drivers/infiniband/core/uverbs_cmd.c

   1 /*
   2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005, 2006, 2007 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 PathScale, Inc.  All rights reserved.
   5  * Copyright (c) 2006 Mellanox Technologies.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/file.h>
  37 #include <linux/fs.h>
  38 #include <linux/slab.h>
  39 #include <linux/sched.h>
  40
  41 #include <linux/uaccess.h>
  42
  43 #include <rdma/uverbs_types.h>
  44 #include <rdma/uverbs_std_types.h>
  45 #include "rdma_core.h"
  46
  47 #include "uverbs.h"
  48 #include "core_priv.h"
  49
  50 static struct ib_uverbs_completion_event_file *
  51 ib_uverbs_lookup_comp_file(int fd, struct ib_ucontext *context)
  52 {
  53         struct ib_uobject *uobj = uobj_get_read(uobj_get_type(comp_channel),
  54                                                 fd, context);
  55         struct ib_uobject_file *uobj_file;
  56
  57         if (IS_ERR(uobj))
  58                 return (void *)uobj;
  59
  60         uverbs_uobject_get(uobj);
  61         uobj_put_read(uobj);
  62
  63         uobj_file = container_of(uobj, struct ib_uobject_file, uobj);
  64         return container_of(uobj_file, struct ib_uverbs_completion_event_file,
  65                             uobj_file);
  66 }
  67
  68 ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
  69                               struct ib_device *ib_dev,
  70                               const char __user *buf,
  71                               int in_len, int out_len)
  72 {
  73         struct ib_uverbs_get_context      cmd;
  74         struct ib_uverbs_get_context_resp resp;
  75         struct ib_udata                   udata;
  76         struct ib_ucontext               *ucontext;
  77         struct file                      *filp;
  78         struct ib_rdmacg_object          cg_obj;
  79         int ret;
  80
  81         if (out_len < sizeof resp)
  82                 return -ENOSPC;
  83
  84         if (copy_from_user(&cmd, buf, sizeof cmd))
  85                 return -EFAULT;
  86
  87         mutex_lock(&file->mutex);
  88
  89         if (file->ucontext) {
  90                 ret = -EINVAL;
  91                 goto err;
  92         }
  93
  94         INIT_UDATA(&udata, buf + sizeof cmd,
  95                    (unsigned long) cmd.response + sizeof resp,
  96                    in_len - sizeof cmd, out_len - sizeof resp);
  97
  98         ret = ib_rdmacg_try_charge(&cg_obj, ib_dev, RDMACG_RESOURCE_HCA_HANDLE);
  99         if (ret)
 100                 goto err;
 101
 102         ucontext = ib_dev->alloc_ucontext(ib_dev, &udata);
 103         if (IS_ERR(ucontext)) {
 104                 ret = PTR_ERR(ucontext);
 105                 goto err_alloc;
 106         }
 107
 108         ucontext->device = ib_dev;
 109         ucontext->cg_obj = cg_obj;
 110         /* ufile is required when some objects are released */
 111         ucontext->ufile = file;
 112         uverbs_initialize_ucontext(ucontext);
 113
 114         rcu_read_lock();
 115         ucontext->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
 116         rcu_read_unlock();
 117         ucontext->closing = 0;
 118
 119 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 120         ucontext->umem_tree = RB_ROOT;
 121         init_rwsem(&ucontext->umem_rwsem);
 122         ucontext->odp_mrs_count = 0;
 123         INIT_LIST_HEAD(&ucontext->no_private_counters);
 124
 125         if (!(ib_dev->attrs.device_cap_flags & IB_DEVICE_ON_DEMAND_PAGING))
 126                 ucontext->invalidate_range = NULL;
 127
 128 #endif
 129
 130         resp.num_comp_vectors = file->device->num_comp_vectors;
 131
 132         ret = get_unused_fd_flags(O_CLOEXEC);
 133         if (ret < 0)
 134                 goto err_free;
 135         resp.async_fd = ret;
 136
 137         filp = ib_uverbs_alloc_async_event_file(file, ib_dev);
 138         if (IS_ERR(filp)) {
 139                 ret = PTR_ERR(filp);
 140                 goto err_fd;
 141         }
 142
 143         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 144                          &resp, sizeof resp)) {
 145                 ret = -EFAULT;
 146                 goto err_file;
 147         }
 148
 149         file->ucontext = ucontext;
 150
 151         fd_install(resp.async_fd, filp);
 152
 153         mutex_unlock(&file->mutex);
 154
 155         return in_len;
 156
 157 err_file:
 158         ib_uverbs_free_async_event_file(file);
 159         fput(filp);
 160
 161 err_fd:
 162         put_unused_fd(resp.async_fd);
 163
 164 err_free:
 165         put_pid(ucontext->tgid);
 166         ib_dev->dealloc_ucontext(ucontext);
 167
 168 err_alloc:
 169         ib_rdmacg_uncharge(&cg_obj, ib_dev, RDMACG_RESOURCE_HCA_HANDLE);
 170
 171 err:
 172         mutex_unlock(&file->mutex);
 173         return ret;
 174 }
 175
 176 static void copy_query_dev_fields(struct ib_uverbs_file *file,
 177                                   struct ib_device *ib_dev,
 178                                   struct ib_uverbs_query_device_resp *resp,
 179                                   struct ib_device_attr *attr)
 180 {
 181         resp->fw_ver            = attr->fw_ver;
 182         resp->node_guid         = ib_dev->node_guid;
 183         resp->sys_image_guid    = attr->sys_image_guid;
 184         resp->max_mr_size       = attr->max_mr_size;
 185         resp->page_size_cap     = attr->page_size_cap;
 186         resp->vendor_id         = attr->vendor_id;
 187         resp->vendor_part_id    = attr->vendor_part_id;
 188         resp->hw_ver            = attr->hw_ver;
 189         resp->max_qp            = attr->max_qp;
 190         resp->max_qp_wr         = attr->max_qp_wr;
 191         resp->device_cap_flags  = lower_32_bits(attr->device_cap_flags);
 192         resp->max_sge           = attr->max_sge;
 193         resp->max_sge_rd        = attr->max_sge_rd;
 194         resp->max_cq            = attr->max_cq;
 195         resp->max_cqe           = attr->max_cqe;
 196         resp->max_mr            = attr->max_mr;
 197         resp->max_pd            = attr->max_pd;
 198         resp->max_qp_rd_atom    = attr->max_qp_rd_atom;
 199         resp->max_ee_rd_atom    = attr->max_ee_rd_atom;
 200         resp->max_res_rd_atom   = attr->max_res_rd_atom;
 201         resp->max_qp_init_rd_atom       = attr->max_qp_init_rd_atom;
 202         resp->max_ee_init_rd_atom       = attr->max_ee_init_rd_atom;
 203         resp->atomic_cap                = attr->atomic_cap;
 204         resp->max_ee                    = attr->max_ee;
 205         resp->max_rdd                   = attr->max_rdd;
 206         resp->max_mw                    = attr->max_mw;
 207         resp->max_raw_ipv6_qp           = attr->max_raw_ipv6_qp;
 208         resp->max_raw_ethy_qp           = attr->max_raw_ethy_qp;
 209         resp->max_mcast_grp             = attr->max_mcast_grp;
 210         resp->max_mcast_qp_attach       = attr->max_mcast_qp_attach;
 211         resp->max_total_mcast_qp_attach = attr->max_total_mcast_qp_attach;
 212         resp->max_ah                    = attr->max_ah;
 213         resp->max_fmr                   = attr->max_fmr;
 214         resp->max_map_per_fmr           = attr->max_map_per_fmr;
 215         resp->max_srq                   = attr->max_srq;
 216         resp->max_srq_wr                = attr->max_srq_wr;
 217         resp->max_srq_sge               = attr->max_srq_sge;
 218         resp->max_pkeys                 = attr->max_pkeys;
 219         resp->local_ca_ack_delay        = attr->local_ca_ack_delay;
 220         resp->phys_port_cnt             = ib_dev->phys_port_cnt;
 221 }
 222
 223 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 224                                struct ib_device *ib_dev,
 225                                const char __user *buf,
 226                                int in_len, int out_len)
 227 {
 228         struct ib_uverbs_query_device      cmd;
 229         struct ib_uverbs_query_device_resp resp;
 230
 231         if (out_len < sizeof resp)
 232                 return -ENOSPC;
 233
 234         if (copy_from_user(&cmd, buf, sizeof cmd))
 235                 return -EFAULT;
 236
 237         memset(&resp, 0, sizeof resp);
 238         copy_query_dev_fields(file, ib_dev, &resp, &ib_dev->attrs);
 239
 240         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 241                          &resp, sizeof resp))
 242                 return -EFAULT;
 243
 244         return in_len;
 245 }
 246
 247 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 248                              struct ib_device *ib_dev,
 249                              const char __user *buf,
 250                              int in_len, int out_len)
 251 {
 252         struct ib_uverbs_query_port      cmd;
 253         struct ib_uverbs_query_port_resp resp;
 254         struct ib_port_attr              attr;
 255         int                              ret;
 256
 257         if (out_len < sizeof resp)
 258                 return -ENOSPC;
 259
 260         if (copy_from_user(&cmd, buf, sizeof cmd))
 261                 return -EFAULT;
 262
 263         ret = ib_query_port(ib_dev, cmd.port_num, &attr);
 264         if (ret)
 265                 return ret;
 266
 267         memset(&resp, 0, sizeof resp);
 268
 269         resp.state           = attr.state;
 270         resp.max_mtu         = attr.max_mtu;
 271         resp.active_mtu      = attr.active_mtu;
 272         resp.gid_tbl_len     = attr.gid_tbl_len;
 273         resp.port_cap_flags  = attr.port_cap_flags;
 274         resp.max_msg_sz      = attr.max_msg_sz;
 275         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 276         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 277         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 278         resp.lid             = attr.lid;
 279         resp.sm_lid          = attr.sm_lid;
 280         resp.lmc             = attr.lmc;
 281         resp.max_vl_num      = attr.max_vl_num;
 282         resp.sm_sl           = attr.sm_sl;
 283         resp.subnet_timeout  = attr.subnet_timeout;
 284         resp.init_type_reply = attr.init_type_reply;
 285         resp.active_width    = attr.active_width;
 286         resp.active_speed    = attr.active_speed;
 287         resp.phys_state      = attr.phys_state;
 288         resp.link_layer      = rdma_port_get_link_layer(ib_dev,
 289                                                         cmd.port_num);
 290
 291         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 292                          &resp, sizeof resp))
 293                 return -EFAULT;
 294
 295         return in_len;
 296 }
 297
 298 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 299                            struct ib_device *ib_dev,
 300                            const char __user *buf,
 301                            int in_len, int out_len)
 302 {
 303         struct ib_uverbs_alloc_pd      cmd;
 304         struct ib_uverbs_alloc_pd_resp resp;
 305         struct ib_udata                udata;
 306         struct ib_uobject             *uobj;
 307         struct ib_pd                  *pd;
 308         int                            ret;
 309
 310         if (out_len < sizeof resp)
 311                 return -ENOSPC;
 312
 313         if (copy_from_user(&cmd, buf, sizeof cmd))
 314                 return -EFAULT;
 315
 316         INIT_UDATA(&udata, buf + sizeof cmd,
 317                    (unsigned long) cmd.response + sizeof resp,
 318                    in_len - sizeof cmd, out_len - sizeof resp);
 319
 320         uobj  = uobj_alloc(uobj_get_type(pd), file->ucontext);
 321         if (IS_ERR(uobj))
 322                 return PTR_ERR(uobj);
 323
 324         pd = ib_dev->alloc_pd(ib_dev, file->ucontext, &udata);
 325         if (IS_ERR(pd)) {
 326                 ret = PTR_ERR(pd);
 327                 goto err;
 328         }
 329
 330         pd->device  = ib_dev;
 331         pd->uobject = uobj;
 332         pd->__internal_mr = NULL;
 333         atomic_set(&pd->usecnt, 0);
 334
 335         uobj->object = pd;
 336         memset(&resp, 0, sizeof resp);
 337         resp.pd_handle = uobj->id;
 338
 339         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 340                          &resp, sizeof resp)) {
 341                 ret = -EFAULT;
 342                 goto err_copy;
 343         }
 344
 345         uobj_alloc_commit(uobj);
 346
 347         return in_len;
 348
 349 err_copy:
 350         ib_dealloc_pd(pd);
 351
 352 err:
 353         uobj_alloc_abort(uobj);
 354         return ret;
 355 }
 356
 357 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 358                              struct ib_device *ib_dev,
 359                              const char __user *buf,
 360                              int in_len, int out_len)
 361 {
 362         struct ib_uverbs_dealloc_pd cmd;
 363         struct ib_uobject          *uobj;
 364         int                         ret;
 365
 366         if (copy_from_user(&cmd, buf, sizeof cmd))
 367                 return -EFAULT;
 368
 369         uobj  = uobj_get_write(uobj_get_type(pd), cmd.pd_handle,
 370                                file->ucontext);
 371         if (IS_ERR(uobj))
 372                 return PTR_ERR(uobj);
 373
 374         ret = uobj_remove_commit(uobj);
 375
 376         return ret ?: in_len;
 377 }
 378
 379 struct xrcd_table_entry {
 380         struct rb_node  node;
 381         struct ib_xrcd *xrcd;
 382         struct inode   *inode;
 383 };
 384
 385 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 386                             struct inode *inode,
 387                             struct ib_xrcd *xrcd)
 388 {
 389         struct xrcd_table_entry *entry, *scan;
 390         struct rb_node **p = &dev->xrcd_tree.rb_node;
 391         struct rb_node *parent = NULL;
 392
 393         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 394         if (!entry)
 395                 return -ENOMEM;
 396
 397         entry->xrcd  = xrcd;
 398         entry->inode = inode;
 399
 400         while (*p) {
 401                 parent = *p;
 402                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 403
 404                 if (inode < scan->inode) {
 405                         p = &(*p)->rb_left;
 406                 } else if (inode > scan->inode) {
 407                         p = &(*p)->rb_right;
 408                 } else {
 409                         kfree(entry);
 410                         return -EEXIST;
 411                 }
 412         }
 413
 414         rb_link_node(&entry->node, parent, p);
 415         rb_insert_color(&entry->node, &dev->xrcd_tree);
 416         igrab(inode);
 417         return 0;
 418 }
 419
 420 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 421                                                   struct inode *inode)
 422 {
 423         struct xrcd_table_entry *entry;
 424         struct rb_node *p = dev->xrcd_tree.rb_node;
 425
 426         while (p) {
 427                 entry = rb_entry(p, struct xrcd_table_entry, node);
 428
 429                 if (inode < entry->inode)
 430                         p = p->rb_left;
 431                 else if (inode > entry->inode)
 432                         p = p->rb_right;
 433                 else
 434                         return entry;
 435         }
 436
 437         return NULL;
 438 }
 439
 440 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 441 {
 442         struct xrcd_table_entry *entry;
 443
 444         entry = xrcd_table_search(dev, inode);
 445         if (!entry)
 446                 return NULL;
 447
 448         return entry->xrcd;
 449 }
 450
 451 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 452                               struct inode *inode)
 453 {
 454         struct xrcd_table_entry *entry;
 455
 456         entry = xrcd_table_search(dev, inode);
 457         if (entry) {
 458                 iput(inode);
 459                 rb_erase(&entry->node, &dev->xrcd_tree);
 460                 kfree(entry);
 461         }
 462 }
 463
 464 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 465                             struct ib_device *ib_dev,
 466                             const char __user *buf, int in_len,
 467                             int out_len)
 468 {
 469         struct ib_uverbs_open_xrcd      cmd;
 470         struct ib_uverbs_open_xrcd_resp resp;
 471         struct ib_udata                 udata;
 472         struct ib_uxrcd_object         *obj;
 473         struct ib_xrcd                 *xrcd = NULL;
 474         struct fd                       f = {NULL, 0};
 475         struct inode                   *inode = NULL;
 476         int                             ret = 0;
 477         int                             new_xrcd = 0;
 478
 479         if (out_len < sizeof resp)
 480                 return -ENOSPC;
 481
 482         if (copy_from_user(&cmd, buf, sizeof cmd))
 483                 return -EFAULT;
 484
 485         INIT_UDATA(&udata, buf + sizeof cmd,
 486                    (unsigned long) cmd.response + sizeof resp,
 487                    in_len - sizeof cmd, out_len - sizeof  resp);
 488
 489         mutex_lock(&file->device->xrcd_tree_mutex);
 490
 491         if (cmd.fd != -1) {
 492                 /* search for file descriptor */
 493                 f = fdget(cmd.fd);
 494                 if (!f.file) {
 495                         ret = -EBADF;
 496                         goto err_tree_mutex_unlock;
 497                 }
 498
 499                 inode = file_inode(f.file);
 500                 xrcd = find_xrcd(file->device, inode);
 501                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 502                         /* no file descriptor. Need CREATE flag */
 503                         ret = -EAGAIN;
 504                         goto err_tree_mutex_unlock;
 505                 }
 506
 507                 if (xrcd && cmd.oflags & O_EXCL) {
 508                         ret = -EINVAL;
 509                         goto err_tree_mutex_unlock;
 510                 }
 511         }
 512
 513         obj  = (struct ib_uxrcd_object *)uobj_alloc(uobj_get_type(xrcd),
 514                                                     file->ucontext);
 515         if (IS_ERR(obj)) {
 516                 ret = PTR_ERR(obj);
 517                 goto err_tree_mutex_unlock;
 518         }
 519
 520         if (!xrcd) {
 521                 xrcd = ib_dev->alloc_xrcd(ib_dev, file->ucontext, &udata);
 522                 if (IS_ERR(xrcd)) {
 523                         ret = PTR_ERR(xrcd);
 524                         goto err;
 525                 }
 526
 527                 xrcd->inode   = inode;
 528                 xrcd->device  = ib_dev;
 529                 atomic_set(&xrcd->usecnt, 0);
 530                 mutex_init(&xrcd->tgt_qp_mutex);
 531                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 532                 new_xrcd = 1;
 533         }
 534
 535         atomic_set(&obj->refcnt, 0);
 536         obj->uobject.object = xrcd;
 537         memset(&resp, 0, sizeof resp);
 538         resp.xrcd_handle = obj->uobject.id;
 539
 540         if (inode) {
 541                 if (new_xrcd) {
 542                         /* create new inode/xrcd table entry */
 543                         ret = xrcd_table_insert(file->device, inode, xrcd);
 544                         if (ret)
 545                                 goto err_dealloc_xrcd;
 546                 }
 547                 atomic_inc(&xrcd->usecnt);
 548         }
 549
 550         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 551                          &resp, sizeof resp)) {
 552                 ret = -EFAULT;
 553                 goto err_copy;
 554         }
 555
 556         if (f.file)
 557                 fdput(f);
 558
 559         uobj_alloc_commit(&obj->uobject);
 560
 561         mutex_unlock(&file->device->xrcd_tree_mutex);
 562         return in_len;
 563
 564 err_copy:
 565         if (inode) {
 566                 if (new_xrcd)
 567                         xrcd_table_delete(file->device, inode);
 568                 atomic_dec(&xrcd->usecnt);
 569         }
 570
 571 err_dealloc_xrcd:
 572         ib_dealloc_xrcd(xrcd);
 573
 574 err:
 575         uobj_alloc_abort(&obj->uobject);
 576
 577 err_tree_mutex_unlock:
 578         if (f.file)
 579                 fdput(f);
 580
 581         mutex_unlock(&file->device->xrcd_tree_mutex);
 582
 583         return ret;
 584 }
 585
 586 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 587                              struct ib_device *ib_dev,
 588                              const char __user *buf, int in_len,
 589                              int out_len)
 590 {
 591         struct ib_uverbs_close_xrcd cmd;
 592         struct ib_uobject           *uobj;
 593         int                         ret = 0;
 594
 595         if (copy_from_user(&cmd, buf, sizeof cmd))
 596                 return -EFAULT;
 597
 598         uobj  = uobj_get_write(uobj_get_type(xrcd), cmd.xrcd_handle,
 599                                file->ucontext);
 600         if (IS_ERR(uobj)) {
 601                 mutex_unlock(&file->device->xrcd_tree_mutex);
 602                 return PTR_ERR(uobj);
 603         }
 604
 605         ret = uobj_remove_commit(uobj);
 606         return ret ?: in_len;
 607 }
 608
 609 int ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 610                            struct ib_xrcd *xrcd,
 611                            enum rdma_remove_reason why)
 612 {
 613         struct inode *inode;
 614         int ret;
 615
 616         inode = xrcd->inode;
 617         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 618                 return 0;
 619
 620         ret = ib_dealloc_xrcd(xrcd);
 621
 622         if (why == RDMA_REMOVE_DESTROY && ret)
 623                 atomic_inc(&xrcd->usecnt);
 624         else if (inode)
 625                 xrcd_table_delete(dev, inode);
 626
 627         return ret;
 628 }
 629
 630 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 631                          struct ib_device *ib_dev,
 632                          const char __user *buf, int in_len,
 633                          int out_len)
 634 {
 635         struct ib_uverbs_reg_mr      cmd;
 636         struct ib_uverbs_reg_mr_resp resp;
 637         struct ib_udata              udata;
 638         struct ib_uobject           *uobj;
 639         struct ib_pd                *pd;
 640         struct ib_mr                *mr;
 641         int                          ret;
 642
 643         if (out_len < sizeof resp)
 644                 return -ENOSPC;
 645
 646         if (copy_from_user(&cmd, buf, sizeof cmd))
 647                 return -EFAULT;
 648
 649         INIT_UDATA(&udata, buf + sizeof cmd,
 650                    (unsigned long) cmd.response + sizeof resp,
 651                    in_len - sizeof cmd, out_len - sizeof resp);
 652
 653         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 654                 return -EINVAL;
 655
 656         ret = ib_check_mr_access(cmd.access_flags);
 657         if (ret)
 658                 return ret;
 659
 660         uobj  = uobj_alloc(uobj_get_type(mr), file->ucontext);
 661         if (IS_ERR(uobj))
 662                 return PTR_ERR(uobj);
 663
 664         pd = uobj_get_obj_read(pd, cmd.pd_handle, file->ucontext);
 665         if (!pd) {
 666                 ret = -EINVAL;
 667                 goto err_free;
 668         }
 669
 670         if (cmd.access_flags & IB_ACCESS_ON_DEMAND) {
 671                 if (!(pd->device->attrs.device_cap_flags &
 672                       IB_DEVICE_ON_DEMAND_PAGING)) {
 673                         pr_debug("ODP support not available\n");
 674                         ret = -EINVAL;
 675                         goto err_put;
 676                 }
 677         }
 678
 679         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 680                                      cmd.access_flags, &udata);
 681         if (IS_ERR(mr)) {
 682                 ret = PTR_ERR(mr);
 683                 goto err_put;
 684         }
 685
 686         mr->device  = pd->device;
 687         mr->pd      = pd;
 688         mr->uobject = uobj;
 689         atomic_inc(&pd->usecnt);
 690
 691         uobj->object = mr;
 692
 693         memset(&resp, 0, sizeof resp);
 694         resp.lkey      = mr->lkey;
 695         resp.rkey      = mr->rkey;
 696         resp.mr_handle = uobj->id;
 697
 698         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 699                          &resp, sizeof resp)) {
 700                 ret = -EFAULT;
 701                 goto err_copy;
 702         }
 703
 704         uobj_put_obj_read(pd);
 705
 706         uobj_alloc_commit(uobj);
 707
 708         return in_len;
 709
 710 err_copy:
 711         ib_dereg_mr(mr);
 712
 713 err_put:
 714         uobj_put_obj_read(pd);
 715
 716 err_free:
 717         uobj_alloc_abort(uobj);
 718         return ret;
 719 }
 720
 721 ssize_t ib_uverbs_rereg_mr(struct ib_uverbs_file *file,
 722                            struct ib_device *ib_dev,
 723                            const char __user *buf, int in_len,
 724                            int out_len)
 725 {
 726         struct ib_uverbs_rereg_mr      cmd;
 727         struct ib_uverbs_rereg_mr_resp resp;
 728         struct ib_udata              udata;
 729         struct ib_pd                *pd = NULL;
 730         struct ib_mr                *mr;
 731         struct ib_pd                *old_pd;
 732         int                          ret;
 733         struct ib_uobject           *uobj;
 734
 735         if (out_len < sizeof(resp))
 736                 return -ENOSPC;
 737
 738         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 739                 return -EFAULT;
 740
 741         INIT_UDATA(&udata, buf + sizeof(cmd),
 742                    (unsigned long) cmd.response + sizeof(resp),
 743                    in_len - sizeof(cmd), out_len - sizeof(resp));
 744
 745         if (cmd.flags & ~IB_MR_REREG_SUPPORTED || !cmd.flags)
 746                 return -EINVAL;
 747
 748         if ((cmd.flags & IB_MR_REREG_TRANS) &&
 749             (!cmd.start || !cmd.hca_va || 0 >= cmd.length ||
 750              (cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK)))
 751                         return -EINVAL;
 752
 753         uobj  = uobj_get_write(uobj_get_type(mr), cmd.mr_handle,
 754                                file->ucontext);
 755         if (IS_ERR(uobj))
 756                 return PTR_ERR(uobj);
 757
 758         mr = uobj->object;
 759
 760         if (cmd.flags & IB_MR_REREG_ACCESS) {
 761                 ret = ib_check_mr_access(cmd.access_flags);
 762                 if (ret)
 763                         goto put_uobjs;
 764         }
 765
 766         if (cmd.flags & IB_MR_REREG_PD) {
 767                 pd = uobj_get_obj_read(pd, cmd.pd_handle, file->ucontext);
 768                 if (!pd) {
 769                         ret = -EINVAL;
 770                         goto put_uobjs;
 771                 }
 772         }
 773
 774         old_pd = mr->pd;
 775         ret = mr->device->rereg_user_mr(mr, cmd.flags, cmd.start,
 776                                         cmd.length, cmd.hca_va,
 777                                         cmd.access_flags, pd, &udata);
 778         if (!ret) {
 779                 if (cmd.flags & IB_MR_REREG_PD) {
 780                         atomic_inc(&pd->usecnt);
 781                         mr->pd = pd;
 782                         atomic_dec(&old_pd->usecnt);
 783                 }
 784         } else {
 785                 goto put_uobj_pd;
 786         }
 787
 788         memset(&resp, 0, sizeof(resp));
 789         resp.lkey      = mr->lkey;
 790         resp.rkey      = mr->rkey;
 791
 792         if (copy_to_user((void __user *)(unsigned long)cmd.response,
 793                          &resp, sizeof(resp)))
 794                 ret = -EFAULT;
 795         else
 796                 ret = in_len;
 797
 798 put_uobj_pd:
 799         if (cmd.flags & IB_MR_REREG_PD)
 800                 uobj_put_obj_read(pd);
 801
 802 put_uobjs:
 803         uobj_put_write(uobj);
 804
 805         return ret;
 806 }
 807
 808 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
 809                            struct ib_device *ib_dev,
 810                            const char __user *buf, int in_len,
 811                            int out_len)
 812 {
 813         struct ib_uverbs_dereg_mr cmd;
 814         struct ib_uobject        *uobj;
 815         int                       ret = -EINVAL;
 816
 817         if (copy_from_user(&cmd, buf, sizeof cmd))
 818                 return -EFAULT;
 819
 820         uobj  = uobj_get_write(uobj_get_type(mr), cmd.mr_handle,
 821                                file->ucontext);
 822         if (IS_ERR(uobj))
 823                 return PTR_ERR(uobj);
 824
 825         ret = uobj_remove_commit(uobj);
 826
 827         return ret ?: in_len;
 828 }
 829
 830 ssize_t ib_uverbs_alloc_mw(struct ib_uverbs_file *file,
 831                            struct ib_device *ib_dev,
 832                            const char __user *buf, int in_len,
 833                            int out_len)
 834 {
 835         struct ib_uverbs_alloc_mw      cmd;
 836         struct ib_uverbs_alloc_mw_resp resp;
 837         struct ib_uobject             *uobj;
 838         struct ib_pd                  *pd;
 839         struct ib_mw                  *mw;
 840         struct ib_udata                udata;
 841         int                            ret;
 842
 843         if (out_len < sizeof(resp))
 844                 return -ENOSPC;
 845
 846         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 847                 return -EFAULT;
 848
 849         uobj  = uobj_alloc(uobj_get_type(mw), file->ucontext);
 850         if (IS_ERR(uobj))
 851                 return PTR_ERR(uobj);
 852
 853         pd = uobj_get_obj_read(pd, cmd.pd_handle, file->ucontext);
 854         if (!pd) {
 855                 ret = -EINVAL;
 856                 goto err_free;
 857         }
 858
 859         INIT_UDATA(&udata, buf + sizeof(cmd),
 860                    (unsigned long)cmd.response + sizeof(resp),
 861                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 862                    out_len - sizeof(resp));
 863
 864         mw = pd->device->alloc_mw(pd, cmd.mw_type, &udata);
 865         if (IS_ERR(mw)) {
 866                 ret = PTR_ERR(mw);
 867                 goto err_put;
 868         }
 869
 870         mw->device  = pd->device;
 871         mw->pd      = pd;
 872         mw->uobject = uobj;
 873         atomic_inc(&pd->usecnt);
 874
 875         uobj->object = mw;
 876
 877         memset(&resp, 0, sizeof(resp));
 878         resp.rkey      = mw->rkey;
 879         resp.mw_handle = uobj->id;
 880
 881         if (copy_to_user((void __user *)(unsigned long)cmd.response,
 882                          &resp, sizeof(resp))) {
 883                 ret = -EFAULT;
 884                 goto err_copy;
 885         }
 886
 887         uobj_put_obj_read(pd);
 888         uobj_alloc_commit(uobj);
 889
 890         return in_len;
 891
 892 err_copy:
 893         uverbs_dealloc_mw(mw);
 894 err_put:
 895         uobj_put_obj_read(pd);
 896 err_free:
 897         uobj_alloc_abort(uobj);
 898         return ret;
 899 }
 900
 901 ssize_t ib_uverbs_dealloc_mw(struct ib_uverbs_file *file,
 902                              struct ib_device *ib_dev,
 903                              const char __user *buf, int in_len,
 904                              int out_len)
 905 {
 906         struct ib_uverbs_dealloc_mw cmd;
 907         struct ib_uobject          *uobj;
 908         int                         ret = -EINVAL;
 909
 910         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 911                 return -EFAULT;
 912
 913         uobj  = uobj_get_write(uobj_get_type(mw), cmd.mw_handle,
 914                                file->ucontext);
 915         if (IS_ERR(uobj))
 916                 return PTR_ERR(uobj);
 917
 918         ret = uobj_remove_commit(uobj);
 919         return ret ?: in_len;
 920 }
 921
 922 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
 923                                       struct ib_device *ib_dev,
 924                                       const char __user *buf, int in_len,
 925                                       int out_len)
 926 {
 927         struct ib_uverbs_create_comp_channel       cmd;
 928         struct ib_uverbs_create_comp_channel_resp  resp;
 929         struct ib_uobject                         *uobj;
 930         struct ib_uverbs_completion_event_file    *ev_file;
 931
 932         if (out_len < sizeof resp)
 933                 return -ENOSPC;
 934
 935         if (copy_from_user(&cmd, buf, sizeof cmd))
 936                 return -EFAULT;
 937
 938         uobj = uobj_alloc(uobj_get_type(comp_channel), file->ucontext);
 939         if (IS_ERR(uobj))
 940                 return PTR_ERR(uobj);
 941
 942         resp.fd = uobj->id;
 943
 944         ev_file = container_of(uobj, struct ib_uverbs_completion_event_file,
 945                                uobj_file.uobj);
 946         ib_uverbs_init_event_queue(&ev_file->ev_queue);
 947
 948         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 949                          &resp, sizeof resp)) {
 950                 uobj_alloc_abort(uobj);
 951                 return -EFAULT;
 952         }
 953
 954         uobj_alloc_commit(uobj);
 955         return in_len;
 956 }
 957
 958 static struct ib_ucq_object *create_cq(struct ib_uverbs_file *file,
 959                                         struct ib_device *ib_dev,
 960                                        struct ib_udata *ucore,
 961                                        struct ib_udata *uhw,
 962                                        struct ib_uverbs_ex_create_cq *cmd,
 963                                        size_t cmd_sz,
 964                                        int (*cb)(struct ib_uverbs_file *file,
 965                                                  struct ib_ucq_object *obj,
 966                                                  struct ib_uverbs_ex_create_cq_resp *resp,
 967                                                  struct ib_udata *udata,
 968                                                  void *context),
 969                                        void *context)
 970 {
 971         struct ib_ucq_object           *obj;
 972         struct ib_uverbs_completion_event_file    *ev_file = NULL;
 973         struct ib_cq                   *cq;
 974         int                             ret;
 975         struct ib_uverbs_ex_create_cq_resp resp;
 976         struct ib_cq_init_attr attr = {};
 977
 978         if (cmd->comp_vector >= file->device->num_comp_vectors)
 979                 return ERR_PTR(-EINVAL);
 980
 981         obj  = (struct ib_ucq_object *)uobj_alloc(uobj_get_type(cq),
 982                                                   file->ucontext);
 983         if (IS_ERR(obj))
 984                 return obj;
 985
 986         if (cmd->comp_channel >= 0) {
 987                 ev_file = ib_uverbs_lookup_comp_file(cmd->comp_channel,
 988                                                      file->ucontext);
 989                 if (IS_ERR(ev_file)) {
 990                         ret = PTR_ERR(ev_file);
 991                         goto err;
 992                 }
 993         }
 994
 995         obj->uobject.user_handle = cmd->user_handle;
 996         obj->uverbs_file           = file;
 997         obj->comp_events_reported  = 0;
 998         obj->async_events_reported = 0;
 999         INIT_LIST_HEAD(&obj->comp_list);
1000         INIT_LIST_HEAD(&obj->async_list);
1001
1002         attr.cqe = cmd->cqe;
1003         attr.comp_vector = cmd->comp_vector;
1004
1005         if (cmd_sz > offsetof(typeof(*cmd), flags) + sizeof(cmd->flags))
1006                 attr.flags = cmd->flags;
1007
1008         cq = ib_dev->create_cq(ib_dev, &attr, file->ucontext, uhw);
1009         if (IS_ERR(cq)) {
1010                 ret = PTR_ERR(cq);
1011                 goto err_file;
1012         }
1013
1014         cq->device        = ib_dev;
1015         cq->uobject       = &obj->uobject;
1016         cq->comp_handler  = ib_uverbs_comp_handler;
1017         cq->event_handler = ib_uverbs_cq_event_handler;
1018         cq->cq_context    = &ev_file->ev_queue;
1019         atomic_set(&cq->usecnt, 0);
1020
1021         obj->uobject.object = cq;
1022         memset(&resp, 0, sizeof resp);
1023         resp.base.cq_handle = obj->uobject.id;
1024         resp.base.cqe       = cq->cqe;
1025
1026         resp.response_length = offsetof(typeof(resp), response_length) +
1027                 sizeof(resp.response_length);
1028
1029         ret = cb(file, obj, &resp, ucore, context);
1030         if (ret)
1031                 goto err_cb;
1032
1033         uobj_alloc_commit(&obj->uobject);
1034
1035         return obj;
1036
1037 err_cb:
1038         ib_destroy_cq(cq);
1039
1040 err_file:
1041         if (ev_file)
1042                 ib_uverbs_release_ucq(file, ev_file, obj);
1043
1044 err:
1045         uobj_alloc_abort(&obj->uobject);
1046
1047         return ERR_PTR(ret);
1048 }
1049
1050 static int ib_uverbs_create_cq_cb(struct ib_uverbs_file *file,
1051                                   struct ib_ucq_object *obj,
1052                                   struct ib_uverbs_ex_create_cq_resp *resp,
1053                                   struct ib_udata *ucore, void *context)
1054 {
1055         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1056                 return -EFAULT;
1057
1058         return 0;
1059 }
1060
1061 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1062                             struct ib_device *ib_dev,
1063                             const char __user *buf, int in_len,
1064                             int out_len)
1065 {
1066         struct ib_uverbs_create_cq      cmd;
1067         struct ib_uverbs_ex_create_cq   cmd_ex;
1068         struct ib_uverbs_create_cq_resp resp;
1069         struct ib_udata                 ucore;
1070         struct ib_udata                 uhw;
1071         struct ib_ucq_object           *obj;
1072
1073         if (out_len < sizeof(resp))
1074                 return -ENOSPC;
1075
1076         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1077                 return -EFAULT;
1078
1079         INIT_UDATA(&ucore, buf, (unsigned long)cmd.response, sizeof(cmd), sizeof(resp));
1080
1081         INIT_UDATA(&uhw, buf + sizeof(cmd),
1082                    (unsigned long)cmd.response + sizeof(resp),
1083                    in_len - sizeof(cmd), out_len - sizeof(resp));
1084
1085         memset(&cmd_ex, 0, sizeof(cmd_ex));
1086         cmd_ex.user_handle = cmd.user_handle;
1087         cmd_ex.cqe = cmd.cqe;
1088         cmd_ex.comp_vector = cmd.comp_vector;
1089         cmd_ex.comp_channel = cmd.comp_channel;
1090
1091         obj = create_cq(file, ib_dev, &ucore, &uhw, &cmd_ex,
1092                         offsetof(typeof(cmd_ex), comp_channel) +
1093                         sizeof(cmd.comp_channel), ib_uverbs_create_cq_cb,
1094                         NULL);
1095
1096         if (IS_ERR(obj))
1097                 return PTR_ERR(obj);
1098
1099         return in_len;
1100 }
1101
1102 static int ib_uverbs_ex_create_cq_cb(struct ib_uverbs_file *file,
1103                                      struct ib_ucq_object *obj,
1104                                      struct ib_uverbs_ex_create_cq_resp *resp,
1105                                      struct ib_udata *ucore, void *context)
1106 {
1107         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1108                 return -EFAULT;
1109
1110         return 0;
1111 }
1112
1113 int ib_uverbs_ex_create_cq(struct ib_uverbs_file *file,
1114                          struct ib_device *ib_dev,
1115                            struct ib_udata *ucore,
1116                            struct ib_udata *uhw)
1117 {
1118         struct ib_uverbs_ex_create_cq_resp resp;
1119         struct ib_uverbs_ex_create_cq  cmd;
1120         struct ib_ucq_object           *obj;
1121         int err;
1122
1123         if (ucore->inlen < sizeof(cmd))
1124                 return -EINVAL;
1125
1126         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
1127         if (err)
1128                 return err;
1129
1130         if (cmd.comp_mask)
1131                 return -EINVAL;
1132
1133         if (cmd.reserved)
1134                 return -EINVAL;
1135
1136         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1137                              sizeof(resp.response_length)))
1138                 return -ENOSPC;
1139
1140         obj = create_cq(file, ib_dev, ucore, uhw, &cmd,
1141                         min(ucore->inlen, sizeof(cmd)),
1142                         ib_uverbs_ex_create_cq_cb, NULL);
1143
1144         if (IS_ERR(obj))
1145                 return PTR_ERR(obj);
1146
1147         return 0;
1148 }
1149
1150 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1151                             struct ib_device *ib_dev,
1152                             const char __user *buf, int in_len,
1153                             int out_len)
1154 {
1155         struct ib_uverbs_resize_cq      cmd;
1156         struct ib_uverbs_resize_cq_resp resp;
1157         struct ib_udata                 udata;
1158         struct ib_cq                    *cq;
1159         int                             ret = -EINVAL;
1160
1161         if (copy_from_user(&cmd, buf, sizeof cmd))
1162                 return -EFAULT;
1163
1164         INIT_UDATA(&udata, buf + sizeof cmd,
1165                    (unsigned long) cmd.response + sizeof resp,
1166                    in_len - sizeof cmd, out_len - sizeof resp);
1167
1168         cq = uobj_get_obj_read(cq, cmd.cq_handle, file->ucontext);
1169         if (!cq)
1170                 return -EINVAL;
1171
1172         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1173         if (ret)
1174                 goto out;
1175
1176         resp.cqe = cq->cqe;
1177
1178         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1179                          &resp, sizeof resp.cqe))
1180                 ret = -EFAULT;
1181
1182 out:
1183         uobj_put_obj_read(cq);
1184
1185         return ret ? ret : in_len;
1186 }
1187
1188 static int copy_wc_to_user(void __user *dest, struct ib_wc *wc)
1189 {
1190         struct ib_uverbs_wc tmp;
1191
1192         tmp.wr_id               = wc->wr_id;
1193         tmp.status              = wc->status;
1194         tmp.opcode              = wc->opcode;
1195         tmp.vendor_err          = wc->vendor_err;
1196         tmp.byte_len            = wc->byte_len;
1197         tmp.ex.imm_data         = (__u32 __force) wc->ex.imm_data;
1198         tmp.qp_num              = wc->qp->qp_num;
1199         tmp.src_qp              = wc->src_qp;
1200         tmp.wc_flags            = wc->wc_flags;
1201         tmp.pkey_index          = wc->pkey_index;
1202         tmp.slid                = wc->slid;
1203         tmp.sl                  = wc->sl;
1204         tmp.dlid_path_bits      = wc->dlid_path_bits;
1205         tmp.port_num            = wc->port_num;
1206         tmp.reserved            = 0;
1207
1208         if (copy_to_user(dest, &tmp, sizeof tmp))
1209                 return -EFAULT;
1210
1211         return 0;
1212 }
1213
1214 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1215                           struct ib_device *ib_dev,
1216                           const char __user *buf, int in_len,
1217                           int out_len)
1218 {
1219         struct ib_uverbs_poll_cq       cmd;
1220         struct ib_uverbs_poll_cq_resp  resp;
1221         u8 __user                     *header_ptr;
1222         u8 __user                     *data_ptr;
1223         struct ib_cq                  *cq;
1224         struct ib_wc                   wc;
1225         int                            ret;
1226
1227         if (copy_from_user(&cmd, buf, sizeof cmd))
1228                 return -EFAULT;
1229
1230         cq = uobj_get_obj_read(cq, cmd.cq_handle, file->ucontext);
1231         if (!cq)
1232                 return -EINVAL;
1233
1234         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1235         header_ptr = (void __user *)(unsigned long) cmd.response;
1236         data_ptr = header_ptr + sizeof resp;
1237
1238         memset(&resp, 0, sizeof resp);
1239         while (resp.count < cmd.ne) {
1240                 ret = ib_poll_cq(cq, 1, &wc);
1241                 if (ret < 0)
1242                         goto out_put;
1243                 if (!ret)
1244                         break;
1245
1246                 ret = copy_wc_to_user(data_ptr, &wc);
1247                 if (ret)
1248                         goto out_put;
1249
1250                 data_ptr += sizeof(struct ib_uverbs_wc);
1251                 ++resp.count;
1252         }
1253
1254         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1255                 ret = -EFAULT;
1256                 goto out_put;
1257         }
1258
1259         ret = in_len;
1260
1261 out_put:
1262         uobj_put_obj_read(cq);
1263         return ret;
1264 }
1265
1266 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1267                                 struct ib_device *ib_dev,
1268                                 const char __user *buf, int in_len,
1269                                 int out_len)
1270 {
1271         struct ib_uverbs_req_notify_cq cmd;
1272         struct ib_cq                  *cq;
1273
1274         if (copy_from_user(&cmd, buf, sizeof cmd))
1275                 return -EFAULT;
1276
1277         cq = uobj_get_obj_read(cq, cmd.cq_handle, file->ucontext);
1278         if (!cq)
1279                 return -EINVAL;
1280
1281         ib_req_notify_cq(cq, cmd.solicited_only ?
1282                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1283
1284         uobj_put_obj_read(cq);
1285
1286         return in_len;
1287 }
1288
1289 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1290                              struct ib_device *ib_dev,
1291                              const char __user *buf, int in_len,
1292                              int out_len)
1293 {
1294         struct ib_uverbs_destroy_cq      cmd;
1295         struct ib_uverbs_destroy_cq_resp resp;
1296         struct ib_uobject               *uobj;
1297         struct ib_cq                    *cq;
1298         struct ib_ucq_object            *obj;
1299         struct ib_uverbs_event_queue    *ev_queue;
1300         int                              ret = -EINVAL;
1301
1302         if (copy_from_user(&cmd, buf, sizeof cmd))
1303                 return -EFAULT;
1304
1305         uobj  = uobj_get_write(uobj_get_type(cq), cmd.cq_handle,
1306                                file->ucontext);
1307         if (IS_ERR(uobj))
1308                 return PTR_ERR(uobj);
1309
1310         /*
1311          * Make sure we don't free the memory in remove_commit as we still
1312          * needs the uobject memory to create the response.
1313          */
1314         uverbs_uobject_get(uobj);
1315         cq      = uobj->object;
1316         ev_queue = cq->cq_context;
1317         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1318
1319         memset(&resp, 0, sizeof(resp));
1320
1321         ret = uobj_remove_commit(uobj);
1322         if (ret) {
1323                 uverbs_uobject_put(uobj);
1324                 return ret;
1325         }
1326
1327         resp.comp_events_reported  = obj->comp_events_reported;
1328         resp.async_events_reported = obj->async_events_reported;
1329
1330         uverbs_uobject_put(uobj);
1331         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1332                          &resp, sizeof resp))
1333                 return -EFAULT;
1334
1335         return in_len;
1336 }
1337
1338 static int create_qp(struct ib_uverbs_file *file,
1339                      struct ib_udata *ucore,
1340                      struct ib_udata *uhw,
1341                      struct ib_uverbs_ex_create_qp *cmd,
1342                      size_t cmd_sz,
1343                      int (*cb)(struct ib_uverbs_file *file,
1344                                struct ib_uverbs_ex_create_qp_resp *resp,
1345                                struct ib_udata *udata),
1346                      void *context)
1347 {
1348         struct ib_uqp_object            *obj;
1349         struct ib_device                *device;
1350         struct ib_pd                    *pd = NULL;
1351         struct ib_xrcd                  *xrcd = NULL;
1352         struct ib_uobject               *xrcd_uobj = ERR_PTR(-ENOENT);
1353         struct ib_cq                    *scq = NULL, *rcq = NULL;
1354         struct ib_srq                   *srq = NULL;
1355         struct ib_qp                    *qp;
1356         char                            *buf;
1357         struct ib_qp_init_attr          attr = {};
1358         struct ib_uverbs_ex_create_qp_resp resp;
1359         int                             ret;
1360         struct ib_rwq_ind_table *ind_tbl = NULL;
1361         bool has_sq = true;
1362
1363         if (cmd->qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1364                 return -EPERM;
1365
1366         obj  = (struct ib_uqp_object *)uobj_alloc(uobj_get_type(qp),
1367                                                   file->ucontext);
1368         if (IS_ERR(obj))
1369                 return PTR_ERR(obj);
1370         obj->uxrcd = NULL;
1371         obj->uevent.uobject.user_handle = cmd->user_handle;
1372         mutex_init(&obj->mcast_lock);
1373
1374         if (cmd_sz >= offsetof(typeof(*cmd), rwq_ind_tbl_handle) +
1375                       sizeof(cmd->rwq_ind_tbl_handle) &&
1376                       (cmd->comp_mask & IB_UVERBS_CREATE_QP_MASK_IND_TABLE)) {
1377                 ind_tbl = uobj_get_obj_read(rwq_ind_table,
1378                                             cmd->rwq_ind_tbl_handle,
1379                                             file->ucontext);
1380                 if (!ind_tbl) {
1381                         ret = -EINVAL;
1382                         goto err_put;
1383                 }
1384
1385                 attr.rwq_ind_tbl = ind_tbl;
1386         }
1387
1388         if ((cmd_sz >= offsetof(typeof(*cmd), reserved1) +
1389                        sizeof(cmd->reserved1)) && cmd->reserved1) {
1390                 ret = -EOPNOTSUPP;
1391                 goto err_put;
1392         }
1393
1394         if (ind_tbl && (cmd->max_recv_wr || cmd->max_recv_sge || cmd->is_srq)) {
1395                 ret = -EINVAL;
1396                 goto err_put;
1397         }
1398
1399         if (ind_tbl && !cmd->max_send_wr)
1400                 has_sq = false;
1401
1402         if (cmd->qp_type == IB_QPT_XRC_TGT) {
1403                 xrcd_uobj = uobj_get_read(uobj_get_type(xrcd), cmd->pd_handle,
1404                                           file->ucontext);
1405
1406                 if (IS_ERR(xrcd_uobj)) {
1407                         ret = -EINVAL;
1408                         goto err_put;
1409                 }
1410
1411                 xrcd = (struct ib_xrcd *)xrcd_uobj->object;
1412                 if (!xrcd) {
1413                         ret = -EINVAL;
1414                         goto err_put;
1415                 }
1416                 device = xrcd->device;
1417         } else {
1418                 if (cmd->qp_type == IB_QPT_XRC_INI) {
1419                         cmd->max_recv_wr = 0;
1420                         cmd->max_recv_sge = 0;
1421                 } else {
1422                         if (cmd->is_srq) {
1423                                 srq = uobj_get_obj_read(srq, cmd->srq_handle,
1424                                                         file->ucontext);
1425                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1426                                         ret = -EINVAL;
1427                                         goto err_put;
1428                                 }
1429                         }
1430
1431                         if (!ind_tbl) {
1432                                 if (cmd->recv_cq_handle != cmd->send_cq_handle) {
1433                                         rcq = uobj_get_obj_read(cq, cmd->recv_cq_handle,
1434                                                                 file->ucontext);
1435                                         if (!rcq) {
1436                                                 ret = -EINVAL;
1437                                                 goto err_put;
1438                                         }
1439                                 }
1440                         }
1441                 }
1442
1443                 if (has_sq)
1444                         scq = uobj_get_obj_read(cq, cmd->send_cq_handle,
1445                                                 file->ucontext);
1446                 if (!ind_tbl)
1447                         rcq = rcq ?: scq;
1448                 pd  = uobj_get_obj_read(pd, cmd->pd_handle, file->ucontext);
1449                 if (!pd || (!scq && has_sq)) {
1450                         ret = -EINVAL;
1451                         goto err_put;
1452                 }
1453
1454                 device = pd->device;
1455         }
1456
1457         attr.event_handler = ib_uverbs_qp_event_handler;
1458         attr.qp_context    = file;
1459         attr.send_cq       = scq;
1460         attr.recv_cq       = rcq;
1461         attr.srq           = srq;
1462         attr.xrcd          = xrcd;
1463         attr.sq_sig_type   = cmd->sq_sig_all ? IB_SIGNAL_ALL_WR :
1464                                               IB_SIGNAL_REQ_WR;
1465         attr.qp_type       = cmd->qp_type;
1466         attr.create_flags  = 0;
1467
1468         attr.cap.max_send_wr     = cmd->max_send_wr;
1469         attr.cap.max_recv_wr     = cmd->max_recv_wr;
1470         attr.cap.max_send_sge    = cmd->max_send_sge;
1471         attr.cap.max_recv_sge    = cmd->max_recv_sge;
1472         attr.cap.max_inline_data = cmd->max_inline_data;
1473
1474         obj->uevent.events_reported     = 0;
1475         INIT_LIST_HEAD(&obj->uevent.event_list);
1476         INIT_LIST_HEAD(&obj->mcast_list);
1477
1478         if (cmd_sz >= offsetof(typeof(*cmd), create_flags) +
1479                       sizeof(cmd->create_flags))
1480                 attr.create_flags = cmd->create_flags;
1481
1482         if (attr.create_flags & ~(IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
1483                                 IB_QP_CREATE_CROSS_CHANNEL |
1484                                 IB_QP_CREATE_MANAGED_SEND |
1485                                 IB_QP_CREATE_MANAGED_RECV |
1486                                 IB_QP_CREATE_SCATTER_FCS |
1487                                 IB_QP_CREATE_CVLAN_STRIPPING)) {
1488                 ret = -EINVAL;
1489                 goto err_put;
1490         }
1491
1492         buf = (void *)cmd + sizeof(*cmd);
1493         if (cmd_sz > sizeof(*cmd))
1494                 if (!(buf[0] == 0 && !memcmp(buf, buf + 1,
1495                                              cmd_sz - sizeof(*cmd) - 1))) {
1496                         ret = -EINVAL;
1497                         goto err_put;
1498                 }
1499
1500         if (cmd->qp_type == IB_QPT_XRC_TGT)
1501                 qp = ib_create_qp(pd, &attr);
1502         else
1503                 qp = device->create_qp(pd, &attr, uhw);
1504
1505         if (IS_ERR(qp)) {
1506                 ret = PTR_ERR(qp);
1507                 goto err_put;
1508         }
1509
1510         if (cmd->qp_type != IB_QPT_XRC_TGT) {
1511                 ret = ib_create_qp_security(qp, device);
1512                 if (ret)
1513                         goto err_cb;
1514
1515                 qp->real_qp       = qp;
1516                 qp->device        = device;
1517                 qp->pd            = pd;
1518                 qp->send_cq       = attr.send_cq;
1519                 qp->recv_cq       = attr.recv_cq;
1520                 qp->srq           = attr.srq;
1521                 qp->rwq_ind_tbl   = ind_tbl;
1522                 qp->event_handler = attr.event_handler;
1523                 qp->qp_context    = attr.qp_context;
1524                 qp->qp_type       = attr.qp_type;
1525                 atomic_set(&qp->usecnt, 0);
1526                 atomic_inc(&pd->usecnt);
1527                 if (attr.send_cq)
1528                         atomic_inc(&attr.send_cq->usecnt);
1529                 if (attr.recv_cq)
1530                         atomic_inc(&attr.recv_cq->usecnt);
1531                 if (attr.srq)
1532                         atomic_inc(&attr.srq->usecnt);
1533                 if (ind_tbl)
1534                         atomic_inc(&ind_tbl->usecnt);
1535         }
1536         qp->uobject = &obj->uevent.uobject;
1537
1538         obj->uevent.uobject.object = qp;
1539
1540         memset(&resp, 0, sizeof resp);
1541         resp.base.qpn             = qp->qp_num;
1542         resp.base.qp_handle       = obj->uevent.uobject.id;
1543         resp.base.max_recv_sge    = attr.cap.max_recv_sge;
1544         resp.base.max_send_sge    = attr.cap.max_send_sge;
1545         resp.base.max_recv_wr     = attr.cap.max_recv_wr;
1546         resp.base.max_send_wr     = attr.cap.max_send_wr;
1547         resp.base.max_inline_data = attr.cap.max_inline_data;
1548
1549         resp.response_length = offsetof(typeof(resp), response_length) +
1550                                sizeof(resp.response_length);
1551
1552         ret = cb(file, &resp, ucore);
1553         if (ret)
1554                 goto err_cb;
1555
1556         if (xrcd) {
1557                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object,
1558                                           uobject);
1559                 atomic_inc(&obj->uxrcd->refcnt);
1560                 uobj_put_read(xrcd_uobj);
1561         }
1562
1563         if (pd)
1564                 uobj_put_obj_read(pd);
1565         if (scq)
1566                 uobj_put_obj_read(scq);
1567         if (rcq && rcq != scq)
1568                 uobj_put_obj_read(rcq);
1569         if (srq)
1570                 uobj_put_obj_read(srq);
1571         if (ind_tbl)
1572                 uobj_put_obj_read(ind_tbl);
1573
1574         uobj_alloc_commit(&obj->uevent.uobject);
1575
1576         return 0;
1577 err_cb:
1578         ib_destroy_qp(qp);
1579
1580 err_put:
1581         if (!IS_ERR(xrcd_uobj))
1582                 uobj_put_read(xrcd_uobj);
1583         if (pd)
1584                 uobj_put_obj_read(pd);
1585         if (scq)
1586                 uobj_put_obj_read(scq);
1587         if (rcq && rcq != scq)
1588                 uobj_put_obj_read(rcq);
1589         if (srq)
1590                 uobj_put_obj_read(srq);
1591         if (ind_tbl)
1592                 uobj_put_obj_read(ind_tbl);
1593
1594         uobj_alloc_abort(&obj->uevent.uobject);
1595         return ret;
1596 }
1597
1598 static int ib_uverbs_create_qp_cb(struct ib_uverbs_file *file,
1599                                   struct ib_uverbs_ex_create_qp_resp *resp,
1600                                   struct ib_udata *ucore)
1601 {
1602         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1603                 return -EFAULT;
1604
1605         return 0;
1606 }
1607
1608 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1609                             struct ib_device *ib_dev,
1610                             const char __user *buf, int in_len,
1611                             int out_len)
1612 {
1613         struct ib_uverbs_create_qp      cmd;
1614         struct ib_uverbs_ex_create_qp   cmd_ex;
1615         struct ib_udata                 ucore;
1616         struct ib_udata                 uhw;
1617         ssize_t resp_size = sizeof(struct ib_uverbs_create_qp_resp);
1618         int                             err;
1619
1620         if (out_len < resp_size)
1621                 return -ENOSPC;
1622
1623         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1624                 return -EFAULT;
1625
1626         INIT_UDATA(&ucore, buf, (unsigned long)cmd.response, sizeof(cmd),
1627                    resp_size);
1628         INIT_UDATA(&uhw, buf + sizeof(cmd),
1629                    (unsigned long)cmd.response + resp_size,
1630                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1631                    out_len - resp_size);
1632
1633         memset(&cmd_ex, 0, sizeof(cmd_ex));
1634         cmd_ex.user_handle = cmd.user_handle;
1635         cmd_ex.pd_handle = cmd.pd_handle;
1636         cmd_ex.send_cq_handle = cmd.send_cq_handle;
1637         cmd_ex.recv_cq_handle = cmd.recv_cq_handle;
1638         cmd_ex.srq_handle = cmd.srq_handle;
1639         cmd_ex.max_send_wr = cmd.max_send_wr;
1640         cmd_ex.max_recv_wr = cmd.max_recv_wr;
1641         cmd_ex.max_send_sge = cmd.max_send_sge;
1642         cmd_ex.max_recv_sge = cmd.max_recv_sge;
1643         cmd_ex.max_inline_data = cmd.max_inline_data;
1644         cmd_ex.sq_sig_all = cmd.sq_sig_all;
1645         cmd_ex.qp_type = cmd.qp_type;
1646         cmd_ex.is_srq = cmd.is_srq;
1647
1648         err = create_qp(file, &ucore, &uhw, &cmd_ex,
1649                         offsetof(typeof(cmd_ex), is_srq) +
1650                         sizeof(cmd.is_srq), ib_uverbs_create_qp_cb,
1651                         NULL);
1652
1653         if (err)
1654                 return err;
1655
1656         return in_len;
1657 }
1658
1659 static int ib_uverbs_ex_create_qp_cb(struct ib_uverbs_file *file,
1660                                      struct ib_uverbs_ex_create_qp_resp *resp,
1661                                      struct ib_udata *ucore)
1662 {
1663         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1664                 return -EFAULT;
1665
1666         return 0;
1667 }
1668
1669 int ib_uverbs_ex_create_qp(struct ib_uverbs_file *file,
1670                            struct ib_device *ib_dev,
1671                            struct ib_udata *ucore,
1672                            struct ib_udata *uhw)
1673 {
1674         struct ib_uverbs_ex_create_qp_resp resp;
1675         struct ib_uverbs_ex_create_qp cmd = {0};
1676         int err;
1677
1678         if (ucore->inlen < (offsetof(typeof(cmd), comp_mask) +
1679                             sizeof(cmd.comp_mask)))
1680                 return -EINVAL;
1681
1682         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
1683         if (err)
1684                 return err;
1685
1686         if (cmd.comp_mask & ~IB_UVERBS_CREATE_QP_SUP_COMP_MASK)
1687                 return -EINVAL;
1688
1689         if (cmd.reserved)
1690                 return -EINVAL;
1691
1692         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1693                              sizeof(resp.response_length)))
1694                 return -ENOSPC;
1695
1696         err = create_qp(file, ucore, uhw, &cmd,
1697                         min(ucore->inlen, sizeof(cmd)),
1698                         ib_uverbs_ex_create_qp_cb, NULL);
1699
1700         if (err)
1701                 return err;
1702
1703         return 0;
1704 }
1705
1706 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1707                           struct ib_device *ib_dev,
1708                           const char __user *buf, int in_len, int out_len)
1709 {
1710         struct ib_uverbs_open_qp        cmd;
1711         struct ib_uverbs_create_qp_resp resp;
1712         struct ib_udata                 udata;
1713         struct ib_uqp_object           *obj;
1714         struct ib_xrcd                 *xrcd;
1715         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1716         struct ib_qp                   *qp;
1717         struct ib_qp_open_attr          attr;
1718         int ret;
1719
1720         if (out_len < sizeof resp)
1721                 return -ENOSPC;
1722
1723         if (copy_from_user(&cmd, buf, sizeof cmd))
1724                 return -EFAULT;
1725
1726         INIT_UDATA(&udata, buf + sizeof cmd,
1727                    (unsigned long) cmd.response + sizeof resp,
1728                    in_len - sizeof cmd, out_len - sizeof resp);
1729
1730         obj  = (struct ib_uqp_object *)uobj_alloc(uobj_get_type(qp),
1731                                                   file->ucontext);
1732         if (IS_ERR(obj))
1733                 return PTR_ERR(obj);
1734
1735         xrcd_uobj = uobj_get_read(uobj_get_type(xrcd), cmd.pd_handle,
1736                                   file->ucontext);
1737         if (IS_ERR(xrcd_uobj)) {
1738                 ret = -EINVAL;
1739                 goto err_put;
1740         }
1741
1742         xrcd = (struct ib_xrcd *)xrcd_uobj->object;
1743         if (!xrcd) {
1744                 ret = -EINVAL;
1745                 goto err_xrcd;
1746         }
1747
1748         attr.event_handler = ib_uverbs_qp_event_handler;
1749         attr.qp_context    = file;
1750         attr.qp_num        = cmd.qpn;
1751         attr.qp_type       = cmd.qp_type;
1752
1753         obj->uevent.events_reported = 0;
1754         INIT_LIST_HEAD(&obj->uevent.event_list);
1755         INIT_LIST_HEAD(&obj->mcast_list);
1756
1757         qp = ib_open_qp(xrcd, &attr);
1758         if (IS_ERR(qp)) {
1759                 ret = PTR_ERR(qp);
1760                 goto err_xrcd;
1761         }
1762
1763         obj->uevent.uobject.object = qp;
1764         obj->uevent.uobject.user_handle = cmd.user_handle;
1765
1766         memset(&resp, 0, sizeof resp);
1767         resp.qpn       = qp->qp_num;
1768         resp.qp_handle = obj->uevent.uobject.id;
1769
1770         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1771                          &resp, sizeof resp)) {
1772                 ret = -EFAULT;
1773                 goto err_destroy;
1774         }
1775
1776         obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
1777         atomic_inc(&obj->uxrcd->refcnt);
1778         qp->uobject = &obj->uevent.uobject;
1779         uobj_put_read(xrcd_uobj);
1780
1781
1782         uobj_alloc_commit(&obj->uevent.uobject);
1783
1784         return in_len;
1785
1786 err_destroy:
1787         ib_destroy_qp(qp);
1788 err_xrcd:
1789         uobj_put_read(xrcd_uobj);
1790 err_put:
1791         uobj_alloc_abort(&obj->uevent.uobject);
1792         return ret;
1793 }
1794
1795 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1796                            struct ib_device *ib_dev,
1797                            const char __user *buf, int in_len,
1798                            int out_len)
1799 {
1800         struct ib_uverbs_query_qp      cmd;
1801         struct ib_uverbs_query_qp_resp resp;
1802         struct ib_qp                   *qp;
1803         struct ib_qp_attr              *attr;
1804         struct ib_qp_init_attr         *init_attr;
1805         const struct ib_global_route   *grh;
1806         int                            ret;
1807
1808         if (copy_from_user(&cmd, buf, sizeof cmd))
1809                 return -EFAULT;
1810
1811         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1812         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1813         if (!attr || !init_attr) {
1814                 ret = -ENOMEM;
1815                 goto out;
1816         }
1817
1818         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
1819         if (!qp) {
1820                 ret = -EINVAL;
1821                 goto out;
1822         }
1823
1824         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1825
1826         uobj_put_obj_read(qp);
1827
1828         if (ret)
1829                 goto out;
1830
1831         memset(&resp, 0, sizeof resp);
1832
1833         resp.qp_state               = attr->qp_state;
1834         resp.cur_qp_state           = attr->cur_qp_state;
1835         resp.path_mtu               = attr->path_mtu;
1836         resp.path_mig_state         = attr->path_mig_state;
1837         resp.qkey                   = attr->qkey;
1838         resp.rq_psn                 = attr->rq_psn;
1839         resp.sq_psn                 = attr->sq_psn;
1840         resp.dest_qp_num            = attr->dest_qp_num;
1841         resp.qp_access_flags        = attr->qp_access_flags;
1842         resp.pkey_index             = attr->pkey_index;
1843         resp.alt_pkey_index         = attr->alt_pkey_index;
1844         resp.sq_draining            = attr->sq_draining;
1845         resp.max_rd_atomic          = attr->max_rd_atomic;
1846         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1847         resp.min_rnr_timer          = attr->min_rnr_timer;
1848         resp.port_num               = attr->port_num;
1849         resp.timeout                = attr->timeout;
1850         resp.retry_cnt              = attr->retry_cnt;
1851         resp.rnr_retry              = attr->rnr_retry;
1852         resp.alt_port_num           = attr->alt_port_num;
1853         resp.alt_timeout            = attr->alt_timeout;
1854
1855         resp.dest.dlid              = rdma_ah_get_dlid(&attr->ah_attr);
1856         resp.dest.sl                = rdma_ah_get_sl(&attr->ah_attr);
1857         resp.dest.src_path_bits     = rdma_ah_get_path_bits(&attr->ah_attr);
1858         resp.dest.static_rate       = rdma_ah_get_static_rate(&attr->ah_attr);
1859         resp.dest.is_global         = !!(rdma_ah_get_ah_flags(&attr->ah_attr) &
1860                                          IB_AH_GRH);
1861         if (resp.dest.is_global) {
1862                 grh = rdma_ah_read_grh(&attr->ah_attr);
1863                 memcpy(resp.dest.dgid, grh->dgid.raw, 16);
1864                 resp.dest.flow_label        = grh->flow_label;
1865                 resp.dest.sgid_index        = grh->sgid_index;
1866                 resp.dest.hop_limit         = grh->hop_limit;
1867                 resp.dest.traffic_class     = grh->traffic_class;
1868         }
1869         resp.dest.port_num          = rdma_ah_get_port_num(&attr->ah_attr);
1870
1871         resp.alt_dest.dlid          = rdma_ah_get_dlid(&attr->alt_ah_attr);
1872         resp.alt_dest.sl            = rdma_ah_get_sl(&attr->alt_ah_attr);
1873         resp.alt_dest.src_path_bits = rdma_ah_get_path_bits(&attr->alt_ah_attr);
1874         resp.alt_dest.static_rate
1875                         = rdma_ah_get_static_rate(&attr->alt_ah_attr);
1876         resp.alt_dest.is_global
1877                         = !!(rdma_ah_get_ah_flags(&attr->alt_ah_attr) &
1878                                                   IB_AH_GRH);
1879         if (resp.alt_dest.is_global) {
1880                 grh = rdma_ah_read_grh(&attr->alt_ah_attr);
1881                 memcpy(resp.alt_dest.dgid, grh->dgid.raw, 16);
1882                 resp.alt_dest.flow_label    = grh->flow_label;
1883                 resp.alt_dest.sgid_index    = grh->sgid_index;
1884                 resp.alt_dest.hop_limit     = grh->hop_limit;
1885                 resp.alt_dest.traffic_class = grh->traffic_class;
1886         }
1887         resp.alt_dest.port_num      = rdma_ah_get_port_num(&attr->alt_ah_attr);
1888
1889         resp.max_send_wr            = init_attr->cap.max_send_wr;
1890         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1891         resp.max_send_sge           = init_attr->cap.max_send_sge;
1892         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1893         resp.max_inline_data        = init_attr->cap.max_inline_data;
1894         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1895
1896         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1897                          &resp, sizeof resp))
1898                 ret = -EFAULT;
1899
1900 out:
1901         kfree(attr);
1902         kfree(init_attr);
1903
1904         return ret ? ret : in_len;
1905 }
1906
1907 /* Remove ignored fields set in the attribute mask */
1908 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1909 {
1910         switch (qp_type) {
1911         case IB_QPT_XRC_INI:
1912                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1913         case IB_QPT_XRC_TGT:
1914                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1915                                 IB_QP_RNR_RETRY);
1916         default:
1917                 return mask;
1918         }
1919 }
1920
1921 static int modify_qp(struct ib_uverbs_file *file,
1922                      struct ib_uverbs_ex_modify_qp *cmd, struct ib_udata *udata)
1923 {
1924         struct ib_qp_attr *attr;
1925         struct ib_qp *qp;
1926         int ret;
1927
1928         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1929         if (!attr)
1930                 return -ENOMEM;
1931
1932         qp = uobj_get_obj_read(qp, cmd->base.qp_handle, file->ucontext);
1933         if (!qp) {
1934                 ret = -EINVAL;
1935                 goto out;
1936         }
1937
1938         if (!rdma_is_port_valid(qp->device, cmd->base.port_num)) {
1939                 ret = -EINVAL;
1940                 goto release_qp;
1941         }
1942
1943         attr->qp_state            = cmd->base.qp_state;
1944         attr->cur_qp_state        = cmd->base.cur_qp_state;
1945         attr->path_mtu            = cmd->base.path_mtu;
1946         attr->path_mig_state      = cmd->base.path_mig_state;
1947         attr->qkey                = cmd->base.qkey;
1948         attr->rq_psn              = cmd->base.rq_psn;
1949         attr->sq_psn              = cmd->base.sq_psn;
1950         attr->dest_qp_num         = cmd->base.dest_qp_num;
1951         attr->qp_access_flags     = cmd->base.qp_access_flags;
1952         attr->pkey_index          = cmd->base.pkey_index;
1953         attr->alt_pkey_index      = cmd->base.alt_pkey_index;
1954         attr->en_sqd_async_notify = cmd->base.en_sqd_async_notify;
1955         attr->max_rd_atomic       = cmd->base.max_rd_atomic;
1956         attr->max_dest_rd_atomic  = cmd->base.max_dest_rd_atomic;
1957         attr->min_rnr_timer       = cmd->base.min_rnr_timer;
1958         attr->port_num            = cmd->base.port_num;
1959         attr->timeout             = cmd->base.timeout;
1960         attr->retry_cnt           = cmd->base.retry_cnt;
1961         attr->rnr_retry           = cmd->base.rnr_retry;
1962         attr->alt_port_num        = cmd->base.alt_port_num;
1963         attr->alt_timeout         = cmd->base.alt_timeout;
1964         attr->rate_limit          = cmd->rate_limit;
1965
1966         attr->ah_attr.type = rdma_ah_find_type(qp->device,
1967                                                cmd->base.dest.port_num);
1968         if (cmd->base.dest.is_global) {
1969                 rdma_ah_set_grh(&attr->ah_attr, NULL,
1970                                 cmd->base.dest.flow_label,
1971                                 cmd->base.dest.sgid_index,
1972                                 cmd->base.dest.hop_limit,
1973                                 cmd->base.dest.traffic_class);
1974                 rdma_ah_set_dgid_raw(&attr->ah_attr, cmd->base.dest.dgid);
1975         } else {
1976                 rdma_ah_set_ah_flags(&attr->ah_attr, 0);
1977         }
1978         rdma_ah_set_dlid(&attr->ah_attr, cmd->base.dest.dlid);
1979         rdma_ah_set_sl(&attr->ah_attr, cmd->base.dest.sl);
1980         rdma_ah_set_path_bits(&attr->ah_attr, cmd->base.dest.src_path_bits);
1981         rdma_ah_set_static_rate(&attr->ah_attr, cmd->base.dest.static_rate);
1982         rdma_ah_set_port_num(&attr->ah_attr,
1983                              cmd->base.dest.port_num);
1984
1985         attr->alt_ah_attr.type = rdma_ah_find_type(qp->device,
1986                                                    cmd->base.dest.port_num);
1987         if (cmd->base.alt_dest.is_global) {
1988                 rdma_ah_set_grh(&attr->alt_ah_attr, NULL,
1989                                 cmd->base.alt_dest.flow_label,
1990                                 cmd->base.alt_dest.sgid_index,
1991                                 cmd->base.alt_dest.hop_limit,
1992                                 cmd->base.alt_dest.traffic_class);
1993                 rdma_ah_set_dgid_raw(&attr->alt_ah_attr,
1994                                      cmd->base.alt_dest.dgid);
1995         } else {
1996                 rdma_ah_set_ah_flags(&attr->alt_ah_attr, 0);
1997         }
1998
1999         rdma_ah_set_dlid(&attr->alt_ah_attr, cmd->base.alt_dest.dlid);
2000         rdma_ah_set_sl(&attr->alt_ah_attr, cmd->base.alt_dest.sl);
2001         rdma_ah_set_path_bits(&attr->alt_ah_attr,
2002                               cmd->base.alt_dest.src_path_bits);
2003         rdma_ah_set_static_rate(&attr->alt_ah_attr,
2004                                 cmd->base.alt_dest.static_rate);
2005         rdma_ah_set_port_num(&attr->alt_ah_attr,
2006                              cmd->base.alt_dest.port_num);
2007
2008         if (qp->real_qp == qp) {
2009                 if (cmd->base.attr_mask & IB_QP_AV) {
2010                         ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
2011                         if (ret)
2012                                 goto release_qp;
2013                 }
2014                 ret = ib_security_modify_qp(qp,
2015                                             attr,
2016                                             modify_qp_mask(qp->qp_type,
2017                                                            cmd->base.attr_mask),
2018                                             udata);
2019         } else {
2020                 ret = ib_security_modify_qp(qp,
2021                                             attr,
2022                                             modify_qp_mask(qp->qp_type,
2023                                                            cmd->base.attr_mask),
2024                                             NULL);
2025         }
2026
2027 release_qp:
2028         uobj_put_obj_read(qp);
2029
2030 out:
2031         kfree(attr);
2032
2033         return ret;
2034 }
2035
2036 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
2037                             struct ib_device *ib_dev,
2038                             const char __user *buf, int in_len,
2039                             int out_len)
2040 {
2041         struct ib_uverbs_ex_modify_qp cmd = {};
2042         struct ib_udata udata;
2043         int ret;
2044
2045         if (copy_from_user(&cmd.base, buf, sizeof(cmd.base)))
2046                 return -EFAULT;
2047
2048         if (cmd.base.attr_mask &
2049             ~((IB_USER_LEGACY_LAST_QP_ATTR_MASK << 1) - 1))
2050                 return -EOPNOTSUPP;
2051
2052         INIT_UDATA(&udata, buf + sizeof(cmd.base), NULL,
2053                    in_len - sizeof(cmd.base), out_len);
2054
2055         ret = modify_qp(file, &cmd, &udata);
2056         if (ret)
2057                 return ret;
2058
2059         return in_len;
2060 }
2061
2062 int ib_uverbs_ex_modify_qp(struct ib_uverbs_file *file,
2063                            struct ib_device *ib_dev,
2064                            struct ib_udata *ucore,
2065                            struct ib_udata *uhw)
2066 {
2067         struct ib_uverbs_ex_modify_qp cmd = {};
2068         int ret;
2069
2070         /*
2071          * Last bit is reserved for extending the attr_mask by
2072          * using another field.
2073          */
2074         BUILD_BUG_ON(IB_USER_LAST_QP_ATTR_MASK == (1 << 31));
2075
2076         if (ucore->inlen < sizeof(cmd.base))
2077                 return -EINVAL;
2078
2079         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2080         if (ret)
2081                 return ret;
2082
2083         if (cmd.base.attr_mask &
2084             ~((IB_USER_LAST_QP_ATTR_MASK << 1) - 1))
2085                 return -EOPNOTSUPP;
2086
2087         if (ucore->inlen > sizeof(cmd)) {
2088                 if (ib_is_udata_cleared(ucore, sizeof(cmd),
2089                                         ucore->inlen - sizeof(cmd)))
2090                         return -EOPNOTSUPP;
2091         }
2092
2093         ret = modify_qp(file, &cmd, uhw);
2094
2095         return ret;
2096 }
2097
2098 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
2099                              struct ib_device *ib_dev,
2100                              const char __user *buf, int in_len,
2101                              int out_len)
2102 {
2103         struct ib_uverbs_destroy_qp      cmd;
2104         struct ib_uverbs_destroy_qp_resp resp;
2105         struct ib_uobject               *uobj;
2106         struct ib_qp                    *qp;
2107         struct ib_uqp_object            *obj;
2108         int                              ret = -EINVAL;
2109
2110         if (copy_from_user(&cmd, buf, sizeof cmd))
2111                 return -EFAULT;
2112
2113         memset(&resp, 0, sizeof resp);
2114
2115         uobj  = uobj_get_write(uobj_get_type(qp), cmd.qp_handle,
2116                                file->ucontext);
2117         if (IS_ERR(uobj))
2118                 return PTR_ERR(uobj);
2119
2120         qp  = uobj->object;
2121         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2122         /*
2123          * Make sure we don't free the memory in remove_commit as we still
2124          * needs the uobject memory to create the response.
2125          */
2126         uverbs_uobject_get(uobj);
2127
2128         ret = uobj_remove_commit(uobj);
2129         if (ret) {
2130                 uverbs_uobject_put(uobj);
2131                 return ret;
2132         }
2133
2134         resp.events_reported = obj->uevent.events_reported;
2135         uverbs_uobject_put(uobj);
2136
2137         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2138                          &resp, sizeof resp))
2139                 return -EFAULT;
2140
2141         return in_len;
2142 }
2143
2144 static void *alloc_wr(size_t wr_size, __u32 num_sge)
2145 {
2146         if (num_sge >= (U32_MAX - ALIGN(wr_size, sizeof (struct ib_sge))) /
2147                        sizeof (struct ib_sge))
2148                 return NULL;
2149
2150         return kmalloc(ALIGN(wr_size, sizeof (struct ib_sge)) +
2151                          num_sge * sizeof (struct ib_sge), GFP_KERNEL);
2152 }
2153
2154 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2155                             struct ib_device *ib_dev,
2156                             const char __user *buf, int in_len,
2157                             int out_len)
2158 {
2159         struct ib_uverbs_post_send      cmd;
2160         struct ib_uverbs_post_send_resp resp;
2161         struct ib_uverbs_send_wr       *user_wr;
2162         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
2163         struct ib_qp                   *qp;
2164         int                             i, sg_ind;
2165         int                             is_ud;
2166         ssize_t                         ret = -EINVAL;
2167         size_t                          next_size;
2168
2169         if (copy_from_user(&cmd, buf, sizeof cmd))
2170                 return -EFAULT;
2171
2172         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2173             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2174                 return -EINVAL;
2175
2176         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2177                 return -EINVAL;
2178
2179         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2180         if (!user_wr)
2181                 return -ENOMEM;
2182
2183         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
2184         if (!qp)
2185                 goto out;
2186
2187         is_ud = qp->qp_type == IB_QPT_UD;
2188         sg_ind = 0;
2189         last = NULL;
2190         for (i = 0; i < cmd.wr_count; ++i) {
2191                 if (copy_from_user(user_wr,
2192                                    buf + sizeof cmd + i * cmd.wqe_size,
2193                                    cmd.wqe_size)) {
2194                         ret = -EFAULT;
2195                         goto out_put;
2196                 }
2197
2198                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2199                         ret = -EINVAL;
2200                         goto out_put;
2201                 }
2202
2203                 if (is_ud) {
2204                         struct ib_ud_wr *ud;
2205
2206                         if (user_wr->opcode != IB_WR_SEND &&
2207                             user_wr->opcode != IB_WR_SEND_WITH_IMM) {
2208                                 ret = -EINVAL;
2209                                 goto out_put;
2210                         }
2211
2212                         next_size = sizeof(*ud);
2213                         ud = alloc_wr(next_size, user_wr->num_sge);
2214                         if (!ud) {
2215                                 ret = -ENOMEM;
2216                                 goto out_put;
2217                         }
2218
2219                         ud->ah = uobj_get_obj_read(ah, user_wr->wr.ud.ah,
2220                                                    file->ucontext);
2221                         if (!ud->ah) {
2222                                 kfree(ud);
2223                                 ret = -EINVAL;
2224                                 goto out_put;
2225                         }
2226                         ud->remote_qpn = user_wr->wr.ud.remote_qpn;
2227                         ud->remote_qkey = user_wr->wr.ud.remote_qkey;
2228
2229                         next = &ud->wr;
2230                 } else if (user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
2231                            user_wr->opcode == IB_WR_RDMA_WRITE ||
2232                            user_wr->opcode == IB_WR_RDMA_READ) {
2233                         struct ib_rdma_wr *rdma;
2234
2235                         next_size = sizeof(*rdma);
2236                         rdma = alloc_wr(next_size, user_wr->num_sge);
2237                         if (!rdma) {
2238                                 ret = -ENOMEM;
2239                                 goto out_put;
2240                         }
2241
2242                         rdma->remote_addr = user_wr->wr.rdma.remote_addr;
2243                         rdma->rkey = user_wr->wr.rdma.rkey;
2244
2245                         next = &rdma->wr;
2246                 } else if (user_wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2247                            user_wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2248                         struct ib_atomic_wr *atomic;
2249
2250                         next_size = sizeof(*atomic);
2251                         atomic = alloc_wr(next_size, user_wr->num_sge);
2252                         if (!atomic) {
2253                                 ret = -ENOMEM;
2254                                 goto out_put;
2255                         }
2256
2257                         atomic->remote_addr = user_wr->wr.atomic.remote_addr;
2258                         atomic->compare_add = user_wr->wr.atomic.compare_add;
2259                         atomic->swap = user_wr->wr.atomic.swap;
2260                         atomic->rkey = user_wr->wr.atomic.rkey;
2261
2262                         next = &atomic->wr;
2263                 } else if (user_wr->opcode == IB_WR_SEND ||
2264                            user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2265                            user_wr->opcode == IB_WR_SEND_WITH_INV) {
2266                         next_size = sizeof(*next);
2267                         next = alloc_wr(next_size, user_wr->num_sge);
2268                         if (!next) {
2269                                 ret = -ENOMEM;
2270                                 goto out_put;
2271                         }
2272                 } else {
2273                         ret = -EINVAL;
2274                         goto out_put;
2275                 }
2276
2277                 if (user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2278                     user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
2279                         next->ex.imm_data =
2280                                         (__be32 __force) user_wr->ex.imm_data;
2281                 } else if (user_wr->opcode == IB_WR_SEND_WITH_INV) {
2282                         next->ex.invalidate_rkey = user_wr->ex.invalidate_rkey;
2283                 }
2284
2285                 if (!last)
2286                         wr = next;
2287                 else
2288                         last->next = next;
2289                 last = next;
2290
2291                 next->next       = NULL;
2292                 next->wr_id      = user_wr->wr_id;
2293                 next->num_sge    = user_wr->num_sge;
2294                 next->opcode     = user_wr->opcode;
2295                 next->send_flags = user_wr->send_flags;
2296
2297                 if (next->num_sge) {
2298                         next->sg_list = (void *) next +
2299                                 ALIGN(next_size, sizeof(struct ib_sge));
2300                         if (copy_from_user(next->sg_list,
2301                                            buf + sizeof cmd +
2302                                            cmd.wr_count * cmd.wqe_size +
2303                                            sg_ind * sizeof (struct ib_sge),
2304                                            next->num_sge * sizeof (struct ib_sge))) {
2305                                 ret = -EFAULT;
2306                                 goto out_put;
2307                         }
2308                         sg_ind += next->num_sge;
2309                 } else
2310                         next->sg_list = NULL;
2311         }
2312
2313         resp.bad_wr = 0;
2314         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2315         if (ret)
2316                 for (next = wr; next; next = next->next) {
2317                         ++resp.bad_wr;
2318                         if (next == bad_wr)
2319                                 break;
2320                 }
2321
2322         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2323                          &resp, sizeof resp))
2324                 ret = -EFAULT;
2325
2326 out_put:
2327         uobj_put_obj_read(qp);
2328
2329         while (wr) {
2330                 if (is_ud && ud_wr(wr)->ah)
2331                         uobj_put_obj_read(ud_wr(wr)->ah);
2332                 next = wr->next;
2333                 kfree(wr);
2334                 wr = next;
2335         }
2336
2337 out:
2338         kfree(user_wr);
2339
2340         return ret ? ret : in_len;
2341 }
2342
2343 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2344                                                     int in_len,
2345                                                     u32 wr_count,
2346                                                     u32 sge_count,
2347                                                     u32 wqe_size)
2348 {
2349         struct ib_uverbs_recv_wr *user_wr;
2350         struct ib_recv_wr        *wr = NULL, *last, *next;
2351         int                       sg_ind;
2352         int                       i;
2353         int                       ret;
2354
2355         if (in_len < wqe_size * wr_count +
2356             sge_count * sizeof (struct ib_uverbs_sge))
2357                 return ERR_PTR(-EINVAL);
2358
2359         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2360                 return ERR_PTR(-EINVAL);
2361
2362         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2363         if (!user_wr)
2364                 return ERR_PTR(-ENOMEM);
2365
2366         sg_ind = 0;
2367         last = NULL;
2368         for (i = 0; i < wr_count; ++i) {
2369                 if (copy_from_user(user_wr, buf + i * wqe_size,
2370                                    wqe_size)) {
2371                         ret = -EFAULT;
2372                         goto err;
2373                 }
2374
2375                 if (user_wr->num_sge + sg_ind > sge_count) {
2376                         ret = -EINVAL;
2377                         goto err;
2378                 }
2379
2380                 if (user_wr->num_sge >=
2381                     (U32_MAX - ALIGN(sizeof *next, sizeof (struct ib_sge))) /
2382                     sizeof (struct ib_sge)) {
2383                         ret = -EINVAL;
2384                         goto err;
2385                 }
2386
2387                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2388                                user_wr->num_sge * sizeof (struct ib_sge),
2389                                GFP_KERNEL);
2390                 if (!next) {
2391                         ret = -ENOMEM;
2392                         goto err;
2393                 }
2394
2395                 if (!last)
2396                         wr = next;
2397                 else
2398                         last->next = next;
2399                 last = next;
2400
2401                 next->next       = NULL;
2402                 next->wr_id      = user_wr->wr_id;
2403                 next->num_sge    = user_wr->num_sge;
2404
2405                 if (next->num_sge) {
2406                         next->sg_list = (void *) next +
2407                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2408                         if (copy_from_user(next->sg_list,
2409                                            buf + wr_count * wqe_size +
2410                                            sg_ind * sizeof (struct ib_sge),
2411                                            next->num_sge * sizeof (struct ib_sge))) {
2412                                 ret = -EFAULT;
2413                                 goto err;
2414                         }
2415                         sg_ind += next->num_sge;
2416                 } else
2417                         next->sg_list = NULL;
2418         }
2419
2420         kfree(user_wr);
2421         return wr;
2422
2423 err:
2424         kfree(user_wr);
2425
2426         while (wr) {
2427                 next = wr->next;
2428                 kfree(wr);
2429                 wr = next;
2430         }
2431
2432         return ERR_PTR(ret);
2433 }
2434
2435 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2436                             struct ib_device *ib_dev,
2437                             const char __user *buf, int in_len,
2438                             int out_len)
2439 {
2440         struct ib_uverbs_post_recv      cmd;
2441         struct ib_uverbs_post_recv_resp resp;
2442         struct ib_recv_wr              *wr, *next, *bad_wr;
2443         struct ib_qp                   *qp;
2444         ssize_t                         ret = -EINVAL;
2445
2446         if (copy_from_user(&cmd, buf, sizeof cmd))
2447                 return -EFAULT;
2448
2449         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2450                                        in_len - sizeof cmd, cmd.wr_count,
2451                                        cmd.sge_count, cmd.wqe_size);
2452         if (IS_ERR(wr))
2453                 return PTR_ERR(wr);
2454
2455         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
2456         if (!qp)
2457                 goto out;
2458
2459         resp.bad_wr = 0;
2460         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2461
2462         uobj_put_obj_read(qp);
2463         if (ret) {
2464                 for (next = wr; next; next = next->next) {
2465                         ++resp.bad_wr;
2466                         if (next == bad_wr)
2467                                 break;
2468                 }
2469         }
2470
2471         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2472                          &resp, sizeof resp))
2473                 ret = -EFAULT;
2474
2475 out:
2476         while (wr) {
2477                 next = wr->next;
2478                 kfree(wr);
2479                 wr = next;
2480         }
2481
2482         return ret ? ret : in_len;
2483 }
2484
2485 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2486                                 struct ib_device *ib_dev,
2487                                 const char __user *buf, int in_len,
2488                                 int out_len)
2489 {
2490         struct ib_uverbs_post_srq_recv      cmd;
2491         struct ib_uverbs_post_srq_recv_resp resp;
2492         struct ib_recv_wr                  *wr, *next, *bad_wr;
2493         struct ib_srq                      *srq;
2494         ssize_t                             ret = -EINVAL;
2495
2496         if (copy_from_user(&cmd, buf, sizeof cmd))
2497                 return -EFAULT;
2498
2499         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2500                                        in_len - sizeof cmd, cmd.wr_count,
2501                                        cmd.sge_count, cmd.wqe_size);
2502         if (IS_ERR(wr))
2503                 return PTR_ERR(wr);
2504
2505         srq = uobj_get_obj_read(srq, cmd.srq_handle, file->ucontext);
2506         if (!srq)
2507                 goto out;
2508
2509         resp.bad_wr = 0;
2510         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2511
2512         uobj_put_obj_read(srq);
2513
2514         if (ret)
2515                 for (next = wr; next; next = next->next) {
2516                         ++resp.bad_wr;
2517                         if (next == bad_wr)
2518                                 break;
2519                 }
2520
2521         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2522                          &resp, sizeof resp))
2523                 ret = -EFAULT;
2524
2525 out:
2526         while (wr) {
2527                 next = wr->next;
2528                 kfree(wr);
2529                 wr = next;
2530         }
2531
2532         return ret ? ret : in_len;
2533 }
2534
2535 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2536                             struct ib_device *ib_dev,
2537                             const char __user *buf, int in_len,
2538                             int out_len)
2539 {
2540         struct ib_uverbs_create_ah       cmd;
2541         struct ib_uverbs_create_ah_resp  resp;
2542         struct ib_uobject               *uobj;
2543         struct ib_pd                    *pd;
2544         struct ib_ah                    *ah;
2545         struct rdma_ah_attr             attr;
2546         int ret;
2547         struct ib_udata                   udata;
2548         u8                              *dmac;
2549
2550         if (out_len < sizeof resp)
2551                 return -ENOSPC;
2552
2553         if (copy_from_user(&cmd, buf, sizeof cmd))
2554                 return -EFAULT;
2555
2556         if (!rdma_is_port_valid(ib_dev, cmd.attr.port_num))
2557                 return -EINVAL;
2558
2559         INIT_UDATA(&udata, buf + sizeof(cmd),
2560                    (unsigned long)cmd.response + sizeof(resp),
2561                    in_len - sizeof(cmd), out_len - sizeof(resp));
2562
2563         uobj  = uobj_alloc(uobj_get_type(ah), file->ucontext);
2564         if (IS_ERR(uobj))
2565                 return PTR_ERR(uobj);
2566
2567         pd = uobj_get_obj_read(pd, cmd.pd_handle, file->ucontext);
2568         if (!pd) {
2569                 ret = -EINVAL;
2570                 goto err;
2571         }
2572
2573         attr.type = rdma_ah_find_type(ib_dev, cmd.attr.port_num);
2574         rdma_ah_set_dlid(&attr, cmd.attr.dlid);
2575         rdma_ah_set_sl(&attr, cmd.attr.sl);
2576         rdma_ah_set_path_bits(&attr, cmd.attr.src_path_bits);
2577         rdma_ah_set_static_rate(&attr, cmd.attr.static_rate);
2578         rdma_ah_set_port_num(&attr, cmd.attr.port_num);
2579
2580         if (cmd.attr.is_global) {
2581                 rdma_ah_set_grh(&attr, NULL, cmd.attr.grh.flow_label,
2582                                 cmd.attr.grh.sgid_index,
2583                                 cmd.attr.grh.hop_limit,
2584                                 cmd.attr.grh.traffic_class);
2585                 rdma_ah_set_dgid_raw(&attr, cmd.attr.grh.dgid);
2586         } else {
2587                 rdma_ah_set_ah_flags(&attr, 0);
2588         }
2589         dmac = rdma_ah_retrieve_dmac(&attr);
2590         if (dmac)
2591                 memset(dmac, 0, ETH_ALEN);
2592
2593         ah = pd->device->create_ah(pd, &attr, &udata);
2594
2595         if (IS_ERR(ah)) {
2596                 ret = PTR_ERR(ah);
2597                 goto err_put;
2598         }
2599
2600         ah->device  = pd->device;
2601         ah->pd      = pd;
2602         atomic_inc(&pd->usecnt);
2603         ah->uobject  = uobj;
2604         uobj->user_handle = cmd.user_handle;
2605         uobj->object = ah;
2606
2607         resp.ah_handle = uobj->id;
2608
2609         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2610                          &resp, sizeof resp)) {
2611                 ret = -EFAULT;
2612                 goto err_copy;
2613         }
2614
2615         uobj_put_obj_read(pd);
2616         uobj_alloc_commit(uobj);
2617
2618         return in_len;
2619
2620 err_copy:
2621         rdma_destroy_ah(ah);
2622
2623 err_put:
2624         uobj_put_obj_read(pd);
2625
2626 err:
2627         uobj_alloc_abort(uobj);
2628         return ret;
2629 }
2630
2631 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2632                              struct ib_device *ib_dev,
2633                              const char __user *buf, int in_len, int out_len)
2634 {
2635         struct ib_uverbs_destroy_ah cmd;
2636         struct ib_uobject          *uobj;
2637         int                         ret;
2638
2639         if (copy_from_user(&cmd, buf, sizeof cmd))
2640                 return -EFAULT;
2641
2642         uobj  = uobj_get_write(uobj_get_type(ah), cmd.ah_handle,
2643                                file->ucontext);
2644         if (IS_ERR(uobj))
2645                 return PTR_ERR(uobj);
2646
2647         ret = uobj_remove_commit(uobj);
2648         return ret ?: in_len;
2649 }
2650
2651 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2652                                struct ib_device *ib_dev,
2653                                const char __user *buf, int in_len,
2654                                int out_len)
2655 {
2656         struct ib_uverbs_attach_mcast cmd;
2657         struct ib_qp                 *qp;
2658         struct ib_uqp_object         *obj;
2659         struct ib_uverbs_mcast_entry *mcast;
2660         int                           ret;
2661
2662         if (copy_from_user(&cmd, buf, sizeof cmd))
2663                 return -EFAULT;
2664
2665         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
2666         if (!qp)
2667                 return -EINVAL;
2668
2669         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2670
2671         mutex_lock(&obj->mcast_lock);
2672         list_for_each_entry(mcast, &obj->mcast_list, list)
2673                 if (cmd.mlid == mcast->lid &&
2674                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2675                         ret = 0;
2676                         goto out_put;
2677                 }
2678
2679         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2680         if (!mcast) {
2681                 ret = -ENOMEM;
2682                 goto out_put;
2683         }
2684
2685         mcast->lid = cmd.mlid;
2686         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2687
2688         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2689         if (!ret)
2690                 list_add_tail(&mcast->list, &obj->mcast_list);
2691         else
2692                 kfree(mcast);
2693
2694 out_put:
2695         mutex_unlock(&obj->mcast_lock);
2696         uobj_put_obj_read(qp);
2697
2698         return ret ? ret : in_len;
2699 }
2700
2701 ssize_t ib_uverbs_detach_mcast(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_detach_mcast cmd;
2707         struct ib_uqp_object         *obj;
2708         struct ib_qp                 *qp;
2709         struct ib_uverbs_mcast_entry *mcast;
2710         int                           ret = -EINVAL;
2711         bool                          found = false;
2712
2713         if (copy_from_user(&cmd, buf, sizeof cmd))
2714                 return -EFAULT;
2715
2716         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
2717         if (!qp)
2718                 return -EINVAL;
2719
2720         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2721         mutex_lock(&obj->mcast_lock);
2722
2723         list_for_each_entry(mcast, &obj->mcast_list, list)
2724                 if (cmd.mlid == mcast->lid &&
2725                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2726                         list_del(&mcast->list);
2727                         kfree(mcast);
2728                         found = true;
2729                         break;
2730                 }
2731
2732         if (!found) {
2733                 ret = -EINVAL;
2734                 goto out_put;
2735         }
2736
2737         ret = ib_detach_mcast(qp, (union ib_gid *)cmd.gid, cmd.mlid);
2738
2739 out_put:
2740         mutex_unlock(&obj->mcast_lock);
2741         uobj_put_obj_read(qp);
2742         return ret ? ret : in_len;
2743 }
2744
2745 static int kern_spec_to_ib_spec_action(struct ib_uverbs_flow_spec *kern_spec,
2746                                        union ib_flow_spec *ib_spec)
2747 {
2748         ib_spec->type = kern_spec->type;
2749         switch (ib_spec->type) {
2750         case IB_FLOW_SPEC_ACTION_TAG:
2751                 if (kern_spec->flow_tag.size !=
2752                     sizeof(struct ib_uverbs_flow_spec_action_tag))
2753                         return -EINVAL;
2754
2755                 ib_spec->flow_tag.size = sizeof(struct ib_flow_spec_action_tag);
2756                 ib_spec->flow_tag.tag_id = kern_spec->flow_tag.tag_id;
2757                 break;
2758         case IB_FLOW_SPEC_ACTION_DROP:
2759                 if (kern_spec->drop.size !=
2760                     sizeof(struct ib_uverbs_flow_spec_action_drop))
2761                         return -EINVAL;
2762
2763                 ib_spec->drop.size = sizeof(struct ib_flow_spec_action_drop);
2764                 break;
2765         default:
2766                 return -EINVAL;
2767         }
2768         return 0;
2769 }
2770
2771 static size_t kern_spec_filter_sz(struct ib_uverbs_flow_spec_hdr *spec)
2772 {
2773         /* Returns user space filter size, includes padding */
2774         return (spec->size - sizeof(struct ib_uverbs_flow_spec_hdr)) / 2;
2775 }
2776
2777 static ssize_t spec_filter_size(void *kern_spec_filter, u16 kern_filter_size,
2778                                 u16 ib_real_filter_sz)
2779 {
2780         /*
2781          * User space filter structures must be 64 bit aligned, otherwise this
2782          * may pass, but we won't handle additional new attributes.
2783          */
2784
2785         if (kern_filter_size > ib_real_filter_sz) {
2786                 if (memchr_inv(kern_spec_filter +
2787                                ib_real_filter_sz, 0,
2788                                kern_filter_size - ib_real_filter_sz))
2789                         return -EINVAL;
2790                 return ib_real_filter_sz;
2791         }
2792         return kern_filter_size;
2793 }
2794
2795 static int kern_spec_to_ib_spec_filter(struct ib_uverbs_flow_spec *kern_spec,
2796                                        union ib_flow_spec *ib_spec)
2797 {
2798         ssize_t actual_filter_sz;
2799         ssize_t kern_filter_sz;
2800         ssize_t ib_filter_sz;
2801         void *kern_spec_mask;
2802         void *kern_spec_val;
2803
2804         if (kern_spec->reserved)
2805                 return -EINVAL;
2806
2807         ib_spec->type = kern_spec->type;
2808
2809         kern_filter_sz = kern_spec_filter_sz(&kern_spec->hdr);
2810         /* User flow spec size must be aligned to 4 bytes */
2811         if (kern_filter_sz != ALIGN(kern_filter_sz, 4))
2812                 return -EINVAL;
2813
2814         kern_spec_val = (void *)kern_spec +
2815                 sizeof(struct ib_uverbs_flow_spec_hdr);
2816         kern_spec_mask = kern_spec_val + kern_filter_sz;
2817         if (ib_spec->type == (IB_FLOW_SPEC_INNER | IB_FLOW_SPEC_VXLAN_TUNNEL))
2818                 return -EINVAL;
2819
2820         switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2821         case IB_FLOW_SPEC_ETH:
2822                 ib_filter_sz = offsetof(struct ib_flow_eth_filter, real_sz);
2823                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2824                                                     kern_filter_sz,
2825                                                     ib_filter_sz);
2826                 if (actual_filter_sz <= 0)
2827                         return -EINVAL;
2828                 ib_spec->size = sizeof(struct ib_flow_spec_eth);
2829                 memcpy(&ib_spec->eth.val, kern_spec_val, actual_filter_sz);
2830                 memcpy(&ib_spec->eth.mask, kern_spec_mask, actual_filter_sz);
2831                 break;
2832         case IB_FLOW_SPEC_IPV4:
2833                 ib_filter_sz = offsetof(struct ib_flow_ipv4_filter, real_sz);
2834                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2835                                                     kern_filter_sz,
2836                                                     ib_filter_sz);
2837                 if (actual_filter_sz <= 0)
2838                         return -EINVAL;
2839                 ib_spec->size = sizeof(struct ib_flow_spec_ipv4);
2840                 memcpy(&ib_spec->ipv4.val, kern_spec_val, actual_filter_sz);
2841                 memcpy(&ib_spec->ipv4.mask, kern_spec_mask, actual_filter_sz);
2842                 break;
2843         case IB_FLOW_SPEC_IPV6:
2844                 ib_filter_sz = offsetof(struct ib_flow_ipv6_filter, real_sz);
2845                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2846                                                     kern_filter_sz,
2847                                                     ib_filter_sz);
2848                 if (actual_filter_sz <= 0)
2849                         return -EINVAL;
2850                 ib_spec->size = sizeof(struct ib_flow_spec_ipv6);
2851                 memcpy(&ib_spec->ipv6.val, kern_spec_val, actual_filter_sz);
2852                 memcpy(&ib_spec->ipv6.mask, kern_spec_mask, actual_filter_sz);
2853
2854                 if ((ntohl(ib_spec->ipv6.mask.flow_label)) >= BIT(20) ||
2855                     (ntohl(ib_spec->ipv6.val.flow_label)) >= BIT(20))
2856                         return -EINVAL;
2857                 break;
2858         case IB_FLOW_SPEC_TCP:
2859         case IB_FLOW_SPEC_UDP:
2860                 ib_filter_sz = offsetof(struct ib_flow_tcp_udp_filter, real_sz);
2861                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2862                                                     kern_filter_sz,
2863                                                     ib_filter_sz);
2864                 if (actual_filter_sz <= 0)
2865                         return -EINVAL;
2866                 ib_spec->size = sizeof(struct ib_flow_spec_tcp_udp);
2867                 memcpy(&ib_spec->tcp_udp.val, kern_spec_val, actual_filter_sz);
2868                 memcpy(&ib_spec->tcp_udp.mask, kern_spec_mask, actual_filter_sz);
2869                 break;
2870         case IB_FLOW_SPEC_VXLAN_TUNNEL:
2871                 ib_filter_sz = offsetof(struct ib_flow_tunnel_filter, real_sz);
2872                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2873                                                     kern_filter_sz,
2874                                                     ib_filter_sz);
2875                 if (actual_filter_sz <= 0)
2876                         return -EINVAL;
2877                 ib_spec->tunnel.size = sizeof(struct ib_flow_spec_tunnel);
2878                 memcpy(&ib_spec->tunnel.val, kern_spec_val, actual_filter_sz);
2879                 memcpy(&ib_spec->tunnel.mask, kern_spec_mask, actual_filter_sz);
2880
2881                 if ((ntohl(ib_spec->tunnel.mask.tunnel_id)) >= BIT(24) ||
2882                     (ntohl(ib_spec->tunnel.val.tunnel_id)) >= BIT(24))
2883                         return -EINVAL;
2884                 break;
2885         default:
2886                 return -EINVAL;
2887         }
2888         return 0;
2889 }
2890
2891 static int kern_spec_to_ib_spec(struct ib_uverbs_flow_spec *kern_spec,
2892                                 union ib_flow_spec *ib_spec)
2893 {
2894         if (kern_spec->reserved)
2895                 return -EINVAL;
2896
2897         if (kern_spec->type >= IB_FLOW_SPEC_ACTION_TAG)
2898                 return kern_spec_to_ib_spec_action(kern_spec, ib_spec);
2899         else
2900                 return kern_spec_to_ib_spec_filter(kern_spec, ib_spec);
2901 }
2902
2903 int ib_uverbs_ex_create_wq(struct ib_uverbs_file *file,
2904                            struct ib_device *ib_dev,
2905                            struct ib_udata *ucore,
2906                            struct ib_udata *uhw)
2907 {
2908         struct ib_uverbs_ex_create_wq     cmd = {};
2909         struct ib_uverbs_ex_create_wq_resp resp = {};
2910         struct ib_uwq_object           *obj;
2911         int err = 0;
2912         struct ib_cq *cq;
2913         struct ib_pd *pd;
2914         struct ib_wq *wq;
2915         struct ib_wq_init_attr wq_init_attr = {};
2916         size_t required_cmd_sz;
2917         size_t required_resp_len;
2918
2919         required_cmd_sz = offsetof(typeof(cmd), max_sge) + sizeof(cmd.max_sge);
2920         required_resp_len = offsetof(typeof(resp), wqn) + sizeof(resp.wqn);
2921
2922         if (ucore->inlen < required_cmd_sz)
2923                 return -EINVAL;
2924
2925         if (ucore->outlen < required_resp_len)
2926                 return -ENOSPC;
2927
2928         if (ucore->inlen > sizeof(cmd) &&
2929             !ib_is_udata_cleared(ucore, sizeof(cmd),
2930                                  ucore->inlen - sizeof(cmd)))
2931                 return -EOPNOTSUPP;
2932
2933         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2934         if (err)
2935                 return err;
2936
2937         if (cmd.comp_mask)
2938                 return -EOPNOTSUPP;
2939
2940         obj  = (struct ib_uwq_object *)uobj_alloc(uobj_get_type(wq),
2941                                                   file->ucontext);
2942         if (IS_ERR(obj))
2943                 return PTR_ERR(obj);
2944
2945         pd  = uobj_get_obj_read(pd, cmd.pd_handle, file->ucontext);
2946         if (!pd) {
2947                 err = -EINVAL;
2948                 goto err_uobj;
2949         }
2950
2951         cq = uobj_get_obj_read(cq, cmd.cq_handle, file->ucontext);
2952         if (!cq) {
2953                 err = -EINVAL;
2954                 goto err_put_pd;
2955         }
2956
2957         wq_init_attr.cq = cq;
2958         wq_init_attr.max_sge = cmd.max_sge;
2959         wq_init_attr.max_wr = cmd.max_wr;
2960         wq_init_attr.wq_context = file;
2961         wq_init_attr.wq_type = cmd.wq_type;
2962         wq_init_attr.event_handler = ib_uverbs_wq_event_handler;
2963         if (ucore->inlen >= (offsetof(typeof(cmd), create_flags) +
2964                              sizeof(cmd.create_flags)))
2965                 wq_init_attr.create_flags = cmd.create_flags;
2966         obj->uevent.events_reported = 0;
2967         INIT_LIST_HEAD(&obj->uevent.event_list);
2968         wq = pd->device->create_wq(pd, &wq_init_attr, uhw);
2969         if (IS_ERR(wq)) {
2970                 err = PTR_ERR(wq);
2971                 goto err_put_cq;
2972         }
2973
2974         wq->uobject = &obj->uevent.uobject;
2975         obj->uevent.uobject.object = wq;
2976         wq->wq_type = wq_init_attr.wq_type;
2977         wq->cq = cq;
2978         wq->pd = pd;
2979         wq->device = pd->device;
2980         wq->wq_context = wq_init_attr.wq_context;
2981         atomic_set(&wq->usecnt, 0);
2982         atomic_inc(&pd->usecnt);
2983         atomic_inc(&cq->usecnt);
2984         wq->uobject = &obj->uevent.uobject;
2985         obj->uevent.uobject.object = wq;
2986
2987         memset(&resp, 0, sizeof(resp));
2988         resp.wq_handle = obj->uevent.uobject.id;
2989         resp.max_sge = wq_init_attr.max_sge;
2990         resp.max_wr = wq_init_attr.max_wr;
2991         resp.wqn = wq->wq_num;
2992         resp.response_length = required_resp_len;
2993         err = ib_copy_to_udata(ucore,
2994                                &resp, resp.response_length);
2995         if (err)
2996                 goto err_copy;
2997
2998         uobj_put_obj_read(pd);
2999         uobj_put_obj_read(cq);
3000         uobj_alloc_commit(&obj->uevent.uobject);
3001         return 0;
3002
3003 err_copy:
3004         ib_destroy_wq(wq);
3005 err_put_cq:
3006         uobj_put_obj_read(cq);
3007 err_put_pd:
3008         uobj_put_obj_read(pd);
3009 err_uobj:
3010         uobj_alloc_abort(&obj->uevent.uobject);
3011
3012         return err;
3013 }
3014
3015 int ib_uverbs_ex_destroy_wq(struct ib_uverbs_file *file,
3016                             struct ib_device *ib_dev,
3017                             struct ib_udata *ucore,
3018                             struct ib_udata *uhw)
3019 {
3020         struct ib_uverbs_ex_destroy_wq  cmd = {};
3021         struct ib_uverbs_ex_destroy_wq_resp     resp = {};
3022         struct ib_wq                    *wq;
3023         struct ib_uobject               *uobj;
3024         struct ib_uwq_object            *obj;
3025         size_t required_cmd_sz;
3026         size_t required_resp_len;
3027         int                             ret;
3028
3029         required_cmd_sz = offsetof(typeof(cmd), wq_handle) + sizeof(cmd.wq_handle);
3030         required_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
3031
3032         if (ucore->inlen < required_cmd_sz)
3033                 return -EINVAL;
3034
3035         if (ucore->outlen < required_resp_len)
3036                 return -ENOSPC;
3037
3038         if (ucore->inlen > sizeof(cmd) &&
3039             !ib_is_udata_cleared(ucore, sizeof(cmd),
3040                                  ucore->inlen - sizeof(cmd)))
3041                 return -EOPNOTSUPP;
3042
3043         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3044         if (ret)
3045                 return ret;
3046
3047         if (cmd.comp_mask)
3048                 return -EOPNOTSUPP;
3049
3050         resp.response_length = required_resp_len;
3051         uobj  = uobj_get_write(uobj_get_type(wq), cmd.wq_handle,
3052                                file->ucontext);
3053         if (IS_ERR(uobj))
3054                 return PTR_ERR(uobj);
3055
3056         wq = uobj->object;
3057         obj = container_of(uobj, struct ib_uwq_object, uevent.uobject);
3058         /*
3059          * Make sure we don't free the memory in remove_commit as we still
3060          * needs the uobject memory to create the response.
3061          */
3062         uverbs_uobject_get(uobj);
3063
3064         ret = uobj_remove_commit(uobj);
3065         resp.events_reported = obj->uevent.events_reported;
3066         uverbs_uobject_put(uobj);
3067         if (ret)
3068                 return ret;
3069
3070         return ib_copy_to_udata(ucore, &resp, resp.response_length);
3071 }
3072
3073 int ib_uverbs_ex_modify_wq(struct ib_uverbs_file *file,
3074                            struct ib_device *ib_dev,
3075                            struct ib_udata *ucore,
3076                            struct ib_udata *uhw)
3077 {
3078         struct ib_uverbs_ex_modify_wq cmd = {};
3079         struct ib_wq *wq;
3080         struct ib_wq_attr wq_attr = {};
3081         size_t required_cmd_sz;
3082         int ret;
3083
3084         required_cmd_sz = offsetof(typeof(cmd), curr_wq_state) + sizeof(cmd.curr_wq_state);
3085         if (ucore->inlen < required_cmd_sz)
3086                 return -EINVAL;
3087
3088         if (ucore->inlen > sizeof(cmd) &&
3089             !ib_is_udata_cleared(ucore, sizeof(cmd),
3090                                  ucore->inlen - sizeof(cmd)))
3091                 return -EOPNOTSUPP;
3092
3093         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3094         if (ret)
3095                 return ret;
3096
3097         if (!cmd.attr_mask)
3098                 return -EINVAL;
3099
3100         if (cmd.attr_mask > (IB_WQ_STATE | IB_WQ_CUR_STATE | IB_WQ_FLAGS))
3101                 return -EINVAL;
3102
3103         wq = uobj_get_obj_read(wq, cmd.wq_handle, file->ucontext);
3104         if (!wq)
3105                 return -EINVAL;
3106
3107         wq_attr.curr_wq_state = cmd.curr_wq_state;
3108         wq_attr.wq_state = cmd.wq_state;
3109         if (cmd.attr_mask & IB_WQ_FLAGS) {
3110                 wq_attr.flags = cmd.flags;
3111                 wq_attr.flags_mask = cmd.flags_mask;
3112         }
3113         ret = wq->device->modify_wq(wq, &wq_attr, cmd.attr_mask, uhw);
3114         uobj_put_obj_read(wq);
3115         return ret;
3116 }
3117
3118 int ib_uverbs_ex_create_rwq_ind_table(struct ib_uverbs_file *file,
3119                                       struct ib_device *ib_dev,
3120                                       struct ib_udata *ucore,
3121                                       struct ib_udata *uhw)
3122 {
3123         struct ib_uverbs_ex_create_rwq_ind_table          cmd = {};
3124         struct ib_uverbs_ex_create_rwq_ind_table_resp  resp = {};
3125         struct ib_uobject                 *uobj;
3126         int err = 0;
3127         struct ib_rwq_ind_table_init_attr init_attr = {};
3128         struct ib_rwq_ind_table *rwq_ind_tbl;
3129         struct ib_wq    **wqs = NULL;
3130         u32 *wqs_handles = NULL;
3131         struct ib_wq    *wq = NULL;
3132         int i, j, num_read_wqs;
3133         u32 num_wq_handles;
3134         u32 expected_in_size;
3135         size_t required_cmd_sz_header;
3136         size_t required_resp_len;
3137
3138         required_cmd_sz_header = offsetof(typeof(cmd), log_ind_tbl_size) + sizeof(cmd.log_ind_tbl_size);
3139         required_resp_len = offsetof(typeof(resp), ind_tbl_num) + sizeof(resp.ind_tbl_num);
3140
3141         if (ucore->inlen < required_cmd_sz_header)
3142                 return -EINVAL;
3143
3144         if (ucore->outlen < required_resp_len)
3145                 return -ENOSPC;
3146
3147         err = ib_copy_from_udata(&cmd, ucore, required_cmd_sz_header);
3148         if (err)
3149                 return err;
3150
3151         ucore->inbuf += required_cmd_sz_header;
3152         ucore->inlen -= required_cmd_sz_header;
3153
3154         if (cmd.comp_mask)
3155                 return -EOPNOTSUPP;
3156
3157         if (cmd.log_ind_tbl_size > IB_USER_VERBS_MAX_LOG_IND_TBL_SIZE)
3158                 return -EINVAL;
3159
3160         num_wq_handles = 1 << cmd.log_ind_tbl_size;
3161         expected_in_size = num_wq_handles * sizeof(__u32);
3162         if (num_wq_handles == 1)
3163                 /* input size for wq handles is u64 aligned */
3164                 expected_in_size += sizeof(__u32);
3165
3166         if (ucore->inlen < expected_in_size)
3167                 return -EINVAL;
3168
3169         if (ucore->inlen > expected_in_size &&
3170             !ib_is_udata_cleared(ucore, expected_in_size,
3171                                  ucore->inlen - expected_in_size))
3172                 return -EOPNOTSUPP;
3173
3174         wqs_handles = kcalloc(num_wq_handles, sizeof(*wqs_handles),
3175                               GFP_KERNEL);
3176         if (!wqs_handles)
3177                 return -ENOMEM;
3178
3179         err = ib_copy_from_udata(wqs_handles, ucore,
3180                                  num_wq_handles * sizeof(__u32));
3181         if (err)
3182                 goto err_free;
3183
3184         wqs = kcalloc(num_wq_handles, sizeof(*wqs), GFP_KERNEL);
3185         if (!wqs) {
3186                 err = -ENOMEM;
3187                 goto  err_free;
3188         }
3189
3190         for (num_read_wqs = 0; num_read_wqs < num_wq_handles;
3191                         num_read_wqs++) {
3192                 wq = uobj_get_obj_read(wq, wqs_handles[num_read_wqs],
3193                                        file->ucontext);
3194                 if (!wq) {
3195                         err = -EINVAL;
3196                         goto put_wqs;
3197                 }
3198
3199                 wqs[num_read_wqs] = wq;
3200         }
3201
3202         uobj  = uobj_alloc(uobj_get_type(rwq_ind_table), file->ucontext);
3203         if (IS_ERR(uobj)) {
3204                 err = PTR_ERR(uobj);
3205                 goto put_wqs;
3206         }
3207
3208         init_attr.log_ind_tbl_size = cmd.log_ind_tbl_size;
3209         init_attr.ind_tbl = wqs;
3210         rwq_ind_tbl = ib_dev->create_rwq_ind_table(ib_dev, &init_attr, uhw);
3211
3212         if (IS_ERR(rwq_ind_tbl)) {
3213                 err = PTR_ERR(rwq_ind_tbl);
3214                 goto err_uobj;
3215         }
3216
3217         rwq_ind_tbl->ind_tbl = wqs;
3218         rwq_ind_tbl->log_ind_tbl_size = init_attr.log_ind_tbl_size;
3219         rwq_ind_tbl->uobject = uobj;
3220         uobj->object = rwq_ind_tbl;
3221         rwq_ind_tbl->device = ib_dev;
3222         atomic_set(&rwq_ind_tbl->usecnt, 0);
3223
3224         for (i = 0; i < num_wq_handles; i++)
3225                 atomic_inc(&wqs[i]->usecnt);
3226
3227         resp.ind_tbl_handle = uobj->id;
3228         resp.ind_tbl_num = rwq_ind_tbl->ind_tbl_num;
3229         resp.response_length = required_resp_len;
3230
3231         err = ib_copy_to_udata(ucore,
3232                                &resp, resp.response_length);
3233         if (err)
3234                 goto err_copy;
3235
3236         kfree(wqs_handles);
3237
3238         for (j = 0; j < num_read_wqs; j++)
3239                 uobj_put_obj_read(wqs[j]);
3240
3241         uobj_alloc_commit(uobj);
3242         return 0;
3243
3244 err_copy:
3245         ib_destroy_rwq_ind_table(rwq_ind_tbl);
3246 err_uobj:
3247         uobj_alloc_abort(uobj);
3248 put_wqs:
3249         for (j = 0; j < num_read_wqs; j++)
3250                 uobj_put_obj_read(wqs[j]);
3251 err_free:
3252         kfree(wqs_handles);
3253         kfree(wqs);
3254         return err;
3255 }
3256
3257 int ib_uverbs_ex_destroy_rwq_ind_table(struct ib_uverbs_file *file,
3258                                        struct ib_device *ib_dev,
3259                                        struct ib_udata *ucore,
3260                                        struct ib_udata *uhw)
3261 {
3262         struct ib_uverbs_ex_destroy_rwq_ind_table       cmd = {};
3263         struct ib_uobject               *uobj;
3264         int                     ret;
3265         size_t required_cmd_sz;
3266
3267         required_cmd_sz = offsetof(typeof(cmd), ind_tbl_handle) + sizeof(cmd.ind_tbl_handle);
3268
3269         if (ucore->inlen < required_cmd_sz)
3270                 return -EINVAL;
3271
3272         if (ucore->inlen > sizeof(cmd) &&
3273             !ib_is_udata_cleared(ucore, sizeof(cmd),
3274                                  ucore->inlen - sizeof(cmd)))
3275                 return -EOPNOTSUPP;
3276
3277         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3278         if (ret)
3279                 return ret;
3280
3281         if (cmd.comp_mask)
3282                 return -EOPNOTSUPP;
3283
3284         uobj  = uobj_get_write(uobj_get_type(rwq_ind_table), cmd.ind_tbl_handle,
3285                                file->ucontext);
3286         if (IS_ERR(uobj))
3287                 return PTR_ERR(uobj);
3288
3289         return uobj_remove_commit(uobj);
3290 }
3291
3292 int ib_uverbs_ex_create_flow(struct ib_uverbs_file *file,
3293                              struct ib_device *ib_dev,
3294                              struct ib_udata *ucore,
3295                              struct ib_udata *uhw)
3296 {
3297         struct ib_uverbs_create_flow      cmd;
3298         struct ib_uverbs_create_flow_resp resp;
3299         struct ib_uobject                 *uobj;
3300         struct ib_flow                    *flow_id;
3301         struct ib_uverbs_flow_attr        *kern_flow_attr;
3302         struct ib_flow_attr               *flow_attr;
3303         struct ib_qp                      *qp;
3304         int err = 0;
3305         void *kern_spec;
3306         void *ib_spec;
3307         int i;
3308
3309         if (ucore->inlen < sizeof(cmd))
3310                 return -EINVAL;
3311
3312         if (ucore->outlen < sizeof(resp))
3313                 return -ENOSPC;
3314
3315         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3316         if (err)
3317                 return err;
3318
3319         ucore->inbuf += sizeof(cmd);
3320         ucore->inlen -= sizeof(cmd);
3321
3322         if (cmd.comp_mask)
3323                 return -EINVAL;
3324
3325         if (!capable(CAP_NET_RAW))
3326                 return -EPERM;
3327
3328         if (cmd.flow_attr.flags >= IB_FLOW_ATTR_FLAGS_RESERVED)
3329                 return -EINVAL;
3330
3331         if ((cmd.flow_attr.flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) &&
3332             ((cmd.flow_attr.type == IB_FLOW_ATTR_ALL_DEFAULT) ||
3333              (cmd.flow_attr.type == IB_FLOW_ATTR_MC_DEFAULT)))
3334                 return -EINVAL;
3335
3336         if (cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
3337                 return -EINVAL;
3338
3339         if (cmd.flow_attr.size > ucore->inlen ||
3340             cmd.flow_attr.size >
3341             (cmd.flow_attr.num_of_specs * sizeof(struct ib_uverbs_flow_spec)))
3342                 return -EINVAL;
3343
3344         if (cmd.flow_attr.reserved[0] ||
3345             cmd.flow_attr.reserved[1])
3346                 return -EINVAL;
3347
3348         if (cmd.flow_attr.num_of_specs) {
3349                 kern_flow_attr = kmalloc(sizeof(*kern_flow_attr) + cmd.flow_attr.size,
3350                                          GFP_KERNEL);
3351                 if (!kern_flow_attr)
3352                         return -ENOMEM;
3353
3354                 memcpy(kern_flow_attr, &cmd.flow_attr, sizeof(*kern_flow_attr));
3355                 err = ib_copy_from_udata(kern_flow_attr + 1, ucore,
3356                                          cmd.flow_attr.size);
3357                 if (err)
3358                         goto err_free_attr;
3359         } else {
3360                 kern_flow_attr = &cmd.flow_attr;
3361         }
3362
3363         uobj  = uobj_alloc(uobj_get_type(flow), file->ucontext);
3364         if (IS_ERR(uobj)) {
3365                 err = PTR_ERR(uobj);
3366                 goto err_free_attr;
3367         }
3368
3369         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
3370         if (!qp) {
3371                 err = -EINVAL;
3372                 goto err_uobj;
3373         }
3374
3375         flow_attr = kzalloc(sizeof(*flow_attr) + cmd.flow_attr.num_of_specs *
3376                             sizeof(union ib_flow_spec), GFP_KERNEL);
3377         if (!flow_attr) {
3378                 err = -ENOMEM;
3379                 goto err_put;
3380         }
3381
3382         flow_attr->type = kern_flow_attr->type;
3383         flow_attr->priority = kern_flow_attr->priority;
3384         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
3385         flow_attr->port = kern_flow_attr->port;
3386         flow_attr->flags = kern_flow_attr->flags;
3387         flow_attr->size = sizeof(*flow_attr);
3388
3389         kern_spec = kern_flow_attr + 1;
3390         ib_spec = flow_attr + 1;
3391         for (i = 0; i < flow_attr->num_of_specs &&
3392              cmd.flow_attr.size > offsetof(struct ib_uverbs_flow_spec, reserved) &&
3393              cmd.flow_attr.size >=
3394              ((struct ib_uverbs_flow_spec *)kern_spec)->size; i++) {
3395                 err = kern_spec_to_ib_spec(kern_spec, ib_spec);
3396                 if (err)
3397                         goto err_free;
3398                 flow_attr->size +=
3399                         ((union ib_flow_spec *) ib_spec)->size;
3400                 cmd.flow_attr.size -= ((struct ib_uverbs_flow_spec *)kern_spec)->size;
3401                 kern_spec += ((struct ib_uverbs_flow_spec *) kern_spec)->size;
3402                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
3403         }
3404         if (cmd.flow_attr.size || (i != flow_attr->num_of_specs)) {
3405                 pr_warn("create flow failed, flow %d: %d bytes left from uverb cmd\n",
3406                         i, cmd.flow_attr.size);
3407                 err = -EINVAL;
3408                 goto err_free;
3409         }
3410         flow_id = ib_create_flow(qp, flow_attr, IB_FLOW_DOMAIN_USER);
3411         if (IS_ERR(flow_id)) {
3412                 err = PTR_ERR(flow_id);
3413                 goto err_free;
3414         }
3415         flow_id->uobject = uobj;
3416         uobj->object = flow_id;
3417
3418         memset(&resp, 0, sizeof(resp));
3419         resp.flow_handle = uobj->id;
3420
3421         err = ib_copy_to_udata(ucore,
3422                                &resp, sizeof(resp));
3423         if (err)
3424                 goto err_copy;
3425
3426         uobj_put_obj_read(qp);
3427         uobj_alloc_commit(uobj);
3428         kfree(flow_attr);
3429         if (cmd.flow_attr.num_of_specs)
3430                 kfree(kern_flow_attr);
3431         return 0;
3432 err_copy:
3433         ib_destroy_flow(flow_id);
3434 err_free:
3435         kfree(flow_attr);
3436 err_put:
3437         uobj_put_obj_read(qp);
3438 err_uobj:
3439         uobj_alloc_abort(uobj);
3440 err_free_attr:
3441         if (cmd.flow_attr.num_of_specs)
3442                 kfree(kern_flow_attr);
3443         return err;
3444 }
3445
3446 int ib_uverbs_ex_destroy_flow(struct ib_uverbs_file *file,
3447                               struct ib_device *ib_dev,
3448                               struct ib_udata *ucore,
3449                               struct ib_udata *uhw)
3450 {
3451         struct ib_uverbs_destroy_flow   cmd;
3452         struct ib_uobject               *uobj;
3453         int                             ret;
3454
3455         if (ucore->inlen < sizeof(cmd))
3456                 return -EINVAL;
3457
3458         ret = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3459         if (ret)
3460                 return ret;
3461
3462         if (cmd.comp_mask)
3463                 return -EINVAL;
3464
3465         uobj  = uobj_get_write(uobj_get_type(flow), cmd.flow_handle,
3466                                file->ucontext);
3467         if (IS_ERR(uobj))
3468                 return PTR_ERR(uobj);
3469
3470         ret = uobj_remove_commit(uobj);
3471         return ret;
3472 }
3473
3474 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
3475                                 struct ib_device *ib_dev,
3476                                 struct ib_uverbs_create_xsrq *cmd,
3477                                 struct ib_udata *udata)
3478 {
3479         struct ib_uverbs_create_srq_resp resp;
3480         struct ib_usrq_object           *obj;
3481         struct ib_pd                    *pd;
3482         struct ib_srq                   *srq;
3483         struct ib_uobject               *uninitialized_var(xrcd_uobj);
3484         struct ib_srq_init_attr          attr;
3485         int ret;
3486
3487         obj  = (struct ib_usrq_object *)uobj_alloc(uobj_get_type(srq),
3488                                                    file->ucontext);
3489         if (IS_ERR(obj))
3490                 return PTR_ERR(obj);
3491
3492         if (cmd->srq_type == IB_SRQT_XRC) {
3493                 xrcd_uobj = uobj_get_read(uobj_get_type(xrcd), cmd->xrcd_handle,
3494                                           file->ucontext);
3495                 if (IS_ERR(xrcd_uobj)) {
3496                         ret = -EINVAL;
3497                         goto err;
3498                 }
3499
3500                 attr.ext.xrc.xrcd = (struct ib_xrcd *)xrcd_uobj->object;
3501                 if (!attr.ext.xrc.xrcd) {
3502                         ret = -EINVAL;
3503                         goto err_put_xrcd;
3504                 }
3505
3506                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
3507                 atomic_inc(&obj->uxrcd->refcnt);
3508
3509                 attr.ext.xrc.cq  = uobj_get_obj_read(cq, cmd->cq_handle,
3510                                                      file->ucontext);
3511                 if (!attr.ext.xrc.cq) {
3512                         ret = -EINVAL;
3513                         goto err_put_xrcd;
3514                 }
3515         }
3516
3517         pd  = uobj_get_obj_read(pd, cmd->pd_handle, file->ucontext);
3518         if (!pd) {
3519                 ret = -EINVAL;
3520                 goto err_put_cq;
3521         }
3522
3523         attr.event_handler  = ib_uverbs_srq_event_handler;
3524         attr.srq_context    = file;
3525         attr.srq_type       = cmd->srq_type;
3526         attr.attr.max_wr    = cmd->max_wr;
3527         attr.attr.max_sge   = cmd->max_sge;
3528         attr.attr.srq_limit = cmd->srq_limit;
3529
3530         obj->uevent.events_reported = 0;
3531         INIT_LIST_HEAD(&obj->uevent.event_list);
3532
3533         srq = pd->device->create_srq(pd, &attr, udata);
3534         if (IS_ERR(srq)) {
3535                 ret = PTR_ERR(srq);
3536                 goto err_put;
3537         }
3538
3539         srq->device        = pd->device;
3540         srq->pd            = pd;
3541         srq->srq_type      = cmd->srq_type;
3542         srq->uobject       = &obj->uevent.uobject;
3543         srq->event_handler = attr.event_handler;
3544         srq->srq_context   = attr.srq_context;
3545
3546         if (cmd->srq_type == IB_SRQT_XRC) {
3547                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
3548                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
3549                 atomic_inc(&attr.ext.xrc.cq->usecnt);
3550                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
3551         }
3552
3553         atomic_inc(&pd->usecnt);
3554         atomic_set(&srq->usecnt, 0);
3555
3556         obj->uevent.uobject.object = srq;
3557         obj->uevent.uobject.user_handle = cmd->user_handle;
3558
3559         memset(&resp, 0, sizeof resp);
3560         resp.srq_handle = obj->uevent.uobject.id;
3561         resp.max_wr     = attr.attr.max_wr;
3562         resp.max_sge    = attr.attr.max_sge;
3563         if (cmd->srq_type == IB_SRQT_XRC)
3564                 resp.srqn = srq->ext.xrc.srq_num;
3565
3566         if (copy_to_user((void __user *) (unsigned long) cmd->response,
3567                          &resp, sizeof resp)) {
3568                 ret = -EFAULT;
3569                 goto err_copy;
3570         }
3571
3572         if (cmd->srq_type == IB_SRQT_XRC) {
3573                 uobj_put_read(xrcd_uobj);
3574                 uobj_put_obj_read(attr.ext.xrc.cq);
3575         }
3576         uobj_put_obj_read(pd);
3577         uobj_alloc_commit(&obj->uevent.uobject);
3578
3579         return 0;
3580
3581 err_copy:
3582         ib_destroy_srq(srq);
3583
3584 err_put:
3585         uobj_put_obj_read(pd);
3586
3587 err_put_cq:
3588         if (cmd->srq_type == IB_SRQT_XRC)
3589                 uobj_put_obj_read(attr.ext.xrc.cq);
3590
3591 err_put_xrcd:
3592         if (cmd->srq_type == IB_SRQT_XRC) {
3593                 atomic_dec(&obj->uxrcd->refcnt);
3594                 uobj_put_read(xrcd_uobj);
3595         }
3596
3597 err:
3598         uobj_alloc_abort(&obj->uevent.uobject);
3599         return ret;
3600 }
3601
3602 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
3603                              struct ib_device *ib_dev,
3604                              const char __user *buf, int in_len,
3605                              int out_len)
3606 {
3607         struct ib_uverbs_create_srq      cmd;
3608         struct ib_uverbs_create_xsrq     xcmd;
3609         struct ib_uverbs_create_srq_resp resp;
3610         struct ib_udata                  udata;
3611         int ret;
3612
3613         if (out_len < sizeof resp)
3614                 return -ENOSPC;
3615
3616         if (copy_from_user(&cmd, buf, sizeof cmd))
3617                 return -EFAULT;
3618
3619         xcmd.response    = cmd.response;
3620         xcmd.user_handle = cmd.user_handle;
3621         xcmd.srq_type    = IB_SRQT_BASIC;
3622         xcmd.pd_handle   = cmd.pd_handle;
3623         xcmd.max_wr      = cmd.max_wr;
3624         xcmd.max_sge     = cmd.max_sge;
3625         xcmd.srq_limit   = cmd.srq_limit;
3626
3627         INIT_UDATA(&udata, buf + sizeof cmd,
3628                    (unsigned long) cmd.response + sizeof resp,
3629                    in_len - sizeof cmd - sizeof(struct ib_uverbs_cmd_hdr),
3630                    out_len - sizeof resp);
3631
3632         ret = __uverbs_create_xsrq(file, ib_dev, &xcmd, &udata);
3633         if (ret)
3634                 return ret;
3635
3636         return in_len;
3637 }
3638
3639 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
3640                               struct ib_device *ib_dev,
3641                               const char __user *buf, int in_len, int out_len)
3642 {
3643         struct ib_uverbs_create_xsrq     cmd;
3644         struct ib_uverbs_create_srq_resp resp;
3645         struct ib_udata                  udata;
3646         int ret;
3647
3648         if (out_len < sizeof resp)
3649                 return -ENOSPC;
3650
3651         if (copy_from_user(&cmd, buf, sizeof cmd))
3652                 return -EFAULT;
3653
3654         INIT_UDATA(&udata, buf + sizeof cmd,
3655                    (unsigned long) cmd.response + sizeof resp,
3656                    in_len - sizeof cmd - sizeof(struct ib_uverbs_cmd_hdr),
3657                    out_len - sizeof resp);
3658
3659         ret = __uverbs_create_xsrq(file, ib_dev, &cmd, &udata);
3660         if (ret)
3661                 return ret;
3662
3663         return in_len;
3664 }
3665
3666 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
3667                              struct ib_device *ib_dev,
3668                              const char __user *buf, int in_len,
3669                              int out_len)
3670 {
3671         struct ib_uverbs_modify_srq cmd;
3672         struct ib_udata             udata;
3673         struct ib_srq              *srq;
3674         struct ib_srq_attr          attr;
3675         int                         ret;
3676
3677         if (copy_from_user(&cmd, buf, sizeof cmd))
3678                 return -EFAULT;
3679
3680         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
3681                    out_len);
3682
3683         srq = uobj_get_obj_read(srq, cmd.srq_handle, file->ucontext);
3684         if (!srq)
3685                 return -EINVAL;
3686
3687         attr.max_wr    = cmd.max_wr;
3688         attr.srq_limit = cmd.srq_limit;
3689
3690         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
3691
3692         uobj_put_obj_read(srq);
3693
3694         return ret ? ret : in_len;
3695 }
3696
3697 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
3698                             struct ib_device *ib_dev,
3699                             const char __user *buf,
3700                             int in_len, int out_len)
3701 {
3702         struct ib_uverbs_query_srq      cmd;
3703         struct ib_uverbs_query_srq_resp resp;
3704         struct ib_srq_attr              attr;
3705         struct ib_srq                   *srq;
3706         int                             ret;
3707
3708         if (out_len < sizeof resp)
3709                 return -ENOSPC;
3710
3711         if (copy_from_user(&cmd, buf, sizeof cmd))
3712                 return -EFAULT;
3713
3714         srq = uobj_get_obj_read(srq, cmd.srq_handle, file->ucontext);
3715         if (!srq)
3716                 return -EINVAL;
3717
3718         ret = ib_query_srq(srq, &attr);
3719
3720         uobj_put_obj_read(srq);
3721
3722         if (ret)
3723                 return ret;
3724
3725         memset(&resp, 0, sizeof resp);
3726
3727         resp.max_wr    = attr.max_wr;
3728         resp.max_sge   = attr.max_sge;
3729         resp.srq_limit = attr.srq_limit;
3730
3731         if (copy_to_user((void __user *) (unsigned long) cmd.response,
3732                          &resp, sizeof resp))
3733                 return -EFAULT;
3734
3735         return in_len;
3736 }
3737
3738 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
3739                               struct ib_device *ib_dev,
3740                               const char __user *buf, int in_len,
3741                               int out_len)
3742 {
3743         struct ib_uverbs_destroy_srq      cmd;
3744         struct ib_uverbs_destroy_srq_resp resp;
3745         struct ib_uobject                *uobj;
3746         struct ib_srq                    *srq;
3747         struct ib_uevent_object          *obj;
3748         int                               ret = -EINVAL;
3749         enum ib_srq_type                  srq_type;
3750
3751         if (copy_from_user(&cmd, buf, sizeof cmd))
3752                 return -EFAULT;
3753
3754         uobj  = uobj_get_write(uobj_get_type(srq), cmd.srq_handle,
3755                                file->ucontext);
3756         if (IS_ERR(uobj))
3757                 return PTR_ERR(uobj);
3758
3759         srq = uobj->object;
3760         obj = container_of(uobj, struct ib_uevent_object, uobject);
3761         srq_type = srq->srq_type;
3762         /*
3763          * Make sure we don't free the memory in remove_commit as we still
3764          * needs the uobject memory to create the response.
3765          */
3766         uverbs_uobject_get(uobj);
3767
3768         memset(&resp, 0, sizeof(resp));
3769
3770         ret = uobj_remove_commit(uobj);
3771         if (ret) {
3772                 uverbs_uobject_put(uobj);
3773                 return ret;
3774         }
3775         resp.events_reported = obj->events_reported;
3776         uverbs_uobject_put(uobj);
3777         if (copy_to_user((void __user *)(unsigned long)cmd.response,
3778                          &resp, sizeof(resp)))
3779                 return -EFAULT;
3780
3781         return in_len;
3782 }
3783
3784 int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
3785                               struct ib_device *ib_dev,
3786                               struct ib_udata *ucore,
3787                               struct ib_udata *uhw)
3788 {
3789         struct ib_uverbs_ex_query_device_resp resp = { {0} };
3790         struct ib_uverbs_ex_query_device  cmd;
3791         struct ib_device_attr attr = {0};
3792         int err;
3793
3794         if (ucore->inlen < sizeof(cmd))
3795                 return -EINVAL;
3796
3797         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3798         if (err)
3799                 return err;
3800
3801         if (cmd.comp_mask)
3802                 return -EINVAL;
3803
3804         if (cmd.reserved)
3805                 return -EINVAL;
3806
3807         resp.response_length = offsetof(typeof(resp), odp_caps);
3808
3809         if (ucore->outlen < resp.response_length)
3810                 return -ENOSPC;
3811
3812         err = ib_dev->query_device(ib_dev, &attr, uhw);
3813         if (err)
3814                 return err;
3815
3816         copy_query_dev_fields(file, ib_dev, &resp.base, &attr);
3817
3818         if (ucore->outlen < resp.response_length + sizeof(resp.odp_caps))
3819                 goto end;
3820
3821 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
3822         resp.odp_caps.general_caps = attr.odp_caps.general_caps;
3823         resp.odp_caps.per_transport_caps.rc_odp_caps =
3824                 attr.odp_caps.per_transport_caps.rc_odp_caps;
3825         resp.odp_caps.per_transport_caps.uc_odp_caps =
3826                 attr.odp_caps.per_transport_caps.uc_odp_caps;
3827         resp.odp_caps.per_transport_caps.ud_odp_caps =
3828                 attr.odp_caps.per_transport_caps.ud_odp_caps;
3829 #endif
3830         resp.response_length += sizeof(resp.odp_caps);
3831
3832         if (ucore->outlen < resp.response_length + sizeof(resp.timestamp_mask))
3833                 goto end;
3834
3835         resp.timestamp_mask = attr.timestamp_mask;
3836         resp.response_length += sizeof(resp.timestamp_mask);
3837
3838         if (ucore->outlen < resp.response_length + sizeof(resp.hca_core_clock))
3839                 goto end;
3840
3841         resp.hca_core_clock = attr.hca_core_clock;
3842         resp.response_length += sizeof(resp.hca_core_clock);
3843
3844         if (ucore->outlen < resp.response_length + sizeof(resp.device_cap_flags_ex))
3845                 goto end;
3846
3847         resp.device_cap_flags_ex = attr.device_cap_flags;
3848         resp.response_length += sizeof(resp.device_cap_flags_ex);
3849
3850         if (ucore->outlen < resp.response_length + sizeof(resp.rss_caps))
3851                 goto end;
3852
3853         resp.rss_caps.supported_qpts = attr.rss_caps.supported_qpts;
3854         resp.rss_caps.max_rwq_indirection_tables =
3855                 attr.rss_caps.max_rwq_indirection_tables;
3856         resp.rss_caps.max_rwq_indirection_table_size =
3857                 attr.rss_caps.max_rwq_indirection_table_size;
3858
3859         resp.response_length += sizeof(resp.rss_caps);
3860
3861         if (ucore->outlen < resp.response_length + sizeof(resp.max_wq_type_rq))
3862                 goto end;
3863
3864         resp.max_wq_type_rq = attr.max_wq_type_rq;
3865         resp.response_length += sizeof(resp.max_wq_type_rq);
3866
3867         if (ucore->outlen < resp.response_length + sizeof(resp.raw_packet_caps))
3868                 goto end;
3869
3870         resp.raw_packet_caps = attr.raw_packet_caps;
3871         resp.response_length += sizeof(resp.raw_packet_caps);
3872 end:
3873         err = ib_copy_to_udata(ucore, &resp, resp.response_length);
3874         return err;
3875 }