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.qp_state > IB_QPS_ERR) {
2031                 ret = -EINVAL;
2032                 goto release_qp;
2033         }
2034
2035         attr->qp_state            = cmd->base.qp_state;
2036         attr->cur_qp_state        = cmd->base.cur_qp_state;
2037         attr->path_mtu            = cmd->base.path_mtu;
2038         attr->path_mig_state      = cmd->base.path_mig_state;
2039         attr->qkey                = cmd->base.qkey;
2040         attr->rq_psn              = cmd->base.rq_psn;
2041         attr->sq_psn              = cmd->base.sq_psn;
2042         attr->dest_qp_num         = cmd->base.dest_qp_num;
2043         attr->qp_access_flags     = cmd->base.qp_access_flags;
2044         attr->pkey_index          = cmd->base.pkey_index;
2045         attr->alt_pkey_index      = cmd->base.alt_pkey_index;
2046         attr->en_sqd_async_notify = cmd->base.en_sqd_async_notify;
2047         attr->max_rd_atomic       = cmd->base.max_rd_atomic;
2048         attr->max_dest_rd_atomic  = cmd->base.max_dest_rd_atomic;
2049         attr->min_rnr_timer       = cmd->base.min_rnr_timer;
2050         attr->port_num            = cmd->base.port_num;
2051         attr->timeout             = cmd->base.timeout;
2052         attr->retry_cnt           = cmd->base.retry_cnt;
2053         attr->rnr_retry           = cmd->base.rnr_retry;
2054         attr->alt_port_num        = cmd->base.alt_port_num;
2055         attr->alt_timeout         = cmd->base.alt_timeout;
2056         attr->rate_limit          = cmd->rate_limit;
2057
2058         if (cmd->base.attr_mask & IB_QP_AV)
2059                 copy_ah_attr_from_uverbs(qp->device, &attr->ah_attr,
2060                                          &cmd->base.dest);
2061
2062         if (cmd->base.attr_mask & IB_QP_ALT_PATH)
2063                 copy_ah_attr_from_uverbs(qp->device, &attr->alt_ah_attr,
2064                                          &cmd->base.alt_dest);
2065
2066         ret = ib_modify_qp_with_udata(qp, attr,
2067                                       modify_qp_mask(qp->qp_type,
2068                                                      cmd->base.attr_mask),
2069                                       udata);
2070
2071 release_qp:
2072         uobj_put_obj_read(qp);
2073 out:
2074         kfree(attr);
2075
2076         return ret;
2077 }
2078
2079 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
2080                             const char __user *buf, int in_len,
2081                             int out_len)
2082 {
2083         struct ib_uverbs_ex_modify_qp cmd = {};
2084         struct ib_udata udata;
2085         int ret;
2086
2087         if (copy_from_user(&cmd.base, buf, sizeof(cmd.base)))
2088                 return -EFAULT;
2089
2090         if (cmd.base.attr_mask &
2091             ~((IB_USER_LEGACY_LAST_QP_ATTR_MASK << 1) - 1))
2092                 return -EOPNOTSUPP;
2093
2094         ib_uverbs_init_udata(&udata, buf + sizeof(cmd.base), NULL,
2095                    in_len - sizeof(cmd.base) - sizeof(struct ib_uverbs_cmd_hdr),
2096                    out_len);
2097
2098         ret = modify_qp(file, &cmd, &udata);
2099         if (ret)
2100                 return ret;
2101
2102         return in_len;
2103 }
2104
2105 int ib_uverbs_ex_modify_qp(struct ib_uverbs_file *file,
2106                            struct ib_udata *ucore,
2107                            struct ib_udata *uhw)
2108 {
2109         struct ib_uverbs_ex_modify_qp cmd = {};
2110         int ret;
2111
2112         /*
2113          * Last bit is reserved for extending the attr_mask by
2114          * using another field.
2115          */
2116         BUILD_BUG_ON(IB_USER_LAST_QP_ATTR_MASK == (1 << 31));
2117
2118         if (ucore->inlen < sizeof(cmd.base))
2119                 return -EINVAL;
2120
2121         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2122         if (ret)
2123                 return ret;
2124
2125         if (cmd.base.attr_mask &
2126             ~((IB_USER_LAST_QP_ATTR_MASK << 1) - 1))
2127                 return -EOPNOTSUPP;
2128
2129         if (ucore->inlen > sizeof(cmd)) {
2130                 if (!ib_is_udata_cleared(ucore, sizeof(cmd),
2131                                          ucore->inlen - sizeof(cmd)))
2132                         return -EOPNOTSUPP;
2133         }
2134
2135         ret = modify_qp(file, &cmd, uhw);
2136
2137         return ret;
2138 }
2139
2140 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
2141                              const char __user *buf, int in_len,
2142                              int out_len)
2143 {
2144         struct ib_uverbs_destroy_qp      cmd;
2145         struct ib_uverbs_destroy_qp_resp resp;
2146         struct ib_uobject               *uobj;
2147         struct ib_uqp_object            *obj;
2148
2149         if (copy_from_user(&cmd, buf, sizeof cmd))
2150                 return -EFAULT;
2151
2152         uobj = uobj_get_destroy(UVERBS_OBJECT_QP, cmd.qp_handle, file);
2153         if (IS_ERR(uobj))
2154                 return PTR_ERR(uobj);
2155
2156         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2157         memset(&resp, 0, sizeof(resp));
2158         resp.events_reported = obj->uevent.events_reported;
2159
2160         uobj_put_destroy(uobj);
2161
2162         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2163                 return -EFAULT;
2164
2165         return in_len;
2166 }
2167
2168 static void *alloc_wr(size_t wr_size, __u32 num_sge)
2169 {
2170         if (num_sge >= (U32_MAX - ALIGN(wr_size, sizeof (struct ib_sge))) /
2171                        sizeof (struct ib_sge))
2172                 return NULL;
2173
2174         return kmalloc(ALIGN(wr_size, sizeof (struct ib_sge)) +
2175                          num_sge * sizeof (struct ib_sge), GFP_KERNEL);
2176 }
2177
2178 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2179                             const char __user *buf, int in_len,
2180                             int out_len)
2181 {
2182         struct ib_uverbs_post_send      cmd;
2183         struct ib_uverbs_post_send_resp resp;
2184         struct ib_uverbs_send_wr       *user_wr;
2185         struct ib_send_wr              *wr = NULL, *last, *next;
2186         const struct ib_send_wr        *bad_wr;
2187         struct ib_qp                   *qp;
2188         int                             i, sg_ind;
2189         int                             is_ud;
2190         ssize_t                         ret = -EINVAL;
2191         size_t                          next_size;
2192
2193         if (copy_from_user(&cmd, buf, sizeof cmd))
2194                 return -EFAULT;
2195
2196         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2197             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2198                 return -EINVAL;
2199
2200         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2201                 return -EINVAL;
2202
2203         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2204         if (!user_wr)
2205                 return -ENOMEM;
2206
2207         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2208         if (!qp)
2209                 goto out;
2210
2211         is_ud = qp->qp_type == IB_QPT_UD;
2212         sg_ind = 0;
2213         last = NULL;
2214         for (i = 0; i < cmd.wr_count; ++i) {
2215                 if (copy_from_user(user_wr,
2216                                    buf + sizeof cmd + i * cmd.wqe_size,
2217                                    cmd.wqe_size)) {
2218                         ret = -EFAULT;
2219                         goto out_put;
2220                 }
2221
2222                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2223                         ret = -EINVAL;
2224                         goto out_put;
2225                 }
2226
2227                 if (is_ud) {
2228                         struct ib_ud_wr *ud;
2229
2230                         if (user_wr->opcode != IB_WR_SEND &&
2231                             user_wr->opcode != IB_WR_SEND_WITH_IMM) {
2232                                 ret = -EINVAL;
2233                                 goto out_put;
2234                         }
2235
2236                         next_size = sizeof(*ud);
2237                         ud = alloc_wr(next_size, user_wr->num_sge);
2238                         if (!ud) {
2239                                 ret = -ENOMEM;
2240                                 goto out_put;
2241                         }
2242
2243                         ud->ah = uobj_get_obj_read(ah, UVERBS_OBJECT_AH,
2244                                                    user_wr->wr.ud.ah, file);
2245                         if (!ud->ah) {
2246                                 kfree(ud);
2247                                 ret = -EINVAL;
2248                                 goto out_put;
2249                         }
2250                         ud->remote_qpn = user_wr->wr.ud.remote_qpn;
2251                         ud->remote_qkey = user_wr->wr.ud.remote_qkey;
2252
2253                         next = &ud->wr;
2254                 } else if (user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
2255                            user_wr->opcode == IB_WR_RDMA_WRITE ||
2256                            user_wr->opcode == IB_WR_RDMA_READ) {
2257                         struct ib_rdma_wr *rdma;
2258
2259                         next_size = sizeof(*rdma);
2260                         rdma = alloc_wr(next_size, user_wr->num_sge);
2261                         if (!rdma) {
2262                                 ret = -ENOMEM;
2263                                 goto out_put;
2264                         }
2265
2266                         rdma->remote_addr = user_wr->wr.rdma.remote_addr;
2267                         rdma->rkey = user_wr->wr.rdma.rkey;
2268
2269                         next = &rdma->wr;
2270                 } else if (user_wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2271                            user_wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2272                         struct ib_atomic_wr *atomic;
2273
2274                         next_size = sizeof(*atomic);
2275                         atomic = alloc_wr(next_size, user_wr->num_sge);
2276                         if (!atomic) {
2277                                 ret = -ENOMEM;
2278                                 goto out_put;
2279                         }
2280
2281                         atomic->remote_addr = user_wr->wr.atomic.remote_addr;
2282                         atomic->compare_add = user_wr->wr.atomic.compare_add;
2283                         atomic->swap = user_wr->wr.atomic.swap;
2284                         atomic->rkey = user_wr->wr.atomic.rkey;
2285
2286                         next = &atomic->wr;
2287                 } else if (user_wr->opcode == IB_WR_SEND ||
2288                            user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2289                            user_wr->opcode == IB_WR_SEND_WITH_INV) {
2290                         next_size = sizeof(*next);
2291                         next = alloc_wr(next_size, user_wr->num_sge);
2292                         if (!next) {
2293                                 ret = -ENOMEM;
2294                                 goto out_put;
2295                         }
2296                 } else {
2297                         ret = -EINVAL;
2298                         goto out_put;
2299                 }
2300
2301                 if (user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2302                     user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
2303                         next->ex.imm_data =
2304                                         (__be32 __force) user_wr->ex.imm_data;
2305                 } else if (user_wr->opcode == IB_WR_SEND_WITH_INV) {
2306                         next->ex.invalidate_rkey = user_wr->ex.invalidate_rkey;
2307                 }
2308
2309                 if (!last)
2310                         wr = next;
2311                 else
2312                         last->next = next;
2313                 last = next;
2314
2315                 next->next       = NULL;
2316                 next->wr_id      = user_wr->wr_id;
2317                 next->num_sge    = user_wr->num_sge;
2318                 next->opcode     = user_wr->opcode;
2319                 next->send_flags = user_wr->send_flags;
2320
2321                 if (next->num_sge) {
2322                         next->sg_list = (void *) next +
2323                                 ALIGN(next_size, sizeof(struct ib_sge));
2324                         if (copy_from_user(next->sg_list,
2325                                            buf + sizeof cmd +
2326                                            cmd.wr_count * cmd.wqe_size +
2327                                            sg_ind * sizeof (struct ib_sge),
2328                                            next->num_sge * sizeof (struct ib_sge))) {
2329                                 ret = -EFAULT;
2330                                 goto out_put;
2331                         }
2332                         sg_ind += next->num_sge;
2333                 } else
2334                         next->sg_list = NULL;
2335         }
2336
2337         resp.bad_wr = 0;
2338         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2339         if (ret)
2340                 for (next = wr; next; next = next->next) {
2341                         ++resp.bad_wr;
2342                         if (next == bad_wr)
2343                                 break;
2344                 }
2345
2346         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2347                 ret = -EFAULT;
2348
2349 out_put:
2350         uobj_put_obj_read(qp);
2351
2352         while (wr) {
2353                 if (is_ud && ud_wr(wr)->ah)
2354                         uobj_put_obj_read(ud_wr(wr)->ah);
2355                 next = wr->next;
2356                 kfree(wr);
2357                 wr = next;
2358         }
2359
2360 out:
2361         kfree(user_wr);
2362
2363         return ret ? ret : in_len;
2364 }
2365
2366 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2367                                                     int in_len,
2368                                                     u32 wr_count,
2369                                                     u32 sge_count,
2370                                                     u32 wqe_size)
2371 {
2372         struct ib_uverbs_recv_wr *user_wr;
2373         struct ib_recv_wr        *wr = NULL, *last, *next;
2374         int                       sg_ind;
2375         int                       i;
2376         int                       ret;
2377
2378         if (in_len < wqe_size * wr_count +
2379             sge_count * sizeof (struct ib_uverbs_sge))
2380                 return ERR_PTR(-EINVAL);
2381
2382         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2383                 return ERR_PTR(-EINVAL);
2384
2385         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2386         if (!user_wr)
2387                 return ERR_PTR(-ENOMEM);
2388
2389         sg_ind = 0;
2390         last = NULL;
2391         for (i = 0; i < wr_count; ++i) {
2392                 if (copy_from_user(user_wr, buf + i * wqe_size,
2393                                    wqe_size)) {
2394                         ret = -EFAULT;
2395                         goto err;
2396                 }
2397
2398                 if (user_wr->num_sge + sg_ind > sge_count) {
2399                         ret = -EINVAL;
2400                         goto err;
2401                 }
2402
2403                 if (user_wr->num_sge >=
2404                     (U32_MAX - ALIGN(sizeof *next, sizeof (struct ib_sge))) /
2405                     sizeof (struct ib_sge)) {
2406                         ret = -EINVAL;
2407                         goto err;
2408                 }
2409
2410                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2411                                user_wr->num_sge * sizeof (struct ib_sge),
2412                                GFP_KERNEL);
2413                 if (!next) {
2414                         ret = -ENOMEM;
2415                         goto err;
2416                 }
2417
2418                 if (!last)
2419                         wr = next;
2420                 else
2421                         last->next = next;
2422                 last = next;
2423
2424                 next->next       = NULL;
2425                 next->wr_id      = user_wr->wr_id;
2426                 next->num_sge    = user_wr->num_sge;
2427
2428                 if (next->num_sge) {
2429                         next->sg_list = (void *) next +
2430                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2431                         if (copy_from_user(next->sg_list,
2432                                            buf + wr_count * wqe_size +
2433                                            sg_ind * sizeof (struct ib_sge),
2434                                            next->num_sge * sizeof (struct ib_sge))) {
2435                                 ret = -EFAULT;
2436                                 goto err;
2437                         }
2438                         sg_ind += next->num_sge;
2439                 } else
2440                         next->sg_list = NULL;
2441         }
2442
2443         kfree(user_wr);
2444         return wr;
2445
2446 err:
2447         kfree(user_wr);
2448
2449         while (wr) {
2450                 next = wr->next;
2451                 kfree(wr);
2452                 wr = next;
2453         }
2454
2455         return ERR_PTR(ret);
2456 }
2457
2458 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2459                             const char __user *buf, int in_len,
2460                             int out_len)
2461 {
2462         struct ib_uverbs_post_recv      cmd;
2463         struct ib_uverbs_post_recv_resp resp;
2464         struct ib_recv_wr              *wr, *next;
2465         const struct ib_recv_wr        *bad_wr;
2466         struct ib_qp                   *qp;
2467         ssize_t                         ret = -EINVAL;
2468
2469         if (copy_from_user(&cmd, buf, sizeof cmd))
2470                 return -EFAULT;
2471
2472         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2473                                        in_len - sizeof cmd, cmd.wr_count,
2474                                        cmd.sge_count, cmd.wqe_size);
2475         if (IS_ERR(wr))
2476                 return PTR_ERR(wr);
2477
2478         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2479         if (!qp)
2480                 goto out;
2481
2482         resp.bad_wr = 0;
2483         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2484
2485         uobj_put_obj_read(qp);
2486         if (ret) {
2487                 for (next = wr; next; next = next->next) {
2488                         ++resp.bad_wr;
2489                         if (next == bad_wr)
2490                                 break;
2491                 }
2492         }
2493
2494         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2495                 ret = -EFAULT;
2496
2497 out:
2498         while (wr) {
2499                 next = wr->next;
2500                 kfree(wr);
2501                 wr = next;
2502         }
2503
2504         return ret ? ret : in_len;
2505 }
2506
2507 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2508                                 const char __user *buf, int in_len,
2509                                 int out_len)
2510 {
2511         struct ib_uverbs_post_srq_recv      cmd;
2512         struct ib_uverbs_post_srq_recv_resp resp;
2513         struct ib_recv_wr                  *wr, *next;
2514         const struct ib_recv_wr            *bad_wr;
2515         struct ib_srq                      *srq;
2516         ssize_t                             ret = -EINVAL;
2517
2518         if (copy_from_user(&cmd, buf, sizeof cmd))
2519                 return -EFAULT;
2520
2521         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2522                                        in_len - sizeof cmd, cmd.wr_count,
2523                                        cmd.sge_count, cmd.wqe_size);
2524         if (IS_ERR(wr))
2525                 return PTR_ERR(wr);
2526
2527         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
2528         if (!srq)
2529                 goto out;
2530
2531         resp.bad_wr = 0;
2532         ret = srq->device->post_srq_recv ?
2533                 srq->device->post_srq_recv(srq, wr, &bad_wr) : -EOPNOTSUPP;
2534
2535         uobj_put_obj_read(srq);
2536
2537         if (ret)
2538                 for (next = wr; next; next = next->next) {
2539                         ++resp.bad_wr;
2540                         if (next == bad_wr)
2541                                 break;
2542                 }
2543
2544         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2545                 ret = -EFAULT;
2546
2547 out:
2548         while (wr) {
2549                 next = wr->next;
2550                 kfree(wr);
2551                 wr = next;
2552         }
2553
2554         return ret ? ret : in_len;
2555 }
2556
2557 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2558                             const char __user *buf, int in_len,
2559                             int out_len)
2560 {
2561         struct ib_uverbs_create_ah       cmd;
2562         struct ib_uverbs_create_ah_resp  resp;
2563         struct ib_uobject               *uobj;
2564         struct ib_pd                    *pd;
2565         struct ib_ah                    *ah;
2566         struct rdma_ah_attr             attr = {};
2567         int ret;
2568         struct ib_udata                   udata;
2569         struct ib_device *ib_dev;
2570
2571         if (out_len < sizeof resp)
2572                 return -ENOSPC;
2573
2574         if (copy_from_user(&cmd, buf, sizeof cmd))
2575                 return -EFAULT;
2576
2577         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
2578                    u64_to_user_ptr(cmd.response) + sizeof(resp),
2579                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
2580                    out_len - sizeof(resp));
2581
2582         uobj = uobj_alloc(UVERBS_OBJECT_AH, file, &ib_dev);
2583         if (IS_ERR(uobj))
2584                 return PTR_ERR(uobj);
2585
2586         if (!rdma_is_port_valid(ib_dev, cmd.attr.port_num)) {
2587                 ret = -EINVAL;
2588                 goto err;
2589         }
2590
2591         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file);
2592         if (!pd) {
2593                 ret = -EINVAL;
2594                 goto err;
2595         }
2596
2597         attr.type = rdma_ah_find_type(ib_dev, cmd.attr.port_num);
2598         rdma_ah_set_make_grd(&attr, false);
2599         rdma_ah_set_dlid(&attr, cmd.attr.dlid);
2600         rdma_ah_set_sl(&attr, cmd.attr.sl);
2601         rdma_ah_set_path_bits(&attr, cmd.attr.src_path_bits);
2602         rdma_ah_set_static_rate(&attr, cmd.attr.static_rate);
2603         rdma_ah_set_port_num(&attr, cmd.attr.port_num);
2604
2605         if (cmd.attr.is_global) {
2606                 rdma_ah_set_grh(&attr, NULL, cmd.attr.grh.flow_label,
2607                                 cmd.attr.grh.sgid_index,
2608                                 cmd.attr.grh.hop_limit,
2609                                 cmd.attr.grh.traffic_class);
2610                 rdma_ah_set_dgid_raw(&attr, cmd.attr.grh.dgid);
2611         } else {
2612                 rdma_ah_set_ah_flags(&attr, 0);
2613         }
2614
2615         ah = rdma_create_user_ah(pd, &attr, &udata);
2616         if (IS_ERR(ah)) {
2617                 ret = PTR_ERR(ah);
2618                 goto err_put;
2619         }
2620
2621         ah->uobject  = uobj;
2622         uobj->user_handle = cmd.user_handle;
2623         uobj->object = ah;
2624
2625         resp.ah_handle = uobj->id;
2626
2627         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
2628                 ret = -EFAULT;
2629                 goto err_copy;
2630         }
2631
2632         uobj_put_obj_read(pd);
2633         return uobj_alloc_commit(uobj, in_len);
2634
2635 err_copy:
2636         rdma_destroy_ah(ah);
2637
2638 err_put:
2639         uobj_put_obj_read(pd);
2640
2641 err:
2642         uobj_alloc_abort(uobj);
2643         return ret;
2644 }
2645
2646 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2647                              const char __user *buf, int in_len, int out_len)
2648 {
2649         struct ib_uverbs_destroy_ah cmd;
2650
2651         if (copy_from_user(&cmd, buf, sizeof cmd))
2652                 return -EFAULT;
2653
2654         return uobj_perform_destroy(UVERBS_OBJECT_AH, cmd.ah_handle, file,
2655                                     in_len);
2656 }
2657
2658 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2659                                const char __user *buf, int in_len,
2660                                int out_len)
2661 {
2662         struct ib_uverbs_attach_mcast cmd;
2663         struct ib_qp                 *qp;
2664         struct ib_uqp_object         *obj;
2665         struct ib_uverbs_mcast_entry *mcast;
2666         int                           ret;
2667
2668         if (copy_from_user(&cmd, buf, sizeof cmd))
2669                 return -EFAULT;
2670
2671         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2672         if (!qp)
2673                 return -EINVAL;
2674
2675         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2676
2677         mutex_lock(&obj->mcast_lock);
2678         list_for_each_entry(mcast, &obj->mcast_list, list)
2679                 if (cmd.mlid == mcast->lid &&
2680                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2681                         ret = 0;
2682                         goto out_put;
2683                 }
2684
2685         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2686         if (!mcast) {
2687                 ret = -ENOMEM;
2688                 goto out_put;
2689         }
2690
2691         mcast->lid = cmd.mlid;
2692         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2693
2694         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2695         if (!ret)
2696                 list_add_tail(&mcast->list, &obj->mcast_list);
2697         else
2698                 kfree(mcast);
2699
2700 out_put:
2701         mutex_unlock(&obj->mcast_lock);
2702         uobj_put_obj_read(qp);
2703
2704         return ret ? ret : in_len;
2705 }
2706
2707 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2708                                const char __user *buf, int in_len,
2709                                int out_len)
2710 {
2711         struct ib_uverbs_detach_mcast cmd;
2712         struct ib_uqp_object         *obj;
2713         struct ib_qp                 *qp;
2714         struct ib_uverbs_mcast_entry *mcast;
2715         int                           ret = -EINVAL;
2716         bool                          found = false;
2717
2718         if (copy_from_user(&cmd, buf, sizeof cmd))
2719                 return -EFAULT;
2720
2721         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2722         if (!qp)
2723                 return -EINVAL;
2724
2725         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2726         mutex_lock(&obj->mcast_lock);
2727
2728         list_for_each_entry(mcast, &obj->mcast_list, list)
2729                 if (cmd.mlid == mcast->lid &&
2730                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2731                         list_del(&mcast->list);
2732                         kfree(mcast);
2733                         found = true;
2734                         break;
2735                 }
2736
2737         if (!found) {
2738                 ret = -EINVAL;
2739                 goto out_put;
2740         }
2741
2742         ret = ib_detach_mcast(qp, (union ib_gid *)cmd.gid, cmd.mlid);
2743
2744 out_put:
2745         mutex_unlock(&obj->mcast_lock);
2746         uobj_put_obj_read(qp);
2747         return ret ? ret : in_len;
2748 }
2749
2750 struct ib_uflow_resources {
2751         size_t                  max;
2752         size_t                  num;
2753         size_t                  collection_num;
2754         size_t                  counters_num;
2755         struct ib_counters      **counters;
2756         struct ib_flow_action   **collection;
2757 };
2758
2759 static struct ib_uflow_resources *flow_resources_alloc(size_t num_specs)
2760 {
2761         struct ib_uflow_resources *resources;
2762
2763         resources = kzalloc(sizeof(*resources), GFP_KERNEL);
2764
2765         if (!resources)
2766                 return NULL;
2767
2768         if (!num_specs)
2769                 goto out;
2770
2771         resources->counters =
2772                 kcalloc(num_specs, sizeof(*resources->counters), GFP_KERNEL);
2773         resources->collection =
2774                 kcalloc(num_specs, sizeof(*resources->collection), GFP_KERNEL);
2775
2776         if (!resources->counters || !resources->collection)
2777                 goto err;
2778
2779 out:
2780         resources->max = num_specs;
2781         return resources;
2782
2783 err:
2784         kfree(resources->counters);
2785         kfree(resources);
2786
2787         return NULL;
2788 }
2789
2790 void ib_uverbs_flow_resources_free(struct ib_uflow_resources *uflow_res)
2791 {
2792         unsigned int i;
2793
2794         if (!uflow_res)
2795                 return;
2796
2797         for (i = 0; i < uflow_res->collection_num; i++)
2798                 atomic_dec(&uflow_res->collection[i]->usecnt);
2799
2800         for (i = 0; i < uflow_res->counters_num; i++)
2801                 atomic_dec(&uflow_res->counters[i]->usecnt);
2802
2803         kfree(uflow_res->collection);
2804         kfree(uflow_res->counters);
2805         kfree(uflow_res);
2806 }
2807
2808 static void flow_resources_add(struct ib_uflow_resources *uflow_res,
2809                                enum ib_flow_spec_type type,
2810                                void *ibobj)
2811 {
2812         WARN_ON(uflow_res->num >= uflow_res->max);
2813
2814         switch (type) {
2815         case IB_FLOW_SPEC_ACTION_HANDLE:
2816                 atomic_inc(&((struct ib_flow_action *)ibobj)->usecnt);
2817                 uflow_res->collection[uflow_res->collection_num++] =
2818                         (struct ib_flow_action *)ibobj;
2819                 break;
2820         case IB_FLOW_SPEC_ACTION_COUNT:
2821                 atomic_inc(&((struct ib_counters *)ibobj)->usecnt);
2822                 uflow_res->counters[uflow_res->counters_num++] =
2823                         (struct ib_counters *)ibobj;
2824                 break;
2825         default:
2826                 WARN_ON(1);
2827         }
2828
2829         uflow_res->num++;
2830 }
2831
2832 static int kern_spec_to_ib_spec_action(struct ib_uverbs_file *ufile,
2833                                        struct ib_uverbs_flow_spec *kern_spec,
2834                                        union ib_flow_spec *ib_spec,
2835                                        struct ib_uflow_resources *uflow_res)
2836 {
2837         ib_spec->type = kern_spec->type;
2838         switch (ib_spec->type) {
2839         case IB_FLOW_SPEC_ACTION_TAG:
2840                 if (kern_spec->flow_tag.size !=
2841                     sizeof(struct ib_uverbs_flow_spec_action_tag))
2842                         return -EINVAL;
2843
2844                 ib_spec->flow_tag.size = sizeof(struct ib_flow_spec_action_tag);
2845                 ib_spec->flow_tag.tag_id = kern_spec->flow_tag.tag_id;
2846                 break;
2847         case IB_FLOW_SPEC_ACTION_DROP:
2848                 if (kern_spec->drop.size !=
2849                     sizeof(struct ib_uverbs_flow_spec_action_drop))
2850                         return -EINVAL;
2851
2852                 ib_spec->drop.size = sizeof(struct ib_flow_spec_action_drop);
2853                 break;
2854         case IB_FLOW_SPEC_ACTION_HANDLE:
2855                 if (kern_spec->action.size !=
2856                     sizeof(struct ib_uverbs_flow_spec_action_handle))
2857                         return -EOPNOTSUPP;
2858                 ib_spec->action.act = uobj_get_obj_read(flow_action,
2859                                                         UVERBS_OBJECT_FLOW_ACTION,
2860                                                         kern_spec->action.handle,
2861                                                         ufile);
2862                 if (!ib_spec->action.act)
2863                         return -EINVAL;
2864                 ib_spec->action.size =
2865                         sizeof(struct ib_flow_spec_action_handle);
2866                 flow_resources_add(uflow_res,
2867                                    IB_FLOW_SPEC_ACTION_HANDLE,
2868                                    ib_spec->action.act);
2869                 uobj_put_obj_read(ib_spec->action.act);
2870                 break;
2871         case IB_FLOW_SPEC_ACTION_COUNT:
2872                 if (kern_spec->flow_count.size !=
2873                         sizeof(struct ib_uverbs_flow_spec_action_count))
2874                         return -EINVAL;
2875                 ib_spec->flow_count.counters =
2876                         uobj_get_obj_read(counters,
2877                                           UVERBS_OBJECT_COUNTERS,
2878                                           kern_spec->flow_count.handle,
2879                                           ufile);
2880                 if (!ib_spec->flow_count.counters)
2881                         return -EINVAL;
2882                 ib_spec->flow_count.size =
2883                                 sizeof(struct ib_flow_spec_action_count);
2884                 flow_resources_add(uflow_res,
2885                                    IB_FLOW_SPEC_ACTION_COUNT,
2886                                    ib_spec->flow_count.counters);
2887                 uobj_put_obj_read(ib_spec->flow_count.counters);
2888                 break;
2889         default:
2890                 return -EINVAL;
2891         }
2892         return 0;
2893 }
2894
2895 static size_t kern_spec_filter_sz(const struct ib_uverbs_flow_spec_hdr *spec)
2896 {
2897         /* Returns user space filter size, includes padding */
2898         return (spec->size - sizeof(struct ib_uverbs_flow_spec_hdr)) / 2;
2899 }
2900
2901 static ssize_t spec_filter_size(const void *kern_spec_filter, u16 kern_filter_size,
2902                                 u16 ib_real_filter_sz)
2903 {
2904         /*
2905          * User space filter structures must be 64 bit aligned, otherwise this
2906          * may pass, but we won't handle additional new attributes.
2907          */
2908
2909         if (kern_filter_size > ib_real_filter_sz) {
2910                 if (memchr_inv(kern_spec_filter +
2911                                ib_real_filter_sz, 0,
2912                                kern_filter_size - ib_real_filter_sz))
2913                         return -EINVAL;
2914                 return ib_real_filter_sz;
2915         }
2916         return kern_filter_size;
2917 }
2918
2919 int ib_uverbs_kern_spec_to_ib_spec_filter(enum ib_flow_spec_type type,
2920                                           const void *kern_spec_mask,
2921                                           const void *kern_spec_val,
2922                                           size_t kern_filter_sz,
2923                                           union ib_flow_spec *ib_spec)
2924 {
2925         ssize_t actual_filter_sz;
2926         ssize_t ib_filter_sz;
2927
2928         /* User flow spec size must be aligned to 4 bytes */
2929         if (kern_filter_sz != ALIGN(kern_filter_sz, 4))
2930                 return -EINVAL;
2931
2932         ib_spec->type = type;
2933
2934         if (ib_spec->type == (IB_FLOW_SPEC_INNER | IB_FLOW_SPEC_VXLAN_TUNNEL))
2935                 return -EINVAL;
2936
2937         switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2938         case IB_FLOW_SPEC_ETH:
2939                 ib_filter_sz = offsetof(struct ib_flow_eth_filter, real_sz);
2940                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2941                                                     kern_filter_sz,
2942                                                     ib_filter_sz);
2943                 if (actual_filter_sz <= 0)
2944                         return -EINVAL;
2945                 ib_spec->size = sizeof(struct ib_flow_spec_eth);
2946                 memcpy(&ib_spec->eth.val, kern_spec_val, actual_filter_sz);
2947                 memcpy(&ib_spec->eth.mask, kern_spec_mask, actual_filter_sz);
2948                 break;
2949         case IB_FLOW_SPEC_IPV4:
2950                 ib_filter_sz = offsetof(struct ib_flow_ipv4_filter, real_sz);
2951                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2952                                                     kern_filter_sz,
2953                                                     ib_filter_sz);
2954                 if (actual_filter_sz <= 0)
2955                         return -EINVAL;
2956                 ib_spec->size = sizeof(struct ib_flow_spec_ipv4);
2957                 memcpy(&ib_spec->ipv4.val, kern_spec_val, actual_filter_sz);
2958                 memcpy(&ib_spec->ipv4.mask, kern_spec_mask, actual_filter_sz);
2959                 break;
2960         case IB_FLOW_SPEC_IPV6:
2961                 ib_filter_sz = offsetof(struct ib_flow_ipv6_filter, real_sz);
2962                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2963                                                     kern_filter_sz,
2964                                                     ib_filter_sz);
2965                 if (actual_filter_sz <= 0)
2966                         return -EINVAL;
2967                 ib_spec->size = sizeof(struct ib_flow_spec_ipv6);
2968                 memcpy(&ib_spec->ipv6.val, kern_spec_val, actual_filter_sz);
2969                 memcpy(&ib_spec->ipv6.mask, kern_spec_mask, actual_filter_sz);
2970
2971                 if ((ntohl(ib_spec->ipv6.mask.flow_label)) >= BIT(20) ||
2972                     (ntohl(ib_spec->ipv6.val.flow_label)) >= BIT(20))
2973                         return -EINVAL;
2974                 break;
2975         case IB_FLOW_SPEC_TCP:
2976         case IB_FLOW_SPEC_UDP:
2977                 ib_filter_sz = offsetof(struct ib_flow_tcp_udp_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_tcp_udp);
2984                 memcpy(&ib_spec->tcp_udp.val, kern_spec_val, actual_filter_sz);
2985                 memcpy(&ib_spec->tcp_udp.mask, kern_spec_mask, actual_filter_sz);
2986                 break;
2987         case IB_FLOW_SPEC_VXLAN_TUNNEL:
2988                 ib_filter_sz = offsetof(struct ib_flow_tunnel_filter, real_sz);
2989                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2990                                                     kern_filter_sz,
2991                                                     ib_filter_sz);
2992                 if (actual_filter_sz <= 0)
2993                         return -EINVAL;
2994                 ib_spec->tunnel.size = sizeof(struct ib_flow_spec_tunnel);
2995                 memcpy(&ib_spec->tunnel.val, kern_spec_val, actual_filter_sz);
2996                 memcpy(&ib_spec->tunnel.mask, kern_spec_mask, actual_filter_sz);
2997
2998                 if ((ntohl(ib_spec->tunnel.mask.tunnel_id)) >= BIT(24) ||
2999                     (ntohl(ib_spec->tunnel.val.tunnel_id)) >= BIT(24))
3000                         return -EINVAL;
3001                 break;
3002         case IB_FLOW_SPEC_ESP:
3003                 ib_filter_sz = offsetof(struct ib_flow_esp_filter, real_sz);
3004                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3005                                                     kern_filter_sz,
3006                                                     ib_filter_sz);
3007                 if (actual_filter_sz <= 0)
3008                         return -EINVAL;
3009                 ib_spec->esp.size = sizeof(struct ib_flow_spec_esp);
3010                 memcpy(&ib_spec->esp.val, kern_spec_val, actual_filter_sz);
3011                 memcpy(&ib_spec->esp.mask, kern_spec_mask, actual_filter_sz);
3012                 break;
3013         case IB_FLOW_SPEC_GRE:
3014                 ib_filter_sz = offsetof(struct ib_flow_gre_filter, real_sz);
3015                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3016                                                     kern_filter_sz,
3017                                                     ib_filter_sz);
3018                 if (actual_filter_sz <= 0)
3019                         return -EINVAL;
3020                 ib_spec->gre.size = sizeof(struct ib_flow_spec_gre);
3021                 memcpy(&ib_spec->gre.val, kern_spec_val, actual_filter_sz);
3022                 memcpy(&ib_spec->gre.mask, kern_spec_mask, actual_filter_sz);
3023                 break;
3024         case IB_FLOW_SPEC_MPLS:
3025                 ib_filter_sz = offsetof(struct ib_flow_mpls_filter, real_sz);
3026                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3027                                                     kern_filter_sz,
3028                                                     ib_filter_sz);
3029                 if (actual_filter_sz <= 0)
3030                         return -EINVAL;
3031                 ib_spec->mpls.size = sizeof(struct ib_flow_spec_mpls);
3032                 memcpy(&ib_spec->mpls.val, kern_spec_val, actual_filter_sz);
3033                 memcpy(&ib_spec->mpls.mask, kern_spec_mask, actual_filter_sz);
3034                 break;
3035         default:
3036                 return -EINVAL;
3037         }
3038         return 0;
3039 }
3040
3041 static int kern_spec_to_ib_spec_filter(struct ib_uverbs_flow_spec *kern_spec,
3042                                        union ib_flow_spec *ib_spec)
3043 {
3044         ssize_t kern_filter_sz;
3045         void *kern_spec_mask;
3046         void *kern_spec_val;
3047
3048         kern_filter_sz = kern_spec_filter_sz(&kern_spec->hdr);
3049
3050         kern_spec_val = (void *)kern_spec +
3051                 sizeof(struct ib_uverbs_flow_spec_hdr);
3052         kern_spec_mask = kern_spec_val + kern_filter_sz;
3053
3054         return ib_uverbs_kern_spec_to_ib_spec_filter(kern_spec->type,
3055                                                      kern_spec_mask,
3056                                                      kern_spec_val,
3057                                                      kern_filter_sz, ib_spec);
3058 }
3059
3060 static int kern_spec_to_ib_spec(struct ib_uverbs_file *ufile,
3061                                 struct ib_uverbs_flow_spec *kern_spec,
3062                                 union ib_flow_spec *ib_spec,
3063                                 struct ib_uflow_resources *uflow_res)
3064 {
3065         if (kern_spec->reserved)
3066                 return -EINVAL;
3067
3068         if (kern_spec->type >= IB_FLOW_SPEC_ACTION_TAG)
3069                 return kern_spec_to_ib_spec_action(ufile, kern_spec, ib_spec,
3070                                                    uflow_res);
3071         else
3072                 return kern_spec_to_ib_spec_filter(kern_spec, ib_spec);
3073 }
3074
3075 int ib_uverbs_ex_create_wq(struct ib_uverbs_file *file,
3076                            struct ib_udata *ucore,
3077                            struct ib_udata *uhw)
3078 {
3079         struct ib_uverbs_ex_create_wq     cmd = {};
3080         struct ib_uverbs_ex_create_wq_resp resp = {};
3081         struct ib_uwq_object           *obj;
3082         int err = 0;
3083         struct ib_cq *cq;
3084         struct ib_pd *pd;
3085         struct ib_wq *wq;
3086         struct ib_wq_init_attr wq_init_attr = {};
3087         size_t required_cmd_sz;
3088         size_t required_resp_len;
3089         struct ib_device *ib_dev;
3090
3091         required_cmd_sz = offsetof(typeof(cmd), max_sge) + sizeof(cmd.max_sge);
3092         required_resp_len = offsetof(typeof(resp), wqn) + sizeof(resp.wqn);
3093
3094         if (ucore->inlen < required_cmd_sz)
3095                 return -EINVAL;
3096
3097         if (ucore->outlen < required_resp_len)
3098                 return -ENOSPC;
3099
3100         if (ucore->inlen > sizeof(cmd) &&
3101             !ib_is_udata_cleared(ucore, sizeof(cmd),
3102                                  ucore->inlen - sizeof(cmd)))
3103                 return -EOPNOTSUPP;
3104
3105         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3106         if (err)
3107                 return err;
3108
3109         if (cmd.comp_mask)
3110                 return -EOPNOTSUPP;
3111
3112         obj = (struct ib_uwq_object *)uobj_alloc(UVERBS_OBJECT_WQ, file,
3113                                                  &ib_dev);
3114         if (IS_ERR(obj))
3115                 return PTR_ERR(obj);
3116
3117         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file);
3118         if (!pd) {
3119                 err = -EINVAL;
3120                 goto err_uobj;
3121         }
3122
3123         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
3124         if (!cq) {
3125                 err = -EINVAL;
3126                 goto err_put_pd;
3127         }
3128
3129         wq_init_attr.cq = cq;
3130         wq_init_attr.max_sge = cmd.max_sge;
3131         wq_init_attr.max_wr = cmd.max_wr;
3132         wq_init_attr.wq_context = file;
3133         wq_init_attr.wq_type = cmd.wq_type;
3134         wq_init_attr.event_handler = ib_uverbs_wq_event_handler;
3135         if (ucore->inlen >= (offsetof(typeof(cmd), create_flags) +
3136                              sizeof(cmd.create_flags)))
3137                 wq_init_attr.create_flags = cmd.create_flags;
3138         obj->uevent.events_reported = 0;
3139         INIT_LIST_HEAD(&obj->uevent.event_list);
3140
3141         if (!pd->device->create_wq) {
3142                 err = -EOPNOTSUPP;
3143                 goto err_put_cq;
3144         }
3145         wq = pd->device->create_wq(pd, &wq_init_attr, uhw);
3146         if (IS_ERR(wq)) {
3147                 err = PTR_ERR(wq);
3148                 goto err_put_cq;
3149         }
3150
3151         wq->uobject = &obj->uevent.uobject;
3152         obj->uevent.uobject.object = wq;
3153         wq->wq_type = wq_init_attr.wq_type;
3154         wq->cq = cq;
3155         wq->pd = pd;
3156         wq->device = pd->device;
3157         wq->wq_context = wq_init_attr.wq_context;
3158         atomic_set(&wq->usecnt, 0);
3159         atomic_inc(&pd->usecnt);
3160         atomic_inc(&cq->usecnt);
3161         wq->uobject = &obj->uevent.uobject;
3162         obj->uevent.uobject.object = wq;
3163
3164         memset(&resp, 0, sizeof(resp));
3165         resp.wq_handle = obj->uevent.uobject.id;
3166         resp.max_sge = wq_init_attr.max_sge;
3167         resp.max_wr = wq_init_attr.max_wr;
3168         resp.wqn = wq->wq_num;
3169         resp.response_length = required_resp_len;
3170         err = ib_copy_to_udata(ucore,
3171                                &resp, resp.response_length);
3172         if (err)
3173                 goto err_copy;
3174
3175         uobj_put_obj_read(pd);
3176         uobj_put_obj_read(cq);
3177         return uobj_alloc_commit(&obj->uevent.uobject, 0);
3178
3179 err_copy:
3180         ib_destroy_wq(wq);
3181 err_put_cq:
3182         uobj_put_obj_read(cq);
3183 err_put_pd:
3184         uobj_put_obj_read(pd);
3185 err_uobj:
3186         uobj_alloc_abort(&obj->uevent.uobject);
3187
3188         return err;
3189 }
3190
3191 int ib_uverbs_ex_destroy_wq(struct ib_uverbs_file *file,
3192                             struct ib_udata *ucore,
3193                             struct ib_udata *uhw)
3194 {
3195         struct ib_uverbs_ex_destroy_wq  cmd = {};
3196         struct ib_uverbs_ex_destroy_wq_resp     resp = {};
3197         struct ib_uobject               *uobj;
3198         struct ib_uwq_object            *obj;
3199         size_t required_cmd_sz;
3200         size_t required_resp_len;
3201         int                             ret;
3202
3203         required_cmd_sz = offsetof(typeof(cmd), wq_handle) + sizeof(cmd.wq_handle);
3204         required_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
3205
3206         if (ucore->inlen < required_cmd_sz)
3207                 return -EINVAL;
3208
3209         if (ucore->outlen < required_resp_len)
3210                 return -ENOSPC;
3211
3212         if (ucore->inlen > sizeof(cmd) &&
3213             !ib_is_udata_cleared(ucore, sizeof(cmd),
3214                                  ucore->inlen - sizeof(cmd)))
3215                 return -EOPNOTSUPP;
3216
3217         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3218         if (ret)
3219                 return ret;
3220
3221         if (cmd.comp_mask)
3222                 return -EOPNOTSUPP;
3223
3224         resp.response_length = required_resp_len;
3225         uobj = uobj_get_destroy(UVERBS_OBJECT_WQ, cmd.wq_handle, file);
3226         if (IS_ERR(uobj))
3227                 return PTR_ERR(uobj);
3228
3229         obj = container_of(uobj, struct ib_uwq_object, uevent.uobject);
3230         resp.events_reported = obj->uevent.events_reported;
3231
3232         uobj_put_destroy(uobj);
3233
3234         return ib_copy_to_udata(ucore, &resp, resp.response_length);
3235 }
3236
3237 int ib_uverbs_ex_modify_wq(struct ib_uverbs_file *file,
3238                            struct ib_udata *ucore,
3239                            struct ib_udata *uhw)
3240 {
3241         struct ib_uverbs_ex_modify_wq cmd = {};
3242         struct ib_wq *wq;
3243         struct ib_wq_attr wq_attr = {};
3244         size_t required_cmd_sz;
3245         int ret;
3246
3247         required_cmd_sz = offsetof(typeof(cmd), curr_wq_state) + sizeof(cmd.curr_wq_state);
3248         if (ucore->inlen < required_cmd_sz)
3249                 return -EINVAL;
3250
3251         if (ucore->inlen > sizeof(cmd) &&
3252             !ib_is_udata_cleared(ucore, sizeof(cmd),
3253                                  ucore->inlen - sizeof(cmd)))
3254                 return -EOPNOTSUPP;
3255
3256         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3257         if (ret)
3258                 return ret;
3259
3260         if (!cmd.attr_mask)
3261                 return -EINVAL;
3262
3263         if (cmd.attr_mask > (IB_WQ_STATE | IB_WQ_CUR_STATE | IB_WQ_FLAGS))
3264                 return -EINVAL;
3265
3266         wq = uobj_get_obj_read(wq, UVERBS_OBJECT_WQ, cmd.wq_handle, file);
3267         if (!wq)
3268                 return -EINVAL;
3269
3270         wq_attr.curr_wq_state = cmd.curr_wq_state;
3271         wq_attr.wq_state = cmd.wq_state;
3272         if (cmd.attr_mask & IB_WQ_FLAGS) {
3273                 wq_attr.flags = cmd.flags;
3274                 wq_attr.flags_mask = cmd.flags_mask;
3275         }
3276         if (!wq->device->modify_wq) {
3277                 ret = -EOPNOTSUPP;
3278                 goto out;
3279         }
3280         ret = wq->device->modify_wq(wq, &wq_attr, cmd.attr_mask, uhw);
3281 out:
3282         uobj_put_obj_read(wq);
3283         return ret;
3284 }
3285
3286 int ib_uverbs_ex_create_rwq_ind_table(struct ib_uverbs_file *file,
3287                                       struct ib_udata *ucore,
3288                                       struct ib_udata *uhw)
3289 {
3290         struct ib_uverbs_ex_create_rwq_ind_table          cmd = {};
3291         struct ib_uverbs_ex_create_rwq_ind_table_resp  resp = {};
3292         struct ib_uobject                 *uobj;
3293         int err = 0;
3294         struct ib_rwq_ind_table_init_attr init_attr = {};
3295         struct ib_rwq_ind_table *rwq_ind_tbl;
3296         struct ib_wq    **wqs = NULL;
3297         u32 *wqs_handles = NULL;
3298         struct ib_wq    *wq = NULL;
3299         int i, j, num_read_wqs;
3300         u32 num_wq_handles;
3301         u32 expected_in_size;
3302         size_t required_cmd_sz_header;
3303         size_t required_resp_len;
3304         struct ib_device *ib_dev;
3305
3306         required_cmd_sz_header = offsetof(typeof(cmd), log_ind_tbl_size) + sizeof(cmd.log_ind_tbl_size);
3307         required_resp_len = offsetof(typeof(resp), ind_tbl_num) + sizeof(resp.ind_tbl_num);
3308
3309         if (ucore->inlen < required_cmd_sz_header)
3310                 return -EINVAL;
3311
3312         if (ucore->outlen < required_resp_len)
3313                 return -ENOSPC;
3314
3315         err = ib_copy_from_udata(&cmd, ucore, required_cmd_sz_header);
3316         if (err)
3317                 return err;
3318
3319         ucore->inbuf += required_cmd_sz_header;
3320         ucore->inlen -= required_cmd_sz_header;
3321
3322         if (cmd.comp_mask)
3323                 return -EOPNOTSUPP;
3324
3325         if (cmd.log_ind_tbl_size > IB_USER_VERBS_MAX_LOG_IND_TBL_SIZE)
3326                 return -EINVAL;
3327
3328         num_wq_handles = 1 << cmd.log_ind_tbl_size;
3329         expected_in_size = num_wq_handles * sizeof(__u32);
3330         if (num_wq_handles == 1)
3331                 /* input size for wq handles is u64 aligned */
3332                 expected_in_size += sizeof(__u32);
3333
3334         if (ucore->inlen < expected_in_size)
3335                 return -EINVAL;
3336
3337         if (ucore->inlen > expected_in_size &&
3338             !ib_is_udata_cleared(ucore, expected_in_size,
3339                                  ucore->inlen - expected_in_size))
3340                 return -EOPNOTSUPP;
3341
3342         wqs_handles = kcalloc(num_wq_handles, sizeof(*wqs_handles),
3343                               GFP_KERNEL);
3344         if (!wqs_handles)
3345                 return -ENOMEM;
3346
3347         err = ib_copy_from_udata(wqs_handles, ucore,
3348                                  num_wq_handles * sizeof(__u32));
3349         if (err)
3350                 goto err_free;
3351
3352         wqs = kcalloc(num_wq_handles, sizeof(*wqs), GFP_KERNEL);
3353         if (!wqs) {
3354                 err = -ENOMEM;
3355                 goto  err_free;
3356         }
3357
3358         for (num_read_wqs = 0; num_read_wqs < num_wq_handles;
3359                         num_read_wqs++) {
3360                 wq = uobj_get_obj_read(wq, UVERBS_OBJECT_WQ,
3361                                        wqs_handles[num_read_wqs], file);
3362                 if (!wq) {
3363                         err = -EINVAL;
3364                         goto put_wqs;
3365                 }
3366
3367                 wqs[num_read_wqs] = wq;
3368         }
3369
3370         uobj = uobj_alloc(UVERBS_OBJECT_RWQ_IND_TBL, file, &ib_dev);
3371         if (IS_ERR(uobj)) {
3372                 err = PTR_ERR(uobj);
3373                 goto put_wqs;
3374         }
3375
3376         init_attr.log_ind_tbl_size = cmd.log_ind_tbl_size;
3377         init_attr.ind_tbl = wqs;
3378
3379         if (!ib_dev->create_rwq_ind_table) {
3380                 err = -EOPNOTSUPP;
3381                 goto err_uobj;
3382         }
3383         rwq_ind_tbl = ib_dev->create_rwq_ind_table(ib_dev, &init_attr, uhw);
3384
3385         if (IS_ERR(rwq_ind_tbl)) {
3386                 err = PTR_ERR(rwq_ind_tbl);
3387                 goto err_uobj;
3388         }
3389
3390         rwq_ind_tbl->ind_tbl = wqs;
3391         rwq_ind_tbl->log_ind_tbl_size = init_attr.log_ind_tbl_size;
3392         rwq_ind_tbl->uobject = uobj;
3393         uobj->object = rwq_ind_tbl;
3394         rwq_ind_tbl->device = ib_dev;
3395         atomic_set(&rwq_ind_tbl->usecnt, 0);
3396
3397         for (i = 0; i < num_wq_handles; i++)
3398                 atomic_inc(&wqs[i]->usecnt);
3399
3400         resp.ind_tbl_handle = uobj->id;
3401         resp.ind_tbl_num = rwq_ind_tbl->ind_tbl_num;
3402         resp.response_length = required_resp_len;
3403
3404         err = ib_copy_to_udata(ucore,
3405                                &resp, resp.response_length);
3406         if (err)
3407                 goto err_copy;
3408
3409         kfree(wqs_handles);
3410
3411         for (j = 0; j < num_read_wqs; j++)
3412                 uobj_put_obj_read(wqs[j]);
3413
3414         return uobj_alloc_commit(uobj, 0);
3415
3416 err_copy:
3417         ib_destroy_rwq_ind_table(rwq_ind_tbl);
3418 err_uobj:
3419         uobj_alloc_abort(uobj);
3420 put_wqs:
3421         for (j = 0; j < num_read_wqs; j++)
3422                 uobj_put_obj_read(wqs[j]);
3423 err_free:
3424         kfree(wqs_handles);
3425         kfree(wqs);
3426         return err;
3427 }
3428
3429 int ib_uverbs_ex_destroy_rwq_ind_table(struct ib_uverbs_file *file,
3430                                        struct ib_udata *ucore,
3431                                        struct ib_udata *uhw)
3432 {
3433         struct ib_uverbs_ex_destroy_rwq_ind_table       cmd = {};
3434         int                     ret;
3435         size_t required_cmd_sz;
3436
3437         required_cmd_sz = offsetof(typeof(cmd), ind_tbl_handle) + sizeof(cmd.ind_tbl_handle);
3438
3439         if (ucore->inlen < required_cmd_sz)
3440                 return -EINVAL;
3441
3442         if (ucore->inlen > sizeof(cmd) &&
3443             !ib_is_udata_cleared(ucore, sizeof(cmd),
3444                                  ucore->inlen - sizeof(cmd)))
3445                 return -EOPNOTSUPP;
3446
3447         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3448         if (ret)
3449                 return ret;
3450
3451         if (cmd.comp_mask)
3452                 return -EOPNOTSUPP;
3453
3454         return uobj_perform_destroy(UVERBS_OBJECT_RWQ_IND_TBL,
3455                                     cmd.ind_tbl_handle, file, 0);
3456 }
3457
3458 int ib_uverbs_ex_create_flow(struct ib_uverbs_file *file,
3459                              struct ib_udata *ucore,
3460                              struct ib_udata *uhw)
3461 {
3462         struct ib_uverbs_create_flow      cmd;
3463         struct ib_uverbs_create_flow_resp resp;
3464         struct ib_uobject                 *uobj;
3465         struct ib_uflow_object            *uflow;
3466         struct ib_flow                    *flow_id;
3467         struct ib_uverbs_flow_attr        *kern_flow_attr;
3468         struct ib_flow_attr               *flow_attr;
3469         struct ib_qp                      *qp;
3470         struct ib_uflow_resources         *uflow_res;
3471         struct ib_uverbs_flow_spec_hdr    *kern_spec;
3472         int err = 0;
3473         void *ib_spec;
3474         int i;
3475         struct ib_device *ib_dev;
3476
3477         if (ucore->inlen < sizeof(cmd))
3478                 return -EINVAL;
3479
3480         if (ucore->outlen < sizeof(resp))
3481                 return -ENOSPC;
3482
3483         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3484         if (err)
3485                 return err;
3486
3487         ucore->inbuf += sizeof(cmd);
3488         ucore->inlen -= sizeof(cmd);
3489
3490         if (cmd.comp_mask)
3491                 return -EINVAL;
3492
3493         if (!capable(CAP_NET_RAW))
3494                 return -EPERM;
3495
3496         if (cmd.flow_attr.flags >= IB_FLOW_ATTR_FLAGS_RESERVED)
3497                 return -EINVAL;
3498
3499         if ((cmd.flow_attr.flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) &&
3500             ((cmd.flow_attr.type == IB_FLOW_ATTR_ALL_DEFAULT) ||
3501              (cmd.flow_attr.type == IB_FLOW_ATTR_MC_DEFAULT)))
3502                 return -EINVAL;
3503
3504         if (cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
3505                 return -EINVAL;
3506
3507         if (cmd.flow_attr.size > ucore->inlen ||
3508             cmd.flow_attr.size >
3509             (cmd.flow_attr.num_of_specs * sizeof(struct ib_uverbs_flow_spec)))
3510                 return -EINVAL;
3511
3512         if (cmd.flow_attr.reserved[0] ||
3513             cmd.flow_attr.reserved[1])
3514                 return -EINVAL;
3515
3516         if (cmd.flow_attr.num_of_specs) {
3517                 kern_flow_attr = kmalloc(sizeof(*kern_flow_attr) + cmd.flow_attr.size,
3518                                          GFP_KERNEL);
3519                 if (!kern_flow_attr)
3520                         return -ENOMEM;
3521
3522                 *kern_flow_attr = cmd.flow_attr;
3523                 err = ib_copy_from_udata(&kern_flow_attr->flow_specs, ucore,
3524                                          cmd.flow_attr.size);
3525                 if (err)
3526                         goto err_free_attr;
3527         } else {
3528                 kern_flow_attr = &cmd.flow_attr;
3529         }
3530
3531         uobj = uobj_alloc(UVERBS_OBJECT_FLOW, file, &ib_dev);
3532         if (IS_ERR(uobj)) {
3533                 err = PTR_ERR(uobj);
3534                 goto err_free_attr;
3535         }
3536
3537         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
3538         if (!qp) {
3539                 err = -EINVAL;
3540                 goto err_uobj;
3541         }
3542
3543         if (qp->qp_type != IB_QPT_UD && qp->qp_type != IB_QPT_RAW_PACKET) {
3544                 err = -EINVAL;
3545                 goto err_put;
3546         }
3547
3548         if (!qp->device->create_flow) {
3549                 err = -EOPNOTSUPP;
3550                 goto err_put;
3551         }
3552
3553         flow_attr = kzalloc(struct_size(flow_attr, flows,
3554                                 cmd.flow_attr.num_of_specs), GFP_KERNEL);
3555         if (!flow_attr) {
3556                 err = -ENOMEM;
3557                 goto err_put;
3558         }
3559         uflow_res = flow_resources_alloc(cmd.flow_attr.num_of_specs);
3560         if (!uflow_res) {
3561                 err = -ENOMEM;
3562                 goto err_free_flow_attr;
3563         }
3564
3565         flow_attr->type = kern_flow_attr->type;
3566         flow_attr->priority = kern_flow_attr->priority;
3567         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
3568         flow_attr->port = kern_flow_attr->port;
3569         flow_attr->flags = kern_flow_attr->flags;
3570         flow_attr->size = sizeof(*flow_attr);
3571
3572         kern_spec = kern_flow_attr->flow_specs;
3573         ib_spec = flow_attr + 1;
3574         for (i = 0; i < flow_attr->num_of_specs &&
3575                         cmd.flow_attr.size >= sizeof(*kern_spec) &&
3576                         cmd.flow_attr.size >= kern_spec->size;
3577              i++) {
3578                 err = kern_spec_to_ib_spec(
3579                                 file, (struct ib_uverbs_flow_spec *)kern_spec,
3580                                 ib_spec, uflow_res);
3581                 if (err)
3582                         goto err_free;
3583
3584                 flow_attr->size +=
3585                         ((union ib_flow_spec *) ib_spec)->size;
3586                 cmd.flow_attr.size -= kern_spec->size;
3587                 kern_spec = ((void *)kern_spec) + kern_spec->size;
3588                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
3589         }
3590         if (cmd.flow_attr.size || (i != flow_attr->num_of_specs)) {
3591                 pr_warn("create flow failed, flow %d: %d bytes left from uverb cmd\n",
3592                         i, cmd.flow_attr.size);
3593                 err = -EINVAL;
3594                 goto err_free;
3595         }
3596
3597         flow_id = qp->device->create_flow(qp, flow_attr,
3598                                           IB_FLOW_DOMAIN_USER, uhw);
3599
3600         if (IS_ERR(flow_id)) {
3601                 err = PTR_ERR(flow_id);
3602                 goto err_free;
3603         }
3604         atomic_inc(&qp->usecnt);
3605         flow_id->qp = qp;
3606         flow_id->device = qp->device;
3607         flow_id->uobject = uobj;
3608         uobj->object = flow_id;
3609         uflow = container_of(uobj, typeof(*uflow), uobject);
3610         uflow->resources = uflow_res;
3611
3612         memset(&resp, 0, sizeof(resp));
3613         resp.flow_handle = uobj->id;
3614
3615         err = ib_copy_to_udata(ucore,
3616                                &resp, sizeof(resp));
3617         if (err)
3618                 goto err_copy;
3619
3620         uobj_put_obj_read(qp);
3621         kfree(flow_attr);
3622         if (cmd.flow_attr.num_of_specs)
3623                 kfree(kern_flow_attr);
3624         return uobj_alloc_commit(uobj, 0);
3625 err_copy:
3626         if (!qp->device->destroy_flow(flow_id))
3627                 atomic_dec(&qp->usecnt);
3628 err_free:
3629         ib_uverbs_flow_resources_free(uflow_res);
3630 err_free_flow_attr:
3631         kfree(flow_attr);
3632 err_put:
3633         uobj_put_obj_read(qp);
3634 err_uobj:
3635         uobj_alloc_abort(uobj);
3636 err_free_attr:
3637         if (cmd.flow_attr.num_of_specs)
3638                 kfree(kern_flow_attr);
3639         return err;
3640 }
3641
3642 int ib_uverbs_ex_destroy_flow(struct ib_uverbs_file *file,
3643                               struct ib_udata *ucore,
3644                               struct ib_udata *uhw)
3645 {
3646         struct ib_uverbs_destroy_flow   cmd;
3647         int                             ret;
3648
3649         if (ucore->inlen < sizeof(cmd))
3650                 return -EINVAL;
3651
3652         ret = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3653         if (ret)
3654                 return ret;
3655
3656         if (cmd.comp_mask)
3657                 return -EINVAL;
3658
3659         return uobj_perform_destroy(UVERBS_OBJECT_FLOW, cmd.flow_handle, file,
3660                                     0);
3661 }
3662
3663 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
3664                                 struct ib_uverbs_create_xsrq *cmd,
3665                                 struct ib_udata *udata)
3666 {
3667         struct ib_uverbs_create_srq_resp resp;
3668         struct ib_usrq_object           *obj;
3669         struct ib_pd                    *pd;
3670         struct ib_srq                   *srq;
3671         struct ib_uobject               *uninitialized_var(xrcd_uobj);
3672         struct ib_srq_init_attr          attr;
3673         int ret;
3674         struct ib_device *ib_dev;
3675
3676         obj = (struct ib_usrq_object *)uobj_alloc(UVERBS_OBJECT_SRQ, file,
3677                                                   &ib_dev);
3678         if (IS_ERR(obj))
3679                 return PTR_ERR(obj);
3680
3681         if (cmd->srq_type == IB_SRQT_TM)
3682                 attr.ext.tag_matching.max_num_tags = cmd->max_num_tags;
3683
3684         if (cmd->srq_type == IB_SRQT_XRC) {
3685                 xrcd_uobj = uobj_get_read(UVERBS_OBJECT_XRCD, cmd->xrcd_handle,
3686                                           file);
3687                 if (IS_ERR(xrcd_uobj)) {
3688                         ret = -EINVAL;
3689                         goto err;
3690                 }
3691
3692                 attr.ext.xrc.xrcd = (struct ib_xrcd *)xrcd_uobj->object;
3693                 if (!attr.ext.xrc.xrcd) {
3694                         ret = -EINVAL;
3695                         goto err_put_xrcd;
3696                 }
3697
3698                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
3699                 atomic_inc(&obj->uxrcd->refcnt);
3700         }
3701
3702         if (ib_srq_has_cq(cmd->srq_type)) {
3703                 attr.ext.cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ,
3704                                                 cmd->cq_handle, file);
3705                 if (!attr.ext.cq) {
3706                         ret = -EINVAL;
3707                         goto err_put_xrcd;
3708                 }
3709         }
3710
3711         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd->pd_handle, file);
3712         if (!pd) {
3713                 ret = -EINVAL;
3714                 goto err_put_cq;
3715         }
3716
3717         attr.event_handler  = ib_uverbs_srq_event_handler;
3718         attr.srq_context    = file;
3719         attr.srq_type       = cmd->srq_type;
3720         attr.attr.max_wr    = cmd->max_wr;
3721         attr.attr.max_sge   = cmd->max_sge;
3722         attr.attr.srq_limit = cmd->srq_limit;
3723
3724         obj->uevent.events_reported = 0;
3725         INIT_LIST_HEAD(&obj->uevent.event_list);
3726
3727         srq = pd->device->create_srq(pd, &attr, udata);
3728         if (IS_ERR(srq)) {
3729                 ret = PTR_ERR(srq);
3730                 goto err_put;
3731         }
3732
3733         srq->device        = pd->device;
3734         srq->pd            = pd;
3735         srq->srq_type      = cmd->srq_type;
3736         srq->uobject       = &obj->uevent.uobject;
3737         srq->event_handler = attr.event_handler;
3738         srq->srq_context   = attr.srq_context;
3739
3740         if (ib_srq_has_cq(cmd->srq_type)) {
3741                 srq->ext.cq       = attr.ext.cq;
3742                 atomic_inc(&attr.ext.cq->usecnt);
3743         }
3744
3745         if (cmd->srq_type == IB_SRQT_XRC) {
3746                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
3747                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
3748         }
3749
3750         atomic_inc(&pd->usecnt);
3751         atomic_set(&srq->usecnt, 0);
3752
3753         obj->uevent.uobject.object = srq;
3754         obj->uevent.uobject.user_handle = cmd->user_handle;
3755
3756         memset(&resp, 0, sizeof resp);
3757         resp.srq_handle = obj->uevent.uobject.id;
3758         resp.max_wr     = attr.attr.max_wr;
3759         resp.max_sge    = attr.attr.max_sge;
3760         if (cmd->srq_type == IB_SRQT_XRC)
3761                 resp.srqn = srq->ext.xrc.srq_num;
3762
3763         if (copy_to_user(u64_to_user_ptr(cmd->response),
3764                          &resp, sizeof resp)) {
3765                 ret = -EFAULT;
3766                 goto err_copy;
3767         }
3768
3769         if (cmd->srq_type == IB_SRQT_XRC)
3770                 uobj_put_read(xrcd_uobj);
3771
3772         if (ib_srq_has_cq(cmd->srq_type))
3773                 uobj_put_obj_read(attr.ext.cq);
3774
3775         uobj_put_obj_read(pd);
3776         return uobj_alloc_commit(&obj->uevent.uobject, 0);
3777
3778 err_copy:
3779         ib_destroy_srq(srq);
3780
3781 err_put:
3782         uobj_put_obj_read(pd);
3783
3784 err_put_cq:
3785         if (ib_srq_has_cq(cmd->srq_type))
3786                 uobj_put_obj_read(attr.ext.cq);
3787
3788 err_put_xrcd:
3789         if (cmd->srq_type == IB_SRQT_XRC) {
3790                 atomic_dec(&obj->uxrcd->refcnt);
3791                 uobj_put_read(xrcd_uobj);
3792         }
3793
3794 err:
3795         uobj_alloc_abort(&obj->uevent.uobject);
3796         return ret;
3797 }
3798
3799 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
3800                              const char __user *buf, int in_len,
3801                              int out_len)
3802 {
3803         struct ib_uverbs_create_srq      cmd;
3804         struct ib_uverbs_create_xsrq     xcmd;
3805         struct ib_uverbs_create_srq_resp resp;
3806         struct ib_udata                  udata;
3807         int ret;
3808
3809         if (out_len < sizeof resp)
3810                 return -ENOSPC;
3811
3812         if (copy_from_user(&cmd, buf, sizeof cmd))
3813                 return -EFAULT;
3814
3815         memset(&xcmd, 0, sizeof(xcmd));
3816         xcmd.response    = cmd.response;
3817         xcmd.user_handle = cmd.user_handle;
3818         xcmd.srq_type    = IB_SRQT_BASIC;
3819         xcmd.pd_handle   = cmd.pd_handle;
3820         xcmd.max_wr      = cmd.max_wr;
3821         xcmd.max_sge     = cmd.max_sge;
3822         xcmd.srq_limit   = cmd.srq_limit;
3823
3824         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
3825                    u64_to_user_ptr(cmd.response) + sizeof(resp),
3826                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
3827                    out_len - sizeof(resp));
3828
3829         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
3830         if (ret)
3831                 return ret;
3832
3833         return in_len;
3834 }
3835
3836 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
3837                               const char __user *buf, int in_len, int out_len)
3838 {
3839         struct ib_uverbs_create_xsrq     cmd;
3840         struct ib_uverbs_create_srq_resp resp;
3841         struct ib_udata                  udata;
3842         int ret;
3843
3844         if (out_len < sizeof resp)
3845                 return -ENOSPC;
3846
3847         if (copy_from_user(&cmd, buf, sizeof cmd))
3848                 return -EFAULT;
3849
3850         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
3851                    u64_to_user_ptr(cmd.response) + sizeof(resp),
3852                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
3853                    out_len - sizeof(resp));
3854
3855         ret = __uverbs_create_xsrq(file, &cmd, &udata);
3856         if (ret)
3857                 return ret;
3858
3859         return in_len;
3860 }
3861
3862 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
3863                              const char __user *buf, int in_len,
3864                              int out_len)
3865 {
3866         struct ib_uverbs_modify_srq cmd;
3867         struct ib_udata             udata;
3868         struct ib_srq              *srq;
3869         struct ib_srq_attr          attr;
3870         int                         ret;
3871
3872         if (copy_from_user(&cmd, buf, sizeof cmd))
3873                 return -EFAULT;
3874
3875         ib_uverbs_init_udata(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
3876                    out_len);
3877
3878         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
3879         if (!srq)
3880                 return -EINVAL;
3881
3882         attr.max_wr    = cmd.max_wr;
3883         attr.srq_limit = cmd.srq_limit;
3884
3885         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
3886
3887         uobj_put_obj_read(srq);
3888
3889         return ret ? ret : in_len;
3890 }
3891
3892 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
3893                             const char __user *buf,
3894                             int in_len, int out_len)
3895 {
3896         struct ib_uverbs_query_srq      cmd;
3897         struct ib_uverbs_query_srq_resp resp;
3898         struct ib_srq_attr              attr;
3899         struct ib_srq                   *srq;
3900         int                             ret;
3901
3902         if (out_len < sizeof resp)
3903                 return -ENOSPC;
3904
3905         if (copy_from_user(&cmd, buf, sizeof cmd))
3906                 return -EFAULT;
3907
3908         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
3909         if (!srq)
3910                 return -EINVAL;
3911
3912         ret = ib_query_srq(srq, &attr);
3913
3914         uobj_put_obj_read(srq);
3915
3916         if (ret)
3917                 return ret;
3918
3919         memset(&resp, 0, sizeof resp);
3920
3921         resp.max_wr    = attr.max_wr;
3922         resp.max_sge   = attr.max_sge;
3923         resp.srq_limit = attr.srq_limit;
3924
3925         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
3926                 return -EFAULT;
3927
3928         return in_len;
3929 }
3930
3931 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
3932                               const char __user *buf, int in_len,
3933                               int out_len)
3934 {
3935         struct ib_uverbs_destroy_srq      cmd;
3936         struct ib_uverbs_destroy_srq_resp resp;
3937         struct ib_uobject                *uobj;
3938         struct ib_uevent_object          *obj;
3939
3940         if (copy_from_user(&cmd, buf, sizeof cmd))
3941                 return -EFAULT;
3942
3943         uobj = uobj_get_destroy(UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
3944         if (IS_ERR(uobj))
3945                 return PTR_ERR(uobj);
3946
3947         obj = container_of(uobj, struct ib_uevent_object, uobject);
3948         memset(&resp, 0, sizeof(resp));
3949         resp.events_reported = obj->events_reported;
3950
3951         uobj_put_destroy(uobj);
3952
3953         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof(resp)))
3954                 return -EFAULT;
3955
3956         return in_len;
3957 }
3958
3959 int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
3960                               struct ib_udata *ucore,
3961                               struct ib_udata *uhw)
3962 {
3963         struct ib_uverbs_ex_query_device_resp resp = { {0} };
3964         struct ib_uverbs_ex_query_device  cmd;
3965         struct ib_device_attr attr = {0};
3966         struct ib_ucontext *ucontext;
3967         struct ib_device *ib_dev;
3968         int err;
3969
3970         ucontext = ib_uverbs_get_ucontext(file);
3971         if (IS_ERR(ucontext))
3972                 return PTR_ERR(ucontext);
3973         ib_dev = ucontext->device;
3974
3975         if (!ib_dev->query_device)
3976                 return -EOPNOTSUPP;
3977
3978         if (ucore->inlen < sizeof(cmd))
3979                 return -EINVAL;
3980
3981         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3982         if (err)
3983                 return err;
3984
3985         if (cmd.comp_mask)
3986                 return -EINVAL;
3987
3988         if (cmd.reserved)
3989                 return -EINVAL;
3990
3991         resp.response_length = offsetof(typeof(resp), odp_caps);
3992
3993         if (ucore->outlen < resp.response_length)
3994                 return -ENOSPC;
3995
3996         err = ib_dev->query_device(ib_dev, &attr, uhw);
3997         if (err)
3998                 return err;
3999
4000         copy_query_dev_fields(ucontext, &resp.base, &attr);
4001
4002         if (ucore->outlen < resp.response_length + sizeof(resp.odp_caps))
4003                 goto end;
4004
4005 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4006         resp.odp_caps.general_caps = attr.odp_caps.general_caps;
4007         resp.odp_caps.per_transport_caps.rc_odp_caps =
4008                 attr.odp_caps.per_transport_caps.rc_odp_caps;
4009         resp.odp_caps.per_transport_caps.uc_odp_caps =
4010                 attr.odp_caps.per_transport_caps.uc_odp_caps;
4011         resp.odp_caps.per_transport_caps.ud_odp_caps =
4012                 attr.odp_caps.per_transport_caps.ud_odp_caps;
4013 #endif
4014         resp.response_length += sizeof(resp.odp_caps);
4015
4016         if (ucore->outlen < resp.response_length + sizeof(resp.timestamp_mask))
4017                 goto end;
4018
4019         resp.timestamp_mask = attr.timestamp_mask;
4020         resp.response_length += sizeof(resp.timestamp_mask);
4021
4022         if (ucore->outlen < resp.response_length + sizeof(resp.hca_core_clock))
4023                 goto end;
4024
4025         resp.hca_core_clock = attr.hca_core_clock;
4026         resp.response_length += sizeof(resp.hca_core_clock);
4027
4028         if (ucore->outlen < resp.response_length + sizeof(resp.device_cap_flags_ex))
4029                 goto end;
4030
4031         resp.device_cap_flags_ex = attr.device_cap_flags;
4032         resp.response_length += sizeof(resp.device_cap_flags_ex);
4033
4034         if (ucore->outlen < resp.response_length + sizeof(resp.rss_caps))
4035                 goto end;
4036
4037         resp.rss_caps.supported_qpts = attr.rss_caps.supported_qpts;
4038         resp.rss_caps.max_rwq_indirection_tables =
4039                 attr.rss_caps.max_rwq_indirection_tables;
4040         resp.rss_caps.max_rwq_indirection_table_size =
4041                 attr.rss_caps.max_rwq_indirection_table_size;
4042
4043         resp.response_length += sizeof(resp.rss_caps);
4044
4045         if (ucore->outlen < resp.response_length + sizeof(resp.max_wq_type_rq))
4046                 goto end;
4047
4048         resp.max_wq_type_rq = attr.max_wq_type_rq;
4049         resp.response_length += sizeof(resp.max_wq_type_rq);
4050
4051         if (ucore->outlen < resp.response_length + sizeof(resp.raw_packet_caps))
4052                 goto end;
4053
4054         resp.raw_packet_caps = attr.raw_packet_caps;
4055         resp.response_length += sizeof(resp.raw_packet_caps);
4056
4057         if (ucore->outlen < resp.response_length + sizeof(resp.tm_caps))
4058                 goto end;
4059
4060         resp.tm_caps.max_rndv_hdr_size  = attr.tm_caps.max_rndv_hdr_size;
4061         resp.tm_caps.max_num_tags       = attr.tm_caps.max_num_tags;
4062         resp.tm_caps.max_ops            = attr.tm_caps.max_ops;
4063         resp.tm_caps.max_sge            = attr.tm_caps.max_sge;
4064         resp.tm_caps.flags              = attr.tm_caps.flags;
4065         resp.response_length += sizeof(resp.tm_caps);
4066
4067         if (ucore->outlen < resp.response_length + sizeof(resp.cq_moderation_caps))
4068                 goto end;
4069
4070         resp.cq_moderation_caps.max_cq_moderation_count  =
4071                 attr.cq_caps.max_cq_moderation_count;
4072         resp.cq_moderation_caps.max_cq_moderation_period =
4073                 attr.cq_caps.max_cq_moderation_period;
4074         resp.response_length += sizeof(resp.cq_moderation_caps);
4075
4076         if (ucore->outlen < resp.response_length + sizeof(resp.max_dm_size))
4077                 goto end;
4078
4079         resp.max_dm_size = attr.max_dm_size;
4080         resp.response_length += sizeof(resp.max_dm_size);
4081 end:
4082         err = ib_copy_to_udata(ucore, &resp, resp.response_length);
4083         return err;
4084 }
4085
4086 int ib_uverbs_ex_modify_cq(struct ib_uverbs_file *file,
4087                            struct ib_udata *ucore,
4088                            struct ib_udata *uhw)
4089 {
4090         struct ib_uverbs_ex_modify_cq cmd = {};
4091         struct ib_cq *cq;
4092         size_t required_cmd_sz;
4093         int ret;
4094
4095         required_cmd_sz = offsetof(typeof(cmd), reserved) +
4096                                 sizeof(cmd.reserved);
4097         if (ucore->inlen < required_cmd_sz)
4098                 return -EINVAL;
4099
4100         /* sanity checks */
4101         if (ucore->inlen > sizeof(cmd) &&
4102             !ib_is_udata_cleared(ucore, sizeof(cmd),
4103                                  ucore->inlen - sizeof(cmd)))
4104                 return -EOPNOTSUPP;
4105
4106         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
4107         if (ret)
4108                 return ret;
4109
4110         if (!cmd.attr_mask || cmd.reserved)
4111                 return -EINVAL;
4112
4113         if (cmd.attr_mask > IB_CQ_MODERATE)
4114                 return -EOPNOTSUPP;
4115
4116         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
4117         if (!cq)
4118                 return -EINVAL;
4119
4120         ret = rdma_set_cq_moderation(cq, cmd.attr.cq_count, cmd.attr.cq_period);
4121
4122         uobj_put_obj_read(cq);
4123
4124         return ret;
4125 }