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)
266 ib_uverbs_detach_umcast(qp, uqp);
268 ib_uverbs_release_uevent(file, &uqp->uevent);
272 list_for_each_entry_safe(uobj, tmp, &context->rwq_ind_tbl_list, list) {
273 struct ib_rwq_ind_table *rwq_ind_tbl = uobj->object;
274 struct ib_wq **ind_tbl = rwq_ind_tbl->ind_tbl;
276 idr_remove_uobj(&ib_uverbs_rwq_ind_tbl_idr, uobj);
277 ib_destroy_rwq_ind_table(rwq_ind_tbl);
282 list_for_each_entry_safe(uobj, tmp, &context->wq_list, list) {
283 struct ib_wq *wq = uobj->object;
284 struct ib_uwq_object *uwq =
285 container_of(uobj, struct ib_uwq_object, uevent.uobject);
287 idr_remove_uobj(&ib_uverbs_wq_idr, uobj);
289 ib_uverbs_release_uevent(file, &uwq->uevent);
293 list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
294 struct ib_srq *srq = uobj->object;
295 struct ib_uevent_object *uevent =
296 container_of(uobj, struct ib_uevent_object, uobject);
298 idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
300 ib_uverbs_release_uevent(file, uevent);
304 list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
305 struct ib_cq *cq = uobj->object;
306 struct ib_uverbs_event_file *ev_file = cq->cq_context;
307 struct ib_ucq_object *ucq =
308 container_of(uobj, struct ib_ucq_object, uobject);
310 idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
312 ib_uverbs_release_ucq(file, ev_file, ucq);
316 list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
317 struct ib_mr *mr = uobj->object;
319 idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
324 mutex_lock(&file->device->xrcd_tree_mutex);
325 list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
326 struct ib_xrcd *xrcd = uobj->object;
327 struct ib_uxrcd_object *uxrcd =
328 container_of(uobj, struct ib_uxrcd_object, uobject);
330 idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
331 ib_uverbs_dealloc_xrcd(file->device, xrcd);
334 mutex_unlock(&file->device->xrcd_tree_mutex);
336 list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
337 struct ib_pd *pd = uobj->object;
339 idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
344 put_pid(context->tgid);
346 return context->device->dealloc_ucontext(context);
349 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
351 complete(&dev->comp);
354 static void ib_uverbs_release_file(struct kref *ref)
356 struct ib_uverbs_file *file =
357 container_of(ref, struct ib_uverbs_file, ref);
358 struct ib_device *ib_dev;
361 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
362 ib_dev = srcu_dereference(file->device->ib_dev,
363 &file->device->disassociate_srcu);
364 if (ib_dev && !ib_dev->disassociate_ucontext)
365 module_put(ib_dev->owner);
366 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
368 if (atomic_dec_and_test(&file->device->refcount))
369 ib_uverbs_comp_dev(file->device);
374 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
375 size_t count, loff_t *pos)
377 struct ib_uverbs_event_file *file = filp->private_data;
378 struct ib_uverbs_event *event;
382 spin_lock_irq(&file->lock);
384 while (list_empty(&file->event_list)) {
385 spin_unlock_irq(&file->lock);
387 if (filp->f_flags & O_NONBLOCK)
390 if (wait_event_interruptible(file->poll_wait,
391 (!list_empty(&file->event_list) ||
392 /* The barriers built into wait_event_interruptible()
393 * and wake_up() guarentee this will see the null set
396 !file->uverbs_file->device->ib_dev)))
399 /* If device was disassociated and no event exists set an error */
400 if (list_empty(&file->event_list) &&
401 !file->uverbs_file->device->ib_dev)
404 spin_lock_irq(&file->lock);
407 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
410 eventsz = sizeof (struct ib_uverbs_async_event_desc);
412 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
414 if (eventsz > count) {
418 list_del(file->event_list.next);
419 if (event->counter) {
421 list_del(&event->obj_list);
425 spin_unlock_irq(&file->lock);
428 if (copy_to_user(buf, event, eventsz))
439 static unsigned int ib_uverbs_event_poll(struct file *filp,
440 struct poll_table_struct *wait)
442 unsigned int pollflags = 0;
443 struct ib_uverbs_event_file *file = filp->private_data;
445 poll_wait(filp, &file->poll_wait, wait);
447 spin_lock_irq(&file->lock);
448 if (!list_empty(&file->event_list))
449 pollflags = POLLIN | POLLRDNORM;
450 spin_unlock_irq(&file->lock);
455 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
457 struct ib_uverbs_event_file *file = filp->private_data;
459 return fasync_helper(fd, filp, on, &file->async_queue);
462 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
464 struct ib_uverbs_event_file *file = filp->private_data;
465 struct ib_uverbs_event *entry, *tmp;
466 int closed_already = 0;
468 mutex_lock(&file->uverbs_file->device->lists_mutex);
469 spin_lock_irq(&file->lock);
470 closed_already = file->is_closed;
472 list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
474 list_del(&entry->obj_list);
477 spin_unlock_irq(&file->lock);
478 if (!closed_already) {
479 list_del(&file->list);
481 ib_unregister_event_handler(&file->uverbs_file->
484 mutex_unlock(&file->uverbs_file->device->lists_mutex);
486 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
487 kref_put(&file->ref, ib_uverbs_release_event_file);
492 static const struct file_operations uverbs_event_fops = {
493 .owner = THIS_MODULE,
494 .read = ib_uverbs_event_read,
495 .poll = ib_uverbs_event_poll,
496 .release = ib_uverbs_event_close,
497 .fasync = ib_uverbs_event_fasync,
501 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
503 struct ib_uverbs_event_file *file = cq_context;
504 struct ib_ucq_object *uobj;
505 struct ib_uverbs_event *entry;
511 spin_lock_irqsave(&file->lock, flags);
512 if (file->is_closed) {
513 spin_unlock_irqrestore(&file->lock, flags);
517 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
519 spin_unlock_irqrestore(&file->lock, flags);
523 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
525 entry->desc.comp.cq_handle = cq->uobject->user_handle;
526 entry->counter = &uobj->comp_events_reported;
528 list_add_tail(&entry->list, &file->event_list);
529 list_add_tail(&entry->obj_list, &uobj->comp_list);
530 spin_unlock_irqrestore(&file->lock, flags);
532 wake_up_interruptible(&file->poll_wait);
533 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
536 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
537 __u64 element, __u64 event,
538 struct list_head *obj_list,
541 struct ib_uverbs_event *entry;
544 spin_lock_irqsave(&file->async_file->lock, flags);
545 if (file->async_file->is_closed) {
546 spin_unlock_irqrestore(&file->async_file->lock, flags);
550 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
552 spin_unlock_irqrestore(&file->async_file->lock, flags);
556 entry->desc.async.element = element;
557 entry->desc.async.event_type = event;
558 entry->desc.async.reserved = 0;
559 entry->counter = counter;
561 list_add_tail(&entry->list, &file->async_file->event_list);
563 list_add_tail(&entry->obj_list, obj_list);
564 spin_unlock_irqrestore(&file->async_file->lock, flags);
566 wake_up_interruptible(&file->async_file->poll_wait);
567 kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
570 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
572 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
573 struct ib_ucq_object, uobject);
575 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
576 event->event, &uobj->async_list,
577 &uobj->async_events_reported);
580 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
582 struct ib_uevent_object *uobj;
584 /* for XRC target qp's, check that qp is live */
585 if (!event->element.qp->uobject || !event->element.qp->uobject->live)
588 uobj = container_of(event->element.qp->uobject,
589 struct ib_uevent_object, uobject);
591 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
592 event->event, &uobj->event_list,
593 &uobj->events_reported);
596 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
598 struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
599 struct ib_uevent_object, uobject);
601 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
602 event->event, &uobj->event_list,
603 &uobj->events_reported);
606 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
608 struct ib_uevent_object *uobj;
610 uobj = container_of(event->element.srq->uobject,
611 struct ib_uevent_object, uobject);
613 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
614 event->event, &uobj->event_list,
615 &uobj->events_reported);
618 void ib_uverbs_event_handler(struct ib_event_handler *handler,
619 struct ib_event *event)
621 struct ib_uverbs_file *file =
622 container_of(handler, struct ib_uverbs_file, event_handler);
624 ib_uverbs_async_handler(file, event->element.port_num, event->event,
628 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
630 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
631 file->async_file = NULL;
634 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
635 struct ib_device *ib_dev,
638 struct ib_uverbs_event_file *ev_file;
642 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
644 return ERR_PTR(-ENOMEM);
646 kref_init(&ev_file->ref);
647 spin_lock_init(&ev_file->lock);
648 INIT_LIST_HEAD(&ev_file->event_list);
649 init_waitqueue_head(&ev_file->poll_wait);
650 ev_file->uverbs_file = uverbs_file;
651 kref_get(&ev_file->uverbs_file->ref);
652 ev_file->async_queue = NULL;
653 ev_file->is_closed = 0;
655 filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
660 mutex_lock(&uverbs_file->device->lists_mutex);
661 list_add_tail(&ev_file->list,
662 &uverbs_file->device->uverbs_events_file_list);
663 mutex_unlock(&uverbs_file->device->lists_mutex);
666 WARN_ON(uverbs_file->async_file);
667 uverbs_file->async_file = ev_file;
668 kref_get(&uverbs_file->async_file->ref);
669 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
671 ib_uverbs_event_handler);
672 ret = ib_register_event_handler(&uverbs_file->event_handler);
676 /* At that point async file stuff was fully set */
677 ev_file->is_async = 1;
684 kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
685 uverbs_file->async_file = NULL;
689 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
690 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
695 * Look up a completion event file by FD. If lookup is successful,
696 * takes a ref to the event file struct that it returns; if
697 * unsuccessful, returns NULL.
699 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
701 struct ib_uverbs_event_file *ev_file = NULL;
702 struct fd f = fdget(fd);
707 if (f.file->f_op != &uverbs_event_fops)
710 ev_file = f.file->private_data;
711 if (ev_file->is_async) {
716 kref_get(&ev_file->ref);
723 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
727 if (command <= IB_USER_VERBS_CMD_OPEN_QP)
728 mask = ib_dev->uverbs_cmd_mask;
730 mask = ib_dev->uverbs_ex_cmd_mask;
732 if (mask & ((u64)1 << command))
738 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
739 size_t count, loff_t *pos)
741 struct ib_uverbs_file *file = filp->private_data;
742 struct ib_device *ib_dev;
743 struct ib_uverbs_cmd_hdr hdr;
749 if (WARN_ON_ONCE(!ib_safe_file_access(filp)))
752 if (count < sizeof hdr)
755 if (copy_from_user(&hdr, buf, sizeof hdr))
758 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
759 ib_dev = srcu_dereference(file->device->ib_dev,
760 &file->device->disassociate_srcu);
766 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
767 IB_USER_VERBS_CMD_COMMAND_MASK)) {
772 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
773 if (verify_command_mask(ib_dev, command)) {
778 if (!file->ucontext &&
779 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
784 flags = (hdr.command &
785 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
788 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
789 !uverbs_cmd_table[command]) {
794 if (hdr.in_words * 4 != count) {
799 ret = uverbs_cmd_table[command](file, ib_dev,
804 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
805 struct ib_uverbs_ex_cmd_hdr ex_hdr;
806 struct ib_udata ucore;
808 size_t written_count = count;
810 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
811 !uverbs_ex_cmd_table[command]) {
816 if (!file->ucontext) {
821 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
826 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
831 count -= sizeof(hdr) + sizeof(ex_hdr);
832 buf += sizeof(hdr) + sizeof(ex_hdr);
834 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
839 if (ex_hdr.cmd_hdr_reserved) {
844 if (ex_hdr.response) {
845 if (!hdr.out_words && !ex_hdr.provider_out_words) {
850 if (!access_ok(VERIFY_WRITE,
851 (void __user *) (unsigned long) ex_hdr.response,
852 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
857 if (hdr.out_words || ex_hdr.provider_out_words) {
863 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
864 hdr.in_words * 8, hdr.out_words * 8);
866 INIT_UDATA_BUF_OR_NULL(&uhw,
868 (unsigned long) ex_hdr.response + ucore.outlen,
869 ex_hdr.provider_in_words * 8,
870 ex_hdr.provider_out_words * 8);
872 ret = uverbs_ex_cmd_table[command](file,
883 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
887 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
889 struct ib_uverbs_file *file = filp->private_data;
890 struct ib_device *ib_dev;
894 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
895 ib_dev = srcu_dereference(file->device->ib_dev,
896 &file->device->disassociate_srcu);
905 ret = ib_dev->mmap(file->ucontext, vma);
907 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
912 * ib_uverbs_open() does not need the BKL:
914 * - the ib_uverbs_device structures are properly reference counted and
915 * everything else is purely local to the file being created, so
916 * races against other open calls are not a problem;
917 * - there is no ioctl method to race against;
918 * - the open method will either immediately run -ENXIO, or all
919 * required initialization will be done.
921 static int ib_uverbs_open(struct inode *inode, struct file *filp)
923 struct ib_uverbs_device *dev;
924 struct ib_uverbs_file *file;
925 struct ib_device *ib_dev;
927 int module_dependent;
930 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
931 if (!atomic_inc_not_zero(&dev->refcount))
934 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
935 mutex_lock(&dev->lists_mutex);
936 ib_dev = srcu_dereference(dev->ib_dev,
937 &dev->disassociate_srcu);
943 /* In case IB device supports disassociate ucontext, there is no hard
944 * dependency between uverbs device and its low level device.
946 module_dependent = !(ib_dev->disassociate_ucontext);
948 if (module_dependent) {
949 if (!try_module_get(ib_dev->owner)) {
955 file = kzalloc(sizeof(*file), GFP_KERNEL);
958 if (module_dependent)
965 file->ucontext = NULL;
966 file->async_file = NULL;
967 kref_init(&file->ref);
968 mutex_init(&file->mutex);
969 mutex_init(&file->cleanup_mutex);
971 filp->private_data = file;
972 kobject_get(&dev->kobj);
973 list_add_tail(&file->list, &dev->uverbs_file_list);
974 mutex_unlock(&dev->lists_mutex);
975 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
977 return nonseekable_open(inode, filp);
980 module_put(ib_dev->owner);
983 mutex_unlock(&dev->lists_mutex);
984 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
985 if (atomic_dec_and_test(&dev->refcount))
986 ib_uverbs_comp_dev(dev);
991 static int ib_uverbs_close(struct inode *inode, struct file *filp)
993 struct ib_uverbs_file *file = filp->private_data;
994 struct ib_uverbs_device *dev = file->device;
996 mutex_lock(&file->cleanup_mutex);
997 if (file->ucontext) {
998 ib_uverbs_cleanup_ucontext(file, file->ucontext);
999 file->ucontext = NULL;
1001 mutex_unlock(&file->cleanup_mutex);
1003 mutex_lock(&file->device->lists_mutex);
1004 if (!file->is_closed) {
1005 list_del(&file->list);
1006 file->is_closed = 1;
1008 mutex_unlock(&file->device->lists_mutex);
1010 if (file->async_file)
1011 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
1013 kref_put(&file->ref, ib_uverbs_release_file);
1014 kobject_put(&dev->kobj);
1019 static const struct file_operations uverbs_fops = {
1020 .owner = THIS_MODULE,
1021 .write = ib_uverbs_write,
1022 .open = ib_uverbs_open,
1023 .release = ib_uverbs_close,
1024 .llseek = no_llseek,
1027 static const struct file_operations uverbs_mmap_fops = {
1028 .owner = THIS_MODULE,
1029 .write = ib_uverbs_write,
1030 .mmap = ib_uverbs_mmap,
1031 .open = ib_uverbs_open,
1032 .release = ib_uverbs_close,
1033 .llseek = no_llseek,
1036 static struct ib_client uverbs_client = {
1038 .add = ib_uverbs_add_one,
1039 .remove = ib_uverbs_remove_one
1042 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
1047 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1048 struct ib_device *ib_dev;
1053 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1054 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1056 ret = sprintf(buf, "%s\n", ib_dev->name);
1057 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1061 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1063 static ssize_t show_dev_abi_version(struct device *device,
1064 struct device_attribute *attr, char *buf)
1066 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1069 struct ib_device *ib_dev;
1073 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1074 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1076 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1077 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1081 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1083 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1084 __stringify(IB_USER_VERBS_ABI_VERSION));
1086 static dev_t overflow_maj;
1087 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1090 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1091 * requesting a new major number and doubling the number of max devices we
1092 * support. It's stupid, but simple.
1094 static int find_overflow_devnum(void)
1098 if (!overflow_maj) {
1099 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1100 "infiniband_verbs");
1102 pr_err("user_verbs: couldn't register dynamic device number\n");
1107 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1108 if (ret >= IB_UVERBS_MAX_DEVICES)
1114 static void ib_uverbs_add_one(struct ib_device *device)
1118 struct ib_uverbs_device *uverbs_dev;
1121 if (!device->alloc_ucontext)
1124 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1128 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1134 atomic_set(&uverbs_dev->refcount, 1);
1135 init_completion(&uverbs_dev->comp);
1136 uverbs_dev->xrcd_tree = RB_ROOT;
1137 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1138 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1139 mutex_init(&uverbs_dev->lists_mutex);
1140 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1141 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1143 spin_lock(&map_lock);
1144 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1145 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1146 spin_unlock(&map_lock);
1147 devnum = find_overflow_devnum();
1151 spin_lock(&map_lock);
1152 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1153 base = devnum + overflow_maj;
1154 set_bit(devnum, overflow_map);
1156 uverbs_dev->devnum = devnum;
1157 base = devnum + IB_UVERBS_BASE_DEV;
1158 set_bit(devnum, dev_map);
1160 spin_unlock(&map_lock);
1162 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1163 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1165 cdev_init(&uverbs_dev->cdev, NULL);
1166 uverbs_dev->cdev.owner = THIS_MODULE;
1167 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1168 uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1169 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1170 if (cdev_add(&uverbs_dev->cdev, base, 1))
1173 uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
1174 uverbs_dev->cdev.dev, uverbs_dev,
1175 "uverbs%d", uverbs_dev->devnum);
1176 if (IS_ERR(uverbs_dev->dev))
1179 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1181 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1184 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1189 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1192 cdev_del(&uverbs_dev->cdev);
1193 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1194 clear_bit(devnum, dev_map);
1196 clear_bit(devnum, overflow_map);
1199 if (atomic_dec_and_test(&uverbs_dev->refcount))
1200 ib_uverbs_comp_dev(uverbs_dev);
1201 wait_for_completion(&uverbs_dev->comp);
1202 kobject_put(&uverbs_dev->kobj);
1206 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1207 struct ib_device *ib_dev)
1209 struct ib_uverbs_file *file;
1210 struct ib_uverbs_event_file *event_file;
1211 struct ib_event event;
1213 /* Pending running commands to terminate */
1214 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1215 event.event = IB_EVENT_DEVICE_FATAL;
1216 event.element.port_num = 0;
1217 event.device = ib_dev;
1219 mutex_lock(&uverbs_dev->lists_mutex);
1220 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1221 struct ib_ucontext *ucontext;
1222 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1223 struct ib_uverbs_file, list);
1224 file->is_closed = 1;
1225 list_del(&file->list);
1226 kref_get(&file->ref);
1227 mutex_unlock(&uverbs_dev->lists_mutex);
1229 ib_uverbs_event_handler(&file->event_handler, &event);
1231 mutex_lock(&file->cleanup_mutex);
1232 ucontext = file->ucontext;
1233 file->ucontext = NULL;
1234 mutex_unlock(&file->cleanup_mutex);
1236 /* At this point ib_uverbs_close cannot be running
1237 * ib_uverbs_cleanup_ucontext
1240 /* We must release the mutex before going ahead and
1241 * calling disassociate_ucontext. disassociate_ucontext
1242 * might end up indirectly calling uverbs_close,
1243 * for example due to freeing the resources
1246 ib_dev->disassociate_ucontext(ucontext);
1247 ib_uverbs_cleanup_ucontext(file, ucontext);
1250 mutex_lock(&uverbs_dev->lists_mutex);
1251 kref_put(&file->ref, ib_uverbs_release_file);
1254 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1255 event_file = list_first_entry(&uverbs_dev->
1256 uverbs_events_file_list,
1257 struct ib_uverbs_event_file,
1259 spin_lock_irq(&event_file->lock);
1260 event_file->is_closed = 1;
1261 spin_unlock_irq(&event_file->lock);
1263 list_del(&event_file->list);
1264 if (event_file->is_async) {
1265 ib_unregister_event_handler(&event_file->uverbs_file->
1267 event_file->uverbs_file->event_handler.device = NULL;
1270 wake_up_interruptible(&event_file->poll_wait);
1271 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1273 mutex_unlock(&uverbs_dev->lists_mutex);
1276 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1278 struct ib_uverbs_device *uverbs_dev = client_data;
1279 int wait_clients = 1;
1284 dev_set_drvdata(uverbs_dev->dev, NULL);
1285 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1286 cdev_del(&uverbs_dev->cdev);
1288 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1289 clear_bit(uverbs_dev->devnum, dev_map);
1291 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1293 if (device->disassociate_ucontext) {
1294 /* We disassociate HW resources and immediately return.
1295 * Userspace will see a EIO errno for all future access.
1296 * Upon returning, ib_device may be freed internally and is not
1298 * uverbs_device is still available until all clients close
1299 * their files, then the uverbs device ref count will be zero
1300 * and its resources will be freed.
1301 * Note: At this point no more files can be opened since the
1302 * cdev was deleted, however active clients can still issue
1303 * commands and close their open files.
1305 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1306 ib_uverbs_free_hw_resources(uverbs_dev, device);
1310 if (atomic_dec_and_test(&uverbs_dev->refcount))
1311 ib_uverbs_comp_dev(uverbs_dev);
1313 wait_for_completion(&uverbs_dev->comp);
1314 kobject_put(&uverbs_dev->kobj);
1317 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1321 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1324 static int __init ib_uverbs_init(void)
1328 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1329 "infiniband_verbs");
1331 pr_err("user_verbs: couldn't register device number\n");
1335 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1336 if (IS_ERR(uverbs_class)) {
1337 ret = PTR_ERR(uverbs_class);
1338 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1342 uverbs_class->devnode = uverbs_devnode;
1344 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1346 pr_err("user_verbs: couldn't create abi_version attribute\n");
1350 ret = ib_register_client(&uverbs_client);
1352 pr_err("user_verbs: couldn't register client\n");
1359 class_destroy(uverbs_class);
1362 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1368 static void __exit ib_uverbs_cleanup(void)
1370 ib_unregister_client(&uverbs_client);
1371 class_destroy(uverbs_class);
1372 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1374 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1375 idr_destroy(&ib_uverbs_pd_idr);
1376 idr_destroy(&ib_uverbs_mr_idr);
1377 idr_destroy(&ib_uverbs_mw_idr);
1378 idr_destroy(&ib_uverbs_ah_idr);
1379 idr_destroy(&ib_uverbs_cq_idr);
1380 idr_destroy(&ib_uverbs_qp_idr);
1381 idr_destroy(&ib_uverbs_srq_idr);
1384 module_init(ib_uverbs_init);
1385 module_exit(ib_uverbs_cleanup);