freebsd stable - Jun 2009 - kern/135412: [zfs] [nfs] zfs(v13)+nfs and open(..., O_WRONLY|O_CREAT|O_EXCL, ...)

hi,
	This pr is realy holding me back, I can't upgrade this server, and
telling serveral tens of users to us cp, etc is not an option. The open
works fine if not using O_EXCL. 

Thanks,
	danny

Quoting "Danny Braniss" <danny@cs.huji.ac.il>:
> hi,
> 	This pr is realy holding me back, I can't upgrade this server, and
> telling serveral tens of users to us cp, etc is not an option. The open
> works fine if not using O_EXCL.
>
> Thanks,
> 	danny
This affects procmail deliveries as well.   Any chance of having this
looked at asap? So I'm on zfs 13, I can't drop back zfs versions, and my
production data is stuck.  Yes, I know this isn't a release kernel, but
I was having ungodly slowness with ZFS and was trying everything I could
think of.  I had, due daily to archiving, 50-100thousand 3k files in
each of about 200 directories. Once I moved them all to another machine
(on UFS), the ZFS partition returned to normal speeds...But now I'm
stuck with this delivery issue (fortunately it's only procmail) due to
the upgrade attempt.  I just ran into the 'mv' problem yesterday.

Any relief in sight?

Thanks,
Rick

On 2009-06-30, Danny Braniss wrote:
> 	This pr is realy holding me back, I can't upgrade this server, and
> telling serveral tens of users to us cp, etc is not an option. The open
> works fine if not using O_EXCL. 
I guess that r185586 needs to be MFCd to stable/7. Here's an untested
patch against stable/7:

%%%
Index: sys/nfsserver/nfs_serv.c
==================================================================---
sys/nfsserver/nfs_serv.c	(revision 195189)
+++ sys/nfsserver/nfs_serv.c	(working copy)
@@ -1656,13 +1656,12 @@ nfsrv_create(struct nfsrv_descript *nfsd
 	caddr_t bpos;
 	int error = 0, rdev, len, tsize, dirfor_ret = 1, diraft_ret = 1;
 	int v3 = (nfsd->nd_flag & ND_NFSV3), how, exclusive_flag = 0;
-	caddr_t cp;
 	struct mbuf *mb, *mreq;
 	struct vnode *dirp = NULL;
 	nfsfh_t nfh;
 	fhandle_t *fhp;
 	u_quad_t tempsize;
-	u_char cverf[NFSX_V3CREATEVERF];
+	struct timespec cverf;
 	struct mount *mp = NULL;
 	int tvfslocked;
 	int vfslocked;
@@ -1741,8 +1740,11 @@ nfsrv_create(struct nfsrv_descript *nfsd
 			nfsm_srvsattr(vap);
 			break;
 		case NFSV3CREATE_EXCLUSIVE:
-			cp = nfsm_dissect_nonblock(caddr_t, NFSX_V3CREATEVERF);
-			bcopy(cp, cverf, NFSX_V3CREATEVERF);
+			tl = nfsm_dissect_nonblock(u_int32_t *,
+			    NFSX_V3CREATEVERF);
+			/* Unique bytes, endianness is not important. */
+			cverf.tv_sec  = tl[0];
+			cverf.tv_nsec = tl[1];
 			exclusive_flag = 1;
 			break;
 		};
@@ -1788,8 +1790,7 @@ nfsrv_create(struct nfsrv_descript *nfsd
 				if (exclusive_flag) {
 					exclusive_flag = 0;
 					VATTR_NULL(vap);
-					bcopy(cverf, (caddr_t)&vap->va_atime,
-						NFSX_V3CREATEVERF);
+					vap->va_atime = cverf;
 					error = VOP_SETATTR(nd.ni_vp, vap, cred,
 						td);
 				}
@@ -1873,7 +1874,7 @@ nfsrv_create(struct nfsrv_descript *nfsd
 	}
 	if (v3) {
 		if (exclusive_flag && !error &&
-			bcmp(cverf, (caddr_t)&vap->va_atime, NFSX_V3CREATEVERF))
+		    bcmp(&cverf, &vap->va_atime, sizeof (cverf)))
 			error = EEXIST;
 		if (dirp == nd.ni_dvp)
 			diraft_ret = VOP_GETATTR(dirp, &diraft, cred, td);
%%%

Please report back if you try the patch.

-- 
Jaakko