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_ring_ctx *ctx = rsrc_data->ctx;
162 struct io_rsrc_put *prsrc, *tmp;
164 if (ref_node->inline_items)
165 io_rsrc_put_work_one(rsrc_data, &ref_node->item);
167 list_for_each_entry_safe(prsrc, tmp, &ref_node->item_list, list) {
168 list_del(&prsrc->list);
169 io_rsrc_put_work_one(rsrc_data, prsrc);
173 io_rsrc_node_destroy(rsrc_data->ctx, ref_node);
174 if (io_put_rsrc_data_ref(rsrc_data))
175 wake_up_all(&ctx->rsrc_quiesce_wq);
178 void io_wait_rsrc_data(struct io_rsrc_data *data)
181 WARN_ON_ONCE(!io_put_rsrc_data_ref(data));
184 void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *node)
186 if (!io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache))
190 void io_rsrc_node_ref_zero(struct io_rsrc_node *node)
191 __must_hold(&node->rsrc_data->ctx->uring_lock)
193 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 */
201 list_del(&node->node);
202 __io_rsrc_put_work(node);
206 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
208 struct io_rsrc_node *ref_node;
209 struct io_cache_entry *entry;
211 entry = io_alloc_cache_get(&ctx->rsrc_node_cache);
213 ref_node = container_of(entry, struct io_rsrc_node, cache);
215 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
220 ref_node->rsrc_data = NULL;
222 INIT_LIST_HEAD(&ref_node->node);
223 INIT_LIST_HEAD(&ref_node->item_list);
224 ref_node->inline_items = 0;
228 void io_rsrc_node_switch(struct io_ring_ctx *ctx,
229 struct io_rsrc_data *data_to_kill)
230 __must_hold(&ctx->uring_lock)
232 struct io_rsrc_node *node = ctx->rsrc_node;
233 struct io_rsrc_node *backup = io_rsrc_node_alloc(ctx);
235 if (WARN_ON_ONCE(!backup))
238 data_to_kill->refs++;
239 node->rsrc_data = data_to_kill;
240 list_add_tail(&node->node, &ctx->rsrc_ref_list);
242 io_put_rsrc_node(ctx, node);
243 ctx->rsrc_node = backup;
246 int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
248 if (io_alloc_cache_empty(&ctx->rsrc_node_cache)) {
249 struct io_rsrc_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
253 io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache);
258 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
259 struct io_ring_ctx *ctx)
264 /* As we may drop ->uring_lock, other task may have started quiesce */
267 ret = io_rsrc_node_switch_start(ctx);
270 io_rsrc_node_switch(ctx, data);
272 /* kill initial ref */
273 if (io_put_rsrc_data_ref(data))
276 data->quiesce = true;
278 prepare_to_wait(&ctx->rsrc_quiesce_wq, &we, TASK_INTERRUPTIBLE);
279 mutex_unlock(&ctx->uring_lock);
281 ret = io_run_task_work_sig(ctx);
283 mutex_lock(&ctx->uring_lock);
287 /* restore the master reference */
294 __set_current_state(TASK_RUNNING);
295 mutex_lock(&ctx->uring_lock);
297 } while (data->refs);
299 finish_wait(&ctx->rsrc_quiesce_wq, &we);
300 data->quiesce = false;
304 static void io_free_page_table(void **table, size_t size)
306 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
308 for (i = 0; i < nr_tables; i++)
313 static void io_rsrc_data_free(struct io_rsrc_data *data)
315 size_t size = data->nr * sizeof(data->tags[0][0]);
318 io_free_page_table((void **)data->tags, size);
322 static __cold void **io_alloc_page_table(size_t size)
324 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
325 size_t init_size = size;
328 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
332 for (i = 0; i < nr_tables; i++) {
333 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
335 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
337 io_free_page_table(table, init_size);
345 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx,
346 rsrc_put_fn *do_put, u64 __user *utags,
347 unsigned nr, struct io_rsrc_data **pdata)
349 struct io_rsrc_data *data;
353 data = kzalloc(sizeof(*data), GFP_KERNEL);
356 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
364 data->do_put = do_put;
368 for (i = 0; i < nr; i++) {
369 u64 *tag_slot = io_get_tag_slot(data, i);
371 if (copy_from_user(tag_slot, &utags[i],
379 io_rsrc_data_free(data);
383 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
384 struct io_uring_rsrc_update2 *up,
387 u64 __user *tags = u64_to_user_ptr(up->tags);
388 __s32 __user *fds = u64_to_user_ptr(up->data);
389 struct io_rsrc_data *data = ctx->file_data;
390 struct io_fixed_file *file_slot;
394 bool needs_switch = false;
398 if (up->offset + nr_args > ctx->nr_user_files)
401 for (done = 0; done < nr_args; done++) {
404 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
405 copy_from_user(&fd, &fds[done], sizeof(fd))) {
409 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
413 if (fd == IORING_REGISTER_FILES_SKIP)
416 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
417 file_slot = io_fixed_file_slot(&ctx->file_table, i);
419 if (file_slot->file_ptr) {
420 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
421 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
424 file_slot->file_ptr = 0;
425 io_file_bitmap_clear(&ctx->file_table, i);
435 * Don't allow io_uring instances to be registered. If
436 * UNIX isn't enabled, then this causes a reference
437 * cycle and this instance can never get freed. If UNIX
438 * is enabled we'll handle it just fine, but there's
439 * still no point in allowing a ring fd as it doesn't
440 * support regular read/write anyway.
442 if (io_is_uring_fops(file)) {
447 err = io_scm_file_account(ctx, file);
452 *io_get_tag_slot(data, i) = tag;
453 io_fixed_file_set(file_slot, file);
454 io_file_bitmap_set(&ctx->file_table, i);
459 io_rsrc_node_switch(ctx, data);
460 return done ? done : err;
463 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
464 struct io_uring_rsrc_update2 *up,
465 unsigned int nr_args)
467 u64 __user *tags = u64_to_user_ptr(up->tags);
468 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
469 struct page *last_hpage = NULL;
470 bool needs_switch = false;
476 if (up->offset + nr_args > ctx->nr_user_bufs)
479 for (done = 0; done < nr_args; done++) {
480 struct io_mapped_ubuf *imu;
481 int offset = up->offset + done;
484 err = io_copy_iov(ctx, &iov, iovs, done);
487 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
491 err = io_buffer_validate(&iov);
494 if (!iov.iov_base && tag) {
498 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
502 i = array_index_nospec(offset, ctx->nr_user_bufs);
503 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
504 err = io_queue_rsrc_removal(ctx->buf_data, i,
505 ctx->rsrc_node, ctx->user_bufs[i]);
507 io_buffer_unmap(ctx, &imu);
510 ctx->user_bufs[i] = ctx->dummy_ubuf;
514 ctx->user_bufs[i] = imu;
515 *io_get_tag_slot(ctx->buf_data, i) = tag;
519 io_rsrc_node_switch(ctx, ctx->buf_data);
520 return done ? done : err;
523 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
524 struct io_uring_rsrc_update2 *up,
530 lockdep_assert_held(&ctx->uring_lock);
532 if (check_add_overflow(up->offset, nr_args, &tmp))
534 err = io_rsrc_node_switch_start(ctx);
539 case IORING_RSRC_FILE:
540 return __io_sqe_files_update(ctx, up, nr_args);
541 case IORING_RSRC_BUFFER:
542 return __io_sqe_buffers_update(ctx, up, nr_args);
547 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
550 struct io_uring_rsrc_update2 up;
554 memset(&up, 0, sizeof(up));
555 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
557 if (up.resv || up.resv2)
559 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
562 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
563 unsigned size, unsigned type)
565 struct io_uring_rsrc_update2 up;
567 if (size != sizeof(up))
569 if (copy_from_user(&up, arg, sizeof(up)))
571 if (!up.nr || up.resv || up.resv2)
573 return __io_register_rsrc_update(ctx, type, &up, up.nr);
576 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
577 unsigned int size, unsigned int type)
579 struct io_uring_rsrc_register rr;
581 /* keep it extendible */
582 if (size != sizeof(rr))
585 memset(&rr, 0, sizeof(rr));
586 if (copy_from_user(&rr, arg, size))
588 if (!rr.nr || rr.resv2)
590 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
594 case IORING_RSRC_FILE:
595 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
597 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
598 rr.nr, u64_to_user_ptr(rr.tags));
599 case IORING_RSRC_BUFFER:
600 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
602 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
603 rr.nr, u64_to_user_ptr(rr.tags));
608 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
610 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
612 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
614 if (sqe->rw_flags || sqe->splice_fd_in)
617 up->offset = READ_ONCE(sqe->off);
618 up->nr_args = READ_ONCE(sqe->len);
621 up->arg = READ_ONCE(sqe->addr);
625 static int io_files_update_with_index_alloc(struct io_kiocb *req,
626 unsigned int issue_flags)
628 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
629 __s32 __user *fds = u64_to_user_ptr(up->arg);
634 if (!req->ctx->file_data)
637 for (done = 0; done < up->nr_args; done++) {
638 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
648 ret = io_fixed_fd_install(req, issue_flags, file,
649 IORING_FILE_INDEX_ALLOC);
652 if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
653 __io_close_fixed(req->ctx, issue_flags, ret);
664 int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
666 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
667 struct io_ring_ctx *ctx = req->ctx;
668 struct io_uring_rsrc_update2 up2;
671 up2.offset = up->offset;
678 if (up->offset == IORING_FILE_INDEX_ALLOC) {
679 ret = io_files_update_with_index_alloc(req, issue_flags);
681 io_ring_submit_lock(ctx, issue_flags);
682 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
684 io_ring_submit_unlock(ctx, issue_flags);
689 io_req_set_res(req, ret, 0);
693 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
694 struct io_rsrc_node *node, void *rsrc)
696 u64 *tag_slot = io_get_tag_slot(data, idx);
697 struct io_rsrc_put *prsrc;
698 bool inline_item = true;
700 if (!node->inline_items) {
702 node->inline_items++;
704 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
710 prsrc->tag = *tag_slot;
714 list_add(&prsrc->list, &node->item_list);
718 void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
722 for (i = 0; i < ctx->nr_user_files; i++) {
723 struct file *file = io_file_from_index(&ctx->file_table, i);
725 /* skip scm accounted files, they'll be freed by ->ring_sock */
726 if (!file || io_file_need_scm(file))
728 io_file_bitmap_clear(&ctx->file_table, i);
732 #if defined(CONFIG_UNIX)
733 if (ctx->ring_sock) {
734 struct sock *sock = ctx->ring_sock->sk;
737 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
741 io_free_file_tables(&ctx->file_table);
742 io_file_table_set_alloc_range(ctx, 0, 0);
743 io_rsrc_data_free(ctx->file_data);
744 ctx->file_data = NULL;
745 ctx->nr_user_files = 0;
748 int io_sqe_files_unregister(struct io_ring_ctx *ctx)
750 unsigned nr = ctx->nr_user_files;
757 * Quiesce may unlock ->uring_lock, and while it's not held
758 * prevent new requests using the table.
760 ctx->nr_user_files = 0;
761 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
762 ctx->nr_user_files = nr;
764 __io_sqe_files_unregister(ctx);
769 * Ensure the UNIX gc is aware of our file set, so we are certain that
770 * the io_uring can be safely unregistered on process exit, even if we have
771 * loops in the file referencing. We account only files that can hold other
772 * files because otherwise they can't form a loop and so are not interesting
775 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
777 #if defined(CONFIG_UNIX)
778 struct sock *sk = ctx->ring_sock->sk;
779 struct sk_buff_head *head = &sk->sk_receive_queue;
780 struct scm_fp_list *fpl;
783 if (likely(!io_file_need_scm(file)))
787 * See if we can merge this file into an existing skb SCM_RIGHTS
788 * file set. If there's no room, fall back to allocating a new skb
791 spin_lock_irq(&head->lock);
792 skb = skb_peek(head);
793 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
794 __skb_unlink(skb, head);
797 spin_unlock_irq(&head->lock);
800 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
804 skb = alloc_skb(0, GFP_KERNEL);
810 fpl->user = get_uid(current_user());
811 fpl->max = SCM_MAX_FD;
814 UNIXCB(skb).fp = fpl;
816 skb->scm_io_uring = 1;
817 skb->destructor = unix_destruct_scm;
818 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
821 fpl = UNIXCB(skb).fp;
822 fpl->fp[fpl->count++] = get_file(file);
823 unix_inflight(fpl->user, file);
824 skb_queue_head(head, skb);
830 static __cold void io_rsrc_file_scm_put(struct io_ring_ctx *ctx, struct file *file)
832 #if defined(CONFIG_UNIX)
833 struct sock *sock = ctx->ring_sock->sk;
834 struct sk_buff_head list, *head = &sock->sk_receive_queue;
838 __skb_queue_head_init(&list);
841 * Find the skb that holds this file in its SCM_RIGHTS. When found,
842 * remove this entry and rearrange the file array.
844 skb = skb_dequeue(head);
846 struct scm_fp_list *fp;
849 for (i = 0; i < fp->count; i++) {
852 if (fp->fp[i] != file)
855 unix_notinflight(fp->user, fp->fp[i]);
856 left = fp->count - 1 - i;
858 memmove(&fp->fp[i], &fp->fp[i + 1],
859 left * sizeof(struct file *));
866 __skb_queue_tail(&list, skb);
876 __skb_queue_tail(&list, skb);
878 skb = skb_dequeue(head);
881 if (skb_peek(&list)) {
882 spin_lock_irq(&head->lock);
883 while ((skb = __skb_dequeue(&list)) != NULL)
884 __skb_queue_tail(head, skb);
885 spin_unlock_irq(&head->lock);
890 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
892 struct file *file = prsrc->file;
894 if (likely(!io_file_need_scm(file)))
897 io_rsrc_file_scm_put(ctx, file);
900 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
901 unsigned nr_args, u64 __user *tags)
903 __s32 __user *fds = (__s32 __user *) arg;
912 if (nr_args > IORING_MAX_FIXED_FILES)
914 if (nr_args > rlimit(RLIMIT_NOFILE))
916 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
921 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
922 io_rsrc_data_free(ctx->file_data);
923 ctx->file_data = NULL;
927 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
928 struct io_fixed_file *file_slot;
930 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
934 /* allow sparse sets */
935 if (!fds || fd == -1) {
937 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
948 * Don't allow io_uring instances to be registered. If UNIX
949 * isn't enabled, then this causes a reference cycle and this
950 * instance can never get freed. If UNIX is enabled we'll
951 * handle it just fine, but there's still no point in allowing
952 * a ring fd as it doesn't support regular read/write anyway.
954 if (io_is_uring_fops(file)) {
958 ret = io_scm_file_account(ctx, file);
963 file_slot = io_fixed_file_slot(&ctx->file_table, i);
964 io_fixed_file_set(file_slot, file);
965 io_file_bitmap_set(&ctx->file_table, i);
968 /* default it to the whole table */
969 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
972 __io_sqe_files_unregister(ctx);
976 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
978 io_buffer_unmap(ctx, &prsrc->buf);
982 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
986 for (i = 0; i < ctx->nr_user_bufs; i++)
987 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
988 kfree(ctx->user_bufs);
989 io_rsrc_data_free(ctx->buf_data);
990 ctx->user_bufs = NULL;
991 ctx->buf_data = NULL;
992 ctx->nr_user_bufs = 0;
995 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
997 unsigned nr = ctx->nr_user_bufs;
1004 * Quiesce may unlock ->uring_lock, and while it's not held
1005 * prevent new requests using the table.
1007 ctx->nr_user_bufs = 0;
1008 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
1009 ctx->nr_user_bufs = nr;
1011 __io_sqe_buffers_unregister(ctx);
1016 * Not super efficient, but this is just a registration time. And we do cache
1017 * the last compound head, so generally we'll only do a full search if we don't
1020 * We check if the given compound head page has already been accounted, to
1021 * avoid double accounting it. This allows us to account the full size of the
1022 * page, not just the constituent pages of a huge page.
1024 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
1025 int nr_pages, struct page *hpage)
1029 /* check current page array */
1030 for (i = 0; i < nr_pages; i++) {
1031 if (!PageCompound(pages[i]))
1033 if (compound_head(pages[i]) == hpage)
1037 /* check previously registered pages */
1038 for (i = 0; i < ctx->nr_user_bufs; i++) {
1039 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
1041 for (j = 0; j < imu->nr_bvecs; j++) {
1042 if (!PageCompound(imu->bvec[j].bv_page))
1044 if (compound_head(imu->bvec[j].bv_page) == hpage)
1052 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
1053 int nr_pages, struct io_mapped_ubuf *imu,
1054 struct page **last_hpage)
1058 imu->acct_pages = 0;
1059 for (i = 0; i < nr_pages; i++) {
1060 if (!PageCompound(pages[i])) {
1065 hpage = compound_head(pages[i]);
1066 if (hpage == *last_hpage)
1068 *last_hpage = hpage;
1069 if (headpage_already_acct(ctx, pages, i, hpage))
1071 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1075 if (!imu->acct_pages)
1078 ret = io_account_mem(ctx, imu->acct_pages);
1080 imu->acct_pages = 0;
1084 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1086 unsigned long start, end, nr_pages;
1087 struct vm_area_struct **vmas = NULL;
1088 struct page **pages = NULL;
1089 int i, pret, ret = -ENOMEM;
1091 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1092 start = ubuf >> PAGE_SHIFT;
1093 nr_pages = end - start;
1095 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1099 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
1105 mmap_read_lock(current->mm);
1106 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
1108 if (pret == nr_pages) {
1109 struct file *file = vmas[0]->vm_file;
1111 /* don't support file backed memory */
1112 for (i = 0; i < nr_pages; i++) {
1113 if (vmas[i]->vm_file != file) {
1119 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) {
1126 ret = pret < 0 ? pret : -EFAULT;
1128 mmap_read_unlock(current->mm);
1131 * if we did partial map, or found file backed vmas,
1132 * release any pages we did get
1135 unpin_user_pages(pages, pret);
1143 pages = ERR_PTR(ret);
1148 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1149 struct io_mapped_ubuf **pimu,
1150 struct page **last_hpage)
1152 struct io_mapped_ubuf *imu = NULL;
1153 struct page **pages = NULL;
1156 int ret, nr_pages, i;
1157 struct folio *folio = NULL;
1159 *pimu = ctx->dummy_ubuf;
1164 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1166 if (IS_ERR(pages)) {
1167 ret = PTR_ERR(pages);
1172 /* If it's a huge page, try to coalesce them into a single bvec entry */
1174 folio = page_folio(pages[0]);
1175 for (i = 1; i < nr_pages; i++) {
1176 if (page_folio(pages[i]) != folio) {
1183 * The pages are bound to the folio, it doesn't
1184 * actually unpin them but drops all but one reference,
1185 * which is usually put down by io_buffer_unmap().
1186 * Note, needs a better helper.
1188 unpin_user_pages(&pages[1], nr_pages - 1);
1193 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1197 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1199 unpin_user_pages(pages, nr_pages);
1203 off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1204 size = iov->iov_len;
1205 /* store original address for later verification */
1206 imu->ubuf = (unsigned long) iov->iov_base;
1207 imu->ubuf_end = imu->ubuf + iov->iov_len;
1208 imu->nr_bvecs = nr_pages;
1213 bvec_set_page(&imu->bvec[0], pages[0], size, off);
1216 for (i = 0; i < nr_pages; i++) {
1219 vec_len = min_t(size_t, size, PAGE_SIZE - off);
1220 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off);
1231 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1233 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1234 return ctx->user_bufs ? 0 : -ENOMEM;
1237 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1238 unsigned int nr_args, u64 __user *tags)
1240 struct page *last_hpage = NULL;
1241 struct io_rsrc_data *data;
1245 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1249 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1251 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
1254 ret = io_buffers_map_alloc(ctx, nr_args);
1256 io_rsrc_data_free(data);
1260 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1262 ret = io_copy_iov(ctx, &iov, arg, i);
1265 ret = io_buffer_validate(&iov);
1269 memset(&iov, 0, sizeof(iov));
1272 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1277 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1283 WARN_ON_ONCE(ctx->buf_data);
1285 ctx->buf_data = data;
1287 __io_sqe_buffers_unregister(ctx);
1291 int io_import_fixed(int ddir, struct iov_iter *iter,
1292 struct io_mapped_ubuf *imu,
1293 u64 buf_addr, size_t len)
1298 if (WARN_ON_ONCE(!imu))
1300 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1302 /* not inside the mapped region */
1303 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1307 * Might not be a start of buffer, set size appropriately
1308 * and advance us to the beginning.
1310 offset = buf_addr - imu->ubuf;
1311 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1315 * Don't use iov_iter_advance() here, as it's really slow for
1316 * using the latter parts of a big fixed buffer - it iterates
1317 * over each segment manually. We can cheat a bit here, because
1320 * 1) it's a BVEC iter, we set it up
1321 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1322 * first and last bvec
1324 * So just find our index, and adjust the iterator afterwards.
1325 * If the offset is within the first bvec (or the whole first
1326 * bvec, just use iov_iter_advance(). This makes it easier
1327 * since we can just skip the first segment, which may not
1328 * be PAGE_SIZE aligned.
1330 const struct bio_vec *bvec = imu->bvec;
1332 if (offset <= bvec->bv_len) {
1334 * Note, huge pages buffers consists of one large
1335 * bvec entry and should always go this way. The other
1336 * branch doesn't expect non PAGE_SIZE'd chunks.
1339 iter->nr_segs = bvec->bv_len;
1340 iter->count -= offset;
1341 iter->iov_offset = offset;
1343 unsigned long seg_skip;
1345 /* skip first vec */
1346 offset -= bvec->bv_len;
1347 seg_skip = 1 + (offset >> PAGE_SHIFT);
1349 iter->bvec = bvec + seg_skip;
1350 iter->nr_segs -= seg_skip;
1351 iter->count -= bvec->bv_len + offset;
1352 iter->iov_offset = offset & ~PAGE_MASK;