Scrubbing root pools may deadlock on kernels without elevator_change()

[loli10K]

Motivation and Context

Fix #8664
This fix is an alternative to #9317 and will be hopefully easier to backport to the 0.8 release branch.

Description

Originally the zfs_vdev_elevator module option was added as a
convenience so the requested elevator would be automatically set
on the underlying block devices. At the time this was simple
because the kernel provided an API function which did exactly this.

This API was then removed in the Linux 4.12 kernel which prompted
us to add compatibly code to set the elevator via a usermodehelper.

Unfortunately changing the evelator via usermodehelper requires reading
some userland binaries, most notably modprobe(8) or sh(1), from a zfs
dataset on systems with root-on-zfs. This can deadlock the system if
used during the following call path because it may need, if the data
is not already cached in the ARC, reading directly from disk while
holding the spa config lock as a writer:

  zfs_ioc_pool_scan()
    -> spa_scan()
      -> spa_scan()
        -> vdev_reopen()
          -> vdev_elevator_switch()
            -> call_usermodehelper()

While the usermodehelper waits sh(1), modprobe(8) is blocked in the ZIO pipeline trying to read from disk:

  INFO: task modprobe:2650 blocked for more than 10 seconds.
       Tainted: P           OE     5.2.14
  modprobe        D    0  2650    206 0x00000000
  Call Trace:
   ? __schedule+0x244/0x5f0
   schedule+0x2f/0xa0
   cv_wait_common+0x156/0x290 [spl]
   ? do_wait_intr_irq+0xb0/0xb0
   spa_config_enter+0x13b/0x1e0 [zfs]
   zio_vdev_io_start+0x51d/0x590 [zfs]
   ? tsd_get_by_thread+0x3b/0x80 [spl]
   zio_nowait+0x142/0x2f0 [zfs]
   arc_read+0xb2d/0x19d0 [zfs]
   ...
   zpl_iter_read+0xfa/0x170 [zfs]
   new_sync_read+0x124/0x1b0
   vfs_read+0x91/0x140
   ksys_read+0x59/0xd0
   do_syscall_64+0x4f/0x130
   entry_SYSCALL_64_after_hwframe+0x44/0xa9

This commit changes how we use the usermodehelper functionality from
synchronous (UMH_WAIT_PROC) to asynchronous (UMH_NO_WAIT) which prevents
scrubs, and other vdev_elevator_switch() consumers, from triggering the
aforementioned issue.

How Has This Been Tested?

Tested on local Debian builder with root pool (mirror) with primarycache=none running zpool scrub -s/zpool scrub -s rpool in a loop.

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)
• Documentation (a change to man pages or other documentation)

Checklist:

• My code follows the ZFS on Linux code style requirements.
• I have updated the documentation accordingly.
• I have read the contributing document.
• I have added tests to cover my changes.
• All new and existing tests passed.
• All commit messages are properly formatted and contain Signed-off-by.

[codecov]

Codecov Report

Merging #9321 into master will decrease coverage by 0.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #9321      +/-   ##
==========================================
- Coverage   79.09%   79.07%   -0.02%     
==========================================
  Files         401      401              
  Lines      122089   122074      -15     
==========================================
- Hits        96568    96532      -36     
- Misses      25521    25542      +21

Flag	Coverage Δ
#kernel	79.74% <100%> (+0.01%)	⬆️
#user	66.73% <ø> (+0.1%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 7238cbd...1fd69a3. Read the comment docs.