1 /*- 2 * Copyright (c) 2013-2015, Mellanox Technologies, Ltd. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 * $FreeBSD$ 26 */ 27 28 #include <linux/module.h> 29 #include <dev/mlx5/qp.h> 30 #include <dev/mlx5/srq.h> 31 #include <linux/slab.h> 32 #include <rdma/ib_umem.h> 33 #include <rdma/ib_user_verbs.h> 34 35 #include "mlx5_ib.h" 36 37 /* not supported currently */ 38 static int srq_signature; 39 40 static void *get_wqe(struct mlx5_ib_srq *srq, int n) 41 { 42 return mlx5_buf_offset(&srq->buf, n << srq->msrq.wqe_shift); 43 } 44 45 static void mlx5_ib_srq_event(struct mlx5_core_srq *srq, int 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 MLX5_EVENT_TYPE_SRQ_RQ_LIMIT: 55 event.event = IB_EVENT_SRQ_LIMIT_REACHED; 56 break; 57 case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR: 58 event.event = IB_EVENT_SRQ_ERR; 59 break; 60 default: 61 pr_warn("mlx5_ib: Unexpected event type %d on SRQ %06x\n", 62 type, srq->srqn); 63 return; 64 } 65 66 ibsrq->event_handler(&event, ibsrq->srq_context); 67 } 68 } 69 70 static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq, 71 struct mlx5_srq_attr *in, 72 struct ib_udata *udata, int buf_size) 73 { 74 struct mlx5_ib_dev *dev = to_mdev(pd->device); 75 struct mlx5_ib_create_srq ucmd = {}; 76 size_t ucmdlen; 77 int err; 78 int npages; 79 int page_shift; 80 int ncont; 81 u32 offset; 82 u32 uidx = MLX5_IB_DEFAULT_UIDX; 83 84 ucmdlen = min(udata->inlen, sizeof(ucmd)); 85 86 if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) { 87 mlx5_ib_dbg(dev, "failed copy udata\n"); 88 return -EFAULT; 89 } 90 91 if (ucmd.reserved0 || ucmd.reserved1) 92 return -EINVAL; 93 94 if (udata->inlen > sizeof(ucmd) && 95 !ib_is_udata_cleared(udata, sizeof(ucmd), 96 udata->inlen - sizeof(ucmd))) 97 return -EINVAL; 98 99 if (in->type == IB_SRQT_XRC) { 100 err = get_srq_user_index(to_mucontext(pd->uobject->context), 101 &ucmd, udata->inlen, &uidx); 102 if (err) 103 return err; 104 } 105 106 srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE); 107 108 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, buf_size, 109 0, 0); 110 if (IS_ERR(srq->umem)) { 111 mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size); 112 err = PTR_ERR(srq->umem); 113 return err; 114 } 115 116 mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, &npages, 117 &page_shift, &ncont, NULL); 118 err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift, 119 &offset); 120 if (err) { 121 mlx5_ib_warn(dev, "bad offset\n"); 122 goto err_umem; 123 } 124 125 in->pas = mlx5_vzalloc(sizeof(*in->pas) * ncont); 126 if (!in->pas) { 127 err = -ENOMEM; 128 goto err_umem; 129 } 130 131 mlx5_ib_populate_pas(dev, srq->umem, page_shift, in->pas, 0); 132 133 err = mlx5_ib_db_map_user(to_mucontext(pd->uobject->context), 134 ucmd.db_addr, &srq->db); 135 if (err) { 136 mlx5_ib_dbg(dev, "map doorbell failed\n"); 137 goto err_in; 138 } 139 140 in->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT; 141 in->page_offset = offset; 142 if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 && 143 in->type == IB_SRQT_XRC) 144 in->user_index = uidx; 145 return 0; 146 147 err_in: 148 kvfree(in->pas); 149 150 err_umem: 151 ib_umem_release(srq->umem); 152 153 return err; 154 } 155 156 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq, 157 struct mlx5_srq_attr *in, int buf_size) 158 { 159 int err; 160 int i; 161 struct mlx5_wqe_srq_next_seg *next; 162 int page_shift; 163 int npages; 164 165 err = mlx5_db_alloc(dev->mdev, &srq->db); 166 if (err) { 167 mlx5_ib_warn(dev, "alloc dbell rec failed\n"); 168 return err; 169 } 170 171 if (mlx5_buf_alloc(dev->mdev, buf_size, 2 * PAGE_SIZE, &srq->buf)) { 172 mlx5_ib_dbg(dev, "buf alloc failed\n"); 173 err = -ENOMEM; 174 goto err_db; 175 } 176 page_shift = srq->buf.page_shift; 177 178 srq->head = 0; 179 srq->tail = srq->msrq.max - 1; 180 srq->wqe_ctr = 0; 181 182 for (i = 0; i < srq->msrq.max; i++) { 183 next = get_wqe(srq, i); 184 next->next_wqe_index = 185 cpu_to_be16((i + 1) & (srq->msrq.max - 1)); 186 } 187 188 npages = DIV_ROUND_UP(srq->buf.npages, 1 << (page_shift - PAGE_SHIFT)); 189 mlx5_ib_dbg(dev, "buf_size %d, page_shift %d, npages %d, calc npages %d\n", 190 buf_size, page_shift, srq->buf.npages, npages); 191 in->pas = mlx5_vzalloc(sizeof(*in->pas) * npages); 192 if (!in->pas) { 193 err = -ENOMEM; 194 goto err_buf; 195 } 196 mlx5_fill_page_array(&srq->buf, in->pas); 197 198 srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL); 199 if (!srq->wrid) { 200 mlx5_ib_dbg(dev, "kmalloc failed %lu\n", 201 (unsigned long)(srq->msrq.max * sizeof(u64))); 202 err = -ENOMEM; 203 goto err_in; 204 } 205 srq->wq_sig = !!srq_signature; 206 207 in->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT; 208 if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 && 209 in->type == IB_SRQT_XRC) 210 in->user_index = MLX5_IB_DEFAULT_UIDX; 211 212 return 0; 213 214 err_in: 215 kvfree(in->pas); 216 217 err_buf: 218 mlx5_buf_free(dev->mdev, &srq->buf); 219 220 err_db: 221 mlx5_db_free(dev->mdev, &srq->db); 222 return err; 223 } 224 225 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq) 226 { 227 mlx5_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db); 228 ib_umem_release(srq->umem); 229 } 230 231 232 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq) 233 { 234 kfree(srq->wrid); 235 mlx5_buf_free(dev->mdev, &srq->buf); 236 mlx5_db_free(dev->mdev, &srq->db); 237 } 238 239 struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd, 240 struct ib_srq_init_attr *init_attr, 241 struct ib_udata *udata) 242 { 243 struct mlx5_ib_dev *dev = to_mdev(pd->device); 244 struct mlx5_ib_srq *srq; 245 int desc_size; 246 int buf_size; 247 int err; 248 struct mlx5_srq_attr in = {0}; 249 __u32 max_srq_wqes = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz); 250 251 /* Sanity check SRQ size before proceeding */ 252 if (init_attr->attr.max_wr >= max_srq_wqes) { 253 mlx5_ib_dbg(dev, "max_wr %d, cap %d\n", 254 init_attr->attr.max_wr, 255 max_srq_wqes); 256 return ERR_PTR(-EINVAL); 257 } 258 259 srq = kmalloc(sizeof(*srq), GFP_KERNEL); 260 if (!srq) 261 return ERR_PTR(-ENOMEM); 262 263 mutex_init(&srq->mutex); 264 spin_lock_init(&srq->lock); 265 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1); 266 srq->msrq.max_gs = init_attr->attr.max_sge; 267 268 desc_size = sizeof(struct mlx5_wqe_srq_next_seg) + 269 srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg); 270 desc_size = roundup_pow_of_two(desc_size); 271 desc_size = max_t(int, 32, desc_size); 272 srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) / 273 sizeof(struct mlx5_wqe_data_seg); 274 srq->msrq.wqe_shift = ilog2(desc_size); 275 buf_size = srq->msrq.max * desc_size; 276 mlx5_ib_dbg(dev, "desc_size 0x%x, req wr 0x%x, srq size 0x%x, max_gs 0x%x, max_avail_gather 0x%x\n", 277 desc_size, init_attr->attr.max_wr, srq->msrq.max, srq->msrq.max_gs, 278 srq->msrq.max_avail_gather); 279 280 if (pd->uobject) 281 err = create_srq_user(pd, srq, &in, udata, buf_size); 282 else 283 err = create_srq_kernel(dev, srq, &in, buf_size); 284 285 if (err) { 286 mlx5_ib_warn(dev, "create srq %s failed, err %d\n", 287 pd->uobject ? "user" : "kernel", err); 288 goto err_srq; 289 } 290 291 in.type = init_attr->srq_type; 292 in.log_size = ilog2(srq->msrq.max); 293 in.wqe_shift = srq->msrq.wqe_shift - 4; 294 if (srq->wq_sig) 295 in.flags |= MLX5_SRQ_FLAG_WQ_SIG; 296 if (init_attr->srq_type == IB_SRQT_XRC) { 297 in.xrcd = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn; 298 in.cqn = to_mcq(init_attr->ext.xrc.cq)->mcq.cqn; 299 } else if (init_attr->srq_type == IB_SRQT_BASIC) { 300 in.xrcd = to_mxrcd(dev->devr.x0)->xrcdn; 301 in.cqn = to_mcq(dev->devr.c0)->mcq.cqn; 302 } 303 304 in.pd = to_mpd(pd)->pdn; 305 in.db_record = srq->db.dma; 306 err = mlx5_core_create_srq(dev->mdev, &srq->msrq, &in); 307 kvfree(in.pas); 308 if (err) { 309 mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err); 310 goto err_usr_kern_srq; 311 } 312 313 mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn); 314 315 srq->msrq.event = mlx5_ib_srq_event; 316 srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn; 317 318 if (pd->uobject) 319 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) { 320 mlx5_ib_dbg(dev, "copy to user failed\n"); 321 err = -EFAULT; 322 goto err_core; 323 } 324 325 init_attr->attr.max_wr = srq->msrq.max - 1; 326 327 return &srq->ibsrq; 328 329 err_core: 330 mlx5_core_destroy_srq(dev->mdev, &srq->msrq); 331 332 err_usr_kern_srq: 333 if (pd->uobject) 334 destroy_srq_user(pd, srq); 335 else 336 destroy_srq_kernel(dev, srq); 337 338 err_srq: 339 kfree(srq); 340 341 return ERR_PTR(err); 342 } 343 344 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 345 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata) 346 { 347 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 348 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 349 int ret; 350 351 /* We don't support resizing SRQs yet */ 352 if (attr_mask & IB_SRQ_MAX_WR) 353 return -EINVAL; 354 355 if (attr_mask & IB_SRQ_LIMIT) { 356 if (attr->srq_limit >= srq->msrq.max) 357 return -EINVAL; 358 359 mutex_lock(&srq->mutex); 360 ret = mlx5_core_arm_srq(dev->mdev, &srq->msrq, attr->srq_limit, 1); 361 mutex_unlock(&srq->mutex); 362 363 if (ret) 364 return ret; 365 } 366 367 return 0; 368 } 369 370 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr) 371 { 372 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 373 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 374 int ret; 375 struct mlx5_srq_attr *out; 376 377 out = kzalloc(sizeof(*out), GFP_KERNEL); 378 if (!out) 379 return -ENOMEM; 380 381 ret = mlx5_core_query_srq(dev->mdev, &srq->msrq, out); 382 if (ret) 383 goto out_box; 384 385 srq_attr->srq_limit = out->lwm; 386 srq_attr->max_wr = srq->msrq.max - 1; 387 srq_attr->max_sge = srq->msrq.max_gs; 388 389 out_box: 390 kfree(out); 391 return ret; 392 } 393 394 int mlx5_ib_destroy_srq(struct ib_srq *srq) 395 { 396 struct mlx5_ib_dev *dev = to_mdev(srq->device); 397 struct mlx5_ib_srq *msrq = to_msrq(srq); 398 399 mlx5_core_destroy_srq(dev->mdev, &msrq->msrq); 400 401 if (srq->uobject) { 402 mlx5_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db); 403 ib_umem_release(msrq->umem); 404 } else { 405 destroy_srq_kernel(dev, msrq); 406 } 407 408 kfree(srq); 409 return 0; 410 } 411 412 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index) 413 { 414 struct mlx5_wqe_srq_next_seg *next; 415 416 /* always called with interrupts disabled. */ 417 spin_lock(&srq->lock); 418 419 next = get_wqe(srq, srq->tail); 420 next->next_wqe_index = cpu_to_be16(wqe_index); 421 srq->tail = wqe_index; 422 423 spin_unlock(&srq->lock); 424 } 425 426 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 427 struct ib_recv_wr **bad_wr) 428 { 429 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 430 struct mlx5_wqe_srq_next_seg *next; 431 struct mlx5_wqe_data_seg *scat; 432 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 433 struct mlx5_core_dev *mdev = dev->mdev; 434 unsigned long flags; 435 int err = 0; 436 int nreq; 437 int i; 438 439 spin_lock_irqsave(&srq->lock, flags); 440 441 if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) { 442 err = -EIO; 443 *bad_wr = wr; 444 goto out; 445 } 446 447 for (nreq = 0; wr; nreq++, wr = wr->next) { 448 if (unlikely(wr->num_sge > srq->msrq.max_gs)) { 449 err = -EINVAL; 450 *bad_wr = wr; 451 break; 452 } 453 454 if (unlikely(srq->head == srq->tail)) { 455 err = -ENOMEM; 456 *bad_wr = wr; 457 break; 458 } 459 460 srq->wrid[srq->head] = wr->wr_id; 461 462 next = get_wqe(srq, srq->head); 463 srq->head = be16_to_cpu(next->next_wqe_index); 464 scat = (struct mlx5_wqe_data_seg *)(next + 1); 465 466 for (i = 0; i < wr->num_sge; i++) { 467 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length); 468 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey); 469 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr); 470 } 471 472 if (i < srq->msrq.max_avail_gather) { 473 scat[i].byte_count = 0; 474 scat[i].lkey = cpu_to_be32(MLX5_INVALID_LKEY); 475 scat[i].addr = 0; 476 } 477 } 478 479 if (likely(nreq)) { 480 srq->wqe_ctr += nreq; 481 482 /* Make sure that descriptors are written before 483 * doorbell record. 484 */ 485 wmb(); 486 487 *srq->db.db = cpu_to_be32(srq->wqe_ctr); 488 } 489 out: 490 spin_unlock_irqrestore(&srq->lock, flags); 491 492 return err; 493 } 494