Always flush buffers after sending a message

[dctrwatson]

Fixes issue of NGX_AGAIN when terminating chunk isn't sent (mainly when sending messages via HTTPS)

[dctrwatson]

Hrm, this is actually not well placed to handle more than my very specific conditions in which I'm hitting this issue

[quentinsf]

Many thanks! This fixed a problem of mysteriously empty responses from the server.

Greatly appreciated!

[wandenberg]

Hi @dctrwatson and @quentinsf ,

I was looking the pull request and also the patch @dctrwatson sent do nginx-devel list. Just to be sure, they are related, right?

Can you help me understand a bit better what is happening and how to reproduce? I know that the module close the connection if the nginx answer with a NGX_AGAIN, and want to do a better treatment to this. But this patch has two problems.

It always allocate a new buffer in request pool, which will be released only when the connection closes. This is a small structure, but multiplied by the number of times that this happens and the number of users connected to the server can be a big problem.

ngx_http_send_special(ngx_http_request_t *r, ngx_uint_t flags)
{
    ngx_buf_t    *b;
    ngx_chain_t   out;

    b = ngx_calloc_buf(r->pool); 

The other problem in fact is that I didn't understand what is the difference between what I already do, since I set the buffer to be flushed on ngx_http_push_stream_send_response_text function, and the use of ngx_http_send_special function.

May be that the problem itself was solved with the patch John applied on nginx chunked filter.

I really need your help to understand the problem and how to reproduce it here to do a proper solution. Together I'm sure we can find a solution.

[wandenberg]

Hi @dctrwatson and @quentinsf

can you test the code on fix_ngx_again_on_read_and_write branch (commit df6b37a). It should solve the problem with the ngx_again without use the ngx_http_send_special function which allocates extra memory.

[wandenberg]

Fixed with a48ca949427b782fba1722932f4881850c50b9f1