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);
80 DEFINE_IDR(ib_uverbs_rwq_ind_tbl_idr);
82 static DEFINE_SPINLOCK(map_lock);
83 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
85 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
86 struct ib_device *ib_dev,
87 const char __user *buf, int in_len,
89 [IB_USER_VERBS_CMD_GET_CONTEXT] = ib_uverbs_get_context,
90 [IB_USER_VERBS_CMD_QUERY_DEVICE] = ib_uverbs_query_device,
91 [IB_USER_VERBS_CMD_QUERY_PORT] = ib_uverbs_query_port,
92 [IB_USER_VERBS_CMD_ALLOC_PD] = ib_uverbs_alloc_pd,
93 [IB_USER_VERBS_CMD_DEALLOC_PD] = ib_uverbs_dealloc_pd,
94 [IB_USER_VERBS_CMD_REG_MR] = ib_uverbs_reg_mr,
95 [IB_USER_VERBS_CMD_REREG_MR] = ib_uverbs_rereg_mr,
96 [IB_USER_VERBS_CMD_DEREG_MR] = ib_uverbs_dereg_mr,
97 [IB_USER_VERBS_CMD_ALLOC_MW] = ib_uverbs_alloc_mw,
98 [IB_USER_VERBS_CMD_DEALLOC_MW] = ib_uverbs_dealloc_mw,
99 [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
100 [IB_USER_VERBS_CMD_CREATE_CQ] = ib_uverbs_create_cq,
101 [IB_USER_VERBS_CMD_RESIZE_CQ] = ib_uverbs_resize_cq,
102 [IB_USER_VERBS_CMD_POLL_CQ] = ib_uverbs_poll_cq,
103 [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ] = ib_uverbs_req_notify_cq,
104 [IB_USER_VERBS_CMD_DESTROY_CQ] = ib_uverbs_destroy_cq,
105 [IB_USER_VERBS_CMD_CREATE_QP] = ib_uverbs_create_qp,
106 [IB_USER_VERBS_CMD_QUERY_QP] = ib_uverbs_query_qp,
107 [IB_USER_VERBS_CMD_MODIFY_QP] = ib_uverbs_modify_qp,
108 [IB_USER_VERBS_CMD_DESTROY_QP] = ib_uverbs_destroy_qp,
109 [IB_USER_VERBS_CMD_POST_SEND] = ib_uverbs_post_send,
110 [IB_USER_VERBS_CMD_POST_RECV] = ib_uverbs_post_recv,
111 [IB_USER_VERBS_CMD_POST_SRQ_RECV] = ib_uverbs_post_srq_recv,
112 [IB_USER_VERBS_CMD_CREATE_AH] = ib_uverbs_create_ah,
113 [IB_USER_VERBS_CMD_DESTROY_AH] = ib_uverbs_destroy_ah,
114 [IB_USER_VERBS_CMD_ATTACH_MCAST] = ib_uverbs_attach_mcast,
115 [IB_USER_VERBS_CMD_DETACH_MCAST] = ib_uverbs_detach_mcast,
116 [IB_USER_VERBS_CMD_CREATE_SRQ] = ib_uverbs_create_srq,
117 [IB_USER_VERBS_CMD_MODIFY_SRQ] = ib_uverbs_modify_srq,
118 [IB_USER_VERBS_CMD_QUERY_SRQ] = ib_uverbs_query_srq,
119 [IB_USER_VERBS_CMD_DESTROY_SRQ] = ib_uverbs_destroy_srq,
120 [IB_USER_VERBS_CMD_OPEN_XRCD] = ib_uverbs_open_xrcd,
121 [IB_USER_VERBS_CMD_CLOSE_XRCD] = ib_uverbs_close_xrcd,
122 [IB_USER_VERBS_CMD_CREATE_XSRQ] = ib_uverbs_create_xsrq,
123 [IB_USER_VERBS_CMD_OPEN_QP] = ib_uverbs_open_qp,
126 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
127 struct ib_device *ib_dev,
128 struct ib_udata *ucore,
129 struct ib_udata *uhw) = {
130 [IB_USER_VERBS_EX_CMD_CREATE_FLOW] = ib_uverbs_ex_create_flow,
131 [IB_USER_VERBS_EX_CMD_DESTROY_FLOW] = ib_uverbs_ex_destroy_flow,
132 [IB_USER_VERBS_EX_CMD_QUERY_DEVICE] = ib_uverbs_ex_query_device,
133 [IB_USER_VERBS_EX_CMD_CREATE_CQ] = ib_uverbs_ex_create_cq,
134 [IB_USER_VERBS_EX_CMD_CREATE_QP] = ib_uverbs_ex_create_qp,
135 [IB_USER_VERBS_EX_CMD_CREATE_WQ] = ib_uverbs_ex_create_wq,
136 [IB_USER_VERBS_EX_CMD_MODIFY_WQ] = ib_uverbs_ex_modify_wq,
137 [IB_USER_VERBS_EX_CMD_DESTROY_WQ] = ib_uverbs_ex_destroy_wq,
138 [IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL] = ib_uverbs_ex_create_rwq_ind_table,
139 [IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL] = ib_uverbs_ex_destroy_rwq_ind_table,
142 static void ib_uverbs_add_one(struct ib_device *device);
143 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
145 int uverbs_dealloc_mw(struct ib_mw *mw)
147 struct ib_pd *pd = mw->pd;
150 ret = mw->device->dealloc_mw(mw);
152 atomic_dec(&pd->usecnt);
156 static void ib_uverbs_release_dev(struct kobject *kobj)
158 struct ib_uverbs_device *dev =
159 container_of(kobj, struct ib_uverbs_device, kobj);
161 cleanup_srcu_struct(&dev->disassociate_srcu);
165 static struct kobj_type ib_uverbs_dev_ktype = {
166 .release = ib_uverbs_release_dev,
169 static void ib_uverbs_release_event_file(struct kref *ref)
171 struct ib_uverbs_event_file *file =
172 container_of(ref, struct ib_uverbs_event_file, ref);
177 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
178 struct ib_uverbs_event_file *ev_file,
179 struct ib_ucq_object *uobj)
181 struct ib_uverbs_event *evt, *tmp;
184 spin_lock_irq(&ev_file->lock);
185 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
186 list_del(&evt->list);
189 spin_unlock_irq(&ev_file->lock);
191 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
194 spin_lock_irq(&file->async_file->lock);
195 list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
196 list_del(&evt->list);
199 spin_unlock_irq(&file->async_file->lock);
202 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
203 struct ib_uevent_object *uobj)
205 struct ib_uverbs_event *evt, *tmp;
207 spin_lock_irq(&file->async_file->lock);
208 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
209 list_del(&evt->list);
212 spin_unlock_irq(&file->async_file->lock);
215 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
216 struct ib_uqp_object *uobj)
218 struct ib_uverbs_mcast_entry *mcast, *tmp;
220 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
221 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
222 list_del(&mcast->list);
227 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
228 struct ib_ucontext *context)
230 struct ib_uobject *uobj, *tmp;
232 context->closing = 1;
234 list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
235 struct ib_ah *ah = uobj->object;
237 idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
242 /* Remove MWs before QPs, in order to support type 2A MWs. */
243 list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
244 struct ib_mw *mw = uobj->object;
246 idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
247 uverbs_dealloc_mw(mw);
251 list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
252 struct ib_flow *flow_id = uobj->object;
254 idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
255 ib_destroy_flow(flow_id);
259 list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
260 struct ib_qp *qp = uobj->object;
261 struct ib_uqp_object *uqp =
262 container_of(uobj, struct ib_uqp_object, uevent.uobject);
264 idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
265 if (qp != qp->real_qp) {
268 ib_uverbs_detach_umcast(qp, uqp);
271 ib_uverbs_release_uevent(file, &uqp->uevent);
275 list_for_each_entry_safe(uobj, tmp, &context->rwq_ind_tbl_list, list) {
276 struct ib_rwq_ind_table *rwq_ind_tbl = uobj->object;
277 struct ib_wq **ind_tbl = rwq_ind_tbl->ind_tbl;
279 idr_remove_uobj(&ib_uverbs_rwq_ind_tbl_idr, uobj);
280 ib_destroy_rwq_ind_table(rwq_ind_tbl);
285 list_for_each_entry_safe(uobj, tmp, &context->wq_list, list) {
286 struct ib_wq *wq = uobj->object;
287 struct ib_uwq_object *uwq =
288 container_of(uobj, struct ib_uwq_object, uevent.uobject);
290 idr_remove_uobj(&ib_uverbs_wq_idr, uobj);
292 ib_uverbs_release_uevent(file, &uwq->uevent);
296 list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
297 struct ib_srq *srq = uobj->object;
298 struct ib_uevent_object *uevent =
299 container_of(uobj, struct ib_uevent_object, uobject);
301 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
303 ib_uverbs_release_uevent(file, uevent);
307 list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
308 struct ib_cq *cq = uobj->object;
309 struct ib_uverbs_event_file *ev_file = cq->cq_context;
310 struct ib_ucq_object *ucq =
311 container_of(uobj, struct ib_ucq_object, uobject);
313 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
315 ib_uverbs_release_ucq(file, ev_file, ucq);
319 list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
320 struct ib_mr *mr = uobj->object;
322 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
327 mutex_lock(&file->device->xrcd_tree_mutex);
328 list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
329 struct ib_xrcd *xrcd = uobj->object;
330 struct ib_uxrcd_object *uxrcd =
331 container_of(uobj, struct ib_uxrcd_object, uobject);
333 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
334 ib_uverbs_dealloc_xrcd(file->device, xrcd);
337 mutex_unlock(&file->device->xrcd_tree_mutex);
339 list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
340 struct ib_pd *pd = uobj->object;
342 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
347 put_pid(context->tgid);
349 return context->device->dealloc_ucontext(context);
352 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
354 complete(&dev->comp);
357 static void ib_uverbs_release_file(struct kref *ref)
359 struct ib_uverbs_file *file =
360 container_of(ref, struct ib_uverbs_file, ref);
361 struct ib_device *ib_dev;
364 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
365 ib_dev = srcu_dereference(file->device->ib_dev,
366 &file->device->disassociate_srcu);
367 if (ib_dev && !ib_dev->disassociate_ucontext)
368 module_put(ib_dev->owner);
369 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
371 if (atomic_dec_and_test(&file->device->refcount))
372 ib_uverbs_comp_dev(file->device);
377 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
378 size_t count, loff_t *pos)
380 struct ib_uverbs_event_file *file = filp->private_data;
381 struct ib_uverbs_event *event;
385 spin_lock_irq(&file->lock);
387 while (list_empty(&file->event_list)) {
388 spin_unlock_irq(&file->lock);
390 if (filp->f_flags & O_NONBLOCK)
393 if (wait_event_interruptible(file->poll_wait,
394 (!list_empty(&file->event_list) ||
395 /* The barriers built into wait_event_interruptible()
396 * and wake_up() guarentee this will see the null set
399 !file->uverbs_file->device->ib_dev)))
402 /* If device was disassociated and no event exists set an error */
403 if (list_empty(&file->event_list) &&
404 !file->uverbs_file->device->ib_dev)
407 spin_lock_irq(&file->lock);
410 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
413 eventsz = sizeof (struct ib_uverbs_async_event_desc);
415 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
417 if (eventsz > count) {
421 list_del(file->event_list.next);
422 if (event->counter) {
424 list_del(&event->obj_list);
428 spin_unlock_irq(&file->lock);
431 if (copy_to_user(buf, event, eventsz))
442 static unsigned int ib_uverbs_event_poll(struct file *filp,
443 struct poll_table_struct *wait)
445 unsigned int pollflags = 0;
446 struct ib_uverbs_event_file *file = filp->private_data;
448 poll_wait(filp, &file->poll_wait, wait);
450 spin_lock_irq(&file->lock);
451 if (!list_empty(&file->event_list))
452 pollflags = POLLIN | POLLRDNORM;
453 spin_unlock_irq(&file->lock);
458 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
460 struct ib_uverbs_event_file *file = filp->private_data;
462 return fasync_helper(fd, filp, on, &file->async_queue);
465 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
467 struct ib_uverbs_event_file *file = filp->private_data;
468 struct ib_uverbs_event *entry, *tmp;
469 int closed_already = 0;
471 mutex_lock(&file->uverbs_file->device->lists_mutex);
472 spin_lock_irq(&file->lock);
473 closed_already = file->is_closed;
475 list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
477 list_del(&entry->obj_list);
480 spin_unlock_irq(&file->lock);
481 if (!closed_already) {
482 list_del(&file->list);
484 ib_unregister_event_handler(&file->uverbs_file->
487 mutex_unlock(&file->uverbs_file->device->lists_mutex);
489 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
490 kref_put(&file->ref, ib_uverbs_release_event_file);
495 static const struct file_operations uverbs_event_fops = {
496 .owner = THIS_MODULE,
497 .read = ib_uverbs_event_read,
498 .poll = ib_uverbs_event_poll,
499 .release = ib_uverbs_event_close,
500 .fasync = ib_uverbs_event_fasync,
504 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
506 struct ib_uverbs_event_file *file = cq_context;
507 struct ib_ucq_object *uobj;
508 struct ib_uverbs_event *entry;
514 spin_lock_irqsave(&file->lock, flags);
515 if (file->is_closed) {
516 spin_unlock_irqrestore(&file->lock, flags);
520 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
522 spin_unlock_irqrestore(&file->lock, flags);
526 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
528 entry->desc.comp.cq_handle = cq->uobject->user_handle;
529 entry->counter = &uobj->comp_events_reported;
531 list_add_tail(&entry->list, &file->event_list);
532 list_add_tail(&entry->obj_list, &uobj->comp_list);
533 spin_unlock_irqrestore(&file->lock, flags);
535 wake_up_interruptible(&file->poll_wait);
536 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
539 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
540 __u64 element, __u64 event,
541 struct list_head *obj_list,
544 struct ib_uverbs_event *entry;
547 spin_lock_irqsave(&file->async_file->lock, flags);
548 if (file->async_file->is_closed) {
549 spin_unlock_irqrestore(&file->async_file->lock, flags);
553 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
555 spin_unlock_irqrestore(&file->async_file->lock, flags);
559 entry->desc.async.element = element;
560 entry->desc.async.event_type = event;
561 entry->desc.async.reserved = 0;
562 entry->counter = counter;
564 list_add_tail(&entry->list, &file->async_file->event_list);
566 list_add_tail(&entry->obj_list, obj_list);
567 spin_unlock_irqrestore(&file->async_file->lock, flags);
569 wake_up_interruptible(&file->async_file->poll_wait);
570 kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
573 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
575 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
576 struct ib_ucq_object, uobject);
578 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
579 event->event, &uobj->async_list,
580 &uobj->async_events_reported);
583 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
585 struct ib_uevent_object *uobj;
587 /* for XRC target qp's, check that qp is live */
588 if (!event->element.qp->uobject || !event->element.qp->uobject->live)
591 uobj = container_of(event->element.qp->uobject,
592 struct ib_uevent_object, uobject);
594 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
595 event->event, &uobj->event_list,
596 &uobj->events_reported);
599 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
601 struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
602 struct ib_uevent_object, uobject);
604 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
605 event->event, &uobj->event_list,
606 &uobj->events_reported);
609 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
611 struct ib_uevent_object *uobj;
613 uobj = container_of(event->element.srq->uobject,
614 struct ib_uevent_object, uobject);
616 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
617 event->event, &uobj->event_list,
618 &uobj->events_reported);
621 void ib_uverbs_event_handler(struct ib_event_handler *handler,
622 struct ib_event *event)
624 struct ib_uverbs_file *file =
625 container_of(handler, struct ib_uverbs_file, event_handler);
627 ib_uverbs_async_handler(file, event->element.port_num, event->event,
631 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
633 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
634 file->async_file = NULL;
637 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
638 struct ib_device *ib_dev,
641 struct ib_uverbs_event_file *ev_file;
645 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
647 return ERR_PTR(-ENOMEM);
649 kref_init(&ev_file->ref);
650 spin_lock_init(&ev_file->lock);
651 INIT_LIST_HEAD(&ev_file->event_list);
652 init_waitqueue_head(&ev_file->poll_wait);
653 ev_file->uverbs_file = uverbs_file;
654 kref_get(&ev_file->uverbs_file->ref);
655 ev_file->async_queue = NULL;
656 ev_file->is_closed = 0;
658 filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
663 mutex_lock(&uverbs_file->device->lists_mutex);
664 list_add_tail(&ev_file->list,
665 &uverbs_file->device->uverbs_events_file_list);
666 mutex_unlock(&uverbs_file->device->lists_mutex);
669 WARN_ON(uverbs_file->async_file);
670 uverbs_file->async_file = ev_file;
671 kref_get(&uverbs_file->async_file->ref);
672 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
674 ib_uverbs_event_handler);
675 ret = ib_register_event_handler(&uverbs_file->event_handler);
679 /* At that point async file stuff was fully set */
680 ev_file->is_async = 1;
687 kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
688 uverbs_file->async_file = NULL;
692 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
693 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
698 * Look up a completion event file by FD. If lookup is successful,
699 * takes a ref to the event file struct that it returns; if
700 * unsuccessful, returns NULL.
702 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
704 struct ib_uverbs_event_file *ev_file = NULL;
705 struct fd f = fdget(fd);
710 if (f.file->f_op != &uverbs_event_fops)
713 ev_file = f.file->private_data;
714 if (ev_file->is_async) {
719 kref_get(&ev_file->ref);
726 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
730 if (command <= IB_USER_VERBS_CMD_OPEN_QP)
731 mask = ib_dev->uverbs_cmd_mask;
733 mask = ib_dev->uverbs_ex_cmd_mask;
735 if (mask & ((u64)1 << command))
741 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
742 size_t count, loff_t *pos)
744 struct ib_uverbs_file *file = filp->private_data;
745 struct ib_device *ib_dev;
746 struct ib_uverbs_cmd_hdr hdr;
752 if (WARN_ON_ONCE(!ib_safe_file_access(filp)))
755 if (count < sizeof hdr)
758 if (copy_from_user(&hdr, buf, sizeof hdr))
761 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
762 ib_dev = srcu_dereference(file->device->ib_dev,
763 &file->device->disassociate_srcu);
769 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
770 IB_USER_VERBS_CMD_COMMAND_MASK)) {
775 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
776 if (verify_command_mask(ib_dev, command)) {
781 if (!file->ucontext &&
782 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
787 flags = (hdr.command &
788 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
791 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
792 !uverbs_cmd_table[command]) {
797 if (hdr.in_words * 4 != count) {
802 ret = uverbs_cmd_table[command](file, ib_dev,
807 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
808 struct ib_uverbs_ex_cmd_hdr ex_hdr;
809 struct ib_udata ucore;
811 size_t written_count = count;
813 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
814 !uverbs_ex_cmd_table[command]) {
819 if (!file->ucontext) {
824 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
829 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
834 count -= sizeof(hdr) + sizeof(ex_hdr);
835 buf += sizeof(hdr) + sizeof(ex_hdr);
837 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
842 if (ex_hdr.cmd_hdr_reserved) {
847 if (ex_hdr.response) {
848 if (!hdr.out_words && !ex_hdr.provider_out_words) {
853 if (!access_ok(VERIFY_WRITE,
854 (void __user *) (unsigned long) ex_hdr.response,
855 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
860 if (hdr.out_words || ex_hdr.provider_out_words) {
866 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
867 hdr.in_words * 8, hdr.out_words * 8);
869 INIT_UDATA_BUF_OR_NULL(&uhw,
871 (unsigned long) ex_hdr.response + ucore.outlen,
872 ex_hdr.provider_in_words * 8,
873 ex_hdr.provider_out_words * 8);
875 ret = uverbs_ex_cmd_table[command](file,
886 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
890 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
892 struct ib_uverbs_file *file = filp->private_data;
893 struct ib_device *ib_dev;
897 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
898 ib_dev = srcu_dereference(file->device->ib_dev,
899 &file->device->disassociate_srcu);
908 ret = ib_dev->mmap(file->ucontext, vma);
910 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
915 * ib_uverbs_open() does not need the BKL:
917 * - the ib_uverbs_device structures are properly reference counted and
918 * everything else is purely local to the file being created, so
919 * races against other open calls are not a problem;
920 * - there is no ioctl method to race against;
921 * - the open method will either immediately run -ENXIO, or all
922 * required initialization will be done.
924 static int ib_uverbs_open(struct inode *inode, struct file *filp)
926 struct ib_uverbs_device *dev;
927 struct ib_uverbs_file *file;
928 struct ib_device *ib_dev;
930 int module_dependent;
933 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
934 if (!atomic_inc_not_zero(&dev->refcount))
937 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
938 mutex_lock(&dev->lists_mutex);
939 ib_dev = srcu_dereference(dev->ib_dev,
940 &dev->disassociate_srcu);
946 /* In case IB device supports disassociate ucontext, there is no hard
947 * dependency between uverbs device and its low level device.
949 module_dependent = !(ib_dev->disassociate_ucontext);
951 if (module_dependent) {
952 if (!try_module_get(ib_dev->owner)) {
958 file = kzalloc(sizeof(*file), GFP_KERNEL);
961 if (module_dependent)
968 file->ucontext = NULL;
969 file->async_file = NULL;
970 kref_init(&file->ref);
971 mutex_init(&file->mutex);
972 mutex_init(&file->cleanup_mutex);
974 filp->private_data = file;
975 kobject_get(&dev->kobj);
976 list_add_tail(&file->list, &dev->uverbs_file_list);
977 mutex_unlock(&dev->lists_mutex);
978 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
980 return nonseekable_open(inode, filp);
983 module_put(ib_dev->owner);
986 mutex_unlock(&dev->lists_mutex);
987 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
988 if (atomic_dec_and_test(&dev->refcount))
989 ib_uverbs_comp_dev(dev);
994 static int ib_uverbs_close(struct inode *inode, struct file *filp)
996 struct ib_uverbs_file *file = filp->private_data;
997 struct ib_uverbs_device *dev = file->device;
999 mutex_lock(&file->cleanup_mutex);
1000 if (file->ucontext) {
1001 ib_uverbs_cleanup_ucontext(file, file->ucontext);
1002 file->ucontext = NULL;
1004 mutex_unlock(&file->cleanup_mutex);
1006 mutex_lock(&file->device->lists_mutex);
1007 if (!file->is_closed) {
1008 list_del(&file->list);
1009 file->is_closed = 1;
1011 mutex_unlock(&file->device->lists_mutex);
1013 if (file->async_file)
1014 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
1016 kref_put(&file->ref, ib_uverbs_release_file);
1017 kobject_put(&dev->kobj);
1022 static const struct file_operations uverbs_fops = {
1023 .owner = THIS_MODULE,
1024 .write = ib_uverbs_write,
1025 .open = ib_uverbs_open,
1026 .release = ib_uverbs_close,
1027 .llseek = no_llseek,
1030 static const struct file_operations uverbs_mmap_fops = {
1031 .owner = THIS_MODULE,
1032 .write = ib_uverbs_write,
1033 .mmap = ib_uverbs_mmap,
1034 .open = ib_uverbs_open,
1035 .release = ib_uverbs_close,
1036 .llseek = no_llseek,
1039 static struct ib_client uverbs_client = {
1041 .add = ib_uverbs_add_one,
1042 .remove = ib_uverbs_remove_one
1045 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
1050 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1051 struct ib_device *ib_dev;
1056 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1057 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1059 ret = sprintf(buf, "%s\n", ib_dev->name);
1060 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1064 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1066 static ssize_t show_dev_abi_version(struct device *device,
1067 struct device_attribute *attr, char *buf)
1069 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1072 struct ib_device *ib_dev;
1076 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1077 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1079 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1080 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1084 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1086 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1087 __stringify(IB_USER_VERBS_ABI_VERSION));
1089 static dev_t overflow_maj;
1090 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1093 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1094 * requesting a new major number and doubling the number of max devices we
1095 * support. It's stupid, but simple.
1097 static int find_overflow_devnum(void)
1101 if (!overflow_maj) {
1102 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1103 "infiniband_verbs");
1105 pr_err("user_verbs: couldn't register dynamic device number\n");
1110 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1111 if (ret >= IB_UVERBS_MAX_DEVICES)
1117 static void ib_uverbs_add_one(struct ib_device *device)
1121 struct ib_uverbs_device *uverbs_dev;
1124 if (!device->alloc_ucontext)
1127 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1131 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1137 atomic_set(&uverbs_dev->refcount, 1);
1138 init_completion(&uverbs_dev->comp);
1139 uverbs_dev->xrcd_tree = RB_ROOT;
1140 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1141 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1142 mutex_init(&uverbs_dev->lists_mutex);
1143 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1144 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1146 spin_lock(&map_lock);
1147 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1148 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1149 spin_unlock(&map_lock);
1150 devnum = find_overflow_devnum();
1154 spin_lock(&map_lock);
1155 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1156 base = devnum + overflow_maj;
1157 set_bit(devnum, overflow_map);
1159 uverbs_dev->devnum = devnum;
1160 base = devnum + IB_UVERBS_BASE_DEV;
1161 set_bit(devnum, dev_map);
1163 spin_unlock(&map_lock);
1165 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1166 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1168 cdev_init(&uverbs_dev->cdev, NULL);
1169 uverbs_dev->cdev.owner = THIS_MODULE;
1170 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1171 uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1172 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1173 if (cdev_add(&uverbs_dev->cdev, base, 1))
1176 uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
1177 uverbs_dev->cdev.dev, uverbs_dev,
1178 "uverbs%d", uverbs_dev->devnum);
1179 if (IS_ERR(uverbs_dev->dev))
1182 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1184 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1187 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1192 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1195 cdev_del(&uverbs_dev->cdev);
1196 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1197 clear_bit(devnum, dev_map);
1199 clear_bit(devnum, overflow_map);
1202 if (atomic_dec_and_test(&uverbs_dev->refcount))
1203 ib_uverbs_comp_dev(uverbs_dev);
1204 wait_for_completion(&uverbs_dev->comp);
1205 kobject_put(&uverbs_dev->kobj);
1209 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1210 struct ib_device *ib_dev)
1212 struct ib_uverbs_file *file;
1213 struct ib_uverbs_event_file *event_file;
1214 struct ib_event event;
1216 /* Pending running commands to terminate */
1217 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1218 event.event = IB_EVENT_DEVICE_FATAL;
1219 event.element.port_num = 0;
1220 event.device = ib_dev;
1222 mutex_lock(&uverbs_dev->lists_mutex);
1223 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1224 struct ib_ucontext *ucontext;
1225 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1226 struct ib_uverbs_file, list);
1227 file->is_closed = 1;
1228 list_del(&file->list);
1229 kref_get(&file->ref);
1230 mutex_unlock(&uverbs_dev->lists_mutex);
1232 ib_uverbs_event_handler(&file->event_handler, &event);
1234 mutex_lock(&file->cleanup_mutex);
1235 ucontext = file->ucontext;
1236 file->ucontext = NULL;
1237 mutex_unlock(&file->cleanup_mutex);
1239 /* At this point ib_uverbs_close cannot be running
1240 * ib_uverbs_cleanup_ucontext
1243 /* We must release the mutex before going ahead and
1244 * calling disassociate_ucontext. disassociate_ucontext
1245 * might end up indirectly calling uverbs_close,
1246 * for example due to freeing the resources
1249 ib_dev->disassociate_ucontext(ucontext);
1250 ib_uverbs_cleanup_ucontext(file, ucontext);
1253 mutex_lock(&uverbs_dev->lists_mutex);
1254 kref_put(&file->ref, ib_uverbs_release_file);
1257 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1258 event_file = list_first_entry(&uverbs_dev->
1259 uverbs_events_file_list,
1260 struct ib_uverbs_event_file,
1262 spin_lock_irq(&event_file->lock);
1263 event_file->is_closed = 1;
1264 spin_unlock_irq(&event_file->lock);
1266 list_del(&event_file->list);
1267 if (event_file->is_async) {
1268 ib_unregister_event_handler(&event_file->uverbs_file->
1270 event_file->uverbs_file->event_handler.device = NULL;
1273 wake_up_interruptible(&event_file->poll_wait);
1274 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1276 mutex_unlock(&uverbs_dev->lists_mutex);
1279 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1281 struct ib_uverbs_device *uverbs_dev = client_data;
1282 int wait_clients = 1;
1287 dev_set_drvdata(uverbs_dev->dev, NULL);
1288 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1289 cdev_del(&uverbs_dev->cdev);
1291 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1292 clear_bit(uverbs_dev->devnum, dev_map);
1294 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1296 if (device->disassociate_ucontext) {
1297 /* We disassociate HW resources and immediately return.
1298 * Userspace will see a EIO errno for all future access.
1299 * Upon returning, ib_device may be freed internally and is not
1301 * uverbs_device is still available until all clients close
1302 * their files, then the uverbs device ref count will be zero
1303 * and its resources will be freed.
1304 * Note: At this point no more files can be opened since the
1305 * cdev was deleted, however active clients can still issue
1306 * commands and close their open files.
1308 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1309 ib_uverbs_free_hw_resources(uverbs_dev, device);
1313 if (atomic_dec_and_test(&uverbs_dev->refcount))
1314 ib_uverbs_comp_dev(uverbs_dev);
1316 wait_for_completion(&uverbs_dev->comp);
1317 kobject_put(&uverbs_dev->kobj);
1320 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1324 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1327 static int __init ib_uverbs_init(void)
1331 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1332 "infiniband_verbs");
1334 pr_err("user_verbs: couldn't register device number\n");
1338 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1339 if (IS_ERR(uverbs_class)) {
1340 ret = PTR_ERR(uverbs_class);
1341 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1345 uverbs_class->devnode = uverbs_devnode;
1347 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1349 pr_err("user_verbs: couldn't create abi_version attribute\n");
1353 ret = ib_register_client(&uverbs_client);
1355 pr_err("user_verbs: couldn't register client\n");
1362 class_destroy(uverbs_class);
1365 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1371 static void __exit ib_uverbs_cleanup(void)
1373 ib_unregister_client(&uverbs_client);
1374 class_destroy(uverbs_class);
1375 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1377 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1378 idr_destroy(&ib_uverbs_pd_idr);
1379 idr_destroy(&ib_uverbs_mr_idr);
1380 idr_destroy(&ib_uverbs_mw_idr);
1381 idr_destroy(&ib_uverbs_ah_idr);
1382 idr_destroy(&ib_uverbs_cq_idr);
1383 idr_destroy(&ib_uverbs_qp_idr);
1384 idr_destroy(&ib_uverbs_srq_idr);
1387 module_init(ib_uverbs_init);
1388 module_exit(ib_uverbs_cleanup);