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 		u32 num;
80 		int log_num;
81 	};
82 
83 	u64 total_size = 0;
84 	struct mlx4_resource *profile;
85 	struct sysinfo si;
86 	int i, j;
87 
88 	profile = kcalloc(MLX4_RES_NUM, sizeof(*profile), GFP_KERNEL);
89 	if (!profile)
90 		return -ENOMEM;
91 
92 	/*
93 	 * We want to scale the number of MTTs with the size of the
94 	 * system memory, since it makes sense to register a lot of
95 	 * memory on a system with a lot of memory.  As a heuristic,
96 	 * make sure we have enough MTTs to cover twice the system
97 	 * memory (with PAGE_SIZE entries).
98 	 *
99 	 * This number has to be a power of two and fit into 32 bits
100 	 * due to device limitations, so cap this at 2^31 as well.
101 	 * That limits us to 8TB of memory registration per HCA with
102 	 * 4KB pages, which is probably OK for the next few months.
103 	 */
104 	si_meminfo(&si);
105 	request->num_mtt =
106 		roundup_pow_of_two(max_t(unsigned, request->num_mtt,
107 					 min(1UL << (31 - log_mtts_per_seg),
108 					     si.totalram >> (log_mtts_per_seg - 1))));
109 
110 	profile[MLX4_RES_QP].size     = dev_cap->qpc_entry_sz;
111 	profile[MLX4_RES_RDMARC].size = dev_cap->rdmarc_entry_sz;
112 	profile[MLX4_RES_ALTC].size   = dev_cap->altc_entry_sz;
113 	profile[MLX4_RES_AUXC].size   = dev_cap->aux_entry_sz;
114 	profile[MLX4_RES_SRQ].size    = dev_cap->srq_entry_sz;
115 	profile[MLX4_RES_CQ].size     = dev_cap->cqc_entry_sz;
116 	profile[MLX4_RES_EQ].size     = dev_cap->eqc_entry_sz;
117 	profile[MLX4_RES_DMPT].size   = dev_cap->dmpt_entry_sz;
118 	profile[MLX4_RES_CMPT].size   = dev_cap->cmpt_entry_sz;
119 	profile[MLX4_RES_MTT].size    = dev_cap->mtt_entry_sz;
120 	profile[MLX4_RES_MCG].size    = mlx4_get_mgm_entry_size(dev);
121 
122 	profile[MLX4_RES_QP].num      = request->num_qp;
123 	profile[MLX4_RES_RDMARC].num  = request->num_qp * request->rdmarc_per_qp;
124 	profile[MLX4_RES_ALTC].num    = request->num_qp;
125 	profile[MLX4_RES_AUXC].num    = request->num_qp;
126 	profile[MLX4_RES_SRQ].num     = request->num_srq;
127 	profile[MLX4_RES_CQ].num      = request->num_cq;
128 	profile[MLX4_RES_EQ].num = mlx4_is_mfunc(dev) ? dev->phys_caps.num_phys_eqs :
129 					min_t(unsigned, dev_cap->max_eqs, MAX_MSIX);
130 	profile[MLX4_RES_DMPT].num    = request->num_mpt;
131 	profile[MLX4_RES_CMPT].num    = MLX4_NUM_CMPTS;
132 	profile[MLX4_RES_MTT].num     = request->num_mtt * (1 << log_mtts_per_seg);
133 	profile[MLX4_RES_MCG].num     = request->num_mcg;
134 
135 	for (i = 0; i < MLX4_RES_NUM; ++i) {
136 		profile[i].type     = i;
137 		profile[i].num      = roundup_pow_of_two(profile[i].num);
138 		profile[i].log_num  = ilog2(profile[i].num);
139 		profile[i].size    *= profile[i].num;
140 		profile[i].size     = max(profile[i].size, (u64) PAGE_SIZE);
141 	}
142 
143 	/*
144 	 * Sort the resources in decreasing order of size.  Since they
145 	 * all have sizes that are powers of 2, we'll be able to keep
146 	 * resources aligned to their size and pack them without gaps
147 	 * using the sorted order.
148 	 */
149 	for (i = MLX4_RES_NUM; i > 0; --i)
150 		for (j = 1; j < i; ++j) {
151 			if (profile[j].size > profile[j - 1].size)
152 				swap(profile[j], profile[j - 1]);
153 		}
154 
155 	for (i = 0; i < MLX4_RES_NUM; ++i) {
156 		if (profile[i].size) {
157 			profile[i].start = total_size;
158 			total_size	+= profile[i].size;
159 		}
160 
161 		if (total_size > dev_cap->max_icm_sz) {
162 			mlx4_err(dev, "Profile requires 0x%llx bytes; won't fit in 0x%llx bytes of context memory\n",
163 				 (unsigned long long) total_size,
164 				 (unsigned long long) dev_cap->max_icm_sz);
165 			kfree(profile);
166 			return -ENOMEM;
167 		}
168 
169 		if (profile[i].size)
170 			mlx4_dbg(dev, "  profile[%2d] (%6s): 2^%02d entries @ 0x%10llx, size 0x%10llx\n",
171 				 i, res_name[profile[i].type],
172 				 profile[i].log_num,
173 				 (unsigned long long) profile[i].start,
174 				 (unsigned long long) profile[i].size);
175 	}
176 
177 	mlx4_dbg(dev, "HCA context memory: reserving %d KB\n",
178 		 (int) (total_size >> 10));
179 
180 	for (i = 0; i < MLX4_RES_NUM; ++i) {
181 		switch (profile[i].type) {
182 		case MLX4_RES_QP:
183 			dev->caps.num_qps     = profile[i].num;
184 			init_hca->qpc_base    = profile[i].start;
185 			init_hca->log_num_qps = profile[i].log_num;
186 			break;
187 		case MLX4_RES_RDMARC:
188 			for (priv->qp_table.rdmarc_shift = 0;
189 			     request->num_qp << priv->qp_table.rdmarc_shift < profile[i].num;
190 			     ++priv->qp_table.rdmarc_shift)
191 				; /* nothing */
192 			dev->caps.max_qp_dest_rdma = 1 << priv->qp_table.rdmarc_shift;
193 			priv->qp_table.rdmarc_base   = (u32) profile[i].start;
194 			init_hca->rdmarc_base	     = profile[i].start;
195 			init_hca->log_rd_per_qp	     = priv->qp_table.rdmarc_shift;
196 			break;
197 		case MLX4_RES_ALTC:
198 			init_hca->altc_base = profile[i].start;
199 			break;
200 		case MLX4_RES_AUXC:
201 			init_hca->auxc_base = profile[i].start;
202 			break;
203 		case MLX4_RES_SRQ:
204 			dev->caps.num_srqs     = profile[i].num;
205 			init_hca->srqc_base    = profile[i].start;
206 			init_hca->log_num_srqs = profile[i].log_num;
207 			break;
208 		case MLX4_RES_CQ:
209 			dev->caps.num_cqs     = profile[i].num;
210 			init_hca->cqc_base    = profile[i].start;
211 			init_hca->log_num_cqs = profile[i].log_num;
212 			break;
213 		case MLX4_RES_EQ:
214 			if (dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_SYS_EQS) {
215 				init_hca->log_num_eqs = 0x1f;
216 				init_hca->eqc_base    = profile[i].start;
217 				init_hca->num_sys_eqs = dev_cap->num_sys_eqs;
218 			} else {
219 				dev->caps.num_eqs     = roundup_pow_of_two(
220 								min_t(unsigned,
221 								      dev_cap->max_eqs,
222 								      MAX_MSIX));
223 				init_hca->eqc_base    = profile[i].start;
224 				init_hca->log_num_eqs = ilog2(dev->caps.num_eqs);
225 			}
226 			break;
227 		case MLX4_RES_DMPT:
228 			dev->caps.num_mpts	= profile[i].num;
229 			priv->mr_table.mpt_base = profile[i].start;
230 			init_hca->dmpt_base	= profile[i].start;
231 			init_hca->log_mpt_sz	= profile[i].log_num;
232 			break;
233 		case MLX4_RES_CMPT:
234 			init_hca->cmpt_base	 = profile[i].start;
235 			break;
236 		case MLX4_RES_MTT:
237 			dev->caps.num_mtts	 = profile[i].num;
238 			priv->mr_table.mtt_base	 = profile[i].start;
239 			init_hca->mtt_base	 = profile[i].start;
240 			break;
241 		case MLX4_RES_MCG:
242 			init_hca->mc_base	  = profile[i].start;
243 			init_hca->log_mc_entry_sz =
244 					ilog2(mlx4_get_mgm_entry_size(dev));
245 			init_hca->log_mc_table_sz = profile[i].log_num;
246 			if (dev->caps.steering_mode ==
247 			    MLX4_STEERING_MODE_DEVICE_MANAGED) {
248 				dev->caps.num_mgms = profile[i].num;
249 			} else {
250 				init_hca->log_mc_hash_sz =
251 						profile[i].log_num - 1;
252 				dev->caps.num_mgms = profile[i].num >> 1;
253 				dev->caps.num_amgms = profile[i].num >> 1;
254 			}
255 			break;
256 		default:
257 			break;
258 		}
259 	}
260 
261 	/*
262 	 * PDs don't take any HCA memory, but we assign them as part
263 	 * of the HCA profile anyway.
264 	 */
265 	dev->caps.num_pds = MLX4_NUM_PDS;
266 
267 	kfree(profile);
268 	return total_size;
269 }
270