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 struct uverbs_lock_class {
  45         struct lock_class_key   key;
  46         char                    name[16];
  47 };
  48
  49 static struct uverbs_lock_class pd_lock_class   = { .name = "PD-uobj" };
  50 static struct uverbs_lock_class mr_lock_class   = { .name = "MR-uobj" };
  51 static struct uverbs_lock_class cq_lock_class   = { .name = "CQ-uobj" };
  52 static struct uverbs_lock_class qp_lock_class   = { .name = "QP-uobj" };
  53 static struct uverbs_lock_class ah_lock_class   = { .name = "AH-uobj" };
  54 static struct uverbs_lock_class srq_lock_class  = { .name = "SRQ-uobj" };
  55 static struct uverbs_lock_class xrcd_lock_class = { .name = "XRCD-uobj" };
  56
  57 #define INIT_UDATA(udata, ibuf, obuf, ilen, olen)                       \
  58         do {                                                            \
  59                 (udata)->inbuf  = (void __user *) (ibuf);               \
  60                 (udata)->outbuf = (void __user *) (obuf);               \
  61                 (udata)->inlen  = (ilen);                               \
  62                 (udata)->outlen = (olen);                               \
  63         } while (0)
  64
  65 /*
  66  * The ib_uobject locking scheme is as follows:
  67  *
  68  * - ib_uverbs_idr_lock protects the uverbs idrs themselves, so it
  69  *   needs to be held during all idr operations.  When an object is
  70  *   looked up, a reference must be taken on the object's kref before
  71  *   dropping this lock.
  72  *
  73  * - Each object also has an rwsem.  This rwsem must be held for
  74  *   reading while an operation that uses the object is performed.
  75  *   For example, while registering an MR, the associated PD's
  76  *   uobject.mutex must be held for reading.  The rwsem must be held
  77  *   for writing while initializing or destroying an object.
  78  *
  79  * - In addition, each object has a "live" flag.  If this flag is not
  80  *   set, then lookups of the object will fail even if it is found in
  81  *   the idr.  This handles a reader that blocks and does not acquire
  82  *   the rwsem until after the object is destroyed.  The destroy
  83  *   operation will set the live flag to 0 and then drop the rwsem;
  84  *   this will allow the reader to acquire the rwsem, see that the
  85  *   live flag is 0, and then drop the rwsem and its reference to
  86  *   object.  The underlying storage will not be freed until the last
  87  *   reference to the object is dropped.
  88  */
  89
  90 static void init_uobj(struct ib_uobject *uobj, u64 user_handle,
  91                       struct ib_ucontext *context, struct uverbs_lock_class *c)
  92 {
  93         uobj->user_handle = user_handle;
  94         uobj->context     = context;
  95         kref_init(&uobj->ref);
  96         init_rwsem(&uobj->mutex);
  97         lockdep_set_class_and_name(&uobj->mutex, &c->key, c->name);
  98         uobj->live        = 0;
  99 }
 100
 101 static void release_uobj(struct kref *kref)
 102 {
 103         kfree(container_of(kref, struct ib_uobject, ref));
 104 }
 105
 106 static void put_uobj(struct ib_uobject *uobj)
 107 {
 108         kref_put(&uobj->ref, release_uobj);
 109 }
 110
 111 static void put_uobj_read(struct ib_uobject *uobj)
 112 {
 113         up_read(&uobj->mutex);
 114         put_uobj(uobj);
 115 }
 116
 117 static void put_uobj_write(struct ib_uobject *uobj)
 118 {
 119         up_write(&uobj->mutex);
 120         put_uobj(uobj);
 121 }
 122
 123 static int idr_add_uobj(struct idr *idr, struct ib_uobject *uobj)
 124 {
 125         int ret;
 126
 127 retry:
 128         if (!idr_pre_get(idr, GFP_KERNEL))
 129                 return -ENOMEM;
 130
 131         spin_lock(&ib_uverbs_idr_lock);
 132         ret = idr_get_new(idr, uobj, &uobj->id);
 133         spin_unlock(&ib_uverbs_idr_lock);
 134
 135         if (ret == -EAGAIN)
 136                 goto retry;
 137
 138         return ret;
 139 }
 140
 141 void idr_remove_uobj(struct idr *idr, struct ib_uobject *uobj)
 142 {
 143         spin_lock(&ib_uverbs_idr_lock);
 144         idr_remove(idr, uobj->id);
 145         spin_unlock(&ib_uverbs_idr_lock);
 146 }
 147
 148 static struct ib_uobject *__idr_get_uobj(struct idr *idr, int id,
 149                                          struct ib_ucontext *context)
 150 {
 151         struct ib_uobject *uobj;
 152
 153         spin_lock(&ib_uverbs_idr_lock);
 154         uobj = idr_find(idr, id);
 155         if (uobj) {
 156                 if (uobj->context == context)
 157                         kref_get(&uobj->ref);
 158                 else
 159                         uobj = NULL;
 160         }
 161         spin_unlock(&ib_uverbs_idr_lock);
 162
 163         return uobj;
 164 }
 165
 166 static struct ib_uobject *idr_read_uobj(struct idr *idr, int id,
 167                                         struct ib_ucontext *context, int nested)
 168 {
 169         struct ib_uobject *uobj;
 170
 171         uobj = __idr_get_uobj(idr, id, context);
 172         if (!uobj)
 173                 return NULL;
 174
 175         if (nested)
 176                 down_read_nested(&uobj->mutex, SINGLE_DEPTH_NESTING);
 177         else
 178                 down_read(&uobj->mutex);
 179         if (!uobj->live) {
 180                 put_uobj_read(uobj);
 181                 return NULL;
 182         }
 183
 184         return uobj;
 185 }
 186
 187 static struct ib_uobject *idr_write_uobj(struct idr *idr, int id,
 188                                          struct ib_ucontext *context)
 189 {
 190         struct ib_uobject *uobj;
 191
 192         uobj = __idr_get_uobj(idr, id, context);
 193         if (!uobj)
 194                 return NULL;
 195
 196         down_write(&uobj->mutex);
 197         if (!uobj->live) {
 198                 put_uobj_write(uobj);
 199                 return NULL;
 200         }
 201
 202         return uobj;
 203 }
 204
 205 static void *idr_read_obj(struct idr *idr, int id, struct ib_ucontext *context,
 206                           int nested)
 207 {
 208         struct ib_uobject *uobj;
 209
 210         uobj = idr_read_uobj(idr, id, context, nested);
 211         return uobj ? uobj->object : NULL;
 212 }
 213
 214 static struct ib_pd *idr_read_pd(int pd_handle, struct ib_ucontext *context)
 215 {
 216         return idr_read_obj(&ib_uverbs_pd_idr, pd_handle, context, 0);
 217 }
 218
 219 static void put_pd_read(struct ib_pd *pd)
 220 {
 221         put_uobj_read(pd->uobject);
 222 }
 223
 224 static struct ib_cq *idr_read_cq(int cq_handle, struct ib_ucontext *context, int nested)
 225 {
 226         return idr_read_obj(&ib_uverbs_cq_idr, cq_handle, context, nested);
 227 }
 228
 229 static void put_cq_read(struct ib_cq *cq)
 230 {
 231         put_uobj_read(cq->uobject);
 232 }
 233
 234 static struct ib_ah *idr_read_ah(int ah_handle, struct ib_ucontext *context)
 235 {
 236         return idr_read_obj(&ib_uverbs_ah_idr, ah_handle, context, 0);
 237 }
 238
 239 static void put_ah_read(struct ib_ah *ah)
 240 {
 241         put_uobj_read(ah->uobject);
 242 }
 243
 244 static struct ib_qp *idr_read_qp(int qp_handle, struct ib_ucontext *context)
 245 {
 246         return idr_read_obj(&ib_uverbs_qp_idr, qp_handle, context, 0);
 247 }
 248
 249 static struct ib_qp *idr_write_qp(int qp_handle, struct ib_ucontext *context)
 250 {
 251         struct ib_uobject *uobj;
 252
 253         uobj = idr_write_uobj(&ib_uverbs_qp_idr, qp_handle, context);
 254         return uobj ? uobj->object : NULL;
 255 }
 256
 257 static void put_qp_read(struct ib_qp *qp)
 258 {
 259         put_uobj_read(qp->uobject);
 260 }
 261
 262 static void put_qp_write(struct ib_qp *qp)
 263 {
 264         put_uobj_write(qp->uobject);
 265 }
 266
 267 static struct ib_srq *idr_read_srq(int srq_handle, struct ib_ucontext *context)
 268 {
 269         return idr_read_obj(&ib_uverbs_srq_idr, srq_handle, context, 0);
 270 }
 271
 272 static void put_srq_read(struct ib_srq *srq)
 273 {
 274         put_uobj_read(srq->uobject);
 275 }
 276
 277 static struct ib_xrcd *idr_read_xrcd(int xrcd_handle, struct ib_ucontext *context,
 278                                      struct ib_uobject **uobj)
 279 {
 280         *uobj = idr_read_uobj(&ib_uverbs_xrcd_idr, xrcd_handle, context, 0);
 281         return *uobj ? (*uobj)->object : NULL;
 282 }
 283
 284 static void put_xrcd_read(struct ib_uobject *uobj)
 285 {
 286         put_uobj_read(uobj);
 287 }
 288
 289 ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
 290                               const char __user *buf,
 291                               int in_len, int out_len)
 292 {
 293         struct ib_uverbs_get_context      cmd;
 294         struct ib_uverbs_get_context_resp resp;
 295         struct ib_udata                   udata;
 296         struct ib_device                 *ibdev = file->device->ib_dev;
 297         struct ib_ucontext               *ucontext;
 298         struct file                      *filp;
 299         int ret;
 300
 301         if (out_len < sizeof resp)
 302                 return -ENOSPC;
 303
 304         if (copy_from_user(&cmd, buf, sizeof cmd))
 305                 return -EFAULT;
 306
 307         mutex_lock(&file->mutex);
 308
 309         if (file->ucontext) {
 310                 ret = -EINVAL;
 311                 goto err;
 312         }
 313
 314         INIT_UDATA(&udata, buf + sizeof cmd,
 315                    (unsigned long) cmd.response + sizeof resp,
 316                    in_len - sizeof cmd, out_len - sizeof resp);
 317
 318         ucontext = ibdev->alloc_ucontext(ibdev, &udata);
 319         if (IS_ERR(ucontext)) {
 320                 ret = PTR_ERR(ucontext);
 321                 goto err;
 322         }
 323
 324         ucontext->device = ibdev;
 325         INIT_LIST_HEAD(&ucontext->pd_list);
 326         INIT_LIST_HEAD(&ucontext->mr_list);
 327         INIT_LIST_HEAD(&ucontext->mw_list);
 328         INIT_LIST_HEAD(&ucontext->cq_list);
 329         INIT_LIST_HEAD(&ucontext->qp_list);
 330         INIT_LIST_HEAD(&ucontext->srq_list);
 331         INIT_LIST_HEAD(&ucontext->ah_list);
 332         INIT_LIST_HEAD(&ucontext->xrcd_list);
 333         ucontext->closing = 0;
 334
 335         resp.num_comp_vectors = file->device->num_comp_vectors;
 336
 337         ret = get_unused_fd();
 338         if (ret < 0)
 339                 goto err_free;
 340         resp.async_fd = ret;
 341
 342         filp = ib_uverbs_alloc_event_file(file, 1);
 343         if (IS_ERR(filp)) {
 344                 ret = PTR_ERR(filp);
 345                 goto err_fd;
 346         }
 347
 348         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 349                          &resp, sizeof resp)) {
 350                 ret = -EFAULT;
 351                 goto err_file;
 352         }
 353
 354         file->async_file = filp->private_data;
 355
 356         INIT_IB_EVENT_HANDLER(&file->event_handler, file->device->ib_dev,
 357                               ib_uverbs_event_handler);
 358         ret = ib_register_event_handler(&file->event_handler);
 359         if (ret)
 360                 goto err_file;
 361
 362         kref_get(&file->async_file->ref);
 363         kref_get(&file->ref);
 364         file->ucontext = ucontext;
 365
 366         fd_install(resp.async_fd, filp);
 367
 368         mutex_unlock(&file->mutex);
 369
 370         return in_len;
 371
 372 err_file:
 373         fput(filp);
 374
 375 err_fd:
 376         put_unused_fd(resp.async_fd);
 377
 378 err_free:
 379         ibdev->dealloc_ucontext(ucontext);
 380
 381 err:
 382         mutex_unlock(&file->mutex);
 383         return ret;
 384 }
 385
 386 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 387                                const char __user *buf,
 388                                int in_len, int out_len)
 389 {
 390         struct ib_uverbs_query_device      cmd;
 391         struct ib_uverbs_query_device_resp resp;
 392         struct ib_device_attr              attr;
 393         int                                ret;
 394
 395         if (out_len < sizeof resp)
 396                 return -ENOSPC;
 397
 398         if (copy_from_user(&cmd, buf, sizeof cmd))
 399                 return -EFAULT;
 400
 401         ret = ib_query_device(file->device->ib_dev, &attr);
 402         if (ret)
 403                 return ret;
 404
 405         memset(&resp, 0, sizeof resp);
 406
 407         resp.fw_ver                    = attr.fw_ver;
 408         resp.node_guid                 = file->device->ib_dev->node_guid;
 409         resp.sys_image_guid            = attr.sys_image_guid;
 410         resp.max_mr_size               = attr.max_mr_size;
 411         resp.page_size_cap             = attr.page_size_cap;
 412         resp.vendor_id                 = attr.vendor_id;
 413         resp.vendor_part_id            = attr.vendor_part_id;
 414         resp.hw_ver                    = attr.hw_ver;
 415         resp.max_qp                    = attr.max_qp;
 416         resp.max_qp_wr                 = attr.max_qp_wr;
 417         resp.device_cap_flags          = attr.device_cap_flags;
 418         resp.max_sge                   = attr.max_sge;
 419         resp.max_sge_rd                = attr.max_sge_rd;
 420         resp.max_cq                    = attr.max_cq;
 421         resp.max_cqe                   = attr.max_cqe;
 422         resp.max_mr                    = attr.max_mr;
 423         resp.max_pd                    = attr.max_pd;
 424         resp.max_qp_rd_atom            = attr.max_qp_rd_atom;
 425         resp.max_ee_rd_atom            = attr.max_ee_rd_atom;
 426         resp.max_res_rd_atom           = attr.max_res_rd_atom;
 427         resp.max_qp_init_rd_atom       = attr.max_qp_init_rd_atom;
 428         resp.max_ee_init_rd_atom       = attr.max_ee_init_rd_atom;
 429         resp.atomic_cap                = attr.atomic_cap;
 430         resp.max_ee                    = attr.max_ee;
 431         resp.max_rdd                   = attr.max_rdd;
 432         resp.max_mw                    = attr.max_mw;
 433         resp.max_raw_ipv6_qp           = attr.max_raw_ipv6_qp;
 434         resp.max_raw_ethy_qp           = attr.max_raw_ethy_qp;
 435         resp.max_mcast_grp             = attr.max_mcast_grp;
 436         resp.max_mcast_qp_attach       = attr.max_mcast_qp_attach;
 437         resp.max_total_mcast_qp_attach = attr.max_total_mcast_qp_attach;
 438         resp.max_ah                    = attr.max_ah;
 439         resp.max_fmr                   = attr.max_fmr;
 440         resp.max_map_per_fmr           = attr.max_map_per_fmr;
 441         resp.max_srq                   = attr.max_srq;
 442         resp.max_srq_wr                = attr.max_srq_wr;
 443         resp.max_srq_sge               = attr.max_srq_sge;
 444         resp.max_pkeys                 = attr.max_pkeys;
 445         resp.local_ca_ack_delay        = attr.local_ca_ack_delay;
 446         resp.phys_port_cnt             = file->device->ib_dev->phys_port_cnt;
 447
 448         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 449                          &resp, sizeof resp))
 450                 return -EFAULT;
 451
 452         return in_len;
 453 }
 454
 455 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 456                              const char __user *buf,
 457                              int in_len, int out_len)
 458 {
 459         struct ib_uverbs_query_port      cmd;
 460         struct ib_uverbs_query_port_resp resp;
 461         struct ib_port_attr              attr;
 462         int                              ret;
 463
 464         if (out_len < sizeof resp)
 465                 return -ENOSPC;
 466
 467         if (copy_from_user(&cmd, buf, sizeof cmd))
 468                 return -EFAULT;
 469
 470         ret = ib_query_port(file->device->ib_dev, cmd.port_num, &attr);
 471         if (ret)
 472                 return ret;
 473
 474         memset(&resp, 0, sizeof resp);
 475
 476         resp.state           = attr.state;
 477         resp.max_mtu         = attr.max_mtu;
 478         resp.active_mtu      = attr.active_mtu;
 479         resp.gid_tbl_len     = attr.gid_tbl_len;
 480         resp.port_cap_flags  = attr.port_cap_flags;
 481         resp.max_msg_sz      = attr.max_msg_sz;
 482         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 483         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 484         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 485         resp.lid             = attr.lid;
 486         resp.sm_lid          = attr.sm_lid;
 487         resp.lmc             = attr.lmc;
 488         resp.max_vl_num      = attr.max_vl_num;
 489         resp.sm_sl           = attr.sm_sl;
 490         resp.subnet_timeout  = attr.subnet_timeout;
 491         resp.init_type_reply = attr.init_type_reply;
 492         resp.active_width    = attr.active_width;
 493         resp.active_speed    = attr.active_speed;
 494         resp.phys_state      = attr.phys_state;
 495         resp.link_layer      = rdma_port_get_link_layer(file->device->ib_dev,
 496                                                         cmd.port_num);
 497
 498         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 499                          &resp, sizeof resp))
 500                 return -EFAULT;
 501
 502         return in_len;
 503 }
 504
 505 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 506                            const char __user *buf,
 507                            int in_len, int out_len)
 508 {
 509         struct ib_uverbs_alloc_pd      cmd;
 510         struct ib_uverbs_alloc_pd_resp resp;
 511         struct ib_udata                udata;
 512         struct ib_uobject             *uobj;
 513         struct ib_pd                  *pd;
 514         int                            ret;
 515
 516         if (out_len < sizeof resp)
 517                 return -ENOSPC;
 518
 519         if (copy_from_user(&cmd, buf, sizeof cmd))
 520                 return -EFAULT;
 521
 522         INIT_UDATA(&udata, buf + sizeof cmd,
 523                    (unsigned long) cmd.response + sizeof resp,
 524                    in_len - sizeof cmd, out_len - sizeof resp);
 525
 526         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 527         if (!uobj)
 528                 return -ENOMEM;
 529
 530         init_uobj(uobj, 0, file->ucontext, &pd_lock_class);
 531         down_write(&uobj->mutex);
 532
 533         pd = file->device->ib_dev->alloc_pd(file->device->ib_dev,
 534                                             file->ucontext, &udata);
 535         if (IS_ERR(pd)) {
 536                 ret = PTR_ERR(pd);
 537                 goto err;
 538         }
 539
 540         pd->device  = file->device->ib_dev;
 541         pd->uobject = uobj;
 542         atomic_set(&pd->usecnt, 0);
 543
 544         uobj->object = pd;
 545         ret = idr_add_uobj(&ib_uverbs_pd_idr, uobj);
 546         if (ret)
 547                 goto err_idr;
 548
 549         memset(&resp, 0, sizeof resp);
 550         resp.pd_handle = uobj->id;
 551
 552         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 553                          &resp, sizeof resp)) {
 554                 ret = -EFAULT;
 555                 goto err_copy;
 556         }
 557
 558         mutex_lock(&file->mutex);
 559         list_add_tail(&uobj->list, &file->ucontext->pd_list);
 560         mutex_unlock(&file->mutex);
 561
 562         uobj->live = 1;
 563
 564         up_write(&uobj->mutex);
 565
 566         return in_len;
 567
 568 err_copy:
 569         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 570
 571 err_idr:
 572         ib_dealloc_pd(pd);
 573
 574 err:
 575         put_uobj_write(uobj);
 576         return ret;
 577 }
 578
 579 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 580                              const char __user *buf,
 581                              int in_len, int out_len)
 582 {
 583         struct ib_uverbs_dealloc_pd cmd;
 584         struct ib_uobject          *uobj;
 585         int                         ret;
 586
 587         if (copy_from_user(&cmd, buf, sizeof cmd))
 588                 return -EFAULT;
 589
 590         uobj = idr_write_uobj(&ib_uverbs_pd_idr, cmd.pd_handle, file->ucontext);
 591         if (!uobj)
 592                 return -EINVAL;
 593
 594         ret = ib_dealloc_pd(uobj->object);
 595         if (!ret)
 596                 uobj->live = 0;
 597
 598         put_uobj_write(uobj);
 599
 600         if (ret)
 601                 return ret;
 602
 603         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 604
 605         mutex_lock(&file->mutex);
 606         list_del(&uobj->list);
 607         mutex_unlock(&file->mutex);
 608
 609         put_uobj(uobj);
 610
 611         return in_len;
 612 }
 613
 614 struct xrcd_table_entry {
 615         struct rb_node  node;
 616         struct ib_xrcd *xrcd;
 617         struct inode   *inode;
 618 };
 619
 620 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 621                             struct inode *inode,
 622                             struct ib_xrcd *xrcd)
 623 {
 624         struct xrcd_table_entry *entry, *scan;
 625         struct rb_node **p = &dev->xrcd_tree.rb_node;
 626         struct rb_node *parent = NULL;
 627
 628         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 629         if (!entry)
 630                 return -ENOMEM;
 631
 632         entry->xrcd  = xrcd;
 633         entry->inode = inode;
 634
 635         while (*p) {
 636                 parent = *p;
 637                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 638
 639                 if (inode < scan->inode) {
 640                         p = &(*p)->rb_left;
 641                 } else if (inode > scan->inode) {
 642                         p = &(*p)->rb_right;
 643                 } else {
 644                         kfree(entry);
 645                         return -EEXIST;
 646                 }
 647         }
 648
 649         rb_link_node(&entry->node, parent, p);
 650         rb_insert_color(&entry->node, &dev->xrcd_tree);
 651         igrab(inode);
 652         return 0;
 653 }
 654
 655 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 656                                                   struct inode *inode)
 657 {
 658         struct xrcd_table_entry *entry;
 659         struct rb_node *p = dev->xrcd_tree.rb_node;
 660
 661         while (p) {
 662                 entry = rb_entry(p, struct xrcd_table_entry, node);
 663
 664                 if (inode < entry->inode)
 665                         p = p->rb_left;
 666                 else if (inode > entry->inode)
 667                         p = p->rb_right;
 668                 else
 669                         return entry;
 670         }
 671
 672         return NULL;
 673 }
 674
 675 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 676 {
 677         struct xrcd_table_entry *entry;
 678
 679         entry = xrcd_table_search(dev, inode);
 680         if (!entry)
 681                 return NULL;
 682
 683         return entry->xrcd;
 684 }
 685
 686 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 687                               struct inode *inode)
 688 {
 689         struct xrcd_table_entry *entry;
 690
 691         entry = xrcd_table_search(dev, inode);
 692         if (entry) {
 693                 iput(inode);
 694                 rb_erase(&entry->node, &dev->xrcd_tree);
 695                 kfree(entry);
 696         }
 697 }
 698
 699 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 700                             const char __user *buf, int in_len,
 701                             int out_len)
 702 {
 703         struct ib_uverbs_open_xrcd      cmd;
 704         struct ib_uverbs_open_xrcd_resp resp;
 705         struct ib_udata                 udata;
 706         struct ib_uxrcd_object         *obj;
 707         struct ib_xrcd                 *xrcd = NULL;
 708         struct file                    *f = NULL;
 709         struct inode                   *inode = NULL;
 710         int                             ret = 0;
 711         int                             new_xrcd = 0;
 712
 713         if (out_len < sizeof resp)
 714                 return -ENOSPC;
 715
 716         if (copy_from_user(&cmd, buf, sizeof cmd))
 717                 return -EFAULT;
 718
 719         INIT_UDATA(&udata, buf + sizeof cmd,
 720                    (unsigned long) cmd.response + sizeof resp,
 721                    in_len - sizeof cmd, out_len - sizeof  resp);
 722
 723         mutex_lock(&file->device->xrcd_tree_mutex);
 724
 725         if (cmd.fd != -1) {
 726                 /* search for file descriptor */
 727                 f = fget(cmd.fd);
 728                 if (!f) {
 729                         ret = -EBADF;
 730                         goto err_tree_mutex_unlock;
 731                 }
 732
 733                 inode = f->f_dentry->d_inode;
 734                 if (!inode) {
 735                         ret = -EBADF;
 736                         goto err_tree_mutex_unlock;
 737                 }
 738
 739                 xrcd = find_xrcd(file->device, inode);
 740                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 741                         /* no file descriptor. Need CREATE flag */
 742                         ret = -EAGAIN;
 743                         goto err_tree_mutex_unlock;
 744                 }
 745
 746                 if (xrcd && cmd.oflags & O_EXCL) {
 747                         ret = -EINVAL;
 748                         goto err_tree_mutex_unlock;
 749                 }
 750         }
 751
 752         obj = kmalloc(sizeof *obj, GFP_KERNEL);
 753         if (!obj) {
 754                 ret = -ENOMEM;
 755                 goto err_tree_mutex_unlock;
 756         }
 757
 758         init_uobj(&obj->uobject, 0, file->ucontext, &xrcd_lock_class);
 759
 760         down_write(&obj->uobject.mutex);
 761
 762         if (!xrcd) {
 763                 xrcd = file->device->ib_dev->alloc_xrcd(file->device->ib_dev,
 764                                                         file->ucontext, &udata);
 765                 if (IS_ERR(xrcd)) {
 766                         ret = PTR_ERR(xrcd);
 767                         goto err;
 768                 }
 769
 770                 xrcd->inode   = inode;
 771                 xrcd->device  = file->device->ib_dev;
 772                 atomic_set(&xrcd->usecnt, 0);
 773                 mutex_init(&xrcd->tgt_qp_mutex);
 774                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 775                 new_xrcd = 1;
 776         }
 777
 778         atomic_set(&obj->refcnt, 0);
 779         obj->uobject.object = xrcd;
 780         ret = idr_add_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 781         if (ret)
 782                 goto err_idr;
 783
 784         memset(&resp, 0, sizeof resp);
 785         resp.xrcd_handle = obj->uobject.id;
 786
 787         if (inode) {
 788                 if (new_xrcd) {
 789                         /* create new inode/xrcd table entry */
 790                         ret = xrcd_table_insert(file->device, inode, xrcd);
 791                         if (ret)
 792                                 goto err_insert_xrcd;
 793                 }
 794                 atomic_inc(&xrcd->usecnt);
 795         }
 796
 797         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 798                          &resp, sizeof resp)) {
 799                 ret = -EFAULT;
 800                 goto err_copy;
 801         }
 802
 803         if (f)
 804                 fput(f);
 805
 806         mutex_lock(&file->mutex);
 807         list_add_tail(&obj->uobject.list, &file->ucontext->xrcd_list);
 808         mutex_unlock(&file->mutex);
 809
 810         obj->uobject.live = 1;
 811         up_write(&obj->uobject.mutex);
 812
 813         mutex_unlock(&file->device->xrcd_tree_mutex);
 814         return in_len;
 815
 816 err_copy:
 817         if (inode) {
 818                 if (new_xrcd)
 819                         xrcd_table_delete(file->device, inode);
 820                 atomic_dec(&xrcd->usecnt);
 821         }
 822
 823 err_insert_xrcd:
 824         idr_remove_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 825
 826 err_idr:
 827         ib_dealloc_xrcd(xrcd);
 828
 829 err:
 830         put_uobj_write(&obj->uobject);
 831
 832 err_tree_mutex_unlock:
 833         if (f)
 834                 fput(f);
 835
 836         mutex_unlock(&file->device->xrcd_tree_mutex);
 837
 838         return ret;
 839 }
 840
 841 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 842                              const char __user *buf, int in_len,
 843                              int out_len)
 844 {
 845         struct ib_uverbs_close_xrcd cmd;
 846         struct ib_uobject           *uobj;
 847         struct ib_xrcd              *xrcd = NULL;
 848         struct inode                *inode = NULL;
 849         struct ib_uxrcd_object      *obj;
 850         int                         live;
 851         int                         ret = 0;
 852
 853         if (copy_from_user(&cmd, buf, sizeof cmd))
 854                 return -EFAULT;
 855
 856         mutex_lock(&file->device->xrcd_tree_mutex);
 857         uobj = idr_write_uobj(&ib_uverbs_xrcd_idr, cmd.xrcd_handle, file->ucontext);
 858         if (!uobj) {
 859                 ret = -EINVAL;
 860                 goto out;
 861         }
 862
 863         xrcd  = uobj->object;
 864         inode = xrcd->inode;
 865         obj   = container_of(uobj, struct ib_uxrcd_object, uobject);
 866         if (atomic_read(&obj->refcnt)) {
 867                 put_uobj_write(uobj);
 868                 ret = -EBUSY;
 869                 goto out;
 870         }
 871
 872         if (!inode || atomic_dec_and_test(&xrcd->usecnt)) {
 873                 ret = ib_dealloc_xrcd(uobj->object);
 874                 if (!ret)
 875                         uobj->live = 0;
 876         }
 877
 878         live = uobj->live;
 879         if (inode && ret)
 880                 atomic_inc(&xrcd->usecnt);
 881
 882         put_uobj_write(uobj);
 883
 884         if (ret)
 885                 goto out;
 886
 887         if (inode && !live)
 888                 xrcd_table_delete(file->device, inode);
 889
 890         idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 891         mutex_lock(&file->mutex);
 892         list_del(&uobj->list);
 893         mutex_unlock(&file->mutex);
 894
 895         put_uobj(uobj);
 896         ret = in_len;
 897
 898 out:
 899         mutex_unlock(&file->device->xrcd_tree_mutex);
 900         return ret;
 901 }
 902
 903 void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 904                             struct ib_xrcd *xrcd)
 905 {
 906         struct inode *inode;
 907
 908         inode = xrcd->inode;
 909         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 910                 return;
 911
 912         ib_dealloc_xrcd(xrcd);
 913
 914         if (inode)
 915                 xrcd_table_delete(dev, inode);
 916 }
 917
 918 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 919                          const char __user *buf, int in_len,
 920                          int out_len)
 921 {
 922         struct ib_uverbs_reg_mr      cmd;
 923         struct ib_uverbs_reg_mr_resp resp;
 924         struct ib_udata              udata;
 925         struct ib_uobject           *uobj;
 926         struct ib_pd                *pd;
 927         struct ib_mr                *mr;
 928         int                          ret;
 929
 930         if (out_len < sizeof resp)
 931                 return -ENOSPC;
 932
 933         if (copy_from_user(&cmd, buf, sizeof cmd))
 934                 return -EFAULT;
 935
 936         INIT_UDATA(&udata, buf + sizeof cmd,
 937                    (unsigned long) cmd.response + sizeof resp,
 938                    in_len - sizeof cmd, out_len - sizeof resp);
 939
 940         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 941                 return -EINVAL;
 942
 943         /*
 944          * Local write permission is required if remote write or
 945          * remote atomic permission is also requested.
 946          */
 947         if (cmd.access_flags & (IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_REMOTE_WRITE) &&
 948             !(cmd.access_flags & IB_ACCESS_LOCAL_WRITE))
 949                 return -EINVAL;
 950
 951         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 952         if (!uobj)
 953                 return -ENOMEM;
 954
 955         init_uobj(uobj, 0, file->ucontext, &mr_lock_class);
 956         down_write(&uobj->mutex);
 957
 958         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
 959         if (!pd) {
 960                 ret = -EINVAL;
 961                 goto err_free;
 962         }
 963
 964         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 965                                      cmd.access_flags, &udata);
 966         if (IS_ERR(mr)) {
 967                 ret = PTR_ERR(mr);
 968                 goto err_put;
 969         }
 970
 971         mr->device  = pd->device;
 972         mr->pd      = pd;
 973         mr->uobject = uobj;
 974         atomic_inc(&pd->usecnt);
 975         atomic_set(&mr->usecnt, 0);
 976
 977         uobj->object = mr;
 978         ret = idr_add_uobj(&ib_uverbs_mr_idr, uobj);
 979         if (ret)
 980                 goto err_unreg;
 981
 982         memset(&resp, 0, sizeof resp);
 983         resp.lkey      = mr->lkey;
 984         resp.rkey      = mr->rkey;
 985         resp.mr_handle = uobj->id;
 986
 987         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 988                          &resp, sizeof resp)) {
 989                 ret = -EFAULT;
 990                 goto err_copy;
 991         }
 992
 993         put_pd_read(pd);
 994
 995         mutex_lock(&file->mutex);
 996         list_add_tail(&uobj->list, &file->ucontext->mr_list);
 997         mutex_unlock(&file->mutex);
 998
 999         uobj->live = 1;
1000
1001         up_write(&uobj->mutex);
1002
1003         return in_len;
1004
1005 err_copy:
1006         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1007
1008 err_unreg:
1009         ib_dereg_mr(mr);
1010
1011 err_put:
1012         put_pd_read(pd);
1013
1014 err_free:
1015         put_uobj_write(uobj);
1016         return ret;
1017 }
1018
1019 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
1020                            const char __user *buf, int in_len,
1021                            int out_len)
1022 {
1023         struct ib_uverbs_dereg_mr cmd;
1024         struct ib_mr             *mr;
1025         struct ib_uobject        *uobj;
1026         int                       ret = -EINVAL;
1027
1028         if (copy_from_user(&cmd, buf, sizeof cmd))
1029                 return -EFAULT;
1030
1031         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle, file->ucontext);
1032         if (!uobj)
1033                 return -EINVAL;
1034
1035         mr = uobj->object;
1036
1037         ret = ib_dereg_mr(mr);
1038         if (!ret)
1039                 uobj->live = 0;
1040
1041         put_uobj_write(uobj);
1042
1043         if (ret)
1044                 return ret;
1045
1046         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1047
1048         mutex_lock(&file->mutex);
1049         list_del(&uobj->list);
1050         mutex_unlock(&file->mutex);
1051
1052         put_uobj(uobj);
1053
1054         return in_len;
1055 }
1056
1057 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
1058                                       const char __user *buf, int in_len,
1059                                       int out_len)
1060 {
1061         struct ib_uverbs_create_comp_channel       cmd;
1062         struct ib_uverbs_create_comp_channel_resp  resp;
1063         struct file                               *filp;
1064         int ret;
1065
1066         if (out_len < sizeof resp)
1067                 return -ENOSPC;
1068
1069         if (copy_from_user(&cmd, buf, sizeof cmd))
1070                 return -EFAULT;
1071
1072         ret = get_unused_fd();
1073         if (ret < 0)
1074                 return ret;
1075         resp.fd = ret;
1076
1077         filp = ib_uverbs_alloc_event_file(file, 0);
1078         if (IS_ERR(filp)) {
1079                 put_unused_fd(resp.fd);
1080                 return PTR_ERR(filp);
1081         }
1082
1083         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1084                          &resp, sizeof resp)) {
1085                 put_unused_fd(resp.fd);
1086                 fput(filp);
1087                 return -EFAULT;
1088         }
1089
1090         fd_install(resp.fd, filp);
1091         return in_len;
1092 }
1093
1094 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1095                             const char __user *buf, int in_len,
1096                             int out_len)
1097 {
1098         struct ib_uverbs_create_cq      cmd;
1099         struct ib_uverbs_create_cq_resp resp;
1100         struct ib_udata                 udata;
1101         struct ib_ucq_object           *obj;
1102         struct ib_uverbs_event_file    *ev_file = NULL;
1103         struct ib_cq                   *cq;
1104         int                             ret;
1105
1106         if (out_len < sizeof resp)
1107                 return -ENOSPC;
1108
1109         if (copy_from_user(&cmd, buf, sizeof cmd))
1110                 return -EFAULT;
1111
1112         INIT_UDATA(&udata, buf + sizeof cmd,
1113                    (unsigned long) cmd.response + sizeof resp,
1114                    in_len - sizeof cmd, out_len - sizeof resp);
1115
1116         if (cmd.comp_vector >= file->device->num_comp_vectors)
1117                 return -EINVAL;
1118
1119         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1120         if (!obj)
1121                 return -ENOMEM;
1122
1123         init_uobj(&obj->uobject, cmd.user_handle, file->ucontext, &cq_lock_class);
1124         down_write(&obj->uobject.mutex);
1125
1126         if (cmd.comp_channel >= 0) {
1127                 ev_file = ib_uverbs_lookup_comp_file(cmd.comp_channel);
1128                 if (!ev_file) {
1129                         ret = -EINVAL;
1130                         goto err;
1131                 }
1132         }
1133
1134         obj->uverbs_file           = file;
1135         obj->comp_events_reported  = 0;
1136         obj->async_events_reported = 0;
1137         INIT_LIST_HEAD(&obj->comp_list);
1138         INIT_LIST_HEAD(&obj->async_list);
1139
1140         cq = file->device->ib_dev->create_cq(file->device->ib_dev, cmd.cqe,
1141                                              cmd.comp_vector,
1142                                              file->ucontext, &udata);
1143         if (IS_ERR(cq)) {
1144                 ret = PTR_ERR(cq);
1145                 goto err_file;
1146         }
1147
1148         cq->device        = file->device->ib_dev;
1149         cq->uobject       = &obj->uobject;
1150         cq->comp_handler  = ib_uverbs_comp_handler;
1151         cq->event_handler = ib_uverbs_cq_event_handler;
1152         cq->cq_context    = ev_file;
1153         atomic_set(&cq->usecnt, 0);
1154
1155         obj->uobject.object = cq;
1156         ret = idr_add_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1157         if (ret)
1158                 goto err_free;
1159
1160         memset(&resp, 0, sizeof resp);
1161         resp.cq_handle = obj->uobject.id;
1162         resp.cqe       = cq->cqe;
1163
1164         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1165                          &resp, sizeof resp)) {
1166                 ret = -EFAULT;
1167                 goto err_copy;
1168         }
1169
1170         mutex_lock(&file->mutex);
1171         list_add_tail(&obj->uobject.list, &file->ucontext->cq_list);
1172         mutex_unlock(&file->mutex);
1173
1174         obj->uobject.live = 1;
1175
1176         up_write(&obj->uobject.mutex);
1177
1178         return in_len;
1179
1180 err_copy:
1181         idr_remove_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1182
1183 err_free:
1184         ib_destroy_cq(cq);
1185
1186 err_file:
1187         if (ev_file)
1188                 ib_uverbs_release_ucq(file, ev_file, obj);
1189
1190 err:
1191         put_uobj_write(&obj->uobject);
1192         return ret;
1193 }
1194
1195 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1196                             const char __user *buf, int in_len,
1197                             int out_len)
1198 {
1199         struct ib_uverbs_resize_cq      cmd;
1200         struct ib_uverbs_resize_cq_resp resp;
1201         struct ib_udata                 udata;
1202         struct ib_cq                    *cq;
1203         int                             ret = -EINVAL;
1204
1205         if (copy_from_user(&cmd, buf, sizeof cmd))
1206                 return -EFAULT;
1207
1208         INIT_UDATA(&udata, buf + sizeof cmd,
1209                    (unsigned long) cmd.response + sizeof resp,
1210                    in_len - sizeof cmd, out_len - sizeof resp);
1211
1212         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1213         if (!cq)
1214                 return -EINVAL;
1215
1216         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1217         if (ret)
1218                 goto out;
1219
1220         resp.cqe = cq->cqe;
1221
1222         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1223                          &resp, sizeof resp.cqe))
1224                 ret = -EFAULT;
1225
1226 out:
1227         put_cq_read(cq);
1228
1229         return ret ? ret : in_len;
1230 }
1231
1232 static int copy_wc_to_user(void __user *dest, struct ib_wc *wc)
1233 {
1234         struct ib_uverbs_wc tmp;
1235
1236         tmp.wr_id               = wc->wr_id;
1237         tmp.status              = wc->status;
1238         tmp.opcode              = wc->opcode;
1239         tmp.vendor_err          = wc->vendor_err;
1240         tmp.byte_len            = wc->byte_len;
1241         tmp.ex.imm_data         = (__u32 __force) wc->ex.imm_data;
1242         tmp.qp_num              = wc->qp->qp_num;
1243         tmp.src_qp              = wc->src_qp;
1244         tmp.wc_flags            = wc->wc_flags;
1245         tmp.pkey_index          = wc->pkey_index;
1246         tmp.slid                = wc->slid;
1247         tmp.sl                  = wc->sl;
1248         tmp.dlid_path_bits      = wc->dlid_path_bits;
1249         tmp.port_num            = wc->port_num;
1250         tmp.reserved            = 0;
1251
1252         if (copy_to_user(dest, &tmp, sizeof tmp))
1253                 return -EFAULT;
1254
1255         return 0;
1256 }
1257
1258 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1259                           const char __user *buf, int in_len,
1260                           int out_len)
1261 {
1262         struct ib_uverbs_poll_cq       cmd;
1263         struct ib_uverbs_poll_cq_resp  resp;
1264         u8 __user                     *header_ptr;
1265         u8 __user                     *data_ptr;
1266         struct ib_cq                  *cq;
1267         struct ib_wc                   wc;
1268         int                            ret;
1269
1270         if (copy_from_user(&cmd, buf, sizeof cmd))
1271                 return -EFAULT;
1272
1273         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1274         if (!cq)
1275                 return -EINVAL;
1276
1277         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1278         header_ptr = (void __user *)(unsigned long) cmd.response;
1279         data_ptr = header_ptr + sizeof resp;
1280
1281         memset(&resp, 0, sizeof resp);
1282         while (resp.count < cmd.ne) {
1283                 ret = ib_poll_cq(cq, 1, &wc);
1284                 if (ret < 0)
1285                         goto out_put;
1286                 if (!ret)
1287                         break;
1288
1289                 ret = copy_wc_to_user(data_ptr, &wc);
1290                 if (ret)
1291                         goto out_put;
1292
1293                 data_ptr += sizeof(struct ib_uverbs_wc);
1294                 ++resp.count;
1295         }
1296
1297         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1298                 ret = -EFAULT;
1299                 goto out_put;
1300         }
1301
1302         ret = in_len;
1303
1304 out_put:
1305         put_cq_read(cq);
1306         return ret;
1307 }
1308
1309 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1310                                 const char __user *buf, int in_len,
1311                                 int out_len)
1312 {
1313         struct ib_uverbs_req_notify_cq cmd;
1314         struct ib_cq                  *cq;
1315
1316         if (copy_from_user(&cmd, buf, sizeof cmd))
1317                 return -EFAULT;
1318
1319         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1320         if (!cq)
1321                 return -EINVAL;
1322
1323         ib_req_notify_cq(cq, cmd.solicited_only ?
1324                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1325
1326         put_cq_read(cq);
1327
1328         return in_len;
1329 }
1330
1331 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1332                              const char __user *buf, int in_len,
1333                              int out_len)
1334 {
1335         struct ib_uverbs_destroy_cq      cmd;
1336         struct ib_uverbs_destroy_cq_resp resp;
1337         struct ib_uobject               *uobj;
1338         struct ib_cq                    *cq;
1339         struct ib_ucq_object            *obj;
1340         struct ib_uverbs_event_file     *ev_file;
1341         int                              ret = -EINVAL;
1342
1343         if (copy_from_user(&cmd, buf, sizeof cmd))
1344                 return -EFAULT;
1345
1346         uobj = idr_write_uobj(&ib_uverbs_cq_idr, cmd.cq_handle, file->ucontext);
1347         if (!uobj)
1348                 return -EINVAL;
1349         cq      = uobj->object;
1350         ev_file = cq->cq_context;
1351         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1352
1353         ret = ib_destroy_cq(cq);
1354         if (!ret)
1355                 uobj->live = 0;
1356
1357         put_uobj_write(uobj);
1358
1359         if (ret)
1360                 return ret;
1361
1362         idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
1363
1364         mutex_lock(&file->mutex);
1365         list_del(&uobj->list);
1366         mutex_unlock(&file->mutex);
1367
1368         ib_uverbs_release_ucq(file, ev_file, obj);
1369
1370         memset(&resp, 0, sizeof resp);
1371         resp.comp_events_reported  = obj->comp_events_reported;
1372         resp.async_events_reported = obj->async_events_reported;
1373
1374         put_uobj(uobj);
1375
1376         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1377                          &resp, sizeof resp))
1378                 return -EFAULT;
1379
1380         return in_len;
1381 }
1382
1383 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1384                             const char __user *buf, int in_len,
1385                             int out_len)
1386 {
1387         struct ib_uverbs_create_qp      cmd;
1388         struct ib_uverbs_create_qp_resp resp;
1389         struct ib_udata                 udata;
1390         struct ib_uqp_object           *obj;
1391         struct ib_device               *device;
1392         struct ib_pd                   *pd = NULL;
1393         struct ib_xrcd                 *xrcd = NULL;
1394         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1395         struct ib_cq                   *scq = NULL, *rcq = NULL;
1396         struct ib_srq                  *srq = NULL;
1397         struct ib_qp                   *qp;
1398         struct ib_qp_init_attr          attr;
1399         int ret;
1400
1401         if (out_len < sizeof resp)
1402                 return -ENOSPC;
1403
1404         if (copy_from_user(&cmd, buf, sizeof cmd))
1405                 return -EFAULT;
1406
1407         if (cmd.qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1408                 return -EPERM;
1409
1410         INIT_UDATA(&udata, buf + sizeof cmd,
1411                    (unsigned long) cmd.response + sizeof resp,
1412                    in_len - sizeof cmd, out_len - sizeof resp);
1413
1414         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1415         if (!obj)
1416                 return -ENOMEM;
1417
1418         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1419         down_write(&obj->uevent.uobject.mutex);
1420
1421         if (cmd.qp_type == IB_QPT_XRC_TGT) {
1422                 xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1423                 if (!xrcd) {
1424                         ret = -EINVAL;
1425                         goto err_put;
1426                 }
1427                 device = xrcd->device;
1428         } else {
1429                 if (cmd.qp_type == IB_QPT_XRC_INI) {
1430                         cmd.max_recv_wr = cmd.max_recv_sge = 0;
1431                 } else {
1432                         if (cmd.is_srq) {
1433                                 srq = idr_read_srq(cmd.srq_handle, file->ucontext);
1434                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1435                                         ret = -EINVAL;
1436                                         goto err_put;
1437                                 }
1438                         }
1439
1440                         if (cmd.recv_cq_handle != cmd.send_cq_handle) {
1441                                 rcq = idr_read_cq(cmd.recv_cq_handle, file->ucontext, 0);
1442                                 if (!rcq) {
1443                                         ret = -EINVAL;
1444                                         goto err_put;
1445                                 }
1446                         }
1447                 }
1448
1449                 scq = idr_read_cq(cmd.send_cq_handle, file->ucontext, !!rcq);
1450                 rcq = rcq ?: scq;
1451                 pd  = idr_read_pd(cmd.pd_handle, file->ucontext);
1452                 if (!pd || !scq) {
1453                         ret = -EINVAL;
1454                         goto err_put;
1455                 }
1456
1457                 device = pd->device;
1458         }
1459
1460         attr.event_handler = ib_uverbs_qp_event_handler;
1461         attr.qp_context    = file;
1462         attr.send_cq       = scq;
1463         attr.recv_cq       = rcq;
1464         attr.srq           = srq;
1465         attr.xrcd          = xrcd;
1466         attr.sq_sig_type   = cmd.sq_sig_all ? IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
1467         attr.qp_type       = cmd.qp_type;
1468         attr.create_flags  = 0;
1469
1470         attr.cap.max_send_wr     = cmd.max_send_wr;
1471         attr.cap.max_recv_wr     = cmd.max_recv_wr;
1472         attr.cap.max_send_sge    = cmd.max_send_sge;
1473         attr.cap.max_recv_sge    = cmd.max_recv_sge;
1474         attr.cap.max_inline_data = cmd.max_inline_data;
1475
1476         obj->uevent.events_reported     = 0;
1477         INIT_LIST_HEAD(&obj->uevent.event_list);
1478         INIT_LIST_HEAD(&obj->mcast_list);
1479
1480         if (cmd.qp_type == IB_QPT_XRC_TGT)
1481                 qp = ib_create_qp(pd, &attr);
1482         else
1483                 qp = device->create_qp(pd, &attr, &udata);
1484
1485         if (IS_ERR(qp)) {
1486                 ret = PTR_ERR(qp);
1487                 goto err_put;
1488         }
1489
1490         if (cmd.qp_type != IB_QPT_XRC_TGT) {
1491                 qp->real_qp       = qp;
1492                 qp->device        = device;
1493                 qp->pd            = pd;
1494                 qp->send_cq       = attr.send_cq;
1495                 qp->recv_cq       = attr.recv_cq;
1496                 qp->srq           = attr.srq;
1497                 qp->event_handler = attr.event_handler;
1498                 qp->qp_context    = attr.qp_context;
1499                 qp->qp_type       = attr.qp_type;
1500                 atomic_set(&qp->usecnt, 0);
1501                 atomic_inc(&pd->usecnt);
1502                 atomic_inc(&attr.send_cq->usecnt);
1503                 if (attr.recv_cq)
1504                         atomic_inc(&attr.recv_cq->usecnt);
1505                 if (attr.srq)
1506                         atomic_inc(&attr.srq->usecnt);
1507         }
1508         qp->uobject = &obj->uevent.uobject;
1509
1510         obj->uevent.uobject.object = qp;
1511         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1512         if (ret)
1513                 goto err_destroy;
1514
1515         memset(&resp, 0, sizeof resp);
1516         resp.qpn             = qp->qp_num;
1517         resp.qp_handle       = obj->uevent.uobject.id;
1518         resp.max_recv_sge    = attr.cap.max_recv_sge;
1519         resp.max_send_sge    = attr.cap.max_send_sge;
1520         resp.max_recv_wr     = attr.cap.max_recv_wr;
1521         resp.max_send_wr     = attr.cap.max_send_wr;
1522         resp.max_inline_data = attr.cap.max_inline_data;
1523
1524         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1525                          &resp, sizeof resp)) {
1526                 ret = -EFAULT;
1527                 goto err_copy;
1528         }
1529
1530         if (xrcd)
1531                 put_xrcd_read(xrcd_uobj);
1532         if (pd)
1533                 put_pd_read(pd);
1534         if (scq)
1535                 put_cq_read(scq);
1536         if (rcq && rcq != scq)
1537                 put_cq_read(rcq);
1538         if (srq)
1539                 put_srq_read(srq);
1540
1541         mutex_lock(&file->mutex);
1542         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1543         mutex_unlock(&file->mutex);
1544
1545         obj->uevent.uobject.live = 1;
1546
1547         up_write(&obj->uevent.uobject.mutex);
1548
1549         return in_len;
1550
1551 err_copy:
1552         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1553
1554 err_destroy:
1555         ib_destroy_qp(qp);
1556
1557 err_put:
1558         if (xrcd)
1559                 put_xrcd_read(xrcd_uobj);
1560         if (pd)
1561                 put_pd_read(pd);
1562         if (scq)
1563                 put_cq_read(scq);
1564         if (rcq && rcq != scq)
1565                 put_cq_read(rcq);
1566         if (srq)
1567                 put_srq_read(srq);
1568
1569         put_uobj_write(&obj->uevent.uobject);
1570         return ret;
1571 }
1572
1573 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1574                           const char __user *buf, int in_len, int out_len)
1575 {
1576         struct ib_uverbs_open_qp        cmd;
1577         struct ib_uverbs_create_qp_resp resp;
1578         struct ib_udata                 udata;
1579         struct ib_uqp_object           *obj;
1580         struct ib_xrcd                 *xrcd;
1581         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1582         struct ib_qp                   *qp;
1583         struct ib_qp_open_attr          attr;
1584         int ret;
1585
1586         if (out_len < sizeof resp)
1587                 return -ENOSPC;
1588
1589         if (copy_from_user(&cmd, buf, sizeof cmd))
1590                 return -EFAULT;
1591
1592         INIT_UDATA(&udata, buf + sizeof cmd,
1593                    (unsigned long) cmd.response + sizeof resp,
1594                    in_len - sizeof cmd, out_len - sizeof resp);
1595
1596         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1597         if (!obj)
1598                 return -ENOMEM;
1599
1600         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1601         down_write(&obj->uevent.uobject.mutex);
1602
1603         xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1604         if (!xrcd) {
1605                 ret = -EINVAL;
1606                 goto err_put;
1607         }
1608
1609         attr.event_handler = ib_uverbs_qp_event_handler;
1610         attr.qp_context    = file;
1611         attr.qp_num        = cmd.qpn;
1612         attr.qp_type       = cmd.qp_type;
1613
1614         obj->uevent.events_reported = 0;
1615         INIT_LIST_HEAD(&obj->uevent.event_list);
1616         INIT_LIST_HEAD(&obj->mcast_list);
1617
1618         qp = ib_open_qp(xrcd, &attr);
1619         if (IS_ERR(qp)) {
1620                 ret = PTR_ERR(qp);
1621                 goto err_put;
1622         }
1623
1624         qp->uobject = &obj->uevent.uobject;
1625
1626         obj->uevent.uobject.object = qp;
1627         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1628         if (ret)
1629                 goto err_destroy;
1630
1631         memset(&resp, 0, sizeof resp);
1632         resp.qpn       = qp->qp_num;
1633         resp.qp_handle = obj->uevent.uobject.id;
1634
1635         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1636                          &resp, sizeof resp)) {
1637                 ret = -EFAULT;
1638                 goto err_remove;
1639         }
1640
1641         put_xrcd_read(xrcd_uobj);
1642
1643         mutex_lock(&file->mutex);
1644         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1645         mutex_unlock(&file->mutex);
1646
1647         obj->uevent.uobject.live = 1;
1648
1649         up_write(&obj->uevent.uobject.mutex);
1650
1651         return in_len;
1652
1653 err_remove:
1654         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1655
1656 err_destroy:
1657         ib_destroy_qp(qp);
1658
1659 err_put:
1660         put_xrcd_read(xrcd_uobj);
1661         put_uobj_write(&obj->uevent.uobject);
1662         return ret;
1663 }
1664
1665 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1666                            const char __user *buf, int in_len,
1667                            int out_len)
1668 {
1669         struct ib_uverbs_query_qp      cmd;
1670         struct ib_uverbs_query_qp_resp resp;
1671         struct ib_qp                   *qp;
1672         struct ib_qp_attr              *attr;
1673         struct ib_qp_init_attr         *init_attr;
1674         int                            ret;
1675
1676         if (copy_from_user(&cmd, buf, sizeof cmd))
1677                 return -EFAULT;
1678
1679         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1680         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1681         if (!attr || !init_attr) {
1682                 ret = -ENOMEM;
1683                 goto out;
1684         }
1685
1686         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1687         if (!qp) {
1688                 ret = -EINVAL;
1689                 goto out;
1690         }
1691
1692         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1693
1694         put_qp_read(qp);
1695
1696         if (ret)
1697                 goto out;
1698
1699         memset(&resp, 0, sizeof resp);
1700
1701         resp.qp_state               = attr->qp_state;
1702         resp.cur_qp_state           = attr->cur_qp_state;
1703         resp.path_mtu               = attr->path_mtu;
1704         resp.path_mig_state         = attr->path_mig_state;
1705         resp.qkey                   = attr->qkey;
1706         resp.rq_psn                 = attr->rq_psn;
1707         resp.sq_psn                 = attr->sq_psn;
1708         resp.dest_qp_num            = attr->dest_qp_num;
1709         resp.qp_access_flags        = attr->qp_access_flags;
1710         resp.pkey_index             = attr->pkey_index;
1711         resp.alt_pkey_index         = attr->alt_pkey_index;
1712         resp.sq_draining            = attr->sq_draining;
1713         resp.max_rd_atomic          = attr->max_rd_atomic;
1714         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1715         resp.min_rnr_timer          = attr->min_rnr_timer;
1716         resp.port_num               = attr->port_num;
1717         resp.timeout                = attr->timeout;
1718         resp.retry_cnt              = attr->retry_cnt;
1719         resp.rnr_retry              = attr->rnr_retry;
1720         resp.alt_port_num           = attr->alt_port_num;
1721         resp.alt_timeout            = attr->alt_timeout;
1722
1723         memcpy(resp.dest.dgid, attr->ah_attr.grh.dgid.raw, 16);
1724         resp.dest.flow_label        = attr->ah_attr.grh.flow_label;
1725         resp.dest.sgid_index        = attr->ah_attr.grh.sgid_index;
1726         resp.dest.hop_limit         = attr->ah_attr.grh.hop_limit;
1727         resp.dest.traffic_class     = attr->ah_attr.grh.traffic_class;
1728         resp.dest.dlid              = attr->ah_attr.dlid;
1729         resp.dest.sl                = attr->ah_attr.sl;
1730         resp.dest.src_path_bits     = attr->ah_attr.src_path_bits;
1731         resp.dest.static_rate       = attr->ah_attr.static_rate;
1732         resp.dest.is_global         = !!(attr->ah_attr.ah_flags & IB_AH_GRH);
1733         resp.dest.port_num          = attr->ah_attr.port_num;
1734
1735         memcpy(resp.alt_dest.dgid, attr->alt_ah_attr.grh.dgid.raw, 16);
1736         resp.alt_dest.flow_label    = attr->alt_ah_attr.grh.flow_label;
1737         resp.alt_dest.sgid_index    = attr->alt_ah_attr.grh.sgid_index;
1738         resp.alt_dest.hop_limit     = attr->alt_ah_attr.grh.hop_limit;
1739         resp.alt_dest.traffic_class = attr->alt_ah_attr.grh.traffic_class;
1740         resp.alt_dest.dlid          = attr->alt_ah_attr.dlid;
1741         resp.alt_dest.sl            = attr->alt_ah_attr.sl;
1742         resp.alt_dest.src_path_bits = attr->alt_ah_attr.src_path_bits;
1743         resp.alt_dest.static_rate   = attr->alt_ah_attr.static_rate;
1744         resp.alt_dest.is_global     = !!(attr->alt_ah_attr.ah_flags & IB_AH_GRH);
1745         resp.alt_dest.port_num      = attr->alt_ah_attr.port_num;
1746
1747         resp.max_send_wr            = init_attr->cap.max_send_wr;
1748         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1749         resp.max_send_sge           = init_attr->cap.max_send_sge;
1750         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1751         resp.max_inline_data        = init_attr->cap.max_inline_data;
1752         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1753
1754         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1755                          &resp, sizeof resp))
1756                 ret = -EFAULT;
1757
1758 out:
1759         kfree(attr);
1760         kfree(init_attr);
1761
1762         return ret ? ret : in_len;
1763 }
1764
1765 /* Remove ignored fields set in the attribute mask */
1766 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1767 {
1768         switch (qp_type) {
1769         case IB_QPT_XRC_INI:
1770                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1771         case IB_QPT_XRC_TGT:
1772                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1773                                 IB_QP_RNR_RETRY);
1774         default:
1775                 return mask;
1776         }
1777 }
1778
1779 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
1780                             const char __user *buf, int in_len,
1781                             int out_len)
1782 {
1783         struct ib_uverbs_modify_qp cmd;
1784         struct ib_udata            udata;
1785         struct ib_qp              *qp;
1786         struct ib_qp_attr         *attr;
1787         int                        ret;
1788
1789         if (copy_from_user(&cmd, buf, sizeof cmd))
1790                 return -EFAULT;
1791
1792         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
1793                    out_len);
1794
1795         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1796         if (!attr)
1797                 return -ENOMEM;
1798
1799         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1800         if (!qp) {
1801                 ret = -EINVAL;
1802                 goto out;
1803         }
1804
1805         attr->qp_state            = cmd.qp_state;
1806         attr->cur_qp_state        = cmd.cur_qp_state;
1807         attr->path_mtu            = cmd.path_mtu;
1808         attr->path_mig_state      = cmd.path_mig_state;
1809         attr->qkey                = cmd.qkey;
1810         attr->rq_psn              = cmd.rq_psn;
1811         attr->sq_psn              = cmd.sq_psn;
1812         attr->dest_qp_num         = cmd.dest_qp_num;
1813         attr->qp_access_flags     = cmd.qp_access_flags;
1814         attr->pkey_index          = cmd.pkey_index;
1815         attr->alt_pkey_index      = cmd.alt_pkey_index;
1816         attr->en_sqd_async_notify = cmd.en_sqd_async_notify;
1817         attr->max_rd_atomic       = cmd.max_rd_atomic;
1818         attr->max_dest_rd_atomic  = cmd.max_dest_rd_atomic;
1819         attr->min_rnr_timer       = cmd.min_rnr_timer;
1820         attr->port_num            = cmd.port_num;
1821         attr->timeout             = cmd.timeout;
1822         attr->retry_cnt           = cmd.retry_cnt;
1823         attr->rnr_retry           = cmd.rnr_retry;
1824         attr->alt_port_num        = cmd.alt_port_num;
1825         attr->alt_timeout         = cmd.alt_timeout;
1826
1827         memcpy(attr->ah_attr.grh.dgid.raw, cmd.dest.dgid, 16);
1828         attr->ah_attr.grh.flow_label        = cmd.dest.flow_label;
1829         attr->ah_attr.grh.sgid_index        = cmd.dest.sgid_index;
1830         attr->ah_attr.grh.hop_limit         = cmd.dest.hop_limit;
1831         attr->ah_attr.grh.traffic_class     = cmd.dest.traffic_class;
1832         attr->ah_attr.dlid                  = cmd.dest.dlid;
1833         attr->ah_attr.sl                    = cmd.dest.sl;
1834         attr->ah_attr.src_path_bits         = cmd.dest.src_path_bits;
1835         attr->ah_attr.static_rate           = cmd.dest.static_rate;
1836         attr->ah_attr.ah_flags              = cmd.dest.is_global ? IB_AH_GRH : 0;
1837         attr->ah_attr.port_num              = cmd.dest.port_num;
1838
1839         memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd.alt_dest.dgid, 16);
1840         attr->alt_ah_attr.grh.flow_label    = cmd.alt_dest.flow_label;
1841         attr->alt_ah_attr.grh.sgid_index    = cmd.alt_dest.sgid_index;
1842         attr->alt_ah_attr.grh.hop_limit     = cmd.alt_dest.hop_limit;
1843         attr->alt_ah_attr.grh.traffic_class = cmd.alt_dest.traffic_class;
1844         attr->alt_ah_attr.dlid              = cmd.alt_dest.dlid;
1845         attr->alt_ah_attr.sl                = cmd.alt_dest.sl;
1846         attr->alt_ah_attr.src_path_bits     = cmd.alt_dest.src_path_bits;
1847         attr->alt_ah_attr.static_rate       = cmd.alt_dest.static_rate;
1848         attr->alt_ah_attr.ah_flags          = cmd.alt_dest.is_global ? IB_AH_GRH : 0;
1849         attr->alt_ah_attr.port_num          = cmd.alt_dest.port_num;
1850
1851         if (qp->real_qp == qp) {
1852                 ret = qp->device->modify_qp(qp, attr,
1853                         modify_qp_mask(qp->qp_type, cmd.attr_mask), &udata);
1854         } else {
1855                 ret = ib_modify_qp(qp, attr, modify_qp_mask(qp->qp_type, cmd.attr_mask));
1856         }
1857
1858         put_qp_read(qp);
1859
1860         if (ret)
1861                 goto out;
1862
1863         ret = in_len;
1864
1865 out:
1866         kfree(attr);
1867
1868         return ret;
1869 }
1870
1871 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
1872                              const char __user *buf, int in_len,
1873                              int out_len)
1874 {
1875         struct ib_uverbs_destroy_qp      cmd;
1876         struct ib_uverbs_destroy_qp_resp resp;
1877         struct ib_uobject               *uobj;
1878         struct ib_qp                    *qp;
1879         struct ib_uqp_object            *obj;
1880         int                              ret = -EINVAL;
1881
1882         if (copy_from_user(&cmd, buf, sizeof cmd))
1883                 return -EFAULT;
1884
1885         memset(&resp, 0, sizeof resp);
1886
1887         uobj = idr_write_uobj(&ib_uverbs_qp_idr, cmd.qp_handle, file->ucontext);
1888         if (!uobj)
1889                 return -EINVAL;
1890         qp  = uobj->object;
1891         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
1892
1893         if (!list_empty(&obj->mcast_list)) {
1894                 put_uobj_write(uobj);
1895                 return -EBUSY;
1896         }
1897
1898         ret = ib_destroy_qp(qp);
1899         if (!ret)
1900                 uobj->live = 0;
1901
1902         put_uobj_write(uobj);
1903
1904         if (ret)
1905                 return ret;
1906
1907         idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
1908
1909         mutex_lock(&file->mutex);
1910         list_del(&uobj->list);
1911         mutex_unlock(&file->mutex);
1912
1913         ib_uverbs_release_uevent(file, &obj->uevent);
1914
1915         resp.events_reported = obj->uevent.events_reported;
1916
1917         put_uobj(uobj);
1918
1919         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1920                          &resp, sizeof resp))
1921                 return -EFAULT;
1922
1923         return in_len;
1924 }
1925
1926 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
1927                             const char __user *buf, int in_len,
1928                             int out_len)
1929 {
1930         struct ib_uverbs_post_send      cmd;
1931         struct ib_uverbs_post_send_resp resp;
1932         struct ib_uverbs_send_wr       *user_wr;
1933         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
1934         struct ib_qp                   *qp;
1935         int                             i, sg_ind;
1936         int                             is_ud;
1937         ssize_t                         ret = -EINVAL;
1938
1939         if (copy_from_user(&cmd, buf, sizeof cmd))
1940                 return -EFAULT;
1941
1942         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
1943             cmd.sge_count * sizeof (struct ib_uverbs_sge))
1944                 return -EINVAL;
1945
1946         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
1947                 return -EINVAL;
1948
1949         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
1950         if (!user_wr)
1951                 return -ENOMEM;
1952
1953         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1954         if (!qp)
1955                 goto out;
1956
1957         is_ud = qp->qp_type == IB_QPT_UD;
1958         sg_ind = 0;
1959         last = NULL;
1960         for (i = 0; i < cmd.wr_count; ++i) {
1961                 if (copy_from_user(user_wr,
1962                                    buf + sizeof cmd + i * cmd.wqe_size,
1963                                    cmd.wqe_size)) {
1964                         ret = -EFAULT;
1965                         goto out_put;
1966                 }
1967
1968                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
1969                         ret = -EINVAL;
1970                         goto out_put;
1971                 }
1972
1973                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
1974                                user_wr->num_sge * sizeof (struct ib_sge),
1975                                GFP_KERNEL);
1976                 if (!next) {
1977                         ret = -ENOMEM;
1978                         goto out_put;
1979                 }
1980
1981                 if (!last)
1982                         wr = next;
1983                 else
1984                         last->next = next;
1985                 last = next;
1986
1987                 next->next       = NULL;
1988                 next->wr_id      = user_wr->wr_id;
1989                 next->num_sge    = user_wr->num_sge;
1990                 next->opcode     = user_wr->opcode;
1991                 next->send_flags = user_wr->send_flags;
1992
1993                 if (is_ud) {
1994                         next->wr.ud.ah = idr_read_ah(user_wr->wr.ud.ah,
1995                                                      file->ucontext);
1996                         if (!next->wr.ud.ah) {
1997                                 ret = -EINVAL;
1998                                 goto out_put;
1999                         }
2000                         next->wr.ud.remote_qpn  = user_wr->wr.ud.remote_qpn;
2001                         next->wr.ud.remote_qkey = user_wr->wr.ud.remote_qkey;
2002                 } else {
2003                         switch (next->opcode) {
2004                         case IB_WR_RDMA_WRITE_WITH_IMM:
2005                                 next->ex.imm_data =
2006                                         (__be32 __force) user_wr->ex.imm_data;
2007                         case IB_WR_RDMA_WRITE:
2008                         case IB_WR_RDMA_READ:
2009                                 next->wr.rdma.remote_addr =
2010                                         user_wr->wr.rdma.remote_addr;
2011                                 next->wr.rdma.rkey        =
2012                                         user_wr->wr.rdma.rkey;
2013                                 break;
2014                         case IB_WR_SEND_WITH_IMM:
2015                                 next->ex.imm_data =
2016                                         (__be32 __force) user_wr->ex.imm_data;
2017                                 break;
2018                         case IB_WR_SEND_WITH_INV:
2019                                 next->ex.invalidate_rkey =
2020                                         user_wr->ex.invalidate_rkey;
2021                                 break;
2022                         case IB_WR_ATOMIC_CMP_AND_SWP:
2023                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2024                                 next->wr.atomic.remote_addr =
2025                                         user_wr->wr.atomic.remote_addr;
2026                                 next->wr.atomic.compare_add =
2027                                         user_wr->wr.atomic.compare_add;
2028                                 next->wr.atomic.swap = user_wr->wr.atomic.swap;
2029                                 next->wr.atomic.rkey = user_wr->wr.atomic.rkey;
2030                                 break;
2031                         default:
2032                                 break;
2033                         }
2034                 }
2035
2036                 if (next->num_sge) {
2037                         next->sg_list = (void *) next +
2038                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2039                         if (copy_from_user(next->sg_list,
2040                                            buf + sizeof cmd +
2041                                            cmd.wr_count * cmd.wqe_size +
2042                                            sg_ind * sizeof (struct ib_sge),
2043                                            next->num_sge * sizeof (struct ib_sge))) {
2044                                 ret = -EFAULT;
2045                                 goto out_put;
2046                         }
2047                         sg_ind += next->num_sge;
2048                 } else
2049                         next->sg_list = NULL;
2050         }
2051
2052         resp.bad_wr = 0;
2053         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2054         if (ret)
2055                 for (next = wr; next; next = next->next) {
2056                         ++resp.bad_wr;
2057                         if (next == bad_wr)
2058                                 break;
2059                 }
2060
2061         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2062                          &resp, sizeof resp))
2063                 ret = -EFAULT;
2064
2065 out_put:
2066         put_qp_read(qp);
2067
2068         while (wr) {
2069                 if (is_ud && wr->wr.ud.ah)
2070                         put_ah_read(wr->wr.ud.ah);
2071                 next = wr->next;
2072                 kfree(wr);
2073                 wr = next;
2074         }
2075
2076 out:
2077         kfree(user_wr);
2078
2079         return ret ? ret : in_len;
2080 }
2081
2082 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2083                                                     int in_len,
2084                                                     u32 wr_count,
2085                                                     u32 sge_count,
2086                                                     u32 wqe_size)
2087 {
2088         struct ib_uverbs_recv_wr *user_wr;
2089         struct ib_recv_wr        *wr = NULL, *last, *next;
2090         int                       sg_ind;
2091         int                       i;
2092         int                       ret;
2093
2094         if (in_len < wqe_size * wr_count +
2095             sge_count * sizeof (struct ib_uverbs_sge))
2096                 return ERR_PTR(-EINVAL);
2097
2098         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2099                 return ERR_PTR(-EINVAL);
2100
2101         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2102         if (!user_wr)
2103                 return ERR_PTR(-ENOMEM);
2104
2105         sg_ind = 0;
2106         last = NULL;
2107         for (i = 0; i < wr_count; ++i) {
2108                 if (copy_from_user(user_wr, buf + i * wqe_size,
2109                                    wqe_size)) {
2110                         ret = -EFAULT;
2111                         goto err;
2112                 }
2113
2114                 if (user_wr->num_sge + sg_ind > sge_count) {
2115                         ret = -EINVAL;
2116                         goto err;
2117                 }
2118
2119                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2120                                user_wr->num_sge * sizeof (struct ib_sge),
2121                                GFP_KERNEL);
2122                 if (!next) {
2123                         ret = -ENOMEM;
2124                         goto err;
2125                 }
2126
2127                 if (!last)
2128                         wr = next;
2129                 else
2130                         last->next = next;
2131                 last = next;
2132
2133                 next->next       = NULL;
2134                 next->wr_id      = user_wr->wr_id;
2135                 next->num_sge    = user_wr->num_sge;
2136
2137                 if (next->num_sge) {
2138                         next->sg_list = (void *) next +
2139                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2140                         if (copy_from_user(next->sg_list,
2141                                            buf + wr_count * wqe_size +
2142                                            sg_ind * sizeof (struct ib_sge),
2143                                            next->num_sge * sizeof (struct ib_sge))) {
2144                                 ret = -EFAULT;
2145                                 goto err;
2146                         }
2147                         sg_ind += next->num_sge;
2148                 } else
2149                         next->sg_list = NULL;
2150         }
2151
2152         kfree(user_wr);
2153         return wr;
2154
2155 err:
2156         kfree(user_wr);
2157
2158         while (wr) {
2159                 next = wr->next;
2160                 kfree(wr);
2161                 wr = next;
2162         }
2163
2164         return ERR_PTR(ret);
2165 }
2166
2167 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2168                             const char __user *buf, int in_len,
2169                             int out_len)
2170 {
2171         struct ib_uverbs_post_recv      cmd;
2172         struct ib_uverbs_post_recv_resp resp;
2173         struct ib_recv_wr              *wr, *next, *bad_wr;
2174         struct ib_qp                   *qp;
2175         ssize_t                         ret = -EINVAL;
2176
2177         if (copy_from_user(&cmd, buf, sizeof cmd))
2178                 return -EFAULT;
2179
2180         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2181                                        in_len - sizeof cmd, cmd.wr_count,
2182                                        cmd.sge_count, cmd.wqe_size);
2183         if (IS_ERR(wr))
2184                 return PTR_ERR(wr);
2185
2186         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2187         if (!qp)
2188                 goto out;
2189
2190         resp.bad_wr = 0;
2191         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2192
2193         put_qp_read(qp);
2194
2195         if (ret)
2196                 for (next = wr; next; next = next->next) {
2197                         ++resp.bad_wr;
2198                         if (next == bad_wr)
2199                                 break;
2200                 }
2201
2202         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2203                          &resp, sizeof resp))
2204                 ret = -EFAULT;
2205
2206 out:
2207         while (wr) {
2208                 next = wr->next;
2209                 kfree(wr);
2210                 wr = next;
2211         }
2212
2213         return ret ? ret : in_len;
2214 }
2215
2216 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2217                                 const char __user *buf, int in_len,
2218                                 int out_len)
2219 {
2220         struct ib_uverbs_post_srq_recv      cmd;
2221         struct ib_uverbs_post_srq_recv_resp resp;
2222         struct ib_recv_wr                  *wr, *next, *bad_wr;
2223         struct ib_srq                      *srq;
2224         ssize_t                             ret = -EINVAL;
2225
2226         if (copy_from_user(&cmd, buf, sizeof cmd))
2227                 return -EFAULT;
2228
2229         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2230                                        in_len - sizeof cmd, cmd.wr_count,
2231                                        cmd.sge_count, cmd.wqe_size);
2232         if (IS_ERR(wr))
2233                 return PTR_ERR(wr);
2234
2235         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2236         if (!srq)
2237                 goto out;
2238
2239         resp.bad_wr = 0;
2240         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2241
2242         put_srq_read(srq);
2243
2244         if (ret)
2245                 for (next = wr; next; next = next->next) {
2246                         ++resp.bad_wr;
2247                         if (next == bad_wr)
2248                                 break;
2249                 }
2250
2251         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2252                          &resp, sizeof resp))
2253                 ret = -EFAULT;
2254
2255 out:
2256         while (wr) {
2257                 next = wr->next;
2258                 kfree(wr);
2259                 wr = next;
2260         }
2261
2262         return ret ? ret : in_len;
2263 }
2264
2265 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2266                             const char __user *buf, int in_len,
2267                             int out_len)
2268 {
2269         struct ib_uverbs_create_ah       cmd;
2270         struct ib_uverbs_create_ah_resp  resp;
2271         struct ib_uobject               *uobj;
2272         struct ib_pd                    *pd;
2273         struct ib_ah                    *ah;
2274         struct ib_ah_attr               attr;
2275         int ret;
2276
2277         if (out_len < sizeof resp)
2278                 return -ENOSPC;
2279
2280         if (copy_from_user(&cmd, buf, sizeof cmd))
2281                 return -EFAULT;
2282
2283         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
2284         if (!uobj)
2285                 return -ENOMEM;
2286
2287         init_uobj(uobj, cmd.user_handle, file->ucontext, &ah_lock_class);
2288         down_write(&uobj->mutex);
2289
2290         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
2291         if (!pd) {
2292                 ret = -EINVAL;
2293                 goto err;
2294         }
2295
2296         attr.dlid              = cmd.attr.dlid;
2297         attr.sl                = cmd.attr.sl;
2298         attr.src_path_bits     = cmd.attr.src_path_bits;
2299         attr.static_rate       = cmd.attr.static_rate;
2300         attr.ah_flags          = cmd.attr.is_global ? IB_AH_GRH : 0;
2301         attr.port_num          = cmd.attr.port_num;
2302         attr.grh.flow_label    = cmd.attr.grh.flow_label;
2303         attr.grh.sgid_index    = cmd.attr.grh.sgid_index;
2304         attr.grh.hop_limit     = cmd.attr.grh.hop_limit;
2305         attr.grh.traffic_class = cmd.attr.grh.traffic_class;
2306         memcpy(attr.grh.dgid.raw, cmd.attr.grh.dgid, 16);
2307
2308         ah = ib_create_ah(pd, &attr);
2309         if (IS_ERR(ah)) {
2310                 ret = PTR_ERR(ah);
2311                 goto err_put;
2312         }
2313
2314         ah->uobject  = uobj;
2315         uobj->object = ah;
2316
2317         ret = idr_add_uobj(&ib_uverbs_ah_idr, uobj);
2318         if (ret)
2319                 goto err_destroy;
2320
2321         resp.ah_handle = uobj->id;
2322
2323         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2324                          &resp, sizeof resp)) {
2325                 ret = -EFAULT;
2326                 goto err_copy;
2327         }
2328
2329         put_pd_read(pd);
2330
2331         mutex_lock(&file->mutex);
2332         list_add_tail(&uobj->list, &file->ucontext->ah_list);
2333         mutex_unlock(&file->mutex);
2334
2335         uobj->live = 1;
2336
2337         up_write(&uobj->mutex);
2338
2339         return in_len;
2340
2341 err_copy:
2342         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2343
2344 err_destroy:
2345         ib_destroy_ah(ah);
2346
2347 err_put:
2348         put_pd_read(pd);
2349
2350 err:
2351         put_uobj_write(uobj);
2352         return ret;
2353 }
2354
2355 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2356                              const char __user *buf, int in_len, int out_len)
2357 {
2358         struct ib_uverbs_destroy_ah cmd;
2359         struct ib_ah               *ah;
2360         struct ib_uobject          *uobj;
2361         int                         ret;
2362
2363         if (copy_from_user(&cmd, buf, sizeof cmd))
2364                 return -EFAULT;
2365
2366         uobj = idr_write_uobj(&ib_uverbs_ah_idr, cmd.ah_handle, file->ucontext);
2367         if (!uobj)
2368                 return -EINVAL;
2369         ah = uobj->object;
2370
2371         ret = ib_destroy_ah(ah);
2372         if (!ret)
2373                 uobj->live = 0;
2374
2375         put_uobj_write(uobj);
2376
2377         if (ret)
2378                 return ret;
2379
2380         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2381
2382         mutex_lock(&file->mutex);
2383         list_del(&uobj->list);
2384         mutex_unlock(&file->mutex);
2385
2386         put_uobj(uobj);
2387
2388         return in_len;
2389 }
2390
2391 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2392                                const char __user *buf, int in_len,
2393                                int out_len)
2394 {
2395         struct ib_uverbs_attach_mcast cmd;
2396         struct ib_qp                 *qp;
2397         struct ib_uqp_object         *obj;
2398         struct ib_uverbs_mcast_entry *mcast;
2399         int                           ret;
2400
2401         if (copy_from_user(&cmd, buf, sizeof cmd))
2402                 return -EFAULT;
2403
2404         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2405         if (!qp)
2406                 return -EINVAL;
2407
2408         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2409
2410         list_for_each_entry(mcast, &obj->mcast_list, list)
2411                 if (cmd.mlid == mcast->lid &&
2412                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2413                         ret = 0;
2414                         goto out_put;
2415                 }
2416
2417         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2418         if (!mcast) {
2419                 ret = -ENOMEM;
2420                 goto out_put;
2421         }
2422
2423         mcast->lid = cmd.mlid;
2424         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2425
2426         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2427         if (!ret)
2428                 list_add_tail(&mcast->list, &obj->mcast_list);
2429         else
2430                 kfree(mcast);
2431
2432 out_put:
2433         put_qp_write(qp);
2434
2435         return ret ? ret : in_len;
2436 }
2437
2438 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2439                                const char __user *buf, int in_len,
2440                                int out_len)
2441 {
2442         struct ib_uverbs_detach_mcast cmd;
2443         struct ib_uqp_object         *obj;
2444         struct ib_qp                 *qp;
2445         struct ib_uverbs_mcast_entry *mcast;
2446         int                           ret = -EINVAL;
2447
2448         if (copy_from_user(&cmd, buf, sizeof cmd))
2449                 return -EFAULT;
2450
2451         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2452         if (!qp)
2453                 return -EINVAL;
2454
2455         ret = ib_detach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
2456         if (ret)
2457                 goto out_put;
2458
2459         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2460
2461         list_for_each_entry(mcast, &obj->mcast_list, list)
2462                 if (cmd.mlid == mcast->lid &&
2463                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2464                         list_del(&mcast->list);
2465                         kfree(mcast);
2466                         break;
2467                 }
2468
2469 out_put:
2470         put_qp_write(qp);
2471
2472         return ret ? ret : in_len;
2473 }
2474
2475 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
2476                                 struct ib_uverbs_create_xsrq *cmd,
2477                                 struct ib_udata *udata)
2478 {
2479         struct ib_uverbs_create_srq_resp resp;
2480         struct ib_usrq_object           *obj;
2481         struct ib_pd                    *pd;
2482         struct ib_srq                   *srq;
2483         struct ib_uobject               *uninitialized_var(xrcd_uobj);
2484         struct ib_srq_init_attr          attr;
2485         int ret;
2486
2487         obj = kmalloc(sizeof *obj, GFP_KERNEL);
2488         if (!obj)
2489                 return -ENOMEM;
2490
2491         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext, &srq_lock_class);
2492         down_write(&obj->uevent.uobject.mutex);
2493
2494         if (cmd->srq_type == IB_SRQT_XRC) {
2495                 attr.ext.xrc.xrcd  = idr_read_xrcd(cmd->xrcd_handle, file->ucontext, &xrcd_uobj);
2496                 if (!attr.ext.xrc.xrcd) {
2497                         ret = -EINVAL;
2498                         goto err;
2499                 }
2500
2501                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
2502                 atomic_inc(&obj->uxrcd->refcnt);
2503
2504                 attr.ext.xrc.cq  = idr_read_cq(cmd->cq_handle, file->ucontext, 0);
2505                 if (!attr.ext.xrc.cq) {
2506                         ret = -EINVAL;
2507                         goto err_put_xrcd;
2508                 }
2509         }
2510
2511         pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
2512         if (!pd) {
2513                 ret = -EINVAL;
2514                 goto err_put_cq;
2515         }
2516
2517         attr.event_handler  = ib_uverbs_srq_event_handler;
2518         attr.srq_context    = file;
2519         attr.srq_type       = cmd->srq_type;
2520         attr.attr.max_wr    = cmd->max_wr;
2521         attr.attr.max_sge   = cmd->max_sge;
2522         attr.attr.srq_limit = cmd->srq_limit;
2523
2524         obj->uevent.events_reported = 0;
2525         INIT_LIST_HEAD(&obj->uevent.event_list);
2526
2527         srq = pd->device->create_srq(pd, &attr, udata);
2528         if (IS_ERR(srq)) {
2529                 ret = PTR_ERR(srq);
2530                 goto err_put;
2531         }
2532
2533         srq->device        = pd->device;
2534         srq->pd            = pd;
2535         srq->srq_type      = cmd->srq_type;
2536         srq->uobject       = &obj->uevent.uobject;
2537         srq->event_handler = attr.event_handler;
2538         srq->srq_context   = attr.srq_context;
2539
2540         if (cmd->srq_type == IB_SRQT_XRC) {
2541                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
2542                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
2543                 atomic_inc(&attr.ext.xrc.cq->usecnt);
2544                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
2545         }
2546
2547         atomic_inc(&pd->usecnt);
2548         atomic_set(&srq->usecnt, 0);
2549
2550         obj->uevent.uobject.object = srq;
2551         ret = idr_add_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2552         if (ret)
2553                 goto err_destroy;
2554
2555         memset(&resp, 0, sizeof resp);
2556         resp.srq_handle = obj->uevent.uobject.id;
2557         resp.max_wr     = attr.attr.max_wr;
2558         resp.max_sge    = attr.attr.max_sge;
2559         if (cmd->srq_type == IB_SRQT_XRC)
2560                 resp.srqn = srq->ext.xrc.srq_num;
2561
2562         if (copy_to_user((void __user *) (unsigned long) cmd->response,
2563                          &resp, sizeof resp)) {
2564                 ret = -EFAULT;
2565                 goto err_copy;
2566         }
2567
2568         if (cmd->srq_type == IB_SRQT_XRC) {
2569                 put_uobj_read(xrcd_uobj);
2570                 put_cq_read(attr.ext.xrc.cq);
2571         }
2572         put_pd_read(pd);
2573
2574         mutex_lock(&file->mutex);
2575         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->srq_list);
2576         mutex_unlock(&file->mutex);
2577
2578         obj->uevent.uobject.live = 1;
2579
2580         up_write(&obj->uevent.uobject.mutex);
2581
2582         return 0;
2583
2584 err_copy:
2585         idr_remove_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2586
2587 err_destroy:
2588         ib_destroy_srq(srq);
2589
2590 err_put:
2591         put_pd_read(pd);
2592
2593 err_put_cq:
2594         if (cmd->srq_type == IB_SRQT_XRC)
2595                 put_cq_read(attr.ext.xrc.cq);
2596
2597 err_put_xrcd:
2598         if (cmd->srq_type == IB_SRQT_XRC) {
2599                 atomic_dec(&obj->uxrcd->refcnt);
2600                 put_uobj_read(xrcd_uobj);
2601         }
2602
2603 err:
2604         put_uobj_write(&obj->uevent.uobject);
2605         return ret;
2606 }
2607
2608 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
2609                              const char __user *buf, int in_len,
2610                              int out_len)
2611 {
2612         struct ib_uverbs_create_srq      cmd;
2613         struct ib_uverbs_create_xsrq     xcmd;
2614         struct ib_uverbs_create_srq_resp resp;
2615         struct ib_udata                  udata;
2616         int ret;
2617
2618         if (out_len < sizeof resp)
2619                 return -ENOSPC;
2620
2621         if (copy_from_user(&cmd, buf, sizeof cmd))
2622                 return -EFAULT;
2623
2624         xcmd.response    = cmd.response;
2625         xcmd.user_handle = cmd.user_handle;
2626         xcmd.srq_type    = IB_SRQT_BASIC;
2627         xcmd.pd_handle   = cmd.pd_handle;
2628         xcmd.max_wr      = cmd.max_wr;
2629         xcmd.max_sge     = cmd.max_sge;
2630         xcmd.srq_limit   = cmd.srq_limit;
2631
2632         INIT_UDATA(&udata, buf + sizeof cmd,
2633                    (unsigned long) cmd.response + sizeof resp,
2634                    in_len - sizeof cmd, out_len - sizeof resp);
2635
2636         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
2637         if (ret)
2638                 return ret;
2639
2640         return in_len;
2641 }
2642
2643 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
2644                               const char __user *buf, int in_len, int out_len)
2645 {
2646         struct ib_uverbs_create_xsrq     cmd;
2647         struct ib_uverbs_create_srq_resp resp;
2648         struct ib_udata                  udata;
2649         int ret;
2650
2651         if (out_len < sizeof resp)
2652                 return -ENOSPC;
2653
2654         if (copy_from_user(&cmd, buf, sizeof cmd))
2655                 return -EFAULT;
2656
2657         INIT_UDATA(&udata, buf + sizeof cmd,
2658                    (unsigned long) cmd.response + sizeof resp,
2659                    in_len - sizeof cmd, out_len - sizeof resp);
2660
2661         ret = __uverbs_create_xsrq(file, &cmd, &udata);
2662         if (ret)
2663                 return ret;
2664
2665         return in_len;
2666 }
2667
2668 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
2669                              const char __user *buf, int in_len,
2670                              int out_len)
2671 {
2672         struct ib_uverbs_modify_srq cmd;
2673         struct ib_udata             udata;
2674         struct ib_srq              *srq;
2675         struct ib_srq_attr          attr;
2676         int                         ret;
2677
2678         if (copy_from_user(&cmd, buf, sizeof cmd))
2679                 return -EFAULT;
2680
2681         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
2682                    out_len);
2683
2684         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2685         if (!srq)
2686                 return -EINVAL;
2687
2688         attr.max_wr    = cmd.max_wr;
2689         attr.srq_limit = cmd.srq_limit;
2690
2691         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
2692
2693         put_srq_read(srq);
2694
2695         return ret ? ret : in_len;
2696 }
2697
2698 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
2699                             const char __user *buf,
2700                             int in_len, int out_len)
2701 {
2702         struct ib_uverbs_query_srq      cmd;
2703         struct ib_uverbs_query_srq_resp resp;
2704         struct ib_srq_attr              attr;
2705         struct ib_srq                   *srq;
2706         int                             ret;
2707
2708         if (out_len < sizeof resp)
2709                 return -ENOSPC;
2710
2711         if (copy_from_user(&cmd, buf, sizeof cmd))
2712                 return -EFAULT;
2713
2714         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2715         if (!srq)
2716                 return -EINVAL;
2717
2718         ret = ib_query_srq(srq, &attr);
2719
2720         put_srq_read(srq);
2721
2722         if (ret)
2723                 return ret;
2724
2725         memset(&resp, 0, sizeof resp);
2726
2727         resp.max_wr    = attr.max_wr;
2728         resp.max_sge   = attr.max_sge;
2729         resp.srq_limit = attr.srq_limit;
2730
2731         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2732                          &resp, sizeof resp))
2733                 return -EFAULT;
2734
2735         return in_len;
2736 }
2737
2738 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
2739                               const char __user *buf, int in_len,
2740                               int out_len)
2741 {
2742         struct ib_uverbs_destroy_srq      cmd;
2743         struct ib_uverbs_destroy_srq_resp resp;
2744         struct ib_uobject                *uobj;
2745         struct ib_srq                    *srq;
2746         struct ib_uevent_object          *obj;
2747         int                               ret = -EINVAL;
2748
2749         if (copy_from_user(&cmd, buf, sizeof cmd))
2750                 return -EFAULT;
2751
2752         uobj = idr_write_uobj(&ib_uverbs_srq_idr, cmd.srq_handle, file->ucontext);
2753         if (!uobj)
2754                 return -EINVAL;
2755         srq = uobj->object;
2756         obj = container_of(uobj, struct ib_uevent_object, uobject);
2757
2758         ret = ib_destroy_srq(srq);
2759         if (!ret)
2760                 uobj->live = 0;
2761
2762         put_uobj_write(uobj);
2763
2764         if (ret)
2765                 return ret;
2766
2767         idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
2768
2769         mutex_lock(&file->mutex);
2770         list_del(&uobj->list);
2771         mutex_unlock(&file->mutex);
2772
2773         ib_uverbs_release_uevent(file, obj);
2774
2775         memset(&resp, 0, sizeof resp);
2776         resp.events_reported = obj->events_reported;
2777
2778         put_uobj(uobj);
2779
2780         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2781                          &resp, sizeof resp))
2782                 ret = -EFAULT;
2783
2784         return ret ? ret : in_len;
2785 }