass)
for st in util.statfiles([join(f) for f in auditpass]):
f = next(nf)
if st or f in dmap:
results[f] = st
for s in subrepos:
del results[s]
del results[b'.hg']
return results
def overridestatus(
orig,
self,
node1=b'.',
node2=None,
match=None,
ignored=False,
clean=False,
unknown=False,
listsubrepos=False,
):
listignored = ignored
listclean = clean
listunknown = unknown
def _cmpsets(l1, l2):
try:
if b'FSMONITOR_LOG_FILE' in encoding.environ:
fn = encoding.environ[b'FSMONITOR_LOG_FILE']
f = open(fn, b'wb')
else:
fn = b'fsmonitorfail.log'
f = self.vfs.open(fn, b'wb')
except (IOError, OSError):
self.ui.warn(_(b'warning: unable to write to %s\n') % fn)
return
try:
for i, (s1, s2) in enumerate(zip(l1, l2)):
if set(s1) != set(s2):
f.write(b'sets at position %d are unequal\n' % i)
f.write(b'watchman returned: %s\n' % s1)
f.write(b'stat returned: %s\n' % s2)
finally:
f.close()
if isinstance(node1, context.changectx):
ctx1 = node1
else:
ctx1 = self[node1]
if isinstance(node2, context.changectx):
ctx2 = node2
else:
ctx2 = self[node2]
working = ctx2.rev() is None
parentworking = working and ctx1 == self[b'.']
match = match or matchmod.always()
# Maybe we can use this opportunity to update Watchman's state.
# Mercurial uses workingcommitctx and/or memctx to represent the part of
# the workingctx that is to be committed. So don't update the state in
# that case.
# HG_PENDING is set in the environment when the dirstate is being updated
# in the middle of a transaction; we must not update our state in that
# case, or we risk forgetting about changes in the working copy.
updatestate = (
parentworking
and match.always()
and not isinstance(ctx2, (context.workingcommitctx, context.memctx))
and b'HG_PENDING' not in encoding.environ
)
try:
if self._fsmonitorstate.walk_on_invalidate:
# Use a short timeout to query the current clock. If that
# takes too long then we assume that the service will be slow
# to answer our query.
# walk_on_invalidate indicates that we prefer to walk the
# tree ourselves because we can ignore portions that Watchman
# cannot and we tend to be faster in the warmer buffer cache
# cases.
self._watchmanclient.settimeout(0.1)
else:
# Give Watchman more time to potentially complete its walk
# and return the initial clock. In this mode we assume that
# the filesystem will be slower than parsing a potentially
# very large Watchman result set.
self._watchmanclient.settimeout(self._fsmonitorstate.timeout + 0.1)
startclock = self._watchmanclient.getcurrentclock()
except Exception as ex:
self._watchmanclient.clearconnection()
_handleunavailable(self.ui, self._fsmonitorstate, ex)
# boo, Watchman failed. bail
return orig(
node1,
node2,
match,
listignored,
listclean,
listunknown,
listsubrepos,
)
if updatestate:
# We need info about unknown files. This may make things slower the
# first time, but whatever.
stateunknown = True
else:
stateunknown = listunknown
if updatestate:
ps = poststatus(startclock)
self.addpostdsstatus(ps)
r = orig(
node1, node2, match, listignored, listclean, stateunknown, listsubrepos
)
modified, added, removed, deleted, unknown, ignored, clean = r
if not listunknown:
unknown = []
# don't do paranoid checks if we're not going to query Watchman anyway
full = listclean or match.traversedir is not None
if self._fsmonitorstate.mode == b'paranoid' and not full:
# run status again and fall back to the old walk this time
self.dirstate._fsmonitordisable = True
# shut the UI up
quiet = self.ui.quiet
self.ui.quiet = True
fout, ferr = self.ui.fout, self.ui.ferr
self.ui.fout = self.ui.ferr = open(os.devnull, b'wb')
try:
rv2 = orig(
node1,
node2,
match,
listignored,
listclean,
listunknown,
listsubrepos,
)
finally:
self.dirstate._fsmonitordisable = False
self.ui.quiet = quiet
self.ui.fout, self.ui.ferr = fout, ferr
# clean isn't tested since it's set to True above
with self.wlock():
_cmpsets(
[modified, added, removed, deleted, unknown, ignored, clean],
rv2,
)
modified, added, removed, deleted, unknown, ignored, clean = rv2
return scmutil.status(
modified, added, removed, deleted, unknown, ignored, clean
)
class poststatus(object):
def __init__(self, startclock):
self._startclock = pycompat.sysbytes(startclock)
def __call__(self, wctx, status):
clock = wctx.repo()._fsmonitorstate.getlastclock() or self._startclock
hashignore = _hashignore(wctx.repo().dirstate._ignore)
notefiles = (
status.modified
+ status.added
+ status.removed
+ status.deleted
+ status.unknown
)
wctx.repo()._fsmonitorstate.set(clock, hashignore, notefiles)
def makedirstate(repo, dirstate):
class fsmonitordirstate(dirstate.__class__):
def _fsmonitorinit(self, repo):
# _fsmonitordisable is used in paranoid mode
self._fsmonitordisable = False
self._fsmonitorstate = repo._fsmonitorstate
self._watchmanclient = repo._watchmanclient
self._repo = weakref.proxy(repo)
def walk(self, *args, **kwargs):
orig = super(fsmonitordirstate, self).walk
if self._fsmonitordisable:
return orig(*args, **kwargs)
return overridewalk(orig, self, *args, **kwargs)
def rebuild(self, *args, **kwargs):
self._fsmonitorstate.invalidate()
return super(fsmonitordirstate, self).rebuild(*args, **kwargs)
def invalidate(self, *args, **kwargs):
self._fsmonitorstate.invalidate()
return super(fsmonitordirstate, self).invalidate(*args, **kwargs)
dirstate.__class__ = fsmonitordirstate
dirstate._fsmonitorinit(repo)
def wrapdirstate(orig, self):
ds = orig(self)
# only override the dirstate when Watchman is available for the repo
if util.safehasattr(self, b'_fsmonitorstate'):
makedirstate(self, ds)
return ds
def extsetup(ui):
extensions.wrapfilecache(
localrepo.localrepository, b'dirstate', wrapdirstate
)
if pycompat.isdarwin:
# An assist for avoiding the dangling-symlink fsevents bug
extensions.wrapfunction(os, b'symlink', wrapsymlink)
extensions.wrapfunction(merge, b'_update', wrapupdate)
def wrapsymlink(orig, source, link_name):
''' if we create a dangling symlink, also touch the parent dir
to encourage fsevents notifications to work more correctly '''
try:
return orig(source, link_name)
finally:
try:
os.utime(os.path.dirname(link_name), None)
except OSError:
pass
class state_update(object):
''' This context manager is responsible for dispatching the state-enter
and state-leave signals to the watchman service. The enter and leave
methods can be invoked manually (for scenarios where context manager
semantics are not possible). If parameters oldnode and newnode are None,
they will be populated based on current working copy in enter and
leave, respectively. Similarly, if the distance is none, it will be
calculated based on the oldnode and newnode in the leave method.'''
def __init__(
self,
repo,
name,
oldnode=None,
newnode=None,
distance=None,
partial=False,
):
self.repo = repo.unfiltered()
self.name = name
self.oldnode = oldnode
self.newnode = newnode
self.distance = distance
self.partial = partial
self._lock = None
self.need_leave = False
def __enter__(self):
self.enter()
def enter(self):
# Make sure we have a wlock prior to sending notifications to watchman.
# We don't want to race with other actors. In the update case,
# merge.update is going to take the wlock almost immediately. We are
# effectively extending the lock around several short sanity checks.
if self.oldnode is None:
self.oldnode = self.repo[b'.'].node()
if self.repo.currentwlock() is None:
if util.safehasattr(self.repo, b'wlocknostateupdate'):
self._lock = self.repo.wlocknostateupdate()
else:
self._lock = self.repo.wlock()
self.need_leave = self._state(b'state-enter', hex(self.oldnode))
return self
def __exit__(self, type_, value, tb):
abort = True if type_ else False
self.exit(abort=abort)
def exit(self, abort=False):
try:
if self.need_leave:
status = b'failed' if abort else b'ok'
if self.newnode is None:
self.newnode = self.repo[b'.'].node()
if self.distance is None:
self.distance = calcdistance(
self.repo, self.oldnode, self.newnode
)
self._state(b'state-leave', hex(self.newnode), status=status)
finally:
self.need_leave = False
if self._lock:
self._lock.release()
def _state(self, cmd, commithash, status=b'ok'):
if not util.safehasattr(self.repo, b'_watchmanclient'):
return False
try:
self.repo._watchmanclient.command(
cmd,
{
b'name': self.name,
b'metadata': {
# the target revision
b'rev': commithash,
# approximate number of commits between current and target
b'distance': self.distance if self.distance else 0,
# success/failure (only really meaningful for state-leave)
b'status': status,
# whether the working copy parent is changing
b'partial': self.partial,
},
},
)
return True
except Exception as e:
# Swallow any errors; fire and forget
self.repo.ui.log(
b'watchman', b'Exception %s while running %s\n', e, cmd
)
return False
# Estimate the distance between two nodes
def calcdistance(repo, oldnode, newnode):
anc = repo.changelog.ancestor(oldnode, newnode)
ancrev = repo[anc].rev()
distance = abs(repo[oldnode].rev() - ancrev) + abs(
repo[newnode].rev() - ancrev
)
return distance
# Bracket working copy updates with calls to the watchman state-enter
# and state-leave commands. This allows clients to perform more intelligent
# settling during bulk file change scenarios
# https://facebook.github.io/watchman/docs/cmd/subscribe.html#advanced-settling
def wrapupdate(
orig,
repo,
node,
branchmerge,
force,
ancestor=None,
mergeancestor=False,
labels=None,
matcher=None,
**kwargs
):
distance = 0
partial = True
oldnode = repo[b'.'].node()
newnode = repo[node].node()
if matcher is None or matcher.always():
partial = False
distance = calcdistance(repo.unfiltered(), oldnode, newnode)
with state_update(
repo,
name=b"hg.update",
oldnode=oldnode,
newnode=newnode,
distance=distance,
partial=partial,
):
return orig(
repo,
node,
branchmerge,
force,
ancestor,
mergeancestor,
labels,
matcher,
**kwargs
)
def repo_has_depth_one_nested_repo(repo):
for f in repo.wvfs.listdir():
if os.path.isdir(os.path.join(repo.root, f, b'.hg')):
msg = b'fsmonitor: sub-repository %r detected, fsmonitor disabled\n'
repo.ui.debug(msg % f)
return True
return False
def reposetup(ui, repo):
# We don't work with largefiles or inotify
exts = extensions.enabled()
for ext in _blacklist:
if ext in exts:
ui.warn(
_(
b'The fsmonitor extension is incompatible with the %s '
b'extension and has been disabled.\n'
)
% ext
)
return
if repo.local():
# We don't work with subrepos either.
#
# if repo[None].substate can cause a dirstate parse, which is too
# slow. Instead, look for a file called hgsubstate,
if repo.wvfs.exists(b'.hgsubstate') or repo.wvfs.exists(b'.hgsub'):
return
if repo_has_depth_one_nested_repo(repo):
return
fsmonitorstate = state.state(repo)
if fsmonitorstate.mode == b'off':
return
try:
client = watchmanclient.client(repo.ui, repo.root)
except Exception as ex:
_handleunavailable(ui, fsmonitorstate, ex)
return
repo._fsmonitorstate = fsmonitorstate
repo._watchmanclient = client
dirstate, cached = localrepo.isfilecached(repo, b'dirstate')
if cached:
# at this point since fsmonitorstate wasn't present,
# repo.dirstate is not a fsmonitordirstate
makedirstate(repo, dirstate)
class fsmonitorrepo(repo.__class__):
def status(self, *args, **kwargs):
orig = super(fsmonitorrepo, self).status
return overridestatus(orig, self, *args, **kwargs)
def wlocknostateupdate(self, *args, **kwargs):
return super(fsmonitorrepo, self).wlock(*args, **kwargs)
def wlock(self, *args, **kwargs):
l = super(fsmonitorrepo, self).wlock(*args, **kwargs)
if not ui.configbool(
b"experimental", b"fsmonitor.transaction_notify"
):
return l
if l.held != 1:
return l
origrelease = l.releasefn
def staterelease():
if origrelease:
origrelease()
if l.stateupdate:
l.stateupdate.exit()
l.stateupdate = None
try:
l.stateupdate = None
l.stateupdate = state_update(self, name=b"hg.transaction")
l.stateupdate.enter()
l.releasefn = staterelease
except Exception as e:
# Swallow any errors; fire and forget
self.ui.log(
b'watchman', b'Exception in state update %s\n', e
)
return l
repo.__class__ = fsmonitorrepo