Re: Updated I/O Consistency & Semantics Section

Steve Landherr (landherr@mailhost.rsn.hp.com)
Thu, 29 May 1997 14:11:07 -0500

William Gropp wrote:
>
> This still contains text about the data being "written to permanent storage"
> (page 47, line 45). I'm uncomfortable with this. For example, it is stronger
> than my reading of POSIX 1003.1, 6.6, File Synchronization. See in particular
> lines 599-601 and 614-617 in that document. Basically, POSIX defines "file
> synchronization" by its observable effect, not a particular implementation
> strategy. I believe we should do the same.

The POSIX systems to which I have access all include moving modified
file data and attributes to permanent storage in their description of
the fsync(2) system call. And from my knowledge of the internals of
three of these systems (HP-UX, SPP-UX, and ConvexOS), I know that the
system call does not return until the writes have completed.

This functionality is necessary to support applications that
periodically checkpoint themselves. They don't want to remove the old
checkpoint file until they are sure the new file is on disk.

-- 
Steve Landherr -- landherr@convex.hp.com -- (972)497-4104
Kernel Software Development -- File Systems
Convex Division of Hewlett-Packard Company -- Richardson, Texas