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