fix growpart to avoid race in partition table reload

[bcressey]

Issue number:
#845

Description of changes:
udevd watches files in /dev via inotify, and when a block disk device is closed after writing, it tells the kernel to reload the partition table.

This was racing with the reload code in growpart; if we were unlucky in our timing, the kernel would be busy with udevd's reload and our call would fail.

To fix this, we stop reloading the partition table in growpart, and use inotify to determine when udevd has deleted and recreated the partition links.

Testing done:
Tested launches and reboots of 250 instances in a cluster. All came up successfully and resized the partition to the expected 20 GiB size.

# systemctl status prepare-local
● prepare-local.service - Prepare Local Directory (/local)
     Loaded: loaded (/x86_64-bottlerocket-linux-gnu/sys-root/usr/lib/systemd/system/prepare-local.service; enabled; vendor preset: enabled)
     Active: active (exited) since Fri 2020-08-21 19:05:08 UTC; 22min ago
   Main PID: 2049 (code=exited, status=0/SUCCESS)
      Tasks: 0 (limit: 9185)
     Memory: 0B
     CGroup: /system.slice/prepare-local.service

Aug 21 19:05:07 localhost growpart[1911]: read partition table from nvme1n1
Aug 21 19:05:07 localhost growpart[1911]: wrote partition table to /dev/nvme1n1
Aug 21 19:05:07 localhost growpart[1911]: saw BOTTLEROCKET-DATA link deleted
Aug 21 19:05:07 localhost growpart[1911]: saw BOTTLEROCKET-DATA link created
Aug 21 19:05:07 localhost growpart[2008]: read partition table from nvme1n1
Aug 21 19:05:07 localhost growpart[2008]: wrote partition table to /dev/nvme1n1
Aug 21 19:05:07 localhost growpart[2008]: saw BOTTLEROCKET-DATA link deleted
Aug 21 19:05:07 localhost growpart[2008]: saw BOTTLEROCKET-DATA link created
Aug 21 19:05:08 localhost systemd-growfs[2026]: Successfully resized "/local" to 19.9G bytes.
Aug 21 19:05:08 localhost systemd[1]: Finished Prepare Local Directory (/local).

Terms of contribution:

By submitting this pull request, I agree that this contribution is dual-licensed under the terms of both the Apache License, version 2.0, and the MIT license.