1 /* 2 * Copyright (c) 2013, Mellanox Technologies inc. 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 39 #include "mlx5_ib.h" 40 #include "user.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_create_srq_mbox_in **in, 77 struct ib_udata *udata, int buf_size, int *inlen) 78 { 79 struct mlx5_ib_dev *dev = to_mdev(pd->device); 80 struct mlx5_ib_create_srq ucmd; 81 int err; 82 int npages; 83 int page_shift; 84 int ncont; 85 u32 offset; 86 87 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) { 88 mlx5_ib_dbg(dev, "failed copy udata\n"); 89 return -EFAULT; 90 } 91 srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE); 92 93 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, buf_size, 94 0, 0); 95 if (IS_ERR(srq->umem)) { 96 mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size); 97 err = PTR_ERR(srq->umem); 98 return err; 99 } 100 101 mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, &npages, 102 &page_shift, &ncont, NULL); 103 err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift, 104 &offset); 105 if (err) { 106 mlx5_ib_warn(dev, "bad offset\n"); 107 goto err_umem; 108 } 109 110 *inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont; 111 *in = mlx5_vzalloc(*inlen); 112 if (!(*in)) { 113 err = -ENOMEM; 114 goto err_umem; 115 } 116 117 mlx5_ib_populate_pas(dev, srq->umem, page_shift, (*in)->pas, 0); 118 119 err = mlx5_ib_db_map_user(to_mucontext(pd->uobject->context), 120 ucmd.db_addr, &srq->db); 121 if (err) { 122 mlx5_ib_dbg(dev, "map doorbell failed\n"); 123 goto err_in; 124 } 125 126 (*in)->ctx.log_pg_sz = page_shift - PAGE_SHIFT; 127 (*in)->ctx.pgoff_cqn = cpu_to_be32(offset << 26); 128 129 return 0; 130 131 err_in: 132 mlx5_vfree(*in); 133 134 err_umem: 135 ib_umem_release(srq->umem); 136 137 return err; 138 } 139 140 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq, 141 struct mlx5_create_srq_mbox_in **in, int buf_size, 142 int *inlen) 143 { 144 int err; 145 int i; 146 struct mlx5_wqe_srq_next_seg *next; 147 int page_shift; 148 int npages; 149 150 err = mlx5_db_alloc(&dev->mdev, &srq->db); 151 if (err) { 152 mlx5_ib_warn(dev, "alloc dbell rec failed\n"); 153 return err; 154 } 155 156 *srq->db.db = 0; 157 158 if (mlx5_buf_alloc(&dev->mdev, buf_size, PAGE_SIZE * 2, &srq->buf)) { 159 mlx5_ib_dbg(dev, "buf alloc failed\n"); 160 err = -ENOMEM; 161 goto err_db; 162 } 163 page_shift = srq->buf.page_shift; 164 165 srq->head = 0; 166 srq->tail = srq->msrq.max - 1; 167 srq->wqe_ctr = 0; 168 169 for (i = 0; i < srq->msrq.max; i++) { 170 next = get_wqe(srq, i); 171 next->next_wqe_index = 172 cpu_to_be16((i + 1) & (srq->msrq.max - 1)); 173 } 174 175 npages = DIV_ROUND_UP(srq->buf.npages, 1 << (page_shift - PAGE_SHIFT)); 176 mlx5_ib_dbg(dev, "buf_size %d, page_shift %d, npages %d, calc npages %d\n", 177 buf_size, page_shift, srq->buf.npages, npages); 178 *inlen = sizeof(**in) + sizeof(*(*in)->pas) * npages; 179 *in = mlx5_vzalloc(*inlen); 180 if (!*in) { 181 err = -ENOMEM; 182 goto err_buf; 183 } 184 mlx5_fill_page_array(&srq->buf, (*in)->pas); 185 186 srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL); 187 if (!srq->wrid) { 188 mlx5_ib_dbg(dev, "kmalloc failed %lu\n", 189 (unsigned long)(srq->msrq.max * sizeof(u64))); 190 err = -ENOMEM; 191 goto err_in; 192 } 193 srq->wq_sig = !!srq_signature; 194 195 (*in)->ctx.log_pg_sz = page_shift - PAGE_SHIFT; 196 197 return 0; 198 199 err_in: 200 mlx5_vfree(*in); 201 202 err_buf: 203 mlx5_buf_free(&dev->mdev, &srq->buf); 204 205 err_db: 206 mlx5_db_free(&dev->mdev, &srq->db); 207 return err; 208 } 209 210 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq) 211 { 212 mlx5_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db); 213 ib_umem_release(srq->umem); 214 } 215 216 217 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq) 218 { 219 kfree(srq->wrid); 220 mlx5_buf_free(&dev->mdev, &srq->buf); 221 mlx5_db_free(&dev->mdev, &srq->db); 222 } 223 224 struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd, 225 struct ib_srq_init_attr *init_attr, 226 struct ib_udata *udata) 227 { 228 struct mlx5_ib_dev *dev = to_mdev(pd->device); 229 struct mlx5_ib_srq *srq; 230 int desc_size; 231 int buf_size; 232 int err; 233 struct mlx5_create_srq_mbox_in *uninitialized_var(in); 234 int uninitialized_var(inlen); 235 int is_xrc; 236 u32 flgs, xrcdn; 237 238 /* Sanity check SRQ size before proceeding */ 239 if (init_attr->attr.max_wr >= dev->mdev.caps.max_srq_wqes) { 240 mlx5_ib_dbg(dev, "max_wr %d, cap %d\n", 241 init_attr->attr.max_wr, 242 dev->mdev.caps.max_srq_wqes); 243 return ERR_PTR(-EINVAL); 244 } 245 246 srq = kmalloc(sizeof(*srq), GFP_KERNEL); 247 if (!srq) 248 return ERR_PTR(-ENOMEM); 249 250 mutex_init(&srq->mutex); 251 spin_lock_init(&srq->lock); 252 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1); 253 srq->msrq.max_gs = init_attr->attr.max_sge; 254 255 desc_size = sizeof(struct mlx5_wqe_srq_next_seg) + 256 srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg); 257 desc_size = roundup_pow_of_two(desc_size); 258 desc_size = max_t(int, 32, desc_size); 259 srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) / 260 sizeof(struct mlx5_wqe_data_seg); 261 srq->msrq.wqe_shift = ilog2(desc_size); 262 buf_size = srq->msrq.max * desc_size; 263 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", 264 desc_size, init_attr->attr.max_wr, srq->msrq.max, srq->msrq.max_gs, 265 srq->msrq.max_avail_gather); 266 267 if (pd->uobject) 268 err = create_srq_user(pd, srq, &in, udata, buf_size, &inlen); 269 else 270 err = create_srq_kernel(dev, srq, &in, buf_size, &inlen); 271 272 if (err) { 273 mlx5_ib_warn(dev, "create srq %s failed, err %d\n", 274 pd->uobject ? "user" : "kernel", err); 275 goto err_srq; 276 } 277 278 is_xrc = (init_attr->srq_type == IB_SRQT_XRC); 279 in->ctx.state_log_sz = ilog2(srq->msrq.max); 280 flgs = ((srq->msrq.wqe_shift - 4) | (is_xrc << 5) | (srq->wq_sig << 7)) << 24; 281 xrcdn = 0; 282 if (is_xrc) { 283 xrcdn = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn; 284 in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(init_attr->ext.xrc.cq)->mcq.cqn); 285 } else if (init_attr->srq_type == IB_SRQT_BASIC) { 286 xrcdn = to_mxrcd(dev->devr.x0)->xrcdn; 287 in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(dev->devr.c0)->mcq.cqn); 288 } 289 290 in->ctx.flags_xrcd = cpu_to_be32((flgs & 0xFF000000) | (xrcdn & 0xFFFFFF)); 291 292 in->ctx.pd = cpu_to_be32(to_mpd(pd)->pdn); 293 in->ctx.db_record = cpu_to_be64(srq->db.dma); 294 err = mlx5_core_create_srq(&dev->mdev, &srq->msrq, in, inlen); 295 mlx5_vfree(in); 296 if (err) { 297 mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err); 298 goto err_srq; 299 } 300 301 mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn); 302 303 srq->msrq.event = mlx5_ib_srq_event; 304 srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn; 305 306 if (pd->uobject) 307 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) { 308 mlx5_ib_dbg(dev, "copy to user failed\n"); 309 err = -EFAULT; 310 goto err_core; 311 } 312 313 init_attr->attr.max_wr = srq->msrq.max - 1; 314 315 return &srq->ibsrq; 316 317 err_core: 318 mlx5_core_destroy_srq(&dev->mdev, &srq->msrq); 319 if (pd->uobject) 320 destroy_srq_user(pd, srq); 321 else 322 destroy_srq_kernel(dev, srq); 323 324 err_srq: 325 kfree(srq); 326 327 return ERR_PTR(err); 328 } 329 330 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr, 331 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata) 332 { 333 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 334 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 335 int ret; 336 337 /* We don't support resizing SRQs yet */ 338 if (attr_mask & IB_SRQ_MAX_WR) 339 return -EINVAL; 340 341 if (attr_mask & IB_SRQ_LIMIT) { 342 if (attr->srq_limit >= srq->msrq.max) 343 return -EINVAL; 344 345 mutex_lock(&srq->mutex); 346 ret = mlx5_core_arm_srq(&dev->mdev, &srq->msrq, attr->srq_limit, 1); 347 mutex_unlock(&srq->mutex); 348 349 if (ret) 350 return ret; 351 } 352 353 return 0; 354 } 355 356 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr) 357 { 358 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device); 359 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 360 int ret; 361 struct mlx5_query_srq_mbox_out *out; 362 363 out = kzalloc(sizeof(*out), GFP_KERNEL); 364 if (!out) 365 return -ENOMEM; 366 367 ret = mlx5_core_query_srq(&dev->mdev, &srq->msrq, out); 368 if (ret) 369 goto out_box; 370 371 srq_attr->srq_limit = be16_to_cpu(out->ctx.lwm); 372 srq_attr->max_wr = srq->msrq.max - 1; 373 srq_attr->max_sge = srq->msrq.max_gs; 374 375 out_box: 376 kfree(out); 377 return ret; 378 } 379 380 int mlx5_ib_destroy_srq(struct ib_srq *srq) 381 { 382 struct mlx5_ib_dev *dev = to_mdev(srq->device); 383 struct mlx5_ib_srq *msrq = to_msrq(srq); 384 385 mlx5_core_destroy_srq(&dev->mdev, &msrq->msrq); 386 387 if (srq->uobject) { 388 mlx5_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db); 389 ib_umem_release(msrq->umem); 390 } else { 391 kfree(msrq->wrid); 392 mlx5_buf_free(&dev->mdev, &msrq->buf); 393 mlx5_db_free(&dev->mdev, &msrq->db); 394 } 395 396 kfree(srq); 397 return 0; 398 } 399 400 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index) 401 { 402 struct mlx5_wqe_srq_next_seg *next; 403 404 /* always called with interrupts disabled. */ 405 spin_lock(&srq->lock); 406 407 next = get_wqe(srq, srq->tail); 408 next->next_wqe_index = cpu_to_be16(wqe_index); 409 srq->tail = wqe_index; 410 411 spin_unlock(&srq->lock); 412 } 413 414 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, 415 struct ib_recv_wr **bad_wr) 416 { 417 struct mlx5_ib_srq *srq = to_msrq(ibsrq); 418 struct mlx5_wqe_srq_next_seg *next; 419 struct mlx5_wqe_data_seg *scat; 420 unsigned long flags; 421 int err = 0; 422 int nreq; 423 int i; 424 425 spin_lock_irqsave(&srq->lock, flags); 426 427 for (nreq = 0; wr; nreq++, wr = wr->next) { 428 if (unlikely(wr->num_sge > srq->msrq.max_gs)) { 429 err = -EINVAL; 430 *bad_wr = wr; 431 break; 432 } 433 434 if (unlikely(srq->head == srq->tail)) { 435 err = -ENOMEM; 436 *bad_wr = wr; 437 break; 438 } 439 440 srq->wrid[srq->head] = wr->wr_id; 441 442 next = get_wqe(srq, srq->head); 443 srq->head = be16_to_cpu(next->next_wqe_index); 444 scat = (struct mlx5_wqe_data_seg *)(next + 1); 445 446 for (i = 0; i < wr->num_sge; i++) { 447 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length); 448 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey); 449 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr); 450 } 451 452 if (i < srq->msrq.max_avail_gather) { 453 scat[i].byte_count = 0; 454 scat[i].lkey = cpu_to_be32(MLX5_INVALID_LKEY); 455 scat[i].addr = 0; 456 } 457 } 458 459 if (likely(nreq)) { 460 srq->wqe_ctr += nreq; 461 462 /* Make sure that descriptors are written before 463 * doorbell record. 464 */ 465 wmb(); 466 467 *srq->db.db = cpu_to_be32(srq->wqe_ctr); 468 } 469 470 spin_unlock_irqrestore(&srq->lock, flags); 471 472 return err; 473 } 474