Loading...
1/*
2 * Copyright (c) 2016, Mellanox Technologies inc. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33#include <linux/file.h>
34#include <linux/anon_inodes.h>
35#include <linux/sched/mm.h>
36#include <rdma/ib_verbs.h>
37#include <rdma/uverbs_types.h>
38#include <linux/rcupdate.h>
39#include <rdma/uverbs_ioctl.h>
40#include <rdma/rdma_user_ioctl.h>
41#include "uverbs.h"
42#include "core_priv.h"
43#include "rdma_core.h"
44
45static void uverbs_uobject_free(struct kref *ref)
46{
47 kfree_rcu(container_of(ref, struct ib_uobject, ref), rcu);
48}
49
50/*
51 * In order to indicate we no longer needs this uobject, uverbs_uobject_put
52 * is called. When the reference count is decreased, the uobject is freed.
53 * For example, this is used when attaching a completion channel to a CQ.
54 */
55void uverbs_uobject_put(struct ib_uobject *uobject)
56{
57 kref_put(&uobject->ref, uverbs_uobject_free);
58}
59EXPORT_SYMBOL(uverbs_uobject_put);
60
61int uverbs_try_lock_object(struct ib_uobject *uobj,
62 enum rdma_lookup_mode mode)
63{
64 /*
65 * When a shared access is required, we use a positive counter. Each
66 * shared access request checks that the value != -1 and increment it.
67 * Exclusive access is required for operations like write or destroy.
68 * In exclusive access mode, we check that the counter is zero (nobody
69 * claimed this object) and we set it to -1. Releasing a shared access
70 * lock is done simply by decreasing the counter. As for exclusive
71 * access locks, since only a single one of them is allowed
72 * concurrently, setting the counter to zero is enough for releasing
73 * this lock.
74 */
75 switch (mode) {
76 case UVERBS_LOOKUP_READ:
77 return atomic_fetch_add_unless(&uobj->usecnt, 1, -1) == -1 ?
78 -EBUSY : 0;
79 case UVERBS_LOOKUP_WRITE:
80 /* lock is exclusive */
81 return atomic_cmpxchg(&uobj->usecnt, 0, -1) == 0 ? 0 : -EBUSY;
82 case UVERBS_LOOKUP_DESTROY:
83 return 0;
84 }
85 return 0;
86}
87EXPORT_SYMBOL(uverbs_try_lock_object);
88
89static void assert_uverbs_usecnt(struct ib_uobject *uobj,
90 enum rdma_lookup_mode mode)
91{
92#ifdef CONFIG_LOCKDEP
93 switch (mode) {
94 case UVERBS_LOOKUP_READ:
95 WARN_ON(atomic_read(&uobj->usecnt) <= 0);
96 break;
97 case UVERBS_LOOKUP_WRITE:
98 WARN_ON(atomic_read(&uobj->usecnt) != -1);
99 break;
100 case UVERBS_LOOKUP_DESTROY:
101 break;
102 }
103#endif
104}
105
106/*
107 * This must be called with the hw_destroy_rwsem locked for read or write,
108 * also the uobject itself must be locked for write.
109 *
110 * Upon return the HW object is guaranteed to be destroyed.
111 *
112 * For RDMA_REMOVE_ABORT, the hw_destroy_rwsem is not required to be held,
113 * however the type's allocat_commit function cannot have been called and the
114 * uobject cannot be on the uobjects_lists
115 *
116 * For RDMA_REMOVE_DESTROY the caller should be holding a kref (eg via
117 * rdma_lookup_get_uobject) and the object is left in a state where the caller
118 * needs to call rdma_lookup_put_uobject.
119 *
120 * For all other destroy modes this function internally unlocks the uobject
121 * and consumes the kref on the uobj.
122 */
123static int uverbs_destroy_uobject(struct ib_uobject *uobj,
124 enum rdma_remove_reason reason,
125 struct uverbs_attr_bundle *attrs)
126{
127 struct ib_uverbs_file *ufile = attrs->ufile;
128 unsigned long flags;
129 int ret;
130
131 lockdep_assert_held(&ufile->hw_destroy_rwsem);
132 assert_uverbs_usecnt(uobj, UVERBS_LOOKUP_WRITE);
133
134 if (reason == RDMA_REMOVE_ABORT) {
135 WARN_ON(!list_empty(&uobj->list));
136 WARN_ON(!uobj->context);
137 uobj->uapi_object->type_class->alloc_abort(uobj);
138 } else if (uobj->object) {
139 ret = uobj->uapi_object->type_class->destroy_hw(uobj, reason,
140 attrs);
141 if (ret)
142 /* Nothing to be done, wait till ucontext will clean it */
143 return ret;
144
145 uobj->object = NULL;
146 }
147
148 uobj->context = NULL;
149
150 /*
151 * For DESTROY the usecnt is not changed, the caller is expected to
152 * manage it via uobj_put_destroy(). Only DESTROY can remove the IDR
153 * handle.
154 */
155 if (reason != RDMA_REMOVE_DESTROY)
156 atomic_set(&uobj->usecnt, 0);
157 else
158 uobj->uapi_object->type_class->remove_handle(uobj);
159
160 if (!list_empty(&uobj->list)) {
161 spin_lock_irqsave(&ufile->uobjects_lock, flags);
162 list_del_init(&uobj->list);
163 spin_unlock_irqrestore(&ufile->uobjects_lock, flags);
164
165 /*
166 * Pairs with the get in rdma_alloc_commit_uobject(), could
167 * destroy uobj.
168 */
169 uverbs_uobject_put(uobj);
170 }
171
172 /*
173 * When aborting the stack kref remains owned by the core code, and is
174 * not transferred into the type. Pairs with the get in alloc_uobj
175 */
176 if (reason == RDMA_REMOVE_ABORT)
177 uverbs_uobject_put(uobj);
178
179 return 0;
180}
181
182/*
183 * This calls uverbs_destroy_uobject() using the RDMA_REMOVE_DESTROY
184 * sequence. It should only be used from command callbacks. On success the
185 * caller must pair this with uobj_put_destroy(). This
186 * version requires the caller to have already obtained an
187 * LOOKUP_DESTROY uobject kref.
188 */
189int uobj_destroy(struct ib_uobject *uobj, struct uverbs_attr_bundle *attrs)
190{
191 struct ib_uverbs_file *ufile = attrs->ufile;
192 int ret;
193
194 down_read(&ufile->hw_destroy_rwsem);
195
196 /*
197 * Once the uobject is destroyed by RDMA_REMOVE_DESTROY then it is left
198 * write locked as the callers put it back with UVERBS_LOOKUP_DESTROY.
199 * This is because any other concurrent thread can still see the object
200 * in the xarray due to RCU. Leaving it locked ensures nothing else will
201 * touch it.
202 */
203 ret = uverbs_try_lock_object(uobj, UVERBS_LOOKUP_WRITE);
204 if (ret)
205 goto out_unlock;
206
207 ret = uverbs_destroy_uobject(uobj, RDMA_REMOVE_DESTROY, attrs);
208 if (ret) {
209 atomic_set(&uobj->usecnt, 0);
210 goto out_unlock;
211 }
212
213out_unlock:
214 up_read(&ufile->hw_destroy_rwsem);
215 return ret;
216}
217
218/*
219 * uobj_get_destroy destroys the HW object and returns a handle to the uobj
220 * with a NULL object pointer. The caller must pair this with
221 * uobj_put_destroy().
222 */
223struct ib_uobject *__uobj_get_destroy(const struct uverbs_api_object *obj,
224 u32 id, struct uverbs_attr_bundle *attrs)
225{
226 struct ib_uobject *uobj;
227 int ret;
228
229 uobj = rdma_lookup_get_uobject(obj, attrs->ufile, id,
230 UVERBS_LOOKUP_DESTROY, attrs);
231 if (IS_ERR(uobj))
232 return uobj;
233
234 ret = uobj_destroy(uobj, attrs);
235 if (ret) {
236 rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_DESTROY);
237 return ERR_PTR(ret);
238 }
239
240 return uobj;
241}
242
243/*
244 * Does both uobj_get_destroy() and uobj_put_destroy(). Returns 0 on success
245 * (negative errno on failure). For use by callers that do not need the uobj.
246 */
247int __uobj_perform_destroy(const struct uverbs_api_object *obj, u32 id,
248 struct uverbs_attr_bundle *attrs)
249{
250 struct ib_uobject *uobj;
251
252 uobj = __uobj_get_destroy(obj, id, attrs);
253 if (IS_ERR(uobj))
254 return PTR_ERR(uobj);
255 uobj_put_destroy(uobj);
256 return 0;
257}
258
259/* alloc_uobj must be undone by uverbs_destroy_uobject() */
260static struct ib_uobject *alloc_uobj(struct uverbs_attr_bundle *attrs,
261 const struct uverbs_api_object *obj)
262{
263 struct ib_uverbs_file *ufile = attrs->ufile;
264 struct ib_uobject *uobj;
265
266 if (!attrs->context) {
267 struct ib_ucontext *ucontext =
268 ib_uverbs_get_ucontext_file(ufile);
269
270 if (IS_ERR(ucontext))
271 return ERR_CAST(ucontext);
272 attrs->context = ucontext;
273 }
274
275 uobj = kzalloc(obj->type_attrs->obj_size, GFP_KERNEL);
276 if (!uobj)
277 return ERR_PTR(-ENOMEM);
278 /*
279 * user_handle should be filled by the handler,
280 * The object is added to the list in the commit stage.
281 */
282 uobj->ufile = ufile;
283 uobj->context = attrs->context;
284 INIT_LIST_HEAD(&uobj->list);
285 uobj->uapi_object = obj;
286 /*
287 * Allocated objects start out as write locked to deny any other
288 * syscalls from accessing them until they are committed. See
289 * rdma_alloc_commit_uobject
290 */
291 atomic_set(&uobj->usecnt, -1);
292 kref_init(&uobj->ref);
293
294 return uobj;
295}
296
297static int idr_add_uobj(struct ib_uobject *uobj)
298{
299 /*
300 * We start with allocating an idr pointing to NULL. This represents an
301 * object which isn't initialized yet. We'll replace it later on with
302 * the real object once we commit.
303 */
304 return xa_alloc(&uobj->ufile->idr, &uobj->id, NULL, xa_limit_32b,
305 GFP_KERNEL);
306}
307
308/* Returns the ib_uobject or an error. The caller should check for IS_ERR. */
309static struct ib_uobject *
310lookup_get_idr_uobject(const struct uverbs_api_object *obj,
311 struct ib_uverbs_file *ufile, s64 id,
312 enum rdma_lookup_mode mode)
313{
314 struct ib_uobject *uobj;
315
316 if (id < 0 || id > ULONG_MAX)
317 return ERR_PTR(-EINVAL);
318
319 rcu_read_lock();
320 /*
321 * The idr_find is guaranteed to return a pointer to something that
322 * isn't freed yet, or NULL, as the free after idr_remove goes through
323 * kfree_rcu(). However the object may still have been released and
324 * kfree() could be called at any time.
325 */
326 uobj = xa_load(&ufile->idr, id);
327 if (!uobj || !kref_get_unless_zero(&uobj->ref))
328 uobj = ERR_PTR(-ENOENT);
329 rcu_read_unlock();
330 return uobj;
331}
332
333static struct ib_uobject *
334lookup_get_fd_uobject(const struct uverbs_api_object *obj,
335 struct ib_uverbs_file *ufile, s64 id,
336 enum rdma_lookup_mode mode)
337{
338 const struct uverbs_obj_fd_type *fd_type;
339 struct file *f;
340 struct ib_uobject *uobject;
341 int fdno = id;
342
343 if (fdno != id)
344 return ERR_PTR(-EINVAL);
345
346 if (mode != UVERBS_LOOKUP_READ)
347 return ERR_PTR(-EOPNOTSUPP);
348
349 if (!obj->type_attrs)
350 return ERR_PTR(-EIO);
351 fd_type =
352 container_of(obj->type_attrs, struct uverbs_obj_fd_type, type);
353
354 f = fget(fdno);
355 if (!f)
356 return ERR_PTR(-EBADF);
357
358 uobject = f->private_data;
359 /*
360 * fget(id) ensures we are not currently running
361 * uverbs_uobject_fd_release(), and the caller is expected to ensure
362 * that release is never done while a call to lookup is possible.
363 */
364 if (f->f_op != fd_type->fops || uobject->ufile != ufile) {
365 fput(f);
366 return ERR_PTR(-EBADF);
367 }
368
369 uverbs_uobject_get(uobject);
370 return uobject;
371}
372
373struct ib_uobject *rdma_lookup_get_uobject(const struct uverbs_api_object *obj,
374 struct ib_uverbs_file *ufile, s64 id,
375 enum rdma_lookup_mode mode,
376 struct uverbs_attr_bundle *attrs)
377{
378 struct ib_uobject *uobj;
379 int ret;
380
381 if (obj == ERR_PTR(-ENOMSG)) {
382 /* must be UVERBS_IDR_ANY_OBJECT, see uapi_get_object() */
383 uobj = lookup_get_idr_uobject(NULL, ufile, id, mode);
384 if (IS_ERR(uobj))
385 return uobj;
386 } else {
387 if (IS_ERR(obj))
388 return ERR_PTR(-EINVAL);
389
390 uobj = obj->type_class->lookup_get(obj, ufile, id, mode);
391 if (IS_ERR(uobj))
392 return uobj;
393
394 if (uobj->uapi_object != obj) {
395 ret = -EINVAL;
396 goto free;
397 }
398 }
399
400 /*
401 * If we have been disassociated block every command except for
402 * DESTROY based commands.
403 */
404 if (mode != UVERBS_LOOKUP_DESTROY &&
405 !srcu_dereference(ufile->device->ib_dev,
406 &ufile->device->disassociate_srcu)) {
407 ret = -EIO;
408 goto free;
409 }
410
411 ret = uverbs_try_lock_object(uobj, mode);
412 if (ret)
413 goto free;
414 if (attrs)
415 attrs->context = uobj->context;
416
417 return uobj;
418free:
419 uobj->uapi_object->type_class->lookup_put(uobj, mode);
420 uverbs_uobject_put(uobj);
421 return ERR_PTR(ret);
422}
423
424static struct ib_uobject *
425alloc_begin_idr_uobject(const struct uverbs_api_object *obj,
426 struct uverbs_attr_bundle *attrs)
427{
428 int ret;
429 struct ib_uobject *uobj;
430
431 uobj = alloc_uobj(attrs, obj);
432 if (IS_ERR(uobj))
433 return uobj;
434
435 ret = idr_add_uobj(uobj);
436 if (ret)
437 goto uobj_put;
438
439 ret = ib_rdmacg_try_charge(&uobj->cg_obj, uobj->context->device,
440 RDMACG_RESOURCE_HCA_OBJECT);
441 if (ret)
442 goto remove;
443
444 return uobj;
445
446remove:
447 xa_erase(&attrs->ufile->idr, uobj->id);
448uobj_put:
449 uverbs_uobject_put(uobj);
450 return ERR_PTR(ret);
451}
452
453static struct ib_uobject *
454alloc_begin_fd_uobject(const struct uverbs_api_object *obj,
455 struct uverbs_attr_bundle *attrs)
456{
457 const struct uverbs_obj_fd_type *fd_type;
458 int new_fd;
459 struct ib_uobject *uobj, *ret;
460 struct file *filp;
461
462 uobj = alloc_uobj(attrs, obj);
463 if (IS_ERR(uobj))
464 return uobj;
465
466 fd_type =
467 container_of(obj->type_attrs, struct uverbs_obj_fd_type, type);
468 if (WARN_ON(fd_type->fops->release != &uverbs_uobject_fd_release &&
469 fd_type->fops->release != &uverbs_async_event_release)) {
470 ret = ERR_PTR(-EINVAL);
471 goto err_fd;
472 }
473
474 new_fd = get_unused_fd_flags(O_CLOEXEC);
475 if (new_fd < 0) {
476 ret = ERR_PTR(new_fd);
477 goto err_fd;
478 }
479
480 /* Note that uverbs_uobject_fd_release() is called during abort */
481 filp = anon_inode_getfile(fd_type->name, fd_type->fops, NULL,
482 fd_type->flags);
483 if (IS_ERR(filp)) {
484 ret = ERR_CAST(filp);
485 goto err_getfile;
486 }
487 uobj->object = filp;
488
489 uobj->id = new_fd;
490 return uobj;
491
492err_getfile:
493 put_unused_fd(new_fd);
494err_fd:
495 uverbs_uobject_put(uobj);
496 return ret;
497}
498
499struct ib_uobject *rdma_alloc_begin_uobject(const struct uverbs_api_object *obj,
500 struct uverbs_attr_bundle *attrs)
501{
502 struct ib_uverbs_file *ufile = attrs->ufile;
503 struct ib_uobject *ret;
504
505 if (IS_ERR(obj))
506 return ERR_PTR(-EINVAL);
507
508 /*
509 * The hw_destroy_rwsem is held across the entire object creation and
510 * released during rdma_alloc_commit_uobject or
511 * rdma_alloc_abort_uobject
512 */
513 if (!down_read_trylock(&ufile->hw_destroy_rwsem))
514 return ERR_PTR(-EIO);
515
516 ret = obj->type_class->alloc_begin(obj, attrs);
517 if (IS_ERR(ret)) {
518 up_read(&ufile->hw_destroy_rwsem);
519 return ret;
520 }
521 return ret;
522}
523
524static void alloc_abort_idr_uobject(struct ib_uobject *uobj)
525{
526 ib_rdmacg_uncharge(&uobj->cg_obj, uobj->context->device,
527 RDMACG_RESOURCE_HCA_OBJECT);
528
529 xa_erase(&uobj->ufile->idr, uobj->id);
530}
531
532static int __must_check destroy_hw_idr_uobject(struct ib_uobject *uobj,
533 enum rdma_remove_reason why,
534 struct uverbs_attr_bundle *attrs)
535{
536 const struct uverbs_obj_idr_type *idr_type =
537 container_of(uobj->uapi_object->type_attrs,
538 struct uverbs_obj_idr_type, type);
539 int ret = idr_type->destroy_object(uobj, why, attrs);
540
541 if (ret)
542 return ret;
543
544 if (why == RDMA_REMOVE_ABORT)
545 return 0;
546
547 ib_rdmacg_uncharge(&uobj->cg_obj, uobj->context->device,
548 RDMACG_RESOURCE_HCA_OBJECT);
549
550 return 0;
551}
552
553static void remove_handle_idr_uobject(struct ib_uobject *uobj)
554{
555 xa_erase(&uobj->ufile->idr, uobj->id);
556 /* Matches the kref in alloc_commit_idr_uobject */
557 uverbs_uobject_put(uobj);
558}
559
560static void alloc_abort_fd_uobject(struct ib_uobject *uobj)
561{
562 struct file *filp = uobj->object;
563
564 fput(filp);
565 put_unused_fd(uobj->id);
566}
567
568static int __must_check destroy_hw_fd_uobject(struct ib_uobject *uobj,
569 enum rdma_remove_reason why,
570 struct uverbs_attr_bundle *attrs)
571{
572 const struct uverbs_obj_fd_type *fd_type = container_of(
573 uobj->uapi_object->type_attrs, struct uverbs_obj_fd_type, type);
574
575 fd_type->destroy_object(uobj, why);
576 return 0;
577}
578
579static void remove_handle_fd_uobject(struct ib_uobject *uobj)
580{
581}
582
583static void alloc_commit_idr_uobject(struct ib_uobject *uobj)
584{
585 struct ib_uverbs_file *ufile = uobj->ufile;
586 void *old;
587
588 /*
589 * We already allocated this IDR with a NULL object, so
590 * this shouldn't fail.
591 *
592 * NOTE: Storing the uobj transfers our kref on uobj to the XArray.
593 * It will be put by remove_commit_idr_uobject()
594 */
595 old = xa_store(&ufile->idr, uobj->id, uobj, GFP_KERNEL);
596 WARN_ON(old != NULL);
597}
598
599static void swap_idr_uobjects(struct ib_uobject *obj_old,
600 struct ib_uobject *obj_new)
601{
602 struct ib_uverbs_file *ufile = obj_old->ufile;
603 void *old;
604
605 /*
606 * New must be an object that been allocated but not yet committed, this
607 * moves the pre-committed state to obj_old, new still must be comitted.
608 */
609 old = xa_cmpxchg(&ufile->idr, obj_old->id, obj_old, XA_ZERO_ENTRY,
610 GFP_KERNEL);
611 if (WARN_ON(old != obj_old))
612 return;
613
614 swap(obj_old->id, obj_new->id);
615
616 old = xa_cmpxchg(&ufile->idr, obj_old->id, NULL, obj_old, GFP_KERNEL);
617 WARN_ON(old != NULL);
618}
619
620static void alloc_commit_fd_uobject(struct ib_uobject *uobj)
621{
622 int fd = uobj->id;
623 struct file *filp = uobj->object;
624
625 /* Matching put will be done in uverbs_uobject_fd_release() */
626 kref_get(&uobj->ufile->ref);
627
628 /* This shouldn't be used anymore. Use the file object instead */
629 uobj->id = 0;
630
631 /*
632 * NOTE: Once we install the file we loose ownership of our kref on
633 * uobj. It will be put by uverbs_uobject_fd_release()
634 */
635 filp->private_data = uobj;
636 fd_install(fd, filp);
637}
638
639/*
640 * In all cases rdma_alloc_commit_uobject() consumes the kref to uobj and the
641 * caller can no longer assume uobj is valid. If this function fails it
642 * destroys the uboject, including the attached HW object.
643 */
644void rdma_alloc_commit_uobject(struct ib_uobject *uobj,
645 struct uverbs_attr_bundle *attrs)
646{
647 struct ib_uverbs_file *ufile = attrs->ufile;
648
649 /* kref is held so long as the uobj is on the uobj list. */
650 uverbs_uobject_get(uobj);
651 spin_lock_irq(&ufile->uobjects_lock);
652 list_add(&uobj->list, &ufile->uobjects);
653 spin_unlock_irq(&ufile->uobjects_lock);
654
655 /* matches atomic_set(-1) in alloc_uobj */
656 atomic_set(&uobj->usecnt, 0);
657
658 /* alloc_commit consumes the uobj kref */
659 uobj->uapi_object->type_class->alloc_commit(uobj);
660
661 /* Matches the down_read in rdma_alloc_begin_uobject */
662 up_read(&ufile->hw_destroy_rwsem);
663}
664
665/*
666 * new_uobj will be assigned to the handle currently used by to_uobj, and
667 * to_uobj will be destroyed.
668 *
669 * Upon return the caller must do:
670 * rdma_alloc_commit_uobject(new_uobj)
671 * uobj_put_destroy(to_uobj)
672 *
673 * to_uobj must have a write get but the put mode switches to destroy once
674 * this is called.
675 */
676void rdma_assign_uobject(struct ib_uobject *to_uobj, struct ib_uobject *new_uobj,
677 struct uverbs_attr_bundle *attrs)
678{
679 assert_uverbs_usecnt(new_uobj, UVERBS_LOOKUP_WRITE);
680
681 if (WARN_ON(to_uobj->uapi_object != new_uobj->uapi_object ||
682 !to_uobj->uapi_object->type_class->swap_uobjects))
683 return;
684
685 to_uobj->uapi_object->type_class->swap_uobjects(to_uobj, new_uobj);
686
687 /*
688 * If this fails then the uobject is still completely valid (though with
689 * a new ID) and we leak it until context close.
690 */
691 uverbs_destroy_uobject(to_uobj, RDMA_REMOVE_DESTROY, attrs);
692}
693
694/*
695 * This consumes the kref for uobj. It is up to the caller to unwind the HW
696 * object and anything else connected to uobj before calling this.
697 */
698void rdma_alloc_abort_uobject(struct ib_uobject *uobj,
699 struct uverbs_attr_bundle *attrs,
700 bool hw_obj_valid)
701{
702 struct ib_uverbs_file *ufile = uobj->ufile;
703 int ret;
704
705 if (hw_obj_valid) {
706 ret = uobj->uapi_object->type_class->destroy_hw(
707 uobj, RDMA_REMOVE_ABORT, attrs);
708 /*
709 * If the driver couldn't destroy the object then go ahead and
710 * commit it. Leaking objects that can't be destroyed is only
711 * done during FD close after the driver has a few more tries to
712 * destroy it.
713 */
714 if (WARN_ON(ret))
715 return rdma_alloc_commit_uobject(uobj, attrs);
716 }
717
718 uverbs_destroy_uobject(uobj, RDMA_REMOVE_ABORT, attrs);
719
720 /* Matches the down_read in rdma_alloc_begin_uobject */
721 up_read(&ufile->hw_destroy_rwsem);
722}
723
724static void lookup_put_idr_uobject(struct ib_uobject *uobj,
725 enum rdma_lookup_mode mode)
726{
727}
728
729static void lookup_put_fd_uobject(struct ib_uobject *uobj,
730 enum rdma_lookup_mode mode)
731{
732 struct file *filp = uobj->object;
733
734 WARN_ON(mode != UVERBS_LOOKUP_READ);
735 /*
736 * This indirectly calls uverbs_uobject_fd_release() and free the
737 * object
738 */
739 fput(filp);
740}
741
742void rdma_lookup_put_uobject(struct ib_uobject *uobj,
743 enum rdma_lookup_mode mode)
744{
745 assert_uverbs_usecnt(uobj, mode);
746 /*
747 * In order to unlock an object, either decrease its usecnt for
748 * read access or zero it in case of exclusive access. See
749 * uverbs_try_lock_object for locking schema information.
750 */
751 switch (mode) {
752 case UVERBS_LOOKUP_READ:
753 atomic_dec(&uobj->usecnt);
754 break;
755 case UVERBS_LOOKUP_WRITE:
756 atomic_set(&uobj->usecnt, 0);
757 break;
758 case UVERBS_LOOKUP_DESTROY:
759 break;
760 }
761
762 uobj->uapi_object->type_class->lookup_put(uobj, mode);
763 /* Pairs with the kref obtained by type->lookup_get */
764 uverbs_uobject_put(uobj);
765}
766
767void setup_ufile_idr_uobject(struct ib_uverbs_file *ufile)
768{
769 xa_init_flags(&ufile->idr, XA_FLAGS_ALLOC);
770}
771
772void release_ufile_idr_uobject(struct ib_uverbs_file *ufile)
773{
774 struct ib_uobject *entry;
775 unsigned long id;
776
777 /*
778 * At this point uverbs_cleanup_ufile() is guaranteed to have run, and
779 * there are no HW objects left, however the xarray is still populated
780 * with anything that has not been cleaned up by userspace. Since the
781 * kref on ufile is 0, nothing is allowed to call lookup_get.
782 *
783 * This is an optimized equivalent to remove_handle_idr_uobject
784 */
785 xa_for_each(&ufile->idr, id, entry) {
786 WARN_ON(entry->object);
787 uverbs_uobject_put(entry);
788 }
789
790 xa_destroy(&ufile->idr);
791}
792
793const struct uverbs_obj_type_class uverbs_idr_class = {
794 .alloc_begin = alloc_begin_idr_uobject,
795 .lookup_get = lookup_get_idr_uobject,
796 .alloc_commit = alloc_commit_idr_uobject,
797 .alloc_abort = alloc_abort_idr_uobject,
798 .lookup_put = lookup_put_idr_uobject,
799 .destroy_hw = destroy_hw_idr_uobject,
800 .remove_handle = remove_handle_idr_uobject,
801 .swap_uobjects = swap_idr_uobjects,
802};
803EXPORT_SYMBOL(uverbs_idr_class);
804
805/*
806 * Users of UVERBS_TYPE_ALLOC_FD should set this function as the struct
807 * file_operations release method.
808 */
809int uverbs_uobject_fd_release(struct inode *inode, struct file *filp)
810{
811 struct ib_uverbs_file *ufile;
812 struct ib_uobject *uobj;
813
814 /*
815 * This can only happen if the fput came from alloc_abort_fd_uobject()
816 */
817 if (!filp->private_data)
818 return 0;
819 uobj = filp->private_data;
820 ufile = uobj->ufile;
821
822 if (down_read_trylock(&ufile->hw_destroy_rwsem)) {
823 struct uverbs_attr_bundle attrs = {
824 .context = uobj->context,
825 .ufile = ufile,
826 };
827
828 /*
829 * lookup_get_fd_uobject holds the kref on the struct file any
830 * time a FD uobj is locked, which prevents this release
831 * method from being invoked. Meaning we can always get the
832 * write lock here, or we have a kernel bug.
833 */
834 WARN_ON(uverbs_try_lock_object(uobj, UVERBS_LOOKUP_WRITE));
835 uverbs_destroy_uobject(uobj, RDMA_REMOVE_CLOSE, &attrs);
836 up_read(&ufile->hw_destroy_rwsem);
837 }
838
839 /* Matches the get in alloc_commit_fd_uobject() */
840 kref_put(&ufile->ref, ib_uverbs_release_file);
841
842 /* Pairs with filp->private_data in alloc_begin_fd_uobject */
843 uverbs_uobject_put(uobj);
844 return 0;
845}
846EXPORT_SYMBOL(uverbs_uobject_fd_release);
847
848/*
849 * Drop the ucontext off the ufile and completely disconnect it from the
850 * ib_device
851 */
852static void ufile_destroy_ucontext(struct ib_uverbs_file *ufile,
853 enum rdma_remove_reason reason)
854{
855 struct ib_ucontext *ucontext = ufile->ucontext;
856 struct ib_device *ib_dev = ucontext->device;
857
858 /*
859 * If we are closing the FD then the user mmap VMAs must have
860 * already been destroyed as they hold on to the filep, otherwise
861 * they need to be zap'd.
862 */
863 if (reason == RDMA_REMOVE_DRIVER_REMOVE) {
864 uverbs_user_mmap_disassociate(ufile);
865 if (ib_dev->ops.disassociate_ucontext)
866 ib_dev->ops.disassociate_ucontext(ucontext);
867 }
868
869 ib_rdmacg_uncharge(&ucontext->cg_obj, ib_dev,
870 RDMACG_RESOURCE_HCA_HANDLE);
871
872 rdma_restrack_del(&ucontext->res);
873
874 ib_dev->ops.dealloc_ucontext(ucontext);
875 WARN_ON(!xa_empty(&ucontext->mmap_xa));
876 kfree(ucontext);
877
878 ufile->ucontext = NULL;
879}
880
881static int __uverbs_cleanup_ufile(struct ib_uverbs_file *ufile,
882 enum rdma_remove_reason reason)
883{
884 struct uverbs_attr_bundle attrs = { .ufile = ufile };
885 struct ib_ucontext *ucontext = ufile->ucontext;
886 struct ib_device *ib_dev = ucontext->device;
887 struct ib_uobject *obj, *next_obj;
888 int ret = -EINVAL;
889
890 if (ib_dev->ops.ufile_hw_cleanup)
891 ib_dev->ops.ufile_hw_cleanup(ufile);
892
893 /*
894 * This shouldn't run while executing other commands on this
895 * context. Thus, the only thing we should take care of is
896 * releasing a FD while traversing this list. The FD could be
897 * closed and released from the _release fop of this FD.
898 * In order to mitigate this, we add a lock.
899 * We take and release the lock per traversal in order to let
900 * other threads (which might still use the FDs) chance to run.
901 */
902 list_for_each_entry_safe(obj, next_obj, &ufile->uobjects, list) {
903 attrs.context = obj->context;
904 /*
905 * if we hit this WARN_ON, that means we are
906 * racing with a lookup_get.
907 */
908 WARN_ON(uverbs_try_lock_object(obj, UVERBS_LOOKUP_WRITE));
909 if (reason == RDMA_REMOVE_DRIVER_FAILURE)
910 obj->object = NULL;
911 if (!uverbs_destroy_uobject(obj, reason, &attrs))
912 ret = 0;
913 else
914 atomic_set(&obj->usecnt, 0);
915 }
916
917 if (reason == RDMA_REMOVE_DRIVER_FAILURE) {
918 WARN_ON(!list_empty(&ufile->uobjects));
919 return 0;
920 }
921 return ret;
922}
923
924/*
925 * Destroy the ucontext and every uobject associated with it.
926 *
927 * This is internally locked and can be called in parallel from multiple
928 * contexts.
929 */
930void uverbs_destroy_ufile_hw(struct ib_uverbs_file *ufile,
931 enum rdma_remove_reason reason)
932{
933 down_write(&ufile->hw_destroy_rwsem);
934
935 /*
936 * If a ucontext was never created then we can't have any uobjects to
937 * cleanup, nothing to do.
938 */
939 if (!ufile->ucontext)
940 goto done;
941
942 while (!list_empty(&ufile->uobjects) &&
943 !__uverbs_cleanup_ufile(ufile, reason)) {
944 }
945
946 if (WARN_ON(!list_empty(&ufile->uobjects)))
947 __uverbs_cleanup_ufile(ufile, RDMA_REMOVE_DRIVER_FAILURE);
948 ufile_destroy_ucontext(ufile, reason);
949
950done:
951 up_write(&ufile->hw_destroy_rwsem);
952}
953
954const struct uverbs_obj_type_class uverbs_fd_class = {
955 .alloc_begin = alloc_begin_fd_uobject,
956 .lookup_get = lookup_get_fd_uobject,
957 .alloc_commit = alloc_commit_fd_uobject,
958 .alloc_abort = alloc_abort_fd_uobject,
959 .lookup_put = lookup_put_fd_uobject,
960 .destroy_hw = destroy_hw_fd_uobject,
961 .remove_handle = remove_handle_fd_uobject,
962};
963EXPORT_SYMBOL(uverbs_fd_class);
964
965struct ib_uobject *
966uverbs_get_uobject_from_file(u16 object_id, enum uverbs_obj_access access,
967 s64 id, struct uverbs_attr_bundle *attrs)
968{
969 const struct uverbs_api_object *obj =
970 uapi_get_object(attrs->ufile->device->uapi, object_id);
971
972 switch (access) {
973 case UVERBS_ACCESS_READ:
974 return rdma_lookup_get_uobject(obj, attrs->ufile, id,
975 UVERBS_LOOKUP_READ, attrs);
976 case UVERBS_ACCESS_DESTROY:
977 /* Actual destruction is done inside uverbs_handle_method */
978 return rdma_lookup_get_uobject(obj, attrs->ufile, id,
979 UVERBS_LOOKUP_DESTROY, attrs);
980 case UVERBS_ACCESS_WRITE:
981 return rdma_lookup_get_uobject(obj, attrs->ufile, id,
982 UVERBS_LOOKUP_WRITE, attrs);
983 case UVERBS_ACCESS_NEW:
984 return rdma_alloc_begin_uobject(obj, attrs);
985 default:
986 WARN_ON(true);
987 return ERR_PTR(-EOPNOTSUPP);
988 }
989}
990
991void uverbs_finalize_object(struct ib_uobject *uobj,
992 enum uverbs_obj_access access, bool hw_obj_valid,
993 bool commit, struct uverbs_attr_bundle *attrs)
994{
995 /*
996 * refcounts should be handled at the object level and not at the
997 * uobject level. Refcounts of the objects themselves are done in
998 * handlers.
999 */
1000
1001 switch (access) {
1002 case UVERBS_ACCESS_READ:
1003 rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_READ);
1004 break;
1005 case UVERBS_ACCESS_WRITE:
1006 rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_WRITE);
1007 break;
1008 case UVERBS_ACCESS_DESTROY:
1009 if (uobj)
1010 rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_DESTROY);
1011 break;
1012 case UVERBS_ACCESS_NEW:
1013 if (commit)
1014 rdma_alloc_commit_uobject(uobj, attrs);
1015 else
1016 rdma_alloc_abort_uobject(uobj, attrs, hw_obj_valid);
1017 break;
1018 default:
1019 WARN_ON(true);
1020 }
1021}
1/*
2 * Copyright (c) 2016, Mellanox Technologies inc. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33#include <linux/file.h>
34#include <linux/anon_inodes.h>
35#include <rdma/ib_verbs.h>
36#include <rdma/uverbs_types.h>
37#include <linux/rcupdate.h>
38#include <rdma/uverbs_ioctl.h>
39#include <rdma/rdma_user_ioctl.h>
40#include "uverbs.h"
41#include "core_priv.h"
42#include "rdma_core.h"
43
44int uverbs_ns_idx(u16 *id, unsigned int ns_count)
45{
46 int ret = (*id & UVERBS_ID_NS_MASK) >> UVERBS_ID_NS_SHIFT;
47
48 if (ret >= ns_count)
49 return -EINVAL;
50
51 *id &= ~UVERBS_ID_NS_MASK;
52 return ret;
53}
54
55const struct uverbs_object_spec *uverbs_get_object(const struct ib_device *ibdev,
56 uint16_t object)
57{
58 const struct uverbs_root_spec *object_hash = ibdev->specs_root;
59 const struct uverbs_object_spec_hash *objects;
60 int ret = uverbs_ns_idx(&object, object_hash->num_buckets);
61
62 if (ret < 0)
63 return NULL;
64
65 objects = object_hash->object_buckets[ret];
66
67 if (object >= objects->num_objects)
68 return NULL;
69
70 return objects->objects[object];
71}
72
73const struct uverbs_method_spec *uverbs_get_method(const struct uverbs_object_spec *object,
74 uint16_t method)
75{
76 const struct uverbs_method_spec_hash *methods;
77 int ret = uverbs_ns_idx(&method, object->num_buckets);
78
79 if (ret < 0)
80 return NULL;
81
82 methods = object->method_buckets[ret];
83 if (method >= methods->num_methods)
84 return NULL;
85
86 return methods->methods[method];
87}
88
89void uverbs_uobject_get(struct ib_uobject *uobject)
90{
91 kref_get(&uobject->ref);
92}
93
94static void uverbs_uobject_free(struct kref *ref)
95{
96 struct ib_uobject *uobj =
97 container_of(ref, struct ib_uobject, ref);
98
99 if (uobj->type->type_class->needs_kfree_rcu)
100 kfree_rcu(uobj, rcu);
101 else
102 kfree(uobj);
103}
104
105void uverbs_uobject_put(struct ib_uobject *uobject)
106{
107 kref_put(&uobject->ref, uverbs_uobject_free);
108}
109
110static int uverbs_try_lock_object(struct ib_uobject *uobj, bool exclusive)
111{
112 /*
113 * When a shared access is required, we use a positive counter. Each
114 * shared access request checks that the value != -1 and increment it.
115 * Exclusive access is required for operations like write or destroy.
116 * In exclusive access mode, we check that the counter is zero (nobody
117 * claimed this object) and we set it to -1. Releasing a shared access
118 * lock is done simply by decreasing the counter. As for exclusive
119 * access locks, since only a single one of them is is allowed
120 * concurrently, setting the counter to zero is enough for releasing
121 * this lock.
122 */
123 if (!exclusive)
124 return __atomic_add_unless(&uobj->usecnt, 1, -1) == -1 ?
125 -EBUSY : 0;
126
127 /* lock is either WRITE or DESTROY - should be exclusive */
128 return atomic_cmpxchg(&uobj->usecnt, 0, -1) == 0 ? 0 : -EBUSY;
129}
130
131static struct ib_uobject *alloc_uobj(struct ib_ucontext *context,
132 const struct uverbs_obj_type *type)
133{
134 struct ib_uobject *uobj = kzalloc(type->obj_size, GFP_KERNEL);
135
136 if (!uobj)
137 return ERR_PTR(-ENOMEM);
138 /*
139 * user_handle should be filled by the handler,
140 * The object is added to the list in the commit stage.
141 */
142 uobj->context = context;
143 uobj->type = type;
144 /*
145 * Allocated objects start out as write locked to deny any other
146 * syscalls from accessing them until they are committed. See
147 * rdma_alloc_commit_uobject
148 */
149 atomic_set(&uobj->usecnt, -1);
150 kref_init(&uobj->ref);
151
152 return uobj;
153}
154
155static int idr_add_uobj(struct ib_uobject *uobj)
156{
157 int ret;
158
159 idr_preload(GFP_KERNEL);
160 spin_lock(&uobj->context->ufile->idr_lock);
161
162 /*
163 * We start with allocating an idr pointing to NULL. This represents an
164 * object which isn't initialized yet. We'll replace it later on with
165 * the real object once we commit.
166 */
167 ret = idr_alloc(&uobj->context->ufile->idr, NULL, 0,
168 min_t(unsigned long, U32_MAX - 1, INT_MAX), GFP_NOWAIT);
169 if (ret >= 0)
170 uobj->id = ret;
171
172 spin_unlock(&uobj->context->ufile->idr_lock);
173 idr_preload_end();
174
175 return ret < 0 ? ret : 0;
176}
177
178/*
179 * It only removes it from the uobjects list, uverbs_uobject_put() is still
180 * required.
181 */
182static void uverbs_idr_remove_uobj(struct ib_uobject *uobj)
183{
184 spin_lock(&uobj->context->ufile->idr_lock);
185 idr_remove(&uobj->context->ufile->idr, uobj->id);
186 spin_unlock(&uobj->context->ufile->idr_lock);
187}
188
189/* Returns the ib_uobject or an error. The caller should check for IS_ERR. */
190static struct ib_uobject *lookup_get_idr_uobject(const struct uverbs_obj_type *type,
191 struct ib_ucontext *ucontext,
192 int id, bool exclusive)
193{
194 struct ib_uobject *uobj;
195
196 rcu_read_lock();
197 /* object won't be released as we're protected in rcu */
198 uobj = idr_find(&ucontext->ufile->idr, id);
199 if (!uobj) {
200 uobj = ERR_PTR(-ENOENT);
201 goto free;
202 }
203
204 /*
205 * The idr_find is guaranteed to return a pointer to something that
206 * isn't freed yet, or NULL, as the free after idr_remove goes through
207 * kfree_rcu(). However the object may still have been released and
208 * kfree() could be called at any time.
209 */
210 if (!kref_get_unless_zero(&uobj->ref))
211 uobj = ERR_PTR(-ENOENT);
212
213free:
214 rcu_read_unlock();
215 return uobj;
216}
217
218static struct ib_uobject *lookup_get_fd_uobject(const struct uverbs_obj_type *type,
219 struct ib_ucontext *ucontext,
220 int id, bool exclusive)
221{
222 struct file *f;
223 struct ib_uobject *uobject;
224 const struct uverbs_obj_fd_type *fd_type =
225 container_of(type, struct uverbs_obj_fd_type, type);
226
227 if (exclusive)
228 return ERR_PTR(-EOPNOTSUPP);
229
230 f = fget(id);
231 if (!f)
232 return ERR_PTR(-EBADF);
233
234 uobject = f->private_data;
235 /*
236 * fget(id) ensures we are not currently running uverbs_close_fd,
237 * and the caller is expected to ensure that uverbs_close_fd is never
238 * done while a call top lookup is possible.
239 */
240 if (f->f_op != fd_type->fops) {
241 fput(f);
242 return ERR_PTR(-EBADF);
243 }
244
245 uverbs_uobject_get(uobject);
246 return uobject;
247}
248
249struct ib_uobject *rdma_lookup_get_uobject(const struct uverbs_obj_type *type,
250 struct ib_ucontext *ucontext,
251 int id, bool exclusive)
252{
253 struct ib_uobject *uobj;
254 int ret;
255
256 uobj = type->type_class->lookup_get(type, ucontext, id, exclusive);
257 if (IS_ERR(uobj))
258 return uobj;
259
260 if (uobj->type != type) {
261 ret = -EINVAL;
262 goto free;
263 }
264
265 ret = uverbs_try_lock_object(uobj, exclusive);
266 if (ret) {
267 WARN(ucontext->cleanup_reason,
268 "ib_uverbs: Trying to lookup_get while cleanup context\n");
269 goto free;
270 }
271
272 return uobj;
273free:
274 uobj->type->type_class->lookup_put(uobj, exclusive);
275 uverbs_uobject_put(uobj);
276 return ERR_PTR(ret);
277}
278
279static struct ib_uobject *alloc_begin_idr_uobject(const struct uverbs_obj_type *type,
280 struct ib_ucontext *ucontext)
281{
282 int ret;
283 struct ib_uobject *uobj;
284
285 uobj = alloc_uobj(ucontext, type);
286 if (IS_ERR(uobj))
287 return uobj;
288
289 ret = idr_add_uobj(uobj);
290 if (ret)
291 goto uobj_put;
292
293 ret = ib_rdmacg_try_charge(&uobj->cg_obj, ucontext->device,
294 RDMACG_RESOURCE_HCA_OBJECT);
295 if (ret)
296 goto idr_remove;
297
298 return uobj;
299
300idr_remove:
301 uverbs_idr_remove_uobj(uobj);
302uobj_put:
303 uverbs_uobject_put(uobj);
304 return ERR_PTR(ret);
305}
306
307static struct ib_uobject *alloc_begin_fd_uobject(const struct uverbs_obj_type *type,
308 struct ib_ucontext *ucontext)
309{
310 const struct uverbs_obj_fd_type *fd_type =
311 container_of(type, struct uverbs_obj_fd_type, type);
312 int new_fd;
313 struct ib_uobject *uobj;
314 struct ib_uobject_file *uobj_file;
315 struct file *filp;
316
317 new_fd = get_unused_fd_flags(O_CLOEXEC);
318 if (new_fd < 0)
319 return ERR_PTR(new_fd);
320
321 uobj = alloc_uobj(ucontext, type);
322 if (IS_ERR(uobj)) {
323 put_unused_fd(new_fd);
324 return uobj;
325 }
326
327 uobj_file = container_of(uobj, struct ib_uobject_file, uobj);
328 filp = anon_inode_getfile(fd_type->name,
329 fd_type->fops,
330 uobj_file,
331 fd_type->flags);
332 if (IS_ERR(filp)) {
333 put_unused_fd(new_fd);
334 uverbs_uobject_put(uobj);
335 return (void *)filp;
336 }
337
338 uobj_file->uobj.id = new_fd;
339 uobj_file->uobj.object = filp;
340 uobj_file->ufile = ucontext->ufile;
341 INIT_LIST_HEAD(&uobj->list);
342 kref_get(&uobj_file->ufile->ref);
343
344 return uobj;
345}
346
347struct ib_uobject *rdma_alloc_begin_uobject(const struct uverbs_obj_type *type,
348 struct ib_ucontext *ucontext)
349{
350 return type->type_class->alloc_begin(type, ucontext);
351}
352
353static int __must_check remove_commit_idr_uobject(struct ib_uobject *uobj,
354 enum rdma_remove_reason why)
355{
356 const struct uverbs_obj_idr_type *idr_type =
357 container_of(uobj->type, struct uverbs_obj_idr_type,
358 type);
359 int ret = idr_type->destroy_object(uobj, why);
360
361 /*
362 * We can only fail gracefully if the user requested to destroy the
363 * object. In the rest of the cases, just remove whatever you can.
364 */
365 if (why == RDMA_REMOVE_DESTROY && ret)
366 return ret;
367
368 ib_rdmacg_uncharge(&uobj->cg_obj, uobj->context->device,
369 RDMACG_RESOURCE_HCA_OBJECT);
370 uverbs_idr_remove_uobj(uobj);
371
372 return ret;
373}
374
375static void alloc_abort_fd_uobject(struct ib_uobject *uobj)
376{
377 struct ib_uobject_file *uobj_file =
378 container_of(uobj, struct ib_uobject_file, uobj);
379 struct file *filp = uobj->object;
380 int id = uobj_file->uobj.id;
381
382 /* Unsuccessful NEW */
383 fput(filp);
384 put_unused_fd(id);
385}
386
387static int __must_check remove_commit_fd_uobject(struct ib_uobject *uobj,
388 enum rdma_remove_reason why)
389{
390 const struct uverbs_obj_fd_type *fd_type =
391 container_of(uobj->type, struct uverbs_obj_fd_type, type);
392 struct ib_uobject_file *uobj_file =
393 container_of(uobj, struct ib_uobject_file, uobj);
394 int ret = fd_type->context_closed(uobj_file, why);
395
396 if (why == RDMA_REMOVE_DESTROY && ret)
397 return ret;
398
399 if (why == RDMA_REMOVE_DURING_CLEANUP) {
400 alloc_abort_fd_uobject(uobj);
401 return ret;
402 }
403
404 uobj_file->uobj.context = NULL;
405 return ret;
406}
407
408static void assert_uverbs_usecnt(struct ib_uobject *uobj, bool exclusive)
409{
410#ifdef CONFIG_LOCKDEP
411 if (exclusive)
412 WARN_ON(atomic_read(&uobj->usecnt) != -1);
413 else
414 WARN_ON(atomic_read(&uobj->usecnt) <= 0);
415#endif
416}
417
418static int __must_check _rdma_remove_commit_uobject(struct ib_uobject *uobj,
419 enum rdma_remove_reason why)
420{
421 int ret;
422 struct ib_ucontext *ucontext = uobj->context;
423
424 ret = uobj->type->type_class->remove_commit(uobj, why);
425 if (ret && why == RDMA_REMOVE_DESTROY) {
426 /* We couldn't remove the object, so just unlock the uobject */
427 atomic_set(&uobj->usecnt, 0);
428 uobj->type->type_class->lookup_put(uobj, true);
429 } else {
430 mutex_lock(&ucontext->uobjects_lock);
431 list_del(&uobj->list);
432 mutex_unlock(&ucontext->uobjects_lock);
433 /* put the ref we took when we created the object */
434 uverbs_uobject_put(uobj);
435 }
436
437 return ret;
438}
439
440/* This is called only for user requested DESTROY reasons */
441int __must_check rdma_remove_commit_uobject(struct ib_uobject *uobj)
442{
443 int ret;
444 struct ib_ucontext *ucontext = uobj->context;
445
446 /* put the ref count we took at lookup_get */
447 uverbs_uobject_put(uobj);
448 /* Cleanup is running. Calling this should have been impossible */
449 if (!down_read_trylock(&ucontext->cleanup_rwsem)) {
450 WARN(true, "ib_uverbs: Cleanup is running while removing an uobject\n");
451 return 0;
452 }
453 assert_uverbs_usecnt(uobj, true);
454 ret = _rdma_remove_commit_uobject(uobj, RDMA_REMOVE_DESTROY);
455
456 up_read(&ucontext->cleanup_rwsem);
457 return ret;
458}
459
460static int null_obj_type_class_remove_commit(struct ib_uobject *uobj,
461 enum rdma_remove_reason why)
462{
463 return 0;
464}
465
466static const struct uverbs_obj_type null_obj_type = {
467 .type_class = &((const struct uverbs_obj_type_class){
468 .remove_commit = null_obj_type_class_remove_commit,
469 /* be cautious */
470 .needs_kfree_rcu = true}),
471};
472
473int rdma_explicit_destroy(struct ib_uobject *uobject)
474{
475 int ret;
476 struct ib_ucontext *ucontext = uobject->context;
477
478 /* Cleanup is running. Calling this should have been impossible */
479 if (!down_read_trylock(&ucontext->cleanup_rwsem)) {
480 WARN(true, "ib_uverbs: Cleanup is running while removing an uobject\n");
481 return 0;
482 }
483 assert_uverbs_usecnt(uobject, true);
484 ret = uobject->type->type_class->remove_commit(uobject,
485 RDMA_REMOVE_DESTROY);
486 if (ret)
487 goto out;
488
489 uobject->type = &null_obj_type;
490
491out:
492 up_read(&ucontext->cleanup_rwsem);
493 return ret;
494}
495
496static void alloc_commit_idr_uobject(struct ib_uobject *uobj)
497{
498 spin_lock(&uobj->context->ufile->idr_lock);
499 /*
500 * We already allocated this IDR with a NULL object, so
501 * this shouldn't fail.
502 */
503 WARN_ON(idr_replace(&uobj->context->ufile->idr,
504 uobj, uobj->id));
505 spin_unlock(&uobj->context->ufile->idr_lock);
506}
507
508static void alloc_commit_fd_uobject(struct ib_uobject *uobj)
509{
510 struct ib_uobject_file *uobj_file =
511 container_of(uobj, struct ib_uobject_file, uobj);
512
513 fd_install(uobj_file->uobj.id, uobj->object);
514 /* This shouldn't be used anymore. Use the file object instead */
515 uobj_file->uobj.id = 0;
516 /* Get another reference as we export this to the fops */
517 uverbs_uobject_get(&uobj_file->uobj);
518}
519
520int rdma_alloc_commit_uobject(struct ib_uobject *uobj)
521{
522 /* Cleanup is running. Calling this should have been impossible */
523 if (!down_read_trylock(&uobj->context->cleanup_rwsem)) {
524 int ret;
525
526 WARN(true, "ib_uverbs: Cleanup is running while allocating an uobject\n");
527 ret = uobj->type->type_class->remove_commit(uobj,
528 RDMA_REMOVE_DURING_CLEANUP);
529 if (ret)
530 pr_warn("ib_uverbs: cleanup of idr object %d failed\n",
531 uobj->id);
532 return ret;
533 }
534
535 /* matches atomic_set(-1) in alloc_uobj */
536 assert_uverbs_usecnt(uobj, true);
537 atomic_set(&uobj->usecnt, 0);
538
539 mutex_lock(&uobj->context->uobjects_lock);
540 list_add(&uobj->list, &uobj->context->uobjects);
541 mutex_unlock(&uobj->context->uobjects_lock);
542
543 uobj->type->type_class->alloc_commit(uobj);
544 up_read(&uobj->context->cleanup_rwsem);
545
546 return 0;
547}
548
549static void alloc_abort_idr_uobject(struct ib_uobject *uobj)
550{
551 uverbs_idr_remove_uobj(uobj);
552 ib_rdmacg_uncharge(&uobj->cg_obj, uobj->context->device,
553 RDMACG_RESOURCE_HCA_OBJECT);
554 uverbs_uobject_put(uobj);
555}
556
557void rdma_alloc_abort_uobject(struct ib_uobject *uobj)
558{
559 uobj->type->type_class->alloc_abort(uobj);
560}
561
562static void lookup_put_idr_uobject(struct ib_uobject *uobj, bool exclusive)
563{
564}
565
566static void lookup_put_fd_uobject(struct ib_uobject *uobj, bool exclusive)
567{
568 struct file *filp = uobj->object;
569
570 WARN_ON(exclusive);
571 /* This indirectly calls uverbs_close_fd and free the object */
572 fput(filp);
573}
574
575void rdma_lookup_put_uobject(struct ib_uobject *uobj, bool exclusive)
576{
577 assert_uverbs_usecnt(uobj, exclusive);
578 uobj->type->type_class->lookup_put(uobj, exclusive);
579 /*
580 * In order to unlock an object, either decrease its usecnt for
581 * read access or zero it in case of exclusive access. See
582 * uverbs_try_lock_object for locking schema information.
583 */
584 if (!exclusive)
585 atomic_dec(&uobj->usecnt);
586 else
587 atomic_set(&uobj->usecnt, 0);
588
589 uverbs_uobject_put(uobj);
590}
591
592const struct uverbs_obj_type_class uverbs_idr_class = {
593 .alloc_begin = alloc_begin_idr_uobject,
594 .lookup_get = lookup_get_idr_uobject,
595 .alloc_commit = alloc_commit_idr_uobject,
596 .alloc_abort = alloc_abort_idr_uobject,
597 .lookup_put = lookup_put_idr_uobject,
598 .remove_commit = remove_commit_idr_uobject,
599 /*
600 * When we destroy an object, we first just lock it for WRITE and
601 * actually DESTROY it in the finalize stage. So, the problematic
602 * scenario is when we just started the finalize stage of the
603 * destruction (nothing was executed yet). Now, the other thread
604 * fetched the object for READ access, but it didn't lock it yet.
605 * The DESTROY thread continues and starts destroying the object.
606 * When the other thread continue - without the RCU, it would
607 * access freed memory. However, the rcu_read_lock delays the free
608 * until the rcu_read_lock of the READ operation quits. Since the
609 * exclusive lock of the object is still taken by the DESTROY flow, the
610 * READ operation will get -EBUSY and it'll just bail out.
611 */
612 .needs_kfree_rcu = true,
613};
614
615static void _uverbs_close_fd(struct ib_uobject_file *uobj_file)
616{
617 struct ib_ucontext *ucontext;
618 struct ib_uverbs_file *ufile = uobj_file->ufile;
619 int ret;
620
621 mutex_lock(&uobj_file->ufile->cleanup_mutex);
622
623 /* uobject was either already cleaned up or is cleaned up right now anyway */
624 if (!uobj_file->uobj.context ||
625 !down_read_trylock(&uobj_file->uobj.context->cleanup_rwsem))
626 goto unlock;
627
628 ucontext = uobj_file->uobj.context;
629 ret = _rdma_remove_commit_uobject(&uobj_file->uobj, RDMA_REMOVE_CLOSE);
630 up_read(&ucontext->cleanup_rwsem);
631 if (ret)
632 pr_warn("uverbs: unable to clean up uobject file in uverbs_close_fd.\n");
633unlock:
634 mutex_unlock(&ufile->cleanup_mutex);
635}
636
637void uverbs_close_fd(struct file *f)
638{
639 struct ib_uobject_file *uobj_file = f->private_data;
640 struct kref *uverbs_file_ref = &uobj_file->ufile->ref;
641
642 _uverbs_close_fd(uobj_file);
643 uverbs_uobject_put(&uobj_file->uobj);
644 kref_put(uverbs_file_ref, ib_uverbs_release_file);
645}
646
647void uverbs_cleanup_ucontext(struct ib_ucontext *ucontext, bool device_removed)
648{
649 enum rdma_remove_reason reason = device_removed ?
650 RDMA_REMOVE_DRIVER_REMOVE : RDMA_REMOVE_CLOSE;
651 unsigned int cur_order = 0;
652
653 ucontext->cleanup_reason = reason;
654 /*
655 * Waits for all remove_commit and alloc_commit to finish. Logically, We
656 * want to hold this forever as the context is going to be destroyed,
657 * but we'll release it since it causes a "held lock freed" BUG message.
658 */
659 down_write(&ucontext->cleanup_rwsem);
660
661 while (!list_empty(&ucontext->uobjects)) {
662 struct ib_uobject *obj, *next_obj;
663 unsigned int next_order = UINT_MAX;
664
665 /*
666 * This shouldn't run while executing other commands on this
667 * context. Thus, the only thing we should take care of is
668 * releasing a FD while traversing this list. The FD could be
669 * closed and released from the _release fop of this FD.
670 * In order to mitigate this, we add a lock.
671 * We take and release the lock per order traversal in order
672 * to let other threads (which might still use the FDs) chance
673 * to run.
674 */
675 mutex_lock(&ucontext->uobjects_lock);
676 list_for_each_entry_safe(obj, next_obj, &ucontext->uobjects,
677 list) {
678 if (obj->type->destroy_order == cur_order) {
679 int ret;
680
681 /*
682 * if we hit this WARN_ON, that means we are
683 * racing with a lookup_get.
684 */
685 WARN_ON(uverbs_try_lock_object(obj, true));
686 ret = obj->type->type_class->remove_commit(obj,
687 reason);
688 list_del(&obj->list);
689 if (ret)
690 pr_warn("ib_uverbs: failed to remove uobject id %d order %u\n",
691 obj->id, cur_order);
692 /* put the ref we took when we created the object */
693 uverbs_uobject_put(obj);
694 } else {
695 next_order = min(next_order,
696 obj->type->destroy_order);
697 }
698 }
699 mutex_unlock(&ucontext->uobjects_lock);
700 cur_order = next_order;
701 }
702 up_write(&ucontext->cleanup_rwsem);
703}
704
705void uverbs_initialize_ucontext(struct ib_ucontext *ucontext)
706{
707 ucontext->cleanup_reason = 0;
708 mutex_init(&ucontext->uobjects_lock);
709 INIT_LIST_HEAD(&ucontext->uobjects);
710 init_rwsem(&ucontext->cleanup_rwsem);
711}
712
713const struct uverbs_obj_type_class uverbs_fd_class = {
714 .alloc_begin = alloc_begin_fd_uobject,
715 .lookup_get = lookup_get_fd_uobject,
716 .alloc_commit = alloc_commit_fd_uobject,
717 .alloc_abort = alloc_abort_fd_uobject,
718 .lookup_put = lookup_put_fd_uobject,
719 .remove_commit = remove_commit_fd_uobject,
720 .needs_kfree_rcu = false,
721};
722
723struct ib_uobject *uverbs_get_uobject_from_context(const struct uverbs_obj_type *type_attrs,
724 struct ib_ucontext *ucontext,
725 enum uverbs_obj_access access,
726 int id)
727{
728 switch (access) {
729 case UVERBS_ACCESS_READ:
730 return rdma_lookup_get_uobject(type_attrs, ucontext, id, false);
731 case UVERBS_ACCESS_DESTROY:
732 case UVERBS_ACCESS_WRITE:
733 return rdma_lookup_get_uobject(type_attrs, ucontext, id, true);
734 case UVERBS_ACCESS_NEW:
735 return rdma_alloc_begin_uobject(type_attrs, ucontext);
736 default:
737 WARN_ON(true);
738 return ERR_PTR(-EOPNOTSUPP);
739 }
740}
741
742int uverbs_finalize_object(struct ib_uobject *uobj,
743 enum uverbs_obj_access access,
744 bool commit)
745{
746 int ret = 0;
747
748 /*
749 * refcounts should be handled at the object level and not at the
750 * uobject level. Refcounts of the objects themselves are done in
751 * handlers.
752 */
753
754 switch (access) {
755 case UVERBS_ACCESS_READ:
756 rdma_lookup_put_uobject(uobj, false);
757 break;
758 case UVERBS_ACCESS_WRITE:
759 rdma_lookup_put_uobject(uobj, true);
760 break;
761 case UVERBS_ACCESS_DESTROY:
762 if (commit)
763 ret = rdma_remove_commit_uobject(uobj);
764 else
765 rdma_lookup_put_uobject(uobj, true);
766 break;
767 case UVERBS_ACCESS_NEW:
768 if (commit)
769 ret = rdma_alloc_commit_uobject(uobj);
770 else
771 rdma_alloc_abort_uobject(uobj);
772 break;
773 default:
774 WARN_ON(true);
775 ret = -EOPNOTSUPP;
776 }
777
778 return ret;
779}
780
781int uverbs_finalize_objects(struct uverbs_attr_bundle *attrs_bundle,
782 struct uverbs_attr_spec_hash * const *spec_hash,
783 size_t num,
784 bool commit)
785{
786 unsigned int i;
787 int ret = 0;
788
789 for (i = 0; i < num; i++) {
790 struct uverbs_attr_bundle_hash *curr_bundle =
791 &attrs_bundle->hash[i];
792 const struct uverbs_attr_spec_hash *curr_spec_bucket =
793 spec_hash[i];
794 unsigned int j;
795
796 for (j = 0; j < curr_bundle->num_attrs; j++) {
797 struct uverbs_attr *attr;
798 const struct uverbs_attr_spec *spec;
799
800 if (!uverbs_attr_is_valid_in_hash(curr_bundle, j))
801 continue;
802
803 attr = &curr_bundle->attrs[j];
804 spec = &curr_spec_bucket->attrs[j];
805
806 if (spec->type == UVERBS_ATTR_TYPE_IDR ||
807 spec->type == UVERBS_ATTR_TYPE_FD) {
808 int current_ret;
809
810 current_ret = uverbs_finalize_object(attr->obj_attr.uobject,
811 spec->obj.access,
812 commit);
813 if (!ret)
814 ret = current_ret;
815 }
816 }
817 }
818 return ret;
819}