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_create_srq_mbox_in **in, 72 struct ib_udata *udata, int buf_size, int *inlen, 73 int type) 74 { 75 struct mlx5_ib_dev *dev = to_mdev(pd->device); 76 struct mlx5_ib_create_srq ucmd = {}; 77 size_t ucmdlen; 78 int err; 79 int npages; 80 int page_shift; 81 int ncont; 82 u32 offset; 83 u32 uidx = MLX5_IB_DEFAULT_UIDX; 84 85 ucmdlen = min(udata->inlen, sizeof(ucmd)); 86 87 if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) { 88 mlx5_ib_dbg(dev, "failed copy udata\n"); 89 return -EFAULT; 90 } 91 92 if (ucmd.reserved0 || ucmd.reserved1) 93 return -EINVAL; 94 95 if (udata->inlen > sizeof(ucmd) && 96 !ib_is_udata_cleared(udata, sizeof(ucmd), 97 udata->inlen - sizeof(ucmd))) 98 return -EINVAL; 99 100 if (type == IB_SRQT_XRC) { 101 err = get_srq_user_index(to_mucontext(pd->uobject->context), 102 &ucmd, udata->inlen, &uidx); 103 if (err) 104 return err; 105 } 106 107 srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE); 108 109 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, buf_size, 110 0, 0); 111 if (IS_ERR(srq->umem)) { 112 mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size); 113 err = PTR_ERR(srq->umem); 114 return err; 115 } 116 117 mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, &npages, 118 &page_shift, &ncont, NULL); 119 err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift, 120 &offset); 121 if (err) { 122 mlx5_ib_warn(dev, "bad offset\n"); 123 goto err_umem; 124 } 125 126 *inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont; 127 *in = mlx5_vzalloc(*inlen); 128 if (!(*in)) { 129 err = -ENOMEM; 130 goto err_umem; 131 } 132 133 mlx5_ib_populate_pas(dev, srq->umem, page_shift, (*in)->pas, 0); 134 135 err = mlx5_ib_db_map_user(to_mucontext(pd->uobject->context), 136 ucmd.db_addr, &srq->db); 137 if (err) { 138 mlx5_ib_dbg(dev, "map doorbell failed\n"); 139 goto err_in; 140 } 141 142 (*in)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT; 143 (*in)->ctx.pgoff_cqn = cpu_to_be32(offset << 26); 144 145 if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 && 146 type == IB_SRQT_XRC) { 147 void *xsrqc = MLX5_ADDR_OF(create_xrc_srq_in, *in, 148 xrc_srq_context_entry); 149 MLX5_SET(xrc_srqc, xsrqc, user_index, uidx); 150 } 151 152 return 0; 153 154 err_in: 155 kvfree(*in); 156 157 err_umem: 158 ib_umem_release(srq->umem); 159 160 return err; 161 } 162 163 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq, 164 struct mlx5_create_srq_mbox_in **in, int buf_size, 165 int *inlen, int type) 166 { 167 int err; 168 int i; 169 struct mlx5_wqe_srq_next_seg *next; 170 int page_shift; 171 int npages; 172 173 err = mlx5_db_alloc(dev->mdev, &srq->db); 174 if (err) { 175 mlx5_ib_warn(dev, "alloc dbell rec failed\n"); 176 return err; 177 } 178 179 if (mlx5_buf_alloc(dev->mdev, buf_size, 2 * PAGE_SIZE, &srq->buf)) { 180 mlx5_ib_dbg(dev, "buf alloc failed\n"); 181 err = -ENOMEM; 182 goto err_db; 183 } 184 page_shift = srq->buf.page_shift; 185 186 srq->head = 0; 187 srq->tail = srq->msrq.max - 1; 188 srq->wqe_ctr = 0; 189 190 for (i = 0; i < srq->msrq.max; i++) { 191 next = get_wqe(srq, i); 192 next->next_wqe_index = 193 cpu_to_be16((i + 1) & (srq->msrq.max - 1)); 194 } 195 196 npages = DIV_ROUND_UP(srq->buf.npages, 1 << (page_shift - PAGE_SHIFT)); 197 mlx5_ib_dbg(dev, "buf_size %d, page_shift %d, npages %d, calc npages %d\n", 198 buf_size, page_shift, srq->buf.npages, npages); 199 *inlen = sizeof(**in) + sizeof(*(*in)->pas) * npages; 200 *in = mlx5_vzalloc(*inlen); 201 if (!*in) { 202 err = -ENOMEM; 203 goto err_buf; 204 } 205 mlx5_fill_page_array(&srq->buf, (*in)->pas); 206 207 srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL); 208 if (!srq->wrid) { 209 mlx5_ib_dbg(dev, "kmalloc failed %lu\n", 210 (unsigned long)(srq->msrq.max * sizeof(u64))); 211 err = -ENOMEM; 212 goto err_in; 213 } 214 srq->wq_sig = !!srq_signature; 215 216 (*in)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT; 217 218 if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 && 219 type == IB_SRQT_XRC) { 220 void *xsrqc = MLX5_ADDR_OF(create_xrc_srq_in, *in, 221 xrc_srq_context_entry); 222 MLX5_SET(xrc_srqc, xsrqc, user_index, MLX5_IB_DEFAULT_UIDX); 223 } 224 225 return 0; 226 227 err_in: 228 kvfree(*in); 229 230 err_buf: 231 mlx5_buf_free(dev->mdev, &srq->buf); 232 233 err_db: 234 mlx5_db_free(dev->mdev, &srq->db); 235 return err; 236 } 237 238 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq) 239 { 240 mlx5_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db); 241 ib_umem_release(srq->umem); 242 } 243 244 245 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq) 246 { 247 kfree(srq->wrid); 248 mlx5_buf_free(dev->mdev, &srq->buf); 249 mlx5_db_free(dev->mdev, &srq->db); 250 } 251 252 struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd, 253 struct ib_srq_init_attr *init_attr, 254 struct ib_udata *udata) 255 { 256 struct mlx5_ib_dev *dev = to_mdev(pd->device); 257 struct mlx5_ib_srq *srq; 258 int desc_size; 259 int buf_size; 260 int err; 261 struct mlx5_create_srq_mbox_in *uninitialized_var(in); 262 int uninitialized_var(inlen); 263 int is_xrc; 264 u32 flgs, xrcdn; 265 __u32 max_srq_wqes = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz); 266 267 /* Sanity check SRQ size before proceeding */ 268 if (init_attr->attr.max_wr >= max_srq_wqes) { 269 mlx5_ib_dbg(dev, "max_wr %d, cap %d\n", 270 init_attr->attr.max_wr, 271 max_srq_wqes); 272 return ERR_PTR(-EINVAL); 273 } 274 275 srq = kmalloc(sizeof(*srq), GFP_KERNEL); 276 if (!srq) 277 return ERR_PTR(-ENOMEM); 278 279 mutex_init(&srq->mutex); 280 spin_lock_init(&srq->lock); 281 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1); 282 srq->msrq.max_gs = init_attr->attr.max_sge; 283 284 desc_size = sizeof(struct mlx5_wqe_srq_next_seg) + 285 srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg); 286 desc_size = roundup_pow_of_two(desc_size); 287 desc_size = max_t(int, 32, desc_size); 288 srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) / 289 sizeof(struct mlx5_wqe_data_seg); 290 srq->msrq.wqe_shift = ilog2(desc_size); 291 buf_size = srq->msrq.max * desc_size; 292 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", 293 desc_size, init_attr->attr.max_wr, srq->msrq.max, srq->msrq.max_gs, 294 srq->msrq.max_avail_gather); 295 296 if (pd->uobject) 297 err = create_srq_user(pd, srq, &in, udata, buf_size, &inlen, init_attr->srq_type); 298 else 299 err = create_srq_kernel(dev, srq, &in, buf_size, &inlen, init_attr->srq_type); 300 301 if (err) { 302 mlx5_ib_warn(dev, "create srq %s failed, err %d\n", 303 pd->uobject ? "user" : "kernel", err); 304 goto err_srq; 305 } 306 307 is_xrc = (init_attr->srq_type == IB_SRQT_XRC); 308 in->ctx.state_log_sz = ilog2(srq->msrq.max); 309 flgs = ((srq->msrq.wqe_shift - 4) | (is_xrc << 5) | (srq->wq_sig << 7)) << 24; 310 xrcdn = 0; 311 if (is_xrc) { 312 xrcdn = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn; 313 in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(init_attr->ext.xrc.cq)->mcq.cqn); 314 } else if (init_attr->srq_type == IB_SRQT_BASIC) { 315 xrcdn = to_mxrcd(dev->devr.x0)->xrcdn; 316 in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(dev->devr.c0)->mcq.cqn); 317 } 318 319 in->ctx.flags_xrcd = cpu_to_be32((flgs & 0xFF000000) | (xrcdn & 0xFFFFFF)); 320 321 in->ctx.pd = cpu_to_be32(to_mpd(pd)->pdn); 322 in->ctx.db_record = cpu_to_be64(srq->db.dma); 323 err = mlx5_core_create_srq(dev->mdev, &srq->msrq, in, inlen, is_xrc); 324 kvfree(in); 325 if (err) { 326 mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err); 327 goto err_usr_kern_srq; 328 } 329 330 mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn); 331 332 srq->msrq.event = mlx5_ib_srq_event; 333 srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn; 334 335 if (pd->uobject) 336 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) { 337 mlx5_ib_dbg(dev, "copy to user failed\n"); 338 err = -EFAULT; 339 goto err_core; 340 } 341 342 init_attr->attr.max_wr = srq->msrq.max - 1; 343 344 return &srq->ibsrq; 345 346 err_core: 347 mlx5_core_destroy_srq(dev->mdev, &srq->msrq); 348 349 err_usr_kern_srq: 350 if (pd->uobject) 351 destroy_srq_user(pd, srq); 352 else 353 destroy_srq_kernel(dev, srq); 354 355 err_srq: 356 kfree(srq); 357 358 return ERR_PTR(err); 359 } 360 361 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 362 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata) 363 { 364 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 365 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 366 int ret; 367 368 /* We don't support resizing SRQs yet */ 369 if (attr_mask & IB_SRQ_MAX_WR) 370 return -EINVAL; 371 372 if (attr_mask & IB_SRQ_LIMIT) { 373 if (attr->srq_limit >= srq->msrq.max) 374 return -EINVAL; 375 376 mutex_lock(&srq->mutex); 377 ret = mlx5_core_arm_srq(dev->mdev, &srq->msrq, attr->srq_limit, 1); 378 mutex_unlock(&srq->mutex); 379 380 if (ret) 381 return ret; 382 } 383 384 return 0; 385 } 386 387 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr) 388 { 389 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 390 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 391 int ret; 392 struct mlx5_query_srq_mbox_out *out; 393 394 out = kzalloc(sizeof(*out), GFP_KERNEL); 395 if (!out) 396 return -ENOMEM; 397 398 ret = mlx5_core_query_srq(dev->mdev, &srq->msrq, out); 399 if (ret) 400 goto out_box; 401 402 srq_attr->srq_limit = be16_to_cpu(out->ctx.lwm); 403 srq_attr->max_wr = srq->msrq.max - 1; 404 srq_attr->max_sge = srq->msrq.max_gs; 405 406 out_box: 407 kfree(out); 408 return ret; 409 } 410 411 int mlx5_ib_destroy_srq(struct ib_srq *srq) 412 { 413 struct mlx5_ib_dev *dev = to_mdev(srq->device); 414 struct mlx5_ib_srq *msrq = to_msrq(srq); 415 416 mlx5_core_destroy_srq(dev->mdev, &msrq->msrq); 417 418 if (srq->uobject) { 419 mlx5_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db); 420 ib_umem_release(msrq->umem); 421 } else { 422 destroy_srq_kernel(dev, msrq); 423 } 424 425 kfree(srq); 426 return 0; 427 } 428 429 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index) 430 { 431 struct mlx5_wqe_srq_next_seg *next; 432 433 /* always called with interrupts disabled. */ 434 spin_lock(&srq->lock); 435 436 next = get_wqe(srq, srq->tail); 437 next->next_wqe_index = cpu_to_be16(wqe_index); 438 srq->tail = wqe_index; 439 440 spin_unlock(&srq->lock); 441 } 442 443 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 444 struct ib_recv_wr **bad_wr) 445 { 446 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 447 struct mlx5_wqe_srq_next_seg *next; 448 struct mlx5_wqe_data_seg *scat; 449 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 450 struct mlx5_core_dev *mdev = dev->mdev; 451 unsigned long flags; 452 int err = 0; 453 int nreq; 454 int i; 455 456 spin_lock_irqsave(&srq->lock, flags); 457 458 if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) { 459 err = -EIO; 460 *bad_wr = wr; 461 goto out; 462 } 463 464 for (nreq = 0; wr; nreq++, wr = wr->next) { 465 if (unlikely(wr->num_sge > srq->msrq.max_gs)) { 466 err = -EINVAL; 467 *bad_wr = wr; 468 break; 469 } 470 471 if (unlikely(srq->head == srq->tail)) { 472 err = -ENOMEM; 473 *bad_wr = wr; 474 break; 475 } 476 477 srq->wrid[srq->head] = wr->wr_id; 478 479 next = get_wqe(srq, srq->head); 480 srq->head = be16_to_cpu(next->next_wqe_index); 481 scat = (struct mlx5_wqe_data_seg *)(next + 1); 482 483 for (i = 0; i < wr->num_sge; i++) { 484 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length); 485 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey); 486 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr); 487 } 488 489 if (i < srq->msrq.max_avail_gather) { 490 scat[i].byte_count = 0; 491 scat[i].lkey = cpu_to_be32(MLX5_INVALID_LKEY); 492 scat[i].addr = 0; 493 } 494 } 495 496 if (likely(nreq)) { 497 srq->wqe_ctr += nreq; 498 499 /* Make sure that descriptors are written before 500 * doorbell record. 501 */ 502 wmb(); 503 504 *srq->db.db = cpu_to_be32(srq->wqe_ctr); 505 } 506 out: 507 spin_unlock_irqrestore(&srq->lock, flags); 508 509 return err; 510 } 511