[ 
https://issues.apache.org/jira/browse/THRIFT-4552?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16437123#comment-16437123
 ] 

gansteed commented on THRIFT-4552:
----------------------------------

That's what I mean in last comment, just like you demo. in 
https://issues.apache.org/jira/browse/THRIFT-4243 , it does things below:

- go serv.Serve()
- serv.Stop()

wg.Add and wg.Done will only occurred in `innerAccept` which is called by 
`AcceptLoop`
wg.Wait will only occurred in `Stop`

so, if you call `go serv.Serve()`, it will spawn a goroutine, and the goroutine 
will block on `client, err := p.serverTransport.Accept()` until there comes a 
new request. and the main goroutine which execute `serv.Stop` will do:

1. load p.closed and check if it is already closed, if not, set it as closed
2. call `p.serverTransport.Interrupt()`
3. call `p.wg.Wait` to wait all spawned goroutines(which handle RPC requests)

the only possible race panic will occur in such a way:

1. main goroutine call `serv.Stop`, which then call `if 
atomic.LoadInt32(&p.closed) != 0`, in the mean time, there comes a request, and 
the goroutine which handle it call `closed := atomic.LoadInt32(&p.closed)`
2. the main goroutine call `atomic.StoreInt32(&p.closed, 1)`, 
`p.serverTransport.Interrupt()`, then `p.wg.Wait()`
3. and the goroutine which handle the request call `p.wg.Add(1)`

then it panic!

so, the solution is use https://golang.org/pkg/sync/atomic/#CompareAndSwapInt32 
instead of two steps(first load, then store).

may I sent a PR in Github to solve this problem?

> why acquire a lock in TSimpleServer implementation for Go?
> ----------------------------------------------------------
>
>                 Key: THRIFT-4552
>                 URL: https://issues.apache.org/jira/browse/THRIFT-4552
>             Project: Thrift
>          Issue Type: Improvement
>    Affects Versions: 0.11.0
>            Reporter: gansteed
>            Priority: Major
>
> I've sent a email to groups, but I think maybe here will be better?
>  
> I'm using Thrift and I'm reading thrift implementation for Go, I found code 
> in `TSimpleServer.AcceptLoop` like this:
>  
> {code:go}
> func (p *TSimpleServer) AcceptLoop() error {
>         for {
>                 client, err := p.serverTransport.Accept()
>                 p.mu.Lock()
>                 if atomic.LoadInt32(&p.closed) != 0 {
>                         return nil
>                 }
>                 if err != nil {
>                         return err
>                 }
>                 if client != nil {
>                         p.wg.Add(1)
>                         go func() {
>                                 defer p.wg.Done()
>                                 if err := p.processRequests(client); err != 
> nil {
>                                         log.Println("error processing 
> request:", err)
>                                 }
>                         }()
>                 }
>                 p.mu.Unlock()
>         }
> }
> {code}
>  
> every time it accept a request,it:
>  
> 1. read if protocol had been closed, this step is atomic, it does not need a 
> lock.
> 2. p.wg.Add(1) to accumulate a goroutine? this step is atomic, too, it does 
> not need a lock
> 3. after processor processed the request, it do p.wg.Done(), it's atomic, 
> too, and it does not need a lock.
>  
> by the way, it seems that `p.wg.Done()` do not need to put in defer? just put 
> it after p.processRequests(client)?
>  
> so is there any particular to do it in this way?if not, I would like to 
> submit a PR to reduce unneccessary performance overhead in TSimpleServer 
> implementation.
>  
>  



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to