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>
53 MODULE_AUTHOR("Roland Dreier");
54 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
55 MODULE_LICENSE("Dual BSD/GPL");
58 IB_UVERBS_MAJOR = 231,
59 IB_UVERBS_BASE_MINOR = 192,
60 IB_UVERBS_MAX_DEVICES = 32
63 #define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
65 static struct class *uverbs_class;
67 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
68 DEFINE_IDR(ib_uverbs_pd_idr);
69 DEFINE_IDR(ib_uverbs_mr_idr);
70 DEFINE_IDR(ib_uverbs_mw_idr);
71 DEFINE_IDR(ib_uverbs_ah_idr);
72 DEFINE_IDR(ib_uverbs_cq_idr);
73 DEFINE_IDR(ib_uverbs_qp_idr);
74 DEFINE_IDR(ib_uverbs_srq_idr);
75 DEFINE_IDR(ib_uverbs_xrcd_idr);
76 DEFINE_IDR(ib_uverbs_rule_idr);
78 static DEFINE_SPINLOCK(map_lock);
79 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
81 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
82 struct ib_device *ib_dev,
83 const char __user *buf, int in_len,
85 [IB_USER_VERBS_CMD_GET_CONTEXT] = ib_uverbs_get_context,
86 [IB_USER_VERBS_CMD_QUERY_DEVICE] = ib_uverbs_query_device,
87 [IB_USER_VERBS_CMD_QUERY_PORT] = ib_uverbs_query_port,
88 [IB_USER_VERBS_CMD_ALLOC_PD] = ib_uverbs_alloc_pd,
89 [IB_USER_VERBS_CMD_DEALLOC_PD] = ib_uverbs_dealloc_pd,
90 [IB_USER_VERBS_CMD_REG_MR] = ib_uverbs_reg_mr,
91 [IB_USER_VERBS_CMD_REREG_MR] = ib_uverbs_rereg_mr,
92 [IB_USER_VERBS_CMD_DEREG_MR] = ib_uverbs_dereg_mr,
93 [IB_USER_VERBS_CMD_ALLOC_MW] = ib_uverbs_alloc_mw,
94 [IB_USER_VERBS_CMD_DEALLOC_MW] = ib_uverbs_dealloc_mw,
95 [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
96 [IB_USER_VERBS_CMD_CREATE_CQ] = ib_uverbs_create_cq,
97 [IB_USER_VERBS_CMD_RESIZE_CQ] = ib_uverbs_resize_cq,
98 [IB_USER_VERBS_CMD_POLL_CQ] = ib_uverbs_poll_cq,
99 [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ] = ib_uverbs_req_notify_cq,
100 [IB_USER_VERBS_CMD_DESTROY_CQ] = ib_uverbs_destroy_cq,
101 [IB_USER_VERBS_CMD_CREATE_QP] = ib_uverbs_create_qp,
102 [IB_USER_VERBS_CMD_QUERY_QP] = ib_uverbs_query_qp,
103 [IB_USER_VERBS_CMD_MODIFY_QP] = ib_uverbs_modify_qp,
104 [IB_USER_VERBS_CMD_DESTROY_QP] = ib_uverbs_destroy_qp,
105 [IB_USER_VERBS_CMD_POST_SEND] = ib_uverbs_post_send,
106 [IB_USER_VERBS_CMD_POST_RECV] = ib_uverbs_post_recv,
107 [IB_USER_VERBS_CMD_POST_SRQ_RECV] = ib_uverbs_post_srq_recv,
108 [IB_USER_VERBS_CMD_CREATE_AH] = ib_uverbs_create_ah,
109 [IB_USER_VERBS_CMD_DESTROY_AH] = ib_uverbs_destroy_ah,
110 [IB_USER_VERBS_CMD_ATTACH_MCAST] = ib_uverbs_attach_mcast,
111 [IB_USER_VERBS_CMD_DETACH_MCAST] = ib_uverbs_detach_mcast,
112 [IB_USER_VERBS_CMD_CREATE_SRQ] = ib_uverbs_create_srq,
113 [IB_USER_VERBS_CMD_MODIFY_SRQ] = ib_uverbs_modify_srq,
114 [IB_USER_VERBS_CMD_QUERY_SRQ] = ib_uverbs_query_srq,
115 [IB_USER_VERBS_CMD_DESTROY_SRQ] = ib_uverbs_destroy_srq,
116 [IB_USER_VERBS_CMD_OPEN_XRCD] = ib_uverbs_open_xrcd,
117 [IB_USER_VERBS_CMD_CLOSE_XRCD] = ib_uverbs_close_xrcd,
118 [IB_USER_VERBS_CMD_CREATE_XSRQ] = ib_uverbs_create_xsrq,
119 [IB_USER_VERBS_CMD_OPEN_QP] = ib_uverbs_open_qp,
122 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
123 struct ib_device *ib_dev,
124 struct ib_udata *ucore,
125 struct ib_udata *uhw) = {
126 [IB_USER_VERBS_EX_CMD_CREATE_FLOW] = ib_uverbs_ex_create_flow,
127 [IB_USER_VERBS_EX_CMD_DESTROY_FLOW] = ib_uverbs_ex_destroy_flow,
128 [IB_USER_VERBS_EX_CMD_QUERY_DEVICE] = ib_uverbs_ex_query_device,
129 [IB_USER_VERBS_EX_CMD_CREATE_CQ] = ib_uverbs_ex_create_cq,
130 [IB_USER_VERBS_EX_CMD_CREATE_QP] = ib_uverbs_ex_create_qp,
133 static void ib_uverbs_add_one(struct ib_device *device);
134 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
136 int uverbs_dealloc_mw(struct ib_mw *mw)
138 struct ib_pd *pd = mw->pd;
141 ret = mw->device->dealloc_mw(mw);
143 atomic_dec(&pd->usecnt);
147 static void ib_uverbs_release_dev(struct kobject *kobj)
149 struct ib_uverbs_device *dev =
150 container_of(kobj, struct ib_uverbs_device, kobj);
152 cleanup_srcu_struct(&dev->disassociate_srcu);
156 static struct kobj_type ib_uverbs_dev_ktype = {
157 .release = ib_uverbs_release_dev,
160 static void ib_uverbs_release_event_file(struct kref *ref)
162 struct ib_uverbs_event_file *file =
163 container_of(ref, struct ib_uverbs_event_file, ref);
168 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
169 struct ib_uverbs_event_file *ev_file,
170 struct ib_ucq_object *uobj)
172 struct ib_uverbs_event *evt, *tmp;
175 spin_lock_irq(&ev_file->lock);
176 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
177 list_del(&evt->list);
180 spin_unlock_irq(&ev_file->lock);
182 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
185 spin_lock_irq(&file->async_file->lock);
186 list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
187 list_del(&evt->list);
190 spin_unlock_irq(&file->async_file->lock);
193 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
194 struct ib_uevent_object *uobj)
196 struct ib_uverbs_event *evt, *tmp;
198 spin_lock_irq(&file->async_file->lock);
199 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
200 list_del(&evt->list);
203 spin_unlock_irq(&file->async_file->lock);
206 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
207 struct ib_uqp_object *uobj)
209 struct ib_uverbs_mcast_entry *mcast, *tmp;
211 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
212 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
213 list_del(&mcast->list);
218 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
219 struct ib_ucontext *context)
221 struct ib_uobject *uobj, *tmp;
223 context->closing = 1;
225 list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
226 struct ib_ah *ah = uobj->object;
228 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
233 /* Remove MWs before QPs, in order to support type 2A MWs. */
234 list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
235 struct ib_mw *mw = uobj->object;
237 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
238 uverbs_dealloc_mw(mw);
242 list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
243 struct ib_flow *flow_id = uobj->object;
245 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
246 ib_destroy_flow(flow_id);
250 list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
251 struct ib_qp *qp = uobj->object;
252 struct ib_uqp_object *uqp =
253 container_of(uobj, struct ib_uqp_object, uevent.uobject);
255 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
256 if (qp != qp->real_qp) {
259 ib_uverbs_detach_umcast(qp, uqp);
262 ib_uverbs_release_uevent(file, &uqp->uevent);
266 list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
267 struct ib_srq *srq = uobj->object;
268 struct ib_uevent_object *uevent =
269 container_of(uobj, struct ib_uevent_object, uobject);
271 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
273 ib_uverbs_release_uevent(file, uevent);
277 list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
278 struct ib_cq *cq = uobj->object;
279 struct ib_uverbs_event_file *ev_file = cq->cq_context;
280 struct ib_ucq_object *ucq =
281 container_of(uobj, struct ib_ucq_object, uobject);
283 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
285 ib_uverbs_release_ucq(file, ev_file, ucq);
289 list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
290 struct ib_mr *mr = uobj->object;
292 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
297 mutex_lock(&file->device->xrcd_tree_mutex);
298 list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
299 struct ib_xrcd *xrcd = uobj->object;
300 struct ib_uxrcd_object *uxrcd =
301 container_of(uobj, struct ib_uxrcd_object, uobject);
303 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
304 ib_uverbs_dealloc_xrcd(file->device, xrcd);
307 mutex_unlock(&file->device->xrcd_tree_mutex);
309 list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
310 struct ib_pd *pd = uobj->object;
312 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
317 put_pid(context->tgid);
319 return context->device->dealloc_ucontext(context);
322 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
324 complete(&dev->comp);
327 static void ib_uverbs_release_file(struct kref *ref)
329 struct ib_uverbs_file *file =
330 container_of(ref, struct ib_uverbs_file, ref);
331 struct ib_device *ib_dev;
334 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
335 ib_dev = srcu_dereference(file->device->ib_dev,
336 &file->device->disassociate_srcu);
337 if (ib_dev && !ib_dev->disassociate_ucontext)
338 module_put(ib_dev->owner);
339 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
341 if (atomic_dec_and_test(&file->device->refcount))
342 ib_uverbs_comp_dev(file->device);
347 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
348 size_t count, loff_t *pos)
350 struct ib_uverbs_event_file *file = filp->private_data;
351 struct ib_uverbs_event *event;
355 spin_lock_irq(&file->lock);
357 while (list_empty(&file->event_list)) {
358 spin_unlock_irq(&file->lock);
360 if (filp->f_flags & O_NONBLOCK)
363 if (wait_event_interruptible(file->poll_wait,
364 (!list_empty(&file->event_list) ||
365 /* The barriers built into wait_event_interruptible()
366 * and wake_up() guarentee this will see the null set
369 !file->uverbs_file->device->ib_dev)))
372 /* If device was disassociated and no event exists set an error */
373 if (list_empty(&file->event_list) &&
374 !file->uverbs_file->device->ib_dev)
377 spin_lock_irq(&file->lock);
380 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
383 eventsz = sizeof (struct ib_uverbs_async_event_desc);
385 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
387 if (eventsz > count) {
391 list_del(file->event_list.next);
392 if (event->counter) {
394 list_del(&event->obj_list);
398 spin_unlock_irq(&file->lock);
401 if (copy_to_user(buf, event, eventsz))
412 static unsigned int ib_uverbs_event_poll(struct file *filp,
413 struct poll_table_struct *wait)
415 unsigned int pollflags = 0;
416 struct ib_uverbs_event_file *file = filp->private_data;
418 poll_wait(filp, &file->poll_wait, wait);
420 spin_lock_irq(&file->lock);
421 if (!list_empty(&file->event_list))
422 pollflags = POLLIN | POLLRDNORM;
423 spin_unlock_irq(&file->lock);
428 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
430 struct ib_uverbs_event_file *file = filp->private_data;
432 return fasync_helper(fd, filp, on, &file->async_queue);
435 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
437 struct ib_uverbs_event_file *file = filp->private_data;
438 struct ib_uverbs_event *entry, *tmp;
439 int closed_already = 0;
441 mutex_lock(&file->uverbs_file->device->lists_mutex);
442 spin_lock_irq(&file->lock);
443 closed_already = file->is_closed;
445 list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
447 list_del(&entry->obj_list);
450 spin_unlock_irq(&file->lock);
451 if (!closed_already) {
452 list_del(&file->list);
454 ib_unregister_event_handler(&file->uverbs_file->
457 mutex_unlock(&file->uverbs_file->device->lists_mutex);
459 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
460 kref_put(&file->ref, ib_uverbs_release_event_file);
465 static const struct file_operations uverbs_event_fops = {
466 .owner = THIS_MODULE,
467 .read = ib_uverbs_event_read,
468 .poll = ib_uverbs_event_poll,
469 .release = ib_uverbs_event_close,
470 .fasync = ib_uverbs_event_fasync,
474 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
476 struct ib_uverbs_event_file *file = cq_context;
477 struct ib_ucq_object *uobj;
478 struct ib_uverbs_event *entry;
484 spin_lock_irqsave(&file->lock, flags);
485 if (file->is_closed) {
486 spin_unlock_irqrestore(&file->lock, flags);
490 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
492 spin_unlock_irqrestore(&file->lock, flags);
496 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
498 entry->desc.comp.cq_handle = cq->uobject->user_handle;
499 entry->counter = &uobj->comp_events_reported;
501 list_add_tail(&entry->list, &file->event_list);
502 list_add_tail(&entry->obj_list, &uobj->comp_list);
503 spin_unlock_irqrestore(&file->lock, flags);
505 wake_up_interruptible(&file->poll_wait);
506 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
509 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
510 __u64 element, __u64 event,
511 struct list_head *obj_list,
514 struct ib_uverbs_event *entry;
517 spin_lock_irqsave(&file->async_file->lock, flags);
518 if (file->async_file->is_closed) {
519 spin_unlock_irqrestore(&file->async_file->lock, flags);
523 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
525 spin_unlock_irqrestore(&file->async_file->lock, flags);
529 entry->desc.async.element = element;
530 entry->desc.async.event_type = event;
531 entry->desc.async.reserved = 0;
532 entry->counter = counter;
534 list_add_tail(&entry->list, &file->async_file->event_list);
536 list_add_tail(&entry->obj_list, obj_list);
537 spin_unlock_irqrestore(&file->async_file->lock, flags);
539 wake_up_interruptible(&file->async_file->poll_wait);
540 kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
543 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
545 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
546 struct ib_ucq_object, uobject);
548 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
549 event->event, &uobj->async_list,
550 &uobj->async_events_reported);
553 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
555 struct ib_uevent_object *uobj;
557 /* for XRC target qp's, check that qp is live */
558 if (!event->element.qp->uobject || !event->element.qp->uobject->live)
561 uobj = container_of(event->element.qp->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_srq_event_handler(struct ib_event *event, void *context_ptr)
571 struct ib_uevent_object *uobj;
573 uobj = container_of(event->element.srq->uobject,
574 struct ib_uevent_object, uobject);
576 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
577 event->event, &uobj->event_list,
578 &uobj->events_reported);
581 void ib_uverbs_event_handler(struct ib_event_handler *handler,
582 struct ib_event *event)
584 struct ib_uverbs_file *file =
585 container_of(handler, struct ib_uverbs_file, event_handler);
587 ib_uverbs_async_handler(file, event->element.port_num, event->event,
591 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
593 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
594 file->async_file = NULL;
597 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
598 struct ib_device *ib_dev,
601 struct ib_uverbs_event_file *ev_file;
605 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
607 return ERR_PTR(-ENOMEM);
609 kref_init(&ev_file->ref);
610 spin_lock_init(&ev_file->lock);
611 INIT_LIST_HEAD(&ev_file->event_list);
612 init_waitqueue_head(&ev_file->poll_wait);
613 ev_file->uverbs_file = uverbs_file;
614 kref_get(&ev_file->uverbs_file->ref);
615 ev_file->async_queue = NULL;
616 ev_file->is_closed = 0;
618 filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
623 mutex_lock(&uverbs_file->device->lists_mutex);
624 list_add_tail(&ev_file->list,
625 &uverbs_file->device->uverbs_events_file_list);
626 mutex_unlock(&uverbs_file->device->lists_mutex);
629 WARN_ON(uverbs_file->async_file);
630 uverbs_file->async_file = ev_file;
631 kref_get(&uverbs_file->async_file->ref);
632 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
634 ib_uverbs_event_handler);
635 ret = ib_register_event_handler(&uverbs_file->event_handler);
639 /* At that point async file stuff was fully set */
640 ev_file->is_async = 1;
647 kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
648 uverbs_file->async_file = NULL;
652 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
653 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
658 * Look up a completion event file by FD. If lookup is successful,
659 * takes a ref to the event file struct that it returns; if
660 * unsuccessful, returns NULL.
662 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
664 struct ib_uverbs_event_file *ev_file = NULL;
665 struct fd f = fdget(fd);
670 if (f.file->f_op != &uverbs_event_fops)
673 ev_file = f.file->private_data;
674 if (ev_file->is_async) {
679 kref_get(&ev_file->ref);
686 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
687 size_t count, loff_t *pos)
689 struct ib_uverbs_file *file = filp->private_data;
690 struct ib_device *ib_dev;
691 struct ib_uverbs_cmd_hdr hdr;
696 if (count < sizeof hdr)
699 if (copy_from_user(&hdr, buf, sizeof hdr))
702 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
703 ib_dev = srcu_dereference(file->device->ib_dev,
704 &file->device->disassociate_srcu);
710 flags = (hdr.command &
711 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
716 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
717 IB_USER_VERBS_CMD_COMMAND_MASK)) {
722 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
724 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
725 !uverbs_cmd_table[command]) {
730 if (!file->ucontext &&
731 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
736 if (!(ib_dev->uverbs_cmd_mask & (1ull << command))) {
741 if (hdr.in_words * 4 != count) {
746 ret = uverbs_cmd_table[command](file, ib_dev,
751 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
754 struct ib_uverbs_ex_cmd_hdr ex_hdr;
755 struct ib_udata ucore;
757 size_t written_count = count;
759 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
760 IB_USER_VERBS_CMD_COMMAND_MASK)) {
765 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
767 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
768 !uverbs_ex_cmd_table[command]) {
773 if (!file->ucontext) {
778 if (!(ib_dev->uverbs_ex_cmd_mask & (1ull << command))) {
783 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
788 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
793 count -= sizeof(hdr) + sizeof(ex_hdr);
794 buf += sizeof(hdr) + sizeof(ex_hdr);
796 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
801 if (ex_hdr.cmd_hdr_reserved) {
806 if (ex_hdr.response) {
807 if (!hdr.out_words && !ex_hdr.provider_out_words) {
812 if (!access_ok(VERIFY_WRITE,
813 (void __user *) (unsigned long) ex_hdr.response,
814 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
819 if (hdr.out_words || ex_hdr.provider_out_words) {
825 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
826 hdr.in_words * 8, hdr.out_words * 8);
828 INIT_UDATA_BUF_OR_NULL(&uhw,
830 (unsigned long) ex_hdr.response + ucore.outlen,
831 ex_hdr.provider_in_words * 8,
832 ex_hdr.provider_out_words * 8);
834 ret = uverbs_ex_cmd_table[command](file,
845 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
849 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
851 struct ib_uverbs_file *file = filp->private_data;
852 struct ib_device *ib_dev;
856 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
857 ib_dev = srcu_dereference(file->device->ib_dev,
858 &file->device->disassociate_srcu);
867 ret = ib_dev->mmap(file->ucontext, vma);
869 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
874 * ib_uverbs_open() does not need the BKL:
876 * - the ib_uverbs_device structures are properly reference counted and
877 * everything else is purely local to the file being created, so
878 * races against other open calls are not a problem;
879 * - there is no ioctl method to race against;
880 * - the open method will either immediately run -ENXIO, or all
881 * required initialization will be done.
883 static int ib_uverbs_open(struct inode *inode, struct file *filp)
885 struct ib_uverbs_device *dev;
886 struct ib_uverbs_file *file;
887 struct ib_device *ib_dev;
889 int module_dependent;
892 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
893 if (!atomic_inc_not_zero(&dev->refcount))
896 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
897 mutex_lock(&dev->lists_mutex);
898 ib_dev = srcu_dereference(dev->ib_dev,
899 &dev->disassociate_srcu);
905 /* In case IB device supports disassociate ucontext, there is no hard
906 * dependency between uverbs device and its low level device.
908 module_dependent = !(ib_dev->disassociate_ucontext);
910 if (module_dependent) {
911 if (!try_module_get(ib_dev->owner)) {
917 file = kzalloc(sizeof(*file), GFP_KERNEL);
920 if (module_dependent)
927 file->ucontext = NULL;
928 file->async_file = NULL;
929 kref_init(&file->ref);
930 mutex_init(&file->mutex);
932 filp->private_data = file;
933 kobject_get(&dev->kobj);
934 list_add_tail(&file->list, &dev->uverbs_file_list);
935 mutex_unlock(&dev->lists_mutex);
936 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
938 return nonseekable_open(inode, filp);
941 module_put(ib_dev->owner);
944 mutex_unlock(&dev->lists_mutex);
945 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
946 if (atomic_dec_and_test(&dev->refcount))
947 ib_uverbs_comp_dev(dev);
952 static int ib_uverbs_close(struct inode *inode, struct file *filp)
954 struct ib_uverbs_file *file = filp->private_data;
955 struct ib_uverbs_device *dev = file->device;
956 struct ib_ucontext *ucontext = NULL;
958 mutex_lock(&file->device->lists_mutex);
959 ucontext = file->ucontext;
960 file->ucontext = NULL;
961 if (!file->is_closed) {
962 list_del(&file->list);
965 mutex_unlock(&file->device->lists_mutex);
967 ib_uverbs_cleanup_ucontext(file, ucontext);
969 if (file->async_file)
970 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
972 kref_put(&file->ref, ib_uverbs_release_file);
973 kobject_put(&dev->kobj);
978 static const struct file_operations uverbs_fops = {
979 .owner = THIS_MODULE,
980 .write = ib_uverbs_write,
981 .open = ib_uverbs_open,
982 .release = ib_uverbs_close,
986 static const struct file_operations uverbs_mmap_fops = {
987 .owner = THIS_MODULE,
988 .write = ib_uverbs_write,
989 .mmap = ib_uverbs_mmap,
990 .open = ib_uverbs_open,
991 .release = ib_uverbs_close,
995 static struct ib_client uverbs_client = {
997 .add = ib_uverbs_add_one,
998 .remove = ib_uverbs_remove_one
1001 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
1006 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1007 struct ib_device *ib_dev;
1012 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1013 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1015 ret = sprintf(buf, "%s\n", ib_dev->name);
1016 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1020 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1022 static ssize_t show_dev_abi_version(struct device *device,
1023 struct device_attribute *attr, char *buf)
1025 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1028 struct ib_device *ib_dev;
1032 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1033 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1035 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1036 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1040 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1042 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1043 __stringify(IB_USER_VERBS_ABI_VERSION));
1045 static dev_t overflow_maj;
1046 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1049 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1050 * requesting a new major number and doubling the number of max devices we
1051 * support. It's stupid, but simple.
1053 static int find_overflow_devnum(void)
1057 if (!overflow_maj) {
1058 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1059 "infiniband_verbs");
1061 printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
1066 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1067 if (ret >= IB_UVERBS_MAX_DEVICES)
1073 static void ib_uverbs_add_one(struct ib_device *device)
1077 struct ib_uverbs_device *uverbs_dev;
1080 if (!device->alloc_ucontext)
1083 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1087 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1093 atomic_set(&uverbs_dev->refcount, 1);
1094 init_completion(&uverbs_dev->comp);
1095 uverbs_dev->xrcd_tree = RB_ROOT;
1096 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1097 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1098 mutex_init(&uverbs_dev->lists_mutex);
1099 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1100 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1102 spin_lock(&map_lock);
1103 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1104 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1105 spin_unlock(&map_lock);
1106 devnum = find_overflow_devnum();
1110 spin_lock(&map_lock);
1111 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1112 base = devnum + overflow_maj;
1113 set_bit(devnum, overflow_map);
1115 uverbs_dev->devnum = devnum;
1116 base = devnum + IB_UVERBS_BASE_DEV;
1117 set_bit(devnum, dev_map);
1119 spin_unlock(&map_lock);
1121 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1122 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1124 cdev_init(&uverbs_dev->cdev, NULL);
1125 uverbs_dev->cdev.owner = THIS_MODULE;
1126 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1127 uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1128 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1129 if (cdev_add(&uverbs_dev->cdev, base, 1))
1132 uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
1133 uverbs_dev->cdev.dev, uverbs_dev,
1134 "uverbs%d", uverbs_dev->devnum);
1135 if (IS_ERR(uverbs_dev->dev))
1138 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1140 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1143 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1148 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1151 cdev_del(&uverbs_dev->cdev);
1152 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1153 clear_bit(devnum, dev_map);
1155 clear_bit(devnum, overflow_map);
1158 if (atomic_dec_and_test(&uverbs_dev->refcount))
1159 ib_uverbs_comp_dev(uverbs_dev);
1160 wait_for_completion(&uverbs_dev->comp);
1161 kobject_put(&uverbs_dev->kobj);
1165 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1166 struct ib_device *ib_dev)
1168 struct ib_uverbs_file *file;
1169 struct ib_uverbs_event_file *event_file;
1170 struct ib_event event;
1172 /* Pending running commands to terminate */
1173 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1174 event.event = IB_EVENT_DEVICE_FATAL;
1175 event.element.port_num = 0;
1176 event.device = ib_dev;
1178 mutex_lock(&uverbs_dev->lists_mutex);
1179 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1180 struct ib_ucontext *ucontext;
1182 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1183 struct ib_uverbs_file, list);
1184 file->is_closed = 1;
1185 ucontext = file->ucontext;
1186 list_del(&file->list);
1187 file->ucontext = NULL;
1188 kref_get(&file->ref);
1189 mutex_unlock(&uverbs_dev->lists_mutex);
1190 /* We must release the mutex before going ahead and calling
1191 * disassociate_ucontext. disassociate_ucontext might end up
1192 * indirectly calling uverbs_close, for example due to freeing
1193 * the resources (e.g mmput).
1195 ib_uverbs_event_handler(&file->event_handler, &event);
1197 ib_dev->disassociate_ucontext(ucontext);
1198 ib_uverbs_cleanup_ucontext(file, ucontext);
1201 mutex_lock(&uverbs_dev->lists_mutex);
1202 kref_put(&file->ref, ib_uverbs_release_file);
1205 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1206 event_file = list_first_entry(&uverbs_dev->
1207 uverbs_events_file_list,
1208 struct ib_uverbs_event_file,
1210 spin_lock_irq(&event_file->lock);
1211 event_file->is_closed = 1;
1212 spin_unlock_irq(&event_file->lock);
1214 list_del(&event_file->list);
1215 if (event_file->is_async) {
1216 ib_unregister_event_handler(&event_file->uverbs_file->
1218 event_file->uverbs_file->event_handler.device = NULL;
1221 wake_up_interruptible(&event_file->poll_wait);
1222 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1224 mutex_unlock(&uverbs_dev->lists_mutex);
1227 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1229 struct ib_uverbs_device *uverbs_dev = client_data;
1230 int wait_clients = 1;
1235 dev_set_drvdata(uverbs_dev->dev, NULL);
1236 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1237 cdev_del(&uverbs_dev->cdev);
1239 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1240 clear_bit(uverbs_dev->devnum, dev_map);
1242 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1244 if (device->disassociate_ucontext) {
1245 /* We disassociate HW resources and immediately return.
1246 * Userspace will see a EIO errno for all future access.
1247 * Upon returning, ib_device may be freed internally and is not
1249 * uverbs_device is still available until all clients close
1250 * their files, then the uverbs device ref count will be zero
1251 * and its resources will be freed.
1252 * Note: At this point no more files can be opened since the
1253 * cdev was deleted, however active clients can still issue
1254 * commands and close their open files.
1256 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1257 ib_uverbs_free_hw_resources(uverbs_dev, device);
1261 if (atomic_dec_and_test(&uverbs_dev->refcount))
1262 ib_uverbs_comp_dev(uverbs_dev);
1264 wait_for_completion(&uverbs_dev->comp);
1265 kobject_put(&uverbs_dev->kobj);
1268 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1272 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1275 static int __init ib_uverbs_init(void)
1279 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1280 "infiniband_verbs");
1282 printk(KERN_ERR "user_verbs: couldn't register device number\n");
1286 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1287 if (IS_ERR(uverbs_class)) {
1288 ret = PTR_ERR(uverbs_class);
1289 printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
1293 uverbs_class->devnode = uverbs_devnode;
1295 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1297 printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
1301 ret = ib_register_client(&uverbs_client);
1303 printk(KERN_ERR "user_verbs: couldn't register client\n");
1310 class_destroy(uverbs_class);
1313 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1319 static void __exit ib_uverbs_cleanup(void)
1321 ib_unregister_client(&uverbs_client);
1322 class_destroy(uverbs_class);
1323 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1325 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1326 idr_destroy(&ib_uverbs_pd_idr);
1327 idr_destroy(&ib_uverbs_mr_idr);
1328 idr_destroy(&ib_uverbs_mw_idr);
1329 idr_destroy(&ib_uverbs_ah_idr);
1330 idr_destroy(&ib_uverbs_cq_idr);
1331 idr_destroy(&ib_uverbs_qp_idr);
1332 idr_destroy(&ib_uverbs_srq_idr);
1335 module_init(ib_uverbs_init);
1336 module_exit(ib_uverbs_cleanup);