[Patch v6 00/18] Introduce Microsoft Azure Network Adatper (MANA) PMD

[Patch v6 00/18] Introduce Microsoft Azure Network Adatper (MANA) PMD

[longli]

From: Long Li <longli@microsoft.com>

MANA is a network interface card to be used in the Azure cloud environment.
MANA provides safe access to user memory through memory registration. It has
IOMMU built into the hardware.

MANA uses IB verbs and RDMA layer to configure hardware resources. It
requires the corresponding RDMA kernel-mode and user-mode drivers.

The MANA RDMA kernel-mode driver is being reviewed at:
https://patchwork.kernel.org/project/netdevbpf/cover/1655345240-26411-1-git-send-email-longli@linuxonhyperv.com/

The MANA RDMA user-mode driver is being reviewed at:
https://github.com/linux-rdma/rdma-core/pull/1177


Long Li (18):
  net/mana: add basic driver, build environment and doc
  net/mana: add device configuration and stop
  net/mana: add function to report support ptypes
  net/mana: add link update
  net/mana: add function for device removal interrupts
  net/mana: add device info
  net/mana: add function to configure RSS
  net/mana: add function to configure RX queues
  net/mana: add function to configure TX queues
  net/mana: implement memory registration
  net/mana: implement the hardware layer operations
  net/mana: add function to start/stop TX queues
  net/mana: add function to start/stop RX queues
  net/mana: add function to receive packets
  net/mana: add function to send packets
  net/mana: add function to start/stop device
  net/mana: add function to report queue stats
  net/mana: add function to support RX interrupts

 MAINTAINERS                       |    6 +
 doc/guides/nics/features/mana.ini |   21 +
 doc/guides/nics/index.rst         |    1 +
 doc/guides/nics/mana.rst          |   66 ++
 drivers/net/mana/gdma.c           |  289 ++++++
 drivers/net/mana/mana.c           | 1449 +++++++++++++++++++++++++++++
 drivers/net/mana/mana.h           |  553 +++++++++++
 drivers/net/mana/meson.build      |   48 +
 drivers/net/mana/mp.c             |  323 +++++++
 drivers/net/mana/mr.c             |  324 +++++++
 drivers/net/mana/rx.c             |  519 +++++++++++
 drivers/net/mana/tx.c             |  405 ++++++++
 drivers/net/mana/version.map      |    3 +
 drivers/net/meson.build           |    1 +
 14 files changed, 4008 insertions(+)
 create mode 100644 doc/guides/nics/features/mana.ini
 create mode 100644 doc/guides/nics/mana.rst
 create mode 100644 drivers/net/mana/gdma.c
 create mode 100644 drivers/net/mana/mana.c
 create mode 100644 drivers/net/mana/mana.h
 create mode 100644 drivers/net/mana/meson.build
 create mode 100644 drivers/net/mana/mp.c
 create mode 100644 drivers/net/mana/mr.c
 create mode 100644 drivers/net/mana/rx.c
 create mode 100644 drivers/net/mana/tx.c
 create mode 100644 drivers/net/mana/version.map

-- 
2.17.1

[Patch v6 01/18] net/mana: add basic driver, build environment and doc

[longli]

From: Long Li <longli@microsoft.com>

MANA is a PCI device. It uses IB verbs to access hardware through the
kernel RDMA layer. This patch introduces build environment and basic
device probe functions.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v2:
Fix typos.
Make the driver build only on x86-64 and Linux.
Remove unused header files.
Change port definition to uint16_t or uint8_t (for IB).
Use getline() in place of fgets() to read and truncate a line.
v3:
Add meson build check for required functions from RDMA direct verb header file
v4:
Remove extra "\n" in logging code.
Use "r" in place of "rb" in fopen() to read text files.

 MAINTAINERS                       |   6 +
 doc/guides/nics/features/mana.ini |  10 +
 doc/guides/nics/index.rst         |   1 +
 doc/guides/nics/mana.rst          |  66 +++
 drivers/net/mana/mana.c           | 704 ++++++++++++++++++++++++++++++
 drivers/net/mana/mana.h           | 210 +++++++++
 drivers/net/mana/meson.build      |  44 ++
 drivers/net/mana/mp.c             | 235 ++++++++++
 drivers/net/mana/version.map      |   3 +
 drivers/net/meson.build           |   1 +
 10 files changed, 1280 insertions(+)
 create mode 100644 doc/guides/nics/features/mana.ini
 create mode 100644 doc/guides/nics/mana.rst
 create mode 100644 drivers/net/mana/mana.c
 create mode 100644 drivers/net/mana/mana.h
 create mode 100644 drivers/net/mana/meson.build
 create mode 100644 drivers/net/mana/mp.c
 create mode 100644 drivers/net/mana/version.map

diff --git a/MAINTAINERS b/MAINTAINERS
index 18d9edaf88..b8bda48a33 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -837,6 +837,12 @@ F: buildtools/options-ibverbs-static.sh
 F: doc/guides/nics/mlx5.rst
 F: doc/guides/nics/features/mlx5.ini
 
+Microsoft mana
+M: Long Li <longli@microsoft.com>
+F: drivers/net/mana
+F: doc/guides/nics/mana.rst
+F: doc/guides/nics/features/mana.ini
+
 Microsoft vdev_netvsc - EXPERIMENTAL
 M: Matan Azrad <matan@nvidia.com>
 F: drivers/net/vdev_netvsc/
diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
new file mode 100644
index 0000000000..b92a27374c
--- /dev/null
+++ b/doc/guides/nics/features/mana.ini
@@ -0,0 +1,10 @@
+;
+; Supported features of the 'mana' network poll mode driver.
+;
+; Refer to default.ini for the full list of available PMD features.
+;
+[Features]
+Linux                = Y
+Multiprocess aware   = Y
+Usage doc            = Y
+x86-64               = Y
diff --git a/doc/guides/nics/index.rst b/doc/guides/nics/index.rst
index 1c94caccea..2725d1d9f0 100644
--- a/doc/guides/nics/index.rst
+++ b/doc/guides/nics/index.rst
@@ -41,6 +41,7 @@ Network Interface Controller Drivers
     intel_vf
     kni
     liquidio
+    mana
     memif
     mlx4
     mlx5
diff --git a/doc/guides/nics/mana.rst b/doc/guides/nics/mana.rst
new file mode 100644
index 0000000000..40e18fe810
--- /dev/null
+++ b/doc/guides/nics/mana.rst
@@ -0,0 +1,66 @@
+..  SPDX-License-Identifier: BSD-3-Clause
+    Copyright 2022 Microsoft Corporation
+
+MANA poll mode driver library
+=============================
+
+The MANA poll mode driver library (**librte_net_mana**) implements support
+for Microsoft Azure Network Adapter VF in SR-IOV context.
+
+Features
+--------
+
+Features of the MANA Ethdev PMD are:
+
+Prerequisites
+-------------
+
+This driver relies on external libraries and kernel drivers for resources
+allocations and initialization. The following dependencies are not part of
+DPDK and must be installed separately:
+
+- **libibverbs** (provided by rdma-core package)
+
+  User space verbs framework used by librte_net_mana. This library provides
+  a generic interface between the kernel and low-level user space drivers
+  such as libmana.
+
+  It allows slow and privileged operations (context initialization, hardware
+  resources allocations) to be managed by the kernel and fast operations to
+  never leave user space.
+
+- **libmana** (provided by rdma-core package)
+
+  Low-level user space driver library for Microsoft Azure Network Adapter
+  devices, it is automatically loaded by libibverbs.
+
+- **Kernel modules**
+
+  They provide the kernel-side verbs API and low level device drivers that
+  manage actual hardware initialization and resources sharing with user
+  space processes.
+
+  Unlike most other PMDs, these modules must remain loaded and bound to
+  their devices:
+
+  - mana: Ethernet device driver that provides kernel network interfaces.
+  - mana_ib: InifiniBand device driver.
+  - ib_uverbs: user space driver for verbs (entry point for libibverbs).
+
+Driver compilation and testing
+------------------------------
+
+Refer to the document :ref:`compiling and testing a PMD for a NIC <pmd_build_and_test>`
+for details.
+
+Netvsc PMD arguments
+--------------------
+
+The user can specify below argument in devargs.
+
+#.  ``mac``:
+
+    Specify the MAC address for this device. If it is set, the driver
+    probes and loads the NIC with a matching mac address. If it is not
+    set, the driver probes on all the NICs on the PCI device. The default
+    value is not set, meaning all the NICs will be probed and loaded.
diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
new file mode 100644
index 0000000000..cb59eb6882
--- /dev/null
+++ b/drivers/net/mana/mana.c
@@ -0,0 +1,704 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2022 Microsoft Corporation
+ */
+
+#include <unistd.h>
+#include <dirent.h>
+#include <fcntl.h>
+#include <sys/mman.h>
+
+#include <ethdev_driver.h>
+#include <ethdev_pci.h>
+#include <rte_kvargs.h>
+#include <rte_eal_paging.h>
+
+#include <infiniband/verbs.h>
+#include <infiniband/manadv.h>
+
+#include <assert.h>
+
+#include "mana.h"
+
+/* Shared memory between primary/secondary processes, per driver */
+struct mana_shared_data *mana_shared_data;
+const struct rte_memzone *mana_shared_mz;
+static const char *MZ_MANA_SHARED_DATA = "mana_shared_data";
+
+struct mana_shared_data mana_local_data;
+
+/* Spinlock for mana_shared_data */
+static rte_spinlock_t mana_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
+
+/* Allocate a buffer on the stack and fill it with a printf format string. */
+#define MKSTR(name, ...) \
+	int mkstr_size_##name = snprintf(NULL, 0, "" __VA_ARGS__); \
+	char name[mkstr_size_##name + 1]; \
+	\
+	memset(name, 0, mkstr_size_##name + 1); \
+	snprintf(name, sizeof(name), "" __VA_ARGS__)
+
+int mana_logtype_driver;
+int mana_logtype_init;
+
+const struct eth_dev_ops mana_dev_ops = {
+};
+
+const struct eth_dev_ops mana_dev_sec_ops = {
+};
+
+uint16_t
+mana_rx_burst_removed(void *dpdk_rxq __rte_unused,
+		      struct rte_mbuf **pkts __rte_unused,
+		      uint16_t pkts_n __rte_unused)
+{
+	rte_mb();
+	return 0;
+}
+
+uint16_t
+mana_tx_burst_removed(void *dpdk_rxq __rte_unused,
+		      struct rte_mbuf **pkts __rte_unused,
+		      uint16_t pkts_n __rte_unused)
+{
+	rte_mb();
+	return 0;
+}
+
+static const char *mana_init_args[] = {
+	"mac",
+	NULL,
+};
+
+/* Support of parsing up to 8 mac address from EAL command line */
+#define MAX_NUM_ADDRESS 8
+struct mana_conf {
+	struct rte_ether_addr mac_array[MAX_NUM_ADDRESS];
+	unsigned int index;
+};
+
+static int mana_arg_parse_callback(const char *key, const char *val,
+				   void *private)
+{
+	struct mana_conf *conf = (struct mana_conf *)private;
+	int ret;
+
+	DRV_LOG(INFO, "key=%s value=%s index=%d", key, val, conf->index);
+
+	if (conf->index >= MAX_NUM_ADDRESS) {
+		DRV_LOG(ERR, "Exceeding max MAC address");
+		return 1;
+	}
+
+	ret = rte_ether_unformat_addr(val, &conf->mac_array[conf->index]);
+	if (ret) {
+		DRV_LOG(ERR, "Invalid MAC address %s", val);
+		return ret;
+	}
+
+	conf->index++;
+
+	return 0;
+}
+
+static int mana_parse_args(struct rte_devargs *devargs, struct mana_conf *conf)
+{
+	struct rte_kvargs *kvlist;
+	unsigned int arg_count;
+	int ret = 0;
+
+	kvlist = rte_kvargs_parse(devargs->args, mana_init_args);
+	if (!kvlist) {
+		DRV_LOG(ERR, "failed to parse kvargs args=%s", devargs->args);
+		return -EINVAL;
+	}
+
+	arg_count = rte_kvargs_count(kvlist, mana_init_args[0]);
+	if (arg_count > MAX_NUM_ADDRESS) {
+		ret = -EINVAL;
+		goto free_kvlist;
+	}
+	ret = rte_kvargs_process(kvlist, mana_init_args[0],
+				 mana_arg_parse_callback, conf);
+	if (ret) {
+		DRV_LOG(ERR, "error parsing args");
+		goto free_kvlist;
+	}
+
+free_kvlist:
+	rte_kvargs_free(kvlist);
+	return ret;
+}
+
+static int get_port_mac(struct ibv_device *device, unsigned int port,
+			struct rte_ether_addr *addr)
+{
+	FILE *file;
+	int ret = 0;
+	DIR *dir;
+	struct dirent *dent;
+	unsigned int dev_port;
+	char mac[20];
+
+	MKSTR(path, "%s/device/net", device->ibdev_path);
+
+	dir = opendir(path);
+	if (!dir)
+		return -ENOENT;
+
+	while ((dent = readdir(dir))) {
+		char *name = dent->d_name;
+
+		MKSTR(filepath, "%s/%s/dev_port", path, name);
+
+		/* Ignore . and .. */
+		if ((name[0] == '.') &&
+		    ((name[1] == '\0') ||
+		     ((name[1] == '.') && (name[2] == '\0'))))
+			continue;
+
+		file = fopen(filepath, "r");
+		if (!file)
+			continue;
+
+		ret = fscanf(file, "%u", &dev_port);
+		fclose(file);
+
+		if (ret != 1)
+			continue;
+
+		/* Ethernet ports start at 0, IB port start at 1 */
+		if (dev_port == port - 1) {
+			MKSTR(filepath, "%s/%s/address", path, name);
+
+			file = fopen(filepath, "r");
+			if (!file)
+				continue;
+
+			ret = fscanf(file, "%s", mac);
+			fclose(file);
+
+			if (ret < 0)
+				break;
+
+			ret = rte_ether_unformat_addr(mac, addr);
+			if (ret)
+				DRV_LOG(ERR, "unrecognized mac addr %s", mac);
+			break;
+		}
+	}
+
+	closedir(dir);
+	return ret;
+}
+
+static int mana_ibv_device_to_pci_addr(const struct ibv_device *device,
+				       struct rte_pci_addr *pci_addr)
+{
+	FILE *file;
+	char *line = NULL;
+	size_t len = 0;
+
+	MKSTR(path, "%s/device/uevent", device->ibdev_path);
+
+	file = fopen(path, "r");
+	if (!file)
+		return -errno;
+
+	while (getline(&line, &len, file) != -1) {
+		/* Extract information. */
+		if (sscanf(line,
+			   "PCI_SLOT_NAME="
+			   "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
+			   &pci_addr->domain,
+			   &pci_addr->bus,
+			   &pci_addr->devid,
+			   &pci_addr->function) == 4) {
+			break;
+		}
+	}
+
+	free(line);
+	fclose(file);
+	return 0;
+}
+
+static int mana_proc_priv_init(struct rte_eth_dev *dev)
+{
+	struct mana_process_priv *priv;
+
+	priv = rte_zmalloc_socket("mana_proc_priv",
+				  sizeof(struct mana_process_priv),
+				  RTE_CACHE_LINE_SIZE,
+				  dev->device->numa_node);
+	if (!priv)
+		return -ENOMEM;
+
+	dev->process_private = priv;
+	return 0;
+}
+
+static int mana_map_doorbell_secondary(struct rte_eth_dev *eth_dev, int fd)
+{
+	struct mana_process_priv *priv = eth_dev->process_private;
+
+	void *addr;
+
+	addr = mmap(NULL, rte_mem_page_size(), PROT_WRITE, MAP_SHARED, fd, 0);
+	if (addr == MAP_FAILED) {
+		DRV_LOG(ERR, "Failed to map secondary doorbell port %u",
+			eth_dev->data->port_id);
+		return -ENOMEM;
+	}
+
+	DRV_LOG(INFO, "Secondary doorbell mapped to %p", addr);
+
+	priv->db_page = addr;
+
+	return 0;
+}
+
+/* Initialize shared data for the driver (all devices) */
+static int mana_init_shared_data(void)
+{
+	int ret =  0;
+	const struct rte_memzone *secondary_mz;
+
+	rte_spinlock_lock(&mana_shared_data_lock);
+
+	/* Skip if shared data is already initialized */
+	if (mana_shared_data)
+		goto exit;
+
+	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+		mana_shared_mz = rte_memzone_reserve(MZ_MANA_SHARED_DATA,
+						     sizeof(*mana_shared_data),
+						     SOCKET_ID_ANY, 0);
+		if (!mana_shared_mz) {
+			DRV_LOG(ERR, "Cannot allocate mana shared data");
+			ret = -rte_errno;
+			goto exit;
+		}
+
+		mana_shared_data = mana_shared_mz->addr;
+		memset(mana_shared_data, 0, sizeof(*mana_shared_data));
+		rte_spinlock_init(&mana_shared_data->lock);
+	} else {
+		secondary_mz = rte_memzone_lookup(MZ_MANA_SHARED_DATA);
+		if (!secondary_mz) {
+			DRV_LOG(ERR, "Cannot attach mana shared data");
+			ret = -rte_errno;
+			goto exit;
+		}
+
+		mana_shared_data = secondary_mz->addr;
+		memset(&mana_local_data, 0, sizeof(mana_local_data));
+	}
+
+exit:
+	rte_spinlock_unlock(&mana_shared_data_lock);
+
+	return ret;
+}
+
+static int mana_init_once(void)
+{
+	int ret;
+
+	ret = mana_init_shared_data();
+	if (ret)
+		return ret;
+
+	rte_spinlock_lock(&mana_shared_data->lock);
+
+	switch (rte_eal_process_type()) {
+	case RTE_PROC_PRIMARY:
+		if (mana_shared_data->init_done)
+			break;
+
+		ret = mana_mp_init_primary();
+		if (ret)
+			break;
+		DRV_LOG(ERR, "MP INIT PRIMARY");
+
+		mana_shared_data->init_done = 1;
+		break;
+
+	case RTE_PROC_SECONDARY:
+
+		if (mana_local_data.init_done)
+			break;
+
+		ret = mana_mp_init_secondary();
+		if (ret)
+			break;
+
+		DRV_LOG(ERR, "MP INIT SECONDARY");
+
+		mana_local_data.init_done = 1;
+		break;
+
+	default:
+		/* Impossible, internal error */
+		ret = -EPROTO;
+		break;
+	}
+
+	rte_spinlock_unlock(&mana_shared_data->lock);
+
+	return ret;
+}
+
+static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv __rte_unused,
+			      struct rte_pci_device *pci_dev,
+			      struct rte_ether_addr *mac_addr)
+{
+	struct ibv_device **ibv_list;
+	int ibv_idx;
+	struct ibv_context *ctx;
+	struct ibv_device_attr_ex dev_attr;
+	int num_devices;
+	int ret = 0;
+	uint8_t port;
+	struct mana_priv *priv = NULL;
+	struct rte_eth_dev *eth_dev = NULL;
+	bool found_port;
+
+	ibv_list = ibv_get_device_list(&num_devices);
+	for (ibv_idx = 0; ibv_idx < num_devices; ibv_idx++) {
+		struct ibv_device *ibdev = ibv_list[ibv_idx];
+		struct rte_pci_addr pci_addr;
+
+		DRV_LOG(INFO, "Probe device name %s dev_name %s ibdev_path %s",
+			ibdev->name, ibdev->dev_name, ibdev->ibdev_path);
+
+		if (mana_ibv_device_to_pci_addr(ibdev, &pci_addr))
+			continue;
+
+		/* Ignore if this IB device is not this PCI device */
+		if (pci_dev->addr.domain != pci_addr.domain ||
+		    pci_dev->addr.bus != pci_addr.bus ||
+		    pci_dev->addr.devid != pci_addr.devid ||
+		    pci_dev->addr.function != pci_addr.function)
+			continue;
+
+		ctx = ibv_open_device(ibdev);
+		if (!ctx) {
+			DRV_LOG(ERR, "Failed to open IB device %s",
+				ibdev->name);
+			continue;
+		}
+
+		ret = ibv_query_device_ex(ctx, NULL, &dev_attr);
+		DRV_LOG(INFO, "dev_attr.orig_attr.phys_port_cnt %u",
+			dev_attr.orig_attr.phys_port_cnt);
+		found_port = false;
+
+		for (port = 1; port <= dev_attr.orig_attr.phys_port_cnt;
+		     port++) {
+			struct ibv_parent_domain_init_attr attr = {};
+			struct rte_ether_addr addr;
+			char address[64];
+			char name[RTE_ETH_NAME_MAX_LEN];
+
+			ret = get_port_mac(ibdev, port, &addr);
+			if (ret)
+				continue;
+
+			if (mac_addr && !rte_is_same_ether_addr(&addr, mac_addr))
+				continue;
+
+			rte_ether_format_addr(address, sizeof(address), &addr);
+			DRV_LOG(INFO, "device located port %u address %s",
+				port, address);
+			found_port = true;
+
+			priv = rte_zmalloc_socket(NULL, sizeof(*priv),
+						  RTE_CACHE_LINE_SIZE,
+						  SOCKET_ID_ANY);
+			if (!priv) {
+				ret = -ENOMEM;
+				goto failed;
+			}
+
+			snprintf(name, sizeof(name), "%s_port%d",
+				 pci_dev->device.name, port);
+
+			if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
+				int fd;
+
+				eth_dev = rte_eth_dev_attach_secondary(name);
+				if (!eth_dev) {
+					DRV_LOG(ERR, "Can't attach to dev %s",
+						name);
+					ret = -ENOMEM;
+					goto failed;
+				}
+
+				eth_dev->device = &pci_dev->device;
+				eth_dev->dev_ops = &mana_dev_sec_ops;
+				ret = mana_proc_priv_init(eth_dev);
+				if (ret)
+					goto failed;
+				priv->process_priv = eth_dev->process_private;
+
+				/* Get the IB FD from the primary process */
+				fd = mana_mp_req_verbs_cmd_fd(eth_dev);
+				if (fd < 0) {
+					DRV_LOG(ERR, "Failed to get FD %d", fd);
+					ret = -ENODEV;
+					goto failed;
+				}
+
+				ret = mana_map_doorbell_secondary(eth_dev, fd);
+				if (ret) {
+					DRV_LOG(ERR, "Failed secondary map %d",
+						fd);
+					goto failed;
+				}
+
+				/* fd is no not used after mapping doorbell */
+				close(fd);
+
+				rte_spinlock_lock(&mana_shared_data->lock);
+				mana_shared_data->secondary_cnt++;
+				mana_local_data.secondary_cnt++;
+				rte_spinlock_unlock(&mana_shared_data->lock);
+
+				rte_eth_copy_pci_info(eth_dev, pci_dev);
+				rte_eth_dev_probing_finish(eth_dev);
+
+				/* Impossible to have more than one port
+				 * matching a MAC address
+				 */
+				continue;
+			}
+
+			eth_dev = rte_eth_dev_allocate(name);
+			if (!eth_dev) {
+				ret = -ENOMEM;
+				goto failed;
+			}
+
+			eth_dev->data->mac_addrs =
+				rte_calloc("mana_mac", 1,
+					   sizeof(struct rte_ether_addr), 0);
+			if (!eth_dev->data->mac_addrs) {
+				ret = -ENOMEM;
+				goto failed;
+			}
+
+			rte_ether_addr_copy(&addr, eth_dev->data->mac_addrs);
+
+			priv->ib_pd = ibv_alloc_pd(ctx);
+			if (!priv->ib_pd) {
+				DRV_LOG(ERR, "ibv_alloc_pd failed port %d", port);
+				ret = -ENOMEM;
+				goto failed;
+			}
+
+			/* Create a parent domain with the port number */
+			attr.pd = priv->ib_pd;
+			attr.comp_mask = IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT;
+			attr.pd_context = (void *)(uint64_t)port;
+			priv->ib_parent_pd = ibv_alloc_parent_domain(ctx, &attr);
+			if (!priv->ib_parent_pd) {
+				DRV_LOG(ERR,
+					"ibv_alloc_parent_domain failed port %d",
+					port);
+				ret = -ENOMEM;
+				goto failed;
+			}
+
+			priv->ib_ctx = ctx;
+			priv->port_id = eth_dev->data->port_id;
+			priv->dev_port = port;
+			eth_dev->data->dev_private = priv;
+			priv->dev_data = eth_dev->data;
+
+			priv->max_rx_queues = dev_attr.orig_attr.max_qp;
+			priv->max_tx_queues = dev_attr.orig_attr.max_qp;
+
+			priv->max_rx_desc =
+				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
+					dev_attr.orig_attr.max_cqe);
+			priv->max_tx_desc =
+				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
+					dev_attr.orig_attr.max_cqe);
+
+			priv->max_send_sge = dev_attr.orig_attr.max_sge;
+			priv->max_recv_sge = dev_attr.orig_attr.max_sge;
+
+			priv->max_mr = dev_attr.orig_attr.max_mr;
+			priv->max_mr_size = dev_attr.orig_attr.max_mr_size;
+
+			DRV_LOG(INFO, "dev %s max queues %d desc %d sge %d",
+				name, priv->max_rx_queues, priv->max_rx_desc,
+				priv->max_send_sge);
+
+			rte_spinlock_lock(&mana_shared_data->lock);
+			mana_shared_data->primary_cnt++;
+			rte_spinlock_unlock(&mana_shared_data->lock);
+
+			eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
+
+			eth_dev->device = &pci_dev->device;
+			eth_dev->data->dev_flags |=
+				RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
+
+			DRV_LOG(INFO, "device %s at port %u",
+				name, eth_dev->data->port_id);
+
+			eth_dev->rx_pkt_burst = mana_rx_burst_removed;
+			eth_dev->tx_pkt_burst = mana_tx_burst_removed;
+			eth_dev->dev_ops = &mana_dev_ops;
+
+			rte_eth_copy_pci_info(eth_dev, pci_dev);
+			rte_eth_dev_probing_finish(eth_dev);
+		}
+
+		/* Secondary process doesn't need an ibv_ctx. It maps the
+		 * doorbell pages using the IB cmd_fd passed from the primary
+		 * process and send messages to primary process for memory
+		 * registartions.
+		 */
+		if (!found_port || rte_eal_process_type() == RTE_PROC_SECONDARY)
+			ibv_close_device(ctx);
+	}
+
+	ibv_free_device_list(ibv_list);
+	return 0;
+
+failed:
+	/* Free the resource for the port failed */
+	if (priv) {
+		if (priv->ib_parent_pd)
+			ibv_dealloc_pd(priv->ib_parent_pd);
+
+		if (priv->ib_pd)
+			ibv_dealloc_pd(priv->ib_pd);
+	}
+
+	if (eth_dev)
+		rte_eth_dev_release_port(eth_dev);
+
+	rte_free(priv);
+
+	ibv_close_device(ctx);
+	ibv_free_device_list(ibv_list);
+
+	return ret;
+}
+
+static int mana_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
+			  struct rte_pci_device *pci_dev)
+{
+	struct rte_devargs *args = pci_dev->device.devargs;
+	struct mana_conf conf = {};
+	unsigned int i;
+	int ret;
+
+	if (args && args->args) {
+		ret = mana_parse_args(args, &conf);
+		if (ret) {
+			DRV_LOG(ERR, "failed to parse parameters args = %s",
+				args->args);
+			return ret;
+		}
+	}
+
+	ret = mana_init_once();
+	if (ret) {
+		DRV_LOG(ERR, "Failed to init PMD global data %d", ret);
+		return ret;
+	}
+
+	/* If there are no driver parameters, probe on all ports */
+	if (!conf.index)
+		return mana_pci_probe_mac(pci_drv, pci_dev, NULL);
+
+	for (i = 0; i < conf.index; i++) {
+		ret = mana_pci_probe_mac(pci_drv, pci_dev, &conf.mac_array[i]);
+		if (ret)
+			return ret;
+	}
+
+	return 0;
+}
+
+static int mana_dev_uninit(struct rte_eth_dev *dev)
+{
+	RTE_SET_USED(dev);
+	return 0;
+}
+
+static int mana_pci_remove(struct rte_pci_device *pci_dev)
+{
+	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+		rte_spinlock_lock(&mana_shared_data_lock);
+
+		rte_spinlock_lock(&mana_shared_data->lock);
+
+		RTE_VERIFY(mana_shared_data->primary_cnt > 0);
+		mana_shared_data->primary_cnt--;
+		if (!mana_shared_data->primary_cnt) {
+			DRV_LOG(DEBUG, "mp uninit primary");
+			mana_mp_uninit_primary();
+		}
+
+		rte_spinlock_unlock(&mana_shared_data->lock);
+
+		/* Also free the shared memory if this is the last */
+		if (!mana_shared_data->primary_cnt) {
+			DRV_LOG(DEBUG, "free shared memezone data");
+			rte_memzone_free(mana_shared_mz);
+		}
+
+		rte_spinlock_unlock(&mana_shared_data_lock);
+	} else {
+		rte_spinlock_lock(&mana_shared_data_lock);
+
+		rte_spinlock_lock(&mana_shared_data->lock);
+		RTE_VERIFY(mana_shared_data->secondary_cnt > 0);
+		mana_shared_data->secondary_cnt--;
+		rte_spinlock_unlock(&mana_shared_data->lock);
+
+		RTE_VERIFY(mana_local_data.secondary_cnt > 0);
+		mana_local_data.secondary_cnt--;
+		if (!mana_local_data.secondary_cnt) {
+			DRV_LOG(DEBUG, "mp uninit secondary");
+			mana_mp_uninit_secondary();
+		}
+
+		rte_spinlock_unlock(&mana_shared_data_lock);
+	}
+
+	return rte_eth_dev_pci_generic_remove(pci_dev, mana_dev_uninit);
+}
+
+static const struct rte_pci_id mana_pci_id_map[] = {
+	{
+		RTE_PCI_DEVICE(PCI_VENDOR_ID_MICROSOFT,
+			       PCI_DEVICE_ID_MICROSOFT_MANA)
+	},
+};
+
+static struct rte_pci_driver mana_pci_driver = {
+	.driver = {
+		.name = "mana_pci",
+	},
+	.id_table = mana_pci_id_map,
+	.probe = mana_pci_probe,
+	.remove = mana_pci_remove,
+	.drv_flags = RTE_PCI_DRV_INTR_RMV,
+};
+
+RTE_INIT(rte_mana_pmd_init)
+{
+	rte_pci_register(&mana_pci_driver);
+}
+
+RTE_PMD_EXPORT_NAME(net_mana, __COUNTER__);
+RTE_PMD_REGISTER_PCI_TABLE(net_mana, mana_pci_id_map);
+RTE_PMD_REGISTER_KMOD_DEP(net_mana, "* ib_uverbs & mana_ib");
+RTE_LOG_REGISTER_SUFFIX(mana_logtype_init, init, NOTICE);
+RTE_LOG_REGISTER_SUFFIX(mana_logtype_driver, driver, NOTICE);
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
new file mode 100644
index 0000000000..e30c030b4e
--- /dev/null
+++ b/drivers/net/mana/mana.h
@@ -0,0 +1,210 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2022 Microsoft Corporation
+ */
+
+#ifndef __MANA_H__
+#define __MANA_H__
+
+enum {
+	PCI_VENDOR_ID_MICROSOFT = 0x1414,
+};
+
+enum {
+	PCI_DEVICE_ID_MICROSOFT_MANA = 0x00ba,
+};
+
+/* Shared data between primary/secondary processes */
+struct mana_shared_data {
+	rte_spinlock_t lock;
+	int init_done;
+	unsigned int primary_cnt;
+	unsigned int secondary_cnt;
+};
+
+#define MIN_RX_BUF_SIZE	1024
+#define MAX_FRAME_SIZE	RTE_ETHER_MAX_LEN
+#define BNIC_MAX_MAC_ADDR 1
+
+#define BNIC_DEV_RX_OFFLOAD_SUPPORT ( \
+		DEV_RX_OFFLOAD_CHECKSUM | \
+		DEV_RX_OFFLOAD_RSS_HASH)
+
+#define BNIC_DEV_TX_OFFLOAD_SUPPORT ( \
+		RTE_ETH_TX_OFFLOAD_MULTI_SEGS | \
+		RTE_ETH_TX_OFFLOAD_IPV4_CKSUM | \
+		RTE_ETH_TX_OFFLOAD_TCP_CKSUM | \
+		RTE_ETH_TX_OFFLOAD_UDP_CKSUM | \
+		RTE_ETH_TX_OFFLOAD_TCP_TSO)
+
+#define INDIRECTION_TABLE_NUM_ELEMENTS 64
+#define TOEPLITZ_HASH_KEY_SIZE_IN_BYTES 40
+#define BNIC_ETH_RSS_SUPPORT ( \
+	ETH_RSS_IPV4 |	     \
+	ETH_RSS_NONFRAG_IPV4_TCP | \
+	ETH_RSS_NONFRAG_IPV4_UDP | \
+	ETH_RSS_IPV6 |	     \
+	ETH_RSS_NONFRAG_IPV6_TCP | \
+	ETH_RSS_NONFRAG_IPV6_UDP)
+
+#define MIN_BUFFERS_PER_QUEUE		64
+#define MAX_RECEIVE_BUFFERS_PER_QUEUE	256
+#define MAX_SEND_BUFFERS_PER_QUEUE	256
+
+struct mana_process_priv {
+	void *db_page;
+};
+
+struct mana_priv {
+	struct rte_eth_dev_data *dev_data;
+	struct mana_process_priv *process_priv;
+	int num_queues;
+
+	/* DPDK port */
+	uint16_t port_id;
+
+	/* IB device port */
+	uint8_t dev_port;
+
+	struct ibv_context *ib_ctx;
+	struct ibv_pd *ib_pd;
+	struct ibv_pd *ib_parent_pd;
+	struct ibv_rwq_ind_table *ind_table;
+	uint8_t ind_table_key[40];
+	struct ibv_qp *rwq_qp;
+	void *db_page;
+	int max_rx_queues;
+	int max_tx_queues;
+	int max_rx_desc;
+	int max_tx_desc;
+	int max_send_sge;
+	int max_recv_sge;
+	int max_mr;
+	uint64_t max_mr_size;
+};
+
+struct mana_txq_desc {
+	struct rte_mbuf *pkt;
+	uint32_t wqe_size_in_bu;
+};
+
+struct mana_rxq_desc {
+	struct rte_mbuf *pkt;
+	uint32_t wqe_size_in_bu;
+};
+
+struct mana_gdma_queue {
+	void *buffer;
+	uint32_t count;	/* in entries */
+	uint32_t size;	/* in bytes */
+	uint32_t id;
+	uint32_t head;
+	uint32_t tail;
+};
+
+struct mana_stats {
+	uint64_t packets;
+	uint64_t bytes;
+	uint64_t errors;
+	uint64_t nombuf;
+};
+
+#define MANA_MR_BTREE_PER_QUEUE_N	64
+struct mana_txq {
+	struct mana_priv *priv;
+	uint32_t num_desc;
+	struct ibv_cq *cq;
+	struct ibv_qp *qp;
+
+	struct mana_gdma_queue gdma_sq;
+	struct mana_gdma_queue gdma_cq;
+
+	uint32_t tx_vp_offset;
+
+	/* For storing pending requests */
+	struct mana_txq_desc *desc_ring;
+
+	/* desc_ring_head is where we put pending requests to ring,
+	 * completion pull off desc_ring_tail
+	 */
+	uint32_t desc_ring_head, desc_ring_tail;
+
+	struct mana_stats stats;
+	unsigned int socket;
+};
+
+struct mana_rxq {
+	struct mana_priv *priv;
+	uint32_t num_desc;
+	struct rte_mempool *mp;
+	struct ibv_cq *cq;
+	struct ibv_wq *wq;
+
+	/* For storing pending requests */
+	struct mana_rxq_desc *desc_ring;
+
+	/* desc_ring_head is where we put pending requests to ring,
+	 * completion pull off desc_ring_tail
+	 */
+	uint32_t desc_ring_head, desc_ring_tail;
+
+	struct mana_gdma_queue gdma_rq;
+	struct mana_gdma_queue gdma_cq;
+
+	struct mana_stats stats;
+
+	unsigned int socket;
+};
+
+extern int mana_logtype_driver;
+extern int mana_logtype_init;
+
+#define DRV_LOG(level, fmt, args...) \
+	rte_log(RTE_LOG_ ## level, mana_logtype_driver, "%s(): " fmt "\n", \
+		__func__, ## args)
+
+#define PMD_INIT_LOG(level, fmt, args...) \
+	rte_log(RTE_LOG_ ## level, mana_logtype_init, "%s(): " fmt "\n",\
+		__func__, ## args)
+
+#define PMD_INIT_FUNC_TRACE() PMD_INIT_LOG(DEBUG, " >>")
+
+const uint32_t *mana_supported_ptypes(struct rte_eth_dev *dev);
+
+uint16_t mana_rx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
+			       uint16_t pkts_n);
+
+uint16_t mana_tx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
+			       uint16_t pkts_n);
+
+/** Request timeout for IPC. */
+#define MANA_MP_REQ_TIMEOUT_SEC 5
+
+/* Request types for IPC. */
+enum mana_mp_req_type {
+	MANA_MP_REQ_VERBS_CMD_FD = 1,
+	MANA_MP_REQ_CREATE_MR,
+	MANA_MP_REQ_START_RXTX,
+	MANA_MP_REQ_STOP_RXTX,
+};
+
+/* Pameters for IPC. */
+struct mana_mp_param {
+	enum mana_mp_req_type type;
+	int port_id;
+	int result;
+
+	/* MANA_MP_REQ_CREATE_MR */
+	uintptr_t addr;
+	uint32_t len;
+};
+
+#define MANA_MP_NAME	"net_mana_mp"
+int mana_mp_init_primary(void);
+int mana_mp_init_secondary(void);
+void mana_mp_uninit_primary(void);
+void mana_mp_uninit_secondary(void);
+int mana_mp_req_verbs_cmd_fd(struct rte_eth_dev *dev);
+
+void mana_mp_req_on_rxtx(struct rte_eth_dev *dev, enum mana_mp_req_type type);
+
+#endif
diff --git a/drivers/net/mana/meson.build b/drivers/net/mana/meson.build
new file mode 100644
index 0000000000..81c4118f53
--- /dev/null
+++ b/drivers/net/mana/meson.build
@@ -0,0 +1,44 @@
+# SPDX-License-Identifier: BSD-3-Clause
+# Copyright(c) 2022 Microsoft Corporation
+
+if not is_linux or not dpdk_conf.has('RTE_ARCH_X86_64')
+    build = false
+    reason = 'mana is supported on Linux X86_64'
+    subdir_done()
+endif
+
+deps += ['pci', 'bus_pci', 'net', 'eal', 'kvargs']
+
+sources += files(
+	'mana.c',
+	'mp.c',
+)
+
+libnames = ['ibverbs', 'mana' ]
+foreach libname:libnames
+    lib = cc.find_library(libname, required:false)
+    if lib.found()
+        ext_deps += lib
+    else
+        build = false
+        reason = 'missing dependency, "' + libname + '"'
+        subdir_done()
+    endif
+endforeach
+
+required_symbols = [
+    ['infiniband/manadv.h', 'manadv_set_context_attr'],
+    ['infiniband/manadv.h', 'manadv_init_obj'],
+    ['infiniband/manadv.h', 'MANADV_CTX_ATTR_BUF_ALLOCATORS'],
+    ['infiniband/manadv.h', 'MANADV_OBJ_QP'],
+    ['infiniband/manadv.h', 'MANADV_OBJ_CQ'],
+    ['infiniband/manadv.h', 'MANADV_OBJ_RWQ'],
+]
+
+foreach arg:required_symbols
+    if not cc.has_header_symbol(arg[0], arg[1])
+        build = false
+        reason = 'missing symbol "' + arg[1] + '" in "' + arg[0] + '"'
+        subdir_done()
+    endif
+endforeach
diff --git a/drivers/net/mana/mp.c b/drivers/net/mana/mp.c
new file mode 100644
index 0000000000..d7580e8a28
--- /dev/null
+++ b/drivers/net/mana/mp.c
@@ -0,0 +1,235 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2022 Microsoft Corporation
+ */
+
+#include <rte_malloc.h>
+#include <ethdev_driver.h>
+#include <rte_log.h>
+
+#include <infiniband/verbs.h>
+
+#include "mana.h"
+
+extern struct mana_shared_data *mana_shared_data;
+
+static void mp_init_msg(struct rte_mp_msg *msg, enum mana_mp_req_type type,
+			int port_id)
+{
+	struct mana_mp_param *param;
+
+	strlcpy(msg->name, MANA_MP_NAME, sizeof(msg->name));
+	msg->len_param = sizeof(*param);
+
+	param = (struct mana_mp_param *)msg->param;
+	param->type = type;
+	param->port_id = port_id;
+}
+
+static int mana_mp_primary_handle(const struct rte_mp_msg *mp_msg,
+				  const void *peer)
+{
+	struct rte_eth_dev *dev;
+	const struct mana_mp_param *param =
+		(const struct mana_mp_param *)mp_msg->param;
+	struct rte_mp_msg mp_res = { 0 };
+	struct mana_mp_param *res = (struct mana_mp_param *)mp_res.param;
+	int ret;
+	struct mana_priv *priv;
+
+	if (!rte_eth_dev_is_valid_port(param->port_id)) {
+		DRV_LOG(ERR, "MP handle port ID %u invalid", param->port_id);
+		return -ENODEV;
+	}
+
+	dev = &rte_eth_devices[param->port_id];
+	priv = dev->data->dev_private;
+
+	mp_init_msg(&mp_res, param->type, param->port_id);
+
+	switch (param->type) {
+	case MANA_MP_REQ_VERBS_CMD_FD:
+		mp_res.num_fds = 1;
+		mp_res.fds[0] = priv->ib_ctx->cmd_fd;
+		res->result = 0;
+		ret = rte_mp_reply(&mp_res, peer);
+		break;
+
+	default:
+		DRV_LOG(ERR, "Port %u unknown primary MP type %u",
+			param->port_id, param->type);
+		ret = -EINVAL;
+	}
+
+	return ret;
+}
+
+static int mana_mp_secondary_handle(const struct rte_mp_msg *mp_msg,
+				    const void *peer)
+{
+	struct rte_mp_msg mp_res = { 0 };
+	struct mana_mp_param *res = (struct mana_mp_param *)mp_res.param;
+	const struct mana_mp_param *param =
+		(const struct mana_mp_param *)mp_msg->param;
+	struct rte_eth_dev *dev;
+	int ret;
+
+	if (!rte_eth_dev_is_valid_port(param->port_id)) {
+		DRV_LOG(ERR, "MP handle port ID %u invalid", param->port_id);
+		return -ENODEV;
+	}
+
+	dev = &rte_eth_devices[param->port_id];
+
+	mp_init_msg(&mp_res, param->type, param->port_id);
+
+	switch (param->type) {
+	case MANA_MP_REQ_START_RXTX:
+		DRV_LOG(INFO, "Port %u starting datapath", dev->data->port_id);
+
+		rte_mb();
+
+		res->result = 0;
+		ret = rte_mp_reply(&mp_res, peer);
+		break;
+
+	case MANA_MP_REQ_STOP_RXTX:
+		DRV_LOG(INFO, "Port %u stopping datapath", dev->data->port_id);
+
+		dev->tx_pkt_burst = mana_tx_burst_removed;
+		dev->rx_pkt_burst = mana_rx_burst_removed;
+
+		rte_mb();
+
+		res->result = 0;
+		ret = rte_mp_reply(&mp_res, peer);
+		break;
+
+	default:
+		DRV_LOG(ERR, "Port %u unknown secondary MP type %u",
+			param->port_id, param->type);
+		ret = -EINVAL;
+	}
+
+	return ret;
+}
+
+int mana_mp_init_primary(void)
+{
+	int ret;
+
+	ret = rte_mp_action_register(MANA_MP_NAME, mana_mp_primary_handle);
+	if (ret && rte_errno != ENOTSUP) {
+		DRV_LOG(ERR, "Failed to register primary handler %d %d",
+			ret, rte_errno);
+		return -1;
+	}
+
+	return 0;
+}
+
+void mana_mp_uninit_primary(void)
+{
+	rte_mp_action_unregister(MANA_MP_NAME);
+}
+
+int mana_mp_init_secondary(void)
+{
+	return rte_mp_action_register(MANA_MP_NAME, mana_mp_secondary_handle);
+}
+
+void mana_mp_uninit_secondary(void)
+{
+	rte_mp_action_unregister(MANA_MP_NAME);
+}
+
+int mana_mp_req_verbs_cmd_fd(struct rte_eth_dev *dev)
+{
+	struct rte_mp_msg mp_req = { 0 };
+	struct rte_mp_msg *mp_res;
+	struct rte_mp_reply mp_rep;
+	struct mana_mp_param *res;
+	struct timespec ts = {.tv_sec = MANA_MP_REQ_TIMEOUT_SEC, .tv_nsec = 0};
+	int ret;
+
+	mp_init_msg(&mp_req, MANA_MP_REQ_VERBS_CMD_FD, dev->data->port_id);
+
+	ret = rte_mp_request_sync(&mp_req, &mp_rep, &ts);
+	if (ret) {
+		DRV_LOG(ERR, "port %u request to primary process failed",
+			dev->data->port_id);
+		return ret;
+	}
+
+	if (mp_rep.nb_received != 1) {
+		DRV_LOG(ERR, "primary replied %u messages", mp_rep.nb_received);
+		ret = -EPROTO;
+		goto exit;
+	}
+
+	mp_res = &mp_rep.msgs[0];
+	res = (struct mana_mp_param *)mp_res->param;
+	if (res->result) {
+		DRV_LOG(ERR, "failed to get CMD FD, port %u",
+			dev->data->port_id);
+		ret = res->result;
+		goto exit;
+	}
+
+	if (mp_res->num_fds != 1) {
+		DRV_LOG(ERR, "got FDs %d unexpected", mp_res->num_fds);
+		ret = -EPROTO;
+		goto exit;
+	}
+
+	ret = mp_res->fds[0];
+	DRV_LOG(ERR, "port %u command FD from primary is %d",
+		dev->data->port_id, ret);
+exit:
+	free(mp_rep.msgs);
+	return ret;
+}
+
+void mana_mp_req_on_rxtx(struct rte_eth_dev *dev, enum mana_mp_req_type type)
+{
+	struct rte_mp_msg mp_req = { 0 };
+	struct rte_mp_msg *mp_res;
+	struct rte_mp_reply mp_rep;
+	struct mana_mp_param *res;
+	struct timespec ts = {.tv_sec = MANA_MP_REQ_TIMEOUT_SEC, .tv_nsec = 0};
+	int i, ret;
+
+	if (type != MANA_MP_REQ_START_RXTX && type != MANA_MP_REQ_STOP_RXTX) {
+		DRV_LOG(ERR, "port %u unknown request (req_type %d)",
+			dev->data->port_id, type);
+		return;
+	}
+
+	if (!mana_shared_data->secondary_cnt)
+		return;
+
+	mp_init_msg(&mp_req, type, dev->data->port_id);
+
+	ret = rte_mp_request_sync(&mp_req, &mp_rep, &ts);
+	if (ret) {
+		if (rte_errno != ENOTSUP)
+			DRV_LOG(ERR, "port %u failed to request Rx/Tx (%d)",
+				dev->data->port_id, type);
+		goto exit;
+	}
+	if (mp_rep.nb_sent != mp_rep.nb_received) {
+		DRV_LOG(ERR, "port %u not all secondaries responded (%d)",
+			dev->data->port_id, type);
+		goto exit;
+	}
+	for (i = 0; i < mp_rep.nb_received; i++) {
+		mp_res = &mp_rep.msgs[i];
+		res = (struct mana_mp_param *)mp_res->param;
+		if (res->result) {
+			DRV_LOG(ERR, "port %u request failed on secondary %d",
+				dev->data->port_id, i);
+			goto exit;
+		}
+	}
+exit:
+	free(mp_rep.msgs);
+}
diff --git a/drivers/net/mana/version.map b/drivers/net/mana/version.map
new file mode 100644
index 0000000000..c2e0723b4c
--- /dev/null
+++ b/drivers/net/mana/version.map
@@ -0,0 +1,3 @@
+DPDK_22 {
+	local: *;
+};
diff --git a/drivers/net/meson.build b/drivers/net/meson.build
index 2355d1cde8..0b111a6ebb 100644
--- a/drivers/net/meson.build
+++ b/drivers/net/meson.build
@@ -34,6 +34,7 @@ drivers = [
         'ixgbe',
         'kni',
         'liquidio',
+        'mana',
         'memif',
         'mlx4',
         'mlx5',
-- 
2.17.1

[Patch v6 02/18] net/mana: add device configuration and stop

[longli]

From: Long Li <longli@microsoft.com>

MANA defines its memory allocation functions to override IB layer default
functions to allocate device queues. This patch adds the code for device
configuration and stop.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v2:
Removed validation for offload settings in mana_dev_configure().

 drivers/net/mana/mana.c | 75 +++++++++++++++++++++++++++++++++++++++--
 drivers/net/mana/mana.h |  3 ++
 2 files changed, 76 insertions(+), 2 deletions(-)

diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index cb59eb6882..147ab144d5 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -40,7 +40,79 @@ static rte_spinlock_t mana_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
 int mana_logtype_driver;
 int mana_logtype_init;
 
+void *mana_alloc_verbs_buf(size_t size, void *data)
+{
+	void *ret;
+	size_t alignment = rte_mem_page_size();
+	int socket = (int)(uintptr_t)data;
+
+	DRV_LOG(DEBUG, "size=%zu socket=%d", size, socket);
+
+	if (alignment == (size_t)-1) {
+		DRV_LOG(ERR, "Failed to get mem page size");
+		rte_errno = ENOMEM;
+		return NULL;
+	}
+
+	ret = rte_zmalloc_socket("mana_verb_buf", size, alignment, socket);
+	if (!ret && size)
+		rte_errno = ENOMEM;
+	return ret;
+}
+
+void mana_free_verbs_buf(void *ptr, void *data __rte_unused)
+{
+	rte_free(ptr);
+}
+
+static int mana_dev_configure(struct rte_eth_dev *dev)
+{
+	struct mana_priv *priv = dev->data->dev_private;
+	struct rte_eth_conf *dev_conf = &dev->data->dev_conf;
+
+	if (dev_conf->rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
+		dev_conf->rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH;
+
+	if (dev->data->nb_rx_queues != dev->data->nb_tx_queues) {
+		DRV_LOG(ERR, "Only support equal number of rx/tx queues");
+		return -EINVAL;
+	}
+
+	if (!rte_is_power_of_2(dev->data->nb_rx_queues)) {
+		DRV_LOG(ERR, "number of TX/RX queues must be power of 2");
+		return -EINVAL;
+	}
+
+	priv->num_queues = dev->data->nb_rx_queues;
+
+	manadv_set_context_attr(priv->ib_ctx, MANADV_CTX_ATTR_BUF_ALLOCATORS,
+				(void *)((uintptr_t)&(struct manadv_ctx_allocators){
+					.alloc = &mana_alloc_verbs_buf,
+					.free = &mana_free_verbs_buf,
+					.data = 0,
+				}));
+
+	return 0;
+}
+
+static int
+mana_dev_close(struct rte_eth_dev *dev)
+{
+	struct mana_priv *priv = dev->data->dev_private;
+	int ret;
+
+	ret = ibv_close_device(priv->ib_ctx);
+	if (ret) {
+		ret = errno;
+		return ret;
+	}
+
+	return 0;
+}
+
 const struct eth_dev_ops mana_dev_ops = {
+	.dev_configure		= mana_dev_configure,
+	.dev_close		= mana_dev_close,
 };
 
 const struct eth_dev_ops mana_dev_sec_ops = {
@@ -627,8 +699,7 @@ static int mana_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 
 static int mana_dev_uninit(struct rte_eth_dev *dev)
 {
-	RTE_SET_USED(dev);
-	return 0;
+	return mana_dev_close(dev);
 }
 
 static int mana_pci_remove(struct rte_pci_device *pci_dev)
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index e30c030b4e..66873394b9 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -207,4 +207,7 @@ int mana_mp_req_verbs_cmd_fd(struct rte_eth_dev *dev);
 
 void mana_mp_req_on_rxtx(struct rte_eth_dev *dev, enum mana_mp_req_type type);
 
+void *mana_alloc_verbs_buf(size_t size, void *data);
+void mana_free_verbs_buf(void *ptr, void *data __rte_unused);
+
 #endif
-- 
2.17.1

[Patch v6 03/18] net/mana: add function to report support ptypes

[longli]

From: Long Li <longli@microsoft.com>

Report supported protocol types.

Signed-off-by: Long Li <longli@microsoft.com>
---
 drivers/net/mana/mana.c | 16 ++++++++++++++++
 drivers/net/mana/mana.h |  2 --
 2 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index 147ab144d5..4559632056 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -110,9 +110,25 @@ mana_dev_close(struct rte_eth_dev *dev)
 	return 0;
 }
 
+static const uint32_t *mana_supported_ptypes(struct rte_eth_dev *dev __rte_unused)
+{
+	static const uint32_t ptypes[] = {
+		RTE_PTYPE_L2_ETHER,
+		RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
+		RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
+		RTE_PTYPE_L4_FRAG,
+		RTE_PTYPE_L4_TCP,
+		RTE_PTYPE_L4_UDP,
+		RTE_PTYPE_UNKNOWN
+	};
+
+	return ptypes;
+}
+
 const struct eth_dev_ops mana_dev_ops = {
 	.dev_configure		= mana_dev_configure,
 	.dev_close		= mana_dev_close,
+	.dev_supported_ptypes_get = mana_supported_ptypes,
 };
 
 const struct eth_dev_ops mana_dev_sec_ops = {
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index 66873394b9..c433940022 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -168,8 +168,6 @@ extern int mana_logtype_init;
 
 #define PMD_INIT_FUNC_TRACE() PMD_INIT_LOG(DEBUG, " >>")
 
-const uint32_t *mana_supported_ptypes(struct rte_eth_dev *dev);
-
 uint16_t mana_rx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
 			       uint16_t pkts_n);
 
-- 
2.17.1

[Patch v6 04/18] net/mana: add link update

[longli]

From: Long Li <longli@microsoft.com>

The carrier state is managed by the Azure host. MANA runs as a VF and
always reports "up".

Signed-off-by: Long Li <longli@microsoft.com>
---
 doc/guides/nics/features/mana.ini |  1 +
 drivers/net/mana/mana.c           | 17 +++++++++++++++++
 2 files changed, 18 insertions(+)

diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
index b92a27374c..62554b0a0a 100644
--- a/doc/guides/nics/features/mana.ini
+++ b/doc/guides/nics/features/mana.ini
@@ -4,6 +4,7 @@
 ; Refer to default.ini for the full list of available PMD features.
 ;
 [Features]
+Link status          = P
 Linux                = Y
 Multiprocess aware   = Y
 Usage doc            = Y
diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index 4559632056..b77d0c29b0 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -125,10 +125,27 @@ static const uint32_t *mana_supported_ptypes(struct rte_eth_dev *dev __rte_unuse
 	return ptypes;
 }
 
+static int mana_dev_link_update(struct rte_eth_dev *dev,
+				int wait_to_complete __rte_unused)
+{
+	struct rte_eth_link link;
+
+	/* MANA has no concept of carrier state, always reporting UP */
+	link = (struct rte_eth_link) {
+		.link_duplex = RTE_ETH_LINK_FULL_DUPLEX,
+		.link_autoneg = RTE_ETH_LINK_SPEED_FIXED,
+		.link_speed = RTE_ETH_SPEED_NUM_200G,
+		.link_status = RTE_ETH_LINK_UP,
+	};
+
+	return rte_eth_linkstatus_set(dev, &link);
+}
+
 const struct eth_dev_ops mana_dev_ops = {
 	.dev_configure		= mana_dev_configure,
 	.dev_close		= mana_dev_close,
 	.dev_supported_ptypes_get = mana_supported_ptypes,
+	.link_update		= mana_dev_link_update,
 };
 
 const struct eth_dev_ops mana_dev_sec_ops = {
-- 
2.17.1

[Patch v6 05/18] net/mana: add function for device removal interrupts

[longli]

From: Long Li <longli@microsoft.com>

MANA supports PCI hot plug events. Add this interrupt to DPDK core so its
parent PMD can detect device removal during Azure servicing or live
migration.

Signed-off-by: Long Li <longli@microsoft.com>
---
 doc/guides/nics/features/mana.ini |  1 +
 drivers/net/mana/mana.c           | 97 +++++++++++++++++++++++++++++++
 drivers/net/mana/mana.h           |  1 +
 3 files changed, 99 insertions(+)

diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
index 62554b0a0a..8043e11f99 100644
--- a/doc/guides/nics/features/mana.ini
+++ b/doc/guides/nics/features/mana.ini
@@ -7,5 +7,6 @@
 Link status          = P
 Linux                = Y
 Multiprocess aware   = Y
+Removal event        = Y
 Usage doc            = Y
 x86-64               = Y
diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index b77d0c29b0..c9591035ac 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -95,12 +95,18 @@ static int mana_dev_configure(struct rte_eth_dev *dev)
 	return 0;
 }
 
+static int mana_intr_uninstall(struct mana_priv *priv);
+
 static int
 mana_dev_close(struct rte_eth_dev *dev)
 {
 	struct mana_priv *priv = dev->data->dev_private;
 	int ret;
 
+	ret = mana_intr_uninstall(priv);
+	if (ret)
+		return ret;
+
 	ret = ibv_close_device(priv->ib_ctx);
 	if (ret) {
 		ret = errno;
@@ -327,6 +333,90 @@ static int mana_ibv_device_to_pci_addr(const struct ibv_device *device,
 	return 0;
 }
 
+static void mana_intr_handler(void *arg)
+{
+	struct mana_priv *priv = arg;
+	struct ibv_context *ctx = priv->ib_ctx;
+	struct ibv_async_event event;
+
+	/* Read and ack all messages from IB device */
+	while (true) {
+		if (ibv_get_async_event(ctx, &event))
+			break;
+
+		if (event.event_type == IBV_EVENT_DEVICE_FATAL) {
+			struct rte_eth_dev *dev;
+
+			dev = &rte_eth_devices[priv->port_id];
+			if (dev->data->dev_conf.intr_conf.rmv)
+				rte_eth_dev_callback_process(dev,
+					RTE_ETH_EVENT_INTR_RMV, NULL);
+		}
+
+		ibv_ack_async_event(&event);
+	}
+}
+
+static int mana_intr_uninstall(struct mana_priv *priv)
+{
+	int ret;
+
+	ret = rte_intr_callback_unregister(priv->intr_handle,
+					   mana_intr_handler, priv);
+	if (ret <= 0) {
+		DRV_LOG(ERR, "Failed to unregister intr callback ret %d", ret);
+		return ret;
+	}
+
+	rte_intr_instance_free(priv->intr_handle);
+
+	return 0;
+}
+
+static int mana_intr_install(struct mana_priv *priv)
+{
+	int ret, flags;
+	struct ibv_context *ctx = priv->ib_ctx;
+
+	priv->intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
+	if (!priv->intr_handle) {
+		DRV_LOG(ERR, "Failed to allocate intr_handle");
+		rte_errno = ENOMEM;
+		return -ENOMEM;
+	}
+
+	rte_intr_fd_set(priv->intr_handle, -1);
+
+	flags = fcntl(ctx->async_fd, F_GETFL);
+	ret = fcntl(ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
+	if (ret) {
+		DRV_LOG(ERR, "Failed to change async_fd to NONBLOCK");
+		goto free_intr;
+	}
+
+	rte_intr_fd_set(priv->intr_handle, ctx->async_fd);
+	rte_intr_type_set(priv->intr_handle, RTE_INTR_HANDLE_EXT);
+
+	ret = rte_intr_callback_register(priv->intr_handle,
+					 mana_intr_handler, priv);
+	if (ret) {
+		DRV_LOG(ERR, "Failed to register intr callback");
+		rte_intr_fd_set(priv->intr_handle, -1);
+		goto restore_fd;
+	}
+
+	return 0;
+
+restore_fd:
+	fcntl(ctx->async_fd, F_SETFL, flags);
+
+free_intr:
+	rte_intr_instance_free(priv->intr_handle);
+	priv->intr_handle = NULL;
+
+	return ret;
+}
+
 static int mana_proc_priv_init(struct rte_eth_dev *dev)
 {
 	struct mana_process_priv *priv;
@@ -640,6 +730,13 @@ static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv __rte_unused,
 				name, priv->max_rx_queues, priv->max_rx_desc,
 				priv->max_send_sge);
 
+			/* Create async interrupt handler */
+			ret = mana_intr_install(priv);
+			if (ret) {
+				DRV_LOG(ERR, "Failed to install intr handler");
+				goto failed;
+			}
+
 			rte_spinlock_lock(&mana_shared_data->lock);
 			mana_shared_data->primary_cnt++;
 			rte_spinlock_unlock(&mana_shared_data->lock);
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index c433940022..f97eed2e81 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -72,6 +72,7 @@ struct mana_priv {
 	uint8_t ind_table_key[40];
 	struct ibv_qp *rwq_qp;
 	void *db_page;
+	struct rte_intr_handle *intr_handle;
 	int max_rx_queues;
 	int max_tx_queues;
 	int max_rx_desc;
-- 
2.17.1

[Patch v6 06/18] net/mana: add device info

[longli]

From: Long Li <longli@microsoft.com>

Add the function to get device info.

Signed-off-by: Long Li <longli@microsoft.com>
---
 doc/guides/nics/features/mana.ini |  1 +
 drivers/net/mana/mana.c           | 82 +++++++++++++++++++++++++++++++
 2 files changed, 83 insertions(+)

diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
index 8043e11f99..566b3e8770 100644
--- a/doc/guides/nics/features/mana.ini
+++ b/doc/guides/nics/features/mana.ini
@@ -8,5 +8,6 @@ Link status          = P
 Linux                = Y
 Multiprocess aware   = Y
 Removal event        = Y
+Speed capabilities   = P
 Usage doc            = Y
 x86-64               = Y
diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index c9591035ac..e1550b3c08 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -116,6 +116,86 @@ mana_dev_close(struct rte_eth_dev *dev)
 	return 0;
 }
 
+static int mana_dev_info_get(struct rte_eth_dev *dev,
+			     struct rte_eth_dev_info *dev_info)
+{
+	struct mana_priv *priv = dev->data->dev_private;
+
+	dev_info->max_mtu = RTE_ETHER_MTU;
+
+	/* RX params */
+	dev_info->min_rx_bufsize = MIN_RX_BUF_SIZE;
+	dev_info->max_rx_pktlen = MAX_FRAME_SIZE;
+
+	dev_info->max_rx_queues = priv->max_rx_queues;
+	dev_info->max_tx_queues = priv->max_tx_queues;
+
+	dev_info->max_mac_addrs = BNIC_MAX_MAC_ADDR;
+	dev_info->max_hash_mac_addrs = 0;
+
+	dev_info->max_vfs = 1;
+
+	/* Offload params */
+	dev_info->rx_offload_capa = BNIC_DEV_RX_OFFLOAD_SUPPORT;
+
+	dev_info->tx_offload_capa = BNIC_DEV_TX_OFFLOAD_SUPPORT;
+
+	/* RSS */
+	dev_info->reta_size = INDIRECTION_TABLE_NUM_ELEMENTS;
+	dev_info->hash_key_size = TOEPLITZ_HASH_KEY_SIZE_IN_BYTES;
+	dev_info->flow_type_rss_offloads = BNIC_ETH_RSS_SUPPORT;
+
+	/* Thresholds */
+	dev_info->default_rxconf = (struct rte_eth_rxconf){
+		.rx_thresh = {
+			.pthresh = 8,
+			.hthresh = 8,
+			.wthresh = 0,
+		},
+		.rx_free_thresh = 32,
+		/* If no descriptors available, pkts are dropped by default */
+		.rx_drop_en = 1,
+	};
+
+	dev_info->default_txconf = (struct rte_eth_txconf){
+		.tx_thresh = {
+			.pthresh = 32,
+			.hthresh = 0,
+			.wthresh = 0,
+		},
+		.tx_rs_thresh = 32,
+		.tx_free_thresh = 32,
+	};
+
+	/* Buffer limits */
+	dev_info->rx_desc_lim.nb_min = MIN_BUFFERS_PER_QUEUE;
+	dev_info->rx_desc_lim.nb_max = priv->max_rx_desc;
+	dev_info->rx_desc_lim.nb_align = MIN_BUFFERS_PER_QUEUE;
+	dev_info->rx_desc_lim.nb_seg_max = priv->max_recv_sge;
+	dev_info->rx_desc_lim.nb_mtu_seg_max = priv->max_recv_sge;
+
+	dev_info->tx_desc_lim.nb_min = MIN_BUFFERS_PER_QUEUE;
+	dev_info->tx_desc_lim.nb_max = priv->max_tx_desc;
+	dev_info->tx_desc_lim.nb_align = MIN_BUFFERS_PER_QUEUE;
+	dev_info->tx_desc_lim.nb_seg_max = priv->max_send_sge;
+	dev_info->rx_desc_lim.nb_mtu_seg_max = priv->max_recv_sge;
+
+	/* Speed */
+	dev_info->speed_capa = ETH_LINK_SPEED_100G;
+
+	/* RX params */
+	dev_info->default_rxportconf.burst_size = 1;
+	dev_info->default_rxportconf.ring_size = MAX_RECEIVE_BUFFERS_PER_QUEUE;
+	dev_info->default_rxportconf.nb_queues = 1;
+
+	/* TX params */
+	dev_info->default_txportconf.burst_size = 1;
+	dev_info->default_txportconf.ring_size = MAX_SEND_BUFFERS_PER_QUEUE;
+	dev_info->default_txportconf.nb_queues = 1;
+
+	return 0;
+}
+
 static const uint32_t *mana_supported_ptypes(struct rte_eth_dev *dev __rte_unused)
 {
 	static const uint32_t ptypes[] = {
@@ -150,11 +230,13 @@ static int mana_dev_link_update(struct rte_eth_dev *dev,
 const struct eth_dev_ops mana_dev_ops = {
 	.dev_configure		= mana_dev_configure,
 	.dev_close		= mana_dev_close,
+	.dev_infos_get		= mana_dev_info_get,
 	.dev_supported_ptypes_get = mana_supported_ptypes,
 	.link_update		= mana_dev_link_update,
 };
 
 const struct eth_dev_ops mana_dev_sec_ops = {
+	.dev_infos_get = mana_dev_info_get,
 };
 
 uint16_t
-- 
2.17.1

[Patch v6 07/18] net/mana: add function to configure RSS

[longli]

From: Long Li <longli@microsoft.com>

Currently this PMD supports RSS configuration when the device is stopped.
Configuring RSS in running state will be supported in the future.

Signed-off-by: Long Li <longli@microsoft.com>
---
 doc/guides/nics/features/mana.ini |  1 +
 drivers/net/mana/mana.c           | 61 +++++++++++++++++++++++++++++++
 drivers/net/mana/mana.h           |  1 +
 3 files changed, 63 insertions(+)

diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
index 566b3e8770..a59c21cc10 100644
--- a/doc/guides/nics/features/mana.ini
+++ b/doc/guides/nics/features/mana.ini
@@ -8,6 +8,7 @@ Link status          = P
 Linux                = Y
 Multiprocess aware   = Y
 Removal event        = Y
+RSS hash             = Y
 Speed capabilities   = P
 Usage doc            = Y
 x86-64               = Y
diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index e1550b3c08..cb136f24c1 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -211,6 +211,65 @@ static const uint32_t *mana_supported_ptypes(struct rte_eth_dev *dev __rte_unuse
 	return ptypes;
 }
 
+static int mana_rss_hash_update(struct rte_eth_dev *dev,
+				struct rte_eth_rss_conf *rss_conf)
+{
+	struct mana_priv *priv = dev->data->dev_private;
+
+	/* Currently can only update RSS hash when device is stopped */
+	if (dev->data->dev_started) {
+		DRV_LOG(ERR, "Can't update RSS after device has started");
+		return -ENODEV;
+	}
+
+	if (rss_conf->rss_hf & ~BNIC_ETH_RSS_SUPPORT) {
+		DRV_LOG(ERR, "Port %u invalid RSS HF 0x%" PRIx64,
+			dev->data->port_id, rss_conf->rss_hf);
+		return -EINVAL;
+	}
+
+	if (rss_conf->rss_key && rss_conf->rss_key_len) {
+		if (rss_conf->rss_key_len != TOEPLITZ_HASH_KEY_SIZE_IN_BYTES) {
+			DRV_LOG(ERR, "Port %u key len must be %u long",
+				dev->data->port_id,
+				TOEPLITZ_HASH_KEY_SIZE_IN_BYTES);
+			return -EINVAL;
+		}
+
+		priv->rss_conf.rss_key_len = rss_conf->rss_key_len;
+		priv->rss_conf.rss_key =
+			rte_zmalloc("mana_rss", rss_conf->rss_key_len,
+				    RTE_CACHE_LINE_SIZE);
+		if (!priv->rss_conf.rss_key)
+			return -ENOMEM;
+		memcpy(priv->rss_conf.rss_key, rss_conf->rss_key,
+		       rss_conf->rss_key_len);
+	}
+	priv->rss_conf.rss_hf = rss_conf->rss_hf;
+
+	return 0;
+}
+
+static int mana_rss_hash_conf_get(struct rte_eth_dev *dev,
+				  struct rte_eth_rss_conf *rss_conf)
+{
+	struct mana_priv *priv = dev->data->dev_private;
+
+	if (!rss_conf)
+		return -EINVAL;
+
+	if (rss_conf->rss_key &&
+	    rss_conf->rss_key_len >= priv->rss_conf.rss_key_len) {
+		memcpy(rss_conf->rss_key, priv->rss_conf.rss_key,
+		       priv->rss_conf.rss_key_len);
+	}
+
+	rss_conf->rss_key_len = priv->rss_conf.rss_key_len;
+	rss_conf->rss_hf = priv->rss_conf.rss_hf;
+
+	return 0;
+}
+
 static int mana_dev_link_update(struct rte_eth_dev *dev,
 				int wait_to_complete __rte_unused)
 {
@@ -232,6 +291,8 @@ const struct eth_dev_ops mana_dev_ops = {
 	.dev_close		= mana_dev_close,
 	.dev_infos_get		= mana_dev_info_get,
 	.dev_supported_ptypes_get = mana_supported_ptypes,
+	.rss_hash_update	= mana_rss_hash_update,
+	.rss_hash_conf_get	= mana_rss_hash_conf_get,
 	.link_update		= mana_dev_link_update,
 };
 
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index f97eed2e81..33f68b3d1b 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -72,6 +72,7 @@ struct mana_priv {
 	uint8_t ind_table_key[40];
 	struct ibv_qp *rwq_qp;
 	void *db_page;
+	struct rte_eth_rss_conf rss_conf;
 	struct rte_intr_handle *intr_handle;
 	int max_rx_queues;
 	int max_tx_queues;
-- 
2.17.1

[Patch v6 08/18] net/mana: add function to configure RX queues

[longli]

From: Long Li <longli@microsoft.com>

RX hardware queue is allocated when starting the queue. This function is
for queue configuration pre starting.

Signed-off-by: Long Li <longli@microsoft.com>
---
 drivers/net/mana/mana.c | 68 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 68 insertions(+)

diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index cb136f24c1..d03adab041 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -196,6 +196,16 @@ static int mana_dev_info_get(struct rte_eth_dev *dev,
 	return 0;
 }
 
+static void mana_dev_rx_queue_info(struct rte_eth_dev *dev, uint16_t queue_id,
+				   struct rte_eth_rxq_info *qinfo)
+{
+	struct mana_rxq *rxq = dev->data->rx_queues[queue_id];
+
+	qinfo->mp = rxq->mp;
+	qinfo->nb_desc = rxq->num_desc;
+	qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
+}
+
 static const uint32_t *mana_supported_ptypes(struct rte_eth_dev *dev __rte_unused)
 {
 	static const uint32_t ptypes[] = {
@@ -270,6 +280,61 @@ static int mana_rss_hash_conf_get(struct rte_eth_dev *dev,
 	return 0;
 }
 
+static int mana_dev_rx_queue_setup(struct rte_eth_dev *dev,
+				   uint16_t queue_idx, uint16_t nb_desc,
+				   unsigned int socket_id,
+				   const struct rte_eth_rxconf *rx_conf __rte_unused,
+				   struct rte_mempool *mp)
+{
+	struct mana_priv *priv = dev->data->dev_private;
+	struct mana_rxq *rxq;
+	int ret;
+
+	rxq = rte_zmalloc_socket("mana_rxq", sizeof(*rxq), 0, socket_id);
+	if (!rxq) {
+		DRV_LOG(ERR, "failed to allocate rxq");
+		return -ENOMEM;
+	}
+
+	DRV_LOG(DEBUG, "idx %u nb_desc %u socket %u",
+		queue_idx, nb_desc, socket_id);
+
+	rxq->socket = socket_id;
+
+	rxq->desc_ring = rte_zmalloc_socket("mana_rx_mbuf_ring",
+					    sizeof(struct mana_rxq_desc) *
+						nb_desc,
+					    RTE_CACHE_LINE_SIZE, socket_id);
+
+	if (!rxq->desc_ring) {
+		DRV_LOG(ERR, "failed to allocate rxq desc_ring");
+		ret = -ENOMEM;
+		goto fail;
+	}
+
+	rxq->num_desc = nb_desc;
+
+	rxq->priv = priv;
+	rxq->num_desc = nb_desc;
+	rxq->mp = mp;
+	dev->data->rx_queues[queue_idx] = rxq;
+
+	return 0;
+
+fail:
+	rte_free(rxq->desc_ring);
+	rte_free(rxq);
+	return ret;
+}
+
+static void mana_dev_rx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
+{
+	struct mana_rxq *rxq = dev->data->rx_queues[qid];
+
+	rte_free(rxq->desc_ring);
+	rte_free(rxq);
+}
+
 static int mana_dev_link_update(struct rte_eth_dev *dev,
 				int wait_to_complete __rte_unused)
 {
@@ -290,9 +355,12 @@ const struct eth_dev_ops mana_dev_ops = {
 	.dev_configure		= mana_dev_configure,
 	.dev_close		= mana_dev_close,
 	.dev_infos_get		= mana_dev_info_get,
+	.rxq_info_get		= mana_dev_rx_queue_info,
 	.dev_supported_ptypes_get = mana_supported_ptypes,
 	.rss_hash_update	= mana_rss_hash_update,
 	.rss_hash_conf_get	= mana_rss_hash_conf_get,
+	.rx_queue_setup		= mana_dev_rx_queue_setup,
+	.rx_queue_release	= mana_dev_rx_queue_release,
 	.link_update		= mana_dev_link_update,
 };
 
-- 
2.17.1

[Patch v6 09/18] net/mana: add function to configure TX queues

[longli]

From: Long Li <longli@microsoft.com>

TX hardware queue is allocated when starting the queue, this is for
pre configuration.

Signed-off-by: Long Li <longli@microsoft.com>
---
 drivers/net/mana/mana.c | 65 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 65 insertions(+)

diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index d03adab041..490686f404 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -196,6 +196,15 @@ static int mana_dev_info_get(struct rte_eth_dev *dev,
 	return 0;
 }
 
+static void mana_dev_tx_queue_info(struct rte_eth_dev *dev, uint16_t queue_id,
+			    struct rte_eth_txq_info *qinfo)
+{
+	struct mana_txq *txq = dev->data->tx_queues[queue_id];
+
+	qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads;
+	qinfo->nb_desc = txq->num_desc;
+}
+
 static void mana_dev_rx_queue_info(struct rte_eth_dev *dev, uint16_t queue_id,
 				   struct rte_eth_rxq_info *qinfo)
 {
@@ -280,6 +289,59 @@ static int mana_rss_hash_conf_get(struct rte_eth_dev *dev,
 	return 0;
 }
 
+static int mana_dev_tx_queue_setup(struct rte_eth_dev *dev,
+				   uint16_t queue_idx, uint16_t nb_desc,
+				   unsigned int socket_id,
+				   const struct rte_eth_txconf *tx_conf __rte_unused)
+
+{
+	struct mana_priv *priv = dev->data->dev_private;
+	struct mana_txq *txq;
+	int ret;
+
+	txq = rte_zmalloc_socket("mana_txq", sizeof(*txq), 0, socket_id);
+	if (!txq) {
+		DRV_LOG(ERR, "failed to allocate txq");
+		return -ENOMEM;
+	}
+
+	txq->socket = socket_id;
+
+	txq->desc_ring = rte_malloc_socket("mana_tx_desc_ring",
+					   sizeof(struct mana_txq_desc) *
+						nb_desc,
+					   RTE_CACHE_LINE_SIZE, socket_id);
+	if (!txq->desc_ring) {
+		DRV_LOG(ERR, "failed to allocate txq desc_ring");
+		ret = -ENOMEM;
+		goto fail;
+	}
+
+	DRV_LOG(DEBUG, "idx %u nb_desc %u socket %u txq->desc_ring %p",
+		queue_idx, nb_desc, socket_id, txq->desc_ring);
+
+	txq->desc_ring_head = 0;
+	txq->desc_ring_tail = 0;
+	txq->priv = priv;
+	txq->num_desc = nb_desc;
+	dev->data->tx_queues[queue_idx] = txq;
+
+	return 0;
+
+fail:
+	rte_free(txq->desc_ring);
+	rte_free(txq);
+	return ret;
+}
+
+static void mana_dev_tx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
+{
+	struct mana_txq *txq = dev->data->tx_queues[qid];
+
+	rte_free(txq->desc_ring);
+	rte_free(txq);
+}
+
 static int mana_dev_rx_queue_setup(struct rte_eth_dev *dev,
 				   uint16_t queue_idx, uint16_t nb_desc,
 				   unsigned int socket_id,
@@ -355,10 +417,13 @@ const struct eth_dev_ops mana_dev_ops = {
 	.dev_configure		= mana_dev_configure,
 	.dev_close		= mana_dev_close,
 	.dev_infos_get		= mana_dev_info_get,
+	.txq_info_get		= mana_dev_tx_queue_info,
 	.rxq_info_get		= mana_dev_rx_queue_info,
 	.dev_supported_ptypes_get = mana_supported_ptypes,
 	.rss_hash_update	= mana_rss_hash_update,
 	.rss_hash_conf_get	= mana_rss_hash_conf_get,
+	.tx_queue_setup		= mana_dev_tx_queue_setup,
+	.tx_queue_release	= mana_dev_tx_queue_release,
 	.rx_queue_setup		= mana_dev_rx_queue_setup,
 	.rx_queue_release	= mana_dev_rx_queue_release,
 	.link_update		= mana_dev_link_update,
-- 
2.17.1

[Patch v6 10/18] net/mana: implement memory registration

[longli]

From: Long Li <longli@microsoft.com>

MANA hardware has iommu built-in, that provides hardware safe access to
user memory through memory registration. Since memory registration is an
expensive operation, this patch implements a two level memory registration
cache mechanisum for each queue and for each port.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v2:
Change all header file functions to start with mana_.
Use spinlock in place of rwlock to memory cache access.
Remove unused header files.
v4:
Remove extra "\n" in logging function.

 drivers/net/mana/mana.c      |  20 +++
 drivers/net/mana/mana.h      |  39 +++++
 drivers/net/mana/meson.build |   1 +
 drivers/net/mana/mp.c        |  85 +++++++++
 drivers/net/mana/mr.c        | 324 +++++++++++++++++++++++++++++++++++
 5 files changed, 469 insertions(+)
 create mode 100644 drivers/net/mana/mr.c

diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index 490686f404..d18cc4ab0e 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -103,6 +103,8 @@ mana_dev_close(struct rte_eth_dev *dev)
 	struct mana_priv *priv = dev->data->dev_private;
 	int ret;
 
+	mana_remove_all_mr(priv);
+
 	ret = mana_intr_uninstall(priv);
 	if (ret)
 		return ret;
@@ -317,6 +319,13 @@ static int mana_dev_tx_queue_setup(struct rte_eth_dev *dev,
 		goto fail;
 	}
 
+	ret = mana_mr_btree_init(&txq->mr_btree,
+				 MANA_MR_BTREE_PER_QUEUE_N, socket_id);
+	if (ret) {
+		DRV_LOG(ERR, "Failed to init TXQ MR btree");
+		goto fail;
+	}
+
 	DRV_LOG(DEBUG, "idx %u nb_desc %u socket %u txq->desc_ring %p",
 		queue_idx, nb_desc, socket_id, txq->desc_ring);
 
@@ -338,6 +347,8 @@ static void mana_dev_tx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
 {
 	struct mana_txq *txq = dev->data->tx_queues[qid];
 
+	mana_mr_btree_free(&txq->mr_btree);
+
 	rte_free(txq->desc_ring);
 	rte_free(txq);
 }
@@ -374,6 +385,13 @@ static int mana_dev_rx_queue_setup(struct rte_eth_dev *dev,
 		goto fail;
 	}
 
+	ret = mana_mr_btree_init(&rxq->mr_btree,
+				 MANA_MR_BTREE_PER_QUEUE_N, socket_id);
+	if (ret) {
+		DRV_LOG(ERR, "Failed to init RXQ MR btree");
+		goto fail;
+	}
+
 	rxq->num_desc = nb_desc;
 
 	rxq->priv = priv;
@@ -393,6 +411,8 @@ static void mana_dev_rx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
 {
 	struct mana_rxq *rxq = dev->data->rx_queues[qid];
 
+	mana_mr_btree_free(&rxq->mr_btree);
+
 	rte_free(rxq->desc_ring);
 	rte_free(rxq);
 }
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index 33f68b3d1b..9e15b43275 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -50,6 +50,22 @@ struct mana_shared_data {
 #define MAX_RECEIVE_BUFFERS_PER_QUEUE	256
 #define MAX_SEND_BUFFERS_PER_QUEUE	256
 
+struct mana_mr_cache {
+	uint32_t	lkey;
+	uintptr_t	addr;
+	size_t		len;
+	void		*verb_obj;
+};
+
+#define MANA_MR_BTREE_CACHE_N	512
+struct mana_mr_btree {
+	uint16_t	len;	/* Used entries */
+	uint16_t	size;	/* Total entries */
+	int		overflow;
+	int		socket;
+	struct mana_mr_cache *table;
+};
+
 struct mana_process_priv {
 	void *db_page;
 };
@@ -82,6 +98,8 @@ struct mana_priv {
 	int max_recv_sge;
 	int max_mr;
 	uint64_t max_mr_size;
+	struct mana_mr_btree mr_btree;
+	rte_spinlock_t	mr_btree_lock;
 };
 
 struct mana_txq_desc {
@@ -131,6 +149,7 @@ struct mana_txq {
 	uint32_t desc_ring_head, desc_ring_tail;
 
 	struct mana_stats stats;
+	struct mana_mr_btree mr_btree;
 	unsigned int socket;
 };
 
@@ -153,6 +172,7 @@ struct mana_rxq {
 	struct mana_gdma_queue gdma_cq;
 
 	struct mana_stats stats;
+	struct mana_mr_btree mr_btree;
 
 	unsigned int socket;
 };
@@ -176,6 +196,24 @@ uint16_t mana_rx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
 uint16_t mana_tx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
 			       uint16_t pkts_n);
 
+struct mana_mr_cache *mana_find_pmd_mr(struct mana_mr_btree *local_tree,
+				       struct mana_priv *priv,
+				       struct rte_mbuf *mbuf);
+int mana_new_pmd_mr(struct mana_mr_btree *local_tree, struct mana_priv *priv,
+		    struct rte_mempool *pool);
+void mana_remove_all_mr(struct mana_priv *priv);
+void mana_del_pmd_mr(struct mana_mr_cache *mr);
+
+void mana_mempool_chunk_cb(struct rte_mempool *mp, void *opaque,
+			   struct rte_mempool_memhdr *memhdr, unsigned int idx);
+
+struct mana_mr_cache *mana_mr_btree_lookup(struct mana_mr_btree *bt,
+					   uint16_t *idx,
+					   uintptr_t addr, size_t len);
+int mana_mr_btree_insert(struct mana_mr_btree *bt, struct mana_mr_cache *entry);
+int mana_mr_btree_init(struct mana_mr_btree *bt, int n, int socket);
+void mana_mr_btree_free(struct mana_mr_btree *bt);
+
 /** Request timeout for IPC. */
 #define MANA_MP_REQ_TIMEOUT_SEC 5
 
@@ -204,6 +242,7 @@ int mana_mp_init_secondary(void);
 void mana_mp_uninit_primary(void);
 void mana_mp_uninit_secondary(void);
 int mana_mp_req_verbs_cmd_fd(struct rte_eth_dev *dev);
+int mana_mp_req_mr_create(struct mana_priv *priv, uintptr_t addr, uint32_t len);
 
 void mana_mp_req_on_rxtx(struct rte_eth_dev *dev, enum mana_mp_req_type type);
 
diff --git a/drivers/net/mana/meson.build b/drivers/net/mana/meson.build
index 81c4118f53..9771394370 100644
--- a/drivers/net/mana/meson.build
+++ b/drivers/net/mana/meson.build
@@ -11,6 +11,7 @@ deps += ['pci', 'bus_pci', 'net', 'eal', 'kvargs']
 
 sources += files(
 	'mana.c',
+	'mr.c',
 	'mp.c',
 )
 
diff --git a/drivers/net/mana/mp.c b/drivers/net/mana/mp.c
index d7580e8a28..f4f78d2787 100644
--- a/drivers/net/mana/mp.c
+++ b/drivers/net/mana/mp.c
@@ -12,6 +12,52 @@
 
 extern struct mana_shared_data *mana_shared_data;
 
+static int mana_mp_mr_create(struct mana_priv *priv, uintptr_t addr,
+			     uint32_t len)
+{
+	struct ibv_mr *ibv_mr;
+	int ret;
+	struct mana_mr_cache *mr;
+
+	ibv_mr = ibv_reg_mr(priv->ib_pd, (void *)addr, len,
+			    IBV_ACCESS_LOCAL_WRITE);
+
+	if (!ibv_mr)
+		return -errno;
+
+	DRV_LOG(DEBUG, "MR (2nd) lkey %u addr %p len %zu",
+		ibv_mr->lkey, ibv_mr->addr, ibv_mr->length);
+
+	mr = rte_calloc("MANA MR", 1, sizeof(*mr), 0);
+	if (!mr) {
+		DRV_LOG(ERR, "(2nd) Failed to allocate MR");
+		ret = -ENOMEM;
+		goto fail_alloc;
+	}
+	mr->lkey = ibv_mr->lkey;
+	mr->addr = (uintptr_t)ibv_mr->addr;
+	mr->len = ibv_mr->length;
+	mr->verb_obj = ibv_mr;
+
+	rte_spinlock_lock(&priv->mr_btree_lock);
+	ret = mana_mr_btree_insert(&priv->mr_btree, mr);
+	rte_spinlock_unlock(&priv->mr_btree_lock);
+	if (ret) {
+		DRV_LOG(ERR, "(2nd) Failed to add to global MR btree");
+		goto fail_btree;
+	}
+
+	return 0;
+
+fail_btree:
+	rte_free(mr);
+
+fail_alloc:
+	ibv_dereg_mr(ibv_mr);
+
+	return ret;
+}
+
 static void mp_init_msg(struct rte_mp_msg *msg, enum mana_mp_req_type type,
 			int port_id)
 {
@@ -47,6 +93,12 @@ static int mana_mp_primary_handle(const struct rte_mp_msg *mp_msg,
 	mp_init_msg(&mp_res, param->type, param->port_id);
 
 	switch (param->type) {
+	case MANA_MP_REQ_CREATE_MR:
+		ret = mana_mp_mr_create(priv, param->addr, param->len);
+		res->result = ret;
+		ret = rte_mp_reply(&mp_res, peer);
+		break;
+
 	case MANA_MP_REQ_VERBS_CMD_FD:
 		mp_res.num_fds = 1;
 		mp_res.fds[0] = priv->ib_ctx->cmd_fd;
@@ -189,6 +241,39 @@ int mana_mp_req_verbs_cmd_fd(struct rte_eth_dev *dev)
 	return ret;
 }
 
+int mana_mp_req_mr_create(struct mana_priv *priv, uintptr_t addr, uint32_t len)
+{
+	struct rte_mp_msg mp_req = { 0 };
+	struct rte_mp_msg *mp_res;
+	struct rte_mp_reply mp_rep;
+	struct mana_mp_param *req = (struct mana_mp_param *)mp_req.param;
+	struct mana_mp_param *res;
+	struct timespec ts = {.tv_sec = MANA_MP_REQ_TIMEOUT_SEC, .tv_nsec = 0};
+	int ret;
+
+	mp_init_msg(&mp_req, MANA_MP_REQ_CREATE_MR, priv->port_id);
+	req->addr = addr;
+	req->len = len;
+
+	ret = rte_mp_request_sync(&mp_req, &mp_rep, &ts);
+	if (ret) {
+		DRV_LOG(ERR, "Port %u request to primary failed",
+			req->port_id);
+		return ret;
+	}
+
+	if (mp_rep.nb_received != 1)
+		return -EPROTO;
+
+	mp_res = &mp_rep.msgs[0];
+	res = (struct mana_mp_param *)mp_res->param;
+	ret = res->result;
+
+	free(mp_rep.msgs);
+
+	return ret;
+}
+
 void mana_mp_req_on_rxtx(struct rte_eth_dev *dev, enum mana_mp_req_type type)
 {
 	struct rte_mp_msg mp_req = { 0 };
diff --git a/drivers/net/mana/mr.c b/drivers/net/mana/mr.c
new file mode 100644
index 0000000000..81b64b840f
--- /dev/null
+++ b/drivers/net/mana/mr.c
@@ -0,0 +1,324 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2022 Microsoft Corporation
+ */
+
+#include <rte_malloc.h>
+#include <ethdev_driver.h>
+#include <rte_eal_paging.h>
+
+#include <infiniband/verbs.h>
+
+#include "mana.h"
+
+struct mana_range {
+	uintptr_t	start;
+	uintptr_t	end;
+	uint32_t	len;
+};
+
+void mana_mempool_chunk_cb(struct rte_mempool *mp __rte_unused, void *opaque,
+			   struct rte_mempool_memhdr *memhdr, unsigned int idx)
+{
+	struct mana_range *ranges = opaque;
+	struct mana_range *range = &ranges[idx];
+	uint64_t page_size = rte_mem_page_size();
+
+	range->start = RTE_ALIGN_FLOOR((uintptr_t)memhdr->addr, page_size);
+	range->end = RTE_ALIGN_CEIL((uintptr_t)memhdr->addr + memhdr->len,
+				    page_size);
+	range->len = range->end - range->start;
+}
+
+int mana_new_pmd_mr(struct mana_mr_btree *local_tree, struct mana_priv *priv,
+		    struct rte_mempool *pool)
+{
+	struct ibv_mr *ibv_mr;
+	struct mana_range ranges[pool->nb_mem_chunks];
+	uint32_t i;
+	struct mana_mr_cache *mr;
+	int ret;
+
+	rte_mempool_mem_iter(pool, mana_mempool_chunk_cb, ranges);
+
+	for (i = 0; i < pool->nb_mem_chunks; i++) {
+		if (ranges[i].len > priv->max_mr_size) {
+			DRV_LOG(ERR, "memory chunk size %u exceeding max MR",
+				ranges[i].len);
+			return -ENOMEM;
+		}
+
+		DRV_LOG(DEBUG,
+			"registering memory chunk start 0x%" PRIx64 " len %u",
+			ranges[i].start, ranges[i].len);
+
+		if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
+			/* Send a message to the primary to do MR */
+			ret = mana_mp_req_mr_create(priv, ranges[i].start,
+						    ranges[i].len);
+			if (ret) {
+				DRV_LOG(ERR,
+					"MR failed start 0x%" PRIx64 " len %u",
+					ranges[i].start, ranges[i].len);
+				return ret;
+			}
+			continue;
+		}
+
+		ibv_mr = ibv_reg_mr(priv->ib_pd, (void *)ranges[i].start,
+				    ranges[i].len, IBV_ACCESS_LOCAL_WRITE);
+		if (ibv_mr) {
+			DRV_LOG(DEBUG, "MR lkey %u addr %p len %" PRIu64,
+				ibv_mr->lkey, ibv_mr->addr, ibv_mr->length);
+
+			mr = rte_calloc("MANA MR", 1, sizeof(*mr), 0);
+			mr->lkey = ibv_mr->lkey;
+			mr->addr = (uintptr_t)ibv_mr->addr;
+			mr->len = ibv_mr->length;
+			mr->verb_obj = ibv_mr;
+
+			rte_spinlock_lock(&priv->mr_btree_lock);
+			ret = mana_mr_btree_insert(&priv->mr_btree, mr);
+			rte_spinlock_unlock(&priv->mr_btree_lock);
+			if (ret) {
+				ibv_dereg_mr(ibv_mr);
+				DRV_LOG(ERR, "Failed to add to global MR btree");
+				return ret;
+			}
+
+			ret = mana_mr_btree_insert(local_tree, mr);
+			if (ret) {
+				/* Don't need to clean up MR as it's already
+				 * in the global tree
+				 */
+				DRV_LOG(ERR, "Failed to add to local MR btree");
+				return ret;
+			}
+		} else {
+			DRV_LOG(ERR, "MR failed at 0x%" PRIx64 " len %u",
+				ranges[i].start, ranges[i].len);
+			return -errno;
+		}
+	}
+	return 0;
+}
+
+void mana_del_pmd_mr(struct mana_mr_cache *mr)
+{
+	int ret;
+	struct ibv_mr *ibv_mr = (struct ibv_mr *)mr->verb_obj;
+
+	ret = ibv_dereg_mr(ibv_mr);
+	if (ret)
+		DRV_LOG(ERR, "dereg MR failed ret %d", ret);
+}
+
+struct mana_mr_cache *mana_find_pmd_mr(struct mana_mr_btree *local_mr_btree,
+				       struct mana_priv *priv,
+				       struct rte_mbuf *mbuf)
+{
+	struct rte_mempool *pool = mbuf->pool;
+	int ret, second_try = 0;
+	struct mana_mr_cache *mr;
+	uint16_t idx;
+
+	DRV_LOG(DEBUG, "finding mr for mbuf addr %p len %d",
+		mbuf->buf_addr, mbuf->buf_len);
+
+try_again:
+	/* First try to find the MR in local queue tree */
+	mr = mana_mr_btree_lookup(local_mr_btree, &idx,
+				  (uintptr_t)mbuf->buf_addr, mbuf->buf_len);
+	if (mr) {
+		DRV_LOG(DEBUG,
+			"Local mr lkey %u addr 0x%" PRIx64 " len %" PRIu64,
+			mr->lkey, mr->addr, mr->len);
+		return mr;
+	}
+
+	/* If not found, try to find the MR in global tree */
+	rte_spinlock_lock(&priv->mr_btree_lock);
+	mr = mana_mr_btree_lookup(&priv->mr_btree, &idx,
+				  (uintptr_t)mbuf->buf_addr,
+				  mbuf->buf_len);
+	rte_spinlock_unlock(&priv->mr_btree_lock);
+
+	/* If found in the global tree, add it to the local tree */
+	if (mr) {
+		ret = mana_mr_btree_insert(local_mr_btree, mr);
+		if (ret) {
+			DRV_LOG(DEBUG, "Failed to add MR to local tree.");
+			return NULL;
+		}
+
+		DRV_LOG(DEBUG,
+			"Added local MR key %u addr 0x%" PRIx64 " len %" PRIu64,
+			mr->lkey, mr->addr, mr->len);
+		return mr;
+	}
+
+	if (second_try) {
+		DRV_LOG(ERR, "Internal error second try failed");
+		return NULL;
+	}
+
+	ret = mana_new_pmd_mr(local_mr_btree, priv, pool);
+	if (ret) {
+		DRV_LOG(ERR, "Failed to allocate MR ret %d addr %p len %d",
+			ret, mbuf->buf_addr, mbuf->buf_len);
+		return NULL;
+	}
+
+	second_try = 1;
+	goto try_again;
+}
+
+void mana_remove_all_mr(struct mana_priv *priv)
+{
+	struct mana_mr_btree *bt = &priv->mr_btree;
+	struct mana_mr_cache *mr;
+	struct ibv_mr *ibv_mr;
+	uint16_t i;
+
+	rte_spinlock_lock(&priv->mr_btree_lock);
+	/* Start with index 1 as the 1st entry is always NULL */
+	for (i = 1; i < bt->len; i++) {
+		mr = &bt->table[i];
+		ibv_mr = mr->verb_obj;
+		ibv_dereg_mr(ibv_mr);
+	}
+	bt->len = 1;
+	rte_spinlock_unlock(&priv->mr_btree_lock);
+}
+
+static int mana_mr_btree_expand(struct mana_mr_btree *bt, int n)
+{
+	void *mem;
+
+	mem = rte_realloc_socket(bt->table, n * sizeof(struct mana_mr_cache),
+				 0, bt->socket);
+	if (!mem) {
+		DRV_LOG(ERR, "Failed to expand btree size %d", n);
+		return -1;
+	}
+
+	DRV_LOG(ERR, "Expanded btree to size %d", n);
+	bt->table = mem;
+	bt->size = n;
+
+	return 0;
+}
+
+struct mana_mr_cache *mana_mr_btree_lookup(struct mana_mr_btree *bt,
+					   uint16_t *idx,
+					   uintptr_t addr, size_t len)
+{
+	struct mana_mr_cache *table;
+	uint16_t n;
+	uint16_t base = 0;
+	int ret;
+
+	n = bt->len;
+
+	/* Try to double the cache if it's full */
+	if (n == bt->size) {
+		ret = mana_mr_btree_expand(bt, bt->size << 1);
+		if (ret)
+			return NULL;
+	}
+
+	table = bt->table;
+
+	/* Do binary search on addr */
+	do {
+		uint16_t delta = n >> 1;
+
+		if (addr < table[base + delta].addr) {
+			n = delta;
+		} else {
+			base += delta;
+			n -= delta;
+		}
+	} while (n > 1);
+
+	*idx = base;
+
+	if (addr + len <= table[base].addr + table[base].len)
+		return &table[base];
+
+	DRV_LOG(DEBUG,
+		"addr 0x%" PRIx64 " len %zu idx %u sum 0x%" PRIx64 " not found",
+		addr, len, *idx, addr + len);
+
+	return NULL;
+}
+
+int mana_mr_btree_init(struct mana_mr_btree *bt, int n, int socket)
+{
+	memset(bt, 0, sizeof(*bt));
+	bt->table = rte_calloc_socket("MANA B-tree table",
+				      n,
+				      sizeof(struct mana_mr_cache),
+				      0, socket);
+	if (!bt->table) {
+		DRV_LOG(ERR, "Failed to allocate B-tree n %d socket %d",
+			n, socket);
+		return -ENOMEM;
+	}
+
+	bt->socket = socket;
+	bt->size = n;
+
+	/* First entry must be NULL for binary search to work */
+	bt->table[0] = (struct mana_mr_cache) {
+		.lkey = UINT32_MAX,
+	};
+	bt->len = 1;
+
+	DRV_LOG(ERR, "B-tree initialized table %p size %d len %d",
+		bt->table, n, bt->len);
+
+	return 0;
+}
+
+void mana_mr_btree_free(struct mana_mr_btree *bt)
+{
+	rte_free(bt->table);
+	memset(bt, 0, sizeof(*bt));
+}
+
+int mana_mr_btree_insert(struct mana_mr_btree *bt, struct mana_mr_cache *entry)
+{
+	struct mana_mr_cache *table;
+	uint16_t idx = 0;
+	uint16_t shift;
+
+	if (mana_mr_btree_lookup(bt, &idx, entry->addr, entry->len)) {
+		DRV_LOG(DEBUG, "Addr 0x%" PRIx64 " len %zu exists in btree",
+			entry->addr, entry->len);
+		return 0;
+	}
+
+	if (bt->len >= bt->size) {
+		bt->overflow = 1;
+		return -1;
+	}
+
+	table = bt->table;
+
+	idx++;
+	shift = (bt->len - idx) * sizeof(struct mana_mr_cache);
+	if (shift) {
+		DRV_LOG(DEBUG, "Moving %u bytes from idx %u to %u",
+			shift, idx, idx + 1);
+		memmove(&table[idx + 1], &table[idx], shift);
+	}
+
+	table[idx] = *entry;
+	bt->len++;
+
+	DRV_LOG(DEBUG,
+		"Inserted MR b-tree table %p idx %d addr 0x%" PRIx64 " len %zu",
+		table, idx, entry->addr, entry->len);
+
+	return 0;
+}
-- 
2.17.1

[Patch v6 11/18] net/mana: implement the hardware layer operations

[longli]

From: Long Li <longli@microsoft.com>

The hardware layer of MANA understands the device queue and doorbell
formats. Those functions are implemented for use by packet RX/TX code.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v2:
Remove unused header files.
Rename a camel case.
v5:
Use RTE_BIT32() instead of defining a new BIT()
v6:
add rte_rmb() after reading owner bits

 drivers/net/mana/gdma.c      | 289 +++++++++++++++++++++++++++++++++++
 drivers/net/mana/mana.h      | 183 ++++++++++++++++++++++
 drivers/net/mana/meson.build |   1 +
 3 files changed, 473 insertions(+)
 create mode 100644 drivers/net/mana/gdma.c

diff --git a/drivers/net/mana/gdma.c b/drivers/net/mana/gdma.c
new file mode 100644
index 0000000000..7ad175651e
--- /dev/null
+++ b/drivers/net/mana/gdma.c
@@ -0,0 +1,289 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2022 Microsoft Corporation
+ */
+
+#include <ethdev_driver.h>
+#include <rte_io.h>
+
+#include "mana.h"
+
+uint8_t *gdma_get_wqe_pointer(struct mana_gdma_queue *queue)
+{
+	uint32_t offset_in_bytes =
+		(queue->head * GDMA_WQE_ALIGNMENT_UNIT_SIZE) &
+		(queue->size - 1);
+
+	DRV_LOG(DEBUG, "txq sq_head %u sq_size %u offset_in_bytes %u",
+		queue->head, queue->size, offset_in_bytes);
+
+	if (offset_in_bytes + GDMA_WQE_ALIGNMENT_UNIT_SIZE > queue->size)
+		DRV_LOG(ERR, "fatal error: offset_in_bytes %u too big",
+			offset_in_bytes);
+
+	return ((uint8_t *)queue->buffer) + offset_in_bytes;
+}
+
+static uint32_t
+write_dma_client_oob(uint8_t *work_queue_buffer_pointer,
+		     const struct gdma_work_request *work_request,
+		     uint32_t client_oob_size)
+{
+	uint8_t *p = work_queue_buffer_pointer;
+
+	struct gdma_wqe_dma_oob *header = (struct gdma_wqe_dma_oob *)p;
+
+	memset(header, 0, sizeof(struct gdma_wqe_dma_oob));
+	header->num_sgl_entries = work_request->num_sgl_elements;
+	header->inline_client_oob_size_in_dwords =
+		client_oob_size / sizeof(uint32_t);
+	header->client_data_unit = work_request->client_data_unit;
+
+	DRV_LOG(DEBUG, "queue buf %p sgl %u oob_h %u du %u oob_buf %p oob_b %u",
+		work_queue_buffer_pointer, header->num_sgl_entries,
+		header->inline_client_oob_size_in_dwords,
+		header->client_data_unit, work_request->inline_oob_data,
+		work_request->inline_oob_size_in_bytes);
+
+	p += sizeof(struct gdma_wqe_dma_oob);
+	if (work_request->inline_oob_data &&
+	    work_request->inline_oob_size_in_bytes > 0) {
+		memcpy(p, work_request->inline_oob_data,
+		       work_request->inline_oob_size_in_bytes);
+		if (client_oob_size > work_request->inline_oob_size_in_bytes)
+			memset(p + work_request->inline_oob_size_in_bytes, 0,
+			       client_oob_size -
+			       work_request->inline_oob_size_in_bytes);
+	}
+
+	return sizeof(struct gdma_wqe_dma_oob) + client_oob_size;
+}
+
+static uint32_t
+write_scatter_gather_list(uint8_t *work_queue_head_pointer,
+			  uint8_t *work_queue_end_pointer,
+			  uint8_t *work_queue_cur_pointer,
+			  struct gdma_work_request *work_request)
+{
+	struct gdma_sgl_element *sge_list;
+	struct gdma_sgl_element dummy_sgl[1];
+	uint8_t *address;
+	uint32_t size;
+	uint32_t num_sge;
+	uint32_t size_to_queue_end;
+	uint32_t sge_list_size;
+
+	DRV_LOG(DEBUG, "work_queue_cur_pointer %p work_request->flags %x",
+		work_queue_cur_pointer, work_request->flags);
+
+	num_sge = work_request->num_sgl_elements;
+	sge_list = work_request->sgl;
+	size_to_queue_end = (uint32_t)(work_queue_end_pointer -
+				       work_queue_cur_pointer);
+
+	if (num_sge == 0) {
+		/* Per spec, the case of an empty SGL should be handled as
+		 * follows to avoid corrupted WQE errors:
+		 * Write one dummy SGL entry
+		 * Set the address to 1, leave the rest as 0
+		 */
+		dummy_sgl[num_sge].address = 1;
+		dummy_sgl[num_sge].size = 0;
+		dummy_sgl[num_sge].memory_key = 0;
+		num_sge++;
+		sge_list = dummy_sgl;
+	}
+
+	sge_list_size = 0;
+	{
+		address = (uint8_t *)sge_list;
+		size = sizeof(struct gdma_sgl_element) * num_sge;
+		if (size_to_queue_end < size) {
+			memcpy(work_queue_cur_pointer, address,
+			       size_to_queue_end);
+			work_queue_cur_pointer = work_queue_head_pointer;
+			address += size_to_queue_end;
+			size -= size_to_queue_end;
+		}
+
+		memcpy(work_queue_cur_pointer, address, size);
+		sge_list_size = size;
+	}
+
+	DRV_LOG(DEBUG, "sge %u address 0x%" PRIx64 " size %u key %u list_s %u",
+		num_sge, sge_list->address, sge_list->size,
+		sge_list->memory_key, sge_list_size);
+
+	return sge_list_size;
+}
+
+int gdma_post_work_request(struct mana_gdma_queue *queue,
+			   struct gdma_work_request *work_req,
+			   struct gdma_posted_wqe_info *wqe_info)
+{
+	uint32_t client_oob_size =
+		work_req->inline_oob_size_in_bytes >
+				INLINE_OOB_SMALL_SIZE_IN_BYTES ?
+			INLINE_OOB_LARGE_SIZE_IN_BYTES :
+			INLINE_OOB_SMALL_SIZE_IN_BYTES;
+
+	uint32_t sgl_data_size = sizeof(struct gdma_sgl_element) *
+			RTE_MAX((uint32_t)1, work_req->num_sgl_elements);
+	uint32_t wqe_size =
+		RTE_ALIGN(sizeof(struct gdma_wqe_dma_oob) +
+				client_oob_size + sgl_data_size,
+			  GDMA_WQE_ALIGNMENT_UNIT_SIZE);
+	uint8_t *wq_buffer_pointer;
+	uint32_t queue_free_units = queue->count - (queue->head - queue->tail);
+
+	if (wqe_size / GDMA_WQE_ALIGNMENT_UNIT_SIZE > queue_free_units) {
+		DRV_LOG(DEBUG, "WQE size %u queue count %u head %u tail %u",
+			wqe_size, queue->count, queue->head, queue->tail);
+		return -EBUSY;
+	}
+
+	DRV_LOG(DEBUG, "client_oob_size %u sgl_data_size %u wqe_size %u",
+		client_oob_size, sgl_data_size, wqe_size);
+
+	if (wqe_info) {
+		wqe_info->wqe_index =
+			((queue->head * GDMA_WQE_ALIGNMENT_UNIT_SIZE) &
+			 (queue->size - 1)) / GDMA_WQE_ALIGNMENT_UNIT_SIZE;
+		wqe_info->unmasked_queue_offset = queue->head;
+		wqe_info->wqe_size_in_bu =
+			wqe_size / GDMA_WQE_ALIGNMENT_UNIT_SIZE;
+	}
+
+	wq_buffer_pointer = gdma_get_wqe_pointer(queue);
+	wq_buffer_pointer += write_dma_client_oob(wq_buffer_pointer, work_req,
+						  client_oob_size);
+	if (wq_buffer_pointer >= ((uint8_t *)queue->buffer) + queue->size)
+		wq_buffer_pointer -= queue->size;
+
+	write_scatter_gather_list((uint8_t *)queue->buffer,
+				  (uint8_t *)queue->buffer + queue->size,
+				  wq_buffer_pointer, work_req);
+
+	queue->head += wqe_size / GDMA_WQE_ALIGNMENT_UNIT_SIZE;
+
+	return 0;
+}
+
+union gdma_doorbell_entry {
+	uint64_t     as_uint64;
+
+	struct {
+		uint64_t id	  : 24;
+		uint64_t reserved    : 8;
+		uint64_t tail_ptr    : 31;
+		uint64_t arm	 : 1;
+	} cq;
+
+	struct {
+		uint64_t id	  : 24;
+		uint64_t wqe_cnt     : 8;
+		uint64_t tail_ptr    : 32;
+	} rq;
+
+	struct {
+		uint64_t id	  : 24;
+		uint64_t reserved    : 8;
+		uint64_t tail_ptr    : 32;
+	} sq;
+
+	struct {
+		uint64_t id	  : 16;
+		uint64_t reserved    : 16;
+		uint64_t tail_ptr    : 31;
+		uint64_t arm	 : 1;
+	} eq;
+}; /* HW DATA */
+
+#define DOORBELL_OFFSET_SQ      0x0
+#define DOORBELL_OFFSET_RQ      0x400
+#define DOORBELL_OFFSET_CQ      0x800
+#define DOORBELL_OFFSET_EQ      0xFF8
+
+int mana_ring_doorbell(void *db_page, enum gdma_queue_types queue_type,
+		       uint32_t queue_id, uint32_t tail)
+{
+	uint8_t *addr = db_page;
+	union gdma_doorbell_entry e = {};
+
+	switch (queue_type) {
+	case gdma_queue_send:
+		e.sq.id = queue_id;
+		e.sq.tail_ptr = tail;
+		addr += DOORBELL_OFFSET_SQ;
+		break;
+
+	case gdma_queue_receive:
+		e.rq.id = queue_id;
+		e.rq.tail_ptr = tail;
+		e.rq.wqe_cnt = 1;
+		addr += DOORBELL_OFFSET_RQ;
+		break;
+
+	case gdma_queue_completion:
+		e.cq.id = queue_id;
+		e.cq.tail_ptr = tail;
+		e.cq.arm = 1;
+		addr += DOORBELL_OFFSET_CQ;
+		break;
+
+	default:
+		DRV_LOG(ERR, "Unsupported queue type %d", queue_type);
+		return -1;
+	}
+
+	/* Ensure all writes are done before ringing doorbell */
+	rte_wmb();
+
+	DRV_LOG(DEBUG, "db_page %p addr %p queue_id %u type %u tail %u",
+		db_page, addr, queue_id, queue_type, tail);
+
+	rte_write64(e.as_uint64, addr);
+	return 0;
+}
+
+int gdma_poll_completion_queue(struct mana_gdma_queue *cq,
+			       struct gdma_comp *comp)
+{
+	struct gdma_hardware_completion_entry *cqe;
+	uint32_t head = cq->head % cq->count;
+	uint32_t new_owner_bits, old_owner_bits;
+	uint32_t cqe_owner_bits;
+	struct gdma_hardware_completion_entry *buffer = cq->buffer;
+
+	cqe = &buffer[head];
+	new_owner_bits = (cq->head / cq->count) & COMPLETION_QUEUE_OWNER_MASK;
+	old_owner_bits = (cq->head / cq->count - 1) &
+				COMPLETION_QUEUE_OWNER_MASK;
+	cqe_owner_bits = cqe->owner_bits;
+
+	DRV_LOG(DEBUG, "comp cqe bits 0x%x owner bits 0x%x",
+		cqe_owner_bits, old_owner_bits);
+
+	if (cqe_owner_bits == old_owner_bits)
+		return 0; /* No new entry */
+
+	if (cqe_owner_bits != new_owner_bits) {
+		DRV_LOG(ERR, "CQ overflowed, ID %u cqe 0x%x new 0x%x",
+			cq->id, cqe_owner_bits, new_owner_bits);
+		return -1;
+	}
+
+	/* Ensure checking owner bits happens before reading from CQE */
+	rte_rmb();
+
+	comp->work_queue_number = cqe->wq_num;
+	comp->send_work_queue = cqe->is_sq;
+
+	memcpy(comp->completion_data, cqe->dma_client_data, GDMA_COMP_DATA_SIZE);
+
+	cq->head++;
+
+	DRV_LOG(DEBUG, "comp new 0x%x old 0x%x cqe 0x%x wq %u sq %u head %u",
+		new_owner_bits, old_owner_bits, cqe_owner_bits,
+		comp->work_queue_number, comp->send_work_queue, cq->head);
+	return 1;
+}
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index 9e15b43275..d87358ab15 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -50,6 +50,178 @@ struct mana_shared_data {
 #define MAX_RECEIVE_BUFFERS_PER_QUEUE	256
 #define MAX_SEND_BUFFERS_PER_QUEUE	256
 
+#define GDMA_WQE_ALIGNMENT_UNIT_SIZE 32
+
+#define COMP_ENTRY_SIZE 64
+#define MAX_TX_WQE_SIZE 512
+#define MAX_RX_WQE_SIZE 256
+
+/* Values from the GDMA specification document, WQE format description */
+#define INLINE_OOB_SMALL_SIZE_IN_BYTES 8
+#define INLINE_OOB_LARGE_SIZE_IN_BYTES 24
+
+#define NOT_USING_CLIENT_DATA_UNIT 0
+
+enum gdma_queue_types {
+	gdma_queue_type_invalid = 0,
+	gdma_queue_send,
+	gdma_queue_receive,
+	gdma_queue_completion,
+	gdma_queue_event,
+	gdma_queue_type_max = 16,
+	/*Room for expansion */
+
+	/* This enum can be expanded to add more queue types but
+	 * it's expected to be done in a contiguous manner.
+	 * Failing that will result in unexpected behavior.
+	 */
+};
+
+#define WORK_QUEUE_NUMBER_BASE_BITS 10
+
+struct gdma_header {
+	/* size of the entire gdma structure, including the entire length of
+	 * the struct that is formed by extending other gdma struct. i.e.
+	 * GDMA_BASE_SPEC extends gdma_header, GDMA_EVENT_QUEUE_SPEC extends
+	 * GDMA_BASE_SPEC, StructSize for GDMA_EVENT_QUEUE_SPEC will be size of
+	 * GDMA_EVENT_QUEUE_SPEC which includes size of GDMA_BASE_SPEC and size
+	 * of gdma_header.
+	 * Above example is for illustration purpose and is not in code
+	 */
+	size_t struct_size;
+};
+
+/* The following macros are from GDMA SPEC 3.6, "Table 2: CQE data structure"
+ * and "Table 4: Event Queue Entry (EQE) data format"
+ */
+#define GDMA_COMP_DATA_SIZE 0x3C /* Must be a multiple of 4 */
+#define GDMA_COMP_DATA_SIZE_IN_UINT32 (GDMA_COMP_DATA_SIZE / 4)
+
+#define COMPLETION_QUEUE_ENTRY_WORK_QUEUE_INDEX 0
+#define COMPLETION_QUEUE_ENTRY_WORK_QUEUE_SIZE 24
+#define COMPLETION_QUEUE_ENTRY_SEND_WORK_QUEUE_INDEX 24
+#define COMPLETION_QUEUE_ENTRY_SEND_WORK_QUEUE_SIZE 1
+#define COMPLETION_QUEUE_ENTRY_OWNER_BITS_INDEX 29
+#define COMPLETION_QUEUE_ENTRY_OWNER_BITS_SIZE 3
+
+#define COMPLETION_QUEUE_OWNER_MASK \
+	((1 << (COMPLETION_QUEUE_ENTRY_OWNER_BITS_SIZE)) - 1)
+
+struct gdma_comp {
+	struct gdma_header gdma_header;
+
+	/* Filled by GDMA core */
+	uint32_t completion_data[GDMA_COMP_DATA_SIZE_IN_UINT32];
+
+	/* Filled by GDMA core */
+	uint32_t work_queue_number;
+
+	/* Filled by GDMA core */
+	bool send_work_queue;
+};
+
+struct gdma_hardware_completion_entry {
+	char dma_client_data[GDMA_COMP_DATA_SIZE];
+	union {
+		uint32_t work_queue_owner_bits;
+		struct {
+			uint32_t wq_num		: 24;
+			uint32_t is_sq		: 1;
+			uint32_t reserved	: 4;
+			uint32_t owner_bits	: 3;
+		};
+	};
+}; /* HW DATA */
+
+struct gdma_posted_wqe_info {
+	struct gdma_header gdma_header;
+
+	/* size of the written wqe in basic units (32B), filled by GDMA core.
+	 * Use this value to progress the work queue after the wqe is processed
+	 * by hardware.
+	 */
+	uint32_t wqe_size_in_bu;
+
+	/* At the time of writing the wqe to the work queue, the offset in the
+	 * work queue buffer where by the wqe will be written. Each unit
+	 * represents 32B of buffer space.
+	 */
+	uint32_t wqe_index;
+
+	/* Unmasked offset in the queue to which the WQE was written.
+	 * In 32 byte units.
+	 */
+	uint32_t unmasked_queue_offset;
+};
+
+struct gdma_sgl_element {
+	uint64_t address;
+	uint32_t memory_key;
+	uint32_t size;
+};
+
+#define MAX_SGL_ENTRIES_FOR_TRANSMIT 30
+
+struct one_sgl {
+	struct gdma_sgl_element gdma_sgl[MAX_SGL_ENTRIES_FOR_TRANSMIT];
+};
+
+struct gdma_work_request {
+	struct gdma_header gdma_header;
+	struct gdma_sgl_element *sgl;
+	uint32_t num_sgl_elements;
+	uint32_t inline_oob_size_in_bytes;
+	void *inline_oob_data;
+	uint32_t flags; /* From _gdma_work_request_FLAGS */
+	uint32_t client_data_unit; /* For LSO, this is the MTU of the data */
+};
+
+enum mana_cqe_type {
+	CQE_INVALID                     = 0,
+};
+
+struct mana_cqe_header {
+	uint32_t cqe_type    : 6;
+	uint32_t client_type : 2;
+	uint32_t vendor_err  : 24;
+}; /* HW DATA */
+
+/* NDIS HASH Types */
+#define BIT(nr)		(1 << (nr))
+#define NDIS_HASH_IPV4          BIT(0)
+#define NDIS_HASH_TCP_IPV4      BIT(1)
+#define NDIS_HASH_UDP_IPV4      BIT(2)
+#define NDIS_HASH_IPV6          BIT(3)
+#define NDIS_HASH_TCP_IPV6      BIT(4)
+#define NDIS_HASH_UDP_IPV6      BIT(5)
+#define NDIS_HASH_IPV6_EX       BIT(6)
+#define NDIS_HASH_TCP_IPV6_EX   BIT(7)
+#define NDIS_HASH_UDP_IPV6_EX   BIT(8)
+
+#define MANA_HASH_L3 (NDIS_HASH_IPV4 | NDIS_HASH_IPV6 | NDIS_HASH_IPV6_EX)
+#define MANA_HASH_L4                                                         \
+	(NDIS_HASH_TCP_IPV4 | NDIS_HASH_UDP_IPV4 | NDIS_HASH_TCP_IPV6 |      \
+	 NDIS_HASH_UDP_IPV6 | NDIS_HASH_TCP_IPV6_EX | NDIS_HASH_UDP_IPV6_EX)
+
+struct gdma_wqe_dma_oob {
+	uint32_t reserved:24;
+	uint32_t last_v_bytes:8;
+	union {
+		uint32_t flags;
+		struct {
+			uint32_t num_sgl_entries:8;
+			uint32_t inline_client_oob_size_in_dwords:3;
+			uint32_t client_oob_in_sgl:1;
+			uint32_t consume_credit:1;
+			uint32_t fence:1;
+			uint32_t reserved1:2;
+			uint32_t client_data_unit:14;
+			uint32_t check_sn:1;
+			uint32_t sgl_direct:1;
+		};
+	};
+};
+
 struct mana_mr_cache {
 	uint32_t	lkey;
 	uintptr_t	addr;
@@ -190,12 +362,23 @@ extern int mana_logtype_init;
 
 #define PMD_INIT_FUNC_TRACE() PMD_INIT_LOG(DEBUG, " >>")
 
+int mana_ring_doorbell(void *db_page, enum gdma_queue_types queue_type,
+		       uint32_t queue_id, uint32_t tail);
+
+int gdma_post_work_request(struct mana_gdma_queue *queue,
+			   struct gdma_work_request *work_req,
+			   struct gdma_posted_wqe_info *wqe_info);
+uint8_t *gdma_get_wqe_pointer(struct mana_gdma_queue *queue);
+
 uint16_t mana_rx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
 			       uint16_t pkts_n);
 
 uint16_t mana_tx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
 			       uint16_t pkts_n);
 
+int gdma_poll_completion_queue(struct mana_gdma_queue *cq,
+			       struct gdma_comp *comp);
+
 struct mana_mr_cache *mana_find_pmd_mr(struct mana_mr_btree *local_tree,
 				       struct mana_priv *priv,
 				       struct rte_mbuf *mbuf);
diff --git a/drivers/net/mana/meson.build b/drivers/net/mana/meson.build
index 9771394370..364d57a619 100644
--- a/drivers/net/mana/meson.build
+++ b/drivers/net/mana/meson.build
@@ -12,6 +12,7 @@ deps += ['pci', 'bus_pci', 'net', 'eal', 'kvargs']
 sources += files(
 	'mana.c',
 	'mr.c',
+	'gdma.c',
 	'mp.c',
 )
 
-- 
2.17.1

[Patch v6 12/18] net/mana: add function to start/stop TX queues

[longli]

From: Long Li <longli@microsoft.com>

MANA allocate device queues through the IB layer when starting TX queues.
When device is stopped all the queues are unmapped and freed.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v2:
Add prefix mana_ to all function names.
Remove unused header files.

 doc/guides/nics/features/mana.ini |   1 +
 drivers/net/mana/mana.h           |   4 +
 drivers/net/mana/meson.build      |   1 +
 drivers/net/mana/tx.c             | 163 ++++++++++++++++++++++++++++++
 4 files changed, 169 insertions(+)
 create mode 100644 drivers/net/mana/tx.c

diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
index a59c21cc10..821443b292 100644
--- a/doc/guides/nics/features/mana.ini
+++ b/doc/guides/nics/features/mana.ini
@@ -7,6 +7,7 @@
 Link status          = P
 Linux                = Y
 Multiprocess aware   = Y
+Queue start/stop     = Y
 Removal event        = Y
 RSS hash             = Y
 Speed capabilities   = P
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index d87358ab15..3613ba7ca2 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -379,6 +379,10 @@ uint16_t mana_tx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
 int gdma_poll_completion_queue(struct mana_gdma_queue *cq,
 			       struct gdma_comp *comp);
 
+int mana_start_tx_queues(struct rte_eth_dev *dev);
+
+int mana_stop_tx_queues(struct rte_eth_dev *dev);
+
 struct mana_mr_cache *mana_find_pmd_mr(struct mana_mr_btree *local_tree,
 				       struct mana_priv *priv,
 				       struct rte_mbuf *mbuf);
diff --git a/drivers/net/mana/meson.build b/drivers/net/mana/meson.build
index 364d57a619..031f443d16 100644
--- a/drivers/net/mana/meson.build
+++ b/drivers/net/mana/meson.build
@@ -11,6 +11,7 @@ deps += ['pci', 'bus_pci', 'net', 'eal', 'kvargs']
 
 sources += files(
 	'mana.c',
+	'tx.c',
 	'mr.c',
 	'gdma.c',
 	'mp.c',
diff --git a/drivers/net/mana/tx.c b/drivers/net/mana/tx.c
new file mode 100644
index 0000000000..fbeea40ef2
--- /dev/null
+++ b/drivers/net/mana/tx.c
@@ -0,0 +1,163 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2022 Microsoft Corporation
+ */
+
+#include <ethdev_driver.h>
+
+#include <infiniband/verbs.h>
+#include <infiniband/manadv.h>
+
+#include "mana.h"
+
+int mana_stop_tx_queues(struct rte_eth_dev *dev)
+{
+	struct mana_priv *priv = dev->data->dev_private;
+	int i, ret;
+
+	for (i = 0; i < priv->num_queues; i++) {
+		struct mana_txq *txq = dev->data->tx_queues[i];
+
+		if (txq->qp) {
+			ret = ibv_destroy_qp(txq->qp);
+			if (ret)
+				DRV_LOG(ERR, "tx_queue destroy_qp failed %d",
+					ret);
+			txq->qp = NULL;
+		}
+
+		if (txq->cq) {
+			ret = ibv_destroy_cq(txq->cq);
+			if (ret)
+				DRV_LOG(ERR, "tx_queue destroy_cp failed %d",
+					ret);
+			txq->cq = NULL;
+		}
+
+		/* Drain and free posted WQEs */
+		while (txq->desc_ring_tail != txq->desc_ring_head) {
+			struct mana_txq_desc *desc =
+				&txq->desc_ring[txq->desc_ring_tail];
+
+			rte_pktmbuf_free(desc->pkt);
+
+			txq->desc_ring_tail =
+				(txq->desc_ring_tail + 1) % txq->num_desc;
+		}
+		txq->desc_ring_head = 0;
+		txq->desc_ring_tail = 0;
+
+		memset(&txq->gdma_sq, 0, sizeof(txq->gdma_sq));
+		memset(&txq->gdma_cq, 0, sizeof(txq->gdma_cq));
+	}
+
+	return 0;
+}
+
+int mana_start_tx_queues(struct rte_eth_dev *dev)
+{
+	struct mana_priv *priv = dev->data->dev_private;
+	int ret, i;
+
+	/* start TX queues */
+	for (i = 0; i < priv->num_queues; i++) {
+		struct mana_txq *txq;
+		struct ibv_qp_init_attr qp_attr = { 0 };
+		struct manadv_obj obj = {};
+		struct manadv_qp dv_qp;
+		struct manadv_cq dv_cq;
+
+		txq = dev->data->tx_queues[i];
+
+		manadv_set_context_attr(priv->ib_ctx,
+			MANADV_CTX_ATTR_BUF_ALLOCATORS,
+			(void *)((uintptr_t)&(struct manadv_ctx_allocators){
+				.alloc = &mana_alloc_verbs_buf,
+				.free = &mana_free_verbs_buf,
+				.data = (void *)(uintptr_t)txq->socket,
+			}));
+
+		txq->cq = ibv_create_cq(priv->ib_ctx, txq->num_desc,
+					NULL, NULL, 0);
+		if (!txq->cq) {
+			DRV_LOG(ERR, "failed to create cq queue index %d", i);
+			ret = -errno;
+			goto fail;
+		}
+
+		qp_attr.send_cq = txq->cq;
+		qp_attr.recv_cq = txq->cq;
+		qp_attr.cap.max_send_wr = txq->num_desc;
+		qp_attr.cap.max_send_sge = priv->max_send_sge;
+
+		/* Skip setting qp_attr.cap.max_inline_data */
+
+		qp_attr.qp_type = IBV_QPT_RAW_PACKET;
+		qp_attr.sq_sig_all = 0;
+
+		txq->qp = ibv_create_qp(priv->ib_parent_pd, &qp_attr);
+		if (!txq->qp) {
+			DRV_LOG(ERR, "Failed to create qp queue index %d", i);
+			ret = -errno;
+			goto fail;
+		}
+
+		/* Get the addresses of CQ, QP and DB */
+		obj.qp.in = txq->qp;
+		obj.qp.out = &dv_qp;
+		obj.cq.in = txq->cq;
+		obj.cq.out = &dv_cq;
+		ret = manadv_init_obj(&obj, MANADV_OBJ_QP | MANADV_OBJ_CQ);
+		if (ret) {
+			DRV_LOG(ERR, "Failed to get manadv objects");
+			goto fail;
+		}
+
+		txq->gdma_sq.buffer = obj.qp.out->sq_buf;
+		txq->gdma_sq.count = obj.qp.out->sq_count;
+		txq->gdma_sq.size = obj.qp.out->sq_size;
+		txq->gdma_sq.id = obj.qp.out->sq_id;
+
+		txq->tx_vp_offset = obj.qp.out->tx_vp_offset;
+		priv->db_page = obj.qp.out->db_page;
+		DRV_LOG(INFO, "txq sq id %u vp_offset %u db_page %p "
+				" buf %p count %u size %u",
+				txq->gdma_sq.id, txq->tx_vp_offset,
+				priv->db_page,
+				txq->gdma_sq.buffer, txq->gdma_sq.count,
+				txq->gdma_sq.size);
+
+		txq->gdma_cq.buffer = obj.cq.out->buf;
+		txq->gdma_cq.count = obj.cq.out->count;
+		txq->gdma_cq.size = txq->gdma_cq.count * COMP_ENTRY_SIZE;
+		txq->gdma_cq.id = obj.cq.out->cq_id;
+
+		/* CQ head starts with count (not 0) */
+		txq->gdma_cq.head = txq->gdma_cq.count;
+
+		DRV_LOG(INFO, "txq cq id %u buf %p count %u size %u head %u",
+			txq->gdma_cq.id, txq->gdma_cq.buffer,
+			txq->gdma_cq.count, txq->gdma_cq.size,
+			txq->gdma_cq.head);
+	}
+
+	return 0;
+
+fail:
+	mana_stop_tx_queues(dev);
+	return ret;
+}
+
+static inline uint16_t get_vsq_frame_num(uint32_t vsq)
+{
+	union {
+		uint32_t gdma_txq_id;
+		struct {
+			uint32_t reserved1	: 10;
+			uint32_t vsq_frame	: 14;
+			uint32_t reserved2	: 8;
+		};
+	} v;
+
+	v.gdma_txq_id = vsq;
+	return v.vsq_frame;
+}
-- 
2.17.1

[Patch v6 13/18] net/mana: add function to start/stop RX queues

[longli]

From: Long Li <longli@microsoft.com>

MANA allocates device queues through the IB layer when starting RX queues.
When device is stopped all the queues are unmapped and freed.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v2:
Add prefix mana_ to all function names.
Remove unused header files.
v4:
Move defition "uint32_t i" from inside "for ()" to outside

 drivers/net/mana/mana.h      |   3 +
 drivers/net/mana/meson.build |   1 +
 drivers/net/mana/rx.c        | 346 +++++++++++++++++++++++++++++++++++
 3 files changed, 350 insertions(+)
 create mode 100644 drivers/net/mana/rx.c

diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index 3613ba7ca2..dc808d363f 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -364,6 +364,7 @@ extern int mana_logtype_init;
 
 int mana_ring_doorbell(void *db_page, enum gdma_queue_types queue_type,
 		       uint32_t queue_id, uint32_t tail);
+int mana_rq_ring_doorbell(struct mana_rxq *rxq);
 
 int gdma_post_work_request(struct mana_gdma_queue *queue,
 			   struct gdma_work_request *work_req,
@@ -379,8 +380,10 @@ uint16_t mana_tx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
 int gdma_poll_completion_queue(struct mana_gdma_queue *cq,
 			       struct gdma_comp *comp);
 
+int mana_start_rx_queues(struct rte_eth_dev *dev);
 int mana_start_tx_queues(struct rte_eth_dev *dev);
 
+int mana_stop_rx_queues(struct rte_eth_dev *dev);
 int mana_stop_tx_queues(struct rte_eth_dev *dev);
 
 struct mana_mr_cache *mana_find_pmd_mr(struct mana_mr_btree *local_tree,
diff --git a/drivers/net/mana/meson.build b/drivers/net/mana/meson.build
index 031f443d16..62e103a510 100644
--- a/drivers/net/mana/meson.build
+++ b/drivers/net/mana/meson.build
@@ -11,6 +11,7 @@ deps += ['pci', 'bus_pci', 'net', 'eal', 'kvargs']
 
 sources += files(
 	'mana.c',
+	'rx.c',
 	'tx.c',
 	'mr.c',
 	'gdma.c',
diff --git a/drivers/net/mana/rx.c b/drivers/net/mana/rx.c
new file mode 100644
index 0000000000..41d0fc9f11
--- /dev/null
+++ b/drivers/net/mana/rx.c
@@ -0,0 +1,346 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2022 Microsoft Corporation
+ */
+#include <ethdev_driver.h>
+
+#include <infiniband/verbs.h>
+#include <infiniband/manadv.h>
+
+#include "mana.h"
+
+static uint8_t mana_rss_hash_key_default[TOEPLITZ_HASH_KEY_SIZE_IN_BYTES] = {
+	0x2c, 0xc6, 0x81, 0xd1,
+	0x5b, 0xdb, 0xf4, 0xf7,
+	0xfc, 0xa2, 0x83, 0x19,
+	0xdb, 0x1a, 0x3e, 0x94,
+	0x6b, 0x9e, 0x38, 0xd9,
+	0x2c, 0x9c, 0x03, 0xd1,
+	0xad, 0x99, 0x44, 0xa7,
+	0xd9, 0x56, 0x3d, 0x59,
+	0x06, 0x3c, 0x25, 0xf3,
+	0xfc, 0x1f, 0xdc, 0x2a,
+};
+
+int mana_rq_ring_doorbell(struct mana_rxq *rxq)
+{
+	struct mana_priv *priv = rxq->priv;
+	int ret;
+	void *db_page = priv->db_page;
+
+	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
+		struct rte_eth_dev *dev =
+			&rte_eth_devices[priv->dev_data->port_id];
+		struct mana_process_priv *process_priv = dev->process_private;
+
+		db_page = process_priv->db_page;
+	}
+
+	ret = mana_ring_doorbell(db_page, gdma_queue_receive,
+				 rxq->gdma_rq.id,
+				 rxq->gdma_rq.head *
+					GDMA_WQE_ALIGNMENT_UNIT_SIZE);
+
+	if (ret)
+		DRV_LOG(ERR, "failed to ring RX doorbell ret %d", ret);
+
+	return ret;
+}
+
+static int mana_alloc_and_post_rx_wqe(struct mana_rxq *rxq)
+{
+	struct rte_mbuf *mbuf = NULL;
+	struct gdma_sgl_element sgl[1];
+	struct gdma_work_request request = {0};
+	struct gdma_posted_wqe_info wqe_info = {0};
+	struct mana_priv *priv = rxq->priv;
+	int ret;
+	struct mana_mr_cache *mr;
+
+	mbuf = rte_pktmbuf_alloc(rxq->mp);
+	if (!mbuf) {
+		rxq->stats.nombuf++;
+		return -ENOMEM;
+	}
+
+	mr = mana_find_pmd_mr(&rxq->mr_btree, priv, mbuf);
+	if (!mr) {
+		DRV_LOG(ERR, "failed to register RX MR");
+		rte_pktmbuf_free(mbuf);
+		return -ENOMEM;
+	}
+
+	request.gdma_header.struct_size = sizeof(request);
+	wqe_info.gdma_header.struct_size = sizeof(wqe_info);
+
+	sgl[0].address = rte_cpu_to_le_64(rte_pktmbuf_mtod(mbuf, uint64_t));
+	sgl[0].memory_key = mr->lkey;
+	sgl[0].size =
+		rte_pktmbuf_data_room_size(rxq->mp) -
+		RTE_PKTMBUF_HEADROOM;
+
+	request.sgl = sgl;
+	request.num_sgl_elements = 1;
+	request.inline_oob_data = NULL;
+	request.inline_oob_size_in_bytes = 0;
+	request.flags = 0;
+	request.client_data_unit = NOT_USING_CLIENT_DATA_UNIT;
+
+	ret = gdma_post_work_request(&rxq->gdma_rq, &request, &wqe_info);
+	if (!ret) {
+		struct mana_rxq_desc *desc =
+			&rxq->desc_ring[rxq->desc_ring_head];
+
+		/* update queue for tracking pending packets */
+		desc->pkt = mbuf;
+		desc->wqe_size_in_bu = wqe_info.wqe_size_in_bu;
+		rxq->desc_ring_head = (rxq->desc_ring_head + 1) % rxq->num_desc;
+	} else {
+		DRV_LOG(ERR, "failed to post recv ret %d", ret);
+		return ret;
+	}
+
+	return 0;
+}
+
+static int mana_alloc_and_post_rx_wqes(struct mana_rxq *rxq)
+{
+	int ret;
+	uint32_t i;
+
+	for (i = 0; i < rxq->num_desc; i++) {
+		ret = mana_alloc_and_post_rx_wqe(rxq);
+		if (ret) {
+			DRV_LOG(ERR, "failed to post RX ret = %d", ret);
+			return ret;
+		}
+	}
+
+	mana_rq_ring_doorbell(rxq);
+
+	return ret;
+}
+
+int mana_stop_rx_queues(struct rte_eth_dev *dev)
+{
+	struct mana_priv *priv = dev->data->dev_private;
+	int ret, i;
+
+	if (priv->rwq_qp) {
+		ret = ibv_destroy_qp(priv->rwq_qp);
+		if (ret)
+			DRV_LOG(ERR, "rx_queue destroy_qp failed %d", ret);
+		priv->rwq_qp = NULL;
+	}
+
+	if (priv->ind_table) {
+		ret = ibv_destroy_rwq_ind_table(priv->ind_table);
+		if (ret)
+			DRV_LOG(ERR, "destroy rwq ind table failed %d", ret);
+		priv->ind_table = NULL;
+	}
+
+	for (i = 0; i < priv->num_queues; i++) {
+		struct mana_rxq *rxq = dev->data->rx_queues[i];
+
+		if (rxq->wq) {
+			ret = ibv_destroy_wq(rxq->wq);
+			if (ret)
+				DRV_LOG(ERR,
+					"rx_queue destroy_wq failed %d", ret);
+			rxq->wq = NULL;
+		}
+
+		if (rxq->cq) {
+			ret = ibv_destroy_cq(rxq->cq);
+			if (ret)
+				DRV_LOG(ERR,
+					"rx_queue destroy_cq failed %d", ret);
+			rxq->cq = NULL;
+		}
+
+		/* Drain and free posted WQEs */
+		while (rxq->desc_ring_tail != rxq->desc_ring_head) {
+			struct mana_rxq_desc *desc =
+				&rxq->desc_ring[rxq->desc_ring_tail];
+
+			rte_pktmbuf_free(desc->pkt);
+
+			rxq->desc_ring_tail =
+				(rxq->desc_ring_tail + 1) % rxq->num_desc;
+		}
+		rxq->desc_ring_head = 0;
+		rxq->desc_ring_tail = 0;
+
+		memset(&rxq->gdma_rq, 0, sizeof(rxq->gdma_rq));
+		memset(&rxq->gdma_cq, 0, sizeof(rxq->gdma_cq));
+	}
+	return 0;
+}
+
+int mana_start_rx_queues(struct rte_eth_dev *dev)
+{
+	struct mana_priv *priv = dev->data->dev_private;
+	int ret, i;
+	struct ibv_wq *ind_tbl[priv->num_queues];
+
+	DRV_LOG(INFO, "start rx queues");
+	for (i = 0; i < priv->num_queues; i++) {
+		struct mana_rxq *rxq = dev->data->rx_queues[i];
+		struct ibv_wq_init_attr wq_attr = {};
+
+		manadv_set_context_attr(priv->ib_ctx,
+			MANADV_CTX_ATTR_BUF_ALLOCATORS,
+			(void *)((uintptr_t)&(struct manadv_ctx_allocators){
+				.alloc = &mana_alloc_verbs_buf,
+				.free = &mana_free_verbs_buf,
+				.data = (void *)(uintptr_t)rxq->socket,
+			}));
+
+		rxq->cq = ibv_create_cq(priv->ib_ctx, rxq->num_desc,
+					NULL, NULL, 0);
+		if (!rxq->cq) {
+			ret = -errno;
+			DRV_LOG(ERR, "failed to create rx cq queue %d", i);
+			goto fail;
+		}
+
+		wq_attr.wq_type = IBV_WQT_RQ;
+		wq_attr.max_wr = rxq->num_desc;
+		wq_attr.max_sge = 1;
+		wq_attr.pd = priv->ib_parent_pd;
+		wq_attr.cq = rxq->cq;
+
+		rxq->wq = ibv_create_wq(priv->ib_ctx, &wq_attr);
+		if (!rxq->wq) {
+			ret = -errno;
+			DRV_LOG(ERR, "failed to create rx wq %d", i);
+			goto fail;
+		}
+
+		ind_tbl[i] = rxq->wq;
+	}
+
+	struct ibv_rwq_ind_table_init_attr ind_table_attr = {
+		.log_ind_tbl_size = rte_log2_u32(RTE_DIM(ind_tbl)),
+		.ind_tbl = ind_tbl,
+		.comp_mask = 0,
+	};
+
+	priv->ind_table = ibv_create_rwq_ind_table(priv->ib_ctx,
+						   &ind_table_attr);
+	if (!priv->ind_table) {
+		ret = -errno;
+		DRV_LOG(ERR, "failed to create ind_table ret %d", ret);
+		goto fail;
+	}
+
+	DRV_LOG(INFO, "ind_table handle %d num %d",
+		priv->ind_table->ind_tbl_handle,
+		priv->ind_table->ind_tbl_num);
+
+	struct ibv_qp_init_attr_ex qp_attr_ex = {
+		.comp_mask = IBV_QP_INIT_ATTR_PD |
+			     IBV_QP_INIT_ATTR_RX_HASH |
+			     IBV_QP_INIT_ATTR_IND_TABLE,
+		.qp_type = IBV_QPT_RAW_PACKET,
+		.pd = priv->ib_parent_pd,
+		.rwq_ind_tbl = priv->ind_table,
+		.rx_hash_conf = {
+			.rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
+			.rx_hash_key_len = TOEPLITZ_HASH_KEY_SIZE_IN_BYTES,
+			.rx_hash_key = mana_rss_hash_key_default,
+			.rx_hash_fields_mask =
+				IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4,
+		},
+
+	};
+
+	/* overwrite default if rss key is set */
+	if (priv->rss_conf.rss_key_len && priv->rss_conf.rss_key)
+		qp_attr_ex.rx_hash_conf.rx_hash_key =
+			priv->rss_conf.rss_key;
+
+	/* overwrite default if rss hash fields are set */
+	if (priv->rss_conf.rss_hf) {
+		qp_attr_ex.rx_hash_conf.rx_hash_fields_mask = 0;
+
+		if (priv->rss_conf.rss_hf & ETH_RSS_IPV4)
+			qp_attr_ex.rx_hash_conf.rx_hash_fields_mask |=
+				IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4;
+
+		if (priv->rss_conf.rss_hf & ETH_RSS_IPV6)
+			qp_attr_ex.rx_hash_conf.rx_hash_fields_mask |=
+				IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_SRC_IPV6;
+
+		if (priv->rss_conf.rss_hf &
+		    (ETH_RSS_NONFRAG_IPV4_TCP | ETH_RSS_NONFRAG_IPV6_TCP))
+			qp_attr_ex.rx_hash_conf.rx_hash_fields_mask |=
+				IBV_RX_HASH_SRC_PORT_TCP |
+				IBV_RX_HASH_DST_PORT_TCP;
+
+		if (priv->rss_conf.rss_hf &
+		    (ETH_RSS_NONFRAG_IPV4_UDP | ETH_RSS_NONFRAG_IPV6_UDP))
+			qp_attr_ex.rx_hash_conf.rx_hash_fields_mask |=
+				IBV_RX_HASH_SRC_PORT_UDP |
+				IBV_RX_HASH_DST_PORT_UDP;
+	}
+
+	priv->rwq_qp = ibv_create_qp_ex(priv->ib_ctx, &qp_attr_ex);
+	if (!priv->rwq_qp) {
+		ret = -errno;
+		DRV_LOG(ERR, "rx ibv_create_qp_ex failed");
+		goto fail;
+	}
+
+	for (i = 0; i < priv->num_queues; i++) {
+		struct mana_rxq *rxq = dev->data->rx_queues[i];
+		struct manadv_obj obj = {};
+		struct manadv_cq dv_cq;
+		struct manadv_rwq dv_wq;
+
+		obj.cq.in = rxq->cq;
+		obj.cq.out = &dv_cq;
+		obj.rwq.in = rxq->wq;
+		obj.rwq.out = &dv_wq;
+		ret = manadv_init_obj(&obj, MANADV_OBJ_CQ | MANADV_OBJ_RWQ);
+		if (ret) {
+			DRV_LOG(ERR, "manadv_init_obj failed ret %d", ret);
+			goto fail;
+		}
+
+		rxq->gdma_cq.buffer = obj.cq.out->buf;
+		rxq->gdma_cq.count = obj.cq.out->count;
+		rxq->gdma_cq.size = rxq->gdma_cq.count * COMP_ENTRY_SIZE;
+		rxq->gdma_cq.id = obj.cq.out->cq_id;
+
+		/* CQ head starts with count */
+		rxq->gdma_cq.head = rxq->gdma_cq.count;
+
+		DRV_LOG(INFO, "rxq cq id %u buf %p count %u size %u",
+			rxq->gdma_cq.id, rxq->gdma_cq.buffer,
+			rxq->gdma_cq.count, rxq->gdma_cq.size);
+
+		priv->db_page = obj.rwq.out->db_page;
+
+		rxq->gdma_rq.buffer = obj.rwq.out->buf;
+		rxq->gdma_rq.count = obj.rwq.out->count;
+		rxq->gdma_rq.size = obj.rwq.out->size;
+		rxq->gdma_rq.id = obj.rwq.out->wq_id;
+
+		DRV_LOG(INFO, "rxq rq id %u buf %p count %u size %u",
+			rxq->gdma_rq.id, rxq->gdma_rq.buffer,
+			rxq->gdma_rq.count, rxq->gdma_rq.size);
+	}
+
+	for (i = 0; i < priv->num_queues; i++) {
+		ret = mana_alloc_and_post_rx_wqes(dev->data->rx_queues[i]);
+		if (ret)
+			goto fail;
+	}
+
+	return 0;
+
+fail:
+	mana_stop_rx_queues(dev);
+	return ret;
+}
-- 
2.17.1

[Patch v6 14/18] net/mana: add function to receive packets

[longli]

From: Long Li <longli@microsoft.com>

With all the RX queues created, MANA can use those queues to receive
packets.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v2:
Add mana_ to all function names.
Rename a camel case.

 doc/guides/nics/features/mana.ini |   2 +
 drivers/net/mana/mana.c           |   2 +
 drivers/net/mana/mana.h           |  37 +++++++++++
 drivers/net/mana/mp.c             |   2 +
 drivers/net/mana/rx.c             | 104 ++++++++++++++++++++++++++++++
 5 files changed, 147 insertions(+)

diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
index 821443b292..fdbf22d335 100644
--- a/doc/guides/nics/features/mana.ini
+++ b/doc/guides/nics/features/mana.ini
@@ -6,6 +6,8 @@
 [Features]
 Link status          = P
 Linux                = Y
+L3 checksum offload  = Y
+L4 checksum offload  = Y
 Multiprocess aware   = Y
 Queue start/stop     = Y
 Removal event        = Y
diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index d18cc4ab0e..c349822991 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -950,6 +950,8 @@ static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv __rte_unused,
 				/* fd is no not used after mapping doorbell */
 				close(fd);
 
+				eth_dev->rx_pkt_burst = mana_rx_burst;
+
 				rte_spinlock_lock(&mana_shared_data->lock);
 				mana_shared_data->secondary_cnt++;
 				mana_local_data.secondary_cnt++;
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index dc808d363f..bafc4d6082 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -178,6 +178,11 @@ struct gdma_work_request {
 
 enum mana_cqe_type {
 	CQE_INVALID                     = 0,
+
+	CQE_RX_OKAY                     = 1,
+	CQE_RX_COALESCED_4              = 2,
+	CQE_RX_OBJECT_FENCE             = 3,
+	CQE_RX_TRUNCATED                = 4,
 };
 
 struct mana_cqe_header {
@@ -203,6 +208,35 @@ struct mana_cqe_header {
 	(NDIS_HASH_TCP_IPV4 | NDIS_HASH_UDP_IPV4 | NDIS_HASH_TCP_IPV6 |      \
 	 NDIS_HASH_UDP_IPV6 | NDIS_HASH_TCP_IPV6_EX | NDIS_HASH_UDP_IPV6_EX)
 
+struct mana_rx_comp_per_packet_info {
+	uint32_t packet_length	: 16;
+	uint32_t reserved0	: 16;
+	uint32_t reserved1;
+	uint32_t packet_hash;
+}; /* HW DATA */
+#define RX_COM_OOB_NUM_PACKETINFO_SEGMENTS 4
+
+struct mana_rx_comp_oob {
+	struct mana_cqe_header cqe_hdr;
+
+	uint32_t rx_vlan_id				: 12;
+	uint32_t rx_vlan_tag_present			: 1;
+	uint32_t rx_outer_ip_header_checksum_succeeded	: 1;
+	uint32_t rx_outer_ip_header_checksum_failed	: 1;
+	uint32_t reserved				: 1;
+	uint32_t rx_hash_type				: 9;
+	uint32_t rx_ip_header_checksum_succeeded	: 1;
+	uint32_t rx_ip_header_checksum_failed		: 1;
+	uint32_t rx_tcp_checksum_succeeded		: 1;
+	uint32_t rx_tcp_checksum_failed			: 1;
+	uint32_t rx_udp_checksum_succeeded		: 1;
+	uint32_t rx_udp_checksum_failed			: 1;
+	uint32_t reserved1				: 1;
+	struct mana_rx_comp_per_packet_info
+		packet_info[RX_COM_OOB_NUM_PACKETINFO_SEGMENTS];
+	uint32_t received_wqe_offset;
+}; /* HW DATA */
+
 struct gdma_wqe_dma_oob {
 	uint32_t reserved:24;
 	uint32_t last_v_bytes:8;
@@ -371,6 +405,9 @@ int gdma_post_work_request(struct mana_gdma_queue *queue,
 			   struct gdma_posted_wqe_info *wqe_info);
 uint8_t *gdma_get_wqe_pointer(struct mana_gdma_queue *queue);
 
+uint16_t mana_rx_burst(void *dpdk_rxq, struct rte_mbuf **rx_pkts,
+		       uint16_t pkts_n);
+
 uint16_t mana_rx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
 			       uint16_t pkts_n);
 
diff --git a/drivers/net/mana/mp.c b/drivers/net/mana/mp.c
index f4f78d2787..36a88c561a 100644
--- a/drivers/net/mana/mp.c
+++ b/drivers/net/mana/mp.c
@@ -138,6 +138,8 @@ static int mana_mp_secondary_handle(const struct rte_mp_msg *mp_msg,
 	case MANA_MP_REQ_START_RXTX:
 		DRV_LOG(INFO, "Port %u starting datapath", dev->data->port_id);
 
+		dev->rx_pkt_burst = mana_rx_burst;
+
 		rte_mb();
 
 		res->result = 0;
diff --git a/drivers/net/mana/rx.c b/drivers/net/mana/rx.c
index 41d0fc9f11..f2573a6d06 100644
--- a/drivers/net/mana/rx.c
+++ b/drivers/net/mana/rx.c
@@ -344,3 +344,107 @@ int mana_start_rx_queues(struct rte_eth_dev *dev)
 	mana_stop_rx_queues(dev);
 	return ret;
 }
+
+uint16_t mana_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+	uint16_t pkt_received = 0, cqe_processed = 0;
+	struct mana_rxq *rxq = dpdk_rxq;
+	struct mana_priv *priv = rxq->priv;
+	struct gdma_comp comp;
+	struct rte_mbuf *mbuf;
+	int ret;
+
+	while (pkt_received < pkts_n &&
+	       gdma_poll_completion_queue(&rxq->gdma_cq, &comp) == 1) {
+		struct mana_rxq_desc *desc;
+		struct mana_rx_comp_oob *oob =
+			(struct mana_rx_comp_oob *)&comp.completion_data[0];
+
+		if (comp.work_queue_number != rxq->gdma_rq.id) {
+			DRV_LOG(ERR, "rxq comp id mismatch wqid=0x%x rcid=0x%x",
+				comp.work_queue_number, rxq->gdma_rq.id);
+			rxq->stats.errors++;
+			break;
+		}
+
+		desc = &rxq->desc_ring[rxq->desc_ring_tail];
+		rxq->gdma_rq.tail += desc->wqe_size_in_bu;
+		mbuf = desc->pkt;
+
+		switch (oob->cqe_hdr.cqe_type) {
+		case CQE_RX_OKAY:
+			/* Proceed to process mbuf */
+			break;
+
+		case CQE_RX_TRUNCATED:
+			DRV_LOG(ERR, "Drop a truncated packet");
+			rxq->stats.errors++;
+			rte_pktmbuf_free(mbuf);
+			goto drop;
+
+		case CQE_RX_COALESCED_4:
+			DRV_LOG(ERR, "RX coalescing is not supported");
+			continue;
+
+		default:
+			DRV_LOG(ERR, "Unknown RX CQE type %d",
+				oob->cqe_hdr.cqe_type);
+			continue;
+		}
+
+		DRV_LOG(DEBUG, "mana_rx_comp_oob CQE_RX_OKAY rxq %p", rxq);
+
+		mbuf->data_off = RTE_PKTMBUF_HEADROOM;
+		mbuf->nb_segs = 1;
+		mbuf->next = NULL;
+		mbuf->pkt_len = oob->packet_info[0].packet_length;
+		mbuf->data_len = oob->packet_info[0].packet_length;
+		mbuf->port = priv->port_id;
+
+		if (oob->rx_ip_header_checksum_succeeded)
+			mbuf->ol_flags |= RTE_MBUF_F_RX_IP_CKSUM_GOOD;
+
+		if (oob->rx_ip_header_checksum_failed)
+			mbuf->ol_flags |= RTE_MBUF_F_RX_IP_CKSUM_BAD;
+
+		if (oob->rx_outer_ip_header_checksum_failed)
+			mbuf->ol_flags |= RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD;
+
+		if (oob->rx_tcp_checksum_succeeded ||
+		    oob->rx_udp_checksum_succeeded)
+			mbuf->ol_flags |= RTE_MBUF_F_RX_L4_CKSUM_GOOD;
+
+		if (oob->rx_tcp_checksum_failed ||
+		    oob->rx_udp_checksum_failed)
+			mbuf->ol_flags |= RTE_MBUF_F_RX_L4_CKSUM_BAD;
+
+		if (oob->rx_hash_type == MANA_HASH_L3 ||
+		    oob->rx_hash_type == MANA_HASH_L4) {
+			mbuf->ol_flags |= RTE_MBUF_F_RX_RSS_HASH;
+			mbuf->hash.rss = oob->packet_info[0].packet_hash;
+		}
+
+		pkts[pkt_received++] = mbuf;
+		rxq->stats.packets++;
+		rxq->stats.bytes += mbuf->data_len;
+
+drop:
+		rxq->desc_ring_tail++;
+		if (rxq->desc_ring_tail >= rxq->num_desc)
+			rxq->desc_ring_tail = 0;
+
+		cqe_processed++;
+
+		/* Post another request */
+		ret = mana_alloc_and_post_rx_wqe(rxq);
+		if (ret) {
+			DRV_LOG(ERR, "failed to post rx wqe ret=%d", ret);
+			break;
+		}
+	}
+
+	if (cqe_processed)
+		mana_rq_ring_doorbell(rxq);
+
+	return pkt_received;
+}
-- 
2.17.1

[Patch v6 15/18] net/mana: add function to send packets

[longli]

From: Long Li <longli@microsoft.com>

With all the TX queues created, MANA can send packets over those queues.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v2:
Rename all camel cases.

 doc/guides/nics/features/mana.ini |   1 +
 drivers/net/mana/mana.c           |   1 +
 drivers/net/mana/mana.h           |  65 ++++++++
 drivers/net/mana/mp.c             |   1 +
 drivers/net/mana/tx.c             | 241 ++++++++++++++++++++++++++++++
 5 files changed, 309 insertions(+)

diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
index fdbf22d335..7922816d66 100644
--- a/doc/guides/nics/features/mana.ini
+++ b/doc/guides/nics/features/mana.ini
@@ -4,6 +4,7 @@
 ; Refer to default.ini for the full list of available PMD features.
 ;
 [Features]
+Free Tx mbuf on demand = Y
 Link status          = P
 Linux                = Y
 L3 checksum offload  = Y
diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index c349822991..0dcd3f3124 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -950,6 +950,7 @@ static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv __rte_unused,
 				/* fd is no not used after mapping doorbell */
 				close(fd);
 
+				eth_dev->tx_pkt_burst = mana_tx_burst;
 				eth_dev->rx_pkt_burst = mana_rx_burst;
 
 				rte_spinlock_lock(&mana_shared_data->lock);
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index bafc4d6082..b4056bd50b 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -62,6 +62,47 @@ struct mana_shared_data {
 
 #define NOT_USING_CLIENT_DATA_UNIT 0
 
+enum tx_packet_format_v2 {
+	short_packet_format = 0,
+	long_packet_format = 1
+};
+
+struct transmit_short_oob_v2 {
+	enum tx_packet_format_v2 packet_format : 2;
+	uint32_t tx_is_outer_ipv4 : 1;
+	uint32_t tx_is_outer_ipv6 : 1;
+	uint32_t tx_compute_IP_header_checksum : 1;
+	uint32_t tx_compute_TCP_checksum : 1;
+	uint32_t tx_compute_UDP_checksum : 1;
+	uint32_t suppress_tx_CQE_generation : 1;
+	uint32_t VCQ_number : 24;
+	uint32_t tx_transport_header_offset : 10;
+	uint32_t VSQ_frame_num : 14;
+	uint32_t short_vport_offset : 8;
+};
+
+struct transmit_long_oob_v2 {
+	uint32_t tx_is_encapsulated_packet : 1;
+	uint32_t tx_inner_is_ipv6 : 1;
+	uint32_t tx_inner_TCP_options_present : 1;
+	uint32_t inject_vlan_prior_tag : 1;
+	uint32_t reserved1 : 12;
+	uint32_t priority_code_point : 3;
+	uint32_t drop_eligible_indicator : 1;
+	uint32_t vlan_identifier : 12;
+	uint32_t tx_inner_frame_offset : 10;
+	uint32_t tx_inner_IP_header_relative_offset : 6;
+	uint32_t long_vport_offset : 12;
+	uint32_t reserved3 : 4;
+	uint32_t reserved4 : 32;
+	uint32_t reserved5 : 32;
+};
+
+struct transmit_oob_v2 {
+	struct transmit_short_oob_v2 short_oob;
+	struct transmit_long_oob_v2 long_oob;
+};
+
 enum gdma_queue_types {
 	gdma_queue_type_invalid = 0,
 	gdma_queue_send,
@@ -183,6 +224,17 @@ enum mana_cqe_type {
 	CQE_RX_COALESCED_4              = 2,
 	CQE_RX_OBJECT_FENCE             = 3,
 	CQE_RX_TRUNCATED                = 4,
+
+	CQE_TX_OKAY                     = 32,
+	CQE_TX_SA_DROP                  = 33,
+	CQE_TX_MTU_DROP                 = 34,
+	CQE_TX_INVALID_OOB              = 35,
+	CQE_TX_INVALID_ETH_TYPE         = 36,
+	CQE_TX_HDR_PROCESSING_ERROR     = 37,
+	CQE_TX_VF_DISABLED              = 38,
+	CQE_TX_VPORT_IDX_OUT_OF_RANGE   = 39,
+	CQE_TX_VPORT_DISABLED           = 40,
+	CQE_TX_VLAN_TAGGING_VIOLATION   = 41,
 };
 
 struct mana_cqe_header {
@@ -191,6 +243,17 @@ struct mana_cqe_header {
 	uint32_t vendor_err  : 24;
 }; /* HW DATA */
 
+struct mana_tx_comp_oob {
+	struct mana_cqe_header cqe_hdr;
+
+	uint32_t tx_data_offset;
+
+	uint32_t tx_sgl_offset       : 5;
+	uint32_t tx_wqe_offset       : 27;
+
+	uint32_t reserved[12];
+}; /* HW DATA */
+
 /* NDIS HASH Types */
 #define BIT(nr)		(1 << (nr))
 #define NDIS_HASH_IPV4          BIT(0)
@@ -407,6 +470,8 @@ uint8_t *gdma_get_wqe_pointer(struct mana_gdma_queue *queue);
 
 uint16_t mana_rx_burst(void *dpdk_rxq, struct rte_mbuf **rx_pkts,
 		       uint16_t pkts_n);
+uint16_t mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts,
+		       uint16_t pkts_n);
 
 uint16_t mana_rx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts,
 			       uint16_t pkts_n);
diff --git a/drivers/net/mana/mp.c b/drivers/net/mana/mp.c
index 36a88c561a..da9c0f36a1 100644
--- a/drivers/net/mana/mp.c
+++ b/drivers/net/mana/mp.c
@@ -138,6 +138,7 @@ static int mana_mp_secondary_handle(const struct rte_mp_msg *mp_msg,
 	case MANA_MP_REQ_START_RXTX:
 		DRV_LOG(INFO, "Port %u starting datapath", dev->data->port_id);
 
+		dev->tx_pkt_burst = mana_tx_burst;
 		dev->rx_pkt_burst = mana_rx_burst;
 
 		rte_mb();
diff --git a/drivers/net/mana/tx.c b/drivers/net/mana/tx.c
index fbeea40ef2..0aefab1b6c 100644
--- a/drivers/net/mana/tx.c
+++ b/drivers/net/mana/tx.c
@@ -161,3 +161,244 @@ static inline uint16_t get_vsq_frame_num(uint32_t vsq)
 	v.gdma_txq_id = vsq;
 	return v.vsq_frame;
 }
+
+uint16_t mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts,
+		       uint16_t nb_pkts)
+{
+	struct mana_txq *txq = dpdk_txq;
+	struct mana_priv *priv = txq->priv;
+	struct gdma_comp comp;
+	int ret;
+	void *db_page;
+
+	/* Process send completions from GDMA */
+	while (gdma_poll_completion_queue(&txq->gdma_cq, &comp) == 1) {
+		struct mana_txq_desc *desc =
+			&txq->desc_ring[txq->desc_ring_tail];
+		struct mana_tx_comp_oob *oob =
+			(struct mana_tx_comp_oob *)&comp.completion_data[0];
+
+		if (oob->cqe_hdr.cqe_type != CQE_TX_OKAY) {
+			DRV_LOG(ERR,
+				"mana_tx_comp_oob cqe_type %u vendor_err %u",
+				oob->cqe_hdr.cqe_type, oob->cqe_hdr.vendor_err);
+			txq->stats.errors++;
+		} else {
+			DRV_LOG(DEBUG, "mana_tx_comp_oob CQE_TX_OKAY");
+			txq->stats.packets++;
+		}
+
+		if (!desc->pkt) {
+			DRV_LOG(ERR, "mana_txq_desc has a NULL pkt");
+		} else {
+			txq->stats.bytes += desc->pkt->data_len;
+			rte_pktmbuf_free(desc->pkt);
+		}
+
+		desc->pkt = NULL;
+		txq->desc_ring_tail = (txq->desc_ring_tail + 1) % txq->num_desc;
+		txq->gdma_sq.tail += desc->wqe_size_in_bu;
+	}
+
+	/* Post send requests to GDMA */
+	uint16_t pkt_idx;
+
+	for (pkt_idx = 0; pkt_idx < nb_pkts; pkt_idx++) {
+		struct rte_mbuf *m_pkt = tx_pkts[pkt_idx];
+		struct rte_mbuf *m_seg = m_pkt;
+		struct transmit_oob_v2 tx_oob = {0};
+		struct one_sgl sgl = {0};
+
+		/* Drop the packet if it exceeds max segments */
+		if (m_pkt->nb_segs > priv->max_send_sge) {
+			DRV_LOG(ERR, "send packet segments %d exceeding max",
+				m_pkt->nb_segs);
+			continue;
+		}
+
+		/* Fill in the oob */
+		tx_oob.short_oob.packet_format = short_packet_format;
+		tx_oob.short_oob.tx_is_outer_ipv4 =
+			m_pkt->ol_flags & RTE_MBUF_F_TX_IPV4 ? 1 : 0;
+		tx_oob.short_oob.tx_is_outer_ipv6 =
+			m_pkt->ol_flags & RTE_MBUF_F_TX_IPV6 ? 1 : 0;
+
+		tx_oob.short_oob.tx_compute_IP_header_checksum =
+			m_pkt->ol_flags & RTE_MBUF_F_TX_IP_CKSUM ? 1 : 0;
+
+		if ((m_pkt->ol_flags & RTE_MBUF_F_TX_L4_MASK) ==
+				RTE_MBUF_F_TX_TCP_CKSUM) {
+			struct rte_tcp_hdr *tcp_hdr;
+
+			/* HW needs partial TCP checksum */
+
+			tcp_hdr = rte_pktmbuf_mtod_offset(m_pkt,
+					  struct rte_tcp_hdr *,
+					  m_pkt->l2_len + m_pkt->l3_len);
+
+			if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV4) {
+				struct rte_ipv4_hdr *ip_hdr;
+
+				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
+						struct rte_ipv4_hdr *,
+						m_pkt->l2_len);
+				tcp_hdr->cksum = rte_ipv4_phdr_cksum(ip_hdr,
+							m_pkt->ol_flags);
+
+			} else if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV6) {
+				struct rte_ipv6_hdr *ip_hdr;
+
+				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
+						struct rte_ipv6_hdr *,
+						m_pkt->l2_len);
+				tcp_hdr->cksum = rte_ipv6_phdr_cksum(ip_hdr,
+							m_pkt->ol_flags);
+			} else {
+				DRV_LOG(ERR, "Invalid input for TCP CKSUM");
+			}
+
+			tx_oob.short_oob.tx_compute_TCP_checksum = 1;
+			tx_oob.short_oob.tx_transport_header_offset =
+				m_pkt->l2_len + m_pkt->l3_len;
+		}
+
+		if ((m_pkt->ol_flags & RTE_MBUF_F_TX_L4_MASK) ==
+				RTE_MBUF_F_TX_UDP_CKSUM) {
+			struct rte_udp_hdr *udp_hdr;
+
+			/* HW needs partial UDP checksum */
+			udp_hdr = rte_pktmbuf_mtod_offset(m_pkt,
+					struct rte_udp_hdr *,
+					m_pkt->l2_len + m_pkt->l3_len);
+
+			if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV4) {
+				struct rte_ipv4_hdr *ip_hdr;
+
+				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
+						struct rte_ipv4_hdr *,
+						m_pkt->l2_len);
+
+				udp_hdr->dgram_cksum =
+					rte_ipv4_phdr_cksum(ip_hdr,
+							    m_pkt->ol_flags);
+
+			} else if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV6) {
+				struct rte_ipv6_hdr *ip_hdr;
+
+				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
+						struct rte_ipv6_hdr *,
+						m_pkt->l2_len);
+
+				udp_hdr->dgram_cksum =
+					rte_ipv6_phdr_cksum(ip_hdr,
+							    m_pkt->ol_flags);
+
+			} else {
+				DRV_LOG(ERR, "Invalid input for UDP CKSUM");
+			}
+
+			tx_oob.short_oob.tx_compute_UDP_checksum = 1;
+		}
+
+		tx_oob.short_oob.suppress_tx_CQE_generation = 0;
+		tx_oob.short_oob.VCQ_number = txq->gdma_cq.id;
+
+		tx_oob.short_oob.VSQ_frame_num =
+			get_vsq_frame_num(txq->gdma_sq.id);
+		tx_oob.short_oob.short_vport_offset = txq->tx_vp_offset;
+
+		DRV_LOG(DEBUG, "tx_oob packet_format %u ipv4 %u ipv6 %u",
+			tx_oob.short_oob.packet_format,
+			tx_oob.short_oob.tx_is_outer_ipv4,
+			tx_oob.short_oob.tx_is_outer_ipv6);
+
+		DRV_LOG(DEBUG, "tx_oob checksum ip %u tcp %u udp %u offset %u",
+			tx_oob.short_oob.tx_compute_IP_header_checksum,
+			tx_oob.short_oob.tx_compute_TCP_checksum,
+			tx_oob.short_oob.tx_compute_UDP_checksum,
+			tx_oob.short_oob.tx_transport_header_offset);
+
+		DRV_LOG(DEBUG, "pkt[%d]: buf_addr 0x%p, nb_segs %d, pkt_len %d",
+			pkt_idx, m_pkt->buf_addr, m_pkt->nb_segs,
+			m_pkt->pkt_len);
+
+		/* Create SGL for packet data buffers */
+		for (uint16_t seg_idx = 0; seg_idx < m_pkt->nb_segs; seg_idx++) {
+			struct mana_mr_cache *mr =
+				mana_find_pmd_mr(&txq->mr_btree, priv, m_seg);
+
+			if (!mr) {
+				DRV_LOG(ERR, "failed to get MR, pkt_idx %u",
+					pkt_idx);
+				return pkt_idx;
+			}
+
+			sgl.gdma_sgl[seg_idx].address =
+				rte_cpu_to_le_64(rte_pktmbuf_mtod(m_seg,
+								  uint64_t));
+			sgl.gdma_sgl[seg_idx].size = m_seg->data_len;
+			sgl.gdma_sgl[seg_idx].memory_key = mr->lkey;
+
+			DRV_LOG(DEBUG,
+				"seg idx %u addr 0x%" PRIx64 " size %x key %x",
+				seg_idx, sgl.gdma_sgl[seg_idx].address,
+				sgl.gdma_sgl[seg_idx].size,
+				sgl.gdma_sgl[seg_idx].memory_key);
+
+			m_seg = m_seg->next;
+		}
+
+		struct gdma_work_request work_req = {0};
+		struct gdma_posted_wqe_info wqe_info = {0};
+
+		work_req.gdma_header.struct_size = sizeof(work_req);
+		wqe_info.gdma_header.struct_size = sizeof(wqe_info);
+
+		work_req.sgl = sgl.gdma_sgl;
+		work_req.num_sgl_elements = m_pkt->nb_segs;
+		work_req.inline_oob_size_in_bytes =
+			sizeof(struct transmit_short_oob_v2);
+		work_req.inline_oob_data = &tx_oob;
+		work_req.flags = 0;
+		work_req.client_data_unit = NOT_USING_CLIENT_DATA_UNIT;
+
+		ret = gdma_post_work_request(&txq->gdma_sq, &work_req,
+					     &wqe_info);
+		if (!ret) {
+			struct mana_txq_desc *desc =
+				&txq->desc_ring[txq->desc_ring_head];
+
+			/* Update queue for tracking pending requests */
+			desc->pkt = m_pkt;
+			desc->wqe_size_in_bu = wqe_info.wqe_size_in_bu;
+			txq->desc_ring_head =
+				(txq->desc_ring_head + 1) % txq->num_desc;
+
+			DRV_LOG(DEBUG, "nb_pkts %u pkt[%d] sent",
+				nb_pkts, pkt_idx);
+		} else {
+			DRV_LOG(INFO, "pkt[%d] failed to post send ret %d",
+				pkt_idx, ret);
+			break;
+		}
+	}
+
+	/* Ring hardware door bell */
+	db_page = priv->db_page;
+	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
+		struct rte_eth_dev *dev =
+			&rte_eth_devices[priv->dev_data->port_id];
+		struct mana_process_priv *process_priv = dev->process_private;
+
+		db_page = process_priv->db_page;
+	}
+
+	ret = mana_ring_doorbell(db_page, gdma_queue_send,
+				 txq->gdma_sq.id,
+				 txq->gdma_sq.head *
+					GDMA_WQE_ALIGNMENT_UNIT_SIZE);
+	if (ret)
+		DRV_LOG(ERR, "mana_ring_doorbell failed ret %d", ret);
+
+	return pkt_idx;
+}
-- 
2.17.1

[Patch v6 16/18] net/mana: add function to start/stop device

[longli]

From: Long Li <longli@microsoft.com>

Add support for starting/stopping the device.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v2:
Use spinlock for memory registration cache.
Add prefix mana_ to all function names.
v6:
Roll back device state on error in mana_dev_start()

 drivers/net/mana/mana.c | 77 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 77 insertions(+)

diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index 0dcd3f3124..eb37f359db 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -97,6 +97,81 @@ static int mana_dev_configure(struct rte_eth_dev *dev)
 
 static int mana_intr_uninstall(struct mana_priv *priv);
 
+static int
+mana_dev_start(struct rte_eth_dev *dev)
+{
+	int ret;
+	struct mana_priv *priv = dev->data->dev_private;
+
+	rte_spinlock_init(&priv->mr_btree_lock);
+	ret = mana_mr_btree_init(&priv->mr_btree, MANA_MR_BTREE_CACHE_N,
+				 dev->device->numa_node);
+	if (ret) {
+		DRV_LOG(ERR, "Failed to init device MR btree %d", ret);
+		return ret;
+	}
+
+	ret = mana_start_tx_queues(dev);
+	if (ret) {
+		DRV_LOG(ERR, "failed to start tx queues %d", ret);
+		goto failed_tx;
+	}
+
+	ret = mana_start_rx_queues(dev);
+	if (ret) {
+		DRV_LOG(ERR, "failed to start rx queues %d", ret);
+		goto failed_rx;
+	}
+
+	rte_wmb();
+
+	dev->tx_pkt_burst = mana_tx_burst;
+	dev->rx_pkt_burst = mana_rx_burst;
+
+	DRV_LOG(INFO, "TX/RX queues have started");
+
+	/* Enable datapath for secondary processes */
+	mana_mp_req_on_rxtx(dev, MANA_MP_REQ_START_RXTX);
+
+	return 0;
+
+failed_rx:
+	mana_stop_tx_queues(dev);
+
+failed_tx:
+	mana_mr_btree_free(&priv->mr_btree);
+
+	return ret;
+}
+
+static int
+mana_dev_stop(struct rte_eth_dev *dev __rte_unused)
+{
+	int ret;
+
+	dev->tx_pkt_burst = mana_tx_burst_removed;
+	dev->rx_pkt_burst = mana_rx_burst_removed;
+
+	/* Stop datapath on secondary processes */
+	mana_mp_req_on_rxtx(dev, MANA_MP_REQ_STOP_RXTX);
+
+	rte_wmb();
+
+	ret = mana_stop_tx_queues(dev);
+	if (ret) {
+		DRV_LOG(ERR, "failed to stop tx queues");
+		return ret;
+	}
+
+	ret = mana_stop_rx_queues(dev);
+	if (ret) {
+		DRV_LOG(ERR, "failed to stop tx queues");
+		return ret;
+	}
+
+	return 0;
+}
+
 static int
 mana_dev_close(struct rte_eth_dev *dev)
 {
@@ -435,6 +510,8 @@ static int mana_dev_link_update(struct rte_eth_dev *dev,
 
 const struct eth_dev_ops mana_dev_ops = {
 	.dev_configure		= mana_dev_configure,
+	.dev_start		= mana_dev_start,
+	.dev_stop		= mana_dev_stop,
 	.dev_close		= mana_dev_close,
 	.dev_infos_get		= mana_dev_info_get,
 	.txq_info_get		= mana_dev_tx_queue_info,
-- 
2.17.1

[Patch v6 17/18] net/mana: add function to report queue stats

[longli]

From: Long Li <longli@microsoft.com>

Report packet statistics.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v5:
Fixed calculation of stats packets/bytes/errors by adding them over the queue stats.

 doc/guides/nics/features/mana.ini |  2 +
 drivers/net/mana/mana.c           | 77 +++++++++++++++++++++++++++++++
 2 files changed, 79 insertions(+)

diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
index 7922816d66..b2729aba3a 100644
--- a/doc/guides/nics/features/mana.ini
+++ b/doc/guides/nics/features/mana.ini
@@ -4,6 +4,7 @@
 ; Refer to default.ini for the full list of available PMD features.
 ;
 [Features]
+Basic stats          = Y
 Free Tx mbuf on demand = Y
 Link status          = P
 Linux                = Y
@@ -14,5 +15,6 @@ Queue start/stop     = Y
 Removal event        = Y
 RSS hash             = Y
 Speed capabilities   = P
+Stats per queue      = Y
 Usage doc            = Y
 x86-64               = Y
diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index eb37f359db..bb8ef652bf 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -508,6 +508,79 @@ static int mana_dev_link_update(struct rte_eth_dev *dev,
 	return rte_eth_linkstatus_set(dev, &link);
 }
 
+static int mana_dev_stats_get(struct rte_eth_dev *dev,
+			      struct rte_eth_stats *stats)
+{
+	unsigned int i;
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		struct mana_txq *txq = dev->data->tx_queues[i];
+
+		if (!txq)
+			continue;
+
+		stats->opackets = txq->stats.packets;
+		stats->obytes = txq->stats.bytes;
+		stats->oerrors = txq->stats.errors;
+
+		if (i < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
+			stats->q_opackets[i] = txq->stats.packets;
+			stats->q_obytes[i] = txq->stats.bytes;
+		}
+	}
+
+	stats->rx_nombuf = 0;
+	for (i = 0; i < dev->data->nb_rx_queues; i++) {
+		struct mana_rxq *rxq = dev->data->rx_queues[i];
+
+		if (!rxq)
+			continue;
+
+		stats->ipackets = rxq->stats.packets;
+		stats->ibytes = rxq->stats.bytes;
+		stats->ierrors = rxq->stats.errors;
+
+		/* There is no good way to get stats->imissed, not setting it */
+
+		if (i < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
+			stats->q_ipackets[i] = rxq->stats.packets;
+			stats->q_ibytes[i] = rxq->stats.bytes;
+		}
+
+		stats->rx_nombuf += rxq->stats.nombuf;
+	}
+
+	return 0;
+}
+
+static int
+mana_dev_stats_reset(struct rte_eth_dev *dev __rte_unused)
+{
+	unsigned int i;
+
+	PMD_INIT_FUNC_TRACE();
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		struct mana_txq *txq = dev->data->tx_queues[i];
+
+		if (!txq)
+			continue;
+
+		memset(&txq->stats, 0, sizeof(txq->stats));
+	}
+
+	for (i = 0; i < dev->data->nb_rx_queues; i++) {
+		struct mana_rxq *rxq = dev->data->rx_queues[i];
+
+		if (!rxq)
+			continue;
+
+		memset(&rxq->stats, 0, sizeof(rxq->stats));
+	}
+
+	return 0;
+}
+
 const struct eth_dev_ops mana_dev_ops = {
 	.dev_configure		= mana_dev_configure,
 	.dev_start		= mana_dev_start,
@@ -524,9 +597,13 @@ const struct eth_dev_ops mana_dev_ops = {
 	.rx_queue_setup		= mana_dev_rx_queue_setup,
 	.rx_queue_release	= mana_dev_rx_queue_release,
 	.link_update		= mana_dev_link_update,
+	.stats_get		= mana_dev_stats_get,
+	.stats_reset		= mana_dev_stats_reset,
 };
 
 const struct eth_dev_ops mana_dev_sec_ops = {
+	.stats_get = mana_dev_stats_get,
+	.stats_reset = mana_dev_stats_reset,
 	.dev_infos_get = mana_dev_info_get,
 };
 
-- 
2.17.1

[Patch v6 18/18] net/mana: add function to support RX interrupts

[longli]

From: Long Li <longli@microsoft.com>

mana can receive RX interrupts from kernel through RDMA verbs interface.
Implement RX interrupts in the driver.

Signed-off-by: Long Li <longli@microsoft.com>
---
Change log:
v5:
New patch added to the series

 doc/guides/nics/features/mana.ini |   1 +
 drivers/net/mana/gdma.c           |  10 +--
 drivers/net/mana/mana.c           | 125 ++++++++++++++++++++++++++----
 drivers/net/mana/mana.h           |  13 +++-
 drivers/net/mana/rx.c             |  91 +++++++++++++++++++---
 drivers/net/mana/tx.c             |   3 +-
 6 files changed, 207 insertions(+), 36 deletions(-)

diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
index b2729aba3a..42d78ac6b1 100644
--- a/doc/guides/nics/features/mana.ini
+++ b/doc/guides/nics/features/mana.ini
@@ -14,6 +14,7 @@ Multiprocess aware   = Y
 Queue start/stop     = Y
 Removal event        = Y
 RSS hash             = Y
+Rx interrupt         = Y
 Speed capabilities   = P
 Stats per queue      = Y
 Usage doc            = Y
diff --git a/drivers/net/mana/gdma.c b/drivers/net/mana/gdma.c
index 7ad175651e..275520bff5 100644
--- a/drivers/net/mana/gdma.c
+++ b/drivers/net/mana/gdma.c
@@ -204,7 +204,7 @@ union gdma_doorbell_entry {
 #define DOORBELL_OFFSET_EQ      0xFF8
 
 int mana_ring_doorbell(void *db_page, enum gdma_queue_types queue_type,
-		       uint32_t queue_id, uint32_t tail)
+		       uint32_t queue_id, uint32_t tail, uint8_t arm)
 {
 	uint8_t *addr = db_page;
 	union gdma_doorbell_entry e = {};
@@ -219,14 +219,14 @@ int mana_ring_doorbell(void *db_page, enum gdma_queue_types queue_type,
 	case gdma_queue_receive:
 		e.rq.id = queue_id;
 		e.rq.tail_ptr = tail;
-		e.rq.wqe_cnt = 1;
+		e.rq.wqe_cnt = arm;
 		addr += DOORBELL_OFFSET_RQ;
 		break;
 
 	case gdma_queue_completion:
 		e.cq.id = queue_id;
 		e.cq.tail_ptr = tail;
-		e.cq.arm = 1;
+		e.cq.arm = arm;
 		addr += DOORBELL_OFFSET_CQ;
 		break;
 
@@ -238,8 +238,8 @@ int mana_ring_doorbell(void *db_page, enum gdma_queue_types queue_type,
 	/* Ensure all writes are done before ringing doorbell */
 	rte_wmb();
 
-	DRV_LOG(DEBUG, "db_page %p addr %p queue_id %u type %u tail %u",
-		db_page, addr, queue_id, queue_type, tail);
+	DRV_LOG(DEBUG, "db_page %p addr %p queue_id %u type %u tail %u arm %u",
+		db_page, addr, queue_id, queue_type, tail, arm);
 
 	rte_write64(e.as_uint64, addr);
 	return 0;
diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
index bb8ef652bf..935811ea01 100644
--- a/drivers/net/mana/mana.c
+++ b/drivers/net/mana/mana.c
@@ -95,7 +95,68 @@ static int mana_dev_configure(struct rte_eth_dev *dev)
 	return 0;
 }
 
-static int mana_intr_uninstall(struct mana_priv *priv);
+static void rx_intr_vec_disable(struct mana_priv *priv)
+{
+	struct rte_intr_handle *intr_handle = priv->intr_handle;
+
+	rte_intr_free_epoll_fd(intr_handle);
+	rte_intr_vec_list_free(intr_handle);
+	rte_intr_nb_efd_set(intr_handle, 0);
+}
+
+static int rx_intr_vec_enable(struct mana_priv *priv)
+{
+	unsigned int i;
+	unsigned int rxqs_n = priv->dev_data->nb_rx_queues;
+	unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
+	struct rte_intr_handle *intr_handle = priv->intr_handle;
+	int ret;
+
+	rx_intr_vec_disable(priv);
+
+	if (rte_intr_vec_list_alloc(intr_handle, NULL, n)) {
+		DRV_LOG(ERR, "Failed to allocate memory for interrupt vector");
+		return -ENOMEM;
+	}
+
+	for (i = 0; i < n; i++) {
+		struct mana_rxq *rxq = priv->dev_data->rx_queues[i];
+
+		ret = rte_intr_vec_list_index_set(intr_handle, i,
+						  RTE_INTR_VEC_RXTX_OFFSET + i);
+		if (ret) {
+			DRV_LOG(ERR, "Failed to set intr vec %u", i);
+			return ret;
+		}
+
+		ret = rte_intr_efds_index_set(intr_handle, i, rxq->channel->fd);
+		if (ret) {
+			DRV_LOG(ERR, "Failed to set FD at intr %u", i);
+			return ret;
+		}
+	}
+
+	return rte_intr_nb_efd_set(intr_handle, n);
+}
+
+static void rxq_intr_disable(struct mana_priv *priv)
+{
+	int err = rte_errno;
+
+	rx_intr_vec_disable(priv);
+	rte_errno = err;
+}
+
+static int rxq_intr_enable(struct mana_priv *priv)
+{
+	const struct rte_eth_intr_conf *const intr_conf =
+		&priv->dev_data->dev_conf.intr_conf;
+
+	if (!intr_conf->rxq)
+		return 0;
+
+	return rx_intr_vec_enable(priv);
+}
 
 static int
 mana_dev_start(struct rte_eth_dev *dev)
@@ -133,8 +194,17 @@ mana_dev_start(struct rte_eth_dev *dev)
 	/* Enable datapath for secondary processes */
 	mana_mp_req_on_rxtx(dev, MANA_MP_REQ_START_RXTX);
 
+	ret = rxq_intr_enable(priv);
+	if (ret) {
+		DRV_LOG(ERR, "Failed to enable RX interrupts");
+		goto failed_intr;
+	}
+
 	return 0;
 
+failed_intr:
+	mana_stop_rx_queues(dev);
+
 failed_rx:
 	mana_stop_tx_queues(dev);
 
@@ -145,9 +215,12 @@ mana_dev_start(struct rte_eth_dev *dev)
 }
 
 static int
-mana_dev_stop(struct rte_eth_dev *dev __rte_unused)
+mana_dev_stop(struct rte_eth_dev *dev)
 {
 	int ret;
+	struct mana_priv *priv = dev->data->dev_private;
+
+	rxq_intr_disable(priv);
 
 	dev->tx_pkt_burst = mana_tx_burst_removed;
 	dev->rx_pkt_burst = mana_rx_burst_removed;
@@ -596,6 +669,8 @@ const struct eth_dev_ops mana_dev_ops = {
 	.tx_queue_release	= mana_dev_tx_queue_release,
 	.rx_queue_setup		= mana_dev_rx_queue_setup,
 	.rx_queue_release	= mana_dev_rx_queue_release,
+	.rx_queue_intr_enable	= mana_rx_intr_enable,
+	.rx_queue_intr_disable	= mana_rx_intr_disable,
 	.link_update		= mana_dev_link_update,
 	.stats_get		= mana_dev_stats_get,
 	.stats_reset		= mana_dev_stats_reset,
@@ -783,7 +858,7 @@ static int mana_ibv_device_to_pci_addr(const struct ibv_device *device,
 	return 0;
 }
 
-static void mana_intr_handler(void *arg)
+void mana_intr_handler(void *arg)
 {
 	struct mana_priv *priv = arg;
 	struct ibv_context *ctx = priv->ib_ctx;
@@ -807,7 +882,7 @@ static void mana_intr_handler(void *arg)
 	}
 }
 
-static int mana_intr_uninstall(struct mana_priv *priv)
+int mana_intr_uninstall(struct mana_priv *priv)
 {
 	int ret;
 
@@ -823,9 +898,20 @@ static int mana_intr_uninstall(struct mana_priv *priv)
 	return 0;
 }
 
-static int mana_intr_install(struct mana_priv *priv)
+int mana_fd_set_non_blocking(int fd)
+{
+	int ret = fcntl(fd, F_GETFL);
+
+	if (ret != -1 && !fcntl(fd, F_SETFL, ret | O_NONBLOCK))
+		return 0;
+
+	rte_errno = errno;
+	return -rte_errno;
+}
+
+int mana_intr_install(struct rte_eth_dev *eth_dev, struct mana_priv *priv)
 {
-	int ret, flags;
+	int ret;
 	struct ibv_context *ctx = priv->ib_ctx;
 
 	priv->intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
@@ -835,31 +921,35 @@ static int mana_intr_install(struct mana_priv *priv)
 		return -ENOMEM;
 	}
 
-	rte_intr_fd_set(priv->intr_handle, -1);
+	ret = rte_intr_fd_set(priv->intr_handle, -1);
+	if (ret)
+		goto free_intr;
 
-	flags = fcntl(ctx->async_fd, F_GETFL);
-	ret = fcntl(ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
+	ret = mana_fd_set_non_blocking(ctx->async_fd);
 	if (ret) {
 		DRV_LOG(ERR, "Failed to change async_fd to NONBLOCK");
 		goto free_intr;
 	}
 
-	rte_intr_fd_set(priv->intr_handle, ctx->async_fd);
-	rte_intr_type_set(priv->intr_handle, RTE_INTR_HANDLE_EXT);
+	ret = rte_intr_fd_set(priv->intr_handle, ctx->async_fd);
+	if (ret)
+		goto free_intr;
+
+	ret = rte_intr_type_set(priv->intr_handle, RTE_INTR_HANDLE_EXT);
+	if (ret)
+		goto free_intr;
 
 	ret = rte_intr_callback_register(priv->intr_handle,
 					 mana_intr_handler, priv);
 	if (ret) {
 		DRV_LOG(ERR, "Failed to register intr callback");
 		rte_intr_fd_set(priv->intr_handle, -1);
-		goto restore_fd;
+		goto free_intr;
 	}
 
+	eth_dev->intr_handle = priv->intr_handle;
 	return 0;
 
-restore_fd:
-	fcntl(ctx->async_fd, F_SETFL, flags);
-
 free_intr:
 	rte_intr_instance_free(priv->intr_handle);
 	priv->intr_handle = NULL;
@@ -1183,8 +1273,10 @@ static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv __rte_unused,
 				name, priv->max_rx_queues, priv->max_rx_desc,
 				priv->max_send_sge);
 
+			rte_eth_copy_pci_info(eth_dev, pci_dev);
+
 			/* Create async interrupt handler */
-			ret = mana_intr_install(priv);
+			ret = mana_intr_install(eth_dev, priv);
 			if (ret) {
 				DRV_LOG(ERR, "Failed to install intr handler");
 				goto failed;
@@ -1207,7 +1299,6 @@ static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv __rte_unused,
 			eth_dev->tx_pkt_burst = mana_tx_burst_removed;
 			eth_dev->dev_ops = &mana_dev_ops;
 
-			rte_eth_copy_pci_info(eth_dev, pci_dev);
 			rte_eth_dev_probing_finish(eth_dev);
 		}
 
diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index b4056bd50b..9cc98fee4c 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -427,6 +427,7 @@ struct mana_rxq {
 	uint32_t num_desc;
 	struct rte_mempool *mp;
 	struct ibv_cq *cq;
+	struct ibv_comp_channel *channel;
 	struct ibv_wq *wq;
 
 	/* For storing pending requests */
@@ -460,8 +461,8 @@ extern int mana_logtype_init;
 #define PMD_INIT_FUNC_TRACE() PMD_INIT_LOG(DEBUG, " >>")
 
 int mana_ring_doorbell(void *db_page, enum gdma_queue_types queue_type,
-		       uint32_t queue_id, uint32_t tail);
-int mana_rq_ring_doorbell(struct mana_rxq *rxq);
+		       uint32_t queue_id, uint32_t tail, uint8_t arm);
+int mana_rq_ring_doorbell(struct mana_rxq *rxq, uint8_t arm);
 
 int gdma_post_work_request(struct mana_gdma_queue *queue,
 			   struct gdma_work_request *work_req,
@@ -496,6 +497,10 @@ int mana_new_pmd_mr(struct mana_mr_btree *local_tree, struct mana_priv *priv,
 void mana_remove_all_mr(struct mana_priv *priv);
 void mana_del_pmd_mr(struct mana_mr_cache *mr);
 
+void mana_intr_handler(void *arg);
+int mana_intr_install(struct rte_eth_dev *eth_dev, struct mana_priv *priv);
+int mana_intr_uninstall(struct mana_priv *priv);
+
 void mana_mempool_chunk_cb(struct rte_mempool *mp, void *opaque,
 			   struct rte_mempool_memhdr *memhdr, unsigned int idx);
 
@@ -541,4 +546,8 @@ void mana_mp_req_on_rxtx(struct rte_eth_dev *dev, enum mana_mp_req_type type);
 void *mana_alloc_verbs_buf(size_t size, void *data);
 void mana_free_verbs_buf(void *ptr, void *data __rte_unused);
 
+int mana_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
+int mana_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
+int mana_fd_set_non_blocking(int fd);
+
 #endif
diff --git a/drivers/net/mana/rx.c b/drivers/net/mana/rx.c
index f2573a6d06..1a61fc59b1 100644
--- a/drivers/net/mana/rx.c
+++ b/drivers/net/mana/rx.c
@@ -21,7 +21,7 @@ static uint8_t mana_rss_hash_key_default[TOEPLITZ_HASH_KEY_SIZE_IN_BYTES] = {
 	0xfc, 0x1f, 0xdc, 0x2a,
 };
 
-int mana_rq_ring_doorbell(struct mana_rxq *rxq)
+int mana_rq_ring_doorbell(struct mana_rxq *rxq, uint8_t arm)
 {
 	struct mana_priv *priv = rxq->priv;
 	int ret;
@@ -36,9 +36,9 @@ int mana_rq_ring_doorbell(struct mana_rxq *rxq)
 	}
 
 	ret = mana_ring_doorbell(db_page, gdma_queue_receive,
-				 rxq->gdma_rq.id,
-				 rxq->gdma_rq.head *
-					GDMA_WQE_ALIGNMENT_UNIT_SIZE);
+			 rxq->gdma_rq.id,
+			 rxq->gdma_rq.head * GDMA_WQE_ALIGNMENT_UNIT_SIZE,
+			 arm);
 
 	if (ret)
 		DRV_LOG(ERR, "failed to ring RX doorbell ret %d", ret);
@@ -115,7 +115,7 @@ static int mana_alloc_and_post_rx_wqes(struct mana_rxq *rxq)
 		}
 	}
 
-	mana_rq_ring_doorbell(rxq);
+	mana_rq_ring_doorbell(rxq, rxq->num_desc);
 
 	return ret;
 }
@@ -156,6 +156,14 @@ int mana_stop_rx_queues(struct rte_eth_dev *dev)
 				DRV_LOG(ERR,
 					"rx_queue destroy_cq failed %d", ret);
 			rxq->cq = NULL;
+
+			if (rxq->channel) {
+				ret = ibv_destroy_comp_channel(rxq->channel);
+				if (ret)
+					DRV_LOG(ERR, "failed destroy comp %d",
+						ret);
+				rxq->channel = NULL;
+			}
 		}
 
 		/* Drain and free posted WQEs */
@@ -196,8 +204,24 @@ int mana_start_rx_queues(struct rte_eth_dev *dev)
 				.data = (void *)(uintptr_t)rxq->socket,
 			}));
 
+		if (dev->data->dev_conf.intr_conf.rxq) {
+			rxq->channel = ibv_create_comp_channel(priv->ib_ctx);
+			if (!rxq->channel) {
+				ret = -errno;
+				DRV_LOG(ERR, "Queue %d comp channel failed", i);
+				goto fail;
+			}
+
+			ret = mana_fd_set_non_blocking(rxq->channel->fd);
+			if (ret) {
+				DRV_LOG(ERR, "Failed to set comp non-blocking");
+				goto fail;
+			}
+		}
+
 		rxq->cq = ibv_create_cq(priv->ib_ctx, rxq->num_desc,
-					NULL, NULL, 0);
+					NULL, rxq->channel,
+					rxq->channel ? i : 0);
 		if (!rxq->cq) {
 			ret = -errno;
 			DRV_LOG(ERR, "failed to create rx cq queue %d", i);
@@ -347,7 +371,8 @@ int mana_start_rx_queues(struct rte_eth_dev *dev)
 
 uint16_t mana_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 {
-	uint16_t pkt_received = 0, cqe_processed = 0;
+	uint16_t pkt_received = 0;
+	uint8_t wqe_posted = 0;
 	struct mana_rxq *rxq = dpdk_rxq;
 	struct mana_priv *priv = rxq->priv;
 	struct gdma_comp comp;
@@ -433,18 +458,62 @@ uint16_t mana_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 		if (rxq->desc_ring_tail >= rxq->num_desc)
 			rxq->desc_ring_tail = 0;
 
-		cqe_processed++;
-
 		/* Post another request */
 		ret = mana_alloc_and_post_rx_wqe(rxq);
 		if (ret) {
 			DRV_LOG(ERR, "failed to post rx wqe ret=%d", ret);
 			break;
 		}
+
+		wqe_posted++;
 	}
 
-	if (cqe_processed)
-		mana_rq_ring_doorbell(rxq);
+	if (wqe_posted)
+		mana_rq_ring_doorbell(rxq, wqe_posted);
 
 	return pkt_received;
 }
+
+static int mana_arm_cq(struct mana_rxq *rxq, uint8_t arm)
+{
+	struct mana_priv *priv = rxq->priv;
+	uint32_t head = rxq->gdma_cq.head %
+		(rxq->gdma_cq.count << COMPLETION_QUEUE_ENTRY_OWNER_BITS_SIZE);
+
+	DRV_LOG(ERR, "Ringing completion queue ID %u head %u arm %d",
+		rxq->gdma_cq.id, head, arm);
+
+	return mana_ring_doorbell(priv->db_page, gdma_queue_completion,
+				  rxq->gdma_cq.id, head, arm);
+}
+
+int mana_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
+{
+	struct mana_rxq *rxq = dev->data->rx_queues[rx_queue_id];
+
+	return mana_arm_cq(rxq, 1);
+}
+
+int mana_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
+{
+	struct mana_rxq *rxq = dev->data->rx_queues[rx_queue_id];
+	struct ibv_cq *ev_cq;
+	void *ev_ctx;
+	int ret;
+
+	ret = ibv_get_cq_event(rxq->channel, &ev_cq, &ev_ctx);
+	if (ret)
+		ret = errno;
+	else if (ev_cq != rxq->cq)
+		ret = EINVAL;
+
+	if (ret) {
+		if (ret != EAGAIN)
+			DRV_LOG(ERR, "Can't disable RX intr queue %d",
+				rx_queue_id);
+	} else {
+		ibv_ack_cq_events(rxq->cq, 1);
+	}
+
+	return -ret;
+}
diff --git a/drivers/net/mana/tx.c b/drivers/net/mana/tx.c
index 0aefab1b6c..5878ea62f9 100644
--- a/drivers/net/mana/tx.c
+++ b/drivers/net/mana/tx.c
@@ -396,7 +396,8 @@ uint16_t mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts,
 	ret = mana_ring_doorbell(db_page, gdma_queue_send,
 				 txq->gdma_sq.id,
 				 txq->gdma_sq.head *
-					GDMA_WQE_ALIGNMENT_UNIT_SIZE);
+					GDMA_WQE_ALIGNMENT_UNIT_SIZE,
+				 0);
 	if (ret)
 		DRV_LOG(ERR, "mana_ring_doorbell failed ret %d", ret);
 
-- 
2.17.1

Re: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[lihuisong (C)]

在 2022/8/31 6:51, longli@linuxonhyperv.com 写道:
> From: Long Li <longli@microsoft.com>
>
> MANA is a PCI device. It uses IB verbs to access hardware through the
> kernel RDMA layer. This patch introduces build environment and basic
> device probe functions.
>
> Signed-off-by: Long Li <longli@microsoft.com>
> ---
> Change log:
> v2:
> Fix typos.
> Make the driver build only on x86-64 and Linux.
> Remove unused header files.
> Change port definition to uint16_t or uint8_t (for IB).
> Use getline() in place of fgets() to read and truncate a line.
> v3:
> Add meson build check for required functions from RDMA direct verb header file
> v4:
> Remove extra "\n" in logging code.
> Use "r" in place of "rb" in fopen() to read text files.
>
> [snip]
> +
> +static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv __rte_unused,
> +			      struct rte_pci_device *pci_dev,
> +			      struct rte_ether_addr *mac_addr)
> +{
> +	struct ibv_device **ibv_list;
> +	int ibv_idx;
> +	struct ibv_context *ctx;
> +	struct ibv_device_attr_ex dev_attr;
> +	int num_devices;
> +	int ret = 0;
> +	uint8_t port;
> +	struct mana_priv *priv = NULL;
> +	struct rte_eth_dev *eth_dev = NULL;
> +	bool found_port;
> +
> +	ibv_list = ibv_get_device_list(&num_devices);
> +	for (ibv_idx = 0; ibv_idx < num_devices; ibv_idx++) {
> +		struct ibv_device *ibdev = ibv_list[ibv_idx];
> +		struct rte_pci_addr pci_addr;
> +
> +		DRV_LOG(INFO, "Probe device name %s dev_name %s ibdev_path %s",
> +			ibdev->name, ibdev->dev_name, ibdev->ibdev_path);
> +
> +		if (mana_ibv_device_to_pci_addr(ibdev, &pci_addr))
> +			continue;
> +
> +		/* Ignore if this IB device is not this PCI device */
> +		if (pci_dev->addr.domain != pci_addr.domain ||
> +		    pci_dev->addr.bus != pci_addr.bus ||
> +		    pci_dev->addr.devid != pci_addr.devid ||
> +		    pci_dev->addr.function != pci_addr.function)
> +			continue;
> +
> +		ctx = ibv_open_device(ibdev);
> +		if (!ctx) {
> +			DRV_LOG(ERR, "Failed to open IB device %s",
> +				ibdev->name);
> +			continue;
> +		}
> +
> +		ret = ibv_query_device_ex(ctx, NULL, &dev_attr);
> +		DRV_LOG(INFO, "dev_attr.orig_attr.phys_port_cnt %u",
> +			dev_attr.orig_attr.phys_port_cnt);
> +		found_port = false;
> +
> +		for (port = 1; port <= dev_attr.orig_attr.phys_port_cnt;
> +		     port++) {
> +			struct ibv_parent_domain_init_attr attr = {};
> +			struct rte_ether_addr addr;
> +			char address[64];
> +			char name[RTE_ETH_NAME_MAX_LEN];
> +
> +			ret = get_port_mac(ibdev, port, &addr);
> +			if (ret)
> +				continue;
> +
> +			if (mac_addr && !rte_is_same_ether_addr(&addr, mac_addr))
> +				continue;
> +
> +			rte_ether_format_addr(address, sizeof(address), &addr);
> +			DRV_LOG(INFO, "device located port %u address %s",
> +				port, address);
> +			found_port = true;
> +
> +			priv = rte_zmalloc_socket(NULL, sizeof(*priv),
> +						  RTE_CACHE_LINE_SIZE,
> +						  SOCKET_ID_ANY);
> +			if (!priv) {
> +				ret = -ENOMEM;
> +				goto failed;
> +			}
> +
> +			snprintf(name, sizeof(name), "%s_port%d",
> +				 pci_dev->device.name, port);
> +
> +			if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
> +				int fd;
> +
> +				eth_dev = rte_eth_dev_attach_secondary(name);
> +				if (!eth_dev) {
> +					DRV_LOG(ERR, "Can't attach to dev %s",
> +						name);
> +					ret = -ENOMEM;
> +					goto failed;
> +				}
> +
> +				eth_dev->device = &pci_dev->device;
> +				eth_dev->dev_ops = &mana_dev_sec_ops;
> +				ret = mana_proc_priv_init(eth_dev);
> +				if (ret)
> +					goto failed;
> +				priv->process_priv = eth_dev->process_private;
> +
> +				/* Get the IB FD from the primary process */
> +				fd = mana_mp_req_verbs_cmd_fd(eth_dev);
> +				if (fd < 0) {
> +					DRV_LOG(ERR, "Failed to get FD %d", fd);
> +					ret = -ENODEV;
> +					goto failed;
> +				}
> +
> +				ret = mana_map_doorbell_secondary(eth_dev, fd);
> +				if (ret) {
> +					DRV_LOG(ERR, "Failed secondary map %d",
> +						fd);
> +					goto failed;
> +				}
> +
> +				/* fd is no not used after mapping doorbell */
> +				close(fd);
> +
> +				rte_spinlock_lock(&mana_shared_data->lock);
> +				mana_shared_data->secondary_cnt++;
> +				mana_local_data.secondary_cnt++;
> +				rte_spinlock_unlock(&mana_shared_data->lock);
> +
> +				rte_eth_copy_pci_info(eth_dev, pci_dev);
> +				rte_eth_dev_probing_finish(eth_dev);
> +
> +				/* Impossible to have more than one port
> +				 * matching a MAC address
> +				 */
> +				continue;
> +			}
> +
> +			eth_dev = rte_eth_dev_allocate(name);
> +			if (!eth_dev) {
> +				ret = -ENOMEM;
> +				goto failed;
> +			}
> +
> +			eth_dev->data->mac_addrs =
> +				rte_calloc("mana_mac", 1,
> +					   sizeof(struct rte_ether_addr), 0);
> +			if (!eth_dev->data->mac_addrs) {
> +				ret = -ENOMEM;
> +				goto failed;
> +			}
> +
> +			rte_ether_addr_copy(&addr, eth_dev->data->mac_addrs);
> +
> +			priv->ib_pd = ibv_alloc_pd(ctx);
> +			if (!priv->ib_pd) {
> +				DRV_LOG(ERR, "ibv_alloc_pd failed port %d", port);
> +				ret = -ENOMEM;
> +				goto failed;
> +			}
> +
> +			/* Create a parent domain with the port number */
> +			attr.pd = priv->ib_pd;
> +			attr.comp_mask = IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT;
> +			attr.pd_context = (void *)(uint64_t)port;
> +			priv->ib_parent_pd = ibv_alloc_parent_domain(ctx, &attr);
> +			if (!priv->ib_parent_pd) {
> +				DRV_LOG(ERR,
> +					"ibv_alloc_parent_domain failed port %d",
> +					port);
> +				ret = -ENOMEM;
> +				goto failed;
> +			}
> +
> +			priv->ib_ctx = ctx;
> +			priv->port_id = eth_dev->data->port_id;
> +			priv->dev_port = port;
> +			eth_dev->data->dev_private = priv;
> +			priv->dev_data = eth_dev->data;
> +
> +			priv->max_rx_queues = dev_attr.orig_attr.max_qp;
> +			priv->max_tx_queues = dev_attr.orig_attr.max_qp;
> +
> +			priv->max_rx_desc =
> +				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
> +					dev_attr.orig_attr.max_cqe);
> +			priv->max_tx_desc =
> +				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
> +					dev_attr.orig_attr.max_cqe);
> +
> +			priv->max_send_sge = dev_attr.orig_attr.max_sge;
> +			priv->max_recv_sge = dev_attr.orig_attr.max_sge;
> +
> +			priv->max_mr = dev_attr.orig_attr.max_mr;
> +			priv->max_mr_size = dev_attr.orig_attr.max_mr_size;
> +
> +			DRV_LOG(INFO, "dev %s max queues %d desc %d sge %d",
> +				name, priv->max_rx_queues, priv->max_rx_desc,
> +				priv->max_send_sge);
> +
> +			rte_spinlock_lock(&mana_shared_data->lock);
> +			mana_shared_data->primary_cnt++;
> +			rte_spinlock_unlock(&mana_shared_data->lock);
> +
> +			eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
> +
> +			eth_dev->device = &pci_dev->device;
> +			eth_dev->data->dev_flags |=
> +				RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
> +

Please do not use the temporary macro. Please review this patch:

f30e69b41f94 ("ethdev: add device flag to bypass auto-filled queue xstats")

This patch requires that per queue statistics are filled in 
.xstats_get() by PMD.

> +			DRV_LOG(INFO, "device %s at port %u",
> +				name, eth_dev->data->port_id);
> +
> +			eth_dev->rx_pkt_burst = mana_rx_burst_removed;
> +			eth_dev->tx_pkt_burst = mana_tx_burst_removed;
> +			eth_dev->dev_ops = &mana_dev_ops;
> +
> +			rte_eth_copy_pci_info(eth_dev, pci_dev);
> +			rte_eth_dev_probing_finish(eth_dev);
> +		}
> +
> +		/* Secondary process doesn't need an ibv_ctx. It maps the
> +		 * doorbell pages using the IB cmd_fd passed from the primary
> +		 * process and send messages to primary process for memory
> +		 * registartions.
> +		 */
> +		if (!found_port || rte_eal_process_type() == RTE_PROC_SECONDARY)
> +			ibv_close_device(ctx);
> +	}
> +
> +	ibv_free_device_list(ibv_list);
> +	return 0;
> +
> +failed:
> +	/* Free the resource for the port failed */
> +	if (priv) {
> +		if (priv->ib_parent_pd)
> +			ibv_dealloc_pd(priv->ib_parent_pd);
> +
> +		if (priv->ib_pd)
> +			ibv_dealloc_pd(priv->ib_pd);
> +	}
> +
> +	if (eth_dev)
> +		rte_eth_dev_release_port(eth_dev);
> +
> +	rte_free(priv);
> +
> +	ibv_close_device(ctx);
> +	ibv_free_device_list(ibv_list);
> +
> +	return ret;
> +}
> [snip]
>

RE: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[Long Li]

> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build environment
> and doc
> 
> 
> 在 2022/8/31 6:51, longli@linuxonhyperv.com 写道:
> > From: Long Li <longli@microsoft.com>
> >
> > MANA is a PCI device. It uses IB verbs to access hardware through the
> > kernel RDMA layer. This patch introduces build environment and basic
> > device probe functions.
> >
> > Signed-off-by: Long Li <longli@microsoft.com>
> > ---
> > Change log:
> > v2:
> > Fix typos.
> > Make the driver build only on x86-64 and Linux.
> > Remove unused header files.
> > Change port definition to uint16_t or uint8_t (for IB).
> > Use getline() in place of fgets() to read and truncate a line.
> > v3:
> > Add meson build check for required functions from RDMA direct verb
> > header file
> > v4:
> > Remove extra "\n" in logging code.
> > Use "r" in place of "rb" in fopen() to read text files.
> >
> > [snip]
> > +
> > +static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv
> __rte_unused,
> > +			      struct rte_pci_device *pci_dev,
> > +			      struct rte_ether_addr *mac_addr) {
> > +	struct ibv_device **ibv_list;
> > +	int ibv_idx;
> > +	struct ibv_context *ctx;
> > +	struct ibv_device_attr_ex dev_attr;
> > +	int num_devices;
> > +	int ret = 0;
> > +	uint8_t port;
> > +	struct mana_priv *priv = NULL;
> > +	struct rte_eth_dev *eth_dev = NULL;
> > +	bool found_port;
> > +
> > +	ibv_list = ibv_get_device_list(&num_devices);
> > +	for (ibv_idx = 0; ibv_idx < num_devices; ibv_idx++) {
> > +		struct ibv_device *ibdev = ibv_list[ibv_idx];
> > +		struct rte_pci_addr pci_addr;
> > +
> > +		DRV_LOG(INFO, "Probe device name %s dev_name %s
> ibdev_path %s",
> > +			ibdev->name, ibdev->dev_name, ibdev-
> >ibdev_path);
> > +
> > +		if (mana_ibv_device_to_pci_addr(ibdev, &pci_addr))
> > +			continue;
> > +
> > +		/* Ignore if this IB device is not this PCI device */
> > +		if (pci_dev->addr.domain != pci_addr.domain ||
> > +		    pci_dev->addr.bus != pci_addr.bus ||
> > +		    pci_dev->addr.devid != pci_addr.devid ||
> > +		    pci_dev->addr.function != pci_addr.function)
> > +			continue;
> > +
> > +		ctx = ibv_open_device(ibdev);
> > +		if (!ctx) {
> > +			DRV_LOG(ERR, "Failed to open IB device %s",
> > +				ibdev->name);
> > +			continue;
> > +		}
> > +
> > +		ret = ibv_query_device_ex(ctx, NULL, &dev_attr);
> > +		DRV_LOG(INFO, "dev_attr.orig_attr.phys_port_cnt %u",
> > +			dev_attr.orig_attr.phys_port_cnt);
> > +		found_port = false;
> > +
> > +		for (port = 1; port <= dev_attr.orig_attr.phys_port_cnt;
> > +		     port++) {
> > +			struct ibv_parent_domain_init_attr attr = {};
> > +			struct rte_ether_addr addr;
> > +			char address[64];
> > +			char name[RTE_ETH_NAME_MAX_LEN];
> > +
> > +			ret = get_port_mac(ibdev, port, &addr);
> > +			if (ret)
> > +				continue;
> > +
> > +			if (mac_addr && !rte_is_same_ether_addr(&addr,
> mac_addr))
> > +				continue;
> > +
> > +			rte_ether_format_addr(address, sizeof(address),
> &addr);
> > +			DRV_LOG(INFO, "device located port %u address
> %s",
> > +				port, address);
> > +			found_port = true;
> > +
> > +			priv = rte_zmalloc_socket(NULL, sizeof(*priv),
> > +						  RTE_CACHE_LINE_SIZE,
> > +						  SOCKET_ID_ANY);
> > +			if (!priv) {
> > +				ret = -ENOMEM;
> > +				goto failed;
> > +			}
> > +
> > +			snprintf(name, sizeof(name), "%s_port%d",
> > +				 pci_dev->device.name, port);
> > +
> > +			if (rte_eal_process_type() ==
> RTE_PROC_SECONDARY) {
> > +				int fd;
> > +
> > +				eth_dev =
> rte_eth_dev_attach_secondary(name);
> > +				if (!eth_dev) {
> > +					DRV_LOG(ERR, "Can't attach to dev
> %s",
> > +						name);
> > +					ret = -ENOMEM;
> > +					goto failed;
> > +				}
> > +
> > +				eth_dev->device = &pci_dev->device;
> > +				eth_dev->dev_ops = &mana_dev_sec_ops;
> > +				ret = mana_proc_priv_init(eth_dev);
> > +				if (ret)
> > +					goto failed;
> > +				priv->process_priv = eth_dev-
> >process_private;
> > +
> > +				/* Get the IB FD from the primary process */
> > +				fd =
> mana_mp_req_verbs_cmd_fd(eth_dev);
> > +				if (fd < 0) {
> > +					DRV_LOG(ERR, "Failed to get FD %d",
> fd);
> > +					ret = -ENODEV;
> > +					goto failed;
> > +				}
> > +
> > +				ret =
> mana_map_doorbell_secondary(eth_dev, fd);
> > +				if (ret) {
> > +					DRV_LOG(ERR, "Failed secondary
> map %d",
> > +						fd);
> > +					goto failed;
> > +				}
> > +
> > +				/* fd is no not used after mapping doorbell */
> > +				close(fd);
> > +
> > +				rte_spinlock_lock(&mana_shared_data-
> >lock);
> > +				mana_shared_data->secondary_cnt++;
> > +				mana_local_data.secondary_cnt++;
> > +				rte_spinlock_unlock(&mana_shared_data-
> >lock);
> > +
> > +				rte_eth_copy_pci_info(eth_dev, pci_dev);
> > +				rte_eth_dev_probing_finish(eth_dev);
> > +
> > +				/* Impossible to have more than one port
> > +				 * matching a MAC address
> > +				 */
> > +				continue;
> > +			}
> > +
> > +			eth_dev = rte_eth_dev_allocate(name);
> > +			if (!eth_dev) {
> > +				ret = -ENOMEM;
> > +				goto failed;
> > +			}
> > +
> > +			eth_dev->data->mac_addrs =
> > +				rte_calloc("mana_mac", 1,
> > +					   sizeof(struct rte_ether_addr), 0);
> > +			if (!eth_dev->data->mac_addrs) {
> > +				ret = -ENOMEM;
> > +				goto failed;
> > +			}
> > +
> > +			rte_ether_addr_copy(&addr, eth_dev->data-
> >mac_addrs);
> > +
> > +			priv->ib_pd = ibv_alloc_pd(ctx);
> > +			if (!priv->ib_pd) {
> > +				DRV_LOG(ERR, "ibv_alloc_pd failed port %d",
> port);
> > +				ret = -ENOMEM;
> > +				goto failed;
> > +			}
> > +
> > +			/* Create a parent domain with the port number */
> > +			attr.pd = priv->ib_pd;
> > +			attr.comp_mask =
> IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT;
> > +			attr.pd_context = (void *)(uint64_t)port;
> > +			priv->ib_parent_pd = ibv_alloc_parent_domain(ctx,
> &attr);
> > +			if (!priv->ib_parent_pd) {
> > +				DRV_LOG(ERR,
> > +					"ibv_alloc_parent_domain failed port
> %d",
> > +					port);
> > +				ret = -ENOMEM;
> > +				goto failed;
> > +			}
> > +
> > +			priv->ib_ctx = ctx;
> > +			priv->port_id = eth_dev->data->port_id;
> > +			priv->dev_port = port;
> > +			eth_dev->data->dev_private = priv;
> > +			priv->dev_data = eth_dev->data;
> > +
> > +			priv->max_rx_queues = dev_attr.orig_attr.max_qp;
> > +			priv->max_tx_queues = dev_attr.orig_attr.max_qp;
> > +
> > +			priv->max_rx_desc =
> > +				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
> > +					dev_attr.orig_attr.max_cqe);
> > +			priv->max_tx_desc =
> > +				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
> > +					dev_attr.orig_attr.max_cqe);
> > +
> > +			priv->max_send_sge = dev_attr.orig_attr.max_sge;
> > +			priv->max_recv_sge = dev_attr.orig_attr.max_sge;
> > +
> > +			priv->max_mr = dev_attr.orig_attr.max_mr;
> > +			priv->max_mr_size =
> dev_attr.orig_attr.max_mr_size;
> > +
> > +			DRV_LOG(INFO, "dev %s max queues %d desc %d
> sge %d",
> > +				name, priv->max_rx_queues, priv-
> >max_rx_desc,
> > +				priv->max_send_sge);
> > +
> > +			rte_spinlock_lock(&mana_shared_data->lock);
> > +			mana_shared_data->primary_cnt++;
> > +			rte_spinlock_unlock(&mana_shared_data->lock);
> > +
> > +			eth_dev->data->dev_flags |=
> RTE_ETH_DEV_INTR_RMV;
> > +
> > +			eth_dev->device = &pci_dev->device;
> > +			eth_dev->data->dev_flags |=
> > +				RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
> > +
> 
> Please do not use the temporary macro. Please review this patch:
> 
> f30e69b41f94 ("ethdev: add device flag to bypass auto-filled queue xstats")
> 
> This patch requires that per queue statistics are filled in
> .xstats_get() by PMD.

Thanks for pointing this out.

It seems some PMDs are still depending on this flag for xstats.

MANA doesn't implement xstats_get() currently, this flag is useful. Is it okay to keep using this flag before it's finally the time to remove it from all PMDs, or when MANA implements xstats?

Re: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[fengchengwen]

On 2022/8/31 6:51, longli@linuxonhyperv.com wrote:
> From: Long Li <longli@microsoft.com>
> 
> MANA is a PCI device. It uses IB verbs to access hardware through the
> kernel RDMA layer. This patch introduces build environment and basic
> device probe functions.
> 
> Signed-off-by: Long Li <longli@microsoft.com>

...

> +static int mana_mp_primary_handle(const struct rte_mp_msg *mp_msg,
> +				  const void *peer)
> +{
> +	struct rte_eth_dev *dev;
> +	const struct mana_mp_param *param =
> +		(const struct mana_mp_param *)mp_msg->param;
> +	struct rte_mp_msg mp_res = { 0 };
> +	struct mana_mp_param *res = (struct mana_mp_param *)mp_res.param;
> +	int ret;
> +	struct mana_priv *priv;
> +
> +	if (!rte_eth_dev_is_valid_port(param->port_id)) {
> +		DRV_LOG(ERR, "MP handle port ID %u invalid", param->port_id);
> +		return -ENODEV;
> +	}
> +
> +	dev = &rte_eth_devices[param->port_id];
> +	priv = dev->data->dev_private;
> +
> +	mp_init_msg(&mp_res, param->type, param->port_id);
> +
> +	switch (param->type) {
> +	case MANA_MP_REQ_VERBS_CMD_FD:
> +		mp_res.num_fds = 1;
> +		mp_res.fds[0] = priv->ib_ctx->cmd_fd;

The cmd_fd is system level handler?

If it's process private handler, it should not used directly in another process.

> +		res->result = 0;
> +		ret = rte_mp_reply(&mp_res, peer);
> +		break;
> +
> +	default:
> +		DRV_LOG(ERR, "Port %u unknown primary MP type %u",
> +			param->port_id, param->type);
> +		ret = -EINVAL;
> +	}
> +
> +	return ret;
> +}
> +

Re: [Patch v6 06/18] net/mana: add device info

[fengchengwen]

On 2022/8/31 6:51, longli@linuxonhyperv.com wrote:
> From: Long Li <longli@microsoft.com>
> 
> Add the function to get device info.
> 
> Signed-off-by: Long Li <longli@microsoft.com>
> ---
>  doc/guides/nics/features/mana.ini |  1 +
>  drivers/net/mana/mana.c           | 82 +++++++++++++++++++++++++++++++
>  2 files changed, 83 insertions(+)
> 
> diff --git a/doc/guides/nics/features/mana.ini b/doc/guides/nics/features/mana.ini
> index 8043e11f99..566b3e8770 100644
> --- a/doc/guides/nics/features/mana.ini
> +++ b/doc/guides/nics/features/mana.ini
> @@ -8,5 +8,6 @@ Link status          = P
>  Linux                = Y
>  Multiprocess aware   = Y
>  Removal event        = Y
> +Speed capabilities   = P
>  Usage doc            = Y
>  x86-64               = Y
> diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c
> index c9591035ac..e1550b3c08 100644
> --- a/drivers/net/mana/mana.c
> +++ b/drivers/net/mana/mana.c
> @@ -116,6 +116,86 @@ mana_dev_close(struct rte_eth_dev *dev)
>  	return 0;
>  }
>  
> +static int mana_dev_info_get(struct rte_eth_dev *dev,
> +			     struct rte_eth_dev_info *dev_info)
> +{
> +	struct mana_priv *priv = dev->data->dev_private;
> +
> +	dev_info->max_mtu = RTE_ETHER_MTU;

...

> +	dev_info->tx_desc_lim.nb_max = priv->max_tx_desc;
> +	dev_info->tx_desc_lim.nb_align = MIN_BUFFERS_PER_QUEUE;
> +	dev_info->tx_desc_lim.nb_seg_max = priv->max_send_sge;
> +	dev_info->rx_desc_lim.nb_mtu_seg_max = priv->max_recv_sge;
> +
> +	/* Speed */
> +	dev_info->speed_capa = ETH_LINK_SPEED_100G;

I notice "[Patch v6 04/18] net/mana: add link update" report always 200G, why here is 100G?

> +
> +	/* RX params */
> +	dev_info->default_rxportconf.burst_size = 1;
> +	dev_info->default_rxportconf.ring_size = MAX_RECEIVE_BUFFERS_PER_QUEUE;
> +	dev_info->default_rxportconf.nb_queues = 1;
> +
> +	/* TX params */
> +	dev_info->default_txportconf.burst_size = 1;
> +	dev_info->default_txportconf.ring_size = MAX_SEND_BUFFERS_PER_QUEUE;
> +	dev_info->default_txportconf.nb_queues = 1;
> +
> +	return 0;
> +}
> +

...

Re: [Patch v6 15/18] net/mana: add function to send packets

[fengchengwen]

On 2022/8/31 6:51, longli@linuxonhyperv.com wrote:
> From: Long Li <longli@microsoft.com>
> 
> With all the TX queues created, MANA can send packets over those queues.
> 
> Signed-off-by: Long Li <longli@microsoft.com>

...

>  }
> +
> +uint16_t mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts,
> +		       uint16_t nb_pkts)
> +{
> +	struct mana_txq *txq = dpdk_txq;
> +	struct mana_priv *priv = txq->priv;
> +	struct gdma_comp comp;
> +	int ret;
> +	void *db_page;
> +
> +	/* Process send completions from GDMA */
> +	while (gdma_poll_completion_queue(&txq->gdma_cq, &comp) == 1) {
> +		struct mana_txq_desc *desc =
> +			&txq->desc_ring[txq->desc_ring_tail];
> +		struct mana_tx_comp_oob *oob =
> +			(struct mana_tx_comp_oob *)&comp.completion_data[0];
> +
> +		if (oob->cqe_hdr.cqe_type != CQE_TX_OKAY) {
> +			DRV_LOG(ERR,
> +				"mana_tx_comp_oob cqe_type %u vendor_err %u",
> +				oob->cqe_hdr.cqe_type, oob->cqe_hdr.vendor_err);
> +			txq->stats.errors++;
> +		} else {
> +			DRV_LOG(DEBUG, "mana_tx_comp_oob CQE_TX_OKAY");
> +			txq->stats.packets++;
> +		}
> +
> +		if (!desc->pkt) {
> +			DRV_LOG(ERR, "mana_txq_desc has a NULL pkt");
> +		} else {
> +			txq->stats.bytes += desc->pkt->data_len;
> +			rte_pktmbuf_free(desc->pkt);
> +		}
> +
> +		desc->pkt = NULL;
> +		txq->desc_ring_tail = (txq->desc_ring_tail + 1) % txq->num_desc;
> +		txq->gdma_sq.tail += desc->wqe_size_in_bu;
> +	}
> +
> +	/* Post send requests to GDMA */
> +	uint16_t pkt_idx;
> +
> +	for (pkt_idx = 0; pkt_idx < nb_pkts; pkt_idx++) {
> +		struct rte_mbuf *m_pkt = tx_pkts[pkt_idx];
> +		struct rte_mbuf *m_seg = m_pkt;
> +		struct transmit_oob_v2 tx_oob = {0};
> +		struct one_sgl sgl = {0};
> +
> +		/* Drop the packet if it exceeds max segments */
> +		if (m_pkt->nb_segs > priv->max_send_sge) {
> +			DRV_LOG(ERR, "send packet segments %d exceeding max",
> +				m_pkt->nb_segs);

This branch violate rte_eth_tx_burst definition, which defined the return value is " *   The maximum number of packets to transmit."

Also I notice the driver didn't implement tx-prepare, which could hold such checking in framework's definition.

> +			continue;
> +		}
> +
> +		/* Fill in the oob */
> +		tx_oob.short_oob.packet_format = short_packet_format;
> +		tx_oob.short_oob.tx_is_outer_ipv4 =
> +			m_pkt->ol_flags & RTE_MBUF_F_TX_IPV4 ? 1 : 0;
> +		tx_oob.short_oob.tx_is_outer_ipv6 =
> +			m_pkt->ol_flags & RTE_MBUF_F_TX_IPV6 ? 1 : 0;
> +
> +		tx_oob.short_oob.tx_compute_IP_header_checksum =
> +			m_pkt->ol_flags & RTE_MBUF_F_TX_IP_CKSUM ? 1 : 0;
> +
> +		if ((m_pkt->ol_flags & RTE_MBUF_F_TX_L4_MASK) ==
> +				RTE_MBUF_F_TX_TCP_CKSUM) {
> +			struct rte_tcp_hdr *tcp_hdr;
> +
> +			/* HW needs partial TCP checksum */
> +
> +			tcp_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> +					  struct rte_tcp_hdr *,
> +					  m_pkt->l2_len + m_pkt->l3_len);
> +
> +			if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV4) {
> +				struct rte_ipv4_hdr *ip_hdr;
> +
> +				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> +						struct rte_ipv4_hdr *,
> +						m_pkt->l2_len);
> +				tcp_hdr->cksum = rte_ipv4_phdr_cksum(ip_hdr,
> +							m_pkt->ol_flags);
> +
> +			} else if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV6) {
> +				struct rte_ipv6_hdr *ip_hdr;
> +
> +				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> +						struct rte_ipv6_hdr *,
> +						m_pkt->l2_len);
> +				tcp_hdr->cksum = rte_ipv6_phdr_cksum(ip_hdr,
> +							m_pkt->ol_flags);
> +			} else {
> +				DRV_LOG(ERR, "Invalid input for TCP CKSUM");
> +			}
> +
> +			tx_oob.short_oob.tx_compute_TCP_checksum = 1;
> +			tx_oob.short_oob.tx_transport_header_offset =
> +				m_pkt->l2_len + m_pkt->l3_len;
> +		}
> +
> +		if ((m_pkt->ol_flags & RTE_MBUF_F_TX_L4_MASK) ==
> +				RTE_MBUF_F_TX_UDP_CKSUM) {
> +			struct rte_udp_hdr *udp_hdr;
> +
> +			/* HW needs partial UDP checksum */
> +			udp_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> +					struct rte_udp_hdr *,
> +					m_pkt->l2_len + m_pkt->l3_len);
> +
> +			if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV4) {
> +				struct rte_ipv4_hdr *ip_hdr;
> +
> +				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> +						struct rte_ipv4_hdr *,
> +						m_pkt->l2_len);
> +
> +				udp_hdr->dgram_cksum =
> +					rte_ipv4_phdr_cksum(ip_hdr,
> +							    m_pkt->ol_flags);
> +
> +			} else if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV6) {
> +				struct rte_ipv6_hdr *ip_hdr;
> +
> +				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> +						struct rte_ipv6_hdr *,
> +						m_pkt->l2_len);
> +
> +				udp_hdr->dgram_cksum =
> +					rte_ipv6_phdr_cksum(ip_hdr,
> +							    m_pkt->ol_flags);
> +
> +			} else {
> +				DRV_LOG(ERR, "Invalid input for UDP CKSUM");
> +			}
> +
> +			tx_oob.short_oob.tx_compute_UDP_checksum = 1;
> +		}

Why not have TSO process, which was reported support in dev_info_get ?

> +
> +		tx_oob.short_oob.suppress_tx_CQE_generation = 0;
> +		tx_oob.short_oob.VCQ_number = txq->gdma_cq.id;
> +
> +		tx_oob.short_oob.VSQ_frame_num =
> +			get_vsq_frame_num(txq->gdma_sq.id);
> +		tx_oob.short_oob.short_vport_offset = txq->tx_vp_offset;
> +
> +		DRV_LOG(DEBUG, "tx_oob packet_format %u ipv4 %u ipv6 %u",
> +			tx_oob.short_oob.packet_format,
> +			tx_oob.short_oob.tx_is_outer_ipv4,
> +			tx_oob.short_oob.tx_is_outer_ipv6);
> +
> +		DRV_LOG(DEBUG, "tx_oob checksum ip %u tcp %u udp %u offset %u",
> +			tx_oob.short_oob.tx_compute_IP_header_checksum,
> +			tx_oob.short_oob.tx_compute_TCP_checksum,
> +			tx_oob.short_oob.tx_compute_UDP_checksum,
> +			tx_oob.short_oob.tx_transport_header_offset);
> +
> +		DRV_LOG(DEBUG, "pkt[%d]: buf_addr 0x%p, nb_segs %d, pkt_len %d",
> +			pkt_idx, m_pkt->buf_addr, m_pkt->nb_segs,
> +			m_pkt->pkt_len);
> +
> +		/* Create SGL for packet data buffers */
> +		for (uint16_t seg_idx = 0; seg_idx < m_pkt->nb_segs; seg_idx++) {
> +			struct mana_mr_cache *mr =
> +				mana_find_pmd_mr(&txq->mr_btree, priv, m_seg);
> +
> +			if (!mr) {
> +				DRV_LOG(ERR, "failed to get MR, pkt_idx %u",
> +					pkt_idx);
> +				return pkt_idx;
> +			}
> +
> +			sgl.gdma_sgl[seg_idx].address =
> +				rte_cpu_to_le_64(rte_pktmbuf_mtod(m_seg,
> +								  uint64_t));
> +			sgl.gdma_sgl[seg_idx].size = m_seg->data_len;
> +			sgl.gdma_sgl[seg_idx].memory_key = mr->lkey;
> +
> +			DRV_LOG(DEBUG,
> +				"seg idx %u addr 0x%" PRIx64 " size %x key %x",
> +				seg_idx, sgl.gdma_sgl[seg_idx].address,
> +				sgl.gdma_sgl[seg_idx].size,
> +				sgl.gdma_sgl[seg_idx].memory_key);
> +
> +			m_seg = m_seg->next;
> +		}
> +
> +		struct gdma_work_request work_req = {0};
> +		struct gdma_posted_wqe_info wqe_info = {0};
> +
> +		work_req.gdma_header.struct_size = sizeof(work_req);
> +		wqe_info.gdma_header.struct_size = sizeof(wqe_info);
> +
> +		work_req.sgl = sgl.gdma_sgl;
> +		work_req.num_sgl_elements = m_pkt->nb_segs;
> +		work_req.inline_oob_size_in_bytes =
> +			sizeof(struct transmit_short_oob_v2);
> +		work_req.inline_oob_data = &tx_oob;
> +		work_req.flags = 0;
> +		work_req.client_data_unit = NOT_USING_CLIENT_DATA_UNIT;
> +
> +		ret = gdma_post_work_request(&txq->gdma_sq, &work_req,
> +					     &wqe_info);
> +		if (!ret) {
> +			struct mana_txq_desc *desc =
> +				&txq->desc_ring[txq->desc_ring_head];
> +
> +			/* Update queue for tracking pending requests */
> +			desc->pkt = m_pkt;
> +			desc->wqe_size_in_bu = wqe_info.wqe_size_in_bu;
> +			txq->desc_ring_head =
> +				(txq->desc_ring_head + 1) % txq->num_desc;
> +
> +			DRV_LOG(DEBUG, "nb_pkts %u pkt[%d] sent",
> +				nb_pkts, pkt_idx);
> +		} else {
> +			DRV_LOG(INFO, "pkt[%d] failed to post send ret %d",
> +				pkt_idx, ret);
> +			break;
> +		}
> +	}
> +
> +	/* Ring hardware door bell */
> +	db_page = priv->db_page;
> +	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
> +		struct rte_eth_dev *dev =
> +			&rte_eth_devices[priv->dev_data->port_id];
> +		struct mana_process_priv *process_priv = dev->process_private;
> +
> +		db_page = process_priv->db_page;
> +	}
> +
> +	ret = mana_ring_doorbell(db_page, gdma_queue_send,
> +				 txq->gdma_sq.id,
> +				 txq->gdma_sq.head *
> +					GDMA_WQE_ALIGNMENT_UNIT_SIZE);
> +	if (ret)
> +		DRV_LOG(ERR, "mana_ring_doorbell failed ret %d", ret);
> +
> +	return pkt_idx;
> +}
> 

RE: [Patch v6 06/18] net/mana: add device info

[Long Li]

> Subject: Re: [Patch v6 06/18] net/mana: add device info
> 
> On 2022/8/31 6:51, longli@linuxonhyperv.com wrote:
> > From: Long Li <longli@microsoft.com>
> >
> > Add the function to get device info.
> >
> > Signed-off-by: Long Li <longli@microsoft.com>
> > ---
> >  doc/guides/nics/features/mana.ini |  1 +
> >  drivers/net/mana/mana.c           | 82 +++++++++++++++++++++++++++++++
> >  2 files changed, 83 insertions(+)
> >
> > diff --git a/doc/guides/nics/features/mana.ini
> > b/doc/guides/nics/features/mana.ini
> > index 8043e11f99..566b3e8770 100644
> > --- a/doc/guides/nics/features/mana.ini
> > +++ b/doc/guides/nics/features/mana.ini
> > @@ -8,5 +8,6 @@ Link status          = P
> >  Linux                = Y
> >  Multiprocess aware   = Y
> >  Removal event        = Y
> > +Speed capabilities   = P
> >  Usage doc            = Y
> >  x86-64               = Y
> > diff --git a/drivers/net/mana/mana.c b/drivers/net/mana/mana.c index
> > c9591035ac..e1550b3c08 100644
> > --- a/drivers/net/mana/mana.c
> > +++ b/drivers/net/mana/mana.c
> > @@ -116,6 +116,86 @@ mana_dev_close(struct rte_eth_dev *dev)
> >  	return 0;
> >  }
> >
> > +static int mana_dev_info_get(struct rte_eth_dev *dev,
> > +			     struct rte_eth_dev_info *dev_info) {
> > +	struct mana_priv *priv = dev->data->dev_private;
> > +
> > +	dev_info->max_mtu = RTE_ETHER_MTU;
> 
> ...
> 
> > +	dev_info->tx_desc_lim.nb_max = priv->max_tx_desc;
> > +	dev_info->tx_desc_lim.nb_align = MIN_BUFFERS_PER_QUEUE;
> > +	dev_info->tx_desc_lim.nb_seg_max = priv->max_send_sge;
> > +	dev_info->rx_desc_lim.nb_mtu_seg_max = priv->max_recv_sge;
> > +
> > +	/* Speed */
> > +	dev_info->speed_capa = ETH_LINK_SPEED_100G;
> 
> I notice "[Patch v6 04/18] net/mana: add link update" report always 200G, why
> here is 100G?

Thanks for pointing this out. I will fix patch 04.

> 
> > +
> > +	/* RX params */
> > +	dev_info->default_rxportconf.burst_size = 1;
> > +	dev_info->default_rxportconf.ring_size =
> MAX_RECEIVE_BUFFERS_PER_QUEUE;
> > +	dev_info->default_rxportconf.nb_queues = 1;
> > +
> > +	/* TX params */
> > +	dev_info->default_txportconf.burst_size = 1;
> > +	dev_info->default_txportconf.ring_size =
> MAX_SEND_BUFFERS_PER_QUEUE;
> > +	dev_info->default_txportconf.nb_queues = 1;
> > +
> > +	return 0;
> > +}
> > +
> 
> ...

RE: [Patch v6 15/18] net/mana: add function to send packets

[Long Li]

> Subject: Re: [Patch v6 15/18] net/mana: add function to send packets
> 
> On 2022/8/31 6:51, longli@linuxonhyperv.com wrote:
> > From: Long Li <longli@microsoft.com>
> >
> > With all the TX queues created, MANA can send packets over those queues.
> >
> > Signed-off-by: Long Li <longli@microsoft.com>
> 
> ...
> 
> >  }
> > +
> > +uint16_t mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts,
> > +		       uint16_t nb_pkts)
> > +{
> > +	struct mana_txq *txq = dpdk_txq;
> > +	struct mana_priv *priv = txq->priv;
> > +	struct gdma_comp comp;
> > +	int ret;
> > +	void *db_page;
> > +
> > +	/* Process send completions from GDMA */
> > +	while (gdma_poll_completion_queue(&txq->gdma_cq, &comp) == 1) {
> > +		struct mana_txq_desc *desc =
> > +			&txq->desc_ring[txq->desc_ring_tail];
> > +		struct mana_tx_comp_oob *oob =
> > +			(struct mana_tx_comp_oob
> *)&comp.completion_data[0];
> > +
> > +		if (oob->cqe_hdr.cqe_type != CQE_TX_OKAY) {
> > +			DRV_LOG(ERR,
> > +				"mana_tx_comp_oob cqe_type %u
> vendor_err %u",
> > +				oob->cqe_hdr.cqe_type, oob-
> >cqe_hdr.vendor_err);
> > +			txq->stats.errors++;
> > +		} else {
> > +			DRV_LOG(DEBUG, "mana_tx_comp_oob
> CQE_TX_OKAY");
> > +			txq->stats.packets++;
> > +		}
> > +
> > +		if (!desc->pkt) {
> > +			DRV_LOG(ERR, "mana_txq_desc has a NULL pkt");
> > +		} else {
> > +			txq->stats.bytes += desc->pkt->data_len;
> > +			rte_pktmbuf_free(desc->pkt);
> > +		}
> > +
> > +		desc->pkt = NULL;
> > +		txq->desc_ring_tail = (txq->desc_ring_tail + 1) % txq->num_desc;
> > +		txq->gdma_sq.tail += desc->wqe_size_in_bu;
> > +	}
> > +
> > +	/* Post send requests to GDMA */
> > +	uint16_t pkt_idx;
> > +
> > +	for (pkt_idx = 0; pkt_idx < nb_pkts; pkt_idx++) {
> > +		struct rte_mbuf *m_pkt = tx_pkts[pkt_idx];
> > +		struct rte_mbuf *m_seg = m_pkt;
> > +		struct transmit_oob_v2 tx_oob = {0};
> > +		struct one_sgl sgl = {0};
> > +
> > +		/* Drop the packet if it exceeds max segments */
> > +		if (m_pkt->nb_segs > priv->max_send_sge) {
> > +			DRV_LOG(ERR, "send packet segments %d exceeding
> max",
> > +				m_pkt->nb_segs);
> 
> This branch violate rte_eth_tx_burst definition, which defined the return value is
> " *   The maximum number of packets to transmit."

Will fix this.

> 
> Also I notice the driver didn't implement tx-prepare, which could hold such
> checking in framework's definition.
> 
> > +			continue;
> > +		}
> > +
> > +		/* Fill in the oob */
> > +		tx_oob.short_oob.packet_format = short_packet_format;
> > +		tx_oob.short_oob.tx_is_outer_ipv4 =
> > +			m_pkt->ol_flags & RTE_MBUF_F_TX_IPV4 ? 1 : 0;
> > +		tx_oob.short_oob.tx_is_outer_ipv6 =
> > +			m_pkt->ol_flags & RTE_MBUF_F_TX_IPV6 ? 1 : 0;
> > +
> > +		tx_oob.short_oob.tx_compute_IP_header_checksum =
> > +			m_pkt->ol_flags & RTE_MBUF_F_TX_IP_CKSUM ? 1 : 0;
> > +
> > +		if ((m_pkt->ol_flags & RTE_MBUF_F_TX_L4_MASK) ==
> > +				RTE_MBUF_F_TX_TCP_CKSUM) {
> > +			struct rte_tcp_hdr *tcp_hdr;
> > +
> > +			/* HW needs partial TCP checksum */
> > +
> > +			tcp_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> > +					  struct rte_tcp_hdr *,
> > +					  m_pkt->l2_len + m_pkt->l3_len);
> > +
> > +			if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV4) {
> > +				struct rte_ipv4_hdr *ip_hdr;
> > +
> > +				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> > +						struct rte_ipv4_hdr *,
> > +						m_pkt->l2_len);
> > +				tcp_hdr->cksum = rte_ipv4_phdr_cksum(ip_hdr,
> > +							m_pkt->ol_flags);
> > +
> > +			} else if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV6) {
> > +				struct rte_ipv6_hdr *ip_hdr;
> > +
> > +				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> > +						struct rte_ipv6_hdr *,
> > +						m_pkt->l2_len);
> > +				tcp_hdr->cksum = rte_ipv6_phdr_cksum(ip_hdr,
> > +							m_pkt->ol_flags);
> > +			} else {
> > +				DRV_LOG(ERR, "Invalid input for TCP CKSUM");
> > +			}
> > +
> > +			tx_oob.short_oob.tx_compute_TCP_checksum = 1;
> > +			tx_oob.short_oob.tx_transport_header_offset =
> > +				m_pkt->l2_len + m_pkt->l3_len;
> > +		}
> > +
> > +		if ((m_pkt->ol_flags & RTE_MBUF_F_TX_L4_MASK) ==
> > +				RTE_MBUF_F_TX_UDP_CKSUM) {
> > +			struct rte_udp_hdr *udp_hdr;
> > +
> > +			/* HW needs partial UDP checksum */
> > +			udp_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> > +					struct rte_udp_hdr *,
> > +					m_pkt->l2_len + m_pkt->l3_len);
> > +
> > +			if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV4) {
> > +				struct rte_ipv4_hdr *ip_hdr;
> > +
> > +				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> > +						struct rte_ipv4_hdr *,
> > +						m_pkt->l2_len);
> > +
> > +				udp_hdr->dgram_cksum =
> > +					rte_ipv4_phdr_cksum(ip_hdr,
> > +							    m_pkt->ol_flags);
> > +
> > +			} else if (m_pkt->ol_flags & RTE_MBUF_F_TX_IPV6) {
> > +				struct rte_ipv6_hdr *ip_hdr;
> > +
> > +				ip_hdr = rte_pktmbuf_mtod_offset(m_pkt,
> > +						struct rte_ipv6_hdr *,
> > +						m_pkt->l2_len);
> > +
> > +				udp_hdr->dgram_cksum =
> > +					rte_ipv6_phdr_cksum(ip_hdr,
> > +							    m_pkt->ol_flags);
> > +
> > +			} else {
> > +				DRV_LOG(ERR, "Invalid input for UDP CKSUM");
> > +			}
> > +
> > +			tx_oob.short_oob.tx_compute_UDP_checksum = 1;
> > +		}
> 
> Why not have TSO process, which was reported support in dev_info_get ?

TSO is currently not implemented. I'm changing BNIC_DEV_TX_OFFLOAD_SUPPORT to fix this.

> 
> > +
> > +		tx_oob.short_oob.suppress_tx_CQE_generation = 0;
> > +		tx_oob.short_oob.VCQ_number = txq->gdma_cq.id;
> > +
> > +		tx_oob.short_oob.VSQ_frame_num =
> > +			get_vsq_frame_num(txq->gdma_sq.id);
> > +		tx_oob.short_oob.short_vport_offset = txq->tx_vp_offset;
> > +
> > +		DRV_LOG(DEBUG, "tx_oob packet_format %u ipv4 %u ipv6 %u",
> > +			tx_oob.short_oob.packet_format,
> > +			tx_oob.short_oob.tx_is_outer_ipv4,
> > +			tx_oob.short_oob.tx_is_outer_ipv6);
> > +
> > +		DRV_LOG(DEBUG, "tx_oob checksum ip %u tcp %u udp %u
> offset %u",
> > +			tx_oob.short_oob.tx_compute_IP_header_checksum,
> > +			tx_oob.short_oob.tx_compute_TCP_checksum,
> > +			tx_oob.short_oob.tx_compute_UDP_checksum,
> > +			tx_oob.short_oob.tx_transport_header_offset);
> > +
> > +		DRV_LOG(DEBUG, "pkt[%d]: buf_addr 0x%p, nb_segs %d,
> pkt_len %d",
> > +			pkt_idx, m_pkt->buf_addr, m_pkt->nb_segs,
> > +			m_pkt->pkt_len);
> > +
> > +		/* Create SGL for packet data buffers */
> > +		for (uint16_t seg_idx = 0; seg_idx < m_pkt->nb_segs; seg_idx++)
> {
> > +			struct mana_mr_cache *mr =
> > +				mana_find_pmd_mr(&txq->mr_btree, priv,
> m_seg);
> > +
> > +			if (!mr) {
> > +				DRV_LOG(ERR, "failed to get MR, pkt_idx %u",
> > +					pkt_idx);
> > +				return pkt_idx;
> > +			}
> > +
> > +			sgl.gdma_sgl[seg_idx].address =
> > +				rte_cpu_to_le_64(rte_pktmbuf_mtod(m_seg,
> > +								  uint64_t));
> > +			sgl.gdma_sgl[seg_idx].size = m_seg->data_len;
> > +			sgl.gdma_sgl[seg_idx].memory_key = mr->lkey;
> > +
> > +			DRV_LOG(DEBUG,
> > +				"seg idx %u addr 0x%" PRIx64 " size %x key %x",
> > +				seg_idx, sgl.gdma_sgl[seg_idx].address,
> > +				sgl.gdma_sgl[seg_idx].size,
> > +				sgl.gdma_sgl[seg_idx].memory_key);
> > +
> > +			m_seg = m_seg->next;
> > +		}
> > +
> > +		struct gdma_work_request work_req = {0};
> > +		struct gdma_posted_wqe_info wqe_info = {0};
> > +
> > +		work_req.gdma_header.struct_size = sizeof(work_req);
> > +		wqe_info.gdma_header.struct_size = sizeof(wqe_info);
> > +
> > +		work_req.sgl = sgl.gdma_sgl;
> > +		work_req.num_sgl_elements = m_pkt->nb_segs;
> > +		work_req.inline_oob_size_in_bytes =
> > +			sizeof(struct transmit_short_oob_v2);
> > +		work_req.inline_oob_data = &tx_oob;
> > +		work_req.flags = 0;
> > +		work_req.client_data_unit = NOT_USING_CLIENT_DATA_UNIT;
> > +
> > +		ret = gdma_post_work_request(&txq->gdma_sq, &work_req,
> > +					     &wqe_info);
> > +		if (!ret) {
> > +			struct mana_txq_desc *desc =
> > +				&txq->desc_ring[txq->desc_ring_head];
> > +
> > +			/* Update queue for tracking pending requests */
> > +			desc->pkt = m_pkt;
> > +			desc->wqe_size_in_bu = wqe_info.wqe_size_in_bu;
> > +			txq->desc_ring_head =
> > +				(txq->desc_ring_head + 1) % txq->num_desc;
> > +
> > +			DRV_LOG(DEBUG, "nb_pkts %u pkt[%d] sent",
> > +				nb_pkts, pkt_idx);
> > +		} else {
> > +			DRV_LOG(INFO, "pkt[%d] failed to post send ret %d",
> > +				pkt_idx, ret);
> > +			break;
> > +		}
> > +	}
> > +
> > +	/* Ring hardware door bell */
> > +	db_page = priv->db_page;
> > +	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
> > +		struct rte_eth_dev *dev =
> > +			&rte_eth_devices[priv->dev_data->port_id];
> > +		struct mana_process_priv *process_priv = dev->process_private;
> > +
> > +		db_page = process_priv->db_page;
> > +	}
> > +
> > +	ret = mana_ring_doorbell(db_page, gdma_queue_send,
> > +				 txq->gdma_sq.id,
> > +				 txq->gdma_sq.head *
> > +					GDMA_WQE_ALIGNMENT_UNIT_SIZE);
> > +	if (ret)
> > +		DRV_LOG(ERR, "mana_ring_doorbell failed ret %d", ret);
> > +
> > +	return pkt_idx;
> > +}
> >

RE: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[Long Li]

> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build environment and
> doc
> 
> On 2022/8/31 6:51, longli@linuxonhyperv.com wrote:
> > From: Long Li <longli@microsoft.com>
> >
> > MANA is a PCI device. It uses IB verbs to access hardware through the
> > kernel RDMA layer. This patch introduces build environment and basic
> > device probe functions.
> >
> > Signed-off-by: Long Li <longli@microsoft.com>
> 
> ...
> 
> > +static int mana_mp_primary_handle(const struct rte_mp_msg *mp_msg,
> > +				  const void *peer)
> > +{
> > +	struct rte_eth_dev *dev;
> > +	const struct mana_mp_param *param =
> > +		(const struct mana_mp_param *)mp_msg->param;
> > +	struct rte_mp_msg mp_res = { 0 };
> > +	struct mana_mp_param *res = (struct mana_mp_param
> *)mp_res.param;
> > +	int ret;
> > +	struct mana_priv *priv;
> > +
> > +	if (!rte_eth_dev_is_valid_port(param->port_id)) {
> > +		DRV_LOG(ERR, "MP handle port ID %u invalid", param->port_id);
> > +		return -ENODEV;
> > +	}
> > +
> > +	dev = &rte_eth_devices[param->port_id];
> > +	priv = dev->data->dev_private;
> > +
> > +	mp_init_msg(&mp_res, param->type, param->port_id);
> > +
> > +	switch (param->type) {
> > +	case MANA_MP_REQ_VERBS_CMD_FD:
> > +		mp_res.num_fds = 1;
> > +		mp_res.fds[0] = priv->ib_ctx->cmd_fd;
> 
> The cmd_fd is system level handler?
> 
> If it's process private handler, it should not used directly in another process.

According to rte_mp_xxx semantics, the file handle is duplicated to another process. It's not directly used. It's required for secondary process to map to the same doorbell pages.

> 
> > +		res->result = 0;
> > +		ret = rte_mp_reply(&mp_res, peer);
> > +		break;
> > +
> > +	default:
> > +		DRV_LOG(ERR, "Port %u unknown primary MP type %u",
> > +			param->port_id, param->type);
> > +		ret = -EINVAL;
> > +	}
> > +
> > +	return ret;
> > +}
> > +

Re: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[fengchengwen]

On 2022/9/3 3:45, Long Li wrote:
>> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build environment and
>> doc
>>
>> On 2022/8/31 6:51, longli@linuxonhyperv.com wrote:
>>> From: Long Li <longli@microsoft.com>
>>>
>>> MANA is a PCI device. It uses IB verbs to access hardware through the
>>> kernel RDMA layer. This patch introduces build environment and basic
>>> device probe functions.
>>>
>>> Signed-off-by: Long Li <longli@microsoft.com>
>> ...
>>
>>> +static int mana_mp_primary_handle(const struct rte_mp_msg *mp_msg,
>>> +				  const void *peer)
>>> +{
>>> +	struct rte_eth_dev *dev;
>>> +	const struct mana_mp_param *param =
>>> +		(const struct mana_mp_param *)mp_msg->param;
>>> +	struct rte_mp_msg mp_res = { 0 };
>>> +	struct mana_mp_param *res = (struct mana_mp_param
>> *)mp_res.param;
>>> +	int ret;
>>> +	struct mana_priv *priv;
>>> +
>>> +	if (!rte_eth_dev_is_valid_port(param->port_id)) {
>>> +		DRV_LOG(ERR, "MP handle port ID %u invalid", param->port_id);
>>> +		return -ENODEV;
>>> +	}
>>> +
>>> +	dev = &rte_eth_devices[param->port_id];
>>> +	priv = dev->data->dev_private;
>>> +
>>> +	mp_init_msg(&mp_res, param->type, param->port_id);
>>> +
>>> +	switch (param->type) {
>>> +	case MANA_MP_REQ_VERBS_CMD_FD:
>>> +		mp_res.num_fds = 1;
>>> +		mp_res.fds[0] = priv->ib_ctx->cmd_fd;
>> The cmd_fd is system level handler?
>>
>> If it's process private handler, it should not used directly in another process.
> According to rte_mp_xxx semantics, the file handle is duplicated to another process. It's not directly used. It's required for secondary process to map to the same doorbell pages.


You are right. I notice the rte_mp_xx use AF_UNIX which could process it.

Thanks for clarify.


>
>>> +		res->result = 0;
>>> +		ret = rte_mp_reply(&mp_res, peer);
>>> +		break;
>>> +
>>> +	default:
>>> +		DRV_LOG(ERR, "Port %u unknown primary MP type %u",
>>> +			param->port_id, param->type);
>>> +		ret = -EINVAL;
>>> +	}
>>> +
>>> +	return ret;
>>> +}
>>> +

Re: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[lihuisong (C)]

在 2022/9/1 2:05, Long Li 写道:
>> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build environment
>> and doc
>>
>>
>> 在 2022/8/31 6:51, longli@linuxonhyperv.com 写道:
>>> From: Long Li <longli@microsoft.com>
>>>
>>> MANA is a PCI device. It uses IB verbs to access hardware through the
>>> kernel RDMA layer. This patch introduces build environment and basic
>>> device probe functions.
>>>
>>> Signed-off-by: Long Li <longli@microsoft.com>
>>> ---
>>> Change log:
>>> v2:
>>> Fix typos.
>>> Make the driver build only on x86-64 and Linux.
>>> Remove unused header files.
>>> Change port definition to uint16_t or uint8_t (for IB).
>>> Use getline() in place of fgets() to read and truncate a line.
>>> v3:
>>> Add meson build check for required functions from RDMA direct verb
>>> header file
>>> v4:
>>> Remove extra "\n" in logging code.
>>> Use "r" in place of "rb" in fopen() to read text files.
>>>
>>> [snip]
>>> +
>>> +static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv
>> __rte_unused,
>>> +			      struct rte_pci_device *pci_dev,
>>> +			      struct rte_ether_addr *mac_addr) {
>>> +	struct ibv_device **ibv_list;
>>> +	int ibv_idx;
>>> +	struct ibv_context *ctx;
>>> +	struct ibv_device_attr_ex dev_attr;
>>> +	int num_devices;
>>> +	int ret = 0;
>>> +	uint8_t port;
>>> +	struct mana_priv *priv = NULL;
>>> +	struct rte_eth_dev *eth_dev = NULL;
>>> +	bool found_port;
>>> +
>>> +	ibv_list = ibv_get_device_list(&num_devices);
>>> +	for (ibv_idx = 0; ibv_idx < num_devices; ibv_idx++) {
>>> +		struct ibv_device *ibdev = ibv_list[ibv_idx];
>>> +		struct rte_pci_addr pci_addr;
>>> +
>>> +		DRV_LOG(INFO, "Probe device name %s dev_name %s
>> ibdev_path %s",
>>> +			ibdev->name, ibdev->dev_name, ibdev-
>>> ibdev_path);
>>> +
>>> +		if (mana_ibv_device_to_pci_addr(ibdev, &pci_addr))
>>> +			continue;
>>> +
>>> +		/* Ignore if this IB device is not this PCI device */
>>> +		if (pci_dev->addr.domain != pci_addr.domain ||
>>> +		    pci_dev->addr.bus != pci_addr.bus ||
>>> +		    pci_dev->addr.devid != pci_addr.devid ||
>>> +		    pci_dev->addr.function != pci_addr.function)
>>> +			continue;
>>> +
>>> +		ctx = ibv_open_device(ibdev);
>>> +		if (!ctx) {
>>> +			DRV_LOG(ERR, "Failed to open IB device %s",
>>> +				ibdev->name);
>>> +			continue;
>>> +		}
>>> +
>>> +		ret = ibv_query_device_ex(ctx, NULL, &dev_attr);
>>> +		DRV_LOG(INFO, "dev_attr.orig_attr.phys_port_cnt %u",
>>> +			dev_attr.orig_attr.phys_port_cnt);
>>> +		found_port = false;
>>> +
>>> +		for (port = 1; port <= dev_attr.orig_attr.phys_port_cnt;
>>> +		     port++) {
>>> +			struct ibv_parent_domain_init_attr attr = {};
>>> +			struct rte_ether_addr addr;
>>> +			char address[64];
>>> +			char name[RTE_ETH_NAME_MAX_LEN];
>>> +
>>> +			ret = get_port_mac(ibdev, port, &addr);
>>> +			if (ret)
>>> +				continue;
>>> +
>>> +			if (mac_addr && !rte_is_same_ether_addr(&addr,
>> mac_addr))
>>> +				continue;
>>> +
>>> +			rte_ether_format_addr(address, sizeof(address),
>> &addr);
>>> +			DRV_LOG(INFO, "device located port %u address
>> %s",
>>> +				port, address);
>>> +			found_port = true;
>>> +
>>> +			priv = rte_zmalloc_socket(NULL, sizeof(*priv),
>>> +						  RTE_CACHE_LINE_SIZE,
>>> +						  SOCKET_ID_ANY);
>>> +			if (!priv) {
>>> +				ret = -ENOMEM;
>>> +				goto failed;
>>> +			}
>>> +
>>> +			snprintf(name, sizeof(name), "%s_port%d",
>>> +				 pci_dev->device.name, port);
>>> +
>>> +			if (rte_eal_process_type() ==
>> RTE_PROC_SECONDARY) {
>>> +				int fd;
>>> +
>>> +				eth_dev =
>> rte_eth_dev_attach_secondary(name);
>>> +				if (!eth_dev) {
>>> +					DRV_LOG(ERR, "Can't attach to dev
>> %s",
>>> +						name);
>>> +					ret = -ENOMEM;
>>> +					goto failed;
>>> +				}
>>> +
>>> +				eth_dev->device = &pci_dev->device;
>>> +				eth_dev->dev_ops = &mana_dev_sec_ops;
>>> +				ret = mana_proc_priv_init(eth_dev);
>>> +				if (ret)
>>> +					goto failed;
>>> +				priv->process_priv = eth_dev-
>>> process_private;
>>> +
>>> +				/* Get the IB FD from the primary process */
>>> +				fd =
>> mana_mp_req_verbs_cmd_fd(eth_dev);
>>> +				if (fd < 0) {
>>> +					DRV_LOG(ERR, "Failed to get FD %d",
>> fd);
>>> +					ret = -ENODEV;
>>> +					goto failed;
>>> +				}
>>> +
>>> +				ret =
>> mana_map_doorbell_secondary(eth_dev, fd);
>>> +				if (ret) {
>>> +					DRV_LOG(ERR, "Failed secondary
>> map %d",
>>> +						fd);
>>> +					goto failed;
>>> +				}
>>> +
>>> +				/* fd is no not used after mapping doorbell */
>>> +				close(fd);
>>> +
>>> +				rte_spinlock_lock(&mana_shared_data-
>>> lock);
>>> +				mana_shared_data->secondary_cnt++;
>>> +				mana_local_data.secondary_cnt++;
>>> +				rte_spinlock_unlock(&mana_shared_data-
>>> lock);
>>> +
>>> +				rte_eth_copy_pci_info(eth_dev, pci_dev);
>>> +				rte_eth_dev_probing_finish(eth_dev);
>>> +
>>> +				/* Impossible to have more than one port
>>> +				 * matching a MAC address
>>> +				 */
>>> +				continue;
>>> +			}
>>> +
>>> +			eth_dev = rte_eth_dev_allocate(name);
>>> +			if (!eth_dev) {
>>> +				ret = -ENOMEM;
>>> +				goto failed;
>>> +			}
>>> +
>>> +			eth_dev->data->mac_addrs =
>>> +				rte_calloc("mana_mac", 1,
>>> +					   sizeof(struct rte_ether_addr), 0);
>>> +			if (!eth_dev->data->mac_addrs) {
>>> +				ret = -ENOMEM;
>>> +				goto failed;
>>> +			}
>>> +
>>> +			rte_ether_addr_copy(&addr, eth_dev->data-
>>> mac_addrs);
>>> +
>>> +			priv->ib_pd = ibv_alloc_pd(ctx);
>>> +			if (!priv->ib_pd) {
>>> +				DRV_LOG(ERR, "ibv_alloc_pd failed port %d",
>> port);
>>> +				ret = -ENOMEM;
>>> +				goto failed;
>>> +			}
>>> +
>>> +			/* Create a parent domain with the port number */
>>> +			attr.pd = priv->ib_pd;
>>> +			attr.comp_mask =
>> IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT;
>>> +			attr.pd_context = (void *)(uint64_t)port;
>>> +			priv->ib_parent_pd = ibv_alloc_parent_domain(ctx,
>> &attr);
>>> +			if (!priv->ib_parent_pd) {
>>> +				DRV_LOG(ERR,
>>> +					"ibv_alloc_parent_domain failed port
>> %d",
>>> +					port);
>>> +				ret = -ENOMEM;
>>> +				goto failed;
>>> +			}
>>> +
>>> +			priv->ib_ctx = ctx;
>>> +			priv->port_id = eth_dev->data->port_id;
>>> +			priv->dev_port = port;
>>> +			eth_dev->data->dev_private = priv;
>>> +			priv->dev_data = eth_dev->data;
>>> +
>>> +			priv->max_rx_queues = dev_attr.orig_attr.max_qp;
>>> +			priv->max_tx_queues = dev_attr.orig_attr.max_qp;
>>> +
>>> +			priv->max_rx_desc =
>>> +				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
>>> +					dev_attr.orig_attr.max_cqe);
>>> +			priv->max_tx_desc =
>>> +				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
>>> +					dev_attr.orig_attr.max_cqe);
>>> +
>>> +			priv->max_send_sge = dev_attr.orig_attr.max_sge;
>>> +			priv->max_recv_sge = dev_attr.orig_attr.max_sge;
>>> +
>>> +			priv->max_mr = dev_attr.orig_attr.max_mr;
>>> +			priv->max_mr_size =
>> dev_attr.orig_attr.max_mr_size;
>>> +
>>> +			DRV_LOG(INFO, "dev %s max queues %d desc %d
>> sge %d",
>>> +				name, priv->max_rx_queues, priv-
>>> max_rx_desc,
>>> +				priv->max_send_sge);
>>> +
>>> +			rte_spinlock_lock(&mana_shared_data->lock);
>>> +			mana_shared_data->primary_cnt++;
>>> +			rte_spinlock_unlock(&mana_shared_data->lock);
>>> +
>>> +			eth_dev->data->dev_flags |=
>> RTE_ETH_DEV_INTR_RMV;
>>> +
>>> +			eth_dev->device = &pci_dev->device;
>>> +			eth_dev->data->dev_flags |=
>>> +				RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
>>> +
>> Please do not use the temporary macro. Please review this patch:
>>
>> f30e69b41f94 ("ethdev: add device flag to bypass auto-filled queue xstats")
>>
>> This patch requires that per queue statistics are filled in
>> .xstats_get() by PMD.
> Thanks for pointing this out.
>
> It seems some PMDs are still depending on this flag for xstats.
>
> MANA doesn't implement xstats_get() currently, this flag is useful. Is it okay to keep using this flag before it's finally the time to remove it from all PMDs, or when MANA implements xstats?
Yes, your xstats doesn't implement now. Per queue stats should be filled 
in xstats API,
and the stats API cannot see per queue stats, so stats API in driver 
shouldn't fill
it(suggest that delete it from patch 17/18).

I guess this flag can be removed if PMD does not support xstats.
>
>

RE: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[Long Li]

> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build environment
> and doc
> 
> 
> 在 2022/9/1 2:05, Long Li 写道:
> >> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build
> >> environment and doc
> >>
> >>
> >> 在 2022/8/31 6:51, longli@linuxonhyperv.com 写道:
> >>> From: Long Li <longli@microsoft.com>
> >>>
> >>> MANA is a PCI device. It uses IB verbs to access hardware through
> >>> the kernel RDMA layer. This patch introduces build environment and
> >>> basic device probe functions.
> >>>
> >>> Signed-off-by: Long Li <longli@microsoft.com>
> >>> ---
> >>> Change log:
> >>> v2:
> >>> Fix typos.
> >>> Make the driver build only on x86-64 and Linux.
> >>> Remove unused header files.
> >>> Change port definition to uint16_t or uint8_t (for IB).
> >>> Use getline() in place of fgets() to read and truncate a line.
> >>> v3:
> >>> Add meson build check for required functions from RDMA direct verb
> >>> header file
> >>> v4:
> >>> Remove extra "\n" in logging code.
> >>> Use "r" in place of "rb" in fopen() to read text files.
> >>>
> >>> [snip]
> >>> +
> >>> +static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv
> >> __rte_unused,
> >>> +			      struct rte_pci_device *pci_dev,
> >>> +			      struct rte_ether_addr *mac_addr) {
> >>> +	struct ibv_device **ibv_list;
> >>> +	int ibv_idx;
> >>> +	struct ibv_context *ctx;
> >>> +	struct ibv_device_attr_ex dev_attr;
> >>> +	int num_devices;
> >>> +	int ret = 0;
> >>> +	uint8_t port;
> >>> +	struct mana_priv *priv = NULL;
> >>> +	struct rte_eth_dev *eth_dev = NULL;
> >>> +	bool found_port;
> >>> +
> >>> +	ibv_list = ibv_get_device_list(&num_devices);
> >>> +	for (ibv_idx = 0; ibv_idx < num_devices; ibv_idx++) {
> >>> +		struct ibv_device *ibdev = ibv_list[ibv_idx];
> >>> +		struct rte_pci_addr pci_addr;
> >>> +
> >>> +		DRV_LOG(INFO, "Probe device name %s dev_name %s
> >> ibdev_path %s",
> >>> +			ibdev->name, ibdev->dev_name, ibdev-
> >>> ibdev_path);
> >>> +
> >>> +		if (mana_ibv_device_to_pci_addr(ibdev, &pci_addr))
> >>> +			continue;
> >>> +
> >>> +		/* Ignore if this IB device is not this PCI device */
> >>> +		if (pci_dev->addr.domain != pci_addr.domain ||
> >>> +		    pci_dev->addr.bus != pci_addr.bus ||
> >>> +		    pci_dev->addr.devid != pci_addr.devid ||
> >>> +		    pci_dev->addr.function != pci_addr.function)
> >>> +			continue;
> >>> +
> >>> +		ctx = ibv_open_device(ibdev);
> >>> +		if (!ctx) {
> >>> +			DRV_LOG(ERR, "Failed to open IB device %s",
> >>> +				ibdev->name);
> >>> +			continue;
> >>> +		}
> >>> +
> >>> +		ret = ibv_query_device_ex(ctx, NULL, &dev_attr);
> >>> +		DRV_LOG(INFO, "dev_attr.orig_attr.phys_port_cnt %u",
> >>> +			dev_attr.orig_attr.phys_port_cnt);
> >>> +		found_port = false;
> >>> +
> >>> +		for (port = 1; port <= dev_attr.orig_attr.phys_port_cnt;
> >>> +		     port++) {
> >>> +			struct ibv_parent_domain_init_attr attr = {};
> >>> +			struct rte_ether_addr addr;
> >>> +			char address[64];
> >>> +			char name[RTE_ETH_NAME_MAX_LEN];
> >>> +
> >>> +			ret = get_port_mac(ibdev, port, &addr);
> >>> +			if (ret)
> >>> +				continue;
> >>> +
> >>> +			if (mac_addr && !rte_is_same_ether_addr(&addr,
> >> mac_addr))
> >>> +				continue;
> >>> +
> >>> +			rte_ether_format_addr(address, sizeof(address),
> >> &addr);
> >>> +			DRV_LOG(INFO, "device located port %u address
> >> %s",
> >>> +				port, address);
> >>> +			found_port = true;
> >>> +
> >>> +			priv = rte_zmalloc_socket(NULL, sizeof(*priv),
> >>> +						  RTE_CACHE_LINE_SIZE,
> >>> +						  SOCKET_ID_ANY);
> >>> +			if (!priv) {
> >>> +				ret = -ENOMEM;
> >>> +				goto failed;
> >>> +			}
> >>> +
> >>> +			snprintf(name, sizeof(name), "%s_port%d",
> >>> +				 pci_dev->device.name, port);
> >>> +
> >>> +			if (rte_eal_process_type() ==
> >> RTE_PROC_SECONDARY) {
> >>> +				int fd;
> >>> +
> >>> +				eth_dev =
> >> rte_eth_dev_attach_secondary(name);
> >>> +				if (!eth_dev) {
> >>> +					DRV_LOG(ERR, "Can't attach to dev
> >> %s",
> >>> +						name);
> >>> +					ret = -ENOMEM;
> >>> +					goto failed;
> >>> +				}
> >>> +
> >>> +				eth_dev->device = &pci_dev->device;
> >>> +				eth_dev->dev_ops = &mana_dev_sec_ops;
> >>> +				ret = mana_proc_priv_init(eth_dev);
> >>> +				if (ret)
> >>> +					goto failed;
> >>> +				priv->process_priv = eth_dev-
> >>> process_private;
> >>> +
> >>> +				/* Get the IB FD from the primary process */
> >>> +				fd =
> >> mana_mp_req_verbs_cmd_fd(eth_dev);
> >>> +				if (fd < 0) {
> >>> +					DRV_LOG(ERR, "Failed to get FD %d",
> >> fd);
> >>> +					ret = -ENODEV;
> >>> +					goto failed;
> >>> +				}
> >>> +
> >>> +				ret =
> >> mana_map_doorbell_secondary(eth_dev, fd);
> >>> +				if (ret) {
> >>> +					DRV_LOG(ERR, "Failed secondary
> >> map %d",
> >>> +						fd);
> >>> +					goto failed;
> >>> +				}
> >>> +
> >>> +				/* fd is no not used after mapping doorbell */
> >>> +				close(fd);
> >>> +
> >>> +				rte_spinlock_lock(&mana_shared_data-
> >>> lock);
> >>> +				mana_shared_data->secondary_cnt++;
> >>> +				mana_local_data.secondary_cnt++;
> >>> +				rte_spinlock_unlock(&mana_shared_data-
> >>> lock);
> >>> +
> >>> +				rte_eth_copy_pci_info(eth_dev, pci_dev);
> >>> +				rte_eth_dev_probing_finish(eth_dev);
> >>> +
> >>> +				/* Impossible to have more than one port
> >>> +				 * matching a MAC address
> >>> +				 */
> >>> +				continue;
> >>> +			}
> >>> +
> >>> +			eth_dev = rte_eth_dev_allocate(name);
> >>> +			if (!eth_dev) {
> >>> +				ret = -ENOMEM;
> >>> +				goto failed;
> >>> +			}
> >>> +
> >>> +			eth_dev->data->mac_addrs =
> >>> +				rte_calloc("mana_mac", 1,
> >>> +					   sizeof(struct rte_ether_addr), 0);
> >>> +			if (!eth_dev->data->mac_addrs) {
> >>> +				ret = -ENOMEM;
> >>> +				goto failed;
> >>> +			}
> >>> +
> >>> +			rte_ether_addr_copy(&addr, eth_dev->data-
> >>> mac_addrs);
> >>> +
> >>> +			priv->ib_pd = ibv_alloc_pd(ctx);
> >>> +			if (!priv->ib_pd) {
> >>> +				DRV_LOG(ERR, "ibv_alloc_pd failed port %d",
> >> port);
> >>> +				ret = -ENOMEM;
> >>> +				goto failed;
> >>> +			}
> >>> +
> >>> +			/* Create a parent domain with the port number */
> >>> +			attr.pd = priv->ib_pd;
> >>> +			attr.comp_mask =
> >> IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT;
> >>> +			attr.pd_context = (void *)(uint64_t)port;
> >>> +			priv->ib_parent_pd = ibv_alloc_parent_domain(ctx,
> >> &attr);
> >>> +			if (!priv->ib_parent_pd) {
> >>> +				DRV_LOG(ERR,
> >>> +					"ibv_alloc_parent_domain failed port
> >> %d",
> >>> +					port);
> >>> +				ret = -ENOMEM;
> >>> +				goto failed;
> >>> +			}
> >>> +
> >>> +			priv->ib_ctx = ctx;
> >>> +			priv->port_id = eth_dev->data->port_id;
> >>> +			priv->dev_port = port;
> >>> +			eth_dev->data->dev_private = priv;
> >>> +			priv->dev_data = eth_dev->data;
> >>> +
> >>> +			priv->max_rx_queues = dev_attr.orig_attr.max_qp;
> >>> +			priv->max_tx_queues = dev_attr.orig_attr.max_qp;
> >>> +
> >>> +			priv->max_rx_desc =
> >>> +				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
> >>> +					dev_attr.orig_attr.max_cqe);
> >>> +			priv->max_tx_desc =
> >>> +				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
> >>> +					dev_attr.orig_attr.max_cqe);
> >>> +
> >>> +			priv->max_send_sge = dev_attr.orig_attr.max_sge;
> >>> +			priv->max_recv_sge = dev_attr.orig_attr.max_sge;
> >>> +
> >>> +			priv->max_mr = dev_attr.orig_attr.max_mr;
> >>> +			priv->max_mr_size =
> >> dev_attr.orig_attr.max_mr_size;
> >>> +
> >>> +			DRV_LOG(INFO, "dev %s max queues %d desc %d
> >> sge %d",
> >>> +				name, priv->max_rx_queues, priv-
> >>> max_rx_desc,
> >>> +				priv->max_send_sge);
> >>> +
> >>> +			rte_spinlock_lock(&mana_shared_data->lock);
> >>> +			mana_shared_data->primary_cnt++;
> >>> +			rte_spinlock_unlock(&mana_shared_data->lock);
> >>> +
> >>> +			eth_dev->data->dev_flags |=
> >> RTE_ETH_DEV_INTR_RMV;
> >>> +
> >>> +			eth_dev->device = &pci_dev->device;
> >>> +			eth_dev->data->dev_flags |=
> >>> +				RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
> >>> +
> >> Please do not use the temporary macro. Please review this patch:
> >>
> >> f30e69b41f94 ("ethdev: add device flag to bypass auto-filled queue
> >> xstats")
> >>
> >> This patch requires that per queue statistics are filled in
> >> .xstats_get() by PMD.
> > Thanks for pointing this out.
> >
> > It seems some PMDs are still depending on this flag for xstats.
> >
> > MANA doesn't implement xstats_get() currently, this flag is useful. Is it
> okay to keep using this flag before it's finally the time to remove it from all
> PMDs, or when MANA implements xstats?
> Yes, your xstats doesn't implement now. Per queue stats should be filled in
> xstats API, and the stats API cannot see per queue stats, so stats API in driver
> shouldn't fill it(suggest that delete it from patch 17/18).
> 
> I guess this flag can be removed if PMD does not support xstats.
> >
> >

I don't understand your suggestion. An application can call rte_eth_stats_get() to get port stats, and this will call into stats_get() in the driver, as implemented in patch 17/18.

When flag RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS  is set, an application can also use rte_eth_xstats_get() to get port stats even the driver doesn't implement xstats_get().

Re: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[lihuisong (C)]

在 2022/9/7 9:36, Long Li 写道:
>> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build environment
>> and doc
>>
>>
>> 在 2022/9/1 2:05, Long Li 写道:
>>>> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build
>>>> environment and doc
>>>>
>>>>
>>>> 在 2022/8/31 6:51, longli@linuxonhyperv.com 写道:
>>>>> From: Long Li <longli@microsoft.com>
>>>>>
>>>>> MANA is a PCI device. It uses IB verbs to access hardware through
>>>>> the kernel RDMA layer. This patch introduces build environment and
>>>>> basic device probe functions.
>>>>>
>>>>> Signed-off-by: Long Li <longli@microsoft.com>
>>>>> ---
>>>>> Change log:
>>>>> v2:
>>>>> Fix typos.
>>>>> Make the driver build only on x86-64 and Linux.
>>>>> Remove unused header files.
>>>>> Change port definition to uint16_t or uint8_t (for IB).
>>>>> Use getline() in place of fgets() to read and truncate a line.
>>>>> v3:
>>>>> Add meson build check for required functions from RDMA direct verb
>>>>> header file
>>>>> v4:
>>>>> Remove extra "\n" in logging code.
>>>>> Use "r" in place of "rb" in fopen() to read text files.
>>>>>
>>>>> [snip]
>>>>> +
>>>>> +static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv
>>>> __rte_unused,
>>>>> +			      struct rte_pci_device *pci_dev,
>>>>> +			      struct rte_ether_addr *mac_addr) {
>>>>> +	struct ibv_device **ibv_list;
>>>>> +	int ibv_idx;
>>>>> +	struct ibv_context *ctx;
>>>>> +	struct ibv_device_attr_ex dev_attr;
>>>>> +	int num_devices;
>>>>> +	int ret = 0;
>>>>> +	uint8_t port;
>>>>> +	struct mana_priv *priv = NULL;
>>>>> +	struct rte_eth_dev *eth_dev = NULL;
>>>>> +	bool found_port;
>>>>> +
>>>>> +	ibv_list = ibv_get_device_list(&num_devices);
>>>>> +	for (ibv_idx = 0; ibv_idx < num_devices; ibv_idx++) {
>>>>> +		struct ibv_device *ibdev = ibv_list[ibv_idx];
>>>>> +		struct rte_pci_addr pci_addr;
>>>>> +
>>>>> +		DRV_LOG(INFO, "Probe device name %s dev_name %s
>>>> ibdev_path %s",
>>>>> +			ibdev->name, ibdev->dev_name, ibdev-
>>>>> ibdev_path);
>>>>> +
>>>>> +		if (mana_ibv_device_to_pci_addr(ibdev, &pci_addr))
>>>>> +			continue;
>>>>> +
>>>>> +		/* Ignore if this IB device is not this PCI device */
>>>>> +		if (pci_dev->addr.domain != pci_addr.domain ||
>>>>> +		    pci_dev->addr.bus != pci_addr.bus ||
>>>>> +		    pci_dev->addr.devid != pci_addr.devid ||
>>>>> +		    pci_dev->addr.function != pci_addr.function)
>>>>> +			continue;
>>>>> +
>>>>> +		ctx = ibv_open_device(ibdev);
>>>>> +		if (!ctx) {
>>>>> +			DRV_LOG(ERR, "Failed to open IB device %s",
>>>>> +				ibdev->name);
>>>>> +			continue;
>>>>> +		}
>>>>> +
>>>>> +		ret = ibv_query_device_ex(ctx, NULL, &dev_attr);
>>>>> +		DRV_LOG(INFO, "dev_attr.orig_attr.phys_port_cnt %u",
>>>>> +			dev_attr.orig_attr.phys_port_cnt);
>>>>> +		found_port = false;
>>>>> +
>>>>> +		for (port = 1; port <= dev_attr.orig_attr.phys_port_cnt;
>>>>> +		     port++) {
>>>>> +			struct ibv_parent_domain_init_attr attr = {};
>>>>> +			struct rte_ether_addr addr;
>>>>> +			char address[64];
>>>>> +			char name[RTE_ETH_NAME_MAX_LEN];
>>>>> +
>>>>> +			ret = get_port_mac(ibdev, port, &addr);
>>>>> +			if (ret)
>>>>> +				continue;
>>>>> +
>>>>> +			if (mac_addr && !rte_is_same_ether_addr(&addr,
>>>> mac_addr))
>>>>> +				continue;
>>>>> +
>>>>> +			rte_ether_format_addr(address, sizeof(address),
>>>> &addr);
>>>>> +			DRV_LOG(INFO, "device located port %u address
>>>> %s",
>>>>> +				port, address);
>>>>> +			found_port = true;
>>>>> +
>>>>> +			priv = rte_zmalloc_socket(NULL, sizeof(*priv),
>>>>> +						  RTE_CACHE_LINE_SIZE,
>>>>> +						  SOCKET_ID_ANY);
>>>>> +			if (!priv) {
>>>>> +				ret = -ENOMEM;
>>>>> +				goto failed;
>>>>> +			}
>>>>> +
>>>>> +			snprintf(name, sizeof(name), "%s_port%d",
>>>>> +				 pci_dev->device.name, port);
>>>>> +
>>>>> +			if (rte_eal_process_type() ==
>>>> RTE_PROC_SECONDARY) {
>>>>> +				int fd;
>>>>> +
>>>>> +				eth_dev =
>>>> rte_eth_dev_attach_secondary(name);
>>>>> +				if (!eth_dev) {
>>>>> +					DRV_LOG(ERR, "Can't attach to dev
>>>> %s",
>>>>> +						name);
>>>>> +					ret = -ENOMEM;
>>>>> +					goto failed;
>>>>> +				}
>>>>> +
>>>>> +				eth_dev->device = &pci_dev->device;
>>>>> +				eth_dev->dev_ops = &mana_dev_sec_ops;
>>>>> +				ret = mana_proc_priv_init(eth_dev);
>>>>> +				if (ret)
>>>>> +					goto failed;
>>>>> +				priv->process_priv = eth_dev-
>>>>> process_private;
>>>>> +
>>>>> +				/* Get the IB FD from the primary process */
>>>>> +				fd =
>>>> mana_mp_req_verbs_cmd_fd(eth_dev);
>>>>> +				if (fd < 0) {
>>>>> +					DRV_LOG(ERR, "Failed to get FD %d",
>>>> fd);
>>>>> +					ret = -ENODEV;
>>>>> +					goto failed;
>>>>> +				}
>>>>> +
>>>>> +				ret =
>>>> mana_map_doorbell_secondary(eth_dev, fd);
>>>>> +				if (ret) {
>>>>> +					DRV_LOG(ERR, "Failed secondary
>>>> map %d",
>>>>> +						fd);
>>>>> +					goto failed;
>>>>> +				}
>>>>> +
>>>>> +				/* fd is no not used after mapping doorbell */
>>>>> +				close(fd);
>>>>> +
>>>>> +				rte_spinlock_lock(&mana_shared_data-
>>>>> lock);
>>>>> +				mana_shared_data->secondary_cnt++;
>>>>> +				mana_local_data.secondary_cnt++;
>>>>> +				rte_spinlock_unlock(&mana_shared_data-
>>>>> lock);
>>>>> +
>>>>> +				rte_eth_copy_pci_info(eth_dev, pci_dev);
>>>>> +				rte_eth_dev_probing_finish(eth_dev);
>>>>> +
>>>>> +				/* Impossible to have more than one port
>>>>> +				 * matching a MAC address
>>>>> +				 */
>>>>> +				continue;
>>>>> +			}
>>>>> +
>>>>> +			eth_dev = rte_eth_dev_allocate(name);
>>>>> +			if (!eth_dev) {
>>>>> +				ret = -ENOMEM;
>>>>> +				goto failed;
>>>>> +			}
>>>>> +
>>>>> +			eth_dev->data->mac_addrs =
>>>>> +				rte_calloc("mana_mac", 1,
>>>>> +					   sizeof(struct rte_ether_addr), 0);
>>>>> +			if (!eth_dev->data->mac_addrs) {
>>>>> +				ret = -ENOMEM;
>>>>> +				goto failed;
>>>>> +			}
>>>>> +
>>>>> +			rte_ether_addr_copy(&addr, eth_dev->data-
>>>>> mac_addrs);
>>>>> +
>>>>> +			priv->ib_pd = ibv_alloc_pd(ctx);
>>>>> +			if (!priv->ib_pd) {
>>>>> +				DRV_LOG(ERR, "ibv_alloc_pd failed port %d",
>>>> port);
>>>>> +				ret = -ENOMEM;
>>>>> +				goto failed;
>>>>> +			}
>>>>> +
>>>>> +			/* Create a parent domain with the port number */
>>>>> +			attr.pd = priv->ib_pd;
>>>>> +			attr.comp_mask =
>>>> IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT;
>>>>> +			attr.pd_context = (void *)(uint64_t)port;
>>>>> +			priv->ib_parent_pd = ibv_alloc_parent_domain(ctx,
>>>> &attr);
>>>>> +			if (!priv->ib_parent_pd) {
>>>>> +				DRV_LOG(ERR,
>>>>> +					"ibv_alloc_parent_domain failed port
>>>> %d",
>>>>> +					port);
>>>>> +				ret = -ENOMEM;
>>>>> +				goto failed;
>>>>> +			}
>>>>> +
>>>>> +			priv->ib_ctx = ctx;
>>>>> +			priv->port_id = eth_dev->data->port_id;
>>>>> +			priv->dev_port = port;
>>>>> +			eth_dev->data->dev_private = priv;
>>>>> +			priv->dev_data = eth_dev->data;
>>>>> +
>>>>> +			priv->max_rx_queues = dev_attr.orig_attr.max_qp;
>>>>> +			priv->max_tx_queues = dev_attr.orig_attr.max_qp;
>>>>> +
>>>>> +			priv->max_rx_desc =
>>>>> +				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
>>>>> +					dev_attr.orig_attr.max_cqe);
>>>>> +			priv->max_tx_desc =
>>>>> +				RTE_MIN(dev_attr.orig_attr.max_qp_wr,
>>>>> +					dev_attr.orig_attr.max_cqe);
>>>>> +
>>>>> +			priv->max_send_sge = dev_attr.orig_attr.max_sge;
>>>>> +			priv->max_recv_sge = dev_attr.orig_attr.max_sge;
>>>>> +
>>>>> +			priv->max_mr = dev_attr.orig_attr.max_mr;
>>>>> +			priv->max_mr_size =
>>>> dev_attr.orig_attr.max_mr_size;
>>>>> +
>>>>> +			DRV_LOG(INFO, "dev %s max queues %d desc %d
>>>> sge %d",
>>>>> +				name, priv->max_rx_queues, priv-
>>>>> max_rx_desc,
>>>>> +				priv->max_send_sge);
>>>>> +
>>>>> +			rte_spinlock_lock(&mana_shared_data->lock);
>>>>> +			mana_shared_data->primary_cnt++;
>>>>> +			rte_spinlock_unlock(&mana_shared_data->lock);
>>>>> +
>>>>> +			eth_dev->data->dev_flags |=
>>>> RTE_ETH_DEV_INTR_RMV;
>>>>> +
>>>>> +			eth_dev->device = &pci_dev->device;
>>>>> +			eth_dev->data->dev_flags |=
>>>>> +				RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
>>>>> +
>>>> Please do not use the temporary macro. Please review this patch:
>>>>
>>>> f30e69b41f94 ("ethdev: add device flag to bypass auto-filled queue
>>>> xstats")
>>>>
>>>> This patch requires that per queue statistics are filled in
>>>> .xstats_get() by PMD.
>>> Thanks for pointing this out.
>>>
>>> It seems some PMDs are still depending on this flag for xstats.
>>>
>>> MANA doesn't implement xstats_get() currently, this flag is useful. Is it
>> okay to keep using this flag before it's finally the time to remove it from all
>> PMDs, or when MANA implements xstats?
>> Yes, your xstats doesn't implement now. Per queue stats should be filled in
>> xstats API, and the stats API cannot see per queue stats, so stats API in driver
>> shouldn't fill it(suggest that delete it from patch 17/18).
>>
>> I guess this flag can be removed if PMD does not support xstats.
>>>
> I don't understand your suggestion. An application can call rte_eth_stats_get() to get port stats, and this will call into stats_get() in the driver, as implemented in patch 17/18.
>
> When flag RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS  is set, an application can also use rte_eth_xstats_get() to get port stats even the driver doesn't implement xstats_get().

I think new PMD should follow the announced switch Ferruh mentioned, 
otherwise, the switch will never be completed.

Suggest that mana driver can implement a simple xstats_get() to fill per 
queue stats if you want to support per queue stats.

@Ferruh, what do you think?

RE: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[Long Li]

> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build environment
> and doc
> 
> 
> 在 2022/9/7 9:36, Long Li 写道:
> >> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build
> >> environment and doc
> >>
> >>
> >> 在 2022/9/1 2:05, Long Li 写道:
> >>>> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build
> >>>> environment and doc
> >>>>
> >>>>
> >>>> 在 2022/8/31 6:51, longli@linuxonhyperv.com 写道:
> >>>>> From: Long Li <longli@microsoft.com>
> >>>>>
> >>>>> MANA is a PCI device. It uses IB verbs to access hardware through
> >>>>> the kernel RDMA layer. This patch introduces build environment and
> >>>>> basic device probe functions.
> >>>>>
> >>>>> Signed-off-by: Long Li <longli@microsoft.com>
> >>>>> ---
> >>>>> Change log:
> >>>>> v2:
> >>>>> Fix typos.
> >>>>> Make the driver build only on x86-64 and Linux.
> >>>>> Remove unused header files.
> >>>>> Change port definition to uint16_t or uint8_t (for IB).
> >>>>> Use getline() in place of fgets() to read and truncate a line.
> >>>>> v3:
> >>>>> Add meson build check for required functions from RDMA direct verb
> >>>>> header file
> >>>>> v4:
> >>>>> Remove extra "\n" in logging code.
> >>>>> Use "r" in place of "rb" in fopen() to read text files.
> >>>>>
> >>>>> [snip]
> >>>>> +
> >>>>> +static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv
> >>>> __rte_unused,
> >>>>> +			      struct rte_pci_device *pci_dev,
> >>>>> +			      struct rte_ether_addr *mac_addr) {
> >>>>> +	struct ibv_device **ibv_list;
> >>>>> +	int ibv_idx;
> >>>>> +	struct ibv_context *ctx;
> >>>>> +	struct ibv_device_attr_ex dev_attr;
> >>>>> +	int num_devices;
> >>>>> +	int ret = 0;
> >>>>> +	uint8_t port;
> >>>>> +	struct mana_priv *priv = NULL;
> >>>>> +	struct rte_eth_dev *eth_dev = NULL;
> >>>>> +	bool found_port;
> >>>>> +
> >>>>> +	ibv_list = ibv_get_device_list(&num_devices);
> >>>>> +	for (ibv_idx = 0; ibv_idx < num_devices; ibv_idx++) {
> >>>>> +		struct ibv_device *ibdev = ibv_list[ibv_idx];
> >>>>> +		struct rte_pci_addr pci_addr;
> >>>>> +
> >>>>> +		DRV_LOG(INFO, "Probe device name %s
> dev_name %s
> >>>> ibdev_path %s",
> >>>>> +			ibdev->name, ibdev->dev_name, ibdev-
> >>>>> ibdev_path);
> >>>>> +
> >>>>> +		if (mana_ibv_device_to_pci_addr(ibdev, &pci_addr))
> >>>>> +			continue;
> >>>>> +
> >>>>> +		/* Ignore if this IB device is not this PCI device */
> >>>>> +		if (pci_dev->addr.domain != pci_addr.domain ||
> >>>>> +		    pci_dev->addr.bus != pci_addr.bus ||
> >>>>> +		    pci_dev->addr.devid != pci_addr.devid ||
> >>>>> +		    pci_dev->addr.function != pci_addr.function)
> >>>>> +			continue;
> >>>>> +
> >>>>> +		ctx = ibv_open_device(ibdev);
> >>>>> +		if (!ctx) {
> >>>>> +			DRV_LOG(ERR, "Failed to open IB device %s",
> >>>>> +				ibdev->name);
> >>>>> +			continue;
> >>>>> +		}
> >>>>> +
> >>>>> +		ret = ibv_query_device_ex(ctx, NULL, &dev_attr);
> >>>>> +		DRV_LOG(INFO,
> "dev_attr.orig_attr.phys_port_cnt %u",
> >>>>> +			dev_attr.orig_attr.phys_port_cnt);
> >>>>> +		found_port = false;
> >>>>> +
> >>>>> +		for (port = 1; port <=
> dev_attr.orig_attr.phys_port_cnt;
> >>>>> +		     port++) {
> >>>>> +			struct ibv_parent_domain_init_attr attr = {};
> >>>>> +			struct rte_ether_addr addr;
> >>>>> +			char address[64];
> >>>>> +			char name[RTE_ETH_NAME_MAX_LEN];
> >>>>> +
> >>>>> +			ret = get_port_mac(ibdev, port, &addr);
> >>>>> +			if (ret)
> >>>>> +				continue;
> >>>>> +
> >>>>> +			if (mac_addr
> && !rte_is_same_ether_addr(&addr,
> >>>> mac_addr))
> >>>>> +				continue;
> >>>>> +
> >>>>> +			rte_ether_format_addr(address,
> sizeof(address),
> >>>> &addr);
> >>>>> +			DRV_LOG(INFO, "device located port %u
> address
> >>>> %s",
> >>>>> +				port, address);
> >>>>> +			found_port = true;
> >>>>> +
> >>>>> +			priv = rte_zmalloc_socket(NULL, sizeof(*priv),
> >>>>> +
> RTE_CACHE_LINE_SIZE,
> >>>>> +						  SOCKET_ID_ANY);
> >>>>> +			if (!priv) {
> >>>>> +				ret = -ENOMEM;
> >>>>> +				goto failed;
> >>>>> +			}
> >>>>> +
> >>>>> +			snprintf(name, sizeof(name), "%s_port%d",
> >>>>> +				 pci_dev->device.name, port);
> >>>>> +
> >>>>> +			if (rte_eal_process_type() ==
> >>>> RTE_PROC_SECONDARY) {
> >>>>> +				int fd;
> >>>>> +
> >>>>> +				eth_dev =
> >>>> rte_eth_dev_attach_secondary(name);
> >>>>> +				if (!eth_dev) {
> >>>>> +					DRV_LOG(ERR, "Can't attach
> to dev
> >>>> %s",
> >>>>> +						name);
> >>>>> +					ret = -ENOMEM;
> >>>>> +					goto failed;
> >>>>> +				}
> >>>>> +
> >>>>> +				eth_dev->device = &pci_dev->device;
> >>>>> +				eth_dev->dev_ops =
> &mana_dev_sec_ops;
> >>>>> +				ret = mana_proc_priv_init(eth_dev);
> >>>>> +				if (ret)
> >>>>> +					goto failed;
> >>>>> +				priv->process_priv = eth_dev-
> >>>>> process_private;
> >>>>> +
> >>>>> +				/* Get the IB FD from the primary
> process */
> >>>>> +				fd =
> >>>> mana_mp_req_verbs_cmd_fd(eth_dev);
> >>>>> +				if (fd < 0) {
> >>>>> +					DRV_LOG(ERR, "Failed to get
> FD %d",
> >>>> fd);
> >>>>> +					ret = -ENODEV;
> >>>>> +					goto failed;
> >>>>> +				}
> >>>>> +
> >>>>> +				ret =
> >>>> mana_map_doorbell_secondary(eth_dev, fd);
> >>>>> +				if (ret) {
> >>>>> +					DRV_LOG(ERR, "Failed
> secondary
> >>>> map %d",
> >>>>> +						fd);
> >>>>> +					goto failed;
> >>>>> +				}
> >>>>> +
> >>>>> +				/* fd is no not used after mapping
> doorbell */
> >>>>> +				close(fd);
> >>>>> +
> >>>>> +
> 	rte_spinlock_lock(&mana_shared_data-
> >>>>> lock);
> >>>>> +				mana_shared_data-
> >secondary_cnt++;
> >>>>> +				mana_local_data.secondary_cnt++;
> >>>>> +
> 	rte_spinlock_unlock(&mana_shared_data-
> >>>>> lock);
> >>>>> +
> >>>>> +				rte_eth_copy_pci_info(eth_dev,
> pci_dev);
> >>>>> +
> 	rte_eth_dev_probing_finish(eth_dev);
> >>>>> +
> >>>>> +				/* Impossible to have more than one
> port
> >>>>> +				 * matching a MAC address
> >>>>> +				 */
> >>>>> +				continue;
> >>>>> +			}
> >>>>> +
> >>>>> +			eth_dev = rte_eth_dev_allocate(name);
> >>>>> +			if (!eth_dev) {
> >>>>> +				ret = -ENOMEM;
> >>>>> +				goto failed;
> >>>>> +			}
> >>>>> +
> >>>>> +			eth_dev->data->mac_addrs =
> >>>>> +				rte_calloc("mana_mac", 1,
> >>>>> +					   sizeof(struct
> rte_ether_addr), 0);
> >>>>> +			if (!eth_dev->data->mac_addrs) {
> >>>>> +				ret = -ENOMEM;
> >>>>> +				goto failed;
> >>>>> +			}
> >>>>> +
> >>>>> +			rte_ether_addr_copy(&addr, eth_dev-
> >data-
> >>>>> mac_addrs);
> >>>>> +
> >>>>> +			priv->ib_pd = ibv_alloc_pd(ctx);
> >>>>> +			if (!priv->ib_pd) {
> >>>>> +				DRV_LOG(ERR, "ibv_alloc_pd failed
> port %d",
> >>>> port);
> >>>>> +				ret = -ENOMEM;
> >>>>> +				goto failed;
> >>>>> +			}
> >>>>> +
> >>>>> +			/* Create a parent domain with the port
> number */
> >>>>> +			attr.pd = priv->ib_pd;
> >>>>> +			attr.comp_mask =
> >>>> IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT;
> >>>>> +			attr.pd_context = (void *)(uint64_t)port;
> >>>>> +			priv->ib_parent_pd =
> ibv_alloc_parent_domain(ctx,
> >>>> &attr);
> >>>>> +			if (!priv->ib_parent_pd) {
> >>>>> +				DRV_LOG(ERR,
> >>>>> +					"ibv_alloc_parent_domain
> failed port
> >>>> %d",
> >>>>> +					port);
> >>>>> +				ret = -ENOMEM;
> >>>>> +				goto failed;
> >>>>> +			}
> >>>>> +
> >>>>> +			priv->ib_ctx = ctx;
> >>>>> +			priv->port_id = eth_dev->data->port_id;
> >>>>> +			priv->dev_port = port;
> >>>>> +			eth_dev->data->dev_private = priv;
> >>>>> +			priv->dev_data = eth_dev->data;
> >>>>> +
> >>>>> +			priv->max_rx_queues =
> dev_attr.orig_attr.max_qp;
> >>>>> +			priv->max_tx_queues =
> dev_attr.orig_attr.max_qp;
> >>>>> +
> >>>>> +			priv->max_rx_desc =
> >>>>> +
> 	RTE_MIN(dev_attr.orig_attr.max_qp_wr,
> >>>>> +					dev_attr.orig_attr.max_cqe);
> >>>>> +			priv->max_tx_desc =
> >>>>> +
> 	RTE_MIN(dev_attr.orig_attr.max_qp_wr,
> >>>>> +					dev_attr.orig_attr.max_cqe);
> >>>>> +
> >>>>> +			priv->max_send_sge =
> dev_attr.orig_attr.max_sge;
> >>>>> +			priv->max_recv_sge =
> dev_attr.orig_attr.max_sge;
> >>>>> +
> >>>>> +			priv->max_mr = dev_attr.orig_attr.max_mr;
> >>>>> +			priv->max_mr_size =
> >>>> dev_attr.orig_attr.max_mr_size;
> >>>>> +
> >>>>> +			DRV_LOG(INFO, "dev %s max queues %d
> desc %d
> >>>> sge %d",
> >>>>> +				name, priv->max_rx_queues, priv-
> >>>>> max_rx_desc,
> >>>>> +				priv->max_send_sge);
> >>>>> +
> >>>>> +			rte_spinlock_lock(&mana_shared_data-
> >lock);
> >>>>> +			mana_shared_data->primary_cnt++;
> >>>>> +			rte_spinlock_unlock(&mana_shared_data-
> >lock);
> >>>>> +
> >>>>> +			eth_dev->data->dev_flags |=
> >>>> RTE_ETH_DEV_INTR_RMV;
> >>>>> +
> >>>>> +			eth_dev->device = &pci_dev->device;
> >>>>> +			eth_dev->data->dev_flags |=
> >>>>> +
> 	RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
> >>>>> +
> >>>> Please do not use the temporary macro. Please review this patch:
> >>>>
> >>>> f30e69b41f94 ("ethdev: add device flag to bypass auto-filled queue
> >>>> xstats")
> >>>>
> >>>> This patch requires that per queue statistics are filled in
> >>>> .xstats_get() by PMD.
> >>> Thanks for pointing this out.
> >>>
> >>> It seems some PMDs are still depending on this flag for xstats.
> >>>
> >>> MANA doesn't implement xstats_get() currently, this flag is useful.
> >>> Is it
> >> okay to keep using this flag before it's finally the time to remove
> >> it from all PMDs, or when MANA implements xstats?
> >> Yes, your xstats doesn't implement now. Per queue stats should be
> >> filled in xstats API, and the stats API cannot see per queue stats,
> >> so stats API in driver shouldn't fill it(suggest that delete it from patch
> 17/18).
> >>
> >> I guess this flag can be removed if PMD does not support xstats.
> >>>
> > I don't understand your suggestion. An application can call
> rte_eth_stats_get() to get port stats, and this will call into stats_get() in the
> driver, as implemented in patch 17/18.
> >
> > When flag RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS  is set, an application
> can also use rte_eth_xstats_get() to get port stats even the driver doesn't
> implement xstats_get().
> 
> I think new PMD should follow the announced switch Ferruh mentioned,
> otherwise, the switch will never be completed.

I see now I understand more. I prefer to keep patch 17/18, or an application calling rte_eth_stats_get() will fail. (before the switch happens)

> 
> Suggest that mana driver can implement a simple xstats_get() to fill per
> queue stats if you want to support per queue stats.
> 
> @Ferruh, what do you think?

Re: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[Ferruh Yigit]

On 9/7/2022 3:16 AM, lihuisong (C) wrote:
> CAUTION: This message has originated from an External Source. Please use 
> proper judgment and caution when opening attachments, clicking links, or 
> responding to this email.
> 
> 
> 在 2022/9/7 9:36, Long Li 写道:
>>> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build 
>>> environment
>>> and doc
>>>
>>>
>>> 在 2022/9/1 2:05, Long Li 写道:
>>>>> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build
>>>>> environment and doc
>>>>>
>>>>>
>>>>> 在 2022/8/31 6:51, longli@linuxonhyperv.com 写道:
>>>>>> From: Long Li <longli@microsoft.com>
>>>>>>
>>>>>> MANA is a PCI device. It uses IB verbs to access hardware through
>>>>>> the kernel RDMA layer. This patch introduces build environment and
>>>>>> basic device probe functions.
>>>>>>
>>>>>> Signed-off-by: Long Li <longli@microsoft.com>
>>>>>> ---
>>>>>> Change log:
>>>>>> v2:
>>>>>> Fix typos.
>>>>>> Make the driver build only on x86-64 and Linux.
>>>>>> Remove unused header files.
>>>>>> Change port definition to uint16_t or uint8_t (for IB).
>>>>>> Use getline() in place of fgets() to read and truncate a line.
>>>>>> v3:
>>>>>> Add meson build check for required functions from RDMA direct verb
>>>>>> header file
>>>>>> v4:
>>>>>> Remove extra "\n" in logging code.
>>>>>> Use "r" in place of "rb" in fopen() to read text files.
>>>>>>
>>>>>> [snip]
>>>>>> +
>>>>>> +static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv
>>>>> __rte_unused,
>>>>>> +                       struct rte_pci_device *pci_dev,
>>>>>> +                       struct rte_ether_addr *mac_addr) {
>>>>>> + struct ibv_device **ibv_list;
>>>>>> + int ibv_idx;
>>>>>> + struct ibv_context *ctx;
>>>>>> + struct ibv_device_attr_ex dev_attr;
>>>>>> + int num_devices;
>>>>>> + int ret = 0;
>>>>>> + uint8_t port;
>>>>>> + struct mana_priv *priv = NULL;
>>>>>> + struct rte_eth_dev *eth_dev = NULL;
>>>>>> + bool found_port;
>>>>>> +
>>>>>> + ibv_list = ibv_get_device_list(&num_devices);
>>>>>> + for (ibv_idx = 0; ibv_idx < num_devices; ibv_idx++) {
>>>>>> +         struct ibv_device *ibdev = ibv_list[ibv_idx];
>>>>>> +         struct rte_pci_addr pci_addr;
>>>>>> +
>>>>>> +         DRV_LOG(INFO, "Probe device name %s dev_name %s
>>>>> ibdev_path %s",
>>>>>> +                 ibdev->name, ibdev->dev_name, ibdev-
>>>>>> ibdev_path);
>>>>>> +
>>>>>> +         if (mana_ibv_device_to_pci_addr(ibdev, &pci_addr))
>>>>>> +                 continue;
>>>>>> +
>>>>>> +         /* Ignore if this IB device is not this PCI device */
>>>>>> +         if (pci_dev->addr.domain != pci_addr.domain ||
>>>>>> +             pci_dev->addr.bus != pci_addr.bus ||
>>>>>> +             pci_dev->addr.devid != pci_addr.devid ||
>>>>>> +             pci_dev->addr.function != pci_addr.function)
>>>>>> +                 continue;
>>>>>> +
>>>>>> +         ctx = ibv_open_device(ibdev);
>>>>>> +         if (!ctx) {
>>>>>> +                 DRV_LOG(ERR, "Failed to open IB device %s",
>>>>>> +                         ibdev->name);
>>>>>> +                 continue;
>>>>>> +         }
>>>>>> +
>>>>>> +         ret = ibv_query_device_ex(ctx, NULL, &dev_attr);
>>>>>> +         DRV_LOG(INFO, "dev_attr.orig_attr.phys_port_cnt %u",
>>>>>> +                 dev_attr.orig_attr.phys_port_cnt);
>>>>>> +         found_port = false;
>>>>>> +
>>>>>> +         for (port = 1; port <= dev_attr.orig_attr.phys_port_cnt;
>>>>>> +              port++) {
>>>>>> +                 struct ibv_parent_domain_init_attr attr = {};
>>>>>> +                 struct rte_ether_addr addr;
>>>>>> +                 char address[64];
>>>>>> +                 char name[RTE_ETH_NAME_MAX_LEN];
>>>>>> +
>>>>>> +                 ret = get_port_mac(ibdev, port, &addr);
>>>>>> +                 if (ret)
>>>>>> +                         continue;
>>>>>> +
>>>>>> +                 if (mac_addr && !rte_is_same_ether_addr(&addr,
>>>>> mac_addr))
>>>>>> +                         continue;
>>>>>> +
>>>>>> +                 rte_ether_format_addr(address, sizeof(address),
>>>>> &addr);
>>>>>> +                 DRV_LOG(INFO, "device located port %u address
>>>>> %s",
>>>>>> +                         port, address);
>>>>>> +                 found_port = true;
>>>>>> +
>>>>>> +                 priv = rte_zmalloc_socket(NULL, sizeof(*priv),
>>>>>> +                                           RTE_CACHE_LINE_SIZE,
>>>>>> +                                           SOCKET_ID_ANY);
>>>>>> +                 if (!priv) {
>>>>>> +                         ret = -ENOMEM;
>>>>>> +                         goto failed;
>>>>>> +                 }
>>>>>> +
>>>>>> +                 snprintf(name, sizeof(name), "%s_port%d",
>>>>>> +                          pci_dev->device.name, port);
>>>>>> +
>>>>>> +                 if (rte_eal_process_type() ==
>>>>> RTE_PROC_SECONDARY) {
>>>>>> +                         int fd;
>>>>>> +
>>>>>> +                         eth_dev =
>>>>> rte_eth_dev_attach_secondary(name);
>>>>>> +                         if (!eth_dev) {
>>>>>> +                                 DRV_LOG(ERR, "Can't attach to dev
>>>>> %s",
>>>>>> +                                         name);
>>>>>> +                                 ret = -ENOMEM;
>>>>>> +                                 goto failed;
>>>>>> +                         }
>>>>>> +
>>>>>> +                         eth_dev->device = &pci_dev->device;
>>>>>> +                         eth_dev->dev_ops = &mana_dev_sec_ops;
>>>>>> +                         ret = mana_proc_priv_init(eth_dev);
>>>>>> +                         if (ret)
>>>>>> +                                 goto failed;
>>>>>> +                         priv->process_priv = eth_dev-
>>>>>> process_private;
>>>>>> +
>>>>>> +                         /* Get the IB FD from the primary 
>>>>>> process */
>>>>>> +                         fd =
>>>>> mana_mp_req_verbs_cmd_fd(eth_dev);
>>>>>> +                         if (fd < 0) {
>>>>>> +                                 DRV_LOG(ERR, "Failed to get FD %d",
>>>>> fd);
>>>>>> +                                 ret = -ENODEV;
>>>>>> +                                 goto failed;
>>>>>> +                         }
>>>>>> +
>>>>>> +                         ret =
>>>>> mana_map_doorbell_secondary(eth_dev, fd);
>>>>>> +                         if (ret) {
>>>>>> +                                 DRV_LOG(ERR, "Failed secondary
>>>>> map %d",
>>>>>> +                                         fd);
>>>>>> +                                 goto failed;
>>>>>> +                         }
>>>>>> +
>>>>>> +                         /* fd is no not used after mapping 
>>>>>> doorbell */
>>>>>> +                         close(fd);
>>>>>> +
>>>>>> +                         rte_spinlock_lock(&mana_shared_data-
>>>>>> lock);
>>>>>> +                         mana_shared_data->secondary_cnt++;
>>>>>> +                         mana_local_data.secondary_cnt++;
>>>>>> +                         rte_spinlock_unlock(&mana_shared_data-
>>>>>> lock);
>>>>>> +
>>>>>> +                         rte_eth_copy_pci_info(eth_dev, pci_dev);
>>>>>> +                         rte_eth_dev_probing_finish(eth_dev);
>>>>>> +
>>>>>> +                         /* Impossible to have more than one port
>>>>>> +                          * matching a MAC address
>>>>>> +                          */
>>>>>> +                         continue;
>>>>>> +                 }
>>>>>> +
>>>>>> +                 eth_dev = rte_eth_dev_allocate(name);
>>>>>> +                 if (!eth_dev) {
>>>>>> +                         ret = -ENOMEM;
>>>>>> +                         goto failed;
>>>>>> +                 }
>>>>>> +
>>>>>> +                 eth_dev->data->mac_addrs =
>>>>>> +                         rte_calloc("mana_mac", 1,
>>>>>> +                                    sizeof(struct 
>>>>>> rte_ether_addr), 0);
>>>>>> +                 if (!eth_dev->data->mac_addrs) {
>>>>>> +                         ret = -ENOMEM;
>>>>>> +                         goto failed;
>>>>>> +                 }
>>>>>> +
>>>>>> +                 rte_ether_addr_copy(&addr, eth_dev->data-
>>>>>> mac_addrs);
>>>>>> +
>>>>>> +                 priv->ib_pd = ibv_alloc_pd(ctx);
>>>>>> +                 if (!priv->ib_pd) {
>>>>>> +                         DRV_LOG(ERR, "ibv_alloc_pd failed port %d",
>>>>> port);
>>>>>> +                         ret = -ENOMEM;
>>>>>> +                         goto failed;
>>>>>> +                 }
>>>>>> +
>>>>>> +                 /* Create a parent domain with the port number */
>>>>>> +                 attr.pd = priv->ib_pd;
>>>>>> +                 attr.comp_mask =
>>>>> IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT;
>>>>>> +                 attr.pd_context = (void *)(uint64_t)port;
>>>>>> +                 priv->ib_parent_pd = ibv_alloc_parent_domain(ctx,
>>>>> &attr);
>>>>>> +                 if (!priv->ib_parent_pd) {
>>>>>> +                         DRV_LOG(ERR,
>>>>>> +                                 "ibv_alloc_parent_domain failed 
>>>>>> port
>>>>> %d",
>>>>>> +                                 port);
>>>>>> +                         ret = -ENOMEM;
>>>>>> +                         goto failed;
>>>>>> +                 }
>>>>>> +
>>>>>> +                 priv->ib_ctx = ctx;
>>>>>> +                 priv->port_id = eth_dev->data->port_id;
>>>>>> +                 priv->dev_port = port;
>>>>>> +                 eth_dev->data->dev_private = priv;
>>>>>> +                 priv->dev_data = eth_dev->data;
>>>>>> +
>>>>>> +                 priv->max_rx_queues = dev_attr.orig_attr.max_qp;
>>>>>> +                 priv->max_tx_queues = dev_attr.orig_attr.max_qp;
>>>>>> +
>>>>>> +                 priv->max_rx_desc =
>>>>>> +                         RTE_MIN(dev_attr.orig_attr.max_qp_wr,
>>>>>> +                                 dev_attr.orig_attr.max_cqe);
>>>>>> +                 priv->max_tx_desc =
>>>>>> +                         RTE_MIN(dev_attr.orig_attr.max_qp_wr,
>>>>>> +                                 dev_attr.orig_attr.max_cqe);
>>>>>> +
>>>>>> +                 priv->max_send_sge = dev_attr.orig_attr.max_sge;
>>>>>> +                 priv->max_recv_sge = dev_attr.orig_attr.max_sge;
>>>>>> +
>>>>>> +                 priv->max_mr = dev_attr.orig_attr.max_mr;
>>>>>> +                 priv->max_mr_size =
>>>>> dev_attr.orig_attr.max_mr_size;
>>>>>> +
>>>>>> +                 DRV_LOG(INFO, "dev %s max queues %d desc %d
>>>>> sge %d",
>>>>>> +                         name, priv->max_rx_queues, priv-
>>>>>> max_rx_desc,
>>>>>> +                         priv->max_send_sge);
>>>>>> +
>>>>>> +                 rte_spinlock_lock(&mana_shared_data->lock);
>>>>>> +                 mana_shared_data->primary_cnt++;
>>>>>> +                 rte_spinlock_unlock(&mana_shared_data->lock);
>>>>>> +
>>>>>> +                 eth_dev->data->dev_flags |=
>>>>> RTE_ETH_DEV_INTR_RMV;
>>>>>> +
>>>>>> +                 eth_dev->device = &pci_dev->device;
>>>>>> +                 eth_dev->data->dev_flags |=
>>>>>> +                         RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
>>>>>> +
>>>>> Please do not use the temporary macro. Please review this patch:
>>>>>
>>>>> f30e69b41f94 ("ethdev: add device flag to bypass auto-filled queue
>>>>> xstats")
>>>>>
>>>>> This patch requires that per queue statistics are filled in
>>>>> .xstats_get() by PMD.
>>>> Thanks for pointing this out.
>>>>
>>>> It seems some PMDs are still depending on this flag for xstats.
>>>>
>>>> MANA doesn't implement xstats_get() currently, this flag is useful. 
>>>> Is it
>>> okay to keep using this flag before it's finally the time to remove 
>>> it from all
>>> PMDs, or when MANA implements xstats?
>>> Yes, your xstats doesn't implement now. Per queue stats should be 
>>> filled in
>>> xstats API, and the stats API cannot see per queue stats, so stats 
>>> API in driver
>>> shouldn't fill it(suggest that delete it from patch 17/18).
>>>
>>> I guess this flag can be removed if PMD does not support xstats.
>>>>
>> I don't understand your suggestion. An application can call 
>> rte_eth_stats_get() to get port stats, and this will call into 
>> stats_get() in the driver, as implemented in patch 17/18.
>>
>> When flag RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS  is set, an application 
>> can also use rte_eth_xstats_get() to get port stats even the driver 
>> doesn't implement xstats_get().
> 
> I think new PMD should follow the announced switch Ferruh mentioned,
> otherwise, the switch will never be completed.
> 
> Suggest that mana driver can implement a simple xstats_get() to fill per
> queue stats if you want to support per queue stats.
> 
> @Ferruh, what do you think?
> 

Hi Huisong,

Thanks for reminding it, yes it makes sense to implement new method in 
new drivers.


Long,

There is a long term plan to move queue stats from basic stats structure 
to xstats. The reason behind is increasing number of queues makes basis 
stats struct too big, on the other hand xstats is more flexible and no 
fixes size array is required.

You can remove the 'RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS' flag and driver 
won't support queue stats in xstats by default. Instead to have that 
support you will need to implement xstats, later or in this set.
When queue stats implemented in xstats, please remember to remove 
updating 'stats->q_*' in basic stats.

Also in 17/18, the feature 'Stats per queue' seems added, but that is 
not correct, feature name is misleading here. But it is about queue 
stats mapping, please check 'doc/guides/nics/features.rst'.
So can you please drop that feature.

RE: [Patch v6 01/18] net/mana: add basic driver, build environment and doc

[Long Li]

> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build environment
> and doc
> 
> On 9/7/2022 3:16 AM, lihuisong (C) wrote:
> > CAUTION: This message has originated from an External Source. Please
> > use proper judgment and caution when opening attachments, clicking
> > links, or responding to this email.
> >
> >
> > 在 2022/9/7 9:36, Long Li 写道:
> >>> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build
> >>> environment and doc
> >>>
> >>>
> >>> 在 2022/9/1 2:05, Long Li 写道:
> >>>>> Subject: Re: [Patch v6 01/18] net/mana: add basic driver, build
> >>>>> environment and doc
> >>>>>
> >>>>>
> >>>>> 在 2022/8/31 6:51, longli@linuxonhyperv.com 写道:
> >>>>>> From: Long Li <longli@microsoft.com>
> >>>>>>
> >>>>>> MANA is a PCI device. It uses IB verbs to access hardware through
> >>>>>> the kernel RDMA layer. This patch introduces build environment and
> >>>>>> basic device probe functions.
> >>>>>>
> >>>>>> Signed-off-by: Long Li <longli@microsoft.com>
> >>>>>> ---
> >>>>>> Change log:
> >>>>>> v2:
> >>>>>> Fix typos.
> >>>>>> Make the driver build only on x86-64 and Linux.
> >>>>>> Remove unused header files.
> >>>>>> Change port definition to uint16_t or uint8_t (for IB).
> >>>>>> Use getline() in place of fgets() to read and truncate a line.
> >>>>>> v3:
> >>>>>> Add meson build check for required functions from RDMA direct
> verb
> >>>>>> header file
> >>>>>> v4:
> >>>>>> Remove extra "\n" in logging code.
> >>>>>> Use "r" in place of "rb" in fopen() to read text files.
> >>>>>>
> >>>>>> [snip]
> >>>>>> +
> >>>>>> +static int mana_pci_probe_mac(struct rte_pci_driver *pci_drv
> >>>>> __rte_unused,
> >>>>>> +                       struct rte_pci_device *pci_dev,
> >>>>>> +                       struct rte_ether_addr *mac_addr) {
> >>>>>> + struct ibv_device **ibv_list;
> >>>>>> + int ibv_idx;
> >>>>>> + struct ibv_context *ctx;
> >>>>>> + struct ibv_device_attr_ex dev_attr;
> >>>>>> + int num_devices;
> >>>>>> + int ret = 0;
> >>>>>> + uint8_t port;
> >>>>>> + struct mana_priv *priv = NULL;
> >>>>>> + struct rte_eth_dev *eth_dev = NULL;
> >>>>>> + bool found_port;
> >>>>>> +
> >>>>>> + ibv_list = ibv_get_device_list(&num_devices);
> >>>>>> + for (ibv_idx = 0; ibv_idx < num_devices; ibv_idx++) {
> >>>>>> +         struct ibv_device *ibdev = ibv_list[ibv_idx];
> >>>>>> +         struct rte_pci_addr pci_addr;
> >>>>>> +
> >>>>>> +         DRV_LOG(INFO, "Probe device name %s dev_name %s
> >>>>> ibdev_path %s",
> >>>>>> +                 ibdev->name, ibdev->dev_name, ibdev-
> >>>>>> ibdev_path);
> >>>>>> +
> >>>>>> +         if (mana_ibv_device_to_pci_addr(ibdev, &pci_addr))
> >>>>>> +                 continue;
> >>>>>> +
> >>>>>> +         /* Ignore if this IB device is not this PCI device */
> >>>>>> +         if (pci_dev->addr.domain != pci_addr.domain ||
> >>>>>> +             pci_dev->addr.bus != pci_addr.bus ||
> >>>>>> +             pci_dev->addr.devid != pci_addr.devid ||
> >>>>>> +             pci_dev->addr.function != pci_addr.function)
> >>>>>> +                 continue;
> >>>>>> +
> >>>>>> +         ctx = ibv_open_device(ibdev);
> >>>>>> +         if (!ctx) {
> >>>>>> +                 DRV_LOG(ERR, "Failed to open IB device %s",
> >>>>>> +                         ibdev->name);
> >>>>>> +                 continue;
> >>>>>> +         }
> >>>>>> +
> >>>>>> +         ret = ibv_query_device_ex(ctx, NULL, &dev_attr);
> >>>>>> +         DRV_LOG(INFO, "dev_attr.orig_attr.phys_port_cnt %u",
> >>>>>> +                 dev_attr.orig_attr.phys_port_cnt);
> >>>>>> +         found_port = false;
> >>>>>> +
> >>>>>> +         for (port = 1; port <= dev_attr.orig_attr.phys_port_cnt;
> >>>>>> +              port++) {
> >>>>>> +                 struct ibv_parent_domain_init_attr attr = {};
> >>>>>> +                 struct rte_ether_addr addr;
> >>>>>> +                 char address[64];
> >>>>>> +                 char name[RTE_ETH_NAME_MAX_LEN];
> >>>>>> +
> >>>>>> +                 ret = get_port_mac(ibdev, port, &addr);
> >>>>>> +                 if (ret)
> >>>>>> +                         continue;
> >>>>>> +
> >>>>>> +                 if (mac_addr && !rte_is_same_ether_addr(&addr,
> >>>>> mac_addr))
> >>>>>> +                         continue;
> >>>>>> +
> >>>>>> +                 rte_ether_format_addr(address, sizeof(address),
> >>>>> &addr);
> >>>>>> +                 DRV_LOG(INFO, "device located port %u address
> >>>>> %s",
> >>>>>> +                         port, address);
> >>>>>> +                 found_port = true;
> >>>>>> +
> >>>>>> +                 priv = rte_zmalloc_socket(NULL, sizeof(*priv),
> >>>>>> +                                           RTE_CACHE_LINE_SIZE,
> >>>>>> +                                           SOCKET_ID_ANY);
> >>>>>> +                 if (!priv) {
> >>>>>> +                         ret = -ENOMEM;
> >>>>>> +                         goto failed;
> >>>>>> +                 }
> >>>>>> +
> >>>>>> +                 snprintf(name, sizeof(name), "%s_port%d",
> >>>>>> +                          pci_dev->device.name, port);
> >>>>>> +
> >>>>>> +                 if (rte_eal_process_type() ==
> >>>>> RTE_PROC_SECONDARY) {
> >>>>>> +                         int fd;
> >>>>>> +
> >>>>>> +                         eth_dev =
> >>>>> rte_eth_dev_attach_secondary(name);
> >>>>>> +                         if (!eth_dev) {
> >>>>>> +                                 DRV_LOG(ERR, "Can't attach to dev
> >>>>> %s",
> >>>>>> +                                         name);
> >>>>>> +                                 ret = -ENOMEM;
> >>>>>> +                                 goto failed;
> >>>>>> +                         }
> >>>>>> +
> >>>>>> +                         eth_dev->device = &pci_dev->device;
> >>>>>> +                         eth_dev->dev_ops = &mana_dev_sec_ops;
> >>>>>> +                         ret = mana_proc_priv_init(eth_dev);
> >>>>>> +                         if (ret)
> >>>>>> +                                 goto failed;
> >>>>>> +                         priv->process_priv = eth_dev-
> >>>>>> process_private;
> >>>>>> +
> >>>>>> +                         /* Get the IB FD from the primary
> >>>>>> process */
> >>>>>> +                         fd =
> >>>>> mana_mp_req_verbs_cmd_fd(eth_dev);
> >>>>>> +                         if (fd < 0) {
> >>>>>> +                                 DRV_LOG(ERR, "Failed to get FD %d",
> >>>>> fd);
> >>>>>> +                                 ret = -ENODEV;
> >>>>>> +                                 goto failed;
> >>>>>> +                         }
> >>>>>> +
> >>>>>> +                         ret =
> >>>>> mana_map_doorbell_secondary(eth_dev, fd);
> >>>>>> +                         if (ret) {
> >>>>>> +                                 DRV_LOG(ERR, "Failed secondary
> >>>>> map %d",
> >>>>>> +                                         fd);
> >>>>>> +                                 goto failed;
> >>>>>> +                         }
> >>>>>> +
> >>>>>> +                         /* fd is no not used after mapping
> >>>>>> doorbell */
> >>>>>> +                         close(fd);
> >>>>>> +
> >>>>>> +                         rte_spinlock_lock(&mana_shared_data-
> >>>>>> lock);
> >>>>>> +                         mana_shared_data->secondary_cnt++;
> >>>>>> +                         mana_local_data.secondary_cnt++;
> >>>>>> +                         rte_spinlock_unlock(&mana_shared_data-
> >>>>>> lock);
> >>>>>> +
> >>>>>> +                         rte_eth_copy_pci_info(eth_dev, pci_dev);
> >>>>>> +                         rte_eth_dev_probing_finish(eth_dev);
> >>>>>> +
> >>>>>> +                         /* Impossible to have more than one port
> >>>>>> +                          * matching a MAC address
> >>>>>> +                          */
> >>>>>> +                         continue;
> >>>>>> +                 }
> >>>>>> +
> >>>>>> +                 eth_dev = rte_eth_dev_allocate(name);
> >>>>>> +                 if (!eth_dev) {
> >>>>>> +                         ret = -ENOMEM;
> >>>>>> +                         goto failed;
> >>>>>> +                 }
> >>>>>> +
> >>>>>> +                 eth_dev->data->mac_addrs =
> >>>>>> +                         rte_calloc("mana_mac", 1,
> >>>>>> +                                    sizeof(struct
> >>>>>> rte_ether_addr), 0);
> >>>>>> +                 if (!eth_dev->data->mac_addrs) {
> >>>>>> +                         ret = -ENOMEM;
> >>>>>> +                         goto failed;
> >>>>>> +                 }
> >>>>>> +
> >>>>>> +                 rte_ether_addr_copy(&addr, eth_dev->data-
> >>>>>> mac_addrs);
> >>>>>> +
> >>>>>> +                 priv->ib_pd = ibv_alloc_pd(ctx);
> >>>>>> +                 if (!priv->ib_pd) {
> >>>>>> +                         DRV_LOG(ERR, "ibv_alloc_pd failed port %d",
> >>>>> port);
> >>>>>> +                         ret = -ENOMEM;
> >>>>>> +                         goto failed;
> >>>>>> +                 }
> >>>>>> +
> >>>>>> +                 /* Create a parent domain with the port number */
> >>>>>> +                 attr.pd = priv->ib_pd;
> >>>>>> +                 attr.comp_mask =
> >>>>> IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT;
> >>>>>> +                 attr.pd_context = (void *)(uint64_t)port;
> >>>>>> +                 priv->ib_parent_pd = ibv_alloc_parent_domain(ctx,
> >>>>> &attr);
> >>>>>> +                 if (!priv->ib_parent_pd) {
> >>>>>> +                         DRV_LOG(ERR,
> >>>>>> +                                 "ibv_alloc_parent_domain failed
> >>>>>> port
> >>>>> %d",
> >>>>>> +                                 port);
> >>>>>> +                         ret = -ENOMEM;
> >>>>>> +                         goto failed;
> >>>>>> +                 }
> >>>>>> +
> >>>>>> +                 priv->ib_ctx = ctx;
> >>>>>> +                 priv->port_id = eth_dev->data->port_id;
> >>>>>> +                 priv->dev_port = port;
> >>>>>> +                 eth_dev->data->dev_private = priv;
> >>>>>> +                 priv->dev_data = eth_dev->data;
> >>>>>> +
> >>>>>> +                 priv->max_rx_queues = dev_attr.orig_attr.max_qp;
> >>>>>> +                 priv->max_tx_queues = dev_attr.orig_attr.max_qp;
> >>>>>> +
> >>>>>> +                 priv->max_rx_desc =
> >>>>>> +                         RTE_MIN(dev_attr.orig_attr.max_qp_wr,
> >>>>>> +                                 dev_attr.orig_attr.max_cqe);
> >>>>>> +                 priv->max_tx_desc =
> >>>>>> +                         RTE_MIN(dev_attr.orig_attr.max_qp_wr,
> >>>>>> +                                 dev_attr.orig_attr.max_cqe);
> >>>>>> +
> >>>>>> +                 priv->max_send_sge = dev_attr.orig_attr.max_sge;
> >>>>>> +                 priv->max_recv_sge = dev_attr.orig_attr.max_sge;
> >>>>>> +
> >>>>>> +                 priv->max_mr = dev_attr.orig_attr.max_mr;
> >>>>>> +                 priv->max_mr_size =
> >>>>> dev_attr.orig_attr.max_mr_size;
> >>>>>> +
> >>>>>> +                 DRV_LOG(INFO, "dev %s max queues %d desc %d
> >>>>> sge %d",
> >>>>>> +                         name, priv->max_rx_queues, priv-
> >>>>>> max_rx_desc,
> >>>>>> +                         priv->max_send_sge);
> >>>>>> +
> >>>>>> +                 rte_spinlock_lock(&mana_shared_data->lock);
> >>>>>> +                 mana_shared_data->primary_cnt++;
> >>>>>> +                 rte_spinlock_unlock(&mana_shared_data->lock);
> >>>>>> +
> >>>>>> +                 eth_dev->data->dev_flags |=
> >>>>> RTE_ETH_DEV_INTR_RMV;
> >>>>>> +
> >>>>>> +                 eth_dev->device = &pci_dev->device;
> >>>>>> +                 eth_dev->data->dev_flags |=
> >>>>>> +                         RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
> >>>>>> +
> >>>>> Please do not use the temporary macro. Please review this patch:
> >>>>>
> >>>>> f30e69b41f94 ("ethdev: add device flag to bypass auto-filled queue
> >>>>> xstats")
> >>>>>
> >>>>> This patch requires that per queue statistics are filled in
> >>>>> .xstats_get() by PMD.
> >>>> Thanks for pointing this out.
> >>>>
> >>>> It seems some PMDs are still depending on this flag for xstats.
> >>>>
> >>>> MANA doesn't implement xstats_get() currently, this flag is useful.
> >>>> Is it
> >>> okay to keep using this flag before it's finally the time to remove
> >>> it from all
> >>> PMDs, or when MANA implements xstats?
> >>> Yes, your xstats doesn't implement now. Per queue stats should be
> >>> filled in
> >>> xstats API, and the stats API cannot see per queue stats, so stats
> >>> API in driver
> >>> shouldn't fill it(suggest that delete it from patch 17/18).
> >>>
> >>> I guess this flag can be removed if PMD does not support xstats.
> >>>>
> >> I don't understand your suggestion. An application can call
> >> rte_eth_stats_get() to get port stats, and this will call into
> >> stats_get() in the driver, as implemented in patch 17/18.
> >>
> >> When flag RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS  is set, an application
> >> can also use rte_eth_xstats_get() to get port stats even the driver
> >> doesn't implement xstats_get().
> >
> > I think new PMD should follow the announced switch Ferruh mentioned,
> > otherwise, the switch will never be completed.
> >
> > Suggest that mana driver can implement a simple xstats_get() to fill per
> > queue stats if you want to support per queue stats.
> >
> > @Ferruh, what do you think?
> >
> 
> Hi Huisong,
> 
> Thanks for reminding it, yes it makes sense to implement new method in
> new drivers.
> 
> 
> Long,
> 
> There is a long term plan to move queue stats from basic stats structure
> to xstats. The reason behind is increasing number of queues makes basis
> stats struct too big, on the other hand xstats is more flexible and no
> fixes size array is required.
> 
> You can remove the 'RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS' flag and
> driver
> won't support queue stats in xstats by default. Instead to have that
> support you will need to implement xstats, later or in this set.
> When queue stats implemented in xstats, please remember to remove
> updating 'stats->q_*' in basic stats.
> 
> Also in 17/18, the feature 'Stats per queue' seems added, but that is
> not correct, feature name is misleading here. But it is about queue
> stats mapping, please check 'doc/guides/nics/features.rst'.
> So can you please drop that feature.

Thank you for clearing the confusion. I'm removing "Stats per queue" in doc/guides/nics/features/mana.ini. The other parts of patch 17/18 are still relevant so not changing them.

Also removing 'RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS' as suggested.