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 int 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);
114 atomic_dec(&pd->usecnt);
119 static void ib_uverbs_release_dev(struct device *device)
121 struct ib_uverbs_device *dev =
122 container_of(device, struct ib_uverbs_device, dev);
124 uverbs_destroy_api(dev->uapi);
125 cleanup_srcu_struct(&dev->disassociate_srcu);
126 mutex_destroy(&dev->lists_mutex);
127 mutex_destroy(&dev->xrcd_tree_mutex);
131 void ib_uverbs_release_ucq(struct ib_uverbs_completion_event_file *ev_file,
132 struct ib_ucq_object *uobj)
134 struct ib_uverbs_event *evt, *tmp;
137 spin_lock_irq(&ev_file->ev_queue.lock);
138 list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
139 list_del(&evt->list);
142 spin_unlock_irq(&ev_file->ev_queue.lock);
144 uverbs_uobject_put(&ev_file->uobj);
147 ib_uverbs_release_uevent(&uobj->uevent);
150 void ib_uverbs_release_uevent(struct ib_uevent_object *uobj)
152 struct ib_uverbs_async_event_file *async_file = uobj->event_file;
153 struct ib_uverbs_event *evt, *tmp;
158 spin_lock_irq(&async_file->ev_queue.lock);
159 list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
160 list_del(&evt->list);
163 spin_unlock_irq(&async_file->ev_queue.lock);
164 uverbs_uobject_put(&async_file->uobj);
167 void ib_uverbs_detach_umcast(struct ib_qp *qp,
168 struct ib_uqp_object *uobj)
170 struct ib_uverbs_mcast_entry *mcast, *tmp;
172 list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
173 ib_detach_mcast(qp, &mcast->gid, mcast->lid);
174 list_del(&mcast->list);
179 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
181 complete(&dev->comp);
184 void ib_uverbs_release_file(struct kref *ref)
186 struct ib_uverbs_file *file =
187 container_of(ref, struct ib_uverbs_file, ref);
188 struct ib_device *ib_dev;
191 release_ufile_idr_uobject(file);
193 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
194 ib_dev = srcu_dereference(file->device->ib_dev,
195 &file->device->disassociate_srcu);
196 if (ib_dev && !ib_dev->ops.disassociate_ucontext)
197 module_put(ib_dev->ops.owner);
198 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
200 if (atomic_dec_and_test(&file->device->refcount))
201 ib_uverbs_comp_dev(file->device);
203 if (file->default_async_file)
204 uverbs_uobject_put(&file->default_async_file->uobj);
205 put_device(&file->device->dev);
207 if (file->disassociate_page)
208 __free_pages(file->disassociate_page, 0);
209 mutex_destroy(&file->umap_lock);
210 mutex_destroy(&file->ucontext_lock);
214 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
215 struct file *filp, char __user *buf,
216 size_t count, loff_t *pos,
219 struct ib_uverbs_event *event;
222 spin_lock_irq(&ev_queue->lock);
224 while (list_empty(&ev_queue->event_list)) {
225 if (ev_queue->is_closed) {
226 spin_unlock_irq(&ev_queue->lock);
230 spin_unlock_irq(&ev_queue->lock);
231 if (filp->f_flags & O_NONBLOCK)
234 if (wait_event_interruptible(ev_queue->poll_wait,
235 (!list_empty(&ev_queue->event_list) ||
236 ev_queue->is_closed)))
239 spin_lock_irq(&ev_queue->lock);
242 event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
244 if (eventsz > count) {
248 list_del(ev_queue->event_list.next);
249 if (event->counter) {
251 list_del(&event->obj_list);
255 spin_unlock_irq(&ev_queue->lock);
258 if (copy_to_user(buf, event, eventsz))
269 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
270 size_t count, loff_t *pos)
272 struct ib_uverbs_async_event_file *file = filp->private_data;
274 return ib_uverbs_event_read(&file->ev_queue, filp, buf, count, pos,
275 sizeof(struct ib_uverbs_async_event_desc));
278 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
279 size_t count, loff_t *pos)
281 struct ib_uverbs_completion_event_file *comp_ev_file =
284 return ib_uverbs_event_read(&comp_ev_file->ev_queue, filp, buf, count,
286 sizeof(struct ib_uverbs_comp_event_desc));
289 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
291 struct poll_table_struct *wait)
293 __poll_t pollflags = 0;
295 poll_wait(filp, &ev_queue->poll_wait, wait);
297 spin_lock_irq(&ev_queue->lock);
298 if (!list_empty(&ev_queue->event_list))
299 pollflags = EPOLLIN | EPOLLRDNORM;
300 else if (ev_queue->is_closed)
301 pollflags = EPOLLERR;
302 spin_unlock_irq(&ev_queue->lock);
307 static __poll_t ib_uverbs_async_event_poll(struct file *filp,
308 struct poll_table_struct *wait)
310 struct ib_uverbs_async_event_file *file = filp->private_data;
312 return ib_uverbs_event_poll(&file->ev_queue, filp, wait);
315 static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
316 struct poll_table_struct *wait)
318 struct ib_uverbs_completion_event_file *comp_ev_file =
321 return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
324 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
326 struct ib_uverbs_async_event_file *file = filp->private_data;
328 return fasync_helper(fd, filp, on, &file->ev_queue.async_queue);
331 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
333 struct ib_uverbs_completion_event_file *comp_ev_file =
336 return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
339 const struct file_operations uverbs_event_fops = {
340 .owner = THIS_MODULE,
341 .read = ib_uverbs_comp_event_read,
342 .poll = ib_uverbs_comp_event_poll,
343 .release = uverbs_uobject_fd_release,
344 .fasync = ib_uverbs_comp_event_fasync,
348 const struct file_operations uverbs_async_event_fops = {
349 .owner = THIS_MODULE,
350 .read = ib_uverbs_async_event_read,
351 .poll = ib_uverbs_async_event_poll,
352 .release = uverbs_async_event_release,
353 .fasync = ib_uverbs_async_event_fasync,
357 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
359 struct ib_uverbs_event_queue *ev_queue = cq_context;
360 struct ib_ucq_object *uobj;
361 struct ib_uverbs_event *entry;
367 spin_lock_irqsave(&ev_queue->lock, flags);
368 if (ev_queue->is_closed) {
369 spin_unlock_irqrestore(&ev_queue->lock, flags);
373 entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
375 spin_unlock_irqrestore(&ev_queue->lock, flags);
381 entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle;
382 entry->counter = &uobj->comp_events_reported;
384 list_add_tail(&entry->list, &ev_queue->event_list);
385 list_add_tail(&entry->obj_list, &uobj->comp_list);
386 spin_unlock_irqrestore(&ev_queue->lock, flags);
388 wake_up_interruptible(&ev_queue->poll_wait);
389 kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
392 void ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file,
393 __u64 element, __u64 event,
394 struct list_head *obj_list, u32 *counter)
396 struct ib_uverbs_event *entry;
402 spin_lock_irqsave(&async_file->ev_queue.lock, flags);
403 if (async_file->ev_queue.is_closed) {
404 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
408 entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
410 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
414 entry->desc.async.element = element;
415 entry->desc.async.event_type = event;
416 entry->desc.async.reserved = 0;
417 entry->counter = counter;
419 list_add_tail(&entry->list, &async_file->ev_queue.event_list);
421 list_add_tail(&entry->obj_list, obj_list);
422 spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
424 wake_up_interruptible(&async_file->ev_queue.poll_wait);
425 kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN);
428 static void uverbs_uobj_event(struct ib_uevent_object *eobj,
429 struct ib_event *event)
431 ib_uverbs_async_handler(eobj->event_file,
432 eobj->uobject.user_handle, event->event,
433 &eobj->event_list, &eobj->events_reported);
436 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
438 uverbs_uobj_event(&event->element.cq->uobject->uevent, event);
441 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
443 /* for XRC target qp's, check that qp is live */
444 if (!event->element.qp->uobject)
447 uverbs_uobj_event(&event->element.qp->uobject->uevent, event);
450 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
452 uverbs_uobj_event(&event->element.wq->uobject->uevent, event);
455 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
457 uverbs_uobj_event(&event->element.srq->uobject->uevent, event);
460 static void ib_uverbs_event_handler(struct ib_event_handler *handler,
461 struct ib_event *event)
463 ib_uverbs_async_handler(
464 container_of(handler, struct ib_uverbs_async_event_file,
466 event->element.port_num, event->event, NULL, NULL);
469 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
471 spin_lock_init(&ev_queue->lock);
472 INIT_LIST_HEAD(&ev_queue->event_list);
473 init_waitqueue_head(&ev_queue->poll_wait);
474 ev_queue->is_closed = 0;
475 ev_queue->async_queue = NULL;
478 void ib_uverbs_init_async_event_file(
479 struct ib_uverbs_async_event_file *async_file)
481 struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile;
482 struct ib_device *ib_dev = async_file->uobj.context->device;
484 ib_uverbs_init_event_queue(&async_file->ev_queue);
486 /* The first async_event_file becomes the default one for the file. */
487 mutex_lock(&uverbs_file->ucontext_lock);
488 if (!uverbs_file->default_async_file) {
489 /* Pairs with the put in ib_uverbs_release_file */
490 uverbs_uobject_get(&async_file->uobj);
491 smp_store_release(&uverbs_file->default_async_file, async_file);
493 mutex_unlock(&uverbs_file->ucontext_lock);
495 INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev,
496 ib_uverbs_event_handler);
497 ib_register_event_handler(&async_file->event_handler);
500 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
501 struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count,
502 const struct uverbs_api_write_method *method_elm)
504 if (method_elm->is_ex) {
505 count -= sizeof(*hdr) + sizeof(*ex_hdr);
507 if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
510 if (hdr->in_words * 8 < method_elm->req_size)
513 if (ex_hdr->cmd_hdr_reserved)
516 if (ex_hdr->response) {
517 if (!hdr->out_words && !ex_hdr->provider_out_words)
520 if (hdr->out_words * 8 < method_elm->resp_size)
523 if (!access_ok(u64_to_user_ptr(ex_hdr->response),
524 (hdr->out_words + ex_hdr->provider_out_words) * 8))
527 if (hdr->out_words || ex_hdr->provider_out_words)
534 /* not extended command */
535 if (hdr->in_words * 4 != count)
538 if (count < method_elm->req_size + sizeof(*hdr)) {
540 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
541 * with a 16 byte write instead of 24. Old kernels didn't
542 * check the size so they allowed this. Now that the size is
543 * checked provide a compatibility work around to not break
546 if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
553 if (hdr->out_words * 4 < method_elm->resp_size)
559 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
560 size_t count, loff_t *pos)
562 struct ib_uverbs_file *file = filp->private_data;
563 const struct uverbs_api_write_method *method_elm;
564 struct uverbs_api *uapi = file->device->uapi;
565 struct ib_uverbs_ex_cmd_hdr ex_hdr;
566 struct ib_uverbs_cmd_hdr hdr;
567 struct uverbs_attr_bundle bundle;
571 if (!ib_safe_file_access(filp)) {
572 pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
573 task_tgid_vnr(current), current->comm);
577 if (count < sizeof(hdr))
580 if (copy_from_user(&hdr, buf, sizeof(hdr)))
583 method_elm = uapi_get_method(uapi, hdr.command);
584 if (IS_ERR(method_elm))
585 return PTR_ERR(method_elm);
587 if (method_elm->is_ex) {
588 if (count < (sizeof(hdr) + sizeof(ex_hdr)))
590 if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
594 ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
598 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
602 memset(bundle.attr_present, 0, sizeof(bundle.attr_present));
604 bundle.context = NULL; /* only valid if bundle has uobject */
605 bundle.uobject = NULL;
606 if (!method_elm->is_ex) {
607 size_t in_len = hdr.in_words * 4 - sizeof(hdr);
608 size_t out_len = hdr.out_words * 4;
611 if (method_elm->has_udata) {
612 bundle.driver_udata.inlen =
613 in_len - method_elm->req_size;
614 in_len = method_elm->req_size;
615 if (bundle.driver_udata.inlen)
616 bundle.driver_udata.inbuf = buf + in_len;
618 bundle.driver_udata.inbuf = NULL;
620 memset(&bundle.driver_udata, 0,
621 sizeof(bundle.driver_udata));
624 if (method_elm->has_resp) {
626 * The macros check that if has_resp is set
627 * then the command request structure starts
628 * with a '__aligned u64 response' member.
630 ret = get_user(response, (const u64 __user *)buf);
634 if (method_elm->has_udata) {
635 bundle.driver_udata.outlen =
636 out_len - method_elm->resp_size;
637 out_len = method_elm->resp_size;
638 if (bundle.driver_udata.outlen)
639 bundle.driver_udata.outbuf =
640 u64_to_user_ptr(response +
643 bundle.driver_udata.outbuf = NULL;
646 bundle.driver_udata.outlen = 0;
647 bundle.driver_udata.outbuf = NULL;
650 ib_uverbs_init_udata_buf_or_null(
651 &bundle.ucore, buf, u64_to_user_ptr(response),
654 buf += sizeof(ex_hdr);
656 ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
657 u64_to_user_ptr(ex_hdr.response),
658 hdr.in_words * 8, hdr.out_words * 8);
660 ib_uverbs_init_udata_buf_or_null(
661 &bundle.driver_udata, buf + bundle.ucore.inlen,
662 u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
663 ex_hdr.provider_in_words * 8,
664 ex_hdr.provider_out_words * 8);
668 ret = method_elm->handler(&bundle);
670 uverbs_finalize_object(bundle.uobject, UVERBS_ACCESS_NEW, true,
673 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
674 return (ret) ? : count;
677 static const struct vm_operations_struct rdma_umap_ops;
679 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
681 struct ib_uverbs_file *file = filp->private_data;
682 struct ib_ucontext *ucontext;
686 srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
687 ucontext = ib_uverbs_get_ucontext_file(file);
688 if (IS_ERR(ucontext)) {
689 ret = PTR_ERR(ucontext);
692 vma->vm_ops = &rdma_umap_ops;
693 ret = ucontext->device->ops.mmap(ucontext, vma);
695 srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
700 * The VMA has been dup'd, initialize the vm_private_data with a new tracking
703 static void rdma_umap_open(struct vm_area_struct *vma)
705 struct ib_uverbs_file *ufile = vma->vm_file->private_data;
706 struct rdma_umap_priv *opriv = vma->vm_private_data;
707 struct rdma_umap_priv *priv;
712 /* We are racing with disassociation */
713 if (!down_read_trylock(&ufile->hw_destroy_rwsem))
716 * Disassociation already completed, the VMA should already be zapped.
718 if (!ufile->ucontext)
721 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
724 rdma_umap_priv_init(priv, vma, opriv->entry);
726 up_read(&ufile->hw_destroy_rwsem);
730 up_read(&ufile->hw_destroy_rwsem);
733 * We can't allow the VMA to be created with the actual IO pages, that
734 * would break our API contract, and it can't be stopped at this
737 vma->vm_private_data = NULL;
738 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
741 static void rdma_umap_close(struct vm_area_struct *vma)
743 struct ib_uverbs_file *ufile = vma->vm_file->private_data;
744 struct rdma_umap_priv *priv = vma->vm_private_data;
750 * The vma holds a reference on the struct file that created it, which
751 * in turn means that the ib_uverbs_file is guaranteed to exist at
754 mutex_lock(&ufile->umap_lock);
756 rdma_user_mmap_entry_put(priv->entry);
758 list_del(&priv->list);
759 mutex_unlock(&ufile->umap_lock);
764 * Once the zap_vma_ptes has been called touches to the VMA will come here and
765 * we return a dummy writable zero page for all the pfns.
767 static vm_fault_t rdma_umap_fault(struct vm_fault *vmf)
769 struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data;
770 struct rdma_umap_priv *priv = vmf->vma->vm_private_data;
774 return VM_FAULT_SIGBUS;
776 /* Read only pages can just use the system zero page. */
777 if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) {
778 vmf->page = ZERO_PAGE(vmf->address);
783 mutex_lock(&ufile->umap_lock);
784 if (!ufile->disassociate_page)
785 ufile->disassociate_page =
786 alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0);
788 if (ufile->disassociate_page) {
790 * This VMA is forced to always be shared so this doesn't have
791 * to worry about COW.
793 vmf->page = ufile->disassociate_page;
796 ret = VM_FAULT_SIGBUS;
798 mutex_unlock(&ufile->umap_lock);
803 static const struct vm_operations_struct rdma_umap_ops = {
804 .open = rdma_umap_open,
805 .close = rdma_umap_close,
806 .fault = rdma_umap_fault,
809 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
811 struct rdma_umap_priv *priv, *next_priv;
813 lockdep_assert_held(&ufile->hw_destroy_rwsem);
816 struct mm_struct *mm = NULL;
818 /* Get an arbitrary mm pointer that hasn't been cleaned yet */
819 mutex_lock(&ufile->umap_lock);
820 while (!list_empty(&ufile->umaps)) {
823 priv = list_first_entry(&ufile->umaps,
824 struct rdma_umap_priv, list);
825 mm = priv->vma->vm_mm;
826 ret = mmget_not_zero(mm);
828 list_del_init(&priv->list);
830 rdma_user_mmap_entry_put(priv->entry);
838 mutex_unlock(&ufile->umap_lock);
843 * The umap_lock is nested under mmap_lock since it used within
844 * the vma_ops callbacks, so we have to clean the list one mm
845 * at a time to get the lock ordering right. Typically there
846 * will only be one mm, so no big deal.
849 mutex_lock(&ufile->umap_lock);
850 list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
852 struct vm_area_struct *vma = priv->vma;
854 if (vma->vm_mm != mm)
856 list_del_init(&priv->list);
858 zap_vma_ptes(vma, vma->vm_start,
859 vma->vm_end - vma->vm_start);
862 rdma_user_mmap_entry_put(priv->entry);
866 mutex_unlock(&ufile->umap_lock);
867 mmap_read_unlock(mm);
873 * ib_uverbs_open() does not need the BKL:
875 * - the ib_uverbs_device structures are properly reference counted and
876 * everything else is purely local to the file being created, so
877 * races against other open calls are not a problem;
878 * - there is no ioctl method to race against;
879 * - the open method will either immediately run -ENXIO, or all
880 * required initialization will be done.
882 static int ib_uverbs_open(struct inode *inode, struct file *filp)
884 struct ib_uverbs_device *dev;
885 struct ib_uverbs_file *file;
886 struct ib_device *ib_dev;
888 int module_dependent;
891 dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
892 if (!atomic_inc_not_zero(&dev->refcount))
895 get_device(&dev->dev);
896 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
897 mutex_lock(&dev->lists_mutex);
898 ib_dev = srcu_dereference(dev->ib_dev,
899 &dev->disassociate_srcu);
905 if (!rdma_dev_access_netns(ib_dev, current->nsproxy->net_ns)) {
910 /* In case IB device supports disassociate ucontext, there is no hard
911 * dependency between uverbs device and its low level device.
913 module_dependent = !(ib_dev->ops.disassociate_ucontext);
915 if (module_dependent) {
916 if (!try_module_get(ib_dev->ops.owner)) {
922 file = kzalloc(sizeof(*file), GFP_KERNEL);
925 if (module_dependent)
932 kref_init(&file->ref);
933 mutex_init(&file->ucontext_lock);
935 spin_lock_init(&file->uobjects_lock);
936 INIT_LIST_HEAD(&file->uobjects);
937 init_rwsem(&file->hw_destroy_rwsem);
938 mutex_init(&file->umap_lock);
939 INIT_LIST_HEAD(&file->umaps);
941 filp->private_data = file;
942 list_add_tail(&file->list, &dev->uverbs_file_list);
943 mutex_unlock(&dev->lists_mutex);
944 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
946 setup_ufile_idr_uobject(file);
948 return stream_open(inode, filp);
951 module_put(ib_dev->ops.owner);
954 mutex_unlock(&dev->lists_mutex);
955 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
956 if (atomic_dec_and_test(&dev->refcount))
957 ib_uverbs_comp_dev(dev);
959 put_device(&dev->dev);
963 static int ib_uverbs_close(struct inode *inode, struct file *filp)
965 struct ib_uverbs_file *file = filp->private_data;
967 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
969 mutex_lock(&file->device->lists_mutex);
970 list_del_init(&file->list);
971 mutex_unlock(&file->device->lists_mutex);
973 kref_put(&file->ref, ib_uverbs_release_file);
978 static const struct file_operations uverbs_fops = {
979 .owner = THIS_MODULE,
980 .write = ib_uverbs_write,
981 .open = ib_uverbs_open,
982 .release = ib_uverbs_close,
984 .unlocked_ioctl = ib_uverbs_ioctl,
985 .compat_ioctl = compat_ptr_ioctl,
988 static const struct file_operations uverbs_mmap_fops = {
989 .owner = THIS_MODULE,
990 .write = ib_uverbs_write,
991 .mmap = ib_uverbs_mmap,
992 .open = ib_uverbs_open,
993 .release = ib_uverbs_close,
995 .unlocked_ioctl = ib_uverbs_ioctl,
996 .compat_ioctl = compat_ptr_ioctl,
999 static int ib_uverbs_get_nl_info(struct ib_device *ibdev, void *client_data,
1000 struct ib_client_nl_info *res)
1002 struct ib_uverbs_device *uverbs_dev = client_data;
1005 if (res->port != -1)
1008 res->abi = ibdev->ops.uverbs_abi_ver;
1009 res->cdev = &uverbs_dev->dev;
1012 * To support DRIVER_ID binding in userspace some of the driver need
1013 * upgrading to expose their PCI dependent revision information
1014 * through get_context instead of relying on modalias matching. When
1015 * the drivers are fixed they can drop this flag.
1017 if (!ibdev->ops.uverbs_no_driver_id_binding) {
1018 ret = nla_put_u32(res->nl_msg, RDMA_NLDEV_ATTR_UVERBS_DRIVER_ID,
1019 ibdev->ops.driver_id);
1026 static struct ib_client uverbs_client = {
1028 .no_kverbs_req = true,
1029 .add = ib_uverbs_add_one,
1030 .remove = ib_uverbs_remove_one,
1031 .get_nl_info = ib_uverbs_get_nl_info,
1033 MODULE_ALIAS_RDMA_CLIENT("uverbs");
1035 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1038 struct ib_uverbs_device *dev =
1039 container_of(device, struct ib_uverbs_device, dev);
1042 struct ib_device *ib_dev;
1044 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1045 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1047 ret = sprintf(buf, "%s\n", dev_name(&ib_dev->dev));
1048 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1052 static DEVICE_ATTR_RO(ibdev);
1054 static ssize_t abi_version_show(struct device *device,
1055 struct device_attribute *attr, char *buf)
1057 struct ib_uverbs_device *dev =
1058 container_of(device, struct ib_uverbs_device, dev);
1061 struct ib_device *ib_dev;
1063 srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1064 ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1066 ret = sprintf(buf, "%u\n", ib_dev->ops.uverbs_abi_ver);
1067 srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1071 static DEVICE_ATTR_RO(abi_version);
1073 static struct attribute *ib_dev_attrs[] = {
1074 &dev_attr_abi_version.attr,
1075 &dev_attr_ibdev.attr,
1079 static const struct attribute_group dev_attr_group = {
1080 .attrs = ib_dev_attrs,
1083 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1084 __stringify(IB_USER_VERBS_ABI_VERSION));
1086 static int ib_uverbs_create_uapi(struct ib_device *device,
1087 struct ib_uverbs_device *uverbs_dev)
1089 struct uverbs_api *uapi;
1091 uapi = uverbs_alloc_api(device);
1093 return PTR_ERR(uapi);
1095 uverbs_dev->uapi = uapi;
1099 static int ib_uverbs_add_one(struct ib_device *device)
1103 struct ib_uverbs_device *uverbs_dev;
1106 if (!device->ops.alloc_ucontext)
1109 uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1113 ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1119 device_initialize(&uverbs_dev->dev);
1120 uverbs_dev->dev.class = uverbs_class;
1121 uverbs_dev->dev.parent = device->dev.parent;
1122 uverbs_dev->dev.release = ib_uverbs_release_dev;
1123 uverbs_dev->groups[0] = &dev_attr_group;
1124 uverbs_dev->dev.groups = uverbs_dev->groups;
1125 atomic_set(&uverbs_dev->refcount, 1);
1126 init_completion(&uverbs_dev->comp);
1127 uverbs_dev->xrcd_tree = RB_ROOT;
1128 mutex_init(&uverbs_dev->xrcd_tree_mutex);
1129 mutex_init(&uverbs_dev->lists_mutex);
1130 INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1131 rcu_assign_pointer(uverbs_dev->ib_dev, device);
1132 uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1134 devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1140 uverbs_dev->devnum = devnum;
1141 if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1142 base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1144 base = IB_UVERBS_BASE_DEV + devnum;
1146 ret = ib_uverbs_create_uapi(device, uverbs_dev);
1150 uverbs_dev->dev.devt = base;
1151 dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1153 cdev_init(&uverbs_dev->cdev,
1154 device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1155 uverbs_dev->cdev.owner = THIS_MODULE;
1157 ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1161 ib_set_client_data(device, &uverbs_client, uverbs_dev);
1165 ida_free(&uverbs_ida, devnum);
1167 if (atomic_dec_and_test(&uverbs_dev->refcount))
1168 ib_uverbs_comp_dev(uverbs_dev);
1169 wait_for_completion(&uverbs_dev->comp);
1170 put_device(&uverbs_dev->dev);
1174 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1175 struct ib_device *ib_dev)
1177 struct ib_uverbs_file *file;
1179 /* Pending running commands to terminate */
1180 uverbs_disassociate_api_pre(uverbs_dev);
1182 mutex_lock(&uverbs_dev->lists_mutex);
1183 while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1184 file = list_first_entry(&uverbs_dev->uverbs_file_list,
1185 struct ib_uverbs_file, list);
1186 list_del_init(&file->list);
1187 kref_get(&file->ref);
1189 /* We must release the mutex before going ahead and calling
1190 * uverbs_cleanup_ufile, as it might end up indirectly calling
1191 * uverbs_close, for example due to freeing the resources (e.g
1194 mutex_unlock(&uverbs_dev->lists_mutex);
1196 uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1197 kref_put(&file->ref, ib_uverbs_release_file);
1199 mutex_lock(&uverbs_dev->lists_mutex);
1201 mutex_unlock(&uverbs_dev->lists_mutex);
1203 uverbs_disassociate_api(uverbs_dev->uapi);
1206 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1208 struct ib_uverbs_device *uverbs_dev = client_data;
1209 int wait_clients = 1;
1211 cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1212 ida_free(&uverbs_ida, uverbs_dev->devnum);
1214 if (device->ops.disassociate_ucontext) {
1215 /* We disassociate HW resources and immediately return.
1216 * Userspace will see a EIO errno for all future access.
1217 * Upon returning, ib_device may be freed internally and is not
1219 * uverbs_device is still available until all clients close
1220 * their files, then the uverbs device ref count will be zero
1221 * and its resources will be freed.
1222 * Note: At this point no more files can be opened since the
1223 * cdev was deleted, however active clients can still issue
1224 * commands and close their open files.
1226 ib_uverbs_free_hw_resources(uverbs_dev, device);
1230 if (atomic_dec_and_test(&uverbs_dev->refcount))
1231 ib_uverbs_comp_dev(uverbs_dev);
1233 wait_for_completion(&uverbs_dev->comp);
1235 put_device(&uverbs_dev->dev);
1238 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1242 return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1245 static int __init ib_uverbs_init(void)
1249 ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1250 IB_UVERBS_NUM_FIXED_MINOR,
1251 "infiniband_verbs");
1253 pr_err("user_verbs: couldn't register device number\n");
1257 ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1258 IB_UVERBS_NUM_DYNAMIC_MINOR,
1259 "infiniband_verbs");
1261 pr_err("couldn't register dynamic device number\n");
1265 uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1266 if (IS_ERR(uverbs_class)) {
1267 ret = PTR_ERR(uverbs_class);
1268 pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1272 uverbs_class->devnode = uverbs_devnode;
1274 ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1276 pr_err("user_verbs: couldn't create abi_version attribute\n");
1280 ret = ib_register_client(&uverbs_client);
1282 pr_err("user_verbs: couldn't register client\n");
1289 class_destroy(uverbs_class);
1292 unregister_chrdev_region(dynamic_uverbs_dev,
1293 IB_UVERBS_NUM_DYNAMIC_MINOR);
1296 unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1297 IB_UVERBS_NUM_FIXED_MINOR);
1303 static void __exit ib_uverbs_cleanup(void)
1305 ib_unregister_client(&uverbs_client);
1306 class_destroy(uverbs_class);
1307 unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1308 IB_UVERBS_NUM_FIXED_MINOR);
1309 unregister_chrdev_region(dynamic_uverbs_dev,
1310 IB_UVERBS_NUM_DYNAMIC_MINOR);
1311 mmu_notifier_synchronize();
1314 module_init(ib_uverbs_init);
1315 module_exit(ib_uverbs_cleanup);