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 <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);
75DEFINE_IDR(ib_uverbs_xrcd_idr);
76
77static DEFINE_SPINLOCK(map_lock);
78static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
79
80static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
81 const char __user *buf, int in_len,
82 int out_len) = {
83 [IB_USER_VERBS_CMD_GET_CONTEXT] = ib_uverbs_get_context,
84 [IB_USER_VERBS_CMD_QUERY_DEVICE] = ib_uverbs_query_device,
85 [IB_USER_VERBS_CMD_QUERY_PORT] = ib_uverbs_query_port,
86 [IB_USER_VERBS_CMD_ALLOC_PD] = ib_uverbs_alloc_pd,
87 [IB_USER_VERBS_CMD_DEALLOC_PD] = ib_uverbs_dealloc_pd,
88 [IB_USER_VERBS_CMD_REG_MR] = ib_uverbs_reg_mr,
89 [IB_USER_VERBS_CMD_DEREG_MR] = ib_uverbs_dereg_mr,
90 [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
91 [IB_USER_VERBS_CMD_CREATE_CQ] = ib_uverbs_create_cq,
92 [IB_USER_VERBS_CMD_RESIZE_CQ] = ib_uverbs_resize_cq,
93 [IB_USER_VERBS_CMD_POLL_CQ] = ib_uverbs_poll_cq,
94 [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ] = ib_uverbs_req_notify_cq,
95 [IB_USER_VERBS_CMD_DESTROY_CQ] = ib_uverbs_destroy_cq,
96 [IB_USER_VERBS_CMD_CREATE_QP] = ib_uverbs_create_qp,
97 [IB_USER_VERBS_CMD_QUERY_QP] = ib_uverbs_query_qp,
98 [IB_USER_VERBS_CMD_MODIFY_QP] = ib_uverbs_modify_qp,
99 [IB_USER_VERBS_CMD_DESTROY_QP] = ib_uverbs_destroy_qp,
100 [IB_USER_VERBS_CMD_POST_SEND] = ib_uverbs_post_send,
101 [IB_USER_VERBS_CMD_POST_RECV] = ib_uverbs_post_recv,
102 [IB_USER_VERBS_CMD_POST_SRQ_RECV] = ib_uverbs_post_srq_recv,
103 [IB_USER_VERBS_CMD_CREATE_AH] = ib_uverbs_create_ah,
104 [IB_USER_VERBS_CMD_DESTROY_AH] = ib_uverbs_destroy_ah,
105 [IB_USER_VERBS_CMD_ATTACH_MCAST] = ib_uverbs_attach_mcast,
106 [IB_USER_VERBS_CMD_DETACH_MCAST] = ib_uverbs_detach_mcast,
107 [IB_USER_VERBS_CMD_CREATE_SRQ] = ib_uverbs_create_srq,
108 [IB_USER_VERBS_CMD_MODIFY_SRQ] = ib_uverbs_modify_srq,
109 [IB_USER_VERBS_CMD_QUERY_SRQ] = ib_uverbs_query_srq,
110 [IB_USER_VERBS_CMD_DESTROY_SRQ] = ib_uverbs_destroy_srq,
111 [IB_USER_VERBS_CMD_OPEN_XRCD] = ib_uverbs_open_xrcd,
112 [IB_USER_VERBS_CMD_CLOSE_XRCD] = ib_uverbs_close_xrcd,
113 [IB_USER_VERBS_CMD_CREATE_XSRQ] = ib_uverbs_create_xsrq,
114 [IB_USER_VERBS_CMD_OPEN_QP] = ib_uverbs_open_qp
115};
116
117static void ib_uverbs_add_one(struct ib_device *device);
118static void ib_uverbs_remove_one(struct ib_device *device);
119
120static void ib_uverbs_release_dev(struct kref *ref)
121{
122 struct ib_uverbs_device *dev =
123 container_of(ref, struct ib_uverbs_device, ref);
124
125 complete(&dev->comp);
126}
127
128static void ib_uverbs_release_event_file(struct kref *ref)
129{
130 struct ib_uverbs_event_file *file =
131 container_of(ref, struct ib_uverbs_event_file, ref);
132
133 kfree(file);
134}
135
136void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
137 struct ib_uverbs_event_file *ev_file,
138 struct ib_ucq_object *uobj)
139{
140 struct ib_uverbs_event *evt, *tmp;
141
142 if (ev_file) {
143 spin_lock_irq(&ev_file->lock);
144 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
145 list_del(&evt->list);
146 kfree(evt);
147 }
148 spin_unlock_irq(&ev_file->lock);
149
150 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
151 }
152
153 spin_lock_irq(&file->async_file->lock);
154 list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
155 list_del(&evt->list);
156 kfree(evt);
157 }
158 spin_unlock_irq(&file->async_file->lock);
159}
160
161void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
162 struct ib_uevent_object *uobj)
163{
164 struct ib_uverbs_event *evt, *tmp;
165
166 spin_lock_irq(&file->async_file->lock);
167 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
168 list_del(&evt->list);
169 kfree(evt);
170 }
171 spin_unlock_irq(&file->async_file->lock);
172}
173
174static void ib_uverbs_detach_umcast(struct ib_qp *qp,
175 struct ib_uqp_object *uobj)
176{
177 struct ib_uverbs_mcast_entry *mcast, *tmp;
178
179 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
180 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
181 list_del(&mcast->list);
182 kfree(mcast);
183 }
184}
185
186static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
187 struct ib_ucontext *context)
188{
189 struct ib_uobject *uobj, *tmp;
190
191 if (!context)
192 return 0;
193
194 context->closing = 1;
195
196 list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
197 struct ib_ah *ah = uobj->object;
198
199 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
200 ib_destroy_ah(ah);
201 kfree(uobj);
202 }
203
204 list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
205 struct ib_qp *qp = uobj->object;
206 struct ib_uqp_object *uqp =
207 container_of(uobj, struct ib_uqp_object, uevent.uobject);
208
209 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
210 if (qp != qp->real_qp) {
211 ib_close_qp(qp);
212 } else {
213 ib_uverbs_detach_umcast(qp, uqp);
214 ib_destroy_qp(qp);
215 }
216 ib_uverbs_release_uevent(file, &uqp->uevent);
217 kfree(uqp);
218 }
219
220 list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
221 struct ib_cq *cq = uobj->object;
222 struct ib_uverbs_event_file *ev_file = cq->cq_context;
223 struct ib_ucq_object *ucq =
224 container_of(uobj, struct ib_ucq_object, uobject);
225
226 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
227 ib_destroy_cq(cq);
228 ib_uverbs_release_ucq(file, ev_file, ucq);
229 kfree(ucq);
230 }
231
232 list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
233 struct ib_srq *srq = uobj->object;
234 struct ib_uevent_object *uevent =
235 container_of(uobj, struct ib_uevent_object, uobject);
236
237 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
238 ib_destroy_srq(srq);
239 ib_uverbs_release_uevent(file, uevent);
240 kfree(uevent);
241 }
242
243 /* XXX Free MWs */
244
245 list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
246 struct ib_mr *mr = uobj->object;
247
248 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
249 ib_dereg_mr(mr);
250 kfree(uobj);
251 }
252
253 mutex_lock(&file->device->xrcd_tree_mutex);
254 list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
255 struct ib_xrcd *xrcd = uobj->object;
256 struct ib_uxrcd_object *uxrcd =
257 container_of(uobj, struct ib_uxrcd_object, uobject);
258
259 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
260 ib_uverbs_dealloc_xrcd(file->device, xrcd);
261 kfree(uxrcd);
262 }
263 mutex_unlock(&file->device->xrcd_tree_mutex);
264
265 list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
266 struct ib_pd *pd = uobj->object;
267
268 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
269 ib_dealloc_pd(pd);
270 kfree(uobj);
271 }
272
273 return context->device->dealloc_ucontext(context);
274}
275
276static void ib_uverbs_release_file(struct kref *ref)
277{
278 struct ib_uverbs_file *file =
279 container_of(ref, struct ib_uverbs_file, ref);
280
281 module_put(file->device->ib_dev->owner);
282 kref_put(&file->device->ref, ib_uverbs_release_dev);
283
284 kfree(file);
285}
286
287static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
288 size_t count, loff_t *pos)
289{
290 struct ib_uverbs_event_file *file = filp->private_data;
291 struct ib_uverbs_event *event;
292 int eventsz;
293 int ret = 0;
294
295 spin_lock_irq(&file->lock);
296
297 while (list_empty(&file->event_list)) {
298 spin_unlock_irq(&file->lock);
299
300 if (filp->f_flags & O_NONBLOCK)
301 return -EAGAIN;
302
303 if (wait_event_interruptible(file->poll_wait,
304 !list_empty(&file->event_list)))
305 return -ERESTARTSYS;
306
307 spin_lock_irq(&file->lock);
308 }
309
310 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
311
312 if (file->is_async)
313 eventsz = sizeof (struct ib_uverbs_async_event_desc);
314 else
315 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
316
317 if (eventsz > count) {
318 ret = -EINVAL;
319 event = NULL;
320 } else {
321 list_del(file->event_list.next);
322 if (event->counter) {
323 ++(*event->counter);
324 list_del(&event->obj_list);
325 }
326 }
327
328 spin_unlock_irq(&file->lock);
329
330 if (event) {
331 if (copy_to_user(buf, event, eventsz))
332 ret = -EFAULT;
333 else
334 ret = eventsz;
335 }
336
337 kfree(event);
338
339 return ret;
340}
341
342static unsigned int ib_uverbs_event_poll(struct file *filp,
343 struct poll_table_struct *wait)
344{
345 unsigned int pollflags = 0;
346 struct ib_uverbs_event_file *file = filp->private_data;
347
348 poll_wait(filp, &file->poll_wait, wait);
349
350 spin_lock_irq(&file->lock);
351 if (!list_empty(&file->event_list))
352 pollflags = POLLIN | POLLRDNORM;
353 spin_unlock_irq(&file->lock);
354
355 return pollflags;
356}
357
358static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
359{
360 struct ib_uverbs_event_file *file = filp->private_data;
361
362 return fasync_helper(fd, filp, on, &file->async_queue);
363}
364
365static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
366{
367 struct ib_uverbs_event_file *file = filp->private_data;
368 struct ib_uverbs_event *entry, *tmp;
369
370 spin_lock_irq(&file->lock);
371 file->is_closed = 1;
372 list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
373 if (entry->counter)
374 list_del(&entry->obj_list);
375 kfree(entry);
376 }
377 spin_unlock_irq(&file->lock);
378
379 if (file->is_async) {
380 ib_unregister_event_handler(&file->uverbs_file->event_handler);
381 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
382 }
383 kref_put(&file->ref, ib_uverbs_release_event_file);
384
385 return 0;
386}
387
388static const struct file_operations uverbs_event_fops = {
389 .owner = THIS_MODULE,
390 .read = ib_uverbs_event_read,
391 .poll = ib_uverbs_event_poll,
392 .release = ib_uverbs_event_close,
393 .fasync = ib_uverbs_event_fasync,
394 .llseek = no_llseek,
395};
396
397void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
398{
399 struct ib_uverbs_event_file *file = cq_context;
400 struct ib_ucq_object *uobj;
401 struct ib_uverbs_event *entry;
402 unsigned long flags;
403
404 if (!file)
405 return;
406
407 spin_lock_irqsave(&file->lock, flags);
408 if (file->is_closed) {
409 spin_unlock_irqrestore(&file->lock, flags);
410 return;
411 }
412
413 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
414 if (!entry) {
415 spin_unlock_irqrestore(&file->lock, flags);
416 return;
417 }
418
419 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
420
421 entry->desc.comp.cq_handle = cq->uobject->user_handle;
422 entry->counter = &uobj->comp_events_reported;
423
424 list_add_tail(&entry->list, &file->event_list);
425 list_add_tail(&entry->obj_list, &uobj->comp_list);
426 spin_unlock_irqrestore(&file->lock, flags);
427
428 wake_up_interruptible(&file->poll_wait);
429 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
430}
431
432static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
433 __u64 element, __u64 event,
434 struct list_head *obj_list,
435 u32 *counter)
436{
437 struct ib_uverbs_event *entry;
438 unsigned long flags;
439
440 spin_lock_irqsave(&file->async_file->lock, flags);
441 if (file->async_file->is_closed) {
442 spin_unlock_irqrestore(&file->async_file->lock, flags);
443 return;
444 }
445
446 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
447 if (!entry) {
448 spin_unlock_irqrestore(&file->async_file->lock, flags);
449 return;
450 }
451
452 entry->desc.async.element = element;
453 entry->desc.async.event_type = event;
454 entry->counter = counter;
455
456 list_add_tail(&entry->list, &file->async_file->event_list);
457 if (obj_list)
458 list_add_tail(&entry->obj_list, obj_list);
459 spin_unlock_irqrestore(&file->async_file->lock, flags);
460
461 wake_up_interruptible(&file->async_file->poll_wait);
462 kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
463}
464
465void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
466{
467 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
468 struct ib_ucq_object, uobject);
469
470 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
471 event->event, &uobj->async_list,
472 &uobj->async_events_reported);
473}
474
475void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
476{
477 struct ib_uevent_object *uobj;
478
479 uobj = container_of(event->element.qp->uobject,
480 struct ib_uevent_object, uobject);
481
482 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
483 event->event, &uobj->event_list,
484 &uobj->events_reported);
485}
486
487void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
488{
489 struct ib_uevent_object *uobj;
490
491 uobj = container_of(event->element.srq->uobject,
492 struct ib_uevent_object, uobject);
493
494 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
495 event->event, &uobj->event_list,
496 &uobj->events_reported);
497}
498
499void ib_uverbs_event_handler(struct ib_event_handler *handler,
500 struct ib_event *event)
501{
502 struct ib_uverbs_file *file =
503 container_of(handler, struct ib_uverbs_file, event_handler);
504
505 ib_uverbs_async_handler(file, event->element.port_num, event->event,
506 NULL, NULL);
507}
508
509struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
510 int is_async)
511{
512 struct ib_uverbs_event_file *ev_file;
513 struct file *filp;
514
515 ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
516 if (!ev_file)
517 return ERR_PTR(-ENOMEM);
518
519 kref_init(&ev_file->ref);
520 spin_lock_init(&ev_file->lock);
521 INIT_LIST_HEAD(&ev_file->event_list);
522 init_waitqueue_head(&ev_file->poll_wait);
523 ev_file->uverbs_file = uverbs_file;
524 ev_file->async_queue = NULL;
525 ev_file->is_async = is_async;
526 ev_file->is_closed = 0;
527
528 filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
529 ev_file, O_RDONLY);
530 if (IS_ERR(filp))
531 kfree(ev_file);
532
533 return filp;
534}
535
536/*
537 * Look up a completion event file by FD. If lookup is successful,
538 * takes a ref to the event file struct that it returns; if
539 * unsuccessful, returns NULL.
540 */
541struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
542{
543 struct ib_uverbs_event_file *ev_file = NULL;
544 struct file *filp;
545
546 filp = fget(fd);
547 if (!filp)
548 return NULL;
549
550 if (filp->f_op != &uverbs_event_fops)
551 goto out;
552
553 ev_file = filp->private_data;
554 if (ev_file->is_async) {
555 ev_file = NULL;
556 goto out;
557 }
558
559 kref_get(&ev_file->ref);
560
561out:
562 fput(filp);
563 return ev_file;
564}
565
566static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
567 size_t count, loff_t *pos)
568{
569 struct ib_uverbs_file *file = filp->private_data;
570 struct ib_uverbs_cmd_hdr hdr;
571
572 if (count < sizeof hdr)
573 return -EINVAL;
574
575 if (copy_from_user(&hdr, buf, sizeof hdr))
576 return -EFAULT;
577
578 if (hdr.in_words * 4 != count)
579 return -EINVAL;
580
581 if (hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
582 !uverbs_cmd_table[hdr.command])
583 return -EINVAL;
584
585 if (!file->ucontext &&
586 hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
587 return -EINVAL;
588
589 if (!(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
590 return -ENOSYS;
591
592 return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
593 hdr.in_words * 4, hdr.out_words * 4);
594}
595
596static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
597{
598 struct ib_uverbs_file *file = filp->private_data;
599
600 if (!file->ucontext)
601 return -ENODEV;
602 else
603 return file->device->ib_dev->mmap(file->ucontext, vma);
604}
605
606/*
607 * ib_uverbs_open() does not need the BKL:
608 *
609 * - the ib_uverbs_device structures are properly reference counted and
610 * everything else is purely local to the file being created, so
611 * races against other open calls are not a problem;
612 * - there is no ioctl method to race against;
613 * - the open method will either immediately run -ENXIO, or all
614 * required initialization will be done.
615 */
616static int ib_uverbs_open(struct inode *inode, struct file *filp)
617{
618 struct ib_uverbs_device *dev;
619 struct ib_uverbs_file *file;
620 int ret;
621
622 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
623 if (dev)
624 kref_get(&dev->ref);
625 else
626 return -ENXIO;
627
628 if (!try_module_get(dev->ib_dev->owner)) {
629 ret = -ENODEV;
630 goto err;
631 }
632
633 file = kmalloc(sizeof *file, GFP_KERNEL);
634 if (!file) {
635 ret = -ENOMEM;
636 goto err_module;
637 }
638
639 file->device = dev;
640 file->ucontext = NULL;
641 file->async_file = NULL;
642 kref_init(&file->ref);
643 mutex_init(&file->mutex);
644
645 filp->private_data = file;
646
647 return nonseekable_open(inode, filp);
648
649err_module:
650 module_put(dev->ib_dev->owner);
651
652err:
653 kref_put(&dev->ref, ib_uverbs_release_dev);
654 return ret;
655}
656
657static int ib_uverbs_close(struct inode *inode, struct file *filp)
658{
659 struct ib_uverbs_file *file = filp->private_data;
660
661 ib_uverbs_cleanup_ucontext(file, file->ucontext);
662
663 if (file->async_file)
664 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
665
666 kref_put(&file->ref, ib_uverbs_release_file);
667
668 return 0;
669}
670
671static const struct file_operations uverbs_fops = {
672 .owner = THIS_MODULE,
673 .write = ib_uverbs_write,
674 .open = ib_uverbs_open,
675 .release = ib_uverbs_close,
676 .llseek = no_llseek,
677};
678
679static const struct file_operations uverbs_mmap_fops = {
680 .owner = THIS_MODULE,
681 .write = ib_uverbs_write,
682 .mmap = ib_uverbs_mmap,
683 .open = ib_uverbs_open,
684 .release = ib_uverbs_close,
685 .llseek = no_llseek,
686};
687
688static struct ib_client uverbs_client = {
689 .name = "uverbs",
690 .add = ib_uverbs_add_one,
691 .remove = ib_uverbs_remove_one
692};
693
694static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
695 char *buf)
696{
697 struct ib_uverbs_device *dev = dev_get_drvdata(device);
698
699 if (!dev)
700 return -ENODEV;
701
702 return sprintf(buf, "%s\n", dev->ib_dev->name);
703}
704static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
705
706static ssize_t show_dev_abi_version(struct device *device,
707 struct device_attribute *attr, char *buf)
708{
709 struct ib_uverbs_device *dev = dev_get_drvdata(device);
710
711 if (!dev)
712 return -ENODEV;
713
714 return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
715}
716static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
717
718static CLASS_ATTR_STRING(abi_version, S_IRUGO,
719 __stringify(IB_USER_VERBS_ABI_VERSION));
720
721static dev_t overflow_maj;
722static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
723
724/*
725 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
726 * requesting a new major number and doubling the number of max devices we
727 * support. It's stupid, but simple.
728 */
729static int find_overflow_devnum(void)
730{
731 int ret;
732
733 if (!overflow_maj) {
734 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
735 "infiniband_verbs");
736 if (ret) {
737 printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
738 return ret;
739 }
740 }
741
742 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
743 if (ret >= IB_UVERBS_MAX_DEVICES)
744 return -1;
745
746 return ret;
747}
748
749static void ib_uverbs_add_one(struct ib_device *device)
750{
751 int devnum;
752 dev_t base;
753 struct ib_uverbs_device *uverbs_dev;
754
755 if (!device->alloc_ucontext)
756 return;
757
758 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
759 if (!uverbs_dev)
760 return;
761
762 kref_init(&uverbs_dev->ref);
763 init_completion(&uverbs_dev->comp);
764 uverbs_dev->xrcd_tree = RB_ROOT;
765 mutex_init(&uverbs_dev->xrcd_tree_mutex);
766
767 spin_lock(&map_lock);
768 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
769 if (devnum >= IB_UVERBS_MAX_DEVICES) {
770 spin_unlock(&map_lock);
771 devnum = find_overflow_devnum();
772 if (devnum < 0)
773 goto err;
774
775 spin_lock(&map_lock);
776 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
777 base = devnum + overflow_maj;
778 set_bit(devnum, overflow_map);
779 } else {
780 uverbs_dev->devnum = devnum;
781 base = devnum + IB_UVERBS_BASE_DEV;
782 set_bit(devnum, dev_map);
783 }
784 spin_unlock(&map_lock);
785
786 uverbs_dev->ib_dev = device;
787 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
788
789 cdev_init(&uverbs_dev->cdev, NULL);
790 uverbs_dev->cdev.owner = THIS_MODULE;
791 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
792 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
793 if (cdev_add(&uverbs_dev->cdev, base, 1))
794 goto err_cdev;
795
796 uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
797 uverbs_dev->cdev.dev, uverbs_dev,
798 "uverbs%d", uverbs_dev->devnum);
799 if (IS_ERR(uverbs_dev->dev))
800 goto err_cdev;
801
802 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
803 goto err_class;
804 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
805 goto err_class;
806
807 ib_set_client_data(device, &uverbs_client, uverbs_dev);
808
809 return;
810
811err_class:
812 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
813
814err_cdev:
815 cdev_del(&uverbs_dev->cdev);
816 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
817 clear_bit(devnum, dev_map);
818 else
819 clear_bit(devnum, overflow_map);
820
821err:
822 kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
823 wait_for_completion(&uverbs_dev->comp);
824 kfree(uverbs_dev);
825 return;
826}
827
828static void ib_uverbs_remove_one(struct ib_device *device)
829{
830 struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
831
832 if (!uverbs_dev)
833 return;
834
835 dev_set_drvdata(uverbs_dev->dev, NULL);
836 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
837 cdev_del(&uverbs_dev->cdev);
838
839 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
840 clear_bit(uverbs_dev->devnum, dev_map);
841 else
842 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
843
844 kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
845 wait_for_completion(&uverbs_dev->comp);
846 kfree(uverbs_dev);
847}
848
849static char *uverbs_devnode(struct device *dev, umode_t *mode)
850{
851 if (mode)
852 *mode = 0666;
853 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
854}
855
856static int __init ib_uverbs_init(void)
857{
858 int ret;
859
860 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
861 "infiniband_verbs");
862 if (ret) {
863 printk(KERN_ERR "user_verbs: couldn't register device number\n");
864 goto out;
865 }
866
867 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
868 if (IS_ERR(uverbs_class)) {
869 ret = PTR_ERR(uverbs_class);
870 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
871 goto out_chrdev;
872 }
873
874 uverbs_class->devnode = uverbs_devnode;
875
876 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
877 if (ret) {
878 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
879 goto out_class;
880 }
881
882 ret = ib_register_client(&uverbs_client);
883 if (ret) {
884 printk(KERN_ERR "user_verbs: couldn't register client\n");
885 goto out_class;
886 }
887
888 return 0;
889
890out_class:
891 class_destroy(uverbs_class);
892
893out_chrdev:
894 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
895
896out:
897 return ret;
898}
899
900static void __exit ib_uverbs_cleanup(void)
901{
902 ib_unregister_client(&uverbs_client);
903 class_destroy(uverbs_class);
904 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
905 if (overflow_maj)
906 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
907 idr_destroy(&ib_uverbs_pd_idr);
908 idr_destroy(&ib_uverbs_mr_idr);
909 idr_destroy(&ib_uverbs_mw_idr);
910 idr_destroy(&ib_uverbs_ah_idr);
911 idr_destroy(&ib_uverbs_cq_idr);
912 idr_destroy(&ib_uverbs_qp_idr);
913 idr_destroy(&ib_uverbs_srq_idr);
914}
915
916module_init(ib_uverbs_init);
917module_exit(ib_uverbs_cleanup);