drivers/infiniband/core/uverbs_cmd.c

   1 /*
   2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005, 2006, 2007 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 PathScale, Inc.  All rights reserved.
   5  * Copyright (c) 2006 Mellanox Technologies.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/file.h>
  37 #include <linux/fs.h>
  38 #include <linux/slab.h>
  39
  40 #include <asm/uaccess.h>
  41
  42 #include "uverbs.h"
  43
  44 static struct lock_class_key pd_lock_key;
  45 static struct lock_class_key mr_lock_key;
  46 static struct lock_class_key cq_lock_key;
  47 static struct lock_class_key qp_lock_key;
  48 static struct lock_class_key ah_lock_key;
  49 static struct lock_class_key srq_lock_key;
  50 static struct lock_class_key xrcd_lock_key;
  51
  52 #define INIT_UDATA(udata, ibuf, obuf, ilen, olen)                       \
  53         do {                                                            \
  54                 (udata)->inbuf  = (void __user *) (ibuf);               \
  55                 (udata)->outbuf = (void __user *) (obuf);               \
  56                 (udata)->inlen  = (ilen);                               \
  57                 (udata)->outlen = (olen);                               \
  58         } while (0)
  59
  60 /*
  61  * The ib_uobject locking scheme is as follows:
  62  *
  63  * - ib_uverbs_idr_lock protects the uverbs idrs themselves, so it
  64  *   needs to be held during all idr operations.  When an object is
  65  *   looked up, a reference must be taken on the object's kref before
  66  *   dropping this lock.
  67  *
  68  * - Each object also has an rwsem.  This rwsem must be held for
  69  *   reading while an operation that uses the object is performed.
  70  *   For example, while registering an MR, the associated PD's
  71  *   uobject.mutex must be held for reading.  The rwsem must be held
  72  *   for writing while initializing or destroying an object.
  73  *
  74  * - In addition, each object has a "live" flag.  If this flag is not
  75  *   set, then lookups of the object will fail even if it is found in
  76  *   the idr.  This handles a reader that blocks and does not acquire
  77  *   the rwsem until after the object is destroyed.  The destroy
  78  *   operation will set the live flag to 0 and then drop the rwsem;
  79  *   this will allow the reader to acquire the rwsem, see that the
  80  *   live flag is 0, and then drop the rwsem and its reference to
  81  *   object.  The underlying storage will not be freed until the last
  82  *   reference to the object is dropped.
  83  */
  84
  85 static void init_uobj(struct ib_uobject *uobj, u64 user_handle,
  86                       struct ib_ucontext *context, struct lock_class_key *key)
  87 {
  88         uobj->user_handle = user_handle;
  89         uobj->context     = context;
  90         kref_init(&uobj->ref);
  91         init_rwsem(&uobj->mutex);
  92         lockdep_set_class(&uobj->mutex, key);
  93         uobj->live        = 0;
  94 }
  95
  96 static void release_uobj(struct kref *kref)
  97 {
  98         kfree(container_of(kref, struct ib_uobject, ref));
  99 }
 100
 101 static void put_uobj(struct ib_uobject *uobj)
 102 {
 103         kref_put(&uobj->ref, release_uobj);
 104 }
 105
 106 static void put_uobj_read(struct ib_uobject *uobj)
 107 {
 108         up_read(&uobj->mutex);
 109         put_uobj(uobj);
 110 }
 111
 112 static void put_uobj_write(struct ib_uobject *uobj)
 113 {
 114         up_write(&uobj->mutex);
 115         put_uobj(uobj);
 116 }
 117
 118 static int idr_add_uobj(struct idr *idr, struct ib_uobject *uobj)
 119 {
 120         int ret;
 121
 122 retry:
 123         if (!idr_pre_get(idr, GFP_KERNEL))
 124                 return -ENOMEM;
 125
 126         spin_lock(&ib_uverbs_idr_lock);
 127         ret = idr_get_new(idr, uobj, &uobj->id);
 128         spin_unlock(&ib_uverbs_idr_lock);
 129
 130         if (ret == -EAGAIN)
 131                 goto retry;
 132
 133         return ret;
 134 }
 135
 136 void idr_remove_uobj(struct idr *idr, struct ib_uobject *uobj)
 137 {
 138         spin_lock(&ib_uverbs_idr_lock);
 139         idr_remove(idr, uobj->id);
 140         spin_unlock(&ib_uverbs_idr_lock);
 141 }
 142
 143 static struct ib_uobject *__idr_get_uobj(struct idr *idr, int id,
 144                                          struct ib_ucontext *context)
 145 {
 146         struct ib_uobject *uobj;
 147
 148         spin_lock(&ib_uverbs_idr_lock);
 149         uobj = idr_find(idr, id);
 150         if (uobj) {
 151                 if (uobj->context == context)
 152                         kref_get(&uobj->ref);
 153                 else
 154                         uobj = NULL;
 155         }
 156         spin_unlock(&ib_uverbs_idr_lock);
 157
 158         return uobj;
 159 }
 160
 161 static struct ib_uobject *idr_read_uobj(struct idr *idr, int id,
 162                                         struct ib_ucontext *context, int nested)
 163 {
 164         struct ib_uobject *uobj;
 165
 166         uobj = __idr_get_uobj(idr, id, context);
 167         if (!uobj)
 168                 return NULL;
 169
 170         if (nested)
 171                 down_read_nested(&uobj->mutex, SINGLE_DEPTH_NESTING);
 172         else
 173                 down_read(&uobj->mutex);
 174         if (!uobj->live) {
 175                 put_uobj_read(uobj);
 176                 return NULL;
 177         }
 178
 179         return uobj;
 180 }
 181
 182 static struct ib_uobject *idr_write_uobj(struct idr *idr, int id,
 183                                          struct ib_ucontext *context)
 184 {
 185         struct ib_uobject *uobj;
 186
 187         uobj = __idr_get_uobj(idr, id, context);
 188         if (!uobj)
 189                 return NULL;
 190
 191         down_write(&uobj->mutex);
 192         if (!uobj->live) {
 193                 put_uobj_write(uobj);
 194                 return NULL;
 195         }
 196
 197         return uobj;
 198 }
 199
 200 static void *idr_read_obj(struct idr *idr, int id, struct ib_ucontext *context,
 201                           int nested)
 202 {
 203         struct ib_uobject *uobj;
 204
 205         uobj = idr_read_uobj(idr, id, context, nested);
 206         return uobj ? uobj->object : NULL;
 207 }
 208
 209 static struct ib_pd *idr_read_pd(int pd_handle, struct ib_ucontext *context)
 210 {
 211         return idr_read_obj(&ib_uverbs_pd_idr, pd_handle, context, 0);
 212 }
 213
 214 static void put_pd_read(struct ib_pd *pd)
 215 {
 216         put_uobj_read(pd->uobject);
 217 }
 218
 219 static struct ib_cq *idr_read_cq(int cq_handle, struct ib_ucontext *context, int nested)
 220 {
 221         return idr_read_obj(&ib_uverbs_cq_idr, cq_handle, context, nested);
 222 }
 223
 224 static void put_cq_read(struct ib_cq *cq)
 225 {
 226         put_uobj_read(cq->uobject);
 227 }
 228
 229 static struct ib_ah *idr_read_ah(int ah_handle, struct ib_ucontext *context)
 230 {
 231         return idr_read_obj(&ib_uverbs_ah_idr, ah_handle, context, 0);
 232 }
 233
 234 static void put_ah_read(struct ib_ah *ah)
 235 {
 236         put_uobj_read(ah->uobject);
 237 }
 238
 239 static struct ib_qp *idr_read_qp(int qp_handle, struct ib_ucontext *context)
 240 {
 241         return idr_read_obj(&ib_uverbs_qp_idr, qp_handle, context, 0);
 242 }
 243
 244 static struct ib_qp *idr_write_qp(int qp_handle, struct ib_ucontext *context)
 245 {
 246         struct ib_uobject *uobj;
 247
 248         uobj = idr_write_uobj(&ib_uverbs_qp_idr, qp_handle, context);
 249         return uobj ? uobj->object : NULL;
 250 }
 251
 252 static void put_qp_read(struct ib_qp *qp)
 253 {
 254         put_uobj_read(qp->uobject);
 255 }
 256
 257 static void put_qp_write(struct ib_qp *qp)
 258 {
 259         put_uobj_write(qp->uobject);
 260 }
 261
 262 static struct ib_srq *idr_read_srq(int srq_handle, struct ib_ucontext *context)
 263 {
 264         return idr_read_obj(&ib_uverbs_srq_idr, srq_handle, context, 0);
 265 }
 266
 267 static void put_srq_read(struct ib_srq *srq)
 268 {
 269         put_uobj_read(srq->uobject);
 270 }
 271
 272 static struct ib_xrcd *idr_read_xrcd(int xrcd_handle, struct ib_ucontext *context,
 273                                      struct ib_uobject **uobj)
 274 {
 275         *uobj = idr_read_uobj(&ib_uverbs_xrcd_idr, xrcd_handle, context, 0);
 276         return *uobj ? (*uobj)->object : NULL;
 277 }
 278
 279 static void put_xrcd_read(struct ib_uobject *uobj)
 280 {
 281         put_uobj_read(uobj);
 282 }
 283
 284 ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
 285                               const char __user *buf,
 286                               int in_len, int out_len)
 287 {
 288         struct ib_uverbs_get_context      cmd;
 289         struct ib_uverbs_get_context_resp resp;
 290         struct ib_udata                   udata;
 291         struct ib_device                 *ibdev = file->device->ib_dev;
 292         struct ib_ucontext               *ucontext;
 293         struct file                      *filp;
 294         int ret;
 295
 296         if (out_len < sizeof resp)
 297                 return -ENOSPC;
 298
 299         if (copy_from_user(&cmd, buf, sizeof cmd))
 300                 return -EFAULT;
 301
 302         mutex_lock(&file->mutex);
 303
 304         if (file->ucontext) {
 305                 ret = -EINVAL;
 306                 goto err;
 307         }
 308
 309         INIT_UDATA(&udata, buf + sizeof cmd,
 310                    (unsigned long) cmd.response + sizeof resp,
 311                    in_len - sizeof cmd, out_len - sizeof resp);
 312
 313         ucontext = ibdev->alloc_ucontext(ibdev, &udata);
 314         if (IS_ERR(ucontext)) {
 315                 ret = PTR_ERR(ucontext);
 316                 goto err;
 317         }
 318
 319         ucontext->device = ibdev;
 320         INIT_LIST_HEAD(&ucontext->pd_list);
 321         INIT_LIST_HEAD(&ucontext->mr_list);
 322         INIT_LIST_HEAD(&ucontext->mw_list);
 323         INIT_LIST_HEAD(&ucontext->cq_list);
 324         INIT_LIST_HEAD(&ucontext->qp_list);
 325         INIT_LIST_HEAD(&ucontext->srq_list);
 326         INIT_LIST_HEAD(&ucontext->ah_list);
 327         INIT_LIST_HEAD(&ucontext->xrcd_list);
 328         ucontext->closing = 0;
 329
 330         resp.num_comp_vectors = file->device->num_comp_vectors;
 331
 332         ret = get_unused_fd();
 333         if (ret < 0)
 334                 goto err_free;
 335         resp.async_fd = ret;
 336
 337         filp = ib_uverbs_alloc_event_file(file, 1);
 338         if (IS_ERR(filp)) {
 339                 ret = PTR_ERR(filp);
 340                 goto err_fd;
 341         }
 342
 343         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 344                          &resp, sizeof resp)) {
 345                 ret = -EFAULT;
 346                 goto err_file;
 347         }
 348
 349         file->async_file = filp->private_data;
 350
 351         INIT_IB_EVENT_HANDLER(&file->event_handler, file->device->ib_dev,
 352                               ib_uverbs_event_handler);
 353         ret = ib_register_event_handler(&file->event_handler);
 354         if (ret)
 355                 goto err_file;
 356
 357         kref_get(&file->async_file->ref);
 358         kref_get(&file->ref);
 359         file->ucontext = ucontext;
 360
 361         fd_install(resp.async_fd, filp);
 362
 363         mutex_unlock(&file->mutex);
 364
 365         return in_len;
 366
 367 err_file:
 368         fput(filp);
 369
 370 err_fd:
 371         put_unused_fd(resp.async_fd);
 372
 373 err_free:
 374         ibdev->dealloc_ucontext(ucontext);
 375
 376 err:
 377         mutex_unlock(&file->mutex);
 378         return ret;
 379 }
 380
 381 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 382                                const char __user *buf,
 383                                int in_len, int out_len)
 384 {
 385         struct ib_uverbs_query_device      cmd;
 386         struct ib_uverbs_query_device_resp resp;
 387         struct ib_device_attr              attr;
 388         int                                ret;
 389
 390         if (out_len < sizeof resp)
 391                 return -ENOSPC;
 392
 393         if (copy_from_user(&cmd, buf, sizeof cmd))
 394                 return -EFAULT;
 395
 396         ret = ib_query_device(file->device->ib_dev, &attr);
 397         if (ret)
 398                 return ret;
 399
 400         memset(&resp, 0, sizeof resp);
 401
 402         resp.fw_ver                    = attr.fw_ver;
 403         resp.node_guid                 = file->device->ib_dev->node_guid;
 404         resp.sys_image_guid            = attr.sys_image_guid;
 405         resp.max_mr_size               = attr.max_mr_size;
 406         resp.page_size_cap             = attr.page_size_cap;
 407         resp.vendor_id                 = attr.vendor_id;
 408         resp.vendor_part_id            = attr.vendor_part_id;
 409         resp.hw_ver                    = attr.hw_ver;
 410         resp.max_qp                    = attr.max_qp;
 411         resp.max_qp_wr                 = attr.max_qp_wr;
 412         resp.device_cap_flags          = attr.device_cap_flags;
 413         resp.max_sge                   = attr.max_sge;
 414         resp.max_sge_rd                = attr.max_sge_rd;
 415         resp.max_cq                    = attr.max_cq;
 416         resp.max_cqe                   = attr.max_cqe;
 417         resp.max_mr                    = attr.max_mr;
 418         resp.max_pd                    = attr.max_pd;
 419         resp.max_qp_rd_atom            = attr.max_qp_rd_atom;
 420         resp.max_ee_rd_atom            = attr.max_ee_rd_atom;
 421         resp.max_res_rd_atom           = attr.max_res_rd_atom;
 422         resp.max_qp_init_rd_atom       = attr.max_qp_init_rd_atom;
 423         resp.max_ee_init_rd_atom       = attr.max_ee_init_rd_atom;
 424         resp.atomic_cap                = attr.atomic_cap;
 425         resp.max_ee                    = attr.max_ee;
 426         resp.max_rdd                   = attr.max_rdd;
 427         resp.max_mw                    = attr.max_mw;
 428         resp.max_raw_ipv6_qp           = attr.max_raw_ipv6_qp;
 429         resp.max_raw_ethy_qp           = attr.max_raw_ethy_qp;
 430         resp.max_mcast_grp             = attr.max_mcast_grp;
 431         resp.max_mcast_qp_attach       = attr.max_mcast_qp_attach;
 432         resp.max_total_mcast_qp_attach = attr.max_total_mcast_qp_attach;
 433         resp.max_ah                    = attr.max_ah;
 434         resp.max_fmr                   = attr.max_fmr;
 435         resp.max_map_per_fmr           = attr.max_map_per_fmr;
 436         resp.max_srq                   = attr.max_srq;
 437         resp.max_srq_wr                = attr.max_srq_wr;
 438         resp.max_srq_sge               = attr.max_srq_sge;
 439         resp.max_pkeys                 = attr.max_pkeys;
 440         resp.local_ca_ack_delay        = attr.local_ca_ack_delay;
 441         resp.phys_port_cnt             = file->device->ib_dev->phys_port_cnt;
 442
 443         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 444                          &resp, sizeof resp))
 445                 return -EFAULT;
 446
 447         return in_len;
 448 }
 449
 450 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 451                              const char __user *buf,
 452                              int in_len, int out_len)
 453 {
 454         struct ib_uverbs_query_port      cmd;
 455         struct ib_uverbs_query_port_resp resp;
 456         struct ib_port_attr              attr;
 457         int                              ret;
 458
 459         if (out_len < sizeof resp)
 460                 return -ENOSPC;
 461
 462         if (copy_from_user(&cmd, buf, sizeof cmd))
 463                 return -EFAULT;
 464
 465         ret = ib_query_port(file->device->ib_dev, cmd.port_num, &attr);
 466         if (ret)
 467                 return ret;
 468
 469         memset(&resp, 0, sizeof resp);
 470
 471         resp.state           = attr.state;
 472         resp.max_mtu         = attr.max_mtu;
 473         resp.active_mtu      = attr.active_mtu;
 474         resp.gid_tbl_len     = attr.gid_tbl_len;
 475         resp.port_cap_flags  = attr.port_cap_flags;
 476         resp.max_msg_sz      = attr.max_msg_sz;
 477         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 478         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 479         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 480         resp.lid             = attr.lid;
 481         resp.sm_lid          = attr.sm_lid;
 482         resp.lmc             = attr.lmc;
 483         resp.max_vl_num      = attr.max_vl_num;
 484         resp.sm_sl           = attr.sm_sl;
 485         resp.subnet_timeout  = attr.subnet_timeout;
 486         resp.init_type_reply = attr.init_type_reply;
 487         resp.active_width    = attr.active_width;
 488         resp.active_speed    = attr.active_speed;
 489         resp.phys_state      = attr.phys_state;
 490         resp.link_layer      = rdma_port_get_link_layer(file->device->ib_dev,
 491                                                         cmd.port_num);
 492
 493         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 494                          &resp, sizeof resp))
 495                 return -EFAULT;
 496
 497         return in_len;
 498 }
 499
 500 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 501                            const char __user *buf,
 502                            int in_len, int out_len)
 503 {
 504         struct ib_uverbs_alloc_pd      cmd;
 505         struct ib_uverbs_alloc_pd_resp resp;
 506         struct ib_udata                udata;
 507         struct ib_uobject             *uobj;
 508         struct ib_pd                  *pd;
 509         int                            ret;
 510
 511         if (out_len < sizeof resp)
 512                 return -ENOSPC;
 513
 514         if (copy_from_user(&cmd, buf, sizeof cmd))
 515                 return -EFAULT;
 516
 517         INIT_UDATA(&udata, buf + sizeof cmd,
 518                    (unsigned long) cmd.response + sizeof resp,
 519                    in_len - sizeof cmd, out_len - sizeof resp);
 520
 521         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 522         if (!uobj)
 523                 return -ENOMEM;
 524
 525         init_uobj(uobj, 0, file->ucontext, &pd_lock_key);
 526         down_write(&uobj->mutex);
 527
 528         pd = file->device->ib_dev->alloc_pd(file->device->ib_dev,
 529                                             file->ucontext, &udata);
 530         if (IS_ERR(pd)) {
 531                 ret = PTR_ERR(pd);
 532                 goto err;
 533         }
 534
 535         pd->device  = file->device->ib_dev;
 536         pd->uobject = uobj;
 537         atomic_set(&pd->usecnt, 0);
 538
 539         uobj->object = pd;
 540         ret = idr_add_uobj(&ib_uverbs_pd_idr, uobj);
 541         if (ret)
 542                 goto err_idr;
 543
 544         memset(&resp, 0, sizeof resp);
 545         resp.pd_handle = uobj->id;
 546
 547         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 548                          &resp, sizeof resp)) {
 549                 ret = -EFAULT;
 550                 goto err_copy;
 551         }
 552
 553         mutex_lock(&file->mutex);
 554         list_add_tail(&uobj->list, &file->ucontext->pd_list);
 555         mutex_unlock(&file->mutex);
 556
 557         uobj->live = 1;
 558
 559         up_write(&uobj->mutex);
 560
 561         return in_len;
 562
 563 err_copy:
 564         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 565
 566 err_idr:
 567         ib_dealloc_pd(pd);
 568
 569 err:
 570         put_uobj_write(uobj);
 571         return ret;
 572 }
 573
 574 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 575                              const char __user *buf,
 576                              int in_len, int out_len)
 577 {
 578         struct ib_uverbs_dealloc_pd cmd;
 579         struct ib_uobject          *uobj;
 580         int                         ret;
 581
 582         if (copy_from_user(&cmd, buf, sizeof cmd))
 583                 return -EFAULT;
 584
 585         uobj = idr_write_uobj(&ib_uverbs_pd_idr, cmd.pd_handle, file->ucontext);
 586         if (!uobj)
 587                 return -EINVAL;
 588
 589         ret = ib_dealloc_pd(uobj->object);
 590         if (!ret)
 591                 uobj->live = 0;
 592
 593         put_uobj_write(uobj);
 594
 595         if (ret)
 596                 return ret;
 597
 598         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 599
 600         mutex_lock(&file->mutex);
 601         list_del(&uobj->list);
 602         mutex_unlock(&file->mutex);
 603
 604         put_uobj(uobj);
 605
 606         return in_len;
 607 }
 608
 609 struct xrcd_table_entry {
 610         struct rb_node  node;
 611         struct ib_xrcd *xrcd;
 612         struct inode   *inode;
 613 };
 614
 615 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 616                             struct inode *inode,
 617                             struct ib_xrcd *xrcd)
 618 {
 619         struct xrcd_table_entry *entry, *scan;
 620         struct rb_node **p = &dev->xrcd_tree.rb_node;
 621         struct rb_node *parent = NULL;
 622
 623         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 624         if (!entry)
 625                 return -ENOMEM;
 626
 627         entry->xrcd  = xrcd;
 628         entry->inode = inode;
 629
 630         while (*p) {
 631                 parent = *p;
 632                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 633
 634                 if (inode < scan->inode) {
 635                         p = &(*p)->rb_left;
 636                 } else if (inode > scan->inode) {
 637                         p = &(*p)->rb_right;
 638                 } else {
 639                         kfree(entry);
 640                         return -EEXIST;
 641                 }
 642         }
 643
 644         rb_link_node(&entry->node, parent, p);
 645         rb_insert_color(&entry->node, &dev->xrcd_tree);
 646         igrab(inode);
 647         return 0;
 648 }
 649
 650 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 651                                                   struct inode *inode)
 652 {
 653         struct xrcd_table_entry *entry;
 654         struct rb_node *p = dev->xrcd_tree.rb_node;
 655
 656         while (p) {
 657                 entry = rb_entry(p, struct xrcd_table_entry, node);
 658
 659                 if (inode < entry->inode)
 660                         p = p->rb_left;
 661                 else if (inode > entry->inode)
 662                         p = p->rb_right;
 663                 else
 664                         return entry;
 665         }
 666
 667         return NULL;
 668 }
 669
 670 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 671 {
 672         struct xrcd_table_entry *entry;
 673
 674         entry = xrcd_table_search(dev, inode);
 675         if (!entry)
 676                 return NULL;
 677
 678         return entry->xrcd;
 679 }
 680
 681 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 682                               struct inode *inode)
 683 {
 684         struct xrcd_table_entry *entry;
 685
 686         entry = xrcd_table_search(dev, inode);
 687         if (entry) {
 688                 iput(inode);
 689                 rb_erase(&entry->node, &dev->xrcd_tree);
 690                 kfree(entry);
 691         }
 692 }
 693
 694 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 695                             const char __user *buf, int in_len,
 696                             int out_len)
 697 {
 698         struct ib_uverbs_open_xrcd      cmd;
 699         struct ib_uverbs_open_xrcd_resp resp;
 700         struct ib_udata                 udata;
 701         struct ib_uxrcd_object         *obj;
 702         struct ib_xrcd                 *xrcd = NULL;
 703         struct file                    *f = NULL;
 704         struct inode                   *inode = NULL;
 705         int                             ret = 0;
 706         int                             new_xrcd = 0;
 707
 708         if (out_len < sizeof resp)
 709                 return -ENOSPC;
 710
 711         if (copy_from_user(&cmd, buf, sizeof cmd))
 712                 return -EFAULT;
 713
 714         INIT_UDATA(&udata, buf + sizeof cmd,
 715                    (unsigned long) cmd.response + sizeof resp,
 716                    in_len - sizeof cmd, out_len - sizeof  resp);
 717
 718         mutex_lock(&file->device->xrcd_tree_mutex);
 719
 720         if (cmd.fd != -1) {
 721                 /* search for file descriptor */
 722                 f = fget(cmd.fd);
 723                 if (!f) {
 724                         ret = -EBADF;
 725                         goto err_tree_mutex_unlock;
 726                 }
 727
 728                 inode = f->f_dentry->d_inode;
 729                 if (!inode) {
 730                         ret = -EBADF;
 731                         goto err_tree_mutex_unlock;
 732                 }
 733
 734                 xrcd = find_xrcd(file->device, inode);
 735                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 736                         /* no file descriptor. Need CREATE flag */
 737                         ret = -EAGAIN;
 738                         goto err_tree_mutex_unlock;
 739                 }
 740
 741                 if (xrcd && cmd.oflags & O_EXCL) {
 742                         ret = -EINVAL;
 743                         goto err_tree_mutex_unlock;
 744                 }
 745         }
 746
 747         obj = kmalloc(sizeof *obj, GFP_KERNEL);
 748         if (!obj) {
 749                 ret = -ENOMEM;
 750                 goto err_tree_mutex_unlock;
 751         }
 752
 753         init_uobj(&obj->uobject, 0, file->ucontext, &xrcd_lock_key);
 754
 755         down_write(&obj->uobject.mutex);
 756
 757         if (!xrcd) {
 758                 xrcd = file->device->ib_dev->alloc_xrcd(file->device->ib_dev,
 759                                                         file->ucontext, &udata);
 760                 if (IS_ERR(xrcd)) {
 761                         ret = PTR_ERR(xrcd);
 762                         goto err;
 763                 }
 764
 765                 xrcd->inode   = inode;
 766                 xrcd->device  = file->device->ib_dev;
 767                 atomic_set(&xrcd->usecnt, 0);
 768                 mutex_init(&xrcd->tgt_qp_mutex);
 769                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 770                 new_xrcd = 1;
 771         }
 772
 773         atomic_set(&obj->refcnt, 0);
 774         obj->uobject.object = xrcd;
 775         ret = idr_add_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 776         if (ret)
 777                 goto err_idr;
 778
 779         memset(&resp, 0, sizeof resp);
 780         resp.xrcd_handle = obj->uobject.id;
 781
 782         if (inode) {
 783                 if (new_xrcd) {
 784                         /* create new inode/xrcd table entry */
 785                         ret = xrcd_table_insert(file->device, inode, xrcd);
 786                         if (ret)
 787                                 goto err_insert_xrcd;
 788                 }
 789                 atomic_inc(&xrcd->usecnt);
 790         }
 791
 792         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 793                          &resp, sizeof resp)) {
 794                 ret = -EFAULT;
 795                 goto err_copy;
 796         }
 797
 798         if (f)
 799                 fput(f);
 800
 801         mutex_lock(&file->mutex);
 802         list_add_tail(&obj->uobject.list, &file->ucontext->xrcd_list);
 803         mutex_unlock(&file->mutex);
 804
 805         obj->uobject.live = 1;
 806         up_write(&obj->uobject.mutex);
 807
 808         mutex_unlock(&file->device->xrcd_tree_mutex);
 809         return in_len;
 810
 811 err_copy:
 812         if (inode) {
 813                 if (new_xrcd)
 814                         xrcd_table_delete(file->device, inode);
 815                 atomic_dec(&xrcd->usecnt);
 816         }
 817
 818 err_insert_xrcd:
 819         idr_remove_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 820
 821 err_idr:
 822         ib_dealloc_xrcd(xrcd);
 823
 824 err:
 825         put_uobj_write(&obj->uobject);
 826
 827 err_tree_mutex_unlock:
 828         if (f)
 829                 fput(f);
 830
 831         mutex_unlock(&file->device->xrcd_tree_mutex);
 832
 833         return ret;
 834 }
 835
 836 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 837                              const char __user *buf, int in_len,
 838                              int out_len)
 839 {
 840         struct ib_uverbs_close_xrcd cmd;
 841         struct ib_uobject           *uobj;
 842         struct ib_xrcd              *xrcd = NULL;
 843         struct inode                *inode = NULL;
 844         struct ib_uxrcd_object      *obj;
 845         int                         live;
 846         int                         ret = 0;
 847
 848         if (copy_from_user(&cmd, buf, sizeof cmd))
 849                 return -EFAULT;
 850
 851         mutex_lock(&file->device->xrcd_tree_mutex);
 852         uobj = idr_write_uobj(&ib_uverbs_xrcd_idr, cmd.xrcd_handle, file->ucontext);
 853         if (!uobj) {
 854                 ret = -EINVAL;
 855                 goto out;
 856         }
 857
 858         xrcd  = uobj->object;
 859         inode = xrcd->inode;
 860         obj   = container_of(uobj, struct ib_uxrcd_object, uobject);
 861         if (atomic_read(&obj->refcnt)) {
 862                 put_uobj_write(uobj);
 863                 ret = -EBUSY;
 864                 goto out;
 865         }
 866
 867         if (!inode || atomic_dec_and_test(&xrcd->usecnt)) {
 868                 ret = ib_dealloc_xrcd(uobj->object);
 869                 if (!ret)
 870                         uobj->live = 0;
 871         }
 872
 873         live = uobj->live;
 874         if (inode && ret)
 875                 atomic_inc(&xrcd->usecnt);
 876
 877         put_uobj_write(uobj);
 878
 879         if (ret)
 880                 goto out;
 881
 882         if (inode && !live)
 883                 xrcd_table_delete(file->device, inode);
 884
 885         idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 886         mutex_lock(&file->mutex);
 887         list_del(&uobj->list);
 888         mutex_unlock(&file->mutex);
 889
 890         put_uobj(uobj);
 891         ret = in_len;
 892
 893 out:
 894         mutex_unlock(&file->device->xrcd_tree_mutex);
 895         return ret;
 896 }
 897
 898 void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 899                             struct ib_xrcd *xrcd)
 900 {
 901         struct inode *inode;
 902
 903         inode = xrcd->inode;
 904         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 905                 return;
 906
 907         ib_dealloc_xrcd(xrcd);
 908
 909         if (inode)
 910                 xrcd_table_delete(dev, inode);
 911 }
 912
 913 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 914                          const char __user *buf, int in_len,
 915                          int out_len)
 916 {
 917         struct ib_uverbs_reg_mr      cmd;
 918         struct ib_uverbs_reg_mr_resp resp;
 919         struct ib_udata              udata;
 920         struct ib_uobject           *uobj;
 921         struct ib_pd                *pd;
 922         struct ib_mr                *mr;
 923         int                          ret;
 924
 925         if (out_len < sizeof resp)
 926                 return -ENOSPC;
 927
 928         if (copy_from_user(&cmd, buf, sizeof cmd))
 929                 return -EFAULT;
 930
 931         INIT_UDATA(&udata, buf + sizeof cmd,
 932                    (unsigned long) cmd.response + sizeof resp,
 933                    in_len - sizeof cmd, out_len - sizeof resp);
 934
 935         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 936                 return -EINVAL;
 937
 938         /*
 939          * Local write permission is required if remote write or
 940          * remote atomic permission is also requested.
 941          */
 942         if (cmd.access_flags & (IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_REMOTE_WRITE) &&
 943             !(cmd.access_flags & IB_ACCESS_LOCAL_WRITE))
 944                 return -EINVAL;
 945
 946         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 947         if (!uobj)
 948                 return -ENOMEM;
 949
 950         init_uobj(uobj, 0, file->ucontext, &mr_lock_key);
 951         down_write(&uobj->mutex);
 952
 953         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
 954         if (!pd) {
 955                 ret = -EINVAL;
 956                 goto err_free;
 957         }
 958
 959         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 960                                      cmd.access_flags, &udata);
 961         if (IS_ERR(mr)) {
 962                 ret = PTR_ERR(mr);
 963                 goto err_put;
 964         }
 965
 966         mr->device  = pd->device;
 967         mr->pd      = pd;
 968         mr->uobject = uobj;
 969         atomic_inc(&pd->usecnt);
 970         atomic_set(&mr->usecnt, 0);
 971
 972         uobj->object = mr;
 973         ret = idr_add_uobj(&ib_uverbs_mr_idr, uobj);
 974         if (ret)
 975                 goto err_unreg;
 976
 977         memset(&resp, 0, sizeof resp);
 978         resp.lkey      = mr->lkey;
 979         resp.rkey      = mr->rkey;
 980         resp.mr_handle = uobj->id;
 981
 982         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 983                          &resp, sizeof resp)) {
 984                 ret = -EFAULT;
 985                 goto err_copy;
 986         }
 987
 988         put_pd_read(pd);
 989
 990         mutex_lock(&file->mutex);
 991         list_add_tail(&uobj->list, &file->ucontext->mr_list);
 992         mutex_unlock(&file->mutex);
 993
 994         uobj->live = 1;
 995
 996         up_write(&uobj->mutex);
 997
 998         return in_len;
 999
1000 err_copy:
1001         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1002
1003 err_unreg:
1004         ib_dereg_mr(mr);
1005
1006 err_put:
1007         put_pd_read(pd);
1008
1009 err_free:
1010         put_uobj_write(uobj);
1011         return ret;
1012 }
1013
1014 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
1015                            const char __user *buf, int in_len,
1016                            int out_len)
1017 {
1018         struct ib_uverbs_dereg_mr cmd;
1019         struct ib_mr             *mr;
1020         struct ib_uobject        *uobj;
1021         int                       ret = -EINVAL;
1022
1023         if (copy_from_user(&cmd, buf, sizeof cmd))
1024                 return -EFAULT;
1025
1026         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle, file->ucontext);
1027         if (!uobj)
1028                 return -EINVAL;
1029
1030         mr = uobj->object;
1031
1032         ret = ib_dereg_mr(mr);
1033         if (!ret)
1034                 uobj->live = 0;
1035
1036         put_uobj_write(uobj);
1037
1038         if (ret)
1039                 return ret;
1040
1041         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1042
1043         mutex_lock(&file->mutex);
1044         list_del(&uobj->list);
1045         mutex_unlock(&file->mutex);
1046
1047         put_uobj(uobj);
1048
1049         return in_len;
1050 }
1051
1052 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
1053                                       const char __user *buf, int in_len,
1054                                       int out_len)
1055 {
1056         struct ib_uverbs_create_comp_channel       cmd;
1057         struct ib_uverbs_create_comp_channel_resp  resp;
1058         struct file                               *filp;
1059         int ret;
1060
1061         if (out_len < sizeof resp)
1062                 return -ENOSPC;
1063
1064         if (copy_from_user(&cmd, buf, sizeof cmd))
1065                 return -EFAULT;
1066
1067         ret = get_unused_fd();
1068         if (ret < 0)
1069                 return ret;
1070         resp.fd = ret;
1071
1072         filp = ib_uverbs_alloc_event_file(file, 0);
1073         if (IS_ERR(filp)) {
1074                 put_unused_fd(resp.fd);
1075                 return PTR_ERR(filp);
1076         }
1077
1078         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1079                          &resp, sizeof resp)) {
1080                 put_unused_fd(resp.fd);
1081                 fput(filp);
1082                 return -EFAULT;
1083         }
1084
1085         fd_install(resp.fd, filp);
1086         return in_len;
1087 }
1088
1089 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1090                             const char __user *buf, int in_len,
1091                             int out_len)
1092 {
1093         struct ib_uverbs_create_cq      cmd;
1094         struct ib_uverbs_create_cq_resp resp;
1095         struct ib_udata                 udata;
1096         struct ib_ucq_object           *obj;
1097         struct ib_uverbs_event_file    *ev_file = NULL;
1098         struct ib_cq                   *cq;
1099         int                             ret;
1100
1101         if (out_len < sizeof resp)
1102                 return -ENOSPC;
1103
1104         if (copy_from_user(&cmd, buf, sizeof cmd))
1105                 return -EFAULT;
1106
1107         INIT_UDATA(&udata, buf + sizeof cmd,
1108                    (unsigned long) cmd.response + sizeof resp,
1109                    in_len - sizeof cmd, out_len - sizeof resp);
1110
1111         if (cmd.comp_vector >= file->device->num_comp_vectors)
1112                 return -EINVAL;
1113
1114         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1115         if (!obj)
1116                 return -ENOMEM;
1117
1118         init_uobj(&obj->uobject, cmd.user_handle, file->ucontext, &cq_lock_key);
1119         down_write(&obj->uobject.mutex);
1120
1121         if (cmd.comp_channel >= 0) {
1122                 ev_file = ib_uverbs_lookup_comp_file(cmd.comp_channel);
1123                 if (!ev_file) {
1124                         ret = -EINVAL;
1125                         goto err;
1126                 }
1127         }
1128
1129         obj->uverbs_file           = file;
1130         obj->comp_events_reported  = 0;
1131         obj->async_events_reported = 0;
1132         INIT_LIST_HEAD(&obj->comp_list);
1133         INIT_LIST_HEAD(&obj->async_list);
1134
1135         cq = file->device->ib_dev->create_cq(file->device->ib_dev, cmd.cqe,
1136                                              cmd.comp_vector,
1137                                              file->ucontext, &udata);
1138         if (IS_ERR(cq)) {
1139                 ret = PTR_ERR(cq);
1140                 goto err_file;
1141         }
1142
1143         cq->device        = file->device->ib_dev;
1144         cq->uobject       = &obj->uobject;
1145         cq->comp_handler  = ib_uverbs_comp_handler;
1146         cq->event_handler = ib_uverbs_cq_event_handler;
1147         cq->cq_context    = ev_file;
1148         atomic_set(&cq->usecnt, 0);
1149
1150         obj->uobject.object = cq;
1151         ret = idr_add_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1152         if (ret)
1153                 goto err_free;
1154
1155         memset(&resp, 0, sizeof resp);
1156         resp.cq_handle = obj->uobject.id;
1157         resp.cqe       = cq->cqe;
1158
1159         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1160                          &resp, sizeof resp)) {
1161                 ret = -EFAULT;
1162                 goto err_copy;
1163         }
1164
1165         mutex_lock(&file->mutex);
1166         list_add_tail(&obj->uobject.list, &file->ucontext->cq_list);
1167         mutex_unlock(&file->mutex);
1168
1169         obj->uobject.live = 1;
1170
1171         up_write(&obj->uobject.mutex);
1172
1173         return in_len;
1174
1175 err_copy:
1176         idr_remove_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1177
1178 err_free:
1179         ib_destroy_cq(cq);
1180
1181 err_file:
1182         if (ev_file)
1183                 ib_uverbs_release_ucq(file, ev_file, obj);
1184
1185 err:
1186         put_uobj_write(&obj->uobject);
1187         return ret;
1188 }
1189
1190 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1191                             const char __user *buf, int in_len,
1192                             int out_len)
1193 {
1194         struct ib_uverbs_resize_cq      cmd;
1195         struct ib_uverbs_resize_cq_resp resp;
1196         struct ib_udata                 udata;
1197         struct ib_cq                    *cq;
1198         int                             ret = -EINVAL;
1199
1200         if (copy_from_user(&cmd, buf, sizeof cmd))
1201                 return -EFAULT;
1202
1203         INIT_UDATA(&udata, buf + sizeof cmd,
1204                    (unsigned long) cmd.response + sizeof resp,
1205                    in_len - sizeof cmd, out_len - sizeof resp);
1206
1207         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1208         if (!cq)
1209                 return -EINVAL;
1210
1211         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1212         if (ret)
1213                 goto out;
1214
1215         resp.cqe = cq->cqe;
1216
1217         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1218                          &resp, sizeof resp.cqe))
1219                 ret = -EFAULT;
1220
1221 out:
1222         put_cq_read(cq);
1223
1224         return ret ? ret : in_len;
1225 }
1226
1227 static int copy_wc_to_user(void __user *dest, struct ib_wc *wc)
1228 {
1229         struct ib_uverbs_wc tmp;
1230
1231         tmp.wr_id               = wc->wr_id;
1232         tmp.status              = wc->status;
1233         tmp.opcode              = wc->opcode;
1234         tmp.vendor_err          = wc->vendor_err;
1235         tmp.byte_len            = wc->byte_len;
1236         tmp.ex.imm_data         = (__u32 __force) wc->ex.imm_data;
1237         tmp.qp_num              = wc->qp->qp_num;
1238         tmp.src_qp              = wc->src_qp;
1239         tmp.wc_flags            = wc->wc_flags;
1240         tmp.pkey_index          = wc->pkey_index;
1241         tmp.slid                = wc->slid;
1242         tmp.sl                  = wc->sl;
1243         tmp.dlid_path_bits      = wc->dlid_path_bits;
1244         tmp.port_num            = wc->port_num;
1245         tmp.reserved            = 0;
1246
1247         if (copy_to_user(dest, &tmp, sizeof tmp))
1248                 return -EFAULT;
1249
1250         return 0;
1251 }
1252
1253 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1254                           const char __user *buf, int in_len,
1255                           int out_len)
1256 {
1257         struct ib_uverbs_poll_cq       cmd;
1258         struct ib_uverbs_poll_cq_resp  resp;
1259         u8 __user                     *header_ptr;
1260         u8 __user                     *data_ptr;
1261         struct ib_cq                  *cq;
1262         struct ib_wc                   wc;
1263         int                            ret;
1264
1265         if (copy_from_user(&cmd, buf, sizeof cmd))
1266                 return -EFAULT;
1267
1268         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1269         if (!cq)
1270                 return -EINVAL;
1271
1272         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1273         header_ptr = (void __user *)(unsigned long) cmd.response;
1274         data_ptr = header_ptr + sizeof resp;
1275
1276         memset(&resp, 0, sizeof resp);
1277         while (resp.count < cmd.ne) {
1278                 ret = ib_poll_cq(cq, 1, &wc);
1279                 if (ret < 0)
1280                         goto out_put;
1281                 if (!ret)
1282                         break;
1283
1284                 ret = copy_wc_to_user(data_ptr, &wc);
1285                 if (ret)
1286                         goto out_put;
1287
1288                 data_ptr += sizeof(struct ib_uverbs_wc);
1289                 ++resp.count;
1290         }
1291
1292         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1293                 ret = -EFAULT;
1294                 goto out_put;
1295         }
1296
1297         ret = in_len;
1298
1299 out_put:
1300         put_cq_read(cq);
1301         return ret;
1302 }
1303
1304 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1305                                 const char __user *buf, int in_len,
1306                                 int out_len)
1307 {
1308         struct ib_uverbs_req_notify_cq cmd;
1309         struct ib_cq                  *cq;
1310
1311         if (copy_from_user(&cmd, buf, sizeof cmd))
1312                 return -EFAULT;
1313
1314         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1315         if (!cq)
1316                 return -EINVAL;
1317
1318         ib_req_notify_cq(cq, cmd.solicited_only ?
1319                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1320
1321         put_cq_read(cq);
1322
1323         return in_len;
1324 }
1325
1326 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1327                              const char __user *buf, int in_len,
1328                              int out_len)
1329 {
1330         struct ib_uverbs_destroy_cq      cmd;
1331         struct ib_uverbs_destroy_cq_resp resp;
1332         struct ib_uobject               *uobj;
1333         struct ib_cq                    *cq;
1334         struct ib_ucq_object            *obj;
1335         struct ib_uverbs_event_file     *ev_file;
1336         int                              ret = -EINVAL;
1337
1338         if (copy_from_user(&cmd, buf, sizeof cmd))
1339                 return -EFAULT;
1340
1341         uobj = idr_write_uobj(&ib_uverbs_cq_idr, cmd.cq_handle, file->ucontext);
1342         if (!uobj)
1343                 return -EINVAL;
1344         cq      = uobj->object;
1345         ev_file = cq->cq_context;
1346         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1347
1348         ret = ib_destroy_cq(cq);
1349         if (!ret)
1350                 uobj->live = 0;
1351
1352         put_uobj_write(uobj);
1353
1354         if (ret)
1355                 return ret;
1356
1357         idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
1358
1359         mutex_lock(&file->mutex);
1360         list_del(&uobj->list);
1361         mutex_unlock(&file->mutex);
1362
1363         ib_uverbs_release_ucq(file, ev_file, obj);
1364
1365         memset(&resp, 0, sizeof resp);
1366         resp.comp_events_reported  = obj->comp_events_reported;
1367         resp.async_events_reported = obj->async_events_reported;
1368
1369         put_uobj(uobj);
1370
1371         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1372                          &resp, sizeof resp))
1373                 return -EFAULT;
1374
1375         return in_len;
1376 }
1377
1378 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1379                             const char __user *buf, int in_len,
1380                             int out_len)
1381 {
1382         struct ib_uverbs_create_qp      cmd;
1383         struct ib_uverbs_create_qp_resp resp;
1384         struct ib_udata                 udata;
1385         struct ib_uqp_object           *obj;
1386         struct ib_device               *device;
1387         struct ib_pd                   *pd = NULL;
1388         struct ib_xrcd                 *xrcd = NULL;
1389         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1390         struct ib_cq                   *scq = NULL, *rcq = NULL;
1391         struct ib_srq                  *srq = NULL;
1392         struct ib_qp                   *qp;
1393         struct ib_qp_init_attr          attr;
1394         int ret;
1395
1396         if (out_len < sizeof resp)
1397                 return -ENOSPC;
1398
1399         if (copy_from_user(&cmd, buf, sizeof cmd))
1400                 return -EFAULT;
1401
1402         INIT_UDATA(&udata, buf + sizeof cmd,
1403                    (unsigned long) cmd.response + sizeof resp,
1404                    in_len - sizeof cmd, out_len - sizeof resp);
1405
1406         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1407         if (!obj)
1408                 return -ENOMEM;
1409
1410         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_key);
1411         down_write(&obj->uevent.uobject.mutex);
1412
1413         if (cmd.qp_type == IB_QPT_XRC_TGT) {
1414                 xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1415                 if (!xrcd) {
1416                         ret = -EINVAL;
1417                         goto err_put;
1418                 }
1419                 device = xrcd->device;
1420         } else {
1421                 pd  = idr_read_pd(cmd.pd_handle, file->ucontext);
1422                 scq = idr_read_cq(cmd.send_cq_handle, file->ucontext, 0);
1423                 if (!pd || !scq) {
1424                         ret = -EINVAL;
1425                         goto err_put;
1426                 }
1427
1428                 if (cmd.qp_type == IB_QPT_XRC_INI) {
1429                         cmd.max_recv_wr = cmd.max_recv_sge = 0;
1430                 } else {
1431                         if (cmd.is_srq) {
1432                                 srq = idr_read_srq(cmd.srq_handle, file->ucontext);
1433                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1434                                         ret = -EINVAL;
1435                                         goto err_put;
1436                                 }
1437                         }
1438                         rcq = (cmd.recv_cq_handle == cmd.send_cq_handle) ?
1439                                scq : idr_read_cq(cmd.recv_cq_handle, file->ucontext, 1);
1440                         if (!rcq) {
1441                                 ret = -EINVAL;
1442                                 goto err_put;
1443                         }
1444                 }
1445                 device = pd->device;
1446         }
1447
1448         attr.event_handler = ib_uverbs_qp_event_handler;
1449         attr.qp_context    = file;
1450         attr.send_cq       = scq;
1451         attr.recv_cq       = rcq;
1452         attr.srq           = srq;
1453         attr.xrcd          = xrcd;
1454         attr.sq_sig_type   = cmd.sq_sig_all ? IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
1455         attr.qp_type       = cmd.qp_type;
1456         attr.create_flags  = 0;
1457
1458         attr.cap.max_send_wr     = cmd.max_send_wr;
1459         attr.cap.max_recv_wr     = cmd.max_recv_wr;
1460         attr.cap.max_send_sge    = cmd.max_send_sge;
1461         attr.cap.max_recv_sge    = cmd.max_recv_sge;
1462         attr.cap.max_inline_data = cmd.max_inline_data;
1463
1464         obj->uevent.events_reported     = 0;
1465         INIT_LIST_HEAD(&obj->uevent.event_list);
1466         INIT_LIST_HEAD(&obj->mcast_list);
1467
1468         if (cmd.qp_type == IB_QPT_XRC_TGT)
1469                 qp = ib_create_qp(pd, &attr);
1470         else
1471                 qp = device->create_qp(pd, &attr, &udata);
1472
1473         if (IS_ERR(qp)) {
1474                 ret = PTR_ERR(qp);
1475                 goto err_put;
1476         }
1477
1478         if (cmd.qp_type != IB_QPT_XRC_TGT) {
1479                 qp->real_qp       = qp;
1480                 qp->device        = device;
1481                 qp->pd            = pd;
1482                 qp->send_cq       = attr.send_cq;
1483                 qp->recv_cq       = attr.recv_cq;
1484                 qp->srq           = attr.srq;
1485                 qp->event_handler = attr.event_handler;
1486                 qp->qp_context    = attr.qp_context;
1487                 qp->qp_type       = attr.qp_type;
1488                 atomic_set(&qp->usecnt, 0);
1489                 atomic_inc(&pd->usecnt);
1490                 atomic_inc(&attr.send_cq->usecnt);
1491                 if (attr.recv_cq)
1492                         atomic_inc(&attr.recv_cq->usecnt);
1493                 if (attr.srq)
1494                         atomic_inc(&attr.srq->usecnt);
1495         }
1496         qp->uobject = &obj->uevent.uobject;
1497
1498         obj->uevent.uobject.object = qp;
1499         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1500         if (ret)
1501                 goto err_destroy;
1502
1503         memset(&resp, 0, sizeof resp);
1504         resp.qpn             = qp->qp_num;
1505         resp.qp_handle       = obj->uevent.uobject.id;
1506         resp.max_recv_sge    = attr.cap.max_recv_sge;
1507         resp.max_send_sge    = attr.cap.max_send_sge;
1508         resp.max_recv_wr     = attr.cap.max_recv_wr;
1509         resp.max_send_wr     = attr.cap.max_send_wr;
1510         resp.max_inline_data = attr.cap.max_inline_data;
1511
1512         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1513                          &resp, sizeof resp)) {
1514                 ret = -EFAULT;
1515                 goto err_copy;
1516         }
1517
1518         if (xrcd)
1519                 put_xrcd_read(xrcd_uobj);
1520         if (pd)
1521                 put_pd_read(pd);
1522         if (scq)
1523                 put_cq_read(scq);
1524         if (rcq && rcq != scq)
1525                 put_cq_read(rcq);
1526         if (srq)
1527                 put_srq_read(srq);
1528
1529         mutex_lock(&file->mutex);
1530         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1531         mutex_unlock(&file->mutex);
1532
1533         obj->uevent.uobject.live = 1;
1534
1535         up_write(&obj->uevent.uobject.mutex);
1536
1537         return in_len;
1538
1539 err_copy:
1540         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1541
1542 err_destroy:
1543         ib_destroy_qp(qp);
1544
1545 err_put:
1546         if (xrcd)
1547                 put_xrcd_read(xrcd_uobj);
1548         if (pd)
1549                 put_pd_read(pd);
1550         if (scq)
1551                 put_cq_read(scq);
1552         if (rcq && rcq != scq)
1553                 put_cq_read(rcq);
1554         if (srq)
1555                 put_srq_read(srq);
1556
1557         put_uobj_write(&obj->uevent.uobject);
1558         return ret;
1559 }
1560
1561 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1562                           const char __user *buf, int in_len, int out_len)
1563 {
1564         struct ib_uverbs_open_qp        cmd;
1565         struct ib_uverbs_create_qp_resp resp;
1566         struct ib_udata                 udata;
1567         struct ib_uqp_object           *obj;
1568         struct ib_xrcd                 *xrcd;
1569         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1570         struct ib_qp                   *qp;
1571         struct ib_qp_open_attr          attr;
1572         int ret;
1573
1574         if (out_len < sizeof resp)
1575                 return -ENOSPC;
1576
1577         if (copy_from_user(&cmd, buf, sizeof cmd))
1578                 return -EFAULT;
1579
1580         INIT_UDATA(&udata, buf + sizeof cmd,
1581                    (unsigned long) cmd.response + sizeof resp,
1582                    in_len - sizeof cmd, out_len - sizeof resp);
1583
1584         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1585         if (!obj)
1586                 return -ENOMEM;
1587
1588         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_key);
1589         down_write(&obj->uevent.uobject.mutex);
1590
1591         xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1592         if (!xrcd) {
1593                 ret = -EINVAL;
1594                 goto err_put;
1595         }
1596
1597         attr.event_handler = ib_uverbs_qp_event_handler;
1598         attr.qp_context    = file;
1599         attr.qp_num        = cmd.qpn;
1600         attr.qp_type       = cmd.qp_type;
1601
1602         obj->uevent.events_reported = 0;
1603         INIT_LIST_HEAD(&obj->uevent.event_list);
1604         INIT_LIST_HEAD(&obj->mcast_list);
1605
1606         qp = ib_open_qp(xrcd, &attr);
1607         if (IS_ERR(qp)) {
1608                 ret = PTR_ERR(qp);
1609                 goto err_put;
1610         }
1611
1612         qp->uobject = &obj->uevent.uobject;
1613
1614         obj->uevent.uobject.object = qp;
1615         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1616         if (ret)
1617                 goto err_destroy;
1618
1619         memset(&resp, 0, sizeof resp);
1620         resp.qpn       = qp->qp_num;
1621         resp.qp_handle = obj->uevent.uobject.id;
1622
1623         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1624                          &resp, sizeof resp)) {
1625                 ret = -EFAULT;
1626                 goto err_remove;
1627         }
1628
1629         put_xrcd_read(xrcd_uobj);
1630
1631         mutex_lock(&file->mutex);
1632         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1633         mutex_unlock(&file->mutex);
1634
1635         obj->uevent.uobject.live = 1;
1636
1637         up_write(&obj->uevent.uobject.mutex);
1638
1639         return in_len;
1640
1641 err_remove:
1642         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1643
1644 err_destroy:
1645         ib_destroy_qp(qp);
1646
1647 err_put:
1648         put_xrcd_read(xrcd_uobj);
1649         put_uobj_write(&obj->uevent.uobject);
1650         return ret;
1651 }
1652
1653 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1654                            const char __user *buf, int in_len,
1655                            int out_len)
1656 {
1657         struct ib_uverbs_query_qp      cmd;
1658         struct ib_uverbs_query_qp_resp resp;
1659         struct ib_qp                   *qp;
1660         struct ib_qp_attr              *attr;
1661         struct ib_qp_init_attr         *init_attr;
1662         int                            ret;
1663
1664         if (copy_from_user(&cmd, buf, sizeof cmd))
1665                 return -EFAULT;
1666
1667         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1668         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1669         if (!attr || !init_attr) {
1670                 ret = -ENOMEM;
1671                 goto out;
1672         }
1673
1674         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1675         if (!qp) {
1676                 ret = -EINVAL;
1677                 goto out;
1678         }
1679
1680         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1681
1682         put_qp_read(qp);
1683
1684         if (ret)
1685                 goto out;
1686
1687         memset(&resp, 0, sizeof resp);
1688
1689         resp.qp_state               = attr->qp_state;
1690         resp.cur_qp_state           = attr->cur_qp_state;
1691         resp.path_mtu               = attr->path_mtu;
1692         resp.path_mig_state         = attr->path_mig_state;
1693         resp.qkey                   = attr->qkey;
1694         resp.rq_psn                 = attr->rq_psn;
1695         resp.sq_psn                 = attr->sq_psn;
1696         resp.dest_qp_num            = attr->dest_qp_num;
1697         resp.qp_access_flags        = attr->qp_access_flags;
1698         resp.pkey_index             = attr->pkey_index;
1699         resp.alt_pkey_index         = attr->alt_pkey_index;
1700         resp.sq_draining            = attr->sq_draining;
1701         resp.max_rd_atomic          = attr->max_rd_atomic;
1702         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1703         resp.min_rnr_timer          = attr->min_rnr_timer;
1704         resp.port_num               = attr->port_num;
1705         resp.timeout                = attr->timeout;
1706         resp.retry_cnt              = attr->retry_cnt;
1707         resp.rnr_retry              = attr->rnr_retry;
1708         resp.alt_port_num           = attr->alt_port_num;
1709         resp.alt_timeout            = attr->alt_timeout;
1710
1711         memcpy(resp.dest.dgid, attr->ah_attr.grh.dgid.raw, 16);
1712         resp.dest.flow_label        = attr->ah_attr.grh.flow_label;
1713         resp.dest.sgid_index        = attr->ah_attr.grh.sgid_index;
1714         resp.dest.hop_limit         = attr->ah_attr.grh.hop_limit;
1715         resp.dest.traffic_class     = attr->ah_attr.grh.traffic_class;
1716         resp.dest.dlid              = attr->ah_attr.dlid;
1717         resp.dest.sl                = attr->ah_attr.sl;
1718         resp.dest.src_path_bits     = attr->ah_attr.src_path_bits;
1719         resp.dest.static_rate       = attr->ah_attr.static_rate;
1720         resp.dest.is_global         = !!(attr->ah_attr.ah_flags & IB_AH_GRH);
1721         resp.dest.port_num          = attr->ah_attr.port_num;
1722
1723         memcpy(resp.alt_dest.dgid, attr->alt_ah_attr.grh.dgid.raw, 16);
1724         resp.alt_dest.flow_label    = attr->alt_ah_attr.grh.flow_label;
1725         resp.alt_dest.sgid_index    = attr->alt_ah_attr.grh.sgid_index;
1726         resp.alt_dest.hop_limit     = attr->alt_ah_attr.grh.hop_limit;
1727         resp.alt_dest.traffic_class = attr->alt_ah_attr.grh.traffic_class;
1728         resp.alt_dest.dlid          = attr->alt_ah_attr.dlid;
1729         resp.alt_dest.sl            = attr->alt_ah_attr.sl;
1730         resp.alt_dest.src_path_bits = attr->alt_ah_attr.src_path_bits;
1731         resp.alt_dest.static_rate   = attr->alt_ah_attr.static_rate;
1732         resp.alt_dest.is_global     = !!(attr->alt_ah_attr.ah_flags & IB_AH_GRH);
1733         resp.alt_dest.port_num      = attr->alt_ah_attr.port_num;
1734
1735         resp.max_send_wr            = init_attr->cap.max_send_wr;
1736         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1737         resp.max_send_sge           = init_attr->cap.max_send_sge;
1738         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1739         resp.max_inline_data        = init_attr->cap.max_inline_data;
1740         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1741
1742         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1743                          &resp, sizeof resp))
1744                 ret = -EFAULT;
1745
1746 out:
1747         kfree(attr);
1748         kfree(init_attr);
1749
1750         return ret ? ret : in_len;
1751 }
1752
1753 /* Remove ignored fields set in the attribute mask */
1754 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1755 {
1756         switch (qp_type) {
1757         case IB_QPT_XRC_INI:
1758                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1759         case IB_QPT_XRC_TGT:
1760                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1761                                 IB_QP_RNR_RETRY);
1762         default:
1763                 return mask;
1764         }
1765 }
1766
1767 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
1768                             const char __user *buf, int in_len,
1769                             int out_len)
1770 {
1771         struct ib_uverbs_modify_qp cmd;
1772         struct ib_udata            udata;
1773         struct ib_qp              *qp;
1774         struct ib_qp_attr         *attr;
1775         int                        ret;
1776
1777         if (copy_from_user(&cmd, buf, sizeof cmd))
1778                 return -EFAULT;
1779
1780         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
1781                    out_len);
1782
1783         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1784         if (!attr)
1785                 return -ENOMEM;
1786
1787         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1788         if (!qp) {
1789                 ret = -EINVAL;
1790                 goto out;
1791         }
1792
1793         if (!rdma_is_port_valid(qp->device, cmd.port_num)) {
1794                 ret = -EINVAL;
1795                 goto release_qp;
1796         }
1797
1798         attr->qp_state            = cmd.qp_state;
1799         attr->cur_qp_state        = cmd.cur_qp_state;
1800         attr->path_mtu            = cmd.path_mtu;
1801         attr->path_mig_state      = cmd.path_mig_state;
1802         attr->qkey                = cmd.qkey;
1803         attr->rq_psn              = cmd.rq_psn;
1804         attr->sq_psn              = cmd.sq_psn;
1805         attr->dest_qp_num         = cmd.dest_qp_num;
1806         attr->qp_access_flags     = cmd.qp_access_flags;
1807         attr->pkey_index          = cmd.pkey_index;
1808         attr->alt_pkey_index      = cmd.alt_pkey_index;
1809         attr->en_sqd_async_notify = cmd.en_sqd_async_notify;
1810         attr->max_rd_atomic       = cmd.max_rd_atomic;
1811         attr->max_dest_rd_atomic  = cmd.max_dest_rd_atomic;
1812         attr->min_rnr_timer       = cmd.min_rnr_timer;
1813         attr->port_num            = cmd.port_num;
1814         attr->timeout             = cmd.timeout;
1815         attr->retry_cnt           = cmd.retry_cnt;
1816         attr->rnr_retry           = cmd.rnr_retry;
1817         attr->alt_port_num        = cmd.alt_port_num;
1818         attr->alt_timeout         = cmd.alt_timeout;
1819
1820         memcpy(attr->ah_attr.grh.dgid.raw, cmd.dest.dgid, 16);
1821         attr->ah_attr.grh.flow_label        = cmd.dest.flow_label;
1822         attr->ah_attr.grh.sgid_index        = cmd.dest.sgid_index;
1823         attr->ah_attr.grh.hop_limit         = cmd.dest.hop_limit;
1824         attr->ah_attr.grh.traffic_class     = cmd.dest.traffic_class;
1825         attr->ah_attr.dlid                  = cmd.dest.dlid;
1826         attr->ah_attr.sl                    = cmd.dest.sl;
1827         attr->ah_attr.src_path_bits         = cmd.dest.src_path_bits;
1828         attr->ah_attr.static_rate           = cmd.dest.static_rate;
1829         attr->ah_attr.ah_flags              = cmd.dest.is_global ? IB_AH_GRH : 0;
1830         attr->ah_attr.port_num              = cmd.dest.port_num;
1831
1832         memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd.alt_dest.dgid, 16);
1833         attr->alt_ah_attr.grh.flow_label    = cmd.alt_dest.flow_label;
1834         attr->alt_ah_attr.grh.sgid_index    = cmd.alt_dest.sgid_index;
1835         attr->alt_ah_attr.grh.hop_limit     = cmd.alt_dest.hop_limit;
1836         attr->alt_ah_attr.grh.traffic_class = cmd.alt_dest.traffic_class;
1837         attr->alt_ah_attr.dlid              = cmd.alt_dest.dlid;
1838         attr->alt_ah_attr.sl                = cmd.alt_dest.sl;
1839         attr->alt_ah_attr.src_path_bits     = cmd.alt_dest.src_path_bits;
1840         attr->alt_ah_attr.static_rate       = cmd.alt_dest.static_rate;
1841         attr->alt_ah_attr.ah_flags          = cmd.alt_dest.is_global ? IB_AH_GRH : 0;
1842         attr->alt_ah_attr.port_num          = cmd.alt_dest.port_num;
1843
1844         if (qp->real_qp == qp) {
1845                 ret = qp->device->modify_qp(qp, attr,
1846                         modify_qp_mask(qp->qp_type, cmd.attr_mask), &udata);
1847         } else {
1848                 ret = ib_modify_qp(qp, attr, modify_qp_mask(qp->qp_type, cmd.attr_mask));
1849         }
1850
1851 release_qp:
1852         put_qp_read(qp);
1853
1854         if (ret)
1855                 goto out;
1856
1857         ret = in_len;
1858
1859 out:
1860         kfree(attr);
1861
1862         return ret;
1863 }
1864
1865 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
1866                              const char __user *buf, int in_len,
1867                              int out_len)
1868 {
1869         struct ib_uverbs_destroy_qp      cmd;
1870         struct ib_uverbs_destroy_qp_resp resp;
1871         struct ib_uobject               *uobj;
1872         struct ib_qp                    *qp;
1873         struct ib_uqp_object            *obj;
1874         int                              ret = -EINVAL;
1875
1876         if (copy_from_user(&cmd, buf, sizeof cmd))
1877                 return -EFAULT;
1878
1879         memset(&resp, 0, sizeof resp);
1880
1881         uobj = idr_write_uobj(&ib_uverbs_qp_idr, cmd.qp_handle, file->ucontext);
1882         if (!uobj)
1883                 return -EINVAL;
1884         qp  = uobj->object;
1885         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
1886
1887         if (!list_empty(&obj->mcast_list)) {
1888                 put_uobj_write(uobj);
1889                 return -EBUSY;
1890         }
1891
1892         ret = ib_destroy_qp(qp);
1893         if (!ret)
1894                 uobj->live = 0;
1895
1896         put_uobj_write(uobj);
1897
1898         if (ret)
1899                 return ret;
1900
1901         idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
1902
1903         mutex_lock(&file->mutex);
1904         list_del(&uobj->list);
1905         mutex_unlock(&file->mutex);
1906
1907         ib_uverbs_release_uevent(file, &obj->uevent);
1908
1909         resp.events_reported = obj->uevent.events_reported;
1910
1911         put_uobj(uobj);
1912
1913         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1914                          &resp, sizeof resp))
1915                 return -EFAULT;
1916
1917         return in_len;
1918 }
1919
1920 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
1921                             const char __user *buf, int in_len,
1922                             int out_len)
1923 {
1924         struct ib_uverbs_post_send      cmd;
1925         struct ib_uverbs_post_send_resp resp;
1926         struct ib_uverbs_send_wr       *user_wr;
1927         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
1928         struct ib_qp                   *qp;
1929         int                             i, sg_ind;
1930         int                             is_ud;
1931         ssize_t                         ret = -EINVAL;
1932
1933         if (copy_from_user(&cmd, buf, sizeof cmd))
1934                 return -EFAULT;
1935
1936         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
1937             cmd.sge_count * sizeof (struct ib_uverbs_sge))
1938                 return -EINVAL;
1939
1940         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
1941                 return -EINVAL;
1942
1943         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
1944         if (!user_wr)
1945                 return -ENOMEM;
1946
1947         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1948         if (!qp)
1949                 goto out;
1950
1951         is_ud = qp->qp_type == IB_QPT_UD;
1952         sg_ind = 0;
1953         last = NULL;
1954         for (i = 0; i < cmd.wr_count; ++i) {
1955                 if (copy_from_user(user_wr,
1956                                    buf + sizeof cmd + i * cmd.wqe_size,
1957                                    cmd.wqe_size)) {
1958                         ret = -EFAULT;
1959                         goto out_put;
1960                 }
1961
1962                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
1963                         ret = -EINVAL;
1964                         goto out_put;
1965                 }
1966
1967                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
1968                                user_wr->num_sge * sizeof (struct ib_sge),
1969                                GFP_KERNEL);
1970                 if (!next) {
1971                         ret = -ENOMEM;
1972                         goto out_put;
1973                 }
1974
1975                 if (!last)
1976                         wr = next;
1977                 else
1978                         last->next = next;
1979                 last = next;
1980
1981                 next->next       = NULL;
1982                 next->wr_id      = user_wr->wr_id;
1983                 next->num_sge    = user_wr->num_sge;
1984                 next->opcode     = user_wr->opcode;
1985                 next->send_flags = user_wr->send_flags;
1986
1987                 if (is_ud) {
1988                         if (next->opcode != IB_WR_SEND &&
1989                             next->opcode != IB_WR_SEND_WITH_IMM) {
1990                                 ret = -EINVAL;
1991                                 goto out_put;
1992                         }
1993
1994                         next->wr.ud.ah = idr_read_ah(user_wr->wr.ud.ah,
1995                                                      file->ucontext);
1996                         if (!next->wr.ud.ah) {
1997                                 ret = -EINVAL;
1998                                 goto out_put;
1999                         }
2000                         next->wr.ud.remote_qpn  = user_wr->wr.ud.remote_qpn;
2001                         next->wr.ud.remote_qkey = user_wr->wr.ud.remote_qkey;
2002                 } else {
2003                         switch (next->opcode) {
2004                         case IB_WR_RDMA_WRITE_WITH_IMM:
2005                                 next->ex.imm_data =
2006                                         (__be32 __force) user_wr->ex.imm_data;
2007                         case IB_WR_RDMA_WRITE:
2008                         case IB_WR_RDMA_READ:
2009                                 next->wr.rdma.remote_addr =
2010                                         user_wr->wr.rdma.remote_addr;
2011                                 next->wr.rdma.rkey        =
2012                                         user_wr->wr.rdma.rkey;
2013                                 break;
2014                         case IB_WR_SEND_WITH_IMM:
2015                                 next->ex.imm_data =
2016                                         (__be32 __force) user_wr->ex.imm_data;
2017                                 break;
2018                         case IB_WR_SEND_WITH_INV:
2019                                 next->ex.invalidate_rkey =
2020                                         user_wr->ex.invalidate_rkey;
2021                                 break;
2022                         case IB_WR_ATOMIC_CMP_AND_SWP:
2023                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2024                                 next->wr.atomic.remote_addr =
2025                                         user_wr->wr.atomic.remote_addr;
2026                                 next->wr.atomic.compare_add =
2027                                         user_wr->wr.atomic.compare_add;
2028                                 next->wr.atomic.swap = user_wr->wr.atomic.swap;
2029                                 next->wr.atomic.rkey = user_wr->wr.atomic.rkey;
2030                         case IB_WR_SEND:
2031                                 break;
2032                         default:
2033                                 ret = -EINVAL;
2034                                 goto out_put;
2035                         }
2036                 }
2037
2038                 if (next->num_sge) {
2039                         next->sg_list = (void *) next +
2040                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2041                         if (copy_from_user(next->sg_list,
2042                                            buf + sizeof cmd +
2043                                            cmd.wr_count * cmd.wqe_size +
2044                                            sg_ind * sizeof (struct ib_sge),
2045                                            next->num_sge * sizeof (struct ib_sge))) {
2046                                 ret = -EFAULT;
2047                                 goto out_put;
2048                         }
2049                         sg_ind += next->num_sge;
2050                 } else
2051                         next->sg_list = NULL;
2052         }
2053
2054         resp.bad_wr = 0;
2055         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2056         if (ret)
2057                 for (next = wr; next; next = next->next) {
2058                         ++resp.bad_wr;
2059                         if (next == bad_wr)
2060                                 break;
2061                 }
2062
2063         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2064                          &resp, sizeof resp))
2065                 ret = -EFAULT;
2066
2067 out_put:
2068         put_qp_read(qp);
2069
2070         while (wr) {
2071                 if (is_ud && wr->wr.ud.ah)
2072                         put_ah_read(wr->wr.ud.ah);
2073                 next = wr->next;
2074                 kfree(wr);
2075                 wr = next;
2076         }
2077
2078 out:
2079         kfree(user_wr);
2080
2081         return ret ? ret : in_len;
2082 }
2083
2084 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2085                                                     int in_len,
2086                                                     u32 wr_count,
2087                                                     u32 sge_count,
2088                                                     u32 wqe_size)
2089 {
2090         struct ib_uverbs_recv_wr *user_wr;
2091         struct ib_recv_wr        *wr = NULL, *last, *next;
2092         int                       sg_ind;
2093         int                       i;
2094         int                       ret;
2095
2096         if (in_len < wqe_size * wr_count +
2097             sge_count * sizeof (struct ib_uverbs_sge))
2098                 return ERR_PTR(-EINVAL);
2099
2100         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2101                 return ERR_PTR(-EINVAL);
2102
2103         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2104         if (!user_wr)
2105                 return ERR_PTR(-ENOMEM);
2106
2107         sg_ind = 0;
2108         last = NULL;
2109         for (i = 0; i < wr_count; ++i) {
2110                 if (copy_from_user(user_wr, buf + i * wqe_size,
2111                                    wqe_size)) {
2112                         ret = -EFAULT;
2113                         goto err;
2114                 }
2115
2116                 if (user_wr->num_sge + sg_ind > sge_count) {
2117                         ret = -EINVAL;
2118                         goto err;
2119                 }
2120
2121                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2122                                user_wr->num_sge * sizeof (struct ib_sge),
2123                                GFP_KERNEL);
2124                 if (!next) {
2125                         ret = -ENOMEM;
2126                         goto err;
2127                 }
2128
2129                 if (!last)
2130                         wr = next;
2131                 else
2132                         last->next = next;
2133                 last = next;
2134
2135                 next->next       = NULL;
2136                 next->wr_id      = user_wr->wr_id;
2137                 next->num_sge    = user_wr->num_sge;
2138
2139                 if (next->num_sge) {
2140                         next->sg_list = (void *) next +
2141                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2142                         if (copy_from_user(next->sg_list,
2143                                            buf + wr_count * wqe_size +
2144                                            sg_ind * sizeof (struct ib_sge),
2145                                            next->num_sge * sizeof (struct ib_sge))) {
2146                                 ret = -EFAULT;
2147                                 goto err;
2148                         }
2149                         sg_ind += next->num_sge;
2150                 } else
2151                         next->sg_list = NULL;
2152         }
2153
2154         kfree(user_wr);
2155         return wr;
2156
2157 err:
2158         kfree(user_wr);
2159
2160         while (wr) {
2161                 next = wr->next;
2162                 kfree(wr);
2163                 wr = next;
2164         }
2165
2166         return ERR_PTR(ret);
2167 }
2168
2169 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2170                             const char __user *buf, int in_len,
2171                             int out_len)
2172 {
2173         struct ib_uverbs_post_recv      cmd;
2174         struct ib_uverbs_post_recv_resp resp;
2175         struct ib_recv_wr              *wr, *next, *bad_wr;
2176         struct ib_qp                   *qp;
2177         ssize_t                         ret = -EINVAL;
2178
2179         if (copy_from_user(&cmd, buf, sizeof cmd))
2180                 return -EFAULT;
2181
2182         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2183                                        in_len - sizeof cmd, cmd.wr_count,
2184                                        cmd.sge_count, cmd.wqe_size);
2185         if (IS_ERR(wr))
2186                 return PTR_ERR(wr);
2187
2188         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2189         if (!qp)
2190                 goto out;
2191
2192         resp.bad_wr = 0;
2193         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2194
2195         put_qp_read(qp);
2196
2197         if (ret)
2198                 for (next = wr; next; next = next->next) {
2199                         ++resp.bad_wr;
2200                         if (next == bad_wr)
2201                                 break;
2202                 }
2203
2204         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2205                          &resp, sizeof resp))
2206                 ret = -EFAULT;
2207
2208 out:
2209         while (wr) {
2210                 next = wr->next;
2211                 kfree(wr);
2212                 wr = next;
2213         }
2214
2215         return ret ? ret : in_len;
2216 }
2217
2218 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2219                                 const char __user *buf, int in_len,
2220                                 int out_len)
2221 {
2222         struct ib_uverbs_post_srq_recv      cmd;
2223         struct ib_uverbs_post_srq_recv_resp resp;
2224         struct ib_recv_wr                  *wr, *next, *bad_wr;
2225         struct ib_srq                      *srq;
2226         ssize_t                             ret = -EINVAL;
2227
2228         if (copy_from_user(&cmd, buf, sizeof cmd))
2229                 return -EFAULT;
2230
2231         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2232                                        in_len - sizeof cmd, cmd.wr_count,
2233                                        cmd.sge_count, cmd.wqe_size);
2234         if (IS_ERR(wr))
2235                 return PTR_ERR(wr);
2236
2237         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2238         if (!srq)
2239                 goto out;
2240
2241         resp.bad_wr = 0;
2242         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2243
2244         put_srq_read(srq);
2245
2246         if (ret)
2247                 for (next = wr; next; next = next->next) {
2248                         ++resp.bad_wr;
2249                         if (next == bad_wr)
2250                                 break;
2251                 }
2252
2253         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2254                          &resp, sizeof resp))
2255                 ret = -EFAULT;
2256
2257 out:
2258         while (wr) {
2259                 next = wr->next;
2260                 kfree(wr);
2261                 wr = next;
2262         }
2263
2264         return ret ? ret : in_len;
2265 }
2266
2267 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2268                             const char __user *buf, int in_len,
2269                             int out_len)
2270 {
2271         struct ib_device                *ib_dev = file->device->ib_dev;
2272         struct ib_uverbs_create_ah       cmd;
2273         struct ib_uverbs_create_ah_resp  resp;
2274         struct ib_uobject               *uobj;
2275         struct ib_pd                    *pd;
2276         struct ib_ah                    *ah;
2277         struct ib_ah_attr               attr;
2278         int ret;
2279
2280         if (out_len < sizeof resp)
2281                 return -ENOSPC;
2282
2283         if (copy_from_user(&cmd, buf, sizeof cmd))
2284                 return -EFAULT;
2285
2286         if (!rdma_is_port_valid(ib_dev, cmd.attr.port_num))
2287                 return -EINVAL;
2288
2289         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
2290         if (!uobj)
2291                 return -ENOMEM;
2292
2293         init_uobj(uobj, cmd.user_handle, file->ucontext, &ah_lock_key);
2294         down_write(&uobj->mutex);
2295
2296         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
2297         if (!pd) {
2298                 ret = -EINVAL;
2299                 goto err;
2300         }
2301
2302         attr.dlid              = cmd.attr.dlid;
2303         attr.sl                = cmd.attr.sl;
2304         attr.src_path_bits     = cmd.attr.src_path_bits;
2305         attr.static_rate       = cmd.attr.static_rate;
2306         attr.ah_flags          = cmd.attr.is_global ? IB_AH_GRH : 0;
2307         attr.port_num          = cmd.attr.port_num;
2308         attr.grh.flow_label    = cmd.attr.grh.flow_label;
2309         attr.grh.sgid_index    = cmd.attr.grh.sgid_index;
2310         attr.grh.hop_limit     = cmd.attr.grh.hop_limit;
2311         attr.grh.traffic_class = cmd.attr.grh.traffic_class;
2312         memcpy(attr.grh.dgid.raw, cmd.attr.grh.dgid, 16);
2313
2314         ah = ib_create_ah(pd, &attr);
2315         if (IS_ERR(ah)) {
2316                 ret = PTR_ERR(ah);
2317                 goto err_put;
2318         }
2319
2320         ah->uobject  = uobj;
2321         uobj->object = ah;
2322
2323         ret = idr_add_uobj(&ib_uverbs_ah_idr, uobj);
2324         if (ret)
2325                 goto err_destroy;
2326
2327         resp.ah_handle = uobj->id;
2328
2329         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2330                          &resp, sizeof resp)) {
2331                 ret = -EFAULT;
2332                 goto err_copy;
2333         }
2334
2335         put_pd_read(pd);
2336
2337         mutex_lock(&file->mutex);
2338         list_add_tail(&uobj->list, &file->ucontext->ah_list);
2339         mutex_unlock(&file->mutex);
2340
2341         uobj->live = 1;
2342
2343         up_write(&uobj->mutex);
2344
2345         return in_len;
2346
2347 err_copy:
2348         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2349
2350 err_destroy:
2351         ib_destroy_ah(ah);
2352
2353 err_put:
2354         put_pd_read(pd);
2355
2356 err:
2357         put_uobj_write(uobj);
2358         return ret;
2359 }
2360
2361 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2362                              const char __user *buf, int in_len, int out_len)
2363 {
2364         struct ib_uverbs_destroy_ah cmd;
2365         struct ib_ah               *ah;
2366         struct ib_uobject          *uobj;
2367         int                         ret;
2368
2369         if (copy_from_user(&cmd, buf, sizeof cmd))
2370                 return -EFAULT;
2371
2372         uobj = idr_write_uobj(&ib_uverbs_ah_idr, cmd.ah_handle, file->ucontext);
2373         if (!uobj)
2374                 return -EINVAL;
2375         ah = uobj->object;
2376
2377         ret = ib_destroy_ah(ah);
2378         if (!ret)
2379                 uobj->live = 0;
2380
2381         put_uobj_write(uobj);
2382
2383         if (ret)
2384                 return ret;
2385
2386         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2387
2388         mutex_lock(&file->mutex);
2389         list_del(&uobj->list);
2390         mutex_unlock(&file->mutex);
2391
2392         put_uobj(uobj);
2393
2394         return in_len;
2395 }
2396
2397 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2398                                const char __user *buf, int in_len,
2399                                int out_len)
2400 {
2401         struct ib_uverbs_attach_mcast cmd;
2402         struct ib_qp                 *qp;
2403         struct ib_uqp_object         *obj;
2404         struct ib_uverbs_mcast_entry *mcast;
2405         int                           ret;
2406
2407         if (copy_from_user(&cmd, buf, sizeof cmd))
2408                 return -EFAULT;
2409
2410         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2411         if (!qp)
2412                 return -EINVAL;
2413
2414         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2415
2416         list_for_each_entry(mcast, &obj->mcast_list, list)
2417                 if (cmd.mlid == mcast->lid &&
2418                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2419                         ret = 0;
2420                         goto out_put;
2421                 }
2422
2423         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2424         if (!mcast) {
2425                 ret = -ENOMEM;
2426                 goto out_put;
2427         }
2428
2429         mcast->lid = cmd.mlid;
2430         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2431
2432         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2433         if (!ret)
2434                 list_add_tail(&mcast->list, &obj->mcast_list);
2435         else
2436                 kfree(mcast);
2437
2438 out_put:
2439         put_qp_write(qp);
2440
2441         return ret ? ret : in_len;
2442 }
2443
2444 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2445                                const char __user *buf, int in_len,
2446                                int out_len)
2447 {
2448         struct ib_uverbs_detach_mcast cmd;
2449         struct ib_uqp_object         *obj;
2450         struct ib_qp                 *qp;
2451         struct ib_uverbs_mcast_entry *mcast;
2452         int                           ret = -EINVAL;
2453         bool                          found = false;
2454
2455         if (copy_from_user(&cmd, buf, sizeof cmd))
2456                 return -EFAULT;
2457
2458         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2459         if (!qp)
2460                 return -EINVAL;
2461
2462         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2463
2464         list_for_each_entry(mcast, &obj->mcast_list, list)
2465                 if (cmd.mlid == mcast->lid &&
2466                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2467                         list_del(&mcast->list);
2468                         kfree(mcast);
2469                         found = true;
2470                         break;
2471                 }
2472
2473         if (!found) {
2474                 ret = -EINVAL;
2475                 goto out_put;
2476         }
2477
2478         ret = ib_detach_mcast(qp, (union ib_gid *)cmd.gid, cmd.mlid);
2479
2480 out_put:
2481         put_qp_write(qp);
2482
2483         return ret ? ret : in_len;
2484 }
2485
2486 int __uverbs_create_xsrq(struct ib_uverbs_file *file,
2487                          struct ib_uverbs_create_xsrq *cmd,
2488                          struct ib_udata *udata)
2489 {
2490         struct ib_uverbs_create_srq_resp resp;
2491         struct ib_usrq_object           *obj;
2492         struct ib_pd                    *pd;
2493         struct ib_srq                   *srq;
2494         struct ib_uobject               *uninitialized_var(xrcd_uobj);
2495         struct ib_srq_init_attr          attr;
2496         int ret;
2497
2498         obj = kmalloc(sizeof *obj, GFP_KERNEL);
2499         if (!obj)
2500                 return -ENOMEM;
2501
2502         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext, &srq_lock_key);
2503         down_write(&obj->uevent.uobject.mutex);
2504
2505         pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
2506         if (!pd) {
2507                 ret = -EINVAL;
2508                 goto err;
2509         }
2510
2511         if (cmd->srq_type == IB_SRQT_XRC) {
2512                 attr.ext.xrc.cq  = idr_read_cq(cmd->cq_handle, file->ucontext, 0);
2513                 if (!attr.ext.xrc.cq) {
2514                         ret = -EINVAL;
2515                         goto err_put_pd;
2516                 }
2517
2518                 attr.ext.xrc.xrcd  = idr_read_xrcd(cmd->xrcd_handle, file->ucontext, &xrcd_uobj);
2519                 if (!attr.ext.xrc.xrcd) {
2520                         ret = -EINVAL;
2521                         goto err_put_cq;
2522                 }
2523
2524                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
2525                 atomic_inc(&obj->uxrcd->refcnt);
2526         }
2527
2528         attr.event_handler  = ib_uverbs_srq_event_handler;
2529         attr.srq_context    = file;
2530         attr.srq_type       = cmd->srq_type;
2531         attr.attr.max_wr    = cmd->max_wr;
2532         attr.attr.max_sge   = cmd->max_sge;
2533         attr.attr.srq_limit = cmd->srq_limit;
2534
2535         obj->uevent.events_reported = 0;
2536         INIT_LIST_HEAD(&obj->uevent.event_list);
2537
2538         srq = pd->device->create_srq(pd, &attr, udata);
2539         if (IS_ERR(srq)) {
2540                 ret = PTR_ERR(srq);
2541                 goto err_put;
2542         }
2543
2544         srq->device        = pd->device;
2545         srq->pd            = pd;
2546         srq->srq_type      = cmd->srq_type;
2547         srq->uobject       = &obj->uevent.uobject;
2548         srq->event_handler = attr.event_handler;
2549         srq->srq_context   = attr.srq_context;
2550
2551         if (cmd->srq_type == IB_SRQT_XRC) {
2552                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
2553                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
2554                 atomic_inc(&attr.ext.xrc.cq->usecnt);
2555                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
2556         }
2557
2558         atomic_inc(&pd->usecnt);
2559         atomic_set(&srq->usecnt, 0);
2560
2561         obj->uevent.uobject.object = srq;
2562         ret = idr_add_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2563         if (ret)
2564                 goto err_destroy;
2565
2566         memset(&resp, 0, sizeof resp);
2567         resp.srq_handle = obj->uevent.uobject.id;
2568         resp.max_wr     = attr.attr.max_wr;
2569         resp.max_sge    = attr.attr.max_sge;
2570         if (cmd->srq_type == IB_SRQT_XRC)
2571                 resp.srqn = srq->ext.xrc.srq_num;
2572
2573         if (copy_to_user((void __user *) (unsigned long) cmd->response,
2574                          &resp, sizeof resp)) {
2575                 ret = -EFAULT;
2576                 goto err_copy;
2577         }
2578
2579         if (cmd->srq_type == IB_SRQT_XRC) {
2580                 put_uobj_read(xrcd_uobj);
2581                 put_cq_read(attr.ext.xrc.cq);
2582         }
2583         put_pd_read(pd);
2584
2585         mutex_lock(&file->mutex);
2586         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->srq_list);
2587         mutex_unlock(&file->mutex);
2588
2589         obj->uevent.uobject.live = 1;
2590
2591         up_write(&obj->uevent.uobject.mutex);
2592
2593         return 0;
2594
2595 err_copy:
2596         idr_remove_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2597
2598 err_destroy:
2599         ib_destroy_srq(srq);
2600
2601 err_put:
2602         if (cmd->srq_type == IB_SRQT_XRC) {
2603                 atomic_dec(&obj->uxrcd->refcnt);
2604                 put_uobj_read(xrcd_uobj);
2605         }
2606
2607 err_put_cq:
2608         if (cmd->srq_type == IB_SRQT_XRC)
2609                 put_cq_read(attr.ext.xrc.cq);
2610
2611 err_put_pd:
2612         put_pd_read(pd);
2613
2614 err:
2615         put_uobj_write(&obj->uevent.uobject);
2616         return ret;
2617 }
2618
2619 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
2620                              const char __user *buf, int in_len,
2621                              int out_len)
2622 {
2623         struct ib_uverbs_create_srq      cmd;
2624         struct ib_uverbs_create_xsrq     xcmd;
2625         struct ib_uverbs_create_srq_resp resp;
2626         struct ib_udata                  udata;
2627         int ret;
2628
2629         if (out_len < sizeof resp)
2630                 return -ENOSPC;
2631
2632         if (copy_from_user(&cmd, buf, sizeof cmd))
2633                 return -EFAULT;
2634
2635         xcmd.response    = cmd.response;
2636         xcmd.user_handle = cmd.user_handle;
2637         xcmd.srq_type    = IB_SRQT_BASIC;
2638         xcmd.pd_handle   = cmd.pd_handle;
2639         xcmd.max_wr      = cmd.max_wr;
2640         xcmd.max_sge     = cmd.max_sge;
2641         xcmd.srq_limit   = cmd.srq_limit;
2642
2643         INIT_UDATA(&udata, buf + sizeof cmd,
2644                    (unsigned long) cmd.response + sizeof resp,
2645                    in_len - sizeof cmd, out_len - sizeof resp);
2646
2647         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
2648         if (ret)
2649                 return ret;
2650
2651         return in_len;
2652 }
2653
2654 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
2655                               const char __user *buf, int in_len, int out_len)
2656 {
2657         struct ib_uverbs_create_xsrq     cmd;
2658         struct ib_uverbs_create_srq_resp resp;
2659         struct ib_udata                  udata;
2660         int ret;
2661
2662         if (out_len < sizeof resp)
2663                 return -ENOSPC;
2664
2665         if (copy_from_user(&cmd, buf, sizeof cmd))
2666                 return -EFAULT;
2667
2668         INIT_UDATA(&udata, buf + sizeof cmd,
2669                    (unsigned long) cmd.response + sizeof resp,
2670                    in_len - sizeof cmd, out_len - sizeof resp);
2671
2672         ret = __uverbs_create_xsrq(file, &cmd, &udata);
2673         if (ret)
2674                 return ret;
2675
2676         return in_len;
2677 }
2678
2679 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
2680                              const char __user *buf, int in_len,
2681                              int out_len)
2682 {
2683         struct ib_uverbs_modify_srq cmd;
2684         struct ib_udata             udata;
2685         struct ib_srq              *srq;
2686         struct ib_srq_attr          attr;
2687         int                         ret;
2688
2689         if (copy_from_user(&cmd, buf, sizeof cmd))
2690                 return -EFAULT;
2691
2692         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
2693                    out_len);
2694
2695         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2696         if (!srq)
2697                 return -EINVAL;
2698
2699         attr.max_wr    = cmd.max_wr;
2700         attr.srq_limit = cmd.srq_limit;
2701
2702         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
2703
2704         put_srq_read(srq);
2705
2706         return ret ? ret : in_len;
2707 }
2708
2709 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
2710                             const char __user *buf,
2711                             int in_len, int out_len)
2712 {
2713         struct ib_uverbs_query_srq      cmd;
2714         struct ib_uverbs_query_srq_resp resp;
2715         struct ib_srq_attr              attr;
2716         struct ib_srq                   *srq;
2717         int                             ret;
2718
2719         if (out_len < sizeof resp)
2720                 return -ENOSPC;
2721
2722         if (copy_from_user(&cmd, buf, sizeof cmd))
2723                 return -EFAULT;
2724
2725         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2726         if (!srq)
2727                 return -EINVAL;
2728
2729         ret = ib_query_srq(srq, &attr);
2730
2731         put_srq_read(srq);
2732
2733         if (ret)
2734                 return ret;
2735
2736         memset(&resp, 0, sizeof resp);
2737
2738         resp.max_wr    = attr.max_wr;
2739         resp.max_sge   = attr.max_sge;
2740         resp.srq_limit = attr.srq_limit;
2741
2742         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2743                          &resp, sizeof resp))
2744                 return -EFAULT;
2745
2746         return in_len;
2747 }
2748
2749 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
2750                               const char __user *buf, int in_len,
2751                               int out_len)
2752 {
2753         struct ib_uverbs_destroy_srq      cmd;
2754         struct ib_uverbs_destroy_srq_resp resp;
2755         struct ib_uobject                *uobj;
2756         struct ib_srq                    *srq;
2757         struct ib_uevent_object          *obj;
2758         int                               ret = -EINVAL;
2759
2760         if (copy_from_user(&cmd, buf, sizeof cmd))
2761                 return -EFAULT;
2762
2763         uobj = idr_write_uobj(&ib_uverbs_srq_idr, cmd.srq_handle, file->ucontext);
2764         if (!uobj)
2765                 return -EINVAL;
2766         srq = uobj->object;
2767         obj = container_of(uobj, struct ib_uevent_object, uobject);
2768
2769         ret = ib_destroy_srq(srq);
2770         if (!ret)
2771                 uobj->live = 0;
2772
2773         put_uobj_write(uobj);
2774
2775         if (ret)
2776                 return ret;
2777
2778         idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
2779
2780         mutex_lock(&file->mutex);
2781         list_del(&uobj->list);
2782         mutex_unlock(&file->mutex);
2783
2784         ib_uverbs_release_uevent(file, obj);
2785
2786         memset(&resp, 0, sizeof resp);
2787         resp.events_reported = obj->events_reported;
2788
2789         put_uobj(uobj);
2790
2791         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2792                          &resp, sizeof resp))
2793                 ret = -EFAULT;
2794
2795         return ret ? ret : in_len;
2796 }