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.
is that there are some times when we should return an EINTR from a select,
some times when we should not. As we can take a signal at any time, we
have to eat EINTR's in some selects. This means we need to check for
kernel oplock breaks more often in the main loop, as well as add the
queuing mechanism needed for the changenotify code (due to the mistake
in understanding POSIX semantics w.r.t. setting a signal mask in a
signal handler). This code now passes all my tests.
However, (and IMHO and I know tridge disagrees) - the correct way to
fix this is to run with RT signals blocked and explicitly unblock
them just before the main select, block them after and then process
them all in one place. Just my 2cents :-).
Jeremy.
handling in Samba. This was needed due to several limitations and
races in the previous code - as a side effect the new code is much
cleaner :)
in summary:
- changed sys_select() to avoid a signal/select race condition. It is a
rare race but once we have signals doing notification and oplocks it
is important.
- changed our main processing loop to take advantage of the new
sys_select semantics
- split the notify code into implementaion dependent and general
parts. Added the following structure that defines an implementation:
struct cnotify_fns {
void * (*register_notify)(connection_struct *conn, char *path, uint32 flags);
BOOL (*check_notify)(connection_struct *conn, uint16 vuid, char *path, uint32 flags, void *data, time_t t);
void (*remove_notify)(void *data);
};
then I wrote two implementations, one using hash/poll (like our old
code) and the other using the new Linux kernel change notify. It
should be easy to add other change notify implementations by creating
a sructure of the above type.
- fixed a bug in change notify where we were returning the wrong error
code.
- rewrote the core change notify code to be much simpler
- moved to real-time signals for leases and change notify
Amazingly, it all seems to work. I was very surprised!
I used a trick where CAP_LEASE isn't claimed until it is needed. This
means we avoid a system call per setreuid(), and never call capset()
unless a user tries to get a oplock on a file that they don't own
