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