drivers/infiniband/core/uverbs_main.c

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