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/cdev.h>
  44
  45 #include <asm/uaccess.h>
  46
  47 #include "uverbs.h"
  48
  49 MODULE_AUTHOR("Roland Dreier");
  50 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  51 MODULE_LICENSE("Dual BSD/GPL");
  52
  53 #define INFINIBANDEVENTFS_MAGIC 0x49426576      /* "IBev" */
  54
  55 enum {
  56         IB_UVERBS_MAJOR       = 231,
  57         IB_UVERBS_BASE_MINOR  = 192,
  58         IB_UVERBS_MAX_DEVICES = 32
  59 };
  60
  61 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  62
  63 static struct class *uverbs_class;
  64
  65 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  66 DEFINE_IDR(ib_uverbs_pd_idr);
  67 DEFINE_IDR(ib_uverbs_mr_idr);
  68 DEFINE_IDR(ib_uverbs_mw_idr);
  69 DEFINE_IDR(ib_uverbs_ah_idr);
  70 DEFINE_IDR(ib_uverbs_cq_idr);
  71 DEFINE_IDR(ib_uverbs_qp_idr);
  72 DEFINE_IDR(ib_uverbs_srq_idr);
  73 DEFINE_IDR(ib_uverbs_xrc_domain_idr);
  74
  75 static spinlock_t map_lock;
  76 static struct ib_uverbs_device *dev_table[IB_UVERBS_MAX_DEVICES];
  77 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  78
  79 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  80                                      const char __user *buf, int in_len,
  81                                      int out_len) = {
  82         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  83         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  84         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  85         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  86         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  87         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  88         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  89         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  90         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  91         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  92         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  93         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  94         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  95         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  96         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
  97         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
  98         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
  99         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 100         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 101         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 102         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 103         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 104         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 105         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 106         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 107         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 108         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 109         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 110         [IB_USER_VERBS_CMD_CREATE_XRC_SRQ]      = ib_uverbs_create_xrc_srq,
 111         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrc_domain,
 112         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrc_domain,
 113         [IB_USER_VERBS_CMD_CREATE_XRC_RCV_QP]   = ib_uverbs_create_xrc_rcv_qp,
 114         [IB_USER_VERBS_CMD_MODIFY_XRC_RCV_QP]   = ib_uverbs_modify_xrc_rcv_qp,
 115         [IB_USER_VERBS_CMD_QUERY_XRC_RCV_QP]    = ib_uverbs_query_xrc_rcv_qp,
 116         [IB_USER_VERBS_CMD_REG_XRC_RCV_QP]      = ib_uverbs_reg_xrc_rcv_qp,
 117         [IB_USER_VERBS_CMD_UNREG_XRC_RCV_QP]    = ib_uverbs_unreg_xrc_rcv_qp,
 118 };
 119
 120 #ifdef __linux__
 121 /* BSD Does not require a fake mountpoint for all files. */
 122 static struct vfsmount *uverbs_event_mnt;
 123 #endif
 124
 125 static void ib_uverbs_add_one(struct ib_device *device);
 126 static void ib_uverbs_remove_one(struct ib_device *device);
 127
 128 static void ib_uverbs_release_dev(struct kref *ref)
 129 {
 130         struct ib_uverbs_device *dev =
 131                 container_of(ref, struct ib_uverbs_device, ref);
 132
 133         complete(&dev->comp);
 134 }
 135
 136 static void ib_uverbs_release_event_file(struct kref *ref)
 137 {
 138         struct ib_uverbs_event_file *file =
 139                 container_of(ref, struct ib_uverbs_event_file, ref);
 140
 141         kfree(file);
 142 }
 143
 144 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 145                           struct ib_uverbs_event_file *ev_file,
 146                           struct ib_ucq_object *uobj)
 147 {
 148         struct ib_uverbs_event *evt, *tmp;
 149
 150         if (ev_file) {
 151                 spin_lock_irq(&ev_file->lock);
 152                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 153                         list_del(&evt->list);
 154                         kfree(evt);
 155                 }
 156                 spin_unlock_irq(&ev_file->lock);
 157
 158                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 159         }
 160
 161         spin_lock_irq(&file->async_file->lock);
 162         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 163                 list_del(&evt->list);
 164                 kfree(evt);
 165         }
 166         spin_unlock_irq(&file->async_file->lock);
 167 }
 168
 169 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 170                               struct ib_uevent_object *uobj)
 171 {
 172         struct ib_uverbs_event *evt, *tmp;
 173
 174         spin_lock_irq(&file->async_file->lock);
 175         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 176                 list_del(&evt->list);
 177                 kfree(evt);
 178         }
 179         spin_unlock_irq(&file->async_file->lock);
 180 }
 181
 182 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 183                                     struct ib_uqp_object *uobj)
 184 {
 185         struct ib_uverbs_mcast_entry *mcast, *tmp;
 186
 187         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 188                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 189                 list_del(&mcast->list);
 190                 kfree(mcast);
 191         }
 192 }
 193
 194 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 195                                       struct ib_ucontext *context)
 196 {
 197         struct ib_uobject *uobj, *tmp;
 198
 199         if (!context)
 200                 return 0;
 201
 202         context->closing = 1;
 203
 204         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 205                 struct ib_ah *ah = uobj->object;
 206
 207                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 208                 ib_destroy_ah(ah);
 209                 kfree(uobj);
 210         }
 211
 212         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 213                 struct ib_qp *qp = uobj->object;
 214                 struct ib_uqp_object *uqp =
 215                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 216
 217                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 218                 ib_uverbs_detach_umcast(qp, uqp);
 219                 ib_destroy_qp(qp);
 220                 ib_uverbs_release_uevent(file, &uqp->uevent);
 221                 kfree(uqp);
 222         }
 223
 224
 225         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 226                 struct ib_srq *srq = uobj->object;
 227                 struct ib_uevent_object *uevent =
 228                         container_of(uobj, struct ib_uevent_object, uobject);
 229
 230                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 231                 ib_destroy_srq(srq);
 232                 ib_uverbs_release_uevent(file, uevent);
 233                 kfree(uevent);
 234         }
 235
 236         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 237                 struct ib_cq *cq = uobj->object;
 238                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 239                 struct ib_ucq_object *ucq =
 240                         container_of(uobj, struct ib_ucq_object, uobject);
 241
 242                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 243                 ib_destroy_cq(cq);
 244                 ib_uverbs_release_ucq(file, ev_file, ucq);
 245                 kfree(ucq);
 246         }
 247
 248         /* XXX Free MWs */
 249
 250         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 251                 struct ib_mr *mr = uobj->object;
 252
 253                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 254                 ib_dereg_mr(mr);
 255                 kfree(uobj);
 256         }
 257
 258         mutex_lock(&file->device->ib_dev->xrcd_table_mutex);
 259         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 260                 struct ib_xrcd *xrcd = uobj->object;
 261                 struct ib_uxrc_rcv_object *xrc_qp_obj, *tmp1;
 262                 struct ib_uxrcd_object *xrcd_uobj =
 263                         container_of(uobj, struct ib_uxrcd_object, uobject);
 264
 265                 list_for_each_entry_safe(xrc_qp_obj, tmp1,
 266                                          &xrcd_uobj->xrc_reg_qp_list, list) {
 267                         list_del(&xrc_qp_obj->list);
 268                         ib_uverbs_cleanup_xrc_rcv_qp(file, xrcd,
 269                                                      xrc_qp_obj->qp_num);
 270                         kfree(xrc_qp_obj);
 271                 }
 272
 273                 idr_remove_uobj(&ib_uverbs_xrc_domain_idr, uobj);
 274                 ib_uverbs_dealloc_xrcd(file->device->ib_dev, xrcd);
 275                 kfree(uobj);
 276         }
 277         mutex_unlock(&file->device->ib_dev->xrcd_table_mutex);
 278
 279         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 280                 struct ib_pd *pd = uobj->object;
 281
 282                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 283                 ib_dealloc_pd(pd);
 284                 kfree(uobj);
 285         }
 286
 287         return context->device->dealloc_ucontext(context);
 288 }
 289
 290 static void ib_uverbs_release_file(struct kref *ref)
 291 {
 292         struct ib_uverbs_file *file =
 293                 container_of(ref, struct ib_uverbs_file, ref);
 294
 295         module_put(file->device->ib_dev->owner);
 296         kref_put(&file->device->ref, ib_uverbs_release_dev);
 297
 298         kfree(file);
 299 }
 300
 301 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 302                                     size_t count, loff_t *pos)
 303 {
 304         struct ib_uverbs_event_file *file = filp->private_data;
 305         struct ib_uverbs_event *event;
 306         int eventsz;
 307         int ret = 0;
 308
 309         spin_lock_irq(&file->lock);
 310
 311         while (list_empty(&file->event_list)) {
 312                 spin_unlock_irq(&file->lock);
 313
 314                 if (filp->f_flags & O_NONBLOCK)
 315                         return -EAGAIN;
 316
 317                 if (wait_event_interruptible(file->poll_wait,
 318                                              !list_empty(&file->event_list)))
 319                         return -ERESTARTSYS;
 320
 321                 spin_lock_irq(&file->lock);
 322         }
 323
 324         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 325
 326         if (file->is_async)
 327                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 328         else
 329                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 330
 331         if (eventsz > count) {
 332                 ret   = -EINVAL;
 333                 event = NULL;
 334         } else {
 335                 list_del(file->event_list.next);
 336                 if (event->counter) {
 337                         ++(*event->counter);
 338                         list_del(&event->obj_list);
 339                 }
 340         }
 341
 342         spin_unlock_irq(&file->lock);
 343
 344         if (event) {
 345                 if (copy_to_user(buf, event, eventsz))
 346                         ret = -EFAULT;
 347                 else
 348                         ret = eventsz;
 349         }
 350
 351         kfree(event);
 352
 353         return ret;
 354 }
 355
 356 static unsigned int ib_uverbs_event_poll(struct file *filp,
 357                                          struct poll_table_struct *wait)
 358 {
 359         unsigned int pollflags = 0;
 360         struct ib_uverbs_event_file *file = filp->private_data;
 361
 362         file->filp = filp;
 363         poll_wait(filp, &file->poll_wait, wait);
 364
 365         spin_lock_irq(&file->lock);
 366         if (!list_empty(&file->event_list))
 367                 pollflags = POLLIN | POLLRDNORM;
 368         spin_unlock_irq(&file->lock);
 369
 370         return pollflags;
 371 }
 372
 373 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 374 {
 375         struct ib_uverbs_event_file *file = filp->private_data;
 376
 377         return fasync_helper(fd, filp, on, &file->async_queue);
 378 }
 379
 380 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 381 {
 382         struct ib_uverbs_event_file *file = filp->private_data;
 383         struct ib_uverbs_event *entry, *tmp;
 384
 385         spin_lock_irq(&file->lock);
 386         file->is_closed = 1;
 387         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 388                 if (entry->counter)
 389                         list_del(&entry->obj_list);
 390                 kfree(entry);
 391         }
 392         spin_unlock_irq(&file->lock);
 393
 394         if (file->is_async) {
 395                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 396                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 397         }
 398         kref_put(&file->ref, ib_uverbs_release_event_file);
 399
 400         return 0;
 401 }
 402
 403 static const struct file_operations uverbs_event_fops = {
 404         .owner   = THIS_MODULE,
 405         .read    = ib_uverbs_event_read,
 406         .poll    = ib_uverbs_event_poll,
 407         .release = ib_uverbs_event_close,
 408         .fasync  = ib_uverbs_event_fasync
 409 };
 410
 411 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 412 {
 413         struct ib_uverbs_event_file    *file = cq_context;
 414         struct ib_ucq_object           *uobj;
 415         struct ib_uverbs_event         *entry;
 416         unsigned long                   flags;
 417
 418         if (!file)
 419                 return;
 420
 421         spin_lock_irqsave(&file->lock, flags);
 422         if (file->is_closed) {
 423                 spin_unlock_irqrestore(&file->lock, flags);
 424                 return;
 425         }
 426
 427         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 428         if (!entry) {
 429                 spin_unlock_irqrestore(&file->lock, flags);
 430                 return;
 431         }
 432
 433         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 434
 435         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 436         entry->counter             = &uobj->comp_events_reported;
 437
 438         list_add_tail(&entry->list, &file->event_list);
 439         list_add_tail(&entry->obj_list, &uobj->comp_list);
 440         spin_unlock_irqrestore(&file->lock, flags);
 441
 442         wake_up_interruptible(&file->poll_wait);
 443         if (file->filp)
 444                 selwakeup(&file->filp->f_selinfo);
 445         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 446 }
 447
 448 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 449                                     __u64 element, __u64 event,
 450                                     struct list_head *obj_list,
 451                                     u32 *counter)
 452 {
 453         struct ib_uverbs_event *entry;
 454         unsigned long flags;
 455
 456         spin_lock_irqsave(&file->async_file->lock, flags);
 457         if (file->async_file->is_closed) {
 458                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 459                 return;
 460         }
 461
 462         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 463         if (!entry) {
 464                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 465                 return;
 466         }
 467
 468         entry->desc.async.element    = element;
 469         entry->desc.async.event_type = event;
 470         entry->counter               = counter;
 471
 472         list_add_tail(&entry->list, &file->async_file->event_list);
 473         if (obj_list)
 474                 list_add_tail(&entry->obj_list, obj_list);
 475         spin_unlock_irqrestore(&file->async_file->lock, flags);
 476
 477         wake_up_interruptible(&file->async_file->poll_wait);
 478         if (file->async_file->filp)
 479                 selwakeup(&file->async_file->filp->f_selinfo);
 480         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 481 }
 482
 483 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 484 {
 485         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 486                                                   struct ib_ucq_object, uobject);
 487
 488         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 489                                 event->event, &uobj->async_list,
 490                                 &uobj->async_events_reported);
 491 }
 492
 493 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 494 {
 495         struct ib_uevent_object *uobj;
 496
 497         uobj = container_of(event->element.qp->uobject,
 498                             struct ib_uevent_object, uobject);
 499
 500         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 501                                 event->event, &uobj->event_list,
 502                                 &uobj->events_reported);
 503 }
 504
 505 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 506 {
 507         struct ib_uevent_object *uobj;
 508
 509         uobj = container_of(event->element.srq->uobject,
 510                             struct ib_uevent_object, uobject);
 511
 512         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 513                                 event->event, &uobj->event_list,
 514                                 &uobj->events_reported);
 515 }
 516
 517 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 518                              struct ib_event *event)
 519 {
 520         struct ib_uverbs_file *file =
 521                 container_of(handler, struct ib_uverbs_file, event_handler);
 522
 523         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 524                                 NULL, NULL);
 525 }
 526
 527 void ib_uverbs_xrc_rcv_qp_event_handler(struct ib_event *event,
 528                                         void *context_ptr)
 529 {
 530         ib_uverbs_async_handler(context_ptr, event->element.xrc_qp_num,
 531                                 event->event, NULL, NULL);
 532 }
 533
 534 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 535                                         int is_async, int *fd)
 536 {
 537         struct ib_uverbs_event_file *ev_file;
 538         struct file *filp;
 539         int ret;
 540
 541         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 542         if (!ev_file)
 543                 return ERR_PTR(-ENOMEM);
 544
 545         kref_init(&ev_file->ref);
 546         spin_lock_init(&ev_file->lock);
 547         INIT_LIST_HEAD(&ev_file->event_list);
 548         init_waitqueue_head(&ev_file->poll_wait);
 549         ev_file->uverbs_file = uverbs_file;
 550         ev_file->async_queue = NULL;
 551         ev_file->is_async    = is_async;
 552         ev_file->is_closed   = 0;
 553         ev_file->filp        = NULL;
 554
 555         *fd = get_unused_fd();
 556         if (*fd < 0) {
 557                 ret = *fd;
 558                 goto err;
 559         }
 560
 561         /*
 562          * fops_get() can't fail here, because we're coming from a
 563          * system call on a uverbs file, which will already have a
 564          * module reference.
 565          */
 566 #ifdef __linux__
 567         filp = alloc_file(uverbs_event_mnt, dget(uverbs_event_mnt->mnt_root),
 568                           FMODE_READ, fops_get(&uverbs_event_fops));
 569 #else
 570         filp = alloc_file(FMODE_READ, fops_get(&uverbs_event_fops));
 571 #endif
 572         if (!filp) {
 573                 ret = -ENFILE;
 574                 goto err_fd;
 575         }
 576
 577         filp->private_data = ev_file;
 578
 579         return filp;
 580
 581 err_fd:
 582         put_unused_fd(*fd);
 583
 584 err:
 585         kfree(ev_file);
 586         return ERR_PTR(ret);
 587 }
 588
 589 /*
 590  * Look up a completion event file by FD.  If lookup is successful,
 591  * takes a ref to the event file struct that it returns; if
 592  * unsuccessful, returns NULL.
 593  */
 594 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 595 {
 596         struct ib_uverbs_event_file *ev_file = NULL;
 597         struct file *filp;
 598
 599         filp = fget(fd);
 600         if (!filp)
 601                 return NULL;
 602
 603         if (filp->f_op != &uverbs_event_fops)
 604                 goto out;
 605
 606         ev_file = filp->private_data;
 607         if (ev_file->is_async) {
 608                 ev_file = NULL;
 609                 goto out;
 610         }
 611
 612         kref_get(&ev_file->ref);
 613
 614 out:
 615         fput(filp);
 616         return ev_file;
 617 }
 618
 619 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 620                              size_t count, loff_t *pos)
 621 {
 622         struct ib_uverbs_file *file = filp->private_data;
 623         struct ib_uverbs_cmd_hdr hdr;
 624
 625         if (count < sizeof hdr)
 626                 return -EINVAL;
 627
 628         if (copy_from_user(&hdr, buf, sizeof hdr))
 629                 return -EFAULT;
 630
 631         if (hdr.in_words * 4 != count)
 632                 return -EINVAL;
 633
 634         if (hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 635             !uverbs_cmd_table[hdr.command]              ||
 636             !(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 637                 return -EINVAL;
 638
 639         if (!file->ucontext &&
 640             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 641                 return -EINVAL;
 642
 643         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 644                                              hdr.in_words * 4, hdr.out_words * 4);
 645 }
 646
 647 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 648 {
 649         struct ib_uverbs_file *file = filp->private_data;
 650
 651         if (!file->ucontext)
 652                 return -ENODEV;
 653         else
 654                 return file->device->ib_dev->mmap(file->ucontext, vma);
 655 }
 656
 657 /*
 658  * ib_uverbs_open() does not need the BKL:
 659  *
 660  *  - dev_table[] accesses are protected by map_lock, the
 661  *    ib_uverbs_device structures are properly reference counted, and
 662  *    everything else is purely local to the file being created, so
 663  *    races against other open calls are not a problem;
 664  *  - there is no ioctl method to race against;
 665  *  - the device is added to dev_table[] as the last part of module
 666  *    initialization, the open method will either immediately run
 667  *    -ENXIO, or all required initialization will be done.
 668  */
 669 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 670 {
 671         struct ib_uverbs_device *dev;
 672         struct ib_uverbs_file *file;
 673         int ret;
 674
 675         spin_lock(&map_lock);
 676         dev = dev_table[iminor(inode) - IB_UVERBS_BASE_MINOR];
 677         if (dev)
 678                 kref_get(&dev->ref);
 679         spin_unlock(&map_lock);
 680
 681         if (!dev)
 682                 return -ENXIO;
 683
 684         if (!try_module_get(dev->ib_dev->owner)) {
 685                 ret = -ENODEV;
 686                 goto err;
 687         }
 688
 689         file = kmalloc(sizeof *file, GFP_KERNEL);
 690         if (!file) {
 691                 ret = -ENOMEM;
 692                 goto err_module;
 693         }
 694
 695         file->device     = dev;
 696         file->ucontext   = NULL;
 697         file->async_file = NULL;
 698         kref_init(&file->ref);
 699         mutex_init(&file->mutex);
 700
 701         filp->private_data = file;
 702
 703         return 0;
 704
 705 err_module:
 706         module_put(dev->ib_dev->owner);
 707
 708 err:
 709         kref_put(&dev->ref, ib_uverbs_release_dev);
 710         return ret;
 711 }
 712
 713 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 714 {
 715         struct ib_uverbs_file *file = filp->private_data;
 716
 717         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 718
 719         if (file->async_file)
 720                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 721
 722         kref_put(&file->ref, ib_uverbs_release_file);
 723
 724         return 0;
 725 }
 726
 727 static const struct file_operations uverbs_fops = {
 728         .owner   = THIS_MODULE,
 729         .write   = ib_uverbs_write,
 730         .open    = ib_uverbs_open,
 731         .release = ib_uverbs_close
 732 };
 733
 734 static const struct file_operations uverbs_mmap_fops = {
 735         .owner   = THIS_MODULE,
 736         .write   = ib_uverbs_write,
 737         .mmap    = ib_uverbs_mmap,
 738         .open    = ib_uverbs_open,
 739         .release = ib_uverbs_close
 740 };
 741
 742 static struct ib_client uverbs_client = {
 743         .name   = "uverbs",
 744         .add    = ib_uverbs_add_one,
 745         .remove = ib_uverbs_remove_one
 746 };
 747
 748 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 749                           char *buf)
 750 {
 751         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 752
 753         if (!dev)
 754                 return -ENODEV;
 755
 756         return sprintf(buf, "%s\n", dev->ib_dev->name);
 757 }
 758 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 759
 760 static ssize_t show_dev_abi_version(struct device *device,
 761                                     struct device_attribute *attr, char *buf)
 762 {
 763         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 764
 765         if (!dev)
 766                 return -ENODEV;
 767
 768         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 769 }
 770 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 771
 772 static ssize_t show_abi_version(struct class *class, struct class_attribute *attr, char *buf)
 773 {
 774         return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
 775 }
 776 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 777
 778 #include <linux/pci.h>
 779
 780 static ssize_t
 781 show_dev_device(struct device *device, struct device_attribute *attr, char *buf)
 782 {
 783         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 784
 785         if (!dev)
 786                 return -ENODEV;
 787
 788         return sprintf(buf, "0x%04x\n",
 789             ((struct pci_dev *)dev->ib_dev->dma_device)->device);
 790 }
 791 static DEVICE_ATTR(device, S_IRUGO, show_dev_device, NULL);
 792
 793 static ssize_t
 794 show_dev_vendor(struct device *device, struct device_attribute *attr, char *buf)
 795 {
 796         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 797
 798         if (!dev)
 799                 return -ENODEV;
 800
 801         return sprintf(buf, "0x%04x\n",
 802             ((struct pci_dev *)dev->ib_dev->dma_device)->vendor);
 803 }
 804 static DEVICE_ATTR(vendor, S_IRUGO, show_dev_vendor, NULL);
 805
 806 struct attribute *device_attrs[] =
 807 {
 808         &dev_attr_device.attr,
 809         &dev_attr_vendor.attr,
 810         NULL
 811 };
 812
 813 static struct attribute_group device_group = {
 814         .name  = "device",
 815         .attrs  = device_attrs
 816 };
 817
 818 static void ib_uverbs_add_one(struct ib_device *device)
 819 {
 820         struct ib_uverbs_device *uverbs_dev;
 821
 822         if (!device->alloc_ucontext)
 823                 return;
 824
 825         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 826         if (!uverbs_dev)
 827                 return;
 828
 829         kref_init(&uverbs_dev->ref);
 830         init_completion(&uverbs_dev->comp);
 831
 832         spin_lock(&map_lock);
 833         uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 834         if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
 835                 spin_unlock(&map_lock);
 836                 goto err;
 837         }
 838         set_bit(uverbs_dev->devnum, dev_map);
 839         spin_unlock(&map_lock);
 840
 841         uverbs_dev->ib_dev           = device;
 842         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 843
 844         uverbs_dev->cdev = cdev_alloc();
 845         if (!uverbs_dev->cdev)
 846                 goto err;
 847         uverbs_dev->cdev->owner = THIS_MODULE;
 848         uverbs_dev->cdev->ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 849         kobject_set_name(&uverbs_dev->cdev->kobj, "uverbs%d", uverbs_dev->devnum);
 850         if (cdev_add(uverbs_dev->cdev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
 851                 goto err_cdev;
 852
 853         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 854                                         uverbs_dev->cdev->dev, uverbs_dev,
 855                                         "uverbs%d", uverbs_dev->devnum);
 856         if (IS_ERR(uverbs_dev->dev))
 857                 goto err_cdev;
 858
 859         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 860                 goto err_class;
 861         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 862                 goto err_class;
 863         if (sysfs_create_group(&uverbs_dev->dev->kobj, &device_group))
 864                 goto err_class;
 865
 866         spin_lock(&map_lock);
 867         dev_table[uverbs_dev->devnum] = uverbs_dev;
 868         spin_unlock(&map_lock);
 869
 870         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 871
 872         return;
 873
 874 err_class:
 875         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 876
 877 err_cdev:
 878         cdev_del(uverbs_dev->cdev);
 879         clear_bit(uverbs_dev->devnum, dev_map);
 880
 881 err:
 882         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 883         wait_for_completion(&uverbs_dev->comp);
 884         kfree(uverbs_dev);
 885         return;
 886 }
 887
 888 static void ib_uverbs_remove_one(struct ib_device *device)
 889 {
 890         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 891
 892         if (!uverbs_dev)
 893                 return;
 894
 895         sysfs_remove_group(&uverbs_dev->dev->kobj, &device_group);
 896         dev_set_drvdata(uverbs_dev->dev, NULL);
 897         device_destroy(uverbs_class, uverbs_dev->cdev->dev);
 898         cdev_del(uverbs_dev->cdev);
 899
 900         spin_lock(&map_lock);
 901         dev_table[uverbs_dev->devnum] = NULL;
 902         spin_unlock(&map_lock);
 903
 904         clear_bit(uverbs_dev->devnum, dev_map);
 905
 906         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 907         wait_for_completion(&uverbs_dev->comp);
 908         kfree(uverbs_dev);
 909 }
 910 #ifdef __linux__
 911 static int uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
 912                                const char *dev_name, void *data,
 913                                struct vfsmount *mnt)
 914 {
 915         return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
 916                              INFINIBANDEVENTFS_MAGIC, mnt);
 917 }
 918
 919 static struct file_system_type uverbs_event_fs = {
 920         /* No owner field so module can be unloaded */
 921         .name    = "infinibandeventfs",
 922         .get_sb  = uverbs_event_get_sb,
 923         .kill_sb = kill_litter_super
 924 };
 925 #endif
 926
 927 static int __init ib_uverbs_init(void)
 928 {
 929         int ret;
 930
 931         spin_lock_init(&map_lock);
 932
 933         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 934                                      "infiniband_verbs");
 935         if (ret) {
 936                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 937                 goto out;
 938         }
 939
 940         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 941         if (IS_ERR(uverbs_class)) {
 942                 ret = PTR_ERR(uverbs_class);
 943                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 944                 goto out_chrdev;
 945         }
 946
 947         ret = class_create_file(uverbs_class, &class_attr_abi_version);
 948         if (ret) {
 949                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 950                 goto out_class;
 951         }
 952
 953 #ifdef __linux__
 954         ret = register_filesystem(&uverbs_event_fs);
 955         if (ret) {
 956                 printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
 957                 goto out_class;
 958         }
 959
 960         uverbs_event_mnt = kern_mount(&uverbs_event_fs);
 961         if (IS_ERR(uverbs_event_mnt)) {
 962                 ret = PTR_ERR(uverbs_event_mnt);
 963                 printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
 964                 goto out_fs;
 965         }
 966 #endif
 967
 968         ret = ib_register_client(&uverbs_client);
 969         if (ret) {
 970                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 971                 goto out_mnt;
 972         }
 973
 974         return 0;
 975
 976 out_mnt:
 977 #ifdef __linux__
 978         mntput(uverbs_event_mnt);
 979
 980 out_fs:
 981         unregister_filesystem(&uverbs_event_fs);
 982 #endif
 983
 984 out_class:
 985         class_destroy(uverbs_class);
 986
 987 out_chrdev:
 988         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 989
 990 out:
 991         return ret;
 992 }
 993
 994 static void __exit ib_uverbs_cleanup(void)
 995 {
 996         ib_unregister_client(&uverbs_client);
 997 #ifdef __linux__
 998         mntput(uverbs_event_mnt);
 999         unregister_filesystem(&uverbs_event_fs);
1000 #endif
1001         class_destroy(uverbs_class);
1002         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1003         idr_destroy(&ib_uverbs_pd_idr);
1004         idr_destroy(&ib_uverbs_mr_idr);
1005         idr_destroy(&ib_uverbs_mw_idr);
1006         idr_destroy(&ib_uverbs_ah_idr);
1007         idr_destroy(&ib_uverbs_cq_idr);
1008         idr_destroy(&ib_uverbs_qp_idr);
1009         idr_destroy(&ib_uverbs_srq_idr);
1010 }
1011
1012 module_init(ib_uverbs_init);
1013 module_exit(ib_uverbs_cleanup);