From patchwork Sun Jan 1 21:10:23 2017 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Saeed Mahameed X-Patchwork-Id: 710080 X-Patchwork-Delegate: davem@davemloft.net Return-Path: X-Original-To: patchwork-incoming@ozlabs.org Delivered-To: patchwork-incoming@ozlabs.org Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by ozlabs.org (Postfix) with ESMTP id 3tsCYh3hc1z9sCM for ; Mon, 2 Jan 2017 08:11:20 +1100 (AEDT) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932506AbdAAVKz (ORCPT ); Sun, 1 Jan 2017 16:10:55 -0500 Received: from mail-il-dmz.mellanox.com ([193.47.165.129]:42838 "EHLO mellanox.co.il" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S932476AbdAAVKu (ORCPT ); Sun, 1 Jan 2017 16:10:50 -0500 Received: from Internal Mail-Server by MTLPINE1 (envelope-from saeedm@mellanox.com) with ESMTPS (AES256-SHA encrypted); 1 Jan 2017 23:10:47 +0200 Received: from reg-l-vrt-045-015.mtl.labs.mlnx (reg-l-vrt-045-015.mtl.labs.mlnx [10.135.45.15]) by labmailer.mlnx (8.13.8/8.13.8) with ESMTP id v01LAkt0008081; Sun, 1 Jan 2017 23:10:47 +0200 From: Saeed Mahameed To: "David S. Miller" , Doug Ledford Cc: netdev@vger.kernel.org, linux-rdma@vger.kernel.org, Leon Romanovsky , Ilya Lesokhin , Artemy Kovalyov , Leon Romanovsky , Saeed Mahameed Subject: [PATCH for-next 07/11] IB/mlx5: Add MR cache for large UMR regions Date: Sun, 1 Jan 2017 23:10:23 +0200 Message-Id: <1483305027-2573-8-git-send-email-saeedm@mellanox.com> X-Mailer: git-send-email 2.7.4 In-Reply-To: <1483305027-2573-1-git-send-email-saeedm@mellanox.com> References: <1483305027-2573-1-git-send-email-saeedm@mellanox.com> Sender: netdev-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: netdev@vger.kernel.org From: Artemy Kovalyov In this change we turn mlx5_ib_update_mtt() into generic mlx5_ib_update_xlt() to perfrom HCA translation table modifiactions supporting both atomic and process contexts and not limited by number of modified entries. Using this function we increase preallocated MRs up to 16GB. Signed-off-by: Artemy Kovalyov Signed-off-by: Leon Romanovsky Signed-off-by: Saeed Mahameed --- drivers/infiniband/hw/mlx5/main.c | 14 +- drivers/infiniband/hw/mlx5/mem.c | 32 +- drivers/infiniband/hw/mlx5/mlx5_ib.h | 15 +- drivers/infiniband/hw/mlx5/mr.c | 386 ++++++++++--------------- drivers/infiniband/hw/mlx5/odp.c | 19 +- drivers/net/ethernet/mellanox/mlx5/core/main.c | 20 ++ include/linux/mlx5/driver.h | 2 +- 7 files changed, 240 insertions(+), 248 deletions(-) diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c index 2ab4e32..b871272 100644 --- a/drivers/infiniband/hw/mlx5/main.c +++ b/drivers/infiniband/hw/mlx5/main.c @@ -1112,11 +1112,18 @@ static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev, context->ibucontext.invalidate_range = &mlx5_ib_invalidate_range; #endif + context->upd_xlt_page = __get_free_page(GFP_KERNEL); + if (!context->upd_xlt_page) { + err = -ENOMEM; + goto out_uars; + } + mutex_init(&context->upd_xlt_page_mutex); + if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) { err = mlx5_core_alloc_transport_domain(dev->mdev, &context->tdn); if (err) - goto out_uars; + goto out_page; } INIT_LIST_HEAD(&context->vma_private_list); @@ -1168,6 +1175,9 @@ static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev, if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn); +out_page: + free_page(context->upd_xlt_page); + out_uars: for (i--; i >= 0; i--) mlx5_cmd_free_uar(dev->mdev, uars[i].index); @@ -1195,6 +1205,8 @@ static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext) if (MLX5_CAP_GEN(dev->mdev, log_max_transport_domain)) mlx5_core_dealloc_transport_domain(dev->mdev, context->tdn); + free_page(context->upd_xlt_page); + for (i = 0; i < uuari->num_uars; i++) { if (mlx5_cmd_free_uar(dev->mdev, uuari->uars[i].index)) mlx5_ib_warn(dev, "failed to free UAR 0x%x\n", uuari->uars[i].index); diff --git a/drivers/infiniband/hw/mlx5/mem.c b/drivers/infiniband/hw/mlx5/mem.c index 6851357..778d8a1 100644 --- a/drivers/infiniband/hw/mlx5/mem.c +++ b/drivers/infiniband/hw/mlx5/mem.c @@ -159,7 +159,7 @@ void __mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem, unsigned long umem_page_shift = ilog2(umem->page_size); int shift = page_shift - umem_page_shift; int mask = (1 << shift) - 1; - int i, k; + int i, k, idx; u64 cur = 0; u64 base; int len; @@ -185,18 +185,36 @@ void __mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem, for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) { len = sg_dma_len(sg) >> umem_page_shift; base = sg_dma_address(sg); - for (k = 0; k < len; k++) { + + /* Skip elements below offset */ + if (i + len < offset << shift) { + i += len; + continue; + } + + /* Skip pages below offset */ + if (i < offset << shift) { + k = (offset << shift) - i; + i = offset << shift; + } else { + k = 0; + } + + for (; k < len; k++) { if (!(i & mask)) { cur = base + (k << umem_page_shift); cur |= access_flags; + idx = (i >> shift) - offset; - pas[i >> shift] = cpu_to_be64(cur); + pas[idx] = cpu_to_be64(cur); mlx5_ib_dbg(dev, "pas[%d] 0x%llx\n", - i >> shift, be64_to_cpu(pas[i >> shift])); - } else - mlx5_ib_dbg(dev, "=====> 0x%llx\n", - base + (k << umem_page_shift)); + i >> shift, be64_to_cpu(pas[idx])); + } i++; + + /* Stop after num_pages reached */ + if (i >> shift >= offset + num_pages) + return; } } } diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h index 73bff77..02d9255 100644 --- a/drivers/infiniband/hw/mlx5/mlx5_ib.h +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h @@ -125,6 +125,10 @@ struct mlx5_ib_ucontext { /* Transport Domain number */ u32 tdn; struct list_head vma_private_list; + + unsigned long upd_xlt_page; + /* protect ODP/KSM */ + struct mutex upd_xlt_page_mutex; }; static inline struct mlx5_ib_ucontext *to_mucontext(struct ib_ucontext *ibucontext) @@ -192,6 +196,13 @@ struct mlx5_ib_flow_db { #define MLX5_IB_UMR_OCTOWORD 16 #define MLX5_IB_UMR_XLT_ALIGNMENT 64 +#define MLX5_IB_UPD_XLT_ZAP BIT(0) +#define MLX5_IB_UPD_XLT_ENABLE BIT(1) +#define MLX5_IB_UPD_XLT_ATOMIC BIT(2) +#define MLX5_IB_UPD_XLT_ADDR BIT(3) +#define MLX5_IB_UPD_XLT_PD BIT(4) +#define MLX5_IB_UPD_XLT_ACCESS BIT(5) + /* Private QP creation flags to be passed in ib_qp_init_attr.create_flags. * * These flags are intended for internal use by the mlx5_ib driver, and they @@ -788,8 +799,8 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, struct ib_mw *mlx5_ib_alloc_mw(struct ib_pd *pd, enum ib_mw_type type, struct ib_udata *udata); int mlx5_ib_dealloc_mw(struct ib_mw *mw); -int mlx5_ib_update_mtt(struct mlx5_ib_mr *mr, u64 start_page_index, - int npages, int zap); +int mlx5_ib_update_xlt(struct mlx5_ib_mr *mr, u64 idx, int npages, + int page_shift, int flags); int mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start, u64 length, u64 virt_addr, int access_flags, struct ib_pd *pd, struct ib_udata *udata); diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c index 271b78e..e4333c8 100644 --- a/drivers/infiniband/hw/mlx5/mr.c +++ b/drivers/infiniband/hw/mlx5/mr.c @@ -46,14 +46,9 @@ enum { }; #define MLX5_UMR_ALIGN 2048 -#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING -static __be64 mlx5_ib_update_mtt_emergency_buffer[ - MLX5_UMR_MTT_MIN_CHUNK_SIZE/sizeof(__be64)] - __aligned(MLX5_UMR_ALIGN); -static DEFINE_MUTEX(mlx5_ib_update_mtt_emergency_buffer_mutex); -#endif static int clean_mr(struct mlx5_ib_mr *mr); +static int use_umr(struct mlx5_ib_dev *dev, int order); static int destroy_mkey(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr) { @@ -629,7 +624,8 @@ int mlx5_mr_cache_init(struct mlx5_ib_dev *dev) ent->dev = dev; if ((dev->mdev->profile->mask & MLX5_PROF_MASK_MR_CACHE) && - (mlx5_core_is_pf(dev->mdev))) + mlx5_core_is_pf(dev->mdev) && + use_umr(dev, ent->order)) limit = dev->mdev->profile->mr_cache[i].limit; else limit = 0; @@ -757,98 +753,13 @@ static int get_octo_len(u64 addr, u64 len, int page_size) return (npages + 1) / 2; } -static int use_umr(int order) +static int use_umr(struct mlx5_ib_dev *dev, int order) { + if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset)) + return order < MAX_MR_CACHE_ENTRIES + 2; return order <= MLX5_MAX_UMR_SHIFT; } -static int dma_map_mr_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem, - int npages, int page_shift, int *size, - __be64 **mr_pas, dma_addr_t *dma) -{ - __be64 *pas; - struct device *ddev = dev->ib_dev.dma_device; - - /* - * UMR copies MTTs in units of MLX5_UMR_MTT_ALIGNMENT bytes. - * To avoid copying garbage after the pas array, we allocate - * a little more. - */ - *size = ALIGN(sizeof(struct mlx5_mtt) * npages, MLX5_UMR_MTT_ALIGNMENT); - *mr_pas = kmalloc(*size + MLX5_UMR_ALIGN - 1, GFP_KERNEL); - if (!(*mr_pas)) - return -ENOMEM; - - pas = PTR_ALIGN(*mr_pas, MLX5_UMR_ALIGN); - mlx5_ib_populate_pas(dev, umem, page_shift, pas, MLX5_IB_MTT_PRESENT); - /* Clear padding after the actual pages. */ - memset(pas + npages, 0, *size - npages * sizeof(struct mlx5_mtt)); - - *dma = dma_map_single(ddev, pas, *size, DMA_TO_DEVICE); - if (dma_mapping_error(ddev, *dma)) { - kfree(*mr_pas); - return -ENOMEM; - } - - return 0; -} - -static void prep_umr_wqe_common(struct ib_pd *pd, struct ib_send_wr *wr, - struct ib_sge *sg, u64 dma, int n, u32 key, - int page_shift) -{ - struct mlx5_ib_dev *dev = to_mdev(pd->device); - struct mlx5_umr_wr *umrwr = umr_wr(wr); - - sg->addr = dma; - sg->length = ALIGN(sizeof(struct mlx5_mtt) * n, - MLX5_IB_UMR_XLT_ALIGNMENT); - sg->lkey = dev->umrc.pd->local_dma_lkey; - - wr->next = NULL; - wr->sg_list = sg; - if (n) - wr->num_sge = 1; - else - wr->num_sge = 0; - - wr->opcode = MLX5_IB_WR_UMR; - - umrwr->xlt_size = sg->length; - umrwr->page_shift = page_shift; - umrwr->mkey = key; -} - -static void prep_umr_reg_wqe(struct ib_pd *pd, struct ib_send_wr *wr, - struct ib_sge *sg, u64 dma, int n, u32 key, - int page_shift, u64 virt_addr, u64 len, - int access_flags) -{ - struct mlx5_umr_wr *umrwr = umr_wr(wr); - - prep_umr_wqe_common(pd, wr, sg, dma, n, key, page_shift); - - wr->send_flags = MLX5_IB_SEND_UMR_ENABLE_MR | - MLX5_IB_SEND_UMR_UPDATE_TRANSLATION | - MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS; - - umrwr->virt_addr = virt_addr; - umrwr->length = len; - umrwr->access_flags = access_flags; - umrwr->pd = pd; -} - -static void prep_umr_unreg_wqe(struct mlx5_ib_dev *dev, - struct ib_send_wr *wr, u32 key) -{ - struct mlx5_umr_wr *umrwr = umr_wr(wr); - - wr->send_flags = MLX5_IB_SEND_UMR_DISABLE_MR | - MLX5_IB_SEND_UMR_FAIL_IF_FREE; - wr->opcode = MLX5_IB_WR_UMR; - umrwr->mkey = key; -} - static int mr_umem_get(struct ib_pd *pd, u64 start, u64 length, int access_flags, struct ib_umem **umem, int *npages, int *page_shift, int *ncont, @@ -927,13 +838,7 @@ static struct mlx5_ib_mr *reg_umr(struct ib_pd *pd, struct ib_umem *umem, int page_shift, int order, int access_flags) { struct mlx5_ib_dev *dev = to_mdev(pd->device); - struct device *ddev = dev->ib_dev.dma_device; - struct mlx5_umr_wr umrwr = {}; struct mlx5_ib_mr *mr; - struct ib_sge sg; - int size; - __be64 *mr_pas; - dma_addr_t dma; int err = 0; int i; @@ -952,144 +857,174 @@ static struct mlx5_ib_mr *reg_umr(struct ib_pd *pd, struct ib_umem *umem, if (!mr) return ERR_PTR(-EAGAIN); - err = dma_map_mr_pas(dev, umem, npages, page_shift, &size, &mr_pas, - &dma); - if (err) - goto free_mr; - - prep_umr_reg_wqe(pd, &umrwr.wr, &sg, dma, npages, mr->mmkey.key, - page_shift, virt_addr, len, access_flags); - - err = mlx5_ib_post_send_wait(dev, &umrwr); - if (err && err != -EFAULT) - goto unmap_dma; - + mr->ibmr.pd = pd; + mr->umem = umem; + mr->access_flags = access_flags; + mr->desc_size = sizeof(struct mlx5_mtt); mr->mmkey.iova = virt_addr; mr->mmkey.size = len; mr->mmkey.pd = to_mpd(pd)->pdn; - mr->live = 1; - -unmap_dma: - dma_unmap_single(ddev, dma, size, DMA_TO_DEVICE); + err = mlx5_ib_update_xlt(mr, 0, npages, page_shift, + MLX5_IB_UPD_XLT_ENABLE); - kfree(mr_pas); - -free_mr: if (err) { free_cached_mr(dev, mr); return ERR_PTR(err); } + mr->live = 1; + return mr; } -#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING -int mlx5_ib_update_mtt(struct mlx5_ib_mr *mr, u64 start_page_index, int npages, - int zap) +static inline int populate_xlt(struct mlx5_ib_mr *mr, int idx, int npages, + void *xlt, int page_shift, size_t size, + int flags) { struct mlx5_ib_dev *dev = mr->dev; - struct device *ddev = dev->ib_dev.dma_device; struct ib_umem *umem = mr->umem; + + npages = min_t(size_t, npages, ib_umem_num_pages(umem) - idx); + + if (!(flags & MLX5_IB_UPD_XLT_ZAP)) { + __mlx5_ib_populate_pas(dev, umem, page_shift, + idx, npages, xlt, + MLX5_IB_MTT_PRESENT); + /* Clear padding after the pages + * brought from the umem. + */ + memset(xlt + (npages * sizeof(struct mlx5_mtt)), 0, + size - npages * sizeof(struct mlx5_mtt)); + } + + return npages; +} + +#define MLX5_MAX_UMR_CHUNK ((1 << (MLX5_MAX_UMR_SHIFT + 4)) - \ + MLX5_UMR_MTT_ALIGNMENT) +#define MLX5_SPARE_UMR_CHUNK 0x10000 + +int mlx5_ib_update_xlt(struct mlx5_ib_mr *mr, u64 idx, int npages, + int page_shift, int flags) +{ + struct mlx5_ib_dev *dev = mr->dev; + struct device *ddev = dev->ib_dev.dma_device; + struct mlx5_ib_ucontext *uctx = NULL; int size; - __be64 *pas; + void *xlt; dma_addr_t dma; struct mlx5_umr_wr wr; struct ib_sge sg; int err = 0; - const int page_index_alignment = MLX5_UMR_MTT_ALIGNMENT / - sizeof(struct mlx5_mtt); - const int page_index_mask = page_index_alignment - 1; + int desc_size = sizeof(struct mlx5_mtt); + const int page_align = MLX5_UMR_MTT_ALIGNMENT / desc_size; + const int page_mask = page_align - 1; size_t pages_mapped = 0; size_t pages_to_map = 0; size_t pages_iter = 0; - int use_emergency_buf = 0; + gfp_t gfp; /* UMR copies MTTs in units of MLX5_UMR_MTT_ALIGNMENT bytes, - * so we need to align the offset and length accordingly */ - if (start_page_index & page_index_mask) { - npages += start_page_index & page_index_mask; - start_page_index &= ~page_index_mask; + * so we need to align the offset and length accordingly + */ + if (idx & page_mask) { + npages += idx & page_mask; + idx &= ~page_mask; } - pages_to_map = ALIGN(npages, page_index_alignment); + gfp = flags & MLX5_IB_UPD_XLT_ATOMIC ? GFP_ATOMIC : GFP_KERNEL; + gfp |= __GFP_ZERO | __GFP_NOWARN; - if (start_page_index + pages_to_map > MLX5_MAX_UMR_PAGES) - return -EINVAL; + pages_to_map = ALIGN(npages, page_align); + size = desc_size * pages_to_map; + size = min_t(int, size, MLX5_MAX_UMR_CHUNK); - size = sizeof(struct mlx5_mtt) * pages_to_map; - size = min_t(int, PAGE_SIZE, size); - /* We allocate with GFP_ATOMIC to avoid recursion into page-reclaim - * code, when we are called from an invalidation. The pas buffer must - * be 2k-aligned for Connect-IB. */ - pas = (__be64 *)get_zeroed_page(GFP_ATOMIC); - if (!pas) { - mlx5_ib_warn(dev, "unable to allocate memory during MTT update, falling back to slower chunked mechanism.\n"); - pas = mlx5_ib_update_mtt_emergency_buffer; - size = MLX5_UMR_MTT_MIN_CHUNK_SIZE; - use_emergency_buf = 1; - mutex_lock(&mlx5_ib_update_mtt_emergency_buffer_mutex); - memset(pas, 0, size); + xlt = (void *)__get_free_pages(gfp, get_order(size)); + if (!xlt && size > MLX5_SPARE_UMR_CHUNK) { + mlx5_ib_dbg(dev, "Failed to allocate %d bytes of order %d. fallback to spare UMR allocation od %d bytes\n", + size, get_order(size), MLX5_SPARE_UMR_CHUNK); + + size = MLX5_SPARE_UMR_CHUNK; + xlt = (void *)__get_free_pages(gfp, get_order(size)); } - pages_iter = size / sizeof(struct mlx5_mtt); - dma = dma_map_single(ddev, pas, size, DMA_TO_DEVICE); + + if (!xlt) { + uctx = to_mucontext(mr->ibmr.uobject->context); + mlx5_ib_warn(dev, "Using XLT emergency buffer\n"); + size = PAGE_SIZE; + xlt = (void *)uctx->upd_xlt_page; + mutex_lock(&uctx->upd_xlt_page_mutex); + memset(xlt, 0, size); + } + pages_iter = size / desc_size; + dma = dma_map_single(ddev, xlt, size, DMA_TO_DEVICE); if (dma_mapping_error(ddev, dma)) { - mlx5_ib_err(dev, "unable to map DMA during MTT update.\n"); + mlx5_ib_err(dev, "unable to map DMA during XLT update.\n"); err = -ENOMEM; - goto free_pas; + goto free_xlt; } + sg.addr = dma; + sg.lkey = dev->umrc.pd->local_dma_lkey; + + memset(&wr, 0, sizeof(wr)); + wr.wr.send_flags = MLX5_IB_SEND_UMR_UPDATE_XLT; + if (!(flags & MLX5_IB_UPD_XLT_ENABLE)) + wr.wr.send_flags |= MLX5_IB_SEND_UMR_FAIL_IF_FREE; + wr.wr.sg_list = &sg; + wr.wr.num_sge = 1; + wr.wr.opcode = MLX5_IB_WR_UMR; + + wr.pd = mr->ibmr.pd; + wr.mkey = mr->mmkey.key; + wr.length = mr->mmkey.size; + wr.virt_addr = mr->mmkey.iova; + wr.access_flags = mr->access_flags; + wr.page_shift = page_shift; + for (pages_mapped = 0; pages_mapped < pages_to_map && !err; - pages_mapped += pages_iter, start_page_index += pages_iter) { + pages_mapped += pages_iter, idx += pages_iter) { dma_sync_single_for_cpu(ddev, dma, size, DMA_TO_DEVICE); - - npages = min_t(size_t, - pages_iter, - ib_umem_num_pages(umem) - start_page_index); - - if (!zap) { - __mlx5_ib_populate_pas(dev, umem, PAGE_SHIFT, - start_page_index, npages, pas, - MLX5_IB_MTT_PRESENT); - /* Clear padding after the pages brought from the - * umem. */ - memset(pas + npages, 0, size - npages * - sizeof(struct mlx5_mtt)); - } + npages = populate_xlt(mr, idx, pages_iter, xlt, + page_shift, size, flags); dma_sync_single_for_device(ddev, dma, size, DMA_TO_DEVICE); - memset(&wr, 0, sizeof(wr)); - - sg.addr = dma; - sg.length = ALIGN(npages * sizeof(struct mlx5_mtt), - MLX5_UMR_MTT_ALIGNMENT); - sg.lkey = dev->umrc.pd->local_dma_lkey; + sg.length = ALIGN(npages * desc_size, + MLX5_UMR_MTT_ALIGNMENT); + + if (pages_mapped + pages_iter >= pages_to_map) { + if (flags & MLX5_IB_UPD_XLT_ENABLE) + wr.wr.send_flags |= + MLX5_IB_SEND_UMR_ENABLE_MR | + MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS | + MLX5_IB_SEND_UMR_UPDATE_TRANSLATION; + if (flags & MLX5_IB_UPD_XLT_PD || + flags & MLX5_IB_UPD_XLT_ACCESS) + wr.wr.send_flags |= + MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS; + if (flags & MLX5_IB_UPD_XLT_ADDR) + wr.wr.send_flags |= + MLX5_IB_SEND_UMR_UPDATE_TRANSLATION; + } - wr.wr.send_flags = MLX5_IB_SEND_UMR_FAIL_IF_FREE | - MLX5_IB_SEND_UMR_UPDATE_XLT; - wr.wr.sg_list = &sg; - wr.wr.num_sge = 1; - wr.wr.opcode = MLX5_IB_WR_UMR; + wr.offset = idx * desc_size; wr.xlt_size = sg.length; - wr.page_shift = PAGE_SHIFT; - wr.mkey = mr->mmkey.key; - wr.offset = start_page_index * sizeof(struct mlx5_mtt); err = mlx5_ib_post_send_wait(dev, &wr); } dma_unmap_single(ddev, dma, size, DMA_TO_DEVICE); -free_pas: - if (!use_emergency_buf) - free_page((unsigned long)pas); +free_xlt: + if (uctx) + mutex_unlock(&uctx->upd_xlt_page_mutex); else - mutex_unlock(&mlx5_ib_update_mtt_emergency_buffer_mutex); + free_pages((unsigned long)xlt, get_order(size)); return err; } -#endif /* * If ibmr is NULL it will be allocated by reg_create. @@ -1204,7 +1139,7 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, if (err < 0) return ERR_PTR(err); - if (use_umr(order)) { + if (use_umr(dev, order)) { mr = reg_umr(pd, umem, virt_addr, length, ncont, page_shift, order, access_flags); if (PTR_ERR(mr) == -EAGAIN) { @@ -1254,39 +1189,25 @@ static int unreg_umr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr) if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) return 0; - prep_umr_unreg_wqe(dev, &umrwr.wr, mr->mmkey.key); + umrwr.wr.send_flags = MLX5_IB_SEND_UMR_DISABLE_MR | + MLX5_IB_SEND_UMR_FAIL_IF_FREE; + umrwr.wr.opcode = MLX5_IB_WR_UMR; + umrwr.mkey = mr->mmkey.key; return mlx5_ib_post_send_wait(dev, &umrwr); } -static int rereg_umr(struct ib_pd *pd, struct mlx5_ib_mr *mr, u64 virt_addr, - u64 length, int npages, int page_shift, int order, +static int rereg_umr(struct ib_pd *pd, struct mlx5_ib_mr *mr, int access_flags, int flags) { struct mlx5_ib_dev *dev = to_mdev(pd->device); - struct device *ddev = dev->ib_dev.dma_device; struct mlx5_umr_wr umrwr = {}; - struct ib_sge sg; - dma_addr_t dma = 0; - __be64 *mr_pas = NULL; - int size; int err; umrwr.wr.send_flags = MLX5_IB_SEND_UMR_FAIL_IF_FREE; - if (flags & IB_MR_REREG_TRANS) { - err = dma_map_mr_pas(dev, mr->umem, npages, page_shift, &size, - &mr_pas, &dma); - if (err) - return err; - - umrwr.virt_addr = virt_addr; - umrwr.length = length; - umrwr.wr.send_flags |= MLX5_IB_SEND_UMR_UPDATE_TRANSLATION; - } - - prep_umr_wqe_common(pd, &umrwr.wr, &sg, dma, npages, mr->mmkey.key, - page_shift); + umrwr.wr.opcode = MLX5_IB_WR_UMR; + umrwr.mkey = mr->mmkey.key; if (flags & IB_MR_REREG_PD || flags & IB_MR_REREG_ACCESS) { umrwr.pd = pd; @@ -1294,13 +1215,8 @@ static int rereg_umr(struct ib_pd *pd, struct mlx5_ib_mr *mr, u64 virt_addr, umrwr.wr.send_flags |= MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS; } - /* post send request to UMR QP */ err = mlx5_ib_post_send_wait(dev, &umrwr); - if (flags & IB_MR_REREG_TRANS) { - dma_unmap_single(ddev, dma, size, DMA_TO_DEVICE); - kfree(mr_pas); - } return err; } @@ -1317,6 +1233,7 @@ int mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start, u64 addr = (flags & IB_MR_REREG_TRANS) ? virt_addr : mr->umem->address; u64 len = (flags & IB_MR_REREG_TRANS) ? length : mr->umem->length; int page_shift = 0; + int upd_flags = 0; int npages = 0; int ncont = 0; int order = 0; @@ -1325,6 +1242,8 @@ int mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start, mlx5_ib_dbg(dev, "start 0x%llx, virt_addr 0x%llx, length 0x%llx, access_flags 0x%x\n", start, virt_addr, length, access_flags); + atomic_sub(mr->npages, &dev->mdev->priv.reg_pages); + if (flags != IB_MR_REREG_PD) { /* * Replace umem. This needs to be done whether or not UMR is @@ -1335,7 +1254,7 @@ int mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start, err = mr_umem_get(pd, addr, len, access_flags, &mr->umem, &npages, &page_shift, &ncont, &order); if (err < 0) { - mr->umem = NULL; + clean_mr(mr); return err; } } @@ -1367,32 +1286,37 @@ int mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start, /* * Send a UMR WQE */ - err = rereg_umr(pd, mr, addr, len, npages, page_shift, - order, access_flags, flags); + mr->ibmr.pd = pd; + mr->access_flags = access_flags; + mr->mmkey.iova = addr; + mr->mmkey.size = len; + mr->mmkey.pd = to_mpd(pd)->pdn; + + if (flags & IB_MR_REREG_TRANS) { + upd_flags = MLX5_IB_UPD_XLT_ADDR; + if (flags & IB_MR_REREG_PD) + upd_flags |= MLX5_IB_UPD_XLT_PD; + if (flags & IB_MR_REREG_ACCESS) + upd_flags |= MLX5_IB_UPD_XLT_ACCESS; + err = mlx5_ib_update_xlt(mr, 0, npages, page_shift, + upd_flags); + } else { + err = rereg_umr(pd, mr, access_flags, flags); + } + if (err) { mlx5_ib_warn(dev, "Failed to rereg UMR\n"); + ib_umem_release(mr->umem); + clean_mr(mr); return err; } } - if (flags & IB_MR_REREG_PD) { - ib_mr->pd = pd; - mr->mmkey.pd = to_mpd(pd)->pdn; - } + set_mr_fileds(dev, mr, npages, len, access_flags); - if (flags & IB_MR_REREG_ACCESS) - mr->access_flags = access_flags; - - if (flags & IB_MR_REREG_TRANS) { - atomic_sub(mr->npages, &dev->mdev->priv.reg_pages); - set_mr_fileds(dev, mr, npages, len, access_flags); - mr->mmkey.iova = addr; - mr->mmkey.size = len; - } #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING update_odp_mr(mr); #endif - return 0; } diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c index 1e73c12..cfd7ee5 100644 --- a/drivers/infiniband/hw/mlx5/odp.c +++ b/drivers/infiniband/hw/mlx5/odp.c @@ -91,16 +91,21 @@ void mlx5_ib_invalidate_range(struct ib_umem *umem, unsigned long start, u64 umr_offset = idx & umr_block_mask; if (in_block && umr_offset == 0) { - mlx5_ib_update_mtt(mr, blk_start_idx, - idx - blk_start_idx, 1); + mlx5_ib_update_xlt(mr, blk_start_idx, + idx - blk_start_idx, + PAGE_SHIFT, + MLX5_IB_UPD_XLT_ZAP | + MLX5_IB_UPD_XLT_ATOMIC); in_block = 0; } } } if (in_block) - mlx5_ib_update_mtt(mr, blk_start_idx, idx - blk_start_idx + 1, - 1); - + mlx5_ib_update_xlt(mr, blk_start_idx, + idx - blk_start_idx + 1, + PAGE_SHIFT, + MLX5_IB_UPD_XLT_ZAP | + MLX5_IB_UPD_XLT_ATOMIC); /* * We are now sure that the device will not access the * memory. We can safely unmap it, and mark it as dirty if @@ -257,7 +262,9 @@ static int pagefault_single_data_segment(struct mlx5_ib_qp *qp, * this MR, since ib_umem_odp_map_dma_pages already * checks this. */ - ret = mlx5_ib_update_mtt(mr, start_idx, npages, 0); + ret = mlx5_ib_update_xlt(mr, start_idx, npages, + PAGE_SHIFT, + MLX5_IB_UPD_XLT_ATOMIC); } else { ret = -EAGAIN; } diff --git a/drivers/net/ethernet/mellanox/mlx5/core/main.c b/drivers/net/ethernet/mellanox/mlx5/core/main.c index 54e5a78..1713bd8 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/main.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/main.c @@ -152,6 +152,26 @@ static struct mlx5_profile profile[] = { .size = 8, .limit = 4 }, + .mr_cache[16] = { + .size = 8, + .limit = 4 + }, + .mr_cache[17] = { + .size = 8, + .limit = 4 + }, + .mr_cache[18] = { + .size = 8, + .limit = 4 + }, + .mr_cache[19] = { + .size = 4, + .limit = 2 + }, + .mr_cache[20] = { + .size = 4, + .limit = 2 + }, }, }; diff --git a/include/linux/mlx5/driver.h b/include/linux/mlx5/driver.h index 0ae5536..ec52f3b 100644 --- a/include/linux/mlx5/driver.h +++ b/include/linux/mlx5/driver.h @@ -959,7 +959,7 @@ enum { }; enum { - MAX_MR_CACHE_ENTRIES = 16, + MAX_MR_CACHE_ENTRIES = 21, }; enum {