[SCM] Samba Shared Repository - branch master updated
Volker Lendecke
vlendec at samba.org
Tue Feb 16 05:21:41 MST 2010
The branch, master has been updated
via 83542d9... s3: Slightly increase parallelism in g_lock
via be919d6... s3: Avoid starving locks when many processes die at the same time
via 725b365... s3: Avoid a thundering herd in g_lock_unlock
via 07978bd... s3: Optimize g_lock_lock for a heavily contended case
via f3bdb16... s3: Fix handling of processes that died in g_lock
from eda16f2... s4-kcc: remove a qsort() that snuck into the new topology code
http://gitweb.samba.org/?p=samba.git;a=shortlog;h=master
- Log -----------------------------------------------------------------
commit 83542d973ca771353109c7da4b0391d6ba910f53
Author: Volker Lendecke <vl at samba.org>
Date: Tue Feb 16 12:31:58 2010 +0100
s3: Slightly increase parallelism in g_lock
There's no need to still hold the g_lock tdb-level lock while telling the
waiters to retry
commit be919d6faed198cdc29322a4d9491946c0b044b3
Author: Volker Lendecke <vl at samba.org>
Date: Tue Feb 16 12:28:53 2010 +0100
s3: Avoid starving locks when many processes die at the same time
In g_lock_unlock we have a little race between the process_exists and
messaging_send call: We only send to 5 waiters now, they all might have died
between us checking their existence and sending the message. This change makes
g_lock_lock retry at least once every minute.
commit 725b3654f831fbe0388cc09f46269903c9eef1d7
Author: Volker Lendecke <vl at samba.org>
Date: Tue Feb 16 12:22:08 2010 +0100
s3: Avoid a thundering herd in g_lock_unlock
Only notify the first 5 pending lock waiters. This avoids a thundering herd
problem that is really nasty in a cluster. It also makes acquiring a lock a bit
more FIFO, lock waiters are added to the end of the array.
commit 07978bd175395e0dc770f68fff5b8bd8b0fdeb51
Author: Volker Lendecke <vl at samba.org>
Date: Mon Feb 15 16:57:16 2010 +0100
s3: Optimize g_lock_lock for a heavily contended case
Only check the existence of the lock owner in g_lock_parse, check the rest of
the records only when we got the lock successfully. This reduces the load on
process_exists which can involve a network roundtrip in the clustered case.
commit f3bdb163f461175c50b4930fa3464beaee30f4a8
Author: Volker Lendecke <vl at samba.org>
Date: Mon Feb 15 16:49:46 2010 +0100
s3: Fix handling of processes that died in g_lock
g_lock_parse might have thrown away entries from the locks array because the
processes were not around anymore. Don't store the orphaned entries.
-----------------------------------------------------------------------
Summary of changes:
source3/lib/g_lock.c | 82 +++++++++++++++++++++++++++++++++++++++++--------
1 files changed, 68 insertions(+), 14 deletions(-)
Changeset truncated at 500 lines:
diff --git a/source3/lib/g_lock.c b/source3/lib/g_lock.c
index 26b079d..add670c 100644
--- a/source3/lib/g_lock.c
+++ b/source3/lib/g_lock.c
@@ -108,6 +108,34 @@ static bool g_lock_parse(TALLOC_CTX *mem_ctx, TDB_DATA data,
(locks[i].lock_type & G_LOCK_PENDING) ?
"(pending)" : "(owner)"));
+ if (((locks[i].lock_type & G_LOCK_PENDING) == 0)
+ && !process_exists(locks[i].pid)) {
+
+ DEBUGADD(10, ("lock owner %s died -- discarding\n",
+ procid_str(talloc_tos(),
+ &locks[i].pid)));
+
+ if (i < (num_locks-1)) {
+ locks[i] = locks[num_locks-1];
+ }
+ num_locks -= 1;
+ }
+ }
+
+ *plocks = locks;
+ *pnum_locks = num_locks;
+ return true;
+}
+
+static void g_lock_cleanup(int *pnum_locks, struct g_lock_rec *locks)
+{
+ int i, num_locks;
+
+ num_locks = *pnum_locks;
+
+ DEBUG(10, ("g_lock_cleanup: %d locks\n", num_locks));
+
+ for (i=0; i<num_locks; i++) {
if (process_exists(locks[i].pid)) {
continue;
}
@@ -119,19 +147,18 @@ static bool g_lock_parse(TALLOC_CTX *mem_ctx, TDB_DATA data,
}
num_locks -= 1;
}
-
- *plocks = locks;
*pnum_locks = num_locks;
- return true;
+ return;
}
static struct g_lock_rec *g_lock_addrec(TALLOC_CTX *mem_ctx,
struct g_lock_rec *locks,
- int num_locks,
+ int *pnum_locks,
const struct server_id pid,
enum g_lock_type lock_type)
{
struct g_lock_rec *result;
+ int num_locks = *pnum_locks;
result = talloc_realloc(mem_ctx, locks, struct g_lock_rec,
num_locks+1);
@@ -141,6 +168,7 @@ static struct g_lock_rec *g_lock_addrec(TALLOC_CTX *mem_ctx,
result[num_locks].pid = pid;
result[num_locks].lock_type = lock_type;
+ *pnum_locks += 1;
return result;
}
@@ -221,7 +249,7 @@ again:
if (our_index == -1) {
/* First round, add ourself */
- locks = g_lock_addrec(talloc_tos(), locks, num_locks,
+ locks = g_lock_addrec(talloc_tos(), locks, &num_locks,
self, lock_type);
if (locks == NULL) {
DEBUG(10, ("g_lock_addrec failed\n"));
@@ -237,7 +265,14 @@ again:
locks[our_index].lock_type = lock_type;
}
- data = make_tdb_data((uint8_t *)locks, talloc_get_size(locks));
+ if (NT_STATUS_IS_OK(status) && ((lock_type & G_LOCK_PENDING) == 0)) {
+ /*
+ * Walk through the list of locks, search for dead entries
+ */
+ g_lock_cleanup(&num_locks, locks);
+ }
+
+ data = make_tdb_data((uint8_t *)locks, num_locks * sizeof(*locks));
store_status = rec->store(rec, data, 0);
if (!NT_STATUS_IS_OK(store_status)) {
DEBUG(1, ("rec->store failed: %s\n",
@@ -263,7 +298,6 @@ NTSTATUS g_lock_lock(struct g_lock_ctx *ctx, const char *name,
NTSTATUS status;
bool retry = false;
struct timeval timeout_end;
- struct timeval timeout_remaining;
struct timeval time_now;
DEBUG(10, ("Trying to acquire lock %d for %s\n", (int)lock_type,
@@ -304,6 +338,7 @@ NTSTATUS g_lock_lock(struct g_lock_ctx *ctx, const char *name,
fd_set *r_fds = NULL;
int max_fd = 0;
int ret;
+ struct timeval select_timeout;
status = g_lock_trylock(ctx, name, lock_type);
if (NT_STATUS_IS_OK(status)) {
@@ -360,12 +395,10 @@ NTSTATUS g_lock_lock(struct g_lock_ctx *ctx, const char *name,
}
#endif
- time_now = timeval_current();
- timeout_remaining = timeval_until(&time_now, &timeout_end);
+ select_timeout = timeval_set(60, 0);
ret = sys_select(max_fd + 1, r_fds, NULL, NULL,
- &timeout_remaining);
-
+ &select_timeout);
if (ret == -1) {
if (errno != EINTR) {
DEBUG(1, ("error calling select: %s\n",
@@ -386,7 +419,7 @@ NTSTATUS g_lock_lock(struct g_lock_ctx *ctx, const char *name,
break;
} else {
DEBUG(10, ("select returned 0 but timeout not "
- "not expired: strange - retrying\n"));
+ "not expired, retrying\n"));
}
} else if (ret != 1) {
DEBUG(1, ("invalid return code of select: %d\n", ret));
@@ -494,14 +527,26 @@ static NTSTATUS g_lock_force_unlock(struct g_lock_ctx *ctx, const char *name,
goto done;
}
+ TALLOC_FREE(rec);
+
if ((lock_type & G_LOCK_PENDING) == 0) {
+ int num_wakeups = 0;
+
/*
- * We've been the lock holder. Tell all others to retry.
+ * We've been the lock holder. Others to retry. Don't
+ * tell all others to avoid a thundering herd. In case
+ * this leads to a complete stall because we miss some
+ * processes, the loop in g_lock_lock tries at least
+ * once a minute.
*/
+
for (i=0; i<num_locks; i++) {
if ((locks[i].lock_type & G_LOCK_PENDING) == 0) {
continue;
}
+ if (!process_exists(locks[i].pid)) {
+ continue;
+ }
/*
* Ping all waiters to retry
@@ -514,13 +559,22 @@ static NTSTATUS g_lock_force_unlock(struct g_lock_ctx *ctx, const char *name,
procid_str(talloc_tos(),
&locks[i].pid),
nt_errstr(status)));
+ } else {
+ num_wakeups += 1;
+ }
+ if (num_wakeups > 5) {
+ break;
}
}
}
done:
+ /*
+ * For the error path, TALLOC_FREE(rec) as well. In the good
+ * path we have already freed it.
+ */
+ TALLOC_FREE(rec);
TALLOC_FREE(locks);
- TALLOC_FREE(rec);
return status;
}
--
Samba Shared Repository
More information about the samba-cvs
mailing list