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