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
  52 #include "srq.h"
  53 #include "vt.h"
  54
  55 /**
  56  * rvt_driver_srq_init - init srq resources on a per driver basis
  57  * @rdi: rvt dev structure
  58  *
  59  * Do any initialization needed when a driver registers with rdmavt.
  60  */
  61 void rvt_driver_srq_init(struct rvt_dev_info *rdi)
  62 {
  63         spin_lock_init(&rdi->n_srqs_lock);
  64         rdi->n_srqs_allocated = 0;
  65 }
  66
  67 /**
  68  * rvt_create_srq - create a shared receive queue
  69  * @ibpd: the protection domain of the SRQ to create
  70  * @srq_init_attr: the attributes of the SRQ
  71  * @udata: data from libibverbs when creating a user SRQ
  72  *
  73  * Return: Allocated srq object
  74  */
  75 struct ib_srq *rvt_create_srq(struct ib_pd *ibpd,
  76                               struct ib_srq_init_attr *srq_init_attr,
  77                               struct ib_udata *udata)
  78 {
  79         struct rvt_dev_info *dev = ib_to_rvt(ibpd->device);
  80         struct rvt_srq *srq;
  81         u32 sz;
  82         struct ib_srq *ret;
  83
  84         if (srq_init_attr->srq_type != IB_SRQT_BASIC)
  85                 return ERR_PTR(-EOPNOTSUPP);
  86
  87         if (srq_init_attr->attr.max_sge == 0 ||
  88             srq_init_attr->attr.max_sge > dev->dparms.props.max_srq_sge ||
  89             srq_init_attr->attr.max_wr == 0 ||
  90             srq_init_attr->attr.max_wr > dev->dparms.props.max_srq_wr)
  91                 return ERR_PTR(-EINVAL);
  92
  93         srq = kzalloc_node(sizeof(*srq), GFP_KERNEL, dev->dparms.node);
  94         if (!srq)
  95                 return ERR_PTR(-ENOMEM);
  96
  97         /*
  98          * Need to use vmalloc() if we want to support large #s of entries.
  99          */
 100         srq->rq.size = srq_init_attr->attr.max_wr + 1;
 101         srq->rq.max_sge = srq_init_attr->attr.max_sge;
 102         sz = sizeof(struct ib_sge) * srq->rq.max_sge +
 103                 sizeof(struct rvt_rwqe);
 104         srq->rq.wq = udata ?
 105                 vmalloc_user(sizeof(struct rvt_rwq) + srq->rq.size * sz) :
 106                 vzalloc_node(sizeof(struct rvt_rwq) + srq->rq.size * sz,
 107                              dev->dparms.node);
 108         if (!srq->rq.wq) {
 109                 ret = ERR_PTR(-ENOMEM);
 110                 goto bail_srq;
 111         }
 112
 113         /*
 114          * Return the address of the RWQ as the offset to mmap.
 115          * See rvt_mmap() for details.
 116          */
 117         if (udata && udata->outlen >= sizeof(__u64)) {
 118                 int err;
 119                 u32 s = sizeof(struct rvt_rwq) + srq->rq.size * sz;
 120
 121                 srq->ip =
 122                     rvt_create_mmap_info(dev, s, ibpd->uobject->context,
 123                                          srq->rq.wq);
 124                 if (!srq->ip) {
 125                         ret = ERR_PTR(-ENOMEM);
 126                         goto bail_wq;
 127                 }
 128
 129                 err = ib_copy_to_udata(udata, &srq->ip->offset,
 130                                        sizeof(srq->ip->offset));
 131                 if (err) {
 132                         ret = ERR_PTR(err);
 133                         goto bail_ip;
 134                 }
 135         }
 136
 137         /*
 138          * ib_create_srq() will initialize srq->ibsrq.
 139          */
 140         spin_lock_init(&srq->rq.lock);
 141         srq->limit = srq_init_attr->attr.srq_limit;
 142
 143         spin_lock(&dev->n_srqs_lock);
 144         if (dev->n_srqs_allocated == dev->dparms.props.max_srq) {
 145                 spin_unlock(&dev->n_srqs_lock);
 146                 ret = ERR_PTR(-ENOMEM);
 147                 goto bail_ip;
 148         }
 149
 150         dev->n_srqs_allocated++;
 151         spin_unlock(&dev->n_srqs_lock);
 152
 153         if (srq->ip) {
 154                 spin_lock_irq(&dev->pending_lock);
 155                 list_add(&srq->ip->pending_mmaps, &dev->pending_mmaps);
 156                 spin_unlock_irq(&dev->pending_lock);
 157         }
 158
 159         return &srq->ibsrq;
 160
 161 bail_ip:
 162         kfree(srq->ip);
 163 bail_wq:
 164         vfree(srq->rq.wq);
 165 bail_srq:
 166         kfree(srq);
 167         return ret;
 168 }
 169
 170 /**
 171  * rvt_modify_srq - modify a shared receive queue
 172  * @ibsrq: the SRQ to modify
 173  * @attr: the new attributes of the SRQ
 174  * @attr_mask: indicates which attributes to modify
 175  * @udata: user data for libibverbs.so
 176  *
 177  * Return: 0 on success
 178  */
 179 int rvt_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
 180                    enum ib_srq_attr_mask attr_mask,
 181                    struct ib_udata *udata)
 182 {
 183         struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
 184         struct rvt_dev_info *dev = ib_to_rvt(ibsrq->device);
 185         struct rvt_rwq *wq;
 186         int ret = 0;
 187
 188         if (attr_mask & IB_SRQ_MAX_WR) {
 189                 struct rvt_rwq *owq;
 190                 struct rvt_rwqe *p;
 191                 u32 sz, size, n, head, tail;
 192
 193                 /* Check that the requested sizes are below the limits. */
 194                 if ((attr->max_wr > dev->dparms.props.max_srq_wr) ||
 195                     ((attr_mask & IB_SRQ_LIMIT) ?
 196                      attr->srq_limit : srq->limit) > attr->max_wr)
 197                         return -EINVAL;
 198
 199                 sz = sizeof(struct rvt_rwqe) +
 200                         srq->rq.max_sge * sizeof(struct ib_sge);
 201                 size = attr->max_wr + 1;
 202                 wq = udata ?
 203                         vmalloc_user(sizeof(struct rvt_rwq) + size * sz) :
 204                         vzalloc_node(sizeof(struct rvt_rwq) + size * sz,
 205                                      dev->dparms.node);
 206                 if (!wq)
 207                         return -ENOMEM;
 208
 209                 /* Check that we can write the offset to mmap. */
 210                 if (udata && udata->inlen >= sizeof(__u64)) {
 211                         __u64 offset_addr;
 212                         __u64 offset = 0;
 213
 214                         ret = ib_copy_from_udata(&offset_addr, udata,
 215                                                  sizeof(offset_addr));
 216                         if (ret)
 217                                 goto bail_free;
 218                         udata->outbuf = (void __user *)
 219                                         (unsigned long)offset_addr;
 220                         ret = ib_copy_to_udata(udata, &offset,
 221                                                sizeof(offset));
 222                         if (ret)
 223                                 goto bail_free;
 224                 }
 225
 226                 spin_lock_irq(&srq->rq.lock);
 227                 /*
 228                  * validate head and tail pointer values and compute
 229                  * the number of remaining WQEs.
 230                  */
 231                 owq = srq->rq.wq;
 232                 head = owq->head;
 233                 tail = owq->tail;
 234                 if (head >= srq->rq.size || tail >= srq->rq.size) {
 235                         ret = -EINVAL;
 236                         goto bail_unlock;
 237                 }
 238                 n = head;
 239                 if (n < tail)
 240                         n += srq->rq.size - tail;
 241                 else
 242                         n -= tail;
 243                 if (size <= n) {
 244                         ret = -EINVAL;
 245                         goto bail_unlock;
 246                 }
 247                 n = 0;
 248                 p = wq->wq;
 249                 while (tail != head) {
 250                         struct rvt_rwqe *wqe;
 251                         int i;
 252
 253                         wqe = rvt_get_rwqe_ptr(&srq->rq, tail);
 254                         p->wr_id = wqe->wr_id;
 255                         p->num_sge = wqe->num_sge;
 256                         for (i = 0; i < wqe->num_sge; i++)
 257                                 p->sg_list[i] = wqe->sg_list[i];
 258                         n++;
 259                         p = (struct rvt_rwqe *)((char *)p + sz);
 260                         if (++tail >= srq->rq.size)
 261                                 tail = 0;
 262                 }
 263                 srq->rq.wq = wq;
 264                 srq->rq.size = size;
 265                 wq->head = n;
 266                 wq->tail = 0;
 267                 if (attr_mask & IB_SRQ_LIMIT)
 268                         srq->limit = attr->srq_limit;
 269                 spin_unlock_irq(&srq->rq.lock);
 270
 271                 vfree(owq);
 272
 273                 if (srq->ip) {
 274                         struct rvt_mmap_info *ip = srq->ip;
 275                         struct rvt_dev_info *dev = ib_to_rvt(srq->ibsrq.device);
 276                         u32 s = sizeof(struct rvt_rwq) + size * sz;
 277
 278                         rvt_update_mmap_info(dev, ip, s, wq);
 279
 280                         /*
 281                          * Return the offset to mmap.
 282                          * See rvt_mmap() for details.
 283                          */
 284                         if (udata && udata->inlen >= sizeof(__u64)) {
 285                                 ret = ib_copy_to_udata(udata, &ip->offset,
 286                                                        sizeof(ip->offset));
 287                                 if (ret)
 288                                         return ret;
 289                         }
 290
 291                         /*
 292                          * Put user mapping info onto the pending list
 293                          * unless it already is on the list.
 294                          */
 295                         spin_lock_irq(&dev->pending_lock);
 296                         if (list_empty(&ip->pending_mmaps))
 297                                 list_add(&ip->pending_mmaps,
 298                                          &dev->pending_mmaps);
 299                         spin_unlock_irq(&dev->pending_lock);
 300                 }
 301         } else if (attr_mask & IB_SRQ_LIMIT) {
 302                 spin_lock_irq(&srq->rq.lock);
 303                 if (attr->srq_limit >= srq->rq.size)
 304                         ret = -EINVAL;
 305                 else
 306                         srq->limit = attr->srq_limit;
 307                 spin_unlock_irq(&srq->rq.lock);
 308         }
 309         return ret;
 310
 311 bail_unlock:
 312         spin_unlock_irq(&srq->rq.lock);
 313 bail_free:
 314         vfree(wq);
 315         return ret;
 316 }
 317
 318 /** rvt_query_srq - query srq data
 319  * @ibsrq: srq to query
 320  * @attr: return info in attr
 321  *
 322  * Return: always 0
 323  */
 324 int rvt_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
 325 {
 326         struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
 327
 328         attr->max_wr = srq->rq.size - 1;
 329         attr->max_sge = srq->rq.max_sge;
 330         attr->srq_limit = srq->limit;
 331         return 0;
 332 }
 333
 334 /**
 335  * rvt_destroy_srq - destory an srq
 336  * @ibsrq: srq object to destroy
 337  *
 338  * Return always 0
 339  */
 340 int rvt_destroy_srq(struct ib_srq *ibsrq)
 341 {
 342         struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
 343         struct rvt_dev_info *dev = ib_to_rvt(ibsrq->device);
 344
 345         spin_lock(&dev->n_srqs_lock);
 346         dev->n_srqs_allocated--;
 347         spin_unlock(&dev->n_srqs_lock);
 348         if (srq->ip)
 349                 kref_put(&srq->ip->ref, rvt_release_mmap_info);
 350         else
 351                 vfree(srq->rq.wq);
 352         kfree(srq);
 353
 354         return 0;
 355 }