Don't simply continue after sleeping, it potentially puts us in a state
of spinning doing nothing slowly, if the ratelimit_test() keeps
detecting the need for limiting.
Observed in vms after the host had been suspended for a while, on resume
systemd entered a loop of making zero progress spamming the console
with:
[431942.850090] systemd[1]: Looping too fast. Throttling execution a
little.
I see no reason to have a continue here, the intention should be to
throttle execution, not circumvent it altogether.
Some guests (ARM, AArch64, x86-RHEL) have 'KVM' in the product name.
Look for that first in order to more precisely report "kvm" when
detecting a QEMU/KVM guest. Without this patch we report "qemu",
even if KVM acceleration is in use on ARM/AArch64 guests.
I've only tested a backported version of this and the previous
patch on an AArch64 guest (which worked). Of course it would be
nice to get regression testing on all guest types that depend on
dmi done.
ARM/AArch64 guests now have SMBIOS tables populated (when boot
with a late enough QEMU and a late enough AAVMF is used as the
bootloader). Furthermore, when booting ARM/AArch64 guests with
ACPI, the DT detection obviously no longer works, so we need
dmi detection.
afaict, this will fix a regression caused by commit 75f86906c5.
Where we used to report "kvm" before that patch, without this patch,
we would only report "qemu". The reason is because cpuid detection
must come before dmi detection. Also, both can safely come before
other xen heuristics. Untested.
We already filter out 0, and as -1 is usually special (meaning infinity,
as in USEC_INFINITY) we should better not accept it either. Better safe
than sorry...
It is really unclear if we want to / have the resources to support this fully, so drop it
for now. It can easily be brought back if a killer usecase emerges.
Note that this code was never hooked up, so this does not remove any features.
This is useful to check that compression actually works, and how
compression influences file size in the best-case-scenario for
compression. (The answer is that not as much as one would hope:
there's still a big overhead of the indexing and since every field
is compressed separately, even fields that compress very well
contribute to the file size. This overhead becomes negligible only
for very big fields.)
If we failed to extract a word, then that's how it is, we shouldn't try
to extract it again and again, it's unlikely to work, and we just
deadlock.
This is a fix-up for 52278ad31d.
When we enumerate journal files and encounter an invalid one, remember
which this, and show it to the user.
Note the possibly slightly surprising logic here: we store only one path
per error code. This means we show all error kinds but not every actual
error we encounter. This has the benefit of not requiring us to keep a
potentially unbounded list of errors with their sources around, but can
still provide a pretty complete overview on the errors we encountered.
Fixes#1669.
- Always print a debug log message about files and directories we cannot
open right when it happens instead of the caller, thus reducing the
number of places where we need to generate the debug message.
- Always push the errors we encounter immediately into the error set,
when we run into them, instead of in the caller. Thus, we never forget
to push them in.
- Use stack instead of heap memory where we can.
- Make remove_file() void, since it cannot fail anyway and always
returned 0.
- Make local machine check of journal directories explicit in a
function, to make things more readable.
- Port to all directory listing loops FOREACH_DIRENT_ALL()
- sd-daemon is library code, hence never log at higher log levels than
LOG_DEBUG.
When reading stuff, we should only return EIO when an actual read error
occured, not when we don't like the data for whatever reason.
We already return ENODATA for all other kinds of file truncation, hence
do the same for the most obvious kind, so that callers know what ENODATA
means.
For each service expose how many file descriptors there are currently in
the fd store.
(Also, fix the exporting of the fdstore limit, given that the field is
just an "unsigned" but we exported it as "uint32_t". Not that there way
any effective difference, but let's make this clean...)
memory_erase() so far just called memset(), which the compiler might
optimize away under certain conditions if it feels there's benefit in
it. C11 knows a new memset_s() call that is like memset(), but may not
be optimized away. Ideally, we'd just use that call, but glibc currently
does not support it. Hence, implement our own simplistic version of it.
We use a GCC pragma to turn off optimization for this call, and also use
the "volatile" keyword on the pointers to ensure that gcc will use the
pointers as-is. According to a variety of internet sources, either one
does the trick. However, there are also reports that at least the
volatile thing isn't fully correct, hence let's add some snake oil and
employ both techniques.
https://news.ycombinator.com/item?id=4711346
Michal Schmidt
Vito Caputo
Lennart Poettering
Andrew Jones
Torstein Husebø
Daniel Mack
Evgeny Vereshchagin
Susant Sahani
Tom Gundersen
Zbigniew Jędrzejewski-Szmek
Thomas Hindoe Paaboel Andersen