3 This directory (nightly/) contains a simple, automatic build-and-test
4 system for Valgrind, intended to be run nightly by cron or a similar
10 When run, the system clones and checks out two trees: one from 24 hours ago
11 and the HEAD from now. ("24 hours ago" and "now" are determined when
12 the script starts running, so if any commits happen while the tests are
13 running they will not be tested.)
15 If the two trees are different (i.e. there have been commits in the past 24
16 hours, either to the master or a branch) it builds ("make"), installs ("make
17 install") and runs the regression tests ("make regtest") in both, and
18 compares the results. Note that the "make install" isn't necessary in order
19 to run the tests because the regression tests use the code built (with
20 "make") within the tree, but it's worth doing because it tests that "make
21 install" isn't totally broken. After checking both trees, it emails a
22 summary of the results to a recipient. All this typically takes something
25 If the two trees are identical, the tests are not run and no results are
26 emailed. This avoids spamming people with uninteresting results emails when
27 no commits have happened recently.
29 The requirements for the machine running the nightly tests are the same as
30 for running the regression tests (see README_DEVELOPERS) plus GNU coreutils
31 and the ability to send e-mail. The mail transport agent used depends on the
32 platform and is decribed in step (4) below.
37 To set up nightly testing for a machine, do the following.
39 (1) Clone a shallow copy from the repository, eg:
41 git clone --depth 1 https://sourceware.org/git/valgrind.git/ $DIR
43 where $DIR is the name of the directory you want it to be in.
44 You'll probably want to include "nightly" in the directory name.
46 Note that this doesn't clone the whole repo history, just the latest
47 revision of everything in the repository. Under $DIR, only directory
48 "nightly" is of interest to us - the rest is ignored. In other words: the
49 testing script clones and checks out new trees (within $DIR/nightly) and
50 tests them independently.
52 (2) Choose a tag that identifies the test results. This is usually the
53 machine name. We'll call it $TAG in what follows.
55 (3) Create a configuration file $DIR/nightly/conf/$TAG.conf. It is sourced
56 by the 'nightly' script, and can define any or all of the following
57 environment variables. (In most cases, only ABT_DETAILS is needed.)
59 - ABT_DETAILS: describes the machine in more detail, eg. the OS. The
60 default is empty, but you should define it. An example:
62 export ABT_DETAILS="Ubuntu 9.04, Intel x86-64"
64 You could also use some invocation of 'uname' or something similar
65 to generate this string. Eg. on Ubuntu Linux this works nicely:
67 export ABT_DETAILS="`cat /etc/issue.net`, `uname -m`"
69 And on Mac OS X this works nicely:
71 export ABT_DETAILS=`uname -mrs`
73 The advantage of doing it like this is that if you update the OS on
74 the test machine you won't have to update ABT_DETAILS manually.
76 - ABT_CONFIGURE_OPTIONS: gives extra configure options. The default is
79 - ABT_EVAL: if provided, it must be the name of a shell script that
80 executes the shell command $1 with arguments $2 .. ${$#}. Allows to
81 compile and run the Valgrind regression tests on another system than
82 the system the 'nightly' script runs on. It is assumed that the remote
83 system shares the local filesystem tree through e.g. NFS. It is the
84 responsibility of the shell script to set the remote working directory
85 such that it matches the local current directory ($PWD).
87 - ABT_RUN_REGTEST: if provided, it must be the name of an argumentless
88 shell function (also specified in the $TAG.conf file) that will be used
89 to run the tests. If not specified, the usual "make regtest" will be
92 - ABT_JOBS: allows parallel builds -- it's passed as the argument to
93 "make -j" when building Valgrind and the tests. The default is 1.
95 - ABT_PERF: unset or set to "" mean 'do not run perf tests' (default value)
96 set to "--vg=../valgrind-new" (run perf tests for new tree)
97 set to "--vg=../valgrind-new --vg=../valgrind-old"
98 (run perf tests for "new" and for "24 hours ago",
99 to compare the performances between the 2 trees).
101 - ABT_PERF_TOOLS: --tools=.... option of perf/vg_perf.
102 (default value: all non experimental tools)
104 - ABT_PERF_REPS: --reps=... option of perf/vg_perf
105 (default value: --reps=3)
107 Note that the appropriate syntax to use in this file will depend on the
108 shell from which the $DIR/nightly/bin/nightly script is run (which in turn
109 may depend on what shell is used by cron or any similar program).
111 (4) Create a mailer script $DIR/nightly/conf/$TAG.sendmail. It must be
112 executable. It's used to send email results to the desired recipient (e.g.
113 valgrind-developers@lists.sourceforge.net). It must handle three command
116 - The first argument is the email subject line. It contains
117 $ABT_DETAILS plus some other stuff.
119 - The second argument is the name of the file containing the email's
120 body (which shows the tests that failed, and the differences between now
123 - The third is the name of the file containing all the diffs from
124 failing tests. Depending on the test results you get, you could
125 inline this file into the email body, or attach it, or compress and
126 attach it, or even omit it. The right choice depends on how many
127 failures you typically get -- if you get few failures, inlining the
128 results make them easier to read; if you get many failures,
129 compressing might be a good idea to minimise the size of the emails.
131 The best way to do this depends on how mail is set up on your machine.
132 You might be able to use /usr/bin/mail, or you might need something more
133 elaborate like using Mutt to send mail via an external account.
135 At first, you should probably just send emails to yourself for testing
136 purposes. After it's working, then sending it to others might be
139 (5) To run the tests, execute:
141 $DIR/nightly/bin/nightly $DIR/nightly $TAG
143 You probably want to put this command into a cron file or equivalent
144 so it is run regularly (preferably every night). Actually, it's
145 probably better to put that command inside a script, and run the script
146 from cron, rather than running $DIR/nightly/bin/nightly directly. That way
147 you can put any other configuration stuff that's necessary inside the
148 script (e.g. make sure that programs used by the mailer script are in
154 If the tests are run, the following files are produced:
156 - $DIR/nightly/old.verbose and $DIR/nightly/new.verbose contain full output of
157 the whole process for each of the two trees.
159 - $DIR/nightly/old.short and $DIR/nightly/new.short contain summary output of
160 the process for each of the two trees. The diff between these two files goes
161 in $DIR/nightly/diff.short.
163 - $DIR/nighlty/final contains the overall summary, constructed from
164 $DIR/nightly/old.short, $DIR/nightly/new.short, $DIR/nightly/diff.short and
165 some other bits and pieces. (The name of this file is what's passed as the
166 second argument to $DIR/nightly/conf/$TAG.sendmail.)
168 - $DIR/nightly/diffs holds the diffs from all the failing tests in the newer
169 tree, concatenated together; the diff from each failure is truncated at 100
170 lines to minimise possible size blow-outs. (The name of this file is
171 what's passed as the third argument to $DIR/nightly/conf/$TAG.sendmail.)
173 - $DIR/nightly/sendmail.log contains the output (stdout and stderr) from
174 $DIR/nightly/conf/$TAG.sendmail goes in $DIR/nightly/sendmail.log.
176 - $DIR/nightly/valgrind-old/ and $DIR/nightly/valgrind-new/ contain the tested
177 trees (and $DIR/nightly/valgrind-old/Inst/ and $DIR/nightly/valgrind-new/Inst/
178 contain the installed code).
180 If the tests aren't run, the following file is produced:
182 - $DIR/nightly/unchanged.log is created only if no tests were run because the
183 two trees were identical. It will contain a short explanatory message.
185 Each time the tests are run, all files from previous runs are deleted.
190 If something goes wrong, looking at the output files can be useful. For
191 example, if no email was sent but you expected one, check sendmail.log to
192 see if the mailer script had a problem. Or check if unchanged.log exists.
194 Occasionally the git server isn't available when the tests runs, for either
195 or both trees. When this happens the email will be sent but it won't be
196 very informative. Usually it's just a temporary server problem and it'll
197 run fine the next time without you having to do anything.
199 Note that the test suite is imperfect:
200 - There are very few machines where all tests pass; that's why the old/new
201 diff is produced. Some of the tests may not be as portable as intended.
202 - Some tests are non-deterministic, and so may pass one day and fail the
205 Improving the test suite to avoid these problems is a long-term goal but it
211 The scripts in the nightly/ directory occasionally get updated. If that
212 happens, you can just "git pull" within $DIR to get the updated versions,
213 which will then be used the next time the tests run. (It's possible that
214 the scripts will be changed in a way that requires changes to the files in
215 $DIR/nightly/conf/, but we try to avoid this.)
217 If you want such updates to happen automatically, you could write a script
218 that does all the steps in SETTING UP above, and instead run that script