d/freebsd-skq
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

6676 Race between unique_insert() and unique_remove() causes ZFS fsid change

Browse Source 
illumos/illumos-gate@40510e8eba
40510e8eba

https://www.illumos.org/issues/6676

  The fsid of zfs filesystems might change after reboot or remount. The problem seems to
  be caused by a race between unique_insert() and unique_remove(). The unique_remove()
  is called from dsl_dataset_evict() which is now an asynchronous thread. In a case the
  dsl_dataset_evict() thread is very slow and calls unique_remove() too late we will end
  up with changed fsid on zfs mount.

  This problem is very likely caused by #5056.

  Steps to Reproduce
  Note: I'm able to reproduce this always on a single core (virtual) machine. On multicore
  machines it is not so easy to reproduce.

# uname -a
SunOS openindiana 5.11 illumos-633aa80 i86pc i386 i86pc Solaris
# zfs create rpool/TEST
# FS=$(echo ::fsinfo | mdb -k | grep TEST | awk '{print $1}')
# echo $FS::print vfs_t vfs_fsid | mdb -k
vfs_fsid = {
    vfs_fsid.val = [ 0x54d7028a, 0x70311508 ]
}
# zfs umount rpool/TEST
# zfs mount rpool/TEST
# FS=$(echo ::fsinfo | mdb -k | grep TEST | awk '{print $1}')
# echo $FS::print vfs_t vfs_fsid | mdb -k
vfs_fsid = {
    vfs_fsid.val = [ 0xd9454e49, 0x6b36d08 ]
}
#

  Impact
  The persistent fsid (filesystem id) is essential for proper NFS functionality.
  If the fsid of a filesystem changes on remount (or after reboot) the NFS
  clients might not be able to automatically recover from such event and the
  manual remount of the NFS filesystems on every NFS client might be needed.

Author: Josef 'Jeff' Sipek <josef.sipek@nexenta.com>
Reviewed by: Saso Kiselkov <saso.kiselkov@nexenta.com>
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexenta.com>
Reviewed by: Dan Vatca <dan.vatca@gmail.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Sebastien Roy <sebastien.roy@delphix.com>
Approved by: Robert Mustacchi <rm@joyent.com>
This commit is contained in:
Josh Paetzel 2017-02-25 03:34:22 +00:00
parent 91c038e77a
commit acf1cb602b
9 changed files with 90 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
uts/common/fs/zfs/dbuf.c
View File

@ -54,7 +54,9 @@ static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
#ifndef __lint
extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
dmu_buf_evict_func_t *evict_func, dmu_buf_t **clear_on_evict_dbufp);
dmu_buf_evict_func_t *evict_func_sync,
dmu_buf_evict_func_t *evict_func_async,
dmu_buf_t **clear_on_evict_dbufp);
#endif /* ! __lint */
/*
@ -361,11 +363,24 @@ dbuf_evict_user(dmu_buf_impl_t *db)
#endif
/*
* Invoke the callback from a taskq to avoid lock order reversals
* and limit stack depth.
* There are two eviction callbacks - one that we call synchronously
* and one that we invoke via a taskq. The async one is useful for
* avoiding lock order reversals and limiting stack depth.
*
* Note that if we have a sync callback but no async callback,
* it's likely that the sync callback will free the structure
* containing the dbu. In that case we need to take care to not
* dereference dbu after calling the sync evict func.
*/
taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0,
&dbu->dbu_tqent);
boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
if (dbu->dbu_evict_func_sync != NULL)
dbu->dbu_evict_func_sync(dbu);
if (has_async) {
taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
dbu, 0, &dbu->dbu_tqent);
}
}
boolean_t

6
uts/common/fs/zfs/dnode.c
View File

@ -1009,7 +1009,7 @@ dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
}
static void
dnode_buf_pageout(void *dbu)
dnode_buf_evict_async(void *dbu)
{
dnode_children_t *children_dnodes = dbu;
int i;
@ -1135,8 +1135,8 @@ dnode_hold_impl(objset_t *os, uint64_t object, int flag,
for (i = 0; i < epb; i++) {
zrl_init(&dnh[i].dnh_zrlock);
}
dmu_buf_init_user(&children_dnodes->dnc_dbu,
dnode_buf_pageout, NULL);
dmu_buf_init_user(&children_dnodes->dnc_dbu, NULL,
dnode_buf_evict_async, NULL);
winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
if (winner != NULL) {

35
uts/common/fs/zfs/dsl_dataset.c
View File

@ -20,7 +20,7 @@
*/
/*
* Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
* Copyright (c) 2011, 2015 by Delphix. All rights reserved.
* Copyright (c) 2011, 2016 by Delphix. All rights reserved.
* Copyright (c) 2014, Joyent, Inc. All rights reserved.
* Copyright (c) 2014 RackTop Systems.
* Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
@ -272,8 +272,24 @@ dsl_dataset_block_freeable(dsl_dataset_t *ds, const blkptr_t *bp,
return (B_TRUE);
}
/*
* We have to release the fsid syncronously or we risk that a subsequent
* mount of the same dataset will fail to unique_insert the fsid. This
* failure would manifest itself as the fsid of this dataset changing
* between mounts which makes NFS clients quite unhappy.
*/
static void
dsl_dataset_evict(void *dbu)
dsl_dataset_evict_sync(void *dbu)
{
dsl_dataset_t *ds = dbu;
ASSERT(ds->ds_owner == NULL);
unique_remove(ds->ds_fsid_guid);
}
static void
dsl_dataset_evict_async(void *dbu)
{
dsl_dataset_t *ds = dbu;
@ -281,8 +297,6 @@ dsl_dataset_evict(void *dbu)
ds->ds_dbuf = NULL;
unique_remove(ds->ds_fsid_guid);
if (ds->ds_objset != NULL)
dmu_objset_evict(ds->ds_objset);
@ -517,7 +531,8 @@ dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
ds->ds_reserved = ds->ds_quota = 0;
}
dmu_buf_init_user(&ds->ds_dbu, dsl_dataset_evict, &ds->ds_dbuf);
dmu_buf_init_user(&ds->ds_dbu, dsl_dataset_evict_sync,
dsl_dataset_evict_async, &ds->ds_dbuf);
if (err == 0)
winner = dmu_buf_set_user_ie(dbuf, &ds->ds_dbu);
@ -540,6 +555,16 @@ dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
} else {
ds->ds_fsid_guid =
unique_insert(dsl_dataset_phys(ds)->ds_fsid_guid);
if (ds->ds_fsid_guid !=
dsl_dataset_phys(ds)->ds_fsid_guid) {
zfs_dbgmsg("ds_fsid_guid changed from "
"%llx to %llx for pool %s dataset id %llu",
(long long)
dsl_dataset_phys(ds)->ds_fsid_guid,
(long long)ds->ds_fsid_guid,
spa_name(dp->dp_spa),
dsobj);
}
}
}
ASSERT3P(ds->ds_dbuf, ==, dbuf);

7
uts/common/fs/zfs/dsl_dir.c
View File

@ -20,7 +20,7 @@
*/
/*
* Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
* Copyright (c) 2012, 2014 by Delphix. All rights reserved.
* Copyright (c) 2012, 2016 by Delphix. All rights reserved.
* Copyright (c) 2013 Martin Matuska. All rights reserved.
* Copyright (c) 2014 Joyent, Inc. All rights reserved.
* Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
@ -128,7 +128,7 @@ extern inline dsl_dir_phys_t *dsl_dir_phys(dsl_dir_t *dd);
static uint64_t dsl_dir_space_towrite(dsl_dir_t *dd);
static void
dsl_dir_evict(void *dbu)
dsl_dir_evict_async(void *dbu)
{
dsl_dir_t *dd = dbu;
dsl_pool_t *dp = dd->dd_pool;
@ -235,7 +235,8 @@ dsl_dir_hold_obj(dsl_pool_t *dp, uint64_t ddobj,
dmu_buf_rele(origin_bonus, FTAG);
}
dmu_buf_init_user(&dd->dd_dbu, dsl_dir_evict, &dd->dd_dbuf);
dmu_buf_init_user(&dd->dd_dbu, NULL, dsl_dir_evict_async,
&dd->dd_dbuf);
winner = dmu_buf_set_user_ie(dbuf, &dd->dd_dbu);
if (winner != NULL) {
if (dd->dd_parent)

10
uts/common/fs/zfs/sa.c
View File

@ -22,7 +22,7 @@
/*
* Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
* Portions Copyright 2011 iXsystems, Inc
* Copyright (c) 2013 by Delphix. All rights reserved.
* Copyright (c) 2013, 2016 by Delphix. All rights reserved.
* Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
* Copyright (c) 2014 Integros [integros.com]
*/
@ -1296,7 +1296,7 @@ sa_build_index(sa_handle_t *hdl, sa_buf_type_t buftype)
/*ARGSUSED*/
static void
sa_evict(void *dbu)
sa_evict_sync(void *dbu)
{
panic("evicting sa dbuf\n");
}
@ -1380,7 +1380,8 @@ sa_handle_get_from_db(objset_t *os, dmu_buf_t *db, void *userp,
sa_handle_t *winner = NULL;
handle = kmem_cache_alloc(sa_cache, KM_SLEEP);
handle->sa_dbu.dbu_evict_func = NULL;
handle->sa_dbu.dbu_evict_func_sync = NULL;
handle->sa_dbu.dbu_evict_func_async = NULL;
handle->sa_userp = userp;
handle->sa_bonus = db;
handle->sa_os = os;
@ -1391,7 +1392,8 @@ sa_handle_get_from_db(objset_t *os, dmu_buf_t *db, void *userp,
error = sa_build_index(handle, SA_BONUS);
if (hdl_type == SA_HDL_SHARED) {
dmu_buf_init_user(&handle->sa_dbu, sa_evict, NULL);
dmu_buf_init_user(&handle->sa_dbu, sa_evict_sync, NULL,
NULL);
winner = dmu_buf_set_user_ie(db, &handle->sa_dbu);
}

26
uts/common/fs/zfs/sys/dmu.h
View File

@ -537,8 +537,14 @@ typedef struct dmu_buf_user {
*/
taskq_ent_t dbu_tqent;
/* This instance's eviction function pointer. */
dmu_buf_evict_func_t *dbu_evict_func;
/*
* This instance's eviction function pointers.
*
* dbu_evict_func_sync is called synchronously and then
* dbu_evict_func_async is executed asynchronously on a taskq.
*/
dmu_buf_evict_func_t *dbu_evict_func_sync;
dmu_buf_evict_func_t *dbu_evict_func_async;
#ifdef ZFS_DEBUG
/*
* Pointer to user's dbuf pointer. NULL for clients that do
@ -562,15 +568,19 @@ typedef struct dmu_buf_user {
/* Very ugly, but it beats issuing suppression directives in many Makefiles. */
extern void
dmu_buf_init_user(dmu_buf_user_t *dbu, dmu_buf_evict_func_t *evict_func,
dmu_buf_t **clear_on_evict_dbufp);
dmu_buf_evict_func_t *evict_func_async, dmu_buf_t **clear_on_evict_dbufp);
#else /* __lint */
inline void
dmu_buf_init_user(dmu_buf_user_t *dbu, dmu_buf_evict_func_t *evict_func,
dmu_buf_t **clear_on_evict_dbufp)
dmu_buf_init_user(dmu_buf_user_t *dbu, dmu_buf_evict_func_t *evict_func_sync,
dmu_buf_evict_func_t *evict_func_async, dmu_buf_t **clear_on_evict_dbufp)
{
ASSERT(dbu->dbu_evict_func == NULL);
ASSERT(evict_func != NULL);
dbu->dbu_evict_func = evict_func;
ASSERT(dbu->dbu_evict_func_sync == NULL);
ASSERT(dbu->dbu_evict_func_async == NULL);
/* must have at least one evict func */
IMPLY(evict_func_sync == NULL, evict_func_async != NULL);
dbu->dbu_evict_func_sync = evict_func_sync;
dbu->dbu_evict_func_async = evict_func_async;
#ifdef ZFS_DEBUG
dbu->dbu_clear_on_evict_dbufp = clear_on_evict_dbufp;
#endif

2
uts/common/fs/zfs/sys/zap_impl.h
View File

@ -199,7 +199,7 @@ boolean_t zap_match(zap_name_t *zn, const char *matchname);
int zap_lockdir(objset_t *os, uint64_t obj, dmu_tx_t *tx,
krw_t lti, boolean_t fatreader, boolean_t adding, void *tag, zap_t **zapp);
void zap_unlockdir(zap_t *zap, void *tag);
void zap_evict(void *dbu);
void zap_evict_sync(void *dbu);
zap_name_t *zap_name_alloc(zap_t *zap, const char *key, matchtype_t mt);
void zap_name_free(zap_name_t *zn);
int zap_hashbits(zap_t *zap);

11
uts/common/fs/zfs/zap.c
View File

@ -81,7 +81,8 @@ fzap_upgrade(zap_t *zap, dmu_tx_t *tx, zap_flags_t flags)
ASSERT(RW_WRITE_HELD(&zap->zap_rwlock));
zap->zap_ismicro = FALSE;
zap->zap_dbu.dbu_evict_func = zap_evict;
zap->zap_dbu.dbu_evict_func_sync = zap_evict_sync;
zap->zap_dbu.dbu_evict_func_async = NULL;
mutex_init(&zap->zap_f.zap_num_entries_mtx, 0, 0, 0);
zap->zap_f.zap_block_shift = highbit64(zap->zap_dbuf->db_size) - 1;
@ -399,7 +400,7 @@ zap_allocate_blocks(zap_t *zap, int nblocks)
}
static void
zap_leaf_pageout(void *dbu)
zap_leaf_evict_sync(void *dbu)
{
zap_leaf_t *l = dbu;
@ -423,7 +424,7 @@ zap_create_leaf(zap_t *zap, dmu_tx_t *tx)
VERIFY(0 == dmu_buf_hold(zap->zap_objset, zap->zap_object,
l->l_blkid << FZAP_BLOCK_SHIFT(zap), NULL, &l->l_dbuf,
DMU_READ_NO_PREFETCH));
dmu_buf_init_user(&l->l_dbu, zap_leaf_pageout, &l->l_dbuf);
dmu_buf_init_user(&l->l_dbu, zap_leaf_evict_sync, NULL, &l->l_dbuf);
winner = dmu_buf_set_user(l->l_dbuf, &l->l_dbu);
ASSERT(winner == NULL);
dmu_buf_will_dirty(l->l_dbuf, tx);
@ -470,13 +471,13 @@ zap_open_leaf(uint64_t blkid, dmu_buf_t *db)
l->l_bs = highbit64(db->db_size) - 1;
l->l_dbuf = db;
dmu_buf_init_user(&l->l_dbu, zap_leaf_pageout, &l->l_dbuf);
dmu_buf_init_user(&l->l_dbu, zap_leaf_evict_sync, NULL, &l->l_dbuf);
winner = dmu_buf_set_user(db, &l->l_dbu);
rw_exit(&l->l_rwlock);
if (winner != NULL) {
/* someone else set it first */
zap_leaf_pageout(&l->l_dbu);
zap_leaf_evict_sync(&l->l_dbu);
l = winner;
}

4
uts/common/fs/zfs/zap_micro.c
View File

@ -397,7 +397,7 @@ mzap_open(objset_t *os, uint64_t obj, dmu_buf_t *db)
* it, because zap_lockdir() checks zap_ismicro without the lock
* held.
*/
dmu_buf_init_user(&zap->zap_dbu, zap_evict, &zap->zap_dbuf);
dmu_buf_init_user(&zap->zap_dbu, zap_evict_sync, NULL, &zap->zap_dbuf);
winner = dmu_buf_set_user(db, &zap->zap_dbu);
if (winner != NULL)
@ -736,7 +736,7 @@ zap_destroy(objset_t *os, uint64_t zapobj, dmu_tx_t *tx)
}
void
zap_evict(void *dbu)
zap_evict_sync(void *dbu)
{
zap_t *zap = dbu;