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#include <linux/sched/mm.h>
49
50#include <linux/uaccess.h>
51
52#include <rdma/ib.h>
53#include <rdma/uverbs_std_types.h>
54#include <rdma/rdma_netlink.h>
55
56#include "uverbs.h"
57#include "core_priv.h"
58#include "rdma_core.h"
59
60MODULE_AUTHOR("Roland Dreier");
61MODULE_DESCRIPTION("InfiniBand userspace verbs access");
62MODULE_LICENSE("Dual BSD/GPL");
63
64enum {
65 IB_UVERBS_MAJOR = 231,
66 IB_UVERBS_BASE_MINOR = 192,
67 IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS,
68 IB_UVERBS_NUM_FIXED_MINOR = 32,
69 IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR,
70};
71
72#define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
73
74static dev_t dynamic_uverbs_dev;
75
76static DEFINE_IDA(uverbs_ida);
77static int ib_uverbs_add_one(struct ib_device *device);
78static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
79static struct ib_client uverbs_client;
80
81static char *uverbs_devnode(const struct device *dev, umode_t *mode)
82{
83 if (mode)
84 *mode = 0666;
85 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
86}
87
88static const struct class uverbs_class = {
89 .name = "infiniband_verbs",
90 .devnode = uverbs_devnode,
91};
92
93/*
94 * Must be called with the ufile->device->disassociate_srcu held, and the lock
95 * must be held until use of the ucontext is finished.
96 */
97struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile)
98{
99 /*
100 * We do not hold the hw_destroy_rwsem lock for this flow, instead
101 * srcu is used. It does not matter if someone races this with
102 * get_context, we get NULL or valid ucontext.
103 */
104 struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext);
105
106 if (!srcu_dereference(ufile->device->ib_dev,
107 &ufile->device->disassociate_srcu))
108 return ERR_PTR(-EIO);
109
110 if (!ucontext)
111 return ERR_PTR(-EINVAL);
112
113 return ucontext;
114}
115EXPORT_SYMBOL(ib_uverbs_get_ucontext_file);
116
117int uverbs_dealloc_mw(struct ib_mw *mw)
118{
119 struct ib_pd *pd = mw->pd;
120 int ret;
121
122 ret = mw->device->ops.dealloc_mw(mw);
123 if (ret)
124 return ret;
125
126 atomic_dec(&pd->usecnt);
127 kfree(mw);
128 return ret;
129}
130
131static void ib_uverbs_release_dev(struct device *device)
132{
133 struct ib_uverbs_device *dev =
134 container_of(device, struct ib_uverbs_device, dev);
135
136 uverbs_destroy_api(dev->uapi);
137 cleanup_srcu_struct(&dev->disassociate_srcu);
138 mutex_destroy(&dev->lists_mutex);
139 mutex_destroy(&dev->xrcd_tree_mutex);
140 kfree(dev);
141}
142
143void ib_uverbs_release_ucq(struct ib_uverbs_completion_event_file *ev_file,
144 struct ib_ucq_object *uobj)
145{
146 struct ib_uverbs_event *evt, *tmp;
147
148 if (ev_file) {
149 spin_lock_irq(&ev_file->ev_queue.lock);
150 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
151 list_del(&evt->list);
152 kfree(evt);
153 }
154 spin_unlock_irq(&ev_file->ev_queue.lock);
155
156 uverbs_uobject_put(&ev_file->uobj);
157 }
158
159 ib_uverbs_release_uevent(&uobj->uevent);
160}
161
162void ib_uverbs_release_uevent(struct ib_uevent_object *uobj)
163{
164 struct ib_uverbs_async_event_file *async_file = uobj->event_file;
165 struct ib_uverbs_event *evt, *tmp;
166
167 if (!async_file)
168 return;
169
170 spin_lock_irq(&async_file->ev_queue.lock);
171 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
172 list_del(&evt->list);
173 kfree(evt);
174 }
175 spin_unlock_irq(&async_file->ev_queue.lock);
176 uverbs_uobject_put(&async_file->uobj);
177}
178
179void ib_uverbs_detach_umcast(struct ib_qp *qp,
180 struct ib_uqp_object *uobj)
181{
182 struct ib_uverbs_mcast_entry *mcast, *tmp;
183
184 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
185 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
186 list_del(&mcast->list);
187 kfree(mcast);
188 }
189}
190
191static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
192{
193 complete(&dev->comp);
194}
195
196void ib_uverbs_release_file(struct kref *ref)
197{
198 struct ib_uverbs_file *file =
199 container_of(ref, struct ib_uverbs_file, ref);
200 struct ib_device *ib_dev;
201 int srcu_key;
202
203 release_ufile_idr_uobject(file);
204
205 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
206 ib_dev = srcu_dereference(file->device->ib_dev,
207 &file->device->disassociate_srcu);
208 if (ib_dev && !ib_dev->ops.disassociate_ucontext)
209 module_put(ib_dev->ops.owner);
210 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
211
212 if (refcount_dec_and_test(&file->device->refcount))
213 ib_uverbs_comp_dev(file->device);
214
215 if (file->default_async_file)
216 uverbs_uobject_put(&file->default_async_file->uobj);
217 put_device(&file->device->dev);
218
219 if (file->disassociate_page)
220 __free_pages(file->disassociate_page, 0);
221 mutex_destroy(&file->disassociation_lock);
222 mutex_destroy(&file->umap_lock);
223 mutex_destroy(&file->ucontext_lock);
224 kfree(file);
225}
226
227static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
228 struct file *filp, char __user *buf,
229 size_t count, loff_t *pos,
230 size_t eventsz)
231{
232 struct ib_uverbs_event *event;
233 int ret = 0;
234
235 spin_lock_irq(&ev_queue->lock);
236
237 while (list_empty(&ev_queue->event_list)) {
238 if (ev_queue->is_closed) {
239 spin_unlock_irq(&ev_queue->lock);
240 return -EIO;
241 }
242
243 spin_unlock_irq(&ev_queue->lock);
244 if (filp->f_flags & O_NONBLOCK)
245 return -EAGAIN;
246
247 if (wait_event_interruptible(ev_queue->poll_wait,
248 (!list_empty(&ev_queue->event_list) ||
249 ev_queue->is_closed)))
250 return -ERESTARTSYS;
251
252 spin_lock_irq(&ev_queue->lock);
253 }
254
255 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
256
257 if (eventsz > count) {
258 ret = -EINVAL;
259 event = NULL;
260 } else {
261 list_del(ev_queue->event_list.next);
262 if (event->counter) {
263 ++(*event->counter);
264 list_del(&event->obj_list);
265 }
266 }
267
268 spin_unlock_irq(&ev_queue->lock);
269
270 if (event) {
271 if (copy_to_user(buf, event, eventsz))
272 ret = -EFAULT;
273 else
274 ret = eventsz;
275 }
276
277 kfree(event);
278
279 return ret;
280}
281
282static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
283 size_t count, loff_t *pos)
284{
285 struct ib_uverbs_async_event_file *file = filp->private_data;
286
287 return ib_uverbs_event_read(&file->ev_queue, filp, buf, count, pos,
288 sizeof(struct ib_uverbs_async_event_desc));
289}
290
291static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
292 size_t count, loff_t *pos)
293{
294 struct ib_uverbs_completion_event_file *comp_ev_file =
295 filp->private_data;
296
297 return ib_uverbs_event_read(&comp_ev_file->ev_queue, filp, buf, count,
298 pos,
299 sizeof(struct ib_uverbs_comp_event_desc));
300}
301
302static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
303 struct file *filp,
304 struct poll_table_struct *wait)
305{
306 __poll_t pollflags = 0;
307
308 poll_wait(filp, &ev_queue->poll_wait, wait);
309
310 spin_lock_irq(&ev_queue->lock);
311 if (!list_empty(&ev_queue->event_list))
312 pollflags = EPOLLIN | EPOLLRDNORM;
313 else if (ev_queue->is_closed)
314 pollflags = EPOLLERR;
315 spin_unlock_irq(&ev_queue->lock);
316
317 return pollflags;
318}
319
320static __poll_t ib_uverbs_async_event_poll(struct file *filp,
321 struct poll_table_struct *wait)
322{
323 struct ib_uverbs_async_event_file *file = filp->private_data;
324
325 return ib_uverbs_event_poll(&file->ev_queue, filp, wait);
326}
327
328static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
329 struct poll_table_struct *wait)
330{
331 struct ib_uverbs_completion_event_file *comp_ev_file =
332 filp->private_data;
333
334 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
335}
336
337static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
338{
339 struct ib_uverbs_async_event_file *file = filp->private_data;
340
341 return fasync_helper(fd, filp, on, &file->ev_queue.async_queue);
342}
343
344static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
345{
346 struct ib_uverbs_completion_event_file *comp_ev_file =
347 filp->private_data;
348
349 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
350}
351
352const struct file_operations uverbs_event_fops = {
353 .owner = THIS_MODULE,
354 .read = ib_uverbs_comp_event_read,
355 .poll = ib_uverbs_comp_event_poll,
356 .release = uverbs_uobject_fd_release,
357 .fasync = ib_uverbs_comp_event_fasync,
358};
359
360const struct file_operations uverbs_async_event_fops = {
361 .owner = THIS_MODULE,
362 .read = ib_uverbs_async_event_read,
363 .poll = ib_uverbs_async_event_poll,
364 .release = uverbs_async_event_release,
365 .fasync = ib_uverbs_async_event_fasync,
366};
367
368void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
369{
370 struct ib_uverbs_event_queue *ev_queue = cq_context;
371 struct ib_ucq_object *uobj;
372 struct ib_uverbs_event *entry;
373 unsigned long flags;
374
375 if (!ev_queue)
376 return;
377
378 spin_lock_irqsave(&ev_queue->lock, flags);
379 if (ev_queue->is_closed) {
380 spin_unlock_irqrestore(&ev_queue->lock, flags);
381 return;
382 }
383
384 entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
385 if (!entry) {
386 spin_unlock_irqrestore(&ev_queue->lock, flags);
387 return;
388 }
389
390 uobj = cq->uobject;
391
392 entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle;
393 entry->counter = &uobj->comp_events_reported;
394
395 list_add_tail(&entry->list, &ev_queue->event_list);
396 list_add_tail(&entry->obj_list, &uobj->comp_list);
397 spin_unlock_irqrestore(&ev_queue->lock, flags);
398
399 wake_up_interruptible(&ev_queue->poll_wait);
400 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
401}
402
403void ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file,
404 __u64 element, __u64 event,
405 struct list_head *obj_list, u32 *counter)
406{
407 struct ib_uverbs_event *entry;
408 unsigned long flags;
409
410 if (!async_file)
411 return;
412
413 spin_lock_irqsave(&async_file->ev_queue.lock, flags);
414 if (async_file->ev_queue.is_closed) {
415 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
416 return;
417 }
418
419 entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
420 if (!entry) {
421 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
422 return;
423 }
424
425 entry->desc.async.element = element;
426 entry->desc.async.event_type = event;
427 entry->desc.async.reserved = 0;
428 entry->counter = counter;
429
430 list_add_tail(&entry->list, &async_file->ev_queue.event_list);
431 if (obj_list)
432 list_add_tail(&entry->obj_list, obj_list);
433 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
434
435 wake_up_interruptible(&async_file->ev_queue.poll_wait);
436 kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN);
437}
438
439static void uverbs_uobj_event(struct ib_uevent_object *eobj,
440 struct ib_event *event)
441{
442 ib_uverbs_async_handler(eobj->event_file,
443 eobj->uobject.user_handle, event->event,
444 &eobj->event_list, &eobj->events_reported);
445}
446
447void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
448{
449 uverbs_uobj_event(&event->element.cq->uobject->uevent, event);
450}
451
452void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
453{
454 /* for XRC target qp's, check that qp is live */
455 if (!event->element.qp->uobject)
456 return;
457
458 uverbs_uobj_event(&event->element.qp->uobject->uevent, event);
459}
460
461void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
462{
463 uverbs_uobj_event(&event->element.wq->uobject->uevent, event);
464}
465
466void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
467{
468 uverbs_uobj_event(&event->element.srq->uobject->uevent, event);
469}
470
471static void ib_uverbs_event_handler(struct ib_event_handler *handler,
472 struct ib_event *event)
473{
474 ib_uverbs_async_handler(
475 container_of(handler, struct ib_uverbs_async_event_file,
476 event_handler),
477 event->element.port_num, event->event, NULL, NULL);
478}
479
480void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
481{
482 spin_lock_init(&ev_queue->lock);
483 INIT_LIST_HEAD(&ev_queue->event_list);
484 init_waitqueue_head(&ev_queue->poll_wait);
485 ev_queue->is_closed = 0;
486 ev_queue->async_queue = NULL;
487}
488
489void ib_uverbs_init_async_event_file(
490 struct ib_uverbs_async_event_file *async_file)
491{
492 struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile;
493 struct ib_device *ib_dev = async_file->uobj.context->device;
494
495 ib_uverbs_init_event_queue(&async_file->ev_queue);
496
497 /* The first async_event_file becomes the default one for the file. */
498 mutex_lock(&uverbs_file->ucontext_lock);
499 if (!uverbs_file->default_async_file) {
500 /* Pairs with the put in ib_uverbs_release_file */
501 uverbs_uobject_get(&async_file->uobj);
502 smp_store_release(&uverbs_file->default_async_file, async_file);
503 }
504 mutex_unlock(&uverbs_file->ucontext_lock);
505
506 INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev,
507 ib_uverbs_event_handler);
508 ib_register_event_handler(&async_file->event_handler);
509}
510
511static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
512 struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count,
513 const struct uverbs_api_write_method *method_elm)
514{
515 if (method_elm->is_ex) {
516 count -= sizeof(*hdr) + sizeof(*ex_hdr);
517
518 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
519 return -EINVAL;
520
521 if (hdr->in_words * 8 < method_elm->req_size)
522 return -ENOSPC;
523
524 if (ex_hdr->cmd_hdr_reserved)
525 return -EINVAL;
526
527 if (ex_hdr->response) {
528 if (!hdr->out_words && !ex_hdr->provider_out_words)
529 return -EINVAL;
530
531 if (hdr->out_words * 8 < method_elm->resp_size)
532 return -ENOSPC;
533
534 if (!access_ok(u64_to_user_ptr(ex_hdr->response),
535 (hdr->out_words + ex_hdr->provider_out_words) * 8))
536 return -EFAULT;
537 } else {
538 if (hdr->out_words || ex_hdr->provider_out_words)
539 return -EINVAL;
540 }
541
542 return 0;
543 }
544
545 /* not extended command */
546 if (hdr->in_words * 4 != count)
547 return -EINVAL;
548
549 if (count < method_elm->req_size + sizeof(*hdr)) {
550 /*
551 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
552 * with a 16 byte write instead of 24. Old kernels didn't
553 * check the size so they allowed this. Now that the size is
554 * checked provide a compatibility work around to not break
555 * those userspaces.
556 */
557 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
558 count == 16) {
559 hdr->in_words = 6;
560 return 0;
561 }
562 return -ENOSPC;
563 }
564 if (hdr->out_words * 4 < method_elm->resp_size)
565 return -ENOSPC;
566
567 return 0;
568}
569
570static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
571 size_t count, loff_t *pos)
572{
573 struct ib_uverbs_file *file = filp->private_data;
574 const struct uverbs_api_write_method *method_elm;
575 struct uverbs_api *uapi = file->device->uapi;
576 struct ib_uverbs_ex_cmd_hdr ex_hdr;
577 struct ib_uverbs_cmd_hdr hdr;
578 struct uverbs_attr_bundle bundle;
579 int srcu_key;
580 ssize_t ret;
581
582 if (!ib_safe_file_access(filp)) {
583 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
584 task_tgid_vnr(current), current->comm);
585 return -EACCES;
586 }
587
588 if (count < sizeof(hdr))
589 return -EINVAL;
590
591 if (copy_from_user(&hdr, buf, sizeof(hdr)))
592 return -EFAULT;
593
594 method_elm = uapi_get_method(uapi, hdr.command);
595 if (IS_ERR(method_elm))
596 return PTR_ERR(method_elm);
597
598 if (method_elm->is_ex) {
599 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
600 return -EINVAL;
601 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
602 return -EFAULT;
603 }
604
605 ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
606 if (ret)
607 return ret;
608
609 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
610
611 buf += sizeof(hdr);
612
613 memset(bundle.attr_present, 0, sizeof(bundle.attr_present));
614 bundle.ufile = file;
615 bundle.context = NULL; /* only valid if bundle has uobject */
616 bundle.uobject = NULL;
617 if (!method_elm->is_ex) {
618 size_t in_len = hdr.in_words * 4 - sizeof(hdr);
619 size_t out_len = hdr.out_words * 4;
620 u64 response = 0;
621
622 if (method_elm->has_udata) {
623 bundle.driver_udata.inlen =
624 in_len - method_elm->req_size;
625 in_len = method_elm->req_size;
626 if (bundle.driver_udata.inlen)
627 bundle.driver_udata.inbuf = buf + in_len;
628 else
629 bundle.driver_udata.inbuf = NULL;
630 } else {
631 memset(&bundle.driver_udata, 0,
632 sizeof(bundle.driver_udata));
633 }
634
635 if (method_elm->has_resp) {
636 /*
637 * The macros check that if has_resp is set
638 * then the command request structure starts
639 * with a '__aligned u64 response' member.
640 */
641 ret = get_user(response, (const u64 __user *)buf);
642 if (ret)
643 goto out_unlock;
644
645 if (method_elm->has_udata) {
646 bundle.driver_udata.outlen =
647 out_len - method_elm->resp_size;
648 out_len = method_elm->resp_size;
649 if (bundle.driver_udata.outlen)
650 bundle.driver_udata.outbuf =
651 u64_to_user_ptr(response +
652 out_len);
653 else
654 bundle.driver_udata.outbuf = NULL;
655 }
656 } else {
657 bundle.driver_udata.outlen = 0;
658 bundle.driver_udata.outbuf = NULL;
659 }
660
661 ib_uverbs_init_udata_buf_or_null(
662 &bundle.ucore, buf, u64_to_user_ptr(response),
663 in_len, out_len);
664 } else {
665 buf += sizeof(ex_hdr);
666
667 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
668 u64_to_user_ptr(ex_hdr.response),
669 hdr.in_words * 8, hdr.out_words * 8);
670
671 ib_uverbs_init_udata_buf_or_null(
672 &bundle.driver_udata, buf + bundle.ucore.inlen,
673 u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
674 ex_hdr.provider_in_words * 8,
675 ex_hdr.provider_out_words * 8);
676
677 }
678
679 ret = method_elm->handler(&bundle);
680 if (bundle.uobject)
681 uverbs_finalize_object(bundle.uobject, UVERBS_ACCESS_NEW, true,
682 !ret, &bundle);
683out_unlock:
684 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
685 return (ret) ? : count;
686}
687
688static const struct vm_operations_struct rdma_umap_ops;
689
690static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
691{
692 struct ib_uverbs_file *file = filp->private_data;
693 struct ib_ucontext *ucontext;
694 int ret = 0;
695 int srcu_key;
696
697 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
698 ucontext = ib_uverbs_get_ucontext_file(file);
699 if (IS_ERR(ucontext)) {
700 ret = PTR_ERR(ucontext);
701 goto out;
702 }
703
704 mutex_lock(&file->disassociation_lock);
705
706 vma->vm_ops = &rdma_umap_ops;
707 ret = ucontext->device->ops.mmap(ucontext, vma);
708
709 mutex_unlock(&file->disassociation_lock);
710out:
711 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
712 return ret;
713}
714
715/*
716 * The VMA has been dup'd, initialize the vm_private_data with a new tracking
717 * struct
718 */
719static void rdma_umap_open(struct vm_area_struct *vma)
720{
721 struct ib_uverbs_file *ufile = vma->vm_file->private_data;
722 struct rdma_umap_priv *opriv = vma->vm_private_data;
723 struct rdma_umap_priv *priv;
724
725 if (!opriv)
726 return;
727
728 /* We are racing with disassociation */
729 if (!down_read_trylock(&ufile->hw_destroy_rwsem))
730 goto out_zap;
731 mutex_lock(&ufile->disassociation_lock);
732
733 /*
734 * Disassociation already completed, the VMA should already be zapped.
735 */
736 if (!ufile->ucontext)
737 goto out_unlock;
738
739 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
740 if (!priv)
741 goto out_unlock;
742 rdma_umap_priv_init(priv, vma, opriv->entry);
743
744 mutex_unlock(&ufile->disassociation_lock);
745 up_read(&ufile->hw_destroy_rwsem);
746 return;
747
748out_unlock:
749 mutex_unlock(&ufile->disassociation_lock);
750 up_read(&ufile->hw_destroy_rwsem);
751out_zap:
752 /*
753 * We can't allow the VMA to be created with the actual IO pages, that
754 * would break our API contract, and it can't be stopped at this
755 * point, so zap it.
756 */
757 vma->vm_private_data = NULL;
758 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
759}
760
761static void rdma_umap_close(struct vm_area_struct *vma)
762{
763 struct ib_uverbs_file *ufile = vma->vm_file->private_data;
764 struct rdma_umap_priv *priv = vma->vm_private_data;
765
766 if (!priv)
767 return;
768
769 /*
770 * The vma holds a reference on the struct file that created it, which
771 * in turn means that the ib_uverbs_file is guaranteed to exist at
772 * this point.
773 */
774 mutex_lock(&ufile->umap_lock);
775 if (priv->entry)
776 rdma_user_mmap_entry_put(priv->entry);
777
778 list_del(&priv->list);
779 mutex_unlock(&ufile->umap_lock);
780 kfree(priv);
781}
782
783/*
784 * Once the zap_vma_ptes has been called touches to the VMA will come here and
785 * we return a dummy writable zero page for all the pfns.
786 */
787static vm_fault_t rdma_umap_fault(struct vm_fault *vmf)
788{
789 struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data;
790 struct rdma_umap_priv *priv = vmf->vma->vm_private_data;
791 vm_fault_t ret = 0;
792
793 if (!priv)
794 return VM_FAULT_SIGBUS;
795
796 /* Read only pages can just use the system zero page. */
797 if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) {
798 vmf->page = ZERO_PAGE(vmf->address);
799 get_page(vmf->page);
800 return 0;
801 }
802
803 mutex_lock(&ufile->umap_lock);
804 if (!ufile->disassociate_page)
805 ufile->disassociate_page =
806 alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0);
807
808 if (ufile->disassociate_page) {
809 /*
810 * This VMA is forced to always be shared so this doesn't have
811 * to worry about COW.
812 */
813 vmf->page = ufile->disassociate_page;
814 get_page(vmf->page);
815 } else {
816 ret = VM_FAULT_SIGBUS;
817 }
818 mutex_unlock(&ufile->umap_lock);
819
820 return ret;
821}
822
823static const struct vm_operations_struct rdma_umap_ops = {
824 .open = rdma_umap_open,
825 .close = rdma_umap_close,
826 .fault = rdma_umap_fault,
827};
828
829void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
830{
831 struct rdma_umap_priv *priv, *next_priv;
832
833 mutex_lock(&ufile->disassociation_lock);
834
835 while (1) {
836 struct mm_struct *mm = NULL;
837
838 /* Get an arbitrary mm pointer that hasn't been cleaned yet */
839 mutex_lock(&ufile->umap_lock);
840 while (!list_empty(&ufile->umaps)) {
841 int ret;
842
843 priv = list_first_entry(&ufile->umaps,
844 struct rdma_umap_priv, list);
845 mm = priv->vma->vm_mm;
846 ret = mmget_not_zero(mm);
847 if (!ret) {
848 list_del_init(&priv->list);
849 if (priv->entry) {
850 rdma_user_mmap_entry_put(priv->entry);
851 priv->entry = NULL;
852 }
853 mm = NULL;
854 continue;
855 }
856 break;
857 }
858 mutex_unlock(&ufile->umap_lock);
859 if (!mm) {
860 mutex_unlock(&ufile->disassociation_lock);
861 return;
862 }
863
864 /*
865 * The umap_lock is nested under mmap_lock since it used within
866 * the vma_ops callbacks, so we have to clean the list one mm
867 * at a time to get the lock ordering right. Typically there
868 * will only be one mm, so no big deal.
869 */
870 mmap_read_lock(mm);
871 mutex_lock(&ufile->umap_lock);
872 list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
873 list) {
874 struct vm_area_struct *vma = priv->vma;
875
876 if (vma->vm_mm != mm)
877 continue;
878 list_del_init(&priv->list);
879
880 zap_vma_ptes(vma, vma->vm_start,
881 vma->vm_end - vma->vm_start);
882
883 if (priv->entry) {
884 rdma_user_mmap_entry_put(priv->entry);
885 priv->entry = NULL;
886 }
887 }
888 mutex_unlock(&ufile->umap_lock);
889 mmap_read_unlock(mm);
890 mmput(mm);
891 }
892
893 mutex_unlock(&ufile->disassociation_lock);
894}
895
896/**
897 * rdma_user_mmap_disassociate() - Revoke mmaps for a device
898 * @device: device to revoke
899 *
900 * This function should be called by drivers that need to disable mmaps for the
901 * device, for instance because it is going to be reset.
902 */
903void rdma_user_mmap_disassociate(struct ib_device *device)
904{
905 struct ib_uverbs_device *uverbs_dev =
906 ib_get_client_data(device, &uverbs_client);
907 struct ib_uverbs_file *ufile;
908
909 mutex_lock(&uverbs_dev->lists_mutex);
910 list_for_each_entry(ufile, &uverbs_dev->uverbs_file_list, list) {
911 if (ufile->ucontext)
912 uverbs_user_mmap_disassociate(ufile);
913 }
914 mutex_unlock(&uverbs_dev->lists_mutex);
915}
916EXPORT_SYMBOL(rdma_user_mmap_disassociate);
917
918/*
919 * ib_uverbs_open() does not need the BKL:
920 *
921 * - the ib_uverbs_device structures are properly reference counted and
922 * everything else is purely local to the file being created, so
923 * races against other open calls are not a problem;
924 * - there is no ioctl method to race against;
925 * - the open method will either immediately run -ENXIO, or all
926 * required initialization will be done.
927 */
928static int ib_uverbs_open(struct inode *inode, struct file *filp)
929{
930 struct ib_uverbs_device *dev;
931 struct ib_uverbs_file *file;
932 struct ib_device *ib_dev;
933 int ret;
934 int module_dependent;
935 int srcu_key;
936
937 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
938 if (!refcount_inc_not_zero(&dev->refcount))
939 return -ENXIO;
940
941 get_device(&dev->dev);
942 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
943 mutex_lock(&dev->lists_mutex);
944 ib_dev = srcu_dereference(dev->ib_dev,
945 &dev->disassociate_srcu);
946 if (!ib_dev) {
947 ret = -EIO;
948 goto err;
949 }
950
951 if (!rdma_dev_access_netns(ib_dev, current->nsproxy->net_ns)) {
952 ret = -EPERM;
953 goto err;
954 }
955
956 /* In case IB device supports disassociate ucontext, there is no hard
957 * dependency between uverbs device and its low level device.
958 */
959 module_dependent = !(ib_dev->ops.disassociate_ucontext);
960
961 if (module_dependent) {
962 if (!try_module_get(ib_dev->ops.owner)) {
963 ret = -ENODEV;
964 goto err;
965 }
966 }
967
968 file = kzalloc(sizeof(*file), GFP_KERNEL);
969 if (!file) {
970 ret = -ENOMEM;
971 if (module_dependent)
972 goto err_module;
973
974 goto err;
975 }
976
977 file->device = dev;
978 kref_init(&file->ref);
979 mutex_init(&file->ucontext_lock);
980
981 spin_lock_init(&file->uobjects_lock);
982 INIT_LIST_HEAD(&file->uobjects);
983 init_rwsem(&file->hw_destroy_rwsem);
984 mutex_init(&file->umap_lock);
985 INIT_LIST_HEAD(&file->umaps);
986
987 mutex_init(&file->disassociation_lock);
988
989 filp->private_data = file;
990 list_add_tail(&file->list, &dev->uverbs_file_list);
991 mutex_unlock(&dev->lists_mutex);
992 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
993
994 setup_ufile_idr_uobject(file);
995
996 return stream_open(inode, filp);
997
998err_module:
999 module_put(ib_dev->ops.owner);
1000
1001err:
1002 mutex_unlock(&dev->lists_mutex);
1003 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1004 if (refcount_dec_and_test(&dev->refcount))
1005 ib_uverbs_comp_dev(dev);
1006
1007 put_device(&dev->dev);
1008 return ret;
1009}
1010
1011static int ib_uverbs_close(struct inode *inode, struct file *filp)
1012{
1013 struct ib_uverbs_file *file = filp->private_data;
1014
1015 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
1016
1017 mutex_lock(&file->device->lists_mutex);
1018 list_del_init(&file->list);
1019 mutex_unlock(&file->device->lists_mutex);
1020
1021 kref_put(&file->ref, ib_uverbs_release_file);
1022
1023 return 0;
1024}
1025
1026static const struct file_operations uverbs_fops = {
1027 .owner = THIS_MODULE,
1028 .write = ib_uverbs_write,
1029 .open = ib_uverbs_open,
1030 .release = ib_uverbs_close,
1031 .unlocked_ioctl = ib_uverbs_ioctl,
1032 .compat_ioctl = compat_ptr_ioctl,
1033};
1034
1035static const struct file_operations uverbs_mmap_fops = {
1036 .owner = THIS_MODULE,
1037 .write = ib_uverbs_write,
1038 .mmap = ib_uverbs_mmap,
1039 .open = ib_uverbs_open,
1040 .release = ib_uverbs_close,
1041 .unlocked_ioctl = ib_uverbs_ioctl,
1042 .compat_ioctl = compat_ptr_ioctl,
1043};
1044
1045static int ib_uverbs_get_nl_info(struct ib_device *ibdev, void *client_data,
1046 struct ib_client_nl_info *res)
1047{
1048 struct ib_uverbs_device *uverbs_dev = client_data;
1049 int ret;
1050
1051 if (res->port != -1)
1052 return -EINVAL;
1053
1054 res->abi = ibdev->ops.uverbs_abi_ver;
1055 res->cdev = &uverbs_dev->dev;
1056
1057 /*
1058 * To support DRIVER_ID binding in userspace some of the driver need
1059 * upgrading to expose their PCI dependent revision information
1060 * through get_context instead of relying on modalias matching. When
1061 * the drivers are fixed they can drop this flag.
1062 */
1063 if (!ibdev->ops.uverbs_no_driver_id_binding) {
1064 ret = nla_put_u32(res->nl_msg, RDMA_NLDEV_ATTR_UVERBS_DRIVER_ID,
1065 ibdev->ops.driver_id);
1066 if (ret)
1067 return ret;
1068 }
1069 return 0;
1070}
1071
1072static struct ib_client uverbs_client = {
1073 .name = "uverbs",
1074 .no_kverbs_req = true,
1075 .add = ib_uverbs_add_one,
1076 .remove = ib_uverbs_remove_one,
1077 .get_nl_info = ib_uverbs_get_nl_info,
1078};
1079MODULE_ALIAS_RDMA_CLIENT("uverbs");
1080
1081static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1082 char *buf)
1083{
1084 struct ib_uverbs_device *dev =
1085 container_of(device, struct ib_uverbs_device, dev);
1086 int ret = -ENODEV;
1087 int srcu_key;
1088 struct ib_device *ib_dev;
1089
1090 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1091 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1092 if (ib_dev)
1093 ret = sysfs_emit(buf, "%s\n", dev_name(&ib_dev->dev));
1094 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1095
1096 return ret;
1097}
1098static DEVICE_ATTR_RO(ibdev);
1099
1100static ssize_t abi_version_show(struct device *device,
1101 struct device_attribute *attr, char *buf)
1102{
1103 struct ib_uverbs_device *dev =
1104 container_of(device, struct ib_uverbs_device, dev);
1105 int ret = -ENODEV;
1106 int srcu_key;
1107 struct ib_device *ib_dev;
1108
1109 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1110 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1111 if (ib_dev)
1112 ret = sysfs_emit(buf, "%u\n", ib_dev->ops.uverbs_abi_ver);
1113 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1114
1115 return ret;
1116}
1117static DEVICE_ATTR_RO(abi_version);
1118
1119static struct attribute *ib_dev_attrs[] = {
1120 &dev_attr_abi_version.attr,
1121 &dev_attr_ibdev.attr,
1122 NULL,
1123};
1124
1125static const struct attribute_group dev_attr_group = {
1126 .attrs = ib_dev_attrs,
1127};
1128
1129static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1130 __stringify(IB_USER_VERBS_ABI_VERSION));
1131
1132static int ib_uverbs_create_uapi(struct ib_device *device,
1133 struct ib_uverbs_device *uverbs_dev)
1134{
1135 struct uverbs_api *uapi;
1136
1137 uapi = uverbs_alloc_api(device);
1138 if (IS_ERR(uapi))
1139 return PTR_ERR(uapi);
1140
1141 uverbs_dev->uapi = uapi;
1142 return 0;
1143}
1144
1145static int ib_uverbs_add_one(struct ib_device *device)
1146{
1147 int devnum;
1148 dev_t base;
1149 struct ib_uverbs_device *uverbs_dev;
1150 int ret;
1151
1152 if (!device->ops.alloc_ucontext ||
1153 device->type == RDMA_DEVICE_TYPE_SMI)
1154 return -EOPNOTSUPP;
1155
1156 uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1157 if (!uverbs_dev)
1158 return -ENOMEM;
1159
1160 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1161 if (ret) {
1162 kfree(uverbs_dev);
1163 return -ENOMEM;
1164 }
1165
1166 device_initialize(&uverbs_dev->dev);
1167 uverbs_dev->dev.class = &uverbs_class;
1168 uverbs_dev->dev.parent = device->dev.parent;
1169 uverbs_dev->dev.release = ib_uverbs_release_dev;
1170 uverbs_dev->groups[0] = &dev_attr_group;
1171 uverbs_dev->dev.groups = uverbs_dev->groups;
1172 refcount_set(&uverbs_dev->refcount, 1);
1173 init_completion(&uverbs_dev->comp);
1174 uverbs_dev->xrcd_tree = RB_ROOT;
1175 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1176 mutex_init(&uverbs_dev->lists_mutex);
1177 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1178 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1179 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1180
1181 devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1182 GFP_KERNEL);
1183 if (devnum < 0) {
1184 ret = -ENOMEM;
1185 goto err;
1186 }
1187 uverbs_dev->devnum = devnum;
1188 if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1189 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1190 else
1191 base = IB_UVERBS_BASE_DEV + devnum;
1192
1193 ret = ib_uverbs_create_uapi(device, uverbs_dev);
1194 if (ret)
1195 goto err_uapi;
1196
1197 uverbs_dev->dev.devt = base;
1198 dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1199
1200 cdev_init(&uverbs_dev->cdev,
1201 device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1202 uverbs_dev->cdev.owner = THIS_MODULE;
1203
1204 ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1205 if (ret)
1206 goto err_uapi;
1207
1208 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1209 return 0;
1210
1211err_uapi:
1212 ida_free(&uverbs_ida, devnum);
1213err:
1214 if (refcount_dec_and_test(&uverbs_dev->refcount))
1215 ib_uverbs_comp_dev(uverbs_dev);
1216 wait_for_completion(&uverbs_dev->comp);
1217 put_device(&uverbs_dev->dev);
1218 return ret;
1219}
1220
1221static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1222 struct ib_device *ib_dev)
1223{
1224 struct ib_uverbs_file *file;
1225
1226 /* Pending running commands to terminate */
1227 uverbs_disassociate_api_pre(uverbs_dev);
1228
1229 mutex_lock(&uverbs_dev->lists_mutex);
1230 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1231 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1232 struct ib_uverbs_file, list);
1233 list_del_init(&file->list);
1234 kref_get(&file->ref);
1235
1236 /* We must release the mutex before going ahead and calling
1237 * uverbs_cleanup_ufile, as it might end up indirectly calling
1238 * uverbs_close, for example due to freeing the resources (e.g
1239 * mmput).
1240 */
1241 mutex_unlock(&uverbs_dev->lists_mutex);
1242
1243 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1244 kref_put(&file->ref, ib_uverbs_release_file);
1245
1246 mutex_lock(&uverbs_dev->lists_mutex);
1247 }
1248 mutex_unlock(&uverbs_dev->lists_mutex);
1249
1250 uverbs_disassociate_api(uverbs_dev->uapi);
1251}
1252
1253static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1254{
1255 struct ib_uverbs_device *uverbs_dev = client_data;
1256 int wait_clients = 1;
1257
1258 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1259 ida_free(&uverbs_ida, uverbs_dev->devnum);
1260
1261 if (device->ops.disassociate_ucontext) {
1262 /* We disassociate HW resources and immediately return.
1263 * Userspace will see a EIO errno for all future access.
1264 * Upon returning, ib_device may be freed internally and is not
1265 * valid any more.
1266 * uverbs_device is still available until all clients close
1267 * their files, then the uverbs device ref count will be zero
1268 * and its resources will be freed.
1269 * Note: At this point no more files can be opened since the
1270 * cdev was deleted, however active clients can still issue
1271 * commands and close their open files.
1272 */
1273 ib_uverbs_free_hw_resources(uverbs_dev, device);
1274 wait_clients = 0;
1275 }
1276
1277 if (refcount_dec_and_test(&uverbs_dev->refcount))
1278 ib_uverbs_comp_dev(uverbs_dev);
1279 if (wait_clients)
1280 wait_for_completion(&uverbs_dev->comp);
1281
1282 put_device(&uverbs_dev->dev);
1283}
1284
1285static int __init ib_uverbs_init(void)
1286{
1287 int ret;
1288
1289 ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1290 IB_UVERBS_NUM_FIXED_MINOR,
1291 "infiniband_verbs");
1292 if (ret) {
1293 pr_err("user_verbs: couldn't register device number\n");
1294 goto out;
1295 }
1296
1297 ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1298 IB_UVERBS_NUM_DYNAMIC_MINOR,
1299 "infiniband_verbs");
1300 if (ret) {
1301 pr_err("couldn't register dynamic device number\n");
1302 goto out_alloc;
1303 }
1304
1305 ret = class_register(&uverbs_class);
1306 if (ret) {
1307 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1308 goto out_chrdev;
1309 }
1310
1311 ret = class_create_file(&uverbs_class, &class_attr_abi_version.attr);
1312 if (ret) {
1313 pr_err("user_verbs: couldn't create abi_version attribute\n");
1314 goto out_class;
1315 }
1316
1317 ret = ib_register_client(&uverbs_client);
1318 if (ret) {
1319 pr_err("user_verbs: couldn't register client\n");
1320 goto out_class;
1321 }
1322
1323 return 0;
1324
1325out_class:
1326 class_unregister(&uverbs_class);
1327
1328out_chrdev:
1329 unregister_chrdev_region(dynamic_uverbs_dev,
1330 IB_UVERBS_NUM_DYNAMIC_MINOR);
1331
1332out_alloc:
1333 unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1334 IB_UVERBS_NUM_FIXED_MINOR);
1335
1336out:
1337 return ret;
1338}
1339
1340static void __exit ib_uverbs_cleanup(void)
1341{
1342 ib_unregister_client(&uverbs_client);
1343 class_unregister(&uverbs_class);
1344 unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1345 IB_UVERBS_NUM_FIXED_MINOR);
1346 unregister_chrdev_region(dynamic_uverbs_dev,
1347 IB_UVERBS_NUM_DYNAMIC_MINOR);
1348 mmu_notifier_synchronize();
1349}
1350
1351module_init(ib_uverbs_init);
1352module_exit(ib_uverbs_cleanup);