1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/file.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>
13 #include <uapi/linux/io_uring.h>
16 #include "openclose.h"
19 struct io_rsrc_update {
26 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
27 struct io_mapped_ubuf **pimu,
28 struct page **last_hpage);
30 #define IO_RSRC_REF_BATCH 100
33 #define IORING_MAX_FIXED_FILES (1U << 20)
34 #define IORING_MAX_REG_BUFFERS (1U << 14)
36 void io_rsrc_refs_drop(struct io_ring_ctx *ctx)
37 __must_hold(&ctx->uring_lock)
39 if (ctx->rsrc_cached_refs) {
40 io_rsrc_put_node(ctx->rsrc_node, ctx->rsrc_cached_refs);
41 ctx->rsrc_cached_refs = 0;
45 int __io_account_mem(struct user_struct *user, unsigned long nr_pages)
47 unsigned long page_limit, cur_pages, new_pages;
52 /* Don't allow more pages than we can safely lock */
53 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
55 cur_pages = atomic_long_read(&user->locked_vm);
57 new_pages = cur_pages + nr_pages;
58 if (new_pages > page_limit)
60 } while (!atomic_long_try_cmpxchg(&user->locked_vm,
61 &cur_pages, new_pages));
65 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
68 __io_unaccount_mem(ctx->user, nr_pages);
71 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
74 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
79 ret = __io_account_mem(ctx->user, nr_pages);
85 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
90 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
91 void __user *arg, unsigned index)
93 struct iovec __user *src;
97 struct compat_iovec __user *ciovs;
98 struct compat_iovec ciov;
100 ciovs = (struct compat_iovec __user *) arg;
101 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
104 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
105 dst->iov_len = ciov.iov_len;
109 src = (struct iovec __user *) arg;
110 if (copy_from_user(dst, &src[index], sizeof(*dst)))
115 static int io_buffer_validate(struct iovec *iov)
117 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
120 * Don't impose further limits on the size and buffer
121 * constraints here, we'll -EINVAL later when IO is
122 * submitted if they are wrong.
125 return iov->iov_len ? -EFAULT : 0;
129 /* arbitrary limit, but we need something */
130 if (iov->iov_len > SZ_1G)
133 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
139 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
141 struct io_mapped_ubuf *imu = *slot;
144 if (imu != ctx->dummy_ubuf) {
145 for (i = 0; i < imu->nr_bvecs; i++)
146 unpin_user_page(imu->bvec[i].bv_page);
148 io_unaccount_mem(ctx, imu->acct_pages);
154 void io_rsrc_refs_refill(struct io_ring_ctx *ctx)
155 __must_hold(&ctx->uring_lock)
157 ctx->rsrc_cached_refs += IO_RSRC_REF_BATCH;
158 refcount_add(IO_RSRC_REF_BATCH, &ctx->rsrc_node->refs);
161 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
163 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
164 struct io_ring_ctx *ctx = rsrc_data->ctx;
165 struct io_rsrc_put *prsrc, *tmp;
167 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
168 list_del(&prsrc->list);
171 if (ctx->flags & IORING_SETUP_IOPOLL) {
172 mutex_lock(&ctx->uring_lock);
173 io_post_aux_cqe(ctx, prsrc->tag, 0, 0);
174 mutex_unlock(&ctx->uring_lock);
176 io_post_aux_cqe(ctx, prsrc->tag, 0, 0);
180 rsrc_data->do_put(ctx, prsrc);
184 io_rsrc_node_destroy(ref_node);
185 if (atomic_dec_and_test(&rsrc_data->refs))
186 complete(&rsrc_data->done);
189 void io_rsrc_put_work(struct work_struct *work)
191 struct io_ring_ctx *ctx;
192 struct llist_node *node;
194 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
195 node = llist_del_all(&ctx->rsrc_put_llist);
198 struct io_rsrc_node *ref_node;
199 struct llist_node *next = node->next;
201 ref_node = llist_entry(node, struct io_rsrc_node, llist);
202 __io_rsrc_put_work(ref_node);
207 void io_rsrc_put_tw(struct callback_head *cb)
209 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
212 io_rsrc_put_work(&ctx->rsrc_put_work.work);
215 void io_wait_rsrc_data(struct io_rsrc_data *data)
217 if (data && !atomic_dec_and_test(&data->refs))
218 wait_for_completion(&data->done);
221 void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
226 __cold void io_rsrc_node_ref_zero(struct io_rsrc_node *node)
228 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
230 bool first_add = false;
231 unsigned long delay = HZ;
233 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
236 /* if we are mid-quiesce then do not delay */
237 if (node->rsrc_data->quiesce)
240 while (!list_empty(&ctx->rsrc_ref_list)) {
241 node = list_first_entry(&ctx->rsrc_ref_list,
242 struct io_rsrc_node, node);
243 /* recycle ref nodes in order */
246 list_del(&node->node);
247 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
249 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
254 if (ctx->submitter_task) {
255 if (!task_work_add(ctx->submitter_task, &ctx->rsrc_put_tw,
259 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
262 static struct io_rsrc_node *io_rsrc_node_alloc(void)
264 struct io_rsrc_node *ref_node;
266 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
270 refcount_set(&ref_node->refs, 1);
271 INIT_LIST_HEAD(&ref_node->node);
272 INIT_LIST_HEAD(&ref_node->rsrc_list);
273 ref_node->done = false;
277 void io_rsrc_node_switch(struct io_ring_ctx *ctx,
278 struct io_rsrc_data *data_to_kill)
279 __must_hold(&ctx->uring_lock)
281 WARN_ON_ONCE(!ctx->rsrc_backup_node);
282 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
284 io_rsrc_refs_drop(ctx);
287 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
289 rsrc_node->rsrc_data = data_to_kill;
290 spin_lock_irq(&ctx->rsrc_ref_lock);
291 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
292 spin_unlock_irq(&ctx->rsrc_ref_lock);
294 atomic_inc(&data_to_kill->refs);
296 io_rsrc_put_node(rsrc_node, 1);
297 ctx->rsrc_node = NULL;
300 if (!ctx->rsrc_node) {
301 ctx->rsrc_node = ctx->rsrc_backup_node;
302 ctx->rsrc_backup_node = NULL;
306 int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
308 if (ctx->rsrc_backup_node)
310 ctx->rsrc_backup_node = io_rsrc_node_alloc();
311 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
314 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
315 struct io_ring_ctx *ctx)
319 /* As we may drop ->uring_lock, other task may have started quiesce */
322 ret = io_rsrc_node_switch_start(ctx);
325 io_rsrc_node_switch(ctx, data);
327 /* kill initial ref, already quiesced if zero */
328 if (atomic_dec_and_test(&data->refs))
331 data->quiesce = true;
332 mutex_unlock(&ctx->uring_lock);
334 ret = io_run_task_work_sig(ctx);
336 atomic_inc(&data->refs);
337 /* wait for all works potentially completing data->done */
338 flush_delayed_work(&ctx->rsrc_put_work);
339 reinit_completion(&data->done);
340 mutex_lock(&ctx->uring_lock);
344 flush_delayed_work(&ctx->rsrc_put_work);
345 ret = wait_for_completion_interruptible(&data->done);
347 mutex_lock(&ctx->uring_lock);
348 if (atomic_read(&data->refs) <= 0)
351 * it has been revived by another thread while
354 mutex_unlock(&ctx->uring_lock);
357 data->quiesce = false;
362 static void io_free_page_table(void **table, size_t size)
364 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
366 for (i = 0; i < nr_tables; i++)
371 static void io_rsrc_data_free(struct io_rsrc_data *data)
373 size_t size = data->nr * sizeof(data->tags[0][0]);
376 io_free_page_table((void **)data->tags, size);
380 static __cold void **io_alloc_page_table(size_t size)
382 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
383 size_t init_size = size;
386 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
390 for (i = 0; i < nr_tables; i++) {
391 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
393 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
395 io_free_page_table(table, init_size);
403 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx,
404 rsrc_put_fn *do_put, u64 __user *utags,
405 unsigned nr, struct io_rsrc_data **pdata)
407 struct io_rsrc_data *data;
411 data = kzalloc(sizeof(*data), GFP_KERNEL);
414 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
422 data->do_put = do_put;
425 for (i = 0; i < nr; i++) {
426 u64 *tag_slot = io_get_tag_slot(data, i);
428 if (copy_from_user(tag_slot, &utags[i],
434 atomic_set(&data->refs, 1);
435 init_completion(&data->done);
439 io_rsrc_data_free(data);
443 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
444 struct io_uring_rsrc_update2 *up,
447 u64 __user *tags = u64_to_user_ptr(up->tags);
448 __s32 __user *fds = u64_to_user_ptr(up->data);
449 struct io_rsrc_data *data = ctx->file_data;
450 struct io_fixed_file *file_slot;
454 bool needs_switch = false;
458 if (up->offset + nr_args > ctx->nr_user_files)
461 for (done = 0; done < nr_args; done++) {
464 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
465 copy_from_user(&fd, &fds[done], sizeof(fd))) {
469 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
473 if (fd == IORING_REGISTER_FILES_SKIP)
476 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
477 file_slot = io_fixed_file_slot(&ctx->file_table, i);
479 if (file_slot->file_ptr) {
480 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
481 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
484 file_slot->file_ptr = 0;
485 io_file_bitmap_clear(&ctx->file_table, i);
495 * Don't allow io_uring instances to be registered. If
496 * UNIX isn't enabled, then this causes a reference
497 * cycle and this instance can never get freed. If UNIX
498 * is enabled we'll handle it just fine, but there's
499 * still no point in allowing a ring fd as it doesn't
500 * support regular read/write anyway.
502 if (io_is_uring_fops(file)) {
507 err = io_scm_file_account(ctx, file);
512 *io_get_tag_slot(data, i) = tag;
513 io_fixed_file_set(file_slot, file);
514 io_file_bitmap_set(&ctx->file_table, i);
519 io_rsrc_node_switch(ctx, data);
520 return done ? done : err;
523 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
524 struct io_uring_rsrc_update2 *up,
525 unsigned int nr_args)
527 u64 __user *tags = u64_to_user_ptr(up->tags);
528 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
529 struct page *last_hpage = NULL;
530 bool needs_switch = false;
536 if (up->offset + nr_args > ctx->nr_user_bufs)
539 for (done = 0; done < nr_args; done++) {
540 struct io_mapped_ubuf *imu;
541 int offset = up->offset + done;
544 err = io_copy_iov(ctx, &iov, iovs, done);
547 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
551 err = io_buffer_validate(&iov);
554 if (!iov.iov_base && tag) {
558 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
562 i = array_index_nospec(offset, ctx->nr_user_bufs);
563 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
564 err = io_queue_rsrc_removal(ctx->buf_data, i,
565 ctx->rsrc_node, ctx->user_bufs[i]);
567 io_buffer_unmap(ctx, &imu);
570 ctx->user_bufs[i] = ctx->dummy_ubuf;
574 ctx->user_bufs[i] = imu;
575 *io_get_tag_slot(ctx->buf_data, offset) = tag;
579 io_rsrc_node_switch(ctx, ctx->buf_data);
580 return done ? done : err;
583 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
584 struct io_uring_rsrc_update2 *up,
590 if (check_add_overflow(up->offset, nr_args, &tmp))
592 err = io_rsrc_node_switch_start(ctx);
597 case IORING_RSRC_FILE:
598 return __io_sqe_files_update(ctx, up, nr_args);
599 case IORING_RSRC_BUFFER:
600 return __io_sqe_buffers_update(ctx, up, nr_args);
605 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
608 struct io_uring_rsrc_update2 up;
612 memset(&up, 0, sizeof(up));
613 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
615 if (up.resv || up.resv2)
617 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
620 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
621 unsigned size, unsigned type)
623 struct io_uring_rsrc_update2 up;
625 if (size != sizeof(up))
627 if (copy_from_user(&up, arg, sizeof(up)))
629 if (!up.nr || up.resv || up.resv2)
631 return __io_register_rsrc_update(ctx, type, &up, up.nr);
634 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
635 unsigned int size, unsigned int type)
637 struct io_uring_rsrc_register rr;
639 /* keep it extendible */
640 if (size != sizeof(rr))
643 memset(&rr, 0, sizeof(rr));
644 if (copy_from_user(&rr, arg, size))
646 if (!rr.nr || rr.resv2)
648 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
652 case IORING_RSRC_FILE:
653 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
655 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
656 rr.nr, u64_to_user_ptr(rr.tags));
657 case IORING_RSRC_BUFFER:
658 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
660 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
661 rr.nr, u64_to_user_ptr(rr.tags));
666 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
668 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
670 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
672 if (sqe->rw_flags || sqe->splice_fd_in)
675 up->offset = READ_ONCE(sqe->off);
676 up->nr_args = READ_ONCE(sqe->len);
679 up->arg = READ_ONCE(sqe->addr);
683 static int io_files_update_with_index_alloc(struct io_kiocb *req,
684 unsigned int issue_flags)
686 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
687 __s32 __user *fds = u64_to_user_ptr(up->arg);
692 if (!req->ctx->file_data)
695 for (done = 0; done < up->nr_args; done++) {
696 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
706 ret = io_fixed_fd_install(req, issue_flags, file,
707 IORING_FILE_INDEX_ALLOC);
710 if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
711 __io_close_fixed(req->ctx, issue_flags, ret);
722 int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
724 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
725 struct io_ring_ctx *ctx = req->ctx;
726 struct io_uring_rsrc_update2 up2;
729 up2.offset = up->offset;
736 if (up->offset == IORING_FILE_INDEX_ALLOC) {
737 ret = io_files_update_with_index_alloc(req, issue_flags);
739 io_ring_submit_lock(ctx, issue_flags);
740 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
742 io_ring_submit_unlock(ctx, issue_flags);
747 io_req_set_res(req, ret, 0);
751 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
752 struct io_rsrc_node *node, void *rsrc)
754 u64 *tag_slot = io_get_tag_slot(data, idx);
755 struct io_rsrc_put *prsrc;
757 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
761 prsrc->tag = *tag_slot;
764 list_add(&prsrc->list, &node->rsrc_list);
768 void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
772 for (i = 0; i < ctx->nr_user_files; i++) {
773 struct file *file = io_file_from_index(&ctx->file_table, i);
775 /* skip scm accounted files, they'll be freed by ->ring_sock */
776 if (!file || io_file_need_scm(file))
778 io_file_bitmap_clear(&ctx->file_table, i);
782 #if defined(CONFIG_UNIX)
783 if (ctx->ring_sock) {
784 struct sock *sock = ctx->ring_sock->sk;
787 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
791 io_free_file_tables(&ctx->file_table);
792 io_file_table_set_alloc_range(ctx, 0, 0);
793 io_rsrc_data_free(ctx->file_data);
794 ctx->file_data = NULL;
795 ctx->nr_user_files = 0;
798 int io_sqe_files_unregister(struct io_ring_ctx *ctx)
800 unsigned nr = ctx->nr_user_files;
807 * Quiesce may unlock ->uring_lock, and while it's not held
808 * prevent new requests using the table.
810 ctx->nr_user_files = 0;
811 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
812 ctx->nr_user_files = nr;
814 __io_sqe_files_unregister(ctx);
819 * Ensure the UNIX gc is aware of our file set, so we are certain that
820 * the io_uring can be safely unregistered on process exit, even if we have
821 * loops in the file referencing. We account only files that can hold other
822 * files because otherwise they can't form a loop and so are not interesting
825 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
827 #if defined(CONFIG_UNIX)
828 struct sock *sk = ctx->ring_sock->sk;
829 struct sk_buff_head *head = &sk->sk_receive_queue;
830 struct scm_fp_list *fpl;
833 if (likely(!io_file_need_scm(file)))
837 * See if we can merge this file into an existing skb SCM_RIGHTS
838 * file set. If there's no room, fall back to allocating a new skb
841 spin_lock_irq(&head->lock);
842 skb = skb_peek(head);
843 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
844 __skb_unlink(skb, head);
847 spin_unlock_irq(&head->lock);
850 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
854 skb = alloc_skb(0, GFP_KERNEL);
860 fpl->user = get_uid(current_user());
861 fpl->max = SCM_MAX_FD;
864 UNIXCB(skb).fp = fpl;
866 skb->scm_io_uring = 1;
867 skb->destructor = unix_destruct_scm;
868 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
871 fpl = UNIXCB(skb).fp;
872 fpl->fp[fpl->count++] = get_file(file);
873 unix_inflight(fpl->user, file);
874 skb_queue_head(head, skb);
880 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
882 struct file *file = prsrc->file;
883 #if defined(CONFIG_UNIX)
884 struct sock *sock = ctx->ring_sock->sk;
885 struct sk_buff_head list, *head = &sock->sk_receive_queue;
889 if (!io_file_need_scm(file)) {
894 __skb_queue_head_init(&list);
897 * Find the skb that holds this file in its SCM_RIGHTS. When found,
898 * remove this entry and rearrange the file array.
900 skb = skb_dequeue(head);
902 struct scm_fp_list *fp;
905 for (i = 0; i < fp->count; i++) {
908 if (fp->fp[i] != file)
911 unix_notinflight(fp->user, fp->fp[i]);
912 left = fp->count - 1 - i;
914 memmove(&fp->fp[i], &fp->fp[i + 1],
915 left * sizeof(struct file *));
922 __skb_queue_tail(&list, skb);
932 __skb_queue_tail(&list, skb);
934 skb = skb_dequeue(head);
937 if (skb_peek(&list)) {
938 spin_lock_irq(&head->lock);
939 while ((skb = __skb_dequeue(&list)) != NULL)
940 __skb_queue_tail(head, skb);
941 spin_unlock_irq(&head->lock);
948 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
949 unsigned nr_args, u64 __user *tags)
951 __s32 __user *fds = (__s32 __user *) arg;
960 if (nr_args > IORING_MAX_FIXED_FILES)
962 if (nr_args > rlimit(RLIMIT_NOFILE))
964 ret = io_rsrc_node_switch_start(ctx);
967 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
972 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
973 io_rsrc_data_free(ctx->file_data);
974 ctx->file_data = NULL;
978 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
979 struct io_fixed_file *file_slot;
981 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
985 /* allow sparse sets */
986 if (!fds || fd == -1) {
988 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
999 * Don't allow io_uring instances to be registered. If UNIX
1000 * isn't enabled, then this causes a reference cycle and this
1001 * instance can never get freed. If UNIX is enabled we'll
1002 * handle it just fine, but there's still no point in allowing
1003 * a ring fd as it doesn't support regular read/write anyway.
1005 if (io_is_uring_fops(file)) {
1009 ret = io_scm_file_account(ctx, file);
1014 file_slot = io_fixed_file_slot(&ctx->file_table, i);
1015 io_fixed_file_set(file_slot, file);
1016 io_file_bitmap_set(&ctx->file_table, i);
1019 /* default it to the whole table */
1020 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
1021 io_rsrc_node_switch(ctx, NULL);
1024 __io_sqe_files_unregister(ctx);
1028 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
1030 io_buffer_unmap(ctx, &prsrc->buf);
1034 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
1038 for (i = 0; i < ctx->nr_user_bufs; i++)
1039 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
1040 kfree(ctx->user_bufs);
1041 io_rsrc_data_free(ctx->buf_data);
1042 ctx->user_bufs = NULL;
1043 ctx->buf_data = NULL;
1044 ctx->nr_user_bufs = 0;
1047 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
1049 unsigned nr = ctx->nr_user_bufs;
1056 * Quiesce may unlock ->uring_lock, and while it's not held
1057 * prevent new requests using the table.
1059 ctx->nr_user_bufs = 0;
1060 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
1061 ctx->nr_user_bufs = nr;
1063 __io_sqe_buffers_unregister(ctx);
1068 * Not super efficient, but this is just a registration time. And we do cache
1069 * the last compound head, so generally we'll only do a full search if we don't
1072 * We check if the given compound head page has already been accounted, to
1073 * avoid double accounting it. This allows us to account the full size of the
1074 * page, not just the constituent pages of a huge page.
1076 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
1077 int nr_pages, struct page *hpage)
1081 /* check current page array */
1082 for (i = 0; i < nr_pages; i++) {
1083 if (!PageCompound(pages[i]))
1085 if (compound_head(pages[i]) == hpage)
1089 /* check previously registered pages */
1090 for (i = 0; i < ctx->nr_user_bufs; i++) {
1091 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
1093 for (j = 0; j < imu->nr_bvecs; j++) {
1094 if (!PageCompound(imu->bvec[j].bv_page))
1096 if (compound_head(imu->bvec[j].bv_page) == hpage)
1104 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
1105 int nr_pages, struct io_mapped_ubuf *imu,
1106 struct page **last_hpage)
1110 imu->acct_pages = 0;
1111 for (i = 0; i < nr_pages; i++) {
1112 if (!PageCompound(pages[i])) {
1117 hpage = compound_head(pages[i]);
1118 if (hpage == *last_hpage)
1120 *last_hpage = hpage;
1121 if (headpage_already_acct(ctx, pages, i, hpage))
1123 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1127 if (!imu->acct_pages)
1130 ret = io_account_mem(ctx, imu->acct_pages);
1132 imu->acct_pages = 0;
1136 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1138 unsigned long start, end, nr_pages;
1139 struct vm_area_struct **vmas = NULL;
1140 struct page **pages = NULL;
1141 int i, pret, ret = -ENOMEM;
1143 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1144 start = ubuf >> PAGE_SHIFT;
1145 nr_pages = end - start;
1147 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1151 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
1157 mmap_read_lock(current->mm);
1158 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
1160 if (pret == nr_pages) {
1161 struct file *file = vmas[0]->vm_file;
1163 /* don't support file backed memory */
1164 for (i = 0; i < nr_pages; i++) {
1165 if (vmas[i]->vm_file != file) {
1171 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) {
1178 ret = pret < 0 ? pret : -EFAULT;
1180 mmap_read_unlock(current->mm);
1183 * if we did partial map, or found file backed vmas,
1184 * release any pages we did get
1187 unpin_user_pages(pages, pret);
1195 pages = ERR_PTR(ret);
1200 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1201 struct io_mapped_ubuf **pimu,
1202 struct page **last_hpage)
1204 struct io_mapped_ubuf *imu = NULL;
1205 struct page **pages = NULL;
1208 int ret, nr_pages, i;
1209 struct folio *folio = NULL;
1211 *pimu = ctx->dummy_ubuf;
1216 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1218 if (IS_ERR(pages)) {
1219 ret = PTR_ERR(pages);
1224 /* If it's a huge page, try to coalesce them into a single bvec entry */
1226 folio = page_folio(pages[0]);
1227 for (i = 1; i < nr_pages; i++) {
1228 if (page_folio(pages[i]) != folio) {
1235 * The pages are bound to the folio, it doesn't
1236 * actually unpin them but drops all but one reference,
1237 * which is usually put down by io_buffer_unmap().
1238 * Note, needs a better helper.
1240 unpin_user_pages(&pages[1], nr_pages - 1);
1245 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1249 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1251 unpin_user_pages(pages, nr_pages);
1255 off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1256 size = iov->iov_len;
1257 /* store original address for later verification */
1258 imu->ubuf = (unsigned long) iov->iov_base;
1259 imu->ubuf_end = imu->ubuf + iov->iov_len;
1260 imu->nr_bvecs = nr_pages;
1265 bvec_set_page(&imu->bvec[0], pages[0], size, off);
1268 for (i = 0; i < nr_pages; i++) {
1271 vec_len = min_t(size_t, size, PAGE_SIZE - off);
1272 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off);
1283 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1285 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1286 return ctx->user_bufs ? 0 : -ENOMEM;
1289 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1290 unsigned int nr_args, u64 __user *tags)
1292 struct page *last_hpage = NULL;
1293 struct io_rsrc_data *data;
1297 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1301 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1303 ret = io_rsrc_node_switch_start(ctx);
1306 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
1309 ret = io_buffers_map_alloc(ctx, nr_args);
1311 io_rsrc_data_free(data);
1315 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1317 ret = io_copy_iov(ctx, &iov, arg, i);
1320 ret = io_buffer_validate(&iov);
1324 memset(&iov, 0, sizeof(iov));
1327 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1332 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1338 WARN_ON_ONCE(ctx->buf_data);
1340 ctx->buf_data = data;
1342 __io_sqe_buffers_unregister(ctx);
1344 io_rsrc_node_switch(ctx, NULL);
1348 int io_import_fixed(int ddir, struct iov_iter *iter,
1349 struct io_mapped_ubuf *imu,
1350 u64 buf_addr, size_t len)
1355 if (WARN_ON_ONCE(!imu))
1357 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1359 /* not inside the mapped region */
1360 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1364 * Might not be a start of buffer, set size appropriately
1365 * and advance us to the beginning.
1367 offset = buf_addr - imu->ubuf;
1368 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1372 * Don't use iov_iter_advance() here, as it's really slow for
1373 * using the latter parts of a big fixed buffer - it iterates
1374 * over each segment manually. We can cheat a bit here, because
1377 * 1) it's a BVEC iter, we set it up
1378 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1379 * first and last bvec
1381 * So just find our index, and adjust the iterator afterwards.
1382 * If the offset is within the first bvec (or the whole first
1383 * bvec, just use iov_iter_advance(). This makes it easier
1384 * since we can just skip the first segment, which may not
1385 * be PAGE_SIZE aligned.
1387 const struct bio_vec *bvec = imu->bvec;
1389 if (offset <= bvec->bv_len) {
1391 * Note, huge pages buffers consists of one large
1392 * bvec entry and should always go this way. The other
1393 * branch doesn't expect non PAGE_SIZE'd chunks.
1396 iter->nr_segs = bvec->bv_len;
1397 iter->count -= offset;
1398 iter->iov_offset = offset;
1400 unsigned long seg_skip;
1402 /* skip first vec */
1403 offset -= bvec->bv_len;
1404 seg_skip = 1 + (offset >> PAGE_SHIFT);
1406 iter->bvec = bvec + seg_skip;
1407 iter->nr_segs -= seg_skip;
1408 iter->count -= bvec->bv_len + offset;
1409 iter->iov_offset = offset & ~PAGE_MASK;