[PATCH] scsi: isci: fix use-after-free in device removal path

[Michael Bommarito <michael.bommarito@gmail.com>]

The ISCI completion tasklet is initialized in isci_host_alloc()
(drivers/scsi/isci/init.c:496) and scheduled from both MSI-X and
legacy interrupt handlers (drivers/scsi/isci/host.c:223,613).

isci_host_deinit() stops the controller and waits for stop
completion, but it never kills completion_tasklet before teardown
continues. A top-of-function tasklet_kill() is not sufficient here:
interrupts are only disabled when isci_host_stop_complete() runs, so
until wait_for_stop() returns the IRQ handlers can still requeue the
tasklet. The tasklet callback also re-enables interrupts after
draining completions, so killing the tasklet before the source is
quiesced leaves the same race open.

Once wait_for_stop() returns, no further IRQ-driven scheduling can
occur. Kill completion_tasklet there so teardown cannot race a queued
tasklet running on a dead ihost. On remove or unload, the stale
callback can otherwise dereference ihost and touch ihost->smu_registers
after the host lifetime ends.

A UML + KASAN analogue reproduced the failure class both with no
tasklet_kill() and with tasklet_kill() placed before source quiesce,
and stayed clean once the kill happened after quiescing the scheduling
source.

This mirrors commit f6ab594672d4 ("scsi: aic94xx: fix use-after-free
in device removal path"), but ISCI needs the kill after
wait_for_stop().

Fixes: 6f231dda6808 ("isci: Intel(R) C600 Series Chipset Storage Control Unit Driver")
Cc: stable@vger.kernel.org
Assisted-by: Claude:claude-opus-4-7
Assisted-by: Codex:gpt-5-4
Signed-off-by: Michael Bommarito <michael.bommarito@gmail.com>
---
 drivers/scsi/isci/host.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/scsi/isci/host.c b/drivers/scsi/isci/host.c
index 6d2f4c831df7..ff199bab5d1a 100644
--- a/drivers/scsi/isci/host.c
+++ b/drivers/scsi/isci/host.c
@@ -1252,6 +1252,9 @@ void isci_host_deinit(struct isci_host *ihost)
 
 	wait_for_stop(ihost);
 
+	/* No further IRQ-driven scheduling can happen past wait_for_stop(). */
+	tasklet_kill(&ihost->completion_tasklet);
+
 	/* phy stop is after controller stop to allow port and device to
 	 * go idle before shutting down the phys, but the expectation is
 	 * that i/o has been shut off well before we reach this
-- 
2.53.0