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_file->ev_queue : NULL;
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         int                              ret = -EINVAL;
1300
1301         if (copy_from_user(&cmd, buf, sizeof cmd))
1302                 return -EFAULT;
1303
1304         uobj  = uobj_get_write(uobj_get_type(cq), cmd.cq_handle,
1305                                file->ucontext);
1306         if (IS_ERR(uobj))
1307                 return PTR_ERR(uobj);
1308
1309         /*
1310          * Make sure we don't free the memory in remove_commit as we still
1311          * needs the uobject memory to create the response.
1312          */
1313         uverbs_uobject_get(uobj);
1314         cq      = uobj->object;
1315         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1316
1317         memset(&resp, 0, sizeof(resp));
1318
1319         ret = uobj_remove_commit(uobj);
1320         if (ret) {
1321                 uverbs_uobject_put(uobj);
1322                 return ret;
1323         }
1324
1325         resp.comp_events_reported  = obj->comp_events_reported;
1326         resp.async_events_reported = obj->async_events_reported;
1327
1328         uverbs_uobject_put(uobj);
1329         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1330                          &resp, sizeof resp))
1331                 return -EFAULT;
1332
1333         return in_len;
1334 }
1335
1336 static int create_qp(struct ib_uverbs_file *file,
1337                      struct ib_udata *ucore,
1338                      struct ib_udata *uhw,
1339                      struct ib_uverbs_ex_create_qp *cmd,
1340                      size_t cmd_sz,
1341                      int (*cb)(struct ib_uverbs_file *file,
1342                                struct ib_uverbs_ex_create_qp_resp *resp,
1343                                struct ib_udata *udata),
1344                      void *context)
1345 {
1346         struct ib_uqp_object            *obj;
1347         struct ib_device                *device;
1348         struct ib_pd                    *pd = NULL;
1349         struct ib_xrcd                  *xrcd = NULL;
1350         struct ib_uobject               *xrcd_uobj = ERR_PTR(-ENOENT);
1351         struct ib_cq                    *scq = NULL, *rcq = NULL;
1352         struct ib_srq                   *srq = NULL;
1353         struct ib_qp                    *qp;
1354         char                            *buf;
1355         struct ib_qp_init_attr          attr = {};
1356         struct ib_uverbs_ex_create_qp_resp resp;
1357         int                             ret;
1358         struct ib_rwq_ind_table *ind_tbl = NULL;
1359         bool has_sq = true;
1360
1361         if (cmd->qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1362                 return -EPERM;
1363
1364         obj  = (struct ib_uqp_object *)uobj_alloc(uobj_get_type(qp),
1365                                                   file->ucontext);
1366         if (IS_ERR(obj))
1367                 return PTR_ERR(obj);
1368         obj->uxrcd = NULL;
1369         obj->uevent.uobject.user_handle = cmd->user_handle;
1370         mutex_init(&obj->mcast_lock);
1371
1372         if (cmd_sz >= offsetof(typeof(*cmd), rwq_ind_tbl_handle) +
1373                       sizeof(cmd->rwq_ind_tbl_handle) &&
1374                       (cmd->comp_mask & IB_UVERBS_CREATE_QP_MASK_IND_TABLE)) {
1375                 ind_tbl = uobj_get_obj_read(rwq_ind_table,
1376                                             cmd->rwq_ind_tbl_handle,
1377                                             file->ucontext);
1378                 if (!ind_tbl) {
1379                         ret = -EINVAL;
1380                         goto err_put;
1381                 }
1382
1383                 attr.rwq_ind_tbl = ind_tbl;
1384         }
1385
1386         if ((cmd_sz >= offsetof(typeof(*cmd), reserved1) +
1387                        sizeof(cmd->reserved1)) && cmd->reserved1) {
1388                 ret = -EOPNOTSUPP;
1389                 goto err_put;
1390         }
1391
1392         if (ind_tbl && (cmd->max_recv_wr || cmd->max_recv_sge || cmd->is_srq)) {
1393                 ret = -EINVAL;
1394                 goto err_put;
1395         }
1396
1397         if (ind_tbl && !cmd->max_send_wr)
1398                 has_sq = false;
1399
1400         if (cmd->qp_type == IB_QPT_XRC_TGT) {
1401                 xrcd_uobj = uobj_get_read(uobj_get_type(xrcd), cmd->pd_handle,
1402                                           file->ucontext);
1403
1404                 if (IS_ERR(xrcd_uobj)) {
1405                         ret = -EINVAL;
1406                         goto err_put;
1407                 }
1408
1409                 xrcd = (struct ib_xrcd *)xrcd_uobj->object;
1410                 if (!xrcd) {
1411                         ret = -EINVAL;
1412                         goto err_put;
1413                 }
1414                 device = xrcd->device;
1415         } else {
1416                 if (cmd->qp_type == IB_QPT_XRC_INI) {
1417                         cmd->max_recv_wr = 0;
1418                         cmd->max_recv_sge = 0;
1419                 } else {
1420                         if (cmd->is_srq) {
1421                                 srq = uobj_get_obj_read(srq, cmd->srq_handle,
1422                                                         file->ucontext);
1423                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1424                                         ret = -EINVAL;
1425                                         goto err_put;
1426                                 }
1427                         }
1428
1429                         if (!ind_tbl) {
1430                                 if (cmd->recv_cq_handle != cmd->send_cq_handle) {
1431                                         rcq = uobj_get_obj_read(cq, cmd->recv_cq_handle,
1432                                                                 file->ucontext);
1433                                         if (!rcq) {
1434                                                 ret = -EINVAL;
1435                                                 goto err_put;
1436                                         }
1437                                 }
1438                         }
1439                 }
1440
1441                 if (has_sq)
1442                         scq = uobj_get_obj_read(cq, cmd->send_cq_handle,
1443                                                 file->ucontext);
1444                 if (!ind_tbl)
1445                         rcq = rcq ?: scq;
1446                 pd  = uobj_get_obj_read(pd, cmd->pd_handle, file->ucontext);
1447                 if (!pd || (!scq && has_sq)) {
1448                         ret = -EINVAL;
1449                         goto err_put;
1450                 }
1451
1452                 device = pd->device;
1453         }
1454
1455         attr.event_handler = ib_uverbs_qp_event_handler;
1456         attr.qp_context    = file;
1457         attr.send_cq       = scq;
1458         attr.recv_cq       = rcq;
1459         attr.srq           = srq;
1460         attr.xrcd          = xrcd;
1461         attr.sq_sig_type   = cmd->sq_sig_all ? IB_SIGNAL_ALL_WR :
1462                                               IB_SIGNAL_REQ_WR;
1463         attr.qp_type       = cmd->qp_type;
1464         attr.create_flags  = 0;
1465
1466         attr.cap.max_send_wr     = cmd->max_send_wr;
1467         attr.cap.max_recv_wr     = cmd->max_recv_wr;
1468         attr.cap.max_send_sge    = cmd->max_send_sge;
1469         attr.cap.max_recv_sge    = cmd->max_recv_sge;
1470         attr.cap.max_inline_data = cmd->max_inline_data;
1471
1472         obj->uevent.events_reported     = 0;
1473         INIT_LIST_HEAD(&obj->uevent.event_list);
1474         INIT_LIST_HEAD(&obj->mcast_list);
1475
1476         if (cmd_sz >= offsetof(typeof(*cmd), create_flags) +
1477                       sizeof(cmd->create_flags))
1478                 attr.create_flags = cmd->create_flags;
1479
1480         if (attr.create_flags & ~(IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
1481                                 IB_QP_CREATE_CROSS_CHANNEL |
1482                                 IB_QP_CREATE_MANAGED_SEND |
1483                                 IB_QP_CREATE_MANAGED_RECV |
1484                                 IB_QP_CREATE_SCATTER_FCS |
1485                                 IB_QP_CREATE_CVLAN_STRIPPING)) {
1486                 ret = -EINVAL;
1487                 goto err_put;
1488         }
1489
1490         buf = (void *)cmd + sizeof(*cmd);
1491         if (cmd_sz > sizeof(*cmd))
1492                 if (!(buf[0] == 0 && !memcmp(buf, buf + 1,
1493                                              cmd_sz - sizeof(*cmd) - 1))) {
1494                         ret = -EINVAL;
1495                         goto err_put;
1496                 }
1497
1498         if (cmd->qp_type == IB_QPT_XRC_TGT)
1499                 qp = ib_create_qp(pd, &attr);
1500         else
1501                 qp = device->create_qp(pd, &attr, uhw);
1502
1503         if (IS_ERR(qp)) {
1504                 ret = PTR_ERR(qp);
1505                 goto err_put;
1506         }
1507
1508         if (cmd->qp_type != IB_QPT_XRC_TGT) {
1509                 ret = ib_create_qp_security(qp, device);
1510                 if (ret)
1511                         goto err_cb;
1512
1513                 qp->real_qp       = qp;
1514                 qp->device        = device;
1515                 qp->pd            = pd;
1516                 qp->send_cq       = attr.send_cq;
1517                 qp->recv_cq       = attr.recv_cq;
1518                 qp->srq           = attr.srq;
1519                 qp->rwq_ind_tbl   = ind_tbl;
1520                 qp->event_handler = attr.event_handler;
1521                 qp->qp_context    = attr.qp_context;
1522                 qp->qp_type       = attr.qp_type;
1523                 atomic_set(&qp->usecnt, 0);
1524                 atomic_inc(&pd->usecnt);
1525                 qp->port = 0;
1526                 if (attr.send_cq)
1527                         atomic_inc(&attr.send_cq->usecnt);
1528                 if (attr.recv_cq)
1529                         atomic_inc(&attr.recv_cq->usecnt);
1530                 if (attr.srq)
1531                         atomic_inc(&attr.srq->usecnt);
1532                 if (ind_tbl)
1533                         atomic_inc(&ind_tbl->usecnt);
1534         }
1535         qp->uobject = &obj->uevent.uobject;
1536
1537         obj->uevent.uobject.object = qp;
1538
1539         memset(&resp, 0, sizeof resp);
1540         resp.base.qpn             = qp->qp_num;
1541         resp.base.qp_handle       = obj->uevent.uobject.id;
1542         resp.base.max_recv_sge    = attr.cap.max_recv_sge;
1543         resp.base.max_send_sge    = attr.cap.max_send_sge;
1544         resp.base.max_recv_wr     = attr.cap.max_recv_wr;
1545         resp.base.max_send_wr     = attr.cap.max_send_wr;
1546         resp.base.max_inline_data = attr.cap.max_inline_data;
1547
1548         resp.response_length = offsetof(typeof(resp), response_length) +
1549                                sizeof(resp.response_length);
1550
1551         ret = cb(file, &resp, ucore);
1552         if (ret)
1553                 goto err_cb;
1554
1555         if (xrcd) {
1556                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object,
1557                                           uobject);
1558                 atomic_inc(&obj->uxrcd->refcnt);
1559                 uobj_put_read(xrcd_uobj);
1560         }
1561
1562         if (pd)
1563                 uobj_put_obj_read(pd);
1564         if (scq)
1565                 uobj_put_obj_read(scq);
1566         if (rcq && rcq != scq)
1567                 uobj_put_obj_read(rcq);
1568         if (srq)
1569                 uobj_put_obj_read(srq);
1570         if (ind_tbl)
1571                 uobj_put_obj_read(ind_tbl);
1572
1573         uobj_alloc_commit(&obj->uevent.uobject);
1574
1575         return 0;
1576 err_cb:
1577         ib_destroy_qp(qp);
1578
1579 err_put:
1580         if (!IS_ERR(xrcd_uobj))
1581                 uobj_put_read(xrcd_uobj);
1582         if (pd)
1583                 uobj_put_obj_read(pd);
1584         if (scq)
1585                 uobj_put_obj_read(scq);
1586         if (rcq && rcq != scq)
1587                 uobj_put_obj_read(rcq);
1588         if (srq)
1589                 uobj_put_obj_read(srq);
1590         if (ind_tbl)
1591                 uobj_put_obj_read(ind_tbl);
1592
1593         uobj_alloc_abort(&obj->uevent.uobject);
1594         return ret;
1595 }
1596
1597 static int ib_uverbs_create_qp_cb(struct ib_uverbs_file *file,
1598                                   struct ib_uverbs_ex_create_qp_resp *resp,
1599                                   struct ib_udata *ucore)
1600 {
1601         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1602                 return -EFAULT;
1603
1604         return 0;
1605 }
1606
1607 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1608                             struct ib_device *ib_dev,
1609                             const char __user *buf, int in_len,
1610                             int out_len)
1611 {
1612         struct ib_uverbs_create_qp      cmd;
1613         struct ib_uverbs_ex_create_qp   cmd_ex;
1614         struct ib_udata                 ucore;
1615         struct ib_udata                 uhw;
1616         ssize_t resp_size = sizeof(struct ib_uverbs_create_qp_resp);
1617         int                             err;
1618
1619         if (out_len < resp_size)
1620                 return -ENOSPC;
1621
1622         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1623                 return -EFAULT;
1624
1625         INIT_UDATA(&ucore, buf, (unsigned long)cmd.response, sizeof(cmd),
1626                    resp_size);
1627         INIT_UDATA(&uhw, buf + sizeof(cmd),
1628                    (unsigned long)cmd.response + resp_size,
1629                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1630                    out_len - resp_size);
1631
1632         memset(&cmd_ex, 0, sizeof(cmd_ex));
1633         cmd_ex.user_handle = cmd.user_handle;
1634         cmd_ex.pd_handle = cmd.pd_handle;
1635         cmd_ex.send_cq_handle = cmd.send_cq_handle;
1636         cmd_ex.recv_cq_handle = cmd.recv_cq_handle;
1637         cmd_ex.srq_handle = cmd.srq_handle;
1638         cmd_ex.max_send_wr = cmd.max_send_wr;
1639         cmd_ex.max_recv_wr = cmd.max_recv_wr;
1640         cmd_ex.max_send_sge = cmd.max_send_sge;
1641         cmd_ex.max_recv_sge = cmd.max_recv_sge;
1642         cmd_ex.max_inline_data = cmd.max_inline_data;
1643         cmd_ex.sq_sig_all = cmd.sq_sig_all;
1644         cmd_ex.qp_type = cmd.qp_type;
1645         cmd_ex.is_srq = cmd.is_srq;
1646
1647         err = create_qp(file, &ucore, &uhw, &cmd_ex,
1648                         offsetof(typeof(cmd_ex), is_srq) +
1649                         sizeof(cmd.is_srq), ib_uverbs_create_qp_cb,
1650                         NULL);
1651
1652         if (err)
1653                 return err;
1654
1655         return in_len;
1656 }
1657
1658 static int ib_uverbs_ex_create_qp_cb(struct ib_uverbs_file *file,
1659                                      struct ib_uverbs_ex_create_qp_resp *resp,
1660                                      struct ib_udata *ucore)
1661 {
1662         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1663                 return -EFAULT;
1664
1665         return 0;
1666 }
1667
1668 int ib_uverbs_ex_create_qp(struct ib_uverbs_file *file,
1669                            struct ib_device *ib_dev,
1670                            struct ib_udata *ucore,
1671                            struct ib_udata *uhw)
1672 {
1673         struct ib_uverbs_ex_create_qp_resp resp;
1674         struct ib_uverbs_ex_create_qp cmd = {0};
1675         int err;
1676
1677         if (ucore->inlen < (offsetof(typeof(cmd), comp_mask) +
1678                             sizeof(cmd.comp_mask)))
1679                 return -EINVAL;
1680
1681         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
1682         if (err)
1683                 return err;
1684
1685         if (cmd.comp_mask & ~IB_UVERBS_CREATE_QP_SUP_COMP_MASK)
1686                 return -EINVAL;
1687
1688         if (cmd.reserved)
1689                 return -EINVAL;
1690
1691         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1692                              sizeof(resp.response_length)))
1693                 return -ENOSPC;
1694
1695         err = create_qp(file, ucore, uhw, &cmd,
1696                         min(ucore->inlen, sizeof(cmd)),
1697                         ib_uverbs_ex_create_qp_cb, NULL);
1698
1699         if (err)
1700                 return err;
1701
1702         return 0;
1703 }
1704
1705 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1706                           struct ib_device *ib_dev,
1707                           const char __user *buf, int in_len, int out_len)
1708 {
1709         struct ib_uverbs_open_qp        cmd;
1710         struct ib_uverbs_create_qp_resp resp;
1711         struct ib_udata                 udata;
1712         struct ib_uqp_object           *obj;
1713         struct ib_xrcd                 *xrcd;
1714         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1715         struct ib_qp                   *qp;
1716         struct ib_qp_open_attr          attr;
1717         int ret;
1718
1719         if (out_len < sizeof resp)
1720                 return -ENOSPC;
1721
1722         if (copy_from_user(&cmd, buf, sizeof cmd))
1723                 return -EFAULT;
1724
1725         INIT_UDATA(&udata, buf + sizeof cmd,
1726                    (unsigned long) cmd.response + sizeof resp,
1727                    in_len - sizeof cmd, out_len - sizeof resp);
1728
1729         obj  = (struct ib_uqp_object *)uobj_alloc(uobj_get_type(qp),
1730                                                   file->ucontext);
1731         if (IS_ERR(obj))
1732                 return PTR_ERR(obj);
1733
1734         xrcd_uobj = uobj_get_read(uobj_get_type(xrcd), cmd.pd_handle,
1735                                   file->ucontext);
1736         if (IS_ERR(xrcd_uobj)) {
1737                 ret = -EINVAL;
1738                 goto err_put;
1739         }
1740
1741         xrcd = (struct ib_xrcd *)xrcd_uobj->object;
1742         if (!xrcd) {
1743                 ret = -EINVAL;
1744                 goto err_xrcd;
1745         }
1746
1747         attr.event_handler = ib_uverbs_qp_event_handler;
1748         attr.qp_context    = file;
1749         attr.qp_num        = cmd.qpn;
1750         attr.qp_type       = cmd.qp_type;
1751
1752         obj->uevent.events_reported = 0;
1753         INIT_LIST_HEAD(&obj->uevent.event_list);
1754         INIT_LIST_HEAD(&obj->mcast_list);
1755
1756         qp = ib_open_qp(xrcd, &attr);
1757         if (IS_ERR(qp)) {
1758                 ret = PTR_ERR(qp);
1759                 goto err_xrcd;
1760         }
1761
1762         obj->uevent.uobject.object = qp;
1763         obj->uevent.uobject.user_handle = cmd.user_handle;
1764
1765         memset(&resp, 0, sizeof resp);
1766         resp.qpn       = qp->qp_num;
1767         resp.qp_handle = obj->uevent.uobject.id;
1768
1769         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1770                          &resp, sizeof resp)) {
1771                 ret = -EFAULT;
1772                 goto err_destroy;
1773         }
1774
1775         obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
1776         atomic_inc(&obj->uxrcd->refcnt);
1777         qp->uobject = &obj->uevent.uobject;
1778         uobj_put_read(xrcd_uobj);
1779
1780
1781         uobj_alloc_commit(&obj->uevent.uobject);
1782
1783         return in_len;
1784
1785 err_destroy:
1786         ib_destroy_qp(qp);
1787 err_xrcd:
1788         uobj_put_read(xrcd_uobj);
1789 err_put:
1790         uobj_alloc_abort(&obj->uevent.uobject);
1791         return ret;
1792 }
1793
1794 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1795                            struct ib_device *ib_dev,
1796                            const char __user *buf, int in_len,
1797                            int out_len)
1798 {
1799         struct ib_uverbs_query_qp      cmd;
1800         struct ib_uverbs_query_qp_resp resp;
1801         struct ib_qp                   *qp;
1802         struct ib_qp_attr              *attr;
1803         struct ib_qp_init_attr         *init_attr;
1804         const struct ib_global_route   *grh;
1805         int                            ret;
1806
1807         if (copy_from_user(&cmd, buf, sizeof cmd))
1808                 return -EFAULT;
1809
1810         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1811         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1812         if (!attr || !init_attr) {
1813                 ret = -ENOMEM;
1814                 goto out;
1815         }
1816
1817         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
1818         if (!qp) {
1819                 ret = -EINVAL;
1820                 goto out;
1821         }
1822
1823         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1824
1825         uobj_put_obj_read(qp);
1826
1827         if (ret)
1828                 goto out;
1829
1830         memset(&resp, 0, sizeof resp);
1831
1832         resp.qp_state               = attr->qp_state;
1833         resp.cur_qp_state           = attr->cur_qp_state;
1834         resp.path_mtu               = attr->path_mtu;
1835         resp.path_mig_state         = attr->path_mig_state;
1836         resp.qkey                   = attr->qkey;
1837         resp.rq_psn                 = attr->rq_psn;
1838         resp.sq_psn                 = attr->sq_psn;
1839         resp.dest_qp_num            = attr->dest_qp_num;
1840         resp.qp_access_flags        = attr->qp_access_flags;
1841         resp.pkey_index             = attr->pkey_index;
1842         resp.alt_pkey_index         = attr->alt_pkey_index;
1843         resp.sq_draining            = attr->sq_draining;
1844         resp.max_rd_atomic          = attr->max_rd_atomic;
1845         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1846         resp.min_rnr_timer          = attr->min_rnr_timer;
1847         resp.port_num               = attr->port_num;
1848         resp.timeout                = attr->timeout;
1849         resp.retry_cnt              = attr->retry_cnt;
1850         resp.rnr_retry              = attr->rnr_retry;
1851         resp.alt_port_num           = attr->alt_port_num;
1852         resp.alt_timeout            = attr->alt_timeout;
1853
1854         resp.dest.dlid              = rdma_ah_get_dlid(&attr->ah_attr);
1855         resp.dest.sl                = rdma_ah_get_sl(&attr->ah_attr);
1856         resp.dest.src_path_bits     = rdma_ah_get_path_bits(&attr->ah_attr);
1857         resp.dest.static_rate       = rdma_ah_get_static_rate(&attr->ah_attr);
1858         resp.dest.is_global         = !!(rdma_ah_get_ah_flags(&attr->ah_attr) &
1859                                          IB_AH_GRH);
1860         if (resp.dest.is_global) {
1861                 grh = rdma_ah_read_grh(&attr->ah_attr);
1862                 memcpy(resp.dest.dgid, grh->dgid.raw, 16);
1863                 resp.dest.flow_label        = grh->flow_label;
1864                 resp.dest.sgid_index        = grh->sgid_index;
1865                 resp.dest.hop_limit         = grh->hop_limit;
1866                 resp.dest.traffic_class     = grh->traffic_class;
1867         }
1868         resp.dest.port_num          = rdma_ah_get_port_num(&attr->ah_attr);
1869
1870         resp.alt_dest.dlid          = rdma_ah_get_dlid(&attr->alt_ah_attr);
1871         resp.alt_dest.sl            = rdma_ah_get_sl(&attr->alt_ah_attr);
1872         resp.alt_dest.src_path_bits = rdma_ah_get_path_bits(&attr->alt_ah_attr);
1873         resp.alt_dest.static_rate
1874                         = rdma_ah_get_static_rate(&attr->alt_ah_attr);
1875         resp.alt_dest.is_global
1876                         = !!(rdma_ah_get_ah_flags(&attr->alt_ah_attr) &
1877                                                   IB_AH_GRH);
1878         if (resp.alt_dest.is_global) {
1879                 grh = rdma_ah_read_grh(&attr->alt_ah_attr);
1880                 memcpy(resp.alt_dest.dgid, grh->dgid.raw, 16);
1881                 resp.alt_dest.flow_label    = grh->flow_label;
1882                 resp.alt_dest.sgid_index    = grh->sgid_index;
1883                 resp.alt_dest.hop_limit     = grh->hop_limit;
1884                 resp.alt_dest.traffic_class = grh->traffic_class;
1885         }
1886         resp.alt_dest.port_num      = rdma_ah_get_port_num(&attr->alt_ah_attr);
1887
1888         resp.max_send_wr            = init_attr->cap.max_send_wr;
1889         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1890         resp.max_send_sge           = init_attr->cap.max_send_sge;
1891         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1892         resp.max_inline_data        = init_attr->cap.max_inline_data;
1893         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1894
1895         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1896                          &resp, sizeof resp))
1897                 ret = -EFAULT;
1898
1899 out:
1900         kfree(attr);
1901         kfree(init_attr);
1902
1903         return ret ? ret : in_len;
1904 }
1905
1906 /* Remove ignored fields set in the attribute mask */
1907 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1908 {
1909         switch (qp_type) {
1910         case IB_QPT_XRC_INI:
1911                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1912         case IB_QPT_XRC_TGT:
1913                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1914                                 IB_QP_RNR_RETRY);
1915         default:
1916                 return mask;
1917         }
1918 }
1919
1920 static int modify_qp(struct ib_uverbs_file *file,
1921                      struct ib_uverbs_ex_modify_qp *cmd, struct ib_udata *udata)
1922 {
1923         struct ib_qp_attr *attr;
1924         struct ib_qp *qp;
1925         int ret;
1926
1927         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1928         if (!attr)
1929                 return -ENOMEM;
1930
1931         qp = uobj_get_obj_read(qp, cmd->base.qp_handle, file->ucontext);
1932         if (!qp) {
1933                 ret = -EINVAL;
1934                 goto out;
1935         }
1936
1937         if ((cmd->base.attr_mask & IB_QP_PORT) &&
1938             !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         if (cmd->base.attr_mask & IB_QP_AV)
1967                 attr->ah_attr.type = rdma_ah_find_type(qp->device,
1968                                                        cmd->base.dest.port_num);
1969         if (cmd->base.dest.is_global) {
1970                 rdma_ah_set_grh(&attr->ah_attr, NULL,
1971                                 cmd->base.dest.flow_label,
1972                                 cmd->base.dest.sgid_index,
1973                                 cmd->base.dest.hop_limit,
1974                                 cmd->base.dest.traffic_class);
1975                 rdma_ah_set_dgid_raw(&attr->ah_attr, cmd->base.dest.dgid);
1976         } else {
1977                 rdma_ah_set_ah_flags(&attr->ah_attr, 0);
1978         }
1979         rdma_ah_set_dlid(&attr->ah_attr, cmd->base.dest.dlid);
1980         rdma_ah_set_sl(&attr->ah_attr, cmd->base.dest.sl);
1981         rdma_ah_set_path_bits(&attr->ah_attr, cmd->base.dest.src_path_bits);
1982         rdma_ah_set_static_rate(&attr->ah_attr, cmd->base.dest.static_rate);
1983         rdma_ah_set_port_num(&attr->ah_attr,
1984                              cmd->base.dest.port_num);
1985
1986         if (cmd->base.attr_mask & IB_QP_ALT_PATH)
1987                 attr->alt_ah_attr.type =
1988                         rdma_ah_find_type(qp->device, cmd->base.dest.port_num);
1989         if (cmd->base.alt_dest.is_global) {
1990                 rdma_ah_set_grh(&attr->alt_ah_attr, NULL,
1991                                 cmd->base.alt_dest.flow_label,
1992                                 cmd->base.alt_dest.sgid_index,
1993                                 cmd->base.alt_dest.hop_limit,
1994                                 cmd->base.alt_dest.traffic_class);
1995                 rdma_ah_set_dgid_raw(&attr->alt_ah_attr,
1996                                      cmd->base.alt_dest.dgid);
1997         } else {
1998                 rdma_ah_set_ah_flags(&attr->alt_ah_attr, 0);
1999         }
2000
2001         rdma_ah_set_dlid(&attr->alt_ah_attr, cmd->base.alt_dest.dlid);
2002         rdma_ah_set_sl(&attr->alt_ah_attr, cmd->base.alt_dest.sl);
2003         rdma_ah_set_path_bits(&attr->alt_ah_attr,
2004                               cmd->base.alt_dest.src_path_bits);
2005         rdma_ah_set_static_rate(&attr->alt_ah_attr,
2006                                 cmd->base.alt_dest.static_rate);
2007         rdma_ah_set_port_num(&attr->alt_ah_attr,
2008                              cmd->base.alt_dest.port_num);
2009
2010         ret = ib_modify_qp_with_udata(qp, attr,
2011                                       modify_qp_mask(qp->qp_type,
2012                                                      cmd->base.attr_mask),
2013                                       udata);
2014
2015 release_qp:
2016         uobj_put_obj_read(qp);
2017 out:
2018         kfree(attr);
2019
2020         return ret;
2021 }
2022
2023 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
2024                             struct ib_device *ib_dev,
2025                             const char __user *buf, int in_len,
2026                             int out_len)
2027 {
2028         struct ib_uverbs_ex_modify_qp cmd = {};
2029         struct ib_udata udata;
2030         int ret;
2031
2032         if (copy_from_user(&cmd.base, buf, sizeof(cmd.base)))
2033                 return -EFAULT;
2034
2035         if (cmd.base.attr_mask &
2036             ~((IB_USER_LEGACY_LAST_QP_ATTR_MASK << 1) - 1))
2037                 return -EOPNOTSUPP;
2038
2039         INIT_UDATA(&udata, buf + sizeof(cmd.base), NULL,
2040                    in_len - sizeof(cmd.base), out_len);
2041
2042         ret = modify_qp(file, &cmd, &udata);
2043         if (ret)
2044                 return ret;
2045
2046         return in_len;
2047 }
2048
2049 int ib_uverbs_ex_modify_qp(struct ib_uverbs_file *file,
2050                            struct ib_device *ib_dev,
2051                            struct ib_udata *ucore,
2052                            struct ib_udata *uhw)
2053 {
2054         struct ib_uverbs_ex_modify_qp cmd = {};
2055         int ret;
2056
2057         /*
2058          * Last bit is reserved for extending the attr_mask by
2059          * using another field.
2060          */
2061         BUILD_BUG_ON(IB_USER_LAST_QP_ATTR_MASK == (1 << 31));
2062
2063         if (ucore->inlen < sizeof(cmd.base))
2064                 return -EINVAL;
2065
2066         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2067         if (ret)
2068                 return ret;
2069
2070         if (cmd.base.attr_mask &
2071             ~((IB_USER_LAST_QP_ATTR_MASK << 1) - 1))
2072                 return -EOPNOTSUPP;
2073
2074         if (ucore->inlen > sizeof(cmd)) {
2075                 if (ib_is_udata_cleared(ucore, sizeof(cmd),
2076                                         ucore->inlen - sizeof(cmd)))
2077                         return -EOPNOTSUPP;
2078         }
2079
2080         ret = modify_qp(file, &cmd, uhw);
2081
2082         return ret;
2083 }
2084
2085 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
2086                              struct ib_device *ib_dev,
2087                              const char __user *buf, int in_len,
2088                              int out_len)
2089 {
2090         struct ib_uverbs_destroy_qp      cmd;
2091         struct ib_uverbs_destroy_qp_resp resp;
2092         struct ib_uobject               *uobj;
2093         struct ib_uqp_object            *obj;
2094         int                              ret = -EINVAL;
2095
2096         if (copy_from_user(&cmd, buf, sizeof cmd))
2097                 return -EFAULT;
2098
2099         memset(&resp, 0, sizeof resp);
2100
2101         uobj  = uobj_get_write(uobj_get_type(qp), cmd.qp_handle,
2102                                file->ucontext);
2103         if (IS_ERR(uobj))
2104                 return PTR_ERR(uobj);
2105
2106         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2107         /*
2108          * Make sure we don't free the memory in remove_commit as we still
2109          * needs the uobject memory to create the response.
2110          */
2111         uverbs_uobject_get(uobj);
2112
2113         ret = uobj_remove_commit(uobj);
2114         if (ret) {
2115                 uverbs_uobject_put(uobj);
2116                 return ret;
2117         }
2118
2119         resp.events_reported = obj->uevent.events_reported;
2120         uverbs_uobject_put(uobj);
2121
2122         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2123                          &resp, sizeof resp))
2124                 return -EFAULT;
2125
2126         return in_len;
2127 }
2128
2129 static void *alloc_wr(size_t wr_size, __u32 num_sge)
2130 {
2131         if (num_sge >= (U32_MAX - ALIGN(wr_size, sizeof (struct ib_sge))) /
2132                        sizeof (struct ib_sge))
2133                 return NULL;
2134
2135         return kmalloc(ALIGN(wr_size, sizeof (struct ib_sge)) +
2136                          num_sge * sizeof (struct ib_sge), GFP_KERNEL);
2137 }
2138
2139 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2140                             struct ib_device *ib_dev,
2141                             const char __user *buf, int in_len,
2142                             int out_len)
2143 {
2144         struct ib_uverbs_post_send      cmd;
2145         struct ib_uverbs_post_send_resp resp;
2146         struct ib_uverbs_send_wr       *user_wr;
2147         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
2148         struct ib_qp                   *qp;
2149         int                             i, sg_ind;
2150         int                             is_ud;
2151         ssize_t                         ret = -EINVAL;
2152         size_t                          next_size;
2153
2154         if (copy_from_user(&cmd, buf, sizeof cmd))
2155                 return -EFAULT;
2156
2157         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2158             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2159                 return -EINVAL;
2160
2161         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2162                 return -EINVAL;
2163
2164         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2165         if (!user_wr)
2166                 return -ENOMEM;
2167
2168         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
2169         if (!qp)
2170                 goto out;
2171
2172         is_ud = qp->qp_type == IB_QPT_UD;
2173         sg_ind = 0;
2174         last = NULL;
2175         for (i = 0; i < cmd.wr_count; ++i) {
2176                 if (copy_from_user(user_wr,
2177                                    buf + sizeof cmd + i * cmd.wqe_size,
2178                                    cmd.wqe_size)) {
2179                         ret = -EFAULT;
2180                         goto out_put;
2181                 }
2182
2183                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2184                         ret = -EINVAL;
2185                         goto out_put;
2186                 }
2187
2188                 if (is_ud) {
2189                         struct ib_ud_wr *ud;
2190
2191                         if (user_wr->opcode != IB_WR_SEND &&
2192                             user_wr->opcode != IB_WR_SEND_WITH_IMM) {
2193                                 ret = -EINVAL;
2194                                 goto out_put;
2195                         }
2196
2197                         next_size = sizeof(*ud);
2198                         ud = alloc_wr(next_size, user_wr->num_sge);
2199                         if (!ud) {
2200                                 ret = -ENOMEM;
2201                                 goto out_put;
2202                         }
2203
2204                         ud->ah = uobj_get_obj_read(ah, user_wr->wr.ud.ah,
2205                                                    file->ucontext);
2206                         if (!ud->ah) {
2207                                 kfree(ud);
2208                                 ret = -EINVAL;
2209                                 goto out_put;
2210                         }
2211                         ud->remote_qpn = user_wr->wr.ud.remote_qpn;
2212                         ud->remote_qkey = user_wr->wr.ud.remote_qkey;
2213
2214                         next = &ud->wr;
2215                 } else if (user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
2216                            user_wr->opcode == IB_WR_RDMA_WRITE ||
2217                            user_wr->opcode == IB_WR_RDMA_READ) {
2218                         struct ib_rdma_wr *rdma;
2219
2220                         next_size = sizeof(*rdma);
2221                         rdma = alloc_wr(next_size, user_wr->num_sge);
2222                         if (!rdma) {
2223                                 ret = -ENOMEM;
2224                                 goto out_put;
2225                         }
2226
2227                         rdma->remote_addr = user_wr->wr.rdma.remote_addr;
2228                         rdma->rkey = user_wr->wr.rdma.rkey;
2229
2230                         next = &rdma->wr;
2231                 } else if (user_wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2232                            user_wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2233                         struct ib_atomic_wr *atomic;
2234
2235                         next_size = sizeof(*atomic);
2236                         atomic = alloc_wr(next_size, user_wr->num_sge);
2237                         if (!atomic) {
2238                                 ret = -ENOMEM;
2239                                 goto out_put;
2240                         }
2241
2242                         atomic->remote_addr = user_wr->wr.atomic.remote_addr;
2243                         atomic->compare_add = user_wr->wr.atomic.compare_add;
2244                         atomic->swap = user_wr->wr.atomic.swap;
2245                         atomic->rkey = user_wr->wr.atomic.rkey;
2246
2247                         next = &atomic->wr;
2248                 } else if (user_wr->opcode == IB_WR_SEND ||
2249                            user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2250                            user_wr->opcode == IB_WR_SEND_WITH_INV) {
2251                         next_size = sizeof(*next);
2252                         next = alloc_wr(next_size, user_wr->num_sge);
2253                         if (!next) {
2254                                 ret = -ENOMEM;
2255                                 goto out_put;
2256                         }
2257                 } else {
2258                         ret = -EINVAL;
2259                         goto out_put;
2260                 }
2261
2262                 if (user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2263                     user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
2264                         next->ex.imm_data =
2265                                         (__be32 __force) user_wr->ex.imm_data;
2266                 } else if (user_wr->opcode == IB_WR_SEND_WITH_INV) {
2267                         next->ex.invalidate_rkey = user_wr->ex.invalidate_rkey;
2268                 }
2269
2270                 if (!last)
2271                         wr = next;
2272                 else
2273                         last->next = next;
2274                 last = next;
2275
2276                 next->next       = NULL;
2277                 next->wr_id      = user_wr->wr_id;
2278                 next->num_sge    = user_wr->num_sge;
2279                 next->opcode     = user_wr->opcode;
2280                 next->send_flags = user_wr->send_flags;
2281
2282                 if (next->num_sge) {
2283                         next->sg_list = (void *) next +
2284                                 ALIGN(next_size, sizeof(struct ib_sge));
2285                         if (copy_from_user(next->sg_list,
2286                                            buf + sizeof cmd +
2287                                            cmd.wr_count * cmd.wqe_size +
2288                                            sg_ind * sizeof (struct ib_sge),
2289                                            next->num_sge * sizeof (struct ib_sge))) {
2290                                 ret = -EFAULT;
2291                                 goto out_put;
2292                         }
2293                         sg_ind += next->num_sge;
2294                 } else
2295                         next->sg_list = NULL;
2296         }
2297
2298         resp.bad_wr = 0;
2299         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2300         if (ret)
2301                 for (next = wr; next; next = next->next) {
2302                         ++resp.bad_wr;
2303                         if (next == bad_wr)
2304                                 break;
2305                 }
2306
2307         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2308                          &resp, sizeof resp))
2309                 ret = -EFAULT;
2310
2311 out_put:
2312         uobj_put_obj_read(qp);
2313
2314         while (wr) {
2315                 if (is_ud && ud_wr(wr)->ah)
2316                         uobj_put_obj_read(ud_wr(wr)->ah);
2317                 next = wr->next;
2318                 kfree(wr);
2319                 wr = next;
2320         }
2321
2322 out:
2323         kfree(user_wr);
2324
2325         return ret ? ret : in_len;
2326 }
2327
2328 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2329                                                     int in_len,
2330                                                     u32 wr_count,
2331                                                     u32 sge_count,
2332                                                     u32 wqe_size)
2333 {
2334         struct ib_uverbs_recv_wr *user_wr;
2335         struct ib_recv_wr        *wr = NULL, *last, *next;
2336         int                       sg_ind;
2337         int                       i;
2338         int                       ret;
2339
2340         if (in_len < wqe_size * wr_count +
2341             sge_count * sizeof (struct ib_uverbs_sge))
2342                 return ERR_PTR(-EINVAL);
2343
2344         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2345                 return ERR_PTR(-EINVAL);
2346
2347         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2348         if (!user_wr)
2349                 return ERR_PTR(-ENOMEM);
2350
2351         sg_ind = 0;
2352         last = NULL;
2353         for (i = 0; i < wr_count; ++i) {
2354                 if (copy_from_user(user_wr, buf + i * wqe_size,
2355                                    wqe_size)) {
2356                         ret = -EFAULT;
2357                         goto err;
2358                 }
2359
2360                 if (user_wr->num_sge + sg_ind > sge_count) {
2361                         ret = -EINVAL;
2362                         goto err;
2363                 }
2364
2365                 if (user_wr->num_sge >=
2366                     (U32_MAX - ALIGN(sizeof *next, sizeof (struct ib_sge))) /
2367                     sizeof (struct ib_sge)) {
2368                         ret = -EINVAL;
2369                         goto err;
2370                 }
2371
2372                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2373                                user_wr->num_sge * sizeof (struct ib_sge),
2374                                GFP_KERNEL);
2375                 if (!next) {
2376                         ret = -ENOMEM;
2377                         goto err;
2378                 }
2379
2380                 if (!last)
2381                         wr = next;
2382                 else
2383                         last->next = next;
2384                 last = next;
2385
2386                 next->next       = NULL;
2387                 next->wr_id      = user_wr->wr_id;
2388                 next->num_sge    = user_wr->num_sge;
2389
2390                 if (next->num_sge) {
2391                         next->sg_list = (void *) next +
2392                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2393                         if (copy_from_user(next->sg_list,
2394                                            buf + wr_count * wqe_size +
2395                                            sg_ind * sizeof (struct ib_sge),
2396                                            next->num_sge * sizeof (struct ib_sge))) {
2397                                 ret = -EFAULT;
2398                                 goto err;
2399                         }
2400                         sg_ind += next->num_sge;
2401                 } else
2402                         next->sg_list = NULL;
2403         }
2404
2405         kfree(user_wr);
2406         return wr;
2407
2408 err:
2409         kfree(user_wr);
2410
2411         while (wr) {
2412                 next = wr->next;
2413                 kfree(wr);
2414                 wr = next;
2415         }
2416
2417         return ERR_PTR(ret);
2418 }
2419
2420 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2421                             struct ib_device *ib_dev,
2422                             const char __user *buf, int in_len,
2423                             int out_len)
2424 {
2425         struct ib_uverbs_post_recv      cmd;
2426         struct ib_uverbs_post_recv_resp resp;
2427         struct ib_recv_wr              *wr, *next, *bad_wr;
2428         struct ib_qp                   *qp;
2429         ssize_t                         ret = -EINVAL;
2430
2431         if (copy_from_user(&cmd, buf, sizeof cmd))
2432                 return -EFAULT;
2433
2434         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2435                                        in_len - sizeof cmd, cmd.wr_count,
2436                                        cmd.sge_count, cmd.wqe_size);
2437         if (IS_ERR(wr))
2438                 return PTR_ERR(wr);
2439
2440         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
2441         if (!qp)
2442                 goto out;
2443
2444         resp.bad_wr = 0;
2445         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2446
2447         uobj_put_obj_read(qp);
2448         if (ret) {
2449                 for (next = wr; next; next = next->next) {
2450                         ++resp.bad_wr;
2451                         if (next == bad_wr)
2452                                 break;
2453                 }
2454         }
2455
2456         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2457                          &resp, sizeof resp))
2458                 ret = -EFAULT;
2459
2460 out:
2461         while (wr) {
2462                 next = wr->next;
2463                 kfree(wr);
2464                 wr = next;
2465         }
2466
2467         return ret ? ret : in_len;
2468 }
2469
2470 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2471                                 struct ib_device *ib_dev,
2472                                 const char __user *buf, int in_len,
2473                                 int out_len)
2474 {
2475         struct ib_uverbs_post_srq_recv      cmd;
2476         struct ib_uverbs_post_srq_recv_resp resp;
2477         struct ib_recv_wr                  *wr, *next, *bad_wr;
2478         struct ib_srq                      *srq;
2479         ssize_t                             ret = -EINVAL;
2480
2481         if (copy_from_user(&cmd, buf, sizeof cmd))
2482                 return -EFAULT;
2483
2484         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2485                                        in_len - sizeof cmd, cmd.wr_count,
2486                                        cmd.sge_count, cmd.wqe_size);
2487         if (IS_ERR(wr))
2488                 return PTR_ERR(wr);
2489
2490         srq = uobj_get_obj_read(srq, cmd.srq_handle, file->ucontext);
2491         if (!srq)
2492                 goto out;
2493
2494         resp.bad_wr = 0;
2495         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2496
2497         uobj_put_obj_read(srq);
2498
2499         if (ret)
2500                 for (next = wr; next; next = next->next) {
2501                         ++resp.bad_wr;
2502                         if (next == bad_wr)
2503                                 break;
2504                 }
2505
2506         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2507                          &resp, sizeof resp))
2508                 ret = -EFAULT;
2509
2510 out:
2511         while (wr) {
2512                 next = wr->next;
2513                 kfree(wr);
2514                 wr = next;
2515         }
2516
2517         return ret ? ret : in_len;
2518 }
2519
2520 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2521                             struct ib_device *ib_dev,
2522                             const char __user *buf, int in_len,
2523                             int out_len)
2524 {
2525         struct ib_uverbs_create_ah       cmd;
2526         struct ib_uverbs_create_ah_resp  resp;
2527         struct ib_uobject               *uobj;
2528         struct ib_pd                    *pd;
2529         struct ib_ah                    *ah;
2530         struct rdma_ah_attr             attr;
2531         int ret;
2532         struct ib_udata                   udata;
2533         u8                              *dmac;
2534
2535         if (out_len < sizeof resp)
2536                 return -ENOSPC;
2537
2538         if (copy_from_user(&cmd, buf, sizeof cmd))
2539                 return -EFAULT;
2540
2541         if (!rdma_is_port_valid(ib_dev, cmd.attr.port_num))
2542                 return -EINVAL;
2543
2544         INIT_UDATA(&udata, buf + sizeof(cmd),
2545                    (unsigned long)cmd.response + sizeof(resp),
2546                    in_len - sizeof(cmd), out_len - sizeof(resp));
2547
2548         uobj  = uobj_alloc(uobj_get_type(ah), file->ucontext);
2549         if (IS_ERR(uobj))
2550                 return PTR_ERR(uobj);
2551
2552         pd = uobj_get_obj_read(pd, cmd.pd_handle, file->ucontext);
2553         if (!pd) {
2554                 ret = -EINVAL;
2555                 goto err;
2556         }
2557
2558         attr.type = rdma_ah_find_type(ib_dev, cmd.attr.port_num);
2559         rdma_ah_set_dlid(&attr, cmd.attr.dlid);
2560         rdma_ah_set_sl(&attr, cmd.attr.sl);
2561         rdma_ah_set_path_bits(&attr, cmd.attr.src_path_bits);
2562         rdma_ah_set_static_rate(&attr, cmd.attr.static_rate);
2563         rdma_ah_set_port_num(&attr, cmd.attr.port_num);
2564
2565         if (cmd.attr.is_global) {
2566                 rdma_ah_set_grh(&attr, NULL, cmd.attr.grh.flow_label,
2567                                 cmd.attr.grh.sgid_index,
2568                                 cmd.attr.grh.hop_limit,
2569                                 cmd.attr.grh.traffic_class);
2570                 rdma_ah_set_dgid_raw(&attr, cmd.attr.grh.dgid);
2571         } else {
2572                 rdma_ah_set_ah_flags(&attr, 0);
2573         }
2574         dmac = rdma_ah_retrieve_dmac(&attr);
2575         if (dmac)
2576                 memset(dmac, 0, ETH_ALEN);
2577
2578         ah = pd->device->create_ah(pd, &attr, &udata);
2579
2580         if (IS_ERR(ah)) {
2581                 ret = PTR_ERR(ah);
2582                 goto err_put;
2583         }
2584
2585         ah->device  = pd->device;
2586         ah->pd      = pd;
2587         atomic_inc(&pd->usecnt);
2588         ah->uobject  = uobj;
2589         uobj->user_handle = cmd.user_handle;
2590         uobj->object = ah;
2591
2592         resp.ah_handle = uobj->id;
2593
2594         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2595                          &resp, sizeof resp)) {
2596                 ret = -EFAULT;
2597                 goto err_copy;
2598         }
2599
2600         uobj_put_obj_read(pd);
2601         uobj_alloc_commit(uobj);
2602
2603         return in_len;
2604
2605 err_copy:
2606         rdma_destroy_ah(ah);
2607
2608 err_put:
2609         uobj_put_obj_read(pd);
2610
2611 err:
2612         uobj_alloc_abort(uobj);
2613         return ret;
2614 }
2615
2616 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2617                              struct ib_device *ib_dev,
2618                              const char __user *buf, int in_len, int out_len)
2619 {
2620         struct ib_uverbs_destroy_ah cmd;
2621         struct ib_uobject          *uobj;
2622         int                         ret;
2623
2624         if (copy_from_user(&cmd, buf, sizeof cmd))
2625                 return -EFAULT;
2626
2627         uobj  = uobj_get_write(uobj_get_type(ah), cmd.ah_handle,
2628                                file->ucontext);
2629         if (IS_ERR(uobj))
2630                 return PTR_ERR(uobj);
2631
2632         ret = uobj_remove_commit(uobj);
2633         return ret ?: in_len;
2634 }
2635
2636 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2637                                struct ib_device *ib_dev,
2638                                const char __user *buf, int in_len,
2639                                int out_len)
2640 {
2641         struct ib_uverbs_attach_mcast cmd;
2642         struct ib_qp                 *qp;
2643         struct ib_uqp_object         *obj;
2644         struct ib_uverbs_mcast_entry *mcast;
2645         int                           ret;
2646
2647         if (copy_from_user(&cmd, buf, sizeof cmd))
2648                 return -EFAULT;
2649
2650         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
2651         if (!qp)
2652                 return -EINVAL;
2653
2654         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2655
2656         mutex_lock(&obj->mcast_lock);
2657         list_for_each_entry(mcast, &obj->mcast_list, list)
2658                 if (cmd.mlid == mcast->lid &&
2659                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2660                         ret = 0;
2661                         goto out_put;
2662                 }
2663
2664         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2665         if (!mcast) {
2666                 ret = -ENOMEM;
2667                 goto out_put;
2668         }
2669
2670         mcast->lid = cmd.mlid;
2671         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2672
2673         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2674         if (!ret)
2675                 list_add_tail(&mcast->list, &obj->mcast_list);
2676         else
2677                 kfree(mcast);
2678
2679 out_put:
2680         mutex_unlock(&obj->mcast_lock);
2681         uobj_put_obj_read(qp);
2682
2683         return ret ? ret : in_len;
2684 }
2685
2686 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2687                                struct ib_device *ib_dev,
2688                                const char __user *buf, int in_len,
2689                                int out_len)
2690 {
2691         struct ib_uverbs_detach_mcast cmd;
2692         struct ib_uqp_object         *obj;
2693         struct ib_qp                 *qp;
2694         struct ib_uverbs_mcast_entry *mcast;
2695         int                           ret = -EINVAL;
2696         bool                          found = false;
2697
2698         if (copy_from_user(&cmd, buf, sizeof cmd))
2699                 return -EFAULT;
2700
2701         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
2702         if (!qp)
2703                 return -EINVAL;
2704
2705         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2706         mutex_lock(&obj->mcast_lock);
2707
2708         list_for_each_entry(mcast, &obj->mcast_list, list)
2709                 if (cmd.mlid == mcast->lid &&
2710                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2711                         list_del(&mcast->list);
2712                         kfree(mcast);
2713                         found = true;
2714                         break;
2715                 }
2716
2717         if (!found) {
2718                 ret = -EINVAL;
2719                 goto out_put;
2720         }
2721
2722         ret = ib_detach_mcast(qp, (union ib_gid *)cmd.gid, cmd.mlid);
2723
2724 out_put:
2725         mutex_unlock(&obj->mcast_lock);
2726         uobj_put_obj_read(qp);
2727         return ret ? ret : in_len;
2728 }
2729
2730 static int kern_spec_to_ib_spec_action(struct ib_uverbs_flow_spec *kern_spec,
2731                                        union ib_flow_spec *ib_spec)
2732 {
2733         ib_spec->type = kern_spec->type;
2734         switch (ib_spec->type) {
2735         case IB_FLOW_SPEC_ACTION_TAG:
2736                 if (kern_spec->flow_tag.size !=
2737                     sizeof(struct ib_uverbs_flow_spec_action_tag))
2738                         return -EINVAL;
2739
2740                 ib_spec->flow_tag.size = sizeof(struct ib_flow_spec_action_tag);
2741                 ib_spec->flow_tag.tag_id = kern_spec->flow_tag.tag_id;
2742                 break;
2743         case IB_FLOW_SPEC_ACTION_DROP:
2744                 if (kern_spec->drop.size !=
2745                     sizeof(struct ib_uverbs_flow_spec_action_drop))
2746                         return -EINVAL;
2747
2748                 ib_spec->drop.size = sizeof(struct ib_flow_spec_action_drop);
2749                 break;
2750         default:
2751                 return -EINVAL;
2752         }
2753         return 0;
2754 }
2755
2756 static size_t kern_spec_filter_sz(struct ib_uverbs_flow_spec_hdr *spec)
2757 {
2758         /* Returns user space filter size, includes padding */
2759         return (spec->size - sizeof(struct ib_uverbs_flow_spec_hdr)) / 2;
2760 }
2761
2762 static ssize_t spec_filter_size(void *kern_spec_filter, u16 kern_filter_size,
2763                                 u16 ib_real_filter_sz)
2764 {
2765         /*
2766          * User space filter structures must be 64 bit aligned, otherwise this
2767          * may pass, but we won't handle additional new attributes.
2768          */
2769
2770         if (kern_filter_size > ib_real_filter_sz) {
2771                 if (memchr_inv(kern_spec_filter +
2772                                ib_real_filter_sz, 0,
2773                                kern_filter_size - ib_real_filter_sz))
2774                         return -EINVAL;
2775                 return ib_real_filter_sz;
2776         }
2777         return kern_filter_size;
2778 }
2779
2780 static int kern_spec_to_ib_spec_filter(struct ib_uverbs_flow_spec *kern_spec,
2781                                        union ib_flow_spec *ib_spec)
2782 {
2783         ssize_t actual_filter_sz;
2784         ssize_t kern_filter_sz;
2785         ssize_t ib_filter_sz;
2786         void *kern_spec_mask;
2787         void *kern_spec_val;
2788
2789         if (kern_spec->reserved)
2790                 return -EINVAL;
2791
2792         ib_spec->type = kern_spec->type;
2793
2794         kern_filter_sz = kern_spec_filter_sz(&kern_spec->hdr);
2795         /* User flow spec size must be aligned to 4 bytes */
2796         if (kern_filter_sz != ALIGN(kern_filter_sz, 4))
2797                 return -EINVAL;
2798
2799         kern_spec_val = (void *)kern_spec +
2800                 sizeof(struct ib_uverbs_flow_spec_hdr);
2801         kern_spec_mask = kern_spec_val + kern_filter_sz;
2802         if (ib_spec->type == (IB_FLOW_SPEC_INNER | IB_FLOW_SPEC_VXLAN_TUNNEL))
2803                 return -EINVAL;
2804
2805         switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2806         case IB_FLOW_SPEC_ETH:
2807                 ib_filter_sz = offsetof(struct ib_flow_eth_filter, real_sz);
2808                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2809                                                     kern_filter_sz,
2810                                                     ib_filter_sz);
2811                 if (actual_filter_sz <= 0)
2812                         return -EINVAL;
2813                 ib_spec->size = sizeof(struct ib_flow_spec_eth);
2814                 memcpy(&ib_spec->eth.val, kern_spec_val, actual_filter_sz);
2815                 memcpy(&ib_spec->eth.mask, kern_spec_mask, actual_filter_sz);
2816                 break;
2817         case IB_FLOW_SPEC_IPV4:
2818                 ib_filter_sz = offsetof(struct ib_flow_ipv4_filter, real_sz);
2819                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2820                                                     kern_filter_sz,
2821                                                     ib_filter_sz);
2822                 if (actual_filter_sz <= 0)
2823                         return -EINVAL;
2824                 ib_spec->size = sizeof(struct ib_flow_spec_ipv4);
2825                 memcpy(&ib_spec->ipv4.val, kern_spec_val, actual_filter_sz);
2826                 memcpy(&ib_spec->ipv4.mask, kern_spec_mask, actual_filter_sz);
2827                 break;
2828         case IB_FLOW_SPEC_IPV6:
2829                 ib_filter_sz = offsetof(struct ib_flow_ipv6_filter, real_sz);
2830                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2831                                                     kern_filter_sz,
2832                                                     ib_filter_sz);
2833                 if (actual_filter_sz <= 0)
2834                         return -EINVAL;
2835                 ib_spec->size = sizeof(struct ib_flow_spec_ipv6);
2836                 memcpy(&ib_spec->ipv6.val, kern_spec_val, actual_filter_sz);
2837                 memcpy(&ib_spec->ipv6.mask, kern_spec_mask, actual_filter_sz);
2838
2839                 if ((ntohl(ib_spec->ipv6.mask.flow_label)) >= BIT(20) ||
2840                     (ntohl(ib_spec->ipv6.val.flow_label)) >= BIT(20))
2841                         return -EINVAL;
2842                 break;
2843         case IB_FLOW_SPEC_TCP:
2844         case IB_FLOW_SPEC_UDP:
2845                 ib_filter_sz = offsetof(struct ib_flow_tcp_udp_filter, real_sz);
2846                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2847                                                     kern_filter_sz,
2848                                                     ib_filter_sz);
2849                 if (actual_filter_sz <= 0)
2850                         return -EINVAL;
2851                 ib_spec->size = sizeof(struct ib_flow_spec_tcp_udp);
2852                 memcpy(&ib_spec->tcp_udp.val, kern_spec_val, actual_filter_sz);
2853                 memcpy(&ib_spec->tcp_udp.mask, kern_spec_mask, actual_filter_sz);
2854                 break;
2855         case IB_FLOW_SPEC_VXLAN_TUNNEL:
2856                 ib_filter_sz = offsetof(struct ib_flow_tunnel_filter, real_sz);
2857                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2858                                                     kern_filter_sz,
2859                                                     ib_filter_sz);
2860                 if (actual_filter_sz <= 0)
2861                         return -EINVAL;
2862                 ib_spec->tunnel.size = sizeof(struct ib_flow_spec_tunnel);
2863                 memcpy(&ib_spec->tunnel.val, kern_spec_val, actual_filter_sz);
2864                 memcpy(&ib_spec->tunnel.mask, kern_spec_mask, actual_filter_sz);
2865
2866                 if ((ntohl(ib_spec->tunnel.mask.tunnel_id)) >= BIT(24) ||
2867                     (ntohl(ib_spec->tunnel.val.tunnel_id)) >= BIT(24))
2868                         return -EINVAL;
2869                 break;
2870         default:
2871                 return -EINVAL;
2872         }
2873         return 0;
2874 }
2875
2876 static int kern_spec_to_ib_spec(struct ib_uverbs_flow_spec *kern_spec,
2877                                 union ib_flow_spec *ib_spec)
2878 {
2879         if (kern_spec->reserved)
2880                 return -EINVAL;
2881
2882         if (kern_spec->type >= IB_FLOW_SPEC_ACTION_TAG)
2883                 return kern_spec_to_ib_spec_action(kern_spec, ib_spec);
2884         else
2885                 return kern_spec_to_ib_spec_filter(kern_spec, ib_spec);
2886 }
2887
2888 int ib_uverbs_ex_create_wq(struct ib_uverbs_file *file,
2889                            struct ib_device *ib_dev,
2890                            struct ib_udata *ucore,
2891                            struct ib_udata *uhw)
2892 {
2893         struct ib_uverbs_ex_create_wq     cmd = {};
2894         struct ib_uverbs_ex_create_wq_resp resp = {};
2895         struct ib_uwq_object           *obj;
2896         int err = 0;
2897         struct ib_cq *cq;
2898         struct ib_pd *pd;
2899         struct ib_wq *wq;
2900         struct ib_wq_init_attr wq_init_attr = {};
2901         size_t required_cmd_sz;
2902         size_t required_resp_len;
2903
2904         required_cmd_sz = offsetof(typeof(cmd), max_sge) + sizeof(cmd.max_sge);
2905         required_resp_len = offsetof(typeof(resp), wqn) + sizeof(resp.wqn);
2906
2907         if (ucore->inlen < required_cmd_sz)
2908                 return -EINVAL;
2909
2910         if (ucore->outlen < required_resp_len)
2911                 return -ENOSPC;
2912
2913         if (ucore->inlen > sizeof(cmd) &&
2914             !ib_is_udata_cleared(ucore, sizeof(cmd),
2915                                  ucore->inlen - sizeof(cmd)))
2916                 return -EOPNOTSUPP;
2917
2918         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2919         if (err)
2920                 return err;
2921
2922         if (cmd.comp_mask)
2923                 return -EOPNOTSUPP;
2924
2925         obj  = (struct ib_uwq_object *)uobj_alloc(uobj_get_type(wq),
2926                                                   file->ucontext);
2927         if (IS_ERR(obj))
2928                 return PTR_ERR(obj);
2929
2930         pd  = uobj_get_obj_read(pd, cmd.pd_handle, file->ucontext);
2931         if (!pd) {
2932                 err = -EINVAL;
2933                 goto err_uobj;
2934         }
2935
2936         cq = uobj_get_obj_read(cq, cmd.cq_handle, file->ucontext);
2937         if (!cq) {
2938                 err = -EINVAL;
2939                 goto err_put_pd;
2940         }
2941
2942         wq_init_attr.cq = cq;
2943         wq_init_attr.max_sge = cmd.max_sge;
2944         wq_init_attr.max_wr = cmd.max_wr;
2945         wq_init_attr.wq_context = file;
2946         wq_init_attr.wq_type = cmd.wq_type;
2947         wq_init_attr.event_handler = ib_uverbs_wq_event_handler;
2948         if (ucore->inlen >= (offsetof(typeof(cmd), create_flags) +
2949                              sizeof(cmd.create_flags)))
2950                 wq_init_attr.create_flags = cmd.create_flags;
2951         obj->uevent.events_reported = 0;
2952         INIT_LIST_HEAD(&obj->uevent.event_list);
2953         wq = pd->device->create_wq(pd, &wq_init_attr, uhw);
2954         if (IS_ERR(wq)) {
2955                 err = PTR_ERR(wq);
2956                 goto err_put_cq;
2957         }
2958
2959         wq->uobject = &obj->uevent.uobject;
2960         obj->uevent.uobject.object = wq;
2961         wq->wq_type = wq_init_attr.wq_type;
2962         wq->cq = cq;
2963         wq->pd = pd;
2964         wq->device = pd->device;
2965         wq->wq_context = wq_init_attr.wq_context;
2966         atomic_set(&wq->usecnt, 0);
2967         atomic_inc(&pd->usecnt);
2968         atomic_inc(&cq->usecnt);
2969         wq->uobject = &obj->uevent.uobject;
2970         obj->uevent.uobject.object = wq;
2971
2972         memset(&resp, 0, sizeof(resp));
2973         resp.wq_handle = obj->uevent.uobject.id;
2974         resp.max_sge = wq_init_attr.max_sge;
2975         resp.max_wr = wq_init_attr.max_wr;
2976         resp.wqn = wq->wq_num;
2977         resp.response_length = required_resp_len;
2978         err = ib_copy_to_udata(ucore,
2979                                &resp, resp.response_length);
2980         if (err)
2981                 goto err_copy;
2982
2983         uobj_put_obj_read(pd);
2984         uobj_put_obj_read(cq);
2985         uobj_alloc_commit(&obj->uevent.uobject);
2986         return 0;
2987
2988 err_copy:
2989         ib_destroy_wq(wq);
2990 err_put_cq:
2991         uobj_put_obj_read(cq);
2992 err_put_pd:
2993         uobj_put_obj_read(pd);
2994 err_uobj:
2995         uobj_alloc_abort(&obj->uevent.uobject);
2996
2997         return err;
2998 }
2999
3000 int ib_uverbs_ex_destroy_wq(struct ib_uverbs_file *file,
3001                             struct ib_device *ib_dev,
3002                             struct ib_udata *ucore,
3003                             struct ib_udata *uhw)
3004 {
3005         struct ib_uverbs_ex_destroy_wq  cmd = {};
3006         struct ib_uverbs_ex_destroy_wq_resp     resp = {};
3007         struct ib_uobject               *uobj;
3008         struct ib_uwq_object            *obj;
3009         size_t required_cmd_sz;
3010         size_t required_resp_len;
3011         int                             ret;
3012
3013         required_cmd_sz = offsetof(typeof(cmd), wq_handle) + sizeof(cmd.wq_handle);
3014         required_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
3015
3016         if (ucore->inlen < required_cmd_sz)
3017                 return -EINVAL;
3018
3019         if (ucore->outlen < required_resp_len)
3020                 return -ENOSPC;
3021
3022         if (ucore->inlen > sizeof(cmd) &&
3023             !ib_is_udata_cleared(ucore, sizeof(cmd),
3024                                  ucore->inlen - sizeof(cmd)))
3025                 return -EOPNOTSUPP;
3026
3027         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3028         if (ret)
3029                 return ret;
3030
3031         if (cmd.comp_mask)
3032                 return -EOPNOTSUPP;
3033
3034         resp.response_length = required_resp_len;
3035         uobj  = uobj_get_write(uobj_get_type(wq), cmd.wq_handle,
3036                                file->ucontext);
3037         if (IS_ERR(uobj))
3038                 return PTR_ERR(uobj);
3039
3040         obj = container_of(uobj, struct ib_uwq_object, uevent.uobject);
3041         /*
3042          * Make sure we don't free the memory in remove_commit as we still
3043          * needs the uobject memory to create the response.
3044          */
3045         uverbs_uobject_get(uobj);
3046
3047         ret = uobj_remove_commit(uobj);
3048         resp.events_reported = obj->uevent.events_reported;
3049         uverbs_uobject_put(uobj);
3050         if (ret)
3051                 return ret;
3052
3053         return ib_copy_to_udata(ucore, &resp, resp.response_length);
3054 }
3055
3056 int ib_uverbs_ex_modify_wq(struct ib_uverbs_file *file,
3057                            struct ib_device *ib_dev,
3058                            struct ib_udata *ucore,
3059                            struct ib_udata *uhw)
3060 {
3061         struct ib_uverbs_ex_modify_wq cmd = {};
3062         struct ib_wq *wq;
3063         struct ib_wq_attr wq_attr = {};
3064         size_t required_cmd_sz;
3065         int ret;
3066
3067         required_cmd_sz = offsetof(typeof(cmd), curr_wq_state) + sizeof(cmd.curr_wq_state);
3068         if (ucore->inlen < required_cmd_sz)
3069                 return -EINVAL;
3070
3071         if (ucore->inlen > sizeof(cmd) &&
3072             !ib_is_udata_cleared(ucore, sizeof(cmd),
3073                                  ucore->inlen - sizeof(cmd)))
3074                 return -EOPNOTSUPP;
3075
3076         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3077         if (ret)
3078                 return ret;
3079
3080         if (!cmd.attr_mask)
3081                 return -EINVAL;
3082
3083         if (cmd.attr_mask > (IB_WQ_STATE | IB_WQ_CUR_STATE | IB_WQ_FLAGS))
3084                 return -EINVAL;
3085
3086         wq = uobj_get_obj_read(wq, cmd.wq_handle, file->ucontext);
3087         if (!wq)
3088                 return -EINVAL;
3089
3090         wq_attr.curr_wq_state = cmd.curr_wq_state;
3091         wq_attr.wq_state = cmd.wq_state;
3092         if (cmd.attr_mask & IB_WQ_FLAGS) {
3093                 wq_attr.flags = cmd.flags;
3094                 wq_attr.flags_mask = cmd.flags_mask;
3095         }
3096         ret = wq->device->modify_wq(wq, &wq_attr, cmd.attr_mask, uhw);
3097         uobj_put_obj_read(wq);
3098         return ret;
3099 }
3100
3101 int ib_uverbs_ex_create_rwq_ind_table(struct ib_uverbs_file *file,
3102                                       struct ib_device *ib_dev,
3103                                       struct ib_udata *ucore,
3104                                       struct ib_udata *uhw)
3105 {
3106         struct ib_uverbs_ex_create_rwq_ind_table          cmd = {};
3107         struct ib_uverbs_ex_create_rwq_ind_table_resp  resp = {};
3108         struct ib_uobject                 *uobj;
3109         int err = 0;
3110         struct ib_rwq_ind_table_init_attr init_attr = {};
3111         struct ib_rwq_ind_table *rwq_ind_tbl;
3112         struct ib_wq    **wqs = NULL;
3113         u32 *wqs_handles = NULL;
3114         struct ib_wq    *wq = NULL;
3115         int i, j, num_read_wqs;
3116         u32 num_wq_handles;
3117         u32 expected_in_size;
3118         size_t required_cmd_sz_header;
3119         size_t required_resp_len;
3120
3121         required_cmd_sz_header = offsetof(typeof(cmd), log_ind_tbl_size) + sizeof(cmd.log_ind_tbl_size);
3122         required_resp_len = offsetof(typeof(resp), ind_tbl_num) + sizeof(resp.ind_tbl_num);
3123
3124         if (ucore->inlen < required_cmd_sz_header)
3125                 return -EINVAL;
3126
3127         if (ucore->outlen < required_resp_len)
3128                 return -ENOSPC;
3129
3130         err = ib_copy_from_udata(&cmd, ucore, required_cmd_sz_header);
3131         if (err)
3132                 return err;
3133
3134         ucore->inbuf += required_cmd_sz_header;
3135         ucore->inlen -= required_cmd_sz_header;
3136
3137         if (cmd.comp_mask)
3138                 return -EOPNOTSUPP;
3139
3140         if (cmd.log_ind_tbl_size > IB_USER_VERBS_MAX_LOG_IND_TBL_SIZE)
3141                 return -EINVAL;
3142
3143         num_wq_handles = 1 << cmd.log_ind_tbl_size;
3144         expected_in_size = num_wq_handles * sizeof(__u32);
3145         if (num_wq_handles == 1)
3146                 /* input size for wq handles is u64 aligned */
3147                 expected_in_size += sizeof(__u32);
3148
3149         if (ucore->inlen < expected_in_size)
3150                 return -EINVAL;
3151
3152         if (ucore->inlen > expected_in_size &&
3153             !ib_is_udata_cleared(ucore, expected_in_size,
3154                                  ucore->inlen - expected_in_size))
3155                 return -EOPNOTSUPP;
3156
3157         wqs_handles = kcalloc(num_wq_handles, sizeof(*wqs_handles),
3158                               GFP_KERNEL);
3159         if (!wqs_handles)
3160                 return -ENOMEM;
3161
3162         err = ib_copy_from_udata(wqs_handles, ucore,
3163                                  num_wq_handles * sizeof(__u32));
3164         if (err)
3165                 goto err_free;
3166
3167         wqs = kcalloc(num_wq_handles, sizeof(*wqs), GFP_KERNEL);
3168         if (!wqs) {
3169                 err = -ENOMEM;
3170                 goto  err_free;
3171         }
3172
3173         for (num_read_wqs = 0; num_read_wqs < num_wq_handles;
3174                         num_read_wqs++) {
3175                 wq = uobj_get_obj_read(wq, wqs_handles[num_read_wqs],
3176                                        file->ucontext);
3177                 if (!wq) {
3178                         err = -EINVAL;
3179                         goto put_wqs;
3180                 }
3181
3182                 wqs[num_read_wqs] = wq;
3183         }
3184
3185         uobj  = uobj_alloc(uobj_get_type(rwq_ind_table), file->ucontext);
3186         if (IS_ERR(uobj)) {
3187                 err = PTR_ERR(uobj);
3188                 goto put_wqs;
3189         }
3190
3191         init_attr.log_ind_tbl_size = cmd.log_ind_tbl_size;
3192         init_attr.ind_tbl = wqs;
3193         rwq_ind_tbl = ib_dev->create_rwq_ind_table(ib_dev, &init_attr, uhw);
3194
3195         if (IS_ERR(rwq_ind_tbl)) {
3196                 err = PTR_ERR(rwq_ind_tbl);
3197                 goto err_uobj;
3198         }
3199
3200         rwq_ind_tbl->ind_tbl = wqs;
3201         rwq_ind_tbl->log_ind_tbl_size = init_attr.log_ind_tbl_size;
3202         rwq_ind_tbl->uobject = uobj;
3203         uobj->object = rwq_ind_tbl;
3204         rwq_ind_tbl->device = ib_dev;
3205         atomic_set(&rwq_ind_tbl->usecnt, 0);
3206
3207         for (i = 0; i < num_wq_handles; i++)
3208                 atomic_inc(&wqs[i]->usecnt);
3209
3210         resp.ind_tbl_handle = uobj->id;
3211         resp.ind_tbl_num = rwq_ind_tbl->ind_tbl_num;
3212         resp.response_length = required_resp_len;
3213
3214         err = ib_copy_to_udata(ucore,
3215                                &resp, resp.response_length);
3216         if (err)
3217                 goto err_copy;
3218
3219         kfree(wqs_handles);
3220
3221         for (j = 0; j < num_read_wqs; j++)
3222                 uobj_put_obj_read(wqs[j]);
3223
3224         uobj_alloc_commit(uobj);
3225         return 0;
3226
3227 err_copy:
3228         ib_destroy_rwq_ind_table(rwq_ind_tbl);
3229 err_uobj:
3230         uobj_alloc_abort(uobj);
3231 put_wqs:
3232         for (j = 0; j < num_read_wqs; j++)
3233                 uobj_put_obj_read(wqs[j]);
3234 err_free:
3235         kfree(wqs_handles);
3236         kfree(wqs);
3237         return err;
3238 }
3239
3240 int ib_uverbs_ex_destroy_rwq_ind_table(struct ib_uverbs_file *file,
3241                                        struct ib_device *ib_dev,
3242                                        struct ib_udata *ucore,
3243                                        struct ib_udata *uhw)
3244 {
3245         struct ib_uverbs_ex_destroy_rwq_ind_table       cmd = {};
3246         struct ib_uobject               *uobj;
3247         int                     ret;
3248         size_t required_cmd_sz;
3249
3250         required_cmd_sz = offsetof(typeof(cmd), ind_tbl_handle) + sizeof(cmd.ind_tbl_handle);
3251
3252         if (ucore->inlen < required_cmd_sz)
3253                 return -EINVAL;
3254
3255         if (ucore->inlen > sizeof(cmd) &&
3256             !ib_is_udata_cleared(ucore, sizeof(cmd),
3257                                  ucore->inlen - sizeof(cmd)))
3258                 return -EOPNOTSUPP;
3259
3260         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3261         if (ret)
3262                 return ret;
3263
3264         if (cmd.comp_mask)
3265                 return -EOPNOTSUPP;
3266
3267         uobj  = uobj_get_write(uobj_get_type(rwq_ind_table), cmd.ind_tbl_handle,
3268                                file->ucontext);
3269         if (IS_ERR(uobj))
3270                 return PTR_ERR(uobj);
3271
3272         return uobj_remove_commit(uobj);
3273 }
3274
3275 int ib_uverbs_ex_create_flow(struct ib_uverbs_file *file,
3276                              struct ib_device *ib_dev,
3277                              struct ib_udata *ucore,
3278                              struct ib_udata *uhw)
3279 {
3280         struct ib_uverbs_create_flow      cmd;
3281         struct ib_uverbs_create_flow_resp resp;
3282         struct ib_uobject                 *uobj;
3283         struct ib_flow                    *flow_id;
3284         struct ib_uverbs_flow_attr        *kern_flow_attr;
3285         struct ib_flow_attr               *flow_attr;
3286         struct ib_qp                      *qp;
3287         int err = 0;
3288         void *kern_spec;
3289         void *ib_spec;
3290         int i;
3291
3292         if (ucore->inlen < sizeof(cmd))
3293                 return -EINVAL;
3294
3295         if (ucore->outlen < sizeof(resp))
3296                 return -ENOSPC;
3297
3298         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3299         if (err)
3300                 return err;
3301
3302         ucore->inbuf += sizeof(cmd);
3303         ucore->inlen -= sizeof(cmd);
3304
3305         if (cmd.comp_mask)
3306                 return -EINVAL;
3307
3308         if (!capable(CAP_NET_RAW))
3309                 return -EPERM;
3310
3311         if (cmd.flow_attr.flags >= IB_FLOW_ATTR_FLAGS_RESERVED)
3312                 return -EINVAL;
3313
3314         if ((cmd.flow_attr.flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) &&
3315             ((cmd.flow_attr.type == IB_FLOW_ATTR_ALL_DEFAULT) ||
3316              (cmd.flow_attr.type == IB_FLOW_ATTR_MC_DEFAULT)))
3317                 return -EINVAL;
3318
3319         if (cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
3320                 return -EINVAL;
3321
3322         if (cmd.flow_attr.size > ucore->inlen ||
3323             cmd.flow_attr.size >
3324             (cmd.flow_attr.num_of_specs * sizeof(struct ib_uverbs_flow_spec)))
3325                 return -EINVAL;
3326
3327         if (cmd.flow_attr.reserved[0] ||
3328             cmd.flow_attr.reserved[1])
3329                 return -EINVAL;
3330
3331         if (cmd.flow_attr.num_of_specs) {
3332                 kern_flow_attr = kmalloc(sizeof(*kern_flow_attr) + cmd.flow_attr.size,
3333                                          GFP_KERNEL);
3334                 if (!kern_flow_attr)
3335                         return -ENOMEM;
3336
3337                 memcpy(kern_flow_attr, &cmd.flow_attr, sizeof(*kern_flow_attr));
3338                 err = ib_copy_from_udata(kern_flow_attr + 1, ucore,
3339                                          cmd.flow_attr.size);
3340                 if (err)
3341                         goto err_free_attr;
3342         } else {
3343                 kern_flow_attr = &cmd.flow_attr;
3344         }
3345
3346         uobj  = uobj_alloc(uobj_get_type(flow), file->ucontext);
3347         if (IS_ERR(uobj)) {
3348                 err = PTR_ERR(uobj);
3349                 goto err_free_attr;
3350         }
3351
3352         qp = uobj_get_obj_read(qp, cmd.qp_handle, file->ucontext);
3353         if (!qp) {
3354                 err = -EINVAL;
3355                 goto err_uobj;
3356         }
3357
3358         flow_attr = kzalloc(sizeof(*flow_attr) + cmd.flow_attr.num_of_specs *
3359                             sizeof(union ib_flow_spec), GFP_KERNEL);
3360         if (!flow_attr) {
3361                 err = -ENOMEM;
3362                 goto err_put;
3363         }
3364
3365         flow_attr->type = kern_flow_attr->type;
3366         flow_attr->priority = kern_flow_attr->priority;
3367         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
3368         flow_attr->port = kern_flow_attr->port;
3369         flow_attr->flags = kern_flow_attr->flags;
3370         flow_attr->size = sizeof(*flow_attr);
3371
3372         kern_spec = kern_flow_attr + 1;
3373         ib_spec = flow_attr + 1;
3374         for (i = 0; i < flow_attr->num_of_specs &&
3375              cmd.flow_attr.size > offsetof(struct ib_uverbs_flow_spec, reserved) &&
3376              cmd.flow_attr.size >=
3377              ((struct ib_uverbs_flow_spec *)kern_spec)->size; i++) {
3378                 err = kern_spec_to_ib_spec(kern_spec, ib_spec);
3379                 if (err)
3380                         goto err_free;
3381                 flow_attr->size +=
3382                         ((union ib_flow_spec *) ib_spec)->size;
3383                 cmd.flow_attr.size -= ((struct ib_uverbs_flow_spec *)kern_spec)->size;
3384                 kern_spec += ((struct ib_uverbs_flow_spec *) kern_spec)->size;
3385                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
3386         }
3387         if (cmd.flow_attr.size || (i != flow_attr->num_of_specs)) {
3388                 pr_warn("create flow failed, flow %d: %d bytes left from uverb cmd\n",
3389                         i, cmd.flow_attr.size);
3390                 err = -EINVAL;
3391                 goto err_free;
3392         }
3393         flow_id = ib_create_flow(qp, flow_attr, IB_FLOW_DOMAIN_USER);
3394         if (IS_ERR(flow_id)) {
3395                 err = PTR_ERR(flow_id);
3396                 goto err_free;
3397         }
3398         flow_id->uobject = uobj;
3399         uobj->object = flow_id;
3400
3401         memset(&resp, 0, sizeof(resp));
3402         resp.flow_handle = uobj->id;
3403
3404         err = ib_copy_to_udata(ucore,
3405                                &resp, sizeof(resp));
3406         if (err)
3407                 goto err_copy;
3408
3409         uobj_put_obj_read(qp);
3410         uobj_alloc_commit(uobj);
3411         kfree(flow_attr);
3412         if (cmd.flow_attr.num_of_specs)
3413                 kfree(kern_flow_attr);
3414         return 0;
3415 err_copy:
3416         ib_destroy_flow(flow_id);
3417 err_free:
3418         kfree(flow_attr);
3419 err_put:
3420         uobj_put_obj_read(qp);
3421 err_uobj:
3422         uobj_alloc_abort(uobj);
3423 err_free_attr:
3424         if (cmd.flow_attr.num_of_specs)
3425                 kfree(kern_flow_attr);
3426         return err;
3427 }
3428
3429 int ib_uverbs_ex_destroy_flow(struct ib_uverbs_file *file,
3430                               struct ib_device *ib_dev,
3431                               struct ib_udata *ucore,
3432                               struct ib_udata *uhw)
3433 {
3434         struct ib_uverbs_destroy_flow   cmd;
3435         struct ib_uobject               *uobj;
3436         int                             ret;
3437
3438         if (ucore->inlen < sizeof(cmd))
3439                 return -EINVAL;
3440
3441         ret = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3442         if (ret)
3443                 return ret;
3444
3445         if (cmd.comp_mask)
3446                 return -EINVAL;
3447
3448         uobj  = uobj_get_write(uobj_get_type(flow), cmd.flow_handle,
3449                                file->ucontext);
3450         if (IS_ERR(uobj))
3451                 return PTR_ERR(uobj);
3452
3453         ret = uobj_remove_commit(uobj);
3454         return ret;
3455 }
3456
3457 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
3458                                 struct ib_device *ib_dev,
3459                                 struct ib_uverbs_create_xsrq *cmd,
3460                                 struct ib_udata *udata)
3461 {
3462         struct ib_uverbs_create_srq_resp resp;
3463         struct ib_usrq_object           *obj;
3464         struct ib_pd                    *pd;
3465         struct ib_srq                   *srq;
3466         struct ib_uobject               *uninitialized_var(xrcd_uobj);
3467         struct ib_srq_init_attr          attr;
3468         int ret;
3469
3470         obj  = (struct ib_usrq_object *)uobj_alloc(uobj_get_type(srq),
3471                                                    file->ucontext);
3472         if (IS_ERR(obj))
3473                 return PTR_ERR(obj);
3474
3475         if (cmd->srq_type == IB_SRQT_XRC) {
3476                 xrcd_uobj = uobj_get_read(uobj_get_type(xrcd), cmd->xrcd_handle,
3477                                           file->ucontext);
3478                 if (IS_ERR(xrcd_uobj)) {
3479                         ret = -EINVAL;
3480                         goto err;
3481                 }
3482
3483                 attr.ext.xrc.xrcd = (struct ib_xrcd *)xrcd_uobj->object;
3484                 if (!attr.ext.xrc.xrcd) {
3485                         ret = -EINVAL;
3486                         goto err_put_xrcd;
3487                 }
3488
3489                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
3490                 atomic_inc(&obj->uxrcd->refcnt);
3491
3492                 attr.ext.xrc.cq  = uobj_get_obj_read(cq, cmd->cq_handle,
3493                                                      file->ucontext);
3494                 if (!attr.ext.xrc.cq) {
3495                         ret = -EINVAL;
3496                         goto err_put_xrcd;
3497                 }
3498         }
3499
3500         pd  = uobj_get_obj_read(pd, cmd->pd_handle, file->ucontext);
3501         if (!pd) {
3502                 ret = -EINVAL;
3503                 goto err_put_cq;
3504         }
3505
3506         attr.event_handler  = ib_uverbs_srq_event_handler;
3507         attr.srq_context    = file;
3508         attr.srq_type       = cmd->srq_type;
3509         attr.attr.max_wr    = cmd->max_wr;
3510         attr.attr.max_sge   = cmd->max_sge;
3511         attr.attr.srq_limit = cmd->srq_limit;
3512
3513         obj->uevent.events_reported = 0;
3514         INIT_LIST_HEAD(&obj->uevent.event_list);
3515
3516         srq = pd->device->create_srq(pd, &attr, udata);
3517         if (IS_ERR(srq)) {
3518                 ret = PTR_ERR(srq);
3519                 goto err_put;
3520         }
3521
3522         srq->device        = pd->device;
3523         srq->pd            = pd;
3524         srq->srq_type      = cmd->srq_type;
3525         srq->uobject       = &obj->uevent.uobject;
3526         srq->event_handler = attr.event_handler;
3527         srq->srq_context   = attr.srq_context;
3528
3529         if (cmd->srq_type == IB_SRQT_XRC) {
3530                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
3531                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
3532                 atomic_inc(&attr.ext.xrc.cq->usecnt);
3533                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
3534         }
3535
3536         atomic_inc(&pd->usecnt);
3537         atomic_set(&srq->usecnt, 0);
3538
3539         obj->uevent.uobject.object = srq;
3540         obj->uevent.uobject.user_handle = cmd->user_handle;
3541
3542         memset(&resp, 0, sizeof resp);
3543         resp.srq_handle = obj->uevent.uobject.id;
3544         resp.max_wr     = attr.attr.max_wr;
3545         resp.max_sge    = attr.attr.max_sge;
3546         if (cmd->srq_type == IB_SRQT_XRC)
3547                 resp.srqn = srq->ext.xrc.srq_num;
3548
3549         if (copy_to_user((void __user *) (unsigned long) cmd->response,
3550                          &resp, sizeof resp)) {
3551                 ret = -EFAULT;
3552                 goto err_copy;
3553         }
3554
3555         if (cmd->srq_type == IB_SRQT_XRC) {
3556                 uobj_put_read(xrcd_uobj);
3557                 uobj_put_obj_read(attr.ext.xrc.cq);
3558         }
3559         uobj_put_obj_read(pd);
3560         uobj_alloc_commit(&obj->uevent.uobject);
3561
3562         return 0;
3563
3564 err_copy:
3565         ib_destroy_srq(srq);
3566
3567 err_put:
3568         uobj_put_obj_read(pd);
3569
3570 err_put_cq:
3571         if (cmd->srq_type == IB_SRQT_XRC)
3572                 uobj_put_obj_read(attr.ext.xrc.cq);
3573
3574 err_put_xrcd:
3575         if (cmd->srq_type == IB_SRQT_XRC) {
3576                 atomic_dec(&obj->uxrcd->refcnt);
3577                 uobj_put_read(xrcd_uobj);
3578         }
3579
3580 err:
3581         uobj_alloc_abort(&obj->uevent.uobject);
3582         return ret;
3583 }
3584
3585 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
3586                              struct ib_device *ib_dev,
3587                              const char __user *buf, int in_len,
3588                              int out_len)
3589 {
3590         struct ib_uverbs_create_srq      cmd;
3591         struct ib_uverbs_create_xsrq     xcmd;
3592         struct ib_uverbs_create_srq_resp resp;
3593         struct ib_udata                  udata;
3594         int ret;
3595
3596         if (out_len < sizeof resp)
3597                 return -ENOSPC;
3598
3599         if (copy_from_user(&cmd, buf, sizeof cmd))
3600                 return -EFAULT;
3601
3602         xcmd.response    = cmd.response;
3603         xcmd.user_handle = cmd.user_handle;
3604         xcmd.srq_type    = IB_SRQT_BASIC;
3605         xcmd.pd_handle   = cmd.pd_handle;
3606         xcmd.max_wr      = cmd.max_wr;
3607         xcmd.max_sge     = cmd.max_sge;
3608         xcmd.srq_limit   = cmd.srq_limit;
3609
3610         INIT_UDATA(&udata, buf + sizeof cmd,
3611                    (unsigned long) cmd.response + sizeof resp,
3612                    in_len - sizeof cmd - sizeof(struct ib_uverbs_cmd_hdr),
3613                    out_len - sizeof resp);
3614
3615         ret = __uverbs_create_xsrq(file, ib_dev, &xcmd, &udata);
3616         if (ret)
3617                 return ret;
3618
3619         return in_len;
3620 }
3621
3622 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
3623                               struct ib_device *ib_dev,
3624                               const char __user *buf, int in_len, int out_len)
3625 {
3626         struct ib_uverbs_create_xsrq     cmd;
3627         struct ib_uverbs_create_srq_resp resp;
3628         struct ib_udata                  udata;
3629         int ret;
3630
3631         if (out_len < sizeof resp)
3632                 return -ENOSPC;
3633
3634         if (copy_from_user(&cmd, buf, sizeof cmd))
3635                 return -EFAULT;
3636
3637         INIT_UDATA(&udata, buf + sizeof cmd,
3638                    (unsigned long) cmd.response + sizeof resp,
3639                    in_len - sizeof cmd - sizeof(struct ib_uverbs_cmd_hdr),
3640                    out_len - sizeof resp);
3641
3642         ret = __uverbs_create_xsrq(file, ib_dev, &cmd, &udata);
3643         if (ret)
3644                 return ret;
3645
3646         return in_len;
3647 }
3648
3649 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
3650                              struct ib_device *ib_dev,
3651                              const char __user *buf, int in_len,
3652                              int out_len)
3653 {
3654         struct ib_uverbs_modify_srq cmd;
3655         struct ib_udata             udata;
3656         struct ib_srq              *srq;
3657         struct ib_srq_attr          attr;
3658         int                         ret;
3659
3660         if (copy_from_user(&cmd, buf, sizeof cmd))
3661                 return -EFAULT;
3662
3663         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
3664                    out_len);
3665
3666         srq = uobj_get_obj_read(srq, cmd.srq_handle, file->ucontext);
3667         if (!srq)
3668                 return -EINVAL;
3669
3670         attr.max_wr    = cmd.max_wr;
3671         attr.srq_limit = cmd.srq_limit;
3672
3673         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
3674
3675         uobj_put_obj_read(srq);
3676
3677         return ret ? ret : in_len;
3678 }
3679
3680 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
3681                             struct ib_device *ib_dev,
3682                             const char __user *buf,
3683                             int in_len, int out_len)
3684 {
3685         struct ib_uverbs_query_srq      cmd;
3686         struct ib_uverbs_query_srq_resp resp;
3687         struct ib_srq_attr              attr;
3688         struct ib_srq                   *srq;
3689         int                             ret;
3690
3691         if (out_len < sizeof resp)
3692                 return -ENOSPC;
3693
3694         if (copy_from_user(&cmd, buf, sizeof cmd))
3695                 return -EFAULT;
3696
3697         srq = uobj_get_obj_read(srq, cmd.srq_handle, file->ucontext);
3698         if (!srq)
3699                 return -EINVAL;
3700
3701         ret = ib_query_srq(srq, &attr);
3702
3703         uobj_put_obj_read(srq);
3704
3705         if (ret)
3706                 return ret;
3707
3708         memset(&resp, 0, sizeof resp);
3709
3710         resp.max_wr    = attr.max_wr;
3711         resp.max_sge   = attr.max_sge;
3712         resp.srq_limit = attr.srq_limit;
3713
3714         if (copy_to_user((void __user *) (unsigned long) cmd.response,
3715                          &resp, sizeof resp))
3716                 return -EFAULT;
3717
3718         return in_len;
3719 }
3720
3721 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
3722                               struct ib_device *ib_dev,
3723                               const char __user *buf, int in_len,
3724                               int out_len)
3725 {
3726         struct ib_uverbs_destroy_srq      cmd;
3727         struct ib_uverbs_destroy_srq_resp resp;
3728         struct ib_uobject                *uobj;
3729         struct ib_uevent_object          *obj;
3730         int                               ret = -EINVAL;
3731
3732         if (copy_from_user(&cmd, buf, sizeof cmd))
3733                 return -EFAULT;
3734
3735         uobj  = uobj_get_write(uobj_get_type(srq), cmd.srq_handle,
3736                                file->ucontext);
3737         if (IS_ERR(uobj))
3738                 return PTR_ERR(uobj);
3739
3740         obj = container_of(uobj, struct ib_uevent_object, uobject);
3741         /*
3742          * Make sure we don't free the memory in remove_commit as we still
3743          * needs the uobject memory to create the response.
3744          */
3745         uverbs_uobject_get(uobj);
3746
3747         memset(&resp, 0, sizeof(resp));
3748
3749         ret = uobj_remove_commit(uobj);
3750         if (ret) {
3751                 uverbs_uobject_put(uobj);
3752                 return ret;
3753         }
3754         resp.events_reported = obj->events_reported;
3755         uverbs_uobject_put(uobj);
3756         if (copy_to_user((void __user *)(unsigned long)cmd.response,
3757                          &resp, sizeof(resp)))
3758                 return -EFAULT;
3759
3760         return in_len;
3761 }
3762
3763 int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
3764                               struct ib_device *ib_dev,
3765                               struct ib_udata *ucore,
3766                               struct ib_udata *uhw)
3767 {
3768         struct ib_uverbs_ex_query_device_resp resp = { {0} };
3769         struct ib_uverbs_ex_query_device  cmd;
3770         struct ib_device_attr attr = {0};
3771         int err;
3772
3773         if (ucore->inlen < sizeof(cmd))
3774                 return -EINVAL;
3775
3776         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3777         if (err)
3778                 return err;
3779
3780         if (cmd.comp_mask)
3781                 return -EINVAL;
3782
3783         if (cmd.reserved)
3784                 return -EINVAL;
3785
3786         resp.response_length = offsetof(typeof(resp), odp_caps);
3787
3788         if (ucore->outlen < resp.response_length)
3789                 return -ENOSPC;
3790
3791         err = ib_dev->query_device(ib_dev, &attr, uhw);
3792         if (err)
3793                 return err;
3794
3795         copy_query_dev_fields(file, ib_dev, &resp.base, &attr);
3796
3797         if (ucore->outlen < resp.response_length + sizeof(resp.odp_caps))
3798                 goto end;
3799
3800 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
3801         resp.odp_caps.general_caps = attr.odp_caps.general_caps;
3802         resp.odp_caps.per_transport_caps.rc_odp_caps =
3803                 attr.odp_caps.per_transport_caps.rc_odp_caps;
3804         resp.odp_caps.per_transport_caps.uc_odp_caps =
3805                 attr.odp_caps.per_transport_caps.uc_odp_caps;
3806         resp.odp_caps.per_transport_caps.ud_odp_caps =
3807                 attr.odp_caps.per_transport_caps.ud_odp_caps;
3808 #endif
3809         resp.response_length += sizeof(resp.odp_caps);
3810
3811         if (ucore->outlen < resp.response_length + sizeof(resp.timestamp_mask))
3812                 goto end;
3813
3814         resp.timestamp_mask = attr.timestamp_mask;
3815         resp.response_length += sizeof(resp.timestamp_mask);
3816
3817         if (ucore->outlen < resp.response_length + sizeof(resp.hca_core_clock))
3818                 goto end;
3819
3820         resp.hca_core_clock = attr.hca_core_clock;
3821         resp.response_length += sizeof(resp.hca_core_clock);
3822
3823         if (ucore->outlen < resp.response_length + sizeof(resp.device_cap_flags_ex))
3824                 goto end;
3825
3826         resp.device_cap_flags_ex = attr.device_cap_flags;
3827         resp.response_length += sizeof(resp.device_cap_flags_ex);
3828
3829         if (ucore->outlen < resp.response_length + sizeof(resp.rss_caps))
3830                 goto end;
3831
3832         resp.rss_caps.supported_qpts = attr.rss_caps.supported_qpts;
3833         resp.rss_caps.max_rwq_indirection_tables =
3834                 attr.rss_caps.max_rwq_indirection_tables;
3835         resp.rss_caps.max_rwq_indirection_table_size =
3836                 attr.rss_caps.max_rwq_indirection_table_size;
3837
3838         resp.response_length += sizeof(resp.rss_caps);
3839
3840         if (ucore->outlen < resp.response_length + sizeof(resp.max_wq_type_rq))
3841                 goto end;
3842
3843         resp.max_wq_type_rq = attr.max_wq_type_rq;
3844         resp.response_length += sizeof(resp.max_wq_type_rq);
3845
3846         if (ucore->outlen < resp.response_length + sizeof(resp.raw_packet_caps))
3847                 goto end;
3848
3849         resp.raw_packet_caps = attr.raw_packet_caps;
3850         resp.response_length += sizeof(resp.raw_packet_caps);
3851 end:
3852         err = ib_copy_to_udata(ucore, &resp, resp.response_length);
3853         return err;
3854 }