4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 * lustre/lod/lod_pool.c
37 * This file provides code related to the Logical Object Device (LOD)
38 * handling of OST Pools on the MDT. Pools are named lists of targets
39 * that allow userspace to group targets that share a particlar property
40 * together so that users or kernel helpers can make decisions about file
41 * allocation based on these properties. For example, pools could be
42 * defined based on fault domains (e.g. separate racks of server nodes) so
43 * that RAID-1 mirroring could select targets from independent fault
44 * domains, or pools could define target performance characteristics so
45 * that applicatins could select IOP-optimized storage or stream-optimized
46 * storage for a particular output file.
48 * This file handles creation, lookup, and removal of pools themselves, as
49 * well as adding and removing targets to pools. It also handles lprocfs
50 * display of configured pool. The pools are accessed by name in the pool
51 * hash, and are refcounted to ensure proper pool structure lifetimes.
53 * Author: Jacques-Charles LAFOUCRIERE <jc.lafoucriere@cea.fr>
54 * Author: Alex Lyashkov <Alexey.Lyashkov@Sun.COM>
55 * Author: Nathaniel Rutman <Nathan.Rutman@Sun.COM>
58 #define DEBUG_SUBSYSTEM S_LOV
60 #include <libcfs/libcfs.h>
61 #include <libcfs/linux/linux-hash.h>
63 #include "lod_internal.h"
65 #define pool_tgt(_p, _i) OST_TGT(lu2lod_dev((_p)->pool_lobd->obd_lu_dev), \
66 (_p)->pool_obds.op_array[_i])
69 * Get a reference on the specified pool.
71 * To ensure the pool descriptor is not freed before the caller is finished
72 * with it. Any process that is accessing \a pool directly needs to hold
73 * reference on it, including /proc since a userspace thread may be holding
74 * the /proc file open and busy in the kernel.
76 * \param[in] pool pool descriptor on which to gain reference
78 static void pool_getref(struct pool_desc *pool)
80 CDEBUG(D_INFO, "pool %p\n", pool);
81 atomic_inc(&pool->pool_refcount);
85 * Drop a reference on the specified pool and free its memory if needed.
87 * One reference is held by the LOD OBD device while it is configured, from
88 * the time the configuration log defines the pool until the time when it is
89 * dropped when the LOD OBD is cleaned up or the pool is deleted. This means
90 * that the pool will not be freed while the LOD device is configured, unless
91 * it is explicitly destroyed by the sysadmin. The pool structure is freed
92 * after the last reference on the structure is released.
94 * \param[in] pool pool descriptor to drop reference on and possibly free
96 void lod_pool_putref(struct pool_desc *pool)
98 CDEBUG(D_INFO, "pool %p\n", pool);
99 if (atomic_dec_and_test(&pool->pool_refcount)) {
100 LASSERT(list_empty(&pool->pool_list));
101 LASSERT(pool->pool_proc_entry == NULL);
102 tgt_pool_free(&(pool->pool_rr.lqr_pool));
103 tgt_pool_free(&(pool->pool_obds));
104 kfree_rcu(pool, pool_rcu);
109 static u32 pool_hashfh(const void *data, u32 len, u32 seed)
111 const char *pool_name = data;
113 return hashlen_hash(cfs_hashlen_string((void *)(unsigned long)seed,
117 static int pool_cmpfn(struct rhashtable_compare_arg *arg, const void *obj)
119 const struct pool_desc *pool = obj;
120 const char *pool_name = arg->key;
122 return strcmp(pool_name, pool->pool_name);
125 static const struct rhashtable_params pools_hash_params = {
126 .key_len = 1, /* actually variable */
127 .key_offset = offsetof(struct pool_desc, pool_name),
128 .head_offset = offsetof(struct pool_desc, pool_hash),
129 .hashfn = pool_hashfh,
130 .obj_cmpfn = pool_cmpfn,
131 .automatic_shrinking = true,
135 * Methods for /proc seq_file iteration of the defined pools.
138 #define POOL_IT_MAGIC 0xB001CEA0
139 struct lod_pool_iterator {
140 unsigned int lpi_magic; /* POOL_IT_MAGIC */
141 unsigned int lpi_idx; /* from 0 to pool_tgt_size - 1 */
142 struct pool_desc *lpi_pool;
146 * Return the next configured target within one pool for seq_file iteration.
148 * Iterator is used to go through the target entries of a single pool
149 * (i.e. the list of OSTs configured for a named pool).
150 * lpi_idx is the current target index in the pool's op_array[].
152 * The return type is a void * because this function is one of the
153 * struct seq_operations methods and must match the function template.
155 * \param[in] seq /proc sequence file iteration tracking structure
156 * \param[in] v unused
157 * \param[in] pos position within iteration; 0 to number of targets - 1
159 * \retval struct pool_iterator of the next pool descriptor
161 static void *pool_proc_next(struct seq_file *seq, void *v, loff_t *pos)
163 struct lod_pool_iterator *iter = seq->private;
166 LASSERTF(iter->lpi_magic == POOL_IT_MAGIC, "%08X\n", iter->lpi_magic);
168 /* test if end of file */
169 if (*pos >= pool_tgt_count(iter->lpi_pool))
172 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_LIST_ASSERT, cfs_fail_val);
174 /* iterate to find a non empty entry */
175 prev_idx = iter->lpi_idx;
177 if (iter->lpi_idx >= pool_tgt_count(iter->lpi_pool)) {
178 iter->lpi_idx = prev_idx; /* we stay on the last entry */
182 /* return != NULL to continue */
187 * Start seq_file iteration via /proc for a single pool.
189 * The \a pos parameter may be non-zero, indicating that the iteration
190 * is starting at some offset in the target list. Use the seq_file
191 * private field to memorize the iterator so we can free it at stop().
192 * Need to restore the private pointer to the pool before freeing it.
194 * \param[in] seq new sequence file structure to initialize
195 * \param[in] pos initial target number at which to start iteration
197 * \retval initialized pool iterator private structure
198 * \retval NULL if \a pos exceeds the number of targets in \a pool
199 * \retval negative error number on failure
201 static void *pool_proc_start(struct seq_file *seq, loff_t *pos)
203 struct pool_desc *pool = seq->private;
204 struct lod_pool_iterator *iter;
207 if ((pool_tgt_count(pool) == 0) ||
208 (*pos >= pool_tgt_count(pool))) {
209 /* iter is not created, so stop() has no way to
210 * find pool to dec ref */
211 lod_pool_putref(pool);
217 return ERR_PTR(-ENOMEM);
218 iter->lpi_magic = POOL_IT_MAGIC;
219 iter->lpi_pool = pool;
223 down_read(&pool_tgt_rw_sem(pool));
230 ptr = pool_proc_next(seq, &iter, &i);
231 } while ((i < *pos) && (ptr != NULL));
240 * Finish seq_file iteration for a single pool.
242 * Once iteration has been completed, the pool_iterator struct must be
243 * freed, and the seq_file private pointer restored to the pool, as it
244 * was initially when pool_proc_start() was called.
246 * In some cases the stop() method may be called 2 times, without calling
247 * the start() method (see seq_read() from fs/seq_file.c). We have to free
248 * the private iterator struct only if seq->private points to the iterator.
250 * \param[in] seq sequence file structure to clean up
251 * \param[in] v (unused)
253 static void pool_proc_stop(struct seq_file *seq, void *v)
255 struct lod_pool_iterator *iter = seq->private;
257 if (iter != NULL && iter->lpi_magic == POOL_IT_MAGIC) {
258 up_read(&pool_tgt_rw_sem(iter->lpi_pool));
259 seq->private = iter->lpi_pool;
260 lod_pool_putref(iter->lpi_pool);
266 * Print out one target entry from the pool for seq_file iteration.
268 * The currently referenced pool target is given by op_array[lpi_idx].
270 * \param[in] seq new sequence file structure to initialize
271 * \param[in] v (unused)
273 static int pool_proc_show(struct seq_file *seq, void *v)
275 struct lod_pool_iterator *iter = v;
276 struct lod_tgt_desc *tgt;
278 LASSERTF(iter->lpi_magic == POOL_IT_MAGIC, "%08X\n", iter->lpi_magic);
279 LASSERT(iter->lpi_pool != NULL);
280 LASSERT(iter->lpi_idx <= pool_tgt_count(iter->lpi_pool));
282 tgt = pool_tgt(iter->lpi_pool, iter->lpi_idx);
284 seq_printf(seq, "%s\n", obd_uuid2str(&(tgt->ltd_uuid)));
289 static const struct seq_operations pool_proc_ops = {
290 .start = pool_proc_start,
291 .next = pool_proc_next,
292 .stop = pool_proc_stop,
293 .show = pool_proc_show,
297 * Open a new /proc file for seq_file iteration of targets in one pool.
299 * Initialize the seq_file private pointer to reference the pool.
301 * \param inode inode to store iteration state for /proc
302 * \param file file descriptor to store iteration methods
304 * \retval 0 for success
305 * \retval negative error number on failure
307 static int pool_proc_open(struct inode *inode, struct file *file)
311 rc = seq_open(file, &pool_proc_ops);
313 struct seq_file *seq = file->private_data;
314 seq->private = PDE_DATA(inode);
319 static struct file_operations pool_proc_operations = {
320 .open = pool_proc_open,
323 .release = seq_release,
327 * Dump the pool target list into the Lustre debug log.
329 * This is a debugging function to allow dumping the list of targets
330 * in \a pool to the Lustre kernel debug log at the given \a level.
332 * This is not currently called by any existing code, but can be called
333 * from within gdb/crash to display the contents of the pool, or from
334 * code under development.
336 * \param[in] level Lustre debug level (D_INFO, D_WARN, D_ERROR, etc)
337 * \param[in] pool pool descriptor to be dumped
339 void lod_dump_pool(int level, struct pool_desc *pool)
345 CDEBUG(level, "pool "LOV_POOLNAMEF" has %d members\n",
346 pool->pool_name, pool->pool_obds.op_count);
347 down_read(&pool_tgt_rw_sem(pool));
349 for (i = 0; i < pool_tgt_count(pool) ; i++) {
350 if (!pool_tgt(pool, i) || !(pool_tgt(pool, i))->ltd_exp)
352 CDEBUG(level, "pool "LOV_POOLNAMEF"[%d] = %s\n",
354 obd_uuid2str(&((pool_tgt(pool, i))->ltd_uuid)));
357 up_read(&pool_tgt_rw_sem(pool));
358 lod_pool_putref(pool);
361 static void pools_hash_exit(void *vpool, void *data)
363 struct pool_desc *pool = vpool;
365 lod_pool_putref(pool);
368 int lod_pool_hash_init(struct rhashtable *tbl)
370 return rhashtable_init(tbl, &pools_hash_params);
373 void lod_pool_hash_destroy(struct rhashtable *tbl)
375 rhashtable_free_and_destroy(tbl, pools_hash_exit, NULL);
379 * Allocate a new pool for the specified device.
381 * Allocate a new pool_desc structure for the specified \a new_pool
382 * device to create a pool with the given \a poolname. The new pool
383 * structure is created with a single reference, and is freed when the
384 * reference count drops to zero.
386 * \param[in] obd Lustre OBD device on which to add a pool iterator
387 * \param[in] poolname the name of the pool to be created
389 * \retval 0 in case of success
390 * \retval negative error code in case of error
392 int lod_pool_new(struct obd_device *obd, char *poolname)
394 struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
395 struct pool_desc *new_pool;
399 if (strlen(poolname) > LOV_MAXPOOLNAME)
400 RETURN(-ENAMETOOLONG);
402 /* OBD_ALLOC_* doesn't work with direct kfree_rcu use */
403 new_pool = kmalloc(sizeof(*new_pool), GFP_KERNEL);
404 if (new_pool == NULL)
407 strlcpy(new_pool->pool_name, poolname, sizeof(new_pool->pool_name));
408 new_pool->pool_lobd = obd;
409 atomic_set(&new_pool->pool_refcount, 1);
410 rc = tgt_pool_init(&new_pool->pool_obds, 0);
414 lu_qos_rr_init(&new_pool->pool_rr);
416 rc = tgt_pool_init(&new_pool->pool_rr.lqr_pool, 0);
418 GOTO(out_free_pool_obds, rc);
420 #ifdef CONFIG_PROC_FS
421 pool_getref(new_pool);
422 new_pool->pool_proc_entry = lprocfs_add_simple(lod->lod_pool_proc_entry,
424 &pool_proc_operations);
425 if (IS_ERR(new_pool->pool_proc_entry)) {
426 CDEBUG(D_CONFIG, "%s: cannot add proc entry "LOV_POOLNAMEF"\n",
427 obd->obd_name, poolname);
428 new_pool->pool_proc_entry = NULL;
429 lod_pool_putref(new_pool);
431 CDEBUG(D_INFO, "pool %p - proc %p\n", new_pool,
432 new_pool->pool_proc_entry);
435 spin_lock(&obd->obd_dev_lock);
436 list_add_tail(&new_pool->pool_list, &lod->lod_pool_list);
437 lod->lod_pool_count++;
438 spin_unlock(&obd->obd_dev_lock);
440 /* Add to hash table only when it is fully ready. */
441 rc = rhashtable_lookup_insert_fast(&lod->lod_pools_hash_body,
442 &new_pool->pool_hash,
447 * Hide -E2BIG and -EBUSY which
454 CDEBUG(D_CONFIG, LOV_POOLNAMEF" is pool #%d\n",
455 poolname, lod->lod_pool_count);
460 spin_lock(&obd->obd_dev_lock);
461 list_del_init(&new_pool->pool_list);
462 lod->lod_pool_count--;
463 spin_unlock(&obd->obd_dev_lock);
465 lprocfs_remove(&new_pool->pool_proc_entry);
467 tgt_pool_free(&new_pool->pool_rr.lqr_pool);
469 tgt_pool_free(&new_pool->pool_obds);
470 OBD_FREE_PTR(new_pool);
475 * Remove the named pool from the OBD device.
477 * \param[in] obd OBD device on which pool was previously created
478 * \param[in] poolname name of pool to remove from \a obd
480 * \retval 0 on successfully removing the pool
481 * \retval negative error numbers for failures
483 int lod_pool_del(struct obd_device *obd, char *poolname)
485 struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
486 struct pool_desc *pool;
489 /* lookup and kill hash reference */
491 pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
493 if (pool && rhashtable_remove_fast(&lod->lod_pools_hash_body,
495 pools_hash_params) != 0)
501 if (pool->pool_proc_entry != NULL) {
502 CDEBUG(D_INFO, "proc entry %p\n", pool->pool_proc_entry);
503 lprocfs_remove(&pool->pool_proc_entry);
504 lod_pool_putref(pool);
507 spin_lock(&obd->obd_dev_lock);
508 list_del_init(&pool->pool_list);
509 lod->lod_pool_count--;
510 spin_unlock(&obd->obd_dev_lock);
512 /* release last reference */
513 lod_pool_putref(pool);
519 * Add a single target device to the named pool.
521 * Add the target specified by \a ostname to the specified \a poolname.
523 * \param[in] obd OBD device on which to add the pool
524 * \param[in] poolname name of the pool to which to add the target \a ostname
525 * \param[in] ostname name of the target device to be added
527 * \retval 0 if \a ostname was (previously) added to the named pool
528 * \retval negative error number on failure
530 int lod_pool_add(struct obd_device *obd, char *poolname, char *ostname)
532 struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
533 struct obd_uuid ost_uuid;
534 struct pool_desc *pool;
535 struct lu_tgt_desc *tgt;
540 pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
542 if (pool && !atomic_inc_not_zero(&pool->pool_refcount))
548 obd_str2uuid(&ost_uuid, ostname);
550 /* search ost in lod array */
551 lod_getref(&lod->lod_ost_descs);
552 lod_foreach_ost(lod, tgt) {
553 if (obd_uuid_equals(&ost_uuid, &tgt->ltd_uuid)) {
562 rc = tgt_pool_add(&pool->pool_obds, tgt->ltd_index,
567 pool->pool_rr.lqr_dirty = 1;
569 CDEBUG(D_CONFIG, "Added %s to "LOV_POOLNAMEF" as member %d\n",
570 ostname, poolname, pool_tgt_count(pool));
574 lod_putref(lod, &lod->lod_ost_descs);
575 lod_pool_putref(pool);
580 * Remove the named target from the specified pool.
582 * Remove one target named \a ostname from \a poolname. The \a ostname
583 * is searched for in the lod_device lod_ost_bitmap array, to ensure the
584 * specified name actually exists in the pool.
586 * \param[in] obd OBD device from which to remove \a poolname
587 * \param[in] poolname name of the pool to be changed
588 * \param[in] ostname name of the target to remove from \a poolname
590 * \retval 0 on successfully removing \a ostname from the pool
591 * \retval negative number on error (e.g. \a ostname not in pool)
593 int lod_pool_remove(struct obd_device *obd, char *poolname, char *ostname)
595 struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
596 struct lu_tgt_desc *ost;
597 struct obd_uuid ost_uuid;
598 struct pool_desc *pool;
602 /* lookup and kill hash reference */
604 pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
606 if (pool && !atomic_inc_not_zero(&pool->pool_refcount))
612 obd_str2uuid(&ost_uuid, ostname);
614 lod_getref(&lod->lod_ost_descs);
615 lod_foreach_ost(lod, ost) {
616 if (obd_uuid_equals(&ost_uuid, &ost->ltd_uuid)) {
622 /* test if ost found in lod array */
626 tgt_pool_remove(&pool->pool_obds, ost->ltd_index);
627 pool->pool_rr.lqr_dirty = 1;
629 CDEBUG(D_CONFIG, "%s removed from "LOV_POOLNAMEF"\n", ostname,
634 lod_putref(lod, &lod->lod_ost_descs);
635 lod_pool_putref(pool);
640 * Check if the specified target exists in the pool.
642 * The caller may not have a reference on \a pool if it got the pool without
643 * calling lod_find_pool() (e.g. directly from the lod pool list)
645 * \param[in] idx Target index to check
646 * \param[in] pool Pool in which to check if target is added.
648 * \retval 0 successfully found index in \a pool
649 * \retval negative error if device not found in \a pool
651 int lod_check_index_in_pool(__u32 idx, struct pool_desc *pool)
656 rc = tgt_check_index(idx, &pool->pool_obds);
657 lod_pool_putref(pool);
662 * Find the pool descriptor for the specified pool and return it with a
663 * reference to the caller if found.
665 * \param[in] lod LOD on which the pools are configured
666 * \param[in] poolname NUL-terminated name of the pool
668 * \retval pointer to pool descriptor on success
669 * \retval NULL if \a poolname could not be found or poolname is empty
671 struct pool_desc *lod_find_pool(struct lod_device *lod, char *poolname)
673 struct pool_desc *pool;
676 if (poolname[0] != '\0') {
678 pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
680 if (pool && !atomic_inc_not_zero(&pool->pool_refcount))
685 "%s: request for an unknown pool (" LOV_POOLNAMEF ")\n",
686 lod->lod_child_exp->exp_obd->obd_name, poolname);
687 if (pool != NULL && pool_tgt_count(pool) == 0) {
688 CDEBUG(D_CONFIG, "%s: request for an empty pool ("
690 lod->lod_child_exp->exp_obd->obd_name, poolname);
691 /* pool is ignored, so we remove ref on it */
692 lod_pool_putref(pool);