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