Pileus Git - ~andy/linux/blob - drivers/infiniband/core/uverbs_cmd.c

   1 /*
   2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005, 2006, 2007 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 PathScale, Inc.  All rights reserved.
   5  * Copyright (c) 2006 Mellanox Technologies.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/file.h>
  37 #include <linux/fs.h>
  38 #include <linux/slab.h>
  39
  40 #include <asm/uaccess.h>
  41
  42 #include "uverbs.h"
  43
  44 static struct lock_class_key pd_lock_key;
  45 static struct lock_class_key mr_lock_key;
  46 static struct lock_class_key cq_lock_key;
  47 static struct lock_class_key qp_lock_key;
  48 static struct lock_class_key ah_lock_key;
  49 static struct lock_class_key srq_lock_key;
  50 static struct lock_class_key xrcd_lock_key;
  51
  52 #define INIT_UDATA(udata, ibuf, obuf, ilen, olen)                       \
  53         do {                                                            \
  54                 (udata)->inbuf  = (void __user *) (ibuf);               \
  55                 (udata)->outbuf = (void __user *) (obuf);               \
  56                 (udata)->inlen  = (ilen);                               \
  57                 (udata)->outlen = (olen);                               \
  58         } while (0)
  59
  60 /*
  61  * The ib_uobject locking scheme is as follows:
  62  *
  63  * - ib_uverbs_idr_lock protects the uverbs idrs themselves, so it
  64  *   needs to be held during all idr operations.  When an object is
  65  *   looked up, a reference must be taken on the object's kref before
  66  *   dropping this lock.
  67  *
  68  * - Each object also has an rwsem.  This rwsem must be held for
  69  *   reading while an operation that uses the object is performed.
  70  *   For example, while registering an MR, the associated PD's
  71  *   uobject.mutex must be held for reading.  The rwsem must be held
  72  *   for writing while initializing or destroying an object.
  73  *
  74  * - In addition, each object has a "live" flag.  If this flag is not
  75  *   set, then lookups of the object will fail even if it is found in
  76  *   the idr.  This handles a reader that blocks and does not acquire
  77  *   the rwsem until after the object is destroyed.  The destroy
  78  *   operation will set the live flag to 0 and then drop the rwsem;
  79  *   this will allow the reader to acquire the rwsem, see that the
  80  *   live flag is 0, and then drop the rwsem and its reference to
  81  *   object.  The underlying storage will not be freed until the last
  82  *   reference to the object is dropped.
  83  */
  84
  85 static void init_uobj(struct ib_uobject *uobj, u64 user_handle,
  86                       struct ib_ucontext *context, struct lock_class_key *key)
  87 {
  88         uobj->user_handle = user_handle;
  89         uobj->context     = context;
  90         kref_init(&uobj->ref);
  91         init_rwsem(&uobj->mutex);
  92         lockdep_set_class(&uobj->mutex, key);
  93         uobj->live        = 0;
  94 }
  95
  96 static void release_uobj(struct kref *kref)
  97 {
  98         kfree(container_of(kref, struct ib_uobject, ref));
  99 }
 100
 101 static void put_uobj(struct ib_uobject *uobj)
 102 {
 103         kref_put(&uobj->ref, release_uobj);
 104 }
 105
 106 static void put_uobj_read(struct ib_uobject *uobj)
 107 {
 108         up_read(&uobj->mutex);
 109         put_uobj(uobj);
 110 }
 111
 112 static void put_uobj_write(struct ib_uobject *uobj)
 113 {
 114         up_write(&uobj->mutex);
 115         put_uobj(uobj);
 116 }
 117
 118 static int idr_add_uobj(struct idr *idr, struct ib_uobject *uobj)
 119 {
 120         int ret;
 121
 122 retry:
 123         if (!idr_pre_get(idr, GFP_KERNEL))
 124                 return -ENOMEM;
 125
 126         spin_lock(&ib_uverbs_idr_lock);
 127         ret = idr_get_new(idr, uobj, &uobj->id);
 128         spin_unlock(&ib_uverbs_idr_lock);
 129
 130         if (ret == -EAGAIN)
 131                 goto retry;
 132
 133         return ret;
 134 }
 135
 136 void idr_remove_uobj(struct idr *idr, struct ib_uobject *uobj)
 137 {
 138         spin_lock(&ib_uverbs_idr_lock);
 139         idr_remove(idr, uobj->id);
 140         spin_unlock(&ib_uverbs_idr_lock);
 141 }
 142
 143 static struct ib_uobject *__idr_get_uobj(struct idr *idr, int id,
 144                                          struct ib_ucontext *context)
 145 {
 146         struct ib_uobject *uobj;
 147
 148         spin_lock(&ib_uverbs_idr_lock);
 149         uobj = idr_find(idr, id);
 150         if (uobj) {
 151                 if (uobj->context == context)
 152                         kref_get(&uobj->ref);
 153                 else
 154                         uobj = NULL;
 155         }
 156         spin_unlock(&ib_uverbs_idr_lock);
 157
 158         return uobj;
 159 }
 160
 161 static struct ib_uobject *idr_read_uobj(struct idr *idr, int id,
 162                                         struct ib_ucontext *context, int nested)
 163 {
 164         struct ib_uobject *uobj;
 165
 166         uobj = __idr_get_uobj(idr, id, context);
 167         if (!uobj)
 168                 return NULL;
 169
 170         if (nested)
 171                 down_read_nested(&uobj->mutex, SINGLE_DEPTH_NESTING);
 172         else
 173                 down_read(&uobj->mutex);
 174         if (!uobj->live) {
 175                 put_uobj_read(uobj);
 176                 return NULL;
 177         }
 178
 179         return uobj;
 180 }
 181
 182 static struct ib_uobject *idr_write_uobj(struct idr *idr, int id,
 183                                          struct ib_ucontext *context)
 184 {
 185         struct ib_uobject *uobj;
 186
 187         uobj = __idr_get_uobj(idr, id, context);
 188         if (!uobj)
 189                 return NULL;
 190
 191         down_write(&uobj->mutex);
 192         if (!uobj->live) {
 193                 put_uobj_write(uobj);
 194                 return NULL;
 195         }
 196
 197         return uobj;
 198 }
 199
 200 static void *idr_read_obj(struct idr *idr, int id, struct ib_ucontext *context,
 201                           int nested)
 202 {
 203         struct ib_uobject *uobj;
 204
 205         uobj = idr_read_uobj(idr, id, context, nested);
 206         return uobj ? uobj->object : NULL;
 207 }
 208
 209 static struct ib_pd *idr_read_pd(int pd_handle, struct ib_ucontext *context)
 210 {
 211         return idr_read_obj(&ib_uverbs_pd_idr, pd_handle, context, 0);
 212 }
 213
 214 static void put_pd_read(struct ib_pd *pd)
 215 {
 216         put_uobj_read(pd->uobject);
 217 }
 218
 219 static struct ib_cq *idr_read_cq(int cq_handle, struct ib_ucontext *context, int nested)
 220 {
 221         return idr_read_obj(&ib_uverbs_cq_idr, cq_handle, context, nested);
 222 }
 223
 224 static void put_cq_read(struct ib_cq *cq)
 225 {
 226         put_uobj_read(cq->uobject);
 227 }
 228
 229 static struct ib_ah *idr_read_ah(int ah_handle, struct ib_ucontext *context)
 230 {
 231         return idr_read_obj(&ib_uverbs_ah_idr, ah_handle, context, 0);
 232 }
 233
 234 static void put_ah_read(struct ib_ah *ah)
 235 {
 236         put_uobj_read(ah->uobject);
 237 }
 238
 239 static struct ib_qp *idr_read_qp(int qp_handle, struct ib_ucontext *context)
 240 {
 241         return idr_read_obj(&ib_uverbs_qp_idr, qp_handle, context, 0);
 242 }
 243
 244 static struct ib_qp *idr_write_qp(int qp_handle, struct ib_ucontext *context)
 245 {
 246         struct ib_uobject *uobj;
 247
 248         uobj = idr_write_uobj(&ib_uverbs_qp_idr, qp_handle, context);
 249         return uobj ? uobj->object : NULL;
 250 }
 251
 252 static void put_qp_read(struct ib_qp *qp)
 253 {
 254         put_uobj_read(qp->uobject);
 255 }
 256
 257 static void put_qp_write(struct ib_qp *qp)
 258 {
 259         put_uobj_write(qp->uobject);
 260 }
 261
 262 static struct ib_srq *idr_read_srq(int srq_handle, struct ib_ucontext *context)
 263 {
 264         return idr_read_obj(&ib_uverbs_srq_idr, srq_handle, context, 0);
 265 }
 266
 267 static void put_srq_read(struct ib_srq *srq)
 268 {
 269         put_uobj_read(srq->uobject);
 270 }
 271
 272 static struct ib_xrcd *idr_read_xrcd(int xrcd_handle, struct ib_ucontext *context,
 273                                      struct ib_uobject **uobj)
 274 {
 275         *uobj = idr_read_uobj(&ib_uverbs_xrcd_idr, xrcd_handle, context, 0);
 276         return *uobj ? (*uobj)->object : NULL;
 277 }
 278
 279 static void put_xrcd_read(struct ib_uobject *uobj)
 280 {
 281         put_uobj_read(uobj);
 282 }
 283
 284 ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
 285                               const char __user *buf,
 286                               int in_len, int out_len)
 287 {
 288         struct ib_uverbs_get_context      cmd;
 289         struct ib_uverbs_get_context_resp resp;
 290         struct ib_udata                   udata;
 291         struct ib_device                 *ibdev = file->device->ib_dev;
 292         struct ib_ucontext               *ucontext;
 293         struct file                      *filp;
 294         int ret;
 295
 296         if (out_len < sizeof resp)
 297                 return -ENOSPC;
 298
 299         if (copy_from_user(&cmd, buf, sizeof cmd))
 300                 return -EFAULT;
 301
 302         mutex_lock(&file->mutex);
 303
 304         if (file->ucontext) {
 305                 ret = -EINVAL;
 306                 goto err;
 307         }
 308
 309         INIT_UDATA(&udata, buf + sizeof cmd,
 310                    (unsigned long) cmd.response + sizeof resp,
 311                    in_len - sizeof cmd, out_len - sizeof resp);
 312
 313         ucontext = ibdev->alloc_ucontext(ibdev, &udata);
 314         if (IS_ERR(ucontext)) {
 315                 ret = PTR_ERR(ucontext);
 316                 goto err;
 317         }
 318
 319         ucontext->device = ibdev;
 320         INIT_LIST_HEAD(&ucontext->pd_list);
 321         INIT_LIST_HEAD(&ucontext->mr_list);
 322         INIT_LIST_HEAD(&ucontext->mw_list);
 323         INIT_LIST_HEAD(&ucontext->cq_list);
 324         INIT_LIST_HEAD(&ucontext->qp_list);
 325         INIT_LIST_HEAD(&ucontext->srq_list);
 326         INIT_LIST_HEAD(&ucontext->ah_list);
 327         INIT_LIST_HEAD(&ucontext->xrcd_list);
 328         ucontext->closing = 0;
 329
 330         resp.num_comp_vectors = file->device->num_comp_vectors;
 331
 332         ret = get_unused_fd();
 333         if (ret < 0)
 334                 goto err_free;
 335         resp.async_fd = ret;
 336
 337         filp = ib_uverbs_alloc_event_file(file, 1);
 338         if (IS_ERR(filp)) {
 339                 ret = PTR_ERR(filp);
 340                 goto err_fd;
 341         }
 342
 343         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 344                          &resp, sizeof resp)) {
 345                 ret = -EFAULT;
 346                 goto err_file;
 347         }
 348
 349         file->async_file = filp->private_data;
 350
 351         INIT_IB_EVENT_HANDLER(&file->event_handler, file->device->ib_dev,
 352                               ib_uverbs_event_handler);
 353         ret = ib_register_event_handler(&file->event_handler);
 354         if (ret)
 355                 goto err_file;
 356
 357         kref_get(&file->async_file->ref);
 358         kref_get(&file->ref);
 359         file->ucontext = ucontext;
 360
 361         fd_install(resp.async_fd, filp);
 362
 363         mutex_unlock(&file->mutex);
 364
 365         return in_len;
 366
 367 err_file:
 368         fput(filp);
 369
 370 err_fd:
 371         put_unused_fd(resp.async_fd);
 372
 373 err_free:
 374         ibdev->dealloc_ucontext(ucontext);
 375
 376 err:
 377         mutex_unlock(&file->mutex);
 378         return ret;
 379 }
 380
 381 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 382                                const char __user *buf,
 383                                int in_len, int out_len)
 384 {
 385         struct ib_uverbs_query_device      cmd;
 386         struct ib_uverbs_query_device_resp resp;
 387         struct ib_device_attr              attr;
 388         int                                ret;
 389
 390         if (out_len < sizeof resp)
 391                 return -ENOSPC;
 392
 393         if (copy_from_user(&cmd, buf, sizeof cmd))
 394                 return -EFAULT;
 395
 396         ret = ib_query_device(file->device->ib_dev, &attr);
 397         if (ret)
 398                 return ret;
 399
 400         memset(&resp, 0, sizeof resp);
 401
 402         resp.fw_ver                    = attr.fw_ver;
 403         resp.node_guid                 = file->device->ib_dev->node_guid;
 404         resp.sys_image_guid            = attr.sys_image_guid;
 405         resp.max_mr_size               = attr.max_mr_size;
 406         resp.page_size_cap             = attr.page_size_cap;
 407         resp.vendor_id                 = attr.vendor_id;
 408         resp.vendor_part_id            = attr.vendor_part_id;
 409         resp.hw_ver                    = attr.hw_ver;
 410         resp.max_qp                    = attr.max_qp;
 411         resp.max_qp_wr                 = attr.max_qp_wr;
 412         resp.device_cap_flags          = attr.device_cap_flags;
 413         resp.max_sge                   = attr.max_sge;
 414         resp.max_sge_rd                = attr.max_sge_rd;
 415         resp.max_cq                    = attr.max_cq;
 416         resp.max_cqe                   = attr.max_cqe;
 417         resp.max_mr                    = attr.max_mr;
 418         resp.max_pd                    = attr.max_pd;
 419         resp.max_qp_rd_atom            = attr.max_qp_rd_atom;
 420         resp.max_ee_rd_atom            = attr.max_ee_rd_atom;
 421         resp.max_res_rd_atom           = attr.max_res_rd_atom;
 422         resp.max_qp_init_rd_atom       = attr.max_qp_init_rd_atom;
 423         resp.max_ee_init_rd_atom       = attr.max_ee_init_rd_atom;
 424         resp.atomic_cap                = attr.atomic_cap;
 425         resp.max_ee                    = attr.max_ee;
 426         resp.max_rdd                   = attr.max_rdd;
 427         resp.max_mw                    = attr.max_mw;
 428         resp.max_raw_ipv6_qp           = attr.max_raw_ipv6_qp;
 429         resp.max_raw_ethy_qp           = attr.max_raw_ethy_qp;
 430         resp.max_mcast_grp             = attr.max_mcast_grp;
 431         resp.max_mcast_qp_attach       = attr.max_mcast_qp_attach;
 432         resp.max_total_mcast_qp_attach = attr.max_total_mcast_qp_attach;
 433         resp.max_ah                    = attr.max_ah;
 434         resp.max_fmr                   = attr.max_fmr;
 435         resp.max_map_per_fmr           = attr.max_map_per_fmr;
 436         resp.max_srq                   = attr.max_srq;
 437         resp.max_srq_wr                = attr.max_srq_wr;
 438         resp.max_srq_sge               = attr.max_srq_sge;
 439         resp.max_pkeys                 = attr.max_pkeys;
 440         resp.local_ca_ack_delay        = attr.local_ca_ack_delay;
 441         resp.phys_port_cnt             = file->device->ib_dev->phys_port_cnt;
 442
 443         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 444                          &resp, sizeof resp))
 445                 return -EFAULT;
 446
 447         return in_len;
 448 }
 449
 450 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 451                              const char __user *buf,
 452                              int in_len, int out_len)
 453 {
 454         struct ib_uverbs_query_port      cmd;
 455         struct ib_uverbs_query_port_resp resp;
 456         struct ib_port_attr              attr;
 457         int                              ret;
 458
 459         if (out_len < sizeof resp)
 460                 return -ENOSPC;
 461
 462         if (copy_from_user(&cmd, buf, sizeof cmd))
 463                 return -EFAULT;
 464
 465         ret = ib_query_port(file->device->ib_dev, cmd.port_num, &attr);
 466         if (ret)
 467                 return ret;
 468
 469         memset(&resp, 0, sizeof resp);
 470
 471         resp.state           = attr.state;
 472         resp.max_mtu         = attr.max_mtu;
 473         resp.active_mtu      = attr.active_mtu;
 474         resp.gid_tbl_len     = attr.gid_tbl_len;
 475         resp.port_cap_flags  = attr.port_cap_flags;
 476         resp.max_msg_sz      = attr.max_msg_sz;
 477         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 478         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 479         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 480         resp.lid             = attr.lid;
 481         resp.sm_lid          = attr.sm_lid;
 482         resp.lmc             = attr.lmc;
 483         resp.max_vl_num      = attr.max_vl_num;
 484         resp.sm_sl           = attr.sm_sl;
 485         resp.subnet_timeout  = attr.subnet_timeout;
 486         resp.init_type_reply = attr.init_type_reply;
 487         resp.active_width    = attr.active_width;
 488         resp.active_speed    = attr.active_speed;
 489         resp.phys_state      = attr.phys_state;
 490         resp.link_layer      = rdma_port_get_link_layer(file->device->ib_dev,
 491                                                         cmd.port_num);
 492
 493         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 494                          &resp, sizeof resp))
 495                 return -EFAULT;
 496
 497         return in_len;
 498 }
 499
 500 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 501                            const char __user *buf,
 502                            int in_len, int out_len)
 503 {
 504         struct ib_uverbs_alloc_pd      cmd;
 505         struct ib_uverbs_alloc_pd_resp resp;
 506         struct ib_udata                udata;
 507         struct ib_uobject             *uobj;
 508         struct ib_pd                  *pd;
 509         int                            ret;
 510
 511         if (out_len < sizeof resp)
 512                 return -ENOSPC;
 513
 514         if (copy_from_user(&cmd, buf, sizeof cmd))
 515                 return -EFAULT;
 516
 517         INIT_UDATA(&udata, buf + sizeof cmd,
 518                    (unsigned long) cmd.response + sizeof resp,
 519                    in_len - sizeof cmd, out_len - sizeof resp);
 520
 521         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 522         if (!uobj)
 523                 return -ENOMEM;
 524
 525         init_uobj(uobj, 0, file->ucontext, &pd_lock_key);
 526         down_write(&uobj->mutex);
 527
 528         pd = file->device->ib_dev->alloc_pd(file->device->ib_dev,
 529                                             file->ucontext, &udata);
 530         if (IS_ERR(pd)) {
 531                 ret = PTR_ERR(pd);
 532                 goto err;
 533         }
 534
 535         pd->device  = file->device->ib_dev;
 536         pd->uobject = uobj;
 537         atomic_set(&pd->usecnt, 0);
 538
 539         uobj->object = pd;
 540         ret = idr_add_uobj(&ib_uverbs_pd_idr, uobj);
 541         if (ret)
 542                 goto err_idr;
 543
 544         memset(&resp, 0, sizeof resp);
 545         resp.pd_handle = uobj->id;
 546
 547         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 548                          &resp, sizeof resp)) {
 549                 ret = -EFAULT;
 550                 goto err_copy;
 551         }
 552
 553         mutex_lock(&file->mutex);
 554         list_add_tail(&uobj->list, &file->ucontext->pd_list);
 555         mutex_unlock(&file->mutex);
 556
 557         uobj->live = 1;
 558
 559         up_write(&uobj->mutex);
 560
 561         return in_len;
 562
 563 err_copy:
 564         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 565
 566 err_idr:
 567         ib_dealloc_pd(pd);
 568
 569 err:
 570         put_uobj_write(uobj);
 571         return ret;
 572 }
 573
 574 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 575                              const char __user *buf,
 576                              int in_len, int out_len)
 577 {
 578         struct ib_uverbs_dealloc_pd cmd;
 579         struct ib_uobject          *uobj;
 580         int                         ret;
 581
 582         if (copy_from_user(&cmd, buf, sizeof cmd))
 583                 return -EFAULT;
 584
 585         uobj = idr_write_uobj(&ib_uverbs_pd_idr, cmd.pd_handle, file->ucontext);
 586         if (!uobj)
 587                 return -EINVAL;
 588
 589         ret = ib_dealloc_pd(uobj->object);
 590         if (!ret)
 591                 uobj->live = 0;
 592
 593         put_uobj_write(uobj);
 594
 595         if (ret)
 596                 return ret;
 597
 598         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 599
 600         mutex_lock(&file->mutex);
 601         list_del(&uobj->list);
 602         mutex_unlock(&file->mutex);
 603
 604         put_uobj(uobj);
 605
 606         return in_len;
 607 }
 608
 609 struct xrcd_table_entry {
 610         struct rb_node  node;
 611         struct ib_xrcd *xrcd;
 612         struct inode   *inode;
 613 };
 614
 615 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 616                             struct inode *inode,
 617                             struct ib_xrcd *xrcd)
 618 {
 619         struct xrcd_table_entry *entry, *scan;
 620         struct rb_node **p = &dev->xrcd_tree.rb_node;
 621         struct rb_node *parent = NULL;
 622
 623         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 624         if (!entry)
 625                 return -ENOMEM;
 626
 627         entry->xrcd  = xrcd;
 628         entry->inode = inode;
 629
 630         while (*p) {
 631                 parent = *p;
 632                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 633
 634                 if (inode < scan->inode) {
 635                         p = &(*p)->rb_left;
 636                 } else if (inode > scan->inode) {
 637                         p = &(*p)->rb_right;
 638                 } else {
 639                         kfree(entry);
 640                         return -EEXIST;
 641                 }
 642         }
 643
 644         rb_link_node(&entry->node, parent, p);
 645         rb_insert_color(&entry->node, &dev->xrcd_tree);
 646         igrab(inode);
 647         return 0;
 648 }
 649
 650 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 651                                                   struct inode *inode)
 652 {
 653         struct xrcd_table_entry *entry;
 654         struct rb_node *p = dev->xrcd_tree.rb_node;
 655
 656         while (p) {
 657                 entry = rb_entry(p, struct xrcd_table_entry, node);
 658
 659                 if (inode < entry->inode)
 660                         p = p->rb_left;
 661                 else if (inode > entry->inode)
 662                         p = p->rb_right;
 663                 else
 664                         return entry;
 665         }
 666
 667         return NULL;
 668 }
 669
 670 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 671 {
 672         struct xrcd_table_entry *entry;
 673
 674         entry = xrcd_table_search(dev, inode);
 675         if (!entry)
 676                 return NULL;
 677
 678         return entry->xrcd;
 679 }
 680
 681 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 682                               struct inode *inode)
 683 {
 684         struct xrcd_table_entry *entry;
 685
 686         entry = xrcd_table_search(dev, inode);
 687         if (entry) {
 688                 iput(inode);
 689                 rb_erase(&entry->node, &dev->xrcd_tree);
 690                 kfree(entry);
 691         }
 692 }
 693
 694 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 695                             const char __user *buf, int in_len,
 696                             int out_len)
 697 {
 698         struct ib_uverbs_open_xrcd      cmd;
 699         struct ib_uverbs_open_xrcd_resp resp;
 700         struct ib_udata                 udata;
 701         struct ib_uxrcd_object         *obj;
 702         struct ib_xrcd                 *xrcd = NULL;
 703         struct file                    *f = NULL;
 704         struct inode                   *inode = NULL;
 705         int                             ret = 0;
 706         int                             new_xrcd = 0;
 707
 708         if (out_len < sizeof resp)
 709                 return -ENOSPC;
 710
 711         if (copy_from_user(&cmd, buf, sizeof cmd))
 712                 return -EFAULT;
 713
 714         INIT_UDATA(&udata, buf + sizeof cmd,
 715                    (unsigned long) cmd.response + sizeof resp,
 716                    in_len - sizeof cmd, out_len - sizeof  resp);
 717
 718         mutex_lock(&file->device->xrcd_tree_mutex);
 719
 720         if (cmd.fd != -1) {
 721                 /* search for file descriptor */
 722                 f = fget(cmd.fd);
 723                 if (!f) {
 724                         ret = -EBADF;
 725                         goto err_tree_mutex_unlock;
 726                 }
 727
 728                 inode = f->f_dentry->d_inode;
 729                 if (!inode) {
 730                         ret = -EBADF;
 731                         goto err_tree_mutex_unlock;
 732                 }
 733
 734                 xrcd = find_xrcd(file->device, inode);
 735                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 736                         /* no file descriptor. Need CREATE flag */
 737                         ret = -EAGAIN;
 738                         goto err_tree_mutex_unlock;
 739                 }
 740
 741                 if (xrcd && cmd.oflags & O_EXCL) {
 742                         ret = -EINVAL;
 743                         goto err_tree_mutex_unlock;
 744                 }
 745         }
 746
 747         obj = kmalloc(sizeof *obj, GFP_KERNEL);
 748         if (!obj) {
 749                 ret = -ENOMEM;
 750                 goto err_tree_mutex_unlock;
 751         }
 752
 753         init_uobj(&obj->uobject, 0, file->ucontext, &xrcd_lock_key);
 754
 755         down_write(&obj->uobject.mutex);
 756
 757         if (!xrcd) {
 758                 xrcd = file->device->ib_dev->alloc_xrcd(file->device->ib_dev,
 759                                                         file->ucontext, &udata);
 760                 if (IS_ERR(xrcd)) {
 761                         ret = PTR_ERR(xrcd);
 762                         goto err;
 763                 }
 764
 765                 xrcd->inode   = inode;
 766                 xrcd->device  = file->device->ib_dev;
 767                 atomic_set(&xrcd->usecnt, 0);
 768                 mutex_init(&xrcd->tgt_qp_mutex);
 769                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 770                 new_xrcd = 1;
 771         }
 772
 773         atomic_set(&obj->refcnt, 0);
 774         obj->uobject.object = xrcd;
 775         ret = idr_add_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 776         if (ret)
 777                 goto err_idr;
 778
 779         memset(&resp, 0, sizeof resp);
 780         resp.xrcd_handle = obj->uobject.id;
 781
 782         if (inode) {
 783                 if (new_xrcd) {
 784                         /* create new inode/xrcd table entry */
 785                         ret = xrcd_table_insert(file->device, inode, xrcd);
 786                         if (ret)
 787                                 goto err_insert_xrcd;
 788                 }
 789                 atomic_inc(&xrcd->usecnt);
 790         }
 791
 792         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 793                          &resp, sizeof resp)) {
 794                 ret = -EFAULT;
 795                 goto err_copy;
 796         }
 797
 798         if (f)
 799                 fput(f);
 800
 801         mutex_lock(&file->mutex);
 802         list_add_tail(&obj->uobject.list, &file->ucontext->xrcd_list);
 803         mutex_unlock(&file->mutex);
 804
 805         obj->uobject.live = 1;
 806         up_write(&obj->uobject.mutex);
 807
 808         mutex_unlock(&file->device->xrcd_tree_mutex);
 809         return in_len;
 810
 811 err_copy:
 812         if (inode) {
 813                 if (new_xrcd)
 814                         xrcd_table_delete(file->device, inode);
 815                 atomic_dec(&xrcd->usecnt);
 816         }
 817
 818 err_insert_xrcd:
 819         idr_remove_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 820
 821 err_idr:
 822         ib_dealloc_xrcd(xrcd);
 823
 824 err:
 825         put_uobj_write(&obj->uobject);
 826
 827 err_tree_mutex_unlock:
 828         if (f)
 829                 fput(f);
 830
 831         mutex_unlock(&file->device->xrcd_tree_mutex);
 832
 833         return ret;
 834 }
 835
 836 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 837                              const char __user *buf, int in_len,
 838                              int out_len)
 839 {
 840         struct ib_uverbs_close_xrcd cmd;
 841         struct ib_uobject           *uobj;
 842         struct ib_xrcd              *xrcd = NULL;
 843         struct inode                *inode = NULL;
 844         struct ib_uxrcd_object      *obj;
 845         int                         live;
 846         int                         ret = 0;
 847
 848         if (copy_from_user(&cmd, buf, sizeof cmd))
 849                 return -EFAULT;
 850
 851         mutex_lock(&file->device->xrcd_tree_mutex);
 852         uobj = idr_write_uobj(&ib_uverbs_xrcd_idr, cmd.xrcd_handle, file->ucontext);
 853         if (!uobj) {
 854                 ret = -EINVAL;
 855                 goto out;
 856         }
 857
 858         xrcd  = uobj->object;
 859         inode = xrcd->inode;
 860         obj   = container_of(uobj, struct ib_uxrcd_object, uobject);
 861         if (atomic_read(&obj->refcnt)) {
 862                 put_uobj_write(uobj);
 863                 ret = -EBUSY;
 864                 goto out;
 865         }
 866
 867         if (!inode || atomic_dec_and_test(&xrcd->usecnt)) {
 868                 ret = ib_dealloc_xrcd(uobj->object);
 869                 if (!ret)
 870                         uobj->live = 0;
 871         }
 872
 873         live = uobj->live;
 874         if (inode && ret)
 875                 atomic_inc(&xrcd->usecnt);
 876
 877         put_uobj_write(uobj);
 878
 879         if (ret)
 880                 goto out;
 881
 882         if (inode && !live)
 883                 xrcd_table_delete(file->device, inode);
 884
 885         idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 886         mutex_lock(&file->mutex);
 887         list_del(&uobj->list);
 888         mutex_unlock(&file->mutex);
 889
 890         put_uobj(uobj);
 891         ret = in_len;
 892
 893 out:
 894         mutex_unlock(&file->device->xrcd_tree_mutex);
 895         return ret;
 896 }
 897
 898 void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 899                             struct ib_xrcd *xrcd)
 900 {
 901         struct inode *inode;
 902
 903         inode = xrcd->inode;
 904         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 905                 return;
 906
 907         ib_dealloc_xrcd(xrcd);
 908
 909         if (inode)
 910                 xrcd_table_delete(dev, inode);
 911 }
 912
 913 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 914                          const char __user *buf, int in_len,
 915                          int out_len)
 916 {
 917         struct ib_uverbs_reg_mr      cmd;
 918         struct ib_uverbs_reg_mr_resp resp;
 919         struct ib_udata              udata;
 920         struct ib_uobject           *uobj;
 921         struct ib_pd                *pd;
 922         struct ib_mr                *mr;
 923         int                          ret;
 924
 925         if (out_len < sizeof resp)
 926                 return -ENOSPC;
 927
 928         if (copy_from_user(&cmd, buf, sizeof cmd))
 929                 return -EFAULT;
 930
 931         INIT_UDATA(&udata, buf + sizeof cmd,
 932                    (unsigned long) cmd.response + sizeof resp,
 933                    in_len - sizeof cmd, out_len - sizeof resp);
 934
 935         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 936                 return -EINVAL;
 937
 938         /*
 939          * Local write permission is required if remote write or
 940          * remote atomic permission is also requested.
 941          */
 942         if (cmd.access_flags & (IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_REMOTE_WRITE) &&
 943             !(cmd.access_flags & IB_ACCESS_LOCAL_WRITE))
 944                 return -EINVAL;
 945
 946         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 947         if (!uobj)
 948                 return -ENOMEM;
 949
 950         init_uobj(uobj, 0, file->ucontext, &mr_lock_key);
 951         down_write(&uobj->mutex);
 952
 953         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
 954         if (!pd) {
 955                 ret = -EINVAL;
 956                 goto err_free;
 957         }
 958
 959         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 960                                      cmd.access_flags, &udata);
 961         if (IS_ERR(mr)) {
 962                 ret = PTR_ERR(mr);
 963                 goto err_put;
 964         }
 965
 966         mr->device  = pd->device;
 967         mr->pd      = pd;
 968         mr->uobject = uobj;
 969         atomic_inc(&pd->usecnt);
 970         atomic_set(&mr->usecnt, 0);
 971
 972         uobj->object = mr;
 973         ret = idr_add_uobj(&ib_uverbs_mr_idr, uobj);
 974         if (ret)
 975                 goto err_unreg;
 976
 977         memset(&resp, 0, sizeof resp);
 978         resp.lkey      = mr->lkey;
 979         resp.rkey      = mr->rkey;
 980         resp.mr_handle = uobj->id;
 981
 982         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 983                          &resp, sizeof resp)) {
 984                 ret = -EFAULT;
 985                 goto err_copy;
 986         }
 987
 988         put_pd_read(pd);
 989
 990         mutex_lock(&file->mutex);
 991         list_add_tail(&uobj->list, &file->ucontext->mr_list);
 992         mutex_unlock(&file->mutex);
 993
 994         uobj->live = 1;
 995
 996         up_write(&uobj->mutex);
 997
 998         return in_len;
 999
1000 err_copy:
1001         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1002
1003 err_unreg:
1004         ib_dereg_mr(mr);
1005
1006 err_put:
1007         put_pd_read(pd);
1008
1009 err_free:
1010         put_uobj_write(uobj);
1011         return ret;
1012 }
1013
1014 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
1015                            const char __user *buf, int in_len,
1016                            int out_len)
1017 {
1018         struct ib_uverbs_dereg_mr cmd;
1019         struct ib_mr             *mr;
1020         struct ib_uobject        *uobj;
1021         int                       ret = -EINVAL;
1022
1023         if (copy_from_user(&cmd, buf, sizeof cmd))
1024                 return -EFAULT;
1025
1026         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle, file->ucontext);
1027         if (!uobj)
1028                 return -EINVAL;
1029
1030         mr = uobj->object;
1031
1032         ret = ib_dereg_mr(mr);
1033         if (!ret)
1034                 uobj->live = 0;
1035
1036         put_uobj_write(uobj);
1037
1038         if (ret)
1039                 return ret;
1040
1041         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1042
1043         mutex_lock(&file->mutex);
1044         list_del(&uobj->list);
1045         mutex_unlock(&file->mutex);
1046
1047         put_uobj(uobj);
1048
1049         return in_len;
1050 }
1051
1052 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
1053                                       const char __user *buf, int in_len,
1054                                       int out_len)
1055 {
1056         struct ib_uverbs_create_comp_channel       cmd;
1057         struct ib_uverbs_create_comp_channel_resp  resp;
1058         struct file                               *filp;
1059         int ret;
1060
1061         if (out_len < sizeof resp)
1062                 return -ENOSPC;
1063
1064         if (copy_from_user(&cmd, buf, sizeof cmd))
1065                 return -EFAULT;
1066
1067         ret = get_unused_fd();
1068         if (ret < 0)
1069                 return ret;
1070         resp.fd = ret;
1071
1072         filp = ib_uverbs_alloc_event_file(file, 0);
1073         if (IS_ERR(filp)) {
1074                 put_unused_fd(resp.fd);
1075                 return PTR_ERR(filp);
1076         }
1077
1078         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1079                          &resp, sizeof resp)) {
1080                 put_unused_fd(resp.fd);
1081                 fput(filp);
1082                 return -EFAULT;
1083         }
1084
1085         fd_install(resp.fd, filp);
1086         return in_len;
1087 }
1088
1089 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1090                             const char __user *buf, int in_len,
1091                             int out_len)
1092 {
1093         struct ib_uverbs_create_cq      cmd;
1094         struct ib_uverbs_create_cq_resp resp;
1095         struct ib_udata                 udata;
1096         struct ib_ucq_object           *obj;
1097         struct ib_uverbs_event_file    *ev_file = NULL;
1098         struct ib_cq                   *cq;
1099         int                             ret;
1100
1101         if (out_len < sizeof resp)
1102                 return -ENOSPC;
1103
1104         if (copy_from_user(&cmd, buf, sizeof cmd))
1105                 return -EFAULT;
1106
1107         INIT_UDATA(&udata, buf + sizeof cmd,
1108                    (unsigned long) cmd.response + sizeof resp,
1109                    in_len - sizeof cmd, out_len - sizeof resp);
1110
1111         if (cmd.comp_vector >= file->device->num_comp_vectors)
1112                 return -EINVAL;
1113
1114         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1115         if (!obj)
1116                 return -ENOMEM;
1117
1118         init_uobj(&obj->uobject, cmd.user_handle, file->ucontext, &cq_lock_key);
1119         down_write(&obj->uobject.mutex);
1120
1121         if (cmd.comp_channel >= 0) {
1122                 ev_file = ib_uverbs_lookup_comp_file(cmd.comp_channel);
1123                 if (!ev_file) {
1124                         ret = -EINVAL;
1125                         goto err;
1126                 }
1127         }
1128
1129         obj->uverbs_file           = file;
1130         obj->comp_events_reported  = 0;
1131         obj->async_events_reported = 0;
1132         INIT_LIST_HEAD(&obj->comp_list);
1133         INIT_LIST_HEAD(&obj->async_list);
1134
1135         cq = file->device->ib_dev->create_cq(file->device->ib_dev, cmd.cqe,
1136                                              cmd.comp_vector,
1137                                              file->ucontext, &udata);
1138         if (IS_ERR(cq)) {
1139                 ret = PTR_ERR(cq);
1140                 goto err_file;
1141         }
1142
1143         cq->device        = file->device->ib_dev;
1144         cq->uobject       = &obj->uobject;
1145         cq->comp_handler  = ib_uverbs_comp_handler;
1146         cq->event_handler = ib_uverbs_cq_event_handler;
1147         cq->cq_context    = ev_file;
1148         atomic_set(&cq->usecnt, 0);
1149
1150         obj->uobject.object = cq;
1151         ret = idr_add_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1152         if (ret)
1153                 goto err_free;
1154
1155         memset(&resp, 0, sizeof resp);
1156         resp.cq_handle = obj->uobject.id;
1157         resp.cqe       = cq->cqe;
1158
1159         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1160                          &resp, sizeof resp)) {
1161                 ret = -EFAULT;
1162                 goto err_copy;
1163         }
1164
1165         mutex_lock(&file->mutex);
1166         list_add_tail(&obj->uobject.list, &file->ucontext->cq_list);
1167         mutex_unlock(&file->mutex);
1168
1169         obj->uobject.live = 1;
1170
1171         up_write(&obj->uobject.mutex);
1172
1173         return in_len;
1174
1175 err_copy:
1176         idr_remove_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1177
1178 err_free:
1179         ib_destroy_cq(cq);
1180
1181 err_file:
1182         if (ev_file)
1183                 ib_uverbs_release_ucq(file, ev_file, obj);
1184
1185 err:
1186         put_uobj_write(&obj->uobject);
1187         return ret;
1188 }
1189
1190 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1191                             const char __user *buf, int in_len,
1192                             int out_len)
1193 {
1194         struct ib_uverbs_resize_cq      cmd;
1195         struct ib_uverbs_resize_cq_resp resp;
1196         struct ib_udata                 udata;
1197         struct ib_cq                    *cq;
1198         int                             ret = -EINVAL;
1199
1200         if (copy_from_user(&cmd, buf, sizeof cmd))
1201                 return -EFAULT;
1202
1203         INIT_UDATA(&udata, buf + sizeof cmd,
1204                    (unsigned long) cmd.response + sizeof resp,
1205                    in_len - sizeof cmd, out_len - sizeof resp);
1206
1207         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1208         if (!cq)
1209                 return -EINVAL;
1210
1211         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1212         if (ret)
1213                 goto out;
1214
1215         resp.cqe = cq->cqe;
1216
1217         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1218                          &resp, sizeof resp.cqe))
1219                 ret = -EFAULT;
1220
1221 out:
1222         put_cq_read(cq);
1223
1224         return ret ? ret : in_len;
1225 }
1226
1227 static int copy_wc_to_user(void __user *dest, struct ib_wc *wc)
1228 {
1229         struct ib_uverbs_wc tmp;
1230
1231         tmp.wr_id               = wc->wr_id;
1232         tmp.status              = wc->status;
1233         tmp.opcode              = wc->opcode;
1234         tmp.vendor_err          = wc->vendor_err;
1235         tmp.byte_len            = wc->byte_len;
1236         tmp.ex.imm_data         = (__u32 __force) wc->ex.imm_data;
1237         tmp.qp_num              = wc->qp->qp_num;
1238         tmp.src_qp              = wc->src_qp;
1239         tmp.wc_flags            = wc->wc_flags;
1240         tmp.pkey_index          = wc->pkey_index;
1241         tmp.slid                = wc->slid;
1242         tmp.sl                  = wc->sl;
1243         tmp.dlid_path_bits      = wc->dlid_path_bits;
1244         tmp.port_num            = wc->port_num;
1245         tmp.reserved            = 0;
1246
1247         if (copy_to_user(dest, &tmp, sizeof tmp))
1248                 return -EFAULT;
1249
1250         return 0;
1251 }
1252
1253 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1254                           const char __user *buf, int in_len,
1255                           int out_len)
1256 {
1257         struct ib_uverbs_poll_cq       cmd;
1258         struct ib_uverbs_poll_cq_resp  resp;
1259         u8 __user                     *header_ptr;
1260         u8 __user                     *data_ptr;
1261         struct ib_cq                  *cq;
1262         struct ib_wc                   wc;
1263         int                            ret;
1264
1265         if (copy_from_user(&cmd, buf, sizeof cmd))
1266                 return -EFAULT;
1267
1268         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1269         if (!cq)
1270                 return -EINVAL;
1271
1272         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1273         header_ptr = (void __user *)(unsigned long) cmd.response;
1274         data_ptr = header_ptr + sizeof resp;
1275
1276         memset(&resp, 0, sizeof resp);
1277         while (resp.count < cmd.ne) {
1278                 ret = ib_poll_cq(cq, 1, &wc);
1279                 if (ret < 0)
1280                         goto out_put;
1281                 if (!ret)
1282                         break;
1283
1284                 ret = copy_wc_to_user(data_ptr, &wc);
1285                 if (ret)
1286                         goto out_put;
1287
1288                 data_ptr += sizeof(struct ib_uverbs_wc);
1289                 ++resp.count;
1290         }
1291
1292         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1293                 ret = -EFAULT;
1294                 goto out_put;
1295         }
1296
1297         ret = in_len;
1298
1299 out_put:
1300         put_cq_read(cq);
1301         return ret;
1302 }
1303
1304 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1305                                 const char __user *buf, int in_len,
1306                                 int out_len)
1307 {
1308         struct ib_uverbs_req_notify_cq cmd;
1309         struct ib_cq                  *cq;
1310
1311         if (copy_from_user(&cmd, buf, sizeof cmd))
1312                 return -EFAULT;
1313
1314         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1315         if (!cq)
1316                 return -EINVAL;
1317
1318         ib_req_notify_cq(cq, cmd.solicited_only ?
1319                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1320
1321         put_cq_read(cq);
1322
1323         return in_len;
1324 }
1325
1326 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1327                              const char __user *buf, int in_len,
1328                              int out_len)
1329 {
1330         struct ib_uverbs_destroy_cq      cmd;
1331         struct ib_uverbs_destroy_cq_resp resp;
1332         struct ib_uobject               *uobj;
1333         struct ib_cq                    *cq;
1334         struct ib_ucq_object            *obj;
1335         struct ib_uverbs_event_file     *ev_file;
1336         int                              ret = -EINVAL;
1337
1338         if (copy_from_user(&cmd, buf, sizeof cmd))
1339                 return -EFAULT;
1340
1341         uobj = idr_write_uobj(&ib_uverbs_cq_idr, cmd.cq_handle, file->ucontext);
1342         if (!uobj)
1343                 return -EINVAL;
1344         cq      = uobj->object;
1345         ev_file = cq->cq_context;
1346         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1347
1348         ret = ib_destroy_cq(cq);
1349         if (!ret)
1350                 uobj->live = 0;
1351
1352         put_uobj_write(uobj);
1353
1354         if (ret)
1355                 return ret;
1356
1357         idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
1358
1359         mutex_lock(&file->mutex);
1360         list_del(&uobj->list);
1361         mutex_unlock(&file->mutex);
1362
1363         ib_uverbs_release_ucq(file, ev_file, obj);
1364
1365         memset(&resp, 0, sizeof resp);
1366         resp.comp_events_reported  = obj->comp_events_reported;
1367         resp.async_events_reported = obj->async_events_reported;
1368
1369         put_uobj(uobj);
1370
1371         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1372                          &resp, sizeof resp))
1373                 return -EFAULT;
1374
1375         return in_len;
1376 }
1377
1378 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1379                             const char __user *buf, int in_len,
1380                             int out_len)
1381 {
1382         struct ib_uverbs_create_qp      cmd;
1383         struct ib_uverbs_create_qp_resp resp;
1384         struct ib_udata                 udata;
1385         struct ib_uqp_object           *obj;
1386         struct ib_device               *device;
1387         struct ib_pd                   *pd = NULL;
1388         struct ib_xrcd                 *xrcd = NULL;
1389         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1390         struct ib_cq                   *scq = NULL, *rcq = NULL;
1391         struct ib_srq                  *srq = NULL;
1392         struct ib_qp                   *qp;
1393         struct ib_qp_init_attr          attr;
1394         int ret;
1395
1396         if (out_len < sizeof resp)
1397                 return -ENOSPC;
1398
1399         if (copy_from_user(&cmd, buf, sizeof cmd))
1400                 return -EFAULT;
1401
1402         if (cmd.qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1403                 return -EPERM;
1404
1405         INIT_UDATA(&udata, buf + sizeof cmd,
1406                    (unsigned long) cmd.response + sizeof resp,
1407                    in_len - sizeof cmd, out_len - sizeof resp);
1408
1409         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1410         if (!obj)
1411                 return -ENOMEM;
1412
1413         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_key);
1414         down_write(&obj->uevent.uobject.mutex);
1415
1416         if (cmd.qp_type == IB_QPT_XRC_TGT) {
1417                 xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1418                 if (!xrcd) {
1419                         ret = -EINVAL;
1420                         goto err_put;
1421                 }
1422                 device = xrcd->device;
1423         } else {
1424                 pd  = idr_read_pd(cmd.pd_handle, file->ucontext);
1425                 scq = idr_read_cq(cmd.send_cq_handle, file->ucontext, 0);
1426                 if (!pd || !scq) {
1427                         ret = -EINVAL;
1428                         goto err_put;
1429                 }
1430
1431                 if (cmd.qp_type == IB_QPT_XRC_INI) {
1432                         cmd.max_recv_wr = cmd.max_recv_sge = 0;
1433                 } else {
1434                         if (cmd.is_srq) {
1435                                 srq = idr_read_srq(cmd.srq_handle, file->ucontext);
1436                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1437                                         ret = -EINVAL;
1438                                         goto err_put;
1439                                 }
1440                         }
1441                         rcq = (cmd.recv_cq_handle == cmd.send_cq_handle) ?
1442                                scq : idr_read_cq(cmd.recv_cq_handle, file->ucontext, 1);
1443                         if (!rcq) {
1444                                 ret = -EINVAL;
1445                                 goto err_put;
1446                         }
1447                 }
1448                 device = pd->device;
1449         }
1450
1451         attr.event_handler = ib_uverbs_qp_event_handler;
1452         attr.qp_context    = file;
1453         attr.send_cq       = scq;
1454         attr.recv_cq       = rcq;
1455         attr.srq           = srq;
1456         attr.xrcd          = xrcd;
1457         attr.sq_sig_type   = cmd.sq_sig_all ? IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
1458         attr.qp_type       = cmd.qp_type;
1459         attr.create_flags  = 0;
1460
1461         attr.cap.max_send_wr     = cmd.max_send_wr;
1462         attr.cap.max_recv_wr     = cmd.max_recv_wr;
1463         attr.cap.max_send_sge    = cmd.max_send_sge;
1464         attr.cap.max_recv_sge    = cmd.max_recv_sge;
1465         attr.cap.max_inline_data = cmd.max_inline_data;
1466
1467         obj->uevent.events_reported     = 0;
1468         INIT_LIST_HEAD(&obj->uevent.event_list);
1469         INIT_LIST_HEAD(&obj->mcast_list);
1470
1471         if (cmd.qp_type == IB_QPT_XRC_TGT)
1472                 qp = ib_create_qp(pd, &attr);
1473         else
1474                 qp = device->create_qp(pd, &attr, &udata);
1475
1476         if (IS_ERR(qp)) {
1477                 ret = PTR_ERR(qp);
1478                 goto err_put;
1479         }
1480
1481         if (cmd.qp_type != IB_QPT_XRC_TGT) {
1482                 qp->real_qp       = qp;
1483                 qp->device        = device;
1484                 qp->pd            = pd;
1485                 qp->send_cq       = attr.send_cq;
1486                 qp->recv_cq       = attr.recv_cq;
1487                 qp->srq           = attr.srq;
1488                 qp->event_handler = attr.event_handler;
1489                 qp->qp_context    = attr.qp_context;
1490                 qp->qp_type       = attr.qp_type;
1491                 atomic_set(&qp->usecnt, 0);
1492                 atomic_inc(&pd->usecnt);
1493                 atomic_inc(&attr.send_cq->usecnt);
1494                 if (attr.recv_cq)
1495                         atomic_inc(&attr.recv_cq->usecnt);
1496                 if (attr.srq)
1497                         atomic_inc(&attr.srq->usecnt);
1498         }
1499         qp->uobject = &obj->uevent.uobject;
1500
1501         obj->uevent.uobject.object = qp;
1502         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1503         if (ret)
1504                 goto err_destroy;
1505
1506         memset(&resp, 0, sizeof resp);
1507         resp.qpn             = qp->qp_num;
1508         resp.qp_handle       = obj->uevent.uobject.id;
1509         resp.max_recv_sge    = attr.cap.max_recv_sge;
1510         resp.max_send_sge    = attr.cap.max_send_sge;
1511         resp.max_recv_wr     = attr.cap.max_recv_wr;
1512         resp.max_send_wr     = attr.cap.max_send_wr;
1513         resp.max_inline_data = attr.cap.max_inline_data;
1514
1515         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1516                          &resp, sizeof resp)) {
1517                 ret = -EFAULT;
1518                 goto err_copy;
1519         }
1520
1521         if (xrcd)
1522                 put_xrcd_read(xrcd_uobj);
1523         if (pd)
1524                 put_pd_read(pd);
1525         if (scq)
1526                 put_cq_read(scq);
1527         if (rcq && rcq != scq)
1528                 put_cq_read(rcq);
1529         if (srq)
1530                 put_srq_read(srq);
1531
1532         mutex_lock(&file->mutex);
1533         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1534         mutex_unlock(&file->mutex);
1535
1536         obj->uevent.uobject.live = 1;
1537
1538         up_write(&obj->uevent.uobject.mutex);
1539
1540         return in_len;
1541
1542 err_copy:
1543         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1544
1545 err_destroy:
1546         ib_destroy_qp(qp);
1547
1548 err_put:
1549         if (xrcd)
1550                 put_xrcd_read(xrcd_uobj);
1551         if (pd)
1552                 put_pd_read(pd);
1553         if (scq)
1554                 put_cq_read(scq);
1555         if (rcq && rcq != scq)
1556                 put_cq_read(rcq);
1557         if (srq)
1558                 put_srq_read(srq);
1559
1560         put_uobj_write(&obj->uevent.uobject);
1561         return ret;
1562 }
1563
1564 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1565                           const char __user *buf, int in_len, int out_len)
1566 {
1567         struct ib_uverbs_open_qp        cmd;
1568         struct ib_uverbs_create_qp_resp resp;
1569         struct ib_udata                 udata;
1570         struct ib_uqp_object           *obj;
1571         struct ib_xrcd                 *xrcd;
1572         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1573         struct ib_qp                   *qp;
1574         struct ib_qp_open_attr          attr;
1575         int ret;
1576
1577         if (out_len < sizeof resp)
1578                 return -ENOSPC;
1579
1580         if (copy_from_user(&cmd, buf, sizeof cmd))
1581                 return -EFAULT;
1582
1583         INIT_UDATA(&udata, buf + sizeof cmd,
1584                    (unsigned long) cmd.response + sizeof resp,
1585                    in_len - sizeof cmd, out_len - sizeof resp);
1586
1587         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1588         if (!obj)
1589                 return -ENOMEM;
1590
1591         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_key);
1592         down_write(&obj->uevent.uobject.mutex);
1593
1594         xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1595         if (!xrcd) {
1596                 ret = -EINVAL;
1597                 goto err_put;
1598         }
1599
1600         attr.event_handler = ib_uverbs_qp_event_handler;
1601         attr.qp_context    = file;
1602         attr.qp_num        = cmd.qpn;
1603         attr.qp_type       = cmd.qp_type;
1604
1605         obj->uevent.events_reported = 0;
1606         INIT_LIST_HEAD(&obj->uevent.event_list);
1607         INIT_LIST_HEAD(&obj->mcast_list);
1608
1609         qp = ib_open_qp(xrcd, &attr);
1610         if (IS_ERR(qp)) {
1611                 ret = PTR_ERR(qp);
1612                 goto err_put;
1613         }
1614
1615         qp->uobject = &obj->uevent.uobject;
1616
1617         obj->uevent.uobject.object = qp;
1618         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1619         if (ret)
1620                 goto err_destroy;
1621
1622         memset(&resp, 0, sizeof resp);
1623         resp.qpn       = qp->qp_num;
1624         resp.qp_handle = obj->uevent.uobject.id;
1625
1626         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1627                          &resp, sizeof resp)) {
1628                 ret = -EFAULT;
1629                 goto err_remove;
1630         }
1631
1632         put_xrcd_read(xrcd_uobj);
1633
1634         mutex_lock(&file->mutex);
1635         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1636         mutex_unlock(&file->mutex);
1637
1638         obj->uevent.uobject.live = 1;
1639
1640         up_write(&obj->uevent.uobject.mutex);
1641
1642         return in_len;
1643
1644 err_remove:
1645         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1646
1647 err_destroy:
1648         ib_destroy_qp(qp);
1649
1650 err_put:
1651         put_xrcd_read(xrcd_uobj);
1652         put_uobj_write(&obj->uevent.uobject);
1653         return ret;
1654 }
1655
1656 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1657                            const char __user *buf, int in_len,
1658                            int out_len)
1659 {
1660         struct ib_uverbs_query_qp      cmd;
1661         struct ib_uverbs_query_qp_resp resp;
1662         struct ib_qp                   *qp;
1663         struct ib_qp_attr              *attr;
1664         struct ib_qp_init_attr         *init_attr;
1665         int                            ret;
1666
1667         if (copy_from_user(&cmd, buf, sizeof cmd))
1668                 return -EFAULT;
1669
1670         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1671         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1672         if (!attr || !init_attr) {
1673                 ret = -ENOMEM;
1674                 goto out;
1675         }
1676
1677         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1678         if (!qp) {
1679                 ret = -EINVAL;
1680                 goto out;
1681         }
1682
1683         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1684
1685         put_qp_read(qp);
1686
1687         if (ret)
1688                 goto out;
1689
1690         memset(&resp, 0, sizeof resp);
1691
1692         resp.qp_state               = attr->qp_state;
1693         resp.cur_qp_state           = attr->cur_qp_state;
1694         resp.path_mtu               = attr->path_mtu;
1695         resp.path_mig_state         = attr->path_mig_state;
1696         resp.qkey                   = attr->qkey;
1697         resp.rq_psn                 = attr->rq_psn;
1698         resp.sq_psn                 = attr->sq_psn;
1699         resp.dest_qp_num            = attr->dest_qp_num;
1700         resp.qp_access_flags        = attr->qp_access_flags;
1701         resp.pkey_index             = attr->pkey_index;
1702         resp.alt_pkey_index         = attr->alt_pkey_index;
1703         resp.sq_draining            = attr->sq_draining;
1704         resp.max_rd_atomic          = attr->max_rd_atomic;
1705         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1706         resp.min_rnr_timer          = attr->min_rnr_timer;
1707         resp.port_num               = attr->port_num;
1708         resp.timeout                = attr->timeout;
1709         resp.retry_cnt              = attr->retry_cnt;
1710         resp.rnr_retry              = attr->rnr_retry;
1711         resp.alt_port_num           = attr->alt_port_num;
1712         resp.alt_timeout            = attr->alt_timeout;
1713
1714         memcpy(resp.dest.dgid, attr->ah_attr.grh.dgid.raw, 16);
1715         resp.dest.flow_label        = attr->ah_attr.grh.flow_label;
1716         resp.dest.sgid_index        = attr->ah_attr.grh.sgid_index;
1717         resp.dest.hop_limit         = attr->ah_attr.grh.hop_limit;
1718         resp.dest.traffic_class     = attr->ah_attr.grh.traffic_class;
1719         resp.dest.dlid              = attr->ah_attr.dlid;
1720         resp.dest.sl                = attr->ah_attr.sl;
1721         resp.dest.src_path_bits     = attr->ah_attr.src_path_bits;
1722         resp.dest.static_rate       = attr->ah_attr.static_rate;
1723         resp.dest.is_global         = !!(attr->ah_attr.ah_flags & IB_AH_GRH);
1724         resp.dest.port_num          = attr->ah_attr.port_num;
1725
1726         memcpy(resp.alt_dest.dgid, attr->alt_ah_attr.grh.dgid.raw, 16);
1727         resp.alt_dest.flow_label    = attr->alt_ah_attr.grh.flow_label;
1728         resp.alt_dest.sgid_index    = attr->alt_ah_attr.grh.sgid_index;
1729         resp.alt_dest.hop_limit     = attr->alt_ah_attr.grh.hop_limit;
1730         resp.alt_dest.traffic_class = attr->alt_ah_attr.grh.traffic_class;
1731         resp.alt_dest.dlid          = attr->alt_ah_attr.dlid;
1732         resp.alt_dest.sl            = attr->alt_ah_attr.sl;
1733         resp.alt_dest.src_path_bits = attr->alt_ah_attr.src_path_bits;
1734         resp.alt_dest.static_rate   = attr->alt_ah_attr.static_rate;
1735         resp.alt_dest.is_global     = !!(attr->alt_ah_attr.ah_flags & IB_AH_GRH);
1736         resp.alt_dest.port_num      = attr->alt_ah_attr.port_num;
1737
1738         resp.max_send_wr            = init_attr->cap.max_send_wr;
1739         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1740         resp.max_send_sge           = init_attr->cap.max_send_sge;
1741         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1742         resp.max_inline_data        = init_attr->cap.max_inline_data;
1743         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1744
1745         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1746                          &resp, sizeof resp))
1747                 ret = -EFAULT;
1748
1749 out:
1750         kfree(attr);
1751         kfree(init_attr);
1752
1753         return ret ? ret : in_len;
1754 }
1755
1756 /* Remove ignored fields set in the attribute mask */
1757 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1758 {
1759         switch (qp_type) {
1760         case IB_QPT_XRC_INI:
1761                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1762         case IB_QPT_XRC_TGT:
1763                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1764                                 IB_QP_RNR_RETRY);
1765         default:
1766                 return mask;
1767         }
1768 }
1769
1770 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
1771                             const char __user *buf, int in_len,
1772                             int out_len)
1773 {
1774         struct ib_uverbs_modify_qp cmd;
1775         struct ib_udata            udata;
1776         struct ib_qp              *qp;
1777         struct ib_qp_attr         *attr;
1778         int                        ret;
1779
1780         if (copy_from_user(&cmd, buf, sizeof cmd))
1781                 return -EFAULT;
1782
1783         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
1784                    out_len);
1785
1786         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1787         if (!attr)
1788                 return -ENOMEM;
1789
1790         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1791         if (!qp) {
1792                 ret = -EINVAL;
1793                 goto out;
1794         }
1795
1796         attr->qp_state            = cmd.qp_state;
1797         attr->cur_qp_state        = cmd.cur_qp_state;
1798         attr->path_mtu            = cmd.path_mtu;
1799         attr->path_mig_state      = cmd.path_mig_state;
1800         attr->qkey                = cmd.qkey;
1801         attr->rq_psn              = cmd.rq_psn;
1802         attr->sq_psn              = cmd.sq_psn;
1803         attr->dest_qp_num         = cmd.dest_qp_num;
1804         attr->qp_access_flags     = cmd.qp_access_flags;
1805         attr->pkey_index          = cmd.pkey_index;
1806         attr->alt_pkey_index      = cmd.alt_pkey_index;
1807         attr->en_sqd_async_notify = cmd.en_sqd_async_notify;
1808         attr->max_rd_atomic       = cmd.max_rd_atomic;
1809         attr->max_dest_rd_atomic  = cmd.max_dest_rd_atomic;
1810         attr->min_rnr_timer       = cmd.min_rnr_timer;
1811         attr->port_num            = cmd.port_num;
1812         attr->timeout             = cmd.timeout;
1813         attr->retry_cnt           = cmd.retry_cnt;
1814         attr->rnr_retry           = cmd.rnr_retry;
1815         attr->alt_port_num        = cmd.alt_port_num;
1816         attr->alt_timeout         = cmd.alt_timeout;
1817
1818         memcpy(attr->ah_attr.grh.dgid.raw, cmd.dest.dgid, 16);
1819         attr->ah_attr.grh.flow_label        = cmd.dest.flow_label;
1820         attr->ah_attr.grh.sgid_index        = cmd.dest.sgid_index;
1821         attr->ah_attr.grh.hop_limit         = cmd.dest.hop_limit;
1822         attr->ah_attr.grh.traffic_class     = cmd.dest.traffic_class;
1823         attr->ah_attr.dlid                  = cmd.dest.dlid;
1824         attr->ah_attr.sl                    = cmd.dest.sl;
1825         attr->ah_attr.src_path_bits         = cmd.dest.src_path_bits;
1826         attr->ah_attr.static_rate           = cmd.dest.static_rate;
1827         attr->ah_attr.ah_flags              = cmd.dest.is_global ? IB_AH_GRH : 0;
1828         attr->ah_attr.port_num              = cmd.dest.port_num;
1829
1830         memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd.alt_dest.dgid, 16);
1831         attr->alt_ah_attr.grh.flow_label    = cmd.alt_dest.flow_label;
1832         attr->alt_ah_attr.grh.sgid_index    = cmd.alt_dest.sgid_index;
1833         attr->alt_ah_attr.grh.hop_limit     = cmd.alt_dest.hop_limit;
1834         attr->alt_ah_attr.grh.traffic_class = cmd.alt_dest.traffic_class;
1835         attr->alt_ah_attr.dlid              = cmd.alt_dest.dlid;
1836         attr->alt_ah_attr.sl                = cmd.alt_dest.sl;
1837         attr->alt_ah_attr.src_path_bits     = cmd.alt_dest.src_path_bits;
1838         attr->alt_ah_attr.static_rate       = cmd.alt_dest.static_rate;
1839         attr->alt_ah_attr.ah_flags          = cmd.alt_dest.is_global ? IB_AH_GRH : 0;
1840         attr->alt_ah_attr.port_num          = cmd.alt_dest.port_num;
1841
1842         if (qp->real_qp == qp) {
1843                 ret = qp->device->modify_qp(qp, attr,
1844                         modify_qp_mask(qp->qp_type, cmd.attr_mask), &udata);
1845         } else {
1846                 ret = ib_modify_qp(qp, attr, modify_qp_mask(qp->qp_type, cmd.attr_mask));
1847         }
1848
1849         put_qp_read(qp);
1850
1851         if (ret)
1852                 goto out;
1853
1854         ret = in_len;
1855
1856 out:
1857         kfree(attr);
1858
1859         return ret;
1860 }
1861
1862 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
1863                              const char __user *buf, int in_len,
1864                              int out_len)
1865 {
1866         struct ib_uverbs_destroy_qp      cmd;
1867         struct ib_uverbs_destroy_qp_resp resp;
1868         struct ib_uobject               *uobj;
1869         struct ib_qp                    *qp;
1870         struct ib_uqp_object            *obj;
1871         int                              ret = -EINVAL;
1872
1873         if (copy_from_user(&cmd, buf, sizeof cmd))
1874                 return -EFAULT;
1875
1876         memset(&resp, 0, sizeof resp);
1877
1878         uobj = idr_write_uobj(&ib_uverbs_qp_idr, cmd.qp_handle, file->ucontext);
1879         if (!uobj)
1880                 return -EINVAL;
1881         qp  = uobj->object;
1882         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
1883
1884         if (!list_empty(&obj->mcast_list)) {
1885                 put_uobj_write(uobj);
1886                 return -EBUSY;
1887         }
1888
1889         ret = ib_destroy_qp(qp);
1890         if (!ret)
1891                 uobj->live = 0;
1892
1893         put_uobj_write(uobj);
1894
1895         if (ret)
1896                 return ret;
1897
1898         idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
1899
1900         mutex_lock(&file->mutex);
1901         list_del(&uobj->list);
1902         mutex_unlock(&file->mutex);
1903
1904         ib_uverbs_release_uevent(file, &obj->uevent);
1905
1906         resp.events_reported = obj->uevent.events_reported;
1907
1908         put_uobj(uobj);
1909
1910         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1911                          &resp, sizeof resp))
1912                 return -EFAULT;
1913
1914         return in_len;
1915 }
1916
1917 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
1918                             const char __user *buf, int in_len,
1919                             int out_len)
1920 {
1921         struct ib_uverbs_post_send      cmd;
1922         struct ib_uverbs_post_send_resp resp;
1923         struct ib_uverbs_send_wr       *user_wr;
1924         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
1925         struct ib_qp                   *qp;
1926         int                             i, sg_ind;
1927         int                             is_ud;
1928         ssize_t                         ret = -EINVAL;
1929
1930         if (copy_from_user(&cmd, buf, sizeof cmd))
1931                 return -EFAULT;
1932
1933         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
1934             cmd.sge_count * sizeof (struct ib_uverbs_sge))
1935                 return -EINVAL;
1936
1937         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
1938                 return -EINVAL;
1939
1940         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
1941         if (!user_wr)
1942                 return -ENOMEM;
1943
1944         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1945         if (!qp)
1946                 goto out;
1947
1948         is_ud = qp->qp_type == IB_QPT_UD;
1949         sg_ind = 0;
1950         last = NULL;
1951         for (i = 0; i < cmd.wr_count; ++i) {
1952                 if (copy_from_user(user_wr,
1953                                    buf + sizeof cmd + i * cmd.wqe_size,
1954                                    cmd.wqe_size)) {
1955                         ret = -EFAULT;
1956                         goto out_put;
1957                 }
1958
1959                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
1960                         ret = -EINVAL;
1961                         goto out_put;
1962                 }
1963
1964                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
1965                                user_wr->num_sge * sizeof (struct ib_sge),
1966                                GFP_KERNEL);
1967                 if (!next) {
1968                         ret = -ENOMEM;
1969                         goto out_put;
1970                 }
1971
1972                 if (!last)
1973                         wr = next;
1974                 else
1975                         last->next = next;
1976                 last = next;
1977
1978                 next->next       = NULL;
1979                 next->wr_id      = user_wr->wr_id;
1980                 next->num_sge    = user_wr->num_sge;
1981                 next->opcode     = user_wr->opcode;
1982                 next->send_flags = user_wr->send_flags;
1983
1984                 if (is_ud) {
1985                         next->wr.ud.ah = idr_read_ah(user_wr->wr.ud.ah,
1986                                                      file->ucontext);
1987                         if (!next->wr.ud.ah) {
1988                                 ret = -EINVAL;
1989                                 goto out_put;
1990                         }
1991                         next->wr.ud.remote_qpn  = user_wr->wr.ud.remote_qpn;
1992                         next->wr.ud.remote_qkey = user_wr->wr.ud.remote_qkey;
1993                 } else {
1994                         switch (next->opcode) {
1995                         case IB_WR_RDMA_WRITE_WITH_IMM:
1996                                 next->ex.imm_data =
1997                                         (__be32 __force) user_wr->ex.imm_data;
1998                         case IB_WR_RDMA_WRITE:
1999                         case IB_WR_RDMA_READ:
2000                                 next->wr.rdma.remote_addr =
2001                                         user_wr->wr.rdma.remote_addr;
2002                                 next->wr.rdma.rkey        =
2003                                         user_wr->wr.rdma.rkey;
2004                                 break;
2005                         case IB_WR_SEND_WITH_IMM:
2006                                 next->ex.imm_data =
2007                                         (__be32 __force) user_wr->ex.imm_data;
2008                                 break;
2009                         case IB_WR_SEND_WITH_INV:
2010                                 next->ex.invalidate_rkey =
2011                                         user_wr->ex.invalidate_rkey;
2012                                 break;
2013                         case IB_WR_ATOMIC_CMP_AND_SWP:
2014                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2015                                 next->wr.atomic.remote_addr =
2016                                         user_wr->wr.atomic.remote_addr;
2017                                 next->wr.atomic.compare_add =
2018                                         user_wr->wr.atomic.compare_add;
2019                                 next->wr.atomic.swap = user_wr->wr.atomic.swap;
2020                                 next->wr.atomic.rkey = user_wr->wr.atomic.rkey;
2021                                 break;
2022                         default:
2023                                 break;
2024                         }
2025                 }
2026
2027                 if (next->num_sge) {
2028                         next->sg_list = (void *) next +
2029                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2030                         if (copy_from_user(next->sg_list,
2031                                            buf + sizeof cmd +
2032                                            cmd.wr_count * cmd.wqe_size +
2033                                            sg_ind * sizeof (struct ib_sge),
2034                                            next->num_sge * sizeof (struct ib_sge))) {
2035                                 ret = -EFAULT;
2036                                 goto out_put;
2037                         }
2038                         sg_ind += next->num_sge;
2039                 } else
2040                         next->sg_list = NULL;
2041         }
2042
2043         resp.bad_wr = 0;
2044         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2045         if (ret)
2046                 for (next = wr; next; next = next->next) {
2047                         ++resp.bad_wr;
2048                         if (next == bad_wr)
2049                                 break;
2050                 }
2051
2052         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2053                          &resp, sizeof resp))
2054                 ret = -EFAULT;
2055
2056 out_put:
2057         put_qp_read(qp);
2058
2059         while (wr) {
2060                 if (is_ud && wr->wr.ud.ah)
2061                         put_ah_read(wr->wr.ud.ah);
2062                 next = wr->next;
2063                 kfree(wr);
2064                 wr = next;
2065         }
2066
2067 out:
2068         kfree(user_wr);
2069
2070         return ret ? ret : in_len;
2071 }
2072
2073 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2074                                                     int in_len,
2075                                                     u32 wr_count,
2076                                                     u32 sge_count,
2077                                                     u32 wqe_size)
2078 {
2079         struct ib_uverbs_recv_wr *user_wr;
2080         struct ib_recv_wr        *wr = NULL, *last, *next;
2081         int                       sg_ind;
2082         int                       i;
2083         int                       ret;
2084
2085         if (in_len < wqe_size * wr_count +
2086             sge_count * sizeof (struct ib_uverbs_sge))
2087                 return ERR_PTR(-EINVAL);
2088
2089         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2090                 return ERR_PTR(-EINVAL);
2091
2092         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2093         if (!user_wr)
2094                 return ERR_PTR(-ENOMEM);
2095
2096         sg_ind = 0;
2097         last = NULL;
2098         for (i = 0; i < wr_count; ++i) {
2099                 if (copy_from_user(user_wr, buf + i * wqe_size,
2100                                    wqe_size)) {
2101                         ret = -EFAULT;
2102                         goto err;
2103                 }
2104
2105                 if (user_wr->num_sge + sg_ind > sge_count) {
2106                         ret = -EINVAL;
2107                         goto err;
2108                 }
2109
2110                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2111                                user_wr->num_sge * sizeof (struct ib_sge),
2112                                GFP_KERNEL);
2113                 if (!next) {
2114                         ret = -ENOMEM;
2115                         goto err;
2116                 }
2117
2118                 if (!last)
2119                         wr = next;
2120                 else
2121                         last->next = next;
2122                 last = next;
2123
2124                 next->next       = NULL;
2125                 next->wr_id      = user_wr->wr_id;
2126                 next->num_sge    = user_wr->num_sge;
2127
2128                 if (next->num_sge) {
2129                         next->sg_list = (void *) next +
2130                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2131                         if (copy_from_user(next->sg_list,
2132                                            buf + wr_count * wqe_size +
2133                                            sg_ind * sizeof (struct ib_sge),
2134                                            next->num_sge * sizeof (struct ib_sge))) {
2135                                 ret = -EFAULT;
2136                                 goto err;
2137                         }
2138                         sg_ind += next->num_sge;
2139                 } else
2140                         next->sg_list = NULL;
2141         }
2142
2143         kfree(user_wr);
2144         return wr;
2145
2146 err:
2147         kfree(user_wr);
2148
2149         while (wr) {
2150                 next = wr->next;
2151                 kfree(wr);
2152                 wr = next;
2153         }
2154
2155         return ERR_PTR(ret);
2156 }
2157
2158 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2159                             const char __user *buf, int in_len,
2160                             int out_len)
2161 {
2162         struct ib_uverbs_post_recv      cmd;
2163         struct ib_uverbs_post_recv_resp resp;
2164         struct ib_recv_wr              *wr, *next, *bad_wr;
2165         struct ib_qp                   *qp;
2166         ssize_t                         ret = -EINVAL;
2167
2168         if (copy_from_user(&cmd, buf, sizeof cmd))
2169                 return -EFAULT;
2170
2171         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2172                                        in_len - sizeof cmd, cmd.wr_count,
2173                                        cmd.sge_count, cmd.wqe_size);
2174         if (IS_ERR(wr))
2175                 return PTR_ERR(wr);
2176
2177         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2178         if (!qp)
2179                 goto out;
2180
2181         resp.bad_wr = 0;
2182         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2183
2184         put_qp_read(qp);
2185
2186         if (ret)
2187                 for (next = wr; next; next = next->next) {
2188                         ++resp.bad_wr;
2189                         if (next == bad_wr)
2190                                 break;
2191                 }
2192
2193         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2194                          &resp, sizeof resp))
2195                 ret = -EFAULT;
2196
2197 out:
2198         while (wr) {
2199                 next = wr->next;
2200                 kfree(wr);
2201                 wr = next;
2202         }
2203
2204         return ret ? ret : in_len;
2205 }
2206
2207 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2208                                 const char __user *buf, int in_len,
2209                                 int out_len)
2210 {
2211         struct ib_uverbs_post_srq_recv      cmd;
2212         struct ib_uverbs_post_srq_recv_resp resp;
2213         struct ib_recv_wr                  *wr, *next, *bad_wr;
2214         struct ib_srq                      *srq;
2215         ssize_t                             ret = -EINVAL;
2216
2217         if (copy_from_user(&cmd, buf, sizeof cmd))
2218                 return -EFAULT;
2219
2220         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2221                                        in_len - sizeof cmd, cmd.wr_count,
2222                                        cmd.sge_count, cmd.wqe_size);
2223         if (IS_ERR(wr))
2224                 return PTR_ERR(wr);
2225
2226         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2227         if (!srq)
2228                 goto out;
2229
2230         resp.bad_wr = 0;
2231         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2232
2233         put_srq_read(srq);
2234
2235         if (ret)
2236                 for (next = wr; next; next = next->next) {
2237                         ++resp.bad_wr;
2238                         if (next == bad_wr)
2239                                 break;
2240                 }
2241
2242         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2243                          &resp, sizeof resp))
2244                 ret = -EFAULT;
2245
2246 out:
2247         while (wr) {
2248                 next = wr->next;
2249                 kfree(wr);
2250                 wr = next;
2251         }
2252
2253         return ret ? ret : in_len;
2254 }
2255
2256 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2257                             const char __user *buf, int in_len,
2258                             int out_len)
2259 {
2260         struct ib_uverbs_create_ah       cmd;
2261         struct ib_uverbs_create_ah_resp  resp;
2262         struct ib_uobject               *uobj;
2263         struct ib_pd                    *pd;
2264         struct ib_ah                    *ah;
2265         struct ib_ah_attr               attr;
2266         int ret;
2267
2268         if (out_len < sizeof resp)
2269                 return -ENOSPC;
2270
2271         if (copy_from_user(&cmd, buf, sizeof cmd))
2272                 return -EFAULT;
2273
2274         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
2275         if (!uobj)
2276                 return -ENOMEM;
2277
2278         init_uobj(uobj, cmd.user_handle, file->ucontext, &ah_lock_key);
2279         down_write(&uobj->mutex);
2280
2281         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
2282         if (!pd) {
2283                 ret = -EINVAL;
2284                 goto err;
2285         }
2286
2287         attr.dlid              = cmd.attr.dlid;
2288         attr.sl                = cmd.attr.sl;
2289         attr.src_path_bits     = cmd.attr.src_path_bits;
2290         attr.static_rate       = cmd.attr.static_rate;
2291         attr.ah_flags          = cmd.attr.is_global ? IB_AH_GRH : 0;
2292         attr.port_num          = cmd.attr.port_num;
2293         attr.grh.flow_label    = cmd.attr.grh.flow_label;
2294         attr.grh.sgid_index    = cmd.attr.grh.sgid_index;
2295         attr.grh.hop_limit     = cmd.attr.grh.hop_limit;
2296         attr.grh.traffic_class = cmd.attr.grh.traffic_class;
2297         memcpy(attr.grh.dgid.raw, cmd.attr.grh.dgid, 16);
2298
2299         ah = ib_create_ah(pd, &attr);
2300         if (IS_ERR(ah)) {
2301                 ret = PTR_ERR(ah);
2302                 goto err_put;
2303         }
2304
2305         ah->uobject  = uobj;
2306         uobj->object = ah;
2307
2308         ret = idr_add_uobj(&ib_uverbs_ah_idr, uobj);
2309         if (ret)
2310                 goto err_destroy;
2311
2312         resp.ah_handle = uobj->id;
2313
2314         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2315                          &resp, sizeof resp)) {
2316                 ret = -EFAULT;
2317                 goto err_copy;
2318         }
2319
2320         put_pd_read(pd);
2321
2322         mutex_lock(&file->mutex);
2323         list_add_tail(&uobj->list, &file->ucontext->ah_list);
2324         mutex_unlock(&file->mutex);
2325
2326         uobj->live = 1;
2327
2328         up_write(&uobj->mutex);
2329
2330         return in_len;
2331
2332 err_copy:
2333         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2334
2335 err_destroy:
2336         ib_destroy_ah(ah);
2337
2338 err_put:
2339         put_pd_read(pd);
2340
2341 err:
2342         put_uobj_write(uobj);
2343         return ret;
2344 }
2345
2346 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2347                              const char __user *buf, int in_len, int out_len)
2348 {
2349         struct ib_uverbs_destroy_ah cmd;
2350         struct ib_ah               *ah;
2351         struct ib_uobject          *uobj;
2352         int                         ret;
2353
2354         if (copy_from_user(&cmd, buf, sizeof cmd))
2355                 return -EFAULT;
2356
2357         uobj = idr_write_uobj(&ib_uverbs_ah_idr, cmd.ah_handle, file->ucontext);
2358         if (!uobj)
2359                 return -EINVAL;
2360         ah = uobj->object;
2361
2362         ret = ib_destroy_ah(ah);
2363         if (!ret)
2364                 uobj->live = 0;
2365
2366         put_uobj_write(uobj);
2367
2368         if (ret)
2369                 return ret;
2370
2371         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2372
2373         mutex_lock(&file->mutex);
2374         list_del(&uobj->list);
2375         mutex_unlock(&file->mutex);
2376
2377         put_uobj(uobj);
2378
2379         return in_len;
2380 }
2381
2382 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2383                                const char __user *buf, int in_len,
2384                                int out_len)
2385 {
2386         struct ib_uverbs_attach_mcast cmd;
2387         struct ib_qp                 *qp;
2388         struct ib_uqp_object         *obj;
2389         struct ib_uverbs_mcast_entry *mcast;
2390         int                           ret;
2391
2392         if (copy_from_user(&cmd, buf, sizeof cmd))
2393                 return -EFAULT;
2394
2395         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2396         if (!qp)
2397                 return -EINVAL;
2398
2399         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2400
2401         list_for_each_entry(mcast, &obj->mcast_list, list)
2402                 if (cmd.mlid == mcast->lid &&
2403                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2404                         ret = 0;
2405                         goto out_put;
2406                 }
2407
2408         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2409         if (!mcast) {
2410                 ret = -ENOMEM;
2411                 goto out_put;
2412         }
2413
2414         mcast->lid = cmd.mlid;
2415         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2416
2417         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2418         if (!ret)
2419                 list_add_tail(&mcast->list, &obj->mcast_list);
2420         else
2421                 kfree(mcast);
2422
2423 out_put:
2424         put_qp_write(qp);
2425
2426         return ret ? ret : in_len;
2427 }
2428
2429 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2430                                const char __user *buf, int in_len,
2431                                int out_len)
2432 {
2433         struct ib_uverbs_detach_mcast cmd;
2434         struct ib_uqp_object         *obj;
2435         struct ib_qp                 *qp;
2436         struct ib_uverbs_mcast_entry *mcast;
2437         int                           ret = -EINVAL;
2438
2439         if (copy_from_user(&cmd, buf, sizeof cmd))
2440                 return -EFAULT;
2441
2442         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2443         if (!qp)
2444                 return -EINVAL;
2445
2446         ret = ib_detach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
2447         if (ret)
2448                 goto out_put;
2449
2450         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2451
2452         list_for_each_entry(mcast, &obj->mcast_list, list)
2453                 if (cmd.mlid == mcast->lid &&
2454                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2455                         list_del(&mcast->list);
2456                         kfree(mcast);
2457                         break;
2458                 }
2459
2460 out_put:
2461         put_qp_write(qp);
2462
2463         return ret ? ret : in_len;
2464 }
2465
2466 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
2467                                 struct ib_uverbs_create_xsrq *cmd,
2468                                 struct ib_udata *udata)
2469 {
2470         struct ib_uverbs_create_srq_resp resp;
2471         struct ib_usrq_object           *obj;
2472         struct ib_pd                    *pd;
2473         struct ib_srq                   *srq;
2474         struct ib_uobject               *uninitialized_var(xrcd_uobj);
2475         struct ib_srq_init_attr          attr;
2476         int ret;
2477
2478         obj = kmalloc(sizeof *obj, GFP_KERNEL);
2479         if (!obj)
2480                 return -ENOMEM;
2481
2482         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext, &srq_lock_key);
2483         down_write(&obj->uevent.uobject.mutex);
2484
2485         pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
2486         if (!pd) {
2487                 ret = -EINVAL;
2488                 goto err;
2489         }
2490
2491         if (cmd->srq_type == IB_SRQT_XRC) {
2492                 attr.ext.xrc.cq  = idr_read_cq(cmd->cq_handle, file->ucontext, 0);
2493                 if (!attr.ext.xrc.cq) {
2494                         ret = -EINVAL;
2495                         goto err_put_pd;
2496                 }
2497
2498                 attr.ext.xrc.xrcd  = idr_read_xrcd(cmd->xrcd_handle, file->ucontext, &xrcd_uobj);
2499                 if (!attr.ext.xrc.xrcd) {
2500                         ret = -EINVAL;
2501                         goto err_put_cq;
2502                 }
2503
2504                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
2505                 atomic_inc(&obj->uxrcd->refcnt);
2506         }
2507
2508         attr.event_handler  = ib_uverbs_srq_event_handler;
2509         attr.srq_context    = file;
2510         attr.srq_type       = cmd->srq_type;
2511         attr.attr.max_wr    = cmd->max_wr;
2512         attr.attr.max_sge   = cmd->max_sge;
2513         attr.attr.srq_limit = cmd->srq_limit;
2514
2515         obj->uevent.events_reported = 0;
2516         INIT_LIST_HEAD(&obj->uevent.event_list);
2517
2518         srq = pd->device->create_srq(pd, &attr, udata);
2519         if (IS_ERR(srq)) {
2520                 ret = PTR_ERR(srq);
2521                 goto err_put;
2522         }
2523
2524         srq->device        = pd->device;
2525         srq->pd            = pd;
2526         srq->srq_type      = cmd->srq_type;
2527         srq->uobject       = &obj->uevent.uobject;
2528         srq->event_handler = attr.event_handler;
2529         srq->srq_context   = attr.srq_context;
2530
2531         if (cmd->srq_type == IB_SRQT_XRC) {
2532                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
2533                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
2534                 atomic_inc(&attr.ext.xrc.cq->usecnt);
2535                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
2536         }
2537
2538         atomic_inc(&pd->usecnt);
2539         atomic_set(&srq->usecnt, 0);
2540
2541         obj->uevent.uobject.object = srq;
2542         ret = idr_add_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2543         if (ret)
2544                 goto err_destroy;
2545
2546         memset(&resp, 0, sizeof resp);
2547         resp.srq_handle = obj->uevent.uobject.id;
2548         resp.max_wr     = attr.attr.max_wr;
2549         resp.max_sge    = attr.attr.max_sge;
2550         if (cmd->srq_type == IB_SRQT_XRC)
2551                 resp.srqn = srq->ext.xrc.srq_num;
2552
2553         if (copy_to_user((void __user *) (unsigned long) cmd->response,
2554                          &resp, sizeof resp)) {
2555                 ret = -EFAULT;
2556                 goto err_copy;
2557         }
2558
2559         if (cmd->srq_type == IB_SRQT_XRC) {
2560                 put_uobj_read(xrcd_uobj);
2561                 put_cq_read(attr.ext.xrc.cq);
2562         }
2563         put_pd_read(pd);
2564
2565         mutex_lock(&file->mutex);
2566         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->srq_list);
2567         mutex_unlock(&file->mutex);
2568
2569         obj->uevent.uobject.live = 1;
2570
2571         up_write(&obj->uevent.uobject.mutex);
2572
2573         return 0;
2574
2575 err_copy:
2576         idr_remove_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2577
2578 err_destroy:
2579         ib_destroy_srq(srq);
2580
2581 err_put:
2582         if (cmd->srq_type == IB_SRQT_XRC) {
2583                 atomic_dec(&obj->uxrcd->refcnt);
2584                 put_uobj_read(xrcd_uobj);
2585         }
2586
2587 err_put_cq:
2588         if (cmd->srq_type == IB_SRQT_XRC)
2589                 put_cq_read(attr.ext.xrc.cq);
2590
2591 err_put_pd:
2592         put_pd_read(pd);
2593
2594 err:
2595         put_uobj_write(&obj->uevent.uobject);
2596         return ret;
2597 }
2598
2599 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
2600                              const char __user *buf, int in_len,
2601                              int out_len)
2602 {
2603         struct ib_uverbs_create_srq      cmd;
2604         struct ib_uverbs_create_xsrq     xcmd;
2605         struct ib_uverbs_create_srq_resp resp;
2606         struct ib_udata                  udata;
2607         int ret;
2608
2609         if (out_len < sizeof resp)
2610                 return -ENOSPC;
2611
2612         if (copy_from_user(&cmd, buf, sizeof cmd))
2613                 return -EFAULT;
2614
2615         xcmd.response    = cmd.response;
2616         xcmd.user_handle = cmd.user_handle;
2617         xcmd.srq_type    = IB_SRQT_BASIC;
2618         xcmd.pd_handle   = cmd.pd_handle;
2619         xcmd.max_wr      = cmd.max_wr;
2620         xcmd.max_sge     = cmd.max_sge;
2621         xcmd.srq_limit   = cmd.srq_limit;
2622
2623         INIT_UDATA(&udata, buf + sizeof cmd,
2624                    (unsigned long) cmd.response + sizeof resp,
2625                    in_len - sizeof cmd, out_len - sizeof resp);
2626
2627         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
2628         if (ret)
2629                 return ret;
2630
2631         return in_len;
2632 }
2633
2634 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
2635                               const char __user *buf, int in_len, int out_len)
2636 {
2637         struct ib_uverbs_create_xsrq     cmd;
2638         struct ib_uverbs_create_srq_resp resp;
2639         struct ib_udata                  udata;
2640         int ret;
2641
2642         if (out_len < sizeof resp)
2643                 return -ENOSPC;
2644
2645         if (copy_from_user(&cmd, buf, sizeof cmd))
2646                 return -EFAULT;
2647
2648         INIT_UDATA(&udata, buf + sizeof cmd,
2649                    (unsigned long) cmd.response + sizeof resp,
2650                    in_len - sizeof cmd, out_len - sizeof resp);
2651
2652         ret = __uverbs_create_xsrq(file, &cmd, &udata);
2653         if (ret)
2654                 return ret;
2655
2656         return in_len;
2657 }
2658
2659 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
2660                              const char __user *buf, int in_len,
2661                              int out_len)
2662 {
2663         struct ib_uverbs_modify_srq cmd;
2664         struct ib_udata             udata;
2665         struct ib_srq              *srq;
2666         struct ib_srq_attr          attr;
2667         int                         ret;
2668
2669         if (copy_from_user(&cmd, buf, sizeof cmd))
2670                 return -EFAULT;
2671
2672         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
2673                    out_len);
2674
2675         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2676         if (!srq)
2677                 return -EINVAL;
2678
2679         attr.max_wr    = cmd.max_wr;
2680         attr.srq_limit = cmd.srq_limit;
2681
2682         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
2683
2684         put_srq_read(srq);
2685
2686         return ret ? ret : in_len;
2687 }
2688
2689 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
2690                             const char __user *buf,
2691                             int in_len, int out_len)
2692 {
2693         struct ib_uverbs_query_srq      cmd;
2694         struct ib_uverbs_query_srq_resp resp;
2695         struct ib_srq_attr              attr;
2696         struct ib_srq                   *srq;
2697         int                             ret;
2698
2699         if (out_len < sizeof resp)
2700                 return -ENOSPC;
2701
2702         if (copy_from_user(&cmd, buf, sizeof cmd))
2703                 return -EFAULT;
2704
2705         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2706         if (!srq)
2707                 return -EINVAL;
2708
2709         ret = ib_query_srq(srq, &attr);
2710
2711         put_srq_read(srq);
2712
2713         if (ret)
2714                 return ret;
2715
2716         memset(&resp, 0, sizeof resp);
2717
2718         resp.max_wr    = attr.max_wr;
2719         resp.max_sge   = attr.max_sge;
2720         resp.srq_limit = attr.srq_limit;
2721
2722         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2723                          &resp, sizeof resp))
2724                 return -EFAULT;
2725
2726         return in_len;
2727 }
2728
2729 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
2730                               const char __user *buf, int in_len,
2731                               int out_len)
2732 {
2733         struct ib_uverbs_destroy_srq      cmd;
2734         struct ib_uverbs_destroy_srq_resp resp;
2735         struct ib_uobject                *uobj;
2736         struct ib_srq                    *srq;
2737         struct ib_uevent_object          *obj;
2738         int                               ret = -EINVAL;
2739
2740         if (copy_from_user(&cmd, buf, sizeof cmd))
2741                 return -EFAULT;
2742
2743         uobj = idr_write_uobj(&ib_uverbs_srq_idr, cmd.srq_handle, file->ucontext);
2744         if (!uobj)
2745                 return -EINVAL;
2746         srq = uobj->object;
2747         obj = container_of(uobj, struct ib_uevent_object, uobject);
2748
2749         ret = ib_destroy_srq(srq);
2750         if (!ret)
2751                 uobj->live = 0;
2752
2753         put_uobj_write(uobj);
2754
2755         if (ret)
2756                 return ret;
2757
2758         idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
2759
2760         mutex_lock(&file->mutex);
2761         list_del(&uobj->list);
2762         mutex_unlock(&file->mutex);
2763
2764         ib_uverbs_release_uevent(file, obj);
2765
2766         memset(&resp, 0, sizeof resp);
2767         resp.events_reported = obj->events_reported;
2768
2769         put_uobj(uobj);
2770
2771         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2772                          &resp, sizeof resp))
2773                 ret = -EFAULT;
2774
2775         return ret ? ret : in_len;
2776 }