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         struct ib_rdmacg_object          cg_obj;
 320         int ret;
 321
 322         if (out_len < sizeof resp)
 323                 return -ENOSPC;
 324
 325         if (copy_from_user(&cmd, buf, sizeof cmd))
 326                 return -EFAULT;
 327
 328         mutex_lock(&file->mutex);
 329
 330         if (file->ucontext) {
 331                 ret = -EINVAL;
 332                 goto err;
 333         }
 334
 335         INIT_UDATA(&udata, buf + sizeof cmd,
 336                    (unsigned long) cmd.response + sizeof resp,
 337                    in_len - sizeof cmd, out_len - sizeof resp);
 338
 339         ret = ib_rdmacg_try_charge(&cg_obj, ib_dev, RDMACG_RESOURCE_HCA_HANDLE);
 340         if (ret)
 341                 goto err;
 342
 343         ucontext = ib_dev->alloc_ucontext(ib_dev, &udata);
 344         if (IS_ERR(ucontext)) {
 345                 ret = PTR_ERR(ucontext);
 346                 goto err_alloc;
 347         }
 348
 349         ucontext->device = ib_dev;
 350         ucontext->cg_obj = cg_obj;
 351         INIT_LIST_HEAD(&ucontext->pd_list);
 352         INIT_LIST_HEAD(&ucontext->mr_list);
 353         INIT_LIST_HEAD(&ucontext->mw_list);
 354         INIT_LIST_HEAD(&ucontext->cq_list);
 355         INIT_LIST_HEAD(&ucontext->qp_list);
 356         INIT_LIST_HEAD(&ucontext->srq_list);
 357         INIT_LIST_HEAD(&ucontext->ah_list);
 358         INIT_LIST_HEAD(&ucontext->wq_list);
 359         INIT_LIST_HEAD(&ucontext->rwq_ind_tbl_list);
 360         INIT_LIST_HEAD(&ucontext->xrcd_list);
 361         INIT_LIST_HEAD(&ucontext->rule_list);
 362         rcu_read_lock();
 363         ucontext->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
 364         rcu_read_unlock();
 365         ucontext->closing = 0;
 366
 367 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 368         ucontext->umem_tree = RB_ROOT;
 369         init_rwsem(&ucontext->umem_rwsem);
 370         ucontext->odp_mrs_count = 0;
 371         INIT_LIST_HEAD(&ucontext->no_private_counters);
 372
 373         if (!(ib_dev->attrs.device_cap_flags & IB_DEVICE_ON_DEMAND_PAGING))
 374                 ucontext->invalidate_range = NULL;
 375
 376 #endif
 377
 378         resp.num_comp_vectors = file->device->num_comp_vectors;
 379
 380         ret = get_unused_fd_flags(O_CLOEXEC);
 381         if (ret < 0)
 382                 goto err_free;
 383         resp.async_fd = ret;
 384
 385         filp = ib_uverbs_alloc_event_file(file, ib_dev, 1);
 386         if (IS_ERR(filp)) {
 387                 ret = PTR_ERR(filp);
 388                 goto err_fd;
 389         }
 390
 391         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 392                          &resp, sizeof resp)) {
 393                 ret = -EFAULT;
 394                 goto err_file;
 395         }
 396
 397         file->ucontext = ucontext;
 398
 399         fd_install(resp.async_fd, filp);
 400
 401         mutex_unlock(&file->mutex);
 402
 403         return in_len;
 404
 405 err_file:
 406         ib_uverbs_free_async_event_file(file);
 407         fput(filp);
 408
 409 err_fd:
 410         put_unused_fd(resp.async_fd);
 411
 412 err_free:
 413         put_pid(ucontext->tgid);
 414         ib_dev->dealloc_ucontext(ucontext);
 415
 416 err_alloc:
 417         ib_rdmacg_uncharge(&cg_obj, ib_dev, RDMACG_RESOURCE_HCA_HANDLE);
 418
 419 err:
 420         mutex_unlock(&file->mutex);
 421         return ret;
 422 }
 423
 424 static void copy_query_dev_fields(struct ib_uverbs_file *file,
 425                                   struct ib_device *ib_dev,
 426                                   struct ib_uverbs_query_device_resp *resp,
 427                                   struct ib_device_attr *attr)
 428 {
 429         resp->fw_ver            = attr->fw_ver;
 430         resp->node_guid         = ib_dev->node_guid;
 431         resp->sys_image_guid    = attr->sys_image_guid;
 432         resp->max_mr_size       = attr->max_mr_size;
 433         resp->page_size_cap     = attr->page_size_cap;
 434         resp->vendor_id         = attr->vendor_id;
 435         resp->vendor_part_id    = attr->vendor_part_id;
 436         resp->hw_ver            = attr->hw_ver;
 437         resp->max_qp            = attr->max_qp;
 438         resp->max_qp_wr         = attr->max_qp_wr;
 439         resp->device_cap_flags  = lower_32_bits(attr->device_cap_flags);
 440         resp->max_sge           = attr->max_sge;
 441         resp->max_sge_rd        = attr->max_sge_rd;
 442         resp->max_cq            = attr->max_cq;
 443         resp->max_cqe           = attr->max_cqe;
 444         resp->max_mr            = attr->max_mr;
 445         resp->max_pd            = attr->max_pd;
 446         resp->max_qp_rd_atom    = attr->max_qp_rd_atom;
 447         resp->max_ee_rd_atom    = attr->max_ee_rd_atom;
 448         resp->max_res_rd_atom   = attr->max_res_rd_atom;
 449         resp->max_qp_init_rd_atom       = attr->max_qp_init_rd_atom;
 450         resp->max_ee_init_rd_atom       = attr->max_ee_init_rd_atom;
 451         resp->atomic_cap                = attr->atomic_cap;
 452         resp->max_ee                    = attr->max_ee;
 453         resp->max_rdd                   = attr->max_rdd;
 454         resp->max_mw                    = attr->max_mw;
 455         resp->max_raw_ipv6_qp           = attr->max_raw_ipv6_qp;
 456         resp->max_raw_ethy_qp           = attr->max_raw_ethy_qp;
 457         resp->max_mcast_grp             = attr->max_mcast_grp;
 458         resp->max_mcast_qp_attach       = attr->max_mcast_qp_attach;
 459         resp->max_total_mcast_qp_attach = attr->max_total_mcast_qp_attach;
 460         resp->max_ah                    = attr->max_ah;
 461         resp->max_fmr                   = attr->max_fmr;
 462         resp->max_map_per_fmr           = attr->max_map_per_fmr;
 463         resp->max_srq                   = attr->max_srq;
 464         resp->max_srq_wr                = attr->max_srq_wr;
 465         resp->max_srq_sge               = attr->max_srq_sge;
 466         resp->max_pkeys                 = attr->max_pkeys;
 467         resp->local_ca_ack_delay        = attr->local_ca_ack_delay;
 468         resp->phys_port_cnt             = ib_dev->phys_port_cnt;
 469 }
 470
 471 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 472                                struct ib_device *ib_dev,
 473                                const char __user *buf,
 474                                int in_len, int out_len)
 475 {
 476         struct ib_uverbs_query_device      cmd;
 477         struct ib_uverbs_query_device_resp resp;
 478
 479         if (out_len < sizeof resp)
 480                 return -ENOSPC;
 481
 482         if (copy_from_user(&cmd, buf, sizeof cmd))
 483                 return -EFAULT;
 484
 485         memset(&resp, 0, sizeof resp);
 486         copy_query_dev_fields(file, ib_dev, &resp, &ib_dev->attrs);
 487
 488         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 489                          &resp, sizeof resp))
 490                 return -EFAULT;
 491
 492         return in_len;
 493 }
 494
 495 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 496                              struct ib_device *ib_dev,
 497                              const char __user *buf,
 498                              int in_len, int out_len)
 499 {
 500         struct ib_uverbs_query_port      cmd;
 501         struct ib_uverbs_query_port_resp resp;
 502         struct ib_port_attr              attr;
 503         int                              ret;
 504
 505         if (out_len < sizeof resp)
 506                 return -ENOSPC;
 507
 508         if (copy_from_user(&cmd, buf, sizeof cmd))
 509                 return -EFAULT;
 510
 511         ret = ib_query_port(ib_dev, cmd.port_num, &attr);
 512         if (ret)
 513                 return ret;
 514
 515         memset(&resp, 0, sizeof resp);
 516
 517         resp.state           = attr.state;
 518         resp.max_mtu         = attr.max_mtu;
 519         resp.active_mtu      = attr.active_mtu;
 520         resp.gid_tbl_len     = attr.gid_tbl_len;
 521         resp.port_cap_flags  = attr.port_cap_flags;
 522         resp.max_msg_sz      = attr.max_msg_sz;
 523         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 524         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 525         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 526         resp.lid             = attr.lid;
 527         resp.sm_lid          = attr.sm_lid;
 528         resp.lmc             = attr.lmc;
 529         resp.max_vl_num      = attr.max_vl_num;
 530         resp.sm_sl           = attr.sm_sl;
 531         resp.subnet_timeout  = attr.subnet_timeout;
 532         resp.init_type_reply = attr.init_type_reply;
 533         resp.active_width    = attr.active_width;
 534         resp.active_speed    = attr.active_speed;
 535         resp.phys_state      = attr.phys_state;
 536         resp.link_layer      = rdma_port_get_link_layer(ib_dev,
 537                                                         cmd.port_num);
 538
 539         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 540                          &resp, sizeof resp))
 541                 return -EFAULT;
 542
 543         return in_len;
 544 }
 545
 546 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 547                            struct ib_device *ib_dev,
 548                            const char __user *buf,
 549                            int in_len, int out_len)
 550 {
 551         struct ib_uverbs_alloc_pd      cmd;
 552         struct ib_uverbs_alloc_pd_resp resp;
 553         struct ib_udata                udata;
 554         struct ib_uobject             *uobj;
 555         struct ib_pd                  *pd;
 556         int                            ret;
 557
 558         if (out_len < sizeof resp)
 559                 return -ENOSPC;
 560
 561         if (copy_from_user(&cmd, buf, sizeof cmd))
 562                 return -EFAULT;
 563
 564         INIT_UDATA(&udata, buf + sizeof cmd,
 565                    (unsigned long) cmd.response + sizeof resp,
 566                    in_len - sizeof cmd, out_len - sizeof resp);
 567
 568         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 569         if (!uobj)
 570                 return -ENOMEM;
 571
 572         init_uobj(uobj, 0, file->ucontext, &pd_lock_class);
 573         ret = ib_rdmacg_try_charge(&uobj->cg_obj, ib_dev,
 574                                    RDMACG_RESOURCE_HCA_OBJECT);
 575         if (ret) {
 576                 kfree(uobj);
 577                 return ret;
 578         }
 579
 580         down_write(&uobj->mutex);
 581
 582         pd = ib_dev->alloc_pd(ib_dev, file->ucontext, &udata);
 583         if (IS_ERR(pd)) {
 584                 ret = PTR_ERR(pd);
 585                 goto err;
 586         }
 587
 588         pd->device  = ib_dev;
 589         pd->uobject = uobj;
 590         pd->__internal_mr = NULL;
 591         atomic_set(&pd->usecnt, 0);
 592
 593         uobj->object = pd;
 594         ret = idr_add_uobj(&ib_uverbs_pd_idr, uobj);
 595         if (ret)
 596                 goto err_idr;
 597
 598         memset(&resp, 0, sizeof resp);
 599         resp.pd_handle = uobj->id;
 600
 601         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 602                          &resp, sizeof resp)) {
 603                 ret = -EFAULT;
 604                 goto err_copy;
 605         }
 606
 607         mutex_lock(&file->mutex);
 608         list_add_tail(&uobj->list, &file->ucontext->pd_list);
 609         mutex_unlock(&file->mutex);
 610
 611         uobj->live = 1;
 612
 613         up_write(&uobj->mutex);
 614
 615         return in_len;
 616
 617 err_copy:
 618         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 619
 620 err_idr:
 621         ib_dealloc_pd(pd);
 622
 623 err:
 624         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
 625         put_uobj_write(uobj);
 626         return ret;
 627 }
 628
 629 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 630                              struct ib_device *ib_dev,
 631                              const char __user *buf,
 632                              int in_len, int out_len)
 633 {
 634         struct ib_uverbs_dealloc_pd cmd;
 635         struct ib_uobject          *uobj;
 636         struct ib_pd               *pd;
 637         int                         ret;
 638
 639         if (copy_from_user(&cmd, buf, sizeof cmd))
 640                 return -EFAULT;
 641
 642         uobj = idr_write_uobj(&ib_uverbs_pd_idr, cmd.pd_handle, file->ucontext);
 643         if (!uobj)
 644                 return -EINVAL;
 645         pd = uobj->object;
 646
 647         if (atomic_read(&pd->usecnt)) {
 648                 ret = -EBUSY;
 649                 goto err_put;
 650         }
 651
 652         ret = pd->device->dealloc_pd(uobj->object);
 653         WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
 654         if (ret)
 655                 goto err_put;
 656
 657         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
 658
 659         uobj->live = 0;
 660         put_uobj_write(uobj);
 661
 662         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 663
 664         mutex_lock(&file->mutex);
 665         list_del(&uobj->list);
 666         mutex_unlock(&file->mutex);
 667
 668         put_uobj(uobj);
 669
 670         return in_len;
 671
 672 err_put:
 673         put_uobj_write(uobj);
 674         return ret;
 675 }
 676
 677 struct xrcd_table_entry {
 678         struct rb_node  node;
 679         struct ib_xrcd *xrcd;
 680         struct inode   *inode;
 681 };
 682
 683 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 684                             struct inode *inode,
 685                             struct ib_xrcd *xrcd)
 686 {
 687         struct xrcd_table_entry *entry, *scan;
 688         struct rb_node **p = &dev->xrcd_tree.rb_node;
 689         struct rb_node *parent = NULL;
 690
 691         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 692         if (!entry)
 693                 return -ENOMEM;
 694
 695         entry->xrcd  = xrcd;
 696         entry->inode = inode;
 697
 698         while (*p) {
 699                 parent = *p;
 700                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 701
 702                 if (inode < scan->inode) {
 703                         p = &(*p)->rb_left;
 704                 } else if (inode > scan->inode) {
 705                         p = &(*p)->rb_right;
 706                 } else {
 707                         kfree(entry);
 708                         return -EEXIST;
 709                 }
 710         }
 711
 712         rb_link_node(&entry->node, parent, p);
 713         rb_insert_color(&entry->node, &dev->xrcd_tree);
 714         igrab(inode);
 715         return 0;
 716 }
 717
 718 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 719                                                   struct inode *inode)
 720 {
 721         struct xrcd_table_entry *entry;
 722         struct rb_node *p = dev->xrcd_tree.rb_node;
 723
 724         while (p) {
 725                 entry = rb_entry(p, struct xrcd_table_entry, node);
 726
 727                 if (inode < entry->inode)
 728                         p = p->rb_left;
 729                 else if (inode > entry->inode)
 730                         p = p->rb_right;
 731                 else
 732                         return entry;
 733         }
 734
 735         return NULL;
 736 }
 737
 738 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 739 {
 740         struct xrcd_table_entry *entry;
 741
 742         entry = xrcd_table_search(dev, inode);
 743         if (!entry)
 744                 return NULL;
 745
 746         return entry->xrcd;
 747 }
 748
 749 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 750                               struct inode *inode)
 751 {
 752         struct xrcd_table_entry *entry;
 753
 754         entry = xrcd_table_search(dev, inode);
 755         if (entry) {
 756                 iput(inode);
 757                 rb_erase(&entry->node, &dev->xrcd_tree);
 758                 kfree(entry);
 759         }
 760 }
 761
 762 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 763                             struct ib_device *ib_dev,
 764                             const char __user *buf, int in_len,
 765                             int out_len)
 766 {
 767         struct ib_uverbs_open_xrcd      cmd;
 768         struct ib_uverbs_open_xrcd_resp resp;
 769         struct ib_udata                 udata;
 770         struct ib_uxrcd_object         *obj;
 771         struct ib_xrcd                 *xrcd = NULL;
 772         struct fd                       f = {NULL, 0};
 773         struct inode                   *inode = NULL;
 774         int                             ret = 0;
 775         int                             new_xrcd = 0;
 776
 777         if (out_len < sizeof resp)
 778                 return -ENOSPC;
 779
 780         if (copy_from_user(&cmd, buf, sizeof cmd))
 781                 return -EFAULT;
 782
 783         INIT_UDATA(&udata, buf + sizeof cmd,
 784                    (unsigned long) cmd.response + sizeof resp,
 785                    in_len - sizeof cmd, out_len - sizeof  resp);
 786
 787         mutex_lock(&file->device->xrcd_tree_mutex);
 788
 789         if (cmd.fd != -1) {
 790                 /* search for file descriptor */
 791                 f = fdget(cmd.fd);
 792                 if (!f.file) {
 793                         ret = -EBADF;
 794                         goto err_tree_mutex_unlock;
 795                 }
 796
 797                 inode = file_inode(f.file);
 798                 xrcd = find_xrcd(file->device, inode);
 799                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 800                         /* no file descriptor. Need CREATE flag */
 801                         ret = -EAGAIN;
 802                         goto err_tree_mutex_unlock;
 803                 }
 804
 805                 if (xrcd && cmd.oflags & O_EXCL) {
 806                         ret = -EINVAL;
 807                         goto err_tree_mutex_unlock;
 808                 }
 809         }
 810
 811         obj = kmalloc(sizeof *obj, GFP_KERNEL);
 812         if (!obj) {
 813                 ret = -ENOMEM;
 814                 goto err_tree_mutex_unlock;
 815         }
 816
 817         init_uobj(&obj->uobject, 0, file->ucontext, &xrcd_lock_class);
 818
 819         down_write(&obj->uobject.mutex);
 820
 821         if (!xrcd) {
 822                 xrcd = ib_dev->alloc_xrcd(ib_dev, file->ucontext, &udata);
 823                 if (IS_ERR(xrcd)) {
 824                         ret = PTR_ERR(xrcd);
 825                         goto err;
 826                 }
 827
 828                 xrcd->inode   = inode;
 829                 xrcd->device  = ib_dev;
 830                 atomic_set(&xrcd->usecnt, 0);
 831                 mutex_init(&xrcd->tgt_qp_mutex);
 832                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 833                 new_xrcd = 1;
 834         }
 835
 836         atomic_set(&obj->refcnt, 0);
 837         obj->uobject.object = xrcd;
 838         ret = idr_add_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 839         if (ret)
 840                 goto err_idr;
 841
 842         memset(&resp, 0, sizeof resp);
 843         resp.xrcd_handle = obj->uobject.id;
 844
 845         if (inode) {
 846                 if (new_xrcd) {
 847                         /* create new inode/xrcd table entry */
 848                         ret = xrcd_table_insert(file->device, inode, xrcd);
 849                         if (ret)
 850                                 goto err_insert_xrcd;
 851                 }
 852                 atomic_inc(&xrcd->usecnt);
 853         }
 854
 855         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 856                          &resp, sizeof resp)) {
 857                 ret = -EFAULT;
 858                 goto err_copy;
 859         }
 860
 861         if (f.file)
 862                 fdput(f);
 863
 864         mutex_lock(&file->mutex);
 865         list_add_tail(&obj->uobject.list, &file->ucontext->xrcd_list);
 866         mutex_unlock(&file->mutex);
 867
 868         obj->uobject.live = 1;
 869         up_write(&obj->uobject.mutex);
 870
 871         mutex_unlock(&file->device->xrcd_tree_mutex);
 872         return in_len;
 873
 874 err_copy:
 875         if (inode) {
 876                 if (new_xrcd)
 877                         xrcd_table_delete(file->device, inode);
 878                 atomic_dec(&xrcd->usecnt);
 879         }
 880
 881 err_insert_xrcd:
 882         idr_remove_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 883
 884 err_idr:
 885         ib_dealloc_xrcd(xrcd);
 886
 887 err:
 888         put_uobj_write(&obj->uobject);
 889
 890 err_tree_mutex_unlock:
 891         if (f.file)
 892                 fdput(f);
 893
 894         mutex_unlock(&file->device->xrcd_tree_mutex);
 895
 896         return ret;
 897 }
 898
 899 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 900                              struct ib_device *ib_dev,
 901                              const char __user *buf, int in_len,
 902                              int out_len)
 903 {
 904         struct ib_uverbs_close_xrcd cmd;
 905         struct ib_uobject           *uobj;
 906         struct ib_xrcd              *xrcd = NULL;
 907         struct inode                *inode = NULL;
 908         struct ib_uxrcd_object      *obj;
 909         int                         live;
 910         int                         ret = 0;
 911
 912         if (copy_from_user(&cmd, buf, sizeof cmd))
 913                 return -EFAULT;
 914
 915         mutex_lock(&file->device->xrcd_tree_mutex);
 916         uobj = idr_write_uobj(&ib_uverbs_xrcd_idr, cmd.xrcd_handle, file->ucontext);
 917         if (!uobj) {
 918                 ret = -EINVAL;
 919                 goto out;
 920         }
 921
 922         xrcd  = uobj->object;
 923         inode = xrcd->inode;
 924         obj   = container_of(uobj, struct ib_uxrcd_object, uobject);
 925         if (atomic_read(&obj->refcnt)) {
 926                 put_uobj_write(uobj);
 927                 ret = -EBUSY;
 928                 goto out;
 929         }
 930
 931         if (!inode || atomic_dec_and_test(&xrcd->usecnt)) {
 932                 ret = ib_dealloc_xrcd(uobj->object);
 933                 if (!ret)
 934                         uobj->live = 0;
 935         }
 936
 937         live = uobj->live;
 938         if (inode && ret)
 939                 atomic_inc(&xrcd->usecnt);
 940
 941         put_uobj_write(uobj);
 942
 943         if (ret)
 944                 goto out;
 945
 946         if (inode && !live)
 947                 xrcd_table_delete(file->device, inode);
 948
 949         idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 950         mutex_lock(&file->mutex);
 951         list_del(&uobj->list);
 952         mutex_unlock(&file->mutex);
 953
 954         put_uobj(uobj);
 955         ret = in_len;
 956
 957 out:
 958         mutex_unlock(&file->device->xrcd_tree_mutex);
 959         return ret;
 960 }
 961
 962 void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 963                             struct ib_xrcd *xrcd)
 964 {
 965         struct inode *inode;
 966
 967         inode = xrcd->inode;
 968         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 969                 return;
 970
 971         ib_dealloc_xrcd(xrcd);
 972
 973         if (inode)
 974                 xrcd_table_delete(dev, inode);
 975 }
 976
 977 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 978                          struct ib_device *ib_dev,
 979                          const char __user *buf, int in_len,
 980                          int out_len)
 981 {
 982         struct ib_uverbs_reg_mr      cmd;
 983         struct ib_uverbs_reg_mr_resp resp;
 984         struct ib_udata              udata;
 985         struct ib_uobject           *uobj;
 986         struct ib_pd                *pd;
 987         struct ib_mr                *mr;
 988         int                          ret;
 989
 990         if (out_len < sizeof resp)
 991                 return -ENOSPC;
 992
 993         if (copy_from_user(&cmd, buf, sizeof cmd))
 994                 return -EFAULT;
 995
 996         INIT_UDATA(&udata, buf + sizeof cmd,
 997                    (unsigned long) cmd.response + sizeof resp,
 998                    in_len - sizeof cmd, out_len - sizeof resp);
 999
1000         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
1001                 return -EINVAL;
1002
1003         ret = ib_check_mr_access(cmd.access_flags);
1004         if (ret)
1005                 return ret;
1006
1007         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
1008         if (!uobj)
1009                 return -ENOMEM;
1010
1011         init_uobj(uobj, 0, file->ucontext, &mr_lock_class);
1012         down_write(&uobj->mutex);
1013
1014         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
1015         if (!pd) {
1016                 ret = -EINVAL;
1017                 goto err_free;
1018         }
1019
1020         if (cmd.access_flags & IB_ACCESS_ON_DEMAND) {
1021                 if (!(pd->device->attrs.device_cap_flags &
1022                       IB_DEVICE_ON_DEMAND_PAGING)) {
1023                         pr_debug("ODP support not available\n");
1024                         ret = -EINVAL;
1025                         goto err_put;
1026                 }
1027         }
1028         ret = ib_rdmacg_try_charge(&uobj->cg_obj, ib_dev,
1029                                    RDMACG_RESOURCE_HCA_OBJECT);
1030         if (ret)
1031                 goto err_charge;
1032
1033         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
1034                                      cmd.access_flags, &udata);
1035         if (IS_ERR(mr)) {
1036                 ret = PTR_ERR(mr);
1037                 goto err_put;
1038         }
1039
1040         mr->device  = pd->device;
1041         mr->pd      = pd;
1042         mr->uobject = uobj;
1043         atomic_inc(&pd->usecnt);
1044
1045         uobj->object = mr;
1046         ret = idr_add_uobj(&ib_uverbs_mr_idr, uobj);
1047         if (ret)
1048                 goto err_unreg;
1049
1050         memset(&resp, 0, sizeof resp);
1051         resp.lkey      = mr->lkey;
1052         resp.rkey      = mr->rkey;
1053         resp.mr_handle = uobj->id;
1054
1055         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1056                          &resp, sizeof resp)) {
1057                 ret = -EFAULT;
1058                 goto err_copy;
1059         }
1060
1061         put_pd_read(pd);
1062
1063         mutex_lock(&file->mutex);
1064         list_add_tail(&uobj->list, &file->ucontext->mr_list);
1065         mutex_unlock(&file->mutex);
1066
1067         uobj->live = 1;
1068
1069         up_write(&uobj->mutex);
1070
1071         return in_len;
1072
1073 err_copy:
1074         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1075
1076 err_unreg:
1077         ib_dereg_mr(mr);
1078
1079 err_put:
1080         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
1081
1082 err_charge:
1083         put_pd_read(pd);
1084
1085 err_free:
1086         put_uobj_write(uobj);
1087         return ret;
1088 }
1089
1090 ssize_t ib_uverbs_rereg_mr(struct ib_uverbs_file *file,
1091                            struct ib_device *ib_dev,
1092                            const char __user *buf, int in_len,
1093                            int out_len)
1094 {
1095         struct ib_uverbs_rereg_mr      cmd;
1096         struct ib_uverbs_rereg_mr_resp resp;
1097         struct ib_udata              udata;
1098         struct ib_pd                *pd = NULL;
1099         struct ib_mr                *mr;
1100         struct ib_pd                *old_pd;
1101         int                          ret;
1102         struct ib_uobject           *uobj;
1103
1104         if (out_len < sizeof(resp))
1105                 return -ENOSPC;
1106
1107         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1108                 return -EFAULT;
1109
1110         INIT_UDATA(&udata, buf + sizeof(cmd),
1111                    (unsigned long) cmd.response + sizeof(resp),
1112                    in_len - sizeof(cmd), out_len - sizeof(resp));
1113
1114         if (cmd.flags & ~IB_MR_REREG_SUPPORTED || !cmd.flags)
1115                 return -EINVAL;
1116
1117         if ((cmd.flags & IB_MR_REREG_TRANS) &&
1118             (!cmd.start || !cmd.hca_va || 0 >= cmd.length ||
1119              (cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK)))
1120                         return -EINVAL;
1121
1122         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle,
1123                               file->ucontext);
1124
1125         if (!uobj)
1126                 return -EINVAL;
1127
1128         mr = uobj->object;
1129
1130         if (cmd.flags & IB_MR_REREG_ACCESS) {
1131                 ret = ib_check_mr_access(cmd.access_flags);
1132                 if (ret)
1133                         goto put_uobjs;
1134         }
1135
1136         if (cmd.flags & IB_MR_REREG_PD) {
1137                 pd = idr_read_pd(cmd.pd_handle, file->ucontext);
1138                 if (!pd) {
1139                         ret = -EINVAL;
1140                         goto put_uobjs;
1141                 }
1142         }
1143
1144         old_pd = mr->pd;
1145         ret = mr->device->rereg_user_mr(mr, cmd.flags, cmd.start,
1146                                         cmd.length, cmd.hca_va,
1147                                         cmd.access_flags, pd, &udata);
1148         if (!ret) {
1149                 if (cmd.flags & IB_MR_REREG_PD) {
1150                         atomic_inc(&pd->usecnt);
1151                         mr->pd = pd;
1152                         atomic_dec(&old_pd->usecnt);
1153                 }
1154         } else {
1155                 goto put_uobj_pd;
1156         }
1157
1158         memset(&resp, 0, sizeof(resp));
1159         resp.lkey      = mr->lkey;
1160         resp.rkey      = mr->rkey;
1161
1162         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1163                          &resp, sizeof(resp)))
1164                 ret = -EFAULT;
1165         else
1166                 ret = in_len;
1167
1168 put_uobj_pd:
1169         if (cmd.flags & IB_MR_REREG_PD)
1170                 put_pd_read(pd);
1171
1172 put_uobjs:
1173
1174         put_uobj_write(mr->uobject);
1175
1176         return ret;
1177 }
1178
1179 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
1180                            struct ib_device *ib_dev,
1181                            const char __user *buf, int in_len,
1182                            int out_len)
1183 {
1184         struct ib_uverbs_dereg_mr cmd;
1185         struct ib_mr             *mr;
1186         struct ib_uobject        *uobj;
1187         int                       ret = -EINVAL;
1188
1189         if (copy_from_user(&cmd, buf, sizeof cmd))
1190                 return -EFAULT;
1191
1192         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle, file->ucontext);
1193         if (!uobj)
1194                 return -EINVAL;
1195
1196         mr = uobj->object;
1197
1198         ret = ib_dereg_mr(mr);
1199         if (!ret)
1200                 uobj->live = 0;
1201
1202         put_uobj_write(uobj);
1203
1204         if (ret)
1205                 return ret;
1206
1207         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
1208
1209         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1210
1211         mutex_lock(&file->mutex);
1212         list_del(&uobj->list);
1213         mutex_unlock(&file->mutex);
1214
1215         put_uobj(uobj);
1216
1217         return in_len;
1218 }
1219
1220 ssize_t ib_uverbs_alloc_mw(struct ib_uverbs_file *file,
1221                            struct ib_device *ib_dev,
1222                            const char __user *buf, int in_len,
1223                            int out_len)
1224 {
1225         struct ib_uverbs_alloc_mw      cmd;
1226         struct ib_uverbs_alloc_mw_resp resp;
1227         struct ib_uobject             *uobj;
1228         struct ib_pd                  *pd;
1229         struct ib_mw                  *mw;
1230         struct ib_udata                udata;
1231         int                            ret;
1232
1233         if (out_len < sizeof(resp))
1234                 return -ENOSPC;
1235
1236         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1237                 return -EFAULT;
1238
1239         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
1240         if (!uobj)
1241                 return -ENOMEM;
1242
1243         init_uobj(uobj, 0, file->ucontext, &mw_lock_class);
1244         down_write(&uobj->mutex);
1245
1246         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
1247         if (!pd) {
1248                 ret = -EINVAL;
1249                 goto err_free;
1250         }
1251
1252         INIT_UDATA(&udata, buf + sizeof(cmd),
1253                    (unsigned long)cmd.response + sizeof(resp),
1254                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1255                    out_len - sizeof(resp));
1256
1257         ret = ib_rdmacg_try_charge(&uobj->cg_obj, ib_dev,
1258                                    RDMACG_RESOURCE_HCA_OBJECT);
1259         if (ret)
1260                 goto err_charge;
1261
1262         mw = pd->device->alloc_mw(pd, cmd.mw_type, &udata);
1263         if (IS_ERR(mw)) {
1264                 ret = PTR_ERR(mw);
1265                 goto err_put;
1266         }
1267
1268         mw->device  = pd->device;
1269         mw->pd      = pd;
1270         mw->uobject = uobj;
1271         atomic_inc(&pd->usecnt);
1272
1273         uobj->object = mw;
1274         ret = idr_add_uobj(&ib_uverbs_mw_idr, uobj);
1275         if (ret)
1276                 goto err_unalloc;
1277
1278         memset(&resp, 0, sizeof(resp));
1279         resp.rkey      = mw->rkey;
1280         resp.mw_handle = uobj->id;
1281
1282         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1283                          &resp, sizeof(resp))) {
1284                 ret = -EFAULT;
1285                 goto err_copy;
1286         }
1287
1288         put_pd_read(pd);
1289
1290         mutex_lock(&file->mutex);
1291         list_add_tail(&uobj->list, &file->ucontext->mw_list);
1292         mutex_unlock(&file->mutex);
1293
1294         uobj->live = 1;
1295
1296         up_write(&uobj->mutex);
1297
1298         return in_len;
1299
1300 err_copy:
1301         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1302
1303 err_unalloc:
1304         uverbs_dealloc_mw(mw);
1305
1306 err_put:
1307         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
1308
1309 err_charge:
1310         put_pd_read(pd);
1311
1312 err_free:
1313         put_uobj_write(uobj);
1314         return ret;
1315 }
1316
1317 ssize_t ib_uverbs_dealloc_mw(struct ib_uverbs_file *file,
1318                              struct ib_device *ib_dev,
1319                              const char __user *buf, int in_len,
1320                              int out_len)
1321 {
1322         struct ib_uverbs_dealloc_mw cmd;
1323         struct ib_mw               *mw;
1324         struct ib_uobject          *uobj;
1325         int                         ret = -EINVAL;
1326
1327         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1328                 return -EFAULT;
1329
1330         uobj = idr_write_uobj(&ib_uverbs_mw_idr, cmd.mw_handle, file->ucontext);
1331         if (!uobj)
1332                 return -EINVAL;
1333
1334         mw = uobj->object;
1335
1336         ret = uverbs_dealloc_mw(mw);
1337         if (!ret)
1338                 uobj->live = 0;
1339
1340         put_uobj_write(uobj);
1341
1342         if (ret)
1343                 return ret;
1344
1345         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
1346
1347         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1348
1349         mutex_lock(&file->mutex);
1350         list_del(&uobj->list);
1351         mutex_unlock(&file->mutex);
1352
1353         put_uobj(uobj);
1354
1355         return in_len;
1356 }
1357
1358 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
1359                                       struct ib_device *ib_dev,
1360                                       const char __user *buf, int in_len,
1361                                       int out_len)
1362 {
1363         struct ib_uverbs_create_comp_channel       cmd;
1364         struct ib_uverbs_create_comp_channel_resp  resp;
1365         struct file                               *filp;
1366         int ret;
1367
1368         if (out_len < sizeof resp)
1369                 return -ENOSPC;
1370
1371         if (copy_from_user(&cmd, buf, sizeof cmd))
1372                 return -EFAULT;
1373
1374         ret = get_unused_fd_flags(O_CLOEXEC);
1375         if (ret < 0)
1376                 return ret;
1377         resp.fd = ret;
1378
1379         filp = ib_uverbs_alloc_event_file(file, ib_dev, 0);
1380         if (IS_ERR(filp)) {
1381                 put_unused_fd(resp.fd);
1382                 return PTR_ERR(filp);
1383         }
1384
1385         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1386                          &resp, sizeof resp)) {
1387                 put_unused_fd(resp.fd);
1388                 fput(filp);
1389                 return -EFAULT;
1390         }
1391
1392         fd_install(resp.fd, filp);
1393         return in_len;
1394 }
1395
1396 static struct ib_ucq_object *create_cq(struct ib_uverbs_file *file,
1397                                         struct ib_device *ib_dev,
1398                                        struct ib_udata *ucore,
1399                                        struct ib_udata *uhw,
1400                                        struct ib_uverbs_ex_create_cq *cmd,
1401                                        size_t cmd_sz,
1402                                        int (*cb)(struct ib_uverbs_file *file,
1403                                                  struct ib_ucq_object *obj,
1404                                                  struct ib_uverbs_ex_create_cq_resp *resp,
1405                                                  struct ib_udata *udata,
1406                                                  void *context),
1407                                        void *context)
1408 {
1409         struct ib_ucq_object           *obj;
1410         struct ib_uverbs_event_file    *ev_file = NULL;
1411         struct ib_cq                   *cq;
1412         int                             ret;
1413         struct ib_uverbs_ex_create_cq_resp resp;
1414         struct ib_cq_init_attr attr = {};
1415
1416         if (cmd->comp_vector >= file->device->num_comp_vectors)
1417                 return ERR_PTR(-EINVAL);
1418
1419         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1420         if (!obj)
1421                 return ERR_PTR(-ENOMEM);
1422
1423         init_uobj(&obj->uobject, cmd->user_handle, file->ucontext, &cq_lock_class);
1424         down_write(&obj->uobject.mutex);
1425
1426         if (cmd->comp_channel >= 0) {
1427                 ev_file = ib_uverbs_lookup_comp_file(cmd->comp_channel);
1428                 if (!ev_file) {
1429                         ret = -EINVAL;
1430                         goto err;
1431                 }
1432         }
1433
1434         obj->uverbs_file           = file;
1435         obj->comp_events_reported  = 0;
1436         obj->async_events_reported = 0;
1437         INIT_LIST_HEAD(&obj->comp_list);
1438         INIT_LIST_HEAD(&obj->async_list);
1439
1440         attr.cqe = cmd->cqe;
1441         attr.comp_vector = cmd->comp_vector;
1442
1443         if (cmd_sz > offsetof(typeof(*cmd), flags) + sizeof(cmd->flags))
1444                 attr.flags = cmd->flags;
1445
1446         ret = ib_rdmacg_try_charge(&obj->uobject.cg_obj, ib_dev,
1447                                    RDMACG_RESOURCE_HCA_OBJECT);
1448         if (ret)
1449                 goto err_charge;
1450
1451         cq = ib_dev->create_cq(ib_dev, &attr,
1452                                              file->ucontext, uhw);
1453         if (IS_ERR(cq)) {
1454                 ret = PTR_ERR(cq);
1455                 goto err_file;
1456         }
1457
1458         cq->device        = ib_dev;
1459         cq->uobject       = &obj->uobject;
1460         cq->comp_handler  = ib_uverbs_comp_handler;
1461         cq->event_handler = ib_uverbs_cq_event_handler;
1462         cq->cq_context    = ev_file;
1463         atomic_set(&cq->usecnt, 0);
1464
1465         obj->uobject.object = cq;
1466         ret = idr_add_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1467         if (ret)
1468                 goto err_free;
1469
1470         memset(&resp, 0, sizeof resp);
1471         resp.base.cq_handle = obj->uobject.id;
1472         resp.base.cqe       = cq->cqe;
1473
1474         resp.response_length = offsetof(typeof(resp), response_length) +
1475                 sizeof(resp.response_length);
1476
1477         ret = cb(file, obj, &resp, ucore, context);
1478         if (ret)
1479                 goto err_cb;
1480
1481         mutex_lock(&file->mutex);
1482         list_add_tail(&obj->uobject.list, &file->ucontext->cq_list);
1483         mutex_unlock(&file->mutex);
1484
1485         obj->uobject.live = 1;
1486
1487         up_write(&obj->uobject.mutex);
1488
1489         return obj;
1490
1491 err_cb:
1492         idr_remove_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1493
1494 err_free:
1495         ib_destroy_cq(cq);
1496
1497 err_file:
1498         ib_rdmacg_uncharge(&obj->uobject.cg_obj, ib_dev,
1499                            RDMACG_RESOURCE_HCA_OBJECT);
1500
1501 err_charge:
1502         if (ev_file)
1503                 ib_uverbs_release_ucq(file, ev_file, obj);
1504
1505 err:
1506         put_uobj_write(&obj->uobject);
1507
1508         return ERR_PTR(ret);
1509 }
1510
1511 static int ib_uverbs_create_cq_cb(struct ib_uverbs_file *file,
1512                                   struct ib_ucq_object *obj,
1513                                   struct ib_uverbs_ex_create_cq_resp *resp,
1514                                   struct ib_udata *ucore, void *context)
1515 {
1516         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1517                 return -EFAULT;
1518
1519         return 0;
1520 }
1521
1522 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1523                             struct ib_device *ib_dev,
1524                             const char __user *buf, int in_len,
1525                             int out_len)
1526 {
1527         struct ib_uverbs_create_cq      cmd;
1528         struct ib_uverbs_ex_create_cq   cmd_ex;
1529         struct ib_uverbs_create_cq_resp resp;
1530         struct ib_udata                 ucore;
1531         struct ib_udata                 uhw;
1532         struct ib_ucq_object           *obj;
1533
1534         if (out_len < sizeof(resp))
1535                 return -ENOSPC;
1536
1537         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1538                 return -EFAULT;
1539
1540         INIT_UDATA(&ucore, buf, (unsigned long)cmd.response, sizeof(cmd), sizeof(resp));
1541
1542         INIT_UDATA(&uhw, buf + sizeof(cmd),
1543                    (unsigned long)cmd.response + sizeof(resp),
1544                    in_len - sizeof(cmd), out_len - sizeof(resp));
1545
1546         memset(&cmd_ex, 0, sizeof(cmd_ex));
1547         cmd_ex.user_handle = cmd.user_handle;
1548         cmd_ex.cqe = cmd.cqe;
1549         cmd_ex.comp_vector = cmd.comp_vector;
1550         cmd_ex.comp_channel = cmd.comp_channel;
1551
1552         obj = create_cq(file, ib_dev, &ucore, &uhw, &cmd_ex,
1553                         offsetof(typeof(cmd_ex), comp_channel) +
1554                         sizeof(cmd.comp_channel), ib_uverbs_create_cq_cb,
1555                         NULL);
1556
1557         if (IS_ERR(obj))
1558                 return PTR_ERR(obj);
1559
1560         return in_len;
1561 }
1562
1563 static int ib_uverbs_ex_create_cq_cb(struct ib_uverbs_file *file,
1564                                      struct ib_ucq_object *obj,
1565                                      struct ib_uverbs_ex_create_cq_resp *resp,
1566                                      struct ib_udata *ucore, void *context)
1567 {
1568         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1569                 return -EFAULT;
1570
1571         return 0;
1572 }
1573
1574 int ib_uverbs_ex_create_cq(struct ib_uverbs_file *file,
1575                          struct ib_device *ib_dev,
1576                            struct ib_udata *ucore,
1577                            struct ib_udata *uhw)
1578 {
1579         struct ib_uverbs_ex_create_cq_resp resp;
1580         struct ib_uverbs_ex_create_cq  cmd;
1581         struct ib_ucq_object           *obj;
1582         int err;
1583
1584         if (ucore->inlen < sizeof(cmd))
1585                 return -EINVAL;
1586
1587         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
1588         if (err)
1589                 return err;
1590
1591         if (cmd.comp_mask)
1592                 return -EINVAL;
1593
1594         if (cmd.reserved)
1595                 return -EINVAL;
1596
1597         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1598                              sizeof(resp.response_length)))
1599                 return -ENOSPC;
1600
1601         obj = create_cq(file, ib_dev, ucore, uhw, &cmd,
1602                         min(ucore->inlen, sizeof(cmd)),
1603                         ib_uverbs_ex_create_cq_cb, NULL);
1604
1605         if (IS_ERR(obj))
1606                 return PTR_ERR(obj);
1607
1608         return 0;
1609 }
1610
1611 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1612                             struct ib_device *ib_dev,
1613                             const char __user *buf, int in_len,
1614                             int out_len)
1615 {
1616         struct ib_uverbs_resize_cq      cmd;
1617         struct ib_uverbs_resize_cq_resp resp;
1618         struct ib_udata                 udata;
1619         struct ib_cq                    *cq;
1620         int                             ret = -EINVAL;
1621
1622         if (copy_from_user(&cmd, buf, sizeof cmd))
1623                 return -EFAULT;
1624
1625         INIT_UDATA(&udata, buf + sizeof cmd,
1626                    (unsigned long) cmd.response + sizeof resp,
1627                    in_len - sizeof cmd, out_len - sizeof resp);
1628
1629         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1630         if (!cq)
1631                 return -EINVAL;
1632
1633         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1634         if (ret)
1635                 goto out;
1636
1637         resp.cqe = cq->cqe;
1638
1639         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1640                          &resp, sizeof resp.cqe))
1641                 ret = -EFAULT;
1642
1643 out:
1644         put_cq_read(cq);
1645
1646         return ret ? ret : in_len;
1647 }
1648
1649 static int copy_wc_to_user(void __user *dest, struct ib_wc *wc)
1650 {
1651         struct ib_uverbs_wc tmp;
1652
1653         tmp.wr_id               = wc->wr_id;
1654         tmp.status              = wc->status;
1655         tmp.opcode              = wc->opcode;
1656         tmp.vendor_err          = wc->vendor_err;
1657         tmp.byte_len            = wc->byte_len;
1658         tmp.ex.imm_data         = (__u32 __force) wc->ex.imm_data;
1659         tmp.qp_num              = wc->qp->qp_num;
1660         tmp.src_qp              = wc->src_qp;
1661         tmp.wc_flags            = wc->wc_flags;
1662         tmp.pkey_index          = wc->pkey_index;
1663         tmp.slid                = wc->slid;
1664         tmp.sl                  = wc->sl;
1665         tmp.dlid_path_bits      = wc->dlid_path_bits;
1666         tmp.port_num            = wc->port_num;
1667         tmp.reserved            = 0;
1668
1669         if (copy_to_user(dest, &tmp, sizeof tmp))
1670                 return -EFAULT;
1671
1672         return 0;
1673 }
1674
1675 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1676                           struct ib_device *ib_dev,
1677                           const char __user *buf, int in_len,
1678                           int out_len)
1679 {
1680         struct ib_uverbs_poll_cq       cmd;
1681         struct ib_uverbs_poll_cq_resp  resp;
1682         u8 __user                     *header_ptr;
1683         u8 __user                     *data_ptr;
1684         struct ib_cq                  *cq;
1685         struct ib_wc                   wc;
1686         int                            ret;
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         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1696         header_ptr = (void __user *)(unsigned long) cmd.response;
1697         data_ptr = header_ptr + sizeof resp;
1698
1699         memset(&resp, 0, sizeof resp);
1700         while (resp.count < cmd.ne) {
1701                 ret = ib_poll_cq(cq, 1, &wc);
1702                 if (ret < 0)
1703                         goto out_put;
1704                 if (!ret)
1705                         break;
1706
1707                 ret = copy_wc_to_user(data_ptr, &wc);
1708                 if (ret)
1709                         goto out_put;
1710
1711                 data_ptr += sizeof(struct ib_uverbs_wc);
1712                 ++resp.count;
1713         }
1714
1715         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1716                 ret = -EFAULT;
1717                 goto out_put;
1718         }
1719
1720         ret = in_len;
1721
1722 out_put:
1723         put_cq_read(cq);
1724         return ret;
1725 }
1726
1727 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1728                                 struct ib_device *ib_dev,
1729                                 const char __user *buf, int in_len,
1730                                 int out_len)
1731 {
1732         struct ib_uverbs_req_notify_cq cmd;
1733         struct ib_cq                  *cq;
1734
1735         if (copy_from_user(&cmd, buf, sizeof cmd))
1736                 return -EFAULT;
1737
1738         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1739         if (!cq)
1740                 return -EINVAL;
1741
1742         ib_req_notify_cq(cq, cmd.solicited_only ?
1743                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1744
1745         put_cq_read(cq);
1746
1747         return in_len;
1748 }
1749
1750 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1751                              struct ib_device *ib_dev,
1752                              const char __user *buf, int in_len,
1753                              int out_len)
1754 {
1755         struct ib_uverbs_destroy_cq      cmd;
1756         struct ib_uverbs_destroy_cq_resp resp;
1757         struct ib_uobject               *uobj;
1758         struct ib_cq                    *cq;
1759         struct ib_ucq_object            *obj;
1760         struct ib_uverbs_event_file     *ev_file;
1761         int                              ret = -EINVAL;
1762
1763         if (copy_from_user(&cmd, buf, sizeof cmd))
1764                 return -EFAULT;
1765
1766         uobj = idr_write_uobj(&ib_uverbs_cq_idr, cmd.cq_handle, file->ucontext);
1767         if (!uobj)
1768                 return -EINVAL;
1769         cq      = uobj->object;
1770         ev_file = cq->cq_context;
1771         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1772
1773         ret = ib_destroy_cq(cq);
1774         if (!ret)
1775                 uobj->live = 0;
1776
1777         put_uobj_write(uobj);
1778
1779         if (ret)
1780                 return ret;
1781
1782         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
1783
1784         idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
1785
1786         mutex_lock(&file->mutex);
1787         list_del(&uobj->list);
1788         mutex_unlock(&file->mutex);
1789
1790         ib_uverbs_release_ucq(file, ev_file, obj);
1791
1792         memset(&resp, 0, sizeof resp);
1793         resp.comp_events_reported  = obj->comp_events_reported;
1794         resp.async_events_reported = obj->async_events_reported;
1795
1796         put_uobj(uobj);
1797
1798         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1799                          &resp, sizeof resp))
1800                 return -EFAULT;
1801
1802         return in_len;
1803 }
1804
1805 static int create_qp(struct ib_uverbs_file *file,
1806                      struct ib_udata *ucore,
1807                      struct ib_udata *uhw,
1808                      struct ib_uverbs_ex_create_qp *cmd,
1809                      size_t cmd_sz,
1810                      int (*cb)(struct ib_uverbs_file *file,
1811                                struct ib_uverbs_ex_create_qp_resp *resp,
1812                                struct ib_udata *udata),
1813                      void *context)
1814 {
1815         struct ib_uqp_object            *obj;
1816         struct ib_device                *device;
1817         struct ib_pd                    *pd = NULL;
1818         struct ib_xrcd                  *xrcd = NULL;
1819         struct ib_uobject               *uninitialized_var(xrcd_uobj);
1820         struct ib_cq                    *scq = NULL, *rcq = NULL;
1821         struct ib_srq                   *srq = NULL;
1822         struct ib_qp                    *qp;
1823         char                            *buf;
1824         struct ib_qp_init_attr          attr = {};
1825         struct ib_uverbs_ex_create_qp_resp resp;
1826         int                             ret;
1827         struct ib_rwq_ind_table *ind_tbl = NULL;
1828         bool has_sq = true;
1829
1830         if (cmd->qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1831                 return -EPERM;
1832
1833         obj = kzalloc(sizeof *obj, GFP_KERNEL);
1834         if (!obj)
1835                 return -ENOMEM;
1836
1837         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext,
1838                   &qp_lock_class);
1839         down_write(&obj->uevent.uobject.mutex);
1840         if (cmd_sz >= offsetof(typeof(*cmd), rwq_ind_tbl_handle) +
1841                       sizeof(cmd->rwq_ind_tbl_handle) &&
1842                       (cmd->comp_mask & IB_UVERBS_CREATE_QP_MASK_IND_TABLE)) {
1843                 ind_tbl = idr_read_rwq_indirection_table(cmd->rwq_ind_tbl_handle,
1844                                                          file->ucontext);
1845                 if (!ind_tbl) {
1846                         ret = -EINVAL;
1847                         goto err_put;
1848                 }
1849
1850                 attr.rwq_ind_tbl = ind_tbl;
1851         }
1852
1853         if ((cmd_sz >= offsetof(typeof(*cmd), reserved1) +
1854                        sizeof(cmd->reserved1)) && cmd->reserved1) {
1855                 ret = -EOPNOTSUPP;
1856                 goto err_put;
1857         }
1858
1859         if (ind_tbl && (cmd->max_recv_wr || cmd->max_recv_sge || cmd->is_srq)) {
1860                 ret = -EINVAL;
1861                 goto err_put;
1862         }
1863
1864         if (ind_tbl && !cmd->max_send_wr)
1865                 has_sq = false;
1866
1867         if (cmd->qp_type == IB_QPT_XRC_TGT) {
1868                 xrcd = idr_read_xrcd(cmd->pd_handle, file->ucontext,
1869                                      &xrcd_uobj);
1870                 if (!xrcd) {
1871                         ret = -EINVAL;
1872                         goto err_put;
1873                 }
1874                 device = xrcd->device;
1875         } else {
1876                 if (cmd->qp_type == IB_QPT_XRC_INI) {
1877                         cmd->max_recv_wr = 0;
1878                         cmd->max_recv_sge = 0;
1879                 } else {
1880                         if (cmd->is_srq) {
1881                                 srq = idr_read_srq(cmd->srq_handle,
1882                                                    file->ucontext);
1883                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1884                                         ret = -EINVAL;
1885                                         goto err_put;
1886                                 }
1887                         }
1888
1889                         if (!ind_tbl) {
1890                                 if (cmd->recv_cq_handle != cmd->send_cq_handle) {
1891                                         rcq = idr_read_cq(cmd->recv_cq_handle,
1892                                                           file->ucontext, 0);
1893                                         if (!rcq) {
1894                                                 ret = -EINVAL;
1895                                                 goto err_put;
1896                                         }
1897                                 }
1898                         }
1899                 }
1900
1901                 if (has_sq)
1902                         scq = idr_read_cq(cmd->send_cq_handle, file->ucontext, !!rcq);
1903                 if (!ind_tbl)
1904                         rcq = rcq ?: scq;
1905                 pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
1906                 if (!pd || (!scq && has_sq)) {
1907                         ret = -EINVAL;
1908                         goto err_put;
1909                 }
1910
1911                 device = pd->device;
1912         }
1913
1914         attr.event_handler = ib_uverbs_qp_event_handler;
1915         attr.qp_context    = file;
1916         attr.send_cq       = scq;
1917         attr.recv_cq       = rcq;
1918         attr.srq           = srq;
1919         attr.xrcd          = xrcd;
1920         attr.sq_sig_type   = cmd->sq_sig_all ? IB_SIGNAL_ALL_WR :
1921                                               IB_SIGNAL_REQ_WR;
1922         attr.qp_type       = cmd->qp_type;
1923         attr.create_flags  = 0;
1924
1925         attr.cap.max_send_wr     = cmd->max_send_wr;
1926         attr.cap.max_recv_wr     = cmd->max_recv_wr;
1927         attr.cap.max_send_sge    = cmd->max_send_sge;
1928         attr.cap.max_recv_sge    = cmd->max_recv_sge;
1929         attr.cap.max_inline_data = cmd->max_inline_data;
1930
1931         obj->uevent.events_reported     = 0;
1932         INIT_LIST_HEAD(&obj->uevent.event_list);
1933         INIT_LIST_HEAD(&obj->mcast_list);
1934
1935         if (cmd_sz >= offsetof(typeof(*cmd), create_flags) +
1936                       sizeof(cmd->create_flags))
1937                 attr.create_flags = cmd->create_flags;
1938
1939         if (attr.create_flags & ~(IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
1940                                 IB_QP_CREATE_CROSS_CHANNEL |
1941                                 IB_QP_CREATE_MANAGED_SEND |
1942                                 IB_QP_CREATE_MANAGED_RECV |
1943                                 IB_QP_CREATE_SCATTER_FCS |
1944                                 IB_QP_CREATE_CVLAN_STRIPPING)) {
1945                 ret = -EINVAL;
1946                 goto err_put;
1947         }
1948
1949         buf = (void *)cmd + sizeof(*cmd);
1950         if (cmd_sz > sizeof(*cmd))
1951                 if (!(buf[0] == 0 && !memcmp(buf, buf + 1,
1952                                              cmd_sz - sizeof(*cmd) - 1))) {
1953                         ret = -EINVAL;
1954                         goto err_put;
1955                 }
1956
1957         ret = ib_rdmacg_try_charge(&obj->uevent.uobject.cg_obj, device,
1958                                    RDMACG_RESOURCE_HCA_OBJECT);
1959         if (ret)
1960                 goto err_put;
1961
1962         if (cmd->qp_type == IB_QPT_XRC_TGT)
1963                 qp = ib_create_qp(pd, &attr);
1964         else
1965                 qp = device->create_qp(pd, &attr, uhw);
1966
1967         if (IS_ERR(qp)) {
1968                 ret = PTR_ERR(qp);
1969                 goto err_create;
1970         }
1971
1972         if (cmd->qp_type != IB_QPT_XRC_TGT) {
1973                 qp->real_qp       = qp;
1974                 qp->device        = device;
1975                 qp->pd            = pd;
1976                 qp->send_cq       = attr.send_cq;
1977                 qp->recv_cq       = attr.recv_cq;
1978                 qp->srq           = attr.srq;
1979                 qp->rwq_ind_tbl   = ind_tbl;
1980                 qp->event_handler = attr.event_handler;
1981                 qp->qp_context    = attr.qp_context;
1982                 qp->qp_type       = attr.qp_type;
1983                 atomic_set(&qp->usecnt, 0);
1984                 atomic_inc(&pd->usecnt);
1985                 if (attr.send_cq)
1986                         atomic_inc(&attr.send_cq->usecnt);
1987                 if (attr.recv_cq)
1988                         atomic_inc(&attr.recv_cq->usecnt);
1989                 if (attr.srq)
1990                         atomic_inc(&attr.srq->usecnt);
1991                 if (ind_tbl)
1992                         atomic_inc(&ind_tbl->usecnt);
1993         }
1994         qp->uobject = &obj->uevent.uobject;
1995
1996         obj->uevent.uobject.object = qp;
1997         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1998         if (ret)
1999                 goto err_destroy;
2000
2001         memset(&resp, 0, sizeof resp);
2002         resp.base.qpn             = qp->qp_num;
2003         resp.base.qp_handle       = obj->uevent.uobject.id;
2004         resp.base.max_recv_sge    = attr.cap.max_recv_sge;
2005         resp.base.max_send_sge    = attr.cap.max_send_sge;
2006         resp.base.max_recv_wr     = attr.cap.max_recv_wr;
2007         resp.base.max_send_wr     = attr.cap.max_send_wr;
2008         resp.base.max_inline_data = attr.cap.max_inline_data;
2009
2010         resp.response_length = offsetof(typeof(resp), response_length) +
2011                                sizeof(resp.response_length);
2012
2013         ret = cb(file, &resp, ucore);
2014         if (ret)
2015                 goto err_cb;
2016
2017         if (xrcd) {
2018                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object,
2019                                           uobject);
2020                 atomic_inc(&obj->uxrcd->refcnt);
2021                 put_xrcd_read(xrcd_uobj);
2022         }
2023
2024         if (pd)
2025                 put_pd_read(pd);
2026         if (scq)
2027                 put_cq_read(scq);
2028         if (rcq && rcq != scq)
2029                 put_cq_read(rcq);
2030         if (srq)
2031                 put_srq_read(srq);
2032         if (ind_tbl)
2033                 put_rwq_indirection_table_read(ind_tbl);
2034
2035         mutex_lock(&file->mutex);
2036         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
2037         mutex_unlock(&file->mutex);
2038
2039         obj->uevent.uobject.live = 1;
2040
2041         up_write(&obj->uevent.uobject.mutex);
2042
2043         return 0;
2044 err_cb:
2045         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
2046
2047 err_destroy:
2048         ib_destroy_qp(qp);
2049
2050 err_create:
2051         ib_rdmacg_uncharge(&obj->uevent.uobject.cg_obj, device,
2052                            RDMACG_RESOURCE_HCA_OBJECT);
2053
2054 err_put:
2055         if (xrcd)
2056                 put_xrcd_read(xrcd_uobj);
2057         if (pd)
2058                 put_pd_read(pd);
2059         if (scq)
2060                 put_cq_read(scq);
2061         if (rcq && rcq != scq)
2062                 put_cq_read(rcq);
2063         if (srq)
2064                 put_srq_read(srq);
2065         if (ind_tbl)
2066                 put_rwq_indirection_table_read(ind_tbl);
2067
2068         put_uobj_write(&obj->uevent.uobject);
2069         return ret;
2070 }
2071
2072 static int ib_uverbs_create_qp_cb(struct ib_uverbs_file *file,
2073                                   struct ib_uverbs_ex_create_qp_resp *resp,
2074                                   struct ib_udata *ucore)
2075 {
2076         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
2077                 return -EFAULT;
2078
2079         return 0;
2080 }
2081
2082 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
2083                             struct ib_device *ib_dev,
2084                             const char __user *buf, int in_len,
2085                             int out_len)
2086 {
2087         struct ib_uverbs_create_qp      cmd;
2088         struct ib_uverbs_ex_create_qp   cmd_ex;
2089         struct ib_udata                 ucore;
2090         struct ib_udata                 uhw;
2091         ssize_t resp_size = sizeof(struct ib_uverbs_create_qp_resp);
2092         int                             err;
2093
2094         if (out_len < resp_size)
2095                 return -ENOSPC;
2096
2097         if (copy_from_user(&cmd, buf, sizeof(cmd)))
2098                 return -EFAULT;
2099
2100         INIT_UDATA(&ucore, buf, (unsigned long)cmd.response, sizeof(cmd),
2101                    resp_size);
2102         INIT_UDATA(&uhw, buf + sizeof(cmd),
2103                    (unsigned long)cmd.response + resp_size,
2104                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
2105                    out_len - resp_size);
2106
2107         memset(&cmd_ex, 0, sizeof(cmd_ex));
2108         cmd_ex.user_handle = cmd.user_handle;
2109         cmd_ex.pd_handle = cmd.pd_handle;
2110         cmd_ex.send_cq_handle = cmd.send_cq_handle;
2111         cmd_ex.recv_cq_handle = cmd.recv_cq_handle;
2112         cmd_ex.srq_handle = cmd.srq_handle;
2113         cmd_ex.max_send_wr = cmd.max_send_wr;
2114         cmd_ex.max_recv_wr = cmd.max_recv_wr;
2115         cmd_ex.max_send_sge = cmd.max_send_sge;
2116         cmd_ex.max_recv_sge = cmd.max_recv_sge;
2117         cmd_ex.max_inline_data = cmd.max_inline_data;
2118         cmd_ex.sq_sig_all = cmd.sq_sig_all;
2119         cmd_ex.qp_type = cmd.qp_type;
2120         cmd_ex.is_srq = cmd.is_srq;
2121
2122         err = create_qp(file, &ucore, &uhw, &cmd_ex,
2123                         offsetof(typeof(cmd_ex), is_srq) +
2124                         sizeof(cmd.is_srq), ib_uverbs_create_qp_cb,
2125                         NULL);
2126
2127         if (err)
2128                 return err;
2129
2130         return in_len;
2131 }
2132
2133 static int ib_uverbs_ex_create_qp_cb(struct ib_uverbs_file *file,
2134                                      struct ib_uverbs_ex_create_qp_resp *resp,
2135                                      struct ib_udata *ucore)
2136 {
2137         if (ib_copy_to_udata(ucore, resp, resp->response_length))
2138                 return -EFAULT;
2139
2140         return 0;
2141 }
2142
2143 int ib_uverbs_ex_create_qp(struct ib_uverbs_file *file,
2144                            struct ib_device *ib_dev,
2145                            struct ib_udata *ucore,
2146                            struct ib_udata *uhw)
2147 {
2148         struct ib_uverbs_ex_create_qp_resp resp;
2149         struct ib_uverbs_ex_create_qp cmd = {0};
2150         int err;
2151
2152         if (ucore->inlen < (offsetof(typeof(cmd), comp_mask) +
2153                             sizeof(cmd.comp_mask)))
2154                 return -EINVAL;
2155
2156         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2157         if (err)
2158                 return err;
2159
2160         if (cmd.comp_mask & ~IB_UVERBS_CREATE_QP_SUP_COMP_MASK)
2161                 return -EINVAL;
2162
2163         if (cmd.reserved)
2164                 return -EINVAL;
2165
2166         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
2167                              sizeof(resp.response_length)))
2168                 return -ENOSPC;
2169
2170         err = create_qp(file, ucore, uhw, &cmd,
2171                         min(ucore->inlen, sizeof(cmd)),
2172                         ib_uverbs_ex_create_qp_cb, NULL);
2173
2174         if (err)
2175                 return err;
2176
2177         return 0;
2178 }
2179
2180 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
2181                           struct ib_device *ib_dev,
2182                           const char __user *buf, int in_len, int out_len)
2183 {
2184         struct ib_uverbs_open_qp        cmd;
2185         struct ib_uverbs_create_qp_resp resp;
2186         struct ib_udata                 udata;
2187         struct ib_uqp_object           *obj;
2188         struct ib_xrcd                 *xrcd;
2189         struct ib_uobject              *uninitialized_var(xrcd_uobj);
2190         struct ib_qp                   *qp;
2191         struct ib_qp_open_attr          attr;
2192         int ret;
2193
2194         if (out_len < sizeof resp)
2195                 return -ENOSPC;
2196
2197         if (copy_from_user(&cmd, buf, sizeof cmd))
2198                 return -EFAULT;
2199
2200         INIT_UDATA(&udata, buf + sizeof cmd,
2201                    (unsigned long) cmd.response + sizeof resp,
2202                    in_len - sizeof cmd, out_len - sizeof resp);
2203
2204         obj = kmalloc(sizeof *obj, GFP_KERNEL);
2205         if (!obj)
2206                 return -ENOMEM;
2207
2208         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
2209         down_write(&obj->uevent.uobject.mutex);
2210
2211         xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
2212         if (!xrcd) {
2213                 ret = -EINVAL;
2214                 goto err_put;
2215         }
2216
2217         attr.event_handler = ib_uverbs_qp_event_handler;
2218         attr.qp_context    = file;
2219         attr.qp_num        = cmd.qpn;
2220         attr.qp_type       = cmd.qp_type;
2221
2222         obj->uevent.events_reported = 0;
2223         INIT_LIST_HEAD(&obj->uevent.event_list);
2224         INIT_LIST_HEAD(&obj->mcast_list);
2225
2226         qp = ib_open_qp(xrcd, &attr);
2227         if (IS_ERR(qp)) {
2228                 ret = PTR_ERR(qp);
2229                 goto err_put;
2230         }
2231
2232         qp->uobject = &obj->uevent.uobject;
2233
2234         obj->uevent.uobject.object = qp;
2235         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
2236         if (ret)
2237                 goto err_destroy;
2238
2239         memset(&resp, 0, sizeof resp);
2240         resp.qpn       = qp->qp_num;
2241         resp.qp_handle = obj->uevent.uobject.id;
2242
2243         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2244                          &resp, sizeof resp)) {
2245                 ret = -EFAULT;
2246                 goto err_remove;
2247         }
2248
2249         obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
2250         atomic_inc(&obj->uxrcd->refcnt);
2251         put_xrcd_read(xrcd_uobj);
2252
2253         mutex_lock(&file->mutex);
2254         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
2255         mutex_unlock(&file->mutex);
2256
2257         obj->uevent.uobject.live = 1;
2258
2259         up_write(&obj->uevent.uobject.mutex);
2260
2261         return in_len;
2262
2263 err_remove:
2264         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
2265
2266 err_destroy:
2267         ib_destroy_qp(qp);
2268
2269 err_put:
2270         put_xrcd_read(xrcd_uobj);
2271         put_uobj_write(&obj->uevent.uobject);
2272         return ret;
2273 }
2274
2275 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
2276                            struct ib_device *ib_dev,
2277                            const char __user *buf, int in_len,
2278                            int out_len)
2279 {
2280         struct ib_uverbs_query_qp      cmd;
2281         struct ib_uverbs_query_qp_resp resp;
2282         struct ib_qp                   *qp;
2283         struct ib_qp_attr              *attr;
2284         struct ib_qp_init_attr         *init_attr;
2285         int                            ret;
2286
2287         if (copy_from_user(&cmd, buf, sizeof cmd))
2288                 return -EFAULT;
2289
2290         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
2291         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
2292         if (!attr || !init_attr) {
2293                 ret = -ENOMEM;
2294                 goto out;
2295         }
2296
2297         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2298         if (!qp) {
2299                 ret = -EINVAL;
2300                 goto out;
2301         }
2302
2303         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
2304
2305         put_qp_read(qp);
2306
2307         if (ret)
2308                 goto out;
2309
2310         memset(&resp, 0, sizeof resp);
2311
2312         resp.qp_state               = attr->qp_state;
2313         resp.cur_qp_state           = attr->cur_qp_state;
2314         resp.path_mtu               = attr->path_mtu;
2315         resp.path_mig_state         = attr->path_mig_state;
2316         resp.qkey                   = attr->qkey;
2317         resp.rq_psn                 = attr->rq_psn;
2318         resp.sq_psn                 = attr->sq_psn;
2319         resp.dest_qp_num            = attr->dest_qp_num;
2320         resp.qp_access_flags        = attr->qp_access_flags;
2321         resp.pkey_index             = attr->pkey_index;
2322         resp.alt_pkey_index         = attr->alt_pkey_index;
2323         resp.sq_draining            = attr->sq_draining;
2324         resp.max_rd_atomic          = attr->max_rd_atomic;
2325         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
2326         resp.min_rnr_timer          = attr->min_rnr_timer;
2327         resp.port_num               = attr->port_num;
2328         resp.timeout                = attr->timeout;
2329         resp.retry_cnt              = attr->retry_cnt;
2330         resp.rnr_retry              = attr->rnr_retry;
2331         resp.alt_port_num           = attr->alt_port_num;
2332         resp.alt_timeout            = attr->alt_timeout;
2333
2334         memcpy(resp.dest.dgid, attr->ah_attr.grh.dgid.raw, 16);
2335         resp.dest.flow_label        = attr->ah_attr.grh.flow_label;
2336         resp.dest.sgid_index        = attr->ah_attr.grh.sgid_index;
2337         resp.dest.hop_limit         = attr->ah_attr.grh.hop_limit;
2338         resp.dest.traffic_class     = attr->ah_attr.grh.traffic_class;
2339         resp.dest.dlid              = attr->ah_attr.dlid;
2340         resp.dest.sl                = attr->ah_attr.sl;
2341         resp.dest.src_path_bits     = attr->ah_attr.src_path_bits;
2342         resp.dest.static_rate       = attr->ah_attr.static_rate;
2343         resp.dest.is_global         = !!(attr->ah_attr.ah_flags & IB_AH_GRH);
2344         resp.dest.port_num          = attr->ah_attr.port_num;
2345
2346         memcpy(resp.alt_dest.dgid, attr->alt_ah_attr.grh.dgid.raw, 16);
2347         resp.alt_dest.flow_label    = attr->alt_ah_attr.grh.flow_label;
2348         resp.alt_dest.sgid_index    = attr->alt_ah_attr.grh.sgid_index;
2349         resp.alt_dest.hop_limit     = attr->alt_ah_attr.grh.hop_limit;
2350         resp.alt_dest.traffic_class = attr->alt_ah_attr.grh.traffic_class;
2351         resp.alt_dest.dlid          = attr->alt_ah_attr.dlid;
2352         resp.alt_dest.sl            = attr->alt_ah_attr.sl;
2353         resp.alt_dest.src_path_bits = attr->alt_ah_attr.src_path_bits;
2354         resp.alt_dest.static_rate   = attr->alt_ah_attr.static_rate;
2355         resp.alt_dest.is_global     = !!(attr->alt_ah_attr.ah_flags & IB_AH_GRH);
2356         resp.alt_dest.port_num      = attr->alt_ah_attr.port_num;
2357
2358         resp.max_send_wr            = init_attr->cap.max_send_wr;
2359         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
2360         resp.max_send_sge           = init_attr->cap.max_send_sge;
2361         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
2362         resp.max_inline_data        = init_attr->cap.max_inline_data;
2363         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
2364
2365         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2366                          &resp, sizeof resp))
2367                 ret = -EFAULT;
2368
2369 out:
2370         kfree(attr);
2371         kfree(init_attr);
2372
2373         return ret ? ret : in_len;
2374 }
2375
2376 /* Remove ignored fields set in the attribute mask */
2377 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
2378 {
2379         switch (qp_type) {
2380         case IB_QPT_XRC_INI:
2381                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
2382         case IB_QPT_XRC_TGT:
2383                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
2384                                 IB_QP_RNR_RETRY);
2385         default:
2386                 return mask;
2387         }
2388 }
2389
2390 static int modify_qp(struct ib_uverbs_file *file,
2391                      struct ib_uverbs_ex_modify_qp *cmd, struct ib_udata *udata)
2392 {
2393         struct ib_qp_attr *attr;
2394         struct ib_qp *qp;
2395         int ret;
2396
2397         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2398         if (!attr)
2399                 return -ENOMEM;
2400
2401         qp = idr_read_qp(cmd->base.qp_handle, file->ucontext);
2402         if (!qp) {
2403                 ret = -EINVAL;
2404                 goto out;
2405         }
2406
2407         attr->qp_state            = cmd->base.qp_state;
2408         attr->cur_qp_state        = cmd->base.cur_qp_state;
2409         attr->path_mtu            = cmd->base.path_mtu;
2410         attr->path_mig_state      = cmd->base.path_mig_state;
2411         attr->qkey                = cmd->base.qkey;
2412         attr->rq_psn              = cmd->base.rq_psn;
2413         attr->sq_psn              = cmd->base.sq_psn;
2414         attr->dest_qp_num         = cmd->base.dest_qp_num;
2415         attr->qp_access_flags     = cmd->base.qp_access_flags;
2416         attr->pkey_index          = cmd->base.pkey_index;
2417         attr->alt_pkey_index      = cmd->base.alt_pkey_index;
2418         attr->en_sqd_async_notify = cmd->base.en_sqd_async_notify;
2419         attr->max_rd_atomic       = cmd->base.max_rd_atomic;
2420         attr->max_dest_rd_atomic  = cmd->base.max_dest_rd_atomic;
2421         attr->min_rnr_timer       = cmd->base.min_rnr_timer;
2422         attr->port_num            = cmd->base.port_num;
2423         attr->timeout             = cmd->base.timeout;
2424         attr->retry_cnt           = cmd->base.retry_cnt;
2425         attr->rnr_retry           = cmd->base.rnr_retry;
2426         attr->alt_port_num        = cmd->base.alt_port_num;
2427         attr->alt_timeout         = cmd->base.alt_timeout;
2428         attr->rate_limit          = cmd->rate_limit;
2429
2430         memcpy(attr->ah_attr.grh.dgid.raw, cmd->base.dest.dgid, 16);
2431         attr->ah_attr.grh.flow_label    = cmd->base.dest.flow_label;
2432         attr->ah_attr.grh.sgid_index    = cmd->base.dest.sgid_index;
2433         attr->ah_attr.grh.hop_limit     = cmd->base.dest.hop_limit;
2434         attr->ah_attr.grh.traffic_class = cmd->base.dest.traffic_class;
2435         attr->ah_attr.dlid              = cmd->base.dest.dlid;
2436         attr->ah_attr.sl                = cmd->base.dest.sl;
2437         attr->ah_attr.src_path_bits     = cmd->base.dest.src_path_bits;
2438         attr->ah_attr.static_rate       = cmd->base.dest.static_rate;
2439         attr->ah_attr.ah_flags          = cmd->base.dest.is_global ?
2440                                           IB_AH_GRH : 0;
2441         attr->ah_attr.port_num          = cmd->base.dest.port_num;
2442
2443         memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd->base.alt_dest.dgid, 16);
2444         attr->alt_ah_attr.grh.flow_label    = cmd->base.alt_dest.flow_label;
2445         attr->alt_ah_attr.grh.sgid_index    = cmd->base.alt_dest.sgid_index;
2446         attr->alt_ah_attr.grh.hop_limit     = cmd->base.alt_dest.hop_limit;
2447         attr->alt_ah_attr.grh.traffic_class = cmd->base.alt_dest.traffic_class;
2448         attr->alt_ah_attr.dlid              = cmd->base.alt_dest.dlid;
2449         attr->alt_ah_attr.sl                = cmd->base.alt_dest.sl;
2450         attr->alt_ah_attr.src_path_bits     = cmd->base.alt_dest.src_path_bits;
2451         attr->alt_ah_attr.static_rate       = cmd->base.alt_dest.static_rate;
2452         attr->alt_ah_attr.ah_flags          = cmd->base.alt_dest.is_global ?
2453                                               IB_AH_GRH : 0;
2454         attr->alt_ah_attr.port_num          = cmd->base.alt_dest.port_num;
2455
2456         if (qp->real_qp == qp) {
2457                 if (cmd->base.attr_mask & IB_QP_AV) {
2458                         ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
2459                         if (ret)
2460                                 goto release_qp;
2461                 }
2462                 ret = qp->device->modify_qp(qp, attr,
2463                                             modify_qp_mask(qp->qp_type,
2464                                                            cmd->base.attr_mask),
2465                                             udata);
2466         } else {
2467                 ret = ib_modify_qp(qp, attr,
2468                                    modify_qp_mask(qp->qp_type,
2469                                                   cmd->base.attr_mask));
2470         }
2471
2472 release_qp:
2473         put_qp_read(qp);
2474
2475 out:
2476         kfree(attr);
2477
2478         return ret;
2479 }
2480
2481 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
2482                             struct ib_device *ib_dev,
2483                             const char __user *buf, int in_len,
2484                             int out_len)
2485 {
2486         struct ib_uverbs_ex_modify_qp cmd = {};
2487         struct ib_udata udata;
2488         int ret;
2489
2490         if (copy_from_user(&cmd.base, buf, sizeof(cmd.base)))
2491                 return -EFAULT;
2492
2493         if (cmd.base.attr_mask &
2494             ~((IB_USER_LEGACY_LAST_QP_ATTR_MASK << 1) - 1))
2495                 return -EOPNOTSUPP;
2496
2497         INIT_UDATA(&udata, buf + sizeof(cmd.base), NULL,
2498                    in_len - sizeof(cmd.base), out_len);
2499
2500         ret = modify_qp(file, &cmd, &udata);
2501         if (ret)
2502                 return ret;
2503
2504         return in_len;
2505 }
2506
2507 int ib_uverbs_ex_modify_qp(struct ib_uverbs_file *file,
2508                            struct ib_device *ib_dev,
2509                            struct ib_udata *ucore,
2510                            struct ib_udata *uhw)
2511 {
2512         struct ib_uverbs_ex_modify_qp cmd = {};
2513         int ret;
2514
2515         /*
2516          * Last bit is reserved for extending the attr_mask by
2517          * using another field.
2518          */
2519         BUILD_BUG_ON(IB_USER_LAST_QP_ATTR_MASK == (1 << 31));
2520
2521         if (ucore->inlen < sizeof(cmd.base))
2522                 return -EINVAL;
2523
2524         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2525         if (ret)
2526                 return ret;
2527
2528         if (cmd.base.attr_mask &
2529             ~((IB_USER_LAST_QP_ATTR_MASK << 1) - 1))
2530                 return -EOPNOTSUPP;
2531
2532         if (ucore->inlen > sizeof(cmd)) {
2533                 if (ib_is_udata_cleared(ucore, sizeof(cmd),
2534                                         ucore->inlen - sizeof(cmd)))
2535                         return -EOPNOTSUPP;
2536         }
2537
2538         ret = modify_qp(file, &cmd, uhw);
2539
2540         return ret;
2541 }
2542
2543 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
2544                              struct ib_device *ib_dev,
2545                              const char __user *buf, int in_len,
2546                              int out_len)
2547 {
2548         struct ib_uverbs_destroy_qp      cmd;
2549         struct ib_uverbs_destroy_qp_resp resp;
2550         struct ib_uobject               *uobj;
2551         struct ib_qp                    *qp;
2552         struct ib_uqp_object            *obj;
2553         int                              ret = -EINVAL;
2554
2555         if (copy_from_user(&cmd, buf, sizeof cmd))
2556                 return -EFAULT;
2557
2558         memset(&resp, 0, sizeof resp);
2559
2560         uobj = idr_write_uobj(&ib_uverbs_qp_idr, cmd.qp_handle, file->ucontext);
2561         if (!uobj)
2562                 return -EINVAL;
2563         qp  = uobj->object;
2564         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2565
2566         if (!list_empty(&obj->mcast_list)) {
2567                 put_uobj_write(uobj);
2568                 return -EBUSY;
2569         }
2570
2571         ret = ib_destroy_qp(qp);
2572         if (!ret)
2573                 uobj->live = 0;
2574
2575         put_uobj_write(uobj);
2576
2577         if (ret)
2578                 return ret;
2579
2580         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
2581
2582         if (obj->uxrcd)
2583                 atomic_dec(&obj->uxrcd->refcnt);
2584
2585         idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
2586
2587         mutex_lock(&file->mutex);
2588         list_del(&uobj->list);
2589         mutex_unlock(&file->mutex);
2590
2591         ib_uverbs_release_uevent(file, &obj->uevent);
2592
2593         resp.events_reported = obj->uevent.events_reported;
2594
2595         put_uobj(uobj);
2596
2597         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2598                          &resp, sizeof resp))
2599                 return -EFAULT;
2600
2601         return in_len;
2602 }
2603
2604 static void *alloc_wr(size_t wr_size, __u32 num_sge)
2605 {
2606         return kmalloc(ALIGN(wr_size, sizeof (struct ib_sge)) +
2607                          num_sge * sizeof (struct ib_sge), GFP_KERNEL);
2608 };
2609
2610 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2611                             struct ib_device *ib_dev,
2612                             const char __user *buf, int in_len,
2613                             int out_len)
2614 {
2615         struct ib_uverbs_post_send      cmd;
2616         struct ib_uverbs_post_send_resp resp;
2617         struct ib_uverbs_send_wr       *user_wr;
2618         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
2619         struct ib_qp                   *qp;
2620         int                             i, sg_ind;
2621         int                             is_ud;
2622         ssize_t                         ret = -EINVAL;
2623         size_t                          next_size;
2624
2625         if (copy_from_user(&cmd, buf, sizeof cmd))
2626                 return -EFAULT;
2627
2628         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2629             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2630                 return -EINVAL;
2631
2632         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2633                 return -EINVAL;
2634
2635         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2636         if (!user_wr)
2637                 return -ENOMEM;
2638
2639         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2640         if (!qp)
2641                 goto out;
2642
2643         is_ud = qp->qp_type == IB_QPT_UD;
2644         sg_ind = 0;
2645         last = NULL;
2646         for (i = 0; i < cmd.wr_count; ++i) {
2647                 if (copy_from_user(user_wr,
2648                                    buf + sizeof cmd + i * cmd.wqe_size,
2649                                    cmd.wqe_size)) {
2650                         ret = -EFAULT;
2651                         goto out_put;
2652                 }
2653
2654                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2655                         ret = -EINVAL;
2656                         goto out_put;
2657                 }
2658
2659                 if (is_ud) {
2660                         struct ib_ud_wr *ud;
2661
2662                         if (user_wr->opcode != IB_WR_SEND &&
2663                             user_wr->opcode != IB_WR_SEND_WITH_IMM) {
2664                                 ret = -EINVAL;
2665                                 goto out_put;
2666                         }
2667
2668                         next_size = sizeof(*ud);
2669                         ud = alloc_wr(next_size, user_wr->num_sge);
2670                         if (!ud) {
2671                                 ret = -ENOMEM;
2672                                 goto out_put;
2673                         }
2674
2675                         ud->ah = idr_read_ah(user_wr->wr.ud.ah, file->ucontext);
2676                         if (!ud->ah) {
2677                                 kfree(ud);
2678                                 ret = -EINVAL;
2679                                 goto out_put;
2680                         }
2681                         ud->remote_qpn = user_wr->wr.ud.remote_qpn;
2682                         ud->remote_qkey = user_wr->wr.ud.remote_qkey;
2683
2684                         next = &ud->wr;
2685                 } else if (user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
2686                            user_wr->opcode == IB_WR_RDMA_WRITE ||
2687                            user_wr->opcode == IB_WR_RDMA_READ) {
2688                         struct ib_rdma_wr *rdma;
2689
2690                         next_size = sizeof(*rdma);
2691                         rdma = alloc_wr(next_size, user_wr->num_sge);
2692                         if (!rdma) {
2693                                 ret = -ENOMEM;
2694                                 goto out_put;
2695                         }
2696
2697                         rdma->remote_addr = user_wr->wr.rdma.remote_addr;
2698                         rdma->rkey = user_wr->wr.rdma.rkey;
2699
2700                         next = &rdma->wr;
2701                 } else if (user_wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2702                            user_wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2703                         struct ib_atomic_wr *atomic;
2704
2705                         next_size = sizeof(*atomic);
2706                         atomic = alloc_wr(next_size, user_wr->num_sge);
2707                         if (!atomic) {
2708                                 ret = -ENOMEM;
2709                                 goto out_put;
2710                         }
2711
2712                         atomic->remote_addr = user_wr->wr.atomic.remote_addr;
2713                         atomic->compare_add = user_wr->wr.atomic.compare_add;
2714                         atomic->swap = user_wr->wr.atomic.swap;
2715                         atomic->rkey = user_wr->wr.atomic.rkey;
2716
2717                         next = &atomic->wr;
2718                 } else if (user_wr->opcode == IB_WR_SEND ||
2719                            user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2720                            user_wr->opcode == IB_WR_SEND_WITH_INV) {
2721                         next_size = sizeof(*next);
2722                         next = alloc_wr(next_size, user_wr->num_sge);
2723                         if (!next) {
2724                                 ret = -ENOMEM;
2725                                 goto out_put;
2726                         }
2727                 } else {
2728                         ret = -EINVAL;
2729                         goto out_put;
2730                 }
2731
2732                 if (user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2733                     user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
2734                         next->ex.imm_data =
2735                                         (__be32 __force) user_wr->ex.imm_data;
2736                 } else if (user_wr->opcode == IB_WR_SEND_WITH_INV) {
2737                         next->ex.invalidate_rkey = user_wr->ex.invalidate_rkey;
2738                 }
2739
2740                 if (!last)
2741                         wr = next;
2742                 else
2743                         last->next = next;
2744                 last = next;
2745
2746                 next->next       = NULL;
2747                 next->wr_id      = user_wr->wr_id;
2748                 next->num_sge    = user_wr->num_sge;
2749                 next->opcode     = user_wr->opcode;
2750                 next->send_flags = user_wr->send_flags;
2751
2752                 if (next->num_sge) {
2753                         next->sg_list = (void *) next +
2754                                 ALIGN(next_size, sizeof(struct ib_sge));
2755                         if (copy_from_user(next->sg_list,
2756                                            buf + sizeof cmd +
2757                                            cmd.wr_count * cmd.wqe_size +
2758                                            sg_ind * sizeof (struct ib_sge),
2759                                            next->num_sge * sizeof (struct ib_sge))) {
2760                                 ret = -EFAULT;
2761                                 goto out_put;
2762                         }
2763                         sg_ind += next->num_sge;
2764                 } else
2765                         next->sg_list = NULL;
2766         }
2767
2768         resp.bad_wr = 0;
2769         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2770         if (ret)
2771                 for (next = wr; next; next = next->next) {
2772                         ++resp.bad_wr;
2773                         if (next == bad_wr)
2774                                 break;
2775                 }
2776
2777         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2778                          &resp, sizeof resp))
2779                 ret = -EFAULT;
2780
2781 out_put:
2782         put_qp_read(qp);
2783
2784         while (wr) {
2785                 if (is_ud && ud_wr(wr)->ah)
2786                         put_ah_read(ud_wr(wr)->ah);
2787                 next = wr->next;
2788                 kfree(wr);
2789                 wr = next;
2790         }
2791
2792 out:
2793         kfree(user_wr);
2794
2795         return ret ? ret : in_len;
2796 }
2797
2798 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2799                                                     int in_len,
2800                                                     u32 wr_count,
2801                                                     u32 sge_count,
2802                                                     u32 wqe_size)
2803 {
2804         struct ib_uverbs_recv_wr *user_wr;
2805         struct ib_recv_wr        *wr = NULL, *last, *next;
2806         int                       sg_ind;
2807         int                       i;
2808         int                       ret;
2809
2810         if (in_len < wqe_size * wr_count +
2811             sge_count * sizeof (struct ib_uverbs_sge))
2812                 return ERR_PTR(-EINVAL);
2813
2814         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2815                 return ERR_PTR(-EINVAL);
2816
2817         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2818         if (!user_wr)
2819                 return ERR_PTR(-ENOMEM);
2820
2821         sg_ind = 0;
2822         last = NULL;
2823         for (i = 0; i < wr_count; ++i) {
2824                 if (copy_from_user(user_wr, buf + i * wqe_size,
2825                                    wqe_size)) {
2826                         ret = -EFAULT;
2827                         goto err;
2828                 }
2829
2830                 if (user_wr->num_sge + sg_ind > sge_count) {
2831                         ret = -EINVAL;
2832                         goto err;
2833                 }
2834
2835                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2836                                user_wr->num_sge * sizeof (struct ib_sge),
2837                                GFP_KERNEL);
2838                 if (!next) {
2839                         ret = -ENOMEM;
2840                         goto err;
2841                 }
2842
2843                 if (!last)
2844                         wr = next;
2845                 else
2846                         last->next = next;
2847                 last = next;
2848
2849                 next->next       = NULL;
2850                 next->wr_id      = user_wr->wr_id;
2851                 next->num_sge    = user_wr->num_sge;
2852
2853                 if (next->num_sge) {
2854                         next->sg_list = (void *) next +
2855                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2856                         if (copy_from_user(next->sg_list,
2857                                            buf + wr_count * wqe_size +
2858                                            sg_ind * sizeof (struct ib_sge),
2859                                            next->num_sge * sizeof (struct ib_sge))) {
2860                                 ret = -EFAULT;
2861                                 goto err;
2862                         }
2863                         sg_ind += next->num_sge;
2864                 } else
2865                         next->sg_list = NULL;
2866         }
2867
2868         kfree(user_wr);
2869         return wr;
2870
2871 err:
2872         kfree(user_wr);
2873
2874         while (wr) {
2875                 next = wr->next;
2876                 kfree(wr);
2877                 wr = next;
2878         }
2879
2880         return ERR_PTR(ret);
2881 }
2882
2883 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2884                             struct ib_device *ib_dev,
2885                             const char __user *buf, int in_len,
2886                             int out_len)
2887 {
2888         struct ib_uverbs_post_recv      cmd;
2889         struct ib_uverbs_post_recv_resp resp;
2890         struct ib_recv_wr              *wr, *next, *bad_wr;
2891         struct ib_qp                   *qp;
2892         ssize_t                         ret = -EINVAL;
2893
2894         if (copy_from_user(&cmd, buf, sizeof cmd))
2895                 return -EFAULT;
2896
2897         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2898                                        in_len - sizeof cmd, cmd.wr_count,
2899                                        cmd.sge_count, cmd.wqe_size);
2900         if (IS_ERR(wr))
2901                 return PTR_ERR(wr);
2902
2903         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2904         if (!qp)
2905                 goto out;
2906
2907         resp.bad_wr = 0;
2908         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2909
2910         put_qp_read(qp);
2911
2912         if (ret)
2913                 for (next = wr; next; next = next->next) {
2914                         ++resp.bad_wr;
2915                         if (next == bad_wr)
2916                                 break;
2917                 }
2918
2919         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2920                          &resp, sizeof resp))
2921                 ret = -EFAULT;
2922
2923 out:
2924         while (wr) {
2925                 next = wr->next;
2926                 kfree(wr);
2927                 wr = next;
2928         }
2929
2930         return ret ? ret : in_len;
2931 }
2932
2933 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2934                                 struct ib_device *ib_dev,
2935                                 const char __user *buf, int in_len,
2936                                 int out_len)
2937 {
2938         struct ib_uverbs_post_srq_recv      cmd;
2939         struct ib_uverbs_post_srq_recv_resp resp;
2940         struct ib_recv_wr                  *wr, *next, *bad_wr;
2941         struct ib_srq                      *srq;
2942         ssize_t                             ret = -EINVAL;
2943
2944         if (copy_from_user(&cmd, buf, sizeof cmd))
2945                 return -EFAULT;
2946
2947         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2948                                        in_len - sizeof cmd, cmd.wr_count,
2949                                        cmd.sge_count, cmd.wqe_size);
2950         if (IS_ERR(wr))
2951                 return PTR_ERR(wr);
2952
2953         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2954         if (!srq)
2955                 goto out;
2956
2957         resp.bad_wr = 0;
2958         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2959
2960         put_srq_read(srq);
2961
2962         if (ret)
2963                 for (next = wr; next; next = next->next) {
2964                         ++resp.bad_wr;
2965                         if (next == bad_wr)
2966                                 break;
2967                 }
2968
2969         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2970                          &resp, sizeof resp))
2971                 ret = -EFAULT;
2972
2973 out:
2974         while (wr) {
2975                 next = wr->next;
2976                 kfree(wr);
2977                 wr = next;
2978         }
2979
2980         return ret ? ret : in_len;
2981 }
2982
2983 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2984                             struct ib_device *ib_dev,
2985                             const char __user *buf, int in_len,
2986                             int out_len)
2987 {
2988         struct ib_uverbs_create_ah       cmd;
2989         struct ib_uverbs_create_ah_resp  resp;
2990         struct ib_uobject               *uobj;
2991         struct ib_pd                    *pd;
2992         struct ib_ah                    *ah;
2993         struct ib_ah_attr               attr;
2994         int ret;
2995         struct ib_udata                   udata;
2996
2997         if (out_len < sizeof resp)
2998                 return -ENOSPC;
2999
3000         if (copy_from_user(&cmd, buf, sizeof cmd))
3001                 return -EFAULT;
3002
3003         INIT_UDATA(&udata, buf + sizeof(cmd),
3004                    (unsigned long)cmd.response + sizeof(resp),
3005                    in_len - sizeof(cmd), out_len - sizeof(resp));
3006
3007         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
3008         if (!uobj)
3009                 return -ENOMEM;
3010
3011         init_uobj(uobj, cmd.user_handle, file->ucontext, &ah_lock_class);
3012         down_write(&uobj->mutex);
3013
3014         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
3015         if (!pd) {
3016                 ret = -EINVAL;
3017                 goto err;
3018         }
3019
3020         attr.dlid              = cmd.attr.dlid;
3021         attr.sl                = cmd.attr.sl;
3022         attr.src_path_bits     = cmd.attr.src_path_bits;
3023         attr.static_rate       = cmd.attr.static_rate;
3024         attr.ah_flags          = cmd.attr.is_global ? IB_AH_GRH : 0;
3025         attr.port_num          = cmd.attr.port_num;
3026         attr.grh.flow_label    = cmd.attr.grh.flow_label;
3027         attr.grh.sgid_index    = cmd.attr.grh.sgid_index;
3028         attr.grh.hop_limit     = cmd.attr.grh.hop_limit;
3029         attr.grh.traffic_class = cmd.attr.grh.traffic_class;
3030         memset(&attr.dmac, 0, sizeof(attr.dmac));
3031         memcpy(attr.grh.dgid.raw, cmd.attr.grh.dgid, 16);
3032
3033         ret = ib_rdmacg_try_charge(&uobj->cg_obj, ib_dev,
3034                                    RDMACG_RESOURCE_HCA_OBJECT);
3035         if (ret)
3036                 goto err_charge;
3037
3038         ah = pd->device->create_ah(pd, &attr, &udata);
3039
3040         if (IS_ERR(ah)) {
3041                 ret = PTR_ERR(ah);
3042                 goto err_create;
3043         }
3044
3045         ah->device  = pd->device;
3046         ah->pd      = pd;
3047         atomic_inc(&pd->usecnt);
3048         ah->uobject  = uobj;
3049         uobj->object = ah;
3050
3051         ret = idr_add_uobj(&ib_uverbs_ah_idr, uobj);
3052         if (ret)
3053                 goto err_destroy;
3054
3055         resp.ah_handle = uobj->id;
3056
3057         if (copy_to_user((void __user *) (unsigned long) cmd.response,
3058                          &resp, sizeof resp)) {
3059                 ret = -EFAULT;
3060                 goto err_copy;
3061         }
3062
3063         put_pd_read(pd);
3064
3065         mutex_lock(&file->mutex);
3066         list_add_tail(&uobj->list, &file->ucontext->ah_list);
3067         mutex_unlock(&file->mutex);
3068
3069         uobj->live = 1;
3070
3071         up_write(&uobj->mutex);
3072
3073         return in_len;
3074
3075 err_copy:
3076         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
3077
3078 err_destroy:
3079         ib_destroy_ah(ah);
3080
3081 err_create:
3082         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
3083
3084 err_charge:
3085         put_pd_read(pd);
3086
3087 err:
3088         put_uobj_write(uobj);
3089         return ret;
3090 }
3091
3092 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
3093                              struct ib_device *ib_dev,
3094                              const char __user *buf, int in_len, int out_len)
3095 {
3096         struct ib_uverbs_destroy_ah cmd;
3097         struct ib_ah               *ah;
3098         struct ib_uobject          *uobj;
3099         int                         ret;
3100
3101         if (copy_from_user(&cmd, buf, sizeof cmd))
3102                 return -EFAULT;
3103
3104         uobj = idr_write_uobj(&ib_uverbs_ah_idr, cmd.ah_handle, file->ucontext);
3105         if (!uobj)
3106                 return -EINVAL;
3107         ah = uobj->object;
3108
3109         ret = ib_destroy_ah(ah);
3110         if (!ret)
3111                 uobj->live = 0;
3112
3113         put_uobj_write(uobj);
3114
3115         if (ret)
3116                 return ret;
3117
3118         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
3119
3120         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
3121
3122         mutex_lock(&file->mutex);
3123         list_del(&uobj->list);
3124         mutex_unlock(&file->mutex);
3125
3126         put_uobj(uobj);
3127
3128         return in_len;
3129 }
3130
3131 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
3132                                struct ib_device *ib_dev,
3133                                const char __user *buf, int in_len,
3134                                int out_len)
3135 {
3136         struct ib_uverbs_attach_mcast cmd;
3137         struct ib_qp                 *qp;
3138         struct ib_uqp_object         *obj;
3139         struct ib_uverbs_mcast_entry *mcast;
3140         int                           ret;
3141
3142         if (copy_from_user(&cmd, buf, sizeof cmd))
3143                 return -EFAULT;
3144
3145         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
3146         if (!qp)
3147                 return -EINVAL;
3148
3149         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
3150
3151         list_for_each_entry(mcast, &obj->mcast_list, list)
3152                 if (cmd.mlid == mcast->lid &&
3153                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
3154                         ret = 0;
3155                         goto out_put;
3156                 }
3157
3158         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
3159         if (!mcast) {
3160                 ret = -ENOMEM;
3161                 goto out_put;
3162         }
3163
3164         mcast->lid = cmd.mlid;
3165         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
3166
3167         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
3168         if (!ret)
3169                 list_add_tail(&mcast->list, &obj->mcast_list);
3170         else
3171                 kfree(mcast);
3172
3173 out_put:
3174         put_qp_write(qp);
3175
3176         return ret ? ret : in_len;
3177 }
3178
3179 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
3180                                struct ib_device *ib_dev,
3181                                const char __user *buf, int in_len,
3182                                int out_len)
3183 {
3184         struct ib_uverbs_detach_mcast cmd;
3185         struct ib_uqp_object         *obj;
3186         struct ib_qp                 *qp;
3187         struct ib_uverbs_mcast_entry *mcast;
3188         int                           ret = -EINVAL;
3189
3190         if (copy_from_user(&cmd, buf, sizeof cmd))
3191                 return -EFAULT;
3192
3193         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
3194         if (!qp)
3195                 return -EINVAL;
3196
3197         ret = ib_detach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
3198         if (ret)
3199                 goto out_put;
3200
3201         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
3202
3203         list_for_each_entry(mcast, &obj->mcast_list, list)
3204                 if (cmd.mlid == mcast->lid &&
3205                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
3206                         list_del(&mcast->list);
3207                         kfree(mcast);
3208                         break;
3209                 }
3210
3211 out_put:
3212         put_qp_write(qp);
3213
3214         return ret ? ret : in_len;
3215 }
3216
3217 static int kern_spec_to_ib_spec_action(struct ib_uverbs_flow_spec *kern_spec,
3218                                        union ib_flow_spec *ib_spec)
3219 {
3220         ib_spec->type = kern_spec->type;
3221         switch (ib_spec->type) {
3222         case IB_FLOW_SPEC_ACTION_TAG:
3223                 if (kern_spec->flow_tag.size !=
3224                     sizeof(struct ib_uverbs_flow_spec_action_tag))
3225                         return -EINVAL;
3226
3227                 ib_spec->flow_tag.size = sizeof(struct ib_flow_spec_action_tag);
3228                 ib_spec->flow_tag.tag_id = kern_spec->flow_tag.tag_id;
3229                 break;
3230         default:
3231                 return -EINVAL;
3232         }
3233         return 0;
3234 }
3235
3236 static size_t kern_spec_filter_sz(struct ib_uverbs_flow_spec_hdr *spec)
3237 {
3238         /* Returns user space filter size, includes padding */
3239         return (spec->size - sizeof(struct ib_uverbs_flow_spec_hdr)) / 2;
3240 }
3241
3242 static ssize_t spec_filter_size(void *kern_spec_filter, u16 kern_filter_size,
3243                                 u16 ib_real_filter_sz)
3244 {
3245         /*
3246          * User space filter structures must be 64 bit aligned, otherwise this
3247          * may pass, but we won't handle additional new attributes.
3248          */
3249
3250         if (kern_filter_size > ib_real_filter_sz) {
3251                 if (memchr_inv(kern_spec_filter +
3252                                ib_real_filter_sz, 0,
3253                                kern_filter_size - ib_real_filter_sz))
3254                         return -EINVAL;
3255                 return ib_real_filter_sz;
3256         }
3257         return kern_filter_size;
3258 }
3259
3260 static int kern_spec_to_ib_spec_filter(struct ib_uverbs_flow_spec *kern_spec,
3261                                        union ib_flow_spec *ib_spec)
3262 {
3263         ssize_t actual_filter_sz;
3264         ssize_t kern_filter_sz;
3265         ssize_t ib_filter_sz;
3266         void *kern_spec_mask;
3267         void *kern_spec_val;
3268
3269         if (kern_spec->reserved)
3270                 return -EINVAL;
3271
3272         ib_spec->type = kern_spec->type;
3273
3274         kern_filter_sz = kern_spec_filter_sz(&kern_spec->hdr);
3275         /* User flow spec size must be aligned to 4 bytes */
3276         if (kern_filter_sz != ALIGN(kern_filter_sz, 4))
3277                 return -EINVAL;
3278
3279         kern_spec_val = (void *)kern_spec +
3280                 sizeof(struct ib_uverbs_flow_spec_hdr);
3281         kern_spec_mask = kern_spec_val + kern_filter_sz;
3282         if (ib_spec->type == (IB_FLOW_SPEC_INNER | IB_FLOW_SPEC_VXLAN_TUNNEL))
3283                 return -EINVAL;
3284
3285         switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
3286         case IB_FLOW_SPEC_ETH:
3287                 ib_filter_sz = offsetof(struct ib_flow_eth_filter, real_sz);
3288                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3289                                                     kern_filter_sz,
3290                                                     ib_filter_sz);
3291                 if (actual_filter_sz <= 0)
3292                         return -EINVAL;
3293                 ib_spec->size = sizeof(struct ib_flow_spec_eth);
3294                 memcpy(&ib_spec->eth.val, kern_spec_val, actual_filter_sz);
3295                 memcpy(&ib_spec->eth.mask, kern_spec_mask, actual_filter_sz);
3296                 break;
3297         case IB_FLOW_SPEC_IPV4:
3298                 ib_filter_sz = offsetof(struct ib_flow_ipv4_filter, real_sz);
3299                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3300                                                     kern_filter_sz,
3301                                                     ib_filter_sz);
3302                 if (actual_filter_sz <= 0)
3303                         return -EINVAL;
3304                 ib_spec->size = sizeof(struct ib_flow_spec_ipv4);
3305                 memcpy(&ib_spec->ipv4.val, kern_spec_val, actual_filter_sz);
3306                 memcpy(&ib_spec->ipv4.mask, kern_spec_mask, actual_filter_sz);
3307                 break;
3308         case IB_FLOW_SPEC_IPV6:
3309                 ib_filter_sz = offsetof(struct ib_flow_ipv6_filter, real_sz);
3310                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3311                                                     kern_filter_sz,
3312                                                     ib_filter_sz);
3313                 if (actual_filter_sz <= 0)
3314                         return -EINVAL;
3315                 ib_spec->size = sizeof(struct ib_flow_spec_ipv6);
3316                 memcpy(&ib_spec->ipv6.val, kern_spec_val, actual_filter_sz);
3317                 memcpy(&ib_spec->ipv6.mask, kern_spec_mask, actual_filter_sz);
3318
3319                 if ((ntohl(ib_spec->ipv6.mask.flow_label)) >= BIT(20) ||
3320                     (ntohl(ib_spec->ipv6.val.flow_label)) >= BIT(20))
3321                         return -EINVAL;
3322                 break;
3323         case IB_FLOW_SPEC_TCP:
3324         case IB_FLOW_SPEC_UDP:
3325                 ib_filter_sz = offsetof(struct ib_flow_tcp_udp_filter, real_sz);
3326                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3327                                                     kern_filter_sz,
3328                                                     ib_filter_sz);
3329                 if (actual_filter_sz <= 0)
3330                         return -EINVAL;
3331                 ib_spec->size = sizeof(struct ib_flow_spec_tcp_udp);
3332                 memcpy(&ib_spec->tcp_udp.val, kern_spec_val, actual_filter_sz);
3333                 memcpy(&ib_spec->tcp_udp.mask, kern_spec_mask, actual_filter_sz);
3334                 break;
3335         case IB_FLOW_SPEC_VXLAN_TUNNEL:
3336                 ib_filter_sz = offsetof(struct ib_flow_tunnel_filter, real_sz);
3337                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3338                                                     kern_filter_sz,
3339                                                     ib_filter_sz);
3340                 if (actual_filter_sz <= 0)
3341                         return -EINVAL;
3342                 ib_spec->tunnel.size = sizeof(struct ib_flow_spec_tunnel);
3343                 memcpy(&ib_spec->tunnel.val, kern_spec_val, actual_filter_sz);
3344                 memcpy(&ib_spec->tunnel.mask, kern_spec_mask, actual_filter_sz);
3345
3346                 if ((ntohl(ib_spec->tunnel.mask.tunnel_id)) >= BIT(24) ||
3347                     (ntohl(ib_spec->tunnel.val.tunnel_id)) >= BIT(24))
3348                         return -EINVAL;
3349                 break;
3350         default:
3351                 return -EINVAL;
3352         }
3353         return 0;
3354 }
3355
3356 static int kern_spec_to_ib_spec(struct ib_uverbs_flow_spec *kern_spec,
3357                                 union ib_flow_spec *ib_spec)
3358 {
3359         if (kern_spec->reserved)
3360                 return -EINVAL;
3361
3362         if (kern_spec->type >= IB_FLOW_SPEC_ACTION_TAG)
3363                 return kern_spec_to_ib_spec_action(kern_spec, ib_spec);
3364         else
3365                 return kern_spec_to_ib_spec_filter(kern_spec, ib_spec);
3366 }
3367
3368 int ib_uverbs_ex_create_wq(struct ib_uverbs_file *file,
3369                            struct ib_device *ib_dev,
3370                            struct ib_udata *ucore,
3371                            struct ib_udata *uhw)
3372 {
3373         struct ib_uverbs_ex_create_wq     cmd = {};
3374         struct ib_uverbs_ex_create_wq_resp resp = {};
3375         struct ib_uwq_object           *obj;
3376         int err = 0;
3377         struct ib_cq *cq;
3378         struct ib_pd *pd;
3379         struct ib_wq *wq;
3380         struct ib_wq_init_attr wq_init_attr = {};
3381         size_t required_cmd_sz;
3382         size_t required_resp_len;
3383
3384         required_cmd_sz = offsetof(typeof(cmd), max_sge) + sizeof(cmd.max_sge);
3385         required_resp_len = offsetof(typeof(resp), wqn) + sizeof(resp.wqn);
3386
3387         if (ucore->inlen < required_cmd_sz)
3388                 return -EINVAL;
3389
3390         if (ucore->outlen < required_resp_len)
3391                 return -ENOSPC;
3392
3393         if (ucore->inlen > sizeof(cmd) &&
3394             !ib_is_udata_cleared(ucore, sizeof(cmd),
3395                                  ucore->inlen - sizeof(cmd)))
3396                 return -EOPNOTSUPP;
3397
3398         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3399         if (err)
3400                 return err;
3401
3402         if (cmd.comp_mask)
3403                 return -EOPNOTSUPP;
3404
3405         obj = kmalloc(sizeof(*obj), GFP_KERNEL);
3406         if (!obj)
3407                 return -ENOMEM;
3408
3409         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext,
3410                   &wq_lock_class);
3411         down_write(&obj->uevent.uobject.mutex);
3412         pd  = idr_read_pd(cmd.pd_handle, file->ucontext);
3413         if (!pd) {
3414                 err = -EINVAL;
3415                 goto err_uobj;
3416         }
3417
3418         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
3419         if (!cq) {
3420                 err = -EINVAL;
3421                 goto err_put_pd;
3422         }
3423
3424         wq_init_attr.cq = cq;
3425         wq_init_attr.max_sge = cmd.max_sge;
3426         wq_init_attr.max_wr = cmd.max_wr;
3427         wq_init_attr.wq_context = file;
3428         wq_init_attr.wq_type = cmd.wq_type;
3429         wq_init_attr.event_handler = ib_uverbs_wq_event_handler;
3430         if (ucore->inlen >= (offsetof(typeof(cmd), create_flags) +
3431                              sizeof(cmd.create_flags)))
3432                 wq_init_attr.create_flags = cmd.create_flags;
3433         obj->uevent.events_reported = 0;
3434         INIT_LIST_HEAD(&obj->uevent.event_list);
3435         wq = pd->device->create_wq(pd, &wq_init_attr, uhw);
3436         if (IS_ERR(wq)) {
3437                 err = PTR_ERR(wq);
3438                 goto err_put_cq;
3439         }
3440
3441         wq->uobject = &obj->uevent.uobject;
3442         obj->uevent.uobject.object = wq;
3443         wq->wq_type = wq_init_attr.wq_type;
3444         wq->cq = cq;
3445         wq->pd = pd;
3446         wq->device = pd->device;
3447         wq->wq_context = wq_init_attr.wq_context;
3448         atomic_set(&wq->usecnt, 0);
3449         atomic_inc(&pd->usecnt);
3450         atomic_inc(&cq->usecnt);
3451         wq->uobject = &obj->uevent.uobject;
3452         obj->uevent.uobject.object = wq;
3453         err = idr_add_uobj(&ib_uverbs_wq_idr, &obj->uevent.uobject);
3454         if (err)
3455                 goto destroy_wq;
3456
3457         memset(&resp, 0, sizeof(resp));
3458         resp.wq_handle = obj->uevent.uobject.id;
3459         resp.max_sge = wq_init_attr.max_sge;
3460         resp.max_wr = wq_init_attr.max_wr;
3461         resp.wqn = wq->wq_num;
3462         resp.response_length = required_resp_len;
3463         err = ib_copy_to_udata(ucore,
3464                                &resp, resp.response_length);
3465         if (err)
3466                 goto err_copy;
3467
3468         put_pd_read(pd);
3469         put_cq_read(cq);
3470
3471         mutex_lock(&file->mutex);
3472         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->wq_list);
3473         mutex_unlock(&file->mutex);
3474
3475         obj->uevent.uobject.live = 1;
3476         up_write(&obj->uevent.uobject.mutex);
3477         return 0;
3478
3479 err_copy:
3480         idr_remove_uobj(&ib_uverbs_wq_idr, &obj->uevent.uobject);
3481 destroy_wq:
3482         ib_destroy_wq(wq);
3483 err_put_cq:
3484         put_cq_read(cq);
3485 err_put_pd:
3486         put_pd_read(pd);
3487 err_uobj:
3488         put_uobj_write(&obj->uevent.uobject);
3489
3490         return err;
3491 }
3492
3493 int ib_uverbs_ex_destroy_wq(struct ib_uverbs_file *file,
3494                             struct ib_device *ib_dev,
3495                             struct ib_udata *ucore,
3496                             struct ib_udata *uhw)
3497 {
3498         struct ib_uverbs_ex_destroy_wq  cmd = {};
3499         struct ib_uverbs_ex_destroy_wq_resp     resp = {};
3500         struct ib_wq                    *wq;
3501         struct ib_uobject               *uobj;
3502         struct ib_uwq_object            *obj;
3503         size_t required_cmd_sz;
3504         size_t required_resp_len;
3505         int                             ret;
3506
3507         required_cmd_sz = offsetof(typeof(cmd), wq_handle) + sizeof(cmd.wq_handle);
3508         required_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
3509
3510         if (ucore->inlen < required_cmd_sz)
3511                 return -EINVAL;
3512
3513         if (ucore->outlen < required_resp_len)
3514                 return -ENOSPC;
3515
3516         if (ucore->inlen > sizeof(cmd) &&
3517             !ib_is_udata_cleared(ucore, sizeof(cmd),
3518                                  ucore->inlen - sizeof(cmd)))
3519                 return -EOPNOTSUPP;
3520
3521         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3522         if (ret)
3523                 return ret;
3524
3525         if (cmd.comp_mask)
3526                 return -EOPNOTSUPP;
3527
3528         resp.response_length = required_resp_len;
3529         uobj = idr_write_uobj(&ib_uverbs_wq_idr, cmd.wq_handle,
3530                               file->ucontext);
3531         if (!uobj)
3532                 return -EINVAL;
3533
3534         wq = uobj->object;
3535         obj = container_of(uobj, struct ib_uwq_object, uevent.uobject);
3536         ret = ib_destroy_wq(wq);
3537         if (!ret)
3538                 uobj->live = 0;
3539
3540         put_uobj_write(uobj);
3541         if (ret)
3542                 return ret;
3543
3544         idr_remove_uobj(&ib_uverbs_wq_idr, uobj);
3545
3546         mutex_lock(&file->mutex);
3547         list_del(&uobj->list);
3548         mutex_unlock(&file->mutex);
3549
3550         ib_uverbs_release_uevent(file, &obj->uevent);
3551         resp.events_reported = obj->uevent.events_reported;
3552         put_uobj(uobj);
3553
3554         ret = ib_copy_to_udata(ucore, &resp, resp.response_length);
3555         if (ret)
3556                 return ret;
3557
3558         return 0;
3559 }
3560
3561 int ib_uverbs_ex_modify_wq(struct ib_uverbs_file *file,
3562                            struct ib_device *ib_dev,
3563                            struct ib_udata *ucore,
3564                            struct ib_udata *uhw)
3565 {
3566         struct ib_uverbs_ex_modify_wq cmd = {};
3567         struct ib_wq *wq;
3568         struct ib_wq_attr wq_attr = {};
3569         size_t required_cmd_sz;
3570         int ret;
3571
3572         required_cmd_sz = offsetof(typeof(cmd), curr_wq_state) + sizeof(cmd.curr_wq_state);
3573         if (ucore->inlen < required_cmd_sz)
3574                 return -EINVAL;
3575
3576         if (ucore->inlen > sizeof(cmd) &&
3577             !ib_is_udata_cleared(ucore, sizeof(cmd),
3578                                  ucore->inlen - sizeof(cmd)))
3579                 return -EOPNOTSUPP;
3580
3581         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3582         if (ret)
3583                 return ret;
3584
3585         if (!cmd.attr_mask)
3586                 return -EINVAL;
3587
3588         if (cmd.attr_mask > (IB_WQ_STATE | IB_WQ_CUR_STATE | IB_WQ_FLAGS))
3589                 return -EINVAL;
3590
3591         wq = idr_read_wq(cmd.wq_handle, file->ucontext);
3592         if (!wq)
3593                 return -EINVAL;
3594
3595         wq_attr.curr_wq_state = cmd.curr_wq_state;
3596         wq_attr.wq_state = cmd.wq_state;
3597         if (cmd.attr_mask & IB_WQ_FLAGS) {
3598                 wq_attr.flags = cmd.flags;
3599                 wq_attr.flags_mask = cmd.flags_mask;
3600         }
3601         ret = wq->device->modify_wq(wq, &wq_attr, cmd.attr_mask, uhw);
3602         put_wq_read(wq);
3603         return ret;
3604 }
3605
3606 int ib_uverbs_ex_create_rwq_ind_table(struct ib_uverbs_file *file,
3607                                       struct ib_device *ib_dev,
3608                                       struct ib_udata *ucore,
3609                                       struct ib_udata *uhw)
3610 {
3611         struct ib_uverbs_ex_create_rwq_ind_table          cmd = {};
3612         struct ib_uverbs_ex_create_rwq_ind_table_resp  resp = {};
3613         struct ib_uobject                 *uobj;
3614         int err = 0;
3615         struct ib_rwq_ind_table_init_attr init_attr = {};
3616         struct ib_rwq_ind_table *rwq_ind_tbl;
3617         struct ib_wq    **wqs = NULL;
3618         u32 *wqs_handles = NULL;
3619         struct ib_wq    *wq = NULL;
3620         int i, j, num_read_wqs;
3621         u32 num_wq_handles;
3622         u32 expected_in_size;
3623         size_t required_cmd_sz_header;
3624         size_t required_resp_len;
3625
3626         required_cmd_sz_header = offsetof(typeof(cmd), log_ind_tbl_size) + sizeof(cmd.log_ind_tbl_size);
3627         required_resp_len = offsetof(typeof(resp), ind_tbl_num) + sizeof(resp.ind_tbl_num);
3628
3629         if (ucore->inlen < required_cmd_sz_header)
3630                 return -EINVAL;
3631
3632         if (ucore->outlen < required_resp_len)
3633                 return -ENOSPC;
3634
3635         err = ib_copy_from_udata(&cmd, ucore, required_cmd_sz_header);
3636         if (err)
3637                 return err;
3638
3639         ucore->inbuf += required_cmd_sz_header;
3640         ucore->inlen -= required_cmd_sz_header;
3641
3642         if (cmd.comp_mask)
3643                 return -EOPNOTSUPP;
3644
3645         if (cmd.log_ind_tbl_size > IB_USER_VERBS_MAX_LOG_IND_TBL_SIZE)
3646                 return -EINVAL;
3647
3648         num_wq_handles = 1 << cmd.log_ind_tbl_size;
3649         expected_in_size = num_wq_handles * sizeof(__u32);
3650         if (num_wq_handles == 1)
3651                 /* input size for wq handles is u64 aligned */
3652                 expected_in_size += sizeof(__u32);
3653
3654         if (ucore->inlen < expected_in_size)
3655                 return -EINVAL;
3656
3657         if (ucore->inlen > expected_in_size &&
3658             !ib_is_udata_cleared(ucore, expected_in_size,
3659                                  ucore->inlen - expected_in_size))
3660                 return -EOPNOTSUPP;
3661
3662         wqs_handles = kcalloc(num_wq_handles, sizeof(*wqs_handles),
3663                               GFP_KERNEL);
3664         if (!wqs_handles)
3665                 return -ENOMEM;
3666
3667         err = ib_copy_from_udata(wqs_handles, ucore,
3668                                  num_wq_handles * sizeof(__u32));
3669         if (err)
3670                 goto err_free;
3671
3672         wqs = kcalloc(num_wq_handles, sizeof(*wqs), GFP_KERNEL);
3673         if (!wqs) {
3674                 err = -ENOMEM;
3675                 goto  err_free;
3676         }
3677
3678         for (num_read_wqs = 0; num_read_wqs < num_wq_handles;
3679                         num_read_wqs++) {
3680                 wq = idr_read_wq(wqs_handles[num_read_wqs], file->ucontext);
3681                 if (!wq) {
3682                         err = -EINVAL;
3683                         goto put_wqs;
3684                 }
3685
3686                 wqs[num_read_wqs] = wq;
3687         }
3688
3689         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
3690         if (!uobj) {
3691                 err = -ENOMEM;
3692                 goto put_wqs;
3693         }
3694
3695         init_uobj(uobj, 0, file->ucontext, &rwq_ind_table_lock_class);
3696         down_write(&uobj->mutex);
3697         init_attr.log_ind_tbl_size = cmd.log_ind_tbl_size;
3698         init_attr.ind_tbl = wqs;
3699         rwq_ind_tbl = ib_dev->create_rwq_ind_table(ib_dev, &init_attr, uhw);
3700
3701         if (IS_ERR(rwq_ind_tbl)) {
3702                 err = PTR_ERR(rwq_ind_tbl);
3703                 goto err_uobj;
3704         }
3705
3706         rwq_ind_tbl->ind_tbl = wqs;
3707         rwq_ind_tbl->log_ind_tbl_size = init_attr.log_ind_tbl_size;
3708         rwq_ind_tbl->uobject = uobj;
3709         uobj->object = rwq_ind_tbl;
3710         rwq_ind_tbl->device = ib_dev;
3711         atomic_set(&rwq_ind_tbl->usecnt, 0);
3712
3713         for (i = 0; i < num_wq_handles; i++)
3714                 atomic_inc(&wqs[i]->usecnt);
3715
3716         err = idr_add_uobj(&ib_uverbs_rwq_ind_tbl_idr, uobj);
3717         if (err)
3718                 goto destroy_ind_tbl;
3719
3720         resp.ind_tbl_handle = uobj->id;
3721         resp.ind_tbl_num = rwq_ind_tbl->ind_tbl_num;
3722         resp.response_length = required_resp_len;
3723
3724         err = ib_copy_to_udata(ucore,
3725                                &resp, resp.response_length);
3726         if (err)
3727                 goto err_copy;
3728
3729         kfree(wqs_handles);
3730
3731         for (j = 0; j < num_read_wqs; j++)
3732                 put_wq_read(wqs[j]);
3733
3734         mutex_lock(&file->mutex);
3735         list_add_tail(&uobj->list, &file->ucontext->rwq_ind_tbl_list);
3736         mutex_unlock(&file->mutex);
3737
3738         uobj->live = 1;
3739
3740         up_write(&uobj->mutex);
3741         return 0;
3742
3743 err_copy:
3744         idr_remove_uobj(&ib_uverbs_rwq_ind_tbl_idr, uobj);
3745 destroy_ind_tbl:
3746         ib_destroy_rwq_ind_table(rwq_ind_tbl);
3747 err_uobj:
3748         put_uobj_write(uobj);
3749 put_wqs:
3750         for (j = 0; j < num_read_wqs; j++)
3751                 put_wq_read(wqs[j]);
3752 err_free:
3753         kfree(wqs_handles);
3754         kfree(wqs);
3755         return err;
3756 }
3757
3758 int ib_uverbs_ex_destroy_rwq_ind_table(struct ib_uverbs_file *file,
3759                                        struct ib_device *ib_dev,
3760                                        struct ib_udata *ucore,
3761                                        struct ib_udata *uhw)
3762 {
3763         struct ib_uverbs_ex_destroy_rwq_ind_table       cmd = {};
3764         struct ib_rwq_ind_table *rwq_ind_tbl;
3765         struct ib_uobject               *uobj;
3766         int                     ret;
3767         struct ib_wq    **ind_tbl;
3768         size_t required_cmd_sz;
3769
3770         required_cmd_sz = offsetof(typeof(cmd), ind_tbl_handle) + sizeof(cmd.ind_tbl_handle);
3771
3772         if (ucore->inlen < required_cmd_sz)
3773                 return -EINVAL;
3774
3775         if (ucore->inlen > sizeof(cmd) &&
3776             !ib_is_udata_cleared(ucore, sizeof(cmd),
3777                                  ucore->inlen - sizeof(cmd)))
3778                 return -EOPNOTSUPP;
3779
3780         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3781         if (ret)
3782                 return ret;
3783
3784         if (cmd.comp_mask)
3785                 return -EOPNOTSUPP;
3786
3787         uobj = idr_write_uobj(&ib_uverbs_rwq_ind_tbl_idr, cmd.ind_tbl_handle,
3788                               file->ucontext);
3789         if (!uobj)
3790                 return -EINVAL;
3791         rwq_ind_tbl = uobj->object;
3792         ind_tbl = rwq_ind_tbl->ind_tbl;
3793
3794         ret = ib_destroy_rwq_ind_table(rwq_ind_tbl);
3795         if (!ret)
3796                 uobj->live = 0;
3797
3798         put_uobj_write(uobj);
3799
3800         if (ret)
3801                 return ret;
3802
3803         idr_remove_uobj(&ib_uverbs_rwq_ind_tbl_idr, uobj);
3804
3805         mutex_lock(&file->mutex);
3806         list_del(&uobj->list);
3807         mutex_unlock(&file->mutex);
3808
3809         put_uobj(uobj);
3810         kfree(ind_tbl);
3811         return ret;
3812 }
3813
3814 int ib_uverbs_ex_create_flow(struct ib_uverbs_file *file,
3815                              struct ib_device *ib_dev,
3816                              struct ib_udata *ucore,
3817                              struct ib_udata *uhw)
3818 {
3819         struct ib_uverbs_create_flow      cmd;
3820         struct ib_uverbs_create_flow_resp resp;
3821         struct ib_uobject                 *uobj;
3822         struct ib_flow                    *flow_id;
3823         struct ib_uverbs_flow_attr        *kern_flow_attr;
3824         struct ib_flow_attr               *flow_attr;
3825         struct ib_qp                      *qp;
3826         int err = 0;
3827         void *kern_spec;
3828         void *ib_spec;
3829         int i;
3830
3831         if (ucore->inlen < sizeof(cmd))
3832                 return -EINVAL;
3833
3834         if (ucore->outlen < sizeof(resp))
3835                 return -ENOSPC;
3836
3837         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3838         if (err)
3839                 return err;
3840
3841         ucore->inbuf += sizeof(cmd);
3842         ucore->inlen -= sizeof(cmd);
3843
3844         if (cmd.comp_mask)
3845                 return -EINVAL;
3846
3847         if (!capable(CAP_NET_RAW))
3848                 return -EPERM;
3849
3850         if (cmd.flow_attr.flags >= IB_FLOW_ATTR_FLAGS_RESERVED)
3851                 return -EINVAL;
3852
3853         if ((cmd.flow_attr.flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) &&
3854             ((cmd.flow_attr.type == IB_FLOW_ATTR_ALL_DEFAULT) ||
3855              (cmd.flow_attr.type == IB_FLOW_ATTR_MC_DEFAULT)))
3856                 return -EINVAL;
3857
3858         if (cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
3859                 return -EINVAL;
3860
3861         if (cmd.flow_attr.size > ucore->inlen ||
3862             cmd.flow_attr.size >
3863             (cmd.flow_attr.num_of_specs * sizeof(struct ib_uverbs_flow_spec)))
3864                 return -EINVAL;
3865
3866         if (cmd.flow_attr.reserved[0] ||
3867             cmd.flow_attr.reserved[1])
3868                 return -EINVAL;
3869
3870         if (cmd.flow_attr.num_of_specs) {
3871                 kern_flow_attr = kmalloc(sizeof(*kern_flow_attr) + cmd.flow_attr.size,
3872                                          GFP_KERNEL);
3873                 if (!kern_flow_attr)
3874                         return -ENOMEM;
3875
3876                 memcpy(kern_flow_attr, &cmd.flow_attr, sizeof(*kern_flow_attr));
3877                 err = ib_copy_from_udata(kern_flow_attr + 1, ucore,
3878                                          cmd.flow_attr.size);
3879                 if (err)
3880                         goto err_free_attr;
3881         } else {
3882                 kern_flow_attr = &cmd.flow_attr;
3883         }
3884
3885         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
3886         if (!uobj) {
3887                 err = -ENOMEM;
3888                 goto err_free_attr;
3889         }
3890         init_uobj(uobj, 0, file->ucontext, &rule_lock_class);
3891         down_write(&uobj->mutex);
3892
3893         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
3894         if (!qp) {
3895                 err = -EINVAL;
3896                 goto err_uobj;
3897         }
3898
3899         flow_attr = kzalloc(sizeof(*flow_attr) + cmd.flow_attr.num_of_specs *
3900                             sizeof(union ib_flow_spec), GFP_KERNEL);
3901         if (!flow_attr) {
3902                 err = -ENOMEM;
3903                 goto err_put;
3904         }
3905
3906         flow_attr->type = kern_flow_attr->type;
3907         flow_attr->priority = kern_flow_attr->priority;
3908         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
3909         flow_attr->port = kern_flow_attr->port;
3910         flow_attr->flags = kern_flow_attr->flags;
3911         flow_attr->size = sizeof(*flow_attr);
3912
3913         kern_spec = kern_flow_attr + 1;
3914         ib_spec = flow_attr + 1;
3915         for (i = 0; i < flow_attr->num_of_specs &&
3916              cmd.flow_attr.size > offsetof(struct ib_uverbs_flow_spec, reserved) &&
3917              cmd.flow_attr.size >=
3918              ((struct ib_uverbs_flow_spec *)kern_spec)->size; i++) {
3919                 err = kern_spec_to_ib_spec(kern_spec, ib_spec);
3920                 if (err)
3921                         goto err_free;
3922                 flow_attr->size +=
3923                         ((union ib_flow_spec *) ib_spec)->size;
3924                 cmd.flow_attr.size -= ((struct ib_uverbs_flow_spec *)kern_spec)->size;
3925                 kern_spec += ((struct ib_uverbs_flow_spec *) kern_spec)->size;
3926                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
3927         }
3928         if (cmd.flow_attr.size || (i != flow_attr->num_of_specs)) {
3929                 pr_warn("create flow failed, flow %d: %d bytes left from uverb cmd\n",
3930                         i, cmd.flow_attr.size);
3931                 err = -EINVAL;
3932                 goto err_free;
3933         }
3934
3935         err = ib_rdmacg_try_charge(&uobj->cg_obj, ib_dev,
3936                                    RDMACG_RESOURCE_HCA_OBJECT);
3937         if (err)
3938                 goto err_free;
3939
3940         flow_id = ib_create_flow(qp, flow_attr, IB_FLOW_DOMAIN_USER);
3941         if (IS_ERR(flow_id)) {
3942                 err = PTR_ERR(flow_id);
3943                 goto err_create;
3944         }
3945         flow_id->uobject = uobj;
3946         uobj->object = flow_id;
3947
3948         err = idr_add_uobj(&ib_uverbs_rule_idr, uobj);
3949         if (err)
3950                 goto destroy_flow;
3951
3952         memset(&resp, 0, sizeof(resp));
3953         resp.flow_handle = uobj->id;
3954
3955         err = ib_copy_to_udata(ucore,
3956                                &resp, sizeof(resp));
3957         if (err)
3958                 goto err_copy;
3959
3960         put_qp_read(qp);
3961         mutex_lock(&file->mutex);
3962         list_add_tail(&uobj->list, &file->ucontext->rule_list);
3963         mutex_unlock(&file->mutex);
3964
3965         uobj->live = 1;
3966
3967         up_write(&uobj->mutex);
3968         kfree(flow_attr);
3969         if (cmd.flow_attr.num_of_specs)
3970                 kfree(kern_flow_attr);
3971         return 0;
3972 err_copy:
3973         idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
3974 destroy_flow:
3975         ib_destroy_flow(flow_id);
3976 err_create:
3977         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
3978 err_free:
3979         kfree(flow_attr);
3980 err_put:
3981         put_qp_read(qp);
3982 err_uobj:
3983         put_uobj_write(uobj);
3984 err_free_attr:
3985         if (cmd.flow_attr.num_of_specs)
3986                 kfree(kern_flow_attr);
3987         return err;
3988 }
3989
3990 int ib_uverbs_ex_destroy_flow(struct ib_uverbs_file *file,
3991                               struct ib_device *ib_dev,
3992                               struct ib_udata *ucore,
3993                               struct ib_udata *uhw)
3994 {
3995         struct ib_uverbs_destroy_flow   cmd;
3996         struct ib_flow                  *flow_id;
3997         struct ib_uobject               *uobj;
3998         int                             ret;
3999
4000         if (ucore->inlen < sizeof(cmd))
4001                 return -EINVAL;
4002
4003         ret = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
4004         if (ret)
4005                 return ret;
4006
4007         if (cmd.comp_mask)
4008                 return -EINVAL;
4009
4010         uobj = idr_write_uobj(&ib_uverbs_rule_idr, cmd.flow_handle,
4011                               file->ucontext);
4012         if (!uobj)
4013                 return -EINVAL;
4014         flow_id = uobj->object;
4015
4016         ret = ib_destroy_flow(flow_id);
4017         if (!ret) {
4018                 ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev,
4019                                    RDMACG_RESOURCE_HCA_OBJECT);
4020                 uobj->live = 0;
4021         }
4022
4023         put_uobj_write(uobj);
4024
4025         idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
4026
4027         mutex_lock(&file->mutex);
4028         list_del(&uobj->list);
4029         mutex_unlock(&file->mutex);
4030
4031         put_uobj(uobj);
4032
4033         return ret;
4034 }
4035
4036 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
4037                                 struct ib_device *ib_dev,
4038                                 struct ib_uverbs_create_xsrq *cmd,
4039                                 struct ib_udata *udata)
4040 {
4041         struct ib_uverbs_create_srq_resp resp;
4042         struct ib_usrq_object           *obj;
4043         struct ib_pd                    *pd;
4044         struct ib_srq                   *srq;
4045         struct ib_uobject               *uninitialized_var(xrcd_uobj);
4046         struct ib_srq_init_attr          attr;
4047         int ret;
4048
4049         obj = kmalloc(sizeof *obj, GFP_KERNEL);
4050         if (!obj)
4051                 return -ENOMEM;
4052
4053         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext, &srq_lock_class);
4054         down_write(&obj->uevent.uobject.mutex);
4055
4056         if (cmd->srq_type == IB_SRQT_XRC) {
4057                 attr.ext.xrc.xrcd  = idr_read_xrcd(cmd->xrcd_handle, file->ucontext, &xrcd_uobj);
4058                 if (!attr.ext.xrc.xrcd) {
4059                         ret = -EINVAL;
4060                         goto err;
4061                 }
4062
4063                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
4064                 atomic_inc(&obj->uxrcd->refcnt);
4065
4066                 attr.ext.xrc.cq  = idr_read_cq(cmd->cq_handle, file->ucontext, 0);
4067                 if (!attr.ext.xrc.cq) {
4068                         ret = -EINVAL;
4069                         goto err_put_xrcd;
4070                 }
4071         }
4072
4073         pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
4074         if (!pd) {
4075                 ret = -EINVAL;
4076                 goto err_put_cq;
4077         }
4078
4079         attr.event_handler  = ib_uverbs_srq_event_handler;
4080         attr.srq_context    = file;
4081         attr.srq_type       = cmd->srq_type;
4082         attr.attr.max_wr    = cmd->max_wr;
4083         attr.attr.max_sge   = cmd->max_sge;
4084         attr.attr.srq_limit = cmd->srq_limit;
4085
4086         obj->uevent.events_reported = 0;
4087         INIT_LIST_HEAD(&obj->uevent.event_list);
4088
4089         ret = ib_rdmacg_try_charge(&obj->uevent.uobject.cg_obj, ib_dev,
4090                                    RDMACG_RESOURCE_HCA_OBJECT);
4091         if (ret)
4092                 goto err_put_cq;
4093
4094         srq = pd->device->create_srq(pd, &attr, udata);
4095         if (IS_ERR(srq)) {
4096                 ret = PTR_ERR(srq);
4097                 goto err_put;
4098         }
4099
4100         srq->device        = pd->device;
4101         srq->pd            = pd;
4102         srq->srq_type      = cmd->srq_type;
4103         srq->uobject       = &obj->uevent.uobject;
4104         srq->event_handler = attr.event_handler;
4105         srq->srq_context   = attr.srq_context;
4106
4107         if (cmd->srq_type == IB_SRQT_XRC) {
4108                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
4109                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
4110                 atomic_inc(&attr.ext.xrc.cq->usecnt);
4111                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
4112         }
4113
4114         atomic_inc(&pd->usecnt);
4115         atomic_set(&srq->usecnt, 0);
4116
4117         obj->uevent.uobject.object = srq;
4118         ret = idr_add_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
4119         if (ret)
4120                 goto err_destroy;
4121
4122         memset(&resp, 0, sizeof resp);
4123         resp.srq_handle = obj->uevent.uobject.id;
4124         resp.max_wr     = attr.attr.max_wr;
4125         resp.max_sge    = attr.attr.max_sge;
4126         if (cmd->srq_type == IB_SRQT_XRC)
4127                 resp.srqn = srq->ext.xrc.srq_num;
4128
4129         if (copy_to_user((void __user *) (unsigned long) cmd->response,
4130                          &resp, sizeof resp)) {
4131                 ret = -EFAULT;
4132                 goto err_copy;
4133         }
4134
4135         if (cmd->srq_type == IB_SRQT_XRC) {
4136                 put_uobj_read(xrcd_uobj);
4137                 put_cq_read(attr.ext.xrc.cq);
4138         }
4139         put_pd_read(pd);
4140
4141         mutex_lock(&file->mutex);
4142         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->srq_list);
4143         mutex_unlock(&file->mutex);
4144
4145         obj->uevent.uobject.live = 1;
4146
4147         up_write(&obj->uevent.uobject.mutex);
4148
4149         return 0;
4150
4151 err_copy:
4152         idr_remove_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
4153
4154 err_destroy:
4155         ib_destroy_srq(srq);
4156
4157 err_put:
4158         ib_rdmacg_uncharge(&obj->uevent.uobject.cg_obj, ib_dev,
4159                            RDMACG_RESOURCE_HCA_OBJECT);
4160         put_pd_read(pd);
4161
4162 err_put_cq:
4163         if (cmd->srq_type == IB_SRQT_XRC)
4164                 put_cq_read(attr.ext.xrc.cq);
4165
4166 err_put_xrcd:
4167         if (cmd->srq_type == IB_SRQT_XRC) {
4168                 atomic_dec(&obj->uxrcd->refcnt);
4169                 put_uobj_read(xrcd_uobj);
4170         }
4171
4172 err:
4173         put_uobj_write(&obj->uevent.uobject);
4174         return ret;
4175 }
4176
4177 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
4178                              struct ib_device *ib_dev,
4179                              const char __user *buf, int in_len,
4180                              int out_len)
4181 {
4182         struct ib_uverbs_create_srq      cmd;
4183         struct ib_uverbs_create_xsrq     xcmd;
4184         struct ib_uverbs_create_srq_resp resp;
4185         struct ib_udata                  udata;
4186         int ret;
4187
4188         if (out_len < sizeof resp)
4189                 return -ENOSPC;
4190
4191         if (copy_from_user(&cmd, buf, sizeof cmd))
4192                 return -EFAULT;
4193
4194         xcmd.response    = cmd.response;
4195         xcmd.user_handle = cmd.user_handle;
4196         xcmd.srq_type    = IB_SRQT_BASIC;
4197         xcmd.pd_handle   = cmd.pd_handle;
4198         xcmd.max_wr      = cmd.max_wr;
4199         xcmd.max_sge     = cmd.max_sge;
4200         xcmd.srq_limit   = cmd.srq_limit;
4201
4202         INIT_UDATA(&udata, buf + sizeof cmd,
4203                    (unsigned long) cmd.response + sizeof resp,
4204                    in_len - sizeof cmd - sizeof(struct ib_uverbs_cmd_hdr),
4205                    out_len - sizeof resp);
4206
4207         ret = __uverbs_create_xsrq(file, ib_dev, &xcmd, &udata);
4208         if (ret)
4209                 return ret;
4210
4211         return in_len;
4212 }
4213
4214 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
4215                               struct ib_device *ib_dev,
4216                               const char __user *buf, int in_len, int out_len)
4217 {
4218         struct ib_uverbs_create_xsrq     cmd;
4219         struct ib_uverbs_create_srq_resp resp;
4220         struct ib_udata                  udata;
4221         int ret;
4222
4223         if (out_len < sizeof resp)
4224                 return -ENOSPC;
4225
4226         if (copy_from_user(&cmd, buf, sizeof cmd))
4227                 return -EFAULT;
4228
4229         INIT_UDATA(&udata, buf + sizeof cmd,
4230                    (unsigned long) cmd.response + sizeof resp,
4231                    in_len - sizeof cmd - sizeof(struct ib_uverbs_cmd_hdr),
4232                    out_len - sizeof resp);
4233
4234         ret = __uverbs_create_xsrq(file, ib_dev, &cmd, &udata);
4235         if (ret)
4236                 return ret;
4237
4238         return in_len;
4239 }
4240
4241 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
4242                              struct ib_device *ib_dev,
4243                              const char __user *buf, int in_len,
4244                              int out_len)
4245 {
4246         struct ib_uverbs_modify_srq cmd;
4247         struct ib_udata             udata;
4248         struct ib_srq              *srq;
4249         struct ib_srq_attr          attr;
4250         int                         ret;
4251
4252         if (copy_from_user(&cmd, buf, sizeof cmd))
4253                 return -EFAULT;
4254
4255         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
4256                    out_len);
4257
4258         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
4259         if (!srq)
4260                 return -EINVAL;
4261
4262         attr.max_wr    = cmd.max_wr;
4263         attr.srq_limit = cmd.srq_limit;
4264
4265         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
4266
4267         put_srq_read(srq);
4268
4269         return ret ? ret : in_len;
4270 }
4271
4272 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
4273                             struct ib_device *ib_dev,
4274                             const char __user *buf,
4275                             int in_len, int out_len)
4276 {
4277         struct ib_uverbs_query_srq      cmd;
4278         struct ib_uverbs_query_srq_resp resp;
4279         struct ib_srq_attr              attr;
4280         struct ib_srq                   *srq;
4281         int                             ret;
4282
4283         if (out_len < sizeof resp)
4284                 return -ENOSPC;
4285
4286         if (copy_from_user(&cmd, buf, sizeof cmd))
4287                 return -EFAULT;
4288
4289         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
4290         if (!srq)
4291                 return -EINVAL;
4292
4293         ret = ib_query_srq(srq, &attr);
4294
4295         put_srq_read(srq);
4296
4297         if (ret)
4298                 return ret;
4299
4300         memset(&resp, 0, sizeof resp);
4301
4302         resp.max_wr    = attr.max_wr;
4303         resp.max_sge   = attr.max_sge;
4304         resp.srq_limit = attr.srq_limit;
4305
4306         if (copy_to_user((void __user *) (unsigned long) cmd.response,
4307                          &resp, sizeof resp))
4308                 return -EFAULT;
4309
4310         return in_len;
4311 }
4312
4313 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
4314                               struct ib_device *ib_dev,
4315                               const char __user *buf, int in_len,
4316                               int out_len)
4317 {
4318         struct ib_uverbs_destroy_srq      cmd;
4319         struct ib_uverbs_destroy_srq_resp resp;
4320         struct ib_uobject                *uobj;
4321         struct ib_srq                    *srq;
4322         struct ib_uevent_object          *obj;
4323         int                               ret = -EINVAL;
4324         struct ib_usrq_object            *us;
4325         enum ib_srq_type                  srq_type;
4326
4327         if (copy_from_user(&cmd, buf, sizeof cmd))
4328                 return -EFAULT;
4329
4330         uobj = idr_write_uobj(&ib_uverbs_srq_idr, cmd.srq_handle, file->ucontext);
4331         if (!uobj)
4332                 return -EINVAL;
4333         srq = uobj->object;
4334         obj = container_of(uobj, struct ib_uevent_object, uobject);
4335         srq_type = srq->srq_type;
4336
4337         ret = ib_destroy_srq(srq);
4338         if (!ret)
4339                 uobj->live = 0;
4340
4341         put_uobj_write(uobj);
4342
4343         if (ret)
4344                 return ret;
4345
4346         ib_rdmacg_uncharge(&uobj->cg_obj, ib_dev, RDMACG_RESOURCE_HCA_OBJECT);
4347
4348         if (srq_type == IB_SRQT_XRC) {
4349                 us = container_of(obj, struct ib_usrq_object, uevent);
4350                 atomic_dec(&us->uxrcd->refcnt);
4351         }
4352
4353         idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
4354
4355         mutex_lock(&file->mutex);
4356         list_del(&uobj->list);
4357         mutex_unlock(&file->mutex);
4358
4359         ib_uverbs_release_uevent(file, obj);
4360
4361         memset(&resp, 0, sizeof resp);
4362         resp.events_reported = obj->events_reported;
4363
4364         put_uobj(uobj);
4365
4366         if (copy_to_user((void __user *) (unsigned long) cmd.response,
4367                          &resp, sizeof resp))
4368                 ret = -EFAULT;
4369
4370         return ret ? ret : in_len;
4371 }
4372
4373 int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
4374                               struct ib_device *ib_dev,
4375                               struct ib_udata *ucore,
4376                               struct ib_udata *uhw)
4377 {
4378         struct ib_uverbs_ex_query_device_resp resp = { {0} };
4379         struct ib_uverbs_ex_query_device  cmd;
4380         struct ib_device_attr attr = {0};
4381         int err;
4382
4383         if (ucore->inlen < sizeof(cmd))
4384                 return -EINVAL;
4385
4386         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
4387         if (err)
4388                 return err;
4389
4390         if (cmd.comp_mask)
4391                 return -EINVAL;
4392
4393         if (cmd.reserved)
4394                 return -EINVAL;
4395
4396         resp.response_length = offsetof(typeof(resp), odp_caps);
4397
4398         if (ucore->outlen < resp.response_length)
4399                 return -ENOSPC;
4400
4401         err = ib_dev->query_device(ib_dev, &attr, uhw);
4402         if (err)
4403                 return err;
4404
4405         copy_query_dev_fields(file, ib_dev, &resp.base, &attr);
4406
4407         if (ucore->outlen < resp.response_length + sizeof(resp.odp_caps))
4408                 goto end;
4409
4410 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4411         resp.odp_caps.general_caps = attr.odp_caps.general_caps;
4412         resp.odp_caps.per_transport_caps.rc_odp_caps =
4413                 attr.odp_caps.per_transport_caps.rc_odp_caps;
4414         resp.odp_caps.per_transport_caps.uc_odp_caps =
4415                 attr.odp_caps.per_transport_caps.uc_odp_caps;
4416         resp.odp_caps.per_transport_caps.ud_odp_caps =
4417                 attr.odp_caps.per_transport_caps.ud_odp_caps;
4418 #endif
4419         resp.response_length += sizeof(resp.odp_caps);
4420
4421         if (ucore->outlen < resp.response_length + sizeof(resp.timestamp_mask))
4422                 goto end;
4423
4424         resp.timestamp_mask = attr.timestamp_mask;
4425         resp.response_length += sizeof(resp.timestamp_mask);
4426
4427         if (ucore->outlen < resp.response_length + sizeof(resp.hca_core_clock))
4428                 goto end;
4429
4430         resp.hca_core_clock = attr.hca_core_clock;
4431         resp.response_length += sizeof(resp.hca_core_clock);
4432
4433         if (ucore->outlen < resp.response_length + sizeof(resp.device_cap_flags_ex))
4434                 goto end;
4435
4436         resp.device_cap_flags_ex = attr.device_cap_flags;
4437         resp.response_length += sizeof(resp.device_cap_flags_ex);
4438
4439         if (ucore->outlen < resp.response_length + sizeof(resp.rss_caps))
4440                 goto end;
4441
4442         resp.rss_caps.supported_qpts = attr.rss_caps.supported_qpts;
4443         resp.rss_caps.max_rwq_indirection_tables =
4444                 attr.rss_caps.max_rwq_indirection_tables;
4445         resp.rss_caps.max_rwq_indirection_table_size =
4446                 attr.rss_caps.max_rwq_indirection_table_size;
4447
4448         resp.response_length += sizeof(resp.rss_caps);
4449
4450         if (ucore->outlen < resp.response_length + sizeof(resp.max_wq_type_rq))
4451                 goto end;
4452
4453         resp.max_wq_type_rq = attr.max_wq_type_rq;
4454         resp.response_length += sizeof(resp.max_wq_type_rq);
4455
4456         if (ucore->outlen < resp.response_length + sizeof(resp.raw_packet_caps))
4457                 goto end;
4458
4459         resp.raw_packet_caps = attr.raw_packet_caps;
4460         resp.response_length += sizeof(resp.raw_packet_caps);
4461 end:
4462         err = ib_copy_to_udata(ucore, &resp, resp.response_length);
4463         return err;
4464 }