diff options
Diffstat (limited to 'drivers/infiniband/hw/mlx4')
-rw-r--r-- | drivers/infiniband/hw/mlx4/ah.c | 8 | ||||
-rw-r--r-- | drivers/infiniband/hw/mlx4/cq.c | 10 | ||||
-rw-r--r-- | drivers/infiniband/hw/mlx4/main.c | 23 | ||||
-rw-r--r-- | drivers/infiniband/hw/mlx4/mcg.c | 1 | ||||
-rw-r--r-- | drivers/infiniband/hw/mlx4/mlx4_ib.h | 19 | ||||
-rw-r--r-- | drivers/infiniband/hw/mlx4/mr.c | 284 | ||||
-rw-r--r-- | drivers/infiniband/hw/mlx4/qp.c | 26 |
7 files changed, 325 insertions, 46 deletions
diff --git a/drivers/infiniband/hw/mlx4/ah.c b/drivers/infiniband/hw/mlx4/ah.c index 538c46a73248..6dee4fdc5d67 100644 --- a/drivers/infiniband/hw/mlx4/ah.c +++ b/drivers/infiniband/hw/mlx4/ah.c @@ -92,12 +92,10 @@ static struct ib_ah *create_iboe_ah(struct ib_pd *pd, int ret; memcpy(&in6, grh->dgid.raw, sizeof(in6)); - if (rdma_is_multicast_addr(&in6)) { + if (rdma_is_multicast_addr(&in6)) is_mcast = 1; - rdma_get_mcast_mac(&in6, ah->av.eth.mac); - } else { - memcpy(ah->av.eth.mac, ah_attr->roce.dmac, ETH_ALEN); - } + + memcpy(ah->av.eth.mac, ah_attr->roce.dmac, ETH_ALEN); ret = ib_get_cached_gid(pd->device, rdma_ah_get_port_num(ah_attr), grh->sgid_index, &sgid, &gid_attr); if (ret) diff --git a/drivers/infiniband/hw/mlx4/cq.c b/drivers/infiniband/hw/mlx4/cq.c index cab796341697..bf4f14a1b4fc 100644 --- a/drivers/infiniband/hw/mlx4/cq.c +++ b/drivers/infiniband/hw/mlx4/cq.c @@ -140,14 +140,18 @@ static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *cont { int err; int cqe_size = dev->dev->caps.cqe_size; + int shift; + int n; *umem = ib_umem_get(context, buf_addr, cqe * cqe_size, IB_ACCESS_LOCAL_WRITE, 1); if (IS_ERR(*umem)) return PTR_ERR(*umem); - err = mlx4_mtt_init(dev->dev, ib_umem_page_count(*umem), - (*umem)->page_shift, &buf->mtt); + n = ib_umem_page_count(*umem); + shift = mlx4_ib_umem_calc_optimal_mtt_size(*umem, 0, &n); + err = mlx4_mtt_init(dev->dev, n, shift, &buf->mtt); + if (err) goto err_buf; @@ -768,11 +772,13 @@ repoll: switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) { case MLX4_OPCODE_RDMA_WRITE_IMM: wc->wc_flags |= IB_WC_WITH_IMM; + /* fall through */ case MLX4_OPCODE_RDMA_WRITE: wc->opcode = IB_WC_RDMA_WRITE; break; case MLX4_OPCODE_SEND_IMM: wc->wc_flags |= IB_WC_WITH_IMM; + /* fall through */ case MLX4_OPCODE_SEND: case MLX4_OPCODE_SEND_INVAL: wc->opcode = IB_WC_SEND; diff --git a/drivers/infiniband/hw/mlx4/main.c b/drivers/infiniband/hw/mlx4/main.c index c636842c5be0..8c8a16791a3f 100644 --- a/drivers/infiniband/hw/mlx4/main.c +++ b/drivers/infiniband/hw/mlx4/main.c @@ -563,6 +563,9 @@ static int mlx4_ib_query_device(struct ib_device *ibdev, props->max_wq_type_rq = props->max_qp; } + props->cq_caps.max_cq_moderation_count = MLX4_MAX_CQ_COUNT; + props->cq_caps.max_cq_moderation_period = MLX4_MAX_CQ_PERIOD; + if (!mlx4_is_slave(dev->dev)) err = mlx4_get_internal_clock_params(dev->dev, &clock_params); @@ -581,6 +584,23 @@ static int mlx4_ib_query_device(struct ib_device *ibdev, sizeof(struct mlx4_wqe_data_seg); } + if (uhw->outlen >= resp.response_length + sizeof(resp.rss_caps)) { + resp.response_length += sizeof(resp.rss_caps); + if (props->rss_caps.supported_qpts) { + resp.rss_caps.rx_hash_function = + MLX4_IB_RX_HASH_FUNC_TOEPLITZ; + resp.rss_caps.rx_hash_fields_mask = + MLX4_IB_RX_HASH_SRC_IPV4 | + MLX4_IB_RX_HASH_DST_IPV4 | + MLX4_IB_RX_HASH_SRC_IPV6 | + MLX4_IB_RX_HASH_DST_IPV6 | + MLX4_IB_RX_HASH_SRC_PORT_TCP | + MLX4_IB_RX_HASH_DST_PORT_TCP | + MLX4_IB_RX_HASH_SRC_PORT_UDP | + MLX4_IB_RX_HASH_DST_PORT_UDP; + } + } + if (uhw->outlen) { err = ib_copy_to_udata(uhw, &resp, resp.response_length); if (err) @@ -2733,6 +2753,9 @@ static void *mlx4_ib_add(struct mlx4_dev *dev) ibdev->ib_dev.get_dev_fw_str = get_fw_ver_str; ibdev->ib_dev.disassociate_ucontext = mlx4_ib_disassociate_ucontext; + ibdev->ib_dev.uverbs_ex_cmd_mask |= + (1ull << IB_USER_VERBS_EX_CMD_MODIFY_CQ); + if ((dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS) && ((mlx4_ib_port_link_layer(&ibdev->ib_dev, 1) == IB_LINK_LAYER_ETHERNET) || diff --git a/drivers/infiniband/hw/mlx4/mcg.c b/drivers/infiniband/hw/mlx4/mcg.c index 70eb9f917303..81ffc007e0a1 100644 --- a/drivers/infiniband/hw/mlx4/mcg.c +++ b/drivers/infiniband/hw/mlx4/mcg.c @@ -944,6 +944,7 @@ int mlx4_ib_mcg_multiplex_handler(struct ib_device *ibdev, int port, switch (sa_mad->mad_hdr.method) { case IB_MGMT_METHOD_SET: may_create = 1; + /* fall through */ case IB_SA_METHOD_DELETE: req = kzalloc(sizeof *req, GFP_KERNEL); if (!req) diff --git a/drivers/infiniband/hw/mlx4/mlx4_ib.h b/drivers/infiniband/hw/mlx4/mlx4_ib.h index 1fa19820355a..e14919c15b06 100644 --- a/drivers/infiniband/hw/mlx4/mlx4_ib.h +++ b/drivers/infiniband/hw/mlx4/mlx4_ib.h @@ -47,6 +47,7 @@ #include <linux/mlx4/device.h> #include <linux/mlx4/doorbell.h> #include <linux/mlx4/qp.h> +#include <linux/mlx4/cq.h> #define MLX4_IB_DRV_NAME "mlx4_ib" @@ -644,12 +645,18 @@ enum query_device_resp_mask { QUERY_DEVICE_RESP_MASK_TIMESTAMP = 1UL << 0, }; +struct mlx4_ib_rss_caps { + __u64 rx_hash_fields_mask; /* enum mlx4_rx_hash_fields */ + __u8 rx_hash_function; /* enum mlx4_rx_hash_function_flags */ + __u8 reserved[7]; +}; + struct mlx4_uverbs_ex_query_device_resp { - __u32 comp_mask; - __u32 response_length; - __u64 hca_core_clock_offset; - __u32 max_inl_recv_sz; - __u32 reserved; + __u32 comp_mask; + __u32 response_length; + __u64 hca_core_clock_offset; + __u32 max_inl_recv_sz; + struct mlx4_ib_rss_caps rss_caps; }; static inline struct mlx4_ib_dev *to_mdev(struct ib_device *ibdev) @@ -929,5 +936,7 @@ struct ib_rwq_ind_table struct ib_rwq_ind_table_init_attr *init_attr, struct ib_udata *udata); int mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *wq_ind_table); +int mlx4_ib_umem_calc_optimal_mtt_size(struct ib_umem *umem, u64 start_va, + int *num_of_mtts); #endif /* MLX4_IB_H */ diff --git a/drivers/infiniband/hw/mlx4/mr.c b/drivers/infiniband/hw/mlx4/mr.c index e6f77f63da75..313bfb9ccb71 100644 --- a/drivers/infiniband/hw/mlx4/mr.c +++ b/drivers/infiniband/hw/mlx4/mr.c @@ -87,50 +87,286 @@ err_free: return ERR_PTR(err); } +enum { + MLX4_MAX_MTT_SHIFT = 31 +}; + +static int mlx4_ib_umem_write_mtt_block(struct mlx4_ib_dev *dev, + struct mlx4_mtt *mtt, + u64 mtt_size, u64 mtt_shift, u64 len, + u64 cur_start_addr, u64 *pages, + int *start_index, int *npages) +{ + u64 cur_end_addr = cur_start_addr + len; + u64 cur_end_addr_aligned = 0; + u64 mtt_entries; + int err = 0; + int k; + + len += (cur_start_addr & (mtt_size - 1ULL)); + cur_end_addr_aligned = round_up(cur_end_addr, mtt_size); + len += (cur_end_addr_aligned - cur_end_addr); + if (len & (mtt_size - 1ULL)) { + pr_warn("write_block: len %llx is not aligned to mtt_size %llx\n", + len, mtt_size); + return -EINVAL; + } + + mtt_entries = (len >> mtt_shift); + + /* + * Align the MTT start address to the mtt_size. + * Required to handle cases when the MR starts in the middle of an MTT + * record. Was not required in old code since the physical addresses + * provided by the dma subsystem were page aligned, which was also the + * MTT size. + */ + cur_start_addr = round_down(cur_start_addr, mtt_size); + /* A new block is started ... */ + for (k = 0; k < mtt_entries; ++k) { + pages[*npages] = cur_start_addr + (mtt_size * k); + (*npages)++; + /* + * Be friendly to mlx4_write_mtt() and pass it chunks of + * appropriate size. + */ + if (*npages == PAGE_SIZE / sizeof(u64)) { + err = mlx4_write_mtt(dev->dev, mtt, *start_index, + *npages, pages); + if (err) + return err; + + (*start_index) += *npages; + *npages = 0; + } + } + + return 0; +} + +static inline u64 alignment_of(u64 ptr) +{ + return ilog2(ptr & (~(ptr - 1))); +} + +static int mlx4_ib_umem_calc_block_mtt(u64 next_block_start, + u64 current_block_end, + u64 block_shift) +{ + /* Check whether the alignment of the new block is aligned as well as + * the previous block. + * Block address must start with zeros till size of entity_size. + */ + if ((next_block_start & ((1ULL << block_shift) - 1ULL)) != 0) + /* + * It is not as well aligned as the previous block-reduce the + * mtt size accordingly. Here we take the last right bit which + * is 1. + */ + block_shift = alignment_of(next_block_start); + + /* + * Check whether the alignment of the end of previous block - is it + * aligned as well as the start of the block + */ + if (((current_block_end) & ((1ULL << block_shift) - 1ULL)) != 0) + /* + * It is not as well aligned as the start of the block - + * reduce the mtt size accordingly. + */ + block_shift = alignment_of(current_block_end); + + return block_shift; +} + int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt, struct ib_umem *umem) { u64 *pages; - int i, k, entry; - int n; - int len; + u64 len = 0; int err = 0; + u64 mtt_size; + u64 cur_start_addr = 0; + u64 mtt_shift; + int start_index = 0; + int npages = 0; struct scatterlist *sg; + int i; pages = (u64 *) __get_free_page(GFP_KERNEL); if (!pages) return -ENOMEM; - i = n = 0; + mtt_shift = mtt->page_shift; + mtt_size = 1ULL << mtt_shift; - for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) { - len = sg_dma_len(sg) >> mtt->page_shift; - for (k = 0; k < len; ++k) { - pages[i++] = sg_dma_address(sg) + - (k << umem->page_shift); - /* - * Be friendly to mlx4_write_mtt() and - * pass it chunks of appropriate size. - */ - if (i == PAGE_SIZE / sizeof (u64)) { - err = mlx4_write_mtt(dev->dev, mtt, n, - i, pages); - if (err) - goto out; - n += i; - i = 0; - } + for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) { + if (cur_start_addr + len == sg_dma_address(sg)) { + /* still the same block */ + len += sg_dma_len(sg); + continue; } + /* + * A new block is started ... + * If len is malaligned, write an extra mtt entry to cover the + * misaligned area (round up the division) + */ + err = mlx4_ib_umem_write_mtt_block(dev, mtt, mtt_size, + mtt_shift, len, + cur_start_addr, + pages, &start_index, + &npages); + if (err) + goto out; + + cur_start_addr = sg_dma_address(sg); + len = sg_dma_len(sg); } - if (i) - err = mlx4_write_mtt(dev->dev, mtt, n, i, pages); + /* Handle the last block */ + if (len > 0) { + /* + * If len is malaligned, write an extra mtt entry to cover + * the misaligned area (round up the division) + */ + err = mlx4_ib_umem_write_mtt_block(dev, mtt, mtt_size, + mtt_shift, len, + cur_start_addr, pages, + &start_index, &npages); + if (err) + goto out; + } + + if (npages) + err = mlx4_write_mtt(dev->dev, mtt, start_index, npages, pages); out: free_page((unsigned long) pages); return err; } +/* + * Calculate optimal mtt size based on contiguous pages. + * Function will return also the number of pages that are not aligned to the + * calculated mtt_size to be added to total number of pages. For that we should + * check the first chunk length & last chunk length and if not aligned to + * mtt_size we should increment the non_aligned_pages number. All chunks in the + * middle already handled as part of mtt shift calculation for both their start + * & end addresses. + */ +int mlx4_ib_umem_calc_optimal_mtt_size(struct ib_umem *umem, u64 start_va, + int *num_of_mtts) +{ + u64 block_shift = MLX4_MAX_MTT_SHIFT; + u64 min_shift = umem->page_shift; + u64 last_block_aligned_end = 0; + u64 current_block_start = 0; + u64 first_block_start = 0; + u64 current_block_len = 0; + u64 last_block_end = 0; + struct scatterlist *sg; + u64 current_block_end; + u64 misalignment_bits; + u64 next_block_start; + u64 total_len = 0; + int i; + + for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) { + /* + * Initialization - save the first chunk start as the + * current_block_start - block means contiguous pages. + */ + if (current_block_len == 0 && current_block_start == 0) { + current_block_start = sg_dma_address(sg); + first_block_start = current_block_start; + /* + * Find the bits that are different between the physical + * address and the virtual address for the start of the + * MR. + * umem_get aligned the start_va to a page boundary. + * Therefore, we need to align the start va to the same + * boundary. + * misalignment_bits is needed to handle the case of a + * single memory region. In this case, the rest of the + * logic will not reduce the block size. If we use a + * block size which is bigger than the alignment of the + * misalignment bits, we might use the virtual page + * number instead of the physical page number, resulting + * in access to the wrong data. + */ + misalignment_bits = + (start_va & (~(((u64)(BIT(umem->page_shift))) - 1ULL))) + ^ current_block_start; + block_shift = min(alignment_of(misalignment_bits), + block_shift); + } + + /* + * Go over the scatter entries and check if they continue the + * previous scatter entry. + */ + next_block_start = sg_dma_address(sg); + current_block_end = current_block_start + current_block_len; + /* If we have a split (non-contig.) between two blocks */ + if (current_block_end != next_block_start) { + block_shift = mlx4_ib_umem_calc_block_mtt + (next_block_start, + current_block_end, + block_shift); + + /* + * If we reached the minimum shift for 4k page we stop + * the loop. + */ + if (block_shift <= min_shift) + goto end; + + /* + * If not saved yet we are in first block - we save the + * length of first block to calculate the + * non_aligned_pages number at the end. + */ + total_len += current_block_len; + + /* Start a new block */ + current_block_start = next_block_start; + current_block_len = sg_dma_len(sg); + continue; + } + /* The scatter entry is another part of the current block, + * increase the block size. + * An entry in the scatter can be larger than 4k (page) as of + * dma mapping which merge some blocks together. + */ + current_block_len += sg_dma_len(sg); + } + + /* Account for the last block in the total len */ + total_len += current_block_len; + /* Add to the first block the misalignment that it suffers from. */ + total_len += (first_block_start & ((1ULL << block_shift) - 1ULL)); + last_block_end = current_block_start + current_block_len; + last_block_aligned_end = round_up(last_block_end, 1 << block_shift); + total_len += (last_block_aligned_end - last_block_end); + + if (total_len & ((1ULL << block_shift) - 1ULL)) + pr_warn("misaligned total length detected (%llu, %llu)!", + total_len, block_shift); + + *num_of_mtts = total_len >> block_shift; +end: + if (block_shift < min_shift) { + /* + * If shift is less than the min we set a warning and return the + * min shift. + */ + pr_warn("umem_calc_optimal_mtt_size - unexpected shift %lld\n", block_shift); + + block_shift = min_shift; + } + return block_shift; +} + struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, u64 virt_addr, int access_flags, struct ib_udata *udata) @@ -155,7 +391,7 @@ struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, } n = ib_umem_page_count(mr->umem); - shift = mr->umem->page_shift; + shift = mlx4_ib_umem_calc_optimal_mtt_size(mr->umem, start, &n); err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length, convert_access(access_flags), n, shift, &mr->mmr); diff --git a/drivers/infiniband/hw/mlx4/qp.c b/drivers/infiniband/hw/mlx4/qp.c index b6b33d99b0b4..013049bcdb53 100644 --- a/drivers/infiniband/hw/mlx4/qp.c +++ b/drivers/infiniband/hw/mlx4/qp.c @@ -1038,6 +1038,8 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd, struct mlx4_ib_create_wq wq; } ucmd; size_t copy_len; + int shift; + int n; copy_len = (src == MLX4_IB_QP_SRC) ? sizeof(struct mlx4_ib_create_qp) : @@ -1100,8 +1102,10 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd, goto err; } - err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem), - qp->umem->page_shift, &qp->mtt); + n = ib_umem_page_count(qp->umem); + shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n); + err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt); + if (err) goto err_buf; @@ -2182,11 +2186,6 @@ static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type, context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) | (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16)); - if (rwq_ind_tbl) { - fill_qp_rss_context(context, qp); - context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET); - } - if (!(attr_mask & IB_QP_PATH_MIG_STATE)) context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11); else { @@ -2216,7 +2215,7 @@ static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type, context->mtu_msgmax = (IB_MTU_4096 << 5) | ilog2(dev->dev->caps.max_gso_sz); else - context->mtu_msgmax = (IB_MTU_4096 << 5) | 12; + context->mtu_msgmax = (IB_MTU_4096 << 5) | 13; } else if (attr_mask & IB_QP_PATH_MTU) { if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) { pr_err("path MTU (%u) is invalid\n", @@ -2387,6 +2386,7 @@ static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type, context->pd = cpu_to_be32(pd->pdn); if (!rwq_ind_tbl) { + context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28); get_cqs(qp, src_type, &send_cq, &recv_cq); } else { /* Set dummy CQs to be compatible with HV and PRM */ send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq); @@ -2394,7 +2394,6 @@ static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type, } context->cqn_send = cpu_to_be32(send_cq->mcq.cqn); context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn); - context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28); /* Set "fast registration enabled" for all kernel QPs */ if (!ibuobject) @@ -2513,7 +2512,7 @@ static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type, MLX4_IB_LINK_TYPE_ETH; if (dev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) { /* set QP to receive both tunneled & non-tunneled packets */ - if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET))) + if (!rwq_ind_tbl) context->srqn = cpu_to_be32(7 << 28); } } @@ -2562,6 +2561,13 @@ static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type, } } + if (rwq_ind_tbl && + cur_state == IB_QPS_RESET && + new_state == IB_QPS_INIT) { + fill_qp_rss_context(context, qp); + context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET); + } + err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state), to_mlx4_state(new_state), context, optpar, sqd_event, &qp->mqp); |