permessage_deflate seems not thread safe

[KyleNyenhuis]

For a given server (asio transport) with permessage_deflate turned on it eventually (quickly) sends corrupted messages. I believe this is due to the fact that I'm "send()"ing messages on threads other than my server's io_service thread.

At first glance in extensions/permessage_deflate/enabled.hpp I see the buffer is used for both compress and decompress:
size_t m_compress_buffer_size;
lib::unique_ptr_uchar_array m_compress_buffer;

But when I attach a debugger I'm not finding any locks in my backtrace that would prevent a compress() and decompress() from occurring at the same time. (on different threads)

Moving all send() calls to the main io_service thread corrects the issue for me.
I thought I read that it should be okay to send() outside of the io_service thread. Am I missing somthing?

[pdgendt]

I think we have this issue as well. The websockets are closed at random times, with code 1006 or 1007 and reason, Could not decode a text frame as UTF-8 or Invalid frame header.

If we disable the deflate setting, then the connection is kept open but if we enable it we see a close event after 5~60 minutes.

Our application is multi-threaded where all relevant threads execute io_service.run().

[zaphoyd]

It definitely should be safe to call send on separate threads than the main I/O thread. The situation with the compress buffer definitely violates this. I've applied an update very similar to one @pdgendt posted in #633 to the develop branch that should fix this.