bool
authgss_ctx_hash_set(struct svc_rpc_gss_data *gd)
{
struct rbtree_x_part *t;
struct authgss_x_part *axp;
gss_union_ctx_id_desc *gss_ctx;
bool rslt;
cond_init_authgss_hash();
gss_ctx = (gss_union_ctx_id_desc *) (gd->ctx);
gd->hk.k = gss_ctx_hash(gss_ctx);
++(gd->refcnt); /* locked */
t = rbtx_partition_of_scalar(&authgss_hash_st.xt, gd->hk.k);
mutex_lock(&t->mtx);
rslt =
rbtree_x_cached_insert(&authgss_hash_st.xt, t, &gd->node_k,
gd->hk.k);
/* lru */
axp = (struct authgss_x_part *)t->u1;
TAILQ_INSERT_TAIL(&axp->lru_q, gd, lru_q);
mutex_unlock(&t->mtx);
/* global size */
(void)atomic_inc_uint32_t(&authgss_hash_st.size);
return (rslt);
}
/**
* @brief Check for expired TCP DRCs.
*/
static inline void drc_free_expired(void)
{
drc_t *drc;
time_t now = time(NULL);
struct rbtree_x_part *t;
struct opr_rbtree_node *odrc = NULL;
DRC_ST_LOCK();
if ((drc_st->tcp_drc_recycle_qlen < 1) ||
(now - drc_st->last_expire_check) < 600) /* 10m */
goto unlock;
do {
drc = TAILQ_FIRST(&drc_st->tcp_drc_recycle_q);
if (drc && (drc->d_u.tcp.recycle_time > 0)
&& ((now - drc->d_u.tcp.recycle_time) >
drc_st->expire_delta) && (drc->refcnt == 0)) {
LogFullDebug(COMPONENT_DUPREQ,
"remove expired drc %p from recycle queue",
drc);
t = rbtx_partition_of_scalar(&drc_st->tcp_drc_recycle_t,
drc->d_u.tcp.hk);
odrc =
opr_rbtree_lookup(&t->t, &drc->d_u.tcp.recycle_k);
if (!odrc) {
LogCrit(COMPONENT_DUPREQ,
"BUG: asked to dequeue DRC not on queue");
} else {
(void)opr_rbtree_remove(&t->t,
&drc->d_u.tcp.
recycle_k);
}
TAILQ_REMOVE(&drc_st->tcp_drc_recycle_q, drc,
d_u.tcp.recycle_q);
--(drc_st->tcp_drc_recycle_qlen);
/* expect DRC to be reachable from some xprt(s) */
PTHREAD_MUTEX_lock(&drc->mtx);
drc->flags &= ~DRC_FLAG_RECYCLE;
/* but if not, dispose it */
if (drc->refcnt == 0) {
PTHREAD_MUTEX_unlock(&drc->mtx);
free_tcp_drc(drc);
continue;
}
PTHREAD_MUTEX_unlock(&drc->mtx);
} else {
LogFullDebug(COMPONENT_DUPREQ,
"unexpired drc %p in recycle queue expire check (nothing happens)",
drc);
drc_st->last_expire_check = now;
break;
}
} while (1);
unlock:
DRC_ST_UNLOCK();
}
struct svc_rpc_gss_data *
authgss_ctx_hash_get(struct rpc_gss_cred *gc)
{
struct svc_rpc_gss_data gk, *gd = NULL;
gss_union_ctx_id_desc *gss_ctx;
struct opr_rbtree_node *ngd;
struct authgss_x_part *axp;
struct rbtree_x_part *t;
cond_init_authgss_hash();
gss_ctx = (gss_union_ctx_id_desc *) (gc->gc_ctx.value);
gk.hk.k = gss_ctx_hash(gss_ctx);
t = rbtx_partition_of_scalar(&authgss_hash_st.xt, gk.hk.k);
mutex_lock(&t->mtx);
ngd =
rbtree_x_cached_lookup(&authgss_hash_st.xt, t, &gk.node_k, gk.hk.k);
if (ngd) {
gd = opr_containerof(ngd, struct svc_rpc_gss_data, node_k);
/* lru adjust */
axp = (struct authgss_x_part *)t->u1;
TAILQ_REMOVE(&axp->lru_q, gd, lru_q);
TAILQ_INSERT_TAIL(&axp->lru_q, gd, lru_q);
++(axp->gen);
(void)atomic_inc_uint32_t(&gd->refcnt);
(void)atomic_inc_uint32_t(&gd->gen);
}
mutex_unlock(&t->mtx);
return (gd);
}
struct rpc_dplx_rec *
rpc_dplx_lookup_rec(int fd, uint32_t iflags, uint32_t *oflags)
{
struct rbtree_x_part *t;
struct rpc_dplx_rec rk, *rec = NULL;
struct opr_rbtree_node *nv;
cond_init_rpc_dplx();
rk.fd_k = fd;
t = rbtx_partition_of_scalar(&(rpc_dplx_rec_set.xt), fd);
rwlock_rdlock(&t->lock);
nv = opr_rbtree_lookup(&t->t, &rk.node_k);
/* XXX rework lock+insert case, so that new entries are inserted
* locked, and t->lock critical section is reduced */
if (! nv) {
rwlock_unlock(&t->lock);
rwlock_wrlock(&t->lock);
nv = opr_rbtree_lookup(&t->t, &rk.node_k);
if (! nv) {
rec = alloc_dplx_rec();
if (! rec) {
__warnx(TIRPC_DEBUG_FLAG_LOCK,
"%s: failed allocating rpc_dplx_rec", __func__);
goto unlock;
}
/* tell the caller */
*oflags = RPC_DPLX_LKP_OFLAG_ALLOC;
rec->fd_k = fd;
if (opr_rbtree_insert(&t->t, &rec->node_k)) {
/* cant happen */
__warnx(TIRPC_DEBUG_FLAG_LOCK,
"%s: collision inserting in locked rbtree partition",
__func__);
free_dplx_rec(rec);
}
}
}
else {
rec = opr_containerof(nv, struct rpc_dplx_rec, node_k);
*oflags = RPC_DPLX_LKP_FLAG_NONE;
}
rpc_dplx_ref(rec, (iflags & RPC_DPLX_LKP_IFLAG_LOCKREC) ?
RPC_DPLX_FLAG_LOCK :
RPC_DPLX_FLAG_NONE);
unlock:
rwlock_unlock(&t->lock);
return (rec);
}
/**
*
* @brief Remove an entry (request) from a duplicate request cache.
*
* The expected pattern is that nfs_rpc_execute shall delete requests only
* in error conditions. The refcnt of the corresponding duplicate request
* entry is unchanged (ie., the caller must still call nfs_dupreq_rele).
*
* We assert req->rq_u1 now points to the corresonding duplicate request
* cache entry.
*
* @param[in] req The svc_req structure.
*
* @return DUPREQ_SUCCESS if successful.
*
*/
dupreq_status_t nfs_dupreq_delete(struct svc_req *req)
{
dupreq_entry_t *dv = (dupreq_entry_t *)req->rq_u1;
dupreq_status_t status = DUPREQ_SUCCESS;
struct rbtree_x_part *t;
drc_t *drc;
/* do nothing if req is marked no-cache */
if (dv == (void *)DUPREQ_NOCACHE)
goto out;
/* do nothing if nfs_dupreq_start failed completely */
if (dv == (void *)DUPREQ_BAD_ADDR1)
goto out;
PTHREAD_MUTEX_lock(&dv->mtx);
drc = dv->hin.drc;
dv->state = DUPREQ_DELETED;
PTHREAD_MUTEX_unlock(&dv->mtx);
LogFullDebug(COMPONENT_DUPREQ,
"deleting dv=%p xid=%u on DRC=%p state=%s, status=%s, refcnt=%d",
dv, dv->hin.tcp.rq_xid, drc,
dupreq_state_table[dv->state], dupreq_status_table[status],
dv->refcnt);
/* XXX dv holds a ref on drc */
t = rbtx_partition_of_scalar(&drc->xt, dv->hk);
PTHREAD_MUTEX_lock(&t->mtx);
rbtree_x_cached_remove(&drc->xt, t, &dv->rbt_k, dv->hk);
PTHREAD_MUTEX_unlock(&t->mtx);
PTHREAD_MUTEX_lock(&drc->mtx);
if (TAILQ_IS_ENQUEUED(dv, fifo_q))
TAILQ_REMOVE(&drc->dupreq_q, dv, fifo_q);
--(drc->size);
/* release dv's ref and unlock */
nfs_dupreq_put_drc(req->rq_xprt, drc, DRC_FLAG_LOCKED);
/* !LOCKED */
out:
return status;
}
int32_t
rpc_dplx_unref(struct rpc_dplx_rec *rec, u_int flags)
{
struct rbtree_x_part *t;
struct opr_rbtree_node *nv;
int32_t refcnt;
if (! (flags & RPC_DPLX_FLAG_LOCKED))
REC_LOCK(rec);
refcnt = --(rec->refcnt);
__warnx(TIRPC_DEBUG_FLAG_REFCNT,
"%s: postunref %p rec->refcnt %u",
__func__, rec, refcnt);
if (rec->refcnt == 0) {
t = rbtx_partition_of_scalar(&rpc_dplx_rec_set.xt, rec->fd_k);
REC_UNLOCK(rec);
rwlock_wrlock(&t->lock);
nv = opr_rbtree_lookup(&t->t, &rec->node_k);
rec = NULL;
if (nv) {
rec = opr_containerof(nv, struct rpc_dplx_rec, node_k);
REC_LOCK(rec);
if (rec->refcnt == 0) {
(void) opr_rbtree_remove(&t->t, &rec->node_k);
REC_UNLOCK(rec);
__warnx(TIRPC_DEBUG_FLAG_REFCNT,
"%s: free rec %p rec->refcnt %u",
__func__, rec, refcnt);
free_dplx_rec(rec);
rec = NULL;
} else {
refcnt = rec->refcnt;
}
}
rwlock_unlock(&t->lock);
}
bool
authgss_ctx_hash_del(struct svc_rpc_gss_data *gd)
{
struct rbtree_x_part *t;
struct authgss_x_part *axp;
cond_init_authgss_hash();
t = rbtx_partition_of_scalar(&authgss_hash_st.xt, gd->hk.k);
mutex_lock(&t->mtx);
rbtree_x_cached_remove(&authgss_hash_st.xt, t, &gd->node_k, gd->hk.k);
axp = (struct authgss_x_part *)t->u1;
TAILQ_REMOVE(&axp->lru_q, gd, lru_q);
mutex_unlock(&t->mtx);
/* global size */
(void)atomic_dec_uint32_t(&authgss_hash_st.size);
/* release gd */
unref_svc_rpc_gss_data(gd, SVC_RPC_GSS_FLAG_NONE);
return (true);
}
/**
* @brief Start a duplicate request transaction
*
* Finds any matching request entry in the cache, if one exists, else
* creates one in the START state. On any non-error return, the refcnt
* of the corresponding entry is incremented.
*
* @param[in] reqnfs The NFS request data
* @param[in] req The request to be cached
*
* @retval DUPREQ_SUCCESS if successful.
* @retval DUPREQ_INSERT_MALLOC_ERROR if an error occured during insertion.
*/
dupreq_status_t nfs_dupreq_start(nfs_request_t *reqnfs,
struct svc_req *req)
{
dupreq_status_t status = DUPREQ_SUCCESS;
dupreq_entry_t *dv, *dk = NULL;
bool release_dk = true;
nfs_res_t *res = NULL;
drc_t *drc;
/* Disabled? */
if (nfs_param.core_param.drc.disabled) {
req->rq_u1 = (void *)DUPREQ_NOCACHE;
res = alloc_nfs_res();
goto out;
}
req->rq_u1 = (void *)DUPREQ_BAD_ADDR1;
req->rq_u2 = (void *)DUPREQ_BAD_ADDR1;
drc = nfs_dupreq_get_drc(req);
if (!drc) {
status = DUPREQ_INSERT_MALLOC_ERROR;
goto out;
}
switch (drc->type) {
case DRC_TCP_V4:
if (reqnfs->funcdesc->service_function == nfs4_Compound) {
if (!nfs_dupreq_v4_cacheable(reqnfs)) {
/* for such requests, we merely thread
* the request through for later
* cleanup--all v41 caching is handled
* by the v41 slot reply cache */
req->rq_u1 = (void *)DUPREQ_NOCACHE;
res = alloc_nfs_res();
goto out;
}
}
break;
default:
/* likewise for other protocol requests we may not or choose not
* to cache */
if (!(reqnfs->funcdesc->dispatch_behaviour & CAN_BE_DUP)) {
req->rq_u1 = (void *)DUPREQ_NOCACHE;
res = alloc_nfs_res();
goto out;
}
break;
}
dk = alloc_dupreq();
if (dk == NULL) {
release_dk = false;
status = DUPREQ_ERROR;
goto release_dk;
}
dk->hin.drc = drc; /* trans. call path ref to dv */
switch (drc->type) {
case DRC_TCP_V4:
case DRC_TCP_V3:
dk->hin.tcp.rq_xid = req->rq_xid;
/* XXX needed? */
dk->hin.rq_prog = req->rq_prog;
dk->hin.rq_vers = req->rq_vers;
dk->hin.rq_proc = req->rq_proc;
break;
case DRC_UDP_V234:
dk->hin.tcp.rq_xid = req->rq_xid;
if (unlikely(!copy_xprt_addr(&dk->hin.addr, req->rq_xprt))) {
status = DUPREQ_INSERT_MALLOC_ERROR;
goto release_dk;
}
dk->hin.rq_prog = req->rq_prog;
dk->hin.rq_vers = req->rq_vers;
dk->hin.rq_proc = req->rq_proc;
break;
default:
/* error */
status = DUPREQ_ERROR;
goto release_dk;
}
/* TI-RPC computed checksum */
dk->hk = req->rq_cksum;
dk->state = DUPREQ_START;
dk->timestamp = time(NULL);
{
struct opr_rbtree_node *nv;
struct rbtree_x_part *t =
rbtx_partition_of_scalar(&drc->xt, dk->hk);
PTHREAD_MUTEX_lock(&t->mtx); /* partition lock */
nv = rbtree_x_cached_lookup(&drc->xt, t, &dk->rbt_k, dk->hk);
if (nv) {
/* cached request */
dv = opr_containerof(nv, dupreq_entry_t, rbt_k);
PTHREAD_MUTEX_lock(&dv->mtx);
if (unlikely(dv->state == DUPREQ_START)) {
status = DUPREQ_BEING_PROCESSED;
} else {
/* satisfy req from the DRC, incref,
extend window */
res = dv->res;
PTHREAD_MUTEX_lock(&drc->mtx);
drc_inc_retwnd(drc);
PTHREAD_MUTEX_unlock(&drc->mtx);
status = DUPREQ_EXISTS;
(dv->refcnt)++;
}
LogDebug(COMPONENT_DUPREQ,
"dupreq hit dk=%p, dk xid=%u cksum %" PRIu64
" state=%s", dk, dk->hin.tcp.rq_xid, dk->hk,
dupreq_state_table[dk->state]);
req->rq_u1 = dv;
PTHREAD_MUTEX_unlock(&dv->mtx);
} else {
/* new request */
res = req->rq_u2 = dk->res = alloc_nfs_res();
(void)rbtree_x_cached_insert(&drc->xt, t, &dk->rbt_k,
dk->hk);
(dk->refcnt)++;
/* add to q tail */
PTHREAD_MUTEX_lock(&drc->mtx);
TAILQ_INSERT_TAIL(&drc->dupreq_q, dk, fifo_q);
++(drc->size);
PTHREAD_MUTEX_unlock(&drc->mtx);
req->rq_u1 = dk;
release_dk = false;
dv = dk;
}
PTHREAD_MUTEX_unlock(&t->mtx);
}
LogFullDebug(COMPONENT_DUPREQ,
"starting dv=%p xid=%u on DRC=%p state=%s, status=%s, refcnt=%d",
dv, dk->hin.tcp.rq_xid, drc,
dupreq_state_table[dv->state], dupreq_status_table[status],
dv->refcnt);
release_dk:
if (release_dk)
nfs_dupreq_free_dupreq(dk);
nfs_dupreq_put_drc(req->rq_xprt, drc, DRC_FLAG_NONE); /* dk ref */
out:
if (res)
reqnfs->res_nfs = req->rq_u2 = res;
return status;
}
/**
* @brief Find and reference a DRC to process the supplied svc_req.
*
* @param[in] req The svc_req being processed.
*
* @return The ref'd DRC if sucessfully located, else NULL.
*/
static /* inline */ drc_t *
nfs_dupreq_get_drc(struct svc_req *req)
{
enum drc_type dtype = get_drc_type(req);
drc_t *drc = NULL;
bool drc_check_expired = false;
switch (dtype) {
case DRC_UDP_V234:
LogFullDebug(COMPONENT_DUPREQ, "ref shared UDP DRC");
drc = &(drc_st->udp_drc);
DRC_ST_LOCK();
(void)nfs_dupreq_ref_drc(drc);
DRC_ST_UNLOCK();
goto out;
case DRC_TCP_V4:
case DRC_TCP_V3:
/* Idempotent address, no need for lock;
* xprt will be valid as long as svc_req.
*/
drc = (drc_t *)req->rq_xprt->xp_u2;
if (drc) {
/* found, no danger of removal */
LogFullDebug(COMPONENT_DUPREQ, "ref DRC=%p for xprt=%p",
drc, req->rq_xprt);
PTHREAD_MUTEX_lock(&drc->mtx); /* LOCKED */
} else {
drc_t drc_k;
struct rbtree_x_part *t = NULL;
struct opr_rbtree_node *ndrc = NULL;
drc_t *tdrc = NULL;
memset(&drc_k, 0, sizeof(drc_k));
drc_k.type = dtype;
/* Since the drc can last longer than the xprt,
* copy the address. Read operation of constant data,
* no xprt lock required.
*/
(void)copy_xprt_addr(&drc_k.d_u.tcp.addr, req->rq_xprt);
drc_k.d_u.tcp.hk =
CityHash64WithSeed((char *)&drc_k.d_u.tcp.addr,
sizeof(sockaddr_t), 911);
{
char str[SOCK_NAME_MAX];
sprint_sockaddr(&drc_k.d_u.tcp.addr,
str, sizeof(str));
LogFullDebug(COMPONENT_DUPREQ,
"get drc for addr: %s", str);
}
t = rbtx_partition_of_scalar(&drc_st->tcp_drc_recycle_t,
drc_k.d_u.tcp.hk);
DRC_ST_LOCK();
ndrc =
opr_rbtree_lookup(&t->t, &drc_k.d_u.tcp.recycle_k);
if (ndrc) {
/* reuse old DRC */
tdrc =
opr_containerof(ndrc, drc_t,
d_u.tcp.recycle_k);
PTHREAD_MUTEX_lock(&tdrc->mtx); /* LOCKED */
if (tdrc->flags & DRC_FLAG_RECYCLE) {
TAILQ_REMOVE(&drc_st->tcp_drc_recycle_q,
tdrc, d_u.tcp.recycle_q);
--(drc_st->tcp_drc_recycle_qlen);
tdrc->flags &= ~DRC_FLAG_RECYCLE;
}
drc = tdrc;
LogFullDebug(COMPONENT_DUPREQ,
"recycle TCP DRC=%p for xprt=%p",
tdrc, req->rq_xprt);
}
if (!drc) {
drc = alloc_tcp_drc(dtype);
LogFullDebug(COMPONENT_DUPREQ,
"alloc new TCP DRC=%p for xprt=%p",
drc, req->rq_xprt);
/* assign addr */
memcpy(&drc->d_u.tcp.addr, &drc_k.d_u.tcp.addr,
sizeof(sockaddr_t));
/* assign already-computed hash */
drc->d_u.tcp.hk = drc_k.d_u.tcp.hk;
PTHREAD_MUTEX_lock(&drc->mtx); /* LOCKED */
/* xprt ref */
drc->refcnt = 1;
/* insert dict */
opr_rbtree_insert(&t->t,
&drc->d_u.tcp.recycle_k);
}
DRC_ST_UNLOCK();
drc->d_u.tcp.recycle_time = 0;
(void)nfs_dupreq_ref_drc(drc); /* xprt ref */
/* try to expire unused DRCs somewhat in proportion to
* new connection arrivals */
drc_check_expired = true;
LogFullDebug(COMPONENT_DUPREQ,
"after ref drc %p refcnt==%u ", drc,
drc->refcnt);
/* Idempotent address, no need for lock;
* set once here, never changes.
* No other fields are modified.
* Assumes address stores are atomic.
*/
req->rq_xprt->xp_u2 = (void *)drc;
}
break;
default:
/* XXX error */
break;
}
/* call path ref */
(void)nfs_dupreq_ref_drc(drc);
PTHREAD_MUTEX_unlock(&drc->mtx);
if (drc_check_expired)
drc_free_expired();
out:
return drc;
}
/**
* @brief Completes a request in the cache
*
* Completes a cache insertion operation begun in nfs_dupreq_start.
* The refcnt of the corresponding duplicate request entry is unchanged
* (ie, the caller must still call nfs_dupreq_rele).
*
* In contrast with the prior DRC implementation, completing a request
* in the current implementation may under normal conditions cause one
* or more cached requests to be retired. Requests are retired in the
* order they were inserted. The primary retire algorithm is a high
* water mark, and a windowing heuristic. One or more requests will be
* retired if the water mark/timeout is exceeded, and if a no duplicate
* requests have been found in the cache in a configurable window of
* immediately preceding requests. A timeout may supplement the water mark,
* in future.
*
* req->rq_u1 has either a magic value, or points to a duplicate request
* cache entry allocated in nfs_dupreq_start.
*
* @param[in] req The request
* @param[in] res_nfs The response
*
* @return DUPREQ_SUCCESS if successful.
* @return DUPREQ_INSERT_MALLOC_ERROR if an error occured.
*/
dupreq_status_t nfs_dupreq_finish(struct svc_req *req, nfs_res_t *res_nfs)
{
dupreq_entry_t *ov = NULL, *dv = (dupreq_entry_t *)req->rq_u1;
dupreq_status_t status = DUPREQ_SUCCESS;
struct rbtree_x_part *t;
drc_t *drc = NULL;
/* do nothing if req is marked no-cache */
if (dv == (void *)DUPREQ_NOCACHE)
goto out;
/* do nothing if nfs_dupreq_start failed completely */
if (dv == (void *)DUPREQ_BAD_ADDR1)
goto out;
PTHREAD_MUTEX_lock(&dv->mtx);
dv->res = res_nfs;
dv->timestamp = time(NULL);
dv->state = DUPREQ_COMPLETE;
drc = dv->hin.drc;
PTHREAD_MUTEX_unlock(&dv->mtx);
/* cond. remove from q head */
PTHREAD_MUTEX_lock(&drc->mtx);
LogFullDebug(COMPONENT_DUPREQ,
"completing dv=%p xid=%u on DRC=%p state=%s, status=%s, refcnt=%d",
dv, dv->hin.tcp.rq_xid, drc,
dupreq_state_table[dv->state], dupreq_status_table[status],
dv->refcnt);
/* ok, do the new retwnd calculation here. then, put drc only if
* we retire an entry */
if (drc_should_retire(drc)) {
/* again: */
ov = TAILQ_FIRST(&drc->dupreq_q);
if (likely(ov)) {
/* finished request count against retwnd */
drc_dec_retwnd(drc);
/* check refcnt */
if (ov->refcnt > 0) {
/* ov still in use, apparently */
goto unlock;
}
/* remove q entry */
TAILQ_REMOVE(&drc->dupreq_q, ov, fifo_q);
--(drc->size);
/* remove dict entry */
t = rbtx_partition_of_scalar(&drc->xt, ov->hk);
/* interlock */
PTHREAD_MUTEX_unlock(&drc->mtx);
PTHREAD_MUTEX_lock(&t->mtx); /* partition lock */
rbtree_x_cached_remove(&drc->xt, t, &ov->rbt_k, ov->hk);
PTHREAD_MUTEX_unlock(&t->mtx);
LogDebug(COMPONENT_DUPREQ,
"retiring ov=%p xid=%u on DRC=%p state=%s, status=%s, refcnt=%d",
ov, ov->hin.tcp.rq_xid,
ov->hin.drc, dupreq_state_table[dv->state],
dupreq_status_table[status], ov->refcnt);
/* deep free ov */
nfs_dupreq_free_dupreq(ov);
goto out;
}
}
unlock:
PTHREAD_MUTEX_unlock(&drc->mtx);
out:
return status;
}
/**
* @brief Find and reference a DRC to process the supplied svc_req.
*
* @param[in] req The svc_req being processed.
*
* @return The ref'd DRC if sucessfully located, else NULL.
*/
static /* inline */ drc_t *
nfs_dupreq_get_drc(struct svc_req *req)
{
enum drc_type dtype = get_drc_type(req);
gsh_xprt_private_t *xu = (gsh_xprt_private_t *) req->rq_xprt->xp_u1;
drc_t *drc = NULL;
bool drc_check_expired = false;
switch (dtype) {
case DRC_UDP_V234:
LogFullDebug(COMPONENT_DUPREQ, "ref shared UDP DRC");
drc = &(drc_st->udp_drc);
DRC_ST_LOCK();
(void)nfs_dupreq_ref_drc(drc);
DRC_ST_UNLOCK();
goto out;
break;
case DRC_TCP_V4:
case DRC_TCP_V3:
pthread_mutex_lock(&req->rq_xprt->xp_lock);
if (xu->drc) {
drc = xu->drc;
LogFullDebug(COMPONENT_DUPREQ, "ref DRC=%p for xprt=%p",
drc, req->rq_xprt);
pthread_mutex_lock(&drc->mtx); /* LOCKED */
} else {
drc_t drc_k;
struct rbtree_x_part *t = NULL;
struct opr_rbtree_node *ndrc = NULL;
drc_t *tdrc = NULL;
memset(&drc_k, 0, sizeof(drc_k));
drc_k.type = dtype;
(void)copy_xprt_addr(&drc_k.d_u.tcp.addr, req->rq_xprt);
drc_k.d_u.tcp.hk =
CityHash64WithSeed((char *)&drc_k.d_u.tcp.addr,
sizeof(sockaddr_t), 911);
{
char str[512];
sprint_sockaddr(&drc_k.d_u.tcp.addr, str, 512);
LogFullDebug(COMPONENT_DUPREQ,
"get drc for addr: %s", str);
}
t = rbtx_partition_of_scalar(&drc_st->tcp_drc_recycle_t,
drc_k.d_u.tcp.hk);
DRC_ST_LOCK();
ndrc =
opr_rbtree_lookup(&t->t, &drc_k.d_u.tcp.recycle_k);
if (ndrc) {
/* reuse old DRC */
tdrc =
opr_containerof(ndrc, drc_t,
d_u.tcp.recycle_k);
pthread_mutex_lock(&tdrc->mtx); /* LOCKED */
if (tdrc->flags & DRC_FLAG_RECYCLE) {
TAILQ_REMOVE(&drc_st->tcp_drc_recycle_q,
tdrc, d_u.tcp.recycle_q);
--(drc_st->tcp_drc_recycle_qlen);
tdrc->flags &= ~DRC_FLAG_RECYCLE;
}
drc = tdrc;
LogFullDebug(COMPONENT_DUPREQ,
"recycle TCP DRC=%p for xprt=%p",
tdrc, req->rq_xprt);
}
if (!drc) {
drc = alloc_tcp_drc(dtype);
LogFullDebug(COMPONENT_DUPREQ,
"alloc new TCP DRC=%p for xprt=%p",
drc, req->rq_xprt);
/* assign addr */
memcpy(&drc->d_u.tcp.addr, &drc_k.d_u.tcp.addr,
sizeof(sockaddr_t));
/* assign already-computed hash */
drc->d_u.tcp.hk = drc_k.d_u.tcp.hk;
pthread_mutex_lock(&drc->mtx); /* LOCKED */
/* xprt ref */
drc->refcnt = 1;
/* insert dict */
opr_rbtree_insert(&t->t,
&drc->d_u.tcp.recycle_k);
}
DRC_ST_UNLOCK();
drc->d_u.tcp.recycle_time = 0;
/* xprt drc */
(void)nfs_dupreq_ref_drc(drc); /* xu ref */
/* try to expire unused DRCs somewhat in proportion to
* new connection arrivals */
drc_check_expired = true;
LogFullDebug(COMPONENT_DUPREQ,
"after ref drc %p refcnt==%u ", drc,
drc->refcnt);
xu->drc = drc;
}
pthread_mutex_unlock(&req->rq_xprt->xp_lock);
break;
default:
/* XXX error */
break;
}
/* call path ref */
(void)nfs_dupreq_ref_drc(drc);
pthread_mutex_unlock(&drc->mtx);
if (drc_check_expired)
drc_free_expired();
out:
return drc;
}
