Typically, in reasonably complex programs we want to realease various
caches (such as glibc allocation caches) in case of memory pressure.
Let's add explicit infrastructure for that to sd-event, that can hook
Linux' Pressure Stall Information (PSI) logic with our event loop.
This adds sd_event_add_memory_pressure() as easy, one-step API to
install an even source that is called under memory pressure.
The parameters which file to watch (the per-cgroup PSI file, or the
system-wide file /proc/pressure/memory) can be configured via env vars.
The idea is that the service manager sooner or later gains controls for
setting this up correctly.
Alternatively to the PSI a similar logic is supported but instead of
waiting for POLLPRI on a procfs/cgroupfs fd we'll wait for POLLIN on
FIFO or AF_UNIX sockets. This is useful for testing, and possibly in
other environments, for example to hook up this protocol directly with
GNOME's low memory monitor.
By default this watches on the cgroup-local PSI so that we aren't
affected by pressure on cgroups we are not related to.
Most of the support for valgrind was under HAVE_VALGRIND_VALGRIND_H, i.e. we
would enable if the valgrind headers were found. The operations then we be
conditionalized on RUNNING_UNDER_VALGRIND.
But in a few places we had code which was conditionalized on VALGRIND, i.e. the
config option. I noticed because I compiled with -Dvalgrind=true on a machine
that didn't have valgrind.h, and the build failed because
RUNNING_UNDER_VALGRIND was not defined. My first idea was to add a check that
the header is present if the option is set, but it seems better to just remove
the option. The code to support valgrind is trivial, and if we're
!RUNNING_UNDER_VALGRIND, it has negligible cost. And the case of running under
valgrind is always some special testing/debugging mode, so we should just do
those extra steps to make valgrind output cleaner. Removing the option makes
things simpler and we don't have to think if something should be covered by the
one or the other configuration bit.
I had a vague recollection that in some places we used -Dvalgrind=true not
for valgrind support, but to enable additional cleanup under other sanitizers.
But that code would fail to build without the valgrind headers anyway, so
I'm not sure if that was still used. If there are uses like that, we can
extend the condition for cleanup_pools().
We should be more careful with distinguishing the cases "all bits set in
caps mask" from "cap mask invalid". We so far mostly used UINT64_MAX for
both, which is not correct though (as it would mean
AmbientCapabilities=~0 followed by AmbientCapabilities=0) would result
in capability 63 to be set (which we don't really allow, since that
means unset).
The rest of our codebase stores caps masks in a uint64_t, and also
assumes UINT64_MAX was a suitable value for "unset mask". Hence refuse
any caps outside of 0…62.
(right now the kernel knows 40 caps, hence 22 more to go before we have
to reconsider our life's choices.)
We refuse it otherwise currently, simply because we cannot store it in a
uint64_t caps mask value anymore while retaining the ability to use
UINT64_MAX as "unset" marker.
The check actually was in place already, just one off.
This substantially reworks mempool_cleanup() so that it releases pools
with all freed tiles only, but keeps all pools with still-allocated
tiles around.
This is more correct, as the previous implementation just released all
pools regardless if anything was still used or not. This would make
valgrind shut up but would just hide memory leaks altogether. Moreover
if called during regular runtime of a program would result in bad memory
accesses all over.
Hence, let's add a proper implementation and only trim pools we really
know are empty.
This way we can safely call these functions later, when under memory
pressure, at any time.
To match how we usually do this current allocation code.
(Also, make it accept a NULL pointer, also in order to match behaviour
in the rest of our codebase)
Follow-up for 4453ebe4db.
With the feature list all dandified, the most important part of the
output, i.e. the project name and version, are less visible.
Yu Watanabe
Lennart Poettering
Zbigniew Jędrzejewski-Szmek
Daan De Meyer
Luca Boccassi
Dmitry V. Levin