Linux Audio

Check our new training course

Loading...
Note: File does not exist in v5.9.
  1// SPDX-License-Identifier: GPL-2.0
  2#include <linux/kernel.h>
  3#include <linux/errno.h>
  4#include <linux/fs.h>
  5#include <linux/file.h>
  6#include <linux/mm.h>
  7#include <linux/slab.h>
  8#include <linux/namei.h>
  9#include <linux/poll.h>
 10#include <linux/vmalloc.h>
 11#include <linux/io_uring.h>
 12
 13#include <uapi/linux/io_uring.h>
 14
 15#include "io_uring.h"
 16#include "opdef.h"
 17#include "kbuf.h"
 18#include "memmap.h"
 19
 20/* BIDs are addressed by a 16-bit field in a CQE */
 21#define MAX_BIDS_PER_BGID (1 << 16)
 22
 23struct kmem_cache *io_buf_cachep;
 24
 25struct io_provide_buf {
 26	struct file			*file;
 27	__u64				addr;
 28	__u32				len;
 29	__u32				bgid;
 30	__u32				nbufs;
 31	__u16				bid;
 32};
 33
 34static inline struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
 35							unsigned int bgid)
 36{
 37	lockdep_assert_held(&ctx->uring_lock);
 38
 39	return xa_load(&ctx->io_bl_xa, bgid);
 40}
 41
 42static int io_buffer_add_list(struct io_ring_ctx *ctx,
 43			      struct io_buffer_list *bl, unsigned int bgid)
 44{
 45	/*
 46	 * Store buffer group ID and finally mark the list as visible.
 47	 * The normal lookup doesn't care about the visibility as we're
 48	 * always under the ->uring_lock, but the RCU lookup from mmap does.
 49	 */
 50	bl->bgid = bgid;
 51	atomic_set(&bl->refs, 1);
 52	return xa_err(xa_store(&ctx->io_bl_xa, bgid, bl, GFP_KERNEL));
 53}
 54
 55bool io_kbuf_recycle_legacy(struct io_kiocb *req, unsigned issue_flags)
 56{
 57	struct io_ring_ctx *ctx = req->ctx;
 58	struct io_buffer_list *bl;
 59	struct io_buffer *buf;
 60
 61	io_ring_submit_lock(ctx, issue_flags);
 62
 63	buf = req->kbuf;
 64	bl = io_buffer_get_list(ctx, buf->bgid);
 65	list_add(&buf->list, &bl->buf_list);
 66	req->flags &= ~REQ_F_BUFFER_SELECTED;
 67	req->buf_index = buf->bgid;
 68
 69	io_ring_submit_unlock(ctx, issue_flags);
 70	return true;
 71}
 72
 73void __io_put_kbuf(struct io_kiocb *req, int len, unsigned issue_flags)
 74{
 75	/*
 76	 * We can add this buffer back to two lists:
 77	 *
 78	 * 1) The io_buffers_cache list. This one is protected by the
 79	 *    ctx->uring_lock. If we already hold this lock, add back to this
 80	 *    list as we can grab it from issue as well.
 81	 * 2) The io_buffers_comp list. This one is protected by the
 82	 *    ctx->completion_lock.
 83	 *
 84	 * We migrate buffers from the comp_list to the issue cache list
 85	 * when we need one.
 86	 */
 87	if (issue_flags & IO_URING_F_UNLOCKED) {
 88		struct io_ring_ctx *ctx = req->ctx;
 89
 90		spin_lock(&ctx->completion_lock);
 91		__io_put_kbuf_list(req, len, &ctx->io_buffers_comp);
 92		spin_unlock(&ctx->completion_lock);
 93	} else {
 94		lockdep_assert_held(&req->ctx->uring_lock);
 95
 96		__io_put_kbuf_list(req, len, &req->ctx->io_buffers_cache);
 97	}
 98}
 99
100static void __user *io_provided_buffer_select(struct io_kiocb *req, size_t *len,
101					      struct io_buffer_list *bl)
102{
103	if (!list_empty(&bl->buf_list)) {
104		struct io_buffer *kbuf;
105
106		kbuf = list_first_entry(&bl->buf_list, struct io_buffer, list);
107		list_del(&kbuf->list);
108		if (*len == 0 || *len > kbuf->len)
109			*len = kbuf->len;
110		if (list_empty(&bl->buf_list))
111			req->flags |= REQ_F_BL_EMPTY;
112		req->flags |= REQ_F_BUFFER_SELECTED;
113		req->kbuf = kbuf;
114		req->buf_index = kbuf->bid;
115		return u64_to_user_ptr(kbuf->addr);
116	}
117	return NULL;
118}
119
120static int io_provided_buffers_select(struct io_kiocb *req, size_t *len,
121				      struct io_buffer_list *bl,
122				      struct iovec *iov)
123{
124	void __user *buf;
125
126	buf = io_provided_buffer_select(req, len, bl);
127	if (unlikely(!buf))
128		return -ENOBUFS;
129
130	iov[0].iov_base = buf;
131	iov[0].iov_len = *len;
132	return 1;
133}
134
135static void __user *io_ring_buffer_select(struct io_kiocb *req, size_t *len,
136					  struct io_buffer_list *bl,
137					  unsigned int issue_flags)
138{
139	struct io_uring_buf_ring *br = bl->buf_ring;
140	__u16 tail, head = bl->head;
141	struct io_uring_buf *buf;
142	void __user *ret;
143
144	tail = smp_load_acquire(&br->tail);
145	if (unlikely(tail == head))
146		return NULL;
147
148	if (head + 1 == tail)
149		req->flags |= REQ_F_BL_EMPTY;
150
151	buf = io_ring_head_to_buf(br, head, bl->mask);
152	if (*len == 0 || *len > buf->len)
153		*len = buf->len;
154	req->flags |= REQ_F_BUFFER_RING | REQ_F_BUFFERS_COMMIT;
155	req->buf_list = bl;
156	req->buf_index = buf->bid;
157	ret = u64_to_user_ptr(buf->addr);
158
159	if (issue_flags & IO_URING_F_UNLOCKED || !io_file_can_poll(req)) {
160		/*
161		 * If we came in unlocked, we have no choice but to consume the
162		 * buffer here, otherwise nothing ensures that the buffer won't
163		 * get used by others. This does mean it'll be pinned until the
164		 * IO completes, coming in unlocked means we're being called from
165		 * io-wq context and there may be further retries in async hybrid
166		 * mode. For the locked case, the caller must call commit when
167		 * the transfer completes (or if we get -EAGAIN and must poll of
168		 * retry).
169		 */
170		io_kbuf_commit(req, bl, *len, 1);
171		req->buf_list = NULL;
172	}
173	return ret;
174}
175
176void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
177			      unsigned int issue_flags)
178{
179	struct io_ring_ctx *ctx = req->ctx;
180	struct io_buffer_list *bl;
181	void __user *ret = NULL;
182
183	io_ring_submit_lock(req->ctx, issue_flags);
184
185	bl = io_buffer_get_list(ctx, req->buf_index);
186	if (likely(bl)) {
187		if (bl->flags & IOBL_BUF_RING)
188			ret = io_ring_buffer_select(req, len, bl, issue_flags);
189		else
190			ret = io_provided_buffer_select(req, len, bl);
191	}
192	io_ring_submit_unlock(req->ctx, issue_flags);
193	return ret;
194}
195
196/* cap it at a reasonable 256, will be one page even for 4K */
197#define PEEK_MAX_IMPORT		256
198
199static int io_ring_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg,
200				struct io_buffer_list *bl)
201{
202	struct io_uring_buf_ring *br = bl->buf_ring;
203	struct iovec *iov = arg->iovs;
204	int nr_iovs = arg->nr_iovs;
205	__u16 nr_avail, tail, head;
206	struct io_uring_buf *buf;
207
208	tail = smp_load_acquire(&br->tail);
209	head = bl->head;
210	nr_avail = min_t(__u16, tail - head, UIO_MAXIOV);
211	if (unlikely(!nr_avail))
212		return -ENOBUFS;
213
214	buf = io_ring_head_to_buf(br, head, bl->mask);
215	if (arg->max_len) {
216		u32 len = READ_ONCE(buf->len);
217
218		if (unlikely(!len))
219			return -ENOBUFS;
220		/*
221		 * Limit incremental buffers to 1 segment. No point trying
222		 * to peek ahead and map more than we need, when the buffers
223		 * themselves should be large when setup with
224		 * IOU_PBUF_RING_INC.
225		 */
226		if (bl->flags & IOBL_INC) {
227			nr_avail = 1;
228		} else {
229			size_t needed;
230
231			needed = (arg->max_len + len - 1) / len;
232			needed = min_not_zero(needed, (size_t) PEEK_MAX_IMPORT);
233			if (nr_avail > needed)
234				nr_avail = needed;
235		}
236	}
237
238	/*
239	 * only alloc a bigger array if we know we have data to map, eg not
240	 * a speculative peek operation.
241	 */
242	if (arg->mode & KBUF_MODE_EXPAND && nr_avail > nr_iovs && arg->max_len) {
243		iov = kmalloc_array(nr_avail, sizeof(struct iovec), GFP_KERNEL);
244		if (unlikely(!iov))
245			return -ENOMEM;
246		if (arg->mode & KBUF_MODE_FREE)
247			kfree(arg->iovs);
248		arg->iovs = iov;
249		nr_iovs = nr_avail;
250	} else if (nr_avail < nr_iovs) {
251		nr_iovs = nr_avail;
252	}
253
254	/* set it to max, if not set, so we can use it unconditionally */
255	if (!arg->max_len)
256		arg->max_len = INT_MAX;
257
258	req->buf_index = buf->bid;
259	do {
260		u32 len = buf->len;
261
262		/* truncate end piece, if needed, for non partial buffers */
263		if (len > arg->max_len) {
264			len = arg->max_len;
265			if (!(bl->flags & IOBL_INC))
266				buf->len = len;
267		}
268
269		iov->iov_base = u64_to_user_ptr(buf->addr);
270		iov->iov_len = len;
271		iov++;
272
273		arg->out_len += len;
274		arg->max_len -= len;
275		if (!arg->max_len)
276			break;
277
278		buf = io_ring_head_to_buf(br, ++head, bl->mask);
279	} while (--nr_iovs);
280
281	if (head == tail)
282		req->flags |= REQ_F_BL_EMPTY;
283
284	req->flags |= REQ_F_BUFFER_RING;
285	req->buf_list = bl;
286	return iov - arg->iovs;
287}
288
289int io_buffers_select(struct io_kiocb *req, struct buf_sel_arg *arg,
290		      unsigned int issue_flags)
291{
292	struct io_ring_ctx *ctx = req->ctx;
293	struct io_buffer_list *bl;
294	int ret = -ENOENT;
295
296	io_ring_submit_lock(ctx, issue_flags);
297	bl = io_buffer_get_list(ctx, req->buf_index);
298	if (unlikely(!bl))
299		goto out_unlock;
300
301	if (bl->flags & IOBL_BUF_RING) {
302		ret = io_ring_buffers_peek(req, arg, bl);
303		/*
304		 * Don't recycle these buffers if we need to go through poll.
305		 * Nobody else can use them anyway, and holding on to provided
306		 * buffers for a send/write operation would happen on the app
307		 * side anyway with normal buffers. Besides, we already
308		 * committed them, they cannot be put back in the queue.
309		 */
310		if (ret > 0) {
311			req->flags |= REQ_F_BUFFERS_COMMIT | REQ_F_BL_NO_RECYCLE;
312			io_kbuf_commit(req, bl, arg->out_len, ret);
313		}
314	} else {
315		ret = io_provided_buffers_select(req, &arg->out_len, bl, arg->iovs);
316	}
317out_unlock:
318	io_ring_submit_unlock(ctx, issue_flags);
319	return ret;
320}
321
322int io_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg)
323{
324	struct io_ring_ctx *ctx = req->ctx;
325	struct io_buffer_list *bl;
326	int ret;
327
328	lockdep_assert_held(&ctx->uring_lock);
329
330	bl = io_buffer_get_list(ctx, req->buf_index);
331	if (unlikely(!bl))
332		return -ENOENT;
333
334	if (bl->flags & IOBL_BUF_RING) {
335		ret = io_ring_buffers_peek(req, arg, bl);
336		if (ret > 0)
337			req->flags |= REQ_F_BUFFERS_COMMIT;
338		return ret;
339	}
340
341	/* don't support multiple buffer selections for legacy */
342	return io_provided_buffers_select(req, &arg->max_len, bl, arg->iovs);
343}
344
345static int __io_remove_buffers(struct io_ring_ctx *ctx,
346			       struct io_buffer_list *bl, unsigned nbufs)
347{
348	unsigned i = 0;
349
350	/* shouldn't happen */
351	if (!nbufs)
352		return 0;
353
354	if (bl->flags & IOBL_BUF_RING) {
355		i = bl->buf_ring->tail - bl->head;
356		if (bl->buf_nr_pages) {
357			int j;
358
359			if (!(bl->flags & IOBL_MMAP)) {
360				for (j = 0; j < bl->buf_nr_pages; j++)
361					unpin_user_page(bl->buf_pages[j]);
362			}
363			io_pages_unmap(bl->buf_ring, &bl->buf_pages,
364					&bl->buf_nr_pages, bl->flags & IOBL_MMAP);
365			bl->flags &= ~IOBL_MMAP;
366		}
367		/* make sure it's seen as empty */
368		INIT_LIST_HEAD(&bl->buf_list);
369		bl->flags &= ~IOBL_BUF_RING;
370		return i;
371	}
372
373	/* protects io_buffers_cache */
374	lockdep_assert_held(&ctx->uring_lock);
375
376	while (!list_empty(&bl->buf_list)) {
377		struct io_buffer *nxt;
378
379		nxt = list_first_entry(&bl->buf_list, struct io_buffer, list);
380		list_move(&nxt->list, &ctx->io_buffers_cache);
381		if (++i == nbufs)
382			return i;
383		cond_resched();
384	}
385
386	return i;
387}
388
389void io_put_bl(struct io_ring_ctx *ctx, struct io_buffer_list *bl)
390{
391	if (atomic_dec_and_test(&bl->refs)) {
392		__io_remove_buffers(ctx, bl, -1U);
393		kfree_rcu(bl, rcu);
394	}
395}
396
397void io_destroy_buffers(struct io_ring_ctx *ctx)
398{
399	struct io_buffer_list *bl;
400	struct list_head *item, *tmp;
401	struct io_buffer *buf;
402	unsigned long index;
403
404	xa_for_each(&ctx->io_bl_xa, index, bl) {
405		xa_erase(&ctx->io_bl_xa, bl->bgid);
406		io_put_bl(ctx, bl);
407	}
408
409	/*
410	 * Move deferred locked entries to cache before pruning
411	 */
412	spin_lock(&ctx->completion_lock);
413	if (!list_empty(&ctx->io_buffers_comp))
414		list_splice_init(&ctx->io_buffers_comp, &ctx->io_buffers_cache);
415	spin_unlock(&ctx->completion_lock);
416
417	list_for_each_safe(item, tmp, &ctx->io_buffers_cache) {
418		buf = list_entry(item, struct io_buffer, list);
419		kmem_cache_free(io_buf_cachep, buf);
420	}
421}
422
423static void io_destroy_bl(struct io_ring_ctx *ctx, struct io_buffer_list *bl)
424{
425	xa_erase(&ctx->io_bl_xa, bl->bgid);
426	io_put_bl(ctx, bl);
427}
428
429int io_remove_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
430{
431	struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
432	u64 tmp;
433
434	if (sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
435	    sqe->splice_fd_in)
436		return -EINVAL;
437
438	tmp = READ_ONCE(sqe->fd);
439	if (!tmp || tmp > MAX_BIDS_PER_BGID)
440		return -EINVAL;
441
442	memset(p, 0, sizeof(*p));
443	p->nbufs = tmp;
444	p->bgid = READ_ONCE(sqe->buf_group);
445	return 0;
446}
447
448int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
449{
450	struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
451	struct io_ring_ctx *ctx = req->ctx;
452	struct io_buffer_list *bl;
453	int ret = 0;
454
455	io_ring_submit_lock(ctx, issue_flags);
456
457	ret = -ENOENT;
458	bl = io_buffer_get_list(ctx, p->bgid);
459	if (bl) {
460		ret = -EINVAL;
461		/* can't use provide/remove buffers command on mapped buffers */
462		if (!(bl->flags & IOBL_BUF_RING))
463			ret = __io_remove_buffers(ctx, bl, p->nbufs);
464	}
465	io_ring_submit_unlock(ctx, issue_flags);
466	if (ret < 0)
467		req_set_fail(req);
468	io_req_set_res(req, ret, 0);
469	return IOU_OK;
470}
471
472int io_provide_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
473{
474	unsigned long size, tmp_check;
475	struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
476	u64 tmp;
477
478	if (sqe->rw_flags || sqe->splice_fd_in)
479		return -EINVAL;
480
481	tmp = READ_ONCE(sqe->fd);
482	if (!tmp || tmp > MAX_BIDS_PER_BGID)
483		return -E2BIG;
484	p->nbufs = tmp;
485	p->addr = READ_ONCE(sqe->addr);
486	p->len = READ_ONCE(sqe->len);
487
488	if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
489				&size))
490		return -EOVERFLOW;
491	if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
492		return -EOVERFLOW;
493
494	size = (unsigned long)p->len * p->nbufs;
495	if (!access_ok(u64_to_user_ptr(p->addr), size))
496		return -EFAULT;
497
498	p->bgid = READ_ONCE(sqe->buf_group);
499	tmp = READ_ONCE(sqe->off);
500	if (tmp > USHRT_MAX)
501		return -E2BIG;
502	if (tmp + p->nbufs > MAX_BIDS_PER_BGID)
503		return -EINVAL;
504	p->bid = tmp;
505	return 0;
506}
507
508#define IO_BUFFER_ALLOC_BATCH 64
509
510static int io_refill_buffer_cache(struct io_ring_ctx *ctx)
511{
512	struct io_buffer *bufs[IO_BUFFER_ALLOC_BATCH];
513	int allocated;
514
515	/*
516	 * Completions that don't happen inline (eg not under uring_lock) will
517	 * add to ->io_buffers_comp. If we don't have any free buffers, check
518	 * the completion list and splice those entries first.
519	 */
520	if (!list_empty_careful(&ctx->io_buffers_comp)) {
521		spin_lock(&ctx->completion_lock);
522		if (!list_empty(&ctx->io_buffers_comp)) {
523			list_splice_init(&ctx->io_buffers_comp,
524						&ctx->io_buffers_cache);
525			spin_unlock(&ctx->completion_lock);
526			return 0;
527		}
528		spin_unlock(&ctx->completion_lock);
529	}
530
531	/*
532	 * No free buffers and no completion entries either. Allocate a new
533	 * batch of buffer entries and add those to our freelist.
534	 */
535
536	allocated = kmem_cache_alloc_bulk(io_buf_cachep, GFP_KERNEL_ACCOUNT,
537					  ARRAY_SIZE(bufs), (void **) bufs);
538	if (unlikely(!allocated)) {
539		/*
540		 * Bulk alloc is all-or-nothing. If we fail to get a batch,
541		 * retry single alloc to be on the safe side.
542		 */
543		bufs[0] = kmem_cache_alloc(io_buf_cachep, GFP_KERNEL);
544		if (!bufs[0])
545			return -ENOMEM;
546		allocated = 1;
547	}
548
549	while (allocated)
550		list_add_tail(&bufs[--allocated]->list, &ctx->io_buffers_cache);
551
552	return 0;
553}
554
555static int io_add_buffers(struct io_ring_ctx *ctx, struct io_provide_buf *pbuf,
556			  struct io_buffer_list *bl)
557{
558	struct io_buffer *buf;
559	u64 addr = pbuf->addr;
560	int i, bid = pbuf->bid;
561
562	for (i = 0; i < pbuf->nbufs; i++) {
563		if (list_empty(&ctx->io_buffers_cache) &&
564		    io_refill_buffer_cache(ctx))
565			break;
566		buf = list_first_entry(&ctx->io_buffers_cache, struct io_buffer,
567					list);
568		list_move_tail(&buf->list, &bl->buf_list);
569		buf->addr = addr;
570		buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
571		buf->bid = bid;
572		buf->bgid = pbuf->bgid;
573		addr += pbuf->len;
574		bid++;
575		cond_resched();
576	}
577
578	return i ? 0 : -ENOMEM;
579}
580
581int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
582{
583	struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
584	struct io_ring_ctx *ctx = req->ctx;
585	struct io_buffer_list *bl;
586	int ret = 0;
587
588	io_ring_submit_lock(ctx, issue_flags);
589
590	bl = io_buffer_get_list(ctx, p->bgid);
591	if (unlikely(!bl)) {
592		bl = kzalloc(sizeof(*bl), GFP_KERNEL_ACCOUNT);
593		if (!bl) {
594			ret = -ENOMEM;
595			goto err;
596		}
597		INIT_LIST_HEAD(&bl->buf_list);
598		ret = io_buffer_add_list(ctx, bl, p->bgid);
599		if (ret) {
600			/*
601			 * Doesn't need rcu free as it was never visible, but
602			 * let's keep it consistent throughout.
603			 */
604			kfree_rcu(bl, rcu);
605			goto err;
606		}
607	}
608	/* can't add buffers via this command for a mapped buffer ring */
609	if (bl->flags & IOBL_BUF_RING) {
610		ret = -EINVAL;
611		goto err;
612	}
613
614	ret = io_add_buffers(ctx, p, bl);
615err:
616	io_ring_submit_unlock(ctx, issue_flags);
617
618	if (ret < 0)
619		req_set_fail(req);
620	io_req_set_res(req, ret, 0);
621	return IOU_OK;
622}
623
624static int io_pin_pbuf_ring(struct io_uring_buf_reg *reg,
625			    struct io_buffer_list *bl)
626{
627	struct io_uring_buf_ring *br = NULL;
628	struct page **pages;
629	int nr_pages, ret;
630
631	pages = io_pin_pages(reg->ring_addr,
632			     flex_array_size(br, bufs, reg->ring_entries),
633			     &nr_pages);
634	if (IS_ERR(pages))
635		return PTR_ERR(pages);
636
637	br = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
638	if (!br) {
639		ret = -ENOMEM;
640		goto error_unpin;
641	}
642
643#ifdef SHM_COLOUR
644	/*
645	 * On platforms that have specific aliasing requirements, SHM_COLOUR
646	 * is set and we must guarantee that the kernel and user side align
647	 * nicely. We cannot do that if IOU_PBUF_RING_MMAP isn't set and
648	 * the application mmap's the provided ring buffer. Fail the request
649	 * if we, by chance, don't end up with aligned addresses. The app
650	 * should use IOU_PBUF_RING_MMAP instead, and liburing will handle
651	 * this transparently.
652	 */
653	if ((reg->ring_addr | (unsigned long) br) & (SHM_COLOUR - 1)) {
654		ret = -EINVAL;
655		goto error_unpin;
656	}
657#endif
658	bl->buf_pages = pages;
659	bl->buf_nr_pages = nr_pages;
660	bl->buf_ring = br;
661	bl->flags |= IOBL_BUF_RING;
662	bl->flags &= ~IOBL_MMAP;
663	return 0;
664error_unpin:
665	unpin_user_pages(pages, nr_pages);
666	kvfree(pages);
667	vunmap(br);
668	return ret;
669}
670
671static int io_alloc_pbuf_ring(struct io_ring_ctx *ctx,
672			      struct io_uring_buf_reg *reg,
673			      struct io_buffer_list *bl)
674{
675	size_t ring_size;
676
677	ring_size = reg->ring_entries * sizeof(struct io_uring_buf_ring);
678
679	bl->buf_ring = io_pages_map(&bl->buf_pages, &bl->buf_nr_pages, ring_size);
680	if (IS_ERR(bl->buf_ring)) {
681		bl->buf_ring = NULL;
682		return -ENOMEM;
683	}
684
685	bl->flags |= (IOBL_BUF_RING | IOBL_MMAP);
686	return 0;
687}
688
689int io_register_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
690{
691	struct io_uring_buf_reg reg;
692	struct io_buffer_list *bl, *free_bl = NULL;
693	int ret;
694
695	lockdep_assert_held(&ctx->uring_lock);
696
697	if (copy_from_user(&reg, arg, sizeof(reg)))
698		return -EFAULT;
699
700	if (reg.resv[0] || reg.resv[1] || reg.resv[2])
701		return -EINVAL;
702	if (reg.flags & ~(IOU_PBUF_RING_MMAP | IOU_PBUF_RING_INC))
703		return -EINVAL;
704	if (!(reg.flags & IOU_PBUF_RING_MMAP)) {
705		if (!reg.ring_addr)
706			return -EFAULT;
707		if (reg.ring_addr & ~PAGE_MASK)
708			return -EINVAL;
709	} else {
710		if (reg.ring_addr)
711			return -EINVAL;
712	}
713
714	if (!is_power_of_2(reg.ring_entries))
715		return -EINVAL;
716
717	/* cannot disambiguate full vs empty due to head/tail size */
718	if (reg.ring_entries >= 65536)
719		return -EINVAL;
720
721	bl = io_buffer_get_list(ctx, reg.bgid);
722	if (bl) {
723		/* if mapped buffer ring OR classic exists, don't allow */
724		if (bl->flags & IOBL_BUF_RING || !list_empty(&bl->buf_list))
725			return -EEXIST;
726		io_destroy_bl(ctx, bl);
727	}
728
729	free_bl = bl = kzalloc(sizeof(*bl), GFP_KERNEL);
730	if (!bl)
731		return -ENOMEM;
732
733	if (!(reg.flags & IOU_PBUF_RING_MMAP))
734		ret = io_pin_pbuf_ring(&reg, bl);
735	else
736		ret = io_alloc_pbuf_ring(ctx, &reg, bl);
737
738	if (!ret) {
739		bl->nr_entries = reg.ring_entries;
740		bl->mask = reg.ring_entries - 1;
741		if (reg.flags & IOU_PBUF_RING_INC)
742			bl->flags |= IOBL_INC;
743
744		io_buffer_add_list(ctx, bl, reg.bgid);
745		return 0;
746	}
747
748	kfree_rcu(free_bl, rcu);
749	return ret;
750}
751
752int io_unregister_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
753{
754	struct io_uring_buf_reg reg;
755	struct io_buffer_list *bl;
756
757	lockdep_assert_held(&ctx->uring_lock);
758
759	if (copy_from_user(&reg, arg, sizeof(reg)))
760		return -EFAULT;
761	if (reg.resv[0] || reg.resv[1] || reg.resv[2])
762		return -EINVAL;
763	if (reg.flags)
764		return -EINVAL;
765
766	bl = io_buffer_get_list(ctx, reg.bgid);
767	if (!bl)
768		return -ENOENT;
769	if (!(bl->flags & IOBL_BUF_RING))
770		return -EINVAL;
771
772	xa_erase(&ctx->io_bl_xa, bl->bgid);
773	io_put_bl(ctx, bl);
774	return 0;
775}
776
777int io_register_pbuf_status(struct io_ring_ctx *ctx, void __user *arg)
778{
779	struct io_uring_buf_status buf_status;
780	struct io_buffer_list *bl;
781	int i;
782
783	if (copy_from_user(&buf_status, arg, sizeof(buf_status)))
784		return -EFAULT;
785
786	for (i = 0; i < ARRAY_SIZE(buf_status.resv); i++)
787		if (buf_status.resv[i])
788			return -EINVAL;
789
790	bl = io_buffer_get_list(ctx, buf_status.buf_group);
791	if (!bl)
792		return -ENOENT;
793	if (!(bl->flags & IOBL_BUF_RING))
794		return -EINVAL;
795
796	buf_status.head = bl->head;
797	if (copy_to_user(arg, &buf_status, sizeof(buf_status)))
798		return -EFAULT;
799
800	return 0;
801}
802
803struct io_buffer_list *io_pbuf_get_bl(struct io_ring_ctx *ctx,
804				      unsigned long bgid)
805{
806	struct io_buffer_list *bl;
807	bool ret;
808
809	/*
810	 * We have to be a bit careful here - we're inside mmap and cannot grab
811	 * the uring_lock. This means the buffer_list could be simultaneously
812	 * going away, if someone is trying to be sneaky. Look it up under rcu
813	 * so we know it's not going away, and attempt to grab a reference to
814	 * it. If the ref is already zero, then fail the mapping. If successful,
815	 * the caller will call io_put_bl() to drop the the reference at at the
816	 * end. This may then safely free the buffer_list (and drop the pages)
817	 * at that point, vm_insert_pages() would've already grabbed the
818	 * necessary vma references.
819	 */
820	rcu_read_lock();
821	bl = xa_load(&ctx->io_bl_xa, bgid);
822	/* must be a mmap'able buffer ring and have pages */
823	ret = false;
824	if (bl && bl->flags & IOBL_MMAP)
825		ret = atomic_inc_not_zero(&bl->refs);
826	rcu_read_unlock();
827
828	if (ret)
829		return bl;
830
831	return ERR_PTR(-EINVAL);
832}
833
834int io_pbuf_mmap(struct file *file, struct vm_area_struct *vma)
835{
836	struct io_ring_ctx *ctx = file->private_data;
837	loff_t pgoff = vma->vm_pgoff << PAGE_SHIFT;
838	struct io_buffer_list *bl;
839	int bgid, ret;
840
841	bgid = (pgoff & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
842	bl = io_pbuf_get_bl(ctx, bgid);
843	if (IS_ERR(bl))
844		return PTR_ERR(bl);
845
846	ret = io_uring_mmap_pages(ctx, vma, bl->buf_pages, bl->buf_nr_pages);
847	io_put_bl(ctx, bl);
848	return ret;
849}