Detect a slow raidz child during reads

[robn]

Motivation and Context

Replacing #16900, which was almost finished with review updates but has stalled. I've been asked to take it over.

See original PR for details.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Performance enhancement (non-breaking change which improves efficiency)
• Code cleanup (non-breaking change which makes code smaller or more readable)
• Breaking change (fix or feature that would cause existing functionality to change)
• Library ABI change (libzfs, libzfs_core, libnvpair, libuutil and libzfsbootenv)
• Documentation (a change to man pages or other documentation)

Checklist:

• My code follows the OpenZFS code style requirements.
• I have updated the documentation accordingly.
• I have read the contributing document.
• I have added tests to cover my changes.
• I have run the ZFS Test Suite with this change applied.
• All commit messages are properly formatted and contain Signed-off-by.

[robn]

Because there will be a little bouncing between two PRs, and because there's two different authors involved, I'll be pushing fixup commits to this branch. Once everyone is happy with review, I will squash them down for merge.

I'll close out the remaining review comments on #16900, and would like it if new comments could be added here. Thanks all for your patience; I know its a bit fiddly (it'd be nicer if Github would allow a PR to change branches, alas).

[tonyhutter]

I haven't looked into it, but I see all the FreeBSD runners have:

Tests with results other than PASS that are unexpected:
    SKIP events/setup (expected PASS)
    SKIP events/slow_vdev_sit_out (expected PASS)

[tonyhutter]

@robn this fixed the FreeBSD CI errors for me: tonyhutter@a491397. Try squashing that in and rebasing.

[behlendorf]

Thanks for picking up this work! It'd be great if we can refine this a bit more and get it integrated.

[pcd1193182]

I've updated this branch to fix a few things. First, I've significantly reduced (hopefully eliminated almost entirely) the unnecessary sit-outs Brian was reporting. We reproduced them internally during our performance testing and the updated version doesn't display them at all. The changes here are to use the latency histogram stats instead of the EWMA as a better source of data. We also decrease the check frequency dramatically to reduce noise, decrease the number of outliers to compensate (along with adding a facility for extreme events to increase their outlier count more rapidly), increase the fence value significantly, and add a decay mechanism to prevent random noise from eventually causing a sit-out of healthy disks.

Second, I've added an autosit property to vdevs. When set to on, it activates this code. This allows users to decide if they want this logic enabled or not. This is intended to work with the next change:

Third, I've made the sitout property writeable. This allows individual vdevs to be sat out from userland. This, in conjunction with the autosit property, allows the user to decide if they want no disk sit-outs, the kernel's automatic sit-outs, or to do something more complex. Using zpool iostat latency data, SMART stats, or any other data source they can think of, they could now create a userland daemon that monitors disk health and sits out disks that it feels are unhealthy.

Giving the capability to this in userland has a number of advantages: easier access to high-level languages and their rich libraries, more safe and rapid iteration of complex logic, and the ability to improve the logic using new developments and advanced approaches without requiring a kernel upgrade or downtime. The kernel functionality is left in place as a simple plug-and-play approach.

[amotin]

I agree that histograms should indeed be a better source of data, comparing to EWMA, except as I mention below, we may take a closer look when we update the previous state. Same time, I wonder if histograms may actually give us even more statistical information about the distribution curves, so that we could better estimate the confidence interval on a small number of disks.

[behlendorf]

Given the amount of churn in this PR it'd be nice to squash the commits the next time it's updated.

[robn]

@pcd1193182 thanks for picking up this often-dropped PR and getting it over the line. Thanks all for the feedback!

[stevekstevek]

Not sure anyone will see this comment, but this capability seems like it could be adapted to my (admittedly unusual, and against some recommendations) setup. I have a raidz1 with just 2 disks. I did this because I want to be able to expand it later to more disks. But for the moment it is what I have. I think that the code in this PR could help me get near mirror read performance from this setup, since the parity disks have exactly all the data, and CPU is relatively cheap. Wonder if the people here more knowlegeable than me would agree with this assessment, and whether they think it would be worthwhile to pursue. It seems like it wouldn't be very useful when there's more than 2 disks involved, and I don't know how common this configuration is (though since 2 disk raidz1 became possible, I imagine other people might have been tempted to do what I did).

[amotin]

@stevekstevek The sit out functionality requires at least 5 (LAT_CHILDREN_MIN) disks to get reasonable statistics. With 2 disks it could give too many false positives.

[pcd1193182]

It's worth noting that in the abstract, your read performance already shouldn't be that different from a mirror. With the raidz1 quirk where the parity and data swap every megabyte, it should already be issuing reads to both devices. This would matter more with 3-disk raidz or 4-disk raidz, where that doesn't happen, but it also doesn't require this functionality; the raidz io start logic could switch to mirror-style dispatching for single sector reads, at least in theory.