tech-userlevel archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: mtree weirdness



On Mon, Mar 02, 2009 at 01:06:32PM +0000, Iain Hibbert wrote:
  | [when the current directory is in a unionfs mount]
  | 
  | Ok I found the root cause, the fix might need consideration :)
  | 
  | The problem is in libc/gen/opendir.c - fts(3) uses the internal
  | __opendir2() entry point and always passes the DTF_REWIND flag which
  | causes a pause/rewind halfway through a unionfs scan because there is some
  | special handling for cases when a fchdir() is to be made according to the
  | comment.
  | 
  | Because fts(3) is not always available, it is added to the tools build but
  | lacking the NETBSD_SOURCE definition causing the __opendir2() call to be
  | translated into the standard opendir() expected in the host libc.
  | 
  | Our libc opendir (dating from CSRG sync in 94) is a stub call, comprising:
  | 
  |     return (__opendir2(name, DTF_HIDEW|DTF_NODUP));
  | 
  | and so the essential difference is the lack of the DTF_REWIND flag. If I
  | add DTF_REWIND to that (as below) and rebuild libc, then
  | "/usr/tools/bin/nbmtree -c" works fine in a unionfs directory.
  | 
  | Now, DTF_REWIND is only required in a unionfs mount situtation (I think)
  | but it is present for the nfs case also and so adding it to default
  | opendir() call might not be correct (but fts always uses it anyway)
  | 
  | Thoughts?
  | 
  | iain
  | 
  | --- /usr/src/lib/libc/gen/opendir.c 2008-01-10 09:49:04.000000000 +0000
  | +++ src/lib/libc/gen/opendir.c      2009-03-02 10:54:30.000000000 +0000
  | @@ -66,7 +66,7 @@ opendir(const char *name)
  | 
  |     _DIAGASSERT(name != NULL);
  | 
  | -   return (__opendir2(name, DTF_HIDEW|DTF_NODUP));
  | +   return (__opendir2(name, DTF_HIDEW|DTF_NODUP|DTF_REWIND));
  |  }
  | 
  |  DIR *

I think your fix will only resolve the "host libc" problem
when building on a NetBSD-current system with the fix, and
not other BSD-derived implementations of opendir() without
the fix.

The patch may be too much of a "sledgehammer" approach; it
effectively forces the DTF_REWIND functionality upon all
callers of opendir() that traverse NFS or union file systems.
This may be a "good thing" -- correctness over performance.
Further consideration of the ramifications is required.

On the other hand, it makes unnecessary the "backdoor" call to
__opendir2() that fts_build() is using.  I haven't checked
if there's other callers within the tree using that API.


cheers,
Luke.

Attachment: pgpjkFjar9ghJ.pgp
Description: PGP signature



Home | Main Index | Thread Index | Old Index