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);
79 DEFINE_IDR(ib_uverbs_wq_idr);
81 static DEFINE_SPINLOCK(map_lock);
82 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
84 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
85 struct ib_device *ib_dev,
86 const char __user *buf, int in_len,
88 [IB_USER_VERBS_CMD_GET_CONTEXT] = ib_uverbs_get_context,
89 [IB_USER_VERBS_CMD_QUERY_DEVICE] = ib_uverbs_query_device,
90 [IB_USER_VERBS_CMD_QUERY_PORT] = ib_uverbs_query_port,
91 [IB_USER_VERBS_CMD_ALLOC_PD] = ib_uverbs_alloc_pd,
92 [IB_USER_VERBS_CMD_DEALLOC_PD] = ib_uverbs_dealloc_pd,
93 [IB_USER_VERBS_CMD_REG_MR] = ib_uverbs_reg_mr,
94 [IB_USER_VERBS_CMD_REREG_MR] = ib_uverbs_rereg_mr,
95 [IB_USER_VERBS_CMD_DEREG_MR] = ib_uverbs_dereg_mr,
96 [IB_USER_VERBS_CMD_ALLOC_MW] = ib_uverbs_alloc_mw,
97 [IB_USER_VERBS_CMD_DEALLOC_MW] = ib_uverbs_dealloc_mw,
98 [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
99 [IB_USER_VERBS_CMD_CREATE_CQ] = ib_uverbs_create_cq,
100 [IB_USER_VERBS_CMD_RESIZE_CQ] = ib_uverbs_resize_cq,
101 [IB_USER_VERBS_CMD_POLL_CQ] = ib_uverbs_poll_cq,
102 [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ] = ib_uverbs_req_notify_cq,
103 [IB_USER_VERBS_CMD_DESTROY_CQ] = ib_uverbs_destroy_cq,
104 [IB_USER_VERBS_CMD_CREATE_QP] = ib_uverbs_create_qp,
105 [IB_USER_VERBS_CMD_QUERY_QP] = ib_uverbs_query_qp,
106 [IB_USER_VERBS_CMD_MODIFY_QP] = ib_uverbs_modify_qp,
107 [IB_USER_VERBS_CMD_DESTROY_QP] = ib_uverbs_destroy_qp,
108 [IB_USER_VERBS_CMD_POST_SEND] = ib_uverbs_post_send,
109 [IB_USER_VERBS_CMD_POST_RECV] = ib_uverbs_post_recv,
110 [IB_USER_VERBS_CMD_POST_SRQ_RECV] = ib_uverbs_post_srq_recv,
111 [IB_USER_VERBS_CMD_CREATE_AH] = ib_uverbs_create_ah,
112 [IB_USER_VERBS_CMD_DESTROY_AH] = ib_uverbs_destroy_ah,
113 [IB_USER_VERBS_CMD_ATTACH_MCAST] = ib_uverbs_attach_mcast,
114 [IB_USER_VERBS_CMD_DETACH_MCAST] = ib_uverbs_detach_mcast,
115 [IB_USER_VERBS_CMD_CREATE_SRQ] = ib_uverbs_create_srq,
116 [IB_USER_VERBS_CMD_MODIFY_SRQ] = ib_uverbs_modify_srq,
117 [IB_USER_VERBS_CMD_QUERY_SRQ] = ib_uverbs_query_srq,
118 [IB_USER_VERBS_CMD_DESTROY_SRQ] = ib_uverbs_destroy_srq,
119 [IB_USER_VERBS_CMD_OPEN_XRCD] = ib_uverbs_open_xrcd,
120 [IB_USER_VERBS_CMD_CLOSE_XRCD] = ib_uverbs_close_xrcd,
121 [IB_USER_VERBS_CMD_CREATE_XSRQ] = ib_uverbs_create_xsrq,
122 [IB_USER_VERBS_CMD_OPEN_QP] = ib_uverbs_open_qp,
125 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
126 struct ib_device *ib_dev,
127 struct ib_udata *ucore,
128 struct ib_udata *uhw) = {
129 [IB_USER_VERBS_EX_CMD_CREATE_FLOW] = ib_uverbs_ex_create_flow,
130 [IB_USER_VERBS_EX_CMD_DESTROY_FLOW] = ib_uverbs_ex_destroy_flow,
131 [IB_USER_VERBS_EX_CMD_QUERY_DEVICE] = ib_uverbs_ex_query_device,
132 [IB_USER_VERBS_EX_CMD_CREATE_CQ] = ib_uverbs_ex_create_cq,
133 [IB_USER_VERBS_EX_CMD_CREATE_QP] = ib_uverbs_ex_create_qp,
134 [IB_USER_VERBS_EX_CMD_CREATE_WQ] = ib_uverbs_ex_create_wq,
135 [IB_USER_VERBS_EX_CMD_MODIFY_WQ] = ib_uverbs_ex_modify_wq,
136 [IB_USER_VERBS_EX_CMD_DESTROY_WQ] = ib_uverbs_ex_destroy_wq,
139 static void ib_uverbs_add_one(struct ib_device *device);
140 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
142 int uverbs_dealloc_mw(struct ib_mw *mw)
144 struct ib_pd *pd = mw->pd;
147 ret = mw->device->dealloc_mw(mw);
149 atomic_dec(&pd->usecnt);
153 static void ib_uverbs_release_dev(struct kobject *kobj)
155 struct ib_uverbs_device *dev =
156 container_of(kobj, struct ib_uverbs_device, kobj);
158 cleanup_srcu_struct(&dev->disassociate_srcu);
162 static struct kobj_type ib_uverbs_dev_ktype = {
163 .release = ib_uverbs_release_dev,
166 static void ib_uverbs_release_event_file(struct kref *ref)
168 struct ib_uverbs_event_file *file =
169 container_of(ref, struct ib_uverbs_event_file, ref);
174 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
175 struct ib_uverbs_event_file *ev_file,
176 struct ib_ucq_object *uobj)
178 struct ib_uverbs_event *evt, *tmp;
181 spin_lock_irq(&ev_file->lock);
182 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
183 list_del(&evt->list);
186 spin_unlock_irq(&ev_file->lock);
188 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
191 spin_lock_irq(&file->async_file->lock);
192 list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
193 list_del(&evt->list);
196 spin_unlock_irq(&file->async_file->lock);
199 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
200 struct ib_uevent_object *uobj)
202 struct ib_uverbs_event *evt, *tmp;
204 spin_lock_irq(&file->async_file->lock);
205 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
206 list_del(&evt->list);
209 spin_unlock_irq(&file->async_file->lock);
212 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
213 struct ib_uqp_object *uobj)
215 struct ib_uverbs_mcast_entry *mcast, *tmp;
217 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
218 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
219 list_del(&mcast->list);
224 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
225 struct ib_ucontext *context)
227 struct ib_uobject *uobj, *tmp;
229 context->closing = 1;
231 list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
232 struct ib_ah *ah = uobj->object;
234 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
239 /* Remove MWs before QPs, in order to support type 2A MWs. */
240 list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
241 struct ib_mw *mw = uobj->object;
243 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
244 uverbs_dealloc_mw(mw);
248 list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
249 struct ib_flow *flow_id = uobj->object;
251 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
252 ib_destroy_flow(flow_id);
256 list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
257 struct ib_qp *qp = uobj->object;
258 struct ib_uqp_object *uqp =
259 container_of(uobj, struct ib_uqp_object, uevent.uobject);
261 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
262 if (qp != qp->real_qp) {
265 ib_uverbs_detach_umcast(qp, uqp);
268 ib_uverbs_release_uevent(file, &uqp->uevent);
272 list_for_each_entry_safe(uobj, tmp, &context->wq_list, list) {
273 struct ib_wq *wq = uobj->object;
274 struct ib_uwq_object *uwq =
275 container_of(uobj, struct ib_uwq_object, uevent.uobject);
277 idr_remove_uobj(&ib_uverbs_wq_idr, uobj);
279 ib_uverbs_release_uevent(file, &uwq->uevent);
283 list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
284 struct ib_srq *srq = uobj->object;
285 struct ib_uevent_object *uevent =
286 container_of(uobj, struct ib_uevent_object, uobject);
288 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
290 ib_uverbs_release_uevent(file, uevent);
294 list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
295 struct ib_cq *cq = uobj->object;
296 struct ib_uverbs_event_file *ev_file = cq->cq_context;
297 struct ib_ucq_object *ucq =
298 container_of(uobj, struct ib_ucq_object, uobject);
300 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
302 ib_uverbs_release_ucq(file, ev_file, ucq);
306 list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
307 struct ib_mr *mr = uobj->object;
309 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
314 mutex_lock(&file->device->xrcd_tree_mutex);
315 list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
316 struct ib_xrcd *xrcd = uobj->object;
317 struct ib_uxrcd_object *uxrcd =
318 container_of(uobj, struct ib_uxrcd_object, uobject);
320 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
321 ib_uverbs_dealloc_xrcd(file->device, xrcd);
324 mutex_unlock(&file->device->xrcd_tree_mutex);
326 list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
327 struct ib_pd *pd = uobj->object;
329 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
334 put_pid(context->tgid);
336 return context->device->dealloc_ucontext(context);
339 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
341 complete(&dev->comp);
344 static void ib_uverbs_release_file(struct kref *ref)
346 struct ib_uverbs_file *file =
347 container_of(ref, struct ib_uverbs_file, ref);
348 struct ib_device *ib_dev;
351 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
352 ib_dev = srcu_dereference(file->device->ib_dev,
353 &file->device->disassociate_srcu);
354 if (ib_dev && !ib_dev->disassociate_ucontext)
355 module_put(ib_dev->owner);
356 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
358 if (atomic_dec_and_test(&file->device->refcount))
359 ib_uverbs_comp_dev(file->device);
364 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
365 size_t count, loff_t *pos)
367 struct ib_uverbs_event_file *file = filp->private_data;
368 struct ib_uverbs_event *event;
372 spin_lock_irq(&file->lock);
374 while (list_empty(&file->event_list)) {
375 spin_unlock_irq(&file->lock);
377 if (filp->f_flags & O_NONBLOCK)
380 if (wait_event_interruptible(file->poll_wait,
381 (!list_empty(&file->event_list) ||
382 /* The barriers built into wait_event_interruptible()
383 * and wake_up() guarentee this will see the null set
386 !file->uverbs_file->device->ib_dev)))
389 /* If device was disassociated and no event exists set an error */
390 if (list_empty(&file->event_list) &&
391 !file->uverbs_file->device->ib_dev)
394 spin_lock_irq(&file->lock);
397 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
400 eventsz = sizeof (struct ib_uverbs_async_event_desc);
402 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
404 if (eventsz > count) {
408 list_del(file->event_list.next);
409 if (event->counter) {
411 list_del(&event->obj_list);
415 spin_unlock_irq(&file->lock);
418 if (copy_to_user(buf, event, eventsz))
429 static unsigned int ib_uverbs_event_poll(struct file *filp,
430 struct poll_table_struct *wait)
432 unsigned int pollflags = 0;
433 struct ib_uverbs_event_file *file = filp->private_data;
435 poll_wait(filp, &file->poll_wait, wait);
437 spin_lock_irq(&file->lock);
438 if (!list_empty(&file->event_list))
439 pollflags = POLLIN | POLLRDNORM;
440 spin_unlock_irq(&file->lock);
445 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
447 struct ib_uverbs_event_file *file = filp->private_data;
449 return fasync_helper(fd, filp, on, &file->async_queue);
452 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
454 struct ib_uverbs_event_file *file = filp->private_data;
455 struct ib_uverbs_event *entry, *tmp;
456 int closed_already = 0;
458 mutex_lock(&file->uverbs_file->device->lists_mutex);
459 spin_lock_irq(&file->lock);
460 closed_already = file->is_closed;
462 list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
464 list_del(&entry->obj_list);
467 spin_unlock_irq(&file->lock);
468 if (!closed_already) {
469 list_del(&file->list);
471 ib_unregister_event_handler(&file->uverbs_file->
474 mutex_unlock(&file->uverbs_file->device->lists_mutex);
476 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
477 kref_put(&file->ref, ib_uverbs_release_event_file);
482 static const struct file_operations uverbs_event_fops = {
483 .owner = THIS_MODULE,
484 .read = ib_uverbs_event_read,
485 .poll = ib_uverbs_event_poll,
486 .release = ib_uverbs_event_close,
487 .fasync = ib_uverbs_event_fasync,
491 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
493 struct ib_uverbs_event_file *file = cq_context;
494 struct ib_ucq_object *uobj;
495 struct ib_uverbs_event *entry;
501 spin_lock_irqsave(&file->lock, flags);
502 if (file->is_closed) {
503 spin_unlock_irqrestore(&file->lock, flags);
507 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
509 spin_unlock_irqrestore(&file->lock, flags);
513 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
515 entry->desc.comp.cq_handle = cq->uobject->user_handle;
516 entry->counter = &uobj->comp_events_reported;
518 list_add_tail(&entry->list, &file->event_list);
519 list_add_tail(&entry->obj_list, &uobj->comp_list);
520 spin_unlock_irqrestore(&file->lock, flags);
522 wake_up_interruptible(&file->poll_wait);
523 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
526 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
527 __u64 element, __u64 event,
528 struct list_head *obj_list,
531 struct ib_uverbs_event *entry;
534 spin_lock_irqsave(&file->async_file->lock, flags);
535 if (file->async_file->is_closed) {
536 spin_unlock_irqrestore(&file->async_file->lock, flags);
540 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
542 spin_unlock_irqrestore(&file->async_file->lock, flags);
546 entry->desc.async.element = element;
547 entry->desc.async.event_type = event;
548 entry->desc.async.reserved = 0;
549 entry->counter = counter;
551 list_add_tail(&entry->list, &file->async_file->event_list);
553 list_add_tail(&entry->obj_list, obj_list);
554 spin_unlock_irqrestore(&file->async_file->lock, flags);
556 wake_up_interruptible(&file->async_file->poll_wait);
557 kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
560 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
562 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
563 struct ib_ucq_object, uobject);
565 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
566 event->event, &uobj->async_list,
567 &uobj->async_events_reported);
570 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
572 struct ib_uevent_object *uobj;
574 /* for XRC target qp's, check that qp is live */
575 if (!event->element.qp->uobject || !event->element.qp->uobject->live)
578 uobj = container_of(event->element.qp->uobject,
579 struct ib_uevent_object, uobject);
581 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
582 event->event, &uobj->event_list,
583 &uobj->events_reported);
586 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
588 struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
589 struct ib_uevent_object, uobject);
591 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
592 event->event, &uobj->event_list,
593 &uobj->events_reported);
596 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
598 struct ib_uevent_object *uobj;
600 uobj = container_of(event->element.srq->uobject,
601 struct ib_uevent_object, uobject);
603 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
604 event->event, &uobj->event_list,
605 &uobj->events_reported);
608 void ib_uverbs_event_handler(struct ib_event_handler *handler,
609 struct ib_event *event)
611 struct ib_uverbs_file *file =
612 container_of(handler, struct ib_uverbs_file, event_handler);
614 ib_uverbs_async_handler(file, event->element.port_num, event->event,
618 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
620 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
621 file->async_file = NULL;
624 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
625 struct ib_device *ib_dev,
628 struct ib_uverbs_event_file *ev_file;
632 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
634 return ERR_PTR(-ENOMEM);
636 kref_init(&ev_file->ref);
637 spin_lock_init(&ev_file->lock);
638 INIT_LIST_HEAD(&ev_file->event_list);
639 init_waitqueue_head(&ev_file->poll_wait);
640 ev_file->uverbs_file = uverbs_file;
641 kref_get(&ev_file->uverbs_file->ref);
642 ev_file->async_queue = NULL;
643 ev_file->is_closed = 0;
645 filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
650 mutex_lock(&uverbs_file->device->lists_mutex);
651 list_add_tail(&ev_file->list,
652 &uverbs_file->device->uverbs_events_file_list);
653 mutex_unlock(&uverbs_file->device->lists_mutex);
656 WARN_ON(uverbs_file->async_file);
657 uverbs_file->async_file = ev_file;
658 kref_get(&uverbs_file->async_file->ref);
659 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
661 ib_uverbs_event_handler);
662 ret = ib_register_event_handler(&uverbs_file->event_handler);
666 /* At that point async file stuff was fully set */
667 ev_file->is_async = 1;
674 kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
675 uverbs_file->async_file = NULL;
679 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
680 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
685 * Look up a completion event file by FD. If lookup is successful,
686 * takes a ref to the event file struct that it returns; if
687 * unsuccessful, returns NULL.
689 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
691 struct ib_uverbs_event_file *ev_file = NULL;
692 struct fd f = fdget(fd);
697 if (f.file->f_op != &uverbs_event_fops)
700 ev_file = f.file->private_data;
701 if (ev_file->is_async) {
706 kref_get(&ev_file->ref);
713 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
717 if (command <= IB_USER_VERBS_CMD_OPEN_QP)
718 mask = ib_dev->uverbs_cmd_mask;
720 mask = ib_dev->uverbs_ex_cmd_mask;
722 if (mask & ((u64)1 << command))
728 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
729 size_t count, loff_t *pos)
731 struct ib_uverbs_file *file = filp->private_data;
732 struct ib_device *ib_dev;
733 struct ib_uverbs_cmd_hdr hdr;
739 if (WARN_ON_ONCE(!ib_safe_file_access(filp)))
742 if (count < sizeof hdr)
745 if (copy_from_user(&hdr, buf, sizeof hdr))
748 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
749 ib_dev = srcu_dereference(file->device->ib_dev,
750 &file->device->disassociate_srcu);
756 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
757 IB_USER_VERBS_CMD_COMMAND_MASK)) {
762 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
763 if (verify_command_mask(ib_dev, command)) {
768 if (!file->ucontext &&
769 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
774 flags = (hdr.command &
775 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
778 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
779 !uverbs_cmd_table[command]) {
784 if (hdr.in_words * 4 != count) {
789 ret = uverbs_cmd_table[command](file, ib_dev,
794 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
795 struct ib_uverbs_ex_cmd_hdr ex_hdr;
796 struct ib_udata ucore;
798 size_t written_count = count;
800 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
801 !uverbs_ex_cmd_table[command]) {
806 if (!file->ucontext) {
811 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
816 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
821 count -= sizeof(hdr) + sizeof(ex_hdr);
822 buf += sizeof(hdr) + sizeof(ex_hdr);
824 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
829 if (ex_hdr.cmd_hdr_reserved) {
834 if (ex_hdr.response) {
835 if (!hdr.out_words && !ex_hdr.provider_out_words) {
840 if (!access_ok(VERIFY_WRITE,
841 (void __user *) (unsigned long) ex_hdr.response,
842 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
847 if (hdr.out_words || ex_hdr.provider_out_words) {
853 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
854 hdr.in_words * 8, hdr.out_words * 8);
856 INIT_UDATA_BUF_OR_NULL(&uhw,
858 (unsigned long) ex_hdr.response + ucore.outlen,
859 ex_hdr.provider_in_words * 8,
860 ex_hdr.provider_out_words * 8);
862 ret = uverbs_ex_cmd_table[command](file,
873 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
877 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
879 struct ib_uverbs_file *file = filp->private_data;
880 struct ib_device *ib_dev;
884 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
885 ib_dev = srcu_dereference(file->device->ib_dev,
886 &file->device->disassociate_srcu);
895 ret = ib_dev->mmap(file->ucontext, vma);
897 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
902 * ib_uverbs_open() does not need the BKL:
904 * - the ib_uverbs_device structures are properly reference counted and
905 * everything else is purely local to the file being created, so
906 * races against other open calls are not a problem;
907 * - there is no ioctl method to race against;
908 * - the open method will either immediately run -ENXIO, or all
909 * required initialization will be done.
911 static int ib_uverbs_open(struct inode *inode, struct file *filp)
913 struct ib_uverbs_device *dev;
914 struct ib_uverbs_file *file;
915 struct ib_device *ib_dev;
917 int module_dependent;
920 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
921 if (!atomic_inc_not_zero(&dev->refcount))
924 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
925 mutex_lock(&dev->lists_mutex);
926 ib_dev = srcu_dereference(dev->ib_dev,
927 &dev->disassociate_srcu);
933 /* In case IB device supports disassociate ucontext, there is no hard
934 * dependency between uverbs device and its low level device.
936 module_dependent = !(ib_dev->disassociate_ucontext);
938 if (module_dependent) {
939 if (!try_module_get(ib_dev->owner)) {
945 file = kzalloc(sizeof(*file), GFP_KERNEL);
948 if (module_dependent)
955 file->ucontext = NULL;
956 file->async_file = NULL;
957 kref_init(&file->ref);
958 mutex_init(&file->mutex);
960 filp->private_data = file;
961 kobject_get(&dev->kobj);
962 list_add_tail(&file->list, &dev->uverbs_file_list);
963 mutex_unlock(&dev->lists_mutex);
964 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
966 return nonseekable_open(inode, filp);
969 module_put(ib_dev->owner);
972 mutex_unlock(&dev->lists_mutex);
973 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
974 if (atomic_dec_and_test(&dev->refcount))
975 ib_uverbs_comp_dev(dev);
980 static int ib_uverbs_close(struct inode *inode, struct file *filp)
982 struct ib_uverbs_file *file = filp->private_data;
983 struct ib_uverbs_device *dev = file->device;
984 struct ib_ucontext *ucontext = NULL;
986 mutex_lock(&file->device->lists_mutex);
987 ucontext = file->ucontext;
988 file->ucontext = NULL;
989 if (!file->is_closed) {
990 list_del(&file->list);
993 mutex_unlock(&file->device->lists_mutex);
995 ib_uverbs_cleanup_ucontext(file, ucontext);
997 if (file->async_file)
998 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
1000 kref_put(&file->ref, ib_uverbs_release_file);
1001 kobject_put(&dev->kobj);
1006 static const struct file_operations uverbs_fops = {
1007 .owner = THIS_MODULE,
1008 .write = ib_uverbs_write,
1009 .open = ib_uverbs_open,
1010 .release = ib_uverbs_close,
1011 .llseek = no_llseek,
1014 static const struct file_operations uverbs_mmap_fops = {
1015 .owner = THIS_MODULE,
1016 .write = ib_uverbs_write,
1017 .mmap = ib_uverbs_mmap,
1018 .open = ib_uverbs_open,
1019 .release = ib_uverbs_close,
1020 .llseek = no_llseek,
1023 static struct ib_client uverbs_client = {
1025 .add = ib_uverbs_add_one,
1026 .remove = ib_uverbs_remove_one
1029 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
1034 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1035 struct ib_device *ib_dev;
1040 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1041 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1043 ret = sprintf(buf, "%s\n", ib_dev->name);
1044 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1048 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1050 static ssize_t show_dev_abi_version(struct device *device,
1051 struct device_attribute *attr, char *buf)
1053 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1056 struct ib_device *ib_dev;
1060 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1061 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1063 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1064 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1068 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1070 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1071 __stringify(IB_USER_VERBS_ABI_VERSION));
1073 static dev_t overflow_maj;
1074 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1077 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1078 * requesting a new major number and doubling the number of max devices we
1079 * support. It's stupid, but simple.
1081 static int find_overflow_devnum(void)
1085 if (!overflow_maj) {
1086 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1087 "infiniband_verbs");
1089 pr_err("user_verbs: couldn't register dynamic device number\n");
1094 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1095 if (ret >= IB_UVERBS_MAX_DEVICES)
1101 static void ib_uverbs_add_one(struct ib_device *device)
1105 struct ib_uverbs_device *uverbs_dev;
1108 if (!device->alloc_ucontext)
1111 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1115 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1121 atomic_set(&uverbs_dev->refcount, 1);
1122 init_completion(&uverbs_dev->comp);
1123 uverbs_dev->xrcd_tree = RB_ROOT;
1124 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1125 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1126 mutex_init(&uverbs_dev->lists_mutex);
1127 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1128 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1130 spin_lock(&map_lock);
1131 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1132 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1133 spin_unlock(&map_lock);
1134 devnum = find_overflow_devnum();
1138 spin_lock(&map_lock);
1139 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1140 base = devnum + overflow_maj;
1141 set_bit(devnum, overflow_map);
1143 uverbs_dev->devnum = devnum;
1144 base = devnum + IB_UVERBS_BASE_DEV;
1145 set_bit(devnum, dev_map);
1147 spin_unlock(&map_lock);
1149 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1150 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1152 cdev_init(&uverbs_dev->cdev, NULL);
1153 uverbs_dev->cdev.owner = THIS_MODULE;
1154 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1155 uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1156 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1157 if (cdev_add(&uverbs_dev->cdev, base, 1))
1160 uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
1161 uverbs_dev->cdev.dev, uverbs_dev,
1162 "uverbs%d", uverbs_dev->devnum);
1163 if (IS_ERR(uverbs_dev->dev))
1166 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1168 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1171 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1176 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1179 cdev_del(&uverbs_dev->cdev);
1180 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1181 clear_bit(devnum, dev_map);
1183 clear_bit(devnum, overflow_map);
1186 if (atomic_dec_and_test(&uverbs_dev->refcount))
1187 ib_uverbs_comp_dev(uverbs_dev);
1188 wait_for_completion(&uverbs_dev->comp);
1189 kobject_put(&uverbs_dev->kobj);
1193 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1194 struct ib_device *ib_dev)
1196 struct ib_uverbs_file *file;
1197 struct ib_uverbs_event_file *event_file;
1198 struct ib_event event;
1200 /* Pending running commands to terminate */
1201 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1202 event.event = IB_EVENT_DEVICE_FATAL;
1203 event.element.port_num = 0;
1204 event.device = ib_dev;
1206 mutex_lock(&uverbs_dev->lists_mutex);
1207 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1208 struct ib_ucontext *ucontext;
1210 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1211 struct ib_uverbs_file, list);
1212 file->is_closed = 1;
1213 ucontext = file->ucontext;
1214 list_del(&file->list);
1215 file->ucontext = NULL;
1216 kref_get(&file->ref);
1217 mutex_unlock(&uverbs_dev->lists_mutex);
1218 /* We must release the mutex before going ahead and calling
1219 * disassociate_ucontext. disassociate_ucontext might end up
1220 * indirectly calling uverbs_close, for example due to freeing
1221 * the resources (e.g mmput).
1223 ib_uverbs_event_handler(&file->event_handler, &event);
1225 ib_dev->disassociate_ucontext(ucontext);
1226 ib_uverbs_cleanup_ucontext(file, ucontext);
1229 mutex_lock(&uverbs_dev->lists_mutex);
1230 kref_put(&file->ref, ib_uverbs_release_file);
1233 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1234 event_file = list_first_entry(&uverbs_dev->
1235 uverbs_events_file_list,
1236 struct ib_uverbs_event_file,
1238 spin_lock_irq(&event_file->lock);
1239 event_file->is_closed = 1;
1240 spin_unlock_irq(&event_file->lock);
1242 list_del(&event_file->list);
1243 if (event_file->is_async) {
1244 ib_unregister_event_handler(&event_file->uverbs_file->
1246 event_file->uverbs_file->event_handler.device = NULL;
1249 wake_up_interruptible(&event_file->poll_wait);
1250 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1252 mutex_unlock(&uverbs_dev->lists_mutex);
1255 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1257 struct ib_uverbs_device *uverbs_dev = client_data;
1258 int wait_clients = 1;
1263 dev_set_drvdata(uverbs_dev->dev, NULL);
1264 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1265 cdev_del(&uverbs_dev->cdev);
1267 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1268 clear_bit(uverbs_dev->devnum, dev_map);
1270 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1272 if (device->disassociate_ucontext) {
1273 /* We disassociate HW resources and immediately return.
1274 * Userspace will see a EIO errno for all future access.
1275 * Upon returning, ib_device may be freed internally and is not
1277 * uverbs_device is still available until all clients close
1278 * their files, then the uverbs device ref count will be zero
1279 * and its resources will be freed.
1280 * Note: At this point no more files can be opened since the
1281 * cdev was deleted, however active clients can still issue
1282 * commands and close their open files.
1284 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1285 ib_uverbs_free_hw_resources(uverbs_dev, device);
1289 if (atomic_dec_and_test(&uverbs_dev->refcount))
1290 ib_uverbs_comp_dev(uverbs_dev);
1292 wait_for_completion(&uverbs_dev->comp);
1293 kobject_put(&uverbs_dev->kobj);
1296 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1300 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1303 static int __init ib_uverbs_init(void)
1307 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1308 "infiniband_verbs");
1310 pr_err("user_verbs: couldn't register device number\n");
1314 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1315 if (IS_ERR(uverbs_class)) {
1316 ret = PTR_ERR(uverbs_class);
1317 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1321 uverbs_class->devnode = uverbs_devnode;
1323 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1325 pr_err("user_verbs: couldn't create abi_version attribute\n");
1329 ret = ib_register_client(&uverbs_client);
1331 pr_err("user_verbs: couldn't register client\n");
1338 class_destroy(uverbs_class);
1341 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1347 static void __exit ib_uverbs_cleanup(void)
1349 ib_unregister_client(&uverbs_client);
1350 class_destroy(uverbs_class);
1351 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1353 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1354 idr_destroy(&ib_uverbs_pd_idr);
1355 idr_destroy(&ib_uverbs_mr_idr);
1356 idr_destroy(&ib_uverbs_mw_idr);
1357 idr_destroy(&ib_uverbs_ah_idr);
1358 idr_destroy(&ib_uverbs_cq_idr);
1359 idr_destroy(&ib_uverbs_qp_idr);
1360 idr_destroy(&ib_uverbs_srq_idr);
1363 module_init(ib_uverbs_init);
1364 module_exit(ib_uverbs_cleanup);