drivers/infiniband/core/uverbs_main.c

   1 /*
   2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
   6  * Copyright (c) 2005 PathScale, Inc. All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  */
  36
  37 #include <linux/module.h>
  38 #include <linux/init.h>
  39 #include <linux/device.h>
  40 #include <linux/err.h>
  41 #include <linux/fs.h>
  42 #include <linux/poll.h>
  43 #include <linux/sched.h>
  44 #include <linux/file.h>
  45 #include <linux/cdev.h>
  46 #include <linux/anon_inodes.h>
  47 #include <linux/slab.h>
  48
  49 #include <asm/uaccess.h>
  50
  51 #include "uverbs.h"
  52
  53 MODULE_AUTHOR("Roland Dreier");
  54 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  55 MODULE_LICENSE("Dual BSD/GPL");
  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_xrcd_idr);
  76 DEFINE_IDR(ib_uverbs_rule_idr);
  77
  78 static DEFINE_SPINLOCK(map_lock);
  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_ALLOC_MW]            = ib_uverbs_alloc_mw,
  92         [IB_USER_VERBS_CMD_DEALLOC_MW]          = ib_uverbs_dealloc_mw,
  93         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  94         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  95         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  96         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  97         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  98         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  99         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
 100         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 101         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 102         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 103         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 104         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 105         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 106         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 107         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 108         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 109         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 110         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 111         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 112         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 113         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 114         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 115         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 116         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 117         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp,
 118 };
 119
 120 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
 121                                     struct ib_udata *ucore,
 122                                     struct ib_udata *uhw) = {
 123         [IB_USER_VERBS_EX_CMD_CREATE_FLOW]      = ib_uverbs_ex_create_flow,
 124         [IB_USER_VERBS_EX_CMD_DESTROY_FLOW]     = ib_uverbs_ex_destroy_flow
 125 };
 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         /* Remove MWs before QPs, in order to support type 2A MWs. */
 215         list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
 216                 struct ib_mw *mw = uobj->object;
 217
 218                 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
 219                 ib_dealloc_mw(mw);
 220                 kfree(uobj);
 221         }
 222
 223         list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
 224                 struct ib_flow *flow_id = uobj->object;
 225
 226                 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
 227                 ib_destroy_flow(flow_id);
 228                 kfree(uobj);
 229         }
 230
 231         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 232                 struct ib_qp *qp = uobj->object;
 233                 struct ib_uqp_object *uqp =
 234                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 235
 236                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 237                 if (qp != qp->real_qp) {
 238                         ib_close_qp(qp);
 239                 } else {
 240                         ib_uverbs_detach_umcast(qp, uqp);
 241                         ib_destroy_qp(qp);
 242                 }
 243                 ib_uverbs_release_uevent(file, &uqp->uevent);
 244                 kfree(uqp);
 245         }
 246
 247         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 248                 struct ib_cq *cq = uobj->object;
 249                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 250                 struct ib_ucq_object *ucq =
 251                         container_of(uobj, struct ib_ucq_object, uobject);
 252
 253                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 254                 ib_destroy_cq(cq);
 255                 ib_uverbs_release_ucq(file, ev_file, ucq);
 256                 kfree(ucq);
 257         }
 258
 259         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 260                 struct ib_srq *srq = uobj->object;
 261                 struct ib_uevent_object *uevent =
 262                         container_of(uobj, struct ib_uevent_object, uobject);
 263
 264                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 265                 ib_destroy_srq(srq);
 266                 ib_uverbs_release_uevent(file, uevent);
 267                 kfree(uevent);
 268         }
 269
 270         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 271                 struct ib_mr *mr = uobj->object;
 272
 273                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 274                 ib_dereg_mr(mr);
 275                 kfree(uobj);
 276         }
 277
 278         mutex_lock(&file->device->xrcd_tree_mutex);
 279         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 280                 struct ib_xrcd *xrcd = uobj->object;
 281                 struct ib_uxrcd_object *uxrcd =
 282                         container_of(uobj, struct ib_uxrcd_object, uobject);
 283
 284                 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 285                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 286                 kfree(uxrcd);
 287         }
 288         mutex_unlock(&file->device->xrcd_tree_mutex);
 289
 290         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 291                 struct ib_pd *pd = uobj->object;
 292
 293                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 294                 ib_dealloc_pd(pd);
 295                 kfree(uobj);
 296         }
 297
 298         return context->device->dealloc_ucontext(context);
 299 }
 300
 301 static void ib_uverbs_release_file(struct kref *ref)
 302 {
 303         struct ib_uverbs_file *file =
 304                 container_of(ref, struct ib_uverbs_file, ref);
 305
 306         module_put(file->device->ib_dev->owner);
 307         kref_put(&file->device->ref, ib_uverbs_release_dev);
 308
 309         kfree(file);
 310 }
 311
 312 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 313                                     size_t count, loff_t *pos)
 314 {
 315         struct ib_uverbs_event_file *file = filp->private_data;
 316         struct ib_uverbs_event *event;
 317         int eventsz;
 318         int ret = 0;
 319
 320         spin_lock_irq(&file->lock);
 321
 322         while (list_empty(&file->event_list)) {
 323                 spin_unlock_irq(&file->lock);
 324
 325                 if (filp->f_flags & O_NONBLOCK)
 326                         return -EAGAIN;
 327
 328                 if (wait_event_interruptible(file->poll_wait,
 329                                              !list_empty(&file->event_list)))
 330                         return -ERESTARTSYS;
 331
 332                 spin_lock_irq(&file->lock);
 333         }
 334
 335         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 336
 337         if (file->is_async)
 338                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 339         else
 340                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 341
 342         if (eventsz > count) {
 343                 ret   = -EINVAL;
 344                 event = NULL;
 345         } else {
 346                 list_del(file->event_list.next);
 347                 if (event->counter) {
 348                         ++(*event->counter);
 349                         list_del(&event->obj_list);
 350                 }
 351         }
 352
 353         spin_unlock_irq(&file->lock);
 354
 355         if (event) {
 356                 if (copy_to_user(buf, event, eventsz))
 357                         ret = -EFAULT;
 358                 else
 359                         ret = eventsz;
 360         }
 361
 362         kfree(event);
 363
 364         return ret;
 365 }
 366
 367 static unsigned int ib_uverbs_event_poll(struct file *filp,
 368                                          struct poll_table_struct *wait)
 369 {
 370         unsigned int pollflags = 0;
 371         struct ib_uverbs_event_file *file = filp->private_data;
 372
 373         poll_wait(filp, &file->poll_wait, wait);
 374
 375         spin_lock_irq(&file->lock);
 376         if (!list_empty(&file->event_list))
 377                 pollflags = POLLIN | POLLRDNORM;
 378         spin_unlock_irq(&file->lock);
 379
 380         return pollflags;
 381 }
 382
 383 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 384 {
 385         struct ib_uverbs_event_file *file = filp->private_data;
 386
 387         return fasync_helper(fd, filp, on, &file->async_queue);
 388 }
 389
 390 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 391 {
 392         struct ib_uverbs_event_file *file = filp->private_data;
 393         struct ib_uverbs_event *entry, *tmp;
 394
 395         spin_lock_irq(&file->lock);
 396         file->is_closed = 1;
 397         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 398                 if (entry->counter)
 399                         list_del(&entry->obj_list);
 400                 kfree(entry);
 401         }
 402         spin_unlock_irq(&file->lock);
 403
 404         if (file->is_async) {
 405                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 406                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 407         }
 408         kref_put(&file->ref, ib_uverbs_release_event_file);
 409
 410         return 0;
 411 }
 412
 413 static const struct file_operations uverbs_event_fops = {
 414         .owner   = THIS_MODULE,
 415         .read    = ib_uverbs_event_read,
 416         .poll    = ib_uverbs_event_poll,
 417         .release = ib_uverbs_event_close,
 418         .fasync  = ib_uverbs_event_fasync,
 419         .llseek  = no_llseek,
 420 };
 421
 422 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 423 {
 424         struct ib_uverbs_event_file    *file = cq_context;
 425         struct ib_ucq_object           *uobj;
 426         struct ib_uverbs_event         *entry;
 427         unsigned long                   flags;
 428
 429         if (!file)
 430                 return;
 431
 432         spin_lock_irqsave(&file->lock, flags);
 433         if (file->is_closed) {
 434                 spin_unlock_irqrestore(&file->lock, flags);
 435                 return;
 436         }
 437
 438         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 439         if (!entry) {
 440                 spin_unlock_irqrestore(&file->lock, flags);
 441                 return;
 442         }
 443
 444         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 445
 446         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 447         entry->counter             = &uobj->comp_events_reported;
 448
 449         list_add_tail(&entry->list, &file->event_list);
 450         list_add_tail(&entry->obj_list, &uobj->comp_list);
 451         spin_unlock_irqrestore(&file->lock, flags);
 452
 453         wake_up_interruptible(&file->poll_wait);
 454         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 455 }
 456
 457 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 458                                     __u64 element, __u64 event,
 459                                     struct list_head *obj_list,
 460                                     u32 *counter)
 461 {
 462         struct ib_uverbs_event *entry;
 463         unsigned long flags;
 464
 465         spin_lock_irqsave(&file->async_file->lock, flags);
 466         if (file->async_file->is_closed) {
 467                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 468                 return;
 469         }
 470
 471         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 472         if (!entry) {
 473                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 474                 return;
 475         }
 476
 477         entry->desc.async.element    = element;
 478         entry->desc.async.event_type = event;
 479         entry->counter               = counter;
 480
 481         list_add_tail(&entry->list, &file->async_file->event_list);
 482         if (obj_list)
 483                 list_add_tail(&entry->obj_list, obj_list);
 484         spin_unlock_irqrestore(&file->async_file->lock, flags);
 485
 486         wake_up_interruptible(&file->async_file->poll_wait);
 487         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 488 }
 489
 490 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 491 {
 492         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 493                                                   struct ib_ucq_object, uobject);
 494
 495         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 496                                 event->event, &uobj->async_list,
 497                                 &uobj->async_events_reported);
 498 }
 499
 500 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 501 {
 502         struct ib_uevent_object *uobj;
 503
 504         uobj = container_of(event->element.qp->uobject,
 505                             struct ib_uevent_object, uobject);
 506
 507         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 508                                 event->event, &uobj->event_list,
 509                                 &uobj->events_reported);
 510 }
 511
 512 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 513 {
 514         struct ib_uevent_object *uobj;
 515
 516         uobj = container_of(event->element.srq->uobject,
 517                             struct ib_uevent_object, uobject);
 518
 519         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 520                                 event->event, &uobj->event_list,
 521                                 &uobj->events_reported);
 522 }
 523
 524 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 525                              struct ib_event *event)
 526 {
 527         struct ib_uverbs_file *file =
 528                 container_of(handler, struct ib_uverbs_file, event_handler);
 529
 530         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 531                                 NULL, NULL);
 532 }
 533
 534 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 535                                         int is_async)
 536 {
 537         struct ib_uverbs_event_file *ev_file;
 538         struct file *filp;
 539
 540         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 541         if (!ev_file)
 542                 return ERR_PTR(-ENOMEM);
 543
 544         kref_init(&ev_file->ref);
 545         spin_lock_init(&ev_file->lock);
 546         INIT_LIST_HEAD(&ev_file->event_list);
 547         init_waitqueue_head(&ev_file->poll_wait);
 548         ev_file->uverbs_file = uverbs_file;
 549         ev_file->async_queue = NULL;
 550         ev_file->is_async    = is_async;
 551         ev_file->is_closed   = 0;
 552
 553         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 554                                   ev_file, O_RDONLY);
 555         if (IS_ERR(filp))
 556                 kfree(ev_file);
 557
 558         return filp;
 559 }
 560
 561 /*
 562  * Look up a completion event file by FD.  If lookup is successful,
 563  * takes a ref to the event file struct that it returns; if
 564  * unsuccessful, returns NULL.
 565  */
 566 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 567 {
 568         struct ib_uverbs_event_file *ev_file = NULL;
 569         struct fd f = fdget(fd);
 570
 571         if (!f.file)
 572                 return NULL;
 573
 574         if (f.file->f_op != &uverbs_event_fops)
 575                 goto out;
 576
 577         ev_file = f.file->private_data;
 578         if (ev_file->is_async) {
 579                 ev_file = NULL;
 580                 goto out;
 581         }
 582
 583         kref_get(&ev_file->ref);
 584
 585 out:
 586         fdput(f);
 587         return ev_file;
 588 }
 589
 590 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 591                              size_t count, loff_t *pos)
 592 {
 593         struct ib_uverbs_file *file = filp->private_data;
 594         struct ib_uverbs_cmd_hdr hdr;
 595         __u32 flags;
 596
 597         if (count < sizeof hdr)
 598                 return -EINVAL;
 599
 600         if (copy_from_user(&hdr, buf, sizeof hdr))
 601                 return -EFAULT;
 602
 603         flags = (hdr.command &
 604                  IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
 605
 606         if (!flags) {
 607                 __u32 command;
 608
 609                 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 610                                            IB_USER_VERBS_CMD_COMMAND_MASK))
 611                         return -EINVAL;
 612
 613                 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 614
 615                 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
 616                     !uverbs_cmd_table[command])
 617                         return -EINVAL;
 618
 619                 if (!file->ucontext &&
 620                     command != IB_USER_VERBS_CMD_GET_CONTEXT)
 621                         return -EINVAL;
 622
 623                 if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << command)))
 624                         return -ENOSYS;
 625
 626                 if (hdr.in_words * 4 != count)
 627                         return -EINVAL;
 628
 629                 return uverbs_cmd_table[command](file,
 630                                                  buf + sizeof(hdr),
 631                                                  hdr.in_words * 4,
 632                                                  hdr.out_words * 4);
 633
 634         } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
 635                 __u32 command;
 636
 637                 struct ib_uverbs_ex_cmd_hdr ex_hdr;
 638                 struct ib_udata ucore;
 639                 struct ib_udata uhw;
 640                 int err;
 641                 size_t written_count = count;
 642
 643                 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 644                                            IB_USER_VERBS_CMD_COMMAND_MASK))
 645                         return -EINVAL;
 646
 647                 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 648
 649                 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
 650                     !uverbs_ex_cmd_table[command])
 651                         return -ENOSYS;
 652
 653                 if (!file->ucontext)
 654                         return -EINVAL;
 655
 656                 if (!(file->device->ib_dev->uverbs_ex_cmd_mask & (1ull << command)))
 657                         return -ENOSYS;
 658
 659                 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 660                         return -EINVAL;
 661
 662                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
 663                         return -EFAULT;
 664
 665                 count -= sizeof(hdr) + sizeof(ex_hdr);
 666                 buf += sizeof(hdr) + sizeof(ex_hdr);
 667
 668                 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count)
 669                         return -EINVAL;
 670
 671                 if (ex_hdr.cmd_hdr_reserved)
 672                         return -EINVAL;
 673
 674                 if (ex_hdr.response) {
 675                         if (!hdr.out_words && !ex_hdr.provider_out_words)
 676                                 return -EINVAL;
 677
 678                         if (!access_ok(VERIFY_WRITE,
 679                                        (void __user *) (unsigned long) ex_hdr.response,
 680                                        (hdr.out_words + ex_hdr.provider_out_words) * 8))
 681                                 return -EFAULT;
 682                 } else {
 683                         if (hdr.out_words || ex_hdr.provider_out_words)
 684                                 return -EINVAL;
 685                 }
 686
 687                 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
 688                                        hdr.in_words * 8, hdr.out_words * 8);
 689
 690                 INIT_UDATA_BUF_OR_NULL(&uhw,
 691                                        buf + ucore.inlen,
 692                                        (unsigned long) ex_hdr.response + ucore.outlen,
 693                                        ex_hdr.provider_in_words * 8,
 694                                        ex_hdr.provider_out_words * 8);
 695
 696                 err = uverbs_ex_cmd_table[command](file,
 697                                                    &ucore,
 698                                                    &uhw);
 699
 700                 if (err)
 701                         return err;
 702
 703                 return written_count;
 704         }
 705
 706         return -ENOSYS;
 707 }
 708
 709 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 710 {
 711         struct ib_uverbs_file *file = filp->private_data;
 712
 713         if (!file->ucontext)
 714                 return -ENODEV;
 715         else
 716                 return file->device->ib_dev->mmap(file->ucontext, vma);
 717 }
 718
 719 /*
 720  * ib_uverbs_open() does not need the BKL:
 721  *
 722  *  - the ib_uverbs_device structures are properly reference counted and
 723  *    everything else is purely local to the file being created, so
 724  *    races against other open calls are not a problem;
 725  *  - there is no ioctl method to race against;
 726  *  - the open method will either immediately run -ENXIO, or all
 727  *    required initialization will be done.
 728  */
 729 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 730 {
 731         struct ib_uverbs_device *dev;
 732         struct ib_uverbs_file *file;
 733         int ret;
 734
 735         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 736         if (dev)
 737                 kref_get(&dev->ref);
 738         else
 739                 return -ENXIO;
 740
 741         if (!try_module_get(dev->ib_dev->owner)) {
 742                 ret = -ENODEV;
 743                 goto err;
 744         }
 745
 746         file = kmalloc(sizeof *file, GFP_KERNEL);
 747         if (!file) {
 748                 ret = -ENOMEM;
 749                 goto err_module;
 750         }
 751
 752         file->device     = dev;
 753         file->ucontext   = NULL;
 754         file->async_file = NULL;
 755         kref_init(&file->ref);
 756         mutex_init(&file->mutex);
 757
 758         filp->private_data = file;
 759
 760         return nonseekable_open(inode, filp);
 761
 762 err_module:
 763         module_put(dev->ib_dev->owner);
 764
 765 err:
 766         kref_put(&dev->ref, ib_uverbs_release_dev);
 767         return ret;
 768 }
 769
 770 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 771 {
 772         struct ib_uverbs_file *file = filp->private_data;
 773
 774         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 775
 776         if (file->async_file)
 777                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 778
 779         kref_put(&file->ref, ib_uverbs_release_file);
 780
 781         return 0;
 782 }
 783
 784 static const struct file_operations uverbs_fops = {
 785         .owner   = THIS_MODULE,
 786         .write   = ib_uverbs_write,
 787         .open    = ib_uverbs_open,
 788         .release = ib_uverbs_close,
 789         .llseek  = no_llseek,
 790 };
 791
 792 static const struct file_operations uverbs_mmap_fops = {
 793         .owner   = THIS_MODULE,
 794         .write   = ib_uverbs_write,
 795         .mmap    = ib_uverbs_mmap,
 796         .open    = ib_uverbs_open,
 797         .release = ib_uverbs_close,
 798         .llseek  = no_llseek,
 799 };
 800
 801 static struct ib_client uverbs_client = {
 802         .name   = "uverbs",
 803         .add    = ib_uverbs_add_one,
 804         .remove = ib_uverbs_remove_one
 805 };
 806
 807 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 808                           char *buf)
 809 {
 810         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 811
 812         if (!dev)
 813                 return -ENODEV;
 814
 815         return sprintf(buf, "%s\n", dev->ib_dev->name);
 816 }
 817 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 818
 819 static ssize_t show_dev_abi_version(struct device *device,
 820                                     struct device_attribute *attr, char *buf)
 821 {
 822         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 823
 824         if (!dev)
 825                 return -ENODEV;
 826
 827         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 828 }
 829 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 830
 831 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
 832                          __stringify(IB_USER_VERBS_ABI_VERSION));
 833
 834 static dev_t overflow_maj;
 835 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
 836
 837 /*
 838  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
 839  * requesting a new major number and doubling the number of max devices we
 840  * support. It's stupid, but simple.
 841  */
 842 static int find_overflow_devnum(void)
 843 {
 844         int ret;
 845
 846         if (!overflow_maj) {
 847                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
 848                                           "infiniband_verbs");
 849                 if (ret) {
 850                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
 851                         return ret;
 852                 }
 853         }
 854
 855         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
 856         if (ret >= IB_UVERBS_MAX_DEVICES)
 857                 return -1;
 858
 859         return ret;
 860 }
 861
 862 static void ib_uverbs_add_one(struct ib_device *device)
 863 {
 864         int devnum;
 865         dev_t base;
 866         struct ib_uverbs_device *uverbs_dev;
 867
 868         if (!device->alloc_ucontext)
 869                 return;
 870
 871         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 872         if (!uverbs_dev)
 873                 return;
 874
 875         kref_init(&uverbs_dev->ref);
 876         init_completion(&uverbs_dev->comp);
 877         uverbs_dev->xrcd_tree = RB_ROOT;
 878         mutex_init(&uverbs_dev->xrcd_tree_mutex);
 879
 880         spin_lock(&map_lock);
 881         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 882         if (devnum >= IB_UVERBS_MAX_DEVICES) {
 883                 spin_unlock(&map_lock);
 884                 devnum = find_overflow_devnum();
 885                 if (devnum < 0)
 886                         goto err;
 887
 888                 spin_lock(&map_lock);
 889                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
 890                 base = devnum + overflow_maj;
 891                 set_bit(devnum, overflow_map);
 892         } else {
 893                 uverbs_dev->devnum = devnum;
 894                 base = devnum + IB_UVERBS_BASE_DEV;
 895                 set_bit(devnum, dev_map);
 896         }
 897         spin_unlock(&map_lock);
 898
 899         uverbs_dev->ib_dev           = device;
 900         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 901
 902         cdev_init(&uverbs_dev->cdev, NULL);
 903         uverbs_dev->cdev.owner = THIS_MODULE;
 904         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 905         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
 906         if (cdev_add(&uverbs_dev->cdev, base, 1))
 907                 goto err_cdev;
 908
 909         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 910                                         uverbs_dev->cdev.dev, uverbs_dev,
 911                                         "uverbs%d", uverbs_dev->devnum);
 912         if (IS_ERR(uverbs_dev->dev))
 913                 goto err_cdev;
 914
 915         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 916                 goto err_class;
 917         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 918                 goto err_class;
 919
 920         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 921
 922         return;
 923
 924 err_class:
 925         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 926
 927 err_cdev:
 928         cdev_del(&uverbs_dev->cdev);
 929         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 930                 clear_bit(devnum, dev_map);
 931         else
 932                 clear_bit(devnum, overflow_map);
 933
 934 err:
 935         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 936         wait_for_completion(&uverbs_dev->comp);
 937         kfree(uverbs_dev);
 938         return;
 939 }
 940
 941 static void ib_uverbs_remove_one(struct ib_device *device)
 942 {
 943         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 944
 945         if (!uverbs_dev)
 946                 return;
 947
 948         dev_set_drvdata(uverbs_dev->dev, NULL);
 949         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 950         cdev_del(&uverbs_dev->cdev);
 951
 952         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 953                 clear_bit(uverbs_dev->devnum, dev_map);
 954         else
 955                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
 956
 957         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 958         wait_for_completion(&uverbs_dev->comp);
 959         kfree(uverbs_dev);
 960 }
 961
 962 static char *uverbs_devnode(struct device *dev, umode_t *mode)
 963 {
 964         if (mode)
 965                 *mode = 0666;
 966         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
 967 }
 968
 969 static int __init ib_uverbs_init(void)
 970 {
 971         int ret;
 972
 973         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 974                                      "infiniband_verbs");
 975         if (ret) {
 976                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 977                 goto out;
 978         }
 979
 980         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 981         if (IS_ERR(uverbs_class)) {
 982                 ret = PTR_ERR(uverbs_class);
 983                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 984                 goto out_chrdev;
 985         }
 986
 987         uverbs_class->devnode = uverbs_devnode;
 988
 989         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
 990         if (ret) {
 991                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 992                 goto out_class;
 993         }
 994
 995         ret = ib_register_client(&uverbs_client);
 996         if (ret) {
 997                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 998                 goto out_class;
 999         }
1000
1001         return 0;
1002
1003 out_class:
1004         class_destroy(uverbs_class);
1005
1006 out_chrdev:
1007         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1008
1009 out:
1010         return ret;
1011 }
1012
1013 static void __exit ib_uverbs_cleanup(void)
1014 {
1015         ib_unregister_client(&uverbs_client);
1016         class_destroy(uverbs_class);
1017         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1018         if (overflow_maj)
1019                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1020         idr_destroy(&ib_uverbs_pd_idr);
1021         idr_destroy(&ib_uverbs_mr_idr);
1022         idr_destroy(&ib_uverbs_mw_idr);
1023         idr_destroy(&ib_uverbs_ah_idr);
1024         idr_destroy(&ib_uverbs_cq_idr);
1025         idr_destroy(&ib_uverbs_qp_idr);
1026         idr_destroy(&ib_uverbs_srq_idr);
1027 }
1028
1029 module_init(ib_uverbs_init);
1030 module_exit(ib_uverbs_cleanup);