Skip to contentSkip to conversation

gve: Add modify ring size support

PATCH In review Alex Berliner 3 participants 3 messages Apr 18, 2023 – Apr 19, 2023

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 gve: Add modify ring size support

Add driver support and ethtool bindings for changing the size of the driver's descriptor rings.

- Add logic to change ring size via ethtool
- Handle ring sizes being provided by ring size device option
- Add structures to store max ring size
- Consolidate duplicate variables for storing ring size

---
 drivers/net/ethernet/google/gve/gve.h         | 26 ++++--
 drivers/net/ethernet/google/gve/gve_adminq.c  | 85 +++++++++++++------
 drivers/net/ethernet/google/gve/gve_adminq.h  | 18 +++-
 drivers/net/ethernet/google/gve/gve_ethtool.c | 79 ++++++++++++++++-
 drivers/net/ethernet/google/gve/gve_main.c    | 43 ++++++++--
 drivers/net/ethernet/google/gve/gve_rx.c      | 10 +--
 drivers/net/ethernet/google/gve/gve_rx_dqo.c  |  3 +-
 drivers/net/ethernet/google/gve/gve_tx_dqo.c  |  2 +-
 8 files changed, 206 insertions(+), 60 deletions(-)
drivers/net/ethernet/google/gve/gve.h +18 -8
index e214b51d3c8b..e2196646554f 100644
@@ -40,6 +40,18 @@
 #define NIC_TX_STATS_REPORT_NUM	0
 #define NIC_RX_STATS_REPORT_NUM	4
 
+/* Experiment derived */
+#define GVE_TX_PAGE_COUNT 64
+
+/* Minimum descriptor ring size in bytes */
+#define GVE_RING_SIZE_MIN 4096
+
+/* Sanity check min ring element length for tx and rx queues */
+#define GVE_RING_LENGTH_LIMIT_MIN 64
+
+/* Sanity check max ring element length for tx and rx queues */
+#define GVE_RING_LENGTH_LIMIT_MAX 2048
+
 #define GVE_DATA_SLOT_ADDR_PAGE_MASK (~(PAGE_SIZE - 1))
 
 /* PTYPEs are always 10 bits. */
@@ -504,11 +516,6 @@ struct gve_qpl_config {
 	unsigned long *qpl_id_map; /* bitmap of used qpl ids */
 };
 
-struct gve_options_dqo_rda {
-	u16 tx_comp_ring_entries; /* number of tx_comp descriptors */
-	u16 rx_buff_ring_entries; /* number of rx_buff descriptors */
-};
-
 struct gve_irq_db {
 	__be32 index;
 } ____cacheline_aligned;
@@ -550,13 +557,14 @@ struct gve_priv {
 	u16 num_event_counters;
 	u16 tx_desc_cnt; /* num desc per ring */
 	u16 rx_desc_cnt; /* num desc per ring */
-	u16 tx_pages_per_qpl; /* tx buffer length */
-	u16 rx_data_slot_cnt; /* rx buffer length */
+	u16 max_rx_desc_cnt; /* max num desc per rx ring */
+	u16 max_tx_desc_cnt; /* max num desc per tx ring */
 	u64 max_registered_pages;
 	u64 num_registered_pages; /* num pages registered with NIC */
 	struct bpf_prog *xdp_prog; /* XDP BPF program */
 	u32 rx_copybreak; /* copy packets smaller than this */
 	u16 default_num_queues; /* default num queues to set up */
+	bool modify_ringsize_enabled;
 
 	u16 num_xdp_queues;
 	struct gve_queue_config tx_cfg;
@@ -622,7 +630,6 @@ struct gve_priv {
 	u64 link_speed;
 	bool up_before_suspend; /* True if dev was up before suspend */
 
-	struct gve_options_dqo_rda options_dqo_rda;
 	struct gve_ptype_lut *ptype_lut_dqo;
 
 	/* Must be a power of two. */
@@ -961,6 +968,9 @@ int gve_adjust_queues(struct gve_priv *priv,
 		      struct gve_queue_config new_tx_config);
 /* report stats handling */
 void gve_handle_report_stats(struct gve_priv *priv);
+int gve_adjust_ring_sizes(struct gve_priv *priv,
+			  int new_tx_desc_cnt,
+			  int new_rx_desc_cnt);
 /* exported by ethtool.c */
 extern const struct ethtool_ops gve_ethtool_ops;
 /* needed by ethtool */
drivers/net/ethernet/google/gve/gve_adminq.c +57 -28

200 lines not on this page · raw message

drivers/net/ethernet/google/gve/gve_adminq.h +14 -4
index f894beb3deaf..1ff55f957427 100644
@@ -71,12 +71,12 @@ struct gve_device_descriptor {
 	__be16 default_num_queues;
 	__be16 mtu;
 	__be16 counters;
-	__be16 tx_pages_per_qpl;
+	__be16 reserved2;
 	__be16 rx_pages_per_qpl;
 	u8  mac[ETH_ALEN];
 	__be16 num_device_options;
 	__be16 total_length;
-	u8  reserved2[6];
+	u8  reserved3[6];
 };
 
 static_assert(sizeof(struct gve_device_descriptor) == 40);
@@ -103,12 +103,19 @@ static_assert(sizeof(struct gve_device_option_gqi_qpl) == 4);
 
 struct gve_device_option_dqo_rda {
 	__be32 supported_features_mask;
-	__be16 tx_comp_ring_entries;
-	__be16 rx_buff_ring_entries;
+	__be32 reserved;
 };
 
 static_assert(sizeof(struct gve_device_option_dqo_rda) == 8);
 
+struct gve_device_option_modify_ring {
+	__be32 supported_features_mask;
+	__be16 max_rx_ring_size;
+	__be16 max_tx_ring_size;
+};
+
+static_assert(sizeof(struct gve_device_option_modify_ring) == 8);
+
 struct gve_device_option_jumbo_frames {
 	__be32 supported_features_mask;
 	__be16 max_mtu;
@@ -130,6 +137,7 @@ enum gve_dev_opt_id {
 	GVE_DEV_OPT_ID_GQI_RDA = 0x2,
 	GVE_DEV_OPT_ID_GQI_QPL = 0x3,
 	GVE_DEV_OPT_ID_DQO_RDA = 0x4,
+	GVE_DEV_OPT_ID_MODIFY_RING = 0x6,
 	GVE_DEV_OPT_ID_JUMBO_FRAMES = 0x8,
 };
 
@@ -138,10 +146,12 @@ enum gve_dev_opt_req_feat_mask {
 	GVE_DEV_OPT_REQ_FEAT_MASK_GQI_RDA = 0x0,
 	GVE_DEV_OPT_REQ_FEAT_MASK_GQI_QPL = 0x0,
 	GVE_DEV_OPT_REQ_FEAT_MASK_DQO_RDA = 0x0,
+	GVE_DEV_OPT_REQ_FEAT_MASK_MODIFY_RING = 0x0,
 	GVE_DEV_OPT_REQ_FEAT_MASK_JUMBO_FRAMES = 0x0,
 };
 
 enum gve_sup_feature_mask {
+	GVE_SUP_MODIFY_RING_MASK  = 1 << 0,
 	GVE_SUP_JUMBO_FRAMES_MASK = 1 << 2,
 };
 
drivers/net/ethernet/google/gve/gve_ethtool.c +76 -3

95 lines not on this page · raw message

drivers/net/ethernet/google/gve/gve_main.c +37 -6

73 lines not on this page · raw message

drivers/net/ethernet/google/gve/gve_rx.c +2 -8
index d1da7413dc4d..e52c048fc846 100644
@@ -197,9 +197,9 @@ 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");
@@ -209,7 +209,7 @@ static int gve_rx_alloc_ring(struct gve_priv *priv, int idx)
 	rx->gve = priv;
 	rx->q_num = idx;
 
-	slots = priv->rx_data_slot_cnt;
+	slots = priv->rx_desc_cnt;
 	rx->mask = slots - 1;
 	rx->data.raw_addressing = priv->queue_format == GVE_GQI_RDA_FORMAT;
 
@@ -256,12 +256,6 @@ 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) {
drivers/net/ethernet/google/gve/gve_rx_dqo.c +1 -2
index e57b73eb70f6..90344e6fdff7 100644
@@ -228,8 +228,7 @@ static int gve_rx_alloc_ring_dqo(struct gve_priv *priv, int idx)
 	size_t size;
 	int i;
 
-	const u32 buffer_queue_slots =
-		priv->options_dqo_rda.rx_buff_ring_entries;
+	const u32 buffer_queue_slots = priv->rx_desc_cnt;
 	const u32 completion_queue_slots = priv->rx_desc_cnt;
 
 	netif_dbg(priv, drv, priv->dev, "allocating rx ring DQO\n");
drivers/net/ethernet/google/gve/gve_tx_dqo.c +1 -1
index b76143bfd594..107eca0b5c1c 100644
@@ -154,7 +154,7 @@ static int gve_tx_alloc_ring_dqo(struct gve_priv *priv, int idx)
 
 	/* Queue sizes must be a power of 2 */
 	tx->mask = priv->tx_desc_cnt - 1;
-	tx->dqo.complq_mask = priv->options_dqo_rda.tx_comp_ring_entries - 1;
+	tx->dqo.complq_mask = priv->tx_desc_cnt - 1;
 
 	/* The max number of pending packets determines the maximum number of
 	 * descriptors which maybe written to the completion queue.
Signature · 1 line
2.40.0.634.g4ca3ef3211-goog
#1 lore raw
Quoted 58 lines · Alex Berliner
drivers/net/ethernet/google/gve/gve_ethtool.c +57 -0
+static int gve_set_ringparam(struct net_device *netdev,
+                            struct ethtool_ringparam *cmd,
+                            struct kernel_ethtool_ringparam *kernel_cmd,
+                            struct netlink_ext_ack *extack)
+{
+       struct gve_priv *priv = netdev_priv(netdev);
+       int old_rx_desc_cnt = priv->rx_desc_cnt;
+       int old_tx_desc_cnt = priv->tx_desc_cnt;
+       int new_tx_desc_cnt = cmd->tx_pending;
+       int new_rx_desc_cnt = cmd->rx_pending;
+       int new_max_registered_pages =
+               new_rx_desc_cnt * gve_num_rx_qpls(priv) +
+                       GVE_TX_PAGE_COUNT * gve_num_tx_qpls(priv);
+
+       if (!priv->modify_ringsize_enabled) {
+               dev_err(&priv->pdev->dev, "Modify ringsize disabled\n");
+               return -EINVAL;
+       }
+
+       if (new_tx_desc_cnt < GVE_RING_LENGTH_LIMIT_MIN ||
+               new_rx_desc_cnt < GVE_RING_LENGTH_LIMIT_MIN) {
+               dev_err(&priv->pdev->dev, "Ring size cannot be less than %d\n",
+                       GVE_RING_LENGTH_LIMIT_MIN);
+               return -EINVAL;
+       }
+
+       if (new_tx_desc_cnt > GVE_RING_LENGTH_LIMIT_MAX ||
+               new_rx_desc_cnt > GVE_RING_LENGTH_LIMIT_MAX) {
+               dev_err(&priv->pdev->dev,
+                       "Ring size cannot be greater than %d\n",
+                       GVE_RING_LENGTH_LIMIT_MAX);
+               return -EINVAL;
+       }
+
+       /* Ring size must be a power of 2, will fail if passed values are not
+        * In the future we may want to update to round down to the
+        * closest valid ring size
+        */
+       if ((new_tx_desc_cnt & (new_tx_desc_cnt - 1)) != 0 ||
+               (new_rx_desc_cnt & (new_rx_desc_cnt - 1)) != 0) {
+               dev_err(&priv->pdev->dev, "Ring size must be a power of 2\n");
+               return -EINVAL;
+       }
+
+       if (new_tx_desc_cnt > priv->max_tx_desc_cnt) {
+               dev_err(&priv->pdev->dev,
+                       "Tx ring size passed %d is larger than max tx ring size %u\n",
+                       new_tx_desc_cnt, priv->max_tx_desc_cnt);
+               return -EINVAL;
+       }
+
+       if (new_rx_desc_cnt > priv->max_rx_desc_cnt) {
+               dev_err(&priv->pdev->dev,
+                       "Rx ring size passed %d is larger than max rx ring size %u\n",
+                       new_rx_desc_cnt, priv->max_rx_desc_cnt);
+               return -EINVAL;
+       }

drivers/net/ethernet/google/gve/gve_ethtool.c static void gve_get_ringparam(struct net_device *netdev, · line 550

Quoted 10 lines
drivers/net/ethernet/google/gve/gve_ethtool.c +10 -0
+
+       if (new_max_registered_pages > priv->max_registered_pages) {
+               dev_err(&priv->pdev->dev,
+                               "Allocating too many pages %d; max %llu",
+                               new_max_registered_pages,
+                               priv->max_registered_pages);
+               return -EINVAL;
+       }
+
+       // Nothing to change return success

I think we should have /* */ comments?
Also, as is evident from the checkpatch report, the alignment of the parenthesis needs to be fixed.

drivers/net/ethernet/google/gve/gve_ethtool.c static void gve_get_ringparam(struct net_device *netdev, · line 552

Quoted 2 lines
drivers/net/ethernet/google/gve/gve_ethtool.c +2 -0
+       if (new_tx_desc_cnt == old_tx_desc_cnt && new_rx_desc_cnt == old_rx_desc_cnt)
+               return 0;

Having this condition right at the beginning can avoid unnecessary checks?

Quoted 6 lines
drivers/net/ethernet/google/gve/gve_ethtool.c +3 -1
+
+       return gve_adjust_ring_sizes(priv, new_tx_desc_cnt, new_rx_desc_cnt);
+}
--

2.40.0.634.g4ca3ef3211-goog

maintainer
#2 lore raw
In reply to Alex Berliner

drivers/net/ethernet/google/gve/gve_main.c static int gve_close(struct net_device *dev) · line 1402

Quoted 9 lines · On Tue, 18 Apr 2023 22:13:13 +0000 Alex Berliner wrote:
drivers/net/ethernet/google/gve/gve_main.c +9 -0
+		err = gve_close(priv->dev);
+		if (err)
+			return err;
+		priv->tx_desc_cnt = new_tx_desc_cnt;
+		priv->rx_desc_cnt = new_rx_desc_cnt;
+
+		err = gve_open(priv->dev);
+		if (err)
+			goto err;

please don't close / open. if the machine is low on memory it will fall off the network. I think I had a similar complaint about xdp patches for gve, it'd be great if people working on gve within google talked to each other.