Add L2ARC arcstats for MFU/MRU buffers and buffer content type

Currently the initial ARC state (MFU/MRU) of cached L2ARC buffers at the
time of their caching is unknown. The same applies for the buffer
content type.

This commit adds L2ARC arcstats that display the aligned size (in bytes)
of L2ARC buffers according to their content type (data/metadata) and
according to their ARC state (MRU/MFU) at the time of caching in L2ARC.

This may prove beneficial in adjusting the L2ARC caching policy in the
future.

Signed-off-by: George Amanakis <[email protected]>

Author: gamanakis

cmd/zdb/zdb.c

@@ -4101,6 +4101,8 @@ dump_l2arc_log_entries(uint64_t log_entries,
 		    (u_longlong_t)L2BLK_GET_PREFETCH((&le[j])->le_prop));
 		(void) printf("|\t\t\t\taddress: %llu\n",
 		    (u_longlong_t)le[j].le_daddr);
+		(void) printf("|\t\t\t\tstate: %llu\n",
+		    (u_longlong_t)L2BLK_GET_STATE((&le[j])->le_prop));
 		(void) printf("|\n");
 	}
 	(void) printf("\n");

include/sys/arc_impl.h

@@ -349,6 +349,8 @@ typedef struct l2arc_lb_ptr_buf {
 #define	L2BLK_SET_TYPE(field, x)	BF64_SET((field), 48, 8, x)
 #define	L2BLK_GET_PROTECTED(field)	BF64_GET((field), 56, 1)
 #define	L2BLK_SET_PROTECTED(field, x)	BF64_SET((field), 56, 1, x)
+#define	L2BLK_GET_STATE(field)		BF64_GET((field), 57, 4)
+#define	L2BLK_SET_STATE(field, x)	BF64_SET((field), 57, 4, x)
 
 #define	PTR_SWAP(x, y)		\
 	do {			\
@@ -445,6 +447,7 @@ typedef struct l2arc_buf_hdr {
 	uint64_t		b_daddr;	/* disk address, offset byte */
 	uint32_t		b_hits;
 	list_node_t		b_l2node;
+	arc_state_type_t	b_arcs_state;
 } l2arc_buf_hdr_t;
 
 typedef struct l2arc_write_callback {
@@ -733,6 +736,20 @@ typedef struct arc_stats {
 	kstat_named_t arcstat_mfu_ghost_evictable_metadata;
 	kstat_named_t arcstat_l2_hits;
 	kstat_named_t arcstat_l2_misses;
+	/*
+	 * Aligned size (in bytes) of L2ARC cached buffers according to their
+	 * ARC state at the time of caching in L2ARC.
+	 */
+	kstat_named_t arcstat_l2_mru_asize;
+	kstat_named_t arcstat_l2_mru_ghost_asize;
+	kstat_named_t arcstat_l2_mfu_asize;
+	kstat_named_t arcstat_l2_mfu_ghost_asize;
+	/*
+	 * Aligned size (in bytes) of L2ARC cached buffers by buffer content
+	 * type.
+	 */
+	kstat_named_t arcstat_l2_bufc_data_asize;
+	kstat_named_t arcstat_l2_bufc_metadata_asize;
 	kstat_named_t arcstat_l2_feeds;
 	kstat_named_t arcstat_l2_rw_clash;
 	kstat_named_t arcstat_l2_read_bytes;

module/zfs/arc.c

@@ -527,6 +527,12 @@ arc_stats_t arc_stats = {
 	{ "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
 	{ "l2_hits",			KSTAT_DATA_UINT64 },
 	{ "l2_misses",			KSTAT_DATA_UINT64 },
+	{ "l2_mru_asize",		KSTAT_DATA_UINT64 },
+	{ "l2_mru_ghost_asize",		KSTAT_DATA_UINT64 },
+	{ "l2_mfu_asize",		KSTAT_DATA_UINT64 },
+	{ "l2_mfu_ghost_asize",		KSTAT_DATA_UINT64 },
+	{ "l2_bufc_data_asize",		KSTAT_DATA_UINT64 },
+	{ "l2_bufc_metadata_asize",	KSTAT_DATA_UINT64 },
 	{ "l2_feeds",			KSTAT_DATA_UINT64 },
 	{ "l2_rw_clash",		KSTAT_DATA_UINT64 },
 	{ "l2_read_bytes",		KSTAT_DATA_UINT64 },
@@ -887,6 +893,7 @@ static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
 static void l2arc_read_done(zio_t *);
 static void l2arc_do_free_on_write(void);
+static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr);
 
 /*
  * L2ARC TRIM
@@ -1707,7 +1714,8 @@ arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
 static arc_buf_hdr_t *
 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
     dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
-    enum zio_compress compress, boolean_t protected, boolean_t prefetch)
+    enum zio_compress compress, boolean_t protected, boolean_t prefetch,
+    arc_state_type_t arcs_state)
 {
 	arc_buf_hdr_t	*hdr;
 
@@ -1730,6 +1738,7 @@ arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
 
 	hdr->b_l2hdr.b_dev = dev;
 	hdr->b_l2hdr.b_daddr = daddr;
+	hdr->b_l2hdr.b_arcs_state = arcs_state;
 
 	return (hdr);
 }
@@ -3661,6 +3670,68 @@ arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
 	return (buf);
 }
 
+static void
+l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr)
+{
+	l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
+	l2arc_dev_t *dev = l2hdr->b_dev;
+	uint64_t lsize = HDR_GET_LSIZE(hdr);
+	uint64_t psize = HDR_GET_PSIZE(hdr);
+	uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
+	arc_buf_contents_t type = hdr->b_type;
+	int64_t lsize_s;
+	int64_t psize_s;
+	int64_t asize_s;
+
+	if (!incr) {
+		lsize_s = -lsize;
+		psize_s = -psize;
+		asize_s = -asize;
+	} else {
+		lsize_s = lsize;
+		psize_s = psize;
+		asize_s = asize;
+	}
+
+	ARCSTAT_INCR(arcstat_l2_psize, psize_s);
+	ARCSTAT_INCR(arcstat_l2_lsize, lsize_s);
+
+	/*
+	 * We use the value stored in the L2 header upon initial caching in
+	 * L2ARC. This is because we are interested in the state of the buffer
+	 * in ARC when it was cached in L2ARC for the first time. This also
+	 * solves the problem of a possibly absent L1 header.
+	 */
+	switch (hdr->b_l2hdr.b_arcs_state) {
+		case ARC_STATE_MRU:
+			ARCSTAT_INCR(arcstat_l2_mru_asize, asize_s);
+			break;
+		case ARC_STATE_MRU_GHOST:
+			ARCSTAT_INCR(arcstat_l2_mru_ghost_asize, asize_s);
+			break;
+		case ARC_STATE_MFU:
+			ARCSTAT_INCR(arcstat_l2_mfu_asize, asize_s);
+			break;
+		case ARC_STATE_MFU_GHOST:
+			ARCSTAT_INCR(arcstat_l2_mfu_ghost_asize, asize_s);
+			break;
+		default:
+			break;
+	}
+
+	switch (type) {
+		case ARC_BUFC_DATA:
+			ARCSTAT_INCR(arcstat_l2_bufc_data_asize, asize_s);
+			break;
+		case ARC_BUFC_METADATA:
+			ARCSTAT_INCR(arcstat_l2_bufc_metadata_asize, asize_s);
+			break;
+		default:
+			break;
+	}
+}
+
+
 static void
 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
 {
@@ -3674,9 +3745,7 @@ arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
 
 	list_remove(&dev->l2ad_buflist, hdr);
 
-	ARCSTAT_INCR(arcstat_l2_psize, -psize);
-	ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
-
+	l2arc_hdr_arcstats_update(hdr, B_FALSE);
 	vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
 
 	(void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
@@ -8025,9 +8094,6 @@ l2arc_write_done(zio_t *zio)
 	DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
 	    l2arc_write_callback_t *, cb);
 
-	if (zio->io_error != 0)
-		ARCSTAT_BUMP(arcstat_l2_writes_error);
-
 	/*
 	 * All writes completed, or an error was hit.
 	 */
@@ -8091,8 +8157,7 @@ l2arc_write_done(zio_t *zio)
 			arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
 
 			uint64_t psize = HDR_GET_PSIZE(hdr);
-			ARCSTAT_INCR(arcstat_l2_psize, -psize);
-			ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
+			l2arc_hdr_arcstats_update(hdr, B_FALSE);
 
 			bytes_dropped +=
 			    vdev_psize_to_asize(dev->l2ad_vdev, psize);
@@ -8140,6 +8205,8 @@ l2arc_write_done(zio_t *zio)
 	list_destroy(&cb->l2wcb_abd_list);
 
 	if (zio->io_error != 0) {
+		ARCSTAT_BUMP(arcstat_l2_writes_error);
+
 		/*
 		 * Restore the lbps array in the header to its previous state.
 		 * If the list of log block pointers is empty, zero out the
@@ -9024,6 +9091,8 @@ l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
 			hdr->b_l2hdr.b_hits = 0;
 
 			hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
+			hdr->b_l2hdr.b_arcs_state =
+			    hdr->b_l1hdr.b_state->arcs_state;
 			arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR);
 
 			mutex_enter(&dev->l2ad_mtx);
@@ -9046,6 +9115,7 @@ l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
 			write_psize += psize;
 			write_asize += asize;
 			dev->l2ad_hand += asize;
+			l2arc_hdr_arcstats_update(hdr, B_TRUE);
 			vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
 
 			mutex_exit(hash_lock);
@@ -10023,19 +10093,18 @@ l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
 	    L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
 	    L2BLK_GET_COMPRESS((le)->le_prop),
 	    L2BLK_GET_PROTECTED((le)->le_prop),
-	    L2BLK_GET_PREFETCH((le)->le_prop));
+	    L2BLK_GET_PREFETCH((le)->le_prop),
+	    L2BLK_GET_STATE((le)->le_prop));
 	asize = vdev_psize_to_asize(dev->l2ad_vdev,
 	    L2BLK_GET_PSIZE((le)->le_prop));
 
 	/*
 	 * vdev_space_update() has to be called before arc_hdr_destroy() to
-	 * avoid underflow since the latter also calls the former.
+	 * avoid underflow since the latter also calls vdev_space_update().
 	 */
+	l2arc_hdr_arcstats_update(hdr, B_TRUE);
 	vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
 
-	ARCSTAT_INCR(arcstat_l2_lsize, HDR_GET_LSIZE(hdr));
-	ARCSTAT_INCR(arcstat_l2_psize, HDR_GET_PSIZE(hdr));
-
 	mutex_enter(&dev->l2ad_mtx);
 	list_insert_tail(&dev->l2ad_buflist, hdr);
 	(void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
@@ -10055,14 +10124,15 @@ l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
 			arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
 			exists->b_l2hdr.b_dev = dev;
 			exists->b_l2hdr.b_daddr = le->le_daddr;
+			exists->b_l2hdr.b_arcs_state =
+			    L2BLK_GET_STATE((le)->le_prop);
 			mutex_enter(&dev->l2ad_mtx);
 			list_insert_tail(&dev->l2ad_buflist, exists);
 			(void) zfs_refcount_add_many(&dev->l2ad_alloc,
 			    arc_hdr_size(exists), exists);
 			mutex_exit(&dev->l2ad_mtx);
+			l2arc_hdr_arcstats_update(exists, B_TRUE);
 			vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
-			ARCSTAT_INCR(arcstat_l2_lsize, HDR_GET_LSIZE(exists));
-			ARCSTAT_INCR(arcstat_l2_psize, HDR_GET_PSIZE(exists));
 		}
 		ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
 	}
@@ -10357,6 +10427,7 @@ l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
 	L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
 	L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
 	L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
+	L2BLK_SET_STATE((le)->le_prop, hdr->b_l1hdr.b_state->arcs_state);
 
 	dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
 	    HDR_GET_PSIZE(hdr));