X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Flnet%2Flib-move.c;h=ce7f926367dfc2fb34643c83d217271ecce12ffd;hp=156a09a8289e9efc6616f13791b745f0bcfde6f4;hb=e8a1124e74f027779068eb7931f10035e0bc6f1d;hpb=14a611ad680e89523abbcab0a3310511ab808ba8 diff --git a/lnet/lnet/lib-move.c b/lnet/lnet/lib-move.c index 156a09a..ce7f926 100644 --- a/lnet/lnet/lib-move.c +++ b/lnet/lnet/lib-move.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2012, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -46,176 +46,17 @@ static int local_nid_dist_zero = 1; CFS_MODULE_PARM(local_nid_dist_zero, "i", int, 0444, "Reserved"); -/* forward ref */ -static void lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg); - -#define LNET_MATCHMD_NONE 0 /* Didn't match */ -#define LNET_MATCHMD_OK 1 /* Matched OK */ -#define LNET_MATCHMD_DROP 2 /* Must be discarded */ - -static int -lnet_try_match_md (int index, int op_mask, lnet_process_id_t src, - unsigned int rlength, unsigned int roffset, - __u64 match_bits, lnet_libmd_t *md, lnet_msg_t *msg, - unsigned int *mlength_out, unsigned int *offset_out) -{ - /* ALWAYS called holding the LNET_LOCK, and can't LNET_UNLOCK; - * lnet_match_blocked_msg() relies on this to avoid races */ - unsigned int offset; - unsigned int mlength; - lnet_me_t *me = md->md_me; - - /* mismatched MD op */ - if ((md->md_options & op_mask) == 0) - return LNET_MATCHMD_NONE; - - /* MD exhausted */ - if (lnet_md_exhausted(md)) - return LNET_MATCHMD_NONE; - - /* mismatched ME nid/pid? */ - if (me->me_match_id.nid != LNET_NID_ANY && - me->me_match_id.nid != src.nid) - return LNET_MATCHMD_NONE; - - if (me->me_match_id.pid != LNET_PID_ANY && - me->me_match_id.pid != src.pid) - return LNET_MATCHMD_NONE; - - /* mismatched ME matchbits? */ - if (((me->me_match_bits ^ match_bits) & ~me->me_ignore_bits) != 0) - return LNET_MATCHMD_NONE; - - /* Hurrah! This _is_ a match; check it out... */ - - if ((md->md_options & LNET_MD_MANAGE_REMOTE) == 0) - offset = md->md_offset; - else - offset = roffset; - - if ((md->md_options & LNET_MD_MAX_SIZE) != 0) { - mlength = md->md_max_size; - LASSERT (md->md_offset + mlength <= md->md_length); - } else { - mlength = md->md_length - offset; - } - - if (rlength <= mlength) { /* fits in allowed space */ - mlength = rlength; - } else if ((md->md_options & LNET_MD_TRUNCATE) == 0) { - /* this packet _really_ is too big */ - CERROR("Matching packet from %s, match "LPU64 - " length %d too big: %d left, %d allowed\n", - libcfs_id2str(src), match_bits, rlength, - md->md_length - offset, mlength); - - return LNET_MATCHMD_DROP; - } - - /* Commit to this ME/MD */ - CDEBUG(D_NET, "Incoming %s index %x from %s of " - "length %d/%d into md "LPX64" [%d] + %d\n", - (op_mask == LNET_MD_OP_PUT) ? "put" : "get", - index, libcfs_id2str(src), mlength, rlength, - md->md_lh.lh_cookie, md->md_niov, offset); - - lnet_commit_md(md, msg); - md->md_offset = offset + mlength; - - /* NB Caller will set ev.type and ev.hdr_data */ - msg->msg_ev.initiator = src; - msg->msg_ev.pt_index = index; - msg->msg_ev.match_bits = match_bits; - msg->msg_ev.rlength = rlength; - msg->msg_ev.mlength = mlength; - msg->msg_ev.offset = offset; - - lnet_md_deconstruct(md, &msg->msg_ev.md); - lnet_md2handle(&msg->msg_ev.md_handle, md); - - *offset_out = offset; - *mlength_out = mlength; - - /* Auto-unlink NOW, so the ME gets unlinked if required. - * We bumped md->md_refcount above so the MD just gets flagged - * for unlink when it is finalized. */ - if ((md->md_flags & LNET_MD_FLAG_AUTO_UNLINK) != 0 && - lnet_md_exhausted(md)) { - lnet_md_unlink(md); - } - - return LNET_MATCHMD_OK; -} - -static int -lnet_match_md(int index, int op_mask, lnet_process_id_t src, - unsigned int rlength, unsigned int roffset, - __u64 match_bits, lnet_msg_t *msg, - unsigned int *mlength_out, unsigned int *offset_out, - lnet_libmd_t **md_out) -{ - lnet_portal_t *ptl = &the_lnet.ln_portals[index]; - lnet_me_t *me; - lnet_me_t *tmp; - lnet_libmd_t *md; - int rc; - - CDEBUG (D_NET, "Request from %s of length %d into portal %d " - "MB="LPX64"\n", libcfs_id2str(src), rlength, index, match_bits); - - if (index < 0 || index >= the_lnet.ln_nportals) { - CERROR("Invalid portal %d not in [0-%d]\n", - index, the_lnet.ln_nportals); - return LNET_MATCHMD_DROP; - } - - cfs_list_for_each_entry_safe_typed (me, tmp, &ptl->ptl_ml, - lnet_me_t, me_list) { - md = me->me_md; - - /* ME attached but MD not attached yet */ - if (md == NULL) - continue; - - LASSERT (me == md->md_me); - - rc = lnet_try_match_md(index, op_mask, src, rlength, - roffset, match_bits, md, msg, - mlength_out, offset_out); - switch (rc) { - default: - LBUG(); - - case LNET_MATCHMD_NONE: - continue; - - case LNET_MATCHMD_OK: - *md_out = md; - return LNET_MATCHMD_OK; - - case LNET_MATCHMD_DROP: - return LNET_MATCHMD_DROP; - } - /* not reached */ - } - - if (op_mask == LNET_MD_OP_GET || - (ptl->ptl_options & LNET_PTL_LAZY) == 0) - return LNET_MATCHMD_DROP; - - return LNET_MATCHMD_NONE; -} - int lnet_fail_nid (lnet_nid_t nid, unsigned int threshold) { - lnet_test_peer_t *tp; - struct list_head *el; - struct list_head *next; - struct list_head cull; + lnet_test_peer_t *tp; + cfs_list_t *el; + cfs_list_t *next; + cfs_list_t 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 */ LIBCFS_ALLOC(tp, sizeof(*tp)); @@ -225,35 +66,35 @@ lnet_fail_nid (lnet_nid_t nid, unsigned int threshold) tp->tp_nid = nid; tp->tp_threshold = threshold; - LNET_LOCK(); - list_add_tail (&tp->tp_list, &the_lnet.ln_test_peers); - LNET_UNLOCK(); - return 0; - } + lnet_net_lock(0); + cfs_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 */ + CFS_INIT_LIST_HEAD(&cull); - LNET_LOCK(); + 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); + cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) { + tp = cfs_list_entry (el, lnet_test_peer_t, tp_list); if (tp->tp_threshold == 0 || /* needs culling anyway */ nid == LNET_NID_ANY || /* removing all entries */ tp->tp_nid == nid) /* matched this one */ { - list_del (&tp->tp_list); - list_add (&tp->tp_list, &cull); + cfs_list_del (&tp->tp_list); + cfs_list_add (&tp->tp_list, &cull); } } - LNET_UNLOCK(); + lnet_net_unlock(0); - while (!list_empty (&cull)) { - tp = list_entry (cull.next, lnet_test_peer_t, tp_list); + while (!cfs_list_empty (&cull)) { + tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list); - list_del (&tp->tp_list); + cfs_list_del (&tp->tp_list); LIBCFS_FREE(tp, sizeof (*tp)); } return 0; @@ -262,18 +103,19 @@ lnet_fail_nid (lnet_nid_t nid, unsigned int threshold) static int fail_peer (lnet_nid_t nid, int outgoing) { - lnet_test_peer_t *tp; - struct list_head *el; - struct list_head *next; - struct list_head cull; + lnet_test_peer_t *tp; + cfs_list_t *el; + cfs_list_t *next; + cfs_list_t cull; int fail = 0; CFS_INIT_LIST_HEAD (&cull); - LNET_LOCK(); + /* 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); + cfs_list_for_each_safe (el, next, &the_lnet.ln_test_peers) { + tp = cfs_list_entry (el, lnet_test_peer_t, tp_list); if (tp->tp_threshold == 0) { /* zombie entry */ @@ -281,8 +123,8 @@ fail_peer (lnet_nid_t nid, int 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); + cfs_list_del (&tp->tp_list); + cfs_list_add (&tp->tp_list, &cull); } continue; } @@ -296,19 +138,19 @@ fail_peer (lnet_nid_t nid, int outgoing) if (outgoing && tp->tp_threshold == 0) { /* see above */ - list_del (&tp->tp_list); - list_add (&tp->tp_list, &cull); + cfs_list_del (&tp->tp_list); + cfs_list_add (&tp->tp_list, &cull); } } break; } } - LNET_UNLOCK (); + lnet_net_unlock(0); - while (!list_empty (&cull)) { - tp = list_entry (cull.next, lnet_test_peer_t, tp_list); - list_del (&tp->tp_list); + while (!cfs_list_empty (&cull)) { + tp = cfs_list_entry (cull.next, lnet_test_peer_t, tp_list); + cfs_list_del (&tp->tp_list); LIBCFS_FREE(tp, sizeof (*tp)); } @@ -321,11 +163,13 @@ lnet_iov_nob (unsigned int niov, struct iovec *iov) { unsigned int nob = 0; + LASSERT(niov == 0 || iov != NULL); while (niov-- > 0) nob += (iov++)->iov_len; return (nob); } +EXPORT_SYMBOL(lnet_iov_nob); void lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset, @@ -384,6 +228,7 @@ lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset, } } while (nob > 0); } +EXPORT_SYMBOL(lnet_copy_iov2iov); int lnet_extract_iov (int dst_niov, struct iovec *dst, @@ -430,6 +275,7 @@ lnet_extract_iov (int dst_niov, struct iovec *dst, offset = 0; } } +EXPORT_SYMBOL(lnet_extract_iov); #ifndef __KERNEL__ unsigned int @@ -478,28 +324,30 @@ lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov) { unsigned int nob = 0; + LASSERT(niov == 0 || kiov != NULL); while (niov-- > 0) nob += (kiov++)->kiov_len; return (nob); } +EXPORT_SYMBOL(lnet_kiov_nob); void lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset, unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset, unsigned int nob) { - /* NB diov, siov are READ-ONLY */ - unsigned int this_nob; - char *daddr = NULL; - char *saddr = NULL; + /* NB diov, siov are READ-ONLY */ + unsigned int this_nob; + char *daddr = NULL; + char *saddr = NULL; - if (nob == 0) - return; + if (nob == 0) + return; - LASSERT (!in_interrupt ()); + LASSERT (!in_interrupt ()); - LASSERT (ndiov > 0); + LASSERT (ndiov > 0); while (doffset >= diov->kiov_len) { doffset -= diov->kiov_len; diov++; @@ -522,64 +370,65 @@ lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset siov->kiov_len - soffset); this_nob = MIN(this_nob, nob); - if (daddr == NULL) - daddr = ((char *)cfs_kmap(diov->kiov_page)) + - diov->kiov_offset + doffset; - if (saddr == NULL) - saddr = ((char *)cfs_kmap(siov->kiov_page)) + - siov->kiov_offset + soffset; - - /* Vanishing risk of kmap deadlock when mapping 2 pages. - * However in practice at least one of the kiovs will be mapped - * kernel pages and the map/unmap will be NOOPs */ - - memcpy (daddr, saddr, this_nob); - nob -= this_nob; - - if (diov->kiov_len > doffset + this_nob) { - daddr += this_nob; - doffset += this_nob; - } else { - cfs_kunmap(diov->kiov_page); - daddr = NULL; - diov++; - ndiov--; - doffset = 0; - } - - if (siov->kiov_len > soffset + this_nob) { - saddr += this_nob; - soffset += this_nob; - } else { - cfs_kunmap(siov->kiov_page); - saddr = NULL; - siov++; - nsiov--; - soffset = 0; - } - } while (nob > 0); - - if (daddr != NULL) - cfs_kunmap(diov->kiov_page); - if (saddr != NULL) - cfs_kunmap(siov->kiov_page); + if (daddr == NULL) + daddr = ((char *)kmap(diov->kiov_page)) + + diov->kiov_offset + doffset; + if (saddr == NULL) + saddr = ((char *)kmap(siov->kiov_page)) + + siov->kiov_offset + soffset; + + /* Vanishing risk of kmap deadlock when mapping 2 pages. + * However in practice at least one of the kiovs will be mapped + * kernel pages and the map/unmap will be NOOPs */ + + memcpy (daddr, saddr, this_nob); + nob -= this_nob; + + if (diov->kiov_len > doffset + this_nob) { + daddr += this_nob; + doffset += this_nob; + } else { + kunmap(diov->kiov_page); + daddr = NULL; + diov++; + ndiov--; + doffset = 0; + } + + if (siov->kiov_len > soffset + this_nob) { + saddr += this_nob; + soffset += this_nob; + } else { + kunmap(siov->kiov_page); + saddr = NULL; + siov++; + nsiov--; + soffset = 0; + } + } while (nob > 0); + + if (daddr != NULL) + kunmap(diov->kiov_page); + if (saddr != NULL) + kunmap(siov->kiov_page); } +EXPORT_SYMBOL(lnet_copy_kiov2kiov); void lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset, unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset, unsigned int nob) { - /* NB iov, kiov are READ-ONLY */ - unsigned int this_nob; - char *addr = NULL; + /* NB iov, kiov are READ-ONLY */ + unsigned int this_nob; + char *addr = NULL; - if (nob == 0) - return; + if (nob == 0) + return; - LASSERT (!in_interrupt ()); + LASSERT (!in_interrupt ()); - LASSERT (niov > 0); + LASSERT (niov > 0); while (iovoffset >= iov->iov_len) { iovoffset -= iov->iov_len; iov++; @@ -603,7 +452,7 @@ lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset this_nob = MIN(this_nob, nob); if (addr == NULL) - addr = ((char *)cfs_kmap(kiov->kiov_page)) + + addr = ((char *)kmap(kiov->kiov_page)) + kiov->kiov_offset + kiovoffset; memcpy ((char *)iov->iov_base + iovoffset, addr, this_nob); @@ -621,34 +470,35 @@ lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset addr += this_nob; kiovoffset += this_nob; } else { - cfs_kunmap(kiov->kiov_page); - addr = NULL; - kiov++; - nkiov--; - kiovoffset = 0; - } + kunmap(kiov->kiov_page); + addr = NULL; + kiov++; + nkiov--; + kiovoffset = 0; + } - } while (nob > 0); + } while (nob > 0); - if (addr != NULL) - cfs_kunmap(kiov->kiov_page); + if (addr != NULL) + kunmap(kiov->kiov_page); } +EXPORT_SYMBOL(lnet_copy_kiov2iov); void lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset, unsigned int niov, struct iovec *iov, unsigned int iovoffset, unsigned int nob) { - /* NB kiov, iov are READ-ONLY */ - unsigned int this_nob; - char *addr = NULL; + /* NB kiov, iov are READ-ONLY */ + unsigned int this_nob; + char *addr = NULL; - if (nob == 0) - return; + if (nob == 0) + return; - LASSERT (!in_interrupt ()); + LASSERT (!in_interrupt ()); - LASSERT (nkiov > 0); + LASSERT (nkiov > 0); while (kiovoffset >= kiov->kiov_len) { kiovoffset -= kiov->kiov_len; kiov++; @@ -672,35 +522,36 @@ lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffs this_nob = MIN(this_nob, nob); if (addr == NULL) - addr = ((char *)cfs_kmap(kiov->kiov_page)) + - kiov->kiov_offset + kiovoffset; - - memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob); - nob -= this_nob; - - if (kiov->kiov_len > kiovoffset + this_nob) { - addr += this_nob; - kiovoffset += this_nob; - } else { - cfs_kunmap(kiov->kiov_page); - addr = NULL; - kiov++; - nkiov--; - kiovoffset = 0; - } - - if (iov->iov_len > iovoffset + this_nob) { - iovoffset += this_nob; - } else { - iov++; - niov--; - iovoffset = 0; - } - } while (nob > 0); - - if (addr != NULL) - cfs_kunmap(kiov->kiov_page); + addr = ((char *)kmap(kiov->kiov_page)) + + kiov->kiov_offset + kiovoffset; + + memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob); + nob -= this_nob; + + if (kiov->kiov_len > kiovoffset + this_nob) { + addr += this_nob; + kiovoffset += this_nob; + } else { + kunmap(kiov->kiov_page); + addr = NULL; + kiov++; + nkiov--; + kiovoffset = 0; + } + + if (iov->iov_len > iovoffset + this_nob) { + iovoffset += this_nob; + } else { + iov++; + niov--; + iovoffset = 0; + } + } while (nob > 0); + + if (addr != NULL) + kunmap(kiov->kiov_page); } +EXPORT_SYMBOL(lnet_copy_iov2kiov); int lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst, @@ -733,14 +584,14 @@ lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst, dst->kiov_page = src->kiov_page; dst->kiov_offset = src->kiov_offset + offset; - if (len <= frag_len) { - dst->kiov_len = len; - LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE); - return (niov); - } + 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 <= CFS_PAGE_SIZE); + dst->kiov_len = frag_len; + LASSERT (dst->kiov_offset + dst->kiov_len <= PAGE_CACHE_SIZE); len -= frag_len; dst++; @@ -750,28 +601,29 @@ lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst, offset = 0; } } +EXPORT_SYMBOL(lnet_extract_kiov); #endif void lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed, unsigned int offset, unsigned int mlen, unsigned int rlen) { - unsigned int niov = 0; - struct iovec *iov = NULL; - lnet_kiov_t *kiov = NULL; - int rc; + unsigned int niov = 0; + struct iovec *iov = NULL; + lnet_kiov_t *kiov = NULL; + int rc; - LASSERT (!in_interrupt ()); - LASSERT (mlen == 0 || msg != NULL); + 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_wanted = mlen; - msg->msg_offset = offset; msg->msg_receiving = 0; if (mlen != 0) { @@ -790,34 +642,6 @@ lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed, lnet_finalize(ni, msg, rc); } -int -lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2) -{ - lnet_peer_t *p1 = r1->lr_gateway; - lnet_peer_t *p2 = r2->lr_gateway; - - 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; - - return 0; -} - - void lnet_setpayloadbuffer(lnet_msg_t *msg) { @@ -861,85 +685,69 @@ lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target, void lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg) { - void *priv = msg->msg_private; - int rc; + void *priv = msg->msg_private; + int rc; - LASSERT (!in_interrupt ()); - LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND || - (msg->msg_txcredit && msg->msg_peertxcredit)); + LASSERT (!in_interrupt ()); + LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND || + (msg->msg_txcredit && msg->msg_peertxcredit)); - rc = (ni->ni_lnd->lnd_send)(ni, priv, msg); - if (rc < 0) - lnet_finalize(ni, msg, rc); + rc = (ni->ni_lnd->lnd_send)(ni, priv, msg); + if (rc < 0) + lnet_finalize(ni, msg, rc); } int -lnet_eager_recv_locked(lnet_msg_t *msg) +lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg) { - lnet_peer_t *peer; - lnet_ni_t *ni; - int rc = 0; - - LASSERT (!msg->msg_delayed); - msg->msg_delayed = 1; - - LASSERT (msg->msg_receiving); - LASSERT (!msg->msg_sending); - - peer = msg->msg_rxpeer; - ni = peer->lp_ni; - - if (ni->ni_lnd->lnd_eager_recv != NULL) { - LNET_UNLOCK(); - - rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg, - &msg->msg_private); - if (rc != 0) { - CERROR("recv from %s / send to %s aborted: " - "eager_recv failed %d\n", - libcfs_nid2str(peer->lp_nid), - libcfs_id2str(msg->msg_target), rc); - LASSERT (rc < 0); /* required by my callers */ - } - - LNET_LOCK(); - } - - return rc; + int rc; + + LASSERT(!msg->msg_sending); + LASSERT(msg->msg_receiving); + LASSERT(!msg->msg_rx_ready_delay); + LASSERT(ni->ni_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_LOCK */ +/* NB: caller shall hold a ref on 'lp' as I'd drop lnet_net_lock */ void -lnet_ni_peer_alive(lnet_peer_t *lp) +lnet_ni_query_locked(lnet_ni_t *ni, lnet_peer_t *lp) { - cfs_time_t last_alive = 0; - lnet_ni_t *ni = lp->lp_ni; + cfs_time_t last_alive = 0; - LASSERT (ni != NULL); - LASSERT (ni->ni_peertimeout > 0); - LASSERT (ni->ni_lnd->lnd_query != NULL); + LASSERT(lnet_peer_aliveness_enabled(lp)); + LASSERT(ni->ni_lnd->lnd_query != NULL); - LNET_UNLOCK(); - (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive); - LNET_LOCK(); + lnet_net_unlock(lp->lp_cpt); + (ni->ni_lnd->lnd_query)(ni, lp->lp_nid, &last_alive); + lnet_net_lock(lp->lp_cpt); - lp->lp_last_query = cfs_time_current(); + lp->lp_last_query = cfs_time_current(); - if (last_alive != 0) /* NI has updated timestamp */ - lp->lp_last_alive = last_alive; - return; + if (last_alive != 0) /* NI has updated timestamp */ + lp->lp_last_alive = last_alive; } -/* NB: always called with LNET_LOCK held */ +/* NB: always called with lnet_net_lock held */ static inline int lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now) { - lnet_ni_t *ni = lp->lp_ni; - cfs_time_t deadline; - int alive; + int alive; + cfs_time_t deadline; - LASSERT (ni != NULL); - LASSERT (ni->ni_peertimeout > 0); + 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()). @@ -949,12 +757,15 @@ lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now) return 0; deadline = cfs_time_add(lp->lp_last_alive, - cfs_time_seconds(ni->ni_peertimeout)); + cfs_time_seconds(lp->lp_ni->ni_peertimeout)); alive = cfs_time_after(deadline, now); - /* Update obsolete lp_alive */ - if (alive && !lp->lp_alive && lp->lp_timestamp != 0 && - cfs_time_before(lp->lp_timestamp, lp->lp_last_alive)) + /* 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; @@ -962,16 +773,13 @@ lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now) /* NB: returns 1 when alive, 0 when dead, negative when error; - * may drop the LNET_LOCK */ + * may drop the lnet_net_lock */ int lnet_peer_alive_locked (lnet_peer_t *lp) { - lnet_ni_t *ni = lp->lp_ni; - cfs_time_t now = cfs_time_current(); + cfs_time_t now = cfs_time_current(); - LASSERT (ni != NULL); - - if (ni->ni_peertimeout <= 0) /* disabled */ + if (!lnet_peer_aliveness_enabled(lp)) return -ENODEV; if (lnet_peer_is_alive(lp, now)) @@ -992,13 +800,14 @@ lnet_peer_alive_locked (lnet_peer_t *lp) "%d < %d (%d/%d)\n", libcfs_nid2str(lp->lp_nid), (int)now, (int)next_query, - lnet_queryinterval, ni->ni_peertimeout); + lnet_queryinterval, + lp->lp_ni->ni_peertimeout); return 0; } } /* query NI for latest aliveness news */ - lnet_ni_peer_alive(lp); + lnet_ni_query_locked(lp->lp_ni, lp); if (lnet_peer_is_alive(lp, now)) return 1; @@ -1008,35 +817,44 @@ lnet_peer_alive_locked (lnet_peer_t *lp) } int -lnet_post_send_locked (lnet_msg_t *msg, int do_send) +lnet_post_send_locked(lnet_msg_t *msg, int do_send) { - /* lnet_send is going to LNET_UNLOCK immediately after this, so it sets - * do_send FALSE and I don't do the unlock/send/lock bit. I return - * EAGAIN if msg blocked, EHOSTUNREACH if msg_txpeer appears dead, and - * 0 if sent or OK to send */ - lnet_peer_t *lp = msg->msg_txpeer; - lnet_ni_t *ni = lp->lp_ni; - - /* non-lnet_send() callers have checked before */ - LASSERT (!do_send || msg->msg_delayed); - LASSERT (!msg->msg_receiving); - - /* NB 'lp' is always the next hop */ - if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 && - lnet_peer_alive_locked(lp) == 0) { - LNET_UNLOCK(); - - CDEBUG(D_NETERROR, "Dropping message for %s: peer not alive\n", - libcfs_id2str(msg->msg_target)); - if (do_send) - lnet_finalize(ni, msg, -EHOSTUNREACH); - - LNET_LOCK(); - return EHOSTUNREACH; - } + /* 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. + * I return EAGAIN if msg blocked, EHOSTUNREACH if msg_txpeer + * appears dead, and 0 if sent or OK to send */ + struct lnet_peer *lp = msg->msg_txpeer; + struct lnet_ni *ni = lp->lp_ni; + struct lnet_tx_queue *tq; + int cpt; + + /* non-lnet_send() callers have checked before */ + LASSERT(!do_send || msg->msg_tx_delayed); + LASSERT(!msg->msg_receiving); + LASSERT(msg->msg_tx_committed); + + cpt = msg->msg_tx_cpt; + tq = ni->ni_tx_queues[cpt]; + + /* 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_peertxcredit) { - LASSERT ((lp->lp_txcredits < 0) == !list_empty(&lp->lp_txq)); + LASSERT ((lp->lp_txcredits < 0) == + !cfs_list_empty(&lp->lp_txq)); msg->msg_peertxcredit = 1; lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t); @@ -1046,77 +864,65 @@ lnet_post_send_locked (lnet_msg_t *msg, int do_send) lp->lp_mintxcredits = lp->lp_txcredits; if (lp->lp_txcredits < 0) { - msg->msg_delayed = 1; - list_add_tail (&msg->msg_list, &lp->lp_txq); + msg->msg_tx_delayed = 1; + cfs_list_add_tail(&msg->msg_list, &lp->lp_txq); return EAGAIN; } } if (!msg->msg_txcredit) { - LASSERT ((ni->ni_txcredits < 0) == !list_empty(&ni->ni_txq)); - - msg->msg_txcredit = 1; - ni->ni_txcredits--; - - if (ni->ni_txcredits < ni->ni_mintxcredits) - ni->ni_mintxcredits = ni->ni_txcredits; - - if (ni->ni_txcredits < 0) { - msg->msg_delayed = 1; - list_add_tail (&msg->msg_list, &ni->ni_txq); - return EAGAIN; - } - } - - if (do_send) { - LNET_UNLOCK(); - lnet_ni_send(ni, msg); - LNET_LOCK(); - } - return 0; + LASSERT((tq->tq_credits < 0) == + !cfs_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; + cfs_list_add_tail(&msg->msg_list, &tq->tq_delayed); + return EAGAIN; + } + } + + if (do_send) { + lnet_net_unlock(cpt); + lnet_ni_send(ni, msg); + lnet_net_lock(cpt); + } + return 0; } #ifdef __KERNEL__ -static void -lnet_commit_routedmsg (lnet_msg_t *msg) -{ - /* ALWAYS called holding the LNET_LOCK */ - LASSERT (msg->msg_routing); - - the_lnet.ln_counters.msgs_alloc++; - if (the_lnet.ln_counters.msgs_alloc > - the_lnet.ln_counters.msgs_max) - the_lnet.ln_counters.msgs_max = - the_lnet.ln_counters.msgs_alloc; - - the_lnet.ln_counters.route_count++; - the_lnet.ln_counters.route_length += msg->msg_len; - - LASSERT (!msg->msg_onactivelist); - msg->msg_onactivelist = 1; - list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs); -} lnet_rtrbufpool_t * lnet_msg2bufpool(lnet_msg_t *msg) { - lnet_rtrbufpool_t *rbp = &the_lnet.ln_rtrpools[0]; + lnet_rtrbufpool_t *rbp; + int cpt; - LASSERT (msg->msg_len <= LNET_MTU); - while (msg->msg_len > (unsigned int)rbp->rbp_npages * CFS_PAGE_SIZE) { - rbp++; - LASSERT (rbp < &the_lnet.ln_rtrpools[LNET_NRBPOOLS]); - } + 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; + return rbp; } int lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv) { - /* lnet_parse is going to LNET_UNLOCK immediately after this, so it - * sets do_recv FALSE and I don't do the unlock/send/lock bit. I - * return EAGAIN if msg blocked and 0 if received or OK to receive */ + /* 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 EAGAIN if msg blocked and 0 if received or OK to receive */ lnet_peer_t *lp = msg->msg_rxpeer; lnet_rtrbufpool_t *rbp; lnet_rtrbuf_t *rb; @@ -1128,11 +934,12 @@ lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv) LASSERT (msg->msg_receiving); LASSERT (!msg->msg_sending); - /* non-lnet_parse callers only send delayed messages */ - LASSERT (!do_recv || msg->msg_delayed); + /* 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)); + LASSERT ((lp->lp_rtrcredits < 0) == + !cfs_list_empty(&lp->lp_rtrq)); msg->msg_peerrtrcredit = 1; lp->lp_rtrcredits--; @@ -1141,8 +948,9 @@ lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv) if (lp->lp_rtrcredits < 0) { /* must have checked eager_recv before here */ - LASSERT (msg->msg_delayed); - list_add_tail(&msg->msg_list, &lp->lp_rtrq); + LASSERT(msg->msg_rx_ready_delay); + msg->msg_rx_delayed = 1; + cfs_list_add_tail(&msg->msg_list, &lp->lp_rtrq); return EAGAIN; } } @@ -1150,7 +958,8 @@ lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv) rbp = lnet_msg2bufpool(msg); if (!msg->msg_rtrcredit) { - LASSERT ((rbp->rbp_credits < 0) == !list_empty(&rbp->rbp_msgs)); + LASSERT ((rbp->rbp_credits < 0) == + !cfs_list_empty(&rbp->rbp_msgs)); msg->msg_rtrcredit = 1; rbp->rbp_credits--; @@ -1159,51 +968,56 @@ lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv) if (rbp->rbp_credits < 0) { /* must have checked eager_recv before here */ - LASSERT (msg->msg_delayed); - list_add_tail(&msg->msg_list, &rbp->rbp_msgs); + LASSERT(msg->msg_rx_ready_delay); + msg->msg_rx_delayed = 1; + cfs_list_add_tail(&msg->msg_list, &rbp->rbp_msgs); return EAGAIN; } } - LASSERT (!list_empty(&rbp->rbp_bufs)); - rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list); - list_del(&rb->rb_list); + LASSERT (!cfs_list_empty(&rbp->rbp_bufs)); + rb = cfs_list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list); + cfs_list_del(&rb->rb_list); msg->msg_niov = rbp->rbp_npages; msg->msg_kiov = &rb->rb_kiov[0]; if (do_recv) { - LNET_UNLOCK(); - lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1, - 0, msg->msg_len, msg->msg_len); - LNET_LOCK(); - } - return 0; + 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 0; } #endif void -lnet_return_credits_locked (lnet_msg_t *msg) +lnet_return_tx_credits_locked(lnet_msg_t *msg) { - lnet_peer_t *txpeer = msg->msg_txpeer; - lnet_peer_t *rxpeer = msg->msg_rxpeer; - lnet_msg_t *msg2; - lnet_ni_t *ni; + lnet_peer_t *txpeer = msg->msg_txpeer; + lnet_msg_t *msg2; - if (msg->msg_txcredit) { - /* give back NI txcredits */ - msg->msg_txcredit = 0; - ni = txpeer->lp_ni; + if (msg->msg_txcredit) { + struct lnet_ni *ni = txpeer->lp_ni; + struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt]; - LASSERT((ni->ni_txcredits < 0) == !list_empty(&ni->ni_txq)); + /* give back NI txcredits */ + msg->msg_txcredit = 0; - ni->ni_txcredits++; - if (ni->ni_txcredits <= 0) { - msg2 = list_entry(ni->ni_txq.next, lnet_msg_t, msg_list); - list_del(&msg2->msg_list); + LASSERT((tq->tq_credits < 0) == + !cfs_list_empty(&tq->tq_delayed)); - LASSERT(msg2->msg_txpeer->lp_ni == ni); - LASSERT(msg2->msg_delayed); + tq->tq_credits++; + if (tq->tq_credits <= 0) { + msg2 = cfs_list_entry(tq->tq_delayed.next, + lnet_msg_t, msg_list); + cfs_list_del(&msg2->msg_list); + + LASSERT(msg2->msg_txpeer->lp_ni == ni); + LASSERT(msg2->msg_tx_delayed); (void) lnet_post_send_locked(msg2, 1); } @@ -1213,19 +1027,20 @@ lnet_return_credits_locked (lnet_msg_t *msg) /* give back peer txcredits */ msg->msg_peertxcredit = 0; - LASSERT((txpeer->lp_txcredits < 0) == !list_empty(&txpeer->lp_txq)); + LASSERT((txpeer->lp_txcredits < 0) == + !cfs_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); + msg2 = cfs_list_entry(txpeer->lp_txq.next, + lnet_msg_t, msg_list); + cfs_list_del(&msg2->msg_list); - LASSERT (msg2->msg_txpeer == txpeer); - LASSERT (msg2->msg_delayed); + LASSERT(msg2->msg_txpeer == txpeer); + LASSERT(msg2->msg_tx_delayed); (void) lnet_post_send_locked(msg2, 1); } @@ -1235,8 +1050,15 @@ lnet_return_credits_locked (lnet_msg_t *msg) msg->msg_txpeer = NULL; lnet_peer_decref_locked(txpeer); } +} +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; @@ -1247,22 +1069,24 @@ lnet_return_credits_locked (lnet_msg_t *msg) * itself */ LASSERT (msg->msg_kiov != NULL); - rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]); + rb = cfs_list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]); rbp = rb->rb_pool; LASSERT (rbp == lnet_msg2bufpool(msg)); msg->msg_kiov = NULL; msg->msg_rtrcredit = 0; - LASSERT((rbp->rbp_credits < 0) == !list_empty(&rbp->rbp_msgs)); - LASSERT((rbp->rbp_credits > 0) == !list_empty(&rbp->rbp_bufs)); + LASSERT((rbp->rbp_credits < 0) == + !cfs_list_empty(&rbp->rbp_msgs)); + LASSERT((rbp->rbp_credits > 0) == + !cfs_list_empty(&rbp->rbp_bufs)); - list_add(&rb->rb_list, &rbp->rbp_bufs); + cfs_list_add(&rb->rb_list, &rbp->rbp_bufs); rbp->rbp_credits++; if (rbp->rbp_credits <= 0) { - msg2 = list_entry(rbp->rbp_msgs.next, - lnet_msg_t, msg_list); - list_del(&msg2->msg_list); + msg2 = cfs_list_entry(rbp->rbp_msgs.next, + lnet_msg_t, msg_list); + cfs_list_del(&msg2->msg_list); (void) lnet_post_routed_recv_locked(msg2, 1); } @@ -1272,13 +1096,14 @@ lnet_return_credits_locked (lnet_msg_t *msg) /* give back peer router credits */ msg->msg_peerrtrcredit = 0; - LASSERT((rxpeer->lp_rtrcredits < 0) == !list_empty(&rxpeer->lp_rtrq)); + LASSERT((rxpeer->lp_rtrcredits < 0) == + !cfs_list_empty(&rxpeer->lp_rtrq)); rxpeer->lp_rtrcredits++; if (rxpeer->lp_rtrcredits <= 0) { - msg2 = list_entry(rxpeer->lp_rtrq.next, - lnet_msg_t, msg_list); - list_del(&msg2->msg_list); + msg2 = cfs_list_entry(rxpeer->lp_rtrq.next, + lnet_msg_t, msg_list); + cfs_list_del(&msg2->msg_list); (void) lnet_post_routed_recv_locked(msg2, 1); } @@ -1293,20 +1118,115 @@ lnet_return_credits_locked (lnet_msg_t *msg) } } -int -lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg) +static int +lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2) { - lnet_nid_t dst_nid = msg->msg_target.nid; - lnet_ni_t *src_ni; - lnet_ni_t *local_ni; - lnet_remotenet_t *rnet; - lnet_route_t *route; - lnet_route_t *best_route; - struct list_head *tmp; - lnet_peer_t *lp; - lnet_peer_t *lp2; - int rc; + 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; + cfs_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); @@ -1314,72 +1234,80 @@ lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg) msg->msg_sending = 1; - /* NB! ni != NULL == interface pre-determined (ACK/REPLY) */ - - LNET_LOCK(); - - if (the_lnet.ln_shutdown) { - LNET_UNLOCK(); - return -ESHUTDOWN; - } - - if (src_nid == LNET_NID_ANY) { - src_ni = NULL; - } else { - src_ni = lnet_nid2ni_locked(src_nid); - if (src_ni == NULL) { - LNET_UNLOCK(); - CERROR("Can't send to %s: src %s is not a local nid\n", - libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid)); + 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); } /* Is this for someone on a local network? */ - local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid)); + 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); - } else { - lnet_ni_decref_locked(local_ni); - lnet_ni_decref_locked(src_ni); - LNET_UNLOCK(); - CERROR("No route to %s via from %s\n", - libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid)); - return -EINVAL; - } - - LASSERT (src_nid != LNET_NID_ANY); - - if (!msg->msg_routing) - msg->msg_hdr.src_nid = cpu_to_le64(src_nid); - - if (src_ni == the_lnet.ln_loni) { - /* No send credit hassles with LOLND */ - LNET_UNLOCK(); - lnet_ni_send(src_ni, msg); - lnet_ni_decref(src_ni); - return 0; - } - - rc = lnet_nid2peer_locked(&lp, dst_nid); - lnet_ni_decref_locked(src_ni); /* lp has ref on src_ni; lose mine */ - if (rc != 0) { - LNET_UNLOCK(); - CERROR("Error %d finding peer %s\n", rc, - libcfs_nid2str(dst_nid)); + 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_UNLOCK(); + lnet_net_unlock(cpt); /* NB * - once application finishes computation, check here to update @@ -1389,62 +1317,55 @@ lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg) if (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) lnet_router_checker(); - LNET_LOCK(); + lnet_net_lock(cpt); #endif - /* sending to a remote network */ - rnet = lnet_find_net_locked(LNET_NIDNET(dst_nid)); - if (rnet == NULL) { - if (src_ni != NULL) - lnet_ni_decref_locked(src_ni); - LNET_UNLOCK(); - CERROR("No route to %s\n", libcfs_id2str(msg->msg_target)); + /* 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; } - /* Find the best gateway I can use */ - lp = NULL; - best_route = NULL; - list_for_each(tmp, &rnet->lrn_routes) { - route = list_entry(tmp, lnet_route_t, lr_list); - lp2 = route->lr_gateway; - - if (lp2->lp_alive && - lnet_router_down_ni(lp2, rnet->lrn_net) <= 0 && - (src_ni == NULL || lp2->lp_ni == src_ni) && - (lp == NULL || - lnet_compare_routes(route, best_route) > 0)) { - best_route = route; - lp = lp2; - } - } - - if (lp == NULL) { - if (src_ni != NULL) - lnet_ni_decref_locked(src_ni); - LNET_UNLOCK(); - - CERROR("No route to %s via %s (all routers down)\n", - libcfs_id2str(msg->msg_target), - libcfs_nid2str(src_nid)); - return -EHOSTUNREACH; - } - - /* Place selected route at the end of the route list to ensure - * fairness; everything else being equal... */ - list_del(&best_route->lr_list); - list_add_tail(&best_route->lr_list, &rnet->lrn_routes); + /* 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); - } + lnet_ni_decref_locked(src_ni, cpt); + } - lnet_peer_addref_locked(lp); + lnet_peer_addref_locked(lp); - LASSERT (src_nid != LNET_NID_ANY); + 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 */ @@ -1465,7 +1386,7 @@ lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg) msg->msg_txpeer = lp; /* msg takes my ref on lp */ rc = lnet_post_send_locked(msg, 0); - LNET_UNLOCK(); + lnet_net_unlock(cpt); if (rc == EHOSTUNREACH) return -EHOSTUNREACH; @@ -1477,406 +1398,125 @@ lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg) } static void -lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg) +lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob) { - /* ALWAYS called holding the LNET_LOCK */ - /* Here, we commit the MD to a network OP by marking it busy and - * decrementing its threshold. Come what may, the network "owns" - * the MD until a call to lnet_finalize() signals completion. */ - LASSERT (!msg->msg_routing); + lnet_net_lock(cpt); + the_lnet.ln_counters[cpt]->drop_count++; + the_lnet.ln_counters[cpt]->drop_length += nob; + lnet_net_unlock(cpt); - msg->msg_md = md; - - md->md_refcount++; - if (md->md_threshold != LNET_MD_THRESH_INF) { - LASSERT (md->md_threshold > 0); - md->md_threshold--; - } - - the_lnet.ln_counters.msgs_alloc++; - if (the_lnet.ln_counters.msgs_alloc > - the_lnet.ln_counters.msgs_max) - the_lnet.ln_counters.msgs_max = - the_lnet.ln_counters.msgs_alloc; - - LASSERT (!msg->msg_onactivelist); - msg->msg_onactivelist = 1; - list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs); -} - -static void -lnet_drop_message (lnet_ni_t *ni, void *private, unsigned int nob) -{ - LNET_LOCK(); - the_lnet.ln_counters.drop_count++; - the_lnet.ln_counters.drop_length += nob; - LNET_UNLOCK(); - - lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob); -} - -static void -lnet_drop_delayed_put(lnet_msg_t *msg, char *reason) -{ - lnet_process_id_t id = {0}; - - id.nid = msg->msg_hdr.src_nid; - id.pid = msg->msg_hdr.src_pid; - - LASSERT (msg->msg_md == NULL); - LASSERT (msg->msg_delayed); - LASSERT (msg->msg_rxpeer != NULL); - LASSERT (msg->msg_hdr.type == LNET_MSG_PUT); - - CWARN("Dropping delayed PUT from %s portal %d match "LPU64 - " offset %d length %d: %s\n", - libcfs_id2str(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_private, msg->msg_len); - - LNET_LOCK(); - - lnet_peer_decref_locked(msg->msg_rxpeer); - msg->msg_rxpeer = NULL; - - lnet_msg_free(msg); - - LNET_UNLOCK(); -} - -int -LNetSetLazyPortal(int portal) -{ - lnet_portal_t *ptl = &the_lnet.ln_portals[portal]; - - if (portal < 0 || portal >= the_lnet.ln_nportals) - return -EINVAL; - - CDEBUG(D_NET, "Setting portal %d lazy\n", portal); - - LNET_LOCK(); - - ptl->ptl_options |= LNET_PTL_LAZY; - - LNET_UNLOCK(); - - return 0; -} - -int -LNetClearLazyPortal(int portal) -{ - struct list_head zombies; - lnet_portal_t *ptl = &the_lnet.ln_portals[portal]; - lnet_msg_t *msg; - - if (portal < 0 || portal >= the_lnet.ln_nportals) - return -EINVAL; - - LNET_LOCK(); - - if ((ptl->ptl_options & LNET_PTL_LAZY) == 0) { - LNET_UNLOCK(); - return 0; - } - - if (the_lnet.ln_shutdown) - CWARN ("Active lazy portal %d on exit\n", portal); - else - CDEBUG (D_NET, "clearing portal %d lazy\n", portal); - - /* grab all the blocked messages atomically */ - list_add(&zombies, &ptl->ptl_msgq); - list_del_init(&ptl->ptl_msgq); - - ptl->ptl_msgq_version++; - ptl->ptl_options &= ~LNET_PTL_LAZY; - - LNET_UNLOCK(); - - while (!list_empty(&zombies)) { - msg = list_entry(zombies.next, lnet_msg_t, msg_list); - list_del(&msg->msg_list); - - lnet_drop_delayed_put(msg, "Clearing lazy portal attr"); - } - - return 0; + lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob); } static void -lnet_recv_put(lnet_libmd_t *md, lnet_msg_t *msg, int delayed, - unsigned int offset, unsigned int mlength) -{ - lnet_hdr_t *hdr = &msg->msg_hdr; - - LNET_LOCK(); - - the_lnet.ln_counters.recv_count++; - the_lnet.ln_counters.recv_length += mlength; - - LNET_UNLOCK(); - - if (mlength != 0) - lnet_setpayloadbuffer(msg); - - msg->msg_ev.type = LNET_EVENT_PUT; - msg->msg_ev.target.pid = hdr->dest_pid; - msg->msg_ev.target.nid = hdr->dest_nid; - msg->msg_ev.hdr_data = hdr->msg.put.hdr_data; - - /* Must I ACK? If so I'll grab the ack_wmd out of the header and put - * it back into the ACK during lnet_finalize() */ - msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) && - (md->md_options & LNET_MD_ACK_DISABLE) == 0); - - lnet_ni_recv(msg->msg_rxpeer->lp_ni, - msg->msg_private, - msg, delayed, offset, mlength, - hdr->payload_length); -} - -/* called with LNET_LOCK held */ -void -lnet_match_blocked_msg(lnet_libmd_t *md) +lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg) { - CFS_LIST_HEAD (drops); - CFS_LIST_HEAD (matches); - struct list_head *tmp; - struct list_head *entry; - lnet_msg_t *msg; - lnet_me_t *me = md->md_me; - lnet_portal_t *ptl = &the_lnet.ln_portals[me->me_portal]; - - LASSERT (me->me_portal < (unsigned int)the_lnet.ln_nportals); - - if ((ptl->ptl_options & LNET_PTL_LAZY) == 0) { - LASSERT (list_empty(&ptl->ptl_msgq)); - return; - } - - LASSERT (md->md_refcount == 0); /* a brand new MD */ - - list_for_each_safe (entry, tmp, &ptl->ptl_msgq) { - int rc; - int index; - unsigned int mlength; - unsigned int offset; - lnet_hdr_t *hdr; - lnet_process_id_t src; - - msg = list_entry(entry, lnet_msg_t, msg_list); - - LASSERT (msg->msg_delayed); - - hdr = &msg->msg_hdr; - index = hdr->msg.put.ptl_index; - - src.nid = hdr->src_nid; - src.pid = hdr->src_pid; - - rc = lnet_try_match_md(index, LNET_MD_OP_PUT, src, - hdr->payload_length, - hdr->msg.put.offset, - hdr->msg.put.match_bits, - md, msg, &mlength, &offset); - - if (rc == LNET_MATCHMD_NONE) - continue; - - /* Hurrah! This _is_ a match */ - list_del(&msg->msg_list); - ptl->ptl_msgq_version++; - - if (rc == LNET_MATCHMD_OK) { - list_add_tail(&msg->msg_list, &matches); - - CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d " - "match "LPU64" offset %d length %d.\n", - libcfs_id2str(src), - hdr->msg.put.ptl_index, - hdr->msg.put.match_bits, - hdr->msg.put.offset, - hdr->payload_length); - } else { - LASSERT (rc == LNET_MATCHMD_DROP); - - list_add_tail(&msg->msg_list, &drops); - } - - if (lnet_md_exhausted(md)) - break; - } - - LNET_UNLOCK(); - - list_for_each_safe (entry, tmp, &drops) { - msg = list_entry(entry, lnet_msg_t, msg_list); - - list_del(&msg->msg_list); + lnet_hdr_t *hdr = &msg->msg_hdr; - lnet_drop_delayed_put(msg, "Bad match"); - } - - list_for_each_safe (entry, tmp, &matches) { - msg = list_entry(entry, lnet_msg_t, msg_list); + if (msg->msg_wanted != 0) + lnet_setpayloadbuffer(msg); - list_del(&msg->msg_list); + lnet_build_msg_event(msg, LNET_EVENT_PUT); - /* md won't disappear under me, since each msg - * holds a ref on it */ - lnet_recv_put(md, msg, 1, - msg->msg_ev.offset, - msg->msg_ev.mlength); - } + /* Must I ACK? If so I'll grab the ack_wmd out of the header and put + * it back into the ACK during lnet_finalize() */ + msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) && + (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0); - LNET_LOCK(); + lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed, + msg->msg_offset, msg->msg_wanted, hdr->payload_length); } static int lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg) { - int rc; - int index; - __u64 version; - lnet_hdr_t *hdr = &msg->msg_hdr; - unsigned int rlength = hdr->payload_length; - unsigned int mlength = 0; - unsigned int offset = 0; - lnet_process_id_t src= {0}; - lnet_libmd_t *md; - lnet_portal_t *ptl; + lnet_hdr_t *hdr = &msg->msg_hdr; + struct lnet_match_info info; + int rc; - src.nid = hdr->src_nid; - src.pid = hdr->src_pid; + /* 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); - /* 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; - index = hdr->msg.put.ptl_index; - ptl = &the_lnet.ln_portals[index]; - - LNET_LOCK(); + msg->msg_rx_ready_delay = ni->ni_lnd->lnd_eager_recv == NULL; again: - rc = lnet_match_md(index, LNET_MD_OP_PUT, src, - rlength, hdr->msg.put.offset, - hdr->msg.put.match_bits, msg, - &mlength, &offset, &md); + rc = lnet_ptl_match_md(&info, msg); switch (rc) { default: LBUG(); case LNET_MATCHMD_OK: - LNET_UNLOCK(); - lnet_recv_put(md, msg, msg->msg_delayed, offset, mlength); + lnet_recv_put(ni, msg); return 0; case LNET_MATCHMD_NONE: - version = ptl->ptl_ml_version; - - rc = 0; - if (!msg->msg_delayed) - rc = lnet_eager_recv_locked(msg); - - if (rc == 0 && - !the_lnet.ln_shutdown && - ((ptl->ptl_options & LNET_PTL_LAZY) != 0)) { - if (version != ptl->ptl_ml_version) - goto again; - - list_add_tail(&msg->msg_list, &ptl->ptl_msgq); - ptl->ptl_msgq_version++; - LNET_UNLOCK(); - - CDEBUG(D_NET, "Delaying PUT from %s portal %d match " - LPU64" offset %d length %d: no match \n", - libcfs_id2str(src), index, - hdr->msg.put.match_bits, - hdr->msg.put.offset, rlength); - return 0; - } - /* fall through */ - - case LNET_MATCHMD_DROP: - CDEBUG(D_NETERROR, - "Dropping PUT from %s portal %d match "LPU64 - " offset %d length %d: %d\n", - libcfs_id2str(src), index, - hdr->msg.put.match_bits, - hdr->msg.put.offset, rlength, rc); - LNET_UNLOCK(); - - return ENOENT; /* +ve: OK but no match */ - } + if (msg->msg_rx_delayed) /* attached on delayed list */ + return 0; + + 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 int lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get) { - lnet_hdr_t *hdr = &msg->msg_hdr; - unsigned int mlength = 0; - unsigned int offset = 0; - lnet_process_id_t src = {0}; - lnet_handle_wire_t reply_wmd; - lnet_libmd_t *md; - int rc; - - src.nid = hdr->src_nid; - src.pid = hdr->src_pid; - - /* Convert get fields to host byte order */ - hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits); - hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index); - hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length); - hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset); - - LNET_LOCK(); - - rc = lnet_match_md(hdr->msg.get.ptl_index, LNET_MD_OP_GET, src, - hdr->msg.get.sink_length, hdr->msg.get.src_offset, - hdr->msg.get.match_bits, msg, - &mlength, &offset, &md); - if (rc == LNET_MATCHMD_DROP) { - CDEBUG(D_NETERROR, - "Dropping GET from %s portal %d match "LPU64 - " offset %d length %d\n", - libcfs_id2str(src), - hdr->msg.get.ptl_index, - hdr->msg.get.match_bits, - hdr->msg.get.src_offset, - hdr->msg.get.sink_length); - LNET_UNLOCK(); - return ENOENT; /* +ve: OK but no match */ - } - - LASSERT (rc == LNET_MATCHMD_OK); - - the_lnet.ln_counters.send_count++; - the_lnet.ln_counters.send_length += mlength; - - LNET_UNLOCK(); - - msg->msg_ev.type = LNET_EVENT_GET; - msg->msg_ev.target.pid = hdr->dest_pid; - msg->msg_ev.target.nid = hdr->dest_nid; - msg->msg_ev.hdr_data = 0; - - reply_wmd = hdr->msg.get.return_wmd; - - lnet_prep_send(msg, LNET_MSG_REPLY, src, offset, mlength); + 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; @@ -1890,16 +1530,17 @@ lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get) lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0); msg->msg_receiving = 0; - rc = lnet_send(ni->ni_nid, msg); - if (rc < 0) { - /* didn't get as far as lnet_ni_send() */ - CERROR("%s: Unable to send REPLY for GET from %s: %d\n", - libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), rc); + rc = lnet_send(ni->ni_nid, msg, LNET_NID_ANY); + if (rc < 0) { + /* didn't get as far as lnet_ni_send() */ + CERROR("%s: Unable to send REPLY for GET from %s: %d\n", + libcfs_nid2str(ni->ni_nid), + libcfs_id2str(info.mi_id), rc); - lnet_finalize(ni, msg, rc); - } + lnet_finalize(ni, msg, rc); + } - return 0; + return 0; } static int @@ -1911,8 +1552,10 @@ lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg) lnet_libmd_t *md; int rlength; int mlength; + int cpt; - LNET_LOCK(); + 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; @@ -1920,17 +1563,17 @@ lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg) /* NB handles only looked up by creator (no flips) */ md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd); if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) { - CDEBUG(D_NETERROR, "%s: Dropping REPLY from %s for %s " - "MD "LPX64"."LPX64"\n", - libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), - (md == NULL) ? "invalid" : "inactive", - hdr->msg.reply.dst_wmd.wh_interface_cookie, - hdr->msg.reply.dst_wmd.wh_object_cookie); + 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); - LNET_UNLOCK(); + lnet_res_unlock(cpt); return ENOENT; /* +ve: OK but no match */ } @@ -1941,12 +1584,12 @@ lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg) if (mlength < rlength && (md->md_options & LNET_MD_TRUNCATE) == 0) { - CDEBUG(D_NETERROR, "%s: Dropping REPLY from %s length %d " - "for MD "LPX64" would overflow (%d)\n", - libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), + 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_UNLOCK(); + lnet_res_unlock(cpt); return ENOENT; /* +ve: OK but no match */ } @@ -1954,29 +1597,17 @@ lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg) libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie); - lnet_commit_md(md, msg); - - if (mlength != 0) - lnet_setpayloadbuffer(msg); - - msg->msg_ev.type = LNET_EVENT_REPLY; - msg->msg_ev.target.pid = hdr->dest_pid; - msg->msg_ev.target.nid = hdr->dest_nid; - msg->msg_ev.initiator = src; - msg->msg_ev.rlength = rlength; - msg->msg_ev.mlength = mlength; - msg->msg_ev.offset = 0; + lnet_msg_attach_md(msg, md, 0, mlength); - lnet_md_deconstruct(md, &msg->msg_ev.md); - lnet_md2handle(&msg->msg_ev.md_handle, md); + if (mlength != 0) + lnet_setpayloadbuffer(msg); - the_lnet.ln_counters.recv_count++; - the_lnet.ln_counters.recv_length += mlength; + lnet_res_unlock(cpt); - LNET_UNLOCK(); + lnet_build_msg_event(msg, LNET_EVENT_REPLY); - lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength); - return 0; + lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength); + return 0; } static int @@ -1985,6 +1616,7 @@ lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg) 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; @@ -1993,7 +1625,8 @@ lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg) hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits); hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength); - LNET_LOCK(); + cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie); + lnet_res_lock(cpt); /* NB handles only looked up by creator (no flips) */ md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd); @@ -2009,7 +1642,7 @@ lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg) CERROR("Source MD also attached to portal %d\n", md->md_me->me_portal); - LNET_UNLOCK(); + lnet_res_unlock(cpt); return ENOENT; /* +ve! */ } @@ -2017,24 +1650,39 @@ lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg) libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), hdr->msg.ack.dst_wmd.wh_object_cookie); - lnet_commit_md(md, msg); + lnet_msg_attach_md(msg, md, 0, 0); - msg->msg_ev.type = LNET_EVENT_ACK; - msg->msg_ev.target.pid = hdr->dest_pid; - msg->msg_ev.target.nid = hdr->dest_nid; - msg->msg_ev.initiator = src; - msg->msg_ev.mlength = hdr->msg.ack.mlength; - msg->msg_ev.match_bits = hdr->msg.ack.match_bits; + lnet_res_unlock(cpt); - lnet_md_deconstruct(md, &msg->msg_ev.md); - lnet_md2handle(&msg->msg_ev.md_handle, md); + lnet_build_msg_event(msg, LNET_EVENT_ACK); - the_lnet.ln_counters.recv_count++; + lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len); + return 0; +} - LNET_UNLOCK(); +static int +lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg) +{ + int rc = 0; - lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len); - return 0; +#ifdef __KERNEL__ + 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; } char * @@ -2055,6 +1703,7 @@ lnet_msgtyp2str (int type) return (""); } } +EXPORT_SYMBOL(lnet_msgtyp2str); void lnet_print_hdr(lnet_hdr_t * hdr) @@ -2119,27 +1768,29 @@ lnet_print_hdr(lnet_hdr_t * hdr) } int -lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid, - void *private, int rdma_req) +lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid, + void *private, int rdma_req) { - int rc = 0; - int for_me; - lnet_msg_t *msg; + 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; + lnet_nid_t src_nid; + __u32 payload_length; + __u32 type; - LASSERT (!in_interrupt ()); + 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); + 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: @@ -2174,17 +1825,16 @@ lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid, return -EPROTO; } - if (the_lnet.ln_routing) { - cfs_time_t now = cfs_time_current(); - - LNET_LOCK(); - - ni->ni_last_alive = now; - if (ni->ni_status != NULL && - ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) - ni->ni_status->ns_status = LNET_NI_STATUS_UP; + if (the_lnet.ln_routing && + ni->ni_last_alive != cfs_time_current_sec()) { + lnet_ni_lock(ni); - LNET_UNLOCK(); + /* NB: so far here is the only place to set NI status to "up */ + 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 @@ -2235,7 +1885,7 @@ lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid, /* 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 */ + if (!cfs_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", @@ -2260,56 +1910,51 @@ lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid, msg->msg_len = msg->msg_wanted = payload_length; msg->msg_offset = 0; msg->msg_hdr = *hdr; - - LNET_LOCK(); - rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid); - if (rc != 0) { - LNET_UNLOCK(); - CERROR("%s, src %s: Dropping %s " - "(error %d looking up sender)\n", - libcfs_nid2str(from_nid), libcfs_nid2str(src_nid), - lnet_msgtyp2str(type), rc); - goto free_drop; - } - LNET_UNLOCK(); - -#ifndef __KERNEL__ - LASSERT (for_me); -#else - if (!for_me) { - msg->msg_target.pid = dest_pid; - msg->msg_target.nid = dest_nid; - msg->msg_routing = 1; - msg->msg_offset = 0; - - LNET_LOCK(); - if (msg->msg_rxpeer->lp_rtrcredits <= 0 || - lnet_msg2bufpool(msg)->rbp_credits <= 0) { - rc = lnet_eager_recv_locked(msg); - if (rc != 0) { - LNET_UNLOCK(); - goto free_drop; - } - } - lnet_commit_routedmsg(msg); - rc = lnet_post_routed_recv_locked(msg, 0); - LNET_UNLOCK(); - - if (rc == 0) - lnet_ni_recv(ni, msg->msg_private, msg, 0, - 0, payload_length, payload_length); - return 0; - } -#endif - /* convert common msg->hdr fields to host byteorder */ - msg->msg_hdr.type = type; - msg->msg_hdr.src_nid = src_nid; - msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid); - msg->msg_hdr.dest_nid = dest_nid; - msg->msg_hdr.dest_pid = dest_pid; - msg->msg_hdr.payload_length = payload_length; - - msg->msg_ev.sender = from_nid; + /* 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 == 0) { + 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: @@ -2326,6 +1971,7 @@ lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid, break; default: LASSERT(0); + rc = -EPROTO; goto free_drop; /* prevent an unused label if !kernel */ } @@ -2335,34 +1981,148 @@ lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid, LASSERT (rc == ENOENT); free_drop: - LASSERT (msg->msg_md == NULL); - LNET_LOCK(); - if (msg->msg_rxpeer != NULL) { - lnet_peer_decref_locked(msg->msg_rxpeer); - msg->msg_rxpeer = NULL; - } - lnet_msg_free(msg); /* expects LNET_LOCK held */ - LNET_UNLOCK(); + LASSERT(msg->msg_md == NULL); + lnet_finalize(ni, msg, rc); drop: - lnet_drop_message(ni, private, payload_length); - return 0; + lnet_drop_message(ni, cpt, private, payload_length); + return 0; +} +EXPORT_SYMBOL(lnet_parse); + +void +lnet_drop_delayed_msg_list(cfs_list_t *head, char *reason) +{ + while (!cfs_list_empty(head)) { + lnet_process_id_t id = {0}; + lnet_msg_t *msg; + + msg = cfs_list_entry(head->next, lnet_msg_t, msg_list); + cfs_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); + } } +void +lnet_recv_delayed_msg_list(cfs_list_t *head) +{ + while (!cfs_list_empty(head)) { + lnet_msg_t *msg; + lnet_process_id_t id; + + msg = cfs_list_entry(head->next, lnet_msg_t, msg_list); + cfs_list_del(&msg->msg_list); + + /* md won't disappear under me, since each msg + * holds a ref on it */ + + id.nid = msg->msg_hdr.src_nid; + id.pid = msg->msg_hdr.src_pid; + + LASSERT(msg->msg_rx_delayed); + LASSERT(msg->msg_md != NULL); + LASSERT(msg->msg_rxpeer != NULL); + LASSERT(msg->msg_hdr.type == LNET_MSG_PUT); + + 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); + + 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 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) { - lnet_msg_t *msg; - lnet_libmd_t *md; - int rc; + 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 */ + if (!cfs_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", @@ -2376,27 +2136,28 @@ LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack, libcfs_id2str(target)); return -ENOMEM; } + msg->msg_vmflush = !!memory_pressure_get(); - LNET_LOCK(); + cpt = lnet_cpt_of_cookie(mdh.cookie); + lnet_res_lock(cpt); md = lnet_handle2md(&mdh); if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) { - lnet_msg_free(msg); - CERROR("Dropping PUT ("LPU64":%d:%s): MD (%d) invalid\n", match_bits, portal, libcfs_id2str(target), md == NULL ? -1 : md->md_threshold); if (md != NULL && md->md_me != NULL) CERROR("Source MD also attached to portal %d\n", md->md_me->me_portal); + lnet_res_unlock(cpt); - LNET_UNLOCK(); - return -ENOENT; - } + lnet_msg_free(msg); + return -ENOENT; + } - CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target)); + CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target)); - lnet_commit_md(md, msg); + lnet_msg_attach_md(msg, md, 0, 0); lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length); @@ -2418,29 +2179,13 @@ LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack, LNET_WIRE_HANDLE_COOKIE_NONE; } - msg->msg_ev.type = LNET_EVENT_SEND; - msg->msg_ev.initiator.nid = LNET_NID_ANY; - msg->msg_ev.initiator.pid = the_lnet.ln_pid; - msg->msg_ev.target = target; - msg->msg_ev.sender = LNET_NID_ANY; - msg->msg_ev.pt_index = portal; - msg->msg_ev.match_bits = match_bits; - msg->msg_ev.rlength = md->md_length; - msg->msg_ev.mlength = md->md_length; - msg->msg_ev.offset = offset; - msg->msg_ev.hdr_data = hdr_data; - - lnet_md_deconstruct(md, &msg->msg_ev.md); - lnet_md2handle(&msg->msg_ev.md_handle, md); + lnet_res_unlock(cpt); - the_lnet.ln_counters.send_count++; - the_lnet.ln_counters.send_length += md->md_length; + lnet_build_msg_event(msg, LNET_EVENT_SEND); - LNET_UNLOCK(); - - rc = lnet_send(self, msg); + rc = lnet_send(self, msg, LNET_NID_ANY); if (rc != 0) { - CERROR("Error sending PUT to %s: %d\n", + CNETERR( "Error sending PUT to %s: %d\n", libcfs_id2str(target), rc); lnet_finalize (NULL, msg, rc); } @@ -2448,6 +2193,7 @@ LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack, /* completion will be signalled by an event */ return 0; } +EXPORT_SYMBOL(LNetPut); lnet_msg_t * lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg) @@ -2459,14 +2205,16 @@ lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg) * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when * lnet_finalize() is called on it, so the LND must call this first */ - lnet_msg_t *msg = lnet_msg_alloc(); - lnet_libmd_t *getmd = getmsg->msg_md; - lnet_process_id_t peer_id = getmsg->msg_target; + struct lnet_msg *msg = lnet_msg_alloc(); + struct lnet_libmd *getmd = getmsg->msg_md; + lnet_process_id_t peer_id = getmsg->msg_target; + int cpt; - LASSERT (!getmsg->msg_target_is_router); - LASSERT (!getmsg->msg_routing); + LASSERT(!getmsg->msg_target_is_router); + LASSERT(!getmsg->msg_routing); - LNET_LOCK(); + cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie); + lnet_res_lock(cpt); LASSERT (getmd->md_refcount > 0); @@ -2480,44 +2228,49 @@ lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg) CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n", libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd); - goto drop_msg; - } - - LASSERT (getmd->md_offset == 0); + lnet_res_unlock(cpt); + goto drop; + } - CDEBUG(D_NET, "%s: Reply from %s md %p\n", - libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd); + LASSERT(getmd->md_offset == 0); - lnet_commit_md (getmd, msg); + CDEBUG(D_NET, "%s: Reply from %s md %p\n", + libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd); - msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */ + /* setup information for lnet_build_msg_event */ + msg->msg_from = peer_id.nid; + msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */ + msg->msg_hdr.src_nid = peer_id.nid; + msg->msg_hdr.payload_length = getmd->md_length; + msg->msg_receiving = 1; /* required by lnet_msg_attach_md */ - msg->msg_ev.type = LNET_EVENT_REPLY; - msg->msg_ev.initiator = peer_id; - msg->msg_ev.sender = peer_id.nid; /* optimized GETs can't be routed */ - msg->msg_ev.rlength = msg->msg_ev.mlength = getmd->md_length; - msg->msg_ev.offset = 0; + lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length); + lnet_res_unlock(cpt); - lnet_md_deconstruct(getmd, &msg->msg_ev.md); - lnet_md2handle(&msg->msg_ev.md_handle, getmd); + cpt = lnet_cpt_of_nid(peer_id.nid); - the_lnet.ln_counters.recv_count++; - the_lnet.ln_counters.recv_length += getmd->md_length; + lnet_net_lock(cpt); + lnet_msg_commit(msg, cpt); + lnet_net_unlock(cpt); - LNET_UNLOCK(); + lnet_build_msg_event(msg, LNET_EVENT_REPLY); - return msg; + return msg; - drop_msg: - lnet_msg_free(msg); drop: - the_lnet.ln_counters.drop_count++; - the_lnet.ln_counters.drop_length += getmd->md_length; + cpt = lnet_cpt_of_nid(peer_id.nid); - LNET_UNLOCK (); + 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); - return NULL; + if (msg != NULL) + lnet_msg_free(msg); + + 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) @@ -2534,20 +2287,42 @@ lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len) 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 -LNetGet(lnet_nid_t self, lnet_handle_md_t mdh, - lnet_process_id_t target, unsigned int portal, - __u64 match_bits, unsigned int 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) { - lnet_msg_t *msg; - lnet_libmd_t *md; - int rc; + 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 */ + if (!cfs_list_empty (&the_lnet.ln_test_peers) && /* normally we don't */ fail_peer (target.nid, 1)) /* shall we now? */ { CERROR("Dropping GET to %s: simulated failure\n", @@ -2562,12 +2337,11 @@ LNetGet(lnet_nid_t self, lnet_handle_md_t mdh, return -ENOMEM; } - LNET_LOCK(); + cpt = lnet_cpt_of_cookie(mdh.cookie); + lnet_res_lock(cpt); md = lnet_handle2md(&mdh); if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) { - lnet_msg_free(msg); - CERROR("Dropping GET ("LPU64":%d:%s): MD (%d) invalid\n", match_bits, portal, libcfs_id2str(target), md == NULL ? -1 : md->md_threshold); @@ -2575,13 +2349,16 @@ LNetGet(lnet_nid_t self, lnet_handle_md_t mdh, CERROR("REPLY MD also attached to portal %d\n", md->md_me->me_portal); - LNET_UNLOCK(); + lnet_res_unlock(cpt); + + lnet_msg_free(msg); + return -ENOENT; } CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target)); - lnet_commit_md(md, msg); + lnet_msg_attach_md(msg, md, 0, 0); lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0); @@ -2591,33 +2368,18 @@ LNetGet(lnet_nid_t self, lnet_handle_md_t mdh, msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length); /* NB handles only looked up by creator (no flips) */ - msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie = - the_lnet.ln_interface_cookie; - msg->msg_hdr.msg.get.return_wmd.wh_object_cookie = - md->md_lh.lh_cookie; - - msg->msg_ev.type = LNET_EVENT_SEND; - msg->msg_ev.initiator.nid = LNET_NID_ANY; - msg->msg_ev.initiator.pid = the_lnet.ln_pid; - msg->msg_ev.target = target; - msg->msg_ev.sender = LNET_NID_ANY; - msg->msg_ev.pt_index = portal; - msg->msg_ev.match_bits = match_bits; - msg->msg_ev.rlength = md->md_length; - msg->msg_ev.mlength = md->md_length; - msg->msg_ev.offset = offset; - msg->msg_ev.hdr_data = 0; - - lnet_md_deconstruct(md, &msg->msg_ev.md); - lnet_md2handle(&msg->msg_ev.md_handle, md); - - the_lnet.ln_counters.send_count++; - - LNET_UNLOCK(); - - rc = lnet_send(self, msg); + 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) { - CERROR("error sending GET to %s: %d\n", + CNETERR( "Error sending GET to %s: %d\n", libcfs_id2str(target), rc); lnet_finalize (NULL, msg, rc); } @@ -2625,16 +2387,33 @@ LNetGet(lnet_nid_t self, lnet_handle_md_t mdh, /* completion will be signalled by an event */ return 0; } +EXPORT_SYMBOL(LNetGet); +/** + * Calculate distance to node at \a dstnid. + * + * \param dstnid Target NID. + * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid + * is saved here. + * \param orderp If not NULL, order of the route to reach \a dstnid is saved + * here. + * + * \retval 0 If \a dstnid belongs to a local interface, and reserved option + * local_nid_dist_zero is set, which is the default. + * \retval positives Distance to target NID, i.e. number of hops plus one. + * \retval -EHOSTUNREACH If \a dstnid is not reachable. + */ int -LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp) +LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp) { - struct list_head *e; - lnet_ni_t *ni; - lnet_remotenet_t *rnet; - __u32 dstnet = LNET_NIDNET(dstnid); - int hops; - __u32 order = 2; + cfs_list_t *e; + struct lnet_ni *ni; + lnet_remotenet_t *rnet; + __u32 dstnet = LNET_NIDNET(dstnid); + int hops; + int cpt; + __u32 order = 2; + cfs_list_t *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 @@ -2644,10 +2423,10 @@ LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp) LASSERT (the_lnet.ln_init); LASSERT (the_lnet.ln_refcount > 0); - LNET_LOCK(); + cpt = lnet_net_lock_current(); - list_for_each (e, &the_lnet.ln_nis) { - ni = list_entry(e, lnet_ni_t, ni_list); + cfs_list_for_each (e, &the_lnet.ln_nis) { + ni = cfs_list_entry(e, lnet_ni_t, ni_list); if (ni->ni_nid == dstnid) { if (srcnidp != NULL) @@ -2658,7 +2437,7 @@ LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp) else *orderp = 1; } - LNET_UNLOCK(); + lnet_net_unlock(cpt); return local_nid_dist_zero ? 0 : 1; } @@ -2668,23 +2447,25 @@ LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp) *srcnidp = ni->ni_nid; if (orderp != NULL) *orderp = order; - LNET_UNLOCK(); - return 1; - } + lnet_net_unlock(cpt); + return 1; + } - order++; - } + order++; + } - list_for_each (e, &the_lnet.ln_remote_nets) { - rnet = list_entry(e, lnet_remotenet_t, lrn_list); + rn_list = lnet_net2rnethash(dstnet); + cfs_list_for_each(e, rn_list) { + rnet = cfs_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)); + LASSERT (!cfs_list_empty(&rnet->lrn_routes)); - list_for_each_entry(route, &rnet->lrn_routes, lr_list) { + cfs_list_for_each_entry(route, &rnet->lrn_routes, + lr_list) { if (shortest == NULL || route->lr_hops < shortest->lr_hops) shortest = route; @@ -2696,16 +2477,34 @@ LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp) *srcnidp = shortest->lr_gateway->lp_ni->ni_nid; if (orderp != NULL) *orderp = order; - LNET_UNLOCK(); - return hops + 1; - } - order++; - } - - LNET_UNLOCK(); - return -EHOSTUNREACH; + lnet_net_unlock(cpt); + return hops + 1; + } + order++; + } + + lnet_net_unlock(cpt); + return -EHOSTUNREACH; } +EXPORT_SYMBOL(LNetDist); +/** + * 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) { @@ -2714,16 +2513,16 @@ LNetSetAsync(lnet_process_id_t id, int nasync) #else lnet_ni_t *ni; lnet_remotenet_t *rnet; - struct list_head *tmp; + cfs_list_t *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? */ - + /* Target on a local network? */ ni = lnet_net2ni(LNET_NIDNET(id.nid)); if (ni != NULL) { if (ni->ni_lnd->lnd_setasync != NULL) @@ -2740,22 +2539,22 @@ LNetSetAsync(lnet_process_id_t id, int nasync) nnids = 0; /* Snapshot all the router NIDs */ - LNET_LOCK(); - rnet = lnet_find_net_locked(LNET_NIDNET(id.nid)); - if (rnet != NULL) { - list_for_each(tmp, &rnet->lrn_routes) { - if (nnids == maxnids) { - LNET_UNLOCK(); + cpt = lnet_net_lock_current(); + rnet = lnet_find_net_locked(LNET_NIDNET(id.nid)); + if (rnet != NULL) { + cfs_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); + route = cfs_list_entry(tmp, lnet_route_t, lr_list); nids[nnids++] = route->lr_gateway->lp_nid; } } - LNET_UNLOCK(); + lnet_net_unlock(cpt); /* set async on all the routers */ while (nnids-- > 0) { @@ -2778,3 +2577,4 @@ LNetSetAsync(lnet_process_id_t id, int nasync) return rc; #endif } +EXPORT_SYMBOL(LNetSetAsync);