[Ocfs2-commits] mfasheh commits r2644 - trunk/fs/ocfs2
svn-commits at oss.oracle.com
svn-commits at oss.oracle.com
Tue Oct 11 15:41:57 CDT 2005
Author: mfasheh
Signed-off-by: jlbec
Signed-off-by: zab
Date: 2005-10-11 15:41:55 -0500 (Tue, 11 Oct 2005)
New Revision: 2644
Modified:
trunk/fs/ocfs2/alloc.c
trunk/fs/ocfs2/dlmglue.c
trunk/fs/ocfs2/dlmglue.h
trunk/fs/ocfs2/ocfs2.h
Log:
* Fix a pair of LVB issues:
- i_size_off was prone to overflow. just stuff a full blown i_size
instead.
- lvb sequence numbers could wrap back to a previous value after a node
death (as it will be set to zero then), leading to a bad value for
trustable_clusters. Don't use the sequence numbers anymore and dump
trustable_clusters.
This changes the lvb contents in a backwards compatible way (old versions
won't trust the new lvb and new versions won't trust the old lvb), so
rolling upgrades should be possible.
Signed-off-by: jlbec
Signed-off-by: zab
Modified: trunk/fs/ocfs2/alloc.c
===================================================================
--- trunk/fs/ocfs2/alloc.c 2005-10-07 22:06:05 UTC (rev 2643)
+++ trunk/fs/ocfs2/alloc.c 2005-10-11 20:41:55 UTC (rev 2644)
@@ -1777,9 +1777,6 @@
target_i_clusters = ocfs2_clusters_for_bytes(osb->sb,
i_size_read(inode));
- /* the extent map gets truncated in ocfs2_do_truncate */
- ocfs2_meta_lvb_set_trunc_clusters(inode, target_i_clusters);
-
last_eb_bh = tc->tc_last_eb_bh;
tc->tc_last_eb_bh = NULL;
Modified: trunk/fs/ocfs2/dlmglue.c
===================================================================
--- trunk/fs/ocfs2/dlmglue.c 2005-10-07 22:06:05 UTC (rev 2643)
+++ trunk/fs/ocfs2/dlmglue.c 2005-10-11 20:41:55 UTC (rev 2644)
@@ -181,14 +181,6 @@
} while (0)
static void ocfs2_vote_on_unlock(ocfs2_super *osb,
struct ocfs2_lock_res *lockres);
-/* Called after we refresh our inode, only has any effect if we have
- * an EX lock. This populates the LVB with the initial values for our
- * change set. */
-static void ocfs2_reset_meta_lvb_values(struct inode *inode);
-static void __ocfs2_stuff_meta_lvb(struct inode *inode);
-static void ocfs2_refresh_inode_from_lvb(struct inode *inode);
-static void __ocfs2_lvb_on_downconvert(struct ocfs2_lock_res *lockres,
- int new_level);
static int ocfs2_meta_lock_update(struct inode *inode,
struct buffer_head **bh);
static void ocfs2_drop_osb_locks(ocfs2_super *osb);
@@ -203,67 +195,6 @@
struct ocfs2_lock_res *lockres,
int new_level);
-static inline int ocfs2_lvb_is_trustable(struct ocfs2_lock_res *lockres)
-{
- struct ocfs2_lvb *lvb = (struct ocfs2_lvb *) lockres->l_lksb.lvb;
- unsigned int lvb_seq;
- int ret = 0;
-
- mlog_entry_void();
-
- spin_lock(&lockres->l_lock);
-
- lvb_seq = be32_to_cpu(lvb->lvb_seq);
- if (lvb_seq &&
- lockres->l_local_seq == lvb_seq)
- ret = 1;
-
- spin_unlock(&lockres->l_lock);
-
- mlog_exit(ret);
- return ret;
-}
-
-static inline void ocfs2_set_local_seq_from_lvb(struct ocfs2_lock_res *lockres)
-{
- struct ocfs2_lvb *lvb = (struct ocfs2_lvb *) lockres->l_lksb.lvb;
- unsigned int lvb_seq;
-
- mlog_entry_void();
-
- spin_lock(&lockres->l_lock);
-
- lvb_seq = be32_to_cpu(lvb->lvb_seq);
- if (lvb_seq)
- lockres->l_local_seq = lvb_seq;
-
- spin_unlock(&lockres->l_lock);
-
- mlog_exit_void();
-}
-
-/* fill in new values as we add them to the lvb. */
-static inline void ocfs2_meta_lvb_get_trunc_clusters(struct ocfs2_lock_res *lockres,
- unsigned int *trunc_clusters)
-{
- struct ocfs2_meta_lvb *lvb;
-
- mlog_entry_void();
-
- BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_META);
-
- spin_lock(&lockres->l_lock);
- BUG_ON(lockres->l_level <= LKM_NLMODE);
-
- lvb = (struct ocfs2_meta_lvb *) lockres->l_lksb.lvb;
- if (trunc_clusters)
- *trunc_clusters = be32_to_cpu(lvb->lvb_trunc_clusters);
-
- spin_unlock(&lockres->l_lock);
-
- mlog_exit_void();
-}
-
static void ocfs2_build_lock_name(enum ocfs2_lock_type type,
u64 blkno,
u32 generation,
@@ -380,7 +311,6 @@
"Lockres %s has %u ex holders\n",
res->l_name, res->l_ex_holders);
- res->l_local_seq = 0;
/* Need to clear out the lock status block for the dlm */
memset(&res->l_lksb, 0, sizeof(res->l_lksb));
@@ -1267,12 +1197,9 @@
/* Call this with the lockres locked. I am reasonably sure we don't
* need ip_lock in this function as anyone who would be changing those
- * values is supposed to be blocked in ocfs2_meta_lock right now.
- * This doesn't stuff trunc_clusters as that should already be up to
- * date from the truncate code. */
+ * values is supposed to be blocked in ocfs2_meta_lock right now. */
static void __ocfs2_stuff_meta_lvb(struct inode *inode)
{
- unsigned int isize_off;
struct ocfs2_inode_info *oi = OCFS2_I(inode);
struct ocfs2_lock_res *lockres = &oi->ip_meta_lockres;
struct ocfs2_meta_lvb *lvb;
@@ -1281,13 +1208,12 @@
lvb = (struct ocfs2_meta_lvb *) lockres->l_lksb.lvb;
- /* take advantage of the fact that we don't support holes --
- * we only need the byte offset from i_clusters which nicely
- * cuts our lvb storage requirement to a u32 */
- isize_off = ocfs2_clusters_to_bytes(inode->i_sb, oi->ip_clusters)
- - i_size_read(inode);
+ /* Setting this to zero will ensure that old versions of the
+ * LVB code don't trust our information. */
+ lvb->lvb_old_seq = cpu_to_be32(0);
+ lvb->lvb_version = cpu_to_be32(OCFS2_LVB_VERSION);
- lvb->lvb_isize_off = cpu_to_be32(isize_off);
+ lvb->lvb_isize = cpu_to_be64(i_size_read(inode));
lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters);
lvb->lvb_iuid = cpu_to_be32(inode->i_uid);
lvb->lvb_igid = cpu_to_be32(inode->i_gid);
@@ -1314,7 +1240,6 @@
static void ocfs2_refresh_inode_from_lvb(struct inode *inode)
{
- loff_t isize;
struct ocfs2_inode_info *oi = OCFS2_I(inode);
struct ocfs2_lock_res *lockres = &oi->ip_meta_lockres;
struct ocfs2_meta_lvb *lvb;
@@ -1328,11 +1253,8 @@
/* We're safe here without the lockres lock... */
spin_lock(&oi->ip_lock);
oi->ip_clusters = be32_to_cpu(lvb->lvb_iclusters);
+ i_size_write(inode, be64_to_cpu(lvb->lvb_isize));
- isize = ocfs2_clusters_to_bytes(inode->i_sb, oi->ip_clusters)
- - be32_to_cpu(lvb->lvb_isize_off);
- i_size_write(inode, isize);
-
/* fast-symlinks are a special case */
if (S_ISLNK(inode->i_mode) && !oi->ip_clusters)
inode->i_blocks = 0;
@@ -1355,53 +1277,20 @@
mlog_exit_void();
}
-static void ocfs2_reset_meta_lvb_values(struct inode *inode)
+static inline int ocfs2_meta_lvb_is_trustable(struct ocfs2_lock_res *lockres)
{
- struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_meta_lockres;
- struct ocfs2_meta_lvb *lvb;
- u32 i_clusters;
+ struct ocfs2_meta_lvb *lvb = (struct ocfs2_meta_lvb *) lockres->l_lksb.lvb;
- mlog_entry_void();
-
- lvb = (struct ocfs2_meta_lvb *) lockres->l_lksb.lvb;
-
- spin_lock(&OCFS2_I(inode)->ip_lock);
- i_clusters = OCFS2_I(inode)->ip_clusters;
- spin_unlock(&OCFS2_I(inode)->ip_lock);
-
- spin_lock(&lockres->l_lock);
- if (lockres->l_level == LKM_EXMODE)
- lvb->lvb_trunc_clusters = cpu_to_be32(i_clusters);
- spin_unlock(&lockres->l_lock);
-
- mlog_exit_void();
+ /* Old OCFS2 versions stored a "sequence" in the lvb to
+ * determine whether the information could be trusted. We
+ * don't want to use an lvb populated from a node running the
+ * old code, so check that sequence is not set. */
+ if (!lvb->lvb_old_seq &&
+ be32_to_cpu(lvb->lvb_version) == OCFS2_LVB_VERSION)
+ return 1;
+ return 0;
}
-static void __ocfs2_lvb_on_downconvert(struct ocfs2_lock_res *lockres,
- int new_level)
-{
- struct ocfs2_lvb *lvb = (struct ocfs2_lvb *) lockres->l_lksb.lvb;
- u32 lvb_seq = be32_to_cpu(lvb->lvb_seq);
-
- mlog_entry_void();
-
- if (lockres->l_level == LKM_EXMODE) {
- lvb_seq++;
- /* Overflow? */
- if (!lvb_seq)
- lvb_seq = 1;
- lockres->l_local_seq = lvb_seq;
- if (new_level == LKM_NLMODE)
- lockres->l_local_seq++;
- lvb->lvb_seq = cpu_to_be32(lvb_seq);
- } else if (lockres->l_level == LKM_PRMODE) {
- if (lvb_seq)
- lockres->l_local_seq++;
- }
-
- mlog_exit_void();
-}
-
/* Determine whether a lock resource needs to be refreshed, and
* arbitrate who gets to refresh it.
*
@@ -1462,7 +1351,6 @@
struct buffer_head **bh)
{
int status = 0;
- u32 trustable_clusters = 0;
struct ocfs2_inode_info *oi = OCFS2_I(inode);
struct ocfs2_lock_res *lockres;
ocfs2_dinode *fe;
@@ -1489,10 +1377,13 @@
* for the inode metadata. */
ocfs2_metadata_cache_purge(inode);
- if (ocfs2_lvb_is_trustable(lockres)) {
- /* yay, fastpath! */
- ocfs2_meta_lvb_get_trunc_clusters(lockres,
- &trustable_clusters);
+ /* will do nothing for inode types that don't use the extent
+ * map (directories, bitmap files, etc) */
+ ocfs2_extent_map_trunc(inode, 0);
+
+ if (ocfs2_meta_lvb_is_trustable(lockres)) {
+ mlog(0, "Trusting LVB on inode %"MLFu64"\n",
+ oi->ip_blkno);
ocfs2_refresh_inode_from_lvb(inode);
} else {
/* Boo, we have to go to disk. */
@@ -1534,16 +1425,6 @@
ocfs2_refresh_inode(inode, fe);
}
- mlog(0, "inode %"MLFu64", I can only trust %u clusters\n",
- OCFS2_I(inode)->ip_blkno, trustable_clusters);
-
- /* will do nothing for inode types that don't use the extent
- * map (directories, bitmap files, etc) */
- ocfs2_extent_map_trunc(inode, trustable_clusters);
-
- ocfs2_set_local_seq_from_lvb(lockres);
- ocfs2_reset_meta_lvb_values(inode);
-
#ifdef OCFS2_DELETE_INODE_WORKAROUND
/* We might as well check this here - since the inode is now
* locked, an up to date view will indicate whether this was
@@ -2072,12 +1953,10 @@
/* the metadata lock requires a bit more work as we have an
* LVB to worry about. */
- if (lockres->l_flags & OCFS2_LOCK_ATTACHED) {
- if (lockres->l_level == LKM_EXMODE)
- __ocfs2_stuff_meta_lvb(inode);
- /* fake an NLMODE downconvert for the lvb code. */
- __ocfs2_lvb_on_downconvert(lockres, LKM_NLMODE);
- }
+ if (lockres->l_flags & OCFS2_LOCK_ATTACHED &&
+ lockres->l_level == LKM_EXMODE &&
+ !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
+ __ocfs2_stuff_meta_lvb(inode);
}
int ocfs2_drop_inode_locks(struct inode *inode)
@@ -2284,14 +2163,13 @@
if (lockres->l_level == LKM_EXMODE)
set_lvb = 1;
- /* If the lock hasn't been refreshed yet (rare), then our
- * memory inode values are old and we skip stuffing
- * the lvb. Additionally we cannot incrememnt the lvb
- * sequence numbers */
+ /* If the lock hasn't been refreshed yet (rare), then
+ * our memory inode values are old and we skip
+ * stuffing the lvb. There's no need to actually clear
+ * out the lvb here as it's value is still valid. */
if (!(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) {
if (set_lvb)
__ocfs2_stuff_meta_lvb(inode);
- __ocfs2_lvb_on_downconvert(lockres, new_level);
} else
mlog(0, "lockres %s: downconverting stale lock!\n",
lockres->l_name);
@@ -2573,21 +2451,6 @@
mlog_exit_void();
}
-void ocfs2_meta_lvb_set_trunc_clusters(struct inode *inode,
- unsigned int trunc_clusters)
-{
- struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_meta_lockres;
- struct ocfs2_meta_lvb *lvb;
-
- spin_lock(&lockres->l_lock);
- BUG_ON(lockres->l_level != LKM_EXMODE);
-
- lvb = (struct ocfs2_meta_lvb *) lockres->l_lksb.lvb;
- if (be32_to_cpu(lvb->lvb_trunc_clusters) > trunc_clusters)
- lvb->lvb_trunc_clusters = cpu_to_be32(trunc_clusters);
- spin_unlock(&lockres->l_lock);
-}
-
/* This aids in debugging situations where a bad LVB might be involved. */
void ocfs2_dump_meta_lvb_info(u64 level,
const char *function,
@@ -2598,12 +2461,11 @@
mlog(level, "LVB information for %s (called from %s:%u):\n",
lockres->l_name, function, line);
- mlog(level, "Lockres local seq: %u, seq: %u, size_off: %u, "
- "clusters: %u\n",
- lockres->l_local_seq, be32_to_cpu(lvb->lvb.lvb_seq),
- be32_to_cpu(lvb->lvb_isize_off), be32_to_cpu(lvb->lvb_iclusters));
- mlog(level, "trunc clusters %u, uid %u, gid %u, mode 0x%x\n",
- be32_to_cpu(lvb->lvb_trunc_clusters), be32_to_cpu(lvb->lvb_iuid),
+ mlog(level, "old_seq: %u, version: %u, clusters: %u\n",
+ be32_to_cpu(lvb->lvb_old_seq), be32_to_cpu(lvb->lvb_version),
+ be32_to_cpu(lvb->lvb_iclusters));
+ mlog(level, "size: %"MLFu64", uid %u, gid %u, mode 0x%x\n",
+ be64_to_cpu(lvb->lvb_isize), be32_to_cpu(lvb->lvb_iuid),
be32_to_cpu(lvb->lvb_igid), be16_to_cpu(lvb->lvb_imode));
mlog(level, "nlink %u, atime_packed 0x%"MLFx64", "
"ctime_packed 0x%"MLFx64", mtime_packed 0x%"MLFx64"\n",
Modified: trunk/fs/ocfs2/dlmglue.h
===================================================================
--- trunk/fs/ocfs2/dlmglue.h 2005-10-07 22:06:05 UTC (rev 2643)
+++ trunk/fs/ocfs2/dlmglue.h 2005-10-11 20:41:55 UTC (rev 2644)
@@ -27,45 +27,11 @@
#ifndef DLMGLUE_H
#define DLMGLUE_H
-/*
- * LVB Sequence number rules:
- * local seq and lvb seq are initialized to zero.
- *
- * Note that the lvb is basically invalid until the 1st EX downconvert
- * as he's the only guy that can set it valid. This is ok though as PR
- * holders would have to do an I/O under lock anyway.
- *
- * NL->PR:
- * NL->EX:
- * If LVB is valid:
- * if local seq == lvb seq, then we are up to date with the contents.
- * otherwise, we take the slow path to get up to date and then set our
- * local seq to the lvb seq.
- *
- * PR->NL:
- * If LVB is valid:
- * We increment our local seq. -- this allows up to
- * one set of changes to the lvb before we considers ourselves
- * invalid.
- *
- * PR->EX:
- * Do nothing.
- *
- * EX->NL:
- * EX->PR:
- * Set the LVB as valid.
- * Populate the LVB contents (this is lock type specific)
- * Increment the LVB seq.
- * Set my local seq to the LVB seq.
- * if (EX->NL)
- * do an additional increment of my local seq.
- */
-struct ocfs2_lvb {
- __be32 lvb_seq;
-};
+#define OCFS2_LVB_VERSION 1
+
struct ocfs2_meta_lvb {
- struct ocfs2_lvb lvb;
- __be32 lvb_trunc_clusters;
+ __be32 lvb_old_seq;
+ __be32 lvb_version;
__be32 lvb_iclusters;
__be32 lvb_iuid;
__be32 lvb_igid;
@@ -74,8 +40,8 @@
__be64 lvb_iatime_packed;
__be64 lvb_ictime_packed;
__be64 lvb_imtime_packed;
- __be32 lvb_isize_off;
- __be32 lvb_reserved[3];
+ __be64 lvb_isize;
+ __be32 lvb_reserved[2];
};
int ocfs2_dlm_init(ocfs2_super *osb);
@@ -119,8 +85,6 @@
void ocfs2_process_blocked_lock(ocfs2_super *osb,
struct ocfs2_lock_res *lockres);
-void ocfs2_meta_lvb_set_trunc_clusters(struct inode *inode,
- unsigned int trunc_clusters);
/* aids in debugging and tracking lvbs */
void ocfs2_dump_meta_lvb_info(u64 level,
const char *function,
Modified: trunk/fs/ocfs2/ocfs2.h
===================================================================
--- trunk/fs/ocfs2/ocfs2.h 2005-10-07 22:06:05 UTC (rev 2643)
+++ trunk/fs/ocfs2/ocfs2.h 2005-10-11 20:41:55 UTC (rev 2644)
@@ -134,7 +134,6 @@
unsigned int l_ro_holders;
unsigned int l_ex_holders;
struct dlm_lockstatus l_lksb;
- u32 l_local_seq;
/* used from AST/BAST funcs. */
enum ocfs2_ast_action l_action;
More information about the Ocfs2-commits
mailing list