X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;f=drivers%2Fnet%2Fmlx4%2Fmlx4.c;h=640e06f747adaef2c95b2bb506b2d6fe3b8910dd;hb=35b2d13fd6fdcbd191f2a30d74648faeb1186c65;hp=031b1e667fca56aa82225f76a4b2d05620b6e1cd;hpb=5b4c63bdae942206e4c2c1486ea018c1d12acf29;p=dpdk.git diff --git a/drivers/net/mlx4/mlx4.c b/drivers/net/mlx4/mlx4.c index 031b1e667f..640e06f747 100644 --- a/drivers/net/mlx4/mlx4.c +++ b/drivers/net/mlx4/mlx4.c @@ -1,233 +1,280 @@ -/*- - * BSD LICENSE - * - * Copyright 2012 6WIND S.A. - * Copyright 2012 Mellanox - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions - * are met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above copyright - * notice, this list of conditions and the following disclaimer in - * the documentation and/or other materials provided with the - * distribution. - * * Neither the name of 6WIND S.A. nor the names of its - * contributors may be used to endorse or promote products derived - * from this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +/* SPDX-License-Identifier: BSD-3-Clause + * Copyright 2012 6WIND S.A. + * Copyright 2012 Mellanox Technologies, Ltd */ -/* System headers. */ +/** + * @file + * mlx4 driver initialization. + */ + +#include +#include +#include +#include #include +#include #include #include -#include -#include #include -#include +#include #include -#include -#include -#include -#include +/* Verbs headers do not support -pedantic. */ +#ifdef PEDANTIC +#pragma GCC diagnostic ignored "-Wpedantic" +#endif +#include +#ifdef PEDANTIC +#pragma GCC diagnostic error "-Wpedantic" +#endif + +#include +#include #include -#include #include -#include -#include -#include +#include +#include +#include #include -#include #include -#include +#include +#include +#include -/* PMD headers. */ #include "mlx4.h" +#include "mlx4_glue.h" #include "mlx4_flow.h" +#include "mlx4_mr.h" #include "mlx4_rxtx.h" #include "mlx4_utils.h" +static const char *MZ_MLX4_PMD_SHARED_DATA = "mlx4_pmd_shared_data"; + +/* Shared memory between primary and secondary processes. */ +struct mlx4_shared_data *mlx4_shared_data; + +/* Spinlock for mlx4_shared_data allocation. */ +static rte_spinlock_t mlx4_shared_data_lock = RTE_SPINLOCK_INITIALIZER; + +/* Process local data for secondary processes. */ +static struct mlx4_local_data mlx4_local_data; + /** Configuration structure for device arguments. */ struct mlx4_conf { struct { uint32_t present; /**< Bit-field for existing ports. */ uint32_t enabled; /**< Bit-field for user-enabled ports. */ } ports; + int mr_ext_memseg_en; + /** Whether memseg should be extended for MR creation. */ }; /* Available parameters list. */ const char *pmd_mlx4_init_params[] = { MLX4_PMD_PORT_KVARG, + MLX4_MR_EXT_MEMSEG_EN_KVARG, NULL, }; -/* Device configuration. */ +static void mlx4_dev_stop(struct rte_eth_dev *dev); /** - * DPDK callback for Ethernet device configuration. - * - * Prepare the driver for a given number of TX and RX queues. + * Initialize shared data between primary and secondary process. * - * @param dev - * Pointer to Ethernet device structure. + * A memzone is reserved by primary process and secondary processes attach to + * the memzone. * * @return - * 0 on success, negative errno value otherwise and rte_errno is set. + * 0 on success, a negative errno value otherwise and rte_errno is set. */ static int -mlx4_dev_configure(struct rte_eth_dev *dev) +mlx4_init_shared_data(void) { - struct priv *priv = dev->data->dev_private; - unsigned int rxqs_n = dev->data->nb_rx_queues; - unsigned int txqs_n = dev->data->nb_tx_queues; - - priv->rxqs = (void *)dev->data->rx_queues; - priv->txqs = (void *)dev->data->tx_queues; - if (txqs_n != priv->txqs_n) { - INFO("%p: TX queues number update: %u -> %u", - (void *)dev, priv->txqs_n, txqs_n); - priv->txqs_n = txqs_n; - } - if (rxqs_n != priv->rxqs_n) { - INFO("%p: Rx queues number update: %u -> %u", - (void *)dev, priv->rxqs_n, rxqs_n); - priv->rxqs_n = rxqs_n; + const struct rte_memzone *mz; + int ret = 0; + + rte_spinlock_lock(&mlx4_shared_data_lock); + if (mlx4_shared_data == NULL) { + if (rte_eal_process_type() == RTE_PROC_PRIMARY) { + /* Allocate shared memory. */ + mz = rte_memzone_reserve(MZ_MLX4_PMD_SHARED_DATA, + sizeof(*mlx4_shared_data), + SOCKET_ID_ANY, 0); + if (mz == NULL) { + ERROR("Cannot allocate mlx4 shared data\n"); + ret = -rte_errno; + goto error; + } + mlx4_shared_data = mz->addr; + memset(mlx4_shared_data, 0, sizeof(*mlx4_shared_data)); + rte_spinlock_init(&mlx4_shared_data->lock); + } else { + /* Lookup allocated shared memory. */ + mz = rte_memzone_lookup(MZ_MLX4_PMD_SHARED_DATA); + if (mz == NULL) { + ERROR("Cannot attach mlx4 shared data\n"); + ret = -rte_errno; + goto error; + } + mlx4_shared_data = mz->addr; + memset(&mlx4_local_data, 0, sizeof(mlx4_local_data)); + } } - return 0; +error: + rte_spinlock_unlock(&mlx4_shared_data_lock); + return ret; } -struct mlx4_check_mempool_data { - int ret; - char *start; - char *end; -}; - -/* Called by mlx4_check_mempool() when iterating the memory chunks. */ -static void mlx4_check_mempool_cb(struct rte_mempool *mp, - void *opaque, struct rte_mempool_memhdr *memhdr, - unsigned mem_idx) +#ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS +/** + * Verbs callback to allocate a memory. This function should allocate the space + * according to the size provided residing inside a huge page. + * Please note that all allocation must respect the alignment from libmlx4 + * (i.e. currently sysconf(_SC_PAGESIZE)). + * + * @param[in] size + * The size in bytes of the memory to allocate. + * @param[in] data + * A pointer to the callback data. + * + * @return + * Allocated buffer, NULL otherwise and rte_errno is set. + */ +static void * +mlx4_alloc_verbs_buf(size_t size, void *data) { - struct mlx4_check_mempool_data *data = opaque; + struct mlx4_priv *priv = data; + void *ret; + size_t alignment = sysconf(_SC_PAGESIZE); + unsigned int socket = SOCKET_ID_ANY; - (void)mp; - (void)mem_idx; - /* It already failed, skip the next chunks. */ - if (data->ret != 0) - return; - /* It is the first chunk. */ - if (data->start == NULL && data->end == NULL) { - data->start = memhdr->addr; - data->end = data->start + memhdr->len; - return; - } - if (data->end == memhdr->addr) { - data->end += memhdr->len; - return; - } - if (data->start == (char *)memhdr->addr + memhdr->len) { - data->start -= memhdr->len; - return; + if (priv->verbs_alloc_ctx.type == MLX4_VERBS_ALLOC_TYPE_TX_QUEUE) { + const struct txq *txq = priv->verbs_alloc_ctx.obj; + + socket = txq->socket; + } else if (priv->verbs_alloc_ctx.type == + MLX4_VERBS_ALLOC_TYPE_RX_QUEUE) { + const struct rxq *rxq = priv->verbs_alloc_ctx.obj; + + socket = rxq->socket; } - /* Error, mempool is not virtually contigous. */ - data->ret = -1; + assert(data != NULL); + ret = rte_malloc_socket(__func__, size, alignment, socket); + if (!ret && size) + rte_errno = ENOMEM; + return ret; +} + +/** + * Verbs callback to free a memory. + * + * @param[in] ptr + * A pointer to the memory to free. + * @param[in] data + * A pointer to the callback data. + */ +static void +mlx4_free_verbs_buf(void *ptr, void *data __rte_unused) +{ + assert(data != NULL); + rte_free(ptr); } +#endif /** - * Check if a mempool can be used: it must be virtually contiguous. + * Initialize process private data structure. * - * @param[in] mp - * Pointer to memory pool. - * @param[out] start - * Pointer to the start address of the mempool virtual memory area - * @param[out] end - * Pointer to the end address of the mempool virtual memory area + * @param dev + * Pointer to Ethernet device structure. * * @return - * 0 on success (mempool is virtually contiguous), -1 on error. + * 0 on success, a negative errno value otherwise and rte_errno is set. */ -static int mlx4_check_mempool(struct rte_mempool *mp, uintptr_t *start, - uintptr_t *end) +static int +mlx4_proc_priv_init(struct rte_eth_dev *dev) { - struct mlx4_check_mempool_data data; + struct mlx4_proc_priv *ppriv; + size_t ppriv_size; - memset(&data, 0, sizeof(data)); - rte_mempool_mem_iter(mp, mlx4_check_mempool_cb, &data); - *start = (uintptr_t)data.start; - *end = (uintptr_t)data.end; - return data.ret; + /* + * UAR register table follows the process private structure. BlueFlame + * registers for Tx queues are stored in the table. + */ + ppriv_size = sizeof(struct mlx4_proc_priv) + + dev->data->nb_tx_queues * sizeof(void *); + ppriv = rte_malloc_socket("mlx4_proc_priv", ppriv_size, + RTE_CACHE_LINE_SIZE, dev->device->numa_node); + if (!ppriv) { + rte_errno = ENOMEM; + return -rte_errno; + } + ppriv->uar_table_sz = ppriv_size; + dev->process_private = ppriv; + return 0; } /** - * Register mempool as a memory region. + * Un-initialize process private data structure. * - * @param pd - * Pointer to protection domain. - * @param mp - * Pointer to memory pool. + * @param dev + * Pointer to Ethernet device structure. + */ +static void +mlx4_proc_priv_uninit(struct rte_eth_dev *dev) +{ + if (!dev->process_private) + return; + rte_free(dev->process_private); + dev->process_private = NULL; +} + +/** + * DPDK callback for Ethernet device configuration. + * + * @param dev + * Pointer to Ethernet device structure. * * @return - * Memory region pointer, NULL in case of error and rte_errno is set. + * 0 on success, negative errno value otherwise and rte_errno is set. */ -struct ibv_mr * -mlx4_mp2mr(struct ibv_pd *pd, struct rte_mempool *mp) +static int +mlx4_dev_configure(struct rte_eth_dev *dev) { - const struct rte_memseg *ms = rte_eal_get_physmem_layout(); - uintptr_t start; - uintptr_t end; - unsigned int i; - struct ibv_mr *mr; + struct mlx4_priv *priv = dev->data->dev_private; + struct rte_flow_error error; + int ret; - if (mlx4_check_mempool(mp, &start, &end) != 0) { - rte_errno = EINVAL; - ERROR("mempool %p: not virtually contiguous", - (void *)mp); - return NULL; - } - DEBUG("mempool %p area start=%p end=%p size=%zu", - (void *)mp, (void *)start, (void *)end, - (size_t)(end - start)); - /* Round start and end to page boundary if found in memory segments. */ - for (i = 0; (i < RTE_MAX_MEMSEG) && (ms[i].addr != NULL); ++i) { - uintptr_t addr = (uintptr_t)ms[i].addr; - size_t len = ms[i].len; - unsigned int align = ms[i].hugepage_sz; - - if ((start > addr) && (start < addr + len)) - start = RTE_ALIGN_FLOOR(start, align); - if ((end > addr) && (end < addr + len)) - end = RTE_ALIGN_CEIL(end, align); - } - DEBUG("mempool %p using start=%p end=%p size=%zu for MR", - (void *)mp, (void *)start, (void *)end, - (size_t)(end - start)); - mr = ibv_reg_mr(pd, - (void *)start, - end - start, - IBV_ACCESS_LOCAL_WRITE); - if (!mr) - rte_errno = errno ? errno : EINVAL; - return mr; + /* Prepare internal flow rules. */ + ret = mlx4_flow_sync(priv, &error); + if (ret) { + ERROR("cannot set up internal flow rules (code %d, \"%s\")," + " flow error type %d, cause %p, message: %s", + -ret, strerror(-ret), error.type, error.cause, + error.message ? error.message : "(unspecified)"); + goto exit; + } + ret = mlx4_intr_install(priv); + if (ret) { + ERROR("%p: interrupt handler installation failed", + (void *)dev); + goto exit; + } + ret = mlx4_proc_priv_init(dev); + if (ret) { + ERROR("%p: process private data allocation failed", + (void *)dev); + goto exit; + } +exit: + return ret; } /** * DPDK callback to start the device. * - * Simulate device start by attaching all configured flows. + * Simulate device start by initializing common RSS resources and attaching + * all configured flows. * * @param dev * Pointer to Ethernet device structure. @@ -238,33 +285,46 @@ mlx4_mp2mr(struct ibv_pd *pd, struct rte_mempool *mp) static int mlx4_dev_start(struct rte_eth_dev *dev) { - struct priv *priv = dev->data->dev_private; + struct mlx4_priv *priv = dev->data->dev_private; + struct rte_flow_error error; int ret; if (priv->started) return 0; DEBUG("%p: attaching configured flows to all RX queues", (void *)dev); priv->started = 1; - ret = mlx4_mac_addr_add(priv); - if (ret) + ret = mlx4_rss_init(priv); + if (ret) { + ERROR("%p: cannot initialize RSS resources: %s", + (void *)dev, strerror(-ret)); goto err; - ret = mlx4_intr_install(priv); + } +#ifndef NDEBUG + mlx4_mr_dump_dev(dev); +#endif + ret = mlx4_rxq_intr_enable(priv); if (ret) { ERROR("%p: interrupt handler installation failed", (void *)dev); goto err; } - ret = mlx4_priv_flow_start(priv); + ret = mlx4_flow_sync(priv, &error); if (ret) { - ERROR("%p: flow start failed: %s", - (void *)dev, strerror(ret)); + ERROR("%p: cannot attach flow rules (code %d, \"%s\")," + " flow error type %d, cause %p, message: %s", + (void *)dev, + -ret, strerror(-ret), error.type, error.cause, + error.message ? error.message : "(unspecified)"); goto err; } + rte_wmb(); + dev->tx_pkt_burst = mlx4_tx_burst; + dev->rx_pkt_burst = mlx4_rx_burst; + /* Enable datapath on secondary process. */ + mlx4_mp_req_start_rxtx(dev); return 0; err: - /* Rollback. */ - mlx4_mac_addr_del(priv); - priv->started = 0; + mlx4_dev_stop(dev); return ret; } @@ -279,15 +339,20 @@ err: static void mlx4_dev_stop(struct rte_eth_dev *dev) { - struct priv *priv = dev->data->dev_private; + struct mlx4_priv *priv = dev->data->dev_private; if (!priv->started) return; DEBUG("%p: detaching flows from all RX queues", (void *)dev); priv->started = 0; - mlx4_priv_flow_stop(priv); - mlx4_intr_uninstall(priv); - mlx4_mac_addr_del(priv); + dev->tx_pkt_burst = mlx4_tx_burst_removed; + dev->rx_pkt_burst = mlx4_rx_burst_removed; + rte_wmb(); + /* Disable datapath on secondary process. */ + mlx4_mp_req_stop_rxtx(dev); + mlx4_flow_sync(priv, NULL); + mlx4_rxq_intr_disable(priv); + mlx4_rss_deinit(priv); } /** @@ -301,106 +366,35 @@ mlx4_dev_stop(struct rte_eth_dev *dev) static void mlx4_dev_close(struct rte_eth_dev *dev) { - struct priv *priv = dev->data->dev_private; - void *tmp; + struct mlx4_priv *priv = dev->data->dev_private; unsigned int i; - if (priv == NULL) - return; DEBUG("%p: closing device \"%s\"", (void *)dev, ((priv->ctx != NULL) ? priv->ctx->device->name : "")); - mlx4_mac_addr_del(priv); - /* - * Prevent crashes when queues are still in use. This is unfortunately - * still required for DPDK 1.3 because some programs (such as testpmd) - * never release them before closing the device. - */ dev->rx_pkt_burst = mlx4_rx_burst_removed; dev->tx_pkt_burst = mlx4_tx_burst_removed; - if (priv->rxqs != NULL) { - /* XXX race condition if mlx4_rx_burst() is still running. */ - usleep(1000); - for (i = 0; (i != priv->rxqs_n); ++i) { - tmp = (*priv->rxqs)[i]; - if (tmp == NULL) - continue; - (*priv->rxqs)[i] = NULL; - mlx4_rxq_cleanup(tmp); - rte_free(tmp); - } - priv->rxqs_n = 0; - priv->rxqs = NULL; - } - if (priv->txqs != NULL) { - /* XXX race condition if mlx4_tx_burst() is still running. */ - usleep(1000); - for (i = 0; (i != priv->txqs_n); ++i) { - tmp = (*priv->txqs)[i]; - if (tmp == NULL) - continue; - (*priv->txqs)[i] = NULL; - mlx4_txq_cleanup(tmp); - rte_free(tmp); - } - priv->txqs_n = 0; - priv->txqs = NULL; - } + rte_wmb(); + /* Disable datapath on secondary process. */ + mlx4_mp_req_stop_rxtx(dev); + mlx4_flow_clean(priv); + mlx4_rss_deinit(priv); + for (i = 0; i != dev->data->nb_rx_queues; ++i) + mlx4_rx_queue_release(dev->data->rx_queues[i]); + for (i = 0; i != dev->data->nb_tx_queues; ++i) + mlx4_tx_queue_release(dev->data->tx_queues[i]); + mlx4_proc_priv_uninit(dev); + mlx4_mr_release(dev); if (priv->pd != NULL) { assert(priv->ctx != NULL); - claim_zero(ibv_dealloc_pd(priv->pd)); - claim_zero(ibv_close_device(priv->ctx)); + claim_zero(mlx4_glue->dealloc_pd(priv->pd)); + claim_zero(mlx4_glue->close_device(priv->ctx)); } else assert(priv->ctx == NULL); mlx4_intr_uninstall(priv); memset(priv, 0, sizeof(*priv)); } -const struct rte_flow_ops mlx4_flow_ops = { - .validate = mlx4_flow_validate, - .create = mlx4_flow_create, - .destroy = mlx4_flow_destroy, - .flush = mlx4_flow_flush, - .query = NULL, - .isolate = mlx4_flow_isolate, -}; - -/** - * Manage filter operations. - * - * @param dev - * Pointer to Ethernet device structure. - * @param filter_type - * Filter type. - * @param filter_op - * Operation to perform. - * @param arg - * Pointer to operation-specific structure. - * - * @return - * 0 on success, negative errno value otherwise and rte_errno is set. - */ -static int -mlx4_dev_filter_ctrl(struct rte_eth_dev *dev, - enum rte_filter_type filter_type, - enum rte_filter_op filter_op, - void *arg) -{ - switch (filter_type) { - case RTE_ETH_FILTER_GENERIC: - if (filter_op != RTE_ETH_FILTER_GET) - break; - *(const void **)arg = &mlx4_flow_ops; - return 0; - default: - ERROR("%p: filter type (%d) not supported", - (void *)dev, filter_type); - break; - } - rte_errno = ENOTSUP; - return -rte_errno; -} - static const struct eth_dev_ops mlx4_dev_ops = { .dev_configure = mlx4_dev_configure, .dev_start = mlx4_dev_start, @@ -409,9 +403,20 @@ static const struct eth_dev_ops mlx4_dev_ops = { .dev_set_link_up = mlx4_dev_set_link_up, .dev_close = mlx4_dev_close, .link_update = mlx4_link_update, + .promiscuous_enable = mlx4_promiscuous_enable, + .promiscuous_disable = mlx4_promiscuous_disable, + .allmulticast_enable = mlx4_allmulticast_enable, + .allmulticast_disable = mlx4_allmulticast_disable, + .mac_addr_remove = mlx4_mac_addr_remove, + .mac_addr_add = mlx4_mac_addr_add, + .mac_addr_set = mlx4_mac_addr_set, + .set_mc_addr_list = mlx4_set_mc_addr_list, .stats_get = mlx4_stats_get, .stats_reset = mlx4_stats_reset, + .fw_version_get = mlx4_fw_version_get, .dev_infos_get = mlx4_dev_infos_get, + .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get, + .vlan_filter_set = mlx4_vlan_filter_set, .rx_queue_setup = mlx4_rx_queue_setup, .tx_queue_setup = mlx4_tx_queue_setup, .rx_queue_release = mlx4_rx_queue_release, @@ -419,9 +424,18 @@ static const struct eth_dev_ops mlx4_dev_ops = { .flow_ctrl_get = mlx4_flow_ctrl_get, .flow_ctrl_set = mlx4_flow_ctrl_set, .mtu_set = mlx4_mtu_set, - .filter_ctrl = mlx4_dev_filter_ctrl, + .filter_ctrl = mlx4_filter_ctrl, .rx_queue_intr_enable = mlx4_rx_intr_enable, .rx_queue_intr_disable = mlx4_rx_intr_disable, + .is_removed = mlx4_is_removed, +}; + +/* Available operations from secondary process. */ +static const struct eth_dev_ops mlx4_dev_sec_ops = { + .stats_get = mlx4_stats_get, + .stats_reset = mlx4_stats_reset, + .fw_version_get = mlx4_fw_version_get, + .dev_infos_get = mlx4_dev_infos_get, }; /** @@ -502,7 +516,7 @@ mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf) return -rte_errno; } if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) { - uint32_t ports = rte_log2_u32(conf->ports.present); + uint32_t ports = rte_log2_u32(conf->ports.present + 1); if (tmp >= ports) { ERROR("port index %lu outside range [0,%" PRIu32 ")", @@ -515,6 +529,8 @@ mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf) return -rte_errno; } conf->ports.enabled |= 1 << tmp; + } else if (strcmp(MLX4_MR_EXT_MEMSEG_EN_KVARG, key) == 0) { + conf->mr_ext_memseg_en = !!tmp; } else { rte_errno = EINVAL; WARN("%s: unknown parameter", key); @@ -550,10 +566,10 @@ mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf) } /* Process parameters. */ for (i = 0; pmd_mlx4_init_params[i]; ++i) { - arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG); + arg_count = rte_kvargs_count(kvlist, pmd_mlx4_init_params[i]); while (arg_count-- > 0) { ret = rte_kvargs_process(kvlist, - MLX4_PMD_PORT_KVARG, + pmd_mlx4_init_params[i], (int (*)(const char *, const char *, void *)) @@ -568,8 +584,147 @@ free_kvlist: return ret; } +/** + * Interpret RSS capabilities reported by device. + * + * This function returns the set of usable Verbs RSS hash fields, kernel + * quirks taken into account. + * + * @param ctx + * Verbs context. + * @param pd + * Verbs protection domain. + * @param device_attr_ex + * Extended device attributes to interpret. + * + * @return + * Usable RSS hash fields mask in Verbs format. + */ +static uint64_t +mlx4_hw_rss_sup(struct ibv_context *ctx, struct ibv_pd *pd, + struct ibv_device_attr_ex *device_attr_ex) +{ + uint64_t hw_rss_sup = device_attr_ex->rss_caps.rx_hash_fields_mask; + struct ibv_cq *cq = NULL; + struct ibv_wq *wq = NULL; + struct ibv_rwq_ind_table *ind = NULL; + struct ibv_qp *qp = NULL; + + if (!hw_rss_sup) { + WARN("no RSS capabilities reported; disabling support for UDP" + " RSS and inner VXLAN RSS"); + return IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4 | + IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6 | + IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP; + } + if (!(hw_rss_sup & IBV_RX_HASH_INNER)) + return hw_rss_sup; + /* + * Although reported as supported, missing code in some Linux + * versions (v4.15, v4.16) prevents the creation of hash QPs with + * inner capability. + * + * There is no choice but to attempt to instantiate a temporary RSS + * context in order to confirm its support. + */ + cq = mlx4_glue->create_cq(ctx, 1, NULL, NULL, 0); + wq = cq ? mlx4_glue->create_wq + (ctx, + &(struct ibv_wq_init_attr){ + .wq_type = IBV_WQT_RQ, + .max_wr = 1, + .max_sge = 1, + .pd = pd, + .cq = cq, + }) : NULL; + ind = wq ? mlx4_glue->create_rwq_ind_table + (ctx, + &(struct ibv_rwq_ind_table_init_attr){ + .log_ind_tbl_size = 0, + .ind_tbl = &wq, + .comp_mask = 0, + }) : NULL; + qp = ind ? mlx4_glue->create_qp_ex + (ctx, + &(struct ibv_qp_init_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 = pd, + .rwq_ind_tbl = ind, + .rx_hash_conf = { + .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ, + .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE, + .rx_hash_key = mlx4_rss_hash_key_default, + .rx_hash_fields_mask = hw_rss_sup, + }, + }) : NULL; + if (!qp) { + WARN("disabling unusable inner RSS capability due to kernel" + " quirk"); + hw_rss_sup &= ~IBV_RX_HASH_INNER; + } else { + claim_zero(mlx4_glue->destroy_qp(qp)); + } + if (ind) + claim_zero(mlx4_glue->destroy_rwq_ind_table(ind)); + if (wq) + claim_zero(mlx4_glue->destroy_wq(wq)); + if (cq) + claim_zero(mlx4_glue->destroy_cq(cq)); + return hw_rss_sup; +} + static struct rte_pci_driver mlx4_driver; +/** + * PMD global initialization. + * + * Independent from individual device, this function initializes global + * per-PMD data structures distinguishing primary and secondary processes. + * Hence, each initialization is called once per a process. + * + * @return + * 0 on success, a negative errno value otherwise and rte_errno is set. + */ +static int +mlx4_init_once(void) +{ + struct mlx4_shared_data *sd; + struct mlx4_local_data *ld = &mlx4_local_data; + + if (mlx4_init_shared_data()) + return -rte_errno; + sd = mlx4_shared_data; + assert(sd); + rte_spinlock_lock(&sd->lock); + switch (rte_eal_process_type()) { + case RTE_PROC_PRIMARY: + if (sd->init_done) + break; + LIST_INIT(&sd->mem_event_cb_list); + rte_rwlock_init(&sd->mem_event_rwlock); + rte_mem_event_callback_register("MLX4_MEM_EVENT_CB", + mlx4_mr_mem_event_cb, NULL); + mlx4_mp_init_primary(); + sd->init_done = true; + break; + case RTE_PROC_SECONDARY: + if (ld->init_done) + break; + mlx4_mp_init_secondary(); + ++sd->secondary_cnt; + ld->init_done = true; + break; + default: + break; + } + rte_spinlock_unlock(&sd->lock); + return 0; +} + /** * DPDK callback to register a PCI device. * @@ -592,15 +747,23 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev) int err = 0; struct ibv_context *attr_ctx = NULL; struct ibv_device_attr device_attr; + struct ibv_device_attr_ex device_attr_ex; struct mlx4_conf conf = { .ports.present = 0, + .mr_ext_memseg_en = 1, }; unsigned int vf; int i; (void)pci_drv; + err = mlx4_init_once(); + if (err) { + ERROR("unable to init PMD global data: %s", + strerror(rte_errno)); + return -rte_errno; + } assert(pci_drv == &mlx4_driver); - list = ibv_get_device_list(&i); + list = mlx4_glue->get_device_list(&i); if (list == NULL) { rte_errno = errno; assert(rte_errno); @@ -629,12 +792,12 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev) PCI_DEVICE_ID_MELLANOX_CONNECTX3VF); INFO("PCI information matches, using device \"%s\" (VF: %s)", list[i]->name, (vf ? "true" : "false")); - attr_ctx = ibv_open_device(list[i]); + attr_ctx = mlx4_glue->open_device(list[i]); err = errno; break; } if (attr_ctx == NULL) { - ibv_free_device_list(list); + mlx4_glue->free_device_list(list); switch (err) { case 0: rte_errno = ENODEV; @@ -651,65 +814,121 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev) } ibv_dev = list[i]; DEBUG("device opened"); - if (ibv_query_device(attr_ctx, &device_attr)) { - rte_errno = ENODEV; + if (mlx4_glue->query_device(attr_ctx, &device_attr)) { + err = ENODEV; goto error; } INFO("%u port(s) detected", device_attr.phys_port_cnt); conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1; if (mlx4_args(pci_dev->device.devargs, &conf)) { ERROR("failed to process device arguments"); - rte_errno = EINVAL; + err = EINVAL; goto error; } /* Use all ports when none are defined */ if (!conf.ports.enabled) conf.ports.enabled = conf.ports.present; + /* Retrieve extended device attributes. */ + if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) { + err = ENODEV; + goto error; + } + assert(device_attr.max_sge >= MLX4_MAX_SGE); for (i = 0; i < device_attr.phys_port_cnt; i++) { uint32_t port = i + 1; /* ports are indexed from one */ struct ibv_context *ctx = NULL; struct ibv_port_attr port_attr; struct ibv_pd *pd = NULL; - struct priv *priv = NULL; + struct mlx4_priv *priv = NULL; struct rte_eth_dev *eth_dev = NULL; - struct ether_addr mac; + struct rte_ether_addr mac; + char name[RTE_ETH_NAME_MAX_LEN]; /* If port is not enabled, skip. */ if (!(conf.ports.enabled & (1 << i))) continue; DEBUG("using port %u", port); - ctx = ibv_open_device(ibv_dev); + ctx = mlx4_glue->open_device(ibv_dev); if (ctx == NULL) { - rte_errno = ENODEV; + err = ENODEV; goto port_error; } + snprintf(name, sizeof(name), "%s port %u", + mlx4_glue->get_device_name(ibv_dev), port); + if (rte_eal_process_type() == RTE_PROC_SECONDARY) { + eth_dev = rte_eth_dev_attach_secondary(name); + if (eth_dev == NULL) { + ERROR("can not attach rte ethdev"); + rte_errno = ENOMEM; + err = rte_errno; + goto error; + } + priv = eth_dev->data->dev_private; + if (!priv->verbs_alloc_ctx.enabled) { + ERROR("secondary process is not supported" + " due to lack of external allocator" + " from Verbs"); + rte_errno = ENOTSUP; + err = rte_errno; + goto error; + } + eth_dev->device = &pci_dev->device; + eth_dev->dev_ops = &mlx4_dev_sec_ops; + err = mlx4_proc_priv_init(eth_dev); + if (err) + goto error; + /* Receive command fd from primary process. */ + err = mlx4_mp_req_verbs_cmd_fd(eth_dev); + if (err < 0) { + err = rte_errno; + goto error; + } + /* Remap UAR for Tx queues. */ + err = mlx4_tx_uar_init_secondary(eth_dev, err); + if (err) { + err = rte_errno; + goto error; + } + /* + * Ethdev pointer is still required as input since + * the primary device is not accessible from the + * secondary process. + */ + eth_dev->tx_pkt_burst = mlx4_tx_burst; + eth_dev->rx_pkt_burst = mlx4_rx_burst; + claim_zero(mlx4_glue->close_device(ctx)); + rte_eth_copy_pci_info(eth_dev, pci_dev); + rte_eth_dev_probing_finish(eth_dev); + continue; + } /* Check port status. */ - err = ibv_query_port(ctx, port, &port_attr); + err = mlx4_glue->query_port(ctx, port, &port_attr); if (err) { - rte_errno = err; - ERROR("port query failed: %s", strerror(rte_errno)); + err = ENODEV; + ERROR("port query failed: %s", strerror(err)); goto port_error; } if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) { - rte_errno = ENOTSUP; + err = ENOTSUP; ERROR("port %d is not configured in Ethernet mode", port); goto port_error; } if (port_attr.state != IBV_PORT_ACTIVE) DEBUG("port %d is not active: \"%s\" (%d)", - port, ibv_port_state_str(port_attr.state), + port, mlx4_glue->port_state_str(port_attr.state), port_attr.state); /* Make asynchronous FD non-blocking to handle interrupts. */ - if (mlx4_fd_set_non_blocking(ctx->async_fd) < 0) { + err = mlx4_fd_set_non_blocking(ctx->async_fd); + if (err) { ERROR("cannot make asynchronous FD non-blocking: %s", - strerror(rte_errno)); + strerror(err)); goto port_error; } /* Allocate protection domain. */ - pd = ibv_alloc_pd(ctx); + pd = mlx4_glue->alloc_pd(ctx); if (pd == NULL) { - rte_errno = ENOMEM; + err = ENOMEM; ERROR("PD allocation failure"); goto port_error; } @@ -718,7 +937,7 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev) sizeof(*priv), RTE_CACHE_LINE_SIZE); if (priv == NULL) { - rte_errno = ENOMEM; + err = ENOMEM; ERROR("priv allocation failure"); goto port_error; } @@ -726,12 +945,45 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev) priv->device_attr = device_attr; priv->port = port; priv->pd = pd; - priv->mtu = ETHER_MTU; + priv->mtu = RTE_ETHER_MTU; priv->vf = vf; + priv->hw_csum = !!(device_attr.device_cap_flags & + IBV_DEVICE_RAW_IP_CSUM); + DEBUG("checksum offloading is %ssupported", + (priv->hw_csum ? "" : "not ")); + /* Only ConnectX-3 Pro supports tunneling. */ + priv->hw_csum_l2tun = + priv->hw_csum && + (device_attr.vendor_part_id == + PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO); + DEBUG("L2 tunnel checksum offloads are %ssupported", + priv->hw_csum_l2tun ? "" : "not "); + priv->hw_rss_sup = mlx4_hw_rss_sup(priv->ctx, priv->pd, + &device_attr_ex); + DEBUG("supported RSS hash fields mask: %016" PRIx64, + priv->hw_rss_sup); + priv->hw_rss_max_qps = + device_attr_ex.rss_caps.max_rwq_indirection_table_size; + DEBUG("MAX RSS queues %d", priv->hw_rss_max_qps); + priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps & + IBV_RAW_PACKET_CAP_SCATTER_FCS); + DEBUG("FCS stripping toggling is %ssupported", + priv->hw_fcs_strip ? "" : "not "); + priv->tso = + ((device_attr_ex.tso_caps.max_tso > 0) && + (device_attr_ex.tso_caps.supported_qpts & + (1 << IBV_QPT_RAW_PACKET))); + if (priv->tso) + priv->tso_max_payload_sz = + device_attr_ex.tso_caps.max_tso; + DEBUG("TSO is %ssupported", + priv->tso ? "" : "not "); + priv->mr_ext_memseg_en = conf.mr_ext_memseg_en; /* Configure the first MAC address by default. */ - if (mlx4_get_mac(priv, &mac.addr_bytes)) { + err = mlx4_get_mac(priv, &mac.addr_bytes); + if (err) { ERROR("cannot get MAC address, is mlx4_en loaded?" - " (rte_errno: %s)", strerror(rte_errno)); + " (error: %s)", strerror(err)); goto port_error; } INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x", @@ -740,9 +992,7 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev) mac.addr_bytes[2], mac.addr_bytes[3], mac.addr_bytes[4], mac.addr_bytes[5]); /* Register MAC address. */ - priv->mac = mac; - if (mlx4_mac_addr_add(priv)) - goto port_error; + priv->mac[0] = mac; #ifndef NDEBUG { char ifname[IF_NAMESIZE]; @@ -757,24 +1007,16 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev) /* Get actual MTU if possible. */ mlx4_mtu_get(priv, &priv->mtu); DEBUG("port %u MTU is %u", priv->port, priv->mtu); - /* from rte_ethdev.c */ - { - char name[RTE_ETH_NAME_MAX_LEN]; - - snprintf(name, sizeof(name), "%s port %u", - ibv_get_device_name(ibv_dev), port); - eth_dev = rte_eth_dev_allocate(name); - } + eth_dev = rte_eth_dev_allocate(name); if (eth_dev == NULL) { + err = ENOMEM; ERROR("can not allocate rte ethdev"); - rte_errno = ENOMEM; goto port_error; } eth_dev->data->dev_private = priv; - eth_dev->data->mac_addrs = &priv->mac; + eth_dev->data->mac_addrs = priv->mac; eth_dev->device = &pci_dev->device; rte_eth_copy_pci_info(eth_dev, pci_dev); - eth_dev->device->driver = &mlx4_driver.driver; /* Initialize local interrupt handle for current port. */ priv->intr_handle = (struct rte_intr_handle){ .fd = -1, @@ -793,28 +1035,64 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev) * handled by rte_intr_rx_ctl(). */ eth_dev->intr_handle = &priv->intr_handle; - priv->dev = eth_dev; + priv->dev_data = eth_dev->data; eth_dev->dev_ops = &mlx4_dev_ops; - eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE; +#ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS + /* Hint libmlx4 to use PMD allocator for data plane resources */ + struct mlx4dv_ctx_allocators alctr = { + .alloc = &mlx4_alloc_verbs_buf, + .free = &mlx4_free_verbs_buf, + .data = priv, + }; + err = mlx4_glue->dv_set_context_attr + (ctx, MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS, + (void *)((uintptr_t)&alctr)); + if (err) + WARN("Verbs external allocator is not supported"); + else + priv->verbs_alloc_ctx.enabled = 1; +#endif /* Bring Ethernet device up. */ DEBUG("forcing Ethernet interface up"); - mlx4_dev_set_link_up(priv->dev); + mlx4_dev_set_link_up(eth_dev); /* Update link status once if waiting for LSC. */ if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC) mlx4_link_update(eth_dev, 0); + /* + * Once the device is added to the list of memory event + * callback, its global MR cache table cannot be expanded + * on the fly because of deadlock. If it overflows, lookup + * should be done by searching MR list linearly, which is slow. + */ + err = mlx4_mr_btree_init(&priv->mr.cache, + MLX4_MR_BTREE_CACHE_N * 2, + eth_dev->device->numa_node); + if (err) { + /* rte_errno is already set. */ + goto port_error; + } + /* Add device to memory callback list. */ + rte_rwlock_write_lock(&mlx4_shared_data->mem_event_rwlock); + LIST_INSERT_HEAD(&mlx4_shared_data->mem_event_cb_list, + priv, mem_event_cb); + rte_rwlock_write_unlock(&mlx4_shared_data->mem_event_rwlock); + rte_eth_dev_probing_finish(eth_dev); continue; port_error: rte_free(priv); + if (eth_dev != NULL) + eth_dev->data->dev_private = NULL; if (pd) - claim_zero(ibv_dealloc_pd(pd)); + claim_zero(mlx4_glue->dealloc_pd(pd)); if (ctx) - claim_zero(ibv_close_device(ctx)); - if (eth_dev) + claim_zero(mlx4_glue->close_device(ctx)); + if (eth_dev != NULL) { + /* mac_addrs must not be freed because part of dev_private */ + eth_dev->data->mac_addrs = NULL; rte_eth_dev_release_port(eth_dev); + } break; } - if (i == device_attr.phys_port_cnt) - return 0; /* * XXX if something went wrong in the loop above, there is a resource * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as @@ -823,11 +1101,12 @@ port_error: */ error: if (attr_ctx) - claim_zero(ibv_close_device(attr_ctx)); + claim_zero(mlx4_glue->close_device(attr_ctx)); if (list) - ibv_free_device_list(list); - assert(rte_errno >= 0); - return -rte_errno; + mlx4_glue->free_device_list(list); + if (err) + rte_errno = err; + return -err; } static const struct rte_pci_id mlx4_pci_id_map[] = { @@ -858,13 +1137,147 @@ static struct rte_pci_driver mlx4_driver = { RTE_PCI_DRV_INTR_RMV, }; +#ifdef RTE_IBVERBS_LINK_DLOPEN + +/** + * Suffix RTE_EAL_PMD_PATH with "-glue". + * + * This function performs a sanity check on RTE_EAL_PMD_PATH before + * suffixing its last component. + * + * @param buf[out] + * Output buffer, should be large enough otherwise NULL is returned. + * @param size + * Size of @p out. + * + * @return + * Pointer to @p buf or @p NULL in case suffix cannot be appended. + */ +static char * +mlx4_glue_path(char *buf, size_t size) +{ + static const char *const bad[] = { "/", ".", "..", NULL }; + const char *path = RTE_EAL_PMD_PATH; + size_t len = strlen(path); + size_t off; + int i; + + while (len && path[len - 1] == '/') + --len; + for (off = len; off && path[off - 1] != '/'; --off) + ; + for (i = 0; bad[i]; ++i) + if (!strncmp(path + off, bad[i], (int)(len - off))) + goto error; + i = snprintf(buf, size, "%.*s-glue", (int)len, path); + if (i == -1 || (size_t)i >= size) + goto error; + return buf; +error: + ERROR("unable to append \"-glue\" to last component of" + " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\")," + " please re-configure DPDK"); + return NULL; +} + +/** + * Initialization routine for run-time dependency on rdma-core. + */ +static int +mlx4_glue_init(void) +{ + char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")]; + const char *path[] = { + /* + * A basic security check is necessary before trusting + * MLX4_GLUE_PATH, which may override RTE_EAL_PMD_PATH. + */ + (geteuid() == getuid() && getegid() == getgid() ? + getenv("MLX4_GLUE_PATH") : NULL), + /* + * When RTE_EAL_PMD_PATH is set, use its glue-suffixed + * variant, otherwise let dlopen() look up libraries on its + * own. + */ + (*RTE_EAL_PMD_PATH ? + mlx4_glue_path(glue_path, sizeof(glue_path)) : ""), + }; + unsigned int i = 0; + void *handle = NULL; + void **sym; + const char *dlmsg; + + while (!handle && i != RTE_DIM(path)) { + const char *end; + size_t len; + int ret; + + if (!path[i]) { + ++i; + continue; + } + end = strpbrk(path[i], ":;"); + if (!end) + end = path[i] + strlen(path[i]); + len = end - path[i]; + ret = 0; + do { + char name[ret + 1]; + + ret = snprintf(name, sizeof(name), "%.*s%s" MLX4_GLUE, + (int)len, path[i], + (!len || *(end - 1) == '/') ? "" : "/"); + if (ret == -1) + break; + if (sizeof(name) != (size_t)ret + 1) + continue; + DEBUG("looking for rdma-core glue as \"%s\"", name); + handle = dlopen(name, RTLD_LAZY); + break; + } while (1); + path[i] = end + 1; + if (!*end) + ++i; + } + if (!handle) { + rte_errno = EINVAL; + dlmsg = dlerror(); + if (dlmsg) + WARN("cannot load glue library: %s", dlmsg); + goto glue_error; + } + sym = dlsym(handle, "mlx4_glue"); + if (!sym || !*sym) { + rte_errno = EINVAL; + dlmsg = dlerror(); + if (dlmsg) + ERROR("cannot resolve glue symbol: %s", dlmsg); + goto glue_error; + } + mlx4_glue = *sym; + return 0; +glue_error: + if (handle) + dlclose(handle); + WARN("cannot initialize PMD due to missing run-time" + " dependency on rdma-core libraries (libibverbs," + " libmlx4)"); + return -rte_errno; +} + +#endif + /** * Driver initialization routine. */ -RTE_INIT(rte_mlx4_pmd_init); -static void -rte_mlx4_pmd_init(void) +RTE_INIT(rte_mlx4_pmd_init) { + /* + * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we + * want to get success errno value in case of calling them + * when the device was removed. + */ + setenv("MLX4_DEVICE_FATAL_CLEANUP", "1", 1); /* * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use * huge pages. Calling ibv_fork_init() during init allows @@ -872,7 +1285,26 @@ rte_mlx4_pmd_init(void) * using this PMD, which is not supported in forked processes. */ setenv("RDMAV_HUGEPAGES_SAFE", "1", 1); - ibv_fork_init(); +#ifdef RTE_IBVERBS_LINK_DLOPEN + if (mlx4_glue_init()) + return; + assert(mlx4_glue); +#endif +#ifndef NDEBUG + /* Glue structure must not contain any NULL pointers. */ + { + unsigned int i; + + for (i = 0; i != sizeof(*mlx4_glue) / sizeof(void *); ++i) + assert(((const void *const *)mlx4_glue)[i]); + } +#endif + if (strcmp(mlx4_glue->version, MLX4_GLUE_VERSION)) { + ERROR("rdma-core glue \"%s\" mismatch: \"%s\" is required", + mlx4_glue->version, MLX4_GLUE_VERSION); + return; + } + mlx4_glue->fork_init(); rte_pci_register(&mlx4_driver); }