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>
52 #include <rdma/uverbs_std_types.h>
55 #include "core_priv.h"
56 #include "rdma_core.h"
58 MODULE_AUTHOR("Roland Dreier");
59 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
60 MODULE_LICENSE("Dual BSD/GPL");
63 IB_UVERBS_MAJOR = 231,
64 IB_UVERBS_BASE_MINOR = 192,
65 IB_UVERBS_MAX_DEVICES = 32
68 #define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
70 static struct class *uverbs_class;
72 static DEFINE_SPINLOCK(map_lock);
73 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
75 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
76 struct ib_device *ib_dev,
77 const char __user *buf, int in_len,
79 [IB_USER_VERBS_CMD_GET_CONTEXT] = ib_uverbs_get_context,
80 [IB_USER_VERBS_CMD_QUERY_DEVICE] = ib_uverbs_query_device,
81 [IB_USER_VERBS_CMD_QUERY_PORT] = ib_uverbs_query_port,
82 [IB_USER_VERBS_CMD_ALLOC_PD] = ib_uverbs_alloc_pd,
83 [IB_USER_VERBS_CMD_DEALLOC_PD] = ib_uverbs_dealloc_pd,
84 [IB_USER_VERBS_CMD_REG_MR] = ib_uverbs_reg_mr,
85 [IB_USER_VERBS_CMD_REREG_MR] = ib_uverbs_rereg_mr,
86 [IB_USER_VERBS_CMD_DEREG_MR] = ib_uverbs_dereg_mr,
87 [IB_USER_VERBS_CMD_ALLOC_MW] = ib_uverbs_alloc_mw,
88 [IB_USER_VERBS_CMD_DEALLOC_MW] = ib_uverbs_dealloc_mw,
89 [IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
90 [IB_USER_VERBS_CMD_CREATE_CQ] = ib_uverbs_create_cq,
91 [IB_USER_VERBS_CMD_RESIZE_CQ] = ib_uverbs_resize_cq,
92 [IB_USER_VERBS_CMD_POLL_CQ] = ib_uverbs_poll_cq,
93 [IB_USER_VERBS_CMD_REQ_NOTIFY_CQ] = ib_uverbs_req_notify_cq,
94 [IB_USER_VERBS_CMD_DESTROY_CQ] = ib_uverbs_destroy_cq,
95 [IB_USER_VERBS_CMD_CREATE_QP] = ib_uverbs_create_qp,
96 [IB_USER_VERBS_CMD_QUERY_QP] = ib_uverbs_query_qp,
97 [IB_USER_VERBS_CMD_MODIFY_QP] = ib_uverbs_modify_qp,
98 [IB_USER_VERBS_CMD_DESTROY_QP] = ib_uverbs_destroy_qp,
99 [IB_USER_VERBS_CMD_POST_SEND] = ib_uverbs_post_send,
100 [IB_USER_VERBS_CMD_POST_RECV] = ib_uverbs_post_recv,
101 [IB_USER_VERBS_CMD_POST_SRQ_RECV] = ib_uverbs_post_srq_recv,
102 [IB_USER_VERBS_CMD_CREATE_AH] = ib_uverbs_create_ah,
103 [IB_USER_VERBS_CMD_DESTROY_AH] = ib_uverbs_destroy_ah,
104 [IB_USER_VERBS_CMD_ATTACH_MCAST] = ib_uverbs_attach_mcast,
105 [IB_USER_VERBS_CMD_DETACH_MCAST] = ib_uverbs_detach_mcast,
106 [IB_USER_VERBS_CMD_CREATE_SRQ] = ib_uverbs_create_srq,
107 [IB_USER_VERBS_CMD_MODIFY_SRQ] = ib_uverbs_modify_srq,
108 [IB_USER_VERBS_CMD_QUERY_SRQ] = ib_uverbs_query_srq,
109 [IB_USER_VERBS_CMD_DESTROY_SRQ] = ib_uverbs_destroy_srq,
110 [IB_USER_VERBS_CMD_OPEN_XRCD] = ib_uverbs_open_xrcd,
111 [IB_USER_VERBS_CMD_CLOSE_XRCD] = ib_uverbs_close_xrcd,
112 [IB_USER_VERBS_CMD_CREATE_XSRQ] = ib_uverbs_create_xsrq,
113 [IB_USER_VERBS_CMD_OPEN_QP] = ib_uverbs_open_qp,
116 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
117 struct ib_device *ib_dev,
118 struct ib_udata *ucore,
119 struct ib_udata *uhw) = {
120 [IB_USER_VERBS_EX_CMD_CREATE_FLOW] = ib_uverbs_ex_create_flow,
121 [IB_USER_VERBS_EX_CMD_DESTROY_FLOW] = ib_uverbs_ex_destroy_flow,
122 [IB_USER_VERBS_EX_CMD_QUERY_DEVICE] = ib_uverbs_ex_query_device,
123 [IB_USER_VERBS_EX_CMD_CREATE_CQ] = ib_uverbs_ex_create_cq,
124 [IB_USER_VERBS_EX_CMD_CREATE_QP] = ib_uverbs_ex_create_qp,
125 [IB_USER_VERBS_EX_CMD_CREATE_WQ] = ib_uverbs_ex_create_wq,
126 [IB_USER_VERBS_EX_CMD_MODIFY_WQ] = ib_uverbs_ex_modify_wq,
127 [IB_USER_VERBS_EX_CMD_DESTROY_WQ] = ib_uverbs_ex_destroy_wq,
128 [IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL] = ib_uverbs_ex_create_rwq_ind_table,
129 [IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL] = ib_uverbs_ex_destroy_rwq_ind_table,
130 [IB_USER_VERBS_EX_CMD_MODIFY_QP] = ib_uverbs_ex_modify_qp,
133 static void ib_uverbs_add_one(struct ib_device *device);
134 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
136 int uverbs_dealloc_mw(struct ib_mw *mw)
138 struct ib_pd *pd = mw->pd;
141 ret = mw->device->dealloc_mw(mw);
143 atomic_dec(&pd->usecnt);
147 static void ib_uverbs_release_dev(struct kobject *kobj)
149 struct ib_uverbs_device *dev =
150 container_of(kobj, struct ib_uverbs_device, kobj);
152 cleanup_srcu_struct(&dev->disassociate_srcu);
156 static struct kobj_type ib_uverbs_dev_ktype = {
157 .release = ib_uverbs_release_dev,
160 static void ib_uverbs_release_async_event_file(struct kref *ref)
162 struct ib_uverbs_async_event_file *file =
163 container_of(ref, struct ib_uverbs_async_event_file, ref);
168 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
169 struct ib_uverbs_completion_event_file *ev_file,
170 struct ib_ucq_object *uobj)
172 struct ib_uverbs_event *evt, *tmp;
175 spin_lock_irq(&ev_file->ev_queue.lock);
176 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
177 list_del(&evt->list);
180 spin_unlock_irq(&ev_file->ev_queue.lock);
182 uverbs_uobject_put(&ev_file->uobj_file.uobj);
185 spin_lock_irq(&file->async_file->ev_queue.lock);
186 list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
187 list_del(&evt->list);
190 spin_unlock_irq(&file->async_file->ev_queue.lock);
193 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
194 struct ib_uevent_object *uobj)
196 struct ib_uverbs_event *evt, *tmp;
198 spin_lock_irq(&file->async_file->ev_queue.lock);
199 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
200 list_del(&evt->list);
203 spin_unlock_irq(&file->async_file->ev_queue.lock);
206 void ib_uverbs_detach_umcast(struct ib_qp *qp,
207 struct ib_uqp_object *uobj)
209 struct ib_uverbs_mcast_entry *mcast, *tmp;
211 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
212 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
213 list_del(&mcast->list);
218 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
219 struct ib_ucontext *context,
222 context->closing = 1;
223 uverbs_cleanup_ucontext(context, device_removed);
224 put_pid(context->tgid);
226 ib_rdmacg_uncharge(&context->cg_obj, context->device,
227 RDMACG_RESOURCE_HCA_HANDLE);
229 return context->device->dealloc_ucontext(context);
232 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
234 complete(&dev->comp);
237 void ib_uverbs_release_file(struct kref *ref)
239 struct ib_uverbs_file *file =
240 container_of(ref, struct ib_uverbs_file, ref);
241 struct ib_device *ib_dev;
244 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
245 ib_dev = srcu_dereference(file->device->ib_dev,
246 &file->device->disassociate_srcu);
247 if (ib_dev && !ib_dev->disassociate_ucontext)
248 module_put(ib_dev->owner);
249 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
251 if (atomic_dec_and_test(&file->device->refcount))
252 ib_uverbs_comp_dev(file->device);
254 kobject_put(&file->device->kobj);
258 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
259 struct ib_uverbs_file *uverbs_file,
260 struct file *filp, char __user *buf,
261 size_t count, loff_t *pos,
264 struct ib_uverbs_event *event;
267 spin_lock_irq(&ev_queue->lock);
269 while (list_empty(&ev_queue->event_list)) {
270 spin_unlock_irq(&ev_queue->lock);
272 if (filp->f_flags & O_NONBLOCK)
275 if (wait_event_interruptible(ev_queue->poll_wait,
276 (!list_empty(&ev_queue->event_list) ||
277 /* The barriers built into wait_event_interruptible()
278 * and wake_up() guarentee this will see the null set
281 !uverbs_file->device->ib_dev)))
284 /* If device was disassociated and no event exists set an error */
285 if (list_empty(&ev_queue->event_list) &&
286 !uverbs_file->device->ib_dev)
289 spin_lock_irq(&ev_queue->lock);
292 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
294 if (eventsz > count) {
298 list_del(ev_queue->event_list.next);
299 if (event->counter) {
301 list_del(&event->obj_list);
305 spin_unlock_irq(&ev_queue->lock);
308 if (copy_to_user(buf, event, eventsz))
319 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
320 size_t count, loff_t *pos)
322 struct ib_uverbs_async_event_file *file = filp->private_data;
324 return ib_uverbs_event_read(&file->ev_queue, file->uverbs_file, filp,
326 sizeof(struct ib_uverbs_async_event_desc));
329 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
330 size_t count, loff_t *pos)
332 struct ib_uverbs_completion_event_file *comp_ev_file =
335 return ib_uverbs_event_read(&comp_ev_file->ev_queue,
336 comp_ev_file->uobj_file.ufile, filp,
338 sizeof(struct ib_uverbs_comp_event_desc));
341 static unsigned int ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
343 struct poll_table_struct *wait)
345 unsigned int pollflags = 0;
347 poll_wait(filp, &ev_queue->poll_wait, wait);
349 spin_lock_irq(&ev_queue->lock);
350 if (!list_empty(&ev_queue->event_list))
351 pollflags = POLLIN | POLLRDNORM;
352 spin_unlock_irq(&ev_queue->lock);
357 static unsigned int ib_uverbs_async_event_poll(struct file *filp,
358 struct poll_table_struct *wait)
360 return ib_uverbs_event_poll(filp->private_data, filp, wait);
363 static unsigned int ib_uverbs_comp_event_poll(struct file *filp,
364 struct poll_table_struct *wait)
366 struct ib_uverbs_completion_event_file *comp_ev_file =
369 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
372 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
374 struct ib_uverbs_event_queue *ev_queue = filp->private_data;
376 return fasync_helper(fd, filp, on, &ev_queue->async_queue);
379 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
381 struct ib_uverbs_completion_event_file *comp_ev_file =
384 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
387 static int ib_uverbs_async_event_close(struct inode *inode, struct file *filp)
389 struct ib_uverbs_async_event_file *file = filp->private_data;
390 struct ib_uverbs_file *uverbs_file = file->uverbs_file;
391 struct ib_uverbs_event *entry, *tmp;
392 int closed_already = 0;
394 mutex_lock(&uverbs_file->device->lists_mutex);
395 spin_lock_irq(&file->ev_queue.lock);
396 closed_already = file->ev_queue.is_closed;
397 file->ev_queue.is_closed = 1;
398 list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
400 list_del(&entry->obj_list);
403 spin_unlock_irq(&file->ev_queue.lock);
404 if (!closed_already) {
405 list_del(&file->list);
406 ib_unregister_event_handler(&uverbs_file->event_handler);
408 mutex_unlock(&uverbs_file->device->lists_mutex);
410 kref_put(&uverbs_file->ref, ib_uverbs_release_file);
411 kref_put(&file->ref, ib_uverbs_release_async_event_file);
416 static int ib_uverbs_comp_event_close(struct inode *inode, struct file *filp)
418 struct ib_uverbs_completion_event_file *file = filp->private_data;
419 struct ib_uverbs_event *entry, *tmp;
421 spin_lock_irq(&file->ev_queue.lock);
422 list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
424 list_del(&entry->obj_list);
427 file->ev_queue.is_closed = 1;
428 spin_unlock_irq(&file->ev_queue.lock);
430 uverbs_close_fd(filp);
435 const struct file_operations uverbs_event_fops = {
436 .owner = THIS_MODULE,
437 .read = ib_uverbs_comp_event_read,
438 .poll = ib_uverbs_comp_event_poll,
439 .release = ib_uverbs_comp_event_close,
440 .fasync = ib_uverbs_comp_event_fasync,
444 static const struct file_operations uverbs_async_event_fops = {
445 .owner = THIS_MODULE,
446 .read = ib_uverbs_async_event_read,
447 .poll = ib_uverbs_async_event_poll,
448 .release = ib_uverbs_async_event_close,
449 .fasync = ib_uverbs_async_event_fasync,
453 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
455 struct ib_uverbs_event_queue *ev_queue = cq_context;
456 struct ib_ucq_object *uobj;
457 struct ib_uverbs_event *entry;
463 spin_lock_irqsave(&ev_queue->lock, flags);
464 if (ev_queue->is_closed) {
465 spin_unlock_irqrestore(&ev_queue->lock, flags);
469 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
471 spin_unlock_irqrestore(&ev_queue->lock, flags);
475 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
477 entry->desc.comp.cq_handle = cq->uobject->user_handle;
478 entry->counter = &uobj->comp_events_reported;
480 list_add_tail(&entry->list, &ev_queue->event_list);
481 list_add_tail(&entry->obj_list, &uobj->comp_list);
482 spin_unlock_irqrestore(&ev_queue->lock, flags);
484 wake_up_interruptible(&ev_queue->poll_wait);
485 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
488 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
489 __u64 element, __u64 event,
490 struct list_head *obj_list,
493 struct ib_uverbs_event *entry;
496 spin_lock_irqsave(&file->async_file->ev_queue.lock, flags);
497 if (file->async_file->ev_queue.is_closed) {
498 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
502 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
504 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
508 entry->desc.async.element = element;
509 entry->desc.async.event_type = event;
510 entry->desc.async.reserved = 0;
511 entry->counter = counter;
513 list_add_tail(&entry->list, &file->async_file->ev_queue.event_list);
515 list_add_tail(&entry->obj_list, obj_list);
516 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
518 wake_up_interruptible(&file->async_file->ev_queue.poll_wait);
519 kill_fasync(&file->async_file->ev_queue.async_queue, SIGIO, POLL_IN);
522 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
524 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
525 struct ib_ucq_object, uobject);
527 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
528 event->event, &uobj->async_list,
529 &uobj->async_events_reported);
532 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
534 struct ib_uevent_object *uobj;
536 /* for XRC target qp's, check that qp is live */
537 if (!event->element.qp->uobject)
540 uobj = container_of(event->element.qp->uobject,
541 struct ib_uevent_object, uobject);
543 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
544 event->event, &uobj->event_list,
545 &uobj->events_reported);
548 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
550 struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
551 struct ib_uevent_object, uobject);
553 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
554 event->event, &uobj->event_list,
555 &uobj->events_reported);
558 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
560 struct ib_uevent_object *uobj;
562 uobj = container_of(event->element.srq->uobject,
563 struct ib_uevent_object, uobject);
565 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
566 event->event, &uobj->event_list,
567 &uobj->events_reported);
570 void ib_uverbs_event_handler(struct ib_event_handler *handler,
571 struct ib_event *event)
573 struct ib_uverbs_file *file =
574 container_of(handler, struct ib_uverbs_file, event_handler);
576 ib_uverbs_async_handler(file, event->element.port_num, event->event,
580 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
582 kref_put(&file->async_file->ref, ib_uverbs_release_async_event_file);
583 file->async_file = NULL;
586 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
588 spin_lock_init(&ev_queue->lock);
589 INIT_LIST_HEAD(&ev_queue->event_list);
590 init_waitqueue_head(&ev_queue->poll_wait);
591 ev_queue->is_closed = 0;
592 ev_queue->async_queue = NULL;
595 struct file *ib_uverbs_alloc_async_event_file(struct ib_uverbs_file *uverbs_file,
596 struct ib_device *ib_dev)
598 struct ib_uverbs_async_event_file *ev_file;
601 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
603 return ERR_PTR(-ENOMEM);
605 ib_uverbs_init_event_queue(&ev_file->ev_queue);
606 ev_file->uverbs_file = uverbs_file;
607 kref_get(&ev_file->uverbs_file->ref);
608 kref_init(&ev_file->ref);
609 filp = anon_inode_getfile("[infinibandevent]", &uverbs_async_event_fops,
614 mutex_lock(&uverbs_file->device->lists_mutex);
615 list_add_tail(&ev_file->list,
616 &uverbs_file->device->uverbs_events_file_list);
617 mutex_unlock(&uverbs_file->device->lists_mutex);
619 WARN_ON(uverbs_file->async_file);
620 uverbs_file->async_file = ev_file;
621 kref_get(&uverbs_file->async_file->ref);
622 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
624 ib_uverbs_event_handler);
625 ib_register_event_handler(&uverbs_file->event_handler);
626 /* At that point async file stuff was fully set */
631 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
632 kref_put(&ev_file->ref, ib_uverbs_release_async_event_file);
636 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
640 if (command <= IB_USER_VERBS_CMD_OPEN_QP)
641 mask = ib_dev->uverbs_cmd_mask;
643 mask = ib_dev->uverbs_ex_cmd_mask;
645 if (mask & ((u64)1 << command))
651 static bool verify_command_idx(u32 command, bool extended)
654 return command < ARRAY_SIZE(uverbs_ex_cmd_table);
656 return command < ARRAY_SIZE(uverbs_cmd_table);
659 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
660 size_t count, loff_t *pos)
662 struct ib_uverbs_file *file = filp->private_data;
663 struct ib_device *ib_dev;
664 struct ib_uverbs_cmd_hdr hdr;
665 bool extended_command;
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 flags = (hdr.command &
699 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
701 extended_command = flags & IB_USER_VERBS_CMD_FLAG_EXTENDED;
702 if (!verify_command_idx(command, extended_command)) {
707 if (verify_command_mask(ib_dev, command)) {
712 if (!file->ucontext &&
713 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
719 if (!uverbs_cmd_table[command]) {
724 if (hdr.in_words * 4 != count) {
729 ret = uverbs_cmd_table[command](file, ib_dev,
734 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
735 struct ib_uverbs_ex_cmd_hdr ex_hdr;
736 struct ib_udata ucore;
738 size_t written_count = count;
740 if (!uverbs_ex_cmd_table[command]) {
745 if (!file->ucontext) {
750 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
755 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
760 count -= sizeof(hdr) + sizeof(ex_hdr);
761 buf += sizeof(hdr) + sizeof(ex_hdr);
763 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
768 if (ex_hdr.cmd_hdr_reserved) {
773 if (ex_hdr.response) {
774 if (!hdr.out_words && !ex_hdr.provider_out_words) {
779 if (!access_ok(VERIFY_WRITE,
780 (void __user *) (unsigned long) ex_hdr.response,
781 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
786 if (hdr.out_words || ex_hdr.provider_out_words) {
792 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
793 hdr.in_words * 8, hdr.out_words * 8);
795 INIT_UDATA_BUF_OR_NULL(&uhw,
797 (unsigned long) ex_hdr.response + ucore.outlen,
798 ex_hdr.provider_in_words * 8,
799 ex_hdr.provider_out_words * 8);
801 ret = uverbs_ex_cmd_table[command](file,
812 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
816 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
818 struct ib_uverbs_file *file = filp->private_data;
819 struct ib_device *ib_dev;
823 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
824 ib_dev = srcu_dereference(file->device->ib_dev,
825 &file->device->disassociate_srcu);
834 ret = ib_dev->mmap(file->ucontext, vma);
836 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
841 * ib_uverbs_open() does not need the BKL:
843 * - the ib_uverbs_device structures are properly reference counted and
844 * everything else is purely local to the file being created, so
845 * races against other open calls are not a problem;
846 * - there is no ioctl method to race against;
847 * - the open method will either immediately run -ENXIO, or all
848 * required initialization will be done.
850 static int ib_uverbs_open(struct inode *inode, struct file *filp)
852 struct ib_uverbs_device *dev;
853 struct ib_uverbs_file *file;
854 struct ib_device *ib_dev;
856 int module_dependent;
859 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
860 if (!atomic_inc_not_zero(&dev->refcount))
863 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
864 mutex_lock(&dev->lists_mutex);
865 ib_dev = srcu_dereference(dev->ib_dev,
866 &dev->disassociate_srcu);
872 /* In case IB device supports disassociate ucontext, there is no hard
873 * dependency between uverbs device and its low level device.
875 module_dependent = !(ib_dev->disassociate_ucontext);
877 if (module_dependent) {
878 if (!try_module_get(ib_dev->owner)) {
884 file = kzalloc(sizeof(*file), GFP_KERNEL);
887 if (module_dependent)
894 spin_lock_init(&file->idr_lock);
895 idr_init(&file->idr);
896 file->ucontext = NULL;
897 file->async_file = NULL;
898 kref_init(&file->ref);
899 mutex_init(&file->mutex);
900 mutex_init(&file->cleanup_mutex);
902 filp->private_data = file;
903 kobject_get(&dev->kobj);
904 list_add_tail(&file->list, &dev->uverbs_file_list);
905 mutex_unlock(&dev->lists_mutex);
906 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
908 return nonseekable_open(inode, filp);
911 module_put(ib_dev->owner);
914 mutex_unlock(&dev->lists_mutex);
915 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
916 if (atomic_dec_and_test(&dev->refcount))
917 ib_uverbs_comp_dev(dev);
922 static int ib_uverbs_close(struct inode *inode, struct file *filp)
924 struct ib_uverbs_file *file = filp->private_data;
926 mutex_lock(&file->cleanup_mutex);
927 if (file->ucontext) {
928 ib_uverbs_cleanup_ucontext(file, file->ucontext, false);
929 file->ucontext = NULL;
931 mutex_unlock(&file->cleanup_mutex);
932 idr_destroy(&file->idr);
934 mutex_lock(&file->device->lists_mutex);
935 if (!file->is_closed) {
936 list_del(&file->list);
939 mutex_unlock(&file->device->lists_mutex);
941 if (file->async_file)
942 kref_put(&file->async_file->ref,
943 ib_uverbs_release_async_event_file);
945 kref_put(&file->ref, ib_uverbs_release_file);
950 static const struct file_operations uverbs_fops = {
951 .owner = THIS_MODULE,
952 .write = ib_uverbs_write,
953 .open = ib_uverbs_open,
954 .release = ib_uverbs_close,
956 #if IS_ENABLED(CONFIG_INFINIBAND_EXP_USER_ACCESS)
957 .unlocked_ioctl = ib_uverbs_ioctl,
961 static const struct file_operations uverbs_mmap_fops = {
962 .owner = THIS_MODULE,
963 .write = ib_uverbs_write,
964 .mmap = ib_uverbs_mmap,
965 .open = ib_uverbs_open,
966 .release = ib_uverbs_close,
968 #if IS_ENABLED(CONFIG_INFINIBAND_EXP_USER_ACCESS)
969 .unlocked_ioctl = ib_uverbs_ioctl,
973 static struct ib_client uverbs_client = {
975 .add = ib_uverbs_add_one,
976 .remove = ib_uverbs_remove_one
979 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
984 struct ib_uverbs_device *dev = dev_get_drvdata(device);
985 struct ib_device *ib_dev;
990 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
991 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
993 ret = sprintf(buf, "%s\n", ib_dev->name);
994 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
998 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1000 static ssize_t show_dev_abi_version(struct device *device,
1001 struct device_attribute *attr, char *buf)
1003 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1006 struct ib_device *ib_dev;
1010 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1011 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1013 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1014 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1018 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1020 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1021 __stringify(IB_USER_VERBS_ABI_VERSION));
1023 static dev_t overflow_maj;
1024 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1027 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1028 * requesting a new major number and doubling the number of max devices we
1029 * support. It's stupid, but simple.
1031 static int find_overflow_devnum(void)
1035 if (!overflow_maj) {
1036 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1037 "infiniband_verbs");
1039 pr_err("user_verbs: couldn't register dynamic device number\n");
1044 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1045 if (ret >= IB_UVERBS_MAX_DEVICES)
1051 static void ib_uverbs_add_one(struct ib_device *device)
1055 struct ib_uverbs_device *uverbs_dev;
1058 if (!device->alloc_ucontext)
1061 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1065 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1071 atomic_set(&uverbs_dev->refcount, 1);
1072 init_completion(&uverbs_dev->comp);
1073 uverbs_dev->xrcd_tree = RB_ROOT;
1074 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1075 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1076 mutex_init(&uverbs_dev->lists_mutex);
1077 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1078 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1080 spin_lock(&map_lock);
1081 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1082 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1083 spin_unlock(&map_lock);
1084 devnum = find_overflow_devnum();
1088 spin_lock(&map_lock);
1089 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1090 base = devnum + overflow_maj;
1091 set_bit(devnum, overflow_map);
1093 uverbs_dev->devnum = devnum;
1094 base = devnum + IB_UVERBS_BASE_DEV;
1095 set_bit(devnum, dev_map);
1097 spin_unlock(&map_lock);
1099 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1100 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1102 cdev_init(&uverbs_dev->cdev, NULL);
1103 uverbs_dev->cdev.owner = THIS_MODULE;
1104 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1105 cdev_set_parent(&uverbs_dev->cdev, &uverbs_dev->kobj);
1106 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1107 if (cdev_add(&uverbs_dev->cdev, base, 1))
1110 uverbs_dev->dev = device_create(uverbs_class, device->dev.parent,
1111 uverbs_dev->cdev.dev, uverbs_dev,
1112 "uverbs%d", uverbs_dev->devnum);
1113 if (IS_ERR(uverbs_dev->dev))
1116 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1118 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1121 if (!device->specs_root) {
1122 const struct uverbs_object_tree_def *default_root[] = {
1123 uverbs_default_get_objects()};
1125 uverbs_dev->specs_root = uverbs_alloc_spec_tree(1,
1127 if (IS_ERR(uverbs_dev->specs_root))
1130 device->specs_root = uverbs_dev->specs_root;
1133 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1138 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1141 cdev_del(&uverbs_dev->cdev);
1142 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1143 clear_bit(devnum, dev_map);
1145 clear_bit(devnum, overflow_map);
1148 if (atomic_dec_and_test(&uverbs_dev->refcount))
1149 ib_uverbs_comp_dev(uverbs_dev);
1150 wait_for_completion(&uverbs_dev->comp);
1151 kobject_put(&uverbs_dev->kobj);
1155 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1156 struct ib_device *ib_dev)
1158 struct ib_uverbs_file *file;
1159 struct ib_uverbs_async_event_file *event_file;
1160 struct ib_event event;
1162 /* Pending running commands to terminate */
1163 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1164 event.event = IB_EVENT_DEVICE_FATAL;
1165 event.element.port_num = 0;
1166 event.device = ib_dev;
1168 mutex_lock(&uverbs_dev->lists_mutex);
1169 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1170 struct ib_ucontext *ucontext;
1171 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1172 struct ib_uverbs_file, list);
1173 file->is_closed = 1;
1174 list_del(&file->list);
1175 kref_get(&file->ref);
1176 mutex_unlock(&uverbs_dev->lists_mutex);
1179 mutex_lock(&file->cleanup_mutex);
1180 ucontext = file->ucontext;
1181 file->ucontext = NULL;
1182 mutex_unlock(&file->cleanup_mutex);
1184 /* At this point ib_uverbs_close cannot be running
1185 * ib_uverbs_cleanup_ucontext
1188 /* We must release the mutex before going ahead and
1189 * calling disassociate_ucontext. disassociate_ucontext
1190 * might end up indirectly calling uverbs_close,
1191 * for example due to freeing the resources
1194 ib_uverbs_event_handler(&file->event_handler, &event);
1195 ib_dev->disassociate_ucontext(ucontext);
1196 mutex_lock(&file->cleanup_mutex);
1197 ib_uverbs_cleanup_ucontext(file, ucontext, true);
1198 mutex_unlock(&file->cleanup_mutex);
1201 mutex_lock(&uverbs_dev->lists_mutex);
1202 kref_put(&file->ref, ib_uverbs_release_file);
1205 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1206 event_file = list_first_entry(&uverbs_dev->
1207 uverbs_events_file_list,
1208 struct ib_uverbs_async_event_file,
1210 spin_lock_irq(&event_file->ev_queue.lock);
1211 event_file->ev_queue.is_closed = 1;
1212 spin_unlock_irq(&event_file->ev_queue.lock);
1214 list_del(&event_file->list);
1215 ib_unregister_event_handler(
1216 &event_file->uverbs_file->event_handler);
1217 event_file->uverbs_file->event_handler.device =
1220 wake_up_interruptible(&event_file->ev_queue.poll_wait);
1221 kill_fasync(&event_file->ev_queue.async_queue, SIGIO, POLL_IN);
1223 mutex_unlock(&uverbs_dev->lists_mutex);
1226 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1228 struct ib_uverbs_device *uverbs_dev = client_data;
1229 int wait_clients = 1;
1234 dev_set_drvdata(uverbs_dev->dev, NULL);
1235 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1236 cdev_del(&uverbs_dev->cdev);
1238 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1239 clear_bit(uverbs_dev->devnum, dev_map);
1241 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1243 if (device->disassociate_ucontext) {
1244 /* We disassociate HW resources and immediately return.
1245 * Userspace will see a EIO errno for all future access.
1246 * Upon returning, ib_device may be freed internally and is not
1248 * uverbs_device is still available until all clients close
1249 * their files, then the uverbs device ref count will be zero
1250 * and its resources will be freed.
1251 * Note: At this point no more files can be opened since the
1252 * cdev was deleted, however active clients can still issue
1253 * commands and close their open files.
1255 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1256 ib_uverbs_free_hw_resources(uverbs_dev, device);
1260 if (atomic_dec_and_test(&uverbs_dev->refcount))
1261 ib_uverbs_comp_dev(uverbs_dev);
1263 wait_for_completion(&uverbs_dev->comp);
1264 if (uverbs_dev->specs_root) {
1265 uverbs_free_spec_tree(uverbs_dev->specs_root);
1266 device->specs_root = NULL;
1269 kobject_put(&uverbs_dev->kobj);
1272 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1276 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1279 static int __init ib_uverbs_init(void)
1283 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1284 "infiniband_verbs");
1286 pr_err("user_verbs: couldn't register device number\n");
1290 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1291 if (IS_ERR(uverbs_class)) {
1292 ret = PTR_ERR(uverbs_class);
1293 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1297 uverbs_class->devnode = uverbs_devnode;
1299 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1301 pr_err("user_verbs: couldn't create abi_version attribute\n");
1305 ret = ib_register_client(&uverbs_client);
1307 pr_err("user_verbs: couldn't register client\n");
1314 class_destroy(uverbs_class);
1317 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1323 static void __exit ib_uverbs_cleanup(void)
1325 ib_unregister_client(&uverbs_client);
1326 class_destroy(uverbs_class);
1327 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1329 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1332 module_init(ib_uverbs_init);
1333 module_exit(ib_uverbs_cleanup);