hydra-queue-runner: Limit memory usage

When using a binary cache store, the queue runner receives NARs from
the build machines, compresses them, and uploads them to the
cache. However, keeping multiple large NARs in memory can cause the
queue runner to run out of memory. This can happen for instance when
it's processing multiple ISO images concurrently.

The fix is to use a TokenServer to prevent the builder threads to
store more than a certain total size of NARs concurrently (at the
moment, this is hard-coded at 4 GiB). Builder threads that cause the
limit to be exceeded will block until other threads have finished.

The 4 GiB limit does not include certain other allocations, such as
for xz compression or for FSAccessor::readFile(). But since these are
unlikely to be more than the size of the NARs and hydra.nixos.org has
32 GiB RAM, it should be fine.

src/hydra-queue-runner/build-remote.cc

@@ -283,16 +283,43 @@ void State::buildRemote(ref<Store> destStore,
 
     /* Copy the output paths. */
     if (/* machine->sshName != "localhost" */ true) {
-        printMsg(lvlDebug, format("copying outputs of ‘%1%’ from ‘%2%’") % step->drvPath % machine->sshName);
+        MaintainCount mc(nrStepsCopyingFrom);
+
+        auto now1 = std::chrono::steady_clock::now();
+
         PathSet outputs;
         for (auto & output : step->drv.outputs)
             outputs.insert(output.second.path);
-        MaintainCount mc(nrStepsCopyingFrom);
+
+        /* Query the size of the output paths. */
+        size_t totalNarSize = 0;
+        to << cmdQueryPathInfos << outputs;
+        to.flush();
+        while (true) {
+            if (readString(from) == "") break;
+            readString(from); // deriver
+            readStrings<PathSet>(from); // references
+            readLongLong(from); // download size
+            totalNarSize += readLongLong(from);
+        }
+
+        printMsg(lvlDebug, format("copying outputs of ‘%s’ from ‘%s’ (%d bytes)")
+            % step->drvPath % machine->sshName % totalNarSize);
+
+        /* Block until we have the required amount of memory
+           available. FIXME: only need this for binary cache
+           destination stores. */
+        auto resStart = std::chrono::steady_clock::now();
+        auto memoryReservation(memoryTokens.get(totalNarSize));
+        auto resStop = std::chrono::steady_clock::now();
+
+        auto resMs = std::chrono::duration_cast<std::chrono::milliseconds>(resStop - resStart).count();
+        if (resMs >= 1000)
+            printMsg(lvlError, format("warning: had to wait %d ms for %d memory tokens for %s")
+                % resMs % totalNarSize % step->drvPath);
 
         result.accessor = destStore->getFSAccessor();
 
-        auto now1 = std::chrono::steady_clock::now();
-
         to << cmdExportPaths << 0 << outputs;
         to.flush();
         destStore->importPaths(false, from, result.accessor);

src/hydra-queue-runner/hydra-queue-runner.cc

@@ -17,6 +17,7 @@ using namespace nix;
 
 
 State::State()
+    : memoryTokens(4ULL << 30) // FIXME: make this configurable
 {
     hydraData = getEnv("HYDRA_DATA");
     if (hydraData == "") throw Error("$HYDRA_DATA must be set");
@@ -567,6 +568,8 @@ void State::dumpStatus(Connection & conn, bool log)
         root.attr("dispatchTimeAvgMs", nrDispatcherWakeups == 0 ? 0.0 : (float) dispatchTimeMs / nrDispatcherWakeups);
         root.attr("nrDbConnections", dbPool.count());
         root.attr("nrActiveDbUpdates", nrActiveDbUpdates);
+        root.attr("memoryTokensInUse", memoryTokens.currentUse());
+
         {
             root.attr("machines");
             JSONObject nested(out);
@@ -589,6 +592,7 @@ void State::dumpStatus(Connection & conn, bool log)
                 }
             }
         }
+
         {
             root.attr("jobsets");
             JSONObject nested(out);
@@ -600,6 +604,7 @@ void State::dumpStatus(Connection & conn, bool log)
                 nested2.attr("seconds", jobset.second->getSeconds());
             }
         }
+
         {
             root.attr("machineTypes");
             JSONObject nested(out);

src/hydra-queue-runner/state.hh

@@ -9,14 +9,13 @@
 
 #include "db.hh"
 #include "counter.hh"
+#include "token-server.hh"
+
+#include "derivations.hh"
 #include "pathlocks.hh"
 #include "pool.hh"
-#include "sync.hh"
-
 #include "store-api.hh"
-#include "derivations.hh"
+#include "sync.hh"
-
-#include "binary-cache-store.hh" // FIXME
 
 
 typedef unsigned int BuildID;
@@ -354,6 +353,13 @@ private:
     std::shared_ptr<nix::Store> _localStore;
     std::shared_ptr<nix::Store> _destStore;
 
+    /* Token server to prevent threads from allocating too many big
+       strings concurrently while importing NARs from the build
+       machines. When a thread imports a NAR of size N, it will first
+       acquire N memory tokens, causing it to block until that many
+       tokens are available. */
+    nix::TokenServer memoryTokens;
+
 public:
     State();
 

src/hydra-queue-runner/token-server.hh

@@ -3,21 +3,27 @@
 #include <atomic>
 
 #include "sync.hh"
+#include "types.hh"
+
+namespace nix {
+
+MakeError(NoTokens, Error)
 
 /* This class hands out tokens. There are only ‘maxTokens’ tokens
-   available. Calling get() will return a Token object, representing
+   available. Calling get(N) will return a Token object, representing
-   ownership of a token. If no token is available, get() will sleep
+   ownership of N tokens. If the requested number of tokens is
-   until another thread returns a token. */
+   unavailable, get() will sleep until another thread returns a
+   token. */
 
 class TokenServer
 {
-    unsigned int maxTokens;
+    const size_t maxTokens;
 
-    Sync<unsigned int> curTokens{0};
+    Sync<size_t> inUse{0};
     std::condition_variable wakeup;
 
 public:
-    TokenServer(unsigned int maxTokens) : maxTokens(maxTokens) { }
+    TokenServer(size_t maxTokens) : maxTokens(maxTokens) { }
 
     class Token
     {
@@ -25,19 +31,24 @@ public:
 
         TokenServer * ts;
 
+        size_t tokens;
+
         bool acquired = false;
 
-        Token(TokenServer * ts, unsigned int timeout) : ts(ts)
+        Token(TokenServer * ts, size_t tokens, unsigned int timeout)
+            : ts(ts), tokens(tokens)
         {
-            auto curTokens(ts->curTokens.lock());
+            if (tokens >= ts->maxTokens)
-            while (*curTokens >= ts->maxTokens)
+                throw NoTokens(format("requesting more tokens (%d) than exist (%d)") % tokens);
+            auto inUse(ts->inUse.lock());
+            while (*inUse + tokens > ts->maxTokens)
                 if (timeout) {
-                    if (!curTokens.wait_for(ts->wakeup, std::chrono::seconds(timeout),
+                    if (!inUse.wait_for(ts->wakeup, std::chrono::seconds(timeout),
-                            [&]() { return *curTokens < ts->maxTokens; }))
+                            [&]() { return *inUse + tokens <= ts->maxTokens; }))
                         return;
                 } else
-                    curTokens.wait(ts->wakeup);
+                    inUse.wait(ts->wakeup);
-            (*curTokens)++;
+            *inUse += tokens;
             acquired = true;
         }
 
@@ -50,9 +61,9 @@ public:
         {
             if (!ts || !acquired) return;
             {
-                auto curTokens(ts->curTokens.lock());
+                auto inUse(ts->inUse.lock());
-                assert(*curTokens);
+                assert(*inUse >= tokens);
-                (*curTokens)--;
+                *inUse -= tokens;
             }
             ts->wakeup.notify_one();
         }
@@ -60,8 +71,16 @@ public:
         bool operator ()() { return acquired; }
     };
 
-    Token get(unsigned int timeout = 0)
+    Token get(size_t tokens = 1, unsigned int timeout = 0)
     {
-        return Token(this, timeout);
+        return Token(this, tokens, timeout);
+    }
+
+    size_t currentUse()
+    {
+        auto inUse_(inUse.lock());
+        return *inUse_;
     }
 };
+
+}