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
  40 #include <asm/uaccess.h>
  41
  42 #include "uverbs.h"
  43
  44 struct uverbs_lock_class {
  45         struct lock_class_key   key;
  46         char                    name[16];
  47 };
  48
  49 static struct uverbs_lock_class pd_lock_class   = { .name = "PD-uobj" };
  50 static struct uverbs_lock_class mr_lock_class   = { .name = "MR-uobj" };
  51 static struct uverbs_lock_class mw_lock_class   = { .name = "MW-uobj" };
  52 static struct uverbs_lock_class cq_lock_class   = { .name = "CQ-uobj" };
  53 static struct uverbs_lock_class qp_lock_class   = { .name = "QP-uobj" };
  54 static struct uverbs_lock_class ah_lock_class   = { .name = "AH-uobj" };
  55 static struct uverbs_lock_class srq_lock_class  = { .name = "SRQ-uobj" };
  56 static struct uverbs_lock_class xrcd_lock_class = { .name = "XRCD-uobj" };
  57 #ifdef CONFIG_INFINIBAND_EXPERIMENTAL_UVERBS_FLOW_STEERING
  58 static struct uverbs_lock_class rule_lock_class = { .name = "RULE-uobj" };
  59 #endif /* CONFIG_INFINIBAND_EXPERIMENTAL_UVERBS_FLOW_STEERING */
  60
  61 #define INIT_UDATA(udata, ibuf, obuf, ilen, olen)                       \
  62         do {                                                            \
  63                 (udata)->inbuf  = (void __user *) (ibuf);               \
  64                 (udata)->outbuf = (void __user *) (obuf);               \
  65                 (udata)->inlen  = (ilen);                               \
  66                 (udata)->outlen = (olen);                               \
  67         } while (0)
  68
  69 /*
  70  * The ib_uobject locking scheme is as follows:
  71  *
  72  * - ib_uverbs_idr_lock protects the uverbs idrs themselves, so it
  73  *   needs to be held during all idr operations.  When an object is
  74  *   looked up, a reference must be taken on the object's kref before
  75  *   dropping this lock.
  76  *
  77  * - Each object also has an rwsem.  This rwsem must be held for
  78  *   reading while an operation that uses the object is performed.
  79  *   For example, while registering an MR, the associated PD's
  80  *   uobject.mutex must be held for reading.  The rwsem must be held
  81  *   for writing while initializing or destroying an object.
  82  *
  83  * - In addition, each object has a "live" flag.  If this flag is not
  84  *   set, then lookups of the object will fail even if it is found in
  85  *   the idr.  This handles a reader that blocks and does not acquire
  86  *   the rwsem until after the object is destroyed.  The destroy
  87  *   operation will set the live flag to 0 and then drop the rwsem;
  88  *   this will allow the reader to acquire the rwsem, see that the
  89  *   live flag is 0, and then drop the rwsem and its reference to
  90  *   object.  The underlying storage will not be freed until the last
  91  *   reference to the object is dropped.
  92  */
  93
  94 static void init_uobj(struct ib_uobject *uobj, u64 user_handle,
  95                       struct ib_ucontext *context, struct uverbs_lock_class *c)
  96 {
  97         uobj->user_handle = user_handle;
  98         uobj->context     = context;
  99         kref_init(&uobj->ref);
 100         init_rwsem(&uobj->mutex);
 101         lockdep_set_class_and_name(&uobj->mutex, &c->key, c->name);
 102         uobj->live        = 0;
 103 }
 104
 105 static void release_uobj(struct kref *kref)
 106 {
 107         kfree(container_of(kref, struct ib_uobject, ref));
 108 }
 109
 110 static void put_uobj(struct ib_uobject *uobj)
 111 {
 112         kref_put(&uobj->ref, release_uobj);
 113 }
 114
 115 static void put_uobj_read(struct ib_uobject *uobj)
 116 {
 117         up_read(&uobj->mutex);
 118         put_uobj(uobj);
 119 }
 120
 121 static void put_uobj_write(struct ib_uobject *uobj)
 122 {
 123         up_write(&uobj->mutex);
 124         put_uobj(uobj);
 125 }
 126
 127 static int idr_add_uobj(struct idr *idr, struct ib_uobject *uobj)
 128 {
 129         int ret;
 130
 131         idr_preload(GFP_KERNEL);
 132         spin_lock(&ib_uverbs_idr_lock);
 133
 134         ret = idr_alloc(idr, uobj, 0, 0, GFP_NOWAIT);
 135         if (ret >= 0)
 136                 uobj->id = ret;
 137
 138         spin_unlock(&ib_uverbs_idr_lock);
 139         idr_preload_end();
 140
 141         return ret < 0 ? ret : 0;
 142 }
 143
 144 void idr_remove_uobj(struct idr *idr, struct ib_uobject *uobj)
 145 {
 146         spin_lock(&ib_uverbs_idr_lock);
 147         idr_remove(idr, uobj->id);
 148         spin_unlock(&ib_uverbs_idr_lock);
 149 }
 150
 151 static struct ib_uobject *__idr_get_uobj(struct idr *idr, int id,
 152                                          struct ib_ucontext *context)
 153 {
 154         struct ib_uobject *uobj;
 155
 156         spin_lock(&ib_uverbs_idr_lock);
 157         uobj = idr_find(idr, id);
 158         if (uobj) {
 159                 if (uobj->context == context)
 160                         kref_get(&uobj->ref);
 161                 else
 162                         uobj = NULL;
 163         }
 164         spin_unlock(&ib_uverbs_idr_lock);
 165
 166         return uobj;
 167 }
 168
 169 static struct ib_uobject *idr_read_uobj(struct idr *idr, int id,
 170                                         struct ib_ucontext *context, int nested)
 171 {
 172         struct ib_uobject *uobj;
 173
 174         uobj = __idr_get_uobj(idr, id, context);
 175         if (!uobj)
 176                 return NULL;
 177
 178         if (nested)
 179                 down_read_nested(&uobj->mutex, SINGLE_DEPTH_NESTING);
 180         else
 181                 down_read(&uobj->mutex);
 182         if (!uobj->live) {
 183                 put_uobj_read(uobj);
 184                 return NULL;
 185         }
 186
 187         return uobj;
 188 }
 189
 190 static struct ib_uobject *idr_write_uobj(struct idr *idr, int id,
 191                                          struct ib_ucontext *context)
 192 {
 193         struct ib_uobject *uobj;
 194
 195         uobj = __idr_get_uobj(idr, id, context);
 196         if (!uobj)
 197                 return NULL;
 198
 199         down_write(&uobj->mutex);
 200         if (!uobj->live) {
 201                 put_uobj_write(uobj);
 202                 return NULL;
 203         }
 204
 205         return uobj;
 206 }
 207
 208 static void *idr_read_obj(struct idr *idr, int id, struct ib_ucontext *context,
 209                           int nested)
 210 {
 211         struct ib_uobject *uobj;
 212
 213         uobj = idr_read_uobj(idr, id, context, nested);
 214         return uobj ? uobj->object : NULL;
 215 }
 216
 217 static struct ib_pd *idr_read_pd(int pd_handle, struct ib_ucontext *context)
 218 {
 219         return idr_read_obj(&ib_uverbs_pd_idr, pd_handle, context, 0);
 220 }
 221
 222 static void put_pd_read(struct ib_pd *pd)
 223 {
 224         put_uobj_read(pd->uobject);
 225 }
 226
 227 static struct ib_cq *idr_read_cq(int cq_handle, struct ib_ucontext *context, int nested)
 228 {
 229         return idr_read_obj(&ib_uverbs_cq_idr, cq_handle, context, nested);
 230 }
 231
 232 static void put_cq_read(struct ib_cq *cq)
 233 {
 234         put_uobj_read(cq->uobject);
 235 }
 236
 237 static struct ib_ah *idr_read_ah(int ah_handle, struct ib_ucontext *context)
 238 {
 239         return idr_read_obj(&ib_uverbs_ah_idr, ah_handle, context, 0);
 240 }
 241
 242 static void put_ah_read(struct ib_ah *ah)
 243 {
 244         put_uobj_read(ah->uobject);
 245 }
 246
 247 static struct ib_qp *idr_read_qp(int qp_handle, struct ib_ucontext *context)
 248 {
 249         return idr_read_obj(&ib_uverbs_qp_idr, qp_handle, context, 0);
 250 }
 251
 252 static struct ib_qp *idr_write_qp(int qp_handle, struct ib_ucontext *context)
 253 {
 254         struct ib_uobject *uobj;
 255
 256         uobj = idr_write_uobj(&ib_uverbs_qp_idr, qp_handle, context);
 257         return uobj ? uobj->object : NULL;
 258 }
 259
 260 static void put_qp_read(struct ib_qp *qp)
 261 {
 262         put_uobj_read(qp->uobject);
 263 }
 264
 265 static void put_qp_write(struct ib_qp *qp)
 266 {
 267         put_uobj_write(qp->uobject);
 268 }
 269
 270 static struct ib_srq *idr_read_srq(int srq_handle, struct ib_ucontext *context)
 271 {
 272         return idr_read_obj(&ib_uverbs_srq_idr, srq_handle, context, 0);
 273 }
 274
 275 static void put_srq_read(struct ib_srq *srq)
 276 {
 277         put_uobj_read(srq->uobject);
 278 }
 279
 280 static struct ib_xrcd *idr_read_xrcd(int xrcd_handle, struct ib_ucontext *context,
 281                                      struct ib_uobject **uobj)
 282 {
 283         *uobj = idr_read_uobj(&ib_uverbs_xrcd_idr, xrcd_handle, context, 0);
 284         return *uobj ? (*uobj)->object : NULL;
 285 }
 286
 287 static void put_xrcd_read(struct ib_uobject *uobj)
 288 {
 289         put_uobj_read(uobj);
 290 }
 291
 292 ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
 293                               const char __user *buf,
 294                               int in_len, int out_len)
 295 {
 296         struct ib_uverbs_get_context      cmd;
 297         struct ib_uverbs_get_context_resp resp;
 298         struct ib_udata                   udata;
 299         struct ib_device                 *ibdev = file->device->ib_dev;
 300         struct ib_ucontext               *ucontext;
 301         struct file                      *filp;
 302         int ret;
 303
 304         if (out_len < sizeof resp)
 305                 return -ENOSPC;
 306
 307         if (copy_from_user(&cmd, buf, sizeof cmd))
 308                 return -EFAULT;
 309
 310         mutex_lock(&file->mutex);
 311
 312         if (file->ucontext) {
 313                 ret = -EINVAL;
 314                 goto err;
 315         }
 316
 317         INIT_UDATA(&udata, buf + sizeof cmd,
 318                    (unsigned long) cmd.response + sizeof resp,
 319                    in_len - sizeof cmd, out_len - sizeof resp);
 320
 321         ucontext = ibdev->alloc_ucontext(ibdev, &udata);
 322         if (IS_ERR(ucontext)) {
 323                 ret = PTR_ERR(ucontext);
 324                 goto err;
 325         }
 326
 327         ucontext->device = ibdev;
 328         INIT_LIST_HEAD(&ucontext->pd_list);
 329         INIT_LIST_HEAD(&ucontext->mr_list);
 330         INIT_LIST_HEAD(&ucontext->mw_list);
 331         INIT_LIST_HEAD(&ucontext->cq_list);
 332         INIT_LIST_HEAD(&ucontext->qp_list);
 333         INIT_LIST_HEAD(&ucontext->srq_list);
 334         INIT_LIST_HEAD(&ucontext->ah_list);
 335         INIT_LIST_HEAD(&ucontext->xrcd_list);
 336         INIT_LIST_HEAD(&ucontext->rule_list);
 337         ucontext->closing = 0;
 338
 339         resp.num_comp_vectors = file->device->num_comp_vectors;
 340
 341         ret = get_unused_fd_flags(O_CLOEXEC);
 342         if (ret < 0)
 343                 goto err_free;
 344         resp.async_fd = ret;
 345
 346         filp = ib_uverbs_alloc_event_file(file, 1);
 347         if (IS_ERR(filp)) {
 348                 ret = PTR_ERR(filp);
 349                 goto err_fd;
 350         }
 351
 352         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 353                          &resp, sizeof resp)) {
 354                 ret = -EFAULT;
 355                 goto err_file;
 356         }
 357
 358         file->async_file = filp->private_data;
 359
 360         INIT_IB_EVENT_HANDLER(&file->event_handler, file->device->ib_dev,
 361                               ib_uverbs_event_handler);
 362         ret = ib_register_event_handler(&file->event_handler);
 363         if (ret)
 364                 goto err_file;
 365
 366         kref_get(&file->async_file->ref);
 367         kref_get(&file->ref);
 368         file->ucontext = ucontext;
 369
 370         fd_install(resp.async_fd, filp);
 371
 372         mutex_unlock(&file->mutex);
 373
 374         return in_len;
 375
 376 err_file:
 377         fput(filp);
 378
 379 err_fd:
 380         put_unused_fd(resp.async_fd);
 381
 382 err_free:
 383         ibdev->dealloc_ucontext(ucontext);
 384
 385 err:
 386         mutex_unlock(&file->mutex);
 387         return ret;
 388 }
 389
 390 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 391                                const char __user *buf,
 392                                int in_len, int out_len)
 393 {
 394         struct ib_uverbs_query_device      cmd;
 395         struct ib_uverbs_query_device_resp resp;
 396         struct ib_device_attr              attr;
 397         int                                ret;
 398
 399         if (out_len < sizeof resp)
 400                 return -ENOSPC;
 401
 402         if (copy_from_user(&cmd, buf, sizeof cmd))
 403                 return -EFAULT;
 404
 405         ret = ib_query_device(file->device->ib_dev, &attr);
 406         if (ret)
 407                 return ret;
 408
 409         memset(&resp, 0, sizeof resp);
 410
 411         resp.fw_ver                    = attr.fw_ver;
 412         resp.node_guid                 = file->device->ib_dev->node_guid;
 413         resp.sys_image_guid            = attr.sys_image_guid;
 414         resp.max_mr_size               = attr.max_mr_size;
 415         resp.page_size_cap             = attr.page_size_cap;
 416         resp.vendor_id                 = attr.vendor_id;
 417         resp.vendor_part_id            = attr.vendor_part_id;
 418         resp.hw_ver                    = attr.hw_ver;
 419         resp.max_qp                    = attr.max_qp;
 420         resp.max_qp_wr                 = attr.max_qp_wr;
 421         resp.device_cap_flags          = attr.device_cap_flags;
 422         resp.max_sge                   = attr.max_sge;
 423         resp.max_sge_rd                = attr.max_sge_rd;
 424         resp.max_cq                    = attr.max_cq;
 425         resp.max_cqe                   = attr.max_cqe;
 426         resp.max_mr                    = attr.max_mr;
 427         resp.max_pd                    = attr.max_pd;
 428         resp.max_qp_rd_atom            = attr.max_qp_rd_atom;
 429         resp.max_ee_rd_atom            = attr.max_ee_rd_atom;
 430         resp.max_res_rd_atom           = attr.max_res_rd_atom;
 431         resp.max_qp_init_rd_atom       = attr.max_qp_init_rd_atom;
 432         resp.max_ee_init_rd_atom       = attr.max_ee_init_rd_atom;
 433         resp.atomic_cap                = attr.atomic_cap;
 434         resp.max_ee                    = attr.max_ee;
 435         resp.max_rdd                   = attr.max_rdd;
 436         resp.max_mw                    = attr.max_mw;
 437         resp.max_raw_ipv6_qp           = attr.max_raw_ipv6_qp;
 438         resp.max_raw_ethy_qp           = attr.max_raw_ethy_qp;
 439         resp.max_mcast_grp             = attr.max_mcast_grp;
 440         resp.max_mcast_qp_attach       = attr.max_mcast_qp_attach;
 441         resp.max_total_mcast_qp_attach = attr.max_total_mcast_qp_attach;
 442         resp.max_ah                    = attr.max_ah;
 443         resp.max_fmr                   = attr.max_fmr;
 444         resp.max_map_per_fmr           = attr.max_map_per_fmr;
 445         resp.max_srq                   = attr.max_srq;
 446         resp.max_srq_wr                = attr.max_srq_wr;
 447         resp.max_srq_sge               = attr.max_srq_sge;
 448         resp.max_pkeys                 = attr.max_pkeys;
 449         resp.local_ca_ack_delay        = attr.local_ca_ack_delay;
 450         resp.phys_port_cnt             = file->device->ib_dev->phys_port_cnt;
 451
 452         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 453                          &resp, sizeof resp))
 454                 return -EFAULT;
 455
 456         return in_len;
 457 }
 458
 459 ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
 460                              const char __user *buf,
 461                              int in_len, int out_len)
 462 {
 463         struct ib_uverbs_query_port      cmd;
 464         struct ib_uverbs_query_port_resp resp;
 465         struct ib_port_attr              attr;
 466         int                              ret;
 467
 468         if (out_len < sizeof resp)
 469                 return -ENOSPC;
 470
 471         if (copy_from_user(&cmd, buf, sizeof cmd))
 472                 return -EFAULT;
 473
 474         ret = ib_query_port(file->device->ib_dev, cmd.port_num, &attr);
 475         if (ret)
 476                 return ret;
 477
 478         memset(&resp, 0, sizeof resp);
 479
 480         resp.state           = attr.state;
 481         resp.max_mtu         = attr.max_mtu;
 482         resp.active_mtu      = attr.active_mtu;
 483         resp.gid_tbl_len     = attr.gid_tbl_len;
 484         resp.port_cap_flags  = attr.port_cap_flags;
 485         resp.max_msg_sz      = attr.max_msg_sz;
 486         resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
 487         resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
 488         resp.pkey_tbl_len    = attr.pkey_tbl_len;
 489         resp.lid             = attr.lid;
 490         resp.sm_lid          = attr.sm_lid;
 491         resp.lmc             = attr.lmc;
 492         resp.max_vl_num      = attr.max_vl_num;
 493         resp.sm_sl           = attr.sm_sl;
 494         resp.subnet_timeout  = attr.subnet_timeout;
 495         resp.init_type_reply = attr.init_type_reply;
 496         resp.active_width    = attr.active_width;
 497         resp.active_speed    = attr.active_speed;
 498         resp.phys_state      = attr.phys_state;
 499         resp.link_layer      = rdma_port_get_link_layer(file->device->ib_dev,
 500                                                         cmd.port_num);
 501
 502         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 503                          &resp, sizeof resp))
 504                 return -EFAULT;
 505
 506         return in_len;
 507 }
 508
 509 ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
 510                            const char __user *buf,
 511                            int in_len, int out_len)
 512 {
 513         struct ib_uverbs_alloc_pd      cmd;
 514         struct ib_uverbs_alloc_pd_resp resp;
 515         struct ib_udata                udata;
 516         struct ib_uobject             *uobj;
 517         struct ib_pd                  *pd;
 518         int                            ret;
 519
 520         if (out_len < sizeof resp)
 521                 return -ENOSPC;
 522
 523         if (copy_from_user(&cmd, buf, sizeof cmd))
 524                 return -EFAULT;
 525
 526         INIT_UDATA(&udata, buf + sizeof cmd,
 527                    (unsigned long) cmd.response + sizeof resp,
 528                    in_len - sizeof cmd, out_len - sizeof resp);
 529
 530         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 531         if (!uobj)
 532                 return -ENOMEM;
 533
 534         init_uobj(uobj, 0, file->ucontext, &pd_lock_class);
 535         down_write(&uobj->mutex);
 536
 537         pd = file->device->ib_dev->alloc_pd(file->device->ib_dev,
 538                                             file->ucontext, &udata);
 539         if (IS_ERR(pd)) {
 540                 ret = PTR_ERR(pd);
 541                 goto err;
 542         }
 543
 544         pd->device  = file->device->ib_dev;
 545         pd->uobject = uobj;
 546         atomic_set(&pd->usecnt, 0);
 547
 548         uobj->object = pd;
 549         ret = idr_add_uobj(&ib_uverbs_pd_idr, uobj);
 550         if (ret)
 551                 goto err_idr;
 552
 553         memset(&resp, 0, sizeof resp);
 554         resp.pd_handle = uobj->id;
 555
 556         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 557                          &resp, sizeof resp)) {
 558                 ret = -EFAULT;
 559                 goto err_copy;
 560         }
 561
 562         mutex_lock(&file->mutex);
 563         list_add_tail(&uobj->list, &file->ucontext->pd_list);
 564         mutex_unlock(&file->mutex);
 565
 566         uobj->live = 1;
 567
 568         up_write(&uobj->mutex);
 569
 570         return in_len;
 571
 572 err_copy:
 573         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 574
 575 err_idr:
 576         ib_dealloc_pd(pd);
 577
 578 err:
 579         put_uobj_write(uobj);
 580         return ret;
 581 }
 582
 583 ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
 584                              const char __user *buf,
 585                              int in_len, int out_len)
 586 {
 587         struct ib_uverbs_dealloc_pd cmd;
 588         struct ib_uobject          *uobj;
 589         int                         ret;
 590
 591         if (copy_from_user(&cmd, buf, sizeof cmd))
 592                 return -EFAULT;
 593
 594         uobj = idr_write_uobj(&ib_uverbs_pd_idr, cmd.pd_handle, file->ucontext);
 595         if (!uobj)
 596                 return -EINVAL;
 597
 598         ret = ib_dealloc_pd(uobj->object);
 599         if (!ret)
 600                 uobj->live = 0;
 601
 602         put_uobj_write(uobj);
 603
 604         if (ret)
 605                 return ret;
 606
 607         idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 608
 609         mutex_lock(&file->mutex);
 610         list_del(&uobj->list);
 611         mutex_unlock(&file->mutex);
 612
 613         put_uobj(uobj);
 614
 615         return in_len;
 616 }
 617
 618 struct xrcd_table_entry {
 619         struct rb_node  node;
 620         struct ib_xrcd *xrcd;
 621         struct inode   *inode;
 622 };
 623
 624 static int xrcd_table_insert(struct ib_uverbs_device *dev,
 625                             struct inode *inode,
 626                             struct ib_xrcd *xrcd)
 627 {
 628         struct xrcd_table_entry *entry, *scan;
 629         struct rb_node **p = &dev->xrcd_tree.rb_node;
 630         struct rb_node *parent = NULL;
 631
 632         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 633         if (!entry)
 634                 return -ENOMEM;
 635
 636         entry->xrcd  = xrcd;
 637         entry->inode = inode;
 638
 639         while (*p) {
 640                 parent = *p;
 641                 scan = rb_entry(parent, struct xrcd_table_entry, node);
 642
 643                 if (inode < scan->inode) {
 644                         p = &(*p)->rb_left;
 645                 } else if (inode > scan->inode) {
 646                         p = &(*p)->rb_right;
 647                 } else {
 648                         kfree(entry);
 649                         return -EEXIST;
 650                 }
 651         }
 652
 653         rb_link_node(&entry->node, parent, p);
 654         rb_insert_color(&entry->node, &dev->xrcd_tree);
 655         igrab(inode);
 656         return 0;
 657 }
 658
 659 static struct xrcd_table_entry *xrcd_table_search(struct ib_uverbs_device *dev,
 660                                                   struct inode *inode)
 661 {
 662         struct xrcd_table_entry *entry;
 663         struct rb_node *p = dev->xrcd_tree.rb_node;
 664
 665         while (p) {
 666                 entry = rb_entry(p, struct xrcd_table_entry, node);
 667
 668                 if (inode < entry->inode)
 669                         p = p->rb_left;
 670                 else if (inode > entry->inode)
 671                         p = p->rb_right;
 672                 else
 673                         return entry;
 674         }
 675
 676         return NULL;
 677 }
 678
 679 static struct ib_xrcd *find_xrcd(struct ib_uverbs_device *dev, struct inode *inode)
 680 {
 681         struct xrcd_table_entry *entry;
 682
 683         entry = xrcd_table_search(dev, inode);
 684         if (!entry)
 685                 return NULL;
 686
 687         return entry->xrcd;
 688 }
 689
 690 static void xrcd_table_delete(struct ib_uverbs_device *dev,
 691                               struct inode *inode)
 692 {
 693         struct xrcd_table_entry *entry;
 694
 695         entry = xrcd_table_search(dev, inode);
 696         if (entry) {
 697                 iput(inode);
 698                 rb_erase(&entry->node, &dev->xrcd_tree);
 699                 kfree(entry);
 700         }
 701 }
 702
 703 ssize_t ib_uverbs_open_xrcd(struct ib_uverbs_file *file,
 704                             const char __user *buf, int in_len,
 705                             int out_len)
 706 {
 707         struct ib_uverbs_open_xrcd      cmd;
 708         struct ib_uverbs_open_xrcd_resp resp;
 709         struct ib_udata                 udata;
 710         struct ib_uxrcd_object         *obj;
 711         struct ib_xrcd                 *xrcd = NULL;
 712         struct fd                       f = {NULL, 0};
 713         struct inode                   *inode = NULL;
 714         int                             ret = 0;
 715         int                             new_xrcd = 0;
 716
 717         if (out_len < sizeof resp)
 718                 return -ENOSPC;
 719
 720         if (copy_from_user(&cmd, buf, sizeof cmd))
 721                 return -EFAULT;
 722
 723         INIT_UDATA(&udata, buf + sizeof cmd,
 724                    (unsigned long) cmd.response + sizeof resp,
 725                    in_len - sizeof cmd, out_len - sizeof  resp);
 726
 727         mutex_lock(&file->device->xrcd_tree_mutex);
 728
 729         if (cmd.fd != -1) {
 730                 /* search for file descriptor */
 731                 f = fdget(cmd.fd);
 732                 if (!f.file) {
 733                         ret = -EBADF;
 734                         goto err_tree_mutex_unlock;
 735                 }
 736
 737                 inode = file_inode(f.file);
 738                 xrcd = find_xrcd(file->device, inode);
 739                 if (!xrcd && !(cmd.oflags & O_CREAT)) {
 740                         /* no file descriptor. Need CREATE flag */
 741                         ret = -EAGAIN;
 742                         goto err_tree_mutex_unlock;
 743                 }
 744
 745                 if (xrcd && cmd.oflags & O_EXCL) {
 746                         ret = -EINVAL;
 747                         goto err_tree_mutex_unlock;
 748                 }
 749         }
 750
 751         obj = kmalloc(sizeof *obj, GFP_KERNEL);
 752         if (!obj) {
 753                 ret = -ENOMEM;
 754                 goto err_tree_mutex_unlock;
 755         }
 756
 757         init_uobj(&obj->uobject, 0, file->ucontext, &xrcd_lock_class);
 758
 759         down_write(&obj->uobject.mutex);
 760
 761         if (!xrcd) {
 762                 xrcd = file->device->ib_dev->alloc_xrcd(file->device->ib_dev,
 763                                                         file->ucontext, &udata);
 764                 if (IS_ERR(xrcd)) {
 765                         ret = PTR_ERR(xrcd);
 766                         goto err;
 767                 }
 768
 769                 xrcd->inode   = inode;
 770                 xrcd->device  = file->device->ib_dev;
 771                 atomic_set(&xrcd->usecnt, 0);
 772                 mutex_init(&xrcd->tgt_qp_mutex);
 773                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
 774                 new_xrcd = 1;
 775         }
 776
 777         atomic_set(&obj->refcnt, 0);
 778         obj->uobject.object = xrcd;
 779         ret = idr_add_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 780         if (ret)
 781                 goto err_idr;
 782
 783         memset(&resp, 0, sizeof resp);
 784         resp.xrcd_handle = obj->uobject.id;
 785
 786         if (inode) {
 787                 if (new_xrcd) {
 788                         /* create new inode/xrcd table entry */
 789                         ret = xrcd_table_insert(file->device, inode, xrcd);
 790                         if (ret)
 791                                 goto err_insert_xrcd;
 792                 }
 793                 atomic_inc(&xrcd->usecnt);
 794         }
 795
 796         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 797                          &resp, sizeof resp)) {
 798                 ret = -EFAULT;
 799                 goto err_copy;
 800         }
 801
 802         if (f.file)
 803                 fdput(f);
 804
 805         mutex_lock(&file->mutex);
 806         list_add_tail(&obj->uobject.list, &file->ucontext->xrcd_list);
 807         mutex_unlock(&file->mutex);
 808
 809         obj->uobject.live = 1;
 810         up_write(&obj->uobject.mutex);
 811
 812         mutex_unlock(&file->device->xrcd_tree_mutex);
 813         return in_len;
 814
 815 err_copy:
 816         if (inode) {
 817                 if (new_xrcd)
 818                         xrcd_table_delete(file->device, inode);
 819                 atomic_dec(&xrcd->usecnt);
 820         }
 821
 822 err_insert_xrcd:
 823         idr_remove_uobj(&ib_uverbs_xrcd_idr, &obj->uobject);
 824
 825 err_idr:
 826         ib_dealloc_xrcd(xrcd);
 827
 828 err:
 829         put_uobj_write(&obj->uobject);
 830
 831 err_tree_mutex_unlock:
 832         if (f.file)
 833                 fdput(f);
 834
 835         mutex_unlock(&file->device->xrcd_tree_mutex);
 836
 837         return ret;
 838 }
 839
 840 ssize_t ib_uverbs_close_xrcd(struct ib_uverbs_file *file,
 841                              const char __user *buf, int in_len,
 842                              int out_len)
 843 {
 844         struct ib_uverbs_close_xrcd cmd;
 845         struct ib_uobject           *uobj;
 846         struct ib_xrcd              *xrcd = NULL;
 847         struct inode                *inode = NULL;
 848         struct ib_uxrcd_object      *obj;
 849         int                         live;
 850         int                         ret = 0;
 851
 852         if (copy_from_user(&cmd, buf, sizeof cmd))
 853                 return -EFAULT;
 854
 855         mutex_lock(&file->device->xrcd_tree_mutex);
 856         uobj = idr_write_uobj(&ib_uverbs_xrcd_idr, cmd.xrcd_handle, file->ucontext);
 857         if (!uobj) {
 858                 ret = -EINVAL;
 859                 goto out;
 860         }
 861
 862         xrcd  = uobj->object;
 863         inode = xrcd->inode;
 864         obj   = container_of(uobj, struct ib_uxrcd_object, uobject);
 865         if (atomic_read(&obj->refcnt)) {
 866                 put_uobj_write(uobj);
 867                 ret = -EBUSY;
 868                 goto out;
 869         }
 870
 871         if (!inode || atomic_dec_and_test(&xrcd->usecnt)) {
 872                 ret = ib_dealloc_xrcd(uobj->object);
 873                 if (!ret)
 874                         uobj->live = 0;
 875         }
 876
 877         live = uobj->live;
 878         if (inode && ret)
 879                 atomic_inc(&xrcd->usecnt);
 880
 881         put_uobj_write(uobj);
 882
 883         if (ret)
 884                 goto out;
 885
 886         if (inode && !live)
 887                 xrcd_table_delete(file->device, inode);
 888
 889         idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 890         mutex_lock(&file->mutex);
 891         list_del(&uobj->list);
 892         mutex_unlock(&file->mutex);
 893
 894         put_uobj(uobj);
 895         ret = in_len;
 896
 897 out:
 898         mutex_unlock(&file->device->xrcd_tree_mutex);
 899         return ret;
 900 }
 901
 902 void ib_uverbs_dealloc_xrcd(struct ib_uverbs_device *dev,
 903                             struct ib_xrcd *xrcd)
 904 {
 905         struct inode *inode;
 906
 907         inode = xrcd->inode;
 908         if (inode && !atomic_dec_and_test(&xrcd->usecnt))
 909                 return;
 910
 911         ib_dealloc_xrcd(xrcd);
 912
 913         if (inode)
 914                 xrcd_table_delete(dev, inode);
 915 }
 916
 917 ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
 918                          const char __user *buf, int in_len,
 919                          int out_len)
 920 {
 921         struct ib_uverbs_reg_mr      cmd;
 922         struct ib_uverbs_reg_mr_resp resp;
 923         struct ib_udata              udata;
 924         struct ib_uobject           *uobj;
 925         struct ib_pd                *pd;
 926         struct ib_mr                *mr;
 927         int                          ret;
 928
 929         if (out_len < sizeof resp)
 930                 return -ENOSPC;
 931
 932         if (copy_from_user(&cmd, buf, sizeof cmd))
 933                 return -EFAULT;
 934
 935         INIT_UDATA(&udata, buf + sizeof cmd,
 936                    (unsigned long) cmd.response + sizeof resp,
 937                    in_len - sizeof cmd, out_len - sizeof resp);
 938
 939         if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
 940                 return -EINVAL;
 941
 942         /*
 943          * Local write permission is required if remote write or
 944          * remote atomic permission is also requested.
 945          */
 946         if (cmd.access_flags & (IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_REMOTE_WRITE) &&
 947             !(cmd.access_flags & IB_ACCESS_LOCAL_WRITE))
 948                 return -EINVAL;
 949
 950         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
 951         if (!uobj)
 952                 return -ENOMEM;
 953
 954         init_uobj(uobj, 0, file->ucontext, &mr_lock_class);
 955         down_write(&uobj->mutex);
 956
 957         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
 958         if (!pd) {
 959                 ret = -EINVAL;
 960                 goto err_free;
 961         }
 962
 963         mr = pd->device->reg_user_mr(pd, cmd.start, cmd.length, cmd.hca_va,
 964                                      cmd.access_flags, &udata);
 965         if (IS_ERR(mr)) {
 966                 ret = PTR_ERR(mr);
 967                 goto err_put;
 968         }
 969
 970         mr->device  = pd->device;
 971         mr->pd      = pd;
 972         mr->uobject = uobj;
 973         atomic_inc(&pd->usecnt);
 974         atomic_set(&mr->usecnt, 0);
 975
 976         uobj->object = mr;
 977         ret = idr_add_uobj(&ib_uverbs_mr_idr, uobj);
 978         if (ret)
 979                 goto err_unreg;
 980
 981         memset(&resp, 0, sizeof resp);
 982         resp.lkey      = mr->lkey;
 983         resp.rkey      = mr->rkey;
 984         resp.mr_handle = uobj->id;
 985
 986         if (copy_to_user((void __user *) (unsigned long) cmd.response,
 987                          &resp, sizeof resp)) {
 988                 ret = -EFAULT;
 989                 goto err_copy;
 990         }
 991
 992         put_pd_read(pd);
 993
 994         mutex_lock(&file->mutex);
 995         list_add_tail(&uobj->list, &file->ucontext->mr_list);
 996         mutex_unlock(&file->mutex);
 997
 998         uobj->live = 1;
 999
1000         up_write(&uobj->mutex);
1001
1002         return in_len;
1003
1004 err_copy:
1005         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1006
1007 err_unreg:
1008         ib_dereg_mr(mr);
1009
1010 err_put:
1011         put_pd_read(pd);
1012
1013 err_free:
1014         put_uobj_write(uobj);
1015         return ret;
1016 }
1017
1018 ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
1019                            const char __user *buf, int in_len,
1020                            int out_len)
1021 {
1022         struct ib_uverbs_dereg_mr cmd;
1023         struct ib_mr             *mr;
1024         struct ib_uobject        *uobj;
1025         int                       ret = -EINVAL;
1026
1027         if (copy_from_user(&cmd, buf, sizeof cmd))
1028                 return -EFAULT;
1029
1030         uobj = idr_write_uobj(&ib_uverbs_mr_idr, cmd.mr_handle, file->ucontext);
1031         if (!uobj)
1032                 return -EINVAL;
1033
1034         mr = uobj->object;
1035
1036         ret = ib_dereg_mr(mr);
1037         if (!ret)
1038                 uobj->live = 0;
1039
1040         put_uobj_write(uobj);
1041
1042         if (ret)
1043                 return ret;
1044
1045         idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
1046
1047         mutex_lock(&file->mutex);
1048         list_del(&uobj->list);
1049         mutex_unlock(&file->mutex);
1050
1051         put_uobj(uobj);
1052
1053         return in_len;
1054 }
1055
1056 ssize_t ib_uverbs_alloc_mw(struct ib_uverbs_file *file,
1057                          const char __user *buf, int in_len,
1058                          int out_len)
1059 {
1060         struct ib_uverbs_alloc_mw      cmd;
1061         struct ib_uverbs_alloc_mw_resp resp;
1062         struct ib_uobject             *uobj;
1063         struct ib_pd                  *pd;
1064         struct ib_mw                  *mw;
1065         int                            ret;
1066
1067         if (out_len < sizeof(resp))
1068                 return -ENOSPC;
1069
1070         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1071                 return -EFAULT;
1072
1073         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
1074         if (!uobj)
1075                 return -ENOMEM;
1076
1077         init_uobj(uobj, 0, file->ucontext, &mw_lock_class);
1078         down_write(&uobj->mutex);
1079
1080         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
1081         if (!pd) {
1082                 ret = -EINVAL;
1083                 goto err_free;
1084         }
1085
1086         mw = pd->device->alloc_mw(pd, cmd.mw_type);
1087         if (IS_ERR(mw)) {
1088                 ret = PTR_ERR(mw);
1089                 goto err_put;
1090         }
1091
1092         mw->device  = pd->device;
1093         mw->pd      = pd;
1094         mw->uobject = uobj;
1095         atomic_inc(&pd->usecnt);
1096
1097         uobj->object = mw;
1098         ret = idr_add_uobj(&ib_uverbs_mw_idr, uobj);
1099         if (ret)
1100                 goto err_unalloc;
1101
1102         memset(&resp, 0, sizeof(resp));
1103         resp.rkey      = mw->rkey;
1104         resp.mw_handle = uobj->id;
1105
1106         if (copy_to_user((void __user *)(unsigned long)cmd.response,
1107                          &resp, sizeof(resp))) {
1108                 ret = -EFAULT;
1109                 goto err_copy;
1110         }
1111
1112         put_pd_read(pd);
1113
1114         mutex_lock(&file->mutex);
1115         list_add_tail(&uobj->list, &file->ucontext->mw_list);
1116         mutex_unlock(&file->mutex);
1117
1118         uobj->live = 1;
1119
1120         up_write(&uobj->mutex);
1121
1122         return in_len;
1123
1124 err_copy:
1125         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1126
1127 err_unalloc:
1128         ib_dealloc_mw(mw);
1129
1130 err_put:
1131         put_pd_read(pd);
1132
1133 err_free:
1134         put_uobj_write(uobj);
1135         return ret;
1136 }
1137
1138 ssize_t ib_uverbs_dealloc_mw(struct ib_uverbs_file *file,
1139                            const char __user *buf, int in_len,
1140                            int out_len)
1141 {
1142         struct ib_uverbs_dealloc_mw cmd;
1143         struct ib_mw               *mw;
1144         struct ib_uobject          *uobj;
1145         int                         ret = -EINVAL;
1146
1147         if (copy_from_user(&cmd, buf, sizeof(cmd)))
1148                 return -EFAULT;
1149
1150         uobj = idr_write_uobj(&ib_uverbs_mw_idr, cmd.mw_handle, file->ucontext);
1151         if (!uobj)
1152                 return -EINVAL;
1153
1154         mw = uobj->object;
1155
1156         ret = ib_dealloc_mw(mw);
1157         if (!ret)
1158                 uobj->live = 0;
1159
1160         put_uobj_write(uobj);
1161
1162         if (ret)
1163                 return ret;
1164
1165         idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
1166
1167         mutex_lock(&file->mutex);
1168         list_del(&uobj->list);
1169         mutex_unlock(&file->mutex);
1170
1171         put_uobj(uobj);
1172
1173         return in_len;
1174 }
1175
1176 ssize_t ib_uverbs_create_comp_channel(struct ib_uverbs_file *file,
1177                                       const char __user *buf, int in_len,
1178                                       int out_len)
1179 {
1180         struct ib_uverbs_create_comp_channel       cmd;
1181         struct ib_uverbs_create_comp_channel_resp  resp;
1182         struct file                               *filp;
1183         int ret;
1184
1185         if (out_len < sizeof resp)
1186                 return -ENOSPC;
1187
1188         if (copy_from_user(&cmd, buf, sizeof cmd))
1189                 return -EFAULT;
1190
1191         ret = get_unused_fd_flags(O_CLOEXEC);
1192         if (ret < 0)
1193                 return ret;
1194         resp.fd = ret;
1195
1196         filp = ib_uverbs_alloc_event_file(file, 0);
1197         if (IS_ERR(filp)) {
1198                 put_unused_fd(resp.fd);
1199                 return PTR_ERR(filp);
1200         }
1201
1202         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1203                          &resp, sizeof resp)) {
1204                 put_unused_fd(resp.fd);
1205                 fput(filp);
1206                 return -EFAULT;
1207         }
1208
1209         fd_install(resp.fd, filp);
1210         return in_len;
1211 }
1212
1213 ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
1214                             const char __user *buf, int in_len,
1215                             int out_len)
1216 {
1217         struct ib_uverbs_create_cq      cmd;
1218         struct ib_uverbs_create_cq_resp resp;
1219         struct ib_udata                 udata;
1220         struct ib_ucq_object           *obj;
1221         struct ib_uverbs_event_file    *ev_file = NULL;
1222         struct ib_cq                   *cq;
1223         int                             ret;
1224
1225         if (out_len < sizeof resp)
1226                 return -ENOSPC;
1227
1228         if (copy_from_user(&cmd, buf, sizeof cmd))
1229                 return -EFAULT;
1230
1231         INIT_UDATA(&udata, buf + sizeof cmd,
1232                    (unsigned long) cmd.response + sizeof resp,
1233                    in_len - sizeof cmd, out_len - sizeof resp);
1234
1235         if (cmd.comp_vector >= file->device->num_comp_vectors)
1236                 return -EINVAL;
1237
1238         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1239         if (!obj)
1240                 return -ENOMEM;
1241
1242         init_uobj(&obj->uobject, cmd.user_handle, file->ucontext, &cq_lock_class);
1243         down_write(&obj->uobject.mutex);
1244
1245         if (cmd.comp_channel >= 0) {
1246                 ev_file = ib_uverbs_lookup_comp_file(cmd.comp_channel);
1247                 if (!ev_file) {
1248                         ret = -EINVAL;
1249                         goto err;
1250                 }
1251         }
1252
1253         obj->uverbs_file           = file;
1254         obj->comp_events_reported  = 0;
1255         obj->async_events_reported = 0;
1256         INIT_LIST_HEAD(&obj->comp_list);
1257         INIT_LIST_HEAD(&obj->async_list);
1258
1259         cq = file->device->ib_dev->create_cq(file->device->ib_dev, cmd.cqe,
1260                                              cmd.comp_vector,
1261                                              file->ucontext, &udata);
1262         if (IS_ERR(cq)) {
1263                 ret = PTR_ERR(cq);
1264                 goto err_file;
1265         }
1266
1267         cq->device        = file->device->ib_dev;
1268         cq->uobject       = &obj->uobject;
1269         cq->comp_handler  = ib_uverbs_comp_handler;
1270         cq->event_handler = ib_uverbs_cq_event_handler;
1271         cq->cq_context    = ev_file;
1272         atomic_set(&cq->usecnt, 0);
1273
1274         obj->uobject.object = cq;
1275         ret = idr_add_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1276         if (ret)
1277                 goto err_free;
1278
1279         memset(&resp, 0, sizeof resp);
1280         resp.cq_handle = obj->uobject.id;
1281         resp.cqe       = cq->cqe;
1282
1283         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1284                          &resp, sizeof resp)) {
1285                 ret = -EFAULT;
1286                 goto err_copy;
1287         }
1288
1289         mutex_lock(&file->mutex);
1290         list_add_tail(&obj->uobject.list, &file->ucontext->cq_list);
1291         mutex_unlock(&file->mutex);
1292
1293         obj->uobject.live = 1;
1294
1295         up_write(&obj->uobject.mutex);
1296
1297         return in_len;
1298
1299 err_copy:
1300         idr_remove_uobj(&ib_uverbs_cq_idr, &obj->uobject);
1301
1302 err_free:
1303         ib_destroy_cq(cq);
1304
1305 err_file:
1306         if (ev_file)
1307                 ib_uverbs_release_ucq(file, ev_file, obj);
1308
1309 err:
1310         put_uobj_write(&obj->uobject);
1311         return ret;
1312 }
1313
1314 ssize_t ib_uverbs_resize_cq(struct ib_uverbs_file *file,
1315                             const char __user *buf, int in_len,
1316                             int out_len)
1317 {
1318         struct ib_uverbs_resize_cq      cmd;
1319         struct ib_uverbs_resize_cq_resp resp;
1320         struct ib_udata                 udata;
1321         struct ib_cq                    *cq;
1322         int                             ret = -EINVAL;
1323
1324         if (copy_from_user(&cmd, buf, sizeof cmd))
1325                 return -EFAULT;
1326
1327         INIT_UDATA(&udata, buf + sizeof cmd,
1328                    (unsigned long) cmd.response + sizeof resp,
1329                    in_len - sizeof cmd, out_len - sizeof resp);
1330
1331         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1332         if (!cq)
1333                 return -EINVAL;
1334
1335         ret = cq->device->resize_cq(cq, cmd.cqe, &udata);
1336         if (ret)
1337                 goto out;
1338
1339         resp.cqe = cq->cqe;
1340
1341         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1342                          &resp, sizeof resp.cqe))
1343                 ret = -EFAULT;
1344
1345 out:
1346         put_cq_read(cq);
1347
1348         return ret ? ret : in_len;
1349 }
1350
1351 static int copy_wc_to_user(void __user *dest, struct ib_wc *wc)
1352 {
1353         struct ib_uverbs_wc tmp;
1354
1355         tmp.wr_id               = wc->wr_id;
1356         tmp.status              = wc->status;
1357         tmp.opcode              = wc->opcode;
1358         tmp.vendor_err          = wc->vendor_err;
1359         tmp.byte_len            = wc->byte_len;
1360         tmp.ex.imm_data         = (__u32 __force) wc->ex.imm_data;
1361         tmp.qp_num              = wc->qp->qp_num;
1362         tmp.src_qp              = wc->src_qp;
1363         tmp.wc_flags            = wc->wc_flags;
1364         tmp.pkey_index          = wc->pkey_index;
1365         tmp.slid                = wc->slid;
1366         tmp.sl                  = wc->sl;
1367         tmp.dlid_path_bits      = wc->dlid_path_bits;
1368         tmp.port_num            = wc->port_num;
1369         tmp.reserved            = 0;
1370
1371         if (copy_to_user(dest, &tmp, sizeof tmp))
1372                 return -EFAULT;
1373
1374         return 0;
1375 }
1376
1377 ssize_t ib_uverbs_poll_cq(struct ib_uverbs_file *file,
1378                           const char __user *buf, int in_len,
1379                           int out_len)
1380 {
1381         struct ib_uverbs_poll_cq       cmd;
1382         struct ib_uverbs_poll_cq_resp  resp;
1383         u8 __user                     *header_ptr;
1384         u8 __user                     *data_ptr;
1385         struct ib_cq                  *cq;
1386         struct ib_wc                   wc;
1387         int                            ret;
1388
1389         if (copy_from_user(&cmd, buf, sizeof cmd))
1390                 return -EFAULT;
1391
1392         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1393         if (!cq)
1394                 return -EINVAL;
1395
1396         /* we copy a struct ib_uverbs_poll_cq_resp to user space */
1397         header_ptr = (void __user *)(unsigned long) cmd.response;
1398         data_ptr = header_ptr + sizeof resp;
1399
1400         memset(&resp, 0, sizeof resp);
1401         while (resp.count < cmd.ne) {
1402                 ret = ib_poll_cq(cq, 1, &wc);
1403                 if (ret < 0)
1404                         goto out_put;
1405                 if (!ret)
1406                         break;
1407
1408                 ret = copy_wc_to_user(data_ptr, &wc);
1409                 if (ret)
1410                         goto out_put;
1411
1412                 data_ptr += sizeof(struct ib_uverbs_wc);
1413                 ++resp.count;
1414         }
1415
1416         if (copy_to_user(header_ptr, &resp, sizeof resp)) {
1417                 ret = -EFAULT;
1418                 goto out_put;
1419         }
1420
1421         ret = in_len;
1422
1423 out_put:
1424         put_cq_read(cq);
1425         return ret;
1426 }
1427
1428 ssize_t ib_uverbs_req_notify_cq(struct ib_uverbs_file *file,
1429                                 const char __user *buf, int in_len,
1430                                 int out_len)
1431 {
1432         struct ib_uverbs_req_notify_cq cmd;
1433         struct ib_cq                  *cq;
1434
1435         if (copy_from_user(&cmd, buf, sizeof cmd))
1436                 return -EFAULT;
1437
1438         cq = idr_read_cq(cmd.cq_handle, file->ucontext, 0);
1439         if (!cq)
1440                 return -EINVAL;
1441
1442         ib_req_notify_cq(cq, cmd.solicited_only ?
1443                          IB_CQ_SOLICITED : IB_CQ_NEXT_COMP);
1444
1445         put_cq_read(cq);
1446
1447         return in_len;
1448 }
1449
1450 ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
1451                              const char __user *buf, int in_len,
1452                              int out_len)
1453 {
1454         struct ib_uverbs_destroy_cq      cmd;
1455         struct ib_uverbs_destroy_cq_resp resp;
1456         struct ib_uobject               *uobj;
1457         struct ib_cq                    *cq;
1458         struct ib_ucq_object            *obj;
1459         struct ib_uverbs_event_file     *ev_file;
1460         int                              ret = -EINVAL;
1461
1462         if (copy_from_user(&cmd, buf, sizeof cmd))
1463                 return -EFAULT;
1464
1465         uobj = idr_write_uobj(&ib_uverbs_cq_idr, cmd.cq_handle, file->ucontext);
1466         if (!uobj)
1467                 return -EINVAL;
1468         cq      = uobj->object;
1469         ev_file = cq->cq_context;
1470         obj     = container_of(cq->uobject, struct ib_ucq_object, uobject);
1471
1472         ret = ib_destroy_cq(cq);
1473         if (!ret)
1474                 uobj->live = 0;
1475
1476         put_uobj_write(uobj);
1477
1478         if (ret)
1479                 return ret;
1480
1481         idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
1482
1483         mutex_lock(&file->mutex);
1484         list_del(&uobj->list);
1485         mutex_unlock(&file->mutex);
1486
1487         ib_uverbs_release_ucq(file, ev_file, obj);
1488
1489         memset(&resp, 0, sizeof resp);
1490         resp.comp_events_reported  = obj->comp_events_reported;
1491         resp.async_events_reported = obj->async_events_reported;
1492
1493         put_uobj(uobj);
1494
1495         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1496                          &resp, sizeof resp))
1497                 return -EFAULT;
1498
1499         return in_len;
1500 }
1501
1502 ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
1503                             const char __user *buf, int in_len,
1504                             int out_len)
1505 {
1506         struct ib_uverbs_create_qp      cmd;
1507         struct ib_uverbs_create_qp_resp resp;
1508         struct ib_udata                 udata;
1509         struct ib_uqp_object           *obj;
1510         struct ib_device               *device;
1511         struct ib_pd                   *pd = NULL;
1512         struct ib_xrcd                 *xrcd = NULL;
1513         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1514         struct ib_cq                   *scq = NULL, *rcq = NULL;
1515         struct ib_srq                  *srq = NULL;
1516         struct ib_qp                   *qp;
1517         struct ib_qp_init_attr          attr;
1518         int ret;
1519
1520         if (out_len < sizeof resp)
1521                 return -ENOSPC;
1522
1523         if (copy_from_user(&cmd, buf, sizeof cmd))
1524                 return -EFAULT;
1525
1526         if (cmd.qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
1527                 return -EPERM;
1528
1529         INIT_UDATA(&udata, buf + sizeof cmd,
1530                    (unsigned long) cmd.response + sizeof resp,
1531                    in_len - sizeof cmd, out_len - sizeof resp);
1532
1533         obj = kzalloc(sizeof *obj, GFP_KERNEL);
1534         if (!obj)
1535                 return -ENOMEM;
1536
1537         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1538         down_write(&obj->uevent.uobject.mutex);
1539
1540         if (cmd.qp_type == IB_QPT_XRC_TGT) {
1541                 xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1542                 if (!xrcd) {
1543                         ret = -EINVAL;
1544                         goto err_put;
1545                 }
1546                 device = xrcd->device;
1547         } else {
1548                 if (cmd.qp_type == IB_QPT_XRC_INI) {
1549                         cmd.max_recv_wr = cmd.max_recv_sge = 0;
1550                 } else {
1551                         if (cmd.is_srq) {
1552                                 srq = idr_read_srq(cmd.srq_handle, file->ucontext);
1553                                 if (!srq || srq->srq_type != IB_SRQT_BASIC) {
1554                                         ret = -EINVAL;
1555                                         goto err_put;
1556                                 }
1557                         }
1558
1559                         if (cmd.recv_cq_handle != cmd.send_cq_handle) {
1560                                 rcq = idr_read_cq(cmd.recv_cq_handle, file->ucontext, 0);
1561                                 if (!rcq) {
1562                                         ret = -EINVAL;
1563                                         goto err_put;
1564                                 }
1565                         }
1566                 }
1567
1568                 scq = idr_read_cq(cmd.send_cq_handle, file->ucontext, !!rcq);
1569                 rcq = rcq ?: scq;
1570                 pd  = idr_read_pd(cmd.pd_handle, file->ucontext);
1571                 if (!pd || !scq) {
1572                         ret = -EINVAL;
1573                         goto err_put;
1574                 }
1575
1576                 device = pd->device;
1577         }
1578
1579         attr.event_handler = ib_uverbs_qp_event_handler;
1580         attr.qp_context    = file;
1581         attr.send_cq       = scq;
1582         attr.recv_cq       = rcq;
1583         attr.srq           = srq;
1584         attr.xrcd          = xrcd;
1585         attr.sq_sig_type   = cmd.sq_sig_all ? IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
1586         attr.qp_type       = cmd.qp_type;
1587         attr.create_flags  = 0;
1588
1589         attr.cap.max_send_wr     = cmd.max_send_wr;
1590         attr.cap.max_recv_wr     = cmd.max_recv_wr;
1591         attr.cap.max_send_sge    = cmd.max_send_sge;
1592         attr.cap.max_recv_sge    = cmd.max_recv_sge;
1593         attr.cap.max_inline_data = cmd.max_inline_data;
1594
1595         obj->uevent.events_reported     = 0;
1596         INIT_LIST_HEAD(&obj->uevent.event_list);
1597         INIT_LIST_HEAD(&obj->mcast_list);
1598
1599         if (cmd.qp_type == IB_QPT_XRC_TGT)
1600                 qp = ib_create_qp(pd, &attr);
1601         else
1602                 qp = device->create_qp(pd, &attr, &udata);
1603
1604         if (IS_ERR(qp)) {
1605                 ret = PTR_ERR(qp);
1606                 goto err_put;
1607         }
1608
1609         if (cmd.qp_type != IB_QPT_XRC_TGT) {
1610                 qp->real_qp       = qp;
1611                 qp->device        = device;
1612                 qp->pd            = pd;
1613                 qp->send_cq       = attr.send_cq;
1614                 qp->recv_cq       = attr.recv_cq;
1615                 qp->srq           = attr.srq;
1616                 qp->event_handler = attr.event_handler;
1617                 qp->qp_context    = attr.qp_context;
1618                 qp->qp_type       = attr.qp_type;
1619                 atomic_set(&qp->usecnt, 0);
1620                 atomic_inc(&pd->usecnt);
1621                 atomic_inc(&attr.send_cq->usecnt);
1622                 if (attr.recv_cq)
1623                         atomic_inc(&attr.recv_cq->usecnt);
1624                 if (attr.srq)
1625                         atomic_inc(&attr.srq->usecnt);
1626         }
1627         qp->uobject = &obj->uevent.uobject;
1628
1629         obj->uevent.uobject.object = qp;
1630         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1631         if (ret)
1632                 goto err_destroy;
1633
1634         memset(&resp, 0, sizeof resp);
1635         resp.qpn             = qp->qp_num;
1636         resp.qp_handle       = obj->uevent.uobject.id;
1637         resp.max_recv_sge    = attr.cap.max_recv_sge;
1638         resp.max_send_sge    = attr.cap.max_send_sge;
1639         resp.max_recv_wr     = attr.cap.max_recv_wr;
1640         resp.max_send_wr     = attr.cap.max_send_wr;
1641         resp.max_inline_data = attr.cap.max_inline_data;
1642
1643         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1644                          &resp, sizeof resp)) {
1645                 ret = -EFAULT;
1646                 goto err_copy;
1647         }
1648
1649         if (xrcd) {
1650                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object,
1651                                           uobject);
1652                 atomic_inc(&obj->uxrcd->refcnt);
1653                 put_xrcd_read(xrcd_uobj);
1654         }
1655
1656         if (pd)
1657                 put_pd_read(pd);
1658         if (scq)
1659                 put_cq_read(scq);
1660         if (rcq && rcq != scq)
1661                 put_cq_read(rcq);
1662         if (srq)
1663                 put_srq_read(srq);
1664
1665         mutex_lock(&file->mutex);
1666         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1667         mutex_unlock(&file->mutex);
1668
1669         obj->uevent.uobject.live = 1;
1670
1671         up_write(&obj->uevent.uobject.mutex);
1672
1673         return in_len;
1674
1675 err_copy:
1676         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1677
1678 err_destroy:
1679         ib_destroy_qp(qp);
1680
1681 err_put:
1682         if (xrcd)
1683                 put_xrcd_read(xrcd_uobj);
1684         if (pd)
1685                 put_pd_read(pd);
1686         if (scq)
1687                 put_cq_read(scq);
1688         if (rcq && rcq != scq)
1689                 put_cq_read(rcq);
1690         if (srq)
1691                 put_srq_read(srq);
1692
1693         put_uobj_write(&obj->uevent.uobject);
1694         return ret;
1695 }
1696
1697 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1698                           const char __user *buf, int in_len, int out_len)
1699 {
1700         struct ib_uverbs_open_qp        cmd;
1701         struct ib_uverbs_create_qp_resp resp;
1702         struct ib_udata                 udata;
1703         struct ib_uqp_object           *obj;
1704         struct ib_xrcd                 *xrcd;
1705         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1706         struct ib_qp                   *qp;
1707         struct ib_qp_open_attr          attr;
1708         int ret;
1709
1710         if (out_len < sizeof resp)
1711                 return -ENOSPC;
1712
1713         if (copy_from_user(&cmd, buf, sizeof cmd))
1714                 return -EFAULT;
1715
1716         INIT_UDATA(&udata, buf + sizeof cmd,
1717                    (unsigned long) cmd.response + sizeof resp,
1718                    in_len - sizeof cmd, out_len - sizeof resp);
1719
1720         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1721         if (!obj)
1722                 return -ENOMEM;
1723
1724         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1725         down_write(&obj->uevent.uobject.mutex);
1726
1727         xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1728         if (!xrcd) {
1729                 ret = -EINVAL;
1730                 goto err_put;
1731         }
1732
1733         attr.event_handler = ib_uverbs_qp_event_handler;
1734         attr.qp_context    = file;
1735         attr.qp_num        = cmd.qpn;
1736         attr.qp_type       = cmd.qp_type;
1737
1738         obj->uevent.events_reported = 0;
1739         INIT_LIST_HEAD(&obj->uevent.event_list);
1740         INIT_LIST_HEAD(&obj->mcast_list);
1741
1742         qp = ib_open_qp(xrcd, &attr);
1743         if (IS_ERR(qp)) {
1744                 ret = PTR_ERR(qp);
1745                 goto err_put;
1746         }
1747
1748         qp->uobject = &obj->uevent.uobject;
1749
1750         obj->uevent.uobject.object = qp;
1751         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1752         if (ret)
1753                 goto err_destroy;
1754
1755         memset(&resp, 0, sizeof resp);
1756         resp.qpn       = qp->qp_num;
1757         resp.qp_handle = obj->uevent.uobject.id;
1758
1759         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1760                          &resp, sizeof resp)) {
1761                 ret = -EFAULT;
1762                 goto err_remove;
1763         }
1764
1765         obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
1766         atomic_inc(&obj->uxrcd->refcnt);
1767         put_xrcd_read(xrcd_uobj);
1768
1769         mutex_lock(&file->mutex);
1770         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1771         mutex_unlock(&file->mutex);
1772
1773         obj->uevent.uobject.live = 1;
1774
1775         up_write(&obj->uevent.uobject.mutex);
1776
1777         return in_len;
1778
1779 err_remove:
1780         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1781
1782 err_destroy:
1783         ib_destroy_qp(qp);
1784
1785 err_put:
1786         put_xrcd_read(xrcd_uobj);
1787         put_uobj_write(&obj->uevent.uobject);
1788         return ret;
1789 }
1790
1791 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1792                            const char __user *buf, int in_len,
1793                            int out_len)
1794 {
1795         struct ib_uverbs_query_qp      cmd;
1796         struct ib_uverbs_query_qp_resp resp;
1797         struct ib_qp                   *qp;
1798         struct ib_qp_attr              *attr;
1799         struct ib_qp_init_attr         *init_attr;
1800         int                            ret;
1801
1802         if (copy_from_user(&cmd, buf, sizeof cmd))
1803                 return -EFAULT;
1804
1805         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1806         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1807         if (!attr || !init_attr) {
1808                 ret = -ENOMEM;
1809                 goto out;
1810         }
1811
1812         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1813         if (!qp) {
1814                 ret = -EINVAL;
1815                 goto out;
1816         }
1817
1818         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1819
1820         put_qp_read(qp);
1821
1822         if (ret)
1823                 goto out;
1824
1825         memset(&resp, 0, sizeof resp);
1826
1827         resp.qp_state               = attr->qp_state;
1828         resp.cur_qp_state           = attr->cur_qp_state;
1829         resp.path_mtu               = attr->path_mtu;
1830         resp.path_mig_state         = attr->path_mig_state;
1831         resp.qkey                   = attr->qkey;
1832         resp.rq_psn                 = attr->rq_psn;
1833         resp.sq_psn                 = attr->sq_psn;
1834         resp.dest_qp_num            = attr->dest_qp_num;
1835         resp.qp_access_flags        = attr->qp_access_flags;
1836         resp.pkey_index             = attr->pkey_index;
1837         resp.alt_pkey_index         = attr->alt_pkey_index;
1838         resp.sq_draining            = attr->sq_draining;
1839         resp.max_rd_atomic          = attr->max_rd_atomic;
1840         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1841         resp.min_rnr_timer          = attr->min_rnr_timer;
1842         resp.port_num               = attr->port_num;
1843         resp.timeout                = attr->timeout;
1844         resp.retry_cnt              = attr->retry_cnt;
1845         resp.rnr_retry              = attr->rnr_retry;
1846         resp.alt_port_num           = attr->alt_port_num;
1847         resp.alt_timeout            = attr->alt_timeout;
1848
1849         memcpy(resp.dest.dgid, attr->ah_attr.grh.dgid.raw, 16);
1850         resp.dest.flow_label        = attr->ah_attr.grh.flow_label;
1851         resp.dest.sgid_index        = attr->ah_attr.grh.sgid_index;
1852         resp.dest.hop_limit         = attr->ah_attr.grh.hop_limit;
1853         resp.dest.traffic_class     = attr->ah_attr.grh.traffic_class;
1854         resp.dest.dlid              = attr->ah_attr.dlid;
1855         resp.dest.sl                = attr->ah_attr.sl;
1856         resp.dest.src_path_bits     = attr->ah_attr.src_path_bits;
1857         resp.dest.static_rate       = attr->ah_attr.static_rate;
1858         resp.dest.is_global         = !!(attr->ah_attr.ah_flags & IB_AH_GRH);
1859         resp.dest.port_num          = attr->ah_attr.port_num;
1860
1861         memcpy(resp.alt_dest.dgid, attr->alt_ah_attr.grh.dgid.raw, 16);
1862         resp.alt_dest.flow_label    = attr->alt_ah_attr.grh.flow_label;
1863         resp.alt_dest.sgid_index    = attr->alt_ah_attr.grh.sgid_index;
1864         resp.alt_dest.hop_limit     = attr->alt_ah_attr.grh.hop_limit;
1865         resp.alt_dest.traffic_class = attr->alt_ah_attr.grh.traffic_class;
1866         resp.alt_dest.dlid          = attr->alt_ah_attr.dlid;
1867         resp.alt_dest.sl            = attr->alt_ah_attr.sl;
1868         resp.alt_dest.src_path_bits = attr->alt_ah_attr.src_path_bits;
1869         resp.alt_dest.static_rate   = attr->alt_ah_attr.static_rate;
1870         resp.alt_dest.is_global     = !!(attr->alt_ah_attr.ah_flags & IB_AH_GRH);
1871         resp.alt_dest.port_num      = attr->alt_ah_attr.port_num;
1872
1873         resp.max_send_wr            = init_attr->cap.max_send_wr;
1874         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1875         resp.max_send_sge           = init_attr->cap.max_send_sge;
1876         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1877         resp.max_inline_data        = init_attr->cap.max_inline_data;
1878         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1879
1880         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1881                          &resp, sizeof resp))
1882                 ret = -EFAULT;
1883
1884 out:
1885         kfree(attr);
1886         kfree(init_attr);
1887
1888         return ret ? ret : in_len;
1889 }
1890
1891 /* Remove ignored fields set in the attribute mask */
1892 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1893 {
1894         switch (qp_type) {
1895         case IB_QPT_XRC_INI:
1896                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1897         case IB_QPT_XRC_TGT:
1898                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1899                                 IB_QP_RNR_RETRY);
1900         default:
1901                 return mask;
1902         }
1903 }
1904
1905 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
1906                             const char __user *buf, int in_len,
1907                             int out_len)
1908 {
1909         struct ib_uverbs_modify_qp cmd;
1910         struct ib_udata            udata;
1911         struct ib_qp              *qp;
1912         struct ib_qp_attr         *attr;
1913         int                        ret;
1914
1915         if (copy_from_user(&cmd, buf, sizeof cmd))
1916                 return -EFAULT;
1917
1918         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
1919                    out_len);
1920
1921         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1922         if (!attr)
1923                 return -ENOMEM;
1924
1925         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1926         if (!qp) {
1927                 ret = -EINVAL;
1928                 goto out;
1929         }
1930
1931         attr->qp_state            = cmd.qp_state;
1932         attr->cur_qp_state        = cmd.cur_qp_state;
1933         attr->path_mtu            = cmd.path_mtu;
1934         attr->path_mig_state      = cmd.path_mig_state;
1935         attr->qkey                = cmd.qkey;
1936         attr->rq_psn              = cmd.rq_psn;
1937         attr->sq_psn              = cmd.sq_psn;
1938         attr->dest_qp_num         = cmd.dest_qp_num;
1939         attr->qp_access_flags     = cmd.qp_access_flags;
1940         attr->pkey_index          = cmd.pkey_index;
1941         attr->alt_pkey_index      = cmd.alt_pkey_index;
1942         attr->en_sqd_async_notify = cmd.en_sqd_async_notify;
1943         attr->max_rd_atomic       = cmd.max_rd_atomic;
1944         attr->max_dest_rd_atomic  = cmd.max_dest_rd_atomic;
1945         attr->min_rnr_timer       = cmd.min_rnr_timer;
1946         attr->port_num            = cmd.port_num;
1947         attr->timeout             = cmd.timeout;
1948         attr->retry_cnt           = cmd.retry_cnt;
1949         attr->rnr_retry           = cmd.rnr_retry;
1950         attr->alt_port_num        = cmd.alt_port_num;
1951         attr->alt_timeout         = cmd.alt_timeout;
1952
1953         memcpy(attr->ah_attr.grh.dgid.raw, cmd.dest.dgid, 16);
1954         attr->ah_attr.grh.flow_label        = cmd.dest.flow_label;
1955         attr->ah_attr.grh.sgid_index        = cmd.dest.sgid_index;
1956         attr->ah_attr.grh.hop_limit         = cmd.dest.hop_limit;
1957         attr->ah_attr.grh.traffic_class     = cmd.dest.traffic_class;
1958         attr->ah_attr.dlid                  = cmd.dest.dlid;
1959         attr->ah_attr.sl                    = cmd.dest.sl;
1960         attr->ah_attr.src_path_bits         = cmd.dest.src_path_bits;
1961         attr->ah_attr.static_rate           = cmd.dest.static_rate;
1962         attr->ah_attr.ah_flags              = cmd.dest.is_global ? IB_AH_GRH : 0;
1963         attr->ah_attr.port_num              = cmd.dest.port_num;
1964
1965         memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd.alt_dest.dgid, 16);
1966         attr->alt_ah_attr.grh.flow_label    = cmd.alt_dest.flow_label;
1967         attr->alt_ah_attr.grh.sgid_index    = cmd.alt_dest.sgid_index;
1968         attr->alt_ah_attr.grh.hop_limit     = cmd.alt_dest.hop_limit;
1969         attr->alt_ah_attr.grh.traffic_class = cmd.alt_dest.traffic_class;
1970         attr->alt_ah_attr.dlid              = cmd.alt_dest.dlid;
1971         attr->alt_ah_attr.sl                = cmd.alt_dest.sl;
1972         attr->alt_ah_attr.src_path_bits     = cmd.alt_dest.src_path_bits;
1973         attr->alt_ah_attr.static_rate       = cmd.alt_dest.static_rate;
1974         attr->alt_ah_attr.ah_flags          = cmd.alt_dest.is_global ? IB_AH_GRH : 0;
1975         attr->alt_ah_attr.port_num          = cmd.alt_dest.port_num;
1976
1977         if (qp->real_qp == qp) {
1978                 ret = qp->device->modify_qp(qp, attr,
1979                         modify_qp_mask(qp->qp_type, cmd.attr_mask), &udata);
1980         } else {
1981                 ret = ib_modify_qp(qp, attr, modify_qp_mask(qp->qp_type, cmd.attr_mask));
1982         }
1983
1984         put_qp_read(qp);
1985
1986         if (ret)
1987                 goto out;
1988
1989         ret = in_len;
1990
1991 out:
1992         kfree(attr);
1993
1994         return ret;
1995 }
1996
1997 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
1998                              const char __user *buf, int in_len,
1999                              int out_len)
2000 {
2001         struct ib_uverbs_destroy_qp      cmd;
2002         struct ib_uverbs_destroy_qp_resp resp;
2003         struct ib_uobject               *uobj;
2004         struct ib_qp                    *qp;
2005         struct ib_uqp_object            *obj;
2006         int                              ret = -EINVAL;
2007
2008         if (copy_from_user(&cmd, buf, sizeof cmd))
2009                 return -EFAULT;
2010
2011         memset(&resp, 0, sizeof resp);
2012
2013         uobj = idr_write_uobj(&ib_uverbs_qp_idr, cmd.qp_handle, file->ucontext);
2014         if (!uobj)
2015                 return -EINVAL;
2016         qp  = uobj->object;
2017         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2018
2019         if (!list_empty(&obj->mcast_list)) {
2020                 put_uobj_write(uobj);
2021                 return -EBUSY;
2022         }
2023
2024         ret = ib_destroy_qp(qp);
2025         if (!ret)
2026                 uobj->live = 0;
2027
2028         put_uobj_write(uobj);
2029
2030         if (ret)
2031                 return ret;
2032
2033         if (obj->uxrcd)
2034                 atomic_dec(&obj->uxrcd->refcnt);
2035
2036         idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
2037
2038         mutex_lock(&file->mutex);
2039         list_del(&uobj->list);
2040         mutex_unlock(&file->mutex);
2041
2042         ib_uverbs_release_uevent(file, &obj->uevent);
2043
2044         resp.events_reported = obj->uevent.events_reported;
2045
2046         put_uobj(uobj);
2047
2048         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2049                          &resp, sizeof resp))
2050                 return -EFAULT;
2051
2052         return in_len;
2053 }
2054
2055 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2056                             const char __user *buf, int in_len,
2057                             int out_len)
2058 {
2059         struct ib_uverbs_post_send      cmd;
2060         struct ib_uverbs_post_send_resp resp;
2061         struct ib_uverbs_send_wr       *user_wr;
2062         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
2063         struct ib_qp                   *qp;
2064         int                             i, sg_ind;
2065         int                             is_ud;
2066         ssize_t                         ret = -EINVAL;
2067
2068         if (copy_from_user(&cmd, buf, sizeof cmd))
2069                 return -EFAULT;
2070
2071         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2072             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2073                 return -EINVAL;
2074
2075         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2076                 return -EINVAL;
2077
2078         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2079         if (!user_wr)
2080                 return -ENOMEM;
2081
2082         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2083         if (!qp)
2084                 goto out;
2085
2086         is_ud = qp->qp_type == IB_QPT_UD;
2087         sg_ind = 0;
2088         last = NULL;
2089         for (i = 0; i < cmd.wr_count; ++i) {
2090                 if (copy_from_user(user_wr,
2091                                    buf + sizeof cmd + i * cmd.wqe_size,
2092                                    cmd.wqe_size)) {
2093                         ret = -EFAULT;
2094                         goto out_put;
2095                 }
2096
2097                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2098                         ret = -EINVAL;
2099                         goto out_put;
2100                 }
2101
2102                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2103                                user_wr->num_sge * sizeof (struct ib_sge),
2104                                GFP_KERNEL);
2105                 if (!next) {
2106                         ret = -ENOMEM;
2107                         goto out_put;
2108                 }
2109
2110                 if (!last)
2111                         wr = next;
2112                 else
2113                         last->next = next;
2114                 last = next;
2115
2116                 next->next       = NULL;
2117                 next->wr_id      = user_wr->wr_id;
2118                 next->num_sge    = user_wr->num_sge;
2119                 next->opcode     = user_wr->opcode;
2120                 next->send_flags = user_wr->send_flags;
2121
2122                 if (is_ud) {
2123                         next->wr.ud.ah = idr_read_ah(user_wr->wr.ud.ah,
2124                                                      file->ucontext);
2125                         if (!next->wr.ud.ah) {
2126                                 ret = -EINVAL;
2127                                 goto out_put;
2128                         }
2129                         next->wr.ud.remote_qpn  = user_wr->wr.ud.remote_qpn;
2130                         next->wr.ud.remote_qkey = user_wr->wr.ud.remote_qkey;
2131                 } else {
2132                         switch (next->opcode) {
2133                         case IB_WR_RDMA_WRITE_WITH_IMM:
2134                                 next->ex.imm_data =
2135                                         (__be32 __force) user_wr->ex.imm_data;
2136                         case IB_WR_RDMA_WRITE:
2137                         case IB_WR_RDMA_READ:
2138                                 next->wr.rdma.remote_addr =
2139                                         user_wr->wr.rdma.remote_addr;
2140                                 next->wr.rdma.rkey        =
2141                                         user_wr->wr.rdma.rkey;
2142                                 break;
2143                         case IB_WR_SEND_WITH_IMM:
2144                                 next->ex.imm_data =
2145                                         (__be32 __force) user_wr->ex.imm_data;
2146                                 break;
2147                         case IB_WR_SEND_WITH_INV:
2148                                 next->ex.invalidate_rkey =
2149                                         user_wr->ex.invalidate_rkey;
2150                                 break;
2151                         case IB_WR_ATOMIC_CMP_AND_SWP:
2152                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2153                                 next->wr.atomic.remote_addr =
2154                                         user_wr->wr.atomic.remote_addr;
2155                                 next->wr.atomic.compare_add =
2156                                         user_wr->wr.atomic.compare_add;
2157                                 next->wr.atomic.swap = user_wr->wr.atomic.swap;
2158                                 next->wr.atomic.rkey = user_wr->wr.atomic.rkey;
2159                                 break;
2160                         default:
2161                                 break;
2162                         }
2163                 }
2164
2165                 if (next->num_sge) {
2166                         next->sg_list = (void *) next +
2167                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2168                         if (copy_from_user(next->sg_list,
2169                                            buf + sizeof cmd +
2170                                            cmd.wr_count * cmd.wqe_size +
2171                                            sg_ind * sizeof (struct ib_sge),
2172                                            next->num_sge * sizeof (struct ib_sge))) {
2173                                 ret = -EFAULT;
2174                                 goto out_put;
2175                         }
2176                         sg_ind += next->num_sge;
2177                 } else
2178                         next->sg_list = NULL;
2179         }
2180
2181         resp.bad_wr = 0;
2182         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2183         if (ret)
2184                 for (next = wr; next; next = next->next) {
2185                         ++resp.bad_wr;
2186                         if (next == bad_wr)
2187                                 break;
2188                 }
2189
2190         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2191                          &resp, sizeof resp))
2192                 ret = -EFAULT;
2193
2194 out_put:
2195         put_qp_read(qp);
2196
2197         while (wr) {
2198                 if (is_ud && wr->wr.ud.ah)
2199                         put_ah_read(wr->wr.ud.ah);
2200                 next = wr->next;
2201                 kfree(wr);
2202                 wr = next;
2203         }
2204
2205 out:
2206         kfree(user_wr);
2207
2208         return ret ? ret : in_len;
2209 }
2210
2211 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2212                                                     int in_len,
2213                                                     u32 wr_count,
2214                                                     u32 sge_count,
2215                                                     u32 wqe_size)
2216 {
2217         struct ib_uverbs_recv_wr *user_wr;
2218         struct ib_recv_wr        *wr = NULL, *last, *next;
2219         int                       sg_ind;
2220         int                       i;
2221         int                       ret;
2222
2223         if (in_len < wqe_size * wr_count +
2224             sge_count * sizeof (struct ib_uverbs_sge))
2225                 return ERR_PTR(-EINVAL);
2226
2227         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2228                 return ERR_PTR(-EINVAL);
2229
2230         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2231         if (!user_wr)
2232                 return ERR_PTR(-ENOMEM);
2233
2234         sg_ind = 0;
2235         last = NULL;
2236         for (i = 0; i < wr_count; ++i) {
2237                 if (copy_from_user(user_wr, buf + i * wqe_size,
2238                                    wqe_size)) {
2239                         ret = -EFAULT;
2240                         goto err;
2241                 }
2242
2243                 if (user_wr->num_sge + sg_ind > sge_count) {
2244                         ret = -EINVAL;
2245                         goto err;
2246                 }
2247
2248                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2249                                user_wr->num_sge * sizeof (struct ib_sge),
2250                                GFP_KERNEL);
2251                 if (!next) {
2252                         ret = -ENOMEM;
2253                         goto err;
2254                 }
2255
2256                 if (!last)
2257                         wr = next;
2258                 else
2259                         last->next = next;
2260                 last = next;
2261
2262                 next->next       = NULL;
2263                 next->wr_id      = user_wr->wr_id;
2264                 next->num_sge    = user_wr->num_sge;
2265
2266                 if (next->num_sge) {
2267                         next->sg_list = (void *) next +
2268                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2269                         if (copy_from_user(next->sg_list,
2270                                            buf + wr_count * wqe_size +
2271                                            sg_ind * sizeof (struct ib_sge),
2272                                            next->num_sge * sizeof (struct ib_sge))) {
2273                                 ret = -EFAULT;
2274                                 goto err;
2275                         }
2276                         sg_ind += next->num_sge;
2277                 } else
2278                         next->sg_list = NULL;
2279         }
2280
2281         kfree(user_wr);
2282         return wr;
2283
2284 err:
2285         kfree(user_wr);
2286
2287         while (wr) {
2288                 next = wr->next;
2289                 kfree(wr);
2290                 wr = next;
2291         }
2292
2293         return ERR_PTR(ret);
2294 }
2295
2296 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2297                             const char __user *buf, int in_len,
2298                             int out_len)
2299 {
2300         struct ib_uverbs_post_recv      cmd;
2301         struct ib_uverbs_post_recv_resp resp;
2302         struct ib_recv_wr              *wr, *next, *bad_wr;
2303         struct ib_qp                   *qp;
2304         ssize_t                         ret = -EINVAL;
2305
2306         if (copy_from_user(&cmd, buf, sizeof cmd))
2307                 return -EFAULT;
2308
2309         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2310                                        in_len - sizeof cmd, cmd.wr_count,
2311                                        cmd.sge_count, cmd.wqe_size);
2312         if (IS_ERR(wr))
2313                 return PTR_ERR(wr);
2314
2315         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2316         if (!qp)
2317                 goto out;
2318
2319         resp.bad_wr = 0;
2320         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2321
2322         put_qp_read(qp);
2323
2324         if (ret)
2325                 for (next = wr; next; next = next->next) {
2326                         ++resp.bad_wr;
2327                         if (next == bad_wr)
2328                                 break;
2329                 }
2330
2331         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2332                          &resp, sizeof resp))
2333                 ret = -EFAULT;
2334
2335 out:
2336         while (wr) {
2337                 next = wr->next;
2338                 kfree(wr);
2339                 wr = next;
2340         }
2341
2342         return ret ? ret : in_len;
2343 }
2344
2345 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2346                                 const char __user *buf, int in_len,
2347                                 int out_len)
2348 {
2349         struct ib_uverbs_post_srq_recv      cmd;
2350         struct ib_uverbs_post_srq_recv_resp resp;
2351         struct ib_recv_wr                  *wr, *next, *bad_wr;
2352         struct ib_srq                      *srq;
2353         ssize_t                             ret = -EINVAL;
2354
2355         if (copy_from_user(&cmd, buf, sizeof cmd))
2356                 return -EFAULT;
2357
2358         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2359                                        in_len - sizeof cmd, cmd.wr_count,
2360                                        cmd.sge_count, cmd.wqe_size);
2361         if (IS_ERR(wr))
2362                 return PTR_ERR(wr);
2363
2364         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2365         if (!srq)
2366                 goto out;
2367
2368         resp.bad_wr = 0;
2369         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2370
2371         put_srq_read(srq);
2372
2373         if (ret)
2374                 for (next = wr; next; next = next->next) {
2375                         ++resp.bad_wr;
2376                         if (next == bad_wr)
2377                                 break;
2378                 }
2379
2380         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2381                          &resp, sizeof resp))
2382                 ret = -EFAULT;
2383
2384 out:
2385         while (wr) {
2386                 next = wr->next;
2387                 kfree(wr);
2388                 wr = next;
2389         }
2390
2391         return ret ? ret : in_len;
2392 }
2393
2394 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2395                             const char __user *buf, int in_len,
2396                             int out_len)
2397 {
2398         struct ib_uverbs_create_ah       cmd;
2399         struct ib_uverbs_create_ah_resp  resp;
2400         struct ib_uobject               *uobj;
2401         struct ib_pd                    *pd;
2402         struct ib_ah                    *ah;
2403         struct ib_ah_attr               attr;
2404         int ret;
2405
2406         if (out_len < sizeof resp)
2407                 return -ENOSPC;
2408
2409         if (copy_from_user(&cmd, buf, sizeof cmd))
2410                 return -EFAULT;
2411
2412         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
2413         if (!uobj)
2414                 return -ENOMEM;
2415
2416         init_uobj(uobj, cmd.user_handle, file->ucontext, &ah_lock_class);
2417         down_write(&uobj->mutex);
2418
2419         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
2420         if (!pd) {
2421                 ret = -EINVAL;
2422                 goto err;
2423         }
2424
2425         attr.dlid              = cmd.attr.dlid;
2426         attr.sl                = cmd.attr.sl;
2427         attr.src_path_bits     = cmd.attr.src_path_bits;
2428         attr.static_rate       = cmd.attr.static_rate;
2429         attr.ah_flags          = cmd.attr.is_global ? IB_AH_GRH : 0;
2430         attr.port_num          = cmd.attr.port_num;
2431         attr.grh.flow_label    = cmd.attr.grh.flow_label;
2432         attr.grh.sgid_index    = cmd.attr.grh.sgid_index;
2433         attr.grh.hop_limit     = cmd.attr.grh.hop_limit;
2434         attr.grh.traffic_class = cmd.attr.grh.traffic_class;
2435         memcpy(attr.grh.dgid.raw, cmd.attr.grh.dgid, 16);
2436
2437         ah = ib_create_ah(pd, &attr);
2438         if (IS_ERR(ah)) {
2439                 ret = PTR_ERR(ah);
2440                 goto err_put;
2441         }
2442
2443         ah->uobject  = uobj;
2444         uobj->object = ah;
2445
2446         ret = idr_add_uobj(&ib_uverbs_ah_idr, uobj);
2447         if (ret)
2448                 goto err_destroy;
2449
2450         resp.ah_handle = uobj->id;
2451
2452         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2453                          &resp, sizeof resp)) {
2454                 ret = -EFAULT;
2455                 goto err_copy;
2456         }
2457
2458         put_pd_read(pd);
2459
2460         mutex_lock(&file->mutex);
2461         list_add_tail(&uobj->list, &file->ucontext->ah_list);
2462         mutex_unlock(&file->mutex);
2463
2464         uobj->live = 1;
2465
2466         up_write(&uobj->mutex);
2467
2468         return in_len;
2469
2470 err_copy:
2471         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2472
2473 err_destroy:
2474         ib_destroy_ah(ah);
2475
2476 err_put:
2477         put_pd_read(pd);
2478
2479 err:
2480         put_uobj_write(uobj);
2481         return ret;
2482 }
2483
2484 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2485                              const char __user *buf, int in_len, int out_len)
2486 {
2487         struct ib_uverbs_destroy_ah cmd;
2488         struct ib_ah               *ah;
2489         struct ib_uobject          *uobj;
2490         int                         ret;
2491
2492         if (copy_from_user(&cmd, buf, sizeof cmd))
2493                 return -EFAULT;
2494
2495         uobj = idr_write_uobj(&ib_uverbs_ah_idr, cmd.ah_handle, file->ucontext);
2496         if (!uobj)
2497                 return -EINVAL;
2498         ah = uobj->object;
2499
2500         ret = ib_destroy_ah(ah);
2501         if (!ret)
2502                 uobj->live = 0;
2503
2504         put_uobj_write(uobj);
2505
2506         if (ret)
2507                 return ret;
2508
2509         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2510
2511         mutex_lock(&file->mutex);
2512         list_del(&uobj->list);
2513         mutex_unlock(&file->mutex);
2514
2515         put_uobj(uobj);
2516
2517         return in_len;
2518 }
2519
2520 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2521                                const char __user *buf, int in_len,
2522                                int out_len)
2523 {
2524         struct ib_uverbs_attach_mcast cmd;
2525         struct ib_qp                 *qp;
2526         struct ib_uqp_object         *obj;
2527         struct ib_uverbs_mcast_entry *mcast;
2528         int                           ret;
2529
2530         if (copy_from_user(&cmd, buf, sizeof cmd))
2531                 return -EFAULT;
2532
2533         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2534         if (!qp)
2535                 return -EINVAL;
2536
2537         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2538
2539         list_for_each_entry(mcast, &obj->mcast_list, list)
2540                 if (cmd.mlid == mcast->lid &&
2541                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2542                         ret = 0;
2543                         goto out_put;
2544                 }
2545
2546         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2547         if (!mcast) {
2548                 ret = -ENOMEM;
2549                 goto out_put;
2550         }
2551
2552         mcast->lid = cmd.mlid;
2553         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2554
2555         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2556         if (!ret)
2557                 list_add_tail(&mcast->list, &obj->mcast_list);
2558         else
2559                 kfree(mcast);
2560
2561 out_put:
2562         put_qp_write(qp);
2563
2564         return ret ? ret : in_len;
2565 }
2566
2567 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2568                                const char __user *buf, int in_len,
2569                                int out_len)
2570 {
2571         struct ib_uverbs_detach_mcast cmd;
2572         struct ib_uqp_object         *obj;
2573         struct ib_qp                 *qp;
2574         struct ib_uverbs_mcast_entry *mcast;
2575         int                           ret = -EINVAL;
2576
2577         if (copy_from_user(&cmd, buf, sizeof cmd))
2578                 return -EFAULT;
2579
2580         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2581         if (!qp)
2582                 return -EINVAL;
2583
2584         ret = ib_detach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
2585         if (ret)
2586                 goto out_put;
2587
2588         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2589
2590         list_for_each_entry(mcast, &obj->mcast_list, list)
2591                 if (cmd.mlid == mcast->lid &&
2592                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2593                         list_del(&mcast->list);
2594                         kfree(mcast);
2595                         break;
2596                 }
2597
2598 out_put:
2599         put_qp_write(qp);
2600
2601         return ret ? ret : in_len;
2602 }
2603
2604 #ifdef CONFIG_INFINIBAND_EXPERIMENTAL_UVERBS_FLOW_STEERING
2605 static int kern_spec_to_ib_spec(struct ib_kern_spec *kern_spec,
2606                                 union ib_flow_spec *ib_spec)
2607 {
2608         ib_spec->type = kern_spec->type;
2609
2610         switch (ib_spec->type) {
2611         case IB_FLOW_SPEC_ETH:
2612                 ib_spec->eth.size = sizeof(struct ib_flow_spec_eth);
2613                 if (ib_spec->eth.size != kern_spec->eth.size)
2614                         return -EINVAL;
2615                 memcpy(&ib_spec->eth.val, &kern_spec->eth.val,
2616                        sizeof(struct ib_flow_eth_filter));
2617                 memcpy(&ib_spec->eth.mask, &kern_spec->eth.mask,
2618                        sizeof(struct ib_flow_eth_filter));
2619                 break;
2620         case IB_FLOW_SPEC_IPV4:
2621                 ib_spec->ipv4.size = sizeof(struct ib_flow_spec_ipv4);
2622                 if (ib_spec->ipv4.size != kern_spec->ipv4.size)
2623                         return -EINVAL;
2624                 memcpy(&ib_spec->ipv4.val, &kern_spec->ipv4.val,
2625                        sizeof(struct ib_flow_ipv4_filter));
2626                 memcpy(&ib_spec->ipv4.mask, &kern_spec->ipv4.mask,
2627                        sizeof(struct ib_flow_ipv4_filter));
2628                 break;
2629         case IB_FLOW_SPEC_TCP:
2630         case IB_FLOW_SPEC_UDP:
2631                 ib_spec->tcp_udp.size = sizeof(struct ib_flow_spec_tcp_udp);
2632                 if (ib_spec->tcp_udp.size != kern_spec->tcp_udp.size)
2633                         return -EINVAL;
2634                 memcpy(&ib_spec->tcp_udp.val, &kern_spec->tcp_udp.val,
2635                        sizeof(struct ib_flow_tcp_udp_filter));
2636                 memcpy(&ib_spec->tcp_udp.mask, &kern_spec->tcp_udp.mask,
2637                        sizeof(struct ib_flow_tcp_udp_filter));
2638                 break;
2639         default:
2640                 return -EINVAL;
2641         }
2642         return 0;
2643 }
2644
2645 ssize_t ib_uverbs_create_flow(struct ib_uverbs_file *file,
2646                               const char __user *buf, int in_len,
2647                               int out_len)
2648 {
2649         struct ib_uverbs_create_flow      cmd;
2650         struct ib_uverbs_create_flow_resp resp;
2651         struct ib_uobject                 *uobj;
2652         struct ib_flow                    *flow_id;
2653         struct ib_kern_flow_attr          *kern_flow_attr;
2654         struct ib_flow_attr               *flow_attr;
2655         struct ib_qp                      *qp;
2656         int err = 0;
2657         void *kern_spec;
2658         void *ib_spec;
2659         int i;
2660         int kern_attr_size;
2661
2662         if (out_len < sizeof(resp))
2663                 return -ENOSPC;
2664
2665         if (copy_from_user(&cmd, buf, sizeof(cmd)))
2666                 return -EFAULT;
2667
2668         if (cmd.comp_mask)
2669                 return -EINVAL;
2670
2671         if ((cmd.flow_attr.type == IB_FLOW_ATTR_SNIFFER &&
2672              !capable(CAP_NET_ADMIN)) || !capable(CAP_NET_RAW))
2673                 return -EPERM;
2674
2675         if (cmd.flow_attr.num_of_specs < 0 ||
2676             cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
2677                 return -EINVAL;
2678
2679         kern_attr_size = cmd.flow_attr.size - sizeof(cmd) -
2680                          sizeof(struct ib_uverbs_cmd_hdr_ex);
2681
2682         if (cmd.flow_attr.size < 0 || cmd.flow_attr.size > in_len ||
2683             kern_attr_size < 0 || kern_attr_size >
2684             (cmd.flow_attr.num_of_specs * sizeof(struct ib_kern_spec)))
2685                 return -EINVAL;
2686
2687         if (cmd.flow_attr.num_of_specs) {
2688                 kern_flow_attr = kmalloc(cmd.flow_attr.size, GFP_KERNEL);
2689                 if (!kern_flow_attr)
2690                         return -ENOMEM;
2691
2692                 memcpy(kern_flow_attr, &cmd.flow_attr, sizeof(*kern_flow_attr));
2693                 if (copy_from_user(kern_flow_attr + 1, buf + sizeof(cmd),
2694                                    kern_attr_size)) {
2695                         err = -EFAULT;
2696                         goto err_free_attr;
2697                 }
2698         } else {
2699                 kern_flow_attr = &cmd.flow_attr;
2700                 kern_attr_size = sizeof(cmd.flow_attr);
2701         }
2702
2703         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
2704         if (!uobj) {
2705                 err = -ENOMEM;
2706                 goto err_free_attr;
2707         }
2708         init_uobj(uobj, 0, file->ucontext, &rule_lock_class);
2709         down_write(&uobj->mutex);
2710
2711         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2712         if (!qp) {
2713                 err = -EINVAL;
2714                 goto err_uobj;
2715         }
2716
2717         flow_attr = kmalloc(cmd.flow_attr.size, GFP_KERNEL);
2718         if (!flow_attr) {
2719                 err = -ENOMEM;
2720                 goto err_put;
2721         }
2722
2723         flow_attr->type = kern_flow_attr->type;
2724         flow_attr->priority = kern_flow_attr->priority;
2725         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
2726         flow_attr->port = kern_flow_attr->port;
2727         flow_attr->flags = kern_flow_attr->flags;
2728         flow_attr->size = sizeof(*flow_attr);
2729
2730         kern_spec = kern_flow_attr + 1;
2731         ib_spec = flow_attr + 1;
2732         for (i = 0; i < flow_attr->num_of_specs && kern_attr_size > 0; i++) {
2733                 err = kern_spec_to_ib_spec(kern_spec, ib_spec);
2734                 if (err)
2735                         goto err_free;
2736                 flow_attr->size +=
2737                         ((union ib_flow_spec *) ib_spec)->size;
2738                 kern_attr_size -= ((struct ib_kern_spec *) kern_spec)->size;
2739                 kern_spec += ((struct ib_kern_spec *) kern_spec)->size;
2740                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
2741         }
2742         if (kern_attr_size) {
2743                 pr_warn("create flow failed, %d bytes left from uverb cmd\n",
2744                         kern_attr_size);
2745                 goto err_free;
2746         }
2747         flow_id = ib_create_flow(qp, flow_attr, IB_FLOW_DOMAIN_USER);
2748         if (IS_ERR(flow_id)) {
2749                 err = PTR_ERR(flow_id);
2750                 goto err_free;
2751         }
2752         flow_id->qp = qp;
2753         flow_id->uobject = uobj;
2754         uobj->object = flow_id;
2755
2756         err = idr_add_uobj(&ib_uverbs_rule_idr, uobj);
2757         if (err)
2758                 goto destroy_flow;
2759
2760         memset(&resp, 0, sizeof(resp));
2761         resp.flow_handle = uobj->id;
2762
2763         if (copy_to_user((void __user *)(unsigned long) cmd.response,
2764                          &resp, sizeof(resp))) {
2765                 err = -EFAULT;
2766                 goto err_copy;
2767         }
2768
2769         put_qp_read(qp);
2770         mutex_lock(&file->mutex);
2771         list_add_tail(&uobj->list, &file->ucontext->rule_list);
2772         mutex_unlock(&file->mutex);
2773
2774         uobj->live = 1;
2775
2776         up_write(&uobj->mutex);
2777         kfree(flow_attr);
2778         if (cmd.flow_attr.num_of_specs)
2779                 kfree(kern_flow_attr);
2780         return in_len;
2781 err_copy:
2782         idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
2783 destroy_flow:
2784         ib_destroy_flow(flow_id);
2785 err_free:
2786         kfree(flow_attr);
2787 err_put:
2788         put_qp_read(qp);
2789 err_uobj:
2790         put_uobj_write(uobj);
2791 err_free_attr:
2792         if (cmd.flow_attr.num_of_specs)
2793                 kfree(kern_flow_attr);
2794         return err;
2795 }
2796
2797 ssize_t ib_uverbs_destroy_flow(struct ib_uverbs_file *file,
2798                                const char __user *buf, int in_len,
2799                                int out_len) {
2800         struct ib_uverbs_destroy_flow   cmd;
2801         struct ib_flow                  *flow_id;
2802         struct ib_uobject               *uobj;
2803         int                             ret;
2804
2805         if (copy_from_user(&cmd, buf, sizeof(cmd)))
2806                 return -EFAULT;
2807
2808         uobj = idr_write_uobj(&ib_uverbs_rule_idr, cmd.flow_handle,
2809                               file->ucontext);
2810         if (!uobj)
2811                 return -EINVAL;
2812         flow_id = uobj->object;
2813
2814         ret = ib_destroy_flow(flow_id);
2815         if (!ret)
2816                 uobj->live = 0;
2817
2818         put_uobj_write(uobj);
2819
2820         idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
2821
2822         mutex_lock(&file->mutex);
2823         list_del(&uobj->list);
2824         mutex_unlock(&file->mutex);
2825
2826         put_uobj(uobj);
2827
2828         return ret ? ret : in_len;
2829 }
2830 #endif /* CONFIG_INFINIBAND_EXPERIMENTAL_UVERBS_FLOW_STEERING */
2831
2832 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
2833                                 struct ib_uverbs_create_xsrq *cmd,
2834                                 struct ib_udata *udata)
2835 {
2836         struct ib_uverbs_create_srq_resp resp;
2837         struct ib_usrq_object           *obj;
2838         struct ib_pd                    *pd;
2839         struct ib_srq                   *srq;
2840         struct ib_uobject               *uninitialized_var(xrcd_uobj);
2841         struct ib_srq_init_attr          attr;
2842         int ret;
2843
2844         obj = kmalloc(sizeof *obj, GFP_KERNEL);
2845         if (!obj)
2846                 return -ENOMEM;
2847
2848         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext, &srq_lock_class);
2849         down_write(&obj->uevent.uobject.mutex);
2850
2851         if (cmd->srq_type == IB_SRQT_XRC) {
2852                 attr.ext.xrc.xrcd  = idr_read_xrcd(cmd->xrcd_handle, file->ucontext, &xrcd_uobj);
2853                 if (!attr.ext.xrc.xrcd) {
2854                         ret = -EINVAL;
2855                         goto err;
2856                 }
2857
2858                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
2859                 atomic_inc(&obj->uxrcd->refcnt);
2860
2861                 attr.ext.xrc.cq  = idr_read_cq(cmd->cq_handle, file->ucontext, 0);
2862                 if (!attr.ext.xrc.cq) {
2863                         ret = -EINVAL;
2864                         goto err_put_xrcd;
2865                 }
2866         }
2867
2868         pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
2869         if (!pd) {
2870                 ret = -EINVAL;
2871                 goto err_put_cq;
2872         }
2873
2874         attr.event_handler  = ib_uverbs_srq_event_handler;
2875         attr.srq_context    = file;
2876         attr.srq_type       = cmd->srq_type;
2877         attr.attr.max_wr    = cmd->max_wr;
2878         attr.attr.max_sge   = cmd->max_sge;
2879         attr.attr.srq_limit = cmd->srq_limit;
2880
2881         obj->uevent.events_reported = 0;
2882         INIT_LIST_HEAD(&obj->uevent.event_list);
2883
2884         srq = pd->device->create_srq(pd, &attr, udata);
2885         if (IS_ERR(srq)) {
2886                 ret = PTR_ERR(srq);
2887                 goto err_put;
2888         }
2889
2890         srq->device        = pd->device;
2891         srq->pd            = pd;
2892         srq->srq_type      = cmd->srq_type;
2893         srq->uobject       = &obj->uevent.uobject;
2894         srq->event_handler = attr.event_handler;
2895         srq->srq_context   = attr.srq_context;
2896
2897         if (cmd->srq_type == IB_SRQT_XRC) {
2898                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
2899                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
2900                 atomic_inc(&attr.ext.xrc.cq->usecnt);
2901                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
2902         }
2903
2904         atomic_inc(&pd->usecnt);
2905         atomic_set(&srq->usecnt, 0);
2906
2907         obj->uevent.uobject.object = srq;
2908         ret = idr_add_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2909         if (ret)
2910                 goto err_destroy;
2911
2912         memset(&resp, 0, sizeof resp);
2913         resp.srq_handle = obj->uevent.uobject.id;
2914         resp.max_wr     = attr.attr.max_wr;
2915         resp.max_sge    = attr.attr.max_sge;
2916         if (cmd->srq_type == IB_SRQT_XRC)
2917                 resp.srqn = srq->ext.xrc.srq_num;
2918
2919         if (copy_to_user((void __user *) (unsigned long) cmd->response,
2920                          &resp, sizeof resp)) {
2921                 ret = -EFAULT;
2922                 goto err_copy;
2923         }
2924
2925         if (cmd->srq_type == IB_SRQT_XRC) {
2926                 put_uobj_read(xrcd_uobj);
2927                 put_cq_read(attr.ext.xrc.cq);
2928         }
2929         put_pd_read(pd);
2930
2931         mutex_lock(&file->mutex);
2932         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->srq_list);
2933         mutex_unlock(&file->mutex);
2934
2935         obj->uevent.uobject.live = 1;
2936
2937         up_write(&obj->uevent.uobject.mutex);
2938
2939         return 0;
2940
2941 err_copy:
2942         idr_remove_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2943
2944 err_destroy:
2945         ib_destroy_srq(srq);
2946
2947 err_put:
2948         put_pd_read(pd);
2949
2950 err_put_cq:
2951         if (cmd->srq_type == IB_SRQT_XRC)
2952                 put_cq_read(attr.ext.xrc.cq);
2953
2954 err_put_xrcd:
2955         if (cmd->srq_type == IB_SRQT_XRC) {
2956                 atomic_dec(&obj->uxrcd->refcnt);
2957                 put_uobj_read(xrcd_uobj);
2958         }
2959
2960 err:
2961         put_uobj_write(&obj->uevent.uobject);
2962         return ret;
2963 }
2964
2965 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
2966                              const char __user *buf, int in_len,
2967                              int out_len)
2968 {
2969         struct ib_uverbs_create_srq      cmd;
2970         struct ib_uverbs_create_xsrq     xcmd;
2971         struct ib_uverbs_create_srq_resp resp;
2972         struct ib_udata                  udata;
2973         int ret;
2974
2975         if (out_len < sizeof resp)
2976                 return -ENOSPC;
2977
2978         if (copy_from_user(&cmd, buf, sizeof cmd))
2979                 return -EFAULT;
2980
2981         xcmd.response    = cmd.response;
2982         xcmd.user_handle = cmd.user_handle;
2983         xcmd.srq_type    = IB_SRQT_BASIC;
2984         xcmd.pd_handle   = cmd.pd_handle;
2985         xcmd.max_wr      = cmd.max_wr;
2986         xcmd.max_sge     = cmd.max_sge;
2987         xcmd.srq_limit   = cmd.srq_limit;
2988
2989         INIT_UDATA(&udata, buf + sizeof cmd,
2990                    (unsigned long) cmd.response + sizeof resp,
2991                    in_len - sizeof cmd, out_len - sizeof resp);
2992
2993         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
2994         if (ret)
2995                 return ret;
2996
2997         return in_len;
2998 }
2999
3000 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
3001                               const char __user *buf, int in_len, int out_len)
3002 {
3003         struct ib_uverbs_create_xsrq     cmd;
3004         struct ib_uverbs_create_srq_resp resp;
3005         struct ib_udata                  udata;
3006         int ret;
3007
3008         if (out_len < sizeof resp)
3009                 return -ENOSPC;
3010
3011         if (copy_from_user(&cmd, buf, sizeof cmd))
3012                 return -EFAULT;
3013
3014         INIT_UDATA(&udata, buf + sizeof cmd,
3015                    (unsigned long) cmd.response + sizeof resp,
3016                    in_len - sizeof cmd, out_len - sizeof resp);
3017
3018         ret = __uverbs_create_xsrq(file, &cmd, &udata);
3019         if (ret)
3020                 return ret;
3021
3022         return in_len;
3023 }
3024
3025 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
3026                              const char __user *buf, int in_len,
3027                              int out_len)
3028 {
3029         struct ib_uverbs_modify_srq cmd;
3030         struct ib_udata             udata;
3031         struct ib_srq              *srq;
3032         struct ib_srq_attr          attr;
3033         int                         ret;
3034
3035         if (copy_from_user(&cmd, buf, sizeof cmd))
3036                 return -EFAULT;
3037
3038         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
3039                    out_len);
3040
3041         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
3042         if (!srq)
3043                 return -EINVAL;
3044
3045         attr.max_wr    = cmd.max_wr;
3046         attr.srq_limit = cmd.srq_limit;
3047
3048         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
3049
3050         put_srq_read(srq);
3051
3052         return ret ? ret : in_len;
3053 }
3054
3055 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
3056                             const char __user *buf,
3057                             int in_len, int out_len)
3058 {
3059         struct ib_uverbs_query_srq      cmd;
3060         struct ib_uverbs_query_srq_resp resp;
3061         struct ib_srq_attr              attr;
3062         struct ib_srq                   *srq;
3063         int                             ret;
3064
3065         if (out_len < sizeof resp)
3066                 return -ENOSPC;
3067
3068         if (copy_from_user(&cmd, buf, sizeof cmd))
3069                 return -EFAULT;
3070
3071         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
3072         if (!srq)
3073                 return -EINVAL;
3074
3075         ret = ib_query_srq(srq, &attr);
3076
3077         put_srq_read(srq);
3078
3079         if (ret)
3080                 return ret;
3081
3082         memset(&resp, 0, sizeof resp);
3083
3084         resp.max_wr    = attr.max_wr;
3085         resp.max_sge   = attr.max_sge;
3086         resp.srq_limit = attr.srq_limit;
3087
3088         if (copy_to_user((void __user *) (unsigned long) cmd.response,
3089                          &resp, sizeof resp))
3090                 return -EFAULT;
3091
3092         return in_len;
3093 }
3094
3095 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
3096                               const char __user *buf, int in_len,
3097                               int out_len)
3098 {
3099         struct ib_uverbs_destroy_srq      cmd;
3100         struct ib_uverbs_destroy_srq_resp resp;
3101         struct ib_uobject                *uobj;
3102         struct ib_srq                    *srq;
3103         struct ib_uevent_object          *obj;
3104         int                               ret = -EINVAL;
3105         struct ib_usrq_object            *us;
3106         enum ib_srq_type                  srq_type;
3107
3108         if (copy_from_user(&cmd, buf, sizeof cmd))
3109                 return -EFAULT;
3110
3111         uobj = idr_write_uobj(&ib_uverbs_srq_idr, cmd.srq_handle, file->ucontext);
3112         if (!uobj)
3113                 return -EINVAL;
3114         srq = uobj->object;
3115         obj = container_of(uobj, struct ib_uevent_object, uobject);
3116         srq_type = srq->srq_type;
3117
3118         ret = ib_destroy_srq(srq);
3119         if (!ret)
3120                 uobj->live = 0;
3121
3122         put_uobj_write(uobj);
3123
3124         if (ret)
3125                 return ret;
3126
3127         if (srq_type == IB_SRQT_XRC) {
3128                 us = container_of(obj, struct ib_usrq_object, uevent);
3129                 atomic_dec(&us->uxrcd->refcnt);
3130         }
3131
3132         idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
3133
3134         mutex_lock(&file->mutex);
3135         list_del(&uobj->list);
3136         mutex_unlock(&file->mutex);
3137
3138         ib_uverbs_release_uevent(file, obj);
3139
3140         memset(&resp, 0, sizeof resp);
3141         resp.events_reported = obj->events_reported;
3142
3143         put_uobj(uobj);
3144
3145         if (copy_to_user((void __user *) (unsigned long) cmd.response,
3146                          &resp, sizeof resp))
3147                 ret = -EFAULT;
3148
3149         return ret ? ret : in_len;
3150 }