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         ucontext->closing = false;
 121         ucontext->cleanup_retryable = false;
 122
 123 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 124         mutex_init(&ucontext->per_mm_list_lock);
 125         INIT_LIST_HEAD(&ucontext->per_mm_list);
 126         if (!(ib_dev->attrs.device_cap_flags & IB_DEVICE_ON_DEMAND_PAGING))
 127                 ucontext->invalidate_range = NULL;
 128
 129 #endif
 130
 131         resp.num_comp_vectors = file->device->num_comp_vectors;
 132
 133         ret = get_unused_fd_flags(O_CLOEXEC);
 134         if (ret < 0)
 135                 goto err_free;
 136         resp.async_fd = ret;
 137
 138         filp = ib_uverbs_alloc_async_event_file(file, ib_dev);
 139         if (IS_ERR(filp)) {
 140                 ret = PTR_ERR(filp);
 141                 goto err_fd;
 142         }
 143
 144         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 145                 ret = -EFAULT;
 146                 goto err_file;
 147         }
 148
 149         fd_install(resp.async_fd, filp);
 150
 151         /*
 152          * Make sure that ib_uverbs_get_ucontext() sees the pointer update
 153          * only after all writes to setup the ucontext have completed
 154          */
 155         smp_store_release(&file->ucontext, ucontext);
 156
 157         mutex_unlock(&file->ucontext_lock);
 158
 159         return in_len;
 160
 161 err_file:
 162         ib_uverbs_free_async_event_file(file);
 163         fput(filp);
 164
 165 err_fd:
 166         put_unused_fd(resp.async_fd);
 167
 168 err_free:
 169         ib_dev->dealloc_ucontext(ucontext);
 170
 171 err_alloc:
 172         ib_rdmacg_uncharge(&cg_obj, ib_dev, RDMACG_RESOURCE_HCA_HANDLE);
 173
 174 err:
 175         mutex_unlock(&file->ucontext_lock);
 176         return ret;
 177 }
 178
 179 static void copy_query_dev_fields(struct ib_ucontext *ucontext,
 180                                   struct ib_uverbs_query_device_resp *resp,
 181                                   struct ib_device_attr *attr)
 182 {
 183         struct ib_device *ib_dev = ucontext->device;
 184
 185         resp->fw_ver            = attr->fw_ver;
 186         resp->node_guid         = ib_dev->node_guid;
 187         resp->sys_image_guid    = attr->sys_image_guid;
 188         resp->max_mr_size       = attr->max_mr_size;
 189         resp->page_size_cap     = attr->page_size_cap;
 190         resp->vendor_id         = attr->vendor_id;
 191         resp->vendor_part_id    = attr->vendor_part_id;
 192         resp->hw_ver            = attr->hw_ver;
 193         resp->max_qp            = attr->max_qp;
 194         resp->max_qp_wr         = attr->max_qp_wr;
 195         resp->device_cap_flags  = lower_32_bits(attr->device_cap_flags);
 196         resp->max_sge           = min(attr->max_send_sge, attr->max_recv_sge);
 197         resp->max_sge_rd        = attr->max_sge_rd;
 198         resp->max_cq            = attr->max_cq;
 199         resp->max_cqe           = attr->max_cqe;
 200         resp->max_mr            = attr->max_mr;
 201         resp->max_pd            = attr->max_pd;
 202         resp->max_qp_rd_atom    = attr->max_qp_rd_atom;
 203         resp->max_ee_rd_atom    = attr->max_ee_rd_atom;
 204         resp->max_res_rd_atom   = attr->max_res_rd_atom;
 205         resp->max_qp_init_rd_atom       = attr->max_qp_init_rd_atom;
 206         resp->max_ee_init_rd_atom       = attr->max_ee_init_rd_atom;
 207         resp->atomic_cap                = attr->atomic_cap;
 208         resp->max_ee                    = attr->max_ee;
 209         resp->max_rdd                   = attr->max_rdd;
 210         resp->max_mw                    = attr->max_mw;
 211         resp->max_raw_ipv6_qp           = attr->max_raw_ipv6_qp;
 212         resp->max_raw_ethy_qp           = attr->max_raw_ethy_qp;
 213         resp->max_mcast_grp             = attr->max_mcast_grp;
 214         resp->max_mcast_qp_attach       = attr->max_mcast_qp_attach;
 215         resp->max_total_mcast_qp_attach = attr->max_total_mcast_qp_attach;
 216         resp->max_ah                    = attr->max_ah;
 217         resp->max_fmr                   = attr->max_fmr;
 218         resp->max_map_per_fmr           = attr->max_map_per_fmr;
 219         resp->max_srq                   = attr->max_srq;
 220         resp->max_srq_wr                = attr->max_srq_wr;
 221         resp->max_srq_sge               = attr->max_srq_sge;
 222         resp->max_pkeys                 = attr->max_pkeys;
 223         resp->local_ca_ack_delay        = attr->local_ca_ack_delay;
 224         resp->phys_port_cnt             = ib_dev->phys_port_cnt;
 225 }
 226
 227 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 228                                const char __user *buf,
 229                                int in_len, int out_len)
 230 {
 231         struct ib_uverbs_query_device      cmd;
 232         struct ib_uverbs_query_device_resp resp;
 233         struct ib_ucontext *ucontext;
 234
 235         ucontext = ib_uverbs_get_ucontext(file);
 236         if (IS_ERR(ucontext))
 237                 return PTR_ERR(ucontext);
 238
 239         if (out_len < sizeof resp)
 240                 return -ENOSPC;
 241
 242         if (copy_from_user(&cmd, buf, sizeof cmd))
 243                 return -EFAULT;
 244
 245         memset(&resp, 0, sizeof resp);
 246         copy_query_dev_fields(ucontext, &resp, &ucontext->device->attrs);
 247
 248         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
 249                 return -EFAULT;
 250
 251         return in_len;
 252 }
 253
 254 /*
 255  * ib_uverbs_query_port_resp.port_cap_flags started out as just a copy of the
 256  * PortInfo CapabilityMask, but was extended with unique bits.
 257  */
 258 static u32 make_port_cap_flags(const struct ib_port_attr *attr)
 259 {
 260         u32 res;
 261
 262         /* All IBA CapabilityMask bits are passed through here, except bit 26,
 263          * which is overridden with IP_BASED_GIDS. This is due to a historical
 264          * mistake in the implementation of IP_BASED_GIDS. Otherwise all other
 265          * bits match the IBA definition across all kernel versions.
 266          */
 267         res = attr->port_cap_flags & ~(u32)IB_UVERBS_PCF_IP_BASED_GIDS;
 268
 269         if (attr->ip_gids)
 270                 res |= IB_UVERBS_PCF_IP_BASED_GIDS;
 271
 272         return res;
 273 }
 274
 275 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 276                              const char __user *buf,
 277                              int in_len, int out_len)
 278 {
 279         struct ib_uverbs_query_port      cmd;
 280         struct ib_uverbs_query_port_resp resp;
 281         struct ib_port_attr              attr;
 282         int                              ret;
 283         struct ib_ucontext *ucontext;
 284         struct ib_device *ib_dev;
 285
 286         ucontext = ib_uverbs_get_ucontext(file);
 287         if (IS_ERR(ucontext))
 288                 return PTR_ERR(ucontext);
 289         ib_dev = ucontext->device;
 290
 291         if (out_len < sizeof resp)
 292                 return -ENOSPC;
 293
 294         if (copy_from_user(&cmd, buf, sizeof cmd))
 295                 return -EFAULT;
 296
 297         ret = ib_query_port(ib_dev, cmd.port_num, &attr);
 298         if (ret)
 299                 return ret;
 300
 301         memset(&resp, 0, sizeof resp);
 302
 303         resp.state           = attr.state;
 304         resp.max_mtu         = attr.max_mtu;
 305         resp.active_mtu      = attr.active_mtu;
 306         resp.gid_tbl_len     = attr.gid_tbl_len;
 307         resp.port_cap_flags  = make_port_cap_flags(&attr);
 308         resp.max_msg_sz      = attr.max_msg_sz;
 309         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 310         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 311         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 312
 313         if (rdma_is_grh_required(ib_dev, cmd.port_num))
 314                 resp.flags |= IB_UVERBS_QPF_GRH_REQUIRED;
 315
 316         if (rdma_cap_opa_ah(ib_dev, cmd.port_num)) {
 317                 resp.lid     = OPA_TO_IB_UCAST_LID(attr.lid);
 318                 resp.sm_lid  = OPA_TO_IB_UCAST_LID(attr.sm_lid);
 319         } else {
 320                 resp.lid     = ib_lid_cpu16(attr.lid);
 321                 resp.sm_lid  = ib_lid_cpu16(attr.sm_lid);
 322         }
 323         resp.lmc             = attr.lmc;
 324         resp.max_vl_num      = attr.max_vl_num;
 325         resp.sm_sl           = attr.sm_sl;
 326         resp.subnet_timeout  = attr.subnet_timeout;
 327         resp.init_type_reply = attr.init_type_reply;
 328         resp.active_width    = attr.active_width;
 329         resp.active_speed    = attr.active_speed;
 330         resp.phys_state      = attr.phys_state;
 331         resp.link_layer      = rdma_port_get_link_layer(ib_dev,
 332                                                         cmd.port_num);
 333
 334         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
 335                 return -EFAULT;
 336
 337         return in_len;
 338 }
 339
 340 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 341                            const char __user *buf,
 342                            int in_len, int out_len)
 343 {
 344         struct ib_uverbs_alloc_pd      cmd;
 345         struct ib_uverbs_alloc_pd_resp resp;
 346         struct ib_udata                udata;
 347         struct ib_uobject             *uobj;
 348         struct ib_pd                  *pd;
 349         int                            ret;
 350         struct ib_device *ib_dev;
 351
 352         if (out_len < sizeof resp)
 353                 return -ENOSPC;
 354
 355         if (copy_from_user(&cmd, buf, sizeof cmd))
 356                 return -EFAULT;
 357
 358         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 359                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 360                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 361                    out_len - sizeof(resp));
 362
 363         uobj = uobj_alloc(UVERBS_OBJECT_PD, file, &ib_dev);
 364         if (IS_ERR(uobj))
 365                 return PTR_ERR(uobj);
 366
 367         pd = ib_dev->alloc_pd(ib_dev, uobj->context, &udata);
 368         if (IS_ERR(pd)) {
 369                 ret = PTR_ERR(pd);
 370                 goto err;
 371         }
 372
 373         pd->device  = ib_dev;
 374         pd->uobject = uobj;
 375         pd->__internal_mr = NULL;
 376         atomic_set(&pd->usecnt, 0);
 377
 378         uobj->object = pd;
 379         memset(&resp, 0, sizeof resp);
 380         resp.pd_handle = uobj->id;
 381         pd->res.type = RDMA_RESTRACK_PD;
 382         rdma_restrack_add(&pd->res);
 383
 384         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 385                 ret = -EFAULT;
 386                 goto err_copy;
 387         }
 388
 389         return uobj_alloc_commit(uobj, in_len);
 390
 391 err_copy:
 392         ib_dealloc_pd(pd);
 393
 394 err:
 395         uobj_alloc_abort(uobj);
 396         return ret;
 397 }
 398
 399 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 400                              const char __user *buf,
 401                              int in_len, int out_len)
 402 {
 403         struct ib_uverbs_dealloc_pd cmd;
 404
 405         if (copy_from_user(&cmd, buf, sizeof cmd))
 406                 return -EFAULT;
 407
 408         return uobj_perform_destroy(UVERBS_OBJECT_PD, cmd.pd_handle, file,
 409                                     in_len);
 410 }
 411
 412 struct xrcd_table_entry {
 413         struct rb_node  node;
 414         struct ib_xrcd *xrcd;
 415         struct inode   *inode;
 416 };
 417
 418 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 419                             struct inode *inode,
 420                             struct ib_xrcd *xrcd)
 421 {
 422         struct xrcd_table_entry *entry, *scan;
 423         struct rb_node **p = &dev->xrcd_tree.rb_node;
 424         struct rb_node *parent = NULL;
 425
 426         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 427         if (!entry)
 428                 return -ENOMEM;
 429
 430         entry->xrcd  = xrcd;
 431         entry->inode = inode;
 432
 433         while (*p) {
 434                 parent = *p;
 435                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 436
 437                 if (inode < scan->inode) {
 438                         p = &(*p)->rb_left;
 439                 } else if (inode > scan->inode) {
 440                         p = &(*p)->rb_right;
 441                 } else {
 442                         kfree(entry);
 443                         return -EEXIST;
 444                 }
 445         }
 446
 447         rb_link_node(&entry->node, parent, p);
 448         rb_insert_color(&entry->node, &dev->xrcd_tree);
 449         igrab(inode);
 450         return 0;
 451 }
 452
 453 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 454                                                   struct inode *inode)
 455 {
 456         struct xrcd_table_entry *entry;
 457         struct rb_node *p = dev->xrcd_tree.rb_node;
 458
 459         while (p) {
 460                 entry = rb_entry(p, struct xrcd_table_entry, node);
 461
 462                 if (inode < entry->inode)
 463                         p = p->rb_left;
 464                 else if (inode > entry->inode)
 465                         p = p->rb_right;
 466                 else
 467                         return entry;
 468         }
 469
 470         return NULL;
 471 }
 472
 473 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 474 {
 475         struct xrcd_table_entry *entry;
 476
 477         entry = xrcd_table_search(dev, inode);
 478         if (!entry)
 479                 return NULL;
 480
 481         return entry->xrcd;
 482 }
 483
 484 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 485                               struct inode *inode)
 486 {
 487         struct xrcd_table_entry *entry;
 488
 489         entry = xrcd_table_search(dev, inode);
 490         if (entry) {
 491                 iput(inode);
 492                 rb_erase(&entry->node, &dev->xrcd_tree);
 493                 kfree(entry);
 494         }
 495 }
 496
 497 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 498                             const char __user *buf, int in_len,
 499                             int out_len)
 500 {
 501         struct ib_uverbs_open_xrcd      cmd;
 502         struct ib_uverbs_open_xrcd_resp resp;
 503         struct ib_udata                 udata;
 504         struct ib_uxrcd_object         *obj;
 505         struct ib_xrcd                 *xrcd = NULL;
 506         struct fd                       f = {NULL, 0};
 507         struct inode                   *inode = NULL;
 508         int                             ret = 0;
 509         int                             new_xrcd = 0;
 510         struct ib_device *ib_dev;
 511
 512         if (out_len < sizeof resp)
 513                 return -ENOSPC;
 514
 515         if (copy_from_user(&cmd, buf, sizeof cmd))
 516                 return -EFAULT;
 517
 518         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 519                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 520                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 521                    out_len - sizeof(resp));
 522
 523         mutex_lock(&file->device->xrcd_tree_mutex);
 524
 525         if (cmd.fd != -1) {
 526                 /* search for file descriptor */
 527                 f = fdget(cmd.fd);
 528                 if (!f.file) {
 529                         ret = -EBADF;
 530                         goto err_tree_mutex_unlock;
 531                 }
 532
 533                 inode = file_inode(f.file);
 534                 xrcd = find_xrcd(file->device, inode);
 535                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 536                         /* no file descriptor. Need CREATE flag */
 537                         ret = -EAGAIN;
 538                         goto err_tree_mutex_unlock;
 539                 }
 540
 541                 if (xrcd && cmd.oflags & O_EXCL) {
 542                         ret = -EINVAL;
 543                         goto err_tree_mutex_unlock;
 544                 }
 545         }
 546
 547         obj = (struct ib_uxrcd_object *)uobj_alloc(UVERBS_OBJECT_XRCD, file,
 548                                                    &ib_dev);
 549         if (IS_ERR(obj)) {
 550                 ret = PTR_ERR(obj);
 551                 goto err_tree_mutex_unlock;
 552         }
 553
 554         if (!xrcd) {
 555                 xrcd = ib_dev->alloc_xrcd(ib_dev, obj->uobject.context, &udata);
 556                 if (IS_ERR(xrcd)) {
 557                         ret = PTR_ERR(xrcd);
 558                         goto err;
 559                 }
 560
 561                 xrcd->inode   = inode;
 562                 xrcd->device  = ib_dev;
 563                 atomic_set(&xrcd->usecnt, 0);
 564                 mutex_init(&xrcd->tgt_qp_mutex);
 565                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 566                 new_xrcd = 1;
 567         }
 568
 569         atomic_set(&obj->refcnt, 0);
 570         obj->uobject.object = xrcd;
 571         memset(&resp, 0, sizeof resp);
 572         resp.xrcd_handle = obj->uobject.id;
 573
 574         if (inode) {
 575                 if (new_xrcd) {
 576                         /* create new inode/xrcd table entry */
 577                         ret = xrcd_table_insert(file->device, inode, xrcd);
 578                         if (ret)
 579                                 goto err_dealloc_xrcd;
 580                 }
 581                 atomic_inc(&xrcd->usecnt);
 582         }
 583
 584         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 585                 ret = -EFAULT;
 586                 goto err_copy;
 587         }
 588
 589         if (f.file)
 590                 fdput(f);
 591
 592         mutex_unlock(&file->device->xrcd_tree_mutex);
 593
 594         return uobj_alloc_commit(&obj->uobject, in_len);
 595
 596 err_copy:
 597         if (inode) {
 598                 if (new_xrcd)
 599                         xrcd_table_delete(file->device, inode);
 600                 atomic_dec(&xrcd->usecnt);
 601         }
 602
 603 err_dealloc_xrcd:
 604         ib_dealloc_xrcd(xrcd);
 605
 606 err:
 607         uobj_alloc_abort(&obj->uobject);
 608
 609 err_tree_mutex_unlock:
 610         if (f.file)
 611                 fdput(f);
 612
 613         mutex_unlock(&file->device->xrcd_tree_mutex);
 614
 615         return ret;
 616 }
 617
 618 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 619                              const char __user *buf, int in_len,
 620                              int out_len)
 621 {
 622         struct ib_uverbs_close_xrcd cmd;
 623
 624         if (copy_from_user(&cmd, buf, sizeof cmd))
 625                 return -EFAULT;
 626
 627         return uobj_perform_destroy(UVERBS_OBJECT_XRCD, cmd.xrcd_handle, file,
 628                                     in_len);
 629 }
 630
 631 int ib_uverbs_dealloc_xrcd(struct ib_uobject *uobject,
 632                            struct ib_xrcd *xrcd,
 633                            enum rdma_remove_reason why)
 634 {
 635         struct inode *inode;
 636         int ret;
 637         struct ib_uverbs_device *dev = uobject->context->ufile->device;
 638
 639         inode = xrcd->inode;
 640         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 641                 return 0;
 642
 643         ret = ib_dealloc_xrcd(xrcd);
 644
 645         if (ib_is_destroy_retryable(ret, why, uobject)) {
 646                 atomic_inc(&xrcd->usecnt);
 647                 return ret;
 648         }
 649
 650         if (inode)
 651                 xrcd_table_delete(dev, inode);
 652
 653         return ret;
 654 }
 655
 656 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 657                          const char __user *buf, int in_len,
 658                          int out_len)
 659 {
 660         struct ib_uverbs_reg_mr      cmd;
 661         struct ib_uverbs_reg_mr_resp resp;
 662         struct ib_udata              udata;
 663         struct ib_uobject           *uobj;
 664         struct ib_pd                *pd;
 665         struct ib_mr                *mr;
 666         int                          ret;
 667         struct ib_device *ib_dev;
 668
 669         if (out_len < sizeof resp)
 670                 return -ENOSPC;
 671
 672         if (copy_from_user(&cmd, buf, sizeof cmd))
 673                 return -EFAULT;
 674
 675         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 676                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 677                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 678                    out_len - sizeof(resp));
 679
 680         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 681                 return -EINVAL;
 682
 683         ret = ib_check_mr_access(cmd.access_flags);
 684         if (ret)
 685                 return ret;
 686
 687         uobj = uobj_alloc(UVERBS_OBJECT_MR, file, &ib_dev);
 688         if (IS_ERR(uobj))
 689                 return PTR_ERR(uobj);
 690
 691         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file);
 692         if (!pd) {
 693                 ret = -EINVAL;
 694                 goto err_free;
 695         }
 696
 697         if (cmd.access_flags & IB_ACCESS_ON_DEMAND) {
 698                 if (!(pd->device->attrs.device_cap_flags &
 699                       IB_DEVICE_ON_DEMAND_PAGING)) {
 700                         pr_debug("ODP support not available\n");
 701                         ret = -EINVAL;
 702                         goto err_put;
 703                 }
 704         }
 705
 706         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 707                                      cmd.access_flags, &udata);
 708         if (IS_ERR(mr)) {
 709                 ret = PTR_ERR(mr);
 710                 goto err_put;
 711         }
 712
 713         mr->device  = pd->device;
 714         mr->pd      = pd;
 715         mr->dm      = NULL;
 716         mr->uobject = uobj;
 717         atomic_inc(&pd->usecnt);
 718         mr->res.type = RDMA_RESTRACK_MR;
 719         rdma_restrack_add(&mr->res);
 720
 721         uobj->object = mr;
 722
 723         memset(&resp, 0, sizeof resp);
 724         resp.lkey      = mr->lkey;
 725         resp.rkey      = mr->rkey;
 726         resp.mr_handle = uobj->id;
 727
 728         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 729                 ret = -EFAULT;
 730                 goto err_copy;
 731         }
 732
 733         uobj_put_obj_read(pd);
 734
 735         return uobj_alloc_commit(uobj, in_len);
 736
 737 err_copy:
 738         ib_dereg_mr(mr);
 739
 740 err_put:
 741         uobj_put_obj_read(pd);
 742
 743 err_free:
 744         uobj_alloc_abort(uobj);
 745         return ret;
 746 }
 747
 748 ssize_t ib_uverbs_rereg_mr(struct ib_uverbs_file *file,
 749                            const char __user *buf, int in_len,
 750                            int out_len)
 751 {
 752         struct ib_uverbs_rereg_mr      cmd;
 753         struct ib_uverbs_rereg_mr_resp resp;
 754         struct ib_udata              udata;
 755         struct ib_pd                *pd = NULL;
 756         struct ib_mr                *mr;
 757         struct ib_pd                *old_pd;
 758         int                          ret;
 759         struct ib_uobject           *uobj;
 760
 761         if (out_len < sizeof(resp))
 762                 return -ENOSPC;
 763
 764         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 765                 return -EFAULT;
 766
 767         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 768                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 769                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 770                    out_len - sizeof(resp));
 771
 772         if (cmd.flags & ~IB_MR_REREG_SUPPORTED || !cmd.flags)
 773                 return -EINVAL;
 774
 775         if ((cmd.flags & IB_MR_REREG_TRANS) &&
 776             (!cmd.start || !cmd.hca_va || 0 >= cmd.length ||
 777              (cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK)))
 778                         return -EINVAL;
 779
 780         uobj = uobj_get_write(UVERBS_OBJECT_MR, cmd.mr_handle, file);
 781         if (IS_ERR(uobj))
 782                 return PTR_ERR(uobj);
 783
 784         mr = uobj->object;
 785
 786         if (mr->dm) {
 787                 ret = -EINVAL;
 788                 goto put_uobjs;
 789         }
 790
 791         if (cmd.flags & IB_MR_REREG_ACCESS) {
 792                 ret = ib_check_mr_access(cmd.access_flags);
 793                 if (ret)
 794                         goto put_uobjs;
 795         }
 796
 797         if (cmd.flags & IB_MR_REREG_PD) {
 798                 pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle,
 799                                        file);
 800                 if (!pd) {
 801                         ret = -EINVAL;
 802                         goto put_uobjs;
 803                 }
 804         }
 805
 806         old_pd = mr->pd;
 807         ret = mr->device->rereg_user_mr(mr, cmd.flags, cmd.start,
 808                                         cmd.length, cmd.hca_va,
 809                                         cmd.access_flags, pd, &udata);
 810         if (!ret) {
 811                 if (cmd.flags & IB_MR_REREG_PD) {
 812                         atomic_inc(&pd->usecnt);
 813                         mr->pd = pd;
 814                         atomic_dec(&old_pd->usecnt);
 815                 }
 816         } else {
 817                 goto put_uobj_pd;
 818         }
 819
 820         memset(&resp, 0, sizeof(resp));
 821         resp.lkey      = mr->lkey;
 822         resp.rkey      = mr->rkey;
 823
 824         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof(resp)))
 825                 ret = -EFAULT;
 826         else
 827                 ret = in_len;
 828
 829 put_uobj_pd:
 830         if (cmd.flags & IB_MR_REREG_PD)
 831                 uobj_put_obj_read(pd);
 832
 833 put_uobjs:
 834         uobj_put_write(uobj);
 835
 836         return ret;
 837 }
 838
 839 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
 840                            const char __user *buf, int in_len,
 841                            int out_len)
 842 {
 843         struct ib_uverbs_dereg_mr cmd;
 844
 845         if (copy_from_user(&cmd, buf, sizeof cmd))
 846                 return -EFAULT;
 847
 848         return uobj_perform_destroy(UVERBS_OBJECT_MR, cmd.mr_handle, file,
 849                                     in_len);
 850 }
 851
 852 ssize_t ib_uverbs_alloc_mw(struct ib_uverbs_file *file,
 853                            const char __user *buf, int in_len,
 854                            int out_len)
 855 {
 856         struct ib_uverbs_alloc_mw      cmd;
 857         struct ib_uverbs_alloc_mw_resp resp;
 858         struct ib_uobject             *uobj;
 859         struct ib_pd                  *pd;
 860         struct ib_mw                  *mw;
 861         struct ib_udata                udata;
 862         int                            ret;
 863         struct ib_device *ib_dev;
 864
 865         if (out_len < sizeof(resp))
 866                 return -ENOSPC;
 867
 868         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 869                 return -EFAULT;
 870
 871         uobj = uobj_alloc(UVERBS_OBJECT_MW, file, &ib_dev);
 872         if (IS_ERR(uobj))
 873                 return PTR_ERR(uobj);
 874
 875         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file);
 876         if (!pd) {
 877                 ret = -EINVAL;
 878                 goto err_free;
 879         }
 880
 881         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
 882                    u64_to_user_ptr(cmd.response) + sizeof(resp),
 883                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
 884                    out_len - sizeof(resp));
 885
 886         mw = pd->device->alloc_mw(pd, cmd.mw_type, &udata);
 887         if (IS_ERR(mw)) {
 888                 ret = PTR_ERR(mw);
 889                 goto err_put;
 890         }
 891
 892         mw->device  = pd->device;
 893         mw->pd      = pd;
 894         mw->uobject = uobj;
 895         atomic_inc(&pd->usecnt);
 896
 897         uobj->object = mw;
 898
 899         memset(&resp, 0, sizeof(resp));
 900         resp.rkey      = mw->rkey;
 901         resp.mw_handle = uobj->id;
 902
 903         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof(resp))) {
 904                 ret = -EFAULT;
 905                 goto err_copy;
 906         }
 907
 908         uobj_put_obj_read(pd);
 909         return uobj_alloc_commit(uobj, in_len);
 910
 911 err_copy:
 912         uverbs_dealloc_mw(mw);
 913 err_put:
 914         uobj_put_obj_read(pd);
 915 err_free:
 916         uobj_alloc_abort(uobj);
 917         return ret;
 918 }
 919
 920 ssize_t ib_uverbs_dealloc_mw(struct ib_uverbs_file *file,
 921                              const char __user *buf, int in_len,
 922                              int out_len)
 923 {
 924         struct ib_uverbs_dealloc_mw cmd;
 925
 926         if (copy_from_user(&cmd, buf, sizeof(cmd)))
 927                 return -EFAULT;
 928
 929         return uobj_perform_destroy(UVERBS_OBJECT_MW, cmd.mw_handle, file,
 930                                     in_len);
 931 }
 932
 933 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
 934                                       const char __user *buf, int in_len,
 935                                       int out_len)
 936 {
 937         struct ib_uverbs_create_comp_channel       cmd;
 938         struct ib_uverbs_create_comp_channel_resp  resp;
 939         struct ib_uobject                         *uobj;
 940         struct ib_uverbs_completion_event_file    *ev_file;
 941         struct ib_device *ib_dev;
 942
 943         if (out_len < sizeof resp)
 944                 return -ENOSPC;
 945
 946         if (copy_from_user(&cmd, buf, sizeof cmd))
 947                 return -EFAULT;
 948
 949         uobj = uobj_alloc(UVERBS_OBJECT_COMP_CHANNEL, file, &ib_dev);
 950         if (IS_ERR(uobj))
 951                 return PTR_ERR(uobj);
 952
 953         resp.fd = uobj->id;
 954
 955         ev_file = container_of(uobj, struct ib_uverbs_completion_event_file,
 956                                uobj);
 957         ib_uverbs_init_event_queue(&ev_file->ev_queue);
 958
 959         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
 960                 uobj_alloc_abort(uobj);
 961                 return -EFAULT;
 962         }
 963
 964         return uobj_alloc_commit(uobj, in_len);
 965 }
 966
 967 static struct ib_ucq_object *create_cq(struct ib_uverbs_file *file,
 968                                        struct ib_udata *ucore,
 969                                        struct ib_udata *uhw,
 970                                        struct ib_uverbs_ex_create_cq *cmd,
 971                                        size_t cmd_sz,
 972                                        int (*cb)(struct ib_uverbs_file *file,
 973                                                  struct ib_ucq_object *obj,
 974                                                  struct ib_uverbs_ex_create_cq_resp *resp,
 975                                                  struct ib_udata *udata,
 976                                                  void *context),
 977                                        void *context)
 978 {
 979         struct ib_ucq_object           *obj;
 980         struct ib_uverbs_completion_event_file    *ev_file = NULL;
 981         struct ib_cq                   *cq;
 982         int                             ret;
 983         struct ib_uverbs_ex_create_cq_resp resp;
 984         struct ib_cq_init_attr attr = {};
 985         struct ib_device *ib_dev;
 986
 987         if (cmd->comp_vector >= file->device->num_comp_vectors)
 988                 return ERR_PTR(-EINVAL);
 989
 990         obj = (struct ib_ucq_object *)uobj_alloc(UVERBS_OBJECT_CQ, file,
 991                                                  &ib_dev);
 992         if (IS_ERR(obj))
 993                 return obj;
 994
 995         if (!ib_dev->create_cq) {
 996                 ret = -EOPNOTSUPP;
 997                 goto err;
 998         }
 999
1000         if (cmd->comp_channel >= 0) {
1001                 ev_file = ib_uverbs_lookup_comp_file(cmd->comp_channel, file);
1002                 if (IS_ERR(ev_file)) {
1003                         ret = PTR_ERR(ev_file);
1004                         goto err;
1005                 }
1006         }
1007
1008         obj->uobject.user_handle = cmd->user_handle;
1009         obj->comp_events_reported  = 0;
1010         obj->async_events_reported = 0;
1011         INIT_LIST_HEAD(&obj->comp_list);
1012         INIT_LIST_HEAD(&obj->async_list);
1013
1014         attr.cqe = cmd->cqe;
1015         attr.comp_vector = cmd->comp_vector;
1016
1017         if (cmd_sz > offsetof(typeof(*cmd), flags) + sizeof(cmd->flags))
1018                 attr.flags = cmd->flags;
1019
1020         cq = ib_dev->create_cq(ib_dev, &attr, obj->uobject.context, uhw);
1021         if (IS_ERR(cq)) {
1022                 ret = PTR_ERR(cq);
1023                 goto err_file;
1024         }
1025
1026         cq->device        = ib_dev;
1027         cq->uobject       = &obj->uobject;
1028         cq->comp_handler  = ib_uverbs_comp_handler;
1029         cq->event_handler = ib_uverbs_cq_event_handler;
1030         cq->cq_context    = ev_file ? &ev_file->ev_queue : NULL;
1031         atomic_set(&cq->usecnt, 0);
1032
1033         obj->uobject.object = cq;
1034         memset(&resp, 0, sizeof resp);
1035         resp.base.cq_handle = obj->uobject.id;
1036         resp.base.cqe       = cq->cqe;
1037
1038         resp.response_length = offsetof(typeof(resp), response_length) +
1039                 sizeof(resp.response_length);
1040
1041         cq->res.type = RDMA_RESTRACK_CQ;
1042         rdma_restrack_add(&cq->res);
1043
1044         ret = cb(file, obj, &resp, ucore, context);
1045         if (ret)
1046                 goto err_cb;
1047
1048         ret = uobj_alloc_commit(&obj->uobject, 0);
1049         if (ret)
1050                 return ERR_PTR(ret);
1051         return obj;
1052
1053 err_cb:
1054         ib_destroy_cq(cq);
1055
1056 err_file:
1057         if (ev_file)
1058                 ib_uverbs_release_ucq(file, ev_file, obj);
1059
1060 err:
1061         uobj_alloc_abort(&obj->uobject);
1062
1063         return ERR_PTR(ret);
1064 }
1065
1066 static int ib_uverbs_create_cq_cb(struct ib_uverbs_file *file,
1067                                   struct ib_ucq_object *obj,
1068                                   struct ib_uverbs_ex_create_cq_resp *resp,
1069                                   struct ib_udata *ucore, void *context)
1070 {
1071         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1072                 return -EFAULT;
1073
1074         return 0;
1075 }
1076
1077 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1078                             const char __user *buf, int in_len,
1079                             int out_len)
1080 {
1081         struct ib_uverbs_create_cq      cmd;
1082         struct ib_uverbs_ex_create_cq   cmd_ex;
1083         struct ib_uverbs_create_cq_resp resp;
1084         struct ib_udata                 ucore;
1085         struct ib_udata                 uhw;
1086         struct ib_ucq_object           *obj;
1087
1088         if (out_len < sizeof(resp))
1089                 return -ENOSPC;
1090
1091         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1092                 return -EFAULT;
1093
1094         ib_uverbs_init_udata(&ucore, buf, u64_to_user_ptr(cmd.response),
1095                              sizeof(cmd), sizeof(resp));
1096
1097         ib_uverbs_init_udata(&uhw, buf + sizeof(cmd),
1098                    u64_to_user_ptr(cmd.response) + sizeof(resp),
1099                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1100                    out_len - sizeof(resp));
1101
1102         memset(&cmd_ex, 0, sizeof(cmd_ex));
1103         cmd_ex.user_handle = cmd.user_handle;
1104         cmd_ex.cqe = cmd.cqe;
1105         cmd_ex.comp_vector = cmd.comp_vector;
1106         cmd_ex.comp_channel = cmd.comp_channel;
1107
1108         obj = create_cq(file, &ucore, &uhw, &cmd_ex,
1109                         offsetof(typeof(cmd_ex), comp_channel) +
1110                         sizeof(cmd.comp_channel), ib_uverbs_create_cq_cb,
1111                         NULL);
1112
1113         if (IS_ERR(obj))
1114                 return PTR_ERR(obj);
1115
1116         return in_len;
1117 }
1118
1119 static int ib_uverbs_ex_create_cq_cb(struct ib_uverbs_file *file,
1120                                      struct ib_ucq_object *obj,
1121                                      struct ib_uverbs_ex_create_cq_resp *resp,
1122                                      struct ib_udata *ucore, void *context)
1123 {
1124         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1125                 return -EFAULT;
1126
1127         return 0;
1128 }
1129
1130 int ib_uverbs_ex_create_cq(struct ib_uverbs_file *file,
1131                            struct ib_udata *ucore,
1132                            struct ib_udata *uhw)
1133 {
1134         struct ib_uverbs_ex_create_cq_resp resp;
1135         struct ib_uverbs_ex_create_cq  cmd;
1136         struct ib_ucq_object           *obj;
1137         int err;
1138
1139         if (ucore->inlen < sizeof(cmd))
1140                 return -EINVAL;
1141
1142         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
1143         if (err)
1144                 return err;
1145
1146         if (cmd.comp_mask)
1147                 return -EINVAL;
1148
1149         if (cmd.reserved)
1150                 return -EINVAL;
1151
1152         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1153                              sizeof(resp.response_length)))
1154                 return -ENOSPC;
1155
1156         obj = create_cq(file, ucore, uhw, &cmd,
1157                         min(ucore->inlen, sizeof(cmd)),
1158                         ib_uverbs_ex_create_cq_cb, NULL);
1159
1160         return PTR_ERR_OR_ZERO(obj);
1161 }
1162
1163 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1164                             const char __user *buf, int in_len,
1165                             int out_len)
1166 {
1167         struct ib_uverbs_resize_cq      cmd;
1168         struct ib_uverbs_resize_cq_resp resp = {};
1169         struct ib_udata                 udata;
1170         struct ib_cq                    *cq;
1171         int                             ret = -EINVAL;
1172
1173         if (copy_from_user(&cmd, buf, sizeof cmd))
1174                 return -EFAULT;
1175
1176         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
1177                    u64_to_user_ptr(cmd.response) + sizeof(resp),
1178                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1179                    out_len - sizeof(resp));
1180
1181         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
1182         if (!cq)
1183                 return -EINVAL;
1184
1185         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1186         if (ret)
1187                 goto out;
1188
1189         resp.cqe = cq->cqe;
1190
1191         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp.cqe))
1192                 ret = -EFAULT;
1193
1194 out:
1195         uobj_put_obj_read(cq);
1196
1197         return ret ? ret : in_len;
1198 }
1199
1200 static int copy_wc_to_user(struct ib_device *ib_dev, void __user *dest,
1201                            struct ib_wc *wc)
1202 {
1203         struct ib_uverbs_wc tmp;
1204
1205         tmp.wr_id               = wc->wr_id;
1206         tmp.status              = wc->status;
1207         tmp.opcode              = wc->opcode;
1208         tmp.vendor_err          = wc->vendor_err;
1209         tmp.byte_len            = wc->byte_len;
1210         tmp.ex.imm_data         = wc->ex.imm_data;
1211         tmp.qp_num              = wc->qp->qp_num;
1212         tmp.src_qp              = wc->src_qp;
1213         tmp.wc_flags            = wc->wc_flags;
1214         tmp.pkey_index          = wc->pkey_index;
1215         if (rdma_cap_opa_ah(ib_dev, wc->port_num))
1216                 tmp.slid        = OPA_TO_IB_UCAST_LID(wc->slid);
1217         else
1218                 tmp.slid        = ib_lid_cpu16(wc->slid);
1219         tmp.sl                  = wc->sl;
1220         tmp.dlid_path_bits      = wc->dlid_path_bits;
1221         tmp.port_num            = wc->port_num;
1222         tmp.reserved            = 0;
1223
1224         if (copy_to_user(dest, &tmp, sizeof tmp))
1225                 return -EFAULT;
1226
1227         return 0;
1228 }
1229
1230 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1231                           const char __user *buf, int in_len,
1232                           int out_len)
1233 {
1234         struct ib_uverbs_poll_cq       cmd;
1235         struct ib_uverbs_poll_cq_resp  resp;
1236         u8 __user                     *header_ptr;
1237         u8 __user                     *data_ptr;
1238         struct ib_cq                  *cq;
1239         struct ib_wc                   wc;
1240         int                            ret;
1241
1242         if (copy_from_user(&cmd, buf, sizeof cmd))
1243                 return -EFAULT;
1244
1245         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
1246         if (!cq)
1247                 return -EINVAL;
1248
1249         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1250         header_ptr = u64_to_user_ptr(cmd.response);
1251         data_ptr = header_ptr + sizeof resp;
1252
1253         memset(&resp, 0, sizeof resp);
1254         while (resp.count < cmd.ne) {
1255                 ret = ib_poll_cq(cq, 1, &wc);
1256                 if (ret < 0)
1257                         goto out_put;
1258                 if (!ret)
1259                         break;
1260
1261                 ret = copy_wc_to_user(cq->device, data_ptr, &wc);
1262                 if (ret)
1263                         goto out_put;
1264
1265                 data_ptr += sizeof(struct ib_uverbs_wc);
1266                 ++resp.count;
1267         }
1268
1269         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1270                 ret = -EFAULT;
1271                 goto out_put;
1272         }
1273
1274         ret = in_len;
1275
1276 out_put:
1277         uobj_put_obj_read(cq);
1278         return ret;
1279 }
1280
1281 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1282                                 const char __user *buf, int in_len,
1283                                 int out_len)
1284 {
1285         struct ib_uverbs_req_notify_cq cmd;
1286         struct ib_cq                  *cq;
1287
1288         if (copy_from_user(&cmd, buf, sizeof cmd))
1289                 return -EFAULT;
1290
1291         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
1292         if (!cq)
1293                 return -EINVAL;
1294
1295         ib_req_notify_cq(cq, cmd.solicited_only ?
1296                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1297
1298         uobj_put_obj_read(cq);
1299
1300         return in_len;
1301 }
1302
1303 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1304                              const char __user *buf, int in_len,
1305                              int out_len)
1306 {
1307         struct ib_uverbs_destroy_cq      cmd;
1308         struct ib_uverbs_destroy_cq_resp resp;
1309         struct ib_uobject               *uobj;
1310         struct ib_ucq_object            *obj;
1311
1312         if (copy_from_user(&cmd, buf, sizeof cmd))
1313                 return -EFAULT;
1314
1315         uobj = uobj_get_destroy(UVERBS_OBJECT_CQ, cmd.cq_handle, file);
1316         if (IS_ERR(uobj))
1317                 return PTR_ERR(uobj);
1318
1319         obj = container_of(uobj, struct ib_ucq_object, uobject);
1320         memset(&resp, 0, sizeof(resp));
1321         resp.comp_events_reported  = obj->comp_events_reported;
1322         resp.async_events_reported = obj->async_events_reported;
1323
1324         uobj_put_destroy(uobj);
1325
1326         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
1327                 return -EFAULT;
1328
1329         return in_len;
1330 }
1331
1332 static int create_qp(struct ib_uverbs_file *file,
1333                      struct ib_udata *ucore,
1334                      struct ib_udata *uhw,
1335                      struct ib_uverbs_ex_create_qp *cmd,
1336                      size_t cmd_sz,
1337                      int (*cb)(struct ib_uverbs_file *file,
1338                                struct ib_uverbs_ex_create_qp_resp *resp,
1339                                struct ib_udata *udata),
1340                      void *context)
1341 {
1342         struct ib_uqp_object            *obj;
1343         struct ib_device                *device;
1344         struct ib_pd                    *pd = NULL;
1345         struct ib_xrcd                  *xrcd = NULL;
1346         struct ib_uobject               *xrcd_uobj = ERR_PTR(-ENOENT);
1347         struct ib_cq                    *scq = NULL, *rcq = NULL;
1348         struct ib_srq                   *srq = NULL;
1349         struct ib_qp                    *qp;
1350         char                            *buf;
1351         struct ib_qp_init_attr          attr = {};
1352         struct ib_uverbs_ex_create_qp_resp resp;
1353         int                             ret;
1354         struct ib_rwq_ind_table *ind_tbl = NULL;
1355         bool has_sq = true;
1356         struct ib_device *ib_dev;
1357
1358         if (cmd->qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1359                 return -EPERM;
1360
1361         obj = (struct ib_uqp_object *)uobj_alloc(UVERBS_OBJECT_QP, file,
1362                                                  &ib_dev);
1363         if (IS_ERR(obj))
1364                 return PTR_ERR(obj);
1365         obj->uxrcd = NULL;
1366         obj->uevent.uobject.user_handle = cmd->user_handle;
1367         mutex_init(&obj->mcast_lock);
1368
1369         if (cmd_sz >= offsetof(typeof(*cmd), rwq_ind_tbl_handle) +
1370                       sizeof(cmd->rwq_ind_tbl_handle) &&
1371                       (cmd->comp_mask & IB_UVERBS_CREATE_QP_MASK_IND_TABLE)) {
1372                 ind_tbl = uobj_get_obj_read(rwq_ind_table,
1373                                             UVERBS_OBJECT_RWQ_IND_TBL,
1374                                             cmd->rwq_ind_tbl_handle, file);
1375                 if (!ind_tbl) {
1376                         ret = -EINVAL;
1377                         goto err_put;
1378                 }
1379
1380                 attr.rwq_ind_tbl = ind_tbl;
1381         }
1382
1383         if (cmd_sz > sizeof(*cmd) &&
1384             !ib_is_udata_cleared(ucore, sizeof(*cmd),
1385                                  cmd_sz - sizeof(*cmd))) {
1386                 ret = -EOPNOTSUPP;
1387                 goto err_put;
1388         }
1389
1390         if (ind_tbl && (cmd->max_recv_wr || cmd->max_recv_sge || cmd->is_srq)) {
1391                 ret = -EINVAL;
1392                 goto err_put;
1393         }
1394
1395         if (ind_tbl && !cmd->max_send_wr)
1396                 has_sq = false;
1397
1398         if (cmd->qp_type == IB_QPT_XRC_TGT) {
1399                 xrcd_uobj = uobj_get_read(UVERBS_OBJECT_XRCD, cmd->pd_handle,
1400                                           file);
1401
1402                 if (IS_ERR(xrcd_uobj)) {
1403                         ret = -EINVAL;
1404                         goto err_put;
1405                 }
1406
1407                 xrcd = (struct ib_xrcd *)xrcd_uobj->object;
1408                 if (!xrcd) {
1409                         ret = -EINVAL;
1410                         goto err_put;
1411                 }
1412                 device = xrcd->device;
1413         } else {
1414                 if (cmd->qp_type == IB_QPT_XRC_INI) {
1415                         cmd->max_recv_wr = 0;
1416                         cmd->max_recv_sge = 0;
1417                 } else {
1418                         if (cmd->is_srq) {
1419                                 srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ,
1420                                                         cmd->srq_handle, file);
1421                                 if (!srq || srq->srq_type == IB_SRQT_XRC) {
1422                                         ret = -EINVAL;
1423                                         goto err_put;
1424                                 }
1425                         }
1426
1427                         if (!ind_tbl) {
1428                                 if (cmd->recv_cq_handle != cmd->send_cq_handle) {
1429                                         rcq = uobj_get_obj_read(
1430                                                 cq, UVERBS_OBJECT_CQ,
1431                                                 cmd->recv_cq_handle, file);
1432                                         if (!rcq) {
1433                                                 ret = -EINVAL;
1434                                                 goto err_put;
1435                                         }
1436                                 }
1437                         }
1438                 }
1439
1440                 if (has_sq)
1441                         scq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ,
1442                                                 cmd->send_cq_handle, file);
1443                 if (!ind_tbl)
1444                         rcq = rcq ?: scq;
1445                 pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd->pd_handle,
1446                                        file);
1447                 if (!pd || (!scq && has_sq)) {
1448                         ret = -EINVAL;
1449                         goto err_put;
1450                 }
1451
1452                 device = pd->device;
1453         }
1454
1455         attr.event_handler = ib_uverbs_qp_event_handler;
1456         attr.qp_context    = file;
1457         attr.send_cq       = scq;
1458         attr.recv_cq       = rcq;
1459         attr.srq           = srq;
1460         attr.xrcd          = xrcd;
1461         attr.sq_sig_type   = cmd->sq_sig_all ? IB_SIGNAL_ALL_WR :
1462                                               IB_SIGNAL_REQ_WR;
1463         attr.qp_type       = cmd->qp_type;
1464         attr.create_flags  = 0;
1465
1466         attr.cap.max_send_wr     = cmd->max_send_wr;
1467         attr.cap.max_recv_wr     = cmd->max_recv_wr;
1468         attr.cap.max_send_sge    = cmd->max_send_sge;
1469         attr.cap.max_recv_sge    = cmd->max_recv_sge;
1470         attr.cap.max_inline_data = cmd->max_inline_data;
1471
1472         obj->uevent.events_reported     = 0;
1473         INIT_LIST_HEAD(&obj->uevent.event_list);
1474         INIT_LIST_HEAD(&obj->mcast_list);
1475
1476         if (cmd_sz >= offsetof(typeof(*cmd), create_flags) +
1477                       sizeof(cmd->create_flags))
1478                 attr.create_flags = cmd->create_flags;
1479
1480         if (attr.create_flags & ~(IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
1481                                 IB_QP_CREATE_CROSS_CHANNEL |
1482                                 IB_QP_CREATE_MANAGED_SEND |
1483                                 IB_QP_CREATE_MANAGED_RECV |
1484                                 IB_QP_CREATE_SCATTER_FCS |
1485                                 IB_QP_CREATE_CVLAN_STRIPPING |
1486                                 IB_QP_CREATE_SOURCE_QPN |
1487                                 IB_QP_CREATE_PCI_WRITE_END_PADDING)) {
1488                 ret = -EINVAL;
1489                 goto err_put;
1490         }
1491
1492         if (attr.create_flags & IB_QP_CREATE_SOURCE_QPN) {
1493                 if (!capable(CAP_NET_RAW)) {
1494                         ret = -EPERM;
1495                         goto err_put;
1496                 }
1497
1498                 attr.source_qpn = cmd->source_qpn;
1499         }
1500
1501         buf = (void *)cmd + sizeof(*cmd);
1502         if (cmd_sz > sizeof(*cmd))
1503                 if (!(buf[0] == 0 && !memcmp(buf, buf + 1,
1504                                              cmd_sz - sizeof(*cmd) - 1))) {
1505                         ret = -EINVAL;
1506                         goto err_put;
1507                 }
1508
1509         if (cmd->qp_type == IB_QPT_XRC_TGT)
1510                 qp = ib_create_qp(pd, &attr);
1511         else
1512                 qp = _ib_create_qp(device, pd, &attr, uhw,
1513                                    &obj->uevent.uobject);
1514
1515         if (IS_ERR(qp)) {
1516                 ret = PTR_ERR(qp);
1517                 goto err_put;
1518         }
1519
1520         if (cmd->qp_type != IB_QPT_XRC_TGT) {
1521                 ret = ib_create_qp_security(qp, device);
1522                 if (ret)
1523                         goto err_cb;
1524
1525                 qp->real_qp       = qp;
1526                 qp->pd            = pd;
1527                 qp->send_cq       = attr.send_cq;
1528                 qp->recv_cq       = attr.recv_cq;
1529                 qp->srq           = attr.srq;
1530                 qp->rwq_ind_tbl   = ind_tbl;
1531                 qp->event_handler = attr.event_handler;
1532                 qp->qp_context    = attr.qp_context;
1533                 qp->qp_type       = attr.qp_type;
1534                 atomic_set(&qp->usecnt, 0);
1535                 atomic_inc(&pd->usecnt);
1536                 qp->port = 0;
1537                 if (attr.send_cq)
1538                         atomic_inc(&attr.send_cq->usecnt);
1539                 if (attr.recv_cq)
1540                         atomic_inc(&attr.recv_cq->usecnt);
1541                 if (attr.srq)
1542                         atomic_inc(&attr.srq->usecnt);
1543                 if (ind_tbl)
1544                         atomic_inc(&ind_tbl->usecnt);
1545         } else {
1546                 /* It is done in _ib_create_qp for other QP types */
1547                 qp->uobject = &obj->uevent.uobject;
1548         }
1549
1550         obj->uevent.uobject.object = qp;
1551
1552         memset(&resp, 0, sizeof resp);
1553         resp.base.qpn             = qp->qp_num;
1554         resp.base.qp_handle       = obj->uevent.uobject.id;
1555         resp.base.max_recv_sge    = attr.cap.max_recv_sge;
1556         resp.base.max_send_sge    = attr.cap.max_send_sge;
1557         resp.base.max_recv_wr     = attr.cap.max_recv_wr;
1558         resp.base.max_send_wr     = attr.cap.max_send_wr;
1559         resp.base.max_inline_data = attr.cap.max_inline_data;
1560
1561         resp.response_length = offsetof(typeof(resp), response_length) +
1562                                sizeof(resp.response_length);
1563
1564         ret = cb(file, &resp, ucore);
1565         if (ret)
1566                 goto err_cb;
1567
1568         if (xrcd) {
1569                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object,
1570                                           uobject);
1571                 atomic_inc(&obj->uxrcd->refcnt);
1572                 uobj_put_read(xrcd_uobj);
1573         }
1574
1575         if (pd)
1576                 uobj_put_obj_read(pd);
1577         if (scq)
1578                 uobj_put_obj_read(scq);
1579         if (rcq && rcq != scq)
1580                 uobj_put_obj_read(rcq);
1581         if (srq)
1582                 uobj_put_obj_read(srq);
1583         if (ind_tbl)
1584                 uobj_put_obj_read(ind_tbl);
1585
1586         return uobj_alloc_commit(&obj->uevent.uobject, 0);
1587 err_cb:
1588         ib_destroy_qp(qp);
1589
1590 err_put:
1591         if (!IS_ERR(xrcd_uobj))
1592                 uobj_put_read(xrcd_uobj);
1593         if (pd)
1594                 uobj_put_obj_read(pd);
1595         if (scq)
1596                 uobj_put_obj_read(scq);
1597         if (rcq && rcq != scq)
1598                 uobj_put_obj_read(rcq);
1599         if (srq)
1600                 uobj_put_obj_read(srq);
1601         if (ind_tbl)
1602                 uobj_put_obj_read(ind_tbl);
1603
1604         uobj_alloc_abort(&obj->uevent.uobject);
1605         return ret;
1606 }
1607
1608 static int ib_uverbs_create_qp_cb(struct ib_uverbs_file *file,
1609                                   struct ib_uverbs_ex_create_qp_resp *resp,
1610                                   struct ib_udata *ucore)
1611 {
1612         if (ib_copy_to_udata(ucore, &resp->base, sizeof(resp->base)))
1613                 return -EFAULT;
1614
1615         return 0;
1616 }
1617
1618 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1619                             const char __user *buf, int in_len,
1620                             int out_len)
1621 {
1622         struct ib_uverbs_create_qp      cmd;
1623         struct ib_uverbs_ex_create_qp   cmd_ex;
1624         struct ib_udata                 ucore;
1625         struct ib_udata                 uhw;
1626         ssize_t resp_size = sizeof(struct ib_uverbs_create_qp_resp);
1627         int                             err;
1628
1629         if (out_len < resp_size)
1630                 return -ENOSPC;
1631
1632         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1633                 return -EFAULT;
1634
1635         ib_uverbs_init_udata(&ucore, buf, u64_to_user_ptr(cmd.response),
1636                    sizeof(cmd), resp_size);
1637         ib_uverbs_init_udata(&uhw, buf + sizeof(cmd),
1638                    u64_to_user_ptr(cmd.response) + resp_size,
1639                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1640                    out_len - resp_size);
1641
1642         memset(&cmd_ex, 0, sizeof(cmd_ex));
1643         cmd_ex.user_handle = cmd.user_handle;
1644         cmd_ex.pd_handle = cmd.pd_handle;
1645         cmd_ex.send_cq_handle = cmd.send_cq_handle;
1646         cmd_ex.recv_cq_handle = cmd.recv_cq_handle;
1647         cmd_ex.srq_handle = cmd.srq_handle;
1648         cmd_ex.max_send_wr = cmd.max_send_wr;
1649         cmd_ex.max_recv_wr = cmd.max_recv_wr;
1650         cmd_ex.max_send_sge = cmd.max_send_sge;
1651         cmd_ex.max_recv_sge = cmd.max_recv_sge;
1652         cmd_ex.max_inline_data = cmd.max_inline_data;
1653         cmd_ex.sq_sig_all = cmd.sq_sig_all;
1654         cmd_ex.qp_type = cmd.qp_type;
1655         cmd_ex.is_srq = cmd.is_srq;
1656
1657         err = create_qp(file, &ucore, &uhw, &cmd_ex,
1658                         offsetof(typeof(cmd_ex), is_srq) +
1659                         sizeof(cmd.is_srq), ib_uverbs_create_qp_cb,
1660                         NULL);
1661
1662         if (err)
1663                 return err;
1664
1665         return in_len;
1666 }
1667
1668 static int ib_uverbs_ex_create_qp_cb(struct ib_uverbs_file *file,
1669                                      struct ib_uverbs_ex_create_qp_resp *resp,
1670                                      struct ib_udata *ucore)
1671 {
1672         if (ib_copy_to_udata(ucore, resp, resp->response_length))
1673                 return -EFAULT;
1674
1675         return 0;
1676 }
1677
1678 int ib_uverbs_ex_create_qp(struct ib_uverbs_file *file,
1679                            struct ib_udata *ucore,
1680                            struct ib_udata *uhw)
1681 {
1682         struct ib_uverbs_ex_create_qp_resp resp;
1683         struct ib_uverbs_ex_create_qp cmd = {0};
1684         int err;
1685
1686         if (ucore->inlen < (offsetof(typeof(cmd), comp_mask) +
1687                             sizeof(cmd.comp_mask)))
1688                 return -EINVAL;
1689
1690         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
1691         if (err)
1692                 return err;
1693
1694         if (cmd.comp_mask & ~IB_UVERBS_CREATE_QP_SUP_COMP_MASK)
1695                 return -EINVAL;
1696
1697         if (cmd.reserved)
1698                 return -EINVAL;
1699
1700         if (ucore->outlen < (offsetof(typeof(resp), response_length) +
1701                              sizeof(resp.response_length)))
1702                 return -ENOSPC;
1703
1704         err = create_qp(file, ucore, uhw, &cmd,
1705                         min(ucore->inlen, sizeof(cmd)),
1706                         ib_uverbs_ex_create_qp_cb, NULL);
1707
1708         if (err)
1709                 return err;
1710
1711         return 0;
1712 }
1713
1714 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1715                           const char __user *buf, int in_len, int out_len)
1716 {
1717         struct ib_uverbs_open_qp        cmd;
1718         struct ib_uverbs_create_qp_resp resp;
1719         struct ib_udata                 udata;
1720         struct ib_uqp_object           *obj;
1721         struct ib_xrcd                 *xrcd;
1722         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1723         struct ib_qp                   *qp;
1724         struct ib_qp_open_attr          attr;
1725         int ret;
1726         struct ib_device *ib_dev;
1727
1728         if (out_len < sizeof resp)
1729                 return -ENOSPC;
1730
1731         if (copy_from_user(&cmd, buf, sizeof cmd))
1732                 return -EFAULT;
1733
1734         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
1735                    u64_to_user_ptr(cmd.response) + sizeof(resp),
1736                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
1737                    out_len - sizeof(resp));
1738
1739         obj = (struct ib_uqp_object *)uobj_alloc(UVERBS_OBJECT_QP, file,
1740                                                  &ib_dev);
1741         if (IS_ERR(obj))
1742                 return PTR_ERR(obj);
1743
1744         xrcd_uobj = uobj_get_read(UVERBS_OBJECT_XRCD, cmd.pd_handle, file);
1745         if (IS_ERR(xrcd_uobj)) {
1746                 ret = -EINVAL;
1747                 goto err_put;
1748         }
1749
1750         xrcd = (struct ib_xrcd *)xrcd_uobj->object;
1751         if (!xrcd) {
1752                 ret = -EINVAL;
1753                 goto err_xrcd;
1754         }
1755
1756         attr.event_handler = ib_uverbs_qp_event_handler;
1757         attr.qp_context    = file;
1758         attr.qp_num        = cmd.qpn;
1759         attr.qp_type       = cmd.qp_type;
1760
1761         obj->uevent.events_reported = 0;
1762         INIT_LIST_HEAD(&obj->uevent.event_list);
1763         INIT_LIST_HEAD(&obj->mcast_list);
1764
1765         qp = ib_open_qp(xrcd, &attr);
1766         if (IS_ERR(qp)) {
1767                 ret = PTR_ERR(qp);
1768                 goto err_xrcd;
1769         }
1770
1771         obj->uevent.uobject.object = qp;
1772         obj->uevent.uobject.user_handle = cmd.user_handle;
1773
1774         memset(&resp, 0, sizeof resp);
1775         resp.qpn       = qp->qp_num;
1776         resp.qp_handle = obj->uevent.uobject.id;
1777
1778         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
1779                 ret = -EFAULT;
1780                 goto err_destroy;
1781         }
1782
1783         obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
1784         atomic_inc(&obj->uxrcd->refcnt);
1785         qp->uobject = &obj->uevent.uobject;
1786         uobj_put_read(xrcd_uobj);
1787
1788         return uobj_alloc_commit(&obj->uevent.uobject, in_len);
1789
1790 err_destroy:
1791         ib_destroy_qp(qp);
1792 err_xrcd:
1793         uobj_put_read(xrcd_uobj);
1794 err_put:
1795         uobj_alloc_abort(&obj->uevent.uobject);
1796         return ret;
1797 }
1798
1799 static void copy_ah_attr_to_uverbs(struct ib_uverbs_qp_dest *uverb_attr,
1800                                    struct rdma_ah_attr *rdma_attr)
1801 {
1802         const struct ib_global_route   *grh;
1803
1804         uverb_attr->dlid              = rdma_ah_get_dlid(rdma_attr);
1805         uverb_attr->sl                = rdma_ah_get_sl(rdma_attr);
1806         uverb_attr->src_path_bits     = rdma_ah_get_path_bits(rdma_attr);
1807         uverb_attr->static_rate       = rdma_ah_get_static_rate(rdma_attr);
1808         uverb_attr->is_global         = !!(rdma_ah_get_ah_flags(rdma_attr) &
1809                                          IB_AH_GRH);
1810         if (uverb_attr->is_global) {
1811                 grh = rdma_ah_read_grh(rdma_attr);
1812                 memcpy(uverb_attr->dgid, grh->dgid.raw, 16);
1813                 uverb_attr->flow_label        = grh->flow_label;
1814                 uverb_attr->sgid_index        = grh->sgid_index;
1815                 uverb_attr->hop_limit         = grh->hop_limit;
1816                 uverb_attr->traffic_class     = grh->traffic_class;
1817         }
1818         uverb_attr->port_num          = rdma_ah_get_port_num(rdma_attr);
1819 }
1820
1821 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1822                            const char __user *buf, int in_len,
1823                            int out_len)
1824 {
1825         struct ib_uverbs_query_qp      cmd;
1826         struct ib_uverbs_query_qp_resp resp;
1827         struct ib_qp                   *qp;
1828         struct ib_qp_attr              *attr;
1829         struct ib_qp_init_attr         *init_attr;
1830         int                            ret;
1831
1832         if (copy_from_user(&cmd, buf, sizeof cmd))
1833                 return -EFAULT;
1834
1835         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1836         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1837         if (!attr || !init_attr) {
1838                 ret = -ENOMEM;
1839                 goto out;
1840         }
1841
1842         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
1843         if (!qp) {
1844                 ret = -EINVAL;
1845                 goto out;
1846         }
1847
1848         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1849
1850         uobj_put_obj_read(qp);
1851
1852         if (ret)
1853                 goto out;
1854
1855         memset(&resp, 0, sizeof resp);
1856
1857         resp.qp_state               = attr->qp_state;
1858         resp.cur_qp_state           = attr->cur_qp_state;
1859         resp.path_mtu               = attr->path_mtu;
1860         resp.path_mig_state         = attr->path_mig_state;
1861         resp.qkey                   = attr->qkey;
1862         resp.rq_psn                 = attr->rq_psn;
1863         resp.sq_psn                 = attr->sq_psn;
1864         resp.dest_qp_num            = attr->dest_qp_num;
1865         resp.qp_access_flags        = attr->qp_access_flags;
1866         resp.pkey_index             = attr->pkey_index;
1867         resp.alt_pkey_index         = attr->alt_pkey_index;
1868         resp.sq_draining            = attr->sq_draining;
1869         resp.max_rd_atomic          = attr->max_rd_atomic;
1870         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1871         resp.min_rnr_timer          = attr->min_rnr_timer;
1872         resp.port_num               = attr->port_num;
1873         resp.timeout                = attr->timeout;
1874         resp.retry_cnt              = attr->retry_cnt;
1875         resp.rnr_retry              = attr->rnr_retry;
1876         resp.alt_port_num           = attr->alt_port_num;
1877         resp.alt_timeout            = attr->alt_timeout;
1878
1879         copy_ah_attr_to_uverbs(&resp.dest, &attr->ah_attr);
1880         copy_ah_attr_to_uverbs(&resp.alt_dest, &attr->alt_ah_attr);
1881
1882         resp.max_send_wr            = init_attr->cap.max_send_wr;
1883         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1884         resp.max_send_sge           = init_attr->cap.max_send_sge;
1885         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1886         resp.max_inline_data        = init_attr->cap.max_inline_data;
1887         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1888
1889         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
1890                 ret = -EFAULT;
1891
1892 out:
1893         kfree(attr);
1894         kfree(init_attr);
1895
1896         return ret ? ret : in_len;
1897 }
1898
1899 /* Remove ignored fields set in the attribute mask */
1900 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1901 {
1902         switch (qp_type) {
1903         case IB_QPT_XRC_INI:
1904                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1905         case IB_QPT_XRC_TGT:
1906                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1907                                 IB_QP_RNR_RETRY);
1908         default:
1909                 return mask;
1910         }
1911 }
1912
1913 static void copy_ah_attr_from_uverbs(struct ib_device *dev,
1914                                      struct rdma_ah_attr *rdma_attr,
1915                                      struct ib_uverbs_qp_dest *uverb_attr)
1916 {
1917         rdma_attr->type = rdma_ah_find_type(dev, uverb_attr->port_num);
1918         if (uverb_attr->is_global) {
1919                 rdma_ah_set_grh(rdma_attr, NULL,
1920                                 uverb_attr->flow_label,
1921                                 uverb_attr->sgid_index,
1922                                 uverb_attr->hop_limit,
1923                                 uverb_attr->traffic_class);
1924                 rdma_ah_set_dgid_raw(rdma_attr, uverb_attr->dgid);
1925         } else {
1926                 rdma_ah_set_ah_flags(rdma_attr, 0);
1927         }
1928         rdma_ah_set_dlid(rdma_attr, uverb_attr->dlid);
1929         rdma_ah_set_sl(rdma_attr, uverb_attr->sl);
1930         rdma_ah_set_path_bits(rdma_attr, uverb_attr->src_path_bits);
1931         rdma_ah_set_static_rate(rdma_attr, uverb_attr->static_rate);
1932         rdma_ah_set_port_num(rdma_attr, uverb_attr->port_num);
1933         rdma_ah_set_make_grd(rdma_attr, false);
1934 }
1935
1936 static int modify_qp(struct ib_uverbs_file *file,
1937                      struct ib_uverbs_ex_modify_qp *cmd, struct ib_udata *udata)
1938 {
1939         struct ib_qp_attr *attr;
1940         struct ib_qp *qp;
1941         int ret;
1942
1943         attr = kzalloc(sizeof(*attr), GFP_KERNEL);
1944         if (!attr)
1945                 return -ENOMEM;
1946
1947         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd->base.qp_handle, file);
1948         if (!qp) {
1949                 ret = -EINVAL;
1950                 goto out;
1951         }
1952
1953         if ((cmd->base.attr_mask & IB_QP_PORT) &&
1954             !rdma_is_port_valid(qp->device, cmd->base.port_num)) {
1955                 ret = -EINVAL;
1956                 goto release_qp;
1957         }
1958
1959         if ((cmd->base.attr_mask & IB_QP_AV)) {
1960                 if (!rdma_is_port_valid(qp->device, cmd->base.dest.port_num)) {
1961                         ret = -EINVAL;
1962                         goto release_qp;
1963                 }
1964
1965                 if (cmd->base.attr_mask & IB_QP_STATE &&
1966                     cmd->base.qp_state == IB_QPS_RTR) {
1967                 /* We are in INIT->RTR TRANSITION (if we are not,
1968                  * this transition will be rejected in subsequent checks).
1969                  * In the INIT->RTR transition, we cannot have IB_QP_PORT set,
1970                  * but the IB_QP_STATE flag is required.
1971                  *
1972                  * Since kernel 3.14 (commit dbf727de7440), the uverbs driver,
1973                  * when IB_QP_AV is set, has required inclusion of a valid
1974                  * port number in the primary AV. (AVs are created and handled
1975                  * differently for infiniband and ethernet (RoCE) ports).
1976                  *
1977                  * Check the port number included in the primary AV against
1978                  * the port number in the qp struct, which was set (and saved)
1979                  * in the RST->INIT transition.
1980                  */
1981                         if (cmd->base.dest.port_num != qp->real_qp->port) {
1982                                 ret = -EINVAL;
1983                                 goto release_qp;
1984                         }
1985                 } else {
1986                 /* We are in SQD->SQD. (If we are not, this transition will
1987                  * be rejected later in the verbs layer checks).
1988                  * Check for both IB_QP_PORT and IB_QP_AV, these can be set
1989                  * together in the SQD->SQD transition.
1990                  *
1991                  * If only IP_QP_AV was set, add in IB_QP_PORT as well (the
1992                  * verbs layer driver does not track primary port changes
1993                  * resulting from path migration. Thus, in SQD, if the primary
1994                  * AV is modified, the primary port should also be modified).
1995                  *
1996                  * Note that in this transition, the IB_QP_STATE flag
1997                  * is not allowed.
1998                  */
1999                         if (((cmd->base.attr_mask & (IB_QP_AV | IB_QP_PORT))
2000                              == (IB_QP_AV | IB_QP_PORT)) &&
2001                             cmd->base.port_num != cmd->base.dest.port_num) {
2002                                 ret = -EINVAL;
2003                                 goto release_qp;
2004                         }
2005                         if ((cmd->base.attr_mask & (IB_QP_AV | IB_QP_PORT))
2006                             == IB_QP_AV) {
2007                                 cmd->base.attr_mask |= IB_QP_PORT;
2008                                 cmd->base.port_num = cmd->base.dest.port_num;
2009                         }
2010                 }
2011         }
2012
2013         if ((cmd->base.attr_mask & IB_QP_ALT_PATH) &&
2014             (!rdma_is_port_valid(qp->device, cmd->base.alt_port_num) ||
2015             !rdma_is_port_valid(qp->device, cmd->base.alt_dest.port_num) ||
2016             cmd->base.alt_port_num != cmd->base.alt_dest.port_num)) {
2017                 ret = -EINVAL;
2018                 goto release_qp;
2019         }
2020
2021         if ((cmd->base.attr_mask & IB_QP_CUR_STATE &&
2022             cmd->base.cur_qp_state > IB_QPS_ERR) ||
2023             (cmd->base.attr_mask & IB_QP_STATE &&
2024             cmd->base.qp_state > IB_QPS_ERR)) {
2025                 ret = -EINVAL;
2026                 goto release_qp;
2027         }
2028
2029         if (cmd->base.attr_mask & IB_QP_STATE)
2030                 attr->qp_state = cmd->base.qp_state;
2031         if (cmd->base.attr_mask & IB_QP_CUR_STATE)
2032                 attr->cur_qp_state = cmd->base.cur_qp_state;
2033         if (cmd->base.attr_mask & IB_QP_PATH_MTU)
2034                 attr->path_mtu = cmd->base.path_mtu;
2035         if (cmd->base.attr_mask & IB_QP_PATH_MIG_STATE)
2036                 attr->path_mig_state = cmd->base.path_mig_state;
2037         if (cmd->base.attr_mask & IB_QP_QKEY)
2038                 attr->qkey = cmd->base.qkey;
2039         if (cmd->base.attr_mask & IB_QP_RQ_PSN)
2040                 attr->rq_psn = cmd->base.rq_psn;
2041         if (cmd->base.attr_mask & IB_QP_SQ_PSN)
2042                 attr->sq_psn = cmd->base.sq_psn;
2043         if (cmd->base.attr_mask & IB_QP_DEST_QPN)
2044                 attr->dest_qp_num = cmd->base.dest_qp_num;
2045         if (cmd->base.attr_mask & IB_QP_ACCESS_FLAGS)
2046                 attr->qp_access_flags = cmd->base.qp_access_flags;
2047         if (cmd->base.attr_mask & IB_QP_PKEY_INDEX)
2048                 attr->pkey_index = cmd->base.pkey_index;
2049         if (cmd->base.attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY)
2050                 attr->en_sqd_async_notify = cmd->base.en_sqd_async_notify;
2051         if (cmd->base.attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
2052                 attr->max_rd_atomic = cmd->base.max_rd_atomic;
2053         if (cmd->base.attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2054                 attr->max_dest_rd_atomic = cmd->base.max_dest_rd_atomic;
2055         if (cmd->base.attr_mask & IB_QP_MIN_RNR_TIMER)
2056                 attr->min_rnr_timer = cmd->base.min_rnr_timer;
2057         if (cmd->base.attr_mask & IB_QP_PORT)
2058                 attr->port_num = cmd->base.port_num;
2059         if (cmd->base.attr_mask & IB_QP_TIMEOUT)
2060                 attr->timeout = cmd->base.timeout;
2061         if (cmd->base.attr_mask & IB_QP_RETRY_CNT)
2062                 attr->retry_cnt = cmd->base.retry_cnt;
2063         if (cmd->base.attr_mask & IB_QP_RNR_RETRY)
2064                 attr->rnr_retry = cmd->base.rnr_retry;
2065         if (cmd->base.attr_mask & IB_QP_ALT_PATH) {
2066                 attr->alt_port_num = cmd->base.alt_port_num;
2067                 attr->alt_timeout = cmd->base.alt_timeout;
2068                 attr->alt_pkey_index = cmd->base.alt_pkey_index;
2069         }
2070         if (cmd->base.attr_mask & IB_QP_RATE_LIMIT)
2071                 attr->rate_limit = cmd->rate_limit;
2072
2073         if (cmd->base.attr_mask & IB_QP_AV)
2074                 copy_ah_attr_from_uverbs(qp->device, &attr->ah_attr,
2075                                          &cmd->base.dest);
2076
2077         if (cmd->base.attr_mask & IB_QP_ALT_PATH)
2078                 copy_ah_attr_from_uverbs(qp->device, &attr->alt_ah_attr,
2079                                          &cmd->base.alt_dest);
2080
2081         ret = ib_modify_qp_with_udata(qp, attr,
2082                                       modify_qp_mask(qp->qp_type,
2083                                                      cmd->base.attr_mask),
2084                                       udata);
2085
2086 release_qp:
2087         uobj_put_obj_read(qp);
2088 out:
2089         kfree(attr);
2090
2091         return ret;
2092 }
2093
2094 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
2095                             const char __user *buf, int in_len,
2096                             int out_len)
2097 {
2098         struct ib_uverbs_ex_modify_qp cmd = {};
2099         struct ib_udata udata;
2100         int ret;
2101
2102         if (copy_from_user(&cmd.base, buf, sizeof(cmd.base)))
2103                 return -EFAULT;
2104
2105         if (cmd.base.attr_mask &
2106             ~((IB_USER_LEGACY_LAST_QP_ATTR_MASK << 1) - 1))
2107                 return -EOPNOTSUPP;
2108
2109         ib_uverbs_init_udata(&udata, buf + sizeof(cmd.base), NULL,
2110                    in_len - sizeof(cmd.base) - sizeof(struct ib_uverbs_cmd_hdr),
2111                    out_len);
2112
2113         ret = modify_qp(file, &cmd, &udata);
2114         if (ret)
2115                 return ret;
2116
2117         return in_len;
2118 }
2119
2120 int ib_uverbs_ex_modify_qp(struct ib_uverbs_file *file,
2121                            struct ib_udata *ucore,
2122                            struct ib_udata *uhw)
2123 {
2124         struct ib_uverbs_ex_modify_qp cmd = {};
2125         int ret;
2126
2127         /*
2128          * Last bit is reserved for extending the attr_mask by
2129          * using another field.
2130          */
2131         BUILD_BUG_ON(IB_USER_LAST_QP_ATTR_MASK == (1 << 31));
2132
2133         if (ucore->inlen < sizeof(cmd.base))
2134                 return -EINVAL;
2135
2136         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
2137         if (ret)
2138                 return ret;
2139
2140         if (cmd.base.attr_mask &
2141             ~((IB_USER_LAST_QP_ATTR_MASK << 1) - 1))
2142                 return -EOPNOTSUPP;
2143
2144         if (ucore->inlen > sizeof(cmd)) {
2145                 if (!ib_is_udata_cleared(ucore, sizeof(cmd),
2146                                          ucore->inlen - sizeof(cmd)))
2147                         return -EOPNOTSUPP;
2148         }
2149
2150         ret = modify_qp(file, &cmd, uhw);
2151
2152         return ret;
2153 }
2154
2155 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
2156                              const char __user *buf, int in_len,
2157                              int out_len)
2158 {
2159         struct ib_uverbs_destroy_qp      cmd;
2160         struct ib_uverbs_destroy_qp_resp resp;
2161         struct ib_uobject               *uobj;
2162         struct ib_uqp_object            *obj;
2163
2164         if (copy_from_user(&cmd, buf, sizeof cmd))
2165                 return -EFAULT;
2166
2167         uobj = uobj_get_destroy(UVERBS_OBJECT_QP, cmd.qp_handle, file);
2168         if (IS_ERR(uobj))
2169                 return PTR_ERR(uobj);
2170
2171         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2172         memset(&resp, 0, sizeof(resp));
2173         resp.events_reported = obj->uevent.events_reported;
2174
2175         uobj_put_destroy(uobj);
2176
2177         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2178                 return -EFAULT;
2179
2180         return in_len;
2181 }
2182
2183 static void *alloc_wr(size_t wr_size, __u32 num_sge)
2184 {
2185         if (num_sge >= (U32_MAX - ALIGN(wr_size, sizeof (struct ib_sge))) /
2186                        sizeof (struct ib_sge))
2187                 return NULL;
2188
2189         return kmalloc(ALIGN(wr_size, sizeof (struct ib_sge)) +
2190                          num_sge * sizeof (struct ib_sge), GFP_KERNEL);
2191 }
2192
2193 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2194                             const char __user *buf, int in_len,
2195                             int out_len)
2196 {
2197         struct ib_uverbs_post_send      cmd;
2198         struct ib_uverbs_post_send_resp resp;
2199         struct ib_uverbs_send_wr       *user_wr;
2200         struct ib_send_wr              *wr = NULL, *last, *next;
2201         const struct ib_send_wr        *bad_wr;
2202         struct ib_qp                   *qp;
2203         int                             i, sg_ind;
2204         int                             is_ud;
2205         ssize_t                         ret = -EINVAL;
2206         size_t                          next_size;
2207
2208         if (copy_from_user(&cmd, buf, sizeof cmd))
2209                 return -EFAULT;
2210
2211         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2212             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2213                 return -EINVAL;
2214
2215         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2216                 return -EINVAL;
2217
2218         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2219         if (!user_wr)
2220                 return -ENOMEM;
2221
2222         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2223         if (!qp)
2224                 goto out;
2225
2226         is_ud = qp->qp_type == IB_QPT_UD;
2227         sg_ind = 0;
2228         last = NULL;
2229         for (i = 0; i < cmd.wr_count; ++i) {
2230                 if (copy_from_user(user_wr,
2231                                    buf + sizeof cmd + i * cmd.wqe_size,
2232                                    cmd.wqe_size)) {
2233                         ret = -EFAULT;
2234                         goto out_put;
2235                 }
2236
2237                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2238                         ret = -EINVAL;
2239                         goto out_put;
2240                 }
2241
2242                 if (is_ud) {
2243                         struct ib_ud_wr *ud;
2244
2245                         if (user_wr->opcode != IB_WR_SEND &&
2246                             user_wr->opcode != IB_WR_SEND_WITH_IMM) {
2247                                 ret = -EINVAL;
2248                                 goto out_put;
2249                         }
2250
2251                         next_size = sizeof(*ud);
2252                         ud = alloc_wr(next_size, user_wr->num_sge);
2253                         if (!ud) {
2254                                 ret = -ENOMEM;
2255                                 goto out_put;
2256                         }
2257
2258                         ud->ah = uobj_get_obj_read(ah, UVERBS_OBJECT_AH,
2259                                                    user_wr->wr.ud.ah, file);
2260                         if (!ud->ah) {
2261                                 kfree(ud);
2262                                 ret = -EINVAL;
2263                                 goto out_put;
2264                         }
2265                         ud->remote_qpn = user_wr->wr.ud.remote_qpn;
2266                         ud->remote_qkey = user_wr->wr.ud.remote_qkey;
2267
2268                         next = &ud->wr;
2269                 } else if (user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
2270                            user_wr->opcode == IB_WR_RDMA_WRITE ||
2271                            user_wr->opcode == IB_WR_RDMA_READ) {
2272                         struct ib_rdma_wr *rdma;
2273
2274                         next_size = sizeof(*rdma);
2275                         rdma = alloc_wr(next_size, user_wr->num_sge);
2276                         if (!rdma) {
2277                                 ret = -ENOMEM;
2278                                 goto out_put;
2279                         }
2280
2281                         rdma->remote_addr = user_wr->wr.rdma.remote_addr;
2282                         rdma->rkey = user_wr->wr.rdma.rkey;
2283
2284                         next = &rdma->wr;
2285                 } else if (user_wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2286                            user_wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2287                         struct ib_atomic_wr *atomic;
2288
2289                         next_size = sizeof(*atomic);
2290                         atomic = alloc_wr(next_size, user_wr->num_sge);
2291                         if (!atomic) {
2292                                 ret = -ENOMEM;
2293                                 goto out_put;
2294                         }
2295
2296                         atomic->remote_addr = user_wr->wr.atomic.remote_addr;
2297                         atomic->compare_add = user_wr->wr.atomic.compare_add;
2298                         atomic->swap = user_wr->wr.atomic.swap;
2299                         atomic->rkey = user_wr->wr.atomic.rkey;
2300
2301                         next = &atomic->wr;
2302                 } else if (user_wr->opcode == IB_WR_SEND ||
2303                            user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2304                            user_wr->opcode == IB_WR_SEND_WITH_INV) {
2305                         next_size = sizeof(*next);
2306                         next = alloc_wr(next_size, user_wr->num_sge);
2307                         if (!next) {
2308                                 ret = -ENOMEM;
2309                                 goto out_put;
2310                         }
2311                 } else {
2312                         ret = -EINVAL;
2313                         goto out_put;
2314                 }
2315
2316                 if (user_wr->opcode == IB_WR_SEND_WITH_IMM ||
2317                     user_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
2318                         next->ex.imm_data =
2319                                         (__be32 __force) user_wr->ex.imm_data;
2320                 } else if (user_wr->opcode == IB_WR_SEND_WITH_INV) {
2321                         next->ex.invalidate_rkey = user_wr->ex.invalidate_rkey;
2322                 }
2323
2324                 if (!last)
2325                         wr = next;
2326                 else
2327                         last->next = next;
2328                 last = next;
2329
2330                 next->next       = NULL;
2331                 next->wr_id      = user_wr->wr_id;
2332                 next->num_sge    = user_wr->num_sge;
2333                 next->opcode     = user_wr->opcode;
2334                 next->send_flags = user_wr->send_flags;
2335
2336                 if (next->num_sge) {
2337                         next->sg_list = (void *) next +
2338                                 ALIGN(next_size, sizeof(struct ib_sge));
2339                         if (copy_from_user(next->sg_list,
2340                                            buf + sizeof cmd +
2341                                            cmd.wr_count * cmd.wqe_size +
2342                                            sg_ind * sizeof (struct ib_sge),
2343                                            next->num_sge * sizeof (struct ib_sge))) {
2344                                 ret = -EFAULT;
2345                                 goto out_put;
2346                         }
2347                         sg_ind += next->num_sge;
2348                 } else
2349                         next->sg_list = NULL;
2350         }
2351
2352         resp.bad_wr = 0;
2353         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2354         if (ret)
2355                 for (next = wr; next; next = next->next) {
2356                         ++resp.bad_wr;
2357                         if (next == bad_wr)
2358                                 break;
2359                 }
2360
2361         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2362                 ret = -EFAULT;
2363
2364 out_put:
2365         uobj_put_obj_read(qp);
2366
2367         while (wr) {
2368                 if (is_ud && ud_wr(wr)->ah)
2369                         uobj_put_obj_read(ud_wr(wr)->ah);
2370                 next = wr->next;
2371                 kfree(wr);
2372                 wr = next;
2373         }
2374
2375 out:
2376         kfree(user_wr);
2377
2378         return ret ? ret : in_len;
2379 }
2380
2381 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2382                                                     int in_len,
2383                                                     u32 wr_count,
2384                                                     u32 sge_count,
2385                                                     u32 wqe_size)
2386 {
2387         struct ib_uverbs_recv_wr *user_wr;
2388         struct ib_recv_wr        *wr = NULL, *last, *next;
2389         int                       sg_ind;
2390         int                       i;
2391         int                       ret;
2392
2393         if (in_len < wqe_size * wr_count +
2394             sge_count * sizeof (struct ib_uverbs_sge))
2395                 return ERR_PTR(-EINVAL);
2396
2397         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2398                 return ERR_PTR(-EINVAL);
2399
2400         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2401         if (!user_wr)
2402                 return ERR_PTR(-ENOMEM);
2403
2404         sg_ind = 0;
2405         last = NULL;
2406         for (i = 0; i < wr_count; ++i) {
2407                 if (copy_from_user(user_wr, buf + i * wqe_size,
2408                                    wqe_size)) {
2409                         ret = -EFAULT;
2410                         goto err;
2411                 }
2412
2413                 if (user_wr->num_sge + sg_ind > sge_count) {
2414                         ret = -EINVAL;
2415                         goto err;
2416                 }
2417
2418                 if (user_wr->num_sge >=
2419                     (U32_MAX - ALIGN(sizeof *next, sizeof (struct ib_sge))) /
2420                     sizeof (struct ib_sge)) {
2421                         ret = -EINVAL;
2422                         goto err;
2423                 }
2424
2425                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2426                                user_wr->num_sge * sizeof (struct ib_sge),
2427                                GFP_KERNEL);
2428                 if (!next) {
2429                         ret = -ENOMEM;
2430                         goto err;
2431                 }
2432
2433                 if (!last)
2434                         wr = next;
2435                 else
2436                         last->next = next;
2437                 last = next;
2438
2439                 next->next       = NULL;
2440                 next->wr_id      = user_wr->wr_id;
2441                 next->num_sge    = user_wr->num_sge;
2442
2443                 if (next->num_sge) {
2444                         next->sg_list = (void *) next +
2445                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2446                         if (copy_from_user(next->sg_list,
2447                                            buf + wr_count * wqe_size +
2448                                            sg_ind * sizeof (struct ib_sge),
2449                                            next->num_sge * sizeof (struct ib_sge))) {
2450                                 ret = -EFAULT;
2451                                 goto err;
2452                         }
2453                         sg_ind += next->num_sge;
2454                 } else
2455                         next->sg_list = NULL;
2456         }
2457
2458         kfree(user_wr);
2459         return wr;
2460
2461 err:
2462         kfree(user_wr);
2463
2464         while (wr) {
2465                 next = wr->next;
2466                 kfree(wr);
2467                 wr = next;
2468         }
2469
2470         return ERR_PTR(ret);
2471 }
2472
2473 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2474                             const char __user *buf, int in_len,
2475                             int out_len)
2476 {
2477         struct ib_uverbs_post_recv      cmd;
2478         struct ib_uverbs_post_recv_resp resp;
2479         struct ib_recv_wr              *wr, *next;
2480         const struct ib_recv_wr        *bad_wr;
2481         struct ib_qp                   *qp;
2482         ssize_t                         ret = -EINVAL;
2483
2484         if (copy_from_user(&cmd, buf, sizeof cmd))
2485                 return -EFAULT;
2486
2487         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2488                                        in_len - sizeof cmd, cmd.wr_count,
2489                                        cmd.sge_count, cmd.wqe_size);
2490         if (IS_ERR(wr))
2491                 return PTR_ERR(wr);
2492
2493         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2494         if (!qp)
2495                 goto out;
2496
2497         resp.bad_wr = 0;
2498         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2499
2500         uobj_put_obj_read(qp);
2501         if (ret) {
2502                 for (next = wr; next; next = next->next) {
2503                         ++resp.bad_wr;
2504                         if (next == bad_wr)
2505                                 break;
2506                 }
2507         }
2508
2509         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2510                 ret = -EFAULT;
2511
2512 out:
2513         while (wr) {
2514                 next = wr->next;
2515                 kfree(wr);
2516                 wr = next;
2517         }
2518
2519         return ret ? ret : in_len;
2520 }
2521
2522 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2523                                 const char __user *buf, int in_len,
2524                                 int out_len)
2525 {
2526         struct ib_uverbs_post_srq_recv      cmd;
2527         struct ib_uverbs_post_srq_recv_resp resp;
2528         struct ib_recv_wr                  *wr, *next;
2529         const struct ib_recv_wr            *bad_wr;
2530         struct ib_srq                      *srq;
2531         ssize_t                             ret = -EINVAL;
2532
2533         if (copy_from_user(&cmd, buf, sizeof cmd))
2534                 return -EFAULT;
2535
2536         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2537                                        in_len - sizeof cmd, cmd.wr_count,
2538                                        cmd.sge_count, cmd.wqe_size);
2539         if (IS_ERR(wr))
2540                 return PTR_ERR(wr);
2541
2542         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
2543         if (!srq)
2544                 goto out;
2545
2546         resp.bad_wr = 0;
2547         ret = srq->device->post_srq_recv ?
2548                 srq->device->post_srq_recv(srq, wr, &bad_wr) : -EOPNOTSUPP;
2549
2550         uobj_put_obj_read(srq);
2551
2552         if (ret)
2553                 for (next = wr; next; next = next->next) {
2554                         ++resp.bad_wr;
2555                         if (next == bad_wr)
2556                                 break;
2557                 }
2558
2559         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
2560                 ret = -EFAULT;
2561
2562 out:
2563         while (wr) {
2564                 next = wr->next;
2565                 kfree(wr);
2566                 wr = next;
2567         }
2568
2569         return ret ? ret : in_len;
2570 }
2571
2572 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2573                             const char __user *buf, int in_len,
2574                             int out_len)
2575 {
2576         struct ib_uverbs_create_ah       cmd;
2577         struct ib_uverbs_create_ah_resp  resp;
2578         struct ib_uobject               *uobj;
2579         struct ib_pd                    *pd;
2580         struct ib_ah                    *ah;
2581         struct rdma_ah_attr             attr = {};
2582         int ret;
2583         struct ib_udata                   udata;
2584         struct ib_device *ib_dev;
2585
2586         if (out_len < sizeof resp)
2587                 return -ENOSPC;
2588
2589         if (copy_from_user(&cmd, buf, sizeof cmd))
2590                 return -EFAULT;
2591
2592         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
2593                    u64_to_user_ptr(cmd.response) + sizeof(resp),
2594                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
2595                    out_len - sizeof(resp));
2596
2597         uobj = uobj_alloc(UVERBS_OBJECT_AH, file, &ib_dev);
2598         if (IS_ERR(uobj))
2599                 return PTR_ERR(uobj);
2600
2601         if (!rdma_is_port_valid(ib_dev, cmd.attr.port_num)) {
2602                 ret = -EINVAL;
2603                 goto err;
2604         }
2605
2606         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file);
2607         if (!pd) {
2608                 ret = -EINVAL;
2609                 goto err;
2610         }
2611
2612         attr.type = rdma_ah_find_type(ib_dev, cmd.attr.port_num);
2613         rdma_ah_set_make_grd(&attr, false);
2614         rdma_ah_set_dlid(&attr, cmd.attr.dlid);
2615         rdma_ah_set_sl(&attr, cmd.attr.sl);
2616         rdma_ah_set_path_bits(&attr, cmd.attr.src_path_bits);
2617         rdma_ah_set_static_rate(&attr, cmd.attr.static_rate);
2618         rdma_ah_set_port_num(&attr, cmd.attr.port_num);
2619
2620         if (cmd.attr.is_global) {
2621                 rdma_ah_set_grh(&attr, NULL, cmd.attr.grh.flow_label,
2622                                 cmd.attr.grh.sgid_index,
2623                                 cmd.attr.grh.hop_limit,
2624                                 cmd.attr.grh.traffic_class);
2625                 rdma_ah_set_dgid_raw(&attr, cmd.attr.grh.dgid);
2626         } else {
2627                 rdma_ah_set_ah_flags(&attr, 0);
2628         }
2629
2630         ah = rdma_create_user_ah(pd, &attr, &udata);
2631         if (IS_ERR(ah)) {
2632                 ret = PTR_ERR(ah);
2633                 goto err_put;
2634         }
2635
2636         ah->uobject  = uobj;
2637         uobj->user_handle = cmd.user_handle;
2638         uobj->object = ah;
2639
2640         resp.ah_handle = uobj->id;
2641
2642         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp)) {
2643                 ret = -EFAULT;
2644                 goto err_copy;
2645         }
2646
2647         uobj_put_obj_read(pd);
2648         return uobj_alloc_commit(uobj, in_len);
2649
2650 err_copy:
2651         rdma_destroy_ah(ah);
2652
2653 err_put:
2654         uobj_put_obj_read(pd);
2655
2656 err:
2657         uobj_alloc_abort(uobj);
2658         return ret;
2659 }
2660
2661 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2662                              const char __user *buf, int in_len, int out_len)
2663 {
2664         struct ib_uverbs_destroy_ah cmd;
2665
2666         if (copy_from_user(&cmd, buf, sizeof cmd))
2667                 return -EFAULT;
2668
2669         return uobj_perform_destroy(UVERBS_OBJECT_AH, cmd.ah_handle, file,
2670                                     in_len);
2671 }
2672
2673 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2674                                const char __user *buf, int in_len,
2675                                int out_len)
2676 {
2677         struct ib_uverbs_attach_mcast cmd;
2678         struct ib_qp                 *qp;
2679         struct ib_uqp_object         *obj;
2680         struct ib_uverbs_mcast_entry *mcast;
2681         int                           ret;
2682
2683         if (copy_from_user(&cmd, buf, sizeof cmd))
2684                 return -EFAULT;
2685
2686         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2687         if (!qp)
2688                 return -EINVAL;
2689
2690         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2691
2692         mutex_lock(&obj->mcast_lock);
2693         list_for_each_entry(mcast, &obj->mcast_list, list)
2694                 if (cmd.mlid == mcast->lid &&
2695                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2696                         ret = 0;
2697                         goto out_put;
2698                 }
2699
2700         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2701         if (!mcast) {
2702                 ret = -ENOMEM;
2703                 goto out_put;
2704         }
2705
2706         mcast->lid = cmd.mlid;
2707         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2708
2709         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2710         if (!ret)
2711                 list_add_tail(&mcast->list, &obj->mcast_list);
2712         else
2713                 kfree(mcast);
2714
2715 out_put:
2716         mutex_unlock(&obj->mcast_lock);
2717         uobj_put_obj_read(qp);
2718
2719         return ret ? ret : in_len;
2720 }
2721
2722 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2723                                const char __user *buf, int in_len,
2724                                int out_len)
2725 {
2726         struct ib_uverbs_detach_mcast cmd;
2727         struct ib_uqp_object         *obj;
2728         struct ib_qp                 *qp;
2729         struct ib_uverbs_mcast_entry *mcast;
2730         int                           ret = -EINVAL;
2731         bool                          found = false;
2732
2733         if (copy_from_user(&cmd, buf, sizeof cmd))
2734                 return -EFAULT;
2735
2736         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
2737         if (!qp)
2738                 return -EINVAL;
2739
2740         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2741         mutex_lock(&obj->mcast_lock);
2742
2743         list_for_each_entry(mcast, &obj->mcast_list, list)
2744                 if (cmd.mlid == mcast->lid &&
2745                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2746                         list_del(&mcast->list);
2747                         kfree(mcast);
2748                         found = true;
2749                         break;
2750                 }
2751
2752         if (!found) {
2753                 ret = -EINVAL;
2754                 goto out_put;
2755         }
2756
2757         ret = ib_detach_mcast(qp, (union ib_gid *)cmd.gid, cmd.mlid);
2758
2759 out_put:
2760         mutex_unlock(&obj->mcast_lock);
2761         uobj_put_obj_read(qp);
2762         return ret ? ret : in_len;
2763 }
2764
2765 struct ib_uflow_resources *flow_resources_alloc(size_t num_specs)
2766 {
2767         struct ib_uflow_resources *resources;
2768
2769         resources = kzalloc(sizeof(*resources), GFP_KERNEL);
2770
2771         if (!resources)
2772                 return NULL;
2773
2774         if (!num_specs)
2775                 goto out;
2776
2777         resources->counters =
2778                 kcalloc(num_specs, sizeof(*resources->counters), GFP_KERNEL);
2779         resources->collection =
2780                 kcalloc(num_specs, sizeof(*resources->collection), GFP_KERNEL);
2781
2782         if (!resources->counters || !resources->collection)
2783                 goto err;
2784
2785 out:
2786         resources->max = num_specs;
2787         return resources;
2788
2789 err:
2790         kfree(resources->counters);
2791         kfree(resources);
2792
2793         return NULL;
2794 }
2795 EXPORT_SYMBOL(flow_resources_alloc);
2796
2797 void ib_uverbs_flow_resources_free(struct ib_uflow_resources *uflow_res)
2798 {
2799         unsigned int i;
2800
2801         if (!uflow_res)
2802                 return;
2803
2804         for (i = 0; i < uflow_res->collection_num; i++)
2805                 atomic_dec(&uflow_res->collection[i]->usecnt);
2806
2807         for (i = 0; i < uflow_res->counters_num; i++)
2808                 atomic_dec(&uflow_res->counters[i]->usecnt);
2809
2810         kfree(uflow_res->collection);
2811         kfree(uflow_res->counters);
2812         kfree(uflow_res);
2813 }
2814 EXPORT_SYMBOL(ib_uverbs_flow_resources_free);
2815
2816 void flow_resources_add(struct ib_uflow_resources *uflow_res,
2817                         enum ib_flow_spec_type type,
2818                         void *ibobj)
2819 {
2820         WARN_ON(uflow_res->num >= uflow_res->max);
2821
2822         switch (type) {
2823         case IB_FLOW_SPEC_ACTION_HANDLE:
2824                 atomic_inc(&((struct ib_flow_action *)ibobj)->usecnt);
2825                 uflow_res->collection[uflow_res->collection_num++] =
2826                         (struct ib_flow_action *)ibobj;
2827                 break;
2828         case IB_FLOW_SPEC_ACTION_COUNT:
2829                 atomic_inc(&((struct ib_counters *)ibobj)->usecnt);
2830                 uflow_res->counters[uflow_res->counters_num++] =
2831                         (struct ib_counters *)ibobj;
2832                 break;
2833         default:
2834                 WARN_ON(1);
2835         }
2836
2837         uflow_res->num++;
2838 }
2839 EXPORT_SYMBOL(flow_resources_add);
2840
2841 static int kern_spec_to_ib_spec_action(struct ib_uverbs_file *ufile,
2842                                        struct ib_uverbs_flow_spec *kern_spec,
2843                                        union ib_flow_spec *ib_spec,
2844                                        struct ib_uflow_resources *uflow_res)
2845 {
2846         ib_spec->type = kern_spec->type;
2847         switch (ib_spec->type) {
2848         case IB_FLOW_SPEC_ACTION_TAG:
2849                 if (kern_spec->flow_tag.size !=
2850                     sizeof(struct ib_uverbs_flow_spec_action_tag))
2851                         return -EINVAL;
2852
2853                 ib_spec->flow_tag.size = sizeof(struct ib_flow_spec_action_tag);
2854                 ib_spec->flow_tag.tag_id = kern_spec->flow_tag.tag_id;
2855                 break;
2856         case IB_FLOW_SPEC_ACTION_DROP:
2857                 if (kern_spec->drop.size !=
2858                     sizeof(struct ib_uverbs_flow_spec_action_drop))
2859                         return -EINVAL;
2860
2861                 ib_spec->drop.size = sizeof(struct ib_flow_spec_action_drop);
2862                 break;
2863         case IB_FLOW_SPEC_ACTION_HANDLE:
2864                 if (kern_spec->action.size !=
2865                     sizeof(struct ib_uverbs_flow_spec_action_handle))
2866                         return -EOPNOTSUPP;
2867                 ib_spec->action.act = uobj_get_obj_read(flow_action,
2868                                                         UVERBS_OBJECT_FLOW_ACTION,
2869                                                         kern_spec->action.handle,
2870                                                         ufile);
2871                 if (!ib_spec->action.act)
2872                         return -EINVAL;
2873                 ib_spec->action.size =
2874                         sizeof(struct ib_flow_spec_action_handle);
2875                 flow_resources_add(uflow_res,
2876                                    IB_FLOW_SPEC_ACTION_HANDLE,
2877                                    ib_spec->action.act);
2878                 uobj_put_obj_read(ib_spec->action.act);
2879                 break;
2880         case IB_FLOW_SPEC_ACTION_COUNT:
2881                 if (kern_spec->flow_count.size !=
2882                         sizeof(struct ib_uverbs_flow_spec_action_count))
2883                         return -EINVAL;
2884                 ib_spec->flow_count.counters =
2885                         uobj_get_obj_read(counters,
2886                                           UVERBS_OBJECT_COUNTERS,
2887                                           kern_spec->flow_count.handle,
2888                                           ufile);
2889                 if (!ib_spec->flow_count.counters)
2890                         return -EINVAL;
2891                 ib_spec->flow_count.size =
2892                                 sizeof(struct ib_flow_spec_action_count);
2893                 flow_resources_add(uflow_res,
2894                                    IB_FLOW_SPEC_ACTION_COUNT,
2895                                    ib_spec->flow_count.counters);
2896                 uobj_put_obj_read(ib_spec->flow_count.counters);
2897                 break;
2898         default:
2899                 return -EINVAL;
2900         }
2901         return 0;
2902 }
2903
2904 static size_t kern_spec_filter_sz(const struct ib_uverbs_flow_spec_hdr *spec)
2905 {
2906         /* Returns user space filter size, includes padding */
2907         return (spec->size - sizeof(struct ib_uverbs_flow_spec_hdr)) / 2;
2908 }
2909
2910 static ssize_t spec_filter_size(const void *kern_spec_filter, u16 kern_filter_size,
2911                                 u16 ib_real_filter_sz)
2912 {
2913         /*
2914          * User space filter structures must be 64 bit aligned, otherwise this
2915          * may pass, but we won't handle additional new attributes.
2916          */
2917
2918         if (kern_filter_size > ib_real_filter_sz) {
2919                 if (memchr_inv(kern_spec_filter +
2920                                ib_real_filter_sz, 0,
2921                                kern_filter_size - ib_real_filter_sz))
2922                         return -EINVAL;
2923                 return ib_real_filter_sz;
2924         }
2925         return kern_filter_size;
2926 }
2927
2928 int ib_uverbs_kern_spec_to_ib_spec_filter(enum ib_flow_spec_type type,
2929                                           const void *kern_spec_mask,
2930                                           const void *kern_spec_val,
2931                                           size_t kern_filter_sz,
2932                                           union ib_flow_spec *ib_spec)
2933 {
2934         ssize_t actual_filter_sz;
2935         ssize_t ib_filter_sz;
2936
2937         /* User flow spec size must be aligned to 4 bytes */
2938         if (kern_filter_sz != ALIGN(kern_filter_sz, 4))
2939                 return -EINVAL;
2940
2941         ib_spec->type = type;
2942
2943         if (ib_spec->type == (IB_FLOW_SPEC_INNER | IB_FLOW_SPEC_VXLAN_TUNNEL))
2944                 return -EINVAL;
2945
2946         switch (ib_spec->type & ~IB_FLOW_SPEC_INNER) {
2947         case IB_FLOW_SPEC_ETH:
2948                 ib_filter_sz = offsetof(struct ib_flow_eth_filter, real_sz);
2949                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2950                                                     kern_filter_sz,
2951                                                     ib_filter_sz);
2952                 if (actual_filter_sz <= 0)
2953                         return -EINVAL;
2954                 ib_spec->size = sizeof(struct ib_flow_spec_eth);
2955                 memcpy(&ib_spec->eth.val, kern_spec_val, actual_filter_sz);
2956                 memcpy(&ib_spec->eth.mask, kern_spec_mask, actual_filter_sz);
2957                 break;
2958         case IB_FLOW_SPEC_IPV4:
2959                 ib_filter_sz = offsetof(struct ib_flow_ipv4_filter, real_sz);
2960                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2961                                                     kern_filter_sz,
2962                                                     ib_filter_sz);
2963                 if (actual_filter_sz <= 0)
2964                         return -EINVAL;
2965                 ib_spec->size = sizeof(struct ib_flow_spec_ipv4);
2966                 memcpy(&ib_spec->ipv4.val, kern_spec_val, actual_filter_sz);
2967                 memcpy(&ib_spec->ipv4.mask, kern_spec_mask, actual_filter_sz);
2968                 break;
2969         case IB_FLOW_SPEC_IPV6:
2970                 ib_filter_sz = offsetof(struct ib_flow_ipv6_filter, real_sz);
2971                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2972                                                     kern_filter_sz,
2973                                                     ib_filter_sz);
2974                 if (actual_filter_sz <= 0)
2975                         return -EINVAL;
2976                 ib_spec->size = sizeof(struct ib_flow_spec_ipv6);
2977                 memcpy(&ib_spec->ipv6.val, kern_spec_val, actual_filter_sz);
2978                 memcpy(&ib_spec->ipv6.mask, kern_spec_mask, actual_filter_sz);
2979
2980                 if ((ntohl(ib_spec->ipv6.mask.flow_label)) >= BIT(20) ||
2981                     (ntohl(ib_spec->ipv6.val.flow_label)) >= BIT(20))
2982                         return -EINVAL;
2983                 break;
2984         case IB_FLOW_SPEC_TCP:
2985         case IB_FLOW_SPEC_UDP:
2986                 ib_filter_sz = offsetof(struct ib_flow_tcp_udp_filter, real_sz);
2987                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2988                                                     kern_filter_sz,
2989                                                     ib_filter_sz);
2990                 if (actual_filter_sz <= 0)
2991                         return -EINVAL;
2992                 ib_spec->size = sizeof(struct ib_flow_spec_tcp_udp);
2993                 memcpy(&ib_spec->tcp_udp.val, kern_spec_val, actual_filter_sz);
2994                 memcpy(&ib_spec->tcp_udp.mask, kern_spec_mask, actual_filter_sz);
2995                 break;
2996         case IB_FLOW_SPEC_VXLAN_TUNNEL:
2997                 ib_filter_sz = offsetof(struct ib_flow_tunnel_filter, real_sz);
2998                 actual_filter_sz = spec_filter_size(kern_spec_mask,
2999                                                     kern_filter_sz,
3000                                                     ib_filter_sz);
3001                 if (actual_filter_sz <= 0)
3002                         return -EINVAL;
3003                 ib_spec->tunnel.size = sizeof(struct ib_flow_spec_tunnel);
3004                 memcpy(&ib_spec->tunnel.val, kern_spec_val, actual_filter_sz);
3005                 memcpy(&ib_spec->tunnel.mask, kern_spec_mask, actual_filter_sz);
3006
3007                 if ((ntohl(ib_spec->tunnel.mask.tunnel_id)) >= BIT(24) ||
3008                     (ntohl(ib_spec->tunnel.val.tunnel_id)) >= BIT(24))
3009                         return -EINVAL;
3010                 break;
3011         case IB_FLOW_SPEC_ESP:
3012                 ib_filter_sz = offsetof(struct ib_flow_esp_filter, real_sz);
3013                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3014                                                     kern_filter_sz,
3015                                                     ib_filter_sz);
3016                 if (actual_filter_sz <= 0)
3017                         return -EINVAL;
3018                 ib_spec->esp.size = sizeof(struct ib_flow_spec_esp);
3019                 memcpy(&ib_spec->esp.val, kern_spec_val, actual_filter_sz);
3020                 memcpy(&ib_spec->esp.mask, kern_spec_mask, actual_filter_sz);
3021                 break;
3022         case IB_FLOW_SPEC_GRE:
3023                 ib_filter_sz = offsetof(struct ib_flow_gre_filter, real_sz);
3024                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3025                                                     kern_filter_sz,
3026                                                     ib_filter_sz);
3027                 if (actual_filter_sz <= 0)
3028                         return -EINVAL;
3029                 ib_spec->gre.size = sizeof(struct ib_flow_spec_gre);
3030                 memcpy(&ib_spec->gre.val, kern_spec_val, actual_filter_sz);
3031                 memcpy(&ib_spec->gre.mask, kern_spec_mask, actual_filter_sz);
3032                 break;
3033         case IB_FLOW_SPEC_MPLS:
3034                 ib_filter_sz = offsetof(struct ib_flow_mpls_filter, real_sz);
3035                 actual_filter_sz = spec_filter_size(kern_spec_mask,
3036                                                     kern_filter_sz,
3037                                                     ib_filter_sz);
3038                 if (actual_filter_sz <= 0)
3039                         return -EINVAL;
3040                 ib_spec->mpls.size = sizeof(struct ib_flow_spec_mpls);
3041                 memcpy(&ib_spec->mpls.val, kern_spec_val, actual_filter_sz);
3042                 memcpy(&ib_spec->mpls.mask, kern_spec_mask, actual_filter_sz);
3043                 break;
3044         default:
3045                 return -EINVAL;
3046         }
3047         return 0;
3048 }
3049
3050 static int kern_spec_to_ib_spec_filter(struct ib_uverbs_flow_spec *kern_spec,
3051                                        union ib_flow_spec *ib_spec)
3052 {
3053         ssize_t kern_filter_sz;
3054         void *kern_spec_mask;
3055         void *kern_spec_val;
3056
3057         kern_filter_sz = kern_spec_filter_sz(&kern_spec->hdr);
3058
3059         kern_spec_val = (void *)kern_spec +
3060                 sizeof(struct ib_uverbs_flow_spec_hdr);
3061         kern_spec_mask = kern_spec_val + kern_filter_sz;
3062
3063         return ib_uverbs_kern_spec_to_ib_spec_filter(kern_spec->type,
3064                                                      kern_spec_mask,
3065                                                      kern_spec_val,
3066                                                      kern_filter_sz, ib_spec);
3067 }
3068
3069 static int kern_spec_to_ib_spec(struct ib_uverbs_file *ufile,
3070                                 struct ib_uverbs_flow_spec *kern_spec,
3071                                 union ib_flow_spec *ib_spec,
3072                                 struct ib_uflow_resources *uflow_res)
3073 {
3074         if (kern_spec->reserved)
3075                 return -EINVAL;
3076
3077         if (kern_spec->type >= IB_FLOW_SPEC_ACTION_TAG)
3078                 return kern_spec_to_ib_spec_action(ufile, kern_spec, ib_spec,
3079                                                    uflow_res);
3080         else
3081                 return kern_spec_to_ib_spec_filter(kern_spec, ib_spec);
3082 }
3083
3084 int ib_uverbs_ex_create_wq(struct ib_uverbs_file *file,
3085                            struct ib_udata *ucore,
3086                            struct ib_udata *uhw)
3087 {
3088         struct ib_uverbs_ex_create_wq     cmd = {};
3089         struct ib_uverbs_ex_create_wq_resp resp = {};
3090         struct ib_uwq_object           *obj;
3091         int err = 0;
3092         struct ib_cq *cq;
3093         struct ib_pd *pd;
3094         struct ib_wq *wq;
3095         struct ib_wq_init_attr wq_init_attr = {};
3096         size_t required_cmd_sz;
3097         size_t required_resp_len;
3098         struct ib_device *ib_dev;
3099
3100         required_cmd_sz = offsetof(typeof(cmd), max_sge) + sizeof(cmd.max_sge);
3101         required_resp_len = offsetof(typeof(resp), wqn) + sizeof(resp.wqn);
3102
3103         if (ucore->inlen < required_cmd_sz)
3104                 return -EINVAL;
3105
3106         if (ucore->outlen < required_resp_len)
3107                 return -ENOSPC;
3108
3109         if (ucore->inlen > sizeof(cmd) &&
3110             !ib_is_udata_cleared(ucore, sizeof(cmd),
3111                                  ucore->inlen - sizeof(cmd)))
3112                 return -EOPNOTSUPP;
3113
3114         err = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3115         if (err)
3116                 return err;
3117
3118         if (cmd.comp_mask)
3119                 return -EOPNOTSUPP;
3120
3121         obj = (struct ib_uwq_object *)uobj_alloc(UVERBS_OBJECT_WQ, file,
3122                                                  &ib_dev);
3123         if (IS_ERR(obj))
3124                 return PTR_ERR(obj);
3125
3126         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd.pd_handle, file);
3127         if (!pd) {
3128                 err = -EINVAL;
3129                 goto err_uobj;
3130         }
3131
3132         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
3133         if (!cq) {
3134                 err = -EINVAL;
3135                 goto err_put_pd;
3136         }
3137
3138         wq_init_attr.cq = cq;
3139         wq_init_attr.max_sge = cmd.max_sge;
3140         wq_init_attr.max_wr = cmd.max_wr;
3141         wq_init_attr.wq_context = file;
3142         wq_init_attr.wq_type = cmd.wq_type;
3143         wq_init_attr.event_handler = ib_uverbs_wq_event_handler;
3144         if (ucore->inlen >= (offsetof(typeof(cmd), create_flags) +
3145                              sizeof(cmd.create_flags)))
3146                 wq_init_attr.create_flags = cmd.create_flags;
3147         obj->uevent.events_reported = 0;
3148         INIT_LIST_HEAD(&obj->uevent.event_list);
3149
3150         if (!pd->device->create_wq) {
3151                 err = -EOPNOTSUPP;
3152                 goto err_put_cq;
3153         }
3154         wq = pd->device->create_wq(pd, &wq_init_attr, uhw);
3155         if (IS_ERR(wq)) {
3156                 err = PTR_ERR(wq);
3157                 goto err_put_cq;
3158         }
3159
3160         wq->uobject = &obj->uevent.uobject;
3161         obj->uevent.uobject.object = wq;
3162         wq->wq_type = wq_init_attr.wq_type;
3163         wq->cq = cq;
3164         wq->pd = pd;
3165         wq->device = pd->device;
3166         wq->wq_context = wq_init_attr.wq_context;
3167         atomic_set(&wq->usecnt, 0);
3168         atomic_inc(&pd->usecnt);
3169         atomic_inc(&cq->usecnt);
3170         wq->uobject = &obj->uevent.uobject;
3171         obj->uevent.uobject.object = wq;
3172
3173         memset(&resp, 0, sizeof(resp));
3174         resp.wq_handle = obj->uevent.uobject.id;
3175         resp.max_sge = wq_init_attr.max_sge;
3176         resp.max_wr = wq_init_attr.max_wr;
3177         resp.wqn = wq->wq_num;
3178         resp.response_length = required_resp_len;
3179         err = ib_copy_to_udata(ucore,
3180                                &resp, resp.response_length);
3181         if (err)
3182                 goto err_copy;
3183
3184         uobj_put_obj_read(pd);
3185         uobj_put_obj_read(cq);
3186         return uobj_alloc_commit(&obj->uevent.uobject, 0);
3187
3188 err_copy:
3189         ib_destroy_wq(wq);
3190 err_put_cq:
3191         uobj_put_obj_read(cq);
3192 err_put_pd:
3193         uobj_put_obj_read(pd);
3194 err_uobj:
3195         uobj_alloc_abort(&obj->uevent.uobject);
3196
3197         return err;
3198 }
3199
3200 int ib_uverbs_ex_destroy_wq(struct ib_uverbs_file *file,
3201                             struct ib_udata *ucore,
3202                             struct ib_udata *uhw)
3203 {
3204         struct ib_uverbs_ex_destroy_wq  cmd = {};
3205         struct ib_uverbs_ex_destroy_wq_resp     resp = {};
3206         struct ib_uobject               *uobj;
3207         struct ib_uwq_object            *obj;
3208         size_t required_cmd_sz;
3209         size_t required_resp_len;
3210         int                             ret;
3211
3212         required_cmd_sz = offsetof(typeof(cmd), wq_handle) + sizeof(cmd.wq_handle);
3213         required_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
3214
3215         if (ucore->inlen < required_cmd_sz)
3216                 return -EINVAL;
3217
3218         if (ucore->outlen < required_resp_len)
3219                 return -ENOSPC;
3220
3221         if (ucore->inlen > sizeof(cmd) &&
3222             !ib_is_udata_cleared(ucore, sizeof(cmd),
3223                                  ucore->inlen - sizeof(cmd)))
3224                 return -EOPNOTSUPP;
3225
3226         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3227         if (ret)
3228                 return ret;
3229
3230         if (cmd.comp_mask)
3231                 return -EOPNOTSUPP;
3232
3233         resp.response_length = required_resp_len;
3234         uobj = uobj_get_destroy(UVERBS_OBJECT_WQ, cmd.wq_handle, file);
3235         if (IS_ERR(uobj))
3236                 return PTR_ERR(uobj);
3237
3238         obj = container_of(uobj, struct ib_uwq_object, uevent.uobject);
3239         resp.events_reported = obj->uevent.events_reported;
3240
3241         uobj_put_destroy(uobj);
3242
3243         return ib_copy_to_udata(ucore, &resp, resp.response_length);
3244 }
3245
3246 int ib_uverbs_ex_modify_wq(struct ib_uverbs_file *file,
3247                            struct ib_udata *ucore,
3248                            struct ib_udata *uhw)
3249 {
3250         struct ib_uverbs_ex_modify_wq cmd = {};
3251         struct ib_wq *wq;
3252         struct ib_wq_attr wq_attr = {};
3253         size_t required_cmd_sz;
3254         int ret;
3255
3256         required_cmd_sz = offsetof(typeof(cmd), curr_wq_state) + sizeof(cmd.curr_wq_state);
3257         if (ucore->inlen < required_cmd_sz)
3258                 return -EINVAL;
3259
3260         if (ucore->inlen > sizeof(cmd) &&
3261             !ib_is_udata_cleared(ucore, sizeof(cmd),
3262                                  ucore->inlen - sizeof(cmd)))
3263                 return -EOPNOTSUPP;
3264
3265         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3266         if (ret)
3267                 return ret;
3268
3269         if (!cmd.attr_mask)
3270                 return -EINVAL;
3271
3272         if (cmd.attr_mask > (IB_WQ_STATE | IB_WQ_CUR_STATE | IB_WQ_FLAGS))
3273                 return -EINVAL;
3274
3275         wq = uobj_get_obj_read(wq, UVERBS_OBJECT_WQ, cmd.wq_handle, file);
3276         if (!wq)
3277                 return -EINVAL;
3278
3279         wq_attr.curr_wq_state = cmd.curr_wq_state;
3280         wq_attr.wq_state = cmd.wq_state;
3281         if (cmd.attr_mask & IB_WQ_FLAGS) {
3282                 wq_attr.flags = cmd.flags;
3283                 wq_attr.flags_mask = cmd.flags_mask;
3284         }
3285         if (!wq->device->modify_wq) {
3286                 ret = -EOPNOTSUPP;
3287                 goto out;
3288         }
3289         ret = wq->device->modify_wq(wq, &wq_attr, cmd.attr_mask, uhw);
3290 out:
3291         uobj_put_obj_read(wq);
3292         return ret;
3293 }
3294
3295 int ib_uverbs_ex_create_rwq_ind_table(struct ib_uverbs_file *file,
3296                                       struct ib_udata *ucore,
3297                                       struct ib_udata *uhw)
3298 {
3299         struct ib_uverbs_ex_create_rwq_ind_table          cmd = {};
3300         struct ib_uverbs_ex_create_rwq_ind_table_resp  resp = {};
3301         struct ib_uobject                 *uobj;
3302         int err = 0;
3303         struct ib_rwq_ind_table_init_attr init_attr = {};
3304         struct ib_rwq_ind_table *rwq_ind_tbl;
3305         struct ib_wq    **wqs = NULL;
3306         u32 *wqs_handles = NULL;
3307         struct ib_wq    *wq = NULL;
3308         int i, j, num_read_wqs;
3309         u32 num_wq_handles;
3310         u32 expected_in_size;
3311         size_t required_cmd_sz_header;
3312         size_t required_resp_len;
3313         struct ib_device *ib_dev;
3314
3315         required_cmd_sz_header = offsetof(typeof(cmd), log_ind_tbl_size) + sizeof(cmd.log_ind_tbl_size);
3316         required_resp_len = offsetof(typeof(resp), ind_tbl_num) + sizeof(resp.ind_tbl_num);
3317
3318         if (ucore->inlen < required_cmd_sz_header)
3319                 return -EINVAL;
3320
3321         if (ucore->outlen < required_resp_len)
3322                 return -ENOSPC;
3323
3324         err = ib_copy_from_udata(&cmd, ucore, required_cmd_sz_header);
3325         if (err)
3326                 return err;
3327
3328         ucore->inbuf += required_cmd_sz_header;
3329         ucore->inlen -= required_cmd_sz_header;
3330
3331         if (cmd.comp_mask)
3332                 return -EOPNOTSUPP;
3333
3334         if (cmd.log_ind_tbl_size > IB_USER_VERBS_MAX_LOG_IND_TBL_SIZE)
3335                 return -EINVAL;
3336
3337         num_wq_handles = 1 << cmd.log_ind_tbl_size;
3338         expected_in_size = num_wq_handles * sizeof(__u32);
3339         if (num_wq_handles == 1)
3340                 /* input size for wq handles is u64 aligned */
3341                 expected_in_size += sizeof(__u32);
3342
3343         if (ucore->inlen < expected_in_size)
3344                 return -EINVAL;
3345
3346         if (ucore->inlen > expected_in_size &&
3347             !ib_is_udata_cleared(ucore, expected_in_size,
3348                                  ucore->inlen - expected_in_size))
3349                 return -EOPNOTSUPP;
3350
3351         wqs_handles = kcalloc(num_wq_handles, sizeof(*wqs_handles),
3352                               GFP_KERNEL);
3353         if (!wqs_handles)
3354                 return -ENOMEM;
3355
3356         err = ib_copy_from_udata(wqs_handles, ucore,
3357                                  num_wq_handles * sizeof(__u32));
3358         if (err)
3359                 goto err_free;
3360
3361         wqs = kcalloc(num_wq_handles, sizeof(*wqs), GFP_KERNEL);
3362         if (!wqs) {
3363                 err = -ENOMEM;
3364                 goto  err_free;
3365         }
3366
3367         for (num_read_wqs = 0; num_read_wqs < num_wq_handles;
3368                         num_read_wqs++) {
3369                 wq = uobj_get_obj_read(wq, UVERBS_OBJECT_WQ,
3370                                        wqs_handles[num_read_wqs], file);
3371                 if (!wq) {
3372                         err = -EINVAL;
3373                         goto put_wqs;
3374                 }
3375
3376                 wqs[num_read_wqs] = wq;
3377         }
3378
3379         uobj = uobj_alloc(UVERBS_OBJECT_RWQ_IND_TBL, file, &ib_dev);
3380         if (IS_ERR(uobj)) {
3381                 err = PTR_ERR(uobj);
3382                 goto put_wqs;
3383         }
3384
3385         init_attr.log_ind_tbl_size = cmd.log_ind_tbl_size;
3386         init_attr.ind_tbl = wqs;
3387
3388         if (!ib_dev->create_rwq_ind_table) {
3389                 err = -EOPNOTSUPP;
3390                 goto err_uobj;
3391         }
3392         rwq_ind_tbl = ib_dev->create_rwq_ind_table(ib_dev, &init_attr, uhw);
3393
3394         if (IS_ERR(rwq_ind_tbl)) {
3395                 err = PTR_ERR(rwq_ind_tbl);
3396                 goto err_uobj;
3397         }
3398
3399         rwq_ind_tbl->ind_tbl = wqs;
3400         rwq_ind_tbl->log_ind_tbl_size = init_attr.log_ind_tbl_size;
3401         rwq_ind_tbl->uobject = uobj;
3402         uobj->object = rwq_ind_tbl;
3403         rwq_ind_tbl->device = ib_dev;
3404         atomic_set(&rwq_ind_tbl->usecnt, 0);
3405
3406         for (i = 0; i < num_wq_handles; i++)
3407                 atomic_inc(&wqs[i]->usecnt);
3408
3409         resp.ind_tbl_handle = uobj->id;
3410         resp.ind_tbl_num = rwq_ind_tbl->ind_tbl_num;
3411         resp.response_length = required_resp_len;
3412
3413         err = ib_copy_to_udata(ucore,
3414                                &resp, resp.response_length);
3415         if (err)
3416                 goto err_copy;
3417
3418         kfree(wqs_handles);
3419
3420         for (j = 0; j < num_read_wqs; j++)
3421                 uobj_put_obj_read(wqs[j]);
3422
3423         return uobj_alloc_commit(uobj, 0);
3424
3425 err_copy:
3426         ib_destroy_rwq_ind_table(rwq_ind_tbl);
3427 err_uobj:
3428         uobj_alloc_abort(uobj);
3429 put_wqs:
3430         for (j = 0; j < num_read_wqs; j++)
3431                 uobj_put_obj_read(wqs[j]);
3432 err_free:
3433         kfree(wqs_handles);
3434         kfree(wqs);
3435         return err;
3436 }
3437
3438 int ib_uverbs_ex_destroy_rwq_ind_table(struct ib_uverbs_file *file,
3439                                        struct ib_udata *ucore,
3440                                        struct ib_udata *uhw)
3441 {
3442         struct ib_uverbs_ex_destroy_rwq_ind_table       cmd = {};
3443         int                     ret;
3444         size_t required_cmd_sz;
3445
3446         required_cmd_sz = offsetof(typeof(cmd), ind_tbl_handle) + sizeof(cmd.ind_tbl_handle);
3447
3448         if (ucore->inlen < required_cmd_sz)
3449                 return -EINVAL;
3450
3451         if (ucore->inlen > sizeof(cmd) &&
3452             !ib_is_udata_cleared(ucore, sizeof(cmd),
3453                                  ucore->inlen - sizeof(cmd)))
3454                 return -EOPNOTSUPP;
3455
3456         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
3457         if (ret)
3458                 return ret;
3459
3460         if (cmd.comp_mask)
3461                 return -EOPNOTSUPP;
3462
3463         return uobj_perform_destroy(UVERBS_OBJECT_RWQ_IND_TBL,
3464                                     cmd.ind_tbl_handle, file, 0);
3465 }
3466
3467 int ib_uverbs_ex_create_flow(struct ib_uverbs_file *file,
3468                              struct ib_udata *ucore,
3469                              struct ib_udata *uhw)
3470 {
3471         struct ib_uverbs_create_flow      cmd;
3472         struct ib_uverbs_create_flow_resp resp;
3473         struct ib_uobject                 *uobj;
3474         struct ib_flow                    *flow_id;
3475         struct ib_uverbs_flow_attr        *kern_flow_attr;
3476         struct ib_flow_attr               *flow_attr;
3477         struct ib_qp                      *qp;
3478         struct ib_uflow_resources         *uflow_res;
3479         struct ib_uverbs_flow_spec_hdr    *kern_spec;
3480         int err = 0;
3481         void *ib_spec;
3482         int i;
3483         struct ib_device *ib_dev;
3484
3485         if (ucore->inlen < sizeof(cmd))
3486                 return -EINVAL;
3487
3488         if (ucore->outlen < sizeof(resp))
3489                 return -ENOSPC;
3490
3491         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3492         if (err)
3493                 return err;
3494
3495         ucore->inbuf += sizeof(cmd);
3496         ucore->inlen -= sizeof(cmd);
3497
3498         if (cmd.comp_mask)
3499                 return -EINVAL;
3500
3501         if (!capable(CAP_NET_RAW))
3502                 return -EPERM;
3503
3504         if (cmd.flow_attr.flags >= IB_FLOW_ATTR_FLAGS_RESERVED)
3505                 return -EINVAL;
3506
3507         if ((cmd.flow_attr.flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) &&
3508             ((cmd.flow_attr.type == IB_FLOW_ATTR_ALL_DEFAULT) ||
3509              (cmd.flow_attr.type == IB_FLOW_ATTR_MC_DEFAULT)))
3510                 return -EINVAL;
3511
3512         if (cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
3513                 return -EINVAL;
3514
3515         if (cmd.flow_attr.size > ucore->inlen ||
3516             cmd.flow_attr.size >
3517             (cmd.flow_attr.num_of_specs * sizeof(struct ib_uverbs_flow_spec)))
3518                 return -EINVAL;
3519
3520         if (cmd.flow_attr.reserved[0] ||
3521             cmd.flow_attr.reserved[1])
3522                 return -EINVAL;
3523
3524         if (cmd.flow_attr.num_of_specs) {
3525                 kern_flow_attr = kmalloc(sizeof(*kern_flow_attr) + cmd.flow_attr.size,
3526                                          GFP_KERNEL);
3527                 if (!kern_flow_attr)
3528                         return -ENOMEM;
3529
3530                 *kern_flow_attr = cmd.flow_attr;
3531                 err = ib_copy_from_udata(&kern_flow_attr->flow_specs, ucore,
3532                                          cmd.flow_attr.size);
3533                 if (err)
3534                         goto err_free_attr;
3535         } else {
3536                 kern_flow_attr = &cmd.flow_attr;
3537         }
3538
3539         uobj = uobj_alloc(UVERBS_OBJECT_FLOW, file, &ib_dev);
3540         if (IS_ERR(uobj)) {
3541                 err = PTR_ERR(uobj);
3542                 goto err_free_attr;
3543         }
3544
3545         qp = uobj_get_obj_read(qp, UVERBS_OBJECT_QP, cmd.qp_handle, file);
3546         if (!qp) {
3547                 err = -EINVAL;
3548                 goto err_uobj;
3549         }
3550
3551         if (qp->qp_type != IB_QPT_UD && qp->qp_type != IB_QPT_RAW_PACKET) {
3552                 err = -EINVAL;
3553                 goto err_put;
3554         }
3555
3556         if (!qp->device->create_flow) {
3557                 err = -EOPNOTSUPP;
3558                 goto err_put;
3559         }
3560
3561         flow_attr = kzalloc(struct_size(flow_attr, flows,
3562                                 cmd.flow_attr.num_of_specs), GFP_KERNEL);
3563         if (!flow_attr) {
3564                 err = -ENOMEM;
3565                 goto err_put;
3566         }
3567         uflow_res = flow_resources_alloc(cmd.flow_attr.num_of_specs);
3568         if (!uflow_res) {
3569                 err = -ENOMEM;
3570                 goto err_free_flow_attr;
3571         }
3572
3573         flow_attr->type = kern_flow_attr->type;
3574         flow_attr->priority = kern_flow_attr->priority;
3575         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
3576         flow_attr->port = kern_flow_attr->port;
3577         flow_attr->flags = kern_flow_attr->flags;
3578         flow_attr->size = sizeof(*flow_attr);
3579
3580         kern_spec = kern_flow_attr->flow_specs;
3581         ib_spec = flow_attr + 1;
3582         for (i = 0; i < flow_attr->num_of_specs &&
3583                         cmd.flow_attr.size >= sizeof(*kern_spec) &&
3584                         cmd.flow_attr.size >= kern_spec->size;
3585              i++) {
3586                 err = kern_spec_to_ib_spec(
3587                                 file, (struct ib_uverbs_flow_spec *)kern_spec,
3588                                 ib_spec, uflow_res);
3589                 if (err)
3590                         goto err_free;
3591
3592                 flow_attr->size +=
3593                         ((union ib_flow_spec *) ib_spec)->size;
3594                 cmd.flow_attr.size -= kern_spec->size;
3595                 kern_spec = ((void *)kern_spec) + kern_spec->size;
3596                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
3597         }
3598         if (cmd.flow_attr.size || (i != flow_attr->num_of_specs)) {
3599                 pr_warn("create flow failed, flow %d: %d bytes left from uverb cmd\n",
3600                         i, cmd.flow_attr.size);
3601                 err = -EINVAL;
3602                 goto err_free;
3603         }
3604
3605         flow_id = qp->device->create_flow(qp, flow_attr,
3606                                           IB_FLOW_DOMAIN_USER, uhw);
3607
3608         if (IS_ERR(flow_id)) {
3609                 err = PTR_ERR(flow_id);
3610                 goto err_free;
3611         }
3612
3613         ib_set_flow(uobj, flow_id, qp, qp->device, uflow_res);
3614
3615         memset(&resp, 0, sizeof(resp));
3616         resp.flow_handle = uobj->id;
3617
3618         err = ib_copy_to_udata(ucore,
3619                                &resp, sizeof(resp));
3620         if (err)
3621                 goto err_copy;
3622
3623         uobj_put_obj_read(qp);
3624         kfree(flow_attr);
3625         if (cmd.flow_attr.num_of_specs)
3626                 kfree(kern_flow_attr);
3627         return uobj_alloc_commit(uobj, 0);
3628 err_copy:
3629         if (!qp->device->destroy_flow(flow_id))
3630                 atomic_dec(&qp->usecnt);
3631 err_free:
3632         ib_uverbs_flow_resources_free(uflow_res);
3633 err_free_flow_attr:
3634         kfree(flow_attr);
3635 err_put:
3636         uobj_put_obj_read(qp);
3637 err_uobj:
3638         uobj_alloc_abort(uobj);
3639 err_free_attr:
3640         if (cmd.flow_attr.num_of_specs)
3641                 kfree(kern_flow_attr);
3642         return err;
3643 }
3644
3645 int ib_uverbs_ex_destroy_flow(struct ib_uverbs_file *file,
3646                               struct ib_udata *ucore,
3647                               struct ib_udata *uhw)
3648 {
3649         struct ib_uverbs_destroy_flow   cmd;
3650         int                             ret;
3651
3652         if (ucore->inlen < sizeof(cmd))
3653                 return -EINVAL;
3654
3655         ret = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3656         if (ret)
3657                 return ret;
3658
3659         if (cmd.comp_mask)
3660                 return -EINVAL;
3661
3662         return uobj_perform_destroy(UVERBS_OBJECT_FLOW, cmd.flow_handle, file,
3663                                     0);
3664 }
3665
3666 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
3667                                 struct ib_uverbs_create_xsrq *cmd,
3668                                 struct ib_udata *udata)
3669 {
3670         struct ib_uverbs_create_srq_resp resp;
3671         struct ib_usrq_object           *obj;
3672         struct ib_pd                    *pd;
3673         struct ib_srq                   *srq;
3674         struct ib_uobject               *uninitialized_var(xrcd_uobj);
3675         struct ib_srq_init_attr          attr;
3676         int ret;
3677         struct ib_device *ib_dev;
3678
3679         obj = (struct ib_usrq_object *)uobj_alloc(UVERBS_OBJECT_SRQ, file,
3680                                                   &ib_dev);
3681         if (IS_ERR(obj))
3682                 return PTR_ERR(obj);
3683
3684         if (cmd->srq_type == IB_SRQT_TM)
3685                 attr.ext.tag_matching.max_num_tags = cmd->max_num_tags;
3686
3687         if (cmd->srq_type == IB_SRQT_XRC) {
3688                 xrcd_uobj = uobj_get_read(UVERBS_OBJECT_XRCD, cmd->xrcd_handle,
3689                                           file);
3690                 if (IS_ERR(xrcd_uobj)) {
3691                         ret = -EINVAL;
3692                         goto err;
3693                 }
3694
3695                 attr.ext.xrc.xrcd = (struct ib_xrcd *)xrcd_uobj->object;
3696                 if (!attr.ext.xrc.xrcd) {
3697                         ret = -EINVAL;
3698                         goto err_put_xrcd;
3699                 }
3700
3701                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
3702                 atomic_inc(&obj->uxrcd->refcnt);
3703         }
3704
3705         if (ib_srq_has_cq(cmd->srq_type)) {
3706                 attr.ext.cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ,
3707                                                 cmd->cq_handle, file);
3708                 if (!attr.ext.cq) {
3709                         ret = -EINVAL;
3710                         goto err_put_xrcd;
3711                 }
3712         }
3713
3714         pd = uobj_get_obj_read(pd, UVERBS_OBJECT_PD, cmd->pd_handle, file);
3715         if (!pd) {
3716                 ret = -EINVAL;
3717                 goto err_put_cq;
3718         }
3719
3720         attr.event_handler  = ib_uverbs_srq_event_handler;
3721         attr.srq_context    = file;
3722         attr.srq_type       = cmd->srq_type;
3723         attr.attr.max_wr    = cmd->max_wr;
3724         attr.attr.max_sge   = cmd->max_sge;
3725         attr.attr.srq_limit = cmd->srq_limit;
3726
3727         obj->uevent.events_reported = 0;
3728         INIT_LIST_HEAD(&obj->uevent.event_list);
3729
3730         srq = pd->device->create_srq(pd, &attr, udata);
3731         if (IS_ERR(srq)) {
3732                 ret = PTR_ERR(srq);
3733                 goto err_put;
3734         }
3735
3736         srq->device        = pd->device;
3737         srq->pd            = pd;
3738         srq->srq_type      = cmd->srq_type;
3739         srq->uobject       = &obj->uevent.uobject;
3740         srq->event_handler = attr.event_handler;
3741         srq->srq_context   = attr.srq_context;
3742
3743         if (ib_srq_has_cq(cmd->srq_type)) {
3744                 srq->ext.cq       = attr.ext.cq;
3745                 atomic_inc(&attr.ext.cq->usecnt);
3746         }
3747
3748         if (cmd->srq_type == IB_SRQT_XRC) {
3749                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
3750                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
3751         }
3752
3753         atomic_inc(&pd->usecnt);
3754         atomic_set(&srq->usecnt, 0);
3755
3756         obj->uevent.uobject.object = srq;
3757         obj->uevent.uobject.user_handle = cmd->user_handle;
3758
3759         memset(&resp, 0, sizeof resp);
3760         resp.srq_handle = obj->uevent.uobject.id;
3761         resp.max_wr     = attr.attr.max_wr;
3762         resp.max_sge    = attr.attr.max_sge;
3763         if (cmd->srq_type == IB_SRQT_XRC)
3764                 resp.srqn = srq->ext.xrc.srq_num;
3765
3766         if (copy_to_user(u64_to_user_ptr(cmd->response),
3767                          &resp, sizeof resp)) {
3768                 ret = -EFAULT;
3769                 goto err_copy;
3770         }
3771
3772         if (cmd->srq_type == IB_SRQT_XRC)
3773                 uobj_put_read(xrcd_uobj);
3774
3775         if (ib_srq_has_cq(cmd->srq_type))
3776                 uobj_put_obj_read(attr.ext.cq);
3777
3778         uobj_put_obj_read(pd);
3779         return uobj_alloc_commit(&obj->uevent.uobject, 0);
3780
3781 err_copy:
3782         ib_destroy_srq(srq);
3783
3784 err_put:
3785         uobj_put_obj_read(pd);
3786
3787 err_put_cq:
3788         if (ib_srq_has_cq(cmd->srq_type))
3789                 uobj_put_obj_read(attr.ext.cq);
3790
3791 err_put_xrcd:
3792         if (cmd->srq_type == IB_SRQT_XRC) {
3793                 atomic_dec(&obj->uxrcd->refcnt);
3794                 uobj_put_read(xrcd_uobj);
3795         }
3796
3797 err:
3798         uobj_alloc_abort(&obj->uevent.uobject);
3799         return ret;
3800 }
3801
3802 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
3803                              const char __user *buf, int in_len,
3804                              int out_len)
3805 {
3806         struct ib_uverbs_create_srq      cmd;
3807         struct ib_uverbs_create_xsrq     xcmd;
3808         struct ib_uverbs_create_srq_resp resp;
3809         struct ib_udata                  udata;
3810         int ret;
3811
3812         if (out_len < sizeof resp)
3813                 return -ENOSPC;
3814
3815         if (copy_from_user(&cmd, buf, sizeof cmd))
3816                 return -EFAULT;
3817
3818         memset(&xcmd, 0, sizeof(xcmd));
3819         xcmd.response    = cmd.response;
3820         xcmd.user_handle = cmd.user_handle;
3821         xcmd.srq_type    = IB_SRQT_BASIC;
3822         xcmd.pd_handle   = cmd.pd_handle;
3823         xcmd.max_wr      = cmd.max_wr;
3824         xcmd.max_sge     = cmd.max_sge;
3825         xcmd.srq_limit   = cmd.srq_limit;
3826
3827         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
3828                    u64_to_user_ptr(cmd.response) + sizeof(resp),
3829                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
3830                    out_len - sizeof(resp));
3831
3832         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
3833         if (ret)
3834                 return ret;
3835
3836         return in_len;
3837 }
3838
3839 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
3840                               const char __user *buf, int in_len, int out_len)
3841 {
3842         struct ib_uverbs_create_xsrq     cmd;
3843         struct ib_uverbs_create_srq_resp resp;
3844         struct ib_udata                  udata;
3845         int ret;
3846
3847         if (out_len < sizeof resp)
3848                 return -ENOSPC;
3849
3850         if (copy_from_user(&cmd, buf, sizeof cmd))
3851                 return -EFAULT;
3852
3853         ib_uverbs_init_udata(&udata, buf + sizeof(cmd),
3854                    u64_to_user_ptr(cmd.response) + sizeof(resp),
3855                    in_len - sizeof(cmd) - sizeof(struct ib_uverbs_cmd_hdr),
3856                    out_len - sizeof(resp));
3857
3858         ret = __uverbs_create_xsrq(file, &cmd, &udata);
3859         if (ret)
3860                 return ret;
3861
3862         return in_len;
3863 }
3864
3865 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
3866                              const char __user *buf, int in_len,
3867                              int out_len)
3868 {
3869         struct ib_uverbs_modify_srq cmd;
3870         struct ib_udata             udata;
3871         struct ib_srq              *srq;
3872         struct ib_srq_attr          attr;
3873         int                         ret;
3874
3875         if (copy_from_user(&cmd, buf, sizeof cmd))
3876                 return -EFAULT;
3877
3878         ib_uverbs_init_udata(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
3879                    out_len);
3880
3881         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
3882         if (!srq)
3883                 return -EINVAL;
3884
3885         attr.max_wr    = cmd.max_wr;
3886         attr.srq_limit = cmd.srq_limit;
3887
3888         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
3889
3890         uobj_put_obj_read(srq);
3891
3892         return ret ? ret : in_len;
3893 }
3894
3895 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
3896                             const char __user *buf,
3897                             int in_len, int out_len)
3898 {
3899         struct ib_uverbs_query_srq      cmd;
3900         struct ib_uverbs_query_srq_resp resp;
3901         struct ib_srq_attr              attr;
3902         struct ib_srq                   *srq;
3903         int                             ret;
3904
3905         if (out_len < sizeof resp)
3906                 return -ENOSPC;
3907
3908         if (copy_from_user(&cmd, buf, sizeof cmd))
3909                 return -EFAULT;
3910
3911         srq = uobj_get_obj_read(srq, UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
3912         if (!srq)
3913                 return -EINVAL;
3914
3915         ret = ib_query_srq(srq, &attr);
3916
3917         uobj_put_obj_read(srq);
3918
3919         if (ret)
3920                 return ret;
3921
3922         memset(&resp, 0, sizeof resp);
3923
3924         resp.max_wr    = attr.max_wr;
3925         resp.max_sge   = attr.max_sge;
3926         resp.srq_limit = attr.srq_limit;
3927
3928         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof resp))
3929                 return -EFAULT;
3930
3931         return in_len;
3932 }
3933
3934 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
3935                               const char __user *buf, int in_len,
3936                               int out_len)
3937 {
3938         struct ib_uverbs_destroy_srq      cmd;
3939         struct ib_uverbs_destroy_srq_resp resp;
3940         struct ib_uobject                *uobj;
3941         struct ib_uevent_object          *obj;
3942
3943         if (copy_from_user(&cmd, buf, sizeof cmd))
3944                 return -EFAULT;
3945
3946         uobj = uobj_get_destroy(UVERBS_OBJECT_SRQ, cmd.srq_handle, file);
3947         if (IS_ERR(uobj))
3948                 return PTR_ERR(uobj);
3949
3950         obj = container_of(uobj, struct ib_uevent_object, uobject);
3951         memset(&resp, 0, sizeof(resp));
3952         resp.events_reported = obj->events_reported;
3953
3954         uobj_put_destroy(uobj);
3955
3956         if (copy_to_user(u64_to_user_ptr(cmd.response), &resp, sizeof(resp)))
3957                 return -EFAULT;
3958
3959         return in_len;
3960 }
3961
3962 int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
3963                               struct ib_udata *ucore,
3964                               struct ib_udata *uhw)
3965 {
3966         struct ib_uverbs_ex_query_device_resp resp = { {0} };
3967         struct ib_uverbs_ex_query_device  cmd;
3968         struct ib_device_attr attr = {0};
3969         struct ib_ucontext *ucontext;
3970         struct ib_device *ib_dev;
3971         int err;
3972
3973         ucontext = ib_uverbs_get_ucontext(file);
3974         if (IS_ERR(ucontext))
3975                 return PTR_ERR(ucontext);
3976         ib_dev = ucontext->device;
3977
3978         if (!ib_dev->query_device)
3979                 return -EOPNOTSUPP;
3980
3981         if (ucore->inlen < sizeof(cmd))
3982                 return -EINVAL;
3983
3984         err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
3985         if (err)
3986                 return err;
3987
3988         if (cmd.comp_mask)
3989                 return -EINVAL;
3990
3991         if (cmd.reserved)
3992                 return -EINVAL;
3993
3994         resp.response_length = offsetof(typeof(resp), odp_caps);
3995
3996         if (ucore->outlen < resp.response_length)
3997                 return -ENOSPC;
3998
3999         err = ib_dev->query_device(ib_dev, &attr, uhw);
4000         if (err)
4001                 return err;
4002
4003         copy_query_dev_fields(ucontext, &resp.base, &attr);
4004
4005         if (ucore->outlen < resp.response_length + sizeof(resp.odp_caps))
4006                 goto end;
4007
4008 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4009         resp.odp_caps.general_caps = attr.odp_caps.general_caps;
4010         resp.odp_caps.per_transport_caps.rc_odp_caps =
4011                 attr.odp_caps.per_transport_caps.rc_odp_caps;
4012         resp.odp_caps.per_transport_caps.uc_odp_caps =
4013                 attr.odp_caps.per_transport_caps.uc_odp_caps;
4014         resp.odp_caps.per_transport_caps.ud_odp_caps =
4015                 attr.odp_caps.per_transport_caps.ud_odp_caps;
4016 #endif
4017         resp.response_length += sizeof(resp.odp_caps);
4018
4019         if (ucore->outlen < resp.response_length + sizeof(resp.timestamp_mask))
4020                 goto end;
4021
4022         resp.timestamp_mask = attr.timestamp_mask;
4023         resp.response_length += sizeof(resp.timestamp_mask);
4024
4025         if (ucore->outlen < resp.response_length + sizeof(resp.hca_core_clock))
4026                 goto end;
4027
4028         resp.hca_core_clock = attr.hca_core_clock;
4029         resp.response_length += sizeof(resp.hca_core_clock);
4030
4031         if (ucore->outlen < resp.response_length + sizeof(resp.device_cap_flags_ex))
4032                 goto end;
4033
4034         resp.device_cap_flags_ex = attr.device_cap_flags;
4035         resp.response_length += sizeof(resp.device_cap_flags_ex);
4036
4037         if (ucore->outlen < resp.response_length + sizeof(resp.rss_caps))
4038                 goto end;
4039
4040         resp.rss_caps.supported_qpts = attr.rss_caps.supported_qpts;
4041         resp.rss_caps.max_rwq_indirection_tables =
4042                 attr.rss_caps.max_rwq_indirection_tables;
4043         resp.rss_caps.max_rwq_indirection_table_size =
4044                 attr.rss_caps.max_rwq_indirection_table_size;
4045
4046         resp.response_length += sizeof(resp.rss_caps);
4047
4048         if (ucore->outlen < resp.response_length + sizeof(resp.max_wq_type_rq))
4049                 goto end;
4050
4051         resp.max_wq_type_rq = attr.max_wq_type_rq;
4052         resp.response_length += sizeof(resp.max_wq_type_rq);
4053
4054         if (ucore->outlen < resp.response_length + sizeof(resp.raw_packet_caps))
4055                 goto end;
4056
4057         resp.raw_packet_caps = attr.raw_packet_caps;
4058         resp.response_length += sizeof(resp.raw_packet_caps);
4059
4060         if (ucore->outlen < resp.response_length + sizeof(resp.tm_caps))
4061                 goto end;
4062
4063         resp.tm_caps.max_rndv_hdr_size  = attr.tm_caps.max_rndv_hdr_size;
4064         resp.tm_caps.max_num_tags       = attr.tm_caps.max_num_tags;
4065         resp.tm_caps.max_ops            = attr.tm_caps.max_ops;
4066         resp.tm_caps.max_sge            = attr.tm_caps.max_sge;
4067         resp.tm_caps.flags              = attr.tm_caps.flags;
4068         resp.response_length += sizeof(resp.tm_caps);
4069
4070         if (ucore->outlen < resp.response_length + sizeof(resp.cq_moderation_caps))
4071                 goto end;
4072
4073         resp.cq_moderation_caps.max_cq_moderation_count  =
4074                 attr.cq_caps.max_cq_moderation_count;
4075         resp.cq_moderation_caps.max_cq_moderation_period =
4076                 attr.cq_caps.max_cq_moderation_period;
4077         resp.response_length += sizeof(resp.cq_moderation_caps);
4078
4079         if (ucore->outlen < resp.response_length + sizeof(resp.max_dm_size))
4080                 goto end;
4081
4082         resp.max_dm_size = attr.max_dm_size;
4083         resp.response_length += sizeof(resp.max_dm_size);
4084 end:
4085         err = ib_copy_to_udata(ucore, &resp, resp.response_length);
4086         return err;
4087 }
4088
4089 int ib_uverbs_ex_modify_cq(struct ib_uverbs_file *file,
4090                            struct ib_udata *ucore,
4091                            struct ib_udata *uhw)
4092 {
4093         struct ib_uverbs_ex_modify_cq cmd = {};
4094         struct ib_cq *cq;
4095         size_t required_cmd_sz;
4096         int ret;
4097
4098         required_cmd_sz = offsetof(typeof(cmd), reserved) +
4099                                 sizeof(cmd.reserved);
4100         if (ucore->inlen < required_cmd_sz)
4101                 return -EINVAL;
4102
4103         /* sanity checks */
4104         if (ucore->inlen > sizeof(cmd) &&
4105             !ib_is_udata_cleared(ucore, sizeof(cmd),
4106                                  ucore->inlen - sizeof(cmd)))
4107                 return -EOPNOTSUPP;
4108
4109         ret = ib_copy_from_udata(&cmd, ucore, min(sizeof(cmd), ucore->inlen));
4110         if (ret)
4111                 return ret;
4112
4113         if (!cmd.attr_mask || cmd.reserved)
4114                 return -EINVAL;
4115
4116         if (cmd.attr_mask > IB_CQ_MODERATE)
4117                 return -EOPNOTSUPP;
4118
4119         cq = uobj_get_obj_read(cq, UVERBS_OBJECT_CQ, cmd.cq_handle, file);
4120         if (!cq)
4121                 return -EINVAL;
4122
4123         ret = rdma_set_cq_moderation(cq, cmd.attr.cq_count, cmd.attr.cq_period);
4124
4125         uobj_put_obj_read(cq);
4126
4127         return ret;
4128 }