This code was in compat_threads, since it was _used_ for efficiently
notifying the main libevent thread from another thread. But in
spite of its usage, it's fundamentally a part of the network code.
Our previous algorithm had a nonzero probability of picking no
events to cancel, which is of course incorrect. The new code uses
Vitter's good old reservoir sampling "algorithm R" from 1985.
Fixes bug 26008; bugfix on 0.2.6.3-alpha.
* ADD new /src/common/crypto_rand.[ch] module.
* ADD new /src/common/crypto_util.[ch] module (contains the memwipe()
function, since all crypto_* modules need this).
* FIXES part of #24658: https://bugs.torproject.org/24658
This change makes cpuworker and test_workqueue no longer need to
include event2/event.h. Now workqueue.c needs to include it, but
that is at least somewhat logical here.
Each piece of queued work now has an associated priority value; each
priority goes on a separate queue.
With probability (N-1)/N, the workers will take work from the highest
priority nonempty queue. Otherwise, they'll look for work in a
queue of lower priority. This behavior is meant to prevent
starvation for lower-priority tasks.
Only some very ancient distributions don't ship with Libevent 2 anymore,
even the oldest supported Ubuntu LTS version has it. This allows us to
get rid of a lot of compat code.
This is a fairly easy way for us to get our test coverage up on
compat_threads.c and workqueue.c -- I already implemented these
tests, so we might as well enable them.
These functions must really never fail; so have crypto_rand() assert
that it's working okay, and have crypto_seed_rng() demand that
callers check its return value. Also have crypto_seed_rng() check
RAND_status() before returning.
We don't want to accept any work after one of our worker functions has
returned WQ_RPL_SHUTDOWN. This testcase currently fails, because we do
not actually stop any of the worker threads.
The workqueue test help message has two issues. First, the message uses 4 space
indentation when 2 space indentation seems more common. Second, the help
message misses some options.
This commit fixes both issues.
Previously I used one queue per worker; now I use one queue for
everyone. The "broadcast" code is gone, replaced with an idempotent
'update' operation.
