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(s32 fd, struct ib_uverbs_file *ufile)
  52 {
  53         struct ib_uobject *uobj = ufd_get_read(UVERBS_OBJECT_COMP_CHANNEL,
  54                                                fd, ufile);
  55
  56         if (IS_ERR(uobj))
  57                 return (void *)uobj;
  58
  59         uverbs_uobject_get(uobj);
  60         uobj_put_read(uobj);
  61
  62         return container_of(uobj, struct ib_uverbs_completion_event_file,
  63                             uobj);
  64 }
  65 #define ib_uverbs_lookup_comp_file(_fd, _ufile)                                \
  66         _ib_uverbs_lookup_comp_file((_fd)*typecheck(s32, _fd), _ufile)
  67
  68 ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
  69                               const char __user *buf,
  70                               int in_len, int out_len)
  71 {
  72         struct ib_uverbs_get_context      cmd;
  73         struct ib_uverbs_get_context_resp resp;
  74         struct ib_udata                   udata;
  75         struct ib_ucontext               *ucontext;
  76         struct file                      *filp;
  77         struct ib_rdmacg_object          cg_obj;
  78         struct ib_device *ib_dev;
  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->ucontext_lock);
  88         ib_dev = srcu_dereference(file->device->ib_dev,
  89                                   &file->device->disassociate_srcu);
  90         if (!ib_dev) {
  91                 ret = -EIO;
  92                 goto err;
  93         }
  94
  95         if (file->ucontext) {
  96                 ret = -EINVAL;
  97                 goto err;
  98         }
  99
 100         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 101                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 102                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 103                    out_len - sizeof(resp));
 104
 105         ret = ib_rdmacg_try_charge(&cg_obj, ib_dev, RDMACG_RESOURCE_HCA_HANDLE);
 106         if (ret)
 107                 goto err;
 108
 109         ucontext = ib_dev->alloc_ucontext(ib_dev, &udata);
 110         if (IS_ERR(ucontext)) {
 111                 ret = PTR_ERR(ucontext);
 112                 goto err_alloc;
 113         }
 114
 115         ucontext->device = ib_dev;
 116         ucontext->cg_obj = cg_obj;
 117         /* ufile is required when some objects are released */
 118         ucontext->ufile = file;
 119
 120         rcu_read_lock();
 121         ucontext->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
 122         rcu_read_unlock();
 123         ucontext->closing = 0;
 124         ucontext->cleanup_retryable = false;
 125
 126 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 127         ucontext->umem_tree = RB_ROOT_CACHED;
 128         init_rwsem(&ucontext->umem_rwsem);
 129         ucontext->odp_mrs_count = 0;
 130         INIT_LIST_HEAD(&ucontext->no_private_counters);
 131
 132         if (!(ib_dev->attrs.device_cap_flags & IB_DEVICE_ON_DEMAND_PAGING))
 133                 ucontext->invalidate_range = NULL;
 134
 135 #endif
 136
 137         resp.num_comp_vectors = file->device->num_comp_vectors;
 138
 139         ret = get_unused_fd_flags(O_CLOEXEC);
 140         if (ret < 0)
 141                 goto err_free;
 142         resp.async_fd = ret;
 143
 144         filp = ib_uverbs_alloc_async_event_file(file, ib_dev);
 145         if (IS_ERR(filp)) {
 146                 ret = PTR_ERR(filp);
 147                 goto err_fd;
 148         }
 149
 150         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 151                 ret = -EFAULT;
 152                 goto err_file;
 153         }
 154
 155         fd_install(resp.async_fd, filp);
 156
 157         /*
 158          * Make sure that ib_uverbs_get_ucontext() sees the pointer update
 159          * only after all writes to setup the ucontext have completed
 160          */
 161         smp_store_release(&file->ucontext, ucontext);
 162
 163         mutex_unlock(&file->ucontext_lock);
 164
 165         return in_len;
 166
 167 err_file:
 168         ib_uverbs_free_async_event_file(file);
 169         fput(filp);
 170
 171 err_fd:
 172         put_unused_fd(resp.async_fd);
 173
 174 err_free:
 175         put_pid(ucontext->tgid);
 176         ib_dev->dealloc_ucontext(ucontext);
 177
 178 err_alloc:
 179         ib_rdmacg_uncharge(&cg_obj, ib_dev, RDMACG_RESOURCE_HCA_HANDLE);
 180
 181 err:
 182         mutex_unlock(&file->ucontext_lock);
 183         return ret;
 184 }
 185
 186 static void copy_query_dev_fields(struct ib_ucontext *ucontext,
 187                                   struct ib_uverbs_query_device_resp *resp,
 188                                   struct ib_device_attr *attr)
 189 {
 190         struct ib_device *ib_dev = ucontext->device;
 191
 192         resp->fw_ver            = attr->fw_ver;
 193         resp->node_guid         = ib_dev->node_guid;
 194         resp->sys_image_guid    = attr->sys_image_guid;
 195         resp->max_mr_size       = attr->max_mr_size;
 196         resp->page_size_cap     = attr->page_size_cap;
 197         resp->vendor_id         = attr->vendor_id;
 198         resp->vendor_part_id    = attr->vendor_part_id;
 199         resp->hw_ver            = attr->hw_ver;
 200         resp->max_qp            = attr->max_qp;
 201         resp->max_qp_wr         = attr->max_qp_wr;
 202         resp->device_cap_flags  = lower_32_bits(attr->device_cap_flags);
 203         resp->max_sge           = min(attr->max_send_sge, attr->max_recv_sge);
 204         resp->max_sge_rd        = attr->max_sge_rd;
 205         resp->max_cq            = attr->max_cq;
 206         resp->max_cqe           = attr->max_cqe;
 207         resp->max_mr            = attr->max_mr;
 208         resp->max_pd            = attr->max_pd;
 209         resp->max_qp_rd_atom    = attr->max_qp_rd_atom;
 210         resp->max_ee_rd_atom    = attr->max_ee_rd_atom;
 211         resp->max_res_rd_atom   = attr->max_res_rd_atom;
 212         resp->max_qp_init_rd_atom       = attr->max_qp_init_rd_atom;
 213         resp->max_ee_init_rd_atom       = attr->max_ee_init_rd_atom;
 214         resp->atomic_cap                = attr->atomic_cap;
 215         resp->max_ee                    = attr->max_ee;
 216         resp->max_rdd                   = attr->max_rdd;
 217         resp->max_mw                    = attr->max_mw;
 218         resp->max_raw_ipv6_qp           = attr->max_raw_ipv6_qp;
 219         resp->max_raw_ethy_qp           = attr->max_raw_ethy_qp;
 220         resp->max_mcast_grp             = attr->max_mcast_grp;
 221         resp->max_mcast_qp_attach       = attr->max_mcast_qp_attach;
 222         resp->max_total_mcast_qp_attach = attr->max_total_mcast_qp_attach;
 223         resp->max_ah                    = attr->max_ah;
 224         resp->max_fmr                   = attr->max_fmr;
 225         resp->max_map_per_fmr           = attr->max_map_per_fmr;
 226         resp->max_srq                   = attr->max_srq;
 227         resp->max_srq_wr                = attr->max_srq_wr;
 228         resp->max_srq_sge               = attr->max_srq_sge;
 229         resp->max_pkeys                 = attr->max_pkeys;
 230         resp->local_ca_ack_delay        = attr->local_ca_ack_delay;
 231         resp->phys_port_cnt             = ib_dev->phys_port_cnt;
 232 }
 233
 234 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 235                                const char __user *buf,
 236                                int in_len, int out_len)
 237 {
 238         struct ib_uverbs_query_device      cmd;
 239         struct ib_uverbs_query_device_resp resp;
 240         struct ib_ucontext *ucontext;
 241
 242         ucontext = ib_uverbs_get_ucontext(file);
 243         if (IS_ERR(ucontext))
 244                 return PTR_ERR(ucontext);
 245
 246         if (out_len < sizeof resp)
 247                 return -ENOSPC;
 248
 249         if (copy_from_user(&cmd, buf, sizeof cmd))
 250                 return -EFAULT;
 251
 252         memset(&resp, 0, sizeof resp);
 253         copy_query_dev_fields(ucontext, &resp, &ucontext->device->attrs);
 254
 255         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
 256                 return -EFAULT;
 257
 258         return in_len;
 259 }
 260
 261 /*
 262  * ib_uverbs_query_port_resp.port_cap_flags started out as just a copy of the
 263  * PortInfo CapabilityMask, but was extended with unique bits.
 264  */
 265 static u32 make_port_cap_flags(const struct ib_port_attr *attr)
 266 {
 267         u32 res;
 268
 269         /* All IBA CapabilityMask bits are passed through here, except bit 26,
 270          * which is overridden with IP_BASED_GIDS. This is due to a historical
 271          * mistake in the implementation of IP_BASED_GIDS. Otherwise all other
 272          * bits match the IBA definition across all kernel versions.
 273          */
 274         res = attr->port_cap_flags & ~(u32)IB_UVERBS_PCF_IP_BASED_GIDS;
 275
 276         if (attr->ip_gids)
 277                 res |= IB_UVERBS_PCF_IP_BASED_GIDS;
 278
 279         return res;
 280 }
 281
 282 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 283                              const char __user *buf,
 284                              int in_len, int out_len)
 285 {
 286         struct ib_uverbs_query_port      cmd;
 287         struct ib_uverbs_query_port_resp resp;
 288         struct ib_port_attr              attr;
 289         int                              ret;
 290         struct ib_ucontext *ucontext;
 291         struct ib_device *ib_dev;
 292
 293         ucontext = ib_uverbs_get_ucontext(file);
 294         if (IS_ERR(ucontext))
 295                 return PTR_ERR(ucontext);
 296         ib_dev = ucontext->device;
 297
 298         if (out_len < sizeof resp)
 299                 return -ENOSPC;
 300
 301         if (copy_from_user(&cmd, buf, sizeof cmd))
 302                 return -EFAULT;
 303
 304         ret = ib_query_port(ib_dev, cmd.port_num, &attr);
 305         if (ret)
 306                 return ret;
 307
 308         memset(&resp, 0, sizeof resp);
 309
 310         resp.state           = attr.state;
 311         resp.max_mtu         = attr.max_mtu;
 312         resp.active_mtu      = attr.active_mtu;
 313         resp.gid_tbl_len     = attr.gid_tbl_len;
 314         resp.port_cap_flags  = make_port_cap_flags(&attr);
 315         resp.max_msg_sz      = attr.max_msg_sz;
 316         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 317         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 318         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 319
 320         if (rdma_is_grh_required(ib_dev, cmd.port_num))
 321                 resp.flags |= IB_UVERBS_QPF_GRH_REQUIRED;
 322
 323         if (rdma_cap_opa_ah(ib_dev, cmd.port_num)) {
 324                 resp.lid     = OPA_TO_IB_UCAST_LID(attr.lid);
 325                 resp.sm_lid  = OPA_TO_IB_UCAST_LID(attr.sm_lid);
 326         } else {
 327                 resp.lid     = ib_lid_cpu16(attr.lid);
 328                 resp.sm_lid  = ib_lid_cpu16(attr.sm_lid);
 329         }
 330         resp.lmc             = attr.lmc;
 331         resp.max_vl_num      = attr.max_vl_num;
 332         resp.sm_sl           = attr.sm_sl;
 333         resp.subnet_timeout  = attr.subnet_timeout;
 334         resp.init_type_reply = attr.init_type_reply;
 335         resp.active_width    = attr.active_width;
 336         resp.active_speed    = attr.active_speed;
 337         resp.phys_state      = attr.phys_state;
 338         resp.link_layer      = rdma_port_get_link_layer(ib_dev,
 339                                                         cmd.port_num);
 340
 341         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
 342                 return -EFAULT;
 343
 344         return in_len;
 345 }
 346
 347 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 348                            const char __user *buf,
 349                            int in_len, int out_len)
 350 {
 351         struct ib_uverbs_alloc_pd      cmd;
 352         struct ib_uverbs_alloc_pd_resp resp;
 353         struct ib_udata                udata;
 354         struct ib_uobject             *uobj;
 355         struct ib_pd                  *pd;
 356         int                            ret;
 357         struct ib_device *ib_dev;
 358
 359         if (out_len < sizeof resp)
 360                 return -ENOSPC;
 361
 362         if (copy_from_user(&cmd, buf, sizeof cmd))
 363                 return -EFAULT;
 364
 365         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 366                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 367                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 368                    out_len - sizeof(resp));
 369
 370         uobj = uobj_alloc(UVERBS_OBJECT_PD, file, &ib_dev);
 371         if (IS_ERR(uobj))
 372                 return PTR_ERR(uobj);
 373
 374         pd = ib_dev->alloc_pd(ib_dev, uobj->context, &udata);
 375         if (IS_ERR(pd)) {
 376                 ret = PTR_ERR(pd);
 377                 goto err;
 378         }
 379
 380         pd->device  = ib_dev;
 381         pd->uobject = uobj;
 382         pd->__internal_mr = NULL;
 383         atomic_set(&pd->usecnt, 0);
 384
 385         uobj->object = pd;
 386         memset(&resp, 0, sizeof resp);
 387         resp.pd_handle = uobj->id;
 388         pd->res.type = RDMA_RESTRACK_PD;
 389         rdma_restrack_add(&pd->res);
 390
 391         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 392                 ret = -EFAULT;
 393                 goto err_copy;
 394         }
 395
 396         return uobj_alloc_commit(uobj, in_len);
 397
 398 err_copy:
 399         ib_dealloc_pd(pd);
 400
 401 err:
 402         uobj_alloc_abort(uobj);
 403         return ret;
 404 }
 405
 406 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 407                              const char __user *buf,
 408                              int in_len, int out_len)
 409 {
 410         struct ib_uverbs_dealloc_pd cmd;
 411
 412         if (copy_from_user(&cmd, buf, sizeof cmd))
 413                 return -EFAULT;
 414
 415         return uobj_perform_destroy(UVERBS_OBJECT_PD, cmd.pd_handle, file,
 416                                     in_len);
 417 }
 418
 419 struct xrcd_table_entry {
 420         struct rb_node  node;
 421         struct ib_xrcd *xrcd;
 422         struct inode   *inode;
 423 };
 424
 425 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 426                             struct inode *inode,
 427                             struct ib_xrcd *xrcd)
 428 {
 429         struct xrcd_table_entry *entry, *scan;
 430         struct rb_node **p = &dev->xrcd_tree.rb_node;
 431         struct rb_node *parent = NULL;
 432
 433         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 434         if (!entry)
 435                 return -ENOMEM;
 436
 437         entry->xrcd  = xrcd;
 438         entry->inode = inode;
 439
 440         while (*p) {
 441                 parent = *p;
 442                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 443
 444                 if (inode < scan->inode) {
 445                         p = &(*p)->rb_left;
 446                 } else if (inode > scan->inode) {
 447                         p = &(*p)->rb_right;
 448                 } else {
 449                         kfree(entry);
 450                         return -EEXIST;
 451                 }
 452         }
 453
 454         rb_link_node(&entry->node, parent, p);
 455         rb_insert_color(&entry->node, &dev->xrcd_tree);
 456         igrab(inode);
 457         return 0;
 458 }
 459
 460 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 461                                                   struct inode *inode)
 462 {
 463         struct xrcd_table_entry *entry;
 464         struct rb_node *p = dev->xrcd_tree.rb_node;
 465
 466         while (p) {
 467                 entry = rb_entry(p, struct xrcd_table_entry, node);
 468
 469                 if (inode < entry->inode)
 470                         p = p->rb_left;
 471                 else if (inode > entry->inode)
 472                         p = p->rb_right;
 473                 else
 474                         return entry;
 475         }
 476
 477         return NULL;
 478 }
 479
 480 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 481 {
 482         struct xrcd_table_entry *entry;
 483
 484         entry = xrcd_table_search(dev, inode);
 485         if (!entry)
 486                 return NULL;
 487
 488         return entry->xrcd;
 489 }
 490
 491 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 492                               struct inode *inode)
 493 {
 494         struct xrcd_table_entry *entry;
 495
 496         entry = xrcd_table_search(dev, inode);
 497         if (entry) {
 498                 iput(inode);
 499                 rb_erase(&entry->node, &dev->xrcd_tree);
 500                 kfree(entry);
 501         }
 502 }
 503
 504 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 505                             const char __user *buf, int in_len,
 506                             int out_len)
 507 {
 508         struct ib_uverbs_open_xrcd      cmd;
 509         struct ib_uverbs_open_xrcd_resp resp;
 510         struct ib_udata                 udata;
 511         struct ib_uxrcd_object         *obj;
 512         struct ib_xrcd                 *xrcd = NULL;
 513         struct fd                       f = {NULL, 0};
 514         struct inode                   *inode = NULL;
 515         int                             ret = 0;
 516         int                             new_xrcd = 0;
 517         struct ib_device *ib_dev;
 518
 519         if (out_len < sizeof resp)
 520                 return -ENOSPC;
 521
 522         if (copy_from_user(&cmd, buf, sizeof cmd))
 523                 return -EFAULT;
 524
 525         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 526                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 527                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 528                    out_len - sizeof(resp));
 529
 530         mutex_lock(&file->device->xrcd_tree_mutex);
 531
 532         if (cmd.fd != -1) {
 533                 /* search for file descriptor */
 534                 f = fdget(cmd.fd);
 535                 if (!f.file) {
 536                         ret = -EBADF;
 537                         goto err_tree_mutex_unlock;
 538                 }
 539
 540                 inode = file_inode(f.file);
 541                 xrcd = find_xrcd(file->device, inode);
 542                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 543                         /* no file descriptor. Need CREATE flag */
 544                         ret = -EAGAIN;
 545                         goto err_tree_mutex_unlock;
 546                 }
 547
 548                 if (xrcd && cmd.oflags & O_EXCL) {
 549                         ret = -EINVAL;
 550                         goto err_tree_mutex_unlock;
 551                 }
 552         }
 553
 554         obj = (struct ib_uxrcd_object *)uobj_alloc(UVERBS_OBJECT_XRCD, file,
 555                                                    &ib_dev);
 556         if (IS_ERR(obj)) {
 557                 ret = PTR_ERR(obj);
 558                 goto err_tree_mutex_unlock;
 559         }
 560
 561         if (!xrcd) {
 562                 xrcd = ib_dev->alloc_xrcd(ib_dev, obj->uobject.context, &udata);
 563                 if (IS_ERR(xrcd)) {
 564                         ret = PTR_ERR(xrcd);
 565                         goto err;
 566                 }
 567
 568                 xrcd->inode   = inode;
 569                 xrcd->device  = ib_dev;
 570                 atomic_set(&xrcd->usecnt, 0);
 571                 mutex_init(&xrcd->tgt_qp_mutex);
 572                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 573                 new_xrcd = 1;
 574         }
 575
 576         atomic_set(&obj->refcnt, 0);
 577         obj->uobject.object = xrcd;
 578         memset(&resp, 0, sizeof resp);
 579         resp.xrcd_handle = obj->uobject.id;
 580
 581         if (inode) {
 582                 if (new_xrcd) {
 583                         /* create new inode/xrcd table entry */
 584                         ret = xrcd_table_insert(file->device, inode, xrcd);
 585                         if (ret)
 586                                 goto err_dealloc_xrcd;
 587                 }
 588                 atomic_inc(&xrcd->usecnt);
 589         }
 590
 591         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 592                 ret = -EFAULT;
 593                 goto err_copy;
 594         }
 595
 596         if (f.file)
 597                 fdput(f);
 598
 599         mutex_unlock(&file->device->xrcd_tree_mutex);
 600
 601         return uobj_alloc_commit(&obj->uobject, in_len);
 602
 603 err_copy:
 604         if (inode) {
 605                 if (new_xrcd)
 606                         xrcd_table_delete(file->device, inode);
 607                 atomic_dec(&xrcd->usecnt);
 608         }
 609
 610 err_dealloc_xrcd:
 611         ib_dealloc_xrcd(xrcd);
 612
 613 err:
 614         uobj_alloc_abort(&obj->uobject);
 615
 616 err_tree_mutex_unlock:
 617         if (f.file)
 618                 fdput(f);
 619
 620         mutex_unlock(&file->device->xrcd_tree_mutex);
 621
 622         return ret;
 623 }
 624
 625 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 626                              const char __user *buf, int in_len,
 627                              int out_len)
 628 {
 629         struct ib_uverbs_close_xrcd cmd;
 630
 631         if (copy_from_user(&cmd, buf, sizeof cmd))
 632                 return -EFAULT;
 633
 634         return uobj_perform_destroy(UVERBS_OBJECT_XRCD, cmd.xrcd_handle, file,
 635                                     in_len);
 636 }
 637
 638 int ib_uverbs_dealloc_xrcd(struct ib_uobject *uobject,
 639                            struct ib_xrcd *xrcd,
 640                            enum rdma_remove_reason why)
 641 {
 642         struct inode *inode;
 643         int ret;
 644         struct ib_uverbs_device *dev = uobject->context->ufile->device;
 645
 646         inode = xrcd->inode;
 647         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 648                 return 0;
 649
 650         ret = ib_dealloc_xrcd(xrcd);
 651
 652         if (ib_is_destroy_retryable(ret, why, uobject)) {
 653                 atomic_inc(&xrcd->usecnt);
 654                 return ret;
 655         }
 656
 657         if (inode)
 658                 xrcd_table_delete(dev, inode);
 659
 660         return ret;
 661 }
 662
 663 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 664                          const char __user *buf, int in_len,
 665                          int out_len)
 666 {
 667         struct ib_uverbs_reg_mr      cmd;
 668         struct ib_uverbs_reg_mr_resp resp;
 669         struct ib_udata              udata;
 670         struct ib_uobject           *uobj;
 671         struct ib_pd                *pd;
 672         struct ib_mr                *mr;
 673         int                          ret;
 674         struct ib_device *ib_dev;
 675
 676         if (out_len < sizeof resp)
 677                 return -ENOSPC;
 678
 679         if (copy_from_user(&cmd, buf, sizeof cmd))
 680                 return -EFAULT;
 681
 682         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 683                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 684                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 685                    out_len - sizeof(resp));
 686
 687         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 688                 return -EINVAL;
 689
 690         ret = ib_check_mr_access(cmd.access_flags);
 691         if (ret)
 692                 return ret;
 693
 694         uobj = uobj_alloc(UVERBS_OBJECT_MR, file, &ib_dev);
 695         if (IS_ERR(uobj))
 696                 return PTR_ERR(uobj);
 697
 698         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file);
 699         if (!pd) {
 700                 ret = -EINVAL;
 701                 goto err_free;
 702         }
 703
 704         if (cmd.access_flags & IB_ACCESS_ON_DEMAND) {
 705                 if (!(pd->device->attrs.device_cap_flags &
 706                       IB_DEVICE_ON_DEMAND_PAGING)) {
 707                         pr_debug("ODP support not available\n");
 708                         ret = -EINVAL;
 709                         goto err_put;
 710                 }
 711         }
 712
 713         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 714                                      cmd.access_flags, &udata);
 715         if (IS_ERR(mr)) {
 716                 ret = PTR_ERR(mr);
 717                 goto err_put;
 718         }
 719
 720         mr->device  = pd->device;
 721         mr->pd      = pd;
 722         mr->dm      = NULL;
 723         mr->uobject = uobj;
 724         atomic_inc(&pd->usecnt);
 725         mr->res.type = RDMA_RESTRACK_MR;
 726         rdma_restrack_add(&mr->res);
 727
 728         uobj->object = mr;
 729
 730         memset(&resp, 0, sizeof resp);
 731         resp.lkey      = mr->lkey;
 732         resp.rkey      = mr->rkey;
 733         resp.mr_handle = uobj->id;
 734
 735         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 736                 ret = -EFAULT;
 737                 goto err_copy;
 738         }
 739
 740         uobj_put_obj_read(pd);
 741
 742         return uobj_alloc_commit(uobj, in_len);
 743
 744 err_copy:
 745         ib_dereg_mr(mr);
 746
 747 err_put:
 748         uobj_put_obj_read(pd);
 749
 750 err_free:
 751         uobj_alloc_abort(uobj);
 752         return ret;
 753 }
 754
 755 ssize_t ib_uverbs_rereg_mr(struct ib_uverbs_file *file,
 756                            const char __user *buf, int in_len,
 757                            int out_len)
 758 {
 759         struct ib_uverbs_rereg_mr      cmd;
 760         struct ib_uverbs_rereg_mr_resp resp;
 761         struct ib_udata              udata;
 762         struct ib_pd                *pd = NULL;
 763         struct ib_mr                *mr;
 764         struct ib_pd                *old_pd;
 765         int                          ret;
 766         struct ib_uobject           *uobj;
 767
 768         if (out_len < sizeof(resp))
 769                 return -ENOSPC;
 770
 771         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 772                 return -EFAULT;
 773
 774         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 775                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 776                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 777                    out_len - sizeof(resp));
 778
 779         if (cmd.flags & ~IB_MR_REREG_SUPPORTED || !cmd.flags)
 780                 return -EINVAL;
 781
 782         if ((cmd.flags & IB_MR_REREG_TRANS) &&
 783             (!cmd.start || !cmd.hca_va || 0 >= cmd.length ||
 784              (cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK)))
 785                         return -EINVAL;
 786
 787         uobj = uobj_get_write(UVERBS_OBJECT_MR, cmd.mr_handle, file);
 788         if (IS_ERR(uobj))
 789                 return PTR_ERR(uobj);
 790
 791         mr = uobj->object;
 792
 793         if (mr->dm) {
 794                 ret = -EINVAL;
 795                 goto put_uobjs;
 796         }
 797
 798         if (cmd.flags & IB_MR_REREG_ACCESS) {
 799                 ret = ib_check_mr_access(cmd.access_flags);
 800                 if (ret)
 801                         goto put_uobjs;
 802         }
 803
 804         if (cmd.flags & IB_MR_REREG_PD) {
 805                 pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle,
 806                                        file);
 807                 if (!pd) {
 808                         ret = -EINVAL;
 809                         goto put_uobjs;
 810                 }
 811         }
 812
 813         old_pd = mr->pd;
 814         ret = mr->device->rereg_user_mr(mr, cmd.flags, cmd.start,
 815                                         cmd.length, cmd.hca_va,
 816                                         cmd.access_flags, pd, &udata);
 817         if (!ret) {
 818                 if (cmd.flags & IB_MR_REREG_PD) {
 819                         atomic_inc(&pd->usecnt);
 820                         mr->pd = pd;
 821                         atomic_dec(&old_pd->usecnt);
 822                 }
 823         } else {
 824                 goto put_uobj_pd;
 825         }
 826
 827         memset(&resp, 0, sizeof(resp));
 828         resp.lkey      = mr->lkey;
 829         resp.rkey      = mr->rkey;
 830
 831         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof(resp)))
 832                 ret = -EFAULT;
 833         else
 834                 ret = in_len;
 835
 836 put_uobj_pd:
 837         if (cmd.flags & IB_MR_REREG_PD)
 838                 uobj_put_obj_read(pd);
 839
 840 put_uobjs:
 841         uobj_put_write(uobj);
 842
 843         return ret;
 844 }
 845
 846 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
 847                            const char __user *buf, int in_len,
 848                            int out_len)
 849 {
 850         struct ib_uverbs_dereg_mr cmd;
 851
 852         if (copy_from_user(&cmd, buf, sizeof cmd))
 853                 return -EFAULT;
 854
 855         return uobj_perform_destroy(UVERBS_OBJECT_MR, cmd.mr_handle, file,
 856                                     in_len);
 857 }
 858
 859 ssize_t ib_uverbs_alloc_mw(struct ib_uverbs_file *file,
 860                            const char __user *buf, int in_len,
 861                            int out_len)
 862 {
 863         struct ib_uverbs_alloc_mw      cmd;
 864         struct ib_uverbs_alloc_mw_resp resp;
 865         struct ib_uobject             *uobj;
 866         struct ib_pd                  *pd;
 867         struct ib_mw                  *mw;
 868         struct ib_udata                udata;
 869         int                            ret;
 870         struct ib_device *ib_dev;
 871
 872         if (out_len < sizeof(resp))
 873                 return -ENOSPC;
 874
 875         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 876                 return -EFAULT;
 877
 878         uobj = uobj_alloc(UVERBS_OBJECT_MW, file, &ib_dev);
 879         if (IS_ERR(uobj))
 880                 return PTR_ERR(uobj);
 881
 882         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file);
 883         if (!pd) {
 884                 ret = -EINVAL;
 885                 goto err_free;
 886         }
 887
 888         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 889                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 890                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 891                    out_len - sizeof(resp));
 892
 893         mw = pd->device->alloc_mw(pd, cmd.mw_type, &udata);
 894         if (IS_ERR(mw)) {
 895                 ret = PTR_ERR(mw);
 896                 goto err_put;
 897         }
 898
 899         mw->device  = pd->device;
 900         mw->pd      = pd;
 901         mw->uobject = uobj;
 902         atomic_inc(&pd->usecnt);
 903
 904         uobj->object = mw;
 905
 906         memset(&resp, 0, sizeof(resp));
 907         resp.rkey      = mw->rkey;
 908         resp.mw_handle = uobj->id;
 909
 910         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof(resp))) {
 911                 ret = -EFAULT;
 912                 goto err_copy;
 913         }
 914
 915         uobj_put_obj_read(pd);
 916         return uobj_alloc_commit(uobj, in_len);
 917
 918 err_copy:
 919         uverbs_dealloc_mw(mw);
 920 err_put:
 921         uobj_put_obj_read(pd);
 922 err_free:
 923         uobj_alloc_abort(uobj);
 924         return ret;
 925 }
 926
 927 ssize_t ib_uverbs_dealloc_mw(struct ib_uverbs_file *file,
 928                              const char __user *buf, int in_len,
 929                              int out_len)
 930 {
 931         struct ib_uverbs_dealloc_mw cmd;
 932
 933         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 934                 return -EFAULT;
 935
 936         return uobj_perform_destroy(UVERBS_OBJECT_MW, cmd.mw_handle, file,
 937                                     in_len);
 938 }
 939
 940 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
 941                                       const char __user *buf, int in_len,
 942                                       int out_len)
 943 {
 944         struct ib_uverbs_create_comp_channel       cmd;
 945         struct ib_uverbs_create_comp_channel_resp  resp;
 946         struct ib_uobject                         *uobj;
 947         struct ib_uverbs_completion_event_file    *ev_file;
 948         struct ib_device *ib_dev;
 949
 950         if (out_len < sizeof resp)
 951                 return -ENOSPC;
 952
 953         if (copy_from_user(&cmd, buf, sizeof cmd))
 954                 return -EFAULT;
 955
 956         uobj = uobj_alloc(UVERBS_OBJECT_COMP_CHANNEL, file, &ib_dev);
 957         if (IS_ERR(uobj))
 958                 return PTR_ERR(uobj);
 959
 960         resp.fd = uobj->id;
 961
 962         ev_file = container_of(uobj, struct ib_uverbs_completion_event_file,
 963                                uobj);
 964         ib_uverbs_init_event_queue(&ev_file->ev_queue);
 965
 966         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 967                 uobj_alloc_abort(uobj);
 968                 return -EFAULT;
 969         }
 970
 971         return uobj_alloc_commit(uobj, in_len);
 972 }
 973
 974 static struct ib_ucq_object *create_cq(struct ib_uverbs_file *file,
 975                                        struct ib_udata *ucore,
 976                                        struct ib_udata *uhw,
 977                                        struct ib_uverbs_ex_create_cq *cmd,
 978                                        size_t cmd_sz,
 979                                        int (*cb)(struct ib_uverbs_file *file,
 980                                                  struct ib_ucq_object *obj,
 981                                                  struct ib_uverbs_ex_create_cq_resp *resp,
 982                                                  struct ib_udata *udata,
 983                                                  void *context),
 984                                        void *context)
 985 {
 986         struct ib_ucq_object           *obj;
 987         struct ib_uverbs_completion_event_file    *ev_file = NULL;
 988         struct ib_cq                   *cq;
 989         int                             ret;
 990         struct ib_uverbs_ex_create_cq_resp resp;
 991         struct ib_cq_init_attr attr = {};
 992         struct ib_device *ib_dev;
 993
 994         if (cmd->comp_vector >= file->device->num_comp_vectors)
 995                 return ERR_PTR(-EINVAL);
 996
 997         obj = (struct ib_ucq_object *)uobj_alloc(UVERBS_OBJECT_CQ, file,
 998                                                  &ib_dev);
 999         if (IS_ERR(obj))
1000                 return obj;
1001
1002         if (!ib_dev->create_cq) {
1003                 ret = -EOPNOTSUPP;
1004                 goto err;
1005         }
1006
1007         if (cmd->comp_channel >= 0) {
1008                 ev_file = ib_uverbs_lookup_comp_file(cmd->comp_channel, file);
1009                 if (IS_ERR(ev_file)) {
1010                         ret = PTR_ERR(ev_file);
1011                         goto err;
1012                 }
1013         }
1014
1015         obj->uobject.user_handle = cmd->user_handle;
1016         obj->comp_events_reported  = 0;
1017         obj->async_events_reported = 0;
1018         INIT_LIST_HEAD(&obj->comp_list);
1019         INIT_LIST_HEAD(&obj->async_list);
1020
1021         attr.cqe = cmd->cqe;
1022         attr.comp_vector = cmd->comp_vector;
1023
1024         if (cmd_sz > offsetof(typeof(*cmd), flags) + sizeof(cmd->flags))
1025                 attr.flags = cmd->flags;
1026
1027         cq = ib_dev->create_cq(ib_dev, &attr, obj->uobject.context, uhw);
1028         if (IS_ERR(cq)) {
1029                 ret = PTR_ERR(cq);
1030                 goto err_file;
1031         }
1032
1033         cq->device        = ib_dev;
1034         cq->uobject       = &obj->uobject;
1035         cq->comp_handler  = ib_uverbs_comp_handler;
1036         cq->event_handler = ib_uverbs_cq_event_handler;
1037         cq->cq_context    = ev_file ? &ev_file->ev_queue : NULL;
1038         atomic_set(&cq->usecnt, 0);
1039
1040         obj->uobject.object = cq;
1041         memset(&resp, 0, sizeof resp);
1042         resp.base.cq_handle = obj->uobject.id;
1043         resp.base.cqe       = cq->cqe;
1044
1045         resp.response_length = offsetof(typeof(resp), response_length) +
1046                 sizeof(resp.response_length);
1047
1048         cq->res.type = RDMA_RESTRACK_CQ;
1049         rdma_restrack_add(&cq->res);
1050
1051         ret = cb(file, obj, &resp, ucore, context);
1052         if (ret)
1053                 goto err_cb;
1054
1055         ret = uobj_alloc_commit(&obj->uobject, 0);
1056         if (ret)
1057                 return ERR_PTR(ret);
1058         return obj;
1059
1060 err_cb:
1061         ib_destroy_cq(cq);
1062
1063 err_file:
1064         if (ev_file)
1065                 ib_uverbs_release_ucq(file, ev_file, obj);
1066
1067 err:
1068         uobj_alloc_abort(&obj->uobject);
1069
1070         return ERR_PTR(ret);
1071 }
1072
1073 static int ib_uverbs_create_cq_cb(struct ib_uverbs_file *file,
1074                                   struct ib_ucq_object *obj,
1075                                   struct ib_uverbs_ex_create_cq_resp *resp,
1076                                   struct ib_udata *ucore, void *context)
1077 {
1078         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1079                 return -EFAULT;
1080
1081         return 0;
1082 }
1083
1084 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1085                             const char __user *buf, int in_len,
1086                             int out_len)
1087 {
1088         struct ib_uverbs_create_cq      cmd;
1089         struct ib_uverbs_ex_create_cq   cmd_ex;
1090         struct ib_uverbs_create_cq_resp resp;
1091         struct ib_udata                 ucore;
1092         struct ib_udata                 uhw;
1093         struct ib_ucq_object           *obj;
1094
1095         if (out_len < sizeof(resp))
1096                 return -ENOSPC;
1097
1098         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1099                 return -EFAULT;
1100
1101         ib_uverbs_init_udata(&ucore, buf, u64_to_user_ptr(cmd.response),
1102                              sizeof(cmd), sizeof(resp));
1103
1104         ib_uverbs_init_udata(&uhw, buf + sizeof(cmd),
1105                    u64_to_user_ptr(cmd.response) + sizeof(resp),
1106                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1107                    out_len - sizeof(resp));
1108
1109         memset(&cmd_ex, 0, sizeof(cmd_ex));
1110         cmd_ex.user_handle = cmd.user_handle;
1111         cmd_ex.cqe = cmd.cqe;
1112         cmd_ex.comp_vector = cmd.comp_vector;
1113         cmd_ex.comp_channel = cmd.comp_channel;
1114
1115         obj = create_cq(file, &ucore, &uhw, &cmd_ex,
1116                         offsetof(typeof(cmd_ex), comp_channel) +
1117                         sizeof(cmd.comp_channel), ib_uverbs_create_cq_cb,
1118                         NULL);
1119
1120         if (IS_ERR(obj))
1121                 return PTR_ERR(obj);
1122
1123         return in_len;
1124 }
1125
1126 static int ib_uverbs_ex_create_cq_cb(struct ib_uverbs_file *file,
1127                                      struct ib_ucq_object *obj,
1128                                      struct ib_uverbs_ex_create_cq_resp *resp,
1129                                      struct ib_udata *ucore, void *context)
1130 {
1131         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1132                 return -EFAULT;
1133
1134         return 0;
1135 }
1136
1137 int ib_uverbs_ex_create_cq(struct ib_uverbs_file *file,
1138                            struct ib_udata *ucore,
1139                            struct ib_udata *uhw)
1140 {
1141         struct ib_uverbs_ex_create_cq_resp resp;
1142         struct ib_uverbs_ex_create_cq  cmd;
1143         struct ib_ucq_object           *obj;
1144         int err;
1145
1146         if (ucore->inlen < sizeof(cmd))
1147                 return -EINVAL;
1148
1149         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
1150         if (err)
1151                 return err;
1152
1153         if (cmd.comp_mask)
1154                 return -EINVAL;
1155
1156         if (cmd.reserved)
1157                 return -EINVAL;
1158
1159         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1160                              sizeof(resp.response_length)))
1161                 return -ENOSPC;
1162
1163         obj = create_cq(file, ucore, uhw, &cmd,
1164                         min(ucore->inlen, sizeof(cmd)),
1165                         ib_uverbs_ex_create_cq_cb, NULL);
1166
1167         return PTR_ERR_OR_ZERO(obj);
1168 }
1169
1170 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1171                             const char __user *buf, int in_len,
1172                             int out_len)
1173 {
1174         struct ib_uverbs_resize_cq      cmd;
1175         struct ib_uverbs_resize_cq_resp resp = {};
1176         struct ib_udata                 udata;
1177         struct ib_cq                    *cq;
1178         int                             ret = -EINVAL;
1179
1180         if (copy_from_user(&cmd, buf, sizeof cmd))
1181                 return -EFAULT;
1182
1183         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
1184                    u64_to_user_ptr(cmd.response) + sizeof(resp),
1185                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1186                    out_len - sizeof(resp));
1187
1188         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
1189         if (!cq)
1190                 return -EINVAL;
1191
1192         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1193         if (ret)
1194                 goto out;
1195
1196         resp.cqe = cq->cqe;
1197
1198         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp.cqe))
1199                 ret = -EFAULT;
1200
1201 out:
1202         uobj_put_obj_read(cq);
1203
1204         return ret ? ret : in_len;
1205 }
1206
1207 static int copy_wc_to_user(struct ib_device *ib_dev, void __user *dest,
1208                            struct ib_wc *wc)
1209 {
1210         struct ib_uverbs_wc tmp;
1211
1212         tmp.wr_id               = wc->wr_id;
1213         tmp.status              = wc->status;
1214         tmp.opcode              = wc->opcode;
1215         tmp.vendor_err          = wc->vendor_err;
1216         tmp.byte_len            = wc->byte_len;
1217         tmp.ex.imm_data         = wc->ex.imm_data;
1218         tmp.qp_num              = wc->qp->qp_num;
1219         tmp.src_qp              = wc->src_qp;
1220         tmp.wc_flags            = wc->wc_flags;
1221         tmp.pkey_index          = wc->pkey_index;
1222         if (rdma_cap_opa_ah(ib_dev, wc->port_num))
1223                 tmp.slid        = OPA_TO_IB_UCAST_LID(wc->slid);
1224         else
1225                 tmp.slid        = ib_lid_cpu16(wc->slid);
1226         tmp.sl                  = wc->sl;
1227         tmp.dlid_path_bits      = wc->dlid_path_bits;
1228         tmp.port_num            = wc->port_num;
1229         tmp.reserved            = 0;
1230
1231         if (copy_to_user(dest, &tmp, sizeof tmp))
1232                 return -EFAULT;
1233
1234         return 0;
1235 }
1236
1237 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1238                           const char __user *buf, int in_len,
1239                           int out_len)
1240 {
1241         struct ib_uverbs_poll_cq       cmd;
1242         struct ib_uverbs_poll_cq_resp  resp;
1243         u8 __user                     *header_ptr;
1244         u8 __user                     *data_ptr;
1245         struct ib_cq                  *cq;
1246         struct ib_wc                   wc;
1247         int                            ret;
1248
1249         if (copy_from_user(&cmd, buf, sizeof cmd))
1250                 return -EFAULT;
1251
1252         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
1253         if (!cq)
1254                 return -EINVAL;
1255
1256         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1257         header_ptr = u64_to_user_ptr(cmd.response);
1258         data_ptr = header_ptr + sizeof resp;
1259
1260         memset(&resp, 0, sizeof resp);
1261         while (resp.count < cmd.ne) {
1262                 ret = ib_poll_cq(cq, 1, &wc);
1263                 if (ret < 0)
1264                         goto out_put;
1265                 if (!ret)
1266                         break;
1267
1268                 ret = copy_wc_to_user(cq->device, data_ptr, &wc);
1269                 if (ret)
1270                         goto out_put;
1271
1272                 data_ptr += sizeof(struct ib_uverbs_wc);
1273                 ++resp.count;
1274         }
1275
1276         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1277                 ret = -EFAULT;
1278                 goto out_put;
1279         }
1280
1281         ret = in_len;
1282
1283 out_put:
1284         uobj_put_obj_read(cq);
1285         return ret;
1286 }
1287
1288 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1289                                 const char __user *buf, int in_len,
1290                                 int out_len)
1291 {
1292         struct ib_uverbs_req_notify_cq cmd;
1293         struct ib_cq                  *cq;
1294
1295         if (copy_from_user(&cmd, buf, sizeof cmd))
1296                 return -EFAULT;
1297
1298         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
1299         if (!cq)
1300                 return -EINVAL;
1301
1302         ib_req_notify_cq(cq, cmd.solicited_only ?
1303                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1304
1305         uobj_put_obj_read(cq);
1306
1307         return in_len;
1308 }
1309
1310 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1311                              const char __user *buf, int in_len,
1312                              int out_len)
1313 {
1314         struct ib_uverbs_destroy_cq      cmd;
1315         struct ib_uverbs_destroy_cq_resp resp;
1316         struct ib_uobject               *uobj;
1317         struct ib_ucq_object            *obj;
1318
1319         if (copy_from_user(&cmd, buf, sizeof cmd))
1320                 return -EFAULT;
1321
1322         uobj = uobj_get_destroy(UVERBS_OBJECT_CQ, cmd.cq_handle, file);
1323         if (IS_ERR(uobj))
1324                 return PTR_ERR(uobj);
1325
1326         obj = container_of(uobj, struct ib_ucq_object, uobject);
1327         memset(&resp, 0, sizeof(resp));
1328         resp.comp_events_reported  = obj->comp_events_reported;
1329         resp.async_events_reported = obj->async_events_reported;
1330
1331         uobj_put_destroy(uobj);
1332
1333         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
1334                 return -EFAULT;
1335
1336         return in_len;
1337 }
1338
1339 static int create_qp(struct ib_uverbs_file *file,
1340                      struct ib_udata *ucore,
1341                      struct ib_udata *uhw,
1342                      struct ib_uverbs_ex_create_qp *cmd,
1343                      size_t cmd_sz,
1344                      int (*cb)(struct ib_uverbs_file *file,
1345                                struct ib_uverbs_ex_create_qp_resp *resp,
1346                                struct ib_udata *udata),
1347                      void *context)
1348 {
1349         struct ib_uqp_object            *obj;
1350         struct ib_device                *device;
1351         struct ib_pd                    *pd = NULL;
1352         struct ib_xrcd                  *xrcd = NULL;
1353         struct ib_uobject               *xrcd_uobj = ERR_PTR(-ENOENT);
1354         struct ib_cq                    *scq = NULL, *rcq = NULL;
1355         struct ib_srq                   *srq = NULL;
1356         struct ib_qp                    *qp;
1357         char                            *buf;
1358         struct ib_qp_init_attr          attr = {};
1359         struct ib_uverbs_ex_create_qp_resp resp;
1360         int                             ret;
1361         struct ib_rwq_ind_table *ind_tbl = NULL;
1362         bool has_sq = true;
1363         struct ib_device *ib_dev;
1364
1365         if (cmd->qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1366                 return -EPERM;
1367
1368         obj = (struct ib_uqp_object *)uobj_alloc(UVERBS_OBJECT_QP, file,
1369                                                  &ib_dev);
1370         if (IS_ERR(obj))
1371                 return PTR_ERR(obj);
1372         obj->uxrcd = NULL;
1373         obj->uevent.uobject.user_handle = cmd->user_handle;
1374         mutex_init(&obj->mcast_lock);
1375
1376         if (cmd_sz >= offsetof(typeof(*cmd), rwq_ind_tbl_handle) +
1377                       sizeof(cmd->rwq_ind_tbl_handle) &&
1378                       (cmd->comp_mask & IB_UVERBS_CREATE_QP_MASK_IND_TABLE)) {
1379                 ind_tbl = uobj_get_obj_read(rwq_ind_table,
1380                                             UVERBS_OBJECT_RWQ_IND_TBL,
1381                                             cmd->rwq_ind_tbl_handle, file);
1382                 if (!ind_tbl) {
1383                         ret = -EINVAL;
1384                         goto err_put;
1385                 }
1386
1387                 attr.rwq_ind_tbl = ind_tbl;
1388         }
1389
1390         if (cmd_sz > sizeof(*cmd) &&
1391             !ib_is_udata_cleared(ucore, sizeof(*cmd),
1392                                  cmd_sz - sizeof(*cmd))) {
1393                 ret = -EOPNOTSUPP;
1394                 goto err_put;
1395         }
1396
1397         if (ind_tbl && (cmd->max_recv_wr || cmd->max_recv_sge || cmd->is_srq)) {
1398                 ret = -EINVAL;
1399                 goto err_put;
1400         }
1401
1402         if (ind_tbl && !cmd->max_send_wr)
1403                 has_sq = false;
1404
1405         if (cmd->qp_type == IB_QPT_XRC_TGT) {
1406                 xrcd_uobj = uobj_get_read(UVERBS_OBJECT_XRCD, cmd->pd_handle,
1407                                           file);
1408
1409                 if (IS_ERR(xrcd_uobj)) {
1410                         ret = -EINVAL;
1411                         goto err_put;
1412                 }
1413
1414                 xrcd = (struct ib_xrcd *)xrcd_uobj->object;
1415                 if (!xrcd) {
1416                         ret = -EINVAL;
1417                         goto err_put;
1418                 }
1419                 device = xrcd->device;
1420         } else {
1421                 if (cmd->qp_type == IB_QPT_XRC_INI) {
1422                         cmd->max_recv_wr = 0;
1423                         cmd->max_recv_sge = 0;
1424                 } else {
1425                         if (cmd->is_srq) {
1426                                 srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ,
1427                                                         cmd->srq_handle, file);
1428                                 if (!srq || srq->srq_type == IB_SRQT_XRC) {
1429                                         ret = -EINVAL;
1430                                         goto err_put;
1431                                 }
1432                         }
1433
1434                         if (!ind_tbl) {
1435                                 if (cmd->recv_cq_handle != cmd->send_cq_handle) {
1436                                         rcq = uobj_get_obj_read(
1437                                                 cq, UVERBS_OBJECT_CQ,
1438                                                 cmd->recv_cq_handle, file);
1439                                         if (!rcq) {
1440                                                 ret = -EINVAL;
1441                                                 goto err_put;
1442                                         }
1443                                 }
1444                         }
1445                 }
1446
1447                 if (has_sq)
1448                         scq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ,
1449                                                 cmd->send_cq_handle, file);
1450                 if (!ind_tbl)
1451                         rcq = rcq ?: scq;
1452                 pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd->pd_handle,
1453                                        file);
1454                 if (!pd || (!scq && has_sq)) {
1455                         ret = -EINVAL;
1456                         goto err_put;
1457                 }
1458
1459                 device = pd->device;
1460         }
1461
1462         attr.event_handler = ib_uverbs_qp_event_handler;
1463         attr.qp_context    = file;
1464         attr.send_cq       = scq;
1465         attr.recv_cq       = rcq;
1466         attr.srq           = srq;
1467         attr.xrcd          = xrcd;
1468         attr.sq_sig_type   = cmd->sq_sig_all ? IB_SIGNAL_ALL_WR :
1469                                               IB_SIGNAL_REQ_WR;
1470         attr.qp_type       = cmd->qp_type;
1471         attr.create_flags  = 0;
1472
1473         attr.cap.max_send_wr     = cmd->max_send_wr;
1474         attr.cap.max_recv_wr     = cmd->max_recv_wr;
1475         attr.cap.max_send_sge    = cmd->max_send_sge;
1476         attr.cap.max_recv_sge    = cmd->max_recv_sge;
1477         attr.cap.max_inline_data = cmd->max_inline_data;
1478
1479         obj->uevent.events_reported     = 0;
1480         INIT_LIST_HEAD(&obj->uevent.event_list);
1481         INIT_LIST_HEAD(&obj->mcast_list);
1482
1483         if (cmd_sz >= offsetof(typeof(*cmd), create_flags) +
1484                       sizeof(cmd->create_flags))
1485                 attr.create_flags = cmd->create_flags;
1486
1487         if (attr.create_flags & ~(IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
1488                                 IB_QP_CREATE_CROSS_CHANNEL |
1489                                 IB_QP_CREATE_MANAGED_SEND |
1490                                 IB_QP_CREATE_MANAGED_RECV |
1491                                 IB_QP_CREATE_SCATTER_FCS |
1492                                 IB_QP_CREATE_CVLAN_STRIPPING |
1493                                 IB_QP_CREATE_SOURCE_QPN |
1494                                 IB_QP_CREATE_PCI_WRITE_END_PADDING)) {
1495                 ret = -EINVAL;
1496                 goto err_put;
1497         }
1498
1499         if (attr.create_flags & IB_QP_CREATE_SOURCE_QPN) {
1500                 if (!capable(CAP_NET_RAW)) {
1501                         ret = -EPERM;
1502                         goto err_put;
1503                 }
1504
1505                 attr.source_qpn = cmd->source_qpn;
1506         }
1507
1508         buf = (void *)cmd + sizeof(*cmd);
1509         if (cmd_sz > sizeof(*cmd))
1510                 if (!(buf[0] == 0 && !memcmp(buf, buf + 1,
1511                                              cmd_sz - sizeof(*cmd) - 1))) {
1512                         ret = -EINVAL;
1513                         goto err_put;
1514                 }
1515
1516         if (cmd->qp_type == IB_QPT_XRC_TGT)
1517                 qp = ib_create_qp(pd, &attr);
1518         else
1519                 qp = _ib_create_qp(device, pd, &attr, uhw,
1520                                    &obj->uevent.uobject);
1521
1522         if (IS_ERR(qp)) {
1523                 ret = PTR_ERR(qp);
1524                 goto err_put;
1525         }
1526
1527         if (cmd->qp_type != IB_QPT_XRC_TGT) {
1528                 ret = ib_create_qp_security(qp, device);
1529                 if (ret)
1530                         goto err_cb;
1531
1532                 qp->real_qp       = qp;
1533                 qp->pd            = pd;
1534                 qp->send_cq       = attr.send_cq;
1535                 qp->recv_cq       = attr.recv_cq;
1536                 qp->srq           = attr.srq;
1537                 qp->rwq_ind_tbl   = ind_tbl;
1538                 qp->event_handler = attr.event_handler;
1539                 qp->qp_context    = attr.qp_context;
1540                 qp->qp_type       = attr.qp_type;
1541                 atomic_set(&qp->usecnt, 0);
1542                 atomic_inc(&pd->usecnt);
1543                 qp->port = 0;
1544                 if (attr.send_cq)
1545                         atomic_inc(&attr.send_cq->usecnt);
1546                 if (attr.recv_cq)
1547                         atomic_inc(&attr.recv_cq->usecnt);
1548                 if (attr.srq)
1549                         atomic_inc(&attr.srq->usecnt);
1550                 if (ind_tbl)
1551                         atomic_inc(&ind_tbl->usecnt);
1552         } else {
1553                 /* It is done in _ib_create_qp for other QP types */
1554                 qp->uobject = &obj->uevent.uobject;
1555         }
1556
1557         obj->uevent.uobject.object = qp;
1558
1559         memset(&resp, 0, sizeof resp);
1560         resp.base.qpn             = qp->qp_num;
1561         resp.base.qp_handle       = obj->uevent.uobject.id;
1562         resp.base.max_recv_sge    = attr.cap.max_recv_sge;
1563         resp.base.max_send_sge    = attr.cap.max_send_sge;
1564         resp.base.max_recv_wr     = attr.cap.max_recv_wr;
1565         resp.base.max_send_wr     = attr.cap.max_send_wr;
1566         resp.base.max_inline_data = attr.cap.max_inline_data;
1567
1568         resp.response_length = offsetof(typeof(resp), response_length) +
1569                                sizeof(resp.response_length);
1570
1571         ret = cb(file, &resp, ucore);
1572         if (ret)
1573                 goto err_cb;
1574
1575         if (xrcd) {
1576                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object,
1577                                           uobject);
1578                 atomic_inc(&obj->uxrcd->refcnt);
1579                 uobj_put_read(xrcd_uobj);
1580         }
1581
1582         if (pd)
1583                 uobj_put_obj_read(pd);
1584         if (scq)
1585                 uobj_put_obj_read(scq);
1586         if (rcq && rcq != scq)
1587                 uobj_put_obj_read(rcq);
1588         if (srq)
1589                 uobj_put_obj_read(srq);
1590         if (ind_tbl)
1591                 uobj_put_obj_read(ind_tbl);
1592
1593         return uobj_alloc_commit(&obj->uevent.uobject, 0);
1594 err_cb:
1595         ib_destroy_qp(qp);
1596
1597 err_put:
1598         if (!IS_ERR(xrcd_uobj))
1599                 uobj_put_read(xrcd_uobj);
1600         if (pd)
1601                 uobj_put_obj_read(pd);
1602         if (scq)
1603                 uobj_put_obj_read(scq);
1604         if (rcq && rcq != scq)
1605                 uobj_put_obj_read(rcq);
1606         if (srq)
1607                 uobj_put_obj_read(srq);
1608         if (ind_tbl)
1609                 uobj_put_obj_read(ind_tbl);
1610
1611         uobj_alloc_abort(&obj->uevent.uobject);
1612         return ret;
1613 }
1614
1615 static int ib_uverbs_create_qp_cb(struct ib_uverbs_file *file,
1616                                   struct ib_uverbs_ex_create_qp_resp *resp,
1617                                   struct ib_udata *ucore)
1618 {
1619         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1620                 return -EFAULT;
1621
1622         return 0;
1623 }
1624
1625 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1626                             const char __user *buf, int in_len,
1627                             int out_len)
1628 {
1629         struct ib_uverbs_create_qp      cmd;
1630         struct ib_uverbs_ex_create_qp   cmd_ex;
1631         struct ib_udata                 ucore;
1632         struct ib_udata                 uhw;
1633         ssize_t resp_size = sizeof(struct ib_uverbs_create_qp_resp);
1634         int                             err;
1635
1636         if (out_len < resp_size)
1637                 return -ENOSPC;
1638
1639         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1640                 return -EFAULT;
1641
1642         ib_uverbs_init_udata(&ucore, buf, u64_to_user_ptr(cmd.response),
1643                    sizeof(cmd), resp_size);
1644         ib_uverbs_init_udata(&uhw, buf + sizeof(cmd),
1645                    u64_to_user_ptr(cmd.response) + resp_size,
1646                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1647                    out_len - resp_size);
1648
1649         memset(&cmd_ex, 0, sizeof(cmd_ex));
1650         cmd_ex.user_handle = cmd.user_handle;
1651         cmd_ex.pd_handle = cmd.pd_handle;
1652         cmd_ex.send_cq_handle = cmd.send_cq_handle;
1653         cmd_ex.recv_cq_handle = cmd.recv_cq_handle;
1654         cmd_ex.srq_handle = cmd.srq_handle;
1655         cmd_ex.max_send_wr = cmd.max_send_wr;
1656         cmd_ex.max_recv_wr = cmd.max_recv_wr;
1657         cmd_ex.max_send_sge = cmd.max_send_sge;
1658         cmd_ex.max_recv_sge = cmd.max_recv_sge;
1659         cmd_ex.max_inline_data = cmd.max_inline_data;
1660         cmd_ex.sq_sig_all = cmd.sq_sig_all;
1661         cmd_ex.qp_type = cmd.qp_type;
1662         cmd_ex.is_srq = cmd.is_srq;
1663
1664         err = create_qp(file, &ucore, &uhw, &cmd_ex,
1665                         offsetof(typeof(cmd_ex), is_srq) +
1666                         sizeof(cmd.is_srq), ib_uverbs_create_qp_cb,
1667                         NULL);
1668
1669         if (err)
1670                 return err;
1671
1672         return in_len;
1673 }
1674
1675 static int ib_uverbs_ex_create_qp_cb(struct ib_uverbs_file *file,
1676                                      struct ib_uverbs_ex_create_qp_resp *resp,
1677                                      struct ib_udata *ucore)
1678 {
1679         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1680                 return -EFAULT;
1681
1682         return 0;
1683 }
1684
1685 int ib_uverbs_ex_create_qp(struct ib_uverbs_file *file,
1686                            struct ib_udata *ucore,
1687                            struct ib_udata *uhw)
1688 {
1689         struct ib_uverbs_ex_create_qp_resp resp;
1690         struct ib_uverbs_ex_create_qp cmd = {0};
1691         int err;
1692
1693         if (ucore->inlen < (offsetof(typeof(cmd), comp_mask) +
1694                             sizeof(cmd.comp_mask)))
1695                 return -EINVAL;
1696
1697         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
1698         if (err)
1699                 return err;
1700
1701         if (cmd.comp_mask & ~IB_UVERBS_CREATE_QP_SUP_COMP_MASK)
1702                 return -EINVAL;
1703
1704         if (cmd.reserved)
1705                 return -EINVAL;
1706
1707         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1708                              sizeof(resp.response_length)))
1709                 return -ENOSPC;
1710
1711         err = create_qp(file, ucore, uhw, &cmd,
1712                         min(ucore->inlen, sizeof(cmd)),
1713                         ib_uverbs_ex_create_qp_cb, NULL);
1714
1715         if (err)
1716                 return err;
1717
1718         return 0;
1719 }
1720
1721 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1722                           const char __user *buf, int in_len, int out_len)
1723 {
1724         struct ib_uverbs_open_qp        cmd;
1725         struct ib_uverbs_create_qp_resp resp;
1726         struct ib_udata                 udata;
1727         struct ib_uqp_object           *obj;
1728         struct ib_xrcd                 *xrcd;
1729         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1730         struct ib_qp                   *qp;
1731         struct ib_qp_open_attr          attr;
1732         int ret;
1733         struct ib_device *ib_dev;
1734
1735         if (out_len < sizeof resp)
1736                 return -ENOSPC;
1737
1738         if (copy_from_user(&cmd, buf, sizeof cmd))
1739                 return -EFAULT;
1740
1741         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
1742                    u64_to_user_ptr(cmd.response) + sizeof(resp),
1743                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1744                    out_len - sizeof(resp));
1745
1746         obj = (struct ib_uqp_object *)uobj_alloc(UVERBS_OBJECT_QP, file,
1747                                                  &ib_dev);
1748         if (IS_ERR(obj))
1749                 return PTR_ERR(obj);
1750
1751         xrcd_uobj = uobj_get_read(UVERBS_OBJECT_XRCD, cmd.pd_handle, file);
1752         if (IS_ERR(xrcd_uobj)) {
1753                 ret = -EINVAL;
1754                 goto err_put;
1755         }
1756
1757         xrcd = (struct ib_xrcd *)xrcd_uobj->object;
1758         if (!xrcd) {
1759                 ret = -EINVAL;
1760                 goto err_xrcd;
1761         }
1762
1763         attr.event_handler = ib_uverbs_qp_event_handler;
1764         attr.qp_context    = file;
1765         attr.qp_num        = cmd.qpn;
1766         attr.qp_type       = cmd.qp_type;
1767
1768         obj->uevent.events_reported = 0;
1769         INIT_LIST_HEAD(&obj->uevent.event_list);
1770         INIT_LIST_HEAD(&obj->mcast_list);
1771
1772         qp = ib_open_qp(xrcd, &attr);
1773         if (IS_ERR(qp)) {
1774                 ret = PTR_ERR(qp);
1775                 goto err_xrcd;
1776         }
1777
1778         obj->uevent.uobject.object = qp;
1779         obj->uevent.uobject.user_handle = cmd.user_handle;
1780
1781         memset(&resp, 0, sizeof resp);
1782         resp.qpn       = qp->qp_num;
1783         resp.qp_handle = obj->uevent.uobject.id;
1784
1785         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
1786                 ret = -EFAULT;
1787                 goto err_destroy;
1788         }
1789
1790         obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
1791         atomic_inc(&obj->uxrcd->refcnt);
1792         qp->uobject = &obj->uevent.uobject;
1793         uobj_put_read(xrcd_uobj);
1794
1795         return uobj_alloc_commit(&obj->uevent.uobject, in_len);
1796
1797 err_destroy:
1798         ib_destroy_qp(qp);
1799 err_xrcd:
1800         uobj_put_read(xrcd_uobj);
1801 err_put:
1802         uobj_alloc_abort(&obj->uevent.uobject);
1803         return ret;
1804 }
1805
1806 static void copy_ah_attr_to_uverbs(struct ib_uverbs_qp_dest *uverb_attr,
1807                                    struct rdma_ah_attr *rdma_attr)
1808 {
1809         const struct ib_global_route   *grh;
1810
1811         uverb_attr->dlid              = rdma_ah_get_dlid(rdma_attr);
1812         uverb_attr->sl                = rdma_ah_get_sl(rdma_attr);
1813         uverb_attr->src_path_bits     = rdma_ah_get_path_bits(rdma_attr);
1814         uverb_attr->static_rate       = rdma_ah_get_static_rate(rdma_attr);
1815         uverb_attr->is_global         = !!(rdma_ah_get_ah_flags(rdma_attr) &
1816                                          IB_AH_GRH);
1817         if (uverb_attr->is_global) {
1818                 grh = rdma_ah_read_grh(rdma_attr);
1819                 memcpy(uverb_attr->dgid, grh->dgid.raw, 16);
1820                 uverb_attr->flow_label        = grh->flow_label;
1821                 uverb_attr->sgid_index        = grh->sgid_index;
1822                 uverb_attr->hop_limit         = grh->hop_limit;
1823                 uverb_attr->traffic_class     = grh->traffic_class;
1824         }
1825         uverb_attr->port_num          = rdma_ah_get_port_num(rdma_attr);
1826 }
1827
1828 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1829                            const char __user *buf, int in_len,
1830                            int out_len)
1831 {
1832         struct ib_uverbs_query_qp      cmd;
1833         struct ib_uverbs_query_qp_resp resp;
1834         struct ib_qp                   *qp;
1835         struct ib_qp_attr              *attr;
1836         struct ib_qp_init_attr         *init_attr;
1837         int                            ret;
1838
1839         if (copy_from_user(&cmd, buf, sizeof cmd))
1840                 return -EFAULT;
1841
1842         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1843         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1844         if (!attr || !init_attr) {
1845                 ret = -ENOMEM;
1846                 goto out;
1847         }
1848
1849         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
1850         if (!qp) {
1851                 ret = -EINVAL;
1852                 goto out;
1853         }
1854
1855         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1856
1857         uobj_put_obj_read(qp);
1858
1859         if (ret)
1860                 goto out;
1861
1862         memset(&resp, 0, sizeof resp);
1863
1864         resp.qp_state               = attr->qp_state;
1865         resp.cur_qp_state           = attr->cur_qp_state;
1866         resp.path_mtu               = attr->path_mtu;
1867         resp.path_mig_state         = attr->path_mig_state;
1868         resp.qkey                   = attr->qkey;
1869         resp.rq_psn                 = attr->rq_psn;
1870         resp.sq_psn                 = attr->sq_psn;
1871         resp.dest_qp_num            = attr->dest_qp_num;
1872         resp.qp_access_flags        = attr->qp_access_flags;
1873         resp.pkey_index             = attr->pkey_index;
1874         resp.alt_pkey_index         = attr->alt_pkey_index;
1875         resp.sq_draining            = attr->sq_draining;
1876         resp.max_rd_atomic          = attr->max_rd_atomic;
1877         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1878         resp.min_rnr_timer          = attr->min_rnr_timer;
1879         resp.port_num               = attr->port_num;
1880         resp.timeout                = attr->timeout;
1881         resp.retry_cnt              = attr->retry_cnt;
1882         resp.rnr_retry              = attr->rnr_retry;
1883         resp.alt_port_num           = attr->alt_port_num;
1884         resp.alt_timeout            = attr->alt_timeout;
1885
1886         copy_ah_attr_to_uverbs(&resp.dest, &attr->ah_attr);
1887         copy_ah_attr_to_uverbs(&resp.alt_dest, &attr->alt_ah_attr);
1888
1889         resp.max_send_wr            = init_attr->cap.max_send_wr;
1890         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1891         resp.max_send_sge           = init_attr->cap.max_send_sge;
1892         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1893         resp.max_inline_data        = init_attr->cap.max_inline_data;
1894         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1895
1896         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
1897                 ret = -EFAULT;
1898
1899 out:
1900         kfree(attr);
1901         kfree(init_attr);
1902
1903         return ret ? ret : in_len;
1904 }
1905
1906 /* Remove ignored fields set in the attribute mask */
1907 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1908 {
1909         switch (qp_type) {
1910         case IB_QPT_XRC_INI:
1911                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1912         case IB_QPT_XRC_TGT:
1913                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1914                                 IB_QP_RNR_RETRY);
1915         default:
1916                 return mask;
1917         }
1918 }
1919
1920 static void copy_ah_attr_from_uverbs(struct ib_device *dev,
1921                                      struct rdma_ah_attr *rdma_attr,
1922                                      struct ib_uverbs_qp_dest *uverb_attr)
1923 {
1924         rdma_attr->type = rdma_ah_find_type(dev, uverb_attr->port_num);
1925         if (uverb_attr->is_global) {
1926                 rdma_ah_set_grh(rdma_attr, NULL,
1927                                 uverb_attr->flow_label,
1928                                 uverb_attr->sgid_index,
1929                                 uverb_attr->hop_limit,
1930                                 uverb_attr->traffic_class);
1931                 rdma_ah_set_dgid_raw(rdma_attr, uverb_attr->dgid);
1932         } else {
1933                 rdma_ah_set_ah_flags(rdma_attr, 0);
1934         }
1935         rdma_ah_set_dlid(rdma_attr, uverb_attr->dlid);
1936         rdma_ah_set_sl(rdma_attr, uverb_attr->sl);
1937         rdma_ah_set_path_bits(rdma_attr, uverb_attr->src_path_bits);
1938         rdma_ah_set_static_rate(rdma_attr, uverb_attr->static_rate);
1939         rdma_ah_set_port_num(rdma_attr, uverb_attr->port_num);
1940         rdma_ah_set_make_grd(rdma_attr, false);
1941 }
1942
1943 static int modify_qp(struct ib_uverbs_file *file,
1944                      struct ib_uverbs_ex_modify_qp *cmd, struct ib_udata *udata)
1945 {
1946         struct ib_qp_attr *attr;
1947         struct ib_qp *qp;
1948         int ret;
1949
1950         attr = kzalloc(sizeof(*attr), GFP_KERNEL);
1951         if (!attr)
1952                 return -ENOMEM;
1953
1954         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd->base.qp_handle, file);
1955         if (!qp) {
1956                 ret = -EINVAL;
1957                 goto out;
1958         }
1959
1960         if ((cmd->base.attr_mask & IB_QP_PORT) &&
1961             !rdma_is_port_valid(qp->device, cmd->base.port_num)) {
1962                 ret = -EINVAL;
1963                 goto release_qp;
1964         }
1965
1966         if ((cmd->base.attr_mask & IB_QP_AV)) {
1967                 if (!rdma_is_port_valid(qp->device, cmd->base.dest.port_num)) {
1968                         ret = -EINVAL;
1969                         goto release_qp;
1970                 }
1971
1972                 if (cmd->base.attr_mask & IB_QP_STATE &&
1973                     cmd->base.qp_state == IB_QPS_RTR) {
1974                 /* We are in INIT->RTR TRANSITION (if we are not,
1975                  * this transition will be rejected in subsequent checks).
1976                  * In the INIT->RTR transition, we cannot have IB_QP_PORT set,
1977                  * but the IB_QP_STATE flag is required.
1978                  *
1979                  * Since kernel 3.14 (commit dbf727de7440), the uverbs driver,
1980                  * when IB_QP_AV is set, has required inclusion of a valid
1981                  * port number in the primary AV. (AVs are created and handled
1982                  * differently for infiniband and ethernet (RoCE) ports).
1983                  *
1984                  * Check the port number included in the primary AV against
1985                  * the port number in the qp struct, which was set (and saved)
1986                  * in the RST->INIT transition.
1987                  */
1988                         if (cmd->base.dest.port_num != qp->real_qp->port) {
1989                                 ret = -EINVAL;
1990                                 goto release_qp;
1991                         }
1992                 } else {
1993                 /* We are in SQD->SQD. (If we are not, this transition will
1994                  * be rejected later in the verbs layer checks).
1995                  * Check for both IB_QP_PORT and IB_QP_AV, these can be set
1996                  * together in the SQD->SQD transition.
1997                  *
1998                  * If only IP_QP_AV was set, add in IB_QP_PORT as well (the
1999                  * verbs layer driver does not track primary port changes
2000                  * resulting from path migration. Thus, in SQD, if the primary
2001                  * AV is modified, the primary port should also be modified).
2002                  *
2003                  * Note that in this transition, the IB_QP_STATE flag
2004                  * is not allowed.
2005                  */
2006                         if (((cmd->base.attr_mask & (IB_QP_AV | IB_QP_PORT))
2007                              == (IB_QP_AV | IB_QP_PORT)) &&
2008                             cmd->base.port_num != cmd->base.dest.port_num) {
2009                                 ret = -EINVAL;
2010                                 goto release_qp;
2011                         }
2012                         if ((cmd->base.attr_mask & (IB_QP_AV | IB_QP_PORT))
2013                             == IB_QP_AV) {
2014                                 cmd->base.attr_mask |= IB_QP_PORT;
2015                                 cmd->base.port_num = cmd->base.dest.port_num;
2016                         }
2017                 }
2018         }
2019
2020         if ((cmd->base.attr_mask & IB_QP_ALT_PATH) &&
2021             (!rdma_is_port_valid(qp->device, cmd->base.alt_port_num) ||
2022             !rdma_is_port_valid(qp->device, cmd->base.alt_dest.port_num) ||
2023             cmd->base.alt_port_num != cmd->base.alt_dest.port_num)) {
2024                 ret = -EINVAL;
2025                 goto release_qp;
2026         }
2027
2028         if ((cmd->base.attr_mask & IB_QP_CUR_STATE &&
2029             cmd->base.cur_qp_state > IB_QPS_ERR) ||
2030             (cmd->base.attr_mask & IB_QP_STATE &&
2031             cmd->base.qp_state > IB_QPS_ERR)) {
2032                 ret = -EINVAL;
2033                 goto release_qp;
2034         }
2035
2036         if (cmd->base.attr_mask & IB_QP_STATE)
2037                 attr->qp_state = cmd->base.qp_state;
2038         if (cmd->base.attr_mask & IB_QP_CUR_STATE)
2039                 attr->cur_qp_state = cmd->base.cur_qp_state;
2040         if (cmd->base.attr_mask & IB_QP_PATH_MTU)
2041                 attr->path_mtu = cmd->base.path_mtu;
2042         if (cmd->base.attr_mask & IB_QP_PATH_MIG_STATE)
2043                 attr->path_mig_state = cmd->base.path_mig_state;
2044         if (cmd->base.attr_mask & IB_QP_QKEY)
2045                 attr->qkey = cmd->base.qkey;
2046         if (cmd->base.attr_mask & IB_QP_RQ_PSN)
2047                 attr->rq_psn = cmd->base.rq_psn;
2048         if (cmd->base.attr_mask & IB_QP_SQ_PSN)
2049                 attr->sq_psn = cmd->base.sq_psn;
2050         if (cmd->base.attr_mask & IB_QP_DEST_QPN)
2051                 attr->dest_qp_num = cmd->base.dest_qp_num;
2052         if (cmd->base.attr_mask & IB_QP_ACCESS_FLAGS)
2053                 attr->qp_access_flags = cmd->base.qp_access_flags;
2054         if (cmd->base.attr_mask & IB_QP_PKEY_INDEX)
2055                 attr->pkey_index = cmd->base.pkey_index;
2056         if (cmd->base.attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY)
2057                 attr->en_sqd_async_notify = cmd->base.en_sqd_async_notify;
2058         if (cmd->base.attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
2059                 attr->max_rd_atomic = cmd->base.max_rd_atomic;
2060         if (cmd->base.attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2061                 attr->max_dest_rd_atomic = cmd->base.max_dest_rd_atomic;
2062         if (cmd->base.attr_mask & IB_QP_MIN_RNR_TIMER)
2063                 attr->min_rnr_timer = cmd->base.min_rnr_timer;
2064         if (cmd->base.attr_mask & IB_QP_PORT)
2065                 attr->port_num = cmd->base.port_num;
2066         if (cmd->base.attr_mask & IB_QP_TIMEOUT)
2067                 attr->timeout = cmd->base.timeout;
2068         if (cmd->base.attr_mask & IB_QP_RETRY_CNT)
2069                 attr->retry_cnt = cmd->base.retry_cnt;
2070         if (cmd->base.attr_mask & IB_QP_RNR_RETRY)
2071                 attr->rnr_retry = cmd->base.rnr_retry;
2072         if (cmd->base.attr_mask & IB_QP_ALT_PATH) {
2073                 attr->alt_port_num = cmd->base.alt_port_num;
2074                 attr->alt_timeout = cmd->base.alt_timeout;
2075                 attr->alt_pkey_index = cmd->base.alt_pkey_index;
2076         }
2077         if (cmd->base.attr_mask & IB_QP_RATE_LIMIT)
2078                 attr->rate_limit = cmd->rate_limit;
2079
2080         if (cmd->base.attr_mask & IB_QP_AV)
2081                 copy_ah_attr_from_uverbs(qp->device, &attr->ah_attr,
2082                                          &cmd->base.dest);
2083
2084         if (cmd->base.attr_mask & IB_QP_ALT_PATH)
2085                 copy_ah_attr_from_uverbs(qp->device, &attr->alt_ah_attr,
2086                                          &cmd->base.alt_dest);
2087
2088         ret = ib_modify_qp_with_udata(qp, attr,
2089                                       modify_qp_mask(qp->qp_type,
2090                                                      cmd->base.attr_mask),
2091                                       udata);
2092
2093 release_qp:
2094         uobj_put_obj_read(qp);
2095 out:
2096         kfree(attr);
2097
2098         return ret;
2099 }
2100
2101 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
2102                             const char __user *buf, int in_len,
2103                             int out_len)
2104 {
2105         struct ib_uverbs_ex_modify_qp cmd = {};
2106         struct ib_udata udata;
2107         int ret;
2108
2109         if (copy_from_user(&cmd.base, buf, sizeof(cmd.base)))
2110                 return -EFAULT;
2111
2112         if (cmd.base.attr_mask &
2113             ~((IB_USER_LEGACY_LAST_QP_ATTR_MASK << 1) - 1))
2114                 return -EOPNOTSUPP;
2115
2116         ib_uverbs_init_udata(&udata, buf + sizeof(cmd.base), NULL,
2117                    in_len - sizeof(cmd.base) - sizeof(struct ib_uverbs_cmd_hdr),
2118                    out_len);
2119
2120         ret = modify_qp(file, &cmd, &udata);
2121         if (ret)
2122                 return ret;
2123
2124         return in_len;
2125 }
2126
2127 int ib_uverbs_ex_modify_qp(struct ib_uverbs_file *file,
2128                            struct ib_udata *ucore,
2129                            struct ib_udata *uhw)
2130 {
2131         struct ib_uverbs_ex_modify_qp cmd = {};
2132         int ret;
2133
2134         /*
2135          * Last bit is reserved for extending the attr_mask by
2136          * using another field.
2137          */
2138         BUILD_BUG_ON(IB_USER_LAST_QP_ATTR_MASK == (1 << 31));
2139
2140         if (ucore->inlen < sizeof(cmd.base))
2141                 return -EINVAL;
2142
2143         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2144         if (ret)
2145                 return ret;
2146
2147         if (cmd.base.attr_mask &
2148             ~((IB_USER_LAST_QP_ATTR_MASK << 1) - 1))
2149                 return -EOPNOTSUPP;
2150
2151         if (ucore->inlen > sizeof(cmd)) {
2152                 if (!ib_is_udata_cleared(ucore, sizeof(cmd),
2153                                          ucore->inlen - sizeof(cmd)))
2154                         return -EOPNOTSUPP;
2155         }
2156
2157         ret = modify_qp(file, &cmd, uhw);
2158
2159         return ret;
2160 }
2161
2162 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
2163                              const char __user *buf, int in_len,
2164                              int out_len)
2165 {
2166         struct ib_uverbs_destroy_qp      cmd;
2167         struct ib_uverbs_destroy_qp_resp resp;
2168         struct ib_uobject               *uobj;
2169         struct ib_uqp_object            *obj;
2170
2171         if (copy_from_user(&cmd, buf, sizeof cmd))
2172                 return -EFAULT;
2173
2174         uobj = uobj_get_destroy(UVERBS_OBJECT_QP, cmd.qp_handle, file);
2175         if (IS_ERR(uobj))
2176                 return PTR_ERR(uobj);
2177
2178         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2179         memset(&resp, 0, sizeof(resp));
2180         resp.events_reported = obj->uevent.events_reported;
2181
2182         uobj_put_destroy(uobj);
2183
2184         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2185                 return -EFAULT;
2186
2187         return in_len;
2188 }
2189
2190 static void *alloc_wr(size_t wr_size, __u32 num_sge)
2191 {
2192         if (num_sge >= (U32_MAX - ALIGN(wr_size, sizeof (struct ib_sge))) /
2193                        sizeof (struct ib_sge))
2194                 return NULL;
2195
2196         return kmalloc(ALIGN(wr_size, sizeof (struct ib_sge)) +
2197                          num_sge * sizeof (struct ib_sge), GFP_KERNEL);
2198 }
2199
2200 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2201                             const char __user *buf, int in_len,
2202                             int out_len)
2203 {
2204         struct ib_uverbs_post_send      cmd;
2205         struct ib_uverbs_post_send_resp resp;
2206         struct ib_uverbs_send_wr       *user_wr;
2207         struct ib_send_wr              *wr = NULL, *last, *next;
2208         const struct ib_send_wr        *bad_wr;
2209         struct ib_qp                   *qp;
2210         int                             i, sg_ind;
2211         int                             is_ud;
2212         ssize_t                         ret = -EINVAL;
2213         size_t                          next_size;
2214
2215         if (copy_from_user(&cmd, buf, sizeof cmd))
2216                 return -EFAULT;
2217
2218         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2219             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2220                 return -EINVAL;
2221
2222         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2223                 return -EINVAL;
2224
2225         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2226         if (!user_wr)
2227                 return -ENOMEM;
2228
2229         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2230         if (!qp)
2231                 goto out;
2232
2233         is_ud = qp->qp_type == IB_QPT_UD;
2234         sg_ind = 0;
2235         last = NULL;
2236         for (i = 0; i < cmd.wr_count; ++i) {
2237                 if (copy_from_user(user_wr,
2238                                    buf + sizeof cmd + i * cmd.wqe_size,
2239                                    cmd.wqe_size)) {
2240                         ret = -EFAULT;
2241                         goto out_put;
2242                 }
2243
2244                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2245                         ret = -EINVAL;
2246                         goto out_put;
2247                 }
2248
2249                 if (is_ud) {
2250                         struct ib_ud_wr *ud;
2251
2252                         if (user_wr->opcode != IB_WR_SEND &&
2253                             user_wr->opcode != IB_WR_SEND_WITH_IMM) {
2254                                 ret = -EINVAL;
2255                                 goto out_put;
2256                         }
2257
2258                         next_size = sizeof(*ud);
2259                         ud = alloc_wr(next_size, user_wr->num_sge);
2260                         if (!ud) {
2261                                 ret = -ENOMEM;
2262                                 goto out_put;
2263                         }
2264
2265                         ud->ah = uobj_get_obj_read(ah, UVERBS_OBJECT_AH,
2266                                                    user_wr->wr.ud.ah, file);
2267                         if (!ud->ah) {
2268                                 kfree(ud);
2269                                 ret = -EINVAL;
2270                                 goto out_put;
2271                         }
2272                         ud->remote_qpn = user_wr->wr.ud.remote_qpn;
2273                         ud->remote_qkey = user_wr->wr.ud.remote_qkey;
2274
2275                         next = &ud->wr;
2276                 } else if (user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
2277                            user_wr->opcode == IB_WR_RDMA_WRITE ||
2278                            user_wr->opcode == IB_WR_RDMA_READ) {
2279                         struct ib_rdma_wr *rdma;
2280
2281                         next_size = sizeof(*rdma);
2282                         rdma = alloc_wr(next_size, user_wr->num_sge);
2283                         if (!rdma) {
2284                                 ret = -ENOMEM;
2285                                 goto out_put;
2286                         }
2287
2288                         rdma->remote_addr = user_wr->wr.rdma.remote_addr;
2289                         rdma->rkey = user_wr->wr.rdma.rkey;
2290
2291                         next = &rdma->wr;
2292                 } else if (user_wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2293                            user_wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2294                         struct ib_atomic_wr *atomic;
2295
2296                         next_size = sizeof(*atomic);
2297                         atomic = alloc_wr(next_size, user_wr->num_sge);
2298                         if (!atomic) {
2299                                 ret = -ENOMEM;
2300                                 goto out_put;
2301                         }
2302
2303                         atomic->remote_addr = user_wr->wr.atomic.remote_addr;
2304                         atomic->compare_add = user_wr->wr.atomic.compare_add;
2305                         atomic->swap = user_wr->wr.atomic.swap;
2306                         atomic->rkey = user_wr->wr.atomic.rkey;
2307
2308                         next = &atomic->wr;
2309                 } else if (user_wr->opcode == IB_WR_SEND ||
2310                            user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2311                            user_wr->opcode == IB_WR_SEND_WITH_INV) {
2312                         next_size = sizeof(*next);
2313                         next = alloc_wr(next_size, user_wr->num_sge);
2314                         if (!next) {
2315                                 ret = -ENOMEM;
2316                                 goto out_put;
2317                         }
2318                 } else {
2319                         ret = -EINVAL;
2320                         goto out_put;
2321                 }
2322
2323                 if (user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2324                     user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
2325                         next->ex.imm_data =
2326                                         (__be32 __force) user_wr->ex.imm_data;
2327                 } else if (user_wr->opcode == IB_WR_SEND_WITH_INV) {
2328                         next->ex.invalidate_rkey = user_wr->ex.invalidate_rkey;
2329                 }
2330
2331                 if (!last)
2332                         wr = next;
2333                 else
2334                         last->next = next;
2335                 last = next;
2336
2337                 next->next       = NULL;
2338                 next->wr_id      = user_wr->wr_id;
2339                 next->num_sge    = user_wr->num_sge;
2340                 next->opcode     = user_wr->opcode;
2341                 next->send_flags = user_wr->send_flags;
2342
2343                 if (next->num_sge) {
2344                         next->sg_list = (void *) next +
2345                                 ALIGN(next_size, sizeof(struct ib_sge));
2346                         if (copy_from_user(next->sg_list,
2347                                            buf + sizeof cmd +
2348                                            cmd.wr_count * cmd.wqe_size +
2349                                            sg_ind * sizeof (struct ib_sge),
2350                                            next->num_sge * sizeof (struct ib_sge))) {
2351                                 ret = -EFAULT;
2352                                 goto out_put;
2353                         }
2354                         sg_ind += next->num_sge;
2355                 } else
2356                         next->sg_list = NULL;
2357         }
2358
2359         resp.bad_wr = 0;
2360         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2361         if (ret)
2362                 for (next = wr; next; next = next->next) {
2363                         ++resp.bad_wr;
2364                         if (next == bad_wr)
2365                                 break;
2366                 }
2367
2368         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2369                 ret = -EFAULT;
2370
2371 out_put:
2372         uobj_put_obj_read(qp);
2373
2374         while (wr) {
2375                 if (is_ud && ud_wr(wr)->ah)
2376                         uobj_put_obj_read(ud_wr(wr)->ah);
2377                 next = wr->next;
2378                 kfree(wr);
2379                 wr = next;
2380         }
2381
2382 out:
2383         kfree(user_wr);
2384
2385         return ret ? ret : in_len;
2386 }
2387
2388 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2389                                                     int in_len,
2390                                                     u32 wr_count,
2391                                                     u32 sge_count,
2392                                                     u32 wqe_size)
2393 {
2394         struct ib_uverbs_recv_wr *user_wr;
2395         struct ib_recv_wr        *wr = NULL, *last, *next;
2396         int                       sg_ind;
2397         int                       i;
2398         int                       ret;
2399
2400         if (in_len < wqe_size * wr_count +
2401             sge_count * sizeof (struct ib_uverbs_sge))
2402                 return ERR_PTR(-EINVAL);
2403
2404         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2405                 return ERR_PTR(-EINVAL);
2406
2407         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2408         if (!user_wr)
2409                 return ERR_PTR(-ENOMEM);
2410
2411         sg_ind = 0;
2412         last = NULL;
2413         for (i = 0; i < wr_count; ++i) {
2414                 if (copy_from_user(user_wr, buf + i * wqe_size,
2415                                    wqe_size)) {
2416                         ret = -EFAULT;
2417                         goto err;
2418                 }
2419
2420                 if (user_wr->num_sge + sg_ind > sge_count) {
2421                         ret = -EINVAL;
2422                         goto err;
2423                 }
2424
2425                 if (user_wr->num_sge >=
2426                     (U32_MAX - ALIGN(sizeof *next, sizeof (struct ib_sge))) /
2427                     sizeof (struct ib_sge)) {
2428                         ret = -EINVAL;
2429                         goto err;
2430                 }
2431
2432                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2433                                user_wr->num_sge * sizeof (struct ib_sge),
2434                                GFP_KERNEL);
2435                 if (!next) {
2436                         ret = -ENOMEM;
2437                         goto err;
2438                 }
2439
2440                 if (!last)
2441                         wr = next;
2442                 else
2443                         last->next = next;
2444                 last = next;
2445
2446                 next->next       = NULL;
2447                 next->wr_id      = user_wr->wr_id;
2448                 next->num_sge    = user_wr->num_sge;
2449
2450                 if (next->num_sge) {
2451                         next->sg_list = (void *) next +
2452                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2453                         if (copy_from_user(next->sg_list,
2454                                            buf + wr_count * wqe_size +
2455                                            sg_ind * sizeof (struct ib_sge),
2456                                            next->num_sge * sizeof (struct ib_sge))) {
2457                                 ret = -EFAULT;
2458                                 goto err;
2459                         }
2460                         sg_ind += next->num_sge;
2461                 } else
2462                         next->sg_list = NULL;
2463         }
2464
2465         kfree(user_wr);
2466         return wr;
2467
2468 err:
2469         kfree(user_wr);
2470
2471         while (wr) {
2472                 next = wr->next;
2473                 kfree(wr);
2474                 wr = next;
2475         }
2476
2477         return ERR_PTR(ret);
2478 }
2479
2480 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2481                             const char __user *buf, int in_len,
2482                             int out_len)
2483 {
2484         struct ib_uverbs_post_recv      cmd;
2485         struct ib_uverbs_post_recv_resp resp;
2486         struct ib_recv_wr              *wr, *next;
2487         const struct ib_recv_wr        *bad_wr;
2488         struct ib_qp                   *qp;
2489         ssize_t                         ret = -EINVAL;
2490
2491         if (copy_from_user(&cmd, buf, sizeof cmd))
2492                 return -EFAULT;
2493
2494         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2495                                        in_len - sizeof cmd, cmd.wr_count,
2496                                        cmd.sge_count, cmd.wqe_size);
2497         if (IS_ERR(wr))
2498                 return PTR_ERR(wr);
2499
2500         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2501         if (!qp)
2502                 goto out;
2503
2504         resp.bad_wr = 0;
2505         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2506
2507         uobj_put_obj_read(qp);
2508         if (ret) {
2509                 for (next = wr; next; next = next->next) {
2510                         ++resp.bad_wr;
2511                         if (next == bad_wr)
2512                                 break;
2513                 }
2514         }
2515
2516         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2517                 ret = -EFAULT;
2518
2519 out:
2520         while (wr) {
2521                 next = wr->next;
2522                 kfree(wr);
2523                 wr = next;
2524         }
2525
2526         return ret ? ret : in_len;
2527 }
2528
2529 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2530                                 const char __user *buf, int in_len,
2531                                 int out_len)
2532 {
2533         struct ib_uverbs_post_srq_recv      cmd;
2534         struct ib_uverbs_post_srq_recv_resp resp;
2535         struct ib_recv_wr                  *wr, *next;
2536         const struct ib_recv_wr            *bad_wr;
2537         struct ib_srq                      *srq;
2538         ssize_t                             ret = -EINVAL;
2539
2540         if (copy_from_user(&cmd, buf, sizeof cmd))
2541                 return -EFAULT;
2542
2543         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2544                                        in_len - sizeof cmd, cmd.wr_count,
2545                                        cmd.sge_count, cmd.wqe_size);
2546         if (IS_ERR(wr))
2547                 return PTR_ERR(wr);
2548
2549         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
2550         if (!srq)
2551                 goto out;
2552
2553         resp.bad_wr = 0;
2554         ret = srq->device->post_srq_recv ?
2555                 srq->device->post_srq_recv(srq, wr, &bad_wr) : -EOPNOTSUPP;
2556
2557         uobj_put_obj_read(srq);
2558
2559         if (ret)
2560                 for (next = wr; next; next = next->next) {
2561                         ++resp.bad_wr;
2562                         if (next == bad_wr)
2563                                 break;
2564                 }
2565
2566         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2567                 ret = -EFAULT;
2568
2569 out:
2570         while (wr) {
2571                 next = wr->next;
2572                 kfree(wr);
2573                 wr = next;
2574         }
2575
2576         return ret ? ret : in_len;
2577 }
2578
2579 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2580                             const char __user *buf, int in_len,
2581                             int out_len)
2582 {
2583         struct ib_uverbs_create_ah       cmd;
2584         struct ib_uverbs_create_ah_resp  resp;
2585         struct ib_uobject               *uobj;
2586         struct ib_pd                    *pd;
2587         struct ib_ah                    *ah;
2588         struct rdma_ah_attr             attr = {};
2589         int ret;
2590         struct ib_udata                   udata;
2591         struct ib_device *ib_dev;
2592
2593         if (out_len < sizeof resp)
2594                 return -ENOSPC;
2595
2596         if (copy_from_user(&cmd, buf, sizeof cmd))
2597                 return -EFAULT;
2598
2599         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
2600                    u64_to_user_ptr(cmd.response) + sizeof(resp),
2601                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
2602                    out_len - sizeof(resp));
2603
2604         uobj = uobj_alloc(UVERBS_OBJECT_AH, file, &ib_dev);
2605         if (IS_ERR(uobj))
2606                 return PTR_ERR(uobj);
2607
2608         if (!rdma_is_port_valid(ib_dev, cmd.attr.port_num)) {
2609                 ret = -EINVAL;
2610                 goto err;
2611         }
2612
2613         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file);
2614         if (!pd) {
2615                 ret = -EINVAL;
2616                 goto err;
2617         }
2618
2619         attr.type = rdma_ah_find_type(ib_dev, cmd.attr.port_num);
2620         rdma_ah_set_make_grd(&attr, false);
2621         rdma_ah_set_dlid(&attr, cmd.attr.dlid);
2622         rdma_ah_set_sl(&attr, cmd.attr.sl);
2623         rdma_ah_set_path_bits(&attr, cmd.attr.src_path_bits);
2624         rdma_ah_set_static_rate(&attr, cmd.attr.static_rate);
2625         rdma_ah_set_port_num(&attr, cmd.attr.port_num);
2626
2627         if (cmd.attr.is_global) {
2628                 rdma_ah_set_grh(&attr, NULL, cmd.attr.grh.flow_label,
2629                                 cmd.attr.grh.sgid_index,
2630                                 cmd.attr.grh.hop_limit,
2631                                 cmd.attr.grh.traffic_class);
2632                 rdma_ah_set_dgid_raw(&attr, cmd.attr.grh.dgid);
2633         } else {
2634                 rdma_ah_set_ah_flags(&attr, 0);
2635         }
2636
2637         ah = rdma_create_user_ah(pd, &attr, &udata);
2638         if (IS_ERR(ah)) {
2639                 ret = PTR_ERR(ah);
2640                 goto err_put;
2641         }
2642
2643         ah->uobject  = uobj;
2644         uobj->user_handle = cmd.user_handle;
2645         uobj->object = ah;
2646
2647         resp.ah_handle = uobj->id;
2648
2649         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
2650                 ret = -EFAULT;
2651                 goto err_copy;
2652         }
2653
2654         uobj_put_obj_read(pd);
2655         return uobj_alloc_commit(uobj, in_len);
2656
2657 err_copy:
2658         rdma_destroy_ah(ah);
2659
2660 err_put:
2661         uobj_put_obj_read(pd);
2662
2663 err:
2664         uobj_alloc_abort(uobj);
2665         return ret;
2666 }
2667
2668 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2669                              const char __user *buf, int in_len, int out_len)
2670 {
2671         struct ib_uverbs_destroy_ah cmd;
2672
2673         if (copy_from_user(&cmd, buf, sizeof cmd))
2674                 return -EFAULT;
2675
2676         return uobj_perform_destroy(UVERBS_OBJECT_AH, cmd.ah_handle, file,
2677                                     in_len);
2678 }
2679
2680 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2681                                const char __user *buf, int in_len,
2682                                int out_len)
2683 {
2684         struct ib_uverbs_attach_mcast cmd;
2685         struct ib_qp                 *qp;
2686         struct ib_uqp_object         *obj;
2687         struct ib_uverbs_mcast_entry *mcast;
2688         int                           ret;
2689
2690         if (copy_from_user(&cmd, buf, sizeof cmd))
2691                 return -EFAULT;
2692
2693         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2694         if (!qp)
2695                 return -EINVAL;
2696
2697         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2698
2699         mutex_lock(&obj->mcast_lock);
2700         list_for_each_entry(mcast, &obj->mcast_list, list)
2701                 if (cmd.mlid == mcast->lid &&
2702                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2703                         ret = 0;
2704                         goto out_put;
2705                 }
2706
2707         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2708         if (!mcast) {
2709                 ret = -ENOMEM;
2710                 goto out_put;
2711         }
2712
2713         mcast->lid = cmd.mlid;
2714         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2715
2716         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2717         if (!ret)
2718                 list_add_tail(&mcast->list, &obj->mcast_list);
2719         else
2720                 kfree(mcast);
2721
2722 out_put:
2723         mutex_unlock(&obj->mcast_lock);
2724         uobj_put_obj_read(qp);
2725
2726         return ret ? ret : in_len;
2727 }
2728
2729 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2730                                const char __user *buf, int in_len,
2731                                int out_len)
2732 {
2733         struct ib_uverbs_detach_mcast cmd;
2734         struct ib_uqp_object         *obj;
2735         struct ib_qp                 *qp;
2736         struct ib_uverbs_mcast_entry *mcast;
2737         int                           ret = -EINVAL;
2738         bool                          found = false;
2739
2740         if (copy_from_user(&cmd, buf, sizeof cmd))
2741                 return -EFAULT;
2742
2743         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2744         if (!qp)
2745                 return -EINVAL;
2746
2747         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2748         mutex_lock(&obj->mcast_lock);
2749
2750         list_for_each_entry(mcast, &obj->mcast_list, list)
2751                 if (cmd.mlid == mcast->lid &&
2752                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2753                         list_del(&mcast->list);
2754                         kfree(mcast);
2755                         found = true;
2756                         break;
2757                 }
2758
2759         if (!found) {
2760                 ret = -EINVAL;
2761                 goto out_put;
2762         }
2763
2764         ret = ib_detach_mcast(qp, (union ib_gid *)cmd.gid, cmd.mlid);
2765
2766 out_put:
2767         mutex_unlock(&obj->mcast_lock);
2768         uobj_put_obj_read(qp);
2769         return ret ? ret : in_len;
2770 }
2771
2772 struct ib_uflow_resources {
2773         size_t                  max;
2774         size_t                  num;
2775         size_t                  collection_num;
2776         size_t                  counters_num;
2777         struct ib_counters      **counters;
2778         struct ib_flow_action   **collection;
2779 };
2780
2781 static struct ib_uflow_resources *flow_resources_alloc(size_t num_specs)
2782 {
2783         struct ib_uflow_resources *resources;
2784
2785         resources = kzalloc(sizeof(*resources), GFP_KERNEL);
2786
2787         if (!resources)
2788                 return NULL;
2789
2790         if (!num_specs)
2791                 goto out;
2792
2793         resources->counters =
2794                 kcalloc(num_specs, sizeof(*resources->counters), GFP_KERNEL);
2795         resources->collection =
2796                 kcalloc(num_specs, sizeof(*resources->collection), GFP_KERNEL);
2797
2798         if (!resources->counters || !resources->collection)
2799                 goto err;
2800
2801 out:
2802         resources->max = num_specs;
2803         return resources;
2804
2805 err:
2806         kfree(resources->counters);
2807         kfree(resources);
2808
2809         return NULL;
2810 }
2811
2812 void ib_uverbs_flow_resources_free(struct ib_uflow_resources *uflow_res)
2813 {
2814         unsigned int i;
2815
2816         if (!uflow_res)
2817                 return;
2818
2819         for (i = 0; i < uflow_res->collection_num; i++)
2820                 atomic_dec(&uflow_res->collection[i]->usecnt);
2821
2822         for (i = 0; i < uflow_res->counters_num; i++)
2823                 atomic_dec(&uflow_res->counters[i]->usecnt);
2824
2825         kfree(uflow_res->collection);
2826         kfree(uflow_res->counters);
2827         kfree(uflow_res);
2828 }
2829
2830 static void flow_resources_add(struct ib_uflow_resources *uflow_res,
2831                                enum ib_flow_spec_type type,
2832                                void *ibobj)
2833 {
2834         WARN_ON(uflow_res->num >= uflow_res->max);
2835
2836         switch (type) {
2837         case IB_FLOW_SPEC_ACTION_HANDLE:
2838                 atomic_inc(&((struct ib_flow_action *)ibobj)->usecnt);
2839                 uflow_res->collection[uflow_res->collection_num++] =
2840                         (struct ib_flow_action *)ibobj;
2841                 break;
2842         case IB_FLOW_SPEC_ACTION_COUNT:
2843                 atomic_inc(&((struct ib_counters *)ibobj)->usecnt);
2844                 uflow_res->counters[uflow_res->counters_num++] =
2845                         (struct ib_counters *)ibobj;
2846                 break;
2847         default:
2848                 WARN_ON(1);
2849         }
2850
2851         uflow_res->num++;
2852 }
2853
2854 static int kern_spec_to_ib_spec_action(struct ib_uverbs_file *ufile,
2855                                        struct ib_uverbs_flow_spec *kern_spec,
2856                                        union ib_flow_spec *ib_spec,
2857                                        struct ib_uflow_resources *uflow_res)
2858 {
2859         ib_spec->type = kern_spec->type;
2860         switch (ib_spec->type) {
2861         case IB_FLOW_SPEC_ACTION_TAG:
2862                 if (kern_spec->flow_tag.size !=
2863                     sizeof(struct ib_uverbs_flow_spec_action_tag))
2864                         return -EINVAL;
2865
2866                 ib_spec->flow_tag.size = sizeof(struct ib_flow_spec_action_tag);
2867                 ib_spec->flow_tag.tag_id = kern_spec->flow_tag.tag_id;
2868                 break;
2869         case IB_FLOW_SPEC_ACTION_DROP:
2870                 if (kern_spec->drop.size !=
2871                     sizeof(struct ib_uverbs_flow_spec_action_drop))
2872                         return -EINVAL;
2873
2874                 ib_spec->drop.size = sizeof(struct ib_flow_spec_action_drop);
2875                 break;
2876         case IB_FLOW_SPEC_ACTION_HANDLE:
2877                 if (kern_spec->action.size !=
2878                     sizeof(struct ib_uverbs_flow_spec_action_handle))
2879                         return -EOPNOTSUPP;
2880                 ib_spec->action.act = uobj_get_obj_read(flow_action,
2881                                                         UVERBS_OBJECT_FLOW_ACTION,
2882                                                         kern_spec->action.handle,
2883                                                         ufile);
2884                 if (!ib_spec->action.act)
2885                         return -EINVAL;
2886                 ib_spec->action.size =
2887                         sizeof(struct ib_flow_spec_action_handle);
2888                 flow_resources_add(uflow_res,
2889                                    IB_FLOW_SPEC_ACTION_HANDLE,
2890                                    ib_spec->action.act);
2891                 uobj_put_obj_read(ib_spec->action.act);
2892                 break;
2893         case IB_FLOW_SPEC_ACTION_COUNT:
2894                 if (kern_spec->flow_count.size !=
2895                         sizeof(struct ib_uverbs_flow_spec_action_count))
2896                         return -EINVAL;
2897                 ib_spec->flow_count.counters =
2898                         uobj_get_obj_read(counters,
2899                                           UVERBS_OBJECT_COUNTERS,
2900                                           kern_spec->flow_count.handle,
2901                                           ufile);
2902                 if (!ib_spec->flow_count.counters)
2903                         return -EINVAL;
2904                 ib_spec->flow_count.size =
2905                                 sizeof(struct ib_flow_spec_action_count);
2906                 flow_resources_add(uflow_res,
2907                                    IB_FLOW_SPEC_ACTION_COUNT,
2908                                    ib_spec->flow_count.counters);
2909                 uobj_put_obj_read(ib_spec->flow_count.counters);
2910                 break;
2911         default:
2912                 return -EINVAL;
2913         }
2914         return 0;
2915 }
2916
2917 static ssize_t spec_filter_size(const void *kern_spec_filter, u16 kern_filter_size,
2918                                 u16 ib_real_filter_sz)
2919 {
2920         /*
2921          * User space filter structures must be 64 bit aligned, otherwise this
2922          * may pass, but we won't handle additional new attributes.
2923          */
2924
2925         if (kern_filter_size > ib_real_filter_sz) {
2926                 if (memchr_inv(kern_spec_filter +
2927                                ib_real_filter_sz, 0,
2928                                kern_filter_size - ib_real_filter_sz))
2929                         return -EINVAL;
2930                 return ib_real_filter_sz;
2931         }
2932         return kern_filter_size;
2933 }
2934
2935 int ib_uverbs_kern_spec_to_ib_spec_filter(enum ib_flow_spec_type type,
2936                                           const void *kern_spec_mask,
2937                                           const void *kern_spec_val,
2938                                           size_t kern_filter_sz,
2939                                           union ib_flow_spec *ib_spec)
2940 {
2941         ssize_t actual_filter_sz;
2942         ssize_t ib_filter_sz;
2943
2944         /* User flow spec size must be aligned to 4 bytes */
2945         if (kern_filter_sz != ALIGN(kern_filter_sz, 4))
2946                 return -EINVAL;
2947
2948         ib_spec->type = type;
2949
2950         if (ib_spec->type == (IB_FLOW_SPEC_INNER | IB_FLOW_SPEC_VXLAN_TUNNEL))
2951                 return -EINVAL;
2952
2953         switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2954         case IB_FLOW_SPEC_ETH:
2955                 ib_filter_sz = offsetof(struct ib_flow_eth_filter, real_sz);
2956                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2957                                                     kern_filter_sz,
2958                                                     ib_filter_sz);
2959                 if (actual_filter_sz <= 0)
2960                         return -EINVAL;
2961                 ib_spec->size = sizeof(struct ib_flow_spec_eth);
2962                 memcpy(&ib_spec->eth.val, kern_spec_val, actual_filter_sz);
2963                 memcpy(&ib_spec->eth.mask, kern_spec_mask, actual_filter_sz);
2964                 break;
2965         case IB_FLOW_SPEC_IPV4:
2966                 ib_filter_sz = offsetof(struct ib_flow_ipv4_filter, real_sz);
2967                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2968                                                     kern_filter_sz,
2969                                                     ib_filter_sz);
2970                 if (actual_filter_sz <= 0)
2971                         return -EINVAL;
2972                 ib_spec->size = sizeof(struct ib_flow_spec_ipv4);
2973                 memcpy(&ib_spec->ipv4.val, kern_spec_val, actual_filter_sz);
2974                 memcpy(&ib_spec->ipv4.mask, kern_spec_mask, actual_filter_sz);
2975                 break;
2976         case IB_FLOW_SPEC_IPV6:
2977                 ib_filter_sz = offsetof(struct ib_flow_ipv6_filter, real_sz);
2978                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2979                                                     kern_filter_sz,
2980                                                     ib_filter_sz);
2981                 if (actual_filter_sz <= 0)
2982                         return -EINVAL;
2983                 ib_spec->size = sizeof(struct ib_flow_spec_ipv6);
2984                 memcpy(&ib_spec->ipv6.val, kern_spec_val, actual_filter_sz);
2985                 memcpy(&ib_spec->ipv6.mask, kern_spec_mask, actual_filter_sz);
2986
2987                 if ((ntohl(ib_spec->ipv6.mask.flow_label)) >= BIT(20) ||
2988                     (ntohl(ib_spec->ipv6.val.flow_label)) >= BIT(20))
2989                         return -EINVAL;
2990                 break;
2991         case IB_FLOW_SPEC_TCP:
2992         case IB_FLOW_SPEC_UDP:
2993                 ib_filter_sz = offsetof(struct ib_flow_tcp_udp_filter, real_sz);
2994                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2995                                                     kern_filter_sz,
2996                                                     ib_filter_sz);
2997                 if (actual_filter_sz <= 0)
2998                         return -EINVAL;
2999                 ib_spec->size = sizeof(struct ib_flow_spec_tcp_udp);
3000                 memcpy(&ib_spec->tcp_udp.val, kern_spec_val, actual_filter_sz);
3001                 memcpy(&ib_spec->tcp_udp.mask, kern_spec_mask, actual_filter_sz);
3002                 break;
3003         case IB_FLOW_SPEC_VXLAN_TUNNEL:
3004                 ib_filter_sz = offsetof(struct ib_flow_tunnel_filter, real_sz);
3005                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3006                                                     kern_filter_sz,
3007                                                     ib_filter_sz);
3008                 if (actual_filter_sz <= 0)
3009                         return -EINVAL;
3010                 ib_spec->tunnel.size = sizeof(struct ib_flow_spec_tunnel);
3011                 memcpy(&ib_spec->tunnel.val, kern_spec_val, actual_filter_sz);
3012                 memcpy(&ib_spec->tunnel.mask, kern_spec_mask, actual_filter_sz);
3013
3014                 if ((ntohl(ib_spec->tunnel.mask.tunnel_id)) >= BIT(24) ||
3015                     (ntohl(ib_spec->tunnel.val.tunnel_id)) >= BIT(24))
3016                         return -EINVAL;
3017                 break;
3018         case IB_FLOW_SPEC_ESP:
3019                 ib_filter_sz = offsetof(struct ib_flow_esp_filter, real_sz);
3020                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3021                                                     kern_filter_sz,
3022                                                     ib_filter_sz);
3023                 if (actual_filter_sz <= 0)
3024                         return -EINVAL;
3025                 ib_spec->esp.size = sizeof(struct ib_flow_spec_esp);
3026                 memcpy(&ib_spec->esp.val, kern_spec_val, actual_filter_sz);
3027                 memcpy(&ib_spec->esp.mask, kern_spec_mask, actual_filter_sz);
3028                 break;
3029         case IB_FLOW_SPEC_GRE:
3030                 ib_filter_sz = offsetof(struct ib_flow_gre_filter, real_sz);
3031                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3032                                                     kern_filter_sz,
3033                                                     ib_filter_sz);
3034                 if (actual_filter_sz <= 0)
3035                         return -EINVAL;
3036                 ib_spec->gre.size = sizeof(struct ib_flow_spec_gre);
3037                 memcpy(&ib_spec->gre.val, kern_spec_val, actual_filter_sz);
3038                 memcpy(&ib_spec->gre.mask, kern_spec_mask, actual_filter_sz);
3039                 break;
3040         case IB_FLOW_SPEC_MPLS:
3041                 ib_filter_sz = offsetof(struct ib_flow_mpls_filter, real_sz);
3042                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3043                                                     kern_filter_sz,
3044                                                     ib_filter_sz);
3045                 if (actual_filter_sz <= 0)
3046                         return -EINVAL;
3047                 ib_spec->mpls.size = sizeof(struct ib_flow_spec_mpls);
3048                 memcpy(&ib_spec->mpls.val, kern_spec_val, actual_filter_sz);
3049                 memcpy(&ib_spec->mpls.mask, kern_spec_mask, actual_filter_sz);
3050                 break;
3051         default:
3052                 return -EINVAL;
3053         }
3054         return 0;
3055 }
3056
3057 static int kern_spec_to_ib_spec_filter(struct ib_uverbs_flow_spec *kern_spec,
3058                                        union ib_flow_spec *ib_spec)
3059 {
3060         size_t kern_filter_sz;
3061         void *kern_spec_mask;
3062         void *kern_spec_val;
3063
3064         if (check_sub_overflow((size_t)kern_spec->hdr.size,
3065                                sizeof(struct ib_uverbs_flow_spec_hdr),
3066                                &kern_filter_sz))
3067                 return -EINVAL;
3068
3069         kern_filter_sz /= 2;
3070
3071         kern_spec_val = (void *)kern_spec +
3072                 sizeof(struct ib_uverbs_flow_spec_hdr);
3073         kern_spec_mask = kern_spec_val + kern_filter_sz;
3074
3075         return ib_uverbs_kern_spec_to_ib_spec_filter(kern_spec->type,
3076                                                      kern_spec_mask,
3077                                                      kern_spec_val,
3078                                                      kern_filter_sz, ib_spec);
3079 }
3080
3081 static int kern_spec_to_ib_spec(struct ib_uverbs_file *ufile,
3082                                 struct ib_uverbs_flow_spec *kern_spec,
3083                                 union ib_flow_spec *ib_spec,
3084                                 struct ib_uflow_resources *uflow_res)
3085 {
3086         if (kern_spec->reserved)
3087                 return -EINVAL;
3088
3089         if (kern_spec->type >= IB_FLOW_SPEC_ACTION_TAG)
3090                 return kern_spec_to_ib_spec_action(ufile, kern_spec, ib_spec,
3091                                                    uflow_res);
3092         else
3093                 return kern_spec_to_ib_spec_filter(kern_spec, ib_spec);
3094 }
3095
3096 int ib_uverbs_ex_create_wq(struct ib_uverbs_file *file,
3097                            struct ib_udata *ucore,
3098                            struct ib_udata *uhw)
3099 {
3100         struct ib_uverbs_ex_create_wq     cmd = {};
3101         struct ib_uverbs_ex_create_wq_resp resp = {};
3102         struct ib_uwq_object           *obj;
3103         int err = 0;
3104         struct ib_cq *cq;
3105         struct ib_pd *pd;
3106         struct ib_wq *wq;
3107         struct ib_wq_init_attr wq_init_attr = {};
3108         size_t required_cmd_sz;
3109         size_t required_resp_len;
3110         struct ib_device *ib_dev;
3111
3112         required_cmd_sz = offsetof(typeof(cmd), max_sge) + sizeof(cmd.max_sge);
3113         required_resp_len = offsetof(typeof(resp), wqn) + sizeof(resp.wqn);
3114
3115         if (ucore->inlen < required_cmd_sz)
3116                 return -EINVAL;
3117
3118         if (ucore->outlen < required_resp_len)
3119                 return -ENOSPC;
3120
3121         if (ucore->inlen > sizeof(cmd) &&
3122             !ib_is_udata_cleared(ucore, sizeof(cmd),
3123                                  ucore->inlen - sizeof(cmd)))
3124                 return -EOPNOTSUPP;
3125
3126         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3127         if (err)
3128                 return err;
3129
3130         if (cmd.comp_mask)
3131                 return -EOPNOTSUPP;
3132
3133         obj = (struct ib_uwq_object *)uobj_alloc(UVERBS_OBJECT_WQ, file,
3134                                                  &ib_dev);
3135         if (IS_ERR(obj))
3136                 return PTR_ERR(obj);
3137
3138         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file);
3139         if (!pd) {
3140                 err = -EINVAL;
3141                 goto err_uobj;
3142         }
3143
3144         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
3145         if (!cq) {
3146                 err = -EINVAL;
3147                 goto err_put_pd;
3148         }
3149
3150         wq_init_attr.cq = cq;
3151         wq_init_attr.max_sge = cmd.max_sge;
3152         wq_init_attr.max_wr = cmd.max_wr;
3153         wq_init_attr.wq_context = file;
3154         wq_init_attr.wq_type = cmd.wq_type;
3155         wq_init_attr.event_handler = ib_uverbs_wq_event_handler;
3156         if (ucore->inlen >= (offsetof(typeof(cmd), create_flags) +
3157                              sizeof(cmd.create_flags)))
3158                 wq_init_attr.create_flags = cmd.create_flags;
3159         obj->uevent.events_reported = 0;
3160         INIT_LIST_HEAD(&obj->uevent.event_list);
3161
3162         if (!pd->device->create_wq) {
3163                 err = -EOPNOTSUPP;
3164                 goto err_put_cq;
3165         }
3166         wq = pd->device->create_wq(pd, &wq_init_attr, uhw);
3167         if (IS_ERR(wq)) {
3168                 err = PTR_ERR(wq);
3169                 goto err_put_cq;
3170         }
3171
3172         wq->uobject = &obj->uevent.uobject;
3173         obj->uevent.uobject.object = wq;
3174         wq->wq_type = wq_init_attr.wq_type;
3175         wq->cq = cq;
3176         wq->pd = pd;
3177         wq->device = pd->device;
3178         wq->wq_context = wq_init_attr.wq_context;
3179         atomic_set(&wq->usecnt, 0);
3180         atomic_inc(&pd->usecnt);
3181         atomic_inc(&cq->usecnt);
3182         wq->uobject = &obj->uevent.uobject;
3183         obj->uevent.uobject.object = wq;
3184
3185         memset(&resp, 0, sizeof(resp));
3186         resp.wq_handle = obj->uevent.uobject.id;
3187         resp.max_sge = wq_init_attr.max_sge;
3188         resp.max_wr = wq_init_attr.max_wr;
3189         resp.wqn = wq->wq_num;
3190         resp.response_length = required_resp_len;
3191         err = ib_copy_to_udata(ucore,
3192                                &resp, resp.response_length);
3193         if (err)
3194                 goto err_copy;
3195
3196         uobj_put_obj_read(pd);
3197         uobj_put_obj_read(cq);
3198         return uobj_alloc_commit(&obj->uevent.uobject, 0);
3199
3200 err_copy:
3201         ib_destroy_wq(wq);
3202 err_put_cq:
3203         uobj_put_obj_read(cq);
3204 err_put_pd:
3205         uobj_put_obj_read(pd);
3206 err_uobj:
3207         uobj_alloc_abort(&obj->uevent.uobject);
3208
3209         return err;
3210 }
3211
3212 int ib_uverbs_ex_destroy_wq(struct ib_uverbs_file *file,
3213                             struct ib_udata *ucore,
3214                             struct ib_udata *uhw)
3215 {
3216         struct ib_uverbs_ex_destroy_wq  cmd = {};
3217         struct ib_uverbs_ex_destroy_wq_resp     resp = {};
3218         struct ib_uobject               *uobj;
3219         struct ib_uwq_object            *obj;
3220         size_t required_cmd_sz;
3221         size_t required_resp_len;
3222         int                             ret;
3223
3224         required_cmd_sz = offsetof(typeof(cmd), wq_handle) + sizeof(cmd.wq_handle);
3225         required_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
3226
3227         if (ucore->inlen < required_cmd_sz)
3228                 return -EINVAL;
3229
3230         if (ucore->outlen < required_resp_len)
3231                 return -ENOSPC;
3232
3233         if (ucore->inlen > sizeof(cmd) &&
3234             !ib_is_udata_cleared(ucore, sizeof(cmd),
3235                                  ucore->inlen - sizeof(cmd)))
3236                 return -EOPNOTSUPP;
3237
3238         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3239         if (ret)
3240                 return ret;
3241
3242         if (cmd.comp_mask)
3243                 return -EOPNOTSUPP;
3244
3245         resp.response_length = required_resp_len;
3246         uobj = uobj_get_destroy(UVERBS_OBJECT_WQ, cmd.wq_handle, file);
3247         if (IS_ERR(uobj))
3248                 return PTR_ERR(uobj);
3249
3250         obj = container_of(uobj, struct ib_uwq_object, uevent.uobject);
3251         resp.events_reported = obj->uevent.events_reported;
3252
3253         uobj_put_destroy(uobj);
3254
3255         return ib_copy_to_udata(ucore, &resp, resp.response_length);
3256 }
3257
3258 int ib_uverbs_ex_modify_wq(struct ib_uverbs_file *file,
3259                            struct ib_udata *ucore,
3260                            struct ib_udata *uhw)
3261 {
3262         struct ib_uverbs_ex_modify_wq cmd = {};
3263         struct ib_wq *wq;
3264         struct ib_wq_attr wq_attr = {};
3265         size_t required_cmd_sz;
3266         int ret;
3267
3268         required_cmd_sz = offsetof(typeof(cmd), curr_wq_state) + sizeof(cmd.curr_wq_state);
3269         if (ucore->inlen < required_cmd_sz)
3270                 return -EINVAL;
3271
3272         if (ucore->inlen > sizeof(cmd) &&
3273             !ib_is_udata_cleared(ucore, sizeof(cmd),
3274                                  ucore->inlen - sizeof(cmd)))
3275                 return -EOPNOTSUPP;
3276
3277         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3278         if (ret)
3279                 return ret;
3280
3281         if (!cmd.attr_mask)
3282                 return -EINVAL;
3283
3284         if (cmd.attr_mask > (IB_WQ_STATE | IB_WQ_CUR_STATE | IB_WQ_FLAGS))
3285                 return -EINVAL;
3286
3287         wq = uobj_get_obj_read(wq, UVERBS_OBJECT_WQ, cmd.wq_handle, file);
3288         if (!wq)
3289                 return -EINVAL;
3290
3291         wq_attr.curr_wq_state = cmd.curr_wq_state;
3292         wq_attr.wq_state = cmd.wq_state;
3293         if (cmd.attr_mask & IB_WQ_FLAGS) {
3294                 wq_attr.flags = cmd.flags;
3295                 wq_attr.flags_mask = cmd.flags_mask;
3296         }
3297         if (!wq->device->modify_wq) {
3298                 ret = -EOPNOTSUPP;
3299                 goto out;
3300         }
3301         ret = wq->device->modify_wq(wq, &wq_attr, cmd.attr_mask, uhw);
3302 out:
3303         uobj_put_obj_read(wq);
3304         return ret;
3305 }
3306
3307 int ib_uverbs_ex_create_rwq_ind_table(struct ib_uverbs_file *file,
3308                                       struct ib_udata *ucore,
3309                                       struct ib_udata *uhw)
3310 {
3311         struct ib_uverbs_ex_create_rwq_ind_table          cmd = {};
3312         struct ib_uverbs_ex_create_rwq_ind_table_resp  resp = {};
3313         struct ib_uobject                 *uobj;
3314         int err = 0;
3315         struct ib_rwq_ind_table_init_attr init_attr = {};
3316         struct ib_rwq_ind_table *rwq_ind_tbl;
3317         struct ib_wq    **wqs = NULL;
3318         u32 *wqs_handles = NULL;
3319         struct ib_wq    *wq = NULL;
3320         int i, j, num_read_wqs;
3321         u32 num_wq_handles;
3322         u32 expected_in_size;
3323         size_t required_cmd_sz_header;
3324         size_t required_resp_len;
3325         struct ib_device *ib_dev;
3326
3327         required_cmd_sz_header = offsetof(typeof(cmd), log_ind_tbl_size) + sizeof(cmd.log_ind_tbl_size);
3328         required_resp_len = offsetof(typeof(resp), ind_tbl_num) + sizeof(resp.ind_tbl_num);
3329
3330         if (ucore->inlen < required_cmd_sz_header)
3331                 return -EINVAL;
3332
3333         if (ucore->outlen < required_resp_len)
3334                 return -ENOSPC;
3335
3336         err = ib_copy_from_udata(&cmd, ucore, required_cmd_sz_header);
3337         if (err)
3338                 return err;
3339
3340         ucore->inbuf += required_cmd_sz_header;
3341         ucore->inlen -= required_cmd_sz_header;
3342
3343         if (cmd.comp_mask)
3344                 return -EOPNOTSUPP;
3345
3346         if (cmd.log_ind_tbl_size > IB_USER_VERBS_MAX_LOG_IND_TBL_SIZE)
3347                 return -EINVAL;
3348
3349         num_wq_handles = 1 << cmd.log_ind_tbl_size;
3350         expected_in_size = num_wq_handles * sizeof(__u32);
3351         if (num_wq_handles == 1)
3352                 /* input size for wq handles is u64 aligned */
3353                 expected_in_size += sizeof(__u32);
3354
3355         if (ucore->inlen < expected_in_size)
3356                 return -EINVAL;
3357
3358         if (ucore->inlen > expected_in_size &&
3359             !ib_is_udata_cleared(ucore, expected_in_size,
3360                                  ucore->inlen - expected_in_size))
3361                 return -EOPNOTSUPP;
3362
3363         wqs_handles = kcalloc(num_wq_handles, sizeof(*wqs_handles),
3364                               GFP_KERNEL);
3365         if (!wqs_handles)
3366                 return -ENOMEM;
3367
3368         err = ib_copy_from_udata(wqs_handles, ucore,
3369                                  num_wq_handles * sizeof(__u32));
3370         if (err)
3371                 goto err_free;
3372
3373         wqs = kcalloc(num_wq_handles, sizeof(*wqs), GFP_KERNEL);
3374         if (!wqs) {
3375                 err = -ENOMEM;
3376                 goto  err_free;
3377         }
3378
3379         for (num_read_wqs = 0; num_read_wqs < num_wq_handles;
3380                         num_read_wqs++) {
3381                 wq = uobj_get_obj_read(wq, UVERBS_OBJECT_WQ,
3382                                        wqs_handles[num_read_wqs], file);
3383                 if (!wq) {
3384                         err = -EINVAL;
3385                         goto put_wqs;
3386                 }
3387
3388                 wqs[num_read_wqs] = wq;
3389         }
3390
3391         uobj = uobj_alloc(UVERBS_OBJECT_RWQ_IND_TBL, file, &ib_dev);
3392         if (IS_ERR(uobj)) {
3393                 err = PTR_ERR(uobj);
3394                 goto put_wqs;
3395         }
3396
3397         init_attr.log_ind_tbl_size = cmd.log_ind_tbl_size;
3398         init_attr.ind_tbl = wqs;
3399
3400         if (!ib_dev->create_rwq_ind_table) {
3401                 err = -EOPNOTSUPP;
3402                 goto err_uobj;
3403         }
3404         rwq_ind_tbl = ib_dev->create_rwq_ind_table(ib_dev, &init_attr, uhw);
3405
3406         if (IS_ERR(rwq_ind_tbl)) {
3407                 err = PTR_ERR(rwq_ind_tbl);
3408                 goto err_uobj;
3409         }
3410
3411         rwq_ind_tbl->ind_tbl = wqs;
3412         rwq_ind_tbl->log_ind_tbl_size = init_attr.log_ind_tbl_size;
3413         rwq_ind_tbl->uobject = uobj;
3414         uobj->object = rwq_ind_tbl;
3415         rwq_ind_tbl->device = ib_dev;
3416         atomic_set(&rwq_ind_tbl->usecnt, 0);
3417
3418         for (i = 0; i < num_wq_handles; i++)
3419                 atomic_inc(&wqs[i]->usecnt);
3420
3421         resp.ind_tbl_handle = uobj->id;
3422         resp.ind_tbl_num = rwq_ind_tbl->ind_tbl_num;
3423         resp.response_length = required_resp_len;
3424
3425         err = ib_copy_to_udata(ucore,
3426                                &resp, resp.response_length);
3427         if (err)
3428                 goto err_copy;
3429
3430         kfree(wqs_handles);
3431
3432         for (j = 0; j < num_read_wqs; j++)
3433                 uobj_put_obj_read(wqs[j]);
3434
3435         return uobj_alloc_commit(uobj, 0);
3436
3437 err_copy:
3438         ib_destroy_rwq_ind_table(rwq_ind_tbl);
3439 err_uobj:
3440         uobj_alloc_abort(uobj);
3441 put_wqs:
3442         for (j = 0; j < num_read_wqs; j++)
3443                 uobj_put_obj_read(wqs[j]);
3444 err_free:
3445         kfree(wqs_handles);
3446         kfree(wqs);
3447         return err;
3448 }
3449
3450 int ib_uverbs_ex_destroy_rwq_ind_table(struct ib_uverbs_file *file,
3451                                        struct ib_udata *ucore,
3452                                        struct ib_udata *uhw)
3453 {
3454         struct ib_uverbs_ex_destroy_rwq_ind_table       cmd = {};
3455         int                     ret;
3456         size_t required_cmd_sz;
3457
3458         required_cmd_sz = offsetof(typeof(cmd), ind_tbl_handle) + sizeof(cmd.ind_tbl_handle);
3459
3460         if (ucore->inlen < required_cmd_sz)
3461                 return -EINVAL;
3462
3463         if (ucore->inlen > sizeof(cmd) &&
3464             !ib_is_udata_cleared(ucore, sizeof(cmd),
3465                                  ucore->inlen - sizeof(cmd)))
3466                 return -EOPNOTSUPP;
3467
3468         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3469         if (ret)
3470                 return ret;
3471
3472         if (cmd.comp_mask)
3473                 return -EOPNOTSUPP;
3474
3475         return uobj_perform_destroy(UVERBS_OBJECT_RWQ_IND_TBL,
3476                                     cmd.ind_tbl_handle, file, 0);
3477 }
3478
3479 int ib_uverbs_ex_create_flow(struct ib_uverbs_file *file,
3480                              struct ib_udata *ucore,
3481                              struct ib_udata *uhw)
3482 {
3483         struct ib_uverbs_create_flow      cmd;
3484         struct ib_uverbs_create_flow_resp resp;
3485         struct ib_uobject                 *uobj;
3486         struct ib_uflow_object            *uflow;
3487         struct ib_flow                    *flow_id;
3488         struct ib_uverbs_flow_attr        *kern_flow_attr;
3489         struct ib_flow_attr               *flow_attr;
3490         struct ib_qp                      *qp;
3491         struct ib_uflow_resources         *uflow_res;
3492         struct ib_uverbs_flow_spec_hdr    *kern_spec;
3493         int err = 0;
3494         void *ib_spec;
3495         int i;
3496         struct ib_device *ib_dev;
3497
3498         if (ucore->inlen < sizeof(cmd))
3499                 return -EINVAL;
3500
3501         if (ucore->outlen < sizeof(resp))
3502                 return -ENOSPC;
3503
3504         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3505         if (err)
3506                 return err;
3507
3508         ucore->inbuf += sizeof(cmd);
3509         ucore->inlen -= sizeof(cmd);
3510
3511         if (cmd.comp_mask)
3512                 return -EINVAL;
3513
3514         if (!capable(CAP_NET_RAW))
3515                 return -EPERM;
3516
3517         if (cmd.flow_attr.flags >= IB_FLOW_ATTR_FLAGS_RESERVED)
3518                 return -EINVAL;
3519
3520         if ((cmd.flow_attr.flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) &&
3521             ((cmd.flow_attr.type == IB_FLOW_ATTR_ALL_DEFAULT) ||
3522              (cmd.flow_attr.type == IB_FLOW_ATTR_MC_DEFAULT)))
3523                 return -EINVAL;
3524
3525         if (cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
3526                 return -EINVAL;
3527
3528         if (cmd.flow_attr.size > ucore->inlen ||
3529             cmd.flow_attr.size >
3530             (cmd.flow_attr.num_of_specs * sizeof(struct ib_uverbs_flow_spec)))
3531                 return -EINVAL;
3532
3533         if (cmd.flow_attr.reserved[0] ||
3534             cmd.flow_attr.reserved[1])
3535                 return -EINVAL;
3536
3537         if (cmd.flow_attr.num_of_specs) {
3538                 kern_flow_attr = kmalloc(sizeof(*kern_flow_attr) + cmd.flow_attr.size,
3539                                          GFP_KERNEL);
3540                 if (!kern_flow_attr)
3541                         return -ENOMEM;
3542
3543                 *kern_flow_attr = cmd.flow_attr;
3544                 err = ib_copy_from_udata(&kern_flow_attr->flow_specs, ucore,
3545                                          cmd.flow_attr.size);
3546                 if (err)
3547                         goto err_free_attr;
3548         } else {
3549                 kern_flow_attr = &cmd.flow_attr;
3550         }
3551
3552         uobj = uobj_alloc(UVERBS_OBJECT_FLOW, file, &ib_dev);
3553         if (IS_ERR(uobj)) {
3554                 err = PTR_ERR(uobj);
3555                 goto err_free_attr;
3556         }
3557
3558         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
3559         if (!qp) {
3560                 err = -EINVAL;
3561                 goto err_uobj;
3562         }
3563
3564         if (qp->qp_type != IB_QPT_UD && qp->qp_type != IB_QPT_RAW_PACKET) {
3565                 err = -EINVAL;
3566                 goto err_put;
3567         }
3568
3569         if (!qp->device->create_flow) {
3570                 err = -EOPNOTSUPP;
3571                 goto err_put;
3572         }
3573
3574         flow_attr = kzalloc(struct_size(flow_attr, flows,
3575                                 cmd.flow_attr.num_of_specs), GFP_KERNEL);
3576         if (!flow_attr) {
3577                 err = -ENOMEM;
3578                 goto err_put;
3579         }
3580         uflow_res = flow_resources_alloc(cmd.flow_attr.num_of_specs);
3581         if (!uflow_res) {
3582                 err = -ENOMEM;
3583                 goto err_free_flow_attr;
3584         }
3585
3586         flow_attr->type = kern_flow_attr->type;
3587         flow_attr->priority = kern_flow_attr->priority;
3588         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
3589         flow_attr->port = kern_flow_attr->port;
3590         flow_attr->flags = kern_flow_attr->flags;
3591         flow_attr->size = sizeof(*flow_attr);
3592
3593         kern_spec = kern_flow_attr->flow_specs;
3594         ib_spec = flow_attr + 1;
3595         for (i = 0; i < flow_attr->num_of_specs &&
3596                         cmd.flow_attr.size >= sizeof(*kern_spec) &&
3597                         cmd.flow_attr.size >= kern_spec->size;
3598              i++) {
3599                 err = kern_spec_to_ib_spec(
3600                                 file, (struct ib_uverbs_flow_spec *)kern_spec,
3601                                 ib_spec, uflow_res);
3602                 if (err)
3603                         goto err_free;
3604
3605                 flow_attr->size +=
3606                         ((union ib_flow_spec *) ib_spec)->size;
3607                 cmd.flow_attr.size -= kern_spec->size;
3608                 kern_spec = ((void *)kern_spec) + kern_spec->size;
3609                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
3610         }
3611         if (cmd.flow_attr.size || (i != flow_attr->num_of_specs)) {
3612                 pr_warn("create flow failed, flow %d: %d bytes left from uverb cmd\n",
3613                         i, cmd.flow_attr.size);
3614                 err = -EINVAL;
3615                 goto err_free;
3616         }
3617
3618         flow_id = qp->device->create_flow(qp, flow_attr,
3619                                           IB_FLOW_DOMAIN_USER, uhw);
3620
3621         if (IS_ERR(flow_id)) {
3622                 err = PTR_ERR(flow_id);
3623                 goto err_free;
3624         }
3625         atomic_inc(&qp->usecnt);
3626         flow_id->qp = qp;
3627         flow_id->device = qp->device;
3628         flow_id->uobject = uobj;
3629         uobj->object = flow_id;
3630         uflow = container_of(uobj, typeof(*uflow), uobject);
3631         uflow->resources = uflow_res;
3632
3633         memset(&resp, 0, sizeof(resp));
3634         resp.flow_handle = uobj->id;
3635
3636         err = ib_copy_to_udata(ucore,
3637                                &resp, sizeof(resp));
3638         if (err)
3639                 goto err_copy;
3640
3641         uobj_put_obj_read(qp);
3642         kfree(flow_attr);
3643         if (cmd.flow_attr.num_of_specs)
3644                 kfree(kern_flow_attr);
3645         return uobj_alloc_commit(uobj, 0);
3646 err_copy:
3647         if (!qp->device->destroy_flow(flow_id))
3648                 atomic_dec(&qp->usecnt);
3649 err_free:
3650         ib_uverbs_flow_resources_free(uflow_res);
3651 err_free_flow_attr:
3652         kfree(flow_attr);
3653 err_put:
3654         uobj_put_obj_read(qp);
3655 err_uobj:
3656         uobj_alloc_abort(uobj);
3657 err_free_attr:
3658         if (cmd.flow_attr.num_of_specs)
3659                 kfree(kern_flow_attr);
3660         return err;
3661 }
3662
3663 int ib_uverbs_ex_destroy_flow(struct ib_uverbs_file *file,
3664                               struct ib_udata *ucore,
3665                               struct ib_udata *uhw)
3666 {
3667         struct ib_uverbs_destroy_flow   cmd;
3668         int                             ret;
3669
3670         if (ucore->inlen < sizeof(cmd))
3671                 return -EINVAL;
3672
3673         ret = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3674         if (ret)
3675                 return ret;
3676
3677         if (cmd.comp_mask)
3678                 return -EINVAL;
3679
3680         return uobj_perform_destroy(UVERBS_OBJECT_FLOW, cmd.flow_handle, file,
3681                                     0);
3682 }
3683
3684 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
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         struct ib_device *ib_dev;
3696
3697         obj = (struct ib_usrq_object *)uobj_alloc(UVERBS_OBJECT_SRQ, file,
3698                                                   &ib_dev);
3699         if (IS_ERR(obj))
3700                 return PTR_ERR(obj);
3701
3702         if (cmd->srq_type == IB_SRQT_TM)
3703                 attr.ext.tag_matching.max_num_tags = cmd->max_num_tags;
3704
3705         if (cmd->srq_type == IB_SRQT_XRC) {
3706                 xrcd_uobj = uobj_get_read(UVERBS_OBJECT_XRCD, cmd->xrcd_handle,
3707                                           file);
3708                 if (IS_ERR(xrcd_uobj)) {
3709                         ret = -EINVAL;
3710                         goto err;
3711                 }
3712
3713                 attr.ext.xrc.xrcd = (struct ib_xrcd *)xrcd_uobj->object;
3714                 if (!attr.ext.xrc.xrcd) {
3715                         ret = -EINVAL;
3716                         goto err_put_xrcd;
3717                 }
3718
3719                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
3720                 atomic_inc(&obj->uxrcd->refcnt);
3721         }
3722
3723         if (ib_srq_has_cq(cmd->srq_type)) {
3724                 attr.ext.cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ,
3725                                                 cmd->cq_handle, file);
3726                 if (!attr.ext.cq) {
3727                         ret = -EINVAL;
3728                         goto err_put_xrcd;
3729                 }
3730         }
3731
3732         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd->pd_handle, file);
3733         if (!pd) {
3734                 ret = -EINVAL;
3735                 goto err_put_cq;
3736         }
3737
3738         attr.event_handler  = ib_uverbs_srq_event_handler;
3739         attr.srq_context    = file;
3740         attr.srq_type       = cmd->srq_type;
3741         attr.attr.max_wr    = cmd->max_wr;
3742         attr.attr.max_sge   = cmd->max_sge;
3743         attr.attr.srq_limit = cmd->srq_limit;
3744
3745         obj->uevent.events_reported = 0;
3746         INIT_LIST_HEAD(&obj->uevent.event_list);
3747
3748         srq = pd->device->create_srq(pd, &attr, udata);
3749         if (IS_ERR(srq)) {
3750                 ret = PTR_ERR(srq);
3751                 goto err_put;
3752         }
3753
3754         srq->device        = pd->device;
3755         srq->pd            = pd;
3756         srq->srq_type      = cmd->srq_type;
3757         srq->uobject       = &obj->uevent.uobject;
3758         srq->event_handler = attr.event_handler;
3759         srq->srq_context   = attr.srq_context;
3760
3761         if (ib_srq_has_cq(cmd->srq_type)) {
3762                 srq->ext.cq       = attr.ext.cq;
3763                 atomic_inc(&attr.ext.cq->usecnt);
3764         }
3765
3766         if (cmd->srq_type == IB_SRQT_XRC) {
3767                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
3768                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
3769         }
3770
3771         atomic_inc(&pd->usecnt);
3772         atomic_set(&srq->usecnt, 0);
3773
3774         obj->uevent.uobject.object = srq;
3775         obj->uevent.uobject.user_handle = cmd->user_handle;
3776
3777         memset(&resp, 0, sizeof resp);
3778         resp.srq_handle = obj->uevent.uobject.id;
3779         resp.max_wr     = attr.attr.max_wr;
3780         resp.max_sge    = attr.attr.max_sge;
3781         if (cmd->srq_type == IB_SRQT_XRC)
3782                 resp.srqn = srq->ext.xrc.srq_num;
3783
3784         if (copy_to_user(u64_to_user_ptr(cmd->response),
3785                          &resp, sizeof resp)) {
3786                 ret = -EFAULT;
3787                 goto err_copy;
3788         }
3789
3790         if (cmd->srq_type == IB_SRQT_XRC)
3791                 uobj_put_read(xrcd_uobj);
3792
3793         if (ib_srq_has_cq(cmd->srq_type))
3794                 uobj_put_obj_read(attr.ext.cq);
3795
3796         uobj_put_obj_read(pd);
3797         return uobj_alloc_commit(&obj->uevent.uobject, 0);
3798
3799 err_copy:
3800         ib_destroy_srq(srq);
3801
3802 err_put:
3803         uobj_put_obj_read(pd);
3804
3805 err_put_cq:
3806         if (ib_srq_has_cq(cmd->srq_type))
3807                 uobj_put_obj_read(attr.ext.cq);
3808
3809 err_put_xrcd:
3810         if (cmd->srq_type == IB_SRQT_XRC) {
3811                 atomic_dec(&obj->uxrcd->refcnt);
3812                 uobj_put_read(xrcd_uobj);
3813         }
3814
3815 err:
3816         uobj_alloc_abort(&obj->uevent.uobject);
3817         return ret;
3818 }
3819
3820 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
3821                              const char __user *buf, int in_len,
3822                              int out_len)
3823 {
3824         struct ib_uverbs_create_srq      cmd;
3825         struct ib_uverbs_create_xsrq     xcmd;
3826         struct ib_uverbs_create_srq_resp resp;
3827         struct ib_udata                  udata;
3828         int ret;
3829
3830         if (out_len < sizeof resp)
3831                 return -ENOSPC;
3832
3833         if (copy_from_user(&cmd, buf, sizeof cmd))
3834                 return -EFAULT;
3835
3836         memset(&xcmd, 0, sizeof(xcmd));
3837         xcmd.response    = cmd.response;
3838         xcmd.user_handle = cmd.user_handle;
3839         xcmd.srq_type    = IB_SRQT_BASIC;
3840         xcmd.pd_handle   = cmd.pd_handle;
3841         xcmd.max_wr      = cmd.max_wr;
3842         xcmd.max_sge     = cmd.max_sge;
3843         xcmd.srq_limit   = cmd.srq_limit;
3844
3845         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
3846                    u64_to_user_ptr(cmd.response) + sizeof(resp),
3847                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
3848                    out_len - sizeof(resp));
3849
3850         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
3851         if (ret)
3852                 return ret;
3853
3854         return in_len;
3855 }
3856
3857 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
3858                               const char __user *buf, int in_len, int out_len)
3859 {
3860         struct ib_uverbs_create_xsrq     cmd;
3861         struct ib_uverbs_create_srq_resp resp;
3862         struct ib_udata                  udata;
3863         int ret;
3864
3865         if (out_len < sizeof resp)
3866                 return -ENOSPC;
3867
3868         if (copy_from_user(&cmd, buf, sizeof cmd))
3869                 return -EFAULT;
3870
3871         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
3872                    u64_to_user_ptr(cmd.response) + sizeof(resp),
3873                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
3874                    out_len - sizeof(resp));
3875
3876         ret = __uverbs_create_xsrq(file, &cmd, &udata);
3877         if (ret)
3878                 return ret;
3879
3880         return in_len;
3881 }
3882
3883 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
3884                              const char __user *buf, int in_len,
3885                              int out_len)
3886 {
3887         struct ib_uverbs_modify_srq cmd;
3888         struct ib_udata             udata;
3889         struct ib_srq              *srq;
3890         struct ib_srq_attr          attr;
3891         int                         ret;
3892
3893         if (copy_from_user(&cmd, buf, sizeof cmd))
3894                 return -EFAULT;
3895
3896         ib_uverbs_init_udata(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
3897                    out_len);
3898
3899         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
3900         if (!srq)
3901                 return -EINVAL;
3902
3903         attr.max_wr    = cmd.max_wr;
3904         attr.srq_limit = cmd.srq_limit;
3905
3906         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
3907
3908         uobj_put_obj_read(srq);
3909
3910         return ret ? ret : in_len;
3911 }
3912
3913 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
3914                             const char __user *buf,
3915                             int in_len, int out_len)
3916 {
3917         struct ib_uverbs_query_srq      cmd;
3918         struct ib_uverbs_query_srq_resp resp;
3919         struct ib_srq_attr              attr;
3920         struct ib_srq                   *srq;
3921         int                             ret;
3922
3923         if (out_len < sizeof resp)
3924                 return -ENOSPC;
3925
3926         if (copy_from_user(&cmd, buf, sizeof cmd))
3927                 return -EFAULT;
3928
3929         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
3930         if (!srq)
3931                 return -EINVAL;
3932
3933         ret = ib_query_srq(srq, &attr);
3934
3935         uobj_put_obj_read(srq);
3936
3937         if (ret)
3938                 return ret;
3939
3940         memset(&resp, 0, sizeof resp);
3941
3942         resp.max_wr    = attr.max_wr;
3943         resp.max_sge   = attr.max_sge;
3944         resp.srq_limit = attr.srq_limit;
3945
3946         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
3947                 return -EFAULT;
3948
3949         return in_len;
3950 }
3951
3952 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
3953                               const char __user *buf, int in_len,
3954                               int out_len)
3955 {
3956         struct ib_uverbs_destroy_srq      cmd;
3957         struct ib_uverbs_destroy_srq_resp resp;
3958         struct ib_uobject                *uobj;
3959         struct ib_uevent_object          *obj;
3960
3961         if (copy_from_user(&cmd, buf, sizeof cmd))
3962                 return -EFAULT;
3963
3964         uobj = uobj_get_destroy(UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
3965         if (IS_ERR(uobj))
3966                 return PTR_ERR(uobj);
3967
3968         obj = container_of(uobj, struct ib_uevent_object, uobject);
3969         memset(&resp, 0, sizeof(resp));
3970         resp.events_reported = obj->events_reported;
3971
3972         uobj_put_destroy(uobj);
3973
3974         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof(resp)))
3975                 return -EFAULT;
3976
3977         return in_len;
3978 }
3979
3980 int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
3981                               struct ib_udata *ucore,
3982                               struct ib_udata *uhw)
3983 {
3984         struct ib_uverbs_ex_query_device_resp resp = { {0} };
3985         struct ib_uverbs_ex_query_device  cmd;
3986         struct ib_device_attr attr = {0};
3987         struct ib_ucontext *ucontext;
3988         struct ib_device *ib_dev;
3989         int err;
3990
3991         ucontext = ib_uverbs_get_ucontext(file);
3992         if (IS_ERR(ucontext))
3993                 return PTR_ERR(ucontext);
3994         ib_dev = ucontext->device;
3995
3996         if (!ib_dev->query_device)
3997                 return -EOPNOTSUPP;
3998
3999         if (ucore->inlen < sizeof(cmd))
4000                 return -EINVAL;
4001
4002         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
4003         if (err)
4004                 return err;
4005
4006         if (cmd.comp_mask)
4007                 return -EINVAL;
4008
4009         if (cmd.reserved)
4010                 return -EINVAL;
4011
4012         resp.response_length = offsetof(typeof(resp), odp_caps);
4013
4014         if (ucore->outlen < resp.response_length)
4015                 return -ENOSPC;
4016
4017         err = ib_dev->query_device(ib_dev, &attr, uhw);
4018         if (err)
4019                 return err;
4020
4021         copy_query_dev_fields(ucontext, &resp.base, &attr);
4022
4023         if (ucore->outlen < resp.response_length + sizeof(resp.odp_caps))
4024                 goto end;
4025
4026 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4027         resp.odp_caps.general_caps = attr.odp_caps.general_caps;
4028         resp.odp_caps.per_transport_caps.rc_odp_caps =
4029                 attr.odp_caps.per_transport_caps.rc_odp_caps;
4030         resp.odp_caps.per_transport_caps.uc_odp_caps =
4031                 attr.odp_caps.per_transport_caps.uc_odp_caps;
4032         resp.odp_caps.per_transport_caps.ud_odp_caps =
4033                 attr.odp_caps.per_transport_caps.ud_odp_caps;
4034 #endif
4035         resp.response_length += sizeof(resp.odp_caps);
4036
4037         if (ucore->outlen < resp.response_length + sizeof(resp.timestamp_mask))
4038                 goto end;
4039
4040         resp.timestamp_mask = attr.timestamp_mask;
4041         resp.response_length += sizeof(resp.timestamp_mask);
4042
4043         if (ucore->outlen < resp.response_length + sizeof(resp.hca_core_clock))
4044                 goto end;
4045
4046         resp.hca_core_clock = attr.hca_core_clock;
4047         resp.response_length += sizeof(resp.hca_core_clock);
4048
4049         if (ucore->outlen < resp.response_length + sizeof(resp.device_cap_flags_ex))
4050                 goto end;
4051
4052         resp.device_cap_flags_ex = attr.device_cap_flags;
4053         resp.response_length += sizeof(resp.device_cap_flags_ex);
4054
4055         if (ucore->outlen < resp.response_length + sizeof(resp.rss_caps))
4056                 goto end;
4057
4058         resp.rss_caps.supported_qpts = attr.rss_caps.supported_qpts;
4059         resp.rss_caps.max_rwq_indirection_tables =
4060                 attr.rss_caps.max_rwq_indirection_tables;
4061         resp.rss_caps.max_rwq_indirection_table_size =
4062                 attr.rss_caps.max_rwq_indirection_table_size;
4063
4064         resp.response_length += sizeof(resp.rss_caps);
4065
4066         if (ucore->outlen < resp.response_length + sizeof(resp.max_wq_type_rq))
4067                 goto end;
4068
4069         resp.max_wq_type_rq = attr.max_wq_type_rq;
4070         resp.response_length += sizeof(resp.max_wq_type_rq);
4071
4072         if (ucore->outlen < resp.response_length + sizeof(resp.raw_packet_caps))
4073                 goto end;
4074
4075         resp.raw_packet_caps = attr.raw_packet_caps;
4076         resp.response_length += sizeof(resp.raw_packet_caps);
4077
4078         if (ucore->outlen < resp.response_length + sizeof(resp.tm_caps))
4079                 goto end;
4080
4081         resp.tm_caps.max_rndv_hdr_size  = attr.tm_caps.max_rndv_hdr_size;
4082         resp.tm_caps.max_num_tags       = attr.tm_caps.max_num_tags;
4083         resp.tm_caps.max_ops            = attr.tm_caps.max_ops;
4084         resp.tm_caps.max_sge            = attr.tm_caps.max_sge;
4085         resp.tm_caps.flags              = attr.tm_caps.flags;
4086         resp.response_length += sizeof(resp.tm_caps);
4087
4088         if (ucore->outlen < resp.response_length + sizeof(resp.cq_moderation_caps))
4089                 goto end;
4090
4091         resp.cq_moderation_caps.max_cq_moderation_count  =
4092                 attr.cq_caps.max_cq_moderation_count;
4093         resp.cq_moderation_caps.max_cq_moderation_period =
4094                 attr.cq_caps.max_cq_moderation_period;
4095         resp.response_length += sizeof(resp.cq_moderation_caps);
4096
4097         if (ucore->outlen < resp.response_length + sizeof(resp.max_dm_size))
4098                 goto end;
4099
4100         resp.max_dm_size = attr.max_dm_size;
4101         resp.response_length += sizeof(resp.max_dm_size);
4102 end:
4103         err = ib_copy_to_udata(ucore, &resp, resp.response_length);
4104         return err;
4105 }
4106
4107 int ib_uverbs_ex_modify_cq(struct ib_uverbs_file *file,
4108                            struct ib_udata *ucore,
4109                            struct ib_udata *uhw)
4110 {
4111         struct ib_uverbs_ex_modify_cq cmd = {};
4112         struct ib_cq *cq;
4113         size_t required_cmd_sz;
4114         int ret;
4115
4116         required_cmd_sz = offsetof(typeof(cmd), reserved) +
4117                                 sizeof(cmd.reserved);
4118         if (ucore->inlen < required_cmd_sz)
4119                 return -EINVAL;
4120
4121         /* sanity checks */
4122         if (ucore->inlen > sizeof(cmd) &&
4123             !ib_is_udata_cleared(ucore, sizeof(cmd),
4124                                  ucore->inlen - sizeof(cmd)))
4125                 return -EOPNOTSUPP;
4126
4127         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
4128         if (ret)
4129                 return ret;
4130
4131         if (!cmd.attr_mask || cmd.reserved)
4132                 return -EINVAL;
4133
4134         if (cmd.attr_mask > IB_CQ_MODERATE)
4135                 return -EOPNOTSUPP;
4136
4137         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
4138         if (!cq)
4139                 return -EINVAL;
4140
4141         ret = rdma_set_cq_moderation(cq, cmd.attr.cq_count, cmd.attr.cq_period);
4142
4143         uobj_put_obj_read(cq);
4144
4145         return ret;
4146 }