Hi Paul.

My fuse_daemon implementation does exactly as you describe. 
Experimentally, I have discovered that DiskArbitration services aren't 
available at shutdown time (trying to create a DADiskRef from a BSD name 
fails, at least), but resorting to unmount(const char*, int) does the job.
However, if ntfs-3g manages to catch the signal before fuse_daemon, both 
processes both lock up (in the fuse_daemon case, it happens when 
fuse_daemon is in the process of carrying out the unmount(...) system 
call for the ntfs-3g file system).
In that situation both ntfs-3g and fuse_daemon gets forcibly terminated 
with SIGKILL after 30 seconds.

- Erik

Paul Marks skrev:
> Good afternoon,
>
> When I was working on NTFS-3G, I tried to avoid modifying FUSE or  
> ntfs-3g itself.  Instead, I just wrote a short daemon that blocked on  
> sigsuspend() until it received SIGTERM, SIGQUIT, or SIGINT.
>
> If it received SIGTERM from its parent (i.e. launchd), that implies  
> system shutdown.  Then I would scan mounted filesystems with  
> getmntinfo() looking for those of type "fusefs," and unmount() those,  
> perhaps with MNT_FORCE.  With some of the changes to MacFUSE, you may  
> also want to scan for anything of a type matching "fusefs_*" as well.
>
> This would cause the ntfs-3g binary to terminate cleanly, without  
> needing to handle SIGTERM.  A more advisable option would be to go  
> through DiskArbitration and give DADiskUnmount() a shot before  
> resorting to unmount(), which will assure proper notifications get  
> sent to things like Spotlight before the volume is forcibly unmounted.
>
> It worked for me, at the time.
>
> Hope this helps,
>    - Paul
>
> On Feb 6, 2008, at 4:04 AM, Erik Larsson wrote:
>
>   
>> Hi group, hi Amit.
>>
>> I'm maintaining an ntfs-3g package for OS X at
>> http://macntfs-3g.blogspot.com and I'm struggling with an issue where
>> the ntfs-3g process locks up at shutdown, and has to be SIGKILL'd by  
>> the
>> OS after a delay of 30 seconds.
>> ntfs-3g is not doing any signal handling itself in the current  
>> version,
>> and instead passes that responsibility over the fuse library by  
>> invoking:
>>
>>    fuse_set_signal_handlers(fuse_get_session(struct fuse*))
>>
>> When I send the ntfs-3g process SIGTERM under "normal"  
>> circumstances, in
>> a non-shutdown situation, there is no problem. ntfs-3g terminates
>> gracefully. Some kind of deadlock seems to occur when this is done  
>> in a
>> shutdown situation though.
>>
>> I looked through the MacFUSE patchset and found a modification to
>> fuse_signals.c which seems suspicious to me (scroll down to the end of
>> the mail to see what I'm talking about). I don't have a very good view
>> of the internal structure of fuse, but it seems like a fair assumption
>> that fuse_signals.c takes care of signal handling.
>> In the below code snippet, the exit_handler forks (seems dangerous  
>> to me
>> in a shutdown situation where processes are killed off all the time),
>> and invokes the command /sbin/umount . I'm suggesting that this  
>> behavior
>> is somehow causing the lockup, but I have no real "evidence" for it,
>> just thoughts.
>>
>> As a workaround, I have written a daemon, fuse_daemon (actually a
>> reimplementation of Paul Marks' utility for the older ntfs-3g package)
>> that waits in the background until it gets SIGTERM, and then iterates
>> through all fuse file systems and unmounts them cleanly. This only  
>> seems
>> to work if fuse_daemon gets SIGTERM before ntfs-3g does, so signal
>> handling in ntfs-3g needs to be turned off completely for this  
>> solution
>> work.
>>
>> I hope you realize the serious implication of this problem... if  
>> ntfs-3g
>> (or indeed any other FUSE file system) doesn't terminate gracefully,  
>> it
>> might leave the user's hard drive in an inconsistent state, worst  
>> case.
>> Has anyone encountered this issue before, and solved it without having
>> an external daemon manage unmounting of file systems?
>>
>> - Erik Larsson
>>
>> ---------
>> diff -Naur old/lib/fuse_signals.c new/lib/fuse_signals.c
>> --- old/lib/fuse_signals.c    2007-10-16 09:35:23.000000000 -0700
>> +++ new/lib/fuse_signals.c    2008-01-01 15:56:28.000000000 -0800
>> @@ -13,12 +13,45 @@
>> #include <signal.h>
>>
>> static struct fuse_session *fuse_instance;
>> +#if (__FreeBSD__ >= 10)
>> +extern char *fuse_session_get_mntonname(struct fuse_session *se);
>> +
>> +#include <unistd.h>
>> +
>> +int
>> +fuse_chan_fd_np(void)
>> +{
>> +    if (fuse_instance && !fuse_session_exited(fuse_instance)) {
>> +        return fuse_chan_fd(fuse_session_next_chan(fuse_instance,  
>> NULL));
>> +    } else {
>> +        return -1;
>> +    }
>> +}
>> +
>> +#endif
>>
>> static void exit_handler(int sig)
>> {
>>    (void) sig;
>> +#if (__FreeBSD__ >= 10)
>> +    if (fuse_instance && !fuse_session_exited(fuse_instance)) {
>> +        int fd;
>> +        pid_t pid;
>> +
>> +        fd = fuse_chan_fd(fuse_session_next_chan(fuse_instance,  
>> NULL));
>> +        pid = fork();
>> +        if (pid == 0) { /* child */
>> +             char *mntonname =  
>> fuse_session_get_mntonname(fuse_instance);
>> +             fcntl(fd, F_SETFD, 1); /* close-on-exec */
>> +             execl("/sbin/umount", "/sbin/umount", mntonname, NULL);
>> +        } else {
>> +            /* We do nothing in the parent. */
>> +        }
>> +    }
>> +#else
>>    if (fuse_instance)
>>        fuse_session_exit(fuse_instance);
>> +#endif
>> }
>>
>> static int set_one_signal_handler(int sig, void (*handler)(int))
>> ---------
>>
>>
>>     
>
>
> >
>   


--~--~---------~--~----~------------~-------~--~----~
You received this message because you are subscribed to the Google Groups 
"macfuse-devel" group.
To post to this group, send email to [email protected]
To unsubscribe from this group, send email to [EMAIL PROTECTED]
For more options, visit this group at 
http://groups.google.com/group/macfuse-devel?hl=en
-~----------~----~----~----~------~----~------~--~---

Reply via email to