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 #include <linux/smp_lock.h>
  49
  50 #include <asm/uaccess.h>
  51
  52 #include "uverbs.h"
  53
  54 MODULE_AUTHOR("Roland Dreier");
  55 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  56 MODULE_LICENSE("Dual BSD/GPL");
  57
  58 #define INFINIBANDEVENTFS_MAGIC 0x49426576      /* "IBev" */
  59
  60 enum {
  61         IB_UVERBS_MAJOR       = 231,
  62         IB_UVERBS_BASE_MINOR  = 192,
  63         IB_UVERBS_MAX_DEVICES = 32
  64 };
  65
  66 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  67
  68 static struct class *uverbs_class;
  69
  70 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  71 DEFINE_IDR(ib_uverbs_pd_idr);
  72 DEFINE_IDR(ib_uverbs_mr_idr);
  73 DEFINE_IDR(ib_uverbs_mw_idr);
  74 DEFINE_IDR(ib_uverbs_ah_idr);
  75 DEFINE_IDR(ib_uverbs_cq_idr);
  76 DEFINE_IDR(ib_uverbs_qp_idr);
  77 DEFINE_IDR(ib_uverbs_srq_idr);
  78
  79 static spinlock_t map_lock;
  80 static struct ib_uverbs_device *dev_table[IB_UVERBS_MAX_DEVICES];
  81 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  82
  83 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  84                                      const char __user *buf, int in_len,
  85                                      int out_len) = {
  86         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  87         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  88         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  89         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  90         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  91         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  92         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  93         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  94         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  95         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  96         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  97         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  98         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  99         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
 100         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 101         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 102         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 103         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 104         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 105         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 106         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 107         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 108         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 109         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 110         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 111         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 112         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 113         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 114 };
 115
 116 static struct vfsmount *uverbs_event_mnt;
 117
 118 static void ib_uverbs_add_one(struct ib_device *device);
 119 static void ib_uverbs_remove_one(struct ib_device *device);
 120
 121 static void ib_uverbs_release_dev(struct kref *ref)
 122 {
 123         struct ib_uverbs_device *dev =
 124                 container_of(ref, struct ib_uverbs_device, ref);
 125
 126         complete(&dev->comp);
 127 }
 128
 129 static void ib_uverbs_release_event_file(struct kref *ref)
 130 {
 131         struct ib_uverbs_event_file *file =
 132                 container_of(ref, struct ib_uverbs_event_file, ref);
 133
 134         kfree(file);
 135 }
 136
 137 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 138                           struct ib_uverbs_event_file *ev_file,
 139                           struct ib_ucq_object *uobj)
 140 {
 141         struct ib_uverbs_event *evt, *tmp;
 142
 143         if (ev_file) {
 144                 spin_lock_irq(&ev_file->lock);
 145                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 146                         list_del(&evt->list);
 147                         kfree(evt);
 148                 }
 149                 spin_unlock_irq(&ev_file->lock);
 150
 151                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 152         }
 153
 154         spin_lock_irq(&file->async_file->lock);
 155         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 156                 list_del(&evt->list);
 157                 kfree(evt);
 158         }
 159         spin_unlock_irq(&file->async_file->lock);
 160 }
 161
 162 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 163                               struct ib_uevent_object *uobj)
 164 {
 165         struct ib_uverbs_event *evt, *tmp;
 166
 167         spin_lock_irq(&file->async_file->lock);
 168         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 169                 list_del(&evt->list);
 170                 kfree(evt);
 171         }
 172         spin_unlock_irq(&file->async_file->lock);
 173 }
 174
 175 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 176                                     struct ib_uqp_object *uobj)
 177 {
 178         struct ib_uverbs_mcast_entry *mcast, *tmp;
 179
 180         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 181                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 182                 list_del(&mcast->list);
 183                 kfree(mcast);
 184         }
 185 }
 186
 187 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 188                                       struct ib_ucontext *context)
 189 {
 190         struct ib_uobject *uobj, *tmp;
 191
 192         if (!context)
 193                 return 0;
 194
 195         context->closing = 1;
 196
 197         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 198                 struct ib_ah *ah = uobj->object;
 199
 200                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 201                 ib_destroy_ah(ah);
 202                 kfree(uobj);
 203         }
 204
 205         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 206                 struct ib_qp *qp = uobj->object;
 207                 struct ib_uqp_object *uqp =
 208                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 209
 210                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 211                 ib_uverbs_detach_umcast(qp, uqp);
 212                 ib_destroy_qp(qp);
 213                 ib_uverbs_release_uevent(file, &uqp->uevent);
 214                 kfree(uqp);
 215         }
 216
 217         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 218                 struct ib_cq *cq = uobj->object;
 219                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 220                 struct ib_ucq_object *ucq =
 221                         container_of(uobj, struct ib_ucq_object, uobject);
 222
 223                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 224                 ib_destroy_cq(cq);
 225                 ib_uverbs_release_ucq(file, ev_file, ucq);
 226                 kfree(ucq);
 227         }
 228
 229         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 230                 struct ib_srq *srq = uobj->object;
 231                 struct ib_uevent_object *uevent =
 232                         container_of(uobj, struct ib_uevent_object, uobject);
 233
 234                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 235                 ib_destroy_srq(srq);
 236                 ib_uverbs_release_uevent(file, uevent);
 237                 kfree(uevent);
 238         }
 239
 240         /* XXX Free MWs */
 241
 242         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 243                 struct ib_mr *mr = uobj->object;
 244
 245                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 246                 ib_dereg_mr(mr);
 247                 kfree(uobj);
 248         }
 249
 250         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 251                 struct ib_pd *pd = uobj->object;
 252
 253                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 254                 ib_dealloc_pd(pd);
 255                 kfree(uobj);
 256         }
 257
 258         return context->device->dealloc_ucontext(context);
 259 }
 260
 261 static void ib_uverbs_release_file(struct kref *ref)
 262 {
 263         struct ib_uverbs_file *file =
 264                 container_of(ref, struct ib_uverbs_file, ref);
 265
 266         module_put(file->device->ib_dev->owner);
 267         kref_put(&file->device->ref, ib_uverbs_release_dev);
 268
 269         kfree(file);
 270 }
 271
 272 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 273                                     size_t count, loff_t *pos)
 274 {
 275         struct ib_uverbs_event_file *file = filp->private_data;
 276         struct ib_uverbs_event *event;
 277         int eventsz;
 278         int ret = 0;
 279
 280         spin_lock_irq(&file->lock);
 281
 282         while (list_empty(&file->event_list)) {
 283                 spin_unlock_irq(&file->lock);
 284
 285                 if (filp->f_flags & O_NONBLOCK)
 286                         return -EAGAIN;
 287
 288                 if (wait_event_interruptible(file->poll_wait,
 289                                              !list_empty(&file->event_list)))
 290                         return -ERESTARTSYS;
 291
 292                 spin_lock_irq(&file->lock);
 293         }
 294
 295         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 296
 297         if (file->is_async)
 298                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 299         else
 300                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 301
 302         if (eventsz > count) {
 303                 ret   = -EINVAL;
 304                 event = NULL;
 305         } else {
 306                 list_del(file->event_list.next);
 307                 if (event->counter) {
 308                         ++(*event->counter);
 309                         list_del(&event->obj_list);
 310                 }
 311         }
 312
 313         spin_unlock_irq(&file->lock);
 314
 315         if (event) {
 316                 if (copy_to_user(buf, event, eventsz))
 317                         ret = -EFAULT;
 318                 else
 319                         ret = eventsz;
 320         }
 321
 322         kfree(event);
 323
 324         return ret;
 325 }
 326
 327 static unsigned int ib_uverbs_event_poll(struct file *filp,
 328                                          struct poll_table_struct *wait)
 329 {
 330         unsigned int pollflags = 0;
 331         struct ib_uverbs_event_file *file = filp->private_data;
 332
 333         poll_wait(filp, &file->poll_wait, wait);
 334
 335         spin_lock_irq(&file->lock);
 336         if (!list_empty(&file->event_list))
 337                 pollflags = POLLIN | POLLRDNORM;
 338         spin_unlock_irq(&file->lock);
 339
 340         return pollflags;
 341 }
 342
 343 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 344 {
 345         struct ib_uverbs_event_file *file = filp->private_data;
 346
 347         return fasync_helper(fd, filp, on, &file->async_queue);
 348 }
 349
 350 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 351 {
 352         struct ib_uverbs_event_file *file = filp->private_data;
 353         struct ib_uverbs_event *entry, *tmp;
 354
 355         spin_lock_irq(&file->lock);
 356         file->is_closed = 1;
 357         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 358                 if (entry->counter)
 359                         list_del(&entry->obj_list);
 360                 kfree(entry);
 361         }
 362         spin_unlock_irq(&file->lock);
 363
 364         ib_uverbs_event_fasync(-1, filp, 0);
 365
 366         if (file->is_async) {
 367                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 368                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 369         }
 370         kref_put(&file->ref, ib_uverbs_release_event_file);
 371
 372         return 0;
 373 }
 374
 375 static const struct file_operations uverbs_event_fops = {
 376         .owner   = THIS_MODULE,
 377         .read    = ib_uverbs_event_read,
 378         .poll    = ib_uverbs_event_poll,
 379         .release = ib_uverbs_event_close,
 380         .fasync  = ib_uverbs_event_fasync
 381 };
 382
 383 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 384 {
 385         struct ib_uverbs_event_file    *file = cq_context;
 386         struct ib_ucq_object           *uobj;
 387         struct ib_uverbs_event         *entry;
 388         unsigned long                   flags;
 389
 390         if (!file)
 391                 return;
 392
 393         spin_lock_irqsave(&file->lock, flags);
 394         if (file->is_closed) {
 395                 spin_unlock_irqrestore(&file->lock, flags);
 396                 return;
 397         }
 398
 399         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 400         if (!entry) {
 401                 spin_unlock_irqrestore(&file->lock, flags);
 402                 return;
 403         }
 404
 405         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 406
 407         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 408         entry->counter             = &uobj->comp_events_reported;
 409
 410         list_add_tail(&entry->list, &file->event_list);
 411         list_add_tail(&entry->obj_list, &uobj->comp_list);
 412         spin_unlock_irqrestore(&file->lock, flags);
 413
 414         wake_up_interruptible(&file->poll_wait);
 415         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 416 }
 417
 418 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 419                                     __u64 element, __u64 event,
 420                                     struct list_head *obj_list,
 421                                     u32 *counter)
 422 {
 423         struct ib_uverbs_event *entry;
 424         unsigned long flags;
 425
 426         spin_lock_irqsave(&file->async_file->lock, flags);
 427         if (!file->async_file->is_closed) {
 428                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 429                 return;
 430         }
 431
 432         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 433         if (!entry) {
 434                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 435                 return;
 436         }
 437
 438         entry->desc.async.element    = element;
 439         entry->desc.async.event_type = event;
 440         entry->counter               = counter;
 441
 442         list_add_tail(&entry->list, &file->async_file->event_list);
 443         if (obj_list)
 444                 list_add_tail(&entry->obj_list, obj_list);
 445         spin_unlock_irqrestore(&file->async_file->lock, flags);
 446
 447         wake_up_interruptible(&file->async_file->poll_wait);
 448         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 449 }
 450
 451 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 452 {
 453         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 454                                                   struct ib_ucq_object, uobject);
 455
 456         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 457                                 event->event, &uobj->async_list,
 458                                 &uobj->async_events_reported);
 459 }
 460
 461 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 462 {
 463         struct ib_uevent_object *uobj;
 464
 465         uobj = container_of(event->element.qp->uobject,
 466                             struct ib_uevent_object, uobject);
 467
 468         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 469                                 event->event, &uobj->event_list,
 470                                 &uobj->events_reported);
 471 }
 472
 473 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 474 {
 475         struct ib_uevent_object *uobj;
 476
 477         uobj = container_of(event->element.srq->uobject,
 478                             struct ib_uevent_object, uobject);
 479
 480         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 481                                 event->event, &uobj->event_list,
 482                                 &uobj->events_reported);
 483 }
 484
 485 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 486                              struct ib_event *event)
 487 {
 488         struct ib_uverbs_file *file =
 489                 container_of(handler, struct ib_uverbs_file, event_handler);
 490
 491         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 492                                 NULL, NULL);
 493 }
 494
 495 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 496                                         int is_async, int *fd)
 497 {
 498         struct ib_uverbs_event_file *ev_file;
 499         struct file *filp;
 500         int ret;
 501
 502         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 503         if (!ev_file)
 504                 return ERR_PTR(-ENOMEM);
 505
 506         kref_init(&ev_file->ref);
 507         spin_lock_init(&ev_file->lock);
 508         INIT_LIST_HEAD(&ev_file->event_list);
 509         init_waitqueue_head(&ev_file->poll_wait);
 510         ev_file->uverbs_file = uverbs_file;
 511         ev_file->async_queue = NULL;
 512         ev_file->is_async    = is_async;
 513         ev_file->is_closed   = 0;
 514
 515         *fd = get_unused_fd();
 516         if (*fd < 0) {
 517                 ret = *fd;
 518                 goto err;
 519         }
 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 = alloc_file(uverbs_event_mnt, dget(uverbs_event_mnt->mnt_root),
 527                           FMODE_READ, fops_get(&uverbs_event_fops));
 528         if (!filp) {
 529                 ret = -ENFILE;
 530                 goto err_fd;
 531         }
 532
 533         filp->private_data = ev_file;
 534
 535         return filp;
 536
 537 err_fd:
 538         put_unused_fd(*fd);
 539
 540 err:
 541         kfree(ev_file);
 542         return ERR_PTR(ret);
 543 }
 544
 545 /*
 546  * Look up a completion event file by FD.  If lookup is successful,
 547  * takes a ref to the event file struct that it returns; if
 548  * unsuccessful, returns NULL.
 549  */
 550 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 551 {
 552         struct ib_uverbs_event_file *ev_file = NULL;
 553         struct file *filp;
 554
 555         filp = fget(fd);
 556         if (!filp)
 557                 return NULL;
 558
 559         if (filp->f_op != &uverbs_event_fops)
 560                 goto out;
 561
 562         ev_file = filp->private_data;
 563         if (ev_file->is_async) {
 564                 ev_file = NULL;
 565                 goto out;
 566         }
 567
 568         kref_get(&ev_file->ref);
 569
 570 out:
 571         fput(filp);
 572         return ev_file;
 573 }
 574
 575 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 576                              size_t count, loff_t *pos)
 577 {
 578         struct ib_uverbs_file *file = filp->private_data;
 579         struct ib_uverbs_cmd_hdr hdr;
 580
 581         if (count < sizeof hdr)
 582                 return -EINVAL;
 583
 584         if (copy_from_user(&hdr, buf, sizeof hdr))
 585                 return -EFAULT;
 586
 587         if (hdr.in_words * 4 != count)
 588                 return -EINVAL;
 589
 590         if (hdr.command < 0                             ||
 591             hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 592             !uverbs_cmd_table[hdr.command]              ||
 593             !(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 594                 return -EINVAL;
 595
 596         if (!file->ucontext &&
 597             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 598                 return -EINVAL;
 599
 600         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 601                                              hdr.in_words * 4, hdr.out_words * 4);
 602 }
 603
 604 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 605 {
 606         struct ib_uverbs_file *file = filp->private_data;
 607
 608         if (!file->ucontext)
 609                 return -ENODEV;
 610         else
 611                 return file->device->ib_dev->mmap(file->ucontext, vma);
 612 }
 613
 614 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 615 {
 616         struct ib_uverbs_device *dev;
 617         struct ib_uverbs_file *file;
 618         int ret;
 619
 620         lock_kernel();
 621         spin_lock(&map_lock);
 622         dev = dev_table[iminor(inode) - IB_UVERBS_BASE_MINOR];
 623         if (dev)
 624                 kref_get(&dev->ref);
 625         spin_unlock(&map_lock);
 626
 627         if (!dev) {
 628                 unlock_kernel();
 629                 return -ENXIO;
 630         }
 631
 632         if (!try_module_get(dev->ib_dev->owner)) {
 633                 ret = -ENODEV;
 634                 goto err;
 635         }
 636
 637         file = kmalloc(sizeof *file, GFP_KERNEL);
 638         if (!file) {
 639                 ret = -ENOMEM;
 640                 goto err_module;
 641         }
 642
 643         file->device     = dev;
 644         file->ucontext   = NULL;
 645         file->async_file = NULL;
 646         kref_init(&file->ref);
 647         mutex_init(&file->mutex);
 648
 649         filp->private_data = file;
 650
 651         unlock_kernel();
 652         return 0;
 653
 654 err_module:
 655         module_put(dev->ib_dev->owner);
 656
 657 err:
 658         kref_put(&dev->ref, ib_uverbs_release_dev);
 659         unlock_kernel();
 660         return ret;
 661 }
 662
 663 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 664 {
 665         struct ib_uverbs_file *file = filp->private_data;
 666
 667         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 668
 669         if (file->async_file)
 670                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 671
 672         kref_put(&file->ref, ib_uverbs_release_file);
 673
 674         return 0;
 675 }
 676
 677 static const struct file_operations uverbs_fops = {
 678         .owner   = THIS_MODULE,
 679         .write   = ib_uverbs_write,
 680         .open    = ib_uverbs_open,
 681         .release = ib_uverbs_close
 682 };
 683
 684 static const struct file_operations uverbs_mmap_fops = {
 685         .owner   = THIS_MODULE,
 686         .write   = ib_uverbs_write,
 687         .mmap    = ib_uverbs_mmap,
 688         .open    = ib_uverbs_open,
 689         .release = ib_uverbs_close
 690 };
 691
 692 static struct ib_client uverbs_client = {
 693         .name   = "uverbs",
 694         .add    = ib_uverbs_add_one,
 695         .remove = ib_uverbs_remove_one
 696 };
 697
 698 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 699                           char *buf)
 700 {
 701         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 702
 703         if (!dev)
 704                 return -ENODEV;
 705
 706         return sprintf(buf, "%s\n", dev->ib_dev->name);
 707 }
 708 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 709
 710 static ssize_t show_dev_abi_version(struct device *device,
 711                                     struct device_attribute *attr, char *buf)
 712 {
 713         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 714
 715         if (!dev)
 716                 return -ENODEV;
 717
 718         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 719 }
 720 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 721
 722 static ssize_t show_abi_version(struct class *class, char *buf)
 723 {
 724         return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
 725 }
 726 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 727
 728 static void ib_uverbs_add_one(struct ib_device *device)
 729 {
 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         uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 744         if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
 745                 spin_unlock(&map_lock);
 746                 goto err;
 747         }
 748         set_bit(uverbs_dev->devnum, dev_map);
 749         spin_unlock(&map_lock);
 750
 751         uverbs_dev->ib_dev           = device;
 752         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 753
 754         uverbs_dev->cdev = cdev_alloc();
 755         if (!uverbs_dev->cdev)
 756                 goto err;
 757         uverbs_dev->cdev->owner = THIS_MODULE;
 758         uverbs_dev->cdev->ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 759         kobject_set_name(&uverbs_dev->cdev->kobj, "uverbs%d", uverbs_dev->devnum);
 760         if (cdev_add(uverbs_dev->cdev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
 761                 goto err_cdev;
 762
 763         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 764                                         uverbs_dev->cdev->dev,
 765                                         "uverbs%d", uverbs_dev->devnum);
 766         if (IS_ERR(uverbs_dev->dev))
 767                 goto err_cdev;
 768
 769         dev_set_drvdata(uverbs_dev->dev, uverbs_dev);
 770
 771         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 772                 goto err_class;
 773         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 774                 goto err_class;
 775
 776         spin_lock(&map_lock);
 777         dev_table[uverbs_dev->devnum] = uverbs_dev;
 778         spin_unlock(&map_lock);
 779
 780         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 781
 782         return;
 783
 784 err_class:
 785         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 786
 787 err_cdev:
 788         cdev_del(uverbs_dev->cdev);
 789         clear_bit(uverbs_dev->devnum, dev_map);
 790
 791 err:
 792         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 793         wait_for_completion(&uverbs_dev->comp);
 794         kfree(uverbs_dev);
 795         return;
 796 }
 797
 798 static void ib_uverbs_remove_one(struct ib_device *device)
 799 {
 800         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 801
 802         if (!uverbs_dev)
 803                 return;
 804
 805         dev_set_drvdata(uverbs_dev->dev, NULL);
 806         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 807         cdev_del(uverbs_dev->cdev);
 808
 809         spin_lock(&map_lock);
 810         dev_table[uverbs_dev->devnum] = NULL;
 811         spin_unlock(&map_lock);
 812
 813         clear_bit(uverbs_dev->devnum, dev_map);
 814
 815         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 816         wait_for_completion(&uverbs_dev->comp);
 817         kfree(uverbs_dev);
 818 }
 819
 820 static int uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
 821                                const char *dev_name, void *data,
 822                                struct vfsmount *mnt)
 823 {
 824         return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
 825                              INFINIBANDEVENTFS_MAGIC, mnt);
 826 }
 827
 828 static struct file_system_type uverbs_event_fs = {
 829         /* No owner field so module can be unloaded */
 830         .name    = "infinibandeventfs",
 831         .get_sb  = uverbs_event_get_sb,
 832         .kill_sb = kill_litter_super
 833 };
 834
 835 static int __init ib_uverbs_init(void)
 836 {
 837         int ret;
 838
 839         spin_lock_init(&map_lock);
 840
 841         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 842                                      "infiniband_verbs");
 843         if (ret) {
 844                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 845                 goto out;
 846         }
 847
 848         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 849         if (IS_ERR(uverbs_class)) {
 850                 ret = PTR_ERR(uverbs_class);
 851                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 852                 goto out_chrdev;
 853         }
 854
 855         ret = class_create_file(uverbs_class, &class_attr_abi_version);
 856         if (ret) {
 857                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 858                 goto out_class;
 859         }
 860
 861         ret = register_filesystem(&uverbs_event_fs);
 862         if (ret) {
 863                 printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
 864                 goto out_class;
 865         }
 866
 867         uverbs_event_mnt = kern_mount(&uverbs_event_fs);
 868         if (IS_ERR(uverbs_event_mnt)) {
 869                 ret = PTR_ERR(uverbs_event_mnt);
 870                 printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
 871                 goto out_fs;
 872         }
 873
 874         ret = ib_register_client(&uverbs_client);
 875         if (ret) {
 876                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 877                 goto out_mnt;
 878         }
 879
 880         return 0;
 881
 882 out_mnt:
 883         mntput(uverbs_event_mnt);
 884
 885 out_fs:
 886         unregister_filesystem(&uverbs_event_fs);
 887
 888 out_class:
 889         class_destroy(uverbs_class);
 890
 891 out_chrdev:
 892         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 893
 894 out:
 895         return ret;
 896 }
 897
 898 static void __exit ib_uverbs_cleanup(void)
 899 {
 900         ib_unregister_client(&uverbs_client);
 901         mntput(uverbs_event_mnt);
 902         unregister_filesystem(&uverbs_event_fs);
 903         class_destroy(uverbs_class);
 904         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 905         idr_destroy(&ib_uverbs_pd_idr);
 906         idr_destroy(&ib_uverbs_mr_idr);
 907         idr_destroy(&ib_uverbs_mw_idr);
 908         idr_destroy(&ib_uverbs_ah_idr);
 909         idr_destroy(&ib_uverbs_cq_idr);
 910         idr_destroy(&ib_uverbs_qp_idr);
 911         idr_destroy(&ib_uverbs_srq_idr);
 912 }
 913
 914 module_init(ib_uverbs_init);
 915 module_exit(ib_uverbs_cleanup);