sys/ofed/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/file.h>
  44 #include <linux/mount.h>
  45 #include <linux/cdev.h>
  46
  47 #include <asm/uaccess.h>
  48
  49 #include "uverbs.h"
  50
  51 MODULE_AUTHOR("Roland Dreier");
  52 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  53 MODULE_LICENSE("Dual BSD/GPL");
  54
  55 #define INFINIBANDEVENTFS_MAGIC 0x49426576      /* "IBev" */
  56
  57 enum {
  58         IB_UVERBS_MAJOR       = 231,
  59         IB_UVERBS_BASE_MINOR  = 192,
  60         IB_UVERBS_MAX_DEVICES = 32
  61 };
  62
  63 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  64
  65 static struct class *uverbs_class;
  66
  67 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  68 DEFINE_IDR(ib_uverbs_pd_idr);
  69 DEFINE_IDR(ib_uverbs_mr_idr);
  70 DEFINE_IDR(ib_uverbs_mw_idr);
  71 DEFINE_IDR(ib_uverbs_ah_idr);
  72 DEFINE_IDR(ib_uverbs_cq_idr);
  73 DEFINE_IDR(ib_uverbs_qp_idr);
  74 DEFINE_IDR(ib_uverbs_srq_idr);
  75 DEFINE_IDR(ib_uverbs_xrc_domain_idr);
  76
  77 static spinlock_t 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         [IB_USER_VERBS_CMD_CREATE_XRC_SRQ]      = ib_uverbs_create_xrc_srq,
 113         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrc_domain,
 114         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrc_domain,
 115         [IB_USER_VERBS_CMD_CREATE_XRC_RCV_QP]   = ib_uverbs_create_xrc_rcv_qp,
 116         [IB_USER_VERBS_CMD_MODIFY_XRC_RCV_QP]   = ib_uverbs_modify_xrc_rcv_qp,
 117         [IB_USER_VERBS_CMD_QUERY_XRC_RCV_QP]    = ib_uverbs_query_xrc_rcv_qp,
 118         [IB_USER_VERBS_CMD_REG_XRC_RCV_QP]      = ib_uverbs_reg_xrc_rcv_qp,
 119         [IB_USER_VERBS_CMD_UNREG_XRC_RCV_QP]    = ib_uverbs_unreg_xrc_rcv_qp,
 120 };
 121
 122 #ifdef __linux__
 123 /* BSD Does not require a fake mountpoint for all files. */
 124 static struct vfsmount *uverbs_event_mnt;
 125 #endif
 126
 127 static void ib_uverbs_add_one(struct ib_device *device);
 128 static void ib_uverbs_remove_one(struct ib_device *device);
 129
 130 static void ib_uverbs_release_dev(struct kref *ref)
 131 {
 132         struct ib_uverbs_device *dev =
 133                 container_of(ref, struct ib_uverbs_device, ref);
 134
 135         complete(&dev->comp);
 136 }
 137
 138 static void ib_uverbs_release_event_file(struct kref *ref)
 139 {
 140         struct ib_uverbs_event_file *file =
 141                 container_of(ref, struct ib_uverbs_event_file, ref);
 142
 143         kfree(file);
 144 }
 145
 146 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 147                           struct ib_uverbs_event_file *ev_file,
 148                           struct ib_ucq_object *uobj)
 149 {
 150         struct ib_uverbs_event *evt, *tmp;
 151
 152         if (ev_file) {
 153                 spin_lock_irq(&ev_file->lock);
 154                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 155                         list_del(&evt->list);
 156                         kfree(evt);
 157                 }
 158                 spin_unlock_irq(&ev_file->lock);
 159
 160                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 161         }
 162
 163         spin_lock_irq(&file->async_file->lock);
 164         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 165                 list_del(&evt->list);
 166                 kfree(evt);
 167         }
 168         spin_unlock_irq(&file->async_file->lock);
 169 }
 170
 171 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 172                               struct ib_uevent_object *uobj)
 173 {
 174         struct ib_uverbs_event *evt, *tmp;
 175
 176         spin_lock_irq(&file->async_file->lock);
 177         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 178                 list_del(&evt->list);
 179                 kfree(evt);
 180         }
 181         spin_unlock_irq(&file->async_file->lock);
 182 }
 183
 184 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 185                                     struct ib_uqp_object *uobj)
 186 {
 187         struct ib_uverbs_mcast_entry *mcast, *tmp;
 188
 189         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 190                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 191                 list_del(&mcast->list);
 192                 kfree(mcast);
 193         }
 194 }
 195
 196 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 197                                       struct ib_ucontext *context)
 198 {
 199         struct ib_uobject *uobj, *tmp;
 200
 201         if (!context)
 202                 return 0;
 203
 204         context->closing = 1;
 205
 206         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 207                 struct ib_ah *ah = uobj->object;
 208
 209                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 210                 ib_destroy_ah(ah);
 211                 kfree(uobj);
 212         }
 213
 214         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 215                 struct ib_qp *qp = uobj->object;
 216                 struct ib_uqp_object *uqp =
 217                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 218
 219                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 220                 ib_uverbs_detach_umcast(qp, uqp);
 221                 ib_destroy_qp(qp);
 222                 ib_uverbs_release_uevent(file, &uqp->uevent);
 223                 kfree(uqp);
 224         }
 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         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 239                 struct ib_cq *cq = uobj->object;
 240                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 241                 struct ib_ucq_object *ucq =
 242                         container_of(uobj, struct ib_ucq_object, uobject);
 243
 244                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 245                 ib_destroy_cq(cq);
 246                 ib_uverbs_release_ucq(file, ev_file, ucq);
 247                 kfree(ucq);
 248         }
 249
 250         /* XXX Free MWs */
 251
 252         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 253                 struct ib_mr *mr = uobj->object;
 254
 255                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 256                 ib_dereg_mr(mr);
 257                 kfree(uobj);
 258         }
 259
 260         mutex_lock(&file->device->ib_dev->xrcd_table_mutex);
 261         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 262                 struct ib_xrcd *xrcd = uobj->object;
 263                 struct ib_uxrc_rcv_object *xrc_qp_obj, *tmp1;
 264                 struct ib_uxrcd_object *xrcd_uobj =
 265                         container_of(uobj, struct ib_uxrcd_object, uobject);
 266
 267                 list_for_each_entry_safe(xrc_qp_obj, tmp1,
 268                                          &xrcd_uobj->xrc_reg_qp_list, list) {
 269                         list_del(&xrc_qp_obj->list);
 270                         ib_uverbs_cleanup_xrc_rcv_qp(file, xrcd,
 271                                                      xrc_qp_obj->qp_num);
 272                         kfree(xrc_qp_obj);
 273                 }
 274
 275                 idr_remove_uobj(&ib_uverbs_xrc_domain_idr, uobj);
 276                 ib_uverbs_dealloc_xrcd(file->device->ib_dev, xrcd);
 277                 kfree(uobj);
 278         }
 279         mutex_unlock(&file->device->ib_dev->xrcd_table_mutex);
 280
 281         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 282                 struct ib_pd *pd = uobj->object;
 283
 284                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 285                 ib_dealloc_pd(pd);
 286                 kfree(uobj);
 287         }
 288
 289         return context->device->dealloc_ucontext(context);
 290 }
 291
 292 static void ib_uverbs_release_file(struct kref *ref)
 293 {
 294         struct ib_uverbs_file *file =
 295                 container_of(ref, struct ib_uverbs_file, ref);
 296
 297         module_put(file->device->ib_dev->owner);
 298         kref_put(&file->device->ref, ib_uverbs_release_dev);
 299
 300         kfree(file);
 301 }
 302
 303 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 304                                     size_t count, loff_t *pos)
 305 {
 306         struct ib_uverbs_event_file *file = filp->private_data;
 307         struct ib_uverbs_event *event;
 308         int eventsz;
 309         int ret = 0;
 310
 311         spin_lock_irq(&file->lock);
 312
 313         while (list_empty(&file->event_list)) {
 314                 spin_unlock_irq(&file->lock);
 315
 316                 if (filp->f_flags & O_NONBLOCK)
 317                         return -EAGAIN;
 318
 319                 if (wait_event_interruptible(file->poll_wait,
 320                                              !list_empty(&file->event_list)))
 321                         return -ERESTARTSYS;
 322
 323                 spin_lock_irq(&file->lock);
 324         }
 325
 326         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 327
 328         if (file->is_async)
 329                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 330         else
 331                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 332
 333         if (eventsz > count) {
 334                 ret   = -EINVAL;
 335                 event = NULL;
 336         } else {
 337                 list_del(file->event_list.next);
 338                 if (event->counter) {
 339                         ++(*event->counter);
 340                         list_del(&event->obj_list);
 341                 }
 342         }
 343
 344         spin_unlock_irq(&file->lock);
 345
 346         if (event) {
 347                 if (copy_to_user(buf, event, eventsz))
 348                         ret = -EFAULT;
 349                 else
 350                         ret = eventsz;
 351         }
 352
 353         kfree(event);
 354
 355         return ret;
 356 }
 357
 358 static unsigned int ib_uverbs_event_poll(struct file *filp,
 359                                          struct poll_table_struct *wait)
 360 {
 361         unsigned int pollflags = 0;
 362         struct ib_uverbs_event_file *file = filp->private_data;
 363
 364         file->filp = filp;
 365         poll_wait(filp, &file->poll_wait, wait);
 366
 367         spin_lock_irq(&file->lock);
 368         if (!list_empty(&file->event_list))
 369                 pollflags = POLLIN | POLLRDNORM;
 370         spin_unlock_irq(&file->lock);
 371
 372         return pollflags;
 373 }
 374
 375 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 376 {
 377         struct ib_uverbs_event_file *file = filp->private_data;
 378
 379         return fasync_helper(fd, filp, on, &file->async_queue);
 380 }
 381
 382 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 383 {
 384         struct ib_uverbs_event_file *file = filp->private_data;
 385         struct ib_uverbs_event *entry, *tmp;
 386
 387         spin_lock_irq(&file->lock);
 388         file->is_closed = 1;
 389         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 390                 if (entry->counter)
 391                         list_del(&entry->obj_list);
 392                 kfree(entry);
 393         }
 394         spin_unlock_irq(&file->lock);
 395
 396         if (file->is_async) {
 397                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 398                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 399         }
 400         kref_put(&file->ref, ib_uverbs_release_event_file);
 401
 402         return 0;
 403 }
 404
 405 static const struct file_operations uverbs_event_fops = {
 406         .owner   = THIS_MODULE,
 407         .read    = ib_uverbs_event_read,
 408         .poll    = ib_uverbs_event_poll,
 409         .release = ib_uverbs_event_close,
 410         .fasync  = ib_uverbs_event_fasync
 411 };
 412
 413 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 414 {
 415         struct ib_uverbs_event_file    *file = cq_context;
 416         struct ib_ucq_object           *uobj;
 417         struct ib_uverbs_event         *entry;
 418         unsigned long                   flags;
 419
 420         if (!file)
 421                 return;
 422
 423         spin_lock_irqsave(&file->lock, flags);
 424         if (file->is_closed) {
 425                 spin_unlock_irqrestore(&file->lock, flags);
 426                 return;
 427         }
 428
 429         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 430         if (!entry) {
 431                 spin_unlock_irqrestore(&file->lock, flags);
 432                 return;
 433         }
 434
 435         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 436
 437         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 438         entry->counter             = &uobj->comp_events_reported;
 439
 440         list_add_tail(&entry->list, &file->event_list);
 441         list_add_tail(&entry->obj_list, &uobj->comp_list);
 442         spin_unlock_irqrestore(&file->lock, flags);
 443
 444         wake_up_interruptible(&file->poll_wait);
 445         if (file->filp)
 446                 selwakeup(&file->filp->f_selinfo);
 447         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 448 }
 449
 450 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 451                                     __u64 element, __u64 event,
 452                                     struct list_head *obj_list,
 453                                     u32 *counter)
 454 {
 455         struct ib_uverbs_event *entry;
 456         unsigned long flags;
 457
 458         spin_lock_irqsave(&file->async_file->lock, flags);
 459         if (file->async_file->is_closed) {
 460                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 461                 return;
 462         }
 463
 464         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 465         if (!entry) {
 466                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 467                 return;
 468         }
 469
 470         entry->desc.async.element    = element;
 471         entry->desc.async.event_type = event;
 472         entry->counter               = counter;
 473
 474         list_add_tail(&entry->list, &file->async_file->event_list);
 475         if (obj_list)
 476                 list_add_tail(&entry->obj_list, obj_list);
 477         spin_unlock_irqrestore(&file->async_file->lock, flags);
 478
 479         wake_up_interruptible(&file->async_file->poll_wait);
 480         if (file->async_file->filp)
 481                 selwakeup(&file->async_file->filp->f_selinfo);
 482         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 483 }
 484
 485 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 486 {
 487         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 488                                                   struct ib_ucq_object, uobject);
 489
 490         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 491                                 event->event, &uobj->async_list,
 492                                 &uobj->async_events_reported);
 493 }
 494
 495 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 496 {
 497         struct ib_uevent_object *uobj;
 498
 499         uobj = container_of(event->element.qp->uobject,
 500                             struct ib_uevent_object, uobject);
 501
 502         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 503                                 event->event, &uobj->event_list,
 504                                 &uobj->events_reported);
 505 }
 506
 507 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 508 {
 509         struct ib_uevent_object *uobj;
 510
 511         uobj = container_of(event->element.srq->uobject,
 512                             struct ib_uevent_object, uobject);
 513
 514         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 515                                 event->event, &uobj->event_list,
 516                                 &uobj->events_reported);
 517 }
 518
 519 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 520                              struct ib_event *event)
 521 {
 522         struct ib_uverbs_file *file =
 523                 container_of(handler, struct ib_uverbs_file, event_handler);
 524
 525         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 526                                 NULL, NULL);
 527 }
 528
 529 void ib_uverbs_xrc_rcv_qp_event_handler(struct ib_event *event,
 530                                         void *context_ptr)
 531 {
 532         ib_uverbs_async_handler(context_ptr, event->element.xrc_qp_num,
 533                                 event->event, NULL, NULL);
 534 }
 535
 536 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 537                                         int is_async, int *fd)
 538 {
 539         struct ib_uverbs_event_file *ev_file;
 540         struct file *filp;
 541         int ret;
 542
 543         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 544         if (!ev_file)
 545                 return ERR_PTR(-ENOMEM);
 546
 547         kref_init(&ev_file->ref);
 548         spin_lock_init(&ev_file->lock);
 549         INIT_LIST_HEAD(&ev_file->event_list);
 550         init_waitqueue_head(&ev_file->poll_wait);
 551         ev_file->uverbs_file = uverbs_file;
 552         ev_file->async_queue = NULL;
 553         ev_file->is_async    = is_async;
 554         ev_file->is_closed   = 0;
 555         ev_file->filp        = NULL;
 556
 557         *fd = get_unused_fd();
 558         if (*fd < 0) {
 559                 ret = *fd;
 560                 goto err;
 561         }
 562
 563         /*
 564          * fops_get() can't fail here, because we're coming from a
 565          * system call on a uverbs file, which will already have a
 566          * module reference.
 567          */
 568         filp = alloc_file(uverbs_event_mnt, dget(uverbs_event_mnt->mnt_root),
 569                           FMODE_READ, fops_get(&uverbs_event_fops));
 570         if (!filp) {
 571                 ret = -ENFILE;
 572                 goto err_fd;
 573         }
 574
 575         filp->private_data = ev_file;
 576
 577         return filp;
 578
 579 err_fd:
 580         put_unused_fd(*fd);
 581
 582 err:
 583         kfree(ev_file);
 584         return ERR_PTR(ret);
 585 }
 586
 587 /*
 588  * Look up a completion event file by FD.  If lookup is successful,
 589  * takes a ref to the event file struct that it returns; if
 590  * unsuccessful, returns NULL.
 591  */
 592 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 593 {
 594         struct ib_uverbs_event_file *ev_file = NULL;
 595         struct file *filp;
 596
 597         filp = fget(fd);
 598         if (!filp)
 599                 return NULL;
 600
 601         if (filp->f_op != &uverbs_event_fops)
 602                 goto out;
 603
 604         ev_file = filp->private_data;
 605         if (ev_file->is_async) {
 606                 ev_file = NULL;
 607                 goto out;
 608         }
 609
 610         kref_get(&ev_file->ref);
 611
 612 out:
 613         fput(filp);
 614         return ev_file;
 615 }
 616
 617 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 618                              size_t count, loff_t *pos)
 619 {
 620         struct ib_uverbs_file *file = filp->private_data;
 621         struct ib_uverbs_cmd_hdr hdr;
 622
 623         if (count < sizeof hdr)
 624                 return -EINVAL;
 625
 626         if (copy_from_user(&hdr, buf, sizeof hdr))
 627                 return -EFAULT;
 628
 629         if (hdr.in_words * 4 != count)
 630                 return -EINVAL;
 631
 632         if (hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 633             !uverbs_cmd_table[hdr.command]              ||
 634             !(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 635                 return -EINVAL;
 636
 637         if (!file->ucontext &&
 638             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 639                 return -EINVAL;
 640
 641         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 642                                              hdr.in_words * 4, hdr.out_words * 4);
 643 }
 644
 645 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 646 {
 647         struct ib_uverbs_file *file = filp->private_data;
 648
 649         if (!file->ucontext)
 650                 return -ENODEV;
 651         else
 652                 return file->device->ib_dev->mmap(file->ucontext, vma);
 653 }
 654
 655 /*
 656  * ib_uverbs_open() does not need the BKL:
 657  *
 658  *  - dev_table[] accesses are protected by map_lock, the
 659  *    ib_uverbs_device structures are properly reference counted, and
 660  *    everything else is purely local to the file being created, so
 661  *    races against other open calls are not a problem;
 662  *  - there is no ioctl method to race against;
 663  *  - the device is added to dev_table[] as the last part of module
 664  *    initialization, the open method will either immediately run
 665  *    -ENXIO, or all required initialization will be done.
 666  */
 667 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 668 {
 669         struct ib_uverbs_device *dev;
 670         struct ib_uverbs_file *file;
 671         int ret;
 672
 673         spin_lock(&map_lock);
 674         dev = dev_table[iminor(inode) - IB_UVERBS_BASE_MINOR];
 675         if (dev)
 676                 kref_get(&dev->ref);
 677         spin_unlock(&map_lock);
 678
 679         if (!dev)
 680                 return -ENXIO;
 681
 682         if (!try_module_get(dev->ib_dev->owner)) {
 683                 ret = -ENODEV;
 684                 goto err;
 685         }
 686
 687         file = kmalloc(sizeof *file, GFP_KERNEL);
 688         if (!file) {
 689                 ret = -ENOMEM;
 690                 goto err_module;
 691         }
 692
 693         file->device     = dev;
 694         file->ucontext   = NULL;
 695         file->async_file = NULL;
 696         kref_init(&file->ref);
 697         mutex_init(&file->mutex);
 698
 699         filp->private_data = file;
 700
 701         return 0;
 702
 703 err_module:
 704         module_put(dev->ib_dev->owner);
 705
 706 err:
 707         kref_put(&dev->ref, ib_uverbs_release_dev);
 708         return ret;
 709 }
 710
 711 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 712 {
 713         struct ib_uverbs_file *file = filp->private_data;
 714
 715         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 716
 717         if (file->async_file)
 718                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 719
 720         kref_put(&file->ref, ib_uverbs_release_file);
 721
 722         return 0;
 723 }
 724
 725 static const struct file_operations uverbs_fops = {
 726         .owner   = THIS_MODULE,
 727         .write   = ib_uverbs_write,
 728         .open    = ib_uverbs_open,
 729         .release = ib_uverbs_close
 730 };
 731
 732 static const struct file_operations uverbs_mmap_fops = {
 733         .owner   = THIS_MODULE,
 734         .write   = ib_uverbs_write,
 735         .mmap    = ib_uverbs_mmap,
 736         .open    = ib_uverbs_open,
 737         .release = ib_uverbs_close
 738 };
 739
 740 static struct ib_client uverbs_client = {
 741         .name   = "uverbs",
 742         .add    = ib_uverbs_add_one,
 743         .remove = ib_uverbs_remove_one
 744 };
 745
 746 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 747                           char *buf)
 748 {
 749         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 750
 751         if (!dev)
 752                 return -ENODEV;
 753
 754         return sprintf(buf, "%s\n", dev->ib_dev->name);
 755 }
 756 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 757
 758 static ssize_t show_dev_abi_version(struct device *device,
 759                                     struct device_attribute *attr, char *buf)
 760 {
 761         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 762
 763         if (!dev)
 764                 return -ENODEV;
 765
 766         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 767 }
 768 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 769
 770 static ssize_t show_abi_version(struct class *class, char *buf)
 771 {
 772         return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
 773 }
 774 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 775
 776 #include <linux/pci.h>
 777
 778 static ssize_t
 779 show_dev_device(struct device *device, struct device_attribute *attr, char *buf)
 780 {
 781         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 782
 783         if (!dev)
 784                 return -ENODEV;
 785
 786         return sprintf(buf, "0x%04x\n",
 787             ((struct pci_dev *)dev->ib_dev->dma_device)->device);
 788 }
 789 static DEVICE_ATTR(device, S_IRUGO, show_dev_device, NULL);
 790
 791 static ssize_t
 792 show_dev_vendor(struct device *device, struct device_attribute *attr, char *buf)
 793 {
 794         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 795
 796         if (!dev)
 797                 return -ENODEV;
 798
 799         return sprintf(buf, "0x%04x\n",
 800             ((struct pci_dev *)dev->ib_dev->dma_device)->vendor);
 801 }
 802 static DEVICE_ATTR(vendor, S_IRUGO, show_dev_vendor, NULL);
 803
 804 struct attribute *device_attrs[] =
 805 {
 806         &dev_attr_device.attr,
 807         &dev_attr_vendor.attr,
 808         NULL
 809 };
 810
 811 static struct attribute_group device_group = {
 812         .name  = "device",
 813         .attrs  = device_attrs
 814 };
 815
 816 static void ib_uverbs_add_one(struct ib_device *device)
 817 {
 818         struct ib_uverbs_device *uverbs_dev;
 819
 820         if (!device->alloc_ucontext)
 821                 return;
 822
 823         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 824         if (!uverbs_dev)
 825                 return;
 826
 827         kref_init(&uverbs_dev->ref);
 828         init_completion(&uverbs_dev->comp);
 829
 830         spin_lock(&map_lock);
 831         uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 832         if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
 833                 spin_unlock(&map_lock);
 834                 goto err;
 835         }
 836         set_bit(uverbs_dev->devnum, dev_map);
 837         spin_unlock(&map_lock);
 838
 839         uverbs_dev->ib_dev           = device;
 840         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 841
 842         uverbs_dev->cdev = cdev_alloc();
 843         if (!uverbs_dev->cdev)
 844                 goto err;
 845         uverbs_dev->cdev->owner = THIS_MODULE;
 846         uverbs_dev->cdev->ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 847         kobject_set_name(&uverbs_dev->cdev->kobj, "uverbs%d", uverbs_dev->devnum);
 848         if (cdev_add(uverbs_dev->cdev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
 849                 goto err_cdev;
 850
 851         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 852                                         uverbs_dev->cdev->dev, uverbs_dev,
 853                                         "uverbs%d", uverbs_dev->devnum);
 854         if (IS_ERR(uverbs_dev->dev))
 855                 goto err_cdev;
 856
 857         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 858                 goto err_class;
 859         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 860                 goto err_class;
 861         if (sysfs_create_group(&uverbs_dev->dev->kobj, &device_group))
 862                 goto err_class;
 863
 864         spin_lock(&map_lock);
 865         dev_table[uverbs_dev->devnum] = uverbs_dev;
 866         spin_unlock(&map_lock);
 867
 868         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 869
 870         return;
 871
 872 err_class:
 873         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 874
 875 err_cdev:
 876         cdev_del(uverbs_dev->cdev);
 877         clear_bit(uverbs_dev->devnum, dev_map);
 878
 879 err:
 880         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 881         wait_for_completion(&uverbs_dev->comp);
 882         kfree(uverbs_dev);
 883         return;
 884 }
 885
 886 static void ib_uverbs_remove_one(struct ib_device *device)
 887 {
 888         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 889
 890         if (!uverbs_dev)
 891                 return;
 892
 893         sysfs_remove_group(&uverbs_dev->dev->kobj, &device_group);
 894         dev_set_drvdata(uverbs_dev->dev, NULL);
 895         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 896         cdev_del(uverbs_dev->cdev);
 897
 898         spin_lock(&map_lock);
 899         dev_table[uverbs_dev->devnum] = NULL;
 900         spin_unlock(&map_lock);
 901
 902         clear_bit(uverbs_dev->devnum, dev_map);
 903
 904         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 905         wait_for_completion(&uverbs_dev->comp);
 906         kfree(uverbs_dev);
 907 }
 908 #ifdef __linux__
 909 static int uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
 910                                const char *dev_name, void *data,
 911                                struct vfsmount *mnt)
 912 {
 913         return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
 914                              INFINIBANDEVENTFS_MAGIC, mnt);
 915 }
 916
 917 static struct file_system_type uverbs_event_fs = {
 918         /* No owner field so module can be unloaded */
 919         .name    = "infinibandeventfs",
 920         .get_sb  = uverbs_event_get_sb,
 921         .kill_sb = kill_litter_super
 922 };
 923 #endif
 924
 925 static int __init ib_uverbs_init(void)
 926 {
 927         int ret;
 928
 929         spin_lock_init(&map_lock);
 930
 931         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 932                                      "infiniband_verbs");
 933         if (ret) {
 934                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 935                 goto out;
 936         }
 937
 938         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 939         if (IS_ERR(uverbs_class)) {
 940                 ret = PTR_ERR(uverbs_class);
 941                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 942                 goto out_chrdev;
 943         }
 944
 945         ret = class_create_file(uverbs_class, &class_attr_abi_version);
 946         if (ret) {
 947                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 948                 goto out_class;
 949         }
 950
 951 #ifdef __linux__
 952         ret = register_filesystem(&uverbs_event_fs);
 953         if (ret) {
 954                 printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
 955                 goto out_class;
 956         }
 957
 958         uverbs_event_mnt = kern_mount(&uverbs_event_fs);
 959         if (IS_ERR(uverbs_event_mnt)) {
 960                 ret = PTR_ERR(uverbs_event_mnt);
 961                 printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
 962                 goto out_fs;
 963         }
 964 #endif
 965
 966         ret = ib_register_client(&uverbs_client);
 967         if (ret) {
 968                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 969                 goto out_mnt;
 970         }
 971
 972         return 0;
 973
 974 out_mnt:
 975 #ifdef __linux__
 976         mntput(uverbs_event_mnt);
 977
 978 out_fs:
 979         unregister_filesystem(&uverbs_event_fs);
 980 #endif
 981
 982 out_class:
 983         class_destroy(uverbs_class);
 984
 985 out_chrdev:
 986         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 987
 988 out:
 989         return ret;
 990 }
 991
 992 static void __exit ib_uverbs_cleanup(void)
 993 {
 994         ib_unregister_client(&uverbs_client);
 995 #ifdef __linux__
 996         mntput(uverbs_event_mnt);
 997         unregister_filesystem(&uverbs_event_fs);
 998 #endif
 999         class_destroy(uverbs_class);
1000         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1001         idr_destroy(&ib_uverbs_pd_idr);
1002         idr_destroy(&ib_uverbs_mr_idr);
1003         idr_destroy(&ib_uverbs_mw_idr);
1004         idr_destroy(&ib_uverbs_ah_idr);
1005         idr_destroy(&ib_uverbs_cq_idr);
1006         idr_destroy(&ib_uverbs_qp_idr);
1007         idr_destroy(&ib_uverbs_srq_idr);
1008 }
1009
1010 module_init(ib_uverbs_init);
1011 module_exit(ib_uverbs_cleanup);