SAFE public projects git trees. - safe/jmp/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
  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/mount.h>
  46 #include <linux/cdev.h>
  47
  48 #include <asm/uaccess.h>
  49
  50 #include "uverbs.h"
  51
  52 MODULE_AUTHOR("Roland Dreier");
  53 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  54 MODULE_LICENSE("Dual BSD/GPL");
  55
  56 #define INFINIBANDEVENTFS_MAGIC 0x49426576      /* "IBev" */
  57
  58 enum {
  59         IB_UVERBS_MAJOR       = 231,
  60         IB_UVERBS_BASE_MINOR  = 192,
  61         IB_UVERBS_MAX_DEVICES = 32
  62 };
  63
  64 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  65
  66 static struct class *uverbs_class;
  67
  68 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  69 DEFINE_IDR(ib_uverbs_pd_idr);
  70 DEFINE_IDR(ib_uverbs_mr_idr);
  71 DEFINE_IDR(ib_uverbs_mw_idr);
  72 DEFINE_IDR(ib_uverbs_ah_idr);
  73 DEFINE_IDR(ib_uverbs_cq_idr);
  74 DEFINE_IDR(ib_uverbs_qp_idr);
  75 DEFINE_IDR(ib_uverbs_srq_idr);
  76
  77 static DEFINE_SPINLOCK(map_lock);
  78 static struct ib_uverbs_device *dev_table[IB_UVERBS_MAX_DEVICES];
  79 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  80
  81 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  82                                      const char __user *buf, int in_len,
  83                                      int out_len) = {
  84         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  85         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  86         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  87         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  88         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  89         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  90         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  91         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  92         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  93         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  94         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  95         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  96         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  97         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  98         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
  99         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 100         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 101         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 102         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 103         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 104         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 105         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 106         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 107         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 108         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 109         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 110         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 111         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 112 };
 113
 114 static struct vfsmount *uverbs_event_mnt;
 115
 116 static void ib_uverbs_add_one(struct ib_device *device);
 117 static void ib_uverbs_remove_one(struct ib_device *device);
 118
 119 static void ib_uverbs_release_dev(struct kref *ref)
 120 {
 121         struct ib_uverbs_device *dev =
 122                 container_of(ref, struct ib_uverbs_device, ref);
 123
 124         complete(&dev->comp);
 125 }
 126
 127 static void ib_uverbs_release_event_file(struct kref *ref)
 128 {
 129         struct ib_uverbs_event_file *file =
 130                 container_of(ref, struct ib_uverbs_event_file, ref);
 131
 132         kfree(file);
 133 }
 134
 135 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 136                           struct ib_uverbs_event_file *ev_file,
 137                           struct ib_ucq_object *uobj)
 138 {
 139         struct ib_uverbs_event *evt, *tmp;
 140
 141         if (ev_file) {
 142                 spin_lock_irq(&ev_file->lock);
 143                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 144                         list_del(&evt->list);
 145                         kfree(evt);
 146                 }
 147                 spin_unlock_irq(&ev_file->lock);
 148
 149                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 150         }
 151
 152         spin_lock_irq(&file->async_file->lock);
 153         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 154                 list_del(&evt->list);
 155                 kfree(evt);
 156         }
 157         spin_unlock_irq(&file->async_file->lock);
 158 }
 159
 160 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 161                               struct ib_uevent_object *uobj)
 162 {
 163         struct ib_uverbs_event *evt, *tmp;
 164
 165         spin_lock_irq(&file->async_file->lock);
 166         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 167                 list_del(&evt->list);
 168                 kfree(evt);
 169         }
 170         spin_unlock_irq(&file->async_file->lock);
 171 }
 172
 173 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 174                                     struct ib_uqp_object *uobj)
 175 {
 176         struct ib_uverbs_mcast_entry *mcast, *tmp;
 177
 178         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 179                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 180                 list_del(&mcast->list);
 181                 kfree(mcast);
 182         }
 183 }
 184
 185 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 186                                       struct ib_ucontext *context)
 187 {
 188         struct ib_uobject *uobj, *tmp;
 189
 190         if (!context)
 191                 return 0;
 192
 193         context->closing = 1;
 194
 195         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 196                 struct ib_ah *ah = uobj->object;
 197
 198                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 199                 ib_destroy_ah(ah);
 200                 kfree(uobj);
 201         }
 202
 203         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 204                 struct ib_qp *qp = uobj->object;
 205                 struct ib_uqp_object *uqp =
 206                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 207
 208                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 209                 ib_uverbs_detach_umcast(qp, uqp);
 210                 ib_destroy_qp(qp);
 211                 ib_uverbs_release_uevent(file, &uqp->uevent);
 212                 kfree(uqp);
 213         }
 214
 215         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 216                 struct ib_cq *cq = uobj->object;
 217                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 218                 struct ib_ucq_object *ucq =
 219                         container_of(uobj, struct ib_ucq_object, uobject);
 220
 221                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 222                 ib_destroy_cq(cq);
 223                 ib_uverbs_release_ucq(file, ev_file, ucq);
 224                 kfree(ucq);
 225         }
 226
 227         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 228                 struct ib_srq *srq = uobj->object;
 229                 struct ib_uevent_object *uevent =
 230                         container_of(uobj, struct ib_uevent_object, uobject);
 231
 232                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 233                 ib_destroy_srq(srq);
 234                 ib_uverbs_release_uevent(file, uevent);
 235                 kfree(uevent);
 236         }
 237
 238         /* XXX Free MWs */
 239
 240         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 241                 struct ib_mr *mr = uobj->object;
 242
 243                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 244                 ib_dereg_mr(mr);
 245                 kfree(uobj);
 246         }
 247
 248         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 249                 struct ib_pd *pd = uobj->object;
 250
 251                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 252                 ib_dealloc_pd(pd);
 253                 kfree(uobj);
 254         }
 255
 256         return context->device->dealloc_ucontext(context);
 257 }
 258
 259 static void ib_uverbs_release_file(struct kref *ref)
 260 {
 261         struct ib_uverbs_file *file =
 262                 container_of(ref, struct ib_uverbs_file, ref);
 263
 264         module_put(file->device->ib_dev->owner);
 265         kref_put(&file->device->ref, ib_uverbs_release_dev);
 266
 267         kfree(file);
 268 }
 269
 270 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 271                                     size_t count, loff_t *pos)
 272 {
 273         struct ib_uverbs_event_file *file = filp->private_data;
 274         struct ib_uverbs_event *event;
 275         int eventsz;
 276         int ret = 0;
 277
 278         spin_lock_irq(&file->lock);
 279
 280         while (list_empty(&file->event_list)) {
 281                 spin_unlock_irq(&file->lock);
 282
 283                 if (filp->f_flags & O_NONBLOCK)
 284                         return -EAGAIN;
 285
 286                 if (wait_event_interruptible(file->poll_wait,
 287                                              !list_empty(&file->event_list)))
 288                         return -ERESTARTSYS;
 289
 290                 spin_lock_irq(&file->lock);
 291         }
 292
 293         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 294
 295         if (file->is_async)
 296                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 297         else
 298                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 299
 300         if (eventsz > count) {
 301                 ret   = -EINVAL;
 302                 event = NULL;
 303         } else {
 304                 list_del(file->event_list.next);
 305                 if (event->counter) {
 306                         ++(*event->counter);
 307                         list_del(&event->obj_list);
 308                 }
 309         }
 310
 311         spin_unlock_irq(&file->lock);
 312
 313         if (event) {
 314                 if (copy_to_user(buf, event, eventsz))
 315                         ret = -EFAULT;
 316                 else
 317                         ret = eventsz;
 318         }
 319
 320         kfree(event);
 321
 322         return ret;
 323 }
 324
 325 static unsigned int ib_uverbs_event_poll(struct file *filp,
 326                                          struct poll_table_struct *wait)
 327 {
 328         unsigned int pollflags = 0;
 329         struct ib_uverbs_event_file *file = filp->private_data;
 330
 331         poll_wait(filp, &file->poll_wait, wait);
 332
 333         spin_lock_irq(&file->lock);
 334         if (!list_empty(&file->event_list))
 335                 pollflags = POLLIN | POLLRDNORM;
 336         spin_unlock_irq(&file->lock);
 337
 338         return pollflags;
 339 }
 340
 341 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 342 {
 343         struct ib_uverbs_event_file *file = filp->private_data;
 344
 345         return fasync_helper(fd, filp, on, &file->async_queue);
 346 }
 347
 348 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 349 {
 350         struct ib_uverbs_event_file *file = filp->private_data;
 351         struct ib_uverbs_event *entry, *tmp;
 352
 353         spin_lock_irq(&file->lock);
 354         file->is_closed = 1;
 355         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 356                 if (entry->counter)
 357                         list_del(&entry->obj_list);
 358                 kfree(entry);
 359         }
 360         spin_unlock_irq(&file->lock);
 361
 362         if (file->is_async) {
 363                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 364                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 365         }
 366         kref_put(&file->ref, ib_uverbs_release_event_file);
 367
 368         return 0;
 369 }
 370
 371 static const struct file_operations uverbs_event_fops = {
 372         .owner   = THIS_MODULE,
 373         .read    = ib_uverbs_event_read,
 374         .poll    = ib_uverbs_event_poll,
 375         .release = ib_uverbs_event_close,
 376         .fasync  = ib_uverbs_event_fasync
 377 };
 378
 379 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 380 {
 381         struct ib_uverbs_event_file    *file = cq_context;
 382         struct ib_ucq_object           *uobj;
 383         struct ib_uverbs_event         *entry;
 384         unsigned long                   flags;
 385
 386         if (!file)
 387                 return;
 388
 389         spin_lock_irqsave(&file->lock, flags);
 390         if (file->is_closed) {
 391                 spin_unlock_irqrestore(&file->lock, flags);
 392                 return;
 393         }
 394
 395         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 396         if (!entry) {
 397                 spin_unlock_irqrestore(&file->lock, flags);
 398                 return;
 399         }
 400
 401         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 402
 403         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 404         entry->counter             = &uobj->comp_events_reported;
 405
 406         list_add_tail(&entry->list, &file->event_list);
 407         list_add_tail(&entry->obj_list, &uobj->comp_list);
 408         spin_unlock_irqrestore(&file->lock, flags);
 409
 410         wake_up_interruptible(&file->poll_wait);
 411         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 412 }
 413
 414 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 415                                     __u64 element, __u64 event,
 416                                     struct list_head *obj_list,
 417                                     u32 *counter)
 418 {
 419         struct ib_uverbs_event *entry;
 420         unsigned long flags;
 421
 422         spin_lock_irqsave(&file->async_file->lock, flags);
 423         if (file->async_file->is_closed) {
 424                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 425                 return;
 426         }
 427
 428         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 429         if (!entry) {
 430                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 431                 return;
 432         }
 433
 434         entry->desc.async.element    = element;
 435         entry->desc.async.event_type = event;
 436         entry->counter               = counter;
 437
 438         list_add_tail(&entry->list, &file->async_file->event_list);
 439         if (obj_list)
 440                 list_add_tail(&entry->obj_list, obj_list);
 441         spin_unlock_irqrestore(&file->async_file->lock, flags);
 442
 443         wake_up_interruptible(&file->async_file->poll_wait);
 444         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 445 }
 446
 447 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 448 {
 449         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 450                                                   struct ib_ucq_object, uobject);
 451
 452         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 453                                 event->event, &uobj->async_list,
 454                                 &uobj->async_events_reported);
 455 }
 456
 457 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 458 {
 459         struct ib_uevent_object *uobj;
 460
 461         uobj = container_of(event->element.qp->uobject,
 462                             struct ib_uevent_object, uobject);
 463
 464         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 465                                 event->event, &uobj->event_list,
 466                                 &uobj->events_reported);
 467 }
 468
 469 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 470 {
 471         struct ib_uevent_object *uobj;
 472
 473         uobj = container_of(event->element.srq->uobject,
 474                             struct ib_uevent_object, uobject);
 475
 476         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 477                                 event->event, &uobj->event_list,
 478                                 &uobj->events_reported);
 479 }
 480
 481 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 482                              struct ib_event *event)
 483 {
 484         struct ib_uverbs_file *file =
 485                 container_of(handler, struct ib_uverbs_file, event_handler);
 486
 487         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 488                                 NULL, NULL);
 489 }
 490
 491 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 492                                         int is_async, int *fd)
 493 {
 494         struct ib_uverbs_event_file *ev_file;
 495         struct file *filp;
 496         int ret;
 497
 498         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 499         if (!ev_file)
 500                 return ERR_PTR(-ENOMEM);
 501
 502         kref_init(&ev_file->ref);
 503         spin_lock_init(&ev_file->lock);
 504         INIT_LIST_HEAD(&ev_file->event_list);
 505         init_waitqueue_head(&ev_file->poll_wait);
 506         ev_file->uverbs_file = uverbs_file;
 507         ev_file->async_queue = NULL;
 508         ev_file->is_async    = is_async;
 509         ev_file->is_closed   = 0;
 510
 511         *fd = get_unused_fd();
 512         if (*fd < 0) {
 513                 ret = *fd;
 514                 goto err;
 515         }
 516
 517         /*
 518          * fops_get() can't fail here, because we're coming from a
 519          * system call on a uverbs file, which will already have a
 520          * module reference.
 521          */
 522         filp = alloc_file(uverbs_event_mnt, dget(uverbs_event_mnt->mnt_root),
 523                           FMODE_READ, fops_get(&uverbs_event_fops));
 524         if (!filp) {
 525                 ret = -ENFILE;
 526                 goto err_fd;
 527         }
 528
 529         filp->private_data = ev_file;
 530
 531         return filp;
 532
 533 err_fd:
 534         put_unused_fd(*fd);
 535
 536 err:
 537         kfree(ev_file);
 538         return ERR_PTR(ret);
 539 }
 540
 541 /*
 542  * Look up a completion event file by FD.  If lookup is successful,
 543  * takes a ref to the event file struct that it returns; if
 544  * unsuccessful, returns NULL.
 545  */
 546 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 547 {
 548         struct ib_uverbs_event_file *ev_file = NULL;
 549         struct file *filp;
 550
 551         filp = fget(fd);
 552         if (!filp)
 553                 return NULL;
 554
 555         if (filp->f_op != &uverbs_event_fops)
 556                 goto out;
 557
 558         ev_file = filp->private_data;
 559         if (ev_file->is_async) {
 560                 ev_file = NULL;
 561                 goto out;
 562         }
 563
 564         kref_get(&ev_file->ref);
 565
 566 out:
 567         fput(filp);
 568         return ev_file;
 569 }
 570
 571 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 572                              size_t count, loff_t *pos)
 573 {
 574         struct ib_uverbs_file *file = filp->private_data;
 575         struct ib_uverbs_cmd_hdr hdr;
 576
 577         if (count < sizeof hdr)
 578                 return -EINVAL;
 579
 580         if (copy_from_user(&hdr, buf, sizeof hdr))
 581                 return -EFAULT;
 582
 583         if (hdr.in_words * 4 != count)
 584                 return -EINVAL;
 585
 586         if (hdr.command < 0                             ||
 587             hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 588             !uverbs_cmd_table[hdr.command])
 589                 return -EINVAL;
 590
 591         if (!file->ucontext &&
 592             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 593                 return -EINVAL;
 594
 595         if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 596                 return -ENOSYS;
 597
 598         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 599                                              hdr.in_words * 4, hdr.out_words * 4);
 600 }
 601
 602 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 603 {
 604         struct ib_uverbs_file *file = filp->private_data;
 605
 606         if (!file->ucontext)
 607                 return -ENODEV;
 608         else
 609                 return file->device->ib_dev->mmap(file->ucontext, vma);
 610 }
 611
 612 /*
 613  * ib_uverbs_open() does not need the BKL:
 614  *
 615  *  - dev_table[] accesses are protected by map_lock, the
 616  *    ib_uverbs_device structures are properly reference counted, and
 617  *    everything else is purely local to the file being created, so
 618  *    races against other open calls are not a problem;
 619  *  - there is no ioctl method to race against;
 620  *  - the device is added to dev_table[] as the last part of module
 621  *    initialization, the open method will either immediately run
 622  *    -ENXIO, or all required initialization will be done.
 623  */
 624 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 625 {
 626         struct ib_uverbs_device *dev;
 627         struct ib_uverbs_file *file;
 628         int ret;
 629
 630         spin_lock(&map_lock);
 631         dev = dev_table[iminor(inode) - IB_UVERBS_BASE_MINOR];
 632         if (dev)
 633                 kref_get(&dev->ref);
 634         spin_unlock(&map_lock);
 635
 636         if (!dev)
 637                 return -ENXIO;
 638
 639         if (!try_module_get(dev->ib_dev->owner)) {
 640                 ret = -ENODEV;
 641                 goto err;
 642         }
 643
 644         file = kmalloc(sizeof *file, GFP_KERNEL);
 645         if (!file) {
 646                 ret = -ENOMEM;
 647                 goto err_module;
 648         }
 649
 650         file->device     = dev;
 651         file->ucontext   = NULL;
 652         file->async_file = NULL;
 653         kref_init(&file->ref);
 654         mutex_init(&file->mutex);
 655
 656         filp->private_data = file;
 657
 658         return 0;
 659
 660 err_module:
 661         module_put(dev->ib_dev->owner);
 662
 663 err:
 664         kref_put(&dev->ref, ib_uverbs_release_dev);
 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 const 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 const 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 device *device, struct device_attribute *attr,
 704                           char *buf)
 705 {
 706         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 707
 708         if (!dev)
 709                 return -ENODEV;
 710
 711         return sprintf(buf, "%s\n", dev->ib_dev->name);
 712 }
 713 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 714
 715 static ssize_t show_dev_abi_version(struct device *device,
 716                                     struct device_attribute *attr, char *buf)
 717 {
 718         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 719
 720         if (!dev)
 721                 return -ENODEV;
 722
 723         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 724 }
 725 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 726
 727 static ssize_t show_abi_version(struct class *class, char *buf)
 728 {
 729         return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
 730 }
 731 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 732
 733 static void ib_uverbs_add_one(struct ib_device *device)
 734 {
 735         struct ib_uverbs_device *uverbs_dev;
 736
 737         if (!device->alloc_ucontext)
 738                 return;
 739
 740         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 741         if (!uverbs_dev)
 742                 return;
 743
 744         kref_init(&uverbs_dev->ref);
 745         init_completion(&uverbs_dev->comp);
 746
 747         spin_lock(&map_lock);
 748         uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 749         if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
 750                 spin_unlock(&map_lock);
 751                 goto err;
 752         }
 753         set_bit(uverbs_dev->devnum, dev_map);
 754         spin_unlock(&map_lock);
 755
 756         uverbs_dev->ib_dev           = device;
 757         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 758
 759         uverbs_dev->cdev = cdev_alloc();
 760         if (!uverbs_dev->cdev)
 761                 goto err;
 762         uverbs_dev->cdev->owner = THIS_MODULE;
 763         uverbs_dev->cdev->ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 764         kobject_set_name(&uverbs_dev->cdev->kobj, "uverbs%d", uverbs_dev->devnum);
 765         if (cdev_add(uverbs_dev->cdev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
 766                 goto err_cdev;
 767
 768         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 769                                         uverbs_dev->cdev->dev, uverbs_dev,
 770                                         "uverbs%d", uverbs_dev->devnum);
 771         if (IS_ERR(uverbs_dev->dev))
 772                 goto err_cdev;
 773
 774         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 775                 goto err_class;
 776         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 777                 goto err_class;
 778
 779         spin_lock(&map_lock);
 780         dev_table[uverbs_dev->devnum] = uverbs_dev;
 781         spin_unlock(&map_lock);
 782
 783         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 784
 785         return;
 786
 787 err_class:
 788         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 789
 790 err_cdev:
 791         cdev_del(uverbs_dev->cdev);
 792         clear_bit(uverbs_dev->devnum, dev_map);
 793
 794 err:
 795         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 796         wait_for_completion(&uverbs_dev->comp);
 797         kfree(uverbs_dev);
 798         return;
 799 }
 800
 801 static void ib_uverbs_remove_one(struct ib_device *device)
 802 {
 803         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 804
 805         if (!uverbs_dev)
 806                 return;
 807
 808         dev_set_drvdata(uverbs_dev->dev, NULL);
 809         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 810         cdev_del(uverbs_dev->cdev);
 811
 812         spin_lock(&map_lock);
 813         dev_table[uverbs_dev->devnum] = NULL;
 814         spin_unlock(&map_lock);
 815
 816         clear_bit(uverbs_dev->devnum, dev_map);
 817
 818         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 819         wait_for_completion(&uverbs_dev->comp);
 820         kfree(uverbs_dev);
 821 }
 822
 823 static int uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
 824                                const char *dev_name, void *data,
 825                                struct vfsmount *mnt)
 826 {
 827         return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
 828                              INFINIBANDEVENTFS_MAGIC, mnt);
 829 }
 830
 831 static struct file_system_type uverbs_event_fs = {
 832         /* No owner field so module can be unloaded */
 833         .name    = "infinibandeventfs",
 834         .get_sb  = uverbs_event_get_sb,
 835         .kill_sb = kill_litter_super
 836 };
 837
 838 static int __init ib_uverbs_init(void)
 839 {
 840         int ret;
 841
 842         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 843                                      "infiniband_verbs");
 844         if (ret) {
 845                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 846                 goto out;
 847         }
 848
 849         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 850         if (IS_ERR(uverbs_class)) {
 851                 ret = PTR_ERR(uverbs_class);
 852                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 853                 goto out_chrdev;
 854         }
 855
 856         ret = class_create_file(uverbs_class, &class_attr_abi_version);
 857         if (ret) {
 858                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 859                 goto out_class;
 860         }
 861
 862         ret = register_filesystem(&uverbs_event_fs);
 863         if (ret) {
 864                 printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
 865                 goto out_class;
 866         }
 867
 868         uverbs_event_mnt = kern_mount(&uverbs_event_fs);
 869         if (IS_ERR(uverbs_event_mnt)) {
 870                 ret = PTR_ERR(uverbs_event_mnt);
 871                 printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
 872                 goto out_fs;
 873         }
 874
 875         ret = ib_register_client(&uverbs_client);
 876         if (ret) {
 877                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 878                 goto out_mnt;
 879         }
 880
 881         return 0;
 882
 883 out_mnt:
 884         mntput(uverbs_event_mnt);
 885
 886 out_fs:
 887         unregister_filesystem(&uverbs_event_fs);
 888
 889 out_class:
 890         class_destroy(uverbs_class);
 891
 892 out_chrdev:
 893         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 894
 895 out:
 896         return ret;
 897 }
 898
 899 static void __exit ib_uverbs_cleanup(void)
 900 {
 901         ib_unregister_client(&uverbs_client);
 902         mntput(uverbs_event_mnt);
 903         unregister_filesystem(&uverbs_event_fs);
 904         class_destroy(uverbs_class);
 905         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 906         idr_destroy(&ib_uverbs_pd_idr);
 907         idr_destroy(&ib_uverbs_mr_idr);
 908         idr_destroy(&ib_uverbs_mw_idr);
 909         idr_destroy(&ib_uverbs_ah_idr);
 910         idr_destroy(&ib_uverbs_cq_idr);
 911         idr_destroy(&ib_uverbs_qp_idr);
 912         idr_destroy(&ib_uverbs_srq_idr);
 913 }
 914
 915 module_init(ib_uverbs_init);
 916 module_exit(ib_uverbs_cleanup);