tor-parallel-relay-conn/src/or/eventdns.c

/* Async DNS Library
 * Adam Langley <agl@imperialviolet.org>
 * http://www.imperialviolet.org/eventdns.html
 * Public Domain codenext
 *
 * This software is Public Domain. To view a copy of the public domain dedication,
 * visit http://creativecommons.org/licenses/publicdomain/ or send a letter to
 * Creative Commons, 559 Nathan Abbott Way, Stanford, California 94305, USA.
 *
 * I ask and expect, but do not require, that all derivative works contain an
 * attribution similar to:
 * 	Parts developed by Adam Langley <agl@imperialviolet.org>
 *
 * You may wish to replace the word "Parts" with something else depending on
 * the amount of original code.
 *
 * (Derivative works does not include programs which link against, run or include
 * the source verbatim in their source distributions)
 *
 * Version: 0.1b
 *
 *
 * Welcome, gentle reader
 *
 * Async DNS lookups are really a whole lot harder than they should be,
 * mostly stemming from the fact that the libc resolver has never been
 * very good at them. Before you use this library you should see if libc
 * can do the job for you with the modern async call getaddrinfo_r
 * (Google for it). Otherwise, please continue.
 *
 * This code is based on libevent and you must call event_init before
 * any of the APIs in this file. You must also seed the OpenSSL random
 * source if you are using OpenSSL for ids (see below).
 *
 * This library is designed to be included and shipped with your source
 * code. You statically link with it. You should also test for the
 * existance of strtok_r and define HAVE_STRTOK_R if you have it.
 *
 * The DNS protocol requires a good source of id numbers and these
 * numbers should be unpredictable for spoofing reasons. There are
 * three methods for generating them here and you must define exactly
 * one of them. In increasing order of preference:
 *
 * DNS_USE_GETTIMEOFDAY_FOR_ID:
 *   Using the bottom 16 bits of the usec result from gettimeofday. This
 *   is a pretty poor solution but should work anywhere
 * DNS_USE_CPU_CLOCK_FOR_ID:
 *   Using the bottom 16 bits of the nsec result from the CPU's time
 *   counter. This is better, but may not work everywhere. Requires
 *   POSIX realtime support and you'll need to link against -lrt on
 *   glibc systems at least
 * DNS_USE_OPENSSL_FOR_ID:
 *   Uses the OpenSSL RAND_bytes call to generate the data. You must
 *   have seeded the pool before making any calls to this library.
 *
 * The library keeps track of the state of nameservers and will avoid
 * them when they go down. Otherwise it will round robin between them.
 *
 * Quick start guide:
 *   #include "eventdns.h"
 *   void callback(int result, char type, int count, int ttl,
 *		 void *addresses, void *arg);
 *   eventdns_resolv_conf_parse(DNS_OPTIONS_ALL, "/etc/resolv.conf");
 *   eventdns_resolve("www.hostname.com", 0, callback, NULL);
 *
 * When the lookup is complete the callback function is called. The
 * first argument will be one of the DNS_ERR_* defines in eventdns.h.
 * Hopefully it will be DNS_ERR_NONE, in which case type will be
 * DNS_IPv4_A, count will be the number of IP addresses, ttl is the time
 * which the data can be cached for (in seconds), addresses will point
 * to an array of uint32_t's and arg will be whatever you passed to
 * eventdns_resolve.
 *
 * Searching:
 *
 * In order for this library to be a good replacment for glibc's resolver it
 * supports searching. This involves setting a list of default domains, in
 * which names will be queried for. The number of nots in the query name
 * determines the order in which this list is used.
 *
 * Searching appears to be a single lookup from the point of view of the API,
 * although many DNS queries may be generated from a single call to
 * eventdns_resolve. Searching can also drastically slow down the resolution of
 * names.
 *
 * To disable searching:
 *   1. Never set it up. If you never call eventdns_resolv_conf_parse or
 *   eventdns_search_add then no searching will occur.
 *
 *   2. If you do call eventdns_resolv_conf_parse then don't pass
 *   DNS_OPTION_SEARCH (or DNS_OPTIONS_ALL, which implies it)
 *
 *   3. When calling eventdns_resolve, pass the DNS_QUERY_NO_SEARCH flag
 *
 * The order of searches depends on the number of dots in the name. If the
 * number is greater than the ndots setting then the names is first tried
 * globally. Otherwise each search domain is appended in turn.
 *
 * The ndots setting can either be set from a resolv.conf, or by calling
 * eventdns_search_ndots_set.
 *
 * For example, with ndots set to 1 (the default) and a search domain list of
 * ["myhome.net"]:
 *  Query: www
 *  Order: www.myhome.net, www.
 *
 *  Query: www.abc
 *  Order: www.abc., www.abc.myhome.net
 *
 * API reference:
 *
 * int eventdns_nameserver_add(unsigned long int addresss)
 *   Add a nameserver. The address should be an IP address in
 *   network byte order. The type of address is chosen so that
 *   it match in_addr.s_addr.
 *   Returns non-zero on error.
 *
 * int eventdns_nameserer_ip_add(const char *ip_as_string)
 *   This wraps the above function by parsing a string as an IP
 *   address and add it as a nameserver.
 *   Returns non-zero on error
 *
 * int eventdns_resolve(const char *name, int flags,
 *		      eventdns_callback_type callback,
 *		      void *ptr)
 *   Resolve a name. The name parameter should be a DNS name.
 *   The flags parameter should be 0, or DNS_QUERY_NO_SEARCH
 *   which disables searching for this query. (see defn of
 *   searching above).
 *
 *   The callback argument is a function which is called when
 *   this query completes and ptr is an argument which is passed
 *   to that callback function.
 *
 *   Returns non-zero on error
 *
 * void eventdns_search_clear()
 *   Clears the list of search domains
 *
 * void eventdns_search_add(const char *domain)
 *   Add a domain to the list of search domains
 *
 * void eventdns_search_ndots_set(int ndots)
 *   Set the number of dots which, when found in a name, causes
 *   the first query to be without any search domain.
 *
 * int eventdns_resolv_conf_parse(int flags, const char *filename)
 *   Parse a resolv.conf like file from the given filename.
 *
 *   See the manpage for resolv.conf for the format of this file.
 *   The flags argument determines what information is parsed from
 *   this file:
 *     DNS_OPTION_SEARCH - domain, search and ndots options
 *     DNS_OPTION_NAMESERVERS - nameserver lines
 *     DNS_OPTION_MISC - timeout and attempts options
 *     DNS_OPTIONS_ALL - all of the above
 *   The following directives are not parsed from the file:
 *     sortlist, rotate, no-check-names, inet6, debug
 *
 *   Returns non-zero on error:
 *    0 no errors
 *    1 failed to open file
 *    2 failed to stat file
 *    3 file too large
 *    4 out of memory
 *    5 short read from file
 *
 * Internals:
 *
 * Requests are kept in two queues. The first is the inflight queue. In
 * this queue requests have an allocated transaction id and nameserver.
 * They will soon be transmitted if they haven't already been.
 *
 * The second is the waiting queue. The size of the inflight ring is
 * limited and all other requests wait in waiting queue for space. This
 * bounds the number of concurrent requests so that we don't flood the
 * nameserver. Several algorithms require a full walk of the inflight
 * queue and so bounding its size keeps thing going nicly under huge
 * (many thousands of requests) loads.
 *
 * If a nameserver looses too many requests it is considered down and we
 * try not to use it. After a while we send a probe to that nameserver
 * (a lookup for google.com) and, if it replies, we consider it working
 * again. If the nameserver fails a probe we wait longer to try again
 * with the next probe.
 */

#include "eventdns.h"
#include "eventdns_tor.h"
//#define NDEBUG

#ifndef DNS_USE_CPU_CLOCK_FOR_ID
#ifndef DNS_USE_GETTIMEOFDAY_FOR_ID
#ifndef DNS_USE_OPENSSL_FOR_ID
#error Must configure at least one id generation method.
#error Please see the documentation
#endif
#endif
#endif

// #define _POSIX_C_SOURCE 200507
#define _GNU_SOURCE

#ifdef DNS_USE_CPU_CLOCK_FOR_ID
#ifdef DNS_USE_OPENSSL_FOR_ID
#error Multiple id options selected
#endif
#ifdef DNS_USE_GETTIMEOFDAY_FOR_ID
#error Multiple id options selected
#endif
#include <time.h>
#endif

#ifdef DNS_USE_OPENSSL_FOR_ID
#ifdef DNS_USE_GETTIMEOFDAY_FOR_ID
#error Multiple id options selected
#endif
#include <openssl/rand.h>
#endif

#define _FORTIFY_SOURCE 3

#include <string.h>
#include <sys/types.h>
#include <fcntl.h>
#include <sys/socket.h>
#include <sys/time.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <alloca.h>
#include <assert.h>
#include <netinet/in.h>
#include <arpa/inet.h>
#include <unistd.h>
#include <limits.h>
#include <sys/stat.h>

#ifndef HOST_NAME_MAX
#define HOST_NAME_MAX 255
#endif

#ifndef NDEBUG
#include <stdio.h>
#endif

#undef MIN
#define MIN(a,b) ((a)<(b)?(a):(b))

#ifdef __USE_ISOC99B
// libevent doesn't work without this
typedef uint8_t u_char;
typedef unsigned int uint;
#endif
#include <event.h>

#define u64 uint64_t
#define u32 uint32_t
#define u16 uint16_t
#define u8  uint8_t

#ifndef NDEBUG
#include <stdio.h>
#define log printf
#else
#define log(x,...)
#endif

#include "eventdns.h"

#define MAX_ADDRS 4  // maximum number of addresses from a single packet
// which we bother recording

#define TYPE_A 1
#define CLASS_INET 1

struct request {
	u8 *request;  // the dns packet data
	uint request_len;
	int reissue_count;
	int tx_count;  // the number of times that this packet has been sent
	void *user_pointer;  // the pointer given to us for this request
	eventdns_callback_type user_callback;
	struct nameserver *ns;  // the server which we last sent it

	// elements used by the searching code
	int search_index;
	struct search_state *search_state;
	char *search_origname;  // needs to be free()ed
	int search_flags;

	// these objects are kept in a circular list
	struct request *next, *prev;

	struct event timeout_event;

	u16 trans_id;  // the transaction id
	char request_appended;  // true if the request pointer is data which follows this struct
	char transmit_me;  // needs to be transmitted
};

struct nameserver {
	int socket;  // a connected UDP socket
	u32 address;
	int failed_times;  // number of times which we have given this server a chance
	int timedout;  // number of times in a row a request has timed out
	struct event event;
	// these objects are kept in a circular list
	struct nameserver *next, *prev;
	struct event timeout_event;  // used to keep the timeout for
				     // when we next probe this server.
				     // Valid if state == 0
	char state;  // zero if we think that this server is down
	char choaked;  // true if we have an EAGAIN from this server's socket
	char write_waiting;  // true if we are waiting for EV_WRITE events
};

static struct request *req_head = NULL, *req_waiting_head = NULL;
static struct nameserver *server_head = NULL;

// The number of good nameservers that we have
static int global_good_nameservers = 0;

// inflight requests are contained in the req_head list
// and are actually going out across the network
static int global_requests_inflight = 0;
// requests which aren't inflight are in the waiting list
// and are counted here
static int global_requests_waiting = 0;

static int global_max_requests_inflight = 64;

static struct timeval global_timeout = {3, 0};  // 3 seconds
static int global_max_reissues = 1;  // a reissue occurs when we get some errors from the server
static int global_max_retransmits = 3;  // number of times we'll retransmit a request which timed out
// number of timeouts in a row before we consider this server to be down
static int global_max_nameserver_timeout = 3;

// These are the timeout values for nameservers. If we find a nameserver is down
// we try to probe it at intervals as given below. Values are in seconds.
static const struct timeval global_nameserver_timeouts[] = {{10, 0}, {60, 0}, {300, 0}, {900, 0}, {3600, 0}};
static const int global_nameserver_timeouts_length = sizeof(global_nameserver_timeouts)/sizeof(struct timeval);

const char *const eventdns_error_strings[] = {"no error", "The name server was unable to interpret the query", "The name server suffered an internal error", "The requested domain name does not exist", "The name server refused to reply to the request"};

static struct nameserver *nameserver_pick();
static void eventdns_request_insert(struct request *req, struct request **head);
static void nameserver_ready_callback(int fd, short events, void *arg);
static int eventdns_transmit();
static int eventdns_request_transmit(struct request *req);
static void nameserver_send_probe(struct nameserver *const ns);
static void search_request_finished(struct request *const);
static int search_try_next(struct request *const req);
static int search_request_new(const char *const name, int flags, eventdns_callback_type user_callback, void *user_arg);
static void eventdns_requests_pump_waiting_queue();
static u16 transaction_id_pick();
static struct request *request_new(const char *name, int flags, eventdns_callback_type callback, void *ptr);
static void request_submit(struct request *req);

// This walks the list of inflight requests to find the
// one with a matching transaction id. Returns NULL on
// failure
static struct request *
request_find_from_trans_id(u16 trans_id) {
	struct request *req = req_head, *const started_at = req_head;

	if (req) {
		do {
			if (req->trans_id == trans_id) return req;
			req = req->next;
		} while (req != started_at);
	}

	return NULL;
}

// a libevent callback function which is called when a nameserver
// has gone down and we want to test if it has came back to life yet
static void
nameserver_prod_callback(int fd, short events, void *arg) {
	struct nameserver *const ns = (struct nameserver *) arg;

	nameserver_send_probe(ns);
}

// a libevent callback which is called when a nameserver probe (to see if
// it has come back to life) times out. We increment the count of failed_times
// and wait longer to send the next probe packet.
static void
nameserver_probe_failed(struct nameserver *const ns) {
	assert(ns->state == 0);
	evtimer_del(&ns->timeout_event);
	const struct timeval *const timeout =
	  &global_nameserver_timeouts[MIN(ns->failed_times,
					  global_nameserver_timeouts_length - 1)];
	ns->failed_times++;

	evtimer_set(&ns->timeout_event, nameserver_prod_callback, ns);
	evtimer_add(&ns->timeout_event, (struct timeval *) timeout);
}

// called when a nameserver has been deemed to have failed. For example, too
// many packets have timed out etc
static void
nameserver_failed(struct nameserver *const ns) {
	// if this nameserver has already been marked as failed
	// then don't do anything
	if (!ns->state) return;

	log("Nameserver %lx has failed\n", (unsigned long)ns->address);
	global_good_nameservers--;
	assert(global_good_nameservers >= 0);
	if (global_good_nameservers == 0) {
		log("All nameservers have failed\n");
	}

	ns->state = 0;
	ns->failed_times = 1;

	evtimer_set(&ns->timeout_event, nameserver_prod_callback, ns);
	evtimer_add(&ns->timeout_event, (struct timeval *) &global_nameserver_timeouts[0]);

	// walk the list of inflight requests to see if any can be reassigned to
	// a different server. Requests in the waiting queue don't have a
	// nameserver assigned yet
	
	// if we don't have *any* good nameservers then there's no point 
	// trying to reassign requests to one
	if (!global_good_nameservers) return;

	struct request *req = req_head, *const started_at = req_head;
	if (req) {
		do {
			if (req->tx_count == 0 && req->ns == ns) {
				// still waiting to go out, can be moved
				// to another server
				req->ns = nameserver_pick();
			}
			req = req->next;
		} while (req != started_at);
	}
}

static void
nameserver_up(struct nameserver *const ns) {
	if (ns->state) return;
	log("Nameserver %lx is back up\n", (unsigned long)ns->address);
	evtimer_del(&ns->timeout_event);
	ns->state = 1;
	ns->failed_times = 0;
	global_good_nameservers++;
}

static void
request_trans_id_set(struct request *const req, const u16 trans_id) {
	req->trans_id = trans_id;
	*((u16 *) req->request) = htons(trans_id);
}

// Called to remove a request from a list and dealloc it.
// head is a pointer to the head of the list it should be
// removed from or NULL if the request isn't in a list.
static void
request_finished(struct request *const req, struct request **head) {
	if (head) {
		if (req->next == req) {
			// only item in the list
			*head = NULL;
		} else {
			req->next->prev = req->prev;
			req->prev->next = req->next;
			if (*head == req) *head = req->next;
		}
	}

	log("Removing timeout for %x\n", (unsigned) req);
	evtimer_del(&req->timeout_event);

	search_request_finished(req);
	global_requests_inflight--;

	if (!req->request_appended) {
		// need to free the request data on it's own
		free(req->request);
	} else {
		// the request data is appended onto the header
		// so everything gets free()ed when we:
	}

	free(req);

	eventdns_requests_pump_waiting_queue();
}

// This is called when a server returns a funny error code.
// We try the request again with another server.
//
// return:
//   0 ok
//   1 failed/reissue is pointless
static int
request_reissue(struct request *req) {
	const struct nameserver *const last_ns = req->ns;
	// the last nameserver should have been marked as failing
	// by the caller of this function, therefore pick will try
	// not to return it
	req->ns = nameserver_pick();
	if (req->ns == last_ns) {
		// ... but pick did return it
		// not a lot of point in trying again with the
		// same server
		return 1;
	}

	req->reissue_count++;
	req->tx_count = 0;
	req->transmit_me = 1;

	return 0;
}

// this function looks for space on the inflight queue and promotes
// requests from the waiting queue if it can.
static void
eventdns_requests_pump_waiting_queue() {
	while (global_requests_inflight < global_max_requests_inflight &&
	    global_requests_waiting) {
		// move a request from the waiting queue to the inflight queue
		assert(req_waiting_head);
		struct request *req;
		if (req_waiting_head->next == req_waiting_head) {
			// only one item in the queue
			req = req_waiting_head;
			req_waiting_head = NULL;
		} else {
			req = req_waiting_head;
			req->next->prev = req->prev;
			req->prev->next = req->next;
			req_waiting_head = req->next;
		}

		global_requests_waiting--;
		global_requests_inflight++;
		
		req->ns = nameserver_pick();
		request_trans_id_set(req, transaction_id_pick());

		eventdns_request_insert(req, &req_head);
		eventdns_request_transmit(req);
		eventdns_transmit();
	}
}

// this processes a parsed reply packet
static void
reply_handle(u16 trans_id, u16 flags, u32 ttl, u32 addrcount, u32 *addresses) {
	int error;
	static const int error_codes[] = {DNS_ERR_FORMAT, DNS_ERR_SERVERFAILED, DNS_ERR_NOTEXIST, DNS_ERR_NOTIMPL, DNS_ERR_REFUSED};

	struct request *const req = request_find_from_trans_id(trans_id);
	if (!req) return;
	
	if (flags & 0x020f || !addrcount) {
		// there was an error
		if (flags & 0x0200) {
			error = DNS_ERR_TRUNCATED;
		} else {
			u16 error_code = (flags & 0x000f) - 1;
			if (error_code > 4) {
				error = DNS_ERR_UNKNOWN;
			} else {
				error = error_codes[error_code];
			}
		}

		switch(error) {
		case DNS_ERR_SERVERFAILED:
		case DNS_ERR_NOTIMPL:
		case DNS_ERR_REFUSED:
			// we regard these errors as marking a bad nameserver
			if (req->reissue_count < global_max_reissues) {
				nameserver_failed(req->ns);
				if (!request_reissue(req)) return;
			}
			break;
		default:
			// we got a good reply from the nameserver
			nameserver_up(req->ns);
		}

		if (req->search_state) {
			// if we have a list of domains to search in, try the next one
			if (!search_try_next(req)) {
				// a new request was issued so this request is finished and
				// the user callback will be made when that request (or a
				// child of it) finishes.
				request_finished(req, &req_head);
				return;
			}
		}
		
		// all else failed. Pass the failure up
		req->user_callback(error, 0, 0, 0, NULL, req->user_pointer);
		request_finished(req, &req_head);
	} else {
		// all ok, tell the user
		req->user_callback(DNS_ERR_NONE, DNS_IPv4_A, addrcount, ttl, addresses, req->user_pointer);
		nameserver_up(req->ns);
		request_finished(req, &req_head);
	}
}

// parses a raw packet from the wire
static void
reply_parse(u8 *packet, int length) {
	int j = 0;  // index into packet
	u16 _t;  // used by the macros
	u32 _t32;  // used by the macros

#define GET32(x) do { if (j + 4 > length) return; memcpy(&_t32, packet + j, 4); j += 4; x = ntohl(_t32); } while(0);
#define GET16(x) do { if (j + 2 > length) return; memcpy(&_t, packet + j, 2); j += 2; x = ntohs(_t); } while(0);
#define GET8(x) do { if (j >= length) return; x = packet[j++]; } while(0);
	u16 trans_id, flags, questions, answers, authority, additional, datalength;
	u32 ttl, ttl_r = 0xffffffff;
	u32 addresses[MAX_ADDRS];
	int addresses_done = 0;

	GET16(trans_id);
	GET16(flags);
	GET16(questions);
	GET16(answers);
	GET16(authority);
	GET16(additional);

	if (!(flags & 0x8000)) return;  // must be an answer
	if (flags & 0x020f) {
		// there was an error
		reply_handle(trans_id, flags, 0, 0, NULL);
		return;
	}
	// if (!answers) return;  // must have an answer of some form
	
	// This macro skips a name in the DNS reply. Normally the 
	// names are a series of length prefixed strings terminated with
	// a length of 0 (the lengths are u8's < 63).
	// However, the length can start with a pair of 1 bits and that
	// means that the next 14 bits are a pointer within the current
	// packet. The name stops after a pointer like that.
#define SKIP_NAME \
	for(;;) { \
		u8 label_len; \
		GET8(label_len); \
		if (!label_len) break; \
		if (label_len & 0xc0) { \
			GET8(label_len); \
			break; \
		} \
		if (label_len > 63) return; \
		j += label_len; \
	}

	uint i;
	// skip over each question in the reply
	for (i = 0; i < questions; ++i) {
		// the question looks like
		//   <label:name><u16:type><u16:class>
		SKIP_NAME;
		j += 4;
	}

	// now we have the answer section which looks like
	// <label:name><u16:type><u16:class><u32:ttl><u16:len><data...>
	
	for (i = 0; i < answers; ++i) {
		u16 type, class;

		SKIP_NAME;
		GET16(type);
		GET16(class);
		GET32(ttl);
		GET16(datalength);

		if (type == TYPE_A && class == CLASS_INET) {
			const uint addrcount = datalength >> 2;  // each IP address is 4 bytes
			const uint addrtocopy = MIN(MAX_ADDRS - addresses_done, addrcount);

			ttl_r = MIN(ttl_r, ttl);
			// we only bother with the first four addresses.
			if (j + 4*addrtocopy > length) return;
			memcpy(&addresses[addresses_done], packet + j, 4*addrtocopy);
			j += 4*addrtocopy;
			addresses_done += addrtocopy;
			if (addresses_done == MAX_ADDRS) break;
		} else {
			// skip over any other type of resource
			j += datalength;
		}
	}

	reply_handle(trans_id, flags, ttl_r, addresses_done, addresses);
#undef SKIP_NAME
#undef GET32
#undef GET16
#undef GET8
}

// Try to choose a strong transaction id which isn't already in flight
static u16
transaction_id_pick() {
	for (;;) {
#ifdef DNS_USE_CPU_CLOCK_FOR_ID
		struct timespec ts;
		if (clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts)) abort();
		const u16 trans_id = ts.tv_nsec & 0xffff;
#endif

#ifdef DNS_USE_GETTIMEOFDAY_FOR_ID
		struct timeval tv;
		gettimeofday(&tv, NULL);
		const u16 trans_id = tv.tv_usec & 0xffff;
#endif

#ifdef DNS_USE_OPENSSL_FOR_ID
		u16 trans_id;
		if (RAND_pseudo_bytes((u8 *) &trans_id, 2) == -1) {
			/* // in the case that the RAND call fails we back
			// down to using gettimeofday.
			struct timeval tv;
			gettimeofday(&tv, NULL);
			trans_id = tv.tv_usec & 0xffff; */
			abort();
		}
#endif

		if (trans_id == 0xffff) continue;
		// now check to see if that id is already inflight
		const struct request *req = req_head, *const started_at = req_head;
		if (req) {
			do {
				if (req->trans_id == trans_id) break;
				req = req->next;
			} while (req != started_at);
		}
		// we didn't find it, so this is a good id
		if (req == started_at) return trans_id;
	}
}

// choose a namesever to use. This function will try to ignore
// nameservers which we think are down and load balence across the rest
// by updating the server_head global each time.
static struct nameserver *
nameserver_pick() {
	if (!server_head) return NULL;

	// if we don't have any good nameservers then there's no
	// point in trying to find one.
	if (!global_good_nameservers) {
		server_head = server_head->next;
		return server_head;
	}
	
	// remember that nameservers are in a circular list
	struct nameserver *started_at = server_head, *picked;
	for (;;) {
		if (server_head->state) {
			// we think this server is currently good
			picked = server_head;
			server_head = server_head->next;
			return picked;
		}

		server_head = server_head->next;
		if (server_head == started_at) {
			// all the nameservers seem to be down
			// so we just return this one and hope for the
			// best
			assert(global_good_nameservers == 0);
			picked = server_head;
			server_head = server_head->next;
			return picked;
		}
	}
}

// this is called when a namesever socket is ready for reading
static void
nameserver_read(struct nameserver *ns) {
	u8 packet[1500];

	for (;;) {
          	const int r = recv(ns->socket, packet, sizeof(packet), 0);
		if (r < 0) {
			if (errno == EAGAIN) return;
			nameserver_failed(ns);
			return;
		}
		reply_parse(packet, r);
	}
}

// set if we are waiting for the ability to write to this server.
// if waiting is true then we ask libevent for EV_WRITE events, otherwise
// we stop these events.
static void
nameserver_write_waiting(struct nameserver *ns, char waiting) {
	if (ns->write_waiting == waiting) return;
	
	ns->write_waiting = waiting;
	event_del(&ns->event);
	event_set(&ns->event, ns->socket, EV_READ | (waiting ? EV_WRITE : 0) | EV_PERSIST,
			nameserver_ready_callback, ns);
	event_add(&ns->event, NULL);
}

// a callback function. Called by libevent when the kernel says that
// a nameserver socket is ready for writing or reading
static void
nameserver_ready_callback(int fd, short events, void *arg) {
	struct nameserver *ns = (struct nameserver *) arg;
	
	if (events & EV_WRITE) {
		ns->choaked = 0;
		if (!eventdns_transmit()) {
			nameserver_write_waiting(ns, 0);
		}
	}
	if (events & EV_READ) {
		nameserver_read(ns);
	}
}

// Converts a string to a length-prefixed set of DNS labels.
// @buf must be strlen(name)+2 or longer. name and buf must
// not overlap. name_len should be the length of name
//
// Input: abc.def
// Output: <3>abc<3>def<0>
//
// Returns the length of the data. negative on error
//   -1  label was > 63 bytes
//   -2  name was > 255 bytes
static int
dnsname_to_labels(u8 *const buf, const char *name, const int name_len) { \
	const char *end = name + name_len; \
	int j = 0;  // current offset into buf

	if (name_len > 255) return -2;
	
	for (;;) {
		const char *const start = name;
		name = strchr(name, '.');
		if (!name) {
			const uint label_len = end - start;
			if (label_len > 63) return -1;
			buf[j++] = label_len;

			memcpy(buf + j, start, end - start);
			j += end - start;
			break;
		} else {
			// append length of the label.
			const uint label_len = name - start;
			if (label_len > 63) return -1;
			buf[j++] = label_len;

			memcpy(buf + j, start, name - start);
			j += name - start;
			// hop over the '.'
			name++;
		}
	}

	// the labels must be terminated by a 0.
	// It's possible that the name ended in a .
	// in which case the zero is already there
	if (!j || buf[j-1]) buf[j++] = 0;
	return j;
}

// Finds the length of a dns request for a DNS name of the given
// length. The actual request may be smaller than the value returned
// here
static int
eventdns_request_len(const int name_len) {
	return 96 + // length of the DNS standard header
		name_len + 2 + 
		4;  // space for the resource type
}

// build a dns request packet into buf. buf should be at least as long
// as eventdns_request_len told you it should be.
//
// Returns the amount of space used. Negative on error.
static int
eventdns_request_data_build(const char *const name, const int name_len, const u16 trans_id,
		const u16 type, const u16 class,
		u8 *const buf) {
	int j = 0;  // current offset into buf
	u16 _t;  // used by the macros
	
#define APPEND16(x) do { _t = htons(x); memcpy(buf + j, &_t, 2); j += 2; } while(0);
	APPEND16(trans_id);
	APPEND16(0x0100);  // standard query, recusion needed
	APPEND16(1);  // one question
	APPEND16(0);  // no answers
	APPEND16(0);  // no authority
	APPEND16(0);  // no additional

	u8 *labels = (u8 *) alloca(name_len + 2);
	const int labels_len = dnsname_to_labels(labels, name, name_len);
	if (labels_len < 0) return labels_len;
	memcpy(buf + j, labels, labels_len);
	j += labels_len;

	APPEND16(type);
	APPEND16(class);
#undef APPEND16

	return j;
}

// this is a libevent callback function which is called when a request
// has timed out.
static void
eventdns_request_timeout_callback(int fd, short events, void *arg) {
	struct request *const req = (struct request *) arg;
	
	log("Request %x timed out\n", (unsigned) arg);

	req->ns->timedout++;
	if (req->ns->timedout > global_max_nameserver_timeout) {
		nameserver_failed(req->ns);
	}

	evtimer_del(&req->timeout_event);
	if (req->tx_count >= global_max_retransmits) {
		// this request has failed
		req->user_callback(DNS_ERR_TIMEOUT, 0, 0, 0, NULL, req->user_pointer);
		request_finished(req, &req_head);
	} else {
		// retransmit it
		eventdns_request_transmit(req);
	}
}

// try to send a request to a given server.
//
// return:
//   0 ok
//   1 temporary failure
//   2 other failure
static int
eventdns_request_transmit_to(struct request *req, struct nameserver *server) {
	const int r = send(server->socket, req->request, req->request_len, 0);
	if (r < 0) {
		if (errno == EAGAIN) return 1;
		return 2;
	} else if (r != req->request_len) {
		return 1;  // short write
	} else {
		return 0;
	}
}

// try to send a request, updating the fields of the request
// as needed
//
// return:
//   0 ok
//   1 failed
static int
eventdns_request_transmit(struct request *req) {
	int retcode = 0;
	
	// if we fail to send this packet then this flag marks it
	// for eventdns_transmit
	req->transmit_me = 1;
	if (req->trans_id == 0xffff) abort();

	if (req->ns->choaked) {
		// don't bother trying to write to a socket
		// which we have had EAGAIN from
		return 1;
	}

	const int r = eventdns_request_transmit_to(req, req->ns);
	switch (r) {
	case 1:
		// temp failure
		req->ns->choaked = 1;
		nameserver_write_waiting(req->ns, 1);
		return 1;
	case 2:
		// failed in some other way
		nameserver_failed(req->ns);
		retcode = 1;
		// fall through
	default:
		// all ok
		log("Setting timeout for %x\n", (unsigned) req);
		evtimer_set(&req->timeout_event, eventdns_request_timeout_callback, req);
		evtimer_add(&req->timeout_event, &global_timeout);
		req->tx_count++;
		req->transmit_me = 0;
		return retcode;
	}
}

static void
nameserver_probe_callback(int result, char type, int count, int ttl, void *addresses, void *arg) {
	struct nameserver *const ns = (struct nameserver *) arg;
	if (result == DNS_ERR_NONE || result == DNS_ERR_NOTEXIST) {
		// this is a good reply
		nameserver_up(ns);
	} else nameserver_probe_failed(ns);
}

static void
nameserver_send_probe(struct nameserver *const ns) {
	// here we need to send a probe to a given nameserver
	// in the hope that it is up now.

  	log("Sending probe to %lx\n", (unsigned long)ns->address);

	struct request *const req = request_new("www.google.com", DNS_QUERY_NO_SEARCH, nameserver_probe_callback, ns);
	// we force this into the inflight queue no matter what
	request_trans_id_set(req, transaction_id_pick());
	req->ns = ns;
	request_submit(req);
}

// returns:
//   0 didn't try to transmit anything
//   1 tried to transmit something
static int
eventdns_transmit() {
	char did_try_to_transmit = 0;

	if (req_head) {
		struct request *const started_at = req_head, *req = req_head;
		// first transmit all the requests which are currently waiting
		do {
			if (req->transmit_me) {
				did_try_to_transmit = 1;
				eventdns_request_transmit(req);
			}

			req = req->next;
		} while (req != started_at);
	}

	return did_try_to_transmit;
}

// exported function
int
eventdns_nameserver_add(unsigned long int address) {
	// first check to see if we already have this nameserver
	
	const struct nameserver *server = server_head, *const started_at = server_head;
	if (server) {
		do {
			if (server->address == address) return 3;
			server = server->next;
		} while (server != started_at);
	}
	
	struct nameserver *ns = (struct nameserver *) malloc(sizeof(struct nameserver));
	struct sockaddr_in sin;
	int err = 0;

	memset(ns, 0, sizeof(struct nameserver));

	ns->socket = socket(PF_INET, SOCK_DGRAM, 0);
	if (ns->socket < 0) { err = 1; goto out1; }
        fcntl(ns->socket, F_SETFL, O_NONBLOCK);
	sin.sin_addr.s_addr = address;
	sin.sin_port = htons(53);
	sin.sin_family = AF_INET;
	if (connect(ns->socket, (struct sockaddr *) &sin, sizeof(sin)) != 0) {
		err = 2;
		goto out2;
	}

	ns->address = address;
	ns->state = 1;
	event_set(&ns->event, ns->socket, EV_READ | EV_PERSIST, nameserver_ready_callback, ns);
	event_add(&ns->event, NULL);

	// insert this nameserver into the list of them
	if (!server_head) {
		ns->next = ns->prev = ns;
		server_head = ns;
	} else {
		ns->next = server_head->next;
		ns->prev = server_head;
		server_head->next = ns;
		if (server_head->prev == server_head) {
			server_head->prev = ns;
		}
	}

	global_good_nameservers++;
	
	return 0;

out2:
	close(ns->socket);
out1:
	free(ns);
	return err;
}

// exported function
int
eventdns_nameserver_ip_add(const char *ip_as_string) {
	struct in_addr ina;
	if (!inet_aton(ip_as_string, &ina)) return 4;
	return eventdns_nameserver_add(ina.s_addr);
}

// insert into the tail of the queue
static void
eventdns_request_insert(struct request *req, struct request **head) {
	if (!*head) {
		*head = req;
		req->next = req->prev = req;
		return;
	}

	req->prev = (*head)->prev;
	req->prev->next = req;
	req->next = *head;
	(*head)->prev = req;
}

static int
string_num_dots(const char *s) {
	int count = 0;
	while ((s = strchr(s, '.'))) {
		s++;
		count++;
	}
	return count;
}

static struct request *
request_new(const char *name, int flags, eventdns_callback_type callback, void *ptr) {
	const char issuing_now = (global_requests_inflight < global_max_requests_inflight) ? 1 : 0;

	const int name_len = strlen(name);
	const int request_max_len = eventdns_request_len(name_len);
	const u16 trans_id = issuing_now ? transaction_id_pick() : 0xffff;
	// the request data is alloced in a single block with the header
	struct request *const req = (struct request *) malloc(sizeof(struct request) + request_max_len);
	memset(req, 0, sizeof(struct request));

	// request data lives just after the header
	req->request = ((u8 *) req) + sizeof(struct request);
	req->request_appended = 1;  // denotes that the request data shouldn't be free()ed
	req->request_len = eventdns_request_data_build(name, name_len, trans_id, TYPE_A, CLASS_INET, req->request);
	if (req->request_len < 0) goto err1;
	req->trans_id = trans_id;
	req->tx_count = 0;
	req->user_pointer = ptr;
	req->user_callback = callback;
	req->ns = issuing_now ? nameserver_pick() : NULL;
	req->next = req->prev = NULL;

	return req;
err1:
	free(req->request);
	return NULL;
}

static void
request_submit(struct request *const req) {
	if (req->ns) {
		// if it has a nameserver assigned then this is going
		// straight into the inflight queue
		eventdns_request_insert(req, &req_head);
		global_requests_inflight++;
		eventdns_request_transmit(req);
	} else {
		eventdns_request_insert(req, &req_waiting_head);
		global_requests_waiting++;
	}
}

// exported function
int eventdns_resolve(const char *name, int flags, eventdns_callback_type callback, void *ptr) {
	log("resolve for %s\n", name);
	if (flags & DNS_QUERY_NO_SEARCH) {
		struct request *const req = request_new(name, flags, callback, ptr);
		if (!req) return 1;
		request_submit(req);
		return 0;
	} else {
		return search_request_new(name, flags, callback, ptr);
	}
}

/////////////////////////////////////////////////////////////////////
// Search support
//
// the libc resolver has support for searching a number of domains
// to find a name. If nothing else then it takes the single domain
// from the gethostname() call.
//
// It can also be configured via the domain and search options in a
// resolv.conf.
//
// The ndots option controls how many dots it takes for the resolver
// to decide that a name is non-local and so try a raw lookup first.

struct search_domain {
	int len;
	struct search_domain *next;
	// the text string is appended to this structure
};

struct search_state {
	int refcount;
	int ndots;
	int num_domains;
	struct search_domain *head;
};

static struct search_state *global_search_state = NULL;

static void
search_state_decref(struct search_state *const state) {
	if (!state) return;
	state->refcount--;
	if (!state->refcount) {
		struct search_domain *next, *dom;
		for (dom = state->head; dom; dom = next) {
			next = dom->next;
			free(dom);
		}
		free(state);
	}
};

static struct search_state *
search_state_new() {
	struct search_state *state = (struct search_state *) malloc(sizeof(struct search_state));
	memset(state, 0, sizeof(struct search_state));
	state->refcount = 1;
	state->ndots = 1;

	return state;
}

static void
search_postfix_clear() {
	search_state_decref(global_search_state);
	
	global_search_state = search_state_new();
}

// exported function
void
eventdns_search_clear() {
	search_postfix_clear();
}

static void
search_postfix_add(const char *domain) {
	while (domain[0] == '.') domain++;
	const int domain_len = strlen(domain);

	if (!global_search_state) global_search_state = search_state_new();
	global_search_state->num_domains++;

	struct search_domain *sdomain = (struct search_domain *) malloc(sizeof(struct search_domain) + domain_len);
	memcpy( ((u8 *) sdomain) + sizeof(struct search_domain), domain, domain_len);
	sdomain->next = global_search_state->head;
	sdomain->len = domain_len;

	global_search_state->head = sdomain;
}

// reverse the order of members in the postfix list. This is needed because,
// when parsing resolv.conf we push elements in the wrong order
static void
search_reverse() {
	struct search_domain *cur, *prev = NULL, *next;
	cur = global_search_state->head;
	while (cur) {
		next = cur->next;
		cur->next = prev;
		prev = cur;
		cur = next;
	}

	global_search_state->head = prev;
}

// exported function
void
eventdns_search_add(const char *domain) {
	search_postfix_add(domain);
}

// exported function
void
eventdns_search_ndots_set(const int ndots) {
	if (!global_search_state) global_search_state = search_state_new();
	global_search_state->ndots = ndots;
}

static void
search_set_from_hostname() {
	char hostname[HOST_NAME_MAX + 1];

	search_postfix_clear();
	if (gethostname(hostname, sizeof(hostname))) return;
	char *domainname = strchr(hostname, '.');
	if (!domainname) return;
	search_postfix_add(domainname);
}

// warning: returns malloced string
static char *
search_make_new(const struct search_state *const state, int n, const char *const base_name) {
	const int base_len = strlen(base_name);
	const char need_to_append_dot = base_name[base_len - 1] == '.' ? 0 : 1;
	struct search_domain *dom;

	for (dom = state->head; dom; dom = dom->next) {
		if (!n--) {
			// this is the postfix we want
			// the actual postfix string is kept at the end of the structure
			const u8 *const postfix = ((u8 *) dom) + sizeof(struct search_domain);
			const int postfix_len = dom->len;
			char *const newname = (char *) malloc(base_len + need_to_append_dot + postfix_len + 1);
			memcpy(newname, base_name, base_len);
			if (need_to_append_dot) newname[base_len] = '.';
			memcpy(newname + base_len + need_to_append_dot, postfix, postfix_len);
			newname[base_len + need_to_append_dot + postfix_len] = 0;
			return newname;
		}
	}

	// we ran off the end of the list and still didn't find the requested string
	abort();
}

static int
search_request_new(const char *const name, int flags, eventdns_callback_type user_callback, void *user_arg) {
	if ( ((flags & DNS_QUERY_NO_SEARCH) == 0) &&
	     global_search_state &&
		 global_search_state->num_domains) {
		// we have some domains to search
		struct request *req;
		if (string_num_dots(name) >= global_search_state->ndots) {
			req = request_new(name, flags, user_callback, user_arg);
			if (!req) return 1;
			req->search_index = -1;
		} else {
			char *const new_name = search_make_new(global_search_state, 0, name);
			req = request_new(new_name, flags, user_callback, user_arg);
			free(new_name);
			if (!req) return 1;
			req->search_index = 0;
		}
		req->search_origname = strdup(name);
		req->search_state = global_search_state;
		req->search_flags = flags;
		global_search_state->refcount++;
		request_submit(req);
		return 0;
	} else {
		struct request *const req = request_new(name, flags, user_callback, user_arg);
		if (!req) return 1;
		request_submit(req);
		return 0;
	}
}

// this is called when a request has failed to find a name. We need to check
// if it is part of a search and, if so, try the next name in the list
// returns:
//   0 another request has been submitted
//   1 no more requests needed
static int
search_try_next(struct request *const req) {
	if (req->search_state) {
		// it is part of a search
		req->search_index++;
		if (req->search_index >= req->search_state->num_domains) {
			// no more postfixes to try, however we may need to try
			// this name without a postfix
			if (string_num_dots(req->search_origname) < req->search_state->ndots) {
				// yep, we need to try it raw
				log("search: trying raw query %s\n", req->search_origname);
				struct request *const newreq = request_new(req->search_origname, req->search_flags, req->user_callback, req->user_pointer);
				if (newreq) {
					request_submit(newreq);
					return 0;
				}
			}
			return 1;
		}

		char *const new_name = search_make_new(req->search_state, req->search_index, req->search_origname);
		log("search: now trying %s (%d)\n", new_name, req->search_index);
		struct request *const newreq = request_new(new_name, req->search_flags, req->user_callback, req->user_pointer);
		free(new_name);
		if (!newreq) return 1;
		newreq->search_origname = req->search_origname;
		req->search_origname = NULL;
		newreq->search_state = req->search_state;
		newreq->search_flags = req->search_flags;
		newreq->search_index = req->search_index;
		newreq->search_state->refcount++;
		request_submit(newreq);
		return 0;
	}
	return 1;
}

static void
search_request_finished(struct request *const req) {
	if (req->search_state) {
		search_state_decref(req->search_state);
		req->search_state = NULL;
	}
	if (req->search_origname) {
		free(req->search_origname);
		req->search_origname = NULL;
	}
}

/////////////////////////////////////////////////////////////////////
// Parsing resolv.conf files

static void
eventdns_resolv_set_defaults(int flags) {
	// if the file isn't found then we assume a local resolver
	if (flags & DNS_OPTION_SEARCH) search_set_from_hostname();
	if (flags & DNS_OPTION_NAMESERVERS) eventdns_nameserver_ip_add("127.0.0.1");
}

#ifndef HAVE_STRTOK_R
static char *
strtok_r(char *s, const char *delim, char **state) {
	return strtok(s, delim);
}
#endif

// helper version of atoi which returns -1 on error
static int
strtoint(const char *const str) {
	char *endptr;
	const int r = strtol(str, &endptr, 10);
	if (*endptr) return -1;
	return r;
}

static void
resolv_conf_parse_line(char *const start, int flags) {
	char *strtok_state;
	static const char *const delims = " \t";
#define NEXT_TOKEN strtok_r(NULL, delims, &strtok_state)

	char *const first_token = strtok_r(start, delims, &strtok_state);
	if (!first_token) return;
	 
	if (!strcmp(first_token, "nameserver")) {
		const char *const nameserver = NEXT_TOKEN;
		struct in_addr ina;

		if (inet_aton(nameserver, &ina)) {
			// address is valid
			eventdns_nameserver_add(ina.s_addr);
		}
	} else if (!strcmp(first_token, "domain") && (flags & DNS_OPTION_SEARCH)) {
		const char *const domain = NEXT_TOKEN;
		if (domain) {
			search_postfix_clear();
			search_postfix_add(domain);
		}
	} else if (!strcmp(first_token, "search") && (flags & DNS_OPTION_SEARCH)) {
		search_postfix_clear();
		const char *domain;

		while ((domain = NEXT_TOKEN)) {
			search_postfix_add(domain);
		}
		search_reverse();
	} else if (!strcmp(first_token, "options")) {
		const char *option;

		while ((option = NEXT_TOKEN)) { 
			if (!strncmp(option, "ndots:", 6)) {
				const int ndots = strtoint(&option[6]);
				if (ndots == -1) continue;
				if (!(flags & DNS_OPTION_SEARCH)) continue;
				log("setting ndots to %d\n", ndots);
				if (!global_search_state) global_search_state = search_state_new();
				global_search_state->ndots = ndots;
			} else if (!strncmp(option, "timeout:", 8)) {
				const int timeout = strtoint(&option[8]);
				if (timeout == -1) continue;
				if (!(flags & DNS_OPTION_MISC)) continue;
				log("setting timeout to %d\n", timeout);
				global_timeout.tv_sec = timeout;
			} else if (!strncmp(option, "attempts:", 9)) {
				const int retries = strtoint(&option[9]);
				if (retries == -1) continue;
				if (!(flags & DNS_OPTION_MISC)) continue;
				log("setting retries to %d\n", retries);
				global_max_retransmits = retries;
			}
		}
	}
#undef NEXT_TOKEN
}

// exported function
// returns:
//   0 no errors
//   1 failed to open file
//   2 failed to stat file
//   3 file too large
//   4 out of memory
//   5 short read from file
int
eventdns_resolv_conf_parse(int flags, const char *const filename) {
	struct stat st;
	int err = 0;

	log("parsing resolve file %s\n", filename);

	const int fd = open(filename, O_RDONLY);
	if (fd < 0) {
		eventdns_resolv_set_defaults(flags);
		return 0;
	}
	
	if (fstat(fd, &st)) { err = 2; goto out1; }
	if (!st.st_size) {
		eventdns_resolv_set_defaults(flags);
		err = 0;
		goto out1;
	}
	if (st.st_size > 65535) { err = 3; goto out1; }  // no resolv.conf should be any bigger

	u8 *const resolv = (u8 *) malloc(st.st_size + 1);
	if (!resolv) { err = 4; goto out1; }
	
	if (read(fd, resolv, st.st_size) != st.st_size) { err = 5; goto out2; }
	resolv[st.st_size] = 0;  // we malloced an extra byte
	
	char *start = (char *) resolv;
	for (;;) {
		char *const newline = strchr(start, '\n');
		if (!newline) {
			resolv_conf_parse_line(start, flags);
			break;
		} else {
			*newline = 0;
			resolv_conf_parse_line(start, flags);
			start = newline + 1;
		}
	}

	if (!server_head && (flags & DNS_OPTION_NAMESERVERS)) {
		// no nameservers were configured.
		eventdns_nameserver_ip_add("127.0.0.1");
	}
	if (flags & DNS_OPTION_SEARCH && (!global_search_state || global_search_state->num_domains == 0)) {
		search_set_from_hostname();
	}

out2:
	free(resolv);
out1:
	close(fd);
	return err;
}