file.c

linux 内核源代码

			if (unlikely(err)) {
				if (err == -EEXIST)
					continue;
				goto err_out;
			}
			pages[nr] = *cached_page;
			page_cache_get(*cached_page);
			if (unlikely(!pagevec_add(lru_pvec, *cached_page)))
				__pagevec_lru_add(lru_pvec);
			*cached_page = NULL;
		}
		index++;
		nr++;
	} while (nr < nr_pages);
out:
	return err;
err_out:
	while (nr > 0) {
		unlock_page(pages[--nr]);
		page_cache_release(pages[nr]);
	}
	goto out;
}

static inline int ntfs_submit_bh_for_read(struct buffer_head *bh)
{
	lock_buffer(bh);
	get_bh(bh);
	bh->b_end_io = end_buffer_read_sync;
	return submit_bh(READ, bh);
}

/**
 * ntfs_prepare_pages_for_non_resident_write - prepare pages for receiving data
 * @pages:	array of destination pages
 * @nr_pages:	number of pages in @pages
 * @pos:	byte position in file at which the write begins
 * @bytes:	number of bytes to be written
 *
 * This is called for non-resident attributes from ntfs_file_buffered_write()
 * with i_mutex held on the inode (@pages[0]->mapping->host).  There are
 * @nr_pages pages in @pages which are locked but not kmap()ped.  The source
 * data has not yet been copied into the @pages.
 * 
 * Need to fill any holes with actual clusters, allocate buffers if necessary,
 * ensure all the buffers are mapped, and bring uptodate any buffers that are
 * only partially being written to.
 *
 * If @nr_pages is greater than one, we are guaranteed that the cluster size is
 * greater than PAGE_CACHE_SIZE, that all pages in @pages are entirely inside
 * the same cluster and that they are the entirety of that cluster, and that
 * the cluster is sparse, i.e. we need to allocate a cluster to fill the hole.
 *
 * i_size is not to be modified yet.
 *
 * Return 0 on success or -errno on error.
 */
static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
		unsigned nr_pages, s64 pos, size_t bytes)
{
	VCN vcn, highest_vcn = 0, cpos, cend, bh_cpos, bh_cend;
	LCN lcn;
	s64 bh_pos, vcn_len, end, initialized_size;
	sector_t lcn_block;
	struct page *page;
	struct inode *vi;
	ntfs_inode *ni, *base_ni = NULL;
	ntfs_volume *vol;
	runlist_element *rl, *rl2;
	struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
	ntfs_attr_search_ctx *ctx = NULL;
	MFT_RECORD *m = NULL;
	ATTR_RECORD *a = NULL;
	unsigned long flags;
	u32 attr_rec_len = 0;
	unsigned blocksize, u;
	int err, mp_size;
	bool rl_write_locked, was_hole, is_retry;
	unsigned char blocksize_bits;
	struct {
		u8 runlist_merged:1;
		u8 mft_attr_mapped:1;
		u8 mp_rebuilt:1;
		u8 attr_switched:1;
	} status = { 0, 0, 0, 0 };

	BUG_ON(!nr_pages);
	BUG_ON(!pages);
	BUG_ON(!*pages);
	vi = pages[0]->mapping->host;
	ni = NTFS_I(vi);
	vol = ni->vol;
	ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
			"index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
			vi->i_ino, ni->type, pages[0]->index, nr_pages,
			(long long)pos, bytes);
	blocksize = vol->sb->s_blocksize;
	blocksize_bits = vol->sb->s_blocksize_bits;
	u = 0;
	do {
		page = pages[u];
		BUG_ON(!page);
		/*
		 * create_empty_buffers() will create uptodate/dirty buffers if
		 * the page is uptodate/dirty.
		 */
		if (!page_has_buffers(page)) {
			create_empty_buffers(page, blocksize, 0);
			if (unlikely(!page_has_buffers(page)))
				return -ENOMEM;
		}
	} while (++u < nr_pages);
	rl_write_locked = false;
	rl = NULL;
	err = 0;
	vcn = lcn = -1;
	vcn_len = 0;
	lcn_block = -1;
	was_hole = false;
	cpos = pos >> vol->cluster_size_bits;
	end = pos + bytes;
	cend = (end + vol->cluster_size - 1) >> vol->cluster_size_bits;
	/*
	 * Loop over each page and for each page over each buffer.  Use goto to
	 * reduce indentation.
	 */
	u = 0;
do_next_page:
	page = pages[u];
	bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
	bh = head = page_buffers(page);
	do {
		VCN cdelta;
		s64 bh_end;
		unsigned bh_cofs;

		/* Clear buffer_new on all buffers to reinitialise state. */
		if (buffer_new(bh))
			clear_buffer_new(bh);
		bh_end = bh_pos + blocksize;
		bh_cpos = bh_pos >> vol->cluster_size_bits;
		bh_cofs = bh_pos & vol->cluster_size_mask;
		if (buffer_mapped(bh)) {
			/*
			 * The buffer is already mapped.  If it is uptodate,
			 * ignore it.
			 */
			if (buffer_uptodate(bh))
				continue;
			/*
			 * The buffer is not uptodate.  If the page is uptodate
			 * set the buffer uptodate and otherwise ignore it.
			 */
			if (PageUptodate(page)) {
				set_buffer_uptodate(bh);
				continue;
			}
			/*
			 * Neither the page nor the buffer are uptodate.  If
			 * the buffer is only partially being written to, we
			 * need to read it in before the write, i.e. now.
			 */
			if ((bh_pos < pos && bh_end > pos) ||
					(bh_pos < end && bh_end > end)) {
				/*
				 * If the buffer is fully or partially within
				 * the initialized size, do an actual read.
				 * Otherwise, simply zero the buffer.
				 */
				read_lock_irqsave(&ni->size_lock, flags);
				initialized_size = ni->initialized_size;
				read_unlock_irqrestore(&ni->size_lock, flags);
				if (bh_pos < initialized_size) {
					ntfs_submit_bh_for_read(bh);
					*wait_bh++ = bh;
				} else {
					zero_user_page(page, bh_offset(bh),
							blocksize, KM_USER0);
					set_buffer_uptodate(bh);
				}
			}
			continue;
		}
		/* Unmapped buffer.  Need to map it. */
		bh->b_bdev = vol->sb->s_bdev;
		/*
		 * If the current buffer is in the same clusters as the map
		 * cache, there is no need to check the runlist again.  The
		 * map cache is made up of @vcn, which is the first cached file
		 * cluster, @vcn_len which is the number of cached file
		 * clusters, @lcn is the device cluster corresponding to @vcn,
		 * and @lcn_block is the block number corresponding to @lcn.
		 */
		cdelta = bh_cpos - vcn;
		if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
map_buffer_cached:
			BUG_ON(lcn < 0);
			bh->b_blocknr = lcn_block +
					(cdelta << (vol->cluster_size_bits -
					blocksize_bits)) +
					(bh_cofs >> blocksize_bits);
			set_buffer_mapped(bh);
			/*
			 * If the page is uptodate so is the buffer.  If the
			 * buffer is fully outside the write, we ignore it if
			 * it was already allocated and we mark it dirty so it
			 * gets written out if we allocated it.  On the other
			 * hand, if we allocated the buffer but we are not
			 * marking it dirty we set buffer_new so we can do
			 * error recovery.
			 */
			if (PageUptodate(page)) {
				if (!buffer_uptodate(bh))
					set_buffer_uptodate(bh);
				if (unlikely(was_hole)) {
					/* We allocated the buffer. */
					unmap_underlying_metadata(bh->b_bdev,
							bh->b_blocknr);
					if (bh_end <= pos || bh_pos >= end)
						mark_buffer_dirty(bh);
					else
						set_buffer_new(bh);
				}
				continue;
			}
			/* Page is _not_ uptodate. */
			if (likely(!was_hole)) {
				/*
				 * Buffer was already allocated.  If it is not
				 * uptodate and is only partially being written
				 * to, we need to read it in before the write,
				 * i.e. now.
				 */
				if (!buffer_uptodate(bh) && bh_pos < end &&
						bh_end > pos &&
						(bh_pos < pos ||
						bh_end > end)) {
					/*
					 * If the buffer is fully or partially
					 * within the initialized size, do an
					 * actual read.  Otherwise, simply zero
					 * the buffer.
					 */
					read_lock_irqsave(&ni->size_lock,
							flags);
					initialized_size = ni->initialized_size;
					read_unlock_irqrestore(&ni->size_lock,
							flags);
					if (bh_pos < initialized_size) {
						ntfs_submit_bh_for_read(bh);
						*wait_bh++ = bh;
					} else {
						zero_user_page(page,
							bh_offset(bh),
							blocksize, KM_USER0);
						set_buffer_uptodate(bh);
					}
				}
				continue;
			}
			/* We allocated the buffer. */
			unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
			/*
			 * If the buffer is fully outside the write, zero it,
			 * set it uptodate, and mark it dirty so it gets
			 * written out.  If it is partially being written to,
			 * zero region surrounding the write but leave it to
			 * commit write to do anything else.  Finally, if the
			 * buffer is fully being overwritten, do nothing.
			 */
			if (bh_end <= pos || bh_pos >= end) {
				if (!buffer_uptodate(bh)) {
					zero_user_page(page, bh_offset(bh),
							blocksize, KM_USER0);
					set_buffer_uptodate(bh);
				}
				mark_buffer_dirty(bh);
				continue;
			}
			set_buffer_new(bh);
			if (!buffer_uptodate(bh) &&
					(bh_pos < pos || bh_end > end)) {
				u8 *kaddr;
				unsigned pofs;
					
				kaddr = kmap_atomic(page, KM_USER0);
				if (bh_pos < pos) {
					pofs = bh_pos & ~PAGE_CACHE_MASK;
					memset(kaddr + pofs, 0, pos - bh_pos);
				}
				if (bh_end > end) {
					pofs = end & ~PAGE_CACHE_MASK;
					memset(kaddr + pofs, 0, bh_end - end);
				}
				kunmap_atomic(kaddr, KM_USER0);
				flush_dcache_page(page);
			}
			continue;
		}
		/*
		 * Slow path: this is the first buffer in the cluster.  If it
		 * is outside allocated size and is not uptodate, zero it and
		 * set it uptodate.
		 */
		read_lock_irqsave(&ni->size_lock, flags);
		initialized_size = ni->allocated_size;
		read_unlock_irqrestore(&ni->size_lock, flags);
		if (bh_pos > initialized_size) {
			if (PageUptodate(page)) {
				if (!buffer_uptodate(bh))
					set_buffer_uptodate(bh);
			} else if (!buffer_uptodate(bh)) {
				zero_user_page(page, bh_offset(bh), blocksize,
						KM_USER0);
				set_buffer_uptodate(bh);
			}
			continue;
		}
		is_retry = false;
		if (!rl) {
			down_read(&ni->runlist.lock);
retry_remap:
			rl = ni->runlist.rl;
		}
		if (likely(rl != NULL)) {
			/* Seek to element containing target cluster. */
			while (rl->length && rl[1].vcn <= bh_cpos)
				rl++;
			lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
			if (likely(lcn >= 0)) {
				/*
				 * Successful remap, setup the map cache and
				 * use that to deal with the buffer.
				 */
				was_hole = false;
				vcn = bh_cpos;
				vcn_len = rl[1].vcn - vcn;
				lcn_block = lcn << (vol->cluster_size_bits -
						blocksize_bits);
				cdelta = 0;
				/*
				 * If the number of remaining clusters touched
				 * by the write is smaller or equal to the
				 * number of cached clusters, unlock the
				 * runlist as the map cache will be used from
				 * now on.
				 */
				if (likely(vcn + vcn_len >= cend)) {
					if (rl_write_locked) {
						up_write(&ni->runlist.lock);
						rl_write_locked = false;
					} else
						up_read(&ni->runlist.lock);
					rl = NULL;
				}
				goto map_buffer_cached;
			}
		} else
			lcn = LCN_RL_NOT_MAPPED;
		/*
		 * If it is not a hole and not out of bounds, the runlist is
		 * probably unmapped so try to map it now.
		 */
		if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
			if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
				/* Attempt to map runlist. */
				if (!rl_write_locked) {
					/*
					 * We need the runlist locked for
					 * writing, so if it is locked for
					 * reading relock it now and retry in
					 * case it changed whilst we dropped
					 * the lock.
					 */
					up_read(&ni->runlist.lock);
					down_write(&ni->runlist.lock);
					rl_write_locked = true;
					goto retry_remap;
				}
				err = ntfs_map_runlist_nolock(ni, bh_cpos,
						NULL);
				if (likely(!err)) {
					is_retry = true;
					goto retry_remap;
				}
				/*
				 * If @vcn is out of bounds, pretend @lcn is
				 * LCN_ENOENT.  As long as the buffer is out
				 * of bounds this will work fine.
				 */
				if (err == -ENOENT) {
					lcn = LCN_ENOENT;
					err = 0;
					goto rl_not_mapped_enoent;
				}
			} else
				err = -EIO;
			/* Failed to map the buffer, even after retrying. */
			bh->b_blocknr = -1;
			ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
					"attribute type 0x%x, vcn 0x%llx, "
					"vcn offset 0x%x, because its "
					"location on disk could not be "
					"determined%s (error code %i).",
					ni->mft_no, ni->type,
					(unsigned long long)bh_cpos,
					(unsigned)bh_pos &
					vol->cluster_size_mask,
					is_retry ? " even after retrying" : "",
					err);
			break;
		}
rl_not_mapped_enoent:
		/*
		 * The buffer is in a hole or out of bounds.  We need to fill
		 * the hole, unless the buffer is in a cluster which is not
		 * touched by the write, in which case we just leave the buffer
		 * unmapped.  This can only happen when the cluster size is
		 * less than the page cache size.
		 */
		if (unlikely(vol->cluster_size < PAGE_CACHE_SIZE)) {
			bh_cend = (bh_end + vol->cluster_size - 1) >>
					vol->cluster_size_bits;
			if ((bh_cend <= cpos || bh_cpos >= cend)) {
				bh->b_blocknr = -1;
				/*
				 * If the buffer is uptodate we skip it.  If it
				 * is not but the page is uptodate, we can set
				 * the buffer uptodate.  If the page is not
				 * uptodate, we can clear the buffer and set it
				 * uptodate.  Whether this is worthwhile is
				 * debatable and this could be removed.
				 */