[PATCH] scsi: mpt3sas: fix nr_msix underflow in _base_assign_reply_queues()

[Ivy Lopez]

The fallback path incorrectly subtracts iopoll_q_count when computing
nr_msix, when the intent is to reserve both high_iops_queues and
iopoll_q_count reply queues from the round-robin pool. Since
iopoll_q_count can be positive, the current subtraction inflates
nr_msix instead of reducing it, leaving far more queues in the
round-robin pool than actually available once high-iops and iopoll
queues are accounted for.

Beyond producing an incorrect grouping of cpus onto msix vectors, the
corrected formula can still drive nr_msix to zero or below under
plausible queue configurations (verified arithmetically), which would
either wrap to a large unsigned value and silently break affinity
grouping, or hit a divide-by-zero in the following nr_cpus / nr_msix
computation. Add an explicit check for nr_msix == 0 and also warn if the
reply queue budget is exhausted.

Fixes: 432bc7caef4e ("scsi: mpt3sas: Add io_uring iopoll support")
Signed-off-by: Ivy Lopez <skunkolee@xxxxxxxxx>
---
 drivers/scsi/mpt3sas/mpt3sas_base.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/drivers/scsi/mpt3sas/mpt3sas_base.c b/drivers/scsi/mpt3sas/mpt3sas_base.c
index ce5a5882acc8..31b80ed6ddec 100644
--- a/drivers/scsi/mpt3sas/mpt3sas_base.c
+++ b/drivers/scsi/mpt3sas/mpt3sas_base.c
@@ -3275,7 +3275,11 @@ _base_assign_reply_queues(struct MPT3SAS_ADAPTER *ioc)
 
 fall_back:
 	cpu = cpumask_first(cpu_online_mask);
-	nr_msix -= (ioc->high_iops_queues - iopoll_q_count);
+	nr_msix -= (ioc->high_iops_queues + iopoll_q_count);
+	if (!nr_msix) {
+		ioc_warn(ioc, "high_iops_queues and iopoll_q_count exceed available MSI-X vectors\n");
+		return;
+	}
 	index = 0;
 
 	list_for_each_entry(reply_q, &ioc->reply_queue_list, list) {
-- 
2.55.0