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  * $Id: uverbs_main.c 2733 2005-06-28 19:14:34Z roland $
  37  */
  38
  39 #include <linux/module.h>
  40 #include <linux/init.h>
  41 #include <linux/device.h>
  42 #include <linux/err.h>
  43 #include <linux/fs.h>
  44 #include <linux/poll.h>
  45 #include <linux/file.h>
  46 #include <linux/mount.h>
  47 #include <linux/cdev.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 #define INFINIBANDEVENTFS_MAGIC 0x49426576      /* "IBev" */
  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_MUTEX(ib_uverbs_idr_mutex);
  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
  78 static spinlock_t map_lock;
  79 static struct ib_uverbs_device *dev_table[IB_UVERBS_MAX_DEVICES];
  80 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  81
  82 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  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_DEREG_MR]            = ib_uverbs_dereg_mr,
  92         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  93         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  94         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  95         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  96         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  97         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  98         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  99         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 100         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 101         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 102         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 103         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 104         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 105         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 106         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 107         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 108         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 109         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 110         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 111         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 112 };
 113
 114 static struct vfsmount *uverbs_event_mnt;
 115
 116 static void ib_uverbs_add_one(struct ib_device *device);
 117 static void ib_uverbs_remove_one(struct ib_device *device);
 118
 119 static void ib_uverbs_release_dev(struct kref *ref)
 120 {
 121         struct ib_uverbs_device *dev =
 122                 container_of(ref, struct ib_uverbs_device, ref);
 123
 124         kfree(dev);
 125 }
 126
 127 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 128                           struct ib_uverbs_event_file *ev_file,
 129                           struct ib_ucq_object *uobj)
 130 {
 131         struct ib_uverbs_event *evt, *tmp;
 132
 133         if (ev_file) {
 134                 spin_lock_irq(&ev_file->lock);
 135                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 136                         list_del(&evt->list);
 137                         kfree(evt);
 138                 }
 139                 spin_unlock_irq(&ev_file->lock);
 140
 141                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 142         }
 143
 144         spin_lock_irq(&file->async_file->lock);
 145         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 146                 list_del(&evt->list);
 147                 kfree(evt);
 148         }
 149         spin_unlock_irq(&file->async_file->lock);
 150 }
 151
 152 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 153                               struct ib_uevent_object *uobj)
 154 {
 155         struct ib_uverbs_event *evt, *tmp;
 156
 157         spin_lock_irq(&file->async_file->lock);
 158         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 159                 list_del(&evt->list);
 160                 kfree(evt);
 161         }
 162         spin_unlock_irq(&file->async_file->lock);
 163 }
 164
 165 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 166                                     struct ib_uqp_object *uobj)
 167 {
 168         struct ib_uverbs_mcast_entry *mcast, *tmp;
 169
 170         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 171                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 172                 list_del(&mcast->list);
 173                 kfree(mcast);
 174         }
 175 }
 176
 177 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 178                                       struct ib_ucontext *context)
 179 {
 180         struct ib_uobject *uobj, *tmp;
 181
 182         if (!context)
 183                 return 0;
 184
 185         mutex_lock(&ib_uverbs_idr_mutex);
 186
 187         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 188                 struct ib_ah *ah = idr_find(&ib_uverbs_ah_idr, uobj->id);
 189                 idr_remove(&ib_uverbs_ah_idr, uobj->id);
 190                 ib_destroy_ah(ah);
 191                 list_del(&uobj->list);
 192                 kfree(uobj);
 193         }
 194
 195         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 196                 struct ib_qp *qp = idr_find(&ib_uverbs_qp_idr, uobj->id);
 197                 struct ib_uqp_object *uqp =
 198                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 199                 idr_remove(&ib_uverbs_qp_idr, uobj->id);
 200                 ib_uverbs_detach_umcast(qp, uqp);
 201                 ib_destroy_qp(qp);
 202                 list_del(&uobj->list);
 203                 ib_uverbs_release_uevent(file, &uqp->uevent);
 204                 kfree(uqp);
 205         }
 206
 207         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 208                 struct ib_cq *cq = idr_find(&ib_uverbs_cq_idr, uobj->id);
 209                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 210                 struct ib_ucq_object *ucq =
 211                         container_of(uobj, struct ib_ucq_object, uobject);
 212                 idr_remove(&ib_uverbs_cq_idr, uobj->id);
 213                 ib_destroy_cq(cq);
 214                 list_del(&uobj->list);
 215                 ib_uverbs_release_ucq(file, ev_file, ucq);
 216                 kfree(ucq);
 217         }
 218
 219         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 220                 struct ib_srq *srq = idr_find(&ib_uverbs_srq_idr, uobj->id);
 221                 struct ib_uevent_object *uevent =
 222                         container_of(uobj, struct ib_uevent_object, uobject);
 223                 idr_remove(&ib_uverbs_srq_idr, uobj->id);
 224                 ib_destroy_srq(srq);
 225                 list_del(&uobj->list);
 226                 ib_uverbs_release_uevent(file, uevent);
 227                 kfree(uevent);
 228         }
 229
 230         /* XXX Free MWs */
 231
 232         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 233                 struct ib_mr *mr = idr_find(&ib_uverbs_mr_idr, uobj->id);
 234                 struct ib_device *mrdev = mr->device;
 235                 struct ib_umem_object *memobj;
 236
 237                 idr_remove(&ib_uverbs_mr_idr, uobj->id);
 238                 ib_dereg_mr(mr);
 239
 240                 memobj = container_of(uobj, struct ib_umem_object, uobject);
 241                 ib_umem_release_on_close(mrdev, &memobj->umem);
 242
 243                 list_del(&uobj->list);
 244                 kfree(memobj);
 245         }
 246
 247         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 248                 struct ib_pd *pd = idr_find(&ib_uverbs_pd_idr, uobj->id);
 249                 idr_remove(&ib_uverbs_pd_idr, uobj->id);
 250                 ib_dealloc_pd(pd);
 251                 list_del(&uobj->list);
 252                 kfree(uobj);
 253         }
 254
 255         mutex_unlock(&ib_uverbs_idr_mutex);
 256
 257         return context->device->dealloc_ucontext(context);
 258 }
 259
 260 static void ib_uverbs_release_file(struct kref *ref)
 261 {
 262         struct ib_uverbs_file *file =
 263                 container_of(ref, struct ib_uverbs_file, ref);
 264
 265         module_put(file->device->ib_dev->owner);
 266         kref_put(&file->device->ref, ib_uverbs_release_dev);
 267
 268         kfree(file);
 269 }
 270
 271 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 272                                     size_t count, loff_t *pos)
 273 {
 274         struct ib_uverbs_event_file *file = filp->private_data;
 275         struct ib_uverbs_event *event;
 276         int eventsz;
 277         int ret = 0;
 278
 279         spin_lock_irq(&file->lock);
 280
 281         while (list_empty(&file->event_list)) {
 282                 spin_unlock_irq(&file->lock);
 283
 284                 if (filp->f_flags & O_NONBLOCK)
 285                         return -EAGAIN;
 286
 287                 if (wait_event_interruptible(file->poll_wait,
 288                                              !list_empty(&file->event_list)))
 289                         return -ERESTARTSYS;
 290
 291                 spin_lock_irq(&file->lock);
 292         }
 293
 294         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 295
 296         if (file->is_async)
 297                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 298         else
 299                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 300
 301         if (eventsz > count) {
 302                 ret   = -EINVAL;
 303                 event = NULL;
 304         } else {
 305                 list_del(file->event_list.next);
 306                 if (event->counter) {
 307                         ++(*event->counter);
 308                         list_del(&event->obj_list);
 309                 }
 310         }
 311
 312         spin_unlock_irq(&file->lock);
 313
 314         if (event) {
 315                 if (copy_to_user(buf, event, eventsz))
 316                         ret = -EFAULT;
 317                 else
 318                         ret = eventsz;
 319         }
 320
 321         kfree(event);
 322
 323         return ret;
 324 }
 325
 326 static unsigned int ib_uverbs_event_poll(struct file *filp,
 327                                          struct poll_table_struct *wait)
 328 {
 329         unsigned int pollflags = 0;
 330         struct ib_uverbs_event_file *file = filp->private_data;
 331
 332         poll_wait(filp, &file->poll_wait, wait);
 333
 334         spin_lock_irq(&file->lock);
 335         if (!list_empty(&file->event_list))
 336                 pollflags = POLLIN | POLLRDNORM;
 337         spin_unlock_irq(&file->lock);
 338
 339         return pollflags;
 340 }
 341
 342 void ib_uverbs_release_event_file(struct kref *ref)
 343 {
 344         struct ib_uverbs_event_file *file =
 345                 container_of(ref, struct ib_uverbs_event_file, ref);
 346
 347         kfree(file);
 348 }
 349
 350 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 351 {
 352         struct ib_uverbs_event_file *file = filp->private_data;
 353
 354         return fasync_helper(fd, filp, on, &file->async_queue);
 355 }
 356
 357 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 358 {
 359         struct ib_uverbs_event_file *file = filp->private_data;
 360         struct ib_uverbs_event *entry, *tmp;
 361
 362         spin_lock_irq(&file->lock);
 363         file->file = NULL;
 364         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 365                 if (entry->counter)
 366                         list_del(&entry->obj_list);
 367                 kfree(entry);
 368         }
 369         spin_unlock_irq(&file->lock);
 370
 371         ib_uverbs_event_fasync(-1, filp, 0);
 372
 373         if (file->is_async) {
 374                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 375                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 376         }
 377         kref_put(&file->ref, ib_uverbs_release_event_file);
 378
 379         return 0;
 380 }
 381
 382 static struct file_operations uverbs_event_fops = {
 383         .owner   = THIS_MODULE,
 384         .read    = ib_uverbs_event_read,
 385         .poll    = ib_uverbs_event_poll,
 386         .release = ib_uverbs_event_close,
 387         .fasync  = ib_uverbs_event_fasync
 388 };
 389
 390 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 391 {
 392         struct ib_uverbs_event_file    *file = cq_context;
 393         struct ib_ucq_object           *uobj;
 394         struct ib_uverbs_event         *entry;
 395         unsigned long                   flags;
 396
 397         if (!file)
 398                 return;
 399
 400         spin_lock_irqsave(&file->lock, flags);
 401         if (!file->file) {
 402                 spin_unlock_irqrestore(&file->lock, flags);
 403                 return;
 404         }
 405
 406         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 407         if (!entry) {
 408                 spin_unlock_irqrestore(&file->lock, flags);
 409                 return;
 410         }
 411
 412         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 413
 414         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 415         entry->counter             = &uobj->comp_events_reported;
 416
 417         list_add_tail(&entry->list, &file->event_list);
 418         list_add_tail(&entry->obj_list, &uobj->comp_list);
 419         spin_unlock_irqrestore(&file->lock, flags);
 420
 421         wake_up_interruptible(&file->poll_wait);
 422         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 423 }
 424
 425 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 426                                     __u64 element, __u64 event,
 427                                     struct list_head *obj_list,
 428                                     u32 *counter)
 429 {
 430         struct ib_uverbs_event *entry;
 431         unsigned long flags;
 432
 433         spin_lock_irqsave(&file->async_file->lock, flags);
 434         if (!file->async_file->file) {
 435                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 436                 return;
 437         }
 438
 439         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 440         if (!entry) {
 441                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 442                 return;
 443         }
 444
 445         entry->desc.async.element    = element;
 446         entry->desc.async.event_type = event;
 447         entry->counter               = counter;
 448
 449         list_add_tail(&entry->list, &file->async_file->event_list);
 450         if (obj_list)
 451                 list_add_tail(&entry->obj_list, obj_list);
 452         spin_unlock_irqrestore(&file->async_file->lock, flags);
 453
 454         wake_up_interruptible(&file->async_file->poll_wait);
 455         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 456 }
 457
 458 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 459 {
 460         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 461                                                   struct ib_ucq_object, uobject);
 462
 463         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 464                                 event->event, &uobj->async_list,
 465                                 &uobj->async_events_reported);
 466 }
 467
 468 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 469 {
 470         struct ib_uevent_object *uobj;
 471
 472         uobj = container_of(event->element.qp->uobject,
 473                             struct ib_uevent_object, uobject);
 474
 475         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 476                                 event->event, &uobj->event_list,
 477                                 &uobj->events_reported);
 478 }
 479
 480 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 481 {
 482         struct ib_uevent_object *uobj;
 483
 484         uobj = container_of(event->element.srq->uobject,
 485                             struct ib_uevent_object, uobject);
 486
 487         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 488                                 event->event, &uobj->event_list,
 489                                 &uobj->events_reported);
 490 }
 491
 492 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 493                              struct ib_event *event)
 494 {
 495         struct ib_uverbs_file *file =
 496                 container_of(handler, struct ib_uverbs_file, event_handler);
 497
 498         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 499                                 NULL, NULL);
 500 }
 501
 502 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 503                                         int is_async, int *fd)
 504 {
 505         struct ib_uverbs_event_file *ev_file;
 506         struct file *filp;
 507         int ret;
 508
 509         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 510         if (!ev_file)
 511                 return ERR_PTR(-ENOMEM);
 512
 513         kref_init(&ev_file->ref);
 514         spin_lock_init(&ev_file->lock);
 515         INIT_LIST_HEAD(&ev_file->event_list);
 516         init_waitqueue_head(&ev_file->poll_wait);
 517         ev_file->uverbs_file = uverbs_file;
 518         ev_file->async_queue = NULL;
 519         ev_file->is_async    = is_async;
 520
 521         *fd = get_unused_fd();
 522         if (*fd < 0) {
 523                 ret = *fd;
 524                 goto err;
 525         }
 526
 527         filp = get_empty_filp();
 528         if (!filp) {
 529                 ret = -ENFILE;
 530                 goto err_fd;
 531         }
 532
 533         ev_file->file      = filp;
 534
 535         /*
 536          * fops_get() can't fail here, because we're coming from a
 537          * system call on a uverbs file, which will already have a
 538          * module reference.
 539          */
 540         filp->f_op         = fops_get(&uverbs_event_fops);
 541         filp->f_vfsmnt     = mntget(uverbs_event_mnt);
 542         filp->f_dentry     = dget(uverbs_event_mnt->mnt_root);
 543         filp->f_mapping    = filp->f_dentry->d_inode->i_mapping;
 544         filp->f_flags      = O_RDONLY;
 545         filp->f_mode       = FMODE_READ;
 546         filp->private_data = ev_file;
 547
 548         return filp;
 549
 550 err_fd:
 551         put_unused_fd(*fd);
 552
 553 err:
 554         kfree(ev_file);
 555         return ERR_PTR(ret);
 556 }
 557
 558 /*
 559  * Look up a completion event file by FD.  If lookup is successful,
 560  * takes a ref to the event file struct that it returns; if
 561  * unsuccessful, returns NULL.
 562  */
 563 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 564 {
 565         struct ib_uverbs_event_file *ev_file = NULL;
 566         struct file *filp;
 567
 568         filp = fget(fd);
 569         if (!filp)
 570                 return NULL;
 571
 572         if (filp->f_op != &uverbs_event_fops)
 573                 goto out;
 574
 575         ev_file = filp->private_data;
 576         if (ev_file->is_async) {
 577                 ev_file = NULL;
 578                 goto out;
 579         }
 580
 581         kref_get(&ev_file->ref);
 582
 583 out:
 584         fput(filp);
 585         return ev_file;
 586 }
 587
 588 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 589                              size_t count, loff_t *pos)
 590 {
 591         struct ib_uverbs_file *file = filp->private_data;
 592         struct ib_uverbs_cmd_hdr hdr;
 593
 594         if (count < sizeof hdr)
 595                 return -EINVAL;
 596
 597         if (copy_from_user(&hdr, buf, sizeof hdr))
 598                 return -EFAULT;
 599
 600         if (hdr.in_words * 4 != count)
 601                 return -EINVAL;
 602
 603         if (hdr.command < 0                             ||
 604             hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 605             !uverbs_cmd_table[hdr.command]              ||
 606             !(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 607                 return -EINVAL;
 608
 609         if (!file->ucontext &&
 610             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 611                 return -EINVAL;
 612
 613         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 614                                              hdr.in_words * 4, hdr.out_words * 4);
 615 }
 616
 617 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 618 {
 619         struct ib_uverbs_file *file = filp->private_data;
 620
 621         if (!file->ucontext)
 622                 return -ENODEV;
 623         else
 624                 return file->device->ib_dev->mmap(file->ucontext, vma);
 625 }
 626
 627 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 628 {
 629         struct ib_uverbs_device *dev;
 630         struct ib_uverbs_file *file;
 631         int ret;
 632
 633         spin_lock(&map_lock);
 634         dev = dev_table[iminor(inode) - IB_UVERBS_BASE_MINOR];
 635         if (dev)
 636                 kref_get(&dev->ref);
 637         spin_unlock(&map_lock);
 638
 639         if (!dev)
 640                 return -ENXIO;
 641
 642         if (!try_module_get(dev->ib_dev->owner)) {
 643                 ret = -ENODEV;
 644                 goto err;
 645         }
 646
 647         file = kmalloc(sizeof *file, GFP_KERNEL);
 648         if (!file) {
 649                 ret = -ENOMEM;
 650                 goto err_module;
 651         }
 652
 653         file->device     = dev;
 654         file->ucontext   = NULL;
 655         file->async_file = NULL;
 656         kref_init(&file->ref);
 657         mutex_init(&file->mutex);
 658
 659         filp->private_data = file;
 660
 661         return 0;
 662
 663 err_module:
 664         module_put(dev->ib_dev->owner);
 665
 666 err:
 667         kref_put(&dev->ref, ib_uverbs_release_dev);
 668
 669         return ret;
 670 }
 671
 672 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 673 {
 674         struct ib_uverbs_file *file = filp->private_data;
 675
 676         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 677
 678         if (file->async_file)
 679                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 680
 681         kref_put(&file->ref, ib_uverbs_release_file);
 682
 683         return 0;
 684 }
 685
 686 static struct file_operations uverbs_fops = {
 687         .owner   = THIS_MODULE,
 688         .write   = ib_uverbs_write,
 689         .open    = ib_uverbs_open,
 690         .release = ib_uverbs_close
 691 };
 692
 693 static struct file_operations uverbs_mmap_fops = {
 694         .owner   = THIS_MODULE,
 695         .write   = ib_uverbs_write,
 696         .mmap    = ib_uverbs_mmap,
 697         .open    = ib_uverbs_open,
 698         .release = ib_uverbs_close
 699 };
 700
 701 static struct ib_client uverbs_client = {
 702         .name   = "uverbs",
 703         .add    = ib_uverbs_add_one,
 704         .remove = ib_uverbs_remove_one
 705 };
 706
 707 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
 708 {
 709         struct ib_uverbs_device *dev = class_get_devdata(class_dev);
 710
 711         if (!dev)
 712                 return -ENODEV;
 713
 714         return sprintf(buf, "%s\n", dev->ib_dev->name);
 715 }
 716 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 717
 718 static ssize_t show_dev_abi_version(struct class_device *class_dev, char *buf)
 719 {
 720         struct ib_uverbs_device *dev = class_get_devdata(class_dev);
 721
 722         if (!dev)
 723                 return -ENODEV;
 724
 725         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 726 }
 727 static CLASS_DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 728
 729 static ssize_t show_abi_version(struct class *class, char *buf)
 730 {
 731         return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
 732 }
 733 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 734
 735 static void ib_uverbs_add_one(struct ib_device *device)
 736 {
 737         struct ib_uverbs_device *uverbs_dev;
 738
 739         if (!device->alloc_ucontext)
 740                 return;
 741
 742         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 743         if (!uverbs_dev)
 744                 return;
 745
 746         kref_init(&uverbs_dev->ref);
 747
 748         spin_lock(&map_lock);
 749         uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 750         if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
 751                 spin_unlock(&map_lock);
 752                 goto err;
 753         }
 754         set_bit(uverbs_dev->devnum, dev_map);
 755         spin_unlock(&map_lock);
 756
 757         uverbs_dev->ib_dev           = device;
 758         uverbs_dev->num_comp_vectors = 1;
 759
 760         uverbs_dev->dev = cdev_alloc();
 761         if (!uverbs_dev->dev)
 762                 goto err;
 763         uverbs_dev->dev->owner = THIS_MODULE;
 764         uverbs_dev->dev->ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 765         kobject_set_name(&uverbs_dev->dev->kobj, "uverbs%d", uverbs_dev->devnum);
 766         if (cdev_add(uverbs_dev->dev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
 767                 goto err_cdev;
 768
 769         uverbs_dev->class_dev = class_device_create(uverbs_class, NULL,
 770                                                     uverbs_dev->dev->dev,
 771                                                     device->dma_device,
 772                                                     "uverbs%d", uverbs_dev->devnum);
 773         if (IS_ERR(uverbs_dev->class_dev))
 774                 goto err_cdev;
 775
 776         class_set_devdata(uverbs_dev->class_dev, uverbs_dev);
 777
 778         if (class_device_create_file(uverbs_dev->class_dev, &class_device_attr_ibdev))
 779                 goto err_class;
 780         if (class_device_create_file(uverbs_dev->class_dev, &class_device_attr_abi_version))
 781                 goto err_class;
 782
 783         spin_lock(&map_lock);
 784         dev_table[uverbs_dev->devnum] = uverbs_dev;
 785         spin_unlock(&map_lock);
 786
 787         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 788
 789         return;
 790
 791 err_class:
 792         class_device_destroy(uverbs_class, uverbs_dev->dev->dev);
 793
 794 err_cdev:
 795         cdev_del(uverbs_dev->dev);
 796         clear_bit(uverbs_dev->devnum, dev_map);
 797
 798 err:
 799         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 800         return;
 801 }
 802
 803 static void ib_uverbs_remove_one(struct ib_device *device)
 804 {
 805         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 806
 807         if (!uverbs_dev)
 808                 return;
 809
 810         class_set_devdata(uverbs_dev->class_dev, NULL);
 811         class_device_destroy(uverbs_class, uverbs_dev->dev->dev);
 812         cdev_del(uverbs_dev->dev);
 813
 814         spin_lock(&map_lock);
 815         dev_table[uverbs_dev->devnum] = NULL;
 816         spin_unlock(&map_lock);
 817
 818         clear_bit(uverbs_dev->devnum, dev_map);
 819         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 820 }
 821
 822 static struct super_block *uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
 823                                                const char *dev_name, void *data)
 824 {
 825         return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
 826                              INFINIBANDEVENTFS_MAGIC);
 827 }
 828
 829 static struct file_system_type uverbs_event_fs = {
 830         /* No owner field so module can be unloaded */
 831         .name    = "infinibandeventfs",
 832         .get_sb  = uverbs_event_get_sb,
 833         .kill_sb = kill_litter_super
 834 };
 835
 836 static int __init ib_uverbs_init(void)
 837 {
 838         int ret;
 839
 840         spin_lock_init(&map_lock);
 841
 842         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 843                                      "infiniband_verbs");
 844         if (ret) {
 845                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 846                 goto out;
 847         }
 848
 849         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 850         if (IS_ERR(uverbs_class)) {
 851                 ret = PTR_ERR(uverbs_class);
 852                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 853                 goto out_chrdev;
 854         }
 855
 856         ret = class_create_file(uverbs_class, &class_attr_abi_version);
 857         if (ret) {
 858                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 859                 goto out_class;
 860         }
 861
 862         ret = register_filesystem(&uverbs_event_fs);
 863         if (ret) {
 864                 printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
 865                 goto out_class;
 866         }
 867
 868         uverbs_event_mnt = kern_mount(&uverbs_event_fs);
 869         if (IS_ERR(uverbs_event_mnt)) {
 870                 ret = PTR_ERR(uverbs_event_mnt);
 871                 printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
 872                 goto out_fs;
 873         }
 874
 875         ret = ib_register_client(&uverbs_client);
 876         if (ret) {
 877                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 878                 goto out_mnt;
 879         }
 880
 881         return 0;
 882
 883 out_mnt:
 884         mntput(uverbs_event_mnt);
 885
 886 out_fs:
 887         unregister_filesystem(&uverbs_event_fs);
 888
 889 out_class:
 890         class_destroy(uverbs_class);
 891
 892 out_chrdev:
 893         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 894
 895 out:
 896         return ret;
 897 }
 898
 899 static void __exit ib_uverbs_cleanup(void)
 900 {
 901         ib_unregister_client(&uverbs_client);
 902         mntput(uverbs_event_mnt);
 903         unregister_filesystem(&uverbs_event_fs);
 904         class_destroy(uverbs_class);
 905         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 906         flush_scheduled_work();
 907         idr_destroy(&ib_uverbs_pd_idr);
 908         idr_destroy(&ib_uverbs_mr_idr);
 909         idr_destroy(&ib_uverbs_mw_idr);
 910         idr_destroy(&ib_uverbs_ah_idr);
 911         idr_destroy(&ib_uverbs_cq_idr);
 912         idr_destroy(&ib_uverbs_qp_idr);
 913         idr_destroy(&ib_uverbs_srq_idr);
 914 }
 915
 916 module_init(ib_uverbs_init);
 917 module_exit(ib_uverbs_cleanup);