ptllnd.c

非常经典的一个分布式系统

        CDEBUG(D_NET, "nposted_buffers = %d (before)\n",plni->plni_nposted_buffers);
        CDEBUG(D_NET, "nbuffers = %d (before)\n",plni->plni_nbuffers);

	plni->plni_nmsgs += delta;
	LASSERT(plni->plni_nmsgs >= 0);
	
        nmsgs = plni->plni_nmsgs + plni->plni_msgs_spare;

        nbufs = (nmsgs * plni->plni_max_msg_size + plni->plni_buffer_size - 1) /
                plni->plni_buffer_size;

        while (nbufs > plni->plni_nbuffers) {
                buf = ptllnd_create_buffer(ni);

                if (buf == NULL)
                        return -ENOMEM;

                rc = ptllnd_post_buffer(buf);
                if (rc != 0) {
                        /* TODO - this path seems to orpahn the buffer
                         * in a state where its not posted and will never be
                         * However it does not leak the buffer as it's
                         * already been put onto the global buffer list
                         * and will be cleaned up
                         */
                        return rc;
                }
        }

        CDEBUG(D_NET, "nposted_buffers = %d (after)\n",plni->plni_nposted_buffers);
        CDEBUG(D_NET, "nbuffers = %d (after)\n",plni->plni_nbuffers);
        return 0;
}

void
ptllnd_destroy_buffers (lnet_ni_t *ni)
{
        ptllnd_ni_t       *plni = ni->ni_data;
        ptllnd_buffer_t   *buf;
        struct list_head  *tmp;
        struct list_head  *nxt;

        CDEBUG(D_NET, "nposted_buffers = %d (before)\n",plni->plni_nposted_buffers);
        CDEBUG(D_NET, "nbuffers = %d (before)\n",plni->plni_nbuffers);

        list_for_each_safe(tmp, nxt, &plni->plni_buffers) {
                buf = list_entry(tmp, ptllnd_buffer_t, plb_list);

                //CDEBUG(D_NET, "buf=%p posted=%d\n",buf,buf->plb_posted);

                LASSERT (plni->plni_nbuffers > 0);
                if (buf->plb_posted) {
			time_t   start = cfs_time_current_sec();
			int      w = plni->plni_long_wait;

                        LASSERT (plni->plni_nposted_buffers > 0);

#ifdef LUSTRE_PORTALS_UNLINK_SEMANTICS
                        (void) PtlMDUnlink(buf->plb_md);

			while (buf->plb_posted) {
				if (w > 0 && cfs_time_current_sec() > start + w/1000) {
					CWARN("Waited %ds to unlink buffer\n",
					      (int)(cfs_time_current_sec() - start));
					w *= 2;
				}
				ptllnd_wait(ni, w);
			}
#else
                        while (buf->plb_posted) {
                                rc = PtlMDUnlink(buf->plb_md);
                                if (rc == PTL_OK) {
                                        buf->plb_posted = 0;
                                        plni->plni_nposted_buffers--;
                                        break;
                                }
                                LASSERT (rc == PTL_MD_IN_USE);
				if (w > 0 && cfs_time_current_sec() > start + w/1000) {
					CWARN("Waited %ds to unlink buffer\n",
					      cfs_time_current_sec() - start);
					w *= 2;
				}
				ptllnd_wait(ni, w);
                        }
#endif
                }
                ptllnd_destroy_buffer(buf);
        }

        CDEBUG(D_NET, "nposted_buffers = %d (after)\n",plni->plni_nposted_buffers);
        CDEBUG(D_NET, "nbuffers = %d (after)\n",plni->plni_nbuffers);

        LASSERT (plni->plni_nposted_buffers == 0);
        LASSERT (plni->plni_nbuffers == 0);
}

int
ptllnd_create_peer_hash (lnet_ni_t *ni)
{
        ptllnd_ni_t *plni = ni->ni_data;
        int          i;

        plni->plni_npeers = 0;

        LIBCFS_ALLOC(plni->plni_peer_hash,
                     plni->plni_peer_hash_size * sizeof(*plni->plni_peer_hash));
        if (plni->plni_peer_hash == NULL) {
                CERROR("Can't allocate ptllnd peer hash (size %d)\n",
                       plni->plni_peer_hash_size);
                return -ENOMEM;
        }

        for (i = 0; i < plni->plni_peer_hash_size; i++)
                CFS_INIT_LIST_HEAD(&plni->plni_peer_hash[i]);

        return 0;
}

void
ptllnd_destroy_peer_hash (lnet_ni_t *ni)
{
        ptllnd_ni_t    *plni = ni->ni_data;
        int             i;

        LASSERT( plni->plni_npeers == 0);

        for (i = 0; i < plni->plni_peer_hash_size; i++)
                LASSERT (list_empty(&plni->plni_peer_hash[i]));

        LIBCFS_FREE(plni->plni_peer_hash,
                    plni->plni_peer_hash_size * sizeof(*plni->plni_peer_hash));
}

void
ptllnd_close_peers (lnet_ni_t *ni)
{
        ptllnd_ni_t    *plni = ni->ni_data;
        ptllnd_peer_t  *plp;
        int             i;

        for (i = 0; i < plni->plni_peer_hash_size; i++)
                while (!list_empty(&plni->plni_peer_hash[i])) {
                        plp = list_entry(plni->plni_peer_hash[i].next,
                                         ptllnd_peer_t, plp_list);

                        ptllnd_close_peer(plp, 0);
                }
}

int
ptllnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
{
	switch (cmd) {
	case IOC_LIBCFS_DEBUG_PEER:
		ptllnd_dump_debug(ni, *((lnet_process_id_t *)arg));
		return 0;
		
	default:
		return -EINVAL;
	}
}

__u64
ptllnd_get_timestamp(void)
{
        struct timeval  tv;
        int             rc = gettimeofday(&tv, NULL);

        LASSERT (rc == 0);
        return ((__u64)tv.tv_sec) * 1000000 + tv.tv_usec;
}

void
ptllnd_shutdown (lnet_ni_t *ni)
{
        ptllnd_ni_t *plni = ni->ni_data;
        int          rc;
	time_t       start = cfs_time_current_sec();
	int          w = plni->plni_long_wait;

        LASSERT (ptllnd_ni_count == 1);
	plni->plni_max_tx_history = 0;

	ptllnd_cull_tx_history(plni);

        ptllnd_close_peers(ni);
        ptllnd_destroy_buffers(ni);

        while (plni->plni_npeers > 0) {
		if (w > 0 && cfs_time_current_sec() > start + w/1000) {
			CWARN("Waited %ds for peers to shutdown\n",
			      (int)(cfs_time_current_sec() - start));
			w *= 2;
		}
                ptllnd_wait(ni, w);
	}

        LASSERT (plni->plni_ntxs == 0);
        LASSERT (plni->plni_nrxs == 0);

        rc = PtlEQFree(plni->plni_eqh);
        LASSERT (rc == PTL_OK);

        rc = PtlNIFini(plni->plni_nih);
        LASSERT (rc == PTL_OK);

        ptllnd_destroy_peer_hash(ni);
        LIBCFS_FREE(plni, sizeof(*plni));
        ptllnd_ni_count--;
}

int
ptllnd_startup (lnet_ni_t *ni)
{
        ptllnd_ni_t *plni;
        int          rc;

	/* could get limits from portals I guess... */
	ni->ni_maxtxcredits =
	ni->ni_peertxcredits = 1000;

        if (ptllnd_ni_count != 0) {
                CERROR("Can't have > 1 instance of ptllnd\n");
                return -EPERM;
        }

        ptllnd_ni_count++;

	rc = ptllnd_history_init();
	if (rc != 0) {
		CERROR("Can't init history\n");
		goto failed0;
	}
	
        LIBCFS_ALLOC(plni, sizeof(*plni));
        if (plni == NULL) {
                CERROR("Can't allocate ptllnd state\n");
                rc = -ENOMEM;
                goto failed0;
        }

        ni->ni_data = plni;

        plni->plni_stamp = ptllnd_get_timestamp();
        plni->plni_nrxs = 0;
        plni->plni_ntxs = 0;
	plni->plni_ntx_history = 0;
	plni->plni_watchdog_peeridx = 0;
	plni->plni_watchdog_nextt = cfs_time_current_sec();
        CFS_INIT_LIST_HEAD(&plni->plni_zombie_txs);
        CFS_INIT_LIST_HEAD(&plni->plni_tx_history);

        /*
         *  Initilize buffer related data structures
         */
        CFS_INIT_LIST_HEAD(&plni->plni_buffers);
        plni->plni_nbuffers = 0;
        plni->plni_nposted_buffers = 0;

        rc = ptllnd_get_tunables(ni);
        if (rc != 0)
                goto failed1;

        rc = ptllnd_create_peer_hash(ni);
        if (rc != 0)
                goto failed1;

        /* NB I most probably won't get the PID I requested here.  It doesn't
         * matter because I don't need a fixed PID (only connection acceptors
         * need a "well known" PID). */

        rc = PtlNIInit(PTL_IFACE_DEFAULT, plni->plni_ptllnd_pid,
                       NULL, NULL, &plni->plni_nih);
        if (rc != PTL_OK && rc != PTL_IFACE_DUP) {
                CERROR("PtlNIInit failed: %s(%d)\n",
		       ptllnd_errtype2str(rc), rc);
                rc = -ENODEV;
                goto failed2;
        }

        rc = PtlEQAlloc(plni->plni_nih, plni->plni_eq_size,
                        PTL_EQ_HANDLER_NONE, &plni->plni_eqh);
        if (rc != PTL_OK) {
                CERROR("PtlEQAlloc failed: %s(%d)\n",
		       ptllnd_errtype2str(rc), rc);
                rc = -ENODEV;
                goto failed3;
        }

        /*
         * Fetch the Portals NID
         */
	rc = PtlGetId(plni->plni_nih, &plni->plni_portals_id);
        if (rc != PTL_OK) {
                CERROR ("PtlGetID failed : %s(%d)\n",
			ptllnd_errtype2str(rc), rc);
                rc = -EINVAL;
                goto failed4;
        }

        /*
         * Create the new NID.  Based on the LND network type
         * and the lower ni's address data.
         */
        ni->ni_nid = ptllnd_ptl2lnetnid(ni, plni->plni_portals_id.nid);

        CDEBUG(D_NET, "ptl id  =%s\n", ptllnd_ptlid2str(plni->plni_portals_id));
        CDEBUG(D_NET, "lnet id =%s (passed back)\n",
               libcfs_id2str((lnet_process_id_t) {
                       .nid = ni->ni_nid, .pid = the_lnet.ln_pid}));

        rc = ptllnd_size_buffers(ni, 0);
        if (rc != 0)
                goto failed4;

	return 0;

 failed4:
        ptllnd_destroy_buffers(ni);
        PtlEQFree(plni->plni_eqh);
 failed3:
        PtlNIFini(plni->plni_nih);
 failed2:
        ptllnd_destroy_peer_hash(ni);
 failed1:
        LIBCFS_FREE(plni, sizeof(*plni));
 failed0:
	ptllnd_history_fini();
        ptllnd_ni_count--;
        CDEBUG(D_NET, "<<< rc=%d\n",rc);
        return rc;
}

const char *ptllnd_evtype2str(int type)
{
#define DO_TYPE(x) case x: return #x;
        switch(type)
        {
                DO_TYPE(PTL_EVENT_GET_START);
                DO_TYPE(PTL_EVENT_GET_END);
                DO_TYPE(PTL_EVENT_PUT_START);
                DO_TYPE(PTL_EVENT_PUT_END);
                DO_TYPE(PTL_EVENT_REPLY_START);
                DO_TYPE(PTL_EVENT_REPLY_END);
                DO_TYPE(PTL_EVENT_ACK);
                DO_TYPE(PTL_EVENT_SEND_START);
                DO_TYPE(PTL_EVENT_SEND_END);
                DO_TYPE(PTL_EVENT_UNLINK);
        default:
                return "<unknown event type>";
        }
#undef DO_TYPE
}

const char *ptllnd_msgtype2str(int type)
{
#define DO_TYPE(x) case x: return #x;
        switch(type)
        {
                DO_TYPE(PTLLND_MSG_TYPE_INVALID);
                DO_TYPE(PTLLND_MSG_TYPE_PUT);
                DO_TYPE(PTLLND_MSG_TYPE_GET);
                DO_TYPE(PTLLND_MSG_TYPE_IMMEDIATE);
                DO_TYPE(PTLLND_MSG_TYPE_HELLO);
                DO_TYPE(PTLLND_MSG_TYPE_NOOP);
                DO_TYPE(PTLLND_MSG_TYPE_NAK);
        default:
                return "<unknown msg type>";
        }
#undef DO_TYPE
}

const char *ptllnd_errtype2str(int type)
{
#define DO_TYPE(x) case x: return #x;
        switch(type)
        {
                DO_TYPE(PTL_OK);
                DO_TYPE(PTL_SEGV);
                DO_TYPE(PTL_NO_SPACE);
                DO_TYPE(PTL_ME_IN_USE);
                DO_TYPE(PTL_NAL_FAILED);
                DO_TYPE(PTL_NO_INIT);
                DO_TYPE(PTL_IFACE_DUP);
                DO_TYPE(PTL_IFACE_INVALID);
                DO_TYPE(PTL_HANDLE_INVALID);
                DO_TYPE(PTL_MD_INVALID);
                DO_TYPE(PTL_ME_INVALID);
                DO_TYPE(PTL_PROCESS_INVALID);
                DO_TYPE(PTL_PT_INDEX_INVALID);
                DO_TYPE(PTL_SR_INDEX_INVALID);
                DO_TYPE(PTL_EQ_INVALID);
                DO_TYPE(PTL_EQ_DROPPED);
                DO_TYPE(PTL_EQ_EMPTY);
                DO_TYPE(PTL_MD_NO_UPDATE);
                DO_TYPE(PTL_FAIL);
                DO_TYPE(PTL_AC_INDEX_INVALID);
                DO_TYPE(PTL_MD_ILLEGAL);
                DO_TYPE(PTL_ME_LIST_TOO_LONG);
                DO_TYPE(PTL_MD_IN_USE);
                DO_TYPE(PTL_NI_INVALID);
                DO_TYPE(PTL_PID_INVALID);
                DO_TYPE(PTL_PT_FULL);
                DO_TYPE(PTL_VAL_FAILED);
                DO_TYPE(PTL_NOT_IMPLEMENTED);
                DO_TYPE(PTL_NO_ACK);
                DO_TYPE(PTL_EQ_IN_USE);
                DO_TYPE(PTL_PID_IN_USE);
                DO_TYPE(PTL_INV_EQ_SIZE);
                DO_TYPE(PTL_AGAIN);
        default:
                return "<unknown error type>";
        }
#undef DO_TYPE
}