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
  76 static DEFINE_SPINLOCK(map_lock);
  77 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  78
  79 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  80                                      const char __user *buf, int in_len,
  81                                      int out_len) = {
  82         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  83         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  84         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  85         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  86         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  87         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  88         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  89         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  90         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  91         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  92         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  93         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  94         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  95         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  96         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
  97         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
  98         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
  99         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 100         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 101         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 102         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 103         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 104         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 105         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 106         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 107         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 108         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 109         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 110 };
 111
 112 static void ib_uverbs_add_one(struct ib_device *device);
 113 static void ib_uverbs_remove_one(struct ib_device *device);
 114
 115 static void ib_uverbs_release_dev(struct kref *ref)
 116 {
 117         struct ib_uverbs_device *dev =
 118                 container_of(ref, struct ib_uverbs_device, ref);
 119
 120         complete(&dev->comp);
 121 }
 122
 123 static void ib_uverbs_release_event_file(struct kref *ref)
 124 {
 125         struct ib_uverbs_event_file *file =
 126                 container_of(ref, struct ib_uverbs_event_file, ref);
 127
 128         kfree(file);
 129 }
 130
 131 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 132                           struct ib_uverbs_event_file *ev_file,
 133                           struct ib_ucq_object *uobj)
 134 {
 135         struct ib_uverbs_event *evt, *tmp;
 136
 137         if (ev_file) {
 138                 spin_lock_irq(&ev_file->lock);
 139                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 140                         list_del(&evt->list);
 141                         kfree(evt);
 142                 }
 143                 spin_unlock_irq(&ev_file->lock);
 144
 145                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 146         }
 147
 148         spin_lock_irq(&file->async_file->lock);
 149         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 150                 list_del(&evt->list);
 151                 kfree(evt);
 152         }
 153         spin_unlock_irq(&file->async_file->lock);
 154 }
 155
 156 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 157                               struct ib_uevent_object *uobj)
 158 {
 159         struct ib_uverbs_event *evt, *tmp;
 160
 161         spin_lock_irq(&file->async_file->lock);
 162         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 163                 list_del(&evt->list);
 164                 kfree(evt);
 165         }
 166         spin_unlock_irq(&file->async_file->lock);
 167 }
 168
 169 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 170                                     struct ib_uqp_object *uobj)
 171 {
 172         struct ib_uverbs_mcast_entry *mcast, *tmp;
 173
 174         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 175                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 176                 list_del(&mcast->list);
 177                 kfree(mcast);
 178         }
 179 }
 180
 181 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 182                                       struct ib_ucontext *context)
 183 {
 184         struct ib_uobject *uobj, *tmp;
 185
 186         if (!context)
 187                 return 0;
 188
 189         context->closing = 1;
 190
 191         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 192                 struct ib_ah *ah = uobj->object;
 193
 194                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 195                 ib_destroy_ah(ah);
 196                 kfree(uobj);
 197         }
 198
 199         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 200                 struct ib_qp *qp = uobj->object;
 201                 struct ib_uqp_object *uqp =
 202                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 203
 204                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 205                 ib_uverbs_detach_umcast(qp, uqp);
 206                 ib_destroy_qp(qp);
 207                 ib_uverbs_release_uevent(file, &uqp->uevent);
 208                 kfree(uqp);
 209         }
 210
 211         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 212                 struct ib_cq *cq = uobj->object;
 213                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 214                 struct ib_ucq_object *ucq =
 215                         container_of(uobj, struct ib_ucq_object, uobject);
 216
 217                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 218                 ib_destroy_cq(cq);
 219                 ib_uverbs_release_ucq(file, ev_file, ucq);
 220                 kfree(ucq);
 221         }
 222
 223         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 224                 struct ib_srq *srq = uobj->object;
 225                 struct ib_uevent_object *uevent =
 226                         container_of(uobj, struct ib_uevent_object, uobject);
 227
 228                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 229                 ib_destroy_srq(srq);
 230                 ib_uverbs_release_uevent(file, uevent);
 231                 kfree(uevent);
 232         }
 233
 234         /* XXX Free MWs */
 235
 236         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 237                 struct ib_mr *mr = uobj->object;
 238
 239                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 240                 ib_dereg_mr(mr);
 241                 kfree(uobj);
 242         }
 243
 244         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 245                 struct ib_pd *pd = uobj->object;
 246
 247                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 248                 ib_dealloc_pd(pd);
 249                 kfree(uobj);
 250         }
 251
 252         return context->device->dealloc_ucontext(context);
 253 }
 254
 255 static void ib_uverbs_release_file(struct kref *ref)
 256 {
 257         struct ib_uverbs_file *file =
 258                 container_of(ref, struct ib_uverbs_file, ref);
 259
 260         module_put(file->device->ib_dev->owner);
 261         kref_put(&file->device->ref, ib_uverbs_release_dev);
 262
 263         kfree(file);
 264 }
 265
 266 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 267                                     size_t count, loff_t *pos)
 268 {
 269         struct ib_uverbs_event_file *file = filp->private_data;
 270         struct ib_uverbs_event *event;
 271         int eventsz;
 272         int ret = 0;
 273
 274         spin_lock_irq(&file->lock);
 275
 276         while (list_empty(&file->event_list)) {
 277                 spin_unlock_irq(&file->lock);
 278
 279                 if (filp->f_flags & O_NONBLOCK)
 280                         return -EAGAIN;
 281
 282                 if (wait_event_interruptible(file->poll_wait,
 283                                              !list_empty(&file->event_list)))
 284                         return -ERESTARTSYS;
 285
 286                 spin_lock_irq(&file->lock);
 287         }
 288
 289         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 290
 291         if (file->is_async)
 292                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 293         else
 294                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 295
 296         if (eventsz > count) {
 297                 ret   = -EINVAL;
 298                 event = NULL;
 299         } else {
 300                 list_del(file->event_list.next);
 301                 if (event->counter) {
 302                         ++(*event->counter);
 303                         list_del(&event->obj_list);
 304                 }
 305         }
 306
 307         spin_unlock_irq(&file->lock);
 308
 309         if (event) {
 310                 if (copy_to_user(buf, event, eventsz))
 311                         ret = -EFAULT;
 312                 else
 313                         ret = eventsz;
 314         }
 315
 316         kfree(event);
 317
 318         return ret;
 319 }
 320
 321 static unsigned int ib_uverbs_event_poll(struct file *filp,
 322                                          struct poll_table_struct *wait)
 323 {
 324         unsigned int pollflags = 0;
 325         struct ib_uverbs_event_file *file = filp->private_data;
 326
 327         poll_wait(filp, &file->poll_wait, wait);
 328
 329         spin_lock_irq(&file->lock);
 330         if (!list_empty(&file->event_list))
 331                 pollflags = POLLIN | POLLRDNORM;
 332         spin_unlock_irq(&file->lock);
 333
 334         return pollflags;
 335 }
 336
 337 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 338 {
 339         struct ib_uverbs_event_file *file = filp->private_data;
 340
 341         return fasync_helper(fd, filp, on, &file->async_queue);
 342 }
 343
 344 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 345 {
 346         struct ib_uverbs_event_file *file = filp->private_data;
 347         struct ib_uverbs_event *entry, *tmp;
 348
 349         spin_lock_irq(&file->lock);
 350         file->is_closed = 1;
 351         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 352                 if (entry->counter)
 353                         list_del(&entry->obj_list);
 354                 kfree(entry);
 355         }
 356         spin_unlock_irq(&file->lock);
 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 const 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->is_closed) {
 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->is_closed) {
 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 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 454 {
 455         struct ib_uevent_object *uobj;
 456
 457         uobj = container_of(event->element.qp->uobject,
 458                             struct ib_uevent_object, uobject);
 459
 460         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 461                                 event->event, &uobj->event_list,
 462                                 &uobj->events_reported);
 463 }
 464
 465 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 466 {
 467         struct ib_uevent_object *uobj;
 468
 469         uobj = container_of(event->element.srq->uobject,
 470                             struct ib_uevent_object, uobject);
 471
 472         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 473                                 event->event, &uobj->event_list,
 474                                 &uobj->events_reported);
 475 }
 476
 477 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 478                              struct ib_event *event)
 479 {
 480         struct ib_uverbs_file *file =
 481                 container_of(handler, struct ib_uverbs_file, event_handler);
 482
 483         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 484                                 NULL, NULL);
 485 }
 486
 487 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 488                                         int is_async)
 489 {
 490         struct ib_uverbs_event_file *ev_file;
 491         struct file *filp;
 492
 493         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 494         if (!ev_file)
 495                 return ERR_PTR(-ENOMEM);
 496
 497         kref_init(&ev_file->ref);
 498         spin_lock_init(&ev_file->lock);
 499         INIT_LIST_HEAD(&ev_file->event_list);
 500         init_waitqueue_head(&ev_file->poll_wait);
 501         ev_file->uverbs_file = uverbs_file;
 502         ev_file->async_queue = NULL;
 503         ev_file->is_async    = is_async;
 504         ev_file->is_closed   = 0;
 505
 506         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 507                                   ev_file, O_RDONLY);
 508         if (IS_ERR(filp))
 509                 kfree(ev_file);
 510
 511         return filp;
 512 }
 513
 514 /*
 515  * Look up a completion event file by FD.  If lookup is successful,
 516  * takes a ref to the event file struct that it returns; if
 517  * unsuccessful, returns NULL.
 518  */
 519 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 520 {
 521         struct ib_uverbs_event_file *ev_file = NULL;
 522         struct file *filp;
 523
 524         filp = fget(fd);
 525         if (!filp)
 526                 return NULL;
 527
 528         if (filp->f_op != &uverbs_event_fops)
 529                 goto out;
 530
 531         ev_file = filp->private_data;
 532         if (ev_file->is_async) {
 533                 ev_file = NULL;
 534                 goto out;
 535         }
 536
 537         kref_get(&ev_file->ref);
 538
 539 out:
 540         fput(filp);
 541         return ev_file;
 542 }
 543
 544 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 545                              size_t count, loff_t *pos)
 546 {
 547         struct ib_uverbs_file *file = filp->private_data;
 548         struct ib_uverbs_cmd_hdr hdr;
 549
 550         if (count < sizeof hdr)
 551                 return -EINVAL;
 552
 553         if (copy_from_user(&hdr, buf, sizeof hdr))
 554                 return -EFAULT;
 555
 556         if (hdr.in_words * 4 != count)
 557                 return -EINVAL;
 558
 559         if (hdr.command < 0                             ||
 560             hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 561             !uverbs_cmd_table[hdr.command])
 562                 return -EINVAL;
 563
 564         if (!file->ucontext &&
 565             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 566                 return -EINVAL;
 567
 568         if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 569                 return -ENOSYS;
 570
 571         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 572                                              hdr.in_words * 4, hdr.out_words * 4);
 573 }
 574
 575 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 576 {
 577         struct ib_uverbs_file *file = filp->private_data;
 578
 579         if (!file->ucontext)
 580                 return -ENODEV;
 581         else
 582                 return file->device->ib_dev->mmap(file->ucontext, vma);
 583 }
 584
 585 /*
 586  * ib_uverbs_open() does not need the BKL:
 587  *
 588  *  - the ib_uverbs_device structures are properly reference counted and
 589  *    everything else is purely local to the file being created, so
 590  *    races against other open calls are not a problem;
 591  *  - there is no ioctl method to race against;
 592  *  - the open method will either immediately run -ENXIO, or all
 593  *    required initialization will be done.
 594  */
 595 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 596 {
 597         struct ib_uverbs_device *dev;
 598         struct ib_uverbs_file *file;
 599         int ret;
 600
 601         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 602         if (dev)
 603                 kref_get(&dev->ref);
 604         else
 605                 return -ENXIO;
 606
 607         if (!try_module_get(dev->ib_dev->owner)) {
 608                 ret = -ENODEV;
 609                 goto err;
 610         }
 611
 612         file = kmalloc(sizeof *file, GFP_KERNEL);
 613         if (!file) {
 614                 ret = -ENOMEM;
 615                 goto err_module;
 616         }
 617
 618         file->device     = dev;
 619         file->ucontext   = NULL;
 620         file->async_file = NULL;
 621         kref_init(&file->ref);
 622         mutex_init(&file->mutex);
 623
 624         filp->private_data = file;
 625
 626         return 0;
 627
 628 err_module:
 629         module_put(dev->ib_dev->owner);
 630
 631 err:
 632         kref_put(&dev->ref, ib_uverbs_release_dev);
 633         return ret;
 634 }
 635
 636 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 637 {
 638         struct ib_uverbs_file *file = filp->private_data;
 639
 640         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 641
 642         if (file->async_file)
 643                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 644
 645         kref_put(&file->ref, ib_uverbs_release_file);
 646
 647         return 0;
 648 }
 649
 650 static const struct file_operations uverbs_fops = {
 651         .owner   = THIS_MODULE,
 652         .write   = ib_uverbs_write,
 653         .open    = ib_uverbs_open,
 654         .release = ib_uverbs_close
 655 };
 656
 657 static const struct file_operations uverbs_mmap_fops = {
 658         .owner   = THIS_MODULE,
 659         .write   = ib_uverbs_write,
 660         .mmap    = ib_uverbs_mmap,
 661         .open    = ib_uverbs_open,
 662         .release = ib_uverbs_close
 663 };
 664
 665 static struct ib_client uverbs_client = {
 666         .name   = "uverbs",
 667         .add    = ib_uverbs_add_one,
 668         .remove = ib_uverbs_remove_one
 669 };
 670
 671 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 672                           char *buf)
 673 {
 674         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 675
 676         if (!dev)
 677                 return -ENODEV;
 678
 679         return sprintf(buf, "%s\n", dev->ib_dev->name);
 680 }
 681 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 682
 683 static ssize_t show_dev_abi_version(struct device *device,
 684                                     struct device_attribute *attr, char *buf)
 685 {
 686         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 687
 688         if (!dev)
 689                 return -ENODEV;
 690
 691         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 692 }
 693 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 694
 695 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
 696                          __stringify(IB_USER_VERBS_ABI_VERSION));
 697
 698 static dev_t overflow_maj;
 699 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
 700
 701 /*
 702  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
 703  * requesting a new major number and doubling the number of max devices we
 704  * support. It's stupid, but simple.
 705  */
 706 static int find_overflow_devnum(void)
 707 {
 708         int ret;
 709
 710         if (!overflow_maj) {
 711                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
 712                                           "infiniband_verbs");
 713                 if (ret) {
 714                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
 715                         return ret;
 716                 }
 717         }
 718
 719         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
 720         if (ret >= IB_UVERBS_MAX_DEVICES)
 721                 return -1;
 722
 723         return ret;
 724 }
 725
 726 static void ib_uverbs_add_one(struct ib_device *device)
 727 {
 728         int devnum;
 729         dev_t base;
 730         struct ib_uverbs_device *uverbs_dev;
 731
 732         if (!device->alloc_ucontext)
 733                 return;
 734
 735         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 736         if (!uverbs_dev)
 737                 return;
 738
 739         kref_init(&uverbs_dev->ref);
 740         init_completion(&uverbs_dev->comp);
 741
 742         spin_lock(&map_lock);
 743         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 744         if (devnum >= IB_UVERBS_MAX_DEVICES) {
 745                 spin_unlock(&map_lock);
 746                 devnum = find_overflow_devnum();
 747                 if (devnum < 0)
 748                         goto err;
 749
 750                 spin_lock(&map_lock);
 751                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
 752                 base = devnum + overflow_maj;
 753                 set_bit(devnum, overflow_map);
 754         } else {
 755                 uverbs_dev->devnum = devnum;
 756                 base = devnum + IB_UVERBS_BASE_DEV;
 757                 set_bit(devnum, dev_map);
 758         }
 759         spin_unlock(&map_lock);
 760
 761         uverbs_dev->ib_dev           = device;
 762         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 763
 764         cdev_init(&uverbs_dev->cdev, NULL);
 765         uverbs_dev->cdev.owner = THIS_MODULE;
 766         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 767         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
 768         if (cdev_add(&uverbs_dev->cdev, base, 1))
 769                 goto err_cdev;
 770
 771         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 772                                         uverbs_dev->cdev.dev, uverbs_dev,
 773                                         "uverbs%d", uverbs_dev->devnum);
 774         if (IS_ERR(uverbs_dev->dev))
 775                 goto err_cdev;
 776
 777         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 778                 goto err_class;
 779         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 780                 goto err_class;
 781
 782         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 783
 784         return;
 785
 786 err_class:
 787         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 788
 789 err_cdev:
 790         cdev_del(&uverbs_dev->cdev);
 791         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 792                 clear_bit(devnum, dev_map);
 793         else
 794                 clear_bit(devnum, overflow_map);
 795
 796 err:
 797         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 798         wait_for_completion(&uverbs_dev->comp);
 799         kfree(uverbs_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         dev_set_drvdata(uverbs_dev->dev, NULL);
 811         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 812         cdev_del(&uverbs_dev->cdev);
 813
 814         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 815                 clear_bit(uverbs_dev->devnum, dev_map);
 816         else
 817                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
 818
 819         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 820         wait_for_completion(&uverbs_dev->comp);
 821         kfree(uverbs_dev);
 822 }
 823
 824 static int __init ib_uverbs_init(void)
 825 {
 826         int ret;
 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.attr);
 843         if (ret) {
 844                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 845                 goto out_class;
 846         }
 847
 848         ret = ib_register_client(&uverbs_client);
 849         if (ret) {
 850                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 851                 goto out_class;
 852         }
 853
 854         return 0;
 855
 856 out_class:
 857         class_destroy(uverbs_class);
 858
 859 out_chrdev:
 860         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 861
 862 out:
 863         return ret;
 864 }
 865
 866 static void __exit ib_uverbs_cleanup(void)
 867 {
 868         ib_unregister_client(&uverbs_client);
 869         class_destroy(uverbs_class);
 870         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 871         if (overflow_maj)
 872                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
 873         idr_destroy(&ib_uverbs_pd_idr);
 874         idr_destroy(&ib_uverbs_mr_idr);
 875         idr_destroy(&ib_uverbs_mw_idr);
 876         idr_destroy(&ib_uverbs_ah_idr);
 877         idr_destroy(&ib_uverbs_cq_idr);
 878         idr_destroy(&ib_uverbs_qp_idr);
 879         idr_destroy(&ib_uverbs_srq_idr);
 880 }
 881
 882 module_init(ib_uverbs_init);
 883 module_exit(ib_uverbs_cleanup);