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>
48 #include <linux/sched/mm.h>
50 #include <linux/uaccess.h>
53 #include <rdma/uverbs_std_types.h>
54 #include <rdma/rdma_netlink.h>
57 #include "core_priv.h"
58 #include "rdma_core.h"
60 MODULE_AUTHOR("Roland Dreier");
61 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
62 MODULE_LICENSE("Dual BSD/GPL");
65 IB_UVERBS_MAJOR = 231,
66 IB_UVERBS_BASE_MINOR = 192,
67 IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS,
68 IB_UVERBS_NUM_FIXED_MINOR = 32,
69 IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR,
72 #define IB_UVERBS_BASE_DEV MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
74 static dev_t dynamic_uverbs_dev;
75 static struct class *uverbs_class;
77 static DEFINE_IDA(uverbs_ida);
78 static void ib_uverbs_add_one(struct ib_device *device);
79 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
82 * Must be called with the ufile->device->disassociate_srcu held, and the lock
83 * must be held until use of the ucontext is finished.
85 struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile)
88 * We do not hold the hw_destroy_rwsem lock for this flow, instead
89 * srcu is used. It does not matter if someone races this with
90 * get_context, we get NULL or valid ucontext.
92 struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext);
94 if (!srcu_dereference(ufile->device->ib_dev,
95 &ufile->device->disassociate_srcu))
99 return ERR_PTR(-EINVAL);
103 EXPORT_SYMBOL(ib_uverbs_get_ucontext_file);
105 int uverbs_dealloc_mw(struct ib_mw *mw)
107 struct ib_pd *pd = mw->pd;
110 ret = mw->device->ops.dealloc_mw(mw);
112 atomic_dec(&pd->usecnt);
116 static void ib_uverbs_release_dev(struct device *device)
118 struct ib_uverbs_device *dev =
119 container_of(device, struct ib_uverbs_device, dev);
121 uverbs_destroy_api(dev->uapi);
122 cleanup_srcu_struct(&dev->disassociate_srcu);
123 mutex_destroy(&dev->lists_mutex);
124 mutex_destroy(&dev->xrcd_tree_mutex);
128 void ib_uverbs_release_ucq(struct ib_uverbs_completion_event_file *ev_file,
129 struct ib_ucq_object *uobj)
131 struct ib_uverbs_event *evt, *tmp;
134 spin_lock_irq(&ev_file->ev_queue.lock);
135 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
136 list_del(&evt->list);
139 spin_unlock_irq(&ev_file->ev_queue.lock);
141 uverbs_uobject_put(&ev_file->uobj);
144 ib_uverbs_release_uevent(&uobj->uevent);
147 void ib_uverbs_release_uevent(struct ib_uevent_object *uobj)
149 struct ib_uverbs_async_event_file *async_file =
150 READ_ONCE(uobj->uobject.ufile->async_file);
151 struct ib_uverbs_event *evt, *tmp;
156 spin_lock_irq(&async_file->ev_queue.lock);
157 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
158 list_del(&evt->list);
161 spin_unlock_irq(&async_file->ev_queue.lock);
164 void ib_uverbs_detach_umcast(struct ib_qp *qp,
165 struct ib_uqp_object *uobj)
167 struct ib_uverbs_mcast_entry *mcast, *tmp;
169 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
170 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
171 list_del(&mcast->list);
176 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
178 complete(&dev->comp);
181 void ib_uverbs_release_file(struct kref *ref)
183 struct ib_uverbs_file *file =
184 container_of(ref, struct ib_uverbs_file, ref);
185 struct ib_device *ib_dev;
188 release_ufile_idr_uobject(file);
190 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
191 ib_dev = srcu_dereference(file->device->ib_dev,
192 &file->device->disassociate_srcu);
193 if (ib_dev && !ib_dev->ops.disassociate_ucontext)
194 module_put(ib_dev->ops.owner);
195 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
197 if (atomic_dec_and_test(&file->device->refcount))
198 ib_uverbs_comp_dev(file->device);
200 if (file->async_file)
201 uverbs_uobject_put(&file->async_file->uobj);
202 put_device(&file->device->dev);
204 if (file->disassociate_page)
205 __free_pages(file->disassociate_page, 0);
206 mutex_destroy(&file->umap_lock);
207 mutex_destroy(&file->ucontext_lock);
211 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
212 struct file *filp, char __user *buf,
213 size_t count, loff_t *pos,
216 struct ib_uverbs_event *event;
219 spin_lock_irq(&ev_queue->lock);
221 while (list_empty(&ev_queue->event_list)) {
222 spin_unlock_irq(&ev_queue->lock);
224 if (filp->f_flags & O_NONBLOCK)
227 if (wait_event_interruptible(ev_queue->poll_wait,
228 (!list_empty(&ev_queue->event_list) ||
229 ev_queue->is_closed)))
232 spin_lock_irq(&ev_queue->lock);
234 /* If device was disassociated and no event exists set an error */
235 if (list_empty(&ev_queue->event_list) && ev_queue->is_closed) {
236 spin_unlock_irq(&ev_queue->lock);
241 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
243 if (eventsz > count) {
247 list_del(ev_queue->event_list.next);
248 if (event->counter) {
250 list_del(&event->obj_list);
254 spin_unlock_irq(&ev_queue->lock);
257 if (copy_to_user(buf, event, eventsz))
268 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
269 size_t count, loff_t *pos)
271 struct ib_uverbs_async_event_file *file = filp->private_data;
273 return ib_uverbs_event_read(&file->ev_queue, filp, buf, count, pos,
274 sizeof(struct ib_uverbs_async_event_desc));
277 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
278 size_t count, loff_t *pos)
280 struct ib_uverbs_completion_event_file *comp_ev_file =
283 return ib_uverbs_event_read(&comp_ev_file->ev_queue, filp, buf, count,
285 sizeof(struct ib_uverbs_comp_event_desc));
288 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
290 struct poll_table_struct *wait)
292 __poll_t pollflags = 0;
294 poll_wait(filp, &ev_queue->poll_wait, wait);
296 spin_lock_irq(&ev_queue->lock);
297 if (!list_empty(&ev_queue->event_list))
298 pollflags = EPOLLIN | EPOLLRDNORM;
299 spin_unlock_irq(&ev_queue->lock);
304 static __poll_t ib_uverbs_async_event_poll(struct file *filp,
305 struct poll_table_struct *wait)
307 struct ib_uverbs_async_event_file *file = filp->private_data;
309 return ib_uverbs_event_poll(&file->ev_queue, filp, wait);
312 static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
313 struct poll_table_struct *wait)
315 struct ib_uverbs_completion_event_file *comp_ev_file =
318 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
321 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
323 struct ib_uverbs_async_event_file *file = filp->private_data;
325 return fasync_helper(fd, filp, on, &file->ev_queue.async_queue);
328 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
330 struct ib_uverbs_completion_event_file *comp_ev_file =
333 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
336 const struct file_operations uverbs_event_fops = {
337 .owner = THIS_MODULE,
338 .read = ib_uverbs_comp_event_read,
339 .poll = ib_uverbs_comp_event_poll,
340 .release = uverbs_uobject_fd_release,
341 .fasync = ib_uverbs_comp_event_fasync,
345 const struct file_operations uverbs_async_event_fops = {
346 .owner = THIS_MODULE,
347 .read = ib_uverbs_async_event_read,
348 .poll = ib_uverbs_async_event_poll,
349 .release = uverbs_uobject_fd_release,
350 .fasync = ib_uverbs_async_event_fasync,
354 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
356 struct ib_uverbs_event_queue *ev_queue = cq_context;
357 struct ib_ucq_object *uobj;
358 struct ib_uverbs_event *entry;
364 spin_lock_irqsave(&ev_queue->lock, flags);
365 if (ev_queue->is_closed) {
366 spin_unlock_irqrestore(&ev_queue->lock, flags);
370 entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
372 spin_unlock_irqrestore(&ev_queue->lock, flags);
378 entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle;
379 entry->counter = &uobj->comp_events_reported;
381 list_add_tail(&entry->list, &ev_queue->event_list);
382 list_add_tail(&entry->obj_list, &uobj->comp_list);
383 spin_unlock_irqrestore(&ev_queue->lock, flags);
385 wake_up_interruptible(&ev_queue->poll_wait);
386 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
390 ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file,
391 __u64 element, __u64 event, struct list_head *obj_list,
394 struct ib_uverbs_event *entry;
400 spin_lock_irqsave(&async_file->ev_queue.lock, flags);
401 if (async_file->ev_queue.is_closed) {
402 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
406 entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
408 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
412 entry->desc.async.element = element;
413 entry->desc.async.event_type = event;
414 entry->desc.async.reserved = 0;
415 entry->counter = counter;
417 list_add_tail(&entry->list, &async_file->ev_queue.event_list);
419 list_add_tail(&entry->obj_list, obj_list);
420 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
422 wake_up_interruptible(&async_file->ev_queue.poll_wait);
423 kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN);
426 static void uverbs_uobj_event(struct ib_uevent_object *eobj,
427 struct ib_event *event)
429 ib_uverbs_async_handler(READ_ONCE(eobj->uobject.ufile->async_file),
430 eobj->uobject.user_handle, event->event,
431 &eobj->event_list, &eobj->events_reported);
434 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
436 uverbs_uobj_event(&event->element.cq->uobject->uevent, event);
439 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
441 /* for XRC target qp's, check that qp is live */
442 if (!event->element.qp->uobject)
445 uverbs_uobj_event(&event->element.qp->uobject->uevent, event);
448 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
450 uverbs_uobj_event(&event->element.wq->uobject->uevent, event);
453 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
455 uverbs_uobj_event(&event->element.srq->uobject->uevent, event);
458 static void ib_uverbs_event_handler(struct ib_event_handler *handler,
459 struct ib_event *event)
461 ib_uverbs_async_handler(
462 container_of(handler, struct ib_uverbs_async_event_file,
464 event->element.port_num, event->event, NULL, NULL);
467 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
469 spin_lock_init(&ev_queue->lock);
470 INIT_LIST_HEAD(&ev_queue->event_list);
471 init_waitqueue_head(&ev_queue->poll_wait);
472 ev_queue->is_closed = 0;
473 ev_queue->async_queue = NULL;
476 void ib_uverbs_init_async_event_file(
477 struct ib_uverbs_async_event_file *async_file)
479 struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile;
480 struct ib_device *ib_dev = async_file->uobj.context->device;
482 ib_uverbs_init_event_queue(&async_file->ev_queue);
484 /* The first async_event_file becomes the default one for the file. */
485 mutex_lock(&uverbs_file->ucontext_lock);
486 if (!uverbs_file->async_file) {
487 /* Pairs with the put in ib_uverbs_release_file */
488 uverbs_uobject_get(&async_file->uobj);
489 smp_store_release(&uverbs_file->async_file, async_file);
491 mutex_unlock(&uverbs_file->ucontext_lock);
493 INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev,
494 ib_uverbs_event_handler);
495 ib_register_event_handler(&async_file->event_handler);
498 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
499 struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count,
500 const struct uverbs_api_write_method *method_elm)
502 if (method_elm->is_ex) {
503 count -= sizeof(*hdr) + sizeof(*ex_hdr);
505 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
508 if (hdr->in_words * 8 < method_elm->req_size)
511 if (ex_hdr->cmd_hdr_reserved)
514 if (ex_hdr->response) {
515 if (!hdr->out_words && !ex_hdr->provider_out_words)
518 if (hdr->out_words * 8 < method_elm->resp_size)
521 if (!access_ok(u64_to_user_ptr(ex_hdr->response),
522 (hdr->out_words + ex_hdr->provider_out_words) * 8))
525 if (hdr->out_words || ex_hdr->provider_out_words)
532 /* not extended command */
533 if (hdr->in_words * 4 != count)
536 if (count < method_elm->req_size + sizeof(hdr)) {
538 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
539 * with a 16 byte write instead of 24. Old kernels didn't
540 * check the size so they allowed this. Now that the size is
541 * checked provide a compatibility work around to not break
544 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
551 if (hdr->out_words * 4 < method_elm->resp_size)
557 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
558 size_t count, loff_t *pos)
560 struct ib_uverbs_file *file = filp->private_data;
561 const struct uverbs_api_write_method *method_elm;
562 struct uverbs_api *uapi = file->device->uapi;
563 struct ib_uverbs_ex_cmd_hdr ex_hdr;
564 struct ib_uverbs_cmd_hdr hdr;
565 struct uverbs_attr_bundle bundle;
569 if (!ib_safe_file_access(filp)) {
570 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
571 task_tgid_vnr(current), current->comm);
575 if (count < sizeof(hdr))
578 if (copy_from_user(&hdr, buf, sizeof(hdr)))
581 method_elm = uapi_get_method(uapi, hdr.command);
582 if (IS_ERR(method_elm))
583 return PTR_ERR(method_elm);
585 if (method_elm->is_ex) {
586 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
588 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
592 ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
596 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
600 memset(bundle.attr_present, 0, sizeof(bundle.attr_present));
602 bundle.context = NULL; /* only valid if bundle has uobject */
603 if (!method_elm->is_ex) {
604 size_t in_len = hdr.in_words * 4 - sizeof(hdr);
605 size_t out_len = hdr.out_words * 4;
608 if (method_elm->has_udata) {
609 bundle.driver_udata.inlen =
610 in_len - method_elm->req_size;
611 in_len = method_elm->req_size;
612 if (bundle.driver_udata.inlen)
613 bundle.driver_udata.inbuf = buf + in_len;
615 bundle.driver_udata.inbuf = NULL;
617 memset(&bundle.driver_udata, 0,
618 sizeof(bundle.driver_udata));
621 if (method_elm->has_resp) {
623 * The macros check that if has_resp is set
624 * then the command request structure starts
625 * with a '__aligned u64 response' member.
627 ret = get_user(response, (const u64 __user *)buf);
631 if (method_elm->has_udata) {
632 bundle.driver_udata.outlen =
633 out_len - method_elm->resp_size;
634 out_len = method_elm->resp_size;
635 if (bundle.driver_udata.outlen)
636 bundle.driver_udata.outbuf =
637 u64_to_user_ptr(response +
640 bundle.driver_udata.outbuf = NULL;
643 bundle.driver_udata.outlen = 0;
644 bundle.driver_udata.outbuf = NULL;
647 ib_uverbs_init_udata_buf_or_null(
648 &bundle.ucore, buf, u64_to_user_ptr(response),
651 buf += sizeof(ex_hdr);
653 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
654 u64_to_user_ptr(ex_hdr.response),
655 hdr.in_words * 8, hdr.out_words * 8);
657 ib_uverbs_init_udata_buf_or_null(
658 &bundle.driver_udata, buf + bundle.ucore.inlen,
659 u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
660 ex_hdr.provider_in_words * 8,
661 ex_hdr.provider_out_words * 8);
665 ret = method_elm->handler(&bundle);
667 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
668 return (ret) ? : count;
671 static const struct vm_operations_struct rdma_umap_ops;
673 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
675 struct ib_uverbs_file *file = filp->private_data;
676 struct ib_ucontext *ucontext;
680 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
681 ucontext = ib_uverbs_get_ucontext_file(file);
682 if (IS_ERR(ucontext)) {
683 ret = PTR_ERR(ucontext);
686 vma->vm_ops = &rdma_umap_ops;
687 ret = ucontext->device->ops.mmap(ucontext, vma);
689 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
694 * The VMA has been dup'd, initialize the vm_private_data with a new tracking
697 static void rdma_umap_open(struct vm_area_struct *vma)
699 struct ib_uverbs_file *ufile = vma->vm_file->private_data;
700 struct rdma_umap_priv *opriv = vma->vm_private_data;
701 struct rdma_umap_priv *priv;
706 /* We are racing with disassociation */
707 if (!down_read_trylock(&ufile->hw_destroy_rwsem))
710 * Disassociation already completed, the VMA should already be zapped.
712 if (!ufile->ucontext)
715 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
718 rdma_umap_priv_init(priv, vma, opriv->entry);
720 up_read(&ufile->hw_destroy_rwsem);
724 up_read(&ufile->hw_destroy_rwsem);
727 * We can't allow the VMA to be created with the actual IO pages, that
728 * would break our API contract, and it can't be stopped at this
731 vma->vm_private_data = NULL;
732 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
735 static void rdma_umap_close(struct vm_area_struct *vma)
737 struct ib_uverbs_file *ufile = vma->vm_file->private_data;
738 struct rdma_umap_priv *priv = vma->vm_private_data;
744 * The vma holds a reference on the struct file that created it, which
745 * in turn means that the ib_uverbs_file is guaranteed to exist at
748 mutex_lock(&ufile->umap_lock);
750 rdma_user_mmap_entry_put(priv->entry);
752 list_del(&priv->list);
753 mutex_unlock(&ufile->umap_lock);
758 * Once the zap_vma_ptes has been called touches to the VMA will come here and
759 * we return a dummy writable zero page for all the pfns.
761 static vm_fault_t rdma_umap_fault(struct vm_fault *vmf)
763 struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data;
764 struct rdma_umap_priv *priv = vmf->vma->vm_private_data;
768 return VM_FAULT_SIGBUS;
770 /* Read only pages can just use the system zero page. */
771 if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) {
772 vmf->page = ZERO_PAGE(vmf->address);
777 mutex_lock(&ufile->umap_lock);
778 if (!ufile->disassociate_page)
779 ufile->disassociate_page =
780 alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0);
782 if (ufile->disassociate_page) {
784 * This VMA is forced to always be shared so this doesn't have
785 * to worry about COW.
787 vmf->page = ufile->disassociate_page;
790 ret = VM_FAULT_SIGBUS;
792 mutex_unlock(&ufile->umap_lock);
797 static const struct vm_operations_struct rdma_umap_ops = {
798 .open = rdma_umap_open,
799 .close = rdma_umap_close,
800 .fault = rdma_umap_fault,
803 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
805 struct rdma_umap_priv *priv, *next_priv;
807 lockdep_assert_held(&ufile->hw_destroy_rwsem);
810 struct mm_struct *mm = NULL;
812 /* Get an arbitrary mm pointer that hasn't been cleaned yet */
813 mutex_lock(&ufile->umap_lock);
814 while (!list_empty(&ufile->umaps)) {
817 priv = list_first_entry(&ufile->umaps,
818 struct rdma_umap_priv, list);
819 mm = priv->vma->vm_mm;
820 ret = mmget_not_zero(mm);
822 list_del_init(&priv->list);
828 mutex_unlock(&ufile->umap_lock);
833 * The umap_lock is nested under mmap_sem since it used within
834 * the vma_ops callbacks, so we have to clean the list one mm
835 * at a time to get the lock ordering right. Typically there
836 * will only be one mm, so no big deal.
838 down_read(&mm->mmap_sem);
839 if (!mmget_still_valid(mm))
841 mutex_lock(&ufile->umap_lock);
842 list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
844 struct vm_area_struct *vma = priv->vma;
846 if (vma->vm_mm != mm)
848 list_del_init(&priv->list);
850 zap_vma_ptes(vma, vma->vm_start,
851 vma->vm_end - vma->vm_start);
854 rdma_user_mmap_entry_put(priv->entry);
858 mutex_unlock(&ufile->umap_lock);
860 up_read(&mm->mmap_sem);
866 * ib_uverbs_open() does not need the BKL:
868 * - the ib_uverbs_device structures are properly reference counted and
869 * everything else is purely local to the file being created, so
870 * races against other open calls are not a problem;
871 * - there is no ioctl method to race against;
872 * - the open method will either immediately run -ENXIO, or all
873 * required initialization will be done.
875 static int ib_uverbs_open(struct inode *inode, struct file *filp)
877 struct ib_uverbs_device *dev;
878 struct ib_uverbs_file *file;
879 struct ib_device *ib_dev;
881 int module_dependent;
884 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
885 if (!atomic_inc_not_zero(&dev->refcount))
888 get_device(&dev->dev);
889 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
890 mutex_lock(&dev->lists_mutex);
891 ib_dev = srcu_dereference(dev->ib_dev,
892 &dev->disassociate_srcu);
898 if (!rdma_dev_access_netns(ib_dev, current->nsproxy->net_ns)) {
903 /* In case IB device supports disassociate ucontext, there is no hard
904 * dependency between uverbs device and its low level device.
906 module_dependent = !(ib_dev->ops.disassociate_ucontext);
908 if (module_dependent) {
909 if (!try_module_get(ib_dev->ops.owner)) {
915 file = kzalloc(sizeof(*file), GFP_KERNEL);
918 if (module_dependent)
925 kref_init(&file->ref);
926 mutex_init(&file->ucontext_lock);
928 spin_lock_init(&file->uobjects_lock);
929 INIT_LIST_HEAD(&file->uobjects);
930 init_rwsem(&file->hw_destroy_rwsem);
931 mutex_init(&file->umap_lock);
932 INIT_LIST_HEAD(&file->umaps);
934 filp->private_data = file;
935 list_add_tail(&file->list, &dev->uverbs_file_list);
936 mutex_unlock(&dev->lists_mutex);
937 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
939 setup_ufile_idr_uobject(file);
941 return stream_open(inode, filp);
944 module_put(ib_dev->ops.owner);
947 mutex_unlock(&dev->lists_mutex);
948 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
949 if (atomic_dec_and_test(&dev->refcount))
950 ib_uverbs_comp_dev(dev);
952 put_device(&dev->dev);
956 static int ib_uverbs_close(struct inode *inode, struct file *filp)
958 struct ib_uverbs_file *file = filp->private_data;
960 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
962 mutex_lock(&file->device->lists_mutex);
963 list_del_init(&file->list);
964 mutex_unlock(&file->device->lists_mutex);
966 kref_put(&file->ref, ib_uverbs_release_file);
971 static const struct file_operations uverbs_fops = {
972 .owner = THIS_MODULE,
973 .write = ib_uverbs_write,
974 .open = ib_uverbs_open,
975 .release = ib_uverbs_close,
977 .unlocked_ioctl = ib_uverbs_ioctl,
978 .compat_ioctl = compat_ptr_ioctl,
981 static const struct file_operations uverbs_mmap_fops = {
982 .owner = THIS_MODULE,
983 .write = ib_uverbs_write,
984 .mmap = ib_uverbs_mmap,
985 .open = ib_uverbs_open,
986 .release = ib_uverbs_close,
988 .unlocked_ioctl = ib_uverbs_ioctl,
989 .compat_ioctl = compat_ptr_ioctl,
992 static int ib_uverbs_get_nl_info(struct ib_device *ibdev, void *client_data,
993 struct ib_client_nl_info *res)
995 struct ib_uverbs_device *uverbs_dev = client_data;
1001 res->abi = ibdev->ops.uverbs_abi_ver;
1002 res->cdev = &uverbs_dev->dev;
1005 * To support DRIVER_ID binding in userspace some of the driver need
1006 * upgrading to expose their PCI dependent revision information
1007 * through get_context instead of relying on modalias matching. When
1008 * the drivers are fixed they can drop this flag.
1010 if (!ibdev->ops.uverbs_no_driver_id_binding) {
1011 ret = nla_put_u32(res->nl_msg, RDMA_NLDEV_ATTR_UVERBS_DRIVER_ID,
1012 ibdev->ops.driver_id);
1019 static struct ib_client uverbs_client = {
1021 .no_kverbs_req = true,
1022 .add = ib_uverbs_add_one,
1023 .remove = ib_uverbs_remove_one,
1024 .get_nl_info = ib_uverbs_get_nl_info,
1026 MODULE_ALIAS_RDMA_CLIENT("uverbs");
1028 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1031 struct ib_uverbs_device *dev =
1032 container_of(device, struct ib_uverbs_device, dev);
1035 struct ib_device *ib_dev;
1037 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1038 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1040 ret = sprintf(buf, "%s\n", dev_name(&ib_dev->dev));
1041 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1045 static DEVICE_ATTR_RO(ibdev);
1047 static ssize_t abi_version_show(struct device *device,
1048 struct device_attribute *attr, char *buf)
1050 struct ib_uverbs_device *dev =
1051 container_of(device, struct ib_uverbs_device, dev);
1054 struct ib_device *ib_dev;
1056 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1057 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1059 ret = sprintf(buf, "%u\n", ib_dev->ops.uverbs_abi_ver);
1060 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1064 static DEVICE_ATTR_RO(abi_version);
1066 static struct attribute *ib_dev_attrs[] = {
1067 &dev_attr_abi_version.attr,
1068 &dev_attr_ibdev.attr,
1072 static const struct attribute_group dev_attr_group = {
1073 .attrs = ib_dev_attrs,
1076 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1077 __stringify(IB_USER_VERBS_ABI_VERSION));
1079 static int ib_uverbs_create_uapi(struct ib_device *device,
1080 struct ib_uverbs_device *uverbs_dev)
1082 struct uverbs_api *uapi;
1084 uapi = uverbs_alloc_api(device);
1086 return PTR_ERR(uapi);
1088 uverbs_dev->uapi = uapi;
1092 static void ib_uverbs_add_one(struct ib_device *device)
1096 struct ib_uverbs_device *uverbs_dev;
1099 if (!device->ops.alloc_ucontext)
1102 uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1106 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1112 device_initialize(&uverbs_dev->dev);
1113 uverbs_dev->dev.class = uverbs_class;
1114 uverbs_dev->dev.parent = device->dev.parent;
1115 uverbs_dev->dev.release = ib_uverbs_release_dev;
1116 uverbs_dev->groups[0] = &dev_attr_group;
1117 uverbs_dev->dev.groups = uverbs_dev->groups;
1118 atomic_set(&uverbs_dev->refcount, 1);
1119 init_completion(&uverbs_dev->comp);
1120 uverbs_dev->xrcd_tree = RB_ROOT;
1121 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1122 mutex_init(&uverbs_dev->lists_mutex);
1123 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1124 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1125 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1127 devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1131 uverbs_dev->devnum = devnum;
1132 if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1133 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1135 base = IB_UVERBS_BASE_DEV + devnum;
1137 if (ib_uverbs_create_uapi(device, uverbs_dev))
1140 uverbs_dev->dev.devt = base;
1141 dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1143 cdev_init(&uverbs_dev->cdev,
1144 device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1145 uverbs_dev->cdev.owner = THIS_MODULE;
1147 ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1151 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1155 ida_free(&uverbs_ida, devnum);
1157 if (atomic_dec_and_test(&uverbs_dev->refcount))
1158 ib_uverbs_comp_dev(uverbs_dev);
1159 wait_for_completion(&uverbs_dev->comp);
1160 put_device(&uverbs_dev->dev);
1164 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1165 struct ib_device *ib_dev)
1167 struct ib_uverbs_file *file;
1169 /* Pending running commands to terminate */
1170 uverbs_disassociate_api_pre(uverbs_dev);
1172 mutex_lock(&uverbs_dev->lists_mutex);
1173 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1174 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1175 struct ib_uverbs_file, list);
1176 list_del_init(&file->list);
1177 kref_get(&file->ref);
1179 /* We must release the mutex before going ahead and calling
1180 * uverbs_cleanup_ufile, as it might end up indirectly calling
1181 * uverbs_close, for example due to freeing the resources (e.g
1184 mutex_unlock(&uverbs_dev->lists_mutex);
1186 ib_uverbs_async_handler(READ_ONCE(file->async_file), 0,
1187 IB_EVENT_DEVICE_FATAL, NULL, NULL);
1189 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1190 kref_put(&file->ref, ib_uverbs_release_file);
1192 mutex_lock(&uverbs_dev->lists_mutex);
1194 mutex_unlock(&uverbs_dev->lists_mutex);
1196 uverbs_disassociate_api(uverbs_dev->uapi);
1199 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1201 struct ib_uverbs_device *uverbs_dev = client_data;
1202 int wait_clients = 1;
1207 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1208 ida_free(&uverbs_ida, uverbs_dev->devnum);
1210 if (device->ops.disassociate_ucontext) {
1211 /* We disassociate HW resources and immediately return.
1212 * Userspace will see a EIO errno for all future access.
1213 * Upon returning, ib_device may be freed internally and is not
1215 * uverbs_device is still available until all clients close
1216 * their files, then the uverbs device ref count will be zero
1217 * and its resources will be freed.
1218 * Note: At this point no more files can be opened since the
1219 * cdev was deleted, however active clients can still issue
1220 * commands and close their open files.
1222 ib_uverbs_free_hw_resources(uverbs_dev, device);
1226 if (atomic_dec_and_test(&uverbs_dev->refcount))
1227 ib_uverbs_comp_dev(uverbs_dev);
1229 wait_for_completion(&uverbs_dev->comp);
1231 put_device(&uverbs_dev->dev);
1234 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1238 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1241 static int __init ib_uverbs_init(void)
1245 ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1246 IB_UVERBS_NUM_FIXED_MINOR,
1247 "infiniband_verbs");
1249 pr_err("user_verbs: couldn't register device number\n");
1253 ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1254 IB_UVERBS_NUM_DYNAMIC_MINOR,
1255 "infiniband_verbs");
1257 pr_err("couldn't register dynamic device number\n");
1261 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1262 if (IS_ERR(uverbs_class)) {
1263 ret = PTR_ERR(uverbs_class);
1264 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1268 uverbs_class->devnode = uverbs_devnode;
1270 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1272 pr_err("user_verbs: couldn't create abi_version attribute\n");
1276 ret = ib_register_client(&uverbs_client);
1278 pr_err("user_verbs: couldn't register client\n");
1285 class_destroy(uverbs_class);
1288 unregister_chrdev_region(dynamic_uverbs_dev,
1289 IB_UVERBS_NUM_DYNAMIC_MINOR);
1292 unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1293 IB_UVERBS_NUM_FIXED_MINOR);
1299 static void __exit ib_uverbs_cleanup(void)
1301 ib_unregister_client(&uverbs_client);
1302 class_destroy(uverbs_class);
1303 unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1304 IB_UVERBS_NUM_FIXED_MINOR);
1305 unregister_chrdev_region(dynamic_uverbs_dev,
1306 IB_UVERBS_NUM_DYNAMIC_MINOR);
1307 mmu_notifier_synchronize();
1310 module_init(ib_uverbs_init);
1311 module_exit(ib_uverbs_cleanup);