pread vs. git

[Eric Blake <eblake at redhat dot com>]

Recent git added a hack to explicitly avoid cygwin's pread as unsafe:

http://git.661346.n2.nabble.com/PATCH-v2-1-1-index-pack-Disable-threading-on-cygwin-td7562195.html

Comments like this aren't very re-assuring either:

http://sourceware.org/ml/cygwin/2011-06/msg00057.html

or this big hairy comment in the source of fhandle_disk_file:

   Using this handle for pread/pwrite would break atomicity, because the
   read/write operation would have to be followed by a seek back to the old
   file position.  What we do is to open another handle to the file on the
   first call to either pread or pwrite.  This is used for any subsequent
   pread/pwrite.  Thus the current file position of the "normal" file
   handle is not touched.

   FIXME:

   Note that this is just a hack.  The problem with this approach is that
   a change to the file permissions might disallow to open the file with
   the required permissions to read or write.  This appears to be a
border case,
   but that's exactly what git does.  It creates the file for reading and
   writing and after writing it, it chmods the file to read-only.  Then it
   calls pread on the file to examine the content.  This works, but if git
   would use the original handle to pwrite to the file, it would be broken
   with our approach.

   One way to implement this is to open the pread/pwrite handle right at
   file open time.  We would simply maintain two handles, which wouldn't
   be much of a problem given how we do that for other fhandler types as
   well.

   However, ultimately fhandler_disk_file should become a derived class of
   fhandler_base_overlapped.  Each raw_read or raw_write would fetch the
   actual file position, read/write from there, and then set the file
   position again.  Fortunately, while the file position is not maintained
   by the I/O manager, it can be fetched and set to a new value by all
   processes holding a handle to that file object.  Pread and pwrite differ
   from raw_read and raw_write just by not touching the current file pos.


What still remains to make pread a first-class thread-safe
implementation that obeys POSIX, so that I don't have to cripple my next
build of git to avoid threaded pread?

-- 
Eric Blake   eblake@redhat.com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

Attachment: signature.asc
Description: OpenPGP digital signature