xref: /linux/io_uring/rsrc.c (revision ba6e0e5cb5b2c2e736e16b4aead816450a8718e6)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
4 #include <linux/fs.h>
5 #include <linux/file.h>
6 #include <linux/mm.h>
7 #include <linux/slab.h>
8 #include <linux/nospec.h>
9 #include <linux/hugetlb.h>
10 #include <linux/compat.h>
11 #include <linux/io_uring.h>
12 
13 #include <uapi/linux/io_uring.h>
14 
15 #include "io_uring.h"
16 #include "openclose.h"
17 #include "rsrc.h"
18 
19 struct io_rsrc_update {
20 	struct file			*file;
21 	u64				arg;
22 	u32				nr_args;
23 	u32				offset;
24 };
25 
26 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
27 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
28 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
29 				  struct io_mapped_ubuf **pimu,
30 				  struct page **last_hpage);
31 
32 /* only define max */
33 #define IORING_MAX_FIXED_FILES	(1U << 20)
34 #define IORING_MAX_REG_BUFFERS	(1U << 14)
35 
36 static const struct io_mapped_ubuf dummy_ubuf = {
37 	/* set invalid range, so io_import_fixed() fails meeting it */
38 	.ubuf = -1UL,
39 	.ubuf_end = 0,
40 };
41 
42 int __io_account_mem(struct user_struct *user, unsigned long nr_pages)
43 {
44 	unsigned long page_limit, cur_pages, new_pages;
45 
46 	if (!nr_pages)
47 		return 0;
48 
49 	/* Don't allow more pages than we can safely lock */
50 	page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
51 
52 	cur_pages = atomic_long_read(&user->locked_vm);
53 	do {
54 		new_pages = cur_pages + nr_pages;
55 		if (new_pages > page_limit)
56 			return -ENOMEM;
57 	} while (!atomic_long_try_cmpxchg(&user->locked_vm,
58 					  &cur_pages, new_pages));
59 	return 0;
60 }
61 
62 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
63 {
64 	if (ctx->user)
65 		__io_unaccount_mem(ctx->user, nr_pages);
66 
67 	if (ctx->mm_account)
68 		atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
69 }
70 
71 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
72 {
73 	int ret;
74 
75 	if (ctx->user) {
76 		ret = __io_account_mem(ctx->user, nr_pages);
77 		if (ret)
78 			return ret;
79 	}
80 
81 	if (ctx->mm_account)
82 		atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
83 
84 	return 0;
85 }
86 
87 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
88 		       void __user *arg, unsigned index)
89 {
90 	struct iovec __user *src;
91 
92 #ifdef CONFIG_COMPAT
93 	if (ctx->compat) {
94 		struct compat_iovec __user *ciovs;
95 		struct compat_iovec ciov;
96 
97 		ciovs = (struct compat_iovec __user *) arg;
98 		if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
99 			return -EFAULT;
100 
101 		dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
102 		dst->iov_len = ciov.iov_len;
103 		return 0;
104 	}
105 #endif
106 	src = (struct iovec __user *) arg;
107 	if (copy_from_user(dst, &src[index], sizeof(*dst)))
108 		return -EFAULT;
109 	return 0;
110 }
111 
112 static int io_buffer_validate(struct iovec *iov)
113 {
114 	unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
115 
116 	/*
117 	 * Don't impose further limits on the size and buffer
118 	 * constraints here, we'll -EINVAL later when IO is
119 	 * submitted if they are wrong.
120 	 */
121 	if (!iov->iov_base)
122 		return iov->iov_len ? -EFAULT : 0;
123 	if (!iov->iov_len)
124 		return -EFAULT;
125 
126 	/* arbitrary limit, but we need something */
127 	if (iov->iov_len > SZ_1G)
128 		return -EFAULT;
129 
130 	if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
131 		return -EOVERFLOW;
132 
133 	return 0;
134 }
135 
136 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
137 {
138 	struct io_mapped_ubuf *imu = *slot;
139 	unsigned int i;
140 
141 	if (imu != &dummy_ubuf) {
142 		for (i = 0; i < imu->nr_bvecs; i++)
143 			unpin_user_page(imu->bvec[i].bv_page);
144 		if (imu->acct_pages)
145 			io_unaccount_mem(ctx, imu->acct_pages);
146 		kvfree(imu);
147 	}
148 	*slot = NULL;
149 }
150 
151 static void io_rsrc_put_work(struct io_rsrc_node *node)
152 {
153 	struct io_rsrc_put *prsrc = &node->item;
154 
155 	if (prsrc->tag)
156 		io_post_aux_cqe(node->ctx, prsrc->tag, 0, 0);
157 
158 	switch (node->type) {
159 	case IORING_RSRC_FILE:
160 		io_rsrc_file_put(node->ctx, prsrc);
161 		break;
162 	case IORING_RSRC_BUFFER:
163 		io_rsrc_buf_put(node->ctx, prsrc);
164 		break;
165 	default:
166 		WARN_ON_ONCE(1);
167 		break;
168 	}
169 }
170 
171 void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *node)
172 {
173 	if (!io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache))
174 		kfree(node);
175 }
176 
177 void io_rsrc_node_ref_zero(struct io_rsrc_node *node)
178 	__must_hold(&node->ctx->uring_lock)
179 {
180 	struct io_ring_ctx *ctx = node->ctx;
181 
182 	while (!list_empty(&ctx->rsrc_ref_list)) {
183 		node = list_first_entry(&ctx->rsrc_ref_list,
184 					    struct io_rsrc_node, node);
185 		/* recycle ref nodes in order */
186 		if (node->refs)
187 			break;
188 		list_del(&node->node);
189 
190 		if (likely(!node->empty))
191 			io_rsrc_put_work(node);
192 		io_rsrc_node_destroy(ctx, node);
193 	}
194 	if (list_empty(&ctx->rsrc_ref_list) && unlikely(ctx->rsrc_quiesce))
195 		wake_up_all(&ctx->rsrc_quiesce_wq);
196 }
197 
198 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
199 {
200 	struct io_rsrc_node *ref_node;
201 	struct io_cache_entry *entry;
202 
203 	entry = io_alloc_cache_get(&ctx->rsrc_node_cache);
204 	if (entry) {
205 		ref_node = container_of(entry, struct io_rsrc_node, cache);
206 	} else {
207 		ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
208 		if (!ref_node)
209 			return NULL;
210 	}
211 
212 	ref_node->ctx = ctx;
213 	ref_node->empty = 0;
214 	ref_node->refs = 1;
215 	return ref_node;
216 }
217 
218 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
219 				      struct io_ring_ctx *ctx)
220 {
221 	struct io_rsrc_node *backup;
222 	DEFINE_WAIT(we);
223 	int ret;
224 
225 	/* As We may drop ->uring_lock, other task may have started quiesce */
226 	if (data->quiesce)
227 		return -ENXIO;
228 
229 	backup = io_rsrc_node_alloc(ctx);
230 	if (!backup)
231 		return -ENOMEM;
232 	ctx->rsrc_node->empty = true;
233 	ctx->rsrc_node->type = -1;
234 	list_add_tail(&ctx->rsrc_node->node, &ctx->rsrc_ref_list);
235 	io_put_rsrc_node(ctx, ctx->rsrc_node);
236 	ctx->rsrc_node = backup;
237 
238 	if (list_empty(&ctx->rsrc_ref_list))
239 		return 0;
240 
241 	if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
242 		atomic_set(&ctx->cq_wait_nr, 1);
243 		smp_mb();
244 	}
245 
246 	ctx->rsrc_quiesce++;
247 	data->quiesce = true;
248 	do {
249 		prepare_to_wait(&ctx->rsrc_quiesce_wq, &we, TASK_INTERRUPTIBLE);
250 		mutex_unlock(&ctx->uring_lock);
251 
252 		ret = io_run_task_work_sig(ctx);
253 		if (ret < 0) {
254 			mutex_lock(&ctx->uring_lock);
255 			if (list_empty(&ctx->rsrc_ref_list))
256 				ret = 0;
257 			break;
258 		}
259 
260 		schedule();
261 		__set_current_state(TASK_RUNNING);
262 		mutex_lock(&ctx->uring_lock);
263 		ret = 0;
264 	} while (!list_empty(&ctx->rsrc_ref_list));
265 
266 	finish_wait(&ctx->rsrc_quiesce_wq, &we);
267 	data->quiesce = false;
268 	ctx->rsrc_quiesce--;
269 
270 	if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
271 		atomic_set(&ctx->cq_wait_nr, 0);
272 		smp_mb();
273 	}
274 	return ret;
275 }
276 
277 static void io_free_page_table(void **table, size_t size)
278 {
279 	unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
280 
281 	for (i = 0; i < nr_tables; i++)
282 		kfree(table[i]);
283 	kfree(table);
284 }
285 
286 static void io_rsrc_data_free(struct io_rsrc_data *data)
287 {
288 	size_t size = data->nr * sizeof(data->tags[0][0]);
289 
290 	if (data->tags)
291 		io_free_page_table((void **)data->tags, size);
292 	kfree(data);
293 }
294 
295 static __cold void **io_alloc_page_table(size_t size)
296 {
297 	unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
298 	size_t init_size = size;
299 	void **table;
300 
301 	table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
302 	if (!table)
303 		return NULL;
304 
305 	for (i = 0; i < nr_tables; i++) {
306 		unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
307 
308 		table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
309 		if (!table[i]) {
310 			io_free_page_table(table, init_size);
311 			return NULL;
312 		}
313 		size -= this_size;
314 	}
315 	return table;
316 }
317 
318 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, int type,
319 				     u64 __user *utags,
320 				     unsigned nr, struct io_rsrc_data **pdata)
321 {
322 	struct io_rsrc_data *data;
323 	int ret = 0;
324 	unsigned i;
325 
326 	data = kzalloc(sizeof(*data), GFP_KERNEL);
327 	if (!data)
328 		return -ENOMEM;
329 	data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
330 	if (!data->tags) {
331 		kfree(data);
332 		return -ENOMEM;
333 	}
334 
335 	data->nr = nr;
336 	data->ctx = ctx;
337 	data->rsrc_type = type;
338 	if (utags) {
339 		ret = -EFAULT;
340 		for (i = 0; i < nr; i++) {
341 			u64 *tag_slot = io_get_tag_slot(data, i);
342 
343 			if (copy_from_user(tag_slot, &utags[i],
344 					   sizeof(*tag_slot)))
345 				goto fail;
346 		}
347 	}
348 	*pdata = data;
349 	return 0;
350 fail:
351 	io_rsrc_data_free(data);
352 	return ret;
353 }
354 
355 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
356 				 struct io_uring_rsrc_update2 *up,
357 				 unsigned nr_args)
358 {
359 	u64 __user *tags = u64_to_user_ptr(up->tags);
360 	__s32 __user *fds = u64_to_user_ptr(up->data);
361 	struct io_rsrc_data *data = ctx->file_data;
362 	struct io_fixed_file *file_slot;
363 	int fd, i, err = 0;
364 	unsigned int done;
365 
366 	if (!ctx->file_data)
367 		return -ENXIO;
368 	if (up->offset + nr_args > ctx->nr_user_files)
369 		return -EINVAL;
370 
371 	for (done = 0; done < nr_args; done++) {
372 		u64 tag = 0;
373 
374 		if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
375 		    copy_from_user(&fd, &fds[done], sizeof(fd))) {
376 			err = -EFAULT;
377 			break;
378 		}
379 		if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
380 			err = -EINVAL;
381 			break;
382 		}
383 		if (fd == IORING_REGISTER_FILES_SKIP)
384 			continue;
385 
386 		i = array_index_nospec(up->offset + done, ctx->nr_user_files);
387 		file_slot = io_fixed_file_slot(&ctx->file_table, i);
388 
389 		if (file_slot->file_ptr) {
390 			err = io_queue_rsrc_removal(data, i,
391 						    io_slot_file(file_slot));
392 			if (err)
393 				break;
394 			file_slot->file_ptr = 0;
395 			io_file_bitmap_clear(&ctx->file_table, i);
396 		}
397 		if (fd != -1) {
398 			struct file *file = fget(fd);
399 
400 			if (!file) {
401 				err = -EBADF;
402 				break;
403 			}
404 			/*
405 			 * Don't allow io_uring instances to be registered. If
406 			 * UNIX isn't enabled, then this causes a reference
407 			 * cycle and this instance can never get freed. If UNIX
408 			 * is enabled we'll handle it just fine, but there's
409 			 * still no point in allowing a ring fd as it doesn't
410 			 * support regular read/write anyway.
411 			 */
412 			if (io_is_uring_fops(file)) {
413 				fput(file);
414 				err = -EBADF;
415 				break;
416 			}
417 			err = io_scm_file_account(ctx, file);
418 			if (err) {
419 				fput(file);
420 				break;
421 			}
422 			*io_get_tag_slot(data, i) = tag;
423 			io_fixed_file_set(file_slot, file);
424 			io_file_bitmap_set(&ctx->file_table, i);
425 		}
426 	}
427 	return done ? done : err;
428 }
429 
430 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
431 				   struct io_uring_rsrc_update2 *up,
432 				   unsigned int nr_args)
433 {
434 	u64 __user *tags = u64_to_user_ptr(up->tags);
435 	struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
436 	struct page *last_hpage = NULL;
437 	__u32 done;
438 	int i, err;
439 
440 	if (!ctx->buf_data)
441 		return -ENXIO;
442 	if (up->offset + nr_args > ctx->nr_user_bufs)
443 		return -EINVAL;
444 
445 	for (done = 0; done < nr_args; done++) {
446 		struct io_mapped_ubuf *imu;
447 		u64 tag = 0;
448 
449 		err = io_copy_iov(ctx, &iov, iovs, done);
450 		if (err)
451 			break;
452 		if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
453 			err = -EFAULT;
454 			break;
455 		}
456 		err = io_buffer_validate(&iov);
457 		if (err)
458 			break;
459 		if (!iov.iov_base && tag) {
460 			err = -EINVAL;
461 			break;
462 		}
463 		err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
464 		if (err)
465 			break;
466 
467 		i = array_index_nospec(up->offset + done, ctx->nr_user_bufs);
468 		if (ctx->user_bufs[i] != &dummy_ubuf) {
469 			err = io_queue_rsrc_removal(ctx->buf_data, i,
470 						    ctx->user_bufs[i]);
471 			if (unlikely(err)) {
472 				io_buffer_unmap(ctx, &imu);
473 				break;
474 			}
475 			ctx->user_bufs[i] = (struct io_mapped_ubuf *)&dummy_ubuf;
476 		}
477 
478 		ctx->user_bufs[i] = imu;
479 		*io_get_tag_slot(ctx->buf_data, i) = tag;
480 	}
481 	return done ? done : err;
482 }
483 
484 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
485 				     struct io_uring_rsrc_update2 *up,
486 				     unsigned nr_args)
487 {
488 	__u32 tmp;
489 
490 	lockdep_assert_held(&ctx->uring_lock);
491 
492 	if (check_add_overflow(up->offset, nr_args, &tmp))
493 		return -EOVERFLOW;
494 
495 	switch (type) {
496 	case IORING_RSRC_FILE:
497 		return __io_sqe_files_update(ctx, up, nr_args);
498 	case IORING_RSRC_BUFFER:
499 		return __io_sqe_buffers_update(ctx, up, nr_args);
500 	}
501 	return -EINVAL;
502 }
503 
504 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
505 			     unsigned nr_args)
506 {
507 	struct io_uring_rsrc_update2 up;
508 
509 	if (!nr_args)
510 		return -EINVAL;
511 	memset(&up, 0, sizeof(up));
512 	if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
513 		return -EFAULT;
514 	if (up.resv || up.resv2)
515 		return -EINVAL;
516 	return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
517 }
518 
519 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
520 			    unsigned size, unsigned type)
521 {
522 	struct io_uring_rsrc_update2 up;
523 
524 	if (size != sizeof(up))
525 		return -EINVAL;
526 	if (copy_from_user(&up, arg, sizeof(up)))
527 		return -EFAULT;
528 	if (!up.nr || up.resv || up.resv2)
529 		return -EINVAL;
530 	return __io_register_rsrc_update(ctx, type, &up, up.nr);
531 }
532 
533 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
534 			    unsigned int size, unsigned int type)
535 {
536 	struct io_uring_rsrc_register rr;
537 
538 	/* keep it extendible */
539 	if (size != sizeof(rr))
540 		return -EINVAL;
541 
542 	memset(&rr, 0, sizeof(rr));
543 	if (copy_from_user(&rr, arg, size))
544 		return -EFAULT;
545 	if (!rr.nr || rr.resv2)
546 		return -EINVAL;
547 	if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
548 		return -EINVAL;
549 
550 	switch (type) {
551 	case IORING_RSRC_FILE:
552 		if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
553 			break;
554 		return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
555 					     rr.nr, u64_to_user_ptr(rr.tags));
556 	case IORING_RSRC_BUFFER:
557 		if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
558 			break;
559 		return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
560 					       rr.nr, u64_to_user_ptr(rr.tags));
561 	}
562 	return -EINVAL;
563 }
564 
565 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
566 {
567 	struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
568 
569 	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
570 		return -EINVAL;
571 	if (sqe->rw_flags || sqe->splice_fd_in)
572 		return -EINVAL;
573 
574 	up->offset = READ_ONCE(sqe->off);
575 	up->nr_args = READ_ONCE(sqe->len);
576 	if (!up->nr_args)
577 		return -EINVAL;
578 	up->arg = READ_ONCE(sqe->addr);
579 	return 0;
580 }
581 
582 static int io_files_update_with_index_alloc(struct io_kiocb *req,
583 					    unsigned int issue_flags)
584 {
585 	struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
586 	__s32 __user *fds = u64_to_user_ptr(up->arg);
587 	unsigned int done;
588 	struct file *file;
589 	int ret, fd;
590 
591 	if (!req->ctx->file_data)
592 		return -ENXIO;
593 
594 	for (done = 0; done < up->nr_args; done++) {
595 		if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
596 			ret = -EFAULT;
597 			break;
598 		}
599 
600 		file = fget(fd);
601 		if (!file) {
602 			ret = -EBADF;
603 			break;
604 		}
605 		ret = io_fixed_fd_install(req, issue_flags, file,
606 					  IORING_FILE_INDEX_ALLOC);
607 		if (ret < 0)
608 			break;
609 		if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
610 			__io_close_fixed(req->ctx, issue_flags, ret);
611 			ret = -EFAULT;
612 			break;
613 		}
614 	}
615 
616 	if (done)
617 		return done;
618 	return ret;
619 }
620 
621 int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
622 {
623 	struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
624 	struct io_ring_ctx *ctx = req->ctx;
625 	struct io_uring_rsrc_update2 up2;
626 	int ret;
627 
628 	up2.offset = up->offset;
629 	up2.data = up->arg;
630 	up2.nr = 0;
631 	up2.tags = 0;
632 	up2.resv = 0;
633 	up2.resv2 = 0;
634 
635 	if (up->offset == IORING_FILE_INDEX_ALLOC) {
636 		ret = io_files_update_with_index_alloc(req, issue_flags);
637 	} else {
638 		io_ring_submit_lock(ctx, issue_flags);
639 		ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
640 						&up2, up->nr_args);
641 		io_ring_submit_unlock(ctx, issue_flags);
642 	}
643 
644 	if (ret < 0)
645 		req_set_fail(req);
646 	io_req_set_res(req, ret, 0);
647 	return IOU_OK;
648 }
649 
650 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, void *rsrc)
651 {
652 	struct io_ring_ctx *ctx = data->ctx;
653 	struct io_rsrc_node *node = ctx->rsrc_node;
654 	u64 *tag_slot = io_get_tag_slot(data, idx);
655 
656 	ctx->rsrc_node = io_rsrc_node_alloc(ctx);
657 	if (unlikely(!ctx->rsrc_node)) {
658 		ctx->rsrc_node = node;
659 		return -ENOMEM;
660 	}
661 
662 	node->item.rsrc = rsrc;
663 	node->type = data->rsrc_type;
664 	node->item.tag = *tag_slot;
665 	*tag_slot = 0;
666 	list_add_tail(&node->node, &ctx->rsrc_ref_list);
667 	io_put_rsrc_node(ctx, node);
668 	return 0;
669 }
670 
671 void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
672 {
673 	int i;
674 
675 	for (i = 0; i < ctx->nr_user_files; i++) {
676 		struct file *file = io_file_from_index(&ctx->file_table, i);
677 
678 		/* skip scm accounted files, they'll be freed by ->ring_sock */
679 		if (!file || io_file_need_scm(file))
680 			continue;
681 		io_file_bitmap_clear(&ctx->file_table, i);
682 		fput(file);
683 	}
684 
685 #if defined(CONFIG_UNIX)
686 	if (ctx->ring_sock) {
687 		struct sock *sock = ctx->ring_sock->sk;
688 		struct sk_buff *skb;
689 
690 		while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
691 			kfree_skb(skb);
692 	}
693 #endif
694 	io_free_file_tables(&ctx->file_table);
695 	io_file_table_set_alloc_range(ctx, 0, 0);
696 	io_rsrc_data_free(ctx->file_data);
697 	ctx->file_data = NULL;
698 	ctx->nr_user_files = 0;
699 }
700 
701 int io_sqe_files_unregister(struct io_ring_ctx *ctx)
702 {
703 	unsigned nr = ctx->nr_user_files;
704 	int ret;
705 
706 	if (!ctx->file_data)
707 		return -ENXIO;
708 
709 	/*
710 	 * Quiesce may unlock ->uring_lock, and while it's not held
711 	 * prevent new requests using the table.
712 	 */
713 	ctx->nr_user_files = 0;
714 	ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
715 	ctx->nr_user_files = nr;
716 	if (!ret)
717 		__io_sqe_files_unregister(ctx);
718 	return ret;
719 }
720 
721 /*
722  * Ensure the UNIX gc is aware of our file set, so we are certain that
723  * the io_uring can be safely unregistered on process exit, even if we have
724  * loops in the file referencing. We account only files that can hold other
725  * files because otherwise they can't form a loop and so are not interesting
726  * for GC.
727  */
728 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
729 {
730 #if defined(CONFIG_UNIX)
731 	struct sock *sk = ctx->ring_sock->sk;
732 	struct sk_buff_head *head = &sk->sk_receive_queue;
733 	struct scm_fp_list *fpl;
734 	struct sk_buff *skb;
735 
736 	if (likely(!io_file_need_scm(file)))
737 		return 0;
738 
739 	/*
740 	 * See if we can merge this file into an existing skb SCM_RIGHTS
741 	 * file set. If there's no room, fall back to allocating a new skb
742 	 * and filling it in.
743 	 */
744 	spin_lock_irq(&head->lock);
745 	skb = skb_peek(head);
746 	if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
747 		__skb_unlink(skb, head);
748 	else
749 		skb = NULL;
750 	spin_unlock_irq(&head->lock);
751 
752 	if (!skb) {
753 		fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
754 		if (!fpl)
755 			return -ENOMEM;
756 
757 		skb = alloc_skb(0, GFP_KERNEL);
758 		if (!skb) {
759 			kfree(fpl);
760 			return -ENOMEM;
761 		}
762 
763 		fpl->user = get_uid(current_user());
764 		fpl->max = SCM_MAX_FD;
765 		fpl->count = 0;
766 
767 		UNIXCB(skb).fp = fpl;
768 		skb->sk = sk;
769 		skb->destructor = io_uring_destruct_scm;
770 		refcount_add(skb->truesize, &sk->sk_wmem_alloc);
771 	}
772 
773 	fpl = UNIXCB(skb).fp;
774 	fpl->fp[fpl->count++] = get_file(file);
775 	unix_inflight(fpl->user, file);
776 	skb_queue_head(head, skb);
777 	fput(file);
778 #endif
779 	return 0;
780 }
781 
782 static __cold void io_rsrc_file_scm_put(struct io_ring_ctx *ctx, struct file *file)
783 {
784 #if defined(CONFIG_UNIX)
785 	struct sock *sock = ctx->ring_sock->sk;
786 	struct sk_buff_head list, *head = &sock->sk_receive_queue;
787 	struct sk_buff *skb;
788 	int i;
789 
790 	__skb_queue_head_init(&list);
791 
792 	/*
793 	 * Find the skb that holds this file in its SCM_RIGHTS. When found,
794 	 * remove this entry and rearrange the file array.
795 	 */
796 	skb = skb_dequeue(head);
797 	while (skb) {
798 		struct scm_fp_list *fp;
799 
800 		fp = UNIXCB(skb).fp;
801 		for (i = 0; i < fp->count; i++) {
802 			int left;
803 
804 			if (fp->fp[i] != file)
805 				continue;
806 
807 			unix_notinflight(fp->user, fp->fp[i]);
808 			left = fp->count - 1 - i;
809 			if (left) {
810 				memmove(&fp->fp[i], &fp->fp[i + 1],
811 						left * sizeof(struct file *));
812 			}
813 			fp->count--;
814 			if (!fp->count) {
815 				kfree_skb(skb);
816 				skb = NULL;
817 			} else {
818 				__skb_queue_tail(&list, skb);
819 			}
820 			fput(file);
821 			file = NULL;
822 			break;
823 		}
824 
825 		if (!file)
826 			break;
827 
828 		__skb_queue_tail(&list, skb);
829 
830 		skb = skb_dequeue(head);
831 	}
832 
833 	if (skb_peek(&list)) {
834 		spin_lock_irq(&head->lock);
835 		while ((skb = __skb_dequeue(&list)) != NULL)
836 			__skb_queue_tail(head, skb);
837 		spin_unlock_irq(&head->lock);
838 	}
839 #endif
840 }
841 
842 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
843 {
844 	struct file *file = prsrc->file;
845 
846 	if (likely(!io_file_need_scm(file)))
847 		fput(file);
848 	else
849 		io_rsrc_file_scm_put(ctx, file);
850 }
851 
852 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
853 			  unsigned nr_args, u64 __user *tags)
854 {
855 	__s32 __user *fds = (__s32 __user *) arg;
856 	struct file *file;
857 	int fd, ret;
858 	unsigned i;
859 
860 	if (ctx->file_data)
861 		return -EBUSY;
862 	if (!nr_args)
863 		return -EINVAL;
864 	if (nr_args > IORING_MAX_FIXED_FILES)
865 		return -EMFILE;
866 	if (nr_args > rlimit(RLIMIT_NOFILE))
867 		return -EMFILE;
868 	ret = io_rsrc_data_alloc(ctx, IORING_RSRC_FILE, tags, nr_args,
869 				 &ctx->file_data);
870 	if (ret)
871 		return ret;
872 
873 	if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
874 		io_rsrc_data_free(ctx->file_data);
875 		ctx->file_data = NULL;
876 		return -ENOMEM;
877 	}
878 
879 	for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
880 		struct io_fixed_file *file_slot;
881 
882 		if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
883 			ret = -EFAULT;
884 			goto fail;
885 		}
886 		/* allow sparse sets */
887 		if (!fds || fd == -1) {
888 			ret = -EINVAL;
889 			if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
890 				goto fail;
891 			continue;
892 		}
893 
894 		file = fget(fd);
895 		ret = -EBADF;
896 		if (unlikely(!file))
897 			goto fail;
898 
899 		/*
900 		 * Don't allow io_uring instances to be registered. If UNIX
901 		 * isn't enabled, then this causes a reference cycle and this
902 		 * instance can never get freed. If UNIX is enabled we'll
903 		 * handle it just fine, but there's still no point in allowing
904 		 * a ring fd as it doesn't support regular read/write anyway.
905 		 */
906 		if (io_is_uring_fops(file)) {
907 			fput(file);
908 			goto fail;
909 		}
910 		ret = io_scm_file_account(ctx, file);
911 		if (ret) {
912 			fput(file);
913 			goto fail;
914 		}
915 		file_slot = io_fixed_file_slot(&ctx->file_table, i);
916 		io_fixed_file_set(file_slot, file);
917 		io_file_bitmap_set(&ctx->file_table, i);
918 	}
919 
920 	/* default it to the whole table */
921 	io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
922 	return 0;
923 fail:
924 	__io_sqe_files_unregister(ctx);
925 	return ret;
926 }
927 
928 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
929 {
930 	io_buffer_unmap(ctx, &prsrc->buf);
931 	prsrc->buf = NULL;
932 }
933 
934 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
935 {
936 	unsigned int i;
937 
938 	for (i = 0; i < ctx->nr_user_bufs; i++)
939 		io_buffer_unmap(ctx, &ctx->user_bufs[i]);
940 	kfree(ctx->user_bufs);
941 	io_rsrc_data_free(ctx->buf_data);
942 	ctx->user_bufs = NULL;
943 	ctx->buf_data = NULL;
944 	ctx->nr_user_bufs = 0;
945 }
946 
947 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
948 {
949 	unsigned nr = ctx->nr_user_bufs;
950 	int ret;
951 
952 	if (!ctx->buf_data)
953 		return -ENXIO;
954 
955 	/*
956 	 * Quiesce may unlock ->uring_lock, and while it's not held
957 	 * prevent new requests using the table.
958 	 */
959 	ctx->nr_user_bufs = 0;
960 	ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
961 	ctx->nr_user_bufs = nr;
962 	if (!ret)
963 		__io_sqe_buffers_unregister(ctx);
964 	return ret;
965 }
966 
967 /*
968  * Not super efficient, but this is just a registration time. And we do cache
969  * the last compound head, so generally we'll only do a full search if we don't
970  * match that one.
971  *
972  * We check if the given compound head page has already been accounted, to
973  * avoid double accounting it. This allows us to account the full size of the
974  * page, not just the constituent pages of a huge page.
975  */
976 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
977 				  int nr_pages, struct page *hpage)
978 {
979 	int i, j;
980 
981 	/* check current page array */
982 	for (i = 0; i < nr_pages; i++) {
983 		if (!PageCompound(pages[i]))
984 			continue;
985 		if (compound_head(pages[i]) == hpage)
986 			return true;
987 	}
988 
989 	/* check previously registered pages */
990 	for (i = 0; i < ctx->nr_user_bufs; i++) {
991 		struct io_mapped_ubuf *imu = ctx->user_bufs[i];
992 
993 		for (j = 0; j < imu->nr_bvecs; j++) {
994 			if (!PageCompound(imu->bvec[j].bv_page))
995 				continue;
996 			if (compound_head(imu->bvec[j].bv_page) == hpage)
997 				return true;
998 		}
999 	}
1000 
1001 	return false;
1002 }
1003 
1004 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
1005 				 int nr_pages, struct io_mapped_ubuf *imu,
1006 				 struct page **last_hpage)
1007 {
1008 	int i, ret;
1009 
1010 	imu->acct_pages = 0;
1011 	for (i = 0; i < nr_pages; i++) {
1012 		if (!PageCompound(pages[i])) {
1013 			imu->acct_pages++;
1014 		} else {
1015 			struct page *hpage;
1016 
1017 			hpage = compound_head(pages[i]);
1018 			if (hpage == *last_hpage)
1019 				continue;
1020 			*last_hpage = hpage;
1021 			if (headpage_already_acct(ctx, pages, i, hpage))
1022 				continue;
1023 			imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1024 		}
1025 	}
1026 
1027 	if (!imu->acct_pages)
1028 		return 0;
1029 
1030 	ret = io_account_mem(ctx, imu->acct_pages);
1031 	if (ret)
1032 		imu->acct_pages = 0;
1033 	return ret;
1034 }
1035 
1036 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1037 {
1038 	unsigned long start, end, nr_pages;
1039 	struct page **pages = NULL;
1040 	int ret;
1041 
1042 	end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1043 	start = ubuf >> PAGE_SHIFT;
1044 	nr_pages = end - start;
1045 	WARN_ON(!nr_pages);
1046 
1047 	pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1048 	if (!pages)
1049 		return ERR_PTR(-ENOMEM);
1050 
1051 	mmap_read_lock(current->mm);
1052 	ret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM, pages);
1053 	mmap_read_unlock(current->mm);
1054 
1055 	/* success, mapped all pages */
1056 	if (ret == nr_pages) {
1057 		*npages = nr_pages;
1058 		return pages;
1059 	}
1060 
1061 	/* partial map, or didn't map anything */
1062 	if (ret >= 0) {
1063 		/* if we did partial map, release any pages we did get */
1064 		if (ret)
1065 			unpin_user_pages(pages, ret);
1066 		ret = -EFAULT;
1067 	}
1068 	kvfree(pages);
1069 	return ERR_PTR(ret);
1070 }
1071 
1072 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1073 				  struct io_mapped_ubuf **pimu,
1074 				  struct page **last_hpage)
1075 {
1076 	struct io_mapped_ubuf *imu = NULL;
1077 	struct page **pages = NULL;
1078 	unsigned long off;
1079 	size_t size;
1080 	int ret, nr_pages, i;
1081 	struct folio *folio = NULL;
1082 
1083 	*pimu = (struct io_mapped_ubuf *)&dummy_ubuf;
1084 	if (!iov->iov_base)
1085 		return 0;
1086 
1087 	ret = -ENOMEM;
1088 	pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1089 				&nr_pages);
1090 	if (IS_ERR(pages)) {
1091 		ret = PTR_ERR(pages);
1092 		pages = NULL;
1093 		goto done;
1094 	}
1095 
1096 	/* If it's a huge page, try to coalesce them into a single bvec entry */
1097 	if (nr_pages > 1) {
1098 		folio = page_folio(pages[0]);
1099 		for (i = 1; i < nr_pages; i++) {
1100 			/*
1101 			 * Pages must be consecutive and on the same folio for
1102 			 * this to work
1103 			 */
1104 			if (page_folio(pages[i]) != folio ||
1105 			    pages[i] != pages[i - 1] + 1) {
1106 				folio = NULL;
1107 				break;
1108 			}
1109 		}
1110 		if (folio) {
1111 			/*
1112 			 * The pages are bound to the folio, it doesn't
1113 			 * actually unpin them but drops all but one reference,
1114 			 * which is usually put down by io_buffer_unmap().
1115 			 * Note, needs a better helper.
1116 			 */
1117 			unpin_user_pages(&pages[1], nr_pages - 1);
1118 			nr_pages = 1;
1119 		}
1120 	}
1121 
1122 	imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1123 	if (!imu)
1124 		goto done;
1125 
1126 	ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1127 	if (ret) {
1128 		unpin_user_pages(pages, nr_pages);
1129 		goto done;
1130 	}
1131 
1132 	off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1133 	size = iov->iov_len;
1134 	/* store original address for later verification */
1135 	imu->ubuf = (unsigned long) iov->iov_base;
1136 	imu->ubuf_end = imu->ubuf + iov->iov_len;
1137 	imu->nr_bvecs = nr_pages;
1138 	*pimu = imu;
1139 	ret = 0;
1140 
1141 	if (folio) {
1142 		bvec_set_page(&imu->bvec[0], pages[0], size, off);
1143 		goto done;
1144 	}
1145 	for (i = 0; i < nr_pages; i++) {
1146 		size_t vec_len;
1147 
1148 		vec_len = min_t(size_t, size, PAGE_SIZE - off);
1149 		bvec_set_page(&imu->bvec[i], pages[i], vec_len, off);
1150 		off = 0;
1151 		size -= vec_len;
1152 	}
1153 done:
1154 	if (ret)
1155 		kvfree(imu);
1156 	kvfree(pages);
1157 	return ret;
1158 }
1159 
1160 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1161 {
1162 	ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1163 	return ctx->user_bufs ? 0 : -ENOMEM;
1164 }
1165 
1166 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1167 			    unsigned int nr_args, u64 __user *tags)
1168 {
1169 	struct page *last_hpage = NULL;
1170 	struct io_rsrc_data *data;
1171 	int i, ret;
1172 	struct iovec iov;
1173 
1174 	BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1175 
1176 	if (ctx->user_bufs)
1177 		return -EBUSY;
1178 	if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1179 		return -EINVAL;
1180 	ret = io_rsrc_data_alloc(ctx, IORING_RSRC_BUFFER, tags, nr_args, &data);
1181 	if (ret)
1182 		return ret;
1183 	ret = io_buffers_map_alloc(ctx, nr_args);
1184 	if (ret) {
1185 		io_rsrc_data_free(data);
1186 		return ret;
1187 	}
1188 
1189 	for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1190 		if (arg) {
1191 			ret = io_copy_iov(ctx, &iov, arg, i);
1192 			if (ret)
1193 				break;
1194 			ret = io_buffer_validate(&iov);
1195 			if (ret)
1196 				break;
1197 		} else {
1198 			memset(&iov, 0, sizeof(iov));
1199 		}
1200 
1201 		if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1202 			ret = -EINVAL;
1203 			break;
1204 		}
1205 
1206 		ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1207 					     &last_hpage);
1208 		if (ret)
1209 			break;
1210 	}
1211 
1212 	WARN_ON_ONCE(ctx->buf_data);
1213 
1214 	ctx->buf_data = data;
1215 	if (ret)
1216 		__io_sqe_buffers_unregister(ctx);
1217 	return ret;
1218 }
1219 
1220 int io_import_fixed(int ddir, struct iov_iter *iter,
1221 			   struct io_mapped_ubuf *imu,
1222 			   u64 buf_addr, size_t len)
1223 {
1224 	u64 buf_end;
1225 	size_t offset;
1226 
1227 	if (WARN_ON_ONCE(!imu))
1228 		return -EFAULT;
1229 	if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1230 		return -EFAULT;
1231 	/* not inside the mapped region */
1232 	if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1233 		return -EFAULT;
1234 
1235 	/*
1236 	 * Might not be a start of buffer, set size appropriately
1237 	 * and advance us to the beginning.
1238 	 */
1239 	offset = buf_addr - imu->ubuf;
1240 	iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1241 
1242 	if (offset) {
1243 		/*
1244 		 * Don't use iov_iter_advance() here, as it's really slow for
1245 		 * using the latter parts of a big fixed buffer - it iterates
1246 		 * over each segment manually. We can cheat a bit here, because
1247 		 * we know that:
1248 		 *
1249 		 * 1) it's a BVEC iter, we set it up
1250 		 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1251 		 *    first and last bvec
1252 		 *
1253 		 * So just find our index, and adjust the iterator afterwards.
1254 		 * If the offset is within the first bvec (or the whole first
1255 		 * bvec, just use iov_iter_advance(). This makes it easier
1256 		 * since we can just skip the first segment, which may not
1257 		 * be PAGE_SIZE aligned.
1258 		 */
1259 		const struct bio_vec *bvec = imu->bvec;
1260 
1261 		if (offset <= bvec->bv_len) {
1262 			/*
1263 			 * Note, huge pages buffers consists of one large
1264 			 * bvec entry and should always go this way. The other
1265 			 * branch doesn't expect non PAGE_SIZE'd chunks.
1266 			 */
1267 			iter->bvec = bvec;
1268 			iter->nr_segs = bvec->bv_len;
1269 			iter->count -= offset;
1270 			iter->iov_offset = offset;
1271 		} else {
1272 			unsigned long seg_skip;
1273 
1274 			/* skip first vec */
1275 			offset -= bvec->bv_len;
1276 			seg_skip = 1 + (offset >> PAGE_SHIFT);
1277 
1278 			iter->bvec = bvec + seg_skip;
1279 			iter->nr_segs -= seg_skip;
1280 			iter->count -= bvec->bv_len + offset;
1281 			iter->iov_offset = offset & ~PAGE_MASK;
1282 		}
1283 	}
1284 
1285 	return 0;
1286 }
1287