lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date:   Thu,  9 Sep 2021 11:28:45 +0200
From:   Íñigo Huguet <ihuguet@...hat.com>
To:     ecree.xilinx@...il.com, habetsm.xilinx@...il.com,
        davem@...emloft.net, kuba@...nel.org
Cc:     ast@...nel.org, daniel@...earbox.net, hawk@...nel.org,
        john.fastabend@...il.com, netdev@...r.kernel.org,
        linux-kernel@...r.kernel.org, bpf@...r.kernel.org,
        Íñigo Huguet <ihuguet@...hat.com>
Subject: [PATCH net 1/2] sfc: fallback for lack of xdp tx queues

If there are not enough resources to allocate one TX queue per core for
XDP TX it was completely disabled.

This patch implements a fallback solution for sharing the available
queues using __netif_tx_lock for synchronization. In the normal case that
there is one TX queue per CPU, no locking is done, as it was before.

With this fallback solution, XDP TX will work in much more cases that
were failing, specially in machines with many CPUs. It's hard for XDP
users to know what features are supported across different NICs and
configurations, so they will benefit on having wider support.

Signed-off-by: Íñigo Huguet <ihuguet@...hat.com>
---
 drivers/net/ethernet/sfc/efx_channels.c | 54 ++++++++++++++++++++-----
 drivers/net/ethernet/sfc/net_driver.h   |  8 ++++
 drivers/net/ethernet/sfc/tx.c           | 21 ++++++----
 3 files changed, 64 insertions(+), 19 deletions(-)

diff --git a/drivers/net/ethernet/sfc/efx_channels.c b/drivers/net/ethernet/sfc/efx_channels.c
index e5b0d795c301..e7849f1260a1 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) {
+	if (n_channels >= max_channels) {
+		efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_DISABLED;
 		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;
+			  "XDP_TX and XDP_REDIRECT will not work on this interface\n");
 	} else if (n_channels + n_xdp_tx > efx->max_vis) {
+		efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_DISABLED;
 		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;
+			  "XDP_TX and XDP_REDIRECT will not work on this interface\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_DISABLED) {
 		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 = 0;
 	}
 
 	if (vec_count < n_channels) {
@@ -921,7 +935,25 @@ int efx_set_channels(struct efx_nic *efx)
 			}
 		}
 	}
-	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 less XDP TX queues than CPUs, assign the already existing
+	 * queues to the exceeding CPUs (this means that we will have to use
+	 * locking when transmitting with XDP)
+	 */
+	next_queue = 0;
+	while (xdp_queue_number < efx->xdp_tx_queue_count) {
+		tx_queue = efx->xdp_tx_queues[next_queue++];
+		channel = tx_queue->channel;
+		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 = 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..e731c766f709 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_DISABLED	/* xdp tx not available */
+};
+
 /**
  * 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..c7afd6cda902 100644
--- a/drivers/net/ethernet/sfc/tx.c
+++ b/drivers/net/ethernet/sfc/tx.c
@@ -430,21 +430,23 @@ int efx_xdp_tx_buffers(struct efx_nic *efx, int n, struct xdp_frame **xdpfs,
 	int cpu;
 	int i;
 
-	cpu = raw_smp_processor_id();
+	if (unlikely(efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_DISABLED))
+		return -EINVAL;
+	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 (!n)
-		return 0;
+	if (efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED)
+		HARD_TX_LOCK(efx->net_dev, tx_queue->core_txq, cpu);
 
 	/* Check for available space. We should never need multiple
 	 * descriptors per frame.
@@ -484,6 +486,9 @@ 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);
 
+	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;
 }
 
-- 
2.31.1

Powered by blists - more mailing lists

Powered by Openwall GNU/*/Linux Powered by OpenVZ