commit d20eb72e3c5bcefd1f50336d8b5a4d7fa87bcccc
author David Hastings <dahastin@fuchsia.infra.roller.google.com> Tue Sep 20 01:33:47 2022 +0000
committer Copybara-Service <copybara-worker@google.com> Mon Sep 19 18:35:16 2022 -0700
tree e854b8de7b7dccbb730c4d6860553ae97998c95b
parent 70e0e2bb190b7fd8ecd5708045dbfe8c761c46ee

[roll] Roll fuchsia [virtualization][vsock] Prevent TX from starving RX control

The state lock is write preferring. It seems like a possible ordering
is:

1) Write guard is dropped
2) Readers wake up, await when trying to acquire read lock
3) TX task wakes up and pulls a chain off of the queue
4) TX task acquires a write lock pre-empting the readers

That's generally not a problem in most places (although I might want to
revisit this with a fairer lock), but when handling an RX chain blocking
on a read also blocks control packets.

The integration test is maxing out TX throughput in a tight loop, and it
seems to have happened enough times that we were unable to send our
shutdown packet before the test finished.

Metrics from Debian on QEMU:

(Before) Micro Benchmark Results
------------------------
* Data corruption check: PASSED
* Round trip latency of 4096 bytes:
	Min: 183506ns (0.184ms)
	25th percentile: 524556ns (0.525ms)
	50th percentile: 857401ns (0.857ms)
	75th percentile: 1129873ns (1.130ms)
	99th percentile: 1941657ns (1.942ms)
	Max: 9924678ns (9.925ms)
* TX (guest -> host, unreliable) throughput of 128 MiB:
	Min: 100.94MiB/s
	25th percentile: 108.81MiB/s
	50th percentile: 111.96MiB/s
	75th percentile: 116.58MiB/s
	99thth percentile: 125.79MiB/s
	Max: 126.63MiB/s
* RX (host -> guest, unreliable) throughput of 128 MiB:
	Min: 103.03MiB/s
	25th percentile: 111.21MiB/s
	50th percentile: 115.81MiB/s
	75th percentile: 120.78MiB/s
	99thth percentile: 130.60MiB/s
	Max: 131.09MiB/s
* Single stream unidirectional round trip throughput of 128 MiB:
	Min: 163.80MiB/s
	25th percentile: 185.83MiB/s
	50th percentile: 190.41MiB/s
	75th percentile: 194.30MiB/s
	99thth percentile: 209.05MiB/s
	Max: 209.54MiB/s
* Multistream (5 connections) unidirectional round trip throughput of 128 MiB:
	Min: 32.65MiB/s
	25th percentile: 34.03MiB/s
	50th percentile: 34.59MiB/s
	75th percentile: 35.19MiB/s
	99thth percentile: 36.85MiB/s
	Max: 36.93MiB/s

(After) Micro Benchmark Results
------------------------
* Data corruption check: PASSED
* Round trip latency of 4096 bytes:
	Min: 185399ns (0.185ms)
	25th percentile: 795221ns (0.795ms)
	50th percentile: 1028699ns (1.029ms)
	75th percentile: 1287934ns (1.288ms)
	99th percentile: 2122094ns (2.122ms)
	Max: 7476711ns (7.477ms)
* TX (guest -> host, unreliable) throughput of 128 MiB:
	Min: 101.26MiB/s
	25th percentile: 110.47MiB/s
	50th percentile: 113.87MiB/s
	75th percentile: 118.03MiB/s
	99thth percentile: 124.97MiB/s
	Max: 126.28MiB/s
* RX (host -> guest, unreliable) throughput of 128 MiB:
	Min: 106.17MiB/s
	25th percentile: 114.16MiB/s
	50th percentile: 118.09MiB/s
	75th percentile: 122.71MiB/s
	99thth percentile: 130.44MiB/s
	Max: 132.59MiB/s
* Single stream unidirectional round trip throughput of 128 MiB:
	Min: 168.83MiB/s
	25th percentile: 185.59MiB/s
	50th percentile: 193.18MiB/s
	75th percentile: 200.73MiB/s
	99thth percentile: 218.25MiB/s
	Max: 222.25MiB/s
* Multistream (5 connections) unidirectional round trip throughput of 128 MiB:
	Min: 33.82MiB/s
	25th percentile: 34.65MiB/s
	50th percentile: 35.08MiB/s
	75th percentile: 35.47MiB/s
	99thth percentile: 35.97MiB/s
	Max: 36.49MiB/s

Original-Bug: 109780
Original-Reviewed-on: https://fuchsia-review.googlesource.com/c/fuchsia/+/726887
Original-Revision: 9e74f088457e5282af8eeed699efd3fabfaf8683
GitOrigin-RevId: bd845f7464ea0a704c62957af57f43aee9bd5bc0
Change-Id: I2fdd8c93f30340acfa06dc5d87cf70644487e9e1