tests/bcachefs: test journal flush SRCU liveness

[Komzpa]

Summary

Adds journal_flush_srcu_liveness.ktest for the SRCU/journal-liveness shape in
koverstreet/bcachefs-tools#783.

The test mounts a 3-device debug bcachefs filesystem, delays journal flushes,
constrains move writeback concurrency, runs foreground fio pressure, and
evacuates a source device. It fails if the stress window emits either:

• btree trans held srcu lock
• bch2_journal_flush_seq stuck?

After the stress window it remounts read-only, unmounts, runs
bcachefs fsck -ny, and checks for bcachefs leak/error/device-error reports.

Current Source-PR Status

This was originally paired with
koverstreet/bcachefs-tools#790, but that PR is now
closed after upstream review rejected the broad SRCU-unlock approach and split
the useful pieces out.

The surviving source-side journal-reclaim work is
koverstreet/bcachefs-tools#792. This ktest should not be
read as a red-to-green reproducer for that PR: it is a regression/smoke oracle
for the journal-flush/SRCU liveness invariant from
koverstreet/bcachefs-tools#783.

Scope Update After Coverage Audit

This is not a failing reproducer for
koverstreet/bcachefs-tools#783. It passes unpatched
origin/testing at 838ead8957377100c58b501fc1c50816cb7e7da2 as well as the
historical koverstreet/bcachefs-tools#790 branch, so there is no red-to-green
transition.

I also tried a closer member-stall prototype that suspends a source member with
dmsetup suspend during bcachefs device evacuate. The initial oracle treated
any bch2_journal_flush_seq stuck? as failure, but that is too strict when the
test deliberately suspends a member for longer than the journal warning
threshold. With the oracle narrowed to post-resume non-recovery or
btree trans held srcu lock, the prototype passes current unpatched
koverstreet/bcachefs-tools origin/testing, while still showing the expected
journal-stuck diagnostics during the artificial suspend.

That prototype is therefore not included in this PR as a regression test.

Testing

• git diff --check origin/master..HEAD
• bash -n tests/fs/bcachefs/journal_flush_srcu_liveness.ktest
• tests/fs/bcachefs/journal_flush_srcu_liveness.ktest list-tests
• tests/fs/bcachefs/journal_flush_srcu_liveness.ktest deps
• Full VM run against unpatched koverstreet/bcachefs-tools origin/testing
  at 838ead8957377100c58b501fc1c50816cb7e7da2, confirming this is not a red
  reproducer:
  • bcachefs module loaded in guest: version: v1.38.8-161-g838ead895737
  • result: ========= PASSED journal_flush_srcu_liveness in 69s
  • result: TEST SUCCESS
• Full VM run against the original paired source fix at
  fd929bb4241a885b8ce62f0f5b2412f247b13699:
  • bcachefs module loaded in guest: version: v1.38.8-162-gfd929bb4241a
  • result: ========= PASSED journal_flush_srcu_liveness in 24s
  • result: TEST SUCCESS
  • log oracle found no btree trans held srcu lock
  • log oracle found no bch2_journal_flush_seq stuck?
• Full VM run against historical koverstreet/bcachefs-tools#790 source head
  58f07184dbe66dcfa01d8ef021d1326fa4027a2f:
  • bcachefs module loaded in guest: version: v1.38.8-182-g58f07184dbe6
  • result: ========= PASSED journal_flush_srcu_liveness in 114s
  • result: TEST SUCCESS
  • log oracle found no btree trans held srcu lock
  • log oracle found no bch2_journal_flush_seq stuck?

VM proof

VM run against master bcachefs-tools (kernel 7.2.0-rc4): journal_flush_srcu_liveness PASSED in 32s -- delayed journal flushes plus constrained move-writeback concurrency and foreground fio pressure on a 3-device debug fs completed without an SRCU/journal-liveness warning, source evacuated.