@@ -23,6 +23,22 @@
const struct mlx5_glue *mlx5_glue;
#endif
+/* Environment variable to control the doorbell register mapping. */
+#define MLX5_SHUT_UP_BF "MLX5_SHUT_UP_BF"
+#if defined(RTE_ARCH_ARM64)
+#define MLX5_SHUT_UP_BF_DEFAULT "0"
+#else
+#define MLX5_SHUT_UP_BF_DEFAULT "1"
+#endif
+
+/* Default PMD specific parameter value. */
+#define MLX5_TXDB_UNSET (-1)
+
+/* MLX5_TX_DB_NC supported values. */
+#define MLX5_TXDB_CACHED 0
+#define MLX5_TXDB_NCACHED 1
+#define MLX5_TXDB_HEURISTIC 2
+
int
mlx5_get_pci_addr(const char *dev_path, struct rte_pci_addr *pci_addr)
{
@@ -401,6 +417,127 @@ mlx5_glue_constructor(void)
mlx5_glue = NULL;
}
+static int
+mlx5_config_doorbell_mapping_env(int dbnc)
+{
+ char *env;
+ int value;
+
+ MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
+ /* Get environment variable to store. */
+ env = getenv(MLX5_SHUT_UP_BF);
+ value = env ? !!strcmp(env, "0") : MLX5_TXDB_UNSET;
+ if (dbnc == MLX5_TXDB_UNSET)
+ setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
+ else
+ setenv(MLX5_SHUT_UP_BF,
+ dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
+ return value;
+}
+
+static void
+mlx5_restore_doorbell_mapping_env(int value)
+{
+ MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
+ /* Restore the original environment variable state. */
+ if (value == MLX5_TXDB_UNSET)
+ unsetenv(MLX5_SHUT_UP_BF);
+ else
+ setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
+}
+
+/**
+ * Function API to open IB device using DevX.
+ *
+ * This function calls the Linux glue APIs to open a device.
+ *
+ * @param dev_ctx
+ * Pointer to the context device data structure.
+ * @param dev
+ * Pointer to the generic device.
+ * @param dbnc
+ * Device argument help configure the environment variable.
+ * @param classes
+ * Chosen classes come from device arguments.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_os_devx_open_device(struct mlx5_dev_ctx *dev_ctx, struct rte_device *dev,
+ int dbnc, uint32_t classes)
+{
+ struct ibv_device *ibv;
+ struct ibv_context *ctx = NULL;
+ int dbmap_env;
+
+ ibv = mlx5_os_get_ibv_dev(dev);
+ if (!ibv)
+ return -rte_errno;
+ DRV_LOG(INFO, "Dev information matches for device \"%s\".", ibv->name);
+ /*
+ * Configure environment variable "MLX5_BF_SHUT_UP" before the device
+ * creation. The rdma_core library checks the variable at device
+ * creation and stores the result internally.
+ */
+ dbmap_env = mlx5_config_doorbell_mapping_env(dbnc);
+ /* Try to open IB device with DV. */
+ errno = 0;
+ ctx = mlx5_glue->dv_open_device(ibv);
+ /*
+ * The environment variable is not needed anymore, all device creation
+ * attempts are completed.
+ */
+ mlx5_restore_doorbell_mapping_env(dbmap_env);
+ if (ctx == NULL && classes != MLX5_CLASS_ETH) {
+ DRV_LOG(ERR, "Failed to open IB device \"%s\".", ibv->name);
+ rte_errno = errno ? errno : ENODEV;
+ return -rte_errno;
+ }
+ dev_ctx->ctx = ctx;
+ return 0;
+}
+
+/**
+ * Allocate Protection Domain object and extract its pdn using DV API.
+ *
+ * @param[out] dev_ctx
+ * Pointer to the context device data structure.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_os_pd_create(struct mlx5_dev_ctx *dev_ctx)
+{
+#ifdef HAVE_IBV_FLOW_DV_SUPPORT
+ struct mlx5dv_obj obj;
+ struct mlx5dv_pd pd_info;
+ int ret;
+
+ dev_ctx->pd = mlx5_glue->alloc_pd(dev_ctx->ctx);
+ if (dev_ctx->pd == NULL) {
+ DRV_LOG(ERR, "Failed to allocate PD.");
+ return errno ? -errno : -ENOMEM;
+ }
+ obj.pd.in = dev_ctx->pd;
+ obj.pd.out = &pd_info;
+ ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
+ if (ret != 0) {
+ DRV_LOG(ERR, "Fail to get PD object info.");
+ mlx5_glue->dealloc_pd(dev_ctx->pd);
+ dev_ctx->pd = NULL;
+ return -errno;
+ }
+ dev_ctx->pdn = pd_info.pdn;
+ return 0;
+#else
+ (void)dev_ctx;
+ DRV_LOG(ERR, "Cannot get pdn - no DV support.");
+ return -ENOTSUP;
+#endif /* HAVE_IBV_FLOW_DV_SUPPORT */
+}
+
struct ibv_device *
mlx5_os_get_ibv_device(const struct rte_pci_addr *addr)
{
@@ -423,8 +560,13 @@ mlx5_os_get_ibv_device(const struct rte_pci_addr *addr)
ibv_match = ibv_list[n];
break;
}
- if (ibv_match == NULL)
+ if (ibv_match == NULL) {
+ DRV_LOG(WARNING,
+ "No Verbs device matches PCI device " PCI_PRI_FMT ","
+ " are kernel drivers loaded?",
+ addr->domain, addr->bus, addr->devid, addr->function);
rte_errno = ENOENT;
+ }
mlx5_glue->free_device_list(ibv_list);
return ibv_match;
}
@@ -41,6 +41,20 @@ static inline void mlx5_cpu_id(unsigned int level,
}
#endif
+/*
+ * Device parameter to force doorbell register mapping to non-cahed region
+ * eliminating the extra write memory barrier.
+ */
+#define MLX5_TX_DB_NC "tx_db_nc"
+
+/* Default PMD specific parameter value. */
+#define MLX5_TXDB_UNSET (-1)
+
+/* MLX5_TX_DB_NC supported values. */
+#define MLX5_TXDB_CACHED 0
+#define MLX5_TXDB_NCACHED 1
+#define MLX5_TXDB_HEURISTIC 2
+
RTE_LOG_REGISTER_DEFAULT(mlx5_common_logtype, NOTICE)
/* Head of list of drivers. */
@@ -88,6 +102,83 @@ driver_get(uint32_t class)
return NULL;
}
+/**
+ * Verify and store value for device argument.
+ *
+ * @param[in] key
+ * Key argument to verify.
+ * @param[in] val
+ * Value associated with key.
+ * @param opaque
+ * User data.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static int
+mlx5_common_args_check(const char *key, const char *val, void *opaque)
+{
+ int *dbnc = opaque;
+ signed long tmp;
+
+ errno = 0;
+ tmp = strtol(val, NULL, 0);
+ if (errno) {
+ rte_errno = errno;
+ DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
+ return -rte_errno;
+ }
+ if (strcmp(MLX5_TX_DB_NC, key) == 0) {
+ if (tmp != MLX5_TXDB_CACHED &&
+ tmp != MLX5_TXDB_NCACHED &&
+ tmp != MLX5_TXDB_HEURISTIC) {
+ DRV_LOG(ERR, "Invalid Tx doorbell mapping parameter.");
+ rte_errno = EINVAL;
+ return -rte_errno;
+ }
+ *dbnc = tmp;
+ }
+ return 0;
+}
+
+/**
+ * Parse Tx doorbell mapping parameter.
+ *
+ * @param devargs
+ * Device arguments structure.
+ * @param dbnc
+ * Pointer to get into doorbell mapping parameter.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static int
+mlx5_parse_db_map_arg(struct rte_devargs *devargs, int *dbnc)
+{
+ struct rte_kvargs *kvlist;
+ int ret = 0;
+
+ if (devargs == NULL)
+ return 0;
+ kvlist = rte_kvargs_parse(devargs->args, NULL);
+ if (kvlist == NULL) {
+ rte_errno = EINVAL;
+ return -rte_errno;
+ }
+ if (rte_kvargs_count(kvlist, MLX5_TX_DB_NC)) {
+ ret = rte_kvargs_process(kvlist, MLX5_TX_DB_NC,
+ mlx5_common_args_check, dbnc);
+ if (ret) {
+ rte_errno = EINVAL;
+ rte_kvargs_free(kvlist);
+ return -rte_errno;
+ }
+ }
+ rte_kvargs_free(kvlist);
+ return 0;
+}
+
+
static int
devargs_class_handler(__rte_unused const char *key,
const char *class_names, void *opaque)
@@ -219,6 +310,81 @@ mlx5_dev_to_pci_str(const struct rte_device *dev, char *addr, size_t size)
#endif
}
+/**
+ * Uninitialize context device and release all its resources.
+ *
+ * @param dev_ctx
+ * Pointer to the context device data structure.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+void
+mlx5_dev_ctx_release(struct mlx5_dev_ctx *dev_ctx)
+{
+ if (dev_ctx->pd != NULL) {
+ claim_zero(mlx5_os_dealloc_pd(dev_ctx->pd));
+ dev_ctx->pd = NULL;
+ }
+ if (dev_ctx->ctx != NULL) {
+ claim_zero(mlx5_glue->close_device(dev_ctx->ctx));
+ dev_ctx->ctx = NULL;
+ }
+}
+
+/**
+ * Initialize context device and allocate all its resources.
+ *
+ * @param dev_ctx
+ * Pointer to the context device data structure.
+ * @param dev
+ * Pointer to mlx5 device structure.
+ * @param classes_loaded
+ * Chosen classes come from device arguments.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_dev_ctx_prepare(struct mlx5_dev_ctx *dev_ctx, struct rte_device *dev,
+ uint32_t classes_loaded)
+{
+ int dbnc = MLX5_TXDB_UNSET;
+ int ret;
+
+ dev_ctx->numa_node = dev->numa_node;
+ /*
+ * Parse Tx doorbell mapping parameter. It helps to configure
+ * environment variable "MLX5_BF_SHUT_UP" before the device creation.
+ */
+ ret = mlx5_parse_db_map_arg(dev->devargs, &dbnc);
+ if (ret < 0)
+ return ret;
+ /*
+ * Open device using DevX.
+ * If DevX isn't supported, ctx field remains NULL.
+ */
+ ret = mlx5_os_devx_open_device(dev_ctx, dev, dbnc, classes_loaded);
+ if (ret < 0)
+ return ret;
+ /*
+ * When DevX is not supported and the classes selected by the user can
+ * also work with Verbs, the mlx5_os_devx_open_device function returns
+ * 0 although no device has been created at this time.
+ * Later they will try to create again in Verbs.
+ */
+ if (dev_ctx->ctx == NULL)
+ return 0;
+ /* Allocate Protection Domain object and extract its pdn. */
+ ret = mlx5_os_pd_create(dev_ctx);
+ if (ret)
+ goto error;
+ return ret;
+error:
+ mlx5_dev_ctx_release(dev_ctx);
+ return ret;
+}
+
static void
dev_release(struct mlx5_common_device *dev)
{
@@ -324,6 +324,46 @@ void mlx5_common_init(void);
* from devargs, locating target RDMA device and probing with it.
*/
+/**
+ * Shared device context structure.
+ * Contains HW device objects which belong to same device with multiple drivers.
+ */
+struct mlx5_dev_ctx {
+ void *ctx; /* Verbs/DV/DevX context. */
+ void *pd; /* Protection Domain. */
+ uint32_t pdn; /* Protection Domain Number. */
+ int numa_node; /* Numa node of device. */
+};
+
+/**
+ * Uninitialize context device and release all its resources.
+ *
+ * @param dev_ctx
+ * Pointer to the context device data structure.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+__rte_internal
+void mlx5_dev_ctx_release(struct mlx5_dev_ctx *dev_ctx);
+
+/**
+ * Initialize context device and allocate all its resources.
+ *
+ * @param dev_ctx
+ * Pointer to the context device data structure.
+ * @param dev
+ * Pointer to mlx5 device structure.
+ * @param classes_loaded
+ * Chosen classes come from device arguments.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+__rte_internal
+int mlx5_dev_ctx_prepare(struct mlx5_dev_ctx *dev_ctx, struct rte_device *dev,
+ uint32_t classes_loaded);
+
/**
* Initialization function for the driver called during device probing.
*/
@@ -419,4 +459,12 @@ __rte_internal
bool
mlx5_dev_is_pci(const struct rte_device *dev);
+/* mlx5_common_os.c */
+
+int mlx5_os_devx_open_device(struct mlx5_dev_ctx *dev_ctx,
+ struct rte_device *dev, int dbnc,
+ uint32_t classes);
+int mlx5_os_pd_create(struct mlx5_dev_ctx *dev_ctx);
+
+
#endif /* RTE_PMD_MLX5_COMMON_H_ */
@@ -9,6 +9,9 @@ INTERNAL {
mlx5_common_init;
+ mlx5_dev_ctx_release;
+ mlx5_dev_ctx_prepare;
+
mlx5_common_verbs_reg_mr; # WINDOWS_NO_EXPORT
mlx5_common_verbs_dereg_mr; # WINDOWS_NO_EXPORT
@@ -7,6 +7,7 @@
#include <stdio.h>
#include <rte_mempool.h>
+#include <rte_bus_pci.h>
#include <rte_malloc.h>
#include <rte_errno.h>
@@ -17,7 +18,7 @@
#include "mlx5_malloc.h"
/**
- * Initialization routine for run-time dependency on external lib
+ * Initialization routine for run-time dependency on external lib.
*/
void
mlx5_glue_constructor(void)
@@ -25,7 +26,7 @@ mlx5_glue_constructor(void)
}
/**
- * Allocate PD. Given a devx context object
+ * Allocate PD. Given a DevX context object
* return an mlx5-pd object.
*
* @param[in] ctx
@@ -37,8 +38,8 @@ mlx5_glue_constructor(void)
void *
mlx5_os_alloc_pd(void *ctx)
{
- struct mlx5_pd *ppd = mlx5_malloc(MLX5_MEM_ZERO,
- sizeof(struct mlx5_pd), 0, SOCKET_ID_ANY);
+ struct mlx5_pd *ppd = mlx5_malloc(MLX5_MEM_ZERO, sizeof(struct mlx5_pd),
+ 0, SOCKET_ID_ANY);
if (!ppd)
return NULL;
@@ -60,7 +61,7 @@ mlx5_os_alloc_pd(void *ctx)
* Pointer to mlx5_pd.
*
* @return
- * Zero if pd is released successfully, negative number otherwise.
+ * Zero if pd is released successfully, negative number otherwise.
*/
int
mlx5_os_dealloc_pd(void *pd)
@@ -72,6 +73,202 @@ mlx5_os_dealloc_pd(void *pd)
return 0;
}
+/**
+ * Detect if a devx_device_bdf object has identical DBDF values to the
+ * rte_pci_addr found in bus/pci probing
+ *
+ * @param[in] devx_bdf
+ * Pointer to the devx_device_bdf structure.
+ * @param[in] addr
+ * Pointer to the rte_pci_addr structure.
+ *
+ * @return
+ * 1 on Device match, 0 on mismatch.
+ */
+static int
+mlx5_match_devx_bdf_to_addr(struct devx_device_bdf *devx_bdf,
+ struct rte_pci_addr *addr)
+{
+ if (addr->domain != (devx_bdf->bus_id >> 8) ||
+ addr->bus != (devx_bdf->bus_id & 0xff) ||
+ addr->devid != devx_bdf->dev_id ||
+ addr->function != devx_bdf->fnc_id) {
+ return 0;
+ }
+ return 1;
+}
+
+/**
+ * Detect if a devx_device_bdf object matches the rte_pci_addr
+ * found in bus/pci probing
+ * Compare both the Native/PF BDF and the raw_bdf representing a VF BDF.
+ *
+ * @param[in] devx_bdf
+ * Pointer to the devx_device_bdf structure.
+ * @param[in] addr
+ * Pointer to the rte_pci_addr structure.
+ *
+ * @return
+ * 1 on Device match, 0 on mismatch, rte_errno code on failure.
+ */
+static int
+mlx5_match_devx_devices_to_addr(struct devx_device_bdf *devx_bdf,
+ struct rte_pci_addr *addr)
+{
+ int err;
+ struct devx_device mlx5_dev;
+
+ if (mlx5_match_devx_bdf_to_addr(devx_bdf, addr))
+ return 1;
+ /*
+ * Didn't match on Native/PF BDF, could still match a VF BDF,
+ * check it next.
+ */
+ err = mlx5_glue->query_device(devx_bdf, &mlx5_dev);
+ if (err) {
+ DRV_LOG(ERR, "query_device failed");
+ rte_errno = err;
+ return rte_errno;
+ }
+ if (mlx5_match_devx_bdf_to_addr(&mlx5_dev.raw_bdf, addr))
+ return 1;
+ return 0;
+}
+
+/**
+ * Look for DevX device that match to given rte_device.
+ *
+ * @param dev
+ * Pointer to the generic device.
+ *
+ * @return
+ * A device match on success, NULL otherwise and rte_errno is set.
+ */
+static struct devx_device_bdf *
+mlx5_os_get_devx_device(struct rte_device *dev)
+{
+ int n;
+ struct devx_device_bdf *devx_list;
+ struct devx_device_bdf *orig_devx_list;
+ struct devx_device_bdf *devx_match = NULL;
+ struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev);
+ struct rte_pci_addr *addr = &pci_dev->addr;
+
+ errno = 0;
+ devx_list = mlx5_glue->get_device_list(&n);
+ if (devx_list == NULL) {
+ rte_errno = errno ? errno : ENOSYS;
+ DRV_LOG(ERR, "Cannot list devices, is DevX enabled?");
+ return NULL;
+ }
+ orig_devx_list = devx_list;
+ while (n-- > 0) {
+ int ret = mlx5_match_devx_devices_to_addr(devx_list, addr);
+ if (!ret) {
+ devx_list++;
+ continue;
+ }
+ if (ret != 1) {
+ rte_errno = ret;
+ goto exit;
+ }
+ devx_match = devx_list;
+ break;
+ }
+ if (devx_match == NULL) {
+ /* No device matches, just complain and bail out. */
+ DRV_LOG(WARNING,
+ "No DevX device matches PCI device " PCI_PRI_FMT ","
+ " is DevX Configured?",
+ addr->domain, addr->bus, addr->devid, addr->function);
+ rte_errno = ENOENT;
+ }
+exit:
+ mlx5_glue->free_device_list(orig_devx_list);
+ return devx_match;
+}
+
+/**
+ * Function API open device under Windows.
+ *
+ * This function calls the Windows glue APIs to open a device.
+ *
+ * @param[out] dev_ctx
+ * Pointer to the context device data structure.
+ * @param dev
+ * Pointer to the generic device.
+ * @param dbnc
+ * Device argument help configure the environment variable.
+ * @param classes
+ * Chosen classes come from device arguments.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_os_devx_open_device(struct mlx5_dev_ctx *dev_ctx, struct rte_device *dev,
+ int dbnc, uint32_t classes)
+{
+ RTE_SET_USED(dbnc);
+ struct devx_device_bdf *devx_bdf_dev = NULL;
+ struct mlx5_context *mlx5_ctx;
+
+ if (classes != MLX5_CLASS_ETH) {
+ DRV_LOG(WARNING,
+ "The chosen classes are not supported on Windows.");
+ rte_errno = ENOTSUP;
+ return -rte_errno;
+ }
+ devx_bdf_dev = mlx5_os_get_devx_device(dev);
+ if (devx_bdf_dev == NULL)
+ return -rte_errno;
+ /* Try to open DevX device with DV. */
+ mlx5_ctx = mlx5_glue->open_device(devx_bdf_dev);
+ if (mlx5_ctx) {
+ DRV_LOG(ERR, "Failed to open DevX device.");
+ rte_errno = errno;
+ return -rte_errno;
+ }
+ if (mlx5_glue->query_device(devx_bdf_dev, &mlx5_ctx->mlx5_dev)) {
+ DRV_LOG(ERR, "Failed to query device context fields.");
+ claim_zero(mlx5_glue->close_device(mlx5_ctx));
+ rte_errno = errno;
+ return -rte_errno;
+ }
+ dev_ctx->ctx = mlx5_ctx;
+ return 0;
+}
+
+/**
+ * Allocate Protection Domain object and extract its pdn using DV API.
+ *
+ * @param[out] dev_ctx
+ * Pointer to the context device data structure.
+ *
+ * @return
+ * 0 on success, a negative value otherwise.
+ */
+int
+mlx5_os_pd_create(struct mlx5_dev_ctx *dev_ctx)
+{
+ struct mlx5_pd *pd;
+
+ pd = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*pd), 0, SOCKET_ID_ANY);
+ if (!pd)
+ return -1;
+ struct mlx5_devx_obj *obj = mlx5_devx_cmd_alloc_pd(dev_ctx->ctx);
+ if (!obj) {
+ mlx5_free(pd);
+ return -1;
+ }
+ pd->obj = obj;
+ pd->pdn = obj->id;
+ pd->devx_ctx = dev_ctx->ctx;
+ dev_ctx->pd = pd;
+ dev_ctx->pdn = pd->pdn;
+ return 0;
+}
+
/**
* Register umem.
*