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 <rdma/ib.h>
  52
  53 #include "uverbs.h"
  54
  55 MODULE_AUTHOR("Roland Dreier");
  56 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  57 MODULE_LICENSE("Dual BSD/GPL");
  58
  59 enum {
  60         IB_UVERBS_MAJOR       = 231,
  61         IB_UVERBS_BASE_MINOR  = 192,
  62         IB_UVERBS_MAX_DEVICES = 32
  63 };
  64
  65 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  66
  67 static struct class *uverbs_class;
  68
  69 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  70 DEFINE_IDR(ib_uverbs_pd_idr);
  71 DEFINE_IDR(ib_uverbs_mr_idr);
  72 DEFINE_IDR(ib_uverbs_mw_idr);
  73 DEFINE_IDR(ib_uverbs_ah_idr);
  74 DEFINE_IDR(ib_uverbs_cq_idr);
  75 DEFINE_IDR(ib_uverbs_qp_idr);
  76 DEFINE_IDR(ib_uverbs_srq_idr);
  77 DEFINE_IDR(ib_uverbs_xrcd_idr);
  78
  79 static DEFINE_SPINLOCK(map_lock);
  80 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  81
  82 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  83                                      const char __user *buf, int in_len,
  84                                      int out_len) = {
  85         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  86         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  87         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  88         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  89         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  90         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  91         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  92         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  93         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  94         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  95         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  96         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  97         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  98         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  99         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 100         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 101         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 102         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 103         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 104         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 105         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 106         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 107         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 108         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 109         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 110         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 111         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 112         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 113         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 114         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 115         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 116         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp
 117 };
 118
 119 static void ib_uverbs_add_one(struct ib_device *device);
 120 static void ib_uverbs_remove_one(struct ib_device *device);
 121
 122 static void ib_uverbs_release_dev(struct kobject *kobj)
 123 {
 124         struct ib_uverbs_device *dev =
 125                 container_of(kobj, struct ib_uverbs_device, kobj);
 126
 127         kfree(dev);
 128 }
 129
 130 static struct kobj_type ib_uverbs_dev_ktype = {
 131         .release = ib_uverbs_release_dev,
 132 };
 133
 134 static void ib_uverbs_release_event_file(struct kref *ref)
 135 {
 136         struct ib_uverbs_event_file *file =
 137                 container_of(ref, struct ib_uverbs_event_file, ref);
 138
 139         kfree(file);
 140 }
 141
 142 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 143                           struct ib_uverbs_event_file *ev_file,
 144                           struct ib_ucq_object *uobj)
 145 {
 146         struct ib_uverbs_event *evt, *tmp;
 147
 148         if (ev_file) {
 149                 spin_lock_irq(&ev_file->lock);
 150                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 151                         list_del(&evt->list);
 152                         kfree(evt);
 153                 }
 154                 spin_unlock_irq(&ev_file->lock);
 155
 156                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 157         }
 158
 159         spin_lock_irq(&file->async_file->lock);
 160         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 161                 list_del(&evt->list);
 162                 kfree(evt);
 163         }
 164         spin_unlock_irq(&file->async_file->lock);
 165 }
 166
 167 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 168                               struct ib_uevent_object *uobj)
 169 {
 170         struct ib_uverbs_event *evt, *tmp;
 171
 172         spin_lock_irq(&file->async_file->lock);
 173         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 174                 list_del(&evt->list);
 175                 kfree(evt);
 176         }
 177         spin_unlock_irq(&file->async_file->lock);
 178 }
 179
 180 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 181                                     struct ib_uqp_object *uobj)
 182 {
 183         struct ib_uverbs_mcast_entry *mcast, *tmp;
 184
 185         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 186                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 187                 list_del(&mcast->list);
 188                 kfree(mcast);
 189         }
 190 }
 191
 192 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 193                                       struct ib_ucontext *context)
 194 {
 195         struct ib_uobject *uobj, *tmp;
 196
 197         if (!context)
 198                 return 0;
 199
 200         context->closing = 1;
 201
 202         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 203                 struct ib_ah *ah = uobj->object;
 204
 205                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 206                 ib_destroy_ah(ah);
 207                 kfree(uobj);
 208         }
 209
 210         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 211                 struct ib_qp *qp = uobj->object;
 212                 struct ib_uqp_object *uqp =
 213                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 214
 215                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 216                 if (qp == qp->real_qp)
 217                         ib_uverbs_detach_umcast(qp, uqp);
 218                 ib_destroy_qp(qp);
 219                 ib_uverbs_release_uevent(file, &uqp->uevent);
 220                 kfree(uqp);
 221         }
 222
 223         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 224                 struct ib_cq *cq = uobj->object;
 225                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 226                 struct ib_ucq_object *ucq =
 227                         container_of(uobj, struct ib_ucq_object, uobject);
 228
 229                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 230                 ib_destroy_cq(cq);
 231                 ib_uverbs_release_ucq(file, ev_file, ucq);
 232                 kfree(ucq);
 233         }
 234
 235         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 236                 struct ib_srq *srq = uobj->object;
 237                 struct ib_uevent_object *uevent =
 238                         container_of(uobj, struct ib_uevent_object, uobject);
 239
 240                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 241                 ib_destroy_srq(srq);
 242                 ib_uverbs_release_uevent(file, uevent);
 243                 kfree(uevent);
 244         }
 245
 246         /* XXX Free MWs */
 247
 248         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 249                 struct ib_mr *mr = uobj->object;
 250
 251                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 252                 ib_dereg_mr(mr);
 253                 kfree(uobj);
 254         }
 255
 256         mutex_lock(&file->device->xrcd_tree_mutex);
 257         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 258                 struct ib_xrcd *xrcd = uobj->object;
 259                 struct ib_uxrcd_object *uxrcd =
 260                         container_of(uobj, struct ib_uxrcd_object, uobject);
 261
 262                 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 263                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 264                 kfree(uxrcd);
 265         }
 266         mutex_unlock(&file->device->xrcd_tree_mutex);
 267
 268         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 269                 struct ib_pd *pd = uobj->object;
 270
 271                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 272                 ib_dealloc_pd(pd);
 273                 kfree(uobj);
 274         }
 275
 276         return context->device->dealloc_ucontext(context);
 277 }
 278
 279 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
 280 {
 281         complete(&dev->comp);
 282 }
 283
 284 static void ib_uverbs_release_file(struct kref *ref)
 285 {
 286         struct ib_uverbs_file *file =
 287                 container_of(ref, struct ib_uverbs_file, ref);
 288
 289         module_put(file->device->ib_dev->owner);
 290         if (atomic_dec_and_test(&file->device->refcount))
 291                 ib_uverbs_comp_dev(file->device);
 292
 293         kobject_put(&file->device->kobj);
 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 (WARN_ON_ONCE(!ib_safe_file_access(filp)))
 584                 return -EACCES;
 585
 586         if (count < sizeof hdr)
 587                 return -EINVAL;
 588
 589         if (copy_from_user(&hdr, buf, sizeof hdr))
 590                 return -EFAULT;
 591
 592         if (hdr.in_words * 4 != count)
 593                 return -EINVAL;
 594
 595         if (hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 596             !uverbs_cmd_table[hdr.command])
 597                 return -EINVAL;
 598
 599         if (!file->ucontext &&
 600             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 601                 return -EINVAL;
 602
 603         if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 604                 return -ENOSYS;
 605
 606         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 607                                              hdr.in_words * 4, hdr.out_words * 4);
 608 }
 609
 610 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 611 {
 612         struct ib_uverbs_file *file = filp->private_data;
 613
 614         if (!file->ucontext)
 615                 return -ENODEV;
 616         else
 617                 return file->device->ib_dev->mmap(file->ucontext, vma);
 618 }
 619
 620 /*
 621  * ib_uverbs_open() does not need the BKL:
 622  *
 623  *  - the ib_uverbs_device structures are properly reference counted and
 624  *    everything else is purely local to the file being created, so
 625  *    races against other open calls are not a problem;
 626  *  - there is no ioctl method to race against;
 627  *  - the open method will either immediately run -ENXIO, or all
 628  *    required initialization will be done.
 629  */
 630 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 631 {
 632         struct ib_uverbs_device *dev;
 633         struct ib_uverbs_file *file;
 634         int ret;
 635
 636         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 637         if (!atomic_inc_not_zero(&dev->refcount))
 638                 return -ENXIO;
 639
 640         if (!try_module_get(dev->ib_dev->owner)) {
 641                 ret = -ENODEV;
 642                 goto err;
 643         }
 644
 645         file = kmalloc(sizeof *file, GFP_KERNEL);
 646         if (!file) {
 647                 ret = -ENOMEM;
 648                 goto err_module;
 649         }
 650
 651         file->device     = dev;
 652         file->ucontext   = NULL;
 653         file->async_file = NULL;
 654         kref_init(&file->ref);
 655         mutex_init(&file->mutex);
 656
 657         filp->private_data = file;
 658         kobject_get(&dev->kobj);
 659
 660         return nonseekable_open(inode, filp);
 661
 662 err_module:
 663         module_put(dev->ib_dev->owner);
 664
 665 err:
 666         if (atomic_dec_and_test(&dev->refcount))
 667                 ib_uverbs_comp_dev(dev);
 668
 669         return ret;
 670 }
 671
 672 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 673 {
 674         struct ib_uverbs_file *file = filp->private_data;
 675
 676         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 677
 678         if (file->async_file)
 679                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 680
 681         kref_put(&file->ref, ib_uverbs_release_file);
 682
 683         return 0;
 684 }
 685
 686 static const struct file_operations uverbs_fops = {
 687         .owner   = THIS_MODULE,
 688         .write   = ib_uverbs_write,
 689         .open    = ib_uverbs_open,
 690         .release = ib_uverbs_close,
 691         .llseek  = no_llseek,
 692 };
 693
 694 static const struct file_operations uverbs_mmap_fops = {
 695         .owner   = THIS_MODULE,
 696         .write   = ib_uverbs_write,
 697         .mmap    = ib_uverbs_mmap,
 698         .open    = ib_uverbs_open,
 699         .release = ib_uverbs_close,
 700         .llseek  = no_llseek,
 701 };
 702
 703 static struct ib_client uverbs_client = {
 704         .name   = "uverbs",
 705         .add    = ib_uverbs_add_one,
 706         .remove = ib_uverbs_remove_one
 707 };
 708
 709 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 710                           char *buf)
 711 {
 712         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 713
 714         if (!dev)
 715                 return -ENODEV;
 716
 717         return sprintf(buf, "%s\n", dev->ib_dev->name);
 718 }
 719 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 720
 721 static ssize_t show_dev_abi_version(struct device *device,
 722                                     struct device_attribute *attr, char *buf)
 723 {
 724         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 725
 726         if (!dev)
 727                 return -ENODEV;
 728
 729         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 730 }
 731 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 732
 733 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
 734                          __stringify(IB_USER_VERBS_ABI_VERSION));
 735
 736 static dev_t overflow_maj;
 737 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
 738
 739 /*
 740  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
 741  * requesting a new major number and doubling the number of max devices we
 742  * support. It's stupid, but simple.
 743  */
 744 static int find_overflow_devnum(void)
 745 {
 746         int ret;
 747
 748         if (!overflow_maj) {
 749                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
 750                                           "infiniband_verbs");
 751                 if (ret) {
 752                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
 753                         return ret;
 754                 }
 755         }
 756
 757         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
 758         if (ret >= IB_UVERBS_MAX_DEVICES)
 759                 return -1;
 760
 761         return ret;
 762 }
 763
 764 static void ib_uverbs_add_one(struct ib_device *device)
 765 {
 766         int devnum;
 767         dev_t base;
 768         struct ib_uverbs_device *uverbs_dev;
 769
 770         if (!device->alloc_ucontext)
 771                 return;
 772
 773         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 774         if (!uverbs_dev)
 775                 return;
 776
 777         atomic_set(&uverbs_dev->refcount, 1);
 778         init_completion(&uverbs_dev->comp);
 779         uverbs_dev->xrcd_tree = RB_ROOT;
 780         mutex_init(&uverbs_dev->xrcd_tree_mutex);
 781         kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
 782
 783         spin_lock(&map_lock);
 784         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 785         if (devnum >= IB_UVERBS_MAX_DEVICES) {
 786                 spin_unlock(&map_lock);
 787                 devnum = find_overflow_devnum();
 788                 if (devnum < 0)
 789                         goto err;
 790
 791                 spin_lock(&map_lock);
 792                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
 793                 base = devnum + overflow_maj;
 794                 set_bit(devnum, overflow_map);
 795         } else {
 796                 uverbs_dev->devnum = devnum;
 797                 base = devnum + IB_UVERBS_BASE_DEV;
 798                 set_bit(devnum, dev_map);
 799         }
 800         spin_unlock(&map_lock);
 801
 802         uverbs_dev->ib_dev           = device;
 803         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 804
 805         cdev_init(&uverbs_dev->cdev, NULL);
 806         uverbs_dev->cdev.owner = THIS_MODULE;
 807         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 808         uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
 809         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
 810         if (cdev_add(&uverbs_dev->cdev, base, 1))
 811                 goto err_cdev;
 812
 813         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 814                                         uverbs_dev->cdev.dev, uverbs_dev,
 815                                         "uverbs%d", uverbs_dev->devnum);
 816         if (IS_ERR(uverbs_dev->dev))
 817                 goto err_cdev;
 818
 819         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 820                 goto err_class;
 821         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 822                 goto err_class;
 823
 824         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 825
 826         return;
 827
 828 err_class:
 829         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 830
 831 err_cdev:
 832         cdev_del(&uverbs_dev->cdev);
 833         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 834                 clear_bit(devnum, dev_map);
 835         else
 836                 clear_bit(devnum, overflow_map);
 837
 838 err:
 839         if (atomic_dec_and_test(&uverbs_dev->refcount))
 840                 ib_uverbs_comp_dev(uverbs_dev);
 841         wait_for_completion(&uverbs_dev->comp);
 842         kobject_put(&uverbs_dev->kobj);
 843         return;
 844 }
 845
 846 static void ib_uverbs_remove_one(struct ib_device *device)
 847 {
 848         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 849
 850         if (!uverbs_dev)
 851                 return;
 852
 853         dev_set_drvdata(uverbs_dev->dev, NULL);
 854         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 855         cdev_del(&uverbs_dev->cdev);
 856
 857         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 858                 clear_bit(uverbs_dev->devnum, dev_map);
 859         else
 860                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
 861
 862         if (atomic_dec_and_test(&uverbs_dev->refcount))
 863                 ib_uverbs_comp_dev(uverbs_dev);
 864         wait_for_completion(&uverbs_dev->comp);
 865         kobject_put(&uverbs_dev->kobj);
 866 }
 867
 868 static char *uverbs_devnode(struct device *dev, mode_t *mode)
 869 {
 870         if (mode)
 871                 *mode = 0666;
 872         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
 873 }
 874
 875 static int __init ib_uverbs_init(void)
 876 {
 877         int ret;
 878
 879         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 880                                      "infiniband_verbs");
 881         if (ret) {
 882                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 883                 goto out;
 884         }
 885
 886         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 887         if (IS_ERR(uverbs_class)) {
 888                 ret = PTR_ERR(uverbs_class);
 889                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 890                 goto out_chrdev;
 891         }
 892
 893         uverbs_class->devnode = uverbs_devnode;
 894
 895         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
 896         if (ret) {
 897                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 898                 goto out_class;
 899         }
 900
 901         ret = ib_register_client(&uverbs_client);
 902         if (ret) {
 903                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 904                 goto out_class;
 905         }
 906
 907         return 0;
 908
 909 out_class:
 910         class_destroy(uverbs_class);
 911
 912 out_chrdev:
 913         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 914
 915 out:
 916         return ret;
 917 }
 918
 919 static void __exit ib_uverbs_cleanup(void)
 920 {
 921         ib_unregister_client(&uverbs_client);
 922         class_destroy(uverbs_class);
 923         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 924         if (overflow_maj)
 925                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
 926         idr_destroy(&ib_uverbs_pd_idr);
 927         idr_destroy(&ib_uverbs_mr_idr);
 928         idr_destroy(&ib_uverbs_mw_idr);
 929         idr_destroy(&ib_uverbs_ah_idr);
 930         idr_destroy(&ib_uverbs_cq_idr);
 931         idr_destroy(&ib_uverbs_qp_idr);
 932         idr_destroy(&ib_uverbs_srq_idr);
 933 }
 934
 935 module_init(ib_uverbs_init);
 936 module_exit(ib_uverbs_cleanup);