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#include <linux/slab.h> 37 38#include "mlx4_ib.h" 39#include "user.h" 40 41static void *get_wqe(struct mlx4_ib_srq *srq, int n) 42{ 43 return mlx4_buf_offset(&srq->buf, n << srq->msrq.wqe_shift); 44} 45 46static void mlx4_ib_srq_event(struct mlx4_srq *srq, enum mlx4_event type) 47{ 48 struct ib_event event; 49 struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq; 50 51 if (ibsrq->event_handler) { 52 event.device = ibsrq->device; 53 event.element.srq = ibsrq; 54 switch (type) { 55 case MLX4_EVENT_TYPE_SRQ_LIMIT: 56 event.event = IB_EVENT_SRQ_LIMIT_REACHED; 57 break; 58 case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR: 59 event.event = IB_EVENT_SRQ_ERR; 60 break; 61 default: 62 printk(KERN_WARNING "mlx4_ib: Unexpected event type %d " 63 "on SRQ %06x\n", type, srq->srqn); 64 return; 65 } 66 67 ibsrq->event_handler(&event, ibsrq->srq_context); 68 } 69} 70 71struct ib_srq *mlx4_ib_create_srq(struct ib_pd *pd, 72 struct ib_srq_init_attr *init_attr, 73 struct ib_udata *udata) 74{ 75 struct mlx4_ib_dev *dev = to_mdev(pd->device); 76 struct mlx4_ib_srq *srq; 77 struct mlx4_wqe_srq_next_seg *next; 78 struct mlx4_wqe_data_seg *scatter; 79 int desc_size; 80 int buf_size; 81 int err; 82 int i; 83 84 /* Sanity check SRQ size before proceeding */ 85 if (init_attr->attr.max_wr >= dev->dev->caps.max_srq_wqes || 86 init_attr->attr.max_sge > dev->dev->caps.max_srq_sge) 87 return ERR_PTR(-EINVAL); 88 89 srq = kmalloc(sizeof *srq, GFP_KERNEL); 90 if (!srq) 91 return ERR_PTR(-ENOMEM); 92 93 mutex_init(&srq->mutex); 94 spin_lock_init(&srq->lock); 95 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1); 96 srq->msrq.max_gs = init_attr->attr.max_sge; 97 98 desc_size = max(32UL, 99 roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) + 100 srq->msrq.max_gs * 101 sizeof (struct mlx4_wqe_data_seg))); 102 srq->msrq.wqe_shift = ilog2(desc_size); 103 104 buf_size = srq->msrq.max * desc_size; 105 106 if (pd->uobject) { 107 struct mlx4_ib_create_srq ucmd; 108 109 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) { 110 err = -EFAULT; 111 goto err_srq; 112 } 113 114 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, 115 buf_size, 0, 0); 116 if (IS_ERR(srq->umem)) { 117 err = PTR_ERR(srq->umem); 118 goto err_srq; 119 } 120 121 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(srq->umem), 122 ilog2(srq->umem->page_size), &srq->mtt); 123 if (err) 124 goto err_buf; 125 126 err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem); 127 if (err) 128 goto err_mtt; 129 130 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context), 131 ucmd.db_addr, &srq->db); 132 if (err) 133 goto err_mtt; 134 } else { 135 err = mlx4_db_alloc(dev->dev, &srq->db, 0); 136 if (err) 137 goto err_srq; 138 139 *srq->db.db = 0; 140 141 if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, &srq->buf)) { 142 err = -ENOMEM; 143 goto err_db; 144 } 145 146 srq->head = 0; 147 srq->tail = srq->msrq.max - 1; 148 srq->wqe_ctr = 0; 149 150 for (i = 0; i < srq->msrq.max; ++i) { 151 next = get_wqe(srq, i); 152 next->next_wqe_index = 153 cpu_to_be16((i + 1) & (srq->msrq.max - 1)); 154 155 for (scatter = (void *) (next + 1); 156 (void *) scatter < (void *) next + desc_size; 157 ++scatter) 158 scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY); 159 } 160 161 err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift, 162 &srq->mtt); 163 if (err) 164 goto err_buf; 165 166 err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf); 167 if (err) 168 goto err_mtt; 169 170 srq->wrid = kmalloc(srq->msrq.max * sizeof (u64), GFP_KERNEL); 171 if (!srq->wrid) { 172 err = -ENOMEM; 173 goto err_mtt; 174 } 175 } 176 177 err = mlx4_srq_alloc(dev->dev, to_mpd(pd)->pdn, &srq->mtt, 178 srq->db.dma, &srq->msrq); 179 if (err) 180 goto err_wrid; 181 182 srq->msrq.event = mlx4_ib_srq_event; 183 184 if (pd->uobject) 185 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) { 186 err = -EFAULT; 187 goto err_wrid; 188 } 189 190 init_attr->attr.max_wr = srq->msrq.max - 1; 191 192 return &srq->ibsrq; 193 194err_wrid: 195 if (pd->uobject) 196 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db); 197 else 198 kfree(srq->wrid); 199 200err_mtt: 201 mlx4_mtt_cleanup(dev->dev, &srq->mtt); 202 203err_buf: 204 if (pd->uobject) 205 ib_umem_release(srq->umem); 206 else 207 mlx4_buf_free(dev->dev, buf_size, &srq->buf); 208 209err_db: 210 if (!pd->uobject) 211 mlx4_db_free(dev->dev, &srq->db); 212 213err_srq: 214 kfree(srq); 215 216 return ERR_PTR(err); 217} 218 219int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 220 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata) 221{ 222 struct mlx4_ib_dev *dev = to_mdev(ibsrq->device); 223 struct mlx4_ib_srq *srq = to_msrq(ibsrq); 224 int ret; 225 226 /* We don't support resizing SRQs (yet?) */ 227 if (attr_mask & IB_SRQ_MAX_WR) 228 return -EINVAL; 229 230 if (attr_mask & IB_SRQ_LIMIT) { 231 if (attr->srq_limit >= srq->msrq.max) 232 return -EINVAL; 233 234 mutex_lock(&srq->mutex); 235 ret = mlx4_srq_arm(dev->dev, &srq->msrq, attr->srq_limit); 236 mutex_unlock(&srq->mutex); 237 238 if (ret) 239 return ret; 240 } 241 242 return 0; 243} 244 245int mlx4_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr) 246{ 247 struct mlx4_ib_dev *dev = to_mdev(ibsrq->device); 248 struct mlx4_ib_srq *srq = to_msrq(ibsrq); 249 int ret; 250 int limit_watermark; 251 252 ret = mlx4_srq_query(dev->dev, &srq->msrq, &limit_watermark); 253 if (ret) 254 return ret; 255 256 srq_attr->srq_limit = limit_watermark; 257 srq_attr->max_wr = srq->msrq.max - 1; 258 srq_attr->max_sge = srq->msrq.max_gs; 259 260 return 0; 261} 262 263int mlx4_ib_destroy_srq(struct ib_srq *srq) 264{ 265 struct mlx4_ib_dev *dev = to_mdev(srq->device); 266 struct mlx4_ib_srq *msrq = to_msrq(srq); 267 268 mlx4_srq_free(dev->dev, &msrq->msrq); 269 mlx4_mtt_cleanup(dev->dev, &msrq->mtt); 270 271 if (srq->uobject) { 272 mlx4_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db); 273 ib_umem_release(msrq->umem); 274 } else { 275 kfree(msrq->wrid); 276 mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift, 277 &msrq->buf); 278 mlx4_db_free(dev->dev, &msrq->db); 279 } 280 281 kfree(msrq); 282 283 return 0; 284} 285 286void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index) 287{ 288 struct mlx4_wqe_srq_next_seg *next; 289 290 /* always called with interrupts disabled. */ 291 spin_lock(&srq->lock); 292 293 next = get_wqe(srq, srq->tail); 294 next->next_wqe_index = cpu_to_be16(wqe_index); 295 srq->tail = wqe_index; 296 297 spin_unlock(&srq->lock); 298} 299 300int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 301 struct ib_recv_wr **bad_wr) 302{ 303 struct mlx4_ib_srq *srq = to_msrq(ibsrq); 304 struct mlx4_wqe_srq_next_seg *next; 305 struct mlx4_wqe_data_seg *scat; 306 unsigned long flags; 307 int err = 0; 308 int nreq; 309 int i; 310 311 spin_lock_irqsave(&srq->lock, flags); 312 313 for (nreq = 0; wr; ++nreq, wr = wr->next) { 314 if (unlikely(wr->num_sge > srq->msrq.max_gs)) { 315 err = -EINVAL; 316 *bad_wr = wr; 317 break; 318 } 319 320 if (unlikely(srq->head == srq->tail)) { 321 err = -ENOMEM; 322 *bad_wr = wr; 323 break; 324 } 325 326 srq->wrid[srq->head] = wr->wr_id; 327 328 next = get_wqe(srq, srq->head); 329 srq->head = be16_to_cpu(next->next_wqe_index); 330 scat = (struct mlx4_wqe_data_seg *) (next + 1); 331 332 for (i = 0; i < wr->num_sge; ++i) { 333 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length); 334 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey); 335 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr); 336 } 337 338 if (i < srq->msrq.max_gs) { 339 scat[i].byte_count = 0; 340 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY); 341 scat[i].addr = 0; 342 } 343 } 344 345 if (likely(nreq)) { 346 srq->wqe_ctr += nreq; 347 348 /* 349 * Make sure that descriptors are written before 350 * doorbell record. 351 */ 352 wmb(); 353 354 *srq->db.db = cpu_to_be32(srq->wqe_ctr); 355 } 356 357 spin_unlock_irqrestore(&srq->lock, flags); 358 359 return err; 360} 361