drivers/infiniband/core/uverbs_main.c

   1 /*
   2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
   6  * Copyright (c) 2005 PathScale, Inc. All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  *
  36  * $Id: uverbs_main.c 2733 2005-06-28 19:14:34Z roland $
  37  */
  38
  39 #include <linux/module.h>
  40 #include <linux/init.h>
  41 #include <linux/device.h>
  42 #include <linux/err.h>
  43 #include <linux/fs.h>
  44 #include <linux/poll.h>
  45 #include <linux/file.h>
  46 #include <linux/mount.h>
  47 #include <linux/cdev.h>
  48
  49 #include <asm/uaccess.h>
  50
  51 #include "uverbs.h"
  52
  53 MODULE_AUTHOR("Roland Dreier");
  54 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  55 MODULE_LICENSE("Dual BSD/GPL");
  56
  57 #define INFINIBANDEVENTFS_MAGIC 0x49426576      /* "IBev" */
  58
  59 enum {
  60         IB_UVERBS_MAJOR       = 231,
  61         IB_UVERBS_BASE_MINOR  = 192,
  62         IB_UVERBS_MAX_DEVICES = 32
  63 };
  64
  65 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  66
  67 static struct class *uverbs_class;
  68
  69 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  70 DEFINE_IDR(ib_uverbs_pd_idr);
  71 DEFINE_IDR(ib_uverbs_mr_idr);
  72 DEFINE_IDR(ib_uverbs_mw_idr);
  73 DEFINE_IDR(ib_uverbs_ah_idr);
  74 DEFINE_IDR(ib_uverbs_cq_idr);
  75 DEFINE_IDR(ib_uverbs_qp_idr);
  76 DEFINE_IDR(ib_uverbs_srq_idr);
  77
  78 static spinlock_t map_lock;
  79 static struct ib_uverbs_device *dev_table[IB_UVERBS_MAX_DEVICES];
  80 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  81
  82 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  83                                      const char __user *buf, int in_len,
  84                                      int out_len) = {
  85         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  86         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  87         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  88         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  89         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  90         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  91         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  92         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  93         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  94         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  95         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  96         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  97         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  98         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  99         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 100         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 101         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 102         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 103         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 104         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 105         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 106         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 107         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 108         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 109         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 110         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 111         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 112         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 113 };
 114
 115 static struct vfsmount *uverbs_event_mnt;
 116
 117 static void ib_uverbs_add_one(struct ib_device *device);
 118 static void ib_uverbs_remove_one(struct ib_device *device);
 119
 120 static void ib_uverbs_release_dev(struct kref *ref)
 121 {
 122         struct ib_uverbs_device *dev =
 123                 container_of(ref, struct ib_uverbs_device, ref);
 124
 125         complete(&dev->comp);
 126 }
 127
 128 static void ib_uverbs_release_event_file(struct kref *ref)
 129 {
 130         struct ib_uverbs_event_file *file =
 131                 container_of(ref, struct ib_uverbs_event_file, ref);
 132
 133         kfree(file);
 134 }
 135
 136 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 137                           struct ib_uverbs_event_file *ev_file,
 138                           struct ib_ucq_object *uobj)
 139 {
 140         struct ib_uverbs_event *evt, *tmp;
 141
 142         if (ev_file) {
 143                 spin_lock_irq(&ev_file->lock);
 144                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 145                         list_del(&evt->list);
 146                         kfree(evt);
 147                 }
 148                 spin_unlock_irq(&ev_file->lock);
 149
 150                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 151         }
 152
 153         spin_lock_irq(&file->async_file->lock);
 154         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 155                 list_del(&evt->list);
 156                 kfree(evt);
 157         }
 158         spin_unlock_irq(&file->async_file->lock);
 159 }
 160
 161 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 162                               struct ib_uevent_object *uobj)
 163 {
 164         struct ib_uverbs_event *evt, *tmp;
 165
 166         spin_lock_irq(&file->async_file->lock);
 167         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 168                 list_del(&evt->list);
 169                 kfree(evt);
 170         }
 171         spin_unlock_irq(&file->async_file->lock);
 172 }
 173
 174 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 175                                     struct ib_uqp_object *uobj)
 176 {
 177         struct ib_uverbs_mcast_entry *mcast, *tmp;
 178
 179         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 180                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 181                 list_del(&mcast->list);
 182                 kfree(mcast);
 183         }
 184 }
 185
 186 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 187                                       struct ib_ucontext *context)
 188 {
 189         struct ib_uobject *uobj, *tmp;
 190
 191         if (!context)
 192                 return 0;
 193
 194         context->closing = 1;
 195
 196         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 197                 struct ib_ah *ah = uobj->object;
 198
 199                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 200                 ib_destroy_ah(ah);
 201                 kfree(uobj);
 202         }
 203
 204         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 205                 struct ib_qp *qp = uobj->object;
 206                 struct ib_uqp_object *uqp =
 207                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 208
 209                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 210                 ib_uverbs_detach_umcast(qp, uqp);
 211                 ib_destroy_qp(qp);
 212                 ib_uverbs_release_uevent(file, &uqp->uevent);
 213                 kfree(uqp);
 214         }
 215
 216         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 217                 struct ib_cq *cq = uobj->object;
 218                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 219                 struct ib_ucq_object *ucq =
 220                         container_of(uobj, struct ib_ucq_object, uobject);
 221
 222                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 223                 ib_destroy_cq(cq);
 224                 ib_uverbs_release_ucq(file, ev_file, ucq);
 225                 kfree(ucq);
 226         }
 227
 228         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 229                 struct ib_srq *srq = uobj->object;
 230                 struct ib_uevent_object *uevent =
 231                         container_of(uobj, struct ib_uevent_object, uobject);
 232
 233                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 234                 ib_destroy_srq(srq);
 235                 ib_uverbs_release_uevent(file, uevent);
 236                 kfree(uevent);
 237         }
 238
 239         /* XXX Free MWs */
 240
 241         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 242                 struct ib_mr *mr = uobj->object;
 243
 244                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 245                 ib_dereg_mr(mr);
 246                 kfree(uobj);
 247         }
 248
 249         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 250                 struct ib_pd *pd = uobj->object;
 251
 252                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 253                 ib_dealloc_pd(pd);
 254                 kfree(uobj);
 255         }
 256
 257         return context->device->dealloc_ucontext(context);
 258 }
 259
 260 static void ib_uverbs_release_file(struct kref *ref)
 261 {
 262         struct ib_uverbs_file *file =
 263                 container_of(ref, struct ib_uverbs_file, ref);
 264
 265         module_put(file->device->ib_dev->owner);
 266         kref_put(&file->device->ref, ib_uverbs_release_dev);
 267
 268         kfree(file);
 269 }
 270
 271 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 272                                     size_t count, loff_t *pos)
 273 {
 274         struct ib_uverbs_event_file *file = filp->private_data;
 275         struct ib_uverbs_event *event;
 276         int eventsz;
 277         int ret = 0;
 278
 279         spin_lock_irq(&file->lock);
 280
 281         while (list_empty(&file->event_list)) {
 282                 spin_unlock_irq(&file->lock);
 283
 284                 if (filp->f_flags & O_NONBLOCK)
 285                         return -EAGAIN;
 286
 287                 if (wait_event_interruptible(file->poll_wait,
 288                                              !list_empty(&file->event_list)))
 289                         return -ERESTARTSYS;
 290
 291                 spin_lock_irq(&file->lock);
 292         }
 293
 294         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 295
 296         if (file->is_async)
 297                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 298         else
 299                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 300
 301         if (eventsz > count) {
 302                 ret   = -EINVAL;
 303                 event = NULL;
 304         } else {
 305                 list_del(file->event_list.next);
 306                 if (event->counter) {
 307                         ++(*event->counter);
 308                         list_del(&event->obj_list);
 309                 }
 310         }
 311
 312         spin_unlock_irq(&file->lock);
 313
 314         if (event) {
 315                 if (copy_to_user(buf, event, eventsz))
 316                         ret = -EFAULT;
 317                 else
 318                         ret = eventsz;
 319         }
 320
 321         kfree(event);
 322
 323         return ret;
 324 }
 325
 326 static unsigned int ib_uverbs_event_poll(struct file *filp,
 327                                          struct poll_table_struct *wait)
 328 {
 329         unsigned int pollflags = 0;
 330         struct ib_uverbs_event_file *file = filp->private_data;
 331
 332         poll_wait(filp, &file->poll_wait, wait);
 333
 334         spin_lock_irq(&file->lock);
 335         if (!list_empty(&file->event_list))
 336                 pollflags = POLLIN | POLLRDNORM;
 337         spin_unlock_irq(&file->lock);
 338
 339         return pollflags;
 340 }
 341
 342 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 343 {
 344         struct ib_uverbs_event_file *file = filp->private_data;
 345
 346         return fasync_helper(fd, filp, on, &file->async_queue);
 347 }
 348
 349 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 350 {
 351         struct ib_uverbs_event_file *file = filp->private_data;
 352         struct ib_uverbs_event *entry, *tmp;
 353
 354         spin_lock_irq(&file->lock);
 355         file->is_closed = 1;
 356         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 357                 if (entry->counter)
 358                         list_del(&entry->obj_list);
 359                 kfree(entry);
 360         }
 361         spin_unlock_irq(&file->lock);
 362
 363         ib_uverbs_event_fasync(-1, filp, 0);
 364
 365         if (file->is_async) {
 366                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 367                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 368         }
 369         kref_put(&file->ref, ib_uverbs_release_event_file);
 370
 371         return 0;
 372 }
 373
 374 static const struct file_operations uverbs_event_fops = {
 375         .owner   = THIS_MODULE,
 376         .read    = ib_uverbs_event_read,
 377         .poll    = ib_uverbs_event_poll,
 378         .release = ib_uverbs_event_close,
 379         .fasync  = ib_uverbs_event_fasync
 380 };
 381
 382 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 383 {
 384         struct ib_uverbs_event_file    *file = cq_context;
 385         struct ib_ucq_object           *uobj;
 386         struct ib_uverbs_event         *entry;
 387         unsigned long                   flags;
 388
 389         if (!file)
 390                 return;
 391
 392         spin_lock_irqsave(&file->lock, flags);
 393         if (file->is_closed) {
 394                 spin_unlock_irqrestore(&file->lock, flags);
 395                 return;
 396         }
 397
 398         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 399         if (!entry) {
 400                 spin_unlock_irqrestore(&file->lock, flags);
 401                 return;
 402         }
 403
 404         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 405
 406         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 407         entry->counter             = &uobj->comp_events_reported;
 408
 409         list_add_tail(&entry->list, &file->event_list);
 410         list_add_tail(&entry->obj_list, &uobj->comp_list);
 411         spin_unlock_irqrestore(&file->lock, flags);
 412
 413         wake_up_interruptible(&file->poll_wait);
 414         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 415 }
 416
 417 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 418                                     __u64 element, __u64 event,
 419                                     struct list_head *obj_list,
 420                                     u32 *counter)
 421 {
 422         struct ib_uverbs_event *entry;
 423         unsigned long flags;
 424
 425         spin_lock_irqsave(&file->async_file->lock, flags);
 426         if (file->async_file->is_closed) {
 427                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 428                 return;
 429         }
 430
 431         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 432         if (!entry) {
 433                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 434                 return;
 435         }
 436
 437         entry->desc.async.element    = element;
 438         entry->desc.async.event_type = event;
 439         entry->counter               = counter;
 440
 441         list_add_tail(&entry->list, &file->async_file->event_list);
 442         if (obj_list)
 443                 list_add_tail(&entry->obj_list, obj_list);
 444         spin_unlock_irqrestore(&file->async_file->lock, flags);
 445
 446         wake_up_interruptible(&file->async_file->poll_wait);
 447         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 448 }
 449
 450 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 451 {
 452         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 453                                                   struct ib_ucq_object, uobject);
 454
 455         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 456                                 event->event, &uobj->async_list,
 457                                 &uobj->async_events_reported);
 458 }
 459
 460 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 461 {
 462         struct ib_uevent_object *uobj;
 463
 464         uobj = container_of(event->element.qp->uobject,
 465                             struct ib_uevent_object, uobject);
 466
 467         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 468                                 event->event, &uobj->event_list,
 469                                 &uobj->events_reported);
 470 }
 471
 472 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 473 {
 474         struct ib_uevent_object *uobj;
 475
 476         uobj = container_of(event->element.srq->uobject,
 477                             struct ib_uevent_object, uobject);
 478
 479         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 480                                 event->event, &uobj->event_list,
 481                                 &uobj->events_reported);
 482 }
 483
 484 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 485                              struct ib_event *event)
 486 {
 487         struct ib_uverbs_file *file =
 488                 container_of(handler, struct ib_uverbs_file, event_handler);
 489
 490         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 491                                 NULL, NULL);
 492 }
 493
 494 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 495                                         int is_async, int *fd)
 496 {
 497         struct ib_uverbs_event_file *ev_file;
 498         struct file *filp;
 499         int ret;
 500
 501         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 502         if (!ev_file)
 503                 return ERR_PTR(-ENOMEM);
 504
 505         kref_init(&ev_file->ref);
 506         spin_lock_init(&ev_file->lock);
 507         INIT_LIST_HEAD(&ev_file->event_list);
 508         init_waitqueue_head(&ev_file->poll_wait);
 509         ev_file->uverbs_file = uverbs_file;
 510         ev_file->async_queue = NULL;
 511         ev_file->is_async    = is_async;
 512         ev_file->is_closed   = 0;
 513
 514         *fd = get_unused_fd();
 515         if (*fd < 0) {
 516                 ret = *fd;
 517                 goto err;
 518         }
 519
 520         /*
 521          * fops_get() can't fail here, because we're coming from a
 522          * system call on a uverbs file, which will already have a
 523          * module reference.
 524          */
 525         filp = alloc_file(uverbs_event_mnt, dget(uverbs_event_mnt->mnt_root),
 526                           FMODE_READ, fops_get(&uverbs_event_fops));
 527         if (!filp) {
 528                 ret = -ENFILE;
 529                 goto err_fd;
 530         }
 531
 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 /*
 614  * ib_uverbs_open() does not need the BKL:
 615  *
 616  *  - dev_table[] accesses are protected by map_lock, the
 617  *    ib_uverbs_device structures are properly reference counted, and
 618  *    everything else is purely local to the file being created, so
 619  *    races against other open calls are not a problem;
 620  *  - there is no ioctl method to race against;
 621  *  - the device is added to dev_table[] as the last part of module
 622  *    initialization, the open method will either immediately run
 623  *    -ENXIO, or all required initialization will be done.
 624  */
 625 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 626 {
 627         struct ib_uverbs_device *dev;
 628         struct ib_uverbs_file *file;
 629         int ret;
 630
 631         spin_lock(&map_lock);
 632         dev = dev_table[iminor(inode) - IB_UVERBS_BASE_MINOR];
 633         if (dev)
 634                 kref_get(&dev->ref);
 635         spin_unlock(&map_lock);
 636
 637         if (!dev)
 638                 return -ENXIO;
 639
 640         if (!try_module_get(dev->ib_dev->owner)) {
 641                 ret = -ENODEV;
 642                 goto err;
 643         }
 644
 645         file = kmalloc(sizeof *file, GFP_KERNEL);
 646         if (!file) {
 647                 ret = -ENOMEM;
 648                 goto err_module;
 649         }
 650
 651         file->device     = dev;
 652         file->ucontext   = NULL;
 653         file->async_file = NULL;
 654         kref_init(&file->ref);
 655         mutex_init(&file->mutex);
 656
 657         filp->private_data = file;
 658
 659         return 0;
 660
 661 err_module:
 662         module_put(dev->ib_dev->owner);
 663
 664 err:
 665         kref_put(&dev->ref, ib_uverbs_release_dev);
 666         return ret;
 667 }
 668
 669 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 670 {
 671         struct ib_uverbs_file *file = filp->private_data;
 672
 673         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 674
 675         if (file->async_file)
 676                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 677
 678         kref_put(&file->ref, ib_uverbs_release_file);
 679
 680         return 0;
 681 }
 682
 683 static const struct file_operations uverbs_fops = {
 684         .owner   = THIS_MODULE,
 685         .write   = ib_uverbs_write,
 686         .open    = ib_uverbs_open,
 687         .release = ib_uverbs_close
 688 };
 689
 690 static const struct file_operations uverbs_mmap_fops = {
 691         .owner   = THIS_MODULE,
 692         .write   = ib_uverbs_write,
 693         .mmap    = ib_uverbs_mmap,
 694         .open    = ib_uverbs_open,
 695         .release = ib_uverbs_close
 696 };
 697
 698 static struct ib_client uverbs_client = {
 699         .name   = "uverbs",
 700         .add    = ib_uverbs_add_one,
 701         .remove = ib_uverbs_remove_one
 702 };
 703
 704 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 705                           char *buf)
 706 {
 707         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 708
 709         if (!dev)
 710                 return -ENODEV;
 711
 712         return sprintf(buf, "%s\n", dev->ib_dev->name);
 713 }
 714 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 715
 716 static ssize_t show_dev_abi_version(struct device *device,
 717                                     struct device_attribute *attr, char *buf)
 718 {
 719         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 720
 721         if (!dev)
 722                 return -ENODEV;
 723
 724         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 725 }
 726 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 727
 728 static ssize_t show_abi_version(struct class *class, char *buf)
 729 {
 730         return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
 731 }
 732 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 733
 734 static void ib_uverbs_add_one(struct ib_device *device)
 735 {
 736         struct ib_uverbs_device *uverbs_dev;
 737
 738         if (!device->alloc_ucontext)
 739                 return;
 740
 741         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 742         if (!uverbs_dev)
 743                 return;
 744
 745         kref_init(&uverbs_dev->ref);
 746         init_completion(&uverbs_dev->comp);
 747
 748         spin_lock(&map_lock);
 749         uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 750         if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
 751                 spin_unlock(&map_lock);
 752                 goto err;
 753         }
 754         set_bit(uverbs_dev->devnum, dev_map);
 755         spin_unlock(&map_lock);
 756
 757         uverbs_dev->ib_dev           = device;
 758         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 759
 760         uverbs_dev->cdev = cdev_alloc();
 761         if (!uverbs_dev->cdev)
 762                 goto err;
 763         uverbs_dev->cdev->owner = THIS_MODULE;
 764         uverbs_dev->cdev->ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 765         kobject_set_name(&uverbs_dev->cdev->kobj, "uverbs%d", uverbs_dev->devnum);
 766         if (cdev_add(uverbs_dev->cdev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
 767                 goto err_cdev;
 768
 769         uverbs_dev->dev = device_create_drvdata(uverbs_class,
 770                                                 device->dma_device,
 771                                                 uverbs_dev->cdev->dev,
 772                                                 uverbs_dev,
 773                                                 "uverbs%d",
 774                                                 uverbs_dev->devnum);
 775         if (IS_ERR(uverbs_dev->dev))
 776                 goto err_cdev;
 777
 778         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 779                 goto err_class;
 780         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 781                 goto err_class;
 782
 783         spin_lock(&map_lock);
 784         dev_table[uverbs_dev->devnum] = uverbs_dev;
 785         spin_unlock(&map_lock);
 786
 787         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 788
 789         return;
 790
 791 err_class:
 792         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 793
 794 err_cdev:
 795         cdev_del(uverbs_dev->cdev);
 796         clear_bit(uverbs_dev->devnum, dev_map);
 797
 798 err:
 799         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 800         wait_for_completion(&uverbs_dev->comp);
 801         kfree(uverbs_dev);
 802         return;
 803 }
 804
 805 static void ib_uverbs_remove_one(struct ib_device *device)
 806 {
 807         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 808
 809         if (!uverbs_dev)
 810                 return;
 811
 812         dev_set_drvdata(uverbs_dev->dev, NULL);
 813         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 814         cdev_del(uverbs_dev->cdev);
 815
 816         spin_lock(&map_lock);
 817         dev_table[uverbs_dev->devnum] = NULL;
 818         spin_unlock(&map_lock);
 819
 820         clear_bit(uverbs_dev->devnum, dev_map);
 821
 822         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 823         wait_for_completion(&uverbs_dev->comp);
 824         kfree(uverbs_dev);
 825 }
 826
 827 static int uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
 828                                const char *dev_name, void *data,
 829                                struct vfsmount *mnt)
 830 {
 831         return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
 832                              INFINIBANDEVENTFS_MAGIC, mnt);
 833 }
 834
 835 static struct file_system_type uverbs_event_fs = {
 836         /* No owner field so module can be unloaded */
 837         .name    = "infinibandeventfs",
 838         .get_sb  = uverbs_event_get_sb,
 839         .kill_sb = kill_litter_super
 840 };
 841
 842 static int __init ib_uverbs_init(void)
 843 {
 844         int ret;
 845
 846         spin_lock_init(&map_lock);
 847
 848         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 849                                      "infiniband_verbs");
 850         if (ret) {
 851                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 852                 goto out;
 853         }
 854
 855         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 856         if (IS_ERR(uverbs_class)) {
 857                 ret = PTR_ERR(uverbs_class);
 858                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 859                 goto out_chrdev;
 860         }
 861
 862         ret = class_create_file(uverbs_class, &class_attr_abi_version);
 863         if (ret) {
 864                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 865                 goto out_class;
 866         }
 867
 868         ret = register_filesystem(&uverbs_event_fs);
 869         if (ret) {
 870                 printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
 871                 goto out_class;
 872         }
 873
 874         uverbs_event_mnt = kern_mount(&uverbs_event_fs);
 875         if (IS_ERR(uverbs_event_mnt)) {
 876                 ret = PTR_ERR(uverbs_event_mnt);
 877                 printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
 878                 goto out_fs;
 879         }
 880
 881         ret = ib_register_client(&uverbs_client);
 882         if (ret) {
 883                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 884                 goto out_mnt;
 885         }
 886
 887         return 0;
 888
 889 out_mnt:
 890         mntput(uverbs_event_mnt);
 891
 892 out_fs:
 893         unregister_filesystem(&uverbs_event_fs);
 894
 895 out_class:
 896         class_destroy(uverbs_class);
 897
 898 out_chrdev:
 899         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 900
 901 out:
 902         return ret;
 903 }
 904
 905 static void __exit ib_uverbs_cleanup(void)
 906 {
 907         ib_unregister_client(&uverbs_client);
 908         mntput(uverbs_event_mnt);
 909         unregister_filesystem(&uverbs_event_fs);
 910         class_destroy(uverbs_class);
 911         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 912         idr_destroy(&ib_uverbs_pd_idr);
 913         idr_destroy(&ib_uverbs_mr_idr);
 914         idr_destroy(&ib_uverbs_mw_idr);
 915         idr_destroy(&ib_uverbs_ah_idr);
 916         idr_destroy(&ib_uverbs_cq_idr);
 917         idr_destroy(&ib_uverbs_qp_idr);
 918         idr_destroy(&ib_uverbs_srq_idr);
 919 }
 920
 921 module_init(ib_uverbs_init);
 922 module_exit(ib_uverbs_cleanup);