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