* GPL HEADER END
*/
/*
- * Copyright (c) 2012, Intel Corporation.
+ * Copyright (c) 2012, 2017, 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 = LNET_PTL_ROTOR_HASH_RT;
-CFS_MODULE_PARM(portal_rotor, "i", int, 0644,
- "redirect PUTs to different cpu-partitions");
+int portal_rotor = LNET_PTL_ROTOR_HASH_RT;
+module_param(portal_rotor, int, 0644);
+MODULE_PARM_DESC(portal_rotor, "redirect PUTs to different cpu-partitions");
static int
-lnet_ptl_match_type(unsigned int index, lnet_process_id_t match_id,
+lnet_ptl_match_type(unsigned int index, struct lnet_process_id match_id,
__u64 mbits, __u64 ignore_bits)
{
struct lnet_portal *ptl = the_lnet.ln_portals[index];
}
static int
-lnet_try_match_md(lnet_libmd_t *md,
+lnet_try_match_md(struct lnet_libmd *md,
struct lnet_match_info *info, struct lnet_msg *msg)
{
/* ALWAYS called holding the lnet_res_lock, and can't lnet_res_unlock;
* lnet_match_blocked_msg() relies on this to avoid races */
unsigned int offset;
unsigned int mlength;
- lnet_me_t *me = md->md_me;
+ struct lnet_me *me = md->md_me;
/* MD exhausted */
if (lnet_md_exhausted(md))
mlength = info->mi_rlength;
} else if ((md->md_options & LNET_MD_TRUNCATE) == 0) {
/* this packet _really_ is too big */
- CERROR("Matching packet from %s, match "LPU64
+ CERROR("Matching packet from %s, match %llu"
" length %d too big: %d left, %d allowed\n",
libcfs_id2str(info->mi_id), info->mi_mbits,
info->mi_rlength, md->md_length - offset, mlength);
/* Commit to this ME/MD */
CDEBUG(D_NET, "Incoming %s index %x from %s of "
- "length %d/%d into md "LPX64" [%d] + %d\n",
+ "length %d/%d into md %#llx [%d] + %d\n",
(info->mi_opc == LNET_MD_OP_PUT) ? "put" : "get",
info->mi_portal, libcfs_id2str(info->mi_id), mlength,
info->mi_rlength, md->md_lh.lh_cookie, md->md_niov, offset);
}
static struct lnet_match_table *
-lnet_match2mt(struct lnet_portal *ptl, lnet_process_id_t id, __u64 mbits)
+lnet_match2mt(struct lnet_portal *ptl, struct lnet_process_id id, __u64 mbits)
{
if (LNET_CPT_NUMBER == 1)
return ptl->ptl_mtables[0]; /* the only one */
/* if it's a unique portal, return match-table hashed by NID */
return lnet_ptl_is_unique(ptl) ?
- ptl->ptl_mtables[lnet_cpt_of_nid(id.nid)] : NULL;
+ ptl->ptl_mtables[lnet_cpt_of_nid(id.nid, NULL)] : NULL;
}
struct lnet_match_table *
-lnet_mt_of_attach(unsigned int index, lnet_process_id_t id,
- __u64 mbits, __u64 ignore_bits, lnet_ins_pos_t pos)
+lnet_mt_of_attach(unsigned int index, struct lnet_process_id id,
+ __u64 mbits, __u64 ignore_bits, enum lnet_ins_pos pos)
{
struct lnet_portal *ptl;
struct lnet_match_table *mtable;
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);
+ cpt = info->mi_cpt;
else
cpt = rotor % LNET_CPT_NUMBER;
struct list_head *
lnet_mt_match_head(struct lnet_match_table *mtable,
- lnet_process_id_t id, __u64 mbits)
+ struct lnet_process_id id, __u64 mbits)
{
struct lnet_portal *ptl = the_lnet.ln_portals[mtable->mt_portal];
LASSERT(lnet_ptl_is_unique(ptl));
hash = hash_long(hash, LNET_MT_HASH_BITS);
- return &mtable->mt_mhash[hash];
+ return &mtable->mt_mhash[hash & LNET_MT_HASH_MASK];
}
}
struct lnet_match_info *info, struct lnet_msg *msg)
{
struct list_head *head;
- lnet_me_t *me;
- lnet_me_t *tmp;
+ struct lnet_me *me;
+ struct lnet_me *tmp;
int exhausted = 0;
int rc;
int rc = 0;
int i;
- /* steal buffer from other CPTs, and delay it if nothing to steal,
- * this function is more expensive than a regular match, but we
- * don't expect it can happen a lot */
+ /*
+ * Steal buffer from other CPTs, and delay msg if nothing to
+ * steal. This function is more expensive than a regular
+ * match, but we don't expect it can happen a lot. The return
+ * code contains one of LNET_MATCHMD_OK, LNET_MATCHMD_DROP, or
+ * LNET_MATCHMD_NONE.
+ */
LASSERT(lnet_ptl_is_wildcard(ptl));
for (i = 0; i < LNET_CPT_NUMBER; i++) {
lnet_res_lock(cpt);
lnet_ptl_lock(ptl);
- if (i == 0) { /* the first try, attach on stealing list */
+ if (i == 0) {
+ /* The first try, add to stealing list. */
list_add_tail(&msg->msg_list,
&ptl->ptl_msg_stealing);
}
- if (!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 &&
mtable->mt_enabled)
lnet_ptl_disable_mt(ptl, cpt);
- if ((rc & LNET_MATCHMD_FINISH) != 0)
+ if ((rc & LNET_MATCHMD_FINISH) != 0) {
+ /* Match found, remove from stealing list. */
+ list_del_init(&msg->msg_list);
+ } else if (i == LNET_CPT_NUMBER - 1 || /* (1) */
+ ptl->ptl_mt_nmaps == 0 || /* (2) */
+ (ptl->ptl_mt_nmaps == 1 && /* (3) */
+ ptl->ptl_mt_maps[0] == cpt)) {
+ /*
+ * No match found, and this is either
+ * (1) the last cpt to check, or
+ * (2) there is no active cpt, or
+ * (3) this is the only active cpt.
+ * There is nothing to steal: delay or
+ * drop the message.
+ */
list_del_init(&msg->msg_list);
- } else {
- /* could be matched by lnet_ptl_attach_md()
- * which is called by another thread */
- rc = msg->msg_md == NULL ?
- LNET_MATCHMD_DROP : LNET_MATCHMD_OK;
- }
-
- 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 */
- list_del_init(&msg->msg_list);
-
- if (lnet_ptl_is_lazy(ptl)) {
- msg->msg_rx_delayed = 1;
- list_add_tail(&msg->msg_list,
- &ptl->ptl_msg_delayed);
- rc = LNET_MATCHMD_NONE;
+ if (lnet_ptl_is_lazy(ptl)) {
+ msg->msg_rx_delayed = 1;
+ list_add_tail(&msg->msg_list,
+ &ptl->ptl_msg_delayed);
+ rc = LNET_MATCHMD_NONE;
+ } else {
+ rc = LNET_MATCHMD_DROP;
+ }
} else {
- rc = LNET_MATCHMD_DROP;
+ /* Do another iteration. */
+ rc = 0;
}
+ } else {
+ /*
+ * No longer on stealing list: another thread
+ * matched the message in lnet_ptl_attach_md().
+ * We are now expected to handle the message.
+ */
+ rc = msg->msg_md == NULL ?
+ LNET_MATCHMD_DROP : LNET_MATCHMD_OK;
}
lnet_ptl_unlock(ptl);
lnet_res_unlock(cpt);
- if ((rc & LNET_MATCHMD_FINISH) != 0 || msg->msg_rx_delayed)
+ /*
+ * Note that test (1) above ensures that we always
+ * exit the loop through this break statement.
+ *
+ * LNET_MATCHMD_NONE means msg was added to the
+ * delayed queue, and we may no longer reference it
+ * after lnet_ptl_unlock() and lnet_res_unlock().
+ */
+ if (rc & (LNET_MATCHMD_FINISH | LNET_MATCHMD_NONE))
break;
}
int rc;
CDEBUG(D_NET, "Request from %s of length %d into portal %d "
- "MB="LPX64"\n", libcfs_id2str(info->mi_id),
+ "MB=%#llx\n", libcfs_id2str(info->mi_id),
info->mi_rlength, info->mi_portal, info->mi_mbits);
if (info->mi_portal >= the_lnet.ln_nportals) {
mtable = lnet_mt_of_match(info, msg);
lnet_res_lock(mtable->mt_cpt);
- if (the_lnet.ln_shutdown) {
+ if (the_lnet.ln_state != LNET_STATE_RUNNING) {
rc = LNET_MATCHMD_DROP;
goto out1;
}
lnet_ptl_unlock(ptl);
lnet_res_unlock(mtable->mt_cpt);
-
- } else {
+ rc = LNET_MATCHMD_NONE;
+ } else {
lnet_res_unlock(mtable->mt_cpt);
rc = lnet_ptl_match_delay(ptl, info, msg);
}
- if (msg->msg_rx_delayed) {
+ /* LNET_MATCHMD_NONE means msg was added to the delay queue */
+ if (rc & LNET_MATCHMD_NONE) {
CDEBUG(D_NET,
- "Delaying %s from %s ptl %d MB "LPX64" off %d len %d\n",
+ "Delaying %s from %s ptl %d MB %#llx off %d len %d\n",
info->mi_opc == LNET_MD_OP_PUT ? "PUT" : "GET",
libcfs_id2str(info->mi_id), info->mi_portal,
info->mi_mbits, info->mi_roffset, info->mi_rlength);
}
void
-lnet_ptl_detach_md(lnet_me_t *me, lnet_libmd_t *md)
+lnet_ptl_detach_md(struct lnet_me *me, struct lnet_libmd *md)
{
LASSERT(me->me_md == md && md->md_me == me);
/* called with lnet_res_lock held */
void
-lnet_ptl_attach_md(lnet_me_t *me, lnet_libmd_t *md,
+lnet_ptl_attach_md(struct lnet_me *me, struct lnet_libmd *md,
struct list_head *matches, struct list_head *drops)
{
- struct lnet_portal *ptl = the_lnet.ln_portals[me->me_portal];
+ struct lnet_portal *ptl = the_lnet.ln_portals[me->me_portal];
struct lnet_match_table *mtable;
- struct list_head *head;
- lnet_msg_t *tmp;
- lnet_msg_t *msg;
- int exhausted = 0;
- int cpt;
+ struct list_head *head;
+ struct lnet_msg *tmp;
+ struct lnet_msg *msg;
+ int exhausted = 0;
+ int cpt;
LASSERT(md->md_refcount == 0); /* a brand new MD */
again:
list_for_each_entry_safe(msg, tmp, head, msg_list) {
struct lnet_match_info info;
- lnet_hdr_t *hdr;
+ struct lnet_hdr *hdr;
int rc;
LASSERT(msg->msg_rx_delayed || head == &ptl->ptl_msg_stealing);
hdr = &msg->msg_hdr;
- info.mi_id.nid = hdr->src_nid;
+ /* Multi-Rail: Primary peer NID */
+ info.mi_id.nid = msg->msg_initiator;
info.mi_id.pid = hdr->src_pid;
info.mi_opc = LNET_MD_OP_PUT;
info.mi_portal = hdr->msg.put.ptl_index;
list_add_tail(&msg->msg_list, matches);
CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
- "match "LPU64" offset %d length %d.\n",
+ "match %llu offset %d length %d.\n",
libcfs_id2str(info.mi_id),
info.mi_portal, info.mi_mbits,
info.mi_roffset, info.mi_rlength);
lnet_ptl_unlock(ptl);
}
-void
+static void
lnet_ptl_cleanup(struct lnet_portal *ptl)
{
struct lnet_match_table *mtable;
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) {
struct list_head *mhash;
- lnet_me_t *me;
+ struct lnet_me *me;
int j;
if (mtable->mt_mhash == NULL) /* uninitialized match-table */
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);
+ struct lnet_me, me_list);
CERROR("Active ME %p on exit\n", me);
list_del(&me->me_list);
- lnet_me_free(me);
+ CDEBUG(D_MALLOC,
+ "slab-freed 'me' at %p in cleanup.\n",
+ me);
+ kmem_cache_free(lnet_mes_cachep, me);
}
}
/* the extra entry is for MEs with ignore bits */
- LIBCFS_FREE(mhash, sizeof(*mhash) * (LNET_MT_HASH_SIZE + 1));
+ CFS_FREE_PTR_ARRAY(mhash, LNET_MT_HASH_SIZE + 1);
}
cfs_percpt_free(ptl->ptl_mtables);
ptl->ptl_mtables = NULL;
}
-int
+static int
lnet_ptl_setup(struct lnet_portal *ptl, int index)
{
struct lnet_match_table *mtable;
ptl->ptl_index = index;
INIT_LIST_HEAD(&ptl->ptl_msg_delayed);
INIT_LIST_HEAD(&ptl->ptl_msg_stealing);
-#ifdef __KERNEL__
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,
*
* \param portal Index of the portal to enable the lazy attribute on.
*
- * \retval 0 On success.
+ * \retval 0 On success.
* \retval -EINVAL If \a portal is not a valid index.
*/
int
}
EXPORT_SYMBOL(LNetSetLazyPortal);
-/**
- * Turn off the lazy portal attribute. Delayed requests on the portal,
- * if any, will be all dropped when this function returns.
- *
- * \param portal Index of the portal to disable the lazy attribute on.
- *
- * \retval 0 On success.
- * \retval -EINVAL If \a portal is not a valid index.
- */
int
-LNetClearLazyPortal(int portal)
+lnet_clear_lazy_portal(struct lnet_ni *ni, int portal, char *reason)
{
struct lnet_portal *ptl;
- struct list_head zombies = LIST_HEAD_INIT(zombies);
+ LIST_HEAD(zombies);
if (portal < 0 || portal >= the_lnet.ln_nportals)
return -EINVAL;
return 0;
}
- if (the_lnet.ln_shutdown)
- CWARN("Active lazy portal %d on exit\n", portal);
- else
- CDEBUG(D_NET, "clearing portal %d lazy\n", portal);
+ if (ni != NULL) {
+ struct lnet_msg *msg, *tmp;
- /* grab all the blocked messages atomically */
- list_splice_init(&ptl->ptl_msg_delayed, &zombies);
+ /* grab all messages which are on the NI passed in */
+ list_for_each_entry_safe(msg, tmp, &ptl->ptl_msg_delayed,
+ msg_list) {
+ if (msg->msg_txni == ni || msg->msg_rxni == ni)
+ list_move(&msg->msg_list, &zombies);
+ }
+ } else {
+ if (the_lnet.ln_state != LNET_STATE_RUNNING)
+ CWARN("Active lazy portal %d on exit\n", portal);
+ else
+ CDEBUG(D_NET, "clearing portal %d lazy\n", portal);
- lnet_ptl_unsetopt(ptl, LNET_PTL_LAZY);
+ /* grab all the blocked messages atomically */
+ list_splice_init(&ptl->ptl_msg_delayed, &zombies);
+
+ lnet_ptl_unsetopt(ptl, LNET_PTL_LAZY);
+ }
lnet_ptl_unlock(ptl);
lnet_res_unlock(LNET_LOCK_EX);
- lnet_drop_delayed_msg_list(&zombies, "Clearing lazy portal attr");
+ lnet_drop_delayed_msg_list(&zombies, reason);
return 0;
}
+
+/**
+ * Turn off the lazy portal attribute. Delayed requests on the portal,
+ * if any, will be all dropped when this function returns.
+ *
+ * \param portal Index of the portal to disable the lazy attribute on.
+ *
+ * \retval 0 On success.
+ * \retval -EINVAL If \a portal is not a valid index.
+ */
+int
+LNetClearLazyPortal(int portal)
+{
+ return lnet_clear_lazy_portal(NULL, portal,
+ "Clearing lazy portal attr");
+}
EXPORT_SYMBOL(LNetClearLazyPortal);