Loading...
1/*
2 * Copyright (c) 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5 * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
6 * Copyright (c) 2005 PathScale, Inc. All rights reserved.
7 *
8 * This software is available to you under a choice of one of two
9 * licenses. You may choose to be licensed under the terms of the GNU
10 * General Public License (GPL) Version 2, available from the file
11 * COPYING in the main directory of this source tree, or the
12 * OpenIB.org BSD license below:
13 *
14 * Redistribution and use in source and binary forms, with or
15 * without modification, are permitted provided that the following
16 * conditions are met:
17 *
18 * - Redistributions of source code must retain the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer.
21 *
22 * - Redistributions in binary form must reproduce the above
23 * copyright notice, this list of conditions and the following
24 * disclaimer in the documentation and/or other materials
25 * provided with the distribution.
26 *
27 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 * SOFTWARE.
35 */
36
37#include <linux/module.h>
38#include <linux/init.h>
39#include <linux/device.h>
40#include <linux/err.h>
41#include <linux/fs.h>
42#include <linux/poll.h>
43#include <linux/sched.h>
44#include <linux/file.h>
45#include <linux/cdev.h>
46#include <linux/anon_inodes.h>
47#include <linux/slab.h>
48
49#include <asm/uaccess.h>
50
51#include "uverbs.h"
52
53MODULE_AUTHOR("Roland Dreier");
54MODULE_DESCRIPTION("InfiniBand userspace verbs access");
55MODULE_LICENSE("Dual BSD/GPL");
56
57enum {
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
65static struct class *uverbs_class;
66
67DEFINE_SPINLOCK(ib_uverbs_idr_lock);
68DEFINE_IDR(ib_uverbs_pd_idr);
69DEFINE_IDR(ib_uverbs_mr_idr);
70DEFINE_IDR(ib_uverbs_mw_idr);
71DEFINE_IDR(ib_uverbs_ah_idr);
72DEFINE_IDR(ib_uverbs_cq_idr);
73DEFINE_IDR(ib_uverbs_qp_idr);
74DEFINE_IDR(ib_uverbs_srq_idr);
75
76static DEFINE_SPINLOCK(map_lock);
77static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
78
79static 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
112static void ib_uverbs_add_one(struct ib_device *device);
113static void ib_uverbs_remove_one(struct ib_device *device);
114
115static void ib_uverbs_release_dev(struct kref *ref)
116{
117 struct ib_uverbs_device *dev =
118 container_of(ref, struct ib_uverbs_device, ref);
119
120 complete(&dev->comp);
121}
122
123static void ib_uverbs_release_event_file(struct kref *ref)
124{
125 struct ib_uverbs_event_file *file =
126 container_of(ref, struct ib_uverbs_event_file, ref);
127
128 kfree(file);
129}
130
131void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
132 struct ib_uverbs_event_file *ev_file,
133 struct ib_ucq_object *uobj)
134{
135 struct ib_uverbs_event *evt, *tmp;
136
137 if (ev_file) {
138 spin_lock_irq(&ev_file->lock);
139 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
140 list_del(&evt->list);
141 kfree(evt);
142 }
143 spin_unlock_irq(&ev_file->lock);
144
145 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
146 }
147
148 spin_lock_irq(&file->async_file->lock);
149 list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
150 list_del(&evt->list);
151 kfree(evt);
152 }
153 spin_unlock_irq(&file->async_file->lock);
154}
155
156void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
157 struct ib_uevent_object *uobj)
158{
159 struct ib_uverbs_event *evt, *tmp;
160
161 spin_lock_irq(&file->async_file->lock);
162 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
163 list_del(&evt->list);
164 kfree(evt);
165 }
166 spin_unlock_irq(&file->async_file->lock);
167}
168
169static void ib_uverbs_detach_umcast(struct ib_qp *qp,
170 struct ib_uqp_object *uobj)
171{
172 struct ib_uverbs_mcast_entry *mcast, *tmp;
173
174 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
175 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
176 list_del(&mcast->list);
177 kfree(mcast);
178 }
179}
180
181static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
182 struct ib_ucontext *context)
183{
184 struct ib_uobject *uobj, *tmp;
185
186 if (!context)
187 return 0;
188
189 context->closing = 1;
190
191 list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
192 struct ib_ah *ah = uobj->object;
193
194 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
195 ib_destroy_ah(ah);
196 kfree(uobj);
197 }
198
199 list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
200 struct ib_qp *qp = uobj->object;
201 struct ib_uqp_object *uqp =
202 container_of(uobj, struct ib_uqp_object, uevent.uobject);
203
204 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
205 ib_uverbs_detach_umcast(qp, uqp);
206 ib_destroy_qp(qp);
207 ib_uverbs_release_uevent(file, &uqp->uevent);
208 kfree(uqp);
209 }
210
211 list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
212 struct ib_cq *cq = uobj->object;
213 struct ib_uverbs_event_file *ev_file = cq->cq_context;
214 struct ib_ucq_object *ucq =
215 container_of(uobj, struct ib_ucq_object, uobject);
216
217 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
218 ib_destroy_cq(cq);
219 ib_uverbs_release_ucq(file, ev_file, ucq);
220 kfree(ucq);
221 }
222
223 list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
224 struct ib_srq *srq = uobj->object;
225 struct ib_uevent_object *uevent =
226 container_of(uobj, struct ib_uevent_object, uobject);
227
228 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
229 ib_destroy_srq(srq);
230 ib_uverbs_release_uevent(file, uevent);
231 kfree(uevent);
232 }
233
234 /* XXX Free MWs */
235
236 list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
237 struct ib_mr *mr = uobj->object;
238
239 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
240 ib_dereg_mr(mr);
241 kfree(uobj);
242 }
243
244 list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
245 struct ib_pd *pd = uobj->object;
246
247 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
248 ib_dealloc_pd(pd);
249 kfree(uobj);
250 }
251
252 return context->device->dealloc_ucontext(context);
253}
254
255static void ib_uverbs_release_file(struct kref *ref)
256{
257 struct ib_uverbs_file *file =
258 container_of(ref, struct ib_uverbs_file, ref);
259
260 module_put(file->device->ib_dev->owner);
261 kref_put(&file->device->ref, ib_uverbs_release_dev);
262
263 kfree(file);
264}
265
266static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
267 size_t count, loff_t *pos)
268{
269 struct ib_uverbs_event_file *file = filp->private_data;
270 struct ib_uverbs_event *event;
271 int eventsz;
272 int ret = 0;
273
274 spin_lock_irq(&file->lock);
275
276 while (list_empty(&file->event_list)) {
277 spin_unlock_irq(&file->lock);
278
279 if (filp->f_flags & O_NONBLOCK)
280 return -EAGAIN;
281
282 if (wait_event_interruptible(file->poll_wait,
283 !list_empty(&file->event_list)))
284 return -ERESTARTSYS;
285
286 spin_lock_irq(&file->lock);
287 }
288
289 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
290
291 if (file->is_async)
292 eventsz = sizeof (struct ib_uverbs_async_event_desc);
293 else
294 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
295
296 if (eventsz > count) {
297 ret = -EINVAL;
298 event = NULL;
299 } else {
300 list_del(file->event_list.next);
301 if (event->counter) {
302 ++(*event->counter);
303 list_del(&event->obj_list);
304 }
305 }
306
307 spin_unlock_irq(&file->lock);
308
309 if (event) {
310 if (copy_to_user(buf, event, eventsz))
311 ret = -EFAULT;
312 else
313 ret = eventsz;
314 }
315
316 kfree(event);
317
318 return ret;
319}
320
321static unsigned int ib_uverbs_event_poll(struct file *filp,
322 struct poll_table_struct *wait)
323{
324 unsigned int pollflags = 0;
325 struct ib_uverbs_event_file *file = filp->private_data;
326
327 poll_wait(filp, &file->poll_wait, wait);
328
329 spin_lock_irq(&file->lock);
330 if (!list_empty(&file->event_list))
331 pollflags = POLLIN | POLLRDNORM;
332 spin_unlock_irq(&file->lock);
333
334 return pollflags;
335}
336
337static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
338{
339 struct ib_uverbs_event_file *file = filp->private_data;
340
341 return fasync_helper(fd, filp, on, &file->async_queue);
342}
343
344static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
345{
346 struct ib_uverbs_event_file *file = filp->private_data;
347 struct ib_uverbs_event *entry, *tmp;
348
349 spin_lock_irq(&file->lock);
350 file->is_closed = 1;
351 list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
352 if (entry->counter)
353 list_del(&entry->obj_list);
354 kfree(entry);
355 }
356 spin_unlock_irq(&file->lock);
357
358 if (file->is_async) {
359 ib_unregister_event_handler(&file->uverbs_file->event_handler);
360 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
361 }
362 kref_put(&file->ref, ib_uverbs_release_event_file);
363
364 return 0;
365}
366
367static const struct file_operations uverbs_event_fops = {
368 .owner = THIS_MODULE,
369 .read = ib_uverbs_event_read,
370 .poll = ib_uverbs_event_poll,
371 .release = ib_uverbs_event_close,
372 .fasync = ib_uverbs_event_fasync,
373 .llseek = no_llseek,
374};
375
376void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
377{
378 struct ib_uverbs_event_file *file = cq_context;
379 struct ib_ucq_object *uobj;
380 struct ib_uverbs_event *entry;
381 unsigned long flags;
382
383 if (!file)
384 return;
385
386 spin_lock_irqsave(&file->lock, flags);
387 if (file->is_closed) {
388 spin_unlock_irqrestore(&file->lock, flags);
389 return;
390 }
391
392 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
393 if (!entry) {
394 spin_unlock_irqrestore(&file->lock, flags);
395 return;
396 }
397
398 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
399
400 entry->desc.comp.cq_handle = cq->uobject->user_handle;
401 entry->counter = &uobj->comp_events_reported;
402
403 list_add_tail(&entry->list, &file->event_list);
404 list_add_tail(&entry->obj_list, &uobj->comp_list);
405 spin_unlock_irqrestore(&file->lock, flags);
406
407 wake_up_interruptible(&file->poll_wait);
408 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
409}
410
411static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
412 __u64 element, __u64 event,
413 struct list_head *obj_list,
414 u32 *counter)
415{
416 struct ib_uverbs_event *entry;
417 unsigned long flags;
418
419 spin_lock_irqsave(&file->async_file->lock, flags);
420 if (file->async_file->is_closed) {
421 spin_unlock_irqrestore(&file->async_file->lock, flags);
422 return;
423 }
424
425 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
426 if (!entry) {
427 spin_unlock_irqrestore(&file->async_file->lock, flags);
428 return;
429 }
430
431 entry->desc.async.element = element;
432 entry->desc.async.event_type = event;
433 entry->counter = counter;
434
435 list_add_tail(&entry->list, &file->async_file->event_list);
436 if (obj_list)
437 list_add_tail(&entry->obj_list, obj_list);
438 spin_unlock_irqrestore(&file->async_file->lock, flags);
439
440 wake_up_interruptible(&file->async_file->poll_wait);
441 kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
442}
443
444void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
445{
446 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
447 struct ib_ucq_object, uobject);
448
449 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
450 event->event, &uobj->async_list,
451 &uobj->async_events_reported);
452}
453
454void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
455{
456 struct ib_uevent_object *uobj;
457
458 uobj = container_of(event->element.qp->uobject,
459 struct ib_uevent_object, uobject);
460
461 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
462 event->event, &uobj->event_list,
463 &uobj->events_reported);
464}
465
466void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
467{
468 struct ib_uevent_object *uobj;
469
470 uobj = container_of(event->element.srq->uobject,
471 struct ib_uevent_object, uobject);
472
473 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
474 event->event, &uobj->event_list,
475 &uobj->events_reported);
476}
477
478void ib_uverbs_event_handler(struct ib_event_handler *handler,
479 struct ib_event *event)
480{
481 struct ib_uverbs_file *file =
482 container_of(handler, struct ib_uverbs_file, event_handler);
483
484 ib_uverbs_async_handler(file, event->element.port_num, event->event,
485 NULL, NULL);
486}
487
488struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
489 int is_async)
490{
491 struct ib_uverbs_event_file *ev_file;
492 struct file *filp;
493
494 ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
495 if (!ev_file)
496 return ERR_PTR(-ENOMEM);
497
498 kref_init(&ev_file->ref);
499 spin_lock_init(&ev_file->lock);
500 INIT_LIST_HEAD(&ev_file->event_list);
501 init_waitqueue_head(&ev_file->poll_wait);
502 ev_file->uverbs_file = uverbs_file;
503 ev_file->async_queue = NULL;
504 ev_file->is_async = is_async;
505 ev_file->is_closed = 0;
506
507 filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
508 ev_file, O_RDONLY);
509 if (IS_ERR(filp))
510 kfree(ev_file);
511
512 return filp;
513}
514
515/*
516 * Look up a completion event file by FD. If lookup is successful,
517 * takes a ref to the event file struct that it returns; if
518 * unsuccessful, returns NULL.
519 */
520struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
521{
522 struct ib_uverbs_event_file *ev_file = NULL;
523 struct file *filp;
524
525 filp = fget(fd);
526 if (!filp)
527 return NULL;
528
529 if (filp->f_op != &uverbs_event_fops)
530 goto out;
531
532 ev_file = filp->private_data;
533 if (ev_file->is_async) {
534 ev_file = NULL;
535 goto out;
536 }
537
538 kref_get(&ev_file->ref);
539
540out:
541 fput(filp);
542 return ev_file;
543}
544
545static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
546 size_t count, loff_t *pos)
547{
548 struct ib_uverbs_file *file = filp->private_data;
549 struct ib_uverbs_cmd_hdr hdr;
550
551 if (count < sizeof hdr)
552 return -EINVAL;
553
554 if (copy_from_user(&hdr, buf, sizeof hdr))
555 return -EFAULT;
556
557 if (hdr.in_words * 4 != count)
558 return -EINVAL;
559
560 if (hdr.command < 0 ||
561 hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
562 !uverbs_cmd_table[hdr.command])
563 return -EINVAL;
564
565 if (!file->ucontext &&
566 hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
567 return -EINVAL;
568
569 if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
570 return -ENOSYS;
571
572 return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
573 hdr.in_words * 4, hdr.out_words * 4);
574}
575
576static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
577{
578 struct ib_uverbs_file *file = filp->private_data;
579
580 if (!file->ucontext)
581 return -ENODEV;
582 else
583 return file->device->ib_dev->mmap(file->ucontext, vma);
584}
585
586/*
587 * ib_uverbs_open() does not need the BKL:
588 *
589 * - the ib_uverbs_device structures are properly reference counted and
590 * everything else is purely local to the file being created, so
591 * races against other open calls are not a problem;
592 * - there is no ioctl method to race against;
593 * - the open method will either immediately run -ENXIO, or all
594 * required initialization will be done.
595 */
596static int ib_uverbs_open(struct inode *inode, struct file *filp)
597{
598 struct ib_uverbs_device *dev;
599 struct ib_uverbs_file *file;
600 int ret;
601
602 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
603 if (dev)
604 kref_get(&dev->ref);
605 else
606 return -ENXIO;
607
608 if (!try_module_get(dev->ib_dev->owner)) {
609 ret = -ENODEV;
610 goto err;
611 }
612
613 file = kmalloc(sizeof *file, GFP_KERNEL);
614 if (!file) {
615 ret = -ENOMEM;
616 goto err_module;
617 }
618
619 file->device = dev;
620 file->ucontext = NULL;
621 file->async_file = NULL;
622 kref_init(&file->ref);
623 mutex_init(&file->mutex);
624
625 filp->private_data = file;
626
627 return nonseekable_open(inode, filp);
628
629err_module:
630 module_put(dev->ib_dev->owner);
631
632err:
633 kref_put(&dev->ref, ib_uverbs_release_dev);
634 return ret;
635}
636
637static int ib_uverbs_close(struct inode *inode, struct file *filp)
638{
639 struct ib_uverbs_file *file = filp->private_data;
640
641 ib_uverbs_cleanup_ucontext(file, file->ucontext);
642
643 if (file->async_file)
644 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
645
646 kref_put(&file->ref, ib_uverbs_release_file);
647
648 return 0;
649}
650
651static const struct file_operations uverbs_fops = {
652 .owner = THIS_MODULE,
653 .write = ib_uverbs_write,
654 .open = ib_uverbs_open,
655 .release = ib_uverbs_close,
656 .llseek = no_llseek,
657};
658
659static const struct file_operations uverbs_mmap_fops = {
660 .owner = THIS_MODULE,
661 .write = ib_uverbs_write,
662 .mmap = ib_uverbs_mmap,
663 .open = ib_uverbs_open,
664 .release = ib_uverbs_close,
665 .llseek = no_llseek,
666};
667
668static struct ib_client uverbs_client = {
669 .name = "uverbs",
670 .add = ib_uverbs_add_one,
671 .remove = ib_uverbs_remove_one
672};
673
674static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
675 char *buf)
676{
677 struct ib_uverbs_device *dev = dev_get_drvdata(device);
678
679 if (!dev)
680 return -ENODEV;
681
682 return sprintf(buf, "%s\n", dev->ib_dev->name);
683}
684static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
685
686static ssize_t show_dev_abi_version(struct device *device,
687 struct device_attribute *attr, char *buf)
688{
689 struct ib_uverbs_device *dev = dev_get_drvdata(device);
690
691 if (!dev)
692 return -ENODEV;
693
694 return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
695}
696static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
697
698static CLASS_ATTR_STRING(abi_version, S_IRUGO,
699 __stringify(IB_USER_VERBS_ABI_VERSION));
700
701static dev_t overflow_maj;
702static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
703
704/*
705 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
706 * requesting a new major number and doubling the number of max devices we
707 * support. It's stupid, but simple.
708 */
709static int find_overflow_devnum(void)
710{
711 int ret;
712
713 if (!overflow_maj) {
714 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
715 "infiniband_verbs");
716 if (ret) {
717 printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
718 return ret;
719 }
720 }
721
722 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
723 if (ret >= IB_UVERBS_MAX_DEVICES)
724 return -1;
725
726 return ret;
727}
728
729static void ib_uverbs_add_one(struct ib_device *device)
730{
731 int devnum;
732 dev_t base;
733 struct ib_uverbs_device *uverbs_dev;
734
735 if (!device->alloc_ucontext)
736 return;
737
738 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
739 if (!uverbs_dev)
740 return;
741
742 kref_init(&uverbs_dev->ref);
743 init_completion(&uverbs_dev->comp);
744
745 spin_lock(&map_lock);
746 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
747 if (devnum >= IB_UVERBS_MAX_DEVICES) {
748 spin_unlock(&map_lock);
749 devnum = find_overflow_devnum();
750 if (devnum < 0)
751 goto err;
752
753 spin_lock(&map_lock);
754 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
755 base = devnum + overflow_maj;
756 set_bit(devnum, overflow_map);
757 } else {
758 uverbs_dev->devnum = devnum;
759 base = devnum + IB_UVERBS_BASE_DEV;
760 set_bit(devnum, dev_map);
761 }
762 spin_unlock(&map_lock);
763
764 uverbs_dev->ib_dev = device;
765 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
766
767 cdev_init(&uverbs_dev->cdev, NULL);
768 uverbs_dev->cdev.owner = THIS_MODULE;
769 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
770 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
771 if (cdev_add(&uverbs_dev->cdev, base, 1))
772 goto err_cdev;
773
774 uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
775 uverbs_dev->cdev.dev, uverbs_dev,
776 "uverbs%d", uverbs_dev->devnum);
777 if (IS_ERR(uverbs_dev->dev))
778 goto err_cdev;
779
780 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
781 goto err_class;
782 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
783 goto err_class;
784
785 ib_set_client_data(device, &uverbs_client, uverbs_dev);
786
787 return;
788
789err_class:
790 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
791
792err_cdev:
793 cdev_del(&uverbs_dev->cdev);
794 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
795 clear_bit(devnum, dev_map);
796 else
797 clear_bit(devnum, overflow_map);
798
799err:
800 kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
801 wait_for_completion(&uverbs_dev->comp);
802 kfree(uverbs_dev);
803 return;
804}
805
806static void ib_uverbs_remove_one(struct ib_device *device)
807{
808 struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
809
810 if (!uverbs_dev)
811 return;
812
813 dev_set_drvdata(uverbs_dev->dev, NULL);
814 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
815 cdev_del(&uverbs_dev->cdev);
816
817 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
818 clear_bit(uverbs_dev->devnum, dev_map);
819 else
820 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
821
822 kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
823 wait_for_completion(&uverbs_dev->comp);
824 kfree(uverbs_dev);
825}
826
827static char *uverbs_devnode(struct device *dev, mode_t *mode)
828{
829 if (mode)
830 *mode = 0666;
831 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
832}
833
834static int __init ib_uverbs_init(void)
835{
836 int ret;
837
838 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
839 "infiniband_verbs");
840 if (ret) {
841 printk(KERN_ERR "user_verbs: couldn't register device number\n");
842 goto out;
843 }
844
845 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
846 if (IS_ERR(uverbs_class)) {
847 ret = PTR_ERR(uverbs_class);
848 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
849 goto out_chrdev;
850 }
851
852 uverbs_class->devnode = uverbs_devnode;
853
854 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
855 if (ret) {
856 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
857 goto out_class;
858 }
859
860 ret = ib_register_client(&uverbs_client);
861 if (ret) {
862 printk(KERN_ERR "user_verbs: couldn't register client\n");
863 goto out_class;
864 }
865
866 return 0;
867
868out_class:
869 class_destroy(uverbs_class);
870
871out_chrdev:
872 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
873
874out:
875 return ret;
876}
877
878static void __exit ib_uverbs_cleanup(void)
879{
880 ib_unregister_client(&uverbs_client);
881 class_destroy(uverbs_class);
882 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
883 if (overflow_maj)
884 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
885 idr_destroy(&ib_uverbs_pd_idr);
886 idr_destroy(&ib_uverbs_mr_idr);
887 idr_destroy(&ib_uverbs_mw_idr);
888 idr_destroy(&ib_uverbs_ah_idr);
889 idr_destroy(&ib_uverbs_cq_idr);
890 idr_destroy(&ib_uverbs_qp_idr);
891 idr_destroy(&ib_uverbs_srq_idr);
892}
893
894module_init(ib_uverbs_init);
895module_exit(ib_uverbs_cleanup);
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
55MODULE_AUTHOR("Roland Dreier");
56MODULE_DESCRIPTION("InfiniBand userspace verbs access");
57MODULE_LICENSE("Dual BSD/GPL");
58
59enum {
60 IB_UVERBS_MAJOR = 231,
61 IB_UVERBS_BASE_MINOR = 192,
62 IB_UVERBS_MAX_DEVICES = 32
63};
64
65#define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
66
67static struct class *uverbs_class;
68
69DEFINE_SPINLOCK(ib_uverbs_idr_lock);
70DEFINE_IDR(ib_uverbs_pd_idr);
71DEFINE_IDR(ib_uverbs_mr_idr);
72DEFINE_IDR(ib_uverbs_mw_idr);
73DEFINE_IDR(ib_uverbs_ah_idr);
74DEFINE_IDR(ib_uverbs_cq_idr);
75DEFINE_IDR(ib_uverbs_qp_idr);
76DEFINE_IDR(ib_uverbs_srq_idr);
77DEFINE_IDR(ib_uverbs_xrcd_idr);
78DEFINE_IDR(ib_uverbs_rule_idr);
79DEFINE_IDR(ib_uverbs_wq_idr);
80DEFINE_IDR(ib_uverbs_rwq_ind_tbl_idr);
81
82static DEFINE_SPINLOCK(map_lock);
83static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
84
85static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
86 struct ib_device *ib_dev,
87 const char __user *buf, int in_len,
88 int out_len) = {
89 [IB_USER_VERBS_CMD_GET_CONTEXT] = ib_uverbs_get_context,
90 [IB_USER_VERBS_CMD_QUERY_DEVICE] = ib_uverbs_query_device,
91 [IB_USER_VERBS_CMD_QUERY_PORT] = ib_uverbs_query_port,
92 [IB_USER_VERBS_CMD_ALLOC_PD] = ib_uverbs_alloc_pd,
93 [IB_USER_VERBS_CMD_DEALLOC_PD] = ib_uverbs_dealloc_pd,
94 [IB_USER_VERBS_CMD_REG_MR] = ib_uverbs_reg_mr,
95 [IB_USER_VERBS_CMD_REREG_MR] = ib_uverbs_rereg_mr,
96 [IB_USER_VERBS_CMD_DEREG_MR] = ib_uverbs_dereg_mr,
97 [IB_USER_VERBS_CMD_ALLOC_MW] = ib_uverbs_alloc_mw,
98 [IB_USER_VERBS_CMD_DEALLOC_MW] = ib_uverbs_dealloc_mw,
99 [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
100 [IB_USER_VERBS_CMD_CREATE_CQ] = ib_uverbs_create_cq,
101 [IB_USER_VERBS_CMD_RESIZE_CQ] = ib_uverbs_resize_cq,
102 [IB_USER_VERBS_CMD_POLL_CQ] = ib_uverbs_poll_cq,
103 [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ] = ib_uverbs_req_notify_cq,
104 [IB_USER_VERBS_CMD_DESTROY_CQ] = ib_uverbs_destroy_cq,
105 [IB_USER_VERBS_CMD_CREATE_QP] = ib_uverbs_create_qp,
106 [IB_USER_VERBS_CMD_QUERY_QP] = ib_uverbs_query_qp,
107 [IB_USER_VERBS_CMD_MODIFY_QP] = ib_uverbs_modify_qp,
108 [IB_USER_VERBS_CMD_DESTROY_QP] = ib_uverbs_destroy_qp,
109 [IB_USER_VERBS_CMD_POST_SEND] = ib_uverbs_post_send,
110 [IB_USER_VERBS_CMD_POST_RECV] = ib_uverbs_post_recv,
111 [IB_USER_VERBS_CMD_POST_SRQ_RECV] = ib_uverbs_post_srq_recv,
112 [IB_USER_VERBS_CMD_CREATE_AH] = ib_uverbs_create_ah,
113 [IB_USER_VERBS_CMD_DESTROY_AH] = ib_uverbs_destroy_ah,
114 [IB_USER_VERBS_CMD_ATTACH_MCAST] = ib_uverbs_attach_mcast,
115 [IB_USER_VERBS_CMD_DETACH_MCAST] = ib_uverbs_detach_mcast,
116 [IB_USER_VERBS_CMD_CREATE_SRQ] = ib_uverbs_create_srq,
117 [IB_USER_VERBS_CMD_MODIFY_SRQ] = ib_uverbs_modify_srq,
118 [IB_USER_VERBS_CMD_QUERY_SRQ] = ib_uverbs_query_srq,
119 [IB_USER_VERBS_CMD_DESTROY_SRQ] = ib_uverbs_destroy_srq,
120 [IB_USER_VERBS_CMD_OPEN_XRCD] = ib_uverbs_open_xrcd,
121 [IB_USER_VERBS_CMD_CLOSE_XRCD] = ib_uverbs_close_xrcd,
122 [IB_USER_VERBS_CMD_CREATE_XSRQ] = ib_uverbs_create_xsrq,
123 [IB_USER_VERBS_CMD_OPEN_QP] = ib_uverbs_open_qp,
124};
125
126static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
127 struct ib_device *ib_dev,
128 struct ib_udata *ucore,
129 struct ib_udata *uhw) = {
130 [IB_USER_VERBS_EX_CMD_CREATE_FLOW] = ib_uverbs_ex_create_flow,
131 [IB_USER_VERBS_EX_CMD_DESTROY_FLOW] = ib_uverbs_ex_destroy_flow,
132 [IB_USER_VERBS_EX_CMD_QUERY_DEVICE] = ib_uverbs_ex_query_device,
133 [IB_USER_VERBS_EX_CMD_CREATE_CQ] = ib_uverbs_ex_create_cq,
134 [IB_USER_VERBS_EX_CMD_CREATE_QP] = ib_uverbs_ex_create_qp,
135 [IB_USER_VERBS_EX_CMD_CREATE_WQ] = ib_uverbs_ex_create_wq,
136 [IB_USER_VERBS_EX_CMD_MODIFY_WQ] = ib_uverbs_ex_modify_wq,
137 [IB_USER_VERBS_EX_CMD_DESTROY_WQ] = ib_uverbs_ex_destroy_wq,
138 [IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL] = ib_uverbs_ex_create_rwq_ind_table,
139 [IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL] = ib_uverbs_ex_destroy_rwq_ind_table,
140 [IB_USER_VERBS_EX_CMD_MODIFY_QP] = ib_uverbs_ex_modify_qp,
141};
142
143static void ib_uverbs_add_one(struct ib_device *device);
144static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
145
146int uverbs_dealloc_mw(struct ib_mw *mw)
147{
148 struct ib_pd *pd = mw->pd;
149 int ret;
150
151 ret = mw->device->dealloc_mw(mw);
152 if (!ret)
153 atomic_dec(&pd->usecnt);
154 return ret;
155}
156
157static void ib_uverbs_release_dev(struct kobject *kobj)
158{
159 struct ib_uverbs_device *dev =
160 container_of(kobj, struct ib_uverbs_device, kobj);
161
162 cleanup_srcu_struct(&dev->disassociate_srcu);
163 kfree(dev);
164}
165
166static struct kobj_type ib_uverbs_dev_ktype = {
167 .release = ib_uverbs_release_dev,
168};
169
170static void ib_uverbs_release_event_file(struct kref *ref)
171{
172 struct ib_uverbs_event_file *file =
173 container_of(ref, struct ib_uverbs_event_file, ref);
174
175 kfree(file);
176}
177
178void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
179 struct ib_uverbs_event_file *ev_file,
180 struct ib_ucq_object *uobj)
181{
182 struct ib_uverbs_event *evt, *tmp;
183
184 if (ev_file) {
185 spin_lock_irq(&ev_file->lock);
186 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
187 list_del(&evt->list);
188 kfree(evt);
189 }
190 spin_unlock_irq(&ev_file->lock);
191
192 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
193 }
194
195 spin_lock_irq(&file->async_file->lock);
196 list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
197 list_del(&evt->list);
198 kfree(evt);
199 }
200 spin_unlock_irq(&file->async_file->lock);
201}
202
203void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
204 struct ib_uevent_object *uobj)
205{
206 struct ib_uverbs_event *evt, *tmp;
207
208 spin_lock_irq(&file->async_file->lock);
209 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
210 list_del(&evt->list);
211 kfree(evt);
212 }
213 spin_unlock_irq(&file->async_file->lock);
214}
215
216static void ib_uverbs_detach_umcast(struct ib_qp *qp,
217 struct ib_uqp_object *uobj)
218{
219 struct ib_uverbs_mcast_entry *mcast, *tmp;
220
221 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
222 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
223 list_del(&mcast->list);
224 kfree(mcast);
225 }
226}
227
228static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
229 struct ib_ucontext *context)
230{
231 struct ib_uobject *uobj, *tmp;
232
233 context->closing = 1;
234
235 list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
236 struct ib_ah *ah = uobj->object;
237
238 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
239 ib_destroy_ah(ah);
240 kfree(uobj);
241 }
242
243 /* Remove MWs before QPs, in order to support type 2A MWs. */
244 list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
245 struct ib_mw *mw = uobj->object;
246
247 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
248 uverbs_dealloc_mw(mw);
249 kfree(uobj);
250 }
251
252 list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
253 struct ib_flow *flow_id = uobj->object;
254
255 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
256 ib_destroy_flow(flow_id);
257 kfree(uobj);
258 }
259
260 list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
261 struct ib_qp *qp = uobj->object;
262 struct ib_uqp_object *uqp =
263 container_of(uobj, struct ib_uqp_object, uevent.uobject);
264
265 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
266 if (qp == qp->real_qp)
267 ib_uverbs_detach_umcast(qp, uqp);
268 ib_destroy_qp(qp);
269 ib_uverbs_release_uevent(file, &uqp->uevent);
270 kfree(uqp);
271 }
272
273 list_for_each_entry_safe(uobj, tmp, &context->rwq_ind_tbl_list, list) {
274 struct ib_rwq_ind_table *rwq_ind_tbl = uobj->object;
275 struct ib_wq **ind_tbl = rwq_ind_tbl->ind_tbl;
276
277 idr_remove_uobj(&ib_uverbs_rwq_ind_tbl_idr, uobj);
278 ib_destroy_rwq_ind_table(rwq_ind_tbl);
279 kfree(ind_tbl);
280 kfree(uobj);
281 }
282
283 list_for_each_entry_safe(uobj, tmp, &context->wq_list, list) {
284 struct ib_wq *wq = uobj->object;
285 struct ib_uwq_object *uwq =
286 container_of(uobj, struct ib_uwq_object, uevent.uobject);
287
288 idr_remove_uobj(&ib_uverbs_wq_idr, uobj);
289 ib_destroy_wq(wq);
290 ib_uverbs_release_uevent(file, &uwq->uevent);
291 kfree(uwq);
292 }
293
294 list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
295 struct ib_srq *srq = uobj->object;
296 struct ib_uevent_object *uevent =
297 container_of(uobj, struct ib_uevent_object, uobject);
298
299 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
300 ib_destroy_srq(srq);
301 ib_uverbs_release_uevent(file, uevent);
302 kfree(uevent);
303 }
304
305 list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
306 struct ib_cq *cq = uobj->object;
307 struct ib_uverbs_event_file *ev_file = cq->cq_context;
308 struct ib_ucq_object *ucq =
309 container_of(uobj, struct ib_ucq_object, uobject);
310
311 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
312 ib_destroy_cq(cq);
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(&ib_uverbs_mr_idr, uobj);
321 ib_dereg_mr(mr);
322 kfree(uobj);
323 }
324
325 mutex_lock(&file->device->xrcd_tree_mutex);
326 list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
327 struct ib_xrcd *xrcd = uobj->object;
328 struct ib_uxrcd_object *uxrcd =
329 container_of(uobj, struct ib_uxrcd_object, uobject);
330
331 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
332 ib_uverbs_dealloc_xrcd(file->device, xrcd);
333 kfree(uxrcd);
334 }
335 mutex_unlock(&file->device->xrcd_tree_mutex);
336
337 list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
338 struct ib_pd *pd = uobj->object;
339
340 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
341 ib_dealloc_pd(pd);
342 kfree(uobj);
343 }
344
345 put_pid(context->tgid);
346
347 return context->device->dealloc_ucontext(context);
348}
349
350static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
351{
352 complete(&dev->comp);
353}
354
355static void ib_uverbs_release_file(struct kref *ref)
356{
357 struct ib_uverbs_file *file =
358 container_of(ref, struct ib_uverbs_file, ref);
359 struct ib_device *ib_dev;
360 int srcu_key;
361
362 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
363 ib_dev = srcu_dereference(file->device->ib_dev,
364 &file->device->disassociate_srcu);
365 if (ib_dev && !ib_dev->disassociate_ucontext)
366 module_put(ib_dev->owner);
367 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
368
369 if (atomic_dec_and_test(&file->device->refcount))
370 ib_uverbs_comp_dev(file->device);
371
372 kfree(file);
373}
374
375static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
376 size_t count, loff_t *pos)
377{
378 struct ib_uverbs_event_file *file = filp->private_data;
379 struct ib_uverbs_event *event;
380 int eventsz;
381 int ret = 0;
382
383 spin_lock_irq(&file->lock);
384
385 while (list_empty(&file->event_list)) {
386 spin_unlock_irq(&file->lock);
387
388 if (filp->f_flags & O_NONBLOCK)
389 return -EAGAIN;
390
391 if (wait_event_interruptible(file->poll_wait,
392 (!list_empty(&file->event_list) ||
393 /* The barriers built into wait_event_interruptible()
394 * and wake_up() guarentee this will see the null set
395 * without using RCU
396 */
397 !file->uverbs_file->device->ib_dev)))
398 return -ERESTARTSYS;
399
400 /* If device was disassociated and no event exists set an error */
401 if (list_empty(&file->event_list) &&
402 !file->uverbs_file->device->ib_dev)
403 return -EIO;
404
405 spin_lock_irq(&file->lock);
406 }
407
408 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
409
410 if (file->is_async)
411 eventsz = sizeof (struct ib_uverbs_async_event_desc);
412 else
413 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
414
415 if (eventsz > count) {
416 ret = -EINVAL;
417 event = NULL;
418 } else {
419 list_del(file->event_list.next);
420 if (event->counter) {
421 ++(*event->counter);
422 list_del(&event->obj_list);
423 }
424 }
425
426 spin_unlock_irq(&file->lock);
427
428 if (event) {
429 if (copy_to_user(buf, event, eventsz))
430 ret = -EFAULT;
431 else
432 ret = eventsz;
433 }
434
435 kfree(event);
436
437 return ret;
438}
439
440static unsigned int ib_uverbs_event_poll(struct file *filp,
441 struct poll_table_struct *wait)
442{
443 unsigned int pollflags = 0;
444 struct ib_uverbs_event_file *file = filp->private_data;
445
446 poll_wait(filp, &file->poll_wait, wait);
447
448 spin_lock_irq(&file->lock);
449 if (!list_empty(&file->event_list))
450 pollflags = POLLIN | POLLRDNORM;
451 spin_unlock_irq(&file->lock);
452
453 return pollflags;
454}
455
456static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
457{
458 struct ib_uverbs_event_file *file = filp->private_data;
459
460 return fasync_helper(fd, filp, on, &file->async_queue);
461}
462
463static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
464{
465 struct ib_uverbs_event_file *file = filp->private_data;
466 struct ib_uverbs_event *entry, *tmp;
467 int closed_already = 0;
468
469 mutex_lock(&file->uverbs_file->device->lists_mutex);
470 spin_lock_irq(&file->lock);
471 closed_already = file->is_closed;
472 file->is_closed = 1;
473 list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
474 if (entry->counter)
475 list_del(&entry->obj_list);
476 kfree(entry);
477 }
478 spin_unlock_irq(&file->lock);
479 if (!closed_already) {
480 list_del(&file->list);
481 if (file->is_async)
482 ib_unregister_event_handler(&file->uverbs_file->
483 event_handler);
484 }
485 mutex_unlock(&file->uverbs_file->device->lists_mutex);
486
487 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
488 kref_put(&file->ref, ib_uverbs_release_event_file);
489
490 return 0;
491}
492
493static const struct file_operations uverbs_event_fops = {
494 .owner = THIS_MODULE,
495 .read = ib_uverbs_event_read,
496 .poll = ib_uverbs_event_poll,
497 .release = ib_uverbs_event_close,
498 .fasync = ib_uverbs_event_fasync,
499 .llseek = no_llseek,
500};
501
502void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
503{
504 struct ib_uverbs_event_file *file = cq_context;
505 struct ib_ucq_object *uobj;
506 struct ib_uverbs_event *entry;
507 unsigned long flags;
508
509 if (!file)
510 return;
511
512 spin_lock_irqsave(&file->lock, flags);
513 if (file->is_closed) {
514 spin_unlock_irqrestore(&file->lock, flags);
515 return;
516 }
517
518 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
519 if (!entry) {
520 spin_unlock_irqrestore(&file->lock, flags);
521 return;
522 }
523
524 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
525
526 entry->desc.comp.cq_handle = cq->uobject->user_handle;
527 entry->counter = &uobj->comp_events_reported;
528
529 list_add_tail(&entry->list, &file->event_list);
530 list_add_tail(&entry->obj_list, &uobj->comp_list);
531 spin_unlock_irqrestore(&file->lock, flags);
532
533 wake_up_interruptible(&file->poll_wait);
534 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
535}
536
537static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
538 __u64 element, __u64 event,
539 struct list_head *obj_list,
540 u32 *counter)
541{
542 struct ib_uverbs_event *entry;
543 unsigned long flags;
544
545 spin_lock_irqsave(&file->async_file->lock, flags);
546 if (file->async_file->is_closed) {
547 spin_unlock_irqrestore(&file->async_file->lock, flags);
548 return;
549 }
550
551 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
552 if (!entry) {
553 spin_unlock_irqrestore(&file->async_file->lock, flags);
554 return;
555 }
556
557 entry->desc.async.element = element;
558 entry->desc.async.event_type = event;
559 entry->desc.async.reserved = 0;
560 entry->counter = counter;
561
562 list_add_tail(&entry->list, &file->async_file->event_list);
563 if (obj_list)
564 list_add_tail(&entry->obj_list, obj_list);
565 spin_unlock_irqrestore(&file->async_file->lock, flags);
566
567 wake_up_interruptible(&file->async_file->poll_wait);
568 kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
569}
570
571void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
572{
573 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
574 struct ib_ucq_object, uobject);
575
576 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
577 event->event, &uobj->async_list,
578 &uobj->async_events_reported);
579}
580
581void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
582{
583 struct ib_uevent_object *uobj;
584
585 /* for XRC target qp's, check that qp is live */
586 if (!event->element.qp->uobject || !event->element.qp->uobject->live)
587 return;
588
589 uobj = container_of(event->element.qp->uobject,
590 struct ib_uevent_object, uobject);
591
592 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
593 event->event, &uobj->event_list,
594 &uobj->events_reported);
595}
596
597void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
598{
599 struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
600 struct ib_uevent_object, uobject);
601
602 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
603 event->event, &uobj->event_list,
604 &uobj->events_reported);
605}
606
607void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
608{
609 struct ib_uevent_object *uobj;
610
611 uobj = container_of(event->element.srq->uobject,
612 struct ib_uevent_object, uobject);
613
614 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
615 event->event, &uobj->event_list,
616 &uobj->events_reported);
617}
618
619void ib_uverbs_event_handler(struct ib_event_handler *handler,
620 struct ib_event *event)
621{
622 struct ib_uverbs_file *file =
623 container_of(handler, struct ib_uverbs_file, event_handler);
624
625 ib_uverbs_async_handler(file, event->element.port_num, event->event,
626 NULL, NULL);
627}
628
629void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
630{
631 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
632 file->async_file = NULL;
633}
634
635struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
636 struct ib_device *ib_dev,
637 int is_async)
638{
639 struct ib_uverbs_event_file *ev_file;
640 struct file *filp;
641 int ret;
642
643 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
644 if (!ev_file)
645 return ERR_PTR(-ENOMEM);
646
647 kref_init(&ev_file->ref);
648 spin_lock_init(&ev_file->lock);
649 INIT_LIST_HEAD(&ev_file->event_list);
650 init_waitqueue_head(&ev_file->poll_wait);
651 ev_file->uverbs_file = uverbs_file;
652 kref_get(&ev_file->uverbs_file->ref);
653 ev_file->async_queue = NULL;
654 ev_file->is_closed = 0;
655
656 filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
657 ev_file, O_RDONLY);
658 if (IS_ERR(filp))
659 goto err_put_refs;
660
661 mutex_lock(&uverbs_file->device->lists_mutex);
662 list_add_tail(&ev_file->list,
663 &uverbs_file->device->uverbs_events_file_list);
664 mutex_unlock(&uverbs_file->device->lists_mutex);
665
666 if (is_async) {
667 WARN_ON(uverbs_file->async_file);
668 uverbs_file->async_file = ev_file;
669 kref_get(&uverbs_file->async_file->ref);
670 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
671 ib_dev,
672 ib_uverbs_event_handler);
673 ret = ib_register_event_handler(&uverbs_file->event_handler);
674 if (ret)
675 goto err_put_file;
676
677 /* At that point async file stuff was fully set */
678 ev_file->is_async = 1;
679 }
680
681 return filp;
682
683err_put_file:
684 fput(filp);
685 kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
686 uverbs_file->async_file = NULL;
687 return ERR_PTR(ret);
688
689err_put_refs:
690 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
691 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
692 return filp;
693}
694
695/*
696 * Look up a completion event file by FD. If lookup is successful,
697 * takes a ref to the event file struct that it returns; if
698 * unsuccessful, returns NULL.
699 */
700struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
701{
702 struct ib_uverbs_event_file *ev_file = NULL;
703 struct fd f = fdget(fd);
704
705 if (!f.file)
706 return NULL;
707
708 if (f.file->f_op != &uverbs_event_fops)
709 goto out;
710
711 ev_file = f.file->private_data;
712 if (ev_file->is_async) {
713 ev_file = NULL;
714 goto out;
715 }
716
717 kref_get(&ev_file->ref);
718
719out:
720 fdput(f);
721 return ev_file;
722}
723
724static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
725{
726 u64 mask;
727
728 if (command <= IB_USER_VERBS_CMD_OPEN_QP)
729 mask = ib_dev->uverbs_cmd_mask;
730 else
731 mask = ib_dev->uverbs_ex_cmd_mask;
732
733 if (mask & ((u64)1 << command))
734 return 0;
735
736 return -1;
737}
738
739static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
740 size_t count, loff_t *pos)
741{
742 struct ib_uverbs_file *file = filp->private_data;
743 struct ib_device *ib_dev;
744 struct ib_uverbs_cmd_hdr hdr;
745 __u32 command;
746 __u32 flags;
747 int srcu_key;
748 ssize_t ret;
749
750 if (!ib_safe_file_access(filp)) {
751 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
752 task_tgid_vnr(current), current->comm);
753 return -EACCES;
754 }
755
756 if (count < sizeof hdr)
757 return -EINVAL;
758
759 if (copy_from_user(&hdr, buf, sizeof hdr))
760 return -EFAULT;
761
762 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
763 ib_dev = srcu_dereference(file->device->ib_dev,
764 &file->device->disassociate_srcu);
765 if (!ib_dev) {
766 ret = -EIO;
767 goto out;
768 }
769
770 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
771 IB_USER_VERBS_CMD_COMMAND_MASK)) {
772 ret = -EINVAL;
773 goto out;
774 }
775
776 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
777 if (verify_command_mask(ib_dev, command)) {
778 ret = -EOPNOTSUPP;
779 goto out;
780 }
781
782 if (!file->ucontext &&
783 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
784 ret = -EINVAL;
785 goto out;
786 }
787
788 flags = (hdr.command &
789 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
790
791 if (!flags) {
792 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
793 !uverbs_cmd_table[command]) {
794 ret = -EINVAL;
795 goto out;
796 }
797
798 if (hdr.in_words * 4 != count) {
799 ret = -EINVAL;
800 goto out;
801 }
802
803 ret = uverbs_cmd_table[command](file, ib_dev,
804 buf + sizeof(hdr),
805 hdr.in_words * 4,
806 hdr.out_words * 4);
807
808 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
809 struct ib_uverbs_ex_cmd_hdr ex_hdr;
810 struct ib_udata ucore;
811 struct ib_udata uhw;
812 size_t written_count = count;
813
814 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
815 !uverbs_ex_cmd_table[command]) {
816 ret = -ENOSYS;
817 goto out;
818 }
819
820 if (!file->ucontext) {
821 ret = -EINVAL;
822 goto out;
823 }
824
825 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
826 ret = -EINVAL;
827 goto out;
828 }
829
830 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
831 ret = -EFAULT;
832 goto out;
833 }
834
835 count -= sizeof(hdr) + sizeof(ex_hdr);
836 buf += sizeof(hdr) + sizeof(ex_hdr);
837
838 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
839 ret = -EINVAL;
840 goto out;
841 }
842
843 if (ex_hdr.cmd_hdr_reserved) {
844 ret = -EINVAL;
845 goto out;
846 }
847
848 if (ex_hdr.response) {
849 if (!hdr.out_words && !ex_hdr.provider_out_words) {
850 ret = -EINVAL;
851 goto out;
852 }
853
854 if (!access_ok(VERIFY_WRITE,
855 (void __user *) (unsigned long) ex_hdr.response,
856 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
857 ret = -EFAULT;
858 goto out;
859 }
860 } else {
861 if (hdr.out_words || ex_hdr.provider_out_words) {
862 ret = -EINVAL;
863 goto out;
864 }
865 }
866
867 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
868 hdr.in_words * 8, hdr.out_words * 8);
869
870 INIT_UDATA_BUF_OR_NULL(&uhw,
871 buf + ucore.inlen,
872 (unsigned long) ex_hdr.response + ucore.outlen,
873 ex_hdr.provider_in_words * 8,
874 ex_hdr.provider_out_words * 8);
875
876 ret = uverbs_ex_cmd_table[command](file,
877 ib_dev,
878 &ucore,
879 &uhw);
880 if (!ret)
881 ret = written_count;
882 } else {
883 ret = -ENOSYS;
884 }
885
886out:
887 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
888 return ret;
889}
890
891static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
892{
893 struct ib_uverbs_file *file = filp->private_data;
894 struct ib_device *ib_dev;
895 int ret = 0;
896 int srcu_key;
897
898 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
899 ib_dev = srcu_dereference(file->device->ib_dev,
900 &file->device->disassociate_srcu);
901 if (!ib_dev) {
902 ret = -EIO;
903 goto out;
904 }
905
906 if (!file->ucontext)
907 ret = -ENODEV;
908 else
909 ret = ib_dev->mmap(file->ucontext, vma);
910out:
911 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
912 return ret;
913}
914
915/*
916 * ib_uverbs_open() does not need the BKL:
917 *
918 * - the ib_uverbs_device structures are properly reference counted and
919 * everything else is purely local to the file being created, so
920 * races against other open calls are not a problem;
921 * - there is no ioctl method to race against;
922 * - the open method will either immediately run -ENXIO, or all
923 * required initialization will be done.
924 */
925static int ib_uverbs_open(struct inode *inode, struct file *filp)
926{
927 struct ib_uverbs_device *dev;
928 struct ib_uverbs_file *file;
929 struct ib_device *ib_dev;
930 int ret;
931 int module_dependent;
932 int srcu_key;
933
934 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
935 if (!atomic_inc_not_zero(&dev->refcount))
936 return -ENXIO;
937
938 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
939 mutex_lock(&dev->lists_mutex);
940 ib_dev = srcu_dereference(dev->ib_dev,
941 &dev->disassociate_srcu);
942 if (!ib_dev) {
943 ret = -EIO;
944 goto err;
945 }
946
947 /* In case IB device supports disassociate ucontext, there is no hard
948 * dependency between uverbs device and its low level device.
949 */
950 module_dependent = !(ib_dev->disassociate_ucontext);
951
952 if (module_dependent) {
953 if (!try_module_get(ib_dev->owner)) {
954 ret = -ENODEV;
955 goto err;
956 }
957 }
958
959 file = kzalloc(sizeof(*file), GFP_KERNEL);
960 if (!file) {
961 ret = -ENOMEM;
962 if (module_dependent)
963 goto err_module;
964
965 goto err;
966 }
967
968 file->device = dev;
969 file->ucontext = NULL;
970 file->async_file = NULL;
971 kref_init(&file->ref);
972 mutex_init(&file->mutex);
973 mutex_init(&file->cleanup_mutex);
974
975 filp->private_data = file;
976 kobject_get(&dev->kobj);
977 list_add_tail(&file->list, &dev->uverbs_file_list);
978 mutex_unlock(&dev->lists_mutex);
979 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
980
981 return nonseekable_open(inode, filp);
982
983err_module:
984 module_put(ib_dev->owner);
985
986err:
987 mutex_unlock(&dev->lists_mutex);
988 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
989 if (atomic_dec_and_test(&dev->refcount))
990 ib_uverbs_comp_dev(dev);
991
992 return ret;
993}
994
995static int ib_uverbs_close(struct inode *inode, struct file *filp)
996{
997 struct ib_uverbs_file *file = filp->private_data;
998 struct ib_uverbs_device *dev = file->device;
999
1000 mutex_lock(&file->cleanup_mutex);
1001 if (file->ucontext) {
1002 ib_uverbs_cleanup_ucontext(file, file->ucontext);
1003 file->ucontext = NULL;
1004 }
1005 mutex_unlock(&file->cleanup_mutex);
1006
1007 mutex_lock(&file->device->lists_mutex);
1008 if (!file->is_closed) {
1009 list_del(&file->list);
1010 file->is_closed = 1;
1011 }
1012 mutex_unlock(&file->device->lists_mutex);
1013
1014 if (file->async_file)
1015 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
1016
1017 kref_put(&file->ref, ib_uverbs_release_file);
1018 kobject_put(&dev->kobj);
1019
1020 return 0;
1021}
1022
1023static const struct file_operations uverbs_fops = {
1024 .owner = THIS_MODULE,
1025 .write = ib_uverbs_write,
1026 .open = ib_uverbs_open,
1027 .release = ib_uverbs_close,
1028 .llseek = no_llseek,
1029};
1030
1031static const struct file_operations uverbs_mmap_fops = {
1032 .owner = THIS_MODULE,
1033 .write = ib_uverbs_write,
1034 .mmap = ib_uverbs_mmap,
1035 .open = ib_uverbs_open,
1036 .release = ib_uverbs_close,
1037 .llseek = no_llseek,
1038};
1039
1040static struct ib_client uverbs_client = {
1041 .name = "uverbs",
1042 .add = ib_uverbs_add_one,
1043 .remove = ib_uverbs_remove_one
1044};
1045
1046static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
1047 char *buf)
1048{
1049 int ret = -ENODEV;
1050 int srcu_key;
1051 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1052 struct ib_device *ib_dev;
1053
1054 if (!dev)
1055 return -ENODEV;
1056
1057 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1058 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1059 if (ib_dev)
1060 ret = sprintf(buf, "%s\n", ib_dev->name);
1061 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1062
1063 return ret;
1064}
1065static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1066
1067static ssize_t show_dev_abi_version(struct device *device,
1068 struct device_attribute *attr, char *buf)
1069{
1070 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1071 int ret = -ENODEV;
1072 int srcu_key;
1073 struct ib_device *ib_dev;
1074
1075 if (!dev)
1076 return -ENODEV;
1077 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1078 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1079 if (ib_dev)
1080 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1081 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1082
1083 return ret;
1084}
1085static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1086
1087static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1088 __stringify(IB_USER_VERBS_ABI_VERSION));
1089
1090static dev_t overflow_maj;
1091static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1092
1093/*
1094 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1095 * requesting a new major number and doubling the number of max devices we
1096 * support. It's stupid, but simple.
1097 */
1098static int find_overflow_devnum(void)
1099{
1100 int ret;
1101
1102 if (!overflow_maj) {
1103 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1104 "infiniband_verbs");
1105 if (ret) {
1106 pr_err("user_verbs: couldn't register dynamic device number\n");
1107 return ret;
1108 }
1109 }
1110
1111 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1112 if (ret >= IB_UVERBS_MAX_DEVICES)
1113 return -1;
1114
1115 return ret;
1116}
1117
1118static void ib_uverbs_add_one(struct ib_device *device)
1119{
1120 int devnum;
1121 dev_t base;
1122 struct ib_uverbs_device *uverbs_dev;
1123 int ret;
1124
1125 if (!device->alloc_ucontext)
1126 return;
1127
1128 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1129 if (!uverbs_dev)
1130 return;
1131
1132 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1133 if (ret) {
1134 kfree(uverbs_dev);
1135 return;
1136 }
1137
1138 atomic_set(&uverbs_dev->refcount, 1);
1139 init_completion(&uverbs_dev->comp);
1140 uverbs_dev->xrcd_tree = RB_ROOT;
1141 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1142 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1143 mutex_init(&uverbs_dev->lists_mutex);
1144 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1145 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1146
1147 spin_lock(&map_lock);
1148 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1149 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1150 spin_unlock(&map_lock);
1151 devnum = find_overflow_devnum();
1152 if (devnum < 0)
1153 goto err;
1154
1155 spin_lock(&map_lock);
1156 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1157 base = devnum + overflow_maj;
1158 set_bit(devnum, overflow_map);
1159 } else {
1160 uverbs_dev->devnum = devnum;
1161 base = devnum + IB_UVERBS_BASE_DEV;
1162 set_bit(devnum, dev_map);
1163 }
1164 spin_unlock(&map_lock);
1165
1166 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1167 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1168
1169 cdev_init(&uverbs_dev->cdev, NULL);
1170 uverbs_dev->cdev.owner = THIS_MODULE;
1171 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1172 uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1173 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1174 if (cdev_add(&uverbs_dev->cdev, base, 1))
1175 goto err_cdev;
1176
1177 uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
1178 uverbs_dev->cdev.dev, uverbs_dev,
1179 "uverbs%d", uverbs_dev->devnum);
1180 if (IS_ERR(uverbs_dev->dev))
1181 goto err_cdev;
1182
1183 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1184 goto err_class;
1185 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1186 goto err_class;
1187
1188 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1189
1190 return;
1191
1192err_class:
1193 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1194
1195err_cdev:
1196 cdev_del(&uverbs_dev->cdev);
1197 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1198 clear_bit(devnum, dev_map);
1199 else
1200 clear_bit(devnum, overflow_map);
1201
1202err:
1203 if (atomic_dec_and_test(&uverbs_dev->refcount))
1204 ib_uverbs_comp_dev(uverbs_dev);
1205 wait_for_completion(&uverbs_dev->comp);
1206 kobject_put(&uverbs_dev->kobj);
1207 return;
1208}
1209
1210static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1211 struct ib_device *ib_dev)
1212{
1213 struct ib_uverbs_file *file;
1214 struct ib_uverbs_event_file *event_file;
1215 struct ib_event event;
1216
1217 /* Pending running commands to terminate */
1218 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1219 event.event = IB_EVENT_DEVICE_FATAL;
1220 event.element.port_num = 0;
1221 event.device = ib_dev;
1222
1223 mutex_lock(&uverbs_dev->lists_mutex);
1224 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1225 struct ib_ucontext *ucontext;
1226 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1227 struct ib_uverbs_file, list);
1228 file->is_closed = 1;
1229 list_del(&file->list);
1230 kref_get(&file->ref);
1231 mutex_unlock(&uverbs_dev->lists_mutex);
1232
1233 ib_uverbs_event_handler(&file->event_handler, &event);
1234
1235 mutex_lock(&file->cleanup_mutex);
1236 ucontext = file->ucontext;
1237 file->ucontext = NULL;
1238 mutex_unlock(&file->cleanup_mutex);
1239
1240 /* At this point ib_uverbs_close cannot be running
1241 * ib_uverbs_cleanup_ucontext
1242 */
1243 if (ucontext) {
1244 /* We must release the mutex before going ahead and
1245 * calling disassociate_ucontext. disassociate_ucontext
1246 * might end up indirectly calling uverbs_close,
1247 * for example due to freeing the resources
1248 * (e.g mmput).
1249 */
1250 ib_dev->disassociate_ucontext(ucontext);
1251 ib_uverbs_cleanup_ucontext(file, ucontext);
1252 }
1253
1254 mutex_lock(&uverbs_dev->lists_mutex);
1255 kref_put(&file->ref, ib_uverbs_release_file);
1256 }
1257
1258 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1259 event_file = list_first_entry(&uverbs_dev->
1260 uverbs_events_file_list,
1261 struct ib_uverbs_event_file,
1262 list);
1263 spin_lock_irq(&event_file->lock);
1264 event_file->is_closed = 1;
1265 spin_unlock_irq(&event_file->lock);
1266
1267 list_del(&event_file->list);
1268 if (event_file->is_async) {
1269 ib_unregister_event_handler(&event_file->uverbs_file->
1270 event_handler);
1271 event_file->uverbs_file->event_handler.device = NULL;
1272 }
1273
1274 wake_up_interruptible(&event_file->poll_wait);
1275 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1276 }
1277 mutex_unlock(&uverbs_dev->lists_mutex);
1278}
1279
1280static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1281{
1282 struct ib_uverbs_device *uverbs_dev = client_data;
1283 int wait_clients = 1;
1284
1285 if (!uverbs_dev)
1286 return;
1287
1288 dev_set_drvdata(uverbs_dev->dev, NULL);
1289 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1290 cdev_del(&uverbs_dev->cdev);
1291
1292 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1293 clear_bit(uverbs_dev->devnum, dev_map);
1294 else
1295 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1296
1297 if (device->disassociate_ucontext) {
1298 /* We disassociate HW resources and immediately return.
1299 * Userspace will see a EIO errno for all future access.
1300 * Upon returning, ib_device may be freed internally and is not
1301 * valid any more.
1302 * uverbs_device is still available until all clients close
1303 * their files, then the uverbs device ref count will be zero
1304 * and its resources will be freed.
1305 * Note: At this point no more files can be opened since the
1306 * cdev was deleted, however active clients can still issue
1307 * commands and close their open files.
1308 */
1309 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1310 ib_uverbs_free_hw_resources(uverbs_dev, device);
1311 wait_clients = 0;
1312 }
1313
1314 if (atomic_dec_and_test(&uverbs_dev->refcount))
1315 ib_uverbs_comp_dev(uverbs_dev);
1316 if (wait_clients)
1317 wait_for_completion(&uverbs_dev->comp);
1318 kobject_put(&uverbs_dev->kobj);
1319}
1320
1321static char *uverbs_devnode(struct device *dev, umode_t *mode)
1322{
1323 if (mode)
1324 *mode = 0666;
1325 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1326}
1327
1328static int __init ib_uverbs_init(void)
1329{
1330 int ret;
1331
1332 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1333 "infiniband_verbs");
1334 if (ret) {
1335 pr_err("user_verbs: couldn't register device number\n");
1336 goto out;
1337 }
1338
1339 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1340 if (IS_ERR(uverbs_class)) {
1341 ret = PTR_ERR(uverbs_class);
1342 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1343 goto out_chrdev;
1344 }
1345
1346 uverbs_class->devnode = uverbs_devnode;
1347
1348 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1349 if (ret) {
1350 pr_err("user_verbs: couldn't create abi_version attribute\n");
1351 goto out_class;
1352 }
1353
1354 ret = ib_register_client(&uverbs_client);
1355 if (ret) {
1356 pr_err("user_verbs: couldn't register client\n");
1357 goto out_class;
1358 }
1359
1360 return 0;
1361
1362out_class:
1363 class_destroy(uverbs_class);
1364
1365out_chrdev:
1366 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1367
1368out:
1369 return ret;
1370}
1371
1372static void __exit ib_uverbs_cleanup(void)
1373{
1374 ib_unregister_client(&uverbs_client);
1375 class_destroy(uverbs_class);
1376 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1377 if (overflow_maj)
1378 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1379 idr_destroy(&ib_uverbs_pd_idr);
1380 idr_destroy(&ib_uverbs_mr_idr);
1381 idr_destroy(&ib_uverbs_mw_idr);
1382 idr_destroy(&ib_uverbs_ah_idr);
1383 idr_destroy(&ib_uverbs_cq_idr);
1384 idr_destroy(&ib_uverbs_qp_idr);
1385 idr_destroy(&ib_uverbs_srq_idr);
1386}
1387
1388module_init(ib_uverbs_init);
1389module_exit(ib_uverbs_cleanup);