kern/52700: Testbed VMs hang on wd1 access since recent ATA changes

[gson%gson.org@localhost (Andreas Gustafsson)]

>Number:         52700
>Category:       kern
>Synopsis:       Testbed VMs hang on wd1 access since recent ATA changes
>Confidential:   no
>Severity:       critical
>Priority:       high
>Responsible:    kern-bug-people
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Sun Nov 05 13:20:00 +0000 2017
>Originator:     Andreas Gustafsson
>Release:        NetBSD-current, source date >= 2017.11.01.19.34.46
>Organization:

>Environment:
System: NetBSD
Architecture: i386
Machine: i386
>Description:

The test results at

  http://releng.netbsd.org/b5reports/i386/commits-2017.11.html#2017.11.01.19.15.31 

include "ATF output: raw, xml, html" links as expected up to
and including source date 2017.11.01.19.15.31:

  build: OK with 483383 lines of log, install: OK, tests: 6004 passed, 173 skipped, 123 expected_failure, 8 failed, ATF output: raw, xml, html

but they are missing from the next run, of source date
2017.11.01.21.13.26, and subsequent ones:

  build: OK with 484342 lines of log, install: OK, tests: 6004 passed, 173 skipped, 123 expected_failure, 8 failed

Since the ATF output files are exported from the test VM using a
secondary virtual hard disk (wd1), it seems likely that this is
related to the ATA changes committed between these two dates:

  commit 2017.11.01.19.34.45 mlelstv src/sys/dev/files.dev 1.4
  commit 2017.11.01.19.34.46 mlelstv src/sys/dev/ata/ata_raid_adaptec.c 1.10
  commit 2017.11.01.19.34.46 mlelstv src/sys/dev/ata/ata_raid_intel.c 1.8
  commit 2017.11.01.19.34.46 mlelstv src/sys/dev/ata/ata_raid_jmicron.c 1.6
  commit 2017.11.01.19.34.46 mlelstv src/sys/dev/ata/ata_raid_nvidia.c 1.3
  commit 2017.11.01.19.34.46 mlelstv src/sys/dev/ata/ata_raid_promise.c 1.12
  commit 2017.11.01.19.34.46 mlelstv src/sys/dev/ata/ata_raid_via.c 1.8
  commit 2017.11.01.19.34.46 mlelstv src/sys/dev/ata/wd.c 1.434
  commit 2017.11.01.19.34.46 mlelstv src/sys/dev/ata/wdvar.h 1.45

From the console logs, it looks a silent hang at the point where the
test results should be written to rwd1d immediately after the test run
has finished.

The amd64 port is also affected, but sparc (which uses a virtual
SCSI disk instead of an ATA one) is not.

>How-To-Repeat:

Inspect the test reports.

>Fix: