* GPL HEADER END
*/
/*
- * Copyright (c) 2011, 2012, Whamcloud, Inc.
+ * Copyright (c) 2012, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#include <lnet/lib-lnet.h>
/* NB: add /proc interfaces in upcoming patches */
-int portal_rotor;
+int portal_rotor = LNET_PTL_ROTOR_HASH_RT;
CFS_MODULE_PARM(portal_rotor, "i", int, 0644,
"redirect PUTs to different cpu-partitions");
}
}
-struct lnet_match_table *
-lnet_mt_of_match(unsigned int index, lnet_process_id_t id, __u64 mbits)
+static struct lnet_match_table *
+lnet_mt_of_match(struct lnet_match_info *info, struct lnet_msg *msg)
{
struct lnet_match_table *mtable;
struct lnet_portal *ptl;
int nmaps;
int rotor;
+ int routed;
int cpt;
/* NB: called w/o lock */
- LASSERT(index < the_lnet.ln_nportals);
- ptl = the_lnet.ln_portals[index];
+ LASSERT(info->mi_portal < the_lnet.ln_nportals);
+ ptl = the_lnet.ln_portals[info->mi_portal];
LASSERT(lnet_ptl_is_wildcard(ptl) || lnet_ptl_is_unique(ptl));
- mtable = lnet_match2mt(ptl, id, mbits);
+ mtable = lnet_match2mt(ptl, info->mi_id, info->mi_mbits);
if (mtable != NULL)
return mtable;
/* it's a wildcard portal */
- if (!portal_rotor) {
+ routed = LNET_NIDNET(msg->msg_hdr.src_nid) !=
+ LNET_NIDNET(msg->msg_hdr.dest_nid);
+
+ if (portal_rotor == LNET_PTL_ROTOR_OFF ||
+ (portal_rotor != LNET_PTL_ROTOR_ON && !routed)) {
cpt = lnet_cpt_current();
if (ptl->ptl_mtables[cpt]->mt_enabled)
return ptl->ptl_mtables[cpt];
}
- rotor = ptl->ptl_rotor++;
- cpt = rotor % LNET_CPT_NUMBER;
+ rotor = ptl->ptl_rotor++; /* get round-robin factor */
+ if (portal_rotor == LNET_PTL_ROTOR_HASH_RT && routed)
+ cpt = lnet_cpt_of_nid(msg->msg_hdr.src_nid);
+ else
+ cpt = rotor % LNET_CPT_NUMBER;
if (!ptl->ptl_mtables[cpt]->mt_enabled) {
/* is there any active entry for this portal? */
return ptl->ptl_mtables[cpt];
}
-cfs_list_t *
+static int
+lnet_mt_test_exhausted(struct lnet_match_table *mtable, int pos)
+{
+ __u64 *bmap;
+ int i;
+
+ if (!lnet_ptl_is_wildcard(the_lnet.ln_portals[mtable->mt_portal]))
+ return 0;
+
+ if (pos < 0) { /* check all bits */
+ for (i = 0; i < LNET_MT_EXHAUSTED_BMAP; i++) {
+ if (mtable->mt_exhausted[i] != (__u64)(-1))
+ return 0;
+ }
+ return 1;
+ }
+
+ LASSERT(pos <= LNET_MT_HASH_IGNORE);
+ /* mtable::mt_mhash[pos] is marked as exhausted or not */
+ bmap = &mtable->mt_exhausted[pos >> LNET_MT_BITS_U64];
+ pos &= (1 << LNET_MT_BITS_U64) - 1;
+
+ return ((*bmap) & (1ULL << pos)) != 0;
+}
+
+static void
+lnet_mt_set_exhausted(struct lnet_match_table *mtable, int pos, int exhausted)
+{
+ __u64 *bmap;
+
+ LASSERT(lnet_ptl_is_wildcard(the_lnet.ln_portals[mtable->mt_portal]));
+ LASSERT(pos <= LNET_MT_HASH_IGNORE);
+
+ /* set mtable::mt_mhash[pos] as exhausted/non-exhausted */
+ bmap = &mtable->mt_exhausted[pos >> LNET_MT_BITS_U64];
+ pos &= (1 << LNET_MT_BITS_U64) - 1;
+
+ if (!exhausted)
+ *bmap &= ~(1ULL << pos);
+ else
+ *bmap |= 1ULL << pos;
+}
+
+struct list_head *
lnet_mt_match_head(struct lnet_match_table *mtable,
lnet_process_id_t id, __u64 mbits)
{
struct lnet_portal *ptl = the_lnet.ln_portals[mtable->mt_portal];
if (lnet_ptl_is_wildcard(ptl)) {
- return &mtable->mt_mlist;
-
+ return &mtable->mt_mhash[mbits & LNET_MT_HASH_MASK];
} else {
unsigned long hash = mbits + id.nid + id.pid;
LASSERT(lnet_ptl_is_unique(ptl));
- hash = cfs_hash_long(hash, LNET_MT_HASH_BITS);
+ hash = hash_long(hash, LNET_MT_HASH_BITS);
return &mtable->mt_mhash[hash];
}
}
lnet_mt_match_md(struct lnet_match_table *mtable,
struct lnet_match_info *info, struct lnet_msg *msg)
{
- cfs_list_t *head;
+ struct list_head *head;
lnet_me_t *me;
lnet_me_t *tmp;
int exhausted = 0;
int rc;
- /* NB: only wildcard portal can return LNET_MATCHMD_EXHAUSTED */
+ /* any ME with ignore bits? */
+ if (!list_empty(&mtable->mt_mhash[LNET_MT_HASH_IGNORE]))
+ head = &mtable->mt_mhash[LNET_MT_HASH_IGNORE];
+ else
+ head = lnet_mt_match_head(mtable, info->mi_id, info->mi_mbits);
+ again:
+ /* NB: only wildcard portal needs to return LNET_MATCHMD_EXHAUSTED */
if (lnet_ptl_is_wildcard(the_lnet.ln_portals[mtable->mt_portal]))
exhausted = LNET_MATCHMD_EXHAUSTED;
- head = lnet_mt_match_head(mtable, info->mi_id, info->mi_mbits);
-
- cfs_list_for_each_entry_safe(me, tmp, head, me_list) {
+ list_for_each_entry_safe(me, tmp, head, me_list) {
/* ME attached but MD not attached yet */
if (me->me_md == NULL)
continue;
}
}
+ if (exhausted == LNET_MATCHMD_EXHAUSTED) { /* @head is exhausted */
+ lnet_mt_set_exhausted(mtable, head - mtable->mt_mhash, 1);
+ if (!lnet_mt_test_exhausted(mtable, -1))
+ exhausted = 0;
+ }
+
+ if (exhausted == 0 && head == &mtable->mt_mhash[LNET_MT_HASH_IGNORE]) {
+ head = lnet_mt_match_head(mtable, info->mi_id, info->mi_mbits);
+ goto again; /* re-check MEs w/o ignore-bits */
+ }
+
if (info->mi_opc == LNET_MD_OP_GET ||
!lnet_ptl_is_lazy(the_lnet.ln_portals[info->mi_portal]))
return LNET_MATCHMD_DROP | exhausted;
if (lnet_ptl_is_lazy(ptl)) {
if (msg->msg_rx_ready_delay) {
msg->msg_rx_delayed = 1;
- cfs_list_add_tail(&msg->msg_list,
- &ptl->ptl_msg_delayed);
+ list_add_tail(&msg->msg_list,
+ &ptl->ptl_msg_delayed);
}
rc = LNET_MATCHMD_NONE;
} else {
lnet_ptl_lock(ptl);
if (i == 0) { /* the first try, attach on stealing list */
- cfs_list_add_tail(&msg->msg_list,
- &ptl->ptl_msg_stealing);
+ list_add_tail(&msg->msg_list,
+ &ptl->ptl_msg_stealing);
}
- if (!cfs_list_empty(&msg->msg_list)) { /* on stealing list */
+ if (!list_empty(&msg->msg_list)) { /* on stealing list */
rc = lnet_mt_match_md(mtable, info, msg);
if ((rc & LNET_MATCHMD_EXHAUSTED) != 0 &&
lnet_ptl_disable_mt(ptl, cpt);
if ((rc & LNET_MATCHMD_FINISH) != 0)
- cfs_list_del_init(&msg->msg_list);
+ list_del_init(&msg->msg_list);
} else {
/* could be matched by lnet_ptl_attach_md()
LNET_MATCHMD_DROP : LNET_MATCHMD_OK;
}
- if (!cfs_list_empty(&msg->msg_list) && /* not matched yet */
+ if (!list_empty(&msg->msg_list) && /* not matched yet */
(i == LNET_CPT_NUMBER - 1 || /* the last CPT */
ptl->ptl_mt_nmaps == 0 || /* no active CPT */
(ptl->ptl_mt_nmaps == 1 && /* the only active CPT */
ptl->ptl_mt_maps[0] == cpt))) {
/* nothing to steal, delay or drop */
- cfs_list_del_init(&msg->msg_list);
+ list_del_init(&msg->msg_list);
if (lnet_ptl_is_lazy(ptl)) {
msg->msg_rx_delayed = 1;
- cfs_list_add_tail(&msg->msg_list,
- &ptl->ptl_msg_delayed);
+ list_add_tail(&msg->msg_list,
+ &ptl->ptl_msg_delayed);
rc = LNET_MATCHMD_NONE;
} else {
rc = LNET_MATCHMD_DROP;
if (rc != 0) /* matched or delayed early message */
return rc;
- mtable = lnet_mt_of_match(info->mi_portal,
- info->mi_id, info->mi_mbits);
+ mtable = lnet_mt_of_match(info, msg);
lnet_res_lock(mtable->mt_cpt);
if (the_lnet.ln_shutdown) {
lnet_ptl_lock(ptl);
msg->msg_rx_delayed = 1;
- cfs_list_add_tail(&msg->msg_list, &ptl->ptl_msg_delayed);
+ list_add_tail(&msg->msg_list, &ptl->ptl_msg_delayed);
lnet_ptl_unlock(ptl);
lnet_res_unlock(mtable->mt_cpt);
/* called with lnet_res_lock held */
void
lnet_ptl_attach_md(lnet_me_t *me, lnet_libmd_t *md,
- cfs_list_t *matches, cfs_list_t *drops)
+ struct list_head *matches, struct list_head *drops)
{
struct lnet_portal *ptl = the_lnet.ln_portals[me->me_portal];
struct lnet_match_table *mtable;
- cfs_list_t *head;
+ struct list_head *head;
lnet_msg_t *tmp;
lnet_msg_t *msg;
int exhausted = 0;
cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
mtable = ptl->ptl_mtables[cpt];
- if (cfs_list_empty(&ptl->ptl_msg_stealing) &&
- cfs_list_empty(&ptl->ptl_msg_delayed) &&
- mtable->mt_enabled)
+ if (list_empty(&ptl->ptl_msg_stealing) &&
+ list_empty(&ptl->ptl_msg_delayed) &&
+ !lnet_mt_test_exhausted(mtable, me->me_pos))
return;
lnet_ptl_lock(ptl);
head = &ptl->ptl_msg_stealing;
again:
- cfs_list_for_each_entry_safe(msg, tmp, head, msg_list) {
+ list_for_each_entry_safe(msg, tmp, head, msg_list) {
struct lnet_match_info info;
lnet_hdr_t *hdr;
int rc;
/* Hurrah! This _is_ a match */
LASSERT((rc & LNET_MATCHMD_FINISH) != 0);
- cfs_list_del_init(&msg->msg_list);
+ list_del_init(&msg->msg_list);
if (head == &ptl->ptl_msg_stealing) {
if (exhausted)
}
if ((rc & LNET_MATCHMD_OK) != 0) {
- cfs_list_add_tail(&msg->msg_list, matches);
+ list_add_tail(&msg->msg_list, matches);
CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
"match "LPU64" offset %d length %d.\n",
info.mi_portal, info.mi_mbits,
info.mi_roffset, info.mi_rlength);
} else {
- cfs_list_add_tail(&msg->msg_list, drops);
+ list_add_tail(&msg->msg_list, drops);
}
if (exhausted)
goto again;
}
- if (lnet_ptl_is_wildcard(ptl) && !exhausted && !mtable->mt_enabled)
- lnet_ptl_enable_mt(ptl, cpt);
+ if (lnet_ptl_is_wildcard(ptl) && !exhausted) {
+ lnet_mt_set_exhausted(mtable, me->me_pos, 0);
+ if (!mtable->mt_enabled)
+ lnet_ptl_enable_mt(ptl, cpt);
+ }
lnet_ptl_unlock(ptl);
}
if (ptl->ptl_mtables == NULL) /* uninitialized portal */
return;
- LASSERT(cfs_list_empty(&ptl->ptl_msg_delayed));
- LASSERT(cfs_list_empty(&ptl->ptl_msg_stealing));
+ LASSERT(list_empty(&ptl->ptl_msg_delayed));
+ LASSERT(list_empty(&ptl->ptl_msg_stealing));
#ifndef __KERNEL__
# ifdef HAVE_LIBPTHREAD
pthread_mutex_destroy(&ptl->ptl_lock);
# endif
#endif
cfs_percpt_for_each(mtable, i, ptl->ptl_mtables) {
- cfs_list_t *mhash;
- lnet_me_t *me;
- int j;
+ struct list_head *mhash;
+ lnet_me_t *me;
+ int j;
if (mtable->mt_mhash == NULL) /* uninitialized match-table */
continue;
mhash = mtable->mt_mhash;
/* cleanup ME */
- while (!cfs_list_empty(&mtable->mt_mlist)) {
- me = cfs_list_entry(mtable->mt_mlist.next,
- lnet_me_t, me_list);
- CERROR("Active wildcard ME %p on exit\n", me);
- cfs_list_del(&me->me_list);
- lnet_me_free(me);
- }
-
- for (j = 0; j < LNET_MT_HASH_SIZE; j++) {
- while (!cfs_list_empty(&mhash[j])) {
- me = cfs_list_entry(mhash[j].next,
- lnet_me_t, me_list);
- CERROR("Active unique ME %p on exit\n", me);
- cfs_list_del(&me->me_list);
+ for (j = 0; j < LNET_MT_HASH_SIZE + 1; j++) {
+ while (!list_empty(&mhash[j])) {
+ me = list_entry(mhash[j].next,
+ lnet_me_t, me_list);
+ CERROR("Active ME %p on exit\n", me);
+ list_del(&me->me_list);
lnet_me_free(me);
}
}
-
- LIBCFS_FREE(mhash, sizeof(*mhash) * LNET_MT_HASH_SIZE);
+ /* the extra entry is for MEs with ignore bits */
+ LIBCFS_FREE(mhash, sizeof(*mhash) * (LNET_MT_HASH_SIZE + 1));
}
cfs_percpt_free(ptl->ptl_mtables);
lnet_ptl_setup(struct lnet_portal *ptl, int index)
{
struct lnet_match_table *mtable;
- cfs_list_t *mhash;
+ struct list_head *mhash;
int i;
int j;
}
ptl->ptl_index = index;
- CFS_INIT_LIST_HEAD(&ptl->ptl_msg_delayed);
- CFS_INIT_LIST_HEAD(&ptl->ptl_msg_stealing);
+ INIT_LIST_HEAD(&ptl->ptl_msg_delayed);
+ INIT_LIST_HEAD(&ptl->ptl_msg_stealing);
#ifdef __KERNEL__
- cfs_spin_lock_init(&ptl->ptl_lock);
+ spin_lock_init(&ptl->ptl_lock);
#else
# ifdef HAVE_LIBPTHREAD
pthread_mutex_init(&ptl->ptl_lock, NULL);
# endif
#endif
cfs_percpt_for_each(mtable, i, ptl->ptl_mtables) {
+ /* the extra entry is for MEs with ignore bits */
LIBCFS_CPT_ALLOC(mhash, lnet_cpt_table(), i,
- sizeof(*mhash) * LNET_MT_HASH_SIZE);
+ sizeof(*mhash) * (LNET_MT_HASH_SIZE + 1));
if (mhash == NULL) {
CERROR("Failed to create match hash for portal %d\n",
index);
goto failed;
}
+ memset(&mtable->mt_exhausted[0], -1,
+ sizeof(mtable->mt_exhausted[0]) *
+ LNET_MT_EXHAUSTED_BMAP);
mtable->mt_mhash = mhash;
- for (j = 0; j < LNET_MT_HASH_SIZE; j++)
- CFS_INIT_LIST_HEAD(&mhash[j]);
+ for (j = 0; j < LNET_MT_HASH_SIZE + 1; j++)
+ INIT_LIST_HEAD(&mhash[j]);
- CFS_INIT_LIST_HEAD(&mtable->mt_mlist);
mtable->mt_portal = index;
mtable->mt_cpt = i;
}
return 0;
}
+EXPORT_SYMBOL(LNetSetLazyPortal);
/**
* Turn off the lazy portal attribute. Delayed requests on the portal,
LNetClearLazyPortal(int portal)
{
struct lnet_portal *ptl;
- CFS_LIST_HEAD (zombies);
+ struct list_head zombies = LIST_HEAD_INIT(zombies);
if (portal < 0 || portal >= the_lnet.ln_nportals)
return -EINVAL;
CDEBUG(D_NET, "clearing portal %d lazy\n", portal);
/* grab all the blocked messages atomically */
- cfs_list_splice_init(&ptl->ptl_msg_delayed, &zombies);
+ list_splice_init(&ptl->ptl_msg_delayed, &zombies);
lnet_ptl_unsetopt(ptl, LNET_PTL_LAZY);
return 0;
}
+EXPORT_SYMBOL(LNetClearLazyPortal);