1 /* 2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved. 3 * Copyright (c) 2005 Mellanox Technologies. All rights reserved. 4 * Copyright (c) 2006, 2007 Cisco Systems, Inc. All rights reserved. 5 * 6 * This software is available to you under a choice of one of two 7 * licenses. You may choose to be licensed under the terms of the GNU 8 * General Public License (GPL) Version 2, available from the file 9 * COPYING in the main directory of this source tree, or the 10 * OpenIB.org BSD license below: 11 * 12 * Redistribution and use in source and binary forms, with or 13 * without modification, are permitted provided that the following 14 * conditions are met: 15 * 16 * - Redistributions of source code must retain the above 17 * copyright notice, this list of conditions and the following 18 * disclaimer. 19 * 20 * - Redistributions in binary form must reproduce the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer in the documentation and/or other materials 23 * provided with the distribution. 24 * 25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 32 * SOFTWARE. 33 */ 34 35 #include <linux/slab.h> 36 37 #include "mlx4.h" 38 #include "fw.h" 39 40 enum { 41 MLX4_RES_QP, 42 MLX4_RES_RDMARC, 43 MLX4_RES_ALTC, 44 MLX4_RES_AUXC, 45 MLX4_RES_SRQ, 46 MLX4_RES_CQ, 47 MLX4_RES_EQ, 48 MLX4_RES_DMPT, 49 MLX4_RES_CMPT, 50 MLX4_RES_MTT, 51 MLX4_RES_MCG, 52 MLX4_RES_NUM 53 }; 54 55 static const char *res_name[] = { 56 [MLX4_RES_QP] = "QP", 57 [MLX4_RES_RDMARC] = "RDMARC", 58 [MLX4_RES_ALTC] = "ALTC", 59 [MLX4_RES_AUXC] = "AUXC", 60 [MLX4_RES_SRQ] = "SRQ", 61 [MLX4_RES_CQ] = "CQ", 62 [MLX4_RES_EQ] = "EQ", 63 [MLX4_RES_DMPT] = "DMPT", 64 [MLX4_RES_CMPT] = "CMPT", 65 [MLX4_RES_MTT] = "MTT", 66 [MLX4_RES_MCG] = "MCG", 67 }; 68 69 u64 mlx4_make_profile(struct mlx4_dev *dev, 70 struct mlx4_profile *request, 71 struct mlx4_dev_cap *dev_cap, 72 struct mlx4_init_hca_param *init_hca) 73 { 74 struct mlx4_priv *priv = mlx4_priv(dev); 75 struct mlx4_resource { 76 u64 size; 77 u64 start; 78 int type; 79 int num; 80 int log_num; 81 }; 82 83 u64 total_size = 0; 84 struct mlx4_resource *profile; 85 struct mlx4_resource tmp; 86 struct sysinfo si; 87 int i, j; 88 89 profile = kcalloc(MLX4_RES_NUM, sizeof(*profile), GFP_KERNEL); 90 if (!profile) 91 return -ENOMEM; 92 93 /* 94 * We want to scale the number of MTTs with the size of the 95 * system memory, since it makes sense to register a lot of 96 * memory on a system with a lot of memory. As a heuristic, 97 * make sure we have enough MTTs to cover twice the system 98 * memory (with PAGE_SIZE entries). 99 * 100 * This number has to be a power of two and fit into 32 bits 101 * due to device limitations, so cap this at 2^31 as well. 102 * That limits us to 8TB of memory registration per HCA with 103 * 4KB pages, which is probably OK for the next few months. 104 */ 105 si_meminfo(&si); 106 request->num_mtt = 107 roundup_pow_of_two(max_t(unsigned, request->num_mtt, 108 min(1UL << 31, 109 si.totalram >> (log_mtts_per_seg - 1)))); 110 111 profile[MLX4_RES_QP].size = dev_cap->qpc_entry_sz; 112 profile[MLX4_RES_RDMARC].size = dev_cap->rdmarc_entry_sz; 113 profile[MLX4_RES_ALTC].size = dev_cap->altc_entry_sz; 114 profile[MLX4_RES_AUXC].size = dev_cap->aux_entry_sz; 115 profile[MLX4_RES_SRQ].size = dev_cap->srq_entry_sz; 116 profile[MLX4_RES_CQ].size = dev_cap->cqc_entry_sz; 117 profile[MLX4_RES_EQ].size = dev_cap->eqc_entry_sz; 118 profile[MLX4_RES_DMPT].size = dev_cap->dmpt_entry_sz; 119 profile[MLX4_RES_CMPT].size = dev_cap->cmpt_entry_sz; 120 profile[MLX4_RES_MTT].size = dev_cap->mtt_entry_sz; 121 profile[MLX4_RES_MCG].size = mlx4_get_mgm_entry_size(dev); 122 123 profile[MLX4_RES_QP].num = request->num_qp; 124 profile[MLX4_RES_RDMARC].num = request->num_qp * request->rdmarc_per_qp; 125 profile[MLX4_RES_ALTC].num = request->num_qp; 126 profile[MLX4_RES_AUXC].num = request->num_qp; 127 profile[MLX4_RES_SRQ].num = request->num_srq; 128 profile[MLX4_RES_CQ].num = request->num_cq; 129 profile[MLX4_RES_EQ].num = mlx4_is_mfunc(dev) ? 130 dev->phys_caps.num_phys_eqs : 131 min_t(unsigned, dev_cap->max_eqs, MAX_MSIX); 132 profile[MLX4_RES_DMPT].num = request->num_mpt; 133 profile[MLX4_RES_CMPT].num = MLX4_NUM_CMPTS; 134 profile[MLX4_RES_MTT].num = request->num_mtt * (1 << log_mtts_per_seg); 135 profile[MLX4_RES_MCG].num = request->num_mcg; 136 137 for (i = 0; i < MLX4_RES_NUM; ++i) { 138 profile[i].type = i; 139 profile[i].num = roundup_pow_of_two(profile[i].num); 140 profile[i].log_num = ilog2(profile[i].num); 141 profile[i].size *= profile[i].num; 142 profile[i].size = max(profile[i].size, (u64) PAGE_SIZE); 143 } 144 145 /* 146 * Sort the resources in decreasing order of size. Since they 147 * all have sizes that are powers of 2, we'll be able to keep 148 * resources aligned to their size and pack them without gaps 149 * using the sorted order. 150 */ 151 for (i = MLX4_RES_NUM; i > 0; --i) 152 for (j = 1; j < i; ++j) { 153 if (profile[j].size > profile[j - 1].size) { 154 tmp = profile[j]; 155 profile[j] = profile[j - 1]; 156 profile[j - 1] = tmp; 157 } 158 } 159 160 for (i = 0; i < MLX4_RES_NUM; ++i) { 161 if (profile[i].size) { 162 profile[i].start = total_size; 163 total_size += profile[i].size; 164 } 165 166 if (total_size > dev_cap->max_icm_sz) { 167 mlx4_err(dev, "Profile requires 0x%llx bytes; " 168 "won't fit in 0x%llx bytes of context memory.\n", 169 (unsigned long long) total_size, 170 (unsigned long long) dev_cap->max_icm_sz); 171 kfree(profile); 172 return -ENOMEM; 173 } 174 175 if (profile[i].size) 176 mlx4_dbg(dev, " profile[%2d] (%6s): 2^%02d entries @ 0x%10llx, " 177 "size 0x%10llx\n", 178 i, res_name[profile[i].type], profile[i].log_num, 179 (unsigned long long) profile[i].start, 180 (unsigned long long) profile[i].size); 181 } 182 183 mlx4_dbg(dev, "HCA context memory: reserving %d KB\n", 184 (int) (total_size >> 10)); 185 186 for (i = 0; i < MLX4_RES_NUM; ++i) { 187 switch (profile[i].type) { 188 case MLX4_RES_QP: 189 dev->caps.num_qps = profile[i].num; 190 init_hca->qpc_base = profile[i].start; 191 init_hca->log_num_qps = profile[i].log_num; 192 break; 193 case MLX4_RES_RDMARC: 194 for (priv->qp_table.rdmarc_shift = 0; 195 request->num_qp << priv->qp_table.rdmarc_shift < profile[i].num; 196 ++priv->qp_table.rdmarc_shift) 197 ; /* nothing */ 198 dev->caps.max_qp_dest_rdma = 1 << priv->qp_table.rdmarc_shift; 199 priv->qp_table.rdmarc_base = (u32) profile[i].start; 200 init_hca->rdmarc_base = profile[i].start; 201 init_hca->log_rd_per_qp = priv->qp_table.rdmarc_shift; 202 break; 203 case MLX4_RES_ALTC: 204 init_hca->altc_base = profile[i].start; 205 break; 206 case MLX4_RES_AUXC: 207 init_hca->auxc_base = profile[i].start; 208 break; 209 case MLX4_RES_SRQ: 210 dev->caps.num_srqs = profile[i].num; 211 init_hca->srqc_base = profile[i].start; 212 init_hca->log_num_srqs = profile[i].log_num; 213 break; 214 case MLX4_RES_CQ: 215 dev->caps.num_cqs = profile[i].num; 216 init_hca->cqc_base = profile[i].start; 217 init_hca->log_num_cqs = profile[i].log_num; 218 break; 219 case MLX4_RES_EQ: 220 dev->caps.num_eqs = roundup_pow_of_two(min_t(unsigned, dev_cap->max_eqs, 221 MAX_MSIX)); 222 init_hca->eqc_base = profile[i].start; 223 init_hca->log_num_eqs = ilog2(dev->caps.num_eqs); 224 break; 225 case MLX4_RES_DMPT: 226 dev->caps.num_mpts = profile[i].num; 227 priv->mr_table.mpt_base = profile[i].start; 228 init_hca->dmpt_base = profile[i].start; 229 init_hca->log_mpt_sz = profile[i].log_num; 230 break; 231 case MLX4_RES_CMPT: 232 init_hca->cmpt_base = profile[i].start; 233 break; 234 case MLX4_RES_MTT: 235 dev->caps.num_mtts = profile[i].num; 236 priv->mr_table.mtt_base = profile[i].start; 237 init_hca->mtt_base = profile[i].start; 238 break; 239 case MLX4_RES_MCG: 240 init_hca->mc_base = profile[i].start; 241 init_hca->log_mc_entry_sz = 242 ilog2(mlx4_get_mgm_entry_size(dev)); 243 init_hca->log_mc_table_sz = profile[i].log_num; 244 if (dev->caps.steering_mode == 245 MLX4_STEERING_MODE_DEVICE_MANAGED) { 246 dev->caps.num_mgms = profile[i].num; 247 } else { 248 init_hca->log_mc_hash_sz = 249 profile[i].log_num - 1; 250 dev->caps.num_mgms = profile[i].num >> 1; 251 dev->caps.num_amgms = profile[i].num >> 1; 252 } 253 break; 254 default: 255 break; 256 } 257 } 258 259 /* 260 * PDs don't take any HCA memory, but we assign them as part 261 * of the HCA profile anyway. 262 */ 263 dev->caps.num_pds = MLX4_NUM_PDS; 264 265 kfree(profile); 266 return total_size; 267 } 268