X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Flod%2Flod_pool.c;h=04d81ddf990ed9d0597693b6b47ecc34eaa40879;hb=86fda8611930a6f9ed73cc5e7813c4809a96ab6d;hp=1b15f6cb52d32a2dc6057136d8f435b3acbd1cb3;hpb=c866e2ad4b5fa0e512d1dc7135f37b59f3544205;p=fs%2Flustre-release.git diff --git a/lustre/lod/lod_pool.c b/lustre/lod/lod_pool.c index 1b15f6c..04d81dd 100644 --- a/lustre/lod/lod_pool.c +++ b/lustre/lod/lod_pool.c @@ -23,7 +23,7 @@ * Copyright 2008 Sun Microsystems, Inc. All rights reserved * Use is subject to license terms. * - * Copyright (c) 2012, 2014, Intel Corporation. + * Copyright (c) 2012, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -58,6 +58,7 @@ #define DEBUG_SUBSYSTEM S_LOV #include +#include #include #include "lod_internal.h" @@ -96,130 +97,38 @@ void lod_pool_putref(struct pool_desc *pool) { CDEBUG(D_INFO, "pool %p\n", pool); if (atomic_dec_and_test(&pool->pool_refcount)) { - LASSERT(hlist_unhashed(&pool->pool_hash)); LASSERT(list_empty(&pool->pool_list)); LASSERT(pool->pool_proc_entry == NULL); - lod_ost_pool_free(&(pool->pool_rr.lqr_pool)); - lod_ost_pool_free(&(pool->pool_obds)); - OBD_FREE_PTR(pool); + lod_tgt_pool_free(&(pool->pool_rr.lqr_pool)); + lod_tgt_pool_free(&(pool->pool_obds)); + kfree_rcu(pool, pool_rcu); EXIT; } } -/** - * Drop the refcount in cases where the caller holds a spinlock. - * - * This is needed if the caller cannot be blocked while freeing memory. - * It assumes that there is some other known refcount held on the \a pool - * and the memory cannot actually be freed, but the refcounting needs to - * be kept accurate. - * - * \param[in] pool pool descriptor on which to drop reference - */ -static void pool_putref_locked(struct pool_desc *pool) +static u32 pool_hashfh(const void *data, u32 len, u32 seed) { - CDEBUG(D_INFO, "pool %p\n", pool); - LASSERT(atomic_read(&pool->pool_refcount) > 1); + const char *pool_name = data; - atomic_dec(&pool->pool_refcount); + return hashlen_hash(cfs_hashlen_string((void *)(unsigned long)seed, + pool_name)); } -/* - * Group of functions needed for cfs_hash implementation. This - * includes pool lookup, refcounting, and cleanup. - */ - -/** - * Hash the pool name for use by the cfs_hash handlers. - * - * Use the standard DJB2 hash function for ASCII strings in Lustre. - * - * \param[in] hash_body hash structure where this key is embedded (unused) - * \param[in] key key to be hashed (in this case the pool name) - * \param[in] mask bitmask to limit the hash value to the desired size - * - * \retval computed hash value from \a key and limited by \a mask - */ -static __u32 pool_hashfn(struct cfs_hash *hash_body, const void *key, - unsigned mask) +static int pool_cmpfn(struct rhashtable_compare_arg *arg, const void *obj) { - return cfs_hash_djb2_hash(key, strnlen(key, LOV_MAXPOOLNAME), mask); -} + const struct pool_desc *pool = obj; + const char *pool_name = arg->key; -/** - * Return the actual key (pool name) from the hashed \a hnode. - * - * Allows extracting the key name when iterating over all hash entries. - * - * \param[in] hnode hash node found by lookup or iteration - * - * \retval char array referencing the pool name (no refcount) - */ -static void *pool_key(struct hlist_node *hnode) -{ - struct pool_desc *pool; - - pool = hlist_entry(hnode, struct pool_desc, pool_hash); - return pool->pool_name; -} - -/** - * Check if the specified hash key matches the hash node. - * - * This is needed in case there is a hash key collision, allowing the hash - * table lookup/iteration to distinguish between the two entries. - * - * \param[in] key key (pool name) being searched for - * \param[in] compared current entry being compared - * - * \retval 0 if \a key is the same as the key of \a compared - * \retval 1 if \a key is different from the key of \a compared - */ -static int pool_hashkey_keycmp(const void *key, struct hlist_node *compared) -{ - return !strncmp(key, pool_key(compared), LOV_MAXPOOLNAME); + return strcmp(pool_name, pool->pool_name); } -/** - * Return the actual pool data structure from the hash table entry. - * - * Once the hash table entry is found, extract the pool data from it. - * The return type of this function is void * because it needs to be - * assigned to the generic hash operations table. - * - * \param[in] hnode hash table entry - * - * \retval struct pool_desc for the specified \a hnode - */ -static void *pool_hashobject(struct hlist_node *hnode) -{ - return hlist_entry(hnode, struct pool_desc, pool_hash); -} - -static void pool_hashrefcount_get(struct cfs_hash *hs, struct hlist_node *hnode) -{ - struct pool_desc *pool; - - pool = hlist_entry(hnode, struct pool_desc, pool_hash); - pool_getref(pool); -} - -static void pool_hashrefcount_put_locked(struct cfs_hash *hs, - struct hlist_node *hnode) -{ - struct pool_desc *pool; - - pool = hlist_entry(hnode, struct pool_desc, pool_hash); - pool_putref_locked(pool); -} - -struct cfs_hash_ops pool_hash_operations = { - .hs_hash = pool_hashfn, - .hs_key = pool_key, - .hs_keycmp = pool_hashkey_keycmp, - .hs_object = pool_hashobject, - .hs_get = pool_hashrefcount_get, - .hs_put_locked = pool_hashrefcount_put_locked, +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, }; /* @@ -464,7 +373,7 @@ void lod_dump_pool(int level, struct pool_desc *pool) * \retval negative error number on failure */ #define POOL_INIT_COUNT 2 -int lod_ost_pool_init(struct ost_pool *op, unsigned int count) +int lod_tgt_pool_init(struct lu_tgt_pool *op, unsigned int count) { ENTRY; @@ -496,7 +405,7 @@ int lod_ost_pool_init(struct ost_pool *op, unsigned int count) * \retval 0 on success * \retval negative error number on failure. */ -int lod_ost_pool_extend(struct ost_pool *op, unsigned int min_count) +int lod_tgt_pool_extend(struct lu_tgt_pool *op, unsigned int min_count) { __u32 *new; __u32 new_size; @@ -534,7 +443,7 @@ int lod_ost_pool_extend(struct ost_pool *op, unsigned int min_count) * \retval 0 if target could be added to the pool * \retval negative error if target \a idx was not added */ -int lod_ost_pool_add(struct ost_pool *op, __u32 idx, unsigned int min_count) +int lod_tgt_pool_add(struct lu_tgt_pool *op, __u32 idx, unsigned int min_count) { unsigned int i; int rc = 0; @@ -542,7 +451,7 @@ int lod_ost_pool_add(struct ost_pool *op, __u32 idx, unsigned int min_count) down_write(&op->op_rw_sem); - rc = lod_ost_pool_extend(op, min_count); + rc = lod_tgt_pool_extend(op, min_count); if (rc) GOTO(out, rc); @@ -574,7 +483,7 @@ out: * \retval 0 on success * \retval negative error number on failure */ -int lod_ost_pool_remove(struct ost_pool *op, __u32 idx) +int lod_tgt_pool_remove(struct lu_tgt_pool *op, __u32 idx) { unsigned int i; ENTRY; @@ -608,7 +517,7 @@ int lod_ost_pool_remove(struct ost_pool *op, __u32 idx) * * \retval 0 on success or if pool was already freed */ -int lod_ost_pool_free(struct ost_pool *op) +int lod_tgt_pool_free(struct lu_tgt_pool *op) { ENTRY; @@ -626,6 +535,23 @@ int lod_ost_pool_free(struct ost_pool *op) RETURN(0); } +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. * @@ -650,24 +576,24 @@ int lod_pool_new(struct obd_device *obd, char *poolname) if (strlen(poolname) > LOV_MAXPOOLNAME) RETURN(-ENAMETOOLONG); - OBD_ALLOC_PTR(new_pool); + /* 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 = lod_ost_pool_init(&new_pool->pool_obds, 0); + rc = lod_tgt_pool_init(&new_pool->pool_obds, 0); if (rc) GOTO(out_err, rc); - lod_qos_rr_init(&new_pool->pool_rr); - rc = lod_ost_pool_init(&new_pool->pool_rr.lqr_pool, 0); + lu_qos_rr_init(&new_pool->pool_rr); + + rc = lod_tgt_pool_init(&new_pool->pool_rr.lqr_pool, 0); if (rc) GOTO(out_free_pool_obds, rc); - INIT_HLIST_NODE(&new_pool->pool_hash); - #ifdef CONFIG_PROC_FS pool_getref(new_pool); new_pool->pool_proc_entry = lprocfs_add_simple(lod->lod_pool_proc_entry, @@ -688,11 +614,19 @@ int lod_pool_new(struct obd_device *obd, char *poolname) lod->lod_pool_count++; spin_unlock(&obd->obd_dev_lock); - /* add to find only when it fully ready */ - rc = cfs_hash_add_unique(lod->lod_pools_hash_body, poolname, - &new_pool->pool_hash); - if (rc) - GOTO(out_err, rc = -EEXIST); + /* 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); @@ -707,9 +641,9 @@ out_err: lprocfs_remove(&new_pool->pool_proc_entry); - lod_ost_pool_free(&new_pool->pool_rr.lqr_pool); + lod_tgt_pool_free(&new_pool->pool_rr.lqr_pool); out_free_pool_obds: - lod_ost_pool_free(&new_pool->pool_obds); + lod_tgt_pool_free(&new_pool->pool_obds); OBD_FREE_PTR(new_pool); return rc; } @@ -730,8 +664,15 @@ int lod_pool_del(struct obd_device *obd, char *poolname) ENTRY; /* lookup and kill hash reference */ - pool = cfs_hash_del_key(lod->lod_pools_hash_body, poolname); - if (pool == NULL) + 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) { @@ -765,23 +706,28 @@ int lod_pool_del(struct obd_device *obd, char *poolname) */ 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; - unsigned int idx; - int rc = -EINVAL; + 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; - pool = cfs_hash_lookup(lod->lod_pools_hash_body, poolname); - if (pool == NULL) + 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, idx) { - if (obd_uuid_equals(&ost_uuid, &OST_TGT(lod, idx)->ltd_uuid)) { + lod_foreach_ost(lod, tgt) { + if (obd_uuid_equals(&ost_uuid, &tgt->ltd_uuid)) { rc = 0; break; } @@ -790,7 +736,8 @@ int lod_pool_add(struct obd_device *obd, char *poolname, char *ostname) if (rc) GOTO(out, rc); - rc = lod_ost_pool_add(&pool->pool_obds, idx, lod->lod_osts_size); + rc = lod_tgt_pool_add(&pool->pool_obds, tgt->ltd_index, + lod->lod_ost_count); if (rc) GOTO(out, rc); @@ -822,22 +769,28 @@ out: */ int lod_pool_remove(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; - unsigned int idx; - int rc = -EINVAL; + 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; - pool = cfs_hash_lookup(lod->lod_pools_hash_body, poolname); - if (pool == NULL) + /* 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); - cfs_foreach_bit(lod->lod_ost_bitmap, idx) { - if (obd_uuid_equals(&ost_uuid, &OST_TGT(lod, idx)->ltd_uuid)) { + lod_foreach_ost(lod, ost) { + if (obd_uuid_equals(&ost_uuid, &ost->ltd_uuid)) { rc = 0; break; } @@ -847,8 +800,7 @@ int lod_pool_remove(struct obd_device *obd, char *poolname, char *ostname) if (rc) GOTO(out, rc); - lod_ost_pool_remove(&pool->pool_obds, idx); - + lod_tgt_pool_remove(&pool->pool_obds, ost->ltd_index); pool->pool_rr.lqr_dirty = 1; CDEBUG(D_CONFIG, "%s removed from "LOV_POOLNAMEF"\n", ostname, @@ -912,10 +864,15 @@ struct pool_desc *lod_find_pool(struct lod_device *lod, char *poolname) pool = NULL; if (poolname[0] != '\0') { - pool = cfs_hash_lookup(lod->lod_pools_hash_body, poolname); - if (pool == NULL) - CDEBUG(D_CONFIG, "%s: request for an unknown pool (" - LOV_POOLNAMEF")\n", + 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 ("