Add the full ZFS umbrella group

I've tested this patch quite a bit on powerpc64, where I was seeing the issue a lot. So far, this patch resolves the issue and doesn't cause any new issues. Any status update on this?

Thanks for attacking this Andriy its a big task which really needed some attention.

I don't feel qualified to comment directly on the VFS changes myself without investing quite a bit of time familiarising myself with the ins and outs of our VFS layer and unfortunately I don't have time for that ATM.

Would it be a good idea to add kib to the reviews, IIRC he's very familiar with VFS?

Another victim, disputable perhaps, is support for case insensitive
operations.

What do you mean by "disputable" here?

IMO ths stuff should be committed. Nobody cares except ZFS/VFS interaction besides you, and waiting for a review would only cause aggravation to users suffering from the bug.

In D6533#150525, @swills wrote:

I've tested this patch quite a bit on powerpc64, where I was seeing the issue a lot. So far, this patch resolves the issue and doesn't cause any new issues. Any status update on this?

Thank you very much for testing! I'd like to have a bit more of discussion before I can consider committing this change.

In D6533#151464, @smh wrote:

Would it be a good idea to add kib to the reviews, IIRC he's very familiar with VFS?

Thank you for adding Kostik. The more eyes the better.

In D6533#151510, @emaste wrote:

What do you mean by "disputable" here?

It seems that we formally support insensitive and mixed values of casesensitivity property, but I am not sure how much we really support them now.
I think that mixed is the same as sensitive at the moment as we do not have any other mechanism to control the policy.
Also, I am not sure how well our VFS support case insensitive operations (the name cache, etc).

So, I removed the support for anything else but the case-sensitive lookups (in ZPL) to make the code simpler, but I am not sure if that was a right thing to do.

In D6533#151511, @kib wrote:

IMO ths stuff should be committed. Nobody cares except ZFS/VFS interaction besides you, and waiting for a review would only cause aggravation to users suffering from the bug.

Well, I think that the patch has several points of potential contention. Some functions are changed so much that they are very different from their originals in illumos.
Also, I decided to remove all chunks of non-trivial size guarded with ifdef illumos. In my opinion they served only to obscure the code and didn't help much with correctly merging upstream changes. My opinion in general is that it is inevitable that the "upper half" of the ZPL code would be quite different between OSes.

The patch is also "incomplete". There are still some major changes to be done. For instance, z_teardown_lock that protects things like rollback and unmounting should be acquired before any vnode locks, but currently that is quite hard to implement, so it is always acquired after all vnode locks making the code messier than necessary.
Ideally, however, it would be better to replace z_teardown_lock with a VFS mechanism similar to vfs_write_suspend.

All that said, I do not really mind committing the change. I will try to draw more attention to this review request, so that the contentious points are discussed before the code goes in.

I've now done a number of exp-run builds (all packages) on an amd64 box using this patch and there were no issues, or at least none that I noticed.

For what it's worth, I use a case-insensitive filesystem to build a Mono project.
This won't be missed: the project should be fixed instead.

In D6533#152851, @lifanov_mail.lifanov.com wrote:

For what it's worth, I use a case-insensitive filesystem to build a Mono project.
This won't be missed: the project should be fixed instead.

I've restored the functionality (and even tried to make it more correct).
It would be great if you could test it.
Thank you!

Hmm. I'd like to run this in the freebsd.org cluster on a larger scale. Do you have this checked in somewhere? Dealing with raw .diff files without version history isn't terribly appealing.

For what it's worth, I've been running this on amd64 on top of 11.0-ALPHA2 since June 3rd without a hiccup. Plenty of regular usage, running a production email server as well as a lot of other stuff.

We're running diff 19054 on a bunch of freebsd.org cluster machines now. I'm spinning up the machine that used to regularly explode when building npm/node things right now.

I confirmed that with the latest patch case-insensitive filesystem works.

$ echo foo >> /tmp/m/ab
$ echo bar >> /tmp/m/Ab
$ echo baz >> /tmp/m/aB
$ cat /tmp/m/AB
foo
bar
baz