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.
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:
14 * Redistribution and use in source and binary forms, with or
15 * without modification, are permitted provided that the following
18 * - Redistributions of source code must retain the above
19 * copyright notice, this list of conditions and the following
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.
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
37 #include <linux/module.h>
38 #include <linux/init.h>
39 #include <linux/device.h>
40 #include <linux/err.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>
49 #include <asm/uaccess.h>
55 MODULE_AUTHOR("Roland Dreier");
56 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
57 MODULE_LICENSE("Dual BSD/GPL");
60 IB_UVERBS_MAJOR = 231,
61 IB_UVERBS_BASE_MINOR = 192,
62 IB_UVERBS_MAX_DEVICES = 32
65 #define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
67 static struct class *uverbs_class;
69 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
70 DEFINE_IDR(ib_uverbs_pd_idr);
71 DEFINE_IDR(ib_uverbs_mr_idr);
72 DEFINE_IDR(ib_uverbs_mw_idr);
73 DEFINE_IDR(ib_uverbs_ah_idr);
74 DEFINE_IDR(ib_uverbs_cq_idr);
75 DEFINE_IDR(ib_uverbs_qp_idr);
76 DEFINE_IDR(ib_uverbs_srq_idr);
77 DEFINE_IDR(ib_uverbs_xrcd_idr);
78 DEFINE_IDR(ib_uverbs_rule_idr);
80 static DEFINE_SPINLOCK(map_lock);
81 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
83 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
84 struct ib_device *ib_dev,
85 const char __user *buf, int in_len,
87 [IB_USER_VERBS_CMD_GET_CONTEXT] = ib_uverbs_get_context,
88 [IB_USER_VERBS_CMD_QUERY_DEVICE] = ib_uverbs_query_device,
89 [IB_USER_VERBS_CMD_QUERY_PORT] = ib_uverbs_query_port,
90 [IB_USER_VERBS_CMD_ALLOC_PD] = ib_uverbs_alloc_pd,
91 [IB_USER_VERBS_CMD_DEALLOC_PD] = ib_uverbs_dealloc_pd,
92 [IB_USER_VERBS_CMD_REG_MR] = ib_uverbs_reg_mr,
93 [IB_USER_VERBS_CMD_REREG_MR] = ib_uverbs_rereg_mr,
94 [IB_USER_VERBS_CMD_DEREG_MR] = ib_uverbs_dereg_mr,
95 [IB_USER_VERBS_CMD_ALLOC_MW] = ib_uverbs_alloc_mw,
96 [IB_USER_VERBS_CMD_DEALLOC_MW] = ib_uverbs_dealloc_mw,
97 [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
98 [IB_USER_VERBS_CMD_CREATE_CQ] = ib_uverbs_create_cq,
99 [IB_USER_VERBS_CMD_RESIZE_CQ] = ib_uverbs_resize_cq,
100 [IB_USER_VERBS_CMD_POLL_CQ] = ib_uverbs_poll_cq,
101 [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ] = ib_uverbs_req_notify_cq,
102 [IB_USER_VERBS_CMD_DESTROY_CQ] = ib_uverbs_destroy_cq,
103 [IB_USER_VERBS_CMD_CREATE_QP] = ib_uverbs_create_qp,
104 [IB_USER_VERBS_CMD_QUERY_QP] = ib_uverbs_query_qp,
105 [IB_USER_VERBS_CMD_MODIFY_QP] = ib_uverbs_modify_qp,
106 [IB_USER_VERBS_CMD_DESTROY_QP] = ib_uverbs_destroy_qp,
107 [IB_USER_VERBS_CMD_POST_SEND] = ib_uverbs_post_send,
108 [IB_USER_VERBS_CMD_POST_RECV] = ib_uverbs_post_recv,
109 [IB_USER_VERBS_CMD_POST_SRQ_RECV] = ib_uverbs_post_srq_recv,
110 [IB_USER_VERBS_CMD_CREATE_AH] = ib_uverbs_create_ah,
111 [IB_USER_VERBS_CMD_DESTROY_AH] = ib_uverbs_destroy_ah,
112 [IB_USER_VERBS_CMD_ATTACH_MCAST] = ib_uverbs_attach_mcast,
113 [IB_USER_VERBS_CMD_DETACH_MCAST] = ib_uverbs_detach_mcast,
114 [IB_USER_VERBS_CMD_CREATE_SRQ] = ib_uverbs_create_srq,
115 [IB_USER_VERBS_CMD_MODIFY_SRQ] = ib_uverbs_modify_srq,
116 [IB_USER_VERBS_CMD_QUERY_SRQ] = ib_uverbs_query_srq,
117 [IB_USER_VERBS_CMD_DESTROY_SRQ] = ib_uverbs_destroy_srq,
118 [IB_USER_VERBS_CMD_OPEN_XRCD] = ib_uverbs_open_xrcd,
119 [IB_USER_VERBS_CMD_CLOSE_XRCD] = ib_uverbs_close_xrcd,
120 [IB_USER_VERBS_CMD_CREATE_XSRQ] = ib_uverbs_create_xsrq,
121 [IB_USER_VERBS_CMD_OPEN_QP] = ib_uverbs_open_qp,
124 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
125 struct ib_device *ib_dev,
126 struct ib_udata *ucore,
127 struct ib_udata *uhw) = {
128 [IB_USER_VERBS_EX_CMD_CREATE_FLOW] = ib_uverbs_ex_create_flow,
129 [IB_USER_VERBS_EX_CMD_DESTROY_FLOW] = ib_uverbs_ex_destroy_flow,
130 [IB_USER_VERBS_EX_CMD_QUERY_DEVICE] = ib_uverbs_ex_query_device,
131 [IB_USER_VERBS_EX_CMD_CREATE_CQ] = ib_uverbs_ex_create_cq,
132 [IB_USER_VERBS_EX_CMD_CREATE_QP] = ib_uverbs_ex_create_qp,
135 static void ib_uverbs_add_one(struct ib_device *device);
136 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
138 static void ib_uverbs_release_dev(struct kobject *kobj)
140 struct ib_uverbs_device *dev =
141 container_of(kobj, struct ib_uverbs_device, kobj);
143 cleanup_srcu_struct(&dev->disassociate_srcu);
147 static struct kobj_type ib_uverbs_dev_ktype = {
148 .release = ib_uverbs_release_dev,
151 static void ib_uverbs_release_event_file(struct kref *ref)
153 struct ib_uverbs_event_file *file =
154 container_of(ref, struct ib_uverbs_event_file, ref);
159 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
160 struct ib_uverbs_event_file *ev_file,
161 struct ib_ucq_object *uobj)
163 struct ib_uverbs_event *evt, *tmp;
166 spin_lock_irq(&ev_file->lock);
167 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
168 list_del(&evt->list);
171 spin_unlock_irq(&ev_file->lock);
173 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
176 spin_lock_irq(&file->async_file->lock);
177 list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
178 list_del(&evt->list);
181 spin_unlock_irq(&file->async_file->lock);
184 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
185 struct ib_uevent_object *uobj)
187 struct ib_uverbs_event *evt, *tmp;
189 spin_lock_irq(&file->async_file->lock);
190 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
191 list_del(&evt->list);
194 spin_unlock_irq(&file->async_file->lock);
197 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
198 struct ib_uqp_object *uobj)
200 struct ib_uverbs_mcast_entry *mcast, *tmp;
202 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
203 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
204 list_del(&mcast->list);
209 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
210 struct ib_ucontext *context)
212 struct ib_uobject *uobj, *tmp;
214 context->closing = 1;
216 list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
217 struct ib_ah *ah = uobj->object;
219 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
224 /* Remove MWs before QPs, in order to support type 2A MWs. */
225 list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
226 struct ib_mw *mw = uobj->object;
228 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
233 list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
234 struct ib_flow *flow_id = uobj->object;
236 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
237 ib_destroy_flow(flow_id);
241 list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
242 struct ib_qp *qp = uobj->object;
243 struct ib_uqp_object *uqp =
244 container_of(uobj, struct ib_uqp_object, uevent.uobject);
246 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
247 if (qp != qp->real_qp) {
250 ib_uverbs_detach_umcast(qp, uqp);
253 ib_uverbs_release_uevent(file, &uqp->uevent);
257 list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
258 struct ib_srq *srq = uobj->object;
259 struct ib_uevent_object *uevent =
260 container_of(uobj, struct ib_uevent_object, uobject);
262 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
264 ib_uverbs_release_uevent(file, uevent);
268 list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
269 struct ib_cq *cq = uobj->object;
270 struct ib_uverbs_event_file *ev_file = cq->cq_context;
271 struct ib_ucq_object *ucq =
272 container_of(uobj, struct ib_ucq_object, uobject);
274 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
276 ib_uverbs_release_ucq(file, ev_file, ucq);
280 list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
281 struct ib_mr *mr = uobj->object;
283 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
288 mutex_lock(&file->device->xrcd_tree_mutex);
289 list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
290 struct ib_xrcd *xrcd = uobj->object;
291 struct ib_uxrcd_object *uxrcd =
292 container_of(uobj, struct ib_uxrcd_object, uobject);
294 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
295 ib_uverbs_dealloc_xrcd(file->device, xrcd);
298 mutex_unlock(&file->device->xrcd_tree_mutex);
300 list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
301 struct ib_pd *pd = uobj->object;
303 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
308 put_pid(context->tgid);
310 return context->device->dealloc_ucontext(context);
313 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
315 complete(&dev->comp);
318 static void ib_uverbs_release_file(struct kref *ref)
320 struct ib_uverbs_file *file =
321 container_of(ref, struct ib_uverbs_file, ref);
322 struct ib_device *ib_dev;
325 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
326 ib_dev = srcu_dereference(file->device->ib_dev,
327 &file->device->disassociate_srcu);
328 if (ib_dev && !ib_dev->disassociate_ucontext)
329 module_put(ib_dev->owner);
330 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
332 if (atomic_dec_and_test(&file->device->refcount))
333 ib_uverbs_comp_dev(file->device);
338 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
339 size_t count, loff_t *pos)
341 struct ib_uverbs_event_file *file = filp->private_data;
342 struct ib_uverbs_event *event;
346 spin_lock_irq(&file->lock);
348 while (list_empty(&file->event_list)) {
349 spin_unlock_irq(&file->lock);
351 if (filp->f_flags & O_NONBLOCK)
354 if (wait_event_interruptible(file->poll_wait,
355 (!list_empty(&file->event_list) ||
356 /* The barriers built into wait_event_interruptible()
357 * and wake_up() guarentee this will see the null set
360 !file->uverbs_file->device->ib_dev)))
363 /* If device was disassociated and no event exists set an error */
364 if (list_empty(&file->event_list) &&
365 !file->uverbs_file->device->ib_dev)
368 spin_lock_irq(&file->lock);
371 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
374 eventsz = sizeof (struct ib_uverbs_async_event_desc);
376 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
378 if (eventsz > count) {
382 list_del(file->event_list.next);
383 if (event->counter) {
385 list_del(&event->obj_list);
389 spin_unlock_irq(&file->lock);
392 if (copy_to_user(buf, event, eventsz))
403 static unsigned int ib_uverbs_event_poll(struct file *filp,
404 struct poll_table_struct *wait)
406 unsigned int pollflags = 0;
407 struct ib_uverbs_event_file *file = filp->private_data;
409 poll_wait(filp, &file->poll_wait, wait);
411 spin_lock_irq(&file->lock);
412 if (!list_empty(&file->event_list))
413 pollflags = POLLIN | POLLRDNORM;
414 spin_unlock_irq(&file->lock);
419 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
421 struct ib_uverbs_event_file *file = filp->private_data;
423 return fasync_helper(fd, filp, on, &file->async_queue);
426 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
428 struct ib_uverbs_event_file *file = filp->private_data;
429 struct ib_uverbs_event *entry, *tmp;
430 int closed_already = 0;
432 mutex_lock(&file->uverbs_file->device->lists_mutex);
433 spin_lock_irq(&file->lock);
434 closed_already = file->is_closed;
436 list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
438 list_del(&entry->obj_list);
441 spin_unlock_irq(&file->lock);
442 if (!closed_already) {
443 list_del(&file->list);
445 ib_unregister_event_handler(&file->uverbs_file->
448 mutex_unlock(&file->uverbs_file->device->lists_mutex);
450 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
451 kref_put(&file->ref, ib_uverbs_release_event_file);
456 static const struct file_operations uverbs_event_fops = {
457 .owner = THIS_MODULE,
458 .read = ib_uverbs_event_read,
459 .poll = ib_uverbs_event_poll,
460 .release = ib_uverbs_event_close,
461 .fasync = ib_uverbs_event_fasync,
465 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
467 struct ib_uverbs_event_file *file = cq_context;
468 struct ib_ucq_object *uobj;
469 struct ib_uverbs_event *entry;
475 spin_lock_irqsave(&file->lock, flags);
476 if (file->is_closed) {
477 spin_unlock_irqrestore(&file->lock, flags);
481 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
483 spin_unlock_irqrestore(&file->lock, flags);
487 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
489 entry->desc.comp.cq_handle = cq->uobject->user_handle;
490 entry->counter = &uobj->comp_events_reported;
492 list_add_tail(&entry->list, &file->event_list);
493 list_add_tail(&entry->obj_list, &uobj->comp_list);
494 spin_unlock_irqrestore(&file->lock, flags);
496 wake_up_interruptible(&file->poll_wait);
497 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
500 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
501 __u64 element, __u64 event,
502 struct list_head *obj_list,
505 struct ib_uverbs_event *entry;
508 spin_lock_irqsave(&file->async_file->lock, flags);
509 if (file->async_file->is_closed) {
510 spin_unlock_irqrestore(&file->async_file->lock, flags);
514 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
516 spin_unlock_irqrestore(&file->async_file->lock, flags);
520 entry->desc.async.element = element;
521 entry->desc.async.event_type = event;
522 entry->desc.async.reserved = 0;
523 entry->counter = counter;
525 list_add_tail(&entry->list, &file->async_file->event_list);
527 list_add_tail(&entry->obj_list, obj_list);
528 spin_unlock_irqrestore(&file->async_file->lock, flags);
530 wake_up_interruptible(&file->async_file->poll_wait);
531 kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
534 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
536 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
537 struct ib_ucq_object, uobject);
539 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
540 event->event, &uobj->async_list,
541 &uobj->async_events_reported);
544 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
546 struct ib_uevent_object *uobj;
548 /* for XRC target qp's, check that qp is live */
549 if (!event->element.qp->uobject || !event->element.qp->uobject->live)
552 uobj = container_of(event->element.qp->uobject,
553 struct ib_uevent_object, uobject);
555 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
556 event->event, &uobj->event_list,
557 &uobj->events_reported);
560 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
562 struct ib_uevent_object *uobj;
564 uobj = container_of(event->element.srq->uobject,
565 struct ib_uevent_object, uobject);
567 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
568 event->event, &uobj->event_list,
569 &uobj->events_reported);
572 void ib_uverbs_event_handler(struct ib_event_handler *handler,
573 struct ib_event *event)
575 struct ib_uverbs_file *file =
576 container_of(handler, struct ib_uverbs_file, event_handler);
578 ib_uverbs_async_handler(file, event->element.port_num, event->event,
582 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
584 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
585 file->async_file = NULL;
588 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
589 struct ib_device *ib_dev,
592 struct ib_uverbs_event_file *ev_file;
596 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
598 return ERR_PTR(-ENOMEM);
600 kref_init(&ev_file->ref);
601 spin_lock_init(&ev_file->lock);
602 INIT_LIST_HEAD(&ev_file->event_list);
603 init_waitqueue_head(&ev_file->poll_wait);
604 ev_file->uverbs_file = uverbs_file;
605 kref_get(&ev_file->uverbs_file->ref);
606 ev_file->async_queue = NULL;
607 ev_file->is_closed = 0;
609 filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
614 mutex_lock(&uverbs_file->device->lists_mutex);
615 list_add_tail(&ev_file->list,
616 &uverbs_file->device->uverbs_events_file_list);
617 mutex_unlock(&uverbs_file->device->lists_mutex);
620 WARN_ON(uverbs_file->async_file);
621 uverbs_file->async_file = ev_file;
622 kref_get(&uverbs_file->async_file->ref);
623 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
625 ib_uverbs_event_handler);
626 ret = ib_register_event_handler(&uverbs_file->event_handler);
630 /* At that point async file stuff was fully set */
631 ev_file->is_async = 1;
638 kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
639 uverbs_file->async_file = NULL;
643 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
644 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
649 * Look up a completion event file by FD. If lookup is successful,
650 * takes a ref to the event file struct that it returns; if
651 * unsuccessful, returns NULL.
653 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
655 struct ib_uverbs_event_file *ev_file = NULL;
656 struct fd f = fdget(fd);
661 if (f.file->f_op != &uverbs_event_fops)
664 ev_file = f.file->private_data;
665 if (ev_file->is_async) {
670 kref_get(&ev_file->ref);
677 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
678 size_t count, loff_t *pos)
680 struct ib_uverbs_file *file = filp->private_data;
681 struct ib_device *ib_dev;
682 struct ib_uverbs_cmd_hdr hdr;
687 if (WARN_ON_ONCE(!ib_safe_file_access(filp)))
690 if (count < sizeof hdr)
693 if (copy_from_user(&hdr, buf, sizeof hdr))
696 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
697 ib_dev = srcu_dereference(file->device->ib_dev,
698 &file->device->disassociate_srcu);
704 flags = (hdr.command &
705 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
710 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
711 IB_USER_VERBS_CMD_COMMAND_MASK)) {
716 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
718 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
719 !uverbs_cmd_table[command]) {
724 if (!file->ucontext &&
725 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
730 if (!(ib_dev->uverbs_cmd_mask & (1ull << command))) {
735 if (hdr.in_words * 4 != count) {
740 ret = uverbs_cmd_table[command](file, ib_dev,
745 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
748 struct ib_uverbs_ex_cmd_hdr ex_hdr;
749 struct ib_udata ucore;
751 size_t written_count = count;
753 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
754 IB_USER_VERBS_CMD_COMMAND_MASK)) {
759 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
761 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
762 !uverbs_ex_cmd_table[command]) {
767 if (!file->ucontext) {
772 if (!(ib_dev->uverbs_ex_cmd_mask & (1ull << command))) {
777 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
782 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
787 count -= sizeof(hdr) + sizeof(ex_hdr);
788 buf += sizeof(hdr) + sizeof(ex_hdr);
790 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
795 if (ex_hdr.cmd_hdr_reserved) {
800 if (ex_hdr.response) {
801 if (!hdr.out_words && !ex_hdr.provider_out_words) {
806 if (!access_ok(VERIFY_WRITE,
807 (void __user *) (unsigned long) ex_hdr.response,
808 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
813 if (hdr.out_words || ex_hdr.provider_out_words) {
819 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
820 hdr.in_words * 8, hdr.out_words * 8);
822 INIT_UDATA_BUF_OR_NULL(&uhw,
824 (unsigned long) ex_hdr.response + ucore.outlen,
825 ex_hdr.provider_in_words * 8,
826 ex_hdr.provider_out_words * 8);
828 ret = uverbs_ex_cmd_table[command](file,
839 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
843 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
845 struct ib_uverbs_file *file = filp->private_data;
846 struct ib_device *ib_dev;
850 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
851 ib_dev = srcu_dereference(file->device->ib_dev,
852 &file->device->disassociate_srcu);
861 ret = ib_dev->mmap(file->ucontext, vma);
863 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
868 * ib_uverbs_open() does not need the BKL:
870 * - the ib_uverbs_device structures are properly reference counted and
871 * everything else is purely local to the file being created, so
872 * races against other open calls are not a problem;
873 * - there is no ioctl method to race against;
874 * - the open method will either immediately run -ENXIO, or all
875 * required initialization will be done.
877 static int ib_uverbs_open(struct inode *inode, struct file *filp)
879 struct ib_uverbs_device *dev;
880 struct ib_uverbs_file *file;
881 struct ib_device *ib_dev;
883 int module_dependent;
886 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
887 if (!atomic_inc_not_zero(&dev->refcount))
890 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
891 mutex_lock(&dev->lists_mutex);
892 ib_dev = srcu_dereference(dev->ib_dev,
893 &dev->disassociate_srcu);
899 /* In case IB device supports disassociate ucontext, there is no hard
900 * dependency between uverbs device and its low level device.
902 module_dependent = !(ib_dev->disassociate_ucontext);
904 if (module_dependent) {
905 if (!try_module_get(ib_dev->owner)) {
911 file = kzalloc(sizeof(*file), GFP_KERNEL);
914 if (module_dependent)
921 file->ucontext = NULL;
922 file->async_file = NULL;
923 kref_init(&file->ref);
924 mutex_init(&file->mutex);
926 filp->private_data = file;
927 kobject_get(&dev->kobj);
928 list_add_tail(&file->list, &dev->uverbs_file_list);
929 mutex_unlock(&dev->lists_mutex);
930 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
932 return nonseekable_open(inode, filp);
935 module_put(ib_dev->owner);
938 mutex_unlock(&dev->lists_mutex);
939 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
940 if (atomic_dec_and_test(&dev->refcount))
941 ib_uverbs_comp_dev(dev);
946 static int ib_uverbs_close(struct inode *inode, struct file *filp)
948 struct ib_uverbs_file *file = filp->private_data;
949 struct ib_uverbs_device *dev = file->device;
950 struct ib_ucontext *ucontext = NULL;
952 mutex_lock(&file->device->lists_mutex);
953 ucontext = file->ucontext;
954 file->ucontext = NULL;
955 if (!file->is_closed) {
956 list_del(&file->list);
959 mutex_unlock(&file->device->lists_mutex);
961 ib_uverbs_cleanup_ucontext(file, ucontext);
963 if (file->async_file)
964 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
966 kref_put(&file->ref, ib_uverbs_release_file);
967 kobject_put(&dev->kobj);
972 static const struct file_operations uverbs_fops = {
973 .owner = THIS_MODULE,
974 .write = ib_uverbs_write,
975 .open = ib_uverbs_open,
976 .release = ib_uverbs_close,
980 static const struct file_operations uverbs_mmap_fops = {
981 .owner = THIS_MODULE,
982 .write = ib_uverbs_write,
983 .mmap = ib_uverbs_mmap,
984 .open = ib_uverbs_open,
985 .release = ib_uverbs_close,
989 static struct ib_client uverbs_client = {
991 .add = ib_uverbs_add_one,
992 .remove = ib_uverbs_remove_one
995 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
1000 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1001 struct ib_device *ib_dev;
1006 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1007 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1009 ret = sprintf(buf, "%s\n", ib_dev->name);
1010 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1014 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1016 static ssize_t show_dev_abi_version(struct device *device,
1017 struct device_attribute *attr, char *buf)
1019 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1022 struct ib_device *ib_dev;
1026 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1027 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1029 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1030 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1034 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1036 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1037 __stringify(IB_USER_VERBS_ABI_VERSION));
1039 static dev_t overflow_maj;
1040 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1043 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1044 * requesting a new major number and doubling the number of max devices we
1045 * support. It's stupid, but simple.
1047 static int find_overflow_devnum(void)
1051 if (!overflow_maj) {
1052 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1053 "infiniband_verbs");
1055 printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
1060 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1061 if (ret >= IB_UVERBS_MAX_DEVICES)
1067 static void ib_uverbs_add_one(struct ib_device *device)
1071 struct ib_uverbs_device *uverbs_dev;
1074 if (!device->alloc_ucontext)
1077 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1081 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1087 atomic_set(&uverbs_dev->refcount, 1);
1088 init_completion(&uverbs_dev->comp);
1089 uverbs_dev->xrcd_tree = RB_ROOT;
1090 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1091 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1092 mutex_init(&uverbs_dev->lists_mutex);
1093 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1094 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1096 spin_lock(&map_lock);
1097 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1098 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1099 spin_unlock(&map_lock);
1100 devnum = find_overflow_devnum();
1104 spin_lock(&map_lock);
1105 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1106 base = devnum + overflow_maj;
1107 set_bit(devnum, overflow_map);
1109 uverbs_dev->devnum = devnum;
1110 base = devnum + IB_UVERBS_BASE_DEV;
1111 set_bit(devnum, dev_map);
1113 spin_unlock(&map_lock);
1115 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1116 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1118 cdev_init(&uverbs_dev->cdev, NULL);
1119 uverbs_dev->cdev.owner = THIS_MODULE;
1120 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1121 uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1122 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1123 if (cdev_add(&uverbs_dev->cdev, base, 1))
1126 uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
1127 uverbs_dev->cdev.dev, uverbs_dev,
1128 "uverbs%d", uverbs_dev->devnum);
1129 if (IS_ERR(uverbs_dev->dev))
1132 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1134 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1137 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1142 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1145 cdev_del(&uverbs_dev->cdev);
1146 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1147 clear_bit(devnum, dev_map);
1149 clear_bit(devnum, overflow_map);
1152 if (atomic_dec_and_test(&uverbs_dev->refcount))
1153 ib_uverbs_comp_dev(uverbs_dev);
1154 wait_for_completion(&uverbs_dev->comp);
1155 kobject_put(&uverbs_dev->kobj);
1159 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1160 struct ib_device *ib_dev)
1162 struct ib_uverbs_file *file;
1163 struct ib_uverbs_event_file *event_file;
1164 struct ib_event event;
1166 /* Pending running commands to terminate */
1167 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1168 event.event = IB_EVENT_DEVICE_FATAL;
1169 event.element.port_num = 0;
1170 event.device = ib_dev;
1172 mutex_lock(&uverbs_dev->lists_mutex);
1173 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1174 struct ib_ucontext *ucontext;
1176 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1177 struct ib_uverbs_file, list);
1178 file->is_closed = 1;
1179 ucontext = file->ucontext;
1180 list_del(&file->list);
1181 file->ucontext = NULL;
1182 kref_get(&file->ref);
1183 mutex_unlock(&uverbs_dev->lists_mutex);
1184 /* We must release the mutex before going ahead and calling
1185 * disassociate_ucontext. disassociate_ucontext might end up
1186 * indirectly calling uverbs_close, for example due to freeing
1187 * the resources (e.g mmput).
1189 ib_uverbs_event_handler(&file->event_handler, &event);
1191 ib_dev->disassociate_ucontext(ucontext);
1192 ib_uverbs_cleanup_ucontext(file, ucontext);
1195 mutex_lock(&uverbs_dev->lists_mutex);
1196 kref_put(&file->ref, ib_uverbs_release_file);
1199 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1200 event_file = list_first_entry(&uverbs_dev->
1201 uverbs_events_file_list,
1202 struct ib_uverbs_event_file,
1204 spin_lock_irq(&event_file->lock);
1205 event_file->is_closed = 1;
1206 spin_unlock_irq(&event_file->lock);
1208 list_del(&event_file->list);
1209 if (event_file->is_async) {
1210 ib_unregister_event_handler(&event_file->uverbs_file->
1212 event_file->uverbs_file->event_handler.device = NULL;
1215 wake_up_interruptible(&event_file->poll_wait);
1216 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1218 mutex_unlock(&uverbs_dev->lists_mutex);
1221 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1223 struct ib_uverbs_device *uverbs_dev = client_data;
1224 int wait_clients = 1;
1229 dev_set_drvdata(uverbs_dev->dev, NULL);
1230 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1231 cdev_del(&uverbs_dev->cdev);
1233 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1234 clear_bit(uverbs_dev->devnum, dev_map);
1236 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1238 if (device->disassociate_ucontext) {
1239 /* We disassociate HW resources and immediately return.
1240 * Userspace will see a EIO errno for all future access.
1241 * Upon returning, ib_device may be freed internally and is not
1243 * uverbs_device is still available until all clients close
1244 * their files, then the uverbs device ref count will be zero
1245 * and its resources will be freed.
1246 * Note: At this point no more files can be opened since the
1247 * cdev was deleted, however active clients can still issue
1248 * commands and close their open files.
1250 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1251 ib_uverbs_free_hw_resources(uverbs_dev, device);
1255 if (atomic_dec_and_test(&uverbs_dev->refcount))
1256 ib_uverbs_comp_dev(uverbs_dev);
1258 wait_for_completion(&uverbs_dev->comp);
1259 kobject_put(&uverbs_dev->kobj);
1262 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1266 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1269 static int __init ib_uverbs_init(void)
1273 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1274 "infiniband_verbs");
1276 printk(KERN_ERR "user_verbs: couldn't register device number\n");
1280 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1281 if (IS_ERR(uverbs_class)) {
1282 ret = PTR_ERR(uverbs_class);
1283 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
1287 uverbs_class->devnode = uverbs_devnode;
1289 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1291 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
1295 ret = ib_register_client(&uverbs_client);
1297 printk(KERN_ERR "user_verbs: couldn't register client\n");
1304 class_destroy(uverbs_class);
1307 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1313 static void __exit ib_uverbs_cleanup(void)
1315 ib_unregister_client(&uverbs_client);
1316 class_destroy(uverbs_class);
1317 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1319 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1320 idr_destroy(&ib_uverbs_pd_idr);
1321 idr_destroy(&ib_uverbs_mr_idr);
1322 idr_destroy(&ib_uverbs_mw_idr);
1323 idr_destroy(&ib_uverbs_ah_idr);
1324 idr_destroy(&ib_uverbs_cq_idr);
1325 idr_destroy(&ib_uverbs_qp_idr);
1326 idr_destroy(&ib_uverbs_srq_idr);
1329 module_init(ib_uverbs_init);
1330 module_exit(ib_uverbs_cleanup);