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 <rdma/mlx4-abi.h> 40 #include <rdma/uverbs_ioctl.h> 41 42 static void *get_wqe(struct mlx4_ib_srq *srq, int n) 43 { 44 return mlx4_buf_offset(&srq->buf, n << srq->msrq.wqe_shift); 45 } 46 47 static void mlx4_ib_srq_event(struct mlx4_srq *srq, enum mlx4_event type) 48 { 49 struct ib_event event; 50 struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq; 51 52 if (ibsrq->event_handler) { 53 event.device = ibsrq->device; 54 event.element.srq = ibsrq; 55 switch (type) { 56 case MLX4_EVENT_TYPE_SRQ_LIMIT: 57 event.event = IB_EVENT_SRQ_LIMIT_REACHED; 58 break; 59 case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR: 60 event.event = IB_EVENT_SRQ_ERR; 61 break; 62 default: 63 pr_warn("Unexpected event type %d " 64 "on SRQ %06x\n", type, srq->srqn); 65 return; 66 } 67 68 ibsrq->event_handler(&event, ibsrq->srq_context); 69 } 70 } 71 72 int mlx4_ib_create_srq(struct ib_srq *ib_srq, 73 struct ib_srq_init_attr *init_attr, 74 struct ib_udata *udata) 75 { 76 struct mlx4_ib_dev *dev = to_mdev(ib_srq->device); 77 struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context( 78 udata, struct mlx4_ib_ucontext, ibucontext); 79 struct mlx4_ib_srq *srq = to_msrq(ib_srq); 80 struct mlx4_wqe_srq_next_seg *next; 81 struct mlx4_wqe_data_seg *scatter; 82 u32 cqn; 83 u16 xrcdn; 84 int desc_size; 85 int buf_size; 86 int err; 87 int i; 88 89 /* Sanity check SRQ size before proceeding */ 90 if (init_attr->attr.max_wr >= dev->dev->caps.max_srq_wqes || 91 init_attr->attr.max_sge > dev->dev->caps.max_srq_sge) 92 return -EINVAL; 93 94 mutex_init(&srq->mutex); 95 spin_lock_init(&srq->lock); 96 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1); 97 srq->msrq.max_gs = init_attr->attr.max_sge; 98 99 desc_size = max(32UL, 100 roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) + 101 srq->msrq.max_gs * 102 sizeof (struct mlx4_wqe_data_seg))); 103 srq->msrq.wqe_shift = ilog2(desc_size); 104 105 buf_size = srq->msrq.max * desc_size; 106 107 if (udata) { 108 struct mlx4_ib_create_srq ucmd; 109 110 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) 111 return -EFAULT; 112 113 srq->umem = 114 ib_umem_get(ib_srq->device, ucmd.buf_addr, buf_size, 0); 115 if (IS_ERR(srq->umem)) 116 return PTR_ERR(srq->umem); 117 118 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(srq->umem), 119 PAGE_SHIFT, &srq->mtt); 120 if (err) 121 goto err_buf; 122 123 err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem); 124 if (err) 125 goto err_mtt; 126 127 err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &srq->db); 128 if (err) 129 goto err_mtt; 130 } else { 131 err = mlx4_db_alloc(dev->dev, &srq->db, 0); 132 if (err) 133 return err; 134 135 *srq->db.db = 0; 136 137 if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, 138 &srq->buf)) { 139 err = -ENOMEM; 140 goto err_db; 141 } 142 143 srq->head = 0; 144 srq->tail = srq->msrq.max - 1; 145 srq->wqe_ctr = 0; 146 147 for (i = 0; i < srq->msrq.max; ++i) { 148 next = get_wqe(srq, i); 149 next->next_wqe_index = 150 cpu_to_be16((i + 1) & (srq->msrq.max - 1)); 151 152 for (scatter = (void *) (next + 1); 153 (void *) scatter < (void *) next + desc_size; 154 ++scatter) 155 scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY); 156 } 157 158 err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift, 159 &srq->mtt); 160 if (err) 161 goto err_buf; 162 163 err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf); 164 if (err) 165 goto err_mtt; 166 167 srq->wrid = kvmalloc_array(srq->msrq.max, 168 sizeof(u64), GFP_KERNEL); 169 if (!srq->wrid) { 170 err = -ENOMEM; 171 goto err_mtt; 172 } 173 } 174 175 cqn = ib_srq_has_cq(init_attr->srq_type) ? 176 to_mcq(init_attr->ext.cq)->mcq.cqn : 0; 177 xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ? 178 to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn : 179 (u16) dev->dev->caps.reserved_xrcds; 180 err = mlx4_srq_alloc(dev->dev, to_mpd(ib_srq->pd)->pdn, cqn, xrcdn, 181 &srq->mtt, srq->db.dma, &srq->msrq); 182 if (err) 183 goto err_wrid; 184 185 srq->msrq.event = mlx4_ib_srq_event; 186 srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn; 187 188 if (udata) 189 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) { 190 err = -EFAULT; 191 goto err_wrid; 192 } 193 194 init_attr->attr.max_wr = srq->msrq.max - 1; 195 196 return 0; 197 198 err_wrid: 199 if (udata) 200 mlx4_ib_db_unmap_user(ucontext, &srq->db); 201 else 202 kvfree(srq->wrid); 203 204 err_mtt: 205 mlx4_mtt_cleanup(dev->dev, &srq->mtt); 206 207 err_buf: 208 if (!srq->umem) 209 mlx4_buf_free(dev->dev, buf_size, &srq->buf); 210 ib_umem_release(srq->umem); 211 212 err_db: 213 if (!udata) 214 mlx4_db_free(dev->dev, &srq->db); 215 216 return err; 217 } 218 219 int 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 } 244 245 int 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 } 262 263 void mlx4_ib_destroy_srq(struct ib_srq *srq, struct ib_udata *udata) 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 (udata) { 272 mlx4_ib_db_unmap_user( 273 rdma_udata_to_drv_context( 274 udata, 275 struct mlx4_ib_ucontext, 276 ibucontext), 277 &msrq->db); 278 } else { 279 kvfree(msrq->wrid); 280 mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift, 281 &msrq->buf); 282 mlx4_db_free(dev->dev, &msrq->db); 283 } 284 ib_umem_release(msrq->umem); 285 } 286 287 void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index) 288 { 289 struct mlx4_wqe_srq_next_seg *next; 290 291 /* always called with interrupts disabled. */ 292 spin_lock(&srq->lock); 293 294 next = get_wqe(srq, srq->tail); 295 next->next_wqe_index = cpu_to_be16(wqe_index); 296 srq->tail = wqe_index; 297 298 spin_unlock(&srq->lock); 299 } 300 301 int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr, 302 const struct ib_recv_wr **bad_wr) 303 { 304 struct mlx4_ib_srq *srq = to_msrq(ibsrq); 305 struct mlx4_wqe_srq_next_seg *next; 306 struct mlx4_wqe_data_seg *scat; 307 unsigned long flags; 308 int err = 0; 309 int nreq; 310 int i; 311 struct mlx4_ib_dev *mdev = to_mdev(ibsrq->device); 312 313 spin_lock_irqsave(&srq->lock, flags); 314 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) { 315 err = -EIO; 316 *bad_wr = wr; 317 nreq = 0; 318 goto out; 319 } 320 321 for (nreq = 0; wr; ++nreq, wr = wr->next) { 322 if (unlikely(wr->num_sge > srq->msrq.max_gs)) { 323 err = -EINVAL; 324 *bad_wr = wr; 325 break; 326 } 327 328 if (unlikely(srq->head == srq->tail)) { 329 err = -ENOMEM; 330 *bad_wr = wr; 331 break; 332 } 333 334 srq->wrid[srq->head] = wr->wr_id; 335 336 next = get_wqe(srq, srq->head); 337 srq->head = be16_to_cpu(next->next_wqe_index); 338 scat = (struct mlx4_wqe_data_seg *) (next + 1); 339 340 for (i = 0; i < wr->num_sge; ++i) { 341 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length); 342 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey); 343 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr); 344 } 345 346 if (i < srq->msrq.max_gs) { 347 scat[i].byte_count = 0; 348 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY); 349 scat[i].addr = 0; 350 } 351 } 352 353 if (likely(nreq)) { 354 srq->wqe_ctr += nreq; 355 356 /* 357 * Make sure that descriptors are written before 358 * doorbell record. 359 */ 360 wmb(); 361 362 *srq->db.db = cpu_to_be32(srq->wqe_ctr); 363 } 364 out: 365 366 spin_unlock_irqrestore(&srq->lock, flags); 367 368 return err; 369 } 370