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         [IB_USER_VERBS_EX_CMD_CREATE_CQ]        = ib_uverbs_ex_create_cq,
 128 };
 129
 130 static void ib_uverbs_add_one(struct ib_device *device);
 131 static void ib_uverbs_remove_one(struct ib_device *device);
 132
 133 static void ib_uverbs_release_dev(struct kref *ref)
 134 {
 135         struct ib_uverbs_device *dev =
 136                 container_of(ref, struct ib_uverbs_device, ref);
 137
 138         complete(&dev->comp);
 139 }
 140
 141 static void ib_uverbs_release_event_file(struct kref *ref)
 142 {
 143         struct ib_uverbs_event_file *file =
 144                 container_of(ref, struct ib_uverbs_event_file, ref);
 145
 146         kfree(file);
 147 }
 148
 149 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 150                           struct ib_uverbs_event_file *ev_file,
 151                           struct ib_ucq_object *uobj)
 152 {
 153         struct ib_uverbs_event *evt, *tmp;
 154
 155         if (ev_file) {
 156                 spin_lock_irq(&ev_file->lock);
 157                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 158                         list_del(&evt->list);
 159                         kfree(evt);
 160                 }
 161                 spin_unlock_irq(&ev_file->lock);
 162
 163                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 164         }
 165
 166         spin_lock_irq(&file->async_file->lock);
 167         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 168                 list_del(&evt->list);
 169                 kfree(evt);
 170         }
 171         spin_unlock_irq(&file->async_file->lock);
 172 }
 173
 174 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 175                               struct ib_uevent_object *uobj)
 176 {
 177         struct ib_uverbs_event *evt, *tmp;
 178
 179         spin_lock_irq(&file->async_file->lock);
 180         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 181                 list_del(&evt->list);
 182                 kfree(evt);
 183         }
 184         spin_unlock_irq(&file->async_file->lock);
 185 }
 186
 187 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 188                                     struct ib_uqp_object *uobj)
 189 {
 190         struct ib_uverbs_mcast_entry *mcast, *tmp;
 191
 192         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 193                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 194                 list_del(&mcast->list);
 195                 kfree(mcast);
 196         }
 197 }
 198
 199 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 200                                       struct ib_ucontext *context)
 201 {
 202         struct ib_uobject *uobj, *tmp;
 203
 204         if (!context)
 205                 return 0;
 206
 207         context->closing = 1;
 208
 209         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 210                 struct ib_ah *ah = uobj->object;
 211
 212                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 213                 ib_destroy_ah(ah);
 214                 kfree(uobj);
 215         }
 216
 217         /* Remove MWs before QPs, in order to support type 2A MWs. */
 218         list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
 219                 struct ib_mw *mw = uobj->object;
 220
 221                 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
 222                 ib_dealloc_mw(mw);
 223                 kfree(uobj);
 224         }
 225
 226         list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
 227                 struct ib_flow *flow_id = uobj->object;
 228
 229                 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
 230                 ib_destroy_flow(flow_id);
 231                 kfree(uobj);
 232         }
 233
 234         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 235                 struct ib_qp *qp = uobj->object;
 236                 struct ib_uqp_object *uqp =
 237                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 238
 239                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 240                 if (qp != qp->real_qp) {
 241                         ib_close_qp(qp);
 242                 } else {
 243                         ib_uverbs_detach_umcast(qp, uqp);
 244                         ib_destroy_qp(qp);
 245                 }
 246                 ib_uverbs_release_uevent(file, &uqp->uevent);
 247                 kfree(uqp);
 248         }
 249
 250         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 251                 struct ib_srq *srq = uobj->object;
 252                 struct ib_uevent_object *uevent =
 253                         container_of(uobj, struct ib_uevent_object, uobject);
 254
 255                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 256                 ib_destroy_srq(srq);
 257                 ib_uverbs_release_uevent(file, uevent);
 258                 kfree(uevent);
 259         }
 260
 261         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 262                 struct ib_cq *cq = uobj->object;
 263                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 264                 struct ib_ucq_object *ucq =
 265                         container_of(uobj, struct ib_ucq_object, uobject);
 266
 267                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 268                 ib_destroy_cq(cq);
 269                 ib_uverbs_release_ucq(file, ev_file, ucq);
 270                 kfree(ucq);
 271         }
 272
 273         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 274                 struct ib_mr *mr = uobj->object;
 275
 276                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 277                 ib_dereg_mr(mr);
 278                 kfree(uobj);
 279         }
 280
 281         mutex_lock(&file->device->xrcd_tree_mutex);
 282         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 283                 struct ib_xrcd *xrcd = uobj->object;
 284                 struct ib_uxrcd_object *uxrcd =
 285                         container_of(uobj, struct ib_uxrcd_object, uobject);
 286
 287                 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 288                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 289                 kfree(uxrcd);
 290         }
 291         mutex_unlock(&file->device->xrcd_tree_mutex);
 292
 293         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 294                 struct ib_pd *pd = uobj->object;
 295
 296                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 297                 ib_dealloc_pd(pd);
 298                 kfree(uobj);
 299         }
 300
 301         put_pid(context->tgid);
 302
 303         return context->device->dealloc_ucontext(context);
 304 }
 305
 306 static void ib_uverbs_release_file(struct kref *ref)
 307 {
 308         struct ib_uverbs_file *file =
 309                 container_of(ref, struct ib_uverbs_file, ref);
 310
 311         module_put(file->device->ib_dev->owner);
 312         kref_put(&file->device->ref, ib_uverbs_release_dev);
 313
 314         kfree(file);
 315 }
 316
 317 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 318                                     size_t count, loff_t *pos)
 319 {
 320         struct ib_uverbs_event_file *file = filp->private_data;
 321         struct ib_uverbs_event *event;
 322         int eventsz;
 323         int ret = 0;
 324
 325         spin_lock_irq(&file->lock);
 326
 327         while (list_empty(&file->event_list)) {
 328                 spin_unlock_irq(&file->lock);
 329
 330                 if (filp->f_flags & O_NONBLOCK)
 331                         return -EAGAIN;
 332
 333                 if (wait_event_interruptible(file->poll_wait,
 334                                              !list_empty(&file->event_list)))
 335                         return -ERESTARTSYS;
 336
 337                 spin_lock_irq(&file->lock);
 338         }
 339
 340         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 341
 342         if (file->is_async)
 343                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 344         else
 345                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 346
 347         if (eventsz > count) {
 348                 ret   = -EINVAL;
 349                 event = NULL;
 350         } else {
 351                 list_del(file->event_list.next);
 352                 if (event->counter) {
 353                         ++(*event->counter);
 354                         list_del(&event->obj_list);
 355                 }
 356         }
 357
 358         spin_unlock_irq(&file->lock);
 359
 360         if (event) {
 361                 if (copy_to_user(buf, event, eventsz))
 362                         ret = -EFAULT;
 363                 else
 364                         ret = eventsz;
 365         }
 366
 367         kfree(event);
 368
 369         return ret;
 370 }
 371
 372 static unsigned int ib_uverbs_event_poll(struct file *filp,
 373                                          struct poll_table_struct *wait)
 374 {
 375         unsigned int pollflags = 0;
 376         struct ib_uverbs_event_file *file = filp->private_data;
 377
 378         poll_wait(filp, &file->poll_wait, wait);
 379
 380         spin_lock_irq(&file->lock);
 381         if (!list_empty(&file->event_list))
 382                 pollflags = POLLIN | POLLRDNORM;
 383         spin_unlock_irq(&file->lock);
 384
 385         return pollflags;
 386 }
 387
 388 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 389 {
 390         struct ib_uverbs_event_file *file = filp->private_data;
 391
 392         return fasync_helper(fd, filp, on, &file->async_queue);
 393 }
 394
 395 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 396 {
 397         struct ib_uverbs_event_file *file = filp->private_data;
 398         struct ib_uverbs_event *entry, *tmp;
 399
 400         spin_lock_irq(&file->lock);
 401         file->is_closed = 1;
 402         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 403                 if (entry->counter)
 404                         list_del(&entry->obj_list);
 405                 kfree(entry);
 406         }
 407         spin_unlock_irq(&file->lock);
 408
 409         if (file->is_async) {
 410                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 411                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 412         }
 413         kref_put(&file->ref, ib_uverbs_release_event_file);
 414
 415         return 0;
 416 }
 417
 418 static const struct file_operations uverbs_event_fops = {
 419         .owner   = THIS_MODULE,
 420         .read    = ib_uverbs_event_read,
 421         .poll    = ib_uverbs_event_poll,
 422         .release = ib_uverbs_event_close,
 423         .fasync  = ib_uverbs_event_fasync,
 424         .llseek  = no_llseek,
 425 };
 426
 427 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 428 {
 429         struct ib_uverbs_event_file    *file = cq_context;
 430         struct ib_ucq_object           *uobj;
 431         struct ib_uverbs_event         *entry;
 432         unsigned long                   flags;
 433
 434         if (!file)
 435                 return;
 436
 437         spin_lock_irqsave(&file->lock, flags);
 438         if (file->is_closed) {
 439                 spin_unlock_irqrestore(&file->lock, flags);
 440                 return;
 441         }
 442
 443         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 444         if (!entry) {
 445                 spin_unlock_irqrestore(&file->lock, flags);
 446                 return;
 447         }
 448
 449         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 450
 451         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 452         entry->counter             = &uobj->comp_events_reported;
 453
 454         list_add_tail(&entry->list, &file->event_list);
 455         list_add_tail(&entry->obj_list, &uobj->comp_list);
 456         spin_unlock_irqrestore(&file->lock, flags);
 457
 458         wake_up_interruptible(&file->poll_wait);
 459         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 460 }
 461
 462 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 463                                     __u64 element, __u64 event,
 464                                     struct list_head *obj_list,
 465                                     u32 *counter)
 466 {
 467         struct ib_uverbs_event *entry;
 468         unsigned long flags;
 469
 470         spin_lock_irqsave(&file->async_file->lock, flags);
 471         if (file->async_file->is_closed) {
 472                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 473                 return;
 474         }
 475
 476         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 477         if (!entry) {
 478                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 479                 return;
 480         }
 481
 482         entry->desc.async.element    = element;
 483         entry->desc.async.event_type = event;
 484         entry->desc.async.reserved   = 0;
 485         entry->counter               = counter;
 486
 487         list_add_tail(&entry->list, &file->async_file->event_list);
 488         if (obj_list)
 489                 list_add_tail(&entry->obj_list, obj_list);
 490         spin_unlock_irqrestore(&file->async_file->lock, flags);
 491
 492         wake_up_interruptible(&file->async_file->poll_wait);
 493         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 494 }
 495
 496 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 497 {
 498         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 499                                                   struct ib_ucq_object, uobject);
 500
 501         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 502                                 event->event, &uobj->async_list,
 503                                 &uobj->async_events_reported);
 504 }
 505
 506 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 507 {
 508         struct ib_uevent_object *uobj;
 509
 510         /* for XRC target qp's, check that qp is live */
 511         if (!event->element.qp->uobject || !event->element.qp->uobject->live)
 512                 return;
 513
 514         uobj = container_of(event->element.qp->uobject,
 515                             struct ib_uevent_object, uobject);
 516
 517         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 518                                 event->event, &uobj->event_list,
 519                                 &uobj->events_reported);
 520 }
 521
 522 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 523 {
 524         struct ib_uevent_object *uobj;
 525
 526         uobj = container_of(event->element.srq->uobject,
 527                             struct ib_uevent_object, uobject);
 528
 529         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 530                                 event->event, &uobj->event_list,
 531                                 &uobj->events_reported);
 532 }
 533
 534 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 535                              struct ib_event *event)
 536 {
 537         struct ib_uverbs_file *file =
 538                 container_of(handler, struct ib_uverbs_file, event_handler);
 539
 540         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 541                                 NULL, NULL);
 542 }
 543
 544 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 545                                         int is_async)
 546 {
 547         struct ib_uverbs_event_file *ev_file;
 548         struct file *filp;
 549
 550         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 551         if (!ev_file)
 552                 return ERR_PTR(-ENOMEM);
 553
 554         kref_init(&ev_file->ref);
 555         spin_lock_init(&ev_file->lock);
 556         INIT_LIST_HEAD(&ev_file->event_list);
 557         init_waitqueue_head(&ev_file->poll_wait);
 558         ev_file->uverbs_file = uverbs_file;
 559         ev_file->async_queue = NULL;
 560         ev_file->is_async    = is_async;
 561         ev_file->is_closed   = 0;
 562
 563         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 564                                   ev_file, O_RDONLY);
 565         if (IS_ERR(filp))
 566                 kfree(ev_file);
 567
 568         return filp;
 569 }
 570
 571 /*
 572  * Look up a completion event file by FD.  If lookup is successful,
 573  * takes a ref to the event file struct that it returns; if
 574  * unsuccessful, returns NULL.
 575  */
 576 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 577 {
 578         struct ib_uverbs_event_file *ev_file = NULL;
 579         struct fd f = fdget(fd);
 580
 581         if (!f.file)
 582                 return NULL;
 583
 584         if (f.file->f_op != &uverbs_event_fops)
 585                 goto out;
 586
 587         ev_file = f.file->private_data;
 588         if (ev_file->is_async) {
 589                 ev_file = NULL;
 590                 goto out;
 591         }
 592
 593         kref_get(&ev_file->ref);
 594
 595 out:
 596         fdput(f);
 597         return ev_file;
 598 }
 599
 600 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 601                              size_t count, loff_t *pos)
 602 {
 603         struct ib_uverbs_file *file = filp->private_data;
 604         struct ib_uverbs_cmd_hdr hdr;
 605         __u32 flags;
 606
 607         if (count < sizeof hdr)
 608                 return -EINVAL;
 609
 610         if (copy_from_user(&hdr, buf, sizeof hdr))
 611                 return -EFAULT;
 612
 613         flags = (hdr.command &
 614                  IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
 615
 616         if (!flags) {
 617                 __u32 command;
 618
 619                 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 620                                            IB_USER_VERBS_CMD_COMMAND_MASK))
 621                         return -EINVAL;
 622
 623                 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 624
 625                 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
 626                     !uverbs_cmd_table[command])
 627                         return -EINVAL;
 628
 629                 if (!file->ucontext &&
 630                     command != IB_USER_VERBS_CMD_GET_CONTEXT)
 631                         return -EINVAL;
 632
 633                 if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << command)))
 634                         return -ENOSYS;
 635
 636                 if (hdr.in_words * 4 != count)
 637                         return -EINVAL;
 638
 639                 return uverbs_cmd_table[command](file,
 640                                                  buf + sizeof(hdr),
 641                                                  hdr.in_words * 4,
 642                                                  hdr.out_words * 4);
 643
 644         } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
 645                 __u32 command;
 646
 647                 struct ib_uverbs_ex_cmd_hdr ex_hdr;
 648                 struct ib_udata ucore;
 649                 struct ib_udata uhw;
 650                 int err;
 651                 size_t written_count = count;
 652
 653                 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 654                                            IB_USER_VERBS_CMD_COMMAND_MASK))
 655                         return -EINVAL;
 656
 657                 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 658
 659                 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
 660                     !uverbs_ex_cmd_table[command])
 661                         return -ENOSYS;
 662
 663                 if (!file->ucontext)
 664                         return -EINVAL;
 665
 666                 if (!(file->device->ib_dev->uverbs_ex_cmd_mask & (1ull << command)))
 667                         return -ENOSYS;
 668
 669                 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 670                         return -EINVAL;
 671
 672                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
 673                         return -EFAULT;
 674
 675                 count -= sizeof(hdr) + sizeof(ex_hdr);
 676                 buf += sizeof(hdr) + sizeof(ex_hdr);
 677
 678                 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count)
 679                         return -EINVAL;
 680
 681                 if (ex_hdr.cmd_hdr_reserved)
 682                         return -EINVAL;
 683
 684                 if (ex_hdr.response) {
 685                         if (!hdr.out_words && !ex_hdr.provider_out_words)
 686                                 return -EINVAL;
 687
 688                         if (!access_ok(VERIFY_WRITE,
 689                                        (void __user *) (unsigned long) ex_hdr.response,
 690                                        (hdr.out_words + ex_hdr.provider_out_words) * 8))
 691                                 return -EFAULT;
 692                 } else {
 693                         if (hdr.out_words || ex_hdr.provider_out_words)
 694                                 return -EINVAL;
 695                 }
 696
 697                 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
 698                                        hdr.in_words * 8, hdr.out_words * 8);
 699
 700                 INIT_UDATA_BUF_OR_NULL(&uhw,
 701                                        buf + ucore.inlen,
 702                                        (unsigned long) ex_hdr.response + ucore.outlen,
 703                                        ex_hdr.provider_in_words * 8,
 704                                        ex_hdr.provider_out_words * 8);
 705
 706                 err = uverbs_ex_cmd_table[command](file,
 707                                                    &ucore,
 708                                                    &uhw);
 709
 710                 if (err)
 711                         return err;
 712
 713                 return written_count;
 714         }
 715
 716         return -ENOSYS;
 717 }
 718
 719 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 720 {
 721         struct ib_uverbs_file *file = filp->private_data;
 722
 723         if (!file->ucontext)
 724                 return -ENODEV;
 725         else
 726                 return file->device->ib_dev->mmap(file->ucontext, vma);
 727 }
 728
 729 /*
 730  * ib_uverbs_open() does not need the BKL:
 731  *
 732  *  - the ib_uverbs_device structures are properly reference counted and
 733  *    everything else is purely local to the file being created, so
 734  *    races against other open calls are not a problem;
 735  *  - there is no ioctl method to race against;
 736  *  - the open method will either immediately run -ENXIO, or all
 737  *    required initialization will be done.
 738  */
 739 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 740 {
 741         struct ib_uverbs_device *dev;
 742         struct ib_uverbs_file *file;
 743         int ret;
 744
 745         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 746         if (dev)
 747                 kref_get(&dev->ref);
 748         else
 749                 return -ENXIO;
 750
 751         if (!try_module_get(dev->ib_dev->owner)) {
 752                 ret = -ENODEV;
 753                 goto err;
 754         }
 755
 756         file = kmalloc(sizeof *file, GFP_KERNEL);
 757         if (!file) {
 758                 ret = -ENOMEM;
 759                 goto err_module;
 760         }
 761
 762         file->device     = dev;
 763         file->ucontext   = NULL;
 764         file->async_file = NULL;
 765         kref_init(&file->ref);
 766         mutex_init(&file->mutex);
 767
 768         filp->private_data = file;
 769
 770         return nonseekable_open(inode, filp);
 771
 772 err_module:
 773         module_put(dev->ib_dev->owner);
 774
 775 err:
 776         kref_put(&dev->ref, ib_uverbs_release_dev);
 777         return ret;
 778 }
 779
 780 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 781 {
 782         struct ib_uverbs_file *file = filp->private_data;
 783
 784         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 785
 786         if (file->async_file)
 787                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 788
 789         kref_put(&file->ref, ib_uverbs_release_file);
 790
 791         return 0;
 792 }
 793
 794 static const struct file_operations uverbs_fops = {
 795         .owner   = THIS_MODULE,
 796         .write   = ib_uverbs_write,
 797         .open    = ib_uverbs_open,
 798         .release = ib_uverbs_close,
 799         .llseek  = no_llseek,
 800 };
 801
 802 static const struct file_operations uverbs_mmap_fops = {
 803         .owner   = THIS_MODULE,
 804         .write   = ib_uverbs_write,
 805         .mmap    = ib_uverbs_mmap,
 806         .open    = ib_uverbs_open,
 807         .release = ib_uverbs_close,
 808         .llseek  = no_llseek,
 809 };
 810
 811 static struct ib_client uverbs_client = {
 812         .name   = "uverbs",
 813         .add    = ib_uverbs_add_one,
 814         .remove = ib_uverbs_remove_one
 815 };
 816
 817 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 818                           char *buf)
 819 {
 820         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 821
 822         if (!dev)
 823                 return -ENODEV;
 824
 825         return sprintf(buf, "%s\n", dev->ib_dev->name);
 826 }
 827 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 828
 829 static ssize_t show_dev_abi_version(struct device *device,
 830                                     struct device_attribute *attr, char *buf)
 831 {
 832         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 833
 834         if (!dev)
 835                 return -ENODEV;
 836
 837         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 838 }
 839 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 840
 841 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
 842                          __stringify(IB_USER_VERBS_ABI_VERSION));
 843
 844 static dev_t overflow_maj;
 845 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
 846
 847 /*
 848  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
 849  * requesting a new major number and doubling the number of max devices we
 850  * support. It's stupid, but simple.
 851  */
 852 static int find_overflow_devnum(void)
 853 {
 854         int ret;
 855
 856         if (!overflow_maj) {
 857                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
 858                                           "infiniband_verbs");
 859                 if (ret) {
 860                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
 861                         return ret;
 862                 }
 863         }
 864
 865         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
 866         if (ret >= IB_UVERBS_MAX_DEVICES)
 867                 return -1;
 868
 869         return ret;
 870 }
 871
 872 static void ib_uverbs_add_one(struct ib_device *device)
 873 {
 874         int devnum;
 875         dev_t base;
 876         struct ib_uverbs_device *uverbs_dev;
 877
 878         if (!device->alloc_ucontext)
 879                 return;
 880
 881         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 882         if (!uverbs_dev)
 883                 return;
 884
 885         kref_init(&uverbs_dev->ref);
 886         init_completion(&uverbs_dev->comp);
 887         uverbs_dev->xrcd_tree = RB_ROOT;
 888         mutex_init(&uverbs_dev->xrcd_tree_mutex);
 889
 890         spin_lock(&map_lock);
 891         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 892         if (devnum >= IB_UVERBS_MAX_DEVICES) {
 893                 spin_unlock(&map_lock);
 894                 devnum = find_overflow_devnum();
 895                 if (devnum < 0)
 896                         goto err;
 897
 898                 spin_lock(&map_lock);
 899                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
 900                 base = devnum + overflow_maj;
 901                 set_bit(devnum, overflow_map);
 902         } else {
 903                 uverbs_dev->devnum = devnum;
 904                 base = devnum + IB_UVERBS_BASE_DEV;
 905                 set_bit(devnum, dev_map);
 906         }
 907         spin_unlock(&map_lock);
 908
 909         uverbs_dev->ib_dev           = device;
 910         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 911
 912         cdev_init(&uverbs_dev->cdev, NULL);
 913         uverbs_dev->cdev.owner = THIS_MODULE;
 914         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 915         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
 916         if (cdev_add(&uverbs_dev->cdev, base, 1))
 917                 goto err_cdev;
 918
 919         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 920                                         uverbs_dev->cdev.dev, uverbs_dev,
 921                                         "uverbs%d", uverbs_dev->devnum);
 922         if (IS_ERR(uverbs_dev->dev))
 923                 goto err_cdev;
 924
 925         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 926                 goto err_class;
 927         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 928                 goto err_class;
 929
 930         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 931
 932         return;
 933
 934 err_class:
 935         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 936
 937 err_cdev:
 938         cdev_del(&uverbs_dev->cdev);
 939         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 940                 clear_bit(devnum, dev_map);
 941         else
 942                 clear_bit(devnum, overflow_map);
 943
 944 err:
 945         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 946         wait_for_completion(&uverbs_dev->comp);
 947         kfree(uverbs_dev);
 948         return;
 949 }
 950
 951 static void ib_uverbs_remove_one(struct ib_device *device)
 952 {
 953         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 954
 955         if (!uverbs_dev)
 956                 return;
 957
 958         dev_set_drvdata(uverbs_dev->dev, NULL);
 959         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 960         cdev_del(&uverbs_dev->cdev);
 961
 962         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 963                 clear_bit(uverbs_dev->devnum, dev_map);
 964         else
 965                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
 966
 967         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
 968         wait_for_completion(&uverbs_dev->comp);
 969         kfree(uverbs_dev);
 970 }
 971
 972 static char *uverbs_devnode(struct device *dev, umode_t *mode)
 973 {
 974         if (mode)
 975                 *mode = 0666;
 976         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
 977 }
 978
 979 static int __init ib_uverbs_init(void)
 980 {
 981         int ret;
 982
 983         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 984                                      "infiniband_verbs");
 985         if (ret) {
 986                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 987                 goto out;
 988         }
 989
 990         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 991         if (IS_ERR(uverbs_class)) {
 992                 ret = PTR_ERR(uverbs_class);
 993                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 994                 goto out_chrdev;
 995         }
 996
 997         uverbs_class->devnode = uverbs_devnode;
 998
 999         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1000         if (ret) {
1001                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
1002                 goto out_class;
1003         }
1004
1005         ret = ib_register_client(&uverbs_client);
1006         if (ret) {
1007                 printk(KERN_ERR "user_verbs: couldn't register client\n");
1008                 goto out_class;
1009         }
1010
1011         return 0;
1012
1013 out_class:
1014         class_destroy(uverbs_class);
1015
1016 out_chrdev:
1017         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1018
1019 out:
1020         return ret;
1021 }
1022
1023 static void __exit ib_uverbs_cleanup(void)
1024 {
1025         ib_unregister_client(&uverbs_client);
1026         class_destroy(uverbs_class);
1027         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1028         if (overflow_maj)
1029                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1030         idr_destroy(&ib_uverbs_pd_idr);
1031         idr_destroy(&ib_uverbs_mr_idr);
1032         idr_destroy(&ib_uverbs_mw_idr);
1033         idr_destroy(&ib_uverbs_ah_idr);
1034         idr_destroy(&ib_uverbs_cq_idr);
1035         idr_destroy(&ib_uverbs_qp_idr);
1036         idr_destroy(&ib_uverbs_srq_idr);
1037 }
1038
1039 module_init(ib_uverbs_init);
1040 module_exit(ib_uverbs_cleanup);