safely remove USB hard drive

[Patrick O'Callaghan]

On Thu, 2008-04-24 at 12:47 -0400, Joe Smith wrote:
> Patrick O'Callaghan wrote:
> > On Thu, 2008-04-24 at 10:49 -0400, Joe Smith wrote:
> >> ...
> >> I'm not convinced that using sync(1) would be effective, especially from
> >> the desktop environment where who knows what processes have something
> >> open on a pluggable device. Sync would be prudent if you were moving raw
> >> data blocks to a device, but not for mounted filesystems.
> > 
> > Of course sync is effective for mounted filesystems, why wouldn't it be?
> > It means data are now synched *up to this point*.
> 
> By "not effective" I don't mean that it doesn't work, I mean that it 
> can't guarantee that the device is safe to remove. What's to stop 
> something from starting a write to the device 10us after the kernel 
> syncs all the buffers? The sync+unmount has to be atomic in order to 
> guarantee that the data buffers are flushed and that nothing else can 
> start a new write. unmount(2) is the logical (only?) place to lock the 
> two operations together.

Correct, but show me where it says that unmount does a sync. If some
"official" doc somewhere says that that's what happens, then I'll accept
it, but until then it's speculation *even if it's in the source*. This
point needs to be stressed. The docs are the only collective memory we
have about how things are *supposed* to work. Everything else is written
in sand.

> > ...
> > There is no documentary evidence that unmount calls sync (contrary to
> > what many people think, including me up to about yesterday). The
> > eject(1) page doesn't mention syncing or flushing buffers.
> 
> A man page is usually an afterthought, not a spec. A unix without 
> incomplete man pages wouldn't be unix ;-)
> 
> As you yourself pointed out, even a look at the source code won't prove 
> it. If you have evidence that umount (or eject) alone is not sufficient 
> to flush all pending i/o for the device, then it should be reported as a 
> bug.
> 
> Use "sync && umount" if you like, there's no harm in it, but sync alone 
> is not a substitute for umount. eject(1) can safely substitute for 
> umount, because eject is designed to unmount the device.

Who said sync was a substitute for unmount? The point at issue is
whether you need both of them, or just unmount on its own.

poc