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 <linux/uaccess.h>
54 #include "core_priv.h"
55 #include "rdma_core.h"
57 MODULE_AUTHOR("Roland Dreier");
58 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
59 MODULE_LICENSE("Dual BSD/GPL");
62 IB_UVERBS_MAJOR = 231,
63 IB_UVERBS_BASE_MINOR = 192,
64 IB_UVERBS_MAX_DEVICES = 32
67 #define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
69 static struct class *uverbs_class;
71 static DEFINE_SPINLOCK(map_lock);
72 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
74 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
75 struct ib_device *ib_dev,
76 const char __user *buf, int in_len,
78 [IB_USER_VERBS_CMD_GET_CONTEXT] = ib_uverbs_get_context,
79 [IB_USER_VERBS_CMD_QUERY_DEVICE] = ib_uverbs_query_device,
80 [IB_USER_VERBS_CMD_QUERY_PORT] = ib_uverbs_query_port,
81 [IB_USER_VERBS_CMD_ALLOC_PD] = ib_uverbs_alloc_pd,
82 [IB_USER_VERBS_CMD_DEALLOC_PD] = ib_uverbs_dealloc_pd,
83 [IB_USER_VERBS_CMD_REG_MR] = ib_uverbs_reg_mr,
84 [IB_USER_VERBS_CMD_REREG_MR] = ib_uverbs_rereg_mr,
85 [IB_USER_VERBS_CMD_DEREG_MR] = ib_uverbs_dereg_mr,
86 [IB_USER_VERBS_CMD_ALLOC_MW] = ib_uverbs_alloc_mw,
87 [IB_USER_VERBS_CMD_DEALLOC_MW] = ib_uverbs_dealloc_mw,
88 [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
89 [IB_USER_VERBS_CMD_CREATE_CQ] = ib_uverbs_create_cq,
90 [IB_USER_VERBS_CMD_RESIZE_CQ] = ib_uverbs_resize_cq,
91 [IB_USER_VERBS_CMD_POLL_CQ] = ib_uverbs_poll_cq,
92 [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ] = ib_uverbs_req_notify_cq,
93 [IB_USER_VERBS_CMD_DESTROY_CQ] = ib_uverbs_destroy_cq,
94 [IB_USER_VERBS_CMD_CREATE_QP] = ib_uverbs_create_qp,
95 [IB_USER_VERBS_CMD_QUERY_QP] = ib_uverbs_query_qp,
96 [IB_USER_VERBS_CMD_MODIFY_QP] = ib_uverbs_modify_qp,
97 [IB_USER_VERBS_CMD_DESTROY_QP] = ib_uverbs_destroy_qp,
98 [IB_USER_VERBS_CMD_POST_SEND] = ib_uverbs_post_send,
99 [IB_USER_VERBS_CMD_POST_RECV] = ib_uverbs_post_recv,
100 [IB_USER_VERBS_CMD_POST_SRQ_RECV] = ib_uverbs_post_srq_recv,
101 [IB_USER_VERBS_CMD_CREATE_AH] = ib_uverbs_create_ah,
102 [IB_USER_VERBS_CMD_DESTROY_AH] = ib_uverbs_destroy_ah,
103 [IB_USER_VERBS_CMD_ATTACH_MCAST] = ib_uverbs_attach_mcast,
104 [IB_USER_VERBS_CMD_DETACH_MCAST] = ib_uverbs_detach_mcast,
105 [IB_USER_VERBS_CMD_CREATE_SRQ] = ib_uverbs_create_srq,
106 [IB_USER_VERBS_CMD_MODIFY_SRQ] = ib_uverbs_modify_srq,
107 [IB_USER_VERBS_CMD_QUERY_SRQ] = ib_uverbs_query_srq,
108 [IB_USER_VERBS_CMD_DESTROY_SRQ] = ib_uverbs_destroy_srq,
109 [IB_USER_VERBS_CMD_OPEN_XRCD] = ib_uverbs_open_xrcd,
110 [IB_USER_VERBS_CMD_CLOSE_XRCD] = ib_uverbs_close_xrcd,
111 [IB_USER_VERBS_CMD_CREATE_XSRQ] = ib_uverbs_create_xsrq,
112 [IB_USER_VERBS_CMD_OPEN_QP] = ib_uverbs_open_qp,
115 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
116 struct ib_device *ib_dev,
117 struct ib_udata *ucore,
118 struct ib_udata *uhw) = {
119 [IB_USER_VERBS_EX_CMD_CREATE_FLOW] = ib_uverbs_ex_create_flow,
120 [IB_USER_VERBS_EX_CMD_DESTROY_FLOW] = ib_uverbs_ex_destroy_flow,
121 [IB_USER_VERBS_EX_CMD_QUERY_DEVICE] = ib_uverbs_ex_query_device,
122 [IB_USER_VERBS_EX_CMD_CREATE_CQ] = ib_uverbs_ex_create_cq,
123 [IB_USER_VERBS_EX_CMD_CREATE_QP] = ib_uverbs_ex_create_qp,
124 [IB_USER_VERBS_EX_CMD_CREATE_WQ] = ib_uverbs_ex_create_wq,
125 [IB_USER_VERBS_EX_CMD_MODIFY_WQ] = ib_uverbs_ex_modify_wq,
126 [IB_USER_VERBS_EX_CMD_DESTROY_WQ] = ib_uverbs_ex_destroy_wq,
127 [IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL] = ib_uverbs_ex_create_rwq_ind_table,
128 [IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL] = ib_uverbs_ex_destroy_rwq_ind_table,
129 [IB_USER_VERBS_EX_CMD_MODIFY_QP] = ib_uverbs_ex_modify_qp,
132 static void ib_uverbs_add_one(struct ib_device *device);
133 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
135 int uverbs_dealloc_mw(struct ib_mw *mw)
137 struct ib_pd *pd = mw->pd;
140 ret = mw->device->dealloc_mw(mw);
142 atomic_dec(&pd->usecnt);
146 static void ib_uverbs_release_dev(struct kobject *kobj)
148 struct ib_uverbs_device *dev =
149 container_of(kobj, struct ib_uverbs_device, kobj);
151 cleanup_srcu_struct(&dev->disassociate_srcu);
155 static struct kobj_type ib_uverbs_dev_ktype = {
156 .release = ib_uverbs_release_dev,
159 static void ib_uverbs_release_async_event_file(struct kref *ref)
161 struct ib_uverbs_async_event_file *file =
162 container_of(ref, struct ib_uverbs_async_event_file, ref);
167 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
168 struct ib_uverbs_completion_event_file *ev_file,
169 struct ib_ucq_object *uobj)
171 struct ib_uverbs_event *evt, *tmp;
174 spin_lock_irq(&ev_file->ev_queue.lock);
175 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
176 list_del(&evt->list);
179 spin_unlock_irq(&ev_file->ev_queue.lock);
181 uverbs_uobject_put(&ev_file->uobj_file.uobj);
184 spin_lock_irq(&file->async_file->ev_queue.lock);
185 list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
186 list_del(&evt->list);
189 spin_unlock_irq(&file->async_file->ev_queue.lock);
192 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
193 struct ib_uevent_object *uobj)
195 struct ib_uverbs_event *evt, *tmp;
197 spin_lock_irq(&file->async_file->ev_queue.lock);
198 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
199 list_del(&evt->list);
202 spin_unlock_irq(&file->async_file->ev_queue.lock);
205 void ib_uverbs_detach_umcast(struct ib_qp *qp,
206 struct ib_uqp_object *uobj)
208 struct ib_uverbs_mcast_entry *mcast, *tmp;
210 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
211 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
212 list_del(&mcast->list);
217 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
218 struct ib_ucontext *context,
221 context->closing = 1;
222 uverbs_cleanup_ucontext(context, device_removed);
223 put_pid(context->tgid);
225 ib_rdmacg_uncharge(&context->cg_obj, context->device,
226 RDMACG_RESOURCE_HCA_HANDLE);
228 return context->device->dealloc_ucontext(context);
231 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
233 complete(&dev->comp);
236 void ib_uverbs_release_file(struct kref *ref)
238 struct ib_uverbs_file *file =
239 container_of(ref, struct ib_uverbs_file, ref);
240 struct ib_device *ib_dev;
243 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
244 ib_dev = srcu_dereference(file->device->ib_dev,
245 &file->device->disassociate_srcu);
246 if (ib_dev && !ib_dev->disassociate_ucontext)
247 module_put(ib_dev->owner);
248 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
250 if (atomic_dec_and_test(&file->device->refcount))
251 ib_uverbs_comp_dev(file->device);
253 kobject_put(&file->device->kobj);
257 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
258 struct ib_uverbs_file *uverbs_file,
259 struct file *filp, char __user *buf,
260 size_t count, loff_t *pos,
263 struct ib_uverbs_event *event;
266 spin_lock_irq(&ev_queue->lock);
268 while (list_empty(&ev_queue->event_list)) {
269 spin_unlock_irq(&ev_queue->lock);
271 if (filp->f_flags & O_NONBLOCK)
274 if (wait_event_interruptible(ev_queue->poll_wait,
275 (!list_empty(&ev_queue->event_list) ||
276 /* The barriers built into wait_event_interruptible()
277 * and wake_up() guarentee this will see the null set
280 !uverbs_file->device->ib_dev)))
283 /* If device was disassociated and no event exists set an error */
284 if (list_empty(&ev_queue->event_list) &&
285 !uverbs_file->device->ib_dev)
288 spin_lock_irq(&ev_queue->lock);
291 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
293 if (eventsz > count) {
297 list_del(ev_queue->event_list.next);
298 if (event->counter) {
300 list_del(&event->obj_list);
304 spin_unlock_irq(&ev_queue->lock);
307 if (copy_to_user(buf, event, eventsz))
318 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
319 size_t count, loff_t *pos)
321 struct ib_uverbs_async_event_file *file = filp->private_data;
323 return ib_uverbs_event_read(&file->ev_queue, file->uverbs_file, filp,
325 sizeof(struct ib_uverbs_async_event_desc));
328 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
329 size_t count, loff_t *pos)
331 struct ib_uverbs_completion_event_file *comp_ev_file =
334 return ib_uverbs_event_read(&comp_ev_file->ev_queue,
335 comp_ev_file->uobj_file.ufile, filp,
337 sizeof(struct ib_uverbs_comp_event_desc));
340 static unsigned int ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
342 struct poll_table_struct *wait)
344 unsigned int pollflags = 0;
346 poll_wait(filp, &ev_queue->poll_wait, wait);
348 spin_lock_irq(&ev_queue->lock);
349 if (!list_empty(&ev_queue->event_list))
350 pollflags = POLLIN | POLLRDNORM;
351 spin_unlock_irq(&ev_queue->lock);
356 static unsigned int ib_uverbs_async_event_poll(struct file *filp,
357 struct poll_table_struct *wait)
359 return ib_uverbs_event_poll(filp->private_data, filp, wait);
362 static unsigned int ib_uverbs_comp_event_poll(struct file *filp,
363 struct poll_table_struct *wait)
365 struct ib_uverbs_completion_event_file *comp_ev_file =
368 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
371 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
373 struct ib_uverbs_event_queue *ev_queue = filp->private_data;
375 return fasync_helper(fd, filp, on, &ev_queue->async_queue);
378 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
380 struct ib_uverbs_completion_event_file *comp_ev_file =
383 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
386 static int ib_uverbs_async_event_close(struct inode *inode, struct file *filp)
388 struct ib_uverbs_async_event_file *file = filp->private_data;
389 struct ib_uverbs_file *uverbs_file = file->uverbs_file;
390 struct ib_uverbs_event *entry, *tmp;
391 int closed_already = 0;
393 mutex_lock(&uverbs_file->device->lists_mutex);
394 spin_lock_irq(&file->ev_queue.lock);
395 closed_already = file->ev_queue.is_closed;
396 file->ev_queue.is_closed = 1;
397 list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
399 list_del(&entry->obj_list);
402 spin_unlock_irq(&file->ev_queue.lock);
403 if (!closed_already) {
404 list_del(&file->list);
405 ib_unregister_event_handler(&uverbs_file->event_handler);
407 mutex_unlock(&uverbs_file->device->lists_mutex);
409 kref_put(&uverbs_file->ref, ib_uverbs_release_file);
410 kref_put(&file->ref, ib_uverbs_release_async_event_file);
415 static int ib_uverbs_comp_event_close(struct inode *inode, struct file *filp)
417 struct ib_uverbs_completion_event_file *file = filp->private_data;
418 struct ib_uverbs_event *entry, *tmp;
420 spin_lock_irq(&file->ev_queue.lock);
421 list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
423 list_del(&entry->obj_list);
426 spin_unlock_irq(&file->ev_queue.lock);
428 uverbs_close_fd(filp);
433 const struct file_operations uverbs_event_fops = {
434 .owner = THIS_MODULE,
435 .read = ib_uverbs_comp_event_read,
436 .poll = ib_uverbs_comp_event_poll,
437 .release = ib_uverbs_comp_event_close,
438 .fasync = ib_uverbs_comp_event_fasync,
442 static const struct file_operations uverbs_async_event_fops = {
443 .owner = THIS_MODULE,
444 .read = ib_uverbs_async_event_read,
445 .poll = ib_uverbs_async_event_poll,
446 .release = ib_uverbs_async_event_close,
447 .fasync = ib_uverbs_async_event_fasync,
451 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
453 struct ib_uverbs_event_queue *ev_queue = cq_context;
454 struct ib_ucq_object *uobj;
455 struct ib_uverbs_event *entry;
461 spin_lock_irqsave(&ev_queue->lock, flags);
462 if (ev_queue->is_closed) {
463 spin_unlock_irqrestore(&ev_queue->lock, flags);
467 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
469 spin_unlock_irqrestore(&ev_queue->lock, flags);
473 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
475 entry->desc.comp.cq_handle = cq->uobject->user_handle;
476 entry->counter = &uobj->comp_events_reported;
478 list_add_tail(&entry->list, &ev_queue->event_list);
479 list_add_tail(&entry->obj_list, &uobj->comp_list);
480 spin_unlock_irqrestore(&ev_queue->lock, flags);
482 wake_up_interruptible(&ev_queue->poll_wait);
483 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
486 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
487 __u64 element, __u64 event,
488 struct list_head *obj_list,
491 struct ib_uverbs_event *entry;
494 spin_lock_irqsave(&file->async_file->ev_queue.lock, flags);
495 if (file->async_file->ev_queue.is_closed) {
496 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
500 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
502 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
506 entry->desc.async.element = element;
507 entry->desc.async.event_type = event;
508 entry->desc.async.reserved = 0;
509 entry->counter = counter;
511 list_add_tail(&entry->list, &file->async_file->ev_queue.event_list);
513 list_add_tail(&entry->obj_list, obj_list);
514 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
516 wake_up_interruptible(&file->async_file->ev_queue.poll_wait);
517 kill_fasync(&file->async_file->ev_queue.async_queue, SIGIO, POLL_IN);
520 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
522 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
523 struct ib_ucq_object, uobject);
525 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
526 event->event, &uobj->async_list,
527 &uobj->async_events_reported);
530 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
532 struct ib_uevent_object *uobj;
534 /* for XRC target qp's, check that qp is live */
535 if (!event->element.qp->uobject)
538 uobj = container_of(event->element.qp->uobject,
539 struct ib_uevent_object, uobject);
541 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
542 event->event, &uobj->event_list,
543 &uobj->events_reported);
546 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
548 struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
549 struct ib_uevent_object, uobject);
551 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
552 event->event, &uobj->event_list,
553 &uobj->events_reported);
556 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
558 struct ib_uevent_object *uobj;
560 uobj = container_of(event->element.srq->uobject,
561 struct ib_uevent_object, uobject);
563 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
564 event->event, &uobj->event_list,
565 &uobj->events_reported);
568 void ib_uverbs_event_handler(struct ib_event_handler *handler,
569 struct ib_event *event)
571 struct ib_uverbs_file *file =
572 container_of(handler, struct ib_uverbs_file, event_handler);
574 ib_uverbs_async_handler(file, event->element.port_num, event->event,
578 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
580 kref_put(&file->async_file->ref, ib_uverbs_release_async_event_file);
581 file->async_file = NULL;
584 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
586 spin_lock_init(&ev_queue->lock);
587 INIT_LIST_HEAD(&ev_queue->event_list);
588 init_waitqueue_head(&ev_queue->poll_wait);
589 ev_queue->is_closed = 0;
590 ev_queue->async_queue = NULL;
593 struct file *ib_uverbs_alloc_async_event_file(struct ib_uverbs_file *uverbs_file,
594 struct ib_device *ib_dev)
596 struct ib_uverbs_async_event_file *ev_file;
600 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
602 return ERR_PTR(-ENOMEM);
604 ib_uverbs_init_event_queue(&ev_file->ev_queue);
605 ev_file->uverbs_file = uverbs_file;
606 kref_get(&ev_file->uverbs_file->ref);
607 kref_init(&ev_file->ref);
608 filp = anon_inode_getfile("[infinibandevent]", &uverbs_async_event_fops,
613 mutex_lock(&uverbs_file->device->lists_mutex);
614 list_add_tail(&ev_file->list,
615 &uverbs_file->device->uverbs_events_file_list);
616 mutex_unlock(&uverbs_file->device->lists_mutex);
618 WARN_ON(uverbs_file->async_file);
619 uverbs_file->async_file = ev_file;
620 kref_get(&uverbs_file->async_file->ref);
621 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
623 ib_uverbs_event_handler);
624 ret = ib_register_event_handler(&uverbs_file->event_handler);
628 /* At that point async file stuff was fully set */
634 kref_put(&uverbs_file->async_file->ref,
635 ib_uverbs_release_async_event_file);
636 uverbs_file->async_file = NULL;
640 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
641 kref_put(&ev_file->ref, ib_uverbs_release_async_event_file);
645 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
649 if (command <= IB_USER_VERBS_CMD_OPEN_QP)
650 mask = ib_dev->uverbs_cmd_mask;
652 mask = ib_dev->uverbs_ex_cmd_mask;
654 if (mask & ((u64)1 << command))
660 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
661 size_t count, loff_t *pos)
663 struct ib_uverbs_file *file = filp->private_data;
664 struct ib_device *ib_dev;
665 struct ib_uverbs_cmd_hdr hdr;
671 if (!ib_safe_file_access(filp)) {
672 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
673 task_tgid_vnr(current), current->comm);
677 if (count < sizeof hdr)
680 if (copy_from_user(&hdr, buf, sizeof hdr))
683 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
684 ib_dev = srcu_dereference(file->device->ib_dev,
685 &file->device->disassociate_srcu);
691 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
692 IB_USER_VERBS_CMD_COMMAND_MASK)) {
697 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
698 if (verify_command_mask(ib_dev, command)) {
703 if (!file->ucontext &&
704 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
709 flags = (hdr.command &
710 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
713 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
714 !uverbs_cmd_table[command]) {
719 if (hdr.in_words * 4 != count) {
724 ret = uverbs_cmd_table[command](file, ib_dev,
729 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
730 struct ib_uverbs_ex_cmd_hdr ex_hdr;
731 struct ib_udata ucore;
733 size_t written_count = count;
735 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
736 !uverbs_ex_cmd_table[command]) {
741 if (!file->ucontext) {
746 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
751 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
756 count -= sizeof(hdr) + sizeof(ex_hdr);
757 buf += sizeof(hdr) + sizeof(ex_hdr);
759 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
764 if (ex_hdr.cmd_hdr_reserved) {
769 if (ex_hdr.response) {
770 if (!hdr.out_words && !ex_hdr.provider_out_words) {
775 if (!access_ok(VERIFY_WRITE,
776 (void __user *) (unsigned long) ex_hdr.response,
777 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
782 if (hdr.out_words || ex_hdr.provider_out_words) {
788 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
789 hdr.in_words * 8, hdr.out_words * 8);
791 INIT_UDATA_BUF_OR_NULL(&uhw,
793 (unsigned long) ex_hdr.response + ucore.outlen,
794 ex_hdr.provider_in_words * 8,
795 ex_hdr.provider_out_words * 8);
797 ret = uverbs_ex_cmd_table[command](file,
808 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
812 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
814 struct ib_uverbs_file *file = filp->private_data;
815 struct ib_device *ib_dev;
819 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
820 ib_dev = srcu_dereference(file->device->ib_dev,
821 &file->device->disassociate_srcu);
830 ret = ib_dev->mmap(file->ucontext, vma);
832 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
837 * ib_uverbs_open() does not need the BKL:
839 * - the ib_uverbs_device structures are properly reference counted and
840 * everything else is purely local to the file being created, so
841 * races against other open calls are not a problem;
842 * - there is no ioctl method to race against;
843 * - the open method will either immediately run -ENXIO, or all
844 * required initialization will be done.
846 static int ib_uverbs_open(struct inode *inode, struct file *filp)
848 struct ib_uverbs_device *dev;
849 struct ib_uverbs_file *file;
850 struct ib_device *ib_dev;
852 int module_dependent;
855 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
856 if (!atomic_inc_not_zero(&dev->refcount))
859 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
860 mutex_lock(&dev->lists_mutex);
861 ib_dev = srcu_dereference(dev->ib_dev,
862 &dev->disassociate_srcu);
868 /* In case IB device supports disassociate ucontext, there is no hard
869 * dependency between uverbs device and its low level device.
871 module_dependent = !(ib_dev->disassociate_ucontext);
873 if (module_dependent) {
874 if (!try_module_get(ib_dev->owner)) {
880 file = kzalloc(sizeof(*file), GFP_KERNEL);
883 if (module_dependent)
890 spin_lock_init(&file->idr_lock);
891 idr_init(&file->idr);
892 file->ucontext = NULL;
893 file->async_file = NULL;
894 kref_init(&file->ref);
895 mutex_init(&file->mutex);
896 mutex_init(&file->cleanup_mutex);
898 filp->private_data = file;
899 kobject_get(&dev->kobj);
900 list_add_tail(&file->list, &dev->uverbs_file_list);
901 mutex_unlock(&dev->lists_mutex);
902 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
904 return nonseekable_open(inode, filp);
907 module_put(ib_dev->owner);
910 mutex_unlock(&dev->lists_mutex);
911 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
912 if (atomic_dec_and_test(&dev->refcount))
913 ib_uverbs_comp_dev(dev);
918 static int ib_uverbs_close(struct inode *inode, struct file *filp)
920 struct ib_uverbs_file *file = filp->private_data;
922 mutex_lock(&file->cleanup_mutex);
923 if (file->ucontext) {
924 ib_uverbs_cleanup_ucontext(file, file->ucontext, false);
925 file->ucontext = NULL;
927 mutex_unlock(&file->cleanup_mutex);
928 idr_destroy(&file->idr);
930 mutex_lock(&file->device->lists_mutex);
931 if (!file->is_closed) {
932 list_del(&file->list);
935 mutex_unlock(&file->device->lists_mutex);
937 if (file->async_file)
938 kref_put(&file->async_file->ref,
939 ib_uverbs_release_async_event_file);
941 kref_put(&file->ref, ib_uverbs_release_file);
946 static const struct file_operations uverbs_fops = {
947 .owner = THIS_MODULE,
948 .write = ib_uverbs_write,
949 .open = ib_uverbs_open,
950 .release = ib_uverbs_close,
954 static const struct file_operations uverbs_mmap_fops = {
955 .owner = THIS_MODULE,
956 .write = ib_uverbs_write,
957 .mmap = ib_uverbs_mmap,
958 .open = ib_uverbs_open,
959 .release = ib_uverbs_close,
963 static struct ib_client uverbs_client = {
965 .add = ib_uverbs_add_one,
966 .remove = ib_uverbs_remove_one
969 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
974 struct ib_uverbs_device *dev = dev_get_drvdata(device);
975 struct ib_device *ib_dev;
980 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
981 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
983 ret = sprintf(buf, "%s\n", ib_dev->name);
984 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
988 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
990 static ssize_t show_dev_abi_version(struct device *device,
991 struct device_attribute *attr, char *buf)
993 struct ib_uverbs_device *dev = dev_get_drvdata(device);
996 struct ib_device *ib_dev;
1000 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1001 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1003 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1004 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1008 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1010 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1011 __stringify(IB_USER_VERBS_ABI_VERSION));
1013 static dev_t overflow_maj;
1014 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1017 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1018 * requesting a new major number and doubling the number of max devices we
1019 * support. It's stupid, but simple.
1021 static int find_overflow_devnum(void)
1025 if (!overflow_maj) {
1026 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1027 "infiniband_verbs");
1029 pr_err("user_verbs: couldn't register dynamic device number\n");
1034 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1035 if (ret >= IB_UVERBS_MAX_DEVICES)
1041 static void ib_uverbs_add_one(struct ib_device *device)
1045 struct ib_uverbs_device *uverbs_dev;
1048 if (!device->alloc_ucontext)
1051 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1055 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1061 atomic_set(&uverbs_dev->refcount, 1);
1062 init_completion(&uverbs_dev->comp);
1063 uverbs_dev->xrcd_tree = RB_ROOT;
1064 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1065 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1066 mutex_init(&uverbs_dev->lists_mutex);
1067 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1068 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1070 spin_lock(&map_lock);
1071 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1072 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1073 spin_unlock(&map_lock);
1074 devnum = find_overflow_devnum();
1078 spin_lock(&map_lock);
1079 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1080 base = devnum + overflow_maj;
1081 set_bit(devnum, overflow_map);
1083 uverbs_dev->devnum = devnum;
1084 base = devnum + IB_UVERBS_BASE_DEV;
1085 set_bit(devnum, dev_map);
1087 spin_unlock(&map_lock);
1089 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1090 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1092 cdev_init(&uverbs_dev->cdev, NULL);
1093 uverbs_dev->cdev.owner = THIS_MODULE;
1094 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1095 cdev_set_parent(&uverbs_dev->cdev, &uverbs_dev->kobj);
1096 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1097 if (cdev_add(&uverbs_dev->cdev, base, 1))
1100 uverbs_dev->dev = device_create(uverbs_class, device->dev.parent,
1101 uverbs_dev->cdev.dev, uverbs_dev,
1102 "uverbs%d", uverbs_dev->devnum);
1103 if (IS_ERR(uverbs_dev->dev))
1106 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1108 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1111 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1116 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1119 cdev_del(&uverbs_dev->cdev);
1120 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1121 clear_bit(devnum, dev_map);
1123 clear_bit(devnum, overflow_map);
1126 if (atomic_dec_and_test(&uverbs_dev->refcount))
1127 ib_uverbs_comp_dev(uverbs_dev);
1128 wait_for_completion(&uverbs_dev->comp);
1129 kobject_put(&uverbs_dev->kobj);
1133 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1134 struct ib_device *ib_dev)
1136 struct ib_uverbs_file *file;
1137 struct ib_uverbs_async_event_file *event_file;
1138 struct ib_event event;
1140 /* Pending running commands to terminate */
1141 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1142 event.event = IB_EVENT_DEVICE_FATAL;
1143 event.element.port_num = 0;
1144 event.device = ib_dev;
1146 mutex_lock(&uverbs_dev->lists_mutex);
1147 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1148 struct ib_ucontext *ucontext;
1149 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1150 struct ib_uverbs_file, list);
1151 file->is_closed = 1;
1152 list_del(&file->list);
1153 kref_get(&file->ref);
1154 mutex_unlock(&uverbs_dev->lists_mutex);
1157 mutex_lock(&file->cleanup_mutex);
1158 ucontext = file->ucontext;
1159 file->ucontext = NULL;
1160 mutex_unlock(&file->cleanup_mutex);
1162 /* At this point ib_uverbs_close cannot be running
1163 * ib_uverbs_cleanup_ucontext
1166 /* We must release the mutex before going ahead and
1167 * calling disassociate_ucontext. disassociate_ucontext
1168 * might end up indirectly calling uverbs_close,
1169 * for example due to freeing the resources
1172 ib_uverbs_event_handler(&file->event_handler, &event);
1173 ib_dev->disassociate_ucontext(ucontext);
1174 mutex_lock(&file->cleanup_mutex);
1175 ib_uverbs_cleanup_ucontext(file, ucontext, true);
1176 mutex_unlock(&file->cleanup_mutex);
1179 mutex_lock(&uverbs_dev->lists_mutex);
1180 kref_put(&file->ref, ib_uverbs_release_file);
1183 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1184 event_file = list_first_entry(&uverbs_dev->
1185 uverbs_events_file_list,
1186 struct ib_uverbs_async_event_file,
1188 spin_lock_irq(&event_file->ev_queue.lock);
1189 event_file->ev_queue.is_closed = 1;
1190 spin_unlock_irq(&event_file->ev_queue.lock);
1192 list_del(&event_file->list);
1193 ib_unregister_event_handler(
1194 &event_file->uverbs_file->event_handler);
1195 event_file->uverbs_file->event_handler.device =
1198 wake_up_interruptible(&event_file->ev_queue.poll_wait);
1199 kill_fasync(&event_file->ev_queue.async_queue, SIGIO, POLL_IN);
1201 mutex_unlock(&uverbs_dev->lists_mutex);
1204 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1206 struct ib_uverbs_device *uverbs_dev = client_data;
1207 int wait_clients = 1;
1212 dev_set_drvdata(uverbs_dev->dev, NULL);
1213 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1214 cdev_del(&uverbs_dev->cdev);
1216 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1217 clear_bit(uverbs_dev->devnum, dev_map);
1219 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1221 if (device->disassociate_ucontext) {
1222 /* We disassociate HW resources and immediately return.
1223 * Userspace will see a EIO errno for all future access.
1224 * Upon returning, ib_device may be freed internally and is not
1226 * uverbs_device is still available until all clients close
1227 * their files, then the uverbs device ref count will be zero
1228 * and its resources will be freed.
1229 * Note: At this point no more files can be opened since the
1230 * cdev was deleted, however active clients can still issue
1231 * commands and close their open files.
1233 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1234 ib_uverbs_free_hw_resources(uverbs_dev, device);
1238 if (atomic_dec_and_test(&uverbs_dev->refcount))
1239 ib_uverbs_comp_dev(uverbs_dev);
1241 wait_for_completion(&uverbs_dev->comp);
1242 kobject_put(&uverbs_dev->kobj);
1245 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1249 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1252 static int __init ib_uverbs_init(void)
1256 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1257 "infiniband_verbs");
1259 pr_err("user_verbs: couldn't register device number\n");
1263 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1264 if (IS_ERR(uverbs_class)) {
1265 ret = PTR_ERR(uverbs_class);
1266 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1270 uverbs_class->devnode = uverbs_devnode;
1272 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1274 pr_err("user_verbs: couldn't create abi_version attribute\n");
1278 ret = ib_register_client(&uverbs_client);
1280 pr_err("user_verbs: couldn't register client\n");
1287 class_destroy(uverbs_class);
1290 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1296 static void __exit ib_uverbs_cleanup(void)
1298 ib_unregister_client(&uverbs_client);
1299 class_destroy(uverbs_class);
1300 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1302 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1305 module_init(ib_uverbs_init);
1306 module_exit(ib_uverbs_cleanup);