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 spin_unlock_irq(&file->ev_queue.lock);
429 uverbs_close_fd(filp);
434 const struct file_operations uverbs_event_fops = {
435 .owner = THIS_MODULE,
436 .read = ib_uverbs_comp_event_read,
437 .poll = ib_uverbs_comp_event_poll,
438 .release = ib_uverbs_comp_event_close,
439 .fasync = ib_uverbs_comp_event_fasync,
443 static const struct file_operations uverbs_async_event_fops = {
444 .owner = THIS_MODULE,
445 .read = ib_uverbs_async_event_read,
446 .poll = ib_uverbs_async_event_poll,
447 .release = ib_uverbs_async_event_close,
448 .fasync = ib_uverbs_async_event_fasync,
452 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
454 struct ib_uverbs_event_queue *ev_queue = cq_context;
455 struct ib_ucq_object *uobj;
456 struct ib_uverbs_event *entry;
462 spin_lock_irqsave(&ev_queue->lock, flags);
463 if (ev_queue->is_closed) {
464 spin_unlock_irqrestore(&ev_queue->lock, flags);
468 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
470 spin_unlock_irqrestore(&ev_queue->lock, flags);
474 uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
476 entry->desc.comp.cq_handle = cq->uobject->user_handle;
477 entry->counter = &uobj->comp_events_reported;
479 list_add_tail(&entry->list, &ev_queue->event_list);
480 list_add_tail(&entry->obj_list, &uobj->comp_list);
481 spin_unlock_irqrestore(&ev_queue->lock, flags);
483 wake_up_interruptible(&ev_queue->poll_wait);
484 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
487 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
488 __u64 element, __u64 event,
489 struct list_head *obj_list,
492 struct ib_uverbs_event *entry;
495 spin_lock_irqsave(&file->async_file->ev_queue.lock, flags);
496 if (file->async_file->ev_queue.is_closed) {
497 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
501 entry = kmalloc(sizeof *entry, GFP_ATOMIC);
503 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
507 entry->desc.async.element = element;
508 entry->desc.async.event_type = event;
509 entry->desc.async.reserved = 0;
510 entry->counter = counter;
512 list_add_tail(&entry->list, &file->async_file->ev_queue.event_list);
514 list_add_tail(&entry->obj_list, obj_list);
515 spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
517 wake_up_interruptible(&file->async_file->ev_queue.poll_wait);
518 kill_fasync(&file->async_file->ev_queue.async_queue, SIGIO, POLL_IN);
521 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
523 struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
524 struct ib_ucq_object, uobject);
526 ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
527 event->event, &uobj->async_list,
528 &uobj->async_events_reported);
531 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
533 struct ib_uevent_object *uobj;
535 /* for XRC target qp's, check that qp is live */
536 if (!event->element.qp->uobject)
539 uobj = container_of(event->element.qp->uobject,
540 struct ib_uevent_object, uobject);
542 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
543 event->event, &uobj->event_list,
544 &uobj->events_reported);
547 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
549 struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
550 struct ib_uevent_object, uobject);
552 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
553 event->event, &uobj->event_list,
554 &uobj->events_reported);
557 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
559 struct ib_uevent_object *uobj;
561 uobj = container_of(event->element.srq->uobject,
562 struct ib_uevent_object, uobject);
564 ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
565 event->event, &uobj->event_list,
566 &uobj->events_reported);
569 void ib_uverbs_event_handler(struct ib_event_handler *handler,
570 struct ib_event *event)
572 struct ib_uverbs_file *file =
573 container_of(handler, struct ib_uverbs_file, event_handler);
575 ib_uverbs_async_handler(file, event->element.port_num, event->event,
579 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
581 kref_put(&file->async_file->ref, ib_uverbs_release_async_event_file);
582 file->async_file = NULL;
585 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
587 spin_lock_init(&ev_queue->lock);
588 INIT_LIST_HEAD(&ev_queue->event_list);
589 init_waitqueue_head(&ev_queue->poll_wait);
590 ev_queue->is_closed = 0;
591 ev_queue->async_queue = NULL;
594 struct file *ib_uverbs_alloc_async_event_file(struct ib_uverbs_file *uverbs_file,
595 struct ib_device *ib_dev)
597 struct ib_uverbs_async_event_file *ev_file;
600 ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
602 return ERR_PTR(-ENOMEM);
604 ib_uverbs_init_event_queue(&ev_file->ev_queue);
605 ev_file->uverbs_file = uverbs_file;
606 kref_get(&ev_file->uverbs_file->ref);
607 kref_init(&ev_file->ref);
608 filp = anon_inode_getfile("[infinibandevent]", &uverbs_async_event_fops,
613 mutex_lock(&uverbs_file->device->lists_mutex);
614 list_add_tail(&ev_file->list,
615 &uverbs_file->device->uverbs_events_file_list);
616 mutex_unlock(&uverbs_file->device->lists_mutex);
618 WARN_ON(uverbs_file->async_file);
619 uverbs_file->async_file = ev_file;
620 kref_get(&uverbs_file->async_file->ref);
621 INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
623 ib_uverbs_event_handler);
624 ib_register_event_handler(&uverbs_file->event_handler);
625 /* At that point async file stuff was fully set */
630 kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
631 kref_put(&ev_file->ref, ib_uverbs_release_async_event_file);
635 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
639 if (command <= IB_USER_VERBS_CMD_OPEN_QP)
640 mask = ib_dev->uverbs_cmd_mask;
642 mask = ib_dev->uverbs_ex_cmd_mask;
644 if (mask & ((u64)1 << command))
650 static bool verify_command_idx(u32 command, bool extended)
653 return command < ARRAY_SIZE(uverbs_ex_cmd_table);
655 return command < ARRAY_SIZE(uverbs_cmd_table);
658 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
659 size_t count, loff_t *pos)
661 struct ib_uverbs_file *file = filp->private_data;
662 struct ib_device *ib_dev;
663 struct ib_uverbs_cmd_hdr hdr;
664 bool extended_command;
670 if (!ib_safe_file_access(filp)) {
671 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
672 task_tgid_vnr(current), current->comm);
676 if (count < sizeof hdr)
679 if (copy_from_user(&hdr, buf, sizeof hdr))
682 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
683 ib_dev = srcu_dereference(file->device->ib_dev,
684 &file->device->disassociate_srcu);
690 if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
691 IB_USER_VERBS_CMD_COMMAND_MASK)) {
696 command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
697 flags = (hdr.command &
698 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
700 extended_command = flags & IB_USER_VERBS_CMD_FLAG_EXTENDED;
701 if (!verify_command_idx(command, extended_command)) {
706 if (verify_command_mask(ib_dev, command)) {
711 if (!file->ucontext &&
712 command != IB_USER_VERBS_CMD_GET_CONTEXT) {
718 if (!uverbs_cmd_table[command]) {
723 if (hdr.in_words * 4 != count) {
728 ret = uverbs_cmd_table[command](file, ib_dev,
733 } else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
734 struct ib_uverbs_ex_cmd_hdr ex_hdr;
735 struct ib_udata ucore;
737 size_t written_count = count;
739 if (!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;
925 mutex_lock(&file->cleanup_mutex);
926 if (file->ucontext) {
927 ib_uverbs_cleanup_ucontext(file, file->ucontext, false);
928 file->ucontext = NULL;
930 mutex_unlock(&file->cleanup_mutex);
931 idr_destroy(&file->idr);
933 mutex_lock(&file->device->lists_mutex);
934 if (!file->is_closed) {
935 list_del(&file->list);
938 mutex_unlock(&file->device->lists_mutex);
940 if (file->async_file)
941 kref_put(&file->async_file->ref,
942 ib_uverbs_release_async_event_file);
944 kref_put(&file->ref, ib_uverbs_release_file);
949 static const struct file_operations uverbs_fops = {
950 .owner = THIS_MODULE,
951 .write = ib_uverbs_write,
952 .open = ib_uverbs_open,
953 .release = ib_uverbs_close,
955 #if IS_ENABLED(CONFIG_INFINIBAND_EXP_USER_ACCESS)
956 .unlocked_ioctl = ib_uverbs_ioctl,
960 static const struct file_operations uverbs_mmap_fops = {
961 .owner = THIS_MODULE,
962 .write = ib_uverbs_write,
963 .mmap = ib_uverbs_mmap,
964 .open = ib_uverbs_open,
965 .release = ib_uverbs_close,
967 #if IS_ENABLED(CONFIG_INFINIBAND_EXP_USER_ACCESS)
968 .unlocked_ioctl = ib_uverbs_ioctl,
972 static struct ib_client uverbs_client = {
974 .add = ib_uverbs_add_one,
975 .remove = ib_uverbs_remove_one
978 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
983 struct ib_uverbs_device *dev = dev_get_drvdata(device);
984 struct ib_device *ib_dev;
989 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
990 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
992 ret = sprintf(buf, "%s\n", ib_dev->name);
993 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
997 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
999 static ssize_t show_dev_abi_version(struct device *device,
1000 struct device_attribute *attr, char *buf)
1002 struct ib_uverbs_device *dev = dev_get_drvdata(device);
1005 struct ib_device *ib_dev;
1009 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1010 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1012 ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1013 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1017 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1019 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1020 __stringify(IB_USER_VERBS_ABI_VERSION));
1022 static dev_t overflow_maj;
1023 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1026 * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1027 * requesting a new major number and doubling the number of max devices we
1028 * support. It's stupid, but simple.
1030 static int find_overflow_devnum(void)
1034 if (!overflow_maj) {
1035 ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1036 "infiniband_verbs");
1038 pr_err("user_verbs: couldn't register dynamic device number\n");
1043 ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1044 if (ret >= IB_UVERBS_MAX_DEVICES)
1050 static void ib_uverbs_add_one(struct ib_device *device)
1054 struct ib_uverbs_device *uverbs_dev;
1057 if (!device->alloc_ucontext)
1060 uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1064 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1070 atomic_set(&uverbs_dev->refcount, 1);
1071 init_completion(&uverbs_dev->comp);
1072 uverbs_dev->xrcd_tree = RB_ROOT;
1073 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1074 kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1075 mutex_init(&uverbs_dev->lists_mutex);
1076 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1077 INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1079 spin_lock(&map_lock);
1080 devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1081 if (devnum >= IB_UVERBS_MAX_DEVICES) {
1082 spin_unlock(&map_lock);
1083 devnum = find_overflow_devnum();
1087 spin_lock(&map_lock);
1088 uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1089 base = devnum + overflow_maj;
1090 set_bit(devnum, overflow_map);
1092 uverbs_dev->devnum = devnum;
1093 base = devnum + IB_UVERBS_BASE_DEV;
1094 set_bit(devnum, dev_map);
1096 spin_unlock(&map_lock);
1098 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1099 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1101 cdev_init(&uverbs_dev->cdev, NULL);
1102 uverbs_dev->cdev.owner = THIS_MODULE;
1103 uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1104 cdev_set_parent(&uverbs_dev->cdev, &uverbs_dev->kobj);
1105 kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1106 if (cdev_add(&uverbs_dev->cdev, base, 1))
1109 uverbs_dev->dev = device_create(uverbs_class, device->dev.parent,
1110 uverbs_dev->cdev.dev, uverbs_dev,
1111 "uverbs%d", uverbs_dev->devnum);
1112 if (IS_ERR(uverbs_dev->dev))
1115 if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1117 if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1120 if (!device->specs_root) {
1121 const struct uverbs_object_tree_def *default_root[] = {
1122 uverbs_default_get_objects()};
1124 uverbs_dev->specs_root = uverbs_alloc_spec_tree(1,
1126 if (IS_ERR(uverbs_dev->specs_root))
1129 device->specs_root = uverbs_dev->specs_root;
1132 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1137 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1140 cdev_del(&uverbs_dev->cdev);
1141 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1142 clear_bit(devnum, dev_map);
1144 clear_bit(devnum, overflow_map);
1147 if (atomic_dec_and_test(&uverbs_dev->refcount))
1148 ib_uverbs_comp_dev(uverbs_dev);
1149 wait_for_completion(&uverbs_dev->comp);
1150 kobject_put(&uverbs_dev->kobj);
1154 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1155 struct ib_device *ib_dev)
1157 struct ib_uverbs_file *file;
1158 struct ib_uverbs_async_event_file *event_file;
1159 struct ib_event event;
1161 /* Pending running commands to terminate */
1162 synchronize_srcu(&uverbs_dev->disassociate_srcu);
1163 event.event = IB_EVENT_DEVICE_FATAL;
1164 event.element.port_num = 0;
1165 event.device = ib_dev;
1167 mutex_lock(&uverbs_dev->lists_mutex);
1168 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1169 struct ib_ucontext *ucontext;
1170 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1171 struct ib_uverbs_file, list);
1172 file->is_closed = 1;
1173 list_del(&file->list);
1174 kref_get(&file->ref);
1175 mutex_unlock(&uverbs_dev->lists_mutex);
1178 mutex_lock(&file->cleanup_mutex);
1179 ucontext = file->ucontext;
1180 file->ucontext = NULL;
1181 mutex_unlock(&file->cleanup_mutex);
1183 /* At this point ib_uverbs_close cannot be running
1184 * ib_uverbs_cleanup_ucontext
1187 /* We must release the mutex before going ahead and
1188 * calling disassociate_ucontext. disassociate_ucontext
1189 * might end up indirectly calling uverbs_close,
1190 * for example due to freeing the resources
1193 ib_uverbs_event_handler(&file->event_handler, &event);
1194 ib_dev->disassociate_ucontext(ucontext);
1195 mutex_lock(&file->cleanup_mutex);
1196 ib_uverbs_cleanup_ucontext(file, ucontext, true);
1197 mutex_unlock(&file->cleanup_mutex);
1200 mutex_lock(&uverbs_dev->lists_mutex);
1201 kref_put(&file->ref, ib_uverbs_release_file);
1204 while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1205 event_file = list_first_entry(&uverbs_dev->
1206 uverbs_events_file_list,
1207 struct ib_uverbs_async_event_file,
1209 spin_lock_irq(&event_file->ev_queue.lock);
1210 event_file->ev_queue.is_closed = 1;
1211 spin_unlock_irq(&event_file->ev_queue.lock);
1213 list_del(&event_file->list);
1214 ib_unregister_event_handler(
1215 &event_file->uverbs_file->event_handler);
1216 event_file->uverbs_file->event_handler.device =
1219 wake_up_interruptible(&event_file->ev_queue.poll_wait);
1220 kill_fasync(&event_file->ev_queue.async_queue, SIGIO, POLL_IN);
1222 mutex_unlock(&uverbs_dev->lists_mutex);
1225 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1227 struct ib_uverbs_device *uverbs_dev = client_data;
1228 int wait_clients = 1;
1233 dev_set_drvdata(uverbs_dev->dev, NULL);
1234 device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1235 cdev_del(&uverbs_dev->cdev);
1237 if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1238 clear_bit(uverbs_dev->devnum, dev_map);
1240 clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1242 if (device->disassociate_ucontext) {
1243 /* We disassociate HW resources and immediately return.
1244 * Userspace will see a EIO errno for all future access.
1245 * Upon returning, ib_device may be freed internally and is not
1247 * uverbs_device is still available until all clients close
1248 * their files, then the uverbs device ref count will be zero
1249 * and its resources will be freed.
1250 * Note: At this point no more files can be opened since the
1251 * cdev was deleted, however active clients can still issue
1252 * commands and close their open files.
1254 rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1255 ib_uverbs_free_hw_resources(uverbs_dev, device);
1259 if (atomic_dec_and_test(&uverbs_dev->refcount))
1260 ib_uverbs_comp_dev(uverbs_dev);
1262 wait_for_completion(&uverbs_dev->comp);
1263 if (uverbs_dev->specs_root) {
1264 uverbs_free_spec_tree(uverbs_dev->specs_root);
1265 device->specs_root = NULL;
1268 kobject_put(&uverbs_dev->kobj);
1271 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1275 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1278 static int __init ib_uverbs_init(void)
1282 ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1283 "infiniband_verbs");
1285 pr_err("user_verbs: couldn't register device number\n");
1289 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1290 if (IS_ERR(uverbs_class)) {
1291 ret = PTR_ERR(uverbs_class);
1292 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1296 uverbs_class->devnode = uverbs_devnode;
1298 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1300 pr_err("user_verbs: couldn't create abi_version attribute\n");
1304 ret = ib_register_client(&uverbs_client);
1306 pr_err("user_verbs: couldn't register client\n");
1313 class_destroy(uverbs_class);
1316 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1322 static void __exit ib_uverbs_cleanup(void)
1324 ib_unregister_client(&uverbs_client);
1325 class_destroy(uverbs_class);
1326 unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1328 unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1331 module_init(ib_uverbs_init);
1332 module_exit(ib_uverbs_cleanup);