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 = kzalloc(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                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object,
1649                                           uobject);
1650                 atomic_inc(&obj->uxrcd->refcnt);
1651                 put_xrcd_read(xrcd_uobj);
1652         }
1653
1654         if (pd)
1655                 put_pd_read(pd);
1656         if (scq)
1657                 put_cq_read(scq);
1658         if (rcq && rcq != scq)
1659                 put_cq_read(rcq);
1660         if (srq)
1661                 put_srq_read(srq);
1662
1663         mutex_lock(&file->mutex);
1664         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1665         mutex_unlock(&file->mutex);
1666
1667         obj->uevent.uobject.live = 1;
1668
1669         up_write(&obj->uevent.uobject.mutex);
1670
1671         return in_len;
1672
1673 err_copy:
1674         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1675
1676 err_destroy:
1677         ib_destroy_qp(qp);
1678
1679 err_put:
1680         if (xrcd)
1681                 put_xrcd_read(xrcd_uobj);
1682         if (pd)
1683                 put_pd_read(pd);
1684         if (scq)
1685                 put_cq_read(scq);
1686         if (rcq && rcq != scq)
1687                 put_cq_read(rcq);
1688         if (srq)
1689                 put_srq_read(srq);
1690
1691         put_uobj_write(&obj->uevent.uobject);
1692         return ret;
1693 }
1694
1695 ssize_t ib_uverbs_open_qp(struct ib_uverbs_file *file,
1696                           const char __user *buf, int in_len, int out_len)
1697 {
1698         struct ib_uverbs_open_qp        cmd;
1699         struct ib_uverbs_create_qp_resp resp;
1700         struct ib_udata                 udata;
1701         struct ib_uqp_object           *obj;
1702         struct ib_xrcd                 *xrcd;
1703         struct ib_uobject              *uninitialized_var(xrcd_uobj);
1704         struct ib_qp                   *qp;
1705         struct ib_qp_open_attr          attr;
1706         int ret;
1707
1708         if (out_len < sizeof resp)
1709                 return -ENOSPC;
1710
1711         if (copy_from_user(&cmd, buf, sizeof cmd))
1712                 return -EFAULT;
1713
1714         INIT_UDATA(&udata, buf + sizeof cmd,
1715                    (unsigned long) cmd.response + sizeof resp,
1716                    in_len - sizeof cmd, out_len - sizeof resp);
1717
1718         obj = kmalloc(sizeof *obj, GFP_KERNEL);
1719         if (!obj)
1720                 return -ENOMEM;
1721
1722         init_uobj(&obj->uevent.uobject, cmd.user_handle, file->ucontext, &qp_lock_class);
1723         down_write(&obj->uevent.uobject.mutex);
1724
1725         xrcd = idr_read_xrcd(cmd.pd_handle, file->ucontext, &xrcd_uobj);
1726         if (!xrcd) {
1727                 ret = -EINVAL;
1728                 goto err_put;
1729         }
1730
1731         attr.event_handler = ib_uverbs_qp_event_handler;
1732         attr.qp_context    = file;
1733         attr.qp_num        = cmd.qpn;
1734         attr.qp_type       = cmd.qp_type;
1735
1736         obj->uevent.events_reported = 0;
1737         INIT_LIST_HEAD(&obj->uevent.event_list);
1738         INIT_LIST_HEAD(&obj->mcast_list);
1739
1740         qp = ib_open_qp(xrcd, &attr);
1741         if (IS_ERR(qp)) {
1742                 ret = PTR_ERR(qp);
1743                 goto err_put;
1744         }
1745
1746         qp->uobject = &obj->uevent.uobject;
1747
1748         obj->uevent.uobject.object = qp;
1749         ret = idr_add_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1750         if (ret)
1751                 goto err_destroy;
1752
1753         memset(&resp, 0, sizeof resp);
1754         resp.qpn       = qp->qp_num;
1755         resp.qp_handle = obj->uevent.uobject.id;
1756
1757         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1758                          &resp, sizeof resp)) {
1759                 ret = -EFAULT;
1760                 goto err_remove;
1761         }
1762
1763         obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
1764         atomic_inc(&obj->uxrcd->refcnt);
1765         put_xrcd_read(xrcd_uobj);
1766
1767         mutex_lock(&file->mutex);
1768         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->qp_list);
1769         mutex_unlock(&file->mutex);
1770
1771         obj->uevent.uobject.live = 1;
1772
1773         up_write(&obj->uevent.uobject.mutex);
1774
1775         return in_len;
1776
1777 err_remove:
1778         idr_remove_uobj(&ib_uverbs_qp_idr, &obj->uevent.uobject);
1779
1780 err_destroy:
1781         ib_destroy_qp(qp);
1782
1783 err_put:
1784         put_xrcd_read(xrcd_uobj);
1785         put_uobj_write(&obj->uevent.uobject);
1786         return ret;
1787 }
1788
1789 ssize_t ib_uverbs_query_qp(struct ib_uverbs_file *file,
1790                            const char __user *buf, int in_len,
1791                            int out_len)
1792 {
1793         struct ib_uverbs_query_qp      cmd;
1794         struct ib_uverbs_query_qp_resp resp;
1795         struct ib_qp                   *qp;
1796         struct ib_qp_attr              *attr;
1797         struct ib_qp_init_attr         *init_attr;
1798         int                            ret;
1799
1800         if (copy_from_user(&cmd, buf, sizeof cmd))
1801                 return -EFAULT;
1802
1803         attr      = kmalloc(sizeof *attr, GFP_KERNEL);
1804         init_attr = kmalloc(sizeof *init_attr, GFP_KERNEL);
1805         if (!attr || !init_attr) {
1806                 ret = -ENOMEM;
1807                 goto out;
1808         }
1809
1810         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1811         if (!qp) {
1812                 ret = -EINVAL;
1813                 goto out;
1814         }
1815
1816         ret = ib_query_qp(qp, attr, cmd.attr_mask, init_attr);
1817
1818         put_qp_read(qp);
1819
1820         if (ret)
1821                 goto out;
1822
1823         memset(&resp, 0, sizeof resp);
1824
1825         resp.qp_state               = attr->qp_state;
1826         resp.cur_qp_state           = attr->cur_qp_state;
1827         resp.path_mtu               = attr->path_mtu;
1828         resp.path_mig_state         = attr->path_mig_state;
1829         resp.qkey                   = attr->qkey;
1830         resp.rq_psn                 = attr->rq_psn;
1831         resp.sq_psn                 = attr->sq_psn;
1832         resp.dest_qp_num            = attr->dest_qp_num;
1833         resp.qp_access_flags        = attr->qp_access_flags;
1834         resp.pkey_index             = attr->pkey_index;
1835         resp.alt_pkey_index         = attr->alt_pkey_index;
1836         resp.sq_draining            = attr->sq_draining;
1837         resp.max_rd_atomic          = attr->max_rd_atomic;
1838         resp.max_dest_rd_atomic     = attr->max_dest_rd_atomic;
1839         resp.min_rnr_timer          = attr->min_rnr_timer;
1840         resp.port_num               = attr->port_num;
1841         resp.timeout                = attr->timeout;
1842         resp.retry_cnt              = attr->retry_cnt;
1843         resp.rnr_retry              = attr->rnr_retry;
1844         resp.alt_port_num           = attr->alt_port_num;
1845         resp.alt_timeout            = attr->alt_timeout;
1846
1847         memcpy(resp.dest.dgid, attr->ah_attr.grh.dgid.raw, 16);
1848         resp.dest.flow_label        = attr->ah_attr.grh.flow_label;
1849         resp.dest.sgid_index        = attr->ah_attr.grh.sgid_index;
1850         resp.dest.hop_limit         = attr->ah_attr.grh.hop_limit;
1851         resp.dest.traffic_class     = attr->ah_attr.grh.traffic_class;
1852         resp.dest.dlid              = attr->ah_attr.dlid;
1853         resp.dest.sl                = attr->ah_attr.sl;
1854         resp.dest.src_path_bits     = attr->ah_attr.src_path_bits;
1855         resp.dest.static_rate       = attr->ah_attr.static_rate;
1856         resp.dest.is_global         = !!(attr->ah_attr.ah_flags & IB_AH_GRH);
1857         resp.dest.port_num          = attr->ah_attr.port_num;
1858
1859         memcpy(resp.alt_dest.dgid, attr->alt_ah_attr.grh.dgid.raw, 16);
1860         resp.alt_dest.flow_label    = attr->alt_ah_attr.grh.flow_label;
1861         resp.alt_dest.sgid_index    = attr->alt_ah_attr.grh.sgid_index;
1862         resp.alt_dest.hop_limit     = attr->alt_ah_attr.grh.hop_limit;
1863         resp.alt_dest.traffic_class = attr->alt_ah_attr.grh.traffic_class;
1864         resp.alt_dest.dlid          = attr->alt_ah_attr.dlid;
1865         resp.alt_dest.sl            = attr->alt_ah_attr.sl;
1866         resp.alt_dest.src_path_bits = attr->alt_ah_attr.src_path_bits;
1867         resp.alt_dest.static_rate   = attr->alt_ah_attr.static_rate;
1868         resp.alt_dest.is_global     = !!(attr->alt_ah_attr.ah_flags & IB_AH_GRH);
1869         resp.alt_dest.port_num      = attr->alt_ah_attr.port_num;
1870
1871         resp.max_send_wr            = init_attr->cap.max_send_wr;
1872         resp.max_recv_wr            = init_attr->cap.max_recv_wr;
1873         resp.max_send_sge           = init_attr->cap.max_send_sge;
1874         resp.max_recv_sge           = init_attr->cap.max_recv_sge;
1875         resp.max_inline_data        = init_attr->cap.max_inline_data;
1876         resp.sq_sig_all             = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1877
1878         if (copy_to_user((void __user *) (unsigned long) cmd.response,
1879                          &resp, sizeof resp))
1880                 ret = -EFAULT;
1881
1882 out:
1883         kfree(attr);
1884         kfree(init_attr);
1885
1886         return ret ? ret : in_len;
1887 }
1888
1889 /* Remove ignored fields set in the attribute mask */
1890 static int modify_qp_mask(enum ib_qp_type qp_type, int mask)
1891 {
1892         switch (qp_type) {
1893         case IB_QPT_XRC_INI:
1894                 return mask & ~(IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER);
1895         case IB_QPT_XRC_TGT:
1896                 return mask & ~(IB_QP_MAX_QP_RD_ATOMIC | IB_QP_RETRY_CNT |
1897                                 IB_QP_RNR_RETRY);
1898         default:
1899                 return mask;
1900         }
1901 }
1902
1903 ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
1904                             const char __user *buf, int in_len,
1905                             int out_len)
1906 {
1907         struct ib_uverbs_modify_qp cmd;
1908         struct ib_udata            udata;
1909         struct ib_qp              *qp;
1910         struct ib_qp_attr         *attr;
1911         int                        ret;
1912
1913         if (copy_from_user(&cmd, buf, sizeof cmd))
1914                 return -EFAULT;
1915
1916         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
1917                    out_len);
1918
1919         attr = kmalloc(sizeof *attr, GFP_KERNEL);
1920         if (!attr)
1921                 return -ENOMEM;
1922
1923         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
1924         if (!qp) {
1925                 ret = -EINVAL;
1926                 goto out;
1927         }
1928
1929         attr->qp_state            = cmd.qp_state;
1930         attr->cur_qp_state        = cmd.cur_qp_state;
1931         attr->path_mtu            = cmd.path_mtu;
1932         attr->path_mig_state      = cmd.path_mig_state;
1933         attr->qkey                = cmd.qkey;
1934         attr->rq_psn              = cmd.rq_psn;
1935         attr->sq_psn              = cmd.sq_psn;
1936         attr->dest_qp_num         = cmd.dest_qp_num;
1937         attr->qp_access_flags     = cmd.qp_access_flags;
1938         attr->pkey_index          = cmd.pkey_index;
1939         attr->alt_pkey_index      = cmd.alt_pkey_index;
1940         attr->en_sqd_async_notify = cmd.en_sqd_async_notify;
1941         attr->max_rd_atomic       = cmd.max_rd_atomic;
1942         attr->max_dest_rd_atomic  = cmd.max_dest_rd_atomic;
1943         attr->min_rnr_timer       = cmd.min_rnr_timer;
1944         attr->port_num            = cmd.port_num;
1945         attr->timeout             = cmd.timeout;
1946         attr->retry_cnt           = cmd.retry_cnt;
1947         attr->rnr_retry           = cmd.rnr_retry;
1948         attr->alt_port_num        = cmd.alt_port_num;
1949         attr->alt_timeout         = cmd.alt_timeout;
1950
1951         memcpy(attr->ah_attr.grh.dgid.raw, cmd.dest.dgid, 16);
1952         attr->ah_attr.grh.flow_label        = cmd.dest.flow_label;
1953         attr->ah_attr.grh.sgid_index        = cmd.dest.sgid_index;
1954         attr->ah_attr.grh.hop_limit         = cmd.dest.hop_limit;
1955         attr->ah_attr.grh.traffic_class     = cmd.dest.traffic_class;
1956         attr->ah_attr.dlid                  = cmd.dest.dlid;
1957         attr->ah_attr.sl                    = cmd.dest.sl;
1958         attr->ah_attr.src_path_bits         = cmd.dest.src_path_bits;
1959         attr->ah_attr.static_rate           = cmd.dest.static_rate;
1960         attr->ah_attr.ah_flags              = cmd.dest.is_global ? IB_AH_GRH : 0;
1961         attr->ah_attr.port_num              = cmd.dest.port_num;
1962
1963         memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd.alt_dest.dgid, 16);
1964         attr->alt_ah_attr.grh.flow_label    = cmd.alt_dest.flow_label;
1965         attr->alt_ah_attr.grh.sgid_index    = cmd.alt_dest.sgid_index;
1966         attr->alt_ah_attr.grh.hop_limit     = cmd.alt_dest.hop_limit;
1967         attr->alt_ah_attr.grh.traffic_class = cmd.alt_dest.traffic_class;
1968         attr->alt_ah_attr.dlid              = cmd.alt_dest.dlid;
1969         attr->alt_ah_attr.sl                = cmd.alt_dest.sl;
1970         attr->alt_ah_attr.src_path_bits     = cmd.alt_dest.src_path_bits;
1971         attr->alt_ah_attr.static_rate       = cmd.alt_dest.static_rate;
1972         attr->alt_ah_attr.ah_flags          = cmd.alt_dest.is_global ? IB_AH_GRH : 0;
1973         attr->alt_ah_attr.port_num          = cmd.alt_dest.port_num;
1974
1975         if (qp->real_qp == qp) {
1976                 ret = qp->device->modify_qp(qp, attr,
1977                         modify_qp_mask(qp->qp_type, cmd.attr_mask), &udata);
1978         } else {
1979                 ret = ib_modify_qp(qp, attr, modify_qp_mask(qp->qp_type, cmd.attr_mask));
1980         }
1981
1982         put_qp_read(qp);
1983
1984         if (ret)
1985                 goto out;
1986
1987         ret = in_len;
1988
1989 out:
1990         kfree(attr);
1991
1992         return ret;
1993 }
1994
1995 ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
1996                              const char __user *buf, int in_len,
1997                              int out_len)
1998 {
1999         struct ib_uverbs_destroy_qp      cmd;
2000         struct ib_uverbs_destroy_qp_resp resp;
2001         struct ib_uobject               *uobj;
2002         struct ib_qp                    *qp;
2003         struct ib_uqp_object            *obj;
2004         int                              ret = -EINVAL;
2005
2006         if (copy_from_user(&cmd, buf, sizeof cmd))
2007                 return -EFAULT;
2008
2009         memset(&resp, 0, sizeof resp);
2010
2011         uobj = idr_write_uobj(&ib_uverbs_qp_idr, cmd.qp_handle, file->ucontext);
2012         if (!uobj)
2013                 return -EINVAL;
2014         qp  = uobj->object;
2015         obj = container_of(uobj, struct ib_uqp_object, uevent.uobject);
2016
2017         if (!list_empty(&obj->mcast_list)) {
2018                 put_uobj_write(uobj);
2019                 return -EBUSY;
2020         }
2021
2022         ret = ib_destroy_qp(qp);
2023         if (!ret)
2024                 uobj->live = 0;
2025
2026         put_uobj_write(uobj);
2027
2028         if (ret)
2029                 return ret;
2030
2031         if (obj->uxrcd)
2032                 atomic_dec(&obj->uxrcd->refcnt);
2033
2034         idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
2035
2036         mutex_lock(&file->mutex);
2037         list_del(&uobj->list);
2038         mutex_unlock(&file->mutex);
2039
2040         ib_uverbs_release_uevent(file, &obj->uevent);
2041
2042         resp.events_reported = obj->uevent.events_reported;
2043
2044         put_uobj(uobj);
2045
2046         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2047                          &resp, sizeof resp))
2048                 return -EFAULT;
2049
2050         return in_len;
2051 }
2052
2053 ssize_t ib_uverbs_post_send(struct ib_uverbs_file *file,
2054                             const char __user *buf, int in_len,
2055                             int out_len)
2056 {
2057         struct ib_uverbs_post_send      cmd;
2058         struct ib_uverbs_post_send_resp resp;
2059         struct ib_uverbs_send_wr       *user_wr;
2060         struct ib_send_wr              *wr = NULL, *last, *next, *bad_wr;
2061         struct ib_qp                   *qp;
2062         int                             i, sg_ind;
2063         int                             is_ud;
2064         ssize_t                         ret = -EINVAL;
2065
2066         if (copy_from_user(&cmd, buf, sizeof cmd))
2067                 return -EFAULT;
2068
2069         if (in_len < sizeof cmd + cmd.wqe_size * cmd.wr_count +
2070             cmd.sge_count * sizeof (struct ib_uverbs_sge))
2071                 return -EINVAL;
2072
2073         if (cmd.wqe_size < sizeof (struct ib_uverbs_send_wr))
2074                 return -EINVAL;
2075
2076         user_wr = kmalloc(cmd.wqe_size, GFP_KERNEL);
2077         if (!user_wr)
2078                 return -ENOMEM;
2079
2080         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2081         if (!qp)
2082                 goto out;
2083
2084         is_ud = qp->qp_type == IB_QPT_UD;
2085         sg_ind = 0;
2086         last = NULL;
2087         for (i = 0; i < cmd.wr_count; ++i) {
2088                 if (copy_from_user(user_wr,
2089                                    buf + sizeof cmd + i * cmd.wqe_size,
2090                                    cmd.wqe_size)) {
2091                         ret = -EFAULT;
2092                         goto out_put;
2093                 }
2094
2095                 if (user_wr->num_sge + sg_ind > cmd.sge_count) {
2096                         ret = -EINVAL;
2097                         goto out_put;
2098                 }
2099
2100                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2101                                user_wr->num_sge * sizeof (struct ib_sge),
2102                                GFP_KERNEL);
2103                 if (!next) {
2104                         ret = -ENOMEM;
2105                         goto out_put;
2106                 }
2107
2108                 if (!last)
2109                         wr = next;
2110                 else
2111                         last->next = next;
2112                 last = next;
2113
2114                 next->next       = NULL;
2115                 next->wr_id      = user_wr->wr_id;
2116                 next->num_sge    = user_wr->num_sge;
2117                 next->opcode     = user_wr->opcode;
2118                 next->send_flags = user_wr->send_flags;
2119
2120                 if (is_ud) {
2121                         next->wr.ud.ah = idr_read_ah(user_wr->wr.ud.ah,
2122                                                      file->ucontext);
2123                         if (!next->wr.ud.ah) {
2124                                 ret = -EINVAL;
2125                                 goto out_put;
2126                         }
2127                         next->wr.ud.remote_qpn  = user_wr->wr.ud.remote_qpn;
2128                         next->wr.ud.remote_qkey = user_wr->wr.ud.remote_qkey;
2129                 } else {
2130                         switch (next->opcode) {
2131                         case IB_WR_RDMA_WRITE_WITH_IMM:
2132                                 next->ex.imm_data =
2133                                         (__be32 __force) user_wr->ex.imm_data;
2134                         case IB_WR_RDMA_WRITE:
2135                         case IB_WR_RDMA_READ:
2136                                 next->wr.rdma.remote_addr =
2137                                         user_wr->wr.rdma.remote_addr;
2138                                 next->wr.rdma.rkey        =
2139                                         user_wr->wr.rdma.rkey;
2140                                 break;
2141                         case IB_WR_SEND_WITH_IMM:
2142                                 next->ex.imm_data =
2143                                         (__be32 __force) user_wr->ex.imm_data;
2144                                 break;
2145                         case IB_WR_SEND_WITH_INV:
2146                                 next->ex.invalidate_rkey =
2147                                         user_wr->ex.invalidate_rkey;
2148                                 break;
2149                         case IB_WR_ATOMIC_CMP_AND_SWP:
2150                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2151                                 next->wr.atomic.remote_addr =
2152                                         user_wr->wr.atomic.remote_addr;
2153                                 next->wr.atomic.compare_add =
2154                                         user_wr->wr.atomic.compare_add;
2155                                 next->wr.atomic.swap = user_wr->wr.atomic.swap;
2156                                 next->wr.atomic.rkey = user_wr->wr.atomic.rkey;
2157                                 break;
2158                         default:
2159                                 break;
2160                         }
2161                 }
2162
2163                 if (next->num_sge) {
2164                         next->sg_list = (void *) next +
2165                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2166                         if (copy_from_user(next->sg_list,
2167                                            buf + sizeof cmd +
2168                                            cmd.wr_count * cmd.wqe_size +
2169                                            sg_ind * sizeof (struct ib_sge),
2170                                            next->num_sge * sizeof (struct ib_sge))) {
2171                                 ret = -EFAULT;
2172                                 goto out_put;
2173                         }
2174                         sg_ind += next->num_sge;
2175                 } else
2176                         next->sg_list = NULL;
2177         }
2178
2179         resp.bad_wr = 0;
2180         ret = qp->device->post_send(qp->real_qp, wr, &bad_wr);
2181         if (ret)
2182                 for (next = wr; next; next = next->next) {
2183                         ++resp.bad_wr;
2184                         if (next == bad_wr)
2185                                 break;
2186                 }
2187
2188         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2189                          &resp, sizeof resp))
2190                 ret = -EFAULT;
2191
2192 out_put:
2193         put_qp_read(qp);
2194
2195         while (wr) {
2196                 if (is_ud && wr->wr.ud.ah)
2197                         put_ah_read(wr->wr.ud.ah);
2198                 next = wr->next;
2199                 kfree(wr);
2200                 wr = next;
2201         }
2202
2203 out:
2204         kfree(user_wr);
2205
2206         return ret ? ret : in_len;
2207 }
2208
2209 static struct ib_recv_wr *ib_uverbs_unmarshall_recv(const char __user *buf,
2210                                                     int in_len,
2211                                                     u32 wr_count,
2212                                                     u32 sge_count,
2213                                                     u32 wqe_size)
2214 {
2215         struct ib_uverbs_recv_wr *user_wr;
2216         struct ib_recv_wr        *wr = NULL, *last, *next;
2217         int                       sg_ind;
2218         int                       i;
2219         int                       ret;
2220
2221         if (in_len < wqe_size * wr_count +
2222             sge_count * sizeof (struct ib_uverbs_sge))
2223                 return ERR_PTR(-EINVAL);
2224
2225         if (wqe_size < sizeof (struct ib_uverbs_recv_wr))
2226                 return ERR_PTR(-EINVAL);
2227
2228         user_wr = kmalloc(wqe_size, GFP_KERNEL);
2229         if (!user_wr)
2230                 return ERR_PTR(-ENOMEM);
2231
2232         sg_ind = 0;
2233         last = NULL;
2234         for (i = 0; i < wr_count; ++i) {
2235                 if (copy_from_user(user_wr, buf + i * wqe_size,
2236                                    wqe_size)) {
2237                         ret = -EFAULT;
2238                         goto err;
2239                 }
2240
2241                 if (user_wr->num_sge + sg_ind > sge_count) {
2242                         ret = -EINVAL;
2243                         goto err;
2244                 }
2245
2246                 next = kmalloc(ALIGN(sizeof *next, sizeof (struct ib_sge)) +
2247                                user_wr->num_sge * sizeof (struct ib_sge),
2248                                GFP_KERNEL);
2249                 if (!next) {
2250                         ret = -ENOMEM;
2251                         goto err;
2252                 }
2253
2254                 if (!last)
2255                         wr = next;
2256                 else
2257                         last->next = next;
2258                 last = next;
2259
2260                 next->next       = NULL;
2261                 next->wr_id      = user_wr->wr_id;
2262                 next->num_sge    = user_wr->num_sge;
2263
2264                 if (next->num_sge) {
2265                         next->sg_list = (void *) next +
2266                                 ALIGN(sizeof *next, sizeof (struct ib_sge));
2267                         if (copy_from_user(next->sg_list,
2268                                            buf + wr_count * wqe_size +
2269                                            sg_ind * sizeof (struct ib_sge),
2270                                            next->num_sge * sizeof (struct ib_sge))) {
2271                                 ret = -EFAULT;
2272                                 goto err;
2273                         }
2274                         sg_ind += next->num_sge;
2275                 } else
2276                         next->sg_list = NULL;
2277         }
2278
2279         kfree(user_wr);
2280         return wr;
2281
2282 err:
2283         kfree(user_wr);
2284
2285         while (wr) {
2286                 next = wr->next;
2287                 kfree(wr);
2288                 wr = next;
2289         }
2290
2291         return ERR_PTR(ret);
2292 }
2293
2294 ssize_t ib_uverbs_post_recv(struct ib_uverbs_file *file,
2295                             const char __user *buf, int in_len,
2296                             int out_len)
2297 {
2298         struct ib_uverbs_post_recv      cmd;
2299         struct ib_uverbs_post_recv_resp resp;
2300         struct ib_recv_wr              *wr, *next, *bad_wr;
2301         struct ib_qp                   *qp;
2302         ssize_t                         ret = -EINVAL;
2303
2304         if (copy_from_user(&cmd, buf, sizeof cmd))
2305                 return -EFAULT;
2306
2307         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2308                                        in_len - sizeof cmd, cmd.wr_count,
2309                                        cmd.sge_count, cmd.wqe_size);
2310         if (IS_ERR(wr))
2311                 return PTR_ERR(wr);
2312
2313         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2314         if (!qp)
2315                 goto out;
2316
2317         resp.bad_wr = 0;
2318         ret = qp->device->post_recv(qp->real_qp, wr, &bad_wr);
2319
2320         put_qp_read(qp);
2321
2322         if (ret)
2323                 for (next = wr; next; next = next->next) {
2324                         ++resp.bad_wr;
2325                         if (next == bad_wr)
2326                                 break;
2327                 }
2328
2329         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2330                          &resp, sizeof resp))
2331                 ret = -EFAULT;
2332
2333 out:
2334         while (wr) {
2335                 next = wr->next;
2336                 kfree(wr);
2337                 wr = next;
2338         }
2339
2340         return ret ? ret : in_len;
2341 }
2342
2343 ssize_t ib_uverbs_post_srq_recv(struct ib_uverbs_file *file,
2344                                 const char __user *buf, int in_len,
2345                                 int out_len)
2346 {
2347         struct ib_uverbs_post_srq_recv      cmd;
2348         struct ib_uverbs_post_srq_recv_resp resp;
2349         struct ib_recv_wr                  *wr, *next, *bad_wr;
2350         struct ib_srq                      *srq;
2351         ssize_t                             ret = -EINVAL;
2352
2353         if (copy_from_user(&cmd, buf, sizeof cmd))
2354                 return -EFAULT;
2355
2356         wr = ib_uverbs_unmarshall_recv(buf + sizeof cmd,
2357                                        in_len - sizeof cmd, cmd.wr_count,
2358                                        cmd.sge_count, cmd.wqe_size);
2359         if (IS_ERR(wr))
2360                 return PTR_ERR(wr);
2361
2362         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
2363         if (!srq)
2364                 goto out;
2365
2366         resp.bad_wr = 0;
2367         ret = srq->device->post_srq_recv(srq, wr, &bad_wr);
2368
2369         put_srq_read(srq);
2370
2371         if (ret)
2372                 for (next = wr; next; next = next->next) {
2373                         ++resp.bad_wr;
2374                         if (next == bad_wr)
2375                                 break;
2376                 }
2377
2378         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2379                          &resp, sizeof resp))
2380                 ret = -EFAULT;
2381
2382 out:
2383         while (wr) {
2384                 next = wr->next;
2385                 kfree(wr);
2386                 wr = next;
2387         }
2388
2389         return ret ? ret : in_len;
2390 }
2391
2392 ssize_t ib_uverbs_create_ah(struct ib_uverbs_file *file,
2393                             const char __user *buf, int in_len,
2394                             int out_len)
2395 {
2396         struct ib_uverbs_create_ah       cmd;
2397         struct ib_uverbs_create_ah_resp  resp;
2398         struct ib_uobject               *uobj;
2399         struct ib_pd                    *pd;
2400         struct ib_ah                    *ah;
2401         struct ib_ah_attr               attr;
2402         int ret;
2403
2404         if (out_len < sizeof resp)
2405                 return -ENOSPC;
2406
2407         if (copy_from_user(&cmd, buf, sizeof cmd))
2408                 return -EFAULT;
2409
2410         uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
2411         if (!uobj)
2412                 return -ENOMEM;
2413
2414         init_uobj(uobj, cmd.user_handle, file->ucontext, &ah_lock_class);
2415         down_write(&uobj->mutex);
2416
2417         pd = idr_read_pd(cmd.pd_handle, file->ucontext);
2418         if (!pd) {
2419                 ret = -EINVAL;
2420                 goto err;
2421         }
2422
2423         attr.dlid              = cmd.attr.dlid;
2424         attr.sl                = cmd.attr.sl;
2425         attr.src_path_bits     = cmd.attr.src_path_bits;
2426         attr.static_rate       = cmd.attr.static_rate;
2427         attr.ah_flags          = cmd.attr.is_global ? IB_AH_GRH : 0;
2428         attr.port_num          = cmd.attr.port_num;
2429         attr.grh.flow_label    = cmd.attr.grh.flow_label;
2430         attr.grh.sgid_index    = cmd.attr.grh.sgid_index;
2431         attr.grh.hop_limit     = cmd.attr.grh.hop_limit;
2432         attr.grh.traffic_class = cmd.attr.grh.traffic_class;
2433         memcpy(attr.grh.dgid.raw, cmd.attr.grh.dgid, 16);
2434
2435         ah = ib_create_ah(pd, &attr);
2436         if (IS_ERR(ah)) {
2437                 ret = PTR_ERR(ah);
2438                 goto err_put;
2439         }
2440
2441         ah->uobject  = uobj;
2442         uobj->object = ah;
2443
2444         ret = idr_add_uobj(&ib_uverbs_ah_idr, uobj);
2445         if (ret)
2446                 goto err_destroy;
2447
2448         resp.ah_handle = uobj->id;
2449
2450         if (copy_to_user((void __user *) (unsigned long) cmd.response,
2451                          &resp, sizeof resp)) {
2452                 ret = -EFAULT;
2453                 goto err_copy;
2454         }
2455
2456         put_pd_read(pd);
2457
2458         mutex_lock(&file->mutex);
2459         list_add_tail(&uobj->list, &file->ucontext->ah_list);
2460         mutex_unlock(&file->mutex);
2461
2462         uobj->live = 1;
2463
2464         up_write(&uobj->mutex);
2465
2466         return in_len;
2467
2468 err_copy:
2469         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2470
2471 err_destroy:
2472         ib_destroy_ah(ah);
2473
2474 err_put:
2475         put_pd_read(pd);
2476
2477 err:
2478         put_uobj_write(uobj);
2479         return ret;
2480 }
2481
2482 ssize_t ib_uverbs_destroy_ah(struct ib_uverbs_file *file,
2483                              const char __user *buf, int in_len, int out_len)
2484 {
2485         struct ib_uverbs_destroy_ah cmd;
2486         struct ib_ah               *ah;
2487         struct ib_uobject          *uobj;
2488         int                         ret;
2489
2490         if (copy_from_user(&cmd, buf, sizeof cmd))
2491                 return -EFAULT;
2492
2493         uobj = idr_write_uobj(&ib_uverbs_ah_idr, cmd.ah_handle, file->ucontext);
2494         if (!uobj)
2495                 return -EINVAL;
2496         ah = uobj->object;
2497
2498         ret = ib_destroy_ah(ah);
2499         if (!ret)
2500                 uobj->live = 0;
2501
2502         put_uobj_write(uobj);
2503
2504         if (ret)
2505                 return ret;
2506
2507         idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
2508
2509         mutex_lock(&file->mutex);
2510         list_del(&uobj->list);
2511         mutex_unlock(&file->mutex);
2512
2513         put_uobj(uobj);
2514
2515         return in_len;
2516 }
2517
2518 ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
2519                                const char __user *buf, int in_len,
2520                                int out_len)
2521 {
2522         struct ib_uverbs_attach_mcast cmd;
2523         struct ib_qp                 *qp;
2524         struct ib_uqp_object         *obj;
2525         struct ib_uverbs_mcast_entry *mcast;
2526         int                           ret;
2527
2528         if (copy_from_user(&cmd, buf, sizeof cmd))
2529                 return -EFAULT;
2530
2531         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2532         if (!qp)
2533                 return -EINVAL;
2534
2535         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2536
2537         list_for_each_entry(mcast, &obj->mcast_list, list)
2538                 if (cmd.mlid == mcast->lid &&
2539                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2540                         ret = 0;
2541                         goto out_put;
2542                 }
2543
2544         mcast = kmalloc(sizeof *mcast, GFP_KERNEL);
2545         if (!mcast) {
2546                 ret = -ENOMEM;
2547                 goto out_put;
2548         }
2549
2550         mcast->lid = cmd.mlid;
2551         memcpy(mcast->gid.raw, cmd.gid, sizeof mcast->gid.raw);
2552
2553         ret = ib_attach_mcast(qp, &mcast->gid, cmd.mlid);
2554         if (!ret)
2555                 list_add_tail(&mcast->list, &obj->mcast_list);
2556         else
2557                 kfree(mcast);
2558
2559 out_put:
2560         put_qp_write(qp);
2561
2562         return ret ? ret : in_len;
2563 }
2564
2565 ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
2566                                const char __user *buf, int in_len,
2567                                int out_len)
2568 {
2569         struct ib_uverbs_detach_mcast cmd;
2570         struct ib_uqp_object         *obj;
2571         struct ib_qp                 *qp;
2572         struct ib_uverbs_mcast_entry *mcast;
2573         int                           ret = -EINVAL;
2574
2575         if (copy_from_user(&cmd, buf, sizeof cmd))
2576                 return -EFAULT;
2577
2578         qp = idr_write_qp(cmd.qp_handle, file->ucontext);
2579         if (!qp)
2580                 return -EINVAL;
2581
2582         ret = ib_detach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
2583         if (ret)
2584                 goto out_put;
2585
2586         obj = container_of(qp->uobject, struct ib_uqp_object, uevent.uobject);
2587
2588         list_for_each_entry(mcast, &obj->mcast_list, list)
2589                 if (cmd.mlid == mcast->lid &&
2590                     !memcmp(cmd.gid, mcast->gid.raw, sizeof mcast->gid.raw)) {
2591                         list_del(&mcast->list);
2592                         kfree(mcast);
2593                         break;
2594                 }
2595
2596 out_put:
2597         put_qp_write(qp);
2598
2599         return ret ? ret : in_len;
2600 }
2601
2602 static int kern_spec_to_ib_spec(struct ib_kern_spec *kern_spec,
2603                                 union ib_flow_spec *ib_spec)
2604 {
2605         ib_spec->type = kern_spec->type;
2606
2607         switch (ib_spec->type) {
2608         case IB_FLOW_SPEC_ETH:
2609                 ib_spec->eth.size = sizeof(struct ib_flow_spec_eth);
2610                 if (ib_spec->eth.size != kern_spec->eth.size)
2611                         return -EINVAL;
2612                 memcpy(&ib_spec->eth.val, &kern_spec->eth.val,
2613                        sizeof(struct ib_flow_eth_filter));
2614                 memcpy(&ib_spec->eth.mask, &kern_spec->eth.mask,
2615                        sizeof(struct ib_flow_eth_filter));
2616                 break;
2617         case IB_FLOW_SPEC_IPV4:
2618                 ib_spec->ipv4.size = sizeof(struct ib_flow_spec_ipv4);
2619                 if (ib_spec->ipv4.size != kern_spec->ipv4.size)
2620                         return -EINVAL;
2621                 memcpy(&ib_spec->ipv4.val, &kern_spec->ipv4.val,
2622                        sizeof(struct ib_flow_ipv4_filter));
2623                 memcpy(&ib_spec->ipv4.mask, &kern_spec->ipv4.mask,
2624                        sizeof(struct ib_flow_ipv4_filter));
2625                 break;
2626         case IB_FLOW_SPEC_TCP:
2627         case IB_FLOW_SPEC_UDP:
2628                 ib_spec->tcp_udp.size = sizeof(struct ib_flow_spec_tcp_udp);
2629                 if (ib_spec->tcp_udp.size != kern_spec->tcp_udp.size)
2630                         return -EINVAL;
2631                 memcpy(&ib_spec->tcp_udp.val, &kern_spec->tcp_udp.val,
2632                        sizeof(struct ib_flow_tcp_udp_filter));
2633                 memcpy(&ib_spec->tcp_udp.mask, &kern_spec->tcp_udp.mask,
2634                        sizeof(struct ib_flow_tcp_udp_filter));
2635                 break;
2636         default:
2637                 return -EINVAL;
2638         }
2639         return 0;
2640 }
2641
2642 ssize_t ib_uverbs_create_flow(struct ib_uverbs_file *file,
2643                               const char __user *buf, int in_len,
2644                               int out_len)
2645 {
2646         struct ib_uverbs_create_flow      cmd;
2647         struct ib_uverbs_create_flow_resp resp;
2648         struct ib_uobject                 *uobj;
2649         struct ib_flow                    *flow_id;
2650         struct ib_kern_flow_attr          *kern_flow_attr;
2651         struct ib_flow_attr               *flow_attr;
2652         struct ib_qp                      *qp;
2653         int err = 0;
2654         void *kern_spec;
2655         void *ib_spec;
2656         int i;
2657         int kern_attr_size;
2658
2659         if (out_len < sizeof(resp))
2660                 return -ENOSPC;
2661
2662         if (copy_from_user(&cmd, buf, sizeof(cmd)))
2663                 return -EFAULT;
2664
2665         if (cmd.comp_mask)
2666                 return -EINVAL;
2667
2668         if ((cmd.flow_attr.type == IB_FLOW_ATTR_SNIFFER &&
2669              !capable(CAP_NET_ADMIN)) || !capable(CAP_NET_RAW))
2670                 return -EPERM;
2671
2672         if (cmd.flow_attr.num_of_specs < 0 ||
2673             cmd.flow_attr.num_of_specs > IB_FLOW_SPEC_SUPPORT_LAYERS)
2674                 return -EINVAL;
2675
2676         kern_attr_size = cmd.flow_attr.size - sizeof(cmd) -
2677                          sizeof(struct ib_uverbs_cmd_hdr_ex);
2678
2679         if (cmd.flow_attr.size < 0 || cmd.flow_attr.size > in_len ||
2680             kern_attr_size < 0 || kern_attr_size >
2681             (cmd.flow_attr.num_of_specs * sizeof(struct ib_kern_spec)))
2682                 return -EINVAL;
2683
2684         if (cmd.flow_attr.num_of_specs) {
2685                 kern_flow_attr = kmalloc(cmd.flow_attr.size, GFP_KERNEL);
2686                 if (!kern_flow_attr)
2687                         return -ENOMEM;
2688
2689                 memcpy(kern_flow_attr, &cmd.flow_attr, sizeof(*kern_flow_attr));
2690                 if (copy_from_user(kern_flow_attr + 1, buf + sizeof(cmd),
2691                                    kern_attr_size)) {
2692                         err = -EFAULT;
2693                         goto err_free_attr;
2694                 }
2695         } else {
2696                 kern_flow_attr = &cmd.flow_attr;
2697                 kern_attr_size = sizeof(cmd.flow_attr);
2698         }
2699
2700         uobj = kmalloc(sizeof(*uobj), GFP_KERNEL);
2701         if (!uobj) {
2702                 err = -ENOMEM;
2703                 goto err_free_attr;
2704         }
2705         init_uobj(uobj, 0, file->ucontext, &rule_lock_class);
2706         down_write(&uobj->mutex);
2707
2708         qp = idr_read_qp(cmd.qp_handle, file->ucontext);
2709         if (!qp) {
2710                 err = -EINVAL;
2711                 goto err_uobj;
2712         }
2713
2714         flow_attr = kmalloc(cmd.flow_attr.size, GFP_KERNEL);
2715         if (!flow_attr) {
2716                 err = -ENOMEM;
2717                 goto err_put;
2718         }
2719
2720         flow_attr->type = kern_flow_attr->type;
2721         flow_attr->priority = kern_flow_attr->priority;
2722         flow_attr->num_of_specs = kern_flow_attr->num_of_specs;
2723         flow_attr->port = kern_flow_attr->port;
2724         flow_attr->flags = kern_flow_attr->flags;
2725         flow_attr->size = sizeof(*flow_attr);
2726
2727         kern_spec = kern_flow_attr + 1;
2728         ib_spec = flow_attr + 1;
2729         for (i = 0; i < flow_attr->num_of_specs && kern_attr_size > 0; i++) {
2730                 err = kern_spec_to_ib_spec(kern_spec, ib_spec);
2731                 if (err)
2732                         goto err_free;
2733                 flow_attr->size +=
2734                         ((union ib_flow_spec *) ib_spec)->size;
2735                 kern_attr_size -= ((struct ib_kern_spec *) kern_spec)->size;
2736                 kern_spec += ((struct ib_kern_spec *) kern_spec)->size;
2737                 ib_spec += ((union ib_flow_spec *) ib_spec)->size;
2738         }
2739         if (kern_attr_size) {
2740                 pr_warn("create flow failed, %d bytes left from uverb cmd\n",
2741                         kern_attr_size);
2742                 goto err_free;
2743         }
2744         flow_id = ib_create_flow(qp, flow_attr, IB_FLOW_DOMAIN_USER);
2745         if (IS_ERR(flow_id)) {
2746                 err = PTR_ERR(flow_id);
2747                 goto err_free;
2748         }
2749         flow_id->qp = qp;
2750         flow_id->uobject = uobj;
2751         uobj->object = flow_id;
2752
2753         err = idr_add_uobj(&ib_uverbs_rule_idr, uobj);
2754         if (err)
2755                 goto destroy_flow;
2756
2757         memset(&resp, 0, sizeof(resp));
2758         resp.flow_handle = uobj->id;
2759
2760         if (copy_to_user((void __user *)(unsigned long) cmd.response,
2761                          &resp, sizeof(resp))) {
2762                 err = -EFAULT;
2763                 goto err_copy;
2764         }
2765
2766         put_qp_read(qp);
2767         mutex_lock(&file->mutex);
2768         list_add_tail(&uobj->list, &file->ucontext->rule_list);
2769         mutex_unlock(&file->mutex);
2770
2771         uobj->live = 1;
2772
2773         up_write(&uobj->mutex);
2774         kfree(flow_attr);
2775         if (cmd.flow_attr.num_of_specs)
2776                 kfree(kern_flow_attr);
2777         return in_len;
2778 err_copy:
2779         idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
2780 destroy_flow:
2781         ib_destroy_flow(flow_id);
2782 err_free:
2783         kfree(flow_attr);
2784 err_put:
2785         put_qp_read(qp);
2786 err_uobj:
2787         put_uobj_write(uobj);
2788 err_free_attr:
2789         if (cmd.flow_attr.num_of_specs)
2790                 kfree(kern_flow_attr);
2791         return err;
2792 }
2793
2794 ssize_t ib_uverbs_destroy_flow(struct ib_uverbs_file *file,
2795                                const char __user *buf, int in_len,
2796                                int out_len) {
2797         struct ib_uverbs_destroy_flow   cmd;
2798         struct ib_flow                  *flow_id;
2799         struct ib_uobject               *uobj;
2800         int                             ret;
2801
2802         if (copy_from_user(&cmd, buf, sizeof(cmd)))
2803                 return -EFAULT;
2804
2805         uobj = idr_write_uobj(&ib_uverbs_rule_idr, cmd.flow_handle,
2806                               file->ucontext);
2807         if (!uobj)
2808                 return -EINVAL;
2809         flow_id = uobj->object;
2810
2811         ret = ib_destroy_flow(flow_id);
2812         if (!ret)
2813                 uobj->live = 0;
2814
2815         put_uobj_write(uobj);
2816
2817         idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
2818
2819         mutex_lock(&file->mutex);
2820         list_del(&uobj->list);
2821         mutex_unlock(&file->mutex);
2822
2823         put_uobj(uobj);
2824
2825         return ret ? ret : in_len;
2826 }
2827
2828 static int __uverbs_create_xsrq(struct ib_uverbs_file *file,
2829                                 struct ib_uverbs_create_xsrq *cmd,
2830                                 struct ib_udata *udata)
2831 {
2832         struct ib_uverbs_create_srq_resp resp;
2833         struct ib_usrq_object           *obj;
2834         struct ib_pd                    *pd;
2835         struct ib_srq                   *srq;
2836         struct ib_uobject               *uninitialized_var(xrcd_uobj);
2837         struct ib_srq_init_attr          attr;
2838         int ret;
2839
2840         obj = kmalloc(sizeof *obj, GFP_KERNEL);
2841         if (!obj)
2842                 return -ENOMEM;
2843
2844         init_uobj(&obj->uevent.uobject, cmd->user_handle, file->ucontext, &srq_lock_class);
2845         down_write(&obj->uevent.uobject.mutex);
2846
2847         if (cmd->srq_type == IB_SRQT_XRC) {
2848                 attr.ext.xrc.xrcd  = idr_read_xrcd(cmd->xrcd_handle, file->ucontext, &xrcd_uobj);
2849                 if (!attr.ext.xrc.xrcd) {
2850                         ret = -EINVAL;
2851                         goto err;
2852                 }
2853
2854                 obj->uxrcd = container_of(xrcd_uobj, struct ib_uxrcd_object, uobject);
2855                 atomic_inc(&obj->uxrcd->refcnt);
2856
2857                 attr.ext.xrc.cq  = idr_read_cq(cmd->cq_handle, file->ucontext, 0);
2858                 if (!attr.ext.xrc.cq) {
2859                         ret = -EINVAL;
2860                         goto err_put_xrcd;
2861                 }
2862         }
2863
2864         pd  = idr_read_pd(cmd->pd_handle, file->ucontext);
2865         if (!pd) {
2866                 ret = -EINVAL;
2867                 goto err_put_cq;
2868         }
2869
2870         attr.event_handler  = ib_uverbs_srq_event_handler;
2871         attr.srq_context    = file;
2872         attr.srq_type       = cmd->srq_type;
2873         attr.attr.max_wr    = cmd->max_wr;
2874         attr.attr.max_sge   = cmd->max_sge;
2875         attr.attr.srq_limit = cmd->srq_limit;
2876
2877         obj->uevent.events_reported = 0;
2878         INIT_LIST_HEAD(&obj->uevent.event_list);
2879
2880         srq = pd->device->create_srq(pd, &attr, udata);
2881         if (IS_ERR(srq)) {
2882                 ret = PTR_ERR(srq);
2883                 goto err_put;
2884         }
2885
2886         srq->device        = pd->device;
2887         srq->pd            = pd;
2888         srq->srq_type      = cmd->srq_type;
2889         srq->uobject       = &obj->uevent.uobject;
2890         srq->event_handler = attr.event_handler;
2891         srq->srq_context   = attr.srq_context;
2892
2893         if (cmd->srq_type == IB_SRQT_XRC) {
2894                 srq->ext.xrc.cq   = attr.ext.xrc.cq;
2895                 srq->ext.xrc.xrcd = attr.ext.xrc.xrcd;
2896                 atomic_inc(&attr.ext.xrc.cq->usecnt);
2897                 atomic_inc(&attr.ext.xrc.xrcd->usecnt);
2898         }
2899
2900         atomic_inc(&pd->usecnt);
2901         atomic_set(&srq->usecnt, 0);
2902
2903         obj->uevent.uobject.object = srq;
2904         ret = idr_add_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2905         if (ret)
2906                 goto err_destroy;
2907
2908         memset(&resp, 0, sizeof resp);
2909         resp.srq_handle = obj->uevent.uobject.id;
2910         resp.max_wr     = attr.attr.max_wr;
2911         resp.max_sge    = attr.attr.max_sge;
2912         if (cmd->srq_type == IB_SRQT_XRC)
2913                 resp.srqn = srq->ext.xrc.srq_num;
2914
2915         if (copy_to_user((void __user *) (unsigned long) cmd->response,
2916                          &resp, sizeof resp)) {
2917                 ret = -EFAULT;
2918                 goto err_copy;
2919         }
2920
2921         if (cmd->srq_type == IB_SRQT_XRC) {
2922                 put_uobj_read(xrcd_uobj);
2923                 put_cq_read(attr.ext.xrc.cq);
2924         }
2925         put_pd_read(pd);
2926
2927         mutex_lock(&file->mutex);
2928         list_add_tail(&obj->uevent.uobject.list, &file->ucontext->srq_list);
2929         mutex_unlock(&file->mutex);
2930
2931         obj->uevent.uobject.live = 1;
2932
2933         up_write(&obj->uevent.uobject.mutex);
2934
2935         return 0;
2936
2937 err_copy:
2938         idr_remove_uobj(&ib_uverbs_srq_idr, &obj->uevent.uobject);
2939
2940 err_destroy:
2941         ib_destroy_srq(srq);
2942
2943 err_put:
2944         put_pd_read(pd);
2945
2946 err_put_cq:
2947         if (cmd->srq_type == IB_SRQT_XRC)
2948                 put_cq_read(attr.ext.xrc.cq);
2949
2950 err_put_xrcd:
2951         if (cmd->srq_type == IB_SRQT_XRC) {
2952                 atomic_dec(&obj->uxrcd->refcnt);
2953                 put_uobj_read(xrcd_uobj);
2954         }
2955
2956 err:
2957         put_uobj_write(&obj->uevent.uobject);
2958         return ret;
2959 }
2960
2961 ssize_t ib_uverbs_create_srq(struct ib_uverbs_file *file,
2962                              const char __user *buf, int in_len,
2963                              int out_len)
2964 {
2965         struct ib_uverbs_create_srq      cmd;
2966         struct ib_uverbs_create_xsrq     xcmd;
2967         struct ib_uverbs_create_srq_resp resp;
2968         struct ib_udata                  udata;
2969         int ret;
2970
2971         if (out_len < sizeof resp)
2972                 return -ENOSPC;
2973
2974         if (copy_from_user(&cmd, buf, sizeof cmd))
2975                 return -EFAULT;
2976
2977         xcmd.response    = cmd.response;
2978         xcmd.user_handle = cmd.user_handle;
2979         xcmd.srq_type    = IB_SRQT_BASIC;
2980         xcmd.pd_handle   = cmd.pd_handle;
2981         xcmd.max_wr      = cmd.max_wr;
2982         xcmd.max_sge     = cmd.max_sge;
2983         xcmd.srq_limit   = cmd.srq_limit;
2984
2985         INIT_UDATA(&udata, buf + sizeof cmd,
2986                    (unsigned long) cmd.response + sizeof resp,
2987                    in_len - sizeof cmd, out_len - sizeof resp);
2988
2989         ret = __uverbs_create_xsrq(file, &xcmd, &udata);
2990         if (ret)
2991                 return ret;
2992
2993         return in_len;
2994 }
2995
2996 ssize_t ib_uverbs_create_xsrq(struct ib_uverbs_file *file,
2997                               const char __user *buf, int in_len, int out_len)
2998 {
2999         struct ib_uverbs_create_xsrq     cmd;
3000         struct ib_uverbs_create_srq_resp resp;
3001         struct ib_udata                  udata;
3002         int ret;
3003
3004         if (out_len < sizeof resp)
3005                 return -ENOSPC;
3006
3007         if (copy_from_user(&cmd, buf, sizeof cmd))
3008                 return -EFAULT;
3009
3010         INIT_UDATA(&udata, buf + sizeof cmd,
3011                    (unsigned long) cmd.response + sizeof resp,
3012                    in_len - sizeof cmd, out_len - sizeof resp);
3013
3014         ret = __uverbs_create_xsrq(file, &cmd, &udata);
3015         if (ret)
3016                 return ret;
3017
3018         return in_len;
3019 }
3020
3021 ssize_t ib_uverbs_modify_srq(struct ib_uverbs_file *file,
3022                              const char __user *buf, int in_len,
3023                              int out_len)
3024 {
3025         struct ib_uverbs_modify_srq cmd;
3026         struct ib_udata             udata;
3027         struct ib_srq              *srq;
3028         struct ib_srq_attr          attr;
3029         int                         ret;
3030
3031         if (copy_from_user(&cmd, buf, sizeof cmd))
3032                 return -EFAULT;
3033
3034         INIT_UDATA(&udata, buf + sizeof cmd, NULL, in_len - sizeof cmd,
3035                    out_len);
3036
3037         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
3038         if (!srq)
3039                 return -EINVAL;
3040
3041         attr.max_wr    = cmd.max_wr;
3042         attr.srq_limit = cmd.srq_limit;
3043
3044         ret = srq->device->modify_srq(srq, &attr, cmd.attr_mask, &udata);
3045
3046         put_srq_read(srq);
3047
3048         return ret ? ret : in_len;
3049 }
3050
3051 ssize_t ib_uverbs_query_srq(struct ib_uverbs_file *file,
3052                             const char __user *buf,
3053                             int in_len, int out_len)
3054 {
3055         struct ib_uverbs_query_srq      cmd;
3056         struct ib_uverbs_query_srq_resp resp;
3057         struct ib_srq_attr              attr;
3058         struct ib_srq                   *srq;
3059         int                             ret;
3060
3061         if (out_len < sizeof resp)
3062                 return -ENOSPC;
3063
3064         if (copy_from_user(&cmd, buf, sizeof cmd))
3065                 return -EFAULT;
3066
3067         srq = idr_read_srq(cmd.srq_handle, file->ucontext);
3068         if (!srq)
3069                 return -EINVAL;
3070
3071         ret = ib_query_srq(srq, &attr);
3072
3073         put_srq_read(srq);
3074
3075         if (ret)
3076                 return ret;
3077
3078         memset(&resp, 0, sizeof resp);
3079
3080         resp.max_wr    = attr.max_wr;
3081         resp.max_sge   = attr.max_sge;
3082         resp.srq_limit = attr.srq_limit;
3083
3084         if (copy_to_user((void __user *) (unsigned long) cmd.response,
3085                          &resp, sizeof resp))
3086                 return -EFAULT;
3087
3088         return in_len;
3089 }
3090
3091 ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
3092                               const char __user *buf, int in_len,
3093                               int out_len)
3094 {
3095         struct ib_uverbs_destroy_srq      cmd;
3096         struct ib_uverbs_destroy_srq_resp resp;
3097         struct ib_uobject                *uobj;
3098         struct ib_srq                    *srq;
3099         struct ib_uevent_object          *obj;
3100         int                               ret = -EINVAL;
3101         struct ib_usrq_object            *us;
3102         enum ib_srq_type                  srq_type;
3103
3104         if (copy_from_user(&cmd, buf, sizeof cmd))
3105                 return -EFAULT;
3106
3107         uobj = idr_write_uobj(&ib_uverbs_srq_idr, cmd.srq_handle, file->ucontext);
3108         if (!uobj)
3109                 return -EINVAL;
3110         srq = uobj->object;
3111         obj = container_of(uobj, struct ib_uevent_object, uobject);
3112         srq_type = srq->srq_type;
3113
3114         ret = ib_destroy_srq(srq);
3115         if (!ret)
3116                 uobj->live = 0;
3117
3118         put_uobj_write(uobj);
3119
3120         if (ret)
3121                 return ret;
3122
3123         if (srq_type == IB_SRQT_XRC) {
3124                 us = container_of(obj, struct ib_usrq_object, uevent);
3125                 atomic_dec(&us->uxrcd->refcnt);
3126         }
3127
3128         idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
3129
3130         mutex_lock(&file->mutex);
3131         list_del(&uobj->list);
3132         mutex_unlock(&file->mutex);
3133
3134         ib_uverbs_release_uevent(file, obj);
3135
3136         memset(&resp, 0, sizeof resp);
3137         resp.events_reported = obj->events_reported;
3138
3139         put_uobj(uobj);
3140
3141         if (copy_to_user((void __user *) (unsigned long) cmd.response,
3142                          &resp, sizeof resp))
3143                 ret = -EFAULT;
3144
3145         return ret ? ret : in_len;
3146 }