2 // Replicated state machine implementation with a primary and several
3 // backups. The primary receives requests, assigns each a view stamp (a
4 // vid, and a sequence number) in the order of reception, and forwards
5 // them to all backups. A backup executes requests in the order that
6 // the primary stamps them and replies with an OK to the primary. The
7 // primary executes the request after it receives OKs from all backups,
8 // and sends the reply back to the client.
10 // The config module will tell the RSM about a new view. If the
11 // primary in the previous view is a member of the new view, then it
12 // will stay the primary. Otherwise, the smallest numbered node of
13 // the previous view will be the new primary. In either case, the new
14 // primary will be a node from the previous view. The configuration
15 // module constructs the sequence of views for the RSM and the RSM
16 // ensures there will be always one primary, who was a member of the
19 // When a new node starts, the recovery thread is in charge of joining
20 // the RSM. It will collect the internal RSM state from the primary;
21 // the primary asks the config module to add the new node and returns
22 // to the joining the internal RSM state (e.g., paxos log). Since
23 // there is only one primary, all joins happen in well-defined total
26 // The recovery thread also runs during a view change (e.g, when a node
27 // has failed). After a failure some of the backups could have
28 // processed a request that the primary has not, but those results are
29 // not visible to clients (since the primary responds). If the
30 // primary of the previous view is in the current view, then it will
31 // be the primary and its state is authoritive: the backups download
32 // from the primary the current state. A primary waits until all
33 // backups have downloaded the state. Once the RSM is in sync, the
34 // primary accepts requests again from clients. If one of the backups
35 // is the new primary, then its state is authoritative. In either
36 // scenario, the next view uses a node as primary that has the state
37 // resulting from processing all acknowledged client requests.
38 // Therefore, if the nodes sync up before processing the next request,
39 // the next view will have the correct state.
41 // While the RSM in a view change (i.e., a node has failed, a new view
42 // has been formed, but the sync hasn't completed), another failure
43 // could happen, which complicates a view change. During syncing the
44 // primary or backups can timeout, and initiate another Paxos round.
45 // There are 2 variables that RSM uses to keep track in what state it
47 // - inviewchange: a node has failed and the RSM is performing a view change
48 // - insync: this node is syncing its state
50 // If inviewchange is false and a node is the primary, then it can
51 // process client requests. If it is true, clients are told to retry
52 // later again. While inviewchange is true, the RSM may go through several
53 // member list changes, one by one. After a member list
54 // change completes, the nodes tries to sync. If the sync complets,
55 // the view change completes (and inviewchange is set to false). If
56 // the sync fails, the node may start another member list change
57 // (inviewchange = true and insync = false).
59 // The implementation should be used only with servers that run all
60 // requests run to completion; in particular, a request shouldn't
61 // block. If a request blocks, the backup won't respond to the
62 // primary, and the primary won't execute the request. A request may
63 // send an RPC to another host, but the RPC should be a one-way
64 // message to that host; the backup shouldn't do anything based on the
65 // response or execute after the response, because it is not
66 // guaranteed that all backup will receive the same response and
67 // execute in the same order.
69 // The implementation can be viewed as a layered system:
70 // RSM module ---- in charge of replication
71 // config module ---- in charge of view management
72 // Paxos module ---- in charge of running Paxos to agree on a value
74 // Each module has threads and internal locks. Furthermore, a thread
75 // may call down through the layers (e.g., to run Paxos's proposer).
76 // When Paxos's acceptor accepts a new value for an instance, a thread
77 // will invoke an upcall to inform higher layers of the new value.
78 // The rule is that a module releases its internal locks before it
79 // upcalls, but can keep its locks when calling down.
81 #include <sys/types.h>
87 #include "rsm_client.h"
89 rsm::rsm(const string & _first, const string & _me) :
90 stf(0), primary(_first), insync (false), inviewchange (true), vid_commit(0),
91 partitioned (false), dopartition(false), break1(false), break2(false)
93 cfg = unique_ptr<config>(new config(_first, _me, this));
96 // Commit the first view here. We can not have acceptor::acceptor
97 // do the commit, since at that time this->cfg is not initialized
100 rsmrpc = cfg->get_rpcs();
101 rsmrpc->reg(rsm_client_protocol::invoke, &rsm::client_invoke, this);
102 rsmrpc->reg(rsm_client_protocol::members, &rsm::client_members, this);
103 rsmrpc->reg(rsm_protocol::invoke, &rsm::invoke, this);
104 rsmrpc->reg(rsm_protocol::transferreq, &rsm::transferreq, this);
105 rsmrpc->reg(rsm_protocol::transferdonereq, &rsm::transferdonereq, this);
106 rsmrpc->reg(rsm_protocol::joinreq, &rsm::joinreq, this);
108 // tester must be on different port, otherwise it may partition itself
109 testsvr = unique_ptr<rpcs>(new rpcs((in_port_t)stoi(_me) + 1));
110 testsvr->reg(rsm_test_protocol::net_repair, &rsm::test_net_repairreq, this);
111 testsvr->reg(rsm_test_protocol::breakpoint, &rsm::breakpointreq, this);
118 thread(&rsm::recovery, this).detach();
121 void rsm::reg1(int proc, handler *h) {
126 // The recovery thread runs this function
127 void rsm::recovery() [[noreturn]] {
132 while (!cfg->ismember(cfg->myaddr(), vid_commit)) {
133 // XXX iannucci 2013/09/15 -- I don't understand whether accessing
134 // cfg->view_id in this manner involves a race. I suspect not.
135 if (join(primary, ml)) {
136 LOG("recovery: joined");
137 commit_change(cfg->view_id(), ml);
140 this_thread::sleep_for(seconds(3)); // XXX make another node in cfg primary?
144 vid_insync = vid_commit;
145 LOG("recovery: sync vid_insync " << vid_insync);
146 if (primary == cfg->myaddr()) {
147 r = sync_with_backups(ml);
149 r = sync_with_primary(ml);
151 LOG("recovery: sync done");
153 // If there was a commited viewchange during the synchronization, restart
155 if (vid_insync != vid_commit)
159 myvs.vid = vid_commit;
161 inviewchange = false;
163 LOG("recovery: go to sleep " << insync << " " << inviewchange);
164 recovery_cond.wait(ml);
168 bool rsm::sync_with_backups(lock & rsm_mutex_lock) {
169 rsm_mutex_lock.unlock();
171 // Make sure that the state of lock_server is stable during
172 // synchronization; otherwise, the primary's state may be more recent
173 // than replicas after the synchronization.
174 lock invoke_mutex_lock(invoke_mutex);
175 // By acquiring and releasing the invoke_mutex once, we make sure that
176 // the state of lock_server will not be changed until all
177 // replicas are synchronized. The reason is that client_invoke arrives
178 // after this point of time will see inviewchange == true, and returns
181 rsm_mutex_lock.lock();
182 // Start accepting synchronization request (statetransferreq) now!
184 cfg->get_view(vid_insync, backups);
185 backups.erase(find(backups.begin(), backups.end(), cfg->myaddr()));
186 LOG("backups " << backups);
187 sync_cond.wait(rsm_mutex_lock);
193 bool rsm::sync_with_primary(lock & rsm_mutex_lock) {
194 // Remember the primary of vid_insync
196 while (vid_insync == vid_commit) {
197 if (statetransfer(m, rsm_mutex_lock))
200 return statetransferdone(m, rsm_mutex_lock);
205 // Call to transfer state from m to the local node.
206 // Assumes that rsm_mutex is already held.
208 bool rsm::statetransfer(const string & m, lock & rsm_mutex_lock)
210 rsm_protocol::transferres r;
213 LOG("contact " << m << " w. my last_myvs(" << last_myvs.vid << "," << last_myvs.seqno << ")");
216 rsm_mutex_lock.unlock();
219 ret = cl->call_timeout(rsm_protocol::transferreq, milliseconds(100),
220 r, cfg->myaddr(), last_myvs, vid_insync);
222 rsm_mutex_lock.lock();
224 if (cl == 0 || ret != rsm_protocol::OK) {
225 LOG("couldn't reach " << m << " " << hex << cl << " " << dec << ret);
228 if (stf && last_myvs != r.last) {
229 stf->unmarshal_state(r.state);
232 LOG("transfer from " << m << " success, vs(" << last_myvs.vid << "," << last_myvs.seqno << ")");
236 bool rsm::statetransferdone(const string & m, lock & rsm_mutex_lock) {
237 rsm_mutex_lock.unlock();
239 rpcc *cl = h.safebind();
243 auto ret = (rsm_protocol::status)cl->call(rsm_protocol::transferdonereq, r, cfg->myaddr(), vid_insync);
244 done = (ret == rsm_protocol::OK);
246 rsm_mutex_lock.lock();
251 bool rsm::join(const string & m, lock & rsm_mutex_lock) {
256 LOG("contacting " << m << " mylast (" << last_myvs.vid << "," << last_myvs.seqno << ")");
259 rsm_mutex_lock.unlock();
262 ret = cl->call_timeout(rsm_protocol::joinreq, milliseconds(12000), log,
263 cfg->myaddr(), last_myvs);
265 rsm_mutex_lock.lock();
268 if (cl == 0 || ret != rsm_protocol::OK) {
269 LOG("couldn't reach " << m << " " << hex << cl << " " << dec << ret);
272 LOG("succeeded " << log);
278 // Config informs rsm whenever it has successfully
279 // completed a view change
281 void rsm::commit_change(unsigned vid) {
283 commit_change(vid, ml);
284 if (cfg->ismember(cfg->myaddr(), vid_commit))
288 void rsm::commit_change(unsigned vid, lock &) {
289 if (vid <= vid_commit)
291 LOG("commit_change: new view (" << vid << ") last vs (" << last_myvs.vid << "," <<
292 last_myvs.seqno << ") " << primary << " insync " << insync);
296 recovery_cond.notify_one();
297 sync_cond.notify_one();
298 if (cfg->ismember(cfg->myaddr(), vid_commit))
303 void rsm::execute(int procno, const string & req, string & r) {
305 handler *h = procs[procno];
307 unmarshall args(req, false);
309 auto ret = (rsm_protocol::status)(*h)(args, rep);
310 r = marshall{ret, rep.content()}.content();
314 // Clients call client_invoke to invoke a procedure on the replicated state
315 // machine: the primary receives the request, assigns it a sequence
316 // number, and invokes it on all members of the replicated state
319 rsm_client_protocol::status rsm::client_invoke(string & r, int procno, const string & req) {
320 LOG("invoke procno 0x" << hex << procno);
321 lock ml(invoke_mutex);
327 LOG("Checking for inviewchange");
329 return rsm_client_protocol::BUSY;
330 LOG("Checking for primacy");
331 myaddr = cfg->myaddr();
332 if (primary != myaddr)
333 return rsm_client_protocol::NOTPRIMARY;
334 LOG("Assigning a viewstamp");
335 cfg->get_view(vid_commit, m);
336 // assign the RPC the next viewstamp number
341 // send an invoke RPC to all slaves in the current view with a timeout of 1 second
342 LOG("Invoking slaves");
343 for (unsigned i = 0; i < m.size(); i++) {
344 if (m[i] != myaddr) {
345 // if invoke on slave fails, return rsm_client_protocol::BUSY
347 LOG("Sending invoke to " << m[i]);
348 rpcc *cl = h.safebind();
350 return rsm_client_protocol::BUSY;
352 auto ret = (rsm_protocol::status)cl->call_timeout(rsm_protocol::invoke, milliseconds(100), ignored_rval, procno, vs, req);
353 LOG("Invoke returned " << ret);
354 if (ret != rsm_protocol::OK)
355 return rsm_client_protocol::BUSY;
357 lock rsm_mutex_lock(rsm_mutex);
358 partition1(rsm_mutex_lock);
361 execute(procno, req, r);
363 return rsm_client_protocol::OK;
367 // The primary calls the internal invoke at each member of the
368 // replicated state machine
370 // the replica must execute requests in order (with no gaps)
371 // according to requests' seqno
373 rsm_protocol::status rsm::invoke(int &, int proc, viewstamp vs, const string & req) {
374 LOG("invoke procno 0x" << hex << proc);
375 lock ml(invoke_mutex);
380 // check if !inviewchange
381 LOG("Checking for view change");
383 return rsm_protocol::ERR;
385 LOG("Checking for slave status");
386 myaddr = cfg->myaddr();
387 if (primary == myaddr)
388 return rsm_protocol::ERR;
389 cfg->get_view(vid_commit, m);
390 if (find(m.begin(), m.end(), myaddr) == m.end())
391 return rsm_protocol::ERR;
392 // check sequence number
393 LOG("Checking sequence number");
395 return rsm_protocol::ERR;
399 execute(proc, req, r);
402 return rsm_protocol::OK;
406 // RPC handler: Send back the local node's state to the caller
408 rsm_protocol::status rsm::transferreq(rsm_protocol::transferres &r, const string & src,
409 viewstamp last, unsigned vid) {
411 LOG("transferreq from " << src << " (" << last.vid << "," << last.seqno << ") vs (" <<
412 last_myvs.vid << "," << last_myvs.seqno << ")");
413 if (!insync || vid != vid_insync)
414 return rsm_protocol::BUSY;
415 if (stf && last != last_myvs)
416 r.state = stf->marshal_state();
418 return rsm_protocol::OK;
422 // RPC handler: Inform the local node (the primary) that node m has synchronized
425 rsm_protocol::status rsm::transferdonereq(int &, const string & m, unsigned vid) {
427 if (!insync || vid != vid_insync)
428 return rsm_protocol::BUSY;
429 backups.erase(find(backups.begin(), backups.end(), m));
431 sync_cond.notify_one();
432 return rsm_protocol::OK;
435 // a node that wants to join an RSM as a server sends a
436 // joinreq to the RSM's current primary; this is the
437 // handler for that RPC.
438 rsm_protocol::status rsm::joinreq(string & log, const string & m, viewstamp last) {
439 auto ret = rsm_protocol::OK;
442 LOG("join request from " << m << "; last=(" << last.vid << "," << last.seqno << "), mylast=(" <<
443 last_myvs.vid << "," << last_myvs.seqno << ")");
444 if (cfg->ismember(m, vid_commit)) {
445 LOG(m << " is still a member -- nothing to do");
447 } else if (cfg->myaddr() != primary) {
448 LOG("but I, " << cfg->myaddr() << ", am not the primary, " << primary << "!");
449 ret = rsm_protocol::BUSY;
451 // We cache vid_commit to avoid adding m to a view which already contains
452 // m due to race condition
453 LOG("calling down to config layer");
454 unsigned vid_cache = vid_commit;
458 succ = cfg->add(m, vid_cache);
461 if (cfg->ismember(m, cfg->view_id())) {
463 LOG("ret " << ret << " log " << log);
465 LOG("failed; proposer couldn't add " << succ);
466 ret = rsm_protocol::BUSY;
473 // RPC handler: Send back all the nodes this local knows about to client
474 // so the client can switch to a different primary
475 // when it existing primary fails
477 rsm_client_protocol::status rsm::client_members(vector<string> &r, int) {
480 cfg->get_view(vid_commit, m);
481 m.push_back(primary);
483 LOG("return " << m << " m " << primary);
484 return rsm_client_protocol::OK;
487 // if primary is member of new view, that node is primary
488 // otherwise, the lowest number node of the previous view.
489 // caller should hold rsm_mutex
490 void rsm::set_primary(unsigned vid) {
492 cfg->get_view(vid, c);
493 cfg->get_view(vid - 1, p);
494 VERIFY (c.size() > 0);
496 if (isamember(primary,c)) {
497 LOG("set_primary: primary stays " << primary);
501 VERIFY(p.size() > 0);
502 for (unsigned i = 0; i < p.size(); i++) {
503 if (isamember(p[i], c)) {
505 LOG("set_primary: primary is " << primary);
512 bool rsm::amiprimary() {
514 return primary == cfg->myaddr() && !inviewchange;
520 // Simulate partitions
522 // assumes caller holds rsm_mutex
523 void rsm::net_repair(bool heal, lock &) {
525 cfg->get_view(vid_commit, m);
526 for (unsigned i = 0; i < m.size(); i++) {
527 if (m[i] != cfg->myaddr()) {
529 LOG("member " << m[i] << " " << heal);
530 if (h.safebind()) h.safebind()->set_reachable(heal);
533 rsmrpc->set_reachable(heal);
536 rsm_test_protocol::status rsm::test_net_repairreq(rsm_test_protocol::status &r, int heal) {
538 LOG("heal " << heal << " (dopartition " <<
539 dopartition << ", partitioned " << partitioned << ")");
541 net_repair(heal, ml);
547 r = rsm_test_protocol::OK;
551 // simulate failure at breakpoint 1 and 2
553 void rsm::breakpoint1() {
555 LOG("Dying at breakpoint 1 in rsm!");
560 void rsm::breakpoint2() {
562 LOG("Dying at breakpoint 2 in rsm!");
567 void rsm::partition1(lock & rsm_mutex_lock) {
569 net_repair(false, rsm_mutex_lock);
575 rsm_test_protocol::status rsm::breakpointreq(rsm_test_protocol::status &r, int b) {
576 r = rsm_test_protocol::OK;
578 LOG("breakpoint " << b);
579 if (b == 1) break1 = true;
580 else if (b == 2) break2 = true;
581 else if (b == 3 || b == 4) cfg->breakpoint(b);
582 else r = rsm_test_protocol::ERR;