[ceph.git] / ceph / src / common / pick_address.cc

// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
// vim: ts=8 sw=2 smarttab
/*
 * Ceph - scalable distributed file system
 *
 * Copyright (C) 2004-2012 Inktank
 *
 * This is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License version 2.1, as published by the Free Software
 * Foundation.  See file COPYING.
 *
 */

#include "common/pick_address.h"

#include <netdb.h>
#include <netinet/in.h>
#include <string>
#include <string.h>
#include <vector>

#include <boost/algorithm/string/predicate.hpp>
#include <fmt/format.h>

#include "include/ipaddr.h"
#include "include/str_list.h"
#include "common/ceph_context.h"
#ifndef WITH_SEASTAR
#include "common/config.h"
#include "common/config_obs.h"
#endif
#include "common/debug.h"
#include "common/errno.h"
#include "common/numa.h"

#ifndef HAVE_IN_ADDR_T
typedef uint32_t in_addr_t;
#endif

#ifndef IN_LOOPBACKNET
#define IN_LOOPBACKNET 127
#endif

#define dout_subsys ceph_subsys_

using std::string;
using std::vector;

namespace {

bool matches_with_name(const ifaddrs& ifa, const std::string& if_name)
{
  return if_name.compare(ifa.ifa_name) == 0;
}

static int is_loopback_addr(sockaddr* addr)
{
  if (addr->sa_family == AF_INET) {
    const sockaddr_in* sin = (struct sockaddr_in *)(addr);
    const in_addr_t net = ntohl(sin->sin_addr.s_addr) >> IN_CLASSA_NSHIFT;
    return net == IN_LOOPBACKNET ? 1 : 0;
  } else if (addr->sa_family == AF_INET6) {
    sockaddr_in6* sin6 = (struct sockaddr_in6 *)(addr);
    return IN6_IS_ADDR_LOOPBACK(&sin6->sin6_addr) ? 1 : 0;
  } else {
    return -1;
  }
}

static int grade_addr(const ifaddrs& ifa)
{
  if (ifa.ifa_addr == nullptr) {
    return -1;
  }
  int score = 0;
  if (ifa.ifa_flags & IFF_UP) {
    score += 4;
  }
  switch (is_loopback_addr(ifa.ifa_addr)) {
  case 0:
    // prefer non-loopback addresses
    score += 2;
    break;
  case 1:
    score += 0;
    break;
  default:
    score = -1;
    break;
  }
  return score;
}

bool matches_with_net(const ifaddrs& ifa,
                      const sockaddr* net,
                      unsigned int prefix_len,
                      unsigned ipv)
{
  switch (net->sa_family) {
  case AF_INET:
    if (ipv & CEPH_PICK_ADDRESS_IPV4) {
      return matches_ipv4_in_subnet(ifa, (struct sockaddr_in*)net, prefix_len);
    }
    break;
  case AF_INET6:
    if (ipv & CEPH_PICK_ADDRESS_IPV6) {
      return matches_ipv6_in_subnet(ifa, (struct sockaddr_in6*)net, prefix_len);
    }
    break;
  }
  return false;
}

bool matches_with_net(CephContext *cct,
                      const ifaddrs& ifa,
                      const std::string& s,
                      unsigned ipv)
{
  struct sockaddr_storage net;
  unsigned int prefix_len;
  if (!parse_network(s.c_str(), &net, &prefix_len)) {
    lderr(cct) << "unable to parse network: " << s << dendl;
    exit(1);
  }
  return matches_with_net(ifa, (sockaddr*)&net, prefix_len, ipv);
}

int grade_with_numa_node(const ifaddrs& ifa, int numa_node)
{
#if defined(WITH_SEASTAR) || defined(_WIN32)
  return 0;
#else
  if (numa_node < 0) {
    return 0;
  }
  int if_node = -1;
  int r = get_iface_numa_node(ifa.ifa_name, &if_node);
  if (r < 0) {
    return 0;
  }
  return if_node == numa_node ? 1 : 0;
#endif
}
}

const struct sockaddr *find_ip_in_subnet_list(
  CephContext *cct,
  const struct ifaddrs *ifa,
  unsigned ipv,
  const std::string &networks,
  const std::string &interfaces,
  int numa_node)
{
  const auto ifs = get_str_list(interfaces);
  const auto nets = get_str_list(networks);
  if (!ifs.empty() && nets.empty()) {
      lderr(cct) << "interface names specified but not network names" << dendl;
      exit(1);
  }

  int best_score = 0;
  const sockaddr* best_addr = nullptr;
  for (const auto* addr = ifa; addr != nullptr; addr = addr->ifa_next) {
    if (!ifs.empty() &&
	std::none_of(std::begin(ifs), std::end(ifs),
                     [&](const auto& if_name) {
                       return matches_with_name(*addr, if_name);
                     })) {
      continue;
    }
    if (!nets.empty() &&
	std::none_of(std::begin(nets), std::end(nets),
                     [&](const auto& net) {
                       return matches_with_net(cct, *addr, net, ipv);
                     })) {
      continue;
    }
    int score = grade_addr(*addr);
    if (score < 0) {
      continue;
    }
    score += grade_with_numa_node(*addr, numa_node);
    if (score > best_score) {
      best_score = score;
      best_addr = addr->ifa_addr;
    }
  }
  return best_addr;
}

#ifndef WITH_SEASTAR
// observe this change
struct Observer : public md_config_obs_t {
  const char *keys[2];
  explicit Observer(const char *c) {
    keys[0] = c;
    keys[1] = NULL;
  }

  const char** get_tracked_conf_keys() const override {
    return (const char **)keys;
  }
  void handle_conf_change(const ConfigProxy& conf,
			  const std::set <std::string> &changed) override {
    // do nothing.
  }
};

static void fill_in_one_address(CephContext *cct,
				const struct ifaddrs *ifa,
				const string &networks,
				const string &interfaces,
				const char *conf_var,
				int numa_node = -1)
{
  const struct sockaddr *found = find_ip_in_subnet_list(
    cct,
    ifa,
    CEPH_PICK_ADDRESS_IPV4|CEPH_PICK_ADDRESS_IPV6,
    networks,
    interfaces,
    numa_node);
  if (!found) {
    lderr(cct) << "unable to find any IP address in networks '" << networks
	       << "' interfaces '" << interfaces << "'" << dendl;
    exit(1);
  }

  char buf[INET6_ADDRSTRLEN];
  int err;

  err = getnameinfo(found,
		    (found->sa_family == AF_INET)
		    ? sizeof(struct sockaddr_in)
		    : sizeof(struct sockaddr_in6),

		    buf, sizeof(buf),
		    nullptr, 0,
		    NI_NUMERICHOST);
  if (err != 0) {
    lderr(cct) << "unable to convert chosen address to string: " << gai_strerror(err) << dendl;
    exit(1);
  }

  Observer obs(conf_var);

  cct->_conf.add_observer(&obs);

  cct->_conf.set_val_or_die(conf_var, buf);
  cct->_conf.apply_changes(nullptr);

  cct->_conf.remove_observer(&obs);
}

void pick_addresses(CephContext *cct, int needs)
{
  auto public_addr = cct->_conf.get_val<entity_addr_t>("public_addr");
  auto public_network = cct->_conf.get_val<std::string>("public_network");
  auto public_network_interface =
    cct->_conf.get_val<std::string>("public_network_interface");
  auto cluster_addr = cct->_conf.get_val<entity_addr_t>("cluster_addr");
  auto cluster_network = cct->_conf.get_val<std::string>("cluster_network");
  auto cluster_network_interface =
    cct->_conf.get_val<std::string>("cluster_network_interface");

  struct ifaddrs *ifa;
  int r = getifaddrs(&ifa);
  if (r < 0) {
    string err = cpp_strerror(errno);
    lderr(cct) << "unable to fetch interfaces and addresses: " << err << dendl;
    exit(1);
  }
  auto free_ifa = make_scope_guard([ifa] { freeifaddrs(ifa); });
  if ((needs & CEPH_PICK_ADDRESS_PUBLIC) &&
    public_addr.is_blank_ip() && !public_network.empty()) {
    fill_in_one_address(cct, ifa, public_network, public_network_interface,
			"public_addr");
  }

  if ((needs & CEPH_PICK_ADDRESS_CLUSTER) && cluster_addr.is_blank_ip()) {
    if (!cluster_network.empty()) {
      fill_in_one_address(cct, ifa, cluster_network, cluster_network_interface,
			  "cluster_addr");
    } else {
      if (!public_network.empty()) {
        lderr(cct) << "Public network was set, but cluster network was not set " << dendl;
        lderr(cct) << "    Using public network also for cluster network" << dendl;
        fill_in_one_address(cct, ifa, public_network, public_network_interface,
			    "cluster_addr");
      }
    }
  }
}
#endif	// !WITH_SEASTAR

static int fill_in_one_address(
  CephContext *cct,
  const struct ifaddrs *ifa,
  unsigned ipv,
  const string &networks,
  const string &interfaces,
  entity_addrvec_t *addrs,
  int numa_node = -1)
{
  const struct sockaddr *found = find_ip_in_subnet_list(cct, ifa, ipv,
							networks,
							interfaces,
							numa_node);
  if (!found) {
    std::string ip_type = "";
    if ((ipv & CEPH_PICK_ADDRESS_IPV4) && (ipv & CEPH_PICK_ADDRESS_IPV6)) {
      ip_type = "IPv4 or IPv6";
    } else if (ipv & CEPH_PICK_ADDRESS_IPV4) {
      ip_type = "IPv4";
    } else {
      ip_type = "IPv6";
    }
    lderr(cct) << "unable to find any " << ip_type << " address in networks '"
               << networks << "' interfaces '" << interfaces << "'" << dendl;
    return -1;
  }

  char buf[INET6_ADDRSTRLEN];
  int err;

  err = getnameinfo(found,
		    (found->sa_family == AF_INET)
		    ? sizeof(struct sockaddr_in)
		    : sizeof(struct sockaddr_in6),

		    buf, sizeof(buf),
		    nullptr, 0,
		    NI_NUMERICHOST);
  if (err != 0) {
    lderr(cct) << "unable to convert chosen address to string: " << gai_strerror(err) << dendl;
    return -1;
  }

  entity_addr_t addr;
  const char *end = 0;
  bool r = addr.parse(buf, &end);
  if (!r) {
    return -1;
  }
  addrs->v.push_back(addr);
  return 0;
}

int pick_addresses(
  CephContext *cct,
  unsigned flags,
  struct ifaddrs *ifa,
  entity_addrvec_t *addrs,
  int preferred_numa_node)
{
  addrs->v.clear();

  unsigned addrt = (flags & (CEPH_PICK_ADDRESS_PUBLIC |
			     CEPH_PICK_ADDRESS_CLUSTER));
  if (addrt == 0 ||
      addrt == (CEPH_PICK_ADDRESS_PUBLIC |
		CEPH_PICK_ADDRESS_CLUSTER)) {
    return -EINVAL;
  }
  unsigned msgrv = flags & (CEPH_PICK_ADDRESS_MSGR1 |
			    CEPH_PICK_ADDRESS_MSGR2);
  if (msgrv == 0) {
    if (cct->_conf.get_val<bool>("ms_bind_msgr1")) {
      msgrv |= CEPH_PICK_ADDRESS_MSGR1;
    }
    if (cct->_conf.get_val<bool>("ms_bind_msgr2")) {
      msgrv |= CEPH_PICK_ADDRESS_MSGR2;
    }
    if (msgrv == 0) {
      return -EINVAL;
    }
  }
  unsigned ipv = flags & (CEPH_PICK_ADDRESS_IPV4 |
			  CEPH_PICK_ADDRESS_IPV6);
  if (ipv == 0) {
    if (cct->_conf.get_val<bool>("ms_bind_ipv4")) {
      ipv |= CEPH_PICK_ADDRESS_IPV4;
    }
    if (cct->_conf.get_val<bool>("ms_bind_ipv6")) {
      ipv |= CEPH_PICK_ADDRESS_IPV6;
    }
    if (ipv == 0) {
      return -EINVAL;
    }
    if (cct->_conf.get_val<bool>("ms_bind_prefer_ipv4")) {
      flags |= CEPH_PICK_ADDRESS_PREFER_IPV4;
    } else {
      flags &= ~CEPH_PICK_ADDRESS_PREFER_IPV4;
    }
  }

  entity_addr_t addr;
  string networks;
  string interfaces;
  if (addrt & CEPH_PICK_ADDRESS_PUBLIC) {
    addr = cct->_conf.get_val<entity_addr_t>("public_addr");
    networks = cct->_conf.get_val<std::string>("public_network");
    interfaces =
      cct->_conf.get_val<std::string>("public_network_interface");
  } else {
    addr = cct->_conf.get_val<entity_addr_t>("cluster_addr");
    networks = cct->_conf.get_val<std::string>("cluster_network");
    interfaces =
      cct->_conf.get_val<std::string>("cluster_network_interface");
    if (networks.empty()) {
      lderr(cct) << "Falling back to public interface" << dendl;
      // fall back to public_ network and interface if cluster is not set
      networks = cct->_conf.get_val<std::string>("public_network");
      interfaces =
	cct->_conf.get_val<std::string>("public_network_interface");
    }
  }
  if (addr.is_blank_ip() &&
      !networks.empty()) {
    int ipv4_r = !(ipv & CEPH_PICK_ADDRESS_IPV4) ? 0 : -1;
    int ipv6_r = !(ipv & CEPH_PICK_ADDRESS_IPV6) ? 0 : -1;
    // note: pass in ipv to filter the matching addresses
    if ((ipv & CEPH_PICK_ADDRESS_IPV4) &&
	(flags & CEPH_PICK_ADDRESS_PREFER_IPV4)) {
      ipv4_r = fill_in_one_address(cct, ifa, CEPH_PICK_ADDRESS_IPV4,
				   networks, interfaces,
				   addrs,
				   preferred_numa_node);
    }
    if (ipv & CEPH_PICK_ADDRESS_IPV6) {
      ipv6_r = fill_in_one_address(cct, ifa, CEPH_PICK_ADDRESS_IPV6,
				   networks, interfaces,
				   addrs,
				   preferred_numa_node);
    }
    if ((ipv & CEPH_PICK_ADDRESS_IPV4) &&
	!(flags & CEPH_PICK_ADDRESS_PREFER_IPV4)) {
      ipv4_r = fill_in_one_address(cct, ifa, CEPH_PICK_ADDRESS_IPV4,
				   networks, interfaces,
				   addrs,
				   preferred_numa_node);
    }
    if (ipv4_r < 0 || ipv6_r < 0) {
      return -1;
    }
  }

  // note: we may have a blank addr here

  // ipv4 and/or ipv6?
  if (addrs->v.empty()) {
    addr.set_type(entity_addr_t::TYPE_MSGR2);
    if ((ipv & CEPH_PICK_ADDRESS_IPV4) &&
	(flags & CEPH_PICK_ADDRESS_PREFER_IPV4)) {
      addr.set_family(AF_INET);
      addrs->v.push_back(addr);
    }
    if (ipv & CEPH_PICK_ADDRESS_IPV6) {
      addr.set_family(AF_INET6);
      addrs->v.push_back(addr);
    }
    if ((ipv & CEPH_PICK_ADDRESS_IPV4) &&
	!(flags & CEPH_PICK_ADDRESS_PREFER_IPV4)) {
      addr.set_family(AF_INET);
      addrs->v.push_back(addr);
    }
  }

  // msgr2 or legacy or both?
  if (msgrv == (CEPH_PICK_ADDRESS_MSGR1 | CEPH_PICK_ADDRESS_MSGR2)) {
    vector<entity_addr_t> v;
    v.swap(addrs->v);
    for (auto a : v) {
      a.set_type(entity_addr_t::TYPE_MSGR2);
      if (flags & CEPH_PICK_ADDRESS_DEFAULT_MON_PORTS) {
	a.set_port(CEPH_MON_PORT_IANA);
      }
      addrs->v.push_back(a);
      a.set_type(entity_addr_t::TYPE_LEGACY);
      if (flags & CEPH_PICK_ADDRESS_DEFAULT_MON_PORTS) {
	a.set_port(CEPH_MON_PORT_LEGACY);
      }
      addrs->v.push_back(a);
    }
  } else if (msgrv == CEPH_PICK_ADDRESS_MSGR1) {
    for (auto& a : addrs->v) {
      a.set_type(entity_addr_t::TYPE_LEGACY);
    }
  } else {
    for (auto& a : addrs->v) {
      a.set_type(entity_addr_t::TYPE_MSGR2);
    }
  }

  return 0;
}

int pick_addresses(
  CephContext *cct,
  unsigned flags,
  entity_addrvec_t *addrs,
  int preferred_numa_node)
{
  struct ifaddrs *ifa;
  int r = getifaddrs(&ifa);
  if (r < 0) {
    r = -errno;
    string err = cpp_strerror(r);
    lderr(cct) << "unable to fetch interfaces and addresses: "
	       <<  cpp_strerror(r) << dendl;
    return r;
  }
  r = pick_addresses(cct, flags, ifa, addrs, preferred_numa_node);
  freeifaddrs(ifa);
  return r;
}

std::string pick_iface(CephContext *cct, const struct sockaddr_storage &network)
{
  struct ifaddrs *ifa;
  int r = getifaddrs(&ifa);
  if (r < 0) {
    string err = cpp_strerror(errno);
    lderr(cct) << "unable to fetch interfaces and addresses: " << err << dendl;
    return {};
  }
  auto free_ifa = make_scope_guard([ifa] { freeifaddrs(ifa); });
  const unsigned int prefix_len = std::max(sizeof(in_addr::s_addr), sizeof(in6_addr::s6_addr)) * CHAR_BIT;
  for (auto addr = ifa; addr != nullptr; addr = addr->ifa_next) {
    if (matches_with_net(*ifa, (const struct sockaddr *) &network, prefix_len,
			 CEPH_PICK_ADDRESS_IPV4 | CEPH_PICK_ADDRESS_IPV6)) {
      return addr->ifa_name;
    }
  }
  return {};
}


bool have_local_addr(CephContext *cct, const std::list<entity_addr_t>& ls, entity_addr_t *match)
{
  struct ifaddrs *ifa;
  int r = getifaddrs(&ifa);
  if (r < 0) {
    lderr(cct) << "unable to fetch interfaces and addresses: " << cpp_strerror(errno) << dendl;
    exit(1);
  }
  auto free_ifa = make_scope_guard([ifa] { freeifaddrs(ifa); });

  for (struct ifaddrs *addrs = ifa; addrs != nullptr; addrs = addrs->ifa_next) {
    if (addrs->ifa_addr) {
      entity_addr_t a;
      a.set_sockaddr(addrs->ifa_addr);
      for (auto& p : ls) {
        if (a.is_same_host(p)) {
          *match = p;
          return true;
        }
      }
    }
  }
  return false;
}

int get_iface_numa_node(
  const std::string& iface,
  int *node)
{
  enum class iface_t {
    PHY_PORT,
    BOND_PORT
  } ifatype = iface_t::PHY_PORT;
  std::string_view ifa{iface};
  if (auto pos = ifa.find(":"); pos != ifa.npos) {
    ifa.remove_suffix(ifa.size() - pos);
  }
  string fn = fmt::format("/sys/class/net/{}/device/numa_node", ifa);
  int fd = ::open(fn.c_str(), O_RDONLY);
  if (fd < 0) {
    fn = fmt::format("/sys/class/net/{}/bonding/slaves", ifa);
    fd = ::open(fn.c_str(), O_RDONLY);
    if (fd < 0) {
      return -errno;
    }
    ifatype = iface_t::BOND_PORT;
  }

  int r = 0;
  char buf[1024];
  char *endptr = 0;
  r = safe_read(fd, &buf, sizeof(buf));
  if (r < 0) {
    goto out;
  }
  buf[r] = 0;
  while (r > 0 && ::isspace(buf[--r])) {
    buf[r] = 0;
  }

  switch (ifatype) {
  case iface_t::PHY_PORT:
    *node = strtoll(buf, &endptr, 10);
    if (endptr != buf + strlen(buf)) {
      r = -EINVAL;
      goto out;
    }
    r = 0;
    break;
  case iface_t::BOND_PORT:
    int bond_node = -1;
    std::vector<std::string> sv;
    std::string ifacestr = buf;
    get_str_vec(ifacestr, " ", sv);
    for (auto& iter : sv) {
      int bn = -1;
      r = get_iface_numa_node(iter, &bn);
      if (r >= 0) {
        if (bond_node == -1 || bn == bond_node) {
          bond_node = bn;
        } else {
          *node = -2;
          goto out;
        }
      } else {
        goto out;
      }
    }
    *node = bond_node;
    break;
  }

  out:
  ::close(fd);
  return r;
}
Commit	Line	Data
7c673cae FG	1	// -- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t --
	2	// vim: ts=8 sw=2 smarttab
	3	/*
	4	* Ceph - scalable distributed file system
	5	*
	6	* Copyright (C) 2004-2012 Inktank
	7	*
	8	* This is free software; you can redistribute it and/or
	9	* modify it under the terms of the GNU Lesser General Public
	10	* License version 2.1, as published by the Free Software
	11	* Foundation. See file COPYING.
	12	*
	13	*/
	14
	15	#include "common/pick_address.h"
9f95a23c TL	16
9f95a23c TL	17	#include <netdb.h>
522d829b	18	#include <netinet/in.h>
9f95a23c TL	19	#include <string>
	20	#include <string.h>
	21	#include <vector>
	22
522d829b	23	#include <boost/algorithm/string/predicate.hpp>
9f95a23c TL	24	#include <fmt/format.h>
9f95a23c TL	25
7c673cae FG	26	#include "include/ipaddr.h"
7c673cae FG	27	#include "include/str_list.h"
11fdf7f2 TL	28	#include "common/ceph_context.h"
	29	#ifndef WITH_SEASTAR
	30	#include "common/config.h"
	31	#include "common/config_obs.h"
	32	#endif
7c673cae FG	33	#include "common/debug.h"
7c673cae FG	34	#include "common/errno.h"
11fdf7f2	35	#include "common/numa.h"
7c673cae	36
522d829b TL	37	#ifndef HAVE_IN_ADDR_T
	38	typedef uint32_t in_addr_t;
	39	#endif
	40
	41	#ifndef IN_LOOPBACKNET
	42	#define IN_LOOPBACKNET 127
	43	#endif
	44
7c673cae FG	45	#define dout_subsys ceph_subsys_
7c673cae FG	46
f67539c2 TL	47	using std::string;
	48	using std::vector;
	49
522d829b TL	50	namespace {
	51
	52	bool matches_with_name(const ifaddrs& ifa, const std::string& if_name)
	53	{
	54	return if_name.compare(ifa.ifa_name) == 0;
	55	}
	56
	57	static int is_loopback_addr(sockaddr* addr)
	58	{
	59	if (addr->sa_family == AF_INET) {
	60	const sockaddr_in* sin = (struct sockaddr_in *)(addr);
	61	const in_addr_t net = ntohl(sin->sin_addr.s_addr) >> IN_CLASSA_NSHIFT;
	62	return net == IN_LOOPBACKNET ? 1 : 0;
	63	} else if (addr->sa_family == AF_INET6) {
	64	sockaddr_in6* sin6 = (struct sockaddr_in6 *)(addr);
	65	return IN6_IS_ADDR_LOOPBACK(&sin6->sin6_addr) ? 1 : 0;
	66	} else {
	67	return -1;
	68	}
	69	}
	70
	71	static int grade_addr(const ifaddrs& ifa)
	72	{
	73	if (ifa.ifa_addr == nullptr) {
	74	return -1;
	75	}
	76	int score = 0;
	77	if (ifa.ifa_flags & IFF_UP) {
	78	score += 4;
	79	}
	80	switch (is_loopback_addr(ifa.ifa_addr)) {
	81	case 0:
	82	// prefer non-loopback addresses
	83	score += 2;
	84	break;
	85	case 1:
	86	score += 0;
	87	break;
	88	default:
	89	score = -1;
	90	break;
	91	}
	92	return score;
	93	}
	94
	95	bool matches_with_net(const ifaddrs& ifa,
	96	const sockaddr* net,
	97	unsigned int prefix_len,
	98	unsigned ipv)
	99	{
	100	switch (net->sa_family) {
	101	case AF_INET:
	102	if (ipv & CEPH_PICK_ADDRESS_IPV4) {
	103	return matches_ipv4_in_subnet(ifa, (struct sockaddr_in*)net, prefix_len);
	104	}
	105	break;
	106	case AF_INET6:
	107	if (ipv & CEPH_PICK_ADDRESS_IPV6) {
	108	return matches_ipv6_in_subnet(ifa, (struct sockaddr_in6*)net, prefix_len);
	109	}
	110	break;
	111	}
	112	return false;
	113	}
114
115	bool matches_with_net(CephContext *cct,
116	const ifaddrs& ifa,
117	const std::string& s,
118	unsigned ipv)
119	{
120	struct sockaddr_storage net;
121	unsigned int prefix_len;
122	if (!parse_network(s.c_str(), &net, &prefix_len)) {
123	lderr(cct) << "unable to parse network: " << s << dendl;
124	exit(1);
125	}
126	return matches_with_net(ifa, (sockaddr*)&net, prefix_len, ipv);
127	}
128
129	int grade_with_numa_node(const ifaddrs& ifa, int numa_node)
130	{
131	#if defined(WITH_SEASTAR) \|\| defined(_WIN32)
132	return 0;
133	#else
134	if (numa_node < 0) {
135	return 0;
136	}
137	int if_node = -1;
138	int r = get_iface_numa_node(ifa.ifa_name, &if_node);
139	if (r < 0) {
140	return 0;
141	}
142	return if_node == numa_node ? 1 : 0;
143	#endif
144	}
145	}
146
3efd9988 FG	147	const struct sockaddr *find_ip_in_subnet_list(
	148	CephContext *cct,
	149	const struct ifaddrs *ifa,
11fdf7f2	150	unsigned ipv,
3efd9988	151	const std::string &networks,
11fdf7f2 TL	152	const std::string &interfaces,
11fdf7f2 TL	153	int numa_node)
7c673cae	154	{
522d829b TL	155	const auto ifs = get_str_list(interfaces);
	156	const auto nets = get_str_list(networks);
	157	if (!ifs.empty() && nets.empty()) {
3efd9988 FG	158	lderr(cct) << "interface names specified but not network names" << dendl;
3efd9988 FG	159	exit(1);
3efd9988	160	}
7c673cae	161
522d829b TL	162	int best_score = 0;
	163	const sockaddr* best_addr = nullptr;
	164	for (const auto* addr = ifa; addr != nullptr; addr = addr->ifa_next) {
	165	if (!ifs.empty() &&
	166	std::none_of(std::begin(ifs), std::end(ifs),
	167	[&](const auto& if_name) {
	168	return matches_with_name(*addr, if_name);
	169	})) {
	170	continue;
3efd9988	171	}
522d829b TL	172	if (!nets.empty() &&
	173	std::none_of(std::begin(nets), std::end(nets),
	174	[&](const auto& net) {
	175	return matches_with_net(cct, *addr, net, ipv);
	176	})) {
	177	continue;
11fdf7f2	178	}
522d829b TL	179	int score = grade_addr(*addr);
	180	if (score < 0) {
	181	continue;
3efd9988	182	}
522d829b TL	183	score += grade_with_numa_node(*addr, numa_node);
	184	if (score > best_score) {
	185	best_score = score;
	186	best_addr = addr->ifa_addr;
7c673cae	187	}
3efd9988	188	}
522d829b	189	return best_addr;
7c673cae FG	190	}
7c673cae FG	191
11fdf7f2	192	#ifndef WITH_SEASTAR
7c673cae FG	193	// observe this change
	194	struct Observer : public md_config_obs_t {
	195	const char *keys[2];
	196	explicit Observer(const char *c) {
	197	keys[0] = c;
	198	keys[1] = NULL;
	199	}
	200
	201	const char** get_tracked_conf_keys() const override {
	202	return (const char **)keys;
	203	}
11fdf7f2	204	void handle_conf_change(const ConfigProxy& conf,
7c673cae FG	205	const std::set <std::string> &changed) override {
	206	// do nothing.
	207	}
	208	};
	209
	210	static void fill_in_one_address(CephContext *cct,
	211	const struct ifaddrs *ifa,
522d829b TL	212	const string &networks,
522d829b TL	213	const string &interfaces,
11fdf7f2 TL	214	const char *conf_var,
11fdf7f2 TL	215	int numa_node = -1)
7c673cae	216	{
11fdf7f2 TL	217	const struct sockaddr *found = find_ip_in_subnet_list(
	218	cct,
	219	ifa,
	220	CEPH_PICK_ADDRESS_IPV4\|CEPH_PICK_ADDRESS_IPV6,
	221	networks,
	222	interfaces,
	223	numa_node);
7c673cae	224	if (!found) {
3efd9988 FG	225	lderr(cct) << "unable to find any IP address in networks '" << networks
3efd9988 FG	226	<< "' interfaces '" << interfaces << "'" << dendl;
7c673cae FG	227	exit(1);
	228	}
	229
	230	char buf[INET6_ADDRSTRLEN];
	231	int err;
	232
	233	err = getnameinfo(found,
	234	(found->sa_family == AF_INET)
	235	? sizeof(struct sockaddr_in)
	236	: sizeof(struct sockaddr_in6),
	237
	238	buf, sizeof(buf),
11fdf7f2	239	nullptr, 0,
7c673cae FG	240	NI_NUMERICHOST);
	241	if (err != 0) {
	242	lderr(cct) << "unable to convert chosen address to string: " << gai_strerror(err) << dendl;
	243	exit(1);
	244	}
	245
	246	Observer obs(conf_var);
	247
11fdf7f2	248	cct->_conf.add_observer(&obs);
7c673cae	249
11fdf7f2 TL	250	cct->_conf.set_val_or_die(conf_var, buf);
11fdf7f2 TL	251	cct->_conf.apply_changes(nullptr);
7c673cae	252
11fdf7f2	253	cct->_conf.remove_observer(&obs);
7c673cae FG	254	}
	255
	256	void pick_addresses(CephContext *cct, int needs)
	257	{
11fdf7f2 TL	258	auto public_addr = cct->_conf.get_val<entity_addr_t>("public_addr");
	259	auto public_network = cct->_conf.get_val<std::string>("public_network");
	260	auto public_network_interface =
	261	cct->_conf.get_val<std::string>("public_network_interface");
	262	auto cluster_addr = cct->_conf.get_val<entity_addr_t>("cluster_addr");
	263	auto cluster_network = cct->_conf.get_val<std::string>("cluster_network");
	264	auto cluster_network_interface =
	265	cct->_conf.get_val<std::string>("cluster_network_interface");
	266
522d829b TL	267	struct ifaddrs *ifa;
522d829b TL	268	int r = getifaddrs(&ifa);
11fdf7f2	269	if (r < 0) {
7c673cae FG	270	string err = cpp_strerror(errno);
	271	lderr(cct) << "unable to fetch interfaces and addresses: " << err << dendl;
	272	exit(1);
	273	}
522d829b	274	auto free_ifa = make_scope_guard([ifa] { freeifaddrs(ifa); });
11fdf7f2 TL	275	if ((needs & CEPH_PICK_ADDRESS_PUBLIC) &&
	276	public_addr.is_blank_ip() && !public_network.empty()) {
	277	fill_in_one_address(cct, ifa, public_network, public_network_interface,
522d829b	278	"public_addr");
7c673cae FG	279	}
7c673cae FG	280
11fdf7f2 TL	281	if ((needs & CEPH_PICK_ADDRESS_CLUSTER) && cluster_addr.is_blank_ip()) {
	282	if (!cluster_network.empty()) {
	283	fill_in_one_address(cct, ifa, cluster_network, cluster_network_interface,
522d829b	284	"cluster_addr");
7c673cae	285	} else {
11fdf7f2	286	if (!public_network.empty()) {
7c673cae FG	287	lderr(cct) << "Public network was set, but cluster network was not set " << dendl;
7c673cae FG	288	lderr(cct) << " Using public network also for cluster network" << dendl;
11fdf7f2	289	fill_in_one_address(cct, ifa, public_network, public_network_interface,
522d829b	290	"cluster_addr");
7c673cae FG	291	}
	292	}
	293	}
7c673cae	294	}
11fdf7f2 TL	295	#endif // !WITH_SEASTAR
	296
	297	static int fill_in_one_address(
	298	CephContext *cct,
	299	const struct ifaddrs *ifa,
	300	unsigned ipv,
522d829b TL	301	const string &networks,
522d829b TL	302	const string &interfaces,
11fdf7f2 TL	303	entity_addrvec_t *addrs,
	304	int numa_node = -1)
	305	{
522d829b TL	306	const struct sockaddr *found = find_ip_in_subnet_list(cct, ifa, ipv,
	307	networks,
	308	interfaces,
	309	numa_node);
11fdf7f2 TL	310	if (!found) {
	311	std::string ip_type = "";
	312	if ((ipv & CEPH_PICK_ADDRESS_IPV4) && (ipv & CEPH_PICK_ADDRESS_IPV6)) {
	313	ip_type = "IPv4 or IPv6";
	314	} else if (ipv & CEPH_PICK_ADDRESS_IPV4) {
	315	ip_type = "IPv4";
	316	} else {
	317	ip_type = "IPv6";
	318	}
	319	lderr(cct) << "unable to find any " << ip_type << " address in networks '"
	320	<< networks << "' interfaces '" << interfaces << "'" << dendl;
	321	return -1;
	322	}
	323
	324	char buf[INET6_ADDRSTRLEN];
	325	int err;
7c673cae	326
11fdf7f2 TL	327	err = getnameinfo(found,
	328	(found->sa_family == AF_INET)
	329	? sizeof(struct sockaddr_in)
	330	: sizeof(struct sockaddr_in6),
	331
	332	buf, sizeof(buf),
	333	nullptr, 0,
	334	NI_NUMERICHOST);
	335	if (err != 0) {
	336	lderr(cct) << "unable to convert chosen address to string: " << gai_strerror(err) << dendl;
	337	return -1;
	338	}
	339
	340	entity_addr_t addr;
	341	const char *end = 0;
	342	bool r = addr.parse(buf, &end);
	343	if (!r) {
	344	return -1;
	345	}
	346	addrs->v.push_back(addr);
	347	return 0;
	348	}
	349
	350	int pick_addresses(
	351	CephContext *cct,
	352	unsigned flags,
	353	struct ifaddrs *ifa,
	354	entity_addrvec_t *addrs,
	355	int preferred_numa_node)
	356	{
	357	addrs->v.clear();
	358
	359	unsigned addrt = (flags & (CEPH_PICK_ADDRESS_PUBLIC \|
	360	CEPH_PICK_ADDRESS_CLUSTER));
	361	if (addrt == 0 \|\|
	362	addrt == (CEPH_PICK_ADDRESS_PUBLIC \|
	363	CEPH_PICK_ADDRESS_CLUSTER)) {
	364	return -EINVAL;
	365	}
	366	unsigned msgrv = flags & (CEPH_PICK_ADDRESS_MSGR1 \|
	367	CEPH_PICK_ADDRESS_MSGR2);
	368	if (msgrv == 0) {
	369	if (cct->_conf.get_val<bool>("ms_bind_msgr1")) {
	370	msgrv \|= CEPH_PICK_ADDRESS_MSGR1;
	371	}
	372	if (cct->_conf.get_val<bool>("ms_bind_msgr2")) {
	373	msgrv \|= CEPH_PICK_ADDRESS_MSGR2;
	374	}
	375	if (msgrv == 0) {
	376	return -EINVAL;
	377	}
	378	}
	379	unsigned ipv = flags & (CEPH_PICK_ADDRESS_IPV4 \|
	380	CEPH_PICK_ADDRESS_IPV6);
	381	if (ipv == 0) {
	382	if (cct->_conf.get_val<bool>("ms_bind_ipv4")) {
	383	ipv \|= CEPH_PICK_ADDRESS_IPV4;
	384	}
	385	if (cct->_conf.get_val<bool>("ms_bind_ipv6")) {
	386	ipv \|= CEPH_PICK_ADDRESS_IPV6;
	387	}
	388	if (ipv == 0) {
	389	return -EINVAL;
	390	}
391	if (cct->_conf.get_val<bool>("ms_bind_prefer_ipv4")) {
392	flags \|= CEPH_PICK_ADDRESS_PREFER_IPV4;
393	} else {
394	flags &= ~CEPH_PICK_ADDRESS_PREFER_IPV4;
395	}
396	}
397
398	entity_addr_t addr;
399	string networks;
400	string interfaces;
401	if (addrt & CEPH_PICK_ADDRESS_PUBLIC) {
402	addr = cct->_conf.get_val<entity_addr_t>("public_addr");
403	networks = cct->_conf.get_val<std::string>("public_network");
404	interfaces =
405	cct->_conf.get_val<std::string>("public_network_interface");
406	} else {
407	addr = cct->_conf.get_val<entity_addr_t>("cluster_addr");
408	networks = cct->_conf.get_val<std::string>("cluster_network");
409	interfaces =
410	cct->_conf.get_val<std::string>("cluster_network_interface");
411	if (networks.empty()) {
412	lderr(cct) << "Falling back to public interface" << dendl;
413	// fall back to public_ network and interface if cluster is not set
414	networks = cct->_conf.get_val<std::string>("public_network");
415	interfaces =
416	cct->_conf.get_val<std::string>("public_network_interface");
417	}
418	}
419	if (addr.is_blank_ip() &&
420	!networks.empty()) {
421	int ipv4_r = !(ipv & CEPH_PICK_ADDRESS_IPV4) ? 0 : -1;
422	int ipv6_r = !(ipv & CEPH_PICK_ADDRESS_IPV6) ? 0 : -1;
522d829b TL	423	// note: pass in ipv to filter the matching addresses
	424	if ((ipv & CEPH_PICK_ADDRESS_IPV4) &&
	425	(flags & CEPH_PICK_ADDRESS_PREFER_IPV4)) {
	426	ipv4_r = fill_in_one_address(cct, ifa, CEPH_PICK_ADDRESS_IPV4,
	427	networks, interfaces,
	428	addrs,
	429	preferred_numa_node);
	430	}
	431	if (ipv & CEPH_PICK_ADDRESS_IPV6) {
	432	ipv6_r = fill_in_one_address(cct, ifa, CEPH_PICK_ADDRESS_IPV6,
	433	networks, interfaces,
	434	addrs,
	435	preferred_numa_node);
	436	}
	437	if ((ipv & CEPH_PICK_ADDRESS_IPV4) &&
	438	!(flags & CEPH_PICK_ADDRESS_PREFER_IPV4)) {
	439	ipv4_r = fill_in_one_address(cct, ifa, CEPH_PICK_ADDRESS_IPV4,
	440	networks, interfaces,
	441	addrs,
	442	preferred_numa_node);
	443	}
	444	if (ipv4_r < 0 \|\| ipv6_r < 0) {
	445	return -1;
11fdf7f2 TL	446	}
	447	}
	448
	449	// note: we may have a blank addr here
	450
	451	// ipv4 and/or ipv6?
	452	if (addrs->v.empty()) {
11fdf7f2 TL	453	addr.set_type(entity_addr_t::TYPE_MSGR2);
	454	if ((ipv & CEPH_PICK_ADDRESS_IPV4) &&
	455	(flags & CEPH_PICK_ADDRESS_PREFER_IPV4)) {
	456	addr.set_family(AF_INET);
	457	addrs->v.push_back(addr);
	458	}
	459	if (ipv & CEPH_PICK_ADDRESS_IPV6) {
	460	addr.set_family(AF_INET6);
	461	addrs->v.push_back(addr);
	462	}
	463	if ((ipv & CEPH_PICK_ADDRESS_IPV4) &&
	464	!(flags & CEPH_PICK_ADDRESS_PREFER_IPV4)) {
	465	addr.set_family(AF_INET);
	466	addrs->v.push_back(addr);
	467	}
	468	}
	469
	470	// msgr2 or legacy or both?
	471	if (msgrv == (CEPH_PICK_ADDRESS_MSGR1 \| CEPH_PICK_ADDRESS_MSGR2)) {
	472	vector<entity_addr_t> v;
	473	v.swap(addrs->v);
	474	for (auto a : v) {
	475	a.set_type(entity_addr_t::TYPE_MSGR2);
	476	if (flags & CEPH_PICK_ADDRESS_DEFAULT_MON_PORTS) {
	477	a.set_port(CEPH_MON_PORT_IANA);
	478	}
	479	addrs->v.push_back(a);
	480	a.set_type(entity_addr_t::TYPE_LEGACY);
	481	if (flags & CEPH_PICK_ADDRESS_DEFAULT_MON_PORTS) {
	482	a.set_port(CEPH_MON_PORT_LEGACY);
	483	}
	484	addrs->v.push_back(a);
	485	}
	486	} else if (msgrv == CEPH_PICK_ADDRESS_MSGR1) {
	487	for (auto& a : addrs->v) {
	488	a.set_type(entity_addr_t::TYPE_LEGACY);
	489	}
	490	} else {
	491	for (auto& a : addrs->v) {
	492	a.set_type(entity_addr_t::TYPE_MSGR2);
	493	}
	494	}
	495
	496	return 0;
	497	}
	498
	499	int pick_addresses(
	500	CephContext *cct,
	501	unsigned flags,
	502	entity_addrvec_t *addrs,
	503	int preferred_numa_node)
	504	{
	505	struct ifaddrs *ifa;
	506	int r = getifaddrs(&ifa);
	507	if (r < 0) {
	508	r = -errno;
	509	string err = cpp_strerror(r);
	510	lderr(cct) << "unable to fetch interfaces and addresses: "
	511	<< cpp_strerror(r) << dendl;
	512	return r;
	513	}
	514	r = pick_addresses(cct, flags, ifa, addrs, preferred_numa_node);
	515	freeifaddrs(ifa);
	516	return r;
517	}
b5b8bbf5 FG	518
	519	std::string pick_iface(CephContext *cct, const struct sockaddr_storage &network)
	520	{
	521	struct ifaddrs *ifa;
	522	int r = getifaddrs(&ifa);
	523	if (r < 0) {
	524	string err = cpp_strerror(errno);
	525	lderr(cct) << "unable to fetch interfaces and addresses: " << err << dendl;
	526	return {};
	527	}
522d829b	528	auto free_ifa = make_scope_guard([ifa] { freeifaddrs(ifa); });
f67539c2	529	const unsigned int prefix_len = std::max(sizeof(in_addr::s_addr), sizeof(in6_addr::s6_addr)) * CHAR_BIT;
522d829b TL	530	for (auto addr = ifa; addr != nullptr; addr = addr->ifa_next) {
	531	if (matches_with_net(ifa, (const struct sockaddr ) &network, prefix_len,
	532	CEPH_PICK_ADDRESS_IPV4 \| CEPH_PICK_ADDRESS_IPV6)) {
	533	return addr->ifa_name;
	534	}
b5b8bbf5	535	}
522d829b	536	return {};
b5b8bbf5 FG	537	}
	538
	539
f67539c2	540	bool have_local_addr(CephContext cct, const std::list<entity_addr_t>& ls, entity_addr_t match)
7c673cae FG	541	{
	542	struct ifaddrs *ifa;
	543	int r = getifaddrs(&ifa);
	544	if (r < 0) {
	545	lderr(cct) << "unable to fetch interfaces and addresses: " << cpp_strerror(errno) << dendl;
	546	exit(1);
	547	}
522d829b	548	auto free_ifa = make_scope_guard([ifa] { freeifaddrs(ifa); });
7c673cae	549
11fdf7f2	550	for (struct ifaddrs *addrs = ifa; addrs != nullptr; addrs = addrs->ifa_next) {
7c673cae FG	551	if (addrs->ifa_addr) {
	552	entity_addr_t a;
	553	a.set_sockaddr(addrs->ifa_addr);
11fdf7f2 TL	554	for (auto& p : ls) {
	555	if (a.is_same_host(p)) {
	556	*match = p;
522d829b	557	return true;
7c673cae FG	558	}
	559	}
	560	}
	561	}
522d829b	562	return false;
7c673cae	563	}
11fdf7f2 TL	564
	565	int get_iface_numa_node(
	566	const std::string& iface,
	567	int *node)
	568	{
9f95a23c TL	569	enum class iface_t {
	570	PHY_PORT,
	571	BOND_PORT
	572	} ifatype = iface_t::PHY_PORT;
f67539c2	573	std::string_view ifa{iface};
9f95a23c TL	574	if (auto pos = ifa.find(":"); pos != ifa.npos) {
9f95a23c TL	575	ifa.remove_suffix(ifa.size() - pos);
92f5a8d4	576	}
9f95a23c	577	string fn = fmt::format("/sys/class/net/{}/device/numa_node", ifa);
92f5a8d4 TL	578	int fd = ::open(fn.c_str(), O_RDONLY);
92f5a8d4 TL	579	if (fd < 0) {
9f95a23c	580	fn = fmt::format("/sys/class/net/{}/bonding/slaves", ifa);
92f5a8d4 TL	581	fd = ::open(fn.c_str(), O_RDONLY);
	582	if (fd < 0) {
	583	return -errno;
	584	}
9f95a23c	585	ifatype = iface_t::BOND_PORT;
92f5a8d4	586	}
11fdf7f2 TL	587
	588	int r = 0;
	589	char buf[1024];
	590	char *endptr = 0;
11fdf7f2 TL	591	r = safe_read(fd, &buf, sizeof(buf));
	592	if (r < 0) {
	593	goto out;
	594	}
	595	buf[r] = 0;
	596	while (r > 0 && ::isspace(buf[--r])) {
	597	buf[r] = 0;
	598	}
92f5a8d4 TL	599
92f5a8d4 TL	600	switch (ifatype) {
9f95a23c	601	case iface_t::PHY_PORT:
92f5a8d4 TL	602	*node = strtoll(buf, &endptr, 10);
	603	if (endptr != buf + strlen(buf)) {
	604	r = -EINVAL;
	605	goto out;
	606	}
	607	r = 0;
	608	break;
9f95a23c TL	609	case iface_t::BOND_PORT:
9f95a23c TL	610	int bond_node = -1;
92f5a8d4	611	std::vector<std::string> sv;
9f95a23c TL	612	std::string ifacestr = buf;
9f95a23c TL	613	get_str_vec(ifacestr, " ", sv);
92f5a8d4 TL	614	for (auto& iter : sv) {
	615	int bn = -1;
	616	r = get_iface_numa_node(iter, &bn);
	617	if (r >= 0) {
	618	if (bond_node == -1 \|\| bn == bond_node) {
	619	bond_node = bn;
	620	} else {
	621	*node = -2;
	622	goto out;
	623	}
	624	} else {
	625	goto out;
	626	}
	627	}
	628	*node = bond_node;
	629	break;
11fdf7f2	630	}
92f5a8d4 TL	631
92f5a8d4 TL	632	out:
11fdf7f2 TL	633	::close(fd);
	634	return r;
	635	}
92f5a8d4	636