net/mlx5e: Fix interoperability between XSK and ICOSQ recovery flow

[ Upstream commit 17958d7 ]

Both regular RQ and XSKRQ use the same ICOSQ for UMRs. When doing
recovery for the ICOSQ, don't forget to deactivate XSKRQ.

XSK can be opened and closed while channels are active, so a new mutex
prevents the ICOSQ recovery from running at the same time. The ICOSQ
recovery deactivates and reactivates XSKRQ, so any parallel change in
XSK state would break consistency. As the regular RQ is running, it's
not enough to just flush the recovery work, because it can be
rescheduled.

Fixes: be5323c ("net/mlx5e: Report and recover from CQE error on ICOSQ")
Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>

drivers/net/ethernet/mellanox/mlx5/core/en.h

@@ -727,6 +727,8 @@ struct mlx5e_channel {
 	DECLARE_BITMAP(state, MLX5E_CHANNEL_NUM_STATES);
 	int                        ix;
 	int                        cpu;
+	/* Sync between icosq recovery and XSK enable/disable. */
+	struct mutex               icosq_recovery_lock;
 };
 
 struct mlx5e_ptp;

drivers/net/ethernet/mellanox/mlx5/core/en/health.h

@@ -30,6 +30,8 @@ void mlx5e_reporter_rx_destroy(struct mlx5e_priv *priv);
 void mlx5e_reporter_icosq_cqe_err(struct mlx5e_icosq *icosq);
 void mlx5e_reporter_rq_cqe_err(struct mlx5e_rq *rq);
 void mlx5e_reporter_rx_timeout(struct mlx5e_rq *rq);
+void mlx5e_reporter_icosq_suspend_recovery(struct mlx5e_channel *c);
+void mlx5e_reporter_icosq_resume_recovery(struct mlx5e_channel *c);
 
 #define MLX5E_REPORTER_PER_Q_MAX_LEN 256
 #define MLX5E_REPORTER_FLUSH_TIMEOUT_MSEC 2000

drivers/net/ethernet/mellanox/mlx5/core/en/reporter_rx.c

@@ -59,6 +59,7 @@ static void mlx5e_reset_icosq_cc_pc(struct mlx5e_icosq *icosq)
 
 static int mlx5e_rx_reporter_err_icosq_cqe_recover(void *ctx)
 {
+	struct mlx5e_rq *xskrq = NULL;
 	struct mlx5_core_dev *mdev;
 	struct mlx5e_icosq *icosq;
 	struct net_device *dev;
@@ -67,7 +68,13 @@ static int mlx5e_rx_reporter_err_icosq_cqe_recover(void *ctx)
 	int err;
 
 	icosq = ctx;
+
+	mutex_lock(&icosq->channel->icosq_recovery_lock);
+
+	/* mlx5e_close_rq cancels this work before RQ and ICOSQ are killed. */
 	rq = &icosq->channel->rq;
+	if (test_bit(MLX5E_RQ_STATE_ENABLED, &icosq->channel->xskrq.state))
+		xskrq = &icosq->channel->xskrq;
 	mdev = icosq->channel->mdev;
 	dev = icosq->channel->netdev;
 	err = mlx5_core_query_sq_state(mdev, icosq->sqn, &state);
@@ -81,6 +88,9 @@ static int mlx5e_rx_reporter_err_icosq_cqe_recover(void *ctx)
 		goto out;
 
 	mlx5e_deactivate_rq(rq);
+	if (xskrq)
+		mlx5e_deactivate_rq(xskrq);
+
 	err = mlx5e_wait_for_icosq_flush(icosq);
 	if (err)
 		goto out;
@@ -94,15 +104,28 @@ static int mlx5e_rx_reporter_err_icosq_cqe_recover(void *ctx)
 		goto out;
 
 	mlx5e_reset_icosq_cc_pc(icosq);
+
 	mlx5e_free_rx_in_progress_descs(rq);
+	if (xskrq)
+		mlx5e_free_rx_in_progress_descs(xskrq);
+
 	clear_bit(MLX5E_SQ_STATE_RECOVERING, &icosq->state);
 	mlx5e_activate_icosq(icosq);
-	mlx5e_activate_rq(rq);
 
+	mlx5e_activate_rq(rq);
 	rq->stats->recover++;
+
+	if (xskrq) {
+		mlx5e_activate_rq(xskrq);
+		xskrq->stats->recover++;
+	}
+
+	mutex_unlock(&icosq->channel->icosq_recovery_lock);
+
 	return 0;
 out:
 	clear_bit(MLX5E_SQ_STATE_RECOVERING, &icosq->state);
+	mutex_unlock(&icosq->channel->icosq_recovery_lock);
 	return err;
 }
 
@@ -703,6 +726,16 @@ void mlx5e_reporter_icosq_cqe_err(struct mlx5e_icosq *icosq)
 	mlx5e_health_report(priv, priv->rx_reporter, err_str, &err_ctx);
 }
 
+void mlx5e_reporter_icosq_suspend_recovery(struct mlx5e_channel *c)
+{
+	mutex_lock(&c->icosq_recovery_lock);
+}
+
+void mlx5e_reporter_icosq_resume_recovery(struct mlx5e_channel *c)
+{
+	mutex_unlock(&c->icosq_recovery_lock);
+}
+
 static const struct devlink_health_reporter_ops mlx5_rx_reporter_ops = {
 	.name = "rx",
 	.recover = mlx5e_rx_reporter_recover,

drivers/net/ethernet/mellanox/mlx5/core/en/xsk/setup.c

@@ -4,6 +4,7 @@
 #include "setup.h"
 #include "en/params.h"
 #include "en/txrx.h"
+#include "en/health.h"
 
 /* It matches XDP_UMEM_MIN_CHUNK_SIZE, but as this constant is private and may
  * change unexpectedly, and mlx5e has a minimum valid stride size for striding
@@ -170,7 +171,13 @@ void mlx5e_close_xsk(struct mlx5e_channel *c)
 
 void mlx5e_activate_xsk(struct mlx5e_channel *c)
 {
+	/* ICOSQ recovery deactivates RQs. Suspend the recovery to avoid
+	 * activating XSKRQ in the middle of recovery.
+	 */
+	mlx5e_reporter_icosq_suspend_recovery(c);
 	set_bit(MLX5E_RQ_STATE_ENABLED, &c->xskrq.state);
+	mlx5e_reporter_icosq_resume_recovery(c);
+
 	/* TX queue is created active. */
 
 	spin_lock_bh(&c->async_icosq_lock);
@@ -180,6 +187,13 @@ void mlx5e_activate_xsk(struct mlx5e_channel *c)
 
 void mlx5e_deactivate_xsk(struct mlx5e_channel *c)
 {
-	mlx5e_deactivate_rq(&c->xskrq);
+	/* ICOSQ recovery may reactivate XSKRQ if clear_bit is called in the
+	 * middle of recovery. Suspend the recovery to avoid it.
+	 */
+	mlx5e_reporter_icosq_suspend_recovery(c);
+	clear_bit(MLX5E_RQ_STATE_ENABLED, &c->xskrq.state);
+	mlx5e_reporter_icosq_resume_recovery(c);
+	synchronize_net(); /* Sync with NAPI to prevent mlx5e_post_rx_wqes. */
+
 	/* TX queue is disabled on close. */
 }

drivers/net/ethernet/mellanox/mlx5/core/en_main.c

@@ -911,8 +911,6 @@ void mlx5e_deactivate_rq(struct mlx5e_rq *rq)
 void mlx5e_close_rq(struct mlx5e_rq *rq)
 {
 	cancel_work_sync(&rq->dim.work);
-	if (rq->icosq)
-		cancel_work_sync(&rq->icosq->recover_work);
 	cancel_work_sync(&rq->recover_work);
 	mlx5e_destroy_rq(rq);
 	mlx5e_free_rx_descs(rq);
@@ -1875,6 +1873,8 @@ static int mlx5e_open_queues(struct mlx5e_channel *c,
 	if (err)
 		goto err_close_xdpsq_cq;
 
+	mutex_init(&c->icosq_recovery_lock);
+
 	err = mlx5e_open_icosq(c, params, &cparam->icosq, &c->icosq);
 	if (err)
 		goto err_close_async_icosq;
@@ -1943,9 +1943,12 @@ static void mlx5e_close_queues(struct mlx5e_channel *c)
 	mlx5e_close_xdpsq(&c->xdpsq);
 	if (c->xdp)
 		mlx5e_close_xdpsq(&c->rq_xdpsq);
+	/* The same ICOSQ is used for UMRs for both RQ and XSKRQ. */
+	cancel_work_sync(&c->icosq.recover_work);
 	mlx5e_close_rq(&c->rq);
 	mlx5e_close_sqs(c);
 	mlx5e_close_icosq(&c->icosq);
+	mutex_destroy(&c->icosq_recovery_lock);
 	mlx5e_close_icosq(&c->async_icosq);
 	if (c->xdp)
 		mlx5e_close_cq(&c->rq_xdpsq.cq);