mirror of
https://github.com/SCST-project/scst.git
synced 2026-10-22 06:02:33 +00:00
Gleb Chesnokov
60ff16ef4f
Found by manual code inspection, not by an analyzer or a runtime test.
When scst_suspend_activity() times out with commands still active,
__scst_resume_activity() revives the command and management counters.
The old wrapper calls percpu_ref_reinit(), whose zero-count requirement
is enforced by this kernel assertion:
WARN_ON_ONCE(!percpu_ref_is_zero(ref));
Upstream 3.18-4.19 and the retained EL7/UEK4/UEK5 implementations preserve
the per-CPU allocation and use long counters, but reinit still rejects
this legitimate nonzero resurrection. Restore the initial reference
additively, retain outstanding references, and restore per-CPU operation
with the kernel's bias, counter clearing and release-store ordering.
Keep explicitly forced atomic mode unchanged.
SCST serializes kill and resurrection with scst_suspend_mutex and keeps
both counters alive even when their release callbacks run. Require the
same external serialization for mode changes and exit. Concurrent gets
and puts remain allowed.
Kill confirmation can wake the caller before confirm_switch is cleared.
Drain a pending switch callback with rcu_barrier_sched() before reusing
the counters. The kernel switch waitqueue is private, and older kernels
do not export percpu_ref_switch_to_percpu(). This conditional barrier
can also wait for unrelated RCU callbacks; it is skipped once the switch
has completed. Native resurrection on 4.20+ and EL8+ remains unchanged.
Add scst-LIB-percpu-ref to scst.ko with private, real kernel counters.
Keep the backport helper available to KUnit on modern kernels, using
their counter layout and unified RCU API. Cover immediate resurrection
with active refs, resurrection after draining, forced atomic mode, and
repeated suspend cancellations with commands completing while dead.
Also fix the fake Kbuild cleaner used by the Python tooling tests.
CI reported test_cleanup_without_packaged_tooling failing with
"find: cannot delete .../isert-scst.ko: No such file or directory".
iSCSI cleans kernel/ and kernel/isert-scst/ concurrently, so one fixture
find can remove an artifact after the other has enumerated it. Use
-ignore_readdir_race to tolerate only vanished files, matching Kbuild's
idempotent rm -f behavior. Keep other find errors and the injected
FAIL_CLEAN failure fatal; retain parallel cleanup and artifact checks.
Tests, builds and VM validation were not run, as requested.