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 <asm/atomic.h>
34 #include <linux/file.h>
35 #include <linux/lockdep.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 
uverbs_uobject_free(struct kref * ref)45 static 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  */
uverbs_uobject_put(struct ib_uobject * uobject)55 void uverbs_uobject_put(struct ib_uobject *uobject)
56 {
57 	kref_put(&uobject->ref, uverbs_uobject_free);
58 }
59 EXPORT_SYMBOL(uverbs_uobject_put);
60 
uverbs_try_lock_object(struct ib_uobject * uobj,enum rdma_lookup_mode mode)61 static int 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 }
87 
assert_uverbs_usecnt(struct ib_uobject * uobj,enum rdma_lookup_mode mode)88 static void assert_uverbs_usecnt(struct ib_uobject *uobj,
89 				 enum rdma_lookup_mode mode)
90 {
91 #ifdef CONFIG_LOCKDEP
92 	switch (mode) {
93 	case UVERBS_LOOKUP_READ:
94 		WARN_ON(atomic_read(&uobj->usecnt) <= 0);
95 		break;
96 	case UVERBS_LOOKUP_WRITE:
97 		WARN_ON(atomic_read(&uobj->usecnt) != -1);
98 		break;
99 	case UVERBS_LOOKUP_DESTROY:
100 		break;
101 	}
102 #endif
103 }
104 
105 /*
106  * This must be called with the hw_destroy_rwsem locked for read or write,
107  * also the uobject itself must be locked for write.
108  *
109  * Upon return the HW object is guaranteed to be destroyed.
110  *
111  * For RDMA_REMOVE_ABORT, the hw_destroy_rwsem is not required to be held,
112  * however the type's allocat_commit function cannot have been called and the
113  * uobject cannot be on the uobjects_lists
114  *
115  * For RDMA_REMOVE_DESTROY the caller shold be holding a kref (eg via
116  * rdma_lookup_get_uobject) and the object is left in a state where the caller
117  * needs to call rdma_lookup_put_uobject.
118  *
119  * For all other destroy modes this function internally unlocks the uobject
120  * and consumes the kref on the uobj.
121  */
uverbs_destroy_uobject(struct ib_uobject * uobj,enum rdma_remove_reason reason,struct uverbs_attr_bundle * attrs)122 static int uverbs_destroy_uobject(struct ib_uobject *uobj,
123 				  enum rdma_remove_reason reason,
124 				  struct uverbs_attr_bundle *attrs)
125 {
126 	struct ib_uverbs_file *ufile = attrs->ufile;
127 	unsigned long flags;
128 	int ret;
129 
130 	lockdep_assert_held(&ufile->hw_destroy_rwsem);
131 	assert_uverbs_usecnt(uobj, UVERBS_LOOKUP_WRITE);
132 
133 	if (reason == RDMA_REMOVE_ABORT) {
134 		WARN_ON(!list_empty(&uobj->list));
135 		WARN_ON(!uobj->context);
136 		uobj->uapi_object->type_class->alloc_abort(uobj);
137 	} else if (uobj->object) {
138 		ret = uobj->uapi_object->type_class->destroy_hw(uobj, reason,
139 								attrs);
140 		if (ret) {
141 			if (ib_is_destroy_retryable(ret, reason, uobj))
142 				return ret;
143 
144 			/* Nothing to be done, dangle the memory and move on */
145 			WARN(true,
146 			     "ib_uverbs: failed to remove uobject id %d, driver err=%d",
147 			     uobj->id, ret);
148 		}
149 
150 		uobj->object = NULL;
151 	}
152 
153 	uobj->context = NULL;
154 
155 	/*
156 	 * For DESTROY the usecnt is held write locked, the caller is expected
157 	 * to put it unlock and put the object when done with it. Only DESTROY
158 	 * can remove the IDR handle.
159 	 */
160 	if (reason != RDMA_REMOVE_DESTROY)
161 		atomic_set(&uobj->usecnt, 0);
162 	else
163 		uobj->uapi_object->type_class->remove_handle(uobj);
164 
165 	if (!list_empty(&uobj->list)) {
166 		spin_lock_irqsave(&ufile->uobjects_lock, flags);
167 		list_del_init(&uobj->list);
168 		spin_unlock_irqrestore(&ufile->uobjects_lock, flags);
169 
170 		/*
171 		 * Pairs with the get in rdma_alloc_commit_uobject(), could
172 		 * destroy uobj.
173 		 */
174 		uverbs_uobject_put(uobj);
175 	}
176 
177 	/*
178 	 * When aborting the stack kref remains owned by the core code, and is
179 	 * not transferred into the type. Pairs with the get in alloc_uobj
180 	 */
181 	if (reason == RDMA_REMOVE_ABORT)
182 		uverbs_uobject_put(uobj);
183 
184 	return 0;
185 }
186 
187 /*
188  * This calls uverbs_destroy_uobject() using the RDMA_REMOVE_DESTROY
189  * sequence. It should only be used from command callbacks. On success the
190  * caller must pair this with rdma_lookup_put_uobject(LOOKUP_WRITE). This
191  * version requires the caller to have already obtained an
192  * LOOKUP_DESTROY uobject kref.
193  */
uobj_destroy(struct ib_uobject * uobj,struct uverbs_attr_bundle * attrs)194 int uobj_destroy(struct ib_uobject *uobj, struct uverbs_attr_bundle *attrs)
195 {
196 	struct ib_uverbs_file *ufile = attrs->ufile;
197 	int ret;
198 
199 	down_read(&ufile->hw_destroy_rwsem);
200 
201 	ret = uverbs_try_lock_object(uobj, UVERBS_LOOKUP_WRITE);
202 	if (ret)
203 		goto out_unlock;
204 
205 	ret = uverbs_destroy_uobject(uobj, RDMA_REMOVE_DESTROY, attrs);
206 	if (ret) {
207 		atomic_set(&uobj->usecnt, 0);
208 		goto out_unlock;
209 	}
210 
211 out_unlock:
212 	up_read(&ufile->hw_destroy_rwsem);
213 	return ret;
214 }
215 
216 /*
217  * uobj_get_destroy destroys the HW object and returns a handle to the uobj
218  * with a NULL object pointer. The caller must pair this with
219  * uverbs_put_destroy.
220  */
__uobj_get_destroy(const struct uverbs_api_object * obj,u32 id,struct uverbs_attr_bundle * attrs)221 struct ib_uobject *__uobj_get_destroy(const struct uverbs_api_object *obj,
222 				      u32 id, struct uverbs_attr_bundle *attrs)
223 {
224 	struct ib_uobject *uobj;
225 	int ret;
226 
227 	uobj = rdma_lookup_get_uobject(obj, attrs->ufile, id,
228 				       UVERBS_LOOKUP_DESTROY, attrs);
229 	if (IS_ERR(uobj))
230 		return uobj;
231 
232 	ret = uobj_destroy(uobj, attrs);
233 	if (ret) {
234 		rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_DESTROY);
235 		return ERR_PTR(ret);
236 	}
237 
238 	return uobj;
239 }
240 
241 /*
242  * Does both uobj_get_destroy() and uobj_put_destroy().  Returns 0 on success
243  * (negative errno on failure). For use by callers that do not need the uobj.
244  */
__uobj_perform_destroy(const struct uverbs_api_object * obj,u32 id,struct uverbs_attr_bundle * attrs)245 int __uobj_perform_destroy(const struct uverbs_api_object *obj, u32 id,
246 			   struct uverbs_attr_bundle *attrs)
247 {
248 	struct ib_uobject *uobj;
249 
250 	uobj = __uobj_get_destroy(obj, id, attrs);
251 	if (IS_ERR(uobj))
252 		return PTR_ERR(uobj);
253 
254 	rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_WRITE);
255 	return 0;
256 }
257 
258 /* alloc_uobj must be undone by uverbs_destroy_uobject() */
alloc_uobj(struct uverbs_attr_bundle * attrs,const struct uverbs_api_object * obj)259 static struct ib_uobject *alloc_uobj(struct uverbs_attr_bundle *attrs,
260 				     const struct uverbs_api_object *obj)
261 {
262 	struct ib_uverbs_file *ufile = attrs->ufile;
263 	struct ib_uobject *uobj;
264 
265 	if (!attrs->context) {
266 		struct ib_ucontext *ucontext =
267 			ib_uverbs_get_ucontext_file(ufile);
268 
269 		if (IS_ERR(ucontext))
270 			return ERR_CAST(ucontext);
271 		attrs->context = ucontext;
272 	}
273 
274 	uobj = kzalloc(obj->type_attrs->obj_size, GFP_KERNEL);
275 	if (!uobj)
276 		return ERR_PTR(-ENOMEM);
277 	/*
278 	 * user_handle should be filled by the handler,
279 	 * The object is added to the list in the commit stage.
280 	 */
281 	uobj->ufile = ufile;
282 	uobj->context = attrs->context;
283 	INIT_LIST_HEAD(&uobj->list);
284 	uobj->uapi_object = obj;
285 	/*
286 	 * Allocated objects start out as write locked to deny any other
287 	 * syscalls from accessing them until they are committed. See
288 	 * rdma_alloc_commit_uobject
289 	 */
290 	atomic_set(&uobj->usecnt, -1);
291 	kref_init(&uobj->ref);
292 
293 	return uobj;
294 }
295 
idr_add_uobj(struct ib_uobject * uobj)296 static int idr_add_uobj(struct ib_uobject *uobj)
297 {
298        /*
299         * We start with allocating an idr pointing to NULL. This represents an
300         * object which isn't initialized yet. We'll replace it later on with
301         * the real object once we commit.
302         */
303 	return xa_alloc(&uobj->ufile->idr, &uobj->id, NULL, xa_limit_32b,
304 			GFP_KERNEL);
305 }
306 
307 /* Returns the ib_uobject or an error. The caller should check for IS_ERR. */
308 static struct ib_uobject *
lookup_get_idr_uobject(const struct uverbs_api_object * obj,struct ib_uverbs_file * ufile,s64 id,enum rdma_lookup_mode mode)309 lookup_get_idr_uobject(const struct uverbs_api_object *obj,
310 		       struct ib_uverbs_file *ufile, s64 id,
311 		       enum rdma_lookup_mode mode)
312 {
313 	struct ib_uobject *uobj;
314 
315 	if (id < 0 || id > ULONG_MAX)
316 		return ERR_PTR(-EINVAL);
317 
318 	rcu_read_lock();
319 	/*
320 	 * The idr_find is guaranteed to return a pointer to something that
321 	 * isn't freed yet, or NULL, as the free after idr_remove goes through
322 	 * kfree_rcu(). However the object may still have been released and
323 	 * kfree() could be called at any time.
324 	 */
325 	uobj = xa_load(&ufile->idr, id);
326 	if (!uobj || !kref_get_unless_zero(&uobj->ref))
327 		uobj = ERR_PTR(-ENOENT);
328 	rcu_read_unlock();
329 	return uobj;
330 }
331 
332 static struct ib_uobject *
lookup_get_fd_uobject(const struct uverbs_api_object * obj,struct ib_uverbs_file * ufile,s64 id,enum rdma_lookup_mode mode)333 lookup_get_fd_uobject(const struct uverbs_api_object *obj,
334 		      struct ib_uverbs_file *ufile, s64 id,
335 		      enum rdma_lookup_mode mode)
336 {
337 	const struct uverbs_obj_fd_type *fd_type;
338 	struct file *f;
339 	struct ib_uobject *uobject;
340 	int fdno = id;
341 
342 	if (fdno != id)
343 		return ERR_PTR(-EINVAL);
344 
345 	if (mode != UVERBS_LOOKUP_READ)
346 		return ERR_PTR(-EOPNOTSUPP);
347 
348 	if (!obj->type_attrs)
349 		return ERR_PTR(-EIO);
350 	fd_type =
351 		container_of(obj->type_attrs, struct uverbs_obj_fd_type, type);
352 
353 	f = fget(fdno);
354 	if (!f)
355 		return ERR_PTR(-EBADF);
356 
357 	uobject = f->private_data;
358 	/*
359 	 * fget(id) ensures we are not currently running
360 	 * uverbs_uobject_fd_release(), and the caller is expected to ensure
361 	 * that release is never done while a call to lookup is possible.
362 	 */
363 	if (f->f_op != fd_type->fops) {
364 		fput(f);
365 		return ERR_PTR(-EBADF);
366 	}
367 
368 	uverbs_uobject_get(uobject);
369 	return uobject;
370 }
371 
rdma_lookup_get_uobject(const struct uverbs_api_object * obj,struct ib_uverbs_file * ufile,s64 id,enum rdma_lookup_mode mode,struct uverbs_attr_bundle * attrs)372 struct ib_uobject *rdma_lookup_get_uobject(const struct uverbs_api_object *obj,
373 					   struct ib_uverbs_file *ufile, s64 id,
374 					   enum rdma_lookup_mode mode,
375 					   struct uverbs_attr_bundle *attrs)
376 {
377 	struct ib_uobject *uobj;
378 	int ret;
379 
380 	if (obj == ERR_PTR(-ENOMSG)) {
381 		/* must be UVERBS_IDR_ANY_OBJECT, see uapi_get_object() */
382 		uobj = lookup_get_idr_uobject(NULL, ufile, id, mode);
383 		if (IS_ERR(uobj))
384 			return uobj;
385 	} else {
386 		if (IS_ERR(obj))
387 			return ERR_PTR(-EINVAL);
388 
389 		uobj = obj->type_class->lookup_get(obj, ufile, id, mode);
390 		if (IS_ERR(uobj))
391 			return uobj;
392 
393 		if (uobj->uapi_object != obj) {
394 			ret = -EINVAL;
395 			goto free;
396 		}
397 	}
398 
399 	/*
400 	 * If we have been disassociated block every command except for
401 	 * DESTROY based commands.
402 	 */
403 	if (mode != UVERBS_LOOKUP_DESTROY &&
404 	    !srcu_dereference(ufile->device->ib_dev,
405 			      &ufile->device->disassociate_srcu)) {
406 		ret = -EIO;
407 		goto free;
408 	}
409 
410 	ret = uverbs_try_lock_object(uobj, mode);
411 	if (ret)
412 		goto free;
413 	if (attrs)
414 		attrs->context = uobj->context;
415 
416 	return uobj;
417 free:
418 	uobj->uapi_object->type_class->lookup_put(uobj, mode);
419 	uverbs_uobject_put(uobj);
420 	return ERR_PTR(ret);
421 }
422 
423 static struct ib_uobject *
alloc_begin_idr_uobject(const struct uverbs_api_object * obj,struct uverbs_attr_bundle * attrs)424 alloc_begin_idr_uobject(const struct uverbs_api_object *obj,
425 			struct uverbs_attr_bundle *attrs)
426 {
427 	int ret;
428 	struct ib_uobject *uobj;
429 
430 	uobj = alloc_uobj(attrs, obj);
431 	if (IS_ERR(uobj))
432 		return uobj;
433 
434 	ret = idr_add_uobj(uobj);
435 	if (ret)
436 		goto uobj_put;
437 
438 	ret = ib_rdmacg_try_charge(&uobj->cg_obj, uobj->context->device,
439 				   RDMACG_RESOURCE_HCA_OBJECT);
440 	if (ret)
441 		goto remove;
442 
443 	return uobj;
444 
445 remove:
446 	xa_erase(&attrs->ufile->idr, uobj->id);
447 uobj_put:
448 	uverbs_uobject_put(uobj);
449 	return ERR_PTR(ret);
450 }
451 
452 static struct ib_uobject *
alloc_begin_fd_uobject(const struct uverbs_api_object * obj,struct uverbs_attr_bundle * attrs)453 alloc_begin_fd_uobject(const struct uverbs_api_object *obj,
454 		       struct uverbs_attr_bundle *attrs)
455 {
456 	const struct uverbs_obj_fd_type *fd_type =
457 		container_of(obj->type_attrs, struct uverbs_obj_fd_type, type);
458 	int new_fd;
459 	struct ib_uobject *uobj;
460 	struct file *filp;
461 
462 	if (WARN_ON(fd_type->fops->release != &uverbs_uobject_fd_release))
463 		return ERR_PTR(-EINVAL);
464 
465 	new_fd = get_unused_fd_flags(O_CLOEXEC);
466 	if (new_fd < 0)
467 		return ERR_PTR(new_fd);
468 
469 	uobj = alloc_uobj(attrs, obj);
470 	if (IS_ERR(uobj))
471 		goto err_fd;
472 
473 	/* Note that uverbs_uobject_fd_release() is called during abort */
474 	filp = alloc_file(fd_type->flags, fd_type->fops);
475 	if (IS_ERR(filp)) {
476 		uobj = ERR_CAST(filp);
477 		goto err_uobj;
478 	}
479 	uobj->object = filp;
480 
481 	uobj->id = new_fd;
482 	return uobj;
483 
484 err_uobj:
485 	uverbs_uobject_put(uobj);
486 err_fd:
487 	put_unused_fd(new_fd);
488 	return uobj;
489 }
490 
rdma_alloc_begin_uobject(const struct uverbs_api_object * obj,struct uverbs_attr_bundle * attrs)491 struct ib_uobject *rdma_alloc_begin_uobject(const struct uverbs_api_object *obj,
492 					    struct uverbs_attr_bundle *attrs)
493 {
494 	struct ib_uverbs_file *ufile = attrs->ufile;
495 	struct ib_uobject *ret;
496 
497 	if (IS_ERR(obj))
498 		return ERR_PTR(-EINVAL);
499 
500 	/*
501 	 * The hw_destroy_rwsem is held across the entire object creation and
502 	 * released during rdma_alloc_commit_uobject or
503 	 * rdma_alloc_abort_uobject
504 	 */
505 	if (!down_read_trylock(&ufile->hw_destroy_rwsem))
506 		return ERR_PTR(-EIO);
507 
508 	ret = obj->type_class->alloc_begin(obj, attrs);
509 	if (IS_ERR(ret)) {
510 		up_read(&ufile->hw_destroy_rwsem);
511 		return ret;
512 	}
513 	return ret;
514 }
515 
alloc_abort_idr_uobject(struct ib_uobject * uobj)516 static void alloc_abort_idr_uobject(struct ib_uobject *uobj)
517 {
518 	xa_erase(&uobj->ufile->idr, uobj->id);
519 }
520 
destroy_hw_idr_uobject(struct ib_uobject * uobj,enum rdma_remove_reason why,struct uverbs_attr_bundle * attrs)521 static int __must_check destroy_hw_idr_uobject(struct ib_uobject *uobj,
522 					       enum rdma_remove_reason why,
523 					       struct uverbs_attr_bundle *attrs)
524 {
525 	const struct uverbs_obj_idr_type *idr_type =
526 		container_of(uobj->uapi_object->type_attrs,
527 			     struct uverbs_obj_idr_type, type);
528 	int ret = idr_type->destroy_object(uobj, why, attrs);
529 
530 	/*
531 	 * We can only fail gracefully if the user requested to destroy the
532 	 * object or when a retry may be called upon an error.
533 	 * In the rest of the cases, just remove whatever you can.
534 	 */
535 	if (ib_is_destroy_retryable(ret, why, uobj))
536 		return ret;
537 
538 	if (why == RDMA_REMOVE_ABORT)
539 		return 0;
540 
541 	return 0;
542 }
543 
remove_handle_idr_uobject(struct ib_uobject * uobj)544 static void remove_handle_idr_uobject(struct ib_uobject *uobj)
545 {
546 	xa_erase(&uobj->ufile->idr, uobj->id);
547 	/* Matches the kref in alloc_commit_idr_uobject */
548 	uverbs_uobject_put(uobj);
549 }
550 
alloc_abort_fd_uobject(struct ib_uobject * uobj)551 static void alloc_abort_fd_uobject(struct ib_uobject *uobj)
552 {
553 	struct file *filp = uobj->object;
554 
555 	fput(filp);
556 	put_unused_fd(uobj->id);
557 }
558 
destroy_hw_fd_uobject(struct ib_uobject * uobj,enum rdma_remove_reason why,struct uverbs_attr_bundle * attrs)559 static int __must_check destroy_hw_fd_uobject(struct ib_uobject *uobj,
560 					      enum rdma_remove_reason why,
561 					      struct uverbs_attr_bundle *attrs)
562 {
563 	const struct uverbs_obj_fd_type *fd_type = container_of(
564 		uobj->uapi_object->type_attrs, struct uverbs_obj_fd_type, type);
565 	int ret = fd_type->destroy_object(uobj, why);
566 
567 	if (ib_is_destroy_retryable(ret, why, uobj))
568 		return ret;
569 
570 	return 0;
571 }
572 
remove_handle_fd_uobject(struct ib_uobject * uobj)573 static void remove_handle_fd_uobject(struct ib_uobject *uobj)
574 {
575 }
576 
alloc_commit_idr_uobject(struct ib_uobject * uobj)577 static void alloc_commit_idr_uobject(struct ib_uobject *uobj)
578 {
579 	struct ib_uverbs_file *ufile = uobj->ufile;
580 	void *old;
581 
582 	/*
583 	 * We already allocated this IDR with a NULL object, so
584 	 * this shouldn't fail.
585 	 *
586 	 * NOTE: Storing the uobj transfers our kref on uobj to the XArray.
587 	 * It will be put by remove_commit_idr_uobject()
588 	 */
589 	old = xa_store(&ufile->idr, uobj->id, uobj, GFP_KERNEL);
590 	WARN_ON(old != NULL);
591 }
592 
alloc_commit_fd_uobject(struct ib_uobject * uobj)593 static void alloc_commit_fd_uobject(struct ib_uobject *uobj)
594 {
595 	int fd = uobj->id;
596 	struct file *filp = uobj->object;
597 
598 	/* Matching put will be done in uverbs_uobject_fd_release() */
599 	kref_get(&uobj->ufile->ref);
600 
601 	/* This shouldn't be used anymore. Use the file object instead */
602 	uobj->id = 0;
603 
604 	/*
605 	 * NOTE: Once we install the file we loose ownership of our kref on
606 	 * uobj. It will be put by uverbs_uobject_fd_release()
607 	 */
608 	filp->private_data = uobj;
609 	fd_install(fd, filp);
610 }
611 
612 /*
613  * In all cases rdma_alloc_commit_uobject() consumes the kref to uobj and the
614  * caller can no longer assume uobj is valid. If this function fails it
615  * destroys the uboject, including the attached HW object.
616  */
rdma_alloc_commit_uobject(struct ib_uobject * uobj,struct uverbs_attr_bundle * attrs)617 void rdma_alloc_commit_uobject(struct ib_uobject *uobj,
618 			       struct uverbs_attr_bundle *attrs)
619 {
620 	struct ib_uverbs_file *ufile = attrs->ufile;
621 
622 	/* alloc_commit consumes the uobj kref */
623 	uobj->uapi_object->type_class->alloc_commit(uobj);
624 
625 	/* kref is held so long as the uobj is on the uobj list. */
626 	uverbs_uobject_get(uobj);
627 	spin_lock_irq(&ufile->uobjects_lock);
628 	list_add(&uobj->list, &ufile->uobjects);
629 	spin_unlock_irq(&ufile->uobjects_lock);
630 
631 	/* matches atomic_set(-1) in alloc_uobj */
632 	atomic_set(&uobj->usecnt, 0);
633 
634 	/* Matches the down_read in rdma_alloc_begin_uobject */
635 	up_read(&ufile->hw_destroy_rwsem);
636 }
637 
638 /*
639  * This consumes the kref for uobj. It is up to the caller to unwind the HW
640  * object and anything else connected to uobj before calling this.
641  */
rdma_alloc_abort_uobject(struct ib_uobject * uobj,struct uverbs_attr_bundle * attrs)642 void rdma_alloc_abort_uobject(struct ib_uobject *uobj,
643 			      struct uverbs_attr_bundle *attrs)
644 {
645 	struct ib_uverbs_file *ufile = uobj->ufile;
646 
647 	uverbs_destroy_uobject(uobj, RDMA_REMOVE_ABORT, attrs);
648 
649 	/* Matches the down_read in rdma_alloc_begin_uobject */
650 	up_read(&ufile->hw_destroy_rwsem);
651 }
652 
lookup_put_idr_uobject(struct ib_uobject * uobj,enum rdma_lookup_mode mode)653 static void lookup_put_idr_uobject(struct ib_uobject *uobj,
654 				   enum rdma_lookup_mode mode)
655 {
656 }
657 
lookup_put_fd_uobject(struct ib_uobject * uobj,enum rdma_lookup_mode mode)658 static void lookup_put_fd_uobject(struct ib_uobject *uobj,
659 				  enum rdma_lookup_mode mode)
660 {
661 	struct file *filp = uobj->object;
662 
663 	WARN_ON(mode != UVERBS_LOOKUP_READ);
664 	/*
665 	 * This indirectly calls uverbs_uobject_fd_release() and free the
666 	 * object
667 	 */
668 	fput(filp);
669 }
670 
rdma_lookup_put_uobject(struct ib_uobject * uobj,enum rdma_lookup_mode mode)671 void rdma_lookup_put_uobject(struct ib_uobject *uobj,
672 			     enum rdma_lookup_mode mode)
673 {
674 	assert_uverbs_usecnt(uobj, mode);
675 	uobj->uapi_object->type_class->lookup_put(uobj, mode);
676 	/*
677 	 * In order to unlock an object, either decrease its usecnt for
678 	 * read access or zero it in case of exclusive access. See
679 	 * uverbs_try_lock_object for locking schema information.
680 	 */
681 	switch (mode) {
682 	case UVERBS_LOOKUP_READ:
683 		atomic_dec(&uobj->usecnt);
684 		break;
685 	case UVERBS_LOOKUP_WRITE:
686 		atomic_set(&uobj->usecnt, 0);
687 		break;
688 	case UVERBS_LOOKUP_DESTROY:
689 		break;
690 	}
691 
692 	/* Pairs with the kref obtained by type->lookup_get */
693 	uverbs_uobject_put(uobj);
694 }
695 
setup_ufile_idr_uobject(struct ib_uverbs_file * ufile)696 void setup_ufile_idr_uobject(struct ib_uverbs_file *ufile)
697 {
698 	xa_init_flags(&ufile->idr, XA_FLAGS_ALLOC);
699 }
700 
release_ufile_idr_uobject(struct ib_uverbs_file * ufile)701 void release_ufile_idr_uobject(struct ib_uverbs_file *ufile)
702 {
703 	struct ib_uobject *entry;
704 	unsigned long id;
705 
706 	/*
707 	 * At this point uverbs_cleanup_ufile() is guaranteed to have run, and
708 	 * there are no HW objects left, however the xarray is still populated
709 	 * with anything that has not been cleaned up by userspace. Since the
710 	 * kref on ufile is 0, nothing is allowed to call lookup_get.
711 	 *
712 	 * This is an optimized equivalent to remove_handle_idr_uobject
713 	 */
714 	xa_for_each(&ufile->idr, id, entry) {
715 		WARN_ON(entry->object);
716 		uverbs_uobject_put(entry);
717 	}
718 
719 	xa_destroy(&ufile->idr);
720 }
721 
722 const struct uverbs_obj_type_class uverbs_idr_class = {
723 	.alloc_begin = alloc_begin_idr_uobject,
724 	.lookup_get = lookup_get_idr_uobject,
725 	.alloc_commit = alloc_commit_idr_uobject,
726 	.alloc_abort = alloc_abort_idr_uobject,
727 	.lookup_put = lookup_put_idr_uobject,
728 	.destroy_hw = destroy_hw_idr_uobject,
729 	.remove_handle = remove_handle_idr_uobject,
730 };
731 EXPORT_SYMBOL(uverbs_idr_class);
732 
733 /*
734  * Users of UVERBS_TYPE_ALLOC_FD should set this function as the struct
735  * file_operations release method.
736  */
uverbs_uobject_fd_release(struct inode * inode,struct file * filp)737 int uverbs_uobject_fd_release(struct inode *inode, struct file *filp)
738 {
739 	struct ib_uverbs_file *ufile;
740 	struct ib_uobject *uobj;
741 
742 	/*
743 	 * This can only happen if the fput came from alloc_abort_fd_uobject()
744 	 */
745 	if (!filp->private_data)
746 		return 0;
747 	uobj = filp->private_data;
748 	ufile = uobj->ufile;
749 
750 	if (down_read_trylock(&ufile->hw_destroy_rwsem)) {
751 		struct uverbs_attr_bundle attrs = {
752 			.context = uobj->context,
753 			.ufile = ufile,
754 		};
755 
756 		/*
757 		 * lookup_get_fd_uobject holds the kref on the struct file any
758 		 * time a FD uobj is locked, which prevents this release
759 		 * method from being invoked. Meaning we can always get the
760 		 * write lock here, or we have a kernel bug.
761 		 */
762 		WARN_ON(uverbs_try_lock_object(uobj, UVERBS_LOOKUP_WRITE));
763 		uverbs_destroy_uobject(uobj, RDMA_REMOVE_CLOSE, &attrs);
764 		up_read(&ufile->hw_destroy_rwsem);
765 	}
766 
767 	/* Matches the get in alloc_commit_fd_uobject() */
768 	kref_put(&ufile->ref, ib_uverbs_release_file);
769 
770 	/* Pairs with filp->private_data in alloc_begin_fd_uobject */
771 	uverbs_uobject_put(uobj);
772 	return 0;
773 }
774 EXPORT_SYMBOL(uverbs_uobject_fd_release);
775 
776 /*
777  * Drop the ucontext off the ufile and completely disconnect it from the
778  * ib_device
779  */
ufile_destroy_ucontext(struct ib_uverbs_file * ufile,enum rdma_remove_reason reason)780 static void ufile_destroy_ucontext(struct ib_uverbs_file *ufile,
781 				   enum rdma_remove_reason reason)
782 {
783 	struct ib_ucontext *ucontext = ufile->ucontext;
784 	struct ib_device *ib_dev = ucontext->device;
785 
786 	/*
787 	 * If we are closing the FD then the user mmap VMAs must have
788 	 * already been destroyed as they hold on to the filep, otherwise
789 	 * they need to be zap'd.
790 	 */
791 	if (reason == RDMA_REMOVE_DRIVER_REMOVE) {
792 		uverbs_user_mmap_disassociate(ufile);
793 		if (ib_dev->disassociate_ucontext)
794 			ib_dev->disassociate_ucontext(ucontext);
795 	}
796 
797 	ib_dev->dealloc_ucontext(ucontext);
798 	WARN_ON(!xa_empty(&ucontext->mmap_xa));
799 	kfree(ucontext);
800 
801 	ufile->ucontext = NULL;
802 }
803 
__uverbs_cleanup_ufile(struct ib_uverbs_file * ufile,enum rdma_remove_reason reason)804 static int __uverbs_cleanup_ufile(struct ib_uverbs_file *ufile,
805 				  enum rdma_remove_reason reason)
806 {
807 	struct ib_uobject *obj, *next_obj;
808 	int ret = -EINVAL;
809 	struct uverbs_attr_bundle attrs = { .ufile = ufile };
810 
811 	/*
812 	 * This shouldn't run while executing other commands on this
813 	 * context. Thus, the only thing we should take care of is
814 	 * releasing a FD while traversing this list. The FD could be
815 	 * closed and released from the _release fop of this FD.
816 	 * In order to mitigate this, we add a lock.
817 	 * We take and release the lock per traversal in order to let
818 	 * other threads (which might still use the FDs) chance to run.
819 	 */
820 	list_for_each_entry_safe(obj, next_obj, &ufile->uobjects, list) {
821 		attrs.context = obj->context;
822 		/*
823 		 * if we hit this WARN_ON, that means we are
824 		 * racing with a lookup_get.
825 		 */
826 		WARN_ON(uverbs_try_lock_object(obj, UVERBS_LOOKUP_WRITE));
827 		if (!uverbs_destroy_uobject(obj, reason, &attrs))
828 			ret = 0;
829 		else
830 			atomic_set(&obj->usecnt, 0);
831 	}
832 	return ret;
833 }
834 
835 /*
836  * Destroy the uncontext and every uobject associated with it.
837  *
838  * This is internally locked and can be called in parallel from multiple
839  * contexts.
840  */
uverbs_destroy_ufile_hw(struct ib_uverbs_file * ufile,enum rdma_remove_reason reason)841 void uverbs_destroy_ufile_hw(struct ib_uverbs_file *ufile,
842 			     enum rdma_remove_reason reason)
843 {
844 	down_write(&ufile->hw_destroy_rwsem);
845 
846 	/*
847 	 * If a ucontext was never created then we can't have any uobjects to
848 	 * cleanup, nothing to do.
849 	 */
850 	if (!ufile->ucontext)
851 		goto done;
852 
853 	ufile->ucontext->closing = true;
854 	ufile->ucontext->cleanup_retryable = true;
855 	while (!list_empty(&ufile->uobjects))
856 		if (__uverbs_cleanup_ufile(ufile, reason)) {
857 			/*
858 			 * No entry was cleaned-up successfully during this
859 			 * iteration
860 			 */
861 			break;
862 		}
863 
864 	ufile->ucontext->cleanup_retryable = false;
865 	if (!list_empty(&ufile->uobjects))
866 		__uverbs_cleanup_ufile(ufile, reason);
867 
868 	ufile_destroy_ucontext(ufile, reason);
869 
870 done:
871 	up_write(&ufile->hw_destroy_rwsem);
872 }
873 
874 const struct uverbs_obj_type_class uverbs_fd_class = {
875 	.alloc_begin = alloc_begin_fd_uobject,
876 	.lookup_get = lookup_get_fd_uobject,
877 	.alloc_commit = alloc_commit_fd_uobject,
878 	.alloc_abort = alloc_abort_fd_uobject,
879 	.lookup_put = lookup_put_fd_uobject,
880 	.destroy_hw = destroy_hw_fd_uobject,
881 	.remove_handle = remove_handle_fd_uobject,
882 };
883 EXPORT_SYMBOL(uverbs_fd_class);
884 
885 struct ib_uobject *
uverbs_get_uobject_from_file(u16 object_id,enum uverbs_obj_access access,s64 id,struct uverbs_attr_bundle * attrs)886 uverbs_get_uobject_from_file(u16 object_id, enum uverbs_obj_access access,
887 			     s64 id, struct uverbs_attr_bundle *attrs)
888 {
889 	const struct uverbs_api_object *obj =
890 		uapi_get_object(attrs->ufile->device->uapi, object_id);
891 
892 	switch (access) {
893 	case UVERBS_ACCESS_READ:
894 		return rdma_lookup_get_uobject(obj, attrs->ufile, id,
895 					       UVERBS_LOOKUP_READ, attrs);
896 	case UVERBS_ACCESS_DESTROY:
897 		/* Actual destruction is done inside uverbs_handle_method */
898 		return rdma_lookup_get_uobject(obj, attrs->ufile, id,
899 					       UVERBS_LOOKUP_DESTROY, attrs);
900 	case UVERBS_ACCESS_WRITE:
901 		return rdma_lookup_get_uobject(obj, attrs->ufile, id,
902 					       UVERBS_LOOKUP_WRITE, attrs);
903 	case UVERBS_ACCESS_NEW:
904 		return rdma_alloc_begin_uobject(obj, attrs);
905 	default:
906 		WARN_ON(true);
907 		return ERR_PTR(-EOPNOTSUPP);
908 	}
909 }
910 
uverbs_finalize_object(struct ib_uobject * uobj,enum uverbs_obj_access access,bool commit,struct uverbs_attr_bundle * attrs)911 void uverbs_finalize_object(struct ib_uobject *uobj,
912 			    enum uverbs_obj_access access, bool commit,
913 			    struct uverbs_attr_bundle *attrs)
914 {
915 	/*
916 	 * refcounts should be handled at the object level and not at the
917 	 * uobject level. Refcounts of the objects themselves are done in
918 	 * handlers.
919 	 */
920 
921 	switch (access) {
922 	case UVERBS_ACCESS_READ:
923 		rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_READ);
924 		break;
925 	case UVERBS_ACCESS_WRITE:
926 		rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_WRITE);
927 		break;
928 	case UVERBS_ACCESS_DESTROY:
929 		if (uobj)
930 			rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_DESTROY);
931 		break;
932 	case UVERBS_ACCESS_NEW:
933 		if (commit)
934 			rdma_alloc_commit_uobject(uobj, attrs);
935 		else
936 			rdma_alloc_abort_uobject(uobj, attrs);
937 		break;
938 	default:
939 		WARN_ON(true);
940 	}
941 }
942