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_file.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_file.lock);
181 uverbs_uobject_put(&ev_file->uobj_file.uobj);
184 spin_lock_irq(&file->async_file->ev_file.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_file.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_file.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_file.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);
256 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_file *file,
257 struct ib_uverbs_file *uverbs_file,
258 struct file *filp, char __user *buf,
259 size_t count, loff_t *pos,
262 struct ib_uverbs_event *event;
266 spin_lock_irq(&file->lock);
268 while (list_empty(&file->event_list)) {
269 spin_unlock_irq(&file->lock);
271 if (filp->f_flags & O_NONBLOCK)
274 if (wait_event_interruptible(file->poll_wait,
275 (!list_empty(&file->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(&file->event_list) &&
285 !uverbs_file->device->ib_dev)
288 spin_lock_irq(&file->lock);
291 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
294 eventsz = sizeof (struct ib_uverbs_async_event_desc);
296 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
298 if (eventsz > count) {
302 list_del(file->event_list.next);
303 if (event->counter) {
305 list_del(&event->obj_list);
309 spin_unlock_irq(&file->lock);
312 if (copy_to_user(buf, event, eventsz))
323 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
324 size_t count, loff_t *pos)
326 struct ib_uverbs_async_event_file *file = filp->private_data;
328 return ib_uverbs_event_read(&file->ev_file, file->uverbs_file, filp,
329 buf, count, pos, true);
332 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
333 size_t count, loff_t *pos)
335 struct ib_uverbs_completion_event_file *comp_ev_file =
338 return ib_uverbs_event_read(&comp_ev_file->ev_file,
339 comp_ev_file->uobj_file.ufile, filp,
340 buf, count, pos, false);
343 static unsigned int ib_uverbs_event_poll(struct ib_uverbs_event_file *file,
345 struct poll_table_struct *wait)
347 unsigned int pollflags = 0;
349 poll_wait(filp, &file->poll_wait, wait);
351 spin_lock_irq(&file->lock);
352 if (!list_empty(&file->event_list))
353 pollflags = POLLIN | POLLRDNORM;
354 spin_unlock_irq(&file->lock);
359 static unsigned int ib_uverbs_async_event_poll(struct file *filp,
360 struct poll_table_struct *wait)
362 return ib_uverbs_event_poll(filp->private_data, filp, wait);
365 static unsigned int ib_uverbs_comp_event_poll(struct file *filp,
366 struct poll_table_struct *wait)
368 struct ib_uverbs_completion_event_file *comp_ev_file =
371 return ib_uverbs_event_poll(&comp_ev_file->ev_file, filp, wait);
374 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
376 struct ib_uverbs_event_file *file = filp->private_data;
378 return fasync_helper(fd, filp, on, &file->async_queue);
381 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
383 struct ib_uverbs_completion_event_file *comp_ev_file =
386 return fasync_helper(fd, filp, on, &comp_ev_file->ev_file.async_queue);
389 static int ib_uverbs_async_event_close(struct inode *inode, struct file *filp)
391 struct ib_uverbs_async_event_file *file = filp->private_data;
392 struct ib_uverbs_file *uverbs_file = file->uverbs_file;
393 struct ib_uverbs_event *entry, *tmp;
394 int closed_already = 0;
396 mutex_lock(&uverbs_file->device->lists_mutex);
397 spin_lock_irq(&file->ev_file.lock);
398 closed_already = file->ev_file.is_closed;
399 file->ev_file.is_closed = 1;
400 list_for_each_entry_safe(entry, tmp, &file->ev_file.event_list, list) {
402 list_del(&entry->obj_list);
405 spin_unlock_irq(&file->ev_file.lock);
406 if (!closed_already) {
407 list_del(&file->list);
408 ib_unregister_event_handler(&uverbs_file->event_handler);
410 mutex_unlock(&uverbs_file->device->lists_mutex);
412 kref_put(&uverbs_file->ref, ib_uverbs_release_file);
413 kref_put(&file->ref, ib_uverbs_release_async_event_file);
418 static int ib_uverbs_comp_event_close(struct inode *inode, struct file *filp)
420 struct ib_uverbs_completion_event_file *file = filp->private_data;
421 struct ib_uverbs_event *entry, *tmp;
423 spin_lock_irq(&file->ev_file.lock);
424 list_for_each_entry_safe(entry, tmp, &file->ev_file.event_list, list) {
426 list_del(&entry->obj_list);
429 spin_unlock_irq(&file->ev_file.lock);
431 uverbs_close_fd(filp);
436 const struct file_operations uverbs_event_fops = {
437 .owner = THIS_MODULE,
438 .read = ib_uverbs_comp_event_read,
439 .poll = ib_uverbs_comp_event_poll,
440 .release = ib_uverbs_comp_event_close,
441 .fasync = ib_uverbs_comp_event_fasync,
445 static const struct file_operations uverbs_async_event_fops = {
446 .owner = THIS_MODULE,
447 .read = ib_uverbs_async_event_read,
448 .poll = ib_uverbs_async_event_poll,
449 .release = ib_uverbs_async_event_close,
450 .fasync = ib_uverbs_async_event_fasync,
454 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
456 struct ib_uverbs_event_file *file = cq_context;
457 struct ib_ucq_object *uobj;
458 struct ib_uverbs_event *entry;
464 spin_lock_irqsave(&file->lock, flags);
465 if (file->is_closed) {
466 spin_unlock_irqrestore(&file->lock, flags);
470 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
472 spin_unlock_irqrestore(&file->lock, flags);
476 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
478 entry->desc.comp.cq_handle = cq->uobject->user_handle;
479 entry->counter = &uobj->comp_events_reported;
481 list_add_tail(&entry->list, &file->event_list);
482 list_add_tail(&entry->obj_list, &uobj->comp_list);
483 spin_unlock_irqrestore(&file->lock, flags);
485 wake_up_interruptible(&file->poll_wait);
486 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
489 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
490 __u64 element, __u64 event,
491 struct list_head *obj_list,
494 struct ib_uverbs_event *entry;
497 spin_lock_irqsave(&file->async_file->ev_file.lock, flags);
498 if (file->async_file->ev_file.is_closed) {
499 spin_unlock_irqrestore(&file->async_file->ev_file.lock, flags);
503 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
505 spin_unlock_irqrestore(&file->async_file->ev_file.lock, flags);
509 entry->desc.async.element = element;
510 entry->desc.async.event_type = event;
511 entry->desc.async.reserved = 0;
512 entry->counter = counter;
514 list_add_tail(&entry->list, &file->async_file->ev_file.event_list);
516 list_add_tail(&entry->obj_list, obj_list);
517 spin_unlock_irqrestore(&file->async_file->ev_file.lock, flags);
519 wake_up_interruptible(&file->async_file->ev_file.poll_wait);
520 kill_fasync(&file->async_file->ev_file.async_queue, SIGIO, POLL_IN);
523 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
525 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
526 struct ib_ucq_object, uobject);
528 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
529 event->event, &uobj->async_list,
530 &uobj->async_events_reported);
533 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
535 struct ib_uevent_object *uobj;
537 /* for XRC target qp's, check that qp is live */
538 if (!event->element.qp->uobject)
541 uobj = container_of(event->element.qp->uobject,
542 struct ib_uevent_object, uobject);
544 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
545 event->event, &uobj->event_list,
546 &uobj->events_reported);
549 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
551 struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
552 struct ib_uevent_object, uobject);
554 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
555 event->event, &uobj->event_list,
556 &uobj->events_reported);
559 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
561 struct ib_uevent_object *uobj;
563 uobj = container_of(event->element.srq->uobject,
564 struct ib_uevent_object, uobject);
566 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
567 event->event, &uobj->event_list,
568 &uobj->events_reported);
571 void ib_uverbs_event_handler(struct ib_event_handler *handler,
572 struct ib_event *event)
574 struct ib_uverbs_file *file =
575 container_of(handler, struct ib_uverbs_file, event_handler);
577 ib_uverbs_async_handler(file, event->element.port_num, event->event,
581 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
583 kref_put(&file->async_file->ref, ib_uverbs_release_async_event_file);
584 file->async_file = NULL;
587 void ib_uverbs_init_event_file(struct ib_uverbs_event_file *ev_file)
589 spin_lock_init(&ev_file->lock);
590 INIT_LIST_HEAD(&ev_file->event_list);
591 init_waitqueue_head(&ev_file->poll_wait);
592 ev_file->is_closed = 0;
593 ev_file->async_queue = NULL;
596 struct file *ib_uverbs_alloc_async_event_file(struct ib_uverbs_file *uverbs_file,
597 struct ib_device *ib_dev)
599 struct ib_uverbs_async_event_file *ev_file;
603 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
605 return ERR_PTR(-ENOMEM);
607 ib_uverbs_init_event_file(&ev_file->ev_file);
608 ev_file->uverbs_file = uverbs_file;
609 kref_get(&ev_file->uverbs_file->ref);
610 kref_init(&ev_file->ref);
611 filp = anon_inode_getfile("[infinibandevent]", &uverbs_async_event_fops,
616 mutex_lock(&uverbs_file->device->lists_mutex);
617 list_add_tail(&ev_file->list,
618 &uverbs_file->device->uverbs_events_file_list);
619 mutex_unlock(&uverbs_file->device->lists_mutex);
621 WARN_ON(uverbs_file->async_file);
622 uverbs_file->async_file = ev_file;
623 kref_get(&uverbs_file->async_file->ref);
624 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
626 ib_uverbs_event_handler);
627 ret = ib_register_event_handler(&uverbs_file->event_handler);
631 /* At that point async file stuff was fully set */
637 kref_put(&uverbs_file->async_file->ref,
638 ib_uverbs_release_async_event_file);
639 uverbs_file->async_file = NULL;
643 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
644 kref_put(&ev_file->ref, ib_uverbs_release_async_event_file);
648 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
652 if (command <= IB_USER_VERBS_CMD_OPEN_QP)
653 mask = ib_dev->uverbs_cmd_mask;
655 mask = ib_dev->uverbs_ex_cmd_mask;
657 if (mask & ((u64)1 << command))
663 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
664 size_t count, loff_t *pos)
666 struct ib_uverbs_file *file = filp->private_data;
667 struct ib_device *ib_dev;
668 struct ib_uverbs_cmd_hdr hdr;
674 if (!ib_safe_file_access(filp)) {
675 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
676 task_tgid_vnr(current), current->comm);
680 if (count < sizeof hdr)
683 if (copy_from_user(&hdr, buf, sizeof hdr))
686 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
687 ib_dev = srcu_dereference(file->device->ib_dev,
688 &file->device->disassociate_srcu);
694 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
695 IB_USER_VERBS_CMD_COMMAND_MASK)) {
700 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
701 if (verify_command_mask(ib_dev, command)) {
706 if (!file->ucontext &&
707 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
712 flags = (hdr.command &
713 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
716 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
717 !uverbs_cmd_table[command]) {
722 if (hdr.in_words * 4 != count) {
727 ret = uverbs_cmd_table[command](file, ib_dev,
732 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
733 struct ib_uverbs_ex_cmd_hdr ex_hdr;
734 struct ib_udata ucore;
736 size_t written_count = count;
738 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
739 !uverbs_ex_cmd_table[command]) {
744 if (!file->ucontext) {
749 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
754 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
759 count -= sizeof(hdr) + sizeof(ex_hdr);
760 buf += sizeof(hdr) + sizeof(ex_hdr);
762 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
767 if (ex_hdr.cmd_hdr_reserved) {
772 if (ex_hdr.response) {
773 if (!hdr.out_words && !ex_hdr.provider_out_words) {
778 if (!access_ok(VERIFY_WRITE,
779 (void __user *) (unsigned long) ex_hdr.response,
780 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
785 if (hdr.out_words || ex_hdr.provider_out_words) {
791 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
792 hdr.in_words * 8, hdr.out_words * 8);
794 INIT_UDATA_BUF_OR_NULL(&uhw,
796 (unsigned long) ex_hdr.response + ucore.outlen,
797 ex_hdr.provider_in_words * 8,
798 ex_hdr.provider_out_words * 8);
800 ret = uverbs_ex_cmd_table[command](file,
811 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
815 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
817 struct ib_uverbs_file *file = filp->private_data;
818 struct ib_device *ib_dev;
822 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
823 ib_dev = srcu_dereference(file->device->ib_dev,
824 &file->device->disassociate_srcu);
833 ret = ib_dev->mmap(file->ucontext, vma);
835 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
840 * ib_uverbs_open() does not need the BKL:
842 * - the ib_uverbs_device structures are properly reference counted and
843 * everything else is purely local to the file being created, so
844 * races against other open calls are not a problem;
845 * - there is no ioctl method to race against;
846 * - the open method will either immediately run -ENXIO, or all
847 * required initialization will be done.
849 static int ib_uverbs_open(struct inode *inode, struct file *filp)
851 struct ib_uverbs_device *dev;
852 struct ib_uverbs_file *file;
853 struct ib_device *ib_dev;
855 int module_dependent;
858 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
859 if (!atomic_inc_not_zero(&dev->refcount))
862 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
863 mutex_lock(&dev->lists_mutex);
864 ib_dev = srcu_dereference(dev->ib_dev,
865 &dev->disassociate_srcu);
871 /* In case IB device supports disassociate ucontext, there is no hard
872 * dependency between uverbs device and its low level device.
874 module_dependent = !(ib_dev->disassociate_ucontext);
876 if (module_dependent) {
877 if (!try_module_get(ib_dev->owner)) {
883 file = kzalloc(sizeof(*file), GFP_KERNEL);
886 if (module_dependent)
893 spin_lock_init(&file->idr_lock);
894 idr_init(&file->idr);
895 file->ucontext = NULL;
896 file->async_file = NULL;
897 kref_init(&file->ref);
898 mutex_init(&file->mutex);
899 mutex_init(&file->cleanup_mutex);
901 filp->private_data = file;
902 kobject_get(&dev->kobj);
903 list_add_tail(&file->list, &dev->uverbs_file_list);
904 mutex_unlock(&dev->lists_mutex);
905 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
907 return nonseekable_open(inode, filp);
910 module_put(ib_dev->owner);
913 mutex_unlock(&dev->lists_mutex);
914 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
915 if (atomic_dec_and_test(&dev->refcount))
916 ib_uverbs_comp_dev(dev);
921 static int ib_uverbs_close(struct inode *inode, struct file *filp)
923 struct ib_uverbs_file *file = filp->private_data;
924 struct ib_uverbs_device *dev = file->device;
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);
946 kobject_put(&dev->kobj);
951 static const struct file_operations uverbs_fops = {
952 .owner = THIS_MODULE,
953 .write = ib_uverbs_write,
954 .open = ib_uverbs_open,
955 .release = ib_uverbs_close,
959 static const struct file_operations uverbs_mmap_fops = {
960 .owner = THIS_MODULE,
961 .write = ib_uverbs_write,
962 .mmap = ib_uverbs_mmap,
963 .open = ib_uverbs_open,
964 .release = ib_uverbs_close,
968 static struct ib_client uverbs_client = {
970 .add = ib_uverbs_add_one,
971 .remove = ib_uverbs_remove_one
974 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
979 struct ib_uverbs_device *dev = dev_get_drvdata(device);
980 struct ib_device *ib_dev;
985 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
986 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
988 ret = sprintf(buf, "%s\n", ib_dev->name);
989 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
993 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
995 static ssize_t show_dev_abi_version(struct device *device,
996 struct device_attribute *attr, char *buf)
998 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1001 struct ib_device *ib_dev;
1005 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1006 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1008 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1009 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1013 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1015 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1016 __stringify(IB_USER_VERBS_ABI_VERSION));
1018 static dev_t overflow_maj;
1019 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1022 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1023 * requesting a new major number and doubling the number of max devices we
1024 * support. It's stupid, but simple.
1026 static int find_overflow_devnum(void)
1030 if (!overflow_maj) {
1031 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1032 "infiniband_verbs");
1034 pr_err("user_verbs: couldn't register dynamic device number\n");
1039 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1040 if (ret >= IB_UVERBS_MAX_DEVICES)
1046 static void ib_uverbs_add_one(struct ib_device *device)
1050 struct ib_uverbs_device *uverbs_dev;
1053 if (!device->alloc_ucontext)
1056 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1060 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1066 atomic_set(&uverbs_dev->refcount, 1);
1067 init_completion(&uverbs_dev->comp);
1068 uverbs_dev->xrcd_tree = RB_ROOT;
1069 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1070 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1071 mutex_init(&uverbs_dev->lists_mutex);
1072 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1073 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1075 spin_lock(&map_lock);
1076 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1077 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1078 spin_unlock(&map_lock);
1079 devnum = find_overflow_devnum();
1083 spin_lock(&map_lock);
1084 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1085 base = devnum + overflow_maj;
1086 set_bit(devnum, overflow_map);
1088 uverbs_dev->devnum = devnum;
1089 base = devnum + IB_UVERBS_BASE_DEV;
1090 set_bit(devnum, dev_map);
1092 spin_unlock(&map_lock);
1094 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1095 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1097 cdev_init(&uverbs_dev->cdev, NULL);
1098 uverbs_dev->cdev.owner = THIS_MODULE;
1099 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1100 uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1101 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1102 if (cdev_add(&uverbs_dev->cdev, base, 1))
1105 uverbs_dev->dev = device_create(uverbs_class, device->dev.parent,
1106 uverbs_dev->cdev.dev, uverbs_dev,
1107 "uverbs%d", uverbs_dev->devnum);
1108 if (IS_ERR(uverbs_dev->dev))
1111 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1113 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1116 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1121 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1124 cdev_del(&uverbs_dev->cdev);
1125 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1126 clear_bit(devnum, dev_map);
1128 clear_bit(devnum, overflow_map);
1131 if (atomic_dec_and_test(&uverbs_dev->refcount))
1132 ib_uverbs_comp_dev(uverbs_dev);
1133 wait_for_completion(&uverbs_dev->comp);
1134 kobject_put(&uverbs_dev->kobj);
1138 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1139 struct ib_device *ib_dev)
1141 struct ib_uverbs_file *file;
1142 struct ib_uverbs_async_event_file *event_file;
1143 struct ib_event event;
1145 /* Pending running commands to terminate */
1146 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1147 event.event = IB_EVENT_DEVICE_FATAL;
1148 event.element.port_num = 0;
1149 event.device = ib_dev;
1151 mutex_lock(&uverbs_dev->lists_mutex);
1152 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1153 struct ib_ucontext *ucontext;
1154 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1155 struct ib_uverbs_file, list);
1156 file->is_closed = 1;
1157 list_del(&file->list);
1158 kref_get(&file->ref);
1159 mutex_unlock(&uverbs_dev->lists_mutex);
1161 ib_uverbs_event_handler(&file->event_handler, &event);
1163 mutex_lock(&file->cleanup_mutex);
1164 ucontext = file->ucontext;
1165 file->ucontext = NULL;
1166 mutex_unlock(&file->cleanup_mutex);
1168 /* At this point ib_uverbs_close cannot be running
1169 * ib_uverbs_cleanup_ucontext
1172 /* We must release the mutex before going ahead and
1173 * calling disassociate_ucontext. disassociate_ucontext
1174 * might end up indirectly calling uverbs_close,
1175 * for example due to freeing the resources
1178 ib_dev->disassociate_ucontext(ucontext);
1179 mutex_lock(&file->cleanup_mutex);
1180 ib_uverbs_cleanup_ucontext(file, ucontext, true);
1181 mutex_unlock(&file->cleanup_mutex);
1184 mutex_lock(&uverbs_dev->lists_mutex);
1185 kref_put(&file->ref, ib_uverbs_release_file);
1188 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1189 event_file = list_first_entry(&uverbs_dev->
1190 uverbs_events_file_list,
1191 struct ib_uverbs_async_event_file,
1193 spin_lock_irq(&event_file->ev_file.lock);
1194 event_file->ev_file.is_closed = 1;
1195 spin_unlock_irq(&event_file->ev_file.lock);
1197 list_del(&event_file->list);
1198 ib_unregister_event_handler(
1199 &event_file->uverbs_file->event_handler);
1200 event_file->uverbs_file->event_handler.device =
1203 wake_up_interruptible(&event_file->ev_file.poll_wait);
1204 kill_fasync(&event_file->ev_file.async_queue, SIGIO, POLL_IN);
1206 mutex_unlock(&uverbs_dev->lists_mutex);
1209 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1211 struct ib_uverbs_device *uverbs_dev = client_data;
1212 int wait_clients = 1;
1217 dev_set_drvdata(uverbs_dev->dev, NULL);
1218 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1219 cdev_del(&uverbs_dev->cdev);
1221 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1222 clear_bit(uverbs_dev->devnum, dev_map);
1224 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1226 if (device->disassociate_ucontext) {
1227 /* We disassociate HW resources and immediately return.
1228 * Userspace will see a EIO errno for all future access.
1229 * Upon returning, ib_device may be freed internally and is not
1231 * uverbs_device is still available until all clients close
1232 * their files, then the uverbs device ref count will be zero
1233 * and its resources will be freed.
1234 * Note: At this point no more files can be opened since the
1235 * cdev was deleted, however active clients can still issue
1236 * commands and close their open files.
1238 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1239 ib_uverbs_free_hw_resources(uverbs_dev, device);
1243 if (atomic_dec_and_test(&uverbs_dev->refcount))
1244 ib_uverbs_comp_dev(uverbs_dev);
1246 wait_for_completion(&uverbs_dev->comp);
1247 kobject_put(&uverbs_dev->kobj);
1250 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1254 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1257 static int __init ib_uverbs_init(void)
1261 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1262 "infiniband_verbs");
1264 pr_err("user_verbs: couldn't register device number\n");
1268 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1269 if (IS_ERR(uverbs_class)) {
1270 ret = PTR_ERR(uverbs_class);
1271 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1275 uverbs_class->devnode = uverbs_devnode;
1277 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1279 pr_err("user_verbs: couldn't create abi_version attribute\n");
1283 ret = ib_register_client(&uverbs_client);
1285 pr_err("user_verbs: couldn't register client\n");
1292 class_destroy(uverbs_class);
1295 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1301 static void __exit ib_uverbs_cleanup(void)
1303 ib_unregister_client(&uverbs_client);
1304 class_destroy(uverbs_class);
1305 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1307 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1310 module_init(ib_uverbs_init);
1311 module_exit(ib_uverbs_cleanup);