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