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>
62 #include <libcfs/linux/linux-fs.h>
64 #include "lod_internal.h"
66 #define pool_tgt(_p, _i) OST_TGT(lu2lod_dev((_p)->pool_lobd->obd_lu_dev), \
67 (_p)->pool_obds.op_array[_i])
70 * Get a reference on the specified pool.
72 * To ensure the pool descriptor is not freed before the caller is finished
73 * with it. Any process that is accessing \a pool directly needs to hold
74 * reference on it, including /proc since a userspace thread may be holding
75 * the /proc file open and busy in the kernel.
77 * \param[in] pool pool descriptor on which to gain reference
79 static void pool_getref(struct pool_desc *pool)
81 CDEBUG(D_INFO, "pool %p\n", pool);
82 atomic_inc(&pool->pool_refcount);
86 * Drop a reference on the specified pool and free its memory if needed.
88 * One reference is held by the LOD OBD device while it is configured, from
89 * the time the configuration log defines the pool until the time when it is
90 * dropped when the LOD OBD is cleaned up or the pool is deleted. This means
91 * that the pool will not be freed while the LOD device is configured, unless
92 * it is explicitly destroyed by the sysadmin. The pool structure is freed
93 * after the last reference on the structure is released.
95 * \param[in] pool pool descriptor to drop reference on and possibly free
97 void lod_pool_putref(struct pool_desc *pool)
99 CDEBUG(D_INFO, "pool %p\n", pool);
100 if (atomic_dec_and_test(&pool->pool_refcount)) {
101 LASSERT(list_empty(&pool->pool_list));
102 LASSERT(pool->pool_proc_entry == NULL);
103 tgt_pool_free(&(pool->pool_rr.lqr_pool));
104 tgt_pool_free(&(pool->pool_obds));
105 kfree_rcu(pool, pool_rcu);
110 static u32 pool_hashfh(const void *data, u32 len, u32 seed)
112 const char *pool_name = data;
114 return hashlen_hash(cfs_hashlen_string((void *)(unsigned long)seed,
118 static int pool_cmpfn(struct rhashtable_compare_arg *arg, const void *obj)
120 const struct pool_desc *pool = obj;
121 const char *pool_name = arg->key;
123 return strcmp(pool_name, pool->pool_name);
126 static const struct rhashtable_params pools_hash_params = {
127 .key_len = 1, /* actually variable */
128 .key_offset = offsetof(struct pool_desc, pool_name),
129 .head_offset = offsetof(struct pool_desc, pool_hash),
130 .hashfn = pool_hashfh,
131 .obj_cmpfn = pool_cmpfn,
132 .automatic_shrinking = true,
136 * Methods for /proc seq_file iteration of the defined pools.
139 #define POOL_IT_MAGIC 0xB001CEA0
140 struct lod_pool_iterator {
141 unsigned int lpi_magic; /* POOL_IT_MAGIC */
142 unsigned int lpi_idx; /* from 0 to pool_tgt_size - 1 */
143 struct pool_desc *lpi_pool;
147 * Return the next configured target within one pool for seq_file iteration.
149 * Iterator is used to go through the target entries of a single pool
150 * (i.e. the list of OSTs configured for a named pool).
151 * lpi_idx is the current target index in the pool's op_array[].
153 * The return type is a void * because this function is one of the
154 * struct seq_operations methods and must match the function template.
156 * \param[in] seq /proc sequence file iteration tracking structure
157 * \param[in] v unused
158 * \param[in] pos position within iteration; 0 to number of targets - 1
160 * \retval struct pool_iterator of the next pool descriptor
162 static void *pool_proc_next(struct seq_file *seq, void *v, loff_t *pos)
164 struct lod_pool_iterator *iter = seq->private;
167 LASSERTF(iter->lpi_magic == POOL_IT_MAGIC, "%08X\n", iter->lpi_magic);
170 /* test if end of file */
171 if (*pos > pool_tgt_count(iter->lpi_pool))
174 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_LIST_ASSERT, cfs_fail_val);
176 /* iterate to find a non empty entry */
177 prev_idx = iter->lpi_idx;
179 if (iter->lpi_idx >= pool_tgt_count(iter->lpi_pool)) {
180 iter->lpi_idx = prev_idx; /* we stay on the last entry */
184 /* return != NULL to continue */
189 * Start seq_file iteration via /proc for a single pool.
191 * The \a pos parameter may be non-zero, indicating that the iteration
192 * is starting at some offset in the target list. Use the seq_file
193 * private field to memorize the iterator so we can free it at stop().
194 * Need to restore the private pointer to the pool before freeing it.
196 * \param[in] seq new sequence file structure to initialize
197 * \param[in] pos initial target number at which to start iteration
199 * \retval initialized pool iterator private structure
200 * \retval NULL if \a pos exceeds the number of targets in \a pool
201 * \retval negative error number on failure
203 static void *pool_proc_start(struct seq_file *seq, loff_t *pos)
205 struct pool_desc *pool = seq->private;
206 struct lod_pool_iterator *iter;
209 if ((pool_tgt_count(pool) == 0) ||
210 (*pos >= pool_tgt_count(pool))) {
211 /* iter is not created, so stop() has no way to
212 * find pool to dec ref */
213 lod_pool_putref(pool);
219 return ERR_PTR(-ENOMEM);
220 iter->lpi_magic = POOL_IT_MAGIC;
221 iter->lpi_pool = pool;
225 down_read(&pool_tgt_rw_sem(pool));
232 ptr = pool_proc_next(seq, &iter, &i);
233 } while ((i < *pos) && (ptr != NULL));
242 * Finish seq_file iteration for a single pool.
244 * Once iteration has been completed, the pool_iterator struct must be
245 * freed, and the seq_file private pointer restored to the pool, as it
246 * was initially when pool_proc_start() was called.
248 * In some cases the stop() method may be called 2 times, without calling
249 * the start() method (see seq_read() from fs/seq_file.c). We have to free
250 * the private iterator struct only if seq->private points to the iterator.
252 * \param[in] seq sequence file structure to clean up
253 * \param[in] v (unused)
255 static void pool_proc_stop(struct seq_file *seq, void *v)
257 struct lod_pool_iterator *iter = seq->private;
259 if (iter != NULL && iter->lpi_magic == POOL_IT_MAGIC) {
260 up_read(&pool_tgt_rw_sem(iter->lpi_pool));
261 seq->private = iter->lpi_pool;
262 lod_pool_putref(iter->lpi_pool);
268 * Print out one target entry from the pool for seq_file iteration.
270 * The currently referenced pool target is given by op_array[lpi_idx].
272 * \param[in] seq new sequence file structure to initialize
273 * \param[in] v (unused)
275 static int pool_proc_show(struct seq_file *seq, void *v)
277 struct lod_pool_iterator *iter = v;
278 struct lod_tgt_desc *tgt;
280 LASSERTF(iter->lpi_magic == POOL_IT_MAGIC, "%08X\n", iter->lpi_magic);
281 LASSERT(iter->lpi_pool != NULL);
282 LASSERT(iter->lpi_idx <= pool_tgt_count(iter->lpi_pool));
284 tgt = pool_tgt(iter->lpi_pool, iter->lpi_idx);
286 seq_printf(seq, "%s\n", obd_uuid2str(&(tgt->ltd_uuid)));
291 static const struct seq_operations pool_proc_ops = {
292 .start = pool_proc_start,
293 .next = pool_proc_next,
294 .stop = pool_proc_stop,
295 .show = pool_proc_show,
299 * Open a new /proc file for seq_file iteration of targets in one pool.
301 * Initialize the seq_file private pointer to reference the pool.
303 * \param inode inode to store iteration state for /proc
304 * \param file file descriptor to store iteration methods
306 * \retval 0 for success
307 * \retval negative error number on failure
309 static int pool_proc_open(struct inode *inode, struct file *file)
313 rc = seq_open(file, &pool_proc_ops);
315 struct seq_file *seq = file->private_data;
316 seq->private = PDE_DATA(inode);
321 const static struct proc_ops pool_proc_operations = {
322 .open = pool_proc_open,
325 .release = seq_release,
329 * Dump the pool target list into the Lustre debug log.
331 * This is a debugging function to allow dumping the list of targets
332 * in \a pool to the Lustre kernel debug log at the given \a level.
334 * This is not currently called by any existing code, but can be called
335 * from within gdb/crash to display the contents of the pool, or from
336 * code under development.
338 * \param[in] level Lustre debug level (D_INFO, D_WARN, D_ERROR, etc)
339 * \param[in] pool pool descriptor to be dumped
341 void lod_dump_pool(int level, struct pool_desc *pool)
347 CDEBUG(level, "pool "LOV_POOLNAMEF" has %d members\n",
348 pool->pool_name, pool->pool_obds.op_count);
349 down_read(&pool_tgt_rw_sem(pool));
351 for (i = 0; i < pool_tgt_count(pool) ; i++) {
352 if (!pool_tgt(pool, i) || !(pool_tgt(pool, i))->ltd_exp)
354 CDEBUG(level, "pool "LOV_POOLNAMEF"[%d] = %s\n",
356 obd_uuid2str(&((pool_tgt(pool, i))->ltd_uuid)));
359 up_read(&pool_tgt_rw_sem(pool));
360 lod_pool_putref(pool);
363 static void pools_hash_exit(void *vpool, void *data)
365 struct pool_desc *pool = vpool;
367 lod_pool_putref(pool);
370 int lod_pool_hash_init(struct rhashtable *tbl)
372 return rhashtable_init(tbl, &pools_hash_params);
375 void lod_pool_hash_destroy(struct rhashtable *tbl)
377 rhashtable_free_and_destroy(tbl, pools_hash_exit, NULL);
381 * Allocate a new pool for the specified device.
383 * Allocate a new pool_desc structure for the specified \a new_pool
384 * device to create a pool with the given \a poolname. The new pool
385 * structure is created with a single reference, and is freed when the
386 * reference count drops to zero.
388 * \param[in] obd Lustre OBD device on which to add a pool iterator
389 * \param[in] poolname the name of the pool to be created
391 * \retval 0 in case of success
392 * \retval negative error code in case of error
394 int lod_pool_new(struct obd_device *obd, char *poolname)
396 struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
397 struct pool_desc *new_pool;
401 if (strlen(poolname) > LOV_MAXPOOLNAME)
402 RETURN(-ENAMETOOLONG);
404 /* OBD_ALLOC_* doesn't work with direct kfree_rcu use */
405 new_pool = kmalloc(sizeof(*new_pool), GFP_KERNEL);
406 if (new_pool == NULL)
409 strlcpy(new_pool->pool_name, poolname, sizeof(new_pool->pool_name));
410 new_pool->pool_lobd = obd;
411 atomic_set(&new_pool->pool_refcount, 1);
412 rc = tgt_pool_init(&new_pool->pool_obds, 0);
416 lu_qos_rr_init(&new_pool->pool_rr);
418 rc = tgt_pool_init(&new_pool->pool_rr.lqr_pool, 0);
420 GOTO(out_free_pool_obds, rc);
422 #ifdef CONFIG_PROC_FS
423 pool_getref(new_pool);
424 new_pool->pool_proc_entry = lprocfs_add_simple(lod->lod_pool_proc_entry,
426 &pool_proc_operations);
427 if (IS_ERR(new_pool->pool_proc_entry)) {
428 CDEBUG(D_CONFIG, "%s: cannot add proc entry "LOV_POOLNAMEF"\n",
429 obd->obd_name, poolname);
430 new_pool->pool_proc_entry = NULL;
431 lod_pool_putref(new_pool);
433 CDEBUG(D_INFO, "pool %p - proc %p\n", new_pool,
434 new_pool->pool_proc_entry);
437 spin_lock(&obd->obd_dev_lock);
438 list_add_tail(&new_pool->pool_list, &lod->lod_pool_list);
439 lod->lod_pool_count++;
440 spin_unlock(&obd->obd_dev_lock);
442 /* Add to hash table only when it is fully ready. */
443 rc = rhashtable_lookup_insert_fast(&lod->lod_pools_hash_body,
444 &new_pool->pool_hash,
449 * Hide -E2BIG and -EBUSY which
456 CDEBUG(D_CONFIG, LOV_POOLNAMEF" is pool #%d\n",
457 poolname, lod->lod_pool_count);
462 spin_lock(&obd->obd_dev_lock);
463 list_del_init(&new_pool->pool_list);
464 lod->lod_pool_count--;
465 spin_unlock(&obd->obd_dev_lock);
467 lprocfs_remove(&new_pool->pool_proc_entry);
469 tgt_pool_free(&new_pool->pool_rr.lqr_pool);
471 tgt_pool_free(&new_pool->pool_obds);
472 OBD_FREE_PTR(new_pool);
477 * Remove the named pool from the OBD device.
479 * \param[in] obd OBD device on which pool was previously created
480 * \param[in] poolname name of pool to remove from \a obd
482 * \retval 0 on successfully removing the pool
483 * \retval negative error numbers for failures
485 int lod_pool_del(struct obd_device *obd, char *poolname)
487 struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
488 struct pool_desc *pool;
491 /* lookup and kill hash reference */
493 pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
495 if (pool && rhashtable_remove_fast(&lod->lod_pools_hash_body,
497 pools_hash_params) != 0)
503 if (pool->pool_proc_entry != NULL) {
504 CDEBUG(D_INFO, "proc entry %p\n", pool->pool_proc_entry);
505 lprocfs_remove(&pool->pool_proc_entry);
506 lod_pool_putref(pool);
509 spin_lock(&obd->obd_dev_lock);
510 list_del_init(&pool->pool_list);
511 lod->lod_pool_count--;
512 spin_unlock(&obd->obd_dev_lock);
514 /* release last reference */
515 lod_pool_putref(pool);
521 * Add a single target device to the named pool.
523 * Add the target specified by \a ostname to the specified \a poolname.
525 * \param[in] obd OBD device on which to add the pool
526 * \param[in] poolname name of the pool to which to add the target \a ostname
527 * \param[in] ostname name of the target device to be added
529 * \retval 0 if \a ostname was (previously) added to the named pool
530 * \retval negative error number on failure
532 int lod_pool_add(struct obd_device *obd, char *poolname, char *ostname)
534 struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
535 struct obd_uuid ost_uuid;
536 struct pool_desc *pool;
537 struct lu_tgt_desc *tgt;
542 pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
544 if (pool && !atomic_inc_not_zero(&pool->pool_refcount))
550 obd_str2uuid(&ost_uuid, ostname);
552 /* search ost in lod array */
553 lod_getref(&lod->lod_ost_descs);
554 lod_foreach_ost(lod, tgt) {
555 if (obd_uuid_equals(&ost_uuid, &tgt->ltd_uuid)) {
564 rc = tgt_pool_add(&pool->pool_obds, tgt->ltd_index,
569 set_bit(LQ_DIRTY, &pool->pool_rr.lqr_flags);
571 CDEBUG(D_CONFIG, "Added %s to "LOV_POOLNAMEF" as member %d\n",
572 ostname, poolname, pool_tgt_count(pool));
576 lod_putref(lod, &lod->lod_ost_descs);
577 lod_pool_putref(pool);
582 * Remove the named target from the specified pool.
584 * Remove one target named \a ostname from \a poolname. The \a ostname
585 * is searched for in the lod_device lod_ost_bitmap array, to ensure the
586 * specified name actually exists in the pool.
588 * \param[in] obd OBD device from which to remove \a poolname
589 * \param[in] poolname name of the pool to be changed
590 * \param[in] ostname name of the target to remove from \a poolname
592 * \retval 0 on successfully removing \a ostname from the pool
593 * \retval negative number on error (e.g. \a ostname not in pool)
595 int lod_pool_remove(struct obd_device *obd, char *poolname, char *ostname)
597 struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
598 struct lu_tgt_desc *ost;
599 struct obd_uuid ost_uuid;
600 struct pool_desc *pool;
604 /* lookup and kill hash reference */
606 pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
608 if (pool && !atomic_inc_not_zero(&pool->pool_refcount))
614 obd_str2uuid(&ost_uuid, ostname);
616 lod_getref(&lod->lod_ost_descs);
617 lod_foreach_ost(lod, ost) {
618 if (obd_uuid_equals(&ost_uuid, &ost->ltd_uuid)) {
624 /* test if ost found in lod array */
628 tgt_pool_remove(&pool->pool_obds, ost->ltd_index);
629 set_bit(LQ_DIRTY, &pool->pool_rr.lqr_flags);
631 CDEBUG(D_CONFIG, "%s removed from "LOV_POOLNAMEF"\n", ostname,
636 lod_putref(lod, &lod->lod_ost_descs);
637 lod_pool_putref(pool);
642 * Check if the specified target exists in the pool.
644 * The caller may not have a reference on \a pool if it got the pool without
645 * calling lod_find_pool() (e.g. directly from the lod pool list)
647 * \param[in] idx Target index to check
648 * \param[in] pool Pool in which to check if target is added.
650 * \retval 0 successfully found index in \a pool
651 * \retval negative error if device not found in \a pool
653 int lod_check_index_in_pool(__u32 idx, struct pool_desc *pool)
658 rc = tgt_check_index(idx, &pool->pool_obds);
659 lod_pool_putref(pool);
664 * Find the pool descriptor for the specified pool and return it with a
665 * reference to the caller if found.
667 * \param[in] lod LOD on which the pools are configured
668 * \param[in] poolname NUL-terminated name of the pool
670 * \retval pointer to pool descriptor on success
671 * \retval NULL if \a poolname could not be found or poolname is empty
673 struct pool_desc *lod_find_pool(struct lod_device *lod, char *poolname)
675 struct pool_desc *pool;
678 if (poolname[0] != '\0') {
680 pool = rhashtable_lookup(&lod->lod_pools_hash_body, poolname,
682 if (pool && !atomic_inc_not_zero(&pool->pool_refcount))
687 "%s: request for an unknown pool (" LOV_POOLNAMEF ")\n",
688 lod->lod_child_exp->exp_obd->obd_name, poolname);
689 if (pool != NULL && pool_tgt_count(pool) == 0) {
690 CDEBUG(D_CONFIG, "%s: request for an empty pool ("
692 lod->lod_child_exp->exp_obd->obd_name, poolname);
693 /* pool is ignored, so we remove ref on it */
694 lod_pool_putref(pool);