`std::lock_guard` is all we need here, and the added complexity of
`std::unique_lock` is not used and is usually optimized away by the
compiler.  Using `std::lock_guard` directly will reduce the amount of
work that the optimizer needs to do and saves some build CPU cycles.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

The `PerfCounters::inc()` method acquires another lock which can block
the calling thread while holding the `finisher_lock` which can cause a
lot of lock contention.  This can be avoided easily by moving the
`PerfCounters::inc()` call out of the protected code block.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

If `finisher_running` is set, then the `Finisher` thread will
automatically pick up new items queued by other threads.  It is
therefore not needed to wake it up and we can eliminate one system
call.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

Pushing to the queue may take a long time when the `std::vector` needs
to allocate more memory.  We should wake up the `Finisher` thread only
right before unlocking the `finisher_mutex` to reduce lock contention,
because it is the more likely that the mutex can really be acquired
when the thread really wakes up.

This imitates how commit cc7ec3e did it - it refactored only
one of the `queue()` overloads, leaving less-than-optimal copies of
this piece code in all other overloads.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

As noted in commit cc7ec3e, there is only ever a single `Finisher`
thread, therefore the overhead for `notify_all()` can be eliminated.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

These are never changed; `const` prevents accidental changes and
allows further compiler optimizations.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

This aims to speed up compile times because constructor and destructor
contain a lot of code that would be compiled in sources that do not
call them.  Also this allows removing the "common/perf_counters.h"
include.

Since there is now only one instantiation of these for all call sites,
the binary size shrinks by nearly 1 kB.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

This is not only more efficient at runtime, but also shaves 500 bytes
code off the binary.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

This allows eliminating the copy in `ActivePyModule`.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>

This eliminates a temporary `std::string`.

Additionally, convert the `tn` parameter to an rvalue reference and
move it to `thread_name`.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>