Ok - this is a big patch - and it may break smbd a bit (although

I hope not). If you encounter strange file-serving behavior after this
patch then back it out. I analysed our stat() usage and realised we
were doing approx. 3 stat calls per open, and 2 per getattr/setattr.
This patch should fix all that. It causes the stat struct returned
from unix_convert() (which now *must* be passed a valid SMB_STRUCT_STAT
pointer) to be passed through into the open code. This should prevent
the multiple stats that were being done so as not to violate layer
encapsulation in the API's.

Herb - if you could run a NetBench test with this code and do a
padc/par syscall test and also run with the current 2.2.0 code
and test the padc/par syscalls I'd appreciate it - you should
find the number of stat calls reduced - not sure by how much.

The patch depends on unix_convert() actually finding the file
and returning a stat struct, or returning a zero'd out stat
struct if the file didn't exist. I believe we can guarentee this
to be the case - I just wasn't confident enough to make this
an assertion before.

Ok ok - I did write this whilst at the Miami conference.....
sometimes you get a little free time at these things :-).

Jeremy.
(This used to be commit 66a5c05ec46b641224fbe01b30bd7e83571a2a1b)

1 files changed, 2 insertions, 0 deletions

diff --git a/source3/smbd/vfs.c b/source3/smbd/vfs.c
index 467f3e47a2b..2f984aee4f3 100644
--- a/source3/smbd/vfs.c
+++ b/source3/smbd/vfs.c
@@ -323,6 +323,8 @@ BOOL vfs_file_exist(connection_struct *conn,char *fname,SMB_STRUCT_STAT *sbuf)
 
 	if (!sbuf)
 		sbuf = &st;
+
+	ZERO_STRUCTP(sbuf);
  
 	if (vfs_stat(conn,fname,sbuf) != 0) 
 		return(False);