Bug 742596

Description of problem:
(some background: we run several mysql databases with nfs datadirs, and recently upgraded one of our servers to rhel6. After unmounting and remounting the datadir, mysql was starting up with lots of errors about inaccessible tables. Some further troubleshooting found that when mysql starts up it tries to open the ibdata file with O_RDWR|O_CREAT|O_EXCL, and if that call is successful (which would generally indicate that the file didn't exist) it creates a new ibdata file, effectively destroying the relationship between datafiles on disk and the running database.)

Root problem: moving to rhel6, a regression appears to have been introduced, which causes an open() wiht the O_EXCL flag to return incorrect results on file on a newly mounted nfs filesytem.

Version-Release number of selected component (if applicable):
rhel 6.1

How reproducible:
compile attached o_excl.c and follow directions below.

Steps to Reproduce:
(in this case, /db is an nfs filesystem)

[mriddle@db101 src]$ cc -o o_excl o_excl.c 
[mriddle@db101 src]$ sudo touch /db/foo; sudo chown mriddle /db/foo
[mriddle@db101 src]$ ./o_excl /db/foo 
open() of /db/foo returned -1, not creating new file (this is good)
[mriddle@db101 src]$ sudo umount /db
[mriddle@db101 src]$ sudo mount /db
[mriddle@db101 src]$ ./o_excl /db/foo 
open() of /db/foo successful, this is bad!


Actual results:
After mounting an nfs filesystem, an open() that should fail succeeds

Expected results:
open() should fail in this scenario

Additional info: