free another tree

[freeradius.git] / src / main / request_list.c

/*
 * request_list.c       Hide the handling of the REQUEST list from
 *                      the main server.
 *
 * Version:     $Id$
 *
 *   This program is free software; you can redistribute it and/or modify
 *   it under the terms of the GNU General Public License as published by
 *   the Free Software Foundation; either version 2 of the License, or
 *   (at your option) any later version.
 *
 *   This program is distributed in the hope that it will be useful,
 *   but WITHOUT ANY WARRANTY; without even the implied warranty of
 *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *   GNU General Public License for more details.
 *
 *   You should have received a copy of the GNU General Public License
 *   along with this program; if not, write to the Free Software
 *   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 *
 * Copyright 2003-2004  The FreeRADIUS server project
 */
static const char rcsid[] = "$Id$";

#include "autoconf.h"
#include "libradius.h"

#include <stdlib.h>
#include <string.h>

#include "radiusd.h"
#include "rad_assert.h"
#include "request_list.h"
#include "radius_snmp.h"


/*
 *      We keep the incoming requests in an array, indexed by ID.
 *
 *      Each array element contains a linked list of containers of
 *      active requests, a count of the number of requests, and a time
 *      at which the first request in the list must be serviced.
 *
 *      Note that we ALSO keep a tree view of the same data, below.
 *      Both views are needed for the server to work optimally.
 */
typedef struct REQNODE {
        struct REQNODE *prev, *next;
        REQUEST *req;
} REQNODE;

typedef struct REQUESTINFO {
        REQNODE *first_request;
        REQNODE *last_request;
        int request_count;
        time_t last_cleaned_list;
} REQUESTINFO;

static REQUESTINFO      request_list[256];

/*
 *      Remember the next request at which we start walking
 *      the list.
 */
static REQUEST *last_request = NULL;

/*
 *      It MAY make more sense here to key off of the packet ID, just
 *      like the request_list.  Then again, saving another 8 lookups
 *      (on average) isn't much of a problem.
 *
 *      The "request_cmp" function keys off of the packet ID first,
 *      so the first 8 layers of the tree will be the fanned-out
 *      tree for packet ID's.
 */
static rbtree_t         *request_tree;

#ifdef HAVE_PTHREAD_H
static pthread_mutex_t  proxy_mutex;
#else
/*
 *      This is easier than ifdef's throughout the code.
 */
#define pthread_mutex_lock(_x)
#define pthread_mutex_unlock(_x)
#endif

/*
 *      We keep track of packets we're proxying, keyed by
 *      source socket, and destination ip/port, and Id.
 */
static rbtree_t         *proxy_tree;

/*
 *      We keep track of free/used Id's, by destination ip/port.
 *
 *      We need a different tree than above, because this one is NOT
 *      keyed by Id.  Instead, we use this one to allocate Id's.
 */
static rbtree_t         *proxy_id_tree;

/*
 *      We keep the proxy FD's here.  The RADIUS Id's are marked
 *      "allocated" per Id, via a bit per proxy FD.
 */
static int              proxy_fds[32];

static uint32_t         proxy_ipaddr;

/*
 *      We can use 256 RADIUS Id's per dst ipaddr/port, per server
 *      socket.  So, to allocate them, we key off of dst ipaddr/port,
 *      and then search the RADIUS Id's, looking for an unused socket.
 *
 *      We do NOT key off of socket fd's, here, either.  Instead,
 *      we look for a free Id from a sockfd, any sockfd.
 */
typedef struct proxy_id_t {
        uint32_t        dst_ipaddr;
        int             dst_port;

        /*
         *      FIXME: Allocate more proxy sockets when this gets full.
         */
        int             index;
        uint32_t        mask;   /* of FD's we know about. */
        uint32_t        id[1];  /* really id[256] */
} proxy_id_t;


/*
 *      Find a matching entry in the proxy ID tree.
 */
static int proxy_id_cmp(const void *one, const void *two)
{
        const proxy_id_t *a = one;
        const proxy_id_t *b = two;

        /*
         *      The following comparisons look weird, but it's
         *      the only way to make the comparisons work.
         */
        if (a->dst_ipaddr < b->dst_ipaddr) return -1;
        if (a->dst_ipaddr > b->dst_ipaddr) return +1;

        if (a->dst_port < b->dst_port) return -1;
        if (a->dst_port > b->dst_port) return +1;
        
        /*
         *      Everything's equal.  Say so.
         */
        return 0;
}


/*
 *      Compare two REQUEST data structures, based on a number
 *      of criteria.
 */
static int request_cmp(const void *one, const void *two)
{
        const REQUEST *a = one;
        const REQUEST *b = two;

        /*
         *      The following comparisons look weird, but it's
         *      the only way to make the comparisons work.
         */

        /*
         *      If the packets didn't arrive on the same socket,
         *      they're not identical, no matter what their src/dst
         *      ip/ports say.
         */
        if (a->packet->sockfd < b->packet->sockfd) return -1;
        if (a->packet->sockfd > b->packet->sockfd) return +1;

        if (a->packet->id < b->packet->id) return -1;
        if (a->packet->id > b->packet->id) return +1;

        if (a->packet->code < b->packet->code) return -1;
        if (a->packet->code > b->packet->code) return +1;

        if (a->packet->src_ipaddr < b->packet->src_ipaddr) return -1;
        if (a->packet->src_ipaddr > b->packet->src_ipaddr) return +1;

        if (a->packet->src_port < b->packet->src_port) return -1;
        if (a->packet->src_port > b->packet->src_port) return +1;

        /*
         *      Hmm... we may be listening on IPADDR_ANY, in which case
         *      the destination IP is important, too.
         */
        if (a->packet->dst_ipaddr < b->packet->dst_ipaddr) return -1;
        if (a->packet->dst_ipaddr > b->packet->dst_ipaddr) return +1;

        if (a->packet->dst_port < b->packet->dst_port) return -1;
        if (a->packet->dst_port > b->packet->dst_port) return +1;

        /*
         *      Everything's equal.  Say so.
         */
        return 0;
}

/*
 *      Compare two REQUEST data structures, based on a number
 *      of criteria, for proxied packets.
 */
static int proxy_cmp(const void *one, const void *two)
{
        const REQUEST *a = one;
        const REQUEST *b = two;

        rad_assert(a->magic == REQUEST_MAGIC);
        rad_assert(b->magic == REQUEST_MAGIC);

        rad_assert(a->proxy != NULL);
        rad_assert(b->proxy != NULL);

        /*
         *      The following code looks unreasonable, but it's
         *      the only way to make the comparisons work.
         */
        if (a->proxy->sockfd < b->proxy->sockfd) return -1;
        if (a->proxy->sockfd > b->proxy->sockfd) return +1;

        if (a->proxy->id < b->proxy->id) return -1;
        if (a->proxy->id > b->proxy->id) return +1;

        /*
         *      We've got to check packet codes, too.  But
         *      this should be done later, by someone else...
         */

        if (a->proxy->dst_ipaddr < b->proxy->dst_ipaddr) return -1;
        if (a->proxy->dst_ipaddr > b->proxy->dst_ipaddr) return +1;

        if (a->proxy->dst_port < b->proxy->dst_port) return -1;
        if (a->proxy->dst_port > b->proxy->dst_port) return +1;

        /*
         *      FIXME: Check the Proxy-State attribute, too.
         *      This will help cut down on duplicates.
         */

        /*
         *      Everything's equal.  Say so.
         */
        return 0;
}

void rl_free(void)
{
        rbtree_free(request_tree);
        rbtree_free(proxy_id_tree);
        rbtree_free(proxy_tree);
}


/*
 *      Initialize the request list.
 */
int rl_init(void)
{
        /*
         *      Initialize the request_list[] array.
         */
        memset(request_list, 0, sizeof(request_list));

        request_tree = rbtree_create(request_cmp, NULL, 0);
        if (!request_tree) {
                rad_assert("FAIL" == NULL);
        }

        /*
         *      Create the tree for managing proxied requests and
         *      responses.
         */
        proxy_tree = rbtree_create(proxy_cmp, NULL, 1);
        if (!proxy_tree) {
                rad_assert("FAIL" == NULL);
        }

        /*
         *      Create the tree for allocating proxy ID's.
         */
        proxy_id_tree = rbtree_create(proxy_id_cmp, NULL, 0);
        if (!proxy_id_tree) {
                rad_assert("FAIL" == NULL);
        }

#ifdef HAVE_PTHREAD_H
        /*
         *      For now, always create the mutex.
         *
         *      Later, we can only create it if there are multiple threads.
         */
        if (pthread_mutex_init(&proxy_mutex, NULL) != 0) {
                radlog(L_ERR, "FATAL: Failed to initialize proxy mutex: %s",
                       strerror(errno));
                exit(1);
        }
#endif

        /*
         *      The Id allocation table is done by bits, so we have
         *      32 bits per Id.  These bits indicate which entry
         *      in the proxy_fds array is used for that Id.
         *
         *      This design allows 256*32 = 8k requests to be
         *      outstanding to a home server, before something goes
         *      wrong.
         */
        {
                int i;
                rad_listen_t *listener;

                /*
                 *      Mark the Fd's as unused.
                 */
                for (i = 0; i < 32; i++) proxy_fds[i] = -1;

                for (listener = mainconfig.listen;
                     listener != NULL;
                     listener = listener->next) {
                        if (listener->type == RAD_LISTEN_PROXY) {
                                proxy_ipaddr = listener->ipaddr;
                                proxy_fds[listener->fd & 0x1f] = listener->fd;
                                break;
                        }
                }
        }

        return 1;
}


/*
 *      Delete a request from the proxy trees.
 */
static void rl_delete_proxy(REQUEST *request, rbnode_t *node)
{
        proxy_id_t      myid, *entry;

        rad_assert(node != NULL);

        rbtree_delete(proxy_tree, node);
        
        myid.dst_ipaddr = request->proxy->dst_ipaddr;
        myid.dst_port = request->proxy->dst_port;

        /*
         *      Find the Id in the array of allocated Id's,
         *      and delete it.
         */
        entry = rbtree_finddata(proxy_id_tree, &myid);
        if (entry) {
                int i;

                DEBUG3(" proxy: de-allocating %08x:%d %d",
                       entry->dst_ipaddr,
                       entry->dst_port,
                       request->proxy->id);

                /*
                 *      Find the proxy socket associated with this
                 *      Id.  We loop over all 32 proxy fd's, but we
                 *      partially index by proxy fd's, which means
                 *      that we almost always break out of the loop
                 *      quickly.
                 */
                for (i = 0; i < 32; i++) {
                        int offset;

                        offset = (request->proxy->sockfd + i) & 0x1f;
                  
                        if (proxy_fds[offset] == request->proxy->sockfd) {
                                
                                entry->id[request->proxy->id] &= ~(1 << offset);
                                break;
                        }
                } /* else die horribly? */
        } else {
                /*
                 *      Hmm... not sure what to do here.
                 */
                DEBUG3(" proxy: FAILED TO FIND %08x:%d %d",
                       myid.dst_ipaddr,
                       myid.dst_port,
                       request->proxy->id);
        }
}


/*
 *      Delete a particular request.
 */
void rl_delete(REQUEST *request)
{
        int id;
        REQNODE *prev, *next;

        prev = ((REQNODE *) request->container)->prev;
        next = ((REQNODE *) request->container)->next;

        id = request->packet->id;

        /*
         *      Update the last request we touched.
         *
         *      This is so the periodic "walk & clean list"
         *      function, below, doesn't walk over all requests
         *      all of the time.  Rather, it tries to amortize
         *      the cost...
         */
        if (last_request == request) {
                last_request = rl_next(last_request);
        }


        if (prev == NULL) {
                request_list[id].first_request = next;
        } else {
                prev->next = next;
        }

        if (next == NULL) {
                request_list[id].last_request = prev;
        } else {
                next->prev = prev;
        }

        free(request->container);

#ifdef WITH_SNMP
        /*
         *      Update the SNMP statistics.
         *
         *      Note that we do NOT do this in rad_respond(),
         *      as that function is called from child threads.
         *      Instead, we update the stats when a request is
         *      deleted, because only the main server thread calls
         *      this function...
         */
        if (mainconfig.do_snmp) {
                switch (request->reply->code) {
                case PW_AUTHENTICATION_ACK:
                  rad_snmp.auth.total_responses++;
                  rad_snmp.auth.total_access_accepts++;
                  break;

                case PW_AUTHENTICATION_REJECT:
                  rad_snmp.auth.total_responses++;
                  rad_snmp.auth.total_access_rejects++;
                  break;

                case PW_ACCESS_CHALLENGE:
                  rad_snmp.auth.total_responses++;
                  rad_snmp.auth.total_access_challenges++;
                  break;

                case PW_ACCOUNTING_RESPONSE:
                  rad_snmp.acct.total_responses++;
                  break;

                default:
                        break;
                }
        }
#endif

        /*
         *      Delete the request from the tree.
         */
        {
                rbnode_t *node;

                node = rbtree_find(request_tree, request);
                rad_assert(node != NULL);
                rbtree_delete(request_tree, node);


                /*
                 *      If there's a proxied packet, and we're still
                 *      waiting for a reply, then delete the packet
                 *      from the list of outstanding proxied requests.
                 */
                if (request->proxy &&
                    (request->proxy_outstanding > 0)) {
                        pthread_mutex_lock(&proxy_mutex);
                        node = rbtree_find(proxy_tree, request);
                        rl_delete_proxy(request, node);
                        pthread_mutex_unlock(&proxy_mutex);
                }
        }

        request_free(&request);
        request_list[id].request_count--;

}

/*
 *      Add a request to the request list.
 */
void rl_add(REQUEST *request)
{
        int id = request->packet->id;
        REQNODE *node;

        rad_assert(request->container == NULL);

        request->container = rad_malloc(sizeof(REQNODE));
        node = (REQNODE *) request->container;
        node->req = request;

        node->prev = NULL;
        node->next = NULL;

        if (!request_list[id].first_request) {
                rad_assert(request_list[id].request_count == 0);

                request_list[id].first_request = node;
                request_list[id].last_request = node;
        } else {
                rad_assert(request_list[id].request_count != 0);

                node->prev = request_list[id].last_request;
                request_list[id].last_request->next = node;
                request_list[id].last_request = node;
        }

        /*
         *      Insert the request into the tree.
         */
        if (rbtree_insert(request_tree, request) == 0) {
                rad_assert("FAIL" == NULL);
        }

        request_list[id].request_count++;
}

/*
 *      Look up a particular request, using:
 *
 *      Request ID, request code, source IP, source port,
 *
 *      Note that we do NOT use the request vector to look up requests.
 *
 *      We MUST NOT have two requests with identical (id/code/IP/port), and
 *      different vectors.  This is a serious error!
 */
REQUEST *rl_find(RADIUS_PACKET *packet)
{
        REQUEST myrequest;

        myrequest.packet = packet;

        return rbtree_finddata(request_tree, &myrequest);
}

/*
 *      See mainconfig.c
 */
extern int proxy_new_listener(void);

/*
 *      Add an entry to the proxy tree.
 *
 *      This is the ONLY function in this source file which may be called
 *      from a child thread.  It therefore needs mutexes...
 */
int rl_add_proxy(REQUEST *request)
{
        int             i, found, proxy;
        uint32_t        mask;
        proxy_id_t      myid, *entry;

        myid.dst_ipaddr = request->proxy->dst_ipaddr;
        myid.dst_port = request->proxy->dst_port;

        /*
         *      Proxied requests get sent out the proxy FD ONLY.
         *
         *      FIXME: Once we allocate multiple proxy FD's, move this
         *      code to below, so we can have more than 256 requests
         *      outstanding.
         */
        request->proxy_outstanding = 1;

        pthread_mutex_lock(&proxy_mutex);

        /*
         *      Assign a proxy ID.
         */
        entry = rbtree_finddata(proxy_id_tree, &myid);
        if (!entry) {   /* allocate it */
                entry = rad_malloc(sizeof(*entry) + sizeof(entry->id) * 255);
                
                entry->dst_ipaddr = request->proxy->dst_ipaddr;
                entry->dst_port = request->proxy->dst_port;
                entry->index = 0;

                DEBUG3(" proxy: creating %08x:%d",
                       entry->dst_ipaddr,
                       entry->dst_port);
                
                /*
                 *      Insert the new home server entry into
                 *      the tree.
                 *
                 *      FIXME: We don't (currently) delete the
                 *      entries, so this is technically a
                 *      memory leak.
                 */
                if (rbtree_insert(proxy_id_tree, entry) == 0) {
                        DEBUG2("ERROR: Failed to insert entry into proxy Id tree");
                        free(entry);
                        return 0;
                }

                /*
                 *      Clear out bits in the array which DO have
                 *      proxy Fd's associated with them.  We do this
                 *      by getting the mask of bits which have proxy
                 *      fd's...  */
                mask = 0;
                for (i = 0; i < 32; i++) {
                        if (proxy_fds[i] != -1) {
                                mask |= (1 << i);
                        }
                }
                rad_assert(mask != 0);

                /*
                 *      Set bits here indicate that the Fd is in use.
                 */
                entry->mask = mask;

                mask = ~mask;

                /*
                 *      Set the bits which are unused (and therefore
                 *      allocated).  The clear bits indicate that the Id
                 *      for that FD is unused.
                 */
                for (i = 0; i < 256; i++) {
                        entry->id[i] = mask;
                }
        } /* else the entry already existed in the proxy Id tree */
        
 retry:
        /*
         *      Try to find a free Id.
         */
        found = -1;
        for (i = 0; i < 256; i++) {
                /*
                 *      Some bits are still zero..
                 */
                if (entry->id[(i + entry->index) & 0xff] != (uint32_t) ~0) {
                        found = (i + entry->index) & 0xff;
                        break;
                }

                /*
                 *      Hmm... do we want to re-use Id's, when we
                 *      haven't seen all of the responses?
                 */
        }
        
        /*
         *      No free Id, try to get a new FD.
         */
        if (found < 0) {
                /*
                 *      First, see if there were FD's recently allocated,
                 *      which we don't know about.
                 */
                mask = 0;
                for (i = 0; i < 32; i++) {
                        if (proxy_fds[i] < 0) continue;

                        mask |= (1 << i);
                }

                /*
                 *      There ARE more FD's than we know about.
                 *      Update the masks for Id's, and re-try.
                 */
                if (entry->mask != mask) {
                        /*
                         *      New mask always has more bits than
                         *      the old one, but never fewer bits.
                         */
                        rad_assert((entry->mask & mask) == entry->mask);

                        /*
                         *      Clear the bits we already know about,
                         *      and then or in those bits into the
                         *      global mask.
                         */
                        mask ^= entry->mask;
                        entry->mask |= mask;
                        mask = ~mask;
                        
                        /*
                         *      Clear the bits in the Id's for the new
                         *      FD's.
                         */
                        for (i = 0; i < 256; i++) {
                                entry->id[i] &= mask;
                        }
                        
                        /*
                         *      And try again to allocate an Id.
                         */
                        goto retry;
                } /* else no new Fd's were allocated. */

                /*
                 *      If all Fd's are allocated, die.
                 */
                if (~mask == 0) {
                        radlog(L_ERR|L_CONS, "ERROR: More than 8000 proxied requests outstanding for home server %08x:%d",
                               ntohs(entry->dst_ipaddr), entry->dst_port);
                        return 0;
                }
                
                /*
                 *      Allocate a new proxy Fd.  This function adds it
                 *      into the list of listeners.
                 */
                proxy = proxy_new_listener();
                if (proxy < 0) {
                        DEBUG2("ERROR: Failed to create a new socket for proxying requests.");
                        return 0;
                }

                /*
                 *
                 */
                found = -1;
                for (i = 0; i < 32; i++) {
                        /*
                         *      Found a free entry.  Save the socket,
                         *      and remember where we saved it.
                         */
                        if (proxy_fds[(proxy + i) & 0x1f] == -1) {
                                proxy_fds[(proxy + i) & 0x1f] = proxy;
                                found = (proxy + i) & 0x1f;
                                break;
                        }
                }
                rad_assert(found >= 0); /* i.e. the mask had free bits. */

                mask = 1 << found;
                entry->mask |= mask;
                mask = ~mask;

                /*
                 *      Clear the relevant bits in the mask.
                 */
                for (i = 0; i < 256; i++) {
                        entry->id[i] &= mask;
                }

                /*
                 *      Pick a random Id to start from, as we've
                 *      just guaranteed that it's free.
                 */
                found = lrad_rand() & 0xff;
        }
        
        /*
         *      Mark next (hopefully unused) entry.
         */
        entry->index = (found + 1) & 0xff;
        
        /*
         *      We now have to find WHICH proxy fd to use.
         */
        proxy = -1;
        for (i = 0; i < 32; i++) {
                /*
                 *      FIXME: pick a random socket to use?
                 */
                if ((entry->id[found] & (1 << i)) == 0) {
                        proxy = i;
                        break;
                }
        }

        /*
         *      There was no bit clear, which we had just checked above...
         */
        rad_assert(proxy != -1);

        /*
         *      Mark the Id as allocated, for thei Fd.
         */
        entry->id[found] |= (1 << proxy);
        request->proxy->id = found;
        request->proxy->src_ipaddr = proxy_ipaddr;

        rad_assert(proxy_fds[proxy] != -1);
        request->proxy->sockfd = proxy_fds[proxy];

        DEBUG3(" proxy: allocating %08x:%d %d",
               entry->dst_ipaddr,
               entry->dst_port,
               request->proxy->id);
        
        if (!rbtree_insert(proxy_tree, request)) {
                DEBUG2("ERROR: Failed to insert entry into proxy tree");
                return 0;
        }
        
        pthread_mutex_unlock(&proxy_mutex);

        return 1;
}


/*
 *      Look up a particular request, using:
 *
 *      Request Id, request code, source IP, source port,
 *
 *      Note that we do NOT use the request vector to look up requests.
 *
 *      We MUST NOT have two requests with identical (id/code/IP/port), and
 *      different vectors.  This is a serious error!
 */
REQUEST *rl_find_proxy(RADIUS_PACKET *packet)
{
        rbnode_t        *node;
        REQUEST         myrequest, *maybe = NULL;
        RADIUS_PACKET   myproxy;

        /*
         *      If we use the socket FD as an indicator,
         *      then that implicitely contains information
         *      as to our src ipaddr/port, so we don't need
         *      to use that in the comparisons.
         */
        myproxy.sockfd = packet->sockfd;
        myproxy.id = packet->id;
        myproxy.dst_ipaddr = packet->src_ipaddr;
        myproxy.dst_port = packet->src_port;

#ifndef NDEBUG
        myrequest.magic = REQUEST_MAGIC;
#endif
        myrequest.proxy = &myproxy;

        pthread_mutex_lock(&proxy_mutex);
        node = rbtree_find(proxy_tree, &myrequest);

        if (node) {
                maybe = rbtree_node2data(proxy_tree, node);
                rad_assert(maybe->proxy_outstanding > 0);
                maybe->proxy_outstanding--;
                
                /*
                 *      Received all of the replies we expect.
                 *      delete it from both trees.
                 */
                if (maybe->proxy_outstanding == 0) {
                        rl_delete_proxy(&myrequest, node);
                }
        }
        pthread_mutex_unlock(&proxy_mutex);

        return maybe;
}


/*
 *      Walk over all requests, performing a callback for each request.
 */
int rl_walk(RL_WALK_FUNC walker, void *data)
{
        int id, rcode;
        REQNODE *curreq, *next;

        /*
         *      Walk over all 256 ID's.
         */
        for (id = 0; id < 256; id++) {

                /*
                 *      Walk over the request list for each ID.
                 */
                for (curreq = request_list[id].first_request;
                                curreq != NULL ;
                                curreq = next) {
                        /*
                         *      The callback MIGHT delete the current
                         *      request, so we CANNOT depend on curreq->next
                         *      to be there, when going to the next element
                         *      in the 'for' loop.
                         */
                        next = curreq->next;

                        rcode = walker(curreq->req, data);
                        if (rcode != RL_WALK_CONTINUE) {
                                return rcode;
                        }
                }
        }

        return 0;
}


/*
 *      Walk from one request to the next.
 */
REQUEST *rl_next(REQUEST *request)
{
        int id, start_id;
        int count;

        /*
         *      If we were passed a request, then go to the "next" one.
         */
        if (request != NULL) {
                rad_assert(request->magic == REQUEST_MAGIC);

                /*
                 *      It has a "next", return it.
                 */
                if (((REQNODE *)request->container)->next != NULL) {
                        return ((REQNODE *)request->container)->next->req;
                } else {
                        /*
                         *      No "next", increment the ID, and look
                         *      at that one.
                         */
                        start_id = request->packet->id + 1;
                        start_id &= 0xff;
                        count = 255;
                }
        } else {
                /*
                 *      No input request, start looking at ID 0.
                 */
                start_id = 0;
                count = 256;
        }

        /*
         *      Check all ID's, wrapping around at 255.
         */
        for (id = start_id; id < (start_id + count); id++) {

                /*
                 *      This ID has a request, return it.
                 */
                if (request_list[id & 0xff].first_request != NULL) {
                        rad_assert(request_list[id&0xff].first_request->req != request);

                        return request_list[id & 0xff].first_request->req;
                }
        }

        /*
         *      No requests at all in the list. Nothing to do.
         */
        DEBUG3("rl_next:  returning NULL");
        return NULL;
}


/*
 *      Return the number of requests in the request list.
 */
int rl_num_requests(void)
{
        int id;
        int request_count = 0;

        for (id = 0; id < 256; id++) {
                request_count += request_list[id].request_count;
        }

        return request_count;
}


typedef struct rl_walk_t {
        time_t  now;
        time_t  smallest;
} rl_walk_t;


/*
 *  Refresh a request, by using proxy_retry_delay, cleanup_delay,
 *  max_request_time, etc.
 *
 *  When walking over the request list, all of the per-request
 *  magic is done here.
 */
static int refresh_request(REQUEST *request, void *data)
{
        rl_walk_t *info = (rl_walk_t *) data;
        time_t difference;
        child_pid_t child_pid;

        rad_assert(request->magic == REQUEST_MAGIC);

        /*
         *  If the request is marked as a delayed reject, AND it's
         *  time to send the reject, then do so now.
         */
        if (request->finished &&
            ((request->options & RAD_REQUEST_OPTION_DELAYED_REJECT) != 0)) {
                rad_assert(request->child_pid == NO_SUCH_CHILD_PID);

                difference = info->now - request->timestamp;
                if (difference >= (time_t) mainconfig.reject_delay) {

                        /*
                         *  Clear the 'delayed reject' bit, so that we
                         *  don't do this again.
                         */
                        request->options &= ~RAD_REQUEST_OPTION_DELAYED_REJECT;
                        rad_send(request->reply, request->packet,
                                 request->secret);
                }
        }

        /*
         *  If the request has finished processing, AND it's child has
         *  been cleaned up, AND it's time to clean up the request,
         *  OR, it's an accounting request.  THEN, go delete it.
         *
         *  If this is a request which had the "don't cache" option
         *  set, then delete it immediately, as it CANNOT have a
         *  duplicate.
         */
        if (request->finished &&
            ((request->timestamp + mainconfig.cleanup_delay <= info->now) ||
             ((request->options & RAD_REQUEST_OPTION_DONT_CACHE) != 0))) {
                rad_assert(request->child_pid == NO_SUCH_CHILD_PID);

                /*
                 *  Request completed, delete it, and unlink it
                 *  from the currently 'alive' list of requests.
                 */
                DEBUG2("Cleaning up request %d ID %d with timestamp %08lx",
                                request->number, request->packet->id,
                                (unsigned long) request->timestamp);

                /*
                 *  Delete the request.
                 */
                rl_delete(request);
                return RL_WALK_CONTINUE;
        }

        /*
         *  Maybe the child process handling the request has hung:
         *  kill it, and continue.
         */
        if ((request->timestamp + mainconfig.max_request_time) <= info->now) {
                int number;

                child_pid = request->child_pid;
                number = request->number;

                /*
                 *      There MUST be a RAD_PACKET reply.
                 */
                rad_assert(request->reply != NULL);

                /*
                 *      If we've tried to proxy the request, and
                 *      the proxy server hasn't responded, then
                 *      we send a REJECT back to the caller.
                 *
                 *      For safety, we assert that there is no child
                 *      handling the request.  If the assertion fails,
                 *      it means that we've sent a proxied request to
                 *      the home server, and the child thread is still
                 *      sitting on the request!
                 */
                if (request->proxy && !request->proxy_reply) {
                        rad_assert(request->child_pid == NO_SUCH_CHILD_PID);

                        radlog(L_ERR, "Rejecting request %d due to lack of any response from home server %s:%d",
                               request->number,
                               client_name(request->packet->src_ipaddr),
                               request->packet->src_port);
                        request_reject(request);
                        request->finished = TRUE;
                        return RL_WALK_CONTINUE;
                }

                if (mainconfig.kill_unresponsive_children) {
                        if (child_pid != NO_SUCH_CHILD_PID) {
                                /*
                                 *  This request seems to have hung
                                 *   - kill it
                                 */
#ifdef HAVE_PTHREAD_H
                                radlog(L_ERR, "Killing unresponsive thread for request %d",
                                       request->number);
                                pthread_cancel(child_pid);
#endif
                        } /* else no proxy reply, quietly fail */

                        /*
                         *      Maybe we haven't killed it.  In that
                         *      case, print a warning.
                         */
                } else if ((child_pid != NO_SUCH_CHILD_PID) &&
                           ((request->options & RAD_REQUEST_OPTION_LOGGED_CHILD) == 0)) {
                        radlog(L_ERR, "WARNING: Unresponsive child (id %lu) for request %d",
                               (unsigned long)child_pid, number);

                        /*
                         *  Set the option that we've sent a log message,
                         *  so that we don't send more than one message
                         *  per request.
                         */
                        request->options |= RAD_REQUEST_OPTION_LOGGED_CHILD;
                }

                /*
                 *  Send a reject message for the request, mark it
                 *  finished, and forget about the child.
                 */
                request_reject(request);
                request->child_pid = NO_SUCH_CHILD_PID;
                if (mainconfig.kill_unresponsive_children)
                        request->finished = TRUE;
                return RL_WALK_CONTINUE;
        } /* the request has been in the queue for too long */

        /*
         *  If the request is still being processed, then due to the
         *  above check, it's still within it's time limit.  In that
         *  case, don't do anything.
         */
        if (request->child_pid != NO_SUCH_CHILD_PID) {
                return RL_WALK_CONTINUE;
        }

        /*
         *  The request is finished.
         */
        if (request->finished) goto setup_timeout;

        /*
         *  We're not proxying requests at all.
         */
        if (!mainconfig.proxy_requests) goto setup_timeout;

        /*
         *  We're proxying synchronously, so we don't retry it here.
         *  Some other code takes care of retrying the proxy requests.
         */
        if (mainconfig.proxy_synchronous) goto setup_timeout;

        /*
         *  The proxy retry delay is zero, meaning don't retry.
         */
        if (mainconfig.proxy_retry_delay == 0) goto setup_timeout;

        /*
         *  There is no proxied request for this packet, so there's
         *  no proxy retries.
         */
        if (!request->proxy) goto setup_timeout;

        /*
         *  We've already seen the proxy reply, so we don't need
         *  to send another proxy request.
         */
        if (request->proxy_reply) goto setup_timeout;

        /*
         *  It's not yet time to re-send this proxied request.
         */
        if (request->proxy_next_try > info->now) goto setup_timeout;

        /*
         *  If the proxy retry count is zero, then
         *  we've sent the last try, and have NOT received
         *  a reply from the end server.  In that case,
         *  we don't bother trying again, but just mark
         *  the request as finished, and go to the next one.
         */
        if (request->proxy_try_count == 0) {
                rad_assert(request->child_pid == NO_SUCH_CHILD_PID);
                request_reject(request);
                realm_disable(request->proxy->dst_ipaddr,request->proxy->dst_port);
                request->finished = TRUE;
                goto setup_timeout;
        }

        /*
         *  We're trying one more time, so count down
         *  the tries, and set the next try time.
         */
        request->proxy_try_count--;
        request->proxy_next_try = info->now + mainconfig.proxy_retry_delay;

        /* Fix up Acct-Delay-Time */
        if (request->proxy->code == PW_ACCOUNTING_REQUEST) {
                VALUE_PAIR *delaypair;
                delaypair = pairfind(request->proxy->vps, PW_ACCT_DELAY_TIME);

                if (!delaypair) {
                        delaypair = paircreate(PW_ACCT_DELAY_TIME, PW_TYPE_INTEGER);
                        if (!delaypair) {
                                radlog(L_ERR|L_CONS, "no memory");
                                exit(1);
                        }
                        pairadd(&request->proxy->vps, delaypair);
                }
                delaypair->lvalue = info->now - request->proxy->timestamp;

                /* Must recompile the valuepairs to wire format */
                free(request->proxy->data);
                request->proxy->data = NULL;
        } /* proxy accounting request */

        /*
         *  Assert that we have NOT seen the proxy reply yet.
         *
         *  If we HAVE seen it, then we SHOULD NOT be bugging the
         *  home server!
         */
        rad_assert(request->proxy_reply == NULL);

        /*
         *  Send the proxy packet.
         */
        request->proxy_outstanding++;
        rad_send(request->proxy, NULL, request->proxysecret);

setup_timeout:
        /*
         *  Don't do more long-term checks, if we've got to wake
         *  up now.
         */
        if (info->smallest == 0) {
                return RL_WALK_CONTINUE;
        }

        /*
         *  The request is finished.  Wake up when it's time to
         *  clean it up.
         */
        if (request->finished) {
                difference = (request->timestamp + mainconfig.cleanup_delay) - info->now;

                /*
                 *  If the request is marked up to be rejected later,
                 *  then wake up later.
                 */
                if ((request->options & RAD_REQUEST_OPTION_DELAYED_REJECT) != 0) {
                        if (difference >= (time_t) mainconfig.reject_delay) {
                                difference = (time_t) mainconfig.reject_delay;
                        }
                }

        } else if (request->proxy && !request->proxy_reply) {
                /*
                 *  The request is NOT finished, but there is an
                 *  outstanding proxy request, with no matching
                 *  proxy reply.
                 *
                 *  Wake up when it's time to re-send
                 *  the proxy request.
                 *
                 *  But in synchronous proxy, we don't retry but we update
                 *  the next retry time as NAS has not resent the request
                 *  in the given retry window.
                 */
                if (mainconfig.proxy_synchronous) {
                        /*
                         *      If the retry_delay * count has passed,
                         *      then mark the realm dead.
                         */
                        if (info->now > (request->timestamp + (mainconfig.proxy_retry_delay * mainconfig.proxy_retry_count))) {
                                rad_assert(request->child_pid == NO_SUCH_CHILD_PID);
                                request_reject(request);
                                
                                realm_disable(request->proxy->dst_ipaddr,
                                              request->proxy->dst_port);
                                request->finished = TRUE;
                                goto setup_timeout;
                        }
                        request->proxy_next_try = info->now + mainconfig.proxy_retry_delay;
                }
                difference = request->proxy_next_try - info->now;
        } else {
                /*
                 *  The request is NOT finished.
                 *
                 *  Wake up when it's time to kill the errant
                 *  thread/process.
                 */
                difference = (request->timestamp + mainconfig.max_request_time) - info->now;
        }

        /*
         *  If the server is CPU starved, then we CAN miss a time
         *  for servicing requests.  In which case the 'difference'
         *  value will be negative.  select() doesn't like that,
         *  so we fix it.
         */
        if (difference < 0) {
                difference = 0;
        }

        /*
         *  Update the 'smallest' time.
         */
        if ((info->smallest < 0) ||
                (difference < info->smallest)) {
                info->smallest = difference;
        }

        return RL_WALK_CONTINUE;
}


/*
 *  Clean up the request list, every so often.
 *
 *  This is done by walking through ALL of the list, and
 *  - marking any requests which are finished, and expired
 *  - killing any processes which are NOT finished after a delay
 *  - deleting any marked requests.
 */
struct timeval *rl_clean_list(time_t now)
{
        /*
         *  Static variables, so that we don't do all of this work
         *  more than once per second.
         *
         *  Note that we have 'tv' and 'last_tv'.  'last_tv' is
         *  pointed to by 'last_tv_ptr', and depending on the
         *  system implementation of select(), it MAY be modified.
         *
         *  In that was, we want to use the ORIGINAL value, from
         *  'tv', and wipe out the (possibly modified) last_tv.
         */
        static time_t last_cleaned_list = 0;
        static struct timeval tv, *last_tv_ptr = NULL;
        static struct timeval last_tv;

        rl_walk_t info;

        info.now = now;
        info.smallest = -1;

        /*
         *  If we've already set up the timeout or cleaned the
         *  request list this second, then don't do it again.  We
         *  simply return the sleep delay from last time.
         *
         *  Note that if we returned NULL last time, there was nothing
         *  to do.  BUT we've been woken up since then, which can only
         *  happen if we received a packet.  And if we've received a
         *  packet, then there's some work to do in the future.
         *
         *  FIXME: We can probably use gettimeofday() for finer clock
         *  resolution, as the current method will cause it to sleep
         *  too long...
         */
        if ((last_tv_ptr != NULL) &&
            (last_cleaned_list == now) &&
            (tv.tv_sec != 0)) {
                int i;

                /*
                 *  If we're NOT walking the entire request list,
                 *  then we want to iteratively check the request
                 *  list.
                 *
                 *  If there is NO previous request, go look for one.
                 */
                if (!last_request)
                        last_request = rl_next(last_request);

                /*
                 *  On average, there will be one request per
                 *  'cleanup_delay' requests, which needs to be
                 *  serviced.
                 *
                 *  And only do this servicing, if we have a request
                 *  to service.
                 */
                if (last_request)
                        for (i = 0; i < mainconfig.cleanup_delay; i++) {
                                REQUEST *next;

                                /*
                                 *  This function call MAY delete the
                                 *  request pointed to by 'last_request'.
                                 */
                                next = rl_next(last_request);
                                refresh_request(last_request, &info);
                                last_request = next;

                                /*
                                 *  Nothing to do any more, exit.
                                 */
                                if (!last_request)
                                        break;
                        }

                last_tv = tv;
                DEBUG2("Waking up in %d seconds...",
                                (int) last_tv_ptr->tv_sec);
                return last_tv_ptr;
        }
        last_cleaned_list = now;
        last_request = NULL;
        DEBUG2("--- Walking the entire request list ---");

        /*
         *  Hmmm... this is Big Magic.  We make it seem like
         *  there's an additional second to wait, for a whole
         *  host of reasons which I can't explain adequately,
         *  but which cause the code to Just Work Right.
         */
        info.now--;

        rl_walk(refresh_request, &info);

        /*
         *  We haven't found a time at which we need to wake up.
         *  Return NULL, so that the select() call will sleep forever.
         */
        if (info.smallest < 0) {
                /*
                 *  If we're not proxying, then there really isn't anything
                 *  to do.
                 *
                 *  If we ARE proxying, then we can safely sleep
                 *  forever if we're told to NEVER send proxy retries
                 *  ourselves, until the NAS kicks us again.
                 *
                 *  Otherwise, there are no outstanding requests, then
                 *  we can sleep forever.  This happens when we get
                 *  woken up with a bad packet.  It's discarded, so if
                 *  there are no live requests, we can safely sleep
                 *  forever.
                 */
                if ((!mainconfig.proxy_requests) ||
                    mainconfig.proxy_synchronous ||
                    (rl_num_requests() == 0)) {
                        DEBUG2("Nothing to do.  Sleeping until we see a request.");
                        last_tv_ptr = NULL;
                        return NULL;
                }

                /*
                 *  We ARE proxying.  In that case, we avoid a race condition
                 *  where a child thread handling a request proxies the
                 *  packet, and sets the retry delay.  In that case, we're
                 *  supposed to wake up in N seconds, but we can't, as
                 *  we're sleeping forever.
                 *
                 *  Instead, we prevent the problem by waking up anyhow
                 *  at the 'proxy_retry_delay' time, even if there's
                 *  nothing to do.  In the worst case, this will cause
                 *  the server to wake up every N seconds, to do a small
                 *  amount of unnecessary work.
                 */
                info.smallest = mainconfig.proxy_retry_delay;
        }
        /*
         *  Set the time (in seconds) for how long we're
         *  supposed to sleep.
         */
        tv.tv_sec = info.smallest;
        tv.tv_usec = 0;
        DEBUG2("Waking up in %d seconds...", (int) info.smallest);

        /*
         *  Remember how long we should sleep for.
         */
        last_tv = tv;
        last_tv_ptr = &last_tv;
        return last_tv_ptr;
}