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;
76 static DEFINE_IDA(uverbs_ida);
77 static int ib_uverbs_add_one(struct ib_device *device);
78 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
80 static char *uverbs_devnode(const struct device *dev, umode_t *mode)
84 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
87 static const struct class uverbs_class = {
88 .name = "infiniband_verbs",
89 .devnode = uverbs_devnode,
93 * Must be called with the ufile->device->disassociate_srcu held, and the lock
94 * must be held until use of the ucontext is finished.
96 struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile)
99 * We do not hold the hw_destroy_rwsem lock for this flow, instead
100 * srcu is used. It does not matter if someone races this with
101 * get_context, we get NULL or valid ucontext.
103 struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext);
105 if (!srcu_dereference(ufile->device->ib_dev,
106 &ufile->device->disassociate_srcu))
107 return ERR_PTR(-EIO);
110 return ERR_PTR(-EINVAL);
114 EXPORT_SYMBOL(ib_uverbs_get_ucontext_file);
116 int uverbs_dealloc_mw(struct ib_mw *mw)
118 struct ib_pd *pd = mw->pd;
121 ret = mw->device->ops.dealloc_mw(mw);
125 atomic_dec(&pd->usecnt);
130 static void ib_uverbs_release_dev(struct device *device)
132 struct ib_uverbs_device *dev =
133 container_of(device, struct ib_uverbs_device, dev);
135 uverbs_destroy_api(dev->uapi);
136 cleanup_srcu_struct(&dev->disassociate_srcu);
137 mutex_destroy(&dev->lists_mutex);
138 mutex_destroy(&dev->xrcd_tree_mutex);
142 void ib_uverbs_release_ucq(struct ib_uverbs_completion_event_file *ev_file,
143 struct ib_ucq_object *uobj)
145 struct ib_uverbs_event *evt, *tmp;
148 spin_lock_irq(&ev_file->ev_queue.lock);
149 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
150 list_del(&evt->list);
153 spin_unlock_irq(&ev_file->ev_queue.lock);
155 uverbs_uobject_put(&ev_file->uobj);
158 ib_uverbs_release_uevent(&uobj->uevent);
161 void ib_uverbs_release_uevent(struct ib_uevent_object *uobj)
163 struct ib_uverbs_async_event_file *async_file = uobj->event_file;
164 struct ib_uverbs_event *evt, *tmp;
169 spin_lock_irq(&async_file->ev_queue.lock);
170 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
171 list_del(&evt->list);
174 spin_unlock_irq(&async_file->ev_queue.lock);
175 uverbs_uobject_put(&async_file->uobj);
178 void ib_uverbs_detach_umcast(struct ib_qp *qp,
179 struct ib_uqp_object *uobj)
181 struct ib_uverbs_mcast_entry *mcast, *tmp;
183 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
184 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
185 list_del(&mcast->list);
190 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
192 complete(&dev->comp);
195 void ib_uverbs_release_file(struct kref *ref)
197 struct ib_uverbs_file *file =
198 container_of(ref, struct ib_uverbs_file, ref);
199 struct ib_device *ib_dev;
202 release_ufile_idr_uobject(file);
204 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
205 ib_dev = srcu_dereference(file->device->ib_dev,
206 &file->device->disassociate_srcu);
207 if (ib_dev && !ib_dev->ops.disassociate_ucontext)
208 module_put(ib_dev->ops.owner);
209 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
211 if (refcount_dec_and_test(&file->device->refcount))
212 ib_uverbs_comp_dev(file->device);
214 if (file->default_async_file)
215 uverbs_uobject_put(&file->default_async_file->uobj);
216 put_device(&file->device->dev);
218 if (file->disassociate_page)
219 __free_pages(file->disassociate_page, 0);
220 mutex_destroy(&file->umap_lock);
221 mutex_destroy(&file->ucontext_lock);
225 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
226 struct file *filp, char __user *buf,
227 size_t count, loff_t *pos,
230 struct ib_uverbs_event *event;
233 spin_lock_irq(&ev_queue->lock);
235 while (list_empty(&ev_queue->event_list)) {
236 if (ev_queue->is_closed) {
237 spin_unlock_irq(&ev_queue->lock);
241 spin_unlock_irq(&ev_queue->lock);
242 if (filp->f_flags & O_NONBLOCK)
245 if (wait_event_interruptible(ev_queue->poll_wait,
246 (!list_empty(&ev_queue->event_list) ||
247 ev_queue->is_closed)))
250 spin_lock_irq(&ev_queue->lock);
253 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
255 if (eventsz > count) {
259 list_del(ev_queue->event_list.next);
260 if (event->counter) {
262 list_del(&event->obj_list);
266 spin_unlock_irq(&ev_queue->lock);
269 if (copy_to_user(buf, event, eventsz))
280 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
281 size_t count, loff_t *pos)
283 struct ib_uverbs_async_event_file *file = filp->private_data;
285 return ib_uverbs_event_read(&file->ev_queue, filp, buf, count, pos,
286 sizeof(struct ib_uverbs_async_event_desc));
289 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
290 size_t count, loff_t *pos)
292 struct ib_uverbs_completion_event_file *comp_ev_file =
295 return ib_uverbs_event_read(&comp_ev_file->ev_queue, filp, buf, count,
297 sizeof(struct ib_uverbs_comp_event_desc));
300 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
302 struct poll_table_struct *wait)
304 __poll_t pollflags = 0;
306 poll_wait(filp, &ev_queue->poll_wait, wait);
308 spin_lock_irq(&ev_queue->lock);
309 if (!list_empty(&ev_queue->event_list))
310 pollflags = EPOLLIN | EPOLLRDNORM;
311 else if (ev_queue->is_closed)
312 pollflags = EPOLLERR;
313 spin_unlock_irq(&ev_queue->lock);
318 static __poll_t ib_uverbs_async_event_poll(struct file *filp,
319 struct poll_table_struct *wait)
321 struct ib_uverbs_async_event_file *file = filp->private_data;
323 return ib_uverbs_event_poll(&file->ev_queue, filp, wait);
326 static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
327 struct poll_table_struct *wait)
329 struct ib_uverbs_completion_event_file *comp_ev_file =
332 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
335 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
337 struct ib_uverbs_async_event_file *file = filp->private_data;
339 return fasync_helper(fd, filp, on, &file->ev_queue.async_queue);
342 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
344 struct ib_uverbs_completion_event_file *comp_ev_file =
347 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
350 const struct file_operations uverbs_event_fops = {
351 .owner = THIS_MODULE,
352 .read = ib_uverbs_comp_event_read,
353 .poll = ib_uverbs_comp_event_poll,
354 .release = uverbs_uobject_fd_release,
355 .fasync = ib_uverbs_comp_event_fasync,
359 const struct file_operations uverbs_async_event_fops = {
360 .owner = THIS_MODULE,
361 .read = ib_uverbs_async_event_read,
362 .poll = ib_uverbs_async_event_poll,
363 .release = uverbs_async_event_release,
364 .fasync = ib_uverbs_async_event_fasync,
368 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
370 struct ib_uverbs_event_queue *ev_queue = cq_context;
371 struct ib_ucq_object *uobj;
372 struct ib_uverbs_event *entry;
378 spin_lock_irqsave(&ev_queue->lock, flags);
379 if (ev_queue->is_closed) {
380 spin_unlock_irqrestore(&ev_queue->lock, flags);
384 entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
386 spin_unlock_irqrestore(&ev_queue->lock, flags);
392 entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle;
393 entry->counter = &uobj->comp_events_reported;
395 list_add_tail(&entry->list, &ev_queue->event_list);
396 list_add_tail(&entry->obj_list, &uobj->comp_list);
397 spin_unlock_irqrestore(&ev_queue->lock, flags);
399 wake_up_interruptible(&ev_queue->poll_wait);
400 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
403 void ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file,
404 __u64 element, __u64 event,
405 struct list_head *obj_list, u32 *counter)
407 struct ib_uverbs_event *entry;
413 spin_lock_irqsave(&async_file->ev_queue.lock, flags);
414 if (async_file->ev_queue.is_closed) {
415 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
419 entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
421 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
425 entry->desc.async.element = element;
426 entry->desc.async.event_type = event;
427 entry->desc.async.reserved = 0;
428 entry->counter = counter;
430 list_add_tail(&entry->list, &async_file->ev_queue.event_list);
432 list_add_tail(&entry->obj_list, obj_list);
433 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
435 wake_up_interruptible(&async_file->ev_queue.poll_wait);
436 kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN);
439 static void uverbs_uobj_event(struct ib_uevent_object *eobj,
440 struct ib_event *event)
442 ib_uverbs_async_handler(eobj->event_file,
443 eobj->uobject.user_handle, event->event,
444 &eobj->event_list, &eobj->events_reported);
447 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
449 uverbs_uobj_event(&event->element.cq->uobject->uevent, event);
452 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
454 /* for XRC target qp's, check that qp is live */
455 if (!event->element.qp->uobject)
458 uverbs_uobj_event(&event->element.qp->uobject->uevent, event);
461 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
463 uverbs_uobj_event(&event->element.wq->uobject->uevent, event);
466 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
468 uverbs_uobj_event(&event->element.srq->uobject->uevent, event);
471 static void ib_uverbs_event_handler(struct ib_event_handler *handler,
472 struct ib_event *event)
474 ib_uverbs_async_handler(
475 container_of(handler, struct ib_uverbs_async_event_file,
477 event->element.port_num, event->event, NULL, NULL);
480 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
482 spin_lock_init(&ev_queue->lock);
483 INIT_LIST_HEAD(&ev_queue->event_list);
484 init_waitqueue_head(&ev_queue->poll_wait);
485 ev_queue->is_closed = 0;
486 ev_queue->async_queue = NULL;
489 void ib_uverbs_init_async_event_file(
490 struct ib_uverbs_async_event_file *async_file)
492 struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile;
493 struct ib_device *ib_dev = async_file->uobj.context->device;
495 ib_uverbs_init_event_queue(&async_file->ev_queue);
497 /* The first async_event_file becomes the default one for the file. */
498 mutex_lock(&uverbs_file->ucontext_lock);
499 if (!uverbs_file->default_async_file) {
500 /* Pairs with the put in ib_uverbs_release_file */
501 uverbs_uobject_get(&async_file->uobj);
502 smp_store_release(&uverbs_file->default_async_file, async_file);
504 mutex_unlock(&uverbs_file->ucontext_lock);
506 INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev,
507 ib_uverbs_event_handler);
508 ib_register_event_handler(&async_file->event_handler);
511 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
512 struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count,
513 const struct uverbs_api_write_method *method_elm)
515 if (method_elm->is_ex) {
516 count -= sizeof(*hdr) + sizeof(*ex_hdr);
518 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
521 if (hdr->in_words * 8 < method_elm->req_size)
524 if (ex_hdr->cmd_hdr_reserved)
527 if (ex_hdr->response) {
528 if (!hdr->out_words && !ex_hdr->provider_out_words)
531 if (hdr->out_words * 8 < method_elm->resp_size)
534 if (!access_ok(u64_to_user_ptr(ex_hdr->response),
535 (hdr->out_words + ex_hdr->provider_out_words) * 8))
538 if (hdr->out_words || ex_hdr->provider_out_words)
545 /* not extended command */
546 if (hdr->in_words * 4 != count)
549 if (count < method_elm->req_size + sizeof(*hdr)) {
551 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
552 * with a 16 byte write instead of 24. Old kernels didn't
553 * check the size so they allowed this. Now that the size is
554 * checked provide a compatibility work around to not break
557 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
564 if (hdr->out_words * 4 < method_elm->resp_size)
570 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
571 size_t count, loff_t *pos)
573 struct ib_uverbs_file *file = filp->private_data;
574 const struct uverbs_api_write_method *method_elm;
575 struct uverbs_api *uapi = file->device->uapi;
576 struct ib_uverbs_ex_cmd_hdr ex_hdr;
577 struct ib_uverbs_cmd_hdr hdr;
578 struct uverbs_attr_bundle bundle;
582 if (!ib_safe_file_access(filp)) {
583 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
584 task_tgid_vnr(current), current->comm);
588 if (count < sizeof(hdr))
591 if (copy_from_user(&hdr, buf, sizeof(hdr)))
594 method_elm = uapi_get_method(uapi, hdr.command);
595 if (IS_ERR(method_elm))
596 return PTR_ERR(method_elm);
598 if (method_elm->is_ex) {
599 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
601 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
605 ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
609 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
613 memset(bundle.attr_present, 0, sizeof(bundle.attr_present));
615 bundle.context = NULL; /* only valid if bundle has uobject */
616 bundle.uobject = NULL;
617 if (!method_elm->is_ex) {
618 size_t in_len = hdr.in_words * 4 - sizeof(hdr);
619 size_t out_len = hdr.out_words * 4;
622 if (method_elm->has_udata) {
623 bundle.driver_udata.inlen =
624 in_len - method_elm->req_size;
625 in_len = method_elm->req_size;
626 if (bundle.driver_udata.inlen)
627 bundle.driver_udata.inbuf = buf + in_len;
629 bundle.driver_udata.inbuf = NULL;
631 memset(&bundle.driver_udata, 0,
632 sizeof(bundle.driver_udata));
635 if (method_elm->has_resp) {
637 * The macros check that if has_resp is set
638 * then the command request structure starts
639 * with a '__aligned u64 response' member.
641 ret = get_user(response, (const u64 __user *)buf);
645 if (method_elm->has_udata) {
646 bundle.driver_udata.outlen =
647 out_len - method_elm->resp_size;
648 out_len = method_elm->resp_size;
649 if (bundle.driver_udata.outlen)
650 bundle.driver_udata.outbuf =
651 u64_to_user_ptr(response +
654 bundle.driver_udata.outbuf = NULL;
657 bundle.driver_udata.outlen = 0;
658 bundle.driver_udata.outbuf = NULL;
661 ib_uverbs_init_udata_buf_or_null(
662 &bundle.ucore, buf, u64_to_user_ptr(response),
665 buf += sizeof(ex_hdr);
667 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
668 u64_to_user_ptr(ex_hdr.response),
669 hdr.in_words * 8, hdr.out_words * 8);
671 ib_uverbs_init_udata_buf_or_null(
672 &bundle.driver_udata, buf + bundle.ucore.inlen,
673 u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
674 ex_hdr.provider_in_words * 8,
675 ex_hdr.provider_out_words * 8);
679 ret = method_elm->handler(&bundle);
681 uverbs_finalize_object(bundle.uobject, UVERBS_ACCESS_NEW, true,
684 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
685 return (ret) ? : count;
688 static const struct vm_operations_struct rdma_umap_ops;
690 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
692 struct ib_uverbs_file *file = filp->private_data;
693 struct ib_ucontext *ucontext;
697 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
698 ucontext = ib_uverbs_get_ucontext_file(file);
699 if (IS_ERR(ucontext)) {
700 ret = PTR_ERR(ucontext);
703 vma->vm_ops = &rdma_umap_ops;
704 ret = ucontext->device->ops.mmap(ucontext, vma);
706 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
711 * The VMA has been dup'd, initialize the vm_private_data with a new tracking
714 static void rdma_umap_open(struct vm_area_struct *vma)
716 struct ib_uverbs_file *ufile = vma->vm_file->private_data;
717 struct rdma_umap_priv *opriv = vma->vm_private_data;
718 struct rdma_umap_priv *priv;
723 /* We are racing with disassociation */
724 if (!down_read_trylock(&ufile->hw_destroy_rwsem))
727 * Disassociation already completed, the VMA should already be zapped.
729 if (!ufile->ucontext)
732 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
735 rdma_umap_priv_init(priv, vma, opriv->entry);
737 up_read(&ufile->hw_destroy_rwsem);
741 up_read(&ufile->hw_destroy_rwsem);
744 * We can't allow the VMA to be created with the actual IO pages, that
745 * would break our API contract, and it can't be stopped at this
748 vma->vm_private_data = NULL;
749 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
752 static void rdma_umap_close(struct vm_area_struct *vma)
754 struct ib_uverbs_file *ufile = vma->vm_file->private_data;
755 struct rdma_umap_priv *priv = vma->vm_private_data;
761 * The vma holds a reference on the struct file that created it, which
762 * in turn means that the ib_uverbs_file is guaranteed to exist at
765 mutex_lock(&ufile->umap_lock);
767 rdma_user_mmap_entry_put(priv->entry);
769 list_del(&priv->list);
770 mutex_unlock(&ufile->umap_lock);
775 * Once the zap_vma_ptes has been called touches to the VMA will come here and
776 * we return a dummy writable zero page for all the pfns.
778 static vm_fault_t rdma_umap_fault(struct vm_fault *vmf)
780 struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data;
781 struct rdma_umap_priv *priv = vmf->vma->vm_private_data;
785 return VM_FAULT_SIGBUS;
787 /* Read only pages can just use the system zero page. */
788 if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) {
789 vmf->page = ZERO_PAGE(vmf->address);
794 mutex_lock(&ufile->umap_lock);
795 if (!ufile->disassociate_page)
796 ufile->disassociate_page =
797 alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0);
799 if (ufile->disassociate_page) {
801 * This VMA is forced to always be shared so this doesn't have
802 * to worry about COW.
804 vmf->page = ufile->disassociate_page;
807 ret = VM_FAULT_SIGBUS;
809 mutex_unlock(&ufile->umap_lock);
814 static const struct vm_operations_struct rdma_umap_ops = {
815 .open = rdma_umap_open,
816 .close = rdma_umap_close,
817 .fault = rdma_umap_fault,
820 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
822 struct rdma_umap_priv *priv, *next_priv;
824 lockdep_assert_held(&ufile->hw_destroy_rwsem);
827 struct mm_struct *mm = NULL;
829 /* Get an arbitrary mm pointer that hasn't been cleaned yet */
830 mutex_lock(&ufile->umap_lock);
831 while (!list_empty(&ufile->umaps)) {
834 priv = list_first_entry(&ufile->umaps,
835 struct rdma_umap_priv, list);
836 mm = priv->vma->vm_mm;
837 ret = mmget_not_zero(mm);
839 list_del_init(&priv->list);
841 rdma_user_mmap_entry_put(priv->entry);
849 mutex_unlock(&ufile->umap_lock);
854 * The umap_lock is nested under mmap_lock since it used within
855 * the vma_ops callbacks, so we have to clean the list one mm
856 * at a time to get the lock ordering right. Typically there
857 * will only be one mm, so no big deal.
860 mutex_lock(&ufile->umap_lock);
861 list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
863 struct vm_area_struct *vma = priv->vma;
865 if (vma->vm_mm != mm)
867 list_del_init(&priv->list);
869 zap_vma_ptes(vma, vma->vm_start,
870 vma->vm_end - vma->vm_start);
873 rdma_user_mmap_entry_put(priv->entry);
877 mutex_unlock(&ufile->umap_lock);
878 mmap_read_unlock(mm);
884 * ib_uverbs_open() does not need the BKL:
886 * - the ib_uverbs_device structures are properly reference counted and
887 * everything else is purely local to the file being created, so
888 * races against other open calls are not a problem;
889 * - there is no ioctl method to race against;
890 * - the open method will either immediately run -ENXIO, or all
891 * required initialization will be done.
893 static int ib_uverbs_open(struct inode *inode, struct file *filp)
895 struct ib_uverbs_device *dev;
896 struct ib_uverbs_file *file;
897 struct ib_device *ib_dev;
899 int module_dependent;
902 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
903 if (!refcount_inc_not_zero(&dev->refcount))
906 get_device(&dev->dev);
907 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
908 mutex_lock(&dev->lists_mutex);
909 ib_dev = srcu_dereference(dev->ib_dev,
910 &dev->disassociate_srcu);
916 if (!rdma_dev_access_netns(ib_dev, current->nsproxy->net_ns)) {
921 /* In case IB device supports disassociate ucontext, there is no hard
922 * dependency between uverbs device and its low level device.
924 module_dependent = !(ib_dev->ops.disassociate_ucontext);
926 if (module_dependent) {
927 if (!try_module_get(ib_dev->ops.owner)) {
933 file = kzalloc(sizeof(*file), GFP_KERNEL);
936 if (module_dependent)
943 kref_init(&file->ref);
944 mutex_init(&file->ucontext_lock);
946 spin_lock_init(&file->uobjects_lock);
947 INIT_LIST_HEAD(&file->uobjects);
948 init_rwsem(&file->hw_destroy_rwsem);
949 mutex_init(&file->umap_lock);
950 INIT_LIST_HEAD(&file->umaps);
952 filp->private_data = file;
953 list_add_tail(&file->list, &dev->uverbs_file_list);
954 mutex_unlock(&dev->lists_mutex);
955 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
957 setup_ufile_idr_uobject(file);
959 return stream_open(inode, filp);
962 module_put(ib_dev->ops.owner);
965 mutex_unlock(&dev->lists_mutex);
966 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
967 if (refcount_dec_and_test(&dev->refcount))
968 ib_uverbs_comp_dev(dev);
970 put_device(&dev->dev);
974 static int ib_uverbs_close(struct inode *inode, struct file *filp)
976 struct ib_uverbs_file *file = filp->private_data;
978 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
980 mutex_lock(&file->device->lists_mutex);
981 list_del_init(&file->list);
982 mutex_unlock(&file->device->lists_mutex);
984 kref_put(&file->ref, ib_uverbs_release_file);
989 static const struct file_operations uverbs_fops = {
990 .owner = THIS_MODULE,
991 .write = ib_uverbs_write,
992 .open = ib_uverbs_open,
993 .release = ib_uverbs_close,
995 .unlocked_ioctl = ib_uverbs_ioctl,
996 .compat_ioctl = compat_ptr_ioctl,
999 static const struct file_operations uverbs_mmap_fops = {
1000 .owner = THIS_MODULE,
1001 .write = ib_uverbs_write,
1002 .mmap = ib_uverbs_mmap,
1003 .open = ib_uverbs_open,
1004 .release = ib_uverbs_close,
1005 .llseek = no_llseek,
1006 .unlocked_ioctl = ib_uverbs_ioctl,
1007 .compat_ioctl = compat_ptr_ioctl,
1010 static int ib_uverbs_get_nl_info(struct ib_device *ibdev, void *client_data,
1011 struct ib_client_nl_info *res)
1013 struct ib_uverbs_device *uverbs_dev = client_data;
1016 if (res->port != -1)
1019 res->abi = ibdev->ops.uverbs_abi_ver;
1020 res->cdev = &uverbs_dev->dev;
1023 * To support DRIVER_ID binding in userspace some of the driver need
1024 * upgrading to expose their PCI dependent revision information
1025 * through get_context instead of relying on modalias matching. When
1026 * the drivers are fixed they can drop this flag.
1028 if (!ibdev->ops.uverbs_no_driver_id_binding) {
1029 ret = nla_put_u32(res->nl_msg, RDMA_NLDEV_ATTR_UVERBS_DRIVER_ID,
1030 ibdev->ops.driver_id);
1037 static struct ib_client uverbs_client = {
1039 .no_kverbs_req = true,
1040 .add = ib_uverbs_add_one,
1041 .remove = ib_uverbs_remove_one,
1042 .get_nl_info = ib_uverbs_get_nl_info,
1044 MODULE_ALIAS_RDMA_CLIENT("uverbs");
1046 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1049 struct ib_uverbs_device *dev =
1050 container_of(device, struct ib_uverbs_device, dev);
1053 struct ib_device *ib_dev;
1055 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1056 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1058 ret = sysfs_emit(buf, "%s\n", dev_name(&ib_dev->dev));
1059 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1063 static DEVICE_ATTR_RO(ibdev);
1065 static ssize_t abi_version_show(struct device *device,
1066 struct device_attribute *attr, char *buf)
1068 struct ib_uverbs_device *dev =
1069 container_of(device, struct ib_uverbs_device, dev);
1072 struct ib_device *ib_dev;
1074 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1075 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1077 ret = sysfs_emit(buf, "%u\n", ib_dev->ops.uverbs_abi_ver);
1078 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1082 static DEVICE_ATTR_RO(abi_version);
1084 static struct attribute *ib_dev_attrs[] = {
1085 &dev_attr_abi_version.attr,
1086 &dev_attr_ibdev.attr,
1090 static const struct attribute_group dev_attr_group = {
1091 .attrs = ib_dev_attrs,
1094 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1095 __stringify(IB_USER_VERBS_ABI_VERSION));
1097 static int ib_uverbs_create_uapi(struct ib_device *device,
1098 struct ib_uverbs_device *uverbs_dev)
1100 struct uverbs_api *uapi;
1102 uapi = uverbs_alloc_api(device);
1104 return PTR_ERR(uapi);
1106 uverbs_dev->uapi = uapi;
1110 static int ib_uverbs_add_one(struct ib_device *device)
1114 struct ib_uverbs_device *uverbs_dev;
1117 if (!device->ops.alloc_ucontext)
1120 uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1124 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1130 device_initialize(&uverbs_dev->dev);
1131 uverbs_dev->dev.class = &uverbs_class;
1132 uverbs_dev->dev.parent = device->dev.parent;
1133 uverbs_dev->dev.release = ib_uverbs_release_dev;
1134 uverbs_dev->groups[0] = &dev_attr_group;
1135 uverbs_dev->dev.groups = uverbs_dev->groups;
1136 refcount_set(&uverbs_dev->refcount, 1);
1137 init_completion(&uverbs_dev->comp);
1138 uverbs_dev->xrcd_tree = RB_ROOT;
1139 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1140 mutex_init(&uverbs_dev->lists_mutex);
1141 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1142 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1143 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1145 devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1151 uverbs_dev->devnum = devnum;
1152 if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1153 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1155 base = IB_UVERBS_BASE_DEV + devnum;
1157 ret = ib_uverbs_create_uapi(device, uverbs_dev);
1161 uverbs_dev->dev.devt = base;
1162 dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1164 cdev_init(&uverbs_dev->cdev,
1165 device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1166 uverbs_dev->cdev.owner = THIS_MODULE;
1168 ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1172 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1176 ida_free(&uverbs_ida, devnum);
1178 if (refcount_dec_and_test(&uverbs_dev->refcount))
1179 ib_uverbs_comp_dev(uverbs_dev);
1180 wait_for_completion(&uverbs_dev->comp);
1181 put_device(&uverbs_dev->dev);
1185 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1186 struct ib_device *ib_dev)
1188 struct ib_uverbs_file *file;
1190 /* Pending running commands to terminate */
1191 uverbs_disassociate_api_pre(uverbs_dev);
1193 mutex_lock(&uverbs_dev->lists_mutex);
1194 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1195 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1196 struct ib_uverbs_file, list);
1197 list_del_init(&file->list);
1198 kref_get(&file->ref);
1200 /* We must release the mutex before going ahead and calling
1201 * uverbs_cleanup_ufile, as it might end up indirectly calling
1202 * uverbs_close, for example due to freeing the resources (e.g
1205 mutex_unlock(&uverbs_dev->lists_mutex);
1207 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1208 kref_put(&file->ref, ib_uverbs_release_file);
1210 mutex_lock(&uverbs_dev->lists_mutex);
1212 mutex_unlock(&uverbs_dev->lists_mutex);
1214 uverbs_disassociate_api(uverbs_dev->uapi);
1217 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1219 struct ib_uverbs_device *uverbs_dev = client_data;
1220 int wait_clients = 1;
1222 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1223 ida_free(&uverbs_ida, uverbs_dev->devnum);
1225 if (device->ops.disassociate_ucontext) {
1226 /* We disassociate HW resources and immediately return.
1227 * Userspace will see a EIO errno for all future access.
1228 * Upon returning, ib_device may be freed internally and is not
1230 * uverbs_device is still available until all clients close
1231 * their files, then the uverbs device ref count will be zero
1232 * and its resources will be freed.
1233 * Note: At this point no more files can be opened since the
1234 * cdev was deleted, however active clients can still issue
1235 * commands and close their open files.
1237 ib_uverbs_free_hw_resources(uverbs_dev, device);
1241 if (refcount_dec_and_test(&uverbs_dev->refcount))
1242 ib_uverbs_comp_dev(uverbs_dev);
1244 wait_for_completion(&uverbs_dev->comp);
1246 put_device(&uverbs_dev->dev);
1249 static int __init ib_uverbs_init(void)
1253 ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1254 IB_UVERBS_NUM_FIXED_MINOR,
1255 "infiniband_verbs");
1257 pr_err("user_verbs: couldn't register device number\n");
1261 ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1262 IB_UVERBS_NUM_DYNAMIC_MINOR,
1263 "infiniband_verbs");
1265 pr_err("couldn't register dynamic device number\n");
1269 ret = class_register(&uverbs_class);
1271 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1275 ret = class_create_file(&uverbs_class, &class_attr_abi_version.attr);
1277 pr_err("user_verbs: couldn't create abi_version attribute\n");
1281 ret = ib_register_client(&uverbs_client);
1283 pr_err("user_verbs: couldn't register client\n");
1290 class_unregister(&uverbs_class);
1293 unregister_chrdev_region(dynamic_uverbs_dev,
1294 IB_UVERBS_NUM_DYNAMIC_MINOR);
1297 unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1298 IB_UVERBS_NUM_FIXED_MINOR);
1304 static void __exit ib_uverbs_cleanup(void)
1306 ib_unregister_client(&uverbs_client);
1307 class_unregister(&uverbs_class);
1308 unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1309 IB_UVERBS_NUM_FIXED_MINOR);
1310 unregister_chrdev_region(dynamic_uverbs_dev,
1311 IB_UVERBS_NUM_DYNAMIC_MINOR);
1312 mmu_notifier_synchronize();
1315 module_init(ib_uverbs_init);
1316 module_exit(ib_uverbs_cleanup);