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 struct mlx4_wqe_data_seg *scatter; 78 int desc_size; 79 int buf_size; 80 int err; 81 int i; 82 83 /* Sanity check SRQ size before proceeding */ 84 if (init_attr->attr.max_wr >= dev->dev->caps.max_srq_wqes || 85 init_attr->attr.max_sge > dev->dev->caps.max_srq_sge) 86 return ERR_PTR(-EINVAL); 87 88 srq = kmalloc(sizeof *srq, GFP_KERNEL); 89 if (!srq) 90 return ERR_PTR(-ENOMEM); 91 92 mutex_init(&srq->mutex); 93 spin_lock_init(&srq->lock); 94 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1); 95 srq->msrq.max_gs = init_attr->attr.max_sge; 96 97 desc_size = max(32UL, 98 roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) + 99 srq->msrq.max_gs * 100 sizeof (struct mlx4_wqe_data_seg))); 101 srq->msrq.wqe_shift = ilog2(desc_size); 102 103 buf_size = srq->msrq.max * desc_size; 104 105 if (pd->uobject) { 106 struct mlx4_ib_create_srq ucmd; 107 108 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) { 109 err = -EFAULT; 110 goto err_srq; 111 } 112 113 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, 114 buf_size, 0, 0); 115 if (IS_ERR(srq->umem)) { 116 err = PTR_ERR(srq->umem); 117 goto err_srq; 118 } 119 120 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(srq->umem), 121 ilog2(srq->umem->page_size), &srq->mtt); 122 if (err) 123 goto err_buf; 124 125 err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem); 126 if (err) 127 goto err_mtt; 128 129 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context), 130 ucmd.db_addr, &srq->db); 131 if (err) 132 goto err_mtt; 133 } else { 134 err = mlx4_db_alloc(dev->dev, &srq->db, 0); 135 if (err) 136 goto err_srq; 137 138 *srq->db.db = 0; 139 140 if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, &srq->buf)) { 141 err = -ENOMEM; 142 goto err_db; 143 } 144 145 srq->head = 0; 146 srq->tail = srq->msrq.max - 1; 147 srq->wqe_ctr = 0; 148 149 for (i = 0; i < srq->msrq.max; ++i) { 150 next = get_wqe(srq, i); 151 next->next_wqe_index = 152 cpu_to_be16((i + 1) & (srq->msrq.max - 1)); 153 154 for (scatter = (void *) (next + 1); 155 (void *) scatter < (void *) next + desc_size; 156 ++scatter) 157 scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY); 158 } 159 160 err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift, 161 &srq->mtt); 162 if (err) 163 goto err_buf; 164 165 err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf); 166 if (err) 167 goto err_mtt; 168 169 srq->wrid = kmalloc(srq->msrq.max * sizeof (u64), GFP_KERNEL); 170 if (!srq->wrid) { 171 err = -ENOMEM; 172 goto err_mtt; 173 } 174 } 175 176 err = mlx4_srq_alloc(dev->dev, to_mpd(pd)->pdn, &srq->mtt, 177 srq->db.dma, &srq->msrq); 178 if (err) 179 goto err_wrid; 180 181 srq->msrq.event = mlx4_ib_srq_event; 182 183 if (pd->uobject) 184 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) { 185 err = -EFAULT; 186 goto err_wrid; 187 } 188 189 init_attr->attr.max_wr = srq->msrq.max - 1; 190 191 return &srq->ibsrq; 192 193 err_wrid: 194 if (pd->uobject) 195 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db); 196 else 197 kfree(srq->wrid); 198 199 err_mtt: 200 mlx4_mtt_cleanup(dev->dev, &srq->mtt); 201 202 err_buf: 203 if (pd->uobject) 204 ib_umem_release(srq->umem); 205 else 206 mlx4_buf_free(dev->dev, buf_size, &srq->buf); 207 208 err_db: 209 if (!pd->uobject) 210 mlx4_db_free(dev->dev, &srq->db); 211 212 err_srq: 213 kfree(srq); 214 215 return ERR_PTR(err); 216 } 217 218 int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 219 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata) 220 { 221 struct mlx4_ib_dev *dev = to_mdev(ibsrq->device); 222 struct mlx4_ib_srq *srq = to_msrq(ibsrq); 223 int ret; 224 225 /* We don't support resizing SRQs (yet?) */ 226 if (attr_mask & IB_SRQ_MAX_WR) 227 return -EINVAL; 228 229 if (attr_mask & IB_SRQ_LIMIT) { 230 if (attr->srq_limit >= srq->msrq.max) 231 return -EINVAL; 232 233 mutex_lock(&srq->mutex); 234 ret = mlx4_srq_arm(dev->dev, &srq->msrq, attr->srq_limit); 235 mutex_unlock(&srq->mutex); 236 237 if (ret) 238 return ret; 239 } 240 241 return 0; 242 } 243 244 int mlx4_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr) 245 { 246 struct mlx4_ib_dev *dev = to_mdev(ibsrq->device); 247 struct mlx4_ib_srq *srq = to_msrq(ibsrq); 248 int ret; 249 int limit_watermark; 250 251 ret = mlx4_srq_query(dev->dev, &srq->msrq, &limit_watermark); 252 if (ret) 253 return ret; 254 255 srq_attr->srq_limit = limit_watermark; 256 srq_attr->max_wr = srq->msrq.max - 1; 257 srq_attr->max_sge = srq->msrq.max_gs; 258 259 return 0; 260 } 261 262 int mlx4_ib_destroy_srq(struct ib_srq *srq) 263 { 264 struct mlx4_ib_dev *dev = to_mdev(srq->device); 265 struct mlx4_ib_srq *msrq = to_msrq(srq); 266 267 mlx4_srq_free(dev->dev, &msrq->msrq); 268 mlx4_mtt_cleanup(dev->dev, &msrq->mtt); 269 270 if (srq->uobject) { 271 mlx4_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db); 272 ib_umem_release(msrq->umem); 273 } else { 274 kfree(msrq->wrid); 275 mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift, 276 &msrq->buf); 277 mlx4_db_free(dev->dev, &msrq->db); 278 } 279 280 kfree(msrq); 281 282 return 0; 283 } 284 285 void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index) 286 { 287 struct mlx4_wqe_srq_next_seg *next; 288 289 /* always called with interrupts disabled. */ 290 spin_lock(&srq->lock); 291 292 next = get_wqe(srq, srq->tail); 293 next->next_wqe_index = cpu_to_be16(wqe_index); 294 srq->tail = wqe_index; 295 296 spin_unlock(&srq->lock); 297 } 298 299 int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 300 struct ib_recv_wr **bad_wr) 301 { 302 struct mlx4_ib_srq *srq = to_msrq(ibsrq); 303 struct mlx4_wqe_srq_next_seg *next; 304 struct mlx4_wqe_data_seg *scat; 305 unsigned long flags; 306 int err = 0; 307 int nreq; 308 int i; 309 310 spin_lock_irqsave(&srq->lock, flags); 311 312 for (nreq = 0; wr; ++nreq, wr = wr->next) { 313 if (unlikely(wr->num_sge > srq->msrq.max_gs)) { 314 err = -EINVAL; 315 *bad_wr = wr; 316 break; 317 } 318 319 if (unlikely(srq->head == srq->tail)) { 320 err = -ENOMEM; 321 *bad_wr = wr; 322 break; 323 } 324 325 srq->wrid[srq->head] = wr->wr_id; 326 327 next = get_wqe(srq, srq->head); 328 srq->head = be16_to_cpu(next->next_wqe_index); 329 scat = (struct mlx4_wqe_data_seg *) (next + 1); 330 331 for (i = 0; i < wr->num_sge; ++i) { 332 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length); 333 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey); 334 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr); 335 } 336 337 if (i < srq->msrq.max_gs) { 338 scat[i].byte_count = 0; 339 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY); 340 scat[i].addr = 0; 341 } 342 } 343 344 if (likely(nreq)) { 345 srq->wqe_ctr += nreq; 346 347 /* 348 * Make sure that descriptors are written before 349 * doorbell record. 350 */ 351 wmb(); 352 353 *srq->db.db = cpu_to_be32(srq->wqe_ctr); 354 } 355 356 spin_unlock_irqrestore(&srq->lock, flags); 357 358 return err; 359 } 360