random(4): Fortuna: allow increased concurrency
Add experimental feature to increase concurrency in Fortuna. As this
diverges slightly from canonical Fortuna, and due to the security
sensitivity of random(4), it is off by default. To enable it, set the
tunable kern.random.fortuna.concurrent_read="1". The rest of this commit
message describes the behavior when enabled.
Readers continue to update shared Fortuna state under global mutex, as they
do in the status quo implementation of the algorithm, but shift the actual
PRF generation out from under the global lock. This massively reduces the
CPU time readers spend holding the global lock, allowing for increased
concurrency on SMP systems and less bullying of the harvestq kthread.
It is somewhat of a deviation from FS&K. I think the primary difference is
that the specific sequence of AES keys will differ if READ_RANDOM_UIO is
accessed concurrently (as the 2nd thread to take the mutex will no longer
receive a key derived from rekeying the first thread). However, I believe
the goals of rekeying AES are maintained: trivially, we continue to rekey
every 1MB for the statistical property; and each consumer gets a
forward-secret, independent AES key for their PRF.
Since Chacha doesn't need to rekey for sequences of any length, this change
makes no difference to the sequence of Chacha keys and PRF generated when
Chacha is used in place of AES.
On a GENERIC 4-thread VM (so, INVARIANTS/WITNESS, numbers not necessarily
representative), 3x concurrent AES performance jumped from ~55 MiB/s per
thread to ~197 MB/s per thread. Concurrent Chacha20 at 3 threads went from
roughly ~113 MB/s per thread to ~430 MB/s per thread.
Prior to this change, the system was extremely unresponsive with 3-4
concurrent random readers; each thread had high variance in latency and
throughput, depending on who got lucky and won the lock. "rand_harvestq"
thread CPU use was high (double digits), seemingly due to spinning on the
global lock.
After the change, concurrent random readers and the system in general are
much more responsive, and rand_harvestq CPU use dropped to basically zero.
Tests are added to the devrandom suite to ensure the uint128_add64 primitive
utilized by unlocked read functions to specification.
(cherry picked from commit 179f62805cf05c3ae334c03dd94cb4c2da46c716)