/*
* nfs_request - goes something like this
* - fill in request struct
* - links it into list
* - calls nfs_send() for first transmit
* - calls nfs_receive() to get reply
* - break down rpc header and return with nfs reply pointed to
* by mrep or error
* nb: always frees up mreq mbuf list
*/
int
nfs_request(struct vnode *vp, struct mbuf *mreq, int procnum,
struct thread *td, struct ucred *cred, struct mbuf **mrp,
struct mbuf **mdp, caddr_t *dposp)
{
struct mbuf *mrep;
u_int32_t *tl;
struct nfsmount *nmp;
struct mbuf *md;
time_t waituntil;
caddr_t dpos;
int error = 0;
struct timeval now;
AUTH *auth = NULL;
enum nfs_rto_timer_t timer;
struct nfs_feedback_arg nf;
struct rpc_callextra ext;
enum clnt_stat stat;
struct timeval timo;
/* Reject requests while attempting a forced unmount. */
if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
m_freem(mreq);
return (ESTALE);
}
nmp = VFSTONFS(vp->v_mount);
bzero(&nf, sizeof(struct nfs_feedback_arg));
nf.nf_mount = nmp;
nf.nf_td = td;
getmicrouptime(&now);
nf.nf_lastmsg = now.tv_sec -
((nmp->nm_tprintf_delay) - (nmp->nm_tprintf_initial_delay));
/*
* XXX if not already connected call nfs_connect now. Longer
* term, change nfs_mount to call nfs_connect unconditionally
* and let clnt_reconnect_create handle reconnects.
*/
if (!nmp->nm_client)
nfs_connect(nmp);
auth = nfs_getauth(nmp, cred);
if (!auth) {
m_freem(mreq);
return (EACCES);
}
bzero(&ext, sizeof(ext));
ext.rc_auth = auth;
ext.rc_feedback = nfs_feedback;
ext.rc_feedback_arg = &nf;
/*
* Use a conservative timeout for RPCs other than getattr,
* lookup, read or write. The justification for doing "other"
* this way is that these RPCs happen so infrequently that
* timer est. would probably be stale. Also, since many of
* these RPCs are non-idempotent, a conservative timeout is
* desired.
*/
timer = nfs_rto_timer(procnum);
if (timer != NFS_DEFAULT_TIMER)
ext.rc_timers = &nmp->nm_timers[timer - 1];
else
ext.rc_timers = NULL;
#ifdef KDTRACE_HOOKS
if (dtrace_nfsclient_nfs23_start_probe != NULL) {
uint32_t probe_id;
int probe_procnum;
if (nmp->nm_flag & NFSMNT_NFSV3) {
probe_id = nfsclient_nfs3_start_probes[procnum];
probe_procnum = procnum;
} else {
probe_id = nfsclient_nfs2_start_probes[procnum];
probe_procnum = nfsv2_procid[procnum];
}
if (probe_id != 0)
(dtrace_nfsclient_nfs23_start_probe)(probe_id, vp,
mreq, cred, probe_procnum);
}
#endif
nfsstats.rpcrequests++;
tryagain:
timo.tv_sec = nmp->nm_timeo / NFS_HZ;
timo.tv_usec = (nmp->nm_timeo * 1000000) / NFS_HZ;
mrep = NULL;
stat = CLNT_CALL_MBUF(nmp->nm_client, &ext,
(nmp->nm_flag & NFSMNT_NFSV3) ? procnum : nfsv2_procid[procnum],
mreq, &mrep, timo);
/*
* If there was a successful reply and a tprintf msg.
* tprintf a response.
*/
if (stat == RPC_SUCCESS)
error = 0;
else if (stat == RPC_TIMEDOUT)
error = ETIMEDOUT;
else if (stat == RPC_VERSMISMATCH)
error = EOPNOTSUPP;
else if (stat == RPC_PROGVERSMISMATCH)
error = EPROTONOSUPPORT;
else
error = EACCES;
if (error)
goto nfsmout;
KASSERT(mrep != NULL, ("mrep shouldn't be NULL if no error\n"));
/*
* Search for any mbufs that are not a multiple of 4 bytes long
* or with m_data not longword aligned.
* These could cause pointer alignment problems, so copy them to
* well aligned mbufs.
*/
error = nfs_realign(&mrep, M_DONTWAIT);
if (error == ENOMEM) {
m_freem(mrep);
AUTH_DESTROY(auth);
return (error);
}
md = mrep;
dpos = mtod(mrep, caddr_t);
tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED);
if (*tl != 0) {
error = fxdr_unsigned(int, *tl);
if ((nmp->nm_flag & NFSMNT_NFSV3) &&
error == NFSERR_TRYLATER) {
m_freem(mrep);
error = 0;
waituntil = time_second + nfs3_jukebox_delay;
while (time_second < waituntil)
(void)tsleep(&fake_wchan, PSOCK, "nqnfstry",
hz);
goto tryagain;
}
/*
* If the File Handle was stale, invalidate the lookup
* cache, just in case.
*/
if (error == ESTALE)
nfs_purgecache(vp);
/*
* Skip wcc data on NFS errors for now. NetApp filers
* return corrupt postop attrs in the wcc data for NFS
* err EROFS. Not sure if they could return corrupt
* postop attrs for others errors.
*/
if ((nmp->nm_flag & NFSMNT_NFSV3) &&
!nfs_skip_wcc_data_onerr) {
*mrp = mrep;
*mdp = md;
*dposp = dpos;
error |= NFSERR_RETERR;
} else
m_freem(mrep);
goto nfsmout;
}
/*
* nfs_request - goes something like this
* - fill in request struct
* - links it into list
* - calls nfs_send() for first transmit
* - calls nfs_receive() to get reply
* - break down rpc header and return with nfs reply pointed to
* by mrep or error
* nb: always frees up mreq mbuf list
*/
int
nfs_request(struct vnode *vp, struct mbuf *mreq, int procnum,
struct thread *td, struct ucred *cred, struct mbuf **mrp,
struct mbuf **mdp, caddr_t *dposp)
{
struct mbuf *mrep;
u_int32_t *tl;
struct nfsmount *nmp;
struct mbuf *md;
time_t waituntil;
caddr_t dpos;
int error = 0, timeo;
AUTH *auth = NULL;
enum nfs_rto_timer_t timer;
struct nfs_feedback_arg nf;
struct rpc_callextra ext;
enum clnt_stat stat;
struct timeval timo;
/* Reject requests while attempting a forced unmount. */
if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
m_freem(mreq);
return (ESTALE);
}
nmp = VFSTONFS(vp->v_mount);
bzero(&nf, sizeof(struct nfs_feedback_arg));
nf.nf_mount = nmp;
nf.nf_td = td;
nf.nf_lastmsg = time_uptime -
((nmp->nm_tprintf_delay) - (nmp->nm_tprintf_initial_delay));
/*
* XXX if not already connected call nfs_connect now. Longer
* term, change nfs_mount to call nfs_connect unconditionally
* and let clnt_reconnect_create handle reconnects.
*/
if (!nmp->nm_client)
nfs_connect(nmp);
auth = nfs_getauth(nmp, cred);
if (!auth) {
m_freem(mreq);
return (EACCES);
}
bzero(&ext, sizeof(ext));
ext.rc_auth = auth;
ext.rc_feedback = nfs_feedback;
ext.rc_feedback_arg = &nf;
/*
* Use a conservative timeout for RPCs other than getattr,
* lookup, read or write. The justification for doing "other"
* this way is that these RPCs happen so infrequently that
* timer est. would probably be stale. Also, since many of
* these RPCs are non-idempotent, a conservative timeout is
* desired.
*/
timer = nfs_rto_timer(procnum);
if (timer != NFS_DEFAULT_TIMER)
ext.rc_timers = &nmp->nm_timers[timer - 1];
else
ext.rc_timers = NULL;
#ifdef KDTRACE_HOOKS
if (dtrace_nfsclient_nfs23_start_probe != NULL) {
uint32_t probe_id;
int probe_procnum;
if (nmp->nm_flag & NFSMNT_NFSV3) {
probe_id = nfsclient_nfs3_start_probes[procnum];
probe_procnum = procnum;
} else {
probe_id = nfsclient_nfs2_start_probes[procnum];
probe_procnum = nfsv2_procid[procnum];
}
if (probe_id != 0)
(dtrace_nfsclient_nfs23_start_probe)(probe_id, vp,
mreq, cred, probe_procnum);
}
#endif
nfsstats.rpcrequests++;
tryagain:
/*
* This timeout specifies when a new socket should be created,
* along with new xid values. For UDP, this should be done
* infrequently, since retransmits of RPC requests should normally
* use the same xid.
*/
if (nmp->nm_sotype == SOCK_DGRAM) {
if ((nmp->nm_flag & NFSMNT_SOFT) != 0) {
/*
* CLSET_RETRIES is set to 2, so this should be half
* of the total timeout required.
*/
timeo = nmp->nm_retry * nmp->nm_timeo / 2;
if (timeo < 1)
timeo = 1;
timo.tv_sec = timeo / NFS_HZ;
timo.tv_usec = (timeo % NFS_HZ) * 1000000 / NFS_HZ;
} else {
/* For UDP hard mounts, use a large value. */
timo.tv_sec = NFS_MAXTIMEO / NFS_HZ;
timo.tv_usec = 0;
}
} else {
timo.tv_sec = nmp->nm_timeo / NFS_HZ;
timo.tv_usec = (nmp->nm_timeo % NFS_HZ) * 1000000 / NFS_HZ;
}
mrep = NULL;
stat = CLNT_CALL_MBUF(nmp->nm_client, &ext,
(nmp->nm_flag & NFSMNT_NFSV3) ? procnum : nfsv2_procid[procnum],
mreq, &mrep, timo);
/*
* If there was a successful reply and a tprintf msg.
* tprintf a response.
*/
if (stat == RPC_SUCCESS)
error = 0;
else if (stat == RPC_TIMEDOUT) {
nfsstats.rpctimeouts++;
error = ETIMEDOUT;
} else if (stat == RPC_VERSMISMATCH) {
nfsstats.rpcinvalid++;
error = EOPNOTSUPP;
} else if (stat == RPC_PROGVERSMISMATCH) {
nfsstats.rpcinvalid++;
error = EPROTONOSUPPORT;
} else if (stat == RPC_INTR) {
error = EINTR;
} else {
nfsstats.rpcinvalid++;
error = EACCES;
}
if (error)
goto nfsmout;
KASSERT(mrep != NULL, ("mrep shouldn't be NULL if no error\n"));
/*
* Search for any mbufs that are not a multiple of 4 bytes long
* or with m_data not longword aligned.
* These could cause pointer alignment problems, so copy them to
* well aligned mbufs.
*/
error = nfs_realign(&mrep, M_NOWAIT);
if (error == ENOMEM) {
m_freem(mrep);
AUTH_DESTROY(auth);
nfsstats.rpcinvalid++;
return (error);
}
md = mrep;
dpos = mtod(mrep, caddr_t);
tl = nfsm_dissect(u_int32_t *, NFSX_UNSIGNED);
if (*tl != 0) {
error = fxdr_unsigned(int, *tl);
if ((nmp->nm_flag & NFSMNT_NFSV3) &&
error == NFSERR_TRYLATER) {
m_freem(mrep);
error = 0;
waituntil = time_second + nfs3_jukebox_delay;
while (time_second < waituntil)
(void)tsleep(&fake_wchan, PSOCK, "nqnfstry",
hz);
goto tryagain;
}
/*
* Make sure NFSERR_RETERR isn't bogusly set by a server
* such as amd. (No actual NFS error has bit 31 set.)
*/
error &= ~NFSERR_RETERR;
/*
* If the File Handle was stale, invalidate the lookup
* cache, just in case.
*/
if (error == ESTALE)
nfs_purgecache(vp);
/*
* Skip wcc data on non-ENOENT NFS errors for now.
* NetApp filers return corrupt postop attrs in the
* wcc data for NFS err EROFS. Not sure if they could
* return corrupt postop attrs for others errors.
* Blocking ENOENT post-op attributes breaks negative
* name caching, so always allow it through.
*/
if ((nmp->nm_flag & NFSMNT_NFSV3) &&
(!nfs_skip_wcc_data_onerr || error == ENOENT)) {
*mrp = mrep;
*mdp = md;
*dposp = dpos;
error |= NFSERR_RETERR;
} else
m_freem(mrep);
goto nfsmout;
}