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 mw_lock_class   = { .name = "MW-uobj" };
  52 static struct uverbs_lock_class cq_lock_class   = { .name = "CQ-uobj" };
  53 static struct uverbs_lock_class qp_lock_class   = { .name = "QP-uobj" };
  54 static struct uverbs_lock_class ah_lock_class   = { .name = "AH-uobj" };
  55 static struct uverbs_lock_class srq_lock_class  = { .name = "SRQ-uobj" };
  56 static struct uverbs_lock_class xrcd_lock_class = { .name = "XRCD-uobj" };
  57
  58 #define INIT_UDATA(udata, ibuf, obuf, ilen, olen)                       \
  59         do {                                                            \
  60                 (udata)->inbuf  = (void __user *) (ibuf);               \
  61                 (udata)->outbuf = (void __user *) (obuf);               \
  62                 (udata)->inlen  = (ilen);                               \
  63                 (udata)->outlen = (olen);                               \
  64         } while (0)
  65
  66 /*
  67  * The ib_uobject locking scheme is as follows:
  68  *
  69  * - ib_uverbs_idr_lock protects the uverbs idrs themselves, so it
  70  *   needs to be held during all idr operations.  When an object is
  71  *   looked up, a reference must be taken on the object's kref before
  72  *   dropping this lock.
  73  *
  74  * - Each object also has an rwsem.  This rwsem must be held for
  75  *   reading while an operation that uses the object is performed.
  76  *   For example, while registering an MR, the associated PD's
  77  *   uobject.mutex must be held for reading.  The rwsem must be held
  78  *   for writing while initializing or destroying an object.
  79  *
  80  * - In addition, each object has a "live" flag.  If this flag is not
  81  *   set, then lookups of the object will fail even if it is found in
  82  *   the idr.  This handles a reader that blocks and does not acquire
  83  *   the rwsem until after the object is destroyed.  The destroy
  84  *   operation will set the live flag to 0 and then drop the rwsem;
  85  *   this will allow the reader to acquire the rwsem, see that the
  86  *   live flag is 0, and then drop the rwsem and its reference to
  87  *   object.  The underlying storage will not be freed until the last
  88  *   reference to the object is dropped.
  89  */
  90
  91 static void init_uobj(struct ib_uobject *uobj, u64 user_handle,
  92                       struct ib_ucontext *context, struct uverbs_lock_class *c)
  93 {
  94         uobj->user_handle = user_handle;
  95         uobj->context     = context;
  96         kref_init(&uobj->ref);
  97         init_rwsem(&uobj->mutex);
  98         lockdep_set_class_and_name(&uobj->mutex, &c->key, c->name);
  99         uobj->live        = 0;
 100 }
 101
 102 static void release_uobj(struct kref *kref)
 103 {
 104         kfree(container_of(kref, struct ib_uobject, ref));
 105 }
 106
 107 static void put_uobj(struct ib_uobject *uobj)
 108 {
 109         kref_put(&uobj->ref, release_uobj);
 110 }
 111
 112 static void put_uobj_read(struct ib_uobject *uobj)
 113 {
 114         up_read(&uobj->mutex);
 115         put_uobj(uobj);
 116 }
 117
 118 static void put_uobj_write(struct ib_uobject *uobj)
 119 {
 120         up_write(&uobj->mutex);
 121         put_uobj(uobj);
 122 }
 123
 124 static int idr_add_uobj(struct idr *idr, struct ib_uobject *uobj)
 125 {
 126         int ret;
 127
 128         idr_preload(GFP_KERNEL);
 129         spin_lock(&ib_uverbs_idr_lock);
 130
 131         ret = idr_alloc(idr, uobj, 0, 0, GFP_NOWAIT);
 132         if (ret >= 0)
 133                 uobj->id = ret;
 134
 135         spin_unlock(&ib_uverbs_idr_lock);
 136         idr_preload_end();
 137
 138         return ret < 0 ? ret : 0;
 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_flags(O_CLOEXEC);
 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 fd                       f = {NULL, 0};
 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 = fdget(cmd.fd);
 728                 if (!f.file) {
 729                         ret = -EBADF;
 730                         goto err_tree_mutex_unlock;
 731                 }
 732
 733                 inode = file_inode(f.file);
 734                 xrcd = find_xrcd(file->device, inode);
 735                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 736                         /* no file descriptor. Need CREATE flag */
 737                         ret = -EAGAIN;
 738                         goto err_tree_mutex_unlock;
 739                 }
 740
 741                 if (xrcd && cmd.oflags & O_EXCL) {
 742                         ret = -EINVAL;
 743                         goto err_tree_mutex_unlock;
 744                 }
 745         }
 746
 747         obj = kmalloc(sizeof *obj, GFP_KERNEL);
 748         if (!obj) {
 749                 ret = -ENOMEM;
 750                 goto err_tree_mutex_unlock;
 751         }
 752
 753         init_uobj(&obj->uobject, 0, file->ucontext, &xrcd_lock_class);
 754
 755         down_write(&obj->uobject.mutex);
 756
 757         if (!xrcd) {
 758                 xrcd = file->device->ib_dev->alloc_xrcd(file->device->ib_dev,
 759                                                         file->ucontext, &udata);
 760                 if (IS_ERR(xrcd)) {
 761                         ret = PTR_ERR(xrcd);
 762                         goto err;
 763                 }
 764
 765                 xrcd->inode   = inode;
 766                 xrcd->device  = file->device->ib_dev;
 767                 atomic_set(&xrcd->usecnt, 0);
 768                 mutex_init(&xrcd->tgt_qp_mutex);
 769                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 770                 new_xrcd = 1;
 771         }
 772
 773         atomic_set(&obj->refcnt, 0);
 774         obj->uobject.object = xrcd;
 775         ret = idr_add_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 776         if (ret)
 777                 goto err_idr;
 778
 779         memset(&resp, 0, sizeof resp);
 780         resp.xrcd_handle = obj->uobject.id;
 781
 782         if (inode) {
 783                 if (new_xrcd) {
 784                         /* create new inode/xrcd table entry */
 785                         ret = xrcd_table_insert(file->device, inode, xrcd);
 786                         if (ret)
 787                                 goto err_insert_xrcd;
 788                 }
 789                 atomic_inc(&xrcd->usecnt);
 790         }
 791
 792         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 793                          &resp, sizeof resp)) {
 794                 ret = -EFAULT;
 795                 goto err_copy;
 796         }
 797
 798         if (f.file)
 799                 fdput(f);
 800
 801         mutex_lock(&file->mutex);
 802         list_add_tail(&obj->uobject.list, &file->ucontext->xrcd_list);
 803         mutex_unlock(&file->mutex);
 804
 805         obj->uobject.live = 1;
 806         up_write(&obj->uobject.mutex);
 807
 808         mutex_unlock(&file->device->xrcd_tree_mutex);
 809         return in_len;
 810
 811 err_copy:
 812         if (inode) {
 813                 if (new_xrcd)
 814                         xrcd_table_delete(file->device, inode);
 815                 atomic_dec(&xrcd->usecnt);
 816         }
 817
 818 err_insert_xrcd:
 819         idr_remove_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 820
 821 err_idr:
 822         ib_dealloc_xrcd(xrcd);
 823
 824 err:
 825         put_uobj_write(&obj->uobject);
 826
 827 err_tree_mutex_unlock:
 828         if (f.file)
 829                 fdput(f);
 830
 831         mutex_unlock(&file->device->xrcd_tree_mutex);
 832
 833         return ret;
 834 }
 835
 836 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 837                              const char __user *buf, int in_len,
 838                              int out_len)
 839 {
 840         struct ib_uverbs_close_xrcd cmd;
 841         struct ib_uobject           *uobj;
 842         struct ib_xrcd              *xrcd = NULL;
 843         struct inode                *inode = NULL;
 844         struct ib_uxrcd_object      *obj;
 845         int                         live;
 846         int                         ret = 0;
 847
 848         if (copy_from_user(&cmd, buf, sizeof cmd))
 849                 return -EFAULT;
 850
 851         mutex_lock(&file->device->xrcd_tree_mutex);
 852         uobj = idr_write_uobj(&ib_uverbs_xrcd_idr, cmd.xrcd_handle, file->ucontext);
 853         if (!uobj) {
 854                 ret = -EINVAL;
 855                 goto out;
 856         }
 857
 858         xrcd  = uobj->object;
 859         inode = xrcd->inode;
 860         obj   = container_of(uobj, struct ib_uxrcd_object, uobject);
 861         if (atomic_read(&obj->refcnt)) {
 862                 put_uobj_write(uobj);
 863                 ret = -EBUSY;
 864                 goto out;
 865         }
 866
 867         if (!inode || atomic_dec_and_test(&xrcd->usecnt)) {
 868                 ret = ib_dealloc_xrcd(uobj->object);
 869                 if (!ret)
 870                         uobj->live = 0;
 871         }
 872
 873         live = uobj->live;
 874         if (inode && ret)
 875                 atomic_inc(&xrcd->usecnt);
 876
 877         put_uobj_write(uobj);
 878
 879         if (ret)
 880                 goto out;
 881
 882         if (inode && !live)
 883                 xrcd_table_delete(file->device, inode);
 884
 885         idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 886         mutex_lock(&file->mutex);
 887         list_del(&uobj->list);
 888         mutex_unlock(&file->mutex);
 889
 890         put_uobj(uobj);
 891         ret = in_len;
 892
 893 out:
 894         mutex_unlock(&file->device->xrcd_tree_mutex);
 895         return ret;
 896 }
 897
 898 void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 899                             struct ib_xrcd *xrcd)
 900 {
 901         struct inode *inode;
 902
 903         inode = xrcd->inode;
 904         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 905                 return;
 906
 907         ib_dealloc_xrcd(xrcd);
 908
 909         if (inode)
 910                 xrcd_table_delete(dev, inode);
 911 }
 912
 913 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 914                          const char __user *buf, int in_len,
 915                          int out_len)
 916 {
 917         struct ib_uverbs_reg_mr      cmd;
 918         struct ib_uverbs_reg_mr_resp resp;
 919         struct ib_udata              udata;
 920         struct ib_uobject           *uobj;
 921         struct ib_pd                *pd;
 922         struct ib_mr                *mr;
 923         int                          ret;
 924
 925         if (out_len < sizeof resp)
 926                 return -ENOSPC;
 927
 928         if (copy_from_user(&cmd, buf, sizeof cmd))
 929                 return -EFAULT;
 930
 931         INIT_UDATA(&udata, buf + sizeof cmd,
 932                    (unsigned long) cmd.response + sizeof resp,
 933                    in_len - sizeof cmd, out_len - sizeof resp);
 934
 935         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 936                 return -EINVAL;
 937
 938         /*
 939          * Local write permission is required if remote write or
 940          * remote atomic permission is also requested.
 941          */
 942         if (cmd.access_flags & (IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_REMOTE_WRITE) &&
 943             !(cmd.access_flags & IB_ACCESS_LOCAL_WRITE))
 944                 return -EINVAL;
 945
 946         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 947         if (!uobj)
 948                 return -ENOMEM;
 949
 950         init_uobj(uobj, 0, file->ucontext, &mr_lock_class);
 951         down_write(&uobj->mutex);
 952
 953         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
 954         if (!pd) {
 955                 ret = -EINVAL;
 956                 goto err_free;
 957         }
 958
 959         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 960                                      cmd.access_flags, &udata);
 961         if (IS_ERR(mr)) {
 962                 ret = PTR_ERR(mr);
 963                 goto err_put;
 964         }
 965
 966         mr->device  = pd->device;
 967         mr->pd      = pd;
 968         mr->uobject = uobj;
 969         atomic_inc(&pd->usecnt);
 970         atomic_set(&mr->usecnt, 0);
 971
 972         uobj->object = mr;
 973         ret = idr_add_uobj(&ib_uverbs_mr_idr, uobj);
 974         if (ret)
 975                 goto err_unreg;
 976
 977         memset(&resp, 0, sizeof resp);
 978         resp.lkey      = mr->lkey;
 979         resp.rkey      = mr->rkey;
 980         resp.mr_handle = uobj->id;
 981
 982         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 983                          &resp, sizeof resp)) {
 984                 ret = -EFAULT;
 985                 goto err_copy;
 986         }
 987
 988         put_pd_read(pd);
 989
 990         mutex_lock(&file->mutex);
 991         list_add_tail(&uobj->list, &file->ucontext->mr_list);
 992         mutex_unlock(&file->mutex);
 993
 994         uobj->live = 1;
 995
 996         up_write(&uobj->mutex);
 997
 998         return in_len;
 999
1000 err_copy:
1001         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1002
1003 err_unreg:
1004         ib_dereg_mr(mr);
1005
1006 err_put:
1007         put_pd_read(pd);
1008
1009 err_free:
1010         put_uobj_write(uobj);
1011         return ret;
1012 }
1013
1014 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
1015                            const char __user *buf, int in_len,
1016                            int out_len)
1017 {
1018         struct ib_uverbs_dereg_mr cmd;
1019         struct ib_mr             *mr;
1020         struct ib_uobject        *uobj;
1021         int                       ret = -EINVAL;
1022
1023         if (copy_from_user(&cmd, buf, sizeof cmd))
1024                 return -EFAULT;
1025
1026         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle, file->ucontext);
1027         if (!uobj)
1028                 return -EINVAL;
1029
1030         mr = uobj->object;
1031
1032         ret = ib_dereg_mr(mr);
1033         if (!ret)
1034                 uobj->live = 0;
1035
1036         put_uobj_write(uobj);
1037
1038         if (ret)
1039                 return ret;
1040
1041         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1042
1043         mutex_lock(&file->mutex);
1044         list_del(&uobj->list);
1045         mutex_unlock(&file->mutex);
1046
1047         put_uobj(uobj);
1048
1049         return in_len;
1050 }
1051
1052 ssize_t ib_uverbs_alloc_mw(struct ib_uverbs_file *file,
1053                          const char __user *buf, int in_len,
1054                          int out_len)
1055 {
1056         struct ib_uverbs_alloc_mw      cmd;
1057         struct ib_uverbs_alloc_mw_resp resp;
1058         struct ib_uobject             *uobj;
1059         struct ib_pd                  *pd;
1060         struct ib_mw                  *mw;
1061         int                            ret;
1062
1063         if (out_len < sizeof(resp))
1064                 return -ENOSPC;
1065
1066         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1067                 return -EFAULT;
1068
1069         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
1070         if (!uobj)
1071                 return -ENOMEM;
1072
1073         init_uobj(uobj, 0, file->ucontext, &mw_lock_class);
1074         down_write(&uobj->mutex);
1075
1076         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
1077         if (!pd) {
1078                 ret = -EINVAL;
1079                 goto err_free;
1080         }
1081
1082         mw = pd->device->alloc_mw(pd, cmd.mw_type);
1083         if (IS_ERR(mw)) {
1084                 ret = PTR_ERR(mw);
1085                 goto err_put;
1086         }
1087
1088         mw->device  = pd->device;
1089         mw->pd      = pd;
1090         mw->uobject = uobj;
1091         atomic_inc(&pd->usecnt);
1092
1093         uobj->object = mw;
1094         ret = idr_add_uobj(&ib_uverbs_mw_idr, uobj);
1095         if (ret)
1096                 goto err_unalloc;
1097
1098         memset(&resp, 0, sizeof(resp));
1099         resp.rkey      = mw->rkey;
1100         resp.mw_handle = uobj->id;
1101
1102         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1103                          &resp, sizeof(resp))) {
1104                 ret = -EFAULT;
1105                 goto err_copy;
1106         }
1107
1108         put_pd_read(pd);
1109
1110         mutex_lock(&file->mutex);
1111         list_add_tail(&uobj->list, &file->ucontext->mw_list);
1112         mutex_unlock(&file->mutex);
1113
1114         uobj->live = 1;
1115
1116         up_write(&uobj->mutex);
1117
1118         return in_len;
1119
1120 err_copy:
1121         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1122
1123 err_unalloc:
1124         ib_dealloc_mw(mw);
1125
1126 err_put:
1127         put_pd_read(pd);
1128
1129 err_free:
1130         put_uobj_write(uobj);
1131         return ret;
1132 }
1133
1134 ssize_t ib_uverbs_dealloc_mw(struct ib_uverbs_file *file,
1135                            const char __user *buf, int in_len,
1136                            int out_len)
1137 {
1138         struct ib_uverbs_dealloc_mw cmd;
1139         struct ib_mw               *mw;
1140         struct ib_uobject          *uobj;
1141         int                         ret = -EINVAL;
1142
1143         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1144                 return -EFAULT;
1145
1146         uobj = idr_write_uobj(&ib_uverbs_mw_idr, cmd.mw_handle, file->ucontext);
1147         if (!uobj)
1148                 return -EINVAL;
1149
1150         mw = uobj->object;
1151
1152         ret = ib_dealloc_mw(mw);
1153         if (!ret)
1154                 uobj->live = 0;
1155
1156         put_uobj_write(uobj);
1157
1158         if (ret)
1159                 return ret;
1160
1161         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1162
1163         mutex_lock(&file->mutex);
1164         list_del(&uobj->list);
1165         mutex_unlock(&file->mutex);
1166
1167         put_uobj(uobj);
1168
1169         return in_len;
1170 }
1171
1172 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
1173                                       const char __user *buf, int in_len,
1174                                       int out_len)
1175 {
1176         struct ib_uverbs_create_comp_channel       cmd;
1177         struct ib_uverbs_create_comp_channel_resp  resp;
1178         struct file                               *filp;
1179         int ret;
1180
1181         if (out_len < sizeof resp)
1182                 return -ENOSPC;
1183
1184         if (copy_from_user(&cmd, buf, sizeof cmd))
1185                 return -EFAULT;
1186
1187         ret = get_unused_fd_flags(O_CLOEXEC);
1188         if (ret < 0)
1189                 return ret;
1190         resp.fd = ret;
1191
1192         filp = ib_uverbs_alloc_event_file(file, 0);
1193         if (IS_ERR(filp)) {
1194                 put_unused_fd(resp.fd);
1195                 return PTR_ERR(filp);
1196         }
1197
1198         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1199                          &resp, sizeof resp)) {
1200                 put_unused_fd(resp.fd);
1201                 fput(filp);
1202                 return -EFAULT;
1203         }
1204
1205         fd_install(resp.fd, filp);
1206         return in_len;
1207 }
1208
1209 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1210                             const char __user *buf, int in_len,
1211                             int out_len)
1212 {
1213         struct ib_uverbs_create_cq      cmd;
1214         struct ib_uverbs_create_cq_resp resp;
1215         struct ib_udata                 udata;
1216         struct ib_ucq_object           *obj;
1217         struct ib_uverbs_event_file    *ev_file = NULL;
1218         struct ib_cq                   *cq;
1219         int                             ret;
1220
1221         if (out_len < sizeof resp)
1222                 return -ENOSPC;
1223
1224         if (copy_from_user(&cmd, buf, sizeof cmd))
1225                 return -EFAULT;
1226
1227         INIT_UDATA(&udata, buf + sizeof cmd,
1228                    (unsigned long) cmd.response + sizeof resp,
1229                    in_len - sizeof cmd, out_len - sizeof resp);
1230
1231         if (cmd.comp_vector >= file->device->num_comp_vectors)
1232                 return -EINVAL;
1233
1234         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1235         if (!obj)
1236                 return -ENOMEM;
1237
1238         init_uobj(&obj->uobject, cmd.user_handle, file->ucontext, &cq_lock_class);
1239         down_write(&obj->uobject.mutex);
1240
1241         if (cmd.comp_channel >= 0) {
1242                 ev_file = ib_uverbs_lookup_comp_file(cmd.comp_channel);
1243                 if (!ev_file) {
1244                         ret = -EINVAL;
1245                         goto err;
1246                 }
1247         }
1248
1249         obj->uverbs_file           = file;
1250         obj->comp_events_reported  = 0;
1251         obj->async_events_reported = 0;
1252         INIT_LIST_HEAD(&obj->comp_list);
1253         INIT_LIST_HEAD(&obj->async_list);
1254
1255         cq = file->device->ib_dev->create_cq(file->device->ib_dev, cmd.cqe,
1256                                              cmd.comp_vector,
1257                                              file->ucontext, &udata);
1258         if (IS_ERR(cq)) {
1259                 ret = PTR_ERR(cq);
1260                 goto err_file;
1261         }
1262
1263         cq->device        = file->device->ib_dev;
1264         cq->uobject       = &obj->uobject;
1265         cq->comp_handler  = ib_uverbs_comp_handler;
1266         cq->event_handler = ib_uverbs_cq_event_handler;
1267         cq->cq_context    = ev_file;
1268         atomic_set(&cq->usecnt, 0);
1269
1270         obj->uobject.object = cq;
1271         ret = idr_add_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1272         if (ret)
1273                 goto err_free;
1274
1275         memset(&resp, 0, sizeof resp);
1276         resp.cq_handle = obj->uobject.id;
1277         resp.cqe       = cq->cqe;
1278
1279         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1280                          &resp, sizeof resp)) {
1281                 ret = -EFAULT;
1282                 goto err_copy;
1283         }
1284
1285         mutex_lock(&file->mutex);
1286         list_add_tail(&obj->uobject.list, &file->ucontext->cq_list);
1287         mutex_unlock(&file->mutex);
1288
1289         obj->uobject.live = 1;
1290
1291         up_write(&obj->uobject.mutex);
1292
1293         return in_len;
1294
1295 err_copy:
1296         idr_remove_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1297
1298 err_free:
1299         ib_destroy_cq(cq);
1300
1301 err_file:
1302         if (ev_file)
1303                 ib_uverbs_release_ucq(file, ev_file, obj);
1304
1305 err:
1306         put_uobj_write(&obj->uobject);
1307         return ret;
1308 }
1309
1310 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1311                             const char __user *buf, int in_len,
1312                             int out_len)
1313 {
1314         struct ib_uverbs_resize_cq      cmd;
1315         struct ib_uverbs_resize_cq_resp resp;
1316         struct ib_udata                 udata;
1317         struct ib_cq                    *cq;
1318         int                             ret = -EINVAL;
1319
1320         if (copy_from_user(&cmd, buf, sizeof cmd))
1321                 return -EFAULT;
1322
1323         INIT_UDATA(&udata, buf + sizeof cmd,
1324                    (unsigned long) cmd.response + sizeof resp,
1325                    in_len - sizeof cmd, out_len - sizeof resp);
1326
1327         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1328         if (!cq)
1329                 return -EINVAL;
1330
1331         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1332         if (ret)
1333                 goto out;
1334
1335         resp.cqe = cq->cqe;
1336
1337         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1338                          &resp, sizeof resp.cqe))
1339                 ret = -EFAULT;
1340
1341 out:
1342         put_cq_read(cq);
1343
1344         return ret ? ret : in_len;
1345 }
1346
1347 static int copy_wc_to_user(void __user *dest, struct ib_wc *wc)
1348 {
1349         struct ib_uverbs_wc tmp;
1350
1351         tmp.wr_id               = wc->wr_id;
1352         tmp.status              = wc->status;
1353         tmp.opcode              = wc->opcode;
1354         tmp.vendor_err          = wc->vendor_err;
1355         tmp.byte_len            = wc->byte_len;
1356         tmp.ex.imm_data         = (__u32 __force) wc->ex.imm_data;
1357         tmp.qp_num              = wc->qp->qp_num;
1358         tmp.src_qp              = wc->src_qp;
1359         tmp.wc_flags            = wc->wc_flags;
1360         tmp.pkey_index          = wc->pkey_index;
1361         tmp.slid                = wc->slid;
1362         tmp.sl                  = wc->sl;
1363         tmp.dlid_path_bits      = wc->dlid_path_bits;
1364         tmp.port_num            = wc->port_num;
1365         tmp.reserved            = 0;
1366
1367         if (copy_to_user(dest, &tmp, sizeof tmp))
1368                 return -EFAULT;
1369
1370         return 0;
1371 }
1372
1373 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1374                           const char __user *buf, int in_len,
1375                           int out_len)
1376 {
1377         struct ib_uverbs_poll_cq       cmd;
1378         struct ib_uverbs_poll_cq_resp  resp;
1379         u8 __user                     *header_ptr;
1380         u8 __user                     *data_ptr;
1381         struct ib_cq                  *cq;
1382         struct ib_wc                   wc;
1383         int                            ret;
1384
1385         if (copy_from_user(&cmd, buf, sizeof cmd))
1386                 return -EFAULT;
1387
1388         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1389         if (!cq)
1390                 return -EINVAL;
1391
1392         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1393         header_ptr = (void __user *)(unsigned long) cmd.response;
1394         data_ptr = header_ptr + sizeof resp;
1395
1396         memset(&resp, 0, sizeof resp);
1397         while (resp.count < cmd.ne) {
1398                 ret = ib_poll_cq(cq, 1, &wc);
1399                 if (ret < 0)
1400                         goto out_put;
1401                 if (!ret)
1402                         break;
1403
1404                 ret = copy_wc_to_user(data_ptr, &wc);
1405                 if (ret)
1406                         goto out_put;
1407
1408                 data_ptr += sizeof(struct ib_uverbs_wc);
1409                 ++resp.count;
1410         }
1411
1412         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1413                 ret = -EFAULT;
1414                 goto out_put;
1415         }
1416
1417         ret = in_len;
1418
1419 out_put:
1420         put_cq_read(cq);
1421         return ret;
1422 }
1423
1424 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1425                                 const char __user *buf, int in_len,
1426                                 int out_len)
1427 {
1428         struct ib_uverbs_req_notify_cq cmd;
1429         struct ib_cq                  *cq;
1430
1431         if (copy_from_user(&cmd, buf, sizeof cmd))
1432                 return -EFAULT;
1433
1434         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1435         if (!cq)
1436                 return -EINVAL;
1437
1438         ib_req_notify_cq(cq, cmd.solicited_only ?
1439                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1440
1441         put_cq_read(cq);
1442
1443         return in_len;
1444 }
1445
1446 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1447                              const char __user *buf, int in_len,
1448                              int out_len)
1449 {
1450         struct ib_uverbs_destroy_cq      cmd;
1451         struct ib_uverbs_destroy_cq_resp resp;
1452         struct ib_uobject               *uobj;
1453         struct ib_cq                    *cq;
1454         struct ib_ucq_object            *obj;
1455         struct ib_uverbs_event_file     *ev_file;
1456         int                              ret = -EINVAL;
1457
1458         if (copy_from_user(&cmd, buf, sizeof cmd))
1459                 return -EFAULT;
1460
1461         uobj = idr_write_uobj(&ib_uverbs_cq_idr, cmd.cq_handle, file->ucontext);
1462         if (!uobj)
1463                 return -EINVAL;
1464         cq      = uobj->object;
1465         ev_file = cq->cq_context;
1466         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1467
1468         ret = ib_destroy_cq(cq);
1469         if (!ret)
1470                 uobj->live = 0;
1471
1472         put_uobj_write(uobj);
1473
1474         if (ret)
1475                 return ret;
1476
1477         idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
1478
1479         mutex_lock(&file->mutex);
1480         list_del(&uobj->list);
1481         mutex_unlock(&file->mutex);
1482
1483         ib_uverbs_release_ucq(file, ev_file, obj);
1484
1485         memset(&resp, 0, sizeof resp);
1486         resp.comp_events_reported  = obj->comp_events_reported;
1487         resp.async_events_reported = obj->async_events_reported;
1488
1489         put_uobj(uobj);
1490
1491         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1492                          &resp, sizeof resp))
1493                 return -EFAULT;
1494
1495         return in_len;
1496 }
1497
1498 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1499                             const char __user *buf, int in_len,
1500                             int out_len)
1501 {
1502         struct ib_uverbs_create_qp      cmd;
1503         struct ib_uverbs_create_qp_resp resp;
1504         struct ib_udata                 udata;
1505         struct ib_uqp_object           *obj;
1506         struct ib_device               *device;
1507         struct ib_pd                   *pd = NULL;
1508         struct ib_xrcd                 *xrcd = NULL;
1509         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1510         struct ib_cq                   *scq = NULL, *rcq = NULL;
1511         struct ib_srq                  *srq = NULL;
1512         struct ib_qp                   *qp;
1513         struct ib_qp_init_attr          attr;
1514         int ret;
1515
1516         if (out_len < sizeof resp)
1517                 return -ENOSPC;
1518
1519         if (copy_from_user(&cmd, buf, sizeof cmd))
1520                 return -EFAULT;
1521
1522         if (cmd.qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1523                 return -EPERM;
1524
1525         INIT_UDATA(&udata, buf + sizeof cmd,
1526                    (unsigned long) cmd.response + sizeof resp,
1527                    in_len - sizeof cmd, out_len - sizeof resp);
1528
1529         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1530         if (!obj)
1531                 return -ENOMEM;
1532
1533         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1534         down_write(&obj->uevent.uobject.mutex);
1535
1536         if (cmd.qp_type == IB_QPT_XRC_TGT) {
1537                 xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1538                 if (!xrcd) {
1539                         ret = -EINVAL;
1540                         goto err_put;
1541                 }
1542                 device = xrcd->device;
1543         } else {
1544                 if (cmd.qp_type == IB_QPT_XRC_INI) {
1545                         cmd.max_recv_wr = cmd.max_recv_sge = 0;
1546                 } else {
1547                         if (cmd.is_srq) {
1548                                 srq = idr_read_srq(cmd.srq_handle, file->ucontext);
1549                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1550                                         ret = -EINVAL;
1551                                         goto err_put;
1552                                 }
1553                         }
1554
1555                         if (cmd.recv_cq_handle != cmd.send_cq_handle) {
1556                                 rcq = idr_read_cq(cmd.recv_cq_handle, file->ucontext, 0);
1557                                 if (!rcq) {
1558                                         ret = -EINVAL;
1559                                         goto err_put;
1560                                 }
1561                         }
1562                 }
1563
1564                 scq = idr_read_cq(cmd.send_cq_handle, file->ucontext, !!rcq);
1565                 rcq = rcq ?: scq;
1566                 pd  = idr_read_pd(cmd.pd_handle, file->ucontext);
1567                 if (!pd || !scq) {
1568                         ret = -EINVAL;
1569                         goto err_put;
1570                 }
1571
1572                 device = pd->device;
1573         }
1574
1575         attr.event_handler = ib_uverbs_qp_event_handler;
1576         attr.qp_context    = file;
1577         attr.send_cq       = scq;
1578         attr.recv_cq       = rcq;
1579         attr.srq           = srq;
1580         attr.xrcd          = xrcd;
1581         attr.sq_sig_type   = cmd.sq_sig_all ? IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
1582         attr.qp_type       = cmd.qp_type;
1583         attr.create_flags  = 0;
1584
1585         attr.cap.max_send_wr     = cmd.max_send_wr;
1586         attr.cap.max_recv_wr     = cmd.max_recv_wr;
1587         attr.cap.max_send_sge    = cmd.max_send_sge;
1588         attr.cap.max_recv_sge    = cmd.max_recv_sge;
1589         attr.cap.max_inline_data = cmd.max_inline_data;
1590
1591         obj->uevent.events_reported     = 0;
1592         INIT_LIST_HEAD(&obj->uevent.event_list);
1593         INIT_LIST_HEAD(&obj->mcast_list);
1594
1595         if (cmd.qp_type == IB_QPT_XRC_TGT)
1596                 qp = ib_create_qp(pd, &attr);
1597         else
1598                 qp = device->create_qp(pd, &attr, &udata);
1599
1600         if (IS_ERR(qp)) {
1601                 ret = PTR_ERR(qp);
1602                 goto err_put;
1603         }
1604
1605         if (cmd.qp_type != IB_QPT_XRC_TGT) {
1606                 qp->real_qp       = qp;
1607                 qp->device        = device;
1608                 qp->pd            = pd;
1609                 qp->send_cq       = attr.send_cq;
1610                 qp->recv_cq       = attr.recv_cq;
1611                 qp->srq           = attr.srq;
1612                 qp->event_handler = attr.event_handler;
1613                 qp->qp_context    = attr.qp_context;
1614                 qp->qp_type       = attr.qp_type;
1615                 atomic_set(&qp->usecnt, 0);
1616                 atomic_inc(&pd->usecnt);
1617                 atomic_inc(&attr.send_cq->usecnt);
1618                 if (attr.recv_cq)
1619                         atomic_inc(&attr.recv_cq->usecnt);
1620                 if (attr.srq)
1621                         atomic_inc(&attr.srq->usecnt);
1622         }
1623         qp->uobject = &obj->uevent.uobject;
1624
1625         obj->uevent.uobject.object = qp;
1626         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1627         if (ret)
1628                 goto err_destroy;
1629
1630         memset(&resp, 0, sizeof resp);
1631         resp.qpn             = qp->qp_num;
1632         resp.qp_handle       = obj->uevent.uobject.id;
1633         resp.max_recv_sge    = attr.cap.max_recv_sge;
1634         resp.max_send_sge    = attr.cap.max_send_sge;
1635         resp.max_recv_wr     = attr.cap.max_recv_wr;
1636         resp.max_send_wr     = attr.cap.max_send_wr;
1637         resp.max_inline_data = attr.cap.max_inline_data;
1638
1639         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1640                          &resp, sizeof resp)) {
1641                 ret = -EFAULT;
1642                 goto err_copy;
1643         }
1644
1645         if (xrcd)
1646                 put_xrcd_read(xrcd_uobj);
1647         if (pd)
1648                 put_pd_read(pd);
1649         if (scq)
1650                 put_cq_read(scq);
1651         if (rcq && rcq != scq)
1652                 put_cq_read(rcq);
1653         if (srq)
1654                 put_srq_read(srq);
1655
1656         mutex_lock(&file->mutex);
1657         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1658         mutex_unlock(&file->mutex);
1659
1660         obj->uevent.uobject.live = 1;
1661
1662         up_write(&obj->uevent.uobject.mutex);
1663
1664         return in_len;
1665
1666 err_copy:
1667         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1668
1669 err_destroy:
1670         ib_destroy_qp(qp);
1671
1672 err_put:
1673         if (xrcd)
1674                 put_xrcd_read(xrcd_uobj);
1675         if (pd)
1676                 put_pd_read(pd);
1677         if (scq)
1678                 put_cq_read(scq);
1679         if (rcq && rcq != scq)
1680                 put_cq_read(rcq);
1681         if (srq)
1682                 put_srq_read(srq);
1683
1684         put_uobj_write(&obj->uevent.uobject);
1685         return ret;
1686 }
1687
1688 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1689                           const char __user *buf, int in_len, int out_len)
1690 {
1691         struct ib_uverbs_open_qp        cmd;
1692         struct ib_uverbs_create_qp_resp resp;
1693         struct ib_udata                 udata;
1694         struct ib_uqp_object           *obj;
1695         struct ib_xrcd                 *xrcd;
1696         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1697         struct ib_qp                   *qp;
1698         struct ib_qp_open_attr          attr;
1699         int ret;
1700
1701         if (out_len < sizeof resp)
1702                 return -ENOSPC;
1703
1704         if (copy_from_user(&cmd, buf, sizeof cmd))
1705                 return -EFAULT;
1706
1707         INIT_UDATA(&udata, buf + sizeof cmd,
1708                    (unsigned long) cmd.response + sizeof resp,
1709                    in_len - sizeof cmd, out_len - sizeof resp);
1710
1711         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1712         if (!obj)
1713                 return -ENOMEM;
1714
1715         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1716         down_write(&obj->uevent.uobject.mutex);
1717
1718         xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1719         if (!xrcd) {
1720                 ret = -EINVAL;
1721                 goto err_put;
1722         }
1723
1724         attr.event_handler = ib_uverbs_qp_event_handler;
1725         attr.qp_context    = file;
1726         attr.qp_num        = cmd.qpn;
1727         attr.qp_type       = cmd.qp_type;
1728
1729         obj->uevent.events_reported = 0;
1730         INIT_LIST_HEAD(&obj->uevent.event_list);
1731         INIT_LIST_HEAD(&obj->mcast_list);
1732
1733         qp = ib_open_qp(xrcd, &attr);
1734         if (IS_ERR(qp)) {
1735                 ret = PTR_ERR(qp);
1736                 goto err_put;
1737         }
1738
1739         qp->uobject = &obj->uevent.uobject;
1740
1741         obj->uevent.uobject.object = qp;
1742         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1743         if (ret)
1744                 goto err_destroy;
1745
1746         memset(&resp, 0, sizeof resp);
1747         resp.qpn       = qp->qp_num;
1748         resp.qp_handle = obj->uevent.uobject.id;
1749
1750         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1751                          &resp, sizeof resp)) {
1752                 ret = -EFAULT;
1753                 goto err_remove;
1754         }
1755
1756         put_xrcd_read(xrcd_uobj);
1757
1758         mutex_lock(&file->mutex);
1759         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1760         mutex_unlock(&file->mutex);
1761
1762         obj->uevent.uobject.live = 1;
1763
1764         up_write(&obj->uevent.uobject.mutex);
1765
1766         return in_len;
1767
1768 err_remove:
1769         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1770
1771 err_destroy:
1772         ib_destroy_qp(qp);
1773
1774 err_put:
1775         put_xrcd_read(xrcd_uobj);
1776         put_uobj_write(&obj->uevent.uobject);
1777         return ret;
1778 }
1779
1780 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1781                            const char __user *buf, int in_len,
1782                            int out_len)
1783 {
1784         struct ib_uverbs_query_qp      cmd;
1785         struct ib_uverbs_query_qp_resp resp;
1786         struct ib_qp                   *qp;
1787         struct ib_qp_attr              *attr;
1788         struct ib_qp_init_attr         *init_attr;
1789         int                            ret;
1790
1791         if (copy_from_user(&cmd, buf, sizeof cmd))
1792                 return -EFAULT;
1793
1794         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1795         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1796         if (!attr || !init_attr) {
1797                 ret = -ENOMEM;
1798                 goto out;
1799         }
1800
1801         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1802         if (!qp) {
1803                 ret = -EINVAL;
1804                 goto out;
1805         }
1806
1807         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1808
1809         put_qp_read(qp);
1810
1811         if (ret)
1812                 goto out;
1813
1814         memset(&resp, 0, sizeof resp);
1815
1816         resp.qp_state               = attr->qp_state;
1817         resp.cur_qp_state           = attr->cur_qp_state;
1818         resp.path_mtu               = attr->path_mtu;
1819         resp.path_mig_state         = attr->path_mig_state;
1820         resp.qkey                   = attr->qkey;
1821         resp.rq_psn                 = attr->rq_psn;
1822         resp.sq_psn                 = attr->sq_psn;
1823         resp.dest_qp_num            = attr->dest_qp_num;
1824         resp.qp_access_flags        = attr->qp_access_flags;
1825         resp.pkey_index             = attr->pkey_index;
1826         resp.alt_pkey_index         = attr->alt_pkey_index;
1827         resp.sq_draining            = attr->sq_draining;
1828         resp.max_rd_atomic          = attr->max_rd_atomic;
1829         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1830         resp.min_rnr_timer          = attr->min_rnr_timer;
1831         resp.port_num               = attr->port_num;
1832         resp.timeout                = attr->timeout;
1833         resp.retry_cnt              = attr->retry_cnt;
1834         resp.rnr_retry              = attr->rnr_retry;
1835         resp.alt_port_num           = attr->alt_port_num;
1836         resp.alt_timeout            = attr->alt_timeout;
1837
1838         memcpy(resp.dest.dgid, attr->ah_attr.grh.dgid.raw, 16);
1839         resp.dest.flow_label        = attr->ah_attr.grh.flow_label;
1840         resp.dest.sgid_index        = attr->ah_attr.grh.sgid_index;
1841         resp.dest.hop_limit         = attr->ah_attr.grh.hop_limit;
1842         resp.dest.traffic_class     = attr->ah_attr.grh.traffic_class;
1843         resp.dest.dlid              = attr->ah_attr.dlid;
1844         resp.dest.sl                = attr->ah_attr.sl;
1845         resp.dest.src_path_bits     = attr->ah_attr.src_path_bits;
1846         resp.dest.static_rate       = attr->ah_attr.static_rate;
1847         resp.dest.is_global         = !!(attr->ah_attr.ah_flags & IB_AH_GRH);
1848         resp.dest.port_num          = attr->ah_attr.port_num;
1849
1850         memcpy(resp.alt_dest.dgid, attr->alt_ah_attr.grh.dgid.raw, 16);
1851         resp.alt_dest.flow_label    = attr->alt_ah_attr.grh.flow_label;
1852         resp.alt_dest.sgid_index    = attr->alt_ah_attr.grh.sgid_index;
1853         resp.alt_dest.hop_limit     = attr->alt_ah_attr.grh.hop_limit;
1854         resp.alt_dest.traffic_class = attr->alt_ah_attr.grh.traffic_class;
1855         resp.alt_dest.dlid          = attr->alt_ah_attr.dlid;
1856         resp.alt_dest.sl            = attr->alt_ah_attr.sl;
1857         resp.alt_dest.src_path_bits = attr->alt_ah_attr.src_path_bits;
1858         resp.alt_dest.static_rate   = attr->alt_ah_attr.static_rate;
1859         resp.alt_dest.is_global     = !!(attr->alt_ah_attr.ah_flags & IB_AH_GRH);
1860         resp.alt_dest.port_num      = attr->alt_ah_attr.port_num;
1861
1862         resp.max_send_wr            = init_attr->cap.max_send_wr;
1863         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1864         resp.max_send_sge           = init_attr->cap.max_send_sge;
1865         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1866         resp.max_inline_data        = init_attr->cap.max_inline_data;
1867         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1868
1869         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1870                          &resp, sizeof resp))
1871                 ret = -EFAULT;
1872
1873 out:
1874         kfree(attr);
1875         kfree(init_attr);
1876
1877         return ret ? ret : in_len;
1878 }
1879
1880 /* Remove ignored fields set in the attribute mask */
1881 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1882 {
1883         switch (qp_type) {
1884         case IB_QPT_XRC_INI:
1885                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1886         case IB_QPT_XRC_TGT:
1887                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1888                                 IB_QP_RNR_RETRY);
1889         default:
1890                 return mask;
1891         }
1892 }
1893
1894 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
1895                             const char __user *buf, int in_len,
1896                             int out_len)
1897 {
1898         struct ib_uverbs_modify_qp cmd;
1899         struct ib_udata            udata;
1900         struct ib_qp              *qp;
1901         struct ib_qp_attr         *attr;
1902         int                        ret;
1903
1904         if (copy_from_user(&cmd, buf, sizeof cmd))
1905                 return -EFAULT;
1906
1907         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
1908                    out_len);
1909
1910         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1911         if (!attr)
1912                 return -ENOMEM;
1913
1914         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1915         if (!qp) {
1916                 ret = -EINVAL;
1917                 goto out;
1918         }
1919
1920         attr->qp_state            = cmd.qp_state;
1921         attr->cur_qp_state        = cmd.cur_qp_state;
1922         attr->path_mtu            = cmd.path_mtu;
1923         attr->path_mig_state      = cmd.path_mig_state;
1924         attr->qkey                = cmd.qkey;
1925         attr->rq_psn              = cmd.rq_psn;
1926         attr->sq_psn              = cmd.sq_psn;
1927         attr->dest_qp_num         = cmd.dest_qp_num;
1928         attr->qp_access_flags     = cmd.qp_access_flags;
1929         attr->pkey_index          = cmd.pkey_index;
1930         attr->alt_pkey_index      = cmd.alt_pkey_index;
1931         attr->en_sqd_async_notify = cmd.en_sqd_async_notify;
1932         attr->max_rd_atomic       = cmd.max_rd_atomic;
1933         attr->max_dest_rd_atomic  = cmd.max_dest_rd_atomic;
1934         attr->min_rnr_timer       = cmd.min_rnr_timer;
1935         attr->port_num            = cmd.port_num;
1936         attr->timeout             = cmd.timeout;
1937         attr->retry_cnt           = cmd.retry_cnt;
1938         attr->rnr_retry           = cmd.rnr_retry;
1939         attr->alt_port_num        = cmd.alt_port_num;
1940         attr->alt_timeout         = cmd.alt_timeout;
1941
1942         memcpy(attr->ah_attr.grh.dgid.raw, cmd.dest.dgid, 16);
1943         attr->ah_attr.grh.flow_label        = cmd.dest.flow_label;
1944         attr->ah_attr.grh.sgid_index        = cmd.dest.sgid_index;
1945         attr->ah_attr.grh.hop_limit         = cmd.dest.hop_limit;
1946         attr->ah_attr.grh.traffic_class     = cmd.dest.traffic_class;
1947         attr->ah_attr.dlid                  = cmd.dest.dlid;
1948         attr->ah_attr.sl                    = cmd.dest.sl;
1949         attr->ah_attr.src_path_bits         = cmd.dest.src_path_bits;
1950         attr->ah_attr.static_rate           = cmd.dest.static_rate;
1951         attr->ah_attr.ah_flags              = cmd.dest.is_global ? IB_AH_GRH : 0;
1952         attr->ah_attr.port_num              = cmd.dest.port_num;
1953
1954         memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd.alt_dest.dgid, 16);
1955         attr->alt_ah_attr.grh.flow_label    = cmd.alt_dest.flow_label;
1956         attr->alt_ah_attr.grh.sgid_index    = cmd.alt_dest.sgid_index;
1957         attr->alt_ah_attr.grh.hop_limit     = cmd.alt_dest.hop_limit;
1958         attr->alt_ah_attr.grh.traffic_class = cmd.alt_dest.traffic_class;
1959         attr->alt_ah_attr.dlid              = cmd.alt_dest.dlid;
1960         attr->alt_ah_attr.sl                = cmd.alt_dest.sl;
1961         attr->alt_ah_attr.src_path_bits     = cmd.alt_dest.src_path_bits;
1962         attr->alt_ah_attr.static_rate       = cmd.alt_dest.static_rate;
1963         attr->alt_ah_attr.ah_flags          = cmd.alt_dest.is_global ? IB_AH_GRH : 0;
1964         attr->alt_ah_attr.port_num          = cmd.alt_dest.port_num;
1965
1966         if (qp->real_qp == qp) {
1967                 ret = qp->device->modify_qp(qp, attr,
1968                         modify_qp_mask(qp->qp_type, cmd.attr_mask), &udata);
1969         } else {
1970                 ret = ib_modify_qp(qp, attr, modify_qp_mask(qp->qp_type, cmd.attr_mask));
1971         }
1972
1973         put_qp_read(qp);
1974
1975         if (ret)
1976                 goto out;
1977
1978         ret = in_len;
1979
1980 out:
1981         kfree(attr);
1982
1983         return ret;
1984 }
1985
1986 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
1987                              const char __user *buf, int in_len,
1988                              int out_len)
1989 {
1990         struct ib_uverbs_destroy_qp      cmd;
1991         struct ib_uverbs_destroy_qp_resp resp;
1992         struct ib_uobject               *uobj;
1993         struct ib_qp                    *qp;
1994         struct ib_uqp_object            *obj;
1995         int                              ret = -EINVAL;
1996
1997         if (copy_from_user(&cmd, buf, sizeof cmd))
1998                 return -EFAULT;
1999
2000         memset(&resp, 0, sizeof resp);
2001
2002         uobj = idr_write_uobj(&ib_uverbs_qp_idr, cmd.qp_handle, file->ucontext);
2003         if (!uobj)
2004                 return -EINVAL;
2005         qp  = uobj->object;
2006         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2007
2008         if (!list_empty(&obj->mcast_list)) {
2009                 put_uobj_write(uobj);
2010                 return -EBUSY;
2011         }
2012
2013         ret = ib_destroy_qp(qp);
2014         if (!ret)
2015                 uobj->live = 0;
2016
2017         put_uobj_write(uobj);
2018
2019         if (ret)
2020                 return ret;
2021
2022         idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
2023
2024         mutex_lock(&file->mutex);
2025         list_del(&uobj->list);
2026         mutex_unlock(&file->mutex);
2027
2028         ib_uverbs_release_uevent(file, &obj->uevent);
2029
2030         resp.events_reported = obj->uevent.events_reported;
2031
2032         put_uobj(uobj);
2033
2034         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2035                          &resp, sizeof resp))
2036                 return -EFAULT;
2037
2038         return in_len;
2039 }
2040
2041 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2042                             const char __user *buf, int in_len,
2043                             int out_len)
2044 {
2045         struct ib_uverbs_post_send      cmd;
2046         struct ib_uverbs_post_send_resp resp;
2047         struct ib_uverbs_send_wr       *user_wr;
2048         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
2049         struct ib_qp                   *qp;
2050         int                             i, sg_ind;
2051         int                             is_ud;
2052         ssize_t                         ret = -EINVAL;
2053
2054         if (copy_from_user(&cmd, buf, sizeof cmd))
2055                 return -EFAULT;
2056
2057         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2058             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2059                 return -EINVAL;
2060
2061         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2062                 return -EINVAL;
2063
2064         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2065         if (!user_wr)
2066                 return -ENOMEM;
2067
2068         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2069         if (!qp)
2070                 goto out;
2071
2072         is_ud = qp->qp_type == IB_QPT_UD;
2073         sg_ind = 0;
2074         last = NULL;
2075         for (i = 0; i < cmd.wr_count; ++i) {
2076                 if (copy_from_user(user_wr,
2077                                    buf + sizeof cmd + i * cmd.wqe_size,
2078                                    cmd.wqe_size)) {
2079                         ret = -EFAULT;
2080                         goto out_put;
2081                 }
2082
2083                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2084                         ret = -EINVAL;
2085                         goto out_put;
2086                 }
2087
2088                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2089                                user_wr->num_sge * sizeof (struct ib_sge),
2090                                GFP_KERNEL);
2091                 if (!next) {
2092                         ret = -ENOMEM;
2093                         goto out_put;
2094                 }
2095
2096                 if (!last)
2097                         wr = next;
2098                 else
2099                         last->next = next;
2100                 last = next;
2101
2102                 next->next       = NULL;
2103                 next->wr_id      = user_wr->wr_id;
2104                 next->num_sge    = user_wr->num_sge;
2105                 next->opcode     = user_wr->opcode;
2106                 next->send_flags = user_wr->send_flags;
2107
2108                 if (is_ud) {
2109                         next->wr.ud.ah = idr_read_ah(user_wr->wr.ud.ah,
2110                                                      file->ucontext);
2111                         if (!next->wr.ud.ah) {
2112                                 ret = -EINVAL;
2113                                 goto out_put;
2114                         }
2115                         next->wr.ud.remote_qpn  = user_wr->wr.ud.remote_qpn;
2116                         next->wr.ud.remote_qkey = user_wr->wr.ud.remote_qkey;
2117                 } else {
2118                         switch (next->opcode) {
2119                         case IB_WR_RDMA_WRITE_WITH_IMM:
2120                                 next->ex.imm_data =
2121                                         (__be32 __force) user_wr->ex.imm_data;
2122                         case IB_WR_RDMA_WRITE:
2123                         case IB_WR_RDMA_READ:
2124                                 next->wr.rdma.remote_addr =
2125                                         user_wr->wr.rdma.remote_addr;
2126                                 next->wr.rdma.rkey        =
2127                                         user_wr->wr.rdma.rkey;
2128                                 break;
2129                         case IB_WR_SEND_WITH_IMM:
2130                                 next->ex.imm_data =
2131                                         (__be32 __force) user_wr->ex.imm_data;
2132                                 break;
2133                         case IB_WR_SEND_WITH_INV:
2134                                 next->ex.invalidate_rkey =
2135                                         user_wr->ex.invalidate_rkey;
2136                                 break;
2137                         case IB_WR_ATOMIC_CMP_AND_SWP:
2138                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2139                                 next->wr.atomic.remote_addr =
2140                                         user_wr->wr.atomic.remote_addr;
2141                                 next->wr.atomic.compare_add =
2142                                         user_wr->wr.atomic.compare_add;
2143                                 next->wr.atomic.swap = user_wr->wr.atomic.swap;
2144                                 next->wr.atomic.rkey = user_wr->wr.atomic.rkey;
2145                                 break;
2146                         default:
2147                                 break;
2148                         }
2149                 }
2150
2151                 if (next->num_sge) {
2152                         next->sg_list = (void *) next +
2153                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2154                         if (copy_from_user(next->sg_list,
2155                                            buf + sizeof cmd +
2156                                            cmd.wr_count * cmd.wqe_size +
2157                                            sg_ind * sizeof (struct ib_sge),
2158                                            next->num_sge * sizeof (struct ib_sge))) {
2159                                 ret = -EFAULT;
2160                                 goto out_put;
2161                         }
2162                         sg_ind += next->num_sge;
2163                 } else
2164                         next->sg_list = NULL;
2165         }
2166
2167         resp.bad_wr = 0;
2168         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2169         if (ret)
2170                 for (next = wr; next; next = next->next) {
2171                         ++resp.bad_wr;
2172                         if (next == bad_wr)
2173                                 break;
2174                 }
2175
2176         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2177                          &resp, sizeof resp))
2178                 ret = -EFAULT;
2179
2180 out_put:
2181         put_qp_read(qp);
2182
2183         while (wr) {
2184                 if (is_ud && wr->wr.ud.ah)
2185                         put_ah_read(wr->wr.ud.ah);
2186                 next = wr->next;
2187                 kfree(wr);
2188                 wr = next;
2189         }
2190
2191 out:
2192         kfree(user_wr);
2193
2194         return ret ? ret : in_len;
2195 }
2196
2197 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2198                                                     int in_len,
2199                                                     u32 wr_count,
2200                                                     u32 sge_count,
2201                                                     u32 wqe_size)
2202 {
2203         struct ib_uverbs_recv_wr *user_wr;
2204         struct ib_recv_wr        *wr = NULL, *last, *next;
2205         int                       sg_ind;
2206         int                       i;
2207         int                       ret;
2208
2209         if (in_len < wqe_size * wr_count +
2210             sge_count * sizeof (struct ib_uverbs_sge))
2211                 return ERR_PTR(-EINVAL);
2212
2213         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2214                 return ERR_PTR(-EINVAL);
2215
2216         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2217         if (!user_wr)
2218                 return ERR_PTR(-ENOMEM);
2219
2220         sg_ind = 0;
2221         last = NULL;
2222         for (i = 0; i < wr_count; ++i) {
2223                 if (copy_from_user(user_wr, buf + i * wqe_size,
2224                                    wqe_size)) {
2225                         ret = -EFAULT;
2226                         goto err;
2227                 }
2228
2229                 if (user_wr->num_sge + sg_ind > sge_count) {
2230                         ret = -EINVAL;
2231                         goto err;
2232                 }
2233
2234                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2235                                user_wr->num_sge * sizeof (struct ib_sge),
2236                                GFP_KERNEL);
2237                 if (!next) {
2238                         ret = -ENOMEM;
2239                         goto err;
2240                 }
2241
2242                 if (!last)
2243                         wr = next;
2244                 else
2245                         last->next = next;
2246                 last = next;
2247
2248                 next->next       = NULL;
2249                 next->wr_id      = user_wr->wr_id;
2250                 next->num_sge    = user_wr->num_sge;
2251
2252                 if (next->num_sge) {
2253                         next->sg_list = (void *) next +
2254                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2255                         if (copy_from_user(next->sg_list,
2256                                            buf + wr_count * wqe_size +
2257                                            sg_ind * sizeof (struct ib_sge),
2258                                            next->num_sge * sizeof (struct ib_sge))) {
2259                                 ret = -EFAULT;
2260                                 goto err;
2261                         }
2262                         sg_ind += next->num_sge;
2263                 } else
2264                         next->sg_list = NULL;
2265         }
2266
2267         kfree(user_wr);
2268         return wr;
2269
2270 err:
2271         kfree(user_wr);
2272
2273         while (wr) {
2274                 next = wr->next;
2275                 kfree(wr);
2276                 wr = next;
2277         }
2278
2279         return ERR_PTR(ret);
2280 }
2281
2282 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2283                             const char __user *buf, int in_len,
2284                             int out_len)
2285 {
2286         struct ib_uverbs_post_recv      cmd;
2287         struct ib_uverbs_post_recv_resp resp;
2288         struct ib_recv_wr              *wr, *next, *bad_wr;
2289         struct ib_qp                   *qp;
2290         ssize_t                         ret = -EINVAL;
2291
2292         if (copy_from_user(&cmd, buf, sizeof cmd))
2293                 return -EFAULT;
2294
2295         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2296                                        in_len - sizeof cmd, cmd.wr_count,
2297                                        cmd.sge_count, cmd.wqe_size);
2298         if (IS_ERR(wr))
2299                 return PTR_ERR(wr);
2300
2301         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2302         if (!qp)
2303                 goto out;
2304
2305         resp.bad_wr = 0;
2306         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2307
2308         put_qp_read(qp);
2309
2310         if (ret)
2311                 for (next = wr; next; next = next->next) {
2312                         ++resp.bad_wr;
2313                         if (next == bad_wr)
2314                                 break;
2315                 }
2316
2317         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2318                          &resp, sizeof resp))
2319                 ret = -EFAULT;
2320
2321 out:
2322         while (wr) {
2323                 next = wr->next;
2324                 kfree(wr);
2325                 wr = next;
2326         }
2327
2328         return ret ? ret : in_len;
2329 }
2330
2331 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2332                                 const char __user *buf, int in_len,
2333                                 int out_len)
2334 {
2335         struct ib_uverbs_post_srq_recv      cmd;
2336         struct ib_uverbs_post_srq_recv_resp resp;
2337         struct ib_recv_wr                  *wr, *next, *bad_wr;
2338         struct ib_srq                      *srq;
2339         ssize_t                             ret = -EINVAL;
2340
2341         if (copy_from_user(&cmd, buf, sizeof cmd))
2342                 return -EFAULT;
2343
2344         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2345                                        in_len - sizeof cmd, cmd.wr_count,
2346                                        cmd.sge_count, cmd.wqe_size);
2347         if (IS_ERR(wr))
2348                 return PTR_ERR(wr);
2349
2350         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2351         if (!srq)
2352                 goto out;
2353
2354         resp.bad_wr = 0;
2355         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2356
2357         put_srq_read(srq);
2358
2359         if (ret)
2360                 for (next = wr; next; next = next->next) {
2361                         ++resp.bad_wr;
2362                         if (next == bad_wr)
2363                                 break;
2364                 }
2365
2366         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2367                          &resp, sizeof resp))
2368                 ret = -EFAULT;
2369
2370 out:
2371         while (wr) {
2372                 next = wr->next;
2373                 kfree(wr);
2374                 wr = next;
2375         }
2376
2377         return ret ? ret : in_len;
2378 }
2379
2380 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2381                             const char __user *buf, int in_len,
2382                             int out_len)
2383 {
2384         struct ib_uverbs_create_ah       cmd;
2385         struct ib_uverbs_create_ah_resp  resp;
2386         struct ib_uobject               *uobj;
2387         struct ib_pd                    *pd;
2388         struct ib_ah                    *ah;
2389         struct ib_ah_attr               attr;
2390         int ret;
2391
2392         if (out_len < sizeof resp)
2393                 return -ENOSPC;
2394
2395         if (copy_from_user(&cmd, buf, sizeof cmd))
2396                 return -EFAULT;
2397
2398         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
2399         if (!uobj)
2400                 return -ENOMEM;
2401
2402         init_uobj(uobj, cmd.user_handle, file->ucontext, &ah_lock_class);
2403         down_write(&uobj->mutex);
2404
2405         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
2406         if (!pd) {
2407                 ret = -EINVAL;
2408                 goto err;
2409         }
2410
2411         attr.dlid              = cmd.attr.dlid;
2412         attr.sl                = cmd.attr.sl;
2413         attr.src_path_bits     = cmd.attr.src_path_bits;
2414         attr.static_rate       = cmd.attr.static_rate;
2415         attr.ah_flags          = cmd.attr.is_global ? IB_AH_GRH : 0;
2416         attr.port_num          = cmd.attr.port_num;
2417         attr.grh.flow_label    = cmd.attr.grh.flow_label;
2418         attr.grh.sgid_index    = cmd.attr.grh.sgid_index;
2419         attr.grh.hop_limit     = cmd.attr.grh.hop_limit;
2420         attr.grh.traffic_class = cmd.attr.grh.traffic_class;
2421         memcpy(attr.grh.dgid.raw, cmd.attr.grh.dgid, 16);
2422
2423         ah = ib_create_ah(pd, &attr);
2424         if (IS_ERR(ah)) {
2425                 ret = PTR_ERR(ah);
2426                 goto err_put;
2427         }
2428
2429         ah->uobject  = uobj;
2430         uobj->object = ah;
2431
2432         ret = idr_add_uobj(&ib_uverbs_ah_idr, uobj);
2433         if (ret)
2434                 goto err_destroy;
2435
2436         resp.ah_handle = uobj->id;
2437
2438         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2439                          &resp, sizeof resp)) {
2440                 ret = -EFAULT;
2441                 goto err_copy;
2442         }
2443
2444         put_pd_read(pd);
2445
2446         mutex_lock(&file->mutex);
2447         list_add_tail(&uobj->list, &file->ucontext->ah_list);
2448         mutex_unlock(&file->mutex);
2449
2450         uobj->live = 1;
2451
2452         up_write(&uobj->mutex);
2453
2454         return in_len;
2455
2456 err_copy:
2457         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2458
2459 err_destroy:
2460         ib_destroy_ah(ah);
2461
2462 err_put:
2463         put_pd_read(pd);
2464
2465 err:
2466         put_uobj_write(uobj);
2467         return ret;
2468 }
2469
2470 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2471                              const char __user *buf, int in_len, int out_len)
2472 {
2473         struct ib_uverbs_destroy_ah cmd;
2474         struct ib_ah               *ah;
2475         struct ib_uobject          *uobj;
2476         int                         ret;
2477
2478         if (copy_from_user(&cmd, buf, sizeof cmd))
2479                 return -EFAULT;
2480
2481         uobj = idr_write_uobj(&ib_uverbs_ah_idr, cmd.ah_handle, file->ucontext);
2482         if (!uobj)
2483                 return -EINVAL;
2484         ah = uobj->object;
2485
2486         ret = ib_destroy_ah(ah);
2487         if (!ret)
2488                 uobj->live = 0;
2489
2490         put_uobj_write(uobj);
2491
2492         if (ret)
2493                 return ret;
2494
2495         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2496
2497         mutex_lock(&file->mutex);
2498         list_del(&uobj->list);
2499         mutex_unlock(&file->mutex);
2500
2501         put_uobj(uobj);
2502
2503         return in_len;
2504 }
2505
2506 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2507                                const char __user *buf, int in_len,
2508                                int out_len)
2509 {
2510         struct ib_uverbs_attach_mcast cmd;
2511         struct ib_qp                 *qp;
2512         struct ib_uqp_object         *obj;
2513         struct ib_uverbs_mcast_entry *mcast;
2514         int                           ret;
2515
2516         if (copy_from_user(&cmd, buf, sizeof cmd))
2517                 return -EFAULT;
2518
2519         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2520         if (!qp)
2521                 return -EINVAL;
2522
2523         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2524
2525         list_for_each_entry(mcast, &obj->mcast_list, list)
2526                 if (cmd.mlid == mcast->lid &&
2527                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2528                         ret = 0;
2529                         goto out_put;
2530                 }
2531
2532         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2533         if (!mcast) {
2534                 ret = -ENOMEM;
2535                 goto out_put;
2536         }
2537
2538         mcast->lid = cmd.mlid;
2539         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2540
2541         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2542         if (!ret)
2543                 list_add_tail(&mcast->list, &obj->mcast_list);
2544         else
2545                 kfree(mcast);
2546
2547 out_put:
2548         put_qp_write(qp);
2549
2550         return ret ? ret : in_len;
2551 }
2552
2553 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2554                                const char __user *buf, int in_len,
2555                                int out_len)
2556 {
2557         struct ib_uverbs_detach_mcast cmd;
2558         struct ib_uqp_object         *obj;
2559         struct ib_qp                 *qp;
2560         struct ib_uverbs_mcast_entry *mcast;
2561         int                           ret = -EINVAL;
2562
2563         if (copy_from_user(&cmd, buf, sizeof cmd))
2564                 return -EFAULT;
2565
2566         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2567         if (!qp)
2568                 return -EINVAL;
2569
2570         ret = ib_detach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
2571         if (ret)
2572                 goto out_put;
2573
2574         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2575
2576         list_for_each_entry(mcast, &obj->mcast_list, list)
2577                 if (cmd.mlid == mcast->lid &&
2578                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2579                         list_del(&mcast->list);
2580                         kfree(mcast);
2581                         break;
2582                 }
2583
2584 out_put:
2585         put_qp_write(qp);
2586
2587         return ret ? ret : in_len;
2588 }
2589
2590 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
2591                                 struct ib_uverbs_create_xsrq *cmd,
2592                                 struct ib_udata *udata)
2593 {
2594         struct ib_uverbs_create_srq_resp resp;
2595         struct ib_usrq_object           *obj;
2596         struct ib_pd                    *pd;
2597         struct ib_srq                   *srq;
2598         struct ib_uobject               *uninitialized_var(xrcd_uobj);
2599         struct ib_srq_init_attr          attr;
2600         int ret;
2601
2602         obj = kmalloc(sizeof *obj, GFP_KERNEL);
2603         if (!obj)
2604                 return -ENOMEM;
2605
2606         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext, &srq_lock_class);
2607         down_write(&obj->uevent.uobject.mutex);
2608
2609         if (cmd->srq_type == IB_SRQT_XRC) {
2610                 attr.ext.xrc.xrcd  = idr_read_xrcd(cmd->xrcd_handle, file->ucontext, &xrcd_uobj);
2611                 if (!attr.ext.xrc.xrcd) {
2612                         ret = -EINVAL;
2613                         goto err;
2614                 }
2615
2616                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
2617                 atomic_inc(&obj->uxrcd->refcnt);
2618
2619                 attr.ext.xrc.cq  = idr_read_cq(cmd->cq_handle, file->ucontext, 0);
2620                 if (!attr.ext.xrc.cq) {
2621                         ret = -EINVAL;
2622                         goto err_put_xrcd;
2623                 }
2624         }
2625
2626         pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
2627         if (!pd) {
2628                 ret = -EINVAL;
2629                 goto err_put_cq;
2630         }
2631
2632         attr.event_handler  = ib_uverbs_srq_event_handler;
2633         attr.srq_context    = file;
2634         attr.srq_type       = cmd->srq_type;
2635         attr.attr.max_wr    = cmd->max_wr;
2636         attr.attr.max_sge   = cmd->max_sge;
2637         attr.attr.srq_limit = cmd->srq_limit;
2638
2639         obj->uevent.events_reported = 0;
2640         INIT_LIST_HEAD(&obj->uevent.event_list);
2641
2642         srq = pd->device->create_srq(pd, &attr, udata);
2643         if (IS_ERR(srq)) {
2644                 ret = PTR_ERR(srq);
2645                 goto err_put;
2646         }
2647
2648         srq->device        = pd->device;
2649         srq->pd            = pd;
2650         srq->srq_type      = cmd->srq_type;
2651         srq->uobject       = &obj->uevent.uobject;
2652         srq->event_handler = attr.event_handler;
2653         srq->srq_context   = attr.srq_context;
2654
2655         if (cmd->srq_type == IB_SRQT_XRC) {
2656                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
2657                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
2658                 atomic_inc(&attr.ext.xrc.cq->usecnt);
2659                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
2660         }
2661
2662         atomic_inc(&pd->usecnt);
2663         atomic_set(&srq->usecnt, 0);
2664
2665         obj->uevent.uobject.object = srq;
2666         ret = idr_add_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2667         if (ret)
2668                 goto err_destroy;
2669
2670         memset(&resp, 0, sizeof resp);
2671         resp.srq_handle = obj->uevent.uobject.id;
2672         resp.max_wr     = attr.attr.max_wr;
2673         resp.max_sge    = attr.attr.max_sge;
2674         if (cmd->srq_type == IB_SRQT_XRC)
2675                 resp.srqn = srq->ext.xrc.srq_num;
2676
2677         if (copy_to_user((void __user *) (unsigned long) cmd->response,
2678                          &resp, sizeof resp)) {
2679                 ret = -EFAULT;
2680                 goto err_copy;
2681         }
2682
2683         if (cmd->srq_type == IB_SRQT_XRC) {
2684                 put_uobj_read(xrcd_uobj);
2685                 put_cq_read(attr.ext.xrc.cq);
2686         }
2687         put_pd_read(pd);
2688
2689         mutex_lock(&file->mutex);
2690         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->srq_list);
2691         mutex_unlock(&file->mutex);
2692
2693         obj->uevent.uobject.live = 1;
2694
2695         up_write(&obj->uevent.uobject.mutex);
2696
2697         return 0;
2698
2699 err_copy:
2700         idr_remove_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2701
2702 err_destroy:
2703         ib_destroy_srq(srq);
2704
2705 err_put:
2706         put_pd_read(pd);
2707
2708 err_put_cq:
2709         if (cmd->srq_type == IB_SRQT_XRC)
2710                 put_cq_read(attr.ext.xrc.cq);
2711
2712 err_put_xrcd:
2713         if (cmd->srq_type == IB_SRQT_XRC) {
2714                 atomic_dec(&obj->uxrcd->refcnt);
2715                 put_uobj_read(xrcd_uobj);
2716         }
2717
2718 err:
2719         put_uobj_write(&obj->uevent.uobject);
2720         return ret;
2721 }
2722
2723 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
2724                              const char __user *buf, int in_len,
2725                              int out_len)
2726 {
2727         struct ib_uverbs_create_srq      cmd;
2728         struct ib_uverbs_create_xsrq     xcmd;
2729         struct ib_uverbs_create_srq_resp resp;
2730         struct ib_udata                  udata;
2731         int ret;
2732
2733         if (out_len < sizeof resp)
2734                 return -ENOSPC;
2735
2736         if (copy_from_user(&cmd, buf, sizeof cmd))
2737                 return -EFAULT;
2738
2739         xcmd.response    = cmd.response;
2740         xcmd.user_handle = cmd.user_handle;
2741         xcmd.srq_type    = IB_SRQT_BASIC;
2742         xcmd.pd_handle   = cmd.pd_handle;
2743         xcmd.max_wr      = cmd.max_wr;
2744         xcmd.max_sge     = cmd.max_sge;
2745         xcmd.srq_limit   = cmd.srq_limit;
2746
2747         INIT_UDATA(&udata, buf + sizeof cmd,
2748                    (unsigned long) cmd.response + sizeof resp,
2749                    in_len - sizeof cmd, out_len - sizeof resp);
2750
2751         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
2752         if (ret)
2753                 return ret;
2754
2755         return in_len;
2756 }
2757
2758 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
2759                               const char __user *buf, int in_len, int out_len)
2760 {
2761         struct ib_uverbs_create_xsrq     cmd;
2762         struct ib_uverbs_create_srq_resp resp;
2763         struct ib_udata                  udata;
2764         int ret;
2765
2766         if (out_len < sizeof resp)
2767                 return -ENOSPC;
2768
2769         if (copy_from_user(&cmd, buf, sizeof cmd))
2770                 return -EFAULT;
2771
2772         INIT_UDATA(&udata, buf + sizeof cmd,
2773                    (unsigned long) cmd.response + sizeof resp,
2774                    in_len - sizeof cmd, out_len - sizeof resp);
2775
2776         ret = __uverbs_create_xsrq(file, &cmd, &udata);
2777         if (ret)
2778                 return ret;
2779
2780         return in_len;
2781 }
2782
2783 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
2784                              const char __user *buf, int in_len,
2785                              int out_len)
2786 {
2787         struct ib_uverbs_modify_srq cmd;
2788         struct ib_udata             udata;
2789         struct ib_srq              *srq;
2790         struct ib_srq_attr          attr;
2791         int                         ret;
2792
2793         if (copy_from_user(&cmd, buf, sizeof cmd))
2794                 return -EFAULT;
2795
2796         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
2797                    out_len);
2798
2799         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2800         if (!srq)
2801                 return -EINVAL;
2802
2803         attr.max_wr    = cmd.max_wr;
2804         attr.srq_limit = cmd.srq_limit;
2805
2806         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
2807
2808         put_srq_read(srq);
2809
2810         return ret ? ret : in_len;
2811 }
2812
2813 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
2814                             const char __user *buf,
2815                             int in_len, int out_len)
2816 {
2817         struct ib_uverbs_query_srq      cmd;
2818         struct ib_uverbs_query_srq_resp resp;
2819         struct ib_srq_attr              attr;
2820         struct ib_srq                   *srq;
2821         int                             ret;
2822
2823         if (out_len < sizeof resp)
2824                 return -ENOSPC;
2825
2826         if (copy_from_user(&cmd, buf, sizeof cmd))
2827                 return -EFAULT;
2828
2829         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2830         if (!srq)
2831                 return -EINVAL;
2832
2833         ret = ib_query_srq(srq, &attr);
2834
2835         put_srq_read(srq);
2836
2837         if (ret)
2838                 return ret;
2839
2840         memset(&resp, 0, sizeof resp);
2841
2842         resp.max_wr    = attr.max_wr;
2843         resp.max_sge   = attr.max_sge;
2844         resp.srq_limit = attr.srq_limit;
2845
2846         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2847                          &resp, sizeof resp))
2848                 return -EFAULT;
2849
2850         return in_len;
2851 }
2852
2853 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
2854                               const char __user *buf, int in_len,
2855                               int out_len)
2856 {
2857         struct ib_uverbs_destroy_srq      cmd;
2858         struct ib_uverbs_destroy_srq_resp resp;
2859         struct ib_uobject                *uobj;
2860         struct ib_srq                    *srq;
2861         struct ib_uevent_object          *obj;
2862         int                               ret = -EINVAL;
2863
2864         if (copy_from_user(&cmd, buf, sizeof cmd))
2865                 return -EFAULT;
2866
2867         uobj = idr_write_uobj(&ib_uverbs_srq_idr, cmd.srq_handle, file->ucontext);
2868         if (!uobj)
2869                 return -EINVAL;
2870         srq = uobj->object;
2871         obj = container_of(uobj, struct ib_uevent_object, uobject);
2872
2873         ret = ib_destroy_srq(srq);
2874         if (!ret)
2875                 uobj->live = 0;
2876
2877         put_uobj_write(uobj);
2878
2879         if (ret)
2880                 return ret;
2881
2882         idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
2883
2884         mutex_lock(&file->mutex);
2885         list_del(&uobj->list);
2886         mutex_unlock(&file->mutex);
2887
2888         ib_uverbs_release_uevent(file, obj);
2889
2890         memset(&resp, 0, sizeof resp);
2891         resp.events_reported = obj->events_reported;
2892
2893         put_uobj(uobj);
2894
2895         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2896                          &resp, sizeof resp))
2897                 ret = -EFAULT;
2898
2899         return ret ? ret : in_len;
2900 }