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);
31 #define IORING_MAX_FIXED_FILES (1U << 20)
32 #define IORING_MAX_REG_BUFFERS (1U << 14)
34 static inline bool io_put_rsrc_data_ref(struct io_rsrc_data *rsrc_data)
36 return !--rsrc_data->refs;
39 int __io_account_mem(struct user_struct *user, unsigned long nr_pages)
41 unsigned long page_limit, cur_pages, new_pages;
46 /* Don't allow more pages than we can safely lock */
47 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
49 cur_pages = atomic_long_read(&user->locked_vm);
51 new_pages = cur_pages + nr_pages;
52 if (new_pages > page_limit)
54 } while (!atomic_long_try_cmpxchg(&user->locked_vm,
55 &cur_pages, new_pages));
59 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
62 __io_unaccount_mem(ctx->user, nr_pages);
65 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
68 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
73 ret = __io_account_mem(ctx->user, nr_pages);
79 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
84 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
85 void __user *arg, unsigned index)
87 struct iovec __user *src;
91 struct compat_iovec __user *ciovs;
92 struct compat_iovec ciov;
94 ciovs = (struct compat_iovec __user *) arg;
95 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
98 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
99 dst->iov_len = ciov.iov_len;
103 src = (struct iovec __user *) arg;
104 if (copy_from_user(dst, &src[index], sizeof(*dst)))
109 static int io_buffer_validate(struct iovec *iov)
111 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
114 * Don't impose further limits on the size and buffer
115 * constraints here, we'll -EINVAL later when IO is
116 * submitted if they are wrong.
119 return iov->iov_len ? -EFAULT : 0;
123 /* arbitrary limit, but we need something */
124 if (iov->iov_len > SZ_1G)
127 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
133 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
135 struct io_mapped_ubuf *imu = *slot;
138 if (imu != ctx->dummy_ubuf) {
139 for (i = 0; i < imu->nr_bvecs; i++)
140 unpin_user_page(imu->bvec[i].bv_page);
142 io_unaccount_mem(ctx, imu->acct_pages);
148 static void io_rsrc_put_work_one(struct io_rsrc_data *rsrc_data,
149 struct io_rsrc_put *prsrc)
151 struct io_ring_ctx *ctx = rsrc_data->ctx;
154 io_post_aux_cqe(ctx, prsrc->tag, 0, 0);
155 rsrc_data->do_put(ctx, prsrc);
158 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
160 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
161 struct io_rsrc_put *prsrc, *tmp;
163 if (ref_node->inline_items)
164 io_rsrc_put_work_one(rsrc_data, &ref_node->item);
166 list_for_each_entry_safe(prsrc, tmp, &ref_node->item_list, list) {
167 list_del(&prsrc->list);
168 io_rsrc_put_work_one(rsrc_data, prsrc);
172 io_rsrc_node_destroy(rsrc_data->ctx, ref_node);
173 if (io_put_rsrc_data_ref(rsrc_data))
174 complete(&rsrc_data->done);
177 void io_wait_rsrc_data(struct io_rsrc_data *data)
179 if (data && !io_put_rsrc_data_ref(data))
180 wait_for_completion(&data->done);
183 void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *node)
185 if (!io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache))
189 void io_rsrc_node_ref_zero(struct io_rsrc_node *node)
190 __must_hold(&node->rsrc_data->ctx->uring_lock)
192 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
194 while (!list_empty(&ctx->rsrc_ref_list)) {
195 node = list_first_entry(&ctx->rsrc_ref_list,
196 struct io_rsrc_node, node);
197 /* recycle ref nodes in order */
200 list_del(&node->node);
201 __io_rsrc_put_work(node);
205 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
207 struct io_rsrc_node *ref_node;
208 struct io_cache_entry *entry;
210 entry = io_alloc_cache_get(&ctx->rsrc_node_cache);
212 ref_node = container_of(entry, struct io_rsrc_node, cache);
214 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
219 ref_node->rsrc_data = NULL;
221 INIT_LIST_HEAD(&ref_node->node);
222 INIT_LIST_HEAD(&ref_node->item_list);
223 ref_node->inline_items = 0;
227 void io_rsrc_node_switch(struct io_ring_ctx *ctx,
228 struct io_rsrc_data *data_to_kill)
229 __must_hold(&ctx->uring_lock)
231 struct io_rsrc_node *node = ctx->rsrc_node;
232 struct io_rsrc_node *backup = io_rsrc_node_alloc(ctx);
234 if (WARN_ON_ONCE(!backup))
237 data_to_kill->refs++;
238 node->rsrc_data = data_to_kill;
239 list_add_tail(&node->node, &ctx->rsrc_ref_list);
241 io_put_rsrc_node(ctx, node);
242 ctx->rsrc_node = backup;
245 int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
247 if (io_alloc_cache_empty(&ctx->rsrc_node_cache)) {
248 struct io_rsrc_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
252 io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache);
257 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
258 struct io_ring_ctx *ctx)
262 /* As we may drop ->uring_lock, other task may have started quiesce */
265 ret = io_rsrc_node_switch_start(ctx);
268 io_rsrc_node_switch(ctx, data);
270 /* kill initial ref */
271 if (io_put_rsrc_data_ref(data))
274 data->quiesce = true;
275 mutex_unlock(&ctx->uring_lock);
277 ret = io_run_task_work_sig(ctx);
279 mutex_lock(&ctx->uring_lock);
283 /* restore the master reference */
288 ret = wait_for_completion_interruptible(&data->done);
290 mutex_lock(&ctx->uring_lock);
294 * it has been revived by another thread while
297 mutex_unlock(&ctx->uring_lock);
300 data->quiesce = false;
305 static void io_free_page_table(void **table, size_t size)
307 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
309 for (i = 0; i < nr_tables; i++)
314 static void io_rsrc_data_free(struct io_rsrc_data *data)
316 size_t size = data->nr * sizeof(data->tags[0][0]);
319 io_free_page_table((void **)data->tags, size);
323 static __cold void **io_alloc_page_table(size_t size)
325 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
326 size_t init_size = size;
329 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
333 for (i = 0; i < nr_tables; i++) {
334 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
336 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
338 io_free_page_table(table, init_size);
346 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx,
347 rsrc_put_fn *do_put, u64 __user *utags,
348 unsigned nr, struct io_rsrc_data **pdata)
350 struct io_rsrc_data *data;
354 data = kzalloc(sizeof(*data), GFP_KERNEL);
357 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
365 data->do_put = do_put;
369 for (i = 0; i < nr; i++) {
370 u64 *tag_slot = io_get_tag_slot(data, i);
372 if (copy_from_user(tag_slot, &utags[i],
377 init_completion(&data->done);
381 io_rsrc_data_free(data);
385 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
386 struct io_uring_rsrc_update2 *up,
389 u64 __user *tags = u64_to_user_ptr(up->tags);
390 __s32 __user *fds = u64_to_user_ptr(up->data);
391 struct io_rsrc_data *data = ctx->file_data;
392 struct io_fixed_file *file_slot;
396 bool needs_switch = false;
400 if (up->offset + nr_args > ctx->nr_user_files)
403 for (done = 0; done < nr_args; done++) {
406 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
407 copy_from_user(&fd, &fds[done], sizeof(fd))) {
411 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
415 if (fd == IORING_REGISTER_FILES_SKIP)
418 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
419 file_slot = io_fixed_file_slot(&ctx->file_table, i);
421 if (file_slot->file_ptr) {
422 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
423 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
426 file_slot->file_ptr = 0;
427 io_file_bitmap_clear(&ctx->file_table, i);
437 * Don't allow io_uring instances to be registered. If
438 * UNIX isn't enabled, then this causes a reference
439 * cycle and this instance can never get freed. If UNIX
440 * is enabled we'll handle it just fine, but there's
441 * still no point in allowing a ring fd as it doesn't
442 * support regular read/write anyway.
444 if (io_is_uring_fops(file)) {
449 err = io_scm_file_account(ctx, file);
454 *io_get_tag_slot(data, i) = tag;
455 io_fixed_file_set(file_slot, file);
456 io_file_bitmap_set(&ctx->file_table, i);
461 io_rsrc_node_switch(ctx, data);
462 return done ? done : err;
465 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
466 struct io_uring_rsrc_update2 *up,
467 unsigned int nr_args)
469 u64 __user *tags = u64_to_user_ptr(up->tags);
470 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
471 struct page *last_hpage = NULL;
472 bool needs_switch = false;
478 if (up->offset + nr_args > ctx->nr_user_bufs)
481 for (done = 0; done < nr_args; done++) {
482 struct io_mapped_ubuf *imu;
483 int offset = up->offset + done;
486 err = io_copy_iov(ctx, &iov, iovs, done);
489 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
493 err = io_buffer_validate(&iov);
496 if (!iov.iov_base && tag) {
500 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
504 i = array_index_nospec(offset, ctx->nr_user_bufs);
505 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
506 err = io_queue_rsrc_removal(ctx->buf_data, i,
507 ctx->rsrc_node, ctx->user_bufs[i]);
509 io_buffer_unmap(ctx, &imu);
512 ctx->user_bufs[i] = ctx->dummy_ubuf;
516 ctx->user_bufs[i] = imu;
517 *io_get_tag_slot(ctx->buf_data, i) = tag;
521 io_rsrc_node_switch(ctx, ctx->buf_data);
522 return done ? done : err;
525 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
526 struct io_uring_rsrc_update2 *up,
532 lockdep_assert_held(&ctx->uring_lock);
534 if (check_add_overflow(up->offset, nr_args, &tmp))
536 err = io_rsrc_node_switch_start(ctx);
541 case IORING_RSRC_FILE:
542 return __io_sqe_files_update(ctx, up, nr_args);
543 case IORING_RSRC_BUFFER:
544 return __io_sqe_buffers_update(ctx, up, nr_args);
549 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
552 struct io_uring_rsrc_update2 up;
556 memset(&up, 0, sizeof(up));
557 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
559 if (up.resv || up.resv2)
561 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
564 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
565 unsigned size, unsigned type)
567 struct io_uring_rsrc_update2 up;
569 if (size != sizeof(up))
571 if (copy_from_user(&up, arg, sizeof(up)))
573 if (!up.nr || up.resv || up.resv2)
575 return __io_register_rsrc_update(ctx, type, &up, up.nr);
578 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
579 unsigned int size, unsigned int type)
581 struct io_uring_rsrc_register rr;
583 /* keep it extendible */
584 if (size != sizeof(rr))
587 memset(&rr, 0, sizeof(rr));
588 if (copy_from_user(&rr, arg, size))
590 if (!rr.nr || rr.resv2)
592 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
596 case IORING_RSRC_FILE:
597 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
599 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
600 rr.nr, u64_to_user_ptr(rr.tags));
601 case IORING_RSRC_BUFFER:
602 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
604 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
605 rr.nr, u64_to_user_ptr(rr.tags));
610 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
612 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
614 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
616 if (sqe->rw_flags || sqe->splice_fd_in)
619 up->offset = READ_ONCE(sqe->off);
620 up->nr_args = READ_ONCE(sqe->len);
623 up->arg = READ_ONCE(sqe->addr);
627 static int io_files_update_with_index_alloc(struct io_kiocb *req,
628 unsigned int issue_flags)
630 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
631 __s32 __user *fds = u64_to_user_ptr(up->arg);
636 if (!req->ctx->file_data)
639 for (done = 0; done < up->nr_args; done++) {
640 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
650 ret = io_fixed_fd_install(req, issue_flags, file,
651 IORING_FILE_INDEX_ALLOC);
654 if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
655 __io_close_fixed(req->ctx, issue_flags, ret);
666 int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
668 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
669 struct io_ring_ctx *ctx = req->ctx;
670 struct io_uring_rsrc_update2 up2;
673 up2.offset = up->offset;
680 if (up->offset == IORING_FILE_INDEX_ALLOC) {
681 ret = io_files_update_with_index_alloc(req, issue_flags);
683 io_ring_submit_lock(ctx, issue_flags);
684 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
686 io_ring_submit_unlock(ctx, issue_flags);
691 io_req_set_res(req, ret, 0);
695 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
696 struct io_rsrc_node *node, void *rsrc)
698 u64 *tag_slot = io_get_tag_slot(data, idx);
699 struct io_rsrc_put *prsrc;
700 bool inline_item = true;
702 if (!node->inline_items) {
704 node->inline_items++;
706 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
712 prsrc->tag = *tag_slot;
716 list_add(&prsrc->list, &node->item_list);
720 void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
724 for (i = 0; i < ctx->nr_user_files; i++) {
725 struct file *file = io_file_from_index(&ctx->file_table, i);
727 /* skip scm accounted files, they'll be freed by ->ring_sock */
728 if (!file || io_file_need_scm(file))
730 io_file_bitmap_clear(&ctx->file_table, i);
734 #if defined(CONFIG_UNIX)
735 if (ctx->ring_sock) {
736 struct sock *sock = ctx->ring_sock->sk;
739 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
743 io_free_file_tables(&ctx->file_table);
744 io_file_table_set_alloc_range(ctx, 0, 0);
745 io_rsrc_data_free(ctx->file_data);
746 ctx->file_data = NULL;
747 ctx->nr_user_files = 0;
750 int io_sqe_files_unregister(struct io_ring_ctx *ctx)
752 unsigned nr = ctx->nr_user_files;
759 * Quiesce may unlock ->uring_lock, and while it's not held
760 * prevent new requests using the table.
762 ctx->nr_user_files = 0;
763 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
764 ctx->nr_user_files = nr;
766 __io_sqe_files_unregister(ctx);
771 * Ensure the UNIX gc is aware of our file set, so we are certain that
772 * the io_uring can be safely unregistered on process exit, even if we have
773 * loops in the file referencing. We account only files that can hold other
774 * files because otherwise they can't form a loop and so are not interesting
777 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
779 #if defined(CONFIG_UNIX)
780 struct sock *sk = ctx->ring_sock->sk;
781 struct sk_buff_head *head = &sk->sk_receive_queue;
782 struct scm_fp_list *fpl;
785 if (likely(!io_file_need_scm(file)))
789 * See if we can merge this file into an existing skb SCM_RIGHTS
790 * file set. If there's no room, fall back to allocating a new skb
793 spin_lock_irq(&head->lock);
794 skb = skb_peek(head);
795 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
796 __skb_unlink(skb, head);
799 spin_unlock_irq(&head->lock);
802 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
806 skb = alloc_skb(0, GFP_KERNEL);
812 fpl->user = get_uid(current_user());
813 fpl->max = SCM_MAX_FD;
816 UNIXCB(skb).fp = fpl;
818 skb->scm_io_uring = 1;
819 skb->destructor = unix_destruct_scm;
820 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
823 fpl = UNIXCB(skb).fp;
824 fpl->fp[fpl->count++] = get_file(file);
825 unix_inflight(fpl->user, file);
826 skb_queue_head(head, skb);
832 static __cold void io_rsrc_file_scm_put(struct io_ring_ctx *ctx, struct file *file)
834 #if defined(CONFIG_UNIX)
835 struct sock *sock = ctx->ring_sock->sk;
836 struct sk_buff_head list, *head = &sock->sk_receive_queue;
840 __skb_queue_head_init(&list);
843 * Find the skb that holds this file in its SCM_RIGHTS. When found,
844 * remove this entry and rearrange the file array.
846 skb = skb_dequeue(head);
848 struct scm_fp_list *fp;
851 for (i = 0; i < fp->count; i++) {
854 if (fp->fp[i] != file)
857 unix_notinflight(fp->user, fp->fp[i]);
858 left = fp->count - 1 - i;
860 memmove(&fp->fp[i], &fp->fp[i + 1],
861 left * sizeof(struct file *));
868 __skb_queue_tail(&list, skb);
878 __skb_queue_tail(&list, skb);
880 skb = skb_dequeue(head);
883 if (skb_peek(&list)) {
884 spin_lock_irq(&head->lock);
885 while ((skb = __skb_dequeue(&list)) != NULL)
886 __skb_queue_tail(head, skb);
887 spin_unlock_irq(&head->lock);
892 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
894 struct file *file = prsrc->file;
896 if (likely(!io_file_need_scm(file)))
899 io_rsrc_file_scm_put(ctx, file);
902 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
903 unsigned nr_args, u64 __user *tags)
905 __s32 __user *fds = (__s32 __user *) arg;
914 if (nr_args > IORING_MAX_FIXED_FILES)
916 if (nr_args > rlimit(RLIMIT_NOFILE))
918 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
923 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
924 io_rsrc_data_free(ctx->file_data);
925 ctx->file_data = NULL;
929 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
930 struct io_fixed_file *file_slot;
932 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
936 /* allow sparse sets */
937 if (!fds || fd == -1) {
939 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
950 * Don't allow io_uring instances to be registered. If UNIX
951 * isn't enabled, then this causes a reference cycle and this
952 * instance can never get freed. If UNIX is enabled we'll
953 * handle it just fine, but there's still no point in allowing
954 * a ring fd as it doesn't support regular read/write anyway.
956 if (io_is_uring_fops(file)) {
960 ret = io_scm_file_account(ctx, file);
965 file_slot = io_fixed_file_slot(&ctx->file_table, i);
966 io_fixed_file_set(file_slot, file);
967 io_file_bitmap_set(&ctx->file_table, i);
970 /* default it to the whole table */
971 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
974 __io_sqe_files_unregister(ctx);
978 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
980 io_buffer_unmap(ctx, &prsrc->buf);
984 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
988 for (i = 0; i < ctx->nr_user_bufs; i++)
989 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
990 kfree(ctx->user_bufs);
991 io_rsrc_data_free(ctx->buf_data);
992 ctx->user_bufs = NULL;
993 ctx->buf_data = NULL;
994 ctx->nr_user_bufs = 0;
997 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
999 unsigned nr = ctx->nr_user_bufs;
1006 * Quiesce may unlock ->uring_lock, and while it's not held
1007 * prevent new requests using the table.
1009 ctx->nr_user_bufs = 0;
1010 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
1011 ctx->nr_user_bufs = nr;
1013 __io_sqe_buffers_unregister(ctx);
1018 * Not super efficient, but this is just a registration time. And we do cache
1019 * the last compound head, so generally we'll only do a full search if we don't
1022 * We check if the given compound head page has already been accounted, to
1023 * avoid double accounting it. This allows us to account the full size of the
1024 * page, not just the constituent pages of a huge page.
1026 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
1027 int nr_pages, struct page *hpage)
1031 /* check current page array */
1032 for (i = 0; i < nr_pages; i++) {
1033 if (!PageCompound(pages[i]))
1035 if (compound_head(pages[i]) == hpage)
1039 /* check previously registered pages */
1040 for (i = 0; i < ctx->nr_user_bufs; i++) {
1041 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
1043 for (j = 0; j < imu->nr_bvecs; j++) {
1044 if (!PageCompound(imu->bvec[j].bv_page))
1046 if (compound_head(imu->bvec[j].bv_page) == hpage)
1054 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
1055 int nr_pages, struct io_mapped_ubuf *imu,
1056 struct page **last_hpage)
1060 imu->acct_pages = 0;
1061 for (i = 0; i < nr_pages; i++) {
1062 if (!PageCompound(pages[i])) {
1067 hpage = compound_head(pages[i]);
1068 if (hpage == *last_hpage)
1070 *last_hpage = hpage;
1071 if (headpage_already_acct(ctx, pages, i, hpage))
1073 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1077 if (!imu->acct_pages)
1080 ret = io_account_mem(ctx, imu->acct_pages);
1082 imu->acct_pages = 0;
1086 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1088 unsigned long start, end, nr_pages;
1089 struct vm_area_struct **vmas = NULL;
1090 struct page **pages = NULL;
1091 int i, pret, ret = -ENOMEM;
1093 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1094 start = ubuf >> PAGE_SHIFT;
1095 nr_pages = end - start;
1097 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1101 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
1107 mmap_read_lock(current->mm);
1108 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
1110 if (pret == nr_pages) {
1111 struct file *file = vmas[0]->vm_file;
1113 /* don't support file backed memory */
1114 for (i = 0; i < nr_pages; i++) {
1115 if (vmas[i]->vm_file != file) {
1121 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) {
1128 ret = pret < 0 ? pret : -EFAULT;
1130 mmap_read_unlock(current->mm);
1133 * if we did partial map, or found file backed vmas,
1134 * release any pages we did get
1137 unpin_user_pages(pages, pret);
1145 pages = ERR_PTR(ret);
1150 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1151 struct io_mapped_ubuf **pimu,
1152 struct page **last_hpage)
1154 struct io_mapped_ubuf *imu = NULL;
1155 struct page **pages = NULL;
1158 int ret, nr_pages, i;
1159 struct folio *folio = NULL;
1161 *pimu = ctx->dummy_ubuf;
1166 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1168 if (IS_ERR(pages)) {
1169 ret = PTR_ERR(pages);
1174 /* If it's a huge page, try to coalesce them into a single bvec entry */
1176 folio = page_folio(pages[0]);
1177 for (i = 1; i < nr_pages; i++) {
1178 if (page_folio(pages[i]) != folio) {
1185 * The pages are bound to the folio, it doesn't
1186 * actually unpin them but drops all but one reference,
1187 * which is usually put down by io_buffer_unmap().
1188 * Note, needs a better helper.
1190 unpin_user_pages(&pages[1], nr_pages - 1);
1195 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1199 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1201 unpin_user_pages(pages, nr_pages);
1205 off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1206 size = iov->iov_len;
1207 /* store original address for later verification */
1208 imu->ubuf = (unsigned long) iov->iov_base;
1209 imu->ubuf_end = imu->ubuf + iov->iov_len;
1210 imu->nr_bvecs = nr_pages;
1215 bvec_set_page(&imu->bvec[0], pages[0], size, off);
1218 for (i = 0; i < nr_pages; i++) {
1221 vec_len = min_t(size_t, size, PAGE_SIZE - off);
1222 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off);
1233 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1235 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1236 return ctx->user_bufs ? 0 : -ENOMEM;
1239 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1240 unsigned int nr_args, u64 __user *tags)
1242 struct page *last_hpage = NULL;
1243 struct io_rsrc_data *data;
1247 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1251 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1253 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
1256 ret = io_buffers_map_alloc(ctx, nr_args);
1258 io_rsrc_data_free(data);
1262 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1264 ret = io_copy_iov(ctx, &iov, arg, i);
1267 ret = io_buffer_validate(&iov);
1271 memset(&iov, 0, sizeof(iov));
1274 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1279 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1285 WARN_ON_ONCE(ctx->buf_data);
1287 ctx->buf_data = data;
1289 __io_sqe_buffers_unregister(ctx);
1293 int io_import_fixed(int ddir, struct iov_iter *iter,
1294 struct io_mapped_ubuf *imu,
1295 u64 buf_addr, size_t len)
1300 if (WARN_ON_ONCE(!imu))
1302 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1304 /* not inside the mapped region */
1305 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1309 * Might not be a start of buffer, set size appropriately
1310 * and advance us to the beginning.
1312 offset = buf_addr - imu->ubuf;
1313 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1317 * Don't use iov_iter_advance() here, as it's really slow for
1318 * using the latter parts of a big fixed buffer - it iterates
1319 * over each segment manually. We can cheat a bit here, because
1322 * 1) it's a BVEC iter, we set it up
1323 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1324 * first and last bvec
1326 * So just find our index, and adjust the iterator afterwards.
1327 * If the offset is within the first bvec (or the whole first
1328 * bvec, just use iov_iter_advance(). This makes it easier
1329 * since we can just skip the first segment, which may not
1330 * be PAGE_SIZE aligned.
1332 const struct bio_vec *bvec = imu->bvec;
1334 if (offset <= bvec->bv_len) {
1336 * Note, huge pages buffers consists of one large
1337 * bvec entry and should always go this way. The other
1338 * branch doesn't expect non PAGE_SIZE'd chunks.
1341 iter->nr_segs = bvec->bv_len;
1342 iter->count -= offset;
1343 iter->iov_offset = offset;
1345 unsigned long seg_skip;
1347 /* skip first vec */
1348 offset -= bvec->bv_len;
1349 seg_skip = 1 + (offset >> PAGE_SHIFT);
1351 iter->bvec = bvec + seg_skip;
1352 iter->nr_segs -= seg_skip;
1353 iter->count -= bvec->bv_len + offset;
1354 iter->iov_offset = offset & ~PAGE_MASK;