Linux Audio

Check our new training course

Loading...
   1/*
   2 * "splice": joining two ropes together by interweaving their strands.
   3 *
   4 * This is the "extended pipe" functionality, where a pipe is used as
   5 * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6 * buffer that you can use to transfer data from one end to the other.
   7 *
   8 * The traditional unix read/write is extended with a "splice()" operation
   9 * that transfers data buffers to or from a pipe buffer.
  10 *
  11 * Named by Larry McVoy, original implementation from Linus, extended by
  12 * Jens to support splicing to files, network, direct splicing, etc and
  13 * fixing lots of bugs.
  14 *
  15 * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  16 * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17 * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18 *
  19 */
  20#include <linux/bvec.h>
  21#include <linux/fs.h>
  22#include <linux/file.h>
  23#include <linux/pagemap.h>
  24#include <linux/splice.h>
  25#include <linux/memcontrol.h>
  26#include <linux/mm_inline.h>
  27#include <linux/swap.h>
  28#include <linux/writeback.h>
  29#include <linux/export.h>
  30#include <linux/syscalls.h>
  31#include <linux/uio.h>
  32#include <linux/security.h>
  33#include <linux/gfp.h>
  34#include <linux/socket.h>
  35#include <linux/compat.h>
  36#include <linux/sched/signal.h>
  37
  38#include "internal.h"
  39
  40/*
  41 * Attempt to steal a page from a pipe buffer. This should perhaps go into
  42 * a vm helper function, it's already simplified quite a bit by the
  43 * addition of remove_mapping(). If success is returned, the caller may
  44 * attempt to reuse this page for another destination.
  45 */
  46static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  47				     struct pipe_buffer *buf)
  48{
  49	struct page *page = buf->page;
  50	struct address_space *mapping;
  51
  52	lock_page(page);
  53
  54	mapping = page_mapping(page);
  55	if (mapping) {
  56		WARN_ON(!PageUptodate(page));
  57
  58		/*
  59		 * At least for ext2 with nobh option, we need to wait on
  60		 * writeback completing on this page, since we'll remove it
  61		 * from the pagecache.  Otherwise truncate wont wait on the
  62		 * page, allowing the disk blocks to be reused by someone else
  63		 * before we actually wrote our data to them. fs corruption
  64		 * ensues.
  65		 */
  66		wait_on_page_writeback(page);
  67
  68		if (page_has_private(page) &&
  69		    !try_to_release_page(page, GFP_KERNEL))
  70			goto out_unlock;
  71
  72		/*
  73		 * If we succeeded in removing the mapping, set LRU flag
  74		 * and return good.
  75		 */
  76		if (remove_mapping(mapping, page)) {
  77			buf->flags |= PIPE_BUF_FLAG_LRU;
  78			return 0;
  79		}
  80	}
  81
  82	/*
  83	 * Raced with truncate or failed to remove page from current
  84	 * address space, unlock and return failure.
  85	 */
  86out_unlock:
  87	unlock_page(page);
  88	return 1;
  89}
  90
  91static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  92					struct pipe_buffer *buf)
  93{
  94	put_page(buf->page);
  95	buf->flags &= ~PIPE_BUF_FLAG_LRU;
  96}
  97
  98/*
  99 * Check whether the contents of buf is OK to access. Since the content
 100 * is a page cache page, IO may be in flight.
 101 */
 102static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 103				       struct pipe_buffer *buf)
 104{
 105	struct page *page = buf->page;
 106	int err;
 107
 108	if (!PageUptodate(page)) {
 109		lock_page(page);
 110
 111		/*
 112		 * Page got truncated/unhashed. This will cause a 0-byte
 113		 * splice, if this is the first page.
 114		 */
 115		if (!page->mapping) {
 116			err = -ENODATA;
 117			goto error;
 118		}
 119
 120		/*
 121		 * Uh oh, read-error from disk.
 122		 */
 123		if (!PageUptodate(page)) {
 124			err = -EIO;
 125			goto error;
 126		}
 127
 128		/*
 129		 * Page is ok afterall, we are done.
 130		 */
 131		unlock_page(page);
 132	}
 133
 134	return 0;
 135error:
 136	unlock_page(page);
 137	return err;
 138}
 139
 140const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 141	.can_merge = 0,
 142	.confirm = page_cache_pipe_buf_confirm,
 143	.release = page_cache_pipe_buf_release,
 144	.steal = page_cache_pipe_buf_steal,
 145	.get = generic_pipe_buf_get,
 146};
 147
 148static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 149				    struct pipe_buffer *buf)
 150{
 151	if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 152		return 1;
 153
 154	buf->flags |= PIPE_BUF_FLAG_LRU;
 155	return generic_pipe_buf_steal(pipe, buf);
 156}
 157
 158static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 159	.can_merge = 0,
 160	.confirm = generic_pipe_buf_confirm,
 161	.release = page_cache_pipe_buf_release,
 162	.steal = user_page_pipe_buf_steal,
 163	.get = generic_pipe_buf_get,
 164};
 165
 166static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 167{
 168	smp_mb();
 169	if (waitqueue_active(&pipe->wait))
 170		wake_up_interruptible(&pipe->wait);
 171	kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 172}
 173
 174/**
 175 * splice_to_pipe - fill passed data into a pipe
 176 * @pipe:	pipe to fill
 177 * @spd:	data to fill
 178 *
 179 * Description:
 180 *    @spd contains a map of pages and len/offset tuples, along with
 181 *    the struct pipe_buf_operations associated with these pages. This
 182 *    function will link that data to the pipe.
 183 *
 184 */
 185ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 186		       struct splice_pipe_desc *spd)
 187{
 188	unsigned int spd_pages = spd->nr_pages;
 189	int ret = 0, page_nr = 0;
 190
 191	if (!spd_pages)
 192		return 0;
 193
 194	if (unlikely(!pipe->readers)) {
 195		send_sig(SIGPIPE, current, 0);
 196		ret = -EPIPE;
 197		goto out;
 198	}
 199
 200	while (pipe->nrbufs < pipe->buffers) {
 201		int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 202		struct pipe_buffer *buf = pipe->bufs + newbuf;
 203
 204		buf->page = spd->pages[page_nr];
 205		buf->offset = spd->partial[page_nr].offset;
 206		buf->len = spd->partial[page_nr].len;
 207		buf->private = spd->partial[page_nr].private;
 208		buf->ops = spd->ops;
 209		buf->flags = 0;
 210
 211		pipe->nrbufs++;
 212		page_nr++;
 213		ret += buf->len;
 214
 215		if (!--spd->nr_pages)
 216			break;
 217	}
 218
 219	if (!ret)
 220		ret = -EAGAIN;
 221
 222out:
 223	while (page_nr < spd_pages)
 224		spd->spd_release(spd, page_nr++);
 225
 226	return ret;
 227}
 228EXPORT_SYMBOL_GPL(splice_to_pipe);
 229
 230ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 231{
 232	int ret;
 233
 234	if (unlikely(!pipe->readers)) {
 235		send_sig(SIGPIPE, current, 0);
 236		ret = -EPIPE;
 237	} else if (pipe->nrbufs == pipe->buffers) {
 238		ret = -EAGAIN;
 239	} else {
 240		int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 241		pipe->bufs[newbuf] = *buf;
 242		pipe->nrbufs++;
 243		return buf->len;
 244	}
 245	pipe_buf_release(pipe, buf);
 246	return ret;
 247}
 248EXPORT_SYMBOL(add_to_pipe);
 249
 250/*
 251 * Check if we need to grow the arrays holding pages and partial page
 252 * descriptions.
 253 */
 254int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 255{
 256	unsigned int buffers = READ_ONCE(pipe->buffers);
 257
 258	spd->nr_pages_max = buffers;
 259	if (buffers <= PIPE_DEF_BUFFERS)
 260		return 0;
 261
 262	spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 263	spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 264
 265	if (spd->pages && spd->partial)
 266		return 0;
 267
 268	kfree(spd->pages);
 269	kfree(spd->partial);
 270	return -ENOMEM;
 271}
 272
 273void splice_shrink_spd(struct splice_pipe_desc *spd)
 274{
 275	if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 276		return;
 277
 278	kfree(spd->pages);
 279	kfree(spd->partial);
 280}
 281
 282/**
 283 * generic_file_splice_read - splice data from file to a pipe
 284 * @in:		file to splice from
 285 * @ppos:	position in @in
 286 * @pipe:	pipe to splice to
 287 * @len:	number of bytes to splice
 288 * @flags:	splice modifier flags
 289 *
 290 * Description:
 291 *    Will read pages from given file and fill them into a pipe. Can be
 292 *    used as long as it has more or less sane ->read_iter().
 293 *
 294 */
 295ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 296				 struct pipe_inode_info *pipe, size_t len,
 297				 unsigned int flags)
 298{
 299	struct iov_iter to;
 300	struct kiocb kiocb;
 301	int idx, ret;
 302
 303	iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len);
 304	idx = to.idx;
 305	init_sync_kiocb(&kiocb, in);
 306	kiocb.ki_pos = *ppos;
 307	ret = call_read_iter(in, &kiocb, &to);
 308	if (ret > 0) {
 309		*ppos = kiocb.ki_pos;
 310		file_accessed(in);
 311	} else if (ret < 0) {
 312		to.idx = idx;
 313		to.iov_offset = 0;
 314		iov_iter_advance(&to, 0); /* to free what was emitted */
 315		/*
 316		 * callers of ->splice_read() expect -EAGAIN on
 317		 * "can't put anything in there", rather than -EFAULT.
 318		 */
 319		if (ret == -EFAULT)
 320			ret = -EAGAIN;
 321	}
 322
 323	return ret;
 324}
 325EXPORT_SYMBOL(generic_file_splice_read);
 326
 327const struct pipe_buf_operations default_pipe_buf_ops = {
 328	.can_merge = 0,
 329	.confirm = generic_pipe_buf_confirm,
 330	.release = generic_pipe_buf_release,
 331	.steal = generic_pipe_buf_steal,
 332	.get = generic_pipe_buf_get,
 333};
 334
 335static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 336				    struct pipe_buffer *buf)
 337{
 338	return 1;
 339}
 340
 341/* Pipe buffer operations for a socket and similar. */
 342const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 343	.can_merge = 0,
 344	.confirm = generic_pipe_buf_confirm,
 345	.release = generic_pipe_buf_release,
 346	.steal = generic_pipe_buf_nosteal,
 347	.get = generic_pipe_buf_get,
 348};
 349EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 350
 351static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 352			    unsigned long vlen, loff_t offset)
 353{
 354	mm_segment_t old_fs;
 355	loff_t pos = offset;
 356	ssize_t res;
 357
 358	old_fs = get_fs();
 359	set_fs(get_ds());
 360	/* The cast to a user pointer is valid due to the set_fs() */
 361	res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 362	set_fs(old_fs);
 363
 364	return res;
 365}
 366
 367static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 368				 struct pipe_inode_info *pipe, size_t len,
 369				 unsigned int flags)
 370{
 371	struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 372	struct iov_iter to;
 373	struct page **pages;
 374	unsigned int nr_pages;
 375	size_t offset, base, copied = 0;
 376	ssize_t res;
 377	int i;
 378
 379	if (pipe->nrbufs == pipe->buffers)
 380		return -EAGAIN;
 381
 382	/*
 383	 * Try to keep page boundaries matching to source pagecache ones -
 384	 * it probably won't be much help, but...
 385	 */
 386	offset = *ppos & ~PAGE_MASK;
 387
 388	iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
 389
 390	res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &base);
 391	if (res <= 0)
 392		return -ENOMEM;
 393
 394	nr_pages = DIV_ROUND_UP(res + base, PAGE_SIZE);
 395
 396	vec = __vec;
 397	if (nr_pages > PIPE_DEF_BUFFERS) {
 398		vec = kmalloc(nr_pages * sizeof(struct kvec), GFP_KERNEL);
 399		if (unlikely(!vec)) {
 400			res = -ENOMEM;
 401			goto out;
 402		}
 403	}
 404
 405	pipe->bufs[to.idx].offset = offset;
 406	pipe->bufs[to.idx].len -= offset;
 407
 408	for (i = 0; i < nr_pages; i++) {
 409		size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 410		vec[i].iov_base = page_address(pages[i]) + offset;
 411		vec[i].iov_len = this_len;
 412		len -= this_len;
 413		offset = 0;
 414	}
 415
 416	res = kernel_readv(in, vec, nr_pages, *ppos);
 417	if (res > 0) {
 418		copied = res;
 419		*ppos += res;
 420	}
 421
 422	if (vec != __vec)
 423		kfree(vec);
 424out:
 425	for (i = 0; i < nr_pages; i++)
 426		put_page(pages[i]);
 427	kvfree(pages);
 428	iov_iter_advance(&to, copied);	/* truncates and discards */
 429	return res;
 430}
 431
 432/*
 433 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 434 * using sendpage(). Return the number of bytes sent.
 435 */
 436static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 437			    struct pipe_buffer *buf, struct splice_desc *sd)
 438{
 439	struct file *file = sd->u.file;
 440	loff_t pos = sd->pos;
 441	int more;
 442
 443	if (!likely(file->f_op->sendpage))
 444		return -EINVAL;
 445
 446	more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 447
 448	if (sd->len < sd->total_len && pipe->nrbufs > 1)
 449		more |= MSG_SENDPAGE_NOTLAST;
 450
 451	return file->f_op->sendpage(file, buf->page, buf->offset,
 452				    sd->len, &pos, more);
 453}
 454
 455static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 456{
 457	smp_mb();
 458	if (waitqueue_active(&pipe->wait))
 459		wake_up_interruptible(&pipe->wait);
 460	kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 461}
 462
 463/**
 464 * splice_from_pipe_feed - feed available data from a pipe to a file
 465 * @pipe:	pipe to splice from
 466 * @sd:		information to @actor
 467 * @actor:	handler that splices the data
 468 *
 469 * Description:
 470 *    This function loops over the pipe and calls @actor to do the
 471 *    actual moving of a single struct pipe_buffer to the desired
 472 *    destination.  It returns when there's no more buffers left in
 473 *    the pipe or if the requested number of bytes (@sd->total_len)
 474 *    have been copied.  It returns a positive number (one) if the
 475 *    pipe needs to be filled with more data, zero if the required
 476 *    number of bytes have been copied and -errno on error.
 477 *
 478 *    This, together with splice_from_pipe_{begin,end,next}, may be
 479 *    used to implement the functionality of __splice_from_pipe() when
 480 *    locking is required around copying the pipe buffers to the
 481 *    destination.
 482 */
 483static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 484			  splice_actor *actor)
 485{
 486	int ret;
 487
 488	while (pipe->nrbufs) {
 489		struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 490
 491		sd->len = buf->len;
 492		if (sd->len > sd->total_len)
 493			sd->len = sd->total_len;
 494
 495		ret = pipe_buf_confirm(pipe, buf);
 496		if (unlikely(ret)) {
 497			if (ret == -ENODATA)
 498				ret = 0;
 499			return ret;
 500		}
 501
 502		ret = actor(pipe, buf, sd);
 503		if (ret <= 0)
 504			return ret;
 505
 506		buf->offset += ret;
 507		buf->len -= ret;
 508
 509		sd->num_spliced += ret;
 510		sd->len -= ret;
 511		sd->pos += ret;
 512		sd->total_len -= ret;
 513
 514		if (!buf->len) {
 515			pipe_buf_release(pipe, buf);
 516			pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 517			pipe->nrbufs--;
 518			if (pipe->files)
 519				sd->need_wakeup = true;
 520		}
 521
 522		if (!sd->total_len)
 523			return 0;
 524	}
 525
 526	return 1;
 527}
 528
 529/**
 530 * splice_from_pipe_next - wait for some data to splice from
 531 * @pipe:	pipe to splice from
 532 * @sd:		information about the splice operation
 533 *
 534 * Description:
 535 *    This function will wait for some data and return a positive
 536 *    value (one) if pipe buffers are available.  It will return zero
 537 *    or -errno if no more data needs to be spliced.
 538 */
 539static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 540{
 541	/*
 542	 * Check for signal early to make process killable when there are
 543	 * always buffers available
 544	 */
 545	if (signal_pending(current))
 546		return -ERESTARTSYS;
 547
 548	while (!pipe->nrbufs) {
 549		if (!pipe->writers)
 550			return 0;
 551
 552		if (!pipe->waiting_writers && sd->num_spliced)
 553			return 0;
 554
 555		if (sd->flags & SPLICE_F_NONBLOCK)
 556			return -EAGAIN;
 557
 558		if (signal_pending(current))
 559			return -ERESTARTSYS;
 560
 561		if (sd->need_wakeup) {
 562			wakeup_pipe_writers(pipe);
 563			sd->need_wakeup = false;
 564		}
 565
 566		pipe_wait(pipe);
 567	}
 568
 569	return 1;
 570}
 571
 572/**
 573 * splice_from_pipe_begin - start splicing from pipe
 574 * @sd:		information about the splice operation
 575 *
 576 * Description:
 577 *    This function should be called before a loop containing
 578 *    splice_from_pipe_next() and splice_from_pipe_feed() to
 579 *    initialize the necessary fields of @sd.
 580 */
 581static void splice_from_pipe_begin(struct splice_desc *sd)
 582{
 583	sd->num_spliced = 0;
 584	sd->need_wakeup = false;
 585}
 586
 587/**
 588 * splice_from_pipe_end - finish splicing from pipe
 589 * @pipe:	pipe to splice from
 590 * @sd:		information about the splice operation
 591 *
 592 * Description:
 593 *    This function will wake up pipe writers if necessary.  It should
 594 *    be called after a loop containing splice_from_pipe_next() and
 595 *    splice_from_pipe_feed().
 596 */
 597static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 598{
 599	if (sd->need_wakeup)
 600		wakeup_pipe_writers(pipe);
 601}
 602
 603/**
 604 * __splice_from_pipe - splice data from a pipe to given actor
 605 * @pipe:	pipe to splice from
 606 * @sd:		information to @actor
 607 * @actor:	handler that splices the data
 608 *
 609 * Description:
 610 *    This function does little more than loop over the pipe and call
 611 *    @actor to do the actual moving of a single struct pipe_buffer to
 612 *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 613 *    pipe_to_user.
 614 *
 615 */
 616ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 617			   splice_actor *actor)
 618{
 619	int ret;
 620
 621	splice_from_pipe_begin(sd);
 622	do {
 623		cond_resched();
 624		ret = splice_from_pipe_next(pipe, sd);
 625		if (ret > 0)
 626			ret = splice_from_pipe_feed(pipe, sd, actor);
 627	} while (ret > 0);
 628	splice_from_pipe_end(pipe, sd);
 629
 630	return sd->num_spliced ? sd->num_spliced : ret;
 631}
 632EXPORT_SYMBOL(__splice_from_pipe);
 633
 634/**
 635 * splice_from_pipe - splice data from a pipe to a file
 636 * @pipe:	pipe to splice from
 637 * @out:	file to splice to
 638 * @ppos:	position in @out
 639 * @len:	how many bytes to splice
 640 * @flags:	splice modifier flags
 641 * @actor:	handler that splices the data
 642 *
 643 * Description:
 644 *    See __splice_from_pipe. This function locks the pipe inode,
 645 *    otherwise it's identical to __splice_from_pipe().
 646 *
 647 */
 648ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 649			 loff_t *ppos, size_t len, unsigned int flags,
 650			 splice_actor *actor)
 651{
 652	ssize_t ret;
 653	struct splice_desc sd = {
 654		.total_len = len,
 655		.flags = flags,
 656		.pos = *ppos,
 657		.u.file = out,
 658	};
 659
 660	pipe_lock(pipe);
 661	ret = __splice_from_pipe(pipe, &sd, actor);
 662	pipe_unlock(pipe);
 663
 664	return ret;
 665}
 666
 667/**
 668 * iter_file_splice_write - splice data from a pipe to a file
 669 * @pipe:	pipe info
 670 * @out:	file to write to
 671 * @ppos:	position in @out
 672 * @len:	number of bytes to splice
 673 * @flags:	splice modifier flags
 674 *
 675 * Description:
 676 *    Will either move or copy pages (determined by @flags options) from
 677 *    the given pipe inode to the given file.
 678 *    This one is ->write_iter-based.
 679 *
 680 */
 681ssize_t
 682iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 683			  loff_t *ppos, size_t len, unsigned int flags)
 684{
 685	struct splice_desc sd = {
 686		.total_len = len,
 687		.flags = flags,
 688		.pos = *ppos,
 689		.u.file = out,
 690	};
 691	int nbufs = pipe->buffers;
 692	struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 693					GFP_KERNEL);
 694	ssize_t ret;
 695
 696	if (unlikely(!array))
 697		return -ENOMEM;
 698
 699	pipe_lock(pipe);
 700
 701	splice_from_pipe_begin(&sd);
 702	while (sd.total_len) {
 703		struct iov_iter from;
 704		size_t left;
 705		int n, idx;
 706
 707		ret = splice_from_pipe_next(pipe, &sd);
 708		if (ret <= 0)
 709			break;
 710
 711		if (unlikely(nbufs < pipe->buffers)) {
 712			kfree(array);
 713			nbufs = pipe->buffers;
 714			array = kcalloc(nbufs, sizeof(struct bio_vec),
 715					GFP_KERNEL);
 716			if (!array) {
 717				ret = -ENOMEM;
 718				break;
 719			}
 720		}
 721
 722		/* build the vector */
 723		left = sd.total_len;
 724		for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 725			struct pipe_buffer *buf = pipe->bufs + idx;
 726			size_t this_len = buf->len;
 727
 728			if (this_len > left)
 729				this_len = left;
 730
 731			if (idx == pipe->buffers - 1)
 732				idx = -1;
 733
 734			ret = pipe_buf_confirm(pipe, buf);
 735			if (unlikely(ret)) {
 736				if (ret == -ENODATA)
 737					ret = 0;
 738				goto done;
 739			}
 740
 741			array[n].bv_page = buf->page;
 742			array[n].bv_len = this_len;
 743			array[n].bv_offset = buf->offset;
 744			left -= this_len;
 745		}
 746
 747		iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
 748			      sd.total_len - left);
 749		ret = vfs_iter_write(out, &from, &sd.pos, 0);
 750		if (ret <= 0)
 751			break;
 752
 753		sd.num_spliced += ret;
 754		sd.total_len -= ret;
 755		*ppos = sd.pos;
 756
 757		/* dismiss the fully eaten buffers, adjust the partial one */
 758		while (ret) {
 759			struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 760			if (ret >= buf->len) {
 761				ret -= buf->len;
 762				buf->len = 0;
 763				pipe_buf_release(pipe, buf);
 764				pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 765				pipe->nrbufs--;
 766				if (pipe->files)
 767					sd.need_wakeup = true;
 768			} else {
 769				buf->offset += ret;
 770				buf->len -= ret;
 771				ret = 0;
 772			}
 773		}
 774	}
 775done:
 776	kfree(array);
 777	splice_from_pipe_end(pipe, &sd);
 778
 779	pipe_unlock(pipe);
 780
 781	if (sd.num_spliced)
 782		ret = sd.num_spliced;
 783
 784	return ret;
 785}
 786
 787EXPORT_SYMBOL(iter_file_splice_write);
 788
 789static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 790			  struct splice_desc *sd)
 791{
 792	int ret;
 793	void *data;
 794	loff_t tmp = sd->pos;
 795
 796	data = kmap(buf->page);
 797	ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 798	kunmap(buf->page);
 799
 800	return ret;
 801}
 802
 803static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 804					 struct file *out, loff_t *ppos,
 805					 size_t len, unsigned int flags)
 806{
 807	ssize_t ret;
 808
 809	ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 810	if (ret > 0)
 811		*ppos += ret;
 812
 813	return ret;
 814}
 815
 816/**
 817 * generic_splice_sendpage - splice data from a pipe to a socket
 818 * @pipe:	pipe to splice from
 819 * @out:	socket to write to
 820 * @ppos:	position in @out
 821 * @len:	number of bytes to splice
 822 * @flags:	splice modifier flags
 823 *
 824 * Description:
 825 *    Will send @len bytes from the pipe to a network socket. No data copying
 826 *    is involved.
 827 *
 828 */
 829ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 830				loff_t *ppos, size_t len, unsigned int flags)
 831{
 832	return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 833}
 834
 835EXPORT_SYMBOL(generic_splice_sendpage);
 836
 837/*
 838 * Attempt to initiate a splice from pipe to file.
 839 */
 840static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 841			   loff_t *ppos, size_t len, unsigned int flags)
 842{
 843	ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 844				loff_t *, size_t, unsigned int);
 845
 846	if (out->f_op->splice_write)
 847		splice_write = out->f_op->splice_write;
 848	else
 849		splice_write = default_file_splice_write;
 850
 851	return splice_write(pipe, out, ppos, len, flags);
 852}
 853
 854/*
 855 * Attempt to initiate a splice from a file to a pipe.
 856 */
 857static long do_splice_to(struct file *in, loff_t *ppos,
 858			 struct pipe_inode_info *pipe, size_t len,
 859			 unsigned int flags)
 860{
 861	ssize_t (*splice_read)(struct file *, loff_t *,
 862			       struct pipe_inode_info *, size_t, unsigned int);
 863	int ret;
 864
 865	if (unlikely(!(in->f_mode & FMODE_READ)))
 866		return -EBADF;
 867
 868	ret = rw_verify_area(READ, in, ppos, len);
 869	if (unlikely(ret < 0))
 870		return ret;
 871
 872	if (unlikely(len > MAX_RW_COUNT))
 873		len = MAX_RW_COUNT;
 874
 875	if (in->f_op->splice_read)
 876		splice_read = in->f_op->splice_read;
 877	else
 878		splice_read = default_file_splice_read;
 879
 880	return splice_read(in, ppos, pipe, len, flags);
 881}
 882
 883/**
 884 * splice_direct_to_actor - splices data directly between two non-pipes
 885 * @in:		file to splice from
 886 * @sd:		actor information on where to splice to
 887 * @actor:	handles the data splicing
 888 *
 889 * Description:
 890 *    This is a special case helper to splice directly between two
 891 *    points, without requiring an explicit pipe. Internally an allocated
 892 *    pipe is cached in the process, and reused during the lifetime of
 893 *    that process.
 894 *
 895 */
 896ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 897			       splice_direct_actor *actor)
 898{
 899	struct pipe_inode_info *pipe;
 900	long ret, bytes;
 901	umode_t i_mode;
 902	size_t len;
 903	int i, flags, more;
 904
 905	/*
 906	 * We require the input being a regular file, as we don't want to
 907	 * randomly drop data for eg socket -> socket splicing. Use the
 908	 * piped splicing for that!
 909	 */
 910	i_mode = file_inode(in)->i_mode;
 911	if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 912		return -EINVAL;
 913
 914	/*
 915	 * neither in nor out is a pipe, setup an internal pipe attached to
 916	 * 'out' and transfer the wanted data from 'in' to 'out' through that
 917	 */
 918	pipe = current->splice_pipe;
 919	if (unlikely(!pipe)) {
 920		pipe = alloc_pipe_info();
 921		if (!pipe)
 922			return -ENOMEM;
 923
 924		/*
 925		 * We don't have an immediate reader, but we'll read the stuff
 926		 * out of the pipe right after the splice_to_pipe(). So set
 927		 * PIPE_READERS appropriately.
 928		 */
 929		pipe->readers = 1;
 930
 931		current->splice_pipe = pipe;
 932	}
 933
 934	/*
 935	 * Do the splice.
 936	 */
 937	ret = 0;
 938	bytes = 0;
 939	len = sd->total_len;
 940	flags = sd->flags;
 941
 942	/*
 943	 * Don't block on output, we have to drain the direct pipe.
 944	 */
 945	sd->flags &= ~SPLICE_F_NONBLOCK;
 946	more = sd->flags & SPLICE_F_MORE;
 947
 948	while (len) {
 949		size_t read_len;
 950		loff_t pos = sd->pos, prev_pos = pos;
 951
 952		ret = do_splice_to(in, &pos, pipe, len, flags);
 953		if (unlikely(ret <= 0))
 954			goto out_release;
 955
 956		read_len = ret;
 957		sd->total_len = read_len;
 958
 959		/*
 960		 * If more data is pending, set SPLICE_F_MORE
 961		 * If this is the last data and SPLICE_F_MORE was not set
 962		 * initially, clears it.
 963		 */
 964		if (read_len < len)
 965			sd->flags |= SPLICE_F_MORE;
 966		else if (!more)
 967			sd->flags &= ~SPLICE_F_MORE;
 968		/*
 969		 * NOTE: nonblocking mode only applies to the input. We
 970		 * must not do the output in nonblocking mode as then we
 971		 * could get stuck data in the internal pipe:
 972		 */
 973		ret = actor(pipe, sd);
 974		if (unlikely(ret <= 0)) {
 975			sd->pos = prev_pos;
 976			goto out_release;
 977		}
 978
 979		bytes += ret;
 980		len -= ret;
 981		sd->pos = pos;
 982
 983		if (ret < read_len) {
 984			sd->pos = prev_pos + ret;
 985			goto out_release;
 986		}
 987	}
 988
 989done:
 990	pipe->nrbufs = pipe->curbuf = 0;
 991	file_accessed(in);
 992	return bytes;
 993
 994out_release:
 995	/*
 996	 * If we did an incomplete transfer we must release
 997	 * the pipe buffers in question:
 998	 */
 999	for (i = 0; i < pipe->buffers; i++) {
1000		struct pipe_buffer *buf = pipe->bufs + i;
1001
1002		if (buf->ops)
1003			pipe_buf_release(pipe, buf);
1004	}
1005
1006	if (!bytes)
1007		bytes = ret;
1008
1009	goto done;
1010}
1011EXPORT_SYMBOL(splice_direct_to_actor);
1012
1013static int direct_splice_actor(struct pipe_inode_info *pipe,
1014			       struct splice_desc *sd)
1015{
1016	struct file *file = sd->u.file;
1017
1018	return do_splice_from(pipe, file, sd->opos, sd->total_len,
1019			      sd->flags);
1020}
1021
1022/**
1023 * do_splice_direct - splices data directly between two files
1024 * @in:		file to splice from
1025 * @ppos:	input file offset
1026 * @out:	file to splice to
1027 * @opos:	output file offset
1028 * @len:	number of bytes to splice
1029 * @flags:	splice modifier flags
1030 *
1031 * Description:
1032 *    For use by do_sendfile(). splice can easily emulate sendfile, but
1033 *    doing it in the application would incur an extra system call
1034 *    (splice in + splice out, as compared to just sendfile()). So this helper
1035 *    can splice directly through a process-private pipe.
1036 *
1037 */
1038long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1039		      loff_t *opos, size_t len, unsigned int flags)
1040{
1041	struct splice_desc sd = {
1042		.len		= len,
1043		.total_len	= len,
1044		.flags		= flags,
1045		.pos		= *ppos,
1046		.u.file		= out,
1047		.opos		= opos,
1048	};
1049	long ret;
1050
1051	if (unlikely(!(out->f_mode & FMODE_WRITE)))
1052		return -EBADF;
1053
1054	if (unlikely(out->f_flags & O_APPEND))
1055		return -EINVAL;
1056
1057	ret = rw_verify_area(WRITE, out, opos, len);
1058	if (unlikely(ret < 0))
1059		return ret;
1060
1061	ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1062	if (ret > 0)
1063		*ppos = sd.pos;
1064
1065	return ret;
1066}
1067EXPORT_SYMBOL(do_splice_direct);
1068
1069static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1070{
1071	for (;;) {
1072		if (unlikely(!pipe->readers)) {
1073			send_sig(SIGPIPE, current, 0);
1074			return -EPIPE;
1075		}
1076		if (pipe->nrbufs != pipe->buffers)
1077			return 0;
1078		if (flags & SPLICE_F_NONBLOCK)
1079			return -EAGAIN;
1080		if (signal_pending(current))
1081			return -ERESTARTSYS;
1082		pipe->waiting_writers++;
1083		pipe_wait(pipe);
1084		pipe->waiting_writers--;
1085	}
1086}
1087
1088static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1089			       struct pipe_inode_info *opipe,
1090			       size_t len, unsigned int flags);
1091
1092/*
1093 * Determine where to splice to/from.
1094 */
1095static long do_splice(struct file *in, loff_t __user *off_in,
1096		      struct file *out, loff_t __user *off_out,
1097		      size_t len, unsigned int flags)
1098{
1099	struct pipe_inode_info *ipipe;
1100	struct pipe_inode_info *opipe;
1101	loff_t offset;
1102	long ret;
1103
1104	ipipe = get_pipe_info(in);
1105	opipe = get_pipe_info(out);
1106
1107	if (ipipe && opipe) {
1108		if (off_in || off_out)
1109			return -ESPIPE;
1110
1111		if (!(in->f_mode & FMODE_READ))
1112			return -EBADF;
1113
1114		if (!(out->f_mode & FMODE_WRITE))
1115			return -EBADF;
1116
1117		/* Splicing to self would be fun, but... */
1118		if (ipipe == opipe)
1119			return -EINVAL;
1120
1121		return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1122	}
1123
1124	if (ipipe) {
1125		if (off_in)
1126			return -ESPIPE;
1127		if (off_out) {
1128			if (!(out->f_mode & FMODE_PWRITE))
1129				return -EINVAL;
1130			if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1131				return -EFAULT;
1132		} else {
1133			offset = out->f_pos;
1134		}
1135
1136		if (unlikely(!(out->f_mode & FMODE_WRITE)))
1137			return -EBADF;
1138
1139		if (unlikely(out->f_flags & O_APPEND))
1140			return -EINVAL;
1141
1142		ret = rw_verify_area(WRITE, out, &offset, len);
1143		if (unlikely(ret < 0))
1144			return ret;
1145
1146		file_start_write(out);
1147		ret = do_splice_from(ipipe, out, &offset, len, flags);
1148		file_end_write(out);
1149
1150		if (!off_out)
1151			out->f_pos = offset;
1152		else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1153			ret = -EFAULT;
1154
1155		return ret;
1156	}
1157
1158	if (opipe) {
1159		if (off_out)
1160			return -ESPIPE;
1161		if (off_in) {
1162			if (!(in->f_mode & FMODE_PREAD))
1163				return -EINVAL;
1164			if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1165				return -EFAULT;
1166		} else {
1167			offset = in->f_pos;
1168		}
1169
1170		pipe_lock(opipe);
1171		ret = wait_for_space(opipe, flags);
1172		if (!ret)
1173			ret = do_splice_to(in, &offset, opipe, len, flags);
1174		pipe_unlock(opipe);
1175		if (ret > 0)
1176			wakeup_pipe_readers(opipe);
1177		if (!off_in)
1178			in->f_pos = offset;
1179		else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1180			ret = -EFAULT;
1181
1182		return ret;
1183	}
1184
1185	return -EINVAL;
1186}
1187
1188static int iter_to_pipe(struct iov_iter *from,
1189			struct pipe_inode_info *pipe,
1190			unsigned flags)
1191{
1192	struct pipe_buffer buf = {
1193		.ops = &user_page_pipe_buf_ops,
1194		.flags = flags
1195	};
1196	size_t total = 0;
1197	int ret = 0;
1198	bool failed = false;
1199
1200	while (iov_iter_count(from) && !failed) {
1201		struct page *pages[16];
1202		ssize_t copied;
1203		size_t start;
1204		int n;
1205
1206		copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1207		if (copied <= 0) {
1208			ret = copied;
1209			break;
1210		}
1211
1212		for (n = 0; copied; n++, start = 0) {
1213			int size = min_t(int, copied, PAGE_SIZE - start);
1214			if (!failed) {
1215				buf.page = pages[n];
1216				buf.offset = start;
1217				buf.len = size;
1218				ret = add_to_pipe(pipe, &buf);
1219				if (unlikely(ret < 0)) {
1220					failed = true;
1221				} else {
1222					iov_iter_advance(from, ret);
1223					total += ret;
1224				}
1225			} else {
1226				put_page(pages[n]);
1227			}
1228			copied -= size;
1229		}
1230	}
1231	return total ? total : ret;
1232}
1233
1234static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1235			struct splice_desc *sd)
1236{
1237	int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1238	return n == sd->len ? n : -EFAULT;
1239}
1240
1241/*
1242 * For lack of a better implementation, implement vmsplice() to userspace
1243 * as a simple copy of the pipes pages to the user iov.
1244 */
1245static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1246			     unsigned long nr_segs, unsigned int flags)
1247{
1248	struct pipe_inode_info *pipe;
1249	struct splice_desc sd;
1250	long ret;
1251	struct iovec iovstack[UIO_FASTIOV];
1252	struct iovec *iov = iovstack;
1253	struct iov_iter iter;
1254
1255	pipe = get_pipe_info(file);
1256	if (!pipe)
1257		return -EBADF;
1258
1259	ret = import_iovec(READ, uiov, nr_segs,
1260			   ARRAY_SIZE(iovstack), &iov, &iter);
1261	if (ret < 0)
1262		return ret;
1263
1264	sd.total_len = iov_iter_count(&iter);
1265	sd.len = 0;
1266	sd.flags = flags;
1267	sd.u.data = &iter;
1268	sd.pos = 0;
1269
1270	if (sd.total_len) {
1271		pipe_lock(pipe);
1272		ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1273		pipe_unlock(pipe);
1274	}
1275
1276	kfree(iov);
1277	return ret;
1278}
1279
1280/*
1281 * vmsplice splices a user address range into a pipe. It can be thought of
1282 * as splice-from-memory, where the regular splice is splice-from-file (or
1283 * to file). In both cases the output is a pipe, naturally.
1284 */
1285static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
1286			     unsigned long nr_segs, unsigned int flags)
1287{
1288	struct pipe_inode_info *pipe;
1289	struct iovec iovstack[UIO_FASTIOV];
1290	struct iovec *iov = iovstack;
1291	struct iov_iter from;
1292	long ret;
1293	unsigned buf_flag = 0;
1294
1295	if (flags & SPLICE_F_GIFT)
1296		buf_flag = PIPE_BUF_FLAG_GIFT;
1297
1298	pipe = get_pipe_info(file);
1299	if (!pipe)
1300		return -EBADF;
1301
1302	ret = import_iovec(WRITE, uiov, nr_segs,
1303			   ARRAY_SIZE(iovstack), &iov, &from);
1304	if (ret < 0)
1305		return ret;
1306
1307	pipe_lock(pipe);
1308	ret = wait_for_space(pipe, flags);
1309	if (!ret)
1310		ret = iter_to_pipe(&from, pipe, buf_flag);
1311	pipe_unlock(pipe);
1312	if (ret > 0)
1313		wakeup_pipe_readers(pipe);
1314	kfree(iov);
1315	return ret;
1316}
1317
1318/*
1319 * Note that vmsplice only really supports true splicing _from_ user memory
1320 * to a pipe, not the other way around. Splicing from user memory is a simple
1321 * operation that can be supported without any funky alignment restrictions
1322 * or nasty vm tricks. We simply map in the user memory and fill them into
1323 * a pipe. The reverse isn't quite as easy, though. There are two possible
1324 * solutions for that:
1325 *
1326 *	- memcpy() the data internally, at which point we might as well just
1327 *	  do a regular read() on the buffer anyway.
1328 *	- Lots of nasty vm tricks, that are neither fast nor flexible (it
1329 *	  has restriction limitations on both ends of the pipe).
1330 *
1331 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1332 *
1333 */
1334static long do_vmsplice(int fd, const struct iovec __user *iov,
1335			unsigned long nr_segs, unsigned int flags)
1336{
1337	struct fd f;
1338	long error;
1339
1340	if (unlikely(flags & ~SPLICE_F_ALL))
1341		return -EINVAL;
1342	if (unlikely(nr_segs > UIO_MAXIOV))
1343		return -EINVAL;
1344	else if (unlikely(!nr_segs))
1345		return 0;
1346
1347	error = -EBADF;
1348	f = fdget(fd);
1349	if (f.file) {
1350		if (f.file->f_mode & FMODE_WRITE)
1351			error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1352		else if (f.file->f_mode & FMODE_READ)
1353			error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1354
1355		fdput(f);
1356	}
1357
1358	return error;
1359}
1360
1361SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1362		unsigned long, nr_segs, unsigned int, flags)
1363{
1364	return do_vmsplice(fd, iov, nr_segs, flags);
1365}
1366
1367#ifdef CONFIG_COMPAT
1368COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1369		    unsigned int, nr_segs, unsigned int, flags)
1370{
1371	unsigned i;
1372	struct iovec __user *iov;
1373	if (nr_segs > UIO_MAXIOV)
1374		return -EINVAL;
1375	iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1376	for (i = 0; i < nr_segs; i++) {
1377		struct compat_iovec v;
1378		if (get_user(v.iov_base, &iov32[i].iov_base) ||
1379		    get_user(v.iov_len, &iov32[i].iov_len) ||
1380		    put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1381		    put_user(v.iov_len, &iov[i].iov_len))
1382			return -EFAULT;
1383	}
1384	return do_vmsplice(fd, iov, nr_segs, flags);
1385}
1386#endif
1387
1388SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1389		int, fd_out, loff_t __user *, off_out,
1390		size_t, len, unsigned int, flags)
1391{
1392	struct fd in, out;
1393	long error;
1394
1395	if (unlikely(!len))
1396		return 0;
1397
1398	if (unlikely(flags & ~SPLICE_F_ALL))
1399		return -EINVAL;
1400
1401	error = -EBADF;
1402	in = fdget(fd_in);
1403	if (in.file) {
1404		if (in.file->f_mode & FMODE_READ) {
1405			out = fdget(fd_out);
1406			if (out.file) {
1407				if (out.file->f_mode & FMODE_WRITE)
1408					error = do_splice(in.file, off_in,
1409							  out.file, off_out,
1410							  len, flags);
1411				fdput(out);
1412			}
1413		}
1414		fdput(in);
1415	}
1416	return error;
1417}
1418
1419/*
1420 * Make sure there's data to read. Wait for input if we can, otherwise
1421 * return an appropriate error.
1422 */
1423static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1424{
1425	int ret;
1426
1427	/*
1428	 * Check ->nrbufs without the inode lock first. This function
1429	 * is speculative anyways, so missing one is ok.
1430	 */
1431	if (pipe->nrbufs)
1432		return 0;
1433
1434	ret = 0;
1435	pipe_lock(pipe);
1436
1437	while (!pipe->nrbufs) {
1438		if (signal_pending(current)) {
1439			ret = -ERESTARTSYS;
1440			break;
1441		}
1442		if (!pipe->writers)
1443			break;
1444		if (!pipe->waiting_writers) {
1445			if (flags & SPLICE_F_NONBLOCK) {
1446				ret = -EAGAIN;
1447				break;
1448			}
1449		}
1450		pipe_wait(pipe);
1451	}
1452
1453	pipe_unlock(pipe);
1454	return ret;
1455}
1456
1457/*
1458 * Make sure there's writeable room. Wait for room if we can, otherwise
1459 * return an appropriate error.
1460 */
1461static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1462{
1463	int ret;
1464
1465	/*
1466	 * Check ->nrbufs without the inode lock first. This function
1467	 * is speculative anyways, so missing one is ok.
1468	 */
1469	if (pipe->nrbufs < pipe->buffers)
1470		return 0;
1471
1472	ret = 0;
1473	pipe_lock(pipe);
1474
1475	while (pipe->nrbufs >= pipe->buffers) {
1476		if (!pipe->readers) {
1477			send_sig(SIGPIPE, current, 0);
1478			ret = -EPIPE;
1479			break;
1480		}
1481		if (flags & SPLICE_F_NONBLOCK) {
1482			ret = -EAGAIN;
1483			break;
1484		}
1485		if (signal_pending(current)) {
1486			ret = -ERESTARTSYS;
1487			break;
1488		}
1489		pipe->waiting_writers++;
1490		pipe_wait(pipe);
1491		pipe->waiting_writers--;
1492	}
1493
1494	pipe_unlock(pipe);
1495	return ret;
1496}
1497
1498/*
1499 * Splice contents of ipipe to opipe.
1500 */
1501static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1502			       struct pipe_inode_info *opipe,
1503			       size_t len, unsigned int flags)
1504{
1505	struct pipe_buffer *ibuf, *obuf;
1506	int ret = 0, nbuf;
1507	bool input_wakeup = false;
1508
1509
1510retry:
1511	ret = ipipe_prep(ipipe, flags);
1512	if (ret)
1513		return ret;
1514
1515	ret = opipe_prep(opipe, flags);
1516	if (ret)
1517		return ret;
1518
1519	/*
1520	 * Potential ABBA deadlock, work around it by ordering lock
1521	 * grabbing by pipe info address. Otherwise two different processes
1522	 * could deadlock (one doing tee from A -> B, the other from B -> A).
1523	 */
1524	pipe_double_lock(ipipe, opipe);
1525
1526	do {
1527		if (!opipe->readers) {
1528			send_sig(SIGPIPE, current, 0);
1529			if (!ret)
1530				ret = -EPIPE;
1531			break;
1532		}
1533
1534		if (!ipipe->nrbufs && !ipipe->writers)
1535			break;
1536
1537		/*
1538		 * Cannot make any progress, because either the input
1539		 * pipe is empty or the output pipe is full.
1540		 */
1541		if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1542			/* Already processed some buffers, break */
1543			if (ret)
1544				break;
1545
1546			if (flags & SPLICE_F_NONBLOCK) {
1547				ret = -EAGAIN;
1548				break;
1549			}
1550
1551			/*
1552			 * We raced with another reader/writer and haven't
1553			 * managed to process any buffers.  A zero return
1554			 * value means EOF, so retry instead.
1555			 */
1556			pipe_unlock(ipipe);
1557			pipe_unlock(opipe);
1558			goto retry;
1559		}
1560
1561		ibuf = ipipe->bufs + ipipe->curbuf;
1562		nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1563		obuf = opipe->bufs + nbuf;
1564
1565		if (len >= ibuf->len) {
1566			/*
1567			 * Simply move the whole buffer from ipipe to opipe
1568			 */
1569			*obuf = *ibuf;
1570			ibuf->ops = NULL;
1571			opipe->nrbufs++;
1572			ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1573			ipipe->nrbufs--;
1574			input_wakeup = true;
1575		} else {
1576			/*
1577			 * Get a reference to this pipe buffer,
1578			 * so we can copy the contents over.
1579			 */
1580			pipe_buf_get(ipipe, ibuf);
1581			*obuf = *ibuf;
1582
1583			/*
1584			 * Don't inherit the gift flag, we need to
1585			 * prevent multiple steals of this page.
1586			 */
1587			obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1588
1589			obuf->len = len;
1590			opipe->nrbufs++;
1591			ibuf->offset += obuf->len;
1592			ibuf->len -= obuf->len;
1593		}
1594		ret += obuf->len;
1595		len -= obuf->len;
1596	} while (len);
1597
1598	pipe_unlock(ipipe);
1599	pipe_unlock(opipe);
1600
1601	/*
1602	 * If we put data in the output pipe, wakeup any potential readers.
1603	 */
1604	if (ret > 0)
1605		wakeup_pipe_readers(opipe);
1606
1607	if (input_wakeup)
1608		wakeup_pipe_writers(ipipe);
1609
1610	return ret;
1611}
1612
1613/*
1614 * Link contents of ipipe to opipe.
1615 */
1616static int link_pipe(struct pipe_inode_info *ipipe,
1617		     struct pipe_inode_info *opipe,
1618		     size_t len, unsigned int flags)
1619{
1620	struct pipe_buffer *ibuf, *obuf;
1621	int ret = 0, i = 0, nbuf;
1622
1623	/*
1624	 * Potential ABBA deadlock, work around it by ordering lock
1625	 * grabbing by pipe info address. Otherwise two different processes
1626	 * could deadlock (one doing tee from A -> B, the other from B -> A).
1627	 */
1628	pipe_double_lock(ipipe, opipe);
1629
1630	do {
1631		if (!opipe->readers) {
1632			send_sig(SIGPIPE, current, 0);
1633			if (!ret)
1634				ret = -EPIPE;
1635			break;
1636		}
1637
1638		/*
1639		 * If we have iterated all input buffers or ran out of
1640		 * output room, break.
1641		 */
1642		if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1643			break;
1644
1645		ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1646		nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1647
1648		/*
1649		 * Get a reference to this pipe buffer,
1650		 * so we can copy the contents over.
1651		 */
1652		pipe_buf_get(ipipe, ibuf);
1653
1654		obuf = opipe->bufs + nbuf;
1655		*obuf = *ibuf;
1656
1657		/*
1658		 * Don't inherit the gift flag, we need to
1659		 * prevent multiple steals of this page.
1660		 */
1661		obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1662
1663		if (obuf->len > len)
1664			obuf->len = len;
1665
1666		opipe->nrbufs++;
1667		ret += obuf->len;
1668		len -= obuf->len;
1669		i++;
1670	} while (len);
1671
1672	/*
1673	 * return EAGAIN if we have the potential of some data in the
1674	 * future, otherwise just return 0
1675	 */
1676	if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1677		ret = -EAGAIN;
1678
1679	pipe_unlock(ipipe);
1680	pipe_unlock(opipe);
1681
1682	/*
1683	 * If we put data in the output pipe, wakeup any potential readers.
1684	 */
1685	if (ret > 0)
1686		wakeup_pipe_readers(opipe);
1687
1688	return ret;
1689}
1690
1691/*
1692 * This is a tee(1) implementation that works on pipes. It doesn't copy
1693 * any data, it simply references the 'in' pages on the 'out' pipe.
1694 * The 'flags' used are the SPLICE_F_* variants, currently the only
1695 * applicable one is SPLICE_F_NONBLOCK.
1696 */
1697static long do_tee(struct file *in, struct file *out, size_t len,
1698		   unsigned int flags)
1699{
1700	struct pipe_inode_info *ipipe = get_pipe_info(in);
1701	struct pipe_inode_info *opipe = get_pipe_info(out);
1702	int ret = -EINVAL;
1703
1704	/*
1705	 * Duplicate the contents of ipipe to opipe without actually
1706	 * copying the data.
1707	 */
1708	if (ipipe && opipe && ipipe != opipe) {
1709		/*
1710		 * Keep going, unless we encounter an error. The ipipe/opipe
1711		 * ordering doesn't really matter.
1712		 */
1713		ret = ipipe_prep(ipipe, flags);
1714		if (!ret) {
1715			ret = opipe_prep(opipe, flags);
1716			if (!ret)
1717				ret = link_pipe(ipipe, opipe, len, flags);
1718		}
1719	}
1720
1721	return ret;
1722}
1723
1724SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1725{
1726	struct fd in;
1727	int error;
1728
1729	if (unlikely(flags & ~SPLICE_F_ALL))
1730		return -EINVAL;
1731
1732	if (unlikely(!len))
1733		return 0;
1734
1735	error = -EBADF;
1736	in = fdget(fdin);
1737	if (in.file) {
1738		if (in.file->f_mode & FMODE_READ) {
1739			struct fd out = fdget(fdout);
1740			if (out.file) {
1741				if (out.file->f_mode & FMODE_WRITE)
1742					error = do_tee(in.file, out.file,
1743							len, flags);
1744				fdput(out);
1745			}
1746		}
1747 		fdput(in);
1748 	}
1749
1750	return error;
1751}
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * "splice": joining two ropes together by interweaving their strands.
   4 *
   5 * This is the "extended pipe" functionality, where a pipe is used as
   6 * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   7 * buffer that you can use to transfer data from one end to the other.
   8 *
   9 * The traditional unix read/write is extended with a "splice()" operation
  10 * that transfers data buffers to or from a pipe buffer.
  11 *
  12 * Named by Larry McVoy, original implementation from Linus, extended by
  13 * Jens to support splicing to files, network, direct splicing, etc and
  14 * fixing lots of bugs.
  15 *
  16 * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  17 * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  18 * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  19 *
  20 */
  21#include <linux/bvec.h>
  22#include <linux/fs.h>
  23#include <linux/file.h>
  24#include <linux/pagemap.h>
  25#include <linux/splice.h>
  26#include <linux/memcontrol.h>
  27#include <linux/mm_inline.h>
  28#include <linux/swap.h>
  29#include <linux/writeback.h>
  30#include <linux/export.h>
  31#include <linux/syscalls.h>
  32#include <linux/uio.h>
  33#include <linux/security.h>
  34#include <linux/gfp.h>
  35#include <linux/socket.h>
  36#include <linux/sched/signal.h>
  37
  38#include "internal.h"
  39
  40/*
  41 * Attempt to steal a page from a pipe buffer. This should perhaps go into
  42 * a vm helper function, it's already simplified quite a bit by the
  43 * addition of remove_mapping(). If success is returned, the caller may
  44 * attempt to reuse this page for another destination.
  45 */
  46static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
  47		struct pipe_buffer *buf)
  48{
  49	struct page *page = buf->page;
  50	struct address_space *mapping;
  51
  52	lock_page(page);
  53
  54	mapping = page_mapping(page);
  55	if (mapping) {
  56		WARN_ON(!PageUptodate(page));
  57
  58		/*
  59		 * At least for ext2 with nobh option, we need to wait on
  60		 * writeback completing on this page, since we'll remove it
  61		 * from the pagecache.  Otherwise truncate wont wait on the
  62		 * page, allowing the disk blocks to be reused by someone else
  63		 * before we actually wrote our data to them. fs corruption
  64		 * ensues.
  65		 */
  66		wait_on_page_writeback(page);
  67
  68		if (page_has_private(page) &&
  69		    !try_to_release_page(page, GFP_KERNEL))
  70			goto out_unlock;
  71
  72		/*
  73		 * If we succeeded in removing the mapping, set LRU flag
  74		 * and return good.
  75		 */
  76		if (remove_mapping(mapping, page)) {
  77			buf->flags |= PIPE_BUF_FLAG_LRU;
  78			return true;
  79		}
  80	}
  81
  82	/*
  83	 * Raced with truncate or failed to remove page from current
  84	 * address space, unlock and return failure.
  85	 */
  86out_unlock:
  87	unlock_page(page);
  88	return false;
  89}
  90
  91static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  92					struct pipe_buffer *buf)
  93{
  94	put_page(buf->page);
  95	buf->flags &= ~PIPE_BUF_FLAG_LRU;
  96}
  97
  98/*
  99 * Check whether the contents of buf is OK to access. Since the content
 100 * is a page cache page, IO may be in flight.
 101 */
 102static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 103				       struct pipe_buffer *buf)
 104{
 105	struct page *page = buf->page;
 106	int err;
 107
 108	if (!PageUptodate(page)) {
 109		lock_page(page);
 110
 111		/*
 112		 * Page got truncated/unhashed. This will cause a 0-byte
 113		 * splice, if this is the first page.
 114		 */
 115		if (!page->mapping) {
 116			err = -ENODATA;
 117			goto error;
 118		}
 119
 120		/*
 121		 * Uh oh, read-error from disk.
 122		 */
 123		if (!PageUptodate(page)) {
 124			err = -EIO;
 125			goto error;
 126		}
 127
 128		/*
 129		 * Page is ok afterall, we are done.
 130		 */
 131		unlock_page(page);
 132	}
 133
 134	return 0;
 135error:
 136	unlock_page(page);
 137	return err;
 138}
 139
 140const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 141	.confirm	= page_cache_pipe_buf_confirm,
 142	.release	= page_cache_pipe_buf_release,
 143	.try_steal	= page_cache_pipe_buf_try_steal,
 144	.get		= generic_pipe_buf_get,
 145};
 146
 147static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 148		struct pipe_buffer *buf)
 149{
 150	if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 151		return false;
 152
 153	buf->flags |= PIPE_BUF_FLAG_LRU;
 154	return generic_pipe_buf_try_steal(pipe, buf);
 155}
 156
 157static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 158	.release	= page_cache_pipe_buf_release,
 159	.try_steal	= user_page_pipe_buf_try_steal,
 160	.get		= generic_pipe_buf_get,
 161};
 162
 163static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 164{
 165	smp_mb();
 166	if (waitqueue_active(&pipe->rd_wait))
 167		wake_up_interruptible(&pipe->rd_wait);
 168	kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 169}
 170
 171/**
 172 * splice_to_pipe - fill passed data into a pipe
 173 * @pipe:	pipe to fill
 174 * @spd:	data to fill
 175 *
 176 * Description:
 177 *    @spd contains a map of pages and len/offset tuples, along with
 178 *    the struct pipe_buf_operations associated with these pages. This
 179 *    function will link that data to the pipe.
 180 *
 181 */
 182ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 183		       struct splice_pipe_desc *spd)
 184{
 185	unsigned int spd_pages = spd->nr_pages;
 186	unsigned int tail = pipe->tail;
 187	unsigned int head = pipe->head;
 188	unsigned int mask = pipe->ring_size - 1;
 189	int ret = 0, page_nr = 0;
 190
 191	if (!spd_pages)
 192		return 0;
 193
 194	if (unlikely(!pipe->readers)) {
 195		send_sig(SIGPIPE, current, 0);
 196		ret = -EPIPE;
 197		goto out;
 198	}
 199
 200	while (!pipe_full(head, tail, pipe->max_usage)) {
 201		struct pipe_buffer *buf = &pipe->bufs[head & mask];
 202
 203		buf->page = spd->pages[page_nr];
 204		buf->offset = spd->partial[page_nr].offset;
 205		buf->len = spd->partial[page_nr].len;
 206		buf->private = spd->partial[page_nr].private;
 207		buf->ops = spd->ops;
 208		buf->flags = 0;
 209
 210		head++;
 211		pipe->head = head;
 212		page_nr++;
 213		ret += buf->len;
 214
 215		if (!--spd->nr_pages)
 216			break;
 217	}
 218
 219	if (!ret)
 220		ret = -EAGAIN;
 221
 222out:
 223	while (page_nr < spd_pages)
 224		spd->spd_release(spd, page_nr++);
 225
 226	return ret;
 227}
 228EXPORT_SYMBOL_GPL(splice_to_pipe);
 229
 230ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 231{
 232	unsigned int head = pipe->head;
 233	unsigned int tail = pipe->tail;
 234	unsigned int mask = pipe->ring_size - 1;
 235	int ret;
 236
 237	if (unlikely(!pipe->readers)) {
 238		send_sig(SIGPIPE, current, 0);
 239		ret = -EPIPE;
 240	} else if (pipe_full(head, tail, pipe->max_usage)) {
 241		ret = -EAGAIN;
 242	} else {
 243		pipe->bufs[head & mask] = *buf;
 244		pipe->head = head + 1;
 245		return buf->len;
 246	}
 247	pipe_buf_release(pipe, buf);
 248	return ret;
 249}
 250EXPORT_SYMBOL(add_to_pipe);
 251
 252/*
 253 * Check if we need to grow the arrays holding pages and partial page
 254 * descriptions.
 255 */
 256int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 257{
 258	unsigned int max_usage = READ_ONCE(pipe->max_usage);
 259
 260	spd->nr_pages_max = max_usage;
 261	if (max_usage <= PIPE_DEF_BUFFERS)
 262		return 0;
 263
 264	spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 265	spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 266				     GFP_KERNEL);
 267
 268	if (spd->pages && spd->partial)
 269		return 0;
 270
 271	kfree(spd->pages);
 272	kfree(spd->partial);
 273	return -ENOMEM;
 274}
 275
 276void splice_shrink_spd(struct splice_pipe_desc *spd)
 277{
 278	if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 279		return;
 280
 281	kfree(spd->pages);
 282	kfree(spd->partial);
 283}
 284
 285/**
 286 * generic_file_splice_read - splice data from file to a pipe
 287 * @in:		file to splice from
 288 * @ppos:	position in @in
 289 * @pipe:	pipe to splice to
 290 * @len:	number of bytes to splice
 291 * @flags:	splice modifier flags
 292 *
 293 * Description:
 294 *    Will read pages from given file and fill them into a pipe. Can be
 295 *    used as long as it has more or less sane ->read_iter().
 296 *
 297 */
 298ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 299				 struct pipe_inode_info *pipe, size_t len,
 300				 unsigned int flags)
 301{
 302	struct iov_iter to;
 303	struct kiocb kiocb;
 304	unsigned int i_head;
 305	int ret;
 306
 307	iov_iter_pipe(&to, READ, pipe, len);
 308	i_head = to.head;
 309	init_sync_kiocb(&kiocb, in);
 310	kiocb.ki_pos = *ppos;
 311	ret = call_read_iter(in, &kiocb, &to);
 312	if (ret > 0) {
 313		*ppos = kiocb.ki_pos;
 314		file_accessed(in);
 315	} else if (ret < 0) {
 316		to.head = i_head;
 317		to.iov_offset = 0;
 318		iov_iter_advance(&to, 0); /* to free what was emitted */
 319		/*
 320		 * callers of ->splice_read() expect -EAGAIN on
 321		 * "can't put anything in there", rather than -EFAULT.
 322		 */
 323		if (ret == -EFAULT)
 324			ret = -EAGAIN;
 325	}
 326
 327	return ret;
 328}
 329EXPORT_SYMBOL(generic_file_splice_read);
 330
 331const struct pipe_buf_operations default_pipe_buf_ops = {
 332	.release	= generic_pipe_buf_release,
 333	.try_steal	= generic_pipe_buf_try_steal,
 334	.get		= generic_pipe_buf_get,
 335};
 336
 337/* Pipe buffer operations for a socket and similar. */
 338const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 339	.release	= generic_pipe_buf_release,
 340	.get		= generic_pipe_buf_get,
 341};
 342EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 343
 344/*
 345 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 346 * using sendpage(). Return the number of bytes sent.
 347 */
 348static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 349			    struct pipe_buffer *buf, struct splice_desc *sd)
 350{
 351	struct file *file = sd->u.file;
 352	loff_t pos = sd->pos;
 353	int more;
 354
 355	if (!likely(file->f_op->sendpage))
 356		return -EINVAL;
 357
 358	more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 359
 360	if (sd->len < sd->total_len &&
 361	    pipe_occupancy(pipe->head, pipe->tail) > 1)
 362		more |= MSG_SENDPAGE_NOTLAST;
 363
 364	return file->f_op->sendpage(file, buf->page, buf->offset,
 365				    sd->len, &pos, more);
 366}
 367
 368static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 369{
 370	smp_mb();
 371	if (waitqueue_active(&pipe->wr_wait))
 372		wake_up_interruptible(&pipe->wr_wait);
 373	kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 374}
 375
 376/**
 377 * splice_from_pipe_feed - feed available data from a pipe to a file
 378 * @pipe:	pipe to splice from
 379 * @sd:		information to @actor
 380 * @actor:	handler that splices the data
 381 *
 382 * Description:
 383 *    This function loops over the pipe and calls @actor to do the
 384 *    actual moving of a single struct pipe_buffer to the desired
 385 *    destination.  It returns when there's no more buffers left in
 386 *    the pipe or if the requested number of bytes (@sd->total_len)
 387 *    have been copied.  It returns a positive number (one) if the
 388 *    pipe needs to be filled with more data, zero if the required
 389 *    number of bytes have been copied and -errno on error.
 390 *
 391 *    This, together with splice_from_pipe_{begin,end,next}, may be
 392 *    used to implement the functionality of __splice_from_pipe() when
 393 *    locking is required around copying the pipe buffers to the
 394 *    destination.
 395 */
 396static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 397			  splice_actor *actor)
 398{
 399	unsigned int head = pipe->head;
 400	unsigned int tail = pipe->tail;
 401	unsigned int mask = pipe->ring_size - 1;
 402	int ret;
 403
 404	while (!pipe_empty(head, tail)) {
 405		struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 406
 407		sd->len = buf->len;
 408		if (sd->len > sd->total_len)
 409			sd->len = sd->total_len;
 410
 411		ret = pipe_buf_confirm(pipe, buf);
 412		if (unlikely(ret)) {
 413			if (ret == -ENODATA)
 414				ret = 0;
 415			return ret;
 416		}
 417
 418		ret = actor(pipe, buf, sd);
 419		if (ret <= 0)
 420			return ret;
 421
 422		buf->offset += ret;
 423		buf->len -= ret;
 424
 425		sd->num_spliced += ret;
 426		sd->len -= ret;
 427		sd->pos += ret;
 428		sd->total_len -= ret;
 429
 430		if (!buf->len) {
 431			pipe_buf_release(pipe, buf);
 432			tail++;
 433			pipe->tail = tail;
 434			if (pipe->files)
 435				sd->need_wakeup = true;
 436		}
 437
 438		if (!sd->total_len)
 439			return 0;
 440	}
 441
 442	return 1;
 443}
 444
 445/* We know we have a pipe buffer, but maybe it's empty? */
 446static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
 447{
 448	unsigned int tail = pipe->tail;
 449	unsigned int mask = pipe->ring_size - 1;
 450	struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 451
 452	if (unlikely(!buf->len)) {
 453		pipe_buf_release(pipe, buf);
 454		pipe->tail = tail+1;
 455		return true;
 456	}
 457
 458	return false;
 459}
 460
 461/**
 462 * splice_from_pipe_next - wait for some data to splice from
 463 * @pipe:	pipe to splice from
 464 * @sd:		information about the splice operation
 465 *
 466 * Description:
 467 *    This function will wait for some data and return a positive
 468 *    value (one) if pipe buffers are available.  It will return zero
 469 *    or -errno if no more data needs to be spliced.
 470 */
 471static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 472{
 473	/*
 474	 * Check for signal early to make process killable when there are
 475	 * always buffers available
 476	 */
 477	if (signal_pending(current))
 478		return -ERESTARTSYS;
 479
 480repeat:
 481	while (pipe_empty(pipe->head, pipe->tail)) {
 482		if (!pipe->writers)
 483			return 0;
 484
 485		if (sd->num_spliced)
 486			return 0;
 487
 488		if (sd->flags & SPLICE_F_NONBLOCK)
 489			return -EAGAIN;
 490
 491		if (signal_pending(current))
 492			return -ERESTARTSYS;
 493
 494		if (sd->need_wakeup) {
 495			wakeup_pipe_writers(pipe);
 496			sd->need_wakeup = false;
 497		}
 498
 499		pipe_wait_readable(pipe);
 500	}
 501
 502	if (eat_empty_buffer(pipe))
 503		goto repeat;
 504
 505	return 1;
 506}
 507
 508/**
 509 * splice_from_pipe_begin - start splicing from pipe
 510 * @sd:		information about the splice operation
 511 *
 512 * Description:
 513 *    This function should be called before a loop containing
 514 *    splice_from_pipe_next() and splice_from_pipe_feed() to
 515 *    initialize the necessary fields of @sd.
 516 */
 517static void splice_from_pipe_begin(struct splice_desc *sd)
 518{
 519	sd->num_spliced = 0;
 520	sd->need_wakeup = false;
 521}
 522
 523/**
 524 * splice_from_pipe_end - finish splicing from pipe
 525 * @pipe:	pipe to splice from
 526 * @sd:		information about the splice operation
 527 *
 528 * Description:
 529 *    This function will wake up pipe writers if necessary.  It should
 530 *    be called after a loop containing splice_from_pipe_next() and
 531 *    splice_from_pipe_feed().
 532 */
 533static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 534{
 535	if (sd->need_wakeup)
 536		wakeup_pipe_writers(pipe);
 537}
 538
 539/**
 540 * __splice_from_pipe - splice data from a pipe to given actor
 541 * @pipe:	pipe to splice from
 542 * @sd:		information to @actor
 543 * @actor:	handler that splices the data
 544 *
 545 * Description:
 546 *    This function does little more than loop over the pipe and call
 547 *    @actor to do the actual moving of a single struct pipe_buffer to
 548 *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 549 *    pipe_to_user.
 550 *
 551 */
 552ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 553			   splice_actor *actor)
 554{
 555	int ret;
 556
 557	splice_from_pipe_begin(sd);
 558	do {
 559		cond_resched();
 560		ret = splice_from_pipe_next(pipe, sd);
 561		if (ret > 0)
 562			ret = splice_from_pipe_feed(pipe, sd, actor);
 563	} while (ret > 0);
 564	splice_from_pipe_end(pipe, sd);
 565
 566	return sd->num_spliced ? sd->num_spliced : ret;
 567}
 568EXPORT_SYMBOL(__splice_from_pipe);
 569
 570/**
 571 * splice_from_pipe - splice data from a pipe to a file
 572 * @pipe:	pipe to splice from
 573 * @out:	file to splice to
 574 * @ppos:	position in @out
 575 * @len:	how many bytes to splice
 576 * @flags:	splice modifier flags
 577 * @actor:	handler that splices the data
 578 *
 579 * Description:
 580 *    See __splice_from_pipe. This function locks the pipe inode,
 581 *    otherwise it's identical to __splice_from_pipe().
 582 *
 583 */
 584ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 585			 loff_t *ppos, size_t len, unsigned int flags,
 586			 splice_actor *actor)
 587{
 588	ssize_t ret;
 589	struct splice_desc sd = {
 590		.total_len = len,
 591		.flags = flags,
 592		.pos = *ppos,
 593		.u.file = out,
 594	};
 595
 596	pipe_lock(pipe);
 597	ret = __splice_from_pipe(pipe, &sd, actor);
 598	pipe_unlock(pipe);
 599
 600	return ret;
 601}
 602
 603/**
 604 * iter_file_splice_write - splice data from a pipe to a file
 605 * @pipe:	pipe info
 606 * @out:	file to write to
 607 * @ppos:	position in @out
 608 * @len:	number of bytes to splice
 609 * @flags:	splice modifier flags
 610 *
 611 * Description:
 612 *    Will either move or copy pages (determined by @flags options) from
 613 *    the given pipe inode to the given file.
 614 *    This one is ->write_iter-based.
 615 *
 616 */
 617ssize_t
 618iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 619			  loff_t *ppos, size_t len, unsigned int flags)
 620{
 621	struct splice_desc sd = {
 622		.total_len = len,
 623		.flags = flags,
 624		.pos = *ppos,
 625		.u.file = out,
 626	};
 627	int nbufs = pipe->max_usage;
 628	struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 629					GFP_KERNEL);
 630	ssize_t ret;
 631
 632	if (unlikely(!array))
 633		return -ENOMEM;
 634
 635	pipe_lock(pipe);
 636
 637	splice_from_pipe_begin(&sd);
 638	while (sd.total_len) {
 639		struct iov_iter from;
 640		unsigned int head, tail, mask;
 641		size_t left;
 642		int n;
 643
 644		ret = splice_from_pipe_next(pipe, &sd);
 645		if (ret <= 0)
 646			break;
 647
 648		if (unlikely(nbufs < pipe->max_usage)) {
 649			kfree(array);
 650			nbufs = pipe->max_usage;
 651			array = kcalloc(nbufs, sizeof(struct bio_vec),
 652					GFP_KERNEL);
 653			if (!array) {
 654				ret = -ENOMEM;
 655				break;
 656			}
 657		}
 658
 659		head = pipe->head;
 660		tail = pipe->tail;
 661		mask = pipe->ring_size - 1;
 662
 663		/* build the vector */
 664		left = sd.total_len;
 665		for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
 666			struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 667			size_t this_len = buf->len;
 668
 669			/* zero-length bvecs are not supported, skip them */
 670			if (!this_len)
 671				continue;
 672			this_len = min(this_len, left);
 673
 674			ret = pipe_buf_confirm(pipe, buf);
 675			if (unlikely(ret)) {
 676				if (ret == -ENODATA)
 677					ret = 0;
 678				goto done;
 679			}
 680
 681			array[n].bv_page = buf->page;
 682			array[n].bv_len = this_len;
 683			array[n].bv_offset = buf->offset;
 684			left -= this_len;
 685			n++;
 686		}
 687
 688		iov_iter_bvec(&from, WRITE, array, n, sd.total_len - left);
 689		ret = vfs_iter_write(out, &from, &sd.pos, 0);
 690		if (ret <= 0)
 691			break;
 692
 693		sd.num_spliced += ret;
 694		sd.total_len -= ret;
 695		*ppos = sd.pos;
 696
 697		/* dismiss the fully eaten buffers, adjust the partial one */
 698		tail = pipe->tail;
 699		while (ret) {
 700			struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 701			if (ret >= buf->len) {
 702				ret -= buf->len;
 703				buf->len = 0;
 704				pipe_buf_release(pipe, buf);
 705				tail++;
 706				pipe->tail = tail;
 707				if (pipe->files)
 708					sd.need_wakeup = true;
 709			} else {
 710				buf->offset += ret;
 711				buf->len -= ret;
 712				ret = 0;
 713			}
 714		}
 715	}
 716done:
 717	kfree(array);
 718	splice_from_pipe_end(pipe, &sd);
 719
 720	pipe_unlock(pipe);
 721
 722	if (sd.num_spliced)
 723		ret = sd.num_spliced;
 724
 725	return ret;
 726}
 727
 728EXPORT_SYMBOL(iter_file_splice_write);
 729
 730/**
 731 * generic_splice_sendpage - splice data from a pipe to a socket
 732 * @pipe:	pipe to splice from
 733 * @out:	socket to write to
 734 * @ppos:	position in @out
 735 * @len:	number of bytes to splice
 736 * @flags:	splice modifier flags
 737 *
 738 * Description:
 739 *    Will send @len bytes from the pipe to a network socket. No data copying
 740 *    is involved.
 741 *
 742 */
 743ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 744				loff_t *ppos, size_t len, unsigned int flags)
 745{
 746	return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 747}
 748
 749EXPORT_SYMBOL(generic_splice_sendpage);
 750
 751static int warn_unsupported(struct file *file, const char *op)
 752{
 753	pr_debug_ratelimited(
 754		"splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 755		op, file, current->pid, current->comm);
 756	return -EINVAL;
 757}
 758
 759/*
 760 * Attempt to initiate a splice from pipe to file.
 761 */
 762static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 763			   loff_t *ppos, size_t len, unsigned int flags)
 764{
 765	if (unlikely(!out->f_op->splice_write))
 766		return warn_unsupported(out, "write");
 767	return out->f_op->splice_write(pipe, out, ppos, len, flags);
 768}
 769
 770/*
 771 * Attempt to initiate a splice from a file to a pipe.
 772 */
 773static long do_splice_to(struct file *in, loff_t *ppos,
 774			 struct pipe_inode_info *pipe, size_t len,
 775			 unsigned int flags)
 776{
 777	unsigned int p_space;
 778	int ret;
 779
 780	if (unlikely(!(in->f_mode & FMODE_READ)))
 781		return -EBADF;
 782
 783	/* Don't try to read more the pipe has space for. */
 784	p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 785	len = min_t(size_t, len, p_space << PAGE_SHIFT);
 786
 787	ret = rw_verify_area(READ, in, ppos, len);
 788	if (unlikely(ret < 0))
 789		return ret;
 790
 791	if (unlikely(len > MAX_RW_COUNT))
 792		len = MAX_RW_COUNT;
 793
 794	if (unlikely(!in->f_op->splice_read))
 795		return warn_unsupported(in, "read");
 796	return in->f_op->splice_read(in, ppos, pipe, len, flags);
 797}
 798
 799/**
 800 * splice_direct_to_actor - splices data directly between two non-pipes
 801 * @in:		file to splice from
 802 * @sd:		actor information on where to splice to
 803 * @actor:	handles the data splicing
 804 *
 805 * Description:
 806 *    This is a special case helper to splice directly between two
 807 *    points, without requiring an explicit pipe. Internally an allocated
 808 *    pipe is cached in the process, and reused during the lifetime of
 809 *    that process.
 810 *
 811 */
 812ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 813			       splice_direct_actor *actor)
 814{
 815	struct pipe_inode_info *pipe;
 816	long ret, bytes;
 817	umode_t i_mode;
 818	size_t len;
 819	int i, flags, more;
 820
 821	/*
 822	 * We require the input being a regular file, as we don't want to
 823	 * randomly drop data for eg socket -> socket splicing. Use the
 824	 * piped splicing for that!
 825	 */
 826	i_mode = file_inode(in)->i_mode;
 827	if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 828		return -EINVAL;
 829
 830	/*
 831	 * neither in nor out is a pipe, setup an internal pipe attached to
 832	 * 'out' and transfer the wanted data from 'in' to 'out' through that
 833	 */
 834	pipe = current->splice_pipe;
 835	if (unlikely(!pipe)) {
 836		pipe = alloc_pipe_info();
 837		if (!pipe)
 838			return -ENOMEM;
 839
 840		/*
 841		 * We don't have an immediate reader, but we'll read the stuff
 842		 * out of the pipe right after the splice_to_pipe(). So set
 843		 * PIPE_READERS appropriately.
 844		 */
 845		pipe->readers = 1;
 846
 847		current->splice_pipe = pipe;
 848	}
 849
 850	/*
 851	 * Do the splice.
 852	 */
 853	ret = 0;
 854	bytes = 0;
 855	len = sd->total_len;
 856	flags = sd->flags;
 857
 858	/*
 859	 * Don't block on output, we have to drain the direct pipe.
 860	 */
 861	sd->flags &= ~SPLICE_F_NONBLOCK;
 862	more = sd->flags & SPLICE_F_MORE;
 863
 864	WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
 865
 866	while (len) {
 867		size_t read_len;
 868		loff_t pos = sd->pos, prev_pos = pos;
 869
 870		ret = do_splice_to(in, &pos, pipe, len, flags);
 871		if (unlikely(ret <= 0))
 872			goto out_release;
 873
 874		read_len = ret;
 875		sd->total_len = read_len;
 876
 877		/*
 878		 * If more data is pending, set SPLICE_F_MORE
 879		 * If this is the last data and SPLICE_F_MORE was not set
 880		 * initially, clears it.
 881		 */
 882		if (read_len < len)
 883			sd->flags |= SPLICE_F_MORE;
 884		else if (!more)
 885			sd->flags &= ~SPLICE_F_MORE;
 886		/*
 887		 * NOTE: nonblocking mode only applies to the input. We
 888		 * must not do the output in nonblocking mode as then we
 889		 * could get stuck data in the internal pipe:
 890		 */
 891		ret = actor(pipe, sd);
 892		if (unlikely(ret <= 0)) {
 893			sd->pos = prev_pos;
 894			goto out_release;
 895		}
 896
 897		bytes += ret;
 898		len -= ret;
 899		sd->pos = pos;
 900
 901		if (ret < read_len) {
 902			sd->pos = prev_pos + ret;
 903			goto out_release;
 904		}
 905	}
 906
 907done:
 908	pipe->tail = pipe->head = 0;
 909	file_accessed(in);
 910	return bytes;
 911
 912out_release:
 913	/*
 914	 * If we did an incomplete transfer we must release
 915	 * the pipe buffers in question:
 916	 */
 917	for (i = 0; i < pipe->ring_size; i++) {
 918		struct pipe_buffer *buf = &pipe->bufs[i];
 919
 920		if (buf->ops)
 921			pipe_buf_release(pipe, buf);
 922	}
 923
 924	if (!bytes)
 925		bytes = ret;
 926
 927	goto done;
 928}
 929EXPORT_SYMBOL(splice_direct_to_actor);
 930
 931static int direct_splice_actor(struct pipe_inode_info *pipe,
 932			       struct splice_desc *sd)
 933{
 934	struct file *file = sd->u.file;
 935
 936	return do_splice_from(pipe, file, sd->opos, sd->total_len,
 937			      sd->flags);
 938}
 939
 940/**
 941 * do_splice_direct - splices data directly between two files
 942 * @in:		file to splice from
 943 * @ppos:	input file offset
 944 * @out:	file to splice to
 945 * @opos:	output file offset
 946 * @len:	number of bytes to splice
 947 * @flags:	splice modifier flags
 948 *
 949 * Description:
 950 *    For use by do_sendfile(). splice can easily emulate sendfile, but
 951 *    doing it in the application would incur an extra system call
 952 *    (splice in + splice out, as compared to just sendfile()). So this helper
 953 *    can splice directly through a process-private pipe.
 954 *
 955 */
 956long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 957		      loff_t *opos, size_t len, unsigned int flags)
 958{
 959	struct splice_desc sd = {
 960		.len		= len,
 961		.total_len	= len,
 962		.flags		= flags,
 963		.pos		= *ppos,
 964		.u.file		= out,
 965		.opos		= opos,
 966	};
 967	long ret;
 968
 969	if (unlikely(!(out->f_mode & FMODE_WRITE)))
 970		return -EBADF;
 971
 972	if (unlikely(out->f_flags & O_APPEND))
 973		return -EINVAL;
 974
 975	ret = rw_verify_area(WRITE, out, opos, len);
 976	if (unlikely(ret < 0))
 977		return ret;
 978
 979	ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
 980	if (ret > 0)
 981		*ppos = sd.pos;
 982
 983	return ret;
 984}
 985EXPORT_SYMBOL(do_splice_direct);
 986
 987static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
 988{
 989	for (;;) {
 990		if (unlikely(!pipe->readers)) {
 991			send_sig(SIGPIPE, current, 0);
 992			return -EPIPE;
 993		}
 994		if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 995			return 0;
 996		if (flags & SPLICE_F_NONBLOCK)
 997			return -EAGAIN;
 998		if (signal_pending(current))
 999			return -ERESTARTSYS;
1000		pipe_wait_writable(pipe);
1001	}
1002}
1003
1004static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1005			       struct pipe_inode_info *opipe,
1006			       size_t len, unsigned int flags);
1007
1008long splice_file_to_pipe(struct file *in,
1009			 struct pipe_inode_info *opipe,
1010			 loff_t *offset,
1011			 size_t len, unsigned int flags)
1012{
1013	long ret;
1014
1015	pipe_lock(opipe);
1016	ret = wait_for_space(opipe, flags);
1017	if (!ret)
1018		ret = do_splice_to(in, offset, opipe, len, flags);
1019	pipe_unlock(opipe);
1020	if (ret > 0)
1021		wakeup_pipe_readers(opipe);
1022	return ret;
1023}
1024
1025/*
1026 * Determine where to splice to/from.
1027 */
1028long do_splice(struct file *in, loff_t *off_in, struct file *out,
1029	       loff_t *off_out, size_t len, unsigned int flags)
1030{
1031	struct pipe_inode_info *ipipe;
1032	struct pipe_inode_info *opipe;
1033	loff_t offset;
1034	long ret;
1035
1036	if (unlikely(!(in->f_mode & FMODE_READ) ||
1037		     !(out->f_mode & FMODE_WRITE)))
1038		return -EBADF;
1039
1040	ipipe = get_pipe_info(in, true);
1041	opipe = get_pipe_info(out, true);
1042
1043	if (ipipe && opipe) {
1044		if (off_in || off_out)
1045			return -ESPIPE;
1046
1047		/* Splicing to self would be fun, but... */
1048		if (ipipe == opipe)
1049			return -EINVAL;
1050
1051		if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1052			flags |= SPLICE_F_NONBLOCK;
1053
1054		return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1055	}
1056
1057	if (ipipe) {
1058		if (off_in)
1059			return -ESPIPE;
1060		if (off_out) {
1061			if (!(out->f_mode & FMODE_PWRITE))
1062				return -EINVAL;
1063			offset = *off_out;
1064		} else {
1065			offset = out->f_pos;
1066		}
1067
1068		if (unlikely(out->f_flags & O_APPEND))
1069			return -EINVAL;
1070
1071		ret = rw_verify_area(WRITE, out, &offset, len);
1072		if (unlikely(ret < 0))
1073			return ret;
1074
1075		if (in->f_flags & O_NONBLOCK)
1076			flags |= SPLICE_F_NONBLOCK;
1077
1078		file_start_write(out);
1079		ret = do_splice_from(ipipe, out, &offset, len, flags);
1080		file_end_write(out);
1081
1082		if (!off_out)
1083			out->f_pos = offset;
1084		else
1085			*off_out = offset;
1086
1087		return ret;
1088	}
1089
1090	if (opipe) {
1091		if (off_out)
1092			return -ESPIPE;
1093		if (off_in) {
1094			if (!(in->f_mode & FMODE_PREAD))
1095				return -EINVAL;
1096			offset = *off_in;
1097		} else {
1098			offset = in->f_pos;
1099		}
1100
1101		if (out->f_flags & O_NONBLOCK)
1102			flags |= SPLICE_F_NONBLOCK;
1103
1104		ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1105		if (!off_in)
1106			in->f_pos = offset;
1107		else
1108			*off_in = offset;
1109
1110		return ret;
1111	}
1112
1113	return -EINVAL;
1114}
1115
1116static long __do_splice(struct file *in, loff_t __user *off_in,
1117			struct file *out, loff_t __user *off_out,
1118			size_t len, unsigned int flags)
1119{
1120	struct pipe_inode_info *ipipe;
1121	struct pipe_inode_info *opipe;
1122	loff_t offset, *__off_in = NULL, *__off_out = NULL;
1123	long ret;
1124
1125	ipipe = get_pipe_info(in, true);
1126	opipe = get_pipe_info(out, true);
1127
1128	if (ipipe && off_in)
1129		return -ESPIPE;
1130	if (opipe && off_out)
1131		return -ESPIPE;
1132
1133	if (off_out) {
1134		if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1135			return -EFAULT;
1136		__off_out = &offset;
1137	}
1138	if (off_in) {
1139		if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1140			return -EFAULT;
1141		__off_in = &offset;
1142	}
1143
1144	ret = do_splice(in, __off_in, out, __off_out, len, flags);
1145	if (ret < 0)
1146		return ret;
1147
1148	if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1149		return -EFAULT;
1150	if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1151		return -EFAULT;
1152
1153	return ret;
1154}
1155
1156static int iter_to_pipe(struct iov_iter *from,
1157			struct pipe_inode_info *pipe,
1158			unsigned flags)
1159{
1160	struct pipe_buffer buf = {
1161		.ops = &user_page_pipe_buf_ops,
1162		.flags = flags
1163	};
1164	size_t total = 0;
1165	int ret = 0;
1166	bool failed = false;
1167
1168	while (iov_iter_count(from) && !failed) {
1169		struct page *pages[16];
1170		ssize_t copied;
1171		size_t start;
1172		int n;
1173
1174		copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1175		if (copied <= 0) {
1176			ret = copied;
1177			break;
1178		}
1179
1180		for (n = 0; copied; n++, start = 0) {
1181			int size = min_t(int, copied, PAGE_SIZE - start);
1182			if (!failed) {
1183				buf.page = pages[n];
1184				buf.offset = start;
1185				buf.len = size;
1186				ret = add_to_pipe(pipe, &buf);
1187				if (unlikely(ret < 0)) {
1188					failed = true;
1189				} else {
1190					iov_iter_advance(from, ret);
1191					total += ret;
1192				}
1193			} else {
1194				put_page(pages[n]);
1195			}
1196			copied -= size;
1197		}
1198	}
1199	return total ? total : ret;
1200}
1201
1202static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1203			struct splice_desc *sd)
1204{
1205	int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1206	return n == sd->len ? n : -EFAULT;
1207}
1208
1209/*
1210 * For lack of a better implementation, implement vmsplice() to userspace
1211 * as a simple copy of the pipes pages to the user iov.
1212 */
1213static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1214			     unsigned int flags)
1215{
1216	struct pipe_inode_info *pipe = get_pipe_info(file, true);
1217	struct splice_desc sd = {
1218		.total_len = iov_iter_count(iter),
1219		.flags = flags,
1220		.u.data = iter
1221	};
1222	long ret = 0;
1223
1224	if (!pipe)
1225		return -EBADF;
1226
1227	if (sd.total_len) {
1228		pipe_lock(pipe);
1229		ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1230		pipe_unlock(pipe);
1231	}
1232
1233	return ret;
1234}
1235
1236/*
1237 * vmsplice splices a user address range into a pipe. It can be thought of
1238 * as splice-from-memory, where the regular splice is splice-from-file (or
1239 * to file). In both cases the output is a pipe, naturally.
1240 */
1241static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1242			     unsigned int flags)
1243{
1244	struct pipe_inode_info *pipe;
1245	long ret = 0;
1246	unsigned buf_flag = 0;
1247
1248	if (flags & SPLICE_F_GIFT)
1249		buf_flag = PIPE_BUF_FLAG_GIFT;
1250
1251	pipe = get_pipe_info(file, true);
1252	if (!pipe)
1253		return -EBADF;
1254
1255	pipe_lock(pipe);
1256	ret = wait_for_space(pipe, flags);
1257	if (!ret)
1258		ret = iter_to_pipe(iter, pipe, buf_flag);
1259	pipe_unlock(pipe);
1260	if (ret > 0)
1261		wakeup_pipe_readers(pipe);
1262	return ret;
1263}
1264
1265static int vmsplice_type(struct fd f, int *type)
1266{
1267	if (!f.file)
1268		return -EBADF;
1269	if (f.file->f_mode & FMODE_WRITE) {
1270		*type = WRITE;
1271	} else if (f.file->f_mode & FMODE_READ) {
1272		*type = READ;
1273	} else {
1274		fdput(f);
1275		return -EBADF;
1276	}
1277	return 0;
1278}
1279
1280/*
1281 * Note that vmsplice only really supports true splicing _from_ user memory
1282 * to a pipe, not the other way around. Splicing from user memory is a simple
1283 * operation that can be supported without any funky alignment restrictions
1284 * or nasty vm tricks. We simply map in the user memory and fill them into
1285 * a pipe. The reverse isn't quite as easy, though. There are two possible
1286 * solutions for that:
1287 *
1288 *	- memcpy() the data internally, at which point we might as well just
1289 *	  do a regular read() on the buffer anyway.
1290 *	- Lots of nasty vm tricks, that are neither fast nor flexible (it
1291 *	  has restriction limitations on both ends of the pipe).
1292 *
1293 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1294 *
1295 */
1296SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1297		unsigned long, nr_segs, unsigned int, flags)
1298{
1299	struct iovec iovstack[UIO_FASTIOV];
1300	struct iovec *iov = iovstack;
1301	struct iov_iter iter;
1302	ssize_t error;
1303	struct fd f;
1304	int type;
1305
1306	if (unlikely(flags & ~SPLICE_F_ALL))
1307		return -EINVAL;
1308
1309	f = fdget(fd);
1310	error = vmsplice_type(f, &type);
1311	if (error)
1312		return error;
1313
1314	error = import_iovec(type, uiov, nr_segs,
1315			     ARRAY_SIZE(iovstack), &iov, &iter);
1316	if (error < 0)
1317		goto out_fdput;
1318
1319	if (!iov_iter_count(&iter))
1320		error = 0;
1321	else if (iov_iter_rw(&iter) == WRITE)
1322		error = vmsplice_to_pipe(f.file, &iter, flags);
1323	else
1324		error = vmsplice_to_user(f.file, &iter, flags);
1325
1326	kfree(iov);
1327out_fdput:
1328	fdput(f);
1329	return error;
1330}
1331
1332SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1333		int, fd_out, loff_t __user *, off_out,
1334		size_t, len, unsigned int, flags)
1335{
1336	struct fd in, out;
1337	long error;
1338
1339	if (unlikely(!len))
1340		return 0;
1341
1342	if (unlikely(flags & ~SPLICE_F_ALL))
1343		return -EINVAL;
1344
1345	error = -EBADF;
1346	in = fdget(fd_in);
1347	if (in.file) {
1348		out = fdget(fd_out);
1349		if (out.file) {
1350			error = __do_splice(in.file, off_in, out.file, off_out,
1351						len, flags);
1352			fdput(out);
1353		}
1354		fdput(in);
1355	}
1356	return error;
1357}
1358
1359/*
1360 * Make sure there's data to read. Wait for input if we can, otherwise
1361 * return an appropriate error.
1362 */
1363static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1364{
1365	int ret;
1366
1367	/*
1368	 * Check the pipe occupancy without the inode lock first. This function
1369	 * is speculative anyways, so missing one is ok.
1370	 */
1371	if (!pipe_empty(pipe->head, pipe->tail))
1372		return 0;
1373
1374	ret = 0;
1375	pipe_lock(pipe);
1376
1377	while (pipe_empty(pipe->head, pipe->tail)) {
1378		if (signal_pending(current)) {
1379			ret = -ERESTARTSYS;
1380			break;
1381		}
1382		if (!pipe->writers)
1383			break;
1384		if (flags & SPLICE_F_NONBLOCK) {
1385			ret = -EAGAIN;
1386			break;
1387		}
1388		pipe_wait_readable(pipe);
1389	}
1390
1391	pipe_unlock(pipe);
1392	return ret;
1393}
1394
1395/*
1396 * Make sure there's writeable room. Wait for room if we can, otherwise
1397 * return an appropriate error.
1398 */
1399static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1400{
1401	int ret;
1402
1403	/*
1404	 * Check pipe occupancy without the inode lock first. This function
1405	 * is speculative anyways, so missing one is ok.
1406	 */
1407	if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1408		return 0;
1409
1410	ret = 0;
1411	pipe_lock(pipe);
1412
1413	while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1414		if (!pipe->readers) {
1415			send_sig(SIGPIPE, current, 0);
1416			ret = -EPIPE;
1417			break;
1418		}
1419		if (flags & SPLICE_F_NONBLOCK) {
1420			ret = -EAGAIN;
1421			break;
1422		}
1423		if (signal_pending(current)) {
1424			ret = -ERESTARTSYS;
1425			break;
1426		}
1427		pipe_wait_writable(pipe);
1428	}
1429
1430	pipe_unlock(pipe);
1431	return ret;
1432}
1433
1434/*
1435 * Splice contents of ipipe to opipe.
1436 */
1437static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1438			       struct pipe_inode_info *opipe,
1439			       size_t len, unsigned int flags)
1440{
1441	struct pipe_buffer *ibuf, *obuf;
1442	unsigned int i_head, o_head;
1443	unsigned int i_tail, o_tail;
1444	unsigned int i_mask, o_mask;
1445	int ret = 0;
1446	bool input_wakeup = false;
1447
1448
1449retry:
1450	ret = ipipe_prep(ipipe, flags);
1451	if (ret)
1452		return ret;
1453
1454	ret = opipe_prep(opipe, flags);
1455	if (ret)
1456		return ret;
1457
1458	/*
1459	 * Potential ABBA deadlock, work around it by ordering lock
1460	 * grabbing by pipe info address. Otherwise two different processes
1461	 * could deadlock (one doing tee from A -> B, the other from B -> A).
1462	 */
1463	pipe_double_lock(ipipe, opipe);
1464
1465	i_tail = ipipe->tail;
1466	i_mask = ipipe->ring_size - 1;
1467	o_head = opipe->head;
1468	o_mask = opipe->ring_size - 1;
1469
1470	do {
1471		size_t o_len;
1472
1473		if (!opipe->readers) {
1474			send_sig(SIGPIPE, current, 0);
1475			if (!ret)
1476				ret = -EPIPE;
1477			break;
1478		}
1479
1480		i_head = ipipe->head;
1481		o_tail = opipe->tail;
1482
1483		if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1484			break;
1485
1486		/*
1487		 * Cannot make any progress, because either the input
1488		 * pipe is empty or the output pipe is full.
1489		 */
1490		if (pipe_empty(i_head, i_tail) ||
1491		    pipe_full(o_head, o_tail, opipe->max_usage)) {
1492			/* Already processed some buffers, break */
1493			if (ret)
1494				break;
1495
1496			if (flags & SPLICE_F_NONBLOCK) {
1497				ret = -EAGAIN;
1498				break;
1499			}
1500
1501			/*
1502			 * We raced with another reader/writer and haven't
1503			 * managed to process any buffers.  A zero return
1504			 * value means EOF, so retry instead.
1505			 */
1506			pipe_unlock(ipipe);
1507			pipe_unlock(opipe);
1508			goto retry;
1509		}
1510
1511		ibuf = &ipipe->bufs[i_tail & i_mask];
1512		obuf = &opipe->bufs[o_head & o_mask];
1513
1514		if (len >= ibuf->len) {
1515			/*
1516			 * Simply move the whole buffer from ipipe to opipe
1517			 */
1518			*obuf = *ibuf;
1519			ibuf->ops = NULL;
1520			i_tail++;
1521			ipipe->tail = i_tail;
1522			input_wakeup = true;
1523			o_len = obuf->len;
1524			o_head++;
1525			opipe->head = o_head;
1526		} else {
1527			/*
1528			 * Get a reference to this pipe buffer,
1529			 * so we can copy the contents over.
1530			 */
1531			if (!pipe_buf_get(ipipe, ibuf)) {
1532				if (ret == 0)
1533					ret = -EFAULT;
1534				break;
1535			}
1536			*obuf = *ibuf;
1537
1538			/*
1539			 * Don't inherit the gift and merge flags, we need to
1540			 * prevent multiple steals of this page.
1541			 */
1542			obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1543			obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1544
1545			obuf->len = len;
1546			ibuf->offset += len;
1547			ibuf->len -= len;
1548			o_len = len;
1549			o_head++;
1550			opipe->head = o_head;
1551		}
1552		ret += o_len;
1553		len -= o_len;
1554	} while (len);
1555
1556	pipe_unlock(ipipe);
1557	pipe_unlock(opipe);
1558
1559	/*
1560	 * If we put data in the output pipe, wakeup any potential readers.
1561	 */
1562	if (ret > 0)
1563		wakeup_pipe_readers(opipe);
1564
1565	if (input_wakeup)
1566		wakeup_pipe_writers(ipipe);
1567
1568	return ret;
1569}
1570
1571/*
1572 * Link contents of ipipe to opipe.
1573 */
1574static int link_pipe(struct pipe_inode_info *ipipe,
1575		     struct pipe_inode_info *opipe,
1576		     size_t len, unsigned int flags)
1577{
1578	struct pipe_buffer *ibuf, *obuf;
1579	unsigned int i_head, o_head;
1580	unsigned int i_tail, o_tail;
1581	unsigned int i_mask, o_mask;
1582	int ret = 0;
1583
1584	/*
1585	 * Potential ABBA deadlock, work around it by ordering lock
1586	 * grabbing by pipe info address. Otherwise two different processes
1587	 * could deadlock (one doing tee from A -> B, the other from B -> A).
1588	 */
1589	pipe_double_lock(ipipe, opipe);
1590
1591	i_tail = ipipe->tail;
1592	i_mask = ipipe->ring_size - 1;
1593	o_head = opipe->head;
1594	o_mask = opipe->ring_size - 1;
1595
1596	do {
1597		if (!opipe->readers) {
1598			send_sig(SIGPIPE, current, 0);
1599			if (!ret)
1600				ret = -EPIPE;
1601			break;
1602		}
1603
1604		i_head = ipipe->head;
1605		o_tail = opipe->tail;
1606
1607		/*
1608		 * If we have iterated all input buffers or run out of
1609		 * output room, break.
1610		 */
1611		if (pipe_empty(i_head, i_tail) ||
1612		    pipe_full(o_head, o_tail, opipe->max_usage))
1613			break;
1614
1615		ibuf = &ipipe->bufs[i_tail & i_mask];
1616		obuf = &opipe->bufs[o_head & o_mask];
1617
1618		/*
1619		 * Get a reference to this pipe buffer,
1620		 * so we can copy the contents over.
1621		 */
1622		if (!pipe_buf_get(ipipe, ibuf)) {
1623			if (ret == 0)
1624				ret = -EFAULT;
1625			break;
1626		}
1627
1628		*obuf = *ibuf;
1629
1630		/*
1631		 * Don't inherit the gift and merge flag, we need to prevent
1632		 * multiple steals of this page.
1633		 */
1634		obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1635		obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1636
1637		if (obuf->len > len)
1638			obuf->len = len;
1639		ret += obuf->len;
1640		len -= obuf->len;
1641
1642		o_head++;
1643		opipe->head = o_head;
1644		i_tail++;
1645	} while (len);
1646
1647	pipe_unlock(ipipe);
1648	pipe_unlock(opipe);
1649
1650	/*
1651	 * If we put data in the output pipe, wakeup any potential readers.
1652	 */
1653	if (ret > 0)
1654		wakeup_pipe_readers(opipe);
1655
1656	return ret;
1657}
1658
1659/*
1660 * This is a tee(1) implementation that works on pipes. It doesn't copy
1661 * any data, it simply references the 'in' pages on the 'out' pipe.
1662 * The 'flags' used are the SPLICE_F_* variants, currently the only
1663 * applicable one is SPLICE_F_NONBLOCK.
1664 */
1665long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1666{
1667	struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1668	struct pipe_inode_info *opipe = get_pipe_info(out, true);
1669	int ret = -EINVAL;
1670
1671	if (unlikely(!(in->f_mode & FMODE_READ) ||
1672		     !(out->f_mode & FMODE_WRITE)))
1673		return -EBADF;
1674
1675	/*
1676	 * Duplicate the contents of ipipe to opipe without actually
1677	 * copying the data.
1678	 */
1679	if (ipipe && opipe && ipipe != opipe) {
1680		if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1681			flags |= SPLICE_F_NONBLOCK;
1682
1683		/*
1684		 * Keep going, unless we encounter an error. The ipipe/opipe
1685		 * ordering doesn't really matter.
1686		 */
1687		ret = ipipe_prep(ipipe, flags);
1688		if (!ret) {
1689			ret = opipe_prep(opipe, flags);
1690			if (!ret)
1691				ret = link_pipe(ipipe, opipe, len, flags);
1692		}
1693	}
1694
1695	return ret;
1696}
1697
1698SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1699{
1700	struct fd in, out;
1701	int error;
1702
1703	if (unlikely(flags & ~SPLICE_F_ALL))
1704		return -EINVAL;
1705
1706	if (unlikely(!len))
1707		return 0;
1708
1709	error = -EBADF;
1710	in = fdget(fdin);
1711	if (in.file) {
1712		out = fdget(fdout);
1713		if (out.file) {
1714			error = do_tee(in.file, out.file, len, flags);
1715			fdput(out);
1716		}
1717 		fdput(in);
1718 	}
1719
1720	return error;
1721}