Pileus Git - ~andy/linux/blob - drivers/infiniband/core/uverbs_main.c

   1 /*
   2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
   6  * Copyright (c) 2005 PathScale, Inc. All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  */
  36
  37 #include <linux/module.h>
  38 #include <linux/init.h>
  39 #include <linux/device.h>
  40 #include <linux/err.h>
  41 #include <linux/fs.h>
  42 #include <linux/poll.h>
  43 #include <linux/sched.h>
  44 #include <linux/file.h>
  45 #include <linux/cdev.h>
  46 #include <linux/anon_inodes.h>
  47
  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 enum {
  57         IB_UVERBS_MAJOR       = 231,
  58         IB_UVERBS_BASE_MINOR  = 192,
  59         IB_UVERBS_MAX_DEVICES = 32
  60 };
  61
  62 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  63
  64 static struct class *uverbs_class;
  65
  66 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  67 DEFINE_IDR(ib_uverbs_pd_idr);
  68 DEFINE_IDR(ib_uverbs_mr_idr);
  69 DEFINE_IDR(ib_uverbs_mw_idr);
  70 DEFINE_IDR(ib_uverbs_ah_idr);
  71 DEFINE_IDR(ib_uverbs_cq_idr);
  72 DEFINE_IDR(ib_uverbs_qp_idr);
  73 DEFINE_IDR(ib_uverbs_srq_idr);
  74
  75 static DEFINE_SPINLOCK(map_lock);
  76 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  77
  78 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  79                                      const char __user *buf, int in_len,
  80                                      int out_len) = {
  81         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  82         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  83         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  84         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  85         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  86         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  87         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  88         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  89         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  90         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  91         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  92         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  93         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  94         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  95         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
  96         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
  97         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
  98         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
  99         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 100         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 101         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 102         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 103         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 104         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 105         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 106         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 107         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 108         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 109 };
 110
 111 static void ib_uverbs_add_one(struct ib_device *device);
 112 static void ib_uverbs_remove_one(struct ib_device *device);
 113
 114 static void ib_uverbs_release_dev(struct kref *ref)
 115 {
 116         struct ib_uverbs_device *dev =
 117                 container_of(ref, struct ib_uverbs_device, ref);
 118
 119         complete(&dev->comp);
 120 }
 121
 122 static void ib_uverbs_release_event_file(struct kref *ref)
 123 {
 124         struct ib_uverbs_event_file *file =
 125                 container_of(ref, struct ib_uverbs_event_file, ref);
 126
 127         kfree(file);
 128 }
 129
 130 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 131                           struct ib_uverbs_event_file *ev_file,
 132                           struct ib_ucq_object *uobj)
 133 {
 134         struct ib_uverbs_event *evt, *tmp;
 135
 136         if (ev_file) {
 137                 spin_lock_irq(&ev_file->lock);
 138                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 139                         list_del(&evt->list);
 140                         kfree(evt);
 141                 }
 142                 spin_unlock_irq(&ev_file->lock);
 143
 144                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 145         }
 146
 147         spin_lock_irq(&file->async_file->lock);
 148         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 149                 list_del(&evt->list);
 150                 kfree(evt);
 151         }
 152         spin_unlock_irq(&file->async_file->lock);
 153 }
 154
 155 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 156                               struct ib_uevent_object *uobj)
 157 {
 158         struct ib_uverbs_event *evt, *tmp;
 159
 160         spin_lock_irq(&file->async_file->lock);
 161         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 162                 list_del(&evt->list);
 163                 kfree(evt);
 164         }
 165         spin_unlock_irq(&file->async_file->lock);
 166 }
 167
 168 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 169                                     struct ib_uqp_object *uobj)
 170 {
 171         struct ib_uverbs_mcast_entry *mcast, *tmp;
 172
 173         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 174                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 175                 list_del(&mcast->list);
 176                 kfree(mcast);
 177         }
 178 }
 179
 180 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 181                                       struct ib_ucontext *context)
 182 {
 183         struct ib_uobject *uobj, *tmp;
 184
 185         if (!context)
 186                 return 0;
 187
 188         context->closing = 1;
 189
 190         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 191                 struct ib_ah *ah = uobj->object;
 192
 193                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 194                 ib_destroy_ah(ah);
 195                 kfree(uobj);
 196         }
 197
 198         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 199                 struct ib_qp *qp = uobj->object;
 200                 struct ib_uqp_object *uqp =
 201                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 202
 203                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 204                 ib_uverbs_detach_umcast(qp, uqp);
 205                 ib_destroy_qp(qp);
 206                 ib_uverbs_release_uevent(file, &uqp->uevent);
 207                 kfree(uqp);
 208         }
 209
 210         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 211                 struct ib_cq *cq = uobj->object;
 212                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 213                 struct ib_ucq_object *ucq =
 214                         container_of(uobj, struct ib_ucq_object, uobject);
 215
 216                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 217                 ib_destroy_cq(cq);
 218                 ib_uverbs_release_ucq(file, ev_file, ucq);
 219                 kfree(ucq);
 220         }
 221
 222         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 223                 struct ib_srq *srq = uobj->object;
 224                 struct ib_uevent_object *uevent =
 225                         container_of(uobj, struct ib_uevent_object, uobject);
 226
 227                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 228                 ib_destroy_srq(srq);
 229                 ib_uverbs_release_uevent(file, uevent);
 230                 kfree(uevent);
 231         }
 232
 233         /* XXX Free MWs */
 234
 235         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 236                 struct ib_mr *mr = uobj->object;
 237
 238                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 239                 ib_dereg_mr(mr);
 240                 kfree(uobj);
 241         }
 242
 243         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 244                 struct ib_pd *pd = uobj->object;
 245
 246                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 247                 ib_dealloc_pd(pd);
 248                 kfree(uobj);
 249         }
 250
 251         return context->device->dealloc_ucontext(context);
 252 }
 253
 254 static void ib_uverbs_release_file(struct kref *ref)
 255 {
 256         struct ib_uverbs_file *file =
 257                 container_of(ref, struct ib_uverbs_file, ref);
 258
 259         module_put(file->device->ib_dev->owner);
 260         kref_put(&file->device->ref, ib_uverbs_release_dev);
 261
 262         kfree(file);
 263 }
 264
 265 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 266                                     size_t count, loff_t *pos)
 267 {
 268         struct ib_uverbs_event_file *file = filp->private_data;
 269         struct ib_uverbs_event *event;
 270         int eventsz;
 271         int ret = 0;
 272
 273         spin_lock_irq(&file->lock);
 274
 275         while (list_empty(&file->event_list)) {
 276                 spin_unlock_irq(&file->lock);
 277
 278                 if (filp->f_flags & O_NONBLOCK)
 279                         return -EAGAIN;
 280
 281                 if (wait_event_interruptible(file->poll_wait,
 282                                              !list_empty(&file->event_list)))
 283                         return -ERESTARTSYS;
 284
 285                 spin_lock_irq(&file->lock);
 286         }
 287
 288         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 289
 290         if (file->is_async)
 291                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 292         else
 293                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 294
 295         if (eventsz > count) {
 296                 ret   = -EINVAL;
 297                 event = NULL;
 298         } else {
 299                 list_del(file->event_list.next);
 300                 if (event->counter) {
 301                         ++(*event->counter);
 302                         list_del(&event->obj_list);
 303                 }
 304         }
 305
 306         spin_unlock_irq(&file->lock);
 307
 308         if (event) {
 309                 if (copy_to_user(buf, event, eventsz))
 310                         ret = -EFAULT;
 311                 else
 312                         ret = eventsz;
 313         }
 314
 315         kfree(event);
 316
 317         return ret;
 318 }
 319
 320 static unsigned int ib_uverbs_event_poll(struct file *filp,
 321                                          struct poll_table_struct *wait)
 322 {
 323         unsigned int pollflags = 0;
 324         struct ib_uverbs_event_file *file = filp->private_data;
 325
 326         poll_wait(filp, &file->poll_wait, wait);
 327
 328         spin_lock_irq(&file->lock);
 329         if (!list_empty(&file->event_list))
 330                 pollflags = POLLIN | POLLRDNORM;
 331         spin_unlock_irq(&file->lock);
 332
 333         return pollflags;
 334 }
 335
 336 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 337 {
 338         struct ib_uverbs_event_file *file = filp->private_data;
 339
 340         return fasync_helper(fd, filp, on, &file->async_queue);
 341 }
 342
 343 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 344 {
 345         struct ib_uverbs_event_file *file = filp->private_data;
 346         struct ib_uverbs_event *entry, *tmp;
 347
 348         spin_lock_irq(&file->lock);
 349         file->is_closed = 1;
 350         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 351                 if (entry->counter)
 352                         list_del(&entry->obj_list);
 353                 kfree(entry);
 354         }
 355         spin_unlock_irq(&file->lock);
 356
 357         if (file->is_async) {
 358                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 359                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 360         }
 361         kref_put(&file->ref, ib_uverbs_release_event_file);
 362
 363         return 0;
 364 }
 365
 366 static const struct file_operations uverbs_event_fops = {
 367         .owner   = THIS_MODULE,
 368         .read    = ib_uverbs_event_read,
 369         .poll    = ib_uverbs_event_poll,
 370         .release = ib_uverbs_event_close,
 371         .fasync  = ib_uverbs_event_fasync
 372 };
 373
 374 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 375 {
 376         struct ib_uverbs_event_file    *file = cq_context;
 377         struct ib_ucq_object           *uobj;
 378         struct ib_uverbs_event         *entry;
 379         unsigned long                   flags;
 380
 381         if (!file)
 382                 return;
 383
 384         spin_lock_irqsave(&file->lock, flags);
 385         if (file->is_closed) {
 386                 spin_unlock_irqrestore(&file->lock, flags);
 387                 return;
 388         }
 389
 390         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 391         if (!entry) {
 392                 spin_unlock_irqrestore(&file->lock, flags);
 393                 return;
 394         }
 395
 396         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 397
 398         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 399         entry->counter             = &uobj->comp_events_reported;
 400
 401         list_add_tail(&entry->list, &file->event_list);
 402         list_add_tail(&entry->obj_list, &uobj->comp_list);
 403         spin_unlock_irqrestore(&file->lock, flags);
 404
 405         wake_up_interruptible(&file->poll_wait);
 406         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 407 }
 408
 409 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 410                                     __u64 element, __u64 event,
 411                                     struct list_head *obj_list,
 412                                     u32 *counter)
 413 {
 414         struct ib_uverbs_event *entry;
 415         unsigned long flags;
 416
 417         spin_lock_irqsave(&file->async_file->lock, flags);
 418         if (file->async_file->is_closed) {
 419                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 420                 return;
 421         }
 422
 423         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 424         if (!entry) {
 425                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 426                 return;
 427         }
 428
 429         entry->desc.async.element    = element;
 430         entry->desc.async.event_type = event;
 431         entry->counter               = counter;
 432
 433         list_add_tail(&entry->list, &file->async_file->event_list);
 434         if (obj_list)
 435                 list_add_tail(&entry->obj_list, obj_list);
 436         spin_unlock_irqrestore(&file->async_file->lock, flags);
 437
 438         wake_up_interruptible(&file->async_file->poll_wait);
 439         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 440 }
 441
 442 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 443 {
 444         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 445                                                   struct ib_ucq_object, uobject);
 446
 447         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 448                                 event->event, &uobj->async_list,
 449                                 &uobj->async_events_reported);
 450 }
 451
 452 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 453 {
 454         struct ib_uevent_object *uobj;
 455
 456         uobj = container_of(event->element.qp->uobject,
 457                             struct ib_uevent_object, uobject);
 458
 459         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 460                                 event->event, &uobj->event_list,
 461                                 &uobj->events_reported);
 462 }
 463
 464 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 465 {
 466         struct ib_uevent_object *uobj;
 467
 468         uobj = container_of(event->element.srq->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_event_handler(struct ib_event_handler *handler,
 477                              struct ib_event *event)
 478 {
 479         struct ib_uverbs_file *file =
 480                 container_of(handler, struct ib_uverbs_file, event_handler);
 481
 482         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 483                                 NULL, NULL);
 484 }
 485
 486 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 487                                         int is_async, int *fd)
 488 {
 489         struct ib_uverbs_event_file *ev_file;
 490         struct file *filp;
 491         int ret;
 492
 493         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 494         if (!ev_file)
 495                 return ERR_PTR(-ENOMEM);
 496
 497         kref_init(&ev_file->ref);
 498         spin_lock_init(&ev_file->lock);
 499         INIT_LIST_HEAD(&ev_file->event_list);
 500         init_waitqueue_head(&ev_file->poll_wait);
 501         ev_file->uverbs_file = uverbs_file;
 502         ev_file->async_queue = NULL;
 503         ev_file->is_async    = is_async;
 504         ev_file->is_closed   = 0;
 505
 506         *fd = get_unused_fd();
 507         if (*fd < 0) {
 508                 ret = *fd;
 509                 goto err;
 510         }
 511
 512         filp = anon_inode_getfile("[uverbs-event]", &uverbs_event_fops,
 513                                   ev_file, O_RDONLY);
 514         if (!filp) {
 515                 ret = -ENFILE;
 516                 goto err_fd;
 517         }
 518
 519         return filp;
 520
 521 err_fd:
 522         put_unused_fd(*fd);
 523
 524 err:
 525         kfree(ev_file);
 526         return ERR_PTR(ret);
 527 }
 528
 529 /*
 530  * Look up a completion event file by FD.  If lookup is successful,
 531  * takes a ref to the event file struct that it returns; if
 532  * unsuccessful, returns NULL.
 533  */
 534 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 535 {
 536         struct ib_uverbs_event_file *ev_file = NULL;
 537         struct file *filp;
 538
 539         filp = fget(fd);
 540         if (!filp)
 541                 return NULL;
 542
 543         if (filp->f_op != &uverbs_event_fops)
 544                 goto out;
 545
 546         ev_file = filp->private_data;
 547         if (ev_file->is_async) {
 548                 ev_file = NULL;
 549                 goto out;
 550         }
 551
 552         kref_get(&ev_file->ref);
 553
 554 out:
 555         fput(filp);
 556         return ev_file;
 557 }
 558
 559 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 560                              size_t count, loff_t *pos)
 561 {
 562         struct ib_uverbs_file *file = filp->private_data;
 563         struct ib_uverbs_cmd_hdr hdr;
 564
 565         if (count < sizeof hdr)
 566                 return -EINVAL;
 567
 568         if (copy_from_user(&hdr, buf, sizeof hdr))
 569                 return -EFAULT;
 570
 571         if (hdr.in_words * 4 != count)
 572                 return -EINVAL;
 573
 574         if (hdr.command < 0                             ||
 575             hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 576             !uverbs_cmd_table[hdr.command])
 577                 return -EINVAL;
 578
 579         if (!file->ucontext &&
 580             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 581                 return -EINVAL;
 582
 583         if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 584                 return -ENOSYS;
 585
 586         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 587                                              hdr.in_words * 4, hdr.out_words * 4);
 588 }
 589
 590 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 591 {
 592         struct ib_uverbs_file *file = filp->private_data;
 593
 594         if (!file->ucontext)
 595                 return -ENODEV;
 596         else
 597                 return file->device->ib_dev->mmap(file->ucontext, vma);
 598 }
 599
 600 /*
 601  * ib_uverbs_open() does not need the BKL:
 602  *
 603  *  - the ib_uverbs_device structures are properly reference counted and
 604  *    everything else is purely local to the file being created, so
 605  *    races against other open calls are not a problem;
 606  *  - there is no ioctl method to race against;
 607  *  - the open method will either immediately run -ENXIO, or all
 608  *    required initialization will be done.
 609  */
 610 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 611 {
 612         struct ib_uverbs_device *dev;
 613         struct ib_uverbs_file *file;
 614         int ret;
 615
 616         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 617         if (dev)
 618                 kref_get(&dev->ref);
 619         else
 620                 return -ENXIO;
 621
 622         if (!try_module_get(dev->ib_dev->owner)) {
 623                 ret = -ENODEV;
 624                 goto err;
 625         }
 626
 627         file = kmalloc(sizeof *file, GFP_KERNEL);
 628         if (!file) {
 629                 ret = -ENOMEM;
 630                 goto err_module;
 631         }
 632
 633         file->device     = dev;
 634         file->ucontext   = NULL;
 635         file->async_file = NULL;
 636         kref_init(&file->ref);
 637         mutex_init(&file->mutex);
 638
 639         filp->private_data = file;
 640
 641         return 0;
 642
 643 err_module:
 644         module_put(dev->ib_dev->owner);
 645
 646 err:
 647         kref_put(&dev->ref, ib_uverbs_release_dev);
 648         return ret;
 649 }
 650
 651 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 652 {
 653         struct ib_uverbs_file *file = filp->private_data;
 654
 655         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 656
 657         if (file->async_file)
 658                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 659
 660         kref_put(&file->ref, ib_uverbs_release_file);
 661
 662         return 0;
 663 }
 664
 665 static const struct file_operations uverbs_fops = {
 666         .owner   = THIS_MODULE,
 667         .write   = ib_uverbs_write,
 668         .open    = ib_uverbs_open,
 669         .release = ib_uverbs_close
 670 };
 671
 672 static const struct file_operations uverbs_mmap_fops = {
 673         .owner   = THIS_MODULE,
 674         .write   = ib_uverbs_write,
 675         .mmap    = ib_uverbs_mmap,
 676         .open    = ib_uverbs_open,
 677         .release = ib_uverbs_close
 678 };
 679
 680 static struct ib_client uverbs_client = {
 681         .name   = "uverbs",
 682         .add    = ib_uverbs_add_one,
 683         .remove = ib_uverbs_remove_one
 684 };
 685
 686 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 687                           char *buf)
 688 {
 689         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 690
 691         if (!dev)
 692                 return -ENODEV;
 693
 694         return sprintf(buf, "%s\n", dev->ib_dev->name);
 695 }
 696 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 697
 698 static ssize_t show_dev_abi_version(struct device *device,
 699                                     struct device_attribute *attr, char *buf)
 700 {
 701         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 702
 703         if (!dev)
 704                 return -ENODEV;
 705
 706         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 707 }
 708 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 709
 710 static ssize_t show_abi_version(struct class *class, char *buf)
 711 {
 712         return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
 713 }
 714 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 715
 716 static dev_t overflow_maj;
 717 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
 718
 719 /*
 720  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
 721  * requesting a new major number and doubling the number of max devices we
 722  * support. It's stupid, but simple.
 723  */
 724 static int find_overflow_devnum(void)
 725 {
 726         int ret;
 727
 728         if (!overflow_maj) {
 729                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
 730                                           "infiniband_verbs");
 731                 if (ret) {
 732                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
 733                         return ret;
 734                 }
 735         }
 736
 737         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
 738         if (ret >= IB_UVERBS_MAX_DEVICES)
 739                 return -1;
 740
 741         return ret;
 742 }
 743
 744 static void ib_uverbs_add_one(struct ib_device *device)
 745 {
 746         int devnum;
 747         dev_t base;
 748         struct ib_uverbs_device *uverbs_dev;
 749
 750         if (!device->alloc_ucontext)
 751                 return;
 752
 753         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 754         if (!uverbs_dev)
 755                 return;
 756
 757         kref_init(&uverbs_dev->ref);
 758         init_completion(&uverbs_dev->comp);
 759
 760         spin_lock(&map_lock);
 761         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 762         if (devnum >= IB_UVERBS_MAX_DEVICES) {
 763                 spin_unlock(&map_lock);
 764                 devnum = find_overflow_devnum();
 765                 if (devnum < 0)
 766                         goto err;
 767
 768                 spin_lock(&map_lock);
 769                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
 770                 base = devnum + overflow_maj;
 771                 set_bit(devnum, overflow_map);
 772         } else {
 773                 uverbs_dev->devnum = devnum;
 774                 base = devnum + IB_UVERBS_BASE_DEV;
 775                 set_bit(devnum, dev_map);
 776         }
 777         spin_unlock(&map_lock);
 778
 779         uverbs_dev->ib_dev           = device;
 780         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 781
 782         cdev_init(&uverbs_dev->cdev, NULL);
 783         uverbs_dev->cdev.owner = THIS_MODULE;
 784         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 785         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
 786         if (cdev_add(&uverbs_dev->cdev, base, 1))
 787                 goto err_cdev;
 788
 789         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 790                                         uverbs_dev->cdev.dev, uverbs_dev,
 791                                         "uverbs%d", uverbs_dev->devnum);
 792         if (IS_ERR(uverbs_dev->dev))
 793                 goto err_cdev;
 794
 795         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 796                 goto err_class;
 797         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 798                 goto err_class;
 799
 800         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 801
 802         return;
 803
 804 err_class:
 805         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 806
 807 err_cdev:
 808         cdev_del(&uverbs_dev->cdev);
 809         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 810                 clear_bit(devnum, dev_map);
 811         else
 812                 clear_bit(devnum, overflow_map);
 813
 814 err:
 815         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 816         wait_for_completion(&uverbs_dev->comp);
 817         kfree(uverbs_dev);
 818         return;
 819 }
 820
 821 static void ib_uverbs_remove_one(struct ib_device *device)
 822 {
 823         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 824
 825         if (!uverbs_dev)
 826                 return;
 827
 828         dev_set_drvdata(uverbs_dev->dev, NULL);
 829         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 830         cdev_del(&uverbs_dev->cdev);
 831
 832         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 833                 clear_bit(uverbs_dev->devnum, dev_map);
 834         else
 835                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
 836
 837         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 838         wait_for_completion(&uverbs_dev->comp);
 839         kfree(uverbs_dev);
 840 }
 841
 842 static int __init ib_uverbs_init(void)
 843 {
 844         int ret;
 845
 846         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 847                                      "infiniband_verbs");
 848         if (ret) {
 849                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 850                 goto out;
 851         }
 852
 853         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 854         if (IS_ERR(uverbs_class)) {
 855                 ret = PTR_ERR(uverbs_class);
 856                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 857                 goto out_chrdev;
 858         }
 859
 860         ret = class_create_file(uverbs_class, &class_attr_abi_version);
 861         if (ret) {
 862                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 863                 goto out_class;
 864         }
 865
 866         ret = ib_register_client(&uverbs_client);
 867         if (ret) {
 868                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 869                 goto out_class;
 870         }
 871
 872         return 0;
 873
 874 out_class:
 875         class_destroy(uverbs_class);
 876
 877 out_chrdev:
 878         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 879
 880 out:
 881         return ret;
 882 }
 883
 884 static void __exit ib_uverbs_cleanup(void)
 885 {
 886         ib_unregister_client(&uverbs_client);
 887         class_destroy(uverbs_class);
 888         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 889         if (overflow_maj)
 890                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
 891         idr_destroy(&ib_uverbs_pd_idr);
 892         idr_destroy(&ib_uverbs_mr_idr);
 893         idr_destroy(&ib_uverbs_mw_idr);
 894         idr_destroy(&ib_uverbs_ah_idr);
 895         idr_destroy(&ib_uverbs_cq_idr);
 896         idr_destroy(&ib_uverbs_qp_idr);
 897         idr_destroy(&ib_uverbs_srq_idr);
 898 }
 899
 900 module_init(ib_uverbs_init);
 901 module_exit(ib_uverbs_cleanup);