From 399d7921299fc4f146bd62bfa6312382a5429bcc Mon Sep 17 00:00:00 2001 From: Roland Dreier Date: Tue, 31 Jan 2006 21:14:18 -0800 Subject: IB/mthca: Get rid of might_sleep() annotations The might_sleep() annotations in mthca are silly -- they all occur shortly before calls that will end up in core functions like kmalloc() that will print the same warning in an unsafe context anyway. In fact, beyond cluttering the source, we're actually bloating text with CONFIG_DEBUG_SPINLOCK_SLEEP and/or CONFIG_PREEMPT_VOLUNTARY set. With both options set, getting rid of the might_sleep()s saves a lot: add/remove: 0/0 grow/shrink: 0/7 up/down: 0/-171 (-171) function old new delta mthca_pd_alloc 132 109 -23 mthca_init_cq 969 946 -23 mthca_mr_alloc 592 568 -24 mthca_pd_free 67 42 -25 mthca_free_mr 219 194 -25 mthca_free_cq 570 545 -25 mthca_fmr_alloc 742 716 -26 Signed-off-by: Roland Dreier --- drivers/infiniband/hw/mthca/mthca_mr.c | 6 ------ 1 file changed, 6 deletions(-) (limited to 'drivers/infiniband/hw/mthca/mthca_mr.c') diff --git a/drivers/infiniband/hw/mthca/mthca_mr.c b/drivers/infiniband/hw/mthca/mthca_mr.c index e995e2aa016..551320fdc9c 100644 --- a/drivers/infiniband/hw/mthca/mthca_mr.c +++ b/drivers/infiniband/hw/mthca/mthca_mr.c @@ -340,8 +340,6 @@ int mthca_mr_alloc(struct mthca_dev *dev, u32 pd, int buffer_size_shift, int err; u8 status; - might_sleep(); - WARN_ON(buffer_size_shift >= 32); key = mthca_alloc(&dev->mr_table.mpt_alloc); @@ -467,8 +465,6 @@ void mthca_free_mr(struct mthca_dev *dev, struct mthca_mr *mr) int err; u8 status; - might_sleep(); - err = mthca_HW2SW_MPT(dev, NULL, key_to_hw_index(dev, mr->ibmr.lkey) & (dev->limits.num_mpts - 1), @@ -495,8 +491,6 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd, int err = -ENOMEM; int i; - might_sleep(); - if (mr->attr.page_size < 12 || mr->attr.page_size >= 32) return -EINVAL; -- cgit v1.2.3-70-g09d2 From 2fa5e2ebbe2d81f741ba7bed9e07dc38cc734625 Mon Sep 17 00:00:00 2001 From: Roland Dreier Date: Wed, 1 Feb 2006 13:38:24 -0800 Subject: IB/mthca: Whitespace cleanups Remove trailing whitespace and fix indentation that with spaces instead of tabs. Signed-off-by: Roland Dreier --- drivers/infiniband/hw/mthca/mthca_av.c | 2 +- drivers/infiniband/hw/mthca/mthca_cmd.c | 2 +- drivers/infiniband/hw/mthca/mthca_cmd.h | 4 ++-- drivers/infiniband/hw/mthca/mthca_eq.c | 2 +- drivers/infiniband/hw/mthca/mthca_mcg.c | 2 +- drivers/infiniband/hw/mthca/mthca_mr.c | 6 +++--- drivers/infiniband/hw/mthca/mthca_provider.c | 4 ++-- drivers/infiniband/hw/mthca/mthca_qp.c | 28 ++++++++++++++-------------- drivers/infiniband/hw/mthca/mthca_srq.c | 2 +- 9 files changed, 26 insertions(+), 26 deletions(-) (limited to 'drivers/infiniband/hw/mthca/mthca_mr.c') diff --git a/drivers/infiniband/hw/mthca/mthca_av.c b/drivers/infiniband/hw/mthca/mthca_av.c index a19e0ed03d7..198f1267e98 100644 --- a/drivers/infiniband/hw/mthca/mthca_av.c +++ b/drivers/infiniband/hw/mthca/mthca_av.c @@ -147,7 +147,7 @@ int mthca_destroy_ah(struct mthca_dev *dev, struct mthca_ah *ah) switch (ah->type) { case MTHCA_AH_ON_HCA: mthca_free(&dev->av_table.alloc, - (ah->avdma - dev->av_table.ddr_av_base) / + (ah->avdma - dev->av_table.ddr_av_base) / MTHCA_AV_SIZE); break; diff --git a/drivers/infiniband/hw/mthca/mthca_cmd.c b/drivers/infiniband/hw/mthca/mthca_cmd.c index 7780e974cfb..7150fd143bf 100644 --- a/drivers/infiniband/hw/mthca/mthca_cmd.c +++ b/drivers/infiniband/hw/mthca/mthca_cmd.c @@ -275,7 +275,7 @@ static int mthca_cmd_poll(struct mthca_dev *dev, } if (out_is_imm) - *out_param = + *out_param = (u64) be32_to_cpu((__force __be32) __raw_readl(dev->hcr + HCR_OUT_PARAM_OFFSET)) << 32 | (u64) be32_to_cpu((__force __be32) diff --git a/drivers/infiniband/hw/mthca/mthca_cmd.h b/drivers/infiniband/hw/mthca/mthca_cmd.h index 3473b8235ee..a8da84b0a0f 100644 --- a/drivers/infiniband/hw/mthca/mthca_cmd.h +++ b/drivers/infiniband/hw/mthca/mthca_cmd.h @@ -74,9 +74,9 @@ enum { MTHCA_CMD_STAT_REG_BOUND = 0x21, /* HCA local attached memory not present: */ MTHCA_CMD_STAT_LAM_NOT_PRE = 0x22, - /* Bad management packet (silently discarded): */ + /* Bad management packet (silently discarded): */ MTHCA_CMD_STAT_BAD_PKT = 0x30, - /* More outstanding CQEs in CQ than new CQ size: */ + /* More outstanding CQEs in CQ than new CQ size: */ MTHCA_CMD_STAT_BAD_SIZE = 0x40 }; diff --git a/drivers/infiniband/hw/mthca/mthca_eq.c b/drivers/infiniband/hw/mthca/mthca_eq.c index 2eabb27804c..cf43a538839 100644 --- a/drivers/infiniband/hw/mthca/mthca_eq.c +++ b/drivers/infiniband/hw/mthca/mthca_eq.c @@ -497,7 +497,7 @@ static int __devinit mthca_create_eq(struct mthca_dev *dev, eq->dev = dev; eq->nent = roundup_pow_of_two(max(nent, 2)); - npages = ALIGN(eq->nent * MTHCA_EQ_ENTRY_SIZE, PAGE_SIZE) / PAGE_SIZE; + npages = ALIGN(eq->nent * MTHCA_EQ_ENTRY_SIZE, PAGE_SIZE) / PAGE_SIZE; eq->page_list = kmalloc(npages * sizeof *eq->page_list, GFP_KERNEL); diff --git a/drivers/infiniband/hw/mthca/mthca_mcg.c b/drivers/infiniband/hw/mthca/mthca_mcg.c index 321f11e707f..9965bda9afe 100644 --- a/drivers/infiniband/hw/mthca/mthca_mcg.c +++ b/drivers/infiniband/hw/mthca/mthca_mcg.c @@ -187,7 +187,7 @@ int mthca_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid) for (i = 0; i < MTHCA_QP_PER_MGM; ++i) if (mgm->qp[i] == cpu_to_be32(ibqp->qp_num | (1 << 31))) { - mthca_dbg(dev, "QP %06x already a member of MGM\n", + mthca_dbg(dev, "QP %06x already a member of MGM\n", ibqp->qp_num); err = 0; goto out; diff --git a/drivers/infiniband/hw/mthca/mthca_mr.c b/drivers/infiniband/hw/mthca/mthca_mr.c index 551320fdc9c..653f1321a13 100644 --- a/drivers/infiniband/hw/mthca/mthca_mr.c +++ b/drivers/infiniband/hw/mthca/mthca_mr.c @@ -517,7 +517,7 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd, BUG_ON(!mr->mem.arbel.mpt); } else mr->mem.tavor.mpt = dev->mr_table.tavor_fmr.mpt_base + - sizeof *(mr->mem.tavor.mpt) * idx; + sizeof *(mr->mem.tavor.mpt) * idx; mr->mtt = __mthca_alloc_mtt(dev, list_len, dev->mr_table.fmr_mtt_buddy); if (IS_ERR(mr->mtt)) @@ -779,7 +779,7 @@ int __devinit mthca_init_mr_table(struct mthca_dev *dev) } dev->mr_table.tavor_fmr.mpt_base = - ioremap(dev->mr_table.mpt_base, + ioremap(dev->mr_table.mpt_base, (1 << i) * sizeof (struct mthca_mpt_entry)); if (!dev->mr_table.tavor_fmr.mpt_base) { @@ -807,7 +807,7 @@ int __devinit mthca_init_mr_table(struct mthca_dev *dev) goto err_reserve_fmr; dev->mr_table.fmr_mtt_buddy = - &dev->mr_table.tavor_fmr.mtt_buddy; + &dev->mr_table.tavor_fmr.mtt_buddy; } else dev->mr_table.fmr_mtt_buddy = &dev->mr_table.mtt_buddy; diff --git a/drivers/infiniband/hw/mthca/mthca_provider.c b/drivers/infiniband/hw/mthca/mthca_provider.c index cd2038bdd1a..76321f0ef01 100644 --- a/drivers/infiniband/hw/mthca/mthca_provider.c +++ b/drivers/infiniband/hw/mthca/mthca_provider.c @@ -108,12 +108,12 @@ static int mthca_query_device(struct ib_device *ibdev, props->max_srq_wr = mdev->limits.max_srq_wqes; props->max_srq_sge = mdev->limits.max_sg; props->local_ca_ack_delay = mdev->limits.local_ca_ack_delay; - props->atomic_cap = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ? + props->atomic_cap = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ? IB_ATOMIC_HCA : IB_ATOMIC_NONE; props->max_pkeys = mdev->limits.pkey_table_len; props->max_mcast_grp = mdev->limits.num_mgms + mdev->limits.num_amgms; props->max_mcast_qp_attach = MTHCA_QP_PER_MGM; - props->max_total_mcast_qp_attach = props->max_mcast_qp_attach * + props->max_total_mcast_qp_attach = props->max_mcast_qp_attach * props->max_mcast_grp; err = 0; diff --git a/drivers/infiniband/hw/mthca/mthca_qp.c b/drivers/infiniband/hw/mthca/mthca_qp.c index 79245717e98..bdba39a8d9f 100644 --- a/drivers/infiniband/hw/mthca/mthca_qp.c +++ b/drivers/infiniband/hw/mthca/mthca_qp.c @@ -2,7 +2,7 @@ * Copyright (c) 2004 Topspin Communications. All rights reserved. * Copyright (c) 2005 Cisco Systems. All rights reserved. * Copyright (c) 2005 Mellanox Technologies. All rights reserved. - * Copyright (c) 2004 Voltaire, Inc. All rights reserved. + * Copyright (c) 2004 Voltaire, Inc. All rights reserved. * * This software is available to you under a choice of one of two * licenses. You may choose to be licensed under the terms of the GNU @@ -559,9 +559,9 @@ static void mthca_path_set(struct ib_ah_attr *ah, struct mthca_qp_path *path) path->g_mylmc |= 1 << 7; path->mgid_index = ah->grh.sgid_index; path->hop_limit = ah->grh.hop_limit; - path->sl_tclass_flowlabel = + path->sl_tclass_flowlabel = cpu_to_be32((ah->sl << 28) | - (ah->grh.traffic_class << 20) | + (ah->grh.traffic_class << 20) | (ah->grh.flow_label)); memcpy(path->rgid, ah->grh.dgid.raw, 16); } else @@ -597,7 +597,7 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask) } if (attr_mask & IB_QP_STATE) { - if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR) + if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR) return -EINVAL; new_state = attr->qp_state; } else @@ -630,10 +630,10 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask) return -EINVAL; } - if ((attr_mask & IB_QP_PKEY_INDEX) && + if ((attr_mask & IB_QP_PKEY_INDEX) && attr->pkey_index >= dev->limits.pkey_table_len) { mthca_dbg(dev, "PKey index (%u) too large. max is %d\n", - attr->pkey_index,dev->limits.pkey_table_len-1); + attr->pkey_index,dev->limits.pkey_table_len-1); return -EINVAL; } @@ -733,7 +733,7 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask) if (attr_mask & IB_QP_RNR_RETRY) { qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry = attr->rnr_retry << 5; - qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY | + qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY | MTHCA_QP_OPTPAR_ALT_RNR_RETRY); } @@ -749,13 +749,13 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask) if (attr_mask & IB_QP_ALT_PATH) { if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) { - mthca_dbg(dev, "Alternate port number (%u) is invalid\n", + mthca_dbg(dev, "Alternate port number (%u) is invalid\n", attr->alt_port_num); return -EINVAL; } mthca_path_set(&attr->alt_ah_attr, &qp_context->alt_path); - qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index | + qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index | attr->alt_port_num << 24); qp_context->alt_path.ackto = attr->alt_timeout << 3; qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH); @@ -1078,10 +1078,10 @@ static int mthca_map_memfree(struct mthca_dev *dev, if (ret) goto err_qpc; - ret = mthca_table_get(dev, dev->qp_table.rdb_table, - qp->qpn << dev->qp_table.rdb_shift); - if (ret) - goto err_eqpc; + ret = mthca_table_get(dev, dev->qp_table.rdb_table, + qp->qpn << dev->qp_table.rdb_shift); + if (ret) + goto err_eqpc; } @@ -2115,7 +2115,7 @@ int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr, int i; void *wqe; - spin_lock_irqsave(&qp->rq.lock, flags); + spin_lock_irqsave(&qp->rq.lock, flags); /* XXX check that state is OK to post receive */ diff --git a/drivers/infiniband/hw/mthca/mthca_srq.c b/drivers/infiniband/hw/mthca/mthca_srq.c index e7e153d9c4c..f1a1da147d0 100644 --- a/drivers/infiniband/hw/mthca/mthca_srq.c +++ b/drivers/infiniband/hw/mthca/mthca_srq.c @@ -339,7 +339,7 @@ void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq) int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, enum ib_srq_attr_mask attr_mask) -{ +{ struct mthca_dev *dev = to_mdev(ibsrq->device); struct mthca_srq *srq = to_msrq(ibsrq); int ret; -- cgit v1.2.3-70-g09d2 From d36f34aadf184d8cc4c240de2b6319ccea8334bb Mon Sep 17 00:00:00 2001 From: Or Gerlitz Date: Thu, 2 Feb 2006 10:43:45 -0800 Subject: IB: Enable FMR pool user to set page size This patch allows the consumer to set the page size of "pages" mapped by the pool FMRs, which is a feature already existing in the base verbs API. On the cosmetic side it changes ib_fmr_attr.page_size field to be named page_shift. Signed-off-by: Or Gerlitz Signed-off-by: Roland Dreier --- drivers/infiniband/core/fmr_pool.c | 6 +++--- drivers/infiniband/hw/mthca/mthca_mr.c | 10 +++++----- include/rdma/ib_fmr_pool.h | 2 ++ include/rdma/ib_verbs.h | 2 +- 4 files changed, 11 insertions(+), 9 deletions(-) (limited to 'drivers/infiniband/hw/mthca/mthca_mr.c') diff --git a/drivers/infiniband/core/fmr_pool.c b/drivers/infiniband/core/fmr_pool.c index d34a6f1c4f4..838bf54458d 100644 --- a/drivers/infiniband/core/fmr_pool.c +++ b/drivers/infiniband/core/fmr_pool.c @@ -278,9 +278,9 @@ struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd *pd, { struct ib_pool_fmr *fmr; struct ib_fmr_attr attr = { - .max_pages = params->max_pages_per_fmr, - .max_maps = IB_FMR_MAX_REMAPS, - .page_size = PAGE_SHIFT + .max_pages = params->max_pages_per_fmr, + .max_maps = IB_FMR_MAX_REMAPS, + .page_shift = params->page_shift }; for (i = 0; i < params->pool_size; ++i) { diff --git a/drivers/infiniband/hw/mthca/mthca_mr.c b/drivers/infiniband/hw/mthca/mthca_mr.c index 653f1321a13..0b48048ad0f 100644 --- a/drivers/infiniband/hw/mthca/mthca_mr.c +++ b/drivers/infiniband/hw/mthca/mthca_mr.c @@ -491,7 +491,7 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd, int err = -ENOMEM; int i; - if (mr->attr.page_size < 12 || mr->attr.page_size >= 32) + if (mr->attr.page_shift < 12 || mr->attr.page_shift >= 32) return -EINVAL; /* For Arbel, all MTTs must fit in the same page. */ @@ -543,7 +543,7 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd, MTHCA_MPT_FLAG_REGION | access); - mpt_entry->page_size = cpu_to_be32(mr->attr.page_size - 12); + mpt_entry->page_size = cpu_to_be32(mr->attr.page_shift - 12); mpt_entry->key = cpu_to_be32(key); mpt_entry->pd = cpu_to_be32(pd); memset(&mpt_entry->start, 0, @@ -611,7 +611,7 @@ static inline int mthca_check_fmr(struct mthca_fmr *fmr, u64 *page_list, if (list_len > fmr->attr.max_pages) return -EINVAL; - page_mask = (1 << fmr->attr.page_size) - 1; + page_mask = (1 << fmr->attr.page_shift) - 1; /* We are getting page lists, so va must be page aligned. */ if (iova & page_mask) @@ -659,7 +659,7 @@ int mthca_tavor_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list, } mpt_entry.lkey = cpu_to_be32(key); - mpt_entry.length = cpu_to_be64(list_len * (1ull << fmr->attr.page_size)); + mpt_entry.length = cpu_to_be64(list_len * (1ull << fmr->attr.page_shift)); mpt_entry.start = cpu_to_be64(iova); __raw_writel((__force u32) mpt_entry.lkey, &fmr->mem.tavor.mpt->key); @@ -700,7 +700,7 @@ int mthca_arbel_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list, fmr->mem.arbel.mpt->key = cpu_to_be32(key); fmr->mem.arbel.mpt->lkey = cpu_to_be32(key); - fmr->mem.arbel.mpt->length = cpu_to_be64(list_len * (1ull << fmr->attr.page_size)); + fmr->mem.arbel.mpt->length = cpu_to_be64(list_len * (1ull << fmr->attr.page_shift)); fmr->mem.arbel.mpt->start = cpu_to_be64(iova); wmb(); diff --git a/include/rdma/ib_fmr_pool.h b/include/rdma/ib_fmr_pool.h index 86b7e93f198..4ace54cd0cc 100644 --- a/include/rdma/ib_fmr_pool.h +++ b/include/rdma/ib_fmr_pool.h @@ -43,6 +43,7 @@ struct ib_fmr_pool; /** * struct ib_fmr_pool_param - Parameters for creating FMR pool * @max_pages_per_fmr:Maximum number of pages per map request. + * @page_shift: Log2 of sizeof "pages" mapped by this fmr * @access:Access flags for FMRs in pool. * @pool_size:Number of FMRs to allocate for pool. * @dirty_watermark:Flush is triggered when @dirty_watermark dirty @@ -55,6 +56,7 @@ struct ib_fmr_pool; */ struct ib_fmr_pool_param { int max_pages_per_fmr; + int page_shift; enum ib_access_flags access; int pool_size; int dirty_watermark; diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h index 1d31c8cd5ce..61a4390ae9d 100644 --- a/include/rdma/ib_verbs.h +++ b/include/rdma/ib_verbs.h @@ -651,7 +651,7 @@ struct ib_mw_bind { struct ib_fmr_attr { int max_pages; int max_maps; - u8 page_size; + u8 page_shift; }; struct ib_ucontext { -- cgit v1.2.3-70-g09d2 From 651eaac92894f8b6761c51b6637ea9cacea7fba2 Mon Sep 17 00:00:00 2001 From: Eli Cohen Date: Thu, 2 Mar 2006 12:40:46 -0800 Subject: IB/mthca: Optimize large messages on Sinai HCAs Sinai (one-port PCI Express) HCAs get improved throughput for messages bigger than 80 KB in DDR mode if memory keys are formatted in a specific way. The enhancement only works if the memory key table is smaller than 2^24 entries. For larger tables, the enhancement is off and a warning is printed (to avoid silent performance loss). Signed-off-by: Eli Cohen Signed-off-by: Michael Tsirkin Signed-off-by: Roland Dreier --- drivers/infiniband/hw/mthca/mthca_cmd.c | 12 ++++++++---- drivers/infiniband/hw/mthca/mthca_dev.h | 3 ++- drivers/infiniband/hw/mthca/mthca_main.c | 23 +++++++++++++---------- drivers/infiniband/hw/mthca/mthca_mr.c | 20 +++++++++++++++++++- drivers/infiniband/hw/mthca/mthca_profile.c | 10 +++++++++- 5 files changed, 51 insertions(+), 17 deletions(-) (limited to 'drivers/infiniband/hw/mthca/mthca_mr.c') diff --git a/drivers/infiniband/hw/mthca/mthca_cmd.c b/drivers/infiniband/hw/mthca/mthca_cmd.c index 948a2861cae..343eca50787 100644 --- a/drivers/infiniband/hw/mthca/mthca_cmd.c +++ b/drivers/infiniband/hw/mthca/mthca_cmd.c @@ -1277,7 +1277,8 @@ int mthca_INIT_HCA(struct mthca_dev *dev, int err; #define INIT_HCA_IN_SIZE 0x200 -#define INIT_HCA_FLAGS_OFFSET 0x014 +#define INIT_HCA_FLAGS1_OFFSET 0x00c +#define INIT_HCA_FLAGS2_OFFSET 0x014 #define INIT_HCA_QPC_OFFSET 0x020 #define INIT_HCA_QPC_BASE_OFFSET (INIT_HCA_QPC_OFFSET + 0x10) #define INIT_HCA_LOG_QP_OFFSET (INIT_HCA_QPC_OFFSET + 0x17) @@ -1320,15 +1321,18 @@ int mthca_INIT_HCA(struct mthca_dev *dev, memset(inbox, 0, INIT_HCA_IN_SIZE); + if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT) + MTHCA_PUT(inbox, 0x1, INIT_HCA_FLAGS1_OFFSET); + #if defined(__LITTLE_ENDIAN) - *(inbox + INIT_HCA_FLAGS_OFFSET / 4) &= ~cpu_to_be32(1 << 1); + *(inbox + INIT_HCA_FLAGS2_OFFSET / 4) &= ~cpu_to_be32(1 << 1); #elif defined(__BIG_ENDIAN) - *(inbox + INIT_HCA_FLAGS_OFFSET / 4) |= cpu_to_be32(1 << 1); + *(inbox + INIT_HCA_FLAGS2_OFFSET / 4) |= cpu_to_be32(1 << 1); #else #error Host endianness not defined #endif /* Check port for UD address vector: */ - *(inbox + INIT_HCA_FLAGS_OFFSET / 4) |= cpu_to_be32(1); + *(inbox + INIT_HCA_FLAGS2_OFFSET / 4) |= cpu_to_be32(1); /* We leave wqe_quota, responder_exu, etc as 0 (default) */ diff --git a/drivers/infiniband/hw/mthca/mthca_dev.h b/drivers/infiniband/hw/mthca/mthca_dev.h index ea48c897c7c..ad52edbefe9 100644 --- a/drivers/infiniband/hw/mthca/mthca_dev.h +++ b/drivers/infiniband/hw/mthca/mthca_dev.h @@ -64,7 +64,8 @@ enum { MTHCA_FLAG_NO_LAM = 1 << 5, MTHCA_FLAG_FMR = 1 << 6, MTHCA_FLAG_MEMFREE = 1 << 7, - MTHCA_FLAG_PCIE = 1 << 8 + MTHCA_FLAG_PCIE = 1 << 8, + MTHCA_FLAG_SINAI_OPT = 1 << 9 }; enum { diff --git a/drivers/infiniband/hw/mthca/mthca_main.c b/drivers/infiniband/hw/mthca/mthca_main.c index 9c849d27b06..d94837fbf16 100644 --- a/drivers/infiniband/hw/mthca/mthca_main.c +++ b/drivers/infiniband/hw/mthca/mthca_main.c @@ -935,13 +935,19 @@ enum { static struct { u64 latest_fw; - int is_memfree; - int is_pcie; + u32 flags; } mthca_hca_table[] = { - [TAVOR] = { .latest_fw = MTHCA_FW_VER(3, 3, 3), .is_memfree = 0, .is_pcie = 0 }, - [ARBEL_COMPAT] = { .latest_fw = MTHCA_FW_VER(4, 7, 0), .is_memfree = 0, .is_pcie = 1 }, - [ARBEL_NATIVE] = { .latest_fw = MTHCA_FW_VER(5, 1, 0), .is_memfree = 1, .is_pcie = 1 }, - [SINAI] = { .latest_fw = MTHCA_FW_VER(1, 0, 1), .is_memfree = 1, .is_pcie = 1 } + [TAVOR] = { .latest_fw = MTHCA_FW_VER(3, 3, 3), + .flags = 0 }, + [ARBEL_COMPAT] = { .latest_fw = MTHCA_FW_VER(4, 7, 0), + .flags = MTHCA_FLAG_PCIE }, + [ARBEL_NATIVE] = { .latest_fw = MTHCA_FW_VER(5, 1, 0), + .flags = MTHCA_FLAG_MEMFREE | + MTHCA_FLAG_PCIE }, + [SINAI] = { .latest_fw = MTHCA_FW_VER(1, 0, 1), + .flags = MTHCA_FLAG_MEMFREE | + MTHCA_FLAG_PCIE | + MTHCA_FLAG_SINAI_OPT } }; static int __devinit mthca_init_one(struct pci_dev *pdev, @@ -1031,12 +1037,9 @@ static int __devinit mthca_init_one(struct pci_dev *pdev, mdev->pdev = pdev; + mdev->mthca_flags = mthca_hca_table[id->driver_data].flags; if (ddr_hidden) mdev->mthca_flags |= MTHCA_FLAG_DDR_HIDDEN; - if (mthca_hca_table[id->driver_data].is_memfree) - mdev->mthca_flags |= MTHCA_FLAG_MEMFREE; - if (mthca_hca_table[id->driver_data].is_pcie) - mdev->mthca_flags |= MTHCA_FLAG_PCIE; /* * Now reset the HCA before we touch the PCI capabilities or diff --git a/drivers/infiniband/hw/mthca/mthca_mr.c b/drivers/infiniband/hw/mthca/mthca_mr.c index 0b48048ad0f..698b6212576 100644 --- a/drivers/infiniband/hw/mthca/mthca_mr.c +++ b/drivers/infiniband/hw/mthca/mthca_mr.c @@ -76,6 +76,8 @@ struct mthca_mpt_entry { #define MTHCA_MPT_STATUS_SW 0xF0 #define MTHCA_MPT_STATUS_HW 0x00 +#define SINAI_FMR_KEY_INC 0x1000000 + /* * Buddy allocator for MTT segments (currently not very efficient * since it doesn't keep a free list and just searches linearly @@ -330,6 +332,14 @@ static inline u32 key_to_hw_index(struct mthca_dev *dev, u32 key) return tavor_key_to_hw_index(key); } +static inline u32 adjust_key(struct mthca_dev *dev, u32 key) +{ + if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT) + return ((key << 20) & 0x800000) | (key & 0x7fffff); + else + return key; +} + int mthca_mr_alloc(struct mthca_dev *dev, u32 pd, int buffer_size_shift, u64 iova, u64 total_size, u32 access, struct mthca_mr *mr) { @@ -345,6 +355,7 @@ int mthca_mr_alloc(struct mthca_dev *dev, u32 pd, int buffer_size_shift, key = mthca_alloc(&dev->mr_table.mpt_alloc); if (key == -1) return -ENOMEM; + key = adjust_key(dev, key); mr->ibmr.rkey = mr->ibmr.lkey = hw_index_to_key(dev, key); if (mthca_is_memfree(dev)) { @@ -504,6 +515,7 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd, key = mthca_alloc(&dev->mr_table.mpt_alloc); if (key == -1) return -ENOMEM; + key = adjust_key(dev, key); idx = key & (dev->limits.num_mpts - 1); mr->ibmr.rkey = mr->ibmr.lkey = hw_index_to_key(dev, key); @@ -687,7 +699,10 @@ int mthca_arbel_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list, ++fmr->maps; key = arbel_key_to_hw_index(fmr->ibmr.lkey); - key += dev->limits.num_mpts; + if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT) + key += SINAI_FMR_KEY_INC; + else + key += dev->limits.num_mpts; fmr->ibmr.lkey = fmr->ibmr.rkey = arbel_hw_index_to_key(key); *(u8 *) fmr->mem.arbel.mpt = MTHCA_MPT_STATUS_SW; @@ -760,6 +775,9 @@ int __devinit mthca_init_mr_table(struct mthca_dev *dev) else dev->mthca_flags |= MTHCA_FLAG_FMR; + if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT) + mthca_dbg(dev, "Memory key throughput optimization activated.\n"); + err = mthca_buddy_init(&dev->mr_table.mtt_buddy, fls(dev->limits.num_mtt_segs - 1)); diff --git a/drivers/infiniband/hw/mthca/mthca_profile.c b/drivers/infiniband/hw/mthca/mthca_profile.c index 08a909371b0..58d44aa3c30 100644 --- a/drivers/infiniband/hw/mthca/mthca_profile.c +++ b/drivers/infiniband/hw/mthca/mthca_profile.c @@ -152,7 +152,7 @@ u64 mthca_make_profile(struct mthca_dev *dev, } if (total_size > mem_avail) { mthca_err(dev, "Profile requires 0x%llx bytes; " - "won't in 0x%llx bytes of context memory.\n", + "won't fit in 0x%llx bytes of context memory.\n", (unsigned long long) total_size, (unsigned long long) mem_avail); kfree(profile); @@ -262,6 +262,14 @@ u64 mthca_make_profile(struct mthca_dev *dev, */ dev->limits.num_pds = MTHCA_NUM_PDS; + if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT && + init_hca->log_mpt_sz > 23) { + mthca_warn(dev, "MPT table too large (requested size 2^%d >= 2^24)\n", + init_hca->log_mpt_sz); + mthca_warn(dev, "Disabling memory key throughput optimization.\n"); + dev->mthca_flags &= ~MTHCA_FLAG_SINAI_OPT; + } + /* * For Tavor, FMRs use ioremapped PCI memory. For 32 bit * systems it may use too much vmalloc space to map all MTT -- cgit v1.2.3-70-g09d2