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_event_file(struct kref *ref)
161 struct ib_uverbs_event_file *file =
162 container_of(ref, struct ib_uverbs_event_file, ref);
167 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
168 struct ib_uverbs_event_file *ev_file,
169 struct ib_ucq_object *uobj)
171 struct ib_uverbs_event *evt, *tmp;
174 spin_lock_irq(&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->lock);
181 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
184 spin_lock_irq(&file->async_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->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->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->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 static 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 file *filp, char __user *buf,
257 size_t count, loff_t *pos)
259 struct ib_uverbs_event_file *file = filp->private_data;
260 struct ib_uverbs_event *event;
264 spin_lock_irq(&file->lock);
266 while (list_empty(&file->event_list)) {
267 spin_unlock_irq(&file->lock);
269 if (filp->f_flags & O_NONBLOCK)
272 if (wait_event_interruptible(file->poll_wait,
273 (!list_empty(&file->event_list) ||
274 /* The barriers built into wait_event_interruptible()
275 * and wake_up() guarentee this will see the null set
278 !file->uverbs_file->device->ib_dev)))
281 /* If device was disassociated and no event exists set an error */
282 if (list_empty(&file->event_list) &&
283 !file->uverbs_file->device->ib_dev)
286 spin_lock_irq(&file->lock);
289 event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
292 eventsz = sizeof (struct ib_uverbs_async_event_desc);
294 eventsz = sizeof (struct ib_uverbs_comp_event_desc);
296 if (eventsz > count) {
300 list_del(file->event_list.next);
301 if (event->counter) {
303 list_del(&event->obj_list);
307 spin_unlock_irq(&file->lock);
310 if (copy_to_user(buf, event, eventsz))
321 static unsigned int ib_uverbs_event_poll(struct file *filp,
322 struct poll_table_struct *wait)
324 unsigned int pollflags = 0;
325 struct ib_uverbs_event_file *file = filp->private_data;
327 poll_wait(filp, &file->poll_wait, wait);
329 spin_lock_irq(&file->lock);
330 if (!list_empty(&file->event_list))
331 pollflags = POLLIN | POLLRDNORM;
332 spin_unlock_irq(&file->lock);
337 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
339 struct ib_uverbs_event_file *file = filp->private_data;
341 return fasync_helper(fd, filp, on, &file->async_queue);
344 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
346 struct ib_uverbs_event_file *file = filp->private_data;
347 struct ib_uverbs_event *entry, *tmp;
348 int closed_already = 0;
350 mutex_lock(&file->uverbs_file->device->lists_mutex);
351 spin_lock_irq(&file->lock);
352 closed_already = file->is_closed;
354 list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
356 list_del(&entry->obj_list);
359 spin_unlock_irq(&file->lock);
360 if (!closed_already) {
361 list_del(&file->list);
363 ib_unregister_event_handler(&file->uverbs_file->
366 mutex_unlock(&file->uverbs_file->device->lists_mutex);
368 kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
369 kref_put(&file->ref, ib_uverbs_release_event_file);
374 static const struct file_operations uverbs_event_fops = {
375 .owner = THIS_MODULE,
376 .read = ib_uverbs_event_read,
377 .poll = ib_uverbs_event_poll,
378 .release = ib_uverbs_event_close,
379 .fasync = ib_uverbs_event_fasync,
383 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
385 struct ib_uverbs_event_file *file = cq_context;
386 struct ib_ucq_object *uobj;
387 struct ib_uverbs_event *entry;
393 spin_lock_irqsave(&file->lock, flags);
394 if (file->is_closed) {
395 spin_unlock_irqrestore(&file->lock, flags);
399 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
401 spin_unlock_irqrestore(&file->lock, flags);
405 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
407 entry->desc.comp.cq_handle = cq->uobject->user_handle;
408 entry->counter = &uobj->comp_events_reported;
410 list_add_tail(&entry->list, &file->event_list);
411 list_add_tail(&entry->obj_list, &uobj->comp_list);
412 spin_unlock_irqrestore(&file->lock, flags);
414 wake_up_interruptible(&file->poll_wait);
415 kill_fasync(&file->async_queue, SIGIO, POLL_IN);
418 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
419 __u64 element, __u64 event,
420 struct list_head *obj_list,
423 struct ib_uverbs_event *entry;
426 spin_lock_irqsave(&file->async_file->lock, flags);
427 if (file->async_file->is_closed) {
428 spin_unlock_irqrestore(&file->async_file->lock, flags);
432 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
434 spin_unlock_irqrestore(&file->async_file->lock, flags);
438 entry->desc.async.element = element;
439 entry->desc.async.event_type = event;
440 entry->desc.async.reserved = 0;
441 entry->counter = counter;
443 list_add_tail(&entry->list, &file->async_file->event_list);
445 list_add_tail(&entry->obj_list, obj_list);
446 spin_unlock_irqrestore(&file->async_file->lock, flags);
448 wake_up_interruptible(&file->async_file->poll_wait);
449 kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
452 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
454 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
455 struct ib_ucq_object, uobject);
457 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
458 event->event, &uobj->async_list,
459 &uobj->async_events_reported);
462 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
464 struct ib_uevent_object *uobj;
466 /* for XRC target qp's, check that qp is live */
467 if (!event->element.qp->uobject)
470 uobj = container_of(event->element.qp->uobject,
471 struct ib_uevent_object, uobject);
473 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
474 event->event, &uobj->event_list,
475 &uobj->events_reported);
478 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
480 struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
481 struct ib_uevent_object, uobject);
483 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
484 event->event, &uobj->event_list,
485 &uobj->events_reported);
488 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
490 struct ib_uevent_object *uobj;
492 uobj = container_of(event->element.srq->uobject,
493 struct ib_uevent_object, uobject);
495 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
496 event->event, &uobj->event_list,
497 &uobj->events_reported);
500 void ib_uverbs_event_handler(struct ib_event_handler *handler,
501 struct ib_event *event)
503 struct ib_uverbs_file *file =
504 container_of(handler, struct ib_uverbs_file, event_handler);
506 ib_uverbs_async_handler(file, event->element.port_num, event->event,
510 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
512 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
513 file->async_file = NULL;
516 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
517 struct ib_device *ib_dev,
520 struct ib_uverbs_event_file *ev_file;
524 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
526 return ERR_PTR(-ENOMEM);
528 kref_init(&ev_file->ref);
529 spin_lock_init(&ev_file->lock);
530 INIT_LIST_HEAD(&ev_file->event_list);
531 init_waitqueue_head(&ev_file->poll_wait);
532 ev_file->uverbs_file = uverbs_file;
533 kref_get(&ev_file->uverbs_file->ref);
534 ev_file->async_queue = NULL;
535 ev_file->is_closed = 0;
537 filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
542 mutex_lock(&uverbs_file->device->lists_mutex);
543 list_add_tail(&ev_file->list,
544 &uverbs_file->device->uverbs_events_file_list);
545 mutex_unlock(&uverbs_file->device->lists_mutex);
548 WARN_ON(uverbs_file->async_file);
549 uverbs_file->async_file = ev_file;
550 kref_get(&uverbs_file->async_file->ref);
551 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
553 ib_uverbs_event_handler);
554 ret = ib_register_event_handler(&uverbs_file->event_handler);
558 /* At that point async file stuff was fully set */
559 ev_file->is_async = 1;
566 kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
567 uverbs_file->async_file = NULL;
571 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
572 kref_put(&ev_file->ref, ib_uverbs_release_event_file);
577 * Look up a completion event file by FD. If lookup is successful,
578 * takes a ref to the event file struct that it returns; if
579 * unsuccessful, returns NULL.
581 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
583 struct ib_uverbs_event_file *ev_file = NULL;
584 struct fd f = fdget(fd);
589 if (f.file->f_op != &uverbs_event_fops)
592 ev_file = f.file->private_data;
593 if (ev_file->is_async) {
598 kref_get(&ev_file->ref);
605 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
609 if (command <= IB_USER_VERBS_CMD_OPEN_QP)
610 mask = ib_dev->uverbs_cmd_mask;
612 mask = ib_dev->uverbs_ex_cmd_mask;
614 if (mask & ((u64)1 << command))
620 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
621 size_t count, loff_t *pos)
623 struct ib_uverbs_file *file = filp->private_data;
624 struct ib_device *ib_dev;
625 struct ib_uverbs_cmd_hdr hdr;
631 if (!ib_safe_file_access(filp)) {
632 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
633 task_tgid_vnr(current), current->comm);
637 if (count < sizeof hdr)
640 if (copy_from_user(&hdr, buf, sizeof hdr))
643 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
644 ib_dev = srcu_dereference(file->device->ib_dev,
645 &file->device->disassociate_srcu);
651 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
652 IB_USER_VERBS_CMD_COMMAND_MASK)) {
657 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
658 if (verify_command_mask(ib_dev, command)) {
663 if (!file->ucontext &&
664 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
669 flags = (hdr.command &
670 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
673 if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
674 !uverbs_cmd_table[command]) {
679 if (hdr.in_words * 4 != count) {
684 ret = uverbs_cmd_table[command](file, ib_dev,
689 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
690 struct ib_uverbs_ex_cmd_hdr ex_hdr;
691 struct ib_udata ucore;
693 size_t written_count = count;
695 if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
696 !uverbs_ex_cmd_table[command]) {
701 if (!file->ucontext) {
706 if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
711 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
716 count -= sizeof(hdr) + sizeof(ex_hdr);
717 buf += sizeof(hdr) + sizeof(ex_hdr);
719 if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
724 if (ex_hdr.cmd_hdr_reserved) {
729 if (ex_hdr.response) {
730 if (!hdr.out_words && !ex_hdr.provider_out_words) {
735 if (!access_ok(VERIFY_WRITE,
736 (void __user *) (unsigned long) ex_hdr.response,
737 (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
742 if (hdr.out_words || ex_hdr.provider_out_words) {
748 INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
749 hdr.in_words * 8, hdr.out_words * 8);
751 INIT_UDATA_BUF_OR_NULL(&uhw,
753 (unsigned long) ex_hdr.response + ucore.outlen,
754 ex_hdr.provider_in_words * 8,
755 ex_hdr.provider_out_words * 8);
757 ret = uverbs_ex_cmd_table[command](file,
768 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
772 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
774 struct ib_uverbs_file *file = filp->private_data;
775 struct ib_device *ib_dev;
779 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
780 ib_dev = srcu_dereference(file->device->ib_dev,
781 &file->device->disassociate_srcu);
790 ret = ib_dev->mmap(file->ucontext, vma);
792 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
797 * ib_uverbs_open() does not need the BKL:
799 * - the ib_uverbs_device structures are properly reference counted and
800 * everything else is purely local to the file being created, so
801 * races against other open calls are not a problem;
802 * - there is no ioctl method to race against;
803 * - the open method will either immediately run -ENXIO, or all
804 * required initialization will be done.
806 static int ib_uverbs_open(struct inode *inode, struct file *filp)
808 struct ib_uverbs_device *dev;
809 struct ib_uverbs_file *file;
810 struct ib_device *ib_dev;
812 int module_dependent;
815 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
816 if (!atomic_inc_not_zero(&dev->refcount))
819 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
820 mutex_lock(&dev->lists_mutex);
821 ib_dev = srcu_dereference(dev->ib_dev,
822 &dev->disassociate_srcu);
828 /* In case IB device supports disassociate ucontext, there is no hard
829 * dependency between uverbs device and its low level device.
831 module_dependent = !(ib_dev->disassociate_ucontext);
833 if (module_dependent) {
834 if (!try_module_get(ib_dev->owner)) {
840 file = kzalloc(sizeof(*file), GFP_KERNEL);
843 if (module_dependent)
850 spin_lock_init(&file->idr_lock);
851 idr_init(&file->idr);
852 file->ucontext = NULL;
853 file->async_file = NULL;
854 kref_init(&file->ref);
855 mutex_init(&file->mutex);
856 mutex_init(&file->cleanup_mutex);
858 filp->private_data = file;
859 kobject_get(&dev->kobj);
860 list_add_tail(&file->list, &dev->uverbs_file_list);
861 mutex_unlock(&dev->lists_mutex);
862 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
864 return nonseekable_open(inode, filp);
867 module_put(ib_dev->owner);
870 mutex_unlock(&dev->lists_mutex);
871 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
872 if (atomic_dec_and_test(&dev->refcount))
873 ib_uverbs_comp_dev(dev);
878 static int ib_uverbs_close(struct inode *inode, struct file *filp)
880 struct ib_uverbs_file *file = filp->private_data;
881 struct ib_uverbs_device *dev = file->device;
883 mutex_lock(&file->cleanup_mutex);
884 if (file->ucontext) {
885 ib_uverbs_cleanup_ucontext(file, file->ucontext, false);
886 file->ucontext = NULL;
888 mutex_unlock(&file->cleanup_mutex);
889 idr_destroy(&file->idr);
891 mutex_lock(&file->device->lists_mutex);
892 if (!file->is_closed) {
893 list_del(&file->list);
896 mutex_unlock(&file->device->lists_mutex);
898 if (file->async_file)
899 kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
901 kref_put(&file->ref, ib_uverbs_release_file);
902 kobject_put(&dev->kobj);
907 static const struct file_operations uverbs_fops = {
908 .owner = THIS_MODULE,
909 .write = ib_uverbs_write,
910 .open = ib_uverbs_open,
911 .release = ib_uverbs_close,
915 static const struct file_operations uverbs_mmap_fops = {
916 .owner = THIS_MODULE,
917 .write = ib_uverbs_write,
918 .mmap = ib_uverbs_mmap,
919 .open = ib_uverbs_open,
920 .release = ib_uverbs_close,
924 static struct ib_client uverbs_client = {
926 .add = ib_uverbs_add_one,
927 .remove = ib_uverbs_remove_one
930 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
935 struct ib_uverbs_device *dev = dev_get_drvdata(device);
936 struct ib_device *ib_dev;
941 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
942 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
944 ret = sprintf(buf, "%s\n", ib_dev->name);
945 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
949 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
951 static ssize_t show_dev_abi_version(struct device *device,
952 struct device_attribute *attr, char *buf)
954 struct ib_uverbs_device *dev = dev_get_drvdata(device);
957 struct ib_device *ib_dev;
961 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
962 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
964 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
965 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
969 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
971 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
972 __stringify(IB_USER_VERBS_ABI_VERSION));
974 static dev_t overflow_maj;
975 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
978 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
979 * requesting a new major number and doubling the number of max devices we
980 * support. It's stupid, but simple.
982 static int find_overflow_devnum(void)
987 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
990 pr_err("user_verbs: couldn't register dynamic device number\n");
995 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
996 if (ret >= IB_UVERBS_MAX_DEVICES)
1002 static void ib_uverbs_add_one(struct ib_device *device)
1006 struct ib_uverbs_device *uverbs_dev;
1009 if (!device->alloc_ucontext)
1012 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1016 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1022 atomic_set(&uverbs_dev->refcount, 1);
1023 init_completion(&uverbs_dev->comp);
1024 uverbs_dev->xrcd_tree = RB_ROOT;
1025 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1026 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1027 mutex_init(&uverbs_dev->lists_mutex);
1028 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1029 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1031 spin_lock(&map_lock);
1032 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1033 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1034 spin_unlock(&map_lock);
1035 devnum = find_overflow_devnum();
1039 spin_lock(&map_lock);
1040 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1041 base = devnum + overflow_maj;
1042 set_bit(devnum, overflow_map);
1044 uverbs_dev->devnum = devnum;
1045 base = devnum + IB_UVERBS_BASE_DEV;
1046 set_bit(devnum, dev_map);
1048 spin_unlock(&map_lock);
1050 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1051 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1053 cdev_init(&uverbs_dev->cdev, NULL);
1054 uverbs_dev->cdev.owner = THIS_MODULE;
1055 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1056 uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1057 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1058 if (cdev_add(&uverbs_dev->cdev, base, 1))
1061 uverbs_dev->dev = device_create(uverbs_class, device->dev.parent,
1062 uverbs_dev->cdev.dev, uverbs_dev,
1063 "uverbs%d", uverbs_dev->devnum);
1064 if (IS_ERR(uverbs_dev->dev))
1067 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1069 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1072 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1077 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1080 cdev_del(&uverbs_dev->cdev);
1081 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1082 clear_bit(devnum, dev_map);
1084 clear_bit(devnum, overflow_map);
1087 if (atomic_dec_and_test(&uverbs_dev->refcount))
1088 ib_uverbs_comp_dev(uverbs_dev);
1089 wait_for_completion(&uverbs_dev->comp);
1090 kobject_put(&uverbs_dev->kobj);
1094 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1095 struct ib_device *ib_dev)
1097 struct ib_uverbs_file *file;
1098 struct ib_uverbs_event_file *event_file;
1099 struct ib_event event;
1101 /* Pending running commands to terminate */
1102 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1103 event.event = IB_EVENT_DEVICE_FATAL;
1104 event.element.port_num = 0;
1105 event.device = ib_dev;
1107 mutex_lock(&uverbs_dev->lists_mutex);
1108 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1109 struct ib_ucontext *ucontext;
1110 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1111 struct ib_uverbs_file, list);
1112 file->is_closed = 1;
1113 list_del(&file->list);
1114 kref_get(&file->ref);
1115 mutex_unlock(&uverbs_dev->lists_mutex);
1117 ib_uverbs_event_handler(&file->event_handler, &event);
1119 mutex_lock(&file->cleanup_mutex);
1120 ucontext = file->ucontext;
1121 file->ucontext = NULL;
1122 mutex_unlock(&file->cleanup_mutex);
1124 /* At this point ib_uverbs_close cannot be running
1125 * ib_uverbs_cleanup_ucontext
1128 /* We must release the mutex before going ahead and
1129 * calling disassociate_ucontext. disassociate_ucontext
1130 * might end up indirectly calling uverbs_close,
1131 * for example due to freeing the resources
1134 ib_dev->disassociate_ucontext(ucontext);
1135 ib_uverbs_cleanup_ucontext(file, ucontext, true);
1138 mutex_lock(&uverbs_dev->lists_mutex);
1139 kref_put(&file->ref, ib_uverbs_release_file);
1142 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1143 event_file = list_first_entry(&uverbs_dev->
1144 uverbs_events_file_list,
1145 struct ib_uverbs_event_file,
1147 spin_lock_irq(&event_file->lock);
1148 event_file->is_closed = 1;
1149 spin_unlock_irq(&event_file->lock);
1151 list_del(&event_file->list);
1152 if (event_file->is_async) {
1153 ib_unregister_event_handler(&event_file->uverbs_file->
1155 event_file->uverbs_file->event_handler.device = NULL;
1158 wake_up_interruptible(&event_file->poll_wait);
1159 kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1161 mutex_unlock(&uverbs_dev->lists_mutex);
1164 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1166 struct ib_uverbs_device *uverbs_dev = client_data;
1167 int wait_clients = 1;
1172 dev_set_drvdata(uverbs_dev->dev, NULL);
1173 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1174 cdev_del(&uverbs_dev->cdev);
1176 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1177 clear_bit(uverbs_dev->devnum, dev_map);
1179 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1181 if (device->disassociate_ucontext) {
1182 /* We disassociate HW resources and immediately return.
1183 * Userspace will see a EIO errno for all future access.
1184 * Upon returning, ib_device may be freed internally and is not
1186 * uverbs_device is still available until all clients close
1187 * their files, then the uverbs device ref count will be zero
1188 * and its resources will be freed.
1189 * Note: At this point no more files can be opened since the
1190 * cdev was deleted, however active clients can still issue
1191 * commands and close their open files.
1193 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1194 ib_uverbs_free_hw_resources(uverbs_dev, device);
1198 if (atomic_dec_and_test(&uverbs_dev->refcount))
1199 ib_uverbs_comp_dev(uverbs_dev);
1201 wait_for_completion(&uverbs_dev->comp);
1202 kobject_put(&uverbs_dev->kobj);
1205 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1209 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1212 static int __init ib_uverbs_init(void)
1216 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1217 "infiniband_verbs");
1219 pr_err("user_verbs: couldn't register device number\n");
1223 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1224 if (IS_ERR(uverbs_class)) {
1225 ret = PTR_ERR(uverbs_class);
1226 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1230 uverbs_class->devnode = uverbs_devnode;
1232 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1234 pr_err("user_verbs: couldn't create abi_version attribute\n");
1238 ret = ib_register_client(&uverbs_client);
1240 pr_err("user_verbs: couldn't register client\n");
1247 class_destroy(uverbs_class);
1250 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1256 static void __exit ib_uverbs_cleanup(void)
1258 ib_unregister_client(&uverbs_client);
1259 class_destroy(uverbs_class);
1260 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1262 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1265 module_init(ib_uverbs_init);
1266 module_exit(ib_uverbs_cleanup);