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/device.h>
  39 #include <linux/err.h>
  40 #include <linux/fs.h>
  41 #include <linux/poll.h>
  42 #include <linux/file.h>
  43 #include <linux/mount.h>
  44 #include <linux/cdev.h>
  45
  46 #include <asm/uaccess.h>
  47
  48 #include "uverbs.h"
  49
  50 MODULE_AUTHOR("Roland Dreier");
  51 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  52 MODULE_LICENSE("Dual BSD/GPL");
  53
  54 #define INFINIBANDEVENTFS_MAGIC 0x49426576      /* "IBev" */
  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 DEFINE_IDR(ib_uverbs_xrc_domain_idr);
  75
  76 static spinlock_t map_lock;
  77 static struct ib_uverbs_device *dev_table[IB_UVERBS_MAX_DEVICES];
  78 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  79
  80 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  81                                      const char __user *buf, int in_len,
  82                                      int out_len) = {
  83         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  84         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  85         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  86         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  87         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  88         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  89         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  90         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  91         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  92         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  93         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  94         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  95         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  96         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  97         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
  98         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
  99         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 100         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 101         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 102         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 103         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 104         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 105         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 106         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 107         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 108         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 109         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 110         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 111         [IB_USER_VERBS_CMD_CREATE_XRC_SRQ]      = ib_uverbs_create_xrc_srq,
 112         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrc_domain,
 113         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrc_domain,
 114         [IB_USER_VERBS_CMD_CREATE_XRC_RCV_QP]   = ib_uverbs_create_xrc_rcv_qp,
 115         [IB_USER_VERBS_CMD_MODIFY_XRC_RCV_QP]   = ib_uverbs_modify_xrc_rcv_qp,
 116         [IB_USER_VERBS_CMD_QUERY_XRC_RCV_QP]    = ib_uverbs_query_xrc_rcv_qp,
 117         [IB_USER_VERBS_CMD_REG_XRC_RCV_QP]      = ib_uverbs_reg_xrc_rcv_qp,
 118         [IB_USER_VERBS_CMD_UNREG_XRC_RCV_QP]    = ib_uverbs_unreg_xrc_rcv_qp,
 119 };
 120
 121 #ifdef __linux__
 122 /* BSD Does not require a fake mountpoint for all files. */
 123 static struct vfsmount *uverbs_event_mnt;
 124 #endif
 125
 126 static void ib_uverbs_add_one(struct ib_device *device);
 127 static void ib_uverbs_remove_one(struct ib_device *device);
 128
 129 static void ib_uverbs_release_dev(struct kref *ref)
 130 {
 131         struct ib_uverbs_device *dev =
 132                 container_of(ref, struct ib_uverbs_device, ref);
 133
 134         complete(&dev->comp);
 135 }
 136
 137 static void ib_uverbs_release_event_file(struct kref *ref)
 138 {
 139         struct ib_uverbs_event_file *file =
 140                 container_of(ref, struct ib_uverbs_event_file, ref);
 141
 142         kfree(file);
 143 }
 144
 145 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 146                           struct ib_uverbs_event_file *ev_file,
 147                           struct ib_ucq_object *uobj)
 148 {
 149         struct ib_uverbs_event *evt, *tmp;
 150
 151         if (ev_file) {
 152                 spin_lock_irq(&ev_file->lock);
 153                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 154                         list_del(&evt->list);
 155                         kfree(evt);
 156                 }
 157                 spin_unlock_irq(&ev_file->lock);
 158
 159                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 160         }
 161
 162         spin_lock_irq(&file->async_file->lock);
 163         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 164                 list_del(&evt->list);
 165                 kfree(evt);
 166         }
 167         spin_unlock_irq(&file->async_file->lock);
 168 }
 169
 170 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 171                               struct ib_uevent_object *uobj)
 172 {
 173         struct ib_uverbs_event *evt, *tmp;
 174
 175         spin_lock_irq(&file->async_file->lock);
 176         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 177                 list_del(&evt->list);
 178                 kfree(evt);
 179         }
 180         spin_unlock_irq(&file->async_file->lock);
 181 }
 182
 183 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 184                                     struct ib_uqp_object *uobj)
 185 {
 186         struct ib_uverbs_mcast_entry *mcast, *tmp;
 187
 188         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 189                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 190                 list_del(&mcast->list);
 191                 kfree(mcast);
 192         }
 193 }
 194
 195 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 196                                       struct ib_ucontext *context)
 197 {
 198         struct ib_uobject *uobj, *tmp;
 199
 200         if (!context)
 201                 return 0;
 202
 203         context->closing = 1;
 204
 205         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 206                 struct ib_ah *ah = uobj->object;
 207
 208                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 209                 ib_destroy_ah(ah);
 210                 kfree(uobj);
 211         }
 212
 213         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 214                 struct ib_qp *qp = uobj->object;
 215                 struct ib_uqp_object *uqp =
 216                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 217
 218                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 219                 ib_uverbs_detach_umcast(qp, uqp);
 220                 ib_destroy_qp(qp);
 221                 ib_uverbs_release_uevent(file, &uqp->uevent);
 222                 kfree(uqp);
 223         }
 224
 225
 226         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 227                 struct ib_srq *srq = uobj->object;
 228                 struct ib_uevent_object *uevent =
 229                         container_of(uobj, struct ib_uevent_object, uobject);
 230
 231                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 232                 ib_destroy_srq(srq);
 233                 ib_uverbs_release_uevent(file, uevent);
 234                 kfree(uevent);
 235         }
 236
 237         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 238                 struct ib_cq *cq = uobj->object;
 239                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 240                 struct ib_ucq_object *ucq =
 241                         container_of(uobj, struct ib_ucq_object, uobject);
 242
 243                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 244                 ib_destroy_cq(cq);
 245                 ib_uverbs_release_ucq(file, ev_file, ucq);
 246                 kfree(ucq);
 247         }
 248
 249         /* XXX Free MWs */
 250
 251         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 252                 struct ib_mr *mr = uobj->object;
 253
 254                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 255                 ib_dereg_mr(mr);
 256                 kfree(uobj);
 257         }
 258
 259         mutex_lock(&file->device->ib_dev->xrcd_table_mutex);
 260         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 261                 struct ib_xrcd *xrcd = uobj->object;
 262                 struct ib_uxrc_rcv_object *xrc_qp_obj, *tmp1;
 263                 struct ib_uxrcd_object *xrcd_uobj =
 264                         container_of(uobj, struct ib_uxrcd_object, uobject);
 265
 266                 list_for_each_entry_safe(xrc_qp_obj, tmp1,
 267                                          &xrcd_uobj->xrc_reg_qp_list, list) {
 268                         list_del(&xrc_qp_obj->list);
 269                         ib_uverbs_cleanup_xrc_rcv_qp(file, xrcd,
 270                                                      xrc_qp_obj->qp_num);
 271                         kfree(xrc_qp_obj);
 272                 }
 273
 274                 idr_remove_uobj(&ib_uverbs_xrc_domain_idr, uobj);
 275                 ib_uverbs_dealloc_xrcd(file->device->ib_dev, xrcd);
 276                 kfree(uobj);
 277         }
 278         mutex_unlock(&file->device->ib_dev->xrcd_table_mutex);
 279
 280         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 281                 struct ib_pd *pd = uobj->object;
 282
 283                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 284                 ib_dealloc_pd(pd);
 285                 kfree(uobj);
 286         }
 287
 288         return context->device->dealloc_ucontext(context);
 289 }
 290
 291 static void ib_uverbs_release_file(struct kref *ref)
 292 {
 293         struct ib_uverbs_file *file =
 294                 container_of(ref, struct ib_uverbs_file, ref);
 295
 296         module_put(file->device->ib_dev->owner);
 297         kref_put(&file->device->ref, ib_uverbs_release_dev);
 298
 299         kfree(file);
 300 }
 301
 302 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 303                                     size_t count, loff_t *pos)
 304 {
 305         struct ib_uverbs_event_file *file = filp->private_data;
 306         struct ib_uverbs_event *event;
 307         int eventsz;
 308         int ret = 0;
 309
 310         spin_lock_irq(&file->lock);
 311
 312         while (list_empty(&file->event_list)) {
 313                 spin_unlock_irq(&file->lock);
 314
 315                 if (filp->f_flags & O_NONBLOCK)
 316                         return -EAGAIN;
 317
 318                 if (wait_event_interruptible(file->poll_wait,
 319                                              !list_empty(&file->event_list)))
 320                         return -ERESTARTSYS;
 321
 322                 spin_lock_irq(&file->lock);
 323         }
 324
 325         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 326
 327         if (file->is_async)
 328                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 329         else
 330                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 331
 332         if (eventsz > count) {
 333                 ret   = -EINVAL;
 334                 event = NULL;
 335         } else {
 336                 list_del(file->event_list.next);
 337                 if (event->counter) {
 338                         ++(*event->counter);
 339                         list_del(&event->obj_list);
 340                 }
 341         }
 342
 343         spin_unlock_irq(&file->lock);
 344
 345         if (event) {
 346                 if (copy_to_user(buf, event, eventsz))
 347                         ret = -EFAULT;
 348                 else
 349                         ret = eventsz;
 350         }
 351
 352         kfree(event);
 353
 354         return ret;
 355 }
 356
 357 static unsigned int ib_uverbs_event_poll(struct file *filp,
 358                                          struct poll_table_struct *wait)
 359 {
 360         unsigned int pollflags = 0;
 361         struct ib_uverbs_event_file *file = filp->private_data;
 362
 363         file->filp = filp;
 364         poll_wait(filp, &file->poll_wait, wait);
 365
 366         spin_lock_irq(&file->lock);
 367         if (!list_empty(&file->event_list))
 368                 pollflags = POLLIN | POLLRDNORM;
 369         spin_unlock_irq(&file->lock);
 370
 371         return pollflags;
 372 }
 373
 374 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 375 {
 376         struct ib_uverbs_event_file *file = filp->private_data;
 377
 378         return fasync_helper(fd, filp, on, &file->async_queue);
 379 }
 380
 381 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 382 {
 383         struct ib_uverbs_event_file *file = filp->private_data;
 384         struct ib_uverbs_event *entry, *tmp;
 385
 386         spin_lock_irq(&file->lock);
 387         file->is_closed = 1;
 388         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 389                 if (entry->counter)
 390                         list_del(&entry->obj_list);
 391                 kfree(entry);
 392         }
 393         spin_unlock_irq(&file->lock);
 394
 395         if (file->is_async) {
 396                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 397                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 398         }
 399         kref_put(&file->ref, ib_uverbs_release_event_file);
 400
 401         return 0;
 402 }
 403
 404 static const struct file_operations uverbs_event_fops = {
 405         .owner   = THIS_MODULE,
 406         .read    = ib_uverbs_event_read,
 407         .poll    = ib_uverbs_event_poll,
 408         .release = ib_uverbs_event_close,
 409         .fasync  = ib_uverbs_event_fasync
 410 };
 411
 412 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 413 {
 414         struct ib_uverbs_event_file    *file = cq_context;
 415         struct ib_ucq_object           *uobj;
 416         struct ib_uverbs_event         *entry;
 417         unsigned long                   flags;
 418
 419         if (!file)
 420                 return;
 421
 422         spin_lock_irqsave(&file->lock, flags);
 423         if (file->is_closed) {
 424                 spin_unlock_irqrestore(&file->lock, flags);
 425                 return;
 426         }
 427
 428         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 429         if (!entry) {
 430                 spin_unlock_irqrestore(&file->lock, flags);
 431                 return;
 432         }
 433
 434         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 435
 436         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 437         entry->counter             = &uobj->comp_events_reported;
 438
 439         list_add_tail(&entry->list, &file->event_list);
 440         list_add_tail(&entry->obj_list, &uobj->comp_list);
 441         spin_unlock_irqrestore(&file->lock, flags);
 442
 443         wake_up_interruptible(&file->poll_wait);
 444         if (file->filp)
 445                 selwakeup(&file->filp->f_selinfo);
 446         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 447 }
 448
 449 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 450                                     __u64 element, __u64 event,
 451                                     struct list_head *obj_list,
 452                                     u32 *counter)
 453 {
 454         struct ib_uverbs_event *entry;
 455         unsigned long flags;
 456
 457         spin_lock_irqsave(&file->async_file->lock, flags);
 458         if (file->async_file->is_closed) {
 459                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 460                 return;
 461         }
 462
 463         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 464         if (!entry) {
 465                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 466                 return;
 467         }
 468
 469         entry->desc.async.element    = element;
 470         entry->desc.async.event_type = event;
 471         entry->counter               = counter;
 472
 473         list_add_tail(&entry->list, &file->async_file->event_list);
 474         if (obj_list)
 475                 list_add_tail(&entry->obj_list, obj_list);
 476         spin_unlock_irqrestore(&file->async_file->lock, flags);
 477
 478         wake_up_interruptible(&file->async_file->poll_wait);
 479         if (file->async_file->filp)
 480                 selwakeup(&file->async_file->filp->f_selinfo);
 481         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 482 }
 483
 484 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 485 {
 486         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 487                                                   struct ib_ucq_object, uobject);
 488
 489         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 490                                 event->event, &uobj->async_list,
 491                                 &uobj->async_events_reported);
 492 }
 493
 494 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 495 {
 496         struct ib_uevent_object *uobj;
 497
 498         uobj = container_of(event->element.qp->uobject,
 499                             struct ib_uevent_object, uobject);
 500
 501         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 502                                 event->event, &uobj->event_list,
 503                                 &uobj->events_reported);
 504 }
 505
 506 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 507 {
 508         struct ib_uevent_object *uobj;
 509
 510         uobj = container_of(event->element.srq->uobject,
 511                             struct ib_uevent_object, uobject);
 512
 513         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 514                                 event->event, &uobj->event_list,
 515                                 &uobj->events_reported);
 516 }
 517
 518 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 519                              struct ib_event *event)
 520 {
 521         struct ib_uverbs_file *file =
 522                 container_of(handler, struct ib_uverbs_file, event_handler);
 523
 524         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 525                                 NULL, NULL);
 526 }
 527
 528 void ib_uverbs_xrc_rcv_qp_event_handler(struct ib_event *event,
 529                                         void *context_ptr)
 530 {
 531         ib_uverbs_async_handler(context_ptr, event->element.xrc_qp_num,
 532                                 event->event, NULL, NULL);
 533 }
 534
 535 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 536                                         int is_async, int *fd)
 537 {
 538         struct ib_uverbs_event_file *ev_file;
 539         struct file *filp;
 540         int ret;
 541
 542         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 543         if (!ev_file)
 544                 return ERR_PTR(-ENOMEM);
 545
 546         kref_init(&ev_file->ref);
 547         spin_lock_init(&ev_file->lock);
 548         INIT_LIST_HEAD(&ev_file->event_list);
 549         init_waitqueue_head(&ev_file->poll_wait);
 550         ev_file->uverbs_file = uverbs_file;
 551         ev_file->async_queue = NULL;
 552         ev_file->is_async    = is_async;
 553         ev_file->is_closed   = 0;
 554         ev_file->filp        = NULL;
 555
 556         *fd = get_unused_fd();
 557         if (*fd < 0) {
 558                 ret = *fd;
 559                 goto err;
 560         }
 561
 562         /*
 563          * fops_get() can't fail here, because we're coming from a
 564          * system call on a uverbs file, which will already have a
 565          * module reference.
 566          */
 567 #ifdef __linux__
 568         filp = alloc_file(uverbs_event_mnt, dget(uverbs_event_mnt->mnt_root),
 569                           FMODE_READ, fops_get(&uverbs_event_fops));
 570 #else
 571         filp = alloc_file(FMODE_READ, fops_get(&uverbs_event_fops));
 572 #endif
 573         if (!filp) {
 574                 ret = -ENFILE;
 575                 goto err_fd;
 576         }
 577
 578         filp->private_data = ev_file;
 579
 580         return filp;
 581
 582 err_fd:
 583         put_unused_fd(*fd);
 584
 585 err:
 586         kfree(ev_file);
 587         return ERR_PTR(ret);
 588 }
 589
 590 /*
 591  * Look up a completion event file by FD.  If lookup is successful,
 592  * takes a ref to the event file struct that it returns; if
 593  * unsuccessful, returns NULL.
 594  */
 595 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 596 {
 597         struct ib_uverbs_event_file *ev_file = NULL;
 598         struct file *filp;
 599
 600         filp = fget(fd);
 601         if (!filp)
 602                 return NULL;
 603
 604         if (filp->f_op != &uverbs_event_fops)
 605                 goto out;
 606
 607         ev_file = filp->private_data;
 608         if (ev_file->is_async) {
 609                 ev_file = NULL;
 610                 goto out;
 611         }
 612
 613         kref_get(&ev_file->ref);
 614
 615 out:
 616         fput(filp);
 617         return ev_file;
 618 }
 619
 620 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 621                              size_t count, loff_t *pos)
 622 {
 623         struct ib_uverbs_file *file = filp->private_data;
 624         struct ib_uverbs_cmd_hdr hdr;
 625
 626         if (count < sizeof hdr)
 627                 return -EINVAL;
 628
 629         if (copy_from_user(&hdr, buf, sizeof hdr))
 630                 return -EFAULT;
 631
 632         if (hdr.in_words * 4 != count)
 633                 return -EINVAL;
 634
 635         if (hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 636             !uverbs_cmd_table[hdr.command]              ||
 637             !(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 638                 return -EINVAL;
 639
 640         if (!file->ucontext &&
 641             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 642                 return -EINVAL;
 643
 644         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 645                                              hdr.in_words * 4, hdr.out_words * 4);
 646 }
 647
 648 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 649 {
 650         struct ib_uverbs_file *file = filp->private_data;
 651
 652         if (!file->ucontext)
 653                 return -ENODEV;
 654         else
 655                 return file->device->ib_dev->mmap(file->ucontext, vma);
 656 }
 657
 658 /*
 659  * ib_uverbs_open() does not need the BKL:
 660  *
 661  *  - dev_table[] accesses are protected by map_lock, the
 662  *    ib_uverbs_device structures are properly reference counted, and
 663  *    everything else is purely local to the file being created, so
 664  *    races against other open calls are not a problem;
 665  *  - there is no ioctl method to race against;
 666  *  - the device is added to dev_table[] as the last part of module
 667  *    initialization, the open method will either immediately run
 668  *    -ENXIO, or all required initialization will be done.
 669  */
 670 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 671 {
 672         struct ib_uverbs_device *dev;
 673         struct ib_uverbs_file *file;
 674         int ret;
 675
 676         spin_lock(&map_lock);
 677         dev = dev_table[iminor(inode) - IB_UVERBS_BASE_MINOR];
 678         if (dev)
 679                 kref_get(&dev->ref);
 680         spin_unlock(&map_lock);
 681
 682         if (!dev)
 683                 return -ENXIO;
 684
 685         if (!try_module_get(dev->ib_dev->owner)) {
 686                 ret = -ENODEV;
 687                 goto err;
 688         }
 689
 690         file = kmalloc(sizeof *file, GFP_KERNEL);
 691         if (!file) {
 692                 ret = -ENOMEM;
 693                 goto err_module;
 694         }
 695
 696         file->device     = dev;
 697         file->ucontext   = NULL;
 698         file->async_file = NULL;
 699         kref_init(&file->ref);
 700         mutex_init(&file->mutex);
 701
 702         filp->private_data = file;
 703
 704         return 0;
 705
 706 err_module:
 707         module_put(dev->ib_dev->owner);
 708
 709 err:
 710         kref_put(&dev->ref, ib_uverbs_release_dev);
 711         return ret;
 712 }
 713
 714 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 715 {
 716         struct ib_uverbs_file *file = filp->private_data;
 717
 718         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 719
 720         if (file->async_file)
 721                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 722
 723         kref_put(&file->ref, ib_uverbs_release_file);
 724
 725         return 0;
 726 }
 727
 728 static const struct file_operations uverbs_fops = {
 729         .owner   = THIS_MODULE,
 730         .write   = ib_uverbs_write,
 731         .open    = ib_uverbs_open,
 732         .release = ib_uverbs_close
 733 };
 734
 735 static const struct file_operations uverbs_mmap_fops = {
 736         .owner   = THIS_MODULE,
 737         .write   = ib_uverbs_write,
 738         .mmap    = ib_uverbs_mmap,
 739         .open    = ib_uverbs_open,
 740         .release = ib_uverbs_close
 741 };
 742
 743 static struct ib_client uverbs_client = {
 744         .name   = "uverbs",
 745         .add    = ib_uverbs_add_one,
 746         .remove = ib_uverbs_remove_one
 747 };
 748
 749 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 750                           char *buf)
 751 {
 752         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 753
 754         if (!dev)
 755                 return -ENODEV;
 756
 757         return sprintf(buf, "%s\n", dev->ib_dev->name);
 758 }
 759 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 760
 761 static ssize_t show_dev_abi_version(struct device *device,
 762                                     struct device_attribute *attr, char *buf)
 763 {
 764         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 765
 766         if (!dev)
 767                 return -ENODEV;
 768
 769         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 770 }
 771 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 772
 773 static ssize_t show_abi_version(struct class *class, struct class_attribute *attr, char *buf)
 774 {
 775         return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
 776 }
 777 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 778
 779 #include <linux/pci.h>
 780
 781 static ssize_t
 782 show_dev_device(struct device *device, struct device_attribute *attr, char *buf)
 783 {
 784         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 785
 786         if (!dev)
 787                 return -ENODEV;
 788
 789         return sprintf(buf, "0x%04x\n",
 790             ((struct pci_dev *)dev->ib_dev->dma_device)->device);
 791 }
 792 static DEVICE_ATTR(device, S_IRUGO, show_dev_device, NULL);
 793
 794 static ssize_t
 795 show_dev_vendor(struct device *device, struct device_attribute *attr, char *buf)
 796 {
 797         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 798
 799         if (!dev)
 800                 return -ENODEV;
 801
 802         return sprintf(buf, "0x%04x\n",
 803             ((struct pci_dev *)dev->ib_dev->dma_device)->vendor);
 804 }
 805 static DEVICE_ATTR(vendor, S_IRUGO, show_dev_vendor, NULL);
 806
 807 struct attribute *device_attrs[] =
 808 {
 809         &dev_attr_device.attr,
 810         &dev_attr_vendor.attr,
 811         NULL
 812 };
 813
 814 static struct attribute_group device_group = {
 815         .name  = "device",
 816         .attrs  = device_attrs
 817 };
 818
 819 static void ib_uverbs_add_one(struct ib_device *device)
 820 {
 821         struct ib_uverbs_device *uverbs_dev;
 822
 823         if (!device->alloc_ucontext)
 824                 return;
 825
 826         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 827         if (!uverbs_dev)
 828                 return;
 829
 830         kref_init(&uverbs_dev->ref);
 831         init_completion(&uverbs_dev->comp);
 832
 833         spin_lock(&map_lock);
 834         uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 835         if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
 836                 spin_unlock(&map_lock);
 837                 goto err;
 838         }
 839         set_bit(uverbs_dev->devnum, dev_map);
 840         spin_unlock(&map_lock);
 841
 842         uverbs_dev->ib_dev           = device;
 843         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 844
 845         uverbs_dev->cdev = cdev_alloc();
 846         if (!uverbs_dev->cdev)
 847                 goto err;
 848         uverbs_dev->cdev->owner = THIS_MODULE;
 849         uverbs_dev->cdev->ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 850         kobject_set_name(&uverbs_dev->cdev->kobj, "uverbs%d", uverbs_dev->devnum);
 851         if (cdev_add(uverbs_dev->cdev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
 852                 goto err_cdev;
 853
 854         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 855                                         uverbs_dev->cdev->dev, uverbs_dev,
 856                                         "uverbs%d", uverbs_dev->devnum);
 857         if (IS_ERR(uverbs_dev->dev))
 858                 goto err_cdev;
 859
 860         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 861                 goto err_class;
 862         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 863                 goto err_class;
 864         if (sysfs_create_group(&uverbs_dev->dev->kobj, &device_group))
 865                 goto err_class;
 866
 867         spin_lock(&map_lock);
 868         dev_table[uverbs_dev->devnum] = uverbs_dev;
 869         spin_unlock(&map_lock);
 870
 871         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 872
 873         return;
 874
 875 err_class:
 876         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 877
 878 err_cdev:
 879         cdev_del(uverbs_dev->cdev);
 880         clear_bit(uverbs_dev->devnum, dev_map);
 881
 882 err:
 883         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 884         wait_for_completion(&uverbs_dev->comp);
 885         kfree(uverbs_dev);
 886         return;
 887 }
 888
 889 static void ib_uverbs_remove_one(struct ib_device *device)
 890 {
 891         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 892
 893         if (!uverbs_dev)
 894                 return;
 895
 896         sysfs_remove_group(&uverbs_dev->dev->kobj, &device_group);
 897         dev_set_drvdata(uverbs_dev->dev, NULL);
 898         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 899         cdev_del(uverbs_dev->cdev);
 900
 901         spin_lock(&map_lock);
 902         dev_table[uverbs_dev->devnum] = NULL;
 903         spin_unlock(&map_lock);
 904
 905         clear_bit(uverbs_dev->devnum, dev_map);
 906
 907         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 908         wait_for_completion(&uverbs_dev->comp);
 909         kfree(uverbs_dev);
 910 }
 911 #ifdef __linux__
 912 static int uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
 913                                const char *dev_name, void *data,
 914                                struct vfsmount *mnt)
 915 {
 916         return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
 917                              INFINIBANDEVENTFS_MAGIC, mnt);
 918 }
 919
 920 static struct file_system_type uverbs_event_fs = {
 921         /* No owner field so module can be unloaded */
 922         .name    = "infinibandeventfs",
 923         .get_sb  = uverbs_event_get_sb,
 924         .kill_sb = kill_litter_super
 925 };
 926 #endif
 927
 928 static int __init ib_uverbs_init(void)
 929 {
 930         int ret;
 931
 932         spin_lock_init(&map_lock);
 933
 934         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 935                                      "infiniband_verbs");
 936         if (ret) {
 937                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 938                 goto out;
 939         }
 940
 941         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 942         if (IS_ERR(uverbs_class)) {
 943                 ret = PTR_ERR(uverbs_class);
 944                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 945                 goto out_chrdev;
 946         }
 947
 948         ret = class_create_file(uverbs_class, &class_attr_abi_version);
 949         if (ret) {
 950                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 951                 goto out_class;
 952         }
 953
 954 #ifdef __linux__
 955         ret = register_filesystem(&uverbs_event_fs);
 956         if (ret) {
 957                 printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
 958                 goto out_class;
 959         }
 960
 961         uverbs_event_mnt = kern_mount(&uverbs_event_fs);
 962         if (IS_ERR(uverbs_event_mnt)) {
 963                 ret = PTR_ERR(uverbs_event_mnt);
 964                 printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
 965                 goto out_fs;
 966         }
 967 #endif
 968
 969         ret = ib_register_client(&uverbs_client);
 970         if (ret) {
 971                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 972                 goto out_mnt;
 973         }
 974
 975         return 0;
 976
 977 out_mnt:
 978 #ifdef __linux__
 979         mntput(uverbs_event_mnt);
 980
 981 out_fs:
 982         unregister_filesystem(&uverbs_event_fs);
 983 #endif
 984
 985 out_class:
 986         class_destroy(uverbs_class);
 987
 988 out_chrdev:
 989         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 990
 991 out:
 992         return ret;
 993 }
 994
 995 static void __exit ib_uverbs_cleanup(void)
 996 {
 997         ib_unregister_client(&uverbs_client);
 998 #ifdef __linux__
 999         mntput(uverbs_event_mnt);
1000         unregister_filesystem(&uverbs_event_fs);
1001 #endif
1002         class_destroy(uverbs_class);
1003         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1004         idr_destroy(&ib_uverbs_pd_idr);
1005         idr_destroy(&ib_uverbs_mr_idr);
1006         idr_destroy(&ib_uverbs_mw_idr);
1007         idr_destroy(&ib_uverbs_ah_idr);
1008         idr_destroy(&ib_uverbs_cq_idr);
1009         idr_destroy(&ib_uverbs_qp_idr);
1010         idr_destroy(&ib_uverbs_srq_idr);
1011 }
1012
1013 module_init(ib_uverbs_init);
1014 module_exit(ib_uverbs_cleanup);