Add L2ARC Compression

https://www.illumos.org/issues/3137
This issue ticket proposes adding transparent L2ARC compression to
ZFS. The algorithm is hard-coded to be the new LZ4 compression
algorithm, since L2ARC has very specific performance requirements
(high decompression speed).
The feature is controlled via a new secondarycachecompress property
on each dataset, allowing the user to selectively enable
compression only on datasets which can use it.

Patch dated 17/01/2013.

include/sys/arc.h

@@ -20,6 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
  */
 
 #ifndef	_SYS_ARC_H
@@ -78,6 +79,7 @@ typedef enum arc_buf_contents {
 #define	ARC_PREFETCH	(1 << 3)	/* I/O is a prefetch */
 #define	ARC_CACHED	(1 << 4)	/* I/O was already in cache */
 #define	ARC_L2CACHE	(1 << 5)	/* cache in L2ARC */
+#define	ARC_L2COMPRESS	(1 << 6)	/* compress in L2ARC */
 
 /*
  * The following breakdows of arc_size exist for kstat only.
@@ -121,9 +123,9 @@ int arc_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bp,
     arc_done_func_t *done, void *private, int priority, int flags,
     uint32_t *arc_flags, const zbookmark_t *zb);
 zio_t *arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
-    blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
-    arc_done_func_t *ready, arc_done_func_t *done, void *private,
-    int priority, int zio_flags, const zbookmark_t *zb);
+    blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
+    const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *done,
+    void *private, int priority, int zio_flags, const zbookmark_t *zb);
 
 arc_prune_t *arc_add_prune_callback(arc_prune_func_t *func, void *private);
 void arc_remove_prune_callback(arc_prune_t *p);

include/sys/dbuf.h

@@ -20,6 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
  */
 
 #ifndef	_SYS_DBUF_H
@@ -322,6 +323,16 @@ boolean_t dbuf_is_metadata(dmu_buf_impl_t *db);
 	(DBUF_IS_METADATA(_db) &&					\
 	((_db)->db_objset->os_secondary_cache == ZFS_CACHE_METADATA)))
 
+#define	DBUF_IS_L2COMPRESSIBLE(_db)					\
+	((_db)->db_objset->os_secondary_cache_compress ==		\
+	ZFS_CACHE_COMPRESS_ALL || 					\
+	(DBUF_IS_METADATA(_db) &&					\
+	(_db)->db_objset->os_secondary_cache_compress ==		\
+	ZFS_CACHE_COMPRESS_METADATA) || 				\
+	(!DBUF_IS_METADATA(_db) &&					\
+	(_db)->db_objset->os_secondary_cache_compress ==		\
+	ZFS_CACHE_COMPRESS_DATA))
+
 #ifdef ZFS_DEBUG
 
 /*

include/sys/dmu_objset.h

@@ -20,6 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
  */
 
 /* Portions Copyright 2010 Robert Milkowski */
@@ -90,6 +91,7 @@ struct objset {
 	uint8_t os_logbias;
 	uint8_t os_primary_cache;
 	uint8_t os_secondary_cache;
+	uint8_t os_secondary_cache_compress;	/* L2 compression */
 	uint8_t os_sync;
 
 	/* no lock needed: */
@@ -129,6 +131,11 @@ struct objset {
 	((os)->os_secondary_cache == ZFS_CACHE_ALL ||		\
 	(os)->os_secondary_cache == ZFS_CACHE_METADATA)
 
+#define	DMU_OS_IS_L2COMPRESSIBLE(os)					\
+	((os)->os_secondary_cache_compress == ZFS_CACHE_COMPRESS_ALL ||	\
+	(os)->os_secondary_cache_compress == ZFS_CACHE_COMPRESS_DATA ||	\
+	(os)->os_secondary_cache_compress == ZFS_CACHE_COMPRESS_METADATA)
+
 /* called from zpl */
 int dmu_objset_hold(const char *name, void *tag, objset_t **osp);
 int dmu_objset_own(const char *name, dmu_objset_type_t type,

include/sys/fs/zfs.h

@@ -24,6 +24,7 @@
  * Copyright (c) 2012 by Delphix. All rights reserved.
  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
  * Copyright (c) 2012, Joyent, Inc. All rights reserved.
+ * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
  */
 
 /* Portions Copyright 2010 Robert Milkowski */
@@ -128,6 +129,7 @@ typedef enum {
 	ZFS_PROP_REFRATIO,
 	ZFS_PROP_WRITTEN,
 	ZFS_PROP_CLONES,
+	ZFS_PROP_SECONDARYCACHECOMPRESS,
 	ZFS_NUM_PROPS
 } zfs_prop_t;
 
@@ -316,6 +318,13 @@ typedef enum zfs_cache_type {
 	ZFS_CACHE_ALL = 2
 } zfs_cache_type_t;
 
+typedef enum zfs_cache_compress_type {
+	ZFS_CACHE_COMPRESS_NONE = 0,
+	ZFS_CACHE_COMPRESS_DATA = 1,
+	ZFS_CACHE_COMPRESS_METADATA = 2,
+	ZFS_CACHE_COMPRESS_ALL = 3
+} zfs_cache_compress_type_t;
+
 typedef enum {
 	ZFS_SYNC_STANDARD = 0,
 	ZFS_SYNC_ALWAYS = 1,

include/sys/zio.h

@@ -480,6 +480,11 @@ extern zio_t *zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset,
     zio_done_func_t *done, void *private, int priority, enum zio_flag flags,
     boolean_t labels);
 
+extern zio_t *zio_read_null(zio_t *pio, vdev_t *vd, uint64_t offset,
+    uint64_t size, void *data, int checksum,
+    zio_done_func_t *done, void *private, int priority, enum zio_flag flags,
+    boolean_t labels);
+
 extern zio_t *zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset,
     uint64_t size, void *data, int checksum,
     zio_done_func_t *done, void *private, int priority, enum zio_flag flags,

man/man8/zfs.8

@@ -1004,6 +1004,32 @@ This property can also be referred to by its shortened column name, \fBreserv\fR
 Controls what is cached in the secondary cache (L2ARC). If this property is set to \fBall\fR, then both user data and metadata is cached. If this property is set to \fBnone\fR, then neither user data nor metadata is cached. If this property is set to \fBmetadata\fR, then only metadata is cached. The default value is \fBall\fR.
 .RE
 
+.sp
+.ne 2
+.mk
+.na
+\fB\fBsecondarycachecompress\fR=\fBnone\fR | \fBall\fR | \fBmetadata\fR | \fBdata\fR\fR
+.ad
+.sp .6
+.RS 4n
+Controls compression of data on L2ARC devices. The default value is
+\fBnone\fR, which means that no compression is requested for any data
+when it enters the L2ARC. When set to \fBall\fR, all data from this
+dataset is compressed when put into L2ARC and decompressed on cache hit.
+When set to \fBmetadata\fR, only metadata is compressed and conversely
+when set to \fBdata\fR, only user data is compressed. When multiple
+datasets with different values for \fBsecondarycachecompress\fR
+reference the same data in ARC, the data is compressed if at least one
+dataset requests it (i.e. compression is preferred).
+
+Buffers smaller than 2048 bytes (or 4096 bytes when using L2ARC devices
+with 4k sectors) or buffers that cannot be compressed by at least 12.5%
+will be stored uncompressed and thus do not incur any performance
+penalty on cache hits. Please note that this means buffers from datasets
+with a \fBrecordsize\fR smaller than or equal to this limit cannot be
+compressed in L2ARC.
+.RE
+
 .sp
 .ne 2
 .mk

module/zcommon/zfs_prop.c

@@ -180,6 +180,14 @@ zfs_prop_init(void)
 		{ NULL }
 	};
 
+	static zprop_index_t cache_compress_table[] = {
+		{ "none",	ZFS_CACHE_COMPRESS_NONE },
+		{ "metadata",	ZFS_CACHE_COMPRESS_METADATA },
+		{ "data",	ZFS_CACHE_COMPRESS_DATA },
+		{ "all",	ZFS_CACHE_COMPRESS_ALL },
+		{ NULL }
+	};
+
 	static zprop_index_t sync_table[] = {
 		{ "standard",	ZFS_SYNC_STANDARD },
 		{ "always",	ZFS_SYNC_ALWAYS },
@@ -232,6 +240,11 @@ zfs_prop_init(void)
 	    ZFS_CACHE_ALL, PROP_INHERIT,
 	    ZFS_TYPE_FILESYSTEM | ZFS_TYPE_SNAPSHOT | ZFS_TYPE_VOLUME,
 	    "all | none | metadata", "SECONDARYCACHE", cache_table);
+	zprop_register_index(ZFS_PROP_SECONDARYCACHECOMPRESS,
+	    "secondarycachecompress", ZFS_CACHE_COMPRESS_NONE, PROP_INHERIT,
+	    ZFS_TYPE_FILESYSTEM | ZFS_TYPE_SNAPSHOT | ZFS_TYPE_VOLUME,
+	    "all | none | metadata | data", "SECONDARYCACHECOMPRESS",
+	    cache_compress_table);
 	zprop_register_index(ZFS_PROP_LOGBIAS, "logbias", ZFS_LOGBIAS_LATENCY,
 	    PROP_INHERIT, ZFS_TYPE_FILESYSTEM | ZFS_TYPE_VOLUME,
 	    "latency | throughput", "LOGBIAS", logbias_table);