From: Viacheslav Ovsiienko Date: Fri, 8 Nov 2019 15:07:50 +0000 (+0000) Subject: net/mlx5: control transmit doorbell register mapping X-Git-Url: http://git.droids-corp.org/?a=commitdiff_plain;h=8409a28573d39ef15da795cadae88825373f1326;p=dpdk.git net/mlx5: control transmit doorbell register mapping The rdma core library can map doorbell register in two ways, depending on the environment variable "MLX5_SHUT_UP_BF": - as regular cached memory, the variable is either missing or set to zero. This type of mapping may cause the significant doorbell register writing latency and requires explicit memory write barrier to mitigate this issue and prevent write combining. - as non-cached memory, the variable is present and set to not "0" value. This type of mapping may cause performance impact under heavy loading conditions but the explicit write memory barrier is not required and it may improve core performance. The new devarg is introduced "tx_db_nc", if this parameter is set to zero, the doorbell register is forced to be mapped to cached memory and requires explicit memory barrier after writing to. If "tx_db_nc" is set to non-zero value the doorbell will be mapped as non-cached memory, not requiring the memory barrier. If "tx_db_nc" is missing the behaviour will be defined by presence of "MLX5_SHUT_UP_BF" in environment. If variable is missed the default value zero will be set for ARM64 hosts and one for others. In run time the code checks the mapping type and provides the memory barrier after writing to tx doorbell register if it is needed. The mapping type is extracted directly from the uar_mmap_offset field in the queue properties. Fixes: 18a1c20044c0 ("net/mlx5: implement Tx burst template") Cc: stable@dpdk.org Signed-off-by: Viacheslav Ovsiienko Acked-by: Matan Azrad --- diff --git a/doc/guides/nics/mlx5.rst b/doc/guides/nics/mlx5.rst index 3651e82f3c..5fd313c8d5 100644 --- a/doc/guides/nics/mlx5.rst +++ b/doc/guides/nics/mlx5.rst @@ -552,6 +552,29 @@ Run-time configuration Also, if minimal data inlining is requested by non-zero ``txq_inline_min`` option or reported by the NIC, the eMPW feature is disengaged. +- ``tx_db_nc`` parameter [int] + + The rdma core library can map doorbell register in two ways, depending on the + environment variable "MLX5_SHUT_UP_BF": + + - As regular cached memory, if the variable is either missing or set to zero. + - As non-cached memory, if the variable is present and set to not "0" value. + + The type of mapping may slightly affect the Tx performance, the optimal choice + is strongly relied on the host architecture and should be deduced practically. + + If ``tx_db_nc`` is either omitted or set to zero, the doorbell is forced to be + mapped to regular memory, the PMD will perform the extra write memory barrier + after writing to doorbell, it might increase the needed CPU clocks per packet + to send, but latency might be improved. + + If ``tx_db_nc`` is set to not zero, the doorbell is forced to be mapped to + non cached memory, the PMD will not perform the extra write memory barrier + after writing to doorbell, on some architectures it might improve the + performance. + + The default ``tx_db_nc`` value is zero ARM64 hosts and one for others. + - ``tx_vec_en`` parameter [int] A nonzero value enables Tx vector on ConnectX-5, ConnectX-6, ConnectX-6 DX diff --git a/drivers/net/mlx5/Makefile b/drivers/net/mlx5/Makefile index d01fa731fa..5b79631cf5 100644 --- a/drivers/net/mlx5/Makefile +++ b/drivers/net/mlx5/Makefile @@ -205,6 +205,11 @@ mlx5_autoconf.h.new: $(RTE_SDK)/buildtools/auto-config-h.sh infiniband/mlx5dv.h \ func mlx5dv_dr_action_create_flow_meter \ $(AUTOCONF_OUTPUT) + $Q sh -- '$<' '$@' \ + HAVE_MLX5DV_MMAP_GET_NC_PAGES_CMD \ + infiniband/mlx5dv.h \ + enum MLX5_MMAP_GET_NC_PAGES_CMD \ + $(AUTOCONF_OUTPUT) $Q sh -- '$<' '$@' \ HAVE_ETHTOOL_LINK_MODE_25G \ /usr/include/linux/ethtool.h \ diff --git a/drivers/net/mlx5/meson.build b/drivers/net/mlx5/meson.build index 511f5b7ca6..05fadf608b 100644 --- a/drivers/net/mlx5/meson.build +++ b/drivers/net/mlx5/meson.build @@ -134,6 +134,8 @@ if build 'mlx5dv_dr_action_create_dest_devx_tir' ], [ 'HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER', 'infiniband/mlx5dv.h', 'mlx5dv_dr_action_create_flow_meter' ], + [ 'HAVE_MLX5DV_MMAP_GET_NC_PAGES_CMD', 'infiniband/mlx5dv.h', + 'MLX5_MMAP_GET_NC_PAGES_CMD' ], [ 'HAVE_MLX5DV_DR', 'infiniband/mlx5dv.h', 'MLX5DV_DR_DOMAIN_TYPE_NIC_RX' ], [ 'HAVE_MLX5DV_DR_ESWITCH', 'infiniband/mlx5dv.h', diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c index 9a2c711e57..276087d775 100644 --- a/drivers/net/mlx5/mlx5.c +++ b/drivers/net/mlx5/mlx5.c @@ -95,6 +95,12 @@ /* Device parameter to enable multi-packet send WQEs. */ #define MLX5_TXQ_MPW_EN "txq_mpw_en" +/* + * 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" + /* * Device parameter to include 2 dsegs in the title WQEBB. * Deprecated, ignored. @@ -421,6 +427,37 @@ mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused) } #endif /* HAVE_IBV_FLOW_DV_SUPPORT */ +static int +mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config) +{ + char *env; + int value; + + 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_ARG_UNSET; + if (config->dbnc == MLX5_ARG_UNSET) + setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1); + else + setenv(MLX5_SHUT_UP_BF, config->dbnc ? "1" : "0", 1); + return value; +} + +static void +mlx5_restore_doorbell_mapping_env(const struct mlx5_dev_config *config, + int value) +{ + assert(rte_eal_process_type() == RTE_PROC_PRIMARY); + if (config->dbnc == MLX5_ARG_UNSET) + return; + /* Restore the original environment variable state. */ + if (value == MLX5_ARG_UNSET) + unsetenv(MLX5_SHUT_UP_BF); + else + setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1); +} + /** * Allocate shared IB device context. If there is multiport device the * master and representors will share this context, if there is single @@ -434,22 +471,26 @@ mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused) * * @param[in] spawn * Pointer to the IB device attributes (name, port, etc). + * @param[in] config + * Pointer to device configuration structure. * * @return * Pointer to mlx5_ibv_shared object on success, * otherwise NULL and rte_errno is set. */ static struct mlx5_ibv_shared * -mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn) +mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn, + const struct mlx5_dev_config *config) { struct mlx5_ibv_shared *sh; + int dbmap_env; int err = 0; uint32_t i; #ifdef HAVE_IBV_FLOW_DV_SUPPORT struct mlx5_devx_tis_attr tis_attr = { 0 }; #endif -assert(spawn); + assert(spawn); /* Secondary process should not create the shared context. */ assert(rte_eal_process_type() == RTE_PROC_PRIMARY); pthread_mutex_lock(&mlx5_ibv_list_mutex); @@ -472,16 +513,31 @@ assert(spawn); rte_errno = ENOMEM; goto exit; } + /* + * 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(config); /* Try to open IB device with DV first, then usual Verbs. */ errno = 0; sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev); if (sh->ctx) { sh->devx = 1; DRV_LOG(DEBUG, "DevX is supported"); + /* The device is created, no need for environment. */ + mlx5_restore_doorbell_mapping_env(config, dbmap_env); } else { + /* The environment variable is still configured. */ sh->ctx = mlx5_glue->open_device(spawn->ibv_dev); + err = errno ? errno : ENODEV; + /* + * The environment variable is not needed anymore, + * all device creation attempts are completed. + */ + mlx5_restore_doorbell_mapping_env(config, dbmap_env); if (!sh->ctx) { - err = errno ? errno : ENODEV; goto error; } DRV_LOG(DEBUG, "DevX is NOT supported"); @@ -1300,6 +1356,8 @@ mlx5_args_check(const char *key, const char *val, void *opaque) DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key); } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) { config->mps = !!tmp; + } else if (strcmp(MLX5_TX_DB_NC, key) == 0) { + config->dbnc = !!tmp; } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) { DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key); } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) { @@ -1373,6 +1431,7 @@ mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs) MLX5_TXQ_MPW_EN, MLX5_TXQ_MPW_HDR_DSEG_EN, MLX5_TXQ_MAX_INLINE_LEN, + MLX5_TX_DB_NC, MLX5_TX_VEC_EN, MLX5_RX_VEC_EN, MLX5_L3_VXLAN_EN, @@ -1938,7 +1997,20 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev, eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev); return eth_dev; } - sh = mlx5_alloc_shared_ibctx(spawn); + /* + * Some parameters ("tx_db_nc" in particularly) are needed in + * advance to create dv/verbs device context. We proceed the + * devargs here to get ones, and later proceed devargs again + * to override some hardware settings. + */ + err = mlx5_args(&config, dpdk_dev->devargs); + if (err) { + err = rte_errno; + DRV_LOG(ERR, "failed to process device arguments: %s", + strerror(rte_errno)); + goto error; + } + sh = mlx5_alloc_shared_ibctx(spawn, &config); if (!sh) return NULL; config.devx = sh->devx; @@ -2180,13 +2252,8 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev, } own_domain_id = 1; } - err = mlx5_args(&config, dpdk_dev->devargs); - if (err) { - err = rte_errno; - DRV_LOG(ERR, "failed to process device arguments: %s", - strerror(rte_errno)); - goto error; - } + /* Override some values set by hardware configuration. */ + mlx5_args(&config, dpdk_dev->devargs); err = mlx5_dev_check_sibling_config(priv, &config); if (err) goto error; @@ -3031,6 +3098,7 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused, dev_config = (struct mlx5_dev_config){ .hw_padding = 0, .mps = MLX5_ARG_UNSET, + .dbnc = MLX5_ARG_UNSET, .rx_vec_en = 1, .txq_inline_max = MLX5_ARG_UNSET, .txq_inline_min = MLX5_ARG_UNSET, diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h index e8148ced6f..1a92c103d4 100644 --- a/drivers/net/mlx5/mlx5.h +++ b/drivers/net/mlx5/mlx5.h @@ -267,6 +267,7 @@ struct mlx5_dev_config { /* Rx queue count threshold to enable MPRQ. */ } mprq; /* Configurations for Multi-Packet RQ. */ int mps; /* Multi-packet send supported mode. */ + int dbnc; /* Skip doorbell register write barrier. */ unsigned int flow_prio; /* Number of flow priorities. */ enum modify_reg flow_mreg_c[MLX5_MREG_C_NUM]; /* Availibility of mreg_c's. */ diff --git a/drivers/net/mlx5/mlx5_defs.h b/drivers/net/mlx5/mlx5_defs.h index 0ef532f6e8..03060aa087 100644 --- a/drivers/net/mlx5/mlx5_defs.h +++ b/drivers/net/mlx5/mlx5_defs.h @@ -123,6 +123,22 @@ #define MLX5_UAR_PAGE_NUM_MAX 64 #define MLX5_UAR_PAGE_NUM_MASK ((MLX5_UAR_PAGE_NUM_MAX) - 1) +/* Fields of memory mapping type in offset parameter of mmap() */ +#define MLX5_UAR_MMAP_CMD_SHIFT 8 +#define MLX5_UAR_MMAP_CMD_MASK 0xff + +/* 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 + +#ifndef HAVE_MLX5DV_MMAP_GET_NC_PAGES_CMD +#define MLX5_MMAP_GET_NC_PAGES_CMD 3 +#endif + /* Log 2 of the default number of strides per WQE for Multi-Packet RQ. */ #define MLX5_MPRQ_STRIDE_NUM_N 6U diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c index 8bc0542530..1ea5960df6 100644 --- a/drivers/net/mlx5/mlx5_rxtx.c +++ b/drivers/net/mlx5/mlx5_rxtx.c @@ -4754,8 +4754,23 @@ enter_send_single: * to improve latencies. The pure software related data treatment * can be completed after doorbell. Tx CQEs for this SQ are * processed in this thread only by the polling. + * + * The rdma core library can map doorbell register in two ways, + * depending on the environment variable "MLX5_SHUT_UP_BF": + * + * - as regular cached memory, the variable is either missing or + * set to zero. This type of mapping may cause the significant + * doorbell register writing latency and requires explicit + * memory write barrier to mitigate this issue and prevent + * write combining. + * + * - as non-cached memory, the variable is present and set to + * not "0" value. This type of mapping may cause performance + * impact under heavy loading conditions but the explicit write + * memory barrier is not required and it may improve core + * performance. */ - mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, 0); + mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, !txq->db_nc); /* Not all of the mbufs may be stored into elts yet. */ part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy; if (!MLX5_TXOFF_CONFIG(INLINE) && part) { diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h index 559d225a11..88c50aa868 100644 --- a/drivers/net/mlx5/mlx5_rxtx.h +++ b/drivers/net/mlx5/mlx5_rxtx.h @@ -287,6 +287,7 @@ struct mlx5_txq_data { /* When set TX offload for tunneled packets are supported. */ uint16_t swp_en:1; /* Whether SW parser is enabled. */ uint16_t vlan_en:1; /* VLAN insertion in WQE is supported. */ + uint16_t db_nc:1; /* Doorbell mapped to non-cached region. */ uint16_t inlen_send; /* Ordinary send data inline size. */ uint16_t inlen_empw; /* eMPW max packet size to inline. */ uint16_t inlen_mode; /* Minimal data length to inline. */ diff --git a/drivers/net/mlx5/mlx5_txq.c b/drivers/net/mlx5/mlx5_txq.c index 97991f0c04..a0d61644cb 100644 --- a/drivers/net/mlx5/mlx5_txq.c +++ b/drivers/net/mlx5/mlx5_txq.c @@ -18,6 +18,7 @@ #pragma GCC diagnostic ignored "-Wpedantic" #endif #include +#include #ifdef PEDANTIC #pragma GCC diagnostic error "-Wpedantic" #endif @@ -301,6 +302,28 @@ mlx5_tx_queue_release(void *dpdk_txq) } } +/** + * Configure the doorbell register non-cached attribute. + * + * @param txq_ctrl + * Pointer to Tx queue control structure. + * @param page_size + * Systme page size + */ +static void +txq_uar_ncattr_init(struct mlx5_txq_ctrl *txq_ctrl, size_t page_size) +{ + unsigned int cmd; + + txq_ctrl->txq.db_nc = 0; + /* Check the doorbell register mapping type. */ + cmd = txq_ctrl->uar_mmap_offset / page_size; + cmd >>= MLX5_UAR_MMAP_CMD_SHIFT; + cmd &= MLX5_UAR_MMAP_CMD_MASK; + if (cmd == MLX5_MMAP_GET_NC_PAGES_CMD) + txq_ctrl->txq.db_nc = 1; +} + /** * Initialize Tx UAR registers for primary process. * @@ -312,9 +335,9 @@ txq_uar_init(struct mlx5_txq_ctrl *txq_ctrl) { struct mlx5_priv *priv = txq_ctrl->priv; struct mlx5_proc_priv *ppriv = MLX5_PROC_PRIV(PORT_ID(priv)); + const size_t page_size = sysconf(_SC_PAGESIZE); #ifndef RTE_ARCH_64 unsigned int lock_idx; - const size_t page_size = sysconf(_SC_PAGESIZE); #endif if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD) @@ -322,6 +345,7 @@ txq_uar_init(struct mlx5_txq_ctrl *txq_ctrl) assert(rte_eal_process_type() == RTE_PROC_PRIMARY); assert(ppriv); ppriv->uar_table[txq_ctrl->txq.idx] = txq_ctrl->bf_reg; + txq_uar_ncattr_init(txq_ctrl, page_size); #ifndef RTE_ARCH_64 /* Assign an UAR lock according to UAR page number */ lock_idx = (txq_ctrl->uar_mmap_offset / page_size) & @@ -375,6 +399,7 @@ txq_uar_init_secondary(struct mlx5_txq_ctrl *txq_ctrl, int fd) } addr = RTE_PTR_ADD(addr, offset); ppriv->uar_table[txq->idx] = addr; + txq_uar_ncattr_init(txq_ctrl, page_size); return 0; }