Reduce dirty records memory usage

[zfs.git] / module / zfs / zfeature.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 https://opensource.org/licenses/CDDL-1.0.
 * 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 (c) 2011, 2015 by Delphix. All rights reserved.
 */

#include <sys/zfs_context.h>
#include <sys/zfeature.h>
#include <sys/dmu.h>
#include <sys/nvpair.h>
#include <sys/zap.h>
#include <sys/dmu_tx.h>
#include "zfeature_common.h"
#include <sys/spa_impl.h>

/*
 * ZFS Feature Flags
 * -----------------
 *
 * ZFS feature flags are used to provide fine-grained versioning to the ZFS
 * on-disk format. Once enabled on a pool feature flags replace the old
 * spa_version() number.
 *
 * Each new on-disk format change will be given a uniquely identifying string
 * GUID rather than a version number. This avoids the problem of different
 * organizations creating new on-disk formats with the same version number. To
 * keep feature GUIDs unique they should consist of the reverse dns name of the
 * organization which implemented the feature and a short name for the feature,
 * separated by a colon (e.g. com.delphix:async_destroy).
 *
 * Reference Counts
 * ----------------
 *
 * Within each pool features can be in one of three states: disabled, enabled,
 * or active. These states are differentiated by a reference count stored on
 * disk for each feature:
 *
 *   1) If there is no reference count stored on disk the feature is disabled.
 *   2) If the reference count is 0 a system administrator has enabled the
 *      feature, but the feature has not been used yet, so no on-disk
 *      format changes have been made.
 *   3) If the reference count is greater than 0 the feature is active.
 *      The format changes required by the feature are currently on disk.
 *      Note that if the feature's format changes are reversed the feature
 *      may choose to set its reference count back to 0.
 *
 * Feature flags makes no differentiation between non-zero reference counts
 * for an active feature (e.g. a reference count of 1 means the same thing as a
 * reference count of 27834721), but feature implementations may choose to use
 * the reference count to store meaningful information. For example, a new RAID
 * implementation might set the reference count to the number of vdevs using
 * it. If all those disks are removed from the pool the feature goes back to
 * having a reference count of 0.
 *
 * It is the responsibility of the individual features to maintain a non-zero
 * reference count as long as the feature's format changes are present on disk.
 *
 * Dependencies
 * ------------
 *
 * Each feature may depend on other features. The only effect of this
 * relationship is that when a feature is enabled all of its dependencies are
 * automatically enabled as well. Any future work to support disabling of
 * features would need to ensure that features cannot be disabled if other
 * enabled features depend on them.
 *
 * On-disk Format
 * --------------
 *
 * When feature flags are enabled spa_version() is set to SPA_VERSION_FEATURES
 * (5000). In order for this to work the pool is automatically upgraded to
 * SPA_VERSION_BEFORE_FEATURES (28) first, so all pre-feature flags on disk
 * format changes will be in use.
 *
 * Information about features is stored in 3 ZAP objects in the pool's MOS.
 * These objects are linked to by the following names in the pool directory
 * object:
 *
 * 1) features_for_read: feature GUID -> reference count
 *    Features needed to open the pool for reading.
 * 2) features_for_write: feature GUID -> reference count
 *    Features needed to open the pool for writing.
 * 3) feature_descriptions: feature GUID -> descriptive string
 *    A human readable string.
 *
 * All enabled features appear in either features_for_read or
 * features_for_write, but not both.
 *
 * To open a pool in read-only mode only the features listed in
 * features_for_read need to be supported.
 *
 * To open the pool in read-write mode features in both features_for_read and
 * features_for_write need to be supported.
 *
 * Some features may be required to read the ZAP objects containing feature
 * information. To allow software to check for compatibility with these features
 * before the pool is opened their names must be stored in the label in a
 * new "features_for_read" entry (note that features that are only required
 * to write to a pool never need to be stored in the label since the
 * features_for_write ZAP object can be read before the pool is written to).
 * To save space in the label features must be explicitly marked as needing to
 * be written to the label. Also, reference counts are not stored in the label,
 * instead any feature whose reference count drops to 0 is removed from the
 * label.
 *
 * Adding New Features
 * -------------------
 *
 * Features must be registered in zpool_feature_init() function in
 * zfeature_common.c using the zfeature_register() function. This function
 * has arguments to specify if the feature should be stored in the
 * features_for_read or features_for_write ZAP object and if it needs to be
 * written to the label when active.
 *
 * Once a feature is registered it will appear as a "feature@<feature name>"
 * property which can be set by an administrator. Feature implementors should
 * use the spa_feature_is_enabled() and spa_feature_is_active() functions to
 * query the state of a feature and the spa_feature_incr() and
 * spa_feature_decr() functions to change an enabled feature's reference count.
 * Reference counts may only be updated in the syncing context.
 *
 * Features may not perform enable-time initialization. Instead, any such
 * initialization should occur when the feature is first used. This design
 * enforces that on-disk changes be made only when features are used. Code
 * should only check if a feature is enabled using spa_feature_is_enabled(),
 * not by relying on any feature specific metadata existing. If a feature is
 * enabled, but the feature's metadata is not on disk yet then it should be
 * created as needed.
 *
 * As an example, consider the com.delphix:async_destroy feature. This feature
 * relies on the existence of a bptree in the MOS that store blocks for
 * asynchronous freeing. This bptree is not created when async_destroy is
 * enabled. Instead, when a dataset is destroyed spa_feature_is_enabled() is
 * called to check if async_destroy is enabled. If it is and the bptree object
 * does not exist yet, the bptree object is created as part of the dataset
 * destroy and async_destroy's reference count is incremented to indicate it
 * has made an on-disk format change. Later, after the destroyed dataset's
 * blocks have all been asynchronously freed there is no longer any use for the
 * bptree object, so it is destroyed and async_destroy's reference count is
 * decremented back to 0 to indicate that it has undone its on-disk format
 * changes.
 */

typedef enum {
        FEATURE_ACTION_INCR,
        FEATURE_ACTION_DECR,
} feature_action_t;

/*
 * Checks that the active features in the pool are supported by
 * this software.  Adds each unsupported feature (name -> description) to
 * the supplied nvlist.
 */
boolean_t
spa_features_check(spa_t *spa, boolean_t for_write,
    nvlist_t *unsup_feat, nvlist_t *enabled_feat)
{
        objset_t *os = spa->spa_meta_objset;
        boolean_t supported;
        zap_cursor_t *zc;
        zap_attribute_t *za;
        uint64_t obj = for_write ?
            spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
        char *buf;

        zc = kmem_alloc(sizeof (zap_cursor_t), KM_SLEEP);
        za = zap_attribute_alloc();
        buf = kmem_alloc(MAXPATHLEN, KM_SLEEP);

        supported = B_TRUE;
        for (zap_cursor_init(zc, os, obj);
            zap_cursor_retrieve(zc, za) == 0;
            zap_cursor_advance(zc)) {
                ASSERT(za->za_integer_length == sizeof (uint64_t) &&
                    za->za_num_integers == 1);

                if (NULL != enabled_feat) {
                        fnvlist_add_uint64(enabled_feat, za->za_name,
                            za->za_first_integer);
                }

                if (za->za_first_integer != 0 &&
                    !zfeature_is_supported(za->za_name)) {
                        supported = B_FALSE;

                        if (NULL != unsup_feat) {
                                const char *desc = "";

                                if (zap_lookup(os, spa->spa_feat_desc_obj,
                                    za->za_name, 1, MAXPATHLEN, buf) == 0)
                                        desc = buf;

                                VERIFY(nvlist_add_string(unsup_feat,
                                    za->za_name, desc) == 0);
                        }
                }
        }
        zap_cursor_fini(zc);

        kmem_free(buf, MAXPATHLEN);
        zap_attribute_free(za);
        kmem_free(zc, sizeof (zap_cursor_t));

        return (supported);
}

/*
 * Use an in-memory cache of feature refcounts for quick retrieval.
 *
 * Note: well-designed features will not need to use this; they should
 * use spa_feature_is_enabled() and spa_feature_is_active() instead.
 * However, this is non-static for zdb, zhack, and spa_add_feature_stats().
 */
int
feature_get_refcount(spa_t *spa, zfeature_info_t *feature, uint64_t *res)
{
        ASSERT(VALID_FEATURE_FID(feature->fi_feature));
        if (spa->spa_feat_refcount_cache[feature->fi_feature] ==
            SPA_FEATURE_DISABLED) {
                return (SET_ERROR(ENOTSUP));
        }
        *res = spa->spa_feat_refcount_cache[feature->fi_feature];
        return (0);
}

/*
 * Note: well-designed features will not need to use this; they should
 * use spa_feature_is_enabled() and spa_feature_is_active() instead.
 * However, this is non-static for zdb and zhack.
 */
int
feature_get_refcount_from_disk(spa_t *spa, zfeature_info_t *feature,
    uint64_t *res)
{
        int err;
        uint64_t refcount;
        uint64_t zapobj = (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
            spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;

        /*
         * If the pool is currently being created, the feature objects may not
         * have been allocated yet.  Act as though all features are disabled.
         */
        if (zapobj == 0)
                return (SET_ERROR(ENOTSUP));

        err = zap_lookup(spa->spa_meta_objset, zapobj,
            feature->fi_guid, sizeof (uint64_t), 1, &refcount);
        if (err != 0) {
                if (err == ENOENT)
                        return (SET_ERROR(ENOTSUP));
                else
                        return (err);
        }
        *res = refcount;
        return (0);
}


static int
feature_get_enabled_txg(spa_t *spa, zfeature_info_t *feature, uint64_t *res)
{
        uint64_t enabled_txg_obj __maybe_unused = spa->spa_feat_enabled_txg_obj;

        ASSERT(zfeature_depends_on(feature->fi_feature,
            SPA_FEATURE_ENABLED_TXG));

        if (!spa_feature_is_enabled(spa, feature->fi_feature)) {
                return (SET_ERROR(ENOTSUP));
        }

        ASSERT(enabled_txg_obj != 0);

        VERIFY0(zap_lookup(spa->spa_meta_objset, spa->spa_feat_enabled_txg_obj,
            feature->fi_guid, sizeof (uint64_t), 1, res));

        return (0);
}

/*
 * This function is non-static for zhack; it should otherwise not be used
 * outside this file.
 */
void
feature_sync(spa_t *spa, zfeature_info_t *feature, uint64_t refcount,
    dmu_tx_t *tx)
{
        ASSERT(VALID_FEATURE_OR_NONE(feature->fi_feature));
        uint64_t zapobj = (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
            spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;
        VERIFY0(zap_update(spa->spa_meta_objset, zapobj, feature->fi_guid,
            sizeof (uint64_t), 1, &refcount, tx));

        /*
         * feature_sync is called directly from zhack, allowing the
         * creation of arbitrary features whose fi_feature field may
         * be greater than SPA_FEATURES. When called from zhack, the
         * zfeature_info_t object's fi_feature field will be set to
         * SPA_FEATURE_NONE.
         */
        if (feature->fi_feature != SPA_FEATURE_NONE) {
                uint64_t *refcount_cache =
                    &spa->spa_feat_refcount_cache[feature->fi_feature];
                VERIFY3U(*refcount_cache, ==,
                    atomic_swap_64(refcount_cache, refcount));
        }

        if (refcount == 0)
                spa_deactivate_mos_feature(spa, feature->fi_guid);
        else if (feature->fi_flags & ZFEATURE_FLAG_MOS)
                spa_activate_mos_feature(spa, feature->fi_guid, tx);
}

/*
 * This function is non-static for zhack; it should otherwise not be used
 * outside this file.
 */
void
feature_enable_sync(spa_t *spa, zfeature_info_t *feature, dmu_tx_t *tx)
{
        uint64_t initial_refcount =
            (feature->fi_flags & ZFEATURE_FLAG_ACTIVATE_ON_ENABLE) ? 1 : 0;
        uint64_t zapobj = (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
            spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;

        ASSERT(0 != zapobj);
        ASSERT(zfeature_is_valid_guid(feature->fi_guid));
        ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);

        /*
         * If the feature is already enabled, ignore the request.
         */
        if (zap_contains(spa->spa_meta_objset, zapobj, feature->fi_guid) == 0)
                return;

        for (int i = 0; feature->fi_depends[i] != SPA_FEATURE_NONE; i++)
                spa_feature_enable(spa, feature->fi_depends[i], tx);

        VERIFY0(zap_update(spa->spa_meta_objset, spa->spa_feat_desc_obj,
            feature->fi_guid, 1, strlen(feature->fi_desc) + 1,
            feature->fi_desc, tx));

        feature_sync(spa, feature, initial_refcount, tx);

        if (spa_feature_is_enabled(spa, SPA_FEATURE_ENABLED_TXG)) {
                uint64_t enabling_txg = dmu_tx_get_txg(tx);

                if (spa->spa_feat_enabled_txg_obj == 0ULL) {
                        spa->spa_feat_enabled_txg_obj =
                            zap_create_link(spa->spa_meta_objset,
                            DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
                            DMU_POOL_FEATURE_ENABLED_TXG, tx);
                }
                spa_feature_incr(spa, SPA_FEATURE_ENABLED_TXG, tx);

                VERIFY0(zap_add(spa->spa_meta_objset,
                    spa->spa_feat_enabled_txg_obj, feature->fi_guid,
                    sizeof (uint64_t), 1, &enabling_txg, tx));
        }

        /*
         * Errata #4 is mostly a problem with encrypted datasets, but it
         * is also a problem where the old encryption feature did not
         * depend on the bookmark_v2 feature. If the pool does not have
         * any encrypted datasets we can resolve this issue simply by
         * enabling this dependency.
         */
        if (spa->spa_errata == ZPOOL_ERRATA_ZOL_8308_ENCRYPTION &&
            spa_feature_is_enabled(spa, SPA_FEATURE_ENCRYPTION) &&
            !spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION) &&
            feature->fi_feature == SPA_FEATURE_BOOKMARK_V2)
                spa->spa_errata = 0;

        /*
         * Convert the old on-disk error log to the new format when activating
         * the head_errlog feature.
         */
        if (feature->fi_feature == SPA_FEATURE_HEAD_ERRLOG)
                spa_upgrade_errlog(spa, tx);
}

static void
feature_do_action(spa_t *spa, spa_feature_t fid, feature_action_t action,
    dmu_tx_t *tx)
{
        uint64_t refcount = 0;
        zfeature_info_t *feature = &spa_feature_table[fid];
        uint64_t zapobj __maybe_unused =
            (feature->fi_flags & ZFEATURE_FLAG_READONLY_COMPAT) ?
            spa->spa_feat_for_write_obj : spa->spa_feat_for_read_obj;

        ASSERT(VALID_FEATURE_FID(fid));
        ASSERT(0 != zapobj);
        ASSERT(zfeature_is_valid_guid(feature->fi_guid));

        ASSERT(dmu_tx_is_syncing(tx));
        ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);

        VERIFY3U(feature_get_refcount(spa, feature, &refcount), !=, ENOTSUP);

        switch (action) {
        case FEATURE_ACTION_INCR:
                VERIFY3U(refcount, !=, UINT64_MAX);
                refcount++;
                break;
        case FEATURE_ACTION_DECR:
                VERIFY3U(refcount, !=, 0);
                refcount--;
                break;
        default:
                ASSERT(0);
                break;
        }

        feature_sync(spa, feature, refcount, tx);
}

void
spa_feature_create_zap_objects(spa_t *spa, dmu_tx_t *tx)
{
        /*
         * We create feature flags ZAP objects in two instances: during pool
         * creation and during pool upgrade.
         */
        ASSERT((!spa->spa_sync_on && tx->tx_txg == TXG_INITIAL) ||
            dsl_pool_sync_context(spa_get_dsl(spa)));

        spa->spa_feat_for_read_obj = zap_create_link(spa->spa_meta_objset,
            DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
            DMU_POOL_FEATURES_FOR_READ, tx);
        spa->spa_feat_for_write_obj = zap_create_link(spa->spa_meta_objset,
            DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
            DMU_POOL_FEATURES_FOR_WRITE, tx);
        spa->spa_feat_desc_obj = zap_create_link(spa->spa_meta_objset,
            DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
            DMU_POOL_FEATURE_DESCRIPTIONS, tx);
}

/*
 * Enable any required dependencies, then enable the requested feature.
 */
void
spa_feature_enable(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
{
        ASSERT3U(spa_version(spa), >=, SPA_VERSION_FEATURES);
        ASSERT(VALID_FEATURE_FID(fid));
        feature_enable_sync(spa, &spa_feature_table[fid], tx);
}

void
spa_feature_incr(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
{
        feature_do_action(spa, fid, FEATURE_ACTION_INCR, tx);
}

void
spa_feature_decr(spa_t *spa, spa_feature_t fid, dmu_tx_t *tx)
{
        feature_do_action(spa, fid, FEATURE_ACTION_DECR, tx);
}

boolean_t
spa_feature_is_enabled(spa_t *spa, spa_feature_t fid)
{
        int err;
        uint64_t refcount = 0;

        ASSERT(VALID_FEATURE_FID(fid));
        if (spa_version(spa) < SPA_VERSION_FEATURES)
                return (B_FALSE);

        err = feature_get_refcount(spa, &spa_feature_table[fid], &refcount);
        ASSERT(err == 0 || err == ENOTSUP);
        return (err == 0);
}

boolean_t
spa_feature_is_active(spa_t *spa, spa_feature_t fid)
{
        int err;
        uint64_t refcount = 0;

        ASSERT(VALID_FEATURE_FID(fid));
        if (spa_version(spa) < SPA_VERSION_FEATURES)
                return (B_FALSE);

        err = feature_get_refcount(spa, &spa_feature_table[fid], &refcount);
        ASSERT(err == 0 || err == ENOTSUP);
        return (err == 0 && refcount > 0);
}

/*
 * For the feature specified by fid (which must depend on
 * SPA_FEATURE_ENABLED_TXG), return the TXG at which it was enabled in the
 * OUT txg argument.
 *
 * Returns B_TRUE if the feature is enabled, in which case txg will be filled
 * with the transaction group in which the specified feature was enabled.
 * Returns B_FALSE otherwise (i.e. if the feature is not enabled).
 */
boolean_t
spa_feature_enabled_txg(spa_t *spa, spa_feature_t fid, uint64_t *txg)
{
        int err;

        ASSERT(VALID_FEATURE_FID(fid));
        if (spa_version(spa) < SPA_VERSION_FEATURES)
                return (B_FALSE);

        err = feature_get_enabled_txg(spa, &spa_feature_table[fid], txg);
        ASSERT(err == 0 || err == ENOTSUP);

        return (err == 0);
}