drivers/infiniband/sw/rdmavt/srq.c

   1 /*
   2  * Copyright(c) 2016 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47
  48 #include <linux/err.h>
  49 #include <linux/slab.h>
  50 #include <linux/vmalloc.h>
  51 #include <rdma/uverbs_ioctl.h>
  52
  53 #include "srq.h"
  54 #include "vt.h"
  55
  56 /**
  57  * rvt_driver_srq_init - init srq resources on a per driver basis
  58  * @rdi: rvt dev structure
  59  *
  60  * Do any initialization needed when a driver registers with rdmavt.
  61  */
  62 void rvt_driver_srq_init(struct rvt_dev_info *rdi)
  63 {
  64         spin_lock_init(&rdi->n_srqs_lock);
  65         rdi->n_srqs_allocated = 0;
  66 }
  67
  68 /**
  69  * rvt_create_srq - create a shared receive queue
  70  * @ibpd: the protection domain of the SRQ to create
  71  * @srq_init_attr: the attributes of the SRQ
  72  * @udata: data from libibverbs when creating a user SRQ
  73  *
  74  * Return: Allocated srq object
  75  */
  76 struct ib_srq *rvt_create_srq(struct ib_pd *ibpd,
  77                               struct ib_srq_init_attr *srq_init_attr,
  78                               struct ib_udata *udata)
  79 {
  80         struct rvt_dev_info *dev = ib_to_rvt(ibpd->device);
  81         struct rvt_ucontext *ucontext = rdma_udata_to_drv_context(
  82                 udata, struct rvt_ucontext, ibucontext);
  83         struct rvt_srq *srq;
  84         u32 sz;
  85         struct ib_srq *ret;
  86
  87         if (srq_init_attr->srq_type != IB_SRQT_BASIC)
  88                 return ERR_PTR(-EOPNOTSUPP);
  89
  90         if (srq_init_attr->attr.max_sge == 0 ||
  91             srq_init_attr->attr.max_sge > dev->dparms.props.max_srq_sge ||
  92             srq_init_attr->attr.max_wr == 0 ||
  93             srq_init_attr->attr.max_wr > dev->dparms.props.max_srq_wr)
  94                 return ERR_PTR(-EINVAL);
  95
  96         srq = kzalloc_node(sizeof(*srq), GFP_KERNEL, dev->dparms.node);
  97         if (!srq)
  98                 return ERR_PTR(-ENOMEM);
  99
 100         /*
 101          * Need to use vmalloc() if we want to support large #s of entries.
 102          */
 103         srq->rq.size = srq_init_attr->attr.max_wr + 1;
 104         srq->rq.max_sge = srq_init_attr->attr.max_sge;
 105         sz = sizeof(struct ib_sge) * srq->rq.max_sge +
 106                 sizeof(struct rvt_rwqe);
 107         srq->rq.wq = udata ?
 108                 vmalloc_user(sizeof(struct rvt_rwq) + srq->rq.size * sz) :
 109                 vzalloc_node(sizeof(struct rvt_rwq) + srq->rq.size * sz,
 110                              dev->dparms.node);
 111         if (!srq->rq.wq) {
 112                 ret = ERR_PTR(-ENOMEM);
 113                 goto bail_srq;
 114         }
 115
 116         /*
 117          * Return the address of the RWQ as the offset to mmap.
 118          * See rvt_mmap() for details.
 119          */
 120         if (udata && udata->outlen >= sizeof(__u64)) {
 121                 int err;
 122                 u32 s = sizeof(struct rvt_rwq) + srq->rq.size * sz;
 123
 124                 srq->ip =
 125                     rvt_create_mmap_info(dev, s, &ucontext->ibucontext,
 126                                          srq->rq.wq);
 127                 if (!srq->ip) {
 128                         ret = ERR_PTR(-ENOMEM);
 129                         goto bail_wq;
 130                 }
 131
 132                 err = ib_copy_to_udata(udata, &srq->ip->offset,
 133                                        sizeof(srq->ip->offset));
 134                 if (err) {
 135                         ret = ERR_PTR(err);
 136                         goto bail_ip;
 137                 }
 138         }
 139
 140         /*
 141          * ib_create_srq() will initialize srq->ibsrq.
 142          */
 143         spin_lock_init(&srq->rq.lock);
 144         srq->limit = srq_init_attr->attr.srq_limit;
 145
 146         spin_lock(&dev->n_srqs_lock);
 147         if (dev->n_srqs_allocated == dev->dparms.props.max_srq) {
 148                 spin_unlock(&dev->n_srqs_lock);
 149                 ret = ERR_PTR(-ENOMEM);
 150                 goto bail_ip;
 151         }
 152
 153         dev->n_srqs_allocated++;
 154         spin_unlock(&dev->n_srqs_lock);
 155
 156         if (srq->ip) {
 157                 spin_lock_irq(&dev->pending_lock);
 158                 list_add(&srq->ip->pending_mmaps, &dev->pending_mmaps);
 159                 spin_unlock_irq(&dev->pending_lock);
 160         }
 161
 162         return &srq->ibsrq;
 163
 164 bail_ip:
 165         kfree(srq->ip);
 166 bail_wq:
 167         vfree(srq->rq.wq);
 168 bail_srq:
 169         kfree(srq);
 170         return ret;
 171 }
 172
 173 /**
 174  * rvt_modify_srq - modify a shared receive queue
 175  * @ibsrq: the SRQ to modify
 176  * @attr: the new attributes of the SRQ
 177  * @attr_mask: indicates which attributes to modify
 178  * @udata: user data for libibverbs.so
 179  *
 180  * Return: 0 on success
 181  */
 182 int rvt_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
 183                    enum ib_srq_attr_mask attr_mask,
 184                    struct ib_udata *udata)
 185 {
 186         struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
 187         struct rvt_dev_info *dev = ib_to_rvt(ibsrq->device);
 188         struct rvt_rwq *wq;
 189         int ret = 0;
 190
 191         if (attr_mask & IB_SRQ_MAX_WR) {
 192                 struct rvt_rwq *owq;
 193                 struct rvt_rwqe *p;
 194                 u32 sz, size, n, head, tail;
 195
 196                 /* Check that the requested sizes are below the limits. */
 197                 if ((attr->max_wr > dev->dparms.props.max_srq_wr) ||
 198                     ((attr_mask & IB_SRQ_LIMIT) ?
 199                      attr->srq_limit : srq->limit) > attr->max_wr)
 200                         return -EINVAL;
 201
 202                 sz = sizeof(struct rvt_rwqe) +
 203                         srq->rq.max_sge * sizeof(struct ib_sge);
 204                 size = attr->max_wr + 1;
 205                 wq = udata ?
 206                         vmalloc_user(sizeof(struct rvt_rwq) + size * sz) :
 207                         vzalloc_node(sizeof(struct rvt_rwq) + size * sz,
 208                                      dev->dparms.node);
 209                 if (!wq)
 210                         return -ENOMEM;
 211
 212                 /* Check that we can write the offset to mmap. */
 213                 if (udata && udata->inlen >= sizeof(__u64)) {
 214                         __u64 offset_addr;
 215                         __u64 offset = 0;
 216
 217                         ret = ib_copy_from_udata(&offset_addr, udata,
 218                                                  sizeof(offset_addr));
 219                         if (ret)
 220                                 goto bail_free;
 221                         udata->outbuf = (void __user *)
 222                                         (unsigned long)offset_addr;
 223                         ret = ib_copy_to_udata(udata, &offset,
 224                                                sizeof(offset));
 225                         if (ret)
 226                                 goto bail_free;
 227                 }
 228
 229                 spin_lock_irq(&srq->rq.lock);
 230                 /*
 231                  * validate head and tail pointer values and compute
 232                  * the number of remaining WQEs.
 233                  */
 234                 owq = srq->rq.wq;
 235                 head = owq->head;
 236                 tail = owq->tail;
 237                 if (head >= srq->rq.size || tail >= srq->rq.size) {
 238                         ret = -EINVAL;
 239                         goto bail_unlock;
 240                 }
 241                 n = head;
 242                 if (n < tail)
 243                         n += srq->rq.size - tail;
 244                 else
 245                         n -= tail;
 246                 if (size <= n) {
 247                         ret = -EINVAL;
 248                         goto bail_unlock;
 249                 }
 250                 n = 0;
 251                 p = wq->wq;
 252                 while (tail != head) {
 253                         struct rvt_rwqe *wqe;
 254                         int i;
 255
 256                         wqe = rvt_get_rwqe_ptr(&srq->rq, tail);
 257                         p->wr_id = wqe->wr_id;
 258                         p->num_sge = wqe->num_sge;
 259                         for (i = 0; i < wqe->num_sge; i++)
 260                                 p->sg_list[i] = wqe->sg_list[i];
 261                         n++;
 262                         p = (struct rvt_rwqe *)((char *)p + sz);
 263                         if (++tail >= srq->rq.size)
 264                                 tail = 0;
 265                 }
 266                 srq->rq.wq = wq;
 267                 srq->rq.size = size;
 268                 wq->head = n;
 269                 wq->tail = 0;
 270                 if (attr_mask & IB_SRQ_LIMIT)
 271                         srq->limit = attr->srq_limit;
 272                 spin_unlock_irq(&srq->rq.lock);
 273
 274                 vfree(owq);
 275
 276                 if (srq->ip) {
 277                         struct rvt_mmap_info *ip = srq->ip;
 278                         struct rvt_dev_info *dev = ib_to_rvt(srq->ibsrq.device);
 279                         u32 s = sizeof(struct rvt_rwq) + size * sz;
 280
 281                         rvt_update_mmap_info(dev, ip, s, wq);
 282
 283                         /*
 284                          * Return the offset to mmap.
 285                          * See rvt_mmap() for details.
 286                          */
 287                         if (udata && udata->inlen >= sizeof(__u64)) {
 288                                 ret = ib_copy_to_udata(udata, &ip->offset,
 289                                                        sizeof(ip->offset));
 290                                 if (ret)
 291                                         return ret;
 292                         }
 293
 294                         /*
 295                          * Put user mapping info onto the pending list
 296                          * unless it already is on the list.
 297                          */
 298                         spin_lock_irq(&dev->pending_lock);
 299                         if (list_empty(&ip->pending_mmaps))
 300                                 list_add(&ip->pending_mmaps,
 301                                          &dev->pending_mmaps);
 302                         spin_unlock_irq(&dev->pending_lock);
 303                 }
 304         } else if (attr_mask & IB_SRQ_LIMIT) {
 305                 spin_lock_irq(&srq->rq.lock);
 306                 if (attr->srq_limit >= srq->rq.size)
 307                         ret = -EINVAL;
 308                 else
 309                         srq->limit = attr->srq_limit;
 310                 spin_unlock_irq(&srq->rq.lock);
 311         }
 312         return ret;
 313
 314 bail_unlock:
 315         spin_unlock_irq(&srq->rq.lock);
 316 bail_free:
 317         vfree(wq);
 318         return ret;
 319 }
 320
 321 /** rvt_query_srq - query srq data
 322  * @ibsrq: srq to query
 323  * @attr: return info in attr
 324  *
 325  * Return: always 0
 326  */
 327 int rvt_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
 328 {
 329         struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
 330
 331         attr->max_wr = srq->rq.size - 1;
 332         attr->max_sge = srq->rq.max_sge;
 333         attr->srq_limit = srq->limit;
 334         return 0;
 335 }
 336
 337 /**
 338  * rvt_destroy_srq - destory an srq
 339  * @ibsrq: srq object to destroy
 340  *
 341  * Return always 0
 342  */
 343 int rvt_destroy_srq(struct ib_srq *ibsrq)
 344 {
 345         struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
 346         struct rvt_dev_info *dev = ib_to_rvt(ibsrq->device);
 347
 348         spin_lock(&dev->n_srqs_lock);
 349         dev->n_srqs_allocated--;
 350         spin_unlock(&dev->n_srqs_lock);
 351         if (srq->ip)
 352                 kref_put(&srq->ip->ref, rvt_release_mmap_info);
 353         else
 354                 vfree(srq->rq.wq);
 355         kfree(srq);
 356
 357         return 0;
 358 }