/*
* Asynchronous I/O daemons for client nfs.
* They do read-ahead and write-behind operations on the block I/O cache.
* Returns if we hit the timeout defined by the iodmaxidle sysctl.
*/
static void
nfssvc_iod(void *instance)
{
struct buf *bp;
struct nfsmount *nmp;
int myiod, timo;
int error = 0;
mtx_lock(&nfs_iod_mtx);
myiod = (int *)instance - nfs_asyncdaemon;
/*
* Main loop
*/
for (;;) {
while (((nmp = nfs_iodmount[myiod]) == NULL)
|| !TAILQ_FIRST(&nmp->nm_bufq)) {
if (myiod >= nfs_iodmax)
goto finish;
if (nmp)
nmp->nm_bufqiods--;
if (nfs_iodwant[myiod] == NFSIOD_NOT_AVAILABLE)
nfs_iodwant[myiod] = NFSIOD_AVAILABLE;
nfs_iodmount[myiod] = NULL;
/*
* Always keep at least nfs_iodmin kthreads.
*/
timo = (myiod < nfs_iodmin) ? 0 : nfs_iodmaxidle * hz;
error = msleep(&nfs_iodwant[myiod], &nfs_iod_mtx, PWAIT | PCATCH,
"-", timo);
if (error) {
nmp = nfs_iodmount[myiod];
/*
* Rechecking the nm_bufq closes a rare race where the
* nfsiod is woken up at the exact time the idle timeout
* fires
*/
if (nmp && TAILQ_FIRST(&nmp->nm_bufq))
error = 0;
break;
}
}
if (error)
break;
while ((bp = TAILQ_FIRST(&nmp->nm_bufq)) != NULL) {
int giant_locked = 0;
/* Take one off the front of the list */
TAILQ_REMOVE(&nmp->nm_bufq, bp, b_freelist);
nmp->nm_bufqlen--;
if (nmp->nm_bufqwant && nmp->nm_bufqlen <= nfs_numasync) {
nmp->nm_bufqwant = 0;
wakeup(&nmp->nm_bufq);
}
mtx_unlock(&nfs_iod_mtx);
if (NFS_ISV4(bp->b_vp)) {
giant_locked = 1;
mtx_lock(&Giant);
}
if (bp->b_flags & B_DIRECT) {
KASSERT((bp->b_iocmd == BIO_WRITE), ("nfscvs_iod: BIO_WRITE not set"));
(void)nfs_doio_directwrite(bp);
} else {
if (bp->b_iocmd == BIO_READ)
(void) nfs_doio(bp->b_vp, bp, bp->b_rcred, NULL);
else
(void) nfs_doio(bp->b_vp, bp, bp->b_wcred, NULL);
}
if (giant_locked)
mtx_unlock(&Giant);
mtx_lock(&nfs_iod_mtx);
/*
* If there are more than one iod on this mount, then defect
* so that the iods can be shared out fairly between the mounts
*/
if (nfs_defect && nmp->nm_bufqiods > 1) {
NFS_DPF(ASYNCIO,
("nfssvc_iod: iod %d defecting from mount %p\n",
myiod, nmp));
nfs_iodmount[myiod] = NULL;
nmp->nm_bufqiods--;
break;
}
}
}
finish:
nfs_asyncdaemon[myiod] = 0;
if (nmp)
nmp->nm_bufqiods--;
nfs_iodwant[myiod] = NFSIOD_NOT_AVAILABLE;
nfs_iodmount[myiod] = NULL;
/* Someone may be waiting for the last nfsiod to terminate. */
if (--nfs_numasync == 0)
wakeup(&nfs_numasync);
mtx_unlock(&nfs_iod_mtx);
if ((error == 0) || (error == EWOULDBLOCK))
kproc_exit(0);
/* Abnormal termination */
kproc_exit(1);
}
/*
* Look for the request in the cache
* If found then
* return action and optionally reply
* else
* insert it in the cache
*
* The rules are as follows:
* - if in progress, return DROP request
* - if completed within DELAY of the current time, return DROP it
* - if completed a longer time ago return REPLY if the reply was cached or
* return DOIT
* Update/add new request at end of lru list
*/
int
nfsrv_getcache(struct nfsrv_descript *nd, struct nfssvc_sock *slp,
struct mbuf **repp)
{
struct nfsrvcache *rp;
struct mbuf *mb;
struct sockaddr_in *saddr;
caddr_t bpos;
int ret;
/*
* Don't cache recent requests for reliable transport protocols.
* (Maybe we should for the case of a reconnect, but..)
*/
if (!nd->nd_nam2)
return (RC_DOIT);
lwkt_gettoken(&srvcache_token);
loop:
for (rp = NFSRCHASH(nd->nd_retxid)->lh_first; rp != NULL;
rp = rp->rc_hash.le_next) {
if (nd->nd_retxid == rp->rc_xid && nd->nd_procnum == rp->rc_proc &&
netaddr_match(NETFAMILY(rp), &rp->rc_haddr, nd->nd_nam)) {
NFS_DPF(RC, ("H%03x", rp->rc_xid & 0xfff));
if ((rp->rc_flag & RC_LOCKED) != 0) {
rp->rc_flag |= RC_WANTED;
tsleep((caddr_t)rp, 0, "nfsrc", 0);
goto loop;
}
rp->rc_flag |= RC_LOCKED;
/* If not at end of LRU chain, move it there */
if (TAILQ_NEXT(rp, rc_lru) != NULL) {
TAILQ_REMOVE(&nfsrvlruhead, rp, rc_lru);
TAILQ_INSERT_TAIL(&nfsrvlruhead, rp, rc_lru);
}
if (rp->rc_state == RC_UNUSED)
panic("nfsrv cache");
if (rp->rc_state == RC_INPROG) {
nfsstats.srvcache_inproghits++;
ret = RC_DROPIT;
} else if (rp->rc_flag & RC_REPSTATUS) {
nfsstats.srvcache_nonidemdonehits++;
nfs_rephead(0, nd, slp, rp->rc_status,
repp, &mb, &bpos);
ret = RC_REPLY;
} else if (rp->rc_flag & RC_REPMBUF) {
nfsstats.srvcache_nonidemdonehits++;
*repp = m_copym(rp->rc_reply, 0, M_COPYALL,
MB_WAIT);
ret = RC_REPLY;
} else {
nfsstats.srvcache_idemdonehits++;
rp->rc_state = RC_INPROG;
ret = RC_DOIT;
}
rp->rc_flag &= ~RC_LOCKED;
if (rp->rc_flag & RC_WANTED) {
rp->rc_flag &= ~RC_WANTED;
wakeup((caddr_t)rp);
}
lwkt_reltoken(&srvcache_token);
return (ret);
}
}
nfsstats.srvcache_misses++;
NFS_DPF(RC, ("M%03x", nd->nd_retxid & 0xfff));
if (numnfsrvcache < desirednfsrvcache) {
rp = kmalloc((u_long)sizeof *rp, M_NFSD, M_WAITOK | M_ZERO);
numnfsrvcache++;
rp->rc_flag = RC_LOCKED;
} else {
rp = TAILQ_FIRST(&nfsrvlruhead);
while ((rp->rc_flag & RC_LOCKED) != 0) {
rp->rc_flag |= RC_WANTED;
tsleep((caddr_t)rp, 0, "nfsrc", 0);
rp = TAILQ_FIRST(&nfsrvlruhead);
}
rp->rc_flag |= RC_LOCKED;
LIST_REMOVE(rp, rc_hash);
TAILQ_REMOVE(&nfsrvlruhead, rp, rc_lru);
if (rp->rc_flag & RC_REPMBUF) {
m_freem(rp->rc_reply);
rp->rc_reply = NULL;
rp->rc_flag &= ~RC_REPMBUF;
}
if (rp->rc_flag & RC_NAM) {
kfree(rp->rc_nam, M_SONAME);
rp->rc_nam = NULL;
rp->rc_flag &= ~RC_NAM;
}
}
TAILQ_INSERT_TAIL(&nfsrvlruhead, rp, rc_lru);
rp->rc_state = RC_INPROG;
rp->rc_xid = nd->nd_retxid;
saddr = (struct sockaddr_in *)nd->nd_nam;
switch (saddr->sin_family) {
case AF_INET:
rp->rc_flag |= RC_INETADDR;
rp->rc_inetaddr = saddr->sin_addr.s_addr;
break;
case AF_ISO:
default:
rp->rc_flag |= RC_NAM;
rp->rc_nam = dup_sockaddr(nd->nd_nam);
break;
};
rp->rc_proc = nd->nd_procnum;
LIST_INSERT_HEAD(NFSRCHASH(nd->nd_retxid), rp, rc_hash);
rp->rc_flag &= ~RC_LOCKED;
if (rp->rc_flag & RC_WANTED) {
rp->rc_flag &= ~RC_WANTED;
wakeup((caddr_t)rp);
}
lwkt_reltoken(&srvcache_token);
return (RC_DOIT);
}
/*
* Update a request cache entry after the rpc has been done
*/
void
nfsrv_updatecache(struct nfsrv_descript *nd, int repvalid, struct mbuf *repmbuf)
{
struct nfsrvcache *rp;
if (!nd->nd_nam2)
return;
lwkt_gettoken(&srvcache_token);
loop:
for (rp = NFSRCHASH(nd->nd_retxid)->lh_first; rp != NULL;
rp = rp->rc_hash.le_next) {
if (nd->nd_retxid == rp->rc_xid && nd->nd_procnum == rp->rc_proc &&
netaddr_match(NETFAMILY(rp), &rp->rc_haddr, nd->nd_nam)) {
NFS_DPF(RC, ("U%03x", rp->rc_xid & 0xfff));
if ((rp->rc_flag & RC_LOCKED) != 0) {
rp->rc_flag |= RC_WANTED;
tsleep((caddr_t)rp, 0, "nfsrc", 0);
goto loop;
}
rp->rc_flag |= RC_LOCKED;
if (rp->rc_state == RC_DONE) {
/*
* This can occur if the cache is too small.
* Retransmits of the same request aren't
* dropped so we may see the operation
* complete more then once.
*/
if (rp->rc_flag & RC_REPMBUF) {
m_freem(rp->rc_reply);
rp->rc_reply = NULL;
rp->rc_flag &= ~RC_REPMBUF;
}
}
rp->rc_state = RC_DONE;
/*
* If we have a valid reply update status and save
* the reply for non-idempotent rpc's.
*/
if (repvalid && nonidempotent[nd->nd_procnum]) {
if ((nd->nd_flag & ND_NFSV3) == 0 &&
nfsv2_repstat[nfsv2_procid[nd->nd_procnum]]) {
rp->rc_status = nd->nd_repstat;
rp->rc_flag |= RC_REPSTATUS;
} else {
if (rp->rc_flag & RC_REPMBUF) {
m_freem(rp->rc_reply);
rp->rc_reply = NULL;
rp->rc_flag &= ~RC_REPMBUF;
}
rp->rc_reply = m_copym(repmbuf, 0,
M_COPYALL, MB_WAIT);
rp->rc_flag |= RC_REPMBUF;
}
}
rp->rc_flag &= ~RC_LOCKED;
if (rp->rc_flag & RC_WANTED) {
rp->rc_flag &= ~RC_WANTED;
wakeup((caddr_t)rp);
}
break;
}
}
lwkt_reltoken(&srvcache_token);
NFS_DPF(RC, ("L%03x", nd->nd_retxid & 0xfff));
}
/*
* Asynchronous I/O daemons for client nfs.
* They do read-ahead and write-behind operations on the block I/O cache.
* Returns if we hit the timeout defined by the iodmaxidle sysctl.
*/
static void
nfssvc_iod(void *instance)
{
struct buf *bp;
struct nfsmount *nmp;
int myiod, timo;
int error = 0;
mtx_lock(&Giant);
myiod = (int *)instance - nfs_asyncdaemon;
/*
* Main loop
*/
for (;;) {
while (((nmp = nfs_iodmount[myiod]) == NULL
|| !TAILQ_FIRST(&nmp->nm_bufq))
&& error == 0) {
if (myiod >= nfs_iodmax)
goto finish;
if (nmp)
nmp->nm_bufqiods--;
nfs_iodwant[myiod] = curthread->td_proc;
nfs_iodmount[myiod] = NULL;
/*
* Always keep at least nfs_iodmin kthreads.
*/
timo = (myiod < nfs_iodmin) ? 0 : nfs_iodmaxidle * hz;
error = tsleep((caddr_t)&nfs_iodwant[myiod], PWAIT | PCATCH,
"nfsidl", timo);
}
if (error)
break;
while ((bp = TAILQ_FIRST(&nmp->nm_bufq)) != NULL) {
/* Take one off the front of the list */
TAILQ_REMOVE(&nmp->nm_bufq, bp, b_freelist);
nmp->nm_bufqlen--;
if (nmp->nm_bufqwant && nmp->nm_bufqlen <= nfs_numasync) {
nmp->nm_bufqwant = 0;
wakeup(&nmp->nm_bufq);
}
if (bp->b_iocmd == BIO_READ)
(void) nfs_doio(bp, bp->b_rcred, NULL);
else
(void) nfs_doio(bp, bp->b_wcred, NULL);
/*
* If there are more than one iod on this mount, then defect
* so that the iods can be shared out fairly between the mounts
*/
if (nfs_defect && nmp->nm_bufqiods > 1) {
NFS_DPF(ASYNCIO,
("nfssvc_iod: iod %d defecting from mount %p\n",
myiod, nmp));
nfs_iodmount[myiod] = NULL;
nmp->nm_bufqiods--;
break;
}
}
}
finish:
nfs_asyncdaemon[myiod] = 0;
if (nmp)
nmp->nm_bufqiods--;
nfs_iodwant[myiod] = NULL;
nfs_iodmount[myiod] = NULL;
nfs_numasync--;
if ((error == 0) || (error == EWOULDBLOCK))
kthread_exit(0);
/* Abnormal termination */
kthread_exit(1);
}
