drivers/infiniband/core/uverbs_cmd.c

   1 /*
   2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005, 2006, 2007 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 PathScale, Inc.  All rights reserved.
   5  * Copyright (c) 2006 Mellanox Technologies.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/file.h>
  37 #include <linux/fs.h>
  38 #include <linux/slab.h>
  39
  40 #include <asm/uaccess.h>
  41
  42 #include "uverbs.h"
  43
  44 struct uverbs_lock_class {
  45         struct lock_class_key   key;
  46         char                    name[16];
  47 };
  48
  49 static struct uverbs_lock_class pd_lock_class   = { .name = "PD-uobj" };
  50 static struct uverbs_lock_class mr_lock_class   = { .name = "MR-uobj" };
  51 static struct uverbs_lock_class mw_lock_class   = { .name = "MW-uobj" };
  52 static struct uverbs_lock_class cq_lock_class   = { .name = "CQ-uobj" };
  53 static struct uverbs_lock_class qp_lock_class   = { .name = "QP-uobj" };
  54 static struct uverbs_lock_class ah_lock_class   = { .name = "AH-uobj" };
  55 static struct uverbs_lock_class srq_lock_class  = { .name = "SRQ-uobj" };
  56 static struct uverbs_lock_class xrcd_lock_class = { .name = "XRCD-uobj" };
  57 static struct uverbs_lock_class rule_lock_class = { .name = "RULE-uobj" };
  58
  59 /*
  60  * The ib_uobject locking scheme is as follows:
  61  *
  62  * - ib_uverbs_idr_lock protects the uverbs idrs themselves, so it
  63  *   needs to be held during all idr operations.  When an object is
  64  *   looked up, a reference must be taken on the object's kref before
  65  *   dropping this lock.
  66  *
  67  * - Each object also has an rwsem.  This rwsem must be held for
  68  *   reading while an operation that uses the object is performed.
  69  *   For example, while registering an MR, the associated PD's
  70  *   uobject.mutex must be held for reading.  The rwsem must be held
  71  *   for writing while initializing or destroying an object.
  72  *
  73  * - In addition, each object has a "live" flag.  If this flag is not
  74  *   set, then lookups of the object will fail even if it is found in
  75  *   the idr.  This handles a reader that blocks and does not acquire
  76  *   the rwsem until after the object is destroyed.  The destroy
  77  *   operation will set the live flag to 0 and then drop the rwsem;
  78  *   this will allow the reader to acquire the rwsem, see that the
  79  *   live flag is 0, and then drop the rwsem and its reference to
  80  *   object.  The underlying storage will not be freed until the last
  81  *   reference to the object is dropped.
  82  */
  83
  84 static void init_uobj(struct ib_uobject *uobj, u64 user_handle,
  85                       struct ib_ucontext *context, struct uverbs_lock_class *c)
  86 {
  87         uobj->user_handle = user_handle;
  88         uobj->context     = context;
  89         kref_init(&uobj->ref);
  90         init_rwsem(&uobj->mutex);
  91         lockdep_set_class_and_name(&uobj->mutex, &c->key, c->name);
  92         uobj->live        = 0;
  93 }
  94
  95 static void release_uobj(struct kref *kref)
  96 {
  97         kfree(container_of(kref, struct ib_uobject, ref));
  98 }
  99
 100 static void put_uobj(struct ib_uobject *uobj)
 101 {
 102         kref_put(&uobj->ref, release_uobj);
 103 }
 104
 105 static void put_uobj_read(struct ib_uobject *uobj)
 106 {
 107         up_read(&uobj->mutex);
 108         put_uobj(uobj);
 109 }
 110
 111 static void put_uobj_write(struct ib_uobject *uobj)
 112 {
 113         up_write(&uobj->mutex);
 114         put_uobj(uobj);
 115 }
 116
 117 static int idr_add_uobj(struct idr *idr, struct ib_uobject *uobj)
 118 {
 119         int ret;
 120
 121         idr_preload(GFP_KERNEL);
 122         spin_lock(&ib_uverbs_idr_lock);
 123
 124         ret = idr_alloc(idr, uobj, 0, 0, GFP_NOWAIT);
 125         if (ret >= 0)
 126                 uobj->id = ret;
 127
 128         spin_unlock(&ib_uverbs_idr_lock);
 129         idr_preload_end();
 130
 131         return ret < 0 ? ret : 0;
 132 }
 133
 134 void idr_remove_uobj(struct idr *idr, struct ib_uobject *uobj)
 135 {
 136         spin_lock(&ib_uverbs_idr_lock);
 137         idr_remove(idr, uobj->id);
 138         spin_unlock(&ib_uverbs_idr_lock);
 139 }
 140
 141 static struct ib_uobject *__idr_get_uobj(struct idr *idr, int id,
 142                                          struct ib_ucontext *context)
 143 {
 144         struct ib_uobject *uobj;
 145
 146         spin_lock(&ib_uverbs_idr_lock);
 147         uobj = idr_find(idr, id);
 148         if (uobj) {
 149                 if (uobj->context == context)
 150                         kref_get(&uobj->ref);
 151                 else
 152                         uobj = NULL;
 153         }
 154         spin_unlock(&ib_uverbs_idr_lock);
 155
 156         return uobj;
 157 }
 158
 159 static struct ib_uobject *idr_read_uobj(struct idr *idr, int id,
 160                                         struct ib_ucontext *context, int nested)
 161 {
 162         struct ib_uobject *uobj;
 163
 164         uobj = __idr_get_uobj(idr, id, context);
 165         if (!uobj)
 166                 return NULL;
 167
 168         if (nested)
 169                 down_read_nested(&uobj->mutex, SINGLE_DEPTH_NESTING);
 170         else
 171                 down_read(&uobj->mutex);
 172         if (!uobj->live) {
 173                 put_uobj_read(uobj);
 174                 return NULL;
 175         }
 176
 177         return uobj;
 178 }
 179
 180 static struct ib_uobject *idr_write_uobj(struct idr *idr, int id,
 181                                          struct ib_ucontext *context)
 182 {
 183         struct ib_uobject *uobj;
 184
 185         uobj = __idr_get_uobj(idr, id, context);
 186         if (!uobj)
 187                 return NULL;
 188
 189         down_write(&uobj->mutex);
 190         if (!uobj->live) {
 191                 put_uobj_write(uobj);
 192                 return NULL;
 193         }
 194
 195         return uobj;
 196 }
 197
 198 static void *idr_read_obj(struct idr *idr, int id, struct ib_ucontext *context,
 199                           int nested)
 200 {
 201         struct ib_uobject *uobj;
 202
 203         uobj = idr_read_uobj(idr, id, context, nested);
 204         return uobj ? uobj->object : NULL;
 205 }
 206
 207 static struct ib_pd *idr_read_pd(int pd_handle, struct ib_ucontext *context)
 208 {
 209         return idr_read_obj(&ib_uverbs_pd_idr, pd_handle, context, 0);
 210 }
 211
 212 static void put_pd_read(struct ib_pd *pd)
 213 {
 214         put_uobj_read(pd->uobject);
 215 }
 216
 217 static struct ib_cq *idr_read_cq(int cq_handle, struct ib_ucontext *context, int nested)
 218 {
 219         return idr_read_obj(&ib_uverbs_cq_idr, cq_handle, context, nested);
 220 }
 221
 222 static void put_cq_read(struct ib_cq *cq)
 223 {
 224         put_uobj_read(cq->uobject);
 225 }
 226
 227 static struct ib_ah *idr_read_ah(int ah_handle, struct ib_ucontext *context)
 228 {
 229         return idr_read_obj(&ib_uverbs_ah_idr, ah_handle, context, 0);
 230 }
 231
 232 static void put_ah_read(struct ib_ah *ah)
 233 {
 234         put_uobj_read(ah->uobject);
 235 }
 236
 237 static struct ib_qp *idr_read_qp(int qp_handle, struct ib_ucontext *context)
 238 {
 239         return idr_read_obj(&ib_uverbs_qp_idr, qp_handle, context, 0);
 240 }
 241
 242 static struct ib_qp *idr_write_qp(int qp_handle, struct ib_ucontext *context)
 243 {
 244         struct ib_uobject *uobj;
 245
 246         uobj = idr_write_uobj(&ib_uverbs_qp_idr, qp_handle, context);
 247         return uobj ? uobj->object : NULL;
 248 }
 249
 250 static void put_qp_read(struct ib_qp *qp)
 251 {
 252         put_uobj_read(qp->uobject);
 253 }
 254
 255 static void put_qp_write(struct ib_qp *qp)
 256 {
 257         put_uobj_write(qp->uobject);
 258 }
 259
 260 static struct ib_srq *idr_read_srq(int srq_handle, struct ib_ucontext *context)
 261 {
 262         return idr_read_obj(&ib_uverbs_srq_idr, srq_handle, context, 0);
 263 }
 264
 265 static void put_srq_read(struct ib_srq *srq)
 266 {
 267         put_uobj_read(srq->uobject);
 268 }
 269
 270 static struct ib_xrcd *idr_read_xrcd(int xrcd_handle, struct ib_ucontext *context,
 271                                      struct ib_uobject **uobj)
 272 {
 273         *uobj = idr_read_uobj(&ib_uverbs_xrcd_idr, xrcd_handle, context, 0);
 274         return *uobj ? (*uobj)->object : NULL;
 275 }
 276
 277 static void put_xrcd_read(struct ib_uobject *uobj)
 278 {
 279         put_uobj_read(uobj);
 280 }
 281
 282 ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
 283                               const char __user *buf,
 284                               int in_len, int out_len)
 285 {
 286         struct ib_uverbs_get_context      cmd;
 287         struct ib_uverbs_get_context_resp resp;
 288         struct ib_udata                   udata;
 289         struct ib_device                 *ibdev = file->device->ib_dev;
 290         struct ib_ucontext               *ucontext;
 291         struct file                      *filp;
 292         int ret;
 293
 294         if (out_len < sizeof resp)
 295                 return -ENOSPC;
 296
 297         if (copy_from_user(&cmd, buf, sizeof cmd))
 298                 return -EFAULT;
 299
 300         mutex_lock(&file->mutex);
 301
 302         if (file->ucontext) {
 303                 ret = -EINVAL;
 304                 goto err;
 305         }
 306
 307         INIT_UDATA(&udata, buf + sizeof cmd,
 308                    (unsigned long) cmd.response + sizeof resp,
 309                    in_len - sizeof cmd, out_len - sizeof resp);
 310
 311         ucontext = ibdev->alloc_ucontext(ibdev, &udata);
 312         if (IS_ERR(ucontext)) {
 313                 ret = PTR_ERR(ucontext);
 314                 goto err;
 315         }
 316
 317         ucontext->device = ibdev;
 318         INIT_LIST_HEAD(&ucontext->pd_list);
 319         INIT_LIST_HEAD(&ucontext->mr_list);
 320         INIT_LIST_HEAD(&ucontext->mw_list);
 321         INIT_LIST_HEAD(&ucontext->cq_list);
 322         INIT_LIST_HEAD(&ucontext->qp_list);
 323         INIT_LIST_HEAD(&ucontext->srq_list);
 324         INIT_LIST_HEAD(&ucontext->ah_list);
 325         INIT_LIST_HEAD(&ucontext->xrcd_list);
 326         INIT_LIST_HEAD(&ucontext->rule_list);
 327         ucontext->closing = 0;
 328
 329         resp.num_comp_vectors = file->device->num_comp_vectors;
 330
 331         ret = get_unused_fd_flags(O_CLOEXEC);
 332         if (ret < 0)
 333                 goto err_free;
 334         resp.async_fd = ret;
 335
 336         filp = ib_uverbs_alloc_event_file(file, 1);
 337         if (IS_ERR(filp)) {
 338                 ret = PTR_ERR(filp);
 339                 goto err_fd;
 340         }
 341
 342         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 343                          &resp, sizeof resp)) {
 344                 ret = -EFAULT;
 345                 goto err_file;
 346         }
 347
 348         file->async_file = filp->private_data;
 349
 350         INIT_IB_EVENT_HANDLER(&file->event_handler, file->device->ib_dev,
 351                               ib_uverbs_event_handler);
 352         ret = ib_register_event_handler(&file->event_handler);
 353         if (ret)
 354                 goto err_file;
 355
 356         kref_get(&file->async_file->ref);
 357         kref_get(&file->ref);
 358         file->ucontext = ucontext;
 359
 360         fd_install(resp.async_fd, filp);
 361
 362         mutex_unlock(&file->mutex);
 363
 364         return in_len;
 365
 366 err_file:
 367         fput(filp);
 368
 369 err_fd:
 370         put_unused_fd(resp.async_fd);
 371
 372 err_free:
 373         ibdev->dealloc_ucontext(ucontext);
 374
 375 err:
 376         mutex_unlock(&file->mutex);
 377         return ret;
 378 }
 379
 380 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 381                                const char __user *buf,
 382                                int in_len, int out_len)
 383 {
 384         struct ib_uverbs_query_device      cmd;
 385         struct ib_uverbs_query_device_resp resp;
 386         struct ib_device_attr              attr;
 387         int                                ret;
 388
 389         if (out_len < sizeof resp)
 390                 return -ENOSPC;
 391
 392         if (copy_from_user(&cmd, buf, sizeof cmd))
 393                 return -EFAULT;
 394
 395         ret = ib_query_device(file->device->ib_dev, &attr);
 396         if (ret)
 397                 return ret;
 398
 399         memset(&resp, 0, sizeof resp);
 400
 401         resp.fw_ver                    = attr.fw_ver;
 402         resp.node_guid                 = file->device->ib_dev->node_guid;
 403         resp.sys_image_guid            = attr.sys_image_guid;
 404         resp.max_mr_size               = attr.max_mr_size;
 405         resp.page_size_cap             = attr.page_size_cap;
 406         resp.vendor_id                 = attr.vendor_id;
 407         resp.vendor_part_id            = attr.vendor_part_id;
 408         resp.hw_ver                    = attr.hw_ver;
 409         resp.max_qp                    = attr.max_qp;
 410         resp.max_qp_wr                 = attr.max_qp_wr;
 411         resp.device_cap_flags          = attr.device_cap_flags;
 412         resp.max_sge                   = attr.max_sge;
 413         resp.max_sge_rd                = attr.max_sge_rd;
 414         resp.max_cq                    = attr.max_cq;
 415         resp.max_cqe                   = attr.max_cqe;
 416         resp.max_mr                    = attr.max_mr;
 417         resp.max_pd                    = attr.max_pd;
 418         resp.max_qp_rd_atom            = attr.max_qp_rd_atom;
 419         resp.max_ee_rd_atom            = attr.max_ee_rd_atom;
 420         resp.max_res_rd_atom           = attr.max_res_rd_atom;
 421         resp.max_qp_init_rd_atom       = attr.max_qp_init_rd_atom;
 422         resp.max_ee_init_rd_atom       = attr.max_ee_init_rd_atom;
 423         resp.atomic_cap                = attr.atomic_cap;
 424         resp.max_ee                    = attr.max_ee;
 425         resp.max_rdd                   = attr.max_rdd;
 426         resp.max_mw                    = attr.max_mw;
 427         resp.max_raw_ipv6_qp           = attr.max_raw_ipv6_qp;
 428         resp.max_raw_ethy_qp           = attr.max_raw_ethy_qp;
 429         resp.max_mcast_grp             = attr.max_mcast_grp;
 430         resp.max_mcast_qp_attach       = attr.max_mcast_qp_attach;
 431         resp.max_total_mcast_qp_attach = attr.max_total_mcast_qp_attach;
 432         resp.max_ah                    = attr.max_ah;
 433         resp.max_fmr                   = attr.max_fmr;
 434         resp.max_map_per_fmr           = attr.max_map_per_fmr;
 435         resp.max_srq                   = attr.max_srq;
 436         resp.max_srq_wr                = attr.max_srq_wr;
 437         resp.max_srq_sge               = attr.max_srq_sge;
 438         resp.max_pkeys                 = attr.max_pkeys;
 439         resp.local_ca_ack_delay        = attr.local_ca_ack_delay;
 440         resp.phys_port_cnt             = file->device->ib_dev->phys_port_cnt;
 441
 442         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 443                          &resp, sizeof resp))
 444                 return -EFAULT;
 445
 446         return in_len;
 447 }
 448
 449 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 450                              const char __user *buf,
 451                              int in_len, int out_len)
 452 {
 453         struct ib_uverbs_query_port      cmd;
 454         struct ib_uverbs_query_port_resp resp;
 455         struct ib_port_attr              attr;
 456         int                              ret;
 457
 458         if (out_len < sizeof resp)
 459                 return -ENOSPC;
 460
 461         if (copy_from_user(&cmd, buf, sizeof cmd))
 462                 return -EFAULT;
 463
 464         ret = ib_query_port(file->device->ib_dev, cmd.port_num, &attr);
 465         if (ret)
 466                 return ret;
 467
 468         memset(&resp, 0, sizeof resp);
 469
 470         resp.state           = attr.state;
 471         resp.max_mtu         = attr.max_mtu;
 472         resp.active_mtu      = attr.active_mtu;
 473         resp.gid_tbl_len     = attr.gid_tbl_len;
 474         resp.port_cap_flags  = attr.port_cap_flags;
 475         resp.max_msg_sz      = attr.max_msg_sz;
 476         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 477         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 478         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 479         resp.lid             = attr.lid;
 480         resp.sm_lid          = attr.sm_lid;
 481         resp.lmc             = attr.lmc;
 482         resp.max_vl_num      = attr.max_vl_num;
 483         resp.sm_sl           = attr.sm_sl;
 484         resp.subnet_timeout  = attr.subnet_timeout;
 485         resp.init_type_reply = attr.init_type_reply;
 486         resp.active_width    = attr.active_width;
 487         resp.active_speed    = attr.active_speed;
 488         resp.phys_state      = attr.phys_state;
 489         resp.link_layer      = rdma_port_get_link_layer(file->device->ib_dev,
 490                                                         cmd.port_num);
 491
 492         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 493                          &resp, sizeof resp))
 494                 return -EFAULT;
 495
 496         return in_len;
 497 }
 498
 499 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 500                            const char __user *buf,
 501                            int in_len, int out_len)
 502 {
 503         struct ib_uverbs_alloc_pd      cmd;
 504         struct ib_uverbs_alloc_pd_resp resp;
 505         struct ib_udata                udata;
 506         struct ib_uobject             *uobj;
 507         struct ib_pd                  *pd;
 508         int                            ret;
 509
 510         if (out_len < sizeof resp)
 511                 return -ENOSPC;
 512
 513         if (copy_from_user(&cmd, buf, sizeof cmd))
 514                 return -EFAULT;
 515
 516         INIT_UDATA(&udata, buf + sizeof cmd,
 517                    (unsigned long) cmd.response + sizeof resp,
 518                    in_len - sizeof cmd, out_len - sizeof resp);
 519
 520         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 521         if (!uobj)
 522                 return -ENOMEM;
 523
 524         init_uobj(uobj, 0, file->ucontext, &pd_lock_class);
 525         down_write(&uobj->mutex);
 526
 527         pd = file->device->ib_dev->alloc_pd(file->device->ib_dev,
 528                                             file->ucontext, &udata);
 529         if (IS_ERR(pd)) {
 530                 ret = PTR_ERR(pd);
 531                 goto err;
 532         }
 533
 534         pd->device  = file->device->ib_dev;
 535         pd->uobject = uobj;
 536         atomic_set(&pd->usecnt, 0);
 537
 538         uobj->object = pd;
 539         ret = idr_add_uobj(&ib_uverbs_pd_idr, uobj);
 540         if (ret)
 541                 goto err_idr;
 542
 543         memset(&resp, 0, sizeof resp);
 544         resp.pd_handle = uobj->id;
 545
 546         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 547                          &resp, sizeof resp)) {
 548                 ret = -EFAULT;
 549                 goto err_copy;
 550         }
 551
 552         mutex_lock(&file->mutex);
 553         list_add_tail(&uobj->list, &file->ucontext->pd_list);
 554         mutex_unlock(&file->mutex);
 555
 556         uobj->live = 1;
 557
 558         up_write(&uobj->mutex);
 559
 560         return in_len;
 561
 562 err_copy:
 563         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 564
 565 err_idr:
 566         ib_dealloc_pd(pd);
 567
 568 err:
 569         put_uobj_write(uobj);
 570         return ret;
 571 }
 572
 573 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 574                              const char __user *buf,
 575                              int in_len, int out_len)
 576 {
 577         struct ib_uverbs_dealloc_pd cmd;
 578         struct ib_uobject          *uobj;
 579         int                         ret;
 580
 581         if (copy_from_user(&cmd, buf, sizeof cmd))
 582                 return -EFAULT;
 583
 584         uobj = idr_write_uobj(&ib_uverbs_pd_idr, cmd.pd_handle, file->ucontext);
 585         if (!uobj)
 586                 return -EINVAL;
 587
 588         ret = ib_dealloc_pd(uobj->object);
 589         if (!ret)
 590                 uobj->live = 0;
 591
 592         put_uobj_write(uobj);
 593
 594         if (ret)
 595                 return ret;
 596
 597         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 598
 599         mutex_lock(&file->mutex);
 600         list_del(&uobj->list);
 601         mutex_unlock(&file->mutex);
 602
 603         put_uobj(uobj);
 604
 605         return in_len;
 606 }
 607
 608 struct xrcd_table_entry {
 609         struct rb_node  node;
 610         struct ib_xrcd *xrcd;
 611         struct inode   *inode;
 612 };
 613
 614 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 615                             struct inode *inode,
 616                             struct ib_xrcd *xrcd)
 617 {
 618         struct xrcd_table_entry *entry, *scan;
 619         struct rb_node **p = &dev->xrcd_tree.rb_node;
 620         struct rb_node *parent = NULL;
 621
 622         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 623         if (!entry)
 624                 return -ENOMEM;
 625
 626         entry->xrcd  = xrcd;
 627         entry->inode = inode;
 628
 629         while (*p) {
 630                 parent = *p;
 631                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 632
 633                 if (inode < scan->inode) {
 634                         p = &(*p)->rb_left;
 635                 } else if (inode > scan->inode) {
 636                         p = &(*p)->rb_right;
 637                 } else {
 638                         kfree(entry);
 639                         return -EEXIST;
 640                 }
 641         }
 642
 643         rb_link_node(&entry->node, parent, p);
 644         rb_insert_color(&entry->node, &dev->xrcd_tree);
 645         igrab(inode);
 646         return 0;
 647 }
 648
 649 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 650                                                   struct inode *inode)
 651 {
 652         struct xrcd_table_entry *entry;
 653         struct rb_node *p = dev->xrcd_tree.rb_node;
 654
 655         while (p) {
 656                 entry = rb_entry(p, struct xrcd_table_entry, node);
 657
 658                 if (inode < entry->inode)
 659                         p = p->rb_left;
 660                 else if (inode > entry->inode)
 661                         p = p->rb_right;
 662                 else
 663                         return entry;
 664         }
 665
 666         return NULL;
 667 }
 668
 669 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 670 {
 671         struct xrcd_table_entry *entry;
 672
 673         entry = xrcd_table_search(dev, inode);
 674         if (!entry)
 675                 return NULL;
 676
 677         return entry->xrcd;
 678 }
 679
 680 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 681                               struct inode *inode)
 682 {
 683         struct xrcd_table_entry *entry;
 684
 685         entry = xrcd_table_search(dev, inode);
 686         if (entry) {
 687                 iput(inode);
 688                 rb_erase(&entry->node, &dev->xrcd_tree);
 689                 kfree(entry);
 690         }
 691 }
 692
 693 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 694                             const char __user *buf, int in_len,
 695                             int out_len)
 696 {
 697         struct ib_uverbs_open_xrcd      cmd;
 698         struct ib_uverbs_open_xrcd_resp resp;
 699         struct ib_udata                 udata;
 700         struct ib_uxrcd_object         *obj;
 701         struct ib_xrcd                 *xrcd = NULL;
 702         struct fd                       f = {NULL, 0};
 703         struct inode                   *inode = NULL;
 704         int                             ret = 0;
 705         int                             new_xrcd = 0;
 706
 707         if (out_len < sizeof resp)
 708                 return -ENOSPC;
 709
 710         if (copy_from_user(&cmd, buf, sizeof cmd))
 711                 return -EFAULT;
 712
 713         INIT_UDATA(&udata, buf + sizeof cmd,
 714                    (unsigned long) cmd.response + sizeof resp,
 715                    in_len - sizeof cmd, out_len - sizeof  resp);
 716
 717         mutex_lock(&file->device->xrcd_tree_mutex);
 718
 719         if (cmd.fd != -1) {
 720                 /* search for file descriptor */
 721                 f = fdget(cmd.fd);
 722                 if (!f.file) {
 723                         ret = -EBADF;
 724                         goto err_tree_mutex_unlock;
 725                 }
 726
 727                 inode = file_inode(f.file);
 728                 xrcd = find_xrcd(file->device, inode);
 729                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 730                         /* no file descriptor. Need CREATE flag */
 731                         ret = -EAGAIN;
 732                         goto err_tree_mutex_unlock;
 733                 }
 734
 735                 if (xrcd && cmd.oflags & O_EXCL) {
 736                         ret = -EINVAL;
 737                         goto err_tree_mutex_unlock;
 738                 }
 739         }
 740
 741         obj = kmalloc(sizeof *obj, GFP_KERNEL);
 742         if (!obj) {
 743                 ret = -ENOMEM;
 744                 goto err_tree_mutex_unlock;
 745         }
 746
 747         init_uobj(&obj->uobject, 0, file->ucontext, &xrcd_lock_class);
 748
 749         down_write(&obj->uobject.mutex);
 750
 751         if (!xrcd) {
 752                 xrcd = file->device->ib_dev->alloc_xrcd(file->device->ib_dev,
 753                                                         file->ucontext, &udata);
 754                 if (IS_ERR(xrcd)) {
 755                         ret = PTR_ERR(xrcd);
 756                         goto err;
 757                 }
 758
 759                 xrcd->inode   = inode;
 760                 xrcd->device  = file->device->ib_dev;
 761                 atomic_set(&xrcd->usecnt, 0);
 762                 mutex_init(&xrcd->tgt_qp_mutex);
 763                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 764                 new_xrcd = 1;
 765         }
 766
 767         atomic_set(&obj->refcnt, 0);
 768         obj->uobject.object = xrcd;
 769         ret = idr_add_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 770         if (ret)
 771                 goto err_idr;
 772
 773         memset(&resp, 0, sizeof resp);
 774         resp.xrcd_handle = obj->uobject.id;
 775
 776         if (inode) {
 777                 if (new_xrcd) {
 778                         /* create new inode/xrcd table entry */
 779                         ret = xrcd_table_insert(file->device, inode, xrcd);
 780                         if (ret)
 781                                 goto err_insert_xrcd;
 782                 }
 783                 atomic_inc(&xrcd->usecnt);
 784         }
 785
 786         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 787                          &resp, sizeof resp)) {
 788                 ret = -EFAULT;
 789                 goto err_copy;
 790         }
 791
 792         if (f.file)
 793                 fdput(f);
 794
 795         mutex_lock(&file->mutex);
 796         list_add_tail(&obj->uobject.list, &file->ucontext->xrcd_list);
 797         mutex_unlock(&file->mutex);
 798
 799         obj->uobject.live = 1;
 800         up_write(&obj->uobject.mutex);
 801
 802         mutex_unlock(&file->device->xrcd_tree_mutex);
 803         return in_len;
 804
 805 err_copy:
 806         if (inode) {
 807                 if (new_xrcd)
 808                         xrcd_table_delete(file->device, inode);
 809                 atomic_dec(&xrcd->usecnt);
 810         }
 811
 812 err_insert_xrcd:
 813         idr_remove_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 814
 815 err_idr:
 816         ib_dealloc_xrcd(xrcd);
 817
 818 err:
 819         put_uobj_write(&obj->uobject);
 820
 821 err_tree_mutex_unlock:
 822         if (f.file)
 823                 fdput(f);
 824
 825         mutex_unlock(&file->device->xrcd_tree_mutex);
 826
 827         return ret;
 828 }
 829
 830 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 831                              const char __user *buf, int in_len,
 832                              int out_len)
 833 {
 834         struct ib_uverbs_close_xrcd cmd;
 835         struct ib_uobject           *uobj;
 836         struct ib_xrcd              *xrcd = NULL;
 837         struct inode                *inode = NULL;
 838         struct ib_uxrcd_object      *obj;
 839         int                         live;
 840         int                         ret = 0;
 841
 842         if (copy_from_user(&cmd, buf, sizeof cmd))
 843                 return -EFAULT;
 844
 845         mutex_lock(&file->device->xrcd_tree_mutex);
 846         uobj = idr_write_uobj(&ib_uverbs_xrcd_idr, cmd.xrcd_handle, file->ucontext);
 847         if (!uobj) {
 848                 ret = -EINVAL;
 849                 goto out;
 850         }
 851
 852         xrcd  = uobj->object;
 853         inode = xrcd->inode;
 854         obj   = container_of(uobj, struct ib_uxrcd_object, uobject);
 855         if (atomic_read(&obj->refcnt)) {
 856                 put_uobj_write(uobj);
 857                 ret = -EBUSY;
 858                 goto out;
 859         }
 860
 861         if (!inode || atomic_dec_and_test(&xrcd->usecnt)) {
 862                 ret = ib_dealloc_xrcd(uobj->object);
 863                 if (!ret)
 864                         uobj->live = 0;
 865         }
 866
 867         live = uobj->live;
 868         if (inode && ret)
 869                 atomic_inc(&xrcd->usecnt);
 870
 871         put_uobj_write(uobj);
 872
 873         if (ret)
 874                 goto out;
 875
 876         if (inode && !live)
 877                 xrcd_table_delete(file->device, inode);
 878
 879         idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 880         mutex_lock(&file->mutex);
 881         list_del(&uobj->list);
 882         mutex_unlock(&file->mutex);
 883
 884         put_uobj(uobj);
 885         ret = in_len;
 886
 887 out:
 888         mutex_unlock(&file->device->xrcd_tree_mutex);
 889         return ret;
 890 }
 891
 892 void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 893                             struct ib_xrcd *xrcd)
 894 {
 895         struct inode *inode;
 896
 897         inode = xrcd->inode;
 898         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 899                 return;
 900
 901         ib_dealloc_xrcd(xrcd);
 902
 903         if (inode)
 904                 xrcd_table_delete(dev, inode);
 905 }
 906
 907 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 908                          const char __user *buf, int in_len,
 909                          int out_len)
 910 {
 911         struct ib_uverbs_reg_mr      cmd;
 912         struct ib_uverbs_reg_mr_resp resp;
 913         struct ib_udata              udata;
 914         struct ib_uobject           *uobj;
 915         struct ib_pd                *pd;
 916         struct ib_mr                *mr;
 917         int                          ret;
 918
 919         if (out_len < sizeof resp)
 920                 return -ENOSPC;
 921
 922         if (copy_from_user(&cmd, buf, sizeof cmd))
 923                 return -EFAULT;
 924
 925         INIT_UDATA(&udata, buf + sizeof cmd,
 926                    (unsigned long) cmd.response + sizeof resp,
 927                    in_len - sizeof cmd, out_len - sizeof resp);
 928
 929         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 930                 return -EINVAL;
 931
 932         ret = ib_check_mr_access(cmd.access_flags);
 933         if (ret)
 934                 return ret;
 935
 936         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 937         if (!uobj)
 938                 return -ENOMEM;
 939
 940         init_uobj(uobj, 0, file->ucontext, &mr_lock_class);
 941         down_write(&uobj->mutex);
 942
 943         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
 944         if (!pd) {
 945                 ret = -EINVAL;
 946                 goto err_free;
 947         }
 948
 949         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 950                                      cmd.access_flags, &udata);
 951         if (IS_ERR(mr)) {
 952                 ret = PTR_ERR(mr);
 953                 goto err_put;
 954         }
 955
 956         mr->device  = pd->device;
 957         mr->pd      = pd;
 958         mr->uobject = uobj;
 959         atomic_inc(&pd->usecnt);
 960         atomic_set(&mr->usecnt, 0);
 961
 962         uobj->object = mr;
 963         ret = idr_add_uobj(&ib_uverbs_mr_idr, uobj);
 964         if (ret)
 965                 goto err_unreg;
 966
 967         memset(&resp, 0, sizeof resp);
 968         resp.lkey      = mr->lkey;
 969         resp.rkey      = mr->rkey;
 970         resp.mr_handle = uobj->id;
 971
 972         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 973                          &resp, sizeof resp)) {
 974                 ret = -EFAULT;
 975                 goto err_copy;
 976         }
 977
 978         put_pd_read(pd);
 979
 980         mutex_lock(&file->mutex);
 981         list_add_tail(&uobj->list, &file->ucontext->mr_list);
 982         mutex_unlock(&file->mutex);
 983
 984         uobj->live = 1;
 985
 986         up_write(&uobj->mutex);
 987
 988         return in_len;
 989
 990 err_copy:
 991         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 992
 993 err_unreg:
 994         ib_dereg_mr(mr);
 995
 996 err_put:
 997         put_pd_read(pd);
 998
 999 err_free:
1000         put_uobj_write(uobj);
1001         return ret;
1002 }
1003
1004 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
1005                            const char __user *buf, int in_len,
1006                            int out_len)
1007 {
1008         struct ib_uverbs_dereg_mr cmd;
1009         struct ib_mr             *mr;
1010         struct ib_uobject        *uobj;
1011         int                       ret = -EINVAL;
1012
1013         if (copy_from_user(&cmd, buf, sizeof cmd))
1014                 return -EFAULT;
1015
1016         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle, file->ucontext);
1017         if (!uobj)
1018                 return -EINVAL;
1019
1020         mr = uobj->object;
1021
1022         ret = ib_dereg_mr(mr);
1023         if (!ret)
1024                 uobj->live = 0;
1025
1026         put_uobj_write(uobj);
1027
1028         if (ret)
1029                 return ret;
1030
1031         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1032
1033         mutex_lock(&file->mutex);
1034         list_del(&uobj->list);
1035         mutex_unlock(&file->mutex);
1036
1037         put_uobj(uobj);
1038
1039         return in_len;
1040 }
1041
1042 ssize_t ib_uverbs_alloc_mw(struct ib_uverbs_file *file,
1043                          const char __user *buf, int in_len,
1044                          int out_len)
1045 {
1046         struct ib_uverbs_alloc_mw      cmd;
1047         struct ib_uverbs_alloc_mw_resp resp;
1048         struct ib_uobject             *uobj;
1049         struct ib_pd                  *pd;
1050         struct ib_mw                  *mw;
1051         int                            ret;
1052
1053         if (out_len < sizeof(resp))
1054                 return -ENOSPC;
1055
1056         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1057                 return -EFAULT;
1058
1059         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
1060         if (!uobj)
1061                 return -ENOMEM;
1062
1063         init_uobj(uobj, 0, file->ucontext, &mw_lock_class);
1064         down_write(&uobj->mutex);
1065
1066         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
1067         if (!pd) {
1068                 ret = -EINVAL;
1069                 goto err_free;
1070         }
1071
1072         mw = pd->device->alloc_mw(pd, cmd.mw_type);
1073         if (IS_ERR(mw)) {
1074                 ret = PTR_ERR(mw);
1075                 goto err_put;
1076         }
1077
1078         mw->device  = pd->device;
1079         mw->pd      = pd;
1080         mw->uobject = uobj;
1081         atomic_inc(&pd->usecnt);
1082
1083         uobj->object = mw;
1084         ret = idr_add_uobj(&ib_uverbs_mw_idr, uobj);
1085         if (ret)
1086                 goto err_unalloc;
1087
1088         memset(&resp, 0, sizeof(resp));
1089         resp.rkey      = mw->rkey;
1090         resp.mw_handle = uobj->id;
1091
1092         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1093                          &resp, sizeof(resp))) {
1094                 ret = -EFAULT;
1095                 goto err_copy;
1096         }
1097
1098         put_pd_read(pd);
1099
1100         mutex_lock(&file->mutex);
1101         list_add_tail(&uobj->list, &file->ucontext->mw_list);
1102         mutex_unlock(&file->mutex);
1103
1104         uobj->live = 1;
1105
1106         up_write(&uobj->mutex);
1107
1108         return in_len;
1109
1110 err_copy:
1111         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1112
1113 err_unalloc:
1114         ib_dealloc_mw(mw);
1115
1116 err_put:
1117         put_pd_read(pd);
1118
1119 err_free:
1120         put_uobj_write(uobj);
1121         return ret;
1122 }
1123
1124 ssize_t ib_uverbs_dealloc_mw(struct ib_uverbs_file *file,
1125                            const char __user *buf, int in_len,
1126                            int out_len)
1127 {
1128         struct ib_uverbs_dealloc_mw cmd;
1129         struct ib_mw               *mw;
1130         struct ib_uobject          *uobj;
1131         int                         ret = -EINVAL;
1132
1133         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1134                 return -EFAULT;
1135
1136         uobj = idr_write_uobj(&ib_uverbs_mw_idr, cmd.mw_handle, file->ucontext);
1137         if (!uobj)
1138                 return -EINVAL;
1139
1140         mw = uobj->object;
1141
1142         ret = ib_dealloc_mw(mw);
1143         if (!ret)
1144                 uobj->live = 0;
1145
1146         put_uobj_write(uobj);
1147
1148         if (ret)
1149                 return ret;
1150
1151         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1152
1153         mutex_lock(&file->mutex);
1154         list_del(&uobj->list);
1155         mutex_unlock(&file->mutex);
1156
1157         put_uobj(uobj);
1158
1159         return in_len;
1160 }
1161
1162 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
1163                                       const char __user *buf, int in_len,
1164                                       int out_len)
1165 {
1166         struct ib_uverbs_create_comp_channel       cmd;
1167         struct ib_uverbs_create_comp_channel_resp  resp;
1168         struct file                               *filp;
1169         int ret;
1170
1171         if (out_len < sizeof resp)
1172                 return -ENOSPC;
1173
1174         if (copy_from_user(&cmd, buf, sizeof cmd))
1175                 return -EFAULT;
1176
1177         ret = get_unused_fd_flags(O_CLOEXEC);
1178         if (ret < 0)
1179                 return ret;
1180         resp.fd = ret;
1181
1182         filp = ib_uverbs_alloc_event_file(file, 0);
1183         if (IS_ERR(filp)) {
1184                 put_unused_fd(resp.fd);
1185                 return PTR_ERR(filp);
1186         }
1187
1188         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1189                          &resp, sizeof resp)) {
1190                 put_unused_fd(resp.fd);
1191                 fput(filp);
1192                 return -EFAULT;
1193         }
1194
1195         fd_install(resp.fd, filp);
1196         return in_len;
1197 }
1198
1199 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1200                             const char __user *buf, int in_len,
1201                             int out_len)
1202 {
1203         struct ib_uverbs_create_cq      cmd;
1204         struct ib_uverbs_create_cq_resp resp;
1205         struct ib_udata                 udata;
1206         struct ib_ucq_object           *obj;
1207         struct ib_uverbs_event_file    *ev_file = NULL;
1208         struct ib_cq                   *cq;
1209         int                             ret;
1210
1211         if (out_len < sizeof resp)
1212                 return -ENOSPC;
1213
1214         if (copy_from_user(&cmd, buf, sizeof cmd))
1215                 return -EFAULT;
1216
1217         INIT_UDATA(&udata, buf + sizeof cmd,
1218                    (unsigned long) cmd.response + sizeof resp,
1219                    in_len - sizeof cmd, out_len - sizeof resp);
1220
1221         if (cmd.comp_vector >= file->device->num_comp_vectors)
1222                 return -EINVAL;
1223
1224         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1225         if (!obj)
1226                 return -ENOMEM;
1227
1228         init_uobj(&obj->uobject, cmd.user_handle, file->ucontext, &cq_lock_class);
1229         down_write(&obj->uobject.mutex);
1230
1231         if (cmd.comp_channel >= 0) {
1232                 ev_file = ib_uverbs_lookup_comp_file(cmd.comp_channel);
1233                 if (!ev_file) {
1234                         ret = -EINVAL;
1235                         goto err;
1236                 }
1237         }
1238
1239         obj->uverbs_file           = file;
1240         obj->comp_events_reported  = 0;
1241         obj->async_events_reported = 0;
1242         INIT_LIST_HEAD(&obj->comp_list);
1243         INIT_LIST_HEAD(&obj->async_list);
1244
1245         cq = file->device->ib_dev->create_cq(file->device->ib_dev, cmd.cqe,
1246                                              cmd.comp_vector,
1247                                              file->ucontext, &udata);
1248         if (IS_ERR(cq)) {
1249                 ret = PTR_ERR(cq);
1250                 goto err_file;
1251         }
1252
1253         cq->device        = file->device->ib_dev;
1254         cq->uobject       = &obj->uobject;
1255         cq->comp_handler  = ib_uverbs_comp_handler;
1256         cq->event_handler = ib_uverbs_cq_event_handler;
1257         cq->cq_context    = ev_file;
1258         atomic_set(&cq->usecnt, 0);
1259
1260         obj->uobject.object = cq;
1261         ret = idr_add_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1262         if (ret)
1263                 goto err_free;
1264
1265         memset(&resp, 0, sizeof resp);
1266         resp.cq_handle = obj->uobject.id;
1267         resp.cqe       = cq->cqe;
1268
1269         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1270                          &resp, sizeof resp)) {
1271                 ret = -EFAULT;
1272                 goto err_copy;
1273         }
1274
1275         mutex_lock(&file->mutex);
1276         list_add_tail(&obj->uobject.list, &file->ucontext->cq_list);
1277         mutex_unlock(&file->mutex);
1278
1279         obj->uobject.live = 1;
1280
1281         up_write(&obj->uobject.mutex);
1282
1283         return in_len;
1284
1285 err_copy:
1286         idr_remove_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1287
1288 err_free:
1289         ib_destroy_cq(cq);
1290
1291 err_file:
1292         if (ev_file)
1293                 ib_uverbs_release_ucq(file, ev_file, obj);
1294
1295 err:
1296         put_uobj_write(&obj->uobject);
1297         return ret;
1298 }
1299
1300 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1301                             const char __user *buf, int in_len,
1302                             int out_len)
1303 {
1304         struct ib_uverbs_resize_cq      cmd;
1305         struct ib_uverbs_resize_cq_resp resp;
1306         struct ib_udata                 udata;
1307         struct ib_cq                    *cq;
1308         int                             ret = -EINVAL;
1309
1310         if (copy_from_user(&cmd, buf, sizeof cmd))
1311                 return -EFAULT;
1312
1313         INIT_UDATA(&udata, buf + sizeof cmd,
1314                    (unsigned long) cmd.response + sizeof resp,
1315                    in_len - sizeof cmd, out_len - sizeof resp);
1316
1317         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1318         if (!cq)
1319                 return -EINVAL;
1320
1321         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1322         if (ret)
1323                 goto out;
1324
1325         resp.cqe = cq->cqe;
1326
1327         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1328                          &resp, sizeof resp.cqe))
1329                 ret = -EFAULT;
1330
1331 out:
1332         put_cq_read(cq);
1333
1334         return ret ? ret : in_len;
1335 }
1336
1337 static int copy_wc_to_user(void __user *dest, struct ib_wc *wc)
1338 {
1339         struct ib_uverbs_wc tmp;
1340
1341         tmp.wr_id               = wc->wr_id;
1342         tmp.status              = wc->status;
1343         tmp.opcode              = wc->opcode;
1344         tmp.vendor_err          = wc->vendor_err;
1345         tmp.byte_len            = wc->byte_len;
1346         tmp.ex.imm_data         = (__u32 __force) wc->ex.imm_data;
1347         tmp.qp_num              = wc->qp->qp_num;
1348         tmp.src_qp              = wc->src_qp;
1349         tmp.wc_flags            = wc->wc_flags;
1350         tmp.pkey_index          = wc->pkey_index;
1351         tmp.slid                = wc->slid;
1352         tmp.sl                  = wc->sl;
1353         tmp.dlid_path_bits      = wc->dlid_path_bits;
1354         tmp.port_num            = wc->port_num;
1355         tmp.reserved            = 0;
1356
1357         if (copy_to_user(dest, &tmp, sizeof tmp))
1358                 return -EFAULT;
1359
1360         return 0;
1361 }
1362
1363 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1364                           const char __user *buf, int in_len,
1365                           int out_len)
1366 {
1367         struct ib_uverbs_poll_cq       cmd;
1368         struct ib_uverbs_poll_cq_resp  resp;
1369         u8 __user                     *header_ptr;
1370         u8 __user                     *data_ptr;
1371         struct ib_cq                  *cq;
1372         struct ib_wc                   wc;
1373         int                            ret;
1374
1375         if (copy_from_user(&cmd, buf, sizeof cmd))
1376                 return -EFAULT;
1377
1378         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1379         if (!cq)
1380                 return -EINVAL;
1381
1382         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1383         header_ptr = (void __user *)(unsigned long) cmd.response;
1384         data_ptr = header_ptr + sizeof resp;
1385
1386         memset(&resp, 0, sizeof resp);
1387         while (resp.count < cmd.ne) {
1388                 ret = ib_poll_cq(cq, 1, &wc);
1389                 if (ret < 0)
1390                         goto out_put;
1391                 if (!ret)
1392                         break;
1393
1394                 ret = copy_wc_to_user(data_ptr, &wc);
1395                 if (ret)
1396                         goto out_put;
1397
1398                 data_ptr += sizeof(struct ib_uverbs_wc);
1399                 ++resp.count;
1400         }
1401
1402         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1403                 ret = -EFAULT;
1404                 goto out_put;
1405         }
1406
1407         ret = in_len;
1408
1409 out_put:
1410         put_cq_read(cq);
1411         return ret;
1412 }
1413
1414 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1415                                 const char __user *buf, int in_len,
1416                                 int out_len)
1417 {
1418         struct ib_uverbs_req_notify_cq cmd;
1419         struct ib_cq                  *cq;
1420
1421         if (copy_from_user(&cmd, buf, sizeof cmd))
1422                 return -EFAULT;
1423
1424         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1425         if (!cq)
1426                 return -EINVAL;
1427
1428         ib_req_notify_cq(cq, cmd.solicited_only ?
1429                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1430
1431         put_cq_read(cq);
1432
1433         return in_len;
1434 }
1435
1436 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1437                              const char __user *buf, int in_len,
1438                              int out_len)
1439 {
1440         struct ib_uverbs_destroy_cq      cmd;
1441         struct ib_uverbs_destroy_cq_resp resp;
1442         struct ib_uobject               *uobj;
1443         struct ib_cq                    *cq;
1444         struct ib_ucq_object            *obj;
1445         struct ib_uverbs_event_file     *ev_file;
1446         int                              ret = -EINVAL;
1447
1448         if (copy_from_user(&cmd, buf, sizeof cmd))
1449                 return -EFAULT;
1450
1451         uobj = idr_write_uobj(&ib_uverbs_cq_idr, cmd.cq_handle, file->ucontext);
1452         if (!uobj)
1453                 return -EINVAL;
1454         cq      = uobj->object;
1455         ev_file = cq->cq_context;
1456         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1457
1458         ret = ib_destroy_cq(cq);
1459         if (!ret)
1460                 uobj->live = 0;
1461
1462         put_uobj_write(uobj);
1463
1464         if (ret)
1465                 return ret;
1466
1467         idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
1468
1469         mutex_lock(&file->mutex);
1470         list_del(&uobj->list);
1471         mutex_unlock(&file->mutex);
1472
1473         ib_uverbs_release_ucq(file, ev_file, obj);
1474
1475         memset(&resp, 0, sizeof resp);
1476         resp.comp_events_reported  = obj->comp_events_reported;
1477         resp.async_events_reported = obj->async_events_reported;
1478
1479         put_uobj(uobj);
1480
1481         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1482                          &resp, sizeof resp))
1483                 return -EFAULT;
1484
1485         return in_len;
1486 }
1487
1488 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1489                             const char __user *buf, int in_len,
1490                             int out_len)
1491 {
1492         struct ib_uverbs_create_qp      cmd;
1493         struct ib_uverbs_create_qp_resp resp;
1494         struct ib_udata                 udata;
1495         struct ib_uqp_object           *obj;
1496         struct ib_device               *device;
1497         struct ib_pd                   *pd = NULL;
1498         struct ib_xrcd                 *xrcd = NULL;
1499         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1500         struct ib_cq                   *scq = NULL, *rcq = NULL;
1501         struct ib_srq                  *srq = NULL;
1502         struct ib_qp                   *qp;
1503         struct ib_qp_init_attr          attr;
1504         int ret;
1505
1506         if (out_len < sizeof resp)
1507                 return -ENOSPC;
1508
1509         if (copy_from_user(&cmd, buf, sizeof cmd))
1510                 return -EFAULT;
1511
1512         if (cmd.qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1513                 return -EPERM;
1514
1515         INIT_UDATA(&udata, buf + sizeof cmd,
1516                    (unsigned long) cmd.response + sizeof resp,
1517                    in_len - sizeof cmd, out_len - sizeof resp);
1518
1519         obj = kzalloc(sizeof *obj, GFP_KERNEL);
1520         if (!obj)
1521                 return -ENOMEM;
1522
1523         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1524         down_write(&obj->uevent.uobject.mutex);
1525
1526         if (cmd.qp_type == IB_QPT_XRC_TGT) {
1527                 xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1528                 if (!xrcd) {
1529                         ret = -EINVAL;
1530                         goto err_put;
1531                 }
1532                 device = xrcd->device;
1533         } else {
1534                 if (cmd.qp_type == IB_QPT_XRC_INI) {
1535                         cmd.max_recv_wr = cmd.max_recv_sge = 0;
1536                 } else {
1537                         if (cmd.is_srq) {
1538                                 srq = idr_read_srq(cmd.srq_handle, file->ucontext);
1539                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1540                                         ret = -EINVAL;
1541                                         goto err_put;
1542                                 }
1543                         }
1544
1545                         if (cmd.recv_cq_handle != cmd.send_cq_handle) {
1546                                 rcq = idr_read_cq(cmd.recv_cq_handle, file->ucontext, 0);
1547                                 if (!rcq) {
1548                                         ret = -EINVAL;
1549                                         goto err_put;
1550                                 }
1551                         }
1552                 }
1553
1554                 scq = idr_read_cq(cmd.send_cq_handle, file->ucontext, !!rcq);
1555                 rcq = rcq ?: scq;
1556                 pd  = idr_read_pd(cmd.pd_handle, file->ucontext);
1557                 if (!pd || !scq) {
1558                         ret = -EINVAL;
1559                         goto err_put;
1560                 }
1561
1562                 device = pd->device;
1563         }
1564
1565         attr.event_handler = ib_uverbs_qp_event_handler;
1566         attr.qp_context    = file;
1567         attr.send_cq       = scq;
1568         attr.recv_cq       = rcq;
1569         attr.srq           = srq;
1570         attr.xrcd          = xrcd;
1571         attr.sq_sig_type   = cmd.sq_sig_all ? IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
1572         attr.qp_type       = cmd.qp_type;
1573         attr.create_flags  = 0;
1574
1575         attr.cap.max_send_wr     = cmd.max_send_wr;
1576         attr.cap.max_recv_wr     = cmd.max_recv_wr;
1577         attr.cap.max_send_sge    = cmd.max_send_sge;
1578         attr.cap.max_recv_sge    = cmd.max_recv_sge;
1579         attr.cap.max_inline_data = cmd.max_inline_data;
1580
1581         obj->uevent.events_reported     = 0;
1582         INIT_LIST_HEAD(&obj->uevent.event_list);
1583         INIT_LIST_HEAD(&obj->mcast_list);
1584
1585         if (cmd.qp_type == IB_QPT_XRC_TGT)
1586                 qp = ib_create_qp(pd, &attr);
1587         else
1588                 qp = device->create_qp(pd, &attr, &udata);
1589
1590         if (IS_ERR(qp)) {
1591                 ret = PTR_ERR(qp);
1592                 goto err_put;
1593         }
1594
1595         if (cmd.qp_type != IB_QPT_XRC_TGT) {
1596                 qp->real_qp       = qp;
1597                 qp->device        = device;
1598                 qp->pd            = pd;
1599                 qp->send_cq       = attr.send_cq;
1600                 qp->recv_cq       = attr.recv_cq;
1601                 qp->srq           = attr.srq;
1602                 qp->event_handler = attr.event_handler;
1603                 qp->qp_context    = attr.qp_context;
1604                 qp->qp_type       = attr.qp_type;
1605                 atomic_set(&qp->usecnt, 0);
1606                 atomic_inc(&pd->usecnt);
1607                 atomic_inc(&attr.send_cq->usecnt);
1608                 if (attr.recv_cq)
1609                         atomic_inc(&attr.recv_cq->usecnt);
1610                 if (attr.srq)
1611                         atomic_inc(&attr.srq->usecnt);
1612         }
1613         qp->uobject = &obj->uevent.uobject;
1614
1615         obj->uevent.uobject.object = qp;
1616         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1617         if (ret)
1618                 goto err_destroy;
1619
1620         memset(&resp, 0, sizeof resp);
1621         resp.qpn             = qp->qp_num;
1622         resp.qp_handle       = obj->uevent.uobject.id;
1623         resp.max_recv_sge    = attr.cap.max_recv_sge;
1624         resp.max_send_sge    = attr.cap.max_send_sge;
1625         resp.max_recv_wr     = attr.cap.max_recv_wr;
1626         resp.max_send_wr     = attr.cap.max_send_wr;
1627         resp.max_inline_data = attr.cap.max_inline_data;
1628
1629         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1630                          &resp, sizeof resp)) {
1631                 ret = -EFAULT;
1632                 goto err_copy;
1633         }
1634
1635         if (xrcd) {
1636                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object,
1637                                           uobject);
1638                 atomic_inc(&obj->uxrcd->refcnt);
1639                 put_xrcd_read(xrcd_uobj);
1640         }
1641
1642         if (pd)
1643                 put_pd_read(pd);
1644         if (scq)
1645                 put_cq_read(scq);
1646         if (rcq && rcq != scq)
1647                 put_cq_read(rcq);
1648         if (srq)
1649                 put_srq_read(srq);
1650
1651         mutex_lock(&file->mutex);
1652         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1653         mutex_unlock(&file->mutex);
1654
1655         obj->uevent.uobject.live = 1;
1656
1657         up_write(&obj->uevent.uobject.mutex);
1658
1659         return in_len;
1660
1661 err_copy:
1662         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1663
1664 err_destroy:
1665         ib_destroy_qp(qp);
1666
1667 err_put:
1668         if (xrcd)
1669                 put_xrcd_read(xrcd_uobj);
1670         if (pd)
1671                 put_pd_read(pd);
1672         if (scq)
1673                 put_cq_read(scq);
1674         if (rcq && rcq != scq)
1675                 put_cq_read(rcq);
1676         if (srq)
1677                 put_srq_read(srq);
1678
1679         put_uobj_write(&obj->uevent.uobject);
1680         return ret;
1681 }
1682
1683 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1684                           const char __user *buf, int in_len, int out_len)
1685 {
1686         struct ib_uverbs_open_qp        cmd;
1687         struct ib_uverbs_create_qp_resp resp;
1688         struct ib_udata                 udata;
1689         struct ib_uqp_object           *obj;
1690         struct ib_xrcd                 *xrcd;
1691         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1692         struct ib_qp                   *qp;
1693         struct ib_qp_open_attr          attr;
1694         int ret;
1695
1696         if (out_len < sizeof resp)
1697                 return -ENOSPC;
1698
1699         if (copy_from_user(&cmd, buf, sizeof cmd))
1700                 return -EFAULT;
1701
1702         INIT_UDATA(&udata, buf + sizeof cmd,
1703                    (unsigned long) cmd.response + sizeof resp,
1704                    in_len - sizeof cmd, out_len - sizeof resp);
1705
1706         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1707         if (!obj)
1708                 return -ENOMEM;
1709
1710         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1711         down_write(&obj->uevent.uobject.mutex);
1712
1713         xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1714         if (!xrcd) {
1715                 ret = -EINVAL;
1716                 goto err_put;
1717         }
1718
1719         attr.event_handler = ib_uverbs_qp_event_handler;
1720         attr.qp_context    = file;
1721         attr.qp_num        = cmd.qpn;
1722         attr.qp_type       = cmd.qp_type;
1723
1724         obj->uevent.events_reported = 0;
1725         INIT_LIST_HEAD(&obj->uevent.event_list);
1726         INIT_LIST_HEAD(&obj->mcast_list);
1727
1728         qp = ib_open_qp(xrcd, &attr);
1729         if (IS_ERR(qp)) {
1730                 ret = PTR_ERR(qp);
1731                 goto err_put;
1732         }
1733
1734         qp->uobject = &obj->uevent.uobject;
1735
1736         obj->uevent.uobject.object = qp;
1737         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1738         if (ret)
1739                 goto err_destroy;
1740
1741         memset(&resp, 0, sizeof resp);
1742         resp.qpn       = qp->qp_num;
1743         resp.qp_handle = obj->uevent.uobject.id;
1744
1745         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1746                          &resp, sizeof resp)) {
1747                 ret = -EFAULT;
1748                 goto err_remove;
1749         }
1750
1751         obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
1752         atomic_inc(&obj->uxrcd->refcnt);
1753         put_xrcd_read(xrcd_uobj);
1754
1755         mutex_lock(&file->mutex);
1756         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1757         mutex_unlock(&file->mutex);
1758
1759         obj->uevent.uobject.live = 1;
1760
1761         up_write(&obj->uevent.uobject.mutex);
1762
1763         return in_len;
1764
1765 err_remove:
1766         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1767
1768 err_destroy:
1769         ib_destroy_qp(qp);
1770
1771 err_put:
1772         put_xrcd_read(xrcd_uobj);
1773         put_uobj_write(&obj->uevent.uobject);
1774         return ret;
1775 }
1776
1777 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1778                            const char __user *buf, int in_len,
1779                            int out_len)
1780 {
1781         struct ib_uverbs_query_qp      cmd;
1782         struct ib_uverbs_query_qp_resp resp;
1783         struct ib_qp                   *qp;
1784         struct ib_qp_attr              *attr;
1785         struct ib_qp_init_attr         *init_attr;
1786         int                            ret;
1787
1788         if (copy_from_user(&cmd, buf, sizeof cmd))
1789                 return -EFAULT;
1790
1791         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1792         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1793         if (!attr || !init_attr) {
1794                 ret = -ENOMEM;
1795                 goto out;
1796         }
1797
1798         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1799         if (!qp) {
1800                 ret = -EINVAL;
1801                 goto out;
1802         }
1803
1804         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1805
1806         put_qp_read(qp);
1807
1808         if (ret)
1809                 goto out;
1810
1811         memset(&resp, 0, sizeof resp);
1812
1813         resp.qp_state               = attr->qp_state;
1814         resp.cur_qp_state           = attr->cur_qp_state;
1815         resp.path_mtu               = attr->path_mtu;
1816         resp.path_mig_state         = attr->path_mig_state;
1817         resp.qkey                   = attr->qkey;
1818         resp.rq_psn                 = attr->rq_psn;
1819         resp.sq_psn                 = attr->sq_psn;
1820         resp.dest_qp_num            = attr->dest_qp_num;
1821         resp.qp_access_flags        = attr->qp_access_flags;
1822         resp.pkey_index             = attr->pkey_index;
1823         resp.alt_pkey_index         = attr->alt_pkey_index;
1824         resp.sq_draining            = attr->sq_draining;
1825         resp.max_rd_atomic          = attr->max_rd_atomic;
1826         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1827         resp.min_rnr_timer          = attr->min_rnr_timer;
1828         resp.port_num               = attr->port_num;
1829         resp.timeout                = attr->timeout;
1830         resp.retry_cnt              = attr->retry_cnt;
1831         resp.rnr_retry              = attr->rnr_retry;
1832         resp.alt_port_num           = attr->alt_port_num;
1833         resp.alt_timeout            = attr->alt_timeout;
1834
1835         memcpy(resp.dest.dgid, attr->ah_attr.grh.dgid.raw, 16);
1836         resp.dest.flow_label        = attr->ah_attr.grh.flow_label;
1837         resp.dest.sgid_index        = attr->ah_attr.grh.sgid_index;
1838         resp.dest.hop_limit         = attr->ah_attr.grh.hop_limit;
1839         resp.dest.traffic_class     = attr->ah_attr.grh.traffic_class;
1840         resp.dest.dlid              = attr->ah_attr.dlid;
1841         resp.dest.sl                = attr->ah_attr.sl;
1842         resp.dest.src_path_bits     = attr->ah_attr.src_path_bits;
1843         resp.dest.static_rate       = attr->ah_attr.static_rate;
1844         resp.dest.is_global         = !!(attr->ah_attr.ah_flags & IB_AH_GRH);
1845         resp.dest.port_num          = attr->ah_attr.port_num;
1846
1847         memcpy(resp.alt_dest.dgid, attr->alt_ah_attr.grh.dgid.raw, 16);
1848         resp.alt_dest.flow_label    = attr->alt_ah_attr.grh.flow_label;
1849         resp.alt_dest.sgid_index    = attr->alt_ah_attr.grh.sgid_index;
1850         resp.alt_dest.hop_limit     = attr->alt_ah_attr.grh.hop_limit;
1851         resp.alt_dest.traffic_class = attr->alt_ah_attr.grh.traffic_class;
1852         resp.alt_dest.dlid          = attr->alt_ah_attr.dlid;
1853         resp.alt_dest.sl            = attr->alt_ah_attr.sl;
1854         resp.alt_dest.src_path_bits = attr->alt_ah_attr.src_path_bits;
1855         resp.alt_dest.static_rate   = attr->alt_ah_attr.static_rate;
1856         resp.alt_dest.is_global     = !!(attr->alt_ah_attr.ah_flags & IB_AH_GRH);
1857         resp.alt_dest.port_num      = attr->alt_ah_attr.port_num;
1858
1859         resp.max_send_wr            = init_attr->cap.max_send_wr;
1860         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1861         resp.max_send_sge           = init_attr->cap.max_send_sge;
1862         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1863         resp.max_inline_data        = init_attr->cap.max_inline_data;
1864         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1865
1866         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1867                          &resp, sizeof resp))
1868                 ret = -EFAULT;
1869
1870 out:
1871         kfree(attr);
1872         kfree(init_attr);
1873
1874         return ret ? ret : in_len;
1875 }
1876
1877 /* Remove ignored fields set in the attribute mask */
1878 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1879 {
1880         switch (qp_type) {
1881         case IB_QPT_XRC_INI:
1882                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1883         case IB_QPT_XRC_TGT:
1884                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1885                                 IB_QP_RNR_RETRY);
1886         default:
1887                 return mask;
1888         }
1889 }
1890
1891 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
1892                             const char __user *buf, int in_len,
1893                             int out_len)
1894 {
1895         struct ib_uverbs_modify_qp cmd;
1896         struct ib_udata            udata;
1897         struct ib_qp              *qp;
1898         struct ib_qp_attr         *attr;
1899         int                        ret;
1900
1901         if (copy_from_user(&cmd, buf, sizeof cmd))
1902                 return -EFAULT;
1903
1904         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
1905                    out_len);
1906
1907         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1908         if (!attr)
1909                 return -ENOMEM;
1910
1911         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1912         if (!qp) {
1913                 ret = -EINVAL;
1914                 goto out;
1915         }
1916
1917         attr->qp_state            = cmd.qp_state;
1918         attr->cur_qp_state        = cmd.cur_qp_state;
1919         attr->path_mtu            = cmd.path_mtu;
1920         attr->path_mig_state      = cmd.path_mig_state;
1921         attr->qkey                = cmd.qkey;
1922         attr->rq_psn              = cmd.rq_psn;
1923         attr->sq_psn              = cmd.sq_psn;
1924         attr->dest_qp_num         = cmd.dest_qp_num;
1925         attr->qp_access_flags     = cmd.qp_access_flags;
1926         attr->pkey_index          = cmd.pkey_index;
1927         attr->alt_pkey_index      = cmd.alt_pkey_index;
1928         attr->en_sqd_async_notify = cmd.en_sqd_async_notify;
1929         attr->max_rd_atomic       = cmd.max_rd_atomic;
1930         attr->max_dest_rd_atomic  = cmd.max_dest_rd_atomic;
1931         attr->min_rnr_timer       = cmd.min_rnr_timer;
1932         attr->port_num            = cmd.port_num;
1933         attr->timeout             = cmd.timeout;
1934         attr->retry_cnt           = cmd.retry_cnt;
1935         attr->rnr_retry           = cmd.rnr_retry;
1936         attr->alt_port_num        = cmd.alt_port_num;
1937         attr->alt_timeout         = cmd.alt_timeout;
1938
1939         memcpy(attr->ah_attr.grh.dgid.raw, cmd.dest.dgid, 16);
1940         attr->ah_attr.grh.flow_label        = cmd.dest.flow_label;
1941         attr->ah_attr.grh.sgid_index        = cmd.dest.sgid_index;
1942         attr->ah_attr.grh.hop_limit         = cmd.dest.hop_limit;
1943         attr->ah_attr.grh.traffic_class     = cmd.dest.traffic_class;
1944         attr->ah_attr.dlid                  = cmd.dest.dlid;
1945         attr->ah_attr.sl                    = cmd.dest.sl;
1946         attr->ah_attr.src_path_bits         = cmd.dest.src_path_bits;
1947         attr->ah_attr.static_rate           = cmd.dest.static_rate;
1948         attr->ah_attr.ah_flags              = cmd.dest.is_global ? IB_AH_GRH : 0;
1949         attr->ah_attr.port_num              = cmd.dest.port_num;
1950
1951         memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd.alt_dest.dgid, 16);
1952         attr->alt_ah_attr.grh.flow_label    = cmd.alt_dest.flow_label;
1953         attr->alt_ah_attr.grh.sgid_index    = cmd.alt_dest.sgid_index;
1954         attr->alt_ah_attr.grh.hop_limit     = cmd.alt_dest.hop_limit;
1955         attr->alt_ah_attr.grh.traffic_class = cmd.alt_dest.traffic_class;
1956         attr->alt_ah_attr.dlid              = cmd.alt_dest.dlid;
1957         attr->alt_ah_attr.sl                = cmd.alt_dest.sl;
1958         attr->alt_ah_attr.src_path_bits     = cmd.alt_dest.src_path_bits;
1959         attr->alt_ah_attr.static_rate       = cmd.alt_dest.static_rate;
1960         attr->alt_ah_attr.ah_flags          = cmd.alt_dest.is_global ? IB_AH_GRH : 0;
1961         attr->alt_ah_attr.port_num          = cmd.alt_dest.port_num;
1962
1963         if (qp->real_qp == qp) {
1964                 ret = qp->device->modify_qp(qp, attr,
1965                         modify_qp_mask(qp->qp_type, cmd.attr_mask), &udata);
1966         } else {
1967                 ret = ib_modify_qp(qp, attr, modify_qp_mask(qp->qp_type, cmd.attr_mask));
1968         }
1969
1970         put_qp_read(qp);
1971
1972         if (ret)
1973                 goto out;
1974
1975         ret = in_len;
1976
1977 out:
1978         kfree(attr);
1979
1980         return ret;
1981 }
1982
1983 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
1984                              const char __user *buf, int in_len,
1985                              int out_len)
1986 {
1987         struct ib_uverbs_destroy_qp      cmd;
1988         struct ib_uverbs_destroy_qp_resp resp;
1989         struct ib_uobject               *uobj;
1990         struct ib_qp                    *qp;
1991         struct ib_uqp_object            *obj;
1992         int                              ret = -EINVAL;
1993
1994         if (copy_from_user(&cmd, buf, sizeof cmd))
1995                 return -EFAULT;
1996
1997         memset(&resp, 0, sizeof resp);
1998
1999         uobj = idr_write_uobj(&ib_uverbs_qp_idr, cmd.qp_handle, file->ucontext);
2000         if (!uobj)
2001                 return -EINVAL;
2002         qp  = uobj->object;
2003         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2004
2005         if (!list_empty(&obj->mcast_list)) {
2006                 put_uobj_write(uobj);
2007                 return -EBUSY;
2008         }
2009
2010         ret = ib_destroy_qp(qp);
2011         if (!ret)
2012                 uobj->live = 0;
2013
2014         put_uobj_write(uobj);
2015
2016         if (ret)
2017                 return ret;
2018
2019         if (obj->uxrcd)
2020                 atomic_dec(&obj->uxrcd->refcnt);
2021
2022         idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
2023
2024         mutex_lock(&file->mutex);
2025         list_del(&uobj->list);
2026         mutex_unlock(&file->mutex);
2027
2028         ib_uverbs_release_uevent(file, &obj->uevent);
2029
2030         resp.events_reported = obj->uevent.events_reported;
2031
2032         put_uobj(uobj);
2033
2034         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2035                          &resp, sizeof resp))
2036                 return -EFAULT;
2037
2038         return in_len;
2039 }
2040
2041 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2042                             const char __user *buf, int in_len,
2043                             int out_len)
2044 {
2045         struct ib_uverbs_post_send      cmd;
2046         struct ib_uverbs_post_send_resp resp;
2047         struct ib_uverbs_send_wr       *user_wr;
2048         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
2049         struct ib_qp                   *qp;
2050         int                             i, sg_ind;
2051         int                             is_ud;
2052         ssize_t                         ret = -EINVAL;
2053
2054         if (copy_from_user(&cmd, buf, sizeof cmd))
2055                 return -EFAULT;
2056
2057         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2058             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2059                 return -EINVAL;
2060
2061         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2062                 return -EINVAL;
2063
2064         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2065         if (!user_wr)
2066                 return -ENOMEM;
2067
2068         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2069         if (!qp)
2070                 goto out;
2071
2072         is_ud = qp->qp_type == IB_QPT_UD;
2073         sg_ind = 0;
2074         last = NULL;
2075         for (i = 0; i < cmd.wr_count; ++i) {
2076                 if (copy_from_user(user_wr,
2077                                    buf + sizeof cmd + i * cmd.wqe_size,
2078                                    cmd.wqe_size)) {
2079                         ret = -EFAULT;
2080                         goto out_put;
2081                 }
2082
2083                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2084                         ret = -EINVAL;
2085                         goto out_put;
2086                 }
2087
2088                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2089                                user_wr->num_sge * sizeof (struct ib_sge),
2090                                GFP_KERNEL);
2091                 if (!next) {
2092                         ret = -ENOMEM;
2093                         goto out_put;
2094                 }
2095
2096                 if (!last)
2097                         wr = next;
2098                 else
2099                         last->next = next;
2100                 last = next;
2101
2102                 next->next       = NULL;
2103                 next->wr_id      = user_wr->wr_id;
2104                 next->num_sge    = user_wr->num_sge;
2105                 next->opcode     = user_wr->opcode;
2106                 next->send_flags = user_wr->send_flags;
2107
2108                 if (is_ud) {
2109                         next->wr.ud.ah = idr_read_ah(user_wr->wr.ud.ah,
2110                                                      file->ucontext);
2111                         if (!next->wr.ud.ah) {
2112                                 ret = -EINVAL;
2113                                 goto out_put;
2114                         }
2115                         next->wr.ud.remote_qpn  = user_wr->wr.ud.remote_qpn;
2116                         next->wr.ud.remote_qkey = user_wr->wr.ud.remote_qkey;
2117                         if (next->opcode == IB_WR_SEND_WITH_IMM)
2118                                 next->ex.imm_data =
2119                                         (__be32 __force) user_wr->ex.imm_data;
2120                 } else {
2121                         switch (next->opcode) {
2122                         case IB_WR_RDMA_WRITE_WITH_IMM:
2123                                 next->ex.imm_data =
2124                                         (__be32 __force) user_wr->ex.imm_data;
2125                         case IB_WR_RDMA_WRITE:
2126                         case IB_WR_RDMA_READ:
2127                                 next->wr.rdma.remote_addr =
2128                                         user_wr->wr.rdma.remote_addr;
2129                                 next->wr.rdma.rkey        =
2130                                         user_wr->wr.rdma.rkey;
2131                                 break;
2132                         case IB_WR_SEND_WITH_IMM:
2133                                 next->ex.imm_data =
2134                                         (__be32 __force) user_wr->ex.imm_data;
2135                                 break;
2136                         case IB_WR_SEND_WITH_INV:
2137                                 next->ex.invalidate_rkey =
2138                                         user_wr->ex.invalidate_rkey;
2139                                 break;
2140                         case IB_WR_ATOMIC_CMP_AND_SWP:
2141                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2142                                 next->wr.atomic.remote_addr =
2143                                         user_wr->wr.atomic.remote_addr;
2144                                 next->wr.atomic.compare_add =
2145                                         user_wr->wr.atomic.compare_add;
2146                                 next->wr.atomic.swap = user_wr->wr.atomic.swap;
2147                                 next->wr.atomic.rkey = user_wr->wr.atomic.rkey;
2148                                 break;
2149                         default:
2150                                 break;
2151                         }
2152                 }
2153
2154                 if (next->num_sge) {
2155                         next->sg_list = (void *) next +
2156                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2157                         if (copy_from_user(next->sg_list,
2158                                            buf + sizeof cmd +
2159                                            cmd.wr_count * cmd.wqe_size +
2160                                            sg_ind * sizeof (struct ib_sge),
2161                                            next->num_sge * sizeof (struct ib_sge))) {
2162                                 ret = -EFAULT;
2163                                 goto out_put;
2164                         }
2165                         sg_ind += next->num_sge;
2166                 } else
2167                         next->sg_list = NULL;
2168         }
2169
2170         resp.bad_wr = 0;
2171         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2172         if (ret)
2173                 for (next = wr; next; next = next->next) {
2174                         ++resp.bad_wr;
2175                         if (next == bad_wr)
2176                                 break;
2177                 }
2178
2179         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2180                          &resp, sizeof resp))
2181                 ret = -EFAULT;
2182
2183 out_put:
2184         put_qp_read(qp);
2185
2186         while (wr) {
2187                 if (is_ud && wr->wr.ud.ah)
2188                         put_ah_read(wr->wr.ud.ah);
2189                 next = wr->next;
2190                 kfree(wr);
2191                 wr = next;
2192         }
2193
2194 out:
2195         kfree(user_wr);
2196
2197         return ret ? ret : in_len;
2198 }
2199
2200 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2201                                                     int in_len,
2202                                                     u32 wr_count,
2203                                                     u32 sge_count,
2204                                                     u32 wqe_size)
2205 {
2206         struct ib_uverbs_recv_wr *user_wr;
2207         struct ib_recv_wr        *wr = NULL, *last, *next;
2208         int                       sg_ind;
2209         int                       i;
2210         int                       ret;
2211
2212         if (in_len < wqe_size * wr_count +
2213             sge_count * sizeof (struct ib_uverbs_sge))
2214                 return ERR_PTR(-EINVAL);
2215
2216         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2217                 return ERR_PTR(-EINVAL);
2218
2219         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2220         if (!user_wr)
2221                 return ERR_PTR(-ENOMEM);
2222
2223         sg_ind = 0;
2224         last = NULL;
2225         for (i = 0; i < wr_count; ++i) {
2226                 if (copy_from_user(user_wr, buf + i * wqe_size,
2227                                    wqe_size)) {
2228                         ret = -EFAULT;
2229                         goto err;
2230                 }
2231
2232                 if (user_wr->num_sge + sg_ind > sge_count) {
2233                         ret = -EINVAL;
2234                         goto err;
2235                 }
2236
2237                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2238                                user_wr->num_sge * sizeof (struct ib_sge),
2239                                GFP_KERNEL);
2240                 if (!next) {
2241                         ret = -ENOMEM;
2242                         goto err;
2243                 }
2244
2245                 if (!last)
2246                         wr = next;
2247                 else
2248                         last->next = next;
2249                 last = next;
2250
2251                 next->next       = NULL;
2252                 next->wr_id      = user_wr->wr_id;
2253                 next->num_sge    = user_wr->num_sge;
2254
2255                 if (next->num_sge) {
2256                         next->sg_list = (void *) next +
2257                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2258                         if (copy_from_user(next->sg_list,
2259                                            buf + wr_count * wqe_size +
2260                                            sg_ind * sizeof (struct ib_sge),
2261                                            next->num_sge * sizeof (struct ib_sge))) {
2262                                 ret = -EFAULT;
2263                                 goto err;
2264                         }
2265                         sg_ind += next->num_sge;
2266                 } else
2267                         next->sg_list = NULL;
2268         }
2269
2270         kfree(user_wr);
2271         return wr;
2272
2273 err:
2274         kfree(user_wr);
2275
2276         while (wr) {
2277                 next = wr->next;
2278                 kfree(wr);
2279                 wr = next;
2280         }
2281
2282         return ERR_PTR(ret);
2283 }
2284
2285 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2286                             const char __user *buf, int in_len,
2287                             int out_len)
2288 {
2289         struct ib_uverbs_post_recv      cmd;
2290         struct ib_uverbs_post_recv_resp resp;
2291         struct ib_recv_wr              *wr, *next, *bad_wr;
2292         struct ib_qp                   *qp;
2293         ssize_t                         ret = -EINVAL;
2294
2295         if (copy_from_user(&cmd, buf, sizeof cmd))
2296                 return -EFAULT;
2297
2298         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2299                                        in_len - sizeof cmd, cmd.wr_count,
2300                                        cmd.sge_count, cmd.wqe_size);
2301         if (IS_ERR(wr))
2302                 return PTR_ERR(wr);
2303
2304         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2305         if (!qp)
2306                 goto out;
2307
2308         resp.bad_wr = 0;
2309         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2310
2311         put_qp_read(qp);
2312
2313         if (ret)
2314                 for (next = wr; next; next = next->next) {
2315                         ++resp.bad_wr;
2316                         if (next == bad_wr)
2317                                 break;
2318                 }
2319
2320         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2321                          &resp, sizeof resp))
2322                 ret = -EFAULT;
2323
2324 out:
2325         while (wr) {
2326                 next = wr->next;
2327                 kfree(wr);
2328                 wr = next;
2329         }
2330
2331         return ret ? ret : in_len;
2332 }
2333
2334 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2335                                 const char __user *buf, int in_len,
2336                                 int out_len)
2337 {
2338         struct ib_uverbs_post_srq_recv      cmd;
2339         struct ib_uverbs_post_srq_recv_resp resp;
2340         struct ib_recv_wr                  *wr, *next, *bad_wr;
2341         struct ib_srq                      *srq;
2342         ssize_t                             ret = -EINVAL;
2343
2344         if (copy_from_user(&cmd, buf, sizeof cmd))
2345                 return -EFAULT;
2346
2347         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2348                                        in_len - sizeof cmd, cmd.wr_count,
2349                                        cmd.sge_count, cmd.wqe_size);
2350         if (IS_ERR(wr))
2351                 return PTR_ERR(wr);
2352
2353         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2354         if (!srq)
2355                 goto out;
2356
2357         resp.bad_wr = 0;
2358         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2359
2360         put_srq_read(srq);
2361
2362         if (ret)
2363                 for (next = wr; next; next = next->next) {
2364                         ++resp.bad_wr;
2365                         if (next == bad_wr)
2366                                 break;
2367                 }
2368
2369         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2370                          &resp, sizeof resp))
2371                 ret = -EFAULT;
2372
2373 out:
2374         while (wr) {
2375                 next = wr->next;
2376                 kfree(wr);
2377                 wr = next;
2378         }
2379
2380         return ret ? ret : in_len;
2381 }
2382
2383 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2384                             const char __user *buf, int in_len,
2385                             int out_len)
2386 {
2387         struct ib_uverbs_create_ah       cmd;
2388         struct ib_uverbs_create_ah_resp  resp;
2389         struct ib_uobject               *uobj;
2390         struct ib_pd                    *pd;
2391         struct ib_ah                    *ah;
2392         struct ib_ah_attr               attr;
2393         int ret;
2394
2395         if (out_len < sizeof resp)
2396                 return -ENOSPC;
2397
2398         if (copy_from_user(&cmd, buf, sizeof cmd))
2399                 return -EFAULT;
2400
2401         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
2402         if (!uobj)
2403                 return -ENOMEM;
2404
2405         init_uobj(uobj, cmd.user_handle, file->ucontext, &ah_lock_class);
2406         down_write(&uobj->mutex);
2407
2408         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
2409         if (!pd) {
2410                 ret = -EINVAL;
2411                 goto err;
2412         }
2413
2414         attr.dlid              = cmd.attr.dlid;
2415         attr.sl                = cmd.attr.sl;
2416         attr.src_path_bits     = cmd.attr.src_path_bits;
2417         attr.static_rate       = cmd.attr.static_rate;
2418         attr.ah_flags          = cmd.attr.is_global ? IB_AH_GRH : 0;
2419         attr.port_num          = cmd.attr.port_num;
2420         attr.grh.flow_label    = cmd.attr.grh.flow_label;
2421         attr.grh.sgid_index    = cmd.attr.grh.sgid_index;
2422         attr.grh.hop_limit     = cmd.attr.grh.hop_limit;
2423         attr.grh.traffic_class = cmd.attr.grh.traffic_class;
2424         memcpy(attr.grh.dgid.raw, cmd.attr.grh.dgid, 16);
2425
2426         ah = ib_create_ah(pd, &attr);
2427         if (IS_ERR(ah)) {
2428                 ret = PTR_ERR(ah);
2429                 goto err_put;
2430         }
2431
2432         ah->uobject  = uobj;
2433         uobj->object = ah;
2434
2435         ret = idr_add_uobj(&ib_uverbs_ah_idr, uobj);
2436         if (ret)
2437                 goto err_destroy;
2438
2439         resp.ah_handle = uobj->id;
2440
2441         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2442                          &resp, sizeof resp)) {
2443                 ret = -EFAULT;
2444                 goto err_copy;
2445         }
2446
2447         put_pd_read(pd);
2448
2449         mutex_lock(&file->mutex);
2450         list_add_tail(&uobj->list, &file->ucontext->ah_list);
2451         mutex_unlock(&file->mutex);
2452
2453         uobj->live = 1;
2454
2455         up_write(&uobj->mutex);
2456
2457         return in_len;
2458
2459 err_copy:
2460         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2461
2462 err_destroy:
2463         ib_destroy_ah(ah);
2464
2465 err_put:
2466         put_pd_read(pd);
2467
2468 err:
2469         put_uobj_write(uobj);
2470         return ret;
2471 }
2472
2473 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2474                              const char __user *buf, int in_len, int out_len)
2475 {
2476         struct ib_uverbs_destroy_ah cmd;
2477         struct ib_ah               *ah;
2478         struct ib_uobject          *uobj;
2479         int                         ret;
2480
2481         if (copy_from_user(&cmd, buf, sizeof cmd))
2482                 return -EFAULT;
2483
2484         uobj = idr_write_uobj(&ib_uverbs_ah_idr, cmd.ah_handle, file->ucontext);
2485         if (!uobj)
2486                 return -EINVAL;
2487         ah = uobj->object;
2488
2489         ret = ib_destroy_ah(ah);
2490         if (!ret)
2491                 uobj->live = 0;
2492
2493         put_uobj_write(uobj);
2494
2495         if (ret)
2496                 return ret;
2497
2498         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2499
2500         mutex_lock(&file->mutex);
2501         list_del(&uobj->list);
2502         mutex_unlock(&file->mutex);
2503
2504         put_uobj(uobj);
2505
2506         return in_len;
2507 }
2508
2509 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2510                                const char __user *buf, int in_len,
2511                                int out_len)
2512 {
2513         struct ib_uverbs_attach_mcast cmd;
2514         struct ib_qp                 *qp;
2515         struct ib_uqp_object         *obj;
2516         struct ib_uverbs_mcast_entry *mcast;
2517         int                           ret;
2518
2519         if (copy_from_user(&cmd, buf, sizeof cmd))
2520                 return -EFAULT;
2521
2522         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2523         if (!qp)
2524                 return -EINVAL;
2525
2526         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2527
2528         list_for_each_entry(mcast, &obj->mcast_list, list)
2529                 if (cmd.mlid == mcast->lid &&
2530                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2531                         ret = 0;
2532                         goto out_put;
2533                 }
2534
2535         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2536         if (!mcast) {
2537                 ret = -ENOMEM;
2538                 goto out_put;
2539         }
2540
2541         mcast->lid = cmd.mlid;
2542         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2543
2544         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2545         if (!ret)
2546                 list_add_tail(&mcast->list, &obj->mcast_list);
2547         else
2548                 kfree(mcast);
2549
2550 out_put:
2551         put_qp_write(qp);
2552
2553         return ret ? ret : in_len;
2554 }
2555
2556 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2557                                const char __user *buf, int in_len,
2558                                int out_len)
2559 {
2560         struct ib_uverbs_detach_mcast cmd;
2561         struct ib_uqp_object         *obj;
2562         struct ib_qp                 *qp;
2563         struct ib_uverbs_mcast_entry *mcast;
2564         int                           ret = -EINVAL;
2565
2566         if (copy_from_user(&cmd, buf, sizeof cmd))
2567                 return -EFAULT;
2568
2569         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2570         if (!qp)
2571                 return -EINVAL;
2572
2573         ret = ib_detach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
2574         if (ret)
2575                 goto out_put;
2576
2577         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2578
2579         list_for_each_entry(mcast, &obj->mcast_list, list)
2580                 if (cmd.mlid == mcast->lid &&
2581                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2582                         list_del(&mcast->list);
2583                         kfree(mcast);
2584                         break;
2585                 }
2586
2587 out_put:
2588         put_qp_write(qp);
2589
2590         return ret ? ret : in_len;
2591 }
2592
2593 static int kern_spec_to_ib_spec(struct ib_uverbs_flow_spec *kern_spec,
2594                                 union ib_flow_spec *ib_spec)
2595 {
2596         if (kern_spec->reserved)
2597                 return -EINVAL;
2598
2599         ib_spec->type = kern_spec->type;
2600
2601         switch (ib_spec->type) {
2602         case IB_FLOW_SPEC_ETH:
2603                 ib_spec->eth.size = sizeof(struct ib_flow_spec_eth);
2604                 if (ib_spec->eth.size != kern_spec->eth.size)
2605                         return -EINVAL;
2606                 memcpy(&ib_spec->eth.val, &kern_spec->eth.val,
2607                        sizeof(struct ib_flow_eth_filter));
2608                 memcpy(&ib_spec->eth.mask, &kern_spec->eth.mask,
2609                        sizeof(struct ib_flow_eth_filter));
2610                 break;
2611         case IB_FLOW_SPEC_IPV4:
2612                 ib_spec->ipv4.size = sizeof(struct ib_flow_spec_ipv4);
2613                 if (ib_spec->ipv4.size != kern_spec->ipv4.size)
2614                         return -EINVAL;
2615                 memcpy(&ib_spec->ipv4.val, &kern_spec->ipv4.val,
2616                        sizeof(struct ib_flow_ipv4_filter));
2617                 memcpy(&ib_spec->ipv4.mask, &kern_spec->ipv4.mask,
2618                        sizeof(struct ib_flow_ipv4_filter));
2619                 break;
2620         case IB_FLOW_SPEC_TCP:
2621         case IB_FLOW_SPEC_UDP:
2622                 ib_spec->tcp_udp.size = sizeof(struct ib_flow_spec_tcp_udp);
2623                 if (ib_spec->tcp_udp.size != kern_spec->tcp_udp.size)
2624                         return -EINVAL;
2625                 memcpy(&ib_spec->tcp_udp.val, &kern_spec->tcp_udp.val,
2626                        sizeof(struct ib_flow_tcp_udp_filter));
2627                 memcpy(&ib_spec->tcp_udp.mask, &kern_spec->tcp_udp.mask,
2628                        sizeof(struct ib_flow_tcp_udp_filter));
2629                 break;
2630         default:
2631                 return -EINVAL;
2632         }
2633         return 0;
2634 }
2635
2636 int ib_uverbs_ex_create_flow(struct ib_uverbs_file *file,
2637                              struct ib_udata *ucore,
2638                              struct ib_udata *uhw)
2639 {
2640         struct ib_uverbs_create_flow      cmd;
2641         struct ib_uverbs_create_flow_resp resp;
2642         struct ib_uobject                 *uobj;
2643         struct ib_flow                    *flow_id;
2644         struct ib_uverbs_flow_attr        *kern_flow_attr;
2645         struct ib_flow_attr               *flow_attr;
2646         struct ib_qp                      *qp;
2647         int err = 0;
2648         void *kern_spec;
2649         void *ib_spec;
2650         int i;
2651
2652         if (ucore->inlen < sizeof(cmd))
2653                 return -EINVAL;
2654
2655         if (ucore->outlen < sizeof(resp))
2656                 return -ENOSPC;
2657
2658         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
2659         if (err)
2660                 return err;
2661
2662         ucore->inbuf += sizeof(cmd);
2663         ucore->inlen -= sizeof(cmd);
2664
2665         if (cmd.comp_mask)
2666                 return -EINVAL;
2667
2668         if ((cmd.flow_attr.type == IB_FLOW_ATTR_SNIFFER &&
2669              !capable(CAP_NET_ADMIN)) || !capable(CAP_NET_RAW))
2670                 return -EPERM;
2671
2672         if (cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
2673                 return -EINVAL;
2674
2675         if (cmd.flow_attr.size > ucore->inlen ||
2676             cmd.flow_attr.size >
2677             (cmd.flow_attr.num_of_specs * sizeof(struct ib_uverbs_flow_spec)))
2678                 return -EINVAL;
2679
2680         if (cmd.flow_attr.reserved[0] ||
2681             cmd.flow_attr.reserved[1])
2682                 return -EINVAL;
2683
2684         if (cmd.flow_attr.num_of_specs) {
2685                 kern_flow_attr = kmalloc(sizeof(*kern_flow_attr) + cmd.flow_attr.size,
2686                                          GFP_KERNEL);
2687                 if (!kern_flow_attr)
2688                         return -ENOMEM;
2689
2690                 memcpy(kern_flow_attr, &cmd.flow_attr, sizeof(*kern_flow_attr));
2691                 err = ib_copy_from_udata(kern_flow_attr + 1, ucore,
2692                                          cmd.flow_attr.size);
2693                 if (err)
2694                         goto err_free_attr;
2695         } else {
2696                 kern_flow_attr = &cmd.flow_attr;
2697         }
2698
2699         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
2700         if (!uobj) {
2701                 err = -ENOMEM;
2702                 goto err_free_attr;
2703         }
2704         init_uobj(uobj, 0, file->ucontext, &rule_lock_class);
2705         down_write(&uobj->mutex);
2706
2707         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2708         if (!qp) {
2709                 err = -EINVAL;
2710                 goto err_uobj;
2711         }
2712
2713         flow_attr = kmalloc(sizeof(*flow_attr) + cmd.flow_attr.size, GFP_KERNEL);
2714         if (!flow_attr) {
2715                 err = -ENOMEM;
2716                 goto err_put;
2717         }
2718
2719         flow_attr->type = kern_flow_attr->type;
2720         flow_attr->priority = kern_flow_attr->priority;
2721         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
2722         flow_attr->port = kern_flow_attr->port;
2723         flow_attr->flags = kern_flow_attr->flags;
2724         flow_attr->size = sizeof(*flow_attr);
2725
2726         kern_spec = kern_flow_attr + 1;
2727         ib_spec = flow_attr + 1;
2728         for (i = 0; i < flow_attr->num_of_specs &&
2729              cmd.flow_attr.size > offsetof(struct ib_uverbs_flow_spec, reserved) &&
2730              cmd.flow_attr.size >=
2731              ((struct ib_uverbs_flow_spec *)kern_spec)->size; i++) {
2732                 err = kern_spec_to_ib_spec(kern_spec, ib_spec);
2733                 if (err)
2734                         goto err_free;
2735                 flow_attr->size +=
2736                         ((union ib_flow_spec *) ib_spec)->size;
2737                 cmd.flow_attr.size -= ((struct ib_uverbs_flow_spec *)kern_spec)->size;
2738                 kern_spec += ((struct ib_uverbs_flow_spec *) kern_spec)->size;
2739                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
2740         }
2741         if (cmd.flow_attr.size || (i != flow_attr->num_of_specs)) {
2742                 pr_warn("create flow failed, flow %d: %d bytes left from uverb cmd\n",
2743                         i, cmd.flow_attr.size);
2744                 err = -EINVAL;
2745                 goto err_free;
2746         }
2747         flow_id = ib_create_flow(qp, flow_attr, IB_FLOW_DOMAIN_USER);
2748         if (IS_ERR(flow_id)) {
2749                 err = PTR_ERR(flow_id);
2750                 goto err_free;
2751         }
2752         flow_id->qp = qp;
2753         flow_id->uobject = uobj;
2754         uobj->object = flow_id;
2755
2756         err = idr_add_uobj(&ib_uverbs_rule_idr, uobj);
2757         if (err)
2758                 goto destroy_flow;
2759
2760         memset(&resp, 0, sizeof(resp));
2761         resp.flow_handle = uobj->id;
2762
2763         err = ib_copy_to_udata(ucore,
2764                                &resp, sizeof(resp));
2765         if (err)
2766                 goto err_copy;
2767
2768         put_qp_read(qp);
2769         mutex_lock(&file->mutex);
2770         list_add_tail(&uobj->list, &file->ucontext->rule_list);
2771         mutex_unlock(&file->mutex);
2772
2773         uobj->live = 1;
2774
2775         up_write(&uobj->mutex);
2776         kfree(flow_attr);
2777         if (cmd.flow_attr.num_of_specs)
2778                 kfree(kern_flow_attr);
2779         return 0;
2780 err_copy:
2781         idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
2782 destroy_flow:
2783         ib_destroy_flow(flow_id);
2784 err_free:
2785         kfree(flow_attr);
2786 err_put:
2787         put_qp_read(qp);
2788 err_uobj:
2789         put_uobj_write(uobj);
2790 err_free_attr:
2791         if (cmd.flow_attr.num_of_specs)
2792                 kfree(kern_flow_attr);
2793         return err;
2794 }
2795
2796 int ib_uverbs_ex_destroy_flow(struct ib_uverbs_file *file,
2797                               struct ib_udata *ucore,
2798                               struct ib_udata *uhw)
2799 {
2800         struct ib_uverbs_destroy_flow   cmd;
2801         struct ib_flow                  *flow_id;
2802         struct ib_uobject               *uobj;
2803         int                             ret;
2804
2805         if (ucore->inlen < sizeof(cmd))
2806                 return -EINVAL;
2807
2808         ret = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
2809         if (ret)
2810                 return ret;
2811
2812         if (cmd.comp_mask)
2813                 return -EINVAL;
2814
2815         uobj = idr_write_uobj(&ib_uverbs_rule_idr, cmd.flow_handle,
2816                               file->ucontext);
2817         if (!uobj)
2818                 return -EINVAL;
2819         flow_id = uobj->object;
2820
2821         ret = ib_destroy_flow(flow_id);
2822         if (!ret)
2823                 uobj->live = 0;
2824
2825         put_uobj_write(uobj);
2826
2827         idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
2828
2829         mutex_lock(&file->mutex);
2830         list_del(&uobj->list);
2831         mutex_unlock(&file->mutex);
2832
2833         put_uobj(uobj);
2834
2835         return ret;
2836 }
2837
2838 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
2839                                 struct ib_uverbs_create_xsrq *cmd,
2840                                 struct ib_udata *udata)
2841 {
2842         struct ib_uverbs_create_srq_resp resp;
2843         struct ib_usrq_object           *obj;
2844         struct ib_pd                    *pd;
2845         struct ib_srq                   *srq;
2846         struct ib_uobject               *uninitialized_var(xrcd_uobj);
2847         struct ib_srq_init_attr          attr;
2848         int ret;
2849
2850         obj = kmalloc(sizeof *obj, GFP_KERNEL);
2851         if (!obj)
2852                 return -ENOMEM;
2853
2854         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext, &srq_lock_class);
2855         down_write(&obj->uevent.uobject.mutex);
2856
2857         if (cmd->srq_type == IB_SRQT_XRC) {
2858                 attr.ext.xrc.xrcd  = idr_read_xrcd(cmd->xrcd_handle, file->ucontext, &xrcd_uobj);
2859                 if (!attr.ext.xrc.xrcd) {
2860                         ret = -EINVAL;
2861                         goto err;
2862                 }
2863
2864                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
2865                 atomic_inc(&obj->uxrcd->refcnt);
2866
2867                 attr.ext.xrc.cq  = idr_read_cq(cmd->cq_handle, file->ucontext, 0);
2868                 if (!attr.ext.xrc.cq) {
2869                         ret = -EINVAL;
2870                         goto err_put_xrcd;
2871                 }
2872         }
2873
2874         pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
2875         if (!pd) {
2876                 ret = -EINVAL;
2877                 goto err_put_cq;
2878         }
2879
2880         attr.event_handler  = ib_uverbs_srq_event_handler;
2881         attr.srq_context    = file;
2882         attr.srq_type       = cmd->srq_type;
2883         attr.attr.max_wr    = cmd->max_wr;
2884         attr.attr.max_sge   = cmd->max_sge;
2885         attr.attr.srq_limit = cmd->srq_limit;
2886
2887         obj->uevent.events_reported = 0;
2888         INIT_LIST_HEAD(&obj->uevent.event_list);
2889
2890         srq = pd->device->create_srq(pd, &attr, udata);
2891         if (IS_ERR(srq)) {
2892                 ret = PTR_ERR(srq);
2893                 goto err_put;
2894         }
2895
2896         srq->device        = pd->device;
2897         srq->pd            = pd;
2898         srq->srq_type      = cmd->srq_type;
2899         srq->uobject       = &obj->uevent.uobject;
2900         srq->event_handler = attr.event_handler;
2901         srq->srq_context   = attr.srq_context;
2902
2903         if (cmd->srq_type == IB_SRQT_XRC) {
2904                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
2905                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
2906                 atomic_inc(&attr.ext.xrc.cq->usecnt);
2907                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
2908         }
2909
2910         atomic_inc(&pd->usecnt);
2911         atomic_set(&srq->usecnt, 0);
2912
2913         obj->uevent.uobject.object = srq;
2914         ret = idr_add_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2915         if (ret)
2916                 goto err_destroy;
2917
2918         memset(&resp, 0, sizeof resp);
2919         resp.srq_handle = obj->uevent.uobject.id;
2920         resp.max_wr     = attr.attr.max_wr;
2921         resp.max_sge    = attr.attr.max_sge;
2922         if (cmd->srq_type == IB_SRQT_XRC)
2923                 resp.srqn = srq->ext.xrc.srq_num;
2924
2925         if (copy_to_user((void __user *) (unsigned long) cmd->response,
2926                          &resp, sizeof resp)) {
2927                 ret = -EFAULT;
2928                 goto err_copy;
2929         }
2930
2931         if (cmd->srq_type == IB_SRQT_XRC) {
2932                 put_uobj_read(xrcd_uobj);
2933                 put_cq_read(attr.ext.xrc.cq);
2934         }
2935         put_pd_read(pd);
2936
2937         mutex_lock(&file->mutex);
2938         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->srq_list);
2939         mutex_unlock(&file->mutex);
2940
2941         obj->uevent.uobject.live = 1;
2942
2943         up_write(&obj->uevent.uobject.mutex);
2944
2945         return 0;
2946
2947 err_copy:
2948         idr_remove_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2949
2950 err_destroy:
2951         ib_destroy_srq(srq);
2952
2953 err_put:
2954         put_pd_read(pd);
2955
2956 err_put_cq:
2957         if (cmd->srq_type == IB_SRQT_XRC)
2958                 put_cq_read(attr.ext.xrc.cq);
2959
2960 err_put_xrcd:
2961         if (cmd->srq_type == IB_SRQT_XRC) {
2962                 atomic_dec(&obj->uxrcd->refcnt);
2963                 put_uobj_read(xrcd_uobj);
2964         }
2965
2966 err:
2967         put_uobj_write(&obj->uevent.uobject);
2968         return ret;
2969 }
2970
2971 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
2972                              const char __user *buf, int in_len,
2973                              int out_len)
2974 {
2975         struct ib_uverbs_create_srq      cmd;
2976         struct ib_uverbs_create_xsrq     xcmd;
2977         struct ib_uverbs_create_srq_resp resp;
2978         struct ib_udata                  udata;
2979         int ret;
2980
2981         if (out_len < sizeof resp)
2982                 return -ENOSPC;
2983
2984         if (copy_from_user(&cmd, buf, sizeof cmd))
2985                 return -EFAULT;
2986
2987         xcmd.response    = cmd.response;
2988         xcmd.user_handle = cmd.user_handle;
2989         xcmd.srq_type    = IB_SRQT_BASIC;
2990         xcmd.pd_handle   = cmd.pd_handle;
2991         xcmd.max_wr      = cmd.max_wr;
2992         xcmd.max_sge     = cmd.max_sge;
2993         xcmd.srq_limit   = cmd.srq_limit;
2994
2995         INIT_UDATA(&udata, buf + sizeof cmd,
2996                    (unsigned long) cmd.response + sizeof resp,
2997                    in_len - sizeof cmd, out_len - sizeof resp);
2998
2999         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
3000         if (ret)
3001                 return ret;
3002
3003         return in_len;
3004 }
3005
3006 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
3007                               const char __user *buf, int in_len, int out_len)
3008 {
3009         struct ib_uverbs_create_xsrq     cmd;
3010         struct ib_uverbs_create_srq_resp resp;
3011         struct ib_udata                  udata;
3012         int ret;
3013
3014         if (out_len < sizeof resp)
3015                 return -ENOSPC;
3016
3017         if (copy_from_user(&cmd, buf, sizeof cmd))
3018                 return -EFAULT;
3019
3020         INIT_UDATA(&udata, buf + sizeof cmd,
3021                    (unsigned long) cmd.response + sizeof resp,
3022                    in_len - sizeof cmd, out_len - sizeof resp);
3023
3024         ret = __uverbs_create_xsrq(file, &cmd, &udata);
3025         if (ret)
3026                 return ret;
3027
3028         return in_len;
3029 }
3030
3031 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
3032                              const char __user *buf, int in_len,
3033                              int out_len)
3034 {
3035         struct ib_uverbs_modify_srq cmd;
3036         struct ib_udata             udata;
3037         struct ib_srq              *srq;
3038         struct ib_srq_attr          attr;
3039         int                         ret;
3040
3041         if (copy_from_user(&cmd, buf, sizeof cmd))
3042                 return -EFAULT;
3043
3044         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
3045                    out_len);
3046
3047         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
3048         if (!srq)
3049                 return -EINVAL;
3050
3051         attr.max_wr    = cmd.max_wr;
3052         attr.srq_limit = cmd.srq_limit;
3053
3054         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
3055
3056         put_srq_read(srq);
3057
3058         return ret ? ret : in_len;
3059 }
3060
3061 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
3062                             const char __user *buf,
3063                             int in_len, int out_len)
3064 {
3065         struct ib_uverbs_query_srq      cmd;
3066         struct ib_uverbs_query_srq_resp resp;
3067         struct ib_srq_attr              attr;
3068         struct ib_srq                   *srq;
3069         int                             ret;
3070
3071         if (out_len < sizeof resp)
3072                 return -ENOSPC;
3073
3074         if (copy_from_user(&cmd, buf, sizeof cmd))
3075                 return -EFAULT;
3076
3077         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
3078         if (!srq)
3079                 return -EINVAL;
3080
3081         ret = ib_query_srq(srq, &attr);
3082
3083         put_srq_read(srq);
3084
3085         if (ret)
3086                 return ret;
3087
3088         memset(&resp, 0, sizeof resp);
3089
3090         resp.max_wr    = attr.max_wr;
3091         resp.max_sge   = attr.max_sge;
3092         resp.srq_limit = attr.srq_limit;
3093
3094         if (copy_to_user((void __user *) (unsigned long) cmd.response,
3095                          &resp, sizeof resp))
3096                 return -EFAULT;
3097
3098         return in_len;
3099 }
3100
3101 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
3102                               const char __user *buf, int in_len,
3103                               int out_len)
3104 {
3105         struct ib_uverbs_destroy_srq      cmd;
3106         struct ib_uverbs_destroy_srq_resp resp;
3107         struct ib_uobject                *uobj;
3108         struct ib_srq                    *srq;
3109         struct ib_uevent_object          *obj;
3110         int                               ret = -EINVAL;
3111         struct ib_usrq_object            *us;
3112         enum ib_srq_type                  srq_type;
3113
3114         if (copy_from_user(&cmd, buf, sizeof cmd))
3115                 return -EFAULT;
3116
3117         uobj = idr_write_uobj(&ib_uverbs_srq_idr, cmd.srq_handle, file->ucontext);
3118         if (!uobj)
3119                 return -EINVAL;
3120         srq = uobj->object;
3121         obj = container_of(uobj, struct ib_uevent_object, uobject);
3122         srq_type = srq->srq_type;
3123
3124         ret = ib_destroy_srq(srq);
3125         if (!ret)
3126                 uobj->live = 0;
3127
3128         put_uobj_write(uobj);
3129
3130         if (ret)
3131                 return ret;
3132
3133         if (srq_type == IB_SRQT_XRC) {
3134                 us = container_of(obj, struct ib_usrq_object, uevent);
3135                 atomic_dec(&us->uxrcd->refcnt);
3136         }
3137
3138         idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
3139
3140         mutex_lock(&file->mutex);
3141         list_del(&uobj->list);
3142         mutex_unlock(&file->mutex);
3143
3144         ib_uverbs_release_uevent(file, obj);
3145
3146         memset(&resp, 0, sizeof resp);
3147         resp.events_reported = obj->events_reported;
3148
3149         put_uobj(uobj);
3150
3151         if (copy_to_user((void __user *) (unsigned long) cmd.response,
3152                          &resp, sizeof resp))
3153                 ret = -EFAULT;
3154
3155         return ret ? ret : in_len;
3156 }