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/namei.h>
9 #include <linux/poll.h>
10 #include <linux/io_uring.h>
12 #include <uapi/linux/io_uring.h>
18 #define IO_BUFFER_LIST_BUF_PER_PAGE (PAGE_SIZE / sizeof(struct io_uring_buf))
22 /* BIDs are addressed by a 16-bit field in a CQE */
23 #define MAX_BIDS_PER_BGID (1 << 16)
25 struct io_provide_buf {
34 static inline struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
37 if (ctx->io_bl && bgid < BGID_ARRAY)
38 return &ctx->io_bl[bgid];
40 return xa_load(&ctx->io_bl_xa, bgid);
43 static int io_buffer_add_list(struct io_ring_ctx *ctx,
44 struct io_buffer_list *bl, unsigned int bgid)
47 if (bgid < BGID_ARRAY)
50 return xa_err(xa_store(&ctx->io_bl_xa, bgid, bl, GFP_KERNEL));
53 void io_kbuf_recycle_legacy(struct io_kiocb *req, unsigned issue_flags)
55 struct io_ring_ctx *ctx = req->ctx;
56 struct io_buffer_list *bl;
57 struct io_buffer *buf;
60 * For legacy provided buffer mode, don't recycle if we already did
61 * IO to this buffer. For ring-mapped provided buffer mode, we should
62 * increment ring->head to explicitly monopolize the buffer to avoid
65 if (req->flags & REQ_F_PARTIAL_IO)
68 io_ring_submit_lock(ctx, issue_flags);
71 bl = io_buffer_get_list(ctx, buf->bgid);
72 list_add(&buf->list, &bl->buf_list);
73 req->flags &= ~REQ_F_BUFFER_SELECTED;
74 req->buf_index = buf->bgid;
76 io_ring_submit_unlock(ctx, issue_flags);
80 unsigned int __io_put_kbuf(struct io_kiocb *req, unsigned issue_flags)
85 * We can add this buffer back to two lists:
87 * 1) The io_buffers_cache list. This one is protected by the
88 * ctx->uring_lock. If we already hold this lock, add back to this
89 * list as we can grab it from issue as well.
90 * 2) The io_buffers_comp list. This one is protected by the
91 * ctx->completion_lock.
93 * We migrate buffers from the comp_list to the issue cache list
96 if (req->flags & REQ_F_BUFFER_RING) {
97 /* no buffers to recycle for this case */
98 cflags = __io_put_kbuf_list(req, NULL);
99 } else if (issue_flags & IO_URING_F_UNLOCKED) {
100 struct io_ring_ctx *ctx = req->ctx;
102 spin_lock(&ctx->completion_lock);
103 cflags = __io_put_kbuf_list(req, &ctx->io_buffers_comp);
104 spin_unlock(&ctx->completion_lock);
106 lockdep_assert_held(&req->ctx->uring_lock);
108 cflags = __io_put_kbuf_list(req, &req->ctx->io_buffers_cache);
113 static void __user *io_provided_buffer_select(struct io_kiocb *req, size_t *len,
114 struct io_buffer_list *bl)
116 if (!list_empty(&bl->buf_list)) {
117 struct io_buffer *kbuf;
119 kbuf = list_first_entry(&bl->buf_list, struct io_buffer, list);
120 list_del(&kbuf->list);
121 if (*len == 0 || *len > kbuf->len)
123 req->flags |= REQ_F_BUFFER_SELECTED;
125 req->buf_index = kbuf->bid;
126 return u64_to_user_ptr(kbuf->addr);
131 static void __user *io_ring_buffer_select(struct io_kiocb *req, size_t *len,
132 struct io_buffer_list *bl,
133 unsigned int issue_flags)
135 struct io_uring_buf_ring *br = bl->buf_ring;
136 struct io_uring_buf *buf;
137 __u16 head = bl->head;
139 if (unlikely(smp_load_acquire(&br->tail) == head))
143 /* mmaped buffers are always contig */
144 if (bl->is_mmap || head < IO_BUFFER_LIST_BUF_PER_PAGE) {
145 buf = &br->bufs[head];
147 int off = head & (IO_BUFFER_LIST_BUF_PER_PAGE - 1);
148 int index = head / IO_BUFFER_LIST_BUF_PER_PAGE;
149 buf = page_address(bl->buf_pages[index]);
152 if (*len == 0 || *len > buf->len)
154 req->flags |= REQ_F_BUFFER_RING;
156 req->buf_index = buf->bid;
158 if (issue_flags & IO_URING_F_UNLOCKED || !file_can_poll(req->file)) {
160 * If we came in unlocked, we have no choice but to consume the
161 * buffer here, otherwise nothing ensures that the buffer won't
162 * get used by others. This does mean it'll be pinned until the
163 * IO completes, coming in unlocked means we're being called from
164 * io-wq context and there may be further retries in async hybrid
165 * mode. For the locked case, the caller must call commit when
166 * the transfer completes (or if we get -EAGAIN and must poll of
169 req->buf_list = NULL;
172 return u64_to_user_ptr(buf->addr);
175 void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
176 unsigned int issue_flags)
178 struct io_ring_ctx *ctx = req->ctx;
179 struct io_buffer_list *bl;
180 void __user *ret = NULL;
182 io_ring_submit_lock(req->ctx, issue_flags);
184 bl = io_buffer_get_list(ctx, req->buf_index);
187 ret = io_ring_buffer_select(req, len, bl, issue_flags);
189 ret = io_provided_buffer_select(req, len, bl);
191 io_ring_submit_unlock(req->ctx, issue_flags);
195 static __cold int io_init_bl_list(struct io_ring_ctx *ctx)
199 ctx->io_bl = kcalloc(BGID_ARRAY, sizeof(struct io_buffer_list),
204 for (i = 0; i < BGID_ARRAY; i++) {
205 INIT_LIST_HEAD(&ctx->io_bl[i].buf_list);
206 ctx->io_bl[i].bgid = i;
212 static int __io_remove_buffers(struct io_ring_ctx *ctx,
213 struct io_buffer_list *bl, unsigned nbufs)
217 /* shouldn't happen */
222 i = bl->buf_ring->tail - bl->head;
224 folio_put(virt_to_folio(bl->buf_ring));
227 } else if (bl->buf_nr_pages) {
230 for (j = 0; j < bl->buf_nr_pages; j++)
231 unpin_user_page(bl->buf_pages[j]);
232 kvfree(bl->buf_pages);
233 bl->buf_pages = NULL;
234 bl->buf_nr_pages = 0;
236 /* make sure it's seen as empty */
237 INIT_LIST_HEAD(&bl->buf_list);
242 /* protects io_buffers_cache */
243 lockdep_assert_held(&ctx->uring_lock);
245 while (!list_empty(&bl->buf_list)) {
246 struct io_buffer *nxt;
248 nxt = list_first_entry(&bl->buf_list, struct io_buffer, list);
249 list_move(&nxt->list, &ctx->io_buffers_cache);
258 void io_destroy_buffers(struct io_ring_ctx *ctx)
260 struct io_buffer_list *bl;
264 for (i = 0; i < BGID_ARRAY; i++) {
267 __io_remove_buffers(ctx, &ctx->io_bl[i], -1U);
270 xa_for_each(&ctx->io_bl_xa, index, bl) {
271 xa_erase(&ctx->io_bl_xa, bl->bgid);
272 __io_remove_buffers(ctx, bl, -1U);
276 while (!list_empty(&ctx->io_buffers_pages)) {
279 page = list_first_entry(&ctx->io_buffers_pages, struct page, lru);
280 list_del_init(&page->lru);
285 int io_remove_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
287 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
290 if (sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
294 tmp = READ_ONCE(sqe->fd);
295 if (!tmp || tmp > MAX_BIDS_PER_BGID)
298 memset(p, 0, sizeof(*p));
300 p->bgid = READ_ONCE(sqe->buf_group);
304 int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
306 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
307 struct io_ring_ctx *ctx = req->ctx;
308 struct io_buffer_list *bl;
311 io_ring_submit_lock(ctx, issue_flags);
314 bl = io_buffer_get_list(ctx, p->bgid);
317 /* can't use provide/remove buffers command on mapped buffers */
319 ret = __io_remove_buffers(ctx, bl, p->nbufs);
321 io_ring_submit_unlock(ctx, issue_flags);
324 io_req_set_res(req, ret, 0);
328 int io_provide_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
330 unsigned long size, tmp_check;
331 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
334 if (sqe->rw_flags || sqe->splice_fd_in)
337 tmp = READ_ONCE(sqe->fd);
338 if (!tmp || tmp > MAX_BIDS_PER_BGID)
341 p->addr = READ_ONCE(sqe->addr);
342 p->len = READ_ONCE(sqe->len);
344 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
347 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
350 size = (unsigned long)p->len * p->nbufs;
351 if (!access_ok(u64_to_user_ptr(p->addr), size))
354 p->bgid = READ_ONCE(sqe->buf_group);
355 tmp = READ_ONCE(sqe->off);
358 if (tmp + p->nbufs > MAX_BIDS_PER_BGID)
364 static int io_refill_buffer_cache(struct io_ring_ctx *ctx)
366 struct io_buffer *buf;
371 * Completions that don't happen inline (eg not under uring_lock) will
372 * add to ->io_buffers_comp. If we don't have any free buffers, check
373 * the completion list and splice those entries first.
375 if (!list_empty_careful(&ctx->io_buffers_comp)) {
376 spin_lock(&ctx->completion_lock);
377 if (!list_empty(&ctx->io_buffers_comp)) {
378 list_splice_init(&ctx->io_buffers_comp,
379 &ctx->io_buffers_cache);
380 spin_unlock(&ctx->completion_lock);
383 spin_unlock(&ctx->completion_lock);
387 * No free buffers and no completion entries either. Allocate a new
388 * page worth of buffer entries and add those to our freelist.
390 page = alloc_page(GFP_KERNEL_ACCOUNT);
394 list_add(&page->lru, &ctx->io_buffers_pages);
396 buf = page_address(page);
397 bufs_in_page = PAGE_SIZE / sizeof(*buf);
398 while (bufs_in_page) {
399 list_add_tail(&buf->list, &ctx->io_buffers_cache);
407 static int io_add_buffers(struct io_ring_ctx *ctx, struct io_provide_buf *pbuf,
408 struct io_buffer_list *bl)
410 struct io_buffer *buf;
411 u64 addr = pbuf->addr;
412 int i, bid = pbuf->bid;
414 for (i = 0; i < pbuf->nbufs; i++) {
415 if (list_empty(&ctx->io_buffers_cache) &&
416 io_refill_buffer_cache(ctx))
418 buf = list_first_entry(&ctx->io_buffers_cache, struct io_buffer,
420 list_move_tail(&buf->list, &bl->buf_list);
422 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
424 buf->bgid = pbuf->bgid;
430 return i ? 0 : -ENOMEM;
433 int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
435 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
436 struct io_ring_ctx *ctx = req->ctx;
437 struct io_buffer_list *bl;
440 io_ring_submit_lock(ctx, issue_flags);
442 if (unlikely(p->bgid < BGID_ARRAY && !ctx->io_bl)) {
443 ret = io_init_bl_list(ctx);
448 bl = io_buffer_get_list(ctx, p->bgid);
450 bl = kzalloc(sizeof(*bl), GFP_KERNEL_ACCOUNT);
455 INIT_LIST_HEAD(&bl->buf_list);
456 ret = io_buffer_add_list(ctx, bl, p->bgid);
462 /* can't add buffers via this command for a mapped buffer ring */
468 ret = io_add_buffers(ctx, p, bl);
470 io_ring_submit_unlock(ctx, issue_flags);
474 io_req_set_res(req, ret, 0);
478 static int io_pin_pbuf_ring(struct io_uring_buf_reg *reg,
479 struct io_buffer_list *bl)
481 struct io_uring_buf_ring *br;
485 pages = io_pin_pages(reg->ring_addr,
486 flex_array_size(br, bufs, reg->ring_entries),
489 return PTR_ERR(pages);
492 * Apparently some 32-bit boxes (ARM) will return highmem pages,
493 * which then need to be mapped. We could support that, but it'd
494 * complicate the code and slowdown the common cases quite a bit.
495 * So just error out, returning -EINVAL just like we did on kernels
496 * that didn't support mapped buffer rings.
498 for (i = 0; i < nr_pages; i++)
499 if (PageHighMem(pages[i]))
502 br = page_address(pages[0]);
505 * On platforms that have specific aliasing requirements, SHM_COLOUR
506 * is set and we must guarantee that the kernel and user side align
507 * nicely. We cannot do that if IOU_PBUF_RING_MMAP isn't set and
508 * the application mmap's the provided ring buffer. Fail the request
509 * if we, by chance, don't end up with aligned addresses. The app
510 * should use IOU_PBUF_RING_MMAP instead, and liburing will handle
511 * this transparently.
513 if ((reg->ring_addr | (unsigned long) br) & (SHM_COLOUR - 1))
516 bl->buf_pages = pages;
517 bl->buf_nr_pages = nr_pages;
523 for (i = 0; i < nr_pages; i++)
524 unpin_user_page(pages[i]);
529 static int io_alloc_pbuf_ring(struct io_uring_buf_reg *reg,
530 struct io_buffer_list *bl)
532 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
536 ring_size = reg->ring_entries * sizeof(struct io_uring_buf_ring);
537 ptr = (void *) __get_free_pages(gfp, get_order(ring_size));
547 int io_register_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
549 struct io_uring_buf_reg reg;
550 struct io_buffer_list *bl, *free_bl = NULL;
553 if (copy_from_user(®, arg, sizeof(reg)))
556 if (reg.resv[0] || reg.resv[1] || reg.resv[2])
558 if (reg.flags & ~IOU_PBUF_RING_MMAP)
560 if (!(reg.flags & IOU_PBUF_RING_MMAP)) {
563 if (reg.ring_addr & ~PAGE_MASK)
570 if (!is_power_of_2(reg.ring_entries))
573 /* cannot disambiguate full vs empty due to head/tail size */
574 if (reg.ring_entries >= 65536)
577 if (unlikely(reg.bgid < BGID_ARRAY && !ctx->io_bl)) {
578 int ret = io_init_bl_list(ctx);
583 bl = io_buffer_get_list(ctx, reg.bgid);
585 /* if mapped buffer ring OR classic exists, don't allow */
586 if (bl->is_mapped || !list_empty(&bl->buf_list))
589 free_bl = bl = kzalloc(sizeof(*bl), GFP_KERNEL);
594 if (!(reg.flags & IOU_PBUF_RING_MMAP))
595 ret = io_pin_pbuf_ring(®, bl);
597 ret = io_alloc_pbuf_ring(®, bl);
600 bl->nr_entries = reg.ring_entries;
601 bl->mask = reg.ring_entries - 1;
603 io_buffer_add_list(ctx, bl, reg.bgid);
611 int io_unregister_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
613 struct io_uring_buf_reg reg;
614 struct io_buffer_list *bl;
616 if (copy_from_user(®, arg, sizeof(reg)))
618 if (reg.resv[0] || reg.resv[1] || reg.resv[2])
623 bl = io_buffer_get_list(ctx, reg.bgid);
629 __io_remove_buffers(ctx, bl, -1U);
630 if (bl->bgid >= BGID_ARRAY) {
631 xa_erase(&ctx->io_bl_xa, bl->bgid);
637 void *io_pbuf_get_address(struct io_ring_ctx *ctx, unsigned long bgid)
639 struct io_buffer_list *bl;
641 bl = io_buffer_get_list(ctx, bgid);
642 if (!bl || !bl->is_mmap)