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