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 <rdma/ib.h>
  52
  53 #include "uverbs.h"
  54
  55 MODULE_AUTHOR("Roland Dreier");
  56 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  57 MODULE_LICENSE("Dual BSD/GPL");
  58
  59 enum {
  60         IB_UVERBS_MAJOR       = 231,
  61         IB_UVERBS_BASE_MINOR  = 192,
  62         IB_UVERBS_MAX_DEVICES = 32
  63 };
  64
  65 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  66
  67 static struct class *uverbs_class;
  68
  69 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  70 DEFINE_IDR(ib_uverbs_pd_idr);
  71 DEFINE_IDR(ib_uverbs_mr_idr);
  72 DEFINE_IDR(ib_uverbs_mw_idr);
  73 DEFINE_IDR(ib_uverbs_ah_idr);
  74 DEFINE_IDR(ib_uverbs_cq_idr);
  75 DEFINE_IDR(ib_uverbs_qp_idr);
  76 DEFINE_IDR(ib_uverbs_srq_idr);
  77 DEFINE_IDR(ib_uverbs_xrcd_idr);
  78
  79 static DEFINE_SPINLOCK(map_lock);
  80 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  81
  82 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  83                                      const char __user *buf, int in_len,
  84                                      int out_len) = {
  85         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  86         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  87         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  88         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  89         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  90         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  91         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  92         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  93         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  94         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  95         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  96         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  97         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  98         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  99         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 100         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 101         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 102         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 103         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 104         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 105         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 106         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 107         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 108         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 109         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 110         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 111         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 112         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 113         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 114         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 115         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 116         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp
 117 };
 118
 119 static void ib_uverbs_add_one(struct ib_device *device);
 120 static void ib_uverbs_remove_one(struct ib_device *device);
 121
 122 static void ib_uverbs_release_dev(struct kobject *kobj)
 123 {
 124         struct ib_uverbs_device *dev =
 125                 container_of(kobj, struct ib_uverbs_device, kobj);
 126
 127         kfree(dev);
 128 }
 129
 130 static struct kobj_type ib_uverbs_dev_ktype = {
 131         .release = ib_uverbs_release_dev,
 132 };
 133
 134 static void ib_uverbs_release_event_file(struct kref *ref)
 135 {
 136         struct ib_uverbs_event_file *file =
 137                 container_of(ref, struct ib_uverbs_event_file, ref);
 138
 139         kfree(file);
 140 }
 141
 142 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 143                           struct ib_uverbs_event_file *ev_file,
 144                           struct ib_ucq_object *uobj)
 145 {
 146         struct ib_uverbs_event *evt, *tmp;
 147
 148         if (ev_file) {
 149                 spin_lock_irq(&ev_file->lock);
 150                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 151                         list_del(&evt->list);
 152                         kfree(evt);
 153                 }
 154                 spin_unlock_irq(&ev_file->lock);
 155
 156                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 157         }
 158
 159         spin_lock_irq(&file->async_file->lock);
 160         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 161                 list_del(&evt->list);
 162                 kfree(evt);
 163         }
 164         spin_unlock_irq(&file->async_file->lock);
 165 }
 166
 167 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 168                               struct ib_uevent_object *uobj)
 169 {
 170         struct ib_uverbs_event *evt, *tmp;
 171
 172         spin_lock_irq(&file->async_file->lock);
 173         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 174                 list_del(&evt->list);
 175                 kfree(evt);
 176         }
 177         spin_unlock_irq(&file->async_file->lock);
 178 }
 179
 180 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 181                                     struct ib_uqp_object *uobj)
 182 {
 183         struct ib_uverbs_mcast_entry *mcast, *tmp;
 184
 185         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 186                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 187                 list_del(&mcast->list);
 188                 kfree(mcast);
 189         }
 190 }
 191
 192 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 193                                       struct ib_ucontext *context)
 194 {
 195         struct ib_uobject *uobj, *tmp;
 196
 197         if (!context)
 198                 return 0;
 199
 200         context->closing = 1;
 201
 202         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 203                 struct ib_ah *ah = uobj->object;
 204
 205                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 206                 ib_destroy_ah(ah);
 207                 kfree(uobj);
 208         }
 209
 210         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 211                 struct ib_qp *qp = uobj->object;
 212                 struct ib_uqp_object *uqp =
 213                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 214
 215                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 216                 if (qp == qp->real_qp)
 217                         ib_uverbs_detach_umcast(qp, uqp);
 218                 ib_destroy_qp(qp);
 219                 ib_uverbs_release_uevent(file, &uqp->uevent);
 220                 kfree(uqp);
 221         }
 222
 223         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 224                 struct ib_cq *cq = uobj->object;
 225                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 226                 struct ib_ucq_object *ucq =
 227                         container_of(uobj, struct ib_ucq_object, uobject);
 228
 229                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 230                 ib_destroy_cq(cq);
 231                 ib_uverbs_release_ucq(file, ev_file, ucq);
 232                 kfree(ucq);
 233         }
 234
 235         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 236                 struct ib_srq *srq = uobj->object;
 237                 struct ib_uevent_object *uevent =
 238                         container_of(uobj, struct ib_uevent_object, uobject);
 239
 240                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 241                 ib_destroy_srq(srq);
 242                 ib_uverbs_release_uevent(file, uevent);
 243                 kfree(uevent);
 244         }
 245
 246         /* XXX Free MWs */
 247
 248         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 249                 struct ib_mr *mr = uobj->object;
 250
 251                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 252                 ib_dereg_mr(mr);
 253                 kfree(uobj);
 254         }
 255
 256         mutex_lock(&file->device->xrcd_tree_mutex);
 257         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 258                 struct ib_xrcd *xrcd = uobj->object;
 259                 struct ib_uxrcd_object *uxrcd =
 260                         container_of(uobj, struct ib_uxrcd_object, uobject);
 261
 262                 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 263                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 264                 kfree(uxrcd);
 265         }
 266         mutex_unlock(&file->device->xrcd_tree_mutex);
 267
 268         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 269                 struct ib_pd *pd = uobj->object;
 270
 271                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 272                 ib_dealloc_pd(pd);
 273                 kfree(uobj);
 274         }
 275
 276         return context->device->dealloc_ucontext(context);
 277 }
 278
 279 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
 280 {
 281         complete(&dev->comp);
 282 }
 283
 284 static void ib_uverbs_release_file(struct kref *ref)
 285 {
 286         struct ib_uverbs_file *file =
 287                 container_of(ref, struct ib_uverbs_file, ref);
 288
 289         module_put(file->device->ib_dev->owner);
 290         if (atomic_dec_and_test(&file->device->refcount))
 291                 ib_uverbs_comp_dev(file->device);
 292
 293         kfree(file);
 294 }
 295
 296 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 297                                     size_t count, loff_t *pos)
 298 {
 299         struct ib_uverbs_event_file *file = filp->private_data;
 300         struct ib_uverbs_event *event;
 301         int eventsz;
 302         int ret = 0;
 303
 304         spin_lock_irq(&file->lock);
 305
 306         while (list_empty(&file->event_list)) {
 307                 spin_unlock_irq(&file->lock);
 308
 309                 if (filp->f_flags & O_NONBLOCK)
 310                         return -EAGAIN;
 311
 312                 if (wait_event_interruptible(file->poll_wait,
 313                                              !list_empty(&file->event_list)))
 314                         return -ERESTARTSYS;
 315
 316                 spin_lock_irq(&file->lock);
 317         }
 318
 319         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 320
 321         if (file->is_async)
 322                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 323         else
 324                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 325
 326         if (eventsz > count) {
 327                 ret   = -EINVAL;
 328                 event = NULL;
 329         } else {
 330                 list_del(file->event_list.next);
 331                 if (event->counter) {
 332                         ++(*event->counter);
 333                         list_del(&event->obj_list);
 334                 }
 335         }
 336
 337         spin_unlock_irq(&file->lock);
 338
 339         if (event) {
 340                 if (copy_to_user(buf, event, eventsz))
 341                         ret = -EFAULT;
 342                 else
 343                         ret = eventsz;
 344         }
 345
 346         kfree(event);
 347
 348         return ret;
 349 }
 350
 351 static unsigned int ib_uverbs_event_poll(struct file *filp,
 352                                          struct poll_table_struct *wait)
 353 {
 354         unsigned int pollflags = 0;
 355         struct ib_uverbs_event_file *file = filp->private_data;
 356
 357         poll_wait(filp, &file->poll_wait, wait);
 358
 359         spin_lock_irq(&file->lock);
 360         if (!list_empty(&file->event_list))
 361                 pollflags = POLLIN | POLLRDNORM;
 362         spin_unlock_irq(&file->lock);
 363
 364         return pollflags;
 365 }
 366
 367 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 368 {
 369         struct ib_uverbs_event_file *file = filp->private_data;
 370
 371         return fasync_helper(fd, filp, on, &file->async_queue);
 372 }
 373
 374 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 375 {
 376         struct ib_uverbs_event_file *file = filp->private_data;
 377         struct ib_uverbs_event *entry, *tmp;
 378
 379         spin_lock_irq(&file->lock);
 380         file->is_closed = 1;
 381         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 382                 if (entry->counter)
 383                         list_del(&entry->obj_list);
 384                 kfree(entry);
 385         }
 386         spin_unlock_irq(&file->lock);
 387
 388         if (file->is_async) {
 389                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 390                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 391         }
 392         kref_put(&file->ref, ib_uverbs_release_event_file);
 393
 394         return 0;
 395 }
 396
 397 static const struct file_operations uverbs_event_fops = {
 398         .owner   = THIS_MODULE,
 399         .read    = ib_uverbs_event_read,
 400         .poll    = ib_uverbs_event_poll,
 401         .release = ib_uverbs_event_close,
 402         .fasync  = ib_uverbs_event_fasync,
 403         .llseek  = no_llseek,
 404 };
 405
 406 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 407 {
 408         struct ib_uverbs_event_file    *file = cq_context;
 409         struct ib_ucq_object           *uobj;
 410         struct ib_uverbs_event         *entry;
 411         unsigned long                   flags;
 412
 413         if (!file)
 414                 return;
 415
 416         spin_lock_irqsave(&file->lock, flags);
 417         if (file->is_closed) {
 418                 spin_unlock_irqrestore(&file->lock, flags);
 419                 return;
 420         }
 421
 422         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 423         if (!entry) {
 424                 spin_unlock_irqrestore(&file->lock, flags);
 425                 return;
 426         }
 427
 428         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 429
 430         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 431         entry->counter             = &uobj->comp_events_reported;
 432
 433         list_add_tail(&entry->list, &file->event_list);
 434         list_add_tail(&entry->obj_list, &uobj->comp_list);
 435         spin_unlock_irqrestore(&file->lock, flags);
 436
 437         wake_up_interruptible(&file->poll_wait);
 438         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 439 }
 440
 441 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 442                                     __u64 element, __u64 event,
 443                                     struct list_head *obj_list,
 444                                     u32 *counter)
 445 {
 446         struct ib_uverbs_event *entry;
 447         unsigned long flags;
 448
 449         spin_lock_irqsave(&file->async_file->lock, flags);
 450         if (file->async_file->is_closed) {
 451                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 452                 return;
 453         }
 454
 455         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 456         if (!entry) {
 457                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 458                 return;
 459         }
 460
 461         entry->desc.async.element    = element;
 462         entry->desc.async.event_type = event;
 463         entry->desc.async.reserved   = 0;
 464         entry->counter               = counter;
 465
 466         list_add_tail(&entry->list, &file->async_file->event_list);
 467         if (obj_list)
 468                 list_add_tail(&entry->obj_list, obj_list);
 469         spin_unlock_irqrestore(&file->async_file->lock, flags);
 470
 471         wake_up_interruptible(&file->async_file->poll_wait);
 472         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 473 }
 474
 475 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 476 {
 477         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 478                                                   struct ib_ucq_object, uobject);
 479
 480         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 481                                 event->event, &uobj->async_list,
 482                                 &uobj->async_events_reported);
 483 }
 484
 485 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 486 {
 487         struct ib_uevent_object *uobj;
 488
 489         uobj = container_of(event->element.qp->uobject,
 490                             struct ib_uevent_object, uobject);
 491
 492         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 493                                 event->event, &uobj->event_list,
 494                                 &uobj->events_reported);
 495 }
 496
 497 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 498 {
 499         struct ib_uevent_object *uobj;
 500
 501         uobj = container_of(event->element.srq->uobject,
 502                             struct ib_uevent_object, uobject);
 503
 504         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 505                                 event->event, &uobj->event_list,
 506                                 &uobj->events_reported);
 507 }
 508
 509 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 510                              struct ib_event *event)
 511 {
 512         struct ib_uverbs_file *file =
 513                 container_of(handler, struct ib_uverbs_file, event_handler);
 514
 515         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 516                                 NULL, NULL);
 517 }
 518
 519 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 520                                         int is_async)
 521 {
 522         struct ib_uverbs_event_file *ev_file;
 523         struct file *filp;
 524
 525         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 526         if (!ev_file)
 527                 return ERR_PTR(-ENOMEM);
 528
 529         kref_init(&ev_file->ref);
 530         spin_lock_init(&ev_file->lock);
 531         INIT_LIST_HEAD(&ev_file->event_list);
 532         init_waitqueue_head(&ev_file->poll_wait);
 533         ev_file->uverbs_file = uverbs_file;
 534         ev_file->async_queue = NULL;
 535         ev_file->is_async    = is_async;
 536         ev_file->is_closed   = 0;
 537
 538         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 539                                   ev_file, O_RDONLY);
 540         if (IS_ERR(filp))
 541                 kfree(ev_file);
 542
 543         return filp;
 544 }
 545
 546 /*
 547  * Look up a completion event file by FD.  If lookup is successful,
 548  * takes a ref to the event file struct that it returns; if
 549  * unsuccessful, returns NULL.
 550  */
 551 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 552 {
 553         struct ib_uverbs_event_file *ev_file = NULL;
 554         struct file *filp;
 555
 556         filp = fget(fd);
 557         if (!filp)
 558                 return NULL;
 559
 560         if (filp->f_op != &uverbs_event_fops)
 561                 goto out;
 562
 563         ev_file = filp->private_data;
 564         if (ev_file->is_async) {
 565                 ev_file = NULL;
 566                 goto out;
 567         }
 568
 569         kref_get(&ev_file->ref);
 570
 571 out:
 572         fput(filp);
 573         return ev_file;
 574 }
 575
 576 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 577                              size_t count, loff_t *pos)
 578 {
 579         struct ib_uverbs_file *file = filp->private_data;
 580         struct ib_uverbs_cmd_hdr hdr;
 581
 582         if (WARN_ON_ONCE(!ib_safe_file_access(filp)))
 583                 return -EACCES;
 584
 585         if (count < sizeof hdr)
 586                 return -EINVAL;
 587
 588         if (copy_from_user(&hdr, buf, sizeof hdr))
 589                 return -EFAULT;
 590
 591         if (hdr.in_words * 4 != count)
 592                 return -EINVAL;
 593
 594         if (hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
 595             !uverbs_cmd_table[hdr.command])
 596                 return -EINVAL;
 597
 598         if (!file->ucontext &&
 599             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
 600                 return -EINVAL;
 601
 602         if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
 603                 return -ENOSYS;
 604
 605         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
 606                                              hdr.in_words * 4, hdr.out_words * 4);
 607 }
 608
 609 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 610 {
 611         struct ib_uverbs_file *file = filp->private_data;
 612
 613         if (!file->ucontext)
 614                 return -ENODEV;
 615         else
 616                 return file->device->ib_dev->mmap(file->ucontext, vma);
 617 }
 618
 619 /*
 620  * ib_uverbs_open() does not need the BKL:
 621  *
 622  *  - the ib_uverbs_device structures are properly reference counted and
 623  *    everything else is purely local to the file being created, so
 624  *    races against other open calls are not a problem;
 625  *  - there is no ioctl method to race against;
 626  *  - the open method will either immediately run -ENXIO, or all
 627  *    required initialization will be done.
 628  */
 629 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 630 {
 631         struct ib_uverbs_device *dev;
 632         struct ib_uverbs_file *file;
 633         int ret;
 634
 635         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 636         if (!atomic_inc_not_zero(&dev->refcount))
 637                 return -ENXIO;
 638
 639         if (!try_module_get(dev->ib_dev->owner)) {
 640                 ret = -ENODEV;
 641                 goto err;
 642         }
 643
 644         file = kmalloc(sizeof *file, GFP_KERNEL);
 645         if (!file) {
 646                 ret = -ENOMEM;
 647                 goto err_module;
 648         }
 649
 650         file->device     = dev;
 651         file->ucontext   = NULL;
 652         file->async_file = NULL;
 653         kref_init(&file->ref);
 654         mutex_init(&file->mutex);
 655
 656         filp->private_data = file;
 657         kobject_get(&dev->kobj);
 658
 659         return nonseekable_open(inode, filp);
 660
 661 err_module:
 662         module_put(dev->ib_dev->owner);
 663
 664 err:
 665         if (atomic_dec_and_test(&dev->refcount))
 666                 ib_uverbs_comp_dev(dev);
 667
 668         return ret;
 669 }
 670
 671 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 672 {
 673         struct ib_uverbs_file *file = filp->private_data;
 674         struct ib_uverbs_device *dev = file->device;
 675
 676         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 677
 678         if (file->async_file)
 679                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 680
 681         kref_put(&file->ref, ib_uverbs_release_file);
 682         kobject_put(&dev->kobj);
 683
 684         return 0;
 685 }
 686
 687 static const struct file_operations uverbs_fops = {
 688         .owner   = THIS_MODULE,
 689         .write   = ib_uverbs_write,
 690         .open    = ib_uverbs_open,
 691         .release = ib_uverbs_close,
 692         .llseek  = no_llseek,
 693 };
 694
 695 static const struct file_operations uverbs_mmap_fops = {
 696         .owner   = THIS_MODULE,
 697         .write   = ib_uverbs_write,
 698         .mmap    = ib_uverbs_mmap,
 699         .open    = ib_uverbs_open,
 700         .release = ib_uverbs_close,
 701         .llseek  = no_llseek,
 702 };
 703
 704 static struct ib_client uverbs_client = {
 705         .name   = "uverbs",
 706         .add    = ib_uverbs_add_one,
 707         .remove = ib_uverbs_remove_one
 708 };
 709
 710 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 711                           char *buf)
 712 {
 713         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 714
 715         if (!dev)
 716                 return -ENODEV;
 717
 718         return sprintf(buf, "%s\n", dev->ib_dev->name);
 719 }
 720 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 721
 722 static ssize_t show_dev_abi_version(struct device *device,
 723                                     struct device_attribute *attr, char *buf)
 724 {
 725         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 726
 727         if (!dev)
 728                 return -ENODEV;
 729
 730         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 731 }
 732 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 733
 734 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
 735                          __stringify(IB_USER_VERBS_ABI_VERSION));
 736
 737 static dev_t overflow_maj;
 738 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
 739
 740 /*
 741  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
 742  * requesting a new major number and doubling the number of max devices we
 743  * support. It's stupid, but simple.
 744  */
 745 static int find_overflow_devnum(void)
 746 {
 747         int ret;
 748
 749         if (!overflow_maj) {
 750                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
 751                                           "infiniband_verbs");
 752                 if (ret) {
 753                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
 754                         return ret;
 755                 }
 756         }
 757
 758         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
 759         if (ret >= IB_UVERBS_MAX_DEVICES)
 760                 return -1;
 761
 762         return ret;
 763 }
 764
 765 static void ib_uverbs_add_one(struct ib_device *device)
 766 {
 767         int devnum;
 768         dev_t base;
 769         struct ib_uverbs_device *uverbs_dev;
 770
 771         if (!device->alloc_ucontext)
 772                 return;
 773
 774         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 775         if (!uverbs_dev)
 776                 return;
 777
 778         atomic_set(&uverbs_dev->refcount, 1);
 779         init_completion(&uverbs_dev->comp);
 780         uverbs_dev->xrcd_tree = RB_ROOT;
 781         mutex_init(&uverbs_dev->xrcd_tree_mutex);
 782         kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
 783
 784         spin_lock(&map_lock);
 785         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 786         if (devnum >= IB_UVERBS_MAX_DEVICES) {
 787                 spin_unlock(&map_lock);
 788                 devnum = find_overflow_devnum();
 789                 if (devnum < 0)
 790                         goto err;
 791
 792                 spin_lock(&map_lock);
 793                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
 794                 base = devnum + overflow_maj;
 795                 set_bit(devnum, overflow_map);
 796         } else {
 797                 uverbs_dev->devnum = devnum;
 798                 base = devnum + IB_UVERBS_BASE_DEV;
 799                 set_bit(devnum, dev_map);
 800         }
 801         spin_unlock(&map_lock);
 802
 803         uverbs_dev->ib_dev           = device;
 804         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 805
 806         cdev_init(&uverbs_dev->cdev, NULL);
 807         uverbs_dev->cdev.owner = THIS_MODULE;
 808         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 809         uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
 810         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
 811         if (cdev_add(&uverbs_dev->cdev, base, 1))
 812                 goto err_cdev;
 813
 814         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 815                                         uverbs_dev->cdev.dev, uverbs_dev,
 816                                         "uverbs%d", uverbs_dev->devnum);
 817         if (IS_ERR(uverbs_dev->dev))
 818                 goto err_cdev;
 819
 820         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 821                 goto err_class;
 822         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 823                 goto err_class;
 824
 825         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 826
 827         return;
 828
 829 err_class:
 830         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 831
 832 err_cdev:
 833         cdev_del(&uverbs_dev->cdev);
 834         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 835                 clear_bit(devnum, dev_map);
 836         else
 837                 clear_bit(devnum, overflow_map);
 838
 839 err:
 840         if (atomic_dec_and_test(&uverbs_dev->refcount))
 841                 ib_uverbs_comp_dev(uverbs_dev);
 842         wait_for_completion(&uverbs_dev->comp);
 843         kobject_put(&uverbs_dev->kobj);
 844         return;
 845 }
 846
 847 static void ib_uverbs_remove_one(struct ib_device *device)
 848 {
 849         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 850
 851         if (!uverbs_dev)
 852                 return;
 853
 854         dev_set_drvdata(uverbs_dev->dev, NULL);
 855         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 856         cdev_del(&uverbs_dev->cdev);
 857
 858         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 859                 clear_bit(uverbs_dev->devnum, dev_map);
 860         else
 861                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
 862
 863         if (atomic_dec_and_test(&uverbs_dev->refcount))
 864                 ib_uverbs_comp_dev(uverbs_dev);
 865         wait_for_completion(&uverbs_dev->comp);
 866         kobject_put(&uverbs_dev->kobj);
 867 }
 868
 869 static char *uverbs_devnode(struct device *dev, mode_t *mode)
 870 {
 871         if (mode)
 872                 *mode = 0666;
 873         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
 874 }
 875
 876 static int __init ib_uverbs_init(void)
 877 {
 878         int ret;
 879
 880         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
 881                                      "infiniband_verbs");
 882         if (ret) {
 883                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
 884                 goto out;
 885         }
 886
 887         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
 888         if (IS_ERR(uverbs_class)) {
 889                 ret = PTR_ERR(uverbs_class);
 890                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
 891                 goto out_chrdev;
 892         }
 893
 894         uverbs_class->devnode = uverbs_devnode;
 895
 896         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
 897         if (ret) {
 898                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
 899                 goto out_class;
 900         }
 901
 902         ret = ib_register_client(&uverbs_client);
 903         if (ret) {
 904                 printk(KERN_ERR "user_verbs: couldn't register client\n");
 905                 goto out_class;
 906         }
 907
 908         return 0;
 909
 910 out_class:
 911         class_destroy(uverbs_class);
 912
 913 out_chrdev:
 914         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 915
 916 out:
 917         return ret;
 918 }
 919
 920 static void __exit ib_uverbs_cleanup(void)
 921 {
 922         ib_unregister_client(&uverbs_client);
 923         class_destroy(uverbs_class);
 924         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
 925         if (overflow_maj)
 926                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
 927         idr_destroy(&ib_uverbs_pd_idr);
 928         idr_destroy(&ib_uverbs_mr_idr);
 929         idr_destroy(&ib_uverbs_mw_idr);
 930         idr_destroy(&ib_uverbs_ah_idr);
 931         idr_destroy(&ib_uverbs_cq_idr);
 932         idr_destroy(&ib_uverbs_qp_idr);
 933         idr_destroy(&ib_uverbs_srq_idr);
 934 }
 935
 936 module_init(ib_uverbs_init);
 937 module_exit(ib_uverbs_cleanup);