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 "alloc_cache.h"
17 #include "openclose.h"
21 struct io_rsrc_update {
28 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
29 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
30 struct io_mapped_ubuf **pimu,
31 struct page **last_hpage);
34 #define IORING_MAX_FIXED_FILES (1U << 20)
35 #define IORING_MAX_REG_BUFFERS (1U << 14)
37 static const struct io_mapped_ubuf dummy_ubuf = {
38 /* set invalid range, so io_import_fixed() fails meeting it */
43 int __io_account_mem(struct user_struct *user, unsigned long nr_pages)
45 unsigned long page_limit, cur_pages, new_pages;
50 /* Don't allow more pages than we can safely lock */
51 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
53 cur_pages = atomic_long_read(&user->locked_vm);
55 new_pages = cur_pages + nr_pages;
56 if (new_pages > page_limit)
58 } while (!atomic_long_try_cmpxchg(&user->locked_vm,
59 &cur_pages, new_pages));
63 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
66 __io_unaccount_mem(ctx->user, nr_pages);
69 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
72 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
77 ret = __io_account_mem(ctx->user, nr_pages);
83 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
88 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
89 void __user *arg, unsigned index)
91 struct iovec __user *src;
95 struct compat_iovec __user *ciovs;
96 struct compat_iovec ciov;
98 ciovs = (struct compat_iovec __user *) arg;
99 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
102 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
103 dst->iov_len = ciov.iov_len;
107 src = (struct iovec __user *) arg;
108 if (copy_from_user(dst, &src[index], sizeof(*dst)))
113 static int io_buffer_validate(struct iovec *iov)
115 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
118 * Don't impose further limits on the size and buffer
119 * constraints here, we'll -EINVAL later when IO is
120 * submitted if they are wrong.
123 return iov->iov_len ? -EFAULT : 0;
127 /* arbitrary limit, but we need something */
128 if (iov->iov_len > SZ_1G)
131 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
137 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
139 struct io_mapped_ubuf *imu = *slot;
142 if (imu != &dummy_ubuf) {
143 for (i = 0; i < imu->nr_bvecs; i++)
144 unpin_user_page(imu->bvec[i].bv_page);
146 io_unaccount_mem(ctx, imu->acct_pages);
152 static void io_rsrc_put_work(struct io_rsrc_node *node)
154 struct io_rsrc_put *prsrc = &node->item;
157 io_post_aux_cqe(node->ctx, prsrc->tag, 0, 0);
159 switch (node->type) {
160 case IORING_RSRC_FILE:
163 case IORING_RSRC_BUFFER:
164 io_rsrc_buf_put(node->ctx, prsrc);
172 void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *node)
174 if (!io_alloc_cache_put(&ctx->rsrc_node_cache, node))
178 void io_rsrc_node_ref_zero(struct io_rsrc_node *node)
179 __must_hold(&node->ctx->uring_lock)
181 struct io_ring_ctx *ctx = node->ctx;
183 while (!list_empty(&ctx->rsrc_ref_list)) {
184 node = list_first_entry(&ctx->rsrc_ref_list,
185 struct io_rsrc_node, node);
186 /* recycle ref nodes in order */
189 list_del(&node->node);
191 if (likely(!node->empty))
192 io_rsrc_put_work(node);
193 io_rsrc_node_destroy(ctx, node);
195 if (list_empty(&ctx->rsrc_ref_list) && unlikely(ctx->rsrc_quiesce))
196 wake_up_all(&ctx->rsrc_quiesce_wq);
199 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
201 struct io_rsrc_node *ref_node;
203 ref_node = io_alloc_cache_get(&ctx->rsrc_node_cache);
205 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
216 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
217 struct io_ring_ctx *ctx)
219 struct io_rsrc_node *backup;
223 /* As We may drop ->uring_lock, other task may have started quiesce */
227 backup = io_rsrc_node_alloc(ctx);
230 ctx->rsrc_node->empty = true;
231 ctx->rsrc_node->type = -1;
232 list_add_tail(&ctx->rsrc_node->node, &ctx->rsrc_ref_list);
233 io_put_rsrc_node(ctx, ctx->rsrc_node);
234 ctx->rsrc_node = backup;
236 if (list_empty(&ctx->rsrc_ref_list))
239 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
240 atomic_set(&ctx->cq_wait_nr, 1);
245 data->quiesce = true;
247 prepare_to_wait(&ctx->rsrc_quiesce_wq, &we, TASK_INTERRUPTIBLE);
248 mutex_unlock(&ctx->uring_lock);
250 ret = io_run_task_work_sig(ctx);
252 __set_current_state(TASK_RUNNING);
253 mutex_lock(&ctx->uring_lock);
254 if (list_empty(&ctx->rsrc_ref_list))
260 __set_current_state(TASK_RUNNING);
261 mutex_lock(&ctx->uring_lock);
263 } while (!list_empty(&ctx->rsrc_ref_list));
265 finish_wait(&ctx->rsrc_quiesce_wq, &we);
266 data->quiesce = false;
269 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
270 atomic_set(&ctx->cq_wait_nr, 0);
276 static void io_free_page_table(void **table, size_t size)
278 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
280 for (i = 0; i < nr_tables; i++)
285 static void io_rsrc_data_free(struct io_rsrc_data *data)
287 size_t size = data->nr * sizeof(data->tags[0][0]);
290 io_free_page_table((void **)data->tags, size);
294 static __cold void **io_alloc_page_table(size_t size)
296 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
297 size_t init_size = size;
300 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
304 for (i = 0; i < nr_tables; i++) {
305 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
307 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
309 io_free_page_table(table, init_size);
317 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, int type,
319 unsigned nr, struct io_rsrc_data **pdata)
321 struct io_rsrc_data *data;
325 data = kzalloc(sizeof(*data), GFP_KERNEL);
328 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
336 data->rsrc_type = type;
339 for (i = 0; i < nr; i++) {
340 u64 *tag_slot = io_get_tag_slot(data, i);
342 if (copy_from_user(tag_slot, &utags[i],
350 io_rsrc_data_free(data);
354 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
355 struct io_uring_rsrc_update2 *up,
358 u64 __user *tags = u64_to_user_ptr(up->tags);
359 __s32 __user *fds = u64_to_user_ptr(up->data);
360 struct io_rsrc_data *data = ctx->file_data;
361 struct io_fixed_file *file_slot;
367 if (up->offset + nr_args > ctx->nr_user_files)
370 for (done = 0; done < nr_args; done++) {
373 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
374 copy_from_user(&fd, &fds[done], sizeof(fd))) {
378 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
382 if (fd == IORING_REGISTER_FILES_SKIP)
385 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
386 file_slot = io_fixed_file_slot(&ctx->file_table, i);
388 if (file_slot->file_ptr) {
389 err = io_queue_rsrc_removal(data, i,
390 io_slot_file(file_slot));
393 file_slot->file_ptr = 0;
394 io_file_bitmap_clear(&ctx->file_table, i);
397 struct file *file = fget(fd);
404 * Don't allow io_uring instances to be registered.
406 if (io_is_uring_fops(file)) {
411 *io_get_tag_slot(data, i) = tag;
412 io_fixed_file_set(file_slot, file);
413 io_file_bitmap_set(&ctx->file_table, i);
416 return done ? done : err;
419 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
420 struct io_uring_rsrc_update2 *up,
421 unsigned int nr_args)
423 u64 __user *tags = u64_to_user_ptr(up->tags);
424 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
425 struct page *last_hpage = NULL;
431 if (up->offset + nr_args > ctx->nr_user_bufs)
434 for (done = 0; done < nr_args; done++) {
435 struct io_mapped_ubuf *imu;
438 err = io_copy_iov(ctx, &iov, iovs, done);
441 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
445 err = io_buffer_validate(&iov);
448 if (!iov.iov_base && tag) {
452 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
456 i = array_index_nospec(up->offset + done, ctx->nr_user_bufs);
457 if (ctx->user_bufs[i] != &dummy_ubuf) {
458 err = io_queue_rsrc_removal(ctx->buf_data, i,
461 io_buffer_unmap(ctx, &imu);
464 ctx->user_bufs[i] = (struct io_mapped_ubuf *)&dummy_ubuf;
467 ctx->user_bufs[i] = imu;
468 *io_get_tag_slot(ctx->buf_data, i) = tag;
470 return done ? done : err;
473 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
474 struct io_uring_rsrc_update2 *up,
479 lockdep_assert_held(&ctx->uring_lock);
481 if (check_add_overflow(up->offset, nr_args, &tmp))
485 case IORING_RSRC_FILE:
486 return __io_sqe_files_update(ctx, up, nr_args);
487 case IORING_RSRC_BUFFER:
488 return __io_sqe_buffers_update(ctx, up, nr_args);
493 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
496 struct io_uring_rsrc_update2 up;
500 memset(&up, 0, sizeof(up));
501 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
503 if (up.resv || up.resv2)
505 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
508 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
509 unsigned size, unsigned type)
511 struct io_uring_rsrc_update2 up;
513 if (size != sizeof(up))
515 if (copy_from_user(&up, arg, sizeof(up)))
517 if (!up.nr || up.resv || up.resv2)
519 return __io_register_rsrc_update(ctx, type, &up, up.nr);
522 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
523 unsigned int size, unsigned int type)
525 struct io_uring_rsrc_register rr;
527 /* keep it extendible */
528 if (size != sizeof(rr))
531 memset(&rr, 0, sizeof(rr));
532 if (copy_from_user(&rr, arg, size))
534 if (!rr.nr || rr.resv2)
536 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
540 case IORING_RSRC_FILE:
541 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
543 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
544 rr.nr, u64_to_user_ptr(rr.tags));
545 case IORING_RSRC_BUFFER:
546 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
548 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
549 rr.nr, u64_to_user_ptr(rr.tags));
554 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
556 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
558 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
560 if (sqe->rw_flags || sqe->splice_fd_in)
563 up->offset = READ_ONCE(sqe->off);
564 up->nr_args = READ_ONCE(sqe->len);
567 up->arg = READ_ONCE(sqe->addr);
571 static int io_files_update_with_index_alloc(struct io_kiocb *req,
572 unsigned int issue_flags)
574 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
575 __s32 __user *fds = u64_to_user_ptr(up->arg);
580 if (!req->ctx->file_data)
583 for (done = 0; done < up->nr_args; done++) {
584 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
594 ret = io_fixed_fd_install(req, issue_flags, file,
595 IORING_FILE_INDEX_ALLOC);
598 if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
599 __io_close_fixed(req->ctx, issue_flags, ret);
610 int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
612 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
613 struct io_ring_ctx *ctx = req->ctx;
614 struct io_uring_rsrc_update2 up2;
617 up2.offset = up->offset;
624 if (up->offset == IORING_FILE_INDEX_ALLOC) {
625 ret = io_files_update_with_index_alloc(req, issue_flags);
627 io_ring_submit_lock(ctx, issue_flags);
628 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
630 io_ring_submit_unlock(ctx, issue_flags);
635 io_req_set_res(req, ret, 0);
639 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, void *rsrc)
641 struct io_ring_ctx *ctx = data->ctx;
642 struct io_rsrc_node *node = ctx->rsrc_node;
643 u64 *tag_slot = io_get_tag_slot(data, idx);
645 ctx->rsrc_node = io_rsrc_node_alloc(ctx);
646 if (unlikely(!ctx->rsrc_node)) {
647 ctx->rsrc_node = node;
651 node->item.rsrc = rsrc;
652 node->type = data->rsrc_type;
653 node->item.tag = *tag_slot;
655 list_add_tail(&node->node, &ctx->rsrc_ref_list);
656 io_put_rsrc_node(ctx, node);
660 void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
664 for (i = 0; i < ctx->nr_user_files; i++) {
665 struct file *file = io_file_from_index(&ctx->file_table, i);
669 io_file_bitmap_clear(&ctx->file_table, i);
673 io_free_file_tables(&ctx->file_table);
674 io_file_table_set_alloc_range(ctx, 0, 0);
675 io_rsrc_data_free(ctx->file_data);
676 ctx->file_data = NULL;
677 ctx->nr_user_files = 0;
680 int io_sqe_files_unregister(struct io_ring_ctx *ctx)
682 unsigned nr = ctx->nr_user_files;
689 * Quiesce may unlock ->uring_lock, and while it's not held
690 * prevent new requests using the table.
692 ctx->nr_user_files = 0;
693 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
694 ctx->nr_user_files = nr;
696 __io_sqe_files_unregister(ctx);
700 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
701 unsigned nr_args, u64 __user *tags)
703 __s32 __user *fds = (__s32 __user *) arg;
712 if (nr_args > IORING_MAX_FIXED_FILES)
714 if (nr_args > rlimit(RLIMIT_NOFILE))
716 ret = io_rsrc_data_alloc(ctx, IORING_RSRC_FILE, tags, nr_args,
721 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
722 io_rsrc_data_free(ctx->file_data);
723 ctx->file_data = NULL;
727 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
728 struct io_fixed_file *file_slot;
730 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
734 /* allow sparse sets */
735 if (!fds || fd == -1) {
737 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
748 * Don't allow io_uring instances to be registered.
750 if (io_is_uring_fops(file)) {
754 file_slot = io_fixed_file_slot(&ctx->file_table, i);
755 io_fixed_file_set(file_slot, file);
756 io_file_bitmap_set(&ctx->file_table, i);
759 /* default it to the whole table */
760 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
763 __io_sqe_files_unregister(ctx);
767 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
769 io_buffer_unmap(ctx, &prsrc->buf);
773 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
777 for (i = 0; i < ctx->nr_user_bufs; i++)
778 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
779 kfree(ctx->user_bufs);
780 io_rsrc_data_free(ctx->buf_data);
781 ctx->user_bufs = NULL;
782 ctx->buf_data = NULL;
783 ctx->nr_user_bufs = 0;
786 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
788 unsigned nr = ctx->nr_user_bufs;
795 * Quiesce may unlock ->uring_lock, and while it's not held
796 * prevent new requests using the table.
798 ctx->nr_user_bufs = 0;
799 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
800 ctx->nr_user_bufs = nr;
802 __io_sqe_buffers_unregister(ctx);
807 * Not super efficient, but this is just a registration time. And we do cache
808 * the last compound head, so generally we'll only do a full search if we don't
811 * We check if the given compound head page has already been accounted, to
812 * avoid double accounting it. This allows us to account the full size of the
813 * page, not just the constituent pages of a huge page.
815 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
816 int nr_pages, struct page *hpage)
820 /* check current page array */
821 for (i = 0; i < nr_pages; i++) {
822 if (!PageCompound(pages[i]))
824 if (compound_head(pages[i]) == hpage)
828 /* check previously registered pages */
829 for (i = 0; i < ctx->nr_user_bufs; i++) {
830 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
832 for (j = 0; j < imu->nr_bvecs; j++) {
833 if (!PageCompound(imu->bvec[j].bv_page))
835 if (compound_head(imu->bvec[j].bv_page) == hpage)
843 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
844 int nr_pages, struct io_mapped_ubuf *imu,
845 struct page **last_hpage)
850 for (i = 0; i < nr_pages; i++) {
851 if (!PageCompound(pages[i])) {
856 hpage = compound_head(pages[i]);
857 if (hpage == *last_hpage)
860 if (headpage_already_acct(ctx, pages, i, hpage))
862 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
866 if (!imu->acct_pages)
869 ret = io_account_mem(ctx, imu->acct_pages);
875 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
876 struct io_mapped_ubuf **pimu,
877 struct page **last_hpage)
879 struct io_mapped_ubuf *imu = NULL;
880 struct page **pages = NULL;
883 int ret, nr_pages, i;
884 struct folio *folio = NULL;
886 *pimu = (struct io_mapped_ubuf *)&dummy_ubuf;
891 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
894 ret = PTR_ERR(pages);
899 /* If it's a huge page, try to coalesce them into a single bvec entry */
901 folio = page_folio(pages[0]);
902 for (i = 1; i < nr_pages; i++) {
904 * Pages must be consecutive and on the same folio for
907 if (page_folio(pages[i]) != folio ||
908 pages[i] != pages[i - 1] + 1) {
915 * The pages are bound to the folio, it doesn't
916 * actually unpin them but drops all but one reference,
917 * which is usually put down by io_buffer_unmap().
918 * Note, needs a better helper.
920 unpin_user_pages(&pages[1], nr_pages - 1);
925 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
929 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
931 unpin_user_pages(pages, nr_pages);
935 off = (unsigned long) iov->iov_base & ~PAGE_MASK;
937 /* store original address for later verification */
938 imu->ubuf = (unsigned long) iov->iov_base;
939 imu->ubuf_end = imu->ubuf + iov->iov_len;
940 imu->nr_bvecs = nr_pages;
945 bvec_set_page(&imu->bvec[0], pages[0], size, off);
948 for (i = 0; i < nr_pages; i++) {
951 vec_len = min_t(size_t, size, PAGE_SIZE - off);
952 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off);
963 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
965 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
966 return ctx->user_bufs ? 0 : -ENOMEM;
969 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
970 unsigned int nr_args, u64 __user *tags)
972 struct page *last_hpage = NULL;
973 struct io_rsrc_data *data;
977 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
981 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
983 ret = io_rsrc_data_alloc(ctx, IORING_RSRC_BUFFER, tags, nr_args, &data);
986 ret = io_buffers_map_alloc(ctx, nr_args);
988 io_rsrc_data_free(data);
992 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
994 ret = io_copy_iov(ctx, &iov, arg, i);
997 ret = io_buffer_validate(&iov);
1001 memset(&iov, 0, sizeof(iov));
1004 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1009 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1015 WARN_ON_ONCE(ctx->buf_data);
1017 ctx->buf_data = data;
1019 __io_sqe_buffers_unregister(ctx);
1023 int io_import_fixed(int ddir, struct iov_iter *iter,
1024 struct io_mapped_ubuf *imu,
1025 u64 buf_addr, size_t len)
1030 if (WARN_ON_ONCE(!imu))
1032 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1034 /* not inside the mapped region */
1035 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1039 * Might not be a start of buffer, set size appropriately
1040 * and advance us to the beginning.
1042 offset = buf_addr - imu->ubuf;
1043 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1047 * Don't use iov_iter_advance() here, as it's really slow for
1048 * using the latter parts of a big fixed buffer - it iterates
1049 * over each segment manually. We can cheat a bit here, because
1052 * 1) it's a BVEC iter, we set it up
1053 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1054 * first and last bvec
1056 * So just find our index, and adjust the iterator afterwards.
1057 * If the offset is within the first bvec (or the whole first
1058 * bvec, just use iov_iter_advance(). This makes it easier
1059 * since we can just skip the first segment, which may not
1060 * be PAGE_SIZE aligned.
1062 const struct bio_vec *bvec = imu->bvec;
1064 if (offset < bvec->bv_len) {
1066 * Note, huge pages buffers consists of one large
1067 * bvec entry and should always go this way. The other
1068 * branch doesn't expect non PAGE_SIZE'd chunks.
1071 iter->count -= offset;
1072 iter->iov_offset = offset;
1074 unsigned long seg_skip;
1076 /* skip first vec */
1077 offset -= bvec->bv_len;
1078 seg_skip = 1 + (offset >> PAGE_SHIFT);
1080 iter->bvec = bvec + seg_skip;
1081 iter->nr_segs -= seg_skip;
1082 iter->count -= bvec->bv_len + offset;
1083 iter->iov_offset = offset & ~PAGE_MASK;