dist/net_rdma: cleanup now works too

With this longer tests (5sec iperf with kvm) work without problems now.

dist/net_rdma.c

@@ -40,6 +40,9 @@
 
 #include "dist/utils.h"
 
+static const uint64_t kPollReportThreshold = 128;
+static const uint64_t kCleanReportThreshold = 128;
+
 const char *shm_path = NULL;
 size_t shm_size = 256 * 1024 * 1024ULL;  // 256MB
 void *shm_base = NULL;
@@ -195,13 +198,19 @@ static int PeersInitDevs() {
 }
 
 int PeerDevSendIntro(struct Peer *peer) {
+#ifdef DEBUG
   fprintf(stderr, "PeerDevSendIntro(%s)\n", peer->sock_path);
+#endif
 
   struct SimbricksProtoNetDevIntro *di = &peer->dev_intro;
   peer->local_base = (void *) ((uintptr_t) peer->shm_base + di->d2n_offset);
   peer->local_elen = di->d2n_elen;
   peer->local_enum = di->d2n_nentries;
 
+  peer->cleanup_base = (void *) ((uintptr_t) peer->shm_base + di->n2d_offset);
+  peer->cleanup_elen = di->n2d_elen;
+  peer->cleanup_enum = di->n2d_nentries;
+
   struct SimbricksProtoNetNetIntro *ni = &peer->net_intro;
   ssize_t ret = send(peer->sock_fd, ni, sizeof(*ni), 0);
   if (ret < 0) {
@@ -215,7 +224,9 @@ int PeerDevSendIntro(struct Peer *peer) {
 }
 
 int PeerNetSetupQueues(struct Peer *peer) {
+#ifdef DEBUG
   fprintf(stderr, "PeerNetSetupQueues(%s)\n", peer->sock_path);
+#endif
 
   struct SimbricksProtoNetDevIntro *di = &peer->dev_intro;
   if (ShmAlloc(di->d2n_elen * di->d2n_nentries, &di->d2n_offset)) {
@@ -233,6 +244,10 @@ int PeerNetSetupQueues(struct Peer *peer) {
   peer->local_elen = di->n2d_elen;
   peer->local_enum = di->n2d_nentries;
 
+  peer->cleanup_base = (void *) ((uintptr_t) shm_base + di->d2n_offset);
+  peer->cleanup_elen = di->d2n_elen;
+  peer->cleanup_enum = di->d2n_nentries;
+
   if (UxsocketSendFd(peer->sock_fd, di, sizeof(*di), peer->shm_fd)) {
     fprintf(stderr, "PeerNetSetupQueues: sending welcome message failed (%lu)",
             peer - peers);
@@ -241,9 +256,41 @@ int PeerNetSetupQueues(struct Peer *peer) {
   return 0;
 }
 
+int PeerReport(struct Peer *peer, uint32_t written_pos, uint32_t clean_pos) {
+  if (written_pos == peer->cleanup_pos_last &&
+      clean_pos == peer->local_pos_cleaned)
+    return 0;
+
+#ifdef DEBUG
+  fprintf(stderr, "PeerReport: peer %s written %u -> %u, cleaned %u -> %u\n",
+          peer->sock_path, peer->cleanup_pos_last, written_pos,
+          peer->local_pos_cleaned, clean_pos);
+#endif
+
+  peer->cleanup_pos_last = written_pos;
+  while (peer->local_pos_cleaned != clean_pos) {
+    void *entry =
+        (peer->local_base + peer->local_pos_cleaned * peer->local_elen);
+    if (peer->is_dev) {
+      struct SimbricksProtoNetD2NDummy *d2n = entry;
+      d2n->own_type = SIMBRICKS_PROTO_NET_D2N_OWN_DEV;
+    } else {
+      struct SimbricksProtoNetN2DDummy *n2d = entry;
+      n2d->own_type = SIMBRICKS_PROTO_NET_N2D_OWN_NET;
+    }
+
+    peer->local_pos_cleaned += 1;
+    if (peer->local_pos_cleaned >= peer->local_enum)
+      peer->local_pos_cleaned -= peer->local_enum;
+  }
+
+  return 0;
+}
+
 static int PeerEvent(struct Peer *peer, uint32_t events) {
+#ifdef DEBUG
   fprintf(stderr, "PeerEvent(%s)\n", peer->sock_path);
-  fflush(stdout);
+#endif
 
   // disable peer if not an input event
   if (!(events & EPOLLIN)) {
@@ -293,12 +340,8 @@ static int PeerEvent(struct Peer *peer, uint32_t events) {
   return 0;
 }
 
-static void *PollThread(void *data) {
-  while (true) {
-    for (size_t i = 0; i < peer_num; i++) {
-      struct Peer *peer = &peers[i];
-      if (!peer->ready)
-        continue;
+static inline void PollPeerTransfer(struct Peer *peer, bool *report) {
+  // XXX: consider batching this to forward multiple entries at once if possible
 
   void *entry = (peer->local_base + peer->local_pos * peer->local_elen);
   bool ready;
@@ -317,8 +360,64 @@ static void *PollThread(void *data) {
     peer->local_pos += 1;
     if (peer->local_pos >= peer->local_enum)
       peer->local_pos -= peer->local_enum;
+
+    uint64_t unreported = (peer->local_pos - peer->local_pos_reported) %
+                          peer->local_enum;
+    if (unreported >= kPollReportThreshold)
+      *report = true;
   }
+}
+
+static inline void PollPeerCleanup(struct Peer *peer, bool *report) {
+  // XXX: could also be batched
+
+  if (peer->cleanup_pos_next == peer->cleanup_pos_last)
+    return;
+
+  void *entry =
+      (peer->cleanup_base + peer->cleanup_pos_next * peer->cleanup_elen);
+        bool ready;
+  if (peer->is_dev) {
+    struct SimbricksProtoNetN2DDummy *n2d = entry;
+    ready = (n2d->own_type & SIMBRICKS_PROTO_NET_N2D_OWN_MASK) ==
+        SIMBRICKS_PROTO_NET_N2D_OWN_NET;
+  } else {
+    struct SimbricksProtoNetD2NDummy *d2n = entry;
+    ready = (d2n->own_type & SIMBRICKS_PROTO_NET_D2N_OWN_MASK) ==
+        SIMBRICKS_PROTO_NET_D2N_OWN_DEV;
   }
+
+  if (ready) {
+#ifdef DEBUG
+    fprintf(stderr, "PollPeerCleanup: peer %s has clean entry at %u\n",
+            peer->sock_path, peer->cleanup_pos_next);
+#endif
+    peer->cleanup_pos_next += 1;
+    if (peer->cleanup_pos_next >= peer->cleanup_enum)
+      peer->cleanup_pos_next -= peer->cleanup_enum;
+
+    uint64_t unreported = (peer->cleanup_pos_next - peer->cleanup_pos_reported)
+                          % peer->cleanup_enum;
+    if (unreported >= kCleanReportThreshold)
+      *report = true;
+  }
+}
+
+static void *PollThread(void *data) {
+  while (true) {
+    // poll queue for transferring entries
+    bool report = false;
+    for (size_t i = 0; i < peer_num; i++) {
+      struct Peer *peer = &peers[i];
+      if (!peer->ready)
+        continue;
+
+      PollPeerTransfer(peer, &report);
+      PollPeerCleanup(peer, &report);
+    }
+
+    if (report)
+      RdmaPassReport();
   }
   return NULL;
 }
@@ -349,7 +448,10 @@ int main(int argc, char *argv[]) {
   if (ParseArgs(argc, argv))
     return EXIT_FAILURE;
 
+#ifdef DEBUG
   fprintf(stderr, "pid=%d shm=%s\n", getpid(), shm_path);
+#endif
+
   if ((shm_fd = ShmCreate(shm_path, shm_size, &shm_base)) < 0)
     return EXIT_FAILURE;
 

dist/net_rdma.h

@@ -32,21 +32,44 @@
 #include <simbricks/proto/network.h>
 
 struct Peer {
-  struct SimbricksProtoNetDevIntro dev_intro;
-  struct SimbricksProtoNetNetIntro net_intro;
-  const char *sock_path;
-
   /* base address of the local queue we're polling.
      (d2n or n2d depending on is_dev). */
   uint8_t *local_base;
-  uint64_t local_elen;
-  uint64_t local_enum;
-  uint64_t local_pos;
+  uint32_t local_elen;
+  uint32_t local_enum;
+  uint32_t local_pos;
+  // last position reported to our peer
+  uint32_t local_pos_reported;
+  // last position cleaned
+  uint32_t local_pos_cleaned;
 
   // rkey and base address of the remote queue to write to
   uint64_t remote_rkey;
   uint64_t remote_base;
 
+  /* For cleanup we poll the queue just to see when entries get freed. We need
+     to know up to where we can poll, i.e. what entries the peer has written to.
+     The peer communicates this position periodically and we store it in
+     `cleanup_pos_last`. `cleanup_pos_next` refers to the next entry we will
+     poll. Finally we need to report the freed positions to our peer again, so
+     the peer can mark these entries as unused in it's local queue, we again do
+     this periodically and keep track of the last communicated position in
+     `cleanup_pos_reported`. */
+  uint8_t *cleanup_base;
+  uint32_t cleanup_elen;
+  uint32_t cleanup_enum;
+  // next position to be cleaned up
+  uint32_t cleanup_pos_next;
+  // first entry not ready for cleanup yet
+  volatile uint32_t cleanup_pos_last;
+  // last cleanup position reported to peer
+  uint32_t cleanup_pos_reported;
+
+
+  struct SimbricksProtoNetDevIntro dev_intro;
+  struct SimbricksProtoNetNetIntro net_intro;
+  const char *sock_path;
+
   /* RDMA memory region for the shared memory of the queues on this end. Could
      be either our own global SHM region if this is a network peer, or the SHM
      region allocated by the device peer. */
@@ -78,11 +101,13 @@ extern int epfd;
 
 int PeerDevSendIntro(struct Peer *peer);
 int PeerNetSetupQueues(struct Peer *peer);
+int PeerReport(struct Peer *peer, uint32_t written_pos, uint32_t clean_pos);
 
 int RdmaListen(struct sockaddr_in *addr);
 int RdmaConnect(struct sockaddr_in *addr);
 int RdmaPassIntro(struct Peer *peer);
 int RdmaPassEntry(struct Peer *peer);
+int RdmaPassReport();
 int RdmaEvent();
 
 #endif  // DIST_NET_RDMA_NET_RDMA_H_
\ No newline at end of file

dist/rdma.c

@@ -33,11 +33,19 @@
 
 #define MSG_RXBUFS 16
 #define MSG_TXBUFS 16
+#define MAX_PEERS 32
+
+struct NetRdmaReportMsg {
+  uint32_t written_pos[MAX_PEERS];
+  uint32_t clean_pos[MAX_PEERS];
+  bool valid[MAX_PEERS];
+} __attribute__ ((packed));
 
 struct NetRdmaMsg {
   union {
     struct SimbricksProtoNetDevIntro dev;
     struct SimbricksProtoNetNetIntro net;
+    struct NetRdmaReportMsg report;
     struct NetRdmaMsg *next_free;
   };
   uint64_t id;
@@ -47,6 +55,7 @@ struct NetRdmaMsg {
   enum {
     kMsgDev,
     kMsgNet,
+    kMsgReport,
   } msg_type;
 } __attribute__ ((packed));
 
@@ -95,7 +104,7 @@ static int RdmMsgRxEnqueue(struct NetRdmaMsg *msg) {
   return 0;
 }
 
-static int RdmaMsgRx(struct NetRdmaMsg *msg) {
+static int RdmaMsgRxIntro(struct NetRdmaMsg *msg) {
   if (msg->id >= peer_num) {
     fprintf(stderr, "RdmMsgRx: invalid peer id in message (%lu)\n", msg->id);
     abort();
@@ -136,6 +145,30 @@ static int RdmaMsgRx(struct NetRdmaMsg *msg) {
   return 0;
 }
 
+static int RdmaMsgRxReport(struct NetRdmaMsg *msg) {
+  for (size_t i = 0; i < MAX_PEERS && i < peer_num; i++) {
+    if (!msg->report.valid[i])
+      continue;
+
+    if (i >= peer_num) {
+      fprintf(stderr, "RdmaMsgRxReport: invalid ready peer number %zu\n", i);
+      abort();
+    }
+    PeerReport(&peers[i], msg->report.written_pos[i], msg->report.clean_pos[i]);
+  }
+  return 0;
+}
+
+static int RdmaMsgRx(struct NetRdmaMsg *msg) {
+  if (msg->msg_type == kMsgDev || msg->msg_type == kMsgNet)
+    return RdmaMsgRxIntro(msg);
+  else if (msg->msg_type == kMsgReport)
+    return RdmaMsgRxReport(msg);
+
+  fprintf(stderr, "RdmaMsgRx: unexpected message type = %u\n", msg->msg_type);
+  abort();
+}
+
 static int RdmaCommonInit() {
   if (!(pd = ibv_alloc_pd(cm_id->verbs))) {
     perror("RdmaCommonInit: ibv_alloc_pd failed");
@@ -181,7 +214,9 @@ static int RdmaCommonInit() {
     perror("RdmMsgRxEnqueue: ibv_req_notify_cq failed");
     return 1;
   }
+#ifdef RDMA_DEBUG
   fprintf(stderr, "Enqueue rx buffers\n");
+#endif
   // post receive operations for all rx buffers
   for (int i = 0; i < MSG_RXBUFS; i++)
     if (RdmMsgRxEnqueue(&msgs[i]))
@@ -234,7 +269,9 @@ int RdmaListen(struct sockaddr_in *addr) {
     return 1;
   }
 
+#ifdef RDMA_DEBUG
   fprintf(stderr, "RdmaListen: listen done\n");
+#endif
   struct rdma_cm_event *event;
   if (rdma_get_cm_event(cm_channel, &event)) {
     perror("RdmaListen: rdma_get_cm_event failed");
@@ -247,7 +284,9 @@ int RdmaListen(struct sockaddr_in *addr) {
   cm_id = event->id;
   rdma_ack_cm_event(event);
 
+#ifdef RDMA_DEBUG
   fprintf(stderr, "RdmaListen: got conn request\n");
+#endif
 
   if (RdmaCommonInit())
     return 1;
@@ -258,7 +297,10 @@ int RdmaListen(struct sockaddr_in *addr) {
     return 1;
   }
 
+#ifdef RDMA_DEBUG
   fprintf(stderr, "RdmaListen: accept done\n");
+#endif
+
   if (rdma_get_cm_event(cm_channel, &event)) {
     perror("RdmaListen: rdma_get_cm_event failed");
     return 1;
@@ -268,7 +310,10 @@ int RdmaListen(struct sockaddr_in *addr) {
     return 1;
   }
   rdma_ack_cm_event(event);
+
+#ifdef RDMA_DEBUG
   fprintf(stderr, "RdmaListen: conn established\n");
+#endif
 
   if (RdmaCommonSetNonblock())
     return 1;
@@ -302,7 +347,10 @@ int RdmaConnect(struct sockaddr_in *addr) {
     return 1;
   }
   rdma_ack_cm_event(event);
+
+#ifdef RDMA_DEBUG
   fprintf(stderr, "RdmaConnect: address resolved\n");
+#endif
 
   if (rdma_resolve_route(cm_id, 5000)) {
     perror("RdmaConnect: rdma_resolve_route failed");
@@ -318,7 +366,10 @@ int RdmaConnect(struct sockaddr_in *addr) {
     return 1;
   }
   rdma_ack_cm_event(event);
+
+#ifdef RDMA_DEBUG
   fprintf(stderr, "RdmaConnect: route resolved\n");
+#endif
 
   if (RdmaCommonInit())
     return 1;
@@ -330,7 +381,10 @@ int RdmaConnect(struct sockaddr_in *addr) {
     return 1;
   }
 
+#ifdef RDMA_DEBUG
   fprintf(stderr, "RdmaConnect: connect issued\n");
+#endif
+
   if (rdma_get_cm_event(cm_channel, &event)) {
     perror("RdmaConnect: rdma_get_cm_event failed (connect)");
     return 1;
@@ -348,7 +402,9 @@ int RdmaConnect(struct sockaddr_in *addr) {
 }
 
 int RdmaEvent() {
+#ifdef RDMA_DEBUG
   fprintf(stderr, "RdmaEvent [pid=%d]\n", getpid());
+#endif
 
   struct ibv_cq *ecq;
   void *ectx;
@@ -371,10 +427,15 @@ int RdmaEvent() {
       perror("RdmaEvent: ibv_poll_cq failed");
       return 1;
     }
+
+#ifdef RDMA_DEBUG
     fprintf(stderr, "  n=%d\n", n);
+#endif
     for (int i = 0; i < n; i++) {
       if (wcs[i].opcode == IBV_WC_SEND) {
+#ifdef RDMA_DEBUG
         fprintf(stderr, "Send done\n", n);
+#endif
         if (wcs[i].status != IBV_WC_SUCCESS) {
           fprintf(stderr, "RdmaEvent: unsuccessful send (%u)\n", wcs[i].status);
           abort();
@@ -383,7 +444,9 @@ int RdmaEvent() {
         // need to free the send buffer again
         RdmaMsgFree(msgs + wcs[i].wr_id);
       } else if ((wcs[i].opcode & IBV_WC_RECV)) {
+#ifdef RDMA_DEBUG
         fprintf(stderr, "Recv done\n", n);
+#endif
 
         if (wcs[i].status != IBV_WC_SUCCESS) {
           fprintf(stderr, "RdmaEvent: unsuccessful recv (%u)\n", wcs[i].status);
@@ -404,7 +467,9 @@ int RdmaEvent() {
 }
 
 int RdmaPassIntro(struct Peer *peer) {
+#ifdef RDMA_DEBUG
   fprintf(stderr, "RdmaPassIntro(%s)\n", peer->sock_path);
+#endif
 
   // device peers have sent us an SHM region, need to register this an as MR
   if (peer->is_dev) {
@@ -466,15 +531,20 @@ int RdmaPassIntro(struct Peer *peer) {
     perror("RdmaPassIntro: ibv_post_send failed");
     return 1;
   }
-  fprintf(stderr, "RdmaPassIntro: ibv_post_send done\n");
 
+#ifdef RDMA_DEBUG
+  fprintf(stderr, "RdmaPassIntro: ibv_post_send done\n");
+#endif
   return 0;
 }
 
 int RdmaPassEntry(struct Peer *peer) {
-  fprintf(stderr, "RdmaPassEntry(%s,%lu)\n", peer->sock_path, peer->local_pos);
-  fprintf(stderr, "  remote_base=%lx local_base=%lx\n", peer->remote_base,
+#ifdef RDMA_DEBUG
+  fprintf(stderr, "RdmaPassEntry(%s,%u)\n", peer->sock_path, peer->local_pos);
+  fprintf(stderr, "  remote_base=%lx local_base=%p\n", peer->remote_base,
           peer->local_base);
+#endif
+
   uint64_t pos = peer->local_pos * peer->local_elen;
   struct ibv_sge sge;
   sge.addr = (uintptr_t) (peer->local_base + pos);
@@ -496,3 +566,56 @@ int RdmaPassEntry(struct Peer *peer) {
   }
   return 0;
 }
+
+int RdmaPassReport() {
+  if (peer_num > MAX_PEERS) {
+    fprintf(stderr, "RdmaPassReport: peer_num (%zu) larger than max (%u)\n",
+            peer_num, MAX_PEERS);
+    abort();
+  }
+
+  struct NetRdmaMsg *msg = RdmaMsgAlloc();
+  if (!msg)
+    return 1;
+
+  msg->msg_type = kMsgReport;
+  for (size_t i = 0; i < MAX_PEERS; i++) {
+    if (i >= peer_num) {
+      msg->report.valid[i] = false;
+      continue;
+    }
+
+    struct Peer *peer = &peers[i];
+    msg->report.valid[i] = peer->ready;
+    if (!peer->ready)
+      continue;
+
+    peer->cleanup_pos_reported = peer->cleanup_pos_next;
+    msg->report.clean_pos[i] = peer->cleanup_pos_reported;
+    peer->local_pos_reported = peer->local_pos;
+    msg->report.written_pos[i] = peer->local_pos_reported;
+  }
+
+  struct ibv_sge sge;
+  sge.addr = (uintptr_t) msg;
+  sge.length = sizeof(*msg);
+  sge.lkey = mr_msgs->lkey;
+
+  struct ibv_send_wr send_wr = { };
+  send_wr.wr_id = msg - msgs;
+  send_wr.opcode = IBV_WR_SEND;
+  send_wr.send_flags = IBV_SEND_SIGNALED;
+  send_wr.sg_list = &sge;
+  send_wr.num_sge = 1;
+
+  struct ibv_send_wr *bad_send_wr;
+  if (ibv_post_send(cm_id->qp, &send_wr, &bad_send_wr)) {
+    perror("RdmaPassReport: ibv_post_send failed");
+    return 1;
+  }
+
+#ifdef RDMA_DEBUG
+  fprintf(stderr, "RdmaPassReport: ibv_post_send done\n");
+#endif
+  return 0;
+}
\ No newline at end of file