1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Code related to the io_uring_register() syscall
4 *
5 * Copyright (C) 2023 Jens Axboe
6 */
7#include <linux/kernel.h>
8#include <linux/errno.h>
9#include <linux/syscalls.h>
10#include <linux/refcount.h>
11#include <linux/bits.h>
12#include <linux/fs.h>
13#include <linux/file.h>
14#include <linux/slab.h>
15#include <linux/uaccess.h>
16#include <linux/nospec.h>
17#include <linux/compat.h>
18#include <linux/io_uring.h>
19#include <linux/io_uring_types.h>
20
21#include "io_uring.h"
22#include "opdef.h"
23#include "tctx.h"
24#include "rsrc.h"
25#include "sqpoll.h"
26#include "register.h"
27#include "cancel.h"
28#include "kbuf.h"
29#include "napi.h"
30
31#define IORING_MAX_RESTRICTIONS	(IORING_RESTRICTION_LAST + \
32				 IORING_REGISTER_LAST + IORING_OP_LAST)
33
34static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
35			       unsigned int eventfd_async)
36{
37	struct io_ev_fd *ev_fd;
38	__s32 __user *fds = arg;
39	int fd;
40
41	ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
42					lockdep_is_held(&ctx->uring_lock));
43	if (ev_fd)
44		return -EBUSY;
45
46	if (copy_from_user(&fd, fds, sizeof(*fds)))
47		return -EFAULT;
48
49	ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
50	if (!ev_fd)
51		return -ENOMEM;
52
53	ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
54	if (IS_ERR(ev_fd->cq_ev_fd)) {
55		int ret = PTR_ERR(ev_fd->cq_ev_fd);
56		kfree(ev_fd);
57		return ret;
58	}
59
60	spin_lock(&ctx->completion_lock);
61	ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
62	spin_unlock(&ctx->completion_lock);
63
64	ev_fd->eventfd_async = eventfd_async;
65	ctx->has_evfd = true;
66	rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
67	atomic_set(&ev_fd->refs, 1);
68	atomic_set(&ev_fd->ops, 0);
69	return 0;
70}
71
72int io_eventfd_unregister(struct io_ring_ctx *ctx)
73{
74	struct io_ev_fd *ev_fd;
75
76	ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
77					lockdep_is_held(&ctx->uring_lock));
78	if (ev_fd) {
79		ctx->has_evfd = false;
80		rcu_assign_pointer(ctx->io_ev_fd, NULL);
81		if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
82			call_rcu(&ev_fd->rcu, io_eventfd_ops);
83		return 0;
84	}
85
86	return -ENXIO;
87}
88
89static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
90			   unsigned nr_args)
91{
92	struct io_uring_probe *p;
93	size_t size;
94	int i, ret;
95
96	size = struct_size(p, ops, nr_args);
97	if (size == SIZE_MAX)
98		return -EOVERFLOW;
99	p = kzalloc(size, GFP_KERNEL);
100	if (!p)
101		return -ENOMEM;
102
103	ret = -EFAULT;
104	if (copy_from_user(p, arg, size))
105		goto out;
106	ret = -EINVAL;
107	if (memchr_inv(p, 0, size))
108		goto out;
109
110	p->last_op = IORING_OP_LAST - 1;
111	if (nr_args > IORING_OP_LAST)
112		nr_args = IORING_OP_LAST;
113
114	for (i = 0; i < nr_args; i++) {
115		p->ops[i].op = i;
116		if (!io_issue_defs[i].not_supported)
117			p->ops[i].flags = IO_URING_OP_SUPPORTED;
118	}
119	p->ops_len = i;
120
121	ret = 0;
122	if (copy_to_user(arg, p, size))
123		ret = -EFAULT;
124out:
125	kfree(p);
126	return ret;
127}
128
129int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
130{
131	const struct cred *creds;
132
133	creds = xa_erase(&ctx->personalities, id);
134	if (creds) {
135		put_cred(creds);
136		return 0;
137	}
138
139	return -EINVAL;
140}
141
142
143static int io_register_personality(struct io_ring_ctx *ctx)
144{
145	const struct cred *creds;
146	u32 id;
147	int ret;
148
149	creds = get_current_cred();
150
151	ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
152			XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
153	if (ret < 0) {
154		put_cred(creds);
155		return ret;
156	}
157	return id;
158}
159
160static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
161					   void __user *arg, unsigned int nr_args)
162{
163	struct io_uring_restriction *res;
164	size_t size;
165	int i, ret;
166
167	/* Restrictions allowed only if rings started disabled */
168	if (!(ctx->flags & IORING_SETUP_R_DISABLED))
169		return -EBADFD;
170
171	/* We allow only a single restrictions registration */
172	if (ctx->restrictions.registered)
173		return -EBUSY;
174
175	if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
176		return -EINVAL;
177
178	size = array_size(nr_args, sizeof(*res));
179	if (size == SIZE_MAX)
180		return -EOVERFLOW;
181
182	res = memdup_user(arg, size);
183	if (IS_ERR(res))
184		return PTR_ERR(res);
185
186	ret = 0;
187
188	for (i = 0; i < nr_args; i++) {
189		switch (res[i].opcode) {
190		case IORING_RESTRICTION_REGISTER_OP:
191			if (res[i].register_op >= IORING_REGISTER_LAST) {
192				ret = -EINVAL;
193				goto out;
194			}
195
196			__set_bit(res[i].register_op,
197				  ctx->restrictions.register_op);
198			break;
199		case IORING_RESTRICTION_SQE_OP:
200			if (res[i].sqe_op >= IORING_OP_LAST) {
201				ret = -EINVAL;
202				goto out;
203			}
204
205			__set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
206			break;
207		case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
208			ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
209			break;
210		case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
211			ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
212			break;
213		default:
214			ret = -EINVAL;
215			goto out;
216		}
217	}
218
219out:
220	/* Reset all restrictions if an error happened */
221	if (ret != 0)
222		memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
223	else
224		ctx->restrictions.registered = true;
225
226	kfree(res);
227	return ret;
228}
229
230static int io_register_enable_rings(struct io_ring_ctx *ctx)
231{
232	if (!(ctx->flags & IORING_SETUP_R_DISABLED))
233		return -EBADFD;
234
235	if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
236		WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
237		/*
238		 * Lazy activation attempts would fail if it was polled before
239		 * submitter_task is set.
240		 */
241		if (wq_has_sleeper(&ctx->poll_wq))
242			io_activate_pollwq(ctx);
243	}
244
245	if (ctx->restrictions.registered)
246		ctx->restricted = 1;
247
248	ctx->flags &= ~IORING_SETUP_R_DISABLED;
249	if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
250		wake_up(&ctx->sq_data->wait);
251	return 0;
252}
253
254static __cold int __io_register_iowq_aff(struct io_ring_ctx *ctx,
255					 cpumask_var_t new_mask)
256{
257	int ret;
258
259	if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
260		ret = io_wq_cpu_affinity(current->io_uring, new_mask);
261	} else {
262		mutex_unlock(&ctx->uring_lock);
263		ret = io_sqpoll_wq_cpu_affinity(ctx, new_mask);
264		mutex_lock(&ctx->uring_lock);
265	}
266
267	return ret;
268}
269
270static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
271				       void __user *arg, unsigned len)
272{
273	cpumask_var_t new_mask;
274	int ret;
275
276	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
277		return -ENOMEM;
278
279	cpumask_clear(new_mask);
280	if (len > cpumask_size())
281		len = cpumask_size();
282
283#ifdef CONFIG_COMPAT
284	if (in_compat_syscall())
285		ret = compat_get_bitmap(cpumask_bits(new_mask),
286					(const compat_ulong_t __user *)arg,
287					len * 8 /* CHAR_BIT */);
288	else
289#endif
290		ret = copy_from_user(new_mask, arg, len);
291
292	if (ret) {
293		free_cpumask_var(new_mask);
294		return -EFAULT;
295	}
296
297	ret = __io_register_iowq_aff(ctx, new_mask);
298	free_cpumask_var(new_mask);
299	return ret;
300}
301
302static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
303{
304	return __io_register_iowq_aff(ctx, NULL);
305}
306
307static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
308					       void __user *arg)
309	__must_hold(&ctx->uring_lock)
310{
311	struct io_tctx_node *node;
312	struct io_uring_task *tctx = NULL;
313	struct io_sq_data *sqd = NULL;
314	__u32 new_count[2];
315	int i, ret;
316
317	if (copy_from_user(new_count, arg, sizeof(new_count)))
318		return -EFAULT;
319	for (i = 0; i < ARRAY_SIZE(new_count); i++)
320		if (new_count[i] > INT_MAX)
321			return -EINVAL;
322
323	if (ctx->flags & IORING_SETUP_SQPOLL) {
324		sqd = ctx->sq_data;
325		if (sqd) {
326			/*
327			 * Observe the correct sqd->lock -> ctx->uring_lock
328			 * ordering. Fine to drop uring_lock here, we hold
329			 * a ref to the ctx.
330			 */
331			refcount_inc(&sqd->refs);
332			mutex_unlock(&ctx->uring_lock);
333			mutex_lock(&sqd->lock);
334			mutex_lock(&ctx->uring_lock);
335			if (sqd->thread)
336				tctx = sqd->thread->io_uring;
337		}
338	} else {
339		tctx = current->io_uring;
340	}
341
342	BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
343
344	for (i = 0; i < ARRAY_SIZE(new_count); i++)
345		if (new_count[i])
346			ctx->iowq_limits[i] = new_count[i];
347	ctx->iowq_limits_set = true;
348
349	if (tctx && tctx->io_wq) {
350		ret = io_wq_max_workers(tctx->io_wq, new_count);
351		if (ret)
352			goto err;
353	} else {
354		memset(new_count, 0, sizeof(new_count));
355	}
356
357	if (sqd) {
358		mutex_unlock(&sqd->lock);
359		io_put_sq_data(sqd);
360	}
361
362	if (copy_to_user(arg, new_count, sizeof(new_count)))
363		return -EFAULT;
364
365	/* that's it for SQPOLL, only the SQPOLL task creates requests */
366	if (sqd)
367		return 0;
368
369	/* now propagate the restriction to all registered users */
370	list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
371		tctx = node->task->io_uring;
372		if (WARN_ON_ONCE(!tctx->io_wq))
373			continue;
374
375		for (i = 0; i < ARRAY_SIZE(new_count); i++)
376			new_count[i] = ctx->iowq_limits[i];
377		/* ignore errors, it always returns zero anyway */
378		(void)io_wq_max_workers(tctx->io_wq, new_count);
379	}
380	return 0;
381err:
382	if (sqd) {
383		mutex_unlock(&sqd->lock);
384		io_put_sq_data(sqd);
385	}
386	return ret;
387}
388
389static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
390			       void __user *arg, unsigned nr_args)
391	__releases(ctx->uring_lock)
392	__acquires(ctx->uring_lock)
393{
394	int ret;
395
396	/*
397	 * We don't quiesce the refs for register anymore and so it can't be
398	 * dying as we're holding a file ref here.
399	 */
400	if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
401		return -ENXIO;
402
403	if (ctx->submitter_task && ctx->submitter_task != current)
404		return -EEXIST;
405
406	if (ctx->restricted) {
407		opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
408		if (!test_bit(opcode, ctx->restrictions.register_op))
409			return -EACCES;
410	}
411
412	switch (opcode) {
413	case IORING_REGISTER_BUFFERS:
414		ret = -EFAULT;
415		if (!arg)
416			break;
417		ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
418		break;
419	case IORING_UNREGISTER_BUFFERS:
420		ret = -EINVAL;
421		if (arg || nr_args)
422			break;
423		ret = io_sqe_buffers_unregister(ctx);
424		break;
425	case IORING_REGISTER_FILES:
426		ret = -EFAULT;
427		if (!arg)
428			break;
429		ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
430		break;
431	case IORING_UNREGISTER_FILES:
432		ret = -EINVAL;
433		if (arg || nr_args)
434			break;
435		ret = io_sqe_files_unregister(ctx);
436		break;
437	case IORING_REGISTER_FILES_UPDATE:
438		ret = io_register_files_update(ctx, arg, nr_args);
439		break;
440	case IORING_REGISTER_EVENTFD:
441		ret = -EINVAL;
442		if (nr_args != 1)
443			break;
444		ret = io_eventfd_register(ctx, arg, 0);
445		break;
446	case IORING_REGISTER_EVENTFD_ASYNC:
447		ret = -EINVAL;
448		if (nr_args != 1)
449			break;
450		ret = io_eventfd_register(ctx, arg, 1);
451		break;
452	case IORING_UNREGISTER_EVENTFD:
453		ret = -EINVAL;
454		if (arg || nr_args)
455			break;
456		ret = io_eventfd_unregister(ctx);
457		break;
458	case IORING_REGISTER_PROBE:
459		ret = -EINVAL;
460		if (!arg || nr_args > 256)
461			break;
462		ret = io_probe(ctx, arg, nr_args);
463		break;
464	case IORING_REGISTER_PERSONALITY:
465		ret = -EINVAL;
466		if (arg || nr_args)
467			break;
468		ret = io_register_personality(ctx);
469		break;
470	case IORING_UNREGISTER_PERSONALITY:
471		ret = -EINVAL;
472		if (arg)
473			break;
474		ret = io_unregister_personality(ctx, nr_args);
475		break;
476	case IORING_REGISTER_ENABLE_RINGS:
477		ret = -EINVAL;
478		if (arg || nr_args)
479			break;
480		ret = io_register_enable_rings(ctx);
481		break;
482	case IORING_REGISTER_RESTRICTIONS:
483		ret = io_register_restrictions(ctx, arg, nr_args);
484		break;
485	case IORING_REGISTER_FILES2:
486		ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
487		break;
488	case IORING_REGISTER_FILES_UPDATE2:
489		ret = io_register_rsrc_update(ctx, arg, nr_args,
490					      IORING_RSRC_FILE);
491		break;
492	case IORING_REGISTER_BUFFERS2:
493		ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
494		break;
495	case IORING_REGISTER_BUFFERS_UPDATE:
496		ret = io_register_rsrc_update(ctx, arg, nr_args,
497					      IORING_RSRC_BUFFER);
498		break;
499	case IORING_REGISTER_IOWQ_AFF:
500		ret = -EINVAL;
501		if (!arg || !nr_args)
502			break;
503		ret = io_register_iowq_aff(ctx, arg, nr_args);
504		break;
505	case IORING_UNREGISTER_IOWQ_AFF:
506		ret = -EINVAL;
507		if (arg || nr_args)
508			break;
509		ret = io_unregister_iowq_aff(ctx);
510		break;
511	case IORING_REGISTER_IOWQ_MAX_WORKERS:
512		ret = -EINVAL;
513		if (!arg || nr_args != 2)
514			break;
515		ret = io_register_iowq_max_workers(ctx, arg);
516		break;
517	case IORING_REGISTER_RING_FDS:
518		ret = io_ringfd_register(ctx, arg, nr_args);
519		break;
520	case IORING_UNREGISTER_RING_FDS:
521		ret = io_ringfd_unregister(ctx, arg, nr_args);
522		break;
523	case IORING_REGISTER_PBUF_RING:
524		ret = -EINVAL;
525		if (!arg || nr_args != 1)
526			break;
527		ret = io_register_pbuf_ring(ctx, arg);
528		break;
529	case IORING_UNREGISTER_PBUF_RING:
530		ret = -EINVAL;
531		if (!arg || nr_args != 1)
532			break;
533		ret = io_unregister_pbuf_ring(ctx, arg);
534		break;
535	case IORING_REGISTER_SYNC_CANCEL:
536		ret = -EINVAL;
537		if (!arg || nr_args != 1)
538			break;
539		ret = io_sync_cancel(ctx, arg);
540		break;
541	case IORING_REGISTER_FILE_ALLOC_RANGE:
542		ret = -EINVAL;
543		if (!arg || nr_args)
544			break;
545		ret = io_register_file_alloc_range(ctx, arg);
546		break;
547	case IORING_REGISTER_PBUF_STATUS:
548		ret = -EINVAL;
549		if (!arg || nr_args != 1)
550			break;
551		ret = io_register_pbuf_status(ctx, arg);
552		break;
553	case IORING_REGISTER_NAPI:
554		ret = -EINVAL;
555		if (!arg || nr_args != 1)
556			break;
557		ret = io_register_napi(ctx, arg);
558		break;
559	case IORING_UNREGISTER_NAPI:
560		ret = -EINVAL;
561		if (nr_args != 1)
562			break;
563		ret = io_unregister_napi(ctx, arg);
564		break;
565	default:
566		ret = -EINVAL;
567		break;
568	}
569
570	return ret;
571}
572
573SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
574		void __user *, arg, unsigned int, nr_args)
575{
576	struct io_ring_ctx *ctx;
577	long ret = -EBADF;
578	struct file *file;
579	bool use_registered_ring;
580
581	use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
582	opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
583
584	if (opcode >= IORING_REGISTER_LAST)
585		return -EINVAL;
586
587	if (use_registered_ring) {
588		/*
589		 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
590		 * need only dereference our task private array to find it.
591		 */
592		struct io_uring_task *tctx = current->io_uring;
593
594		if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
595			return -EINVAL;
596		fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
597		file = tctx->registered_rings[fd];
598		if (unlikely(!file))
599			return -EBADF;
600	} else {
601		file = fget(fd);
602		if (unlikely(!file))
603			return -EBADF;
604		ret = -EOPNOTSUPP;
605		if (!io_is_uring_fops(file))
606			goto out_fput;
607	}
608
609	ctx = file->private_data;
610
611	mutex_lock(&ctx->uring_lock);
612	ret = __io_uring_register(ctx, opcode, arg, nr_args);
613	mutex_unlock(&ctx->uring_lock);
614	trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
615out_fput:
616	if (!use_registered_ring)
617		fput(file);
618	return ret;
619}
620