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 DEFINE_IDR(ib_uverbs_xrcd_idr);
  76
  77 static DEFINE_SPINLOCK(map_lock);
  78 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  79
  80 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  81                                      const char __user *buf, int in_len,
  82                                      int out_len) = {
  83         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  84         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  85         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  86         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  87         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  88         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  89         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  90         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  91         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  92         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  93         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  94         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  95         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  96         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  97         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
  98         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
  99         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 100         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 101         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 102         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 103         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 104         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 105         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 106         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 107         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 108         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 109         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 110         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 111         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 112         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 113         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 114         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp
 115 };
 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 kobject *kobj)
 121 {
 122         struct ib_uverbs_device *dev =
 123                 container_of(kobj, struct ib_uverbs_device, kobj);
 124
 125         kfree(dev);
 126 }
 127
 128 static struct kobj_type ib_uverbs_dev_ktype = {
 129         .release = ib_uverbs_release_dev,
 130 };
 131
 132 static void ib_uverbs_release_event_file(struct kref *ref)
 133 {
 134         struct ib_uverbs_event_file *file =
 135                 container_of(ref, struct ib_uverbs_event_file, ref);
 136
 137         kfree(file);
 138 }
 139
 140 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 141                           struct ib_uverbs_event_file *ev_file,
 142                           struct ib_ucq_object *uobj)
 143 {
 144         struct ib_uverbs_event *evt, *tmp;
 145
 146         if (ev_file) {
 147                 spin_lock_irq(&ev_file->lock);
 148                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 149                         list_del(&evt->list);
 150                         kfree(evt);
 151                 }
 152                 spin_unlock_irq(&ev_file->lock);
 153
 154                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 155         }
 156
 157         spin_lock_irq(&file->async_file->lock);
 158         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 159                 list_del(&evt->list);
 160                 kfree(evt);
 161         }
 162         spin_unlock_irq(&file->async_file->lock);
 163 }
 164
 165 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 166                               struct ib_uevent_object *uobj)
 167 {
 168         struct ib_uverbs_event *evt, *tmp;
 169
 170         spin_lock_irq(&file->async_file->lock);
 171         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 172                 list_del(&evt->list);
 173                 kfree(evt);
 174         }
 175         spin_unlock_irq(&file->async_file->lock);
 176 }
 177
 178 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 179                                     struct ib_uqp_object *uobj)
 180 {
 181         struct ib_uverbs_mcast_entry *mcast, *tmp;
 182
 183         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 184                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 185                 list_del(&mcast->list);
 186                 kfree(mcast);
 187         }
 188 }
 189
 190 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 191                                       struct ib_ucontext *context)
 192 {
 193         struct ib_uobject *uobj, *tmp;
 194
 195         if (!context)
 196                 return 0;
 197
 198         context->closing = 1;
 199
 200         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 201                 struct ib_ah *ah = uobj->object;
 202
 203                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 204                 ib_destroy_ah(ah);
 205                 kfree(uobj);
 206         }
 207
 208         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 209                 struct ib_qp *qp = uobj->object;
 210                 struct ib_uqp_object *uqp =
 211                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 212
 213                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 214                 if (qp != qp->real_qp) {
 215                         ib_close_qp(qp);
 216                 } else {
 217                         ib_uverbs_detach_umcast(qp, uqp);
 218                         ib_destroy_qp(qp);
 219                 }
 220                 ib_uverbs_release_uevent(file, &uqp->uevent);
 221                 kfree(uqp);
 222         }
 223
 224         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 225                 struct ib_cq *cq = uobj->object;
 226                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 227                 struct ib_ucq_object *ucq =
 228                         container_of(uobj, struct ib_ucq_object, uobject);
 229
 230                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 231                 ib_destroy_cq(cq);
 232                 ib_uverbs_release_ucq(file, ev_file, ucq);
 233                 kfree(ucq);
 234         }
 235
 236         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 237                 struct ib_srq *srq = uobj->object;
 238                 struct ib_uevent_object *uevent =
 239                         container_of(uobj, struct ib_uevent_object, uobject);
 240
 241                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 242                 ib_destroy_srq(srq);
 243                 ib_uverbs_release_uevent(file, uevent);
 244                 kfree(uevent);
 245         }
 246
 247         /* XXX Free MWs */
 248
 249         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 250                 struct ib_mr *mr = uobj->object;
 251
 252                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 253                 ib_dereg_mr(mr);
 254                 kfree(uobj);
 255         }
 256
 257         mutex_lock(&file->device->xrcd_tree_mutex);
 258         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 259                 struct ib_xrcd *xrcd = uobj->object;
 260                 struct ib_uxrcd_object *uxrcd =
 261                         container_of(uobj, struct ib_uxrcd_object, uobject);
 262
 263                 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 264                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 265                 kfree(uxrcd);
 266         }
 267         mutex_unlock(&file->device->xrcd_tree_mutex);
 268
 269         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 270                 struct ib_pd *pd = uobj->object;
 271
 272                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 273                 ib_dealloc_pd(pd);
 274                 kfree(uobj);
 275         }
 276
 277         return context->device->dealloc_ucontext(context);
 278 }
 279
 280 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
 281 {
 282         complete(&dev->comp);
 283 }
 284
 285 static void ib_uverbs_release_file(struct kref *ref)
 286 {
 287         struct ib_uverbs_file *file =
 288                 container_of(ref, struct ib_uverbs_file, ref);
 289
 290         module_put(file->device->ib_dev->owner);
 291         if (atomic_dec_and_test(&file->device->refcount))
 292                 ib_uverbs_comp_dev(file->device);
 293
 294         kfree(file);
 295 }
 296
 297 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 298                                     size_t count, loff_t *pos)
 299 {
 300         struct ib_uverbs_event_file *file = filp->private_data;
 301         struct ib_uverbs_event *event;
 302         int eventsz;
 303         int ret = 0;
 304
 305         spin_lock_irq(&file->lock);
 306
 307         while (list_empty(&file->event_list)) {
 308                 spin_unlock_irq(&file->lock);
 309
 310                 if (filp->f_flags & O_NONBLOCK)
 311                         return -EAGAIN;
 312
 313                 if (wait_event_interruptible(file->poll_wait,
 314                                              !list_empty(&file->event_list)))
 315                         return -ERESTARTSYS;
 316
 317                 spin_lock_irq(&file->lock);
 318         }
 319
 320         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 321
 322         if (file->is_async)
 323                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 324         else
 325                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 326
 327         if (eventsz > count) {
 328                 ret   = -EINVAL;
 329                 event = NULL;
 330         } else {
 331                 list_del(file->event_list.next);
 332                 if (event->counter) {
 333                         ++(*event->counter);
 334                         list_del(&event->obj_list);
 335                 }
 336         }
 337
 338         spin_unlock_irq(&file->lock);
 339
 340         if (event) {
 341                 if (copy_to_user(buf, event, eventsz))
 342                         ret = -EFAULT;
 343                 else
 344                         ret = eventsz;
 345         }
 346
 347         kfree(event);
 348
 349         return ret;
 350 }
 351
 352 static unsigned int ib_uverbs_event_poll(struct file *filp,
 353                                          struct poll_table_struct *wait)
 354 {
 355         unsigned int pollflags = 0;
 356         struct ib_uverbs_event_file *file = filp->private_data;
 357
 358         poll_wait(filp, &file->poll_wait, wait);
 359
 360         spin_lock_irq(&file->lock);
 361         if (!list_empty(&file->event_list))
 362                 pollflags = POLLIN | POLLRDNORM;
 363         spin_unlock_irq(&file->lock);
 364
 365         return pollflags;
 366 }
 367
 368 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 369 {
 370         struct ib_uverbs_event_file *file = filp->private_data;
 371
 372         return fasync_helper(fd, filp, on, &file->async_queue);
 373 }
 374
 375 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 376 {
 377         struct ib_uverbs_event_file *file = filp->private_data;
 378         struct ib_uverbs_event *entry, *tmp;
 379
 380         spin_lock_irq(&file->lock);
 381         file->is_closed = 1;
 382         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 383                 if (entry->counter)
 384                         list_del(&entry->obj_list);
 385                 kfree(entry);
 386         }
 387         spin_unlock_irq(&file->lock);
 388
 389         if (file->is_async) {
 390                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 391                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 392         }
 393         kref_put(&file->ref, ib_uverbs_release_event_file);
 394
 395         return 0;
 396 }
 397
 398 static const struct file_operations uverbs_event_fops = {
 399         .owner   = THIS_MODULE,
 400         .read    = ib_uverbs_event_read,
 401         .poll    = ib_uverbs_event_poll,
 402         .release = ib_uverbs_event_close,
 403         .fasync  = ib_uverbs_event_fasync,
 404         .llseek  = no_llseek,
 405 };
 406
 407 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 408 {
 409         struct ib_uverbs_event_file    *file = cq_context;
 410         struct ib_ucq_object           *uobj;
 411         struct ib_uverbs_event         *entry;
 412         unsigned long                   flags;
 413
 414         if (!file)
 415                 return;
 416
 417         spin_lock_irqsave(&file->lock, flags);
 418         if (file->is_closed) {
 419                 spin_unlock_irqrestore(&file->lock, flags);
 420                 return;
 421         }
 422
 423         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 424         if (!entry) {
 425                 spin_unlock_irqrestore(&file->lock, flags);
 426                 return;
 427         }
 428
 429         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 430
 431         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 432         entry->counter             = &uobj->comp_events_reported;
 433
 434         list_add_tail(&entry->list, &file->event_list);
 435         list_add_tail(&entry->obj_list, &uobj->comp_list);
 436         spin_unlock_irqrestore(&file->lock, flags);
 437
 438         wake_up_interruptible(&file->poll_wait);
 439         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 440 }
 441
 442 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 443                                     __u64 element, __u64 event,
 444                                     struct list_head *obj_list,
 445                                     u32 *counter)
 446 {
 447         struct ib_uverbs_event *entry;
 448         unsigned long flags;
 449
 450         spin_lock_irqsave(&file->async_file->lock, flags);
 451         if (file->async_file->is_closed) {
 452                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 453                 return;
 454         }
 455
 456         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 457         if (!entry) {
 458                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 459                 return;
 460         }
 461
 462         entry->desc.async.element    = element;
 463         entry->desc.async.event_type = event;
 464         entry->desc.async.reserved   = 0;
 465         entry->counter               = counter;
 466
 467         list_add_tail(&entry->list, &file->async_file->event_list);
 468         if (obj_list)
 469                 list_add_tail(&entry->obj_list, obj_list);
 470         spin_unlock_irqrestore(&file->async_file->lock, flags);
 471
 472         wake_up_interruptible(&file->async_file->poll_wait);
 473         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 474 }
 475
 476 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 477 {
 478         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 479                                                   struct ib_ucq_object, uobject);
 480
 481         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 482                                 event->event, &uobj->async_list,
 483                                 &uobj->async_events_reported);
 484 }
 485
 486 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 487 {
 488         struct ib_uevent_object *uobj;
 489
 490         uobj = container_of(event->element.qp->uobject,
 491                             struct ib_uevent_object, uobject);
 492
 493         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 494                                 event->event, &uobj->event_list,
 495                                 &uobj->events_reported);
 496 }
 497
 498 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 499 {
 500         struct ib_uevent_object *uobj;
 501
 502         uobj = container_of(event->element.srq->uobject,
 503                             struct ib_uevent_object, uobject);
 504
 505         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 506                                 event->event, &uobj->event_list,
 507                                 &uobj->events_reported);
 508 }
 509
 510 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 511                              struct ib_event *event)
 512 {
 513         struct ib_uverbs_file *file =
 514                 container_of(handler, struct ib_uverbs_file, event_handler);
 515
 516         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 517                                 NULL, NULL);
 518 }
 519
 520 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 521                                         int is_async)
 522 {
 523         struct ib_uverbs_event_file *ev_file;
 524         struct file *filp;
 525
 526         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 527         if (!ev_file)
 528                 return ERR_PTR(-ENOMEM);
 529
 530         kref_init(&ev_file->ref);
 531         spin_lock_init(&ev_file->lock);
 532         INIT_LIST_HEAD(&ev_file->event_list);
 533         init_waitqueue_head(&ev_file->poll_wait);
 534         ev_file->uverbs_file = uverbs_file;
 535         ev_file->async_queue = NULL;
 536         ev_file->is_async    = is_async;
 537         ev_file->is_closed   = 0;
 538
 539         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 540                                   ev_file, O_RDONLY);
 541         if (IS_ERR(filp))
 542                 kfree(ev_file);
 543
 544         return filp;
 545 }
 546
 547 /*
 548  * Look up a completion event file by FD.  If lookup is successful,
 549  * takes a ref to the event file struct that it returns; if
 550  * unsuccessful, returns NULL.
 551  */
 552 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 553 {
 554         struct ib_uverbs_event_file *ev_file = NULL;
 555         struct file *filp;
 556
 557         filp = fget(fd);
 558         if (!filp)
 559                 return NULL;
 560
 561         if (filp->f_op != &uverbs_event_fops)
 562                 goto out;
 563
 564         ev_file = filp->private_data;
 565         if (ev_file->is_async) {
 566                 ev_file = NULL;
 567                 goto out;
 568         }
 569
 570         kref_get(&ev_file->ref);
 571
 572 out:
 573         fput(filp);
 574         return ev_file;
 575 }
 576
 577 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 578                              size_t count, loff_t *pos)
 579 {
 580         struct ib_uverbs_file *file = filp->private_data;
 581         struct ib_uverbs_cmd_hdr hdr;
 582
 583         if (count < sizeof hdr)
 584                 return -EINVAL;
 585
 586         if (copy_from_user(&hdr, buf, sizeof hdr))
 587                 return -EFAULT;
 588
 589         if (hdr.in_words * 4 != count)
 590                 return -EINVAL;
 591
 592         if (hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 593             !uverbs_cmd_table[hdr.command])
 594                 return -EINVAL;
 595
 596         if (!file->ucontext &&
 597             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 598                 return -EINVAL;
 599
 600         if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 601                 return -ENOSYS;
 602
 603         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 604                                              hdr.in_words * 4, hdr.out_words * 4);
 605 }
 606
 607 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 608 {
 609         struct ib_uverbs_file *file = filp->private_data;
 610
 611         if (!file->ucontext)
 612                 return -ENODEV;
 613         else
 614                 return file->device->ib_dev->mmap(file->ucontext, vma);
 615 }
 616
 617 /*
 618  * ib_uverbs_open() does not need the BKL:
 619  *
 620  *  - the ib_uverbs_device structures are properly reference counted and
 621  *    everything else is purely local to the file being created, so
 622  *    races against other open calls are not a problem;
 623  *  - there is no ioctl method to race against;
 624  *  - the open method will either immediately run -ENXIO, or all
 625  *    required initialization will be done.
 626  */
 627 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 628 {
 629         struct ib_uverbs_device *dev;
 630         struct ib_uverbs_file *file;
 631         int ret;
 632
 633         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 634         if (!atomic_inc_not_zero(&dev->refcount))
 635                 return -ENXIO;
 636
 637         if (!try_module_get(dev->ib_dev->owner)) {
 638                 ret = -ENODEV;
 639                 goto err;
 640         }
 641
 642         file = kmalloc(sizeof *file, GFP_KERNEL);
 643         if (!file) {
 644                 ret = -ENOMEM;
 645                 goto err_module;
 646         }
 647
 648         file->device     = dev;
 649         file->ucontext   = NULL;
 650         file->async_file = NULL;
 651         kref_init(&file->ref);
 652         mutex_init(&file->mutex);
 653
 654         filp->private_data = file;
 655         kobject_get(&dev->kobj);
 656
 657         return nonseekable_open(inode, filp);
 658
 659 err_module:
 660         module_put(dev->ib_dev->owner);
 661
 662 err:
 663         if (atomic_dec_and_test(&dev->refcount))
 664                 ib_uverbs_comp_dev(dev);
 665
 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         struct ib_uverbs_device *dev = file->device;
 673
 674         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 675
 676         if (file->async_file)
 677                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 678
 679         kref_put(&file->ref, ib_uverbs_release_file);
 680         kobject_put(&dev->kobj);
 681
 682         return 0;
 683 }
 684
 685 static const struct file_operations uverbs_fops = {
 686         .owner   = THIS_MODULE,
 687         .write   = ib_uverbs_write,
 688         .open    = ib_uverbs_open,
 689         .release = ib_uverbs_close,
 690         .llseek  = no_llseek,
 691 };
 692
 693 static const struct file_operations uverbs_mmap_fops = {
 694         .owner   = THIS_MODULE,
 695         .write   = ib_uverbs_write,
 696         .mmap    = ib_uverbs_mmap,
 697         .open    = ib_uverbs_open,
 698         .release = ib_uverbs_close,
 699         .llseek  = no_llseek,
 700 };
 701
 702 static struct ib_client uverbs_client = {
 703         .name   = "uverbs",
 704         .add    = ib_uverbs_add_one,
 705         .remove = ib_uverbs_remove_one
 706 };
 707
 708 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 709                           char *buf)
 710 {
 711         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 712
 713         if (!dev)
 714                 return -ENODEV;
 715
 716         return sprintf(buf, "%s\n", dev->ib_dev->name);
 717 }
 718 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 719
 720 static ssize_t show_dev_abi_version(struct device *device,
 721                                     struct device_attribute *attr, char *buf)
 722 {
 723         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 724
 725         if (!dev)
 726                 return -ENODEV;
 727
 728         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 729 }
 730 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 731
 732 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
 733                          __stringify(IB_USER_VERBS_ABI_VERSION));
 734
 735 static dev_t overflow_maj;
 736 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
 737
 738 /*
 739  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
 740  * requesting a new major number and doubling the number of max devices we
 741  * support. It's stupid, but simple.
 742  */
 743 static int find_overflow_devnum(void)
 744 {
 745         int ret;
 746
 747         if (!overflow_maj) {
 748                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
 749                                           "infiniband_verbs");
 750                 if (ret) {
 751                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
 752                         return ret;
 753                 }
 754         }
 755
 756         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
 757         if (ret >= IB_UVERBS_MAX_DEVICES)
 758                 return -1;
 759
 760         return ret;
 761 }
 762
 763 static void ib_uverbs_add_one(struct ib_device *device)
 764 {
 765         int devnum;
 766         dev_t base;
 767         struct ib_uverbs_device *uverbs_dev;
 768
 769         if (!device->alloc_ucontext)
 770                 return;
 771
 772         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 773         if (!uverbs_dev)
 774                 return;
 775
 776         atomic_set(&uverbs_dev->refcount, 1);
 777         init_completion(&uverbs_dev->comp);
 778         uverbs_dev->xrcd_tree = RB_ROOT;
 779         mutex_init(&uverbs_dev->xrcd_tree_mutex);
 780         kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
 781
 782         spin_lock(&map_lock);
 783         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 784         if (devnum >= IB_UVERBS_MAX_DEVICES) {
 785                 spin_unlock(&map_lock);
 786                 devnum = find_overflow_devnum();
 787                 if (devnum < 0)
 788                         goto err;
 789
 790                 spin_lock(&map_lock);
 791                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
 792                 base = devnum + overflow_maj;
 793                 set_bit(devnum, overflow_map);
 794         } else {
 795                 uverbs_dev->devnum = devnum;
 796                 base = devnum + IB_UVERBS_BASE_DEV;
 797                 set_bit(devnum, dev_map);
 798         }
 799         spin_unlock(&map_lock);
 800
 801         uverbs_dev->ib_dev           = device;
 802         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 803
 804         cdev_init(&uverbs_dev->cdev, NULL);
 805         uverbs_dev->cdev.owner = THIS_MODULE;
 806         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 807         uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
 808         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
 809         if (cdev_add(&uverbs_dev->cdev, base, 1))
 810                 goto err_cdev;
 811
 812         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 813                                         uverbs_dev->cdev.dev, uverbs_dev,
 814                                         "uverbs%d", uverbs_dev->devnum);
 815         if (IS_ERR(uverbs_dev->dev))
 816                 goto err_cdev;
 817
 818         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 819                 goto err_class;
 820         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 821                 goto err_class;
 822
 823         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 824
 825         return;
 826
 827 err_class:
 828         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 829
 830 err_cdev:
 831         cdev_del(&uverbs_dev->cdev);
 832         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 833                 clear_bit(devnum, dev_map);
 834         else
 835                 clear_bit(devnum, overflow_map);
 836
 837 err:
 838         if (atomic_dec_and_test(&uverbs_dev->refcount))
 839                 ib_uverbs_comp_dev(uverbs_dev);
 840         wait_for_completion(&uverbs_dev->comp);
 841         kobject_put(&uverbs_dev->kobj);
 842         return;
 843 }
 844
 845 static void ib_uverbs_remove_one(struct ib_device *device)
 846 {
 847         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 848
 849         if (!uverbs_dev)
 850                 return;
 851
 852         dev_set_drvdata(uverbs_dev->dev, NULL);
 853         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 854         cdev_del(&uverbs_dev->cdev);
 855
 856         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 857                 clear_bit(uverbs_dev->devnum, dev_map);
 858         else
 859                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
 860
 861         if (atomic_dec_and_test(&uverbs_dev->refcount))
 862                 ib_uverbs_comp_dev(uverbs_dev);
 863         wait_for_completion(&uverbs_dev->comp);
 864         kobject_put(&uverbs_dev->kobj);
 865 }
 866
 867 static char *uverbs_devnode(struct device *dev, mode_t *mode)
 868 {
 869         if (mode)
 870                 *mode = 0666;
 871         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
 872 }
 873
 874 static int __init ib_uverbs_init(void)
 875 {
 876         int ret;
 877
 878         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 879                                      "infiniband_verbs");
 880         if (ret) {
 881                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 882                 goto out;
 883         }
 884
 885         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 886         if (IS_ERR(uverbs_class)) {
 887                 ret = PTR_ERR(uverbs_class);
 888                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 889                 goto out_chrdev;
 890         }
 891
 892         uverbs_class->devnode = uverbs_devnode;
 893
 894         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
 895         if (ret) {
 896                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 897                 goto out_class;
 898         }
 899
 900         ret = ib_register_client(&uverbs_client);
 901         if (ret) {
 902                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 903                 goto out_class;
 904         }
 905
 906         return 0;
 907
 908 out_class:
 909         class_destroy(uverbs_class);
 910
 911 out_chrdev:
 912         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 913
 914 out:
 915         return ret;
 916 }
 917
 918 static void __exit ib_uverbs_cleanup(void)
 919 {
 920         ib_unregister_client(&uverbs_client);
 921         class_destroy(uverbs_class);
 922         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 923         if (overflow_maj)
 924                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
 925         idr_destroy(&ib_uverbs_pd_idr);
 926         idr_destroy(&ib_uverbs_mr_idr);
 927         idr_destroy(&ib_uverbs_mw_idr);
 928         idr_destroy(&ib_uverbs_ah_idr);
 929         idr_destroy(&ib_uverbs_cq_idr);
 930         idr_destroy(&ib_uverbs_qp_idr);
 931         idr_destroy(&ib_uverbs_srq_idr);
 932 }
 933
 934 module_init(ib_uverbs_init);
 935 module_exit(ib_uverbs_cleanup);