struct sd_rtnl_message would keep two additional pointers into the hdr
field. Every time hdr was realloced, those pointers should be adjusted,
but weren't. It seems less error-prone to keep offsets instead.
Because messages in the read and write queues hold a reference on the
bus, and the bus holds a reference on each of them, we would never
free the bus if the read or write queues were not empty. Explicitly
substract the number of messages in those queue from the bus reference
count when deciding whether to free or not.
A simple test which creates and unrefs simple objects is added.
It is nicer to predefine patterns using configure time check instead of
using casts everywhere.
Since we do not need to use any flags, include "%" in the format instead
of excluding it like PRI* macros.
Information about signals which are not routinely received by systemd
are printed at info level. This should make it easier to see what is
happening in the system.
This adds basic DHCPv4 support. Link-sense is enabled unconditionally,
but the plan is to make that configurable.
I tested this in a VM with lots of NICs and over wifi in the various
coffee shops I found this Christmas, but more testing would definitely
be appreciated.
Static addresses/routes are associated with a network. Dynamic
addresses/routes are associtade with links (as the corresponding network
may be shared by several links).
GREEDY_REALLOC needs to have two size variables: one for the
allocated size, and a second one for the used size. Using
the allocated size only lead to leaving some elements unitialized
and assigning some more than once.
Before 34a3baa4d 'sleep-config: Dereference pointer before check for NULL'
oom conditions would not be detected properly. After that commit, a double
free was performed.
Rework the whole function to be easier to understand, and also replace
strv_split_nulstr with strv_new, since we know the strings anyway.
This fixes a bug pointed out by http://css.csail.mit.edu/stack/
(Optimization-unstable code)
It is a similar fix as f146f5e159 (2013-12-30, core:
Forgot to dereference pointer when checking for NULL)
To explain this bug consider the following similar, but simpler code:
if (!p)
free(*p)
Assume the if condition evaluates to true, then we will access *p,
which means the compiler can assume p is a valid pointer, so it could
dereference p and use the value *p.
Assuming p as a valid pointer, !p will be false.
But initally we assumed the condition evaluates to true.
By this reasoning the optimizing compiler can deduce, we have dead code.
("The if will never be taken, as *p must be valid, because otherwise
accessing *p inside the if would segfault")
This led to an error message of the static code checker, so I checked the
code in question.
As we access *modes and *states before the check in the changed line of
this patch, I assume the line to be wrong and we actually wanted to check
for *modes and *states being both non null.
Directly before the changed line there is:
while ((parent->next_ps && parent->pid != ps->ppid))
parent = parent->next_ps;
which looks one element ahead of the list, hence we can rely on parent
being non null here.
If 'parent' were NULL at that while loop already, it would crash as we're
dereferencing 'parent' when checking for next_ps already.
Signed-off-by: Stefan Beller <stefanbeller@googlemail.com>
The return value of 0 would be treated as failure by mistake,
resulting in " SystemError: error return without exception set".
The way that set_error() is used is changed to be the same
everywhere.
Nspawn has --setenv, and systemd itself accepts systemd.setenv.
It is nice to have the same parameter name everywhere.
Old name is accepted, but not advertised.
scan-build is a static analyzer in llvm. As ususal static analyzers
tend to mostly find theoretical bugs in software that has been in
production for a while. For in-development code it can be useful to
check if new issues is added as there is a chance to spot real problems
before release. For systemd we are now down to 297 issues - the vast
majority are false positives because the tool does not understand the
cleanup attribute.
Running clang's static analyzer scan-build is a bit messy. You have to
run both configure and make "inside" the build-scan tool. To have an
easy shortcut from autogen.sh I thus call both directly from it. This
makes it different from the other options in autogen.sh. I chose 's'
for static analysis.
scan-build is in the package clang-analyzer on fedora.
For a while I have been cleaning up warnings when building with clang.
There are currently only two sources of warnings left: Wcast-align and Wgnu.
I am not convinced that fixing up those up is feasible so I run with them
disabled to spot regressions. E.g. clang is a bit more strict wrt to unused
variables with the cleanup attribute and I have fixed a number of those since.
Like the other options in autogen.sh I have a shortcut for clang as well. I use
'l' for llvm.
