Merge branch 'srp' into for-next
[linux-3.10.git] / 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/mount.h>
46
47 #include <asm/uaccess.h>
48
49 #include "uverbs.h"
50
51 MODULE_AUTHOR("Roland Dreier");
52 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
53 MODULE_LICENSE("Dual BSD/GPL");
54
55 #define INFINIBANDEVENTFS_MAGIC 0x49426576      /* "IBev" */
56
57 enum {
58         IB_UVERBS_MAJOR       = 231,
59         IB_UVERBS_BASE_MINOR  = 192,
60         IB_UVERBS_MAX_DEVICES = 32
61 };
62
63 #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
64
65 static struct class *uverbs_class;
66
67 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
68 DEFINE_IDR(ib_uverbs_pd_idr);
69 DEFINE_IDR(ib_uverbs_mr_idr);
70 DEFINE_IDR(ib_uverbs_mw_idr);
71 DEFINE_IDR(ib_uverbs_ah_idr);
72 DEFINE_IDR(ib_uverbs_cq_idr);
73 DEFINE_IDR(ib_uverbs_qp_idr);
74 DEFINE_IDR(ib_uverbs_srq_idr);
75
76 static DEFINE_SPINLOCK(map_lock);
77 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
78
79 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
80                                      const char __user *buf, int in_len,
81                                      int out_len) = {
82         [IB_USER_VERBS_CMD_GET_CONTEXT]         = ib_uverbs_get_context,
83         [IB_USER_VERBS_CMD_QUERY_DEVICE]        = ib_uverbs_query_device,
84         [IB_USER_VERBS_CMD_QUERY_PORT]          = ib_uverbs_query_port,
85         [IB_USER_VERBS_CMD_ALLOC_PD]            = ib_uverbs_alloc_pd,
86         [IB_USER_VERBS_CMD_DEALLOC_PD]          = ib_uverbs_dealloc_pd,
87         [IB_USER_VERBS_CMD_REG_MR]              = ib_uverbs_reg_mr,
88         [IB_USER_VERBS_CMD_DEREG_MR]            = ib_uverbs_dereg_mr,
89         [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
90         [IB_USER_VERBS_CMD_CREATE_CQ]           = ib_uverbs_create_cq,
91         [IB_USER_VERBS_CMD_RESIZE_CQ]           = ib_uverbs_resize_cq,
92         [IB_USER_VERBS_CMD_POLL_CQ]             = ib_uverbs_poll_cq,
93         [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]       = ib_uverbs_req_notify_cq,
94         [IB_USER_VERBS_CMD_DESTROY_CQ]          = ib_uverbs_destroy_cq,
95         [IB_USER_VERBS_CMD_CREATE_QP]           = ib_uverbs_create_qp,
96         [IB_USER_VERBS_CMD_QUERY_QP]            = ib_uverbs_query_qp,
97         [IB_USER_VERBS_CMD_MODIFY_QP]           = ib_uverbs_modify_qp,
98         [IB_USER_VERBS_CMD_DESTROY_QP]          = ib_uverbs_destroy_qp,
99         [IB_USER_VERBS_CMD_POST_SEND]           = ib_uverbs_post_send,
100         [IB_USER_VERBS_CMD_POST_RECV]           = ib_uverbs_post_recv,
101         [IB_USER_VERBS_CMD_POST_SRQ_RECV]       = ib_uverbs_post_srq_recv,
102         [IB_USER_VERBS_CMD_CREATE_AH]           = ib_uverbs_create_ah,
103         [IB_USER_VERBS_CMD_DESTROY_AH]          = ib_uverbs_destroy_ah,
104         [IB_USER_VERBS_CMD_ATTACH_MCAST]        = ib_uverbs_attach_mcast,
105         [IB_USER_VERBS_CMD_DETACH_MCAST]        = ib_uverbs_detach_mcast,
106         [IB_USER_VERBS_CMD_CREATE_SRQ]          = ib_uverbs_create_srq,
107         [IB_USER_VERBS_CMD_MODIFY_SRQ]          = ib_uverbs_modify_srq,
108         [IB_USER_VERBS_CMD_QUERY_SRQ]           = ib_uverbs_query_srq,
109         [IB_USER_VERBS_CMD_DESTROY_SRQ]         = ib_uverbs_destroy_srq,
110 };
111
112 static struct vfsmount *uverbs_event_mnt;
113
114 static void ib_uverbs_add_one(struct ib_device *device);
115 static void ib_uverbs_remove_one(struct ib_device *device);
116
117 static void ib_uverbs_release_dev(struct kref *ref)
118 {
119         struct ib_uverbs_device *dev =
120                 container_of(ref, struct ib_uverbs_device, ref);
121
122         complete(&dev->comp);
123 }
124
125 static void ib_uverbs_release_event_file(struct kref *ref)
126 {
127         struct ib_uverbs_event_file *file =
128                 container_of(ref, struct ib_uverbs_event_file, ref);
129
130         kfree(file);
131 }
132
133 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
134                           struct ib_uverbs_event_file *ev_file,
135                           struct ib_ucq_object *uobj)
136 {
137         struct ib_uverbs_event *evt, *tmp;
138
139         if (ev_file) {
140                 spin_lock_irq(&ev_file->lock);
141                 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
142                         list_del(&evt->list);
143                         kfree(evt);
144                 }
145                 spin_unlock_irq(&ev_file->lock);
146
147                 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
148         }
149
150         spin_lock_irq(&file->async_file->lock);
151         list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
152                 list_del(&evt->list);
153                 kfree(evt);
154         }
155         spin_unlock_irq(&file->async_file->lock);
156 }
157
158 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
159                               struct ib_uevent_object *uobj)
160 {
161         struct ib_uverbs_event *evt, *tmp;
162
163         spin_lock_irq(&file->async_file->lock);
164         list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
165                 list_del(&evt->list);
166                 kfree(evt);
167         }
168         spin_unlock_irq(&file->async_file->lock);
169 }
170
171 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
172                                     struct ib_uqp_object *uobj)
173 {
174         struct ib_uverbs_mcast_entry *mcast, *tmp;
175
176         list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
177                 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
178                 list_del(&mcast->list);
179                 kfree(mcast);
180         }
181 }
182
183 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
184                                       struct ib_ucontext *context)
185 {
186         struct ib_uobject *uobj, *tmp;
187
188         if (!context)
189                 return 0;
190
191         context->closing = 1;
192
193         list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
194                 struct ib_ah *ah = uobj->object;
195
196                 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
197                 ib_destroy_ah(ah);
198                 kfree(uobj);
199         }
200
201         list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
202                 struct ib_qp *qp = uobj->object;
203                 struct ib_uqp_object *uqp =
204                         container_of(uobj, struct ib_uqp_object, uevent.uobject);
205
206                 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
207                 ib_uverbs_detach_umcast(qp, uqp);
208                 ib_destroy_qp(qp);
209                 ib_uverbs_release_uevent(file, &uqp->uevent);
210                 kfree(uqp);
211         }
212
213         list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
214                 struct ib_cq *cq = uobj->object;
215                 struct ib_uverbs_event_file *ev_file = cq->cq_context;
216                 struct ib_ucq_object *ucq =
217                         container_of(uobj, struct ib_ucq_object, uobject);
218
219                 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
220                 ib_destroy_cq(cq);
221                 ib_uverbs_release_ucq(file, ev_file, ucq);
222                 kfree(ucq);
223         }
224
225         list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
226                 struct ib_srq *srq = uobj->object;
227                 struct ib_uevent_object *uevent =
228                         container_of(uobj, struct ib_uevent_object, uobject);
229
230                 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
231                 ib_destroy_srq(srq);
232                 ib_uverbs_release_uevent(file, uevent);
233                 kfree(uevent);
234         }
235
236         /* XXX Free MWs */
237
238         list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
239                 struct ib_mr *mr = uobj->object;
240
241                 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
242                 ib_dereg_mr(mr);
243                 kfree(uobj);
244         }
245
246         list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
247                 struct ib_pd *pd = uobj->object;
248
249                 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
250                 ib_dealloc_pd(pd);
251                 kfree(uobj);
252         }
253
254         return context->device->dealloc_ucontext(context);
255 }
256
257 static void ib_uverbs_release_file(struct kref *ref)
258 {
259         struct ib_uverbs_file *file =
260                 container_of(ref, struct ib_uverbs_file, ref);
261
262         module_put(file->device->ib_dev->owner);
263         kref_put(&file->device->ref, ib_uverbs_release_dev);
264
265         kfree(file);
266 }
267
268 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
269                                     size_t count, loff_t *pos)
270 {
271         struct ib_uverbs_event_file *file = filp->private_data;
272         struct ib_uverbs_event *event;
273         int eventsz;
274         int ret = 0;
275
276         spin_lock_irq(&file->lock);
277
278         while (list_empty(&file->event_list)) {
279                 spin_unlock_irq(&file->lock);
280
281                 if (filp->f_flags & O_NONBLOCK)
282                         return -EAGAIN;
283
284                 if (wait_event_interruptible(file->poll_wait,
285                                              !list_empty(&file->event_list)))
286                         return -ERESTARTSYS;
287
288                 spin_lock_irq(&file->lock);
289         }
290
291         event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
292
293         if (file->is_async)
294                 eventsz = sizeof (struct ib_uverbs_async_event_desc);
295         else
296                 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
297
298         if (eventsz > count) {
299                 ret   = -EINVAL;
300                 event = NULL;
301         } else {
302                 list_del(file->event_list.next);
303                 if (event->counter) {
304                         ++(*event->counter);
305                         list_del(&event->obj_list);
306                 }
307         }
308
309         spin_unlock_irq(&file->lock);
310
311         if (event) {
312                 if (copy_to_user(buf, event, eventsz))
313                         ret = -EFAULT;
314                 else
315                         ret = eventsz;
316         }
317
318         kfree(event);
319
320         return ret;
321 }
322
323 static unsigned int ib_uverbs_event_poll(struct file *filp,
324                                          struct poll_table_struct *wait)
325 {
326         unsigned int pollflags = 0;
327         struct ib_uverbs_event_file *file = filp->private_data;
328
329         poll_wait(filp, &file->poll_wait, wait);
330
331         spin_lock_irq(&file->lock);
332         if (!list_empty(&file->event_list))
333                 pollflags = POLLIN | POLLRDNORM;
334         spin_unlock_irq(&file->lock);
335
336         return pollflags;
337 }
338
339 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
340 {
341         struct ib_uverbs_event_file *file = filp->private_data;
342
343         return fasync_helper(fd, filp, on, &file->async_queue);
344 }
345
346 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
347 {
348         struct ib_uverbs_event_file *file = filp->private_data;
349         struct ib_uverbs_event *entry, *tmp;
350
351         spin_lock_irq(&file->lock);
352         file->is_closed = 1;
353         list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
354                 if (entry->counter)
355                         list_del(&entry->obj_list);
356                 kfree(entry);
357         }
358         spin_unlock_irq(&file->lock);
359
360         if (file->is_async) {
361                 ib_unregister_event_handler(&file->uverbs_file->event_handler);
362                 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
363         }
364         kref_put(&file->ref, ib_uverbs_release_event_file);
365
366         return 0;
367 }
368
369 static const struct file_operations uverbs_event_fops = {
370         .owner   = THIS_MODULE,
371         .read    = ib_uverbs_event_read,
372         .poll    = ib_uverbs_event_poll,
373         .release = ib_uverbs_event_close,
374         .fasync  = ib_uverbs_event_fasync
375 };
376
377 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
378 {
379         struct ib_uverbs_event_file    *file = cq_context;
380         struct ib_ucq_object           *uobj;
381         struct ib_uverbs_event         *entry;
382         unsigned long                   flags;
383
384         if (!file)
385                 return;
386
387         spin_lock_irqsave(&file->lock, flags);
388         if (file->is_closed) {
389                 spin_unlock_irqrestore(&file->lock, flags);
390                 return;
391         }
392
393         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
394         if (!entry) {
395                 spin_unlock_irqrestore(&file->lock, flags);
396                 return;
397         }
398
399         uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
400
401         entry->desc.comp.cq_handle = cq->uobject->user_handle;
402         entry->counter             = &uobj->comp_events_reported;
403
404         list_add_tail(&entry->list, &file->event_list);
405         list_add_tail(&entry->obj_list, &uobj->comp_list);
406         spin_unlock_irqrestore(&file->lock, flags);
407
408         wake_up_interruptible(&file->poll_wait);
409         kill_fasync(&file->async_queue, SIGIO, POLL_IN);
410 }
411
412 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
413                                     __u64 element, __u64 event,
414                                     struct list_head *obj_list,
415                                     u32 *counter)
416 {
417         struct ib_uverbs_event *entry;
418         unsigned long flags;
419
420         spin_lock_irqsave(&file->async_file->lock, flags);
421         if (file->async_file->is_closed) {
422                 spin_unlock_irqrestore(&file->async_file->lock, flags);
423                 return;
424         }
425
426         entry = kmalloc(sizeof *entry, GFP_ATOMIC);
427         if (!entry) {
428                 spin_unlock_irqrestore(&file->async_file->lock, flags);
429                 return;
430         }
431
432         entry->desc.async.element    = element;
433         entry->desc.async.event_type = event;
434         entry->counter               = counter;
435
436         list_add_tail(&entry->list, &file->async_file->event_list);
437         if (obj_list)
438                 list_add_tail(&entry->obj_list, obj_list);
439         spin_unlock_irqrestore(&file->async_file->lock, flags);
440
441         wake_up_interruptible(&file->async_file->poll_wait);
442         kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
443 }
444
445 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
446 {
447         struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
448                                                   struct ib_ucq_object, uobject);
449
450         ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
451                                 event->event, &uobj->async_list,
452                                 &uobj->async_events_reported);
453 }
454
455 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
456 {
457         struct ib_uevent_object *uobj;
458
459         uobj = container_of(event->element.qp->uobject,
460                             struct ib_uevent_object, uobject);
461
462         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
463                                 event->event, &uobj->event_list,
464                                 &uobj->events_reported);
465 }
466
467 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
468 {
469         struct ib_uevent_object *uobj;
470
471         uobj = container_of(event->element.srq->uobject,
472                             struct ib_uevent_object, uobject);
473
474         ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
475                                 event->event, &uobj->event_list,
476                                 &uobj->events_reported);
477 }
478
479 void ib_uverbs_event_handler(struct ib_event_handler *handler,
480                              struct ib_event *event)
481 {
482         struct ib_uverbs_file *file =
483                 container_of(handler, struct ib_uverbs_file, event_handler);
484
485         ib_uverbs_async_handler(file, event->element.port_num, event->event,
486                                 NULL, NULL);
487 }
488
489 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
490                                         int is_async, int *fd)
491 {
492         struct ib_uverbs_event_file *ev_file;
493         struct path path;
494         struct file *filp;
495         int ret;
496
497         ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
498         if (!ev_file)
499                 return ERR_PTR(-ENOMEM);
500
501         kref_init(&ev_file->ref);
502         spin_lock_init(&ev_file->lock);
503         INIT_LIST_HEAD(&ev_file->event_list);
504         init_waitqueue_head(&ev_file->poll_wait);
505         ev_file->uverbs_file = uverbs_file;
506         ev_file->async_queue = NULL;
507         ev_file->is_async    = is_async;
508         ev_file->is_closed   = 0;
509
510         *fd = get_unused_fd();
511         if (*fd < 0) {
512                 ret = *fd;
513                 goto err;
514         }
515
516         /*
517          * fops_get() can't fail here, because we're coming from a
518          * system call on a uverbs file, which will already have a
519          * module reference.
520          */
521         path.mnt = uverbs_event_mnt;
522         path.dentry = uverbs_event_mnt->mnt_root;
523         path_get(&path);
524         filp = alloc_file(&path, FMODE_READ, fops_get(&uverbs_event_fops));
525         if (!filp) {
526                 ret = -ENFILE;
527                 goto err_fd;
528         }
529
530         filp->private_data = ev_file;
531
532         return filp;
533
534 err_fd:
535         fops_put(&uverbs_event_fops);
536         path_put(&path);
537         put_unused_fd(*fd);
538
539 err:
540         kfree(ev_file);
541         return ERR_PTR(ret);
542 }
543
544 /*
545  * Look up a completion event file by FD.  If lookup is successful,
546  * takes a ref to the event file struct that it returns; if
547  * unsuccessful, returns NULL.
548  */
549 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
550 {
551         struct ib_uverbs_event_file *ev_file = NULL;
552         struct file *filp;
553
554         filp = fget(fd);
555         if (!filp)
556                 return NULL;
557
558         if (filp->f_op != &uverbs_event_fops)
559                 goto out;
560
561         ev_file = filp->private_data;
562         if (ev_file->is_async) {
563                 ev_file = NULL;
564                 goto out;
565         }
566
567         kref_get(&ev_file->ref);
568
569 out:
570         fput(filp);
571         return ev_file;
572 }
573
574 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
575                              size_t count, loff_t *pos)
576 {
577         struct ib_uverbs_file *file = filp->private_data;
578         struct ib_uverbs_cmd_hdr hdr;
579
580         if (count < sizeof hdr)
581                 return -EINVAL;
582
583         if (copy_from_user(&hdr, buf, sizeof hdr))
584                 return -EFAULT;
585
586         if (hdr.in_words * 4 != count)
587                 return -EINVAL;
588
589         if (hdr.command < 0                             ||
590             hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
591             !uverbs_cmd_table[hdr.command])
592                 return -EINVAL;
593
594         if (!file->ucontext &&
595             hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
596                 return -EINVAL;
597
598         if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
599                 return -ENOSYS;
600
601         return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
602                                              hdr.in_words * 4, hdr.out_words * 4);
603 }
604
605 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
606 {
607         struct ib_uverbs_file *file = filp->private_data;
608
609         if (!file->ucontext)
610                 return -ENODEV;
611         else
612                 return file->device->ib_dev->mmap(file->ucontext, vma);
613 }
614
615 /*
616  * ib_uverbs_open() does not need the BKL:
617  *
618  *  - the ib_uverbs_device structures are properly reference counted and
619  *    everything else is purely local to the file being created, so
620  *    races against other open calls are not a problem;
621  *  - there is no ioctl method to race against;
622  *  - the open method will either immediately run -ENXIO, or all
623  *    required initialization will be done.
624  */
625 static int ib_uverbs_open(struct inode *inode, struct file *filp)
626 {
627         struct ib_uverbs_device *dev;
628         struct ib_uverbs_file *file;
629         int ret;
630
631         dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
632         if (dev)
633                 kref_get(&dev->ref);
634         else
635                 return -ENXIO;
636
637         if (!try_module_get(dev->ib_dev->owner)) {
638                 ret = -ENODEV;
639                 goto err;
640         }
641
642         file = kmalloc(sizeof *file, GFP_KERNEL);
643         if (!file) {
644                 ret = -ENOMEM;
645                 goto err_module;
646         }
647
648         file->device     = dev;
649         file->ucontext   = NULL;
650         file->async_file = NULL;
651         kref_init(&file->ref);
652         mutex_init(&file->mutex);
653
654         filp->private_data = file;
655
656         return 0;
657
658 err_module:
659         module_put(dev->ib_dev->owner);
660
661 err:
662         kref_put(&dev->ref, ib_uverbs_release_dev);
663         return ret;
664 }
665
666 static int ib_uverbs_close(struct inode *inode, struct file *filp)
667 {
668         struct ib_uverbs_file *file = filp->private_data;
669
670         ib_uverbs_cleanup_ucontext(file, file->ucontext);
671
672         if (file->async_file)
673                 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
674
675         kref_put(&file->ref, ib_uverbs_release_file);
676
677         return 0;
678 }
679
680 static const struct file_operations uverbs_fops = {
681         .owner   = THIS_MODULE,
682         .write   = ib_uverbs_write,
683         .open    = ib_uverbs_open,
684         .release = ib_uverbs_close
685 };
686
687 static const struct file_operations uverbs_mmap_fops = {
688         .owner   = THIS_MODULE,
689         .write   = ib_uverbs_write,
690         .mmap    = ib_uverbs_mmap,
691         .open    = ib_uverbs_open,
692         .release = ib_uverbs_close
693 };
694
695 static struct ib_client uverbs_client = {
696         .name   = "uverbs",
697         .add    = ib_uverbs_add_one,
698         .remove = ib_uverbs_remove_one
699 };
700
701 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
702                           char *buf)
703 {
704         struct ib_uverbs_device *dev = dev_get_drvdata(device);
705
706         if (!dev)
707                 return -ENODEV;
708
709         return sprintf(buf, "%s\n", dev->ib_dev->name);
710 }
711 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
712
713 static ssize_t show_dev_abi_version(struct device *device,
714                                     struct device_attribute *attr, char *buf)
715 {
716         struct ib_uverbs_device *dev = dev_get_drvdata(device);
717
718         if (!dev)
719                 return -ENODEV;
720
721         return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
722 }
723 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
724
725 static ssize_t show_abi_version(struct class *class, char *buf)
726 {
727         return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
728 }
729 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
730
731 static dev_t overflow_maj;
732 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
733
734 /*
735  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
736  * requesting a new major number and doubling the number of max devices we
737  * support. It's stupid, but simple.
738  */
739 static int find_overflow_devnum(void)
740 {
741         int ret;
742
743         if (!overflow_maj) {
744                 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
745                                           "infiniband_verbs");
746                 if (ret) {
747                         printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
748                         return ret;
749                 }
750         }
751
752         ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
753         if (ret >= IB_UVERBS_MAX_DEVICES)
754                 return -1;
755
756         return ret;
757 }
758
759 static void ib_uverbs_add_one(struct ib_device *device)
760 {
761         int devnum;
762         dev_t base;
763         struct ib_uverbs_device *uverbs_dev;
764
765         if (!device->alloc_ucontext)
766                 return;
767
768         uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
769         if (!uverbs_dev)
770                 return;
771
772         kref_init(&uverbs_dev->ref);
773         init_completion(&uverbs_dev->comp);
774
775         spin_lock(&map_lock);
776         devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
777         if (devnum >= IB_UVERBS_MAX_DEVICES) {
778                 spin_unlock(&map_lock);
779                 devnum = find_overflow_devnum();
780                 if (devnum < 0)
781                         goto err;
782
783                 spin_lock(&map_lock);
784                 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
785                 base = devnum + overflow_maj;
786                 set_bit(devnum, overflow_map);
787         } else {
788                 uverbs_dev->devnum = devnum;
789                 base = devnum + IB_UVERBS_BASE_DEV;
790                 set_bit(devnum, dev_map);
791         }
792         spin_unlock(&map_lock);
793
794         uverbs_dev->ib_dev           = device;
795         uverbs_dev->num_comp_vectors = device->num_comp_vectors;
796
797         cdev_init(&uverbs_dev->cdev, NULL);
798         uverbs_dev->cdev.owner = THIS_MODULE;
799         uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
800         kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
801         if (cdev_add(&uverbs_dev->cdev, base, 1))
802                 goto err_cdev;
803
804         uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
805                                         uverbs_dev->cdev.dev, uverbs_dev,
806                                         "uverbs%d", uverbs_dev->devnum);
807         if (IS_ERR(uverbs_dev->dev))
808                 goto err_cdev;
809
810         if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
811                 goto err_class;
812         if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
813                 goto err_class;
814
815         ib_set_client_data(device, &uverbs_client, uverbs_dev);
816
817         return;
818
819 err_class:
820         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
821
822 err_cdev:
823         cdev_del(&uverbs_dev->cdev);
824         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
825                 clear_bit(devnum, dev_map);
826         else
827                 clear_bit(devnum, overflow_map);
828
829 err:
830         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
831         wait_for_completion(&uverbs_dev->comp);
832         kfree(uverbs_dev);
833         return;
834 }
835
836 static void ib_uverbs_remove_one(struct ib_device *device)
837 {
838         struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
839
840         if (!uverbs_dev)
841                 return;
842
843         dev_set_drvdata(uverbs_dev->dev, NULL);
844         device_destroy(uverbs_class, uverbs_dev->cdev.dev);
845         cdev_del(&uverbs_dev->cdev);
846
847         if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
848                 clear_bit(uverbs_dev->devnum, dev_map);
849         else
850                 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
851
852         kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
853         wait_for_completion(&uverbs_dev->comp);
854         kfree(uverbs_dev);
855 }
856
857 static int uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
858                                const char *dev_name, void *data,
859                                struct vfsmount *mnt)
860 {
861         return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
862                              INFINIBANDEVENTFS_MAGIC, mnt);
863 }
864
865 static struct file_system_type uverbs_event_fs = {
866         /* No owner field so module can be unloaded */
867         .name    = "infinibandeventfs",
868         .get_sb  = uverbs_event_get_sb,
869         .kill_sb = kill_litter_super
870 };
871
872 static int __init ib_uverbs_init(void)
873 {
874         int ret;
875
876         ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
877                                      "infiniband_verbs");
878         if (ret) {
879                 printk(KERN_ERR "user_verbs: couldn't register device number\n");
880                 goto out;
881         }
882
883         uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
884         if (IS_ERR(uverbs_class)) {
885                 ret = PTR_ERR(uverbs_class);
886                 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
887                 goto out_chrdev;
888         }
889
890         ret = class_create_file(uverbs_class, &class_attr_abi_version);
891         if (ret) {
892                 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
893                 goto out_class;
894         }
895
896         ret = register_filesystem(&uverbs_event_fs);
897         if (ret) {
898                 printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
899                 goto out_class;
900         }
901
902         uverbs_event_mnt = kern_mount(&uverbs_event_fs);
903         if (IS_ERR(uverbs_event_mnt)) {
904                 ret = PTR_ERR(uverbs_event_mnt);
905                 printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
906                 goto out_fs;
907         }
908
909         ret = ib_register_client(&uverbs_client);
910         if (ret) {
911                 printk(KERN_ERR "user_verbs: couldn't register client\n");
912                 goto out_mnt;
913         }
914
915         return 0;
916
917 out_mnt:
918         mntput(uverbs_event_mnt);
919
920 out_fs:
921         unregister_filesystem(&uverbs_event_fs);
922
923 out_class:
924         class_destroy(uverbs_class);
925
926 out_chrdev:
927         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
928
929 out:
930         return ret;
931 }
932
933 static void __exit ib_uverbs_cleanup(void)
934 {
935         ib_unregister_client(&uverbs_client);
936         mntput(uverbs_event_mnt);
937         unregister_filesystem(&uverbs_event_fs);
938         class_destroy(uverbs_class);
939         unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
940         if (overflow_maj)
941                 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
942         idr_destroy(&ib_uverbs_pd_idr);
943         idr_destroy(&ib_uverbs_mr_idr);
944         idr_destroy(&ib_uverbs_mw_idr);
945         idr_destroy(&ib_uverbs_ah_idr);
946         idr_destroy(&ib_uverbs_cq_idr);
947         idr_destroy(&ib_uverbs_qp_idr);
948         idr_destroy(&ib_uverbs_srq_idr);
949 }
950
951 module_init(ib_uverbs_init);
952 module_exit(ib_uverbs_cleanup);