drivers/infiniband/hw/mlx5/srq.c

   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_BASIC) {
 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, 0, &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 = kvcalloc(ncont, sizeof(*in->pas), GFP_KERNEL);
 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_BASIC)
 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
 169         err = mlx5_db_alloc(dev->mdev, &srq->db);
 170         if (err) {
 171                 mlx5_ib_warn(dev, "alloc dbell rec failed\n");
 172                 return err;
 173         }
 174
 175         if (mlx5_buf_alloc(dev->mdev, buf_size, &srq->buf)) {
 176                 mlx5_ib_dbg(dev, "buf alloc failed\n");
 177                 err = -ENOMEM;
 178                 goto err_db;
 179         }
 180
 181         srq->head    = 0;
 182         srq->tail    = srq->msrq.max - 1;
 183         srq->wqe_ctr = 0;
 184
 185         for (i = 0; i < srq->msrq.max; i++) {
 186                 next = get_wqe(srq, i);
 187                 next->next_wqe_index =
 188                         cpu_to_be16((i + 1) & (srq->msrq.max - 1));
 189         }
 190
 191         mlx5_ib_dbg(dev, "srq->buf.page_shift = %d\n", srq->buf.page_shift);
 192         in->pas = kvcalloc(srq->buf.npages, sizeof(*in->pas), GFP_KERNEL);
 193         if (!in->pas) {
 194                 err = -ENOMEM;
 195                 goto err_buf;
 196         }
 197         mlx5_fill_page_array(&srq->buf, in->pas);
 198
 199         srq->wrid = kvmalloc_array(srq->msrq.max, sizeof(u64), GFP_KERNEL);
 200         if (!srq->wrid) {
 201                 err = -ENOMEM;
 202                 goto err_in;
 203         }
 204         srq->wq_sig = !!srq_signature;
 205
 206         in->log_page_size = srq->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT;
 207         if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
 208             in->type != IB_SRQT_BASIC)
 209                 in->user_index = MLX5_IB_DEFAULT_UIDX;
 210
 211         return 0;
 212
 213 err_in:
 214         kvfree(in->pas);
 215
 216 err_buf:
 217         mlx5_buf_free(dev->mdev, &srq->buf);
 218
 219 err_db:
 220         mlx5_db_free(dev->mdev, &srq->db);
 221         return err;
 222 }
 223
 224 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq)
 225 {
 226         mlx5_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
 227         ib_umem_release(srq->umem);
 228 }
 229
 230
 231 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq)
 232 {
 233         kvfree(srq->wrid);
 234         mlx5_buf_free(dev->mdev, &srq->buf);
 235         mlx5_db_free(dev->mdev, &srq->db);
 236 }
 237
 238 struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd,
 239                                   struct ib_srq_init_attr *init_attr,
 240                                   struct ib_udata *udata)
 241 {
 242         struct mlx5_ib_dev *dev = to_mdev(pd->device);
 243         struct mlx5_ib_srq *srq;
 244         size_t desc_size;
 245         size_t buf_size;
 246         int err;
 247         struct mlx5_srq_attr in = {0};
 248         __u32 max_srq_wqes = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
 249
 250         /* Sanity check SRQ size before proceeding */
 251         if (init_attr->attr.max_wr >= max_srq_wqes) {
 252                 mlx5_ib_dbg(dev, "max_wr %d, cap %d\n",
 253                             init_attr->attr.max_wr,
 254                             max_srq_wqes);
 255                 return ERR_PTR(-EINVAL);
 256         }
 257
 258         srq = kmalloc(sizeof(*srq), GFP_KERNEL);
 259         if (!srq)
 260                 return ERR_PTR(-ENOMEM);
 261
 262         mutex_init(&srq->mutex);
 263         spin_lock_init(&srq->lock);
 264         srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
 265         srq->msrq.max_gs = init_attr->attr.max_sge;
 266
 267         desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
 268                     srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
 269         if (desc_size == 0 || srq->msrq.max_gs > desc_size)
 270                 return ERR_PTR(-EINVAL);
 271         desc_size = roundup_pow_of_two(desc_size);
 272         desc_size = max_t(size_t, 32, desc_size);
 273         if (desc_size < sizeof(struct mlx5_wqe_srq_next_seg))
 274                 return ERR_PTR(-EINVAL);
 275         srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) /
 276                 sizeof(struct mlx5_wqe_data_seg);
 277         srq->msrq.wqe_shift = ilog2(desc_size);
 278         buf_size = srq->msrq.max * desc_size;
 279         if (buf_size < desc_size)
 280                 return ERR_PTR(-EINVAL);
 281         in.type = init_attr->srq_type;
 282
 283         if (pd->uobject)
 284                 err = create_srq_user(pd, srq, &in, udata, buf_size);
 285         else
 286                 err = create_srq_kernel(dev, srq, &in, buf_size);
 287
 288         if (err) {
 289                 mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
 290                              pd->uobject ? "user" : "kernel", err);
 291                 goto err_srq;
 292         }
 293
 294         in.log_size = ilog2(srq->msrq.max);
 295         in.wqe_shift = srq->msrq.wqe_shift - 4;
 296         if (srq->wq_sig)
 297                 in.flags |= MLX5_SRQ_FLAG_WQ_SIG;
 298
 299         if (init_attr->srq_type == IB_SRQT_XRC)
 300                 in.xrcd = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
 301         else
 302                 in.xrcd = to_mxrcd(dev->devr.x0)->xrcdn;
 303
 304         if (init_attr->srq_type == IB_SRQT_TM) {
 305                 in.tm_log_list_size =
 306                         ilog2(init_attr->ext.tag_matching.max_num_tags) + 1;
 307                 if (in.tm_log_list_size >
 308                     MLX5_CAP_GEN(dev->mdev, log_tag_matching_list_sz)) {
 309                         mlx5_ib_dbg(dev, "TM SRQ max_num_tags exceeding limit\n");
 310                         err = -EINVAL;
 311                         goto err_usr_kern_srq;
 312                 }
 313                 in.flags |= MLX5_SRQ_FLAG_RNDV;
 314         }
 315
 316         if (ib_srq_has_cq(init_attr->srq_type))
 317                 in.cqn = to_mcq(init_attr->ext.cq)->mcq.cqn;
 318         else
 319                 in.cqn = to_mcq(dev->devr.c0)->mcq.cqn;
 320
 321         in.pd = to_mpd(pd)->pdn;
 322         in.db_record = srq->db.dma;
 323         err = mlx5_core_create_srq(dev->mdev, &srq->msrq, &in);
 324         kvfree(in.pas);
 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_srq_attr *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 = out->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 }