drivers/infiniband/hw/ipath/ipath_srq.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/err.h>
  35 #include <linux/vmalloc.h>
  36
  37 #include "ipath_verbs.h"
  38
  39 /**
  40  * ipath_post_srq_receive - post a receive on a shared receive queue
  41  * @ibsrq: the SRQ to post the receive on
  42  * @wr: the list of work requests to post
  43  * @bad_wr: the first WR to cause a problem is put here
  44  *
  45  * This may be called from interrupt context.
  46  */
  47 int ipath_post_srq_receive(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
  48                            struct ib_recv_wr **bad_wr)
  49 {
  50         struct ipath_srq *srq = to_isrq(ibsrq);
  51         struct ipath_rwq *wq;
  52         unsigned long flags;
  53         int ret;
  54
  55         for (; wr; wr = wr->next) {
  56                 struct ipath_rwqe *wqe;
  57                 u32 next;
  58                 int i;
  59
  60                 if ((unsigned) wr->num_sge > srq->rq.max_sge) {
  61                         *bad_wr = wr;
  62                         ret = -ENOMEM;
  63                         goto bail;
  64                 }
  65
  66                 spin_lock_irqsave(&srq->rq.lock, flags);
  67                 wq = srq->rq.wq;
  68                 next = wq->head + 1;
  69                 if (next >= srq->rq.size)
  70                         next = 0;
  71                 if (next == wq->tail) {
  72                         spin_unlock_irqrestore(&srq->rq.lock, flags);
  73                         *bad_wr = wr;
  74                         ret = -ENOMEM;
  75                         goto bail;
  76                 }
  77
  78                 wqe = get_rwqe_ptr(&srq->rq, wq->head);
  79                 wqe->wr_id = wr->wr_id;
  80                 wqe->num_sge = wr->num_sge;
  81                 for (i = 0; i < wr->num_sge; i++)
  82                         wqe->sg_list[i] = wr->sg_list[i];
  83                 wq->head = next;
  84                 spin_unlock_irqrestore(&srq->rq.lock, flags);
  85         }
  86         ret = 0;
  87
  88 bail:
  89         return ret;
  90 }
  91
  92 /**
  93  * ipath_create_srq - create a shared receive queue
  94  * @ibpd: the protection domain of the SRQ to create
  95  * @attr: the attributes of the SRQ
  96  * @udata: not used by the InfiniPath verbs driver
  97  */
  98 struct ib_srq *ipath_create_srq(struct ib_pd *ibpd,
  99                                 struct ib_srq_init_attr *srq_init_attr,
 100                                 struct ib_udata *udata)
 101 {
 102         struct ipath_ibdev *dev = to_idev(ibpd->device);
 103         struct ipath_srq *srq;
 104         u32 sz;
 105         struct ib_srq *ret;
 106
 107         if (srq_init_attr->attr.max_wr == 0) {
 108                 ret = ERR_PTR(-EINVAL);
 109                 goto done;
 110         }
 111
 112         if ((srq_init_attr->attr.max_sge > ib_ipath_max_srq_sges) ||
 113             (srq_init_attr->attr.max_wr > ib_ipath_max_srq_wrs)) {
 114                 ret = ERR_PTR(-EINVAL);
 115                 goto done;
 116         }
 117
 118         srq = kmalloc(sizeof(*srq), GFP_KERNEL);
 119         if (!srq) {
 120                 ret = ERR_PTR(-ENOMEM);
 121                 goto done;
 122         }
 123
 124         /*
 125          * Need to use vmalloc() if we want to support large #s of entries.
 126          */
 127         srq->rq.size = srq_init_attr->attr.max_wr + 1;
 128         srq->rq.max_sge = srq_init_attr->attr.max_sge;
 129         sz = sizeof(struct ib_sge) * srq->rq.max_sge +
 130                 sizeof(struct ipath_rwqe);
 131         srq->rq.wq = vmalloc_user(sizeof(struct ipath_rwq) + srq->rq.size * sz);
 132         if (!srq->rq.wq) {
 133                 ret = ERR_PTR(-ENOMEM);
 134                 goto bail_srq;
 135         }
 136
 137         /*
 138          * Return the address of the RWQ as the offset to mmap.
 139          * See ipath_mmap() for details.
 140          */
 141         if (udata && udata->outlen >= sizeof(__u64)) {
 142                 int err;
 143                 u32 s = sizeof(struct ipath_rwq) + srq->rq.size * sz;
 144
 145                 srq->ip =
 146                     ipath_create_mmap_info(dev, s,
 147                                            ibpd->uobject->context,
 148                                            srq->rq.wq);
 149                 if (!srq->ip) {
 150                         ret = ERR_PTR(-ENOMEM);
 151                         goto bail_wq;
 152                 }
 153
 154                 err = ib_copy_to_udata(udata, &srq->ip->offset,
 155                                        sizeof(srq->ip->offset));
 156                 if (err) {
 157                         ret = ERR_PTR(err);
 158                         goto bail_ip;
 159                 }
 160         } else
 161                 srq->ip = NULL;
 162
 163         /*
 164          * ib_create_srq() will initialize srq->ibsrq.
 165          */
 166         spin_lock_init(&srq->rq.lock);
 167         srq->rq.wq->head = 0;
 168         srq->rq.wq->tail = 0;
 169         srq->limit = srq_init_attr->attr.srq_limit;
 170
 171         spin_lock(&dev->n_srqs_lock);
 172         if (dev->n_srqs_allocated == ib_ipath_max_srqs) {
 173                 spin_unlock(&dev->n_srqs_lock);
 174                 ret = ERR_PTR(-ENOMEM);
 175                 goto bail_ip;
 176         }
 177
 178         dev->n_srqs_allocated++;
 179         spin_unlock(&dev->n_srqs_lock);
 180
 181         if (srq->ip) {
 182                 spin_lock_irq(&dev->pending_lock);
 183                 list_add(&srq->ip->pending_mmaps, &dev->pending_mmaps);
 184                 spin_unlock_irq(&dev->pending_lock);
 185         }
 186
 187         ret = &srq->ibsrq;
 188         goto done;
 189
 190 bail_ip:
 191         kfree(srq->ip);
 192 bail_wq:
 193         vfree(srq->rq.wq);
 194 bail_srq:
 195         kfree(srq);
 196 done:
 197         return ret;
 198 }
 199
 200 /**
 201  * ipath_modify_srq - modify a shared receive queue
 202  * @ibsrq: the SRQ to modify
 203  * @attr: the new attributes of the SRQ
 204  * @attr_mask: indicates which attributes to modify
 205  * @udata: user data for ipathverbs.so
 206  */
 207 int ipath_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
 208                      enum ib_srq_attr_mask attr_mask,
 209                      struct ib_udata *udata)
 210 {
 211         struct ipath_srq *srq = to_isrq(ibsrq);
 212         int ret = 0;
 213
 214         if (attr_mask & IB_SRQ_MAX_WR) {
 215                 struct ipath_rwq *owq;
 216                 struct ipath_rwq *wq;
 217                 struct ipath_rwqe *p;
 218                 u32 sz, size, n, head, tail;
 219
 220                 /* Check that the requested sizes are below the limits. */
 221                 if ((attr->max_wr > ib_ipath_max_srq_wrs) ||
 222                     ((attr_mask & IB_SRQ_LIMIT) ?
 223                      attr->srq_limit : srq->limit) > attr->max_wr) {
 224                         ret = -EINVAL;
 225                         goto bail;
 226                 }
 227
 228                 sz = sizeof(struct ipath_rwqe) +
 229                         srq->rq.max_sge * sizeof(struct ib_sge);
 230                 size = attr->max_wr + 1;
 231                 wq = vmalloc_user(sizeof(struct ipath_rwq) + size * sz);
 232                 if (!wq) {
 233                         ret = -ENOMEM;
 234                         goto bail;
 235                 }
 236
 237                 /*
 238                  * Return the address of the RWQ as the offset to mmap.
 239                  * See ipath_mmap() for details.
 240                  */
 241                 if (udata && udata->inlen >= sizeof(__u64)) {
 242                         __u64 offset_addr;
 243                         __u64 offset = (__u64) wq;
 244
 245                         ret = ib_copy_from_udata(&offset_addr, udata,
 246                                                  sizeof(offset_addr));
 247                         if (ret) {
 248                                 vfree(wq);
 249                                 goto bail;
 250                         }
 251                         udata->outbuf = (void __user *) offset_addr;
 252                         ret = ib_copy_to_udata(udata, &offset,
 253                                                sizeof(offset));
 254                         if (ret) {
 255                                 vfree(wq);
 256                                 goto bail;
 257                         }
 258                 }
 259
 260                 spin_lock_irq(&srq->rq.lock);
 261                 /*
 262                  * validate head pointer value and compute
 263                  * the number of remaining WQEs.
 264                  */
 265                 owq = srq->rq.wq;
 266                 head = owq->head;
 267                 if (head >= srq->rq.size)
 268                         head = 0;
 269                 tail = owq->tail;
 270                 if (tail >= srq->rq.size)
 271                         tail = 0;
 272                 n = head;
 273                 if (n < tail)
 274                         n += srq->rq.size - tail;
 275                 else
 276                         n -= tail;
 277                 if (size <= n) {
 278                         spin_unlock_irq(&srq->rq.lock);
 279                         vfree(wq);
 280                         ret = -EINVAL;
 281                         goto bail;
 282                 }
 283                 n = 0;
 284                 p = wq->wq;
 285                 while (tail != head) {
 286                         struct ipath_rwqe *wqe;
 287                         int i;
 288
 289                         wqe = get_rwqe_ptr(&srq->rq, tail);
 290                         p->wr_id = wqe->wr_id;
 291                         p->num_sge = wqe->num_sge;
 292                         for (i = 0; i < wqe->num_sge; i++)
 293                                 p->sg_list[i] = wqe->sg_list[i];
 294                         n++;
 295                         p = (struct ipath_rwqe *)((char *) p + sz);
 296                         if (++tail >= srq->rq.size)
 297                                 tail = 0;
 298                 }
 299                 srq->rq.wq = wq;
 300                 srq->rq.size = size;
 301                 wq->head = n;
 302                 wq->tail = 0;
 303                 if (attr_mask & IB_SRQ_LIMIT)
 304                         srq->limit = attr->srq_limit;
 305                 spin_unlock_irq(&srq->rq.lock);
 306
 307                 vfree(owq);
 308
 309                 if (srq->ip) {
 310                         struct ipath_mmap_info *ip = srq->ip;
 311                         struct ipath_ibdev *dev = to_idev(srq->ibsrq.device);
 312                         u32 s = sizeof(struct ipath_rwq) + size * sz;
 313
 314                         ipath_update_mmap_info(dev, ip, s, wq);
 315                         spin_lock_irq(&dev->pending_lock);
 316                         if (list_empty(&ip->pending_mmaps))
 317                                 list_add(&ip->pending_mmaps,
 318                                          &dev->pending_mmaps);
 319                         spin_unlock_irq(&dev->pending_lock);
 320                 }
 321         } else if (attr_mask & IB_SRQ_LIMIT) {
 322                 spin_lock_irq(&srq->rq.lock);
 323                 if (attr->srq_limit >= srq->rq.size)
 324                         ret = -EINVAL;
 325                 else
 326                         srq->limit = attr->srq_limit;
 327                 spin_unlock_irq(&srq->rq.lock);
 328         }
 329
 330 bail:
 331         return ret;
 332 }
 333
 334 int ipath_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
 335 {
 336         struct ipath_srq *srq = to_isrq(ibsrq);
 337
 338         attr->max_wr = srq->rq.size - 1;
 339         attr->max_sge = srq->rq.max_sge;
 340         attr->srq_limit = srq->limit;
 341         return 0;
 342 }
 343
 344 /**
 345  * ipath_destroy_srq - destroy a shared receive queue
 346  * @ibsrq: the SRQ to destroy
 347  */
 348 int ipath_destroy_srq(struct ib_srq *ibsrq)
 349 {
 350         struct ipath_srq *srq = to_isrq(ibsrq);
 351         struct ipath_ibdev *dev = to_idev(ibsrq->device);
 352
 353         spin_lock(&dev->n_srqs_lock);
 354         dev->n_srqs_allocated--;
 355         spin_unlock(&dev->n_srqs_lock);
 356         if (srq->ip)
 357                 kref_put(&srq->ip->ref, ipath_release_mmap_info);
 358         else
 359                 vfree(srq->rq.wq);
 360         kfree(srq);
 361
 362         return 0;
 363 }