1 /* 2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved. 3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the 9 * OpenIB.org BSD license below: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * - Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * - Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 34 #include <linux/mlx4/qp.h> 35 #include <linux/mlx4/srq.h> 36 37 #include "mlx4_ib.h" 38 #include "user.h" 39 40 static void *get_wqe(struct mlx4_ib_srq *srq, int n) 41 { 42 return mlx4_buf_offset(&srq->buf, n << srq->msrq.wqe_shift); 43 } 44 45 static void mlx4_ib_srq_event(struct mlx4_srq *srq, enum mlx4_event type) 46 { 47 struct ib_event event; 48 struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq; 49 50 if (ibsrq->event_handler) { 51 event.device = ibsrq->device; 52 event.element.srq = ibsrq; 53 switch (type) { 54 case MLX4_EVENT_TYPE_SRQ_LIMIT: 55 event.event = IB_EVENT_SRQ_LIMIT_REACHED; 56 break; 57 case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR: 58 event.event = IB_EVENT_SRQ_ERR; 59 break; 60 default: 61 printk(KERN_WARNING "mlx4_ib: Unexpected event type %d " 62 "on SRQ %06x\n", type, srq->srqn); 63 return; 64 } 65 66 ibsrq->event_handler(&event, ibsrq->srq_context); 67 } 68 } 69 70 struct ib_srq *mlx4_ib_create_srq(struct ib_pd *pd, 71 struct ib_srq_init_attr *init_attr, 72 struct ib_udata *udata) 73 { 74 struct mlx4_ib_dev *dev = to_mdev(pd->device); 75 struct mlx4_ib_srq *srq; 76 struct mlx4_wqe_srq_next_seg *next; 77 int desc_size; 78 int buf_size; 79 int err; 80 int i; 81 82 /* Sanity check SRQ size before proceeding */ 83 if (init_attr->attr.max_wr >= dev->dev->caps.max_srq_wqes || 84 init_attr->attr.max_sge > dev->dev->caps.max_srq_sge) 85 return ERR_PTR(-EINVAL); 86 87 srq = kmalloc(sizeof *srq, GFP_KERNEL); 88 if (!srq) 89 return ERR_PTR(-ENOMEM); 90 91 mutex_init(&srq->mutex); 92 spin_lock_init(&srq->lock); 93 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1); 94 srq->msrq.max_gs = init_attr->attr.max_sge; 95 96 desc_size = max(32UL, 97 roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) + 98 srq->msrq.max_gs * 99 sizeof (struct mlx4_wqe_data_seg))); 100 srq->msrq.wqe_shift = ilog2(desc_size); 101 102 buf_size = srq->msrq.max * desc_size; 103 104 if (pd->uobject) { 105 struct mlx4_ib_create_srq ucmd; 106 107 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) { 108 err = -EFAULT; 109 goto err_srq; 110 } 111 112 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, 113 buf_size, 0, 0); 114 if (IS_ERR(srq->umem)) { 115 err = PTR_ERR(srq->umem); 116 goto err_srq; 117 } 118 119 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(srq->umem), 120 ilog2(srq->umem->page_size), &srq->mtt); 121 if (err) 122 goto err_buf; 123 124 err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem); 125 if (err) 126 goto err_mtt; 127 128 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context), 129 ucmd.db_addr, &srq->db); 130 if (err) 131 goto err_mtt; 132 } else { 133 err = mlx4_db_alloc(dev->dev, &srq->db, 0); 134 if (err) 135 goto err_srq; 136 137 *srq->db.db = 0; 138 139 if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, &srq->buf)) { 140 err = -ENOMEM; 141 goto err_db; 142 } 143 144 srq->head = 0; 145 srq->tail = srq->msrq.max - 1; 146 srq->wqe_ctr = 0; 147 148 for (i = 0; i < srq->msrq.max; ++i) { 149 next = get_wqe(srq, i); 150 next->next_wqe_index = 151 cpu_to_be16((i + 1) & (srq->msrq.max - 1)); 152 } 153 154 err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift, 155 &srq->mtt); 156 if (err) 157 goto err_buf; 158 159 err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf); 160 if (err) 161 goto err_mtt; 162 163 srq->wrid = kmalloc(srq->msrq.max * sizeof (u64), GFP_KERNEL); 164 if (!srq->wrid) { 165 err = -ENOMEM; 166 goto err_mtt; 167 } 168 } 169 170 err = mlx4_srq_alloc(dev->dev, to_mpd(pd)->pdn, &srq->mtt, 171 srq->db.dma, &srq->msrq); 172 if (err) 173 goto err_wrid; 174 175 srq->msrq.event = mlx4_ib_srq_event; 176 177 if (pd->uobject) 178 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) { 179 err = -EFAULT; 180 goto err_wrid; 181 } 182 183 init_attr->attr.max_wr = srq->msrq.max - 1; 184 185 return &srq->ibsrq; 186 187 err_wrid: 188 if (pd->uobject) 189 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db); 190 else 191 kfree(srq->wrid); 192 193 err_mtt: 194 mlx4_mtt_cleanup(dev->dev, &srq->mtt); 195 196 err_buf: 197 if (pd->uobject) 198 ib_umem_release(srq->umem); 199 else 200 mlx4_buf_free(dev->dev, buf_size, &srq->buf); 201 202 err_db: 203 if (!pd->uobject) 204 mlx4_db_free(dev->dev, &srq->db); 205 206 err_srq: 207 kfree(srq); 208 209 return ERR_PTR(err); 210 } 211 212 int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 213 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata) 214 { 215 struct mlx4_ib_dev *dev = to_mdev(ibsrq->device); 216 struct mlx4_ib_srq *srq = to_msrq(ibsrq); 217 int ret; 218 219 /* We don't support resizing SRQs (yet?) */ 220 if (attr_mask & IB_SRQ_MAX_WR) 221 return -EINVAL; 222 223 if (attr_mask & IB_SRQ_LIMIT) { 224 if (attr->srq_limit >= srq->msrq.max) 225 return -EINVAL; 226 227 mutex_lock(&srq->mutex); 228 ret = mlx4_srq_arm(dev->dev, &srq->msrq, attr->srq_limit); 229 mutex_unlock(&srq->mutex); 230 231 if (ret) 232 return ret; 233 } 234 235 return 0; 236 } 237 238 int mlx4_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr) 239 { 240 struct mlx4_ib_dev *dev = to_mdev(ibsrq->device); 241 struct mlx4_ib_srq *srq = to_msrq(ibsrq); 242 int ret; 243 int limit_watermark; 244 245 ret = mlx4_srq_query(dev->dev, &srq->msrq, &limit_watermark); 246 if (ret) 247 return ret; 248 249 srq_attr->srq_limit = limit_watermark; 250 srq_attr->max_wr = srq->msrq.max - 1; 251 srq_attr->max_sge = srq->msrq.max_gs; 252 253 return 0; 254 } 255 256 int mlx4_ib_destroy_srq(struct ib_srq *srq) 257 { 258 struct mlx4_ib_dev *dev = to_mdev(srq->device); 259 struct mlx4_ib_srq *msrq = to_msrq(srq); 260 261 mlx4_srq_free(dev->dev, &msrq->msrq); 262 mlx4_mtt_cleanup(dev->dev, &msrq->mtt); 263 264 if (srq->uobject) { 265 mlx4_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db); 266 ib_umem_release(msrq->umem); 267 } else { 268 kfree(msrq->wrid); 269 mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift, 270 &msrq->buf); 271 mlx4_db_free(dev->dev, &msrq->db); 272 } 273 274 kfree(msrq); 275 276 return 0; 277 } 278 279 void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index) 280 { 281 struct mlx4_wqe_srq_next_seg *next; 282 283 /* always called with interrupts disabled. */ 284 spin_lock(&srq->lock); 285 286 next = get_wqe(srq, srq->tail); 287 next->next_wqe_index = cpu_to_be16(wqe_index); 288 srq->tail = wqe_index; 289 290 spin_unlock(&srq->lock); 291 } 292 293 int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 294 struct ib_recv_wr **bad_wr) 295 { 296 struct mlx4_ib_srq *srq = to_msrq(ibsrq); 297 struct mlx4_wqe_srq_next_seg *next; 298 struct mlx4_wqe_data_seg *scat; 299 unsigned long flags; 300 int err = 0; 301 int nreq; 302 int i; 303 304 spin_lock_irqsave(&srq->lock, flags); 305 306 for (nreq = 0; wr; ++nreq, wr = wr->next) { 307 if (unlikely(wr->num_sge > srq->msrq.max_gs)) { 308 err = -EINVAL; 309 *bad_wr = wr; 310 break; 311 } 312 313 if (unlikely(srq->head == srq->tail)) { 314 err = -ENOMEM; 315 *bad_wr = wr; 316 break; 317 } 318 319 srq->wrid[srq->head] = wr->wr_id; 320 321 next = get_wqe(srq, srq->head); 322 srq->head = be16_to_cpu(next->next_wqe_index); 323 scat = (struct mlx4_wqe_data_seg *) (next + 1); 324 325 for (i = 0; i < wr->num_sge; ++i) { 326 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length); 327 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey); 328 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr); 329 } 330 331 if (i < srq->msrq.max_gs) { 332 scat[i].byte_count = 0; 333 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY); 334 scat[i].addr = 0; 335 } 336 } 337 338 if (likely(nreq)) { 339 srq->wqe_ctr += nreq; 340 341 /* 342 * Make sure that descriptors are written before 343 * doorbell record. 344 */ 345 wmb(); 346 347 *srq->db.db = cpu_to_be32(srq->wqe_ctr); 348 } 349 350 spin_unlock_irqrestore(&srq->lock, flags); 351 352 return err; 353 } 354