rsm.cc

//
// Replicated state machine implementation with a primary and several
// backups. The primary receives requests, assigns each a view stamp (a
// vid, and a sequence number) in the order of reception, and forwards
// them to all backups. A backup executes requests in the order that
// the primary stamps them and replies with an OK to the primary. The
// primary executes the request after it receives OKs from all backups,
// and sends the reply back to the client.
//
// The config module will tell the RSM about a new view. If the
// primary in the previous view is a member of the new view, then it
// will stay the primary.  Otherwise, the smallest numbered node of
// the previous view will be the new primary.  In either case, the new
// primary will be a node from the previous view.  The configuration
// module constructs the sequence of views for the RSM and the RSM
// ensures there will be always one primary, who was a member of the
// last view.
//
// When a new node starts, the recovery thread is in charge of joining
// the RSM.  It will collect the internal RSM state from the primary;
// the primary asks the config module to add the new node and returns
// to the joining the internal RSM state (e.g., paxos log). Since
// there is only one primary, all joins happen in well-defined total
// order.
//
// The recovery thread also runs during a view change (e.g, when a node
// has failed).  After a failure some of the backups could have
// processed a request that the primary has not, but those results are
// not visible to clients (since the primary responds).  If the
// primary of the previous view is in the current view, then it will
// be the primary and its state is authoritive: the backups download
// from the primary the current state.  A primary waits until all
// backups have downloaded the state.  Once the RSM is in sync, the
// primary accepts requests again from clients.  If one of the backups
// is the new primary, then its state is authoritative.  In either
// scenario, the next view uses a node as primary that has the state
// resulting from processing all acknowledged client requests.
// Therefore, if the nodes sync up before processing the next request,
// the next view will have the correct state.
//
// While the RSM in a view change (i.e., a node has failed, a new view
// has been formed, but the sync hasn't completed), another failure
// could happen, which complicates a view change.  During syncing the
// primary or backups can timeout, and initiate another Paxos round.
// There are 2 variables that RSM uses to keep track in what state it
// is:
//    - inviewchange: a node has failed and the RSM is performing a
//                    view change
//    - insync: this node is syncing its state
//
// If inviewchange is false and a node is the primary, then it can
// process client requests. If it is true, clients are told to retry
// later again.  While inviewchange is true, the RSM may go through several
// member list changes, one by one.   After a member list
// change completes, the nodes tries to sync. If the sync complets,
// the view change completes (and inviewchange is set to false).  If
// the sync fails, the node may start another member list change
// (inviewchange = true and insync = false).
//
// The implementation should be used only with servers that run all
// requests run to completion; in particular, a request shouldn't
// block.  If a request blocks, the backup won't respond to the
// primary, and the primary won't execute the request.  A request may
// send an RPC to another host, but the RPC should be a one-way
// message to that host; the backup shouldn't do anything based on the
// response or execute after the response, because it is not
// guaranteed that all backup will receive the same response and
// execute in the same order.
//
// The implementation can be viewed as a layered system:
//       RSM module     ---- in charge of replication
//       config module  ---- in charge of view management
//       Paxos module   ---- in charge of running Paxos to agree on a value
//
// Each module has threads and internal locks. Furthermore, a thread
// may call down through the layers (e.g., to run Paxos's proposer).
// When Paxos's acceptor accepts a new value for an instance, a thread
// will invoke an upcall to inform higher layers of the new value.
// The rule is that a module releases its internal locks before it
// upcalls, but can keep its locks when calling down.

#include <fstream>
#include <iostream>
#include <unistd.h>

#include "handle.h"
#include "rsm.h"
#include "tprintf.h"
#include "lang/verify.h"
#include "rsm_client.h"

static void *
recoverythread(void *x)
{
  rsm *r = (rsm *) x;
  r->recovery();
  return 0;
}

rsm::rsm(std::string _first, std::string _me)
  : stf(0), primary(_first), insync(false), inviewchange(true), vid_commit(0),
    partitioned (false), dopartition(false), break1(false), break2(false)
{
  pthread_t th;

  last_myvs.vid = 0;
  last_myvs.seqno = 0;
  myvs = last_myvs;
  myvs.seqno = 1;

  pthread_mutex_init(&rsm_mutex, NULL);
  pthread_mutex_init(&invoke_mutex, NULL);
  pthread_cond_init(&recovery_cond, NULL);
  pthread_cond_init(&sync_cond, NULL);

  cfg = new config(_first, _me, this);

  if (_first == _me) {
    // Commit the first view here. We can not have acceptor::acceptor
    // do the commit, since at that time this->cfg is not initialized.
    commit_change(1);
  }
  rsmrpc = cfg->get_rpcs();
  rsmrpc->reg(rsm_client_protocol::invoke, this, &rsm::client_invoke);
  rsmrpc->reg(rsm_client_protocol::members, this, &rsm::client_members);
  rsmrpc->reg(rsm_protocol::invoke, this, &rsm::invoke);
  rsmrpc->reg(rsm_protocol::transferreq, this, &rsm::transferreq);
  rsmrpc->reg(rsm_protocol::transferdonereq, this, &rsm::transferdonereq);
  rsmrpc->reg(rsm_protocol::joinreq, this, &rsm::joinreq);

  // tester must be on different port, otherwise it may partition itself.
  testsvr = new rpcs(atoi(_me.c_str()) + 1);
  testsvr->reg(rsm_test_protocol::net_repair, this, &rsm::test_net_repairreq);
  testsvr->reg(rsm_test_protocol::breakpoint, this, &rsm::breakpointreq);

  {
    ScopedLock ml(&rsm_mutex);
    VERIFY(pthread_create(&th, NULL, &recoverythread, (void *) this) == 0);
  }
}

void
rsm::reg1(int proc, handler *h)
{
  ScopedLock ml(&rsm_mutex);
  procs[proc] = h;
}

// The recovery thread runs this function.
void
rsm::recovery()
{
  bool r = true;
  ScopedLock ml(&rsm_mutex);

  while (1) {
    while (!cfg->ismember(cfg->myaddr(), vid_commit)) {
      if (join(primary)) {
        tprintf("recovery: joined\n");
        commit_change_wo(cfg->vid());
      } else {
        VERIFY(pthread_mutex_unlock(&rsm_mutex) == 0);
        sleep(5); // XXX make another node in cfg primary?
        VERIFY(pthread_mutex_lock(&rsm_mutex) == 0);
      }
    }
    vid_insync = vid_commit;
    if (primary == cfg->myaddr()) {
      tprintf("recovery: sync vid_insync %d, i am primary\n", vid_insync);
      r = sync_with_backups();
    } else {
      tprintf("recovery: sync vid_insync %d, i am backup\n", vid_insync);
      r = sync_with_primary();
    }
    tprintf("recovery: sync done\n");

    // If there was a commited viewchange during the synchronization, restart
    // the recovery.
    if (vid_insync != vid_commit)
      continue;

    if (r) {
      myvs.vid = vid_commit;
      myvs.seqno = 1;
      inviewchange = false;
    }
    tprintf("recovery: go to sleep %d %d\n", insync, inviewchange);
    pthread_cond_wait(&recovery_cond, &rsm_mutex);
  }
}

bool
rsm::sync_with_backups()
{
  // TODO(fb): Is releasing rsm_mutex before acquiring invoke_mutex necessary?
  // pthread_mutex_unlock(&rsm_mutex);
  {
    // Make sure that the state of lock_server_cache_rsm is stable during
    // synchronization; otherwise, the primary's state may be more recent
    // than replicas after the synchronization.
    ScopedLock ml(&invoke_mutex);
    // By acquiring and releasing the invoke_mutex once, we make sure that
    // the state of lock_server_cache_rsm will not be changed until all
    // replicas are synchronized. The reason is that client_invoke arrives
    // after this point of time will see inviewchange == true, and returns
    // BUSY.
  }
  // pthread_mutex_lock(&rsm_mutex);

  // Start accepting synchronization request (statetransferreq) now!
  tprintf("sync_with_backups: insync is true now\n");
  insync = true;

  // Initialize backups.
  std::vector<std::string> members = cfg->get_view(vid_insync);
  backups.clear();
  backups.insert(members.begin(), members.end());
  backups.erase(cfg->myaddr());  // erase primary from backups

  if (!backups.empty()) {
    // Wait until
    //  * all backups in view vid_insync are synchronized;
    //  * or there is a committed viewchange.
    pthread_cond_wait(&recovery_cond, &rsm_mutex);
  }

  tprintf("sync_with_backups: insync is false now\n");
  insync = false;

  return true;
}

bool
rsm::sync_with_primary()
{
  // Remember the primary of vid_insync.
  std::string m = primary;

  // Keep synchronizing with primary until the synchronization succeeds,
  // or there is a commited viewchange.
  while (true) {
    if (vid_commit != vid_insync) {
      return false;
    }
    if (statetransfer(m)) {
      break;
    }
  }
  if (vid_commit != vid_insync) {
    return false;
  }
  statetransferdone(m);
  return true;
}

// Call to transfer state from m to the local node.
// Assumes that rsm_mutex is already held.
bool
rsm::statetransfer(std::string m)
{
  rsm_protocol::transferres r;
  handle h(m);
  rsm_protocol::status ret;

  tprintf("rsm::statetransfer: contact %s w. my last_myvs(%d,%d)\n",
          m.c_str(), last_myvs.vid, last_myvs.seqno);

  VERIFY(pthread_mutex_unlock(&rsm_mutex) == 0);
  rpcc *cl = h.safebind();
  if (cl) {
    ret = cl->call(rsm_protocol::transferreq, cfg->myaddr(), last_myvs,
                   vid_insync, r, rpcc::to(1000));
  }
  VERIFY(pthread_mutex_lock(&rsm_mutex) == 0);

  if (cl == NULL || ret != rsm_protocol::OK) {
    tprintf("rsm::statetransfer: couldn't reach %s %lx %d\n", m.c_str(),
            (long unsigned) cl, ret);
    return false;
  }
  if (stf && last_myvs != r.last) {
    stf->unmarshal_state(r.state);
  }
  last_myvs = r.last;
  tprintf("rsm::statetransfer transfer from %s success, vs(%d,%d)\n",
          m.c_str(), last_myvs.vid, last_myvs.seqno);
  return true;
}

// Inform primary that this slave has synchronized for vid_insync.
// Assumes that rsm_mutex is already held.
bool
rsm::statetransferdone(std::string m)
{
  int r;
  handle h(m);
  rsm_protocol::status ret;

  tprintf("rsm::statetransferdone: contact %s w. my last_myvs(%d,%d)\n",
          m.c_str(), last_myvs.vid, last_myvs.seqno);

  VERIFY(pthread_mutex_unlock(&rsm_mutex) == 0);
  rpcc *cl = h.safebind();
  if (cl) {
    ret = cl->call(rsm_protocol::transferdonereq, cfg->myaddr(), vid_insync,
                   r, rpcc::to(1000));
  }
  VERIFY(pthread_mutex_lock(&rsm_mutex) == 0);

  if (cl == NULL || ret != rsm_protocol::OK) {
    tprintf("rsm::statetransferdone: couldn't reach %s %lx %d\n", m.c_str(),
            (long unsigned) cl, ret);
    return false;
  }

  tprintf("rsm::statetransferdone: %s success\n", m.c_str());

  return true;
}

bool
rsm::join(std::string m)
{
  handle h(m);
  int ret;
  rsm_protocol::joinres r;

  tprintf("rsm::join: %s mylast (%d,%d)\n", m.c_str(), last_myvs.vid,
          last_myvs.seqno);

  VERIFY(pthread_mutex_unlock(&rsm_mutex) == 0);
  rpcc *cl = h.safebind();
  if (cl != 0) {
    ret = cl->call(rsm_protocol::joinreq, cfg->myaddr(), last_myvs, r, rpcc::to(120000));
  }
  VERIFY(pthread_mutex_lock(&rsm_mutex) == 0);

  if (cl == 0 || ret != rsm_protocol::OK) {
    tprintf("rsm::join: couldn't reach %s %p %d\n", m.c_str(), cl, ret);
    return false;
  }
  tprintf("rsm::join: succeeded %s\n", r.log.c_str());
  cfg->restore(r.log);
  return true;
}

// Config informs rsm whenever it has successfully completed a view change.
void
rsm::commit_change(unsigned vid)
{
  ScopedLock ml(&rsm_mutex);
  commit_change_wo(vid);
}

void
rsm::commit_change_wo(unsigned vid)
{
  if (vid <= vid_commit)
    return;

  tprintf("commit_change: new view (%d)  last vs (%d,%d) %s insync %d\n",
          vid, last_myvs.vid, last_myvs.seqno, primary.c_str(), insync);

  vid_commit = vid;
  inviewchange = true;
  set_primary(vid);
  pthread_cond_signal(&recovery_cond);

  if (cfg->ismember(cfg->myaddr(), vid_commit))
    breakpoint2();
}

void
rsm::execute(int procno, std::string req, std::string &r)
{
  tprintf("execute\n");

  handler *h = procs[procno];
  VERIFY(h);
  unmarshall args(req);
  marshall rep;
  std::string reps;
  rsm_protocol::status ret = h->fn(args, rep);
  marshall rep1;
  rep1 << ret;
  rep1 << rep.str();
  r = rep1.str();
}

//
// Clients call client_invoke to invoke a procedure on the replicated state
// machine: the primary receives the request, assigns it a sequence
// number, and invokes it on all members of the replicated state
// machine.
//
rsm_client_protocol::status
rsm::client_invoke(int procno, std::string req, std::string &r)
{
  viewstamp vs;
  std::vector<std::string> members;

  tprintf("client_invoke of procno = %d.\n", procno);

  pthread_mutex_lock(&rsm_mutex);

  if (inviewchange) {
    tprintf("client_invoke: in a view change.\n");
    pthread_mutex_unlock(&rsm_mutex);
    return rsm_client_protocol::BUSY;
  }

  if (primary != cfg->myaddr()) {
    tprintf("client_invoke: not a primary.\n");
    pthread_mutex_unlock(&rsm_mutex);
    return rsm_client_protocol::NOTPRIMARY;
  }

  {
    ScopedLock ml(&invoke_mutex);
    int dummy_r;

    // We are definitely master (primary).
    vs = myvs;
    last_myvs = myvs;
    myvs.seqno += 1;

    members = cfg->get_view(vs.vid);

    // Release rsm_mutex once we have got invoke_mutex.
    pthread_mutex_unlock(&rsm_mutex);

    for (const std::string &member : members) {
      if (member == cfg->myaddr()) {
        continue;
      }

      handle h(member);
      rpcc *cl = h.safebind();

      if (cl == NULL ||
          cl->call(rsm_protocol::invoke, procno, vs, req, dummy_r, rpcc::to(1000)) != rsm_protocol::OK) {
        tprintf("client_invoke: failed to invoke slave %s.\n", member.c_str());
        return rsm_client_protocol::BUSY;
      }

      breakpoint1();
      partition1();
    }

    // Execute the request on master.
    // FIXME (fb): execute should be protected in rsm_mutex.
    // Consider refactoring execute by passing procs[procno] instead.
    execute(procno, req, r);
  }

  return rsm_client_protocol::OK;
}

//
// The primary calls the internal invoke at each member of the
// replicated state machine.
//
// the replica must execute requests in order (with no gaps)
// according to requests' seqno.
//
rsm_protocol::status
rsm::invoke(int proc, viewstamp vs, std::string req, int &)
{
  ScopedLock ml(&rsm_mutex);
  std::string r;

  if (vs != myvs) {
    tprintf("invoke: not expected viewstamp.\n");
    return rsm_protocol::ERR;
  }

  if (inviewchange) {
    tprintf("invoke: in a view change.\n");
    return rsm_protocol::BUSY;
  }

  if (!cfg->ismember(cfg->myaddr(), vs.vid)) {
    tprintf("invoke: not a member.\n");
    return rsm_protocol::ERR;
  }

  last_myvs = vs;
  myvs.seqno += 1;

  execute(proc, req, r);

  breakpoint1();

  return rsm_protocol::OK;
}

// RPC handler: Send back the local node's state to the caller.
rsm_protocol::status
rsm::transferreq(std::string src, viewstamp last, unsigned vid,
                 rsm_protocol::transferres &r)
{
  ScopedLock ml(&rsm_mutex);
  int ret = rsm_protocol::OK;

  tprintf("transferreq from %s (%d,%d) vs (%d,%d), insync %d\n", src.c_str(),
          last.vid, last.seqno, last_myvs.vid, last_myvs.seqno, (int) insync);

  if (!insync || vid != vid_insync) {
     return rsm_protocol::BUSY;
  }
  if (stf && last != last_myvs)
    r.state = stf->marshal_state();
  r.last = last_myvs;
  return ret;
}

// RPC handler: Inform the local node (the primary) that node m
// has synchronized.
rsm_protocol::status
rsm::transferdonereq(std::string m, unsigned vid, int &)
{
  ScopedLock ml(&rsm_mutex);

  // 1. Return BUSY if I am not insync, or if the slave is not synchronizing
  //    for the same view with me.
  // 2. Remove the slave from the list of unsynchronized backups.
  // 3. Wake up recovery thread if all backups are synchronized.

  if (!insync || vid != vid_insync) {
    return rsm_protocol::BUSY;
  }

  backups.erase(m);

  if (backups.empty()) {
    pthread_cond_signal(&recovery_cond);
  }

  return rsm_protocol::OK;
}

// a node that wants to join an RSM as a server sends a
// joinreq to the RSM's current primary; this is the
// handler for that RPC.
rsm_protocol::status
rsm::joinreq(std::string m, viewstamp last, rsm_protocol::joinres &r)
{
  int ret = rsm_protocol::OK;

  ScopedLock ml(&rsm_mutex);

  tprintf("joinreq: src %s last (%d,%d) mylast (%d,%d)\n", m.c_str(),
          last.vid, last.seqno, last_myvs.vid, last_myvs.seqno);

  if (cfg->ismember(m, vid_commit)) {
    tprintf("joinreq: is still a member\n");
    r.log = cfg->dump();
  } else if (cfg->myaddr() != primary) {
    tprintf("joinreq: busy\n");
    ret = rsm_protocol::BUSY;
  } else {
    // We cache vid_commit to avoid adding m to a view which already contains
    // m due to race condition.
    unsigned vid_cache = vid_commit;

    VERIFY(pthread_mutex_unlock(&rsm_mutex) == 0);
    bool succ = cfg->add(m, vid_cache);
    VERIFY(pthread_mutex_lock(&rsm_mutex) == 0);

    if (cfg->ismember(m, cfg->vid())) {
      r.log = cfg->dump();
      tprintf("joinreq: ret %d log %s\n:", ret, r.log.c_str());
    } else {
      tprintf("joinreq: failed; proposer couldn't add %d\n", succ);
      ret = rsm_protocol::BUSY;
    }
  }
  return ret;
}

// RPC handler: Send back all the nodes this local knows about to client
// so the client can switch to a different primary when it existing primary
// fails.
rsm_client_protocol::status
rsm::client_members(int i, std::vector<std::string> &r)
{
  ScopedLock ml(&rsm_mutex);

  std::vector<std::string> m = cfg->get_view(vid_commit);
  m.push_back(primary);
  r = m;
  tprintf("rsm::client_members return %s m %s\n", print_members(m).c_str(),
          primary.c_str());
  return rsm_client_protocol::OK;
}

// If primary is member of new view, that node is primary;
// otherwise, the lowest number node of the previous view.
// caller should hold rsm_mutex.
void
rsm::set_primary(unsigned vid)
{
  std::vector<std::string> c = cfg->get_view(vid);
  std::vector<std::string> p = cfg->get_view(vid - 1);
  VERIFY(c.size() > 0);

  if (isamember(primary, c)) {
    tprintf("set_primary: primary stays %s\n", primary.c_str());
    return;
  }

  VERIFY(p.size() > 0);
  for (unsigned i = 0; i < p.size(); i++) {
    if (isamember(p[i], c)) {
      primary = p[i];
      tprintf("set_primary: primary is %s\n", primary.c_str());
      return;
    }
  }
  VERIFY(0);
}

bool
rsm::amiprimary()
{
  ScopedLock ml(&rsm_mutex);
  return primary == cfg->myaddr() && !inviewchange;
}


// Testing server.

// Simulate partitions.

// Assumes caller holds rsm_mutex.
void
rsm::net_repair_wo(bool heal)
{
  std::vector<std::string> m;
  m = cfg->get_view(vid_commit);
  for (unsigned i  = 0; i < m.size(); i++) {
    if (m[i] != cfg->myaddr()) {
      handle h(m[i]);
      tprintf("rsm::net_repair_wo: %s %d\n", m[i].c_str(), heal);
      if (h.safebind())
        h.safebind()->set_reachable(heal);
    }
  }
  rsmrpc->set_reachable(heal);
}

rsm_test_protocol::status
rsm::test_net_repairreq(int heal, int &r)
{
  ScopedLock ml(&rsm_mutex);

  tprintf("rsm::test_net_repairreq: %d (dopartition %d, partitioned %d)\n",
          heal, dopartition, partitioned);

  if (heal) {
    net_repair_wo(heal);
    partitioned = false;
  } else {
    dopartition = true;
    partitioned = false;
  }
  r = rsm_test_protocol::OK;
  return r;
}

// Simulate failure at breakpoint 1 and 2.

void
rsm::breakpoint1()
{
  if (break1) {
    tprintf("Dying at breakpoint 1 in rsm!\n");
    exit(1);
  }
}

void
rsm::breakpoint2()
{
  if (break2) {
    tprintf("Dying at breakpoint 2 in rsm!\n");
    exit(1);
  }
}

void
rsm::partition1()
{
  if (dopartition) {
    net_repair_wo(false);
    dopartition = false;
    partitioned = true;
  }
}

rsm_test_protocol::status
rsm::breakpointreq(int b, int &r)
{
  r = rsm_test_protocol::OK;
  ScopedLock ml(&rsm_mutex);
  tprintf("rsm::breakpointreq: %d\n", b);
  if (b == 1) break1 = true;
  else if (b == 2) break2 = true;
  else if (b == 3 || b == 4) cfg->breakpoint(b);
  else r = rsm_test_protocol::ERR;
  return r;
}