Re: latest nightly build errors => ebadf problem

[Sam S. <sd...@gn...>]

Joe...@t-... wrote:
> 
>>From the CLISP POV what's important is that CLISP never looses sync
> with the OS about the state of a file handle. Several years ago, a bug was
> fixed in CLISP where the wrong fd was closed (or closed twice?), because sync was lost.
> Imagine if closing a random connection would happen to your long running servers!
> 
> The bug went like this:
> open A -> fd5
> open B -> fd6
> close A (5)
> open C -> fd5 again
> Then somehow, CLISP wanted to close A again (finalization?), i.e. fd5.
> The end result was that the completely unrelated stream C got closed.

this is a realistic scenario.
e.g., close_ichannel calls ChannelStreamLow_close (which calls close(2))
but then it calls ChannelStream_fini (which does some iconv finalization) which 
could signal an error; i.e., the stream could end up not marked as closed even 
though close(2) has already been called.
maybe we should call ChannelStream_fini _before_ ChannelStreamLow_close?
Bruno, why did you chose to close the socket before the encoding?