kernel/drivers/infiniband/core/uverbs_main.c

   1 /*
   2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   5  * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
   6  * Copyright (c) 2005 PathScale, Inc. All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  */
  36
  37 #include <linux/module.h>
  38 #include <linux/init.h>
  39 #include <linux/device.h>
  40 #include <linux/err.h>
  41 #include <linux/fs.h>
  42 #include <linux/poll.h>
  43 #include <linux/sched.h>
  44 #include <linux/file.h>
  45 #include <linux/cdev.h>
  46 #include <linux/anon_inodes.h>
  47 #include <linux/slab.h>
  48
  49 #include <asm/uaccess.h>
  50
  51 #include "uverbs.h"
  52
  53 MODULE_AUTHOR("Roland Dreier");
  54 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  55 MODULE_LICENSE("Dual BSD/GPL");
  56
  57 enum {
  58         IB_UVERBS_MAJOR       = 231,
  59         IB_UVERBS_BASE_MINOR  = 192,
  60         IB_UVERBS_MAX_DEVICES = 32
  61 };
  62
  63 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  64
  65 static struct class *uverbs_class;
  66
  67 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
  68 DEFINE_IDR(ib_uverbs_pd_idr);
  69 DEFINE_IDR(ib_uverbs_mr_idr);
  70 DEFINE_IDR(ib_uverbs_mw_idr);
  71 DEFINE_IDR(ib_uverbs_ah_idr);
  72 DEFINE_IDR(ib_uverbs_cq_idr);
  73 DEFINE_IDR(ib_uverbs_qp_idr);
  74 DEFINE_IDR(ib_uverbs_srq_idr);
  75 DEFINE_IDR(ib_uverbs_xrcd_idr);
  76 DEFINE_IDR(ib_uverbs_rule_idr);
  77
  78 static DEFINE_SPINLOCK(map_lock);
  79 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  80
  81 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  82                                      const char __user *buf, int in_len,
  83                                      int out_len) = {
  84         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  85         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  86         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  87         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  88         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  89         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  90         [IB_USER_VERBS_CMD_REREG_MR]            = ib_uverbs_rereg_mr,
  91         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  92         [IB_USER_VERBS_CMD_ALLOC_MW]            = ib_uverbs_alloc_mw,
  93         [IB_USER_VERBS_CMD_DEALLOC_MW]          = ib_uverbs_dealloc_mw,
  94         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  95         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  96         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  97         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  98         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  99         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
 100         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
 101         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 102         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 103         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 104         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 105         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 106         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 107         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 108         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 109         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 110         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 111         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 112         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 113         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 114         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 115         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 116         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 117         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 118         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp,
 119 };
 120
 121 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
 122                                     struct ib_udata *ucore,
 123                                     struct ib_udata *uhw) = {
 124         [IB_USER_VERBS_EX_CMD_CREATE_FLOW]      = ib_uverbs_ex_create_flow,
 125         [IB_USER_VERBS_EX_CMD_DESTROY_FLOW]     = ib_uverbs_ex_destroy_flow,
 126         [IB_USER_VERBS_EX_CMD_QUERY_DEVICE]     = ib_uverbs_ex_query_device,
 127 };
 128
 129 static void ib_uverbs_add_one(struct ib_device *device);
 130 static void ib_uverbs_remove_one(struct ib_device *device);
 131
 132 static void ib_uverbs_release_dev(struct kobject *kobj)
 133 {
 134         struct ib_uverbs_device *dev =
 135                 container_of(kobj, struct ib_uverbs_device, kobj);
 136
 137         kfree(dev);
 138 }
 139
 140 static struct kobj_type ib_uverbs_dev_ktype = {
 141         .release = ib_uverbs_release_dev,
 142 };
 143
 144 static void ib_uverbs_release_event_file(struct kref *ref)
 145 {
 146         struct ib_uverbs_event_file *file =
 147                 container_of(ref, struct ib_uverbs_event_file, ref);
 148
 149         kfree(file);
 150 }
 151
 152 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 153                           struct ib_uverbs_event_file *ev_file,
 154                           struct ib_ucq_object *uobj)
 155 {
 156         struct ib_uverbs_event *evt, *tmp;
 157
 158         if (ev_file) {
 159                 spin_lock_irq(&ev_file->lock);
 160                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 161                         list_del(&evt->list);
 162                         kfree(evt);
 163                 }
 164                 spin_unlock_irq(&ev_file->lock);
 165
 166                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 167         }
 168
 169         spin_lock_irq(&file->async_file->lock);
 170         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 171                 list_del(&evt->list);
 172                 kfree(evt);
 173         }
 174         spin_unlock_irq(&file->async_file->lock);
 175 }
 176
 177 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 178                               struct ib_uevent_object *uobj)
 179 {
 180         struct ib_uverbs_event *evt, *tmp;
 181
 182         spin_lock_irq(&file->async_file->lock);
 183         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 184                 list_del(&evt->list);
 185                 kfree(evt);
 186         }
 187         spin_unlock_irq(&file->async_file->lock);
 188 }
 189
 190 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 191                                     struct ib_uqp_object *uobj)
 192 {
 193         struct ib_uverbs_mcast_entry *mcast, *tmp;
 194
 195         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 196                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 197                 list_del(&mcast->list);
 198                 kfree(mcast);
 199         }
 200 }
 201
 202 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 203                                       struct ib_ucontext *context)
 204 {
 205         struct ib_uobject *uobj, *tmp;
 206
 207         if (!context)
 208                 return 0;
 209
 210         context->closing = 1;
 211
 212         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 213                 struct ib_ah *ah = uobj->object;
 214
 215                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
 216                 ib_destroy_ah(ah);
 217                 kfree(uobj);
 218         }
 219
 220         /* Remove MWs before QPs, in order to support type 2A MWs. */
 221         list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
 222                 struct ib_mw *mw = uobj->object;
 223
 224                 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
 225                 ib_dealloc_mw(mw);
 226                 kfree(uobj);
 227         }
 228
 229         list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
 230                 struct ib_flow *flow_id = uobj->object;
 231
 232                 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
 233                 ib_destroy_flow(flow_id);
 234                 kfree(uobj);
 235         }
 236
 237         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 238                 struct ib_qp *qp = uobj->object;
 239                 struct ib_uqp_object *uqp =
 240                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 241
 242                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
 243                 if (qp != qp->real_qp) {
 244                         ib_close_qp(qp);
 245                 } else {
 246                         ib_uverbs_detach_umcast(qp, uqp);
 247                         ib_destroy_qp(qp);
 248                 }
 249                 ib_uverbs_release_uevent(file, &uqp->uevent);
 250                 kfree(uqp);
 251         }
 252
 253         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 254                 struct ib_srq *srq = uobj->object;
 255                 struct ib_uevent_object *uevent =
 256                         container_of(uobj, struct ib_uevent_object, uobject);
 257
 258                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
 259                 ib_destroy_srq(srq);
 260                 ib_uverbs_release_uevent(file, uevent);
 261                 kfree(uevent);
 262         }
 263
 264         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 265                 struct ib_cq *cq = uobj->object;
 266                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 267                 struct ib_ucq_object *ucq =
 268                         container_of(uobj, struct ib_ucq_object, uobject);
 269
 270                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
 271                 ib_destroy_cq(cq);
 272                 ib_uverbs_release_ucq(file, ev_file, ucq);
 273                 kfree(ucq);
 274         }
 275
 276         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 277                 struct ib_mr *mr = uobj->object;
 278
 279                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
 280                 ib_dereg_mr(mr);
 281                 kfree(uobj);
 282         }
 283
 284         mutex_lock(&file->device->xrcd_tree_mutex);
 285         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 286                 struct ib_xrcd *xrcd = uobj->object;
 287                 struct ib_uxrcd_object *uxrcd =
 288                         container_of(uobj, struct ib_uxrcd_object, uobject);
 289
 290                 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
 291                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 292                 kfree(uxrcd);
 293         }
 294         mutex_unlock(&file->device->xrcd_tree_mutex);
 295
 296         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 297                 struct ib_pd *pd = uobj->object;
 298
 299                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
 300                 ib_dealloc_pd(pd);
 301                 kfree(uobj);
 302         }
 303
 304         put_pid(context->tgid);
 305
 306         return context->device->dealloc_ucontext(context);
 307 }
 308
 309 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
 310 {
 311         complete(&dev->comp);
 312 }
 313
 314 static void ib_uverbs_release_file(struct kref *ref)
 315 {
 316         struct ib_uverbs_file *file =
 317                 container_of(ref, struct ib_uverbs_file, ref);
 318
 319         module_put(file->device->ib_dev->owner);
 320         if (atomic_dec_and_test(&file->device->refcount))
 321                 ib_uverbs_comp_dev(file->device);
 322
 323         kfree(file);
 324 }
 325
 326 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 327                                     size_t count, loff_t *pos)
 328 {
 329         struct ib_uverbs_event_file *file = filp->private_data;
 330         struct ib_uverbs_event *event;
 331         int eventsz;
 332         int ret = 0;
 333
 334         spin_lock_irq(&file->lock);
 335
 336         while (list_empty(&file->event_list)) {
 337                 spin_unlock_irq(&file->lock);
 338
 339                 if (filp->f_flags & O_NONBLOCK)
 340                         return -EAGAIN;
 341
 342                 if (wait_event_interruptible(file->poll_wait,
 343                                              !list_empty(&file->event_list)))
 344                         return -ERESTARTSYS;
 345
 346                 spin_lock_irq(&file->lock);
 347         }
 348
 349         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 350
 351         if (file->is_async)
 352                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 353         else
 354                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 355
 356         if (eventsz > count) {
 357                 ret   = -EINVAL;
 358                 event = NULL;
 359         } else {
 360                 list_del(file->event_list.next);
 361                 if (event->counter) {
 362                         ++(*event->counter);
 363                         list_del(&event->obj_list);
 364                 }
 365         }
 366
 367         spin_unlock_irq(&file->lock);
 368
 369         if (event) {
 370                 if (copy_to_user(buf, event, eventsz))
 371                         ret = -EFAULT;
 372                 else
 373                         ret = eventsz;
 374         }
 375
 376         kfree(event);
 377
 378         return ret;
 379 }
 380
 381 static unsigned int ib_uverbs_event_poll(struct file *filp,
 382                                          struct poll_table_struct *wait)
 383 {
 384         unsigned int pollflags = 0;
 385         struct ib_uverbs_event_file *file = filp->private_data;
 386
 387         poll_wait(filp, &file->poll_wait, wait);
 388
 389         spin_lock_irq(&file->lock);
 390         if (!list_empty(&file->event_list))
 391                 pollflags = POLLIN | POLLRDNORM;
 392         spin_unlock_irq(&file->lock);
 393
 394         return pollflags;
 395 }
 396
 397 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 398 {
 399         struct ib_uverbs_event_file *file = filp->private_data;
 400
 401         return fasync_helper(fd, filp, on, &file->async_queue);
 402 }
 403
 404 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 405 {
 406         struct ib_uverbs_event_file *file = filp->private_data;
 407         struct ib_uverbs_event *entry, *tmp;
 408
 409         spin_lock_irq(&file->lock);
 410         file->is_closed = 1;
 411         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 412                 if (entry->counter)
 413                         list_del(&entry->obj_list);
 414                 kfree(entry);
 415         }
 416         spin_unlock_irq(&file->lock);
 417
 418         if (file->is_async) {
 419                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
 420                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 421         }
 422         kref_put(&file->ref, ib_uverbs_release_event_file);
 423
 424         return 0;
 425 }
 426
 427 static const struct file_operations uverbs_event_fops = {
 428         .owner   = THIS_MODULE,
 429         .read    = ib_uverbs_event_read,
 430         .poll    = ib_uverbs_event_poll,
 431         .release = ib_uverbs_event_close,
 432         .fasync  = ib_uverbs_event_fasync,
 433         .llseek  = no_llseek,
 434 };
 435
 436 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 437 {
 438         struct ib_uverbs_event_file    *file = cq_context;
 439         struct ib_ucq_object           *uobj;
 440         struct ib_uverbs_event         *entry;
 441         unsigned long                   flags;
 442
 443         if (!file)
 444                 return;
 445
 446         spin_lock_irqsave(&file->lock, flags);
 447         if (file->is_closed) {
 448                 spin_unlock_irqrestore(&file->lock, flags);
 449                 return;
 450         }
 451
 452         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 453         if (!entry) {
 454                 spin_unlock_irqrestore(&file->lock, flags);
 455                 return;
 456         }
 457
 458         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 459
 460         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 461         entry->counter             = &uobj->comp_events_reported;
 462
 463         list_add_tail(&entry->list, &file->event_list);
 464         list_add_tail(&entry->obj_list, &uobj->comp_list);
 465         spin_unlock_irqrestore(&file->lock, flags);
 466
 467         wake_up_interruptible(&file->poll_wait);
 468         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 469 }
 470
 471 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 472                                     __u64 element, __u64 event,
 473                                     struct list_head *obj_list,
 474                                     u32 *counter)
 475 {
 476         struct ib_uverbs_event *entry;
 477         unsigned long flags;
 478
 479         spin_lock_irqsave(&file->async_file->lock, flags);
 480         if (file->async_file->is_closed) {
 481                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 482                 return;
 483         }
 484
 485         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 486         if (!entry) {
 487                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 488                 return;
 489         }
 490
 491         entry->desc.async.element    = element;
 492         entry->desc.async.event_type = event;
 493         entry->desc.async.reserved   = 0;
 494         entry->counter               = counter;
 495
 496         list_add_tail(&entry->list, &file->async_file->event_list);
 497         if (obj_list)
 498                 list_add_tail(&entry->obj_list, obj_list);
 499         spin_unlock_irqrestore(&file->async_file->lock, flags);
 500
 501         wake_up_interruptible(&file->async_file->poll_wait);
 502         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 503 }
 504
 505 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 506 {
 507         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 508                                                   struct ib_ucq_object, uobject);
 509
 510         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 511                                 event->event, &uobj->async_list,
 512                                 &uobj->async_events_reported);
 513 }
 514
 515 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 516 {
 517         struct ib_uevent_object *uobj;
 518
 519         /* for XRC target qp's, check that qp is live */
 520         if (!event->element.qp->uobject || !event->element.qp->uobject->live)
 521                 return;
 522
 523         uobj = container_of(event->element.qp->uobject,
 524                             struct ib_uevent_object, uobject);
 525
 526         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 527                                 event->event, &uobj->event_list,
 528                                 &uobj->events_reported);
 529 }
 530
 531 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 532 {
 533         struct ib_uevent_object *uobj;
 534
 535         uobj = container_of(event->element.srq->uobject,
 536                             struct ib_uevent_object, uobject);
 537
 538         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 539                                 event->event, &uobj->event_list,
 540                                 &uobj->events_reported);
 541 }
 542
 543 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 544                              struct ib_event *event)
 545 {
 546         struct ib_uverbs_file *file =
 547                 container_of(handler, struct ib_uverbs_file, event_handler);
 548
 549         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 550                                 NULL, NULL);
 551 }
 552
 553 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 554                                         int is_async)
 555 {
 556         struct ib_uverbs_event_file *ev_file;
 557         struct file *filp;
 558
 559         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
 560         if (!ev_file)
 561                 return ERR_PTR(-ENOMEM);
 562
 563         kref_init(&ev_file->ref);
 564         spin_lock_init(&ev_file->lock);
 565         INIT_LIST_HEAD(&ev_file->event_list);
 566         init_waitqueue_head(&ev_file->poll_wait);
 567         ev_file->uverbs_file = uverbs_file;
 568         ev_file->async_queue = NULL;
 569         ev_file->is_async    = is_async;
 570         ev_file->is_closed   = 0;
 571
 572         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 573                                   ev_file, O_RDONLY);
 574         if (IS_ERR(filp))
 575                 kfree(ev_file);
 576
 577         return filp;
 578 }
 579
 580 /*
 581  * Look up a completion event file by FD.  If lookup is successful,
 582  * takes a ref to the event file struct that it returns; if
 583  * unsuccessful, returns NULL.
 584  */
 585 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 586 {
 587         struct ib_uverbs_event_file *ev_file = NULL;
 588         struct fd f = fdget(fd);
 589
 590         if (!f.file)
 591                 return NULL;
 592
 593         if (f.file->f_op != &uverbs_event_fops)
 594                 goto out;
 595
 596         ev_file = f.file->private_data;
 597         if (ev_file->is_async) {
 598                 ev_file = NULL;
 599                 goto out;
 600         }
 601
 602         kref_get(&ev_file->ref);
 603
 604 out:
 605         fdput(f);
 606         return ev_file;
 607 }
 608
 609 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 610                              size_t count, loff_t *pos)
 611 {
 612         struct ib_uverbs_file *file = filp->private_data;
 613         struct ib_uverbs_cmd_hdr hdr;
 614         __u32 flags;
 615
 616         if (count < sizeof hdr)
 617                 return -EINVAL;
 618
 619         if (copy_from_user(&hdr, buf, sizeof hdr))
 620                 return -EFAULT;
 621
 622         flags = (hdr.command &
 623                  IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
 624
 625         if (!flags) {
 626                 __u32 command;
 627
 628                 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 629                                            IB_USER_VERBS_CMD_COMMAND_MASK))
 630                         return -EINVAL;
 631
 632                 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 633
 634                 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
 635                     !uverbs_cmd_table[command])
 636                         return -EINVAL;
 637
 638                 if (!file->ucontext &&
 639                     command != IB_USER_VERBS_CMD_GET_CONTEXT)
 640                         return -EINVAL;
 641
 642                 if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << command)))
 643                         return -ENOSYS;
 644
 645                 if (hdr.in_words * 4 != count)
 646                         return -EINVAL;
 647
 648                 return uverbs_cmd_table[command](file,
 649                                                  buf + sizeof(hdr),
 650                                                  hdr.in_words * 4,
 651                                                  hdr.out_words * 4);
 652
 653         } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
 654                 __u32 command;
 655
 656                 struct ib_uverbs_ex_cmd_hdr ex_hdr;
 657                 struct ib_udata ucore;
 658                 struct ib_udata uhw;
 659                 int err;
 660                 size_t written_count = count;
 661
 662                 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 663                                            IB_USER_VERBS_CMD_COMMAND_MASK))
 664                         return -EINVAL;
 665
 666                 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 667
 668                 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
 669                     !uverbs_ex_cmd_table[command])
 670                         return -ENOSYS;
 671
 672                 if (!file->ucontext)
 673                         return -EINVAL;
 674
 675                 if (!(file->device->ib_dev->uverbs_ex_cmd_mask & (1ull << command)))
 676                         return -ENOSYS;
 677
 678                 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 679                         return -EINVAL;
 680
 681                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
 682                         return -EFAULT;
 683
 684                 count -= sizeof(hdr) + sizeof(ex_hdr);
 685                 buf += sizeof(hdr) + sizeof(ex_hdr);
 686
 687                 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count)
 688                         return -EINVAL;
 689
 690                 if (ex_hdr.cmd_hdr_reserved)
 691                         return -EINVAL;
 692
 693                 if (ex_hdr.response) {
 694                         if (!hdr.out_words && !ex_hdr.provider_out_words)
 695                                 return -EINVAL;
 696
 697                         if (!access_ok(VERIFY_WRITE,
 698                                        (void __user *) (unsigned long) ex_hdr.response,
 699                                        (hdr.out_words + ex_hdr.provider_out_words) * 8))
 700                                 return -EFAULT;
 701                 } else {
 702                         if (hdr.out_words || ex_hdr.provider_out_words)
 703                                 return -EINVAL;
 704                 }
 705
 706                 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
 707                                        hdr.in_words * 8, hdr.out_words * 8);
 708
 709                 INIT_UDATA_BUF_OR_NULL(&uhw,
 710                                        buf + ucore.inlen,
 711                                        (unsigned long) ex_hdr.response + ucore.outlen,
 712                                        ex_hdr.provider_in_words * 8,
 713                                        ex_hdr.provider_out_words * 8);
 714
 715                 err = uverbs_ex_cmd_table[command](file,
 716                                                    &ucore,
 717                                                    &uhw);
 718
 719                 if (err)
 720                         return err;
 721
 722                 return written_count;
 723         }
 724
 725         return -ENOSYS;
 726 }
 727
 728 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 729 {
 730         struct ib_uverbs_file *file = filp->private_data;
 731
 732         if (!file->ucontext)
 733                 return -ENODEV;
 734         else
 735                 return file->device->ib_dev->mmap(file->ucontext, vma);
 736 }
 737
 738 /*
 739  * ib_uverbs_open() does not need the BKL:
 740  *
 741  *  - the ib_uverbs_device structures are properly reference counted and
 742  *    everything else is purely local to the file being created, so
 743  *    races against other open calls are not a problem;
 744  *  - there is no ioctl method to race against;
 745  *  - the open method will either immediately run -ENXIO, or all
 746  *    required initialization will be done.
 747  */
 748 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 749 {
 750         struct ib_uverbs_device *dev;
 751         struct ib_uverbs_file *file;
 752         int ret;
 753
 754         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 755         if (!atomic_inc_not_zero(&dev->refcount))
 756                 return -ENXIO;
 757
 758         if (!try_module_get(dev->ib_dev->owner)) {
 759                 ret = -ENODEV;
 760                 goto err;
 761         }
 762
 763         file = kmalloc(sizeof *file, GFP_KERNEL);
 764         if (!file) {
 765                 ret = -ENOMEM;
 766                 goto err_module;
 767         }
 768
 769         file->device     = dev;
 770         file->ucontext   = NULL;
 771         file->async_file = NULL;
 772         kref_init(&file->ref);
 773         mutex_init(&file->mutex);
 774
 775         filp->private_data = file;
 776         kobject_get(&dev->kobj);
 777
 778         return nonseekable_open(inode, filp);
 779
 780 err_module:
 781         module_put(dev->ib_dev->owner);
 782
 783 err:
 784         if (atomic_dec_and_test(&dev->refcount))
 785                 ib_uverbs_comp_dev(dev);
 786
 787         return ret;
 788 }
 789
 790 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 791 {
 792         struct ib_uverbs_file *file = filp->private_data;
 793         struct ib_uverbs_device *dev = file->device;
 794
 795         ib_uverbs_cleanup_ucontext(file, file->ucontext);
 796
 797         if (file->async_file)
 798                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 799
 800         kref_put(&file->ref, ib_uverbs_release_file);
 801         kobject_put(&dev->kobj);
 802
 803         return 0;
 804 }
 805
 806 static const struct file_operations uverbs_fops = {
 807         .owner   = THIS_MODULE,
 808         .write   = ib_uverbs_write,
 809         .open    = ib_uverbs_open,
 810         .release = ib_uverbs_close,
 811         .llseek  = no_llseek,
 812 };
 813
 814 static const struct file_operations uverbs_mmap_fops = {
 815         .owner   = THIS_MODULE,
 816         .write   = ib_uverbs_write,
 817         .mmap    = ib_uverbs_mmap,
 818         .open    = ib_uverbs_open,
 819         .release = ib_uverbs_close,
 820         .llseek  = no_llseek,
 821 };
 822
 823 static struct ib_client uverbs_client = {
 824         .name   = "uverbs",
 825         .add    = ib_uverbs_add_one,
 826         .remove = ib_uverbs_remove_one
 827 };
 828
 829 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 830                           char *buf)
 831 {
 832         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 833
 834         if (!dev)
 835                 return -ENODEV;
 836
 837         return sprintf(buf, "%s\n", dev->ib_dev->name);
 838 }
 839 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 840
 841 static ssize_t show_dev_abi_version(struct device *device,
 842                                     struct device_attribute *attr, char *buf)
 843 {
 844         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 845
 846         if (!dev)
 847                 return -ENODEV;
 848
 849         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
 850 }
 851 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 852
 853 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
 854                          __stringify(IB_USER_VERBS_ABI_VERSION));
 855
 856 static dev_t overflow_maj;
 857 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
 858
 859 /*
 860  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
 861  * requesting a new major number and doubling the number of max devices we
 862  * support. It's stupid, but simple.
 863  */
 864 static int find_overflow_devnum(void)
 865 {
 866         int ret;
 867
 868         if (!overflow_maj) {
 869                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
 870                                           "infiniband_verbs");
 871                 if (ret) {
 872                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
 873                         return ret;
 874                 }
 875         }
 876
 877         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
 878         if (ret >= IB_UVERBS_MAX_DEVICES)
 879                 return -1;
 880
 881         return ret;
 882 }
 883
 884 static void ib_uverbs_add_one(struct ib_device *device)
 885 {
 886         int devnum;
 887         dev_t base;
 888         struct ib_uverbs_device *uverbs_dev;
 889
 890         if (!device->alloc_ucontext)
 891                 return;
 892
 893         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
 894         if (!uverbs_dev)
 895                 return;
 896
 897         atomic_set(&uverbs_dev->refcount, 1);
 898         init_completion(&uverbs_dev->comp);
 899         uverbs_dev->xrcd_tree = RB_ROOT;
 900         mutex_init(&uverbs_dev->xrcd_tree_mutex);
 901         kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
 902
 903         spin_lock(&map_lock);
 904         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
 905         if (devnum >= IB_UVERBS_MAX_DEVICES) {
 906                 spin_unlock(&map_lock);
 907                 devnum = find_overflow_devnum();
 908                 if (devnum < 0)
 909                         goto err;
 910
 911                 spin_lock(&map_lock);
 912                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
 913                 base = devnum + overflow_maj;
 914                 set_bit(devnum, overflow_map);
 915         } else {
 916                 uverbs_dev->devnum = devnum;
 917                 base = devnum + IB_UVERBS_BASE_DEV;
 918                 set_bit(devnum, dev_map);
 919         }
 920         spin_unlock(&map_lock);
 921
 922         uverbs_dev->ib_dev           = device;
 923         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
 924
 925         cdev_init(&uverbs_dev->cdev, NULL);
 926         uverbs_dev->cdev.owner = THIS_MODULE;
 927         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
 928         uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
 929         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
 930         if (cdev_add(&uverbs_dev->cdev, base, 1))
 931                 goto err_cdev;
 932
 933         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
 934                                         uverbs_dev->cdev.dev, uverbs_dev,
 935                                         "uverbs%d", uverbs_dev->devnum);
 936         if (IS_ERR(uverbs_dev->dev))
 937                 goto err_cdev;
 938
 939         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
 940                 goto err_class;
 941         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
 942                 goto err_class;
 943
 944         ib_set_client_data(device, &uverbs_client, uverbs_dev);
 945
 946         return;
 947
 948 err_class:
 949         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 950
 951 err_cdev:
 952         cdev_del(&uverbs_dev->cdev);
 953         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 954                 clear_bit(devnum, dev_map);
 955         else
 956                 clear_bit(devnum, overflow_map);
 957
 958 err:
 959         if (atomic_dec_and_test(&uverbs_dev->refcount))
 960                 ib_uverbs_comp_dev(uverbs_dev);
 961         wait_for_completion(&uverbs_dev->comp);
 962         kobject_put(&uverbs_dev->kobj);
 963         return;
 964 }
 965
 966 static void ib_uverbs_remove_one(struct ib_device *device)
 967 {
 968         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
 969
 970         if (!uverbs_dev)
 971                 return;
 972
 973         dev_set_drvdata(uverbs_dev->dev, NULL);
 974         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
 975         cdev_del(&uverbs_dev->cdev);
 976
 977         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
 978                 clear_bit(uverbs_dev->devnum, dev_map);
 979         else
 980                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
 981
 982         if (atomic_dec_and_test(&uverbs_dev->refcount))
 983                 ib_uverbs_comp_dev(uverbs_dev);
 984         wait_for_completion(&uverbs_dev->comp);
 985         kobject_put(&uverbs_dev->kobj);
 986 }
 987
 988 static char *uverbs_devnode(struct device *dev, umode_t *mode)
 989 {
 990         if (mode)
 991                 *mode = 0666;
 992         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
 993 }
 994
 995 static int __init ib_uverbs_init(void)
 996 {
 997         int ret;
 998
 999         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1000                                      "infiniband_verbs");
1001         if (ret) {
1002                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
1003                 goto out;
1004         }
1005
1006         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1007         if (IS_ERR(uverbs_class)) {
1008                 ret = PTR_ERR(uverbs_class);
1009                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
1010                 goto out_chrdev;
1011         }
1012
1013         uverbs_class->devnode = uverbs_devnode;
1014
1015         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1016         if (ret) {
1017                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
1018                 goto out_class;
1019         }
1020
1021         ret = ib_register_client(&uverbs_client);
1022         if (ret) {
1023                 printk(KERN_ERR "user_verbs: couldn't register client\n");
1024                 goto out_class;
1025         }
1026
1027         return 0;
1028
1029 out_class:
1030         class_destroy(uverbs_class);
1031
1032 out_chrdev:
1033         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1034
1035 out:
1036         return ret;
1037 }
1038
1039 static void __exit ib_uverbs_cleanup(void)
1040 {
1041         ib_unregister_client(&uverbs_client);
1042         class_destroy(uverbs_class);
1043         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1044         if (overflow_maj)
1045                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1046         idr_destroy(&ib_uverbs_pd_idr);
1047         idr_destroy(&ib_uverbs_mr_idr);
1048         idr_destroy(&ib_uverbs_mw_idr);
1049         idr_destroy(&ib_uverbs_ah_idr);
1050         idr_destroy(&ib_uverbs_cq_idr);
1051         idr_destroy(&ib_uverbs_qp_idr);
1052         idr_destroy(&ib_uverbs_srq_idr);
1053 }
1054
1055 module_init(ib_uverbs_init);
1056 module_exit(ib_uverbs_cleanup);