kernel/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_REREG_MR]            = ib_uverbs_rereg_mr,
  91         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  92         [IB_USER_VERBS_CMD_ALLOC_MW]            = ib_uverbs_alloc_mw,
  93         [IB_USER_VERBS_CMD_DEALLOC_MW]          = ib_uverbs_dealloc_mw,
  94         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  95         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  96         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  97         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  98         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  99         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
 100         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
 101         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 102         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 103         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 104         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 105         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 106         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 107         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 108         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 109         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 110         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 111         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 112         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 113         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 114         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 115         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 116         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 117         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 118         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp,
 119 };
 120
 121 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
 122                                     struct ib_udata *ucore,
 123                                     struct ib_udata *uhw) = {
 124         [IB_USER_VERBS_EX_CMD_CREATE_FLOW]      = ib_uverbs_ex_create_flow,
 125         [IB_USER_VERBS_EX_CMD_DESTROY_FLOW]     = ib_uverbs_ex_destroy_flow,
 126         [IB_USER_VERBS_EX_CMD_QUERY_DEVICE]     = ib_uverbs_ex_query_device,
 127 };
 128
 129 static void ib_uverbs_add_one(struct ib_device *device);
 130 static void ib_uverbs_remove_one(struct ib_device *device);
 131
 132 static void ib_uverbs_release_dev(struct kref *ref)
 133 {
 134         struct ib_uverbs_device *dev =
 135                 container_of(ref, struct ib_uverbs_device, ref);
 136
 137         complete(&dev->comp);
 138 }
 139
 140 static void ib_uverbs_release_event_file(struct kref *ref)
 141 {
 142         struct ib_uverbs_event_file *file =
 143                 container_of(ref, struct ib_uverbs_event_file, ref);
 144
 145         kfree(file);
 146 }
 147
 148 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 149                           struct ib_uverbs_event_file *ev_file,
 150                           struct ib_ucq_object *uobj)
 151 {
 152         struct ib_uverbs_event *evt, *tmp;
 153
 154         if (ev_file) {
 155                 spin_lock_irq(&ev_file->lock);
 156                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 157                         list_del(&evt->list);
 158                         kfree(evt);
 159                 }
 160                 spin_unlock_irq(&ev_file->lock);
 161
 162                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 163         }
 164
 165         spin_lock_irq(&file->async_file->lock);
 166         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 167                 list_del(&evt->list);
 168                 kfree(evt);
 169         }
 170         spin_unlock_irq(&file->async_file->lock);
 171 }
 172
 173 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 174                               struct ib_uevent_object *uobj)
 175 {
 176         struct ib_uverbs_event *evt, *tmp;
 177
 178         spin_lock_irq(&file->async_file->lock);
 179         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 180                 list_del(&evt->list);
 181                 kfree(evt);
 182         }
 183         spin_unlock_irq(&file->async_file->lock);
 184 }
 185
 186 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 187                                     struct ib_uqp_object *uobj)
 188 {
 189         struct ib_uverbs_mcast_entry *mcast, *tmp;
 190
 191         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 192                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 193                 list_del(&mcast->list);
 194                 kfree(mcast);
 195         }
 196 }
 197
 198 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 199                                       struct ib_ucontext *context)
 200 {
 201         struct ib_uobject *uobj, *tmp;
 202
 203         if (!context)
 204                 return 0;
 205
 206         context->closing = 1;
 207
 208         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 209                 struct ib_ah *ah = uobj->object;
 210
 211                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 212                 ib_destroy_ah(ah);
 213                 kfree(uobj);
 214         }
 215
 216         /* Remove MWs before QPs, in order to support type 2A MWs. */
 217         list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
 218                 struct ib_mw *mw = uobj->object;
 219
 220                 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
 221                 ib_dealloc_mw(mw);
 222                 kfree(uobj);
 223         }
 224
 225         list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
 226                 struct ib_flow *flow_id = uobj->object;
 227
 228                 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
 229                 ib_destroy_flow(flow_id);
 230                 kfree(uobj);
 231         }
 232
 233         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 234                 struct ib_qp *qp = uobj->object;
 235                 struct ib_uqp_object *uqp =
 236                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 237
 238                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 239                 if (qp != qp->real_qp) {
 240                         ib_close_qp(qp);
 241                 } else {
 242                         ib_uverbs_detach_umcast(qp, uqp);
 243                         ib_destroy_qp(qp);
 244                 }
 245                 ib_uverbs_release_uevent(file, &uqp->uevent);
 246                 kfree(uqp);
 247         }
 248
 249         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 250                 struct ib_srq *srq = uobj->object;
 251                 struct ib_uevent_object *uevent =
 252                         container_of(uobj, struct ib_uevent_object, uobject);
 253
 254                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 255                 ib_destroy_srq(srq);
 256                 ib_uverbs_release_uevent(file, uevent);
 257                 kfree(uevent);
 258         }
 259
 260         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 261                 struct ib_cq *cq = uobj->object;
 262                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 263                 struct ib_ucq_object *ucq =
 264                         container_of(uobj, struct ib_ucq_object, uobject);
 265
 266                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 267                 ib_destroy_cq(cq);
 268                 ib_uverbs_release_ucq(file, ev_file, ucq);
 269                 kfree(ucq);
 270         }
 271
 272         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 273                 struct ib_mr *mr = uobj->object;
 274
 275                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 276                 ib_dereg_mr(mr);
 277                 kfree(uobj);
 278         }
 279
 280         mutex_lock(&file->device->xrcd_tree_mutex);
 281         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 282                 struct ib_xrcd *xrcd = uobj->object;
 283                 struct ib_uxrcd_object *uxrcd =
 284                         container_of(uobj, struct ib_uxrcd_object, uobject);
 285
 286                 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 287                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 288                 kfree(uxrcd);
 289         }
 290         mutex_unlock(&file->device->xrcd_tree_mutex);
 291
 292         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 293                 struct ib_pd *pd = uobj->object;
 294
 295                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 296                 ib_dealloc_pd(pd);
 297                 kfree(uobj);
 298         }
 299
 300         put_pid(context->tgid);
 301
 302         return context->device->dealloc_ucontext(context);
 303 }
 304
 305 static void ib_uverbs_release_file(struct kref *ref)
 306 {
 307         struct ib_uverbs_file *file =
 308                 container_of(ref, struct ib_uverbs_file, ref);
 309
 310         module_put(file->device->ib_dev->owner);
 311         kref_put(&file->device->ref, ib_uverbs_release_dev);
 312
 313         kfree(file);
 314 }
 315
 316 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 317                                     size_t count, loff_t *pos)
 318 {
 319         struct ib_uverbs_event_file *file = filp->private_data;
 320         struct ib_uverbs_event *event;
 321         int eventsz;
 322         int ret = 0;
 323
 324         spin_lock_irq(&file->lock);
 325
 326         while (list_empty(&file->event_list)) {
 327                 spin_unlock_irq(&file->lock);
 328
 329                 if (filp->f_flags & O_NONBLOCK)
 330                         return -EAGAIN;
 331
 332                 if (wait_event_interruptible(file->poll_wait,
 333                                              !list_empty(&file->event_list)))
 334                         return -ERESTARTSYS;
 335
 336                 spin_lock_irq(&file->lock);
 337         }
 338
 339         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 340
 341         if (file->is_async)
 342                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 343         else
 344                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 345
 346         if (eventsz > count) {
 347                 ret   = -EINVAL;
 348                 event = NULL;
 349         } else {
 350                 list_del(file->event_list.next);
 351                 if (event->counter) {
 352                         ++(*event->counter);
 353                         list_del(&event->obj_list);
 354                 }
 355         }
 356
 357         spin_unlock_irq(&file->lock);
 358
 359         if (event) {
 360                 if (copy_to_user(buf, event, eventsz))
 361                         ret = -EFAULT;
 362                 else
 363                         ret = eventsz;
 364         }
 365
 366         kfree(event);
 367
 368         return ret;
 369 }
 370
 371 static unsigned int ib_uverbs_event_poll(struct file *filp,
 372                                          struct poll_table_struct *wait)
 373 {
 374         unsigned int pollflags = 0;
 375         struct ib_uverbs_event_file *file = filp->private_data;
 376
 377         poll_wait(filp, &file->poll_wait, wait);
 378
 379         spin_lock_irq(&file->lock);
 380         if (!list_empty(&file->event_list))
 381                 pollflags = POLLIN | POLLRDNORM;
 382         spin_unlock_irq(&file->lock);
 383
 384         return pollflags;
 385 }
 386
 387 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 388 {
 389         struct ib_uverbs_event_file *file = filp->private_data;
 390
 391         return fasync_helper(fd, filp, on, &file->async_queue);
 392 }
 393
 394 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 395 {
 396         struct ib_uverbs_event_file *file = filp->private_data;
 397         struct ib_uverbs_event *entry, *tmp;
 398
 399         spin_lock_irq(&file->lock);
 400         file->is_closed = 1;
 401         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 402                 if (entry->counter)
 403                         list_del(&entry->obj_list);
 404                 kfree(entry);
 405         }
 406         spin_unlock_irq(&file->lock);
 407
 408         if (file->is_async) {
 409                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 410                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 411         }
 412         kref_put(&file->ref, ib_uverbs_release_event_file);
 413
 414         return 0;
 415 }
 416
 417 static const struct file_operations uverbs_event_fops = {
 418         .owner   = THIS_MODULE,
 419         .read    = ib_uverbs_event_read,
 420         .poll    = ib_uverbs_event_poll,
 421         .release = ib_uverbs_event_close,
 422         .fasync  = ib_uverbs_event_fasync,
 423         .llseek  = no_llseek,
 424 };
 425
 426 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 427 {
 428         struct ib_uverbs_event_file    *file = cq_context;
 429         struct ib_ucq_object           *uobj;
 430         struct ib_uverbs_event         *entry;
 431         unsigned long                   flags;
 432
 433         if (!file)
 434                 return;
 435
 436         spin_lock_irqsave(&file->lock, flags);
 437         if (file->is_closed) {
 438                 spin_unlock_irqrestore(&file->lock, flags);
 439                 return;
 440         }
 441
 442         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 443         if (!entry) {
 444                 spin_unlock_irqrestore(&file->lock, flags);
 445                 return;
 446         }
 447
 448         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 449
 450         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 451         entry->counter             = &uobj->comp_events_reported;
 452
 453         list_add_tail(&entry->list, &file->event_list);
 454         list_add_tail(&entry->obj_list, &uobj->comp_list);
 455         spin_unlock_irqrestore(&file->lock, flags);
 456
 457         wake_up_interruptible(&file->poll_wait);
 458         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 459 }
 460
 461 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 462                                     __u64 element, __u64 event,
 463                                     struct list_head *obj_list,
 464                                     u32 *counter)
 465 {
 466         struct ib_uverbs_event *entry;
 467         unsigned long flags;
 468
 469         spin_lock_irqsave(&file->async_file->lock, flags);
 470         if (file->async_file->is_closed) {
 471                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 472                 return;
 473         }
 474
 475         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 476         if (!entry) {
 477                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 478                 return;
 479         }
 480
 481         entry->desc.async.element    = element;
 482         entry->desc.async.event_type = event;
 483         entry->desc.async.reserved   = 0;
 484         entry->counter               = counter;
 485
 486         list_add_tail(&entry->list, &file->async_file->event_list);
 487         if (obj_list)
 488                 list_add_tail(&entry->obj_list, obj_list);
 489         spin_unlock_irqrestore(&file->async_file->lock, flags);
 490
 491         wake_up_interruptible(&file->async_file->poll_wait);
 492         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 493 }
 494
 495 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 496 {
 497         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 498                                                   struct ib_ucq_object, uobject);
 499
 500         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 501                                 event->event, &uobj->async_list,
 502                                 &uobj->async_events_reported);
 503 }
 504
 505 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 506 {
 507         struct ib_uevent_object *uobj;
 508
 509         /* for XRC target qp's, check that qp is live */
 510         if (!event->element.qp->uobject || !event->element.qp->uobject->live)
 511                 return;
 512
 513         uobj = container_of(event->element.qp->uobject,
 514                             struct ib_uevent_object, uobject);
 515
 516         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 517                                 event->event, &uobj->event_list,
 518                                 &uobj->events_reported);
 519 }
 520
 521 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 522 {
 523         struct ib_uevent_object *uobj;
 524
 525         uobj = container_of(event->element.srq->uobject,
 526                             struct ib_uevent_object, uobject);
 527
 528         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 529                                 event->event, &uobj->event_list,
 530                                 &uobj->events_reported);
 531 }
 532
 533 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 534                              struct ib_event *event)
 535 {
 536         struct ib_uverbs_file *file =
 537                 container_of(handler, struct ib_uverbs_file, event_handler);
 538
 539         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 540                                 NULL, NULL);
 541 }
 542
 543 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 544                                         int is_async)
 545 {
 546         struct ib_uverbs_event_file *ev_file;
 547         struct file *filp;
 548
 549         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 550         if (!ev_file)
 551                 return ERR_PTR(-ENOMEM);
 552
 553         kref_init(&ev_file->ref);
 554         spin_lock_init(&ev_file->lock);
 555         INIT_LIST_HEAD(&ev_file->event_list);
 556         init_waitqueue_head(&ev_file->poll_wait);
 557         ev_file->uverbs_file = uverbs_file;
 558         ev_file->async_queue = NULL;
 559         ev_file->is_async    = is_async;
 560         ev_file->is_closed   = 0;
 561
 562         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 563                                   ev_file, O_RDONLY);
 564         if (IS_ERR(filp))
 565                 kfree(ev_file);
 566
 567         return filp;
 568 }
 569
 570 /*
 571  * Look up a completion event file by FD.  If lookup is successful,
 572  * takes a ref to the event file struct that it returns; if
 573  * unsuccessful, returns NULL.
 574  */
 575 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 576 {
 577         struct ib_uverbs_event_file *ev_file = NULL;
 578         struct fd f = fdget(fd);
 579
 580         if (!f.file)
 581                 return NULL;
 582
 583         if (f.file->f_op != &uverbs_event_fops)
 584                 goto out;
 585
 586         ev_file = f.file->private_data;
 587         if (ev_file->is_async) {
 588                 ev_file = NULL;
 589                 goto out;
 590         }
 591
 592         kref_get(&ev_file->ref);
 593
 594 out:
 595         fdput(f);
 596         return ev_file;
 597 }
 598
 599 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 600                              size_t count, loff_t *pos)
 601 {
 602         struct ib_uverbs_file *file = filp->private_data;
 603         struct ib_uverbs_cmd_hdr hdr;
 604         __u32 flags;
 605
 606         if (count < sizeof hdr)
 607                 return -EINVAL;
 608
 609         if (copy_from_user(&hdr, buf, sizeof hdr))
 610                 return -EFAULT;
 611
 612         flags = (hdr.command &
 613                  IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
 614
 615         if (!flags) {
 616                 __u32 command;
 617
 618                 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 619                                            IB_USER_VERBS_CMD_COMMAND_MASK))
 620                         return -EINVAL;
 621
 622                 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 623
 624                 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
 625                     !uverbs_cmd_table[command])
 626                         return -EINVAL;
 627
 628                 if (!file->ucontext &&
 629                     command != IB_USER_VERBS_CMD_GET_CONTEXT)
 630                         return -EINVAL;
 631
 632                 if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << command)))
 633                         return -ENOSYS;
 634
 635                 if (hdr.in_words * 4 != count)
 636                         return -EINVAL;
 637
 638                 return uverbs_cmd_table[command](file,
 639                                                  buf + sizeof(hdr),
 640                                                  hdr.in_words * 4,
 641                                                  hdr.out_words * 4);
 642
 643         } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
 644                 __u32 command;
 645
 646                 struct ib_uverbs_ex_cmd_hdr ex_hdr;
 647                 struct ib_udata ucore;
 648                 struct ib_udata uhw;
 649                 int err;
 650                 size_t written_count = count;
 651
 652                 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 653                                            IB_USER_VERBS_CMD_COMMAND_MASK))
 654                         return -EINVAL;
 655
 656                 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 657
 658                 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
 659                     !uverbs_ex_cmd_table[command])
 660                         return -ENOSYS;
 661
 662                 if (!file->ucontext)
 663                         return -EINVAL;
 664
 665                 if (!(file->device->ib_dev->uverbs_ex_cmd_mask & (1ull << command)))
 666                         return -ENOSYS;
 667
 668                 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 669                         return -EINVAL;
 670
 671                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
 672                         return -EFAULT;
 673
 674                 count -= sizeof(hdr) + sizeof(ex_hdr);
 675                 buf += sizeof(hdr) + sizeof(ex_hdr);
 676
 677                 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count)
 678                         return -EINVAL;
 679
 680                 if (ex_hdr.cmd_hdr_reserved)
 681                         return -EINVAL;
 682
 683                 if (ex_hdr.response) {
 684                         if (!hdr.out_words && !ex_hdr.provider_out_words)
 685                                 return -EINVAL;
 686
 687                         if (!access_ok(VERIFY_WRITE,
 688                                        (void __user *) (unsigned long) ex_hdr.response,
 689                                        (hdr.out_words + ex_hdr.provider_out_words) * 8))
 690                                 return -EFAULT;
 691                 } else {
 692                         if (hdr.out_words || ex_hdr.provider_out_words)
 693                                 return -EINVAL;
 694                 }
 695
 696                 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
 697                                        hdr.in_words * 8, hdr.out_words * 8);
 698
 699                 INIT_UDATA_BUF_OR_NULL(&uhw,
 700                                        buf + ucore.inlen,
 701                                        (unsigned long) ex_hdr.response + ucore.outlen,
 702                                        ex_hdr.provider_in_words * 8,
 703                                        ex_hdr.provider_out_words * 8);
 704
 705                 err = uverbs_ex_cmd_table[command](file,
 706                                                    &ucore,
 707                                                    &uhw);
 708
 709                 if (err)
 710                         return err;
 711
 712                 return written_count;
 713         }
 714
 715         return -ENOSYS;
 716 }
 717
 718 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 719 {
 720         struct ib_uverbs_file *file = filp->private_data;
 721
 722         if (!file->ucontext)
 723                 return -ENODEV;
 724         else
 725                 return file->device->ib_dev->mmap(file->ucontext, vma);
 726 }
 727
 728 /*
 729  * ib_uverbs_open() does not need the BKL:
 730  *
 731  *  - the ib_uverbs_device structures are properly reference counted and
 732  *    everything else is purely local to the file being created, so
 733  *    races against other open calls are not a problem;
 734  *  - there is no ioctl method to race against;
 735  *  - the open method will either immediately run -ENXIO, or all
 736  *    required initialization will be done.
 737  */
 738 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 739 {
 740         struct ib_uverbs_device *dev;
 741         struct ib_uverbs_file *file;
 742         int ret;
 743
 744         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 745         if (dev)
 746                 kref_get(&dev->ref);
 747         else
 748                 return -ENXIO;
 749
 750         if (!try_module_get(dev->ib_dev->owner)) {
 751                 ret = -ENODEV;
 752                 goto err;
 753         }
 754
 755         file = kmalloc(sizeof *file, GFP_KERNEL);
 756         if (!file) {
 757                 ret = -ENOMEM;
 758                 goto err_module;
 759         }
 760
 761         file->device     = dev;
 762         file->ucontext   = NULL;
 763         file->async_file = NULL;
 764         kref_init(&file->ref);
 765         mutex_init(&file->mutex);
 766
 767         filp->private_data = file;
 768
 769         return nonseekable_open(inode, filp);
 770
 771 err_module:
 772         module_put(dev->ib_dev->owner);
 773
 774 err:
 775         kref_put(&dev->ref, ib_uverbs_release_dev);
 776         return ret;
 777 }
 778
 779 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 780 {
 781         struct ib_uverbs_file *file = filp->private_data;
 782
 783         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 784
 785         if (file->async_file)
 786                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 787
 788         kref_put(&file->ref, ib_uverbs_release_file);
 789
 790         return 0;
 791 }
 792
 793 static const struct file_operations uverbs_fops = {
 794         .owner   = THIS_MODULE,
 795         .write   = ib_uverbs_write,
 796         .open    = ib_uverbs_open,
 797         .release = ib_uverbs_close,
 798         .llseek  = no_llseek,
 799 };
 800
 801 static const struct file_operations uverbs_mmap_fops = {
 802         .owner   = THIS_MODULE,
 803         .write   = ib_uverbs_write,
 804         .mmap    = ib_uverbs_mmap,
 805         .open    = ib_uverbs_open,
 806         .release = ib_uverbs_close,
 807         .llseek  = no_llseek,
 808 };
 809
 810 static struct ib_client uverbs_client = {
 811         .name   = "uverbs",
 812         .add    = ib_uverbs_add_one,
 813         .remove = ib_uverbs_remove_one
 814 };
 815
 816 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 817                           char *buf)
 818 {
 819         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 820
 821         if (!dev)
 822                 return -ENODEV;
 823
 824         return sprintf(buf, "%s\n", dev->ib_dev->name);
 825 }
 826 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 827
 828 static ssize_t show_dev_abi_version(struct device *device,
 829                                     struct device_attribute *attr, char *buf)
 830 {
 831         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 832
 833         if (!dev)
 834                 return -ENODEV;
 835
 836         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 837 }
 838 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 839
 840 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
 841                          __stringify(IB_USER_VERBS_ABI_VERSION));
 842
 843 static dev_t overflow_maj;
 844 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
 845
 846 /*
 847  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
 848  * requesting a new major number and doubling the number of max devices we
 849  * support. It's stupid, but simple.
 850  */
 851 static int find_overflow_devnum(void)
 852 {
 853         int ret;
 854
 855         if (!overflow_maj) {
 856                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
 857                                           "infiniband_verbs");
 858                 if (ret) {
 859                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
 860                         return ret;
 861                 }
 862         }
 863
 864         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
 865         if (ret >= IB_UVERBS_MAX_DEVICES)
 866                 return -1;
 867
 868         return ret;
 869 }
 870
 871 static void ib_uverbs_add_one(struct ib_device *device)
 872 {
 873         int devnum;
 874         dev_t base;
 875         struct ib_uverbs_device *uverbs_dev;
 876
 877         if (!device->alloc_ucontext)
 878                 return;
 879
 880         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 881         if (!uverbs_dev)
 882                 return;
 883
 884         kref_init(&uverbs_dev->ref);
 885         init_completion(&uverbs_dev->comp);
 886         uverbs_dev->xrcd_tree = RB_ROOT;
 887         mutex_init(&uverbs_dev->xrcd_tree_mutex);
 888
 889         spin_lock(&map_lock);
 890         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 891         if (devnum >= IB_UVERBS_MAX_DEVICES) {
 892                 spin_unlock(&map_lock);
 893                 devnum = find_overflow_devnum();
 894                 if (devnum < 0)
 895                         goto err;
 896
 897                 spin_lock(&map_lock);
 898                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
 899                 base = devnum + overflow_maj;
 900                 set_bit(devnum, overflow_map);
 901         } else {
 902                 uverbs_dev->devnum = devnum;
 903                 base = devnum + IB_UVERBS_BASE_DEV;
 904                 set_bit(devnum, dev_map);
 905         }
 906         spin_unlock(&map_lock);
 907
 908         uverbs_dev->ib_dev           = device;
 909         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 910
 911         cdev_init(&uverbs_dev->cdev, NULL);
 912         uverbs_dev->cdev.owner = THIS_MODULE;
 913         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 914         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
 915         if (cdev_add(&uverbs_dev->cdev, base, 1))
 916                 goto err_cdev;
 917
 918         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 919                                         uverbs_dev->cdev.dev, uverbs_dev,
 920                                         "uverbs%d", uverbs_dev->devnum);
 921         if (IS_ERR(uverbs_dev->dev))
 922                 goto err_cdev;
 923
 924         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 925                 goto err_class;
 926         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 927                 goto err_class;
 928
 929         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 930
 931         return;
 932
 933 err_class:
 934         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 935
 936 err_cdev:
 937         cdev_del(&uverbs_dev->cdev);
 938         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 939                 clear_bit(devnum, dev_map);
 940         else
 941                 clear_bit(devnum, overflow_map);
 942
 943 err:
 944         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 945         wait_for_completion(&uverbs_dev->comp);
 946         kfree(uverbs_dev);
 947         return;
 948 }
 949
 950 static void ib_uverbs_remove_one(struct ib_device *device)
 951 {
 952         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 953
 954         if (!uverbs_dev)
 955                 return;
 956
 957         dev_set_drvdata(uverbs_dev->dev, NULL);
 958         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 959         cdev_del(&uverbs_dev->cdev);
 960
 961         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 962                 clear_bit(uverbs_dev->devnum, dev_map);
 963         else
 964                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
 965
 966         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 967         wait_for_completion(&uverbs_dev->comp);
 968         kfree(uverbs_dev);
 969 }
 970
 971 static char *uverbs_devnode(struct device *dev, umode_t *mode)
 972 {
 973         if (mode)
 974                 *mode = 0666;
 975         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
 976 }
 977
 978 static int __init ib_uverbs_init(void)
 979 {
 980         int ret;
 981
 982         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 983                                      "infiniband_verbs");
 984         if (ret) {
 985                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 986                 goto out;
 987         }
 988
 989         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 990         if (IS_ERR(uverbs_class)) {
 991                 ret = PTR_ERR(uverbs_class);
 992                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 993                 goto out_chrdev;
 994         }
 995
 996         uverbs_class->devnode = uverbs_devnode;
 997
 998         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
 999         if (ret) {
1000                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
1001                 goto out_class;
1002         }
1003
1004         ret = ib_register_client(&uverbs_client);
1005         if (ret) {
1006                 printk(KERN_ERR "user_verbs: couldn't register client\n");
1007                 goto out_class;
1008         }
1009
1010         return 0;
1011
1012 out_class:
1013         class_destroy(uverbs_class);
1014
1015 out_chrdev:
1016         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1017
1018 out:
1019         return ret;
1020 }
1021
1022 static void __exit ib_uverbs_cleanup(void)
1023 {
1024         ib_unregister_client(&uverbs_client);
1025         class_destroy(uverbs_class);
1026         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1027         if (overflow_maj)
1028                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1029         idr_destroy(&ib_uverbs_pd_idr);
1030         idr_destroy(&ib_uverbs_mr_idr);
1031         idr_destroy(&ib_uverbs_mw_idr);
1032         idr_destroy(&ib_uverbs_ah_idr);
1033         idr_destroy(&ib_uverbs_cq_idr);
1034         idr_destroy(&ib_uverbs_qp_idr);
1035         idr_destroy(&ib_uverbs_srq_idr);
1036 }
1037
1038 module_init(ib_uverbs_init);
1039 module_exit(ib_uverbs_cleanup);