Without this "meson test" will end up running all tests in the same
cgroup root, and they all will try to manage it. Which usually isn't too
bad, except when they end up clearing up each other's cgroups. This race
is hard to trigger but has caused various CI runs to fail spuriously.
With this change we simply move every test that runs a manager object
into their own private cgroup. Note that we don't clean up the cgroup at
the end, we leave that to the cgroup manager around it.
This fixes races that become visible by test runs throwing out errors
like this:
```
exec-systemcallfilter-failing.service: Passing 0 fds to service
exec-systemcallfilter-failing.service: About to execute: /bin/echo 'This should not be seen'
exec-systemcallfilter-failing.service: Forked /bin/echo as 5693
exec-systemcallfilter-failing.service: Changed dead -> start
exec-systemcallfilter-failing.service: Failed to attach to cgroup /exec-systemcallfilter-failing.service: No such file or directory
Received SIGCHLD from PID 5693 ((echo)).
Child 5693 ((echo)) died (code=exited, status=219/CGROUP)
exec-systemcallfilter-failing.service: Child 5693 belongs to exec-systemcallfilter-failing.service
exec-systemcallfilter-failing.service: Main process exited, code=exited, status=219/CGROUP
exec-systemcallfilter-failing.service: Changed start -> failed
exec-systemcallfilter-failing.service: Unit entered failed state.
exec-systemcallfilter-failing.service: Failed with result 'exit-code'.
exec-systemcallfilter-failing.service: cgroup is empty
Assertion 'service->main_exec_status.status == status_expected' failed at ../src/src/test/test-execute.c:71, function check(). Aborting.
```
BTW, I tracked this race down by using perf:
```
# perf record -e cgroup:cgroup_mkdir,cgroup_rmdir
…
# perf script
```
Thanks a lot @iaguis, @alban for helping me how to use perf for this.
Fixes#5895.
If an error is encountered in any of the Exec* lines, WorkingDirectory,
SELinuxContext, ApparmorProfile, SmackProcessLabel, Service (in .socket
units), User, or Group, refuse to load the unit. If the config stanza
has support, ignore the failure if '-' is present.
For those configuration directives, even if we started the unit, it's
pretty likely that it'll do something unexpected (like write files
in a wrong place, or with a wrong context, or run with wrong permissions,
etc). It seems better to refuse to start the unit and have the admin
clean up the configuration without giving the service a chance to mess
up stuff.
Note that all "security" options that restrict what the unit can do
(Capabilities, AmbientCapabilities, Restrict*, SystemCallFilter, Limit*,
PrivateDevices, Protect*, etc) are _not_ treated like this. Such options are
only supplementary, and are not always available depending on the architecture
and compilation options, so unit authors have to make sure that the service
runs correctly without them anyway.
Fixes#6237, #6277.
According to its manual page, flags given to mkostemp(3) shouldn't include
O_RDWR, O_CREAT or O_EXCL flags as these are always included. Beyond
those, the only flag that all callers (except a few tests where it
probably doesn't matter) use is O_CLOEXEC, so set that unconditionally.
We don#t really support systems where XDG_RUNTIME_DIR is not supported for
systemd --user. Hence, let's always set our own XDG_RUNTIME_DIR for tests that
involve systemd --user, so that we know it is set, and that it doesn't polute
the user's actual runtime dir.
Previously, we had two enums ManagerRunningAs and UnitFileScope, that were
mostly identical and converted from one to the other all the time. The latter
had one more value UNIT_FILE_GLOBAL however.
Let's simplify things, and remove ManagerRunningAs and replace it by
UnitFileScope everywhere, thus making the translation unnecessary. Introduce
two new macros MANAGER_IS_SYSTEM() and MANAGER_IS_USER() to simplify checking
if we are running in one or the user context.
Change the capability bounding set parser and logic so that the bounding
set is kept as a positive set internally. This means that the set
reflects those capabilities that we want to keep instead of drop.
The new parser supports:
<value> - specify both limits to the same value
<soft:hard> - specify both limits
the size or time specific suffixes are supported, for example
LimitRTTIME=1sec
LimitAS=4G:16G
The patch introduces parse_rlimit_range() and rlim type (size, sec,
usec, etc.) specific parsers. No code is duplicated now.
The patch also sync docs for DefaultLimitXXX= and LimitXXX=.
References: https://github.com/systemd/systemd/issues/1769
Previously, the %u, %U, %s and %h specifiers would resolve to the user
name, numeric user ID, shell and home directory of the user configured
in the User= setting of a unit file, or the user of the manager instance
if no User= setting was configured. That at least was the theory. In
real-life this was not ever actually useful:
- For the systemd --user instance it made no sense to ever set User=,
since the instance runs in user context after all, and hence the
privileges to change user IDs don't even exist. The four specifiers
were actually not useful at all in this case.
- For the systemd --system instance we did not allow any resolving that
would require NSS. Hence, %s and %h were not supported, unless
User=root was set, in which case they would be hardcoded to /bin/sh
and /root, to avoid NSS. Then, %u would actually resolve to whatever
was set with User=, but %U would only resolve to the numeric UID of
that setting if the User= was specified in numeric form, or happened
to be root (in which case 0 was hardcoded as mapping). Two of the
specifiers are entirely useless in this case, one is realistically
also useless, and one is pretty pointless.
- Resolving of these settings would only happen if User= was actually
set *before* the specifiers where resolved. This behaviour was
undocumented and is really ugly, as specifiers should actually be
considered something that applies to the whole file equally,
independently of order...
With this change, %u, %U, %s and %h are drastically simplified: they now
always refer to the user that is running the service instance, and the
user configured in the unit file is irrelevant. For the system instance
of systemd this means they always resolve to "root", "0", "/bin/sh" and
"/root", thus avoiding NSS. For the user instance, to the data for the
specific user.
The new behaviour is identical to the old behaviour in all --user cases
and for all units that have no User= set (or set to "0" or "root").
Let's make sure "LimitCPU=30min" can be parsed properly, following the
usual logic how we parse time values. Similar for LimitRTTIME=.
While we are at it, extend a bit on the man page section about resource
limits.
Fixes: #1772
There are more than enough calls doing string manipulations to deserve
its own files, hence do something about it.
This patch also sorts the #include blocks of all files that needed to be
updated, according to the sorting suggestions from CODING_STYLE. Since
pretty much every file needs our string manipulation functions this
effectively means that most files have sorted #include blocks now.
Also touches a few unrelated include files.
This is consistent with how an empty string works in an ExecStart=
statement. We should not differentiate between an empty string and
whitespace only (since they look the same.)
Update the test case with whitespace only to reflect that the list is
reset.
Tested that `test-unit-file` passes and other test cases are not
affected. Installed the patched systemd binaries on a machine, booted
it, looked for out of the usual behavior but did not find any.
Convert config_parse_exec() from using FOREACH_WORD_QUOTED into a loop
of unquote_first_word.
Loop through the arguments only once (the FOREACH_WORD_QUOTED
implementation did it twice, once to count them and another time to
process and store them.)
Use newly introduced flag UNQUOTE_UNESCAPE_RELAX to preserve
unrecognized escape sequences such as regexps matches such as "\w",
"\d", etc. (Valid escape sequences such as "\s" or "\b" still need an
extra backslash if literals are desired for regexps.)
Differences in behavior:
- Handle ; (command separator) in special, so that only ; on its own is
valid for that purpose, an quoted semicolon ";" or ';' will now behave
as a literal semicolon. This is probably what was initially intended.
- Handle \; (to introduce a literal semicolon) in special, so that only \;
is turned into a semicolon but not \\; or "\\;" or "\;" which are kept
as a literal \; in the output. This is probably what was initially
intended.
Known issues:
- Using an empty string (for example, ExecStartPre=<empty>) will empty
the list and remove the existing commands, but using whitespace only
(for example, ExecStartPre=<spaces>) will not. This is a pre-existing
issue and will be dealt with in a follow up commit.
Tested:
- Unit tests passing. Also `make distcheck` still works as expected.
- Installed it on a local machine and booted with it, checked console
output, systemctl and journalctl output, did not notice any issues
running the patched systemd binaries.
Relevant bug: https://bugs.freedesktop.org/show_bug.cgi?id=90794
These tests will be useful to check the cases regarding quoted and
escaped semicolon when we switch to using unquote_first_word.
Additionally, convert some of the tests that have semicolons so that the
argument after the semicolon looks like a path (starts with /) so that
we can see the change of behavior when making config_parse_exec more
strict about what it accepts as a command separator.
An Exec*= line with whitespace after modifiers, like
ExecStart=- /bin/true
is considered to have an empty command path. This is as specified, but causes
systemd to crash with
Assertion 'skip < l' failed at ../src/core/load-fragment.c:607, function config_parse_exec(). Aborting.
Aborted (core dumped)
Fix this by logging an error instead and ignoring the invalid line.
Add corresponding test cases. Also add a test case for a completely empty value
which resets the command list.
https://launchpad.net/bugs/1454173
All other types exported from install.h should be namespaces like this,
hence namespace InstallInfo the same way.
Also, remove external forward definition of UnitFileScope type.
When parsing a unit with a trailing slash after an escaped line break, like
ExecStart=/bin/echo 'foo \
bar'
the split() function (through config_parse()) asserted and crashed pid 1:
Assertion 'current[*l + 1] == quotechars[0]' failed at ../src/shared/util.c:583, function split(). Aborting.
Fix this by returning an error in this case ("trailing garbage").
Add corresponding test case. Also fix the missing "unit" argument of
config_parse_exec() in the comment.
https://launchpad.net/bugs/1447243
This patch removes includes that are not used. The removals were found with
include-what-you-use which checks if any of the symbols from a header is
in use.
The handling of the command name and other arguments is unified. This
simplifies things and should make them more predictable for users.
Incidentally, this makes ExecStart handling match the .desktop file
specification, apart for the requirment for an absolute path.
https://bugs.freedesktop.org/show_bug.cgi?id=86171
It is redundant to store 'hash' and 'compare' function pointers in
struct Hashmap separately. The functions always comprise a pair.
Store a single pointer to struct hash_ops instead.
systemd keeps hundreds of hashmaps, so this saves a little bit of
memory.
Pass on the line on which a section was decleared to the parsers, so they
can distinguish between multiple sections (if they chose to). Currently
no parsers take advantage of this, but a follow-up patch will do that
to distinguish
[Address]
Address=192.168.0.1/24
Label=one
[Address]
Address=192.168.0.2/24
Label=two
from
[Address]
Address=192.168.0.1/24
Label=one
Address=192.168.0.2/24
Label=two
