-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* GPL HEADER START
*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
* in the LICENSE file that accompanied this code).
*
* You should have received a copy of the GNU General Public License
- * version 2 along with this program; If not, see [sun.com URL with a
- * copy of GPLv2].
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * version 2 along with this program; If not, see
+ * http://www.gnu.org/licenses/gpl-2.0.html
*
* GPL HEADER END
*/
/*
- * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2012, 2016, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#define DEBUG_SUBSYSTEM S_LNET
#include <lnet/lib-lnet.h>
+#include <linux/nsproxy.h>
+#include <net/net_namespace.h>
static int local_nid_dist_zero = 1;
-CFS_MODULE_PARM(local_nid_dist_zero, "i", int, 0444,
- "Reserved");
-
-/* forward ref */
-static void lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg);
-static void lnet_drop_delayed_put(lnet_msg_t *msg, char *reason);
-
-#define LNET_MATCHMD_NONE 0 /* Didn't match */
-#define LNET_MATCHMD_OK 1 /* Matched OK */
-#define LNET_MATCHMD_DROP 2 /* Must be discarded */
-
-static int
-lnet_try_match_md (int index, int op_mask, lnet_process_id_t src,
- unsigned int rlength, unsigned int roffset,
- __u64 match_bits, lnet_libmd_t *md, lnet_msg_t *msg,
- unsigned int *mlength_out, unsigned int *offset_out)
-{
- /* ALWAYS called holding the LNET_LOCK, and can't LNET_UNLOCK;
- * lnet_match_blocked_msg() relies on this to avoid races */
- unsigned int offset;
- unsigned int mlength;
- lnet_me_t *me = md->md_me;
-
- /* mismatched MD op */
- if ((md->md_options & op_mask) == 0)
- return LNET_MATCHMD_NONE;
-
- /* MD exhausted */
- if (lnet_md_exhausted(md))
- return LNET_MATCHMD_NONE;
-
- /* mismatched ME nid/pid? */
- if (me->me_match_id.nid != LNET_NID_ANY &&
- me->me_match_id.nid != src.nid)
- return LNET_MATCHMD_NONE;
-
- if (me->me_match_id.pid != LNET_PID_ANY &&
- me->me_match_id.pid != src.pid)
- return LNET_MATCHMD_NONE;
-
- /* mismatched ME matchbits? */
- if (((me->me_match_bits ^ match_bits) & ~me->me_ignore_bits) != 0)
- return LNET_MATCHMD_NONE;
-
- /* Hurrah! This _is_ a match; check it out... */
-
- if ((md->md_options & LNET_MD_MANAGE_REMOTE) == 0)
- offset = md->md_offset;
- else
- offset = roffset;
-
- if ((md->md_options & LNET_MD_MAX_SIZE) != 0) {
- mlength = md->md_max_size;
- LASSERT (md->md_offset + mlength <= md->md_length);
- } else {
- mlength = md->md_length - offset;
- }
-
- if (rlength <= mlength) { /* fits in allowed space */
- mlength = rlength;
- } else if ((md->md_options & LNET_MD_TRUNCATE) == 0) {
- /* this packet _really_ is too big */
- CERROR("Matching packet from %s, match "LPU64
- " length %d too big: %d left, %d allowed\n",
- libcfs_id2str(src), match_bits, rlength,
- md->md_length - offset, mlength);
-
- return LNET_MATCHMD_DROP;
- }
-
- /* Commit to this ME/MD */
- CDEBUG(D_NET, "Incoming %s index %x from %s of "
- "length %d/%d into md "LPX64" [%d] + %d\n",
- (op_mask == LNET_MD_OP_PUT) ? "put" : "get",
- index, libcfs_id2str(src), mlength, rlength,
- md->md_lh.lh_cookie, md->md_niov, offset);
-
- lnet_commit_md(md, msg);
- md->md_offset = offset + mlength;
-
- /* NB Caller will set ev.type and ev.hdr_data */
- msg->msg_ev.initiator = src;
- msg->msg_ev.pt_index = index;
- msg->msg_ev.match_bits = match_bits;
- msg->msg_ev.rlength = rlength;
- msg->msg_ev.mlength = mlength;
- msg->msg_ev.offset = offset;
-
- lnet_md_deconstruct(md, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, md);
-
- *offset_out = offset;
- *mlength_out = mlength;
-
- /* Auto-unlink NOW, so the ME gets unlinked if required.
- * We bumped md->md_refcount above so the MD just gets flagged
- * for unlink when it is finalized. */
- if ((md->md_flags & LNET_MD_FLAG_AUTO_UNLINK) != 0 &&
- lnet_md_exhausted(md)) {
- lnet_md_unlink(md);
- }
-
- return LNET_MATCHMD_OK;
-}
-
-static int
-lnet_match_md(int index, int op_mask, lnet_process_id_t src,
- unsigned int rlength, unsigned int roffset,
- __u64 match_bits, lnet_msg_t *msg,
- unsigned int *mlength_out, unsigned int *offset_out,
- lnet_libmd_t **md_out)
-{
- lnet_portal_t *ptl = &the_lnet.ln_portals[index];
- lnet_me_t *me;
- lnet_me_t *tmp;
- lnet_libmd_t *md;
- int rc;
-
- CDEBUG (D_NET, "Request from %s of length %d into portal %d "
- "MB="LPX64"\n", libcfs_id2str(src), rlength, index, match_bits);
-
- if (index < 0 || index >= the_lnet.ln_nportals) {
- CERROR("Invalid portal %d not in [0-%d]\n",
- index, the_lnet.ln_nportals);
- return LNET_MATCHMD_DROP;
- }
-
- list_for_each_entry_safe (me, tmp, &ptl->ptl_ml, me_list) {
- md = me->me_md;
-
- /* ME attached but MD not attached yet */
- if (md == NULL)
- continue;
-
- LASSERT (me == md->md_me);
-
- rc = lnet_try_match_md(index, op_mask, src, rlength,
- roffset, match_bits, md, msg,
- mlength_out, offset_out);
- switch (rc) {
- default:
- LBUG();
-
- case LNET_MATCHMD_NONE:
- continue;
-
- case LNET_MATCHMD_OK:
- *md_out = md;
- return LNET_MATCHMD_OK;
-
- case LNET_MATCHMD_DROP:
- return LNET_MATCHMD_DROP;
- }
- /* not reached */
- }
-
- if (op_mask == LNET_MD_OP_GET ||
- (ptl->ptl_options & LNET_PTL_LAZY) == 0)
- return LNET_MATCHMD_DROP;
-
- return LNET_MATCHMD_NONE;
-}
+module_param(local_nid_dist_zero, int, 0444);
+MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
int
-lnet_fail_nid (lnet_nid_t nid, unsigned int threshold)
+lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
{
- lnet_test_peer_t *tp;
- struct list_head *el;
- struct list_head *next;
- struct list_head cull;
-
- LASSERT (the_lnet.ln_init);
-
- if (threshold != 0) {
- /* Adding a new entry */
- LIBCFS_ALLOC(tp, sizeof(*tp));
- if (tp == NULL)
- return -ENOMEM;
-
- tp->tp_nid = nid;
- tp->tp_threshold = threshold;
-
- LNET_LOCK();
- list_add_tail (&tp->tp_list, &the_lnet.ln_test_peers);
- LNET_UNLOCK();
- return 0;
- }
-
- /* removing entries */
- CFS_INIT_LIST_HEAD (&cull);
-
- LNET_LOCK();
-
- list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
- tp = list_entry (el, lnet_test_peer_t, tp_list);
-
- if (tp->tp_threshold == 0 || /* needs culling anyway */
- nid == LNET_NID_ANY || /* removing all entries */
- tp->tp_nid == nid) /* matched this one */
- {
- list_del (&tp->tp_list);
- list_add (&tp->tp_list, &cull);
- }
- }
-
- LNET_UNLOCK();
-
- while (!list_empty (&cull)) {
- tp = list_entry (cull.next, lnet_test_peer_t, tp_list);
-
- list_del (&tp->tp_list);
- LIBCFS_FREE(tp, sizeof (*tp));
- }
- return 0;
+ lnet_test_peer_t *tp;
+ struct list_head *el;
+ struct list_head *next;
+ struct list_head cull;
+
+ /* NB: use lnet_net_lock(0) to serialize operations on test peers */
+ if (threshold != 0) {
+ /* Adding a new entry */
+ LIBCFS_ALLOC(tp, sizeof(*tp));
+ if (tp == NULL)
+ return -ENOMEM;
+
+ tp->tp_nid = nid;
+ tp->tp_threshold = threshold;
+
+ lnet_net_lock(0);
+ list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
+ lnet_net_unlock(0);
+ return 0;
+ }
+
+ /* removing entries */
+ INIT_LIST_HEAD(&cull);
+
+ lnet_net_lock(0);
+
+ list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
+ tp = list_entry(el, lnet_test_peer_t, tp_list);
+
+ if (tp->tp_threshold == 0 || /* needs culling anyway */
+ nid == LNET_NID_ANY || /* removing all entries */
+ tp->tp_nid == nid) { /* matched this one */
+ list_del(&tp->tp_list);
+ list_add(&tp->tp_list, &cull);
+ }
+ }
+
+ lnet_net_unlock(0);
+
+ while (!list_empty(&cull)) {
+ tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
+
+ list_del(&tp->tp_list);
+ LIBCFS_FREE(tp, sizeof(*tp));
+ }
+ return 0;
}
static int
fail_peer (lnet_nid_t nid, int outgoing)
{
- lnet_test_peer_t *tp;
- struct list_head *el;
- struct list_head *next;
- struct list_head cull;
- int fail = 0;
-
- CFS_INIT_LIST_HEAD (&cull);
-
- LNET_LOCK();
-
- list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
- tp = list_entry (el, lnet_test_peer_t, tp_list);
-
- if (tp->tp_threshold == 0) {
- /* zombie entry */
- if (outgoing) {
- /* only cull zombies on outgoing tests,
- * since we may be at interrupt priority on
- * incoming messages. */
- list_del (&tp->tp_list);
- list_add (&tp->tp_list, &cull);
- }
- continue;
- }
-
- if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
- nid == tp->tp_nid) { /* fail this peer */
- fail = 1;
-
- if (tp->tp_threshold != LNET_MD_THRESH_INF) {
- tp->tp_threshold--;
- if (outgoing &&
- tp->tp_threshold == 0) {
- /* see above */
- list_del (&tp->tp_list);
- list_add (&tp->tp_list, &cull);
- }
- }
- break;
- }
- }
-
- LNET_UNLOCK ();
-
- while (!list_empty (&cull)) {
- tp = list_entry (cull.next, lnet_test_peer_t, tp_list);
- list_del (&tp->tp_list);
-
- LIBCFS_FREE(tp, sizeof (*tp));
- }
-
- return (fail);
-}
-
-unsigned int
-lnet_iov_nob (unsigned int niov, struct iovec *iov)
-{
- unsigned int nob = 0;
-
- while (niov-- > 0)
- nob += (iov++)->iov_len;
-
- return (nob);
-}
-
-void
-lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset,
- unsigned int nsiov, struct iovec *siov, unsigned int soffset,
- unsigned int nob)
-{
- /* NB diov, siov are READ-ONLY */
- unsigned int this_nob;
-
- if (nob == 0)
- return;
-
- /* skip complete frags before 'doffset' */
- LASSERT (ndiov > 0);
- while (doffset >= diov->iov_len) {
- doffset -= diov->iov_len;
- diov++;
- ndiov--;
- LASSERT (ndiov > 0);
- }
-
- /* skip complete frags before 'soffset' */
- LASSERT (nsiov > 0);
- while (soffset >= siov->iov_len) {
- soffset -= siov->iov_len;
- siov++;
- nsiov--;
- LASSERT (nsiov > 0);
- }
-
- do {
- LASSERT (ndiov > 0);
- LASSERT (nsiov > 0);
- this_nob = MIN(diov->iov_len - doffset,
- siov->iov_len - soffset);
- this_nob = MIN(this_nob, nob);
-
- memcpy ((char *)diov->iov_base + doffset,
- (char *)siov->iov_base + soffset, this_nob);
- nob -= this_nob;
-
- if (diov->iov_len > doffset + this_nob) {
- doffset += this_nob;
- } else {
- diov++;
- ndiov--;
- doffset = 0;
- }
-
- if (siov->iov_len > soffset + this_nob) {
- soffset += this_nob;
- } else {
- siov++;
- nsiov--;
- soffset = 0;
- }
- } while (nob > 0);
+ lnet_test_peer_t *tp;
+ struct list_head *el;
+ struct list_head *next;
+ struct list_head cull;
+ int fail = 0;
+
+ INIT_LIST_HEAD(&cull);
+
+ /* NB: use lnet_net_lock(0) to serialize operations on test peers */
+ lnet_net_lock(0);
+
+ list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
+ tp = list_entry(el, lnet_test_peer_t, tp_list);
+
+ if (tp->tp_threshold == 0) {
+ /* zombie entry */
+ if (outgoing) {
+ /* only cull zombies on outgoing tests,
+ * since we may be at interrupt priority on
+ * incoming messages. */
+ list_del(&tp->tp_list);
+ list_add(&tp->tp_list, &cull);
+ }
+ continue;
+ }
+
+ if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
+ nid == tp->tp_nid) { /* fail this peer */
+ fail = 1;
+
+ if (tp->tp_threshold != LNET_MD_THRESH_INF) {
+ tp->tp_threshold--;
+ if (outgoing &&
+ tp->tp_threshold == 0) {
+ /* see above */
+ list_del(&tp->tp_list);
+ list_add(&tp->tp_list, &cull);
+ }
+ }
+ break;
+ }
+ }
+
+ lnet_net_unlock(0);
+
+ while (!list_empty(&cull)) {
+ tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
+ list_del(&tp->tp_list);
+
+ LIBCFS_FREE(tp, sizeof(*tp));
+ }
+
+ return fail;
}
-int
-lnet_extract_iov (int dst_niov, struct iovec *dst,
- int src_niov, struct iovec *src,
- unsigned int offset, unsigned int len)
-{
- /* Initialise 'dst' to the subset of 'src' starting at 'offset',
- * for exactly 'len' bytes, and return the number of entries.
- * NB not destructive to 'src' */
- unsigned int frag_len;
- unsigned int niov;
-
- if (len == 0) /* no data => */
- return (0); /* no frags */
-
- LASSERT (src_niov > 0);
- while (offset >= src->iov_len) { /* skip initial frags */
- offset -= src->iov_len;
- src_niov--;
- src++;
- LASSERT (src_niov > 0);
- }
-
- niov = 1;
- for (;;) {
- LASSERT (src_niov > 0);
- LASSERT (niov <= dst_niov);
-
- frag_len = src->iov_len - offset;
- dst->iov_base = ((char *)src->iov_base) + offset;
-
- if (len <= frag_len) {
- dst->iov_len = len;
- return (niov);
- }
-
- dst->iov_len = frag_len;
-
- len -= frag_len;
- dst++;
- src++;
- niov++;
- src_niov--;
- offset = 0;
- }
-}
-
-#ifndef __KERNEL__
unsigned int
-lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
+lnet_iov_nob(unsigned int niov, struct kvec *iov)
{
- LASSERT (0);
- return (0);
-}
+ unsigned int nob = 0;
-void
-lnet_copy_kiov2kiov (unsigned int ndkiov, lnet_kiov_t *dkiov, unsigned int doffset,
- unsigned int nskiov, lnet_kiov_t *skiov, unsigned int soffset,
- unsigned int nob)
-{
- LASSERT (0);
-}
+ LASSERT(niov == 0 || iov != NULL);
+ while (niov-- > 0)
+ nob += (iov++)->iov_len;
-void
-lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
- unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
- unsigned int nob)
-{
- LASSERT (0);
+ return (nob);
}
+EXPORT_SYMBOL(lnet_iov_nob);
void
-lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
- unsigned int niov, struct iovec *iov, unsigned int iovoffset,
- unsigned int nob)
+lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
+ unsigned int nsiov, struct kvec *siov, unsigned int soffset,
+ unsigned int nob)
{
- LASSERT (0);
+ /* NB diov, siov are READ-ONLY */
+ unsigned int this_nob;
+
+ if (nob == 0)
+ return;
+
+ /* skip complete frags before 'doffset' */
+ LASSERT(ndiov > 0);
+ while (doffset >= diov->iov_len) {
+ doffset -= diov->iov_len;
+ diov++;
+ ndiov--;
+ LASSERT(ndiov > 0);
+ }
+
+ /* skip complete frags before 'soffset' */
+ LASSERT(nsiov > 0);
+ while (soffset >= siov->iov_len) {
+ soffset -= siov->iov_len;
+ siov++;
+ nsiov--;
+ LASSERT(nsiov > 0);
+ }
+
+ do {
+ LASSERT(ndiov > 0);
+ LASSERT(nsiov > 0);
+ this_nob = MIN(diov->iov_len - doffset,
+ siov->iov_len - soffset);
+ this_nob = MIN(this_nob, nob);
+
+ memcpy((char *)diov->iov_base + doffset,
+ (char *)siov->iov_base + soffset, this_nob);
+ nob -= this_nob;
+
+ if (diov->iov_len > doffset + this_nob) {
+ doffset += this_nob;
+ } else {
+ diov++;
+ ndiov--;
+ doffset = 0;
+ }
+
+ if (siov->iov_len > soffset + this_nob) {
+ soffset += this_nob;
+ } else {
+ siov++;
+ nsiov--;
+ soffset = 0;
+ }
+ } while (nob > 0);
}
+EXPORT_SYMBOL(lnet_copy_iov2iov);
int
-lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
- int src_niov, lnet_kiov_t *src,
- unsigned int offset, unsigned int len)
+lnet_extract_iov(int dst_niov, struct kvec *dst,
+ int src_niov, struct kvec *src,
+ unsigned int offset, unsigned int len)
{
- LASSERT (0);
+ /* Initialise 'dst' to the subset of 'src' starting at 'offset',
+ * for exactly 'len' bytes, and return the number of entries.
+ * NB not destructive to 'src' */
+ unsigned int frag_len;
+ unsigned int niov;
+
+ if (len == 0) /* no data => */
+ return (0); /* no frags */
+
+ LASSERT(src_niov > 0);
+ while (offset >= src->iov_len) { /* skip initial frags */
+ offset -= src->iov_len;
+ src_niov--;
+ src++;
+ LASSERT(src_niov > 0);
+ }
+
+ niov = 1;
+ for (;;) {
+ LASSERT(src_niov > 0);
+ LASSERT((int)niov <= dst_niov);
+
+ frag_len = src->iov_len - offset;
+ dst->iov_base = ((char *)src->iov_base) + offset;
+
+ if (len <= frag_len) {
+ dst->iov_len = len;
+ return (niov);
+ }
+
+ dst->iov_len = frag_len;
+
+ len -= frag_len;
+ dst++;
+ src++;
+ niov++;
+ src_niov--;
+ offset = 0;
+ }
}
+EXPORT_SYMBOL(lnet_extract_iov);
-#else /* __KERNEL__ */
unsigned int
-lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
+lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
{
- unsigned int nob = 0;
+ unsigned int nob = 0;
- while (niov-- > 0)
- nob += (kiov++)->kiov_len;
+ LASSERT(niov == 0 || kiov != NULL);
+ while (niov-- > 0)
+ nob += (kiov++)->kiov_len;
- return (nob);
+ return (nob);
}
+EXPORT_SYMBOL(lnet_kiov_nob);
void
-lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
- unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
- unsigned int nob)
+lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
+ unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
+ unsigned int nob)
{
- /* NB diov, siov are READ-ONLY */
- unsigned int this_nob;
- char *daddr = NULL;
- char *saddr = NULL;
-
- if (nob == 0)
- return;
-
- LASSERT (!in_interrupt ());
-
- LASSERT (ndiov > 0);
- while (doffset >= diov->kiov_len) {
- doffset -= diov->kiov_len;
- diov++;
- ndiov--;
- LASSERT (ndiov > 0);
- }
-
- LASSERT (nsiov > 0);
- while (soffset >= siov->kiov_len) {
- soffset -= siov->kiov_len;
- siov++;
- nsiov--;
- LASSERT (nsiov > 0);
- }
-
- do {
- LASSERT (ndiov > 0);
- LASSERT (nsiov > 0);
- this_nob = MIN(diov->kiov_len - doffset,
- siov->kiov_len - soffset);
- this_nob = MIN(this_nob, nob);
-
- if (daddr == NULL)
- daddr = ((char *)cfs_kmap(diov->kiov_page)) +
- diov->kiov_offset + doffset;
- if (saddr == NULL)
- saddr = ((char *)cfs_kmap(siov->kiov_page)) +
- siov->kiov_offset + soffset;
-
- /* Vanishing risk of kmap deadlock when mapping 2 pages.
- * However in practice at least one of the kiovs will be mapped
- * kernel pages and the map/unmap will be NOOPs */
-
- memcpy (daddr, saddr, this_nob);
- nob -= this_nob;
-
- if (diov->kiov_len > doffset + this_nob) {
- daddr += this_nob;
- doffset += this_nob;
- } else {
- cfs_kunmap(diov->kiov_page);
- daddr = NULL;
- diov++;
- ndiov--;
- doffset = 0;
- }
-
- if (siov->kiov_len > soffset + this_nob) {
- saddr += this_nob;
- soffset += this_nob;
- } else {
- cfs_kunmap(siov->kiov_page);
- saddr = NULL;
- siov++;
- nsiov--;
- soffset = 0;
- }
- } while (nob > 0);
-
- if (daddr != NULL)
- cfs_kunmap(diov->kiov_page);
- if (saddr != NULL)
- cfs_kunmap(siov->kiov_page);
+ /* NB diov, siov are READ-ONLY */
+ unsigned int this_nob;
+ char *daddr = NULL;
+ char *saddr = NULL;
+
+ if (nob == 0)
+ return;
+
+ LASSERT (!in_interrupt ());
+
+ LASSERT (ndiov > 0);
+ while (doffset >= diov->kiov_len) {
+ doffset -= diov->kiov_len;
+ diov++;
+ ndiov--;
+ LASSERT(ndiov > 0);
+ }
+
+ LASSERT(nsiov > 0);
+ while (soffset >= siov->kiov_len) {
+ soffset -= siov->kiov_len;
+ siov++;
+ nsiov--;
+ LASSERT(nsiov > 0);
+ }
+
+ do {
+ LASSERT(ndiov > 0);
+ LASSERT(nsiov > 0);
+ this_nob = MIN(diov->kiov_len - doffset,
+ siov->kiov_len - soffset);
+ this_nob = MIN(this_nob, nob);
+
+ if (daddr == NULL)
+ daddr = ((char *)kmap(diov->kiov_page)) +
+ diov->kiov_offset + doffset;
+ if (saddr == NULL)
+ saddr = ((char *)kmap(siov->kiov_page)) +
+ siov->kiov_offset + soffset;
+
+ /* Vanishing risk of kmap deadlock when mapping 2 pages.
+ * However in practice at least one of the kiovs will be mapped
+ * kernel pages and the map/unmap will be NOOPs */
+
+ memcpy (daddr, saddr, this_nob);
+ nob -= this_nob;
+
+ if (diov->kiov_len > doffset + this_nob) {
+ daddr += this_nob;
+ doffset += this_nob;
+ } else {
+ kunmap(diov->kiov_page);
+ daddr = NULL;
+ diov++;
+ ndiov--;
+ doffset = 0;
+ }
+
+ if (siov->kiov_len > soffset + this_nob) {
+ saddr += this_nob;
+ soffset += this_nob;
+ } else {
+ kunmap(siov->kiov_page);
+ saddr = NULL;
+ siov++;
+ nsiov--;
+ soffset = 0;
+ }
+ } while (nob > 0);
+
+ if (daddr != NULL)
+ kunmap(diov->kiov_page);
+ if (saddr != NULL)
+ kunmap(siov->kiov_page);
}
+EXPORT_SYMBOL(lnet_copy_kiov2kiov);
void
-lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
- unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
- unsigned int nob)
+lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
+ unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
+ unsigned int nob)
{
- /* NB iov, kiov are READ-ONLY */
- unsigned int this_nob;
- char *addr = NULL;
-
- if (nob == 0)
- return;
-
- LASSERT (!in_interrupt ());
-
- LASSERT (niov > 0);
- while (iovoffset >= iov->iov_len) {
- iovoffset -= iov->iov_len;
- iov++;
- niov--;
- LASSERT (niov > 0);
- }
-
- LASSERT (nkiov > 0);
- while (kiovoffset >= kiov->kiov_len) {
- kiovoffset -= kiov->kiov_len;
- kiov++;
- nkiov--;
- LASSERT (nkiov > 0);
- }
-
- do {
- LASSERT (niov > 0);
- LASSERT (nkiov > 0);
- this_nob = MIN(iov->iov_len - iovoffset,
- kiov->kiov_len - kiovoffset);
- this_nob = MIN(this_nob, nob);
-
- if (addr == NULL)
- addr = ((char *)cfs_kmap(kiov->kiov_page)) +
- kiov->kiov_offset + kiovoffset;
-
- memcpy ((char *)iov->iov_base + iovoffset, addr, this_nob);
- nob -= this_nob;
-
- if (iov->iov_len > iovoffset + this_nob) {
- iovoffset += this_nob;
- } else {
- iov++;
- niov--;
- iovoffset = 0;
- }
-
- if (kiov->kiov_len > kiovoffset + this_nob) {
- addr += this_nob;
- kiovoffset += this_nob;
- } else {
- cfs_kunmap(kiov->kiov_page);
- addr = NULL;
- kiov++;
- nkiov--;
- kiovoffset = 0;
- }
-
- } while (nob > 0);
-
- if (addr != NULL)
- cfs_kunmap(kiov->kiov_page);
+ /* NB iov, kiov are READ-ONLY */
+ unsigned int this_nob;
+ char *addr = NULL;
+
+ if (nob == 0)
+ return;
+
+ LASSERT (!in_interrupt ());
+
+ LASSERT (niov > 0);
+ while (iovoffset >= iov->iov_len) {
+ iovoffset -= iov->iov_len;
+ iov++;
+ niov--;
+ LASSERT(niov > 0);
+ }
+
+ LASSERT(nkiov > 0);
+ while (kiovoffset >= kiov->kiov_len) {
+ kiovoffset -= kiov->kiov_len;
+ kiov++;
+ nkiov--;
+ LASSERT(nkiov > 0);
+ }
+
+ do {
+ LASSERT(niov > 0);
+ LASSERT(nkiov > 0);
+ this_nob = MIN(iov->iov_len - iovoffset,
+ kiov->kiov_len - kiovoffset);
+ this_nob = MIN(this_nob, nob);
+
+ if (addr == NULL)
+ addr = ((char *)kmap(kiov->kiov_page)) +
+ kiov->kiov_offset + kiovoffset;
+
+ memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
+ nob -= this_nob;
+
+ if (iov->iov_len > iovoffset + this_nob) {
+ iovoffset += this_nob;
+ } else {
+ iov++;
+ niov--;
+ iovoffset = 0;
+ }
+
+ if (kiov->kiov_len > kiovoffset + this_nob) {
+ addr += this_nob;
+ kiovoffset += this_nob;
+ } else {
+ kunmap(kiov->kiov_page);
+ addr = NULL;
+ kiov++;
+ nkiov--;
+ kiovoffset = 0;
+ }
+
+ } while (nob > 0);
+
+ if (addr != NULL)
+ kunmap(kiov->kiov_page);
}
+EXPORT_SYMBOL(lnet_copy_kiov2iov);
void
-lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
- unsigned int niov, struct iovec *iov, unsigned int iovoffset,
- unsigned int nob)
+lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
+ unsigned int niov, struct kvec *iov, unsigned int iovoffset,
+ unsigned int nob)
{
- /* NB kiov, iov are READ-ONLY */
- unsigned int this_nob;
- char *addr = NULL;
-
- if (nob == 0)
- return;
-
- LASSERT (!in_interrupt ());
-
- LASSERT (nkiov > 0);
- while (kiovoffset >= kiov->kiov_len) {
- kiovoffset -= kiov->kiov_len;
- kiov++;
- nkiov--;
- LASSERT (nkiov > 0);
- }
-
- LASSERT (niov > 0);
- while (iovoffset >= iov->iov_len) {
- iovoffset -= iov->iov_len;
- iov++;
- niov--;
- LASSERT (niov > 0);
- }
-
- do {
- LASSERT (nkiov > 0);
- LASSERT (niov > 0);
- this_nob = MIN(kiov->kiov_len - kiovoffset,
- iov->iov_len - iovoffset);
- this_nob = MIN(this_nob, nob);
-
- if (addr == NULL)
- addr = ((char *)cfs_kmap(kiov->kiov_page)) +
- kiov->kiov_offset + kiovoffset;
-
- memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
- nob -= this_nob;
-
- if (kiov->kiov_len > kiovoffset + this_nob) {
- addr += this_nob;
- kiovoffset += this_nob;
- } else {
- cfs_kunmap(kiov->kiov_page);
- addr = NULL;
- kiov++;
- nkiov--;
- kiovoffset = 0;
- }
-
- if (iov->iov_len > iovoffset + this_nob) {
- iovoffset += this_nob;
- } else {
- iov++;
- niov--;
- iovoffset = 0;
- }
- } while (nob > 0);
-
- if (addr != NULL)
- cfs_kunmap(kiov->kiov_page);
+ /* NB kiov, iov are READ-ONLY */
+ unsigned int this_nob;
+ char *addr = NULL;
+
+ if (nob == 0)
+ return;
+
+ LASSERT (!in_interrupt ());
+
+ LASSERT (nkiov > 0);
+ while (kiovoffset >= kiov->kiov_len) {
+ kiovoffset -= kiov->kiov_len;
+ kiov++;
+ nkiov--;
+ LASSERT(nkiov > 0);
+ }
+
+ LASSERT(niov > 0);
+ while (iovoffset >= iov->iov_len) {
+ iovoffset -= iov->iov_len;
+ iov++;
+ niov--;
+ LASSERT(niov > 0);
+ }
+
+ do {
+ LASSERT(nkiov > 0);
+ LASSERT(niov > 0);
+ this_nob = MIN(kiov->kiov_len - kiovoffset,
+ iov->iov_len - iovoffset);
+ this_nob = MIN(this_nob, nob);
+
+ if (addr == NULL)
+ addr = ((char *)kmap(kiov->kiov_page)) +
+ kiov->kiov_offset + kiovoffset;
+
+ memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
+ nob -= this_nob;
+
+ if (kiov->kiov_len > kiovoffset + this_nob) {
+ addr += this_nob;
+ kiovoffset += this_nob;
+ } else {
+ kunmap(kiov->kiov_page);
+ addr = NULL;
+ kiov++;
+ nkiov--;
+ kiovoffset = 0;
+ }
+
+ if (iov->iov_len > iovoffset + this_nob) {
+ iovoffset += this_nob;
+ } else {
+ iov++;
+ niov--;
+ iovoffset = 0;
+ }
+ } while (nob > 0);
+
+ if (addr != NULL)
+ kunmap(kiov->kiov_page);
}
+EXPORT_SYMBOL(lnet_copy_iov2kiov);
int
-lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
- int src_niov, lnet_kiov_t *src,
- unsigned int offset, unsigned int len)
+lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
+ int src_niov, lnet_kiov_t *src,
+ unsigned int offset, unsigned int len)
{
- /* Initialise 'dst' to the subset of 'src' starting at 'offset',
- * for exactly 'len' bytes, and return the number of entries.
- * NB not destructive to 'src' */
- unsigned int frag_len;
- unsigned int niov;
-
- if (len == 0) /* no data => */
- return (0); /* no frags */
-
- LASSERT (src_niov > 0);
- while (offset >= src->kiov_len) { /* skip initial frags */
- offset -= src->kiov_len;
- src_niov--;
- src++;
- LASSERT (src_niov > 0);
- }
-
- niov = 1;
- for (;;) {
- LASSERT (src_niov > 0);
- LASSERT (niov <= dst_niov);
-
- frag_len = src->kiov_len - offset;
- dst->kiov_page = src->kiov_page;
- dst->kiov_offset = src->kiov_offset + offset;
-
- if (len <= frag_len) {
- dst->kiov_len = len;
- LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
- return (niov);
- }
-
- dst->kiov_len = frag_len;
- LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
-
- len -= frag_len;
- dst++;
- src++;
- niov++;
- src_niov--;
- offset = 0;
- }
+ /* Initialise 'dst' to the subset of 'src' starting at 'offset',
+ * for exactly 'len' bytes, and return the number of entries.
+ * NB not destructive to 'src' */
+ unsigned int frag_len;
+ unsigned int niov;
+
+ if (len == 0) /* no data => */
+ return (0); /* no frags */
+
+ LASSERT(src_niov > 0);
+ while (offset >= src->kiov_len) { /* skip initial frags */
+ offset -= src->kiov_len;
+ src_niov--;
+ src++;
+ LASSERT(src_niov > 0);
+ }
+
+ niov = 1;
+ for (;;) {
+ LASSERT(src_niov > 0);
+ LASSERT((int)niov <= dst_niov);
+
+ frag_len = src->kiov_len - offset;
+ dst->kiov_page = src->kiov_page;
+ dst->kiov_offset = src->kiov_offset + offset;
+
+ if (len <= frag_len) {
+ dst->kiov_len = len;
+ LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
+ return niov;
+ }
+
+ dst->kiov_len = frag_len;
+ LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
+
+ len -= frag_len;
+ dst++;
+ src++;
+ niov++;
+ src_niov--;
+ offset = 0;
+ }
}
-#endif
+EXPORT_SYMBOL(lnet_extract_kiov);
void
lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
- unsigned int offset, unsigned int mlen, unsigned int rlen)
+ unsigned int offset, unsigned int mlen, unsigned int rlen)
{
- unsigned int niov = 0;
- struct iovec *iov = NULL;
- lnet_kiov_t *kiov = NULL;
- int rc;
-
- LASSERT (!in_interrupt ());
- LASSERT (mlen == 0 || msg != NULL);
-
- if (msg != NULL) {
- LASSERT(msg->msg_receiving);
- LASSERT(!msg->msg_sending);
- LASSERT(rlen == msg->msg_len);
- LASSERT(mlen <= msg->msg_len);
-
- msg->msg_wanted = mlen;
- msg->msg_offset = offset;
- msg->msg_receiving = 0;
-
- if (mlen != 0) {
- niov = msg->msg_niov;
- iov = msg->msg_iov;
- kiov = msg->msg_kiov;
-
- LASSERT (niov > 0);
- LASSERT ((iov == NULL) != (kiov == NULL));
- }
- }
-
- rc = (ni->ni_lnd->lnd_recv)(ni, private, msg, delayed,
- niov, iov, kiov, offset, mlen, rlen);
- if (rc < 0)
- lnet_finalize(ni, msg, rc);
+ unsigned int niov = 0;
+ struct kvec *iov = NULL;
+ lnet_kiov_t *kiov = NULL;
+ int rc;
+
+ LASSERT (!in_interrupt ());
+ LASSERT (mlen == 0 || msg != NULL);
+
+ if (msg != NULL) {
+ LASSERT(msg->msg_receiving);
+ LASSERT(!msg->msg_sending);
+ LASSERT(rlen == msg->msg_len);
+ LASSERT(mlen <= msg->msg_len);
+ LASSERT(msg->msg_offset == offset);
+ LASSERT(msg->msg_wanted == mlen);
+
+ msg->msg_receiving = 0;
+
+ if (mlen != 0) {
+ niov = msg->msg_niov;
+ iov = msg->msg_iov;
+ kiov = msg->msg_kiov;
+
+ LASSERT (niov > 0);
+ LASSERT ((iov == NULL) != (kiov == NULL));
+ }
+ }
+
+ rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
+ niov, iov, kiov, offset, mlen,
+ rlen);
+ if (rc < 0)
+ lnet_finalize(ni, msg, rc);
}
-int
-lnet_compare_routers(lnet_peer_t *p1, lnet_peer_t *p2)
-{
- if (p1->lp_txqnob < p2->lp_txqnob)
- return 1;
-
- if (p1->lp_txqnob > p2->lp_txqnob)
- return -1;
-
- if (p1->lp_txcredits > p2->lp_txcredits)
- return 1;
-
- if (p1->lp_txcredits < p2->lp_txcredits)
- return -1;
-
- return 0;
-}
-
-
-void
+static void
lnet_setpayloadbuffer(lnet_msg_t *msg)
{
- lnet_libmd_t *md = msg->msg_md;
-
- LASSERT (msg->msg_len > 0);
- LASSERT (!msg->msg_routing);
- LASSERT (md != NULL);
- LASSERT (msg->msg_niov == 0);
- LASSERT (msg->msg_iov == NULL);
- LASSERT (msg->msg_kiov == NULL);
-
- msg->msg_niov = md->md_niov;
- if ((md->md_options & LNET_MD_KIOV) != 0)
- msg->msg_kiov = md->md_iov.kiov;
- else
- msg->msg_iov = md->md_iov.iov;
+ lnet_libmd_t *md = msg->msg_md;
+
+ LASSERT(msg->msg_len > 0);
+ LASSERT(!msg->msg_routing);
+ LASSERT(md != NULL);
+ LASSERT(msg->msg_niov == 0);
+ LASSERT(msg->msg_iov == NULL);
+ LASSERT(msg->msg_kiov == NULL);
+
+ msg->msg_niov = md->md_niov;
+ if ((md->md_options & LNET_MD_KIOV) != 0)
+ msg->msg_kiov = md->md_iov.kiov;
+ else
+ msg->msg_iov = md->md_iov.iov;
}
void
lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
- unsigned int offset, unsigned int len)
+ unsigned int offset, unsigned int len)
{
- msg->msg_type = type;
- msg->msg_target = target;
- msg->msg_len = len;
- msg->msg_offset = offset;
-
- if (len != 0)
- lnet_setpayloadbuffer(msg);
-
- memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
- msg->msg_hdr.type = cpu_to_le32(type);
- msg->msg_hdr.dest_nid = cpu_to_le64(target.nid);
- msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
- /* src_nid will be set later */
- msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
- msg->msg_hdr.payload_length = cpu_to_le32(len);
+ msg->msg_type = type;
+ msg->msg_target = target;
+ msg->msg_len = len;
+ msg->msg_offset = offset;
+
+ if (len != 0)
+ lnet_setpayloadbuffer(msg);
+
+ memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
+ msg->msg_hdr.type = cpu_to_le32(type);
+ msg->msg_hdr.dest_pid = cpu_to_le32(target.pid);
+ /* src_nid will be set later */
+ msg->msg_hdr.src_pid = cpu_to_le32(the_lnet.ln_pid);
+ msg->msg_hdr.payload_length = cpu_to_le32(len);
}
-void
+static void
lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
{
- void *priv = msg->msg_private;
- int rc;
+ void *priv = msg->msg_private;
+ int rc;
- LASSERT (!in_interrupt ());
- LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
- (msg->msg_txcredit && msg->msg_peertxcredit));
+ LASSERT (!in_interrupt ());
+ LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
+ (msg->msg_txcredit && msg->msg_peertxcredit));
- rc = (ni->ni_lnd->lnd_send)(ni, priv, msg);
- if (rc < 0)
- lnet_finalize(ni, msg, rc);
+ rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
+ if (rc < 0)
+ lnet_finalize(ni, msg, rc);
}
-int
-lnet_eager_recv_locked(lnet_msg_t *msg)
+static int
+lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
{
- lnet_peer_t *peer;
- lnet_ni_t *ni;
- int rc = 0;
-
- LASSERT (!msg->msg_delayed);
- msg->msg_delayed = 1;
-
- LASSERT (msg->msg_receiving);
- LASSERT (!msg->msg_sending);
-
- peer = msg->msg_rxpeer;
- ni = peer->lp_ni;
-
- if (ni->ni_lnd->lnd_eager_recv != NULL) {
- LNET_UNLOCK();
-
- rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
- &msg->msg_private);
- if (rc != 0) {
- CERROR("recv from %s / send to %s aborted: "
- "eager_recv failed %d\n",
- libcfs_nid2str(peer->lp_nid),
- libcfs_id2str(msg->msg_target), rc);
- LASSERT (rc < 0); /* required by my callers */
- }
-
- LNET_LOCK();
- }
-
- return rc;
+ int rc;
+
+ LASSERT(!msg->msg_sending);
+ LASSERT(msg->msg_receiving);
+ LASSERT(!msg->msg_rx_ready_delay);
+ LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
+
+ msg->msg_rx_ready_delay = 1;
+ rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
+ &msg->msg_private);
+ if (rc != 0) {
+ CERROR("recv from %s / send to %s aborted: "
+ "eager_recv failed %d\n",
+ libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
+ libcfs_id2str(msg->msg_target), rc);
+ LASSERT(rc < 0); /* required by my callers */
+ }
+
+ return rc;
}
-int
-lnet_post_send_locked (lnet_msg_t *msg, int do_send)
+/*
+ * This function can be called from two paths:
+ * 1. when sending a message
+ * 2. when decommiting a message (lnet_msg_decommit_tx())
+ * In both these cases the peer_ni should have it's reference count
+ * acquired by the caller and therefore it is safe to drop the spin
+ * lock before calling lnd_query()
+ */
+static void
+lnet_ni_query_locked(lnet_ni_t *ni, struct lnet_peer_ni *lp)
{
- /* lnet_send is going to LNET_UNLOCK immediately after this, so it sets
- * do_send FALSE and I don't do the unlock/send/lock bit. I return
- * EAGAIN if msg blocked and 0 if sent or OK to send */
- lnet_peer_t *lp = msg->msg_txpeer;
- lnet_ni_t *ni = lp->lp_ni;
-
- /* non-lnet_send() callers have checked before */
- LASSERT (!do_send || msg->msg_delayed);
- LASSERT (!msg->msg_receiving);
-
- if (!msg->msg_peertxcredit) {
- LASSERT ((lp->lp_txcredits < 0) == !list_empty(&lp->lp_txq));
-
- msg->msg_peertxcredit = 1;
- lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
- lp->lp_txcredits--;
-
- if (lp->lp_txcredits < lp->lp_mintxcredits)
- lp->lp_mintxcredits = lp->lp_txcredits;
-
- if (lp->lp_txcredits < 0) {
- msg->msg_delayed = 1;
- list_add_tail (&msg->msg_list, &lp->lp_txq);
- return EAGAIN;
- }
- }
+ cfs_time_t last_alive = 0;
+ int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
- if (!msg->msg_txcredit) {
- LASSERT ((ni->ni_txcredits < 0) == !list_empty(&ni->ni_txq));
+ LASSERT(lnet_peer_aliveness_enabled(lp));
+ LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
- msg->msg_txcredit = 1;
- ni->ni_txcredits--;
+ lnet_net_unlock(cpt);
+ (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
+ lnet_net_lock(cpt);
- if (ni->ni_txcredits < ni->ni_mintxcredits)
- ni->ni_mintxcredits = ni->ni_txcredits;
+ lp->lpni_last_query = cfs_time_current();
- if (ni->ni_txcredits < 0) {
- msg->msg_delayed = 1;
- list_add_tail (&msg->msg_list, &ni->ni_txq);
- return EAGAIN;
- }
- }
-
- if (do_send) {
- LNET_UNLOCK();
- lnet_ni_send(ni, msg);
- LNET_LOCK();
- }
- return 0;
+ if (last_alive != 0) /* NI has updated timestamp */
+ lp->lpni_last_alive = last_alive;
}
-#ifdef __KERNEL__
-static void
-lnet_commit_routedmsg (lnet_msg_t *msg)
+/* NB: always called with lnet_net_lock held */
+static inline int
+lnet_peer_is_alive (struct lnet_peer_ni *lp, cfs_time_t now)
{
- /* ALWAYS called holding the LNET_LOCK */
- LASSERT (msg->msg_routing);
-
- the_lnet.ln_counters.msgs_alloc++;
- if (the_lnet.ln_counters.msgs_alloc >
- the_lnet.ln_counters.msgs_max)
- the_lnet.ln_counters.msgs_max =
- the_lnet.ln_counters.msgs_alloc;
-
- the_lnet.ln_counters.route_count++;
- the_lnet.ln_counters.route_length += msg->msg_len;
-
- LASSERT (!msg->msg_onactivelist);
- msg->msg_onactivelist = 1;
- list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs);
+ int alive;
+ cfs_time_t deadline;
+
+ LASSERT (lnet_peer_aliveness_enabled(lp));
+
+ /*
+ * Trust lnet_notify() if it has more recent aliveness news, but
+ * ignore the initial assumed death (see lnet_peers_start_down()).
+ */
+ spin_lock(&lp->lpni_lock);
+ if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
+ cfs_time_aftereq(lp->lpni_timestamp, lp->lpni_last_alive)) {
+ spin_unlock(&lp->lpni_lock);
+ return 0;
+ }
+
+ deadline =
+ cfs_time_add(lp->lpni_last_alive,
+ cfs_time_seconds(lp->lpni_net->net_tunables.
+ lct_peer_timeout));
+ alive = cfs_time_after(deadline, now);
+
+ /*
+ * Update obsolete lp_alive except for routers assumed to be dead
+ * initially, because router checker would update aliveness in this
+ * case, and moreover lpni_last_alive at peer creation is assumed.
+ */
+ if (alive && !lp->lpni_alive &&
+ !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) {
+ spin_unlock(&lp->lpni_lock);
+ lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
+ } else {
+ spin_unlock(&lp->lpni_lock);
+ }
+
+ return alive;
}
-lnet_rtrbufpool_t *
-lnet_msg2bufpool(lnet_msg_t *msg)
-{
- lnet_rtrbufpool_t *rbp = &the_lnet.ln_rtrpools[0];
- LASSERT (msg->msg_len <= LNET_MTU);
- while (msg->msg_len > rbp->rbp_npages * CFS_PAGE_SIZE) {
- rbp++;
- LASSERT (rbp < &the_lnet.ln_rtrpools[LNET_NRBPOOLS]);
- }
-
- return rbp;
+/* NB: returns 1 when alive, 0 when dead, negative when error;
+ * may drop the lnet_net_lock */
+static int
+lnet_peer_alive_locked (struct lnet_ni *ni, struct lnet_peer_ni *lp)
+{
+ cfs_time_t now = cfs_time_current();
+
+ if (!lnet_peer_aliveness_enabled(lp))
+ return -ENODEV;
+
+ if (lnet_peer_is_alive(lp, now))
+ return 1;
+
+ /*
+ * Peer appears dead, but we should avoid frequent NI queries (at
+ * most once per lnet_queryinterval seconds).
+ */
+ if (lp->lpni_last_query != 0) {
+ static const int lnet_queryinterval = 1;
+
+ cfs_time_t next_query =
+ cfs_time_add(lp->lpni_last_query,
+ cfs_time_seconds(lnet_queryinterval));
+
+ if (cfs_time_before(now, next_query)) {
+ if (lp->lpni_alive)
+ CWARN("Unexpected aliveness of peer %s: "
+ "%d < %d (%d/%d)\n",
+ libcfs_nid2str(lp->lpni_nid),
+ (int)now, (int)next_query,
+ lnet_queryinterval,
+ lp->lpni_net->net_tunables.lct_peer_timeout);
+ return 0;
+ }
+ }
+
+ /* query NI for latest aliveness news */
+ lnet_ni_query_locked(ni, lp);
+
+ if (lnet_peer_is_alive(lp, now))
+ return 1;
+
+ lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
+ return 0;
}
-int
-lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
+/**
+ * \param msg The message to be sent.
+ * \param do_send True if lnet_ni_send() should be called in this function.
+ * lnet_send() is going to lnet_net_unlock immediately after this, so
+ * it sets do_send FALSE and I don't do the unlock/send/lock bit.
+ *
+ * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
+ * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
+ * \retval -EHOSTUNREACH If the next hop of the message appears dead.
+ * \retval -ECANCELED If the MD of the message has been unlinked.
+ */
+static int
+lnet_post_send_locked(lnet_msg_t *msg, int do_send)
{
- /* lnet_parse is going to LNET_UNLOCK immediately after this, so it
- * sets do_recv FALSE and I don't do the unlock/send/lock bit. I
- * return EAGAIN if msg blocked and 0 if sent or OK to send */
- lnet_peer_t *lp = msg->msg_rxpeer;
- lnet_rtrbufpool_t *rbp;
- lnet_rtrbuf_t *rb;
-
- LASSERT (msg->msg_iov == NULL);
- LASSERT (msg->msg_kiov == NULL);
- LASSERT (msg->msg_niov == 0);
- LASSERT (msg->msg_routing);
- LASSERT (msg->msg_receiving);
- LASSERT (!msg->msg_sending);
-
- /* non-lnet_parse callers only send delayed messages */
- LASSERT (!do_recv || msg->msg_delayed);
-
- if (!msg->msg_peerrtrcredit) {
- LASSERT ((lp->lp_rtrcredits < 0) == !list_empty(&lp->lp_rtrq));
-
- msg->msg_peerrtrcredit = 1;
- lp->lp_rtrcredits--;
- if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
- lp->lp_minrtrcredits = lp->lp_rtrcredits;
-
- if (lp->lp_rtrcredits < 0) {
- /* must have checked eager_recv before here */
- LASSERT (msg->msg_delayed);
- list_add_tail(&msg->msg_list, &lp->lp_rtrq);
- return EAGAIN;
- }
- }
+ struct lnet_peer_ni *lp = msg->msg_txpeer;
+ struct lnet_ni *ni = msg->msg_txni;
+ int cpt = msg->msg_tx_cpt;
+ struct lnet_tx_queue *tq = ni->ni_tx_queues[cpt];
+
+ /* non-lnet_send() callers have checked before */
+ LASSERT(!do_send || msg->msg_tx_delayed);
+ LASSERT(!msg->msg_receiving);
+ LASSERT(msg->msg_tx_committed);
+
+ /* NB 'lp' is always the next hop */
+ if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
+ lnet_peer_alive_locked(ni, lp) == 0) {
+ the_lnet.ln_counters[cpt]->drop_count++;
+ the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
+ lnet_net_unlock(cpt);
+ if (msg->msg_txpeer)
+ atomic_inc(&msg->msg_txpeer->lpni_stats.drop_count);
+ if (msg->msg_txni)
+ atomic_inc(&msg->msg_txni->ni_stats.drop_count);
+
+ CNETERR("Dropping message for %s: peer not alive\n",
+ libcfs_id2str(msg->msg_target));
+ if (do_send)
+ lnet_finalize(ni, msg, -EHOSTUNREACH);
+
+ lnet_net_lock(cpt);
+ return -EHOSTUNREACH;
+ }
+
+ if (msg->msg_md != NULL &&
+ (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
+ lnet_net_unlock(cpt);
+
+ CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
+ "called on the MD/ME.\n",
+ libcfs_id2str(msg->msg_target));
+ if (do_send)
+ lnet_finalize(ni, msg, -ECANCELED);
+
+ lnet_net_lock(cpt);
+ return -ECANCELED;
+ }
+
+ if (!msg->msg_peertxcredit) {
+ spin_lock(&lp->lpni_lock);
+ LASSERT((lp->lpni_txcredits < 0) ==
+ !list_empty(&lp->lpni_txq));
+
+ msg->msg_peertxcredit = 1;
+ lp->lpni_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
+ lp->lpni_txcredits--;
+
+ if (lp->lpni_txcredits < lp->lpni_mintxcredits)
+ lp->lpni_mintxcredits = lp->lpni_txcredits;
+
+ if (lp->lpni_txcredits < 0) {
+ msg->msg_tx_delayed = 1;
+ list_add_tail(&msg->msg_list, &lp->lpni_txq);
+ spin_unlock(&lp->lpni_lock);
+ return LNET_CREDIT_WAIT;
+ }
+ spin_unlock(&lp->lpni_lock);
+ }
+
+ if (!msg->msg_txcredit) {
+ LASSERT((tq->tq_credits < 0) ==
+ !list_empty(&tq->tq_delayed));
+
+ msg->msg_txcredit = 1;
+ tq->tq_credits--;
+ atomic_dec(&ni->ni_tx_credits);
+
+ if (tq->tq_credits < tq->tq_credits_min)
+ tq->tq_credits_min = tq->tq_credits;
+
+ if (tq->tq_credits < 0) {
+ msg->msg_tx_delayed = 1;
+ list_add_tail(&msg->msg_list, &tq->tq_delayed);
+ return LNET_CREDIT_WAIT;
+ }
+ }
+
+ if (do_send) {
+ lnet_net_unlock(cpt);
+ lnet_ni_send(ni, msg);
+ lnet_net_lock(cpt);
+ }
+ return LNET_CREDIT_OK;
+}
- rbp = lnet_msg2bufpool(msg);
- if (!msg->msg_rtrcredit) {
- LASSERT ((rbp->rbp_credits < 0) == !list_empty(&rbp->rbp_msgs));
+static lnet_rtrbufpool_t *
+lnet_msg2bufpool(lnet_msg_t *msg)
+{
+ lnet_rtrbufpool_t *rbp;
+ int cpt;
- msg->msg_rtrcredit = 1;
- rbp->rbp_credits--;
- if (rbp->rbp_credits < rbp->rbp_mincredits)
- rbp->rbp_mincredits = rbp->rbp_credits;
+ LASSERT(msg->msg_rx_committed);
- if (rbp->rbp_credits < 0) {
- /* must have checked eager_recv before here */
- LASSERT (msg->msg_delayed);
- list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
- return EAGAIN;
- }
- }
+ cpt = msg->msg_rx_cpt;
+ rbp = &the_lnet.ln_rtrpools[cpt][0];
- LASSERT (!list_empty(&rbp->rbp_bufs));
- rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
- list_del(&rb->rb_list);
+ LASSERT(msg->msg_len <= LNET_MTU);
+ while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
+ rbp++;
+ LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
+ }
- msg->msg_niov = rbp->rbp_npages;
- msg->msg_kiov = &rb->rb_kiov[0];
+ return rbp;
+}
- if (do_recv) {
- LNET_UNLOCK();
- lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
- 0, msg->msg_len, msg->msg_len);
- LNET_LOCK();
- }
- return 0;
+static int
+lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
+{
+ /* lnet_parse is going to lnet_net_unlock immediately after this, so it
+ * sets do_recv FALSE and I don't do the unlock/send/lock bit.
+ * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
+ * received or OK to receive */
+ struct lnet_peer_ni *lp = msg->msg_rxpeer;
+ lnet_rtrbufpool_t *rbp;
+ lnet_rtrbuf_t *rb;
+
+ LASSERT (msg->msg_iov == NULL);
+ LASSERT (msg->msg_kiov == NULL);
+ LASSERT (msg->msg_niov == 0);
+ LASSERT (msg->msg_routing);
+ LASSERT (msg->msg_receiving);
+ LASSERT (!msg->msg_sending);
+
+ /* non-lnet_parse callers only receive delayed messages */
+ LASSERT(!do_recv || msg->msg_rx_delayed);
+
+ if (!msg->msg_peerrtrcredit) {
+ spin_lock(&lp->lpni_lock);
+ LASSERT((lp->lpni_rtrcredits < 0) ==
+ !list_empty(&lp->lpni_rtrq));
+
+ msg->msg_peerrtrcredit = 1;
+ lp->lpni_rtrcredits--;
+ if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
+ lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
+
+ if (lp->lpni_rtrcredits < 0) {
+ /* must have checked eager_recv before here */
+ LASSERT(msg->msg_rx_ready_delay);
+ msg->msg_rx_delayed = 1;
+ list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
+ spin_unlock(&lp->lpni_lock);
+ return LNET_CREDIT_WAIT;
+ }
+ spin_unlock(&lp->lpni_lock);
+ }
+
+ rbp = lnet_msg2bufpool(msg);
+
+ if (!msg->msg_rtrcredit) {
+ msg->msg_rtrcredit = 1;
+ rbp->rbp_credits--;
+ if (rbp->rbp_credits < rbp->rbp_mincredits)
+ rbp->rbp_mincredits = rbp->rbp_credits;
+
+ if (rbp->rbp_credits < 0) {
+ /* must have checked eager_recv before here */
+ LASSERT(msg->msg_rx_ready_delay);
+ msg->msg_rx_delayed = 1;
+ list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
+ return LNET_CREDIT_WAIT;
+ }
+ }
+
+ LASSERT(!list_empty(&rbp->rbp_bufs));
+ rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
+ list_del(&rb->rb_list);
+
+ msg->msg_niov = rbp->rbp_npages;
+ msg->msg_kiov = &rb->rb_kiov[0];
+
+ if (do_recv) {
+ int cpt = msg->msg_rx_cpt;
+
+ lnet_net_unlock(cpt);
+ lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
+ 0, msg->msg_len, msg->msg_len);
+ lnet_net_lock(cpt);
+ }
+ return LNET_CREDIT_OK;
}
-#endif
void
-lnet_return_credits_locked (lnet_msg_t *msg)
+lnet_return_tx_credits_locked(lnet_msg_t *msg)
{
- lnet_peer_t *txpeer = msg->msg_txpeer;
- lnet_peer_t *rxpeer = msg->msg_rxpeer;
- lnet_msg_t *msg2;
- lnet_ni_t *ni;
-
- if (msg->msg_txcredit) {
- /* give back NI txcredits */
- msg->msg_txcredit = 0;
- ni = txpeer->lp_ni;
-
- LASSERT((ni->ni_txcredits < 0) == !list_empty(&ni->ni_txq));
-
- ni->ni_txcredits++;
- if (ni->ni_txcredits <= 0) {
- msg2 = list_entry(ni->ni_txq.next, lnet_msg_t, msg_list);
- list_del(&msg2->msg_list);
-
- LASSERT(msg2->msg_txpeer->lp_ni == ni);
- LASSERT(msg2->msg_delayed);
-
- (void) lnet_post_send_locked(msg2, 1);
- }
- }
-
- if (msg->msg_peertxcredit) {
- /* give back peer txcredits */
- msg->msg_peertxcredit = 0;
-
- LASSERT((txpeer->lp_txcredits < 0) == !list_empty(&txpeer->lp_txq));
-
- txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
- LASSERT (txpeer->lp_txqnob >= 0);
-
- txpeer->lp_txcredits++;
- if (txpeer->lp_txcredits <= 0) {
- msg2 = list_entry(txpeer->lp_txq.next,
- lnet_msg_t, msg_list);
- list_del(&msg2->msg_list);
-
- LASSERT (msg2->msg_txpeer == txpeer);
- LASSERT (msg2->msg_delayed);
-
+ struct lnet_peer_ni *txpeer = msg->msg_txpeer;
+ struct lnet_ni *txni = msg->msg_txni;
+ lnet_msg_t *msg2;
+
+ if (msg->msg_txcredit) {
+ struct lnet_ni *ni = msg->msg_txni;
+ struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
+
+ /* give back NI txcredits */
+ msg->msg_txcredit = 0;
+
+ LASSERT((tq->tq_credits < 0) ==
+ !list_empty(&tq->tq_delayed));
+
+ tq->tq_credits++;
+ atomic_inc(&ni->ni_tx_credits);
+ if (tq->tq_credits <= 0) {
+ msg2 = list_entry(tq->tq_delayed.next,
+ lnet_msg_t, msg_list);
+ list_del(&msg2->msg_list);
+
+ LASSERT(msg2->msg_txni == ni);
+ LASSERT(msg2->msg_tx_delayed);
+ LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
+
+ (void) lnet_post_send_locked(msg2, 1);
+ }
+ }
+
+ if (msg->msg_peertxcredit) {
+ /* give back peer txcredits */
+ msg->msg_peertxcredit = 0;
+
+ spin_lock(&txpeer->lpni_lock);
+ LASSERT((txpeer->lpni_txcredits < 0) ==
+ !list_empty(&txpeer->lpni_txq));
+
+ txpeer->lpni_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
+ LASSERT(txpeer->lpni_txqnob >= 0);
+
+ txpeer->lpni_txcredits++;
+ if (txpeer->lpni_txcredits <= 0) {
+ msg2 = list_entry(txpeer->lpni_txq.next,
+ lnet_msg_t, msg_list);
+ list_del(&msg2->msg_list);
+ spin_unlock(&txpeer->lpni_lock);
+
+ LASSERT(msg2->msg_txpeer == txpeer);
+ LASSERT(msg2->msg_tx_delayed);
+
+ if (msg2->msg_tx_cpt != msg->msg_tx_cpt) {
+ lnet_net_unlock(msg->msg_tx_cpt);
+ lnet_net_lock(msg2->msg_tx_cpt);
+ }
(void) lnet_post_send_locked(msg2, 1);
- }
- }
-
- if (txpeer != NULL) {
- msg->msg_txpeer = NULL;
- lnet_peer_decref_locked(txpeer);
- }
-
-#ifdef __KERNEL__
- if (msg->msg_rtrcredit) {
- /* give back global router credits */
- lnet_rtrbuf_t *rb;
- lnet_rtrbufpool_t *rbp;
-
- /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
- * there until it gets one allocated, or aborts the wait
- * itself */
- LASSERT (msg->msg_kiov != NULL);
-
- rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
- rbp = rb->rb_pool;
- LASSERT (rbp == lnet_msg2bufpool(msg));
-
- msg->msg_kiov = NULL;
- msg->msg_rtrcredit = 0;
-
- LASSERT((rbp->rbp_credits < 0) == !list_empty(&rbp->rbp_msgs));
- LASSERT((rbp->rbp_credits > 0) == !list_empty(&rbp->rbp_bufs));
-
- list_add(&rb->rb_list, &rbp->rbp_bufs);
- rbp->rbp_credits++;
- if (rbp->rbp_credits <= 0) {
- msg2 = list_entry(rbp->rbp_msgs.next,
- lnet_msg_t, msg_list);
- list_del(&msg2->msg_list);
-
- (void) lnet_post_routed_recv_locked(msg2, 1);
- }
- }
-
- if (msg->msg_peerrtrcredit) {
- /* give back peer router credits */
- msg->msg_peerrtrcredit = 0;
-
- LASSERT((rxpeer->lp_rtrcredits < 0) == !list_empty(&rxpeer->lp_rtrq));
-
- rxpeer->lp_rtrcredits++;
- if (rxpeer->lp_rtrcredits <= 0) {
- msg2 = list_entry(rxpeer->lp_rtrq.next,
- lnet_msg_t, msg_list);
- list_del(&msg2->msg_list);
-
- (void) lnet_post_routed_recv_locked(msg2, 1);
- }
- }
-#else
- LASSERT (!msg->msg_rtrcredit);
- LASSERT (!msg->msg_peerrtrcredit);
-#endif
- if (rxpeer != NULL) {
- msg->msg_rxpeer = NULL;
- lnet_peer_decref_locked(rxpeer);
- }
+ if (msg2->msg_tx_cpt != msg->msg_tx_cpt) {
+ lnet_net_unlock(msg2->msg_tx_cpt);
+ lnet_net_lock(msg->msg_tx_cpt);
+ }
+ } else {
+ spin_unlock(&txpeer->lpni_lock);
+ }
+ }
+
+ if (txni != NULL) {
+ msg->msg_txni = NULL;
+ lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
+ }
+
+ if (txpeer != NULL) {
+ /*
+ * TODO:
+ * Once the patch for the health comes in we need to set
+ * the health of the peer ni to bad when we fail to send
+ * a message.
+ * int status = msg->msg_ev.status;
+ * if (status != 0)
+ * lnet_set_peer_ni_health_locked(txpeer, false)
+ */
+ msg->msg_txpeer = NULL;
+ lnet_peer_ni_decref_locked(txpeer);
+ }
}
-int
-lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg)
+void
+lnet_schedule_blocked_locked(lnet_rtrbufpool_t *rbp)
{
- lnet_nid_t dst_nid = msg->msg_target.nid;
- lnet_ni_t *src_ni;
- lnet_ni_t *local_ni;
- lnet_remotenet_t *rnet;
- lnet_route_t *route;
- lnet_route_t *best_route;
- struct list_head *tmp;
- lnet_peer_t *lp;
- lnet_peer_t *lp2;
- int rc;
-
- LASSERT (msg->msg_txpeer == NULL);
- LASSERT (!msg->msg_sending);
- LASSERT (!msg->msg_target_is_router);
- LASSERT (!msg->msg_receiving);
-
- msg->msg_sending = 1;
-
- /* NB! ni != NULL == interface pre-determined (ACK/REPLY) */
-
- LNET_LOCK();
-
- if (the_lnet.ln_shutdown) {
- LNET_UNLOCK();
- return -ESHUTDOWN;
- }
+ lnet_msg_t *msg;
- if (src_nid == LNET_NID_ANY) {
- src_ni = NULL;
- } else {
- src_ni = lnet_nid2ni_locked(src_nid);
- if (src_ni == NULL) {
- LNET_UNLOCK();
- CERROR("Can't send to %s: src %s is not a local nid\n",
- libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
- return -EINVAL;
- }
- LASSERT (!msg->msg_routing);
- }
-
- /* Is this for someone on a local network? */
- local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid));
-
- if (local_ni != NULL) {
- if (src_ni == NULL) {
- src_ni = local_ni;
- src_nid = src_ni->ni_nid;
- } else if (src_ni == local_ni) {
- lnet_ni_decref_locked(local_ni);
- } else {
- lnet_ni_decref_locked(local_ni);
- lnet_ni_decref_locked(src_ni);
- LNET_UNLOCK();
- CERROR("no route to %s via from %s\n",
- libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
- return -EINVAL;
- }
-
- LASSERT (src_nid != LNET_NID_ANY);
-
- if (!msg->msg_routing) {
- src_nid = lnet_ptlcompat_srcnid(src_nid, dst_nid);
- msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
- }
-
- if (src_ni == the_lnet.ln_loni) {
- /* No send credit hassles with LOLND */
- LNET_UNLOCK();
- lnet_ni_send(src_ni, msg);
- lnet_ni_decref(src_ni);
- return 0;
- }
-
- rc = lnet_nid2peer_locked(&lp, dst_nid);
- lnet_ni_decref_locked(src_ni); /* lp has ref on src_ni; lose mine */
- if (rc != 0) {
- LNET_UNLOCK();
- CERROR("Error %d finding peer %s\n", rc,
- libcfs_nid2str(dst_nid));
- /* ENOMEM or shutting down */
- return rc;
- }
- LASSERT (lp->lp_ni == src_ni);
- } else {
- /* sending to a remote network */
- rnet = lnet_find_net_locked(LNET_NIDNET(dst_nid));
- if (rnet == NULL) {
- if (src_ni != NULL)
- lnet_ni_decref_locked(src_ni);
- LNET_UNLOCK();
- CERROR("No route to %s\n", libcfs_id2str(msg->msg_target));
- return -EHOSTUNREACH;
- }
-
- /* Find the best gateway I can use */
- lp = NULL;
- best_route = NULL;
- list_for_each(tmp, &rnet->lrn_routes) {
- route = list_entry(tmp, lnet_route_t, lr_list);
- lp2 = route->lr_gateway;
-
- if (lp2->lp_alive &&
- (src_ni == NULL || lp2->lp_ni == src_ni) &&
- (lp == NULL || lnet_compare_routers(lp2, lp) > 0)) {
- best_route = route;
- lp = lp2;
- }
- }
-
- if (lp == NULL) {
- if (src_ni != NULL)
- lnet_ni_decref_locked(src_ni);
- LNET_UNLOCK();
- CERROR("No route to %s (all routers down)\n",
- libcfs_id2str(msg->msg_target));
- return -EHOSTUNREACH;
- }
-
- /* Place selected route at the end of the route list to ensure
- * fairness; everything else being equal... */
- list_del(&best_route->lr_list);
- list_add_tail(&best_route->lr_list, &rnet->lrn_routes);
-
- if (src_ni == NULL) {
- src_ni = lp->lp_ni;
- src_nid = src_ni->ni_nid;
- } else {
- LASSERT (src_ni == lp->lp_ni);
- lnet_ni_decref_locked(src_ni);
- }
-
- lnet_peer_addref_locked(lp);
-
- LASSERT (src_nid != LNET_NID_ANY);
-
- if (!msg->msg_routing) {
- /* I'm the source and now I know which NI to send on */
- src_nid = lnet_ptlcompat_srcnid(src_nid, dst_nid);
- msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
- }
-
- msg->msg_target_is_router = 1;
- msg->msg_target.nid = lp->lp_nid;
- msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
- }
-
- /* 'lp' is our best choice of peer */
-
- LASSERT (!msg->msg_peertxcredit);
- LASSERT (!msg->msg_txcredit);
- LASSERT (msg->msg_txpeer == NULL);
-
- msg->msg_txpeer = lp; /* msg takes my ref on lp */
-
- rc = lnet_post_send_locked(msg, 0);
- LNET_UNLOCK();
+ if (list_empty(&rbp->rbp_msgs))
+ return;
+ msg = list_entry(rbp->rbp_msgs.next,
+ lnet_msg_t, msg_list);
+ list_del(&msg->msg_list);
- if (rc == 0)
- lnet_ni_send(src_ni, msg);
-
- return 0;
+ (void)lnet_post_routed_recv_locked(msg, 1);
}
-static void
-lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg)
+void
+lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
{
- /* ALWAYS called holding the LNET_LOCK */
- /* Here, we commit the MD to a network OP by marking it busy and
- * decrementing its threshold. Come what may, the network "owns"
- * the MD until a call to lnet_finalize() signals completion. */
- LASSERT (!msg->msg_routing);
-
- msg->msg_md = md;
-
- md->md_refcount++;
- if (md->md_threshold != LNET_MD_THRESH_INF) {
- LASSERT (md->md_threshold > 0);
- md->md_threshold--;
- }
+ lnet_msg_t *msg;
+ lnet_msg_t *tmp;
- the_lnet.ln_counters.msgs_alloc++;
- if (the_lnet.ln_counters.msgs_alloc >
- the_lnet.ln_counters.msgs_max)
- the_lnet.ln_counters.msgs_max =
- the_lnet.ln_counters.msgs_alloc;
+ lnet_net_unlock(cpt);
- LASSERT (!msg->msg_onactivelist);
- msg->msg_onactivelist = 1;
- list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs);
-}
+ list_for_each_entry_safe(msg, tmp, list, msg_list) {
+ lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
+ 0, 0, 0, msg->msg_hdr.payload_length);
+ list_del_init(&msg->msg_list);
+ lnet_finalize(NULL, msg, -ECANCELED);
+ }
-static void
-lnet_drop_message (lnet_ni_t *ni, void *private, unsigned int nob)
-{
- LNET_LOCK();
- the_lnet.ln_counters.drop_count++;
- the_lnet.ln_counters.drop_length += nob;
- LNET_UNLOCK();
-
- lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
+ lnet_net_lock(cpt);
}
-static void
-lnet_drop_delayed_put(lnet_msg_t *msg, char *reason)
+void
+lnet_return_rx_credits_locked(lnet_msg_t *msg)
{
- LASSERT (msg->msg_md == NULL);
- LASSERT (msg->msg_delayed);
- LASSERT (msg->msg_rxpeer != NULL);
- LASSERT (msg->msg_hdr.type == LNET_MSG_PUT);
-
- CWARN("Dropping delayed PUT from %s portal %d match "LPU64
- " offset %d length %d: %s\n",
- libcfs_id2str((lnet_process_id_t){
- .nid = msg->msg_hdr.src_nid,
- .pid = msg->msg_hdr.src_pid}),
- msg->msg_hdr.msg.put.ptl_index,
- msg->msg_hdr.msg.put.match_bits,
- msg->msg_hdr.msg.put.offset,
- msg->msg_hdr.payload_length,
- reason);
-
- /* NB I can't drop msg's ref on msg_rxpeer until after I've
- * called lnet_drop_message(), so I just hang onto msg as well
- * until that's done */
-
- lnet_drop_message(msg->msg_rxpeer->lp_ni,
- msg->msg_private, msg->msg_len);
-
- LNET_LOCK();
-
- lnet_peer_decref_locked(msg->msg_rxpeer);
- msg->msg_rxpeer = NULL;
-
- lnet_msg_free(msg);
-
- LNET_UNLOCK();
+ struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
+ struct lnet_ni *rxni = msg->msg_rxni;
+ lnet_msg_t *msg2;
+
+ if (msg->msg_rtrcredit) {
+ /* give back global router credits */
+ lnet_rtrbuf_t *rb;
+ lnet_rtrbufpool_t *rbp;
+
+ /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
+ * there until it gets one allocated, or aborts the wait
+ * itself */
+ LASSERT(msg->msg_kiov != NULL);
+
+ rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
+ rbp = rb->rb_pool;
+
+ msg->msg_kiov = NULL;
+ msg->msg_rtrcredit = 0;
+
+ LASSERT(rbp == lnet_msg2bufpool(msg));
+
+ LASSERT((rbp->rbp_credits > 0) ==
+ !list_empty(&rbp->rbp_bufs));
+
+ /* If routing is now turned off, we just drop this buffer and
+ * don't bother trying to return credits. */
+ if (!the_lnet.ln_routing) {
+ lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
+ goto routing_off;
+ }
+
+ /* It is possible that a user has lowered the desired number of
+ * buffers in this pool. Make sure we never put back
+ * more buffers than the stated number. */
+ if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
+ /* Discard this buffer so we don't have too
+ * many. */
+ lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
+ rbp->rbp_nbuffers--;
+ } else {
+ list_add(&rb->rb_list, &rbp->rbp_bufs);
+ rbp->rbp_credits++;
+ if (rbp->rbp_credits <= 0)
+ lnet_schedule_blocked_locked(rbp);
+ }
+ }
+
+routing_off:
+ if (msg->msg_peerrtrcredit) {
+ /* give back peer router credits */
+ msg->msg_peerrtrcredit = 0;
+
+ spin_lock(&rxpeer->lpni_lock);
+ LASSERT((rxpeer->lpni_rtrcredits < 0) ==
+ !list_empty(&rxpeer->lpni_rtrq));
+
+ rxpeer->lpni_rtrcredits++;
+
+ /* drop all messages which are queued to be routed on that
+ * peer. */
+ if (!the_lnet.ln_routing) {
+ struct list_head drop;
+ INIT_LIST_HEAD(&drop);
+ list_splice_init(&rxpeer->lpni_rtrq, &drop);
+ spin_unlock(&rxpeer->lpni_lock);
+ lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
+ } else if (rxpeer->lpni_rtrcredits <= 0) {
+ msg2 = list_entry(rxpeer->lpni_rtrq.next,
+ lnet_msg_t, msg_list);
+ list_del(&msg2->msg_list);
+ spin_unlock(&rxpeer->lpni_lock);
+ (void) lnet_post_routed_recv_locked(msg2, 1);
+ } else {
+ spin_unlock(&rxpeer->lpni_lock);
+ }
+ }
+ if (rxni != NULL) {
+ msg->msg_rxni = NULL;
+ lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
+ }
+ if (rxpeer != NULL) {
+ msg->msg_rxpeer = NULL;
+ lnet_peer_ni_decref_locked(rxpeer);
+ }
}
-int
-LNetSetLazyPortal(int portal)
+static int
+lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
{
- lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
-
- if (portal < 0 || portal >= the_lnet.ln_nportals)
- return -EINVAL;
+ if (p1->lpni_txqnob < p2->lpni_txqnob)
+ return 1;
- CDEBUG(D_NET, "Setting portal %d lazy\n", portal);
+ if (p1->lpni_txqnob > p2->lpni_txqnob)
+ return -1;
- LNET_LOCK();
+ if (p1->lpni_txcredits > p2->lpni_txcredits)
+ return 1;
- ptl->ptl_options |= LNET_PTL_LAZY;
+ if (p1->lpni_txcredits < p2->lpni_txcredits)
+ return -1;
- LNET_UNLOCK();
-
- return 0;
+ return 0;
}
-int
-LNetClearLazyPortal(int portal)
+static int
+lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
{
- struct list_head zombies;
- lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
- lnet_msg_t *msg;
+ struct lnet_peer_ni *p1 = r1->lr_gateway;
+ struct lnet_peer_ni *p2 = r2->lr_gateway;
+ int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
+ int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
+ int rc;
- if (portal < 0 || portal >= the_lnet.ln_nportals)
- return -EINVAL;
-
- LNET_LOCK();
-
- if ((ptl->ptl_options & LNET_PTL_LAZY) == 0) {
- LNET_UNLOCK();
- return 0;
- }
+ if (r1->lr_priority < r2->lr_priority)
+ return 1;
- 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 (r1->lr_priority > r2->lr_priority)
+ return -1;
- /* grab all the blocked messages atomically */
- list_add(&zombies, &ptl->ptl_msgq);
- list_del_init(&ptl->ptl_msgq);
+ if (r1_hops < r2_hops)
+ return 1;
- ptl->ptl_msgq_version++;
- ptl->ptl_options &= ~LNET_PTL_LAZY;
+ if (r1_hops > r2_hops)
+ return -1;
- LNET_UNLOCK();
+ rc = lnet_compare_peers(p1, p2);
+ if (rc)
+ return rc;
- while (!list_empty(&zombies)) {
- msg = list_entry(zombies.next, lnet_msg_t, msg_list);
- list_del(&msg->msg_list);
+ if (r1->lr_seq - r2->lr_seq <= 0)
+ return 1;
- lnet_drop_delayed_put(msg, "Clearing lazy portal attr");
- }
+ return -1;
+}
- return 0;
+static struct lnet_peer_ni *
+lnet_find_route_locked(struct lnet_net *net, lnet_nid_t target,
+ lnet_nid_t rtr_nid)
+{
+ lnet_remotenet_t *rnet;
+ lnet_route_t *route;
+ lnet_route_t *best_route;
+ lnet_route_t *last_route;
+ struct lnet_peer_ni *lpni_best;
+ struct lnet_peer_ni *lp;
+ int rc;
+
+ /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
+ * rtr_nid nid, otherwise find the best gateway I can use */
+
+ rnet = lnet_find_rnet_locked(LNET_NIDNET(target));
+ if (rnet == NULL)
+ return NULL;
+
+ lpni_best = NULL;
+ best_route = last_route = NULL;
+ list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
+ lp = route->lr_gateway;
+
+ if (!lnet_is_route_alive(route))
+ continue;
+
+ if (net != NULL && lp->lpni_net != net)
+ continue;
+
+ if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
+ return lp;
+
+ if (lpni_best == NULL) {
+ best_route = last_route = route;
+ lpni_best = lp;
+ continue;
+ }
+
+ /* no protection on below fields, but it's harmless */
+ if (last_route->lr_seq - route->lr_seq < 0)
+ last_route = route;
+
+ rc = lnet_compare_routes(route, best_route);
+ if (rc < 0)
+ continue;
+
+ best_route = route;
+ lpni_best = lp;
+ }
+
+ /* set sequence number on the best router to the latest sequence + 1
+ * so we can round-robin all routers, it's race and inaccurate but
+ * harmless and functional */
+ if (best_route != NULL)
+ best_route->lr_seq = last_route->lr_seq + 1;
+ return lpni_best;
}
-static void
-lnet_recv_put(lnet_libmd_t *md, lnet_msg_t *msg, int delayed,
- unsigned int offset, unsigned int mlength)
+static int
+lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
+ struct lnet_msg *msg, lnet_nid_t rtr_nid)
{
- lnet_hdr_t *hdr = &msg->msg_hdr;
+ struct lnet_ni *best_ni;
+ struct lnet_peer_ni *best_lpni;
+ struct lnet_peer_ni *best_gw;
+ struct lnet_peer_ni *lpni;
+ struct lnet_peer *peer;
+ struct lnet_peer_net *peer_net;
+ struct lnet_net *local_net;
+ struct lnet_ni *ni;
+ __u32 seq;
+ int cpt, cpt2, rc;
+ bool routing;
+ bool ni_is_pref;
+ bool preferred;
+ int best_credits;
+ int best_lpni_credits;
+ int md_cpt;
+ int shortest_distance;
+
+ /*
+ * get an initial CPT to use for locking. The idea here is not to
+ * serialize the calls to select_pathway, so that as many
+ * operations can run concurrently as possible. To do that we use
+ * the CPT where this call is being executed. Later on when we
+ * determine the CPT to use in lnet_message_commit, we switch the
+ * lock and check if there was any configuration change. If none,
+ * then we proceed, if there is, then we restart the operation.
+ */
+ cpt = lnet_net_lock_current();
+again:
+ best_ni = NULL;
+ best_lpni = NULL;
+ best_gw = NULL;
+ local_net = NULL;
+ routing = false;
+
+ seq = lnet_get_dlc_seq_locked();
+
+ if (the_lnet.ln_shutdown) {
+ lnet_net_unlock(cpt);
+ return -ESHUTDOWN;
+ }
+
+ if (msg->msg_md != NULL)
+ /* get the cpt of the MD, used during NUMA based selection */
+ md_cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
+ else
+ md_cpt = CFS_CPT_ANY;
+
+ peer = lnet_find_or_create_peer_locked(dst_nid, cpt);
+ if (IS_ERR(peer)) {
+ lnet_net_unlock(cpt);
+ return PTR_ERR(peer);
+ }
+
+ /* If peer is not healthy then can not send anything to it */
+ if (!lnet_is_peer_healthy_locked(peer)) {
+ lnet_net_unlock(cpt);
+ return -EHOSTUNREACH;
+ }
+
+ if (!peer->lp_multi_rail && lnet_get_num_peer_nis(peer) > 1) {
+ CERROR("peer %s is declared to be non MR capable, "
+ "yet configured with more than one NID\n",
+ libcfs_nid2str(dst_nid));
+ return -EINVAL;
+ }
+
+ /*
+ * STEP 1: first jab at determineing best_ni
+ * if src_nid is explicitly specified, then best_ni is already
+ * pre-determiend for us. Otherwise we need to select the best
+ * one to use later on
+ */
+ if (src_nid != LNET_NID_ANY) {
+ best_ni = lnet_nid2ni_locked(src_nid, cpt);
+ if (!best_ni) {
+ lnet_net_unlock(cpt);
+ LCONSOLE_WARN("Can't send to %s: src %s is not a "
+ "local nid\n", libcfs_nid2str(dst_nid),
+ libcfs_nid2str(src_nid));
+ return -EINVAL;
+ }
+
+ if (best_ni->ni_net->net_id != LNET_NIDNET(dst_nid)) {
+ lnet_net_unlock(cpt);
+ LCONSOLE_WARN("No route to %s via from %s\n",
+ libcfs_nid2str(dst_nid),
+ libcfs_nid2str(src_nid));
+ return -EINVAL;
+ }
+ goto pick_peer;
+ }
+
+ /*
+ * Decide whether we need to route to peer_ni.
+ * Get the local net that I need to be on to be able to directly
+ * send to that peer.
+ *
+ * a. Find the peer which the dst_nid belongs to.
+ * b. Iterate through each of the peer_nets/nis to decide
+ * the best peer/local_ni pair to use
+ */
+ list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_on_peer_list) {
+ if (!lnet_is_peer_net_healthy_locked(peer_net))
+ continue;
+
+ local_net = lnet_get_net_locked(peer_net->lpn_net_id);
+ if (!local_net) {
+ struct lnet_peer_ni *net_gw;
+ /*
+ * go through each peer_ni on that peer_net and
+ * determine the best possible gw to go through
+ */
+ list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
+ lpni_on_peer_net_list) {
+ net_gw = lnet_find_route_locked(NULL,
+ lpni->lpni_nid,
+ rtr_nid);
+
+ /*
+ * if no route is found for that network then
+ * move onto the next peer_ni in the peer
+ */
+ if (!net_gw)
+ continue;
+
+ if (!best_gw) {
+ best_gw = net_gw;
+ best_lpni = lpni;
+ } else {
+ rc = lnet_compare_peers(net_gw,
+ best_gw);
+ if (rc > 0) {
+ best_gw = net_gw;
+ best_lpni = lpni;
+ }
+ }
+ }
+
+ if (!best_gw)
+ continue;
+
+ local_net = lnet_get_net_locked
+ (LNET_NIDNET(best_gw->lpni_nid));
+ routing = true;
+ } else {
+ routing = false;
+ best_gw = NULL;
+ }
+
+ /* no routable net found go on to a different net */
+ if (!local_net)
+ continue;
+
+ /*
+ * Iterate through the NIs in this local Net and select
+ * the NI to send from. The selection is determined by
+ * these 3 criterion in the following priority:
+ * 1. NUMA
+ * 2. NI available credits
+ * 3. Round Robin
+ */
+ shortest_distance = INT_MAX;
+ best_credits = INT_MIN;
+ ni = NULL;
+ while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
+ int ni_credits;
+ int distance;
+
+ if (!lnet_is_ni_healthy_locked(ni))
+ continue;
+
+ ni_credits = atomic_read(&ni->ni_tx_credits);
+
+ /*
+ * calculate the distance from the CPT on which
+ * the message memory is allocated to the CPT of
+ * the NI's physical device
+ */
+ distance = cfs_cpt_distance(lnet_cpt_table(),
+ md_cpt,
+ ni->dev_cpt);
+
+ /*
+ * All distances smaller than the NUMA range
+ * are treated equally.
+ */
+ if (distance < lnet_get_numa_range())
+ distance = lnet_get_numa_range();
+
+ /*
+ * Select on shorter distance, then available
+ * credits, then round-robin.
+ */
+ if (distance > shortest_distance) {
+ continue;
+ } else if (distance < shortest_distance) {
+ shortest_distance = distance;
+ } else if (ni_credits < best_credits) {
+ continue;
+ } else if (ni_credits == best_credits) {
+ if (best_ni && best_ni->ni_seq <= ni->ni_seq)
+ continue;
+ }
+ best_ni = ni;
+ best_credits = ni_credits;
+ }
+ }
+
+ /*
+ * if the peer is not MR capable, then we should always send to it
+ * using the first NI in the NET we determined.
+ */
+ if (!peer->lp_multi_rail && local_net != NULL)
+ best_ni = lnet_net2ni_locked(local_net->net_id, cpt);
+
+ if (!best_ni) {
+ lnet_net_unlock(cpt);
+ LCONSOLE_WARN("No local ni found to send from to %s\n",
+ libcfs_nid2str(dst_nid));
+ return -EINVAL;
+ }
+
+ /*
+ * Now that we selected the NI to use increment its sequence
+ * number so the Round Robin algorithm will detect that it has
+ * been used and pick the next NI.
+ */
+ best_ni->ni_seq++;
+
+ if (routing)
+ goto send;
+
+pick_peer:
+ if (best_ni == the_lnet.ln_loni) {
+ /* No send credit hassles with LOLND */
+ lnet_ni_addref_locked(best_ni, cpt);
+ msg->msg_hdr.dest_nid = cpu_to_le64(best_ni->ni_nid);
+ if (!msg->msg_routing)
+ msg->msg_hdr.src_nid = cpu_to_le64(best_ni->ni_nid);
+ msg->msg_target.nid = best_ni->ni_nid;
+ lnet_msg_commit(msg, cpt);
+ msg->msg_txni = best_ni;
+ lnet_net_unlock(cpt);
+
+ return LNET_CREDIT_OK;
+ }
+
+ if (msg->msg_type == LNET_MSG_REPLY ||
+ msg->msg_type == LNET_MSG_ACK) {
+ /*
+ * for replies we want to respond on the same peer_ni we
+ * received the message on if possible. If not, then pick
+ * a peer_ni to send to
+ */
+ best_lpni = lnet_find_peer_ni_locked(dst_nid);
+ if (best_lpni) {
+ lnet_peer_ni_decref_locked(best_lpni);
+ goto send;
+ } else {
+ CDEBUG(D_NET, "unable to send msg_type %d to "
+ "originating %s\n", msg->msg_type,
+ libcfs_nid2str(dst_nid));
+ }
+ }
+
+ peer_net = lnet_peer_get_net_locked(peer,
+ best_ni->ni_net->net_id);
+ /*
+ * peer_net is not available or the src_nid is explicitly defined
+ * and the peer_net for that src_nid is unhealthy. find a route to
+ * the destination nid.
+ */
+ if (!peer_net ||
+ (src_nid != LNET_NID_ANY &&
+ !lnet_is_peer_net_healthy_locked(peer_net))) {
+ best_gw = lnet_find_route_locked(best_ni->ni_net,
+ dst_nid,
+ rtr_nid);
+ /*
+ * if no route is found for that network then
+ * move onto the next peer_ni in the peer
+ */
+ if (!best_gw) {
+ lnet_net_unlock(cpt);
+ LCONSOLE_WARN("No route to peer from %s\n",
+ libcfs_nid2str(best_ni->ni_nid));
+ return -EHOSTUNREACH;
+ }
+
+ CDEBUG(D_NET, "Best route to %s via %s for %s %d\n",
+ libcfs_nid2str(dst_nid),
+ libcfs_nid2str(best_gw->lpni_nid),
+ lnet_msgtyp2str(msg->msg_type), msg->msg_len);
+
+ best_lpni = lnet_find_peer_ni_locked(dst_nid);
+ LASSERT(best_lpni != NULL);
+ lnet_peer_ni_decref_locked(best_lpni);
+
+ routing = true;
+
+ goto send;
+ } else if (!lnet_is_peer_net_healthy_locked(peer_net)) {
+ /*
+ * this peer_net is unhealthy but we still have an opportunity
+ * to find another peer_net that we can use
+ */
+ __u32 net_id = peer_net->lpn_net_id;
+ LCONSOLE_WARN("peer net %s unhealthy\n",
+ libcfs_net2str(net_id));
+ goto again;
+ }
+
+ /*
+ * Look at the peer NIs for the destination peer that connect
+ * to the chosen net. If a peer_ni is preferred when using the
+ * best_ni to communicate, we use that one. If there is no
+ * preferred peer_ni, or there are multiple preferred peer_ni,
+ * the available transmit credits are used. If the transmit
+ * credits are equal, we round-robin over the peer_ni.
+ */
+ lpni = NULL;
+ best_lpni_credits = INT_MIN;
+ preferred = false;
+ while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
+ /*
+ * if this peer ni is not healthy just skip it, no point in
+ * examining it further
+ */
+ if (!lnet_is_peer_ni_healthy_locked(lpni))
+ continue;
+ ni_is_pref = lnet_peer_is_ni_pref_locked(lpni, best_ni);
+
+ /* if this is a preferred peer use it */
+ if (!preferred && ni_is_pref) {
+ preferred = true;
+ } else if (preferred && !ni_is_pref) {
+ /*
+ * this is not the preferred peer so let's ignore
+ * it.
+ */
+ continue;
+ } else if (lpni->lpni_txcredits < best_lpni_credits) {
+ /*
+ * We already have a peer that has more credits
+ * available than this one. No need to consider
+ * this peer further.
+ */
+ continue;
+ } else if (lpni->lpni_txcredits == best_lpni_credits) {
+ /*
+ * The best peer found so far and the current peer
+ * have the same number of available credits let's
+ * make sure to select between them using Round
+ * Robin
+ */
+ if (best_lpni) {
+ if (best_lpni->lpni_seq <= lpni->lpni_seq)
+ continue;
+ }
+ }
+
+ best_lpni = lpni;
+ best_lpni_credits = lpni->lpni_txcredits;
+ }
+
+ /* if we still can't find a peer ni then we can't reach it */
+ if (!best_lpni) {
+ __u32 net_id = peer_net->lpn_net_id;
+ lnet_net_unlock(cpt);
+ LCONSOLE_WARN("no peer_ni found on peer net %s\n",
+ libcfs_net2str(net_id));
+ return -EHOSTUNREACH;
+ }
+
+send:
+ /*
+ * Increment sequence number of the peer selected so that we
+ * pick the next one in Round Robin.
+ */
+ best_lpni->lpni_seq++;
+
+ /*
+ * When routing the best gateway found acts as the best peer
+ * NI to send to.
+ */
+ if (routing)
+ best_lpni = best_gw;
+
+ /*
+ * grab a reference on the peer_ni so it sticks around even if
+ * we need to drop and relock the lnet_net_lock below.
+ */
+ lnet_peer_ni_addref_locked(best_lpni);
+
+ /*
+ * Use lnet_cpt_of_nid() to determine the CPT used to commit the
+ * message. This ensures that we get a CPT that is correct for
+ * the NI when the NI has been restricted to a subset of all CPTs.
+ * If the selected CPT differs from the one currently locked, we
+ * must unlock and relock the lnet_net_lock(), and then check whether
+ * the configuration has changed. We don't have a hold on the best_ni
+ * yet, and it may have vanished.
+ */
+ cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
+ if (cpt != cpt2) {
+ lnet_net_unlock(cpt);
+ cpt = cpt2;
+ lnet_net_lock(cpt);
+ if (seq != lnet_get_dlc_seq_locked()) {
+ lnet_peer_ni_decref_locked(best_lpni);
+ goto again;
+ }
+ }
+
+ /*
+ * store the best_lpni in the message right away to avoid having
+ * to do the same operation under different conditions
+ */
+ msg->msg_txpeer = best_lpni;
+ msg->msg_txni = best_ni;
+
+ /*
+ * grab a reference for the best_ni since now it's in use in this
+ * send. the reference will need to be dropped when the message is
+ * finished in lnet_finalize()
+ */
+ lnet_ni_addref_locked(msg->msg_txni, cpt);
+
+ /*
+ * set the destination nid in the message here because it's
+ * possible that we'd be sending to a different nid than the one
+ * originaly given.
+ */
+ msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
+
+ /*
+ * Always set the target.nid to the best peer picked. Either the
+ * nid will be one of the preconfigured NIDs, or the same NID as
+ * what was originaly set in the target or it will be the NID of
+ * a router if this message should be routed
+ */
+ msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
+
+ /*
+ * lnet_msg_commit assigns the correct cpt to the message, which
+ * is used to decrement the correct refcount on the ni when it's
+ * time to return the credits
+ */
+ lnet_msg_commit(msg, cpt);
+
+ /*
+ * If we are routing the message then we don't need to overwrite
+ * the src_nid since it would've been set at the origin. Otherwise
+ * we are the originator so we need to set it.
+ */
+ if (!msg->msg_routing)
+ msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
+
+ if (routing) {
+ msg->msg_target_is_router = 1;
+ msg->msg_target.pid = LNET_PID_LUSTRE;
+ }
+
+ rc = lnet_post_send_locked(msg, 0);
+
+ lnet_net_unlock(cpt);
+
+ return rc;
+}
- LNET_LOCK();
+int
+lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
+{
+ lnet_nid_t dst_nid = msg->msg_target.nid;
+ int rc;
- the_lnet.ln_counters.recv_count++;
- the_lnet.ln_counters.recv_length += mlength;
+ /*
+ * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
+ * but we might want to use pre-determined router for ACK/REPLY
+ * in the future
+ */
+ /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
+ LASSERT (msg->msg_txpeer == NULL);
+ LASSERT (!msg->msg_sending);
+ LASSERT (!msg->msg_target_is_router);
+ LASSERT (!msg->msg_receiving);
- LNET_UNLOCK();
+ msg->msg_sending = 1;
- if (mlength != 0)
- lnet_setpayloadbuffer(msg);
+ LASSERT(!msg->msg_tx_committed);
- msg->msg_ev.type = LNET_EVENT_PUT;
- msg->msg_ev.target.pid = hdr->dest_pid;
- msg->msg_ev.target.nid = hdr->dest_nid;
- msg->msg_ev.hdr_data = hdr->msg.put.hdr_data;
+ rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
+ if (rc < 0)
+ return rc;
- /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
- * it back into the ACK during lnet_finalize() */
- msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
- (md->md_options & LNET_MD_ACK_DISABLE) == 0);
+ if (rc == LNET_CREDIT_OK)
+ lnet_ni_send(msg->msg_txni, msg);
- lnet_ni_recv(msg->msg_rxpeer->lp_ni,
- msg->msg_private,
- msg, delayed, offset, mlength,
- hdr->payload_length);
+ /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT */
+ return 0;
}
-/* called with LNET_LOCK held */
void
-lnet_match_blocked_msg(lnet_libmd_t *md)
+lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
{
- CFS_LIST_HEAD (drops);
- CFS_LIST_HEAD (matches);
- struct list_head *tmp;
- struct list_head *entry;
- lnet_msg_t *msg;
- lnet_me_t *me = md->md_me;
- lnet_portal_t *ptl = &the_lnet.ln_portals[me->me_portal];
-
- LASSERT (me->me_portal < the_lnet.ln_nportals);
-
- if ((ptl->ptl_options & LNET_PTL_LAZY) == 0) {
- LASSERT (list_empty(&ptl->ptl_msgq));
- return;
- }
-
- LASSERT (md->md_refcount == 0); /* a brand new MD */
-
- list_for_each_safe (entry, tmp, &ptl->ptl_msgq) {
- int rc;
- int index;
- unsigned int mlength;
- unsigned int offset;
- lnet_hdr_t *hdr;
- lnet_process_id_t src;
-
- msg = list_entry(entry, lnet_msg_t, msg_list);
-
- LASSERT (msg->msg_delayed);
+ lnet_net_lock(cpt);
+ the_lnet.ln_counters[cpt]->drop_count++;
+ the_lnet.ln_counters[cpt]->drop_length += nob;
+ lnet_net_unlock(cpt);
- hdr = &msg->msg_hdr;
- index = hdr->msg.put.ptl_index;
-
- src.nid = hdr->src_nid;
- src.pid = hdr->src_pid;
-
- rc = lnet_try_match_md(index, LNET_MD_OP_PUT, src,
- hdr->payload_length,
- hdr->msg.put.offset,
- hdr->msg.put.match_bits,
- md, msg, &mlength, &offset);
-
- if (rc == LNET_MATCHMD_NONE)
- continue;
-
- /* Hurrah! This _is_ a match */
- list_del(&msg->msg_list);
- ptl->ptl_msgq_version++;
-
- if (rc == LNET_MATCHMD_OK) {
- list_add_tail(&msg->msg_list, &matches);
-
- CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
- "match "LPU64" offset %d length %d.\n",
- libcfs_id2str(src),
- hdr->msg.put.ptl_index,
- hdr->msg.put.match_bits,
- hdr->msg.put.offset,
- hdr->payload_length);
- } else {
- LASSERT (rc == LNET_MATCHMD_DROP);
-
- list_add_tail(&msg->msg_list, &drops);
- }
-
- if (lnet_md_exhausted(md))
- break;
- }
-
- LNET_UNLOCK();
-
- list_for_each_safe (entry, tmp, &drops) {
- msg = list_entry(entry, lnet_msg_t, msg_list);
-
- list_del(&msg->msg_list);
+ lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
+}
- lnet_drop_delayed_put(msg, "Bad match");
- }
+static void
+lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
+{
+ lnet_hdr_t *hdr = &msg->msg_hdr;
- list_for_each_safe (entry, tmp, &matches) {
- msg = list_entry(entry, lnet_msg_t, msg_list);
+ if (msg->msg_wanted != 0)
+ lnet_setpayloadbuffer(msg);
- list_del(&msg->msg_list);
+ lnet_build_msg_event(msg, LNET_EVENT_PUT);
- /* md won't disappear under me, since each msg
- * holds a ref on it */
- lnet_recv_put(md, msg, 1,
- msg->msg_ev.offset,
- msg->msg_ev.mlength);
- }
+ /* Must I ACK? If so I'll grab the ack_wmd out of the header and put
+ * it back into the ACK during lnet_finalize() */
+ msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
+ (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
- LNET_LOCK();
+ lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
+ msg->msg_offset, msg->msg_wanted, hdr->payload_length);
}
static int
lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
{
- int rc;
- int index;
- lnet_hdr_t *hdr = &msg->msg_hdr;
- unsigned int rlength = hdr->payload_length;
- unsigned int mlength = 0;
- unsigned int offset = 0;
- lnet_process_id_t src = {/* .nid = */ hdr->src_nid,
- /* .pid = */ hdr->src_pid};
- lnet_libmd_t *md;
-
- /* Convert put fields to host byte order */
- hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
- hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
- hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
-
- index = hdr->msg.put.ptl_index;
-
- LNET_LOCK();
-
- rc = lnet_match_md(index, LNET_MD_OP_PUT, src,
- rlength, hdr->msg.put.offset,
- hdr->msg.put.match_bits, msg,
- &mlength, &offset, &md);
- switch (rc) {
- default:
- LBUG();
-
- case LNET_MATCHMD_OK:
- LNET_UNLOCK();
- lnet_recv_put(md, msg, 0, offset, mlength);
- return 0;
-
- case LNET_MATCHMD_NONE:
- rc = lnet_eager_recv_locked(msg);
- if (rc == 0 && !the_lnet.ln_shutdown) {
- list_add_tail(&msg->msg_list,
- &the_lnet.ln_portals[index].ptl_msgq);
-
- the_lnet.ln_portals[index].ptl_msgq_version++;
-
- CDEBUG(D_NET, "Delaying PUT from %s portal %d match "
- LPU64" offset %d length %d: no match \n",
- libcfs_id2str(src), index,
- hdr->msg.put.match_bits,
- hdr->msg.put.offset, rlength);
-
- LNET_UNLOCK();
- return 0;
- }
- /* fall through */
-
- case LNET_MATCHMD_DROP:
- CDEBUG(D_NETERROR,
- "Dropping PUT from %s portal %d match "LPU64
- " offset %d length %d: %d\n",
- libcfs_id2str(src), index,
- hdr->msg.put.match_bits,
- hdr->msg.put.offset, rlength, rc);
- LNET_UNLOCK();
-
- return ENOENT; /* +ve: OK but no match */
- }
+ lnet_hdr_t *hdr = &msg->msg_hdr;
+ struct lnet_match_info info;
+ int rc;
+ bool ready_delay;
+
+ /* Convert put fields to host byte order */
+ hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
+ hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
+ hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
+
+ /* 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;
+ info.mi_rlength = hdr->payload_length;
+ info.mi_roffset = hdr->msg.put.offset;
+ info.mi_mbits = hdr->msg.put.match_bits;
+ info.mi_cpt = lnet_cpt_of_nid(msg->msg_rxpeer->lpni_nid, ni);
+
+ msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
+ ready_delay = msg->msg_rx_ready_delay;
+
+ again:
+ rc = lnet_ptl_match_md(&info, msg);
+ switch (rc) {
+ default:
+ LBUG();
+
+ case LNET_MATCHMD_OK:
+ lnet_recv_put(ni, msg);
+ return 0;
+
+ case LNET_MATCHMD_NONE:
+ if (ready_delay)
+ /* no eager_recv or has already called it, should
+ * have been attached on delayed list */
+ return 0;
+
+ rc = lnet_ni_eager_recv(ni, msg);
+ if (rc == 0) {
+ ready_delay = true;
+ goto again;
+ }
+ /* fall through */
+
+ case LNET_MATCHMD_DROP:
+ CNETERR("Dropping PUT from %s portal %d match %llu"
+ " offset %d length %d: %d\n",
+ libcfs_id2str(info.mi_id), info.mi_portal,
+ info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
+
+ return -ENOENT; /* -ve: OK but no match */
+ }
}
static int
lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
{
- lnet_hdr_t *hdr = &msg->msg_hdr;
- unsigned int mlength = 0;
- unsigned int offset = 0;
- lnet_process_id_t src = {/* .nid = */ hdr->src_nid,
- /* .pid = */ hdr->src_pid};
- lnet_handle_wire_t reply_wmd;
- lnet_libmd_t *md;
- int rc;
-
- /* Convert get fields to host byte order */
- hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
- hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
- hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
- hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
-
- LNET_LOCK();
-
- rc = lnet_match_md(hdr->msg.get.ptl_index, LNET_MD_OP_GET, src,
- hdr->msg.get.sink_length, hdr->msg.get.src_offset,
- hdr->msg.get.match_bits, msg,
- &mlength, &offset, &md);
- if (rc == LNET_MATCHMD_DROP) {
- CDEBUG(D_NETERROR,
- "Dropping GET from %s portal %d match "LPU64
- " offset %d length %d\n",
- libcfs_id2str(src),
- hdr->msg.get.ptl_index,
- hdr->msg.get.match_bits,
- hdr->msg.get.src_offset,
- hdr->msg.get.sink_length);
- LNET_UNLOCK();
- return ENOENT; /* +ve: OK but no match */
- }
-
- LASSERT (rc == LNET_MATCHMD_OK);
-
- the_lnet.ln_counters.send_count++;
- the_lnet.ln_counters.send_length += mlength;
-
- LNET_UNLOCK();
-
- reply_wmd = hdr->msg.get.return_wmd;
-
- lnet_prep_send(msg, LNET_MSG_REPLY, src, offset, mlength);
-
- msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
-
- msg->msg_ev.type = LNET_EVENT_GET;
- msg->msg_ev.target.pid = hdr->dest_pid;
- msg->msg_ev.target.nid = hdr->dest_nid;
- msg->msg_ev.hdr_data = 0;
-
- if (rdma_get) {
- /* The LND completes the REPLY from her recv procedure */
- lnet_ni_recv(ni, msg->msg_private, msg, 0,
- msg->msg_offset, msg->msg_len, msg->msg_len);
- return 0;
- }
-
- lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
- msg->msg_receiving = 0;
-
- rc = lnet_send(ni->ni_nid, msg);
- if (rc < 0) {
- /* didn't get as far as lnet_ni_send() */
- CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), rc);
-
- lnet_finalize(ni, msg, rc);
- }
-
- return 0;
+ struct lnet_match_info info;
+ lnet_hdr_t *hdr = &msg->msg_hdr;
+ lnet_process_id_t source_id;
+ struct lnet_handle_wire reply_wmd;
+ int rc;
+
+ /* Convert get fields to host byte order */
+ hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
+ hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
+ hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
+ hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
+
+ source_id.nid = hdr->src_nid;
+ source_id.pid = hdr->src_pid;
+ /* Primary peer NID */
+ info.mi_id.nid = msg->msg_initiator;
+ info.mi_id.pid = hdr->src_pid;
+ info.mi_opc = LNET_MD_OP_GET;
+ info.mi_portal = hdr->msg.get.ptl_index;
+ info.mi_rlength = hdr->msg.get.sink_length;
+ info.mi_roffset = hdr->msg.get.src_offset;
+ info.mi_mbits = hdr->msg.get.match_bits;
+
+ rc = lnet_ptl_match_md(&info, msg);
+ if (rc == LNET_MATCHMD_DROP) {
+ CNETERR("Dropping GET from %s portal %d match %llu"
+ " offset %d length %d\n",
+ libcfs_id2str(info.mi_id), info.mi_portal,
+ info.mi_mbits, info.mi_roffset, info.mi_rlength);
+ return -ENOENT; /* -ve: OK but no match */
+ }
+
+ LASSERT(rc == LNET_MATCHMD_OK);
+
+ lnet_build_msg_event(msg, LNET_EVENT_GET);
+
+ reply_wmd = hdr->msg.get.return_wmd;
+
+ lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
+ msg->msg_offset, msg->msg_wanted);
+
+ msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
+
+ if (rdma_get) {
+ /* The LND completes the REPLY from her recv procedure */
+ lnet_ni_recv(ni, msg->msg_private, msg, 0,
+ msg->msg_offset, msg->msg_len, msg->msg_len);
+ return 0;
+ }
+
+ lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
+ msg->msg_receiving = 0;
+
+ rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY);
+ if (rc < 0) {
+ /* didn't get as far as lnet_ni_send() */
+ CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
+ libcfs_nid2str(ni->ni_nid),
+ libcfs_id2str(info.mi_id), rc);
+
+ lnet_finalize(ni, msg, rc);
+ }
+
+ return 0;
}
static int
lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
{
- void *private = msg->msg_private;
- lnet_hdr_t *hdr = &msg->msg_hdr;
- lnet_process_id_t src = {/* .nid = */ hdr->src_nid,
- /* .pid = */ hdr->src_pid};
- lnet_libmd_t *md;
- int rlength;
- int mlength;
-
- LNET_LOCK();
-
- /* NB handles only looked up by creator (no flips) */
- md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
- if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
- CDEBUG(D_NETERROR, "%s: Dropping REPLY from %s for %s "
- "MD "LPX64"."LPX64"\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
- (md == NULL) ? "invalid" : "inactive",
- hdr->msg.reply.dst_wmd.wh_interface_cookie,
- hdr->msg.reply.dst_wmd.wh_object_cookie);
- if (md != NULL && md->md_me != NULL)
- CERROR("REPLY MD also attached to portal %d\n",
- md->md_me->me_portal);
-
- LNET_UNLOCK();
- return ENOENT; /* +ve: OK but no match */
- }
-
- LASSERT (md->md_offset == 0);
-
- rlength = hdr->payload_length;
- mlength = MIN(rlength, md->md_length);
-
- if (mlength < rlength &&
- (md->md_options & LNET_MD_TRUNCATE) == 0) {
- CDEBUG(D_NETERROR, "%s: Dropping REPLY from %s length %d "
- "for MD "LPX64" would overflow (%d)\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
- rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
- mlength);
- LNET_UNLOCK();
- return ENOENT; /* +ve: OK but no match */
- }
-
- CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
- mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
-
- lnet_commit_md(md, msg);
-
- if (mlength != 0)
- lnet_setpayloadbuffer(msg);
-
- msg->msg_ev.type = LNET_EVENT_REPLY;
- msg->msg_ev.target.pid = hdr->dest_pid;
- msg->msg_ev.target.nid = hdr->dest_nid;
- msg->msg_ev.initiator = src;
- msg->msg_ev.rlength = rlength;
- msg->msg_ev.mlength = mlength;
- msg->msg_ev.offset = 0;
-
- lnet_md_deconstruct(md, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, md);
-
- the_lnet.ln_counters.recv_count++;
- the_lnet.ln_counters.recv_length += mlength;
-
- LNET_UNLOCK();
-
- lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
- return 0;
+ void *private = msg->msg_private;
+ lnet_hdr_t *hdr = &msg->msg_hdr;
+ lnet_process_id_t src = {0};
+ lnet_libmd_t *md;
+ int rlength;
+ int mlength;
+ int cpt;
+
+ cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
+ lnet_res_lock(cpt);
+
+ src.nid = hdr->src_nid;
+ src.pid = hdr->src_pid;
+
+ /* NB handles only looked up by creator (no flips) */
+ md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
+ if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
+ CNETERR("%s: Dropping REPLY from %s for %s "
+ "MD %#llx.%#llx\n",
+ libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
+ (md == NULL) ? "invalid" : "inactive",
+ hdr->msg.reply.dst_wmd.wh_interface_cookie,
+ hdr->msg.reply.dst_wmd.wh_object_cookie);
+ if (md != NULL && md->md_me != NULL)
+ CERROR("REPLY MD also attached to portal %d\n",
+ md->md_me->me_portal);
+
+ lnet_res_unlock(cpt);
+ return -ENOENT; /* -ve: OK but no match */
+ }
+
+ LASSERT(md->md_offset == 0);
+
+ rlength = hdr->payload_length;
+ mlength = MIN(rlength, (int)md->md_length);
+
+ if (mlength < rlength &&
+ (md->md_options & LNET_MD_TRUNCATE) == 0) {
+ CNETERR("%s: Dropping REPLY from %s length %d "
+ "for MD %#llx would overflow (%d)\n",
+ libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
+ rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
+ mlength);
+ lnet_res_unlock(cpt);
+ return -ENOENT; /* -ve: OK but no match */
+ }
+
+ CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
+ libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
+ mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
+
+ lnet_msg_attach_md(msg, md, 0, mlength);
+
+ if (mlength != 0)
+ lnet_setpayloadbuffer(msg);
+
+ lnet_res_unlock(cpt);
+
+ lnet_build_msg_event(msg, LNET_EVENT_REPLY);
+
+ lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
+ return 0;
}
static int
lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
{
- lnet_hdr_t *hdr = &msg->msg_hdr;
- lnet_process_id_t src = {/* .nid = */ hdr->src_nid,
- /* .pid = */ hdr->src_pid};
- lnet_libmd_t *md;
-
- /* Convert ack fields to host byte order */
- hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
- hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
-
- LNET_LOCK();
-
- /* NB handles only looked up by creator (no flips) */
- md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
- if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
- /* Don't moan; this is expected */
- CDEBUG(D_NET,
- "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
- (md == NULL) ? "invalid" : "inactive",
- hdr->msg.ack.dst_wmd.wh_interface_cookie,
- hdr->msg.ack.dst_wmd.wh_object_cookie);
- if (md != NULL && md->md_me != NULL)
- CERROR("Source MD also attached to portal %d\n",
- md->md_me->me_portal);
-
- LNET_UNLOCK();
- return ENOENT; /* +ve! */
- }
-
- CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
- hdr->msg.ack.dst_wmd.wh_object_cookie);
-
- lnet_commit_md(md, msg);
-
- msg->msg_ev.type = LNET_EVENT_ACK;
- msg->msg_ev.target.pid = hdr->dest_pid;
- msg->msg_ev.target.nid = hdr->dest_nid;
- msg->msg_ev.initiator = src;
- msg->msg_ev.mlength = hdr->msg.ack.mlength;
- msg->msg_ev.match_bits = hdr->msg.ack.match_bits;
-
- lnet_md_deconstruct(md, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, md);
-
- the_lnet.ln_counters.recv_count++;
+ lnet_hdr_t *hdr = &msg->msg_hdr;
+ lnet_process_id_t src = {0};
+ lnet_libmd_t *md;
+ int cpt;
+
+ src.nid = hdr->src_nid;
+ src.pid = hdr->src_pid;
+
+ /* Convert ack fields to host byte order */
+ hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
+ hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
+
+ cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
+ lnet_res_lock(cpt);
+
+ /* NB handles only looked up by creator (no flips) */
+ md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
+ if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
+ /* Don't moan; this is expected */
+ CDEBUG(D_NET,
+ "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
+ libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
+ (md == NULL) ? "invalid" : "inactive",
+ hdr->msg.ack.dst_wmd.wh_interface_cookie,
+ hdr->msg.ack.dst_wmd.wh_object_cookie);
+ if (md != NULL && md->md_me != NULL)
+ CERROR("Source MD also attached to portal %d\n",
+ md->md_me->me_portal);
+
+ lnet_res_unlock(cpt);
+ return -ENOENT; /* -ve! */
+ }
+
+ CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
+ libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
+ hdr->msg.ack.dst_wmd.wh_object_cookie);
+
+ lnet_msg_attach_md(msg, md, 0, 0);
+
+ lnet_res_unlock(cpt);
+
+ lnet_build_msg_event(msg, LNET_EVENT_ACK);
+
+ lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
+ return 0;
+}
- LNET_UNLOCK();
+/**
+ * \retval LNET_CREDIT_OK If \a msg is forwarded
+ * \retval LNET_CREDIT_WAIT If \a msg is blocked because w/o buffer
+ * \retval -ve error code
+ */
+int
+lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
+{
+ int rc = 0;
+
+ if (!the_lnet.ln_routing)
+ return -ECANCELED;
+
+ if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
+ lnet_msg2bufpool(msg)->rbp_credits <= 0) {
+ if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
+ msg->msg_rx_ready_delay = 1;
+ } else {
+ lnet_net_unlock(msg->msg_rx_cpt);
+ rc = lnet_ni_eager_recv(ni, msg);
+ lnet_net_lock(msg->msg_rx_cpt);
+ }
+ }
+
+ if (rc == 0)
+ rc = lnet_post_routed_recv_locked(msg, 0);
+ return rc;
+}
- lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
- return 0;
+int
+lnet_parse_local(lnet_ni_t *ni, lnet_msg_t *msg)
+{
+ int rc;
+
+ switch (msg->msg_type) {
+ case LNET_MSG_ACK:
+ rc = lnet_parse_ack(ni, msg);
+ break;
+ case LNET_MSG_PUT:
+ rc = lnet_parse_put(ni, msg);
+ break;
+ case LNET_MSG_GET:
+ rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
+ break;
+ case LNET_MSG_REPLY:
+ rc = lnet_parse_reply(ni, msg);
+ break;
+ default: /* prevent an unused label if !kernel */
+ LASSERT(0);
+ return -EPROTO;
+ }
+
+ LASSERT(rc == 0 || rc == -ENOENT);
+ return rc;
}
char *
lnet_msgtyp2str (int type)
{
- switch (type) {
- case LNET_MSG_ACK:
- return ("ACK");
- case LNET_MSG_PUT:
- return ("PUT");
- case LNET_MSG_GET:
- return ("GET");
- case LNET_MSG_REPLY:
- return ("REPLY");
- case LNET_MSG_HELLO:
- return ("HELLO");
- default:
- return ("<UNKNOWN>");
- }
+ switch (type) {
+ case LNET_MSG_ACK:
+ return ("ACK");
+ case LNET_MSG_PUT:
+ return ("PUT");
+ case LNET_MSG_GET:
+ return ("GET");
+ case LNET_MSG_REPLY:
+ return ("REPLY");
+ case LNET_MSG_HELLO:
+ return ("HELLO");
+ default:
+ return ("<UNKNOWN>");
+ }
}
void
lnet_print_hdr(lnet_hdr_t * hdr)
{
- lnet_process_id_t src = {/* .nid = */ hdr->src_nid,
- /* .pid = */ hdr->src_pid};
- lnet_process_id_t dst = {/* .nid = */ hdr->dest_nid,
- /* .pid = */ hdr->dest_pid};
- char *type_str = lnet_msgtyp2str (hdr->type);
-
- CWARN("P3 Header at %p of type %s\n", hdr, type_str);
- CWARN(" From %s\n", libcfs_id2str(src));
- CWARN(" To %s\n", libcfs_id2str(dst));
-
- switch (hdr->type) {
- default:
- break;
-
- case LNET_MSG_PUT:
- CWARN(" Ptl index %d, ack md "LPX64"."LPX64", "
- "match bits "LPU64"\n",
- hdr->msg.put.ptl_index,
- hdr->msg.put.ack_wmd.wh_interface_cookie,
- hdr->msg.put.ack_wmd.wh_object_cookie,
- hdr->msg.put.match_bits);
- CWARN(" Length %d, offset %d, hdr data "LPX64"\n",
- hdr->payload_length, hdr->msg.put.offset,
- hdr->msg.put.hdr_data);
- break;
-
- case LNET_MSG_GET:
- CWARN(" Ptl index %d, return md "LPX64"."LPX64", "
- "match bits "LPU64"\n", hdr->msg.get.ptl_index,
- hdr->msg.get.return_wmd.wh_interface_cookie,
- hdr->msg.get.return_wmd.wh_object_cookie,
- hdr->msg.get.match_bits);
- CWARN(" Length %d, src offset %d\n",
- hdr->msg.get.sink_length,
- hdr->msg.get.src_offset);
- break;
-
- case LNET_MSG_ACK:
- CWARN(" dst md "LPX64"."LPX64", "
- "manipulated length %d\n",
- hdr->msg.ack.dst_wmd.wh_interface_cookie,
- hdr->msg.ack.dst_wmd.wh_object_cookie,
- hdr->msg.ack.mlength);
- break;
-
- case LNET_MSG_REPLY:
- CWARN(" dst md "LPX64"."LPX64", "
- "length %d\n",
- hdr->msg.reply.dst_wmd.wh_interface_cookie,
- hdr->msg.reply.dst_wmd.wh_object_cookie,
- hdr->payload_length);
- }
+ lnet_process_id_t src = {0};
+ lnet_process_id_t dst = {0};
+ char *type_str = lnet_msgtyp2str(hdr->type);
+
+ src.nid = hdr->src_nid;
+ src.pid = hdr->src_pid;
+
+ dst.nid = hdr->dest_nid;
+ dst.pid = hdr->dest_pid;
+
+ CWARN("P3 Header at %p of type %s\n", hdr, type_str);
+ CWARN(" From %s\n", libcfs_id2str(src));
+ CWARN(" To %s\n", libcfs_id2str(dst));
+
+ switch (hdr->type) {
+ default:
+ break;
+
+ case LNET_MSG_PUT:
+ CWARN(" Ptl index %d, ack md %#llx.%#llx, "
+ "match bits %llu\n",
+ hdr->msg.put.ptl_index,
+ hdr->msg.put.ack_wmd.wh_interface_cookie,
+ hdr->msg.put.ack_wmd.wh_object_cookie,
+ hdr->msg.put.match_bits);
+ CWARN(" Length %d, offset %d, hdr data %#llx\n",
+ hdr->payload_length, hdr->msg.put.offset,
+ hdr->msg.put.hdr_data);
+ break;
+
+ case LNET_MSG_GET:
+ CWARN(" Ptl index %d, return md %#llx.%#llx, "
+ "match bits %llu\n", hdr->msg.get.ptl_index,
+ hdr->msg.get.return_wmd.wh_interface_cookie,
+ hdr->msg.get.return_wmd.wh_object_cookie,
+ hdr->msg.get.match_bits);
+ CWARN(" Length %d, src offset %d\n",
+ hdr->msg.get.sink_length,
+ hdr->msg.get.src_offset);
+ break;
+
+ case LNET_MSG_ACK:
+ CWARN(" dst md %#llx.%#llx, "
+ "manipulated length %d\n",
+ hdr->msg.ack.dst_wmd.wh_interface_cookie,
+ hdr->msg.ack.dst_wmd.wh_object_cookie,
+ hdr->msg.ack.mlength);
+ break;
+
+ case LNET_MSG_REPLY:
+ CWARN(" dst md %#llx.%#llx, "
+ "length %d\n",
+ hdr->msg.reply.dst_wmd.wh_interface_cookie,
+ hdr->msg.reply.dst_wmd.wh_object_cookie,
+ hdr->payload_length);
+ }
}
-
int
-lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
- void *private, int rdma_req)
+lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
+ void *private, int rdma_req)
{
- int rc = 0;
- int for_me;
- lnet_msg_t *msg;
- lnet_nid_t dest_nid;
- lnet_nid_t src_nid;
- __u32 payload_length;
- __u32 type;
-
- LASSERT (!in_interrupt ());
-
- type = le32_to_cpu(hdr->type);
- src_nid = le64_to_cpu(hdr->src_nid);
- dest_nid = le64_to_cpu(hdr->dest_nid);
- payload_length = le32_to_cpu(hdr->payload_length);
-
- for_me = lnet_ptlcompat_matchnid(ni->ni_nid, dest_nid);
-
- switch (type) {
- case LNET_MSG_ACK:
- case LNET_MSG_GET:
- if (payload_length > 0) {
- CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
- libcfs_nid2str(from_nid),
- libcfs_nid2str(src_nid),
- lnet_msgtyp2str(type), payload_length);
- return -EPROTO;
- }
- break;
-
- case LNET_MSG_PUT:
- case LNET_MSG_REPLY:
- if (payload_length > (for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
- CERROR("%s, src %s: bad %s payload %d "
- "(%d max expected)\n",
- libcfs_nid2str(from_nid),
- libcfs_nid2str(src_nid),
- lnet_msgtyp2str(type),
- payload_length,
- for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
- return -EPROTO;
- }
- break;
-
- default:
- CERROR("%s, src %s: Bad message type 0x%x\n",
- libcfs_nid2str(from_nid),
- libcfs_nid2str(src_nid), type);
- return -EPROTO;
- }
-
- /* Regard a bad destination NID as a protocol error. Senders should
- * know what they're doing; if they don't they're misconfigured, buggy
- * or malicious so we chop them off at the knees :) */
-
- if (!for_me) {
- if (the_lnet.ln_ptlcompat > 0) {
- /* portals compatibility is single-network */
- CERROR ("%s, src %s: Bad dest nid %s "
- "(routing not supported)\n",
- libcfs_nid2str(from_nid),
- libcfs_nid2str(src_nid),
- libcfs_nid2str(dest_nid));
- return -EPROTO;
- }
-
- if (the_lnet.ln_ptlcompat == 0 &&
- LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
- /* should have gone direct */
- CERROR ("%s, src %s: Bad dest nid %s "
- "(should have been sent direct)\n",
- libcfs_nid2str(from_nid),
- libcfs_nid2str(src_nid),
- libcfs_nid2str(dest_nid));
- return -EPROTO;
- }
-
- if (the_lnet.ln_ptlcompat == 0 &&
- lnet_islocalnid(dest_nid)) {
- /* dest is another local NI; sender should have used
- * this node's NID on its own network */
- CERROR ("%s, src %s: Bad dest nid %s "
- "(it's my nid but on a different network)\n",
- libcfs_nid2str(from_nid),
- libcfs_nid2str(src_nid),
- libcfs_nid2str(dest_nid));
- return -EPROTO;
- }
-
- if (rdma_req && type == LNET_MSG_GET) {
- CERROR ("%s, src %s: Bad optimized GET for %s "
- "(final destination must be me)\n",
- libcfs_nid2str(from_nid),
- libcfs_nid2str(src_nid),
- libcfs_nid2str(dest_nid));
- return -EPROTO;
- }
-
- if (!the_lnet.ln_routing) {
- CERROR ("%s, src %s: Dropping message for %s "
- "(routing not enabled)\n",
- libcfs_nid2str(from_nid),
- libcfs_nid2str(src_nid),
- libcfs_nid2str(dest_nid));
- goto drop;
- }
- }
-
- /* Message looks OK; we're not going to return an error, so we MUST
- * call back lnd_recv() come what may... */
-
- if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
- fail_peer (src_nid, 0)) /* shall we now? */
- {
- CERROR("%s, src %s: Dropping %s to simulate failure\n",
- libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
- lnet_msgtyp2str(type));
- goto drop;
- }
-
- msg = lnet_msg_alloc();
- if (msg == NULL) {
- CERROR("%s, src %s: Dropping %s (out of memory)\n",
- libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
- lnet_msgtyp2str(type));
- goto drop;
- }
-
- /* msg zeroed in lnet_msg_alloc; i.e. flags all clear, pointers NULL etc */
-
- msg->msg_type = type;
- msg->msg_private = private;
- msg->msg_receiving = 1;
- msg->msg_len = msg->msg_wanted = payload_length;
- msg->msg_offset = 0;
- msg->msg_hdr = *hdr;
-
- LNET_LOCK();
- rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid);
- if (rc != 0) {
- LNET_UNLOCK();
- CERROR("%s, src %s: Dropping %s "
- "(error %d looking up sender)\n",
- libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
- lnet_msgtyp2str(type), rc);
- goto free_drop;
- }
- LNET_UNLOCK();
-
-#ifndef __KERNEL__
- LASSERT (for_me);
-#else
- if (!for_me) {
- msg->msg_target.pid = le32_to_cpu(hdr->dest_pid);
- msg->msg_target.nid = dest_nid;
- msg->msg_routing = 1;
- msg->msg_offset = 0;
-
- LNET_LOCK();
- if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
- lnet_msg2bufpool(msg)->rbp_credits <= 0) {
- rc = lnet_eager_recv_locked(msg);
- if (rc != 0) {
- LNET_UNLOCK();
- goto free_drop;
- }
- }
-
- lnet_commit_routedmsg(msg);
- rc = lnet_post_routed_recv_locked(msg, 0);
- LNET_UNLOCK();
-
- if (rc == 0)
- lnet_ni_recv(ni, msg->msg_private, msg, 0,
- 0, payload_length, payload_length);
- return 0;
- }
-#endif
- /* convert common msg->hdr fields to host byteorder */
- msg->msg_hdr.type = type;
- msg->msg_hdr.src_nid = src_nid;
- msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
- msg->msg_hdr.dest_nid = dest_nid;
- msg->msg_hdr.dest_pid = le32_to_cpu(msg->msg_hdr.dest_pid);
- msg->msg_hdr.payload_length = payload_length;
-
- msg->msg_ev.sender = from_nid;
-
- switch (type) {
- case LNET_MSG_ACK:
- rc = lnet_parse_ack(ni, msg);
- break;
- case LNET_MSG_PUT:
- rc = lnet_parse_put(ni, msg);
- break;
- case LNET_MSG_GET:
- rc = lnet_parse_get(ni, msg, rdma_req);
- break;
- case LNET_MSG_REPLY:
- rc = lnet_parse_reply(ni, msg);
- break;
- default:
- LASSERT(0);
- goto free_drop; /* prevent an unused label if !kernel */
- }
-
- if (rc == 0)
- return 0;
-
- LASSERT (rc == ENOENT);
+ int rc = 0;
+ int cpt;
+ int for_me;
+ struct lnet_msg *msg;
+ lnet_pid_t dest_pid;
+ lnet_nid_t dest_nid;
+ lnet_nid_t src_nid;
+ struct lnet_peer_ni *lpni;
+ __u32 payload_length;
+ __u32 type;
+
+ LASSERT (!in_interrupt ());
+
+ type = le32_to_cpu(hdr->type);
+ src_nid = le64_to_cpu(hdr->src_nid);
+ dest_nid = le64_to_cpu(hdr->dest_nid);
+ dest_pid = le32_to_cpu(hdr->dest_pid);
+ payload_length = le32_to_cpu(hdr->payload_length);
+
+ for_me = (ni->ni_nid == dest_nid);
+ cpt = lnet_cpt_of_nid(from_nid, ni);
+
+ switch (type) {
+ case LNET_MSG_ACK:
+ case LNET_MSG_GET:
+ if (payload_length > 0) {
+ CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
+ libcfs_nid2str(from_nid),
+ libcfs_nid2str(src_nid),
+ lnet_msgtyp2str(type), payload_length);
+ return -EPROTO;
+ }
+ break;
+
+ case LNET_MSG_PUT:
+ case LNET_MSG_REPLY:
+ if (payload_length >
+ (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
+ CERROR("%s, src %s: bad %s payload %d "
+ "(%d max expected)\n",
+ libcfs_nid2str(from_nid),
+ libcfs_nid2str(src_nid),
+ lnet_msgtyp2str(type),
+ payload_length,
+ for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
+ return -EPROTO;
+ }
+ break;
+
+ default:
+ CERROR("%s, src %s: Bad message type 0x%x\n",
+ libcfs_nid2str(from_nid),
+ libcfs_nid2str(src_nid), type);
+ return -EPROTO;
+ }
+
+ if (the_lnet.ln_routing &&
+ ni->ni_last_alive != ktime_get_real_seconds()) {
+ /* NB: so far here is the only place to set NI status to "up */
+ lnet_ni_lock(ni);
+ ni->ni_last_alive = ktime_get_real_seconds();
+ if (ni->ni_status != NULL &&
+ ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
+ ni->ni_status->ns_status = LNET_NI_STATUS_UP;
+ lnet_ni_unlock(ni);
+ }
+
+ /* Regard a bad destination NID as a protocol error. Senders should
+ * know what they're doing; if they don't they're misconfigured, buggy
+ * or malicious so we chop them off at the knees :) */
+
+ if (!for_me) {
+ if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
+ /* should have gone direct */
+ CERROR("%s, src %s: Bad dest nid %s "
+ "(should have been sent direct)\n",
+ libcfs_nid2str(from_nid),
+ libcfs_nid2str(src_nid),
+ libcfs_nid2str(dest_nid));
+ return -EPROTO;
+ }
+
+ if (lnet_islocalnid(dest_nid)) {
+ /* dest is another local NI; sender should have used
+ * this node's NID on its own network */
+ CERROR("%s, src %s: Bad dest nid %s "
+ "(it's my nid but on a different network)\n",
+ libcfs_nid2str(from_nid),
+ libcfs_nid2str(src_nid),
+ libcfs_nid2str(dest_nid));
+ return -EPROTO;
+ }
+
+ if (rdma_req && type == LNET_MSG_GET) {
+ CERROR("%s, src %s: Bad optimized GET for %s "
+ "(final destination must be me)\n",
+ libcfs_nid2str(from_nid),
+ libcfs_nid2str(src_nid),
+ libcfs_nid2str(dest_nid));
+ return -EPROTO;
+ }
+
+ if (!the_lnet.ln_routing) {
+ CERROR("%s, src %s: Dropping message for %s "
+ "(routing not enabled)\n",
+ libcfs_nid2str(from_nid),
+ libcfs_nid2str(src_nid),
+ libcfs_nid2str(dest_nid));
+ goto drop;
+ }
+ }
+
+ /* Message looks OK; we're not going to return an error, so we MUST
+ * call back lnd_recv() come what may... */
+
+ if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
+ fail_peer(src_nid, 0)) { /* shall we now? */
+ CERROR("%s, src %s: Dropping %s to simulate failure\n",
+ libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
+ lnet_msgtyp2str(type));
+ goto drop;
+ }
+
+ if (!list_empty(&the_lnet.ln_drop_rules) &&
+ lnet_drop_rule_match(hdr)) {
+ CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
+ "silent message loss\n",
+ libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
+ libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
+ goto drop;
+ }
+
+
+ msg = lnet_msg_alloc();
+ if (msg == NULL) {
+ CERROR("%s, src %s: Dropping %s (out of memory)\n",
+ libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
+ lnet_msgtyp2str(type));
+ goto drop;
+ }
+
+ /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
+ * pointers NULL etc */
+
+ msg->msg_type = type;
+ msg->msg_private = private;
+ msg->msg_receiving = 1;
+ msg->msg_rdma_get = rdma_req;
+ msg->msg_len = msg->msg_wanted = payload_length;
+ msg->msg_offset = 0;
+ msg->msg_hdr = *hdr;
+ /* for building message event */
+ msg->msg_from = from_nid;
+ if (!for_me) {
+ msg->msg_target.pid = dest_pid;
+ msg->msg_target.nid = dest_nid;
+ msg->msg_routing = 1;
+
+ } else {
+ /* convert common msg->hdr fields to host byteorder */
+ msg->msg_hdr.type = type;
+ msg->msg_hdr.src_nid = src_nid;
+ msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
+ msg->msg_hdr.dest_nid = dest_nid;
+ msg->msg_hdr.dest_pid = dest_pid;
+ msg->msg_hdr.payload_length = payload_length;
+ }
+ /* Multi-Rail: Primary NID of source. */
+ msg->msg_initiator = lnet_peer_primary_nid(src_nid);
+
+ lnet_net_lock(cpt);
+ lpni = lnet_nid2peerni_locked(from_nid, cpt);
+ if (IS_ERR(lpni)) {
+ lnet_net_unlock(cpt);
+ CERROR("%s, src %s: Dropping %s "
+ "(error %ld looking up sender)\n",
+ libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
+ lnet_msgtyp2str(type), PTR_ERR(lpni));
+ lnet_msg_free(msg);
+ if (rc == -ESHUTDOWN)
+ /* We are shutting down. Don't do anything more */
+ return 0;
+ goto drop;
+ }
+ msg->msg_rxpeer = lpni;
+ msg->msg_rxni = ni;
+ lnet_ni_addref_locked(ni, cpt);
+
+ if (lnet_isrouter(msg->msg_rxpeer)) {
+ lnet_peer_set_alive(msg->msg_rxpeer);
+ if (avoid_asym_router_failure &&
+ LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
+ /* received a remote message from router, update
+ * remote NI status on this router.
+ * NB: multi-hop routed message will be ignored.
+ */
+ lnet_router_ni_update_locked(msg->msg_rxpeer,
+ LNET_NIDNET(src_nid));
+ }
+ }
+
+ lnet_msg_commit(msg, cpt);
+
+ /* message delay simulation */
+ if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
+ lnet_delay_rule_match_locked(hdr, msg))) {
+ lnet_net_unlock(cpt);
+ return 0;
+ }
+
+ if (!for_me) {
+ rc = lnet_parse_forward_locked(ni, msg);
+ lnet_net_unlock(cpt);
+
+ if (rc < 0)
+ goto free_drop;
+
+ if (rc == LNET_CREDIT_OK) {
+ lnet_ni_recv(ni, msg->msg_private, msg, 0,
+ 0, payload_length, payload_length);
+ }
+ return 0;
+ }
+
+ lnet_net_unlock(cpt);
+
+ rc = lnet_parse_local(ni, msg);
+ if (rc != 0)
+ goto free_drop;
+ return 0;
free_drop:
- LASSERT (msg->msg_md == NULL);
- LNET_LOCK();
- if (msg->msg_rxpeer != NULL) {
- lnet_peer_decref_locked(msg->msg_rxpeer);
- msg->msg_rxpeer = NULL;
- }
- lnet_msg_free(msg); /* expects LNET_LOCK held */
- LNET_UNLOCK();
+ LASSERT(msg->msg_md == NULL);
+ lnet_finalize(ni, msg, rc);
drop:
- lnet_drop_message(ni, private, payload_length);
- return 0;
+ lnet_drop_message(ni, cpt, private, payload_length);
+ return 0;
}
+EXPORT_SYMBOL(lnet_parse);
-int
-LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
- lnet_process_id_t target, unsigned int portal,
- __u64 match_bits, unsigned int offset,
- __u64 hdr_data)
+void
+lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
{
- lnet_msg_t *msg;
- lnet_libmd_t *md;
- int rc;
-
- LASSERT (the_lnet.ln_init);
- LASSERT (the_lnet.ln_refcount > 0);
-
- if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
- fail_peer (target.nid, 1)) /* shall we now? */
- {
- CERROR("Dropping PUT to %s: simulated failure\n",
- libcfs_id2str(target));
- return -EIO;
- }
-
- msg = lnet_msg_alloc();
- if (msg == NULL) {
- CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
- libcfs_id2str(target));
- return -ENOMEM;
- }
-
- LNET_LOCK();
-
- md = lnet_handle2md(&mdh);
- if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
- lnet_msg_free(msg);
-
- CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n",
- match_bits, portal, libcfs_id2str(target),
- md == NULL ? -1 : md->md_threshold);
- if (md != NULL && md->md_me != NULL)
- CERROR("Source MD also attached to portal %d\n",
- md->md_me->me_portal);
-
- LNET_UNLOCK();
- return -ENOENT;
- }
-
- CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
-
- lnet_commit_md(md, msg);
-
- lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
-
- msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
- msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
- msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
- msg->msg_hdr.msg.put.hdr_data = hdr_data;
-
- /* NB handles only looked up by creator (no flips) */
- if (ack == LNET_ACK_REQ) {
- msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
- the_lnet.ln_interface_cookie;
- msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
- md->md_lh.lh_cookie;
- } else {
- msg->msg_hdr.msg.put.ack_wmd = LNET_WIRE_HANDLE_NONE;
- }
+ while (!list_empty(head)) {
+ lnet_process_id_t id = {0};
+ lnet_msg_t *msg;
+
+ msg = list_entry(head->next, lnet_msg_t, msg_list);
+ list_del(&msg->msg_list);
+
+ id.nid = msg->msg_hdr.src_nid;
+ id.pid = msg->msg_hdr.src_pid;
+
+ LASSERT(msg->msg_md == NULL);
+ LASSERT(msg->msg_rx_delayed);
+ LASSERT(msg->msg_rxpeer != NULL);
+ LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
+
+ CWARN("Dropping delayed PUT from %s portal %d match %llu"
+ " offset %d length %d: %s\n",
+ libcfs_id2str(id),
+ msg->msg_hdr.msg.put.ptl_index,
+ msg->msg_hdr.msg.put.match_bits,
+ msg->msg_hdr.msg.put.offset,
+ msg->msg_hdr.payload_length, reason);
+
+ /* NB I can't drop msg's ref on msg_rxpeer until after I've
+ * called lnet_drop_message(), so I just hang onto msg as well
+ * until that's done */
+
+ lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
+ msg->msg_private, msg->msg_len);
+ /*
+ * NB: message will not generate event because w/o attached MD,
+ * but we still should give error code so lnet_msg_decommit()
+ * can skip counters operations and other checks.
+ */
+ lnet_finalize(msg->msg_rxni, msg, -ENOENT);
+ }
+}
- msg->msg_ev.type = LNET_EVENT_SEND;
- msg->msg_ev.initiator.nid = LNET_NID_ANY;
- msg->msg_ev.initiator.pid = the_lnet.ln_pid;
- msg->msg_ev.target = target;
- msg->msg_ev.sender = LNET_NID_ANY;
- msg->msg_ev.pt_index = portal;
- msg->msg_ev.match_bits = match_bits;
- msg->msg_ev.rlength = md->md_length;
- msg->msg_ev.mlength = md->md_length;
- msg->msg_ev.offset = offset;
- msg->msg_ev.hdr_data = hdr_data;
-
- lnet_md_deconstruct(md, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, md);
-
- the_lnet.ln_counters.send_count++;
- the_lnet.ln_counters.send_length += md->md_length;
-
- LNET_UNLOCK();
-
- rc = lnet_send(self, msg);
- if (rc != 0) {
- CERROR("Error sending PUT to %s: %d\n",
- libcfs_id2str(target), rc);
- lnet_finalize (NULL, msg, rc);
- }
+void
+lnet_recv_delayed_msg_list(struct list_head *head)
+{
+ while (!list_empty(head)) {
+ lnet_msg_t *msg;
+ lnet_process_id_t id;
+
+ msg = list_entry(head->next, lnet_msg_t, msg_list);
+ list_del(&msg->msg_list);
+
+ /* md won't disappear under me, since each msg
+ * holds a ref on it */
+
+ id.nid = msg->msg_hdr.src_nid;
+ id.pid = msg->msg_hdr.src_pid;
+
+ LASSERT(msg->msg_rx_delayed);
+ LASSERT(msg->msg_md != NULL);
+ LASSERT(msg->msg_rxpeer != NULL);
+ LASSERT(msg->msg_rxni != NULL);
+ LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
+
+ CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
+ "match %llu offset %d length %d.\n",
+ libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
+ msg->msg_hdr.msg.put.match_bits,
+ msg->msg_hdr.msg.put.offset,
+ msg->msg_hdr.payload_length);
+
+ lnet_recv_put(msg->msg_rxni, msg);
+ }
+}
- /* completion will be signalled by an event */
- return 0;
+/**
+ * Initiate an asynchronous PUT operation.
+ *
+ * There are several events associated with a PUT: completion of the send on
+ * the initiator node (LNET_EVENT_SEND), and when the send completes
+ * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
+ * that the operation was accepted by the target. The event LNET_EVENT_PUT is
+ * used at the target node to indicate the completion of incoming data
+ * delivery.
+ *
+ * The local events will be logged in the EQ associated with the MD pointed to
+ * by \a mdh handle. Using a MD without an associated EQ results in these
+ * events being discarded. In this case, the caller must have another
+ * mechanism (e.g., a higher level protocol) for determining when it is safe
+ * to modify the memory region associated with the MD.
+ *
+ * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
+ * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
+ *
+ * \param self Indicates the NID of a local interface through which to send
+ * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
+ * \param mdh A handle for the MD that describes the memory to be sent. The MD
+ * must be "free floating" (See LNetMDBind()).
+ * \param ack Controls whether an acknowledgment is requested.
+ * Acknowledgments are only sent when they are requested by the initiating
+ * process and the target MD enables them.
+ * \param target A process identifier for the target process.
+ * \param portal The index in the \a target's portal table.
+ * \param match_bits The match bits to use for MD selection at the target
+ * process.
+ * \param offset The offset into the target MD (only used when the target
+ * MD has the LNET_MD_MANAGE_REMOTE option set).
+ * \param hdr_data 64 bits of user data that can be included in the message
+ * header. This data is written to an event queue entry at the target if an
+ * EQ is present on the matching MD.
+ *
+ * \retval 0 Success, and only in this case events will be generated
+ * and logged to EQ (if it exists).
+ * \retval -EIO Simulated failure.
+ * \retval -ENOMEM Memory allocation failure.
+ * \retval -ENOENT Invalid MD object.
+ *
+ * \see lnet_event_t::hdr_data and lnet_event_kind_t.
+ */
+int
+LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
+ lnet_process_id_t target, unsigned int portal,
+ __u64 match_bits, unsigned int offset,
+ __u64 hdr_data)
+{
+ struct lnet_msg *msg;
+ struct lnet_libmd *md;
+ int cpt;
+ int rc;
+
+ LASSERT(the_lnet.ln_refcount > 0);
+
+ if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
+ fail_peer(target.nid, 1)) { /* shall we now? */
+ CERROR("Dropping PUT to %s: simulated failure\n",
+ libcfs_id2str(target));
+ return -EIO;
+ }
+
+ msg = lnet_msg_alloc();
+ if (msg == NULL) {
+ CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
+ libcfs_id2str(target));
+ return -ENOMEM;
+ }
+ msg->msg_vmflush = !!memory_pressure_get();
+
+ cpt = lnet_cpt_of_cookie(mdh.cookie);
+ lnet_res_lock(cpt);
+
+ md = lnet_handle2md(&mdh);
+ if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
+ CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
+ match_bits, portal, libcfs_id2str(target),
+ md == NULL ? -1 : md->md_threshold);
+ if (md != NULL && md->md_me != NULL)
+ CERROR("Source MD also attached to portal %d\n",
+ md->md_me->me_portal);
+ lnet_res_unlock(cpt);
+
+ lnet_msg_free(msg);
+ return -ENOENT;
+ }
+
+ CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
+
+ lnet_msg_attach_md(msg, md, 0, 0);
+
+ lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
+
+ msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
+ msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
+ msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
+ msg->msg_hdr.msg.put.hdr_data = hdr_data;
+
+ /* NB handles only looked up by creator (no flips) */
+ if (ack == LNET_ACK_REQ) {
+ msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
+ the_lnet.ln_interface_cookie;
+ msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
+ md->md_lh.lh_cookie;
+ } else {
+ msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
+ LNET_WIRE_HANDLE_COOKIE_NONE;
+ msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
+ LNET_WIRE_HANDLE_COOKIE_NONE;
+ }
+
+ lnet_res_unlock(cpt);
+
+ lnet_build_msg_event(msg, LNET_EVENT_SEND);
+
+ rc = lnet_send(self, msg, LNET_NID_ANY);
+ if (rc != 0) {
+ CNETERR("Error sending PUT to %s: %d\n",
+ libcfs_id2str(target), rc);
+ lnet_finalize(NULL, msg, rc);
+ }
+
+ /* completion will be signalled by an event */
+ return 0;
}
+EXPORT_SYMBOL(LNetPut);
lnet_msg_t *
lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
{
- /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
- * returns a msg for the LND to pass to lnet_finalize() when the sink
- * data has been received.
- *
- * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
- * lnet_finalize() is called on it, so the LND must call this first */
+ /* The LND can DMA direct to the GET md (i.e. no REPLY msg). This
+ * returns a msg for the LND to pass to lnet_finalize() when the sink
+ * data has been received.
+ *
+ * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
+ * lnet_finalize() is called on it, so the LND must call this first */
- lnet_msg_t *msg = lnet_msg_alloc();
- lnet_libmd_t *getmd = getmsg->msg_md;
- lnet_process_id_t peer_id = getmsg->msg_target;
+ struct lnet_msg *msg = lnet_msg_alloc();
+ struct lnet_libmd *getmd = getmsg->msg_md;
+ lnet_process_id_t peer_id = getmsg->msg_target;
+ int cpt;
- LASSERT (!getmsg->msg_target_is_router);
- LASSERT (!getmsg->msg_routing);
+ LASSERT(!getmsg->msg_target_is_router);
+ LASSERT(!getmsg->msg_routing);
- LNET_LOCK();
+ if (msg == NULL) {
+ CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
+ libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
+ goto drop;
+ }
- LASSERT (getmd->md_refcount > 0);
+ cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
+ lnet_res_lock(cpt);
- if (msg == NULL) {
- CERROR ("%s: Dropping REPLY from %s: can't allocate msg\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
- goto drop;
- }
+ LASSERT(getmd->md_refcount > 0);
- if (getmd->md_threshold == 0) {
- CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
- getmd);
- goto drop_msg;
- }
-
- LASSERT (getmd->md_offset == 0);
+ if (getmd->md_threshold == 0) {
+ CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
+ libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
+ getmd);
+ lnet_res_unlock(cpt);
+ goto drop;
+ }
- CDEBUG(D_NET, "%s: Reply from %s md %p\n",
- libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
+ LASSERT(getmd->md_offset == 0);
- lnet_commit_md (getmd, msg);
+ CDEBUG(D_NET, "%s: Reply from %s md %p\n",
+ libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
- msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
+ /* setup information for lnet_build_msg_event */
+ msg->msg_initiator = lnet_peer_primary_nid(peer_id.nid);
+ /* Cheaper: msg->msg_initiator = getmsg->msg_txpeer->lp_nid; */
+ msg->msg_from = peer_id.nid;
+ msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
+ msg->msg_hdr.src_nid = peer_id.nid;
+ msg->msg_hdr.payload_length = getmd->md_length;
+ msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
- msg->msg_ev.type = LNET_EVENT_REPLY;
- msg->msg_ev.initiator = peer_id;
- msg->msg_ev.sender = peer_id.nid; /* optimized GETs can't be routed */
- msg->msg_ev.rlength = msg->msg_ev.mlength = getmd->md_length;
- msg->msg_ev.offset = 0;
+ lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
+ lnet_res_unlock(cpt);
- lnet_md_deconstruct(getmd, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, getmd);
+ cpt = lnet_cpt_of_nid(peer_id.nid, ni);
- the_lnet.ln_counters.recv_count++;
- the_lnet.ln_counters.recv_length += getmd->md_length;
+ lnet_net_lock(cpt);
+ lnet_msg_commit(msg, cpt);
+ lnet_net_unlock(cpt);
- LNET_UNLOCK();
+ lnet_build_msg_event(msg, LNET_EVENT_REPLY);
- return msg;
+ return msg;
- drop_msg:
- lnet_msg_free(msg);
drop:
- the_lnet.ln_counters.drop_count++;
- the_lnet.ln_counters.drop_length += getmd->md_length;
+ cpt = lnet_cpt_of_nid(peer_id.nid, ni);
+
+ lnet_net_lock(cpt);
+ the_lnet.ln_counters[cpt]->drop_count++;
+ the_lnet.ln_counters[cpt]->drop_length += getmd->md_length;
+ lnet_net_unlock(cpt);
- LNET_UNLOCK ();
+ if (msg != NULL)
+ lnet_msg_free(msg);
- return NULL;
+ return NULL;
}
+EXPORT_SYMBOL(lnet_create_reply_msg);
void
lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
{
- /* Set the REPLY length, now the RDMA that elides the REPLY message has
- * completed and I know it. */
- LASSERT (reply != NULL);
- LASSERT (reply->msg_type == LNET_MSG_GET);
- LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
-
- /* NB I trusted my peer to RDMA. If she tells me she's written beyond
- * the end of my buffer, I might as well be dead. */
- LASSERT (len <= reply->msg_ev.mlength);
-
- reply->msg_ev.mlength = len;
-}
-
-int
-LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
- lnet_process_id_t target, unsigned int portal,
- __u64 match_bits, unsigned int offset)
-{
- lnet_msg_t *msg;
- lnet_libmd_t *md;
- int rc;
-
- LASSERT (the_lnet.ln_init);
- LASSERT (the_lnet.ln_refcount > 0);
-
- if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
- fail_peer (target.nid, 1)) /* shall we now? */
- {
- CERROR("Dropping GET to %s: simulated failure\n",
- libcfs_id2str(target));
- return -EIO;
- }
-
- msg = lnet_msg_alloc();
- if (msg == NULL) {
- CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
- libcfs_id2str(target));
- return -ENOMEM;
- }
-
- LNET_LOCK();
-
- md = lnet_handle2md(&mdh);
- if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
- lnet_msg_free(msg);
-
- CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n",
- match_bits, portal, libcfs_id2str(target),
- md == NULL ? -1 : md->md_threshold);
- if (md != NULL && md->md_me != NULL)
- CERROR("REPLY MD also attached to portal %d\n",
- md->md_me->me_portal);
-
- LNET_UNLOCK();
- return -ENOENT;
- }
-
- CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
-
- lnet_commit_md(md, msg);
+ /* Set the REPLY length, now the RDMA that elides the REPLY message has
+ * completed and I know it. */
+ LASSERT(reply != NULL);
+ LASSERT(reply->msg_type == LNET_MSG_GET);
+ LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
- lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
+ /* NB I trusted my peer to RDMA. If she tells me she's written beyond
+ * the end of my buffer, I might as well be dead. */
+ LASSERT(len <= reply->msg_ev.mlength);
- msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
- msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
- msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
- msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
-
- /* NB handles only looked up by creator (no flips) */
- msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
- the_lnet.ln_interface_cookie;
- msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
- md->md_lh.lh_cookie;
-
- msg->msg_ev.type = LNET_EVENT_SEND;
- msg->msg_ev.initiator.nid = LNET_NID_ANY;
- msg->msg_ev.initiator.pid = the_lnet.ln_pid;
- msg->msg_ev.target = target;
- msg->msg_ev.sender = LNET_NID_ANY;
- msg->msg_ev.pt_index = portal;
- msg->msg_ev.match_bits = match_bits;
- msg->msg_ev.rlength = md->md_length;
- msg->msg_ev.mlength = md->md_length;
- msg->msg_ev.offset = offset;
- msg->msg_ev.hdr_data = 0;
-
- lnet_md_deconstruct(md, &msg->msg_ev.md);
- lnet_md2handle(&msg->msg_ev.md_handle, md);
-
- the_lnet.ln_counters.send_count++;
-
- LNET_UNLOCK();
-
- rc = lnet_send(self, msg);
- if (rc < 0) {
- CERROR("error sending GET to %s: %d\n",
- libcfs_id2str(target), rc);
- lnet_finalize (NULL, msg, rc);
- }
-
- /* completion will be signalled by an event */
- return 0;
+ reply->msg_ev.mlength = len;
}
+EXPORT_SYMBOL(lnet_set_reply_msg_len);
+/**
+ * Initiate an asynchronous GET operation.
+ *
+ * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
+ * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
+ * the target node in the REPLY has been written to local MD.
+ *
+ * On the target node, an LNET_EVENT_GET is logged when the GET request
+ * arrives and is accepted into a MD.
+ *
+ * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
+ * \param mdh A handle for the MD that describes the memory into which the
+ * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
+ *
+ * \retval 0 Success, and only in this case events will be generated
+ * and logged to EQ (if it exists) of the MD.
+ * \retval -EIO Simulated failure.
+ * \retval -ENOMEM Memory allocation failure.
+ * \retval -ENOENT Invalid MD object.
+ */
int
-LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
+LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
+ lnet_process_id_t target, unsigned int portal,
+ __u64 match_bits, unsigned int offset)
{
- struct list_head *e;
- lnet_ni_t *ni;
- lnet_route_t *route;
- lnet_remotenet_t *rnet;
- __u32 dstnet = LNET_NIDNET(dstnid);
- int hops;
- __u32 order = 2;
-
- /* if !local_nid_dist_zero, I don't return a distance of 0 ever
- * (when lustre sees a distance of 0, it substitutes 0@lo), so I
- * keep order 0 free for 0@lo and order 1 free for a local NID
- * match */
-
- LASSERT (the_lnet.ln_init);
- LASSERT (the_lnet.ln_refcount > 0);
-
- LNET_LOCK();
-
- list_for_each (e, &the_lnet.ln_nis) {
- ni = list_entry(e, lnet_ni_t, ni_list);
-
- if (ni->ni_nid == dstnid ||
- (the_lnet.ln_ptlcompat > 0 &&
- LNET_NIDNET(dstnid) == 0 &&
- LNET_NIDADDR(dstnid) == LNET_NIDADDR(ni->ni_nid) &&
- LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) != LOLND)) {
- if (srcnidp != NULL)
- *srcnidp = dstnid;
- if (orderp != NULL) {
- if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
- *orderp = 0;
- else
- *orderp = 1;
- }
- LNET_UNLOCK();
-
- return local_nid_dist_zero ? 0 : 1;
- }
-
- if (LNET_NIDNET(ni->ni_nid) == dstnet ||
- (the_lnet.ln_ptlcompat > 0 &&
- dstnet == 0 &&
- LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) != LOLND)) {
- if (srcnidp != NULL)
- *srcnidp = ni->ni_nid;
- if (orderp != NULL)
- *orderp = order;
- LNET_UNLOCK();
- return 1;
- }
-
- order++;
- }
-
- list_for_each (e, &the_lnet.ln_remote_nets) {
- rnet = list_entry(e, lnet_remotenet_t, lrn_list);
-
- if (rnet->lrn_net == dstnet) {
- LASSERT (!list_empty(&rnet->lrn_routes));
- route = list_entry(rnet->lrn_routes.next,
- lnet_route_t, lr_list);
- hops = rnet->lrn_hops;
- if (srcnidp != NULL)
- *srcnidp = route->lr_gateway->lp_ni->ni_nid;
- if (orderp != NULL)
- *orderp = order;
- LNET_UNLOCK();
- return hops + 1;
- }
- order++;
- }
-
- LNET_UNLOCK();
- return -EHOSTUNREACH;
+ struct lnet_msg *msg;
+ struct lnet_libmd *md;
+ int cpt;
+ int rc;
+
+ LASSERT(the_lnet.ln_refcount > 0);
+
+ if (!list_empty(&the_lnet.ln_test_peers) && /* normally we don't */
+ fail_peer(target.nid, 1)) /* shall we now? */
+ {
+ CERROR("Dropping GET to %s: simulated failure\n",
+ libcfs_id2str(target));
+ return -EIO;
+ }
+
+ msg = lnet_msg_alloc();
+ if (msg == NULL) {
+ CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
+ libcfs_id2str(target));
+ return -ENOMEM;
+ }
+
+ cpt = lnet_cpt_of_cookie(mdh.cookie);
+ lnet_res_lock(cpt);
+
+ md = lnet_handle2md(&mdh);
+ if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
+ CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
+ match_bits, portal, libcfs_id2str(target),
+ md == NULL ? -1 : md->md_threshold);
+ if (md != NULL && md->md_me != NULL)
+ CERROR("REPLY MD also attached to portal %d\n",
+ md->md_me->me_portal);
+
+ lnet_res_unlock(cpt);
+
+ lnet_msg_free(msg);
+ return -ENOENT;
+ }
+
+ CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
+
+ lnet_msg_attach_md(msg, md, 0, 0);
+
+ lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
+
+ msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
+ msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
+ msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
+ msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
+
+ /* NB handles only looked up by creator (no flips) */
+ msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
+ the_lnet.ln_interface_cookie;
+ msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
+ md->md_lh.lh_cookie;
+
+ lnet_res_unlock(cpt);
+
+ lnet_build_msg_event(msg, LNET_EVENT_SEND);
+
+ rc = lnet_send(self, msg, LNET_NID_ANY);
+ if (rc < 0) {
+ CNETERR("Error sending GET to %s: %d\n",
+ libcfs_id2str(target), rc);
+ lnet_finalize(NULL, msg, rc);
+ }
+
+ /* completion will be signalled by an event */
+ return 0;
}
+EXPORT_SYMBOL(LNetGet);
+/**
+ * Calculate distance to node at \a dstnid.
+ *
+ * \param dstnid Target NID.
+ * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
+ * is saved here.
+ * \param orderp If not NULL, order of the route to reach \a dstnid is saved
+ * here.
+ *
+ * \retval 0 If \a dstnid belongs to a local interface, and reserved option
+ * local_nid_dist_zero is set, which is the default.
+ * \retval positives Distance to target NID, i.e. number of hops plus one.
+ * \retval -EHOSTUNREACH If \a dstnid is not reachable.
+ */
int
-LNetSetAsync(lnet_process_id_t id, int nasync)
+LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
{
-#ifdef __KERNEL__
- return 0;
-#else
- lnet_ni_t *ni;
- lnet_remotenet_t *rnet;
- struct list_head *tmp;
- lnet_route_t *route;
- lnet_nid_t *nids;
- int nnids;
- int maxnids = 256;
- int rc = 0;
- int rc2;
-
- /* Target on a local network? */
-
- ni = lnet_net2ni(LNET_NIDNET(id.nid));
- if (ni != NULL) {
- if (ni->ni_lnd->lnd_setasync != NULL)
- rc = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
- lnet_ni_decref(ni);
- return rc;
- }
-
- /* Target on a remote network: apply to routers */
- again:
- LIBCFS_ALLOC(nids, maxnids * sizeof(*nids));
- if (nids == NULL)
- return -ENOMEM;
- nnids = 0;
-
- /* Snapshot all the router NIDs */
- LNET_LOCK();
- rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
- if (rnet != NULL) {
- list_for_each(tmp, &rnet->lrn_routes) {
- if (nnids == maxnids) {
- LNET_UNLOCK();
- LIBCFS_FREE(nids, maxnids * sizeof(*nids));
- maxnids *= 2;
- goto again;
- }
-
- route = list_entry(tmp, lnet_route_t, lr_list);
- nids[nnids++] = route->lr_gateway->lp_nid;
- }
- }
- LNET_UNLOCK();
-
- /* set async on all the routers */
- while (nnids-- > 0) {
- id.pid = LUSTRE_SRV_LNET_PID;
- id.nid = nids[nnids];
-
- ni = lnet_net2ni(LNET_NIDNET(id.nid));
- if (ni == NULL)
- continue;
-
- if (ni->ni_lnd->lnd_setasync != NULL) {
- rc2 = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
- if (rc2 != 0)
- rc = rc2;
- }
- lnet_ni_decref(ni);
- }
-
- LIBCFS_FREE(nids, maxnids * sizeof(*nids));
- return rc;
-#endif
+ struct list_head *e;
+ struct lnet_ni *ni = NULL;
+ lnet_remotenet_t *rnet;
+ __u32 dstnet = LNET_NIDNET(dstnid);
+ int hops;
+ int cpt;
+ __u32 order = 2;
+ struct list_head *rn_list;
+
+ /* if !local_nid_dist_zero, I don't return a distance of 0 ever
+ * (when lustre sees a distance of 0, it substitutes 0@lo), so I
+ * keep order 0 free for 0@lo and order 1 free for a local NID
+ * match */
+
+ LASSERT(the_lnet.ln_refcount > 0);
+
+ cpt = lnet_net_lock_current();
+
+ while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
+ if (ni->ni_nid == dstnid) {
+ if (srcnidp != NULL)
+ *srcnidp = dstnid;
+ if (orderp != NULL) {
+ if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
+ *orderp = 0;
+ else
+ *orderp = 1;
+ }
+ lnet_net_unlock(cpt);
+
+ return local_nid_dist_zero ? 0 : 1;
+ }
+
+ if (LNET_NIDNET(ni->ni_nid) == dstnet) {
+ /* Check if ni was originally created in
+ * current net namespace.
+ * If not, assign order above 0xffff0000,
+ * to make this ni not a priority. */
+ if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
+ order += 0xffff0000;
+
+ if (srcnidp != NULL)
+ *srcnidp = ni->ni_nid;
+ if (orderp != NULL)
+ *orderp = order;
+ lnet_net_unlock(cpt);
+ return 1;
+ }
+
+ order++;
+ }
+
+ rn_list = lnet_net2rnethash(dstnet);
+ list_for_each(e, rn_list) {
+ rnet = list_entry(e, lnet_remotenet_t, lrn_list);
+
+ if (rnet->lrn_net == dstnet) {
+ lnet_route_t *route;
+ lnet_route_t *shortest = NULL;
+ __u32 shortest_hops = LNET_UNDEFINED_HOPS;
+ __u32 route_hops;
+
+ LASSERT(!list_empty(&rnet->lrn_routes));
+
+ list_for_each_entry(route, &rnet->lrn_routes,
+ lr_list) {
+ route_hops = route->lr_hops;
+ if (route_hops == LNET_UNDEFINED_HOPS)
+ route_hops = 1;
+ if (shortest == NULL ||
+ route_hops < shortest_hops) {
+ shortest = route;
+ shortest_hops = route_hops;
+ }
+ }
+
+ LASSERT(shortest != NULL);
+ hops = shortest_hops;
+ if (srcnidp != NULL) {
+ ni = lnet_get_next_ni_locked(
+ shortest->lr_gateway->lpni_net,
+ NULL);
+ *srcnidp = ni->ni_nid;
+ }
+ if (orderp != NULL)
+ *orderp = order;
+ lnet_net_unlock(cpt);
+ return hops + 1;
+ }
+ order++;
+ }
+
+ lnet_net_unlock(cpt);
+ return -EHOSTUNREACH;
}
+EXPORT_SYMBOL(LNetDist);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff