net: ll_temac: Fix RX buffer descriptor handling on GFP_ATOMIC pressure

Failures caused by GFP_ATOMIC memory pressure have been observed, and
due to the missing error handling, results in kernel crash such as

[1876998.350133] kernel BUG at mm/slub.c:3952!
[1876998.350141] invalid opcode: 0000 [#1] PREEMPT SMP PTI
[1876998.350147] CPU: 2 PID: 0 Comm: swapper/2 Not tainted 5.3.0-scnxt #1
[1876998.350150] Hardware name: N/A N/A/COMe-bIP2, BIOS CCR2R920 03/01/2017
[1876998.350160] RIP: 0010:kfree+0x1ca/0x220
[1876998.350164] Code: 85 db 74 49 48 8b 95 68 01 00 00 48 31 c2 48 89 10 e9 d7 fe ff ff 49 8b 04 24 a9 00 00 01 00 75 0b 49 8b 44 24 08 a8 01 75 02 <0f> 0b 49 8b 04 24 31 f6 a9 00 00 01 00 74 06 41 0f b6 74 24
 5b
[1876998.350172] RSP: 0018:ffffc900000f0df0 EFLAGS: 00010246
[1876998.350177] RAX: ffffea00027f0708 RBX: ffff888008d78000 RCX: 0000000000391372
[1876998.350181] RDX: 0000000000000000 RSI: ffffe8ffffd01400 RDI: ffff888008d78000
[1876998.350185] RBP: ffff8881185a5d00 R08: ffffc90000087dd8 R09: 000000000000280a
[1876998.350189] R10: 0000000000000002 R11: 0000000000000000 R12: ffffea0000235e00
[1876998.350193] R13: ffff8881185438a0 R14: 0000000000000000 R15: ffff888118543870
[1876998.350198] FS:  0000000000000000(0000) GS:ffff88811f300000(0000) knlGS:0000000000000000
[1876998.350203] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
s#1 Part1
[1876998.350206] CR2: 00007f8dac7b09f0 CR3: 000000011e20a006 CR4: 00000000001606e0
[1876998.350210] Call Trace:
[1876998.350215]  <IRQ>
[1876998.350224]  ? __netif_receive_skb_core+0x70a/0x920
[1876998.350229]  kfree_skb+0x32/0xb0
[1876998.350234]  __netif_receive_skb_core+0x70a/0x920
[1876998.350240]  __netif_receive_skb_one_core+0x36/0x80
[1876998.350245]  process_backlog+0x8b/0x150
[1876998.350250]  net_rx_action+0xf7/0x340
[1876998.350255]  __do_softirq+0x10f/0x353
[1876998.350262]  irq_exit+0xb2/0xc0
[1876998.350265]  do_IRQ+0x77/0xd0
[1876998.350271]  common_interrupt+0xf/0xf
[1876998.350274]  </IRQ>

In order to handle such failures more graceful, this change splits the
receive loop into one for consuming the received buffers, and one for
allocating new buffers.

When GFP_ATOMIC allocations fail, the receive will continue with the
buffers that is still there, and with the expectation that the allocations
will succeed in a later call to receive.

Fixes: 92744989533c ("net: add Xilinx ll_temac device driver")
Signed-off-by: Esben Haabendal <esben@geanix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

2 files changed, 82 insertions, 31 deletions

diff --git a/drivers/net/ethernet/xilinx/ll_temac.h b/drivers/net/ethernet/xilinx/ll_temac.h
index 276292bca334..99fe059e5c7f 100644
--- a/drivers/net/ethernet/xilinx/ll_temac.h
+++ b/drivers/net/ethernet/xilinx/ll_temac.h
@@ -375,6 +375,7 @@ struct temac_local {
 	int tx_bd_next;
 	int tx_bd_tail;
 	int rx_bd_ci;
+	int rx_bd_tail;
 
 	/* DMA channel control setup */
 	u32 tx_chnl_ctrl;
diff --git a/drivers/net/ethernet/xilinx/ll_temac_main.c b/drivers/net/ethernet/xilinx/ll_temac_main.c
index c368c3914bda..255207f2fd27 100644
--- a/drivers/net/ethernet/xilinx/ll_temac_main.c
+++ b/drivers/net/ethernet/xilinx/ll_temac_main.c
@@ -389,12 +389,13 @@ static int temac_dma_bd_init(struct net_device *ndev)
 	lp->tx_bd_next = 0;
 	lp->tx_bd_tail = 0;
 	lp->rx_bd_ci = 0;
+	lp->rx_bd_tail = RX_BD_NUM - 1;
 
 	/* Enable RX DMA transfers */
 	wmb();
 	lp->dma_out(lp, RX_CURDESC_PTR,  lp->rx_bd_p);
 	lp->dma_out(lp, RX_TAILDESC_PTR,
-		       lp->rx_bd_p + (sizeof(*lp->rx_bd_v) * (RX_BD_NUM - 1)));
+		       lp->rx_bd_p + (sizeof(*lp->rx_bd_v) * lp->rx_bd_tail));
 
 	/* Prepare for TX DMA transfer */
 	lp->dma_out(lp, TX_CURDESC_PTR, lp->tx_bd_p);
@@ -923,27 +924,41 @@ temac_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 static void ll_temac_recv(struct net_device *ndev)
 {
 	struct temac_local *lp = netdev_priv(ndev);
-	struct sk_buff *skb, *new_skb;
-	unsigned int bdstat;
-	struct cdmac_bd *cur_p;
-	dma_addr_t tail_p, skb_dma_addr;
-	int length;
 	unsigned long flags;
+	int rx_bd;
+	bool update_tail = false;
 
 	spin_lock_irqsave(&lp->rx_lock, flags);
 
-	tail_p = lp->rx_bd_p + sizeof(*lp->rx_bd_v) * lp->rx_bd_ci;
-	cur_p = &lp->rx_bd_v[lp->rx_bd_ci];
-
-	bdstat = be32_to_cpu(cur_p->app0);
-	while ((bdstat & STS_CTRL_APP0_CMPLT)) {
+	/* Process all received buffers, passing them on network
+	 * stack.  After this, the buffer descriptors will be in an
+	 * un-allocated stage, where no skb is allocated for it, and
+	 * they are therefore not available for TEMAC/DMA.
+	 */
+	do {
+		struct cdmac_bd *bd = &lp->rx_bd_v[lp->rx_bd_ci];
+		struct sk_buff *skb = lp->rx_skb[lp->rx_bd_ci];
+		unsigned int bdstat = be32_to_cpu(bd->app0);
+		int length;
+
+		/* While this should not normally happen, we can end
+		 * here when GFP_ATOMIC allocations fail, and we
+		 * therefore have un-allocated buffers.
+		 */
+		if (!skb)
+			break;
 
-		skb = lp->rx_skb[lp->rx_bd_ci];
-		length = be32_to_cpu(cur_p->app4) & 0x3FFF;
+		/* Loop over all completed buffer descriptors */
+		if (!(bdstat & STS_CTRL_APP0_CMPLT))
+			break;
 
-		dma_unmap_single(ndev->dev.parent, be32_to_cpu(cur_p->phys),
+		dma_unmap_single(ndev->dev.parent, be32_to_cpu(bd->phys),
 				 XTE_MAX_JUMBO_FRAME_SIZE, DMA_FROM_DEVICE);
+		/* The buffer is not valid for DMA anymore */
+		bd->phys = 0;
+		bd->len = 0;
 
+		length = be32_to_cpu(bd->app4) & 0x3FFF;
 		skb_put(skb, length);
 		skb->protocol = eth_type_trans(skb, ndev);
 		skb_checksum_none_assert(skb);
@@ -958,39 +973,74 @@ static void ll_temac_recv(struct net_device *ndev)
 			 * (back) for proper IP checksum byte order
 			 * (be16).
 			 */
-			skb->csum = htons(be32_to_cpu(cur_p->app3) & 0xFFFF);
+			skb->csum = htons(be32_to_cpu(bd->app3) & 0xFFFF);
 			skb->ip_summed = CHECKSUM_COMPLETE;
 		}
 
 		if (!skb_defer_rx_timestamp(skb))
 			netif_rx(skb);
+		/* The skb buffer is now owned by network stack above */
+		lp->rx_skb[lp->rx_bd_ci] = NULL;
 
 		ndev->stats.rx_packets++;
 		ndev->stats.rx_bytes += length;
 
-		new_skb = netdev_alloc_skb_ip_align(ndev,
-						XTE_MAX_JUMBO_FRAME_SIZE);
-		if (!new_skb) {
-			spin_unlock_irqrestore(&lp->rx_lock, flags);
-			return;
+		rx_bd = lp->rx_bd_ci;
+		if (++lp->rx_bd_ci >= RX_BD_NUM)
+			lp->rx_bd_ci = 0;
+	} while (rx_bd != lp->rx_bd_tail);
+
+	/* Allocate new buffers for those buffer descriptors that were
+	 * passed to network stack.  Note that GFP_ATOMIC allocations
+	 * can fail (e.g. when a larger burst of GFP_ATOMIC
+	 * allocations occurs), so while we try to allocate all
+	 * buffers in the same interrupt where they were processed, we
+	 * continue with what we could get in case of allocation
+	 * failure.  Allocation of remaining buffers will be retried
+	 * in following calls.
+	 */
+	while (1) {
+		struct sk_buff *skb;
+		struct cdmac_bd *bd;
+		dma_addr_t skb_dma_addr;
+
+		rx_bd = lp->rx_bd_tail + 1;
+		if (rx_bd >= RX_BD_NUM)
+			rx_bd = 0;
+		bd = &lp->rx_bd_v[rx_bd];
+
+		if (bd->phys)
+			break;	/* All skb's allocated */
+
+		skb = netdev_alloc_skb_ip_align(ndev, XTE_MAX_JUMBO_FRAME_SIZE);
+		if (!skb) {
+			dev_warn(&ndev->dev, "skb alloc failed\n");
+			break;
 		}
 
-		cur_p->app0 = cpu_to_be32(STS_CTRL_APP0_IRQONEND);
-		skb_dma_addr = dma_map_single(ndev->dev.parent, new_skb->data,
+		skb_dma_addr = dma_map_single(ndev->dev.parent, skb->data,
 					      XTE_MAX_JUMBO_FRAME_SIZE,
 					      DMA_FROM_DEVICE);
-		cur_p->phys = cpu_to_be32(skb_dma_addr);
-		cur_p->len = cpu_to_be32(XTE_MAX_JUMBO_FRAME_SIZE);
-		lp->rx_skb[lp->rx_bd_ci] = new_skb;
+		if (WARN_ON_ONCE(dma_mapping_error(ndev->dev.parent,
+						   skb_dma_addr))) {
+			dev_kfree_skb_any(skb);
+			break;
+		}
 
-		lp->rx_bd_ci++;
-		if (lp->rx_bd_ci >= RX_BD_NUM)
-			lp->rx_bd_ci = 0;
+		bd->phys = cpu_to_be32(skb_dma_addr);
+		bd->len = cpu_to_be32(XTE_MAX_JUMBO_FRAME_SIZE);
+		bd->app0 = cpu_to_be32(STS_CTRL_APP0_IRQONEND);
+		lp->rx_skb[rx_bd] = skb;
+
+		lp->rx_bd_tail = rx_bd;
+		update_tail = true;
+	}
 
-		cur_p = &lp->rx_bd_v[lp->rx_bd_ci];
-		bdstat = be32_to_cpu(cur_p->app0);
+	/* Move tail pointer when buffers have been allocated */
+	if (update_tail) {
+		lp->dma_out(lp, RX_TAILDESC_PTR,
+			lp->rx_bd_p + sizeof(*lp->rx_bd_v) * lp->rx_bd_tail);
 	}
-	lp->dma_out(lp, RX_TAILDESC_PTR, tail_p);
 
 	spin_unlock_irqrestore(&lp->rx_lock, flags);
 }