perf(virtio): use weaker memory barriers from the mem-barrier crate

[mkroening]

This PR replaces the kernel-internal mandatory memory_barrier with weaker memory barriers from the new mem-barriers crate.

This will probably not make a big difference, but this should be an improvement.

What I also noticed is that for packed virtqueues, we currently don't have any memory barriers at all, which should be fixed in the future. For split virtqueues, I think we also have too few at the moment.

[github-actions]

Benchmark Results

Benchmark	Current: 4299cb6	Previous: befcf26	Performance Ratio
startup_benchmark Build Time	116.74 s	111.11 s	1.05
startup_benchmark File Size	0.87 MB	0.87 MB	1.00
Startup Time - 1 core	1.02 s (±0.02 s)	0.98 s (±0.02 s)	1.04
Startup Time - 2 cores	1.01 s (±0.03 s)	0.99 s (±0.03 s)	1.01
Startup Time - 4 cores	1.00 s (±0.03 s)	0.99 s (±0.02 s)	1.01
multithreaded_benchmark Build Time	116.63 s	111.66 s	1.04
multithreaded_benchmark File Size	0.97 MB	0.97 MB	1.00
Multithreaded Pi Efficiency - 2 Threads	93.17 % (±8.19 %)	87.45 % (±8.10 %)	1.07
Multithreaded Pi Efficiency - 4 Threads	44.74 % (±2.94 %)	42.77 % (±3.00 %)	1.05
Multithreaded Pi Efficiency - 8 Threads	25.66 % (±1.52 %)	25.18 % (±1.60 %)	1.02
micro_benchmarks Build Time	300.82 s	296.75 s	1.01
micro_benchmarks File Size	0.98 MB	0.98 MB	1.00
Scheduling time - 1 thread	180.88 ticks (±32.99 ticks)	176.35 ticks (±34.79 ticks)	1.03
Scheduling time - 2 threads	100.91 ticks (±19.37 ticks)	104.10 ticks (±21.43 ticks)	0.97
Micro - Time for syscall (getpid)	11.38 ticks (±4.59 ticks)	11.69 ticks (±5.33 ticks)	0.97
Memcpy speed - (built_in) block size 4096	56883.92 MByte/s (±41662.81 MByte/s)	61256.19 MByte/s (±45369.75 MByte/s)	0.93
Memcpy speed - (built_in) block size 1048576	14792.56 MByte/s (±12874.46 MByte/s)	14154.64 MByte/s (±11768.17 MByte/s)	1.05
Memcpy speed - (built_in) block size 16777216	9290.99 MByte/s (±7457.02 MByte/s)	9650.18 MByte/s (±7829.51 MByte/s)	0.96
Memset speed - (built_in) block size 4096	57150.54 MByte/s (±41824.50 MByte/s)	62658.40 MByte/s (±45879.30 MByte/s)	0.91
Memset speed - (built_in) block size 1048576	15064.46 MByte/s (±12983.79 MByte/s)	14567.64 MByte/s (±12031.17 MByte/s)	1.03
Memset speed - (built_in) block size 16777216	9511.80 MByte/s (±7586.08 MByte/s)	9902.13 MByte/s (±7993.63 MByte/s)	0.96
Memcpy speed - (rust) block size 4096	52416.85 MByte/s (±38777.31 MByte/s)	55637.12 MByte/s (±41593.40 MByte/s)	0.94
Memcpy speed - (rust) block size 1048576	13697.10 MByte/s (±11164.80 MByte/s)	13921.96 MByte/s (±11517.33 MByte/s)	0.98
Memcpy speed - (rust) block size 16777216	9552.87 MByte/s (±7716.97 MByte/s)	9776.63 MByte/s (±7947.05 MByte/s)	0.98
Memset speed - (rust) block size 4096	53016.95 MByte/s (±39192.55 MByte/s)	56255.41 MByte/s (±41950.49 MByte/s)	0.94
Memset speed - (rust) block size 1048576	14026.00 MByte/s (±11349.89 MByte/s)	14238.58 MByte/s (±11680.58 MByte/s)	0.99
Memset speed - (rust) block size 16777216	9821.81 MByte/s (±7892.11 MByte/s)	10072.00 MByte/s (±8153.23 MByte/s)	0.98
alloc_benchmarks Build Time	296.15 s	293.98 s	1.01
alloc_benchmarks File Size	0.95 MB	0.95 MB	1.00
Allocations - Allocation success	100.00 %	100.00 %	1
Allocations - Deallocation success	100.00 %	100.00 %	1
Allocations - Pre-fail Allocations	100.00 %	100.00 %	1
Allocations - Average Allocation time	25753.01 Ticks (±1054.80 Ticks)	25979.67 Ticks (±1072.71 Ticks)	0.99
Allocations - Average Allocation time (no fail)	25753.01 Ticks (±1054.80 Ticks)	25979.67 Ticks (±1072.71 Ticks)	0.99
Allocations - Average Deallocation time	3076.12 Ticks (±1076.97 Ticks)	3078.39 Ticks (±1348.13 Ticks)	1.00
mutex_benchmark Build Time	298.48 s	295.48 s	1.01
mutex_benchmark File Size	0.98 MB	0.98 MB	1.00
Mutex Stress Test Average Time per Iteration - 1 Threads	36.56 ns (±4.36 ns)	36.76 ns (±3.72 ns)	0.99
Mutex Stress Test Average Time per Iteration - 2 Threads	30.46 ns (±3.56 ns)	29.88 ns (±3.19 ns)	1.02

This comment was automatically generated by workflow using github-action-benchmark.

[cagatay-y]

The architectural details at play are a bit over my head but I was wondering if you checked if the PCI configuration spaces and BAR spaces in particular require such barriers. As far as I can see, the explicit references to memory barriers in the VIRTIO specification relate to the structures allocated by the guest and lay outside the aforementioned spaces. Accesses to configuration and BAR spaces should be safe from the compiler as they are behind volatile pointers. I am a bit confused about how the processor treats those accesses but it is surprising to me that the processor would reorder them even though it is aware that they are MMIO spaces.