IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
sharemode db in the following way.
Originally, on startup and shutdown, smbd would scan the share mode
db to ensure it was correct. This lead to scalability issues as
scans lock the db for quite a long time. Andrew had the brainstorm
that we only care about the record we're about to read.
This new code (small change really, but quite significant) causes
get_share_modes() to do a process_exists() call against each pid
in each record, and to delete any that don't and re-write the
entry if any dead records were detected.
This allowed me to remove the startup/shutdown scans of the
db (they can be added into smbstatus if anyone really cares to
have them back). This will please the vfs author who was worried
about the time taken on open() calls, and will lead to much
greater robustness and scalability in the share mode db.
We need much testing of this, and also netbench tests to
ensure the extra process_exists() calls don't hurt performance
(they shouldn't it's a very simple system call).
Jeremy.
was not forced to be 8 byte aligned. Use union to force it to be correctly aligned
for memcpy and use void *, to tell compiler not to optimize aligned copy (this last fix
suggested by Trond @ RedHat). The first fix should be sufficient, but this provides a
"belt and braces" fix.
Jeremy.
we set the DELETE_ON_CLOSE_FLAG on all share modes on the file, which
means the share mode in the fsp will not match the one in the tdb when
we come to close for other file handles, which means we end up with
share modes on files after all handles are closed
fixed by making the comparison function that says if two shares modes
are equal ignore the DELETE_ON_CLOSE_FLAG
major changes include:
- added NSTATUS type
- added automatic mapping between dos and nt error codes
- changed all ERROR() calls to ERROR_DOS() and many to ERROR_NT()
these calls auto-translate to the client error code system
- got rid of the cached error code and the writebmpx code
We eventually will need to also:
- get rid of BOOL, so we don't lose error info
- replace all ERROR_DOS() calls with ERROR_NT() calls
but that is too much for one night
link from Seattle is having problems.
I've added 3 things here to work on the fcntl spin
problem.
1). Check *all* tdb return codes... :-).
2). If we're asking ourselves to break an oplock, and we can't
find a fsp pointer that matches the entry, this is a *logic bug*
and we should abort and panic so someone with gdb can pick up
the pieces.
3). After we've broken an oplock, ensure that the entry itself
has been removed, and if not remove it ourselves. This should
not be neccessary in a correctly working environmen,t, but will
provide an added layer of robustness in error situations.
4). I hate german keyboards :-) :-).
Jeremy.
when copying to a full disk problem, I discovered that we were not allowing
the delete on close flag to be set properly, this led to other things, and
after investigation of the proper delete on close semantics and their relationship
to the file_share_delete flag I discovered there were some cases where we
weren't doing the deny modes properly. And this after only 5 years working
on them..... :-) :-).
So here's the latest attempt. I realised the delete on close flag needs to
be set across all smbds with a dev/ino pair open - in addition, the delete
on close flag, allow share delete and delete access requested all need to
be stored in the share mode tdb.
The "delete_on_close" entry in the fsp struct is now redundant and should
really be removed. This may also mean we can get rid of the "iterate_fsp"
calls that I didn't like adding in the first place. Whilst doing this patch,
I also discovered we needed to do the se_map_generic() call for file opens
and POSIX ACL mapping, so I added that also.
This code, although ugly, now passes the deny mode torture tests plus the
delete on close tests I added. I do need to add one more multiple connection
delete on close test to make sure I got the semantics exactly right, plus we
should also (as Andrew suggested) move to random testing here.
The good news is that NT should now correctly delete the file on disk
full error when copying to a disk :-).
Jeremy.
a --with-spinlocks option to configure, this does mean the on-disk tdb
format has changed, so 2.2alphaX sites will need to re-create their
tdb's. The upside is no more tdb fragmentation and a +5% on netbench.
Swings and roundabouts....
Jeremy.
to find bugs. On 64 bit IRIX, structure packing means that
a
struct {
SMB_DEV_T dev /* 4 bytes */
SMB_INO_T ino /* 8 bytes */
}
has 4 bytes of padding between the two members. If you
don't null the memory before using it as a tdb key,
you randomly can't find keys depending on what is in
the padding. This caused me immense pain and was hard
to track down.... :-)
Jeremy.
This implementation keeps all POSIX lock records in a separate in memory
tdb database only known about in locking/posix.c. In addition, the pending
close fd's are also held in a tdb which has an array of fd's indexed by
device and inode.
The walk-split code uglyness has been moved to posix.c from brlock.c,
which is the only place that needs to know about it, and the extra
functions hacked into brlock to expose internal state have been removed.
This implementation passes smbtorture locktest4, the only thing I need
to check now for completeness is what to do about lock upgrade/downgrades
which Win32 allows under some *very* strange circumstances.
Jeremy.
fd_close now calls fd_close_posix() directly.
set_posix_lock/release_posix_lock() now handle the reference counting.
More changes due when this gets moved to the file locking/posix.c
Jeremy.
When a file is being closed, once it passes the fnum and tid tests then
the locking context should be ignored when removing all locks. This is
what is done in the brl close case, but when you have outstanding
POSIX locks, then you cannot remove all the brl locks in one go, you
have to get the lock list and call do_unlock individually. As this
uses global_smbpid as the locking context, you need to make sure
that this is set correctly for the specific lock being removed. I
now do this by storing the smbpid in each entry in the unlock list returned from
the query call. I removed the smbpid from fsp (not needed) and
things seem ok (even with the stupid smbpid tricks that smbtorture plays :-).
Jeremy.
smbpid used when a file was opened in the files_struct. Else we use
the wrong global_smbpid when we are closing the file and trying to
remove the brl locks - this causes the brl locks to be left when the
file is closed as the samba_context check fails.
Jeremy.
removed from the smbd/open.c code.
We now use a dlink list of structures indexed by dev/inode to store
all pending fd's for close. This could be rewritten to use lib/hash.c
if this is discovered to be too slow in use.
Andrew, please take a look and let me know if this is what you
had in mind.
Jeremy.
test. Was miscounting posix locks, plus was not taking into account
the case where other_fsp == fsp in the 'move locks' case. DOH ! This
code will be re-written anyway :-).
Jeremy.
HEAD should now map brl locks correctly into POSIX locks, including the
really nasty case of large range unlock.
There is a lot of pretty ASCII art in locking/brlock.c explaining
exactly how this code works. If it is unclear, please ask me.
Jeremy.
