It includes the following policy changes:
* All GetRand* functions seed the stack pointer and rdrand result
(in addition to the performance counter)
* The periodic entropy added by the idle scheduler now seeds stack pointer,
rdrand and perfmon data (once every 10 minutes) in addition to
just a sleep timing.
* The entropy added when calling GetStrongRandBytes no longer includes
the once-per-10-minutes perfmon data on windows (it is moved to the
idle scheduler instead, where latency matters less).
Other changes:
* OpenSSL is no longer seeded directly anywhere. Instead, any generated
randomness through our own RNG is fed back to OpenSSL (after an
additional hashing step to prevent leaking our RNG state).
* Seeding that was previously done directly in RandAddSeedSleep is now
moved to SeedSleep(), which is indirectly invoked through ProcRand
from RandAddSeedSleep.
* Seeding that was previously done directly in GetStrongRandBytes()
is now moved to SeedSlow(), which is indirectly invoked through
ProcRand from GetStrongRandBytes().
Note that the CScheduler thread cant be running at this point,
it has already been stopped with the rest of the init threadgroup.
Thus, just calling any remaining loose callbacks during Shutdown()
is sane.
This will be used by CValidationInterface soon.
This requires a bit of work as we need to ensure that most of our
callbacks happen in-order (to avoid synchronization issues in
wallet) - we keep our own internal queue and push things onto it,
scheduling a queue-draining function immediately upon new
callbacks.
Make a copy of the boost time-point to wait for, otherwise the head of
the queue may be deleted by another thread while this one is waiting,
while the boost function still has a reference to it.
Although this problem is in non-test code, this is not an actual problem
outside of the tests because we use the thread scheduler with only one
service thread, so there will never be threads fighting at the head of
the queue.
The old boost fallback escapes this problem because it passes a scalar
value to wait_until instead of a const object reference.
Found by running the tests in LLVM-4.0-master asan.
Don't clear `stopRequested` and `stopWhenEmpty` at the top of
`serviceQueue`, as this results in a race condition: on systems under
heavy load, some of the threads only get scheduled on the CPU when the
other threads have already finished their work. This causes the flags to
be cleared post-hoc and thus those threads to wait forever.
The potential drawback of this change is that the scheduler cannot be
restarted after being stopped (an explicit reset would be needed), but
we don't use this functionality anyway.
On a busy or slow system, the CScheduler unit test could fail because it
assumed all threads would be done after a couple of milliseconds.
Replace the hard-coded sleep with CScheduler stop() method that
will cleanly exit the servicing threads when all tasks are completely
finished.
