worker process suffers CPU 100% usage

[tokers]

Hello!

The worker process CPU usage could be 100% if we use the following configuration pattern:

stream {
    server {
        listen 3105 reuseport;

        lua_check_client_abort on; # must be enabled

        content_by_lua_block {
            local sock, err = ngx.req.socket(true)

            sock:settimeout(1800 * 1000)

            local readline = sock:receiveuntil("\n")

            while not ngx.worker.exiting() do
                local line, err = readline(1)
                ngx.log(ngx.ERR, "line: ", line)
                if not line then
                    if err then
                        ngx.log(ngx.ERR, err)
                        break
                    end
                else
                    ngx.sleep(5000) -- other yieldable API is also applicable.
                end
            end
        }
    }
 }

Building a connection through nc and sending some characters then closing it, in the meanwhile, observing the CPU usage by top.

$ nc 127.0.0.1 3105
asjdkl
zxckj
^C

Now you can find the CPU usage reaches 100%.

It's irrelevant with the version of ngx_stream_lua, you can reproduce this problem even with the newest OpenResty.

After investigating the internal of stream-lua-nginx-module, I found it was caused by the indefinitely trigger of ngx_stram_lua_check_broken_connection, which doesn't the check ev->pending_eof flag.

BTW, this issue was found by @tianchaijz.

[tokers]

I have written a patch to solve this problem, I will push it to the meta-lua-nginx-module repo.

[tokers]

I have written a patch to solve this problem, I will push it to the meta-lua-nginx-module repo.

Nonetheless, I don't know why the ngx_stram_lua_check_broken_connection would be triggerd duplicately under the epoll ET mode.

[tokers]

@dndx Any idea about this?