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 retry:
 129         if (!idr_pre_get(idr, GFP_KERNEL))
 130                 return -ENOMEM;
 131
 132         spin_lock(&ib_uverbs_idr_lock);
 133         ret = idr_get_new(idr, uobj, &uobj->id);
 134         spin_unlock(&ib_uverbs_idr_lock);
 135
 136         if (ret == -EAGAIN)
 137                 goto retry;
 138
 139         return ret;
 140 }
 141
 142 void idr_remove_uobj(struct idr *idr, struct ib_uobject *uobj)
 143 {
 144         spin_lock(&ib_uverbs_idr_lock);
 145         idr_remove(idr, uobj->id);
 146         spin_unlock(&ib_uverbs_idr_lock);
 147 }
 148
 149 static struct ib_uobject *__idr_get_uobj(struct idr *idr, int id,
 150                                          struct ib_ucontext *context)
 151 {
 152         struct ib_uobject *uobj;
 153
 154         spin_lock(&ib_uverbs_idr_lock);
 155         uobj = idr_find(idr, id);
 156         if (uobj) {
 157                 if (uobj->context == context)
 158                         kref_get(&uobj->ref);
 159                 else
 160                         uobj = NULL;
 161         }
 162         spin_unlock(&ib_uverbs_idr_lock);
 163
 164         return uobj;
 165 }
 166
 167 static struct ib_uobject *idr_read_uobj(struct idr *idr, int id,
 168                                         struct ib_ucontext *context, int nested)
 169 {
 170         struct ib_uobject *uobj;
 171
 172         uobj = __idr_get_uobj(idr, id, context);
 173         if (!uobj)
 174                 return NULL;
 175
 176         if (nested)
 177                 down_read_nested(&uobj->mutex, SINGLE_DEPTH_NESTING);
 178         else
 179                 down_read(&uobj->mutex);
 180         if (!uobj->live) {
 181                 put_uobj_read(uobj);
 182                 return NULL;
 183         }
 184
 185         return uobj;
 186 }
 187
 188 static struct ib_uobject *idr_write_uobj(struct idr *idr, int id,
 189                                          struct ib_ucontext *context)
 190 {
 191         struct ib_uobject *uobj;
 192
 193         uobj = __idr_get_uobj(idr, id, context);
 194         if (!uobj)
 195                 return NULL;
 196
 197         down_write(&uobj->mutex);
 198         if (!uobj->live) {
 199                 put_uobj_write(uobj);
 200                 return NULL;
 201         }
 202
 203         return uobj;
 204 }
 205
 206 static void *idr_read_obj(struct idr *idr, int id, struct ib_ucontext *context,
 207                           int nested)
 208 {
 209         struct ib_uobject *uobj;
 210
 211         uobj = idr_read_uobj(idr, id, context, nested);
 212         return uobj ? uobj->object : NULL;
 213 }
 214
 215 static struct ib_pd *idr_read_pd(int pd_handle, struct ib_ucontext *context)
 216 {
 217         return idr_read_obj(&ib_uverbs_pd_idr, pd_handle, context, 0);
 218 }
 219
 220 static void put_pd_read(struct ib_pd *pd)
 221 {
 222         put_uobj_read(pd->uobject);
 223 }
 224
 225 static struct ib_cq *idr_read_cq(int cq_handle, struct ib_ucontext *context, int nested)
 226 {
 227         return idr_read_obj(&ib_uverbs_cq_idr, cq_handle, context, nested);
 228 }
 229
 230 static void put_cq_read(struct ib_cq *cq)
 231 {
 232         put_uobj_read(cq->uobject);
 233 }
 234
 235 static struct ib_ah *idr_read_ah(int ah_handle, struct ib_ucontext *context)
 236 {
 237         return idr_read_obj(&ib_uverbs_ah_idr, ah_handle, context, 0);
 238 }
 239
 240 static void put_ah_read(struct ib_ah *ah)
 241 {
 242         put_uobj_read(ah->uobject);
 243 }
 244
 245 static struct ib_qp *idr_read_qp(int qp_handle, struct ib_ucontext *context)
 246 {
 247         return idr_read_obj(&ib_uverbs_qp_idr, qp_handle, context, 0);
 248 }
 249
 250 static struct ib_qp *idr_write_qp(int qp_handle, struct ib_ucontext *context)
 251 {
 252         struct ib_uobject *uobj;
 253
 254         uobj = idr_write_uobj(&ib_uverbs_qp_idr, qp_handle, context);
 255         return uobj ? uobj->object : NULL;
 256 }
 257
 258 static void put_qp_read(struct ib_qp *qp)
 259 {
 260         put_uobj_read(qp->uobject);
 261 }
 262
 263 static void put_qp_write(struct ib_qp *qp)
 264 {
 265         put_uobj_write(qp->uobject);
 266 }
 267
 268 static struct ib_srq *idr_read_srq(int srq_handle, struct ib_ucontext *context)
 269 {
 270         return idr_read_obj(&ib_uverbs_srq_idr, srq_handle, context, 0);
 271 }
 272
 273 static void put_srq_read(struct ib_srq *srq)
 274 {
 275         put_uobj_read(srq->uobject);
 276 }
 277
 278 static struct ib_xrcd *idr_read_xrcd(int xrcd_handle, struct ib_ucontext *context,
 279                                      struct ib_uobject **uobj)
 280 {
 281         *uobj = idr_read_uobj(&ib_uverbs_xrcd_idr, xrcd_handle, context, 0);
 282         return *uobj ? (*uobj)->object : NULL;
 283 }
 284
 285 static void put_xrcd_read(struct ib_uobject *uobj)
 286 {
 287         put_uobj_read(uobj);
 288 }
 289
 290 ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
 291                               const char __user *buf,
 292                               int in_len, int out_len)
 293 {
 294         struct ib_uverbs_get_context      cmd;
 295         struct ib_uverbs_get_context_resp resp;
 296         struct ib_udata                   udata;
 297         struct ib_device                 *ibdev = file->device->ib_dev;
 298         struct ib_ucontext               *ucontext;
 299         struct file                      *filp;
 300         int ret;
 301
 302         if (out_len < sizeof resp)
 303                 return -ENOSPC;
 304
 305         if (copy_from_user(&cmd, buf, sizeof cmd))
 306                 return -EFAULT;
 307
 308         mutex_lock(&file->mutex);
 309
 310         if (file->ucontext) {
 311                 ret = -EINVAL;
 312                 goto err;
 313         }
 314
 315         INIT_UDATA(&udata, buf + sizeof cmd,
 316                    (unsigned long) cmd.response + sizeof resp,
 317                    in_len - sizeof cmd, out_len - sizeof resp);
 318
 319         ucontext = ibdev->alloc_ucontext(ibdev, &udata);
 320         if (IS_ERR(ucontext)) {
 321                 ret = PTR_ERR(ucontext);
 322                 goto err;
 323         }
 324
 325         ucontext->device = ibdev;
 326         INIT_LIST_HEAD(&ucontext->pd_list);
 327         INIT_LIST_HEAD(&ucontext->mr_list);
 328         INIT_LIST_HEAD(&ucontext->mw_list);
 329         INIT_LIST_HEAD(&ucontext->cq_list);
 330         INIT_LIST_HEAD(&ucontext->qp_list);
 331         INIT_LIST_HEAD(&ucontext->srq_list);
 332         INIT_LIST_HEAD(&ucontext->ah_list);
 333         INIT_LIST_HEAD(&ucontext->xrcd_list);
 334         ucontext->closing = 0;
 335
 336         resp.num_comp_vectors = file->device->num_comp_vectors;
 337
 338         ret = get_unused_fd();
 339         if (ret < 0)
 340                 goto err_free;
 341         resp.async_fd = ret;
 342
 343         filp = ib_uverbs_alloc_event_file(file, 1);
 344         if (IS_ERR(filp)) {
 345                 ret = PTR_ERR(filp);
 346                 goto err_fd;
 347         }
 348
 349         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 350                          &resp, sizeof resp)) {
 351                 ret = -EFAULT;
 352                 goto err_file;
 353         }
 354
 355         file->async_file = filp->private_data;
 356
 357         INIT_IB_EVENT_HANDLER(&file->event_handler, file->device->ib_dev,
 358                               ib_uverbs_event_handler);
 359         ret = ib_register_event_handler(&file->event_handler);
 360         if (ret)
 361                 goto err_file;
 362
 363         kref_get(&file->async_file->ref);
 364         kref_get(&file->ref);
 365         file->ucontext = ucontext;
 366
 367         fd_install(resp.async_fd, filp);
 368
 369         mutex_unlock(&file->mutex);
 370
 371         return in_len;
 372
 373 err_file:
 374         fput(filp);
 375
 376 err_fd:
 377         put_unused_fd(resp.async_fd);
 378
 379 err_free:
 380         ibdev->dealloc_ucontext(ucontext);
 381
 382 err:
 383         mutex_unlock(&file->mutex);
 384         return ret;
 385 }
 386
 387 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 388                                const char __user *buf,
 389                                int in_len, int out_len)
 390 {
 391         struct ib_uverbs_query_device      cmd;
 392         struct ib_uverbs_query_device_resp resp;
 393         struct ib_device_attr              attr;
 394         int                                ret;
 395
 396         if (out_len < sizeof resp)
 397                 return -ENOSPC;
 398
 399         if (copy_from_user(&cmd, buf, sizeof cmd))
 400                 return -EFAULT;
 401
 402         ret = ib_query_device(file->device->ib_dev, &attr);
 403         if (ret)
 404                 return ret;
 405
 406         memset(&resp, 0, sizeof resp);
 407
 408         resp.fw_ver                    = attr.fw_ver;
 409         resp.node_guid                 = file->device->ib_dev->node_guid;
 410         resp.sys_image_guid            = attr.sys_image_guid;
 411         resp.max_mr_size               = attr.max_mr_size;
 412         resp.page_size_cap             = attr.page_size_cap;
 413         resp.vendor_id                 = attr.vendor_id;
 414         resp.vendor_part_id            = attr.vendor_part_id;
 415         resp.hw_ver                    = attr.hw_ver;
 416         resp.max_qp                    = attr.max_qp;
 417         resp.max_qp_wr                 = attr.max_qp_wr;
 418         resp.device_cap_flags          = attr.device_cap_flags;
 419         resp.max_sge                   = attr.max_sge;
 420         resp.max_sge_rd                = attr.max_sge_rd;
 421         resp.max_cq                    = attr.max_cq;
 422         resp.max_cqe                   = attr.max_cqe;
 423         resp.max_mr                    = attr.max_mr;
 424         resp.max_pd                    = attr.max_pd;
 425         resp.max_qp_rd_atom            = attr.max_qp_rd_atom;
 426         resp.max_ee_rd_atom            = attr.max_ee_rd_atom;
 427         resp.max_res_rd_atom           = attr.max_res_rd_atom;
 428         resp.max_qp_init_rd_atom       = attr.max_qp_init_rd_atom;
 429         resp.max_ee_init_rd_atom       = attr.max_ee_init_rd_atom;
 430         resp.atomic_cap                = attr.atomic_cap;
 431         resp.max_ee                    = attr.max_ee;
 432         resp.max_rdd                   = attr.max_rdd;
 433         resp.max_mw                    = attr.max_mw;
 434         resp.max_raw_ipv6_qp           = attr.max_raw_ipv6_qp;
 435         resp.max_raw_ethy_qp           = attr.max_raw_ethy_qp;
 436         resp.max_mcast_grp             = attr.max_mcast_grp;
 437         resp.max_mcast_qp_attach       = attr.max_mcast_qp_attach;
 438         resp.max_total_mcast_qp_attach = attr.max_total_mcast_qp_attach;
 439         resp.max_ah                    = attr.max_ah;
 440         resp.max_fmr                   = attr.max_fmr;
 441         resp.max_map_per_fmr           = attr.max_map_per_fmr;
 442         resp.max_srq                   = attr.max_srq;
 443         resp.max_srq_wr                = attr.max_srq_wr;
 444         resp.max_srq_sge               = attr.max_srq_sge;
 445         resp.max_pkeys                 = attr.max_pkeys;
 446         resp.local_ca_ack_delay        = attr.local_ca_ack_delay;
 447         resp.phys_port_cnt             = file->device->ib_dev->phys_port_cnt;
 448
 449         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 450                          &resp, sizeof resp))
 451                 return -EFAULT;
 452
 453         return in_len;
 454 }
 455
 456 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 457                              const char __user *buf,
 458                              int in_len, int out_len)
 459 {
 460         struct ib_uverbs_query_port      cmd;
 461         struct ib_uverbs_query_port_resp resp;
 462         struct ib_port_attr              attr;
 463         int                              ret;
 464
 465         if (out_len < sizeof resp)
 466                 return -ENOSPC;
 467
 468         if (copy_from_user(&cmd, buf, sizeof cmd))
 469                 return -EFAULT;
 470
 471         ret = ib_query_port(file->device->ib_dev, cmd.port_num, &attr);
 472         if (ret)
 473                 return ret;
 474
 475         memset(&resp, 0, sizeof resp);
 476
 477         resp.state           = attr.state;
 478         resp.max_mtu         = attr.max_mtu;
 479         resp.active_mtu      = attr.active_mtu;
 480         resp.gid_tbl_len     = attr.gid_tbl_len;
 481         resp.port_cap_flags  = attr.port_cap_flags;
 482         resp.max_msg_sz      = attr.max_msg_sz;
 483         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 484         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 485         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 486         resp.lid             = attr.lid;
 487         resp.sm_lid          = attr.sm_lid;
 488         resp.lmc             = attr.lmc;
 489         resp.max_vl_num      = attr.max_vl_num;
 490         resp.sm_sl           = attr.sm_sl;
 491         resp.subnet_timeout  = attr.subnet_timeout;
 492         resp.init_type_reply = attr.init_type_reply;
 493         resp.active_width    = attr.active_width;
 494         resp.active_speed    = attr.active_speed;
 495         resp.phys_state      = attr.phys_state;
 496         resp.link_layer      = rdma_port_get_link_layer(file->device->ib_dev,
 497                                                         cmd.port_num);
 498
 499         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 500                          &resp, sizeof resp))
 501                 return -EFAULT;
 502
 503         return in_len;
 504 }
 505
 506 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 507                            const char __user *buf,
 508                            int in_len, int out_len)
 509 {
 510         struct ib_uverbs_alloc_pd      cmd;
 511         struct ib_uverbs_alloc_pd_resp resp;
 512         struct ib_udata                udata;
 513         struct ib_uobject             *uobj;
 514         struct ib_pd                  *pd;
 515         int                            ret;
 516
 517         if (out_len < sizeof resp)
 518                 return -ENOSPC;
 519
 520         if (copy_from_user(&cmd, buf, sizeof cmd))
 521                 return -EFAULT;
 522
 523         INIT_UDATA(&udata, buf + sizeof cmd,
 524                    (unsigned long) cmd.response + sizeof resp,
 525                    in_len - sizeof cmd, out_len - sizeof resp);
 526
 527         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 528         if (!uobj)
 529                 return -ENOMEM;
 530
 531         init_uobj(uobj, 0, file->ucontext, &pd_lock_class);
 532         down_write(&uobj->mutex);
 533
 534         pd = file->device->ib_dev->alloc_pd(file->device->ib_dev,
 535                                             file->ucontext, &udata);
 536         if (IS_ERR(pd)) {
 537                 ret = PTR_ERR(pd);
 538                 goto err;
 539         }
 540
 541         pd->device  = file->device->ib_dev;
 542         pd->uobject = uobj;
 543         atomic_set(&pd->usecnt, 0);
 544
 545         uobj->object = pd;
 546         ret = idr_add_uobj(&ib_uverbs_pd_idr, uobj);
 547         if (ret)
 548                 goto err_idr;
 549
 550         memset(&resp, 0, sizeof resp);
 551         resp.pd_handle = uobj->id;
 552
 553         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 554                          &resp, sizeof resp)) {
 555                 ret = -EFAULT;
 556                 goto err_copy;
 557         }
 558
 559         mutex_lock(&file->mutex);
 560         list_add_tail(&uobj->list, &file->ucontext->pd_list);
 561         mutex_unlock(&file->mutex);
 562
 563         uobj->live = 1;
 564
 565         up_write(&uobj->mutex);
 566
 567         return in_len;
 568
 569 err_copy:
 570         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 571
 572 err_idr:
 573         ib_dealloc_pd(pd);
 574
 575 err:
 576         put_uobj_write(uobj);
 577         return ret;
 578 }
 579
 580 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 581                              const char __user *buf,
 582                              int in_len, int out_len)
 583 {
 584         struct ib_uverbs_dealloc_pd cmd;
 585         struct ib_uobject          *uobj;
 586         int                         ret;
 587
 588         if (copy_from_user(&cmd, buf, sizeof cmd))
 589                 return -EFAULT;
 590
 591         uobj = idr_write_uobj(&ib_uverbs_pd_idr, cmd.pd_handle, file->ucontext);
 592         if (!uobj)
 593                 return -EINVAL;
 594
 595         ret = ib_dealloc_pd(uobj->object);
 596         if (!ret)
 597                 uobj->live = 0;
 598
 599         put_uobj_write(uobj);
 600
 601         if (ret)
 602                 return ret;
 603
 604         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 605
 606         mutex_lock(&file->mutex);
 607         list_del(&uobj->list);
 608         mutex_unlock(&file->mutex);
 609
 610         put_uobj(uobj);
 611
 612         return in_len;
 613 }
 614
 615 struct xrcd_table_entry {
 616         struct rb_node  node;
 617         struct ib_xrcd *xrcd;
 618         struct inode   *inode;
 619 };
 620
 621 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 622                             struct inode *inode,
 623                             struct ib_xrcd *xrcd)
 624 {
 625         struct xrcd_table_entry *entry, *scan;
 626         struct rb_node **p = &dev->xrcd_tree.rb_node;
 627         struct rb_node *parent = NULL;
 628
 629         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 630         if (!entry)
 631                 return -ENOMEM;
 632
 633         entry->xrcd  = xrcd;
 634         entry->inode = inode;
 635
 636         while (*p) {
 637                 parent = *p;
 638                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 639
 640                 if (inode < scan->inode) {
 641                         p = &(*p)->rb_left;
 642                 } else if (inode > scan->inode) {
 643                         p = &(*p)->rb_right;
 644                 } else {
 645                         kfree(entry);
 646                         return -EEXIST;
 647                 }
 648         }
 649
 650         rb_link_node(&entry->node, parent, p);
 651         rb_insert_color(&entry->node, &dev->xrcd_tree);
 652         igrab(inode);
 653         return 0;
 654 }
 655
 656 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 657                                                   struct inode *inode)
 658 {
 659         struct xrcd_table_entry *entry;
 660         struct rb_node *p = dev->xrcd_tree.rb_node;
 661
 662         while (p) {
 663                 entry = rb_entry(p, struct xrcd_table_entry, node);
 664
 665                 if (inode < entry->inode)
 666                         p = p->rb_left;
 667                 else if (inode > entry->inode)
 668                         p = p->rb_right;
 669                 else
 670                         return entry;
 671         }
 672
 673         return NULL;
 674 }
 675
 676 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 677 {
 678         struct xrcd_table_entry *entry;
 679
 680         entry = xrcd_table_search(dev, inode);
 681         if (!entry)
 682                 return NULL;
 683
 684         return entry->xrcd;
 685 }
 686
 687 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 688                               struct inode *inode)
 689 {
 690         struct xrcd_table_entry *entry;
 691
 692         entry = xrcd_table_search(dev, inode);
 693         if (entry) {
 694                 iput(inode);
 695                 rb_erase(&entry->node, &dev->xrcd_tree);
 696                 kfree(entry);
 697         }
 698 }
 699
 700 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 701                             const char __user *buf, int in_len,
 702                             int out_len)
 703 {
 704         struct ib_uverbs_open_xrcd      cmd;
 705         struct ib_uverbs_open_xrcd_resp resp;
 706         struct ib_udata                 udata;
 707         struct ib_uxrcd_object         *obj;
 708         struct ib_xrcd                 *xrcd = NULL;
 709         struct fd                       f = {NULL, 0};
 710         struct inode                   *inode = NULL;
 711         int                             ret = 0;
 712         int                             new_xrcd = 0;
 713
 714         if (out_len < sizeof resp)
 715                 return -ENOSPC;
 716
 717         if (copy_from_user(&cmd, buf, sizeof cmd))
 718                 return -EFAULT;
 719
 720         INIT_UDATA(&udata, buf + sizeof cmd,
 721                    (unsigned long) cmd.response + sizeof resp,
 722                    in_len - sizeof cmd, out_len - sizeof  resp);
 723
 724         mutex_lock(&file->device->xrcd_tree_mutex);
 725
 726         if (cmd.fd != -1) {
 727                 /* search for file descriptor */
 728                 f = fdget(cmd.fd);
 729                 if (!f.file) {
 730                         ret = -EBADF;
 731                         goto err_tree_mutex_unlock;
 732                 }
 733
 734                 inode = file_inode(f.file);
 735                 xrcd = find_xrcd(file->device, inode);
 736                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 737                         /* no file descriptor. Need CREATE flag */
 738                         ret = -EAGAIN;
 739                         goto err_tree_mutex_unlock;
 740                 }
 741
 742                 if (xrcd && cmd.oflags & O_EXCL) {
 743                         ret = -EINVAL;
 744                         goto err_tree_mutex_unlock;
 745                 }
 746         }
 747
 748         obj = kmalloc(sizeof *obj, GFP_KERNEL);
 749         if (!obj) {
 750                 ret = -ENOMEM;
 751                 goto err_tree_mutex_unlock;
 752         }
 753
 754         init_uobj(&obj->uobject, 0, file->ucontext, &xrcd_lock_class);
 755
 756         down_write(&obj->uobject.mutex);
 757
 758         if (!xrcd) {
 759                 xrcd = file->device->ib_dev->alloc_xrcd(file->device->ib_dev,
 760                                                         file->ucontext, &udata);
 761                 if (IS_ERR(xrcd)) {
 762                         ret = PTR_ERR(xrcd);
 763                         goto err;
 764                 }
 765
 766                 xrcd->inode   = inode;
 767                 xrcd->device  = file->device->ib_dev;
 768                 atomic_set(&xrcd->usecnt, 0);
 769                 mutex_init(&xrcd->tgt_qp_mutex);
 770                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 771                 new_xrcd = 1;
 772         }
 773
 774         atomic_set(&obj->refcnt, 0);
 775         obj->uobject.object = xrcd;
 776         ret = idr_add_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 777         if (ret)
 778                 goto err_idr;
 779
 780         memset(&resp, 0, sizeof resp);
 781         resp.xrcd_handle = obj->uobject.id;
 782
 783         if (inode) {
 784                 if (new_xrcd) {
 785                         /* create new inode/xrcd table entry */
 786                         ret = xrcd_table_insert(file->device, inode, xrcd);
 787                         if (ret)
 788                                 goto err_insert_xrcd;
 789                 }
 790                 atomic_inc(&xrcd->usecnt);
 791         }
 792
 793         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 794                          &resp, sizeof resp)) {
 795                 ret = -EFAULT;
 796                 goto err_copy;
 797         }
 798
 799         if (f.file)
 800                 fdput(f);
 801
 802         mutex_lock(&file->mutex);
 803         list_add_tail(&obj->uobject.list, &file->ucontext->xrcd_list);
 804         mutex_unlock(&file->mutex);
 805
 806         obj->uobject.live = 1;
 807         up_write(&obj->uobject.mutex);
 808
 809         mutex_unlock(&file->device->xrcd_tree_mutex);
 810         return in_len;
 811
 812 err_copy:
 813         if (inode) {
 814                 if (new_xrcd)
 815                         xrcd_table_delete(file->device, inode);
 816                 atomic_dec(&xrcd->usecnt);
 817         }
 818
 819 err_insert_xrcd:
 820         idr_remove_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 821
 822 err_idr:
 823         ib_dealloc_xrcd(xrcd);
 824
 825 err:
 826         put_uobj_write(&obj->uobject);
 827
 828 err_tree_mutex_unlock:
 829         if (f.file)
 830                 fdput(f);
 831
 832         mutex_unlock(&file->device->xrcd_tree_mutex);
 833
 834         return ret;
 835 }
 836
 837 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 838                              const char __user *buf, int in_len,
 839                              int out_len)
 840 {
 841         struct ib_uverbs_close_xrcd cmd;
 842         struct ib_uobject           *uobj;
 843         struct ib_xrcd              *xrcd = NULL;
 844         struct inode                *inode = NULL;
 845         struct ib_uxrcd_object      *obj;
 846         int                         live;
 847         int                         ret = 0;
 848
 849         if (copy_from_user(&cmd, buf, sizeof cmd))
 850                 return -EFAULT;
 851
 852         mutex_lock(&file->device->xrcd_tree_mutex);
 853         uobj = idr_write_uobj(&ib_uverbs_xrcd_idr, cmd.xrcd_handle, file->ucontext);
 854         if (!uobj) {
 855                 ret = -EINVAL;
 856                 goto out;
 857         }
 858
 859         xrcd  = uobj->object;
 860         inode = xrcd->inode;
 861         obj   = container_of(uobj, struct ib_uxrcd_object, uobject);
 862         if (atomic_read(&obj->refcnt)) {
 863                 put_uobj_write(uobj);
 864                 ret = -EBUSY;
 865                 goto out;
 866         }
 867
 868         if (!inode || atomic_dec_and_test(&xrcd->usecnt)) {
 869                 ret = ib_dealloc_xrcd(uobj->object);
 870                 if (!ret)
 871                         uobj->live = 0;
 872         }
 873
 874         live = uobj->live;
 875         if (inode && ret)
 876                 atomic_inc(&xrcd->usecnt);
 877
 878         put_uobj_write(uobj);
 879
 880         if (ret)
 881                 goto out;
 882
 883         if (inode && !live)
 884                 xrcd_table_delete(file->device, inode);
 885
 886         idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 887         mutex_lock(&file->mutex);
 888         list_del(&uobj->list);
 889         mutex_unlock(&file->mutex);
 890
 891         put_uobj(uobj);
 892         ret = in_len;
 893
 894 out:
 895         mutex_unlock(&file->device->xrcd_tree_mutex);
 896         return ret;
 897 }
 898
 899 void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 900                             struct ib_xrcd *xrcd)
 901 {
 902         struct inode *inode;
 903
 904         inode = xrcd->inode;
 905         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 906                 return;
 907
 908         ib_dealloc_xrcd(xrcd);
 909
 910         if (inode)
 911                 xrcd_table_delete(dev, inode);
 912 }
 913
 914 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 915                          const char __user *buf, int in_len,
 916                          int out_len)
 917 {
 918         struct ib_uverbs_reg_mr      cmd;
 919         struct ib_uverbs_reg_mr_resp resp;
 920         struct ib_udata              udata;
 921         struct ib_uobject           *uobj;
 922         struct ib_pd                *pd;
 923         struct ib_mr                *mr;
 924         int                          ret;
 925
 926         if (out_len < sizeof resp)
 927                 return -ENOSPC;
 928
 929         if (copy_from_user(&cmd, buf, sizeof cmd))
 930                 return -EFAULT;
 931
 932         INIT_UDATA(&udata, buf + sizeof cmd,
 933                    (unsigned long) cmd.response + sizeof resp,
 934                    in_len - sizeof cmd, out_len - sizeof resp);
 935
 936         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 937                 return -EINVAL;
 938
 939         /*
 940          * Local write permission is required if remote write or
 941          * remote atomic permission is also requested.
 942          */
 943         if (cmd.access_flags & (IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_REMOTE_WRITE) &&
 944             !(cmd.access_flags & IB_ACCESS_LOCAL_WRITE))
 945                 return -EINVAL;
 946
 947         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 948         if (!uobj)
 949                 return -ENOMEM;
 950
 951         init_uobj(uobj, 0, file->ucontext, &mr_lock_class);
 952         down_write(&uobj->mutex);
 953
 954         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
 955         if (!pd) {
 956                 ret = -EINVAL;
 957                 goto err_free;
 958         }
 959
 960         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 961                                      cmd.access_flags, &udata);
 962         if (IS_ERR(mr)) {
 963                 ret = PTR_ERR(mr);
 964                 goto err_put;
 965         }
 966
 967         mr->device  = pd->device;
 968         mr->pd      = pd;
 969         mr->uobject = uobj;
 970         atomic_inc(&pd->usecnt);
 971         atomic_set(&mr->usecnt, 0);
 972
 973         uobj->object = mr;
 974         ret = idr_add_uobj(&ib_uverbs_mr_idr, uobj);
 975         if (ret)
 976                 goto err_unreg;
 977
 978         memset(&resp, 0, sizeof resp);
 979         resp.lkey      = mr->lkey;
 980         resp.rkey      = mr->rkey;
 981         resp.mr_handle = uobj->id;
 982
 983         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 984                          &resp, sizeof resp)) {
 985                 ret = -EFAULT;
 986                 goto err_copy;
 987         }
 988
 989         put_pd_read(pd);
 990
 991         mutex_lock(&file->mutex);
 992         list_add_tail(&uobj->list, &file->ucontext->mr_list);
 993         mutex_unlock(&file->mutex);
 994
 995         uobj->live = 1;
 996
 997         up_write(&uobj->mutex);
 998
 999         return in_len;
1000
1001 err_copy:
1002         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1003
1004 err_unreg:
1005         ib_dereg_mr(mr);
1006
1007 err_put:
1008         put_pd_read(pd);
1009
1010 err_free:
1011         put_uobj_write(uobj);
1012         return ret;
1013 }
1014
1015 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
1016                            const char __user *buf, int in_len,
1017                            int out_len)
1018 {
1019         struct ib_uverbs_dereg_mr cmd;
1020         struct ib_mr             *mr;
1021         struct ib_uobject        *uobj;
1022         int                       ret = -EINVAL;
1023
1024         if (copy_from_user(&cmd, buf, sizeof cmd))
1025                 return -EFAULT;
1026
1027         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle, file->ucontext);
1028         if (!uobj)
1029                 return -EINVAL;
1030
1031         mr = uobj->object;
1032
1033         ret = ib_dereg_mr(mr);
1034         if (!ret)
1035                 uobj->live = 0;
1036
1037         put_uobj_write(uobj);
1038
1039         if (ret)
1040                 return ret;
1041
1042         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1043
1044         mutex_lock(&file->mutex);
1045         list_del(&uobj->list);
1046         mutex_unlock(&file->mutex);
1047
1048         put_uobj(uobj);
1049
1050         return in_len;
1051 }
1052
1053 ssize_t ib_uverbs_alloc_mw(struct ib_uverbs_file *file,
1054                          const char __user *buf, int in_len,
1055                          int out_len)
1056 {
1057         struct ib_uverbs_alloc_mw      cmd;
1058         struct ib_uverbs_alloc_mw_resp resp;
1059         struct ib_uobject             *uobj;
1060         struct ib_pd                  *pd;
1061         struct ib_mw                  *mw;
1062         int                            ret;
1063
1064         if (out_len < sizeof(resp))
1065                 return -ENOSPC;
1066
1067         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1068                 return -EFAULT;
1069
1070         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
1071         if (!uobj)
1072                 return -ENOMEM;
1073
1074         init_uobj(uobj, 0, file->ucontext, &mw_lock_class);
1075         down_write(&uobj->mutex);
1076
1077         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
1078         if (!pd) {
1079                 ret = -EINVAL;
1080                 goto err_free;
1081         }
1082
1083         mw = pd->device->alloc_mw(pd, cmd.mw_type);
1084         if (IS_ERR(mw)) {
1085                 ret = PTR_ERR(mw);
1086                 goto err_put;
1087         }
1088
1089         mw->device  = pd->device;
1090         mw->pd      = pd;
1091         mw->uobject = uobj;
1092         atomic_inc(&pd->usecnt);
1093
1094         uobj->object = mw;
1095         ret = idr_add_uobj(&ib_uverbs_mw_idr, uobj);
1096         if (ret)
1097                 goto err_unalloc;
1098
1099         memset(&resp, 0, sizeof(resp));
1100         resp.rkey      = mw->rkey;
1101         resp.mw_handle = uobj->id;
1102
1103         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1104                          &resp, sizeof(resp))) {
1105                 ret = -EFAULT;
1106                 goto err_copy;
1107         }
1108
1109         put_pd_read(pd);
1110
1111         mutex_lock(&file->mutex);
1112         list_add_tail(&uobj->list, &file->ucontext->mw_list);
1113         mutex_unlock(&file->mutex);
1114
1115         uobj->live = 1;
1116
1117         up_write(&uobj->mutex);
1118
1119         return in_len;
1120
1121 err_copy:
1122         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1123
1124 err_unalloc:
1125         ib_dealloc_mw(mw);
1126
1127 err_put:
1128         put_pd_read(pd);
1129
1130 err_free:
1131         put_uobj_write(uobj);
1132         return ret;
1133 }
1134
1135 ssize_t ib_uverbs_dealloc_mw(struct ib_uverbs_file *file,
1136                            const char __user *buf, int in_len,
1137                            int out_len)
1138 {
1139         struct ib_uverbs_dealloc_mw cmd;
1140         struct ib_mw               *mw;
1141         struct ib_uobject          *uobj;
1142         int                         ret = -EINVAL;
1143
1144         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1145                 return -EFAULT;
1146
1147         uobj = idr_write_uobj(&ib_uverbs_mw_idr, cmd.mw_handle, file->ucontext);
1148         if (!uobj)
1149                 return -EINVAL;
1150
1151         mw = uobj->object;
1152
1153         ret = ib_dealloc_mw(mw);
1154         if (!ret)
1155                 uobj->live = 0;
1156
1157         put_uobj_write(uobj);
1158
1159         if (ret)
1160                 return ret;
1161
1162         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1163
1164         mutex_lock(&file->mutex);
1165         list_del(&uobj->list);
1166         mutex_unlock(&file->mutex);
1167
1168         put_uobj(uobj);
1169
1170         return in_len;
1171 }
1172
1173 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
1174                                       const char __user *buf, int in_len,
1175                                       int out_len)
1176 {
1177         struct ib_uverbs_create_comp_channel       cmd;
1178         struct ib_uverbs_create_comp_channel_resp  resp;
1179         struct file                               *filp;
1180         int ret;
1181
1182         if (out_len < sizeof resp)
1183                 return -ENOSPC;
1184
1185         if (copy_from_user(&cmd, buf, sizeof cmd))
1186                 return -EFAULT;
1187
1188         ret = get_unused_fd();
1189         if (ret < 0)
1190                 return ret;
1191         resp.fd = ret;
1192
1193         filp = ib_uverbs_alloc_event_file(file, 0);
1194         if (IS_ERR(filp)) {
1195                 put_unused_fd(resp.fd);
1196                 return PTR_ERR(filp);
1197         }
1198
1199         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1200                          &resp, sizeof resp)) {
1201                 put_unused_fd(resp.fd);
1202                 fput(filp);
1203                 return -EFAULT;
1204         }
1205
1206         fd_install(resp.fd, filp);
1207         return in_len;
1208 }
1209
1210 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1211                             const char __user *buf, int in_len,
1212                             int out_len)
1213 {
1214         struct ib_uverbs_create_cq      cmd;
1215         struct ib_uverbs_create_cq_resp resp;
1216         struct ib_udata                 udata;
1217         struct ib_ucq_object           *obj;
1218         struct ib_uverbs_event_file    *ev_file = NULL;
1219         struct ib_cq                   *cq;
1220         int                             ret;
1221
1222         if (out_len < sizeof resp)
1223                 return -ENOSPC;
1224
1225         if (copy_from_user(&cmd, buf, sizeof cmd))
1226                 return -EFAULT;
1227
1228         INIT_UDATA(&udata, buf + sizeof cmd,
1229                    (unsigned long) cmd.response + sizeof resp,
1230                    in_len - sizeof cmd, out_len - sizeof resp);
1231
1232         if (cmd.comp_vector >= file->device->num_comp_vectors)
1233                 return -EINVAL;
1234
1235         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1236         if (!obj)
1237                 return -ENOMEM;
1238
1239         init_uobj(&obj->uobject, cmd.user_handle, file->ucontext, &cq_lock_class);
1240         down_write(&obj->uobject.mutex);
1241
1242         if (cmd.comp_channel >= 0) {
1243                 ev_file = ib_uverbs_lookup_comp_file(cmd.comp_channel);
1244                 if (!ev_file) {
1245                         ret = -EINVAL;
1246                         goto err;
1247                 }
1248         }
1249
1250         obj->uverbs_file           = file;
1251         obj->comp_events_reported  = 0;
1252         obj->async_events_reported = 0;
1253         INIT_LIST_HEAD(&obj->comp_list);
1254         INIT_LIST_HEAD(&obj->async_list);
1255
1256         cq = file->device->ib_dev->create_cq(file->device->ib_dev, cmd.cqe,
1257                                              cmd.comp_vector,
1258                                              file->ucontext, &udata);
1259         if (IS_ERR(cq)) {
1260                 ret = PTR_ERR(cq);
1261                 goto err_file;
1262         }
1263
1264         cq->device        = file->device->ib_dev;
1265         cq->uobject       = &obj->uobject;
1266         cq->comp_handler  = ib_uverbs_comp_handler;
1267         cq->event_handler = ib_uverbs_cq_event_handler;
1268         cq->cq_context    = ev_file;
1269         atomic_set(&cq->usecnt, 0);
1270
1271         obj->uobject.object = cq;
1272         ret = idr_add_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1273         if (ret)
1274                 goto err_free;
1275
1276         memset(&resp, 0, sizeof resp);
1277         resp.cq_handle = obj->uobject.id;
1278         resp.cqe       = cq->cqe;
1279
1280         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1281                          &resp, sizeof resp)) {
1282                 ret = -EFAULT;
1283                 goto err_copy;
1284         }
1285
1286         mutex_lock(&file->mutex);
1287         list_add_tail(&obj->uobject.list, &file->ucontext->cq_list);
1288         mutex_unlock(&file->mutex);
1289
1290         obj->uobject.live = 1;
1291
1292         up_write(&obj->uobject.mutex);
1293
1294         return in_len;
1295
1296 err_copy:
1297         idr_remove_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1298
1299 err_free:
1300         ib_destroy_cq(cq);
1301
1302 err_file:
1303         if (ev_file)
1304                 ib_uverbs_release_ucq(file, ev_file, obj);
1305
1306 err:
1307         put_uobj_write(&obj->uobject);
1308         return ret;
1309 }
1310
1311 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1312                             const char __user *buf, int in_len,
1313                             int out_len)
1314 {
1315         struct ib_uverbs_resize_cq      cmd;
1316         struct ib_uverbs_resize_cq_resp resp;
1317         struct ib_udata                 udata;
1318         struct ib_cq                    *cq;
1319         int                             ret = -EINVAL;
1320
1321         if (copy_from_user(&cmd, buf, sizeof cmd))
1322                 return -EFAULT;
1323
1324         INIT_UDATA(&udata, buf + sizeof cmd,
1325                    (unsigned long) cmd.response + sizeof resp,
1326                    in_len - sizeof cmd, out_len - sizeof resp);
1327
1328         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1329         if (!cq)
1330                 return -EINVAL;
1331
1332         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1333         if (ret)
1334                 goto out;
1335
1336         resp.cqe = cq->cqe;
1337
1338         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1339                          &resp, sizeof resp.cqe))
1340                 ret = -EFAULT;
1341
1342 out:
1343         put_cq_read(cq);
1344
1345         return ret ? ret : in_len;
1346 }
1347
1348 static int copy_wc_to_user(void __user *dest, struct ib_wc *wc)
1349 {
1350         struct ib_uverbs_wc tmp;
1351
1352         tmp.wr_id               = wc->wr_id;
1353         tmp.status              = wc->status;
1354         tmp.opcode              = wc->opcode;
1355         tmp.vendor_err          = wc->vendor_err;
1356         tmp.byte_len            = wc->byte_len;
1357         tmp.ex.imm_data         = (__u32 __force) wc->ex.imm_data;
1358         tmp.qp_num              = wc->qp->qp_num;
1359         tmp.src_qp              = wc->src_qp;
1360         tmp.wc_flags            = wc->wc_flags;
1361         tmp.pkey_index          = wc->pkey_index;
1362         tmp.slid                = wc->slid;
1363         tmp.sl                  = wc->sl;
1364         tmp.dlid_path_bits      = wc->dlid_path_bits;
1365         tmp.port_num            = wc->port_num;
1366         tmp.reserved            = 0;
1367
1368         if (copy_to_user(dest, &tmp, sizeof tmp))
1369                 return -EFAULT;
1370
1371         return 0;
1372 }
1373
1374 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1375                           const char __user *buf, int in_len,
1376                           int out_len)
1377 {
1378         struct ib_uverbs_poll_cq       cmd;
1379         struct ib_uverbs_poll_cq_resp  resp;
1380         u8 __user                     *header_ptr;
1381         u8 __user                     *data_ptr;
1382         struct ib_cq                  *cq;
1383         struct ib_wc                   wc;
1384         int                            ret;
1385
1386         if (copy_from_user(&cmd, buf, sizeof cmd))
1387                 return -EFAULT;
1388
1389         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1390         if (!cq)
1391                 return -EINVAL;
1392
1393         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1394         header_ptr = (void __user *)(unsigned long) cmd.response;
1395         data_ptr = header_ptr + sizeof resp;
1396
1397         memset(&resp, 0, sizeof resp);
1398         while (resp.count < cmd.ne) {
1399                 ret = ib_poll_cq(cq, 1, &wc);
1400                 if (ret < 0)
1401                         goto out_put;
1402                 if (!ret)
1403                         break;
1404
1405                 ret = copy_wc_to_user(data_ptr, &wc);
1406                 if (ret)
1407                         goto out_put;
1408
1409                 data_ptr += sizeof(struct ib_uverbs_wc);
1410                 ++resp.count;
1411         }
1412
1413         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1414                 ret = -EFAULT;
1415                 goto out_put;
1416         }
1417
1418         ret = in_len;
1419
1420 out_put:
1421         put_cq_read(cq);
1422         return ret;
1423 }
1424
1425 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1426                                 const char __user *buf, int in_len,
1427                                 int out_len)
1428 {
1429         struct ib_uverbs_req_notify_cq cmd;
1430         struct ib_cq                  *cq;
1431
1432         if (copy_from_user(&cmd, buf, sizeof cmd))
1433                 return -EFAULT;
1434
1435         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1436         if (!cq)
1437                 return -EINVAL;
1438
1439         ib_req_notify_cq(cq, cmd.solicited_only ?
1440                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1441
1442         put_cq_read(cq);
1443
1444         return in_len;
1445 }
1446
1447 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1448                              const char __user *buf, int in_len,
1449                              int out_len)
1450 {
1451         struct ib_uverbs_destroy_cq      cmd;
1452         struct ib_uverbs_destroy_cq_resp resp;
1453         struct ib_uobject               *uobj;
1454         struct ib_cq                    *cq;
1455         struct ib_ucq_object            *obj;
1456         struct ib_uverbs_event_file     *ev_file;
1457         int                              ret = -EINVAL;
1458
1459         if (copy_from_user(&cmd, buf, sizeof cmd))
1460                 return -EFAULT;
1461
1462         uobj = idr_write_uobj(&ib_uverbs_cq_idr, cmd.cq_handle, file->ucontext);
1463         if (!uobj)
1464                 return -EINVAL;
1465         cq      = uobj->object;
1466         ev_file = cq->cq_context;
1467         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1468
1469         ret = ib_destroy_cq(cq);
1470         if (!ret)
1471                 uobj->live = 0;
1472
1473         put_uobj_write(uobj);
1474
1475         if (ret)
1476                 return ret;
1477
1478         idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
1479
1480         mutex_lock(&file->mutex);
1481         list_del(&uobj->list);
1482         mutex_unlock(&file->mutex);
1483
1484         ib_uverbs_release_ucq(file, ev_file, obj);
1485
1486         memset(&resp, 0, sizeof resp);
1487         resp.comp_events_reported  = obj->comp_events_reported;
1488         resp.async_events_reported = obj->async_events_reported;
1489
1490         put_uobj(uobj);
1491
1492         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1493                          &resp, sizeof resp))
1494                 return -EFAULT;
1495
1496         return in_len;
1497 }
1498
1499 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1500                             const char __user *buf, int in_len,
1501                             int out_len)
1502 {
1503         struct ib_uverbs_create_qp      cmd;
1504         struct ib_uverbs_create_qp_resp resp;
1505         struct ib_udata                 udata;
1506         struct ib_uqp_object           *obj;
1507         struct ib_device               *device;
1508         struct ib_pd                   *pd = NULL;
1509         struct ib_xrcd                 *xrcd = NULL;
1510         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1511         struct ib_cq                   *scq = NULL, *rcq = NULL;
1512         struct ib_srq                  *srq = NULL;
1513         struct ib_qp                   *qp;
1514         struct ib_qp_init_attr          attr;
1515         int ret;
1516
1517         if (out_len < sizeof resp)
1518                 return -ENOSPC;
1519
1520         if (copy_from_user(&cmd, buf, sizeof cmd))
1521                 return -EFAULT;
1522
1523         if (cmd.qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1524                 return -EPERM;
1525
1526         INIT_UDATA(&udata, buf + sizeof cmd,
1527                    (unsigned long) cmd.response + sizeof resp,
1528                    in_len - sizeof cmd, out_len - sizeof resp);
1529
1530         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1531         if (!obj)
1532                 return -ENOMEM;
1533
1534         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1535         down_write(&obj->uevent.uobject.mutex);
1536
1537         if (cmd.qp_type == IB_QPT_XRC_TGT) {
1538                 xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1539                 if (!xrcd) {
1540                         ret = -EINVAL;
1541                         goto err_put;
1542                 }
1543                 device = xrcd->device;
1544         } else {
1545                 if (cmd.qp_type == IB_QPT_XRC_INI) {
1546                         cmd.max_recv_wr = cmd.max_recv_sge = 0;
1547                 } else {
1548                         if (cmd.is_srq) {
1549                                 srq = idr_read_srq(cmd.srq_handle, file->ucontext);
1550                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1551                                         ret = -EINVAL;
1552                                         goto err_put;
1553                                 }
1554                         }
1555
1556                         if (cmd.recv_cq_handle != cmd.send_cq_handle) {
1557                                 rcq = idr_read_cq(cmd.recv_cq_handle, file->ucontext, 0);
1558                                 if (!rcq) {
1559                                         ret = -EINVAL;
1560                                         goto err_put;
1561                                 }
1562                         }
1563                 }
1564
1565                 scq = idr_read_cq(cmd.send_cq_handle, file->ucontext, !!rcq);
1566                 rcq = rcq ?: scq;
1567                 pd  = idr_read_pd(cmd.pd_handle, file->ucontext);
1568                 if (!pd || !scq) {
1569                         ret = -EINVAL;
1570                         goto err_put;
1571                 }
1572
1573                 device = pd->device;
1574         }
1575
1576         attr.event_handler = ib_uverbs_qp_event_handler;
1577         attr.qp_context    = file;
1578         attr.send_cq       = scq;
1579         attr.recv_cq       = rcq;
1580         attr.srq           = srq;
1581         attr.xrcd          = xrcd;
1582         attr.sq_sig_type   = cmd.sq_sig_all ? IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
1583         attr.qp_type       = cmd.qp_type;
1584         attr.create_flags  = 0;
1585
1586         attr.cap.max_send_wr     = cmd.max_send_wr;
1587         attr.cap.max_recv_wr     = cmd.max_recv_wr;
1588         attr.cap.max_send_sge    = cmd.max_send_sge;
1589         attr.cap.max_recv_sge    = cmd.max_recv_sge;
1590         attr.cap.max_inline_data = cmd.max_inline_data;
1591
1592         obj->uevent.events_reported     = 0;
1593         INIT_LIST_HEAD(&obj->uevent.event_list);
1594         INIT_LIST_HEAD(&obj->mcast_list);
1595
1596         if (cmd.qp_type == IB_QPT_XRC_TGT)
1597                 qp = ib_create_qp(pd, &attr);
1598         else
1599                 qp = device->create_qp(pd, &attr, &udata);
1600
1601         if (IS_ERR(qp)) {
1602                 ret = PTR_ERR(qp);
1603                 goto err_put;
1604         }
1605
1606         if (cmd.qp_type != IB_QPT_XRC_TGT) {
1607                 qp->real_qp       = qp;
1608                 qp->device        = device;
1609                 qp->pd            = pd;
1610                 qp->send_cq       = attr.send_cq;
1611                 qp->recv_cq       = attr.recv_cq;
1612                 qp->srq           = attr.srq;
1613                 qp->event_handler = attr.event_handler;
1614                 qp->qp_context    = attr.qp_context;
1615                 qp->qp_type       = attr.qp_type;
1616                 atomic_set(&qp->usecnt, 0);
1617                 atomic_inc(&pd->usecnt);
1618                 atomic_inc(&attr.send_cq->usecnt);
1619                 if (attr.recv_cq)
1620                         atomic_inc(&attr.recv_cq->usecnt);
1621                 if (attr.srq)
1622                         atomic_inc(&attr.srq->usecnt);
1623         }
1624         qp->uobject = &obj->uevent.uobject;
1625
1626         obj->uevent.uobject.object = qp;
1627         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1628         if (ret)
1629                 goto err_destroy;
1630
1631         memset(&resp, 0, sizeof resp);
1632         resp.qpn             = qp->qp_num;
1633         resp.qp_handle       = obj->uevent.uobject.id;
1634         resp.max_recv_sge    = attr.cap.max_recv_sge;
1635         resp.max_send_sge    = attr.cap.max_send_sge;
1636         resp.max_recv_wr     = attr.cap.max_recv_wr;
1637         resp.max_send_wr     = attr.cap.max_send_wr;
1638         resp.max_inline_data = attr.cap.max_inline_data;
1639
1640         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1641                          &resp, sizeof resp)) {
1642                 ret = -EFAULT;
1643                 goto err_copy;
1644         }
1645
1646         if (xrcd)
1647                 put_xrcd_read(xrcd_uobj);
1648         if (pd)
1649                 put_pd_read(pd);
1650         if (scq)
1651                 put_cq_read(scq);
1652         if (rcq && rcq != scq)
1653                 put_cq_read(rcq);
1654         if (srq)
1655                 put_srq_read(srq);
1656
1657         mutex_lock(&file->mutex);
1658         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1659         mutex_unlock(&file->mutex);
1660
1661         obj->uevent.uobject.live = 1;
1662
1663         up_write(&obj->uevent.uobject.mutex);
1664
1665         return in_len;
1666
1667 err_copy:
1668         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1669
1670 err_destroy:
1671         ib_destroy_qp(qp);
1672
1673 err_put:
1674         if (xrcd)
1675                 put_xrcd_read(xrcd_uobj);
1676         if (pd)
1677                 put_pd_read(pd);
1678         if (scq)
1679                 put_cq_read(scq);
1680         if (rcq && rcq != scq)
1681                 put_cq_read(rcq);
1682         if (srq)
1683                 put_srq_read(srq);
1684
1685         put_uobj_write(&obj->uevent.uobject);
1686         return ret;
1687 }
1688
1689 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1690                           const char __user *buf, int in_len, int out_len)
1691 {
1692         struct ib_uverbs_open_qp        cmd;
1693         struct ib_uverbs_create_qp_resp resp;
1694         struct ib_udata                 udata;
1695         struct ib_uqp_object           *obj;
1696         struct ib_xrcd                 *xrcd;
1697         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1698         struct ib_qp                   *qp;
1699         struct ib_qp_open_attr          attr;
1700         int ret;
1701
1702         if (out_len < sizeof resp)
1703                 return -ENOSPC;
1704
1705         if (copy_from_user(&cmd, buf, sizeof cmd))
1706                 return -EFAULT;
1707
1708         INIT_UDATA(&udata, buf + sizeof cmd,
1709                    (unsigned long) cmd.response + sizeof resp,
1710                    in_len - sizeof cmd, out_len - sizeof resp);
1711
1712         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1713         if (!obj)
1714                 return -ENOMEM;
1715
1716         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1717         down_write(&obj->uevent.uobject.mutex);
1718
1719         xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1720         if (!xrcd) {
1721                 ret = -EINVAL;
1722                 goto err_put;
1723         }
1724
1725         attr.event_handler = ib_uverbs_qp_event_handler;
1726         attr.qp_context    = file;
1727         attr.qp_num        = cmd.qpn;
1728         attr.qp_type       = cmd.qp_type;
1729
1730         obj->uevent.events_reported = 0;
1731         INIT_LIST_HEAD(&obj->uevent.event_list);
1732         INIT_LIST_HEAD(&obj->mcast_list);
1733
1734         qp = ib_open_qp(xrcd, &attr);
1735         if (IS_ERR(qp)) {
1736                 ret = PTR_ERR(qp);
1737                 goto err_put;
1738         }
1739
1740         qp->uobject = &obj->uevent.uobject;
1741
1742         obj->uevent.uobject.object = qp;
1743         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1744         if (ret)
1745                 goto err_destroy;
1746
1747         memset(&resp, 0, sizeof resp);
1748         resp.qpn       = qp->qp_num;
1749         resp.qp_handle = obj->uevent.uobject.id;
1750
1751         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1752                          &resp, sizeof resp)) {
1753                 ret = -EFAULT;
1754                 goto err_remove;
1755         }
1756
1757         put_xrcd_read(xrcd_uobj);
1758
1759         mutex_lock(&file->mutex);
1760         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1761         mutex_unlock(&file->mutex);
1762
1763         obj->uevent.uobject.live = 1;
1764
1765         up_write(&obj->uevent.uobject.mutex);
1766
1767         return in_len;
1768
1769 err_remove:
1770         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1771
1772 err_destroy:
1773         ib_destroy_qp(qp);
1774
1775 err_put:
1776         put_xrcd_read(xrcd_uobj);
1777         put_uobj_write(&obj->uevent.uobject);
1778         return ret;
1779 }
1780
1781 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1782                            const char __user *buf, int in_len,
1783                            int out_len)
1784 {
1785         struct ib_uverbs_query_qp      cmd;
1786         struct ib_uverbs_query_qp_resp resp;
1787         struct ib_qp                   *qp;
1788         struct ib_qp_attr              *attr;
1789         struct ib_qp_init_attr         *init_attr;
1790         int                            ret;
1791
1792         if (copy_from_user(&cmd, buf, sizeof cmd))
1793                 return -EFAULT;
1794
1795         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1796         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1797         if (!attr || !init_attr) {
1798                 ret = -ENOMEM;
1799                 goto out;
1800         }
1801
1802         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1803         if (!qp) {
1804                 ret = -EINVAL;
1805                 goto out;
1806         }
1807
1808         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1809
1810         put_qp_read(qp);
1811
1812         if (ret)
1813                 goto out;
1814
1815         memset(&resp, 0, sizeof resp);
1816
1817         resp.qp_state               = attr->qp_state;
1818         resp.cur_qp_state           = attr->cur_qp_state;
1819         resp.path_mtu               = attr->path_mtu;
1820         resp.path_mig_state         = attr->path_mig_state;
1821         resp.qkey                   = attr->qkey;
1822         resp.rq_psn                 = attr->rq_psn;
1823         resp.sq_psn                 = attr->sq_psn;
1824         resp.dest_qp_num            = attr->dest_qp_num;
1825         resp.qp_access_flags        = attr->qp_access_flags;
1826         resp.pkey_index             = attr->pkey_index;
1827         resp.alt_pkey_index         = attr->alt_pkey_index;
1828         resp.sq_draining            = attr->sq_draining;
1829         resp.max_rd_atomic          = attr->max_rd_atomic;
1830         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1831         resp.min_rnr_timer          = attr->min_rnr_timer;
1832         resp.port_num               = attr->port_num;
1833         resp.timeout                = attr->timeout;
1834         resp.retry_cnt              = attr->retry_cnt;
1835         resp.rnr_retry              = attr->rnr_retry;
1836         resp.alt_port_num           = attr->alt_port_num;
1837         resp.alt_timeout            = attr->alt_timeout;
1838
1839         memcpy(resp.dest.dgid, attr->ah_attr.grh.dgid.raw, 16);
1840         resp.dest.flow_label        = attr->ah_attr.grh.flow_label;
1841         resp.dest.sgid_index        = attr->ah_attr.grh.sgid_index;
1842         resp.dest.hop_limit         = attr->ah_attr.grh.hop_limit;
1843         resp.dest.traffic_class     = attr->ah_attr.grh.traffic_class;
1844         resp.dest.dlid              = attr->ah_attr.dlid;
1845         resp.dest.sl                = attr->ah_attr.sl;
1846         resp.dest.src_path_bits     = attr->ah_attr.src_path_bits;
1847         resp.dest.static_rate       = attr->ah_attr.static_rate;
1848         resp.dest.is_global         = !!(attr->ah_attr.ah_flags & IB_AH_GRH);
1849         resp.dest.port_num          = attr->ah_attr.port_num;
1850
1851         memcpy(resp.alt_dest.dgid, attr->alt_ah_attr.grh.dgid.raw, 16);
1852         resp.alt_dest.flow_label    = attr->alt_ah_attr.grh.flow_label;
1853         resp.alt_dest.sgid_index    = attr->alt_ah_attr.grh.sgid_index;
1854         resp.alt_dest.hop_limit     = attr->alt_ah_attr.grh.hop_limit;
1855         resp.alt_dest.traffic_class = attr->alt_ah_attr.grh.traffic_class;
1856         resp.alt_dest.dlid          = attr->alt_ah_attr.dlid;
1857         resp.alt_dest.sl            = attr->alt_ah_attr.sl;
1858         resp.alt_dest.src_path_bits = attr->alt_ah_attr.src_path_bits;
1859         resp.alt_dest.static_rate   = attr->alt_ah_attr.static_rate;
1860         resp.alt_dest.is_global     = !!(attr->alt_ah_attr.ah_flags & IB_AH_GRH);
1861         resp.alt_dest.port_num      = attr->alt_ah_attr.port_num;
1862
1863         resp.max_send_wr            = init_attr->cap.max_send_wr;
1864         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1865         resp.max_send_sge           = init_attr->cap.max_send_sge;
1866         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1867         resp.max_inline_data        = init_attr->cap.max_inline_data;
1868         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1869
1870         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1871                          &resp, sizeof resp))
1872                 ret = -EFAULT;
1873
1874 out:
1875         kfree(attr);
1876         kfree(init_attr);
1877
1878         return ret ? ret : in_len;
1879 }
1880
1881 /* Remove ignored fields set in the attribute mask */
1882 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1883 {
1884         switch (qp_type) {
1885         case IB_QPT_XRC_INI:
1886                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1887         case IB_QPT_XRC_TGT:
1888                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1889                                 IB_QP_RNR_RETRY);
1890         default:
1891                 return mask;
1892         }
1893 }
1894
1895 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
1896                             const char __user *buf, int in_len,
1897                             int out_len)
1898 {
1899         struct ib_uverbs_modify_qp cmd;
1900         struct ib_udata            udata;
1901         struct ib_qp              *qp;
1902         struct ib_qp_attr         *attr;
1903         int                        ret;
1904
1905         if (copy_from_user(&cmd, buf, sizeof cmd))
1906                 return -EFAULT;
1907
1908         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
1909                    out_len);
1910
1911         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1912         if (!attr)
1913                 return -ENOMEM;
1914
1915         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1916         if (!qp) {
1917                 ret = -EINVAL;
1918                 goto out;
1919         }
1920
1921         attr->qp_state            = cmd.qp_state;
1922         attr->cur_qp_state        = cmd.cur_qp_state;
1923         attr->path_mtu            = cmd.path_mtu;
1924         attr->path_mig_state      = cmd.path_mig_state;
1925         attr->qkey                = cmd.qkey;
1926         attr->rq_psn              = cmd.rq_psn;
1927         attr->sq_psn              = cmd.sq_psn;
1928         attr->dest_qp_num         = cmd.dest_qp_num;
1929         attr->qp_access_flags     = cmd.qp_access_flags;
1930         attr->pkey_index          = cmd.pkey_index;
1931         attr->alt_pkey_index      = cmd.alt_pkey_index;
1932         attr->en_sqd_async_notify = cmd.en_sqd_async_notify;
1933         attr->max_rd_atomic       = cmd.max_rd_atomic;
1934         attr->max_dest_rd_atomic  = cmd.max_dest_rd_atomic;
1935         attr->min_rnr_timer       = cmd.min_rnr_timer;
1936         attr->port_num            = cmd.port_num;
1937         attr->timeout             = cmd.timeout;
1938         attr->retry_cnt           = cmd.retry_cnt;
1939         attr->rnr_retry           = cmd.rnr_retry;
1940         attr->alt_port_num        = cmd.alt_port_num;
1941         attr->alt_timeout         = cmd.alt_timeout;
1942
1943         memcpy(attr->ah_attr.grh.dgid.raw, cmd.dest.dgid, 16);
1944         attr->ah_attr.grh.flow_label        = cmd.dest.flow_label;
1945         attr->ah_attr.grh.sgid_index        = cmd.dest.sgid_index;
1946         attr->ah_attr.grh.hop_limit         = cmd.dest.hop_limit;
1947         attr->ah_attr.grh.traffic_class     = cmd.dest.traffic_class;
1948         attr->ah_attr.dlid                  = cmd.dest.dlid;
1949         attr->ah_attr.sl                    = cmd.dest.sl;
1950         attr->ah_attr.src_path_bits         = cmd.dest.src_path_bits;
1951         attr->ah_attr.static_rate           = cmd.dest.static_rate;
1952         attr->ah_attr.ah_flags              = cmd.dest.is_global ? IB_AH_GRH : 0;
1953         attr->ah_attr.port_num              = cmd.dest.port_num;
1954
1955         memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd.alt_dest.dgid, 16);
1956         attr->alt_ah_attr.grh.flow_label    = cmd.alt_dest.flow_label;
1957         attr->alt_ah_attr.grh.sgid_index    = cmd.alt_dest.sgid_index;
1958         attr->alt_ah_attr.grh.hop_limit     = cmd.alt_dest.hop_limit;
1959         attr->alt_ah_attr.grh.traffic_class = cmd.alt_dest.traffic_class;
1960         attr->alt_ah_attr.dlid              = cmd.alt_dest.dlid;
1961         attr->alt_ah_attr.sl                = cmd.alt_dest.sl;
1962         attr->alt_ah_attr.src_path_bits     = cmd.alt_dest.src_path_bits;
1963         attr->alt_ah_attr.static_rate       = cmd.alt_dest.static_rate;
1964         attr->alt_ah_attr.ah_flags          = cmd.alt_dest.is_global ? IB_AH_GRH : 0;
1965         attr->alt_ah_attr.port_num          = cmd.alt_dest.port_num;
1966
1967         if (qp->real_qp == qp) {
1968                 ret = qp->device->modify_qp(qp, attr,
1969                         modify_qp_mask(qp->qp_type, cmd.attr_mask), &udata);
1970         } else {
1971                 ret = ib_modify_qp(qp, attr, modify_qp_mask(qp->qp_type, cmd.attr_mask));
1972         }
1973
1974         put_qp_read(qp);
1975
1976         if (ret)
1977                 goto out;
1978
1979         ret = in_len;
1980
1981 out:
1982         kfree(attr);
1983
1984         return ret;
1985 }
1986
1987 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
1988                              const char __user *buf, int in_len,
1989                              int out_len)
1990 {
1991         struct ib_uverbs_destroy_qp      cmd;
1992         struct ib_uverbs_destroy_qp_resp resp;
1993         struct ib_uobject               *uobj;
1994         struct ib_qp                    *qp;
1995         struct ib_uqp_object            *obj;
1996         int                              ret = -EINVAL;
1997
1998         if (copy_from_user(&cmd, buf, sizeof cmd))
1999                 return -EFAULT;
2000
2001         memset(&resp, 0, sizeof resp);
2002
2003         uobj = idr_write_uobj(&ib_uverbs_qp_idr, cmd.qp_handle, file->ucontext);
2004         if (!uobj)
2005                 return -EINVAL;
2006         qp  = uobj->object;
2007         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2008
2009         if (!list_empty(&obj->mcast_list)) {
2010                 put_uobj_write(uobj);
2011                 return -EBUSY;
2012         }
2013
2014         ret = ib_destroy_qp(qp);
2015         if (!ret)
2016                 uobj->live = 0;
2017
2018         put_uobj_write(uobj);
2019
2020         if (ret)
2021                 return ret;
2022
2023         idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
2024
2025         mutex_lock(&file->mutex);
2026         list_del(&uobj->list);
2027         mutex_unlock(&file->mutex);
2028
2029         ib_uverbs_release_uevent(file, &obj->uevent);
2030
2031         resp.events_reported = obj->uevent.events_reported;
2032
2033         put_uobj(uobj);
2034
2035         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2036                          &resp, sizeof resp))
2037                 return -EFAULT;
2038
2039         return in_len;
2040 }
2041
2042 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2043                             const char __user *buf, int in_len,
2044                             int out_len)
2045 {
2046         struct ib_uverbs_post_send      cmd;
2047         struct ib_uverbs_post_send_resp resp;
2048         struct ib_uverbs_send_wr       *user_wr;
2049         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
2050         struct ib_qp                   *qp;
2051         int                             i, sg_ind;
2052         int                             is_ud;
2053         ssize_t                         ret = -EINVAL;
2054
2055         if (copy_from_user(&cmd, buf, sizeof cmd))
2056                 return -EFAULT;
2057
2058         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2059             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2060                 return -EINVAL;
2061
2062         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2063                 return -EINVAL;
2064
2065         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2066         if (!user_wr)
2067                 return -ENOMEM;
2068
2069         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2070         if (!qp)
2071                 goto out;
2072
2073         is_ud = qp->qp_type == IB_QPT_UD;
2074         sg_ind = 0;
2075         last = NULL;
2076         for (i = 0; i < cmd.wr_count; ++i) {
2077                 if (copy_from_user(user_wr,
2078                                    buf + sizeof cmd + i * cmd.wqe_size,
2079                                    cmd.wqe_size)) {
2080                         ret = -EFAULT;
2081                         goto out_put;
2082                 }
2083
2084                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2085                         ret = -EINVAL;
2086                         goto out_put;
2087                 }
2088
2089                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2090                                user_wr->num_sge * sizeof (struct ib_sge),
2091                                GFP_KERNEL);
2092                 if (!next) {
2093                         ret = -ENOMEM;
2094                         goto out_put;
2095                 }
2096
2097                 if (!last)
2098                         wr = next;
2099                 else
2100                         last->next = next;
2101                 last = next;
2102
2103                 next->next       = NULL;
2104                 next->wr_id      = user_wr->wr_id;
2105                 next->num_sge    = user_wr->num_sge;
2106                 next->opcode     = user_wr->opcode;
2107                 next->send_flags = user_wr->send_flags;
2108
2109                 if (is_ud) {
2110                         next->wr.ud.ah = idr_read_ah(user_wr->wr.ud.ah,
2111                                                      file->ucontext);
2112                         if (!next->wr.ud.ah) {
2113                                 ret = -EINVAL;
2114                                 goto out_put;
2115                         }
2116                         next->wr.ud.remote_qpn  = user_wr->wr.ud.remote_qpn;
2117                         next->wr.ud.remote_qkey = user_wr->wr.ud.remote_qkey;
2118                 } else {
2119                         switch (next->opcode) {
2120                         case IB_WR_RDMA_WRITE_WITH_IMM:
2121                                 next->ex.imm_data =
2122                                         (__be32 __force) user_wr->ex.imm_data;
2123                         case IB_WR_RDMA_WRITE:
2124                         case IB_WR_RDMA_READ:
2125                                 next->wr.rdma.remote_addr =
2126                                         user_wr->wr.rdma.remote_addr;
2127                                 next->wr.rdma.rkey        =
2128                                         user_wr->wr.rdma.rkey;
2129                                 break;
2130                         case IB_WR_SEND_WITH_IMM:
2131                                 next->ex.imm_data =
2132                                         (__be32 __force) user_wr->ex.imm_data;
2133                                 break;
2134                         case IB_WR_SEND_WITH_INV:
2135                                 next->ex.invalidate_rkey =
2136                                         user_wr->ex.invalidate_rkey;
2137                                 break;
2138                         case IB_WR_ATOMIC_CMP_AND_SWP:
2139                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2140                                 next->wr.atomic.remote_addr =
2141                                         user_wr->wr.atomic.remote_addr;
2142                                 next->wr.atomic.compare_add =
2143                                         user_wr->wr.atomic.compare_add;
2144                                 next->wr.atomic.swap = user_wr->wr.atomic.swap;
2145                                 next->wr.atomic.rkey = user_wr->wr.atomic.rkey;
2146                                 break;
2147                         default:
2148                                 break;
2149                         }
2150                 }
2151
2152                 if (next->num_sge) {
2153                         next->sg_list = (void *) next +
2154                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2155                         if (copy_from_user(next->sg_list,
2156                                            buf + sizeof cmd +
2157                                            cmd.wr_count * cmd.wqe_size +
2158                                            sg_ind * sizeof (struct ib_sge),
2159                                            next->num_sge * sizeof (struct ib_sge))) {
2160                                 ret = -EFAULT;
2161                                 goto out_put;
2162                         }
2163                         sg_ind += next->num_sge;
2164                 } else
2165                         next->sg_list = NULL;
2166         }
2167
2168         resp.bad_wr = 0;
2169         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2170         if (ret)
2171                 for (next = wr; next; next = next->next) {
2172                         ++resp.bad_wr;
2173                         if (next == bad_wr)
2174                                 break;
2175                 }
2176
2177         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2178                          &resp, sizeof resp))
2179                 ret = -EFAULT;
2180
2181 out_put:
2182         put_qp_read(qp);
2183
2184         while (wr) {
2185                 if (is_ud && wr->wr.ud.ah)
2186                         put_ah_read(wr->wr.ud.ah);
2187                 next = wr->next;
2188                 kfree(wr);
2189                 wr = next;
2190         }
2191
2192 out:
2193         kfree(user_wr);
2194
2195         return ret ? ret : in_len;
2196 }
2197
2198 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2199                                                     int in_len,
2200                                                     u32 wr_count,
2201                                                     u32 sge_count,
2202                                                     u32 wqe_size)
2203 {
2204         struct ib_uverbs_recv_wr *user_wr;
2205         struct ib_recv_wr        *wr = NULL, *last, *next;
2206         int                       sg_ind;
2207         int                       i;
2208         int                       ret;
2209
2210         if (in_len < wqe_size * wr_count +
2211             sge_count * sizeof (struct ib_uverbs_sge))
2212                 return ERR_PTR(-EINVAL);
2213
2214         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2215                 return ERR_PTR(-EINVAL);
2216
2217         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2218         if (!user_wr)
2219                 return ERR_PTR(-ENOMEM);
2220
2221         sg_ind = 0;
2222         last = NULL;
2223         for (i = 0; i < wr_count; ++i) {
2224                 if (copy_from_user(user_wr, buf + i * wqe_size,
2225                                    wqe_size)) {
2226                         ret = -EFAULT;
2227                         goto err;
2228                 }
2229
2230                 if (user_wr->num_sge + sg_ind > sge_count) {
2231                         ret = -EINVAL;
2232                         goto err;
2233                 }
2234
2235                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2236                                user_wr->num_sge * sizeof (struct ib_sge),
2237                                GFP_KERNEL);
2238                 if (!next) {
2239                         ret = -ENOMEM;
2240                         goto err;
2241                 }
2242
2243                 if (!last)
2244                         wr = next;
2245                 else
2246                         last->next = next;
2247                 last = next;
2248
2249                 next->next       = NULL;
2250                 next->wr_id      = user_wr->wr_id;
2251                 next->num_sge    = user_wr->num_sge;
2252
2253                 if (next->num_sge) {
2254                         next->sg_list = (void *) next +
2255                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2256                         if (copy_from_user(next->sg_list,
2257                                            buf + wr_count * wqe_size +
2258                                            sg_ind * sizeof (struct ib_sge),
2259                                            next->num_sge * sizeof (struct ib_sge))) {
2260                                 ret = -EFAULT;
2261                                 goto err;
2262                         }
2263                         sg_ind += next->num_sge;
2264                 } else
2265                         next->sg_list = NULL;
2266         }
2267
2268         kfree(user_wr);
2269         return wr;
2270
2271 err:
2272         kfree(user_wr);
2273
2274         while (wr) {
2275                 next = wr->next;
2276                 kfree(wr);
2277                 wr = next;
2278         }
2279
2280         return ERR_PTR(ret);
2281 }
2282
2283 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2284                             const char __user *buf, int in_len,
2285                             int out_len)
2286 {
2287         struct ib_uverbs_post_recv      cmd;
2288         struct ib_uverbs_post_recv_resp resp;
2289         struct ib_recv_wr              *wr, *next, *bad_wr;
2290         struct ib_qp                   *qp;
2291         ssize_t                         ret = -EINVAL;
2292
2293         if (copy_from_user(&cmd, buf, sizeof cmd))
2294                 return -EFAULT;
2295
2296         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2297                                        in_len - sizeof cmd, cmd.wr_count,
2298                                        cmd.sge_count, cmd.wqe_size);
2299         if (IS_ERR(wr))
2300                 return PTR_ERR(wr);
2301
2302         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2303         if (!qp)
2304                 goto out;
2305
2306         resp.bad_wr = 0;
2307         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2308
2309         put_qp_read(qp);
2310
2311         if (ret)
2312                 for (next = wr; next; next = next->next) {
2313                         ++resp.bad_wr;
2314                         if (next == bad_wr)
2315                                 break;
2316                 }
2317
2318         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2319                          &resp, sizeof resp))
2320                 ret = -EFAULT;
2321
2322 out:
2323         while (wr) {
2324                 next = wr->next;
2325                 kfree(wr);
2326                 wr = next;
2327         }
2328
2329         return ret ? ret : in_len;
2330 }
2331
2332 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2333                                 const char __user *buf, int in_len,
2334                                 int out_len)
2335 {
2336         struct ib_uverbs_post_srq_recv      cmd;
2337         struct ib_uverbs_post_srq_recv_resp resp;
2338         struct ib_recv_wr                  *wr, *next, *bad_wr;
2339         struct ib_srq                      *srq;
2340         ssize_t                             ret = -EINVAL;
2341
2342         if (copy_from_user(&cmd, buf, sizeof cmd))
2343                 return -EFAULT;
2344
2345         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2346                                        in_len - sizeof cmd, cmd.wr_count,
2347                                        cmd.sge_count, cmd.wqe_size);
2348         if (IS_ERR(wr))
2349                 return PTR_ERR(wr);
2350
2351         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2352         if (!srq)
2353                 goto out;
2354
2355         resp.bad_wr = 0;
2356         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2357
2358         put_srq_read(srq);
2359
2360         if (ret)
2361                 for (next = wr; next; next = next->next) {
2362                         ++resp.bad_wr;
2363                         if (next == bad_wr)
2364                                 break;
2365                 }
2366
2367         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2368                          &resp, sizeof resp))
2369                 ret = -EFAULT;
2370
2371 out:
2372         while (wr) {
2373                 next = wr->next;
2374                 kfree(wr);
2375                 wr = next;
2376         }
2377
2378         return ret ? ret : in_len;
2379 }
2380
2381 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2382                             const char __user *buf, int in_len,
2383                             int out_len)
2384 {
2385         struct ib_uverbs_create_ah       cmd;
2386         struct ib_uverbs_create_ah_resp  resp;
2387         struct ib_uobject               *uobj;
2388         struct ib_pd                    *pd;
2389         struct ib_ah                    *ah;
2390         struct ib_ah_attr               attr;
2391         int ret;
2392
2393         if (out_len < sizeof resp)
2394                 return -ENOSPC;
2395
2396         if (copy_from_user(&cmd, buf, sizeof cmd))
2397                 return -EFAULT;
2398
2399         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
2400         if (!uobj)
2401                 return -ENOMEM;
2402
2403         init_uobj(uobj, cmd.user_handle, file->ucontext, &ah_lock_class);
2404         down_write(&uobj->mutex);
2405
2406         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
2407         if (!pd) {
2408                 ret = -EINVAL;
2409                 goto err;
2410         }
2411
2412         attr.dlid              = cmd.attr.dlid;
2413         attr.sl                = cmd.attr.sl;
2414         attr.src_path_bits     = cmd.attr.src_path_bits;
2415         attr.static_rate       = cmd.attr.static_rate;
2416         attr.ah_flags          = cmd.attr.is_global ? IB_AH_GRH : 0;
2417         attr.port_num          = cmd.attr.port_num;
2418         attr.grh.flow_label    = cmd.attr.grh.flow_label;
2419         attr.grh.sgid_index    = cmd.attr.grh.sgid_index;
2420         attr.grh.hop_limit     = cmd.attr.grh.hop_limit;
2421         attr.grh.traffic_class = cmd.attr.grh.traffic_class;
2422         memcpy(attr.grh.dgid.raw, cmd.attr.grh.dgid, 16);
2423
2424         ah = ib_create_ah(pd, &attr);
2425         if (IS_ERR(ah)) {
2426                 ret = PTR_ERR(ah);
2427                 goto err_put;
2428         }
2429
2430         ah->uobject  = uobj;
2431         uobj->object = ah;
2432
2433         ret = idr_add_uobj(&ib_uverbs_ah_idr, uobj);
2434         if (ret)
2435                 goto err_destroy;
2436
2437         resp.ah_handle = uobj->id;
2438
2439         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2440                          &resp, sizeof resp)) {
2441                 ret = -EFAULT;
2442                 goto err_copy;
2443         }
2444
2445         put_pd_read(pd);
2446
2447         mutex_lock(&file->mutex);
2448         list_add_tail(&uobj->list, &file->ucontext->ah_list);
2449         mutex_unlock(&file->mutex);
2450
2451         uobj->live = 1;
2452
2453         up_write(&uobj->mutex);
2454
2455         return in_len;
2456
2457 err_copy:
2458         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2459
2460 err_destroy:
2461         ib_destroy_ah(ah);
2462
2463 err_put:
2464         put_pd_read(pd);
2465
2466 err:
2467         put_uobj_write(uobj);
2468         return ret;
2469 }
2470
2471 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2472                              const char __user *buf, int in_len, int out_len)
2473 {
2474         struct ib_uverbs_destroy_ah cmd;
2475         struct ib_ah               *ah;
2476         struct ib_uobject          *uobj;
2477         int                         ret;
2478
2479         if (copy_from_user(&cmd, buf, sizeof cmd))
2480                 return -EFAULT;
2481
2482         uobj = idr_write_uobj(&ib_uverbs_ah_idr, cmd.ah_handle, file->ucontext);
2483         if (!uobj)
2484                 return -EINVAL;
2485         ah = uobj->object;
2486
2487         ret = ib_destroy_ah(ah);
2488         if (!ret)
2489                 uobj->live = 0;
2490
2491         put_uobj_write(uobj);
2492
2493         if (ret)
2494                 return ret;
2495
2496         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2497
2498         mutex_lock(&file->mutex);
2499         list_del(&uobj->list);
2500         mutex_unlock(&file->mutex);
2501
2502         put_uobj(uobj);
2503
2504         return in_len;
2505 }
2506
2507 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2508                                const char __user *buf, int in_len,
2509                                int out_len)
2510 {
2511         struct ib_uverbs_attach_mcast cmd;
2512         struct ib_qp                 *qp;
2513         struct ib_uqp_object         *obj;
2514         struct ib_uverbs_mcast_entry *mcast;
2515         int                           ret;
2516
2517         if (copy_from_user(&cmd, buf, sizeof cmd))
2518                 return -EFAULT;
2519
2520         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2521         if (!qp)
2522                 return -EINVAL;
2523
2524         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2525
2526         list_for_each_entry(mcast, &obj->mcast_list, list)
2527                 if (cmd.mlid == mcast->lid &&
2528                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2529                         ret = 0;
2530                         goto out_put;
2531                 }
2532
2533         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2534         if (!mcast) {
2535                 ret = -ENOMEM;
2536                 goto out_put;
2537         }
2538
2539         mcast->lid = cmd.mlid;
2540         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2541
2542         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2543         if (!ret)
2544                 list_add_tail(&mcast->list, &obj->mcast_list);
2545         else
2546                 kfree(mcast);
2547
2548 out_put:
2549         put_qp_write(qp);
2550
2551         return ret ? ret : in_len;
2552 }
2553
2554 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2555                                const char __user *buf, int in_len,
2556                                int out_len)
2557 {
2558         struct ib_uverbs_detach_mcast cmd;
2559         struct ib_uqp_object         *obj;
2560         struct ib_qp                 *qp;
2561         struct ib_uverbs_mcast_entry *mcast;
2562         int                           ret = -EINVAL;
2563
2564         if (copy_from_user(&cmd, buf, sizeof cmd))
2565                 return -EFAULT;
2566
2567         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2568         if (!qp)
2569                 return -EINVAL;
2570
2571         ret = ib_detach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
2572         if (ret)
2573                 goto out_put;
2574
2575         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2576
2577         list_for_each_entry(mcast, &obj->mcast_list, list)
2578                 if (cmd.mlid == mcast->lid &&
2579                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2580                         list_del(&mcast->list);
2581                         kfree(mcast);
2582                         break;
2583                 }
2584
2585 out_put:
2586         put_qp_write(qp);
2587
2588         return ret ? ret : in_len;
2589 }
2590
2591 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
2592                                 struct ib_uverbs_create_xsrq *cmd,
2593                                 struct ib_udata *udata)
2594 {
2595         struct ib_uverbs_create_srq_resp resp;
2596         struct ib_usrq_object           *obj;
2597         struct ib_pd                    *pd;
2598         struct ib_srq                   *srq;
2599         struct ib_uobject               *uninitialized_var(xrcd_uobj);
2600         struct ib_srq_init_attr          attr;
2601         int ret;
2602
2603         obj = kmalloc(sizeof *obj, GFP_KERNEL);
2604         if (!obj)
2605                 return -ENOMEM;
2606
2607         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext, &srq_lock_class);
2608         down_write(&obj->uevent.uobject.mutex);
2609
2610         if (cmd->srq_type == IB_SRQT_XRC) {
2611                 attr.ext.xrc.xrcd  = idr_read_xrcd(cmd->xrcd_handle, file->ucontext, &xrcd_uobj);
2612                 if (!attr.ext.xrc.xrcd) {
2613                         ret = -EINVAL;
2614                         goto err;
2615                 }
2616
2617                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
2618                 atomic_inc(&obj->uxrcd->refcnt);
2619
2620                 attr.ext.xrc.cq  = idr_read_cq(cmd->cq_handle, file->ucontext, 0);
2621                 if (!attr.ext.xrc.cq) {
2622                         ret = -EINVAL;
2623                         goto err_put_xrcd;
2624                 }
2625         }
2626
2627         pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
2628         if (!pd) {
2629                 ret = -EINVAL;
2630                 goto err_put_cq;
2631         }
2632
2633         attr.event_handler  = ib_uverbs_srq_event_handler;
2634         attr.srq_context    = file;
2635         attr.srq_type       = cmd->srq_type;
2636         attr.attr.max_wr    = cmd->max_wr;
2637         attr.attr.max_sge   = cmd->max_sge;
2638         attr.attr.srq_limit = cmd->srq_limit;
2639
2640         obj->uevent.events_reported = 0;
2641         INIT_LIST_HEAD(&obj->uevent.event_list);
2642
2643         srq = pd->device->create_srq(pd, &attr, udata);
2644         if (IS_ERR(srq)) {
2645                 ret = PTR_ERR(srq);
2646                 goto err_put;
2647         }
2648
2649         srq->device        = pd->device;
2650         srq->pd            = pd;
2651         srq->srq_type      = cmd->srq_type;
2652         srq->uobject       = &obj->uevent.uobject;
2653         srq->event_handler = attr.event_handler;
2654         srq->srq_context   = attr.srq_context;
2655
2656         if (cmd->srq_type == IB_SRQT_XRC) {
2657                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
2658                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
2659                 atomic_inc(&attr.ext.xrc.cq->usecnt);
2660                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
2661         }
2662
2663         atomic_inc(&pd->usecnt);
2664         atomic_set(&srq->usecnt, 0);
2665
2666         obj->uevent.uobject.object = srq;
2667         ret = idr_add_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2668         if (ret)
2669                 goto err_destroy;
2670
2671         memset(&resp, 0, sizeof resp);
2672         resp.srq_handle = obj->uevent.uobject.id;
2673         resp.max_wr     = attr.attr.max_wr;
2674         resp.max_sge    = attr.attr.max_sge;
2675         if (cmd->srq_type == IB_SRQT_XRC)
2676                 resp.srqn = srq->ext.xrc.srq_num;
2677
2678         if (copy_to_user((void __user *) (unsigned long) cmd->response,
2679                          &resp, sizeof resp)) {
2680                 ret = -EFAULT;
2681                 goto err_copy;
2682         }
2683
2684         if (cmd->srq_type == IB_SRQT_XRC) {
2685                 put_uobj_read(xrcd_uobj);
2686                 put_cq_read(attr.ext.xrc.cq);
2687         }
2688         put_pd_read(pd);
2689
2690         mutex_lock(&file->mutex);
2691         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->srq_list);
2692         mutex_unlock(&file->mutex);
2693
2694         obj->uevent.uobject.live = 1;
2695
2696         up_write(&obj->uevent.uobject.mutex);
2697
2698         return 0;
2699
2700 err_copy:
2701         idr_remove_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2702
2703 err_destroy:
2704         ib_destroy_srq(srq);
2705
2706 err_put:
2707         put_pd_read(pd);
2708
2709 err_put_cq:
2710         if (cmd->srq_type == IB_SRQT_XRC)
2711                 put_cq_read(attr.ext.xrc.cq);
2712
2713 err_put_xrcd:
2714         if (cmd->srq_type == IB_SRQT_XRC) {
2715                 atomic_dec(&obj->uxrcd->refcnt);
2716                 put_uobj_read(xrcd_uobj);
2717         }
2718
2719 err:
2720         put_uobj_write(&obj->uevent.uobject);
2721         return ret;
2722 }
2723
2724 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
2725                              const char __user *buf, int in_len,
2726                              int out_len)
2727 {
2728         struct ib_uverbs_create_srq      cmd;
2729         struct ib_uverbs_create_xsrq     xcmd;
2730         struct ib_uverbs_create_srq_resp resp;
2731         struct ib_udata                  udata;
2732         int ret;
2733
2734         if (out_len < sizeof resp)
2735                 return -ENOSPC;
2736
2737         if (copy_from_user(&cmd, buf, sizeof cmd))
2738                 return -EFAULT;
2739
2740         xcmd.response    = cmd.response;
2741         xcmd.user_handle = cmd.user_handle;
2742         xcmd.srq_type    = IB_SRQT_BASIC;
2743         xcmd.pd_handle   = cmd.pd_handle;
2744         xcmd.max_wr      = cmd.max_wr;
2745         xcmd.max_sge     = cmd.max_sge;
2746         xcmd.srq_limit   = cmd.srq_limit;
2747
2748         INIT_UDATA(&udata, buf + sizeof cmd,
2749                    (unsigned long) cmd.response + sizeof resp,
2750                    in_len - sizeof cmd, out_len - sizeof resp);
2751
2752         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
2753         if (ret)
2754                 return ret;
2755
2756         return in_len;
2757 }
2758
2759 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
2760                               const char __user *buf, int in_len, int out_len)
2761 {
2762         struct ib_uverbs_create_xsrq     cmd;
2763         struct ib_uverbs_create_srq_resp resp;
2764         struct ib_udata                  udata;
2765         int ret;
2766
2767         if (out_len < sizeof resp)
2768                 return -ENOSPC;
2769
2770         if (copy_from_user(&cmd, buf, sizeof cmd))
2771                 return -EFAULT;
2772
2773         INIT_UDATA(&udata, buf + sizeof cmd,
2774                    (unsigned long) cmd.response + sizeof resp,
2775                    in_len - sizeof cmd, out_len - sizeof resp);
2776
2777         ret = __uverbs_create_xsrq(file, &cmd, &udata);
2778         if (ret)
2779                 return ret;
2780
2781         return in_len;
2782 }
2783
2784 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
2785                              const char __user *buf, int in_len,
2786                              int out_len)
2787 {
2788         struct ib_uverbs_modify_srq cmd;
2789         struct ib_udata             udata;
2790         struct ib_srq              *srq;
2791         struct ib_srq_attr          attr;
2792         int                         ret;
2793
2794         if (copy_from_user(&cmd, buf, sizeof cmd))
2795                 return -EFAULT;
2796
2797         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
2798                    out_len);
2799
2800         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2801         if (!srq)
2802                 return -EINVAL;
2803
2804         attr.max_wr    = cmd.max_wr;
2805         attr.srq_limit = cmd.srq_limit;
2806
2807         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
2808
2809         put_srq_read(srq);
2810
2811         return ret ? ret : in_len;
2812 }
2813
2814 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
2815                             const char __user *buf,
2816                             int in_len, int out_len)
2817 {
2818         struct ib_uverbs_query_srq      cmd;
2819         struct ib_uverbs_query_srq_resp resp;
2820         struct ib_srq_attr              attr;
2821         struct ib_srq                   *srq;
2822         int                             ret;
2823
2824         if (out_len < sizeof resp)
2825                 return -ENOSPC;
2826
2827         if (copy_from_user(&cmd, buf, sizeof cmd))
2828                 return -EFAULT;
2829
2830         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2831         if (!srq)
2832                 return -EINVAL;
2833
2834         ret = ib_query_srq(srq, &attr);
2835
2836         put_srq_read(srq);
2837
2838         if (ret)
2839                 return ret;
2840
2841         memset(&resp, 0, sizeof resp);
2842
2843         resp.max_wr    = attr.max_wr;
2844         resp.max_sge   = attr.max_sge;
2845         resp.srq_limit = attr.srq_limit;
2846
2847         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2848                          &resp, sizeof resp))
2849                 return -EFAULT;
2850
2851         return in_len;
2852 }
2853
2854 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
2855                               const char __user *buf, int in_len,
2856                               int out_len)
2857 {
2858         struct ib_uverbs_destroy_srq      cmd;
2859         struct ib_uverbs_destroy_srq_resp resp;
2860         struct ib_uobject                *uobj;
2861         struct ib_srq                    *srq;
2862         struct ib_uevent_object          *obj;
2863         int                               ret = -EINVAL;
2864
2865         if (copy_from_user(&cmd, buf, sizeof cmd))
2866                 return -EFAULT;
2867
2868         uobj = idr_write_uobj(&ib_uverbs_srq_idr, cmd.srq_handle, file->ucontext);
2869         if (!uobj)
2870                 return -EINVAL;
2871         srq = uobj->object;
2872         obj = container_of(uobj, struct ib_uevent_object, uobject);
2873
2874         ret = ib_destroy_srq(srq);
2875         if (!ret)
2876                 uobj->live = 0;
2877
2878         put_uobj_write(uobj);
2879
2880         if (ret)
2881                 return ret;
2882
2883         idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
2884
2885         mutex_lock(&file->mutex);
2886         list_del(&uobj->list);
2887         mutex_unlock(&file->mutex);
2888
2889         ib_uverbs_release_uevent(file, obj);
2890
2891         memset(&resp, 0, sizeof resp);
2892         resp.events_reported = obj->events_reported;
2893
2894         put_uobj(uobj);
2895
2896         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2897                          &resp, sizeof resp))
2898                 ret = -EFAULT;
2899
2900         return ret ? ret : in_len;
2901 }