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 <linux/uaccess.h>
  50
  51 #include <rdma/ib.h>
  52
  53 #include "uverbs.h"
  54 #include "core_priv.h"
  55
  56 MODULE_AUTHOR("Roland Dreier");
  57 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  58 MODULE_LICENSE("Dual BSD/GPL");
  59
  60 enum {
  61         IB_UVERBS_MAJOR       = 231,
  62         IB_UVERBS_BASE_MINOR  = 192,
  63         IB_UVERBS_MAX_DEVICES = 32
  64 };
  65
  66 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  67
  68 static struct class *uverbs_class;
  69
  70 static DEFINE_SPINLOCK(map_lock);
  71 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  72
  73 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  74                                      struct ib_device *ib_dev,
  75                                      const char __user *buf, int in_len,
  76                                      int out_len) = {
  77         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  78         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  79         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  80         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  81         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  82         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  83         [IB_USER_VERBS_CMD_REREG_MR]            = ib_uverbs_rereg_mr,
  84         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  85         [IB_USER_VERBS_CMD_ALLOC_MW]            = ib_uverbs_alloc_mw,
  86         [IB_USER_VERBS_CMD_DEALLOC_MW]          = ib_uverbs_dealloc_mw,
  87         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  88         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  89         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  90         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  91         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  92         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  93         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
  94         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
  95         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
  96         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
  97         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
  98         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
  99         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 100         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 101         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 102         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 103         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 104         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 105         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 106         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 107         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 108         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 109         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 110         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 111         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp,
 112 };
 113
 114 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
 115                                     struct ib_device *ib_dev,
 116                                     struct ib_udata *ucore,
 117                                     struct ib_udata *uhw) = {
 118         [IB_USER_VERBS_EX_CMD_CREATE_FLOW]      = ib_uverbs_ex_create_flow,
 119         [IB_USER_VERBS_EX_CMD_DESTROY_FLOW]     = ib_uverbs_ex_destroy_flow,
 120         [IB_USER_VERBS_EX_CMD_QUERY_DEVICE]     = ib_uverbs_ex_query_device,
 121         [IB_USER_VERBS_EX_CMD_CREATE_CQ]        = ib_uverbs_ex_create_cq,
 122         [IB_USER_VERBS_EX_CMD_CREATE_QP]        = ib_uverbs_ex_create_qp,
 123         [IB_USER_VERBS_EX_CMD_CREATE_WQ]        = ib_uverbs_ex_create_wq,
 124         [IB_USER_VERBS_EX_CMD_MODIFY_WQ]        = ib_uverbs_ex_modify_wq,
 125         [IB_USER_VERBS_EX_CMD_DESTROY_WQ]       = ib_uverbs_ex_destroy_wq,
 126         [IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL] = ib_uverbs_ex_create_rwq_ind_table,
 127         [IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL] = ib_uverbs_ex_destroy_rwq_ind_table,
 128         [IB_USER_VERBS_EX_CMD_MODIFY_QP]        = ib_uverbs_ex_modify_qp,
 129 };
 130
 131 static void ib_uverbs_add_one(struct ib_device *device);
 132 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
 133
 134 int uverbs_dealloc_mw(struct ib_mw *mw)
 135 {
 136         struct ib_pd *pd = mw->pd;
 137         int ret;
 138
 139         ret = mw->device->dealloc_mw(mw);
 140         if (!ret)
 141                 atomic_dec(&pd->usecnt);
 142         return ret;
 143 }
 144
 145 static void ib_uverbs_release_dev(struct kobject *kobj)
 146 {
 147         struct ib_uverbs_device *dev =
 148                 container_of(kobj, struct ib_uverbs_device, kobj);
 149
 150         cleanup_srcu_struct(&dev->disassociate_srcu);
 151         kfree(dev);
 152 }
 153
 154 static struct kobj_type ib_uverbs_dev_ktype = {
 155         .release = ib_uverbs_release_dev,
 156 };
 157
 158 static void ib_uverbs_release_event_file(struct kref *ref)
 159 {
 160         struct ib_uverbs_event_file *file =
 161                 container_of(ref, struct ib_uverbs_event_file, ref);
 162
 163         kfree(file);
 164 }
 165
 166 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 167                           struct ib_uverbs_event_file *ev_file,
 168                           struct ib_ucq_object *uobj)
 169 {
 170         struct ib_uverbs_event *evt, *tmp;
 171
 172         if (ev_file) {
 173                 spin_lock_irq(&ev_file->lock);
 174                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 175                         list_del(&evt->list);
 176                         kfree(evt);
 177                 }
 178                 spin_unlock_irq(&ev_file->lock);
 179
 180                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 181         }
 182
 183         spin_lock_irq(&file->async_file->lock);
 184         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 185                 list_del(&evt->list);
 186                 kfree(evt);
 187         }
 188         spin_unlock_irq(&file->async_file->lock);
 189 }
 190
 191 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 192                               struct ib_uevent_object *uobj)
 193 {
 194         struct ib_uverbs_event *evt, *tmp;
 195
 196         spin_lock_irq(&file->async_file->lock);
 197         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 198                 list_del(&evt->list);
 199                 kfree(evt);
 200         }
 201         spin_unlock_irq(&file->async_file->lock);
 202 }
 203
 204 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
 205                                     struct ib_uqp_object *uobj)
 206 {
 207         struct ib_uverbs_mcast_entry *mcast, *tmp;
 208
 209         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 210                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 211                 list_del(&mcast->list);
 212                 kfree(mcast);
 213         }
 214 }
 215
 216 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 217                                       struct ib_ucontext *context)
 218 {
 219         struct ib_uobject *uobj, *tmp;
 220
 221         context->closing = 1;
 222
 223         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
 224                 struct ib_ah *ah = uobj->object;
 225
 226                 idr_remove_uobj(uobj);
 227                 ib_destroy_ah(ah);
 228                 ib_rdmacg_uncharge(&uobj->cg_obj, context->device,
 229                                    RDMACG_RESOURCE_HCA_OBJECT);
 230                 kfree(uobj);
 231         }
 232
 233         /* Remove MWs before QPs, in order to support type 2A MWs. */
 234         list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
 235                 struct ib_mw *mw = uobj->object;
 236
 237                 idr_remove_uobj(uobj);
 238                 uverbs_dealloc_mw(mw);
 239                 ib_rdmacg_uncharge(&uobj->cg_obj, context->device,
 240                                    RDMACG_RESOURCE_HCA_OBJECT);
 241                 kfree(uobj);
 242         }
 243
 244         list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
 245                 struct ib_flow *flow_id = uobj->object;
 246
 247                 idr_remove_uobj(uobj);
 248                 ib_destroy_flow(flow_id);
 249                 ib_rdmacg_uncharge(&uobj->cg_obj, context->device,
 250                                    RDMACG_RESOURCE_HCA_OBJECT);
 251                 kfree(uobj);
 252         }
 253
 254         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
 255                 struct ib_qp *qp = uobj->object;
 256                 struct ib_uqp_object *uqp =
 257                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
 258
 259                 idr_remove_uobj(uobj);
 260                 if (qp == qp->real_qp)
 261                         ib_uverbs_detach_umcast(qp, uqp);
 262                 ib_destroy_qp(qp);
 263                 ib_rdmacg_uncharge(&uobj->cg_obj, context->device,
 264                                    RDMACG_RESOURCE_HCA_OBJECT);
 265                 ib_uverbs_release_uevent(file, &uqp->uevent);
 266                 kfree(uqp);
 267         }
 268
 269         list_for_each_entry_safe(uobj, tmp, &context->rwq_ind_tbl_list, list) {
 270                 struct ib_rwq_ind_table *rwq_ind_tbl = uobj->object;
 271                 struct ib_wq **ind_tbl = rwq_ind_tbl->ind_tbl;
 272
 273                 idr_remove_uobj(uobj);
 274                 ib_destroy_rwq_ind_table(rwq_ind_tbl);
 275                 kfree(ind_tbl);
 276                 kfree(uobj);
 277         }
 278
 279         list_for_each_entry_safe(uobj, tmp, &context->wq_list, list) {
 280                 struct ib_wq *wq = uobj->object;
 281                 struct ib_uwq_object *uwq =
 282                         container_of(uobj, struct ib_uwq_object, uevent.uobject);
 283
 284                 idr_remove_uobj(uobj);
 285                 ib_destroy_wq(wq);
 286                 ib_uverbs_release_uevent(file, &uwq->uevent);
 287                 kfree(uwq);
 288         }
 289
 290         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
 291                 struct ib_srq *srq = uobj->object;
 292                 struct ib_uevent_object *uevent =
 293                         container_of(uobj, struct ib_uevent_object, uobject);
 294
 295                 idr_remove_uobj(uobj);
 296                 ib_destroy_srq(srq);
 297                 ib_rdmacg_uncharge(&uobj->cg_obj, context->device,
 298                                    RDMACG_RESOURCE_HCA_OBJECT);
 299                 ib_uverbs_release_uevent(file, uevent);
 300                 kfree(uevent);
 301         }
 302
 303         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
 304                 struct ib_cq *cq = uobj->object;
 305                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
 306                 struct ib_ucq_object *ucq =
 307                         container_of(uobj, struct ib_ucq_object, uobject);
 308
 309                 idr_remove_uobj(uobj);
 310                 ib_destroy_cq(cq);
 311                 ib_rdmacg_uncharge(&uobj->cg_obj, context->device,
 312                                    RDMACG_RESOURCE_HCA_OBJECT);
 313                 ib_uverbs_release_ucq(file, ev_file, ucq);
 314                 kfree(ucq);
 315         }
 316
 317         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
 318                 struct ib_mr *mr = uobj->object;
 319
 320                 idr_remove_uobj(uobj);
 321                 ib_dereg_mr(mr);
 322                 ib_rdmacg_uncharge(&uobj->cg_obj, context->device,
 323                                    RDMACG_RESOURCE_HCA_OBJECT);
 324                 kfree(uobj);
 325         }
 326
 327         mutex_lock(&file->device->xrcd_tree_mutex);
 328         list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
 329                 struct ib_xrcd *xrcd = uobj->object;
 330                 struct ib_uxrcd_object *uxrcd =
 331                         container_of(uobj, struct ib_uxrcd_object, uobject);
 332
 333                 idr_remove_uobj(uobj);
 334                 ib_uverbs_dealloc_xrcd(file->device, xrcd);
 335                 kfree(uxrcd);
 336         }
 337         mutex_unlock(&file->device->xrcd_tree_mutex);
 338
 339         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
 340                 struct ib_pd *pd = uobj->object;
 341
 342                 idr_remove_uobj(uobj);
 343                 ib_dealloc_pd(pd);
 344                 ib_rdmacg_uncharge(&uobj->cg_obj, context->device,
 345                                    RDMACG_RESOURCE_HCA_OBJECT);
 346                 kfree(uobj);
 347         }
 348
 349         put_pid(context->tgid);
 350
 351         ib_rdmacg_uncharge(&context->cg_obj, context->device,
 352                            RDMACG_RESOURCE_HCA_HANDLE);
 353
 354         return context->device->dealloc_ucontext(context);
 355 }
 356
 357 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
 358 {
 359         complete(&dev->comp);
 360 }
 361
 362 static void ib_uverbs_release_file(struct kref *ref)
 363 {
 364         struct ib_uverbs_file *file =
 365                 container_of(ref, struct ib_uverbs_file, ref);
 366         struct ib_device *ib_dev;
 367         int srcu_key;
 368
 369         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 370         ib_dev = srcu_dereference(file->device->ib_dev,
 371                                   &file->device->disassociate_srcu);
 372         if (ib_dev && !ib_dev->disassociate_ucontext)
 373                 module_put(ib_dev->owner);
 374         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 375
 376         if (atomic_dec_and_test(&file->device->refcount))
 377                 ib_uverbs_comp_dev(file->device);
 378
 379         kfree(file);
 380 }
 381
 382 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
 383                                     size_t count, loff_t *pos)
 384 {
 385         struct ib_uverbs_event_file *file = filp->private_data;
 386         struct ib_uverbs_event *event;
 387         int eventsz;
 388         int ret = 0;
 389
 390         spin_lock_irq(&file->lock);
 391
 392         while (list_empty(&file->event_list)) {
 393                 spin_unlock_irq(&file->lock);
 394
 395                 if (filp->f_flags & O_NONBLOCK)
 396                         return -EAGAIN;
 397
 398                 if (wait_event_interruptible(file->poll_wait,
 399                                              (!list_empty(&file->event_list) ||
 400                         /* The barriers built into wait_event_interruptible()
 401                          * and wake_up() guarentee this will see the null set
 402                          * without using RCU
 403                          */
 404                                              !file->uverbs_file->device->ib_dev)))
 405                         return -ERESTARTSYS;
 406
 407                 /* If device was disassociated and no event exists set an error */
 408                 if (list_empty(&file->event_list) &&
 409                     !file->uverbs_file->device->ib_dev)
 410                         return -EIO;
 411
 412                 spin_lock_irq(&file->lock);
 413         }
 414
 415         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
 416
 417         if (file->is_async)
 418                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
 419         else
 420                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
 421
 422         if (eventsz > count) {
 423                 ret   = -EINVAL;
 424                 event = NULL;
 425         } else {
 426                 list_del(file->event_list.next);
 427                 if (event->counter) {
 428                         ++(*event->counter);
 429                         list_del(&event->obj_list);
 430                 }
 431         }
 432
 433         spin_unlock_irq(&file->lock);
 434
 435         if (event) {
 436                 if (copy_to_user(buf, event, eventsz))
 437                         ret = -EFAULT;
 438                 else
 439                         ret = eventsz;
 440         }
 441
 442         kfree(event);
 443
 444         return ret;
 445 }
 446
 447 static unsigned int ib_uverbs_event_poll(struct file *filp,
 448                                          struct poll_table_struct *wait)
 449 {
 450         unsigned int pollflags = 0;
 451         struct ib_uverbs_event_file *file = filp->private_data;
 452
 453         poll_wait(filp, &file->poll_wait, wait);
 454
 455         spin_lock_irq(&file->lock);
 456         if (!list_empty(&file->event_list))
 457                 pollflags = POLLIN | POLLRDNORM;
 458         spin_unlock_irq(&file->lock);
 459
 460         return pollflags;
 461 }
 462
 463 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
 464 {
 465         struct ib_uverbs_event_file *file = filp->private_data;
 466
 467         return fasync_helper(fd, filp, on, &file->async_queue);
 468 }
 469
 470 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
 471 {
 472         struct ib_uverbs_event_file *file = filp->private_data;
 473         struct ib_uverbs_event *entry, *tmp;
 474         int closed_already = 0;
 475
 476         mutex_lock(&file->uverbs_file->device->lists_mutex);
 477         spin_lock_irq(&file->lock);
 478         closed_already = file->is_closed;
 479         file->is_closed = 1;
 480         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
 481                 if (entry->counter)
 482                         list_del(&entry->obj_list);
 483                 kfree(entry);
 484         }
 485         spin_unlock_irq(&file->lock);
 486         if (!closed_already) {
 487                 list_del(&file->list);
 488                 if (file->is_async)
 489                         ib_unregister_event_handler(&file->uverbs_file->
 490                                 event_handler);
 491         }
 492         mutex_unlock(&file->uverbs_file->device->lists_mutex);
 493
 494         kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
 495         kref_put(&file->ref, ib_uverbs_release_event_file);
 496
 497         return 0;
 498 }
 499
 500 static const struct file_operations uverbs_event_fops = {
 501         .owner   = THIS_MODULE,
 502         .read    = ib_uverbs_event_read,
 503         .poll    = ib_uverbs_event_poll,
 504         .release = ib_uverbs_event_close,
 505         .fasync  = ib_uverbs_event_fasync,
 506         .llseek  = no_llseek,
 507 };
 508
 509 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 510 {
 511         struct ib_uverbs_event_file    *file = cq_context;
 512         struct ib_ucq_object           *uobj;
 513         struct ib_uverbs_event         *entry;
 514         unsigned long                   flags;
 515
 516         if (!file)
 517                 return;
 518
 519         spin_lock_irqsave(&file->lock, flags);
 520         if (file->is_closed) {
 521                 spin_unlock_irqrestore(&file->lock, flags);
 522                 return;
 523         }
 524
 525         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 526         if (!entry) {
 527                 spin_unlock_irqrestore(&file->lock, flags);
 528                 return;
 529         }
 530
 531         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 532
 533         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 534         entry->counter             = &uobj->comp_events_reported;
 535
 536         list_add_tail(&entry->list, &file->event_list);
 537         list_add_tail(&entry->obj_list, &uobj->comp_list);
 538         spin_unlock_irqrestore(&file->lock, flags);
 539
 540         wake_up_interruptible(&file->poll_wait);
 541         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
 542 }
 543
 544 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 545                                     __u64 element, __u64 event,
 546                                     struct list_head *obj_list,
 547                                     u32 *counter)
 548 {
 549         struct ib_uverbs_event *entry;
 550         unsigned long flags;
 551
 552         spin_lock_irqsave(&file->async_file->lock, flags);
 553         if (file->async_file->is_closed) {
 554                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 555                 return;
 556         }
 557
 558         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
 559         if (!entry) {
 560                 spin_unlock_irqrestore(&file->async_file->lock, flags);
 561                 return;
 562         }
 563
 564         entry->desc.async.element    = element;
 565         entry->desc.async.event_type = event;
 566         entry->desc.async.reserved   = 0;
 567         entry->counter               = counter;
 568
 569         list_add_tail(&entry->list, &file->async_file->event_list);
 570         if (obj_list)
 571                 list_add_tail(&entry->obj_list, obj_list);
 572         spin_unlock_irqrestore(&file->async_file->lock, flags);
 573
 574         wake_up_interruptible(&file->async_file->poll_wait);
 575         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
 576 }
 577
 578 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 579 {
 580         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 581                                                   struct ib_ucq_object, uobject);
 582
 583         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 584                                 event->event, &uobj->async_list,
 585                                 &uobj->async_events_reported);
 586 }
 587
 588 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 589 {
 590         struct ib_uevent_object *uobj;
 591
 592         /* for XRC target qp's, check that qp is live */
 593         if (!event->element.qp->uobject || !event->element.qp->uobject->live)
 594                 return;
 595
 596         uobj = container_of(event->element.qp->uobject,
 597                             struct ib_uevent_object, uobject);
 598
 599         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 600                                 event->event, &uobj->event_list,
 601                                 &uobj->events_reported);
 602 }
 603
 604 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
 605 {
 606         struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
 607                                                   struct ib_uevent_object, uobject);
 608
 609         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 610                                 event->event, &uobj->event_list,
 611                                 &uobj->events_reported);
 612 }
 613
 614 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 615 {
 616         struct ib_uevent_object *uobj;
 617
 618         uobj = container_of(event->element.srq->uobject,
 619                             struct ib_uevent_object, uobject);
 620
 621         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 622                                 event->event, &uobj->event_list,
 623                                 &uobj->events_reported);
 624 }
 625
 626 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 627                              struct ib_event *event)
 628 {
 629         struct ib_uverbs_file *file =
 630                 container_of(handler, struct ib_uverbs_file, event_handler);
 631
 632         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 633                                 NULL, NULL);
 634 }
 635
 636 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
 637 {
 638         kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
 639         file->async_file = NULL;
 640 }
 641
 642 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
 643                                         struct ib_device        *ib_dev,
 644                                         int is_async)
 645 {
 646         struct ib_uverbs_event_file *ev_file;
 647         struct file *filp;
 648         int ret;
 649
 650         ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
 651         if (!ev_file)
 652                 return ERR_PTR(-ENOMEM);
 653
 654         kref_init(&ev_file->ref);
 655         spin_lock_init(&ev_file->lock);
 656         INIT_LIST_HEAD(&ev_file->event_list);
 657         init_waitqueue_head(&ev_file->poll_wait);
 658         ev_file->uverbs_file = uverbs_file;
 659         kref_get(&ev_file->uverbs_file->ref);
 660         ev_file->async_queue = NULL;
 661         ev_file->is_closed   = 0;
 662
 663         filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
 664                                   ev_file, O_RDONLY);
 665         if (IS_ERR(filp))
 666                 goto err_put_refs;
 667
 668         mutex_lock(&uverbs_file->device->lists_mutex);
 669         list_add_tail(&ev_file->list,
 670                       &uverbs_file->device->uverbs_events_file_list);
 671         mutex_unlock(&uverbs_file->device->lists_mutex);
 672
 673         if (is_async) {
 674                 WARN_ON(uverbs_file->async_file);
 675                 uverbs_file->async_file = ev_file;
 676                 kref_get(&uverbs_file->async_file->ref);
 677                 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
 678                                       ib_dev,
 679                                       ib_uverbs_event_handler);
 680                 ret = ib_register_event_handler(&uverbs_file->event_handler);
 681                 if (ret)
 682                         goto err_put_file;
 683
 684                 /* At that point async file stuff was fully set */
 685                 ev_file->is_async = 1;
 686         }
 687
 688         return filp;
 689
 690 err_put_file:
 691         fput(filp);
 692         kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
 693         uverbs_file->async_file = NULL;
 694         return ERR_PTR(ret);
 695
 696 err_put_refs:
 697         kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
 698         kref_put(&ev_file->ref, ib_uverbs_release_event_file);
 699         return filp;
 700 }
 701
 702 /*
 703  * Look up a completion event file by FD.  If lookup is successful,
 704  * takes a ref to the event file struct that it returns; if
 705  * unsuccessful, returns NULL.
 706  */
 707 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
 708 {
 709         struct ib_uverbs_event_file *ev_file = NULL;
 710         struct fd f = fdget(fd);
 711
 712         if (!f.file)
 713                 return NULL;
 714
 715         if (f.file->f_op != &uverbs_event_fops)
 716                 goto out;
 717
 718         ev_file = f.file->private_data;
 719         if (ev_file->is_async) {
 720                 ev_file = NULL;
 721                 goto out;
 722         }
 723
 724         kref_get(&ev_file->ref);
 725
 726 out:
 727         fdput(f);
 728         return ev_file;
 729 }
 730
 731 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
 732 {
 733         u64 mask;
 734
 735         if (command <= IB_USER_VERBS_CMD_OPEN_QP)
 736                 mask = ib_dev->uverbs_cmd_mask;
 737         else
 738                 mask = ib_dev->uverbs_ex_cmd_mask;
 739
 740         if (mask & ((u64)1 << command))
 741                 return 0;
 742
 743         return -1;
 744 }
 745
 746 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 747                              size_t count, loff_t *pos)
 748 {
 749         struct ib_uverbs_file *file = filp->private_data;
 750         struct ib_device *ib_dev;
 751         struct ib_uverbs_cmd_hdr hdr;
 752         __u32 command;
 753         __u32 flags;
 754         int srcu_key;
 755         ssize_t ret;
 756
 757         if (!ib_safe_file_access(filp)) {
 758                 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
 759                             task_tgid_vnr(current), current->comm);
 760                 return -EACCES;
 761         }
 762
 763         if (count < sizeof hdr)
 764                 return -EINVAL;
 765
 766         if (copy_from_user(&hdr, buf, sizeof hdr))
 767                 return -EFAULT;
 768
 769         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 770         ib_dev = srcu_dereference(file->device->ib_dev,
 771                                   &file->device->disassociate_srcu);
 772         if (!ib_dev) {
 773                 ret = -EIO;
 774                 goto out;
 775         }
 776
 777         if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
 778                                    IB_USER_VERBS_CMD_COMMAND_MASK)) {
 779                 ret = -EINVAL;
 780                 goto out;
 781         }
 782
 783         command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
 784         if (verify_command_mask(ib_dev, command)) {
 785                 ret = -EOPNOTSUPP;
 786                 goto out;
 787         }
 788
 789         if (!file->ucontext &&
 790             command != IB_USER_VERBS_CMD_GET_CONTEXT) {
 791                 ret = -EINVAL;
 792                 goto out;
 793         }
 794
 795         flags = (hdr.command &
 796                  IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
 797
 798         if (!flags) {
 799                 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
 800                     !uverbs_cmd_table[command]) {
 801                         ret = -EINVAL;
 802                         goto out;
 803                 }
 804
 805                 if (hdr.in_words * 4 != count) {
 806                         ret = -EINVAL;
 807                         goto out;
 808                 }
 809
 810                 ret = uverbs_cmd_table[command](file, ib_dev,
 811                                                  buf + sizeof(hdr),
 812                                                  hdr.in_words * 4,
 813                                                  hdr.out_words * 4);
 814
 815         } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
 816                 struct ib_uverbs_ex_cmd_hdr ex_hdr;
 817                 struct ib_udata ucore;
 818                 struct ib_udata uhw;
 819                 size_t written_count = count;
 820
 821                 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
 822                     !uverbs_ex_cmd_table[command]) {
 823                         ret = -ENOSYS;
 824                         goto out;
 825                 }
 826
 827                 if (!file->ucontext) {
 828                         ret = -EINVAL;
 829                         goto out;
 830                 }
 831
 832                 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
 833                         ret = -EINVAL;
 834                         goto out;
 835                 }
 836
 837                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
 838                         ret = -EFAULT;
 839                         goto out;
 840                 }
 841
 842                 count -= sizeof(hdr) + sizeof(ex_hdr);
 843                 buf += sizeof(hdr) + sizeof(ex_hdr);
 844
 845                 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
 846                         ret = -EINVAL;
 847                         goto out;
 848                 }
 849
 850                 if (ex_hdr.cmd_hdr_reserved) {
 851                         ret = -EINVAL;
 852                         goto out;
 853                 }
 854
 855                 if (ex_hdr.response) {
 856                         if (!hdr.out_words && !ex_hdr.provider_out_words) {
 857                                 ret = -EINVAL;
 858                                 goto out;
 859                         }
 860
 861                         if (!access_ok(VERIFY_WRITE,
 862                                        (void __user *) (unsigned long) ex_hdr.response,
 863                                        (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
 864                                 ret = -EFAULT;
 865                                 goto out;
 866                         }
 867                 } else {
 868                         if (hdr.out_words || ex_hdr.provider_out_words) {
 869                                 ret = -EINVAL;
 870                                 goto out;
 871                         }
 872                 }
 873
 874                 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
 875                                        hdr.in_words * 8, hdr.out_words * 8);
 876
 877                 INIT_UDATA_BUF_OR_NULL(&uhw,
 878                                        buf + ucore.inlen,
 879                                        (unsigned long) ex_hdr.response + ucore.outlen,
 880                                        ex_hdr.provider_in_words * 8,
 881                                        ex_hdr.provider_out_words * 8);
 882
 883                 ret = uverbs_ex_cmd_table[command](file,
 884                                                    ib_dev,
 885                                                    &ucore,
 886                                                    &uhw);
 887                 if (!ret)
 888                         ret = written_count;
 889         } else {
 890                 ret = -ENOSYS;
 891         }
 892
 893 out:
 894         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 895         return ret;
 896 }
 897
 898 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 899 {
 900         struct ib_uverbs_file *file = filp->private_data;
 901         struct ib_device *ib_dev;
 902         int ret = 0;
 903         int srcu_key;
 904
 905         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 906         ib_dev = srcu_dereference(file->device->ib_dev,
 907                                   &file->device->disassociate_srcu);
 908         if (!ib_dev) {
 909                 ret = -EIO;
 910                 goto out;
 911         }
 912
 913         if (!file->ucontext)
 914                 ret = -ENODEV;
 915         else
 916                 ret = ib_dev->mmap(file->ucontext, vma);
 917 out:
 918         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 919         return ret;
 920 }
 921
 922 /*
 923  * ib_uverbs_open() does not need the BKL:
 924  *
 925  *  - the ib_uverbs_device structures are properly reference counted and
 926  *    everything else is purely local to the file being created, so
 927  *    races against other open calls are not a problem;
 928  *  - there is no ioctl method to race against;
 929  *  - the open method will either immediately run -ENXIO, or all
 930  *    required initialization will be done.
 931  */
 932 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 933 {
 934         struct ib_uverbs_device *dev;
 935         struct ib_uverbs_file *file;
 936         struct ib_device *ib_dev;
 937         int ret;
 938         int module_dependent;
 939         int srcu_key;
 940
 941         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 942         if (!atomic_inc_not_zero(&dev->refcount))
 943                 return -ENXIO;
 944
 945         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
 946         mutex_lock(&dev->lists_mutex);
 947         ib_dev = srcu_dereference(dev->ib_dev,
 948                                   &dev->disassociate_srcu);
 949         if (!ib_dev) {
 950                 ret = -EIO;
 951                 goto err;
 952         }
 953
 954         /* In case IB device supports disassociate ucontext, there is no hard
 955          * dependency between uverbs device and its low level device.
 956          */
 957         module_dependent = !(ib_dev->disassociate_ucontext);
 958
 959         if (module_dependent) {
 960                 if (!try_module_get(ib_dev->owner)) {
 961                         ret = -ENODEV;
 962                         goto err;
 963                 }
 964         }
 965
 966         file = kzalloc(sizeof(*file), GFP_KERNEL);
 967         if (!file) {
 968                 ret = -ENOMEM;
 969                 if (module_dependent)
 970                         goto err_module;
 971
 972                 goto err;
 973         }
 974
 975         file->device     = dev;
 976         spin_lock_init(&file->idr_lock);
 977         idr_init(&file->idr);
 978         file->ucontext   = NULL;
 979         file->async_file = NULL;
 980         kref_init(&file->ref);
 981         mutex_init(&file->mutex);
 982         mutex_init(&file->cleanup_mutex);
 983
 984         filp->private_data = file;
 985         kobject_get(&dev->kobj);
 986         list_add_tail(&file->list, &dev->uverbs_file_list);
 987         mutex_unlock(&dev->lists_mutex);
 988         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 989
 990         return nonseekable_open(inode, filp);
 991
 992 err_module:
 993         module_put(ib_dev->owner);
 994
 995 err:
 996         mutex_unlock(&dev->lists_mutex);
 997         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 998         if (atomic_dec_and_test(&dev->refcount))
 999                 ib_uverbs_comp_dev(dev);
1000
1001         return ret;
1002 }
1003
1004 static int ib_uverbs_close(struct inode *inode, struct file *filp)
1005 {
1006         struct ib_uverbs_file *file = filp->private_data;
1007         struct ib_uverbs_device *dev = file->device;
1008
1009         mutex_lock(&file->cleanup_mutex);
1010         if (file->ucontext) {
1011                 ib_uverbs_cleanup_ucontext(file, file->ucontext);
1012                 file->ucontext = NULL;
1013         }
1014         mutex_unlock(&file->cleanup_mutex);
1015         idr_destroy(&file->idr);
1016
1017         mutex_lock(&file->device->lists_mutex);
1018         if (!file->is_closed) {
1019                 list_del(&file->list);
1020                 file->is_closed = 1;
1021         }
1022         mutex_unlock(&file->device->lists_mutex);
1023
1024         if (file->async_file)
1025                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
1026
1027         kref_put(&file->ref, ib_uverbs_release_file);
1028         kobject_put(&dev->kobj);
1029
1030         return 0;
1031 }
1032
1033 static const struct file_operations uverbs_fops = {
1034         .owner   = THIS_MODULE,
1035         .write   = ib_uverbs_write,
1036         .open    = ib_uverbs_open,
1037         .release = ib_uverbs_close,
1038         .llseek  = no_llseek,
1039 };
1040
1041 static const struct file_operations uverbs_mmap_fops = {
1042         .owner   = THIS_MODULE,
1043         .write   = ib_uverbs_write,
1044         .mmap    = ib_uverbs_mmap,
1045         .open    = ib_uverbs_open,
1046         .release = ib_uverbs_close,
1047         .llseek  = no_llseek,
1048 };
1049
1050 static struct ib_client uverbs_client = {
1051         .name   = "uverbs",
1052         .add    = ib_uverbs_add_one,
1053         .remove = ib_uverbs_remove_one
1054 };
1055
1056 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
1057                           char *buf)
1058 {
1059         int ret = -ENODEV;
1060         int srcu_key;
1061         struct ib_uverbs_device *dev = dev_get_drvdata(device);
1062         struct ib_device *ib_dev;
1063
1064         if (!dev)
1065                 return -ENODEV;
1066
1067         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1068         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1069         if (ib_dev)
1070                 ret = sprintf(buf, "%s\n", ib_dev->name);
1071         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1072
1073         return ret;
1074 }
1075 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1076
1077 static ssize_t show_dev_abi_version(struct device *device,
1078                                     struct device_attribute *attr, char *buf)
1079 {
1080         struct ib_uverbs_device *dev = dev_get_drvdata(device);
1081         int ret = -ENODEV;
1082         int srcu_key;
1083         struct ib_device *ib_dev;
1084
1085         if (!dev)
1086                 return -ENODEV;
1087         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1088         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1089         if (ib_dev)
1090                 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1091         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1092
1093         return ret;
1094 }
1095 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1096
1097 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1098                          __stringify(IB_USER_VERBS_ABI_VERSION));
1099
1100 static dev_t overflow_maj;
1101 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1102
1103 /*
1104  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1105  * requesting a new major number and doubling the number of max devices we
1106  * support. It's stupid, but simple.
1107  */
1108 static int find_overflow_devnum(void)
1109 {
1110         int ret;
1111
1112         if (!overflow_maj) {
1113                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1114                                           "infiniband_verbs");
1115                 if (ret) {
1116                         pr_err("user_verbs: couldn't register dynamic device number\n");
1117                         return ret;
1118                 }
1119         }
1120
1121         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1122         if (ret >= IB_UVERBS_MAX_DEVICES)
1123                 return -1;
1124
1125         return ret;
1126 }
1127
1128 static void ib_uverbs_add_one(struct ib_device *device)
1129 {
1130         int devnum;
1131         dev_t base;
1132         struct ib_uverbs_device *uverbs_dev;
1133         int ret;
1134
1135         if (!device->alloc_ucontext)
1136                 return;
1137
1138         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1139         if (!uverbs_dev)
1140                 return;
1141
1142         ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1143         if (ret) {
1144                 kfree(uverbs_dev);
1145                 return;
1146         }
1147
1148         atomic_set(&uverbs_dev->refcount, 1);
1149         init_completion(&uverbs_dev->comp);
1150         uverbs_dev->xrcd_tree = RB_ROOT;
1151         mutex_init(&uverbs_dev->xrcd_tree_mutex);
1152         kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1153         mutex_init(&uverbs_dev->lists_mutex);
1154         INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1155         INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1156
1157         spin_lock(&map_lock);
1158         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1159         if (devnum >= IB_UVERBS_MAX_DEVICES) {
1160                 spin_unlock(&map_lock);
1161                 devnum = find_overflow_devnum();
1162                 if (devnum < 0)
1163                         goto err;
1164
1165                 spin_lock(&map_lock);
1166                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1167                 base = devnum + overflow_maj;
1168                 set_bit(devnum, overflow_map);
1169         } else {
1170                 uverbs_dev->devnum = devnum;
1171                 base = devnum + IB_UVERBS_BASE_DEV;
1172                 set_bit(devnum, dev_map);
1173         }
1174         spin_unlock(&map_lock);
1175
1176         rcu_assign_pointer(uverbs_dev->ib_dev, device);
1177         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1178
1179         cdev_init(&uverbs_dev->cdev, NULL);
1180         uverbs_dev->cdev.owner = THIS_MODULE;
1181         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1182         uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1183         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1184         if (cdev_add(&uverbs_dev->cdev, base, 1))
1185                 goto err_cdev;
1186
1187         uverbs_dev->dev = device_create(uverbs_class, device->dev.parent,
1188                                         uverbs_dev->cdev.dev, uverbs_dev,
1189                                         "uverbs%d", uverbs_dev->devnum);
1190         if (IS_ERR(uverbs_dev->dev))
1191                 goto err_cdev;
1192
1193         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1194                 goto err_class;
1195         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1196                 goto err_class;
1197
1198         ib_set_client_data(device, &uverbs_client, uverbs_dev);
1199
1200         return;
1201
1202 err_class:
1203         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1204
1205 err_cdev:
1206         cdev_del(&uverbs_dev->cdev);
1207         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1208                 clear_bit(devnum, dev_map);
1209         else
1210                 clear_bit(devnum, overflow_map);
1211
1212 err:
1213         if (atomic_dec_and_test(&uverbs_dev->refcount))
1214                 ib_uverbs_comp_dev(uverbs_dev);
1215         wait_for_completion(&uverbs_dev->comp);
1216         kobject_put(&uverbs_dev->kobj);
1217         return;
1218 }
1219
1220 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1221                                         struct ib_device *ib_dev)
1222 {
1223         struct ib_uverbs_file *file;
1224         struct ib_uverbs_event_file *event_file;
1225         struct ib_event event;
1226
1227         /* Pending running commands to terminate */
1228         synchronize_srcu(&uverbs_dev->disassociate_srcu);
1229         event.event = IB_EVENT_DEVICE_FATAL;
1230         event.element.port_num = 0;
1231         event.device = ib_dev;
1232
1233         mutex_lock(&uverbs_dev->lists_mutex);
1234         while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1235                 struct ib_ucontext *ucontext;
1236                 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1237                                         struct ib_uverbs_file, list);
1238                 file->is_closed = 1;
1239                 list_del(&file->list);
1240                 kref_get(&file->ref);
1241                 mutex_unlock(&uverbs_dev->lists_mutex);
1242
1243                 ib_uverbs_event_handler(&file->event_handler, &event);
1244
1245                 mutex_lock(&file->cleanup_mutex);
1246                 ucontext = file->ucontext;
1247                 file->ucontext = NULL;
1248                 mutex_unlock(&file->cleanup_mutex);
1249
1250                 /* At this point ib_uverbs_close cannot be running
1251                  * ib_uverbs_cleanup_ucontext
1252                  */
1253                 if (ucontext) {
1254                         /* We must release the mutex before going ahead and
1255                          * calling disassociate_ucontext. disassociate_ucontext
1256                          * might end up indirectly calling uverbs_close,
1257                          * for example due to freeing the resources
1258                          * (e.g mmput).
1259                          */
1260                         ib_dev->disassociate_ucontext(ucontext);
1261                         ib_uverbs_cleanup_ucontext(file, ucontext);
1262                 }
1263
1264                 mutex_lock(&uverbs_dev->lists_mutex);
1265                 kref_put(&file->ref, ib_uverbs_release_file);
1266         }
1267
1268         while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1269                 event_file = list_first_entry(&uverbs_dev->
1270                                               uverbs_events_file_list,
1271                                               struct ib_uverbs_event_file,
1272                                               list);
1273                 spin_lock_irq(&event_file->lock);
1274                 event_file->is_closed = 1;
1275                 spin_unlock_irq(&event_file->lock);
1276
1277                 list_del(&event_file->list);
1278                 if (event_file->is_async) {
1279                         ib_unregister_event_handler(&event_file->uverbs_file->
1280                                                     event_handler);
1281                         event_file->uverbs_file->event_handler.device = NULL;
1282                 }
1283
1284                 wake_up_interruptible(&event_file->poll_wait);
1285                 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1286         }
1287         mutex_unlock(&uverbs_dev->lists_mutex);
1288 }
1289
1290 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1291 {
1292         struct ib_uverbs_device *uverbs_dev = client_data;
1293         int wait_clients = 1;
1294
1295         if (!uverbs_dev)
1296                 return;
1297
1298         dev_set_drvdata(uverbs_dev->dev, NULL);
1299         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1300         cdev_del(&uverbs_dev->cdev);
1301
1302         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1303                 clear_bit(uverbs_dev->devnum, dev_map);
1304         else
1305                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1306
1307         if (device->disassociate_ucontext) {
1308                 /* We disassociate HW resources and immediately return.
1309                  * Userspace will see a EIO errno for all future access.
1310                  * Upon returning, ib_device may be freed internally and is not
1311                  * valid any more.
1312                  * uverbs_device is still available until all clients close
1313                  * their files, then the uverbs device ref count will be zero
1314                  * and its resources will be freed.
1315                  * Note: At this point no more files can be opened since the
1316                  * cdev was deleted, however active clients can still issue
1317                  * commands and close their open files.
1318                  */
1319                 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1320                 ib_uverbs_free_hw_resources(uverbs_dev, device);
1321                 wait_clients = 0;
1322         }
1323
1324         if (atomic_dec_and_test(&uverbs_dev->refcount))
1325                 ib_uverbs_comp_dev(uverbs_dev);
1326         if (wait_clients)
1327                 wait_for_completion(&uverbs_dev->comp);
1328         kobject_put(&uverbs_dev->kobj);
1329 }
1330
1331 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1332 {
1333         if (mode)
1334                 *mode = 0666;
1335         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1336 }
1337
1338 static int __init ib_uverbs_init(void)
1339 {
1340         int ret;
1341
1342         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1343                                      "infiniband_verbs");
1344         if (ret) {
1345                 pr_err("user_verbs: couldn't register device number\n");
1346                 goto out;
1347         }
1348
1349         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1350         if (IS_ERR(uverbs_class)) {
1351                 ret = PTR_ERR(uverbs_class);
1352                 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1353                 goto out_chrdev;
1354         }
1355
1356         uverbs_class->devnode = uverbs_devnode;
1357
1358         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1359         if (ret) {
1360                 pr_err("user_verbs: couldn't create abi_version attribute\n");
1361                 goto out_class;
1362         }
1363
1364         ret = ib_register_client(&uverbs_client);
1365         if (ret) {
1366                 pr_err("user_verbs: couldn't register client\n");
1367                 goto out_class;
1368         }
1369
1370         return 0;
1371
1372 out_class:
1373         class_destroy(uverbs_class);
1374
1375 out_chrdev:
1376         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1377
1378 out:
1379         return ret;
1380 }
1381
1382 static void __exit ib_uverbs_cleanup(void)
1383 {
1384         ib_unregister_client(&uverbs_client);
1385         class_destroy(uverbs_class);
1386         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1387         if (overflow_maj)
1388                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1389 }
1390
1391 module_init(ib_uverbs_init);
1392 module_exit(ib_uverbs_cleanup);