*/
/*
* This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
*/
/*
* lustre/lod/lod_pool.c
#define DEBUG_SUBSYSTEM S_LOV
#include <libcfs/libcfs.h>
+#include <libcfs/linux/linux-hash.h>
+#include <libcfs/linux/linux-fs.h>
#include <obd.h>
#include "lod_internal.h"
{
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);
+ lu_tgt_pool_free(&(pool->pool_rr.lqr_pool));
+ lu_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)
-{
- CDEBUG(D_INFO, "pool %p\n", pool);
- LASSERT(atomic_read(&pool->pool_refcount) > 1);
-
- atomic_dec(&pool->pool_refcount);
-}
-
-/*
- * 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)
-{
- return cfs_hash_djb2_hash(key, strnlen(key, LOV_MAXPOOLNAME), mask);
-}
-
-/**
- * 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 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)
+static u32 pool_hashfh(const void *data, u32 len, u32 seed)
{
- 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;
+ const char *pool_name = data;
- pool = hlist_entry(hnode, struct pool_desc, pool_hash);
- pool_getref(pool);
+ return hashlen_hash(cfs_hashlen_string((void *)(unsigned long)seed,
+ pool_name));
}
-static void pool_hashrefcount_put_locked(struct cfs_hash *hs,
- struct hlist_node *hnode)
+static int pool_cmpfn(struct rhashtable_compare_arg *arg, const void *obj)
{
- struct pool_desc *pool;
+ const struct pool_desc *pool = obj;
+ const char *pool_name = arg->key;
- pool = hlist_entry(hnode, struct pool_desc, pool_hash);
- pool_putref_locked(pool);
+ return strcmp(pool_name, pool->pool_name);
}
-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,
+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,
};
/*
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))
+ if (*pos > pool_tgt_count(iter->lpi_pool))
return NULL;
OBD_FAIL_TIMEOUT(OBD_FAIL_OST_LIST_ASSERT, cfs_fail_val);
iter->lpi_idx = prev_idx; /* we stay on the last entry */
return NULL;
}
- (*pos)++;
+
/* return != NULL to continue */
return iter;
}
return rc;
}
-static struct file_operations pool_proc_operations = {
- .open = pool_proc_open,
- .read = seq_read,
- .llseek = seq_lseek,
- .release = seq_release,
+const static struct proc_ops pool_proc_operations = {
+ .proc_open = pool_proc_open,
+ .proc_read = seq_read,
+ .proc_lseek = seq_lseek,
+ .proc_release = seq_release,
};
/**
lod_pool_putref(pool);
}
-/**
- * Initialize the pool data structures at startup.
- *
- * Allocate and initialize the pool data structures with the specified
- * array size. If pool count is not specified (\a count == 0), then
- * POOL_INIT_COUNT will be used. Allocating a non-zero initial array
- * size avoids the need to reallocate as new pools are added.
- *
- * \param[in] op pool structure
- * \param[in] count initial size of the target op_array[] array
- *
- * \retval 0 indicates successful pool initialization
- * \retval negative error number on failure
- */
-#define POOL_INIT_COUNT 2
-int lod_ost_pool_init(struct lu_tgt_pool *op, unsigned int count)
+static void pools_hash_exit(void *vpool, void *data)
{
- ENTRY;
-
- if (count == 0)
- count = POOL_INIT_COUNT;
- op->op_array = NULL;
- op->op_count = 0;
- init_rwsem(&op->op_rw_sem);
- op->op_size = count * sizeof(op->op_array[0]);
- OBD_ALLOC(op->op_array, op->op_size);
- if (op->op_array == NULL) {
- op->op_size = 0;
- RETURN(-ENOMEM);
- }
- EXIT;
- return 0;
-}
+ struct pool_desc *pool = vpool;
-/**
- * Increase the op_array size to hold more targets in this pool.
- *
- * The size is increased to at least \a min_count, but may be larger
- * for an existing pool since ->op_array[] is growing exponentially.
- * Caller must hold write op_rwlock.
- *
- * \param[in] op pool structure
- * \param[in] min_count minimum number of entries to handle
- *
- * \retval 0 on success
- * \retval negative error number on failure.
- */
-int lod_ost_pool_extend(struct lu_tgt_pool *op, unsigned int min_count)
-{
- __u32 *new;
- __u32 new_size;
-
- LASSERT(min_count != 0);
-
- if (op->op_count * sizeof(op->op_array[0]) < op->op_size)
- return 0;
-
- new_size = max_t(__u32, min_count * sizeof(op->op_array[0]),
- 2 * op->op_size);
- OBD_ALLOC(new, new_size);
- if (new == NULL)
- return -ENOMEM;
-
- /* copy old array to new one */
- memcpy(new, op->op_array, op->op_size);
- OBD_FREE(op->op_array, op->op_size);
- op->op_array = new;
- op->op_size = new_size;
-
- return 0;
-}
-
-/**
- * Add a new target to an existing pool.
- *
- * Add a new target device to the pool previously created and returned by
- * lod_pool_new(). Each target can only be in each pool at most one time.
- *
- * \param[in] op target pool to add new entry
- * \param[in] idx pool index number to add to the \a op array
- * \param[in] min_count minimum number of entries to expect in the pool
- *
- * \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 lu_tgt_pool *op, __u32 idx, unsigned int min_count)
-{
- unsigned int i;
- int rc = 0;
- ENTRY;
-
- down_write(&op->op_rw_sem);
-
- rc = lod_ost_pool_extend(op, min_count);
- if (rc)
- GOTO(out, rc);
-
- /* search ost in pool array */
- for (i = 0; i < op->op_count; i++) {
- if (op->op_array[i] == idx)
- GOTO(out, rc = -EEXIST);
- }
- /* ost not found we add it */
- op->op_array[op->op_count] = idx;
- op->op_count++;
- EXIT;
-out:
- up_write(&op->op_rw_sem);
- return rc;
+ lod_pool_putref(pool);
}
-/**
- * Remove an existing pool from the system.
- *
- * The specified pool must have previously been allocated by
- * lod_pool_new() and not have any target members in the pool.
- * If the removed target is not the last, compact the array
- * to remove empty spaces.
- *
- * \param[in] op pointer to the original data structure
- * \param[in] idx target index to be removed
- *
- * \retval 0 on success
- * \retval negative error number on failure
- */
-int lod_ost_pool_remove(struct lu_tgt_pool *op, __u32 idx)
+int lod_pool_hash_init(struct rhashtable *tbl)
{
- unsigned int i;
- ENTRY;
-
- down_write(&op->op_rw_sem);
-
- for (i = 0; i < op->op_count; i++) {
- if (op->op_array[i] == idx) {
- memmove(&op->op_array[i], &op->op_array[i + 1],
- (op->op_count - i - 1) *
- sizeof(op->op_array[0]));
- op->op_count--;
- up_write(&op->op_rw_sem);
- EXIT;
- return 0;
- }
- }
-
- up_write(&op->op_rw_sem);
- RETURN(-EINVAL);
+ return rhashtable_init(tbl, &pools_hash_params);
}
-/**
- * Free the pool after it was emptied and removed from /proc.
- *
- * Note that all of the child/target entries referenced by this pool
- * must have been removed by lod_ost_pool_remove() before it can be
- * deleted from memory.
- *
- * \param[in] op pool to be freed.
- *
- * \retval 0 on success or if pool was already freed
- */
-int lod_ost_pool_free(struct lu_tgt_pool *op)
+void lod_pool_hash_destroy(struct rhashtable *tbl)
{
- ENTRY;
-
- if (op->op_size == 0)
- RETURN(0);
-
- down_write(&op->op_rw_sem);
-
- OBD_FREE(op->op_array, op->op_size);
- op->op_array = NULL;
- op->op_count = 0;
- op->op_size = 0;
-
- up_write(&op->op_rw_sem);
- RETURN(0);
+ rhashtable_free_and_destroy(tbl, pools_hash_exit, NULL);
}
/**
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_spill_expire = 0;
+ new_pool->pool_spill_is_active = false;
+ new_pool->pool_spill_threshold_pct = 0;
+ new_pool->pool_spill_target[0] = '\0';
new_pool->pool_lobd = obd;
atomic_set(&new_pool->pool_refcount, 1);
- rc = lod_ost_pool_init(&new_pool->pool_obds, 0);
+ rc = lu_tgt_pool_init(&new_pool->pool_obds, 0);
if (rc)
GOTO(out_err, rc);
lu_qos_rr_init(&new_pool->pool_rr);
- rc = lod_ost_pool_init(&new_pool->pool_rr.lqr_pool, 0);
+ rc = lu_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,
new_pool->pool_proc_entry = NULL;
lod_pool_putref(new_pool);
}
+
+ pool_getref(new_pool);
+ new_pool->pool_spill_proc_entry =
+ lprocfs_register(poolname, lod->lod_spill_proc_entry,
+ lprocfs_lod_spill_vars, new_pool);
+ if (IS_ERR(new_pool->pool_spill_proc_entry)) {
+ rc = PTR_ERR(new_pool->pool_spill_proc_entry);
+ 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
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);
lod->lod_pool_count--;
spin_unlock(&obd->obd_dev_lock);
+ lprocfs_remove(&new_pool->pool_spill_proc_entry);
lprocfs_remove(&new_pool->pool_proc_entry);
- lod_ost_pool_free(&new_pool->pool_rr.lqr_pool);
+ lu_tgt_pool_free(&new_pool->pool_rr.lqr_pool);
out_free_pool_obds:
- lod_ost_pool_free(&new_pool->pool_obds);
+ lu_tgt_pool_free(&new_pool->pool_obds);
OBD_FREE_PTR(new_pool);
return rc;
}
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) {
lprocfs_remove(&pool->pool_proc_entry);
lod_pool_putref(pool);
}
+ if (pool->pool_spill_proc_entry != NULL) {
+ CDEBUG(D_INFO, "proc entry %p\n", pool->pool_spill_proc_entry);
+ lprocfs_remove(&pool->pool_spill_proc_entry);
+ lod_pool_putref(pool);
+ }
spin_lock(&obd->obd_dev_lock);
list_del_init(&pool->pool_list);
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);
if (rc)
GOTO(out, rc);
- rc = lod_ost_pool_add(&pool->pool_obds, tgt->ltd_index,
- lod->lod_ost_descs.ltd_tgts_size);
+ rc = lu_tgt_pool_add(&pool->pool_obds, tgt->ltd_index,
+ lod->lod_ost_count);
if (rc)
GOTO(out, rc);
- pool->pool_rr.lqr_dirty = 1;
+ 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));
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);
if (rc)
GOTO(out, rc);
- lod_ost_pool_remove(&pool->pool_obds, ost->ltd_index);
-
- pool->pool_rr.lqr_dirty = 1;
+ lu_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);
*/
int lod_check_index_in_pool(__u32 idx, struct pool_desc *pool)
{
- unsigned int i;
int rc;
- ENTRY;
pool_getref(pool);
-
- down_read(&pool_tgt_rw_sem(pool));
-
- for (i = 0; i < pool_tgt_count(pool); i++) {
- if (pool_tgt_array(pool)[i] == idx)
- GOTO(out, rc = 0);
- }
- rc = -ENOENT;
- EXIT;
-out:
- up_read(&pool_tgt_rw_sem(pool));
-
+ rc = lu_tgt_check_index(idx, &pool->pool_obds);
lod_pool_putref(pool);
return rc;
}
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 ("
return pool;
}
+void lod_spill_target_refresh(const struct lu_env *env, struct lod_device *lod,
+ struct pool_desc *pool)
+{
+ __u64 avail_bytes = 0, total_bytes = 0;
+ struct lu_tgt_pool *osts;
+ int i;
+
+ if (ktime_get_seconds() < pool->pool_spill_expire)
+ return;
+
+ if (pool->pool_spill_threshold_pct == 0)
+ return;
+
+ lod_qos_statfs_update(env, lod, &lod->lod_ost_descs);
+
+ down_write(&pool_tgt_rw_sem(pool));
+ if (ktime_get_seconds() < pool->pool_spill_expire)
+ goto out_sem;
+ pool->pool_spill_expire = ktime_get_seconds() +
+ lod->lod_ost_descs.ltd_lov_desc.ld_qos_maxage;
+
+ osts = &(pool->pool_obds);
+ for (i = 0; i < osts->op_count; i++) {
+ int idx = osts->op_array[i];
+ struct lod_tgt_desc *tgt;
+ struct obd_statfs *sfs;
+
+ if (!test_bit(idx, lod->lod_ost_bitmap))
+ continue;
+ tgt = OST_TGT(lod, idx);
+ if (tgt->ltd_active == 0)
+ continue;
+ sfs = &tgt->ltd_statfs;
+
+ avail_bytes += sfs->os_bavail * sfs->os_bsize;
+ total_bytes += sfs->os_blocks * sfs->os_bsize;
+ }
+ if (total_bytes - avail_bytes >=
+ total_bytes * pool->pool_spill_threshold_pct / 100)
+ pool->pool_spill_is_active = true;
+ else
+ pool->pool_spill_is_active = false;
+
+out_sem:
+ up_write(&pool_tgt_rw_sem(pool));
+}
+
+/*
+ * to prevent infinite loops during spilling, lets limit number of passes
+ */
+#define LOD_SPILL_MAX 10
+
+/*
+ * XXX: consider a better schema to detect loops
+ */
+void lod_check_and_spill_pool(const struct lu_env *env, struct lod_device *lod,
+ char **poolname)
+{
+ struct pool_desc *pool;
+ int replaced = 0;
+
+ if (!poolname || !*poolname || (*poolname)[0] == '\0')
+ return;
+repeat:
+ 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;
+
+ lod_spill_target_refresh(env, lod, pool);
+ if (pool->pool_spill_is_active) {
+ if (++replaced >= LOD_SPILL_MAX)
+ CWARN("%s: more than %d levels of pool spill for '%s->%s'\n",
+ lod2obd(lod)->obd_name, LOD_SPILL_MAX,
+ *poolname, pool->pool_spill_target);
+ lod_set_pool(poolname, pool->pool_spill_target);
+ lod_pool_putref(pool);
+ if (replaced >= LOD_SPILL_MAX)
+ return;
+ goto repeat;
+ }
+
+ lod_pool_putref(pool);
+}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff