[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.
 */

#include <sys/zfs_context.h>
#include <sys/spa.h>
#include <sys/vdev_impl.h>
#include <sys/zio.h>
#include <sys/kstat.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_add_64(&vdc_stats.stat.value.ui64, 1);

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

	if (ve1->ve_offset < ve2->ve_offset)
		return (-1);
	if (ve1->ve_offset > ve2->ve_offset)
		return (1);
	return (0);
}

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

	if (ve1->ve_lastused < ve2->ve_lastused)
		return (-1);
	if (ve1->ve_lastused > ve2->ve_lastused)
		return (1);

	/*
	 * 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));
	ASSERT(ve->ve_fill_io == NULL);
	ASSERT(ve->ve_data != NULL);

	avl_remove(&vc->vc_lastused_tree, ve);
	avl_remove(&vc->vc_offset_tree, ve);
	zio_buf_free(ve->ve_data, VCBS);
	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);
		ASSERT(ve->ve_hits != 0);
		vdev_cache_evict(vc, ve);
	}

	ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_PUSHPAGE);
	ve->ve_offset = offset;
	ve->ve_lastused = ddi_get_lbolt();
	ve->ve_data = zio_buf_alloc(VCBS);

	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));
	ASSERT(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++;
	bcopy(ve->ve_data + cache_phase, zio->io_data, 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;

	ASSERT(fio->io_size == VCBS);

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

	ASSERT(ve->ve_fill_io == fio);
	ASSERT(ve->ve_offset == fio->io_offset);
	ASSERT(ve->ve_data == fio->io_data);

	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.
	 */
	while ((pio = zio_walk_parents(fio)) != 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 0 on cache hit, errno on a miss.
 */
int
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);
	ASSERTV(uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);)
	zio_t *fio;

	ASSERT(zio->io_type == ZIO_TYPE_READ);

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

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

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

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

	mutex_enter(&vc->vc_lock);

	ve_search = kmem_alloc(sizeof(vdev_cache_entry_t), KM_PUSHPAGE);
	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 (ESTALE);
		}

		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 (0);
		}

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

		mutex_exit(&vc->vc_lock);
		VDCSTAT_BUMP(vdc_stat_hits);
		return (0);
	}

	ve = vdev_cache_allocate(zio);

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

	fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
	    ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_CACHE_FILL,
	    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 (0);
}

/*
 * 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;

	ASSERT(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 {
			bcopy((char *)zio->io_data + start - io_start,
			    ve->ve_data + start - ve->ve_offset, 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.
	24	*/
	25
34dc7c2f BB	26	#include <sys/zfs_context.h>
	27	#include <sys/spa.h>
	28	#include <sys/vdev_impl.h>
	29	#include <sys/zio.h>
	30	#include <sys/kstat.h>
	31
	32	/*
	33	* Virtual device read-ahead caching.
	34	*
	35	* This file implements a simple LRU read-ahead cache. When the DMU reads
	36	* a given block, it will often want other, nearby blocks soon thereafter.
	37	* We take advantage of this by reading a larger disk region and caching
	38	* the result. In the best case, this can turn 128 back-to-back 512-byte
	39	* reads into a single 64k read followed by 127 cache hits; this reduces
	40	* latency dramatically. In the worst case, it can turn an isolated 512-byte
	41	* read into a 64k read, which doesn't affect latency all that much but is
	42	* terribly wasteful of bandwidth. A more intelligent version of the cache
	43	* could keep track of access patterns and not do read-ahead unless it sees
	44	* at least two temporally close I/Os to the same region. Currently, only
	45	* metadata I/O is inflated. A futher enhancement could take advantage of
	46	* more semantic information about the I/O. And it could use something
	47	* faster than an AVL tree; that was chosen solely for convenience.
	48	*
	49	* There are five cache operations: allocate, fill, read, write, evict.
	50	*
	51	* (1) Allocate. This reserves a cache entry for the specified region.
	52	* We separate the allocate and fill operations so that multiple threads
	53	* don't generate I/O for the same cache miss.
	54	*
	55	* (2) Fill. When the I/O for a cache miss completes, the fill routine
	56	* places the data in the previously allocated cache entry.
	57	*
	58	* (3) Read. Read data from the cache.
	59	*
	60	* (4) Write. Update cache contents after write completion.
	61	*
	62	* (5) Evict. When allocating a new entry, we evict the oldest (LRU) entry
	63	* if the total cache size exceeds zfs_vdev_cache_size.
	64	*/
	65
	66	/*
	67	* These tunables are for performance analysis.
	68	*/
	69	/*
	70	* All i/os smaller than zfs_vdev_cache_max will be turned into
	71	* 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software
	72	* track buffer). At most zfs_vdev_cache_size bytes will be kept in each
	73	* vdev's vdev_cache.
2cc6c8db GA	74	*
	75	* TODO: Note that with the current ZFS code, it turns out that the
	76	* vdev cache is not helpful, and in some cases actually harmful. It
	77	* is better if we disable this. Once some time has passed, we should
	78	* actually remove this to simplify the code. For now we just disable
	79	* it by setting the zfs_vdev_cache_size to zero. Note that Solaris 11
	80	* has made these same changes.
34dc7c2f BB	81	*/
34dc7c2f BB	82	int zfs_vdev_cache_max = 1<<14; /* 16KB */
2cc6c8db	83	int zfs_vdev_cache_size = 0;
34dc7c2f BB	84	int zfs_vdev_cache_bshift = 16;
	85
	86	#define VCBS (1 << zfs_vdev_cache_bshift) /* 64KB */
	87
	88	kstat_t *vdc_ksp = NULL;
	89
	90	typedef struct vdc_stats {
	91	kstat_named_t vdc_stat_delegations;
	92	kstat_named_t vdc_stat_hits;
	93	kstat_named_t vdc_stat_misses;
	94	} vdc_stats_t;
	95
	96	static vdc_stats_t vdc_stats = {
	97	{ "delegations", KSTAT_DATA_UINT64 },
	98	{ "hits", KSTAT_DATA_UINT64 },
	99	{ "misses", KSTAT_DATA_UINT64 }
	100	};
	101
	102	#define VDCSTAT_BUMP(stat) atomic_add_64(&vdc_stats.stat.value.ui64, 1);
	103
	104	static int
	105	vdev_cache_offset_compare(const void a1, const void a2)
	106	{
	107	const vdev_cache_entry_t *ve1 = a1;
	108	const vdev_cache_entry_t *ve2 = a2;
	109
	110	if (ve1->ve_offset < ve2->ve_offset)
	111	return (-1);
	112	if (ve1->ve_offset > ve2->ve_offset)
	113	return (1);
	114	return (0);
	115	}
	116
	117	static int
	118	vdev_cache_lastused_compare(const void a1, const void a2)
	119	{
	120	const vdev_cache_entry_t *ve1 = a1;
	121	const vdev_cache_entry_t *ve2 = a2;
	122
	123	if (ve1->ve_lastused < ve2->ve_lastused)
	124	return (-1);
	125	if (ve1->ve_lastused > ve2->ve_lastused)
	126	return (1);
	127
	128	/*
	129	* Among equally old entries, sort by offset to ensure uniqueness.
	130	*/
	131	return (vdev_cache_offset_compare(a1, a2));
	132	}
	133
	134	/*
	135	* Evict the specified entry from the cache.
	136	*/
	137	static void
	138	vdev_cache_evict(vdev_cache_t vc, vdev_cache_entry_t ve)
	139	{
	140	ASSERT(MUTEX_HELD(&vc->vc_lock));
	141	ASSERT(ve->ve_fill_io == NULL);
	142	ASSERT(ve->ve_data != NULL);
	143
34dc7c2f BB	144	avl_remove(&vc->vc_lastused_tree, ve);
	145	avl_remove(&vc->vc_offset_tree, ve);
	146	zio_buf_free(ve->ve_data, VCBS);
	147	kmem_free(ve, sizeof (vdev_cache_entry_t));
	148	}
	149
	150	/*
	151	* Allocate an entry in the cache. At the point we don't have the data,
	152	* we're just creating a placeholder so that multiple threads don't all
	153	* go off and read the same blocks.
	154	*/
	155	static vdev_cache_entry_t *
	156	vdev_cache_allocate(zio_t *zio)
	157	{
	158	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
	159	uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
	160	vdev_cache_entry_t *ve;
	161
	162	ASSERT(MUTEX_HELD(&vc->vc_lock));
	163
	164	if (zfs_vdev_cache_size == 0)
	165	return (NULL);
	166
	167	/*
	168	* If adding a new entry would exceed the cache size,
	169	* evict the oldest entry (LRU).
	170	*/
	171	if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
	172	zfs_vdev_cache_size) {
	173	ve = avl_first(&vc->vc_lastused_tree);
b128c09f	174	if (ve->ve_fill_io != NULL)
34dc7c2f	175	return (NULL);
34dc7c2f BB	176	ASSERT(ve->ve_hits != 0);
	177	vdev_cache_evict(vc, ve);
	178	}
	179
b8d06fca	180	ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_PUSHPAGE);
34dc7c2f	181	ve->ve_offset = offset;
428870ff	182	ve->ve_lastused = ddi_get_lbolt();
34dc7c2f BB	183	ve->ve_data = zio_buf_alloc(VCBS);
	184
	185	avl_add(&vc->vc_offset_tree, ve);
	186	avl_add(&vc->vc_lastused_tree, ve);
	187
	188	return (ve);
	189	}
	190
	191	static void
	192	vdev_cache_hit(vdev_cache_t vc, vdev_cache_entry_t ve, zio_t *zio)
	193	{
	194	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
	195
	196	ASSERT(MUTEX_HELD(&vc->vc_lock));
	197	ASSERT(ve->ve_fill_io == NULL);
	198
428870ff	199	if (ve->ve_lastused != ddi_get_lbolt()) {
34dc7c2f	200	avl_remove(&vc->vc_lastused_tree, ve);
428870ff	201	ve->ve_lastused = ddi_get_lbolt();
34dc7c2f BB	202	avl_add(&vc->vc_lastused_tree, ve);
	203	}
	204
	205	ve->ve_hits++;
	206	bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
	207	}
	208
	209	/*
	210	* Fill a previously allocated cache entry with data.
	211	*/
	212	static void
d164b209	213	vdev_cache_fill(zio_t *fio)
34dc7c2f	214	{
d164b209	215	vdev_t *vd = fio->io_vd;
34dc7c2f	216	vdev_cache_t *vc = &vd->vdev_cache;
d164b209 BB	217	vdev_cache_entry_t *ve = fio->io_private;
d164b209 BB	218	zio_t *pio;
34dc7c2f	219
d164b209	220	ASSERT(fio->io_size == VCBS);
34dc7c2f BB	221
	222	/*
	223	* Add data to the cache.
	224	*/
	225	mutex_enter(&vc->vc_lock);
	226
d164b209 BB	227	ASSERT(ve->ve_fill_io == fio);
	228	ASSERT(ve->ve_offset == fio->io_offset);
	229	ASSERT(ve->ve_data == fio->io_data);
34dc7c2f BB	230
	231	ve->ve_fill_io = NULL;
	232
	233	/*
	234	* Even if this cache line was invalidated by a missed write update,
	235	* any reads that were queued up before the missed update are still
	236	* valid, so we can satisfy them from this line before we evict it.
	237	*/
d164b209 BB	238	while ((pio = zio_walk_parents(fio)) != NULL)
d164b209 BB	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	/*
	248	* Read data from the cache. Returns 0 on cache hit, errno on a miss.
	249	*/
	250	int
	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);
1fde1e37	256	ASSERTV(uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);)
34dc7c2f BB	257	zio_t *fio;
	258
	259	ASSERT(zio->io_type == ZIO_TYPE_READ);
	260
	261	if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
	262	return (EINVAL);
	263
	264	if (zio->io_size > zfs_vdev_cache_max)
	265	return (EOVERFLOW);
	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))
34dc7c2f BB	271	return (EXDEV);
	272
	273	ASSERT(cache_phase + zio->io_size <= VCBS);
	274
	275	mutex_enter(&vc->vc_lock);
	276
b8d06fca	277	ve_search = kmem_alloc(sizeof(vdev_cache_entry_t), KM_PUSHPAGE);
5fed499d BB	278	ve_search->ve_offset = cache_offset;
	279	ve = avl_find(&vc->vc_offset_tree, ve_search, NULL);
	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);
	285	return (ESTALE);
	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);
	292	VDCSTAT_BUMP(vdc_stat_delegations);
	293	return (0);
	294	}
	295
	296	vdev_cache_hit(vc, ve, zio);
	297	zio_vdev_io_bypass(zio);
	298
	299	mutex_exit(&vc->vc_lock);
34dc7c2f BB	300	VDCSTAT_BUMP(vdc_stat_hits);
	301	return (0);
	302	}
	303
	304	ve = vdev_cache_allocate(zio);
	305
	306	if (ve == NULL) {
	307	mutex_exit(&vc->vc_lock);
	308	return (ENOMEM);
	309	}
	310
b128c09f	311	fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
34dc7c2f	312	ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_CACHE_FILL,
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
	323	return (0);
	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
	340	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
	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 {
	357	bcopy((char *)zio->io_data + start - io_start,
	358	ve->ve_data + start - ve->ve_offset, end - start);
	359	}
	360	ve = AVL_NEXT(&vc->vc_offset_tree, ve);
	361	}
	362	mutex_exit(&vc->vc_lock);
	363	}
	364
	365	void
	366	vdev_cache_purge(vdev_t *vd)
	367	{
	368	vdev_cache_t *vc = &vd->vdev_cache;
	369	vdev_cache_entry_t *ve;
	370
	371	mutex_enter(&vc->vc_lock);
	372	while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
	373	vdev_cache_evict(vc, ve);
	374	mutex_exit(&vc->vc_lock);
	375	}
	376
	377	void
	378	vdev_cache_init(vdev_t *vd)
	379	{
	380	vdev_cache_t *vc = &vd->vdev_cache;
	381
382	mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
383
384	avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
385	sizeof (vdev_cache_entry_t),
386	offsetof(struct vdev_cache_entry, ve_offset_node));
387
388	avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
389	sizeof (vdev_cache_entry_t),
390	offsetof(struct vdev_cache_entry, ve_lastused_node));
391	}
392
393	void
394	vdev_cache_fini(vdev_t *vd)
395	{
396	vdev_cache_t *vc = &vd->vdev_cache;
397
398	vdev_cache_purge(vd);
399
400	avl_destroy(&vc->vc_offset_tree);
401	avl_destroy(&vc->vc_lastused_tree);
402
403	mutex_destroy(&vc->vc_lock);
404	}
405
406	void
407	vdev_cache_stat_init(void)
408	{
409	vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
410	KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
411	KSTAT_FLAG_VIRTUAL);
412	if (vdc_ksp != NULL) {
413	vdc_ksp->ks_data = &vdc_stats;
414	kstat_install(vdc_ksp);
415	}
416	}
417
418	void
419	vdev_cache_stat_fini(void)
420	{
421	if (vdc_ksp != NULL) {
422	kstat_delete(vdc_ksp);
423	vdc_ksp = NULL;
424	}
425	}
c409e464 BB	426
	427	#if defined(_KERNEL) && defined(HAVE_SPL)
	428	module_param(zfs_vdev_cache_max, int, 0644);
	429	MODULE_PARM_DESC(zfs_vdev_cache_max, "Inflate reads small than max");
	430
	431	module_param(zfs_vdev_cache_size, int, 0444);
	432	MODULE_PARM_DESC(zfs_vdev_cache_size, "Total size of the per-disk cache");
	433
	434	module_param(zfs_vdev_cache_bshift, int, 0644);
	435	MODULE_PARM_DESC(zfs_vdev_cache_bshift, "Shift size to inflate reads too");
	436	#endif