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