Merge tag 'net-5.15-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net

Pull networking fixes from Jakub Kicinski:
 "Including fixes from bpf.

  Current release - regressions:

   - vhost_net: fix OoB on sendmsg() failure

   - mlx5: bridge, fix uninitialized variable usage

   - bnxt_en: fix error recovery regression

  Current release - new code bugs:

   - bpf, mm: fix lockdep warning triggered by stack_map_get_build_id_offset()

  Previous releases - regressions:

   - r6040: restore MDIO clock frequency after MAC reset

   - tcp: fix tp->undo_retrans accounting in tcp_sacktag_one()

   - dsa: flush switchdev workqueue before tearing down CPU/DSA ports

  Previous releases - always broken:

   - ptp: dp83640: don't define PAGE0, avoid compiler warning

   - igc: fix tunnel segmentation offloads

   - phylink: update SFP selected interface on advertising changes

   - stmmac: fix system hang caused by eee_ctrl_timer during suspend/resume

   - mlx5e: fix mutual exclusion between CQE compression and HW TS

  Misc:

   - bpf, cgroups: fix cgroup v2 fallback on v1/v2 mixed mode

   - sfc: fallback for lack of xdp tx queues

   - hns3: add option to turn off page pool feature"

* tag 'net-5.15-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (67 commits)
  mlxbf_gige: clear valid_polarity upon open
  igc: fix tunnel offloading
  net/{mlx5|nfp|bnxt}: Remove unnecessary RTNL lock assert
  net: wan: wanxl: define CROSS_COMPILE_M68K
  selftests: nci: replace unsigned int with int
  net: dsa: flush switchdev workqueue before tearing down CPU/DSA ports
  Revert "net: phy: Uniform PHY driver access"
  net: dsa: destroy the phylink instance on any error in dsa_slave_phy_setup
  ptp: dp83640: don't define PAGE0
  bnx2x: Fix enabling network interfaces without VFs
  Revert "Revert "ipv4: fix memory leaks in ip_cmsg_send() callers""
  tcp: fix tp->undo_retrans accounting in tcp_sacktag_one()
  net-caif: avoid user-triggerable WARN_ON(1)
  bpf, selftests: Add test case for mixed cgroup v1/v2
  bpf, selftests: Add cgroup v1 net_cls classid helpers
  bpf, cgroups: Fix cgroup v2 fallback on v1/v2 mixed mode
  bpf: Add oversize check before call kvcalloc()
  net: hns3: fix the timing issue of VF clearing interrupt sources
  net: hns3: fix the exception when query imp info
  net: hns3: disable mac in flr process
  ...

3 files changed, 103 insertions, 40 deletions

diff --git a/drivers/net/ethernet/sfc/efx_channels.c b/drivers/net/ethernet/sfc/efx_channels.c
index e5b0d795c301..3dbea028b325 100644
--- a/drivers/net/ethernet/sfc/efx_channels.c
+++ b/drivers/net/ethernet/sfc/efx_channels.c
@@ -166,32 +166,46 @@ static int efx_allocate_msix_channels(struct efx_nic *efx,
 	 * We need a channel per event queue, plus a VI per tx queue.
 	 * This may be more pessimistic than it needs to be.
 	 */
-	if (n_channels + n_xdp_ev > max_channels) {
-		netif_err(efx, drv, efx->net_dev,
-			  "Insufficient resources for %d XDP event queues (%d other channels, max %d)\n",
-			  n_xdp_ev, n_channels, max_channels);
-		netif_err(efx, drv, efx->net_dev,
-			  "XDP_TX and XDP_REDIRECT will not work on this interface");
-		efx->n_xdp_channels = 0;
-		efx->xdp_tx_per_channel = 0;
-		efx->xdp_tx_queue_count = 0;
+	if (n_channels >= max_channels) {
+		efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_BORROWED;
+		netif_warn(efx, drv, efx->net_dev,
+			   "Insufficient resources for %d XDP event queues (%d other channels, max %d)\n",
+			   n_xdp_ev, n_channels, max_channels);
+		netif_warn(efx, drv, efx->net_dev,
+			   "XDP_TX and XDP_REDIRECT might decrease device's performance\n");
 	} else if (n_channels + n_xdp_tx > efx->max_vis) {
-		netif_err(efx, drv, efx->net_dev,
-			  "Insufficient resources for %d XDP TX queues (%d other channels, max VIs %d)\n",
-			  n_xdp_tx, n_channels, efx->max_vis);
-		netif_err(efx, drv, efx->net_dev,
-			  "XDP_TX and XDP_REDIRECT will not work on this interface");
-		efx->n_xdp_channels = 0;
-		efx->xdp_tx_per_channel = 0;
-		efx->xdp_tx_queue_count = 0;
+		efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_BORROWED;
+		netif_warn(efx, drv, efx->net_dev,
+			   "Insufficient resources for %d XDP TX queues (%d other channels, max VIs %d)\n",
+			   n_xdp_tx, n_channels, efx->max_vis);
+		netif_warn(efx, drv, efx->net_dev,
+			   "XDP_TX and XDP_REDIRECT might decrease device's performance\n");
+	} else if (n_channels + n_xdp_ev > max_channels) {
+		efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_SHARED;
+		netif_warn(efx, drv, efx->net_dev,
+			   "Insufficient resources for %d XDP event queues (%d other channels, max %d)\n",
+			   n_xdp_ev, n_channels, max_channels);
+
+		n_xdp_ev = max_channels - n_channels;
+		netif_warn(efx, drv, efx->net_dev,
+			   "XDP_TX and XDP_REDIRECT will work with reduced performance (%d cpus/tx_queue)\n",
+			   DIV_ROUND_UP(n_xdp_tx, tx_per_ev * n_xdp_ev));
 	} else {
+		efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_DEDICATED;
+	}
+
+	if (efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_BORROWED) {
 		efx->n_xdp_channels = n_xdp_ev;
 		efx->xdp_tx_per_channel = tx_per_ev;
 		efx->xdp_tx_queue_count = n_xdp_tx;
 		n_channels += n_xdp_ev;
 		netif_dbg(efx, drv, efx->net_dev,
 			  "Allocating %d TX and %d event queues for XDP\n",
-			  n_xdp_tx, n_xdp_ev);
+			  n_xdp_ev * tx_per_ev, n_xdp_ev);
+	} else {
+		efx->n_xdp_channels = 0;
+		efx->xdp_tx_per_channel = 0;
+		efx->xdp_tx_queue_count = n_xdp_tx;
 	}
 
 	if (vec_count < n_channels) {
@@ -858,6 +872,20 @@ rollback:
 	goto out;
 }
 
+static inline int
+efx_set_xdp_tx_queue(struct efx_nic *efx, int xdp_queue_number,
+		     struct efx_tx_queue *tx_queue)
+{
+	if (xdp_queue_number >= efx->xdp_tx_queue_count)
+		return -EINVAL;
+
+	netif_dbg(efx, drv, efx->net_dev, "Channel %u TXQ %u is XDP %u, HW %u\n",
+		  tx_queue->channel->channel, tx_queue->label,
+		  xdp_queue_number, tx_queue->queue);
+	efx->xdp_tx_queues[xdp_queue_number] = tx_queue;
+	return 0;
+}
+
 int efx_set_channels(struct efx_nic *efx)
 {
 	struct efx_tx_queue *tx_queue;
@@ -896,20 +924,9 @@ int efx_set_channels(struct efx_nic *efx)
 			if (efx_channel_is_xdp_tx(channel)) {
 				efx_for_each_channel_tx_queue(tx_queue, channel) {
 					tx_queue->queue = next_queue++;
-
-					/* We may have a few left-over XDP TX
-					 * queues owing to xdp_tx_queue_count
-					 * not dividing evenly by EFX_MAX_TXQ_PER_CHANNEL.
-					 * We still allocate and probe those
-					 * TXQs, but never use them.
-					 */
-					if (xdp_queue_number < efx->xdp_tx_queue_count) {
-						netif_dbg(efx, drv, efx->net_dev, "Channel %u TXQ %u is XDP %u, HW %u\n",
-							  channel->channel, tx_queue->label,
-							  xdp_queue_number, tx_queue->queue);
-						efx->xdp_tx_queues[xdp_queue_number] = tx_queue;
+					rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, tx_queue);
+					if (rc == 0)
 						xdp_queue_number++;
-					}
 				}
 			} else {
 				efx_for_each_channel_tx_queue(tx_queue, channel) {
@@ -918,10 +935,35 @@ int efx_set_channels(struct efx_nic *efx)
 						  channel->channel, tx_queue->label,
 						  tx_queue->queue);
 				}
+
+				/* If XDP is borrowing queues from net stack, it must use the queue
+				 * with no csum offload, which is the first one of the channel
+				 * (note: channel->tx_queue_by_type is not initialized yet)
+				 */
+				if (efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_BORROWED) {
+					tx_queue = &channel->tx_queue[0];
+					rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, tx_queue);
+					if (rc == 0)
+						xdp_queue_number++;
+				}
 			}
 		}
 	}
-	WARN_ON(xdp_queue_number != efx->xdp_tx_queue_count);
+	WARN_ON(efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_DEDICATED &&
+		xdp_queue_number != efx->xdp_tx_queue_count);
+	WARN_ON(efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED &&
+		xdp_queue_number > efx->xdp_tx_queue_count);
+
+	/* If we have more CPUs than assigned XDP TX queues, assign the already
+	 * existing queues to the exceeding CPUs
+	 */
+	next_queue = 0;
+	while (xdp_queue_number < efx->xdp_tx_queue_count) {
+		tx_queue = efx->xdp_tx_queues[next_queue++];
+		rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, tx_queue);
+		if (rc == 0)
+			xdp_queue_number++;
+	}
 
 	rc = netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
 	if (rc)
diff --git a/drivers/net/ethernet/sfc/net_driver.h b/drivers/net/ethernet/sfc/net_driver.h
index 9b4b25704271..f6981810039d 100644
--- a/drivers/net/ethernet/sfc/net_driver.h
+++ b/drivers/net/ethernet/sfc/net_driver.h
@@ -782,6 +782,12 @@ struct efx_async_filter_insertion {
 #define EFX_RPS_MAX_IN_FLIGHT	8
 #endif /* CONFIG_RFS_ACCEL */
 
+enum efx_xdp_tx_queues_mode {
+	EFX_XDP_TX_QUEUES_DEDICATED,	/* one queue per core, locking not needed */
+	EFX_XDP_TX_QUEUES_SHARED,	/* each queue used by more than 1 core */
+	EFX_XDP_TX_QUEUES_BORROWED	/* queues borrowed from net stack */
+};
+
 /**
  * struct efx_nic - an Efx NIC
  * @name: Device name (net device name or bus id before net device registered)
@@ -820,6 +826,7 @@ struct efx_async_filter_insertion {
  *	should be allocated for this NIC
  * @xdp_tx_queue_count: Number of entries in %xdp_tx_queues.
  * @xdp_tx_queues: Array of pointers to tx queues used for XDP transmit.
+ * @xdp_txq_queues_mode: XDP TX queues sharing strategy.
  * @rxq_entries: Size of receive queues requested by user.
  * @txq_entries: Size of transmit queues requested by user.
  * @txq_stop_thresh: TX queue fill level at or above which we stop it.
@@ -979,6 +986,7 @@ struct efx_nic {
 
 	unsigned int xdp_tx_queue_count;
 	struct efx_tx_queue **xdp_tx_queues;
+	enum efx_xdp_tx_queues_mode xdp_txq_queues_mode;
 
 	unsigned rxq_entries;
 	unsigned txq_entries;
diff --git a/drivers/net/ethernet/sfc/tx.c b/drivers/net/ethernet/sfc/tx.c
index 0c6650d2e239..d16e031e95f4 100644
--- a/drivers/net/ethernet/sfc/tx.c
+++ b/drivers/net/ethernet/sfc/tx.c
@@ -428,23 +428,32 @@ int efx_xdp_tx_buffers(struct efx_nic *efx, int n, struct xdp_frame **xdpfs,
 	unsigned int len;
 	int space;
 	int cpu;
-	int i;
+	int i = 0;
 
-	cpu = raw_smp_processor_id();
+	if (unlikely(n && !xdpfs))
+		return -EINVAL;
+	if (unlikely(!n))
+		return 0;
 
-	if (!efx->xdp_tx_queue_count ||
-	    unlikely(cpu >= efx->xdp_tx_queue_count))
+	cpu = raw_smp_processor_id();
+	if (unlikely(cpu >= efx->xdp_tx_queue_count))
 		return -EINVAL;
 
 	tx_queue = efx->xdp_tx_queues[cpu];
 	if (unlikely(!tx_queue))
 		return -EINVAL;
 
-	if (unlikely(n && !xdpfs))
-		return -EINVAL;
+	if (efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED)
+		HARD_TX_LOCK(efx->net_dev, tx_queue->core_txq, cpu);
 
-	if (!n)
-		return 0;
+	/* If we're borrowing net stack queues we have to handle stop-restart
+	 * or we might block the queue and it will be considered as frozen
+	 */
+	if (efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_BORROWED) {
+		if (netif_tx_queue_stopped(tx_queue->core_txq))
+			goto unlock;
+		efx_tx_maybe_stop_queue(tx_queue);
+	}
 
 	/* Check for available space. We should never need multiple
 	 * descriptors per frame.
@@ -484,6 +493,10 @@ int efx_xdp_tx_buffers(struct efx_nic *efx, int n, struct xdp_frame **xdpfs,
 	if (flush && i > 0)
 		efx_nic_push_buffers(tx_queue);
 
+unlock:
+	if (efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED)
+		HARD_TX_UNLOCK(efx->net_dev, tx_queue->core_txq);
+
 	return i == 0 ? -EIO : i;
 }