1 /* 2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 */ 32 33 #include <linux/module.h> 34 #include <linux/mlx5/qp.h> 35 #include <linux/mlx5/srq.h> 36 #include <linux/slab.h> 37 #include <rdma/ib_umem.h> 38 #include <rdma/ib_user_verbs.h> 39 40 #include "mlx5_ib.h" 41 42 /* not supported currently */ 43 static int srq_signature; 44 45 static void *get_wqe(struct mlx5_ib_srq *srq, int n) 46 { 47 return mlx5_buf_offset(&srq->buf, n << srq->msrq.wqe_shift); 48 } 49 50 static void mlx5_ib_srq_event(struct mlx5_core_srq *srq, enum mlx5_event type) 51 { 52 struct ib_event event; 53 struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq; 54 55 if (ibsrq->event_handler) { 56 event.device = ibsrq->device; 57 event.element.srq = ibsrq; 58 switch (type) { 59 case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT: 60 event.event = IB_EVENT_SRQ_LIMIT_REACHED; 61 break; 62 case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR: 63 event.event = IB_EVENT_SRQ_ERR; 64 break; 65 default: 66 pr_warn("mlx5_ib: Unexpected event type %d on SRQ %06x\n", 67 type, srq->srqn); 68 return; 69 } 70 71 ibsrq->event_handler(&event, ibsrq->srq_context); 72 } 73 } 74 75 static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq, 76 struct mlx5_srq_attr *in, 77 struct ib_udata *udata, int buf_size) 78 { 79 struct mlx5_ib_dev *dev = to_mdev(pd->device); 80 struct mlx5_ib_create_srq ucmd = {}; 81 size_t ucmdlen; 82 int err; 83 int npages; 84 int page_shift; 85 int ncont; 86 u32 offset; 87 u32 uidx = MLX5_IB_DEFAULT_UIDX; 88 89 ucmdlen = min(udata->inlen, sizeof(ucmd)); 90 91 if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) { 92 mlx5_ib_dbg(dev, "failed copy udata\n"); 93 return -EFAULT; 94 } 95 96 if (ucmd.reserved0 || ucmd.reserved1) 97 return -EINVAL; 98 99 if (udata->inlen > sizeof(ucmd) && 100 !ib_is_udata_cleared(udata, sizeof(ucmd), 101 udata->inlen - sizeof(ucmd))) 102 return -EINVAL; 103 104 if (in->type == IB_SRQT_XRC) { 105 err = get_srq_user_index(to_mucontext(pd->uobject->context), 106 &ucmd, udata->inlen, &uidx); 107 if (err) 108 return err; 109 } 110 111 srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE); 112 113 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, buf_size, 114 0, 0); 115 if (IS_ERR(srq->umem)) { 116 mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size); 117 err = PTR_ERR(srq->umem); 118 return err; 119 } 120 121 mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, &npages, 122 &page_shift, &ncont, NULL); 123 err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift, 124 &offset); 125 if (err) { 126 mlx5_ib_warn(dev, "bad offset\n"); 127 goto err_umem; 128 } 129 130 in->pas = mlx5_vzalloc(sizeof(*in->pas) * ncont); 131 if (!in->pas) { 132 err = -ENOMEM; 133 goto err_umem; 134 } 135 136 mlx5_ib_populate_pas(dev, srq->umem, page_shift, in->pas, 0); 137 138 err = mlx5_ib_db_map_user(to_mucontext(pd->uobject->context), 139 ucmd.db_addr, &srq->db); 140 if (err) { 141 mlx5_ib_dbg(dev, "map doorbell failed\n"); 142 goto err_in; 143 } 144 145 in->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT; 146 in->page_offset = offset; 147 if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 && 148 in->type == IB_SRQT_XRC) 149 in->user_index = uidx; 150 151 return 0; 152 153 err_in: 154 kvfree(in->pas); 155 156 err_umem: 157 ib_umem_release(srq->umem); 158 159 return err; 160 } 161 162 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq, 163 struct mlx5_srq_attr *in, int buf_size) 164 { 165 int err; 166 int i; 167 struct mlx5_wqe_srq_next_seg *next; 168 int page_shift; 169 int npages; 170 171 err = mlx5_db_alloc(dev->mdev, &srq->db); 172 if (err) { 173 mlx5_ib_warn(dev, "alloc dbell rec failed\n"); 174 return err; 175 } 176 177 if (mlx5_buf_alloc(dev->mdev, buf_size, &srq->buf)) { 178 mlx5_ib_dbg(dev, "buf alloc failed\n"); 179 err = -ENOMEM; 180 goto err_db; 181 } 182 page_shift = srq->buf.page_shift; 183 184 srq->head = 0; 185 srq->tail = srq->msrq.max - 1; 186 srq->wqe_ctr = 0; 187 188 for (i = 0; i < srq->msrq.max; i++) { 189 next = get_wqe(srq, i); 190 next->next_wqe_index = 191 cpu_to_be16((i + 1) & (srq->msrq.max - 1)); 192 } 193 194 npages = DIV_ROUND_UP(srq->buf.npages, 1 << (page_shift - PAGE_SHIFT)); 195 mlx5_ib_dbg(dev, "buf_size %d, page_shift %d, npages %d, calc npages %d\n", 196 buf_size, page_shift, srq->buf.npages, npages); 197 in->pas = mlx5_vzalloc(sizeof(*in->pas) * npages); 198 if (!in->pas) { 199 err = -ENOMEM; 200 goto err_buf; 201 } 202 mlx5_fill_page_array(&srq->buf, in->pas); 203 204 srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL); 205 if (!srq->wrid) { 206 mlx5_ib_dbg(dev, "kmalloc failed %lu\n", 207 (unsigned long)(srq->msrq.max * sizeof(u64))); 208 err = -ENOMEM; 209 goto err_in; 210 } 211 srq->wq_sig = !!srq_signature; 212 213 in->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT; 214 if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 && 215 in->type == IB_SRQT_XRC) 216 in->user_index = MLX5_IB_DEFAULT_UIDX; 217 218 return 0; 219 220 err_in: 221 kvfree(in->pas); 222 223 err_buf: 224 mlx5_buf_free(dev->mdev, &srq->buf); 225 226 err_db: 227 mlx5_db_free(dev->mdev, &srq->db); 228 return err; 229 } 230 231 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq) 232 { 233 mlx5_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db); 234 ib_umem_release(srq->umem); 235 } 236 237 238 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq) 239 { 240 kfree(srq->wrid); 241 mlx5_buf_free(dev->mdev, &srq->buf); 242 mlx5_db_free(dev->mdev, &srq->db); 243 } 244 245 struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd, 246 struct ib_srq_init_attr *init_attr, 247 struct ib_udata *udata) 248 { 249 struct mlx5_ib_dev *dev = to_mdev(pd->device); 250 struct mlx5_ib_srq *srq; 251 int desc_size; 252 int buf_size; 253 int err; 254 struct mlx5_srq_attr in = {0}; 255 __u32 max_srq_wqes = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz); 256 257 /* Sanity check SRQ size before proceeding */ 258 if (init_attr->attr.max_wr >= max_srq_wqes) { 259 mlx5_ib_dbg(dev, "max_wr %d, cap %d\n", 260 init_attr->attr.max_wr, 261 max_srq_wqes); 262 return ERR_PTR(-EINVAL); 263 } 264 265 srq = kmalloc(sizeof(*srq), GFP_KERNEL); 266 if (!srq) 267 return ERR_PTR(-ENOMEM); 268 269 mutex_init(&srq->mutex); 270 spin_lock_init(&srq->lock); 271 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1); 272 srq->msrq.max_gs = init_attr->attr.max_sge; 273 274 desc_size = sizeof(struct mlx5_wqe_srq_next_seg) + 275 srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg); 276 desc_size = roundup_pow_of_two(desc_size); 277 desc_size = max_t(int, 32, desc_size); 278 srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) / 279 sizeof(struct mlx5_wqe_data_seg); 280 srq->msrq.wqe_shift = ilog2(desc_size); 281 buf_size = srq->msrq.max * desc_size; 282 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", 283 desc_size, init_attr->attr.max_wr, srq->msrq.max, srq->msrq.max_gs, 284 srq->msrq.max_avail_gather); 285 286 if (pd->uobject) 287 err = create_srq_user(pd, srq, &in, udata, buf_size); 288 else 289 err = create_srq_kernel(dev, srq, &in, buf_size); 290 291 if (err) { 292 mlx5_ib_warn(dev, "create srq %s failed, err %d\n", 293 pd->uobject ? "user" : "kernel", err); 294 goto err_srq; 295 } 296 297 in.type = init_attr->srq_type; 298 in.log_size = ilog2(srq->msrq.max); 299 in.wqe_shift = srq->msrq.wqe_shift - 4; 300 if (srq->wq_sig) 301 in.flags |= MLX5_SRQ_FLAG_WQ_SIG; 302 if (init_attr->srq_type == IB_SRQT_XRC) { 303 in.xrcd = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn; 304 in.cqn = to_mcq(init_attr->ext.xrc.cq)->mcq.cqn; 305 } else if (init_attr->srq_type == IB_SRQT_BASIC) { 306 in.xrcd = to_mxrcd(dev->devr.x0)->xrcdn; 307 in.cqn = to_mcq(dev->devr.c0)->mcq.cqn; 308 } 309 310 in.pd = to_mpd(pd)->pdn; 311 in.db_record = srq->db.dma; 312 err = mlx5_core_create_srq(dev->mdev, &srq->msrq, &in); 313 kvfree(in.pas); 314 if (err) { 315 mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err); 316 goto err_usr_kern_srq; 317 } 318 319 mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn); 320 321 srq->msrq.event = mlx5_ib_srq_event; 322 srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn; 323 324 if (pd->uobject) 325 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) { 326 mlx5_ib_dbg(dev, "copy to user failed\n"); 327 err = -EFAULT; 328 goto err_core; 329 } 330 331 init_attr->attr.max_wr = srq->msrq.max - 1; 332 333 return &srq->ibsrq; 334 335 err_core: 336 mlx5_core_destroy_srq(dev->mdev, &srq->msrq); 337 338 err_usr_kern_srq: 339 if (pd->uobject) 340 destroy_srq_user(pd, srq); 341 else 342 destroy_srq_kernel(dev, srq); 343 344 err_srq: 345 kfree(srq); 346 347 return ERR_PTR(err); 348 } 349 350 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 351 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata) 352 { 353 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 354 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 355 int ret; 356 357 /* We don't support resizing SRQs yet */ 358 if (attr_mask & IB_SRQ_MAX_WR) 359 return -EINVAL; 360 361 if (attr_mask & IB_SRQ_LIMIT) { 362 if (attr->srq_limit >= srq->msrq.max) 363 return -EINVAL; 364 365 mutex_lock(&srq->mutex); 366 ret = mlx5_core_arm_srq(dev->mdev, &srq->msrq, attr->srq_limit, 1); 367 mutex_unlock(&srq->mutex); 368 369 if (ret) 370 return ret; 371 } 372 373 return 0; 374 } 375 376 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr) 377 { 378 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 379 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 380 int ret; 381 struct mlx5_srq_attr *out; 382 383 out = kzalloc(sizeof(*out), GFP_KERNEL); 384 if (!out) 385 return -ENOMEM; 386 387 ret = mlx5_core_query_srq(dev->mdev, &srq->msrq, out); 388 if (ret) 389 goto out_box; 390 391 srq_attr->srq_limit = out->lwm; 392 srq_attr->max_wr = srq->msrq.max - 1; 393 srq_attr->max_sge = srq->msrq.max_gs; 394 395 out_box: 396 kfree(out); 397 return ret; 398 } 399 400 int mlx5_ib_destroy_srq(struct ib_srq *srq) 401 { 402 struct mlx5_ib_dev *dev = to_mdev(srq->device); 403 struct mlx5_ib_srq *msrq = to_msrq(srq); 404 405 mlx5_core_destroy_srq(dev->mdev, &msrq->msrq); 406 407 if (srq->uobject) { 408 mlx5_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db); 409 ib_umem_release(msrq->umem); 410 } else { 411 destroy_srq_kernel(dev, msrq); 412 } 413 414 kfree(srq); 415 return 0; 416 } 417 418 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index) 419 { 420 struct mlx5_wqe_srq_next_seg *next; 421 422 /* always called with interrupts disabled. */ 423 spin_lock(&srq->lock); 424 425 next = get_wqe(srq, srq->tail); 426 next->next_wqe_index = cpu_to_be16(wqe_index); 427 srq->tail = wqe_index; 428 429 spin_unlock(&srq->lock); 430 } 431 432 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 433 struct ib_recv_wr **bad_wr) 434 { 435 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 436 struct mlx5_wqe_srq_next_seg *next; 437 struct mlx5_wqe_data_seg *scat; 438 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 439 struct mlx5_core_dev *mdev = dev->mdev; 440 unsigned long flags; 441 int err = 0; 442 int nreq; 443 int i; 444 445 spin_lock_irqsave(&srq->lock, flags); 446 447 if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) { 448 err = -EIO; 449 *bad_wr = wr; 450 goto out; 451 } 452 453 for (nreq = 0; wr; nreq++, wr = wr->next) { 454 if (unlikely(wr->num_sge > srq->msrq.max_gs)) { 455 err = -EINVAL; 456 *bad_wr = wr; 457 break; 458 } 459 460 if (unlikely(srq->head == srq->tail)) { 461 err = -ENOMEM; 462 *bad_wr = wr; 463 break; 464 } 465 466 srq->wrid[srq->head] = wr->wr_id; 467 468 next = get_wqe(srq, srq->head); 469 srq->head = be16_to_cpu(next->next_wqe_index); 470 scat = (struct mlx5_wqe_data_seg *)(next + 1); 471 472 for (i = 0; i < wr->num_sge; i++) { 473 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length); 474 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey); 475 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr); 476 } 477 478 if (i < srq->msrq.max_avail_gather) { 479 scat[i].byte_count = 0; 480 scat[i].lkey = cpu_to_be32(MLX5_INVALID_LKEY); 481 scat[i].addr = 0; 482 } 483 } 484 485 if (likely(nreq)) { 486 srq->wqe_ctr += nreq; 487 488 /* Make sure that descriptors are written before 489 * doorbell record. 490 */ 491 wmb(); 492 493 *srq->db.db = cpu_to_be32(srq->wqe_ctr); 494 } 495 out: 496 spin_unlock_irqrestore(&srq->lock, flags); 497 498 return err; 499 } 500