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 #include <linux/sched.h>
  40
  41 #include <linux/uaccess.h>
  42
  43 #include "uverbs.h"
  44 #include "core_priv.h"
  45
  46 struct uverbs_lock_class {
  47         struct lock_class_key   key;
  48         char                    name[16];
  49 };
  50
  51 static struct uverbs_lock_class pd_lock_class   = { .name = "PD-uobj" };
  52 static struct uverbs_lock_class mr_lock_class   = { .name = "MR-uobj" };
  53 static struct uverbs_lock_class mw_lock_class   = { .name = "MW-uobj" };
  54 static struct uverbs_lock_class cq_lock_class   = { .name = "CQ-uobj" };
  55 static struct uverbs_lock_class qp_lock_class   = { .name = "QP-uobj" };
  56 static struct uverbs_lock_class ah_lock_class   = { .name = "AH-uobj" };
  57 static struct uverbs_lock_class srq_lock_class  = { .name = "SRQ-uobj" };
  58 static struct uverbs_lock_class xrcd_lock_class = { .name = "XRCD-uobj" };
  59 static struct uverbs_lock_class rule_lock_class = { .name = "RULE-uobj" };
  60 static struct uverbs_lock_class wq_lock_class = { .name = "WQ-uobj" };
  61 static struct uverbs_lock_class rwq_ind_table_lock_class = { .name = "IND_TBL-uobj" };
  62
  63 /*
  64  * The ib_uobject locking scheme is as follows:
  65  *
  66  * - ib_uverbs_idr_lock protects the uverbs idrs themselves, so it
  67  *   needs to be held during all idr write operations.  When an object is
  68  *   looked up, a reference must be taken on the object's kref before
  69  *   dropping this lock.  For read operations, the rcu_read_lock()
  70  *   and rcu_write_lock() but similarly the kref reference is grabbed
  71  *   before the rcu_read_unlock().
  72  *
  73  * - Each object also has an rwsem.  This rwsem must be held for
  74  *   reading while an operation that uses the object is performed.
  75  *   For example, while registering an MR, the associated PD's
  76  *   uobject.mutex must be held for reading.  The rwsem must be held
  77  *   for writing while initializing or destroying an object.
  78  *
  79  * - In addition, each object has a "live" flag.  If this flag is not
  80  *   set, then lookups of the object will fail even if it is found in
  81  *   the idr.  This handles a reader that blocks and does not acquire
  82  *   the rwsem until after the object is destroyed.  The destroy
  83  *   operation will set the live flag to 0 and then drop the rwsem;
  84  *   this will allow the reader to acquire the rwsem, see that the
  85  *   live flag is 0, and then drop the rwsem and its reference to
  86  *   object.  The underlying storage will not be freed until the last
  87  *   reference to the object is dropped.
  88  */
  89
  90 static void init_uobj(struct ib_uobject *uobj, u64 user_handle,
  91                       struct ib_ucontext *context, struct uverbs_lock_class *c)
  92 {
  93         uobj->user_handle = user_handle;
  94         uobj->context     = context;
  95         kref_init(&uobj->ref);
  96         init_rwsem(&uobj->mutex);
  97         lockdep_set_class_and_name(&uobj->mutex, &c->key, c->name);
  98         uobj->live        = 0;
  99 }
 100
 101 static void release_uobj(struct kref *kref)
 102 {
 103         kfree_rcu(container_of(kref, struct ib_uobject, ref), rcu);
 104 }
 105
 106 static void put_uobj(struct ib_uobject *uobj)
 107 {
 108         kref_put(&uobj->ref, release_uobj);
 109 }
 110
 111 static void put_uobj_read(struct ib_uobject *uobj)
 112 {
 113         up_read(&uobj->mutex);
 114         put_uobj(uobj);
 115 }
 116
 117 static void put_uobj_write(struct ib_uobject *uobj)
 118 {
 119         up_write(&uobj->mutex);
 120         put_uobj(uobj);
 121 }
 122
 123 static int idr_add_uobj(struct idr *idr, struct ib_uobject *uobj)
 124 {
 125         int ret;
 126
 127         idr_preload(GFP_KERNEL);
 128         spin_lock(&ib_uverbs_idr_lock);
 129
 130         ret = idr_alloc(idr, uobj, 0, 0, GFP_NOWAIT);
 131         if (ret >= 0)
 132                 uobj->id = ret;
 133
 134         spin_unlock(&ib_uverbs_idr_lock);
 135         idr_preload_end();
 136
 137         return ret < 0 ? ret : 0;
 138 }
 139
 140 void idr_remove_uobj(struct idr *idr, struct ib_uobject *uobj)
 141 {
 142         spin_lock(&ib_uverbs_idr_lock);
 143         idr_remove(idr, uobj->id);
 144         spin_unlock(&ib_uverbs_idr_lock);
 145 }
 146
 147 static struct ib_uobject *__idr_get_uobj(struct idr *idr, int id,
 148                                          struct ib_ucontext *context)
 149 {
 150         struct ib_uobject *uobj;
 151
 152         rcu_read_lock();
 153         uobj = idr_find(idr, id);
 154         if (uobj) {
 155                 if (uobj->context == context)
 156                         kref_get(&uobj->ref);
 157                 else
 158                         uobj = NULL;
 159         }
 160         rcu_read_unlock();
 161
 162         return uobj;
 163 }
 164
 165 static struct ib_uobject *idr_read_uobj(struct idr *idr, int id,
 166                                         struct ib_ucontext *context, int nested)
 167 {
 168         struct ib_uobject *uobj;
 169
 170         uobj = __idr_get_uobj(idr, id, context);
 171         if (!uobj)
 172                 return NULL;
 173
 174         if (nested)
 175                 down_read_nested(&uobj->mutex, SINGLE_DEPTH_NESTING);
 176         else
 177                 down_read(&uobj->mutex);
 178         if (!uobj->live) {
 179                 put_uobj_read(uobj);
 180                 return NULL;
 181         }
 182
 183         return uobj;
 184 }
 185
 186 static struct ib_uobject *idr_write_uobj(struct idr *idr, int id,
 187                                          struct ib_ucontext *context)
 188 {
 189         struct ib_uobject *uobj;
 190
 191         uobj = __idr_get_uobj(idr, id, context);
 192         if (!uobj)
 193                 return NULL;
 194
 195         down_write(&uobj->mutex);
 196         if (!uobj->live) {
 197                 put_uobj_write(uobj);
 198                 return NULL;
 199         }
 200
 201         return uobj;
 202 }
 203
 204 static void *idr_read_obj(struct idr *idr, int id, struct ib_ucontext *context,
 205                           int nested)
 206 {
 207         struct ib_uobject *uobj;
 208
 209         uobj = idr_read_uobj(idr, id, context, nested);
 210         return uobj ? uobj->object : NULL;
 211 }
 212
 213 static struct ib_pd *idr_read_pd(int pd_handle, struct ib_ucontext *context)
 214 {
 215         return idr_read_obj(&ib_uverbs_pd_idr, pd_handle, context, 0);
 216 }
 217
 218 static void put_pd_read(struct ib_pd *pd)
 219 {
 220         put_uobj_read(pd->uobject);
 221 }
 222
 223 static struct ib_cq *idr_read_cq(int cq_handle, struct ib_ucontext *context, int nested)
 224 {
 225         return idr_read_obj(&ib_uverbs_cq_idr, cq_handle, context, nested);
 226 }
 227
 228 static void put_cq_read(struct ib_cq *cq)
 229 {
 230         put_uobj_read(cq->uobject);
 231 }
 232
 233 static struct ib_ah *idr_read_ah(int ah_handle, struct ib_ucontext *context)
 234 {
 235         return idr_read_obj(&ib_uverbs_ah_idr, ah_handle, context, 0);
 236 }
 237
 238 static void put_ah_read(struct ib_ah *ah)
 239 {
 240         put_uobj_read(ah->uobject);
 241 }
 242
 243 static struct ib_qp *idr_read_qp(int qp_handle, struct ib_ucontext *context)
 244 {
 245         return idr_read_obj(&ib_uverbs_qp_idr, qp_handle, context, 0);
 246 }
 247
 248 static struct ib_wq *idr_read_wq(int wq_handle, struct ib_ucontext *context)
 249 {
 250         return idr_read_obj(&ib_uverbs_wq_idr, wq_handle, context, 0);
 251 }
 252
 253 static void put_wq_read(struct ib_wq *wq)
 254 {
 255         put_uobj_read(wq->uobject);
 256 }
 257
 258 static struct ib_rwq_ind_table *idr_read_rwq_indirection_table(int ind_table_handle,
 259                                                                struct ib_ucontext *context)
 260 {
 261         return idr_read_obj(&ib_uverbs_rwq_ind_tbl_idr, ind_table_handle, context, 0);
 262 }
 263
 264 static void put_rwq_indirection_table_read(struct ib_rwq_ind_table *ind_table)
 265 {
 266         put_uobj_read(ind_table->uobject);
 267 }
 268
 269 static struct ib_qp *idr_write_qp(int qp_handle, struct ib_ucontext *context)
 270 {
 271         struct ib_uobject *uobj;
 272
 273         uobj = idr_write_uobj(&ib_uverbs_qp_idr, qp_handle, context);
 274         return uobj ? uobj->object : NULL;
 275 }
 276
 277 static void put_qp_read(struct ib_qp *qp)
 278 {
 279         put_uobj_read(qp->uobject);
 280 }
 281
 282 static void put_qp_write(struct ib_qp *qp)
 283 {
 284         put_uobj_write(qp->uobject);
 285 }
 286
 287 static struct ib_srq *idr_read_srq(int srq_handle, struct ib_ucontext *context)
 288 {
 289         return idr_read_obj(&ib_uverbs_srq_idr, srq_handle, context, 0);
 290 }
 291
 292 static void put_srq_read(struct ib_srq *srq)
 293 {
 294         put_uobj_read(srq->uobject);
 295 }
 296
 297 static struct ib_xrcd *idr_read_xrcd(int xrcd_handle, struct ib_ucontext *context,
 298                                      struct ib_uobject **uobj)
 299 {
 300         *uobj = idr_read_uobj(&ib_uverbs_xrcd_idr, xrcd_handle, context, 0);
 301         return *uobj ? (*uobj)->object : NULL;
 302 }
 303
 304 static void put_xrcd_read(struct ib_uobject *uobj)
 305 {
 306         put_uobj_read(uobj);
 307 }
 308
 309 ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
 310                               struct ib_device *ib_dev,
 311                               const char __user *buf,
 312                               int in_len, int out_len)
 313 {
 314         struct ib_uverbs_get_context      cmd;
 315         struct ib_uverbs_get_context_resp resp;
 316         struct ib_udata                   udata;
 317         struct ib_ucontext               *ucontext;
 318         struct file                      *filp;
 319         int ret;
 320
 321         if (out_len < sizeof resp)
 322                 return -ENOSPC;
 323
 324         if (copy_from_user(&cmd, buf, sizeof cmd))
 325                 return -EFAULT;
 326
 327         mutex_lock(&file->mutex);
 328
 329         if (file->ucontext) {
 330                 ret = -EINVAL;
 331                 goto err;
 332         }
 333
 334         INIT_UDATA(&udata, buf + sizeof cmd,
 335                    (unsigned long) cmd.response + sizeof resp,
 336                    in_len - sizeof cmd, out_len - sizeof resp);
 337
 338         ucontext = ib_dev->alloc_ucontext(ib_dev, &udata);
 339         if (IS_ERR(ucontext)) {
 340                 ret = PTR_ERR(ucontext);
 341                 goto err;
 342         }
 343
 344         ucontext->device = ib_dev;
 345         INIT_LIST_HEAD(&ucontext->pd_list);
 346         INIT_LIST_HEAD(&ucontext->mr_list);
 347         INIT_LIST_HEAD(&ucontext->mw_list);
 348         INIT_LIST_HEAD(&ucontext->cq_list);
 349         INIT_LIST_HEAD(&ucontext->qp_list);
 350         INIT_LIST_HEAD(&ucontext->srq_list);
 351         INIT_LIST_HEAD(&ucontext->ah_list);
 352         INIT_LIST_HEAD(&ucontext->wq_list);
 353         INIT_LIST_HEAD(&ucontext->rwq_ind_tbl_list);
 354         INIT_LIST_HEAD(&ucontext->xrcd_list);
 355         INIT_LIST_HEAD(&ucontext->rule_list);
 356         rcu_read_lock();
 357         ucontext->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
 358         rcu_read_unlock();
 359         ucontext->closing = 0;
 360
 361 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 362         ucontext->umem_tree = RB_ROOT;
 363         init_rwsem(&ucontext->umem_rwsem);
 364         ucontext->odp_mrs_count = 0;
 365         INIT_LIST_HEAD(&ucontext->no_private_counters);
 366
 367         if (!(ib_dev->attrs.device_cap_flags & IB_DEVICE_ON_DEMAND_PAGING))
 368                 ucontext->invalidate_range = NULL;
 369
 370 #endif
 371
 372         resp.num_comp_vectors = file->device->num_comp_vectors;
 373
 374         ret = get_unused_fd_flags(O_CLOEXEC);
 375         if (ret < 0)
 376                 goto err_free;
 377         resp.async_fd = ret;
 378
 379         filp = ib_uverbs_alloc_event_file(file, ib_dev, 1);
 380         if (IS_ERR(filp)) {
 381                 ret = PTR_ERR(filp);
 382                 goto err_fd;
 383         }
 384
 385         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 386                          &resp, sizeof resp)) {
 387                 ret = -EFAULT;
 388                 goto err_file;
 389         }
 390
 391         file->ucontext = ucontext;
 392
 393         fd_install(resp.async_fd, filp);
 394
 395         mutex_unlock(&file->mutex);
 396
 397         return in_len;
 398
 399 err_file:
 400         ib_uverbs_free_async_event_file(file);
 401         fput(filp);
 402
 403 err_fd:
 404         put_unused_fd(resp.async_fd);
 405
 406 err_free:
 407         put_pid(ucontext->tgid);
 408         ib_dev->dealloc_ucontext(ucontext);
 409
 410 err:
 411         mutex_unlock(&file->mutex);
 412         return ret;
 413 }
 414
 415 static void copy_query_dev_fields(struct ib_uverbs_file *file,
 416                                   struct ib_device *ib_dev,
 417                                   struct ib_uverbs_query_device_resp *resp,
 418                                   struct ib_device_attr *attr)
 419 {
 420         resp->fw_ver            = attr->fw_ver;
 421         resp->node_guid         = ib_dev->node_guid;
 422         resp->sys_image_guid    = attr->sys_image_guid;
 423         resp->max_mr_size       = attr->max_mr_size;
 424         resp->page_size_cap     = attr->page_size_cap;
 425         resp->vendor_id         = attr->vendor_id;
 426         resp->vendor_part_id    = attr->vendor_part_id;
 427         resp->hw_ver            = attr->hw_ver;
 428         resp->max_qp            = attr->max_qp;
 429         resp->max_qp_wr         = attr->max_qp_wr;
 430         resp->device_cap_flags  = lower_32_bits(attr->device_cap_flags);
 431         resp->max_sge           = attr->max_sge;
 432         resp->max_sge_rd        = attr->max_sge_rd;
 433         resp->max_cq            = attr->max_cq;
 434         resp->max_cqe           = attr->max_cqe;
 435         resp->max_mr            = attr->max_mr;
 436         resp->max_pd            = attr->max_pd;
 437         resp->max_qp_rd_atom    = attr->max_qp_rd_atom;
 438         resp->max_ee_rd_atom    = attr->max_ee_rd_atom;
 439         resp->max_res_rd_atom   = attr->max_res_rd_atom;
 440         resp->max_qp_init_rd_atom       = attr->max_qp_init_rd_atom;
 441         resp->max_ee_init_rd_atom       = attr->max_ee_init_rd_atom;
 442         resp->atomic_cap                = attr->atomic_cap;
 443         resp->max_ee                    = attr->max_ee;
 444         resp->max_rdd                   = attr->max_rdd;
 445         resp->max_mw                    = attr->max_mw;
 446         resp->max_raw_ipv6_qp           = attr->max_raw_ipv6_qp;
 447         resp->max_raw_ethy_qp           = attr->max_raw_ethy_qp;
 448         resp->max_mcast_grp             = attr->max_mcast_grp;
 449         resp->max_mcast_qp_attach       = attr->max_mcast_qp_attach;
 450         resp->max_total_mcast_qp_attach = attr->max_total_mcast_qp_attach;
 451         resp->max_ah                    = attr->max_ah;
 452         resp->max_fmr                   = attr->max_fmr;
 453         resp->max_map_per_fmr           = attr->max_map_per_fmr;
 454         resp->max_srq                   = attr->max_srq;
 455         resp->max_srq_wr                = attr->max_srq_wr;
 456         resp->max_srq_sge               = attr->max_srq_sge;
 457         resp->max_pkeys                 = attr->max_pkeys;
 458         resp->local_ca_ack_delay        = attr->local_ca_ack_delay;
 459         resp->phys_port_cnt             = ib_dev->phys_port_cnt;
 460 }
 461
 462 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 463                                struct ib_device *ib_dev,
 464                                const char __user *buf,
 465                                int in_len, int out_len)
 466 {
 467         struct ib_uverbs_query_device      cmd;
 468         struct ib_uverbs_query_device_resp resp;
 469
 470         if (out_len < sizeof resp)
 471                 return -ENOSPC;
 472
 473         if (copy_from_user(&cmd, buf, sizeof cmd))
 474                 return -EFAULT;
 475
 476         memset(&resp, 0, sizeof resp);
 477         copy_query_dev_fields(file, ib_dev, &resp, &ib_dev->attrs);
 478
 479         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 480                          &resp, sizeof resp))
 481                 return -EFAULT;
 482
 483         return in_len;
 484 }
 485
 486 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 487                              struct ib_device *ib_dev,
 488                              const char __user *buf,
 489                              int in_len, int out_len)
 490 {
 491         struct ib_uverbs_query_port      cmd;
 492         struct ib_uverbs_query_port_resp resp;
 493         struct ib_port_attr              attr;
 494         int                              ret;
 495
 496         if (out_len < sizeof resp)
 497                 return -ENOSPC;
 498
 499         if (copy_from_user(&cmd, buf, sizeof cmd))
 500                 return -EFAULT;
 501
 502         ret = ib_query_port(ib_dev, cmd.port_num, &attr);
 503         if (ret)
 504                 return ret;
 505
 506         memset(&resp, 0, sizeof resp);
 507
 508         resp.state           = attr.state;
 509         resp.max_mtu         = attr.max_mtu;
 510         resp.active_mtu      = attr.active_mtu;
 511         resp.gid_tbl_len     = attr.gid_tbl_len;
 512         resp.port_cap_flags  = attr.port_cap_flags;
 513         resp.max_msg_sz      = attr.max_msg_sz;
 514         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 515         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 516         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 517         resp.lid             = attr.lid;
 518         resp.sm_lid          = attr.sm_lid;
 519         resp.lmc             = attr.lmc;
 520         resp.max_vl_num      = attr.max_vl_num;
 521         resp.sm_sl           = attr.sm_sl;
 522         resp.subnet_timeout  = attr.subnet_timeout;
 523         resp.init_type_reply = attr.init_type_reply;
 524         resp.active_width    = attr.active_width;
 525         resp.active_speed    = attr.active_speed;
 526         resp.phys_state      = attr.phys_state;
 527         resp.link_layer      = rdma_port_get_link_layer(ib_dev,
 528                                                         cmd.port_num);
 529
 530         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 531                          &resp, sizeof resp))
 532                 return -EFAULT;
 533
 534         return in_len;
 535 }
 536
 537 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 538                            struct ib_device *ib_dev,
 539                            const char __user *buf,
 540                            int in_len, int out_len)
 541 {
 542         struct ib_uverbs_alloc_pd      cmd;
 543         struct ib_uverbs_alloc_pd_resp resp;
 544         struct ib_udata                udata;
 545         struct ib_uobject             *uobj;
 546         struct ib_pd                  *pd;
 547         int                            ret;
 548
 549         if (out_len < sizeof resp)
 550                 return -ENOSPC;
 551
 552         if (copy_from_user(&cmd, buf, sizeof cmd))
 553                 return -EFAULT;
 554
 555         INIT_UDATA(&udata, buf + sizeof cmd,
 556                    (unsigned long) cmd.response + sizeof resp,
 557                    in_len - sizeof cmd, out_len - sizeof resp);
 558
 559         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 560         if (!uobj)
 561                 return -ENOMEM;
 562
 563         init_uobj(uobj, 0, file->ucontext, &pd_lock_class);
 564         down_write(&uobj->mutex);
 565
 566         pd = ib_dev->alloc_pd(ib_dev, file->ucontext, &udata);
 567         if (IS_ERR(pd)) {
 568                 ret = PTR_ERR(pd);
 569                 goto err;
 570         }
 571
 572         pd->device  = ib_dev;
 573         pd->uobject = uobj;
 574         pd->__internal_mr = NULL;
 575         atomic_set(&pd->usecnt, 0);
 576
 577         uobj->object = pd;
 578         ret = idr_add_uobj(&ib_uverbs_pd_idr, uobj);
 579         if (ret)
 580                 goto err_idr;
 581
 582         memset(&resp, 0, sizeof resp);
 583         resp.pd_handle = uobj->id;
 584
 585         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 586                          &resp, sizeof resp)) {
 587                 ret = -EFAULT;
 588                 goto err_copy;
 589         }
 590
 591         mutex_lock(&file->mutex);
 592         list_add_tail(&uobj->list, &file->ucontext->pd_list);
 593         mutex_unlock(&file->mutex);
 594
 595         uobj->live = 1;
 596
 597         up_write(&uobj->mutex);
 598
 599         return in_len;
 600
 601 err_copy:
 602         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 603
 604 err_idr:
 605         ib_dealloc_pd(pd);
 606
 607 err:
 608         put_uobj_write(uobj);
 609         return ret;
 610 }
 611
 612 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 613                              struct ib_device *ib_dev,
 614                              const char __user *buf,
 615                              int in_len, int out_len)
 616 {
 617         struct ib_uverbs_dealloc_pd cmd;
 618         struct ib_uobject          *uobj;
 619         struct ib_pd               *pd;
 620         int                         ret;
 621
 622         if (copy_from_user(&cmd, buf, sizeof cmd))
 623                 return -EFAULT;
 624
 625         uobj = idr_write_uobj(&ib_uverbs_pd_idr, cmd.pd_handle, file->ucontext);
 626         if (!uobj)
 627                 return -EINVAL;
 628         pd = uobj->object;
 629
 630         if (atomic_read(&pd->usecnt)) {
 631                 ret = -EBUSY;
 632                 goto err_put;
 633         }
 634
 635         ret = pd->device->dealloc_pd(uobj->object);
 636         WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
 637         if (ret)
 638                 goto err_put;
 639
 640         uobj->live = 0;
 641         put_uobj_write(uobj);
 642
 643         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 644
 645         mutex_lock(&file->mutex);
 646         list_del(&uobj->list);
 647         mutex_unlock(&file->mutex);
 648
 649         put_uobj(uobj);
 650
 651         return in_len;
 652
 653 err_put:
 654         put_uobj_write(uobj);
 655         return ret;
 656 }
 657
 658 struct xrcd_table_entry {
 659         struct rb_node  node;
 660         struct ib_xrcd *xrcd;
 661         struct inode   *inode;
 662 };
 663
 664 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 665                             struct inode *inode,
 666                             struct ib_xrcd *xrcd)
 667 {
 668         struct xrcd_table_entry *entry, *scan;
 669         struct rb_node **p = &dev->xrcd_tree.rb_node;
 670         struct rb_node *parent = NULL;
 671
 672         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 673         if (!entry)
 674                 return -ENOMEM;
 675
 676         entry->xrcd  = xrcd;
 677         entry->inode = inode;
 678
 679         while (*p) {
 680                 parent = *p;
 681                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 682
 683                 if (inode < scan->inode) {
 684                         p = &(*p)->rb_left;
 685                 } else if (inode > scan->inode) {
 686                         p = &(*p)->rb_right;
 687                 } else {
 688                         kfree(entry);
 689                         return -EEXIST;
 690                 }
 691         }
 692
 693         rb_link_node(&entry->node, parent, p);
 694         rb_insert_color(&entry->node, &dev->xrcd_tree);
 695         igrab(inode);
 696         return 0;
 697 }
 698
 699 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 700                                                   struct inode *inode)
 701 {
 702         struct xrcd_table_entry *entry;
 703         struct rb_node *p = dev->xrcd_tree.rb_node;
 704
 705         while (p) {
 706                 entry = rb_entry(p, struct xrcd_table_entry, node);
 707
 708                 if (inode < entry->inode)
 709                         p = p->rb_left;
 710                 else if (inode > entry->inode)
 711                         p = p->rb_right;
 712                 else
 713                         return entry;
 714         }
 715
 716         return NULL;
 717 }
 718
 719 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 720 {
 721         struct xrcd_table_entry *entry;
 722
 723         entry = xrcd_table_search(dev, inode);
 724         if (!entry)
 725                 return NULL;
 726
 727         return entry->xrcd;
 728 }
 729
 730 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 731                               struct inode *inode)
 732 {
 733         struct xrcd_table_entry *entry;
 734
 735         entry = xrcd_table_search(dev, inode);
 736         if (entry) {
 737                 iput(inode);
 738                 rb_erase(&entry->node, &dev->xrcd_tree);
 739                 kfree(entry);
 740         }
 741 }
 742
 743 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 744                             struct ib_device *ib_dev,
 745                             const char __user *buf, int in_len,
 746                             int out_len)
 747 {
 748         struct ib_uverbs_open_xrcd      cmd;
 749         struct ib_uverbs_open_xrcd_resp resp;
 750         struct ib_udata                 udata;
 751         struct ib_uxrcd_object         *obj;
 752         struct ib_xrcd                 *xrcd = NULL;
 753         struct fd                       f = {NULL, 0};
 754         struct inode                   *inode = NULL;
 755         int                             ret = 0;
 756         int                             new_xrcd = 0;
 757
 758         if (out_len < sizeof resp)
 759                 return -ENOSPC;
 760
 761         if (copy_from_user(&cmd, buf, sizeof cmd))
 762                 return -EFAULT;
 763
 764         INIT_UDATA(&udata, buf + sizeof cmd,
 765                    (unsigned long) cmd.response + sizeof resp,
 766                    in_len - sizeof cmd, out_len - sizeof  resp);
 767
 768         mutex_lock(&file->device->xrcd_tree_mutex);
 769
 770         if (cmd.fd != -1) {
 771                 /* search for file descriptor */
 772                 f = fdget(cmd.fd);
 773                 if (!f.file) {
 774                         ret = -EBADF;
 775                         goto err_tree_mutex_unlock;
 776                 }
 777
 778                 inode = file_inode(f.file);
 779                 xrcd = find_xrcd(file->device, inode);
 780                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 781                         /* no file descriptor. Need CREATE flag */
 782                         ret = -EAGAIN;
 783                         goto err_tree_mutex_unlock;
 784                 }
 785
 786                 if (xrcd && cmd.oflags & O_EXCL) {
 787                         ret = -EINVAL;
 788                         goto err_tree_mutex_unlock;
 789                 }
 790         }
 791
 792         obj = kmalloc(sizeof *obj, GFP_KERNEL);
 793         if (!obj) {
 794                 ret = -ENOMEM;
 795                 goto err_tree_mutex_unlock;
 796         }
 797
 798         init_uobj(&obj->uobject, 0, file->ucontext, &xrcd_lock_class);
 799
 800         down_write(&obj->uobject.mutex);
 801
 802         if (!xrcd) {
 803                 xrcd = ib_dev->alloc_xrcd(ib_dev, file->ucontext, &udata);
 804                 if (IS_ERR(xrcd)) {
 805                         ret = PTR_ERR(xrcd);
 806                         goto err;
 807                 }
 808
 809                 xrcd->inode   = inode;
 810                 xrcd->device  = ib_dev;
 811                 atomic_set(&xrcd->usecnt, 0);
 812                 mutex_init(&xrcd->tgt_qp_mutex);
 813                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 814                 new_xrcd = 1;
 815         }
 816
 817         atomic_set(&obj->refcnt, 0);
 818         obj->uobject.object = xrcd;
 819         ret = idr_add_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 820         if (ret)
 821                 goto err_idr;
 822
 823         memset(&resp, 0, sizeof resp);
 824         resp.xrcd_handle = obj->uobject.id;
 825
 826         if (inode) {
 827                 if (new_xrcd) {
 828                         /* create new inode/xrcd table entry */
 829                         ret = xrcd_table_insert(file->device, inode, xrcd);
 830                         if (ret)
 831                                 goto err_insert_xrcd;
 832                 }
 833                 atomic_inc(&xrcd->usecnt);
 834         }
 835
 836         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 837                          &resp, sizeof resp)) {
 838                 ret = -EFAULT;
 839                 goto err_copy;
 840         }
 841
 842         if (f.file)
 843                 fdput(f);
 844
 845         mutex_lock(&file->mutex);
 846         list_add_tail(&obj->uobject.list, &file->ucontext->xrcd_list);
 847         mutex_unlock(&file->mutex);
 848
 849         obj->uobject.live = 1;
 850         up_write(&obj->uobject.mutex);
 851
 852         mutex_unlock(&file->device->xrcd_tree_mutex);
 853         return in_len;
 854
 855 err_copy:
 856         if (inode) {
 857                 if (new_xrcd)
 858                         xrcd_table_delete(file->device, inode);
 859                 atomic_dec(&xrcd->usecnt);
 860         }
 861
 862 err_insert_xrcd:
 863         idr_remove_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 864
 865 err_idr:
 866         ib_dealloc_xrcd(xrcd);
 867
 868 err:
 869         put_uobj_write(&obj->uobject);
 870
 871 err_tree_mutex_unlock:
 872         if (f.file)
 873                 fdput(f);
 874
 875         mutex_unlock(&file->device->xrcd_tree_mutex);
 876
 877         return ret;
 878 }
 879
 880 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 881                              struct ib_device *ib_dev,
 882                              const char __user *buf, int in_len,
 883                              int out_len)
 884 {
 885         struct ib_uverbs_close_xrcd cmd;
 886         struct ib_uobject           *uobj;
 887         struct ib_xrcd              *xrcd = NULL;
 888         struct inode                *inode = NULL;
 889         struct ib_uxrcd_object      *obj;
 890         int                         live;
 891         int                         ret = 0;
 892
 893         if (copy_from_user(&cmd, buf, sizeof cmd))
 894                 return -EFAULT;
 895
 896         mutex_lock(&file->device->xrcd_tree_mutex);
 897         uobj = idr_write_uobj(&ib_uverbs_xrcd_idr, cmd.xrcd_handle, file->ucontext);
 898         if (!uobj) {
 899                 ret = -EINVAL;
 900                 goto out;
 901         }
 902
 903         xrcd  = uobj->object;
 904         inode = xrcd->inode;
 905         obj   = container_of(uobj, struct ib_uxrcd_object, uobject);
 906         if (atomic_read(&obj->refcnt)) {
 907                 put_uobj_write(uobj);
 908                 ret = -EBUSY;
 909                 goto out;
 910         }
 911
 912         if (!inode || atomic_dec_and_test(&xrcd->usecnt)) {
 913                 ret = ib_dealloc_xrcd(uobj->object);
 914                 if (!ret)
 915                         uobj->live = 0;
 916         }
 917
 918         live = uobj->live;
 919         if (inode && ret)
 920                 atomic_inc(&xrcd->usecnt);
 921
 922         put_uobj_write(uobj);
 923
 924         if (ret)
 925                 goto out;
 926
 927         if (inode && !live)
 928                 xrcd_table_delete(file->device, inode);
 929
 930         idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 931         mutex_lock(&file->mutex);
 932         list_del(&uobj->list);
 933         mutex_unlock(&file->mutex);
 934
 935         put_uobj(uobj);
 936         ret = in_len;
 937
 938 out:
 939         mutex_unlock(&file->device->xrcd_tree_mutex);
 940         return ret;
 941 }
 942
 943 void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 944                             struct ib_xrcd *xrcd)
 945 {
 946         struct inode *inode;
 947
 948         inode = xrcd->inode;
 949         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 950                 return;
 951
 952         ib_dealloc_xrcd(xrcd);
 953
 954         if (inode)
 955                 xrcd_table_delete(dev, inode);
 956 }
 957
 958 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 959                          struct ib_device *ib_dev,
 960                          const char __user *buf, int in_len,
 961                          int out_len)
 962 {
 963         struct ib_uverbs_reg_mr      cmd;
 964         struct ib_uverbs_reg_mr_resp resp;
 965         struct ib_udata              udata;
 966         struct ib_uobject           *uobj;
 967         struct ib_pd                *pd;
 968         struct ib_mr                *mr;
 969         int                          ret;
 970
 971         if (out_len < sizeof resp)
 972                 return -ENOSPC;
 973
 974         if (copy_from_user(&cmd, buf, sizeof cmd))
 975                 return -EFAULT;
 976
 977         INIT_UDATA(&udata, buf + sizeof cmd,
 978                    (unsigned long) cmd.response + sizeof resp,
 979                    in_len - sizeof cmd, out_len - sizeof resp);
 980
 981         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 982                 return -EINVAL;
 983
 984         ret = ib_check_mr_access(cmd.access_flags);
 985         if (ret)
 986                 return ret;
 987
 988         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 989         if (!uobj)
 990                 return -ENOMEM;
 991
 992         init_uobj(uobj, 0, file->ucontext, &mr_lock_class);
 993         down_write(&uobj->mutex);
 994
 995         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
 996         if (!pd) {
 997                 ret = -EINVAL;
 998                 goto err_free;
 999         }
1000
1001         if (cmd.access_flags & IB_ACCESS_ON_DEMAND) {
1002                 if (!(pd->device->attrs.device_cap_flags &
1003                       IB_DEVICE_ON_DEMAND_PAGING)) {
1004                         pr_debug("ODP support not available\n");
1005                         ret = -EINVAL;
1006                         goto err_put;
1007                 }
1008         }
1009
1010         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
1011                                      cmd.access_flags, &udata);
1012         if (IS_ERR(mr)) {
1013                 ret = PTR_ERR(mr);
1014                 goto err_put;
1015         }
1016
1017         mr->device  = pd->device;
1018         mr->pd      = pd;
1019         mr->uobject = uobj;
1020         atomic_inc(&pd->usecnt);
1021
1022         uobj->object = mr;
1023         ret = idr_add_uobj(&ib_uverbs_mr_idr, uobj);
1024         if (ret)
1025                 goto err_unreg;
1026
1027         memset(&resp, 0, sizeof resp);
1028         resp.lkey      = mr->lkey;
1029         resp.rkey      = mr->rkey;
1030         resp.mr_handle = uobj->id;
1031
1032         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1033                          &resp, sizeof resp)) {
1034                 ret = -EFAULT;
1035                 goto err_copy;
1036         }
1037
1038         put_pd_read(pd);
1039
1040         mutex_lock(&file->mutex);
1041         list_add_tail(&uobj->list, &file->ucontext->mr_list);
1042         mutex_unlock(&file->mutex);
1043
1044         uobj->live = 1;
1045
1046         up_write(&uobj->mutex);
1047
1048         return in_len;
1049
1050 err_copy:
1051         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1052
1053 err_unreg:
1054         ib_dereg_mr(mr);
1055
1056 err_put:
1057         put_pd_read(pd);
1058
1059 err_free:
1060         put_uobj_write(uobj);
1061         return ret;
1062 }
1063
1064 ssize_t ib_uverbs_rereg_mr(struct ib_uverbs_file *file,
1065                            struct ib_device *ib_dev,
1066                            const char __user *buf, int in_len,
1067                            int out_len)
1068 {
1069         struct ib_uverbs_rereg_mr      cmd;
1070         struct ib_uverbs_rereg_mr_resp resp;
1071         struct ib_udata              udata;
1072         struct ib_pd                *pd = NULL;
1073         struct ib_mr                *mr;
1074         struct ib_pd                *old_pd;
1075         int                          ret;
1076         struct ib_uobject           *uobj;
1077
1078         if (out_len < sizeof(resp))
1079                 return -ENOSPC;
1080
1081         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1082                 return -EFAULT;
1083
1084         INIT_UDATA(&udata, buf + sizeof(cmd),
1085                    (unsigned long) cmd.response + sizeof(resp),
1086                    in_len - sizeof(cmd), out_len - sizeof(resp));
1087
1088         if (cmd.flags & ~IB_MR_REREG_SUPPORTED || !cmd.flags)
1089                 return -EINVAL;
1090
1091         if ((cmd.flags & IB_MR_REREG_TRANS) &&
1092             (!cmd.start || !cmd.hca_va || 0 >= cmd.length ||
1093              (cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK)))
1094                         return -EINVAL;
1095
1096         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle,
1097                               file->ucontext);
1098
1099         if (!uobj)
1100                 return -EINVAL;
1101
1102         mr = uobj->object;
1103
1104         if (cmd.flags & IB_MR_REREG_ACCESS) {
1105                 ret = ib_check_mr_access(cmd.access_flags);
1106                 if (ret)
1107                         goto put_uobjs;
1108         }
1109
1110         if (cmd.flags & IB_MR_REREG_PD) {
1111                 pd = idr_read_pd(cmd.pd_handle, file->ucontext);
1112                 if (!pd) {
1113                         ret = -EINVAL;
1114                         goto put_uobjs;
1115                 }
1116         }
1117
1118         old_pd = mr->pd;
1119         ret = mr->device->rereg_user_mr(mr, cmd.flags, cmd.start,
1120                                         cmd.length, cmd.hca_va,
1121                                         cmd.access_flags, pd, &udata);
1122         if (!ret) {
1123                 if (cmd.flags & IB_MR_REREG_PD) {
1124                         atomic_inc(&pd->usecnt);
1125                         mr->pd = pd;
1126                         atomic_dec(&old_pd->usecnt);
1127                 }
1128         } else {
1129                 goto put_uobj_pd;
1130         }
1131
1132         memset(&resp, 0, sizeof(resp));
1133         resp.lkey      = mr->lkey;
1134         resp.rkey      = mr->rkey;
1135
1136         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1137                          &resp, sizeof(resp)))
1138                 ret = -EFAULT;
1139         else
1140                 ret = in_len;
1141
1142 put_uobj_pd:
1143         if (cmd.flags & IB_MR_REREG_PD)
1144                 put_pd_read(pd);
1145
1146 put_uobjs:
1147
1148         put_uobj_write(mr->uobject);
1149
1150         return ret;
1151 }
1152
1153 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
1154                            struct ib_device *ib_dev,
1155                            const char __user *buf, int in_len,
1156                            int out_len)
1157 {
1158         struct ib_uverbs_dereg_mr cmd;
1159         struct ib_mr             *mr;
1160         struct ib_uobject        *uobj;
1161         int                       ret = -EINVAL;
1162
1163         if (copy_from_user(&cmd, buf, sizeof cmd))
1164                 return -EFAULT;
1165
1166         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle, file->ucontext);
1167         if (!uobj)
1168                 return -EINVAL;
1169
1170         mr = uobj->object;
1171
1172         ret = ib_dereg_mr(mr);
1173         if (!ret)
1174                 uobj->live = 0;
1175
1176         put_uobj_write(uobj);
1177
1178         if (ret)
1179                 return ret;
1180
1181         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1182
1183         mutex_lock(&file->mutex);
1184         list_del(&uobj->list);
1185         mutex_unlock(&file->mutex);
1186
1187         put_uobj(uobj);
1188
1189         return in_len;
1190 }
1191
1192 ssize_t ib_uverbs_alloc_mw(struct ib_uverbs_file *file,
1193                            struct ib_device *ib_dev,
1194                            const char __user *buf, int in_len,
1195                            int out_len)
1196 {
1197         struct ib_uverbs_alloc_mw      cmd;
1198         struct ib_uverbs_alloc_mw_resp resp;
1199         struct ib_uobject             *uobj;
1200         struct ib_pd                  *pd;
1201         struct ib_mw                  *mw;
1202         struct ib_udata                udata;
1203         int                            ret;
1204
1205         if (out_len < sizeof(resp))
1206                 return -ENOSPC;
1207
1208         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1209                 return -EFAULT;
1210
1211         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
1212         if (!uobj)
1213                 return -ENOMEM;
1214
1215         init_uobj(uobj, 0, file->ucontext, &mw_lock_class);
1216         down_write(&uobj->mutex);
1217
1218         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
1219         if (!pd) {
1220                 ret = -EINVAL;
1221                 goto err_free;
1222         }
1223
1224         INIT_UDATA(&udata, buf + sizeof(cmd),
1225                    (unsigned long)cmd.response + sizeof(resp),
1226                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1227                    out_len - sizeof(resp));
1228
1229         mw = pd->device->alloc_mw(pd, cmd.mw_type, &udata);
1230         if (IS_ERR(mw)) {
1231                 ret = PTR_ERR(mw);
1232                 goto err_put;
1233         }
1234
1235         mw->device  = pd->device;
1236         mw->pd      = pd;
1237         mw->uobject = uobj;
1238         atomic_inc(&pd->usecnt);
1239
1240         uobj->object = mw;
1241         ret = idr_add_uobj(&ib_uverbs_mw_idr, uobj);
1242         if (ret)
1243                 goto err_unalloc;
1244
1245         memset(&resp, 0, sizeof(resp));
1246         resp.rkey      = mw->rkey;
1247         resp.mw_handle = uobj->id;
1248
1249         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1250                          &resp, sizeof(resp))) {
1251                 ret = -EFAULT;
1252                 goto err_copy;
1253         }
1254
1255         put_pd_read(pd);
1256
1257         mutex_lock(&file->mutex);
1258         list_add_tail(&uobj->list, &file->ucontext->mw_list);
1259         mutex_unlock(&file->mutex);
1260
1261         uobj->live = 1;
1262
1263         up_write(&uobj->mutex);
1264
1265         return in_len;
1266
1267 err_copy:
1268         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1269
1270 err_unalloc:
1271         uverbs_dealloc_mw(mw);
1272
1273 err_put:
1274         put_pd_read(pd);
1275
1276 err_free:
1277         put_uobj_write(uobj);
1278         return ret;
1279 }
1280
1281 ssize_t ib_uverbs_dealloc_mw(struct ib_uverbs_file *file,
1282                              struct ib_device *ib_dev,
1283                              const char __user *buf, int in_len,
1284                              int out_len)
1285 {
1286         struct ib_uverbs_dealloc_mw cmd;
1287         struct ib_mw               *mw;
1288         struct ib_uobject          *uobj;
1289         int                         ret = -EINVAL;
1290
1291         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1292                 return -EFAULT;
1293
1294         uobj = idr_write_uobj(&ib_uverbs_mw_idr, cmd.mw_handle, file->ucontext);
1295         if (!uobj)
1296                 return -EINVAL;
1297
1298         mw = uobj->object;
1299
1300         ret = uverbs_dealloc_mw(mw);
1301         if (!ret)
1302                 uobj->live = 0;
1303
1304         put_uobj_write(uobj);
1305
1306         if (ret)
1307                 return ret;
1308
1309         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1310
1311         mutex_lock(&file->mutex);
1312         list_del(&uobj->list);
1313         mutex_unlock(&file->mutex);
1314
1315         put_uobj(uobj);
1316
1317         return in_len;
1318 }
1319
1320 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
1321                                       struct ib_device *ib_dev,
1322                                       const char __user *buf, int in_len,
1323                                       int out_len)
1324 {
1325         struct ib_uverbs_create_comp_channel       cmd;
1326         struct ib_uverbs_create_comp_channel_resp  resp;
1327         struct file                               *filp;
1328         int ret;
1329
1330         if (out_len < sizeof resp)
1331                 return -ENOSPC;
1332
1333         if (copy_from_user(&cmd, buf, sizeof cmd))
1334                 return -EFAULT;
1335
1336         ret = get_unused_fd_flags(O_CLOEXEC);
1337         if (ret < 0)
1338                 return ret;
1339         resp.fd = ret;
1340
1341         filp = ib_uverbs_alloc_event_file(file, ib_dev, 0);
1342         if (IS_ERR(filp)) {
1343                 put_unused_fd(resp.fd);
1344                 return PTR_ERR(filp);
1345         }
1346
1347         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1348                          &resp, sizeof resp)) {
1349                 put_unused_fd(resp.fd);
1350                 fput(filp);
1351                 return -EFAULT;
1352         }
1353
1354         fd_install(resp.fd, filp);
1355         return in_len;
1356 }
1357
1358 static struct ib_ucq_object *create_cq(struct ib_uverbs_file *file,
1359                                         struct ib_device *ib_dev,
1360                                        struct ib_udata *ucore,
1361                                        struct ib_udata *uhw,
1362                                        struct ib_uverbs_ex_create_cq *cmd,
1363                                        size_t cmd_sz,
1364                                        int (*cb)(struct ib_uverbs_file *file,
1365                                                  struct ib_ucq_object *obj,
1366                                                  struct ib_uverbs_ex_create_cq_resp *resp,
1367                                                  struct ib_udata *udata,
1368                                                  void *context),
1369                                        void *context)
1370 {
1371         struct ib_ucq_object           *obj;
1372         struct ib_uverbs_event_file    *ev_file = NULL;
1373         struct ib_cq                   *cq;
1374         int                             ret;
1375         struct ib_uverbs_ex_create_cq_resp resp;
1376         struct ib_cq_init_attr attr = {};
1377
1378         if (cmd->comp_vector >= file->device->num_comp_vectors)
1379                 return ERR_PTR(-EINVAL);
1380
1381         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1382         if (!obj)
1383                 return ERR_PTR(-ENOMEM);
1384
1385         init_uobj(&obj->uobject, cmd->user_handle, file->ucontext, &cq_lock_class);
1386         down_write(&obj->uobject.mutex);
1387
1388         if (cmd->comp_channel >= 0) {
1389                 ev_file = ib_uverbs_lookup_comp_file(cmd->comp_channel);
1390                 if (!ev_file) {
1391                         ret = -EINVAL;
1392                         goto err;
1393                 }
1394         }
1395
1396         obj->uverbs_file           = file;
1397         obj->comp_events_reported  = 0;
1398         obj->async_events_reported = 0;
1399         INIT_LIST_HEAD(&obj->comp_list);
1400         INIT_LIST_HEAD(&obj->async_list);
1401
1402         attr.cqe = cmd->cqe;
1403         attr.comp_vector = cmd->comp_vector;
1404
1405         if (cmd_sz > offsetof(typeof(*cmd), flags) + sizeof(cmd->flags))
1406                 attr.flags = cmd->flags;
1407
1408         cq = ib_dev->create_cq(ib_dev, &attr,
1409                                              file->ucontext, uhw);
1410         if (IS_ERR(cq)) {
1411                 ret = PTR_ERR(cq);
1412                 goto err_file;
1413         }
1414
1415         cq->device        = ib_dev;
1416         cq->uobject       = &obj->uobject;
1417         cq->comp_handler  = ib_uverbs_comp_handler;
1418         cq->event_handler = ib_uverbs_cq_event_handler;
1419         cq->cq_context    = ev_file;
1420         atomic_set(&cq->usecnt, 0);
1421
1422         obj->uobject.object = cq;
1423         ret = idr_add_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1424         if (ret)
1425                 goto err_free;
1426
1427         memset(&resp, 0, sizeof resp);
1428         resp.base.cq_handle = obj->uobject.id;
1429         resp.base.cqe       = cq->cqe;
1430
1431         resp.response_length = offsetof(typeof(resp), response_length) +
1432                 sizeof(resp.response_length);
1433
1434         ret = cb(file, obj, &resp, ucore, context);
1435         if (ret)
1436                 goto err_cb;
1437
1438         mutex_lock(&file->mutex);
1439         list_add_tail(&obj->uobject.list, &file->ucontext->cq_list);
1440         mutex_unlock(&file->mutex);
1441
1442         obj->uobject.live = 1;
1443
1444         up_write(&obj->uobject.mutex);
1445
1446         return obj;
1447
1448 err_cb:
1449         idr_remove_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1450
1451 err_free:
1452         ib_destroy_cq(cq);
1453
1454 err_file:
1455         if (ev_file)
1456                 ib_uverbs_release_ucq(file, ev_file, obj);
1457
1458 err:
1459         put_uobj_write(&obj->uobject);
1460
1461         return ERR_PTR(ret);
1462 }
1463
1464 static int ib_uverbs_create_cq_cb(struct ib_uverbs_file *file,
1465                                   struct ib_ucq_object *obj,
1466                                   struct ib_uverbs_ex_create_cq_resp *resp,
1467                                   struct ib_udata *ucore, void *context)
1468 {
1469         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1470                 return -EFAULT;
1471
1472         return 0;
1473 }
1474
1475 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1476                             struct ib_device *ib_dev,
1477                             const char __user *buf, int in_len,
1478                             int out_len)
1479 {
1480         struct ib_uverbs_create_cq      cmd;
1481         struct ib_uverbs_ex_create_cq   cmd_ex;
1482         struct ib_uverbs_create_cq_resp resp;
1483         struct ib_udata                 ucore;
1484         struct ib_udata                 uhw;
1485         struct ib_ucq_object           *obj;
1486
1487         if (out_len < sizeof(resp))
1488                 return -ENOSPC;
1489
1490         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1491                 return -EFAULT;
1492
1493         INIT_UDATA(&ucore, buf, (unsigned long)cmd.response, sizeof(cmd), sizeof(resp));
1494
1495         INIT_UDATA(&uhw, buf + sizeof(cmd),
1496                    (unsigned long)cmd.response + sizeof(resp),
1497                    in_len - sizeof(cmd), out_len - sizeof(resp));
1498
1499         memset(&cmd_ex, 0, sizeof(cmd_ex));
1500         cmd_ex.user_handle = cmd.user_handle;
1501         cmd_ex.cqe = cmd.cqe;
1502         cmd_ex.comp_vector = cmd.comp_vector;
1503         cmd_ex.comp_channel = cmd.comp_channel;
1504
1505         obj = create_cq(file, ib_dev, &ucore, &uhw, &cmd_ex,
1506                         offsetof(typeof(cmd_ex), comp_channel) +
1507                         sizeof(cmd.comp_channel), ib_uverbs_create_cq_cb,
1508                         NULL);
1509
1510         if (IS_ERR(obj))
1511                 return PTR_ERR(obj);
1512
1513         return in_len;
1514 }
1515
1516 static int ib_uverbs_ex_create_cq_cb(struct ib_uverbs_file *file,
1517                                      struct ib_ucq_object *obj,
1518                                      struct ib_uverbs_ex_create_cq_resp *resp,
1519                                      struct ib_udata *ucore, void *context)
1520 {
1521         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1522                 return -EFAULT;
1523
1524         return 0;
1525 }
1526
1527 int ib_uverbs_ex_create_cq(struct ib_uverbs_file *file,
1528                          struct ib_device *ib_dev,
1529                            struct ib_udata *ucore,
1530                            struct ib_udata *uhw)
1531 {
1532         struct ib_uverbs_ex_create_cq_resp resp;
1533         struct ib_uverbs_ex_create_cq  cmd;
1534         struct ib_ucq_object           *obj;
1535         int err;
1536
1537         if (ucore->inlen < sizeof(cmd))
1538                 return -EINVAL;
1539
1540         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
1541         if (err)
1542                 return err;
1543
1544         if (cmd.comp_mask)
1545                 return -EINVAL;
1546
1547         if (cmd.reserved)
1548                 return -EINVAL;
1549
1550         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1551                              sizeof(resp.response_length)))
1552                 return -ENOSPC;
1553
1554         obj = create_cq(file, ib_dev, ucore, uhw, &cmd,
1555                         min(ucore->inlen, sizeof(cmd)),
1556                         ib_uverbs_ex_create_cq_cb, NULL);
1557
1558         if (IS_ERR(obj))
1559                 return PTR_ERR(obj);
1560
1561         return 0;
1562 }
1563
1564 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1565                             struct ib_device *ib_dev,
1566                             const char __user *buf, int in_len,
1567                             int out_len)
1568 {
1569         struct ib_uverbs_resize_cq      cmd;
1570         struct ib_uverbs_resize_cq_resp resp;
1571         struct ib_udata                 udata;
1572         struct ib_cq                    *cq;
1573         int                             ret = -EINVAL;
1574
1575         if (copy_from_user(&cmd, buf, sizeof cmd))
1576                 return -EFAULT;
1577
1578         INIT_UDATA(&udata, buf + sizeof cmd,
1579                    (unsigned long) cmd.response + sizeof resp,
1580                    in_len - sizeof cmd, out_len - sizeof resp);
1581
1582         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1583         if (!cq)
1584                 return -EINVAL;
1585
1586         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1587         if (ret)
1588                 goto out;
1589
1590         resp.cqe = cq->cqe;
1591
1592         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1593                          &resp, sizeof resp.cqe))
1594                 ret = -EFAULT;
1595
1596 out:
1597         put_cq_read(cq);
1598
1599         return ret ? ret : in_len;
1600 }
1601
1602 static int copy_wc_to_user(void __user *dest, struct ib_wc *wc)
1603 {
1604         struct ib_uverbs_wc tmp;
1605
1606         tmp.wr_id               = wc->wr_id;
1607         tmp.status              = wc->status;
1608         tmp.opcode              = wc->opcode;
1609         tmp.vendor_err          = wc->vendor_err;
1610         tmp.byte_len            = wc->byte_len;
1611         tmp.ex.imm_data         = (__u32 __force) wc->ex.imm_data;
1612         tmp.qp_num              = wc->qp->qp_num;
1613         tmp.src_qp              = wc->src_qp;
1614         tmp.wc_flags            = wc->wc_flags;
1615         tmp.pkey_index          = wc->pkey_index;
1616         tmp.slid                = wc->slid;
1617         tmp.sl                  = wc->sl;
1618         tmp.dlid_path_bits      = wc->dlid_path_bits;
1619         tmp.port_num            = wc->port_num;
1620         tmp.reserved            = 0;
1621
1622         if (copy_to_user(dest, &tmp, sizeof tmp))
1623                 return -EFAULT;
1624
1625         return 0;
1626 }
1627
1628 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1629                           struct ib_device *ib_dev,
1630                           const char __user *buf, int in_len,
1631                           int out_len)
1632 {
1633         struct ib_uverbs_poll_cq       cmd;
1634         struct ib_uverbs_poll_cq_resp  resp;
1635         u8 __user                     *header_ptr;
1636         u8 __user                     *data_ptr;
1637         struct ib_cq                  *cq;
1638         struct ib_wc                   wc;
1639         int                            ret;
1640
1641         if (copy_from_user(&cmd, buf, sizeof cmd))
1642                 return -EFAULT;
1643
1644         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1645         if (!cq)
1646                 return -EINVAL;
1647
1648         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1649         header_ptr = (void __user *)(unsigned long) cmd.response;
1650         data_ptr = header_ptr + sizeof resp;
1651
1652         memset(&resp, 0, sizeof resp);
1653         while (resp.count < cmd.ne) {
1654                 ret = ib_poll_cq(cq, 1, &wc);
1655                 if (ret < 0)
1656                         goto out_put;
1657                 if (!ret)
1658                         break;
1659
1660                 ret = copy_wc_to_user(data_ptr, &wc);
1661                 if (ret)
1662                         goto out_put;
1663
1664                 data_ptr += sizeof(struct ib_uverbs_wc);
1665                 ++resp.count;
1666         }
1667
1668         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1669                 ret = -EFAULT;
1670                 goto out_put;
1671         }
1672
1673         ret = in_len;
1674
1675 out_put:
1676         put_cq_read(cq);
1677         return ret;
1678 }
1679
1680 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1681                                 struct ib_device *ib_dev,
1682                                 const char __user *buf, int in_len,
1683                                 int out_len)
1684 {
1685         struct ib_uverbs_req_notify_cq cmd;
1686         struct ib_cq                  *cq;
1687
1688         if (copy_from_user(&cmd, buf, sizeof cmd))
1689                 return -EFAULT;
1690
1691         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1692         if (!cq)
1693                 return -EINVAL;
1694
1695         ib_req_notify_cq(cq, cmd.solicited_only ?
1696                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1697
1698         put_cq_read(cq);
1699
1700         return in_len;
1701 }
1702
1703 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1704                              struct ib_device *ib_dev,
1705                              const char __user *buf, int in_len,
1706                              int out_len)
1707 {
1708         struct ib_uverbs_destroy_cq      cmd;
1709         struct ib_uverbs_destroy_cq_resp resp;
1710         struct ib_uobject               *uobj;
1711         struct ib_cq                    *cq;
1712         struct ib_ucq_object            *obj;
1713         struct ib_uverbs_event_file     *ev_file;
1714         int                              ret = -EINVAL;
1715
1716         if (copy_from_user(&cmd, buf, sizeof cmd))
1717                 return -EFAULT;
1718
1719         uobj = idr_write_uobj(&ib_uverbs_cq_idr, cmd.cq_handle, file->ucontext);
1720         if (!uobj)
1721                 return -EINVAL;
1722         cq      = uobj->object;
1723         ev_file = cq->cq_context;
1724         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1725
1726         ret = ib_destroy_cq(cq);
1727         if (!ret)
1728                 uobj->live = 0;
1729
1730         put_uobj_write(uobj);
1731
1732         if (ret)
1733                 return ret;
1734
1735         idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
1736
1737         mutex_lock(&file->mutex);
1738         list_del(&uobj->list);
1739         mutex_unlock(&file->mutex);
1740
1741         ib_uverbs_release_ucq(file, ev_file, obj);
1742
1743         memset(&resp, 0, sizeof resp);
1744         resp.comp_events_reported  = obj->comp_events_reported;
1745         resp.async_events_reported = obj->async_events_reported;
1746
1747         put_uobj(uobj);
1748
1749         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1750                          &resp, sizeof resp))
1751                 return -EFAULT;
1752
1753         return in_len;
1754 }
1755
1756 static int create_qp(struct ib_uverbs_file *file,
1757                      struct ib_udata *ucore,
1758                      struct ib_udata *uhw,
1759                      struct ib_uverbs_ex_create_qp *cmd,
1760                      size_t cmd_sz,
1761                      int (*cb)(struct ib_uverbs_file *file,
1762                                struct ib_uverbs_ex_create_qp_resp *resp,
1763                                struct ib_udata *udata),
1764                      void *context)
1765 {
1766         struct ib_uqp_object            *obj;
1767         struct ib_device                *device;
1768         struct ib_pd                    *pd = NULL;
1769         struct ib_xrcd                  *xrcd = NULL;
1770         struct ib_uobject               *uninitialized_var(xrcd_uobj);
1771         struct ib_cq                    *scq = NULL, *rcq = NULL;
1772         struct ib_srq                   *srq = NULL;
1773         struct ib_qp                    *qp;
1774         char                            *buf;
1775         struct ib_qp_init_attr          attr = {};
1776         struct ib_uverbs_ex_create_qp_resp resp;
1777         int                             ret;
1778         struct ib_rwq_ind_table *ind_tbl = NULL;
1779         bool has_sq = true;
1780
1781         if (cmd->qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1782                 return -EPERM;
1783
1784         obj = kzalloc(sizeof *obj, GFP_KERNEL);
1785         if (!obj)
1786                 return -ENOMEM;
1787
1788         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext,
1789                   &qp_lock_class);
1790         down_write(&obj->uevent.uobject.mutex);
1791         if (cmd_sz >= offsetof(typeof(*cmd), rwq_ind_tbl_handle) +
1792                       sizeof(cmd->rwq_ind_tbl_handle) &&
1793                       (cmd->comp_mask & IB_UVERBS_CREATE_QP_MASK_IND_TABLE)) {
1794                 ind_tbl = idr_read_rwq_indirection_table(cmd->rwq_ind_tbl_handle,
1795                                                          file->ucontext);
1796                 if (!ind_tbl) {
1797                         ret = -EINVAL;
1798                         goto err_put;
1799                 }
1800
1801                 attr.rwq_ind_tbl = ind_tbl;
1802         }
1803
1804         if ((cmd_sz >= offsetof(typeof(*cmd), reserved1) +
1805                        sizeof(cmd->reserved1)) && cmd->reserved1) {
1806                 ret = -EOPNOTSUPP;
1807                 goto err_put;
1808         }
1809
1810         if (ind_tbl && (cmd->max_recv_wr || cmd->max_recv_sge || cmd->is_srq)) {
1811                 ret = -EINVAL;
1812                 goto err_put;
1813         }
1814
1815         if (ind_tbl && !cmd->max_send_wr)
1816                 has_sq = false;
1817
1818         if (cmd->qp_type == IB_QPT_XRC_TGT) {
1819                 xrcd = idr_read_xrcd(cmd->pd_handle, file->ucontext,
1820                                      &xrcd_uobj);
1821                 if (!xrcd) {
1822                         ret = -EINVAL;
1823                         goto err_put;
1824                 }
1825                 device = xrcd->device;
1826         } else {
1827                 if (cmd->qp_type == IB_QPT_XRC_INI) {
1828                         cmd->max_recv_wr = 0;
1829                         cmd->max_recv_sge = 0;
1830                 } else {
1831                         if (cmd->is_srq) {
1832                                 srq = idr_read_srq(cmd->srq_handle,
1833                                                    file->ucontext);
1834                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1835                                         ret = -EINVAL;
1836                                         goto err_put;
1837                                 }
1838                         }
1839
1840                         if (!ind_tbl) {
1841                                 if (cmd->recv_cq_handle != cmd->send_cq_handle) {
1842                                         rcq = idr_read_cq(cmd->recv_cq_handle,
1843                                                           file->ucontext, 0);
1844                                         if (!rcq) {
1845                                                 ret = -EINVAL;
1846                                                 goto err_put;
1847                                         }
1848                                 }
1849                         }
1850                 }
1851
1852                 if (has_sq)
1853                         scq = idr_read_cq(cmd->send_cq_handle, file->ucontext, !!rcq);
1854                 if (!ind_tbl)
1855                         rcq = rcq ?: scq;
1856                 pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
1857                 if (!pd || (!scq && has_sq)) {
1858                         ret = -EINVAL;
1859                         goto err_put;
1860                 }
1861
1862                 device = pd->device;
1863         }
1864
1865         attr.event_handler = ib_uverbs_qp_event_handler;
1866         attr.qp_context    = file;
1867         attr.send_cq       = scq;
1868         attr.recv_cq       = rcq;
1869         attr.srq           = srq;
1870         attr.xrcd          = xrcd;
1871         attr.sq_sig_type   = cmd->sq_sig_all ? IB_SIGNAL_ALL_WR :
1872                                               IB_SIGNAL_REQ_WR;
1873         attr.qp_type       = cmd->qp_type;
1874         attr.create_flags  = 0;
1875
1876         attr.cap.max_send_wr     = cmd->max_send_wr;
1877         attr.cap.max_recv_wr     = cmd->max_recv_wr;
1878         attr.cap.max_send_sge    = cmd->max_send_sge;
1879         attr.cap.max_recv_sge    = cmd->max_recv_sge;
1880         attr.cap.max_inline_data = cmd->max_inline_data;
1881
1882         obj->uevent.events_reported     = 0;
1883         INIT_LIST_HEAD(&obj->uevent.event_list);
1884         INIT_LIST_HEAD(&obj->mcast_list);
1885
1886         if (cmd_sz >= offsetof(typeof(*cmd), create_flags) +
1887                       sizeof(cmd->create_flags))
1888                 attr.create_flags = cmd->create_flags;
1889
1890         if (attr.create_flags & ~(IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
1891                                 IB_QP_CREATE_CROSS_CHANNEL |
1892                                 IB_QP_CREATE_MANAGED_SEND |
1893                                 IB_QP_CREATE_MANAGED_RECV |
1894                                 IB_QP_CREATE_SCATTER_FCS |
1895                                 IB_QP_CREATE_CVLAN_STRIPPING)) {
1896                 ret = -EINVAL;
1897                 goto err_put;
1898         }
1899
1900         buf = (void *)cmd + sizeof(*cmd);
1901         if (cmd_sz > sizeof(*cmd))
1902                 if (!(buf[0] == 0 && !memcmp(buf, buf + 1,
1903                                              cmd_sz - sizeof(*cmd) - 1))) {
1904                         ret = -EINVAL;
1905                         goto err_put;
1906                 }
1907
1908         if (cmd->qp_type == IB_QPT_XRC_TGT)
1909                 qp = ib_create_qp(pd, &attr);
1910         else
1911                 qp = device->create_qp(pd, &attr, uhw);
1912
1913         if (IS_ERR(qp)) {
1914                 ret = PTR_ERR(qp);
1915                 goto err_put;
1916         }
1917
1918         if (cmd->qp_type != IB_QPT_XRC_TGT) {
1919                 qp->real_qp       = qp;
1920                 qp->device        = device;
1921                 qp->pd            = pd;
1922                 qp->send_cq       = attr.send_cq;
1923                 qp->recv_cq       = attr.recv_cq;
1924                 qp->srq           = attr.srq;
1925                 qp->rwq_ind_tbl   = ind_tbl;
1926                 qp->event_handler = attr.event_handler;
1927                 qp->qp_context    = attr.qp_context;
1928                 qp->qp_type       = attr.qp_type;
1929                 atomic_set(&qp->usecnt, 0);
1930                 atomic_inc(&pd->usecnt);
1931                 if (attr.send_cq)
1932                         atomic_inc(&attr.send_cq->usecnt);
1933                 if (attr.recv_cq)
1934                         atomic_inc(&attr.recv_cq->usecnt);
1935                 if (attr.srq)
1936                         atomic_inc(&attr.srq->usecnt);
1937                 if (ind_tbl)
1938                         atomic_inc(&ind_tbl->usecnt);
1939         }
1940         qp->uobject = &obj->uevent.uobject;
1941
1942         obj->uevent.uobject.object = qp;
1943         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1944         if (ret)
1945                 goto err_destroy;
1946
1947         memset(&resp, 0, sizeof resp);
1948         resp.base.qpn             = qp->qp_num;
1949         resp.base.qp_handle       = obj->uevent.uobject.id;
1950         resp.base.max_recv_sge    = attr.cap.max_recv_sge;
1951         resp.base.max_send_sge    = attr.cap.max_send_sge;
1952         resp.base.max_recv_wr     = attr.cap.max_recv_wr;
1953         resp.base.max_send_wr     = attr.cap.max_send_wr;
1954         resp.base.max_inline_data = attr.cap.max_inline_data;
1955
1956         resp.response_length = offsetof(typeof(resp), response_length) +
1957                                sizeof(resp.response_length);
1958
1959         ret = cb(file, &resp, ucore);
1960         if (ret)
1961                 goto err_cb;
1962
1963         if (xrcd) {
1964                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object,
1965                                           uobject);
1966                 atomic_inc(&obj->uxrcd->refcnt);
1967                 put_xrcd_read(xrcd_uobj);
1968         }
1969
1970         if (pd)
1971                 put_pd_read(pd);
1972         if (scq)
1973                 put_cq_read(scq);
1974         if (rcq && rcq != scq)
1975                 put_cq_read(rcq);
1976         if (srq)
1977                 put_srq_read(srq);
1978         if (ind_tbl)
1979                 put_rwq_indirection_table_read(ind_tbl);
1980
1981         mutex_lock(&file->mutex);
1982         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1983         mutex_unlock(&file->mutex);
1984
1985         obj->uevent.uobject.live = 1;
1986
1987         up_write(&obj->uevent.uobject.mutex);
1988
1989         return 0;
1990 err_cb:
1991         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1992
1993 err_destroy:
1994         ib_destroy_qp(qp);
1995
1996 err_put:
1997         if (xrcd)
1998                 put_xrcd_read(xrcd_uobj);
1999         if (pd)
2000                 put_pd_read(pd);
2001         if (scq)
2002                 put_cq_read(scq);
2003         if (rcq && rcq != scq)
2004                 put_cq_read(rcq);
2005         if (srq)
2006                 put_srq_read(srq);
2007         if (ind_tbl)
2008                 put_rwq_indirection_table_read(ind_tbl);
2009
2010         put_uobj_write(&obj->uevent.uobject);
2011         return ret;
2012 }
2013
2014 static int ib_uverbs_create_qp_cb(struct ib_uverbs_file *file,
2015                                   struct ib_uverbs_ex_create_qp_resp *resp,
2016                                   struct ib_udata *ucore)
2017 {
2018         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
2019                 return -EFAULT;
2020
2021         return 0;
2022 }
2023
2024 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
2025                             struct ib_device *ib_dev,
2026                             const char __user *buf, int in_len,
2027                             int out_len)
2028 {
2029         struct ib_uverbs_create_qp      cmd;
2030         struct ib_uverbs_ex_create_qp   cmd_ex;
2031         struct ib_udata                 ucore;
2032         struct ib_udata                 uhw;
2033         ssize_t resp_size = sizeof(struct ib_uverbs_create_qp_resp);
2034         int                             err;
2035
2036         if (out_len < resp_size)
2037                 return -ENOSPC;
2038
2039         if (copy_from_user(&cmd, buf, sizeof(cmd)))
2040                 return -EFAULT;
2041
2042         INIT_UDATA(&ucore, buf, (unsigned long)cmd.response, sizeof(cmd),
2043                    resp_size);
2044         INIT_UDATA(&uhw, buf + sizeof(cmd),
2045                    (unsigned long)cmd.response + resp_size,
2046                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
2047                    out_len - resp_size);
2048
2049         memset(&cmd_ex, 0, sizeof(cmd_ex));
2050         cmd_ex.user_handle = cmd.user_handle;
2051         cmd_ex.pd_handle = cmd.pd_handle;
2052         cmd_ex.send_cq_handle = cmd.send_cq_handle;
2053         cmd_ex.recv_cq_handle = cmd.recv_cq_handle;
2054         cmd_ex.srq_handle = cmd.srq_handle;
2055         cmd_ex.max_send_wr = cmd.max_send_wr;
2056         cmd_ex.max_recv_wr = cmd.max_recv_wr;
2057         cmd_ex.max_send_sge = cmd.max_send_sge;
2058         cmd_ex.max_recv_sge = cmd.max_recv_sge;
2059         cmd_ex.max_inline_data = cmd.max_inline_data;
2060         cmd_ex.sq_sig_all = cmd.sq_sig_all;
2061         cmd_ex.qp_type = cmd.qp_type;
2062         cmd_ex.is_srq = cmd.is_srq;
2063
2064         err = create_qp(file, &ucore, &uhw, &cmd_ex,
2065                         offsetof(typeof(cmd_ex), is_srq) +
2066                         sizeof(cmd.is_srq), ib_uverbs_create_qp_cb,
2067                         NULL);
2068
2069         if (err)
2070                 return err;
2071
2072         return in_len;
2073 }
2074
2075 static int ib_uverbs_ex_create_qp_cb(struct ib_uverbs_file *file,
2076                                      struct ib_uverbs_ex_create_qp_resp *resp,
2077                                      struct ib_udata *ucore)
2078 {
2079         if (ib_copy_to_udata(ucore, resp, resp->response_length))
2080                 return -EFAULT;
2081
2082         return 0;
2083 }
2084
2085 int ib_uverbs_ex_create_qp(struct ib_uverbs_file *file,
2086                            struct ib_device *ib_dev,
2087                            struct ib_udata *ucore,
2088                            struct ib_udata *uhw)
2089 {
2090         struct ib_uverbs_ex_create_qp_resp resp;
2091         struct ib_uverbs_ex_create_qp cmd = {0};
2092         int err;
2093
2094         if (ucore->inlen < (offsetof(typeof(cmd), comp_mask) +
2095                             sizeof(cmd.comp_mask)))
2096                 return -EINVAL;
2097
2098         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2099         if (err)
2100                 return err;
2101
2102         if (cmd.comp_mask & ~IB_UVERBS_CREATE_QP_SUP_COMP_MASK)
2103                 return -EINVAL;
2104
2105         if (cmd.reserved)
2106                 return -EINVAL;
2107
2108         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
2109                              sizeof(resp.response_length)))
2110                 return -ENOSPC;
2111
2112         err = create_qp(file, ucore, uhw, &cmd,
2113                         min(ucore->inlen, sizeof(cmd)),
2114                         ib_uverbs_ex_create_qp_cb, NULL);
2115
2116         if (err)
2117                 return err;
2118
2119         return 0;
2120 }
2121
2122 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
2123                           struct ib_device *ib_dev,
2124                           const char __user *buf, int in_len, int out_len)
2125 {
2126         struct ib_uverbs_open_qp        cmd;
2127         struct ib_uverbs_create_qp_resp resp;
2128         struct ib_udata                 udata;
2129         struct ib_uqp_object           *obj;
2130         struct ib_xrcd                 *xrcd;
2131         struct ib_uobject              *uninitialized_var(xrcd_uobj);
2132         struct ib_qp                   *qp;
2133         struct ib_qp_open_attr          attr;
2134         int ret;
2135
2136         if (out_len < sizeof resp)
2137                 return -ENOSPC;
2138
2139         if (copy_from_user(&cmd, buf, sizeof cmd))
2140                 return -EFAULT;
2141
2142         INIT_UDATA(&udata, buf + sizeof cmd,
2143                    (unsigned long) cmd.response + sizeof resp,
2144                    in_len - sizeof cmd, out_len - sizeof resp);
2145
2146         obj = kmalloc(sizeof *obj, GFP_KERNEL);
2147         if (!obj)
2148                 return -ENOMEM;
2149
2150         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
2151         down_write(&obj->uevent.uobject.mutex);
2152
2153         xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
2154         if (!xrcd) {
2155                 ret = -EINVAL;
2156                 goto err_put;
2157         }
2158
2159         attr.event_handler = ib_uverbs_qp_event_handler;
2160         attr.qp_context    = file;
2161         attr.qp_num        = cmd.qpn;
2162         attr.qp_type       = cmd.qp_type;
2163
2164         obj->uevent.events_reported = 0;
2165         INIT_LIST_HEAD(&obj->uevent.event_list);
2166         INIT_LIST_HEAD(&obj->mcast_list);
2167
2168         qp = ib_open_qp(xrcd, &attr);
2169         if (IS_ERR(qp)) {
2170                 ret = PTR_ERR(qp);
2171                 goto err_put;
2172         }
2173
2174         qp->uobject = &obj->uevent.uobject;
2175
2176         obj->uevent.uobject.object = qp;
2177         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
2178         if (ret)
2179                 goto err_destroy;
2180
2181         memset(&resp, 0, sizeof resp);
2182         resp.qpn       = qp->qp_num;
2183         resp.qp_handle = obj->uevent.uobject.id;
2184
2185         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2186                          &resp, sizeof resp)) {
2187                 ret = -EFAULT;
2188                 goto err_remove;
2189         }
2190
2191         obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
2192         atomic_inc(&obj->uxrcd->refcnt);
2193         put_xrcd_read(xrcd_uobj);
2194
2195         mutex_lock(&file->mutex);
2196         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
2197         mutex_unlock(&file->mutex);
2198
2199         obj->uevent.uobject.live = 1;
2200
2201         up_write(&obj->uevent.uobject.mutex);
2202
2203         return in_len;
2204
2205 err_remove:
2206         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
2207
2208 err_destroy:
2209         ib_destroy_qp(qp);
2210
2211 err_put:
2212         put_xrcd_read(xrcd_uobj);
2213         put_uobj_write(&obj->uevent.uobject);
2214         return ret;
2215 }
2216
2217 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
2218                            struct ib_device *ib_dev,
2219                            const char __user *buf, int in_len,
2220                            int out_len)
2221 {
2222         struct ib_uverbs_query_qp      cmd;
2223         struct ib_uverbs_query_qp_resp resp;
2224         struct ib_qp                   *qp;
2225         struct ib_qp_attr              *attr;
2226         struct ib_qp_init_attr         *init_attr;
2227         int                            ret;
2228
2229         if (copy_from_user(&cmd, buf, sizeof cmd))
2230                 return -EFAULT;
2231
2232         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
2233         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
2234         if (!attr || !init_attr) {
2235                 ret = -ENOMEM;
2236                 goto out;
2237         }
2238
2239         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2240         if (!qp) {
2241                 ret = -EINVAL;
2242                 goto out;
2243         }
2244
2245         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
2246
2247         put_qp_read(qp);
2248
2249         if (ret)
2250                 goto out;
2251
2252         memset(&resp, 0, sizeof resp);
2253
2254         resp.qp_state               = attr->qp_state;
2255         resp.cur_qp_state           = attr->cur_qp_state;
2256         resp.path_mtu               = attr->path_mtu;
2257         resp.path_mig_state         = attr->path_mig_state;
2258         resp.qkey                   = attr->qkey;
2259         resp.rq_psn                 = attr->rq_psn;
2260         resp.sq_psn                 = attr->sq_psn;
2261         resp.dest_qp_num            = attr->dest_qp_num;
2262         resp.qp_access_flags        = attr->qp_access_flags;
2263         resp.pkey_index             = attr->pkey_index;
2264         resp.alt_pkey_index         = attr->alt_pkey_index;
2265         resp.sq_draining            = attr->sq_draining;
2266         resp.max_rd_atomic          = attr->max_rd_atomic;
2267         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
2268         resp.min_rnr_timer          = attr->min_rnr_timer;
2269         resp.port_num               = attr->port_num;
2270         resp.timeout                = attr->timeout;
2271         resp.retry_cnt              = attr->retry_cnt;
2272         resp.rnr_retry              = attr->rnr_retry;
2273         resp.alt_port_num           = attr->alt_port_num;
2274         resp.alt_timeout            = attr->alt_timeout;
2275
2276         memcpy(resp.dest.dgid, attr->ah_attr.grh.dgid.raw, 16);
2277         resp.dest.flow_label        = attr->ah_attr.grh.flow_label;
2278         resp.dest.sgid_index        = attr->ah_attr.grh.sgid_index;
2279         resp.dest.hop_limit         = attr->ah_attr.grh.hop_limit;
2280         resp.dest.traffic_class     = attr->ah_attr.grh.traffic_class;
2281         resp.dest.dlid              = attr->ah_attr.dlid;
2282         resp.dest.sl                = attr->ah_attr.sl;
2283         resp.dest.src_path_bits     = attr->ah_attr.src_path_bits;
2284         resp.dest.static_rate       = attr->ah_attr.static_rate;
2285         resp.dest.is_global         = !!(attr->ah_attr.ah_flags & IB_AH_GRH);
2286         resp.dest.port_num          = attr->ah_attr.port_num;
2287
2288         memcpy(resp.alt_dest.dgid, attr->alt_ah_attr.grh.dgid.raw, 16);
2289         resp.alt_dest.flow_label    = attr->alt_ah_attr.grh.flow_label;
2290         resp.alt_dest.sgid_index    = attr->alt_ah_attr.grh.sgid_index;
2291         resp.alt_dest.hop_limit     = attr->alt_ah_attr.grh.hop_limit;
2292         resp.alt_dest.traffic_class = attr->alt_ah_attr.grh.traffic_class;
2293         resp.alt_dest.dlid          = attr->alt_ah_attr.dlid;
2294         resp.alt_dest.sl            = attr->alt_ah_attr.sl;
2295         resp.alt_dest.src_path_bits = attr->alt_ah_attr.src_path_bits;
2296         resp.alt_dest.static_rate   = attr->alt_ah_attr.static_rate;
2297         resp.alt_dest.is_global     = !!(attr->alt_ah_attr.ah_flags & IB_AH_GRH);
2298         resp.alt_dest.port_num      = attr->alt_ah_attr.port_num;
2299
2300         resp.max_send_wr            = init_attr->cap.max_send_wr;
2301         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
2302         resp.max_send_sge           = init_attr->cap.max_send_sge;
2303         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
2304         resp.max_inline_data        = init_attr->cap.max_inline_data;
2305         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
2306
2307         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2308                          &resp, sizeof resp))
2309                 ret = -EFAULT;
2310
2311 out:
2312         kfree(attr);
2313         kfree(init_attr);
2314
2315         return ret ? ret : in_len;
2316 }
2317
2318 /* Remove ignored fields set in the attribute mask */
2319 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
2320 {
2321         switch (qp_type) {
2322         case IB_QPT_XRC_INI:
2323                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
2324         case IB_QPT_XRC_TGT:
2325                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
2326                                 IB_QP_RNR_RETRY);
2327         default:
2328                 return mask;
2329         }
2330 }
2331
2332 static int modify_qp(struct ib_uverbs_file *file,
2333                      struct ib_uverbs_ex_modify_qp *cmd, struct ib_udata *udata)
2334 {
2335         struct ib_qp_attr *attr;
2336         struct ib_qp *qp;
2337         int ret;
2338
2339         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2340         if (!attr)
2341                 return -ENOMEM;
2342
2343         qp = idr_read_qp(cmd->base.qp_handle, file->ucontext);
2344         if (!qp) {
2345                 ret = -EINVAL;
2346                 goto out;
2347         }
2348
2349         attr->qp_state            = cmd->base.qp_state;
2350         attr->cur_qp_state        = cmd->base.cur_qp_state;
2351         attr->path_mtu            = cmd->base.path_mtu;
2352         attr->path_mig_state      = cmd->base.path_mig_state;
2353         attr->qkey                = cmd->base.qkey;
2354         attr->rq_psn              = cmd->base.rq_psn;
2355         attr->sq_psn              = cmd->base.sq_psn;
2356         attr->dest_qp_num         = cmd->base.dest_qp_num;
2357         attr->qp_access_flags     = cmd->base.qp_access_flags;
2358         attr->pkey_index          = cmd->base.pkey_index;
2359         attr->alt_pkey_index      = cmd->base.alt_pkey_index;
2360         attr->en_sqd_async_notify = cmd->base.en_sqd_async_notify;
2361         attr->max_rd_atomic       = cmd->base.max_rd_atomic;
2362         attr->max_dest_rd_atomic  = cmd->base.max_dest_rd_atomic;
2363         attr->min_rnr_timer       = cmd->base.min_rnr_timer;
2364         attr->port_num            = cmd->base.port_num;
2365         attr->timeout             = cmd->base.timeout;
2366         attr->retry_cnt           = cmd->base.retry_cnt;
2367         attr->rnr_retry           = cmd->base.rnr_retry;
2368         attr->alt_port_num        = cmd->base.alt_port_num;
2369         attr->alt_timeout         = cmd->base.alt_timeout;
2370         attr->rate_limit          = cmd->rate_limit;
2371
2372         memcpy(attr->ah_attr.grh.dgid.raw, cmd->base.dest.dgid, 16);
2373         attr->ah_attr.grh.flow_label    = cmd->base.dest.flow_label;
2374         attr->ah_attr.grh.sgid_index    = cmd->base.dest.sgid_index;
2375         attr->ah_attr.grh.hop_limit     = cmd->base.dest.hop_limit;
2376         attr->ah_attr.grh.traffic_class = cmd->base.dest.traffic_class;
2377         attr->ah_attr.dlid              = cmd->base.dest.dlid;
2378         attr->ah_attr.sl                = cmd->base.dest.sl;
2379         attr->ah_attr.src_path_bits     = cmd->base.dest.src_path_bits;
2380         attr->ah_attr.static_rate       = cmd->base.dest.static_rate;
2381         attr->ah_attr.ah_flags          = cmd->base.dest.is_global ?
2382                                           IB_AH_GRH : 0;
2383         attr->ah_attr.port_num          = cmd->base.dest.port_num;
2384
2385         memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd->base.alt_dest.dgid, 16);
2386         attr->alt_ah_attr.grh.flow_label    = cmd->base.alt_dest.flow_label;
2387         attr->alt_ah_attr.grh.sgid_index    = cmd->base.alt_dest.sgid_index;
2388         attr->alt_ah_attr.grh.hop_limit     = cmd->base.alt_dest.hop_limit;
2389         attr->alt_ah_attr.grh.traffic_class = cmd->base.alt_dest.traffic_class;
2390         attr->alt_ah_attr.dlid              = cmd->base.alt_dest.dlid;
2391         attr->alt_ah_attr.sl                = cmd->base.alt_dest.sl;
2392         attr->alt_ah_attr.src_path_bits     = cmd->base.alt_dest.src_path_bits;
2393         attr->alt_ah_attr.static_rate       = cmd->base.alt_dest.static_rate;
2394         attr->alt_ah_attr.ah_flags          = cmd->base.alt_dest.is_global ?
2395                                               IB_AH_GRH : 0;
2396         attr->alt_ah_attr.port_num          = cmd->base.alt_dest.port_num;
2397
2398         if (qp->real_qp == qp) {
2399                 if (cmd->base.attr_mask & IB_QP_AV) {
2400                         ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
2401                         if (ret)
2402                                 goto release_qp;
2403                 }
2404                 ret = qp->device->modify_qp(qp, attr,
2405                                             modify_qp_mask(qp->qp_type,
2406                                                            cmd->base.attr_mask),
2407                                             udata);
2408         } else {
2409                 ret = ib_modify_qp(qp, attr,
2410                                    modify_qp_mask(qp->qp_type,
2411                                                   cmd->base.attr_mask));
2412         }
2413
2414 release_qp:
2415         put_qp_read(qp);
2416
2417 out:
2418         kfree(attr);
2419
2420         return ret;
2421 }
2422
2423 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
2424                             struct ib_device *ib_dev,
2425                             const char __user *buf, int in_len,
2426                             int out_len)
2427 {
2428         struct ib_uverbs_ex_modify_qp cmd = {};
2429         struct ib_udata udata;
2430         int ret;
2431
2432         if (copy_from_user(&cmd.base, buf, sizeof(cmd.base)))
2433                 return -EFAULT;
2434
2435         if (cmd.base.attr_mask &
2436             ~((IB_USER_LEGACY_LAST_QP_ATTR_MASK << 1) - 1))
2437                 return -EOPNOTSUPP;
2438
2439         INIT_UDATA(&udata, buf + sizeof(cmd.base), NULL,
2440                    in_len - sizeof(cmd.base), out_len);
2441
2442         ret = modify_qp(file, &cmd, &udata);
2443         if (ret)
2444                 return ret;
2445
2446         return in_len;
2447 }
2448
2449 int ib_uverbs_ex_modify_qp(struct ib_uverbs_file *file,
2450                            struct ib_device *ib_dev,
2451                            struct ib_udata *ucore,
2452                            struct ib_udata *uhw)
2453 {
2454         struct ib_uverbs_ex_modify_qp cmd = {};
2455         int ret;
2456
2457         /*
2458          * Last bit is reserved for extending the attr_mask by
2459          * using another field.
2460          */
2461         BUILD_BUG_ON(IB_USER_LAST_QP_ATTR_MASK == (1 << 31));
2462
2463         if (ucore->inlen < sizeof(cmd.base))
2464                 return -EINVAL;
2465
2466         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2467         if (ret)
2468                 return ret;
2469
2470         if (cmd.base.attr_mask &
2471             ~((IB_USER_LAST_QP_ATTR_MASK << 1) - 1))
2472                 return -EOPNOTSUPP;
2473
2474         if (ucore->inlen > sizeof(cmd)) {
2475                 if (ib_is_udata_cleared(ucore, sizeof(cmd),
2476                                         ucore->inlen - sizeof(cmd)))
2477                         return -EOPNOTSUPP;
2478         }
2479
2480         ret = modify_qp(file, &cmd, uhw);
2481
2482         return ret;
2483 }
2484
2485 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
2486                              struct ib_device *ib_dev,
2487                              const char __user *buf, int in_len,
2488                              int out_len)
2489 {
2490         struct ib_uverbs_destroy_qp      cmd;
2491         struct ib_uverbs_destroy_qp_resp resp;
2492         struct ib_uobject               *uobj;
2493         struct ib_qp                    *qp;
2494         struct ib_uqp_object            *obj;
2495         int                              ret = -EINVAL;
2496
2497         if (copy_from_user(&cmd, buf, sizeof cmd))
2498                 return -EFAULT;
2499
2500         memset(&resp, 0, sizeof resp);
2501
2502         uobj = idr_write_uobj(&ib_uverbs_qp_idr, cmd.qp_handle, file->ucontext);
2503         if (!uobj)
2504                 return -EINVAL;
2505         qp  = uobj->object;
2506         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2507
2508         if (!list_empty(&obj->mcast_list)) {
2509                 put_uobj_write(uobj);
2510                 return -EBUSY;
2511         }
2512
2513         ret = ib_destroy_qp(qp);
2514         if (!ret)
2515                 uobj->live = 0;
2516
2517         put_uobj_write(uobj);
2518
2519         if (ret)
2520                 return ret;
2521
2522         if (obj->uxrcd)
2523                 atomic_dec(&obj->uxrcd->refcnt);
2524
2525         idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
2526
2527         mutex_lock(&file->mutex);
2528         list_del(&uobj->list);
2529         mutex_unlock(&file->mutex);
2530
2531         ib_uverbs_release_uevent(file, &obj->uevent);
2532
2533         resp.events_reported = obj->uevent.events_reported;
2534
2535         put_uobj(uobj);
2536
2537         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2538                          &resp, sizeof resp))
2539                 return -EFAULT;
2540
2541         return in_len;
2542 }
2543
2544 static void *alloc_wr(size_t wr_size, __u32 num_sge)
2545 {
2546         return kmalloc(ALIGN(wr_size, sizeof (struct ib_sge)) +
2547                          num_sge * sizeof (struct ib_sge), GFP_KERNEL);
2548 };
2549
2550 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2551                             struct ib_device *ib_dev,
2552                             const char __user *buf, int in_len,
2553                             int out_len)
2554 {
2555         struct ib_uverbs_post_send      cmd;
2556         struct ib_uverbs_post_send_resp resp;
2557         struct ib_uverbs_send_wr       *user_wr;
2558         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
2559         struct ib_qp                   *qp;
2560         int                             i, sg_ind;
2561         int                             is_ud;
2562         ssize_t                         ret = -EINVAL;
2563         size_t                          next_size;
2564
2565         if (copy_from_user(&cmd, buf, sizeof cmd))
2566                 return -EFAULT;
2567
2568         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2569             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2570                 return -EINVAL;
2571
2572         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2573                 return -EINVAL;
2574
2575         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2576         if (!user_wr)
2577                 return -ENOMEM;
2578
2579         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2580         if (!qp)
2581                 goto out;
2582
2583         is_ud = qp->qp_type == IB_QPT_UD;
2584         sg_ind = 0;
2585         last = NULL;
2586         for (i = 0; i < cmd.wr_count; ++i) {
2587                 if (copy_from_user(user_wr,
2588                                    buf + sizeof cmd + i * cmd.wqe_size,
2589                                    cmd.wqe_size)) {
2590                         ret = -EFAULT;
2591                         goto out_put;
2592                 }
2593
2594                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2595                         ret = -EINVAL;
2596                         goto out_put;
2597                 }
2598
2599                 if (is_ud) {
2600                         struct ib_ud_wr *ud;
2601
2602                         if (user_wr->opcode != IB_WR_SEND &&
2603                             user_wr->opcode != IB_WR_SEND_WITH_IMM) {
2604                                 ret = -EINVAL;
2605                                 goto out_put;
2606                         }
2607
2608                         next_size = sizeof(*ud);
2609                         ud = alloc_wr(next_size, user_wr->num_sge);
2610                         if (!ud) {
2611                                 ret = -ENOMEM;
2612                                 goto out_put;
2613                         }
2614
2615                         ud->ah = idr_read_ah(user_wr->wr.ud.ah, file->ucontext);
2616                         if (!ud->ah) {
2617                                 kfree(ud);
2618                                 ret = -EINVAL;
2619                                 goto out_put;
2620                         }
2621                         ud->remote_qpn = user_wr->wr.ud.remote_qpn;
2622                         ud->remote_qkey = user_wr->wr.ud.remote_qkey;
2623
2624                         next = &ud->wr;
2625                 } else if (user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
2626                            user_wr->opcode == IB_WR_RDMA_WRITE ||
2627                            user_wr->opcode == IB_WR_RDMA_READ) {
2628                         struct ib_rdma_wr *rdma;
2629
2630                         next_size = sizeof(*rdma);
2631                         rdma = alloc_wr(next_size, user_wr->num_sge);
2632                         if (!rdma) {
2633                                 ret = -ENOMEM;
2634                                 goto out_put;
2635                         }
2636
2637                         rdma->remote_addr = user_wr->wr.rdma.remote_addr;
2638                         rdma->rkey = user_wr->wr.rdma.rkey;
2639
2640                         next = &rdma->wr;
2641                 } else if (user_wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2642                            user_wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2643                         struct ib_atomic_wr *atomic;
2644
2645                         next_size = sizeof(*atomic);
2646                         atomic = alloc_wr(next_size, user_wr->num_sge);
2647                         if (!atomic) {
2648                                 ret = -ENOMEM;
2649                                 goto out_put;
2650                         }
2651
2652                         atomic->remote_addr = user_wr->wr.atomic.remote_addr;
2653                         atomic->compare_add = user_wr->wr.atomic.compare_add;
2654                         atomic->swap = user_wr->wr.atomic.swap;
2655                         atomic->rkey = user_wr->wr.atomic.rkey;
2656
2657                         next = &atomic->wr;
2658                 } else if (user_wr->opcode == IB_WR_SEND ||
2659                            user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2660                            user_wr->opcode == IB_WR_SEND_WITH_INV) {
2661                         next_size = sizeof(*next);
2662                         next = alloc_wr(next_size, user_wr->num_sge);
2663                         if (!next) {
2664                                 ret = -ENOMEM;
2665                                 goto out_put;
2666                         }
2667                 } else {
2668                         ret = -EINVAL;
2669                         goto out_put;
2670                 }
2671
2672                 if (user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2673                     user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
2674                         next->ex.imm_data =
2675                                         (__be32 __force) user_wr->ex.imm_data;
2676                 } else if (user_wr->opcode == IB_WR_SEND_WITH_INV) {
2677                         next->ex.invalidate_rkey = user_wr->ex.invalidate_rkey;
2678                 }
2679
2680                 if (!last)
2681                         wr = next;
2682                 else
2683                         last->next = next;
2684                 last = next;
2685
2686                 next->next       = NULL;
2687                 next->wr_id      = user_wr->wr_id;
2688                 next->num_sge    = user_wr->num_sge;
2689                 next->opcode     = user_wr->opcode;
2690                 next->send_flags = user_wr->send_flags;
2691
2692                 if (next->num_sge) {
2693                         next->sg_list = (void *) next +
2694                                 ALIGN(next_size, sizeof(struct ib_sge));
2695                         if (copy_from_user(next->sg_list,
2696                                            buf + sizeof cmd +
2697                                            cmd.wr_count * cmd.wqe_size +
2698                                            sg_ind * sizeof (struct ib_sge),
2699                                            next->num_sge * sizeof (struct ib_sge))) {
2700                                 ret = -EFAULT;
2701                                 goto out_put;
2702                         }
2703                         sg_ind += next->num_sge;
2704                 } else
2705                         next->sg_list = NULL;
2706         }
2707
2708         resp.bad_wr = 0;
2709         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2710         if (ret)
2711                 for (next = wr; next; next = next->next) {
2712                         ++resp.bad_wr;
2713                         if (next == bad_wr)
2714                                 break;
2715                 }
2716
2717         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2718                          &resp, sizeof resp))
2719                 ret = -EFAULT;
2720
2721 out_put:
2722         put_qp_read(qp);
2723
2724         while (wr) {
2725                 if (is_ud && ud_wr(wr)->ah)
2726                         put_ah_read(ud_wr(wr)->ah);
2727                 next = wr->next;
2728                 kfree(wr);
2729                 wr = next;
2730         }
2731
2732 out:
2733         kfree(user_wr);
2734
2735         return ret ? ret : in_len;
2736 }
2737
2738 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2739                                                     int in_len,
2740                                                     u32 wr_count,
2741                                                     u32 sge_count,
2742                                                     u32 wqe_size)
2743 {
2744         struct ib_uverbs_recv_wr *user_wr;
2745         struct ib_recv_wr        *wr = NULL, *last, *next;
2746         int                       sg_ind;
2747         int                       i;
2748         int                       ret;
2749
2750         if (in_len < wqe_size * wr_count +
2751             sge_count * sizeof (struct ib_uverbs_sge))
2752                 return ERR_PTR(-EINVAL);
2753
2754         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2755                 return ERR_PTR(-EINVAL);
2756
2757         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2758         if (!user_wr)
2759                 return ERR_PTR(-ENOMEM);
2760
2761         sg_ind = 0;
2762         last = NULL;
2763         for (i = 0; i < wr_count; ++i) {
2764                 if (copy_from_user(user_wr, buf + i * wqe_size,
2765                                    wqe_size)) {
2766                         ret = -EFAULT;
2767                         goto err;
2768                 }
2769
2770                 if (user_wr->num_sge + sg_ind > sge_count) {
2771                         ret = -EINVAL;
2772                         goto err;
2773                 }
2774
2775                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2776                                user_wr->num_sge * sizeof (struct ib_sge),
2777                                GFP_KERNEL);
2778                 if (!next) {
2779                         ret = -ENOMEM;
2780                         goto err;
2781                 }
2782
2783                 if (!last)
2784                         wr = next;
2785                 else
2786                         last->next = next;
2787                 last = next;
2788
2789                 next->next       = NULL;
2790                 next->wr_id      = user_wr->wr_id;
2791                 next->num_sge    = user_wr->num_sge;
2792
2793                 if (next->num_sge) {
2794                         next->sg_list = (void *) next +
2795                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2796                         if (copy_from_user(next->sg_list,
2797                                            buf + wr_count * wqe_size +
2798                                            sg_ind * sizeof (struct ib_sge),
2799                                            next->num_sge * sizeof (struct ib_sge))) {
2800                                 ret = -EFAULT;
2801                                 goto err;
2802                         }
2803                         sg_ind += next->num_sge;
2804                 } else
2805                         next->sg_list = NULL;
2806         }
2807
2808         kfree(user_wr);
2809         return wr;
2810
2811 err:
2812         kfree(user_wr);
2813
2814         while (wr) {
2815                 next = wr->next;
2816                 kfree(wr);
2817                 wr = next;
2818         }
2819
2820         return ERR_PTR(ret);
2821 }
2822
2823 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2824                             struct ib_device *ib_dev,
2825                             const char __user *buf, int in_len,
2826                             int out_len)
2827 {
2828         struct ib_uverbs_post_recv      cmd;
2829         struct ib_uverbs_post_recv_resp resp;
2830         struct ib_recv_wr              *wr, *next, *bad_wr;
2831         struct ib_qp                   *qp;
2832         ssize_t                         ret = -EINVAL;
2833
2834         if (copy_from_user(&cmd, buf, sizeof cmd))
2835                 return -EFAULT;
2836
2837         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2838                                        in_len - sizeof cmd, cmd.wr_count,
2839                                        cmd.sge_count, cmd.wqe_size);
2840         if (IS_ERR(wr))
2841                 return PTR_ERR(wr);
2842
2843         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2844         if (!qp)
2845                 goto out;
2846
2847         resp.bad_wr = 0;
2848         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2849
2850         put_qp_read(qp);
2851
2852         if (ret)
2853                 for (next = wr; next; next = next->next) {
2854                         ++resp.bad_wr;
2855                         if (next == bad_wr)
2856                                 break;
2857                 }
2858
2859         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2860                          &resp, sizeof resp))
2861                 ret = -EFAULT;
2862
2863 out:
2864         while (wr) {
2865                 next = wr->next;
2866                 kfree(wr);
2867                 wr = next;
2868         }
2869
2870         return ret ? ret : in_len;
2871 }
2872
2873 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2874                                 struct ib_device *ib_dev,
2875                                 const char __user *buf, int in_len,
2876                                 int out_len)
2877 {
2878         struct ib_uverbs_post_srq_recv      cmd;
2879         struct ib_uverbs_post_srq_recv_resp resp;
2880         struct ib_recv_wr                  *wr, *next, *bad_wr;
2881         struct ib_srq                      *srq;
2882         ssize_t                             ret = -EINVAL;
2883
2884         if (copy_from_user(&cmd, buf, sizeof cmd))
2885                 return -EFAULT;
2886
2887         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2888                                        in_len - sizeof cmd, cmd.wr_count,
2889                                        cmd.sge_count, cmd.wqe_size);
2890         if (IS_ERR(wr))
2891                 return PTR_ERR(wr);
2892
2893         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2894         if (!srq)
2895                 goto out;
2896
2897         resp.bad_wr = 0;
2898         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2899
2900         put_srq_read(srq);
2901
2902         if (ret)
2903                 for (next = wr; next; next = next->next) {
2904                         ++resp.bad_wr;
2905                         if (next == bad_wr)
2906                                 break;
2907                 }
2908
2909         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2910                          &resp, sizeof resp))
2911                 ret = -EFAULT;
2912
2913 out:
2914         while (wr) {
2915                 next = wr->next;
2916                 kfree(wr);
2917                 wr = next;
2918         }
2919
2920         return ret ? ret : in_len;
2921 }
2922
2923 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2924                             struct ib_device *ib_dev,
2925                             const char __user *buf, int in_len,
2926                             int out_len)
2927 {
2928         struct ib_uverbs_create_ah       cmd;
2929         struct ib_uverbs_create_ah_resp  resp;
2930         struct ib_uobject               *uobj;
2931         struct ib_pd                    *pd;
2932         struct ib_ah                    *ah;
2933         struct ib_ah_attr               attr;
2934         int ret;
2935         struct ib_udata                   udata;
2936
2937         if (out_len < sizeof resp)
2938                 return -ENOSPC;
2939
2940         if (copy_from_user(&cmd, buf, sizeof cmd))
2941                 return -EFAULT;
2942
2943         INIT_UDATA(&udata, buf + sizeof(cmd),
2944                    (unsigned long)cmd.response + sizeof(resp),
2945                    in_len - sizeof(cmd), out_len - sizeof(resp));
2946
2947         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
2948         if (!uobj)
2949                 return -ENOMEM;
2950
2951         init_uobj(uobj, cmd.user_handle, file->ucontext, &ah_lock_class);
2952         down_write(&uobj->mutex);
2953
2954         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
2955         if (!pd) {
2956                 ret = -EINVAL;
2957                 goto err;
2958         }
2959
2960         attr.dlid              = cmd.attr.dlid;
2961         attr.sl                = cmd.attr.sl;
2962         attr.src_path_bits     = cmd.attr.src_path_bits;
2963         attr.static_rate       = cmd.attr.static_rate;
2964         attr.ah_flags          = cmd.attr.is_global ? IB_AH_GRH : 0;
2965         attr.port_num          = cmd.attr.port_num;
2966         attr.grh.flow_label    = cmd.attr.grh.flow_label;
2967         attr.grh.sgid_index    = cmd.attr.grh.sgid_index;
2968         attr.grh.hop_limit     = cmd.attr.grh.hop_limit;
2969         attr.grh.traffic_class = cmd.attr.grh.traffic_class;
2970         memset(&attr.dmac, 0, sizeof(attr.dmac));
2971         memcpy(attr.grh.dgid.raw, cmd.attr.grh.dgid, 16);
2972
2973         ah = pd->device->create_ah(pd, &attr, &udata);
2974
2975         if (IS_ERR(ah)) {
2976                 ret = PTR_ERR(ah);
2977                 goto err_put;
2978         }
2979
2980         ah->device  = pd->device;
2981         ah->pd      = pd;
2982         atomic_inc(&pd->usecnt);
2983         ah->uobject  = uobj;
2984         uobj->object = ah;
2985
2986         ret = idr_add_uobj(&ib_uverbs_ah_idr, uobj);
2987         if (ret)
2988                 goto err_destroy;
2989
2990         resp.ah_handle = uobj->id;
2991
2992         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2993                          &resp, sizeof resp)) {
2994                 ret = -EFAULT;
2995                 goto err_copy;
2996         }
2997
2998         put_pd_read(pd);
2999
3000         mutex_lock(&file->mutex);
3001         list_add_tail(&uobj->list, &file->ucontext->ah_list);
3002         mutex_unlock(&file->mutex);
3003
3004         uobj->live = 1;
3005
3006         up_write(&uobj->mutex);
3007
3008         return in_len;
3009
3010 err_copy:
3011         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
3012
3013 err_destroy:
3014         ib_destroy_ah(ah);
3015
3016 err_put:
3017         put_pd_read(pd);
3018
3019 err:
3020         put_uobj_write(uobj);
3021         return ret;
3022 }
3023
3024 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
3025                              struct ib_device *ib_dev,
3026                              const char __user *buf, int in_len, int out_len)
3027 {
3028         struct ib_uverbs_destroy_ah cmd;
3029         struct ib_ah               *ah;
3030         struct ib_uobject          *uobj;
3031         int                         ret;
3032
3033         if (copy_from_user(&cmd, buf, sizeof cmd))
3034                 return -EFAULT;
3035
3036         uobj = idr_write_uobj(&ib_uverbs_ah_idr, cmd.ah_handle, file->ucontext);
3037         if (!uobj)
3038                 return -EINVAL;
3039         ah = uobj->object;
3040
3041         ret = ib_destroy_ah(ah);
3042         if (!ret)
3043                 uobj->live = 0;
3044
3045         put_uobj_write(uobj);
3046
3047         if (ret)
3048                 return ret;
3049
3050         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
3051
3052         mutex_lock(&file->mutex);
3053         list_del(&uobj->list);
3054         mutex_unlock(&file->mutex);
3055
3056         put_uobj(uobj);
3057
3058         return in_len;
3059 }
3060
3061 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
3062                                struct ib_device *ib_dev,
3063                                const char __user *buf, int in_len,
3064                                int out_len)
3065 {
3066         struct ib_uverbs_attach_mcast cmd;
3067         struct ib_qp                 *qp;
3068         struct ib_uqp_object         *obj;
3069         struct ib_uverbs_mcast_entry *mcast;
3070         int                           ret;
3071
3072         if (copy_from_user(&cmd, buf, sizeof cmd))
3073                 return -EFAULT;
3074
3075         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
3076         if (!qp)
3077                 return -EINVAL;
3078
3079         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
3080
3081         list_for_each_entry(mcast, &obj->mcast_list, list)
3082                 if (cmd.mlid == mcast->lid &&
3083                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
3084                         ret = 0;
3085                         goto out_put;
3086                 }
3087
3088         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
3089         if (!mcast) {
3090                 ret = -ENOMEM;
3091                 goto out_put;
3092         }
3093
3094         mcast->lid = cmd.mlid;
3095         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
3096
3097         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
3098         if (!ret)
3099                 list_add_tail(&mcast->list, &obj->mcast_list);
3100         else
3101                 kfree(mcast);
3102
3103 out_put:
3104         put_qp_write(qp);
3105
3106         return ret ? ret : in_len;
3107 }
3108
3109 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
3110                                struct ib_device *ib_dev,
3111                                const char __user *buf, int in_len,
3112                                int out_len)
3113 {
3114         struct ib_uverbs_detach_mcast cmd;
3115         struct ib_uqp_object         *obj;
3116         struct ib_qp                 *qp;
3117         struct ib_uverbs_mcast_entry *mcast;
3118         int                           ret = -EINVAL;
3119
3120         if (copy_from_user(&cmd, buf, sizeof cmd))
3121                 return -EFAULT;
3122
3123         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
3124         if (!qp)
3125                 return -EINVAL;
3126
3127         ret = ib_detach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
3128         if (ret)
3129                 goto out_put;
3130
3131         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
3132
3133         list_for_each_entry(mcast, &obj->mcast_list, list)
3134                 if (cmd.mlid == mcast->lid &&
3135                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
3136                         list_del(&mcast->list);
3137                         kfree(mcast);
3138                         break;
3139                 }
3140
3141 out_put:
3142         put_qp_write(qp);
3143
3144         return ret ? ret : in_len;
3145 }
3146
3147 static int kern_spec_to_ib_spec_action(struct ib_uverbs_flow_spec *kern_spec,
3148                                        union ib_flow_spec *ib_spec)
3149 {
3150         ib_spec->type = kern_spec->type;
3151         switch (ib_spec->type) {
3152         case IB_FLOW_SPEC_ACTION_TAG:
3153                 if (kern_spec->flow_tag.size !=
3154                     sizeof(struct ib_uverbs_flow_spec_action_tag))
3155                         return -EINVAL;
3156
3157                 ib_spec->flow_tag.size = sizeof(struct ib_flow_spec_action_tag);
3158                 ib_spec->flow_tag.tag_id = kern_spec->flow_tag.tag_id;
3159                 break;
3160         default:
3161                 return -EINVAL;
3162         }
3163         return 0;
3164 }
3165
3166 static size_t kern_spec_filter_sz(struct ib_uverbs_flow_spec_hdr *spec)
3167 {
3168         /* Returns user space filter size, includes padding */
3169         return (spec->size - sizeof(struct ib_uverbs_flow_spec_hdr)) / 2;
3170 }
3171
3172 static ssize_t spec_filter_size(void *kern_spec_filter, u16 kern_filter_size,
3173                                 u16 ib_real_filter_sz)
3174 {
3175         /*
3176          * User space filter structures must be 64 bit aligned, otherwise this
3177          * may pass, but we won't handle additional new attributes.
3178          */
3179
3180         if (kern_filter_size > ib_real_filter_sz) {
3181                 if (memchr_inv(kern_spec_filter +
3182                                ib_real_filter_sz, 0,
3183                                kern_filter_size - ib_real_filter_sz))
3184                         return -EINVAL;
3185                 return ib_real_filter_sz;
3186         }
3187         return kern_filter_size;
3188 }
3189
3190 static int kern_spec_to_ib_spec_filter(struct ib_uverbs_flow_spec *kern_spec,
3191                                        union ib_flow_spec *ib_spec)
3192 {
3193         ssize_t actual_filter_sz;
3194         ssize_t kern_filter_sz;
3195         ssize_t ib_filter_sz;
3196         void *kern_spec_mask;
3197         void *kern_spec_val;
3198
3199         if (kern_spec->reserved)
3200                 return -EINVAL;
3201
3202         ib_spec->type = kern_spec->type;
3203
3204         kern_filter_sz = kern_spec_filter_sz(&kern_spec->hdr);
3205         /* User flow spec size must be aligned to 4 bytes */
3206         if (kern_filter_sz != ALIGN(kern_filter_sz, 4))
3207                 return -EINVAL;
3208
3209         kern_spec_val = (void *)kern_spec +
3210                 sizeof(struct ib_uverbs_flow_spec_hdr);
3211         kern_spec_mask = kern_spec_val + kern_filter_sz;
3212         if (ib_spec->type == (IB_FLOW_SPEC_INNER | IB_FLOW_SPEC_VXLAN_TUNNEL))
3213                 return -EINVAL;
3214
3215         switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
3216         case IB_FLOW_SPEC_ETH:
3217                 ib_filter_sz = offsetof(struct ib_flow_eth_filter, real_sz);
3218                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3219                                                     kern_filter_sz,
3220                                                     ib_filter_sz);
3221                 if (actual_filter_sz <= 0)
3222                         return -EINVAL;
3223                 ib_spec->size = sizeof(struct ib_flow_spec_eth);
3224                 memcpy(&ib_spec->eth.val, kern_spec_val, actual_filter_sz);
3225                 memcpy(&ib_spec->eth.mask, kern_spec_mask, actual_filter_sz);
3226                 break;
3227         case IB_FLOW_SPEC_IPV4:
3228                 ib_filter_sz = offsetof(struct ib_flow_ipv4_filter, real_sz);
3229                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3230                                                     kern_filter_sz,
3231                                                     ib_filter_sz);
3232                 if (actual_filter_sz <= 0)
3233                         return -EINVAL;
3234                 ib_spec->size = sizeof(struct ib_flow_spec_ipv4);
3235                 memcpy(&ib_spec->ipv4.val, kern_spec_val, actual_filter_sz);
3236                 memcpy(&ib_spec->ipv4.mask, kern_spec_mask, actual_filter_sz);
3237                 break;
3238         case IB_FLOW_SPEC_IPV6:
3239                 ib_filter_sz = offsetof(struct ib_flow_ipv6_filter, real_sz);
3240                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3241                                                     kern_filter_sz,
3242                                                     ib_filter_sz);
3243                 if (actual_filter_sz <= 0)
3244                         return -EINVAL;
3245                 ib_spec->size = sizeof(struct ib_flow_spec_ipv6);
3246                 memcpy(&ib_spec->ipv6.val, kern_spec_val, actual_filter_sz);
3247                 memcpy(&ib_spec->ipv6.mask, kern_spec_mask, actual_filter_sz);
3248
3249                 if ((ntohl(ib_spec->ipv6.mask.flow_label)) >= BIT(20) ||
3250                     (ntohl(ib_spec->ipv6.val.flow_label)) >= BIT(20))
3251                         return -EINVAL;
3252                 break;
3253         case IB_FLOW_SPEC_TCP:
3254         case IB_FLOW_SPEC_UDP:
3255                 ib_filter_sz = offsetof(struct ib_flow_tcp_udp_filter, real_sz);
3256                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3257                                                     kern_filter_sz,
3258                                                     ib_filter_sz);
3259                 if (actual_filter_sz <= 0)
3260                         return -EINVAL;
3261                 ib_spec->size = sizeof(struct ib_flow_spec_tcp_udp);
3262                 memcpy(&ib_spec->tcp_udp.val, kern_spec_val, actual_filter_sz);
3263                 memcpy(&ib_spec->tcp_udp.mask, kern_spec_mask, actual_filter_sz);
3264                 break;
3265         case IB_FLOW_SPEC_VXLAN_TUNNEL:
3266                 ib_filter_sz = offsetof(struct ib_flow_tunnel_filter, real_sz);
3267                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3268                                                     kern_filter_sz,
3269                                                     ib_filter_sz);
3270                 if (actual_filter_sz <= 0)
3271                         return -EINVAL;
3272                 ib_spec->tunnel.size = sizeof(struct ib_flow_spec_tunnel);
3273                 memcpy(&ib_spec->tunnel.val, kern_spec_val, actual_filter_sz);
3274                 memcpy(&ib_spec->tunnel.mask, kern_spec_mask, actual_filter_sz);
3275
3276                 if ((ntohl(ib_spec->tunnel.mask.tunnel_id)) >= BIT(24) ||
3277                     (ntohl(ib_spec->tunnel.val.tunnel_id)) >= BIT(24))
3278                         return -EINVAL;
3279                 break;
3280         default:
3281                 return -EINVAL;
3282         }
3283         return 0;
3284 }
3285
3286 static int kern_spec_to_ib_spec(struct ib_uverbs_flow_spec *kern_spec,
3287                                 union ib_flow_spec *ib_spec)
3288 {
3289         if (kern_spec->reserved)
3290                 return -EINVAL;
3291
3292         if (kern_spec->type >= IB_FLOW_SPEC_ACTION_TAG)
3293                 return kern_spec_to_ib_spec_action(kern_spec, ib_spec);
3294         else
3295                 return kern_spec_to_ib_spec_filter(kern_spec, ib_spec);
3296 }
3297
3298 int ib_uverbs_ex_create_wq(struct ib_uverbs_file *file,
3299                            struct ib_device *ib_dev,
3300                            struct ib_udata *ucore,
3301                            struct ib_udata *uhw)
3302 {
3303         struct ib_uverbs_ex_create_wq     cmd = {};
3304         struct ib_uverbs_ex_create_wq_resp resp = {};
3305         struct ib_uwq_object           *obj;
3306         int err = 0;
3307         struct ib_cq *cq;
3308         struct ib_pd *pd;
3309         struct ib_wq *wq;
3310         struct ib_wq_init_attr wq_init_attr = {};
3311         size_t required_cmd_sz;
3312         size_t required_resp_len;
3313
3314         required_cmd_sz = offsetof(typeof(cmd), max_sge) + sizeof(cmd.max_sge);
3315         required_resp_len = offsetof(typeof(resp), wqn) + sizeof(resp.wqn);
3316
3317         if (ucore->inlen < required_cmd_sz)
3318                 return -EINVAL;
3319
3320         if (ucore->outlen < required_resp_len)
3321                 return -ENOSPC;
3322
3323         if (ucore->inlen > sizeof(cmd) &&
3324             !ib_is_udata_cleared(ucore, sizeof(cmd),
3325                                  ucore->inlen - sizeof(cmd)))
3326                 return -EOPNOTSUPP;
3327
3328         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3329         if (err)
3330                 return err;
3331
3332         if (cmd.comp_mask)
3333                 return -EOPNOTSUPP;
3334
3335         obj = kmalloc(sizeof(*obj), GFP_KERNEL);
3336         if (!obj)
3337                 return -ENOMEM;
3338
3339         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext,
3340                   &wq_lock_class);
3341         down_write(&obj->uevent.uobject.mutex);
3342         pd  = idr_read_pd(cmd.pd_handle, file->ucontext);
3343         if (!pd) {
3344                 err = -EINVAL;
3345                 goto err_uobj;
3346         }
3347
3348         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
3349         if (!cq) {
3350                 err = -EINVAL;
3351                 goto err_put_pd;
3352         }
3353
3354         wq_init_attr.cq = cq;
3355         wq_init_attr.max_sge = cmd.max_sge;
3356         wq_init_attr.max_wr = cmd.max_wr;
3357         wq_init_attr.wq_context = file;
3358         wq_init_attr.wq_type = cmd.wq_type;
3359         wq_init_attr.event_handler = ib_uverbs_wq_event_handler;
3360         if (ucore->inlen >= (offsetof(typeof(cmd), create_flags) +
3361                              sizeof(cmd.create_flags)))
3362                 wq_init_attr.create_flags = cmd.create_flags;
3363         obj->uevent.events_reported = 0;
3364         INIT_LIST_HEAD(&obj->uevent.event_list);
3365         wq = pd->device->create_wq(pd, &wq_init_attr, uhw);
3366         if (IS_ERR(wq)) {
3367                 err = PTR_ERR(wq);
3368                 goto err_put_cq;
3369         }
3370
3371         wq->uobject = &obj->uevent.uobject;
3372         obj->uevent.uobject.object = wq;
3373         wq->wq_type = wq_init_attr.wq_type;
3374         wq->cq = cq;
3375         wq->pd = pd;
3376         wq->device = pd->device;
3377         wq->wq_context = wq_init_attr.wq_context;
3378         atomic_set(&wq->usecnt, 0);
3379         atomic_inc(&pd->usecnt);
3380         atomic_inc(&cq->usecnt);
3381         wq->uobject = &obj->uevent.uobject;
3382         obj->uevent.uobject.object = wq;
3383         err = idr_add_uobj(&ib_uverbs_wq_idr, &obj->uevent.uobject);
3384         if (err)
3385                 goto destroy_wq;
3386
3387         memset(&resp, 0, sizeof(resp));
3388         resp.wq_handle = obj->uevent.uobject.id;
3389         resp.max_sge = wq_init_attr.max_sge;
3390         resp.max_wr = wq_init_attr.max_wr;
3391         resp.wqn = wq->wq_num;
3392         resp.response_length = required_resp_len;
3393         err = ib_copy_to_udata(ucore,
3394                                &resp, resp.response_length);
3395         if (err)
3396                 goto err_copy;
3397
3398         put_pd_read(pd);
3399         put_cq_read(cq);
3400
3401         mutex_lock(&file->mutex);
3402         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->wq_list);
3403         mutex_unlock(&file->mutex);
3404
3405         obj->uevent.uobject.live = 1;
3406         up_write(&obj->uevent.uobject.mutex);
3407         return 0;
3408
3409 err_copy:
3410         idr_remove_uobj(&ib_uverbs_wq_idr, &obj->uevent.uobject);
3411 destroy_wq:
3412         ib_destroy_wq(wq);
3413 err_put_cq:
3414         put_cq_read(cq);
3415 err_put_pd:
3416         put_pd_read(pd);
3417 err_uobj:
3418         put_uobj_write(&obj->uevent.uobject);
3419
3420         return err;
3421 }
3422
3423 int ib_uverbs_ex_destroy_wq(struct ib_uverbs_file *file,
3424                             struct ib_device *ib_dev,
3425                             struct ib_udata *ucore,
3426                             struct ib_udata *uhw)
3427 {
3428         struct ib_uverbs_ex_destroy_wq  cmd = {};
3429         struct ib_uverbs_ex_destroy_wq_resp     resp = {};
3430         struct ib_wq                    *wq;
3431         struct ib_uobject               *uobj;
3432         struct ib_uwq_object            *obj;
3433         size_t required_cmd_sz;
3434         size_t required_resp_len;
3435         int                             ret;
3436
3437         required_cmd_sz = offsetof(typeof(cmd), wq_handle) + sizeof(cmd.wq_handle);
3438         required_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
3439
3440         if (ucore->inlen < required_cmd_sz)
3441                 return -EINVAL;
3442
3443         if (ucore->outlen < required_resp_len)
3444                 return -ENOSPC;
3445
3446         if (ucore->inlen > sizeof(cmd) &&
3447             !ib_is_udata_cleared(ucore, sizeof(cmd),
3448                                  ucore->inlen - sizeof(cmd)))
3449                 return -EOPNOTSUPP;
3450
3451         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3452         if (ret)
3453                 return ret;
3454
3455         if (cmd.comp_mask)
3456                 return -EOPNOTSUPP;
3457
3458         resp.response_length = required_resp_len;
3459         uobj = idr_write_uobj(&ib_uverbs_wq_idr, cmd.wq_handle,
3460                               file->ucontext);
3461         if (!uobj)
3462                 return -EINVAL;
3463
3464         wq = uobj->object;
3465         obj = container_of(uobj, struct ib_uwq_object, uevent.uobject);
3466         ret = ib_destroy_wq(wq);
3467         if (!ret)
3468                 uobj->live = 0;
3469
3470         put_uobj_write(uobj);
3471         if (ret)
3472                 return ret;
3473
3474         idr_remove_uobj(&ib_uverbs_wq_idr, uobj);
3475
3476         mutex_lock(&file->mutex);
3477         list_del(&uobj->list);
3478         mutex_unlock(&file->mutex);
3479
3480         ib_uverbs_release_uevent(file, &obj->uevent);
3481         resp.events_reported = obj->uevent.events_reported;
3482         put_uobj(uobj);
3483
3484         ret = ib_copy_to_udata(ucore, &resp, resp.response_length);
3485         if (ret)
3486                 return ret;
3487
3488         return 0;
3489 }
3490
3491 int ib_uverbs_ex_modify_wq(struct ib_uverbs_file *file,
3492                            struct ib_device *ib_dev,
3493                            struct ib_udata *ucore,
3494                            struct ib_udata *uhw)
3495 {
3496         struct ib_uverbs_ex_modify_wq cmd = {};
3497         struct ib_wq *wq;
3498         struct ib_wq_attr wq_attr = {};
3499         size_t required_cmd_sz;
3500         int ret;
3501
3502         required_cmd_sz = offsetof(typeof(cmd), curr_wq_state) + sizeof(cmd.curr_wq_state);
3503         if (ucore->inlen < required_cmd_sz)
3504                 return -EINVAL;
3505
3506         if (ucore->inlen > sizeof(cmd) &&
3507             !ib_is_udata_cleared(ucore, sizeof(cmd),
3508                                  ucore->inlen - sizeof(cmd)))
3509                 return -EOPNOTSUPP;
3510
3511         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3512         if (ret)
3513                 return ret;
3514
3515         if (!cmd.attr_mask)
3516                 return -EINVAL;
3517
3518         if (cmd.attr_mask > (IB_WQ_STATE | IB_WQ_CUR_STATE | IB_WQ_FLAGS))
3519                 return -EINVAL;
3520
3521         wq = idr_read_wq(cmd.wq_handle, file->ucontext);
3522         if (!wq)
3523                 return -EINVAL;
3524
3525         wq_attr.curr_wq_state = cmd.curr_wq_state;
3526         wq_attr.wq_state = cmd.wq_state;
3527         if (cmd.attr_mask & IB_WQ_FLAGS) {
3528                 wq_attr.flags = cmd.flags;
3529                 wq_attr.flags_mask = cmd.flags_mask;
3530         }
3531         ret = wq->device->modify_wq(wq, &wq_attr, cmd.attr_mask, uhw);
3532         put_wq_read(wq);
3533         return ret;
3534 }
3535
3536 int ib_uverbs_ex_create_rwq_ind_table(struct ib_uverbs_file *file,
3537                                       struct ib_device *ib_dev,
3538                                       struct ib_udata *ucore,
3539                                       struct ib_udata *uhw)
3540 {
3541         struct ib_uverbs_ex_create_rwq_ind_table          cmd = {};
3542         struct ib_uverbs_ex_create_rwq_ind_table_resp  resp = {};
3543         struct ib_uobject                 *uobj;
3544         int err = 0;
3545         struct ib_rwq_ind_table_init_attr init_attr = {};
3546         struct ib_rwq_ind_table *rwq_ind_tbl;
3547         struct ib_wq    **wqs = NULL;
3548         u32 *wqs_handles = NULL;
3549         struct ib_wq    *wq = NULL;
3550         int i, j, num_read_wqs;
3551         u32 num_wq_handles;
3552         u32 expected_in_size;
3553         size_t required_cmd_sz_header;
3554         size_t required_resp_len;
3555
3556         required_cmd_sz_header = offsetof(typeof(cmd), log_ind_tbl_size) + sizeof(cmd.log_ind_tbl_size);
3557         required_resp_len = offsetof(typeof(resp), ind_tbl_num) + sizeof(resp.ind_tbl_num);
3558
3559         if (ucore->inlen < required_cmd_sz_header)
3560                 return -EINVAL;
3561
3562         if (ucore->outlen < required_resp_len)
3563                 return -ENOSPC;
3564
3565         err = ib_copy_from_udata(&cmd, ucore, required_cmd_sz_header);
3566         if (err)
3567                 return err;
3568
3569         ucore->inbuf += required_cmd_sz_header;
3570         ucore->inlen -= required_cmd_sz_header;
3571
3572         if (cmd.comp_mask)
3573                 return -EOPNOTSUPP;
3574
3575         if (cmd.log_ind_tbl_size > IB_USER_VERBS_MAX_LOG_IND_TBL_SIZE)
3576                 return -EINVAL;
3577
3578         num_wq_handles = 1 << cmd.log_ind_tbl_size;
3579         expected_in_size = num_wq_handles * sizeof(__u32);
3580         if (num_wq_handles == 1)
3581                 /* input size for wq handles is u64 aligned */
3582                 expected_in_size += sizeof(__u32);
3583
3584         if (ucore->inlen < expected_in_size)
3585                 return -EINVAL;
3586
3587         if (ucore->inlen > expected_in_size &&
3588             !ib_is_udata_cleared(ucore, expected_in_size,
3589                                  ucore->inlen - expected_in_size))
3590                 return -EOPNOTSUPP;
3591
3592         wqs_handles = kcalloc(num_wq_handles, sizeof(*wqs_handles),
3593                               GFP_KERNEL);
3594         if (!wqs_handles)
3595                 return -ENOMEM;
3596
3597         err = ib_copy_from_udata(wqs_handles, ucore,
3598                                  num_wq_handles * sizeof(__u32));
3599         if (err)
3600                 goto err_free;
3601
3602         wqs = kcalloc(num_wq_handles, sizeof(*wqs), GFP_KERNEL);
3603         if (!wqs) {
3604                 err = -ENOMEM;
3605                 goto  err_free;
3606         }
3607
3608         for (num_read_wqs = 0; num_read_wqs < num_wq_handles;
3609                         num_read_wqs++) {
3610                 wq = idr_read_wq(wqs_handles[num_read_wqs], file->ucontext);
3611                 if (!wq) {
3612                         err = -EINVAL;
3613                         goto put_wqs;
3614                 }
3615
3616                 wqs[num_read_wqs] = wq;
3617         }
3618
3619         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
3620         if (!uobj) {
3621                 err = -ENOMEM;
3622                 goto put_wqs;
3623         }
3624
3625         init_uobj(uobj, 0, file->ucontext, &rwq_ind_table_lock_class);
3626         down_write(&uobj->mutex);
3627         init_attr.log_ind_tbl_size = cmd.log_ind_tbl_size;
3628         init_attr.ind_tbl = wqs;
3629         rwq_ind_tbl = ib_dev->create_rwq_ind_table(ib_dev, &init_attr, uhw);
3630
3631         if (IS_ERR(rwq_ind_tbl)) {
3632                 err = PTR_ERR(rwq_ind_tbl);
3633                 goto err_uobj;
3634         }
3635
3636         rwq_ind_tbl->ind_tbl = wqs;
3637         rwq_ind_tbl->log_ind_tbl_size = init_attr.log_ind_tbl_size;
3638         rwq_ind_tbl->uobject = uobj;
3639         uobj->object = rwq_ind_tbl;
3640         rwq_ind_tbl->device = ib_dev;
3641         atomic_set(&rwq_ind_tbl->usecnt, 0);
3642
3643         for (i = 0; i < num_wq_handles; i++)
3644                 atomic_inc(&wqs[i]->usecnt);
3645
3646         err = idr_add_uobj(&ib_uverbs_rwq_ind_tbl_idr, uobj);
3647         if (err)
3648                 goto destroy_ind_tbl;
3649
3650         resp.ind_tbl_handle = uobj->id;
3651         resp.ind_tbl_num = rwq_ind_tbl->ind_tbl_num;
3652         resp.response_length = required_resp_len;
3653
3654         err = ib_copy_to_udata(ucore,
3655                                &resp, resp.response_length);
3656         if (err)
3657                 goto err_copy;
3658
3659         kfree(wqs_handles);
3660
3661         for (j = 0; j < num_read_wqs; j++)
3662                 put_wq_read(wqs[j]);
3663
3664         mutex_lock(&file->mutex);
3665         list_add_tail(&uobj->list, &file->ucontext->rwq_ind_tbl_list);
3666         mutex_unlock(&file->mutex);
3667
3668         uobj->live = 1;
3669
3670         up_write(&uobj->mutex);
3671         return 0;
3672
3673 err_copy:
3674         idr_remove_uobj(&ib_uverbs_rwq_ind_tbl_idr, uobj);
3675 destroy_ind_tbl:
3676         ib_destroy_rwq_ind_table(rwq_ind_tbl);
3677 err_uobj:
3678         put_uobj_write(uobj);
3679 put_wqs:
3680         for (j = 0; j < num_read_wqs; j++)
3681                 put_wq_read(wqs[j]);
3682 err_free:
3683         kfree(wqs_handles);
3684         kfree(wqs);
3685         return err;
3686 }
3687
3688 int ib_uverbs_ex_destroy_rwq_ind_table(struct ib_uverbs_file *file,
3689                                        struct ib_device *ib_dev,
3690                                        struct ib_udata *ucore,
3691                                        struct ib_udata *uhw)
3692 {
3693         struct ib_uverbs_ex_destroy_rwq_ind_table       cmd = {};
3694         struct ib_rwq_ind_table *rwq_ind_tbl;
3695         struct ib_uobject               *uobj;
3696         int                     ret;
3697         struct ib_wq    **ind_tbl;
3698         size_t required_cmd_sz;
3699
3700         required_cmd_sz = offsetof(typeof(cmd), ind_tbl_handle) + sizeof(cmd.ind_tbl_handle);
3701
3702         if (ucore->inlen < required_cmd_sz)
3703                 return -EINVAL;
3704
3705         if (ucore->inlen > sizeof(cmd) &&
3706             !ib_is_udata_cleared(ucore, sizeof(cmd),
3707                                  ucore->inlen - sizeof(cmd)))
3708                 return -EOPNOTSUPP;
3709
3710         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3711         if (ret)
3712                 return ret;
3713
3714         if (cmd.comp_mask)
3715                 return -EOPNOTSUPP;
3716
3717         uobj = idr_write_uobj(&ib_uverbs_rwq_ind_tbl_idr, cmd.ind_tbl_handle,
3718                               file->ucontext);
3719         if (!uobj)
3720                 return -EINVAL;
3721         rwq_ind_tbl = uobj->object;
3722         ind_tbl = rwq_ind_tbl->ind_tbl;
3723
3724         ret = ib_destroy_rwq_ind_table(rwq_ind_tbl);
3725         if (!ret)
3726                 uobj->live = 0;
3727
3728         put_uobj_write(uobj);
3729
3730         if (ret)
3731                 return ret;
3732
3733         idr_remove_uobj(&ib_uverbs_rwq_ind_tbl_idr, uobj);
3734
3735         mutex_lock(&file->mutex);
3736         list_del(&uobj->list);
3737         mutex_unlock(&file->mutex);
3738
3739         put_uobj(uobj);
3740         kfree(ind_tbl);
3741         return ret;
3742 }
3743
3744 int ib_uverbs_ex_create_flow(struct ib_uverbs_file *file,
3745                              struct ib_device *ib_dev,
3746                              struct ib_udata *ucore,
3747                              struct ib_udata *uhw)
3748 {
3749         struct ib_uverbs_create_flow      cmd;
3750         struct ib_uverbs_create_flow_resp resp;
3751         struct ib_uobject                 *uobj;
3752         struct ib_flow                    *flow_id;
3753         struct ib_uverbs_flow_attr        *kern_flow_attr;
3754         struct ib_flow_attr               *flow_attr;
3755         struct ib_qp                      *qp;
3756         int err = 0;
3757         void *kern_spec;
3758         void *ib_spec;
3759         int i;
3760
3761         if (ucore->inlen < sizeof(cmd))
3762                 return -EINVAL;
3763
3764         if (ucore->outlen < sizeof(resp))
3765                 return -ENOSPC;
3766
3767         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3768         if (err)
3769                 return err;
3770
3771         ucore->inbuf += sizeof(cmd);
3772         ucore->inlen -= sizeof(cmd);
3773
3774         if (cmd.comp_mask)
3775                 return -EINVAL;
3776
3777         if (!capable(CAP_NET_RAW))
3778                 return -EPERM;
3779
3780         if (cmd.flow_attr.flags >= IB_FLOW_ATTR_FLAGS_RESERVED)
3781                 return -EINVAL;
3782
3783         if ((cmd.flow_attr.flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) &&
3784             ((cmd.flow_attr.type == IB_FLOW_ATTR_ALL_DEFAULT) ||
3785              (cmd.flow_attr.type == IB_FLOW_ATTR_MC_DEFAULT)))
3786                 return -EINVAL;
3787
3788         if (cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
3789                 return -EINVAL;
3790
3791         if (cmd.flow_attr.size > ucore->inlen ||
3792             cmd.flow_attr.size >
3793             (cmd.flow_attr.num_of_specs * sizeof(struct ib_uverbs_flow_spec)))
3794                 return -EINVAL;
3795
3796         if (cmd.flow_attr.reserved[0] ||
3797             cmd.flow_attr.reserved[1])
3798                 return -EINVAL;
3799
3800         if (cmd.flow_attr.num_of_specs) {
3801                 kern_flow_attr = kmalloc(sizeof(*kern_flow_attr) + cmd.flow_attr.size,
3802                                          GFP_KERNEL);
3803                 if (!kern_flow_attr)
3804                         return -ENOMEM;
3805
3806                 memcpy(kern_flow_attr, &cmd.flow_attr, sizeof(*kern_flow_attr));
3807                 err = ib_copy_from_udata(kern_flow_attr + 1, ucore,
3808                                          cmd.flow_attr.size);
3809                 if (err)
3810                         goto err_free_attr;
3811         } else {
3812                 kern_flow_attr = &cmd.flow_attr;
3813         }
3814
3815         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
3816         if (!uobj) {
3817                 err = -ENOMEM;
3818                 goto err_free_attr;
3819         }
3820         init_uobj(uobj, 0, file->ucontext, &rule_lock_class);
3821         down_write(&uobj->mutex);
3822
3823         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
3824         if (!qp) {
3825                 err = -EINVAL;
3826                 goto err_uobj;
3827         }
3828
3829         flow_attr = kzalloc(sizeof(*flow_attr) + cmd.flow_attr.num_of_specs *
3830                             sizeof(union ib_flow_spec), GFP_KERNEL);
3831         if (!flow_attr) {
3832                 err = -ENOMEM;
3833                 goto err_put;
3834         }
3835
3836         flow_attr->type = kern_flow_attr->type;
3837         flow_attr->priority = kern_flow_attr->priority;
3838         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
3839         flow_attr->port = kern_flow_attr->port;
3840         flow_attr->flags = kern_flow_attr->flags;
3841         flow_attr->size = sizeof(*flow_attr);
3842
3843         kern_spec = kern_flow_attr + 1;
3844         ib_spec = flow_attr + 1;
3845         for (i = 0; i < flow_attr->num_of_specs &&
3846              cmd.flow_attr.size > offsetof(struct ib_uverbs_flow_spec, reserved) &&
3847              cmd.flow_attr.size >=
3848              ((struct ib_uverbs_flow_spec *)kern_spec)->size; i++) {
3849                 err = kern_spec_to_ib_spec(kern_spec, ib_spec);
3850                 if (err)
3851                         goto err_free;
3852                 flow_attr->size +=
3853                         ((union ib_flow_spec *) ib_spec)->size;
3854                 cmd.flow_attr.size -= ((struct ib_uverbs_flow_spec *)kern_spec)->size;
3855                 kern_spec += ((struct ib_uverbs_flow_spec *) kern_spec)->size;
3856                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
3857         }
3858         if (cmd.flow_attr.size || (i != flow_attr->num_of_specs)) {
3859                 pr_warn("create flow failed, flow %d: %d bytes left from uverb cmd\n",
3860                         i, cmd.flow_attr.size);
3861                 err = -EINVAL;
3862                 goto err_free;
3863         }
3864         flow_id = ib_create_flow(qp, flow_attr, IB_FLOW_DOMAIN_USER);
3865         if (IS_ERR(flow_id)) {
3866                 err = PTR_ERR(flow_id);
3867                 goto err_free;
3868         }
3869         flow_id->uobject = uobj;
3870         uobj->object = flow_id;
3871
3872         err = idr_add_uobj(&ib_uverbs_rule_idr, uobj);
3873         if (err)
3874                 goto destroy_flow;
3875
3876         memset(&resp, 0, sizeof(resp));
3877         resp.flow_handle = uobj->id;
3878
3879         err = ib_copy_to_udata(ucore,
3880                                &resp, sizeof(resp));
3881         if (err)
3882                 goto err_copy;
3883
3884         put_qp_read(qp);
3885         mutex_lock(&file->mutex);
3886         list_add_tail(&uobj->list, &file->ucontext->rule_list);
3887         mutex_unlock(&file->mutex);
3888
3889         uobj->live = 1;
3890
3891         up_write(&uobj->mutex);
3892         kfree(flow_attr);
3893         if (cmd.flow_attr.num_of_specs)
3894                 kfree(kern_flow_attr);
3895         return 0;
3896 err_copy:
3897         idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
3898 destroy_flow:
3899         ib_destroy_flow(flow_id);
3900 err_free:
3901         kfree(flow_attr);
3902 err_put:
3903         put_qp_read(qp);
3904 err_uobj:
3905         put_uobj_write(uobj);
3906 err_free_attr:
3907         if (cmd.flow_attr.num_of_specs)
3908                 kfree(kern_flow_attr);
3909         return err;
3910 }
3911
3912 int ib_uverbs_ex_destroy_flow(struct ib_uverbs_file *file,
3913                               struct ib_device *ib_dev,
3914                               struct ib_udata *ucore,
3915                               struct ib_udata *uhw)
3916 {
3917         struct ib_uverbs_destroy_flow   cmd;
3918         struct ib_flow                  *flow_id;
3919         struct ib_uobject               *uobj;
3920         int                             ret;
3921
3922         if (ucore->inlen < sizeof(cmd))
3923                 return -EINVAL;
3924
3925         ret = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3926         if (ret)
3927                 return ret;
3928
3929         if (cmd.comp_mask)
3930                 return -EINVAL;
3931
3932         uobj = idr_write_uobj(&ib_uverbs_rule_idr, cmd.flow_handle,
3933                               file->ucontext);
3934         if (!uobj)
3935                 return -EINVAL;
3936         flow_id = uobj->object;
3937
3938         ret = ib_destroy_flow(flow_id);
3939         if (!ret)
3940                 uobj->live = 0;
3941
3942         put_uobj_write(uobj);
3943
3944         idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
3945
3946         mutex_lock(&file->mutex);
3947         list_del(&uobj->list);
3948         mutex_unlock(&file->mutex);
3949
3950         put_uobj(uobj);
3951
3952         return ret;
3953 }
3954
3955 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
3956                                 struct ib_device *ib_dev,
3957                                 struct ib_uverbs_create_xsrq *cmd,
3958                                 struct ib_udata *udata)
3959 {
3960         struct ib_uverbs_create_srq_resp resp;
3961         struct ib_usrq_object           *obj;
3962         struct ib_pd                    *pd;
3963         struct ib_srq                   *srq;
3964         struct ib_uobject               *uninitialized_var(xrcd_uobj);
3965         struct ib_srq_init_attr          attr;
3966         int ret;
3967
3968         obj = kmalloc(sizeof *obj, GFP_KERNEL);
3969         if (!obj)
3970                 return -ENOMEM;
3971
3972         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext, &srq_lock_class);
3973         down_write(&obj->uevent.uobject.mutex);
3974
3975         if (cmd->srq_type == IB_SRQT_XRC) {
3976                 attr.ext.xrc.xrcd  = idr_read_xrcd(cmd->xrcd_handle, file->ucontext, &xrcd_uobj);
3977                 if (!attr.ext.xrc.xrcd) {
3978                         ret = -EINVAL;
3979                         goto err;
3980                 }
3981
3982                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
3983                 atomic_inc(&obj->uxrcd->refcnt);
3984
3985                 attr.ext.xrc.cq  = idr_read_cq(cmd->cq_handle, file->ucontext, 0);
3986                 if (!attr.ext.xrc.cq) {
3987                         ret = -EINVAL;
3988                         goto err_put_xrcd;
3989                 }
3990         }
3991
3992         pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
3993         if (!pd) {
3994                 ret = -EINVAL;
3995                 goto err_put_cq;
3996         }
3997
3998         attr.event_handler  = ib_uverbs_srq_event_handler;
3999         attr.srq_context    = file;
4000         attr.srq_type       = cmd->srq_type;
4001         attr.attr.max_wr    = cmd->max_wr;
4002         attr.attr.max_sge   = cmd->max_sge;
4003         attr.attr.srq_limit = cmd->srq_limit;
4004
4005         obj->uevent.events_reported = 0;
4006         INIT_LIST_HEAD(&obj->uevent.event_list);
4007
4008         srq = pd->device->create_srq(pd, &attr, udata);
4009         if (IS_ERR(srq)) {
4010                 ret = PTR_ERR(srq);
4011                 goto err_put;
4012         }
4013
4014         srq->device        = pd->device;
4015         srq->pd            = pd;
4016         srq->srq_type      = cmd->srq_type;
4017         srq->uobject       = &obj->uevent.uobject;
4018         srq->event_handler = attr.event_handler;
4019         srq->srq_context   = attr.srq_context;
4020
4021         if (cmd->srq_type == IB_SRQT_XRC) {
4022                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
4023                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
4024                 atomic_inc(&attr.ext.xrc.cq->usecnt);
4025                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
4026         }
4027
4028         atomic_inc(&pd->usecnt);
4029         atomic_set(&srq->usecnt, 0);
4030
4031         obj->uevent.uobject.object = srq;
4032         ret = idr_add_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
4033         if (ret)
4034                 goto err_destroy;
4035
4036         memset(&resp, 0, sizeof resp);
4037         resp.srq_handle = obj->uevent.uobject.id;
4038         resp.max_wr     = attr.attr.max_wr;
4039         resp.max_sge    = attr.attr.max_sge;
4040         if (cmd->srq_type == IB_SRQT_XRC)
4041                 resp.srqn = srq->ext.xrc.srq_num;
4042
4043         if (copy_to_user((void __user *) (unsigned long) cmd->response,
4044                          &resp, sizeof resp)) {
4045                 ret = -EFAULT;
4046                 goto err_copy;
4047         }
4048
4049         if (cmd->srq_type == IB_SRQT_XRC) {
4050                 put_uobj_read(xrcd_uobj);
4051                 put_cq_read(attr.ext.xrc.cq);
4052         }
4053         put_pd_read(pd);
4054
4055         mutex_lock(&file->mutex);
4056         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->srq_list);
4057         mutex_unlock(&file->mutex);
4058
4059         obj->uevent.uobject.live = 1;
4060
4061         up_write(&obj->uevent.uobject.mutex);
4062
4063         return 0;
4064
4065 err_copy:
4066         idr_remove_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
4067
4068 err_destroy:
4069         ib_destroy_srq(srq);
4070
4071 err_put:
4072         put_pd_read(pd);
4073
4074 err_put_cq:
4075         if (cmd->srq_type == IB_SRQT_XRC)
4076                 put_cq_read(attr.ext.xrc.cq);
4077
4078 err_put_xrcd:
4079         if (cmd->srq_type == IB_SRQT_XRC) {
4080                 atomic_dec(&obj->uxrcd->refcnt);
4081                 put_uobj_read(xrcd_uobj);
4082         }
4083
4084 err:
4085         put_uobj_write(&obj->uevent.uobject);
4086         return ret;
4087 }
4088
4089 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
4090                              struct ib_device *ib_dev,
4091                              const char __user *buf, int in_len,
4092                              int out_len)
4093 {
4094         struct ib_uverbs_create_srq      cmd;
4095         struct ib_uverbs_create_xsrq     xcmd;
4096         struct ib_uverbs_create_srq_resp resp;
4097         struct ib_udata                  udata;
4098         int ret;
4099
4100         if (out_len < sizeof resp)
4101                 return -ENOSPC;
4102
4103         if (copy_from_user(&cmd, buf, sizeof cmd))
4104                 return -EFAULT;
4105
4106         xcmd.response    = cmd.response;
4107         xcmd.user_handle = cmd.user_handle;
4108         xcmd.srq_type    = IB_SRQT_BASIC;
4109         xcmd.pd_handle   = cmd.pd_handle;
4110         xcmd.max_wr      = cmd.max_wr;
4111         xcmd.max_sge     = cmd.max_sge;
4112         xcmd.srq_limit   = cmd.srq_limit;
4113
4114         INIT_UDATA(&udata, buf + sizeof cmd,
4115                    (unsigned long) cmd.response + sizeof resp,
4116                    in_len - sizeof cmd - sizeof(struct ib_uverbs_cmd_hdr),
4117                    out_len - sizeof resp);
4118
4119         ret = __uverbs_create_xsrq(file, ib_dev, &xcmd, &udata);
4120         if (ret)
4121                 return ret;
4122
4123         return in_len;
4124 }
4125
4126 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
4127                               struct ib_device *ib_dev,
4128                               const char __user *buf, int in_len, int out_len)
4129 {
4130         struct ib_uverbs_create_xsrq     cmd;
4131         struct ib_uverbs_create_srq_resp resp;
4132         struct ib_udata                  udata;
4133         int ret;
4134
4135         if (out_len < sizeof resp)
4136                 return -ENOSPC;
4137
4138         if (copy_from_user(&cmd, buf, sizeof cmd))
4139                 return -EFAULT;
4140
4141         INIT_UDATA(&udata, buf + sizeof cmd,
4142                    (unsigned long) cmd.response + sizeof resp,
4143                    in_len - sizeof cmd - sizeof(struct ib_uverbs_cmd_hdr),
4144                    out_len - sizeof resp);
4145
4146         ret = __uverbs_create_xsrq(file, ib_dev, &cmd, &udata);
4147         if (ret)
4148                 return ret;
4149
4150         return in_len;
4151 }
4152
4153 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
4154                              struct ib_device *ib_dev,
4155                              const char __user *buf, int in_len,
4156                              int out_len)
4157 {
4158         struct ib_uverbs_modify_srq cmd;
4159         struct ib_udata             udata;
4160         struct ib_srq              *srq;
4161         struct ib_srq_attr          attr;
4162         int                         ret;
4163
4164         if (copy_from_user(&cmd, buf, sizeof cmd))
4165                 return -EFAULT;
4166
4167         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
4168                    out_len);
4169
4170         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
4171         if (!srq)
4172                 return -EINVAL;
4173
4174         attr.max_wr    = cmd.max_wr;
4175         attr.srq_limit = cmd.srq_limit;
4176
4177         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
4178
4179         put_srq_read(srq);
4180
4181         return ret ? ret : in_len;
4182 }
4183
4184 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
4185                             struct ib_device *ib_dev,
4186                             const char __user *buf,
4187                             int in_len, int out_len)
4188 {
4189         struct ib_uverbs_query_srq      cmd;
4190         struct ib_uverbs_query_srq_resp resp;
4191         struct ib_srq_attr              attr;
4192         struct ib_srq                   *srq;
4193         int                             ret;
4194
4195         if (out_len < sizeof resp)
4196                 return -ENOSPC;
4197
4198         if (copy_from_user(&cmd, buf, sizeof cmd))
4199                 return -EFAULT;
4200
4201         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
4202         if (!srq)
4203                 return -EINVAL;
4204
4205         ret = ib_query_srq(srq, &attr);
4206
4207         put_srq_read(srq);
4208
4209         if (ret)
4210                 return ret;
4211
4212         memset(&resp, 0, sizeof resp);
4213
4214         resp.max_wr    = attr.max_wr;
4215         resp.max_sge   = attr.max_sge;
4216         resp.srq_limit = attr.srq_limit;
4217
4218         if (copy_to_user((void __user *) (unsigned long) cmd.response,
4219                          &resp, sizeof resp))
4220                 return -EFAULT;
4221
4222         return in_len;
4223 }
4224
4225 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
4226                               struct ib_device *ib_dev,
4227                               const char __user *buf, int in_len,
4228                               int out_len)
4229 {
4230         struct ib_uverbs_destroy_srq      cmd;
4231         struct ib_uverbs_destroy_srq_resp resp;
4232         struct ib_uobject                *uobj;
4233         struct ib_srq                    *srq;
4234         struct ib_uevent_object          *obj;
4235         int                               ret = -EINVAL;
4236         struct ib_usrq_object            *us;
4237         enum ib_srq_type                  srq_type;
4238
4239         if (copy_from_user(&cmd, buf, sizeof cmd))
4240                 return -EFAULT;
4241
4242         uobj = idr_write_uobj(&ib_uverbs_srq_idr, cmd.srq_handle, file->ucontext);
4243         if (!uobj)
4244                 return -EINVAL;
4245         srq = uobj->object;
4246         obj = container_of(uobj, struct ib_uevent_object, uobject);
4247         srq_type = srq->srq_type;
4248
4249         ret = ib_destroy_srq(srq);
4250         if (!ret)
4251                 uobj->live = 0;
4252
4253         put_uobj_write(uobj);
4254
4255         if (ret)
4256                 return ret;
4257
4258         if (srq_type == IB_SRQT_XRC) {
4259                 us = container_of(obj, struct ib_usrq_object, uevent);
4260                 atomic_dec(&us->uxrcd->refcnt);
4261         }
4262
4263         idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
4264
4265         mutex_lock(&file->mutex);
4266         list_del(&uobj->list);
4267         mutex_unlock(&file->mutex);
4268
4269         ib_uverbs_release_uevent(file, obj);
4270
4271         memset(&resp, 0, sizeof resp);
4272         resp.events_reported = obj->events_reported;
4273
4274         put_uobj(uobj);
4275
4276         if (copy_to_user((void __user *) (unsigned long) cmd.response,
4277                          &resp, sizeof resp))
4278                 ret = -EFAULT;
4279
4280         return ret ? ret : in_len;
4281 }
4282
4283 int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
4284                               struct ib_device *ib_dev,
4285                               struct ib_udata *ucore,
4286                               struct ib_udata *uhw)
4287 {
4288         struct ib_uverbs_ex_query_device_resp resp = { {0} };
4289         struct ib_uverbs_ex_query_device  cmd;
4290         struct ib_device_attr attr = {0};
4291         int err;
4292
4293         if (ucore->inlen < sizeof(cmd))
4294                 return -EINVAL;
4295
4296         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
4297         if (err)
4298                 return err;
4299
4300         if (cmd.comp_mask)
4301                 return -EINVAL;
4302
4303         if (cmd.reserved)
4304                 return -EINVAL;
4305
4306         resp.response_length = offsetof(typeof(resp), odp_caps);
4307
4308         if (ucore->outlen < resp.response_length)
4309                 return -ENOSPC;
4310
4311         err = ib_dev->query_device(ib_dev, &attr, uhw);
4312         if (err)
4313                 return err;
4314
4315         copy_query_dev_fields(file, ib_dev, &resp.base, &attr);
4316
4317         if (ucore->outlen < resp.response_length + sizeof(resp.odp_caps))
4318                 goto end;
4319
4320 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4321         resp.odp_caps.general_caps = attr.odp_caps.general_caps;
4322         resp.odp_caps.per_transport_caps.rc_odp_caps =
4323                 attr.odp_caps.per_transport_caps.rc_odp_caps;
4324         resp.odp_caps.per_transport_caps.uc_odp_caps =
4325                 attr.odp_caps.per_transport_caps.uc_odp_caps;
4326         resp.odp_caps.per_transport_caps.ud_odp_caps =
4327                 attr.odp_caps.per_transport_caps.ud_odp_caps;
4328 #endif
4329         resp.response_length += sizeof(resp.odp_caps);
4330
4331         if (ucore->outlen < resp.response_length + sizeof(resp.timestamp_mask))
4332                 goto end;
4333
4334         resp.timestamp_mask = attr.timestamp_mask;
4335         resp.response_length += sizeof(resp.timestamp_mask);
4336
4337         if (ucore->outlen < resp.response_length + sizeof(resp.hca_core_clock))
4338                 goto end;
4339
4340         resp.hca_core_clock = attr.hca_core_clock;
4341         resp.response_length += sizeof(resp.hca_core_clock);
4342
4343         if (ucore->outlen < resp.response_length + sizeof(resp.device_cap_flags_ex))
4344                 goto end;
4345
4346         resp.device_cap_flags_ex = attr.device_cap_flags;
4347         resp.response_length += sizeof(resp.device_cap_flags_ex);
4348
4349         if (ucore->outlen < resp.response_length + sizeof(resp.rss_caps))
4350                 goto end;
4351
4352         resp.rss_caps.supported_qpts = attr.rss_caps.supported_qpts;
4353         resp.rss_caps.max_rwq_indirection_tables =
4354                 attr.rss_caps.max_rwq_indirection_tables;
4355         resp.rss_caps.max_rwq_indirection_table_size =
4356                 attr.rss_caps.max_rwq_indirection_table_size;
4357
4358         resp.response_length += sizeof(resp.rss_caps);
4359
4360         if (ucore->outlen < resp.response_length + sizeof(resp.max_wq_type_rq))
4361                 goto end;
4362
4363         resp.max_wq_type_rq = attr.max_wq_type_rq;
4364         resp.response_length += sizeof(resp.max_wq_type_rq);
4365
4366         if (ucore->outlen < resp.response_length + sizeof(resp.raw_packet_caps))
4367                 goto end;
4368
4369         resp.raw_packet_caps = attr.raw_packet_caps;
4370         resp.response_length += sizeof(resp.raw_packet_caps);
4371 end:
4372         err = ib_copy_to_udata(ucore, &resp, resp.response_length);
4373         return err;
4374 }