func (c *client) findRegionForRPC(rpc hrpc.Call) (proto.Message, error) { // The region was not in the cache, it // must be looked up in the meta table backoff := backoffStart ctx := rpc.Context() for { // Look up the region in the meta table reg, host, port, err := c.locateRegion(ctx, rpc.Table(), rpc.Key()) if err != nil { if err == TableNotFound { return nil, err } // There was an error with the meta table. Let's sleep for some // backoff amount and retry. backoff, err = sleepAndIncreaseBackoff(ctx, backoff) if err != nil { return nil, err } continue } // Check that the region wasn't added to // the cache while we were looking it up. c.regionsLock.Lock() if existing := c.getRegionFromCache(rpc.Table(), rpc.Key()); existing != nil { // The region was added to the cache while we were looking it // up. Send the RPC to the region that was in the cache. c.regionsLock.Unlock() return c.sendRPCToRegion(rpc, existing) } // The region wasn't added to the cache while we were looking it // up. Mark this one as unavailable and add it to the cache. reg.MarkUnavailable() removed := c.regions.put(reg) for _, r := range removed { c.clients.del(r) } c.regionsLock.Unlock() // Start a goroutine to connect to the region go c.establishRegion(reg, host, port) // Wait for the new region to become // available, and then send the RPC return c.waitOnRegion(rpc, reg) } }
func (c *client) waitOnRegion(rpc hrpc.Call, reg hrpc.RegionInfo) (proto.Message, error) { ch := reg.AvailabilityChan() if ch == nil { // WTF, this region is available? Maybe it was marked as such // since waitOnRegion was called. return c.sendRPC(rpc) } // The region is unavailable. Wait for it to become available, // or for the deadline to be exceeded. select { case <-ch: return c.sendRPC(rpc) case <-rpc.Context().Done(): return nil, ErrDeadline } }
func (c *client) sendRPCToRegion(rpc hrpc.Call, reg hrpc.RegionInfo) (proto.Message, error) { client := reg.Client() // On the first sendRPC to the meta or admin regions, a goroutine must be // manually kicked off for the meta or admin region client if reg == c.adminRegionInfo && client == nil && !c.adminRegionInfo.IsUnavailable() || reg == c.metaRegionInfo && client == nil && !c.metaRegionInfo.IsUnavailable() { if reg.MarkUnavailable() { go c.reestablishRegion(reg) } } // The region was in the cache, check // if the region is marked as available if reg.IsUnavailable() { return c.waitOnRegion(rpc, reg) } rpc.SetRegion(reg) // Queue the RPC to be sent to the region var err error if client == nil { err = errors.New("no client for this region") } else { err = client.QueueRPC(rpc) } if err != nil { // There was an error queueing the RPC. // Mark the region as unavailable. first := reg.MarkUnavailable() // If this was the first goroutine to mark the region as // unavailable, start a goroutine to reestablish a connection if first { go c.reestablishRegion(reg) } // Block until the region becomes available. return c.waitOnRegion(rpc, reg) } // Wait for the response var res hrpc.RPCResult select { case res = <-rpc.ResultChan(): case <-rpc.Context().Done(): return nil, ErrDeadline } // Check for errors if _, ok := res.Error.(region.RetryableError); ok { // There's an error specific to this region, but // our region client is fine. Mark this region as // unavailable (as opposed to all regions sharing // the client), and start a goroutine to reestablish // it. first := reg.MarkUnavailable() if first { go c.reestablishRegion(reg) } if reg != c.metaRegionInfo && reg != c.adminRegionInfo { // The client won't be in the cache if this is the // meta or admin region c.clients.del(reg) } return c.waitOnRegion(rpc, reg) } else if _, ok := res.Error.(region.UnrecoverableError); ok { // If it was an unrecoverable error, the region client is // considered dead. if reg == c.metaRegionInfo || reg == c.adminRegionInfo { // If this is the admin client or the meta table, mark the // region as unavailable and start up a goroutine to // reconnect if it wasn't already marked as such. first := reg.MarkUnavailable() if first { go c.reestablishRegion(reg) } } else { // Else this is a normal region. Mark all the regions // sharing this region's client as unavailable, and start // a goroutine to reconnect for each of them. downregions := c.clients.clientDown(reg) for _, downreg := range downregions { go c.reestablishRegion(downreg) } } // Fall through to the case of the region being unavailable, // which will result in blocking until it's available again. return c.waitOnRegion(rpc, reg) } else { // RPC was successfully sent, or an unknown type of error // occurred. In either case, return the results. return res.Msg, res.Error } }