[mirror_zfs.git] / module / zfs / vdev_cache.c

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */
/*
 * Copyright (c) 2013, 2016 by Delphix. All rights reserved.
 */

#include <sys/zfs_context.h>
#include <sys/spa.h>
#include <sys/vdev_impl.h>
#include <sys/zio.h>
#include <sys/kstat.h>
#include <sys/abd.h>

/*
 * Virtual device read-ahead caching.
 *
 * This file implements a simple LRU read-ahead cache.  When the DMU reads
 * a given block, it will often want other, nearby blocks soon thereafter.
 * We take advantage of this by reading a larger disk region and caching
 * the result.  In the best case, this can turn 128 back-to-back 512-byte
 * reads into a single 64k read followed by 127 cache hits; this reduces
 * latency dramatically.  In the worst case, it can turn an isolated 512-byte
 * read into a 64k read, which doesn't affect latency all that much but is
 * terribly wasteful of bandwidth.  A more intelligent version of the cache
 * could keep track of access patterns and not do read-ahead unless it sees
 * at least two temporally close I/Os to the same region.  Currently, only
 * metadata I/O is inflated.  A futher enhancement could take advantage of
 * more semantic information about the I/O.  And it could use something
 * faster than an AVL tree; that was chosen solely for convenience.
 *
 * There are five cache operations: allocate, fill, read, write, evict.
 *
 * (1) Allocate.  This reserves a cache entry for the specified region.
 *     We separate the allocate and fill operations so that multiple threads
 *     don't generate I/O for the same cache miss.
 *
 * (2) Fill.  When the I/O for a cache miss completes, the fill routine
 *     places the data in the previously allocated cache entry.
 *
 * (3) Read.  Read data from the cache.
 *
 * (4) Write.  Update cache contents after write completion.
 *
 * (5) Evict.  When allocating a new entry, we evict the oldest (LRU) entry
 *     if the total cache size exceeds zfs_vdev_cache_size.
 */

/*
 * These tunables are for performance analysis.
 */
/*
 * All i/os smaller than zfs_vdev_cache_max will be turned into
 * 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software
 * track buffer).  At most zfs_vdev_cache_size bytes will be kept in each
 * vdev's vdev_cache.
 *
 * TODO: Note that with the current ZFS code, it turns out that the
 * vdev cache is not helpful, and in some cases actually harmful.  It
 * is better if we disable this.  Once some time has passed, we should
 * actually remove this to simplify the code.  For now we just disable
 * it by setting the zfs_vdev_cache_size to zero.  Note that Solaris 11
 * has made these same changes.
 */
int zfs_vdev_cache_max = 1<<14;			/* 16KB */
int zfs_vdev_cache_size = 0;
int zfs_vdev_cache_bshift = 16;

#define	VCBS (1 << zfs_vdev_cache_bshift)	/* 64KB */

kstat_t	*vdc_ksp = NULL;

typedef struct vdc_stats {
	kstat_named_t vdc_stat_delegations;
	kstat_named_t vdc_stat_hits;
	kstat_named_t vdc_stat_misses;
} vdc_stats_t;

static vdc_stats_t vdc_stats = {
	{ "delegations",	KSTAT_DATA_UINT64 },
	{ "hits",		KSTAT_DATA_UINT64 },
	{ "misses",		KSTAT_DATA_UINT64 }
};

#define	VDCSTAT_BUMP(stat)	atomic_inc_64(&vdc_stats.stat.value.ui64);

static inline int
vdev_cache_offset_compare(const void *a1, const void *a2)
{
	const vdev_cache_entry_t *ve1 = (const vdev_cache_entry_t *)a1;
	const vdev_cache_entry_t *ve2 = (const vdev_cache_entry_t *)a2;

	return (AVL_CMP(ve1->ve_offset, ve2->ve_offset));
}

static int
vdev_cache_lastused_compare(const void *a1, const void *a2)
{
	const vdev_cache_entry_t *ve1 = (const vdev_cache_entry_t *)a1;
	const vdev_cache_entry_t *ve2 = (const vdev_cache_entry_t *)a2;

	int cmp = AVL_CMP(ve1->ve_lastused, ve2->ve_lastused);
	if (likely(cmp))
		return (cmp);

	/*
	 * Among equally old entries, sort by offset to ensure uniqueness.
	 */
	return (vdev_cache_offset_compare(a1, a2));
}

/*
 * Evict the specified entry from the cache.
 */
static void
vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
{
	ASSERT(MUTEX_HELD(&vc->vc_lock));
	ASSERT3P(ve->ve_fill_io, ==, NULL);
	ASSERT3P(ve->ve_abd, !=, NULL);

	avl_remove(&vc->vc_lastused_tree, ve);
	avl_remove(&vc->vc_offset_tree, ve);
	abd_free(ve->ve_abd);
	kmem_free(ve, sizeof (vdev_cache_entry_t));
}

/*
 * Allocate an entry in the cache.  At the point we don't have the data,
 * we're just creating a placeholder so that multiple threads don't all
 * go off and read the same blocks.
 */
static vdev_cache_entry_t *
vdev_cache_allocate(zio_t *zio)
{
	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
	uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
	vdev_cache_entry_t *ve;

	ASSERT(MUTEX_HELD(&vc->vc_lock));

	if (zfs_vdev_cache_size == 0)
		return (NULL);

	/*
	 * If adding a new entry would exceed the cache size,
	 * evict the oldest entry (LRU).
	 */
	if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
	    zfs_vdev_cache_size) {
		ve = avl_first(&vc->vc_lastused_tree);
		if (ve->ve_fill_io != NULL)
			return (NULL);
		ASSERT3U(ve->ve_hits, !=, 0);
		vdev_cache_evict(vc, ve);
	}

	ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
	ve->ve_offset = offset;
	ve->ve_lastused = ddi_get_lbolt();
	ve->ve_abd = abd_alloc_for_io(VCBS, B_TRUE);

	avl_add(&vc->vc_offset_tree, ve);
	avl_add(&vc->vc_lastused_tree, ve);

	return (ve);
}

static void
vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
{
	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);

	ASSERT(MUTEX_HELD(&vc->vc_lock));
	ASSERT3P(ve->ve_fill_io, ==, NULL);

	if (ve->ve_lastused != ddi_get_lbolt()) {
		avl_remove(&vc->vc_lastused_tree, ve);
		ve->ve_lastused = ddi_get_lbolt();
		avl_add(&vc->vc_lastused_tree, ve);
	}

	ve->ve_hits++;
	abd_copy_off(zio->io_abd, ve->ve_abd, 0, cache_phase, zio->io_size);
}

/*
 * Fill a previously allocated cache entry with data.
 */
static void
vdev_cache_fill(zio_t *fio)
{
	vdev_t *vd = fio->io_vd;
	vdev_cache_t *vc = &vd->vdev_cache;
	vdev_cache_entry_t *ve = fio->io_private;
	zio_t *pio;

	ASSERT3U(fio->io_size, ==, VCBS);

	/*
	 * Add data to the cache.
	 */
	mutex_enter(&vc->vc_lock);

	ASSERT3P(ve->ve_fill_io, ==, fio);
	ASSERT3U(ve->ve_offset, ==, fio->io_offset);
	ASSERT3P(ve->ve_abd, ==, fio->io_abd);

	ve->ve_fill_io = NULL;

	/*
	 * Even if this cache line was invalidated by a missed write update,
	 * any reads that were queued up before the missed update are still
	 * valid, so we can satisfy them from this line before we evict it.
	 */
	zio_link_t *zl = NULL;
	while ((pio = zio_walk_parents(fio, &zl)) != NULL)
		vdev_cache_hit(vc, ve, pio);

	if (fio->io_error || ve->ve_missed_update)
		vdev_cache_evict(vc, ve);

	mutex_exit(&vc->vc_lock);
}

/*
 * Read data from the cache.  Returns B_TRUE cache hit, B_FALSE on miss.
 */
boolean_t
vdev_cache_read(zio_t *zio)
{
	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
	vdev_cache_entry_t *ve, *ve_search;
	uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
	zio_t *fio;
	ASSERTV(uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS));

	ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);

	if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
		return (B_FALSE);

	if (zio->io_size > zfs_vdev_cache_max)
		return (B_FALSE);

	/*
	 * If the I/O straddles two or more cache blocks, don't cache it.
	 */
	if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS))
		return (B_FALSE);

	ASSERT3U(cache_phase + zio->io_size, <=, VCBS);

	mutex_enter(&vc->vc_lock);

	ve_search = kmem_alloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
	ve_search->ve_offset = cache_offset;
	ve = avl_find(&vc->vc_offset_tree, ve_search, NULL);
	kmem_free(ve_search, sizeof (vdev_cache_entry_t));

	if (ve != NULL) {
		if (ve->ve_missed_update) {
			mutex_exit(&vc->vc_lock);
			return (B_FALSE);
		}

		if ((fio = ve->ve_fill_io) != NULL) {
			zio_vdev_io_bypass(zio);
			zio_add_child(zio, fio);
			mutex_exit(&vc->vc_lock);
			VDCSTAT_BUMP(vdc_stat_delegations);
			return (B_TRUE);
		}

		vdev_cache_hit(vc, ve, zio);
		zio_vdev_io_bypass(zio);

		mutex_exit(&vc->vc_lock);
		VDCSTAT_BUMP(vdc_stat_hits);
		return (B_TRUE);
	}

	ve = vdev_cache_allocate(zio);

	if (ve == NULL) {
		mutex_exit(&vc->vc_lock);
		return (B_FALSE);
	}

	fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
	    ve->ve_abd, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_NOW,
	    ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);

	ve->ve_fill_io = fio;
	zio_vdev_io_bypass(zio);
	zio_add_child(zio, fio);

	mutex_exit(&vc->vc_lock);
	zio_nowait(fio);
	VDCSTAT_BUMP(vdc_stat_misses);

	return (B_TRUE);
}

/*
 * Update cache contents upon write completion.
 */
void
vdev_cache_write(zio_t *zio)
{
	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
	vdev_cache_entry_t *ve, ve_search;
	uint64_t io_start = zio->io_offset;
	uint64_t io_end = io_start + zio->io_size;
	uint64_t min_offset = P2ALIGN(io_start, VCBS);
	uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
	avl_index_t where;

	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);

	mutex_enter(&vc->vc_lock);

	ve_search.ve_offset = min_offset;
	ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);

	if (ve == NULL)
		ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);

	while (ve != NULL && ve->ve_offset < max_offset) {
		uint64_t start = MAX(ve->ve_offset, io_start);
		uint64_t end = MIN(ve->ve_offset + VCBS, io_end);

		if (ve->ve_fill_io != NULL) {
			ve->ve_missed_update = 1;
		} else {
			abd_copy_off(ve->ve_abd, zio->io_abd,
			    start - ve->ve_offset, start - io_start,
			    end - start);
		}
		ve = AVL_NEXT(&vc->vc_offset_tree, ve);
	}
	mutex_exit(&vc->vc_lock);
}

void
vdev_cache_purge(vdev_t *vd)
{
	vdev_cache_t *vc = &vd->vdev_cache;
	vdev_cache_entry_t *ve;

	mutex_enter(&vc->vc_lock);
	while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
		vdev_cache_evict(vc, ve);
	mutex_exit(&vc->vc_lock);
}

void
vdev_cache_init(vdev_t *vd)
{
	vdev_cache_t *vc = &vd->vdev_cache;

	mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);

	avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
	    sizeof (vdev_cache_entry_t),
	    offsetof(struct vdev_cache_entry, ve_offset_node));

	avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
	    sizeof (vdev_cache_entry_t),
	    offsetof(struct vdev_cache_entry, ve_lastused_node));
}

void
vdev_cache_fini(vdev_t *vd)
{
	vdev_cache_t *vc = &vd->vdev_cache;

	vdev_cache_purge(vd);

	avl_destroy(&vc->vc_offset_tree);
	avl_destroy(&vc->vc_lastused_tree);

	mutex_destroy(&vc->vc_lock);
}

void
vdev_cache_stat_init(void)
{
	vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
	    KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
	    KSTAT_FLAG_VIRTUAL);
	if (vdc_ksp != NULL) {
		vdc_ksp->ks_data = &vdc_stats;
		kstat_install(vdc_ksp);
	}
}

void
vdev_cache_stat_fini(void)
{
	if (vdc_ksp != NULL) {
		kstat_delete(vdc_ksp);
		vdc_ksp = NULL;
	}
}

#if defined(_KERNEL) && defined(HAVE_SPL)
module_param(zfs_vdev_cache_max, int, 0644);
MODULE_PARM_DESC(zfs_vdev_cache_max, "Inflate reads small than max");

module_param(zfs_vdev_cache_size, int, 0444);
MODULE_PARM_DESC(zfs_vdev_cache_size, "Total size of the per-disk cache");

module_param(zfs_vdev_cache_bshift, int, 0644);
MODULE_PARM_DESC(zfs_vdev_cache_bshift, "Shift size to inflate reads too");
#endif
Commit	Line	Data
34dc7c2f BB	1	/*
	2	* CDDL HEADER START
	3	*
	4	* The contents of this file are subject to the terms of the
	5	* Common Development and Distribution License (the "License").
	6	* You may not use this file except in compliance with the License.
	7	*
	8	* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
	9	* or http://www.opensolaris.org/os/licensing.
	10	* See the License for the specific language governing permissions
	11	* and limitations under the License.
	12	*
	13	* When distributing Covered Code, include this CDDL HEADER in each
	14	* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
	15	* If applicable, add the following below this CDDL HEADER, with the
	16	* fields enclosed by brackets "[]" replaced with your own identifying
	17	* information: Portions Copyright [yyyy] [name of copyright owner]
	18	*
	19	* CDDL HEADER END
	20	*/
	21	/*
d164b209	22	* Copyright 2009 Sun Microsystems, Inc. All rights reserved.
34dc7c2f BB	23	* Use is subject to license terms.
34dc7c2f BB	24	*/
2e528b49	25	/*
a6255b7f	26	* Copyright (c) 2013, 2016 by Delphix. All rights reserved.
2e528b49	27	*/
34dc7c2f	28
34dc7c2f BB	29	#include <sys/zfs_context.h>
	30	#include <sys/spa.h>
	31	#include <sys/vdev_impl.h>
	32	#include <sys/zio.h>
	33	#include <sys/kstat.h>
a6255b7f	34	#include <sys/abd.h>
34dc7c2f BB	35
	36	/*
	37	* Virtual device read-ahead caching.
	38	*
	39	* This file implements a simple LRU read-ahead cache. When the DMU reads
	40	* a given block, it will often want other, nearby blocks soon thereafter.
	41	* We take advantage of this by reading a larger disk region and caching
	42	* the result. In the best case, this can turn 128 back-to-back 512-byte
	43	* reads into a single 64k read followed by 127 cache hits; this reduces
	44	* latency dramatically. In the worst case, it can turn an isolated 512-byte
	45	* read into a 64k read, which doesn't affect latency all that much but is
	46	* terribly wasteful of bandwidth. A more intelligent version of the cache
	47	* could keep track of access patterns and not do read-ahead unless it sees
	48	* at least two temporally close I/Os to the same region. Currently, only
	49	* metadata I/O is inflated. A futher enhancement could take advantage of
	50	* more semantic information about the I/O. And it could use something
	51	* faster than an AVL tree; that was chosen solely for convenience.
	52	*
	53	* There are five cache operations: allocate, fill, read, write, evict.
	54	*
	55	* (1) Allocate. This reserves a cache entry for the specified region.
	56	* We separate the allocate and fill operations so that multiple threads
	57	* don't generate I/O for the same cache miss.
	58	*
	59	* (2) Fill. When the I/O for a cache miss completes, the fill routine
	60	* places the data in the previously allocated cache entry.
	61	*
	62	* (3) Read. Read data from the cache.
	63	*
	64	* (4) Write. Update cache contents after write completion.
	65	*
	66	* (5) Evict. When allocating a new entry, we evict the oldest (LRU) entry
	67	* if the total cache size exceeds zfs_vdev_cache_size.
	68	*/
	69
	70	/*
	71	* These tunables are for performance analysis.
	72	*/
	73	/*
	74	* All i/os smaller than zfs_vdev_cache_max will be turned into
	75	* 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software
	76	* track buffer). At most zfs_vdev_cache_size bytes will be kept in each
	77	* vdev's vdev_cache.
2cc6c8db GA	78	*
	79	* TODO: Note that with the current ZFS code, it turns out that the
	80	* vdev cache is not helpful, and in some cases actually harmful. It
	81	* is better if we disable this. Once some time has passed, we should
	82	* actually remove this to simplify the code. For now we just disable
	83	* it by setting the zfs_vdev_cache_size to zero. Note that Solaris 11
	84	* has made these same changes.
34dc7c2f BB	85	*/
34dc7c2f BB	86	int zfs_vdev_cache_max = 1<<14; /* 16KB */
2cc6c8db	87	int zfs_vdev_cache_size = 0;
34dc7c2f BB	88	int zfs_vdev_cache_bshift = 16;
	89
	90	#define VCBS (1 << zfs_vdev_cache_bshift) /* 64KB */
	91
	92	kstat_t *vdc_ksp = NULL;
	93
	94	typedef struct vdc_stats {
	95	kstat_named_t vdc_stat_delegations;
	96	kstat_named_t vdc_stat_hits;
	97	kstat_named_t vdc_stat_misses;
	98	} vdc_stats_t;
	99
	100	static vdc_stats_t vdc_stats = {
	101	{ "delegations", KSTAT_DATA_UINT64 },
	102	{ "hits", KSTAT_DATA_UINT64 },
	103	{ "misses", KSTAT_DATA_UINT64 }
	104	};
	105
bc89ac84	106	#define VDCSTAT_BUMP(stat) atomic_inc_64(&vdc_stats.stat.value.ui64);
34dc7c2f	107
ee36c709	108	static inline int
34dc7c2f BB	109	vdev_cache_offset_compare(const void a1, const void a2)
34dc7c2f BB	110	{
ee36c709 GN	111	const vdev_cache_entry_t ve1 = (const vdev_cache_entry_t )a1;
ee36c709 GN	112	const vdev_cache_entry_t ve2 = (const vdev_cache_entry_t )a2;
34dc7c2f	113
ee36c709	114	return (AVL_CMP(ve1->ve_offset, ve2->ve_offset));
34dc7c2f BB	115	}
	116
	117	static int
	118	vdev_cache_lastused_compare(const void a1, const void a2)
	119	{
ee36c709 GN	120	const vdev_cache_entry_t ve1 = (const vdev_cache_entry_t )a1;
ee36c709 GN	121	const vdev_cache_entry_t ve2 = (const vdev_cache_entry_t )a2;
34dc7c2f	122
ee36c709 GN	123	int cmp = AVL_CMP(ve1->ve_lastused, ve2->ve_lastused);
	124	if (likely(cmp))
	125	return (cmp);
34dc7c2f BB	126
	127	/*
	128	* Among equally old entries, sort by offset to ensure uniqueness.
	129	*/
	130	return (vdev_cache_offset_compare(a1, a2));
	131	}
	132
	133	/*
	134	* Evict the specified entry from the cache.
	135	*/
	136	static void
	137	vdev_cache_evict(vdev_cache_t vc, vdev_cache_entry_t ve)
	138	{
	139	ASSERT(MUTEX_HELD(&vc->vc_lock));
a6255b7f DQ	140	ASSERT3P(ve->ve_fill_io, ==, NULL);
a6255b7f DQ	141	ASSERT3P(ve->ve_abd, !=, NULL);
34dc7c2f	142
34dc7c2f BB	143	avl_remove(&vc->vc_lastused_tree, ve);
34dc7c2f BB	144	avl_remove(&vc->vc_offset_tree, ve);
a6255b7f	145	abd_free(ve->ve_abd);
34dc7c2f BB	146	kmem_free(ve, sizeof (vdev_cache_entry_t));
	147	}
	148
	149	/*
	150	* Allocate an entry in the cache. At the point we don't have the data,
	151	* we're just creating a placeholder so that multiple threads don't all
	152	* go off and read the same blocks.
	153	*/
	154	static vdev_cache_entry_t *
	155	vdev_cache_allocate(zio_t *zio)
	156	{
	157	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
	158	uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
	159	vdev_cache_entry_t *ve;
	160
	161	ASSERT(MUTEX_HELD(&vc->vc_lock));
	162
	163	if (zfs_vdev_cache_size == 0)
	164	return (NULL);
	165
	166	/*
	167	* If adding a new entry would exceed the cache size,
	168	* evict the oldest entry (LRU).
	169	*/
	170	if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
	171	zfs_vdev_cache_size) {
	172	ve = avl_first(&vc->vc_lastused_tree);
b128c09f	173	if (ve->ve_fill_io != NULL)
34dc7c2f	174	return (NULL);
a6255b7f	175	ASSERT3U(ve->ve_hits, !=, 0);
34dc7c2f BB	176	vdev_cache_evict(vc, ve);
	177	}
	178
79c76d5b	179	ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
34dc7c2f	180	ve->ve_offset = offset;
428870ff	181	ve->ve_lastused = ddi_get_lbolt();
a6255b7f	182	ve->ve_abd = abd_alloc_for_io(VCBS, B_TRUE);
34dc7c2f BB	183
	184	avl_add(&vc->vc_offset_tree, ve);
	185	avl_add(&vc->vc_lastused_tree, ve);
	186
	187	return (ve);
	188	}
	189
	190	static void
	191	vdev_cache_hit(vdev_cache_t vc, vdev_cache_entry_t ve, zio_t *zio)
	192	{
	193	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
	194
	195	ASSERT(MUTEX_HELD(&vc->vc_lock));
a6255b7f	196	ASSERT3P(ve->ve_fill_io, ==, NULL);
34dc7c2f	197
428870ff	198	if (ve->ve_lastused != ddi_get_lbolt()) {
34dc7c2f	199	avl_remove(&vc->vc_lastused_tree, ve);
428870ff	200	ve->ve_lastused = ddi_get_lbolt();
34dc7c2f BB	201	avl_add(&vc->vc_lastused_tree, ve);
	202	}
	203
	204	ve->ve_hits++;
a6255b7f	205	abd_copy_off(zio->io_abd, ve->ve_abd, 0, cache_phase, zio->io_size);
34dc7c2f BB	206	}
	207
	208	/*
	209	* Fill a previously allocated cache entry with data.
	210	*/
	211	static void
d164b209	212	vdev_cache_fill(zio_t *fio)
34dc7c2f	213	{
d164b209	214	vdev_t *vd = fio->io_vd;
34dc7c2f	215	vdev_cache_t *vc = &vd->vdev_cache;
d164b209 BB	216	vdev_cache_entry_t *ve = fio->io_private;
d164b209 BB	217	zio_t *pio;
34dc7c2f	218
a6255b7f	219	ASSERT3U(fio->io_size, ==, VCBS);
34dc7c2f BB	220
	221	/*
	222	* Add data to the cache.
	223	*/
	224	mutex_enter(&vc->vc_lock);
	225
a6255b7f DQ	226	ASSERT3P(ve->ve_fill_io, ==, fio);
	227	ASSERT3U(ve->ve_offset, ==, fio->io_offset);
	228	ASSERT3P(ve->ve_abd, ==, fio->io_abd);
34dc7c2f BB	229
	230	ve->ve_fill_io = NULL;
	231
	232	/*
	233	* Even if this cache line was invalidated by a missed write update,
	234	* any reads that were queued up before the missed update are still
	235	* valid, so we can satisfy them from this line before we evict it.
	236	*/
1c27024e	237	zio_link_t *zl = NULL;
3dfb57a3	238	while ((pio = zio_walk_parents(fio, &zl)) != NULL)
d164b209	239	vdev_cache_hit(vc, ve, pio);
34dc7c2f	240
d164b209	241	if (fio->io_error \|\| ve->ve_missed_update)
34dc7c2f BB	242	vdev_cache_evict(vc, ve);
	243
	244	mutex_exit(&vc->vc_lock);
34dc7c2f BB	245	}
	246
	247	/*
b0bc7a84	248	* Read data from the cache. Returns B_TRUE cache hit, B_FALSE on miss.
34dc7c2f	249	*/
b0bc7a84	250	boolean_t
34dc7c2f BB	251	vdev_cache_read(zio_t *zio)
	252	{
	253	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
5fed499d	254	vdev_cache_entry_t ve, ve_search;
34dc7c2f	255	uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
34dc7c2f	256	zio_t *fio;
d1d7e268	257	ASSERTV(uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS));
34dc7c2f	258
a6255b7f	259	ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
34dc7c2f BB	260
34dc7c2f BB	261	if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
b0bc7a84	262	return (B_FALSE);
34dc7c2f BB	263
34dc7c2f BB	264	if (zio->io_size > zfs_vdev_cache_max)
b0bc7a84	265	return (B_FALSE);
34dc7c2f BB	266
	267	/*
	268	* If the I/O straddles two or more cache blocks, don't cache it.
	269	*/
b128c09f	270	if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS))
b0bc7a84	271	return (B_FALSE);
34dc7c2f	272
a6255b7f	273	ASSERT3U(cache_phase + zio->io_size, <=, VCBS);
34dc7c2f BB	274
	275	mutex_enter(&vc->vc_lock);
	276
79c76d5b	277	ve_search = kmem_alloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
5fed499d BB	278	ve_search->ve_offset = cache_offset;
5fed499d BB	279	ve = avl_find(&vc->vc_offset_tree, ve_search, NULL);
d1d7e268	280	kmem_free(ve_search, sizeof (vdev_cache_entry_t));
34dc7c2f BB	281
	282	if (ve != NULL) {
	283	if (ve->ve_missed_update) {
	284	mutex_exit(&vc->vc_lock);
b0bc7a84	285	return (B_FALSE);
34dc7c2f BB	286	}
	287
	288	if ((fio = ve->ve_fill_io) != NULL) {
34dc7c2f	289	zio_vdev_io_bypass(zio);
d164b209	290	zio_add_child(zio, fio);
34dc7c2f BB	291	mutex_exit(&vc->vc_lock);
34dc7c2f BB	292	VDCSTAT_BUMP(vdc_stat_delegations);
b0bc7a84	293	return (B_TRUE);
34dc7c2f BB	294	}
	295
	296	vdev_cache_hit(vc, ve, zio);
	297	zio_vdev_io_bypass(zio);
	298
	299	mutex_exit(&vc->vc_lock);
34dc7c2f	300	VDCSTAT_BUMP(vdc_stat_hits);
b0bc7a84	301	return (B_TRUE);
34dc7c2f BB	302	}
	303
	304	ve = vdev_cache_allocate(zio);
	305
	306	if (ve == NULL) {
	307	mutex_exit(&vc->vc_lock);
b0bc7a84	308	return (B_FALSE);
34dc7c2f BB	309	}
34dc7c2f BB	310
b128c09f	311	fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
a6255b7f	312	ve->ve_abd, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_NOW,
b128c09f	313	ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);
34dc7c2f BB	314
34dc7c2f BB	315	ve->ve_fill_io = fio;
34dc7c2f	316	zio_vdev_io_bypass(zio);
d164b209	317	zio_add_child(zio, fio);
34dc7c2f BB	318
	319	mutex_exit(&vc->vc_lock);
	320	zio_nowait(fio);
	321	VDCSTAT_BUMP(vdc_stat_misses);
	322
b0bc7a84	323	return (B_TRUE);
34dc7c2f BB	324	}
	325
	326	/*
	327	* Update cache contents upon write completion.
	328	*/
	329	void
	330	vdev_cache_write(zio_t *zio)
	331	{
	332	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
	333	vdev_cache_entry_t *ve, ve_search;
	334	uint64_t io_start = zio->io_offset;
	335	uint64_t io_end = io_start + zio->io_size;
	336	uint64_t min_offset = P2ALIGN(io_start, VCBS);
	337	uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
	338	avl_index_t where;
	339
a6255b7f	340	ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
34dc7c2f BB	341
	342	mutex_enter(&vc->vc_lock);
	343
	344	ve_search.ve_offset = min_offset;
	345	ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
	346
	347	if (ve == NULL)
	348	ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
	349
	350	while (ve != NULL && ve->ve_offset < max_offset) {
	351	uint64_t start = MAX(ve->ve_offset, io_start);
	352	uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
	353
	354	if (ve->ve_fill_io != NULL) {
	355	ve->ve_missed_update = 1;
	356	} else {
12aec7dc CC	357	abd_copy_off(ve->ve_abd, zio->io_abd,
	358	start - ve->ve_offset, start - io_start,
	359	end - start);
34dc7c2f BB	360	}
	361	ve = AVL_NEXT(&vc->vc_offset_tree, ve);
	362	}
	363	mutex_exit(&vc->vc_lock);
	364	}
	365
	366	void
	367	vdev_cache_purge(vdev_t *vd)
	368	{
	369	vdev_cache_t *vc = &vd->vdev_cache;
	370	vdev_cache_entry_t *ve;
	371
	372	mutex_enter(&vc->vc_lock);
	373	while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
	374	vdev_cache_evict(vc, ve);
	375	mutex_exit(&vc->vc_lock);
	376	}
	377
	378	void
	379	vdev_cache_init(vdev_t *vd)
	380	{
	381	vdev_cache_t *vc = &vd->vdev_cache;
	382
	383	mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
	384
	385	avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
	386	sizeof (vdev_cache_entry_t),
	387	offsetof(struct vdev_cache_entry, ve_offset_node));
	388
	389	avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
	390	sizeof (vdev_cache_entry_t),
	391	offsetof(struct vdev_cache_entry, ve_lastused_node));
	392	}
	393
	394	void
	395	vdev_cache_fini(vdev_t *vd)
	396	{
	397	vdev_cache_t *vc = &vd->vdev_cache;
	398
	399	vdev_cache_purge(vd);
	400
	401	avl_destroy(&vc->vc_offset_tree);
	402	avl_destroy(&vc->vc_lastused_tree);
	403
	404	mutex_destroy(&vc->vc_lock);
	405	}
	406
	407	void
	408	vdev_cache_stat_init(void)
	409	{
	410	vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
	411	KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
	412	KSTAT_FLAG_VIRTUAL);
	413	if (vdc_ksp != NULL) {
	414	vdc_ksp->ks_data = &vdc_stats;
	415	kstat_install(vdc_ksp);
	416	}
	417	}
	418
	419	void
	420	vdev_cache_stat_fini(void)
	421	{
	422	if (vdc_ksp != NULL) {
	423	kstat_delete(vdc_ksp);
424	vdc_ksp = NULL;
425	}
426	}
c409e464 BB	427
	428	#if defined(_KERNEL) && defined(HAVE_SPL)
	429	module_param(zfs_vdev_cache_max, int, 0644);
	430	MODULE_PARM_DESC(zfs_vdev_cache_max, "Inflate reads small than max");
	431
	432	module_param(zfs_vdev_cache_size, int, 0444);
	433	MODULE_PARM_DESC(zfs_vdev_cache_size, "Total size of the per-disk cache");
	434
	435	module_param(zfs_vdev_cache_bshift, int, 0644);
	436	MODULE_PARM_DESC(zfs_vdev_cache_bshift, "Shift size to inflate reads too");
	437	#endif