Dimitry Andric wrote:
[lots of stuff snipped]
> I'm also running into this problem, but while using lld.  I must set
> vfs.timestamp_precision to 1 (e.g. sec + ns accurate to 1/HZ) on both
> the client and the server, to make it work.
> Instead of GNU ld, lld uses mmap to write to the output executable.  If
> this executable is more than one page, and resides on an NFS share,
> running it will almost always result in "text file modification", if
> vfs_timestamp_precision >= 2.
> A small test case: http://www.andric.com/freebsd/test-mmap-write.c,
> which writes a simple "hello world" i386-freebsd executable file, using
> the sequence: open() -> ftruncate() -> mmap() -> memcpy() -> munmap() ->
> close().
Hopefully Kostik will correct me if I have this wrong, but I don't believe any
of the above syscalls guarantee that dirty pages have been flushed.
At least for cases without munmap(), the writes of dirty pages can occur after
the file descriptor is closed. I run into this in NFSv4, where there is a Close 
(NFSv4 one)
that can't be done until VOP_INACTIVE().
If you look in the NFS VOP_INACTIVE() { called ncl_inactive() } you'll see:
if (NFS_ISV4(vp) && vp->v_type == VREG) {
237                     /*
238                      * Since mmap()'d files do I/O after VOP_CLOSE(), the 
239                      * Close operations are delayed until now. Any dirty
240                      * buffers/pages must be flushed before the close, so 
that the
241                      * stateid is available for the writes.
242                      */
243                     if (vp->v_object != NULL) {
244                             VM_OBJECT_WLOCK(vp->v_object);
245                             retv = vm_object_page_clean(vp->v_object, 0, 0,
246                                 OBJPC_SYNC);
247                             VM_OBJECT_WUNLOCK(vp->v_object);
248                     } else
249                             retv = TRUE;
250                     if (retv == TRUE) {
251                             (void)ncl_flush(vp, MNT_WAIT, NULL, ap->a_td, 
1, 0);
252                             (void)nfsrpc_close(vp, 1, ap->a_td);
253                     }
254             }
Note that nothing like this is done for NFSv3.
What might work is implementing a VOP_SET_TEXT() vnode op for the NFS
client that does most of the above (except for nfsrpc_close()) and then sets
--> That way, all the dirty pages will be flushed to the server before the 
     starts executing.

> Running this on an NFS share, and then attempting to run the resulting
> 'helloworld' executable will result in the "text file modification"
> error, and it will be killed.  But if you simply copy the executable to
> something else, then it runs fine, even if you use -p to retain the
> properties!
> IMHO this is a rather surprising problem with the NFS code, and Kostik
> remarked that the problem seems to be that the VV_TEXT flag is set too
> early, before the nfs cache is invalidated.  Rick, do you have any ideas
> about this?
I don't think it is the "nfs cache" that needs invalidation, but the dirty
pages written by the mmap'd file need to be flushed, before the VV_TEXT
is set, I think?

If Kostik meant "attribute cache" when he said "nfs cache", I'll note that the
cached attributes (including mtime) are updated by the reply to every write.
(As such, I think it is the writes that must be done before setting VV_TEXT
  that is needed.)

It is a fairly simple patch to create. I'll post one to this thread in a day or 
unless Kostik thinks this isn't correct and not worth trying.


freebsd-current@freebsd.org mailing list
To unsubscribe, send any mail to "freebsd-current-unsubscr...@freebsd.org"

Reply via email to