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)
248 ib_uverbs_detach_umcast(qp, uqp);
250 ib_uverbs_release_uevent(file, &uqp->uevent);
254 list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
255 struct ib_srq *srq = uobj->object;
256 struct ib_uevent_object *uevent =
257 container_of(uobj, struct ib_uevent_object, uobject);
259 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
261 ib_uverbs_release_uevent(file, uevent);
265 list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
266 struct ib_cq *cq = uobj->object;
267 struct ib_uverbs_event_file *ev_file = cq->cq_context;
268 struct ib_ucq_object *ucq =
269 container_of(uobj, struct ib_ucq_object, uobject);
271 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
273 ib_uverbs_release_ucq(file, ev_file, ucq);
277 list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
278 struct ib_mr *mr = uobj->object;
280 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
285 mutex_lock(&file->device->xrcd_tree_mutex);
286 list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
287 struct ib_xrcd *xrcd = uobj->object;
288 struct ib_uxrcd_object *uxrcd =
289 container_of(uobj, struct ib_uxrcd_object, uobject);
291 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
292 ib_uverbs_dealloc_xrcd(file->device, xrcd);
295 mutex_unlock(&file->device->xrcd_tree_mutex);
297 list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
298 struct ib_pd *pd = uobj->object;
300 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
305 put_pid(context->tgid);
307 return context->device->dealloc_ucontext(context);
310 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
312 complete(&dev->comp);
315 static void ib_uverbs_release_file(struct kref *ref)
317 struct ib_uverbs_file *file =
318 container_of(ref, struct ib_uverbs_file, ref);
319 struct ib_device *ib_dev;
322 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
323 ib_dev = srcu_dereference(file->device->ib_dev,
324 &file->device->disassociate_srcu);
325 if (ib_dev && !ib_dev->disassociate_ucontext)
326 module_put(ib_dev->owner);
327 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
329 if (atomic_dec_and_test(&file->device->refcount))
330 ib_uverbs_comp_dev(file->device);
335 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
336 size_t count, loff_t *pos)
338 struct ib_uverbs_event_file *file = filp->private_data;
339 struct ib_uverbs_event *event;
343 spin_lock_irq(&file->lock);
345 while (list_empty(&file->event_list)) {
346 spin_unlock_irq(&file->lock);
348 if (filp->f_flags & O_NONBLOCK)
351 if (wait_event_interruptible(file->poll_wait,
352 (!list_empty(&file->event_list) ||
353 /* The barriers built into wait_event_interruptible()
354 * and wake_up() guarentee this will see the null set
357 !file->uverbs_file->device->ib_dev)))
360 /* If device was disassociated and no event exists set an error */
361 if (list_empty(&file->event_list) &&
362 !file->uverbs_file->device->ib_dev)
365 spin_lock_irq(&file->lock);
368 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
371 eventsz = sizeof (struct ib_uverbs_async_event_desc);
373 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
375 if (eventsz > count) {
379 list_del(file->event_list.next);
380 if (event->counter) {
382 list_del(&event->obj_list);
386 spin_unlock_irq(&file->lock);
389 if (copy_to_user(buf, event, eventsz))
400 static unsigned int ib_uverbs_event_poll(struct file *filp,
401 struct poll_table_struct *wait)
403 unsigned int pollflags = 0;
404 struct ib_uverbs_event_file *file = filp->private_data;
406 poll_wait(filp, &file->poll_wait, wait);
408 spin_lock_irq(&file->lock);
409 if (!list_empty(&file->event_list))
410 pollflags = POLLIN | POLLRDNORM;
411 spin_unlock_irq(&file->lock);
416 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
418 struct ib_uverbs_event_file *file = filp->private_data;
420 return fasync_helper(fd, filp, on, &file->async_queue);
423 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
425 struct ib_uverbs_event_file *file = filp->private_data;
426 struct ib_uverbs_event *entry, *tmp;
427 int closed_already = 0;
429 mutex_lock(&file->uverbs_file->device->lists_mutex);
430 spin_lock_irq(&file->lock);
431 closed_already = file->is_closed;
433 list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
435 list_del(&entry->obj_list);
438 spin_unlock_irq(&file->lock);
439 if (!closed_already) {
440 list_del(&file->list);
442 ib_unregister_event_handler(&file->uverbs_file->
445 mutex_unlock(&file->uverbs_file->device->lists_mutex);
447 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
448 kref_put(&file->ref, ib_uverbs_release_event_file);
453 static const struct file_operations uverbs_event_fops = {
454 .owner = THIS_MODULE,
455 .read = ib_uverbs_event_read,
456 .poll = ib_uverbs_event_poll,
457 .release = ib_uverbs_event_close,
458 .fasync = ib_uverbs_event_fasync,
462 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
464 struct ib_uverbs_event_file *file = cq_context;
465 struct ib_ucq_object *uobj;
466 struct ib_uverbs_event *entry;
472 spin_lock_irqsave(&file->lock, flags);
473 if (file->is_closed) {
474 spin_unlock_irqrestore(&file->lock, flags);
478 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
480 spin_unlock_irqrestore(&file->lock, flags);
484 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
486 entry->desc.comp.cq_handle = cq->uobject->user_handle;
487 entry->counter = &uobj->comp_events_reported;
489 list_add_tail(&entry->list, &file->event_list);
490 list_add_tail(&entry->obj_list, &uobj->comp_list);
491 spin_unlock_irqrestore(&file->lock, flags);
493 wake_up_interruptible(&file->poll_wait);
494 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
497 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
498 __u64 element, __u64 event,
499 struct list_head *obj_list,
502 struct ib_uverbs_event *entry;
505 spin_lock_irqsave(&file->async_file->lock, flags);
506 if (file->async_file->is_closed) {
507 spin_unlock_irqrestore(&file->async_file->lock, flags);
511 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
513 spin_unlock_irqrestore(&file->async_file->lock, flags);
517 entry->desc.async.element = element;
518 entry->desc.async.event_type = event;
519 entry->desc.async.reserved = 0;
520 entry->counter = counter;
522 list_add_tail(&entry->list, &file->async_file->event_list);
524 list_add_tail(&entry->obj_list, obj_list);
525 spin_unlock_irqrestore(&file->async_file->lock, flags);
527 wake_up_interruptible(&file->async_file->poll_wait);
528 kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
531 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
533 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
534 struct ib_ucq_object, uobject);
536 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
537 event->event, &uobj->async_list,
538 &uobj->async_events_reported);
541 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
543 struct ib_uevent_object *uobj;
545 /* for XRC target qp's, check that qp is live */
546 if (!event->element.qp->uobject || !event->element.qp->uobject->live)
549 uobj = container_of(event->element.qp->uobject,
550 struct ib_uevent_object, uobject);
552 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
553 event->event, &uobj->event_list,
554 &uobj->events_reported);
557 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
559 struct ib_uevent_object *uobj;
561 uobj = container_of(event->element.srq->uobject,
562 struct ib_uevent_object, uobject);
564 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
565 event->event, &uobj->event_list,
566 &uobj->events_reported);
569 void ib_uverbs_event_handler(struct ib_event_handler *handler,
570 struct ib_event *event)
572 struct ib_uverbs_file *file =
573 container_of(handler, struct ib_uverbs_file, event_handler);
575 ib_uverbs_async_handler(file, event->element.port_num, event->event,
579 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
581 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
582 file->async_file = NULL;
585 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
586 struct ib_device *ib_dev,
589 struct ib_uverbs_event_file *ev_file;
593 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
595 return ERR_PTR(-ENOMEM);
597 kref_init(&ev_file->ref);
598 spin_lock_init(&ev_file->lock);
599 INIT_LIST_HEAD(&ev_file->event_list);
600 init_waitqueue_head(&ev_file->poll_wait);
601 ev_file->uverbs_file = uverbs_file;
602 kref_get(&ev_file->uverbs_file->ref);
603 ev_file->async_queue = NULL;
604 ev_file->is_closed = 0;
606 filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
611 mutex_lock(&uverbs_file->device->lists_mutex);
612 list_add_tail(&ev_file->list,
613 &uverbs_file->device->uverbs_events_file_list);
614 mutex_unlock(&uverbs_file->device->lists_mutex);
617 WARN_ON(uverbs_file->async_file);
618 uverbs_file->async_file = ev_file;
619 kref_get(&uverbs_file->async_file->ref);
620 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
622 ib_uverbs_event_handler);
623 ret = ib_register_event_handler(&uverbs_file->event_handler);
627 /* At that point async file stuff was fully set */
628 ev_file->is_async = 1;
635 kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
636 uverbs_file->async_file = NULL;
640 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
641 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
646 * Look up a completion event file by FD. If lookup is successful,
647 * takes a ref to the event file struct that it returns; if
648 * unsuccessful, returns NULL.
650 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
652 struct ib_uverbs_event_file *ev_file = NULL;
653 struct fd f = fdget(fd);
658 if (f.file->f_op != &uverbs_event_fops)
661 ev_file = f.file->private_data;
662 if (ev_file->is_async) {
667 kref_get(&ev_file->ref);
674 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
675 size_t count, loff_t *pos)
677 struct ib_uverbs_file *file = filp->private_data;
678 struct ib_device *ib_dev;
679 struct ib_uverbs_cmd_hdr hdr;
684 if (WARN_ON_ONCE(!ib_safe_file_access(filp)))
687 if (count < sizeof hdr)
690 if (copy_from_user(&hdr, buf, sizeof hdr))
693 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
694 ib_dev = srcu_dereference(file->device->ib_dev,
695 &file->device->disassociate_srcu);
701 flags = (hdr.command &
702 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
707 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
708 IB_USER_VERBS_CMD_COMMAND_MASK)) {
713 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
715 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
716 !uverbs_cmd_table[command]) {
721 if (!file->ucontext &&
722 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
727 if (!(ib_dev->uverbs_cmd_mask & (1ull << command))) {
732 if (hdr.in_words * 4 != count) {
737 ret = uverbs_cmd_table[command](file, ib_dev,
742 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
745 struct ib_uverbs_ex_cmd_hdr ex_hdr;
746 struct ib_udata ucore;
748 size_t written_count = count;
750 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
751 IB_USER_VERBS_CMD_COMMAND_MASK)) {
756 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
758 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
759 !uverbs_ex_cmd_table[command]) {
764 if (!file->ucontext) {
769 if (!(ib_dev->uverbs_ex_cmd_mask & (1ull << command))) {
774 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
779 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
784 count -= sizeof(hdr) + sizeof(ex_hdr);
785 buf += sizeof(hdr) + sizeof(ex_hdr);
787 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
792 if (ex_hdr.cmd_hdr_reserved) {
797 if (ex_hdr.response) {
798 if (!hdr.out_words && !ex_hdr.provider_out_words) {
803 if (!access_ok(VERIFY_WRITE,
804 (void __user *) (unsigned long) ex_hdr.response,
805 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
810 if (hdr.out_words || ex_hdr.provider_out_words) {
816 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
817 hdr.in_words * 8, hdr.out_words * 8);
819 INIT_UDATA_BUF_OR_NULL(&uhw,
821 (unsigned long) ex_hdr.response + ucore.outlen,
822 ex_hdr.provider_in_words * 8,
823 ex_hdr.provider_out_words * 8);
825 ret = uverbs_ex_cmd_table[command](file,
836 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
840 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
842 struct ib_uverbs_file *file = filp->private_data;
843 struct ib_device *ib_dev;
847 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
848 ib_dev = srcu_dereference(file->device->ib_dev,
849 &file->device->disassociate_srcu);
858 ret = ib_dev->mmap(file->ucontext, vma);
860 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
865 * ib_uverbs_open() does not need the BKL:
867 * - the ib_uverbs_device structures are properly reference counted and
868 * everything else is purely local to the file being created, so
869 * races against other open calls are not a problem;
870 * - there is no ioctl method to race against;
871 * - the open method will either immediately run -ENXIO, or all
872 * required initialization will be done.
874 static int ib_uverbs_open(struct inode *inode, struct file *filp)
876 struct ib_uverbs_device *dev;
877 struct ib_uverbs_file *file;
878 struct ib_device *ib_dev;
880 int module_dependent;
883 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
884 if (!atomic_inc_not_zero(&dev->refcount))
887 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
888 mutex_lock(&dev->lists_mutex);
889 ib_dev = srcu_dereference(dev->ib_dev,
890 &dev->disassociate_srcu);
896 /* In case IB device supports disassociate ucontext, there is no hard
897 * dependency between uverbs device and its low level device.
899 module_dependent = !(ib_dev->disassociate_ucontext);
901 if (module_dependent) {
902 if (!try_module_get(ib_dev->owner)) {
908 file = kzalloc(sizeof(*file), GFP_KERNEL);
911 if (module_dependent)
918 file->ucontext = NULL;
919 file->async_file = NULL;
920 kref_init(&file->ref);
921 mutex_init(&file->mutex);
922 mutex_init(&file->cleanup_mutex);
924 filp->private_data = file;
925 kobject_get(&dev->kobj);
926 list_add_tail(&file->list, &dev->uverbs_file_list);
927 mutex_unlock(&dev->lists_mutex);
928 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
930 return nonseekable_open(inode, filp);
933 module_put(ib_dev->owner);
936 mutex_unlock(&dev->lists_mutex);
937 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
938 if (atomic_dec_and_test(&dev->refcount))
939 ib_uverbs_comp_dev(dev);
944 static int ib_uverbs_close(struct inode *inode, struct file *filp)
946 struct ib_uverbs_file *file = filp->private_data;
947 struct ib_uverbs_device *dev = file->device;
949 mutex_lock(&file->cleanup_mutex);
950 if (file->ucontext) {
951 ib_uverbs_cleanup_ucontext(file, file->ucontext);
952 file->ucontext = NULL;
954 mutex_unlock(&file->cleanup_mutex);
956 mutex_lock(&file->device->lists_mutex);
957 if (!file->is_closed) {
958 list_del(&file->list);
961 mutex_unlock(&file->device->lists_mutex);
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;
1175 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1176 struct ib_uverbs_file, list);
1177 file->is_closed = 1;
1178 list_del(&file->list);
1179 kref_get(&file->ref);
1180 mutex_unlock(&uverbs_dev->lists_mutex);
1182 ib_uverbs_event_handler(&file->event_handler, &event);
1184 mutex_lock(&file->cleanup_mutex);
1185 ucontext = file->ucontext;
1186 file->ucontext = NULL;
1187 mutex_unlock(&file->cleanup_mutex);
1189 /* At this point ib_uverbs_close cannot be running
1190 * ib_uverbs_cleanup_ucontext
1193 /* We must release the mutex before going ahead and
1194 * calling disassociate_ucontext. disassociate_ucontext
1195 * might end up indirectly calling uverbs_close,
1196 * for example due to freeing the resources
1199 ib_dev->disassociate_ucontext(ucontext);
1200 ib_uverbs_cleanup_ucontext(file, ucontext);
1203 mutex_lock(&uverbs_dev->lists_mutex);
1204 kref_put(&file->ref, ib_uverbs_release_file);
1207 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1208 event_file = list_first_entry(&uverbs_dev->
1209 uverbs_events_file_list,
1210 struct ib_uverbs_event_file,
1212 spin_lock_irq(&event_file->lock);
1213 event_file->is_closed = 1;
1214 spin_unlock_irq(&event_file->lock);
1216 list_del(&event_file->list);
1217 if (event_file->is_async) {
1218 ib_unregister_event_handler(&event_file->uverbs_file->
1220 event_file->uverbs_file->event_handler.device = NULL;
1223 wake_up_interruptible(&event_file->poll_wait);
1224 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1226 mutex_unlock(&uverbs_dev->lists_mutex);
1229 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1231 struct ib_uverbs_device *uverbs_dev = client_data;
1232 int wait_clients = 1;
1237 dev_set_drvdata(uverbs_dev->dev, NULL);
1238 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1239 cdev_del(&uverbs_dev->cdev);
1241 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1242 clear_bit(uverbs_dev->devnum, dev_map);
1244 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1246 if (device->disassociate_ucontext) {
1247 /* We disassociate HW resources and immediately return.
1248 * Userspace will see a EIO errno for all future access.
1249 * Upon returning, ib_device may be freed internally and is not
1251 * uverbs_device is still available until all clients close
1252 * their files, then the uverbs device ref count will be zero
1253 * and its resources will be freed.
1254 * Note: At this point no more files can be opened since the
1255 * cdev was deleted, however active clients can still issue
1256 * commands and close their open files.
1258 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1259 ib_uverbs_free_hw_resources(uverbs_dev, device);
1263 if (atomic_dec_and_test(&uverbs_dev->refcount))
1264 ib_uverbs_comp_dev(uverbs_dev);
1266 wait_for_completion(&uverbs_dev->comp);
1267 kobject_put(&uverbs_dev->kobj);
1270 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1274 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1277 static int __init ib_uverbs_init(void)
1281 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1282 "infiniband_verbs");
1284 printk(KERN_ERR "user_verbs: couldn't register device number\n");
1288 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1289 if (IS_ERR(uverbs_class)) {
1290 ret = PTR_ERR(uverbs_class);
1291 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
1295 uverbs_class->devnode = uverbs_devnode;
1297 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1299 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
1303 ret = ib_register_client(&uverbs_client);
1305 printk(KERN_ERR "user_verbs: couldn't register client\n");
1312 class_destroy(uverbs_class);
1315 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1321 static void __exit ib_uverbs_cleanup(void)
1323 ib_unregister_client(&uverbs_client);
1324 class_destroy(uverbs_class);
1325 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1327 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1328 idr_destroy(&ib_uverbs_pd_idr);
1329 idr_destroy(&ib_uverbs_mr_idr);
1330 idr_destroy(&ib_uverbs_mw_idr);
1331 idr_destroy(&ib_uverbs_ah_idr);
1332 idr_destroy(&ib_uverbs_cq_idr);
1333 idr_destroy(&ib_uverbs_qp_idr);
1334 idr_destroy(&ib_uverbs_srq_idr);
1337 module_init(ib_uverbs_init);
1338 module_exit(ib_uverbs_cleanup);