ptllnd.c

非常经典的一个分布式系统

/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
 * vim:expandtab:shiftwidth=8:tabstop=8:
 *
 * Copyright (C) 2005 Cluster File Systems, Inc. All rights reserved.
 *   Author: Eric Barton <[e-mail]>
 *
 *   This file is part of the Lustre file system, http://www.lustre.org
 *   Lustre is a trademark of Cluster File Systems, Inc.
 *
 *   This file is confidential source code owned by Cluster File Systems.
 *   No viewing, modification, compilation, redistribution, or any other
 *   form of use is permitted except through a signed license agreement.
 *
 *   If you have not signed such an agreement, then you have no rights to
 *   this file.  Please destroy it immediately and contact CFS.
 *
 */

#include "ptllnd.h"

lnd_t               the_ptllnd = {
        .lnd_type       = PTLLND,
        .lnd_startup    = ptllnd_startup,
        .lnd_shutdown   = ptllnd_shutdown,
	.lnd_ctl        = ptllnd_ctl,
        .lnd_send       = ptllnd_send,
        .lnd_recv       = ptllnd_recv,
        .lnd_eager_recv = ptllnd_eager_recv,
        .lnd_notify     = ptllnd_notify,
        .lnd_wait       = ptllnd_wait,
	.lnd_setasync   = ptllnd_setasync,
};

static int ptllnd_ni_count = 0;

static struct list_head ptllnd_idle_history;
static struct list_head ptllnd_history_list;

void
ptllnd_history_fini(void)
{
	ptllnd_he_t *he;

	while (!list_empty(&ptllnd_idle_history)) {
		he = list_entry(ptllnd_idle_history.next,
				ptllnd_he_t, he_list);
		
		list_del(&he->he_list);
		LIBCFS_FREE(he, sizeof(*he));
	}
	
	while (!list_empty(&ptllnd_history_list)) {
		he = list_entry(ptllnd_history_list.next,
				ptllnd_he_t, he_list);
		
		list_del(&he->he_list);
		LIBCFS_FREE(he, sizeof(*he));
	}
}

int
ptllnd_history_init(void)
{
	int          i;
	ptllnd_he_t *he;
	int          n;
	int          rc;
	
	CFS_INIT_LIST_HEAD(&ptllnd_idle_history);
	CFS_INIT_LIST_HEAD(&ptllnd_history_list);
	
	rc = ptllnd_parse_int_tunable(&n, "PTLLND_HISTORY", 0);
	if (rc != 0)
		return rc;
	
	for (i = 0; i < n; i++) {
		LIBCFS_ALLOC(he, sizeof(*he));
		if (he == NULL) {
			ptllnd_history_fini();
			return -ENOMEM;
		}
		
		list_add(&he->he_list, &ptllnd_idle_history);
	}

	PTLLND_HISTORY("Init");

	return 0;
}

void
ptllnd_history(const char *fn, const char *file, const int line,
	       const char *fmt, ...)
{
	static int     seq;
	
        va_list        ap;
	ptllnd_he_t   *he;
	
	if (!list_empty(&ptllnd_idle_history)) {
		he = list_entry(ptllnd_idle_history.next,
				ptllnd_he_t, he_list);
	} else if (!list_empty(&ptllnd_history_list)) {
		he = list_entry(ptllnd_history_list.next,
				ptllnd_he_t, he_list);
	} else {
		return;
	}

	list_del(&he->he_list);
	list_add_tail(&he->he_list, &ptllnd_history_list);

	he->he_seq = seq++;
	he->he_fn = fn;
	he->he_file = file;
	he->he_line = line;
	gettimeofday(&he->he_time, NULL);
	
	va_start(ap, fmt);
	vsnprintf(he->he_msg, sizeof(he->he_msg), fmt, ap);
	va_end(ap);
}

void
ptllnd_dump_history(void)
{
	ptllnd_he_t    *he;

	PTLLND_HISTORY("dumping...");
	
	while (!list_empty(&ptllnd_history_list)) {
		he = list_entry(ptllnd_history_list.next,
				ptllnd_he_t, he_list);

		list_del(&he->he_list);
		
		CDEBUG(D_WARNING, "%d %d.%06d (%s:%d:%s()) %s\n", he->he_seq,
		       (int)he->he_time.tv_sec, (int)he->he_time.tv_usec,
		       he->he_file, he->he_line, he->he_fn, he->he_msg);

		list_add_tail(&he->he_list, &ptllnd_idle_history);
	}

	PTLLND_HISTORY("complete");
}

void 
ptllnd_assert_wire_constants (void)
{
        /* Wire protocol assertions generated by 'wirecheck'
         * running on Linux fedora 2.6.11-co-0.6.4 #1 Mon Jun 19 05:36:13 UTC 2006 i686 i686 i386 GNU
         * with gcc version 4.1.1 20060525 (Red Hat 4.1.1-1) */


        /* Constants... */
        CLASSERT (PTL_RESERVED_MATCHBITS == 0x100);
        CLASSERT (LNET_MSG_MATCHBITS == 0);
        CLASSERT (PTLLND_MSG_MAGIC == 0x50746C4E);
        CLASSERT (PTLLND_MSG_VERSION == 0x04);
        CLASSERT (PTLLND_RDMA_OK == 0x00);
        CLASSERT (PTLLND_RDMA_FAIL == 0x01);
        CLASSERT (PTLLND_MSG_TYPE_INVALID == 0x00);
        CLASSERT (PTLLND_MSG_TYPE_PUT == 0x01);
        CLASSERT (PTLLND_MSG_TYPE_GET == 0x02);
        CLASSERT (PTLLND_MSG_TYPE_IMMEDIATE == 0x03);
        CLASSERT (PTLLND_MSG_TYPE_NOOP == 0x04);
        CLASSERT (PTLLND_MSG_TYPE_HELLO == 0x05);
        CLASSERT (PTLLND_MSG_TYPE_NAK == 0x06);

        /* Checks for struct kptl_msg_t */
        CLASSERT ((int)sizeof(kptl_msg_t) == 136);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_magic) == 0);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_magic) == 4);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_version) == 4);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_version) == 2);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_type) == 6);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_type) == 1);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_credits) == 7);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_credits) == 1);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_nob) == 8);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_nob) == 4);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_cksum) == 12);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_cksum) == 4);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_srcnid) == 16);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_srcnid) == 8);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_srcstamp) == 24);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_srcstamp) == 8);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_dstnid) == 32);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_dstnid) == 8);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_dststamp) == 40);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_dststamp) == 8);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_srcpid) == 48);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_srcpid) == 4);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_dstpid) == 52);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_dstpid) == 4);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_u.immediate) == 56);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_u.immediate) == 72);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_u.rdma) == 56);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_u.rdma) == 80);
        CLASSERT ((int)offsetof(kptl_msg_t, ptlm_u.hello) == 56);
        CLASSERT ((int)sizeof(((kptl_msg_t *)0)->ptlm_u.hello) == 12);

        /* Checks for struct kptl_immediate_msg_t */
        CLASSERT ((int)sizeof(kptl_immediate_msg_t) == 72);
        CLASSERT ((int)offsetof(kptl_immediate_msg_t, kptlim_hdr) == 0);
        CLASSERT ((int)sizeof(((kptl_immediate_msg_t *)0)->kptlim_hdr) == 72);
        CLASSERT ((int)offsetof(kptl_immediate_msg_t, kptlim_payload[13]) == 85);
        CLASSERT ((int)sizeof(((kptl_immediate_msg_t *)0)->kptlim_payload[13]) == 1);

        /* Checks for struct kptl_rdma_msg_t */
        CLASSERT ((int)sizeof(kptl_rdma_msg_t) == 80);
        CLASSERT ((int)offsetof(kptl_rdma_msg_t, kptlrm_hdr) == 0);
        CLASSERT ((int)sizeof(((kptl_rdma_msg_t *)0)->kptlrm_hdr) == 72);
        CLASSERT ((int)offsetof(kptl_rdma_msg_t, kptlrm_matchbits) == 72);
        CLASSERT ((int)sizeof(((kptl_rdma_msg_t *)0)->kptlrm_matchbits) == 8);

        /* Checks for struct kptl_hello_msg_t */
        CLASSERT ((int)sizeof(kptl_hello_msg_t) == 12);
        CLASSERT ((int)offsetof(kptl_hello_msg_t, kptlhm_matchbits) == 0);
        CLASSERT ((int)sizeof(((kptl_hello_msg_t *)0)->kptlhm_matchbits) == 8);
        CLASSERT ((int)offsetof(kptl_hello_msg_t, kptlhm_max_msg_size) == 8);
        CLASSERT ((int)sizeof(((kptl_hello_msg_t *)0)->kptlhm_max_msg_size) == 4);
}

int
ptllnd_parse_int_tunable(int *value, char *name, int dflt)
{
        char    *env = getenv(name);
        char    *end;

        if (env == NULL) {
                *value = dflt;
                return 0;
        }

        *value = strtoull(env, &end, 0);
        if (*end == 0)
                return 0;

        CERROR("Can't parse tunable %s=%s\n", name, env);
        return -EINVAL;
}

int
ptllnd_get_tunables(lnet_ni_t *ni)
{
        ptllnd_ni_t *plni = ni->ni_data;
        int          max_msg_size;
        int          msgs_per_buffer;
        int          rc;
        int          temp;

	/*  Other tunable defaults depend on this */
	rc = ptllnd_parse_int_tunable(&plni->plni_debug, "PTLLND_DEBUG", 0);
	if (rc != 0)
		return rc;

        rc = ptllnd_parse_int_tunable(&plni->plni_portal,
                                      "PTLLND_PORTAL", PTLLND_PORTAL);
        if (rc != 0)
                return rc;

        rc = ptllnd_parse_int_tunable(&temp,
                                      "PTLLND_PID", PTLLND_PID);
        if (rc != 0)
                return rc;
        plni->plni_ptllnd_pid = (ptl_pid_t)temp;

        rc = ptllnd_parse_int_tunable(&plni->plni_peer_credits,
                                      "PTLLND_PEERCREDITS", PTLLND_PEERCREDITS);
        if (rc != 0)
                return rc;

        rc = ptllnd_parse_int_tunable(&max_msg_size,
                                      "PTLLND_MAX_MSG_SIZE",
                                      PTLLND_MAX_ULND_MSG_SIZE);
        if (rc != 0)
                return rc;

        rc = ptllnd_parse_int_tunable(&msgs_per_buffer,
                                      "PTLLND_MSGS_PER_BUFFER", 64);
        if (rc != 0)
                return rc;

        rc = ptllnd_parse_int_tunable(&plni->plni_msgs_spare,
                                      "PTLLND_MSGS_SPARE", 256);
        if (rc != 0)
                return rc;

        rc = ptllnd_parse_int_tunable(&plni->plni_peer_hash_size,
                                      "PTLLND_PEER_HASH_SIZE", 101);
        if (rc != 0)
                return rc;


        rc = ptllnd_parse_int_tunable(&plni->plni_eq_size,
                                      "PTLLND_EQ_SIZE", 1024);
        if (rc != 0)
                return rc;

	rc = ptllnd_parse_int_tunable(&plni->plni_checksum,
				      "PTLLND_CHECKSUM", 0);
	if (rc != 0)
		return rc;

	rc = ptllnd_parse_int_tunable(&plni->plni_max_tx_history,
				      "PTLLND_TX_HISTORY",
				      plni->plni_debug ? 1024 : 0);
	if (rc != 0)
		return rc;

	rc = ptllnd_parse_int_tunable(&plni->plni_abort_on_protocol_mismatch,
				      "PTLLND_ABORT_ON_PROTOCOL_MISMATCH", 1);
	if (rc != 0)
		return rc;

	rc = ptllnd_parse_int_tunable(&plni->plni_abort_on_nak,
				      "PTLLND_ABORT_ON_NAK", 0);
	if (rc != 0)
		return rc;

	rc = ptllnd_parse_int_tunable(&plni->plni_dump_on_nak,
				      "PTLLND_DUMP_ON_NAK", plni->plni_debug);
	if (rc != 0)
		return rc;

	rc = ptllnd_parse_int_tunable(&plni->plni_watchdog_interval,
				      "PTLLND_WATCHDOG_INTERVAL", 1);
	if (rc != 0)
		return rc;
	if (plni->plni_watchdog_interval <= 0)
		plni->plni_watchdog_interval = 1;

	rc = ptllnd_parse_int_tunable(&plni->plni_timeout,
				      "PTLLND_TIMEOUT", 50);
	if (rc != 0)
		return rc;

	rc = ptllnd_parse_int_tunable(&plni->plni_long_wait,
				      "PTLLND_LONG_WAIT",
				      plni->plni_debug ? 5 : plni->plni_timeout);
	if (rc != 0)
		return rc;
	plni->plni_long_wait *= 1000;		/* convert to mS */

        plni->plni_max_msg_size = max_msg_size & ~7;
        if (plni->plni_max_msg_size < PTLLND_MIN_BUFFER_SIZE)
                plni->plni_max_msg_size = PTLLND_MIN_BUFFER_SIZE;
	CLASSERT ((PTLLND_MIN_BUFFER_SIZE & 7) == 0);
	CLASSERT (sizeof(kptl_msg_t) <= PTLLND_MIN_BUFFER_SIZE);

        plni->plni_buffer_size = plni->plni_max_msg_size * msgs_per_buffer;

        CDEBUG(D_NET, "portal          = %d\n",plni->plni_portal);
        CDEBUG(D_NET, "ptllnd_pid      = %d\n",plni->plni_ptllnd_pid);
        CDEBUG(D_NET, "max_msg_size    = %d\n",max_msg_size);
        CDEBUG(D_NET, "msgs_per_buffer = %d\n",msgs_per_buffer);
        CDEBUG(D_NET, "msgs_spare      = %d\n",plni->plni_msgs_spare);
        CDEBUG(D_NET, "peer_hash_size  = %d\n",plni->plni_peer_hash_size);
        CDEBUG(D_NET, "eq_size         = %d\n",plni->plni_eq_size);
        CDEBUG(D_NET, "max_msg_size    = %d\n",plni->plni_max_msg_size);
        CDEBUG(D_NET, "buffer_size     = %d\n",plni->plni_buffer_size);

        return 0;
}

ptllnd_buffer_t *
ptllnd_create_buffer (lnet_ni_t *ni)
{
        ptllnd_ni_t     *plni = ni->ni_data;
        ptllnd_buffer_t *buf;

        LIBCFS_ALLOC(buf, sizeof(*buf));
        if (buf == NULL) {
                CERROR("Can't allocate buffer descriptor\n");
                return NULL;
        }

        buf->plb_ni = ni;
        buf->plb_posted = 0;
        CFS_INIT_LIST_HEAD(&buf->plb_list);

        LIBCFS_ALLOC(buf->plb_buffer, plni->plni_buffer_size);
        if (buf->plb_buffer == NULL) {
                CERROR("Can't allocate buffer size %d\n",
                       plni->plni_buffer_size);
                LIBCFS_FREE(buf, sizeof(*buf));
                return NULL;
        }

        list_add(&buf->plb_list, &plni->plni_buffers);
        plni->plni_nbuffers++;

        return buf;
}

void
ptllnd_destroy_buffer (ptllnd_buffer_t *buf)
{
        ptllnd_ni_t     *plni = buf->plb_ni->ni_data;

        LASSERT (!buf->plb_posted);

        plni->plni_nbuffers--;
        list_del(&buf->plb_list);
        LIBCFS_FREE(buf->plb_buffer, plni->plni_buffer_size);
        LIBCFS_FREE(buf, sizeof(*buf));
}

int
ptllnd_size_buffers (lnet_ni_t *ni, int delta)
{
        ptllnd_ni_t     *plni = ni->ni_data;
        ptllnd_buffer_t *buf;
        int              nmsgs;
        int              nbufs;
        int              rc;