X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Flnet%2Flib-move.c;h=7657ea8c00c9c678aa2a7fb5a036d6006ca35bfb;hp=e9ef01540f41e475925366e9afa51de91883c562;hb=5655ee76f5c7573c9ad076a92df6388f27b65196;hpb=00f255b8c00dff66481a6ab22391869217b5d8af diff --git a/lnet/lnet/lib-move.c b/lnet/lnet/lib-move.c index e9ef015..7657ea8 100644 --- a/lnet/lnet/lib-move.c +++ b/lnet/lnet/lib-move.c @@ -1,359 +1,249 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * lib/lib-move.c - * Data movement routines + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * Copyright (c) 2001-2003 Cluster File Systems, Inc. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * This file is part of Lustre, http://www.lustre.org + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * 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. * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * GPL HEADER END + */ +/* + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2012, 2013, Intel Corporation. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * lnet/lnet/lib-move.c + * + * Data movement routines */ -#define DEBUG_SUBSYSTEM S_PORTALS - -#ifndef __KERNEL__ -# include -#else -# include -#endif -#include -#include - -/* forward ref */ -static void lib_commit_md (lib_nal_t *nal, lib_md_t *md, lib_msg_t *msg); - -static lib_md_t * -lib_match_md(lib_nal_t *nal, int index, int op_mask, - ptl_nid_t src_nid, ptl_pid_t src_pid, - ptl_size_t rlength, ptl_size_t roffset, - ptl_match_bits_t match_bits, lib_msg_t *msg, - ptl_size_t *mlength_out, ptl_size_t *offset_out) -{ - lib_ni_t *ni = &nal->libnal_ni; - struct list_head *match_list = &ni->ni_portals.tbl[index]; - struct list_head *tmp; - lib_me_t *me; - lib_md_t *md; - ptl_size_t mlength; - ptl_size_t offset; - ENTRY; - - CDEBUG (D_NET, "Request from "LPU64".%d of length %d into portal %d " - "MB="LPX64"\n", src_nid, src_pid, rlength, index, match_bits); - - if (index < 0 || index >= ni->ni_portals.size) { - CERROR("Invalid portal %d not in [0-%d]\n", - index, ni->ni_portals.size); - goto failed; - } - - list_for_each (tmp, match_list) { - me = list_entry(tmp, lib_me_t, me_list); - md = me->md; - - /* ME attached but MD not attached yet */ - if (md == NULL) - continue; - - LASSERT (me == md->me); - - /* mismatched MD op */ - if ((md->options & op_mask) == 0) - continue; - - /* MD exhausted */ - if (lib_md_exhausted(md)) - continue; - - /* mismatched ME nid/pid? */ - if (me->match_id.nid != PTL_NID_ANY && - me->match_id.nid != src_nid) - continue; - - CDEBUG(D_NET, "match_id.pid [%x], src_pid [%x]\n", - me->match_id.pid, src_pid); - - if (me->match_id.pid != PTL_PID_ANY && - me->match_id.pid != src_pid) - continue; - - /* mismatched ME matchbits? */ - if (((me->match_bits ^ match_bits) & ~me->ignore_bits) != 0) - continue; - - /* Hurrah! This _is_ a match; check it out... */ - - if ((md->options & PTL_MD_MANAGE_REMOTE) == 0) - offset = md->offset; - else - offset = roffset; - - if ((md->options & PTL_MD_MAX_SIZE) != 0) { - mlength = md->max_size; - LASSERT (md->offset + mlength <= md->length); - } else { - mlength = md->length - offset; - } +#define DEBUG_SUBSYSTEM S_LNET - if (rlength <= mlength) { /* fits in allowed space */ - mlength = rlength; - } else if ((md->options & PTL_MD_TRUNCATE) == 0) { - /* this packet _really_ is too big */ - CERROR("Matching packet %d too big: %d left, " - "%d allowed\n", rlength, md->length - offset, - mlength); - goto failed; - } +#include - /* Commit to this ME/MD */ - CDEBUG(D_NET, "Incoming %s index %x from "LPU64"/%u of " - "length %d/%d into md "LPX64" [%d] + %d\n", - (op_mask == PTL_MD_OP_PUT) ? "put" : "get", - index, src_nid, src_pid, mlength, rlength, - md->md_lh.lh_cookie, md->md_niov, offset); - - lib_commit_md(nal, md, msg); - md->offset = offset + mlength; - - /* NB Caller sets ev.type and ev.hdr_data */ - msg->ev.initiator.nid = src_nid; - msg->ev.initiator.pid = src_pid; - msg->ev.pt_index = index; - msg->ev.match_bits = match_bits; - msg->ev.rlength = rlength; - msg->ev.mlength = mlength; - msg->ev.offset = offset; - - lib_md_deconstruct(nal, md, &msg->ev.md); - ptl_md2handle(&msg->ev.md_handle, nal, md); - - *offset_out = offset; - *mlength_out = mlength; - - /* Auto-unlink NOW, so the ME gets unlinked if required. - * We bumped md->pending above so the MD just gets flagged - * for unlink when it is finalized. */ - if ((md->md_flags & PTL_MD_FLAG_AUTO_UNLINK) != 0 && - lib_md_exhausted(md)) - lib_md_unlink(nal, md); - - RETURN (md); - } +/** lnet message has credit and can be submitted to lnd for send/receive */ +#define LNET_CREDIT_OK 0 +/** lnet message is waiting for credit */ +#define LNET_CREDIT_WAIT 1 - failed: - CERROR (LPU64": Dropping %s from "LPU64".%d portal %d match "LPX64 - " offset %d length %d: no match\n", - ni->ni_pid.nid, (op_mask == PTL_MD_OP_GET) ? "GET" : "PUT", - src_nid, src_pid, index, match_bits, roffset, rlength); - RETURN(NULL); -} +static int local_nid_dist_zero = 1; +CFS_MODULE_PARM(local_nid_dist_zero, "i", int, 0444, + "Reserved"); -int lib_api_fail_nid (nal_t *apinal, ptl_nid_t nid, unsigned int threshold) +int +lnet_fail_nid(lnet_nid_t nid, unsigned int threshold) { - lib_nal_t *nal = apinal->nal_data; - lib_test_peer_t *tp; - unsigned long flags; - struct list_head *el; - struct list_head *next; - struct list_head cull; + lnet_test_peer_t *tp; + struct list_head *el; + struct list_head *next; + struct list_head cull; + + LASSERT(the_lnet.ln_init); + /* NB: use lnet_net_lock(0) to serialize operations on test peers */ if (threshold != 0) { /* Adding a new entry */ - PORTAL_ALLOC(tp, sizeof(*tp)); + LIBCFS_ALLOC(tp, sizeof(*tp)); if (tp == NULL) - return PTL_NO_SPACE; + return -ENOMEM; tp->tp_nid = nid; tp->tp_threshold = threshold; - LIB_LOCK(nal, flags); - list_add_tail (&tp->tp_list, &nal->libnal_ni.ni_test_peers); - LIB_UNLOCK(nal, flags); - return PTL_OK; - } + lnet_net_lock(0); + list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers); + lnet_net_unlock(0); + return 0; + } - /* removing entries */ - CFS_INIT_LIST_HEAD (&cull); + /* removing entries */ + INIT_LIST_HEAD(&cull); - LIB_LOCK(nal, flags); + lnet_net_lock(0); - list_for_each_safe (el, next, &nal->libnal_ni.ni_test_peers) { - tp = list_entry (el, lib_test_peer_t, tp_list); + 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 == PTL_NID_ANY || /* removing all entries */ - tp->tp_nid == nid) /* matched this one */ - { - list_del (&tp->tp_list); - list_add (&tp->tp_list, &cull); - } - } + 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); + } + } - LIB_UNLOCK(nal, flags); + lnet_net_unlock(0); - while (!list_empty (&cull)) { - tp = list_entry (cull.next, lib_test_peer_t, tp_list); + while (!list_empty(&cull)) { + tp = list_entry(cull.next, lnet_test_peer_t, tp_list); - list_del (&tp->tp_list); - PORTAL_FREE(tp, sizeof (*tp)); - } - return PTL_OK; + list_del(&tp->tp_list); + LIBCFS_FREE(tp, sizeof(*tp)); + } + return 0; } static int -fail_peer (lib_nal_t *nal, ptl_nid_t nid, int outgoing) +fail_peer (lnet_nid_t nid, int outgoing) { - lib_test_peer_t *tp; - struct list_head *el; - struct list_head *next; - unsigned long flags; - struct list_head cull; - int fail = 0; - - CFS_INIT_LIST_HEAD (&cull); - - LIB_LOCK (nal, flags); - - list_for_each_safe (el, next, &nal->libnal_ni.ni_test_peers) { - tp = list_entry (el, lib_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 == PTL_NID_ANY || /* fail every peer */ - nid == tp->tp_nid) { /* fail this peer */ - fail = 1; - - if (tp->tp_threshold != PTL_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; - } - } - - LIB_UNLOCK (nal, flags); - - while (!list_empty (&cull)) { - tp = list_entry (cull.next, lib_test_peer_t, tp_list); - list_del (&tp->tp_list); - - PORTAL_FREE(tp, sizeof (*tp)); - } - - return (fail); + 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; } -ptl_size_t -lib_iov_nob (int niov, struct iovec *iov) +unsigned int +lnet_iov_nob (unsigned int niov, struct iovec *iov) { - ptl_size_t nob = 0; + unsigned int nob = 0; + LASSERT(niov == 0 || iov != NULL); while (niov-- > 0) nob += (iov++)->iov_len; return (nob); } +EXPORT_SYMBOL(lnet_iov_nob); void -lib_copy_iov2buf (char *dest, int niov, struct iovec *iov, - ptl_size_t offset, ptl_size_t len) +lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset, + unsigned int nsiov, struct iovec *siov, unsigned int soffset, + unsigned int nob) { - ptl_size_t nob; + /* NB diov, siov are READ-ONLY */ + unsigned int this_nob; - if (len == 0) + if (nob == 0) return; - /* skip complete frags before 'offset' */ - LASSERT (niov > 0); - while (offset >= iov->iov_len) { - offset -= iov->iov_len; - iov++; - niov--; - LASSERT (niov > 0); + /* skip complete frags before 'doffset' */ + LASSERT (ndiov > 0); + while (doffset >= diov->iov_len) { + doffset -= diov->iov_len; + diov++; + ndiov--; + LASSERT (ndiov > 0); } - do { - LASSERT (niov > 0); - nob = MIN (iov->iov_len - offset, len); - memcpy (dest, iov->iov_base + offset, nob); - - len -= nob; - dest += nob; - niov--; - iov++; - offset = 0; - } while (len > 0); -} - -void -lib_copy_buf2iov (int niov, struct iovec *iov, ptl_size_t offset, - char *src, ptl_size_t len) -{ - ptl_size_t nob; - - if (len == 0) - return; - - /* skip complete frags before 'offset' */ - LASSERT (niov > 0); - while (offset >= iov->iov_len) { - offset -= iov->iov_len; - iov++; - niov--; - LASSERT (niov > 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 (niov > 0); - nob = MIN (iov->iov_len - offset, len); - memcpy (iov->iov_base + offset, src, nob); + 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; + } - len -= nob; - src += nob; - niov--; - iov++; - offset = 0; - } while (len > 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 -lib_extract_iov (int dst_niov, struct iovec *dst, - int src_niov, struct iovec *src, - ptl_size_t offset, ptl_size_t len) +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' */ - ptl_size_t frag_len; - int niov; + unsigned int frag_len; + unsigned int niov; if (len == 0) /* no data => */ return (0); /* no frags */ @@ -369,7 +259,7 @@ lib_extract_iov (int dst_niov, struct iovec *dst, niov = 1; for (;;) { LASSERT (src_niov > 0); - LASSERT (niov <= dst_niov); + LASSERT ((int)niov <= dst_niov); frag_len = src->iov_len - offset; dst->iov_base = ((char *)src->iov_base) + offset; @@ -389,132 +279,294 @@ lib_extract_iov (int dst_niov, struct iovec *dst, offset = 0; } } +EXPORT_SYMBOL(lnet_extract_iov); #ifndef __KERNEL__ -ptl_size_t -lib_kiov_nob (int niov, ptl_kiov_t *kiov) +unsigned int +lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov) { LASSERT (0); return (0); } void -lib_copy_kiov2buf (char *dest, int niov, ptl_kiov_t *kiov, - ptl_size_t offset, ptl_size_t len) +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); } void -lib_copy_buf2kiov (int niov, ptl_kiov_t *kiov, ptl_size_t offset, - char *src, ptl_size_t len) +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); +} + +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) { LASSERT (0); } int -lib_extract_kiov (int dst_niov, ptl_kiov_t *dst, - int src_niov, ptl_kiov_t *src, - ptl_size_t offset, ptl_size_t len) +lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst, + int src_niov, lnet_kiov_t *src, + unsigned int offset, unsigned int len) { LASSERT (0); } -#else +#else /* __KERNEL__ */ -ptl_size_t -lib_kiov_nob (int niov, ptl_kiov_t *kiov) +unsigned int +lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov) { - ptl_size_t nob = 0; + unsigned int nob = 0; + LASSERT(niov == 0 || kiov != NULL); while (niov-- > 0) nob += (kiov++)->kiov_len; return (nob); } +EXPORT_SYMBOL(lnet_kiov_nob); void -lib_copy_kiov2buf (char *dest, int niov, ptl_kiov_t *kiov, - ptl_size_t offset, ptl_size_t len) +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) { - ptl_size_t nob; - char *addr; + /* 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); + } - if (len == 0) - return; + 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); - LASSERT (!in_interrupt ()); +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) +{ + /* NB iov, kiov are READ-ONLY */ + unsigned int this_nob; + char *addr = NULL; - LASSERT (niov > 0); - while (offset > kiov->kiov_len) { - offset -= kiov->kiov_len; - kiov++; + if (nob == 0) + return; + + LASSERT (!in_interrupt ()); + + LASSERT (niov > 0); + while (iovoffset >= iov->iov_len) { + iovoffset -= iov->iov_len; + iov++; niov--; LASSERT (niov > 0); } - do{ + LASSERT (nkiov > 0); + while (kiovoffset >= kiov->kiov_len) { + kiovoffset -= kiov->kiov_len; + kiov++; + nkiov--; + LASSERT (nkiov > 0); + } + + do { LASSERT (niov > 0); - nob = MIN (kiov->kiov_len - offset, len); + LASSERT (nkiov > 0); + this_nob = MIN(iov->iov_len - iovoffset, + kiov->kiov_len - kiovoffset); + this_nob = MIN(this_nob, nob); - addr = ((char *)cfs_kmap (kiov->kiov_page)) + kiov->kiov_offset + offset; - memcpy (dest, addr, nob); - cfs_kunmap (kiov->kiov_page); + if (addr == NULL) + addr = ((char *)kmap(kiov->kiov_page)) + + kiov->kiov_offset + kiovoffset; - len -= nob; - dest += nob; - niov--; - kiov++; - offset = 0; - } while (len > 0); + 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 -lib_copy_buf2kiov (int niov, ptl_kiov_t *kiov, ptl_size_t offset, - char *src, ptl_size_t len) +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) { - ptl_size_t nob; - char *addr; + /* NB kiov, iov are READ-ONLY */ + unsigned int this_nob; + char *addr = NULL; - if (len == 0) - return; + if (nob == 0) + return; - LASSERT (!in_interrupt ()); + LASSERT (!in_interrupt ()); - LASSERT (niov > 0); - while (offset >= kiov->kiov_len) { - offset -= kiov->kiov_len; + 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); - nob = MIN (kiov->kiov_len - offset, len); - - addr = ((char *)cfs_kmap (kiov->kiov_page)) + kiov->kiov_offset + offset; - memcpy (addr, src, nob); - cfs_kunmap (kiov->kiov_page); - - len -= nob; - src += nob; - niov--; - kiov++; - offset = 0; - } while (len > 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 -lib_extract_kiov (int dst_niov, ptl_kiov_t *dst, - int src_niov, ptl_kiov_t *src, - ptl_size_t offset, ptl_size_t 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' */ - ptl_size_t frag_len; - int niov; + unsigned int frag_len; + unsigned int niov; if (len == 0) /* no data => */ return (0); /* no frags */ @@ -530,20 +582,20 @@ lib_extract_kiov (int dst_niov, ptl_kiov_t *dst, niov = 1; for (;;) { LASSERT (src_niov > 0); - LASSERT (niov <= dst_niov); + 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); - } + if (len <= frag_len) { + dst->kiov_len = len; + LASSERT (dst->kiov_offset + dst->kiov_len <= PAGE_CACHE_SIZE); + return niov; + } - dst->kiov_len = frag_len; - LASSERT (dst->kiov_offset + dst->kiov_len <= PAGE_SIZE); + dst->kiov_len = frag_len; + LASSERT (dst->kiov_offset + dst->kiov_len <= PAGE_CACHE_SIZE); len -= frag_len; dst++; @@ -553,370 +605,1213 @@ lib_extract_kiov (int dst_niov, ptl_kiov_t *dst, offset = 0; } } +EXPORT_SYMBOL(lnet_extract_kiov); #endif -ptl_err_t -lib_recv (lib_nal_t *nal, void *private, lib_msg_t *msg, lib_md_t *md, - ptl_size_t offset, ptl_size_t mlen, ptl_size_t rlen) +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) { - if (mlen == 0) - return (nal->libnal_recv(nal, private, msg, - 0, NULL, - offset, mlen, rlen)); - - if ((md->options & PTL_MD_KIOV) == 0) - return (nal->libnal_recv(nal, private, msg, - md->md_niov, md->md_iov.iov, - offset, mlen, rlen)); - - return (nal->libnal_recv_pages(nal, private, msg, - md->md_niov, md->md_iov.kiov, - offset, mlen, 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); + 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_lnd->lnd_recv)(ni, private, msg, delayed, + niov, iov, kiov, offset, mlen, rlen); + if (rc < 0) + lnet_finalize(ni, msg, rc); } -ptl_err_t -lib_send (lib_nal_t *nal, void *private, lib_msg_t *msg, - ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid, - lib_md_t *md, ptl_size_t offset, ptl_size_t len) +void +lnet_setpayloadbuffer(lnet_msg_t *msg) { - if (len == 0) - return (nal->libnal_send(nal, private, msg, - hdr, type, nid, pid, - 0, NULL, - offset, len)); - - if ((md->options & PTL_MD_KIOV) == 0) - return (nal->libnal_send(nal, private, msg, - hdr, type, nid, pid, - md->md_niov, md->md_iov.iov, - offset, len)); - - return (nal->libnal_send_pages(nal, private, msg, - hdr, type, nid, pid, - md->md_niov, md->md_iov.kiov, - offset, len)); + 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; } -static void -lib_commit_md (lib_nal_t *nal, lib_md_t *md, lib_msg_t *msg) +void +lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target, + unsigned int offset, unsigned int len) { - /* ALWAYS called holding the LIB_LOCK */ - lib_counters_t *counters = &nal->libnal_ni.ni_counters; - - /* 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 lib_finalize() signals completion. */ - msg->md = md; - - md->pending++; - if (md->threshold != PTL_MD_THRESH_INF) { - LASSERT (md->threshold > 0); - md->threshold--; - } + 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); +} + +void +lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg) +{ + void *priv = msg->msg_private; + int rc; - counters->msgs_alloc++; - if (counters->msgs_alloc > counters->msgs_max) - counters->msgs_max = counters->msgs_alloc; + LASSERT (!in_interrupt ()); + LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND || + (msg->msg_txcredit && msg->msg_peertxcredit)); - list_add (&msg->msg_list, &nal->libnal_ni.ni_active_msgs); + rc = (ni->ni_lnd->lnd_send)(ni, priv, msg); + if (rc < 0) + lnet_finalize(ni, msg, rc); } -static void -lib_drop_message (lib_nal_t *nal, void *private, ptl_hdr_t *hdr) +int +lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg) { - unsigned long flags; + int rc; + + LASSERT(!msg->msg_sending); + LASSERT(msg->msg_receiving); + LASSERT(!msg->msg_rx_ready_delay); + LASSERT(ni->ni_lnd->lnd_eager_recv != NULL); + + msg->msg_rx_ready_delay = 1; + 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(msg->msg_rxpeer->lp_nid), + libcfs_id2str(msg->msg_target), rc); + LASSERT(rc < 0); /* required by my callers */ + } + + return rc; +} + +/* NB: caller shall hold a ref on 'lp' as I'd drop lnet_net_lock */ +void +lnet_ni_query_locked(lnet_ni_t *ni, lnet_peer_t *lp) +{ + cfs_time_t last_alive = 0; + + LASSERT(lnet_peer_aliveness_enabled(lp)); + LASSERT(ni->ni_lnd->lnd_query != NULL); - /* CAVEAT EMPTOR: this only drops messages that we've not committed - * to receive (init_msg() not called) and therefore can't cause an - * event. */ + lnet_net_unlock(lp->lp_cpt); + (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive); + lnet_net_lock(lp->lp_cpt); - LIB_LOCK(nal, flags); - nal->libnal_ni.ni_counters.drop_count++; - nal->libnal_ni.ni_counters.drop_length += hdr->payload_length; - LIB_UNLOCK(nal, flags); + lp->lp_last_query = cfs_time_current(); - /* NULL msg => if NAL calls lib_finalize it will be a noop */ - (void) lib_recv(nal, private, NULL, NULL, 0, 0, hdr->payload_length); + if (last_alive != 0) /* NI has updated timestamp */ + lp->lp_last_alive = last_alive; } -/* - * Incoming messages have a ptl_msg_t object associated with them - * by the library. This object encapsulates the state of the - * message and allows the NAL to do non-blocking receives or sends - * of long messages. +/* NB: always called with lnet_net_lock held */ +static inline int +lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now) +{ + 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()). + */ + if (!lp->lp_alive && lp->lp_alive_count > 0 && + cfs_time_aftereq(lp->lp_timestamp, lp->lp_last_alive)) + return 0; + + deadline = cfs_time_add(lp->lp_last_alive, + cfs_time_seconds(lp->lp_ni->ni_peertimeout)); + 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 lp_last_alive at peer creation is assumed. + */ + if (alive && !lp->lp_alive && + !(lnet_isrouter(lp) && lp->lp_alive_count == 0)) + lnet_notify_locked(lp, 0, 1, lp->lp_last_alive); + + return alive; +} + + +/* NB: returns 1 when alive, 0 when dead, negative when error; + * may drop the lnet_net_lock */ +int +lnet_peer_alive_locked (lnet_peer_t *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->lp_last_query != 0) { + static const int lnet_queryinterval = 1; + + cfs_time_t next_query = + cfs_time_add(lp->lp_last_query, + cfs_time_seconds(lnet_queryinterval)); + + if (cfs_time_before(now, next_query)) { + if (lp->lp_alive) + CWARN("Unexpected aliveness of peer %s: " + "%d < %d (%d/%d)\n", + libcfs_nid2str(lp->lp_nid), + (int)now, (int)next_query, + lnet_queryinterval, + lp->lp_ni->ni_peertimeout); + return 0; + } + } + + /* query NI for latest aliveness news */ + lnet_ni_query_locked(lp->lp_ni, lp); + + if (lnet_peer_is_alive(lp, now)) + return 1; + + lnet_notify_locked(lp, 0, 0, lp->lp_last_alive); + return 0; +} + +/** + * \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 ptl_err_t -parse_put(lib_nal_t *nal, ptl_hdr_t *hdr, void *private, lib_msg_t *msg) +static int +lnet_post_send_locked(lnet_msg_t *msg, int do_send) +{ + lnet_peer_t *lp = msg->msg_txpeer; + lnet_ni_t *ni = lp->lp_ni; + 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(lp) == 0) { + the_lnet.ln_counters[cpt]->drop_count++; + the_lnet.ln_counters[cpt]->drop_length += msg->msg_len; + lnet_net_unlock(cpt); + + 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) { + 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_tx_delayed = 1; + list_add_tail(&msg->msg_list, &lp->lp_txq); + return LNET_CREDIT_WAIT; + } + } + + if (!msg->msg_txcredit) { + LASSERT((tq->tq_credits < 0) == + !list_empty(&tq->tq_delayed)); + + msg->msg_txcredit = 1; + tq->tq_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; +} + +#ifdef __KERNEL__ + +lnet_rtrbufpool_t * +lnet_msg2bufpool(lnet_msg_t *msg) +{ + lnet_rtrbufpool_t *rbp; + int cpt; + + LASSERT(msg->msg_rx_committed); + + cpt = msg->msg_rx_cpt; + rbp = &the_lnet.ln_rtrpools[cpt][0]; + + LASSERT(msg->msg_len <= LNET_MTU); + while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_CACHE_SIZE) { + rbp++; + LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]); + } + + return rbp; +} + +int +lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv) { - lib_ni_t *ni = &nal->libnal_ni; - ptl_size_t mlength = 0; - ptl_size_t offset = 0; - ptl_err_t rc; - lib_md_t *md; - unsigned long flags; - - /* 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); - - LIB_LOCK(nal, flags); - - md = lib_match_md(nal, hdr->msg.put.ptl_index, PTL_MD_OP_PUT, - hdr->src_nid, hdr->src_pid, - hdr->payload_length, hdr->msg.put.offset, - hdr->msg.put.match_bits, msg, - &mlength, &offset); - if (md == NULL) { - LIB_UNLOCK(nal, flags); - return (PTL_FAIL); + /* 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 */ + 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 receive delayed messages */ + LASSERT(!do_recv || msg->msg_rx_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_rx_ready_delay); + msg->msg_rx_delayed = 1; + list_add_tail(&msg->msg_list, &lp->lp_rtrq); + return LNET_CREDIT_WAIT; + } + } + + 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(lp->lp_ni, msg->msg_private, msg, 1, + 0, msg->msg_len, msg->msg_len); + lnet_net_lock(cpt); + } + return LNET_CREDIT_OK; +} +#endif + +void +lnet_return_tx_credits_locked(lnet_msg_t *msg) +{ + lnet_peer_t *txpeer = msg->msg_txpeer; + lnet_msg_t *msg2; + + if (msg->msg_txcredit) { + struct lnet_ni *ni = txpeer->lp_ni; + 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++; + 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_txpeer->lp_ni == ni); + LASSERT(msg2->msg_tx_delayed); + + (void) lnet_post_send_locked(msg2, 1); + } } - msg->ev.type = PTL_EVENT_PUT_END; - msg->ev.hdr_data = hdr->msg.put.hdr_data; + if (msg->msg_peertxcredit) { + /* give back peer txcredits */ + msg->msg_peertxcredit = 0; - if (!ptl_is_wire_handle_none(&hdr->msg.put.ack_wmd) && - !(md->options & PTL_MD_ACK_DISABLE)) { - msg->ack_wmd = hdr->msg.put.ack_wmd; + 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_tx_delayed); + + (void) lnet_post_send_locked(msg2, 1); + } + } + + if (txpeer != NULL) { + msg->msg_txpeer = NULL; + lnet_peer_decref_locked(txpeer); } +} - ni->ni_counters.recv_count++; - ni->ni_counters.recv_length += mlength; +#ifdef __KERNEL__ +void +lnet_schedule_blocked_locked(lnet_rtrbufpool_t *rbp) +{ + lnet_msg_t *msg; - LIB_UNLOCK(nal, flags); + if (list_empty(&rbp->rbp_msgs)) + return; + msg = list_entry(rbp->rbp_msgs.next, + lnet_msg_t, msg_list); + list_del(&msg->msg_list); - rc = lib_recv(nal, private, msg, md, offset, mlength, - hdr->payload_length); - if (rc != PTL_OK) - CERROR(LPU64": error on receiving PUT from "LPU64": %d\n", - ni->ni_pid.nid, hdr->src_nid, rc); + (void)lnet_post_routed_recv_locked(msg, 1); +} +#endif + +void +lnet_drop_routed_msgs_locked(struct list_head *list, int cpt) +{ + lnet_msg_t *msg; + lnet_msg_t *tmp; + struct list_head drop; + + INIT_LIST_HEAD(&drop); + + list_splice_init(list, &drop); + + lnet_net_unlock(cpt); - return (rc); + list_for_each_entry_safe(msg, tmp, &drop, msg_list) { + lnet_ni_recv(msg->msg_rxpeer->lp_ni, msg->msg_private, NULL, + 0, 0, 0, msg->msg_hdr.payload_length); + list_del_init(&msg->msg_list); + lnet_finalize(NULL, msg, -ECANCELED); + } + + lnet_net_lock(cpt); +} + +void +lnet_return_rx_credits_locked(lnet_msg_t *msg) +{ + lnet_peer_t *rxpeer = msg->msg_rxpeer; +#ifdef __KERNEL__ + 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 (rbp->rbp_credits >= rbp->rbp_nbuffers) { + /* Discard this buffer so we don't have too many. */ + lnet_destroy_rtrbuf(rb, rbp->rbp_npages); + } 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; + + LASSERT((rxpeer->lp_rtrcredits < 0) == + !list_empty(&rxpeer->lp_rtrq)); + + rxpeer->lp_rtrcredits++; + + /* drop all messages which are queued to be routed on that + * peer. */ + if (!the_lnet.ln_routing) { + lnet_drop_routed_msgs_locked(&rxpeer->lp_rtrq, + msg->msg_rx_cpt); + } else 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); + } } -static ptl_err_t -parse_get(lib_nal_t *nal, ptl_hdr_t *hdr, void *private, lib_msg_t *msg) +static int +lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2) { - lib_ni_t *ni = &nal->libnal_ni; - ptl_size_t mlength = 0; - ptl_size_t offset = 0; - lib_md_t *md; - ptl_hdr_t reply; - unsigned long flags; - 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); - - LIB_LOCK(nal, flags); - - md = lib_match_md(nal, hdr->msg.get.ptl_index, PTL_MD_OP_GET, - hdr->src_nid, hdr->src_pid, - hdr->msg.get.sink_length, hdr->msg.get.src_offset, - hdr->msg.get.match_bits, msg, - &mlength, &offset); - if (md == NULL) { - LIB_UNLOCK(nal, flags); - return (PTL_FAIL); + lnet_peer_t *p1 = r1->lr_gateway; + lnet_peer_t *p2 = r2->lr_gateway; + + if (r1->lr_priority < r2->lr_priority) + return 1; + + if (r1->lr_priority > r2->lr_priority) + return -1; + + if (r1->lr_hops < r2->lr_hops) + return 1; + + if (r1->lr_hops > r2->lr_hops) + return -1; + + 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; + + if (r1->lr_seq - r2->lr_seq <= 0) + return 1; + + return -1; +} + +static lnet_peer_t * +lnet_find_route_locked(lnet_ni_t *ni, 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 *lp_best; + struct lnet_peer *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_net_locked(LNET_NIDNET(target)); + if (rnet == NULL) + return NULL; + + lp_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 (ni != NULL && lp->lp_ni != ni) + continue; + + if (lp->lp_nid == rtr_nid) /* it's pre-determined router */ + return lp; + + if (lp_best == NULL) { + best_route = last_route = route; + lp_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; + lp_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 lp_best; +} + +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; + struct lnet_ni *src_ni; + struct lnet_ni *local_ni; + struct lnet_peer *lp; + int cpt; + int cpt2; + int rc; + + /* 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); + + msg->msg_sending = 1; + + LASSERT(!msg->msg_tx_committed); + cpt = lnet_cpt_of_nid(rtr_nid == LNET_NID_ANY ? dst_nid : rtr_nid); + again: + lnet_net_lock(cpt); + + if (the_lnet.ln_shutdown) { + lnet_net_unlock(cpt); + return -ESHUTDOWN; + } + + if (src_nid == LNET_NID_ANY) { + src_ni = NULL; + } else { + src_ni = lnet_nid2ni_locked(src_nid, cpt); + if (src_ni == NULL) { + 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; + } + LASSERT (!msg->msg_routing); } - msg->ev.type = PTL_EVENT_GET_END; - msg->ev.hdr_data = 0; + /* Is this for someone on a local network? */ + local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid), cpt); + + 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, cpt); + } else { + lnet_ni_decref_locked(local_ni, cpt); + lnet_ni_decref_locked(src_ni, cpt); + lnet_net_unlock(cpt); + LCONSOLE_WARN("No route to %s via from %s\n", + libcfs_nid2str(dst_nid), + libcfs_nid2str(src_nid)); + return -EINVAL; + } + + LASSERT(src_nid != LNET_NID_ANY); + lnet_msg_commit(msg, cpt); + + if (!msg->msg_routing) + msg->msg_hdr.src_nid = cpu_to_le64(src_nid); + + if (src_ni == the_lnet.ln_loni) { + /* No send credit hassles with LOLND */ + lnet_net_unlock(cpt); + lnet_ni_send(src_ni, msg); + + lnet_net_lock(cpt); + lnet_ni_decref_locked(src_ni, cpt); + lnet_net_unlock(cpt); + return 0; + } + + rc = lnet_nid2peer_locked(&lp, dst_nid, cpt); + /* lp has ref on src_ni; lose mine */ + lnet_ni_decref_locked(src_ni, cpt); + if (rc != 0) { + lnet_net_unlock(cpt); + LCONSOLE_WARN("Error %d finding peer %s\n", rc, + libcfs_nid2str(dst_nid)); + /* ENOMEM or shutting down */ + return rc; + } + LASSERT (lp->lp_ni == src_ni); + } else { +#ifndef __KERNEL__ + lnet_net_unlock(cpt); + + /* NB + * - once application finishes computation, check here to update + * router states before it waits for pending IO in LNetEQPoll + * - recursion breaker: router checker sends no message + * to remote networks */ + if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) + lnet_router_checker(); + + lnet_net_lock(cpt); +#endif + /* sending to a remote network */ + lp = lnet_find_route_locked(src_ni, dst_nid, rtr_nid); + if (lp == NULL) { + if (src_ni != NULL) + lnet_ni_decref_locked(src_ni, cpt); + lnet_net_unlock(cpt); + + LCONSOLE_WARN("No route to %s via %s " + "(all routers down)\n", + libcfs_id2str(msg->msg_target), + libcfs_nid2str(src_nid)); + return -EHOSTUNREACH; + } + + /* rtr_nid is LNET_NID_ANY or NID of pre-determined router, + * it's possible that rtr_nid isn't LNET_NID_ANY and lp isn't + * pre-determined router, this can happen if router table + * was changed when we release the lock */ + if (rtr_nid != lp->lp_nid) { + cpt2 = lnet_cpt_of_nid_locked(lp->lp_nid); + if (cpt2 != cpt) { + if (src_ni != NULL) + lnet_ni_decref_locked(src_ni, cpt); + lnet_net_unlock(cpt); + + rtr_nid = lp->lp_nid; + cpt = cpt2; + goto again; + } + } + + CDEBUG(D_NET, "Best route to %s via %s for %s %d\n", + libcfs_nid2str(dst_nid), libcfs_nid2str(lp->lp_nid), + lnet_msgtyp2str(msg->msg_type), msg->msg_len); + + 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, cpt); + } + + lnet_peer_addref_locked(lp); + + LASSERT(src_nid != LNET_NID_ANY); + lnet_msg_commit(msg, cpt); + + if (!msg->msg_routing) { + /* I'm the source and now I know which NI to send on */ + 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 */ - ni->ni_counters.send_count++; - ni->ni_counters.send_length += mlength; + rc = lnet_post_send_locked(msg, 0); + lnet_net_unlock(cpt); - LIB_UNLOCK(nal, flags); + if (rc < 0) + return rc; - memset (&reply, 0, sizeof (reply)); - reply.type = cpu_to_le32(PTL_MSG_REPLY); - reply.dest_nid = cpu_to_le64(hdr->src_nid); - reply.dest_pid = cpu_to_le32(hdr->src_pid); - reply.src_nid = cpu_to_le64(ni->ni_pid.nid); - reply.src_pid = cpu_to_le32(ni->ni_pid.pid); - reply.payload_length = cpu_to_le32(mlength); + if (rc == LNET_CREDIT_OK) + lnet_ni_send(src_ni, msg); - reply.msg.reply.dst_wmd = hdr->msg.get.return_wmd; + return 0; /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT */ +} + +static void +lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob) +{ + lnet_net_lock(cpt); + the_lnet.ln_counters[cpt]->drop_count++; + the_lnet.ln_counters[cpt]->drop_length += nob; + lnet_net_unlock(cpt); + + lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob); +} + +static void +lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg) +{ + lnet_hdr_t *hdr = &msg->msg_hdr; + + if (msg->msg_wanted != 0) + lnet_setpayloadbuffer(msg); + + lnet_build_msg_event(msg, LNET_EVENT_PUT); + + /* 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_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed, + msg->msg_offset, msg->msg_wanted, hdr->payload_length); +} - /* NB call lib_send() _BEFORE_ lib_recv() completes the incoming - * message. Some NALs _require_ this to implement optimized GET */ +static int +lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg) +{ + lnet_hdr_t *hdr = &msg->msg_hdr; + struct lnet_match_info info; + int rc; + + /* 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); + + info.mi_id.nid = hdr->src_nid; + 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; + + msg->msg_rx_ready_delay = ni->ni_lnd->lnd_eager_recv == NULL; + + again: + rc = lnet_ptl_match_md(&info, msg); + switch (rc) { + default: + LBUG(); - rc = lib_send (nal, private, msg, &reply, PTL_MSG_REPLY, - hdr->src_nid, hdr->src_pid, md, offset, mlength); - if (rc != PTL_OK) - CERROR(LPU64": Unable to send REPLY for GET from "LPU64": %d\n", - ni->ni_pid.nid, hdr->src_nid, rc); + case LNET_MATCHMD_OK: + lnet_recv_put(ni, msg); + return 0; - /* Discard any junk after the hdr */ - (void) lib_recv(nal, private, NULL, NULL, 0, 0, hdr->payload_length); + case LNET_MATCHMD_NONE: + if (msg->msg_rx_delayed) /* attached on delayed list */ + return 0; - return (rc); + rc = lnet_ni_eager_recv(ni, msg); + if (rc == 0) + goto again; + /* fall through */ + + case LNET_MATCHMD_DROP: + CNETERR("Dropping PUT from %s portal %d match "LPU64 + " 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 ptl_err_t -parse_reply(lib_nal_t *nal, ptl_hdr_t *hdr, void *private, lib_msg_t *msg) +static int +lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get) { - lib_ni_t *ni = &nal->libnal_ni; - lib_md_t *md; - int rlength; - int length; - unsigned long flags; - ptl_err_t rc; + struct lnet_match_info info; + lnet_hdr_t *hdr = &msg->msg_hdr; + lnet_handle_wire_t 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); + + info.mi_id.nid = hdr->src_nid; + 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 "LPU64 + " 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, info.mi_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); - LIB_LOCK(nal, flags); + 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 = {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 = ptl_wire_handle2md(&hdr->msg.reply.dst_wmd, nal); - if (md == NULL || md->threshold == 0) { - CERROR (LPU64": Dropping REPLY from "LPU64" for %s MD "LPX64"."LPX64"\n", - ni->ni_pid.nid, hdr->src_nid, - md == NULL ? "invalid" : "inactive", + 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 "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); - LIB_UNLOCK(nal, flags); - return (PTL_FAIL); + lnet_res_unlock(cpt); + return ENOENT; /* +ve: OK but no match */ } - LASSERT (md->offset == 0); + LASSERT (md->md_offset == 0); - length = rlength = hdr->payload_length; + rlength = hdr->payload_length; + mlength = MIN(rlength, (int)md->md_length); - if (length > md->length) { - if ((md->options & PTL_MD_TRUNCATE) == 0) { - CERROR (LPU64": Dropping REPLY from "LPU64 - " length %d for MD "LPX64" would overflow (%d)\n", - ni->ni_pid.nid, hdr->src_nid, length, - hdr->msg.reply.dst_wmd.wh_object_cookie, - md->length); - LIB_UNLOCK(nal, flags); - return (PTL_FAIL); - } - length = md->length; + if (mlength < rlength && + (md->md_options & LNET_MD_TRUNCATE) == 0) { + CNETERR("%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_res_unlock(cpt); + return ENOENT; /* +ve: OK but no match */ } - CDEBUG(D_NET, "Reply from "LPU64" of length %d/%d into md "LPX64"\n", - hdr->src_nid, length, rlength, - hdr->msg.reply.dst_wmd.wh_object_cookie); - - lib_commit_md(nal, md, msg); - - msg->ev.type = PTL_EVENT_REPLY_END; - msg->ev.initiator.nid = hdr->src_nid; - msg->ev.initiator.pid = hdr->src_pid; - msg->ev.rlength = rlength; - msg->ev.mlength = length; - msg->ev.offset = 0; + 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); - lib_md_deconstruct(nal, md, &msg->ev.md); - ptl_md2handle(&msg->ev.md_handle, nal, md); + lnet_msg_attach_md(msg, md, 0, mlength); - ni->ni_counters.recv_count++; - ni->ni_counters.recv_length += length; + if (mlength != 0) + lnet_setpayloadbuffer(msg); - LIB_UNLOCK(nal, flags); + lnet_res_unlock(cpt); - rc = lib_recv(nal, private, msg, md, 0, length, rlength); - if (rc != PTL_OK) - CERROR(LPU64": error on receiving REPLY from "LPU64": %d\n", - ni->ni_pid.nid, hdr->src_nid, rc); + lnet_build_msg_event(msg, LNET_EVENT_REPLY); - return (rc); + lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength); + return 0; } -static ptl_err_t -parse_ack(lib_nal_t *nal, ptl_hdr_t *hdr, void *private, lib_msg_t *msg) +static int +lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg) { - lib_ni_t *ni = &nal->libnal_ni; - lib_md_t *md; - unsigned long flags; + 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); - LIB_LOCK(nal, flags); + 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 = ptl_wire_handle2md(&hdr->msg.ack.dst_wmd, nal); - if (md == NULL || md->threshold == 0) { - CDEBUG(D_INFO, LPU64": Dropping ACK from "LPU64" to %s MD " - LPX64"."LPX64"\n", ni->ni_pid.nid, hdr->src_nid, + 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); - LIB_UNLOCK(nal, flags); - return (PTL_FAIL); + lnet_res_unlock(cpt); + return ENOENT; /* +ve! */ } - CDEBUG(D_NET, LPU64": ACK from "LPU64" into md "LPX64"\n", - ni->ni_pid.nid, hdr->src_nid, + 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); - lib_commit_md(nal, md, msg); - - msg->ev.type = PTL_EVENT_ACK; - msg->ev.initiator.nid = hdr->src_nid; - msg->ev.initiator.pid = hdr->src_pid; - msg->ev.mlength = hdr->msg.ack.mlength; - msg->ev.match_bits = hdr->msg.ack.match_bits; + lnet_msg_attach_md(msg, md, 0, 0); - lib_md_deconstruct(nal, md, &msg->ev.md); - ptl_md2handle(&msg->ev.md_handle, nal, md); + lnet_res_unlock(cpt); - ni->ni_counters.recv_count++; + lnet_build_msg_event(msg, LNET_EVENT_ACK); - LIB_UNLOCK(nal, flags); - - /* We have received and matched up the ack OK, create the - * completion event now... */ - lib_finalize(nal, private, msg, PTL_OK); - - /* ...and now discard any junk after the hdr */ - (void) lib_recv(nal, private, NULL, NULL, 0, 0, hdr->payload_length); + lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len); + return 0; +} - return (PTL_OK); +/** + * \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 + */ +static int +lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg) +{ + int rc = 0; + +#ifdef __KERNEL__ + if (!the_lnet.ln_routing) + return -ECANCELED; + + if (msg->msg_rxpeer->lp_rtrcredits <= 0 || + lnet_msg2bufpool(msg)->rbp_credits <= 0) { + if (ni->ni_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); +#else + LBUG(); +#endif + return rc; } -static char * -hdr_type_string (ptl_hdr_t *hdr) +char * +lnet_msgtyp2str (int type) { - switch (hdr->type) { - case PTL_MSG_ACK: + switch (type) { + case LNET_MSG_ACK: return ("ACK"); - case PTL_MSG_PUT: + case LNET_MSG_PUT: return ("PUT"); - case PTL_MSG_GET: + case LNET_MSG_GET: return ("GET"); - case PTL_MSG_REPLY: + case LNET_MSG_REPLY: return ("REPLY"); - case PTL_MSG_HELLO: + case LNET_MSG_HELLO: return ("HELLO"); default: return (""); } } +EXPORT_SYMBOL(lnet_msgtyp2str); -void print_hdr(lib_nal_t *nal, ptl_hdr_t * hdr) +void +lnet_print_hdr(lnet_hdr_t * hdr) { - char *type_str = hdr_type_string (hdr); + 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 nid/pid "LPX64"/%u", hdr->src_nid, hdr->src_pid); - CWARN(" To nid/pid "LPX64"/%u\n", hdr->dest_nid, hdr->dest_pid); + CWARN(" From %s\n", libcfs_id2str(src)); + CWARN(" To %s\n", libcfs_id2str(dst)); switch (hdr->type) { default: break; - case PTL_MSG_PUT: + case LNET_MSG_PUT: CWARN(" Ptl index %d, ack md "LPX64"."LPX64", " - "match bits "LPX64"\n", + "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, @@ -926,9 +1821,9 @@ void print_hdr(lib_nal_t *nal, ptl_hdr_t * hdr) hdr->msg.put.hdr_data); break; - case PTL_MSG_GET: + case LNET_MSG_GET: CWARN(" Ptl index %d, return md "LPX64"."LPX64", " - "match bits "LPX64"\n", hdr->msg.get.ptl_index, + "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); @@ -937,7 +1832,7 @@ void print_hdr(lib_nal_t *nal, ptl_hdr_t * hdr) hdr->msg.get.src_offset); break; - case PTL_MSG_ACK: + case LNET_MSG_ACK: CWARN(" dst md "LPX64"."LPX64", " "manipulated length %d\n", hdr->msg.ack.dst_wmd.wh_interface_cookie, @@ -945,7 +1840,7 @@ void print_hdr(lib_nal_t *nal, ptl_hdr_t * hdr) hdr->msg.ack.mlength); break; - case PTL_MSG_REPLY: + case LNET_MSG_REPLY: CWARN(" dst md "LPX64"."LPX64", " "length %d\n", hdr->msg.reply.dst_wmd.wh_interface_cookie, @@ -953,475 +1848,815 @@ void print_hdr(lib_nal_t *nal, ptl_hdr_t * hdr) hdr->payload_length); } -} /* end of print_hdr() */ - +} -ptl_err_t -lib_parse(lib_nal_t *nal, ptl_hdr_t *hdr, void *private) +int +lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid, + void *private, int rdma_req) { - unsigned long flags; - ptl_err_t rc; - lib_msg_t *msg; + 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; + __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); + + 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 != cfs_time_current_sec()) { + /* NB: so far here is the only place to set NI status to "up */ + lnet_ni_lock(ni); + ni->ni_last_alive = cfs_time_current_sec(); + 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; + } + + 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; + /* 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; + } + + lnet_net_lock(cpt); + rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid, cpt); + if (rc != 0) { + lnet_net_unlock(cpt); + CERROR("%s, src %s: Dropping %s " + "(error %d looking up sender)\n", + libcfs_nid2str(from_nid), libcfs_nid2str(src_nid), + lnet_msgtyp2str(type), rc); + lnet_msg_free(msg); + goto drop; + } + + lnet_msg_commit(msg, cpt); + + 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); + + 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); + rc = -EPROTO; + goto free_drop; /* prevent an unused label if !kernel */ + } + + if (rc == 0) + return 0; + + LASSERT(rc == ENOENT); + + free_drop: + LASSERT(msg->msg_md == NULL); + lnet_finalize(ni, msg, rc); - /* NB we return PTL_OK if we manage to parse the header and believe - * it looks OK. Anything that goes wrong with receiving the - * message after that point is the responsibility of the NAL */ - - /* convert common fields to host byte order */ - hdr->type = le32_to_cpu(hdr->type); - hdr->src_nid = le64_to_cpu(hdr->src_nid); - hdr->src_pid = le32_to_cpu(hdr->src_pid); - hdr->dest_pid = le32_to_cpu(hdr->dest_pid); - hdr->payload_length = le32_to_cpu(hdr->payload_length); + drop: + lnet_drop_message(ni, cpt, private, payload_length); + return 0; +} +EXPORT_SYMBOL(lnet_parse); - switch (hdr->type) { - case PTL_MSG_HELLO: { - /* dest_nid is really ptl_magicversion_t */ - ptl_magicversion_t *mv = (ptl_magicversion_t *)&hdr->dest_nid; - - mv->magic = le32_to_cpu(mv->magic); - mv->version_major = le16_to_cpu(mv->version_major); - mv->version_minor = le16_to_cpu(mv->version_minor); - - if (mv->magic == PORTALS_PROTO_MAGIC && - mv->version_major == PORTALS_PROTO_VERSION_MAJOR && - mv->version_minor == PORTALS_PROTO_VERSION_MINOR) { - CWARN (LPU64": Dropping unexpected HELLO message: " - "magic %d, version %d.%d from "LPD64"\n", - nal->libnal_ni.ni_pid.nid, mv->magic, - mv->version_major, mv->version_minor, - hdr->src_nid); - - /* it's good but we don't want it */ - lib_drop_message(nal, private, hdr); - return PTL_OK; - } +void +lnet_drop_delayed_msg_list(struct list_head *head, char *reason) +{ + 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 "LPU64 + " 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_rxpeer->lp_ni, + msg->msg_rxpeer->lp_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_rxpeer->lp_ni, msg, -ENOENT); + } +} - /* we got garbage */ - CERROR (LPU64": Bad HELLO message: " - "magic %d, version %d.%d from "LPD64"\n", - nal->libnal_ni.ni_pid.nid, mv->magic, - mv->version_major, mv->version_minor, - hdr->src_nid); - return PTL_FAIL; - } +void +lnet_recv_delayed_msg_list(struct list_head *head) +{ + while (!list_empty(head)) { + lnet_msg_t *msg; + lnet_process_id_t id; - case PTL_MSG_ACK: - case PTL_MSG_PUT: - case PTL_MSG_GET: - case PTL_MSG_REPLY: - hdr->dest_nid = le64_to_cpu(hdr->dest_nid); - if (hdr->dest_nid != nal->libnal_ni.ni_pid.nid) { - CERROR(LPU64": BAD dest NID in %s message from" - LPU64" to "LPU64" (not me)\n", - nal->libnal_ni.ni_pid.nid, hdr_type_string (hdr), - hdr->src_nid, hdr->dest_nid); - return PTL_FAIL; - } - break; + msg = list_entry(head->next, lnet_msg_t, msg_list); + list_del(&msg->msg_list); - default: - CERROR(LPU64": Bad message type 0x%x from "LPU64"\n", - nal->libnal_ni.ni_pid.nid, hdr->type, hdr->src_nid); - return PTL_FAIL; - } + /* md won't disappear under me, since each msg + * holds a ref on it */ - /* We've decided we're not receiving garbage since we can parse the - * header. We will return PTL_OK come what may... */ + id.nid = msg->msg_hdr.src_nid; + id.pid = msg->msg_hdr.src_pid; - if (!list_empty (&nal->libnal_ni.ni_test_peers) && /* normally we don't */ - fail_peer (nal, hdr->src_nid, 0)) /* shall we now? */ - { - CERROR(LPU64": Dropping incoming %s from "LPU64 - ": simulated failure\n", - nal->libnal_ni.ni_pid.nid, hdr_type_string (hdr), - hdr->src_nid); - lib_drop_message(nal, private, hdr); - return PTL_OK; - } + LASSERT(msg->msg_rx_delayed); + LASSERT(msg->msg_md != NULL); + LASSERT(msg->msg_rxpeer != NULL); + LASSERT(msg->msg_hdr.type == LNET_MSG_PUT); - msg = lib_msg_alloc(nal); - if (msg == NULL) { - CERROR(LPU64": Dropping incoming %s from "LPU64 - ": can't allocate a lib_msg_t\n", - nal->libnal_ni.ni_pid.nid, hdr_type_string (hdr), - hdr->src_nid); - lib_drop_message(nal, private, hdr); - return PTL_OK; - } + CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d " + "match "LPU64" 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); - switch (hdr->type) { - case PTL_MSG_ACK: - rc = parse_ack(nal, hdr, private, msg); - break; - case PTL_MSG_PUT: - rc = parse_put(nal, hdr, private, msg); - break; - case PTL_MSG_GET: - rc = parse_get(nal, hdr, private, msg); - break; - case PTL_MSG_REPLY: - rc = parse_reply(nal, hdr, private, msg); - break; - default: - LASSERT(0); - rc = PTL_FAIL; /* no compiler warning please */ - break; - } - - if (rc != PTL_OK) { - if (msg->md != NULL) { - /* committed... */ - lib_finalize(nal, private, msg, rc); - } else { - LIB_LOCK(nal, flags); - lib_msg_free(nal, msg); /* expects LIB_LOCK held */ - LIB_UNLOCK(nal, flags); - - lib_drop_message(nal, private, hdr); - } - } - - return PTL_OK; - /* That's "OK I can parse it", not "OK I like it" :) */ + lnet_recv_put(msg->msg_rxpeer->lp_ni, msg); + } } +/** + * 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 -lib_api_put(nal_t *apinal, ptl_handle_md_t *mdh, - ptl_ack_req_t ack, ptl_process_id_t *id, - ptl_pt_index_t portal, ptl_ac_index_t ac, - ptl_match_bits_t match_bits, - ptl_size_t offset, ptl_hdr_data_t hdr_data) +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) { - lib_nal_t *nal = apinal->nal_data; - lib_ni_t *ni = &nal->libnal_ni; - lib_msg_t *msg; - ptl_hdr_t hdr; - lib_md_t *md; - unsigned long flags; - int rc; - - if (!list_empty (&ni->ni_test_peers) && /* normally we don't */ - fail_peer (nal, id->nid, 1)) /* shall we now? */ - { - CERROR("Dropping PUT to "LPU64": simulated failure\n", - id->nid); - return PTL_PROCESS_INVALID; + struct lnet_msg *msg; + struct lnet_libmd *md; + int cpt; + 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 = lib_msg_alloc(nal); + msg = lnet_msg_alloc(); if (msg == NULL) { - CERROR(LPU64": Dropping PUT to "LPU64": ENOMEM on lib_msg_t\n", - ni->ni_pid.nid, id->nid); - return PTL_NO_SPACE; + CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n", + libcfs_id2str(target)); + return -ENOMEM; } + msg->msg_vmflush = !!memory_pressure_get(); - LIB_LOCK(nal, flags); + cpt = lnet_cpt_of_cookie(mdh.cookie); + lnet_res_lock(cpt); - md = ptl_handle2md(mdh, nal); - if (md == NULL || md->threshold == 0) { - lib_msg_free(nal, msg); - LIB_UNLOCK(nal, flags); + md = lnet_handle2md(&mdh); + if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) { + 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_res_unlock(cpt); - return PTL_MD_INVALID; - } + lnet_msg_free(msg); + return -ENOENT; + } - CDEBUG(D_NET, "PtlPut -> "LPX64"\n", id->nid); + CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target)); - memset (&hdr, 0, sizeof (hdr)); - hdr.type = cpu_to_le32(PTL_MSG_PUT); - hdr.dest_nid = cpu_to_le64(id->nid); - hdr.dest_pid = cpu_to_le32(id->pid); - hdr.src_nid = cpu_to_le64(ni->ni_pid.nid); - hdr.src_pid = cpu_to_le32(ni->ni_pid.pid); - hdr.payload_length = cpu_to_le32(md->length); + 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 == PTL_ACK_REQ) { - hdr.msg.put.ack_wmd.wh_interface_cookie = ni->ni_interface_cookie; - hdr.msg.put.ack_wmd.wh_object_cookie = md->md_lh.lh_cookie; + 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 { - hdr.msg.put.ack_wmd = PTL_WIRE_HANDLE_NONE; + 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; } - hdr.msg.put.match_bits = cpu_to_le64(match_bits); - hdr.msg.put.ptl_index = cpu_to_le32(portal); - hdr.msg.put.offset = cpu_to_le32(offset); - hdr.msg.put.hdr_data = hdr_data; - - lib_commit_md(nal, md, msg); - - msg->ev.type = PTL_EVENT_SEND_END; - msg->ev.initiator.nid = ni->ni_pid.nid; - msg->ev.initiator.pid = ni->ni_pid.pid; - msg->ev.pt_index = portal; - msg->ev.match_bits = match_bits; - msg->ev.rlength = md->length; - msg->ev.mlength = md->length; - msg->ev.offset = offset; - msg->ev.hdr_data = hdr_data; - - lib_md_deconstruct(nal, md, &msg->ev.md); - ptl_md2handle(&msg->ev.md_handle, nal, md); - - ni->ni_counters.send_count++; - ni->ni_counters.send_length += md->length; - - LIB_UNLOCK(nal, flags); - - rc = lib_send (nal, NULL, msg, &hdr, PTL_MSG_PUT, - id->nid, id->pid, md, 0, md->length); - if (rc != PTL_OK) { - CERROR("Error sending PUT to "LPX64": %d\n", - id->nid, rc); - lib_finalize (nal, NULL, msg, rc); + 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 PTL_OK; + return 0; } +EXPORT_SYMBOL(LNetPut); -lib_msg_t * -lib_create_reply_msg (lib_nal_t *nal, ptl_nid_t peer_nid, lib_msg_t *getmsg) +lnet_msg_t * +lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg) { - /* The NAL can DMA direct to the GET md (i.e. no REPLY msg). This - * returns a msg for the NAL to pass to lib_finalize() when the sink + /* 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 - * lib_finalize() is called on it, so the NAL must call this first */ + * lnet_finalize() is called on it, so the LND must call this first */ - lib_ni_t *ni = &nal->libnal_ni; - lib_msg_t *msg = lib_msg_alloc(nal); - lib_md_t *getmd = getmsg->md; - unsigned long flags; + 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; - LIB_LOCK(nal, flags); + LASSERT(!getmsg->msg_target_is_router); + LASSERT(!getmsg->msg_routing); - LASSERT (getmd->pending > 0); + 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; + } - if (msg == NULL) { - CERROR ("Dropping REPLY from "LPU64": can't allocate msg\n", - peer_nid); - goto drop; - } + cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie); + lnet_res_lock(cpt); - if (getmd->threshold == 0) { - CERROR ("Dropping REPLY from "LPU64" for inactive MD %p\n", - peer_nid, getmd); - goto drop_msg; - } + 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); + lnet_res_unlock(cpt); + goto drop; + } - LASSERT (getmd->offset == 0); + LASSERT(getmd->md_offset == 0); - CDEBUG(D_NET, "Reply from "LPU64" md %p\n", peer_nid, getmd); + CDEBUG(D_NET, "%s: Reply from %s md %p\n", + libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd); - lib_commit_md (nal, getmd, msg); + /* setup information for lnet_build_msg_event */ + 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->ev.type = PTL_EVENT_REPLY_END; - msg->ev.initiator.nid = peer_nid; - msg->ev.initiator.pid = 0; /* XXX FIXME!!! */ - msg->ev.rlength = msg->ev.mlength = getmd->length; - msg->ev.offset = 0; + lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length); + lnet_res_unlock(cpt); - lib_md_deconstruct(nal, getmd, &msg->ev.md); - ptl_md2handle(&msg->ev.md_handle, nal, getmd); + cpt = lnet_cpt_of_nid(peer_id.nid); - ni->ni_counters.recv_count++; - ni->ni_counters.recv_length += getmd->length; + lnet_net_lock(cpt); + lnet_msg_commit(msg, cpt); + lnet_net_unlock(cpt); - LIB_UNLOCK(nal, flags); + lnet_build_msg_event(msg, LNET_EVENT_REPLY); - return msg; + return msg; - drop_msg: - lib_msg_free(nal, msg); drop: - nal->libnal_ni.ni_counters.drop_count++; - nal->libnal_ni.ni_counters.drop_length += getmd->length; + cpt = lnet_cpt_of_nid(peer_id.nid); + + 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); - LIB_UNLOCK (nal, flags); + 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; +} +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 -lib_api_get(nal_t *apinal, ptl_handle_md_t *mdh, ptl_process_id_t *id, - ptl_pt_index_t portal, ptl_ac_index_t ac, - ptl_match_bits_t match_bits, ptl_size_t offset) +LNetGet(lnet_nid_t self, lnet_handle_md_t mdh, + lnet_process_id_t target, unsigned int portal, + __u64 match_bits, unsigned int offset) { - lib_nal_t *nal = apinal->nal_data; - lib_ni_t *ni = &nal->libnal_ni; - lib_msg_t *msg; - ptl_hdr_t hdr; - lib_md_t *md; - unsigned long flags; - int rc; - - if (!list_empty (&ni->ni_test_peers) && /* normally we don't */ - fail_peer (nal, id->nid, 1)) /* shall we now? */ + struct lnet_msg *msg; + struct lnet_libmd *md; + int cpt; + 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 "LPX64": simulated failure\n", - id->nid); - return PTL_PROCESS_INVALID; + CERROR("Dropping GET to %s: simulated failure\n", + libcfs_id2str(target)); + return -EIO; } - msg = lib_msg_alloc(nal); + msg = lnet_msg_alloc(); if (msg == NULL) { - CERROR("Dropping GET to "LPU64": ENOMEM on lib_msg_t\n", - id->nid); - return PTL_NO_SPACE; + CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n", + libcfs_id2str(target)); + return -ENOMEM; } - LIB_LOCK(nal, flags); + cpt = lnet_cpt_of_cookie(mdh.cookie); + lnet_res_lock(cpt); - md = ptl_handle2md(mdh, nal); - if (md == NULL || !md->threshold) { - lib_msg_free(nal, msg); - LIB_UNLOCK(nal, flags); + md = lnet_handle2md(&mdh); + if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) { + 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); - return PTL_MD_INVALID; + lnet_res_unlock(cpt); + + lnet_msg_free(msg); + return -ENOENT; } - CDEBUG(D_NET, "PtlGet -> %Lu: %lu\n", (unsigned long long)id->nid, - (unsigned long)id->pid); + 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); - memset (&hdr, 0, sizeof (hdr)); - hdr.type = cpu_to_le32(PTL_MSG_GET); - hdr.dest_nid = cpu_to_le64(id->nid); - hdr.dest_pid = cpu_to_le32(id->pid); - hdr.src_nid = cpu_to_le64(ni->ni_pid.nid); - hdr.src_pid = cpu_to_le32(ni->ni_pid.pid); - hdr.payload_length = 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) */ - hdr.msg.get.return_wmd.wh_interface_cookie = ni->ni_interface_cookie; - hdr.msg.get.return_wmd.wh_object_cookie = md->md_lh.lh_cookie; - - hdr.msg.get.match_bits = cpu_to_le64(match_bits); - hdr.msg.get.ptl_index = cpu_to_le32(portal); - hdr.msg.get.src_offset = cpu_to_le32(offset); - hdr.msg.get.sink_length = cpu_to_le32(md->length); - - lib_commit_md(nal, md, msg); - - msg->ev.type = PTL_EVENT_SEND_END; - msg->ev.initiator = ni->ni_pid; - msg->ev.pt_index = portal; - msg->ev.match_bits = match_bits; - msg->ev.rlength = md->length; - msg->ev.mlength = md->length; - msg->ev.offset = offset; - msg->ev.hdr_data = 0; - - lib_md_deconstruct(nal, md, &msg->ev.md); - ptl_md2handle(&msg->ev.md_handle, nal, md); - - ni->ni_counters.send_count++; - - LIB_UNLOCK(nal, flags); - - rc = lib_send (nal, NULL, msg, &hdr, PTL_MSG_GET, - id->nid, id->pid, NULL, 0, 0); - if (rc != PTL_OK) { - CERROR(LPU64": error sending GET to "LPU64": %d\n", - ni->ni_pid.nid, id->nid, rc); - lib_finalize (nal, NULL, msg, rc); - } + 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 PTL_OK; + 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 +LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp) +{ + struct list_head *e; + struct lnet_ni *ni; + 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_init); + LASSERT (the_lnet.ln_refcount > 0); + + cpt = lnet_net_lock_current(); + + list_for_each(e, &the_lnet.ln_nis) { + ni = list_entry(e, lnet_ni_t, ni_list); + + 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) { + 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; + + LASSERT(!list_empty(&rnet->lrn_routes)); + + list_for_each_entry(route, &rnet->lrn_routes, + lr_list) { + if (shortest == NULL || + route->lr_hops < shortest->lr_hops) + shortest = route; + } + + LASSERT (shortest != NULL); + hops = shortest->lr_hops; + if (srcnidp != NULL) + *srcnidp = shortest->lr_gateway->lp_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); -void lib_assert_wire_constants (void) +/** + * Set the number of asynchronous messages expected from a target process. + * + * This function is only meaningful for userspace callers. It's a no-op when + * called from kernel. + * + * Asynchronous messages are those that can come from a target when the + * userspace process is not waiting for IO to complete; e.g., AST callbacks + * from Lustre servers. Specifying the expected number of such messages + * allows them to be eagerly received when user process is not running in + * LNet; otherwise network errors may occur. + * + * \param id Process ID of the target process. + * \param nasync Number of asynchronous messages expected from the target. + * + * \return 0 on success, and an error code otherwise. + */ +int +LNetSetAsync(lnet_process_id_t id, int nasync) { - /* Wire protocol assertions generated by 'wirecheck' - * running on Linux mdevi 2.4.21-p4smp-55chaos #1 SMP Tue Jun 8 14:38:44 PDT 2004 i686 i686 i - * with gcc version 3.2.3 20030502 (Red Hat Linux 3.2.3-34) */ - - - /* Constants... */ - LASSERT (PORTALS_PROTO_MAGIC == 0xeebc0ded); - LASSERT (PORTALS_PROTO_VERSION_MAJOR == 1); - LASSERT (PORTALS_PROTO_VERSION_MINOR == 0); - LASSERT (PTL_MSG_ACK == 0); - LASSERT (PTL_MSG_PUT == 1); - LASSERT (PTL_MSG_GET == 2); - LASSERT (PTL_MSG_REPLY == 3); - LASSERT (PTL_MSG_HELLO == 4); - - /* Checks for struct ptl_handle_wire_t */ - LASSERT ((int)sizeof(ptl_handle_wire_t) == 16); - LASSERT ((int)offsetof(ptl_handle_wire_t, wh_interface_cookie) == 0); - LASSERT ((int)sizeof(((ptl_handle_wire_t *)0)->wh_interface_cookie) == 8); - LASSERT ((int)offsetof(ptl_handle_wire_t, wh_object_cookie) == 8); - LASSERT ((int)sizeof(((ptl_handle_wire_t *)0)->wh_object_cookie) == 8); - - /* Checks for struct ptl_magicversion_t */ - LASSERT ((int)sizeof(ptl_magicversion_t) == 8); - LASSERT ((int)offsetof(ptl_magicversion_t, magic) == 0); - LASSERT ((int)sizeof(((ptl_magicversion_t *)0)->magic) == 4); - LASSERT ((int)offsetof(ptl_magicversion_t, version_major) == 4); - LASSERT ((int)sizeof(((ptl_magicversion_t *)0)->version_major) == 2); - LASSERT ((int)offsetof(ptl_magicversion_t, version_minor) == 6); - LASSERT ((int)sizeof(((ptl_magicversion_t *)0)->version_minor) == 2); - - /* Checks for struct ptl_hdr_t */ - LASSERT ((int)sizeof(ptl_hdr_t) == 72); - LASSERT ((int)offsetof(ptl_hdr_t, dest_nid) == 0); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->dest_nid) == 8); - LASSERT ((int)offsetof(ptl_hdr_t, src_nid) == 8); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->src_nid) == 8); - LASSERT ((int)offsetof(ptl_hdr_t, dest_pid) == 16); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->dest_pid) == 4); - LASSERT ((int)offsetof(ptl_hdr_t, src_pid) == 20); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->src_pid) == 4); - LASSERT ((int)offsetof(ptl_hdr_t, type) == 24); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->type) == 4); - LASSERT ((int)offsetof(ptl_hdr_t, payload_length) == 28); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->payload_length) == 4); - LASSERT ((int)offsetof(ptl_hdr_t, msg) == 32); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg) == 40); - - /* Ack */ - LASSERT ((int)offsetof(ptl_hdr_t, msg.ack.dst_wmd) == 32); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.ack.dst_wmd) == 16); - LASSERT ((int)offsetof(ptl_hdr_t, msg.ack.match_bits) == 48); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.ack.match_bits) == 8); - LASSERT ((int)offsetof(ptl_hdr_t, msg.ack.mlength) == 56); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.ack.mlength) == 4); - - /* Put */ - LASSERT ((int)offsetof(ptl_hdr_t, msg.put.ack_wmd) == 32); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.put.ack_wmd) == 16); - LASSERT ((int)offsetof(ptl_hdr_t, msg.put.match_bits) == 48); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.put.match_bits) == 8); - LASSERT ((int)offsetof(ptl_hdr_t, msg.put.hdr_data) == 56); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.put.hdr_data) == 8); - LASSERT ((int)offsetof(ptl_hdr_t, msg.put.ptl_index) == 64); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.put.ptl_index) == 4); - LASSERT ((int)offsetof(ptl_hdr_t, msg.put.offset) == 68); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.put.offset) == 4); - - /* Get */ - LASSERT ((int)offsetof(ptl_hdr_t, msg.get.return_wmd) == 32); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.get.return_wmd) == 16); - LASSERT ((int)offsetof(ptl_hdr_t, msg.get.match_bits) == 48); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.get.match_bits) == 8); - LASSERT ((int)offsetof(ptl_hdr_t, msg.get.ptl_index) == 56); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.get.ptl_index) == 4); - LASSERT ((int)offsetof(ptl_hdr_t, msg.get.src_offset) == 60); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.get.src_offset) == 4); - LASSERT ((int)offsetof(ptl_hdr_t, msg.get.sink_length) == 64); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.get.sink_length) == 4); - - /* Reply */ - LASSERT ((int)offsetof(ptl_hdr_t, msg.reply.dst_wmd) == 32); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.reply.dst_wmd) == 16); - - /* Hello */ - LASSERT ((int)offsetof(ptl_hdr_t, msg.hello.incarnation) == 32); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.hello.incarnation) == 8); - LASSERT ((int)offsetof(ptl_hdr_t, msg.hello.type) == 40); - LASSERT ((int)sizeof(((ptl_hdr_t *)0)->msg.hello.type) == 4); +#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; + int cpt; + + /* 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 */ + cpt = lnet_net_lock_current(); + rnet = lnet_find_net_locked(LNET_NIDNET(id.nid)); + if (rnet != NULL) { + list_for_each(tmp, &rnet->lrn_routes) { + if (nnids == maxnids) { + lnet_net_unlock(cpt); + 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_net_unlock(cpt); + + /* 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 } +EXPORT_SYMBOL(LNetSetAsync);