Squashfs: Directly decompress into the page cache for file data

This introduces an implementation of squashfs_readpage_block()
that directly decompresses into the page cache.

This uses the previously added page handler abstraction to push
down the necessary kmap_atomic/kunmap_atomic operations on the
page cache buffers into the decompressors.  This enables
direct copying into the page cache without using the slow
kmap/kunmap calls.

The code detects when multiple threads are racing in
squashfs_readpage() to decompress the same block, and avoids
this regression by falling back to using an intermediate
buffer.

This patch enhances the performance of Squashfs significantly
when multiple processes are accessing the filesystem simultaneously
because it not only reduces memcopying, but it more importantly
eliminates the lock contention on the intermediate buffer.

Using single-thread decompression.

        dd if=file1 of=/dev/null bs=4096 &
        dd if=file2 of=/dev/null bs=4096 &
        dd if=file3 of=/dev/null bs=4096 &
        dd if=file4 of=/dev/null bs=4096

Before:

629145600 bytes (629 MB) copied, 45.8046 s, 13.7 MB/s

After:

629145600 bytes (629 MB) copied, 9.29414 s, 67.7 MB/s

Signed-off-by: Phillip Lougher <phillip@squashfs.org.uk>
Reviewed-by: Minchan Kim <minchan@kernel.org>

5 files changed, 336 insertions, 1 deletions

diff --git a/fs/squashfs/Kconfig b/fs/squashfs/Kconfig
index 159bd66..b6fa865 100644
--- a/fs/squashfs/Kconfig
+++ b/fs/squashfs/Kconfig
@@ -26,6 +26,34 @@ config SQUASHFS
 	  If unsure, say N.
 
 choice
+	prompt "File decompression options"
+	depends on SQUASHFS
+	help
+	  Squashfs now supports two options for decompressing file
+	  data.  Traditionally Squashfs has decompressed into an
+	  intermediate buffer and then memcopied it into the page cache.
+	  Squashfs now supports the ability to decompress directly into
+	  the page cache.
+
+	  If unsure, select "Decompress file data into an intermediate buffer"
+
+config SQUASHFS_FILE_CACHE
+	bool "Decompress file data into an intermediate buffer"
+	help
+	  Decompress file data into an intermediate buffer and then
+	  memcopy it into the page cache.
+
+config SQUASHFS_FILE_DIRECT
+	bool "Decompress files directly into the page cache"
+	help
+	  Directly decompress file data into the page cache.
+	  Doing so can significantly improve performance because
+	  it eliminates a memcpy and it also removes the lock contention
+	  on the single buffer.
+
+endchoice
+
+choice
 	prompt "Decompressor parallelisation options"
 	depends on SQUASHFS
 	help
diff --git a/fs/squashfs/Makefile b/fs/squashfs/Makefile
index e01ba11..4132520 100644
--- a/fs/squashfs/Makefile
+++ b/fs/squashfs/Makefile
@@ -4,7 +4,9 @@
 
 obj-$(CONFIG_SQUASHFS) += squashfs.o
 squashfs-y += block.o cache.o dir.o export.o file.o fragment.o id.o inode.o
-squashfs-y += namei.o super.o symlink.o decompressor.o file_cache.o
+squashfs-y += namei.o super.o symlink.o decompressor.o
+squashfs-$(CONFIG_SQUASHFS_FILE_CACHE) += file_cache.o
+squashfs-$(CONFIG_SQUASHFS_FILE_DIRECT) += file_direct.o page_actor.o
 squashfs-$(CONFIG_SQUASHFS_DECOMP_SINGLE) += decompressor_single.o
 squashfs-$(CONFIG_SQUASHFS_DECOMP_MULTI) += decompressor_multi.o
 squashfs-$(CONFIG_SQUASHFS_DECOMP_MULTI_PERCPU) += decompressor_multi_percpu.o
diff --git a/fs/squashfs/file_direct.c b/fs/squashfs/file_direct.c
new file mode 100644
index 0000000..2943b2b
--- /dev/null
+++ b/fs/squashfs/file_direct.c
@@ -0,0 +1,173 @@
+/*
+ * Copyright (c) 2013
+ * Phillip Lougher <phillip@squashfs.org.uk>
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2. See
+ * the COPYING file in the top-level directory.
+ */
+
+#include <linux/fs.h>
+#include <linux/vfs.h>
+#include <linux/kernel.h>
+#include <linux/slab.h>
+#include <linux/string.h>
+#include <linux/pagemap.h>
+#include <linux/mutex.h>
+
+#include "squashfs_fs.h"
+#include "squashfs_fs_sb.h"
+#include "squashfs_fs_i.h"
+#include "squashfs.h"
+#include "page_actor.h"
+
+static int squashfs_read_cache(struct page *target_page, u64 block, int bsize,
+	int pages, struct page **page);
+
+/* Read separately compressed datablock directly into page cache */
+int squashfs_readpage_block(struct page *target_page, u64 block, int bsize)
+
+{
+	struct inode *inode = target_page->mapping->host;
+	struct squashfs_sb_info *msblk = inode->i_sb->s_fs_info;
+
+	int file_end = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
+	int mask = (1 << (msblk->block_log - PAGE_CACHE_SHIFT)) - 1;
+	int start_index = target_page->index & ~mask;
+	int end_index = start_index | mask;
+	int i, n, pages, missing_pages, bytes, res = -ENOMEM;
+	struct page **page;
+	struct squashfs_page_actor *actor;
+	void *pageaddr;
+
+	if (end_index > file_end)
+		end_index = file_end;
+
+	pages = end_index - start_index + 1;
+
+	page = kmalloc(sizeof(void *) * pages, GFP_KERNEL);
+	if (page == NULL)
+		return res;
+
+	/*
+	 * Create a "page actor" which will kmap and kunmap the
+	 * page cache pages appropriately within the decompressor
+	 */
+	actor = squashfs_page_actor_init_special(page, pages, 0);
+	if (actor == NULL)
+		goto out;
+
+	/* Try to grab all the pages covered by the Squashfs block */
+	for (missing_pages = 0, i = 0, n = start_index; i < pages; i++, n++) {
+		page[i] = (n == target_page->index) ? target_page :
+			grab_cache_page_nowait(target_page->mapping, n);
+
+		if (page[i] == NULL) {
+			missing_pages++;
+			continue;
+		}
+
+		if (PageUptodate(page[i])) {
+			unlock_page(page[i]);
+			page_cache_release(page[i]);
+			page[i] = NULL;
+			missing_pages++;
+		}
+	}
+
+	if (missing_pages) {
+		/*
+		 * Couldn't get one or more pages, this page has either
+		 * been VM reclaimed, but others are still in the page cache
+		 * and uptodate, or we're racing with another thread in
+		 * squashfs_readpage also trying to grab them.  Fall back to
+		 * using an intermediate buffer.
+		 */
+		res = squashfs_read_cache(target_page, block, bsize, pages,
+								page);
+		goto out;
+	}
+
+	/* Decompress directly into the page cache buffers */
+	res = squashfs_read_data(inode->i_sb, block, bsize, NULL, actor);
+	if (res < 0)
+		goto mark_errored;
+
+	/* Last page may have trailing bytes not filled */
+	bytes = res % PAGE_CACHE_SIZE;
+	if (bytes) {
+		pageaddr = kmap_atomic(page[pages - 1]);
+		memset(pageaddr + bytes, 0, PAGE_CACHE_SIZE - bytes);
+		kunmap_atomic(pageaddr);
+	}
+
+	/* Mark pages as uptodate, unlock and release */
+	for (i = 0; i < pages; i++) {
+		flush_dcache_page(page[i]);
+		SetPageUptodate(page[i]);
+		unlock_page(page[i]);
+		if (page[i] != target_page)
+			page_cache_release(page[i]);
+	}
+
+	kfree(actor);
+	kfree(page);
+
+	return 0;
+
+mark_errored:
+	/* Decompression failed, mark pages as errored.  Target_page is
+	 * dealt with by the caller
+	 */
+	for (i = 0; i < pages; i++) {
+		if (page[i] == target_page)
+			continue;
+		flush_dcache_page(page[i]);
+		SetPageError(page[i]);
+		unlock_page(page[i]);
+		page_cache_release(page[i]);
+	}
+
+out:
+	kfree(actor);
+	kfree(page);
+	return res;
+}
+
+
+static int squashfs_read_cache(struct page *target_page, u64 block, int bsize,
+	int pages, struct page **page)
+{
+	struct inode *i = target_page->mapping->host;
+	struct squashfs_cache_entry *buffer = squashfs_get_datablock(i->i_sb,
+						 block, bsize);
+	int bytes = buffer->length, res = buffer->error, n, offset = 0;
+	void *pageaddr;
+
+	if (res) {
+		ERROR("Unable to read page, block %llx, size %x\n", block,
+			bsize);
+		goto out;
+	}
+
+	for (n = 0; n < pages && bytes > 0; n++,
+			bytes -= PAGE_CACHE_SIZE, offset += PAGE_CACHE_SIZE) {
+		int avail = min_t(int, bytes, PAGE_CACHE_SIZE);
+
+		if (page[n] == NULL)
+			continue;
+
+		pageaddr = kmap_atomic(page[n]);
+		squashfs_copy_data(pageaddr, buffer, offset, avail);
+		memset(pageaddr + avail, 0, PAGE_CACHE_SIZE - avail);
+		kunmap_atomic(pageaddr);
+		flush_dcache_page(page[n]);
+		SetPageUptodate(page[n]);
+		unlock_page(page[n]);
+		if (page[n] != target_page)
+			page_cache_release(page[n]);
+	}
+
+out:
+	squashfs_cache_put(buffer);
+	return res;
+}
diff --git a/fs/squashfs/page_actor.c b/fs/squashfs/page_actor.c
new file mode 100644
index 0000000..5a1c11f
--- /dev/null
+++ b/fs/squashfs/page_actor.c
@@ -0,0 +1,100 @@
+/*
+ * Copyright (c) 2013
+ * Phillip Lougher <phillip@squashfs.org.uk>
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2. See
+ * the COPYING file in the top-level directory.
+ */
+
+#include <linux/kernel.h>
+#include <linux/slab.h>
+#include <linux/pagemap.h>
+#include "page_actor.h"
+
+/*
+ * This file contains implementations of page_actor for decompressing into
+ * an intermediate buffer, and for decompressing directly into the
+ * page cache.
+ *
+ * Calling code should avoid sleeping between calls to squashfs_first_page()
+ * and squashfs_finish_page().
+ */
+
+/* Implementation of page_actor for decompressing into intermediate buffer */
+static void *cache_first_page(struct squashfs_page_actor *actor)
+{
+	actor->next_page = 1;
+	return actor->buffer[0];
+}
+
+static void *cache_next_page(struct squashfs_page_actor *actor)
+{
+	if (actor->next_page == actor->pages)
+		return NULL;
+
+	return actor->buffer[actor->next_page++];
+}
+
+static void cache_finish_page(struct squashfs_page_actor *actor)
+{
+	/* empty */
+}
+
+struct squashfs_page_actor *squashfs_page_actor_init(void **buffer,
+	int pages, int length)
+{
+	struct squashfs_page_actor *actor = kmalloc(sizeof(*actor), GFP_KERNEL);
+
+	if (actor == NULL)
+		return NULL;
+
+	actor->length = length ? : pages * PAGE_CACHE_SIZE;
+	actor->buffer = buffer;
+	actor->pages = pages;
+	actor->next_page = 0;
+	actor->squashfs_first_page = cache_first_page;
+	actor->squashfs_next_page = cache_next_page;
+	actor->squashfs_finish_page = cache_finish_page;
+	return actor;
+}
+
+/* Implementation of page_actor for decompressing directly into page cache. */
+static void *direct_first_page(struct squashfs_page_actor *actor)
+{
+	actor->next_page = 1;
+	return actor->pageaddr = kmap_atomic(actor->page[0]);
+}
+
+static void *direct_next_page(struct squashfs_page_actor *actor)
+{
+	if (actor->pageaddr)
+		kunmap_atomic(actor->pageaddr);
+
+	return actor->pageaddr = actor->next_page == actor->pages ? NULL :
+		kmap_atomic(actor->page[actor->next_page++]);
+}
+
+static void direct_finish_page(struct squashfs_page_actor *actor)
+{
+	if (actor->pageaddr)
+		kunmap_atomic(actor->pageaddr);
+}
+
+struct squashfs_page_actor *squashfs_page_actor_init_special(struct page **page,
+	int pages, int length)
+{
+	struct squashfs_page_actor *actor = kmalloc(sizeof(*actor), GFP_KERNEL);
+
+	if (actor == NULL)
+		return NULL;
+
+	actor->length = length ? : pages * PAGE_CACHE_SIZE;
+	actor->page = page;
+	actor->pages = pages;
+	actor->next_page = 0;
+	actor->pageaddr = NULL;
+	actor->squashfs_first_page = direct_first_page;
+	actor->squashfs_next_page = direct_next_page;
+	actor->squashfs_finish_page = direct_finish_page;
+	return actor;
+}
diff --git a/fs/squashfs/page_actor.h b/fs/squashfs/page_actor.h
index 5b0ba5a..26dd820 100644
--- a/fs/squashfs/page_actor.h
+++ b/fs/squashfs/page_actor.h
@@ -8,6 +8,7 @@
  * the COPYING file in the top-level directory.
  */
 
+#ifndef CONFIG_SQUASHFS_FILE_DIRECT
 struct squashfs_page_actor {
 	void	**page;
 	int	pages;
@@ -46,4 +47,35 @@ static inline void squashfs_finish_page(struct squashfs_page_actor *actor)
 {
 	/* empty */
 }
+#else
+struct squashfs_page_actor {
+	union {
+		void		**buffer;
+		struct page	**page;
+	};
+	void	*pageaddr;
+	void    *(*squashfs_first_page)(struct squashfs_page_actor *);
+	void    *(*squashfs_next_page)(struct squashfs_page_actor *);
+	void    (*squashfs_finish_page)(struct squashfs_page_actor *);
+	int	pages;
+	int	length;
+	int	next_page;
+};
+
+extern struct squashfs_page_actor *squashfs_page_actor_init(void **, int, int);
+extern struct squashfs_page_actor *squashfs_page_actor_init_special(struct page
+							 **, int, int);
+static inline void *squashfs_first_page(struct squashfs_page_actor *actor)
+{
+	return actor->squashfs_first_page(actor);
+}
+static inline void *squashfs_next_page(struct squashfs_page_actor *actor)
+{
+	return actor->squashfs_next_page(actor);
+}
+static inline void squashfs_finish_page(struct squashfs_page_actor *actor)
+{
+	actor->squashfs_finish_page(actor);
+}
+#endif
 #endif