Uses SeqLock for CachedAccountInner::hash

[brooksprumo]

Problem

The CachedAccountInner::hash field is protected with an RwLock to ensure the cached account hash is read/written safely. Since Hash is Copy, we can use a lighter-weight synchronization mechanism that is faster than RwLock.

Summary of Changes

Use SeqLock instead of RwLock for CachedAccountInner::hash.

A sequence lock read needs just two atomic loads, and a write locks a mutex plus does two atomic stores. Readers will retry if there is a write in between their two loads. Since we always need a hash value (i.e. non-reference), we're not wasting any time copying bytes around either.

Performance Comparisons

Here's some results micro-benchmarking[1] results from my macbook. This compares SeqLock with RwLock from std and parking_lot.

[1]: Here's the code for the benchmark: https://github.com/brooksprumo/seqlock/blob/master/benches/bench_seqlock.rs

First is when there are only readers:

And here's about 50/50 readers and writers:

I know micro-benchmarks are not really indicative of true speedups, but hopefully this shows that SeqLock is orders of magnitude faster than std's RwLock. Any reduction in lock contention will likely have a positive result on the system as a whole.

[codecov]

Codecov Report

Merging #33696 (18de3ac) into master (0b05e8d) will increase coverage by 0.0%.
Report is 2 commits behind head on master.
The diff coverage is 100.0%.

@@           Coverage Diff           @@
##           master   #33696   +/-   ##
=======================================
  Coverage    81.8%    81.8%           
=======================================
  Files         806      806           
  Lines      217894   217893    -1     
=======================================
+ Hits       178252   178262   +10     
+ Misses      39642    39631   -11     

[alessandrod]

Great stuff, this seems like the ideal case for a seqlock: the hash is small so even in the write contended case readers will likely never spin more than once or twice. Death to syscalls!

[alessandrod]

[1]: Here's the code for the benchmark: https://github.com/brooksprumo/seqlock/blob/master/benches/bench_seqlock.rs

Sorry I'm doing a lot of cargo build today so I have time to kill 😂 You could make the bench more realistic by holding [u8; 32] inside the locks, and making sure the rwlock benches copy the array too

[brooksprumo]

You could make the bench more realistic by holding [u8; 32] inside the locks, and making sure the rwlock benches copy the array too

Yep! I thought about doing this, but initially wanted the seqlock repo to just benchmark the locking and not any effects from copying. I agree that using a [u8; 32] would better match our use case here though!

[brooksprumo]

Sorry I'm doing a lot of cargo build today so I have time to kill 😂

Thank you for taking the time to look/review!

[HaoranYi]

Nice. lgtm.

[brooksprumo]

Now that solana-labs/solana-program-library#5597 has merged, SPL builds should be good too.