noinline vdev_queue_max_async_writes so it can be dtrace'd

[allanjude]

Summary

When trying to use dtrace to investigate ZFS write performance, an important
probe point is not usable because it has been inlined by the compiler.

I am guessing this doesn't happen on IllumOS because they use GCC.

PR: https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=200316
Submitted by: Lacey Powers <[e-mail]>

[delphij]

Have you measured the performance impact of this change?

[allanjude]

In D3472#71090, @delphij wrote:

Have you measured the performance impact of this change?

I ran before and after tests of this patch on a machine with a pair of mirrored 3TB disks, and 24GB of ram.
There was no measurable difference in async write performance.

Running dtrace on the function while writing at max bandwidth (~170mb/s) resulted in 358755 entries into the function in the 60 second measurement period (~6 per ms).
When idle, entries into this function are fewer than 5000 per 60 seconds.

My understanding is that because this function is called when an async write is being queued, to determine queue position, it shouldn't be critical to performance, so un-inlining it shouldn't impact much.

Do you have a recommended benchmark to validate this change?

[avg]

This comment was removed by avg.

[delphij]

No objections from me.