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;
195 while (!list_empty(&ctx->rsrc_ref_list)) {
196 node = list_first_entry(&ctx->rsrc_ref_list,
197 struct io_rsrc_node, node);
198 /* recycle ref nodes in order */
202 list_del(&node->node);
203 __io_rsrc_put_work(node);
207 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
209 struct io_rsrc_node *ref_node;
210 struct io_cache_entry *entry;
212 entry = io_alloc_cache_get(&ctx->rsrc_node_cache);
214 ref_node = container_of(entry, struct io_rsrc_node, cache);
216 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
221 ref_node->rsrc_data = NULL;
223 INIT_LIST_HEAD(&ref_node->node);
224 INIT_LIST_HEAD(&ref_node->item_list);
225 ref_node->done = false;
226 ref_node->inline_items = 0;
230 void io_rsrc_node_switch(struct io_ring_ctx *ctx,
231 struct io_rsrc_data *data_to_kill)
232 __must_hold(&ctx->uring_lock)
234 struct io_rsrc_node *node = ctx->rsrc_node;
235 struct io_rsrc_node *backup = io_rsrc_node_alloc(ctx);
237 if (WARN_ON_ONCE(!backup))
240 data_to_kill->refs++;
241 node->rsrc_data = data_to_kill;
242 list_add_tail(&node->node, &ctx->rsrc_ref_list);
244 io_put_rsrc_node(ctx, node);
245 ctx->rsrc_node = backup;
248 int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
250 if (io_alloc_cache_empty(&ctx->rsrc_node_cache)) {
251 struct io_rsrc_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
255 io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache);
260 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
261 struct io_ring_ctx *ctx)
265 /* As we may drop ->uring_lock, other task may have started quiesce */
268 ret = io_rsrc_node_switch_start(ctx);
271 io_rsrc_node_switch(ctx, data);
273 /* kill initial ref */
274 if (io_put_rsrc_data_ref(data))
277 data->quiesce = true;
278 mutex_unlock(&ctx->uring_lock);
280 ret = io_run_task_work_sig(ctx);
282 mutex_lock(&ctx->uring_lock);
286 /* restore the master reference */
291 ret = wait_for_completion_interruptible(&data->done);
293 mutex_lock(&ctx->uring_lock);
297 * it has been revived by another thread while
300 mutex_unlock(&ctx->uring_lock);
303 data->quiesce = false;
308 static void io_free_page_table(void **table, size_t size)
310 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
312 for (i = 0; i < nr_tables; i++)
317 static void io_rsrc_data_free(struct io_rsrc_data *data)
319 size_t size = data->nr * sizeof(data->tags[0][0]);
322 io_free_page_table((void **)data->tags, size);
326 static __cold void **io_alloc_page_table(size_t size)
328 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
329 size_t init_size = size;
332 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
336 for (i = 0; i < nr_tables; i++) {
337 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
339 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
341 io_free_page_table(table, init_size);
349 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx,
350 rsrc_put_fn *do_put, u64 __user *utags,
351 unsigned nr, struct io_rsrc_data **pdata)
353 struct io_rsrc_data *data;
357 data = kzalloc(sizeof(*data), GFP_KERNEL);
360 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
368 data->do_put = do_put;
372 for (i = 0; i < nr; i++) {
373 u64 *tag_slot = io_get_tag_slot(data, i);
375 if (copy_from_user(tag_slot, &utags[i],
380 init_completion(&data->done);
384 io_rsrc_data_free(data);
388 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
389 struct io_uring_rsrc_update2 *up,
392 u64 __user *tags = u64_to_user_ptr(up->tags);
393 __s32 __user *fds = u64_to_user_ptr(up->data);
394 struct io_rsrc_data *data = ctx->file_data;
395 struct io_fixed_file *file_slot;
399 bool needs_switch = false;
403 if (up->offset + nr_args > ctx->nr_user_files)
406 for (done = 0; done < nr_args; done++) {
409 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
410 copy_from_user(&fd, &fds[done], sizeof(fd))) {
414 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
418 if (fd == IORING_REGISTER_FILES_SKIP)
421 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
422 file_slot = io_fixed_file_slot(&ctx->file_table, i);
424 if (file_slot->file_ptr) {
425 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
426 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
429 file_slot->file_ptr = 0;
430 io_file_bitmap_clear(&ctx->file_table, i);
440 * Don't allow io_uring instances to be registered. If
441 * UNIX isn't enabled, then this causes a reference
442 * cycle and this instance can never get freed. If UNIX
443 * is enabled we'll handle it just fine, but there's
444 * still no point in allowing a ring fd as it doesn't
445 * support regular read/write anyway.
447 if (io_is_uring_fops(file)) {
452 err = io_scm_file_account(ctx, file);
457 *io_get_tag_slot(data, i) = tag;
458 io_fixed_file_set(file_slot, file);
459 io_file_bitmap_set(&ctx->file_table, i);
464 io_rsrc_node_switch(ctx, data);
465 return done ? done : err;
468 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
469 struct io_uring_rsrc_update2 *up,
470 unsigned int nr_args)
472 u64 __user *tags = u64_to_user_ptr(up->tags);
473 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
474 struct page *last_hpage = NULL;
475 bool needs_switch = false;
481 if (up->offset + nr_args > ctx->nr_user_bufs)
484 for (done = 0; done < nr_args; done++) {
485 struct io_mapped_ubuf *imu;
486 int offset = up->offset + done;
489 err = io_copy_iov(ctx, &iov, iovs, done);
492 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
496 err = io_buffer_validate(&iov);
499 if (!iov.iov_base && tag) {
503 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
507 i = array_index_nospec(offset, ctx->nr_user_bufs);
508 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
509 err = io_queue_rsrc_removal(ctx->buf_data, i,
510 ctx->rsrc_node, ctx->user_bufs[i]);
512 io_buffer_unmap(ctx, &imu);
515 ctx->user_bufs[i] = ctx->dummy_ubuf;
519 ctx->user_bufs[i] = imu;
520 *io_get_tag_slot(ctx->buf_data, offset) = tag;
524 io_rsrc_node_switch(ctx, ctx->buf_data);
525 return done ? done : err;
528 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
529 struct io_uring_rsrc_update2 *up,
535 lockdep_assert_held(&ctx->uring_lock);
537 if (check_add_overflow(up->offset, nr_args, &tmp))
539 err = io_rsrc_node_switch_start(ctx);
544 case IORING_RSRC_FILE:
545 return __io_sqe_files_update(ctx, up, nr_args);
546 case IORING_RSRC_BUFFER:
547 return __io_sqe_buffers_update(ctx, up, nr_args);
552 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
555 struct io_uring_rsrc_update2 up;
559 memset(&up, 0, sizeof(up));
560 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
562 if (up.resv || up.resv2)
564 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
567 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
568 unsigned size, unsigned type)
570 struct io_uring_rsrc_update2 up;
572 if (size != sizeof(up))
574 if (copy_from_user(&up, arg, sizeof(up)))
576 if (!up.nr || up.resv || up.resv2)
578 return __io_register_rsrc_update(ctx, type, &up, up.nr);
581 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
582 unsigned int size, unsigned int type)
584 struct io_uring_rsrc_register rr;
586 /* keep it extendible */
587 if (size != sizeof(rr))
590 memset(&rr, 0, sizeof(rr));
591 if (copy_from_user(&rr, arg, size))
593 if (!rr.nr || rr.resv2)
595 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
599 case IORING_RSRC_FILE:
600 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
602 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
603 rr.nr, u64_to_user_ptr(rr.tags));
604 case IORING_RSRC_BUFFER:
605 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
607 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
608 rr.nr, u64_to_user_ptr(rr.tags));
613 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
615 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
617 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
619 if (sqe->rw_flags || sqe->splice_fd_in)
622 up->offset = READ_ONCE(sqe->off);
623 up->nr_args = READ_ONCE(sqe->len);
626 up->arg = READ_ONCE(sqe->addr);
630 static int io_files_update_with_index_alloc(struct io_kiocb *req,
631 unsigned int issue_flags)
633 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
634 __s32 __user *fds = u64_to_user_ptr(up->arg);
639 if (!req->ctx->file_data)
642 for (done = 0; done < up->nr_args; done++) {
643 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
653 ret = io_fixed_fd_install(req, issue_flags, file,
654 IORING_FILE_INDEX_ALLOC);
657 if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
658 __io_close_fixed(req->ctx, issue_flags, ret);
669 int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
671 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
672 struct io_ring_ctx *ctx = req->ctx;
673 struct io_uring_rsrc_update2 up2;
676 up2.offset = up->offset;
683 if (up->offset == IORING_FILE_INDEX_ALLOC) {
684 ret = io_files_update_with_index_alloc(req, issue_flags);
686 io_ring_submit_lock(ctx, issue_flags);
687 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
689 io_ring_submit_unlock(ctx, issue_flags);
694 io_req_set_res(req, ret, 0);
698 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
699 struct io_rsrc_node *node, void *rsrc)
701 u64 *tag_slot = io_get_tag_slot(data, idx);
702 struct io_rsrc_put *prsrc;
703 bool inline_item = true;
705 if (!node->inline_items) {
707 node->inline_items++;
709 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
715 prsrc->tag = *tag_slot;
719 list_add(&prsrc->list, &node->item_list);
723 void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
727 for (i = 0; i < ctx->nr_user_files; i++) {
728 struct file *file = io_file_from_index(&ctx->file_table, i);
730 /* skip scm accounted files, they'll be freed by ->ring_sock */
731 if (!file || io_file_need_scm(file))
733 io_file_bitmap_clear(&ctx->file_table, i);
737 #if defined(CONFIG_UNIX)
738 if (ctx->ring_sock) {
739 struct sock *sock = ctx->ring_sock->sk;
742 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
746 io_free_file_tables(&ctx->file_table);
747 io_file_table_set_alloc_range(ctx, 0, 0);
748 io_rsrc_data_free(ctx->file_data);
749 ctx->file_data = NULL;
750 ctx->nr_user_files = 0;
753 int io_sqe_files_unregister(struct io_ring_ctx *ctx)
755 unsigned nr = ctx->nr_user_files;
762 * Quiesce may unlock ->uring_lock, and while it's not held
763 * prevent new requests using the table.
765 ctx->nr_user_files = 0;
766 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
767 ctx->nr_user_files = nr;
769 __io_sqe_files_unregister(ctx);
774 * Ensure the UNIX gc is aware of our file set, so we are certain that
775 * the io_uring can be safely unregistered on process exit, even if we have
776 * loops in the file referencing. We account only files that can hold other
777 * files because otherwise they can't form a loop and so are not interesting
780 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
782 #if defined(CONFIG_UNIX)
783 struct sock *sk = ctx->ring_sock->sk;
784 struct sk_buff_head *head = &sk->sk_receive_queue;
785 struct scm_fp_list *fpl;
788 if (likely(!io_file_need_scm(file)))
792 * See if we can merge this file into an existing skb SCM_RIGHTS
793 * file set. If there's no room, fall back to allocating a new skb
796 spin_lock_irq(&head->lock);
797 skb = skb_peek(head);
798 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
799 __skb_unlink(skb, head);
802 spin_unlock_irq(&head->lock);
805 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
809 skb = alloc_skb(0, GFP_KERNEL);
815 fpl->user = get_uid(current_user());
816 fpl->max = SCM_MAX_FD;
819 UNIXCB(skb).fp = fpl;
821 skb->scm_io_uring = 1;
822 skb->destructor = unix_destruct_scm;
823 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
826 fpl = UNIXCB(skb).fp;
827 fpl->fp[fpl->count++] = get_file(file);
828 unix_inflight(fpl->user, file);
829 skb_queue_head(head, skb);
835 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
837 struct file *file = prsrc->file;
838 #if defined(CONFIG_UNIX)
839 struct sock *sock = ctx->ring_sock->sk;
840 struct sk_buff_head list, *head = &sock->sk_receive_queue;
844 if (!io_file_need_scm(file)) {
849 __skb_queue_head_init(&list);
852 * Find the skb that holds this file in its SCM_RIGHTS. When found,
853 * remove this entry and rearrange the file array.
855 skb = skb_dequeue(head);
857 struct scm_fp_list *fp;
860 for (i = 0; i < fp->count; i++) {
863 if (fp->fp[i] != file)
866 unix_notinflight(fp->user, fp->fp[i]);
867 left = fp->count - 1 - i;
869 memmove(&fp->fp[i], &fp->fp[i + 1],
870 left * sizeof(struct file *));
877 __skb_queue_tail(&list, skb);
887 __skb_queue_tail(&list, skb);
889 skb = skb_dequeue(head);
892 if (skb_peek(&list)) {
893 spin_lock_irq(&head->lock);
894 while ((skb = __skb_dequeue(&list)) != NULL)
895 __skb_queue_tail(head, skb);
896 spin_unlock_irq(&head->lock);
903 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
904 unsigned nr_args, u64 __user *tags)
906 __s32 __user *fds = (__s32 __user *) arg;
915 if (nr_args > IORING_MAX_FIXED_FILES)
917 if (nr_args > rlimit(RLIMIT_NOFILE))
919 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
924 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
925 io_rsrc_data_free(ctx->file_data);
926 ctx->file_data = NULL;
930 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
931 struct io_fixed_file *file_slot;
933 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
937 /* allow sparse sets */
938 if (!fds || fd == -1) {
940 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
951 * Don't allow io_uring instances to be registered. If UNIX
952 * isn't enabled, then this causes a reference cycle and this
953 * instance can never get freed. If UNIX is enabled we'll
954 * handle it just fine, but there's still no point in allowing
955 * a ring fd as it doesn't support regular read/write anyway.
957 if (io_is_uring_fops(file)) {
961 ret = io_scm_file_account(ctx, file);
966 file_slot = io_fixed_file_slot(&ctx->file_table, i);
967 io_fixed_file_set(file_slot, file);
968 io_file_bitmap_set(&ctx->file_table, i);
971 /* default it to the whole table */
972 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
975 __io_sqe_files_unregister(ctx);
979 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
981 io_buffer_unmap(ctx, &prsrc->buf);
985 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
989 for (i = 0; i < ctx->nr_user_bufs; i++)
990 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
991 kfree(ctx->user_bufs);
992 io_rsrc_data_free(ctx->buf_data);
993 ctx->user_bufs = NULL;
994 ctx->buf_data = NULL;
995 ctx->nr_user_bufs = 0;
998 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
1000 unsigned nr = ctx->nr_user_bufs;
1007 * Quiesce may unlock ->uring_lock, and while it's not held
1008 * prevent new requests using the table.
1010 ctx->nr_user_bufs = 0;
1011 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
1012 ctx->nr_user_bufs = nr;
1014 __io_sqe_buffers_unregister(ctx);
1019 * Not super efficient, but this is just a registration time. And we do cache
1020 * the last compound head, so generally we'll only do a full search if we don't
1023 * We check if the given compound head page has already been accounted, to
1024 * avoid double accounting it. This allows us to account the full size of the
1025 * page, not just the constituent pages of a huge page.
1027 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
1028 int nr_pages, struct page *hpage)
1032 /* check current page array */
1033 for (i = 0; i < nr_pages; i++) {
1034 if (!PageCompound(pages[i]))
1036 if (compound_head(pages[i]) == hpage)
1040 /* check previously registered pages */
1041 for (i = 0; i < ctx->nr_user_bufs; i++) {
1042 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
1044 for (j = 0; j < imu->nr_bvecs; j++) {
1045 if (!PageCompound(imu->bvec[j].bv_page))
1047 if (compound_head(imu->bvec[j].bv_page) == hpage)
1055 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
1056 int nr_pages, struct io_mapped_ubuf *imu,
1057 struct page **last_hpage)
1061 imu->acct_pages = 0;
1062 for (i = 0; i < nr_pages; i++) {
1063 if (!PageCompound(pages[i])) {
1068 hpage = compound_head(pages[i]);
1069 if (hpage == *last_hpage)
1071 *last_hpage = hpage;
1072 if (headpage_already_acct(ctx, pages, i, hpage))
1074 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1078 if (!imu->acct_pages)
1081 ret = io_account_mem(ctx, imu->acct_pages);
1083 imu->acct_pages = 0;
1087 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1089 unsigned long start, end, nr_pages;
1090 struct vm_area_struct **vmas = NULL;
1091 struct page **pages = NULL;
1092 int i, pret, ret = -ENOMEM;
1094 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1095 start = ubuf >> PAGE_SHIFT;
1096 nr_pages = end - start;
1098 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1102 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
1108 mmap_read_lock(current->mm);
1109 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
1111 if (pret == nr_pages) {
1112 struct file *file = vmas[0]->vm_file;
1114 /* don't support file backed memory */
1115 for (i = 0; i < nr_pages; i++) {
1116 if (vmas[i]->vm_file != file) {
1122 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) {
1129 ret = pret < 0 ? pret : -EFAULT;
1131 mmap_read_unlock(current->mm);
1134 * if we did partial map, or found file backed vmas,
1135 * release any pages we did get
1138 unpin_user_pages(pages, pret);
1146 pages = ERR_PTR(ret);
1151 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1152 struct io_mapped_ubuf **pimu,
1153 struct page **last_hpage)
1155 struct io_mapped_ubuf *imu = NULL;
1156 struct page **pages = NULL;
1159 int ret, nr_pages, i;
1160 struct folio *folio = NULL;
1162 *pimu = ctx->dummy_ubuf;
1167 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1169 if (IS_ERR(pages)) {
1170 ret = PTR_ERR(pages);
1175 /* If it's a huge page, try to coalesce them into a single bvec entry */
1177 folio = page_folio(pages[0]);
1178 for (i = 1; i < nr_pages; i++) {
1179 if (page_folio(pages[i]) != folio) {
1186 * The pages are bound to the folio, it doesn't
1187 * actually unpin them but drops all but one reference,
1188 * which is usually put down by io_buffer_unmap().
1189 * Note, needs a better helper.
1191 unpin_user_pages(&pages[1], nr_pages - 1);
1196 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1200 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1202 unpin_user_pages(pages, nr_pages);
1206 off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1207 size = iov->iov_len;
1208 /* store original address for later verification */
1209 imu->ubuf = (unsigned long) iov->iov_base;
1210 imu->ubuf_end = imu->ubuf + iov->iov_len;
1211 imu->nr_bvecs = nr_pages;
1216 bvec_set_page(&imu->bvec[0], pages[0], size, off);
1219 for (i = 0; i < nr_pages; i++) {
1222 vec_len = min_t(size_t, size, PAGE_SIZE - off);
1223 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off);
1234 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1236 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1237 return ctx->user_bufs ? 0 : -ENOMEM;
1240 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1241 unsigned int nr_args, u64 __user *tags)
1243 struct page *last_hpage = NULL;
1244 struct io_rsrc_data *data;
1248 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1252 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1254 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
1257 ret = io_buffers_map_alloc(ctx, nr_args);
1259 io_rsrc_data_free(data);
1263 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1265 ret = io_copy_iov(ctx, &iov, arg, i);
1268 ret = io_buffer_validate(&iov);
1272 memset(&iov, 0, sizeof(iov));
1275 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1280 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1286 WARN_ON_ONCE(ctx->buf_data);
1288 ctx->buf_data = data;
1290 __io_sqe_buffers_unregister(ctx);
1294 int io_import_fixed(int ddir, struct iov_iter *iter,
1295 struct io_mapped_ubuf *imu,
1296 u64 buf_addr, size_t len)
1301 if (WARN_ON_ONCE(!imu))
1303 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1305 /* not inside the mapped region */
1306 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1310 * Might not be a start of buffer, set size appropriately
1311 * and advance us to the beginning.
1313 offset = buf_addr - imu->ubuf;
1314 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1318 * Don't use iov_iter_advance() here, as it's really slow for
1319 * using the latter parts of a big fixed buffer - it iterates
1320 * over each segment manually. We can cheat a bit here, because
1323 * 1) it's a BVEC iter, we set it up
1324 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1325 * first and last bvec
1327 * So just find our index, and adjust the iterator afterwards.
1328 * If the offset is within the first bvec (or the whole first
1329 * bvec, just use iov_iter_advance(). This makes it easier
1330 * since we can just skip the first segment, which may not
1331 * be PAGE_SIZE aligned.
1333 const struct bio_vec *bvec = imu->bvec;
1335 if (offset <= bvec->bv_len) {
1337 * Note, huge pages buffers consists of one large
1338 * bvec entry and should always go this way. The other
1339 * branch doesn't expect non PAGE_SIZE'd chunks.
1342 iter->nr_segs = bvec->bv_len;
1343 iter->count -= offset;
1344 iter->iov_offset = offset;
1346 unsigned long seg_skip;
1348 /* skip first vec */
1349 offset -= bvec->bv_len;
1350 seg_skip = 1 + (offset >> PAGE_SHIFT);
1352 iter->bvec = bvec + seg_skip;
1353 iter->nr_segs -= seg_skip;
1354 iter->count -= bvec->bv_len + offset;
1355 iter->iov_offset = offset & ~PAGE_MASK;