On Fri, Oct 05, 2007 at 01:50:39PM -0400, John Baldwin wrote: > On Thursday 04 October 2007 09:29:25 pm Jeff Roberson wrote: > > On Thu, 4 Oct 2007, John Baldwin wrote: > > > > > On Wednesday 03 October 2007 07:48:00 pm Jeff Roberson wrote: > > >> On Wed, 3 Oct 2007, John Baldwin wrote: > > >> > > >>> jhb 2007-10-03 21:06:05 UTC > > >>> > > >>> FreeBSD src repository > > >>> > > >>> Modified files: (Branch: RELENG_6) > > >>> sys/fs/devfs devfs_vnops.c > > >>> sys/fs/fifofs fifo_vnops.c > > >>> sys/kern uipc_usrreq.c vfs_vnops.c > > >>> sys/vm vnode_pager.c > > >>> Log: > > >>> MFC: Always use an exclusive lock on the leaf vnode during an open() > when > > >>> shared lookups are enabled. This closes a few races including a race > > > where > > >>> concurrent opens of a fifo could result in different v_fifoinfo > > > structures > > >>> in different threads. > > >> > > >> Long term we should really look for a better solution to this problem. > > >> There are a number of was to improve snapshots in ffs by fixing shared > > >> locking. > > > > > > I don't disagree. The fifo case can be fixed easily enough in the fifo > code > > > by using fifo_mtx to protect v_fifoinfo perhaps (or doing an upgrade on > the > > > vnode lock?), but for the MFC I didn't want to have to fix each of the > races > > > with open(2). Probably better to fix it more properly in HEAD first. > > > > Definitely someting for head. Were there any others that you ran into > > besides v_fifoinfo? We should audit this more closely anyhow. I have > > been reluctant to push too much shared locking into VFS because it's not > > been so carefully studied. > > I just saw v_fifoinfo, but Pawel's original commit referenced updates to > v_writecount, etc. The v_writecount one is in vn_open() itself: > > if ((error = VOP_OPEN(vp, fmode, cred, td, fp)) != 0) > goto bad; > > if (fmode & FWRITE) > vp->v_writecount++; > *flagp = fmode; > ASSERT_VOP_ELOCKED(vp, "vn_open_cred"); > if (!mpsafe) > VFS_UNLOCK_GIANT(vfslocked); > return (0); > > If you just held a shared lock there, you could use atomic ops for > vp->v_writecount (and still hold at least a shared vnode lock everywhere > v_writecount is updated) and still be able to read vp->v_writecount safely > while holding an exclusive lock on the vnode.
Another is the vm/vnode_pager.c handling of the v_object.
pgpNU9ql5codJ.pgp
Description: PGP signature