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 DEFINE_IDR(ib_uverbs_rule_idr);
  77
  78 static DEFINE_SPINLOCK(map_lock);
  79 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  80
  81 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  82                                      const char __user *buf, int in_len,
  83                                      int out_len) = {
  84         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  85         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  86         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  87         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  88         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  89         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  90         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  91         [IB_USER_VERBS_CMD_ALLOC_MW]            = ib_uverbs_alloc_mw,
  92         [IB_USER_VERBS_CMD_DEALLOC_MW]          = ib_uverbs_dealloc_mw,
  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         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 115         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 116         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 117         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp,
 118         [IB_USER_VERBS_CMD_CREATE_FLOW]         = ib_uverbs_create_flow,
 119         [IB_USER_VERBS_CMD_DESTROY_FLOW]        = ib_uverbs_destroy_flow
 120 };
 121
 122 static void ib_uverbs_add_one(struct ib_device *device);
 123 static void ib_uverbs_remove_one(struct ib_device *device);
 124
 125 static void ib_uverbs_release_dev(struct kref *ref)
 126 {
 127         struct ib_uverbs_device *dev =
 128                 container_of(ref, struct ib_uverbs_device, ref);
 129
 130         complete(&dev->comp);
 131 }
 132
 133 static void ib_uverbs_release_event_file(struct kref *ref)
 134 {
 135         struct ib_uverbs_event_file *file =
 136                 container_of(ref, struct ib_uverbs_event_file, ref);
 137
 138         kfree(file);
 139 }
 140
 141 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 142                           struct ib_uverbs_event_file *ev_file,
 143                           struct ib_ucq_object *uobj)
 144 {
 145         struct ib_uverbs_event *evt, *tmp;
 146
 147         if (ev_file) {
 148                 spin_lock_irq(&ev_file->lock);
 149                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 150                         list_del(&evt->list);
 151                         kfree(evt);
 152                 }
 153                 spin_unlock_irq(&ev_file->lock);
 154
 155                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 156         }
 157
 158         spin_lock_irq(&file->async_file->lock);
 159         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 160                 list_del(&evt->list);
 161                 kfree(evt);
 162         }
 163         spin_unlock_irq(&file->async_file->lock);
 164 }
 165
 166 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 167                               struct ib_uevent_object *uobj)
 168 {
 169         struct ib_uverbs_event *evt, *tmp;
 170
 171         spin_lock_irq(&file->async_file->lock);
 172         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 173                 list_del(&evt->list);
 174                 kfree(evt);
 175         }
 176         spin_unlock_irq(&file->async_file->lock);
 177 }
 178
 179 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 180                                     struct ib_uqp_object *uobj)
 181 {
 182         struct ib_uverbs_mcast_entry *mcast, *tmp;
 183
 184         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 185                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 186                 list_del(&mcast->list);
 187                 kfree(mcast);
 188         }
 189 }
 190
 191 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 192                                       struct ib_ucontext *context)
 193 {
 194         struct ib_uobject *uobj, *tmp;
 195
 196         if (!context)
 197                 return 0;
 198
 199         context->closing = 1;
 200
 201         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 202                 struct ib_ah *ah = uobj->object;
 203
 204                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 205                 ib_destroy_ah(ah);
 206                 kfree(uobj);
 207         }
 208
 209         /* Remove MWs before QPs, in order to support type 2A MWs. */
 210         list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
 211                 struct ib_mw *mw = uobj->object;
 212
 213                 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
 214                 ib_dealloc_mw(mw);
 215                 kfree(uobj);
 216         }
 217
 218         list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
 219                 struct ib_flow *flow_id = uobj->object;
 220
 221                 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
 222                 ib_destroy_flow(flow_id);
 223                 kfree(uobj);
 224         }
 225
 226         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 227                 struct ib_qp *qp = uobj->object;
 228                 struct ib_uqp_object *uqp =
 229                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 230
 231                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 232                 if (qp != qp->real_qp) {
 233                         ib_close_qp(qp);
 234                 } else {
 235                         ib_uverbs_detach_umcast(qp, uqp);
 236                         ib_destroy_qp(qp);
 237                 }
 238                 ib_uverbs_release_uevent(file, &uqp->uevent);
 239                 kfree(uqp);
 240         }
 241
 242         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 243                 struct ib_cq *cq = uobj->object;
 244                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 245                 struct ib_ucq_object *ucq =
 246                         container_of(uobj, struct ib_ucq_object, uobject);
 247
 248                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 249                 ib_destroy_cq(cq);
 250                 ib_uverbs_release_ucq(file, ev_file, ucq);
 251                 kfree(ucq);
 252         }
 253
 254         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 255                 struct ib_srq *srq = uobj->object;
 256                 struct ib_uevent_object *uevent =
 257                         container_of(uobj, struct ib_uevent_object, uobject);
 258
 259                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 260                 ib_destroy_srq(srq);
 261                 ib_uverbs_release_uevent(file, uevent);
 262                 kfree(uevent);
 263         }
 264
 265         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 266                 struct ib_mr *mr = uobj->object;
 267
 268                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 269                 ib_dereg_mr(mr);
 270                 kfree(uobj);
 271         }
 272
 273         mutex_lock(&file->device->xrcd_tree_mutex);
 274         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 275                 struct ib_xrcd *xrcd = uobj->object;
 276                 struct ib_uxrcd_object *uxrcd =
 277                         container_of(uobj, struct ib_uxrcd_object, uobject);
 278
 279                 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 280                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 281                 kfree(uxrcd);
 282         }
 283         mutex_unlock(&file->device->xrcd_tree_mutex);
 284
 285         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 286                 struct ib_pd *pd = uobj->object;
 287
 288                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 289                 ib_dealloc_pd(pd);
 290                 kfree(uobj);
 291         }
 292
 293         return context->device->dealloc_ucontext(context);
 294 }
 295
 296 static void ib_uverbs_release_file(struct kref *ref)
 297 {
 298         struct ib_uverbs_file *file =
 299                 container_of(ref, struct ib_uverbs_file, ref);
 300
 301         module_put(file->device->ib_dev->owner);
 302         kref_put(&file->device->ref, ib_uverbs_release_dev);
 303
 304         kfree(file);
 305 }
 306
 307 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 308                                     size_t count, loff_t *pos)
 309 {
 310         struct ib_uverbs_event_file *file = filp->private_data;
 311         struct ib_uverbs_event *event;
 312         int eventsz;
 313         int ret = 0;
 314
 315         spin_lock_irq(&file->lock);
 316
 317         while (list_empty(&file->event_list)) {
 318                 spin_unlock_irq(&file->lock);
 319
 320                 if (filp->f_flags & O_NONBLOCK)
 321                         return -EAGAIN;
 322
 323                 if (wait_event_interruptible(file->poll_wait,
 324                                              !list_empty(&file->event_list)))
 325                         return -ERESTARTSYS;
 326
 327                 spin_lock_irq(&file->lock);
 328         }
 329
 330         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 331
 332         if (file->is_async)
 333                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 334         else
 335                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 336
 337         if (eventsz > count) {
 338                 ret   = -EINVAL;
 339                 event = NULL;
 340         } else {
 341                 list_del(file->event_list.next);
 342                 if (event->counter) {
 343                         ++(*event->counter);
 344                         list_del(&event->obj_list);
 345                 }
 346         }
 347
 348         spin_unlock_irq(&file->lock);
 349
 350         if (event) {
 351                 if (copy_to_user(buf, event, eventsz))
 352                         ret = -EFAULT;
 353                 else
 354                         ret = eventsz;
 355         }
 356
 357         kfree(event);
 358
 359         return ret;
 360 }
 361
 362 static unsigned int ib_uverbs_event_poll(struct file *filp,
 363                                          struct poll_table_struct *wait)
 364 {
 365         unsigned int pollflags = 0;
 366         struct ib_uverbs_event_file *file = filp->private_data;
 367
 368         poll_wait(filp, &file->poll_wait, wait);
 369
 370         spin_lock_irq(&file->lock);
 371         if (!list_empty(&file->event_list))
 372                 pollflags = POLLIN | POLLRDNORM;
 373         spin_unlock_irq(&file->lock);
 374
 375         return pollflags;
 376 }
 377
 378 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 379 {
 380         struct ib_uverbs_event_file *file = filp->private_data;
 381
 382         return fasync_helper(fd, filp, on, &file->async_queue);
 383 }
 384
 385 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 386 {
 387         struct ib_uverbs_event_file *file = filp->private_data;
 388         struct ib_uverbs_event *entry, *tmp;
 389
 390         spin_lock_irq(&file->lock);
 391         file->is_closed = 1;
 392         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 393                 if (entry->counter)
 394                         list_del(&entry->obj_list);
 395                 kfree(entry);
 396         }
 397         spin_unlock_irq(&file->lock);
 398
 399         if (file->is_async) {
 400                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 401                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 402         }
 403         kref_put(&file->ref, ib_uverbs_release_event_file);
 404
 405         return 0;
 406 }
 407
 408 static const struct file_operations uverbs_event_fops = {
 409         .owner   = THIS_MODULE,
 410         .read    = ib_uverbs_event_read,
 411         .poll    = ib_uverbs_event_poll,
 412         .release = ib_uverbs_event_close,
 413         .fasync  = ib_uverbs_event_fasync,
 414         .llseek  = no_llseek,
 415 };
 416
 417 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 418 {
 419         struct ib_uverbs_event_file    *file = cq_context;
 420         struct ib_ucq_object           *uobj;
 421         struct ib_uverbs_event         *entry;
 422         unsigned long                   flags;
 423
 424         if (!file)
 425                 return;
 426
 427         spin_lock_irqsave(&file->lock, flags);
 428         if (file->is_closed) {
 429                 spin_unlock_irqrestore(&file->lock, flags);
 430                 return;
 431         }
 432
 433         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 434         if (!entry) {
 435                 spin_unlock_irqrestore(&file->lock, flags);
 436                 return;
 437         }
 438
 439         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 440
 441         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 442         entry->counter             = &uobj->comp_events_reported;
 443
 444         list_add_tail(&entry->list, &file->event_list);
 445         list_add_tail(&entry->obj_list, &uobj->comp_list);
 446         spin_unlock_irqrestore(&file->lock, flags);
 447
 448         wake_up_interruptible(&file->poll_wait);
 449         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 450 }
 451
 452 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 453                                     __u64 element, __u64 event,
 454                                     struct list_head *obj_list,
 455                                     u32 *counter)
 456 {
 457         struct ib_uverbs_event *entry;
 458         unsigned long flags;
 459
 460         spin_lock_irqsave(&file->async_file->lock, flags);
 461         if (file->async_file->is_closed) {
 462                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 463                 return;
 464         }
 465
 466         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 467         if (!entry) {
 468                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 469                 return;
 470         }
 471
 472         entry->desc.async.element    = element;
 473         entry->desc.async.event_type = event;
 474         entry->counter               = counter;
 475
 476         list_add_tail(&entry->list, &file->async_file->event_list);
 477         if (obj_list)
 478                 list_add_tail(&entry->obj_list, obj_list);
 479         spin_unlock_irqrestore(&file->async_file->lock, flags);
 480
 481         wake_up_interruptible(&file->async_file->poll_wait);
 482         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 483 }
 484
 485 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 486 {
 487         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 488                                                   struct ib_ucq_object, uobject);
 489
 490         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 491                                 event->event, &uobj->async_list,
 492                                 &uobj->async_events_reported);
 493 }
 494
 495 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 496 {
 497         struct ib_uevent_object *uobj;
 498
 499         uobj = container_of(event->element.qp->uobject,
 500                             struct ib_uevent_object, uobject);
 501
 502         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 503                                 event->event, &uobj->event_list,
 504                                 &uobj->events_reported);
 505 }
 506
 507 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 508 {
 509         struct ib_uevent_object *uobj;
 510
 511         uobj = container_of(event->element.srq->uobject,
 512                             struct ib_uevent_object, uobject);
 513
 514         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 515                                 event->event, &uobj->event_list,
 516                                 &uobj->events_reported);
 517 }
 518
 519 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 520                              struct ib_event *event)
 521 {
 522         struct ib_uverbs_file *file =
 523                 container_of(handler, struct ib_uverbs_file, event_handler);
 524
 525         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 526                                 NULL, NULL);
 527 }
 528
 529 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 530                                         int is_async)
 531 {
 532         struct ib_uverbs_event_file *ev_file;
 533         struct file *filp;
 534
 535         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 536         if (!ev_file)
 537                 return ERR_PTR(-ENOMEM);
 538
 539         kref_init(&ev_file->ref);
 540         spin_lock_init(&ev_file->lock);
 541         INIT_LIST_HEAD(&ev_file->event_list);
 542         init_waitqueue_head(&ev_file->poll_wait);
 543         ev_file->uverbs_file = uverbs_file;
 544         ev_file->async_queue = NULL;
 545         ev_file->is_async    = is_async;
 546         ev_file->is_closed   = 0;
 547
 548         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 549                                   ev_file, O_RDONLY);
 550         if (IS_ERR(filp))
 551                 kfree(ev_file);
 552
 553         return filp;
 554 }
 555
 556 /*
 557  * Look up a completion event file by FD.  If lookup is successful,
 558  * takes a ref to the event file struct that it returns; if
 559  * unsuccessful, returns NULL.
 560  */
 561 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 562 {
 563         struct ib_uverbs_event_file *ev_file = NULL;
 564         struct fd f = fdget(fd);
 565
 566         if (!f.file)
 567                 return NULL;
 568
 569         if (f.file->f_op != &uverbs_event_fops)
 570                 goto out;
 571
 572         ev_file = f.file->private_data;
 573         if (ev_file->is_async) {
 574                 ev_file = NULL;
 575                 goto out;
 576         }
 577
 578         kref_get(&ev_file->ref);
 579
 580 out:
 581         fdput(f);
 582         return ev_file;
 583 }
 584
 585 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 586                              size_t count, loff_t *pos)
 587 {
 588         struct ib_uverbs_file *file = filp->private_data;
 589         struct ib_uverbs_cmd_hdr hdr;
 590
 591         if (count < sizeof hdr)
 592                 return -EINVAL;
 593
 594         if (copy_from_user(&hdr, buf, sizeof hdr))
 595                 return -EFAULT;
 596
 597         if (hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 598             !uverbs_cmd_table[hdr.command])
 599                 return -EINVAL;
 600
 601         if (!file->ucontext &&
 602             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 603                 return -EINVAL;
 604
 605         if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 606                 return -ENOSYS;
 607
 608         if (hdr.command >= IB_USER_VERBS_CMD_THRESHOLD) {
 609                 struct ib_uverbs_cmd_hdr_ex hdr_ex;
 610
 611                 if (copy_from_user(&hdr_ex, buf, sizeof(hdr_ex)))
 612                         return -EFAULT;
 613
 614                 if (((hdr_ex.in_words + hdr_ex.provider_in_words) * 4) != count)
 615                         return -EINVAL;
 616
 617                 return uverbs_cmd_table[hdr.command](file,
 618                                                      buf + sizeof(hdr_ex),
 619                                                      (hdr_ex.in_words +
 620                                                       hdr_ex.provider_in_words) * 4,
 621                                                      (hdr_ex.out_words +
 622                                                       hdr_ex.provider_out_words) * 4);
 623         } else {
 624                 if (hdr.in_words * 4 != count)
 625                         return -EINVAL;
 626
 627                 return uverbs_cmd_table[hdr.command](file,
 628                                                      buf + sizeof(hdr),
 629                                                      hdr.in_words * 4,
 630                                                      hdr.out_words * 4);
 631         }
 632 }
 633
 634 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 635 {
 636         struct ib_uverbs_file *file = filp->private_data;
 637
 638         if (!file->ucontext)
 639                 return -ENODEV;
 640         else
 641                 return file->device->ib_dev->mmap(file->ucontext, vma);
 642 }
 643
 644 /*
 645  * ib_uverbs_open() does not need the BKL:
 646  *
 647  *  - the ib_uverbs_device structures are properly reference counted and
 648  *    everything else is purely local to the file being created, so
 649  *    races against other open calls are not a problem;
 650  *  - there is no ioctl method to race against;
 651  *  - the open method will either immediately run -ENXIO, or all
 652  *    required initialization will be done.
 653  */
 654 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 655 {
 656         struct ib_uverbs_device *dev;
 657         struct ib_uverbs_file *file;
 658         int ret;
 659
 660         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 661         if (dev)
 662                 kref_get(&dev->ref);
 663         else
 664                 return -ENXIO;
 665
 666         if (!try_module_get(dev->ib_dev->owner)) {
 667                 ret = -ENODEV;
 668                 goto err;
 669         }
 670
 671         file = kmalloc(sizeof *file, GFP_KERNEL);
 672         if (!file) {
 673                 ret = -ENOMEM;
 674                 goto err_module;
 675         }
 676
 677         file->device     = dev;
 678         file->ucontext   = NULL;
 679         file->async_file = NULL;
 680         kref_init(&file->ref);
 681         mutex_init(&file->mutex);
 682
 683         filp->private_data = file;
 684
 685         return nonseekable_open(inode, filp);
 686
 687 err_module:
 688         module_put(dev->ib_dev->owner);
 689
 690 err:
 691         kref_put(&dev->ref, ib_uverbs_release_dev);
 692         return ret;
 693 }
 694
 695 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 696 {
 697         struct ib_uverbs_file *file = filp->private_data;
 698
 699         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 700
 701         if (file->async_file)
 702                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 703
 704         kref_put(&file->ref, ib_uverbs_release_file);
 705
 706         return 0;
 707 }
 708
 709 static const struct file_operations uverbs_fops = {
 710         .owner   = THIS_MODULE,
 711         .write   = ib_uverbs_write,
 712         .open    = ib_uverbs_open,
 713         .release = ib_uverbs_close,
 714         .llseek  = no_llseek,
 715 };
 716
 717 static const struct file_operations uverbs_mmap_fops = {
 718         .owner   = THIS_MODULE,
 719         .write   = ib_uverbs_write,
 720         .mmap    = ib_uverbs_mmap,
 721         .open    = ib_uverbs_open,
 722         .release = ib_uverbs_close,
 723         .llseek  = no_llseek,
 724 };
 725
 726 static struct ib_client uverbs_client = {
 727         .name   = "uverbs",
 728         .add    = ib_uverbs_add_one,
 729         .remove = ib_uverbs_remove_one
 730 };
 731
 732 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 733                           char *buf)
 734 {
 735         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 736
 737         if (!dev)
 738                 return -ENODEV;
 739
 740         return sprintf(buf, "%s\n", dev->ib_dev->name);
 741 }
 742 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 743
 744 static ssize_t show_dev_abi_version(struct device *device,
 745                                     struct device_attribute *attr, char *buf)
 746 {
 747         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 748
 749         if (!dev)
 750                 return -ENODEV;
 751
 752         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 753 }
 754 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 755
 756 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
 757                          __stringify(IB_USER_VERBS_ABI_VERSION));
 758
 759 static dev_t overflow_maj;
 760 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
 761
 762 /*
 763  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
 764  * requesting a new major number and doubling the number of max devices we
 765  * support. It's stupid, but simple.
 766  */
 767 static int find_overflow_devnum(void)
 768 {
 769         int ret;
 770
 771         if (!overflow_maj) {
 772                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
 773                                           "infiniband_verbs");
 774                 if (ret) {
 775                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
 776                         return ret;
 777                 }
 778         }
 779
 780         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
 781         if (ret >= IB_UVERBS_MAX_DEVICES)
 782                 return -1;
 783
 784         return ret;
 785 }
 786
 787 static void ib_uverbs_add_one(struct ib_device *device)
 788 {
 789         int devnum;
 790         dev_t base;
 791         struct ib_uverbs_device *uverbs_dev;
 792
 793         if (!device->alloc_ucontext)
 794                 return;
 795
 796         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 797         if (!uverbs_dev)
 798                 return;
 799
 800         kref_init(&uverbs_dev->ref);
 801         init_completion(&uverbs_dev->comp);
 802         uverbs_dev->xrcd_tree = RB_ROOT;
 803         mutex_init(&uverbs_dev->xrcd_tree_mutex);
 804
 805         spin_lock(&map_lock);
 806         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 807         if (devnum >= IB_UVERBS_MAX_DEVICES) {
 808                 spin_unlock(&map_lock);
 809                 devnum = find_overflow_devnum();
 810                 if (devnum < 0)
 811                         goto err;
 812
 813                 spin_lock(&map_lock);
 814                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
 815                 base = devnum + overflow_maj;
 816                 set_bit(devnum, overflow_map);
 817         } else {
 818                 uverbs_dev->devnum = devnum;
 819                 base = devnum + IB_UVERBS_BASE_DEV;
 820                 set_bit(devnum, dev_map);
 821         }
 822         spin_unlock(&map_lock);
 823
 824         uverbs_dev->ib_dev           = device;
 825         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 826
 827         cdev_init(&uverbs_dev->cdev, NULL);
 828         uverbs_dev->cdev.owner = THIS_MODULE;
 829         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 830         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
 831         if (cdev_add(&uverbs_dev->cdev, base, 1))
 832                 goto err_cdev;
 833
 834         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 835                                         uverbs_dev->cdev.dev, uverbs_dev,
 836                                         "uverbs%d", uverbs_dev->devnum);
 837         if (IS_ERR(uverbs_dev->dev))
 838                 goto err_cdev;
 839
 840         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 841                 goto err_class;
 842         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 843                 goto err_class;
 844
 845         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 846
 847         return;
 848
 849 err_class:
 850         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 851
 852 err_cdev:
 853         cdev_del(&uverbs_dev->cdev);
 854         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 855                 clear_bit(devnum, dev_map);
 856         else
 857                 clear_bit(devnum, overflow_map);
 858
 859 err:
 860         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 861         wait_for_completion(&uverbs_dev->comp);
 862         kfree(uverbs_dev);
 863         return;
 864 }
 865
 866 static void ib_uverbs_remove_one(struct ib_device *device)
 867 {
 868         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 869
 870         if (!uverbs_dev)
 871                 return;
 872
 873         dev_set_drvdata(uverbs_dev->dev, NULL);
 874         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 875         cdev_del(&uverbs_dev->cdev);
 876
 877         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 878                 clear_bit(uverbs_dev->devnum, dev_map);
 879         else
 880                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
 881
 882         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 883         wait_for_completion(&uverbs_dev->comp);
 884         kfree(uverbs_dev);
 885 }
 886
 887 static char *uverbs_devnode(struct device *dev, umode_t *mode)
 888 {
 889         if (mode)
 890                 *mode = 0666;
 891         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
 892 }
 893
 894 static int __init ib_uverbs_init(void)
 895 {
 896         int ret;
 897
 898         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 899                                      "infiniband_verbs");
 900         if (ret) {
 901                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 902                 goto out;
 903         }
 904
 905         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 906         if (IS_ERR(uverbs_class)) {
 907                 ret = PTR_ERR(uverbs_class);
 908                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 909                 goto out_chrdev;
 910         }
 911
 912         uverbs_class->devnode = uverbs_devnode;
 913
 914         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
 915         if (ret) {
 916                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 917                 goto out_class;
 918         }
 919
 920         ret = ib_register_client(&uverbs_client);
 921         if (ret) {
 922                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 923                 goto out_class;
 924         }
 925
 926         return 0;
 927
 928 out_class:
 929         class_destroy(uverbs_class);
 930
 931 out_chrdev:
 932         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 933
 934 out:
 935         return ret;
 936 }
 937
 938 static void __exit ib_uverbs_cleanup(void)
 939 {
 940         ib_unregister_client(&uverbs_client);
 941         class_destroy(uverbs_class);
 942         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 943         if (overflow_maj)
 944                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
 945         idr_destroy(&ib_uverbs_pd_idr);
 946         idr_destroy(&ib_uverbs_mr_idr);
 947         idr_destroy(&ib_uverbs_mw_idr);
 948         idr_destroy(&ib_uverbs_ah_idr);
 949         idr_destroy(&ib_uverbs_cq_idr);
 950         idr_destroy(&ib_uverbs_qp_idr);
 951         idr_destroy(&ib_uverbs_srq_idr);
 952 }
 953
 954 module_init(ib_uverbs_init);
 955 module_exit(ib_uverbs_cleanup);