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(UVERBS_OBJECT_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         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
  95                    u64_to_user_ptr(cmd.response) + sizeof(resp),
  96                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
  97                    out_len - sizeof(resp));
  98
  99         ret = ib_rdmacg_try_charge(&cg_obj, ib_dev, RDMACG_RESOURCE_HCA_HANDLE);
 100         if (ret)
 101                 goto err;
 102
 103         ucontext = ib_dev->alloc_ucontext(ib_dev, &udata);
 104         if (IS_ERR(ucontext)) {
 105                 ret = PTR_ERR(ucontext);
 106                 goto err_alloc;
 107         }
 108
 109         ucontext->device = ib_dev;
 110         ucontext->cg_obj = cg_obj;
 111         /* ufile is required when some objects are released */
 112         ucontext->ufile = file;
 113         uverbs_initialize_ucontext(ucontext);
 114
 115         rcu_read_lock();
 116         ucontext->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
 117         rcu_read_unlock();
 118         ucontext->closing = 0;
 119
 120 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 121         ucontext->umem_tree = RB_ROOT_CACHED;
 122         init_rwsem(&ucontext->umem_rwsem);
 123         ucontext->odp_mrs_count = 0;
 124         INIT_LIST_HEAD(&ucontext->no_private_counters);
 125
 126         if (!(ib_dev->attrs.device_cap_flags & IB_DEVICE_ON_DEMAND_PAGING))
 127                 ucontext->invalidate_range = NULL;
 128
 129 #endif
 130
 131         resp.num_comp_vectors = file->device->num_comp_vectors;
 132
 133         ret = get_unused_fd_flags(O_CLOEXEC);
 134         if (ret < 0)
 135                 goto err_free;
 136         resp.async_fd = ret;
 137
 138         filp = ib_uverbs_alloc_async_event_file(file, ib_dev);
 139         if (IS_ERR(filp)) {
 140                 ret = PTR_ERR(filp);
 141                 goto err_fd;
 142         }
 143
 144         if (copy_to_user(u64_to_user_ptr(cmd.response), &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(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
 241                 return -EFAULT;
 242
 243         return in_len;
 244 }
 245
 246 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 247                              struct ib_device *ib_dev,
 248                              const char __user *buf,
 249                              int in_len, int out_len)
 250 {
 251         struct ib_uverbs_query_port      cmd;
 252         struct ib_uverbs_query_port_resp resp;
 253         struct ib_port_attr              attr;
 254         int                              ret;
 255
 256         if (out_len < sizeof resp)
 257                 return -ENOSPC;
 258
 259         if (copy_from_user(&cmd, buf, sizeof cmd))
 260                 return -EFAULT;
 261
 262         ret = ib_query_port(ib_dev, cmd.port_num, &attr);
 263         if (ret)
 264                 return ret;
 265
 266         memset(&resp, 0, sizeof resp);
 267
 268         resp.state           = attr.state;
 269         resp.max_mtu         = attr.max_mtu;
 270         resp.active_mtu      = attr.active_mtu;
 271         resp.gid_tbl_len     = attr.gid_tbl_len;
 272         resp.port_cap_flags  = attr.port_cap_flags;
 273         resp.max_msg_sz      = attr.max_msg_sz;
 274         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 275         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 276         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 277
 278         if (rdma_cap_opa_ah(ib_dev, cmd.port_num)) {
 279                 resp.lid     = OPA_TO_IB_UCAST_LID(attr.lid);
 280                 resp.sm_lid  = OPA_TO_IB_UCAST_LID(attr.sm_lid);
 281         } else {
 282                 resp.lid     = ib_lid_cpu16(attr.lid);
 283                 resp.sm_lid  = ib_lid_cpu16(attr.sm_lid);
 284         }
 285         resp.lmc             = attr.lmc;
 286         resp.max_vl_num      = attr.max_vl_num;
 287         resp.sm_sl           = attr.sm_sl;
 288         resp.subnet_timeout  = attr.subnet_timeout;
 289         resp.init_type_reply = attr.init_type_reply;
 290         resp.active_width    = attr.active_width;
 291         resp.active_speed    = attr.active_speed;
 292         resp.phys_state      = attr.phys_state;
 293         resp.link_layer      = rdma_port_get_link_layer(ib_dev,
 294                                                         cmd.port_num);
 295
 296         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
 297                 return -EFAULT;
 298
 299         return in_len;
 300 }
 301
 302 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 303                            struct ib_device *ib_dev,
 304                            const char __user *buf,
 305                            int in_len, int out_len)
 306 {
 307         struct ib_uverbs_alloc_pd      cmd;
 308         struct ib_uverbs_alloc_pd_resp resp;
 309         struct ib_udata                udata;
 310         struct ib_uobject             *uobj;
 311         struct ib_pd                  *pd;
 312         int                            ret;
 313
 314         if (out_len < sizeof resp)
 315                 return -ENOSPC;
 316
 317         if (copy_from_user(&cmd, buf, sizeof cmd))
 318                 return -EFAULT;
 319
 320         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 321                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 322                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 323                    out_len - sizeof(resp));
 324
 325         uobj  = uobj_alloc(UVERBS_OBJECT_PD, file->ucontext);
 326         if (IS_ERR(uobj))
 327                 return PTR_ERR(uobj);
 328
 329         pd = ib_dev->alloc_pd(ib_dev, file->ucontext, &udata);
 330         if (IS_ERR(pd)) {
 331                 ret = PTR_ERR(pd);
 332                 goto err;
 333         }
 334
 335         pd->device  = ib_dev;
 336         pd->uobject = uobj;
 337         pd->__internal_mr = NULL;
 338         atomic_set(&pd->usecnt, 0);
 339
 340         uobj->object = pd;
 341         memset(&resp, 0, sizeof resp);
 342         resp.pd_handle = uobj->id;
 343         pd->res.type = RDMA_RESTRACK_PD;
 344         rdma_restrack_add(&pd->res);
 345
 346         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 347                 ret = -EFAULT;
 348                 goto err_copy;
 349         }
 350
 351         uobj_alloc_commit(uobj);
 352
 353         return in_len;
 354
 355 err_copy:
 356         ib_dealloc_pd(pd);
 357
 358 err:
 359         uobj_alloc_abort(uobj);
 360         return ret;
 361 }
 362
 363 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 364                              struct ib_device *ib_dev,
 365                              const char __user *buf,
 366                              int in_len, int out_len)
 367 {
 368         struct ib_uverbs_dealloc_pd cmd;
 369         struct ib_uobject          *uobj;
 370         int                         ret;
 371
 372         if (copy_from_user(&cmd, buf, sizeof cmd))
 373                 return -EFAULT;
 374
 375         uobj  = uobj_get_write(UVERBS_OBJECT_PD, cmd.pd_handle,
 376                                file->ucontext);
 377         if (IS_ERR(uobj))
 378                 return PTR_ERR(uobj);
 379
 380         ret = uobj_remove_commit(uobj);
 381
 382         return ret ?: in_len;
 383 }
 384
 385 struct xrcd_table_entry {
 386         struct rb_node  node;
 387         struct ib_xrcd *xrcd;
 388         struct inode   *inode;
 389 };
 390
 391 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 392                             struct inode *inode,
 393                             struct ib_xrcd *xrcd)
 394 {
 395         struct xrcd_table_entry *entry, *scan;
 396         struct rb_node **p = &dev->xrcd_tree.rb_node;
 397         struct rb_node *parent = NULL;
 398
 399         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 400         if (!entry)
 401                 return -ENOMEM;
 402
 403         entry->xrcd  = xrcd;
 404         entry->inode = inode;
 405
 406         while (*p) {
 407                 parent = *p;
 408                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 409
 410                 if (inode < scan->inode) {
 411                         p = &(*p)->rb_left;
 412                 } else if (inode > scan->inode) {
 413                         p = &(*p)->rb_right;
 414                 } else {
 415                         kfree(entry);
 416                         return -EEXIST;
 417                 }
 418         }
 419
 420         rb_link_node(&entry->node, parent, p);
 421         rb_insert_color(&entry->node, &dev->xrcd_tree);
 422         igrab(inode);
 423         return 0;
 424 }
 425
 426 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 427                                                   struct inode *inode)
 428 {
 429         struct xrcd_table_entry *entry;
 430         struct rb_node *p = dev->xrcd_tree.rb_node;
 431
 432         while (p) {
 433                 entry = rb_entry(p, struct xrcd_table_entry, node);
 434
 435                 if (inode < entry->inode)
 436                         p = p->rb_left;
 437                 else if (inode > entry->inode)
 438                         p = p->rb_right;
 439                 else
 440                         return entry;
 441         }
 442
 443         return NULL;
 444 }
 445
 446 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 447 {
 448         struct xrcd_table_entry *entry;
 449
 450         entry = xrcd_table_search(dev, inode);
 451         if (!entry)
 452                 return NULL;
 453
 454         return entry->xrcd;
 455 }
 456
 457 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 458                               struct inode *inode)
 459 {
 460         struct xrcd_table_entry *entry;
 461
 462         entry = xrcd_table_search(dev, inode);
 463         if (entry) {
 464                 iput(inode);
 465                 rb_erase(&entry->node, &dev->xrcd_tree);
 466                 kfree(entry);
 467         }
 468 }
 469
 470 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 471                             struct ib_device *ib_dev,
 472                             const char __user *buf, int in_len,
 473                             int out_len)
 474 {
 475         struct ib_uverbs_open_xrcd      cmd;
 476         struct ib_uverbs_open_xrcd_resp resp;
 477         struct ib_udata                 udata;
 478         struct ib_uxrcd_object         *obj;
 479         struct ib_xrcd                 *xrcd = NULL;
 480         struct fd                       f = {NULL, 0};
 481         struct inode                   *inode = NULL;
 482         int                             ret = 0;
 483         int                             new_xrcd = 0;
 484
 485         if (out_len < sizeof resp)
 486                 return -ENOSPC;
 487
 488         if (copy_from_user(&cmd, buf, sizeof cmd))
 489                 return -EFAULT;
 490
 491         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 492                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 493                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 494                    out_len - sizeof(resp));
 495
 496         mutex_lock(&file->device->xrcd_tree_mutex);
 497
 498         if (cmd.fd != -1) {
 499                 /* search for file descriptor */
 500                 f = fdget(cmd.fd);
 501                 if (!f.file) {
 502                         ret = -EBADF;
 503                         goto err_tree_mutex_unlock;
 504                 }
 505
 506                 inode = file_inode(f.file);
 507                 xrcd = find_xrcd(file->device, inode);
 508                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 509                         /* no file descriptor. Need CREATE flag */
 510                         ret = -EAGAIN;
 511                         goto err_tree_mutex_unlock;
 512                 }
 513
 514                 if (xrcd && cmd.oflags & O_EXCL) {
 515                         ret = -EINVAL;
 516                         goto err_tree_mutex_unlock;
 517                 }
 518         }
 519
 520         obj  = (struct ib_uxrcd_object *)uobj_alloc(UVERBS_OBJECT_XRCD,
 521                                                     file->ucontext);
 522         if (IS_ERR(obj)) {
 523                 ret = PTR_ERR(obj);
 524                 goto err_tree_mutex_unlock;
 525         }
 526
 527         if (!xrcd) {
 528                 xrcd = ib_dev->alloc_xrcd(ib_dev, file->ucontext, &udata);
 529                 if (IS_ERR(xrcd)) {
 530                         ret = PTR_ERR(xrcd);
 531                         goto err;
 532                 }
 533
 534                 xrcd->inode   = inode;
 535                 xrcd->device  = ib_dev;
 536                 atomic_set(&xrcd->usecnt, 0);
 537                 mutex_init(&xrcd->tgt_qp_mutex);
 538                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 539                 new_xrcd = 1;
 540         }
 541
 542         atomic_set(&obj->refcnt, 0);
 543         obj->uobject.object = xrcd;
 544         memset(&resp, 0, sizeof resp);
 545         resp.xrcd_handle = obj->uobject.id;
 546
 547         if (inode) {
 548                 if (new_xrcd) {
 549                         /* create new inode/xrcd table entry */
 550                         ret = xrcd_table_insert(file->device, inode, xrcd);
 551                         if (ret)
 552                                 goto err_dealloc_xrcd;
 553                 }
 554                 atomic_inc(&xrcd->usecnt);
 555         }
 556
 557         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 558                 ret = -EFAULT;
 559                 goto err_copy;
 560         }
 561
 562         if (f.file)
 563                 fdput(f);
 564
 565         mutex_unlock(&file->device->xrcd_tree_mutex);
 566
 567         uobj_alloc_commit(&obj->uobject);
 568
 569         return in_len;
 570
 571 err_copy:
 572         if (inode) {
 573                 if (new_xrcd)
 574                         xrcd_table_delete(file->device, inode);
 575                 atomic_dec(&xrcd->usecnt);
 576         }
 577
 578 err_dealloc_xrcd:
 579         ib_dealloc_xrcd(xrcd);
 580
 581 err:
 582         uobj_alloc_abort(&obj->uobject);
 583
 584 err_tree_mutex_unlock:
 585         if (f.file)
 586                 fdput(f);
 587
 588         mutex_unlock(&file->device->xrcd_tree_mutex);
 589
 590         return ret;
 591 }
 592
 593 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 594                              struct ib_device *ib_dev,
 595                              const char __user *buf, int in_len,
 596                              int out_len)
 597 {
 598         struct ib_uverbs_close_xrcd cmd;
 599         struct ib_uobject           *uobj;
 600         int                         ret = 0;
 601
 602         if (copy_from_user(&cmd, buf, sizeof cmd))
 603                 return -EFAULT;
 604
 605         uobj  = uobj_get_write(UVERBS_OBJECT_XRCD, cmd.xrcd_handle,
 606                                file->ucontext);
 607         if (IS_ERR(uobj))
 608                 return PTR_ERR(uobj);
 609
 610         ret = uobj_remove_commit(uobj);
 611         return ret ?: in_len;
 612 }
 613
 614 int ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 615                            struct ib_xrcd *xrcd,
 616                            enum rdma_remove_reason why)
 617 {
 618         struct inode *inode;
 619         int ret;
 620
 621         inode = xrcd->inode;
 622         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 623                 return 0;
 624
 625         ret = ib_dealloc_xrcd(xrcd);
 626
 627         if (why == RDMA_REMOVE_DESTROY && ret)
 628                 atomic_inc(&xrcd->usecnt);
 629         else if (inode)
 630                 xrcd_table_delete(dev, inode);
 631
 632         return ret;
 633 }
 634
 635 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 636                          struct ib_device *ib_dev,
 637                          const char __user *buf, int in_len,
 638                          int out_len)
 639 {
 640         struct ib_uverbs_reg_mr      cmd;
 641         struct ib_uverbs_reg_mr_resp resp;
 642         struct ib_udata              udata;
 643         struct ib_uobject           *uobj;
 644         struct ib_pd                *pd;
 645         struct ib_mr                *mr;
 646         int                          ret;
 647
 648         if (out_len < sizeof resp)
 649                 return -ENOSPC;
 650
 651         if (copy_from_user(&cmd, buf, sizeof cmd))
 652                 return -EFAULT;
 653
 654         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 655                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 656                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 657                    out_len - sizeof(resp));
 658
 659         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 660                 return -EINVAL;
 661
 662         ret = ib_check_mr_access(cmd.access_flags);
 663         if (ret)
 664                 return ret;
 665
 666         uobj  = uobj_alloc(UVERBS_OBJECT_MR, file->ucontext);
 667         if (IS_ERR(uobj))
 668                 return PTR_ERR(uobj);
 669
 670         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file->ucontext);
 671         if (!pd) {
 672                 ret = -EINVAL;
 673                 goto err_free;
 674         }
 675
 676         if (cmd.access_flags & IB_ACCESS_ON_DEMAND) {
 677                 if (!(pd->device->attrs.device_cap_flags &
 678                       IB_DEVICE_ON_DEMAND_PAGING)) {
 679                         pr_debug("ODP support not available\n");
 680                         ret = -EINVAL;
 681                         goto err_put;
 682                 }
 683         }
 684
 685         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 686                                      cmd.access_flags, &udata);
 687         if (IS_ERR(mr)) {
 688                 ret = PTR_ERR(mr);
 689                 goto err_put;
 690         }
 691
 692         mr->device  = pd->device;
 693         mr->pd      = pd;
 694         mr->dm      = NULL;
 695         mr->uobject = uobj;
 696         atomic_inc(&pd->usecnt);
 697         mr->res.type = RDMA_RESTRACK_MR;
 698         rdma_restrack_add(&mr->res);
 699
 700         uobj->object = mr;
 701
 702         memset(&resp, 0, sizeof resp);
 703         resp.lkey      = mr->lkey;
 704         resp.rkey      = mr->rkey;
 705         resp.mr_handle = uobj->id;
 706
 707         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 708                 ret = -EFAULT;
 709                 goto err_copy;
 710         }
 711
 712         uobj_put_obj_read(pd);
 713
 714         uobj_alloc_commit(uobj);
 715
 716         return in_len;
 717
 718 err_copy:
 719         ib_dereg_mr(mr);
 720
 721 err_put:
 722         uobj_put_obj_read(pd);
 723
 724 err_free:
 725         uobj_alloc_abort(uobj);
 726         return ret;
 727 }
 728
 729 ssize_t ib_uverbs_rereg_mr(struct ib_uverbs_file *file,
 730                            struct ib_device *ib_dev,
 731                            const char __user *buf, int in_len,
 732                            int out_len)
 733 {
 734         struct ib_uverbs_rereg_mr      cmd;
 735         struct ib_uverbs_rereg_mr_resp resp;
 736         struct ib_udata              udata;
 737         struct ib_pd                *pd = NULL;
 738         struct ib_mr                *mr;
 739         struct ib_pd                *old_pd;
 740         int                          ret;
 741         struct ib_uobject           *uobj;
 742
 743         if (out_len < sizeof(resp))
 744                 return -ENOSPC;
 745
 746         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 747                 return -EFAULT;
 748
 749         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 750                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 751                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 752                    out_len - sizeof(resp));
 753
 754         if (cmd.flags & ~IB_MR_REREG_SUPPORTED || !cmd.flags)
 755                 return -EINVAL;
 756
 757         if ((cmd.flags & IB_MR_REREG_TRANS) &&
 758             (!cmd.start || !cmd.hca_va || 0 >= cmd.length ||
 759              (cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK)))
 760                         return -EINVAL;
 761
 762         uobj  = uobj_get_write(UVERBS_OBJECT_MR, cmd.mr_handle,
 763                                file->ucontext);
 764         if (IS_ERR(uobj))
 765                 return PTR_ERR(uobj);
 766
 767         mr = uobj->object;
 768
 769         if (mr->dm) {
 770                 ret = -EINVAL;
 771                 goto put_uobjs;
 772         }
 773
 774         if (cmd.flags & IB_MR_REREG_ACCESS) {
 775                 ret = ib_check_mr_access(cmd.access_flags);
 776                 if (ret)
 777                         goto put_uobjs;
 778         }
 779
 780         if (cmd.flags & IB_MR_REREG_PD) {
 781                 pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file->ucontext);
 782                 if (!pd) {
 783                         ret = -EINVAL;
 784                         goto put_uobjs;
 785                 }
 786         }
 787
 788         old_pd = mr->pd;
 789         ret = mr->device->rereg_user_mr(mr, cmd.flags, cmd.start,
 790                                         cmd.length, cmd.hca_va,
 791                                         cmd.access_flags, pd, &udata);
 792         if (!ret) {
 793                 if (cmd.flags & IB_MR_REREG_PD) {
 794                         atomic_inc(&pd->usecnt);
 795                         mr->pd = pd;
 796                         atomic_dec(&old_pd->usecnt);
 797                 }
 798         } else {
 799                 goto put_uobj_pd;
 800         }
 801
 802         memset(&resp, 0, sizeof(resp));
 803         resp.lkey      = mr->lkey;
 804         resp.rkey      = mr->rkey;
 805
 806         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof(resp)))
 807                 ret = -EFAULT;
 808         else
 809                 ret = in_len;
 810
 811 put_uobj_pd:
 812         if (cmd.flags & IB_MR_REREG_PD)
 813                 uobj_put_obj_read(pd);
 814
 815 put_uobjs:
 816         uobj_put_write(uobj);
 817
 818         return ret;
 819 }
 820
 821 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
 822                            struct ib_device *ib_dev,
 823                            const char __user *buf, int in_len,
 824                            int out_len)
 825 {
 826         struct ib_uverbs_dereg_mr cmd;
 827         struct ib_uobject        *uobj;
 828         int                       ret = -EINVAL;
 829
 830         if (copy_from_user(&cmd, buf, sizeof cmd))
 831                 return -EFAULT;
 832
 833         uobj  = uobj_get_write(UVERBS_OBJECT_MR, cmd.mr_handle,
 834                                file->ucontext);
 835         if (IS_ERR(uobj))
 836                 return PTR_ERR(uobj);
 837
 838         ret = uobj_remove_commit(uobj);
 839
 840         return ret ?: in_len;
 841 }
 842
 843 ssize_t ib_uverbs_alloc_mw(struct ib_uverbs_file *file,
 844                            struct ib_device *ib_dev,
 845                            const char __user *buf, int in_len,
 846                            int out_len)
 847 {
 848         struct ib_uverbs_alloc_mw      cmd;
 849         struct ib_uverbs_alloc_mw_resp resp;
 850         struct ib_uobject             *uobj;
 851         struct ib_pd                  *pd;
 852         struct ib_mw                  *mw;
 853         struct ib_udata                udata;
 854         int                            ret;
 855
 856         if (out_len < sizeof(resp))
 857                 return -ENOSPC;
 858
 859         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 860                 return -EFAULT;
 861
 862         uobj  = uobj_alloc(UVERBS_OBJECT_MW, file->ucontext);
 863         if (IS_ERR(uobj))
 864                 return PTR_ERR(uobj);
 865
 866         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file->ucontext);
 867         if (!pd) {
 868                 ret = -EINVAL;
 869                 goto err_free;
 870         }
 871
 872         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 873                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 874                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 875                    out_len - sizeof(resp));
 876
 877         mw = pd->device->alloc_mw(pd, cmd.mw_type, &udata);
 878         if (IS_ERR(mw)) {
 879                 ret = PTR_ERR(mw);
 880                 goto err_put;
 881         }
 882
 883         mw->device  = pd->device;
 884         mw->pd      = pd;
 885         mw->uobject = uobj;
 886         atomic_inc(&pd->usecnt);
 887
 888         uobj->object = mw;
 889
 890         memset(&resp, 0, sizeof(resp));
 891         resp.rkey      = mw->rkey;
 892         resp.mw_handle = uobj->id;
 893
 894         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof(resp))) {
 895                 ret = -EFAULT;
 896                 goto err_copy;
 897         }
 898
 899         uobj_put_obj_read(pd);
 900         uobj_alloc_commit(uobj);
 901
 902         return in_len;
 903
 904 err_copy:
 905         uverbs_dealloc_mw(mw);
 906 err_put:
 907         uobj_put_obj_read(pd);
 908 err_free:
 909         uobj_alloc_abort(uobj);
 910         return ret;
 911 }
 912
 913 ssize_t ib_uverbs_dealloc_mw(struct ib_uverbs_file *file,
 914                              struct ib_device *ib_dev,
 915                              const char __user *buf, int in_len,
 916                              int out_len)
 917 {
 918         struct ib_uverbs_dealloc_mw cmd;
 919         struct ib_uobject          *uobj;
 920         int                         ret = -EINVAL;
 921
 922         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 923                 return -EFAULT;
 924
 925         uobj  = uobj_get_write(UVERBS_OBJECT_MW, cmd.mw_handle,
 926                                file->ucontext);
 927         if (IS_ERR(uobj))
 928                 return PTR_ERR(uobj);
 929
 930         ret = uobj_remove_commit(uobj);
 931         return ret ?: in_len;
 932 }
 933
 934 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
 935                                       struct ib_device *ib_dev,
 936                                       const char __user *buf, int in_len,
 937                                       int out_len)
 938 {
 939         struct ib_uverbs_create_comp_channel       cmd;
 940         struct ib_uverbs_create_comp_channel_resp  resp;
 941         struct ib_uobject                         *uobj;
 942         struct ib_uverbs_completion_event_file    *ev_file;
 943
 944         if (out_len < sizeof resp)
 945                 return -ENOSPC;
 946
 947         if (copy_from_user(&cmd, buf, sizeof cmd))
 948                 return -EFAULT;
 949
 950         uobj = uobj_alloc(UVERBS_OBJECT_COMP_CHANNEL, file->ucontext);
 951         if (IS_ERR(uobj))
 952                 return PTR_ERR(uobj);
 953
 954         resp.fd = uobj->id;
 955
 956         ev_file = container_of(uobj, struct ib_uverbs_completion_event_file,
 957                                uobj_file.uobj);
 958         ib_uverbs_init_event_queue(&ev_file->ev_queue);
 959
 960         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 961                 uobj_alloc_abort(uobj);
 962                 return -EFAULT;
 963         }
 964
 965         uobj_alloc_commit(uobj);
 966         return in_len;
 967 }
 968
 969 static struct ib_ucq_object *create_cq(struct ib_uverbs_file *file,
 970                                         struct ib_device *ib_dev,
 971                                        struct ib_udata *ucore,
 972                                        struct ib_udata *uhw,
 973                                        struct ib_uverbs_ex_create_cq *cmd,
 974                                        size_t cmd_sz,
 975                                        int (*cb)(struct ib_uverbs_file *file,
 976                                                  struct ib_ucq_object *obj,
 977                                                  struct ib_uverbs_ex_create_cq_resp *resp,
 978                                                  struct ib_udata *udata,
 979                                                  void *context),
 980                                        void *context)
 981 {
 982         struct ib_ucq_object           *obj;
 983         struct ib_uverbs_completion_event_file    *ev_file = NULL;
 984         struct ib_cq                   *cq;
 985         int                             ret;
 986         struct ib_uverbs_ex_create_cq_resp resp;
 987         struct ib_cq_init_attr attr = {};
 988
 989         if (!ib_dev->create_cq)
 990                 return ERR_PTR(-EOPNOTSUPP);
 991
 992         if (cmd->comp_vector >= file->device->num_comp_vectors)
 993                 return ERR_PTR(-EINVAL);
 994
 995         obj  = (struct ib_ucq_object *)uobj_alloc(UVERBS_OBJECT_CQ,
 996                                                   file->ucontext);
 997         if (IS_ERR(obj))
 998                 return obj;
 999
1000         if (cmd->comp_channel >= 0) {
1001                 ev_file = ib_uverbs_lookup_comp_file(cmd->comp_channel,
1002                                                      file->ucontext);
1003                 if (IS_ERR(ev_file)) {
1004                         ret = PTR_ERR(ev_file);
1005                         goto err;
1006                 }
1007         }
1008
1009         obj->uobject.user_handle = cmd->user_handle;
1010         obj->uverbs_file           = file;
1011         obj->comp_events_reported  = 0;
1012         obj->async_events_reported = 0;
1013         INIT_LIST_HEAD(&obj->comp_list);
1014         INIT_LIST_HEAD(&obj->async_list);
1015
1016         attr.cqe = cmd->cqe;
1017         attr.comp_vector = cmd->comp_vector;
1018
1019         if (cmd_sz > offsetof(typeof(*cmd), flags) + sizeof(cmd->flags))
1020                 attr.flags = cmd->flags;
1021
1022         cq = ib_dev->create_cq(ib_dev, &attr, file->ucontext, uhw);
1023         if (IS_ERR(cq)) {
1024                 ret = PTR_ERR(cq);
1025                 goto err_file;
1026         }
1027
1028         cq->device        = ib_dev;
1029         cq->uobject       = &obj->uobject;
1030         cq->comp_handler  = ib_uverbs_comp_handler;
1031         cq->event_handler = ib_uverbs_cq_event_handler;
1032         cq->cq_context    = ev_file ? &ev_file->ev_queue : NULL;
1033         atomic_set(&cq->usecnt, 0);
1034
1035         obj->uobject.object = cq;
1036         memset(&resp, 0, sizeof resp);
1037         resp.base.cq_handle = obj->uobject.id;
1038         resp.base.cqe       = cq->cqe;
1039
1040         resp.response_length = offsetof(typeof(resp), response_length) +
1041                 sizeof(resp.response_length);
1042
1043         cq->res.type = RDMA_RESTRACK_CQ;
1044         rdma_restrack_add(&cq->res);
1045
1046         ret = cb(file, obj, &resp, ucore, context);
1047         if (ret)
1048                 goto err_cb;
1049
1050         uobj_alloc_commit(&obj->uobject);
1051         return obj;
1052
1053 err_cb:
1054         ib_destroy_cq(cq);
1055
1056 err_file:
1057         if (ev_file)
1058                 ib_uverbs_release_ucq(file, ev_file, obj);
1059
1060 err:
1061         uobj_alloc_abort(&obj->uobject);
1062
1063         return ERR_PTR(ret);
1064 }
1065
1066 static int ib_uverbs_create_cq_cb(struct ib_uverbs_file *file,
1067                                   struct ib_ucq_object *obj,
1068                                   struct ib_uverbs_ex_create_cq_resp *resp,
1069                                   struct ib_udata *ucore, void *context)
1070 {
1071         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1072                 return -EFAULT;
1073
1074         return 0;
1075 }
1076
1077 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1078                             struct ib_device *ib_dev,
1079                             const char __user *buf, int in_len,
1080                             int out_len)
1081 {
1082         struct ib_uverbs_create_cq      cmd;
1083         struct ib_uverbs_ex_create_cq   cmd_ex;
1084         struct ib_uverbs_create_cq_resp resp;
1085         struct ib_udata                 ucore;
1086         struct ib_udata                 uhw;
1087         struct ib_ucq_object           *obj;
1088
1089         if (out_len < sizeof(resp))
1090                 return -ENOSPC;
1091
1092         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1093                 return -EFAULT;
1094
1095         ib_uverbs_init_udata(&ucore, buf, u64_to_user_ptr(cmd.response),
1096                              sizeof(cmd), sizeof(resp));
1097
1098         ib_uverbs_init_udata(&uhw, buf + sizeof(cmd),
1099                    u64_to_user_ptr(cmd.response) + sizeof(resp),
1100                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1101                    out_len - sizeof(resp));
1102
1103         memset(&cmd_ex, 0, sizeof(cmd_ex));
1104         cmd_ex.user_handle = cmd.user_handle;
1105         cmd_ex.cqe = cmd.cqe;
1106         cmd_ex.comp_vector = cmd.comp_vector;
1107         cmd_ex.comp_channel = cmd.comp_channel;
1108
1109         obj = create_cq(file, ib_dev, &ucore, &uhw, &cmd_ex,
1110                         offsetof(typeof(cmd_ex), comp_channel) +
1111                         sizeof(cmd.comp_channel), ib_uverbs_create_cq_cb,
1112                         NULL);
1113
1114         if (IS_ERR(obj))
1115                 return PTR_ERR(obj);
1116
1117         return in_len;
1118 }
1119
1120 static int ib_uverbs_ex_create_cq_cb(struct ib_uverbs_file *file,
1121                                      struct ib_ucq_object *obj,
1122                                      struct ib_uverbs_ex_create_cq_resp *resp,
1123                                      struct ib_udata *ucore, void *context)
1124 {
1125         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1126                 return -EFAULT;
1127
1128         return 0;
1129 }
1130
1131 int ib_uverbs_ex_create_cq(struct ib_uverbs_file *file,
1132                          struct ib_device *ib_dev,
1133                            struct ib_udata *ucore,
1134                            struct ib_udata *uhw)
1135 {
1136         struct ib_uverbs_ex_create_cq_resp resp;
1137         struct ib_uverbs_ex_create_cq  cmd;
1138         struct ib_ucq_object           *obj;
1139         int err;
1140
1141         if (ucore->inlen < sizeof(cmd))
1142                 return -EINVAL;
1143
1144         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
1145         if (err)
1146                 return err;
1147
1148         if (cmd.comp_mask)
1149                 return -EINVAL;
1150
1151         if (cmd.reserved)
1152                 return -EINVAL;
1153
1154         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1155                              sizeof(resp.response_length)))
1156                 return -ENOSPC;
1157
1158         obj = create_cq(file, ib_dev, ucore, uhw, &cmd,
1159                         min(ucore->inlen, sizeof(cmd)),
1160                         ib_uverbs_ex_create_cq_cb, NULL);
1161
1162         return PTR_ERR_OR_ZERO(obj);
1163 }
1164
1165 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1166                             struct ib_device *ib_dev,
1167                             const char __user *buf, int in_len,
1168                             int out_len)
1169 {
1170         struct ib_uverbs_resize_cq      cmd;
1171         struct ib_uverbs_resize_cq_resp resp = {};
1172         struct ib_udata                 udata;
1173         struct ib_cq                    *cq;
1174         int                             ret = -EINVAL;
1175
1176         if (copy_from_user(&cmd, buf, sizeof cmd))
1177                 return -EFAULT;
1178
1179         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
1180                    u64_to_user_ptr(cmd.response) + sizeof(resp),
1181                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1182                    out_len - sizeof(resp));
1183
1184         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file->ucontext);
1185         if (!cq)
1186                 return -EINVAL;
1187
1188         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1189         if (ret)
1190                 goto out;
1191
1192         resp.cqe = cq->cqe;
1193
1194         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp.cqe))
1195                 ret = -EFAULT;
1196
1197 out:
1198         uobj_put_obj_read(cq);
1199
1200         return ret ? ret : in_len;
1201 }
1202
1203 static int copy_wc_to_user(struct ib_device *ib_dev, void __user *dest,
1204                            struct ib_wc *wc)
1205 {
1206         struct ib_uverbs_wc tmp;
1207
1208         tmp.wr_id               = wc->wr_id;
1209         tmp.status              = wc->status;
1210         tmp.opcode              = wc->opcode;
1211         tmp.vendor_err          = wc->vendor_err;
1212         tmp.byte_len            = wc->byte_len;
1213         tmp.ex.imm_data         = wc->ex.imm_data;
1214         tmp.qp_num              = wc->qp->qp_num;
1215         tmp.src_qp              = wc->src_qp;
1216         tmp.wc_flags            = wc->wc_flags;
1217         tmp.pkey_index          = wc->pkey_index;
1218         if (rdma_cap_opa_ah(ib_dev, wc->port_num))
1219                 tmp.slid        = OPA_TO_IB_UCAST_LID(wc->slid);
1220         else
1221                 tmp.slid        = ib_lid_cpu16(wc->slid);
1222         tmp.sl                  = wc->sl;
1223         tmp.dlid_path_bits      = wc->dlid_path_bits;
1224         tmp.port_num            = wc->port_num;
1225         tmp.reserved            = 0;
1226
1227         if (copy_to_user(dest, &tmp, sizeof tmp))
1228                 return -EFAULT;
1229
1230         return 0;
1231 }
1232
1233 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1234                           struct ib_device *ib_dev,
1235                           const char __user *buf, int in_len,
1236                           int out_len)
1237 {
1238         struct ib_uverbs_poll_cq       cmd;
1239         struct ib_uverbs_poll_cq_resp  resp;
1240         u8 __user                     *header_ptr;
1241         u8 __user                     *data_ptr;
1242         struct ib_cq                  *cq;
1243         struct ib_wc                   wc;
1244         int                            ret;
1245
1246         if (copy_from_user(&cmd, buf, sizeof cmd))
1247                 return -EFAULT;
1248
1249         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file->ucontext);
1250         if (!cq)
1251                 return -EINVAL;
1252
1253         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1254         header_ptr = u64_to_user_ptr(cmd.response);
1255         data_ptr = header_ptr + sizeof resp;
1256
1257         memset(&resp, 0, sizeof resp);
1258         while (resp.count < cmd.ne) {
1259                 ret = ib_poll_cq(cq, 1, &wc);
1260                 if (ret < 0)
1261                         goto out_put;
1262                 if (!ret)
1263                         break;
1264
1265                 ret = copy_wc_to_user(ib_dev, data_ptr, &wc);
1266                 if (ret)
1267                         goto out_put;
1268
1269                 data_ptr += sizeof(struct ib_uverbs_wc);
1270                 ++resp.count;
1271         }
1272
1273         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1274                 ret = -EFAULT;
1275                 goto out_put;
1276         }
1277
1278         ret = in_len;
1279
1280 out_put:
1281         uobj_put_obj_read(cq);
1282         return ret;
1283 }
1284
1285 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1286                                 struct ib_device *ib_dev,
1287                                 const char __user *buf, int in_len,
1288                                 int out_len)
1289 {
1290         struct ib_uverbs_req_notify_cq cmd;
1291         struct ib_cq                  *cq;
1292
1293         if (copy_from_user(&cmd, buf, sizeof cmd))
1294                 return -EFAULT;
1295
1296         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file->ucontext);
1297         if (!cq)
1298                 return -EINVAL;
1299
1300         ib_req_notify_cq(cq, cmd.solicited_only ?
1301                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1302
1303         uobj_put_obj_read(cq);
1304
1305         return in_len;
1306 }
1307
1308 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1309                              struct ib_device *ib_dev,
1310                              const char __user *buf, int in_len,
1311                              int out_len)
1312 {
1313         struct ib_uverbs_destroy_cq      cmd;
1314         struct ib_uverbs_destroy_cq_resp resp;
1315         struct ib_uobject               *uobj;
1316         struct ib_cq                    *cq;
1317         struct ib_ucq_object            *obj;
1318         int                              ret = -EINVAL;
1319
1320         if (copy_from_user(&cmd, buf, sizeof cmd))
1321                 return -EFAULT;
1322
1323         uobj  = uobj_get_write(UVERBS_OBJECT_CQ, cmd.cq_handle,
1324                                file->ucontext);
1325         if (IS_ERR(uobj))
1326                 return PTR_ERR(uobj);
1327
1328         /*
1329          * Make sure we don't free the memory in remove_commit as we still
1330          * needs the uobject memory to create the response.
1331          */
1332         uverbs_uobject_get(uobj);
1333         cq      = uobj->object;
1334         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1335
1336         memset(&resp, 0, sizeof(resp));
1337
1338         ret = uobj_remove_commit(uobj);
1339         if (ret) {
1340                 uverbs_uobject_put(uobj);
1341                 return ret;
1342         }
1343
1344         resp.comp_events_reported  = obj->comp_events_reported;
1345         resp.async_events_reported = obj->async_events_reported;
1346
1347         uverbs_uobject_put(uobj);
1348         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
1349                 return -EFAULT;
1350
1351         return in_len;
1352 }
1353
1354 static int create_qp(struct ib_uverbs_file *file,
1355                      struct ib_udata *ucore,
1356                      struct ib_udata *uhw,
1357                      struct ib_uverbs_ex_create_qp *cmd,
1358                      size_t cmd_sz,
1359                      int (*cb)(struct ib_uverbs_file *file,
1360                                struct ib_uverbs_ex_create_qp_resp *resp,
1361                                struct ib_udata *udata),
1362                      void *context)
1363 {
1364         struct ib_uqp_object            *obj;
1365         struct ib_device                *device;
1366         struct ib_pd                    *pd = NULL;
1367         struct ib_xrcd                  *xrcd = NULL;
1368         struct ib_uobject               *xrcd_uobj = ERR_PTR(-ENOENT);
1369         struct ib_cq                    *scq = NULL, *rcq = NULL;
1370         struct ib_srq                   *srq = NULL;
1371         struct ib_qp                    *qp;
1372         char                            *buf;
1373         struct ib_qp_init_attr          attr = {};
1374         struct ib_uverbs_ex_create_qp_resp resp;
1375         int                             ret;
1376         struct ib_rwq_ind_table *ind_tbl = NULL;
1377         bool has_sq = true;
1378
1379         if (cmd->qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1380                 return -EPERM;
1381
1382         obj  = (struct ib_uqp_object *)uobj_alloc(UVERBS_OBJECT_QP,
1383                                                   file->ucontext);
1384         if (IS_ERR(obj))
1385                 return PTR_ERR(obj);
1386         obj->uxrcd = NULL;
1387         obj->uevent.uobject.user_handle = cmd->user_handle;
1388         mutex_init(&obj->mcast_lock);
1389
1390         if (cmd_sz >= offsetof(typeof(*cmd), rwq_ind_tbl_handle) +
1391                       sizeof(cmd->rwq_ind_tbl_handle) &&
1392                       (cmd->comp_mask & IB_UVERBS_CREATE_QP_MASK_IND_TABLE)) {
1393                 ind_tbl = uobj_get_obj_read(rwq_ind_table, UVERBS_OBJECT_RWQ_IND_TBL,
1394                                             cmd->rwq_ind_tbl_handle,
1395                                             file->ucontext);
1396                 if (!ind_tbl) {
1397                         ret = -EINVAL;
1398                         goto err_put;
1399                 }
1400
1401                 attr.rwq_ind_tbl = ind_tbl;
1402         }
1403
1404         if (cmd_sz > sizeof(*cmd) &&
1405             !ib_is_udata_cleared(ucore, sizeof(*cmd),
1406                                  cmd_sz - sizeof(*cmd))) {
1407                 ret = -EOPNOTSUPP;
1408                 goto err_put;
1409         }
1410
1411         if (ind_tbl && (cmd->max_recv_wr || cmd->max_recv_sge || cmd->is_srq)) {
1412                 ret = -EINVAL;
1413                 goto err_put;
1414         }
1415
1416         if (ind_tbl && !cmd->max_send_wr)
1417                 has_sq = false;
1418
1419         if (cmd->qp_type == IB_QPT_XRC_TGT) {
1420                 xrcd_uobj = uobj_get_read(UVERBS_OBJECT_XRCD, cmd->pd_handle,
1421                                           file->ucontext);
1422
1423                 if (IS_ERR(xrcd_uobj)) {
1424                         ret = -EINVAL;
1425                         goto err_put;
1426                 }
1427
1428                 xrcd = (struct ib_xrcd *)xrcd_uobj->object;
1429                 if (!xrcd) {
1430                         ret = -EINVAL;
1431                         goto err_put;
1432                 }
1433                 device = xrcd->device;
1434         } else {
1435                 if (cmd->qp_type == IB_QPT_XRC_INI) {
1436                         cmd->max_recv_wr = 0;
1437                         cmd->max_recv_sge = 0;
1438                 } else {
1439                         if (cmd->is_srq) {
1440                                 srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd->srq_handle,
1441                                                         file->ucontext);
1442                                 if (!srq || srq->srq_type == IB_SRQT_XRC) {
1443                                         ret = -EINVAL;
1444                                         goto err_put;
1445                                 }
1446                         }
1447
1448                         if (!ind_tbl) {
1449                                 if (cmd->recv_cq_handle != cmd->send_cq_handle) {
1450                                         rcq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd->recv_cq_handle,
1451                                                                 file->ucontext);
1452                                         if (!rcq) {
1453                                                 ret = -EINVAL;
1454                                                 goto err_put;
1455                                         }
1456                                 }
1457                         }
1458                 }
1459
1460                 if (has_sq)
1461                         scq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd->send_cq_handle,
1462                                                 file->ucontext);
1463                 if (!ind_tbl)
1464                         rcq = rcq ?: scq;
1465                 pd  = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd->pd_handle, file->ucontext);
1466                 if (!pd || (!scq && has_sq)) {
1467                         ret = -EINVAL;
1468                         goto err_put;
1469                 }
1470
1471                 device = pd->device;
1472         }
1473
1474         attr.event_handler = ib_uverbs_qp_event_handler;
1475         attr.qp_context    = file;
1476         attr.send_cq       = scq;
1477         attr.recv_cq       = rcq;
1478         attr.srq           = srq;
1479         attr.xrcd          = xrcd;
1480         attr.sq_sig_type   = cmd->sq_sig_all ? IB_SIGNAL_ALL_WR :
1481                                               IB_SIGNAL_REQ_WR;
1482         attr.qp_type       = cmd->qp_type;
1483         attr.create_flags  = 0;
1484
1485         attr.cap.max_send_wr     = cmd->max_send_wr;
1486         attr.cap.max_recv_wr     = cmd->max_recv_wr;
1487         attr.cap.max_send_sge    = cmd->max_send_sge;
1488         attr.cap.max_recv_sge    = cmd->max_recv_sge;
1489         attr.cap.max_inline_data = cmd->max_inline_data;
1490
1491         obj->uevent.events_reported     = 0;
1492         INIT_LIST_HEAD(&obj->uevent.event_list);
1493         INIT_LIST_HEAD(&obj->mcast_list);
1494
1495         if (cmd_sz >= offsetof(typeof(*cmd), create_flags) +
1496                       sizeof(cmd->create_flags))
1497                 attr.create_flags = cmd->create_flags;
1498
1499         if (attr.create_flags & ~(IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
1500                                 IB_QP_CREATE_CROSS_CHANNEL |
1501                                 IB_QP_CREATE_MANAGED_SEND |
1502                                 IB_QP_CREATE_MANAGED_RECV |
1503                                 IB_QP_CREATE_SCATTER_FCS |
1504                                 IB_QP_CREATE_CVLAN_STRIPPING |
1505                                 IB_QP_CREATE_SOURCE_QPN |
1506                                 IB_QP_CREATE_PCI_WRITE_END_PADDING)) {
1507                 ret = -EINVAL;
1508                 goto err_put;
1509         }
1510
1511         if (attr.create_flags & IB_QP_CREATE_SOURCE_QPN) {
1512                 if (!capable(CAP_NET_RAW)) {
1513                         ret = -EPERM;
1514                         goto err_put;
1515                 }
1516
1517                 attr.source_qpn = cmd->source_qpn;
1518         }
1519
1520         buf = (void *)cmd + sizeof(*cmd);
1521         if (cmd_sz > sizeof(*cmd))
1522                 if (!(buf[0] == 0 && !memcmp(buf, buf + 1,
1523                                              cmd_sz - sizeof(*cmd) - 1))) {
1524                         ret = -EINVAL;
1525                         goto err_put;
1526                 }
1527
1528         if (cmd->qp_type == IB_QPT_XRC_TGT)
1529                 qp = ib_create_qp(pd, &attr);
1530         else
1531                 qp = _ib_create_qp(device, pd, &attr, uhw,
1532                                    &obj->uevent.uobject);
1533
1534         if (IS_ERR(qp)) {
1535                 ret = PTR_ERR(qp);
1536                 goto err_put;
1537         }
1538
1539         if (cmd->qp_type != IB_QPT_XRC_TGT) {
1540                 ret = ib_create_qp_security(qp, device);
1541                 if (ret)
1542                         goto err_cb;
1543
1544                 qp->real_qp       = qp;
1545                 qp->pd            = pd;
1546                 qp->send_cq       = attr.send_cq;
1547                 qp->recv_cq       = attr.recv_cq;
1548                 qp->srq           = attr.srq;
1549                 qp->rwq_ind_tbl   = ind_tbl;
1550                 qp->event_handler = attr.event_handler;
1551                 qp->qp_context    = attr.qp_context;
1552                 qp->qp_type       = attr.qp_type;
1553                 atomic_set(&qp->usecnt, 0);
1554                 atomic_inc(&pd->usecnt);
1555                 qp->port = 0;
1556                 if (attr.send_cq)
1557                         atomic_inc(&attr.send_cq->usecnt);
1558                 if (attr.recv_cq)
1559                         atomic_inc(&attr.recv_cq->usecnt);
1560                 if (attr.srq)
1561                         atomic_inc(&attr.srq->usecnt);
1562                 if (ind_tbl)
1563                         atomic_inc(&ind_tbl->usecnt);
1564         } else {
1565                 /* It is done in _ib_create_qp for other QP types */
1566                 qp->uobject = &obj->uevent.uobject;
1567         }
1568
1569         obj->uevent.uobject.object = qp;
1570
1571         memset(&resp, 0, sizeof resp);
1572         resp.base.qpn             = qp->qp_num;
1573         resp.base.qp_handle       = obj->uevent.uobject.id;
1574         resp.base.max_recv_sge    = attr.cap.max_recv_sge;
1575         resp.base.max_send_sge    = attr.cap.max_send_sge;
1576         resp.base.max_recv_wr     = attr.cap.max_recv_wr;
1577         resp.base.max_send_wr     = attr.cap.max_send_wr;
1578         resp.base.max_inline_data = attr.cap.max_inline_data;
1579
1580         resp.response_length = offsetof(typeof(resp), response_length) +
1581                                sizeof(resp.response_length);
1582
1583         ret = cb(file, &resp, ucore);
1584         if (ret)
1585                 goto err_cb;
1586
1587         if (xrcd) {
1588                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object,
1589                                           uobject);
1590                 atomic_inc(&obj->uxrcd->refcnt);
1591                 uobj_put_read(xrcd_uobj);
1592         }
1593
1594         if (pd)
1595                 uobj_put_obj_read(pd);
1596         if (scq)
1597                 uobj_put_obj_read(scq);
1598         if (rcq && rcq != scq)
1599                 uobj_put_obj_read(rcq);
1600         if (srq)
1601                 uobj_put_obj_read(srq);
1602         if (ind_tbl)
1603                 uobj_put_obj_read(ind_tbl);
1604
1605         uobj_alloc_commit(&obj->uevent.uobject);
1606
1607         return 0;
1608 err_cb:
1609         ib_destroy_qp(qp);
1610
1611 err_put:
1612         if (!IS_ERR(xrcd_uobj))
1613                 uobj_put_read(xrcd_uobj);
1614         if (pd)
1615                 uobj_put_obj_read(pd);
1616         if (scq)
1617                 uobj_put_obj_read(scq);
1618         if (rcq && rcq != scq)
1619                 uobj_put_obj_read(rcq);
1620         if (srq)
1621                 uobj_put_obj_read(srq);
1622         if (ind_tbl)
1623                 uobj_put_obj_read(ind_tbl);
1624
1625         uobj_alloc_abort(&obj->uevent.uobject);
1626         return ret;
1627 }
1628
1629 static int ib_uverbs_create_qp_cb(struct ib_uverbs_file *file,
1630                                   struct ib_uverbs_ex_create_qp_resp *resp,
1631                                   struct ib_udata *ucore)
1632 {
1633         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1634                 return -EFAULT;
1635
1636         return 0;
1637 }
1638
1639 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1640                             struct ib_device *ib_dev,
1641                             const char __user *buf, int in_len,
1642                             int out_len)
1643 {
1644         struct ib_uverbs_create_qp      cmd;
1645         struct ib_uverbs_ex_create_qp   cmd_ex;
1646         struct ib_udata                 ucore;
1647         struct ib_udata                 uhw;
1648         ssize_t resp_size = sizeof(struct ib_uverbs_create_qp_resp);
1649         int                             err;
1650
1651         if (out_len < resp_size)
1652                 return -ENOSPC;
1653
1654         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1655                 return -EFAULT;
1656
1657         ib_uverbs_init_udata(&ucore, buf, u64_to_user_ptr(cmd.response),
1658                    sizeof(cmd), resp_size);
1659         ib_uverbs_init_udata(&uhw, buf + sizeof(cmd),
1660                    u64_to_user_ptr(cmd.response) + resp_size,
1661                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1662                    out_len - resp_size);
1663
1664         memset(&cmd_ex, 0, sizeof(cmd_ex));
1665         cmd_ex.user_handle = cmd.user_handle;
1666         cmd_ex.pd_handle = cmd.pd_handle;
1667         cmd_ex.send_cq_handle = cmd.send_cq_handle;
1668         cmd_ex.recv_cq_handle = cmd.recv_cq_handle;
1669         cmd_ex.srq_handle = cmd.srq_handle;
1670         cmd_ex.max_send_wr = cmd.max_send_wr;
1671         cmd_ex.max_recv_wr = cmd.max_recv_wr;
1672         cmd_ex.max_send_sge = cmd.max_send_sge;
1673         cmd_ex.max_recv_sge = cmd.max_recv_sge;
1674         cmd_ex.max_inline_data = cmd.max_inline_data;
1675         cmd_ex.sq_sig_all = cmd.sq_sig_all;
1676         cmd_ex.qp_type = cmd.qp_type;
1677         cmd_ex.is_srq = cmd.is_srq;
1678
1679         err = create_qp(file, &ucore, &uhw, &cmd_ex,
1680                         offsetof(typeof(cmd_ex), is_srq) +
1681                         sizeof(cmd.is_srq), ib_uverbs_create_qp_cb,
1682                         NULL);
1683
1684         if (err)
1685                 return err;
1686
1687         return in_len;
1688 }
1689
1690 static int ib_uverbs_ex_create_qp_cb(struct ib_uverbs_file *file,
1691                                      struct ib_uverbs_ex_create_qp_resp *resp,
1692                                      struct ib_udata *ucore)
1693 {
1694         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1695                 return -EFAULT;
1696
1697         return 0;
1698 }
1699
1700 int ib_uverbs_ex_create_qp(struct ib_uverbs_file *file,
1701                            struct ib_device *ib_dev,
1702                            struct ib_udata *ucore,
1703                            struct ib_udata *uhw)
1704 {
1705         struct ib_uverbs_ex_create_qp_resp resp;
1706         struct ib_uverbs_ex_create_qp cmd = {0};
1707         int err;
1708
1709         if (ucore->inlen < (offsetof(typeof(cmd), comp_mask) +
1710                             sizeof(cmd.comp_mask)))
1711                 return -EINVAL;
1712
1713         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
1714         if (err)
1715                 return err;
1716
1717         if (cmd.comp_mask & ~IB_UVERBS_CREATE_QP_SUP_COMP_MASK)
1718                 return -EINVAL;
1719
1720         if (cmd.reserved)
1721                 return -EINVAL;
1722
1723         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1724                              sizeof(resp.response_length)))
1725                 return -ENOSPC;
1726
1727         err = create_qp(file, ucore, uhw, &cmd,
1728                         min(ucore->inlen, sizeof(cmd)),
1729                         ib_uverbs_ex_create_qp_cb, NULL);
1730
1731         if (err)
1732                 return err;
1733
1734         return 0;
1735 }
1736
1737 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1738                           struct ib_device *ib_dev,
1739                           const char __user *buf, int in_len, int out_len)
1740 {
1741         struct ib_uverbs_open_qp        cmd;
1742         struct ib_uverbs_create_qp_resp resp;
1743         struct ib_udata                 udata;
1744         struct ib_uqp_object           *obj;
1745         struct ib_xrcd                 *xrcd;
1746         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1747         struct ib_qp                   *qp;
1748         struct ib_qp_open_attr          attr;
1749         int ret;
1750
1751         if (out_len < sizeof resp)
1752                 return -ENOSPC;
1753
1754         if (copy_from_user(&cmd, buf, sizeof cmd))
1755                 return -EFAULT;
1756
1757         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
1758                    u64_to_user_ptr(cmd.response) + sizeof(resp),
1759                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1760                    out_len - sizeof(resp));
1761
1762         obj  = (struct ib_uqp_object *)uobj_alloc(UVERBS_OBJECT_QP,
1763                                                   file->ucontext);
1764         if (IS_ERR(obj))
1765                 return PTR_ERR(obj);
1766
1767         xrcd_uobj = uobj_get_read(UVERBS_OBJECT_XRCD, cmd.pd_handle,
1768                                   file->ucontext);
1769         if (IS_ERR(xrcd_uobj)) {
1770                 ret = -EINVAL;
1771                 goto err_put;
1772         }
1773
1774         xrcd = (struct ib_xrcd *)xrcd_uobj->object;
1775         if (!xrcd) {
1776                 ret = -EINVAL;
1777                 goto err_xrcd;
1778         }
1779
1780         attr.event_handler = ib_uverbs_qp_event_handler;
1781         attr.qp_context    = file;
1782         attr.qp_num        = cmd.qpn;
1783         attr.qp_type       = cmd.qp_type;
1784
1785         obj->uevent.events_reported = 0;
1786         INIT_LIST_HEAD(&obj->uevent.event_list);
1787         INIT_LIST_HEAD(&obj->mcast_list);
1788
1789         qp = ib_open_qp(xrcd, &attr);
1790         if (IS_ERR(qp)) {
1791                 ret = PTR_ERR(qp);
1792                 goto err_xrcd;
1793         }
1794
1795         obj->uevent.uobject.object = qp;
1796         obj->uevent.uobject.user_handle = cmd.user_handle;
1797
1798         memset(&resp, 0, sizeof resp);
1799         resp.qpn       = qp->qp_num;
1800         resp.qp_handle = obj->uevent.uobject.id;
1801
1802         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
1803                 ret = -EFAULT;
1804                 goto err_destroy;
1805         }
1806
1807         obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
1808         atomic_inc(&obj->uxrcd->refcnt);
1809         qp->uobject = &obj->uevent.uobject;
1810         uobj_put_read(xrcd_uobj);
1811
1812
1813         uobj_alloc_commit(&obj->uevent.uobject);
1814
1815         return in_len;
1816
1817 err_destroy:
1818         ib_destroy_qp(qp);
1819 err_xrcd:
1820         uobj_put_read(xrcd_uobj);
1821 err_put:
1822         uobj_alloc_abort(&obj->uevent.uobject);
1823         return ret;
1824 }
1825
1826 static void copy_ah_attr_to_uverbs(struct ib_uverbs_qp_dest *uverb_attr,
1827                                    struct rdma_ah_attr *rdma_attr)
1828 {
1829         const struct ib_global_route   *grh;
1830
1831         uverb_attr->dlid              = rdma_ah_get_dlid(rdma_attr);
1832         uverb_attr->sl                = rdma_ah_get_sl(rdma_attr);
1833         uverb_attr->src_path_bits     = rdma_ah_get_path_bits(rdma_attr);
1834         uverb_attr->static_rate       = rdma_ah_get_static_rate(rdma_attr);
1835         uverb_attr->is_global         = !!(rdma_ah_get_ah_flags(rdma_attr) &
1836                                          IB_AH_GRH);
1837         if (uverb_attr->is_global) {
1838                 grh = rdma_ah_read_grh(rdma_attr);
1839                 memcpy(uverb_attr->dgid, grh->dgid.raw, 16);
1840                 uverb_attr->flow_label        = grh->flow_label;
1841                 uverb_attr->sgid_index        = grh->sgid_index;
1842                 uverb_attr->hop_limit         = grh->hop_limit;
1843                 uverb_attr->traffic_class     = grh->traffic_class;
1844         }
1845         uverb_attr->port_num          = rdma_ah_get_port_num(rdma_attr);
1846 }
1847
1848 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1849                            struct ib_device *ib_dev,
1850                            const char __user *buf, int in_len,
1851                            int out_len)
1852 {
1853         struct ib_uverbs_query_qp      cmd;
1854         struct ib_uverbs_query_qp_resp resp;
1855         struct ib_qp                   *qp;
1856         struct ib_qp_attr              *attr;
1857         struct ib_qp_init_attr         *init_attr;
1858         int                            ret;
1859
1860         if (copy_from_user(&cmd, buf, sizeof cmd))
1861                 return -EFAULT;
1862
1863         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1864         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1865         if (!attr || !init_attr) {
1866                 ret = -ENOMEM;
1867                 goto out;
1868         }
1869
1870         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file->ucontext);
1871         if (!qp) {
1872                 ret = -EINVAL;
1873                 goto out;
1874         }
1875
1876         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1877
1878         uobj_put_obj_read(qp);
1879
1880         if (ret)
1881                 goto out;
1882
1883         memset(&resp, 0, sizeof resp);
1884
1885         resp.qp_state               = attr->qp_state;
1886         resp.cur_qp_state           = attr->cur_qp_state;
1887         resp.path_mtu               = attr->path_mtu;
1888         resp.path_mig_state         = attr->path_mig_state;
1889         resp.qkey                   = attr->qkey;
1890         resp.rq_psn                 = attr->rq_psn;
1891         resp.sq_psn                 = attr->sq_psn;
1892         resp.dest_qp_num            = attr->dest_qp_num;
1893         resp.qp_access_flags        = attr->qp_access_flags;
1894         resp.pkey_index             = attr->pkey_index;
1895         resp.alt_pkey_index         = attr->alt_pkey_index;
1896         resp.sq_draining            = attr->sq_draining;
1897         resp.max_rd_atomic          = attr->max_rd_atomic;
1898         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1899         resp.min_rnr_timer          = attr->min_rnr_timer;
1900         resp.port_num               = attr->port_num;
1901         resp.timeout                = attr->timeout;
1902         resp.retry_cnt              = attr->retry_cnt;
1903         resp.rnr_retry              = attr->rnr_retry;
1904         resp.alt_port_num           = attr->alt_port_num;
1905         resp.alt_timeout            = attr->alt_timeout;
1906
1907         copy_ah_attr_to_uverbs(&resp.dest, &attr->ah_attr);
1908         copy_ah_attr_to_uverbs(&resp.alt_dest, &attr->alt_ah_attr);
1909
1910         resp.max_send_wr            = init_attr->cap.max_send_wr;
1911         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1912         resp.max_send_sge           = init_attr->cap.max_send_sge;
1913         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1914         resp.max_inline_data        = init_attr->cap.max_inline_data;
1915         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1916
1917         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
1918                 ret = -EFAULT;
1919
1920 out:
1921         kfree(attr);
1922         kfree(init_attr);
1923
1924         return ret ? ret : in_len;
1925 }
1926
1927 /* Remove ignored fields set in the attribute mask */
1928 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1929 {
1930         switch (qp_type) {
1931         case IB_QPT_XRC_INI:
1932                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1933         case IB_QPT_XRC_TGT:
1934                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1935                                 IB_QP_RNR_RETRY);
1936         default:
1937                 return mask;
1938         }
1939 }
1940
1941 static void copy_ah_attr_from_uverbs(struct ib_device *dev,
1942                                      struct rdma_ah_attr *rdma_attr,
1943                                      struct ib_uverbs_qp_dest *uverb_attr)
1944 {
1945         rdma_attr->type = rdma_ah_find_type(dev, uverb_attr->port_num);
1946         if (uverb_attr->is_global) {
1947                 rdma_ah_set_grh(rdma_attr, NULL,
1948                                 uverb_attr->flow_label,
1949                                 uverb_attr->sgid_index,
1950                                 uverb_attr->hop_limit,
1951                                 uverb_attr->traffic_class);
1952                 rdma_ah_set_dgid_raw(rdma_attr, uverb_attr->dgid);
1953         } else {
1954                 rdma_ah_set_ah_flags(rdma_attr, 0);
1955         }
1956         rdma_ah_set_dlid(rdma_attr, uverb_attr->dlid);
1957         rdma_ah_set_sl(rdma_attr, uverb_attr->sl);
1958         rdma_ah_set_path_bits(rdma_attr, uverb_attr->src_path_bits);
1959         rdma_ah_set_static_rate(rdma_attr, uverb_attr->static_rate);
1960         rdma_ah_set_port_num(rdma_attr, uverb_attr->port_num);
1961         rdma_ah_set_make_grd(rdma_attr, false);
1962 }
1963
1964 static int modify_qp(struct ib_uverbs_file *file,
1965                      struct ib_uverbs_ex_modify_qp *cmd, struct ib_udata *udata)
1966 {
1967         struct ib_qp_attr *attr;
1968         struct ib_qp *qp;
1969         int ret;
1970
1971         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1972         if (!attr)
1973                 return -ENOMEM;
1974
1975         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd->base.qp_handle, file->ucontext);
1976         if (!qp) {
1977                 ret = -EINVAL;
1978                 goto out;
1979         }
1980
1981         if ((cmd->base.attr_mask & IB_QP_PORT) &&
1982             !rdma_is_port_valid(qp->device, cmd->base.port_num)) {
1983                 ret = -EINVAL;
1984                 goto release_qp;
1985         }
1986
1987         if ((cmd->base.attr_mask & IB_QP_AV) &&
1988             !rdma_is_port_valid(qp->device, cmd->base.dest.port_num)) {
1989                 ret = -EINVAL;
1990                 goto release_qp;
1991         }
1992
1993         if ((cmd->base.attr_mask & IB_QP_ALT_PATH) &&
1994             (!rdma_is_port_valid(qp->device, cmd->base.alt_port_num) ||
1995             !rdma_is_port_valid(qp->device, cmd->base.alt_dest.port_num))) {
1996                 ret = -EINVAL;
1997                 goto release_qp;
1998         }
1999
2000         if ((cmd->base.attr_mask & IB_QP_CUR_STATE &&
2001             cmd->base.cur_qp_state > IB_QPS_ERR) ||
2002             cmd->base.qp_state > IB_QPS_ERR) {
2003                 ret = -EINVAL;
2004                 goto release_qp;
2005         }
2006
2007         attr->qp_state            = cmd->base.qp_state;
2008         attr->cur_qp_state        = cmd->base.cur_qp_state;
2009         attr->path_mtu            = cmd->base.path_mtu;
2010         attr->path_mig_state      = cmd->base.path_mig_state;
2011         attr->qkey                = cmd->base.qkey;
2012         attr->rq_psn              = cmd->base.rq_psn;
2013         attr->sq_psn              = cmd->base.sq_psn;
2014         attr->dest_qp_num         = cmd->base.dest_qp_num;
2015         attr->qp_access_flags     = cmd->base.qp_access_flags;
2016         attr->pkey_index          = cmd->base.pkey_index;
2017         attr->alt_pkey_index      = cmd->base.alt_pkey_index;
2018         attr->en_sqd_async_notify = cmd->base.en_sqd_async_notify;
2019         attr->max_rd_atomic       = cmd->base.max_rd_atomic;
2020         attr->max_dest_rd_atomic  = cmd->base.max_dest_rd_atomic;
2021         attr->min_rnr_timer       = cmd->base.min_rnr_timer;
2022         attr->port_num            = cmd->base.port_num;
2023         attr->timeout             = cmd->base.timeout;
2024         attr->retry_cnt           = cmd->base.retry_cnt;
2025         attr->rnr_retry           = cmd->base.rnr_retry;
2026         attr->alt_port_num        = cmd->base.alt_port_num;
2027         attr->alt_timeout         = cmd->base.alt_timeout;
2028         attr->rate_limit          = cmd->rate_limit;
2029
2030         if (cmd->base.attr_mask & IB_QP_AV)
2031                 copy_ah_attr_from_uverbs(qp->device, &attr->ah_attr,
2032                                          &cmd->base.dest);
2033
2034         if (cmd->base.attr_mask & IB_QP_ALT_PATH)
2035                 copy_ah_attr_from_uverbs(qp->device, &attr->alt_ah_attr,
2036                                          &cmd->base.alt_dest);
2037
2038         ret = ib_modify_qp_with_udata(qp, attr,
2039                                       modify_qp_mask(qp->qp_type,
2040                                                      cmd->base.attr_mask),
2041                                       udata);
2042
2043 release_qp:
2044         uobj_put_obj_read(qp);
2045 out:
2046         kfree(attr);
2047
2048         return ret;
2049 }
2050
2051 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
2052                             struct ib_device *ib_dev,
2053                             const char __user *buf, int in_len,
2054                             int out_len)
2055 {
2056         struct ib_uverbs_ex_modify_qp cmd = {};
2057         struct ib_udata udata;
2058         int ret;
2059
2060         if (copy_from_user(&cmd.base, buf, sizeof(cmd.base)))
2061                 return -EFAULT;
2062
2063         if (cmd.base.attr_mask &
2064             ~((IB_USER_LEGACY_LAST_QP_ATTR_MASK << 1) - 1))
2065                 return -EOPNOTSUPP;
2066
2067         ib_uverbs_init_udata(&udata, buf + sizeof(cmd.base), NULL,
2068                    in_len - sizeof(cmd.base) - sizeof(struct ib_uverbs_cmd_hdr),
2069                    out_len);
2070
2071         ret = modify_qp(file, &cmd, &udata);
2072         if (ret)
2073                 return ret;
2074
2075         return in_len;
2076 }
2077
2078 int ib_uverbs_ex_modify_qp(struct ib_uverbs_file *file,
2079                            struct ib_device *ib_dev,
2080                            struct ib_udata *ucore,
2081                            struct ib_udata *uhw)
2082 {
2083         struct ib_uverbs_ex_modify_qp cmd = {};
2084         int ret;
2085
2086         /*
2087          * Last bit is reserved for extending the attr_mask by
2088          * using another field.
2089          */
2090         BUILD_BUG_ON(IB_USER_LAST_QP_ATTR_MASK == (1 << 31));
2091
2092         if (ucore->inlen < sizeof(cmd.base))
2093                 return -EINVAL;
2094
2095         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2096         if (ret)
2097                 return ret;
2098
2099         if (cmd.base.attr_mask &
2100             ~((IB_USER_LAST_QP_ATTR_MASK << 1) - 1))
2101                 return -EOPNOTSUPP;
2102
2103         if (ucore->inlen > sizeof(cmd)) {
2104                 if (!ib_is_udata_cleared(ucore, sizeof(cmd),
2105                                          ucore->inlen - sizeof(cmd)))
2106                         return -EOPNOTSUPP;
2107         }
2108
2109         ret = modify_qp(file, &cmd, uhw);
2110
2111         return ret;
2112 }
2113
2114 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
2115                              struct ib_device *ib_dev,
2116                              const char __user *buf, int in_len,
2117                              int out_len)
2118 {
2119         struct ib_uverbs_destroy_qp      cmd;
2120         struct ib_uverbs_destroy_qp_resp resp;
2121         struct ib_uobject               *uobj;
2122         struct ib_uqp_object            *obj;
2123         int                              ret = -EINVAL;
2124
2125         if (copy_from_user(&cmd, buf, sizeof cmd))
2126                 return -EFAULT;
2127
2128         memset(&resp, 0, sizeof resp);
2129
2130         uobj  = uobj_get_write(UVERBS_OBJECT_QP, cmd.qp_handle,
2131                                file->ucontext);
2132         if (IS_ERR(uobj))
2133                 return PTR_ERR(uobj);
2134
2135         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2136         /*
2137          * Make sure we don't free the memory in remove_commit as we still
2138          * needs the uobject memory to create the response.
2139          */
2140         uverbs_uobject_get(uobj);
2141
2142         ret = uobj_remove_commit(uobj);
2143         if (ret) {
2144                 uverbs_uobject_put(uobj);
2145                 return ret;
2146         }
2147
2148         resp.events_reported = obj->uevent.events_reported;
2149         uverbs_uobject_put(uobj);
2150
2151         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2152                 return -EFAULT;
2153
2154         return in_len;
2155 }
2156
2157 static void *alloc_wr(size_t wr_size, __u32 num_sge)
2158 {
2159         if (num_sge >= (U32_MAX - ALIGN(wr_size, sizeof (struct ib_sge))) /
2160                        sizeof (struct ib_sge))
2161                 return NULL;
2162
2163         return kmalloc(ALIGN(wr_size, sizeof (struct ib_sge)) +
2164                          num_sge * sizeof (struct ib_sge), GFP_KERNEL);
2165 }
2166
2167 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2168                             struct ib_device *ib_dev,
2169                             const char __user *buf, int in_len,
2170                             int out_len)
2171 {
2172         struct ib_uverbs_post_send      cmd;
2173         struct ib_uverbs_post_send_resp resp;
2174         struct ib_uverbs_send_wr       *user_wr;
2175         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
2176         struct ib_qp                   *qp;
2177         int                             i, sg_ind;
2178         int                             is_ud;
2179         ssize_t                         ret = -EINVAL;
2180         size_t                          next_size;
2181
2182         if (copy_from_user(&cmd, buf, sizeof cmd))
2183                 return -EFAULT;
2184
2185         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2186             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2187                 return -EINVAL;
2188
2189         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2190                 return -EINVAL;
2191
2192         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2193         if (!user_wr)
2194                 return -ENOMEM;
2195
2196         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file->ucontext);
2197         if (!qp)
2198                 goto out;
2199
2200         is_ud = qp->qp_type == IB_QPT_UD;
2201         sg_ind = 0;
2202         last = NULL;
2203         for (i = 0; i < cmd.wr_count; ++i) {
2204                 if (copy_from_user(user_wr,
2205                                    buf + sizeof cmd + i * cmd.wqe_size,
2206                                    cmd.wqe_size)) {
2207                         ret = -EFAULT;
2208                         goto out_put;
2209                 }
2210
2211                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2212                         ret = -EINVAL;
2213                         goto out_put;
2214                 }
2215
2216                 if (is_ud) {
2217                         struct ib_ud_wr *ud;
2218
2219                         if (user_wr->opcode != IB_WR_SEND &&
2220                             user_wr->opcode != IB_WR_SEND_WITH_IMM) {
2221                                 ret = -EINVAL;
2222                                 goto out_put;
2223                         }
2224
2225                         next_size = sizeof(*ud);
2226                         ud = alloc_wr(next_size, user_wr->num_sge);
2227                         if (!ud) {
2228                                 ret = -ENOMEM;
2229                                 goto out_put;
2230                         }
2231
2232                         ud->ah = uobj_get_obj_read(ah, UVERBS_OBJECT_AH, user_wr->wr.ud.ah,
2233                                                    file->ucontext);
2234                         if (!ud->ah) {
2235                                 kfree(ud);
2236                                 ret = -EINVAL;
2237                                 goto out_put;
2238                         }
2239                         ud->remote_qpn = user_wr->wr.ud.remote_qpn;
2240                         ud->remote_qkey = user_wr->wr.ud.remote_qkey;
2241
2242                         next = &ud->wr;
2243                 } else if (user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
2244                            user_wr->opcode == IB_WR_RDMA_WRITE ||
2245                            user_wr->opcode == IB_WR_RDMA_READ) {
2246                         struct ib_rdma_wr *rdma;
2247
2248                         next_size = sizeof(*rdma);
2249                         rdma = alloc_wr(next_size, user_wr->num_sge);
2250                         if (!rdma) {
2251                                 ret = -ENOMEM;
2252                                 goto out_put;
2253                         }
2254
2255                         rdma->remote_addr = user_wr->wr.rdma.remote_addr;
2256                         rdma->rkey = user_wr->wr.rdma.rkey;
2257
2258                         next = &rdma->wr;
2259                 } else if (user_wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2260                            user_wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2261                         struct ib_atomic_wr *atomic;
2262
2263                         next_size = sizeof(*atomic);
2264                         atomic = alloc_wr(next_size, user_wr->num_sge);
2265                         if (!atomic) {
2266                                 ret = -ENOMEM;
2267                                 goto out_put;
2268                         }
2269
2270                         atomic->remote_addr = user_wr->wr.atomic.remote_addr;
2271                         atomic->compare_add = user_wr->wr.atomic.compare_add;
2272                         atomic->swap = user_wr->wr.atomic.swap;
2273                         atomic->rkey = user_wr->wr.atomic.rkey;
2274
2275                         next = &atomic->wr;
2276                 } else if (user_wr->opcode == IB_WR_SEND ||
2277                            user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2278                            user_wr->opcode == IB_WR_SEND_WITH_INV) {
2279                         next_size = sizeof(*next);
2280                         next = alloc_wr(next_size, user_wr->num_sge);
2281                         if (!next) {
2282                                 ret = -ENOMEM;
2283                                 goto out_put;
2284                         }
2285                 } else {
2286                         ret = -EINVAL;
2287                         goto out_put;
2288                 }
2289
2290                 if (user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2291                     user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
2292                         next->ex.imm_data =
2293                                         (__be32 __force) user_wr->ex.imm_data;
2294                 } else if (user_wr->opcode == IB_WR_SEND_WITH_INV) {
2295                         next->ex.invalidate_rkey = user_wr->ex.invalidate_rkey;
2296                 }
2297
2298                 if (!last)
2299                         wr = next;
2300                 else
2301                         last->next = next;
2302                 last = next;
2303
2304                 next->next       = NULL;
2305                 next->wr_id      = user_wr->wr_id;
2306                 next->num_sge    = user_wr->num_sge;
2307                 next->opcode     = user_wr->opcode;
2308                 next->send_flags = user_wr->send_flags;
2309
2310                 if (next->num_sge) {
2311                         next->sg_list = (void *) next +
2312                                 ALIGN(next_size, sizeof(struct ib_sge));
2313                         if (copy_from_user(next->sg_list,
2314                                            buf + sizeof cmd +
2315                                            cmd.wr_count * cmd.wqe_size +
2316                                            sg_ind * sizeof (struct ib_sge),
2317                                            next->num_sge * sizeof (struct ib_sge))) {
2318                                 ret = -EFAULT;
2319                                 goto out_put;
2320                         }
2321                         sg_ind += next->num_sge;
2322                 } else
2323                         next->sg_list = NULL;
2324         }
2325
2326         resp.bad_wr = 0;
2327         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2328         if (ret)
2329                 for (next = wr; next; next = next->next) {
2330                         ++resp.bad_wr;
2331                         if (next == bad_wr)
2332                                 break;
2333                 }
2334
2335         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2336                 ret = -EFAULT;
2337
2338 out_put:
2339         uobj_put_obj_read(qp);
2340
2341         while (wr) {
2342                 if (is_ud && ud_wr(wr)->ah)
2343                         uobj_put_obj_read(ud_wr(wr)->ah);
2344                 next = wr->next;
2345                 kfree(wr);
2346                 wr = next;
2347         }
2348
2349 out:
2350         kfree(user_wr);
2351
2352         return ret ? ret : in_len;
2353 }
2354
2355 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2356                                                     int in_len,
2357                                                     u32 wr_count,
2358                                                     u32 sge_count,
2359                                                     u32 wqe_size)
2360 {
2361         struct ib_uverbs_recv_wr *user_wr;
2362         struct ib_recv_wr        *wr = NULL, *last, *next;
2363         int                       sg_ind;
2364         int                       i;
2365         int                       ret;
2366
2367         if (in_len < wqe_size * wr_count +
2368             sge_count * sizeof (struct ib_uverbs_sge))
2369                 return ERR_PTR(-EINVAL);
2370
2371         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2372                 return ERR_PTR(-EINVAL);
2373
2374         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2375         if (!user_wr)
2376                 return ERR_PTR(-ENOMEM);
2377
2378         sg_ind = 0;
2379         last = NULL;
2380         for (i = 0; i < wr_count; ++i) {
2381                 if (copy_from_user(user_wr, buf + i * wqe_size,
2382                                    wqe_size)) {
2383                         ret = -EFAULT;
2384                         goto err;
2385                 }
2386
2387                 if (user_wr->num_sge + sg_ind > sge_count) {
2388                         ret = -EINVAL;
2389                         goto err;
2390                 }
2391
2392                 if (user_wr->num_sge >=
2393                     (U32_MAX - ALIGN(sizeof *next, sizeof (struct ib_sge))) /
2394                     sizeof (struct ib_sge)) {
2395                         ret = -EINVAL;
2396                         goto err;
2397                 }
2398
2399                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2400                                user_wr->num_sge * sizeof (struct ib_sge),
2401                                GFP_KERNEL);
2402                 if (!next) {
2403                         ret = -ENOMEM;
2404                         goto err;
2405                 }
2406
2407                 if (!last)
2408                         wr = next;
2409                 else
2410                         last->next = next;
2411                 last = next;
2412
2413                 next->next       = NULL;
2414                 next->wr_id      = user_wr->wr_id;
2415                 next->num_sge    = user_wr->num_sge;
2416
2417                 if (next->num_sge) {
2418                         next->sg_list = (void *) next +
2419                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2420                         if (copy_from_user(next->sg_list,
2421                                            buf + wr_count * wqe_size +
2422                                            sg_ind * sizeof (struct ib_sge),
2423                                            next->num_sge * sizeof (struct ib_sge))) {
2424                                 ret = -EFAULT;
2425                                 goto err;
2426                         }
2427                         sg_ind += next->num_sge;
2428                 } else
2429                         next->sg_list = NULL;
2430         }
2431
2432         kfree(user_wr);
2433         return wr;
2434
2435 err:
2436         kfree(user_wr);
2437
2438         while (wr) {
2439                 next = wr->next;
2440                 kfree(wr);
2441                 wr = next;
2442         }
2443
2444         return ERR_PTR(ret);
2445 }
2446
2447 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2448                             struct ib_device *ib_dev,
2449                             const char __user *buf, int in_len,
2450                             int out_len)
2451 {
2452         struct ib_uverbs_post_recv      cmd;
2453         struct ib_uverbs_post_recv_resp resp;
2454         struct ib_recv_wr              *wr, *next, *bad_wr;
2455         struct ib_qp                   *qp;
2456         ssize_t                         ret = -EINVAL;
2457
2458         if (copy_from_user(&cmd, buf, sizeof cmd))
2459                 return -EFAULT;
2460
2461         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2462                                        in_len - sizeof cmd, cmd.wr_count,
2463                                        cmd.sge_count, cmd.wqe_size);
2464         if (IS_ERR(wr))
2465                 return PTR_ERR(wr);
2466
2467         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file->ucontext);
2468         if (!qp)
2469                 goto out;
2470
2471         resp.bad_wr = 0;
2472         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2473
2474         uobj_put_obj_read(qp);
2475         if (ret) {
2476                 for (next = wr; next; next = next->next) {
2477                         ++resp.bad_wr;
2478                         if (next == bad_wr)
2479                                 break;
2480                 }
2481         }
2482
2483         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2484                 ret = -EFAULT;
2485
2486 out:
2487         while (wr) {
2488                 next = wr->next;
2489                 kfree(wr);
2490                 wr = next;
2491         }
2492
2493         return ret ? ret : in_len;
2494 }
2495
2496 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2497                                 struct ib_device *ib_dev,
2498                                 const char __user *buf, int in_len,
2499                                 int out_len)
2500 {
2501         struct ib_uverbs_post_srq_recv      cmd;
2502         struct ib_uverbs_post_srq_recv_resp resp;
2503         struct ib_recv_wr                  *wr, *next, *bad_wr;
2504         struct ib_srq                      *srq;
2505         ssize_t                             ret = -EINVAL;
2506
2507         if (copy_from_user(&cmd, buf, sizeof cmd))
2508                 return -EFAULT;
2509
2510         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2511                                        in_len - sizeof cmd, cmd.wr_count,
2512                                        cmd.sge_count, cmd.wqe_size);
2513         if (IS_ERR(wr))
2514                 return PTR_ERR(wr);
2515
2516         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file->ucontext);
2517         if (!srq)
2518                 goto out;
2519
2520         resp.bad_wr = 0;
2521         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2522
2523         uobj_put_obj_read(srq);
2524
2525         if (ret)
2526                 for (next = wr; next; next = next->next) {
2527                         ++resp.bad_wr;
2528                         if (next == bad_wr)
2529                                 break;
2530                 }
2531
2532         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2533                 ret = -EFAULT;
2534
2535 out:
2536         while (wr) {
2537                 next = wr->next;
2538                 kfree(wr);
2539                 wr = next;
2540         }
2541
2542         return ret ? ret : in_len;
2543 }
2544
2545 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2546                             struct ib_device *ib_dev,
2547                             const char __user *buf, int in_len,
2548                             int out_len)
2549 {
2550         struct ib_uverbs_create_ah       cmd;
2551         struct ib_uverbs_create_ah_resp  resp;
2552         struct ib_uobject               *uobj;
2553         struct ib_pd                    *pd;
2554         struct ib_ah                    *ah;
2555         struct rdma_ah_attr             attr;
2556         int ret;
2557         struct ib_udata                   udata;
2558
2559         if (out_len < sizeof resp)
2560                 return -ENOSPC;
2561
2562         if (copy_from_user(&cmd, buf, sizeof cmd))
2563                 return -EFAULT;
2564
2565         if (!rdma_is_port_valid(ib_dev, cmd.attr.port_num))
2566                 return -EINVAL;
2567
2568         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
2569                    u64_to_user_ptr(cmd.response) + sizeof(resp),
2570                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
2571                    out_len - sizeof(resp));
2572
2573         uobj  = uobj_alloc(UVERBS_OBJECT_AH, file->ucontext);
2574         if (IS_ERR(uobj))
2575                 return PTR_ERR(uobj);
2576
2577         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file->ucontext);
2578         if (!pd) {
2579                 ret = -EINVAL;
2580                 goto err;
2581         }
2582
2583         attr.type = rdma_ah_find_type(ib_dev, cmd.attr.port_num);
2584         rdma_ah_set_make_grd(&attr, false);
2585         rdma_ah_set_dlid(&attr, cmd.attr.dlid);
2586         rdma_ah_set_sl(&attr, cmd.attr.sl);
2587         rdma_ah_set_path_bits(&attr, cmd.attr.src_path_bits);
2588         rdma_ah_set_static_rate(&attr, cmd.attr.static_rate);
2589         rdma_ah_set_port_num(&attr, cmd.attr.port_num);
2590
2591         if (cmd.attr.is_global) {
2592                 rdma_ah_set_grh(&attr, NULL, cmd.attr.grh.flow_label,
2593                                 cmd.attr.grh.sgid_index,
2594                                 cmd.attr.grh.hop_limit,
2595                                 cmd.attr.grh.traffic_class);
2596                 rdma_ah_set_dgid_raw(&attr, cmd.attr.grh.dgid);
2597         } else {
2598                 rdma_ah_set_ah_flags(&attr, 0);
2599         }
2600
2601         ah = rdma_create_user_ah(pd, &attr, &udata);
2602         if (IS_ERR(ah)) {
2603                 ret = PTR_ERR(ah);
2604                 goto err_put;
2605         }
2606
2607         ah->uobject  = uobj;
2608         uobj->user_handle = cmd.user_handle;
2609         uobj->object = ah;
2610
2611         resp.ah_handle = uobj->id;
2612
2613         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
2614                 ret = -EFAULT;
2615                 goto err_copy;
2616         }
2617
2618         uobj_put_obj_read(pd);
2619         uobj_alloc_commit(uobj);
2620
2621         return in_len;
2622
2623 err_copy:
2624         rdma_destroy_ah(ah);
2625
2626 err_put:
2627         uobj_put_obj_read(pd);
2628
2629 err:
2630         uobj_alloc_abort(uobj);
2631         return ret;
2632 }
2633
2634 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2635                              struct ib_device *ib_dev,
2636                              const char __user *buf, int in_len, int out_len)
2637 {
2638         struct ib_uverbs_destroy_ah cmd;
2639         struct ib_uobject          *uobj;
2640         int                         ret;
2641
2642         if (copy_from_user(&cmd, buf, sizeof cmd))
2643                 return -EFAULT;
2644
2645         uobj  = uobj_get_write(UVERBS_OBJECT_AH, cmd.ah_handle,
2646                                file->ucontext);
2647         if (IS_ERR(uobj))
2648                 return PTR_ERR(uobj);
2649
2650         ret = uobj_remove_commit(uobj);
2651         return ret ?: in_len;
2652 }
2653
2654 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2655                                struct ib_device *ib_dev,
2656                                const char __user *buf, int in_len,
2657                                int out_len)
2658 {
2659         struct ib_uverbs_attach_mcast cmd;
2660         struct ib_qp                 *qp;
2661         struct ib_uqp_object         *obj;
2662         struct ib_uverbs_mcast_entry *mcast;
2663         int                           ret;
2664
2665         if (copy_from_user(&cmd, buf, sizeof cmd))
2666                 return -EFAULT;
2667
2668         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file->ucontext);
2669         if (!qp)
2670                 return -EINVAL;
2671
2672         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2673
2674         mutex_lock(&obj->mcast_lock);
2675         list_for_each_entry(mcast, &obj->mcast_list, list)
2676                 if (cmd.mlid == mcast->lid &&
2677                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2678                         ret = 0;
2679                         goto out_put;
2680                 }
2681
2682         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2683         if (!mcast) {
2684                 ret = -ENOMEM;
2685                 goto out_put;
2686         }
2687
2688         mcast->lid = cmd.mlid;
2689         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2690
2691         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2692         if (!ret)
2693                 list_add_tail(&mcast->list, &obj->mcast_list);
2694         else
2695                 kfree(mcast);
2696
2697 out_put:
2698         mutex_unlock(&obj->mcast_lock);
2699         uobj_put_obj_read(qp);
2700
2701         return ret ? ret : in_len;
2702 }
2703
2704 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2705                                struct ib_device *ib_dev,
2706                                const char __user *buf, int in_len,
2707                                int out_len)
2708 {
2709         struct ib_uverbs_detach_mcast cmd;
2710         struct ib_uqp_object         *obj;
2711         struct ib_qp                 *qp;
2712         struct ib_uverbs_mcast_entry *mcast;
2713         int                           ret = -EINVAL;
2714         bool                          found = false;
2715
2716         if (copy_from_user(&cmd, buf, sizeof cmd))
2717                 return -EFAULT;
2718
2719         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file->ucontext);
2720         if (!qp)
2721                 return -EINVAL;
2722
2723         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2724         mutex_lock(&obj->mcast_lock);
2725
2726         list_for_each_entry(mcast, &obj->mcast_list, list)
2727                 if (cmd.mlid == mcast->lid &&
2728                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2729                         list_del(&mcast->list);
2730                         kfree(mcast);
2731                         found = true;
2732                         break;
2733                 }
2734
2735         if (!found) {
2736                 ret = -EINVAL;
2737                 goto out_put;
2738         }
2739
2740         ret = ib_detach_mcast(qp, (union ib_gid *)cmd.gid, cmd.mlid);
2741
2742 out_put:
2743         mutex_unlock(&obj->mcast_lock);
2744         uobj_put_obj_read(qp);
2745         return ret ? ret : in_len;
2746 }
2747
2748 struct ib_uflow_resources {
2749         size_t                  max;
2750         size_t                  num;
2751         size_t                  collection_num;
2752         size_t                  counters_num;
2753         struct ib_counters      **counters;
2754         struct ib_flow_action   **collection;
2755 };
2756
2757 static struct ib_uflow_resources *flow_resources_alloc(size_t num_specs)
2758 {
2759         struct ib_uflow_resources *resources;
2760
2761         resources = kzalloc(sizeof(*resources), GFP_KERNEL);
2762
2763         if (!resources)
2764                 goto err_res;
2765
2766         resources->counters =
2767                 kcalloc(num_specs, sizeof(*resources->counters), GFP_KERNEL);
2768
2769         if (!resources->counters)
2770                 goto err_cnt;
2771
2772         resources->collection =
2773                 kcalloc(num_specs, sizeof(*resources->collection), GFP_KERNEL);
2774
2775         if (!resources->collection)
2776                 goto err_collection;
2777
2778         resources->max = num_specs;
2779
2780         return resources;
2781
2782 err_collection:
2783         kfree(resources->counters);
2784 err_cnt:
2785         kfree(resources);
2786 err_res:
2787         return NULL;
2788 }
2789
2790 void ib_uverbs_flow_resources_free(struct ib_uflow_resources *uflow_res)
2791 {
2792         unsigned int i;
2793
2794         for (i = 0; i < uflow_res->collection_num; i++)
2795                 atomic_dec(&uflow_res->collection[i]->usecnt);
2796
2797         for (i = 0; i < uflow_res->counters_num; i++)
2798                 atomic_dec(&uflow_res->counters[i]->usecnt);
2799
2800         kfree(uflow_res->collection);
2801         kfree(uflow_res->counters);
2802         kfree(uflow_res);
2803 }
2804
2805 static void flow_resources_add(struct ib_uflow_resources *uflow_res,
2806                                enum ib_flow_spec_type type,
2807                                void *ibobj)
2808 {
2809         WARN_ON(uflow_res->num >= uflow_res->max);
2810
2811         switch (type) {
2812         case IB_FLOW_SPEC_ACTION_HANDLE:
2813                 atomic_inc(&((struct ib_flow_action *)ibobj)->usecnt);
2814                 uflow_res->collection[uflow_res->collection_num++] =
2815                         (struct ib_flow_action *)ibobj;
2816                 break;
2817         case IB_FLOW_SPEC_ACTION_COUNT:
2818                 atomic_inc(&((struct ib_counters *)ibobj)->usecnt);
2819                 uflow_res->counters[uflow_res->counters_num++] =
2820                         (struct ib_counters *)ibobj;
2821                 break;
2822         default:
2823                 WARN_ON(1);
2824         }
2825
2826         uflow_res->num++;
2827 }
2828
2829 static int kern_spec_to_ib_spec_action(struct ib_ucontext *ucontext,
2830                                        struct ib_uverbs_flow_spec *kern_spec,
2831                                        union ib_flow_spec *ib_spec,
2832                                        struct ib_uflow_resources *uflow_res)
2833 {
2834         ib_spec->type = kern_spec->type;
2835         switch (ib_spec->type) {
2836         case IB_FLOW_SPEC_ACTION_TAG:
2837                 if (kern_spec->flow_tag.size !=
2838                     sizeof(struct ib_uverbs_flow_spec_action_tag))
2839                         return -EINVAL;
2840
2841                 ib_spec->flow_tag.size = sizeof(struct ib_flow_spec_action_tag);
2842                 ib_spec->flow_tag.tag_id = kern_spec->flow_tag.tag_id;
2843                 break;
2844         case IB_FLOW_SPEC_ACTION_DROP:
2845                 if (kern_spec->drop.size !=
2846                     sizeof(struct ib_uverbs_flow_spec_action_drop))
2847                         return -EINVAL;
2848
2849                 ib_spec->drop.size = sizeof(struct ib_flow_spec_action_drop);
2850                 break;
2851         case IB_FLOW_SPEC_ACTION_HANDLE:
2852                 if (kern_spec->action.size !=
2853                     sizeof(struct ib_uverbs_flow_spec_action_handle))
2854                         return -EOPNOTSUPP;
2855                 ib_spec->action.act = uobj_get_obj_read(flow_action,
2856                                                         UVERBS_OBJECT_FLOW_ACTION,
2857                                                         kern_spec->action.handle,
2858                                                         ucontext);
2859                 if (!ib_spec->action.act)
2860                         return -EINVAL;
2861                 ib_spec->action.size =
2862                         sizeof(struct ib_flow_spec_action_handle);
2863                 flow_resources_add(uflow_res,
2864                                    IB_FLOW_SPEC_ACTION_HANDLE,
2865                                    ib_spec->action.act);
2866                 uobj_put_obj_read(ib_spec->action.act);
2867                 break;
2868         case IB_FLOW_SPEC_ACTION_COUNT:
2869                 if (kern_spec->flow_count.size !=
2870                         sizeof(struct ib_uverbs_flow_spec_action_count))
2871                         return -EINVAL;
2872                 ib_spec->flow_count.counters =
2873                         uobj_get_obj_read(counters,
2874                                           UVERBS_OBJECT_COUNTERS,
2875                                           kern_spec->flow_count.handle,
2876                                           ucontext);
2877                 if (!ib_spec->flow_count.counters)
2878                         return -EINVAL;
2879                 ib_spec->flow_count.size =
2880                                 sizeof(struct ib_flow_spec_action_count);
2881                 flow_resources_add(uflow_res,
2882                                    IB_FLOW_SPEC_ACTION_COUNT,
2883                                    ib_spec->flow_count.counters);
2884                 uobj_put_obj_read(ib_spec->flow_count.counters);
2885                 break;
2886         default:
2887                 return -EINVAL;
2888         }
2889         return 0;
2890 }
2891
2892 static size_t kern_spec_filter_sz(const struct ib_uverbs_flow_spec_hdr *spec)
2893 {
2894         /* Returns user space filter size, includes padding */
2895         return (spec->size - sizeof(struct ib_uverbs_flow_spec_hdr)) / 2;
2896 }
2897
2898 static ssize_t spec_filter_size(const void *kern_spec_filter, u16 kern_filter_size,
2899                                 u16 ib_real_filter_sz)
2900 {
2901         /*
2902          * User space filter structures must be 64 bit aligned, otherwise this
2903          * may pass, but we won't handle additional new attributes.
2904          */
2905
2906         if (kern_filter_size > ib_real_filter_sz) {
2907                 if (memchr_inv(kern_spec_filter +
2908                                ib_real_filter_sz, 0,
2909                                kern_filter_size - ib_real_filter_sz))
2910                         return -EINVAL;
2911                 return ib_real_filter_sz;
2912         }
2913         return kern_filter_size;
2914 }
2915
2916 int ib_uverbs_kern_spec_to_ib_spec_filter(enum ib_flow_spec_type type,
2917                                           const void *kern_spec_mask,
2918                                           const void *kern_spec_val,
2919                                           size_t kern_filter_sz,
2920                                           union ib_flow_spec *ib_spec)
2921 {
2922         ssize_t actual_filter_sz;
2923         ssize_t ib_filter_sz;
2924
2925         /* User flow spec size must be aligned to 4 bytes */
2926         if (kern_filter_sz != ALIGN(kern_filter_sz, 4))
2927                 return -EINVAL;
2928
2929         ib_spec->type = type;
2930
2931         if (ib_spec->type == (IB_FLOW_SPEC_INNER | IB_FLOW_SPEC_VXLAN_TUNNEL))
2932                 return -EINVAL;
2933
2934         switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2935         case IB_FLOW_SPEC_ETH:
2936                 ib_filter_sz = offsetof(struct ib_flow_eth_filter, real_sz);
2937                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2938                                                     kern_filter_sz,
2939                                                     ib_filter_sz);
2940                 if (actual_filter_sz <= 0)
2941                         return -EINVAL;
2942                 ib_spec->size = sizeof(struct ib_flow_spec_eth);
2943                 memcpy(&ib_spec->eth.val, kern_spec_val, actual_filter_sz);
2944                 memcpy(&ib_spec->eth.mask, kern_spec_mask, actual_filter_sz);
2945                 break;
2946         case IB_FLOW_SPEC_IPV4:
2947                 ib_filter_sz = offsetof(struct ib_flow_ipv4_filter, real_sz);
2948                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2949                                                     kern_filter_sz,
2950                                                     ib_filter_sz);
2951                 if (actual_filter_sz <= 0)
2952                         return -EINVAL;
2953                 ib_spec->size = sizeof(struct ib_flow_spec_ipv4);
2954                 memcpy(&ib_spec->ipv4.val, kern_spec_val, actual_filter_sz);
2955                 memcpy(&ib_spec->ipv4.mask, kern_spec_mask, actual_filter_sz);
2956                 break;
2957         case IB_FLOW_SPEC_IPV6:
2958                 ib_filter_sz = offsetof(struct ib_flow_ipv6_filter, real_sz);
2959                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2960                                                     kern_filter_sz,
2961                                                     ib_filter_sz);
2962                 if (actual_filter_sz <= 0)
2963                         return -EINVAL;
2964                 ib_spec->size = sizeof(struct ib_flow_spec_ipv6);
2965                 memcpy(&ib_spec->ipv6.val, kern_spec_val, actual_filter_sz);
2966                 memcpy(&ib_spec->ipv6.mask, kern_spec_mask, actual_filter_sz);
2967
2968                 if ((ntohl(ib_spec->ipv6.mask.flow_label)) >= BIT(20) ||
2969                     (ntohl(ib_spec->ipv6.val.flow_label)) >= BIT(20))
2970                         return -EINVAL;
2971                 break;
2972         case IB_FLOW_SPEC_TCP:
2973         case IB_FLOW_SPEC_UDP:
2974                 ib_filter_sz = offsetof(struct ib_flow_tcp_udp_filter, real_sz);
2975                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2976                                                     kern_filter_sz,
2977                                                     ib_filter_sz);
2978                 if (actual_filter_sz <= 0)
2979                         return -EINVAL;
2980                 ib_spec->size = sizeof(struct ib_flow_spec_tcp_udp);
2981                 memcpy(&ib_spec->tcp_udp.val, kern_spec_val, actual_filter_sz);
2982                 memcpy(&ib_spec->tcp_udp.mask, kern_spec_mask, actual_filter_sz);
2983                 break;
2984         case IB_FLOW_SPEC_VXLAN_TUNNEL:
2985                 ib_filter_sz = offsetof(struct ib_flow_tunnel_filter, real_sz);
2986                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2987                                                     kern_filter_sz,
2988                                                     ib_filter_sz);
2989                 if (actual_filter_sz <= 0)
2990                         return -EINVAL;
2991                 ib_spec->tunnel.size = sizeof(struct ib_flow_spec_tunnel);
2992                 memcpy(&ib_spec->tunnel.val, kern_spec_val, actual_filter_sz);
2993                 memcpy(&ib_spec->tunnel.mask, kern_spec_mask, actual_filter_sz);
2994
2995                 if ((ntohl(ib_spec->tunnel.mask.tunnel_id)) >= BIT(24) ||
2996                     (ntohl(ib_spec->tunnel.val.tunnel_id)) >= BIT(24))
2997                         return -EINVAL;
2998                 break;
2999         case IB_FLOW_SPEC_ESP:
3000                 ib_filter_sz = offsetof(struct ib_flow_esp_filter, real_sz);
3001                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3002                                                     kern_filter_sz,
3003                                                     ib_filter_sz);
3004                 if (actual_filter_sz <= 0)
3005                         return -EINVAL;
3006                 ib_spec->esp.size = sizeof(struct ib_flow_spec_esp);
3007                 memcpy(&ib_spec->esp.val, kern_spec_val, actual_filter_sz);
3008                 memcpy(&ib_spec->esp.mask, kern_spec_mask, actual_filter_sz);
3009                 break;
3010         case IB_FLOW_SPEC_GRE:
3011                 ib_filter_sz = offsetof(struct ib_flow_gre_filter, real_sz);
3012                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3013                                                     kern_filter_sz,
3014                                                     ib_filter_sz);
3015                 if (actual_filter_sz <= 0)
3016                         return -EINVAL;
3017                 ib_spec->gre.size = sizeof(struct ib_flow_spec_gre);
3018                 memcpy(&ib_spec->gre.val, kern_spec_val, actual_filter_sz);
3019                 memcpy(&ib_spec->gre.mask, kern_spec_mask, actual_filter_sz);
3020                 break;
3021         case IB_FLOW_SPEC_MPLS:
3022                 ib_filter_sz = offsetof(struct ib_flow_mpls_filter, real_sz);
3023                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3024                                                     kern_filter_sz,
3025                                                     ib_filter_sz);
3026                 if (actual_filter_sz <= 0)
3027                         return -EINVAL;
3028                 ib_spec->mpls.size = sizeof(struct ib_flow_spec_mpls);
3029                 memcpy(&ib_spec->mpls.val, kern_spec_val, actual_filter_sz);
3030                 memcpy(&ib_spec->mpls.mask, kern_spec_mask, actual_filter_sz);
3031                 break;
3032         default:
3033                 return -EINVAL;
3034         }
3035         return 0;
3036 }
3037
3038 static int kern_spec_to_ib_spec_filter(struct ib_uverbs_flow_spec *kern_spec,
3039                                        union ib_flow_spec *ib_spec)
3040 {
3041         ssize_t kern_filter_sz;
3042         void *kern_spec_mask;
3043         void *kern_spec_val;
3044
3045         if (kern_spec->reserved)
3046                 return -EINVAL;
3047
3048         kern_filter_sz = kern_spec_filter_sz(&kern_spec->hdr);
3049
3050         kern_spec_val = (void *)kern_spec +
3051                 sizeof(struct ib_uverbs_flow_spec_hdr);
3052         kern_spec_mask = kern_spec_val + kern_filter_sz;
3053
3054         return ib_uverbs_kern_spec_to_ib_spec_filter(kern_spec->type,
3055                                                      kern_spec_mask,
3056                                                      kern_spec_val,
3057                                                      kern_filter_sz, ib_spec);
3058 }
3059
3060 static int kern_spec_to_ib_spec(struct ib_ucontext *ucontext,
3061                                 struct ib_uverbs_flow_spec *kern_spec,
3062                                 union ib_flow_spec *ib_spec,
3063                                 struct ib_uflow_resources *uflow_res)
3064 {
3065         if (kern_spec->reserved)
3066                 return -EINVAL;
3067
3068         if (kern_spec->type >= IB_FLOW_SPEC_ACTION_TAG)
3069                 return kern_spec_to_ib_spec_action(ucontext, kern_spec, ib_spec,
3070                                                    uflow_res);
3071         else
3072                 return kern_spec_to_ib_spec_filter(kern_spec, ib_spec);
3073 }
3074
3075 int ib_uverbs_ex_create_wq(struct ib_uverbs_file *file,
3076                            struct ib_device *ib_dev,
3077                            struct ib_udata *ucore,
3078                            struct ib_udata *uhw)
3079 {
3080         struct ib_uverbs_ex_create_wq     cmd = {};
3081         struct ib_uverbs_ex_create_wq_resp resp = {};
3082         struct ib_uwq_object           *obj;
3083         int err = 0;
3084         struct ib_cq *cq;
3085         struct ib_pd *pd;
3086         struct ib_wq *wq;
3087         struct ib_wq_init_attr wq_init_attr = {};
3088         size_t required_cmd_sz;
3089         size_t required_resp_len;
3090
3091         required_cmd_sz = offsetof(typeof(cmd), max_sge) + sizeof(cmd.max_sge);
3092         required_resp_len = offsetof(typeof(resp), wqn) + sizeof(resp.wqn);
3093
3094         if (ucore->inlen < required_cmd_sz)
3095                 return -EINVAL;
3096
3097         if (ucore->outlen < required_resp_len)
3098                 return -ENOSPC;
3099
3100         if (ucore->inlen > sizeof(cmd) &&
3101             !ib_is_udata_cleared(ucore, sizeof(cmd),
3102                                  ucore->inlen - sizeof(cmd)))
3103                 return -EOPNOTSUPP;
3104
3105         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3106         if (err)
3107                 return err;
3108
3109         if (cmd.comp_mask)
3110                 return -EOPNOTSUPP;
3111
3112         obj  = (struct ib_uwq_object *)uobj_alloc(UVERBS_OBJECT_WQ,
3113                                                   file->ucontext);
3114         if (IS_ERR(obj))
3115                 return PTR_ERR(obj);
3116
3117         pd  = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file->ucontext);
3118         if (!pd) {
3119                 err = -EINVAL;
3120                 goto err_uobj;
3121         }
3122
3123         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file->ucontext);
3124         if (!cq) {
3125                 err = -EINVAL;
3126                 goto err_put_pd;
3127         }
3128
3129         wq_init_attr.cq = cq;
3130         wq_init_attr.max_sge = cmd.max_sge;
3131         wq_init_attr.max_wr = cmd.max_wr;
3132         wq_init_attr.wq_context = file;
3133         wq_init_attr.wq_type = cmd.wq_type;
3134         wq_init_attr.event_handler = ib_uverbs_wq_event_handler;
3135         if (ucore->inlen >= (offsetof(typeof(cmd), create_flags) +
3136                              sizeof(cmd.create_flags)))
3137                 wq_init_attr.create_flags = cmd.create_flags;
3138         obj->uevent.events_reported = 0;
3139         INIT_LIST_HEAD(&obj->uevent.event_list);
3140
3141         if (!pd->device->create_wq) {
3142                 err = -EOPNOTSUPP;
3143                 goto err_put_cq;
3144         }
3145         wq = pd->device->create_wq(pd, &wq_init_attr, uhw);
3146         if (IS_ERR(wq)) {
3147                 err = PTR_ERR(wq);
3148                 goto err_put_cq;
3149         }
3150
3151         wq->uobject = &obj->uevent.uobject;
3152         obj->uevent.uobject.object = wq;
3153         wq->wq_type = wq_init_attr.wq_type;
3154         wq->cq = cq;
3155         wq->pd = pd;
3156         wq->device = pd->device;
3157         wq->wq_context = wq_init_attr.wq_context;
3158         atomic_set(&wq->usecnt, 0);
3159         atomic_inc(&pd->usecnt);
3160         atomic_inc(&cq->usecnt);
3161         wq->uobject = &obj->uevent.uobject;
3162         obj->uevent.uobject.object = wq;
3163
3164         memset(&resp, 0, sizeof(resp));
3165         resp.wq_handle = obj->uevent.uobject.id;
3166         resp.max_sge = wq_init_attr.max_sge;
3167         resp.max_wr = wq_init_attr.max_wr;
3168         resp.wqn = wq->wq_num;
3169         resp.response_length = required_resp_len;
3170         err = ib_copy_to_udata(ucore,
3171                                &resp, resp.response_length);
3172         if (err)
3173                 goto err_copy;
3174
3175         uobj_put_obj_read(pd);
3176         uobj_put_obj_read(cq);
3177         uobj_alloc_commit(&obj->uevent.uobject);
3178         return 0;
3179
3180 err_copy:
3181         ib_destroy_wq(wq);
3182 err_put_cq:
3183         uobj_put_obj_read(cq);
3184 err_put_pd:
3185         uobj_put_obj_read(pd);
3186 err_uobj:
3187         uobj_alloc_abort(&obj->uevent.uobject);
3188
3189         return err;
3190 }
3191
3192 int ib_uverbs_ex_destroy_wq(struct ib_uverbs_file *file,
3193                             struct ib_device *ib_dev,
3194                             struct ib_udata *ucore,
3195                             struct ib_udata *uhw)
3196 {
3197         struct ib_uverbs_ex_destroy_wq  cmd = {};
3198         struct ib_uverbs_ex_destroy_wq_resp     resp = {};
3199         struct ib_uobject               *uobj;
3200         struct ib_uwq_object            *obj;
3201         size_t required_cmd_sz;
3202         size_t required_resp_len;
3203         int                             ret;
3204
3205         required_cmd_sz = offsetof(typeof(cmd), wq_handle) + sizeof(cmd.wq_handle);
3206         required_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
3207
3208         if (ucore->inlen < required_cmd_sz)
3209                 return -EINVAL;
3210
3211         if (ucore->outlen < required_resp_len)
3212                 return -ENOSPC;
3213
3214         if (ucore->inlen > sizeof(cmd) &&
3215             !ib_is_udata_cleared(ucore, sizeof(cmd),
3216                                  ucore->inlen - sizeof(cmd)))
3217                 return -EOPNOTSUPP;
3218
3219         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3220         if (ret)
3221                 return ret;
3222
3223         if (cmd.comp_mask)
3224                 return -EOPNOTSUPP;
3225
3226         resp.response_length = required_resp_len;
3227         uobj  = uobj_get_write(UVERBS_OBJECT_WQ, cmd.wq_handle,
3228                                file->ucontext);
3229         if (IS_ERR(uobj))
3230                 return PTR_ERR(uobj);
3231
3232         obj = container_of(uobj, struct ib_uwq_object, uevent.uobject);
3233         /*
3234          * Make sure we don't free the memory in remove_commit as we still
3235          * needs the uobject memory to create the response.
3236          */
3237         uverbs_uobject_get(uobj);
3238
3239         ret = uobj_remove_commit(uobj);
3240         resp.events_reported = obj->uevent.events_reported;
3241         uverbs_uobject_put(uobj);
3242         if (ret)
3243                 return ret;
3244
3245         return ib_copy_to_udata(ucore, &resp, resp.response_length);
3246 }
3247
3248 int ib_uverbs_ex_modify_wq(struct ib_uverbs_file *file,
3249                            struct ib_device *ib_dev,
3250                            struct ib_udata *ucore,
3251                            struct ib_udata *uhw)
3252 {
3253         struct ib_uverbs_ex_modify_wq cmd = {};
3254         struct ib_wq *wq;
3255         struct ib_wq_attr wq_attr = {};
3256         size_t required_cmd_sz;
3257         int ret;
3258
3259         required_cmd_sz = offsetof(typeof(cmd), curr_wq_state) + sizeof(cmd.curr_wq_state);
3260         if (ucore->inlen < required_cmd_sz)
3261                 return -EINVAL;
3262
3263         if (ucore->inlen > sizeof(cmd) &&
3264             !ib_is_udata_cleared(ucore, sizeof(cmd),
3265                                  ucore->inlen - sizeof(cmd)))
3266                 return -EOPNOTSUPP;
3267
3268         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3269         if (ret)
3270                 return ret;
3271
3272         if (!cmd.attr_mask)
3273                 return -EINVAL;
3274
3275         if (cmd.attr_mask > (IB_WQ_STATE | IB_WQ_CUR_STATE | IB_WQ_FLAGS))
3276                 return -EINVAL;
3277
3278         wq = uobj_get_obj_read(wq, UVERBS_OBJECT_WQ, cmd.wq_handle, file->ucontext);
3279         if (!wq)
3280                 return -EINVAL;
3281
3282         wq_attr.curr_wq_state = cmd.curr_wq_state;
3283         wq_attr.wq_state = cmd.wq_state;
3284         if (cmd.attr_mask & IB_WQ_FLAGS) {
3285                 wq_attr.flags = cmd.flags;
3286                 wq_attr.flags_mask = cmd.flags_mask;
3287         }
3288         if (!wq->device->modify_wq) {
3289                 ret = -EOPNOTSUPP;
3290                 goto out;
3291         }
3292         ret = wq->device->modify_wq(wq, &wq_attr, cmd.attr_mask, uhw);
3293 out:
3294         uobj_put_obj_read(wq);
3295         return ret;
3296 }
3297
3298 int ib_uverbs_ex_create_rwq_ind_table(struct ib_uverbs_file *file,
3299                                       struct ib_device *ib_dev,
3300                                       struct ib_udata *ucore,
3301                                       struct ib_udata *uhw)
3302 {
3303         struct ib_uverbs_ex_create_rwq_ind_table          cmd = {};
3304         struct ib_uverbs_ex_create_rwq_ind_table_resp  resp = {};
3305         struct ib_uobject                 *uobj;
3306         int err = 0;
3307         struct ib_rwq_ind_table_init_attr init_attr = {};
3308         struct ib_rwq_ind_table *rwq_ind_tbl;
3309         struct ib_wq    **wqs = NULL;
3310         u32 *wqs_handles = NULL;
3311         struct ib_wq    *wq = NULL;
3312         int i, j, num_read_wqs;
3313         u32 num_wq_handles;
3314         u32 expected_in_size;
3315         size_t required_cmd_sz_header;
3316         size_t required_resp_len;
3317
3318         required_cmd_sz_header = offsetof(typeof(cmd), log_ind_tbl_size) + sizeof(cmd.log_ind_tbl_size);
3319         required_resp_len = offsetof(typeof(resp), ind_tbl_num) + sizeof(resp.ind_tbl_num);
3320
3321         if (ucore->inlen < required_cmd_sz_header)
3322                 return -EINVAL;
3323
3324         if (ucore->outlen < required_resp_len)
3325                 return -ENOSPC;
3326
3327         err = ib_copy_from_udata(&cmd, ucore, required_cmd_sz_header);
3328         if (err)
3329                 return err;
3330
3331         ucore->inbuf += required_cmd_sz_header;
3332         ucore->inlen -= required_cmd_sz_header;
3333
3334         if (cmd.comp_mask)
3335                 return -EOPNOTSUPP;
3336
3337         if (cmd.log_ind_tbl_size > IB_USER_VERBS_MAX_LOG_IND_TBL_SIZE)
3338                 return -EINVAL;
3339
3340         num_wq_handles = 1 << cmd.log_ind_tbl_size;
3341         expected_in_size = num_wq_handles * sizeof(__u32);
3342         if (num_wq_handles == 1)
3343                 /* input size for wq handles is u64 aligned */
3344                 expected_in_size += sizeof(__u32);
3345
3346         if (ucore->inlen < expected_in_size)
3347                 return -EINVAL;
3348
3349         if (ucore->inlen > expected_in_size &&
3350             !ib_is_udata_cleared(ucore, expected_in_size,
3351                                  ucore->inlen - expected_in_size))
3352                 return -EOPNOTSUPP;
3353
3354         wqs_handles = kcalloc(num_wq_handles, sizeof(*wqs_handles),
3355                               GFP_KERNEL);
3356         if (!wqs_handles)
3357                 return -ENOMEM;
3358
3359         err = ib_copy_from_udata(wqs_handles, ucore,
3360                                  num_wq_handles * sizeof(__u32));
3361         if (err)
3362                 goto err_free;
3363
3364         wqs = kcalloc(num_wq_handles, sizeof(*wqs), GFP_KERNEL);
3365         if (!wqs) {
3366                 err = -ENOMEM;
3367                 goto  err_free;
3368         }
3369
3370         for (num_read_wqs = 0; num_read_wqs < num_wq_handles;
3371                         num_read_wqs++) {
3372                 wq = uobj_get_obj_read(wq, UVERBS_OBJECT_WQ, wqs_handles[num_read_wqs],
3373                                        file->ucontext);
3374                 if (!wq) {
3375                         err = -EINVAL;
3376                         goto put_wqs;
3377                 }
3378
3379                 wqs[num_read_wqs] = wq;
3380         }
3381
3382         uobj  = uobj_alloc(UVERBS_OBJECT_RWQ_IND_TBL, file->ucontext);
3383         if (IS_ERR(uobj)) {
3384                 err = PTR_ERR(uobj);
3385                 goto put_wqs;
3386         }
3387
3388         init_attr.log_ind_tbl_size = cmd.log_ind_tbl_size;
3389         init_attr.ind_tbl = wqs;
3390
3391         if (!ib_dev->create_rwq_ind_table) {
3392                 err = -EOPNOTSUPP;
3393                 goto err_uobj;
3394         }
3395         rwq_ind_tbl = ib_dev->create_rwq_ind_table(ib_dev, &init_attr, uhw);
3396
3397         if (IS_ERR(rwq_ind_tbl)) {
3398                 err = PTR_ERR(rwq_ind_tbl);
3399                 goto err_uobj;
3400         }
3401
3402         rwq_ind_tbl->ind_tbl = wqs;
3403         rwq_ind_tbl->log_ind_tbl_size = init_attr.log_ind_tbl_size;
3404         rwq_ind_tbl->uobject = uobj;
3405         uobj->object = rwq_ind_tbl;
3406         rwq_ind_tbl->device = ib_dev;
3407         atomic_set(&rwq_ind_tbl->usecnt, 0);
3408
3409         for (i = 0; i < num_wq_handles; i++)
3410                 atomic_inc(&wqs[i]->usecnt);
3411
3412         resp.ind_tbl_handle = uobj->id;
3413         resp.ind_tbl_num = rwq_ind_tbl->ind_tbl_num;
3414         resp.response_length = required_resp_len;
3415
3416         err = ib_copy_to_udata(ucore,
3417                                &resp, resp.response_length);
3418         if (err)
3419                 goto err_copy;
3420
3421         kfree(wqs_handles);
3422
3423         for (j = 0; j < num_read_wqs; j++)
3424                 uobj_put_obj_read(wqs[j]);
3425
3426         uobj_alloc_commit(uobj);
3427         return 0;
3428
3429 err_copy:
3430         ib_destroy_rwq_ind_table(rwq_ind_tbl);
3431 err_uobj:
3432         uobj_alloc_abort(uobj);
3433 put_wqs:
3434         for (j = 0; j < num_read_wqs; j++)
3435                 uobj_put_obj_read(wqs[j]);
3436 err_free:
3437         kfree(wqs_handles);
3438         kfree(wqs);
3439         return err;
3440 }
3441
3442 int ib_uverbs_ex_destroy_rwq_ind_table(struct ib_uverbs_file *file,
3443                                        struct ib_device *ib_dev,
3444                                        struct ib_udata *ucore,
3445                                        struct ib_udata *uhw)
3446 {
3447         struct ib_uverbs_ex_destroy_rwq_ind_table       cmd = {};
3448         struct ib_uobject               *uobj;
3449         int                     ret;
3450         size_t required_cmd_sz;
3451
3452         required_cmd_sz = offsetof(typeof(cmd), ind_tbl_handle) + sizeof(cmd.ind_tbl_handle);
3453
3454         if (ucore->inlen < required_cmd_sz)
3455                 return -EINVAL;
3456
3457         if (ucore->inlen > sizeof(cmd) &&
3458             !ib_is_udata_cleared(ucore, sizeof(cmd),
3459                                  ucore->inlen - sizeof(cmd)))
3460                 return -EOPNOTSUPP;
3461
3462         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3463         if (ret)
3464                 return ret;
3465
3466         if (cmd.comp_mask)
3467                 return -EOPNOTSUPP;
3468
3469         uobj  = uobj_get_write(UVERBS_OBJECT_RWQ_IND_TBL, cmd.ind_tbl_handle,
3470                                file->ucontext);
3471         if (IS_ERR(uobj))
3472                 return PTR_ERR(uobj);
3473
3474         return uobj_remove_commit(uobj);
3475 }
3476
3477 int ib_uverbs_ex_create_flow(struct ib_uverbs_file *file,
3478                              struct ib_device *ib_dev,
3479                              struct ib_udata *ucore,
3480                              struct ib_udata *uhw)
3481 {
3482         struct ib_uverbs_create_flow      cmd;
3483         struct ib_uverbs_create_flow_resp resp;
3484         struct ib_uobject                 *uobj;
3485         struct ib_uflow_object            *uflow;
3486         struct ib_flow                    *flow_id;
3487         struct ib_uverbs_flow_attr        *kern_flow_attr;
3488         struct ib_flow_attr               *flow_attr;
3489         struct ib_qp                      *qp;
3490         struct ib_uflow_resources         *uflow_res;
3491         struct ib_uverbs_flow_spec_hdr    *kern_spec;
3492         int err = 0;
3493         void *ib_spec;
3494         int i;
3495
3496         if (ucore->inlen < sizeof(cmd))
3497                 return -EINVAL;
3498
3499         if (ucore->outlen < sizeof(resp))
3500                 return -ENOSPC;
3501
3502         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3503         if (err)
3504                 return err;
3505
3506         ucore->inbuf += sizeof(cmd);
3507         ucore->inlen -= sizeof(cmd);
3508
3509         if (cmd.comp_mask)
3510                 return -EINVAL;
3511
3512         if (!capable(CAP_NET_RAW))
3513                 return -EPERM;
3514
3515         if (cmd.flow_attr.flags >= IB_FLOW_ATTR_FLAGS_RESERVED)
3516                 return -EINVAL;
3517
3518         if ((cmd.flow_attr.flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) &&
3519             ((cmd.flow_attr.type == IB_FLOW_ATTR_ALL_DEFAULT) ||
3520              (cmd.flow_attr.type == IB_FLOW_ATTR_MC_DEFAULT)))
3521                 return -EINVAL;
3522
3523         if (cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
3524                 return -EINVAL;
3525
3526         if (cmd.flow_attr.size > ucore->inlen ||
3527             cmd.flow_attr.size >
3528             (cmd.flow_attr.num_of_specs * sizeof(struct ib_uverbs_flow_spec)))
3529                 return -EINVAL;
3530
3531         if (cmd.flow_attr.reserved[0] ||
3532             cmd.flow_attr.reserved[1])
3533                 return -EINVAL;
3534
3535         if (cmd.flow_attr.num_of_specs) {
3536                 kern_flow_attr = kmalloc(sizeof(*kern_flow_attr) + cmd.flow_attr.size,
3537                                          GFP_KERNEL);
3538                 if (!kern_flow_attr)
3539                         return -ENOMEM;
3540
3541                 *kern_flow_attr = cmd.flow_attr;
3542                 err = ib_copy_from_udata(&kern_flow_attr->flow_specs, ucore,
3543                                          cmd.flow_attr.size);
3544                 if (err)
3545                         goto err_free_attr;
3546         } else {
3547                 kern_flow_attr = &cmd.flow_attr;
3548         }
3549
3550         uobj  = uobj_alloc(UVERBS_OBJECT_FLOW, file->ucontext);
3551         if (IS_ERR(uobj)) {
3552                 err = PTR_ERR(uobj);
3553                 goto err_free_attr;
3554         }
3555
3556         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file->ucontext);
3557         if (!qp) {
3558                 err = -EINVAL;
3559                 goto err_uobj;
3560         }
3561
3562         if (qp->qp_type != IB_QPT_UD && qp->qp_type != IB_QPT_RAW_PACKET) {
3563                 err = -EINVAL;
3564                 goto err_put;
3565         }
3566
3567         flow_attr = kzalloc(struct_size(flow_attr, flows,
3568                                 cmd.flow_attr.num_of_specs), GFP_KERNEL);
3569         if (!flow_attr) {
3570                 err = -ENOMEM;
3571                 goto err_put;
3572         }
3573         uflow_res = flow_resources_alloc(cmd.flow_attr.num_of_specs);
3574         if (!uflow_res) {
3575                 err = -ENOMEM;
3576                 goto err_free_flow_attr;
3577         }
3578
3579         flow_attr->type = kern_flow_attr->type;
3580         flow_attr->priority = kern_flow_attr->priority;
3581         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
3582         flow_attr->port = kern_flow_attr->port;
3583         flow_attr->flags = kern_flow_attr->flags;
3584         flow_attr->size = sizeof(*flow_attr);
3585
3586         kern_spec = kern_flow_attr->flow_specs;
3587         ib_spec = flow_attr + 1;
3588         for (i = 0; i < flow_attr->num_of_specs &&
3589                         cmd.flow_attr.size >= sizeof(*kern_spec) &&
3590                         cmd.flow_attr.size >= kern_spec->size;
3591              i++) {
3592                 err = kern_spec_to_ib_spec(
3593                                 file->ucontext, (struct ib_uverbs_flow_spec *)kern_spec,
3594                                 ib_spec, uflow_res);
3595                 if (err)
3596                         goto err_free;
3597
3598                 flow_attr->size +=
3599                         ((union ib_flow_spec *) ib_spec)->size;
3600                 cmd.flow_attr.size -= kern_spec->size;
3601                 kern_spec = ((void *)kern_spec) + kern_spec->size;
3602                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
3603         }
3604         if (cmd.flow_attr.size || (i != flow_attr->num_of_specs)) {
3605                 pr_warn("create flow failed, flow %d: %d bytes left from uverb cmd\n",
3606                         i, cmd.flow_attr.size);
3607                 err = -EINVAL;
3608                 goto err_free;
3609         }
3610
3611         flow_id = qp->device->create_flow(qp, flow_attr,
3612                                           IB_FLOW_DOMAIN_USER, uhw);
3613
3614         if (IS_ERR(flow_id)) {
3615                 err = PTR_ERR(flow_id);
3616                 goto err_free;
3617         }
3618         atomic_inc(&qp->usecnt);
3619         flow_id->qp = qp;
3620         flow_id->uobject = uobj;
3621         uobj->object = flow_id;
3622         uflow = container_of(uobj, typeof(*uflow), uobject);
3623         uflow->resources = uflow_res;
3624
3625         memset(&resp, 0, sizeof(resp));
3626         resp.flow_handle = uobj->id;
3627
3628         err = ib_copy_to_udata(ucore,
3629                                &resp, sizeof(resp));
3630         if (err)
3631                 goto err_copy;
3632
3633         uobj_put_obj_read(qp);
3634         uobj_alloc_commit(uobj);
3635         kfree(flow_attr);
3636         if (cmd.flow_attr.num_of_specs)
3637                 kfree(kern_flow_attr);
3638         return 0;
3639 err_copy:
3640         ib_destroy_flow(flow_id);
3641 err_free:
3642         ib_uverbs_flow_resources_free(uflow_res);
3643 err_free_flow_attr:
3644         kfree(flow_attr);
3645 err_put:
3646         uobj_put_obj_read(qp);
3647 err_uobj:
3648         uobj_alloc_abort(uobj);
3649 err_free_attr:
3650         if (cmd.flow_attr.num_of_specs)
3651                 kfree(kern_flow_attr);
3652         return err;
3653 }
3654
3655 int ib_uverbs_ex_destroy_flow(struct ib_uverbs_file *file,
3656                               struct ib_device *ib_dev,
3657                               struct ib_udata *ucore,
3658                               struct ib_udata *uhw)
3659 {
3660         struct ib_uverbs_destroy_flow   cmd;
3661         struct ib_uobject               *uobj;
3662         int                             ret;
3663
3664         if (ucore->inlen < sizeof(cmd))
3665                 return -EINVAL;
3666
3667         ret = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3668         if (ret)
3669                 return ret;
3670
3671         if (cmd.comp_mask)
3672                 return -EINVAL;
3673
3674         uobj  = uobj_get_write(UVERBS_OBJECT_FLOW, cmd.flow_handle,
3675                                file->ucontext);
3676         if (IS_ERR(uobj))
3677                 return PTR_ERR(uobj);
3678
3679         ret = uobj_remove_commit(uobj);
3680         return ret;
3681 }
3682
3683 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
3684                                 struct ib_device *ib_dev,
3685                                 struct ib_uverbs_create_xsrq *cmd,
3686                                 struct ib_udata *udata)
3687 {
3688         struct ib_uverbs_create_srq_resp resp;
3689         struct ib_usrq_object           *obj;
3690         struct ib_pd                    *pd;
3691         struct ib_srq                   *srq;
3692         struct ib_uobject               *uninitialized_var(xrcd_uobj);
3693         struct ib_srq_init_attr          attr;
3694         int ret;
3695
3696         obj  = (struct ib_usrq_object *)uobj_alloc(UVERBS_OBJECT_SRQ,
3697                                                    file->ucontext);
3698         if (IS_ERR(obj))
3699                 return PTR_ERR(obj);
3700
3701         if (cmd->srq_type == IB_SRQT_TM)
3702                 attr.ext.tag_matching.max_num_tags = cmd->max_num_tags;
3703
3704         if (cmd->srq_type == IB_SRQT_XRC) {
3705                 xrcd_uobj = uobj_get_read(UVERBS_OBJECT_XRCD, cmd->xrcd_handle,
3706                                           file->ucontext);
3707                 if (IS_ERR(xrcd_uobj)) {
3708                         ret = -EINVAL;
3709                         goto err;
3710                 }
3711
3712                 attr.ext.xrc.xrcd = (struct ib_xrcd *)xrcd_uobj->object;
3713                 if (!attr.ext.xrc.xrcd) {
3714                         ret = -EINVAL;
3715                         goto err_put_xrcd;
3716                 }
3717
3718                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
3719                 atomic_inc(&obj->uxrcd->refcnt);
3720         }
3721
3722         if (ib_srq_has_cq(cmd->srq_type)) {
3723                 attr.ext.cq  = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd->cq_handle,
3724                                                  file->ucontext);
3725                 if (!attr.ext.cq) {
3726                         ret = -EINVAL;
3727                         goto err_put_xrcd;
3728                 }
3729         }
3730
3731         pd  = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd->pd_handle, file->ucontext);
3732         if (!pd) {
3733                 ret = -EINVAL;
3734                 goto err_put_cq;
3735         }
3736
3737         attr.event_handler  = ib_uverbs_srq_event_handler;
3738         attr.srq_context    = file;
3739         attr.srq_type       = cmd->srq_type;
3740         attr.attr.max_wr    = cmd->max_wr;
3741         attr.attr.max_sge   = cmd->max_sge;
3742         attr.attr.srq_limit = cmd->srq_limit;
3743
3744         obj->uevent.events_reported = 0;
3745         INIT_LIST_HEAD(&obj->uevent.event_list);
3746
3747         srq = pd->device->create_srq(pd, &attr, udata);
3748         if (IS_ERR(srq)) {
3749                 ret = PTR_ERR(srq);
3750                 goto err_put;
3751         }
3752
3753         srq->device        = pd->device;
3754         srq->pd            = pd;
3755         srq->srq_type      = cmd->srq_type;
3756         srq->uobject       = &obj->uevent.uobject;
3757         srq->event_handler = attr.event_handler;
3758         srq->srq_context   = attr.srq_context;
3759
3760         if (ib_srq_has_cq(cmd->srq_type)) {
3761                 srq->ext.cq       = attr.ext.cq;
3762                 atomic_inc(&attr.ext.cq->usecnt);
3763         }
3764
3765         if (cmd->srq_type == IB_SRQT_XRC) {
3766                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
3767                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
3768         }
3769
3770         atomic_inc(&pd->usecnt);
3771         atomic_set(&srq->usecnt, 0);
3772
3773         obj->uevent.uobject.object = srq;
3774         obj->uevent.uobject.user_handle = cmd->user_handle;
3775
3776         memset(&resp, 0, sizeof resp);
3777         resp.srq_handle = obj->uevent.uobject.id;
3778         resp.max_wr     = attr.attr.max_wr;
3779         resp.max_sge    = attr.attr.max_sge;
3780         if (cmd->srq_type == IB_SRQT_XRC)
3781                 resp.srqn = srq->ext.xrc.srq_num;
3782
3783         if (copy_to_user(u64_to_user_ptr(cmd->response),
3784                          &resp, sizeof resp)) {
3785                 ret = -EFAULT;
3786                 goto err_copy;
3787         }
3788
3789         if (cmd->srq_type == IB_SRQT_XRC)
3790                 uobj_put_read(xrcd_uobj);
3791
3792         if (ib_srq_has_cq(cmd->srq_type))
3793                 uobj_put_obj_read(attr.ext.cq);
3794
3795         uobj_put_obj_read(pd);
3796         uobj_alloc_commit(&obj->uevent.uobject);
3797
3798         return 0;
3799
3800 err_copy:
3801         ib_destroy_srq(srq);
3802
3803 err_put:
3804         uobj_put_obj_read(pd);
3805
3806 err_put_cq:
3807         if (ib_srq_has_cq(cmd->srq_type))
3808                 uobj_put_obj_read(attr.ext.cq);
3809
3810 err_put_xrcd:
3811         if (cmd->srq_type == IB_SRQT_XRC) {
3812                 atomic_dec(&obj->uxrcd->refcnt);
3813                 uobj_put_read(xrcd_uobj);
3814         }
3815
3816 err:
3817         uobj_alloc_abort(&obj->uevent.uobject);
3818         return ret;
3819 }
3820
3821 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
3822                              struct ib_device *ib_dev,
3823                              const char __user *buf, int in_len,
3824                              int out_len)
3825 {
3826         struct ib_uverbs_create_srq      cmd;
3827         struct ib_uverbs_create_xsrq     xcmd;
3828         struct ib_uverbs_create_srq_resp resp;
3829         struct ib_udata                  udata;
3830         int ret;
3831
3832         if (out_len < sizeof resp)
3833                 return -ENOSPC;
3834
3835         if (copy_from_user(&cmd, buf, sizeof cmd))
3836                 return -EFAULT;
3837
3838         memset(&xcmd, 0, sizeof(xcmd));
3839         xcmd.response    = cmd.response;
3840         xcmd.user_handle = cmd.user_handle;
3841         xcmd.srq_type    = IB_SRQT_BASIC;
3842         xcmd.pd_handle   = cmd.pd_handle;
3843         xcmd.max_wr      = cmd.max_wr;
3844         xcmd.max_sge     = cmd.max_sge;
3845         xcmd.srq_limit   = cmd.srq_limit;
3846
3847         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
3848                    u64_to_user_ptr(cmd.response) + sizeof(resp),
3849                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
3850                    out_len - sizeof(resp));
3851
3852         ret = __uverbs_create_xsrq(file, ib_dev, &xcmd, &udata);
3853         if (ret)
3854                 return ret;
3855
3856         return in_len;
3857 }
3858
3859 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
3860                               struct ib_device *ib_dev,
3861                               const char __user *buf, int in_len, int out_len)
3862 {
3863         struct ib_uverbs_create_xsrq     cmd;
3864         struct ib_uverbs_create_srq_resp resp;
3865         struct ib_udata                  udata;
3866         int ret;
3867
3868         if (out_len < sizeof resp)
3869                 return -ENOSPC;
3870
3871         if (copy_from_user(&cmd, buf, sizeof cmd))
3872                 return -EFAULT;
3873
3874         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
3875                    u64_to_user_ptr(cmd.response) + sizeof(resp),
3876                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
3877                    out_len - sizeof(resp));
3878
3879         ret = __uverbs_create_xsrq(file, ib_dev, &cmd, &udata);
3880         if (ret)
3881                 return ret;
3882
3883         return in_len;
3884 }
3885
3886 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
3887                              struct ib_device *ib_dev,
3888                              const char __user *buf, int in_len,
3889                              int out_len)
3890 {
3891         struct ib_uverbs_modify_srq cmd;
3892         struct ib_udata             udata;
3893         struct ib_srq              *srq;
3894         struct ib_srq_attr          attr;
3895         int                         ret;
3896
3897         if (copy_from_user(&cmd, buf, sizeof cmd))
3898                 return -EFAULT;
3899
3900         ib_uverbs_init_udata(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
3901                    out_len);
3902
3903         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file->ucontext);
3904         if (!srq)
3905                 return -EINVAL;
3906
3907         attr.max_wr    = cmd.max_wr;
3908         attr.srq_limit = cmd.srq_limit;
3909
3910         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
3911
3912         uobj_put_obj_read(srq);
3913
3914         return ret ? ret : in_len;
3915 }
3916
3917 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
3918                             struct ib_device *ib_dev,
3919                             const char __user *buf,
3920                             int in_len, int out_len)
3921 {
3922         struct ib_uverbs_query_srq      cmd;
3923         struct ib_uverbs_query_srq_resp resp;
3924         struct ib_srq_attr              attr;
3925         struct ib_srq                   *srq;
3926         int                             ret;
3927
3928         if (out_len < sizeof resp)
3929                 return -ENOSPC;
3930
3931         if (copy_from_user(&cmd, buf, sizeof cmd))
3932                 return -EFAULT;
3933
3934         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file->ucontext);
3935         if (!srq)
3936                 return -EINVAL;
3937
3938         ret = ib_query_srq(srq, &attr);
3939
3940         uobj_put_obj_read(srq);
3941
3942         if (ret)
3943                 return ret;
3944
3945         memset(&resp, 0, sizeof resp);
3946
3947         resp.max_wr    = attr.max_wr;
3948         resp.max_sge   = attr.max_sge;
3949         resp.srq_limit = attr.srq_limit;
3950
3951         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
3952                 return -EFAULT;
3953
3954         return in_len;
3955 }
3956
3957 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
3958                               struct ib_device *ib_dev,
3959                               const char __user *buf, int in_len,
3960                               int out_len)
3961 {
3962         struct ib_uverbs_destroy_srq      cmd;
3963         struct ib_uverbs_destroy_srq_resp resp;
3964         struct ib_uobject                *uobj;
3965         struct ib_uevent_object          *obj;
3966         int                               ret = -EINVAL;
3967
3968         if (copy_from_user(&cmd, buf, sizeof cmd))
3969                 return -EFAULT;
3970
3971         uobj  = uobj_get_write(UVERBS_OBJECT_SRQ, cmd.srq_handle,
3972                                file->ucontext);
3973         if (IS_ERR(uobj))
3974                 return PTR_ERR(uobj);
3975
3976         obj = container_of(uobj, struct ib_uevent_object, uobject);
3977         /*
3978          * Make sure we don't free the memory in remove_commit as we still
3979          * needs the uobject memory to create the response.
3980          */
3981         uverbs_uobject_get(uobj);
3982
3983         memset(&resp, 0, sizeof(resp));
3984
3985         ret = uobj_remove_commit(uobj);
3986         if (ret) {
3987                 uverbs_uobject_put(uobj);
3988                 return ret;
3989         }
3990         resp.events_reported = obj->events_reported;
3991         uverbs_uobject_put(uobj);
3992         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof(resp)))
3993                 return -EFAULT;
3994
3995         return in_len;
3996 }
3997
3998 int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
3999                               struct ib_device *ib_dev,
4000                               struct ib_udata *ucore,
4001                               struct ib_udata *uhw)
4002 {
4003         struct ib_uverbs_ex_query_device_resp resp = { {0} };
4004         struct ib_uverbs_ex_query_device  cmd;
4005         struct ib_device_attr attr = {0};
4006         int err;
4007
4008         if (!ib_dev->query_device)
4009                 return -EOPNOTSUPP;
4010
4011         if (ucore->inlen < sizeof(cmd))
4012                 return -EINVAL;
4013
4014         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
4015         if (err)
4016                 return err;
4017
4018         if (cmd.comp_mask)
4019                 return -EINVAL;
4020
4021         if (cmd.reserved)
4022                 return -EINVAL;
4023
4024         resp.response_length = offsetof(typeof(resp), odp_caps);
4025
4026         if (ucore->outlen < resp.response_length)
4027                 return -ENOSPC;
4028
4029         err = ib_dev->query_device(ib_dev, &attr, uhw);
4030         if (err)
4031                 return err;
4032
4033         copy_query_dev_fields(file, ib_dev, &resp.base, &attr);
4034
4035         if (ucore->outlen < resp.response_length + sizeof(resp.odp_caps))
4036                 goto end;
4037
4038 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4039         resp.odp_caps.general_caps = attr.odp_caps.general_caps;
4040         resp.odp_caps.per_transport_caps.rc_odp_caps =
4041                 attr.odp_caps.per_transport_caps.rc_odp_caps;
4042         resp.odp_caps.per_transport_caps.uc_odp_caps =
4043                 attr.odp_caps.per_transport_caps.uc_odp_caps;
4044         resp.odp_caps.per_transport_caps.ud_odp_caps =
4045                 attr.odp_caps.per_transport_caps.ud_odp_caps;
4046 #endif
4047         resp.response_length += sizeof(resp.odp_caps);
4048
4049         if (ucore->outlen < resp.response_length + sizeof(resp.timestamp_mask))
4050                 goto end;
4051
4052         resp.timestamp_mask = attr.timestamp_mask;
4053         resp.response_length += sizeof(resp.timestamp_mask);
4054
4055         if (ucore->outlen < resp.response_length + sizeof(resp.hca_core_clock))
4056                 goto end;
4057
4058         resp.hca_core_clock = attr.hca_core_clock;
4059         resp.response_length += sizeof(resp.hca_core_clock);
4060
4061         if (ucore->outlen < resp.response_length + sizeof(resp.device_cap_flags_ex))
4062                 goto end;
4063
4064         resp.device_cap_flags_ex = attr.device_cap_flags;
4065         resp.response_length += sizeof(resp.device_cap_flags_ex);
4066
4067         if (ucore->outlen < resp.response_length + sizeof(resp.rss_caps))
4068                 goto end;
4069
4070         resp.rss_caps.supported_qpts = attr.rss_caps.supported_qpts;
4071         resp.rss_caps.max_rwq_indirection_tables =
4072                 attr.rss_caps.max_rwq_indirection_tables;
4073         resp.rss_caps.max_rwq_indirection_table_size =
4074                 attr.rss_caps.max_rwq_indirection_table_size;
4075
4076         resp.response_length += sizeof(resp.rss_caps);
4077
4078         if (ucore->outlen < resp.response_length + sizeof(resp.max_wq_type_rq))
4079                 goto end;
4080
4081         resp.max_wq_type_rq = attr.max_wq_type_rq;
4082         resp.response_length += sizeof(resp.max_wq_type_rq);
4083
4084         if (ucore->outlen < resp.response_length + sizeof(resp.raw_packet_caps))
4085                 goto end;
4086
4087         resp.raw_packet_caps = attr.raw_packet_caps;
4088         resp.response_length += sizeof(resp.raw_packet_caps);
4089
4090         if (ucore->outlen < resp.response_length + sizeof(resp.tm_caps))
4091                 goto end;
4092
4093         resp.tm_caps.max_rndv_hdr_size  = attr.tm_caps.max_rndv_hdr_size;
4094         resp.tm_caps.max_num_tags       = attr.tm_caps.max_num_tags;
4095         resp.tm_caps.max_ops            = attr.tm_caps.max_ops;
4096         resp.tm_caps.max_sge            = attr.tm_caps.max_sge;
4097         resp.tm_caps.flags              = attr.tm_caps.flags;
4098         resp.response_length += sizeof(resp.tm_caps);
4099
4100         if (ucore->outlen < resp.response_length + sizeof(resp.cq_moderation_caps))
4101                 goto end;
4102
4103         resp.cq_moderation_caps.max_cq_moderation_count  =
4104                 attr.cq_caps.max_cq_moderation_count;
4105         resp.cq_moderation_caps.max_cq_moderation_period =
4106                 attr.cq_caps.max_cq_moderation_period;
4107         resp.response_length += sizeof(resp.cq_moderation_caps);
4108
4109         if (ucore->outlen < resp.response_length + sizeof(resp.max_dm_size))
4110                 goto end;
4111
4112         resp.max_dm_size = attr.max_dm_size;
4113         resp.response_length += sizeof(resp.max_dm_size);
4114 end:
4115         err = ib_copy_to_udata(ucore, &resp, resp.response_length);
4116         return err;
4117 }
4118
4119 int ib_uverbs_ex_modify_cq(struct ib_uverbs_file *file,
4120                            struct ib_device *ib_dev,
4121                            struct ib_udata *ucore,
4122                            struct ib_udata *uhw)
4123 {
4124         struct ib_uverbs_ex_modify_cq cmd = {};
4125         struct ib_cq *cq;
4126         size_t required_cmd_sz;
4127         int ret;
4128
4129         required_cmd_sz = offsetof(typeof(cmd), reserved) +
4130                                 sizeof(cmd.reserved);
4131         if (ucore->inlen < required_cmd_sz)
4132                 return -EINVAL;
4133
4134         /* sanity checks */
4135         if (ucore->inlen > sizeof(cmd) &&
4136             !ib_is_udata_cleared(ucore, sizeof(cmd),
4137                                  ucore->inlen - sizeof(cmd)))
4138                 return -EOPNOTSUPP;
4139
4140         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
4141         if (ret)
4142                 return ret;
4143
4144         if (!cmd.attr_mask || cmd.reserved)
4145                 return -EINVAL;
4146
4147         if (cmd.attr_mask > IB_CQ_MODERATE)
4148                 return -EOPNOTSUPP;
4149
4150         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file->ucontext);
4151         if (!cq)
4152                 return -EINVAL;
4153
4154         ret = rdma_set_cq_moderation(cq, cmd.attr.cq_count, cmd.attr.cq_period);
4155
4156         uobj_put_obj_read(cq);
4157
4158         return ret;
4159 }