scsi: qedf: fix race in qedf_alloc_cmd()

Commit Message

Dan Carpenter Feb. 5, 2022, 9:27 a.m. UTC

The code tests whether there are any free_sqes at the start of the
function but does not update the accounting until later.  This
leaves a window where there is only one sqe available and two threads
could call qedf_alloc_cmd() at the same time and both succeeed.  Even
worse, now if more callers call qedf_alloc_cmd() instead of saying there
is -1 sqes available it will say there is a non-zero number available
and allow it.

The second problem with code is at the end of the function, if "bd_tbl"
is NULL, then the qedf_release_cmd() function will handle some of the
other accounting like "fcport->num_active_ios" and
"cmd_mgr->free_list_cnt" but it does not reset free_sqes back to what
it was.  So that is a resource leak.

Fixes: 61d8658b4a43 ("scsi: qedf: Add QLogic FastLinQ offload FCoE driver framework.")
Signed-off-by: Dan Carpenter <[e-mail]>
---
This patch is based on review and I am not able to test it.  My main
concern with this patch is I may be wrong with paragraph 2 which means
that my patch would just exchange one bug for a different bug.

This really requires someone who understands the code deeply to review
it.

And alternative would be to deliberately leave the potential resource
leak and only fix the race condition.  In other words, if bd_tbl is NULL
then goto out_failed instead of out_inc.

 drivers/scsi/qedf/qedf_io.c | 19 ++++++++-----------
 1 file changed, 8 insertions(+), 11 deletions(-)

Patch

diff --git a/drivers/scsi/qedf/qedf_io.c b/drivers/scsi/qedf/qedf_io.c
index fab43dabe5b3..83b68583230a 100644
--- a/drivers/scsi/qedf/qedf_io.c
+++ b/drivers/scsi/qedf/qedf_io.c
@@ -302,16 +302,12 @@  struct qedf_ioreq *qedf_alloc_cmd(struct qedf_rport *fcport, u8 cmd_type)
 	struct qedf_ioreq *io_req = NULL;
 	struct io_bdt *bd_tbl;
 	u16 xid;
-	uint32_t free_sqes;
 	int i;
 	unsigned long flags;
 
-	free_sqes = atomic_read(&fcport->free_sqes);
-
-	if (!free_sqes) {
+	if (atomic_dec_if_positive(&fcport->free_sqes) < 0) {
 		QEDF_INFO(&(qedf->dbg_ctx), QEDF_LOG_IO,
-		    "Returning NULL, free_sqes=%d.\n ",
-		    free_sqes);
+		    "Returning NULL, no free_sqes.\n ");
 		goto out_failed;
 	}
 
@@ -321,7 +317,7 @@  struct qedf_ioreq *qedf_alloc_cmd(struct qedf_rport *fcport, u8 cmd_type)
 		QEDF_INFO(&(qedf->dbg_ctx), QEDF_LOG_IO,
 		    "Returning NULL, num_active_ios=%d.\n",
 		    atomic_read(&fcport->num_active_ios));
-		goto out_failed;
+		goto out_inc;
 	}
 
 	/* Limit global TIDs certain tasks */
@@ -329,7 +325,7 @@  struct qedf_ioreq *qedf_alloc_cmd(struct qedf_rport *fcport, u8 cmd_type)
 		QEDF_INFO(&(qedf->dbg_ctx), QEDF_LOG_IO,
 		    "Returning NULL, free_list_cnt=%d.\n",
 		    atomic_read(&cmd_mgr->free_list_cnt));
-		goto out_failed;
+		goto out_inc;
 	}
 
 	spin_lock_irqsave(&cmd_mgr->lock, flags);
@@ -346,7 +342,7 @@  struct qedf_ioreq *qedf_alloc_cmd(struct qedf_rport *fcport, u8 cmd_type)
 
 	if (i == FCOE_PARAMS_NUM_TASKS) {
 		spin_unlock_irqrestore(&cmd_mgr->lock, flags);
-		goto out_failed;
+		goto out_inc;
 	}
 
 	if (test_bit(QEDF_CMD_DIRTY, &io_req->flags))
@@ -360,7 +356,6 @@  struct qedf_ioreq *qedf_alloc_cmd(struct qedf_rport *fcport, u8 cmd_type)
 	spin_unlock_irqrestore(&cmd_mgr->lock, flags);
 
 	atomic_inc(&fcport->num_active_ios);
-	atomic_dec(&fcport->free_sqes);
 	xid = io_req->xid;
 	atomic_dec(&cmd_mgr->free_list_cnt);
 
@@ -381,7 +376,7 @@  struct qedf_ioreq *qedf_alloc_cmd(struct qedf_rport *fcport, u8 cmd_type)
 	if (bd_tbl == NULL) {
 		QEDF_ERR(&(qedf->dbg_ctx), "bd_tbl is NULL, xid=%x.\n", xid);
 		kref_put(&io_req->refcount, qedf_release_cmd);
-		goto out_failed;
+		goto out_inc;
 	}
 	bd_tbl->io_req = io_req;
 	io_req->cmd_type = cmd_type;
@@ -394,6 +389,8 @@  struct qedf_ioreq *qedf_alloc_cmd(struct qedf_rport *fcport, u8 cmd_type)
 
 	return io_req;
 
+out_inc:
+	atomic_inc(&fcport->free_sqes);
 out_failed:
 	/* Record failure for stats and return NULL to caller */
 	qedf->alloc_failures++;