drivers/infiniband/core/uverbs_main.c

   1 /*
   2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
   6  * Copyright (c) 2005 PathScale, Inc. All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  */
  36
  37 #include <linux/module.h>
  38 #include <linux/init.h>
  39 #include <linux/device.h>
  40 #include <linux/err.h>
  41 #include <linux/fs.h>
  42 #include <linux/poll.h>
  43 #include <linux/sched.h>
  44 #include <linux/file.h>
  45 #include <linux/cdev.h>
  46 #include <linux/anon_inodes.h>
  47 #include <linux/slab.h>
  48
  49 #include <asm/uaccess.h>
  50
  51 #include <rdma/ib.h>
  52
  53 #include "uverbs.h"
  54
  55 MODULE_AUTHOR("Roland Dreier");
  56 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  57 MODULE_LICENSE("Dual BSD/GPL");
  58
  59 enum {
  60         IB_UVERBS_MAJOR       = 231,
  61         IB_UVERBS_BASE_MINOR  = 192,
  62         IB_UVERBS_MAX_DEVICES = 32
  63 };
  64
  65 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  66
  67 static struct class *uverbs_class;
  68
  69 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  70 DEFINE_IDR(ib_uverbs_pd_idr);
  71 DEFINE_IDR(ib_uverbs_mr_idr);
  72 DEFINE_IDR(ib_uverbs_mw_idr);
  73 DEFINE_IDR(ib_uverbs_ah_idr);
  74 DEFINE_IDR(ib_uverbs_cq_idr);
  75 DEFINE_IDR(ib_uverbs_qp_idr);
  76 DEFINE_IDR(ib_uverbs_srq_idr);
  77 DEFINE_IDR(ib_uverbs_xrcd_idr);
  78 DEFINE_IDR(ib_uverbs_rule_idr);
  79 DEFINE_IDR(ib_uverbs_wq_idr);
  80 DEFINE_IDR(ib_uverbs_rwq_ind_tbl_idr);
  81
  82 static DEFINE_SPINLOCK(map_lock);
  83 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  84
  85 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  86                                      struct ib_device *ib_dev,
  87                                      const char __user *buf, int in_len,
  88                                      int out_len) = {
  89         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  90         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  91         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  92         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  93         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  94         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  95         [IB_USER_VERBS_CMD_REREG_MR]            = ib_uverbs_rereg_mr,
  96         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  97         [IB_USER_VERBS_CMD_ALLOC_MW]            = ib_uverbs_alloc_mw,
  98         [IB_USER_VERBS_CMD_DEALLOC_MW]          = ib_uverbs_dealloc_mw,
  99         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
 100         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
 101         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
 102         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
 103         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
 104         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
 105         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
 106         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 107         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 108         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 109         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 110         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 111         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 112         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 113         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 114         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 115         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 116         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 117         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 118         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 119         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 120         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 121         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 122         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 123         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp,
 124 };
 125
 126 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
 127                                     struct ib_device *ib_dev,
 128                                     struct ib_udata *ucore,
 129                                     struct ib_udata *uhw) = {
 130         [IB_USER_VERBS_EX_CMD_CREATE_FLOW]      = ib_uverbs_ex_create_flow,
 131         [IB_USER_VERBS_EX_CMD_DESTROY_FLOW]     = ib_uverbs_ex_destroy_flow,
 132         [IB_USER_VERBS_EX_CMD_QUERY_DEVICE]     = ib_uverbs_ex_query_device,
 133         [IB_USER_VERBS_EX_CMD_CREATE_CQ]        = ib_uverbs_ex_create_cq,
 134         [IB_USER_VERBS_EX_CMD_CREATE_QP]        = ib_uverbs_ex_create_qp,
 135         [IB_USER_VERBS_EX_CMD_CREATE_WQ]        = ib_uverbs_ex_create_wq,
 136         [IB_USER_VERBS_EX_CMD_MODIFY_WQ]        = ib_uverbs_ex_modify_wq,
 137         [IB_USER_VERBS_EX_CMD_DESTROY_WQ]       = ib_uverbs_ex_destroy_wq,
 138         [IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL] = ib_uverbs_ex_create_rwq_ind_table,
 139         [IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL] = ib_uverbs_ex_destroy_rwq_ind_table,
 140 };
 141
 142 static void ib_uverbs_add_one(struct ib_device *device);
 143 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
 144
 145 int uverbs_dealloc_mw(struct ib_mw *mw)
 146 {
 147         struct ib_pd *pd = mw->pd;
 148         int ret;
 149
 150         ret = mw->device->dealloc_mw(mw);
 151         if (!ret)
 152                 atomic_dec(&pd->usecnt);
 153         return ret;
 154 }
 155
 156 static void ib_uverbs_release_dev(struct kobject *kobj)
 157 {
 158         struct ib_uverbs_device *dev =
 159                 container_of(kobj, struct ib_uverbs_device, kobj);
 160
 161         cleanup_srcu_struct(&dev->disassociate_srcu);
 162         kfree(dev);
 163 }
 164
 165 static struct kobj_type ib_uverbs_dev_ktype = {
 166         .release = ib_uverbs_release_dev,
 167 };
 168
 169 static void ib_uverbs_release_event_file(struct kref *ref)
 170 {
 171         struct ib_uverbs_event_file *file =
 172                 container_of(ref, struct ib_uverbs_event_file, ref);
 173
 174         kfree(file);
 175 }
 176
 177 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 178                           struct ib_uverbs_event_file *ev_file,
 179                           struct ib_ucq_object *uobj)
 180 {
 181         struct ib_uverbs_event *evt, *tmp;
 182
 183         if (ev_file) {
 184                 spin_lock_irq(&ev_file->lock);
 185                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 186                         list_del(&evt->list);
 187                         kfree(evt);
 188                 }
 189                 spin_unlock_irq(&ev_file->lock);
 190
 191                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 192         }
 193
 194         spin_lock_irq(&file->async_file->lock);
 195         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 196                 list_del(&evt->list);
 197                 kfree(evt);
 198         }
 199         spin_unlock_irq(&file->async_file->lock);
 200 }
 201
 202 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 203                               struct ib_uevent_object *uobj)
 204 {
 205         struct ib_uverbs_event *evt, *tmp;
 206
 207         spin_lock_irq(&file->async_file->lock);
 208         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 209                 list_del(&evt->list);
 210                 kfree(evt);
 211         }
 212         spin_unlock_irq(&file->async_file->lock);
 213 }
 214
 215 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 216                                     struct ib_uqp_object *uobj)
 217 {
 218         struct ib_uverbs_mcast_entry *mcast, *tmp;
 219
 220         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 221                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 222                 list_del(&mcast->list);
 223                 kfree(mcast);
 224         }
 225 }
 226
 227 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 228                                       struct ib_ucontext *context)
 229 {
 230         struct ib_uobject *uobj, *tmp;
 231
 232         context->closing = 1;
 233
 234         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 235                 struct ib_ah *ah = uobj->object;
 236
 237                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 238                 ib_destroy_ah(ah);
 239                 kfree(uobj);
 240         }
 241
 242         /* Remove MWs before QPs, in order to support type 2A MWs. */
 243         list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
 244                 struct ib_mw *mw = uobj->object;
 245
 246                 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
 247                 uverbs_dealloc_mw(mw);
 248                 kfree(uobj);
 249         }
 250
 251         list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
 252                 struct ib_flow *flow_id = uobj->object;
 253
 254                 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
 255                 ib_destroy_flow(flow_id);
 256                 kfree(uobj);
 257         }
 258
 259         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 260                 struct ib_qp *qp = uobj->object;
 261                 struct ib_uqp_object *uqp =
 262                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 263
 264                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 265                 if (qp == qp->real_qp)
 266                         ib_uverbs_detach_umcast(qp, uqp);
 267                 ib_destroy_qp(qp);
 268                 ib_uverbs_release_uevent(file, &uqp->uevent);
 269                 kfree(uqp);
 270         }
 271
 272         list_for_each_entry_safe(uobj, tmp, &context->rwq_ind_tbl_list, list) {
 273                 struct ib_rwq_ind_table *rwq_ind_tbl = uobj->object;
 274                 struct ib_wq **ind_tbl = rwq_ind_tbl->ind_tbl;
 275
 276                 idr_remove_uobj(&ib_uverbs_rwq_ind_tbl_idr, uobj);
 277                 ib_destroy_rwq_ind_table(rwq_ind_tbl);
 278                 kfree(ind_tbl);
 279                 kfree(uobj);
 280         }
 281
 282         list_for_each_entry_safe(uobj, tmp, &context->wq_list, list) {
 283                 struct ib_wq *wq = uobj->object;
 284                 struct ib_uwq_object *uwq =
 285                         container_of(uobj, struct ib_uwq_object, uevent.uobject);
 286
 287                 idr_remove_uobj(&ib_uverbs_wq_idr, uobj);
 288                 ib_destroy_wq(wq);
 289                 ib_uverbs_release_uevent(file, &uwq->uevent);
 290                 kfree(uwq);
 291         }
 292
 293         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 294                 struct ib_srq *srq = uobj->object;
 295                 struct ib_uevent_object *uevent =
 296                         container_of(uobj, struct ib_uevent_object, uobject);
 297
 298                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 299                 ib_destroy_srq(srq);
 300                 ib_uverbs_release_uevent(file, uevent);
 301                 kfree(uevent);
 302         }
 303
 304         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 305                 struct ib_cq *cq = uobj->object;
 306                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 307                 struct ib_ucq_object *ucq =
 308                         container_of(uobj, struct ib_ucq_object, uobject);
 309
 310                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 311                 ib_destroy_cq(cq);
 312                 ib_uverbs_release_ucq(file, ev_file, ucq);
 313                 kfree(ucq);
 314         }
 315
 316         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 317                 struct ib_mr *mr = uobj->object;
 318
 319                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 320                 ib_dereg_mr(mr);
 321                 kfree(uobj);
 322         }
 323
 324         mutex_lock(&file->device->xrcd_tree_mutex);
 325         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 326                 struct ib_xrcd *xrcd = uobj->object;
 327                 struct ib_uxrcd_object *uxrcd =
 328                         container_of(uobj, struct ib_uxrcd_object, uobject);
 329
 330                 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 331                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 332                 kfree(uxrcd);
 333         }
 334         mutex_unlock(&file->device->xrcd_tree_mutex);
 335
 336         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 337                 struct ib_pd *pd = uobj->object;
 338
 339                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 340                 ib_dealloc_pd(pd);
 341                 kfree(uobj);
 342         }
 343
 344         put_pid(context->tgid);
 345
 346         return context->device->dealloc_ucontext(context);
 347 }
 348
 349 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
 350 {
 351         complete(&dev->comp);
 352 }
 353
 354 static void ib_uverbs_release_file(struct kref *ref)
 355 {
 356         struct ib_uverbs_file *file =
 357                 container_of(ref, struct ib_uverbs_file, ref);
 358         struct ib_device *ib_dev;
 359         int srcu_key;
 360
 361         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 362         ib_dev = srcu_dereference(file->device->ib_dev,
 363                                   &file->device->disassociate_srcu);
 364         if (ib_dev && !ib_dev->disassociate_ucontext)
 365                 module_put(ib_dev->owner);
 366         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 367
 368         if (atomic_dec_and_test(&file->device->refcount))
 369                 ib_uverbs_comp_dev(file->device);
 370
 371         kfree(file);
 372 }
 373
 374 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 375                                     size_t count, loff_t *pos)
 376 {
 377         struct ib_uverbs_event_file *file = filp->private_data;
 378         struct ib_uverbs_event *event;
 379         int eventsz;
 380         int ret = 0;
 381
 382         spin_lock_irq(&file->lock);
 383
 384         while (list_empty(&file->event_list)) {
 385                 spin_unlock_irq(&file->lock);
 386
 387                 if (filp->f_flags & O_NONBLOCK)
 388                         return -EAGAIN;
 389
 390                 if (wait_event_interruptible(file->poll_wait,
 391                                              (!list_empty(&file->event_list) ||
 392                         /* The barriers built into wait_event_interruptible()
 393                          * and wake_up() guarentee this will see the null set
 394                          * without using RCU
 395                          */
 396                                              !file->uverbs_file->device->ib_dev)))
 397                         return -ERESTARTSYS;
 398
 399                 /* If device was disassociated and no event exists set an error */
 400                 if (list_empty(&file->event_list) &&
 401                     !file->uverbs_file->device->ib_dev)
 402                         return -EIO;
 403
 404                 spin_lock_irq(&file->lock);
 405         }
 406
 407         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 408
 409         if (file->is_async)
 410                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 411         else
 412                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 413
 414         if (eventsz > count) {
 415                 ret   = -EINVAL;
 416                 event = NULL;
 417         } else {
 418                 list_del(file->event_list.next);
 419                 if (event->counter) {
 420                         ++(*event->counter);
 421                         list_del(&event->obj_list);
 422                 }
 423         }
 424
 425         spin_unlock_irq(&file->lock);
 426
 427         if (event) {
 428                 if (copy_to_user(buf, event, eventsz))
 429                         ret = -EFAULT;
 430                 else
 431                         ret = eventsz;
 432         }
 433
 434         kfree(event);
 435
 436         return ret;
 437 }
 438
 439 static unsigned int ib_uverbs_event_poll(struct file *filp,
 440                                          struct poll_table_struct *wait)
 441 {
 442         unsigned int pollflags = 0;
 443         struct ib_uverbs_event_file *file = filp->private_data;
 444
 445         poll_wait(filp, &file->poll_wait, wait);
 446
 447         spin_lock_irq(&file->lock);
 448         if (!list_empty(&file->event_list))
 449                 pollflags = POLLIN | POLLRDNORM;
 450         spin_unlock_irq(&file->lock);
 451
 452         return pollflags;
 453 }
 454
 455 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 456 {
 457         struct ib_uverbs_event_file *file = filp->private_data;
 458
 459         return fasync_helper(fd, filp, on, &file->async_queue);
 460 }
 461
 462 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 463 {
 464         struct ib_uverbs_event_file *file = filp->private_data;
 465         struct ib_uverbs_event *entry, *tmp;
 466         int closed_already = 0;
 467
 468         mutex_lock(&file->uverbs_file->device->lists_mutex);
 469         spin_lock_irq(&file->lock);
 470         closed_already = file->is_closed;
 471         file->is_closed = 1;
 472         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 473                 if (entry->counter)
 474                         list_del(&entry->obj_list);
 475                 kfree(entry);
 476         }
 477         spin_unlock_irq(&file->lock);
 478         if (!closed_already) {
 479                 list_del(&file->list);
 480                 if (file->is_async)
 481                         ib_unregister_event_handler(&file->uverbs_file->
 482                                 event_handler);
 483         }
 484         mutex_unlock(&file->uverbs_file->device->lists_mutex);
 485
 486         kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 487         kref_put(&file->ref, ib_uverbs_release_event_file);
 488
 489         return 0;
 490 }
 491
 492 static const struct file_operations uverbs_event_fops = {
 493         .owner   = THIS_MODULE,
 494         .read    = ib_uverbs_event_read,
 495         .poll    = ib_uverbs_event_poll,
 496         .release = ib_uverbs_event_close,
 497         .fasync  = ib_uverbs_event_fasync,
 498         .llseek  = no_llseek,
 499 };
 500
 501 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 502 {
 503         struct ib_uverbs_event_file    *file = cq_context;
 504         struct ib_ucq_object           *uobj;
 505         struct ib_uverbs_event         *entry;
 506         unsigned long                   flags;
 507
 508         if (!file)
 509                 return;
 510
 511         spin_lock_irqsave(&file->lock, flags);
 512         if (file->is_closed) {
 513                 spin_unlock_irqrestore(&file->lock, flags);
 514                 return;
 515         }
 516
 517         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 518         if (!entry) {
 519                 spin_unlock_irqrestore(&file->lock, flags);
 520                 return;
 521         }
 522
 523         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 524
 525         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 526         entry->counter             = &uobj->comp_events_reported;
 527
 528         list_add_tail(&entry->list, &file->event_list);
 529         list_add_tail(&entry->obj_list, &uobj->comp_list);
 530         spin_unlock_irqrestore(&file->lock, flags);
 531
 532         wake_up_interruptible(&file->poll_wait);
 533         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 534 }
 535
 536 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 537                                     __u64 element, __u64 event,
 538                                     struct list_head *obj_list,
 539                                     u32 *counter)
 540 {
 541         struct ib_uverbs_event *entry;
 542         unsigned long flags;
 543
 544         spin_lock_irqsave(&file->async_file->lock, flags);
 545         if (file->async_file->is_closed) {
 546                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 547                 return;
 548         }
 549
 550         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 551         if (!entry) {
 552                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 553                 return;
 554         }
 555
 556         entry->desc.async.element    = element;
 557         entry->desc.async.event_type = event;
 558         entry->desc.async.reserved   = 0;
 559         entry->counter               = counter;
 560
 561         list_add_tail(&entry->list, &file->async_file->event_list);
 562         if (obj_list)
 563                 list_add_tail(&entry->obj_list, obj_list);
 564         spin_unlock_irqrestore(&file->async_file->lock, flags);
 565
 566         wake_up_interruptible(&file->async_file->poll_wait);
 567         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 568 }
 569
 570 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 571 {
 572         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 573                                                   struct ib_ucq_object, uobject);
 574
 575         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 576                                 event->event, &uobj->async_list,
 577                                 &uobj->async_events_reported);
 578 }
 579
 580 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 581 {
 582         struct ib_uevent_object *uobj;
 583
 584         /* for XRC target qp's, check that qp is live */
 585         if (!event->element.qp->uobject || !event->element.qp->uobject->live)
 586                 return;
 587
 588         uobj = container_of(event->element.qp->uobject,
 589                             struct ib_uevent_object, uobject);
 590
 591         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 592                                 event->event, &uobj->event_list,
 593                                 &uobj->events_reported);
 594 }
 595
 596 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
 597 {
 598         struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
 599                                                   struct ib_uevent_object, uobject);
 600
 601         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 602                                 event->event, &uobj->event_list,
 603                                 &uobj->events_reported);
 604 }
 605
 606 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 607 {
 608         struct ib_uevent_object *uobj;
 609
 610         uobj = container_of(event->element.srq->uobject,
 611                             struct ib_uevent_object, uobject);
 612
 613         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 614                                 event->event, &uobj->event_list,
 615                                 &uobj->events_reported);
 616 }
 617
 618 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 619                              struct ib_event *event)
 620 {
 621         struct ib_uverbs_file *file =
 622                 container_of(handler, struct ib_uverbs_file, event_handler);
 623
 624         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 625                                 NULL, NULL);
 626 }
 627
 628 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
 629 {
 630         kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 631         file->async_file = NULL;
 632 }
 633
 634 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 635                                         struct ib_device        *ib_dev,
 636                                         int is_async)
 637 {
 638         struct ib_uverbs_event_file *ev_file;
 639         struct file *filp;
 640         int ret;
 641
 642         ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
 643         if (!ev_file)
 644                 return ERR_PTR(-ENOMEM);
 645
 646         kref_init(&ev_file->ref);
 647         spin_lock_init(&ev_file->lock);
 648         INIT_LIST_HEAD(&ev_file->event_list);
 649         init_waitqueue_head(&ev_file->poll_wait);
 650         ev_file->uverbs_file = uverbs_file;
 651         kref_get(&ev_file->uverbs_file->ref);
 652         ev_file->async_queue = NULL;
 653         ev_file->is_closed   = 0;
 654
 655         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 656                                   ev_file, O_RDONLY);
 657         if (IS_ERR(filp))
 658                 goto err_put_refs;
 659
 660         mutex_lock(&uverbs_file->device->lists_mutex);
 661         list_add_tail(&ev_file->list,
 662                       &uverbs_file->device->uverbs_events_file_list);
 663         mutex_unlock(&uverbs_file->device->lists_mutex);
 664
 665         if (is_async) {
 666                 WARN_ON(uverbs_file->async_file);
 667                 uverbs_file->async_file = ev_file;
 668                 kref_get(&uverbs_file->async_file->ref);
 669                 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
 670                                       ib_dev,
 671                                       ib_uverbs_event_handler);
 672                 ret = ib_register_event_handler(&uverbs_file->event_handler);
 673                 if (ret)
 674                         goto err_put_file;
 675
 676                 /* At that point async file stuff was fully set */
 677                 ev_file->is_async = 1;
 678         }
 679
 680         return filp;
 681
 682 err_put_file:
 683         fput(filp);
 684         kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
 685         uverbs_file->async_file = NULL;
 686         return ERR_PTR(ret);
 687
 688 err_put_refs:
 689         kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
 690         kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 691         return filp;
 692 }
 693
 694 /*
 695  * Look up a completion event file by FD.  If lookup is successful,
 696  * takes a ref to the event file struct that it returns; if
 697  * unsuccessful, returns NULL.
 698  */
 699 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 700 {
 701         struct ib_uverbs_event_file *ev_file = NULL;
 702         struct fd f = fdget(fd);
 703
 704         if (!f.file)
 705                 return NULL;
 706
 707         if (f.file->f_op != &uverbs_event_fops)
 708                 goto out;
 709
 710         ev_file = f.file->private_data;
 711         if (ev_file->is_async) {
 712                 ev_file = NULL;
 713                 goto out;
 714         }
 715
 716         kref_get(&ev_file->ref);
 717
 718 out:
 719         fdput(f);
 720         return ev_file;
 721 }
 722
 723 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
 724 {
 725         u64 mask;
 726
 727         if (command <= IB_USER_VERBS_CMD_OPEN_QP)
 728                 mask = ib_dev->uverbs_cmd_mask;
 729         else
 730                 mask = ib_dev->uverbs_ex_cmd_mask;
 731
 732         if (mask & ((u64)1 << command))
 733                 return 0;
 734
 735         return -1;
 736 }
 737
 738 static bool verify_command_idx(u32 command, bool extended)
 739 {
 740         if (extended)
 741                 return command < ARRAY_SIZE(uverbs_ex_cmd_table);
 742
 743         return command < ARRAY_SIZE(uverbs_cmd_table);
 744 }
 745
 746 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 747                              size_t count, loff_t *pos)
 748 {
 749         struct ib_uverbs_file *file = filp->private_data;
 750         struct ib_device *ib_dev;
 751         struct ib_uverbs_cmd_hdr hdr;
 752         bool extended_command;
 753         __u32 command;
 754         __u32 flags;
 755         int srcu_key;
 756         ssize_t ret;
 757
 758         if (WARN_ON_ONCE(!ib_safe_file_access(filp)))
 759                 return -EACCES;
 760
 761         if (count < sizeof hdr)
 762                 return -EINVAL;
 763
 764         if (copy_from_user(&hdr, buf, sizeof hdr))
 765                 return -EFAULT;
 766
 767         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 768         ib_dev = srcu_dereference(file->device->ib_dev,
 769                                   &file->device->disassociate_srcu);
 770         if (!ib_dev) {
 771                 ret = -EIO;
 772                 goto out;
 773         }
 774
 775         if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 776                                    IB_USER_VERBS_CMD_COMMAND_MASK)) {
 777                 ret = -EINVAL;
 778                 goto out;
 779         }
 780
 781         command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 782         flags = (hdr.command &
 783                  IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
 784
 785         extended_command = flags & IB_USER_VERBS_CMD_FLAG_EXTENDED;
 786         if (!verify_command_idx(command, extended_command)) {
 787                 ret = -EINVAL;
 788                 goto out;
 789         }
 790
 791         if (verify_command_mask(ib_dev, command)) {
 792                 ret = -EOPNOTSUPP;
 793                 goto out;
 794         }
 795
 796         if (!file->ucontext &&
 797             command != IB_USER_VERBS_CMD_GET_CONTEXT) {
 798                 ret = -EINVAL;
 799                 goto out;
 800         }
 801
 802         if (!flags) {
 803                 if (!uverbs_cmd_table[command]) {
 804                         ret = -EINVAL;
 805                         goto out;
 806                 }
 807
 808                 if (hdr.in_words * 4 != count) {
 809                         ret = -EINVAL;
 810                         goto out;
 811                 }
 812
 813                 ret = uverbs_cmd_table[command](file, ib_dev,
 814                                                  buf + sizeof(hdr),
 815                                                  hdr.in_words * 4,
 816                                                  hdr.out_words * 4);
 817
 818         } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
 819                 struct ib_uverbs_ex_cmd_hdr ex_hdr;
 820                 struct ib_udata ucore;
 821                 struct ib_udata uhw;
 822                 size_t written_count = count;
 823
 824                 if (!uverbs_ex_cmd_table[command]) {
 825                         ret = -ENOSYS;
 826                         goto out;
 827                 }
 828
 829                 if (!file->ucontext) {
 830                         ret = -EINVAL;
 831                         goto out;
 832                 }
 833
 834                 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
 835                         ret = -EINVAL;
 836                         goto out;
 837                 }
 838
 839                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
 840                         ret = -EFAULT;
 841                         goto out;
 842                 }
 843
 844                 count -= sizeof(hdr) + sizeof(ex_hdr);
 845                 buf += sizeof(hdr) + sizeof(ex_hdr);
 846
 847                 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
 848                         ret = -EINVAL;
 849                         goto out;
 850                 }
 851
 852                 if (ex_hdr.cmd_hdr_reserved) {
 853                         ret = -EINVAL;
 854                         goto out;
 855                 }
 856
 857                 if (ex_hdr.response) {
 858                         if (!hdr.out_words && !ex_hdr.provider_out_words) {
 859                                 ret = -EINVAL;
 860                                 goto out;
 861                         }
 862
 863                         if (!access_ok(VERIFY_WRITE,
 864                                        (void __user *) (unsigned long) ex_hdr.response,
 865                                        (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
 866                                 ret = -EFAULT;
 867                                 goto out;
 868                         }
 869                 } else {
 870                         if (hdr.out_words || ex_hdr.provider_out_words) {
 871                                 ret = -EINVAL;
 872                                 goto out;
 873                         }
 874                 }
 875
 876                 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
 877                                        hdr.in_words * 8, hdr.out_words * 8);
 878
 879                 INIT_UDATA_BUF_OR_NULL(&uhw,
 880                                        buf + ucore.inlen,
 881                                        (unsigned long) ex_hdr.response + ucore.outlen,
 882                                        ex_hdr.provider_in_words * 8,
 883                                        ex_hdr.provider_out_words * 8);
 884
 885                 ret = uverbs_ex_cmd_table[command](file,
 886                                                    ib_dev,
 887                                                    &ucore,
 888                                                    &uhw);
 889                 if (!ret)
 890                         ret = written_count;
 891         } else {
 892                 ret = -ENOSYS;
 893         }
 894
 895 out:
 896         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 897         return ret;
 898 }
 899
 900 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 901 {
 902         struct ib_uverbs_file *file = filp->private_data;
 903         struct ib_device *ib_dev;
 904         int ret = 0;
 905         int srcu_key;
 906
 907         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 908         ib_dev = srcu_dereference(file->device->ib_dev,
 909                                   &file->device->disassociate_srcu);
 910         if (!ib_dev) {
 911                 ret = -EIO;
 912                 goto out;
 913         }
 914
 915         if (!file->ucontext)
 916                 ret = -ENODEV;
 917         else
 918                 ret = ib_dev->mmap(file->ucontext, vma);
 919 out:
 920         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 921         return ret;
 922 }
 923
 924 /*
 925  * ib_uverbs_open() does not need the BKL:
 926  *
 927  *  - the ib_uverbs_device structures are properly reference counted and
 928  *    everything else is purely local to the file being created, so
 929  *    races against other open calls are not a problem;
 930  *  - there is no ioctl method to race against;
 931  *  - the open method will either immediately run -ENXIO, or all
 932  *    required initialization will be done.
 933  */
 934 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 935 {
 936         struct ib_uverbs_device *dev;
 937         struct ib_uverbs_file *file;
 938         struct ib_device *ib_dev;
 939         int ret;
 940         int module_dependent;
 941         int srcu_key;
 942
 943         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 944         if (!atomic_inc_not_zero(&dev->refcount))
 945                 return -ENXIO;
 946
 947         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
 948         mutex_lock(&dev->lists_mutex);
 949         ib_dev = srcu_dereference(dev->ib_dev,
 950                                   &dev->disassociate_srcu);
 951         if (!ib_dev) {
 952                 ret = -EIO;
 953                 goto err;
 954         }
 955
 956         /* In case IB device supports disassociate ucontext, there is no hard
 957          * dependency between uverbs device and its low level device.
 958          */
 959         module_dependent = !(ib_dev->disassociate_ucontext);
 960
 961         if (module_dependent) {
 962                 if (!try_module_get(ib_dev->owner)) {
 963                         ret = -ENODEV;
 964                         goto err;
 965                 }
 966         }
 967
 968         file = kzalloc(sizeof(*file), GFP_KERNEL);
 969         if (!file) {
 970                 ret = -ENOMEM;
 971                 if (module_dependent)
 972                         goto err_module;
 973
 974                 goto err;
 975         }
 976
 977         file->device     = dev;
 978         file->ucontext   = NULL;
 979         file->async_file = NULL;
 980         kref_init(&file->ref);
 981         mutex_init(&file->mutex);
 982         mutex_init(&file->cleanup_mutex);
 983
 984         filp->private_data = file;
 985         kobject_get(&dev->kobj);
 986         list_add_tail(&file->list, &dev->uverbs_file_list);
 987         mutex_unlock(&dev->lists_mutex);
 988         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 989
 990         return nonseekable_open(inode, filp);
 991
 992 err_module:
 993         module_put(ib_dev->owner);
 994
 995 err:
 996         mutex_unlock(&dev->lists_mutex);
 997         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 998         if (atomic_dec_and_test(&dev->refcount))
 999                 ib_uverbs_comp_dev(dev);
1000
1001         return ret;
1002 }
1003
1004 static int ib_uverbs_close(struct inode *inode, struct file *filp)
1005 {
1006         struct ib_uverbs_file *file = filp->private_data;
1007         struct ib_uverbs_device *dev = file->device;
1008
1009         mutex_lock(&file->cleanup_mutex);
1010         if (file->ucontext) {
1011                 ib_uverbs_cleanup_ucontext(file, file->ucontext);
1012                 file->ucontext = NULL;
1013         }
1014         mutex_unlock(&file->cleanup_mutex);
1015
1016         mutex_lock(&file->device->lists_mutex);
1017         if (!file->is_closed) {
1018                 list_del(&file->list);
1019                 file->is_closed = 1;
1020         }
1021         mutex_unlock(&file->device->lists_mutex);
1022
1023         if (file->async_file)
1024                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
1025
1026         kref_put(&file->ref, ib_uverbs_release_file);
1027         kobject_put(&dev->kobj);
1028
1029         return 0;
1030 }
1031
1032 static const struct file_operations uverbs_fops = {
1033         .owner   = THIS_MODULE,
1034         .write   = ib_uverbs_write,
1035         .open    = ib_uverbs_open,
1036         .release = ib_uverbs_close,
1037         .llseek  = no_llseek,
1038 };
1039
1040 static const struct file_operations uverbs_mmap_fops = {
1041         .owner   = THIS_MODULE,
1042         .write   = ib_uverbs_write,
1043         .mmap    = ib_uverbs_mmap,
1044         .open    = ib_uverbs_open,
1045         .release = ib_uverbs_close,
1046         .llseek  = no_llseek,
1047 };
1048
1049 static struct ib_client uverbs_client = {
1050         .name   = "uverbs",
1051         .add    = ib_uverbs_add_one,
1052         .remove = ib_uverbs_remove_one
1053 };
1054
1055 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
1056                           char *buf)
1057 {
1058         int ret = -ENODEV;
1059         int srcu_key;
1060         struct ib_uverbs_device *dev = dev_get_drvdata(device);
1061         struct ib_device *ib_dev;
1062
1063         if (!dev)
1064                 return -ENODEV;
1065
1066         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1067         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1068         if (ib_dev)
1069                 ret = sprintf(buf, "%s\n", ib_dev->name);
1070         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1071
1072         return ret;
1073 }
1074 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1075
1076 static ssize_t show_dev_abi_version(struct device *device,
1077                                     struct device_attribute *attr, char *buf)
1078 {
1079         struct ib_uverbs_device *dev = dev_get_drvdata(device);
1080         int ret = -ENODEV;
1081         int srcu_key;
1082         struct ib_device *ib_dev;
1083
1084         if (!dev)
1085                 return -ENODEV;
1086         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1087         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1088         if (ib_dev)
1089                 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1090         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1091
1092         return ret;
1093 }
1094 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1095
1096 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1097                          __stringify(IB_USER_VERBS_ABI_VERSION));
1098
1099 static dev_t overflow_maj;
1100 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1101
1102 /*
1103  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1104  * requesting a new major number and doubling the number of max devices we
1105  * support. It's stupid, but simple.
1106  */
1107 static int find_overflow_devnum(void)
1108 {
1109         int ret;
1110
1111         if (!overflow_maj) {
1112                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1113                                           "infiniband_verbs");
1114                 if (ret) {
1115                         pr_err("user_verbs: couldn't register dynamic device number\n");
1116                         return ret;
1117                 }
1118         }
1119
1120         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1121         if (ret >= IB_UVERBS_MAX_DEVICES)
1122                 return -1;
1123
1124         return ret;
1125 }
1126
1127 static void ib_uverbs_add_one(struct ib_device *device)
1128 {
1129         int devnum;
1130         dev_t base;
1131         struct ib_uverbs_device *uverbs_dev;
1132         int ret;
1133
1134         if (!device->alloc_ucontext)
1135                 return;
1136
1137         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1138         if (!uverbs_dev)
1139                 return;
1140
1141         ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1142         if (ret) {
1143                 kfree(uverbs_dev);
1144                 return;
1145         }
1146
1147         atomic_set(&uverbs_dev->refcount, 1);
1148         init_completion(&uverbs_dev->comp);
1149         uverbs_dev->xrcd_tree = RB_ROOT;
1150         mutex_init(&uverbs_dev->xrcd_tree_mutex);
1151         kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1152         mutex_init(&uverbs_dev->lists_mutex);
1153         INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1154         INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1155
1156         spin_lock(&map_lock);
1157         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1158         if (devnum >= IB_UVERBS_MAX_DEVICES) {
1159                 spin_unlock(&map_lock);
1160                 devnum = find_overflow_devnum();
1161                 if (devnum < 0)
1162                         goto err;
1163
1164                 spin_lock(&map_lock);
1165                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1166                 base = devnum + overflow_maj;
1167                 set_bit(devnum, overflow_map);
1168         } else {
1169                 uverbs_dev->devnum = devnum;
1170                 base = devnum + IB_UVERBS_BASE_DEV;
1171                 set_bit(devnum, dev_map);
1172         }
1173         spin_unlock(&map_lock);
1174
1175         rcu_assign_pointer(uverbs_dev->ib_dev, device);
1176         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1177
1178         cdev_init(&uverbs_dev->cdev, NULL);
1179         uverbs_dev->cdev.owner = THIS_MODULE;
1180         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1181         uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1182         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1183         if (cdev_add(&uverbs_dev->cdev, base, 1))
1184                 goto err_cdev;
1185
1186         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
1187                                         uverbs_dev->cdev.dev, uverbs_dev,
1188                                         "uverbs%d", uverbs_dev->devnum);
1189         if (IS_ERR(uverbs_dev->dev))
1190                 goto err_cdev;
1191
1192         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1193                 goto err_class;
1194         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1195                 goto err_class;
1196
1197         ib_set_client_data(device, &uverbs_client, uverbs_dev);
1198
1199         return;
1200
1201 err_class:
1202         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1203
1204 err_cdev:
1205         cdev_del(&uverbs_dev->cdev);
1206         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1207                 clear_bit(devnum, dev_map);
1208         else
1209                 clear_bit(devnum, overflow_map);
1210
1211 err:
1212         if (atomic_dec_and_test(&uverbs_dev->refcount))
1213                 ib_uverbs_comp_dev(uverbs_dev);
1214         wait_for_completion(&uverbs_dev->comp);
1215         kobject_put(&uverbs_dev->kobj);
1216         return;
1217 }
1218
1219 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1220                                         struct ib_device *ib_dev)
1221 {
1222         struct ib_uverbs_file *file;
1223         struct ib_uverbs_event_file *event_file;
1224         struct ib_event event;
1225
1226         /* Pending running commands to terminate */
1227         synchronize_srcu(&uverbs_dev->disassociate_srcu);
1228         event.event = IB_EVENT_DEVICE_FATAL;
1229         event.element.port_num = 0;
1230         event.device = ib_dev;
1231
1232         mutex_lock(&uverbs_dev->lists_mutex);
1233         while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1234                 struct ib_ucontext *ucontext;
1235                 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1236                                         struct ib_uverbs_file, list);
1237                 file->is_closed = 1;
1238                 list_del(&file->list);
1239                 kref_get(&file->ref);
1240                 mutex_unlock(&uverbs_dev->lists_mutex);
1241
1242                 ib_uverbs_event_handler(&file->event_handler, &event);
1243
1244                 mutex_lock(&file->cleanup_mutex);
1245                 ucontext = file->ucontext;
1246                 file->ucontext = NULL;
1247                 mutex_unlock(&file->cleanup_mutex);
1248
1249                 /* At this point ib_uverbs_close cannot be running
1250                  * ib_uverbs_cleanup_ucontext
1251                  */
1252                 if (ucontext) {
1253                         /* We must release the mutex before going ahead and
1254                          * calling disassociate_ucontext. disassociate_ucontext
1255                          * might end up indirectly calling uverbs_close,
1256                          * for example due to freeing the resources
1257                          * (e.g mmput).
1258                          */
1259                         ib_dev->disassociate_ucontext(ucontext);
1260                         ib_uverbs_cleanup_ucontext(file, ucontext);
1261                 }
1262
1263                 mutex_lock(&uverbs_dev->lists_mutex);
1264                 kref_put(&file->ref, ib_uverbs_release_file);
1265         }
1266
1267         while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1268                 event_file = list_first_entry(&uverbs_dev->
1269                                               uverbs_events_file_list,
1270                                               struct ib_uverbs_event_file,
1271                                               list);
1272                 spin_lock_irq(&event_file->lock);
1273                 event_file->is_closed = 1;
1274                 spin_unlock_irq(&event_file->lock);
1275
1276                 list_del(&event_file->list);
1277                 if (event_file->is_async) {
1278                         ib_unregister_event_handler(&event_file->uverbs_file->
1279                                                     event_handler);
1280                         event_file->uverbs_file->event_handler.device = NULL;
1281                 }
1282
1283                 wake_up_interruptible(&event_file->poll_wait);
1284                 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1285         }
1286         mutex_unlock(&uverbs_dev->lists_mutex);
1287 }
1288
1289 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1290 {
1291         struct ib_uverbs_device *uverbs_dev = client_data;
1292         int wait_clients = 1;
1293
1294         if (!uverbs_dev)
1295                 return;
1296
1297         dev_set_drvdata(uverbs_dev->dev, NULL);
1298         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1299         cdev_del(&uverbs_dev->cdev);
1300
1301         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1302                 clear_bit(uverbs_dev->devnum, dev_map);
1303         else
1304                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1305
1306         if (device->disassociate_ucontext) {
1307                 /* We disassociate HW resources and immediately return.
1308                  * Userspace will see a EIO errno for all future access.
1309                  * Upon returning, ib_device may be freed internally and is not
1310                  * valid any more.
1311                  * uverbs_device is still available until all clients close
1312                  * their files, then the uverbs device ref count will be zero
1313                  * and its resources will be freed.
1314                  * Note: At this point no more files can be opened since the
1315                  * cdev was deleted, however active clients can still issue
1316                  * commands and close their open files.
1317                  */
1318                 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1319                 ib_uverbs_free_hw_resources(uverbs_dev, device);
1320                 wait_clients = 0;
1321         }
1322
1323         if (atomic_dec_and_test(&uverbs_dev->refcount))
1324                 ib_uverbs_comp_dev(uverbs_dev);
1325         if (wait_clients)
1326                 wait_for_completion(&uverbs_dev->comp);
1327         kobject_put(&uverbs_dev->kobj);
1328 }
1329
1330 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1331 {
1332         if (mode)
1333                 *mode = 0666;
1334         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1335 }
1336
1337 static int __init ib_uverbs_init(void)
1338 {
1339         int ret;
1340
1341         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1342                                      "infiniband_verbs");
1343         if (ret) {
1344                 pr_err("user_verbs: couldn't register device number\n");
1345                 goto out;
1346         }
1347
1348         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1349         if (IS_ERR(uverbs_class)) {
1350                 ret = PTR_ERR(uverbs_class);
1351                 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1352                 goto out_chrdev;
1353         }
1354
1355         uverbs_class->devnode = uverbs_devnode;
1356
1357         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1358         if (ret) {
1359                 pr_err("user_verbs: couldn't create abi_version attribute\n");
1360                 goto out_class;
1361         }
1362
1363         ret = ib_register_client(&uverbs_client);
1364         if (ret) {
1365                 pr_err("user_verbs: couldn't register client\n");
1366                 goto out_class;
1367         }
1368
1369         return 0;
1370
1371 out_class:
1372         class_destroy(uverbs_class);
1373
1374 out_chrdev:
1375         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1376
1377 out:
1378         return ret;
1379 }
1380
1381 static void __exit ib_uverbs_cleanup(void)
1382 {
1383         ib_unregister_client(&uverbs_client);
1384         class_destroy(uverbs_class);
1385         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1386         if (overflow_maj)
1387                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1388         idr_destroy(&ib_uverbs_pd_idr);
1389         idr_destroy(&ib_uverbs_mr_idr);
1390         idr_destroy(&ib_uverbs_mw_idr);
1391         idr_destroy(&ib_uverbs_ah_idr);
1392         idr_destroy(&ib_uverbs_cq_idr);
1393         idr_destroy(&ib_uverbs_qp_idr);
1394         idr_destroy(&ib_uverbs_srq_idr);
1395 }
1396
1397 module_init(ib_uverbs_init);
1398 module_exit(ib_uverbs_cleanup);