hydra-eval-jobs: Parallelize

This commit is contained in:
Eelco Dolstra 2020-02-19 21:10:22 +01:00
parent c53cfb5234
commit be8eb9d00d
No known key found for this signature in database
GPG key ID: 8170B4726D7198DE

View file

@ -1,37 +1,71 @@
#include <map>
#include <iostream>
#define GC_LINUX_THREADS 1
#include <gc/gc_allocator.h>
#include "shared.hh"
#include "store-api.hh"
#include "eval.hh"
#include "eval-inline.hh"
#include "util.hh"
#include "json.hh"
#include "get-drvs.hh"
#include "globals.hh"
#include "common-eval-args.hh"
#include "flake/flakeref.hh"
#include "flake/flake.hh"
#include "attr-path.hh"
#include "derivations.hh"
#include "hydra-config.hh"
#include <sys/types.h>
#include <sys/wait.h>
#include <sys/resource.h>
#include <nlohmann/json.hpp>
using namespace nix;
static Path gcRootsDir;
static size_t maxMemorySize;
struct MyArgs : MixEvalArgs, MixCommonArgs
{
Path releaseExpr;
bool flake = false;
bool dryRun = false;
static void findJobs(EvalState & state, JSONObject & top,
Bindings & autoArgs, Value & v, const string & attrPath);
MyArgs() : MixCommonArgs("hydra-eval-jobs")
{
mkFlag()
.longName("help")
.description("show usage information")
.handler([&]() {
printHelp(programName, std::cout);
throw Exit();
});
mkFlag()
.longName("gc-roots-dir")
.description("garbage collector roots directory")
.labels({"path"})
.dest(&gcRootsDir);
static string queryMetaStrings(EvalState & state, DrvInfo & drv, const string & name, const string & subAttribute)
mkFlag()
.longName("dry-run")
.description("don't create store derivations")
.set(&dryRun, true);
mkFlag()
.longName("flake")
.description("build a flake")
.set(&flake, true);
expectArg("expr", &releaseExpr);
}
};
static MyArgs myArgs;
static std::string queryMetaStrings(EvalState & state, DrvInfo & drv, const string & name, const string & subAttribute)
{
Strings res;
std::function<void(Value & v)> rec;
@ -56,116 +90,163 @@ static string queryMetaStrings(EvalState & state, DrvInfo & drv, const string &
return concatStringsSep(", ", res);
}
static void findJobsWrapped(EvalState & state, JSONObject & top,
Bindings & autoArgs, Value & vIn, const string & attrPath)
static void worker(
EvalState & state,
Bindings & autoArgs,
AutoCloseFD & to,
AutoCloseFD & from)
{
debug(format("at path `%1%'") % attrPath);
Value vTop;
checkInterrupt();
if (myArgs.flake) {
using namespace flake;
Value v;
state.autoCallFunction(autoArgs, vIn, v);
auto flakeRef = parseFlakeRef(myArgs.releaseExpr);
if (v.type == tAttrs) {
auto vFlake = state.allocValue();
auto drv = getDerivation(state, v, false);
auto lockedFlake = lockFlake(state, flakeRef,
LockFlags {
.updateLockFile = false,
.useRegistries = false,
.allowMutable = false,
});
if (drv) {
Path drvPath;
callFlake(state, lockedFlake, *vFlake);
DrvInfo::Outputs outputs = drv->queryOutputs();
auto vOutputs = vFlake->attrs->get(state.symbols.create("outputs"))->value;
state.forceValue(*vOutputs);
if (drv->querySystem() == "unknown")
throw EvalError("derivation must have a system attribute");
auto aHydraJobs = vOutputs->attrs->get(state.symbols.create("hydraJobs"));
if (!aHydraJobs)
aHydraJobs = vOutputs->attrs->get(state.symbols.create("checks"));
if (!aHydraJobs)
throw Error("flake '%s' does not provide any Hydra jobs or checks", flakeRef);
{
auto res = top.object(attrPath);
res.attr("nixName", drv->queryName());
res.attr("system", drv->querySystem());
res.attr("drvPath", drvPath = drv->queryDrvPath());
res.attr("description", drv->queryMetaString("description"));
res.attr("license", queryMetaStrings(state, *drv, "license", "shortName"));
res.attr("homepage", drv->queryMetaString("homepage"));
res.attr("maintainers", queryMetaStrings(state, *drv, "maintainers", "email"));
res.attr("schedulingPriority", drv->queryMetaInt("schedulingPriority", 100));
res.attr("timeout", drv->queryMetaInt("timeout", 36000));
res.attr("maxSilent", drv->queryMetaInt("maxSilent", 7200));
res.attr("isChannel", drv->queryMetaBool("isHydraChannel", false));
vTop = *aHydraJobs->value;
/* If this is an aggregate, then get its constituents. */
Bindings::iterator a = v.attrs->find(state.symbols.create("_hydraAggregate"));
if (a != v.attrs->end() && state.forceBool(*a->value, *a->pos)) {
Bindings::iterator a = v.attrs->find(state.symbols.create("constituents"));
if (a == v.attrs->end())
throw EvalError("derivation must have a constituents attribute");
PathSet context;
state.coerceToString(*a->pos, *a->value, context, true, false);
PathSet drvs;
for (auto & i : context)
if (i.at(0) == '!') {
size_t index = i.find("!", 1);
drvs.insert(string(i, index + 1));
} else {
state.evalFile(lookupFileArg(state, myArgs.releaseExpr), vTop);
}
auto vRoot = state.allocValue();
state.autoCallFunction(autoArgs, vTop, *vRoot);
while (true) {
/* Wait for the master to send us a job name. */
writeLine(to.get(), "next");
auto s = readLine(from.get());
if (s == "exit") break;
if (!hasPrefix(s, "do ")) abort();
std::string attrPath(s, 3);
debug("worker process %d at '%s'", getpid(), attrPath);
/* Evaluate it and send info back to the master. */
nlohmann::json reply;
try {
auto v = findAlongAttrPath(state, attrPath, autoArgs, *vRoot).first;
state.forceValue(*v);
if (auto drv = getDerivation(state, *v, false)) {
DrvInfo::Outputs outputs = drv->queryOutputs();
if (drv->querySystem() == "unknown")
throw EvalError("derivation must have a 'system' attribute");
auto drvPath = drv->queryDrvPath();
nlohmann::json job;
job["nixName"] = drv->queryName();
job["system"] =drv->querySystem();
job["drvPath"] = drvPath;
job["description"] = drv->queryMetaString("description");
job["license"] = queryMetaStrings(state, *drv, "license", "shortName");
job["homepage"] = drv->queryMetaString("homepage");
job["maintainers"] = queryMetaStrings(state, *drv, "maintainers", "email");
job["schedulingPriority"] = drv->queryMetaInt("schedulingPriority", 100);
job["timeout"] = drv->queryMetaInt("timeout", 36000);
job["maxSilent"] = drv->queryMetaInt("maxSilent", 7200);
job["isChannel"] = drv->queryMetaBool("isHydraChannel", false);
/* If this is an aggregate, then get its constituents. */
auto a = v->attrs->get(state.symbols.create("_hydraAggregate"));
if (a && state.forceBool(*a->value, *a->pos)) {
auto a = v->attrs->get(state.symbols.create("constituents"));
if (!a)
throw EvalError("derivation must have a constituents attribute");
PathSet context;
state.coerceToString(*a->pos, *a->value, context, true, false);
for (auto & i : context)
if (i.at(0) == '!') {
size_t index = i.find("!", 1);
job["constituents"].push_back(string(i, index + 1));
}
state.forceList(*a->value, *a->pos);
for (unsigned int n = 0; n < a->value->listSize(); ++n) {
auto v = a->value->listElems()[n];
state.forceValue(*v);
if (v->type == tString)
job["namedConstituents"].push_back(state.forceStringNoCtx(*v));
}
res.attr("constituents", concatStringsSep(" ", drvs));
}
/* Register the derivation as a GC root. !!! This
registers roots for jobs that we may have already
done. */
auto localStore = state.store.dynamic_pointer_cast<LocalFSStore>();
if (gcRootsDir != "" && localStore) {
Path root = gcRootsDir + "/" + std::string(baseNameOf(drvPath));
if (!pathExists(root))
localStore->addPermRoot(localStore->parseStorePath(drvPath), root, false);
}
nlohmann::json out;
for (auto & j : outputs)
out[j.first] = j.second;
job["outputs"] = std::move(out);
reply["job"] = std::move(job);
}
/* Register the derivation as a GC root. !!! This
registers roots for jobs that we may have already
done. */
auto localStore = state.store.dynamic_pointer_cast<LocalFSStore>();
if (gcRootsDir != "" && localStore) {
Path root = gcRootsDir + "/" + std::string(baseNameOf(drvPath));
if (!pathExists(root))
localStore->addPermRoot(localStore->parseStorePath(drvPath), root, false);
}
auto res2 = res.object("outputs");
for (auto & j : outputs)
res2.attr(j.first, j.second);
}
}
else {
if (!state.isDerivation(v)) {
for (auto & i : v.attrs->lexicographicOrder()) {
else if (v->type == tAttrs) {
auto attrs = nlohmann::json::array();
StringSet ss;
for (auto & i : v->attrs->lexicographicOrder()) {
std::string name(i->name);
/* Skip jobs with dots in the name. */
if (name.find('.') != std::string::npos) {
if (name.find('.') != std::string::npos || name.find(' ') != std::string::npos) {
printError("skipping job with illegal name '%s'", name);
continue;
}
findJobs(state, top, autoArgs, *i->value,
(attrPath.empty() ? "" : attrPath + ".") + name);
attrs.push_back(name);
}
reply["attrs"] = std::move(attrs);
}
} catch (EvalError & e) {
reply["error"] = filterANSIEscapes(e.msg(), true);
}
writeLine(to.get(), reply.dump());
/* If our RSS exceeds the maximum, exit. The master will
start a new process. */
struct rusage r;
getrusage(RUSAGE_SELF, &r);
if ((size_t) r.ru_maxrss > maxMemorySize * 1024) break;
}
else if (v.type == tNull) {
// allow null values, meaning 'do nothing'
}
else
throw TypeError(format("unsupported value: %1%") % v);
writeLine(to.get(), "restart");
}
static void findJobs(EvalState & state, JSONObject & top,
Bindings & autoArgs, Value & v, const string & attrPath)
{
try {
findJobsWrapped(state, top, autoArgs, v, attrPath);
} catch (EvalError & e) {
auto res = top.object(attrPath);
res.attr("error", filterANSIEscapes(e.msg(), true));
}
}
int main(int argc, char * * argv)
{
/* Prevent undeclared dependencies in the evaluation via
@ -176,54 +257,14 @@ int main(int argc, char * * argv)
auto config = std::make_unique<::Config>();
auto initialHeapSize = config->getStrOption("evaluator_initial_heap_size", "");
if (initialHeapSize != "")
setenv("GC_INITIAL_HEAP_SIZE", initialHeapSize.c_str(), 1);
auto nrWorkers = config->getIntOption("evaluator_workers", 1);
maxMemorySize = config->getIntOption("evaluator_max_memory_size", 4096);
initNix();
initGC();
struct MyArgs : MixEvalArgs, MixCommonArgs
{
Path releaseExpr;
bool flake = false;
MyArgs() : MixCommonArgs("hydra-eval-jobs")
{
mkFlag()
.longName("help")
.description("show usage information")
.handler([&]() {
printHelp(programName, std::cout);
throw Exit();
});
mkFlag()
.longName("gc-roots-dir")
.description("garbage collector roots directory")
.labels({"path"})
.dest(&gcRootsDir);
mkFlag()
.longName("dry-run")
.description("don't create store derivations")
.set(&settings.readOnlyMode, true);
mkFlag()
.longName("flake")
.description("build a flake")
.set(&flake, true);
expectArg("expr", &releaseExpr);
}
};
MyArgs myArgs;
myArgs.parseCmdline(argvToStrings(argc, argv));
JSONObject json(std::cout, true);
std::cout.flush();
/* FIXME: The build hook in conjunction with import-from-derivation is causing "unexpected EOF" during eval */
settings.builders = "";
@ -235,47 +276,196 @@ int main(int argc, char * * argv)
'getEnv', 'currentSystem' etc. */
evalSettings.pureEval = myArgs.flake;
if (myArgs.dryRun) settings.readOnlyMode = true;
if (myArgs.releaseExpr == "") throw UsageError("no expression specified");
if (gcRootsDir == "") printMsg(lvlError, "warning: `--gc-roots-dir' not specified");
EvalState state(myArgs.searchPath, openStore());
struct State
{
std::set<std::string> todo{""};
std::set<std::string> active;
nlohmann::json jobs;
std::exception_ptr exc;
};
Bindings & autoArgs = *myArgs.getAutoArgs(state);
std::condition_variable wakeup;
Value v;
Sync<State> state_;
if (myArgs.flake) {
using namespace flake;
/* Start a handler thread per worker process. */
auto handler = [&]()
{
try {
pid_t pid = -1;
AutoCloseFD from, to;
auto flakeRef = parseFlakeRef(myArgs.releaseExpr);
while (true) {
auto vFlake = state.allocValue();
/* Start a new worker process if necessary. */
if (pid == -1) {
Pipe toPipe, fromPipe;
toPipe.create();
fromPipe.create();
pid = startProcess(
[&,
to{std::make_shared<AutoCloseFD>(std::move(fromPipe.writeSide))},
from{std::make_shared<AutoCloseFD>(std::move(toPipe.readSide))}
]()
{
try {
EvalState state(myArgs.searchPath, openStore());
Bindings & autoArgs = *myArgs.getAutoArgs(state);
worker(state, autoArgs, *to, *from);
} catch (std::exception & e) {
nlohmann::json err;
err["error"] = e.what();
writeLine(to->get(), err.dump());
}
},
ProcessOptions { .allowVfork = false });
from = std::move(fromPipe.readSide);
to = std::move(toPipe.writeSide);
debug("created worker process %d", pid);
}
auto lockedFlake = lockFlake(state, flakeRef,
LockFlags {
.updateLockFile = false,
.useRegistries = false,
.allowMutable = false,
});
/* Check whether the existing worker process is still there. */
auto s = readLine(from.get());
if (s == "restart") {
pid = -1;
continue;
} else if (s != "next") {
auto json = nlohmann::json::parse(s);
throw Error("worker error: %s", (std::string) json["error"]);
}
callFlake(state, lockedFlake, *vFlake);
/* Wait for a job name to become available. */
std::string attrPath;
auto vOutputs = (*vFlake->attrs->get(state.symbols.create("outputs")))->value;
state.forceValue(*vOutputs);
while (true) {
checkInterrupt();
auto state(state_.lock());
if ((state->todo.empty() && state->active.empty()) || state->exc) {
writeLine(to.get(), "exit");
return;
}
if (!state->todo.empty()) {
attrPath = *state->todo.begin();
state->todo.erase(state->todo.begin());
state->active.insert(attrPath);
break;
} else
state.wait(wakeup);
}
auto aHydraJobs = vOutputs->attrs->get(state.symbols.create("hydraJobs"));
if (!aHydraJobs)
aHydraJobs = vOutputs->attrs->get(state.symbols.create("checks"));
if (!aHydraJobs)
throw Error("flake '%s' does not provide any Hydra jobs or checks", flakeRef);
/* Tell the worker to evaluate it. */
writeLine(to.get(), "do " + attrPath);
v = *(*aHydraJobs)->value;
/* Wait for the response. */
auto response = nlohmann::json::parse(readLine(from.get()));
} else {
state.evalFile(lookupFileArg(state, myArgs.releaseExpr), v);
/* Handle the response. */
StringSet newAttrs;
if (response.find("job") != response.end()) {
auto state(state_.lock());
state->jobs[attrPath] = response["job"];
}
if (response.find("attrs") != response.end()) {
for (auto & i : response["attrs"]) {
auto s = (attrPath.empty() ? "" : attrPath + ".") + (std::string) i;
newAttrs.insert(s);
}
}
if (response.find("error") != response.end()) {
auto state(state_.lock());
state->jobs[attrPath]["error"] = response["error"];
}
/* Add newly discovered job names to the queue. */
{
auto state(state_.lock());
state->active.erase(attrPath);
for (auto & s : newAttrs)
state->todo.insert(s);
wakeup.notify_all();
}
}
} catch (...) {
auto state(state_.lock());
state->exc = std::current_exception();
wakeup.notify_all();
}
};
std::vector<std::thread> threads;
for (size_t i = 0; i < nrWorkers; i++)
threads.emplace_back(std::thread(handler));
for (auto & thread : threads)
thread.join();
auto state(state_.lock());
if (state->exc)
std::rethrow_exception(state->exc);
/* For aggregate jobs that have named consistuents
(i.e. constituents that are a job name rather than a
derivation), look up the referenced job and add it to the
dependencies of the aggregate derivation. */
auto store = openStore();
for (auto i = state->jobs.begin(); i != state->jobs.end(); ++i) {
auto jobName = i.key();
auto & job = i.value();
auto named = job.find("namedConstituents");
if (named == job.end()) continue;
if (myArgs.dryRun) {
for (std::string jobName2 : *named) {
auto job2 = state->jobs.find(jobName2);
if (job2 == state->jobs.end())
throw Error("aggregate job '%s' references non-existent job '%s'", jobName, jobName2);
std::string drvPath2 = (*job2)["drvPath"];
job["constituents"].push_back(drvPath2);
}
} else {
std::string drvPath = job["drvPath"];
auto drv = readDerivation(*store, drvPath);
for (std::string jobName2 : *named) {
auto job2 = state->jobs.find(jobName2);
if (job2 == state->jobs.end())
throw Error("aggregate job '%s' references non-existent job '%s'", jobName, jobName2);
std::string drvPath2 = (*job2)["drvPath"];
auto drv2 = readDerivation(*store, drvPath2);
job["constituents"].push_back(drvPath2);
drv.inputDrvs[store->parseStorePath(drvPath2)] = {drv2.outputs.begin()->first};
}
std::string drvName(store->parseStorePath(drvPath).name());
assert(hasSuffix(drvName, drvExtension));
drvName.resize(drvName.size() - drvExtension.size());
auto h = hashDerivationModulo(*store, drv, true);
auto outPath = store->makeOutputPath("out", h, drvName);
drv.env["out"] = store->printStorePath(outPath);
drv.outputs.insert_or_assign("out", DerivationOutput(outPath.clone(), "", ""));
auto newDrvPath = store->printStorePath(writeDerivation(store, drv, drvName));
debug("rewrote aggregate derivation %s -> %s", drvPath, newDrvPath);
job["drvPath"] = newDrvPath;
job["outputs"]["out"] = store->printStorePath(outPath);
}
job.erase("namedConstituents");
}
findJobs(state, json, autoArgs, v, "");
std::cout << state->jobs.dump(2) << "\n";
});
}