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/sched/mm.h>
  45 #include <linux/sched/task.h>
  46 #include <linux/file.h>
  47 #include <linux/cdev.h>
  48 #include <linux/anon_inodes.h>
  49 #include <linux/slab.h>
  50
  51 #include <linux/uaccess.h>
  52
  53 #include <rdma/ib.h>
  54 #include <rdma/uverbs_std_types.h>
  55
  56 #include "uverbs.h"
  57 #include "core_priv.h"
  58 #include "rdma_core.h"
  59
  60 MODULE_AUTHOR("Roland Dreier");
  61 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
  62 MODULE_LICENSE("Dual BSD/GPL");
  63
  64 enum {
  65         IB_UVERBS_MAJOR       = 231,
  66         IB_UVERBS_BASE_MINOR  = 192,
  67         IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS,
  68         IB_UVERBS_NUM_FIXED_MINOR = 32,
  69         IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR,
  70 };
  71
  72 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
  73
  74 static dev_t dynamic_uverbs_dev;
  75 static struct class *uverbs_class;
  76
  77 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
  78
  79 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
  80                                      struct ib_device *ib_dev,
  81                                      const char __user *buf, int in_len,
  82                                      int out_len) = {
  83         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
  84         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
  85         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
  86         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
  87         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
  88         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
  89         [IB_USER_VERBS_CMD_REREG_MR]            = ib_uverbs_rereg_mr,
  90         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
  91         [IB_USER_VERBS_CMD_ALLOC_MW]            = ib_uverbs_alloc_mw,
  92         [IB_USER_VERBS_CMD_DEALLOC_MW]          = ib_uverbs_dealloc_mw,
  93         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
  94         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
  95         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
  96         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
  97         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
  98         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
  99         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
 100         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
 101         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
 102         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
 103         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
 104         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
 105         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
 106         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
 107         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
 108         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
 109         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
 110         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
 111         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
 112         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
 113         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
 114         [IB_USER_VERBS_CMD_OPEN_XRCD]           = ib_uverbs_open_xrcd,
 115         [IB_USER_VERBS_CMD_CLOSE_XRCD]          = ib_uverbs_close_xrcd,
 116         [IB_USER_VERBS_CMD_CREATE_XSRQ]         = ib_uverbs_create_xsrq,
 117         [IB_USER_VERBS_CMD_OPEN_QP]             = ib_uverbs_open_qp,
 118 };
 119
 120 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
 121                                     struct ib_device *ib_dev,
 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         [IB_USER_VERBS_EX_CMD_CREATE_CQ]        = ib_uverbs_ex_create_cq,
 128         [IB_USER_VERBS_EX_CMD_CREATE_QP]        = ib_uverbs_ex_create_qp,
 129         [IB_USER_VERBS_EX_CMD_CREATE_WQ]        = ib_uverbs_ex_create_wq,
 130         [IB_USER_VERBS_EX_CMD_MODIFY_WQ]        = ib_uverbs_ex_modify_wq,
 131         [IB_USER_VERBS_EX_CMD_DESTROY_WQ]       = ib_uverbs_ex_destroy_wq,
 132         [IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL] = ib_uverbs_ex_create_rwq_ind_table,
 133         [IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL] = ib_uverbs_ex_destroy_rwq_ind_table,
 134         [IB_USER_VERBS_EX_CMD_MODIFY_QP]        = ib_uverbs_ex_modify_qp,
 135         [IB_USER_VERBS_EX_CMD_MODIFY_CQ]        = ib_uverbs_ex_modify_cq,
 136 };
 137
 138 static void ib_uverbs_add_one(struct ib_device *device);
 139 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
 140
 141 int uverbs_dealloc_mw(struct ib_mw *mw)
 142 {
 143         struct ib_pd *pd = mw->pd;
 144         int ret;
 145
 146         ret = mw->device->dealloc_mw(mw);
 147         if (!ret)
 148                 atomic_dec(&pd->usecnt);
 149         return ret;
 150 }
 151
 152 static void ib_uverbs_release_dev(struct kobject *kobj)
 153 {
 154         struct ib_uverbs_device *dev =
 155                 container_of(kobj, struct ib_uverbs_device, kobj);
 156
 157         cleanup_srcu_struct(&dev->disassociate_srcu);
 158         kfree(dev);
 159 }
 160
 161 static struct kobj_type ib_uverbs_dev_ktype = {
 162         .release = ib_uverbs_release_dev,
 163 };
 164
 165 static void ib_uverbs_release_async_event_file(struct kref *ref)
 166 {
 167         struct ib_uverbs_async_event_file *file =
 168                 container_of(ref, struct ib_uverbs_async_event_file, ref);
 169
 170         kfree(file);
 171 }
 172
 173 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
 174                           struct ib_uverbs_completion_event_file *ev_file,
 175                           struct ib_ucq_object *uobj)
 176 {
 177         struct ib_uverbs_event *evt, *tmp;
 178
 179         if (ev_file) {
 180                 spin_lock_irq(&ev_file->ev_queue.lock);
 181                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
 182                         list_del(&evt->list);
 183                         kfree(evt);
 184                 }
 185                 spin_unlock_irq(&ev_file->ev_queue.lock);
 186
 187                 uverbs_uobject_put(&ev_file->uobj_file.uobj);
 188         }
 189
 190         spin_lock_irq(&file->async_file->ev_queue.lock);
 191         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
 192                 list_del(&evt->list);
 193                 kfree(evt);
 194         }
 195         spin_unlock_irq(&file->async_file->ev_queue.lock);
 196 }
 197
 198 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
 199                               struct ib_uevent_object *uobj)
 200 {
 201         struct ib_uverbs_event *evt, *tmp;
 202
 203         spin_lock_irq(&file->async_file->ev_queue.lock);
 204         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
 205                 list_del(&evt->list);
 206                 kfree(evt);
 207         }
 208         spin_unlock_irq(&file->async_file->ev_queue.lock);
 209 }
 210
 211 void ib_uverbs_detach_umcast(struct ib_qp *qp,
 212                              struct ib_uqp_object *uobj)
 213 {
 214         struct ib_uverbs_mcast_entry *mcast, *tmp;
 215
 216         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
 217                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
 218                 list_del(&mcast->list);
 219                 kfree(mcast);
 220         }
 221 }
 222
 223 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
 224                                       struct ib_ucontext *context,
 225                                       bool device_removed)
 226 {
 227         context->closing = 1;
 228         uverbs_cleanup_ucontext(context, device_removed);
 229         put_pid(context->tgid);
 230
 231         ib_rdmacg_uncharge(&context->cg_obj, context->device,
 232                            RDMACG_RESOURCE_HCA_HANDLE);
 233
 234         return context->device->dealloc_ucontext(context);
 235 }
 236
 237 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
 238 {
 239         complete(&dev->comp);
 240 }
 241
 242 void ib_uverbs_release_file(struct kref *ref)
 243 {
 244         struct ib_uverbs_file *file =
 245                 container_of(ref, struct ib_uverbs_file, ref);
 246         struct ib_device *ib_dev;
 247         int srcu_key;
 248
 249         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 250         ib_dev = srcu_dereference(file->device->ib_dev,
 251                                   &file->device->disassociate_srcu);
 252         if (ib_dev && !ib_dev->disassociate_ucontext)
 253                 module_put(ib_dev->owner);
 254         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 255
 256         if (atomic_dec_and_test(&file->device->refcount))
 257                 ib_uverbs_comp_dev(file->device);
 258
 259         kobject_put(&file->device->kobj);
 260         kfree(file);
 261 }
 262
 263 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
 264                                     struct ib_uverbs_file *uverbs_file,
 265                                     struct file *filp, char __user *buf,
 266                                     size_t count, loff_t *pos,
 267                                     size_t eventsz)
 268 {
 269         struct ib_uverbs_event *event;
 270         int ret = 0;
 271
 272         spin_lock_irq(&ev_queue->lock);
 273
 274         while (list_empty(&ev_queue->event_list)) {
 275                 spin_unlock_irq(&ev_queue->lock);
 276
 277                 if (filp->f_flags & O_NONBLOCK)
 278                         return -EAGAIN;
 279
 280                 if (wait_event_interruptible(ev_queue->poll_wait,
 281                                              (!list_empty(&ev_queue->event_list) ||
 282                         /* The barriers built into wait_event_interruptible()
 283                          * and wake_up() guarentee this will see the null set
 284                          * without using RCU
 285                          */
 286                                              !uverbs_file->device->ib_dev)))
 287                         return -ERESTARTSYS;
 288
 289                 /* If device was disassociated and no event exists set an error */
 290                 if (list_empty(&ev_queue->event_list) &&
 291                     !uverbs_file->device->ib_dev)
 292                         return -EIO;
 293
 294                 spin_lock_irq(&ev_queue->lock);
 295         }
 296
 297         event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
 298
 299         if (eventsz > count) {
 300                 ret   = -EINVAL;
 301                 event = NULL;
 302         } else {
 303                 list_del(ev_queue->event_list.next);
 304                 if (event->counter) {
 305                         ++(*event->counter);
 306                         list_del(&event->obj_list);
 307                 }
 308         }
 309
 310         spin_unlock_irq(&ev_queue->lock);
 311
 312         if (event) {
 313                 if (copy_to_user(buf, event, eventsz))
 314                         ret = -EFAULT;
 315                 else
 316                         ret = eventsz;
 317         }
 318
 319         kfree(event);
 320
 321         return ret;
 322 }
 323
 324 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
 325                                           size_t count, loff_t *pos)
 326 {
 327         struct ib_uverbs_async_event_file *file = filp->private_data;
 328
 329         return ib_uverbs_event_read(&file->ev_queue, file->uverbs_file, filp,
 330                                     buf, count, pos,
 331                                     sizeof(struct ib_uverbs_async_event_desc));
 332 }
 333
 334 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
 335                                          size_t count, loff_t *pos)
 336 {
 337         struct ib_uverbs_completion_event_file *comp_ev_file =
 338                 filp->private_data;
 339
 340         return ib_uverbs_event_read(&comp_ev_file->ev_queue,
 341                                     comp_ev_file->uobj_file.ufile, filp,
 342                                     buf, count, pos,
 343                                     sizeof(struct ib_uverbs_comp_event_desc));
 344 }
 345
 346 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
 347                                          struct file *filp,
 348                                          struct poll_table_struct *wait)
 349 {
 350         __poll_t pollflags = 0;
 351
 352         poll_wait(filp, &ev_queue->poll_wait, wait);
 353
 354         spin_lock_irq(&ev_queue->lock);
 355         if (!list_empty(&ev_queue->event_list))
 356                 pollflags = EPOLLIN | EPOLLRDNORM;
 357         spin_unlock_irq(&ev_queue->lock);
 358
 359         return pollflags;
 360 }
 361
 362 static __poll_t ib_uverbs_async_event_poll(struct file *filp,
 363                                                struct poll_table_struct *wait)
 364 {
 365         return ib_uverbs_event_poll(filp->private_data, filp, wait);
 366 }
 367
 368 static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
 369                                               struct poll_table_struct *wait)
 370 {
 371         struct ib_uverbs_completion_event_file *comp_ev_file =
 372                 filp->private_data;
 373
 374         return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
 375 }
 376
 377 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
 378 {
 379         struct ib_uverbs_event_queue *ev_queue = filp->private_data;
 380
 381         return fasync_helper(fd, filp, on, &ev_queue->async_queue);
 382 }
 383
 384 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
 385 {
 386         struct ib_uverbs_completion_event_file *comp_ev_file =
 387                 filp->private_data;
 388
 389         return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
 390 }
 391
 392 static int ib_uverbs_async_event_close(struct inode *inode, struct file *filp)
 393 {
 394         struct ib_uverbs_async_event_file *file = filp->private_data;
 395         struct ib_uverbs_file *uverbs_file = file->uverbs_file;
 396         struct ib_uverbs_event *entry, *tmp;
 397         int closed_already = 0;
 398
 399         mutex_lock(&uverbs_file->device->lists_mutex);
 400         spin_lock_irq(&file->ev_queue.lock);
 401         closed_already = file->ev_queue.is_closed;
 402         file->ev_queue.is_closed = 1;
 403         list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
 404                 if (entry->counter)
 405                         list_del(&entry->obj_list);
 406                 kfree(entry);
 407         }
 408         spin_unlock_irq(&file->ev_queue.lock);
 409         if (!closed_already) {
 410                 list_del(&file->list);
 411                 ib_unregister_event_handler(&uverbs_file->event_handler);
 412         }
 413         mutex_unlock(&uverbs_file->device->lists_mutex);
 414
 415         kref_put(&uverbs_file->ref, ib_uverbs_release_file);
 416         kref_put(&file->ref, ib_uverbs_release_async_event_file);
 417
 418         return 0;
 419 }
 420
 421 static int ib_uverbs_comp_event_close(struct inode *inode, struct file *filp)
 422 {
 423         struct ib_uverbs_completion_event_file *file = filp->private_data;
 424         struct ib_uverbs_event *entry, *tmp;
 425
 426         spin_lock_irq(&file->ev_queue.lock);
 427         list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
 428                 if (entry->counter)
 429                         list_del(&entry->obj_list);
 430                 kfree(entry);
 431         }
 432         spin_unlock_irq(&file->ev_queue.lock);
 433
 434         uverbs_close_fd(filp);
 435
 436         return 0;
 437 }
 438
 439 const struct file_operations uverbs_event_fops = {
 440         .owner   = THIS_MODULE,
 441         .read    = ib_uverbs_comp_event_read,
 442         .poll    = ib_uverbs_comp_event_poll,
 443         .release = ib_uverbs_comp_event_close,
 444         .fasync  = ib_uverbs_comp_event_fasync,
 445         .llseek  = no_llseek,
 446 };
 447
 448 static const struct file_operations uverbs_async_event_fops = {
 449         .owner   = THIS_MODULE,
 450         .read    = ib_uverbs_async_event_read,
 451         .poll    = ib_uverbs_async_event_poll,
 452         .release = ib_uverbs_async_event_close,
 453         .fasync  = ib_uverbs_async_event_fasync,
 454         .llseek  = no_llseek,
 455 };
 456
 457 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
 458 {
 459         struct ib_uverbs_event_queue   *ev_queue = cq_context;
 460         struct ib_ucq_object           *uobj;
 461         struct ib_uverbs_event         *entry;
 462         unsigned long                   flags;
 463
 464         if (!ev_queue)
 465                 return;
 466
 467         spin_lock_irqsave(&ev_queue->lock, flags);
 468         if (ev_queue->is_closed) {
 469                 spin_unlock_irqrestore(&ev_queue->lock, flags);
 470                 return;
 471         }
 472
 473         entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 474         if (!entry) {
 475                 spin_unlock_irqrestore(&ev_queue->lock, flags);
 476                 return;
 477         }
 478
 479         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
 480
 481         entry->desc.comp.cq_handle = cq->uobject->user_handle;
 482         entry->counter             = &uobj->comp_events_reported;
 483
 484         list_add_tail(&entry->list, &ev_queue->event_list);
 485         list_add_tail(&entry->obj_list, &uobj->comp_list);
 486         spin_unlock_irqrestore(&ev_queue->lock, flags);
 487
 488         wake_up_interruptible(&ev_queue->poll_wait);
 489         kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
 490 }
 491
 492 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
 493                                     __u64 element, __u64 event,
 494                                     struct list_head *obj_list,
 495                                     u32 *counter)
 496 {
 497         struct ib_uverbs_event *entry;
 498         unsigned long flags;
 499
 500         spin_lock_irqsave(&file->async_file->ev_queue.lock, flags);
 501         if (file->async_file->ev_queue.is_closed) {
 502                 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 503                 return;
 504         }
 505
 506         entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
 507         if (!entry) {
 508                 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 509                 return;
 510         }
 511
 512         entry->desc.async.element    = element;
 513         entry->desc.async.event_type = event;
 514         entry->desc.async.reserved   = 0;
 515         entry->counter               = counter;
 516
 517         list_add_tail(&entry->list, &file->async_file->ev_queue.event_list);
 518         if (obj_list)
 519                 list_add_tail(&entry->obj_list, obj_list);
 520         spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
 521
 522         wake_up_interruptible(&file->async_file->ev_queue.poll_wait);
 523         kill_fasync(&file->async_file->ev_queue.async_queue, SIGIO, POLL_IN);
 524 }
 525
 526 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
 527 {
 528         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
 529                                                   struct ib_ucq_object, uobject);
 530
 531         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
 532                                 event->event, &uobj->async_list,
 533                                 &uobj->async_events_reported);
 534 }
 535
 536 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
 537 {
 538         struct ib_uevent_object *uobj;
 539
 540         /* for XRC target qp's, check that qp is live */
 541         if (!event->element.qp->uobject)
 542                 return;
 543
 544         uobj = container_of(event->element.qp->uobject,
 545                             struct ib_uevent_object, uobject);
 546
 547         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 548                                 event->event, &uobj->event_list,
 549                                 &uobj->events_reported);
 550 }
 551
 552 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
 553 {
 554         struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
 555                                                   struct ib_uevent_object, uobject);
 556
 557         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 558                                 event->event, &uobj->event_list,
 559                                 &uobj->events_reported);
 560 }
 561
 562 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
 563 {
 564         struct ib_uevent_object *uobj;
 565
 566         uobj = container_of(event->element.srq->uobject,
 567                             struct ib_uevent_object, uobject);
 568
 569         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
 570                                 event->event, &uobj->event_list,
 571                                 &uobj->events_reported);
 572 }
 573
 574 void ib_uverbs_event_handler(struct ib_event_handler *handler,
 575                              struct ib_event *event)
 576 {
 577         struct ib_uverbs_file *file =
 578                 container_of(handler, struct ib_uverbs_file, event_handler);
 579
 580         ib_uverbs_async_handler(file, event->element.port_num, event->event,
 581                                 NULL, NULL);
 582 }
 583
 584 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
 585 {
 586         kref_put(&file->async_file->ref, ib_uverbs_release_async_event_file);
 587         file->async_file = NULL;
 588 }
 589
 590 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
 591 {
 592         spin_lock_init(&ev_queue->lock);
 593         INIT_LIST_HEAD(&ev_queue->event_list);
 594         init_waitqueue_head(&ev_queue->poll_wait);
 595         ev_queue->is_closed   = 0;
 596         ev_queue->async_queue = NULL;
 597 }
 598
 599 struct file *ib_uverbs_alloc_async_event_file(struct ib_uverbs_file *uverbs_file,
 600                                               struct ib_device  *ib_dev)
 601 {
 602         struct ib_uverbs_async_event_file *ev_file;
 603         struct file *filp;
 604
 605         ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
 606         if (!ev_file)
 607                 return ERR_PTR(-ENOMEM);
 608
 609         ib_uverbs_init_event_queue(&ev_file->ev_queue);
 610         ev_file->uverbs_file = uverbs_file;
 611         kref_get(&ev_file->uverbs_file->ref);
 612         kref_init(&ev_file->ref);
 613         filp = anon_inode_getfile("[infinibandevent]", &uverbs_async_event_fops,
 614                                   ev_file, O_RDONLY);
 615         if (IS_ERR(filp))
 616                 goto err_put_refs;
 617
 618         mutex_lock(&uverbs_file->device->lists_mutex);
 619         list_add_tail(&ev_file->list,
 620                       &uverbs_file->device->uverbs_events_file_list);
 621         mutex_unlock(&uverbs_file->device->lists_mutex);
 622
 623         WARN_ON(uverbs_file->async_file);
 624         uverbs_file->async_file = ev_file;
 625         kref_get(&uverbs_file->async_file->ref);
 626         INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
 627                               ib_dev,
 628                               ib_uverbs_event_handler);
 629         ib_register_event_handler(&uverbs_file->event_handler);
 630         /* At that point async file stuff was fully set */
 631
 632         return filp;
 633
 634 err_put_refs:
 635         kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
 636         kref_put(&ev_file->ref, ib_uverbs_release_async_event_file);
 637         return filp;
 638 }
 639
 640 static bool verify_command_mask(struct ib_device *ib_dev,
 641                                 u32 command, bool extended)
 642 {
 643         if (!extended)
 644                 return ib_dev->uverbs_cmd_mask & BIT_ULL(command);
 645
 646         return ib_dev->uverbs_ex_cmd_mask & BIT_ULL(command);
 647 }
 648
 649 static bool verify_command_idx(u32 command, bool extended)
 650 {
 651         if (extended)
 652                 return command < ARRAY_SIZE(uverbs_ex_cmd_table) &&
 653                        uverbs_ex_cmd_table[command];
 654
 655         return command < ARRAY_SIZE(uverbs_cmd_table) &&
 656                uverbs_cmd_table[command];
 657 }
 658
 659 static ssize_t process_hdr(struct ib_uverbs_cmd_hdr *hdr,
 660                            u32 *command, bool *extended)
 661 {
 662         if (hdr->command & ~(u32)(IB_USER_VERBS_CMD_FLAG_EXTENDED |
 663                                    IB_USER_VERBS_CMD_COMMAND_MASK))
 664                 return -EINVAL;
 665
 666         *command = hdr->command & IB_USER_VERBS_CMD_COMMAND_MASK;
 667         *extended = hdr->command & IB_USER_VERBS_CMD_FLAG_EXTENDED;
 668
 669         if (!verify_command_idx(*command, *extended))
 670                 return -EOPNOTSUPP;
 671
 672         return 0;
 673 }
 674
 675 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
 676                           struct ib_uverbs_ex_cmd_hdr *ex_hdr,
 677                           size_t count, bool extended)
 678 {
 679         if (extended) {
 680                 count -= sizeof(*hdr) + sizeof(*ex_hdr);
 681
 682                 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
 683                         return -EINVAL;
 684
 685                 if (ex_hdr->cmd_hdr_reserved)
 686                         return -EINVAL;
 687
 688                 if (ex_hdr->response) {
 689                         if (!hdr->out_words && !ex_hdr->provider_out_words)
 690                                 return -EINVAL;
 691
 692                         if (!access_ok(VERIFY_WRITE,
 693                                        u64_to_user_ptr(ex_hdr->response),
 694                                        (hdr->out_words + ex_hdr->provider_out_words) * 8))
 695                                 return -EFAULT;
 696                 } else {
 697                         if (hdr->out_words || ex_hdr->provider_out_words)
 698                                 return -EINVAL;
 699                 }
 700
 701                 return 0;
 702         }
 703
 704         /* not extended command */
 705         if (hdr->in_words * 4 != count)
 706                 return -EINVAL;
 707
 708         return 0;
 709 }
 710
 711 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
 712                              size_t count, loff_t *pos)
 713 {
 714         struct ib_uverbs_file *file = filp->private_data;
 715         struct ib_uverbs_ex_cmd_hdr ex_hdr;
 716         struct ib_device *ib_dev;
 717         struct ib_uverbs_cmd_hdr hdr;
 718         bool extended;
 719         int srcu_key;
 720         u32 command;
 721         ssize_t ret;
 722
 723         if (!ib_safe_file_access(filp)) {
 724                 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
 725                             task_tgid_vnr(current), current->comm);
 726                 return -EACCES;
 727         }
 728
 729         if (count < sizeof(hdr))
 730                 return -EINVAL;
 731
 732         if (copy_from_user(&hdr, buf, sizeof(hdr)))
 733                 return -EFAULT;
 734
 735         ret = process_hdr(&hdr, &command, &extended);
 736         if (ret)
 737                 return ret;
 738
 739         if (!file->ucontext &&
 740             (command != IB_USER_VERBS_CMD_GET_CONTEXT || extended))
 741                 return -EINVAL;
 742
 743         if (extended) {
 744                 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 745                         return -EINVAL;
 746                 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
 747                         return -EFAULT;
 748         }
 749
 750         ret = verify_hdr(&hdr, &ex_hdr, count, extended);
 751         if (ret)
 752                 return ret;
 753
 754         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 755         ib_dev = srcu_dereference(file->device->ib_dev,
 756                                   &file->device->disassociate_srcu);
 757         if (!ib_dev) {
 758                 ret = -EIO;
 759                 goto out;
 760         }
 761
 762         if (!verify_command_mask(ib_dev, command, extended)) {
 763                 ret = -EOPNOTSUPP;
 764                 goto out;
 765         }
 766
 767         buf += sizeof(hdr);
 768
 769         if (!extended) {
 770                 ret = uverbs_cmd_table[command](file, ib_dev, buf,
 771                                                 hdr.in_words * 4,
 772                                                 hdr.out_words * 4);
 773         } else {
 774                 struct ib_udata ucore;
 775                 struct ib_udata uhw;
 776
 777                 buf += sizeof(ex_hdr);
 778
 779                 ib_uverbs_init_udata_buf_or_null(&ucore, buf,
 780                                         u64_to_user_ptr(ex_hdr.response),
 781                                         hdr.in_words * 8, hdr.out_words * 8);
 782
 783                 ib_uverbs_init_udata_buf_or_null(&uhw,
 784                                         buf + ucore.inlen,
 785                                         u64_to_user_ptr(ex_hdr.response) + ucore.outlen,
 786                                         ex_hdr.provider_in_words * 8,
 787                                         ex_hdr.provider_out_words * 8);
 788
 789                 ret = uverbs_ex_cmd_table[command](file, ib_dev, &ucore, &uhw);
 790                 ret = (ret) ? : count;
 791         }
 792
 793 out:
 794         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 795         return ret;
 796 }
 797
 798 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
 799 {
 800         struct ib_uverbs_file *file = filp->private_data;
 801         struct ib_device *ib_dev;
 802         int ret = 0;
 803         int srcu_key;
 804
 805         srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
 806         ib_dev = srcu_dereference(file->device->ib_dev,
 807                                   &file->device->disassociate_srcu);
 808         if (!ib_dev) {
 809                 ret = -EIO;
 810                 goto out;
 811         }
 812
 813         if (!file->ucontext)
 814                 ret = -ENODEV;
 815         else
 816                 ret = ib_dev->mmap(file->ucontext, vma);
 817 out:
 818         srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
 819         return ret;
 820 }
 821
 822 /*
 823  * ib_uverbs_open() does not need the BKL:
 824  *
 825  *  - the ib_uverbs_device structures are properly reference counted and
 826  *    everything else is purely local to the file being created, so
 827  *    races against other open calls are not a problem;
 828  *  - there is no ioctl method to race against;
 829  *  - the open method will either immediately run -ENXIO, or all
 830  *    required initialization will be done.
 831  */
 832 static int ib_uverbs_open(struct inode *inode, struct file *filp)
 833 {
 834         struct ib_uverbs_device *dev;
 835         struct ib_uverbs_file *file;
 836         struct ib_device *ib_dev;
 837         int ret;
 838         int module_dependent;
 839         int srcu_key;
 840
 841         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
 842         if (!atomic_inc_not_zero(&dev->refcount))
 843                 return -ENXIO;
 844
 845         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
 846         mutex_lock(&dev->lists_mutex);
 847         ib_dev = srcu_dereference(dev->ib_dev,
 848                                   &dev->disassociate_srcu);
 849         if (!ib_dev) {
 850                 ret = -EIO;
 851                 goto err;
 852         }
 853
 854         /* In case IB device supports disassociate ucontext, there is no hard
 855          * dependency between uverbs device and its low level device.
 856          */
 857         module_dependent = !(ib_dev->disassociate_ucontext);
 858
 859         if (module_dependent) {
 860                 if (!try_module_get(ib_dev->owner)) {
 861                         ret = -ENODEV;
 862                         goto err;
 863                 }
 864         }
 865
 866         file = kzalloc(sizeof(*file), GFP_KERNEL);
 867         if (!file) {
 868                 ret = -ENOMEM;
 869                 if (module_dependent)
 870                         goto err_module;
 871
 872                 goto err;
 873         }
 874
 875         file->device     = dev;
 876         spin_lock_init(&file->idr_lock);
 877         idr_init(&file->idr);
 878         file->ucontext   = NULL;
 879         file->async_file = NULL;
 880         kref_init(&file->ref);
 881         mutex_init(&file->mutex);
 882         mutex_init(&file->cleanup_mutex);
 883
 884         filp->private_data = file;
 885         kobject_get(&dev->kobj);
 886         list_add_tail(&file->list, &dev->uverbs_file_list);
 887         mutex_unlock(&dev->lists_mutex);
 888         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 889
 890         return nonseekable_open(inode, filp);
 891
 892 err_module:
 893         module_put(ib_dev->owner);
 894
 895 err:
 896         mutex_unlock(&dev->lists_mutex);
 897         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 898         if (atomic_dec_and_test(&dev->refcount))
 899                 ib_uverbs_comp_dev(dev);
 900
 901         return ret;
 902 }
 903
 904 static int ib_uverbs_close(struct inode *inode, struct file *filp)
 905 {
 906         struct ib_uverbs_file *file = filp->private_data;
 907
 908         mutex_lock(&file->cleanup_mutex);
 909         if (file->ucontext) {
 910                 ib_uverbs_cleanup_ucontext(file, file->ucontext, false);
 911                 file->ucontext = NULL;
 912         }
 913         mutex_unlock(&file->cleanup_mutex);
 914         idr_destroy(&file->idr);
 915
 916         mutex_lock(&file->device->lists_mutex);
 917         if (!file->is_closed) {
 918                 list_del(&file->list);
 919                 file->is_closed = 1;
 920         }
 921         mutex_unlock(&file->device->lists_mutex);
 922
 923         if (file->async_file)
 924                 kref_put(&file->async_file->ref,
 925                          ib_uverbs_release_async_event_file);
 926
 927         kref_put(&file->ref, ib_uverbs_release_file);
 928
 929         return 0;
 930 }
 931
 932 static const struct file_operations uverbs_fops = {
 933         .owner   = THIS_MODULE,
 934         .write   = ib_uverbs_write,
 935         .open    = ib_uverbs_open,
 936         .release = ib_uverbs_close,
 937         .llseek  = no_llseek,
 938         .unlocked_ioctl = ib_uverbs_ioctl,
 939         .compat_ioctl = ib_uverbs_ioctl,
 940 };
 941
 942 static const struct file_operations uverbs_mmap_fops = {
 943         .owner   = THIS_MODULE,
 944         .write   = ib_uverbs_write,
 945         .mmap    = ib_uverbs_mmap,
 946         .open    = ib_uverbs_open,
 947         .release = ib_uverbs_close,
 948         .llseek  = no_llseek,
 949         .unlocked_ioctl = ib_uverbs_ioctl,
 950         .compat_ioctl = ib_uverbs_ioctl,
 951 };
 952
 953 static struct ib_client uverbs_client = {
 954         .name   = "uverbs",
 955         .add    = ib_uverbs_add_one,
 956         .remove = ib_uverbs_remove_one
 957 };
 958
 959 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
 960                           char *buf)
 961 {
 962         int ret = -ENODEV;
 963         int srcu_key;
 964         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 965         struct ib_device *ib_dev;
 966
 967         if (!dev)
 968                 return -ENODEV;
 969
 970         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
 971         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
 972         if (ib_dev)
 973                 ret = sprintf(buf, "%s\n", ib_dev->name);
 974         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 975
 976         return ret;
 977 }
 978 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 979
 980 static ssize_t show_dev_abi_version(struct device *device,
 981                                     struct device_attribute *attr, char *buf)
 982 {
 983         struct ib_uverbs_device *dev = dev_get_drvdata(device);
 984         int ret = -ENODEV;
 985         int srcu_key;
 986         struct ib_device *ib_dev;
 987
 988         if (!dev)
 989                 return -ENODEV;
 990         srcu_key = srcu_read_lock(&dev->disassociate_srcu);
 991         ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
 992         if (ib_dev)
 993                 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
 994         srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
 995
 996         return ret;
 997 }
 998 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
 999
1000 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1001                          __stringify(IB_USER_VERBS_ABI_VERSION));
1002
1003 static void ib_uverbs_add_one(struct ib_device *device)
1004 {
1005         int devnum;
1006         dev_t base;
1007         struct ib_uverbs_device *uverbs_dev;
1008         int ret;
1009
1010         if (!device->alloc_ucontext)
1011                 return;
1012
1013         uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1014         if (!uverbs_dev)
1015                 return;
1016
1017         ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1018         if (ret) {
1019                 kfree(uverbs_dev);
1020                 return;
1021         }
1022
1023         atomic_set(&uverbs_dev->refcount, 1);
1024         init_completion(&uverbs_dev->comp);
1025         uverbs_dev->xrcd_tree = RB_ROOT;
1026         mutex_init(&uverbs_dev->xrcd_tree_mutex);
1027         kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1028         mutex_init(&uverbs_dev->lists_mutex);
1029         INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1030         INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1031
1032         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1033         if (devnum >= IB_UVERBS_MAX_DEVICES)
1034                 goto err;
1035         uverbs_dev->devnum = devnum;
1036         set_bit(devnum, dev_map);
1037         if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1038                 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1039         else
1040                 base = IB_UVERBS_BASE_DEV + devnum;
1041
1042         rcu_assign_pointer(uverbs_dev->ib_dev, device);
1043         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1044
1045         cdev_init(&uverbs_dev->cdev, NULL);
1046         uverbs_dev->cdev.owner = THIS_MODULE;
1047         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1048         cdev_set_parent(&uverbs_dev->cdev, &uverbs_dev->kobj);
1049         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1050         if (cdev_add(&uverbs_dev->cdev, base, 1))
1051                 goto err_cdev;
1052
1053         uverbs_dev->dev = device_create(uverbs_class, device->dev.parent,
1054                                         uverbs_dev->cdev.dev, uverbs_dev,
1055                                         "uverbs%d", uverbs_dev->devnum);
1056         if (IS_ERR(uverbs_dev->dev))
1057                 goto err_cdev;
1058
1059         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1060                 goto err_class;
1061         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1062                 goto err_class;
1063
1064         if (!device->specs_root) {
1065                 const struct uverbs_object_tree_def *default_root[] = {
1066                         uverbs_default_get_objects()};
1067
1068                 uverbs_dev->specs_root = uverbs_alloc_spec_tree(1,
1069                                                                 default_root);
1070                 if (IS_ERR(uverbs_dev->specs_root))
1071                         goto err_class;
1072
1073                 device->specs_root = uverbs_dev->specs_root;
1074         }
1075
1076         ib_set_client_data(device, &uverbs_client, uverbs_dev);
1077
1078         return;
1079
1080 err_class:
1081         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1082
1083 err_cdev:
1084         cdev_del(&uverbs_dev->cdev);
1085         clear_bit(devnum, dev_map);
1086
1087 err:
1088         if (atomic_dec_and_test(&uverbs_dev->refcount))
1089                 ib_uverbs_comp_dev(uverbs_dev);
1090         wait_for_completion(&uverbs_dev->comp);
1091         kobject_put(&uverbs_dev->kobj);
1092         return;
1093 }
1094
1095 static void ib_uverbs_disassociate_ucontext(struct ib_ucontext *ibcontext)
1096 {
1097         struct ib_device *ib_dev = ibcontext->device;
1098         struct task_struct *owning_process  = NULL;
1099         struct mm_struct   *owning_mm       = NULL;
1100
1101         owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1102         if (!owning_process)
1103                 return;
1104
1105         owning_mm = get_task_mm(owning_process);
1106         if (!owning_mm) {
1107                 pr_info("no mm, disassociate ucontext is pending task termination\n");
1108                 while (1) {
1109                         put_task_struct(owning_process);
1110                         usleep_range(1000, 2000);
1111                         owning_process = get_pid_task(ibcontext->tgid,
1112                                                       PIDTYPE_PID);
1113                         if (!owning_process ||
1114                             owning_process->state == TASK_DEAD) {
1115                                 pr_info("disassociate ucontext done, task was terminated\n");
1116                                 /* in case task was dead need to release the
1117                                  * task struct.
1118                                  */
1119                                 if (owning_process)
1120                                         put_task_struct(owning_process);
1121                                 return;
1122                         }
1123                 }
1124         }
1125
1126         down_write(&owning_mm->mmap_sem);
1127         ib_dev->disassociate_ucontext(ibcontext);
1128         up_write(&owning_mm->mmap_sem);
1129         mmput(owning_mm);
1130         put_task_struct(owning_process);
1131 }
1132
1133 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1134                                         struct ib_device *ib_dev)
1135 {
1136         struct ib_uverbs_file *file;
1137         struct ib_uverbs_async_event_file *event_file;
1138         struct ib_event event;
1139
1140         /* Pending running commands to terminate */
1141         synchronize_srcu(&uverbs_dev->disassociate_srcu);
1142         event.event = IB_EVENT_DEVICE_FATAL;
1143         event.element.port_num = 0;
1144         event.device = ib_dev;
1145
1146         mutex_lock(&uverbs_dev->lists_mutex);
1147         while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1148                 struct ib_ucontext *ucontext;
1149                 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1150                                         struct ib_uverbs_file, list);
1151                 file->is_closed = 1;
1152                 list_del(&file->list);
1153                 kref_get(&file->ref);
1154                 mutex_unlock(&uverbs_dev->lists_mutex);
1155
1156
1157                 mutex_lock(&file->cleanup_mutex);
1158                 ucontext = file->ucontext;
1159                 file->ucontext = NULL;
1160                 mutex_unlock(&file->cleanup_mutex);
1161
1162                 /* At this point ib_uverbs_close cannot be running
1163                  * ib_uverbs_cleanup_ucontext
1164                  */
1165                 if (ucontext) {
1166                         /* We must release the mutex before going ahead and
1167                          * calling disassociate_ucontext. disassociate_ucontext
1168                          * might end up indirectly calling uverbs_close,
1169                          * for example due to freeing the resources
1170                          * (e.g mmput).
1171                          */
1172                         ib_uverbs_event_handler(&file->event_handler, &event);
1173                         ib_uverbs_disassociate_ucontext(ucontext);
1174                         mutex_lock(&file->cleanup_mutex);
1175                         ib_uverbs_cleanup_ucontext(file, ucontext, true);
1176                         mutex_unlock(&file->cleanup_mutex);
1177                 }
1178
1179                 mutex_lock(&uverbs_dev->lists_mutex);
1180                 kref_put(&file->ref, ib_uverbs_release_file);
1181         }
1182
1183         while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1184                 event_file = list_first_entry(&uverbs_dev->
1185                                               uverbs_events_file_list,
1186                                               struct ib_uverbs_async_event_file,
1187                                               list);
1188                 spin_lock_irq(&event_file->ev_queue.lock);
1189                 event_file->ev_queue.is_closed = 1;
1190                 spin_unlock_irq(&event_file->ev_queue.lock);
1191
1192                 list_del(&event_file->list);
1193                 ib_unregister_event_handler(
1194                         &event_file->uverbs_file->event_handler);
1195                 event_file->uverbs_file->event_handler.device =
1196                         NULL;
1197
1198                 wake_up_interruptible(&event_file->ev_queue.poll_wait);
1199                 kill_fasync(&event_file->ev_queue.async_queue, SIGIO, POLL_IN);
1200         }
1201         mutex_unlock(&uverbs_dev->lists_mutex);
1202 }
1203
1204 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1205 {
1206         struct ib_uverbs_device *uverbs_dev = client_data;
1207         int wait_clients = 1;
1208
1209         if (!uverbs_dev)
1210                 return;
1211
1212         dev_set_drvdata(uverbs_dev->dev, NULL);
1213         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1214         cdev_del(&uverbs_dev->cdev);
1215         clear_bit(uverbs_dev->devnum, dev_map);
1216
1217         if (device->disassociate_ucontext) {
1218                 /* We disassociate HW resources and immediately return.
1219                  * Userspace will see a EIO errno for all future access.
1220                  * Upon returning, ib_device may be freed internally and is not
1221                  * valid any more.
1222                  * uverbs_device is still available until all clients close
1223                  * their files, then the uverbs device ref count will be zero
1224                  * and its resources will be freed.
1225                  * Note: At this point no more files can be opened since the
1226                  * cdev was deleted, however active clients can still issue
1227                  * commands and close their open files.
1228                  */
1229                 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1230                 ib_uverbs_free_hw_resources(uverbs_dev, device);
1231                 wait_clients = 0;
1232         }
1233
1234         if (atomic_dec_and_test(&uverbs_dev->refcount))
1235                 ib_uverbs_comp_dev(uverbs_dev);
1236         if (wait_clients)
1237                 wait_for_completion(&uverbs_dev->comp);
1238         if (uverbs_dev->specs_root) {
1239                 uverbs_free_spec_tree(uverbs_dev->specs_root);
1240                 device->specs_root = NULL;
1241         }
1242
1243         kobject_put(&uverbs_dev->kobj);
1244 }
1245
1246 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1247 {
1248         if (mode)
1249                 *mode = 0666;
1250         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1251 }
1252
1253 static int __init ib_uverbs_init(void)
1254 {
1255         int ret;
1256
1257         ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1258                                      IB_UVERBS_NUM_FIXED_MINOR,
1259                                      "infiniband_verbs");
1260         if (ret) {
1261                 pr_err("user_verbs: couldn't register device number\n");
1262                 goto out;
1263         }
1264
1265         ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1266                                   IB_UVERBS_NUM_DYNAMIC_MINOR,
1267                                   "infiniband_verbs");
1268         if (ret) {
1269                 pr_err("couldn't register dynamic device number\n");
1270                 goto out_alloc;
1271         }
1272
1273         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1274         if (IS_ERR(uverbs_class)) {
1275                 ret = PTR_ERR(uverbs_class);
1276                 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1277                 goto out_chrdev;
1278         }
1279
1280         uverbs_class->devnode = uverbs_devnode;
1281
1282         ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1283         if (ret) {
1284                 pr_err("user_verbs: couldn't create abi_version attribute\n");
1285                 goto out_class;
1286         }
1287
1288         ret = ib_register_client(&uverbs_client);
1289         if (ret) {
1290                 pr_err("user_verbs: couldn't register client\n");
1291                 goto out_class;
1292         }
1293
1294         return 0;
1295
1296 out_class:
1297         class_destroy(uverbs_class);
1298
1299 out_chrdev:
1300         unregister_chrdev_region(dynamic_uverbs_dev,
1301                                  IB_UVERBS_NUM_DYNAMIC_MINOR);
1302
1303 out_alloc:
1304         unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1305                                  IB_UVERBS_NUM_FIXED_MINOR);
1306
1307 out:
1308         return ret;
1309 }
1310
1311 static void __exit ib_uverbs_cleanup(void)
1312 {
1313         ib_unregister_client(&uverbs_client);
1314         class_destroy(uverbs_class);
1315         unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1316                                  IB_UVERBS_NUM_FIXED_MINOR);
1317         unregister_chrdev_region(dynamic_uverbs_dev,
1318                                  IB_UVERBS_NUM_DYNAMIC_MINOR);
1319 }
1320
1321 module_init(ib_uverbs_init);
1322 module_exit(ib_uverbs_cleanup);