ReplicatedPG: block reads on an object until the write is committedwip-5992-3

Signed-off-by: Samuel Just <sam.just@inktank.com>

2 files changed, 197 insertions, 23 deletions

diff --git a/src/osd/ReplicatedPG.cc b/src/osd/ReplicatedPG.cc
index 1684fa0dc53..f097b1c34c7 100644
--- a/src/osd/ReplicatedPG.cc
+++ b/src/osd/ReplicatedPG.cc
@@ -988,21 +988,9 @@ void ReplicatedPG::do_op(OpRequestRef op)
     return;
   }
 
-  if ((op->may_read()) && (obc->obs.oi.lost)) {
-    // This object is lost. Reading from it returns an error.
-    dout(20) << __func__ << ": object " << obc->obs.oi.soid
-	     << " is lost" << dendl;
-    osd->reply_op_error(op, -ENFILE);
-    return;
-  }
   dout(25) << __func__ << ": object " << obc->obs.oi.soid
 	   << " has oi of " << obc->obs.oi << dendl;
   
-  if (!op->may_write() && !obc->obs.exists) {
-    osd->reply_op_error(op, -ENOENT);
-    return;
-  }
-
   // are writes blocked by another object?
   if (obc->blocked_by) {
     dout(10) << "do_op writes for " << obc->obs.oi.soid << " blocked by "
@@ -1121,11 +1109,30 @@ void ReplicatedPG::do_op(OpRequestRef op)
     }
   }
 
-  op->mark_started();
-
   OpContext *ctx = new OpContext(op, m->get_reqid(), m->ops,
 				 &obc->obs, obc->ssc, 
 				 this);
+  if (!get_rw_locks(ctx)) {
+    op->mark_delayed("waiting for rw locks");
+    close_op_ctx(ctx);
+    return;
+  }
+
+  if ((op->may_read()) && (obc->obs.oi.lost)) {
+    // This object is lost. Reading from it returns an error.
+    dout(20) << __func__ << ": object " << obc->obs.oi.soid
+	     << " is lost" << dendl;
+    close_op_ctx(ctx);
+    osd->reply_op_error(op, -ENFILE);
+    return;
+  }
+  if (!op->may_write() && !obc->obs.exists) {
+    close_op_ctx(ctx);
+    osd->reply_op_error(op, -ENOENT);
+    return;
+  }
+
+  op->mark_started();
   ctx->obc = obc;
   ctx->src_obc = src_obc;
 
@@ -1202,7 +1209,7 @@ void ReplicatedPG::execute_ctx(OpContext *ctx)
       if (already_complete(oldv)) {
 	reply_ctx(ctx, 0, oldv, entry->user_version);
       } else {
-	delete ctx;
+	close_op_ctx(ctx);
 
 	if (m->wants_ack()) {
 	  if (already_ack(oldv)) {
@@ -1295,7 +1302,7 @@ void ReplicatedPG::execute_ctx(OpContext *ctx)
 
   if (result == -EAGAIN) {
     // clean up after the ctx
-    delete ctx;
+    close_op_ctx(ctx);
     return;
   }
 
@@ -1347,7 +1354,7 @@ void ReplicatedPG::execute_ctx(OpContext *ctx)
     
     reply->add_flags(CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK);
     osd->send_message_osd_client(reply, m->get_connection());
-    delete ctx;
+    close_op_ctx(ctx);
     return;
   }
 
@@ -1395,13 +1402,13 @@ void ReplicatedPG::execute_ctx(OpContext *ctx)
 void ReplicatedPG::reply_ctx(OpContext *ctx, int r)
 {
   osd->reply_op_error(ctx->op, r);
-  delete ctx;
+  close_op_ctx(ctx);
 }
 
 void ReplicatedPG::reply_ctx(OpContext *ctx, int r, eversion_t v, version_t uv)
 {
   osd->reply_op_error(ctx->op, r, v, uv);
-  delete ctx;
+  close_op_ctx(ctx);
 }
 
 void ReplicatedPG::log_op_stats(OpContext *ctx)
@@ -4589,7 +4596,7 @@ void ReplicatedPG::cancel_copy(CopyOpRef cop)
 
   kick_object_context_blocked(ctx->obc);
 
-  delete ctx;
+  close_op_ctx(ctx);
 }
 
 void ReplicatedPG::cancel_copy_ops()
@@ -4754,6 +4761,8 @@ void ReplicatedPG::op_commit(RepGather *repop)
     eval_repop(repop);
   }
 
+  release_op_ctx_locks(repop->ctx);
+
   repop->put();
   unlock();
 }
@@ -4996,6 +5005,7 @@ ReplicatedPG::RepGather *ReplicatedPG::new_repop(OpContext *ctx, ObjectContextRe
  
 void ReplicatedPG::remove_repop(RepGather *repop)
 {
+  release_op_ctx_locks(repop->ctx);
   repop_map.erase(repop->rep_tid);
   repop->put();
 
@@ -8238,7 +8248,12 @@ void ReplicatedPG::scan_range(
 {
   assert(is_locked());
   dout(10) << "scan_range from " << bi->begin << dendl;
-  bi->version = last_update_applied;
+  if (last_update_applied >= info.log_tail) {
+    bi->version = last_update_applied;
+  } else {
+    osr.flush();
+    bi->version = info.last_update;
+  }
   bi->objects.clear();  // for good measure
 
   vector<hobject_t> ls;
diff --git a/src/osd/ReplicatedPG.h b/src/osd/ReplicatedPG.h
index b398dd2fea4..c229ca68f8c 100644
--- a/src/osd/ReplicatedPG.h
+++ b/src/osd/ReplicatedPG.h
@@ -303,6 +303,8 @@ public:
 
     hobject_t new_temp_oid, discard_temp_oid;  ///< temp objects we should start/stop tracking
 
+    enum { W_LOCK, R_LOCK, NONE } lock_to_release;
+
     OpContext(const OpContext& other);
     const OpContext& operator=(const OpContext& other);
 
@@ -316,7 +318,8 @@ public:
       current_osd_subop_num(0),
       data_off(0), reply(NULL), pg(_pg),
       num_read(0),
-      num_write(0) {
+      num_write(0),
+      lock_to_release(NONE) {
       if (_ssc) {
 	new_snapset = _ssc->snapset;
 	snapset = &_ssc->snapset;
@@ -324,6 +327,7 @@ public:
     }
     ~OpContext() {
       assert(!clone_obc);
+      assert(lock_to_release == NONE);
       if (reply)
 	reply->put();
     }
@@ -382,7 +386,7 @@ public:
       if (--nref == 0) {
 	assert(!obc);
 	assert(src_obc.empty());
-	delete ctx;
+	delete ctx; // must already be unlocked
 	delete this;
 	//generic_dout(0) << "deleting " << this << dendl;
       }
@@ -393,6 +397,161 @@ public:
 
 protected:
 
+  /// Tracks pending readers or writers on an object
+  class RWTracker {
+    struct ObjState {
+      enum State {
+	NONE,
+	READ,
+	WRITE
+      };
+      State state;                 /// rw state
+      uint64_t count;              /// number of readers or writers
+      list<OpRequestRef> waiters;  /// ops waiting on state change
+
+      ObjState() : state(NONE), count(0) {}
+      bool get_read(OpRequestRef op) {
+	switch (state) {
+	case NONE:
+	  assert(count == 0);
+	  state = READ;
+	  // fall through
+	case READ:
+	  count++;
+	  return true;
+	case WRITE:
+	  waiters.push_back(op);
+	  return false;
+	default:
+	  assert(0 == "unhandled case");
+	  return false;
+	}
+      }
+      bool get_write(OpRequestRef op) {
+	switch (state) {
+	case NONE:
+	  assert(count == 0);
+	  state = WRITE;
+	  // fall through
+	case WRITE:
+	  count++;
+	  return true;
+	case READ:
+	  waiters.push_back(op);
+	  return false;
+	default:
+	  assert(0 == "unhandled case");
+	  return false;
+	}
+      }
+      void dec(list<OpRequestRef> *requeue) {
+	assert(count > 0);
+	assert(requeue);
+	assert(requeue->empty());
+	count--;
+	if (count == 0) {
+	  state = NONE;
+	  requeue->swap(waiters);
+	}
+      }
+      void put_read(list<OpRequestRef> *requeue) {
+	assert(state == READ);
+	dec(requeue);
+      }
+      void put_write(list<OpRequestRef> *requeue) {
+	assert(state == WRITE);
+	dec(requeue);
+      }
+      void clear(list<OpRequestRef> *requeue) {
+	state = NONE;
+	count = 0;
+	assert(requeue);
+	assert(requeue->empty());
+	requeue->swap(waiters);
+      }
+      bool empty() const { return state == NONE; }
+    };
+    map<hobject_t, ObjState > obj_state;
+  public:
+    bool get_read(const hobject_t &hoid, OpRequestRef op) {
+      return obj_state[hoid].get_read(op);
+    }
+    bool get_write(const hobject_t &hoid, OpRequestRef op) {
+      return obj_state[hoid].get_write(op);
+    }
+    void put_read(const hobject_t &hoid, list<OpRequestRef> *to_wake) {
+      obj_state[hoid].put_read(to_wake);
+      if (obj_state[hoid].empty()) {
+	obj_state.erase(hoid);
+      }
+    }
+    void put_write(const hobject_t &hoid, list<OpRequestRef> *to_wake) {
+      obj_state[hoid].put_write(to_wake);
+      if (obj_state[hoid].empty()) {
+	obj_state.erase(hoid);
+      }
+    }
+  } rw_manager;
+
+  /**
+   * Grabs locks for OpContext, should be cleaned up in close_op_ctx
+   *
+   * @param ctx [in,out] ctx to get locks for
+   * @return true on success, false if we are queued
+   */
+  bool get_rw_locks(OpContext *ctx) {
+    if (ctx->op->may_write()) {
+      if (rw_manager.get_write(ctx->obs->oi.soid, ctx->op)) {
+	ctx->lock_to_release = OpContext::W_LOCK;
+	return true;
+      } else {
+	assert(0 == "Currently there cannot be a read in flight here");
+	return false;
+      }
+    } else {
+      assert(ctx->op->may_read());
+      if (rw_manager.get_read(ctx->obs->oi.soid, ctx->op)) {
+	ctx->lock_to_release = OpContext::R_LOCK;
+	return true;
+      } else {
+	return false;
+      }
+    }
+  }
+
+  /**
+   * Cleans up OpContext
+   *
+   * @param ctx [in] ctx to clean up
+   */
+  void close_op_ctx(OpContext *ctx) {
+    release_op_ctx_locks(ctx);
+    delete ctx;
+  }
+
+  /**
+   * Releases ctx locks
+   *
+   * @param ctx [in] ctx to clean up
+   */
+  void release_op_ctx_locks(OpContext *ctx) {
+    list<OpRequestRef> to_req;
+    switch (ctx->lock_to_release) {
+    case OpContext::W_LOCK:
+      rw_manager.put_write(ctx->obs->oi.soid, &to_req);
+      break;
+    case OpContext::R_LOCK:
+      rw_manager.put_read(ctx->obs->oi.soid, &to_req);
+      break;
+    case OpContext::NONE:
+      break;
+    default:
+      assert(0);
+    };
+    ctx->lock_to_release = OpContext::NONE;
+    requeue_ops(to_req);
+  }
+
   // replica ops
   // [primary|tail]
   xlist<RepGather*> repop_queue;