fs/netfs/direct_write.c

// SPDX-License-Identifier: GPL-2.0-or-later
/* Unbuffered and direct write support.
 *
 * Copyright (C) 2023 Red Hat, Inc. All Rights Reserved.
 * Written by David Howells (dhowells@redhat.com)
 */

#include <linux/export.h>
#include <linux/uio.h>
#include "internal.h"

static void netfs_cleanup_dio_write(struct netfs_io_request *wreq)
{
	struct inode *inode = wreq->inode;
	unsigned long long end = wreq->start + wreq->transferred;

	if (!wreq->error &&
	    i_size_read(inode) < end) {
		if (wreq->netfs_ops->update_i_size)
			wreq->netfs_ops->update_i_size(inode, end);
		else
			i_size_write(inode, end);
	}
}

/*
 * Write some data from a bounce buffer folio to the server.
 */
static int netfs_write_bounce(struct netfs_io_request *wreq, struct folio *folio,
			      size_t foff, size_t flen)
{
	struct netfs_io_stream *stream = &wreq->io_streams[0];
	size_t fsize = folio_size(folio);
	loff_t fpos = folio_pos(folio);
	bool debug = false;

	_enter("%zx,%zx", foff, flen);

	//stream->sreq_max_len = fsize;
	stream->submit_off = foff;
	stream->submit_len = flen;

	/* Attach the folio to one or more subrequests.  For a big folio, we
	 * could end up with thousands of subrequests if the wsize is small -
	 * but we might need to wait during the creation of subrequests for
	 * network resources (eg. SMB credits).
	 */
	do {
		ssize_t part;

		part = netfs_advance_write(wreq, stream, fpos + stream->submit_off,
					   stream->submit_len, false);
		atomic64_set(&wreq->issued_to, fpos + stream->submit_off);
		stream->submit_off += part;
		stream->sreq_max_len -= part;
		if (part > stream->submit_len)
			stream->submit_len = 0;
		else
			stream->submit_len -= part;
		if (part > 0)
			debug = true;
		cond_resched();
	} while (stream->submit_len > 0);

	atomic64_set(&wreq->issued_to, fpos + fsize);

	if (!debug)
		kdebug("R=%x: No submit", wreq->debug_id);

	_leave(" = 0");
	return 0;
}

/*
 * Insert pages into the bounce buffer and fill them, dispatching writes to
 * cover them as we go.  If there are gaps at either end of the bounce buffer
 * that need to be filled, we download the content for those from the server
 * and perform an RMW cycle.
 */
static ssize_t netfs_write_through_bounce_buffer(struct netfs_io_request *wreq,
						 struct kiocb *iocb, struct iov_iter *iter,
						 struct netfs_group *netfs_group)
{
	struct netfs_io_stream *upload = &wreq->io_streams[0];
	struct netfs_inode *ictx = netfs_inode(wreq->inode);
	struct folio_queue *fq;
	unsigned long long real_size = ictx->remote_i_size;
	unsigned long long start = wreq->start;
	unsigned long long end = start + iov_iter_count(iter);
	size_t bsize = wreq->crypto_bsize;
	size_t bmask = bsize - 1;
	size_t len = iov_iter_count(iter);
	size_t gap_before = start & bmask;
	size_t gap_after = (bsize - (gap_before + len)) & bmask;
	size_t written = 0;
	int ret, slot;

	_enter("%llx-%llx", start, end);

	/* Assume crypto involves units of >1 byte. */
	if (WARN_ON_ONCE(bsize <= 1))
		return -EIO;

	/* The real size must be rounded out to the crypto block size plus
	 * any trailer we might want to attach.
	 */
	if (real_size && wreq->crypto_bsize) {
		if (real_size < wreq->crypto_trailer)
			return -EIO;
		if ((real_size - wreq->crypto_trailer) & bmask)
			return -EIO;
		real_size -= wreq->crypto_trailer;
	}

	if (wreq->origin == NETFS_DIO_WRITE)
		inode_dio_begin(wreq->inode);

	/* Prime the bounce buffer since we're going to need at least
	 * one folio therein.
	 */
	ret = rolling_buffer_init(&wreq->bounce, wreq->debug_id, ITER_SOURCE);
	if (ret < 0)
		goto error;
	wreq->bounce_alloc_to = round_down(start, umax(bsize, PAGE_SIZE));
	ret = netfs_alloc_bounce(wreq, round_down(start, bsize) + bsize, GFP_KERNEL);
	if (ret < 0)
		goto error;
	fq = wreq->bounce.tail;
	slot = 0;

	/* Populate the buffer for any RMW we need to do at either end of the
	 * bounce buffer and perform the read(s).
	 */
	if (gap_before || gap_after) {
		struct folio *folio1 = folioq_folio(fq, slot);
		struct folio *folio2 = NULL;
		unsigned long long gstart = start - gap_before;
		unsigned long long gend = start + len + gap_after;
		bool read_before = false, read_after = false;
		bool overlap = round_down(gend - bsize, PAGE_SIZE) <=
			round_down(gstart, PAGE_SIZE);

		wreq->rmw_gap_before	= gap_before;
		wreq->rmw_gap_after	= gap_after;
		wreq->start		= gstart;
		wreq->len		= gend - gstart;

		__set_bit(NETFS_RREQ_RMW, &ictx->flags);
		if (gstart >= end) {
			/* At or after EOF, nothing to read. */
		} else {
			if (gap_before)
				read_before = true; /* Part of first gap before EOF. */

			if (!gap_after)
				;
			else if (gend - bsize <= gstart)
				; /* All in one crypto block. */
			else if (gend - bsize >= real_size)
				; /* Last block at/after EOF. */
			else
				read_after = true; /* Part of last crypto block before EOF. */
		}

		_debug("gaps=%zx,%zx ov=%u read=%u,%u",
		       gap_before, gap_after, overlap, read_before, read_after);

		if (gap_after && !overlap) {
			unsigned int shift = umax(ilog2(bsize), PAGE_SHIFT);
			unsigned int order = shift - PAGE_SHIFT;

			ret = -ENOMEM;
			folio2 = folio_alloc(GFP_KERNEL, order);
			if (!folio2)
				goto error;

			folio2->index = (gend - folio_size(folio2)) / PAGE_SIZE;
			wreq->rmw_tail = folio2;
			__set_bit(NETFS_RREQ_PUT_RMW_TAIL, &wreq->flags);
		} else if (gap_after && overlap) {
			wreq->rmw_tail = folio1;
			folio2 = folio1;
		}

		if (gap_before && !read_before)
			folio_zero_segment(folio1, 0, offset_in_folio(folio1, start));

		if (gap_after && !read_after)
			folio_zero_segment(folio2, offset_in_folio(folio2, end),
					   folio_size(folio2));

		if (read_before || read_after) {
			ret = netfs_rmw_read(wreq, iocb->ki_filp,
					     gstart, gap_before,
					     gend - gap_after, gap_after);
			if (ret < 0)
				goto error;
		}

		start = wreq->start;
		len = wreq->len;
	}

	/* Find out subreq limits. */
	netfs_prepare_write(wreq, upload, start);

	/* Copy the data in progressively, dispatching writes as we go. */
	while ((len = iov_iter_count(iter))) {
		struct iov_iter tmp = *iter;
		struct folio *folio;
		ssize_t copied;
		size_t foff, fsize, part;

		/* Get the bounce folio we're going to copy/encrypt the data
		 * to.  Note that we use multipage folios if bsize > PAGE_SIZE.
		 */
		ret = netfs_alloc_bounce(wreq, start + umin(len, bsize), GFP_KERNEL);
		if (ret < 0)
			break;

		if (slot == folioq_nr_slots(fq)) {
			fq = fq->next;
			slot = 0;
		}

		folio = folioq_folio(fq, slot);
		fsize = folioq_folio_size(fq, slot);
		foff = offset_in_folio(folio, start + gap_before);
		WARN_ON_ONCE(folio_pos(folio) > start);

		part = umin(fsize - foff, len);

		/* If this next stretch of data aligns with the crypto
		 * algorithm block size, we can get a free copy from the
		 * encryption algorithm.
		 */
		iov_iter_truncate(&tmp, part);
		if (!gap_before &&
		    !(gap_after && part == len) &&
		    netfs_is_crypto_aligned(wreq, &tmp)) {
			if (!netfs_encrypt_to_folio(wreq, folio, iter, start, part, GFP_KERNEL)) {
				ret = wreq->error;
				break;
			}
			copied = part;
		} else {
			copied = copy_folio_from_iter(folio, foff, part, iter);
			if (copied < part)
				break;
			copied += gap_before;
			if (part == len)
				copied += gap_after;
			if (!netfs_encrypt_folio(wreq, folio, folio, start, copied, GFP_KERNEL)) {
				ret = wreq->error;
				break;
			}
		}

		ret = netfs_write_bounce(wreq, folio, offset_in_folio(folio, start), copied);
		if (ret < 0)
			break;
		start += copied;
		if (part < len) {
			written += copied;
			written -= gap_before;
		} else {
			written += part;
		}
		if (foff + part >= fsize)
			slot++;
		gap_before = 0;
		cond_resched();
	}

error:
	netfs_issue_write(wreq, upload);
	smp_wmb(); /* Write lists before ALL_QUEUED. */
	set_bit(NETFS_RREQ_ALL_QUEUED, &wreq->flags);
	if (list_empty(&upload->subrequests))
		netfs_wake_collector(wreq);
	_leave(" = %d [%zx]", ret, written);
	return written ?: ret;
}

/*
 * Perform an unbuffered write where we may have to do an RMW operation on an
 * encrypted file.  This can also be used for direct I/O writes.
 */
ssize_t netfs_unbuffered_write_iter_locked(struct kiocb *iocb, struct iov_iter *iter,
						  struct netfs_group *netfs_group)
{
	struct netfs_io_request *wreq;
	struct netfs_inode *ctx = netfs_inode(file_inode(iocb->ki_filp));
	unsigned long long start = iocb->ki_pos;
	unsigned long long end = start + iov_iter_count(iter);
	ssize_t ret, n;
	size_t len = iov_iter_count(iter);
	bool async = !is_sync_kiocb(iocb);

	_enter("");

	_debug("uw %llx-%llx", start, end);

	wreq = netfs_create_write_req(iocb->ki_filp->f_mapping, iocb->ki_filp, start, NULL,
				      iocb->ki_flags & IOCB_DIRECT ?
				      NETFS_DIO_WRITE : NETFS_UNBUFFERED_WRITE);
	if (IS_ERR(wreq))
		return PTR_ERR(wreq);

	wreq->io_streams[0].avail = true;
	wreq->cleanup = netfs_cleanup_dio_write;
	wreq->len = iov_iter_count(iter);
	trace_netfs_write(wreq, (iocb->ki_flags & IOCB_DIRECT ?
				 netfs_write_trace_dio_write :
				 netfs_write_trace_unbuffered_write));

	/* We're going to need a bounce buffer if what we transmit is going to
	 * be different in some way to the source buffer, e.g. because it gets
	 * encrypted/compressed or because it needs expanding to a block size.
	 */
	if (test_bit(NETFS_ICTX_ENCRYPTED, &ctx->flags)) {
		ret = netfs_write_through_bounce_buffer(wreq, iocb, iter, netfs_group);
		goto done;
	}

	/* If this is an async op and we're not using a bounce buffer, we have
	 * to save the source buffer as the iterator is only good until we
	 * return.  In such a case, extract an iterator to represent as much of
	 * the the output buffer as we can manage.  Note that the extraction
	 * might not be able to allocate a sufficiently large bvec array and
	 * may shorten the request.
	 */
	if (user_backed_iter(iter)) {
		n = netfs_extract_user_iter(iter, len, &wreq->buffer.iter, 0);
		if (n < 0) {
			ret = n;
			goto out;
		}
		wreq->direct_bv = (struct bio_vec *)wreq->buffer.iter.bvec;
		wreq->direct_bv_count = n;
		wreq->direct_bv_unpin = iov_iter_extract_will_pin(iter);
	} else {
		/* If this is a kernel-generated async DIO request, assume that
		 * any resources the iterator points to (eg. a bio_vec array)
		 * will persist till the end of the op.
		 */
		wreq->buffer.iter = *iter;
	}

	__set_bit(NETFS_RREQ_USE_IO_ITER, &wreq->flags);
	if (async)
		__set_bit(NETFS_RREQ_OFFLOAD_COLLECTION, &wreq->flags);

	/* Dispatch the write. */
	__set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags);
	if (async)
		wreq->iocb = iocb;
	wreq->len = iov_iter_count(&wreq->buffer.iter);
	ret = netfs_unbuffered_write(wreq, is_sync_kiocb(iocb), wreq->len);
done:
	if (ret < 0) {
		_debug("begin = %zd", ret);
		goto out;
	}

	if (!async) {
		trace_netfs_rreq(wreq, netfs_rreq_trace_wait_ip);
		ret = netfs_wait_for_write(wreq);
		if (ret > 0)
			iocb->ki_pos += ret;
	} else {
		ret = -EIOCBQUEUED;
	}

out:
	netfs_put_request(wreq, netfs_rreq_trace_put_return);
	return ret;
}
EXPORT_SYMBOL(netfs_unbuffered_write_iter_locked);

/**
 * netfs_unbuffered_write_iter - Unbuffered write to a file
 * @iocb: IO state structure
 * @from: iov_iter with data to write
 *
 * Do an unbuffered write to a file, writing the data directly to the server
 * and not lodging the data in the pagecache.
 *
 * Return:
 * * Negative error code if no data has been written at all of
 *   vfs_fsync_range() failed for a synchronous write
 * * Number of bytes written, even for truncated writes
 */
ssize_t netfs_unbuffered_write_iter(struct kiocb *iocb, struct iov_iter *from)
{
	struct file *file = iocb->ki_filp;
	struct address_space *mapping = file->f_mapping;
	struct inode *inode = mapping->host;
	struct netfs_inode *ictx = netfs_inode(inode);
	ssize_t ret;
	loff_t pos = iocb->ki_pos;
	unsigned long long end = pos + iov_iter_count(from) - 1;

	_enter("%llx,%zx,%llx", pos, iov_iter_count(from), i_size_read(inode));

	if (!iov_iter_count(from))
		return 0;

	trace_netfs_write_iter(iocb, from);
	netfs_stat(&netfs_n_wh_dio_write);

	ret = netfs_start_io_direct(inode);
	if (ret < 0)
		return ret;
	ret = generic_write_checks(iocb, from);
	if (ret <= 0)
		goto out;
	ret = file_remove_privs(file);
	if (ret < 0)
		goto out;
	ret = file_update_time(file);
	if (ret < 0)
		goto out;
	if (iocb->ki_flags & IOCB_NOWAIT) {
		/* We could block if there are any pages in the range. */
		ret = -EAGAIN;
		if (filemap_range_has_page(mapping, pos, end))
			if (filemap_invalidate_inode(inode, true, pos, end))
				goto out;
	} else {
		ret = filemap_write_and_wait_range(mapping, pos, end);
		if (ret < 0)
			goto out;
	}

	/*
	 * After a write we want buffered reads to be sure to go to disk to get
	 * the new data.  We invalidate clean cached page from the region we're
	 * about to write.  We do this *before* the write so that we can return
	 * without clobbering -EIOCBQUEUED from ->direct_IO().
	 */
	ret = filemap_invalidate_inode(inode, true, pos, end);
	if (ret < 0)
		goto out;
	end = iocb->ki_pos + iov_iter_count(from);
	if (end > ictx->zero_point)
		ictx->zero_point = end;

	fscache_invalidate(netfs_i_cookie(ictx), NULL, i_size_read(inode),
			   FSCACHE_INVAL_DIO_WRITE);
	ret = netfs_unbuffered_write_iter_locked(iocb, from, NULL);
out:
	netfs_end_io_direct(inode);
	return ret;
}
EXPORT_SYMBOL(netfs_unbuffered_write_iter);