[fs/lustre-release.git] / lustre / lod / lod_pool.c

/*
 * GPL HEADER START
 *
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License version 2 only,
 * as published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License version 2 for more details (a copy is included
 * in the LICENSE file that accompanied this code).
 *
 * You should have received a copy of the GNU General Public License
 * version 2 along with this program; If not, see
 * http://www.gnu.org/licenses/gpl-2.0.html
 *
 * GPL HEADER END
 */
/*
 * Copyright  2008 Sun Microsystems, Inc. All rights reserved
 * Use is subject to license terms.
 *
 * Copyright (c) 2012, 2017, Intel Corporation.
 */
/*
 * This file is part of Lustre, http://www.lustre.org/
 * Lustre is a trademark of Sun Microsystems, Inc.
 */
/*
 * lustre/lod/lod_pool.c
 *
 * OST pool methods
 *
 * This file provides code related to the Logical Object Device (LOD)
 * handling of OST Pools on the MDT.  Pools are named lists of targets
 * that allow userspace to group targets that share a particlar property
 * together so that users or kernel helpers can make decisions about file
 * allocation based on these properties.  For example, pools could be
 * defined based on fault domains (e.g. separate racks of server nodes) so
 * that RAID-1 mirroring could select targets from independent fault
 * domains, or pools could define target performance characteristics so
 * that applicatins could select IOP-optimized storage or stream-optimized
 * storage for a particular output file.
 *
 * This file handles creation, lookup, and removal of pools themselves, as
 * well as adding and removing targets to pools.  It also handles lprocfs
 * display of configured pool.  The pools are accessed by name in the pool
 * hash, and are refcounted to ensure proper pool structure lifetimes.
 *
 * Author: Jacques-Charles LAFOUCRIERE <jc.lafoucriere@cea.fr>
 * Author: Alex Lyashkov <Alexey.Lyashkov@Sun.COM>
 * Author: Nathaniel Rutman <Nathan.Rutman@Sun.COM>
 */

#define DEBUG_SUBSYSTEM S_LOV

#include <libcfs/libcfs.h>
#include <libcfs/linux/linux-hash.h>
#include <obd.h>
#include "lod_internal.h"

#define pool_tgt(_p, _i) OST_TGT(lu2lod_dev((_p)->pool_lobd->obd_lu_dev), \
                                 (_p)->pool_obds.op_array[_i])

/**
 * Get a reference on the specified pool.
 *
 * To ensure the pool descriptor is not freed before the caller is finished
 * with it.  Any process that is accessing \a pool directly needs to hold
 * reference on it, including /proc since a userspace thread may be holding
 * the /proc file open and busy in the kernel.
 *
 * \param[in] pool      pool descriptor on which to gain reference
 */
static void pool_getref(struct pool_desc *pool)
{
        CDEBUG(D_INFO, "pool %p\n", pool);
        atomic_inc(&pool->pool_refcount);
}

/**
 * Drop a reference on the specified pool and free its memory if needed.
 *
 * One reference is held by the LOD OBD device while it is configured, from
 * the time the configuration log defines the pool until the time when it is
 * dropped when the LOD OBD is cleaned up or the pool is deleted.  This means
 * that the pool will not be freed while the LOD device is configured, unless
 * it is explicitly destroyed by the sysadmin.  The pool structure is freed
 * after the last reference on the structure is released.
 *
 * \param[in] pool      pool descriptor to drop reference on and possibly free
 */
void lod_pool_putref(struct pool_desc *pool)
{
        CDEBUG(D_INFO, "pool %p\n", pool);
        if (atomic_dec_and_test(&pool->pool_refcount)) {
                LASSERT(list_empty(&pool->pool_list));
                LASSERT(pool->pool_proc_entry == NULL);
                tgt_pool_free(&(pool->pool_rr.lqr_pool));
                tgt_pool_free(&(pool->pool_obds));
                kfree_rcu(pool, pool_rcu);
                EXIT;
        }
}

static u32 pool_hashfh(const void *data, u32 len, u32 seed)
{
        const char *pool_name = data;

        return hashlen_hash(cfs_hashlen_string((void *)(unsigned long)seed,
                                               pool_name));
}

static int pool_cmpfn(struct rhashtable_compare_arg *arg, const void *obj)
{
        const struct pool_desc *pool = obj;
        const char *pool_name = arg->key;

        return strcmp(pool_name, pool->pool_name);
}

static const struct rhashtable_params pools_hash_params = {
        .key_len        = 1, /* actually variable */
        .key_offset     = offsetof(struct pool_desc, pool_name),
        .head_offset    = offsetof(struct pool_desc, pool_hash),
        .hashfn         = pool_hashfh,
        .obj_cmpfn      = pool_cmpfn,
        .automatic_shrinking = true,
};

/*
 * Methods for /proc seq_file iteration of the defined pools.
 */

#define POOL_IT_MAGIC 0xB001CEA0
struct lod_pool_iterator {
        unsigned int      lpi_magic;    /* POOL_IT_MAGIC */
        unsigned int      lpi_idx;      /* from 0 to pool_tgt_size - 1 */
        struct pool_desc *lpi_pool;
};

/**
 * Return the next configured target within one pool for seq_file iteration.
 *
 * Iterator is used to go through the target entries of a single pool
 * (i.e. the list of OSTs configured for a named pool).
 * lpi_idx is the current target index in the pool's op_array[].
 *
 * The return type is a void * because this function is one of the
 * struct seq_operations methods and must match the function template.
 *
 * \param[in] seq       /proc sequence file iteration tracking structure
 * \param[in] v         unused
 * \param[in] pos       position within iteration; 0 to number of targets - 1
 *
 * \retval      struct pool_iterator of the next pool descriptor
 */
static void *pool_proc_next(struct seq_file *seq, void *v, loff_t *pos)
{
        struct lod_pool_iterator *iter = seq->private;
        int prev_idx;

        LASSERTF(iter->lpi_magic == POOL_IT_MAGIC, "%08X\n", iter->lpi_magic);

        (*pos)++;
        /* test if end of file */
        if (*pos > pool_tgt_count(iter->lpi_pool))
                return NULL;

        OBD_FAIL_TIMEOUT(OBD_FAIL_OST_LIST_ASSERT, cfs_fail_val);

        /* iterate to find a non empty entry */
        prev_idx = iter->lpi_idx;
        iter->lpi_idx++;
        if (iter->lpi_idx >= pool_tgt_count(iter->lpi_pool)) {
                iter->lpi_idx = prev_idx; /* we stay on the last entry */
                return NULL;
        }

        /* return != NULL to continue */
        return iter;
}

/**
 * Start seq_file iteration via /proc for a single pool.
 *
 * The \a pos parameter may be non-zero, indicating that the iteration
 * is starting at some offset in the target list.  Use the seq_file
 * private field to memorize the iterator so we can free it at stop().
 * Need to restore the private pointer to the pool before freeing it.
 *
 * \param[in] seq       new sequence file structure to initialize
 * \param[in] pos       initial target number at which to start iteration
 *
 * \retval              initialized pool iterator private structure
 * \retval              NULL if \a pos exceeds the number of targets in \a pool
 * \retval              negative error number on failure
 */
static void *pool_proc_start(struct seq_file *seq, loff_t *pos)
{
        struct pool_desc *pool = seq->private;
        struct lod_pool_iterator *iter;

        pool_getref(pool);
        if ((pool_tgt_count(pool) == 0) ||
            (*pos >= pool_tgt_count(pool))) {
                /* iter is not created, so stop() has no way to
                 * find pool to dec ref */
                lod_pool_putref(pool);
                return NULL;
        }

        OBD_ALLOC_PTR(iter);
        if (iter == NULL)
                return ERR_PTR(-ENOMEM);
        iter->lpi_magic = POOL_IT_MAGIC;
        iter->lpi_pool = pool;
        iter->lpi_idx = 0;

        seq->private = iter;
        down_read(&pool_tgt_rw_sem(pool));
        if (*pos > 0) {
                loff_t i;
                void *ptr;

                i = 0;
                do {
                        ptr = pool_proc_next(seq, &iter, &i);
                } while ((i < *pos) && (ptr != NULL));

                return ptr;
        }

        return iter;
}

/**
 * Finish seq_file iteration for a single pool.
 *
 * Once iteration has been completed, the pool_iterator struct must be
 * freed, and the seq_file private pointer restored to the pool, as it
 * was initially when pool_proc_start() was called.
 *
 * In some cases the stop() method may be called 2 times, without calling
 * the start() method (see seq_read() from fs/seq_file.c). We have to free
 * the private iterator struct only if seq->private points to the iterator.
 *
 * \param[in] seq       sequence file structure to clean up
 * \param[in] v         (unused)
 */
static void pool_proc_stop(struct seq_file *seq, void *v)
{
        struct lod_pool_iterator *iter = seq->private;

        if (iter != NULL && iter->lpi_magic == POOL_IT_MAGIC) {
                up_read(&pool_tgt_rw_sem(iter->lpi_pool));
                seq->private = iter->lpi_pool;
                lod_pool_putref(iter->lpi_pool);
                OBD_FREE_PTR(iter);
        }
}

/**
 * Print out one target entry from the pool for seq_file iteration.
 *
 * The currently referenced pool target is given by op_array[lpi_idx].
 *
 * \param[in] seq       new sequence file structure to initialize
 * \param[in] v         (unused)
 */
static int pool_proc_show(struct seq_file *seq, void *v)
{
        struct lod_pool_iterator *iter = v;
        struct lod_tgt_desc  *tgt;

        LASSERTF(iter->lpi_magic == POOL_IT_MAGIC, "%08X\n", iter->lpi_magic);
        LASSERT(iter->lpi_pool != NULL);
        LASSERT(iter->lpi_idx <= pool_tgt_count(iter->lpi_pool));

        tgt = pool_tgt(iter->lpi_pool, iter->lpi_idx);
        if (tgt != NULL)
                seq_printf(seq, "%s\n", obd_uuid2str(&(tgt->ltd_uuid)));

        return 0;
}

static const struct seq_operations pool_proc_ops = {
        .start  = pool_proc_start,
        .next   = pool_proc_next,
        .stop   = pool_proc_stop,
        .show   = pool_proc_show,
};

/**
 * Open a new /proc file for seq_file iteration of targets in one pool.
 *
 * Initialize the seq_file private pointer to reference the pool.
 *
 * \param inode inode to store iteration state for /proc
 * \param file  file descriptor to store iteration methods
 *
 * \retval      0 for success
 * \retval      negative error number on failure
 */
static int pool_proc_open(struct inode *inode, struct file *file)
{
        int rc;

        rc = seq_open(file, &pool_proc_ops);
        if (!rc) {
                struct seq_file *seq = file->private_data;
                seq->private = PDE_DATA(inode);
        }
        return rc;
}

static struct file_operations pool_proc_operations = {
        .open           = pool_proc_open,
        .read           = seq_read,
        .llseek         = seq_lseek,
        .release        = seq_release,
};

/**
 * Dump the pool target list into the Lustre debug log.
 *
 * This is a debugging function to allow dumping the list of targets
 * in \a pool to the Lustre kernel debug log at the given \a level.
 *
 * This is not currently called by any existing code, but can be called
 * from within gdb/crash to display the contents of the pool, or from
 * code under development.
 *
 * \param[in] level     Lustre debug level (D_INFO, D_WARN, D_ERROR, etc)
 * \param[in] pool      pool descriptor to be dumped
 */
void lod_dump_pool(int level, struct pool_desc *pool)
{
        unsigned int i;

        pool_getref(pool);

        CDEBUG(level, "pool "LOV_POOLNAMEF" has %d members\n",
               pool->pool_name, pool->pool_obds.op_count);
        down_read(&pool_tgt_rw_sem(pool));

        for (i = 0; i < pool_tgt_count(pool) ; i++) {
                if (!pool_tgt(pool, i) || !(pool_tgt(pool, i))->ltd_exp)
                        continue;
                CDEBUG(level, "pool "LOV_POOLNAMEF"[%d] = %s\n",
                       pool->pool_name, i,
                       obd_uuid2str(&((pool_tgt(pool, i))->ltd_uuid)));
        }

        up_read(&pool_tgt_rw_sem(pool));
        lod_pool_putref(pool);
}

static void pools_hash_exit(void *vpool, void *data)
{
        struct pool_desc *pool = vpool;

        lod_pool_putref(pool);
}

int lod_pool_hash_init(struct rhashtable *tbl)
{
        return rhashtable_init(tbl, &pools_hash_params);
}

void lod_pool_hash_destroy(struct rhashtable *tbl)
{
        rhashtable_free_and_destroy(tbl, pools_hash_exit, NULL);
}

/**
 * Allocate a new pool for the specified device.
 *
 * Allocate a new pool_desc structure for the specified \a new_pool
 * device to create a pool with the given \a poolname.  The new pool
 * structure is created with a single reference, and is freed when the
 * reference count drops to zero.
 *
 * \param[in] obd       Lustre OBD device on which to add a pool iterator
 * \param[in] poolname  the name of the pool to be created
 *
 * \retval              0 in case of success
 * \retval              negative error code in case of error
 */
int lod_pool_new(struct obd_device *obd, char *poolname)
{
        struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
        struct pool_desc  *new_pool;
        int rc;
        ENTRY;

        if (strlen(poolname) > LOV_MAXPOOLNAME)
                RETURN(-ENAMETOOLONG);

        /* OBD_ALLOC_* doesn't work with direct kfree_rcu use */
        new_pool = kmalloc(sizeof(*new_pool), GFP_KERNEL);
        if (new_pool == NULL)
                RETURN(-ENOMEM);

        strlcpy(new_pool->pool_name, poolname, sizeof(new_pool->pool_name));
        new_pool->pool_lobd = obd;
        atomic_set(&new_pool->pool_refcount, 1);
        rc = tgt_pool_init(&new_pool->pool_obds, 0);
        if (rc)
                GOTO(out_err, rc);

        lu_qos_rr_init(&new_pool->pool_rr);

        rc = tgt_pool_init(&new_pool->pool_rr.lqr_pool, 0);
        if (rc)
                GOTO(out_free_pool_obds, rc);

#ifdef CONFIG_PROC_FS
        pool_getref(new_pool);
        new_pool->pool_proc_entry = lprocfs_add_simple(lod->lod_pool_proc_entry,
                                                       poolname, new_pool,
                                                       &pool_proc_operations);
        if (IS_ERR(new_pool->pool_proc_entry)) {
                CDEBUG(D_CONFIG, "%s: cannot add proc entry "LOV_POOLNAMEF"\n",
                       obd->obd_name, poolname);
                new_pool->pool_proc_entry = NULL;
                lod_pool_putref(new_pool);
        }
        CDEBUG(D_INFO, "pool %p - proc %p\n", new_pool,
               new_pool->pool_proc_entry);
#endif

        spin_lock(&obd->obd_dev_lock);
        list_add_tail(&new_pool->pool_list, &lod->lod_pool_list);
        lod->lod_pool_count++;
        spin_unlock(&obd->obd_dev_lock);

        /* Add to hash table only when it is fully ready. */
        rc = rhashtable_lookup_insert_fast(&lod->lod_pools_hash_body,
                                           &new_pool->pool_hash,
                                           pools_hash_params);
        if (rc) {
                if (rc != -EEXIST)
                        /*
                         * Hide -E2BIG and -EBUSY which
                         * are not helpful.
                         */
                        rc = -ENOMEM;
                GOTO(out_err, rc);
        }

        CDEBUG(D_CONFIG, LOV_POOLNAMEF" is pool #%d\n",
                        poolname, lod->lod_pool_count);

        RETURN(0);

out_err:
        spin_lock(&obd->obd_dev_lock);
        list_del_init(&new_pool->pool_list);
        lod->lod_pool_count--;
        spin_unlock(&obd->obd_dev_lock);

        lprocfs_remove(&new_pool->pool_proc_entry);

        tgt_pool_free(&new_pool->pool_rr.lqr_pool);
out_free_pool_obds:
        tgt_pool_free(&new_pool->pool_obds);
        OBD_FREE_PTR(new_pool);
        return rc;
}

/**
 * Remove the named pool from the OBD device.
 *
 * \param[in] obd       OBD device on which pool was previously created
 * \param[in] poolname  name of pool to remove from \a obd
 *
 * \retval              0 on successfully removing the pool
 * \retval              negative error numbers for failures
 */
int lod_pool_del(struct obd_device *obd, char *poolname)
{
        struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
        struct pool_desc  *pool;
        ENTRY;

        /* lookup and kill hash reference */
        rcu_read_lock();
        pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
                                 pools_hash_params);
        if (pool && rhashtable_remove_fast(&lod->lod_pools_hash_body,
                                           &pool->pool_hash,
                                           pools_hash_params) != 0)
                pool = NULL;
        rcu_read_unlock();
        if (!pool)
                RETURN(-ENOENT);

        if (pool->pool_proc_entry != NULL) {
                CDEBUG(D_INFO, "proc entry %p\n", pool->pool_proc_entry);
                lprocfs_remove(&pool->pool_proc_entry);
                lod_pool_putref(pool);
        }

        spin_lock(&obd->obd_dev_lock);
        list_del_init(&pool->pool_list);
        lod->lod_pool_count--;
        spin_unlock(&obd->obd_dev_lock);

        /* release last reference */
        lod_pool_putref(pool);

        RETURN(0);
}

/**
 * Add a single target device to the named pool.
 *
 * Add the target specified by \a ostname to the specified \a poolname.
 *
 * \param[in] obd       OBD device on which to add the pool
 * \param[in] poolname  name of the pool to which to add the target \a ostname
 * \param[in] ostname   name of the target device to be added
 *
 * \retval              0 if \a ostname was (previously) added to the named pool
 * \retval              negative error number on failure
 */
int lod_pool_add(struct obd_device *obd, char *poolname, char *ostname)
{
        struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
        struct obd_uuid ost_uuid;
        struct pool_desc *pool;
        struct lu_tgt_desc *tgt;
        int rc = -EINVAL;
        ENTRY;

        rcu_read_lock();
        pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
                                 pools_hash_params);
        if (pool && !atomic_inc_not_zero(&pool->pool_refcount))
                pool = NULL;
        rcu_read_unlock();
        if (!pool)
                RETURN(-ENOENT);

        obd_str2uuid(&ost_uuid, ostname);

        /* search ost in lod array */
        lod_getref(&lod->lod_ost_descs);
        lod_foreach_ost(lod, tgt) {
                if (obd_uuid_equals(&ost_uuid, &tgt->ltd_uuid)) {
                        rc = 0;
                        break;
                }
        }

        if (rc)
                GOTO(out, rc);

        rc = tgt_pool_add(&pool->pool_obds, tgt->ltd_index,
                              lod->lod_ost_count);
        if (rc)
                GOTO(out, rc);

        set_bit(LQ_DIRTY, &pool->pool_rr.lqr_flags);

        CDEBUG(D_CONFIG, "Added %s to "LOV_POOLNAMEF" as member %d\n",
                        ostname, poolname,  pool_tgt_count(pool));

        EXIT;
out:
        lod_putref(lod, &lod->lod_ost_descs);
        lod_pool_putref(pool);
        return rc;
}

/**
 * Remove the named target from the specified pool.
 *
 * Remove one target named \a ostname from \a poolname.  The \a ostname
 * is searched for in the lod_device lod_ost_bitmap array, to ensure the
 * specified name actually exists in the pool.
 *
 * \param[in] obd       OBD device from which to remove \a poolname
 * \param[in] poolname  name of the pool to be changed
 * \param[in] ostname   name of the target to remove from \a poolname
 *
 * \retval              0 on successfully removing \a ostname from the pool
 * \retval              negative number on error (e.g. \a ostname not in pool)
 */
int lod_pool_remove(struct obd_device *obd, char *poolname, char *ostname)
{
        struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
        struct lu_tgt_desc *ost;
        struct obd_uuid ost_uuid;
        struct pool_desc *pool;
        int rc = -EINVAL;
        ENTRY;

        /* lookup and kill hash reference */
        rcu_read_lock();
        pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
                                 pools_hash_params);
        if (pool && !atomic_inc_not_zero(&pool->pool_refcount))
                pool = NULL;
        rcu_read_unlock();
        if (!pool)
                RETURN(-ENOENT);

        obd_str2uuid(&ost_uuid, ostname);

        lod_getref(&lod->lod_ost_descs);
        lod_foreach_ost(lod, ost) {
                if (obd_uuid_equals(&ost_uuid, &ost->ltd_uuid)) {
                        rc = 0;
                        break;
                }
        }

        /* test if ost found in lod array */
        if (rc)
                GOTO(out, rc);

        tgt_pool_remove(&pool->pool_obds, ost->ltd_index);
        set_bit(LQ_DIRTY, &pool->pool_rr.lqr_flags);

        CDEBUG(D_CONFIG, "%s removed from "LOV_POOLNAMEF"\n", ostname,
               poolname);

        EXIT;
out:
        lod_putref(lod, &lod->lod_ost_descs);
        lod_pool_putref(pool);
        return rc;
}

/**
 * Check if the specified target exists in the pool.
 *
 * The caller may not have a reference on \a pool if it got the pool without
 * calling lod_find_pool() (e.g. directly from the lod pool list)
 *
 * \param[in] idx       Target index to check
 * \param[in] pool      Pool in which to check if target is added.
 *
 * \retval              0 successfully found index in \a pool
 * \retval              negative error if device not found in \a pool
 */
int lod_check_index_in_pool(__u32 idx, struct pool_desc *pool)
{
        int rc;

        pool_getref(pool);
        rc = tgt_check_index(idx, &pool->pool_obds);
        lod_pool_putref(pool);
        return rc;
}

/**
 * Find the pool descriptor for the specified pool and return it with a
 * reference to the caller if found.
 *
 * \param[in] lod       LOD on which the pools are configured
 * \param[in] poolname  NUL-terminated name of the pool
 *
 * \retval      pointer to pool descriptor on success
 * \retval      NULL if \a poolname could not be found or poolname is empty
 */
struct pool_desc *lod_find_pool(struct lod_device *lod, char *poolname)
{
        struct pool_desc *pool;

        pool = NULL;
        if (poolname[0] != '\0') {
                rcu_read_lock();
                pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
                                         pools_hash_params);
                if (pool && !atomic_inc_not_zero(&pool->pool_refcount))
                        pool = NULL;
                rcu_read_unlock();
                if (!pool)
                        CDEBUG(D_CONFIG,
                               "%s: request for an unknown pool (" LOV_POOLNAMEF ")\n",
                               lod->lod_child_exp->exp_obd->obd_name, poolname);
                if (pool != NULL && pool_tgt_count(pool) == 0) {
                        CDEBUG(D_CONFIG, "%s: request for an empty pool ("
                               LOV_POOLNAMEF")\n",
                               lod->lod_child_exp->exp_obd->obd_name, poolname);
                        /* pool is ignored, so we remove ref on it */
                        lod_pool_putref(pool);
                        pool = NULL;
                }
        }
        return pool;
}