Merge branch 'gve-add-support-for-non-4k-page-sizes'

John Fraker says:

====================
gve: Add support for non-4k page sizes.

This patch series adds support for non-4k page sizes to the driver. Prior
to this patch series, the driver assumes a 4k page size in many small
ways, and will crash in a kernel compiled for a different page size.

This changeset aims to be a minimal changeset that unblocks certain arm
platforms with large page sizes.
====================

Link: https://lore.kernel.org/r/20231128002648.320892-1-jfraker@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

Author: kuba-moo

drivers/net/ethernet/google/gve/gve.h

@@ -8,6 +8,7 @@
 #define _GVE_H_
 
 #include <linux/dma-mapping.h>
+#include <linux/dmapool.h>
 #include <linux/netdevice.h>
 #include <linux/pci.h>
 #include <linux/u64_stats_sync.h>
@@ -41,12 +42,16 @@
 #define NIC_TX_STATS_REPORT_NUM	0
 #define NIC_RX_STATS_REPORT_NUM	4
 
+#define GVE_ADMINQ_BUFFER_SIZE 4096
+
 #define GVE_DATA_SLOT_ADDR_PAGE_MASK (~(PAGE_SIZE - 1))
 
 /* PTYPEs are always 10 bits. */
 #define GVE_NUM_PTYPES	1024
 
-#define GVE_RX_BUFFER_SIZE_DQO 2048
+#define GVE_DEFAULT_RX_BUFFER_SIZE 2048
+
+#define GVE_DEFAULT_RX_BUFFER_OFFSET 2048
 
 #define GVE_XDP_ACTIONS 5
 
@@ -672,6 +677,7 @@ struct gve_priv {
 	/* Admin queue - see gve_adminq.h*/
 	union gve_adminq_command *adminq;
 	dma_addr_t adminq_bus_addr;
+	struct dma_pool *adminq_pool;
 	u32 adminq_mask; /* masks prod_cnt to adminq size */
 	u32 adminq_prod_cnt; /* free-running count of AQ cmds executed */
 	u32 adminq_cmd_fail; /* free-running count of AQ cmds failed */

drivers/net/ethernet/google/gve/gve_adminq.c

@@ -194,12 +194,19 @@ gve_process_device_options(struct gve_priv *priv,
 
 int gve_adminq_alloc(struct device *dev, struct gve_priv *priv)
 {
-	priv->adminq = dma_alloc_coherent(dev, PAGE_SIZE,
-					  &priv->adminq_bus_addr, GFP_KERNEL);
-	if (unlikely(!priv->adminq))
+	priv->adminq_pool = dma_pool_create("adminq_pool", dev,
+					    GVE_ADMINQ_BUFFER_SIZE, 0, 0);
+	if (unlikely(!priv->adminq_pool))
 		return -ENOMEM;
+	priv->adminq = dma_pool_alloc(priv->adminq_pool, GFP_KERNEL,
+				      &priv->adminq_bus_addr);
+	if (unlikely(!priv->adminq)) {
+		dma_pool_destroy(priv->adminq_pool);
+		return -ENOMEM;
+	}
 
-	priv->adminq_mask = (PAGE_SIZE / sizeof(union gve_adminq_command)) - 1;
+	priv->adminq_mask =
+		(GVE_ADMINQ_BUFFER_SIZE / sizeof(union gve_adminq_command)) - 1;
 	priv->adminq_prod_cnt = 0;
 	priv->adminq_cmd_fail = 0;
 	priv->adminq_timeouts = 0;
@@ -218,9 +225,20 @@ int gve_adminq_alloc(struct device *dev, struct gve_priv *priv)
 	priv->adminq_get_ptype_map_cnt = 0;
 
 	/* Setup Admin queue with the device */
-	iowrite32be(priv->adminq_bus_addr / PAGE_SIZE,
-		    &priv->reg_bar0->adminq_pfn);
-
+	if (priv->pdev->revision < 0x1) {
+		iowrite32be(priv->adminq_bus_addr / PAGE_SIZE,
+			    &priv->reg_bar0->adminq_pfn);
+	} else {
+		iowrite16be(GVE_ADMINQ_BUFFER_SIZE,
+			    &priv->reg_bar0->adminq_length);
+#ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
+		iowrite32be(priv->adminq_bus_addr >> 32,
+			    &priv->reg_bar0->adminq_base_address_hi);
+#endif
+		iowrite32be(priv->adminq_bus_addr,
+			    &priv->reg_bar0->adminq_base_address_lo);
+		iowrite32be(GVE_DRIVER_STATUS_RUN_MASK, &priv->reg_bar0->driver_status);
+	}
 	gve_set_admin_queue_ok(priv);
 	return 0;
 }
@@ -230,16 +248,27 @@ void gve_adminq_release(struct gve_priv *priv)
 	int i = 0;
 
 	/* Tell the device the adminq is leaving */
-	iowrite32be(0x0, &priv->reg_bar0->adminq_pfn);
-	while (ioread32be(&priv->reg_bar0->adminq_pfn)) {
-		/* If this is reached the device is unrecoverable and still
-		 * holding memory. Continue looping to avoid memory corruption,
-		 * but WARN so it is visible what is going on.
-		 */
-		if (i == GVE_MAX_ADMINQ_RELEASE_CHECK)
-			WARN(1, "Unrecoverable platform error!");
-		i++;
-		msleep(GVE_ADMINQ_SLEEP_LEN);
+	if (priv->pdev->revision < 0x1) {
+		iowrite32be(0x0, &priv->reg_bar0->adminq_pfn);
+		while (ioread32be(&priv->reg_bar0->adminq_pfn)) {
+			/* If this is reached the device is unrecoverable and still
+			 * holding memory. Continue looping to avoid memory corruption,
+			 * but WARN so it is visible what is going on.
+			 */
+			if (i == GVE_MAX_ADMINQ_RELEASE_CHECK)
+				WARN(1, "Unrecoverable platform error!");
+			i++;
+			msleep(GVE_ADMINQ_SLEEP_LEN);
+		}
+	} else {
+		iowrite32be(GVE_DRIVER_STATUS_RESET_MASK, &priv->reg_bar0->driver_status);
+		while (!(ioread32be(&priv->reg_bar0->device_status)
+				& GVE_DEVICE_STATUS_DEVICE_IS_RESET)) {
+			if (i == GVE_MAX_ADMINQ_RELEASE_CHECK)
+				WARN(1, "Unrecoverable platform error!");
+			i++;
+			msleep(GVE_ADMINQ_SLEEP_LEN);
+		}
 	}
 	gve_clear_device_rings_ok(priv);
 	gve_clear_device_resources_ok(priv);
@@ -251,7 +280,8 @@ void gve_adminq_free(struct device *dev, struct gve_priv *priv)
 	if (!gve_get_admin_queue_ok(priv))
 		return;
 	gve_adminq_release(priv);
-	dma_free_coherent(dev, PAGE_SIZE, priv->adminq, priv->adminq_bus_addr);
+	dma_pool_free(priv->adminq_pool, priv->adminq, priv->adminq_bus_addr);
+	dma_pool_destroy(priv->adminq_pool);
 	gve_clear_admin_queue_ok(priv);
 }
 
@@ -697,18 +727,7 @@ static int gve_set_desc_cnt(struct gve_priv *priv,
 			    struct gve_device_descriptor *descriptor)
 {
 	priv->tx_desc_cnt = be16_to_cpu(descriptor->tx_queue_entries);
-	if (priv->tx_desc_cnt * sizeof(priv->tx->desc[0]) < PAGE_SIZE) {
-		dev_err(&priv->pdev->dev, "Tx desc count %d too low\n",
-			priv->tx_desc_cnt);
-		return -EINVAL;
-	}
 	priv->rx_desc_cnt = be16_to_cpu(descriptor->rx_queue_entries);
-	if (priv->rx_desc_cnt * sizeof(priv->rx->desc.desc_ring[0])
-	    < PAGE_SIZE) {
-		dev_err(&priv->pdev->dev, "Rx desc count %d too low\n",
-			priv->rx_desc_cnt);
-		return -EINVAL;
-	}
 	return 0;
 }
 
@@ -778,16 +797,17 @@ int gve_adminq_describe_device(struct gve_priv *priv)
 	u16 mtu;
 
 	memset(&cmd, 0, sizeof(cmd));
-	descriptor = dma_alloc_coherent(&priv->pdev->dev, PAGE_SIZE,
-					&descriptor_bus, GFP_KERNEL);
+	descriptor = dma_pool_alloc(priv->adminq_pool, GFP_KERNEL,
+				    &descriptor_bus);
 	if (!descriptor)
 		return -ENOMEM;
 	cmd.opcode = cpu_to_be32(GVE_ADMINQ_DESCRIBE_DEVICE);
 	cmd.describe_device.device_descriptor_addr =
 						cpu_to_be64(descriptor_bus);
 	cmd.describe_device.device_descriptor_version =
 			cpu_to_be32(GVE_ADMINQ_DEVICE_DESCRIPTOR_VERSION);
-	cmd.describe_device.available_length = cpu_to_be32(PAGE_SIZE);
+	cmd.describe_device.available_length =
+		cpu_to_be32(GVE_ADMINQ_BUFFER_SIZE);
 
 	err = gve_adminq_execute_cmd(priv, &cmd);
 	if (err)
@@ -868,8 +888,7 @@ int gve_adminq_describe_device(struct gve_priv *priv)
 				      dev_op_jumbo_frames, dev_op_dqo_qpl);
 
 free_device_descriptor:
-	dma_free_coherent(&priv->pdev->dev, PAGE_SIZE, descriptor,
-			  descriptor_bus);
+	dma_pool_free(priv->adminq_pool, descriptor, descriptor_bus);
 	return err;
 }
 
@@ -898,6 +917,7 @@ int gve_adminq_register_page_list(struct gve_priv *priv,
 		.page_list_id = cpu_to_be32(qpl->id),
 		.num_pages = cpu_to_be32(num_entries),
 		.page_address_list_addr = cpu_to_be64(page_list_bus),
+		.page_size = cpu_to_be64(PAGE_SIZE),
 	};
 
 	err = gve_adminq_execute_cmd(priv, &cmd);

drivers/net/ethernet/google/gve/gve_adminq.h

@@ -219,9 +219,10 @@ struct gve_adminq_register_page_list {
 	__be32 page_list_id;
 	__be32 num_pages;
 	__be64 page_address_list_addr;
+	__be64 page_size;
 };
 
-static_assert(sizeof(struct gve_adminq_register_page_list) == 16);
+static_assert(sizeof(struct gve_adminq_register_page_list) == 24);
 
 struct gve_adminq_unregister_page_list {
 	__be32 page_list_id;

drivers/net/ethernet/google/gve/gve_ethtool.c

@@ -519,7 +519,7 @@ static int gve_set_tunable(struct net_device *netdev,
 	case ETHTOOL_RX_COPYBREAK:
 	{
 		u32 max_copybreak = gve_is_gqi(priv) ?
-			(PAGE_SIZE / 2) : priv->data_buffer_size_dqo;
+			GVE_DEFAULT_RX_BUFFER_SIZE : priv->data_buffer_size_dqo;
 
 		len = *(u32 *)value;
 		if (len > max_copybreak)

drivers/net/ethernet/google/gve/gve_main.c

@@ -1328,7 +1328,7 @@ static int gve_open(struct net_device *dev)
 		/* Hard code this for now. This may be tuned in the future for
 		 * performance.
 		 */
-		priv->data_buffer_size_dqo = GVE_RX_BUFFER_SIZE_DQO;
+		priv->data_buffer_size_dqo = GVE_DEFAULT_RX_BUFFER_SIZE;
 	}
 	err = gve_create_rings(priv);
 	if (err)
@@ -1664,7 +1664,7 @@ static int verify_xdp_configuration(struct net_device *dev)
 		return -EOPNOTSUPP;
 	}
 
-	if (dev->mtu > (PAGE_SIZE / 2) - sizeof(struct ethhdr) - GVE_RX_PAD) {
+	if (dev->mtu > GVE_DEFAULT_RX_BUFFER_SIZE - sizeof(struct ethhdr) - GVE_RX_PAD) {
 		netdev_warn(dev, "XDP is not supported for mtu %d.\n",
 			    dev->mtu);
 		return -EOPNOTSUPP;

drivers/net/ethernet/google/gve/gve_register.h

@@ -18,11 +18,20 @@ struct gve_registers {
 	__be32	adminq_event_counter;
 	u8	reserved[3];
 	u8	driver_version;
+	__be32	adminq_base_address_hi;
+	__be32	adminq_base_address_lo;
+	__be16	adminq_length;
 };
 
 enum gve_device_status_flags {
 	GVE_DEVICE_STATUS_RESET_MASK		= BIT(1),
 	GVE_DEVICE_STATUS_LINK_STATUS_MASK	= BIT(2),
 	GVE_DEVICE_STATUS_REPORT_STATS_MASK	= BIT(3),
+	GVE_DEVICE_STATUS_DEVICE_IS_RESET	= BIT(4),
+};
+
+enum gve_driver_status_flags {
+	GVE_DRIVER_STATUS_RUN_MASK		= BIT(0),
+	GVE_DRIVER_STATUS_RESET_MASK		= BIT(1),
 };
 #endif /* _GVE_REGISTER_H_ */

drivers/net/ethernet/google/gve/gve_rx.c

@@ -211,9 +211,9 @@ static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
 {
 	struct gve_rx_ring *rx = &priv->rx[idx];
 	struct device *hdev = &priv->pdev->dev;
-	u32 slots, npages;
 	int filled_pages;
 	size_t bytes;
+	u32 slots;
 	int err;
 
 	netif_dbg(priv, drv, priv->dev, "allocating rx ring\n");
@@ -270,12 +270,6 @@ static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
 
 	/* alloc rx desc ring */
 	bytes = sizeof(struct gve_rx_desc) * priv->rx_desc_cnt;
-	npages = bytes / PAGE_SIZE;
-	if (npages * PAGE_SIZE != bytes) {
-		err = -EIO;
-		goto abort_with_q_resources;
-	}
-
 	rx->desc.desc_ring = dma_alloc_coherent(hdev, bytes, &rx->desc.bus,
 						GFP_KERNEL);
 	if (!rx->desc.desc_ring) {
@@ -289,7 +283,7 @@ static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
 	/* Allocating half-page buffers allows page-flipping which is faster
 	 * than copying or allocating new pages.
 	 */
-	rx->packet_buffer_size = PAGE_SIZE / 2;
+	rx->packet_buffer_size = GVE_DEFAULT_RX_BUFFER_SIZE;
 	gve_rx_ctx_clear(&rx->ctx);
 	gve_rx_add_to_block(priv, idx);
 
@@ -405,10 +399,10 @@ static struct sk_buff *gve_rx_add_frags(struct napi_struct *napi,
 
 static void gve_rx_flip_buff(struct gve_rx_slot_page_info *page_info, __be64 *slot_addr)
 {
-	const __be64 offset = cpu_to_be64(PAGE_SIZE / 2);
+	const __be64 offset = cpu_to_be64(GVE_DEFAULT_RX_BUFFER_OFFSET);
 
 	/* "flip" to other packet buffer on this page */
-	page_info->page_offset ^= PAGE_SIZE / 2;
+	page_info->page_offset ^= GVE_DEFAULT_RX_BUFFER_OFFSET;
 	*(slot_addr) ^= offset;
 }
 
@@ -513,8 +507,7 @@ static struct sk_buff *gve_rx_copy_to_pool(struct gve_rx_ring *rx,
 		return NULL;
 
 	gve_dec_pagecnt_bias(copy_page_info);
-	copy_page_info->page_offset += rx->packet_buffer_size;
-	copy_page_info->page_offset &= (PAGE_SIZE - 1);
+	copy_page_info->page_offset ^= GVE_DEFAULT_RX_BUFFER_OFFSET;
 
 	if (copy_page_info->can_flip) {
 		/* We have used both halves of this copy page, it

drivers/net/ethernet/google/gve/gve_tx.c

@@ -819,7 +819,7 @@ int gve_xdp_xmit_one(struct gve_priv *priv, struct gve_tx_ring *tx,
 	return 0;
 }
 
-#define GVE_TX_START_THRESH	PAGE_SIZE
+#define GVE_TX_START_THRESH	4096
 
 static int gve_clean_tx_done(struct gve_priv *priv, struct gve_tx_ring *tx,
 			     u32 to_do, bool try_to_wake)