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