git.oblomov.eu Git - linux-2.6/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  * $Id: uverbs_main.c 2733 2005-06-28 19:14:34Z roland $
  37  */
  38
  39 #include <linux/module.h>
  40 #include <linux/init.h>
  41 #include <linux/device.h>
  42 #include <linux/err.h>
  43 #include <linux/fs.h>
  44 #include <linux/poll.h>
  45 #include <linux/file.h>
  46 #include <linux/mount.h>
  47 #include <linux/cdev.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 #define INFINIBANDEVENTFS_MAGIC 0x49426576      /* "IBev" */
  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
  78 static spinlock_t map_lock;
  79 static struct ib_uverbs_device *dev_table[IB_UVERBS_MAX_DEVICES];
  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 };
 114
 115 static struct vfsmount *uverbs_event_mnt;
 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 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 129                           struct ib_uverbs_event_file *ev_file,
 130                           struct ib_ucq_object *uobj)
 131 {
 132         struct ib_uverbs_event *evt, *tmp;
 133
 134         if (ev_file) {
 135                 spin_lock_irq(&ev_file->lock);
 136                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 137                         list_del(&evt->list);
 138                         kfree(evt);
 139                 }
 140                 spin_unlock_irq(&ev_file->lock);
 141
 142                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 143         }
 144
 145         spin_lock_irq(&file->async_file->lock);
 146         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 147                 list_del(&evt->list);
 148                 kfree(evt);
 149         }
 150         spin_unlock_irq(&file->async_file->lock);
 151 }
 152
 153 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 154                               struct ib_uevent_object *uobj)
 155 {
 156         struct ib_uverbs_event *evt, *tmp;
 157
 158         spin_lock_irq(&file->async_file->lock);
 159         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 160                 list_del(&evt->list);
 161                 kfree(evt);
 162         }
 163         spin_unlock_irq(&file->async_file->lock);
 164 }
 165
 166 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 167                                     struct ib_uqp_object *uobj)
 168 {
 169         struct ib_uverbs_mcast_entry *mcast, *tmp;
 170
 171         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 172                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 173                 list_del(&mcast->list);
 174                 kfree(mcast);
 175         }
 176 }
 177
 178 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 179                                       struct ib_ucontext *context)
 180 {
 181         struct ib_uobject *uobj, *tmp;
 182
 183         if (!context)
 184                 return 0;
 185
 186         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 187                 struct ib_ah *ah = uobj->object;
 188
 189                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 190                 ib_destroy_ah(ah);
 191                 kfree(uobj);
 192         }
 193
 194         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 195                 struct ib_qp *qp = uobj->object;
 196                 struct ib_uqp_object *uqp =
 197                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 198
 199                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 200                 ib_uverbs_detach_umcast(qp, uqp);
 201                 ib_destroy_qp(qp);
 202                 ib_uverbs_release_uevent(file, &uqp->uevent);
 203                 kfree(uqp);
 204         }
 205
 206         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 207                 struct ib_cq *cq = uobj->object;
 208                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 209                 struct ib_ucq_object *ucq =
 210                         container_of(uobj, struct ib_ucq_object, uobject);
 211
 212                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 213                 ib_destroy_cq(cq);
 214                 ib_uverbs_release_ucq(file, ev_file, ucq);
 215                 kfree(ucq);
 216         }
 217
 218         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 219                 struct ib_srq *srq = uobj->object;
 220                 struct ib_uevent_object *uevent =
 221                         container_of(uobj, struct ib_uevent_object, uobject);
 222
 223                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 224                 ib_destroy_srq(srq);
 225                 ib_uverbs_release_uevent(file, uevent);
 226                 kfree(uevent);
 227         }
 228
 229         /* XXX Free MWs */
 230
 231         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 232                 struct ib_mr *mr = uobj->object;
 233                 struct ib_device *mrdev = mr->device;
 234                 struct ib_umem_object *memobj;
 235
 236                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 237                 ib_dereg_mr(mr);
 238
 239                 memobj = container_of(uobj, struct ib_umem_object, uobject);
 240                 ib_umem_release_on_close(mrdev, &memobj->umem);
 241
 242                 kfree(memobj);
 243         }
 244
 245         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 246                 struct ib_pd *pd = uobj->object;
 247
 248                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 249                 ib_dealloc_pd(pd);
 250                 kfree(uobj);
 251         }
 252
 253         return context->device->dealloc_ucontext(context);
 254 }
 255
 256 static void ib_uverbs_release_file(struct kref *ref)
 257 {
 258         struct ib_uverbs_file *file =
 259                 container_of(ref, struct ib_uverbs_file, ref);
 260
 261         module_put(file->device->ib_dev->owner);
 262         kref_put(&file->device->ref, ib_uverbs_release_dev);
 263
 264         kfree(file);
 265 }
 266
 267 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 268                                     size_t count, loff_t *pos)
 269 {
 270         struct ib_uverbs_event_file *file = filp->private_data;
 271         struct ib_uverbs_event *event;
 272         int eventsz;
 273         int ret = 0;
 274
 275         spin_lock_irq(&file->lock);
 276
 277         while (list_empty(&file->event_list)) {
 278                 spin_unlock_irq(&file->lock);
 279
 280                 if (filp->f_flags & O_NONBLOCK)
 281                         return -EAGAIN;
 282
 283                 if (wait_event_interruptible(file->poll_wait,
 284                                              !list_empty(&file->event_list)))
 285                         return -ERESTARTSYS;
 286
 287                 spin_lock_irq(&file->lock);
 288         }
 289
 290         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 291
 292         if (file->is_async)
 293                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 294         else
 295                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 296
 297         if (eventsz > count) {
 298                 ret   = -EINVAL;
 299                 event = NULL;
 300         } else {
 301                 list_del(file->event_list.next);
 302                 if (event->counter) {
 303                         ++(*event->counter);
 304                         list_del(&event->obj_list);
 305                 }
 306         }
 307
 308         spin_unlock_irq(&file->lock);
 309
 310         if (event) {
 311                 if (copy_to_user(buf, event, eventsz))
 312                         ret = -EFAULT;
 313                 else
 314                         ret = eventsz;
 315         }
 316
 317         kfree(event);
 318
 319         return ret;
 320 }
 321
 322 static unsigned int ib_uverbs_event_poll(struct file *filp,
 323                                          struct poll_table_struct *wait)
 324 {
 325         unsigned int pollflags = 0;
 326         struct ib_uverbs_event_file *file = filp->private_data;
 327
 328         poll_wait(filp, &file->poll_wait, wait);
 329
 330         spin_lock_irq(&file->lock);
 331         if (!list_empty(&file->event_list))
 332                 pollflags = POLLIN | POLLRDNORM;
 333         spin_unlock_irq(&file->lock);
 334
 335         return pollflags;
 336 }
 337
 338 void ib_uverbs_release_event_file(struct kref *ref)
 339 {
 340         struct ib_uverbs_event_file *file =
 341                 container_of(ref, struct ib_uverbs_event_file, ref);
 342
 343         kfree(file);
 344 }
 345
 346 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 347 {
 348         struct ib_uverbs_event_file *file = filp->private_data;
 349
 350         return fasync_helper(fd, filp, on, &file->async_queue);
 351 }
 352
 353 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 354 {
 355         struct ib_uverbs_event_file *file = filp->private_data;
 356         struct ib_uverbs_event *entry, *tmp;
 357
 358         spin_lock_irq(&file->lock);
 359         file->file = NULL;
 360         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 361                 if (entry->counter)
 362                         list_del(&entry->obj_list);
 363                 kfree(entry);
 364         }
 365         spin_unlock_irq(&file->lock);
 366
 367         ib_uverbs_event_fasync(-1, filp, 0);
 368
 369         if (file->is_async) {
 370                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 371                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 372         }
 373         kref_put(&file->ref, ib_uverbs_release_event_file);
 374
 375         return 0;
 376 }
 377
 378 static struct file_operations uverbs_event_fops = {
 379         .owner   = THIS_MODULE,
 380         .read    = ib_uverbs_event_read,
 381         .poll    = ib_uverbs_event_poll,
 382         .release = ib_uverbs_event_close,
 383         .fasync  = ib_uverbs_event_fasync
 384 };
 385
 386 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 387 {
 388         struct ib_uverbs_event_file    *file = cq_context;
 389         struct ib_ucq_object           *uobj;
 390         struct ib_uverbs_event         *entry;
 391         unsigned long                   flags;
 392
 393         if (!file)
 394                 return;
 395
 396         spin_lock_irqsave(&file->lock, flags);
 397         if (!file->file) {
 398                 spin_unlock_irqrestore(&file->lock, flags);
 399                 return;
 400         }
 401
 402         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 403         if (!entry) {
 404                 spin_unlock_irqrestore(&file->lock, flags);
 405                 return;
 406         }
 407
 408         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 409
 410         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 411         entry->counter             = &uobj->comp_events_reported;
 412
 413         list_add_tail(&entry->list, &file->event_list);
 414         list_add_tail(&entry->obj_list, &uobj->comp_list);
 415         spin_unlock_irqrestore(&file->lock, flags);
 416
 417         wake_up_interruptible(&file->poll_wait);
 418         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 419 }
 420
 421 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 422                                     __u64 element, __u64 event,
 423                                     struct list_head *obj_list,
 424                                     u32 *counter)
 425 {
 426         struct ib_uverbs_event *entry;
 427         unsigned long flags;
 428
 429         spin_lock_irqsave(&file->async_file->lock, flags);
 430         if (!file->async_file->file) {
 431                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 432                 return;
 433         }
 434
 435         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 436         if (!entry) {
 437                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 438                 return;
 439         }
 440
 441         entry->desc.async.element    = element;
 442         entry->desc.async.event_type = event;
 443         entry->counter               = counter;
 444
 445         list_add_tail(&entry->list, &file->async_file->event_list);
 446         if (obj_list)
 447                 list_add_tail(&entry->obj_list, obj_list);
 448         spin_unlock_irqrestore(&file->async_file->lock, flags);
 449
 450         wake_up_interruptible(&file->async_file->poll_wait);
 451         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 452 }
 453
 454 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 455 {
 456         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 457                                                   struct ib_ucq_object, uobject);
 458
 459         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 460                                 event->event, &uobj->async_list,
 461                                 &uobj->async_events_reported);
 462 }
 463
 464 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 465 {
 466         struct ib_uevent_object *uobj;
 467
 468         uobj = container_of(event->element.qp->uobject,
 469                             struct ib_uevent_object, uobject);
 470
 471         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 472                                 event->event, &uobj->event_list,
 473                                 &uobj->events_reported);
 474 }
 475
 476 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 477 {
 478         struct ib_uevent_object *uobj;
 479
 480         uobj = container_of(event->element.srq->uobject,
 481                             struct ib_uevent_object, uobject);
 482
 483         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 484                                 event->event, &uobj->event_list,
 485                                 &uobj->events_reported);
 486 }
 487
 488 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 489                              struct ib_event *event)
 490 {
 491         struct ib_uverbs_file *file =
 492                 container_of(handler, struct ib_uverbs_file, event_handler);
 493
 494         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 495                                 NULL, NULL);
 496 }
 497
 498 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 499                                         int is_async, int *fd)
 500 {
 501         struct ib_uverbs_event_file *ev_file;
 502         struct file *filp;
 503         int ret;
 504
 505         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 506         if (!ev_file)
 507                 return ERR_PTR(-ENOMEM);
 508
 509         kref_init(&ev_file->ref);
 510         spin_lock_init(&ev_file->lock);
 511         INIT_LIST_HEAD(&ev_file->event_list);
 512         init_waitqueue_head(&ev_file->poll_wait);
 513         ev_file->uverbs_file = uverbs_file;
 514         ev_file->async_queue = NULL;
 515         ev_file->is_async    = is_async;
 516
 517         *fd = get_unused_fd();
 518         if (*fd < 0) {
 519                 ret = *fd;
 520                 goto err;
 521         }
 522
 523         filp = get_empty_filp();
 524         if (!filp) {
 525                 ret = -ENFILE;
 526                 goto err_fd;
 527         }
 528
 529         ev_file->file      = filp;
 530
 531         /*
 532          * fops_get() can't fail here, because we're coming from a
 533          * system call on a uverbs file, which will already have a
 534          * module reference.
 535          */
 536         filp->f_op         = fops_get(&uverbs_event_fops);
 537         filp->f_vfsmnt     = mntget(uverbs_event_mnt);
 538         filp->f_dentry     = dget(uverbs_event_mnt->mnt_root);
 539         filp->f_mapping    = filp->f_dentry->d_inode->i_mapping;
 540         filp->f_flags      = O_RDONLY;
 541         filp->f_mode       = FMODE_READ;
 542         filp->private_data = ev_file;
 543
 544         return filp;
 545
 546 err_fd:
 547         put_unused_fd(*fd);
 548
 549 err:
 550         kfree(ev_file);
 551         return ERR_PTR(ret);
 552 }
 553
 554 /*
 555  * Look up a completion event file by FD.  If lookup is successful,
 556  * takes a ref to the event file struct that it returns; if
 557  * unsuccessful, returns NULL.
 558  */
 559 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 560 {
 561         struct ib_uverbs_event_file *ev_file = NULL;
 562         struct file *filp;
 563
 564         filp = fget(fd);
 565         if (!filp)
 566                 return NULL;
 567
 568         if (filp->f_op != &uverbs_event_fops)
 569                 goto out;
 570
 571         ev_file = filp->private_data;
 572         if (ev_file->is_async) {
 573                 ev_file = NULL;
 574                 goto out;
 575         }
 576
 577         kref_get(&ev_file->ref);
 578
 579 out:
 580         fput(filp);
 581         return ev_file;
 582 }
 583
 584 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 585                              size_t count, loff_t *pos)
 586 {
 587         struct ib_uverbs_file *file = filp->private_data;
 588         struct ib_uverbs_cmd_hdr hdr;
 589
 590         if (count < sizeof hdr)
 591                 return -EINVAL;
 592
 593         if (copy_from_user(&hdr, buf, sizeof hdr))
 594                 return -EFAULT;
 595
 596         if (hdr.in_words * 4 != count)
 597                 return -EINVAL;
 598
 599         if (hdr.command < 0                             ||
 600             hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 601             !uverbs_cmd_table[hdr.command]              ||
 602             !(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 603                 return -EINVAL;
 604
 605         if (!file->ucontext &&
 606             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 607                 return -EINVAL;
 608
 609         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 610                                              hdr.in_words * 4, hdr.out_words * 4);
 611 }
 612
 613 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 614 {
 615         struct ib_uverbs_file *file = filp->private_data;
 616
 617         if (!file->ucontext)
 618                 return -ENODEV;
 619         else
 620                 return file->device->ib_dev->mmap(file->ucontext, vma);
 621 }
 622
 623 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 624 {
 625         struct ib_uverbs_device *dev;
 626         struct ib_uverbs_file *file;
 627         int ret;
 628
 629         spin_lock(&map_lock);
 630         dev = dev_table[iminor(inode) - IB_UVERBS_BASE_MINOR];
 631         if (dev)
 632                 kref_get(&dev->ref);
 633         spin_unlock(&map_lock);
 634
 635         if (!dev)
 636                 return -ENXIO;
 637
 638         if (!try_module_get(dev->ib_dev->owner)) {
 639                 ret = -ENODEV;
 640                 goto err;
 641         }
 642
 643         file = kmalloc(sizeof *file, GFP_KERNEL);
 644         if (!file) {
 645                 ret = -ENOMEM;
 646                 goto err_module;
 647         }
 648
 649         file->device     = dev;
 650         file->ucontext   = NULL;
 651         file->async_file = NULL;
 652         kref_init(&file->ref);
 653         mutex_init(&file->mutex);
 654
 655         filp->private_data = file;
 656
 657         return 0;
 658
 659 err_module:
 660         module_put(dev->ib_dev->owner);
 661
 662 err:
 663         kref_put(&dev->ref, ib_uverbs_release_dev);
 664
 665         return ret;
 666 }
 667
 668 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 669 {
 670         struct ib_uverbs_file *file = filp->private_data;
 671
 672         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 673
 674         if (file->async_file)
 675                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 676
 677         kref_put(&file->ref, ib_uverbs_release_file);
 678
 679         return 0;
 680 }
 681
 682 static struct file_operations uverbs_fops = {
 683         .owner   = THIS_MODULE,
 684         .write   = ib_uverbs_write,
 685         .open    = ib_uverbs_open,
 686         .release = ib_uverbs_close
 687 };
 688
 689 static struct file_operations uverbs_mmap_fops = {
 690         .owner   = THIS_MODULE,
 691         .write   = ib_uverbs_write,
 692         .mmap    = ib_uverbs_mmap,
 693         .open    = ib_uverbs_open,
 694         .release = ib_uverbs_close
 695 };
 696
 697 static struct ib_client uverbs_client = {
 698         .name   = "uverbs",
 699         .add    = ib_uverbs_add_one,
 700         .remove = ib_uverbs_remove_one
 701 };
 702
 703 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
 704 {
 705         struct ib_uverbs_device *dev = class_get_devdata(class_dev);
 706
 707         if (!dev)
 708                 return -ENODEV;
 709
 710         return sprintf(buf, "%s\n", dev->ib_dev->name);
 711 }
 712 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 713
 714 static ssize_t show_dev_abi_version(struct class_device *class_dev, char *buf)
 715 {
 716         struct ib_uverbs_device *dev = class_get_devdata(class_dev);
 717
 718         if (!dev)
 719                 return -ENODEV;
 720
 721         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 722 }
 723 static CLASS_DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 724
 725 static ssize_t show_abi_version(struct class *class, char *buf)
 726 {
 727         return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
 728 }
 729 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 730
 731 static void ib_uverbs_add_one(struct ib_device *device)
 732 {
 733         struct ib_uverbs_device *uverbs_dev;
 734
 735         if (!device->alloc_ucontext)
 736                 return;
 737
 738         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 739         if (!uverbs_dev)
 740                 return;
 741
 742         kref_init(&uverbs_dev->ref);
 743         init_completion(&uverbs_dev->comp);
 744
 745         spin_lock(&map_lock);
 746         uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 747         if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
 748                 spin_unlock(&map_lock);
 749                 goto err;
 750         }
 751         set_bit(uverbs_dev->devnum, dev_map);
 752         spin_unlock(&map_lock);
 753
 754         uverbs_dev->ib_dev           = device;
 755         uverbs_dev->num_comp_vectors = 1;
 756
 757         uverbs_dev->dev = cdev_alloc();
 758         if (!uverbs_dev->dev)
 759                 goto err;
 760         uverbs_dev->dev->owner = THIS_MODULE;
 761         uverbs_dev->dev->ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 762         kobject_set_name(&uverbs_dev->dev->kobj, "uverbs%d", uverbs_dev->devnum);
 763         if (cdev_add(uverbs_dev->dev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
 764                 goto err_cdev;
 765
 766         uverbs_dev->class_dev = class_device_create(uverbs_class, NULL,
 767                                                     uverbs_dev->dev->dev,
 768                                                     device->dma_device,
 769                                                     "uverbs%d", uverbs_dev->devnum);
 770         if (IS_ERR(uverbs_dev->class_dev))
 771                 goto err_cdev;
 772
 773         class_set_devdata(uverbs_dev->class_dev, uverbs_dev);
 774
 775         if (class_device_create_file(uverbs_dev->class_dev, &class_device_attr_ibdev))
 776                 goto err_class;
 777         if (class_device_create_file(uverbs_dev->class_dev, &class_device_attr_abi_version))
 778                 goto err_class;
 779
 780         spin_lock(&map_lock);
 781         dev_table[uverbs_dev->devnum] = uverbs_dev;
 782         spin_unlock(&map_lock);
 783
 784         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 785
 786         return;
 787
 788 err_class:
 789         class_device_destroy(uverbs_class, uverbs_dev->dev->dev);
 790
 791 err_cdev:
 792         cdev_del(uverbs_dev->dev);
 793         clear_bit(uverbs_dev->devnum, dev_map);
 794
 795 err:
 796         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 797         wait_for_completion(&uverbs_dev->comp);
 798         kfree(uverbs_dev);
 799         return;
 800 }
 801
 802 static void ib_uverbs_remove_one(struct ib_device *device)
 803 {
 804         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 805
 806         if (!uverbs_dev)
 807                 return;
 808
 809         class_set_devdata(uverbs_dev->class_dev, NULL);
 810         class_device_destroy(uverbs_class, uverbs_dev->dev->dev);
 811         cdev_del(uverbs_dev->dev);
 812
 813         spin_lock(&map_lock);
 814         dev_table[uverbs_dev->devnum] = NULL;
 815         spin_unlock(&map_lock);
 816
 817         clear_bit(uverbs_dev->devnum, dev_map);
 818
 819         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 820         wait_for_completion(&uverbs_dev->comp);
 821         kfree(uverbs_dev);
 822 }
 823
 824 static int uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
 825                                const char *dev_name, void *data,
 826                                struct vfsmount *mnt)
 827 {
 828         return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
 829                              INFINIBANDEVENTFS_MAGIC, mnt);
 830 }
 831
 832 static struct file_system_type uverbs_event_fs = {
 833         /* No owner field so module can be unloaded */
 834         .name    = "infinibandeventfs",
 835         .get_sb  = uverbs_event_get_sb,
 836         .kill_sb = kill_litter_super
 837 };
 838
 839 static int __init ib_uverbs_init(void)
 840 {
 841         int ret;
 842
 843         spin_lock_init(&map_lock);
 844
 845         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 846                                      "infiniband_verbs");
 847         if (ret) {
 848                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 849                 goto out;
 850         }
 851
 852         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 853         if (IS_ERR(uverbs_class)) {
 854                 ret = PTR_ERR(uverbs_class);
 855                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 856                 goto out_chrdev;
 857         }
 858
 859         ret = class_create_file(uverbs_class, &class_attr_abi_version);
 860         if (ret) {
 861                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 862                 goto out_class;
 863         }
 864
 865         ret = register_filesystem(&uverbs_event_fs);
 866         if (ret) {
 867                 printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
 868                 goto out_class;
 869         }
 870
 871         uverbs_event_mnt = kern_mount(&uverbs_event_fs);
 872         if (IS_ERR(uverbs_event_mnt)) {
 873                 ret = PTR_ERR(uverbs_event_mnt);
 874                 printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
 875                 goto out_fs;
 876         }
 877
 878         ret = ib_register_client(&uverbs_client);
 879         if (ret) {
 880                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 881                 goto out_mnt;
 882         }
 883
 884         return 0;
 885
 886 out_mnt:
 887         mntput(uverbs_event_mnt);
 888
 889 out_fs:
 890         unregister_filesystem(&uverbs_event_fs);
 891
 892 out_class:
 893         class_destroy(uverbs_class);
 894
 895 out_chrdev:
 896         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 897
 898 out:
 899         return ret;
 900 }
 901
 902 static void __exit ib_uverbs_cleanup(void)
 903 {
 904         ib_unregister_client(&uverbs_client);
 905         mntput(uverbs_event_mnt);
 906         unregister_filesystem(&uverbs_event_fs);
 907         class_destroy(uverbs_class);
 908         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 909         flush_scheduled_work();
 910         idr_destroy(&ib_uverbs_pd_idr);
 911         idr_destroy(&ib_uverbs_mr_idr);
 912         idr_destroy(&ib_uverbs_mw_idr);
 913         idr_destroy(&ib_uverbs_ah_idr);
 914         idr_destroy(&ib_uverbs_cq_idr);
 915         idr_destroy(&ib_uverbs_qp_idr);
 916         idr_destroy(&ib_uverbs_srq_idr);
 917 }
 918
 919 module_init(ib_uverbs_init);
 920 module_exit(ib_uverbs_cleanup);