Cropped response when using body_filter_by_lua with sendfile and aio threads ("task ... already active")

I'm getting incomplete response from OpenResty when using body_filter_by_lua_block together with proxy_cache, sendfile and aio threads. In error log there's alert:

[alert] 1826080#1826080: task #1114 already active

Note 1: it happens only when the traffic goes through physical interface, can't reproduce on loopback
Note 2: remove any of body_filter_by_lua_block, sendfile on or aio threads and the problem no longer appears

Reproduced on OpenResty:

• 1.21.4.1-2~jammy1
• latest openresty with latest lua-nginx-module - wanted to check if this issue was related: #2110 (it's not)

Ubuntu 22.04 and 20.04

nginx.conf:

error_log error.log;

worker_processes auto;

events {
    worker_connections 1024;
}

http {
    sendfile on;
    aio      threads;

    proxy_cache_path cache keys_zone=cache:10m;

    init_by_lua_block {
        static_data = ("."):rep(4000000)
    }
    server {
        listen 127.0.0.1:8001;
        access_log upstream.log;
        location = / {
            content_by_lua_block {
                ngx.header["Content-Length"] = static_data:len()
                ngx.print(static_data)
            }
        }
    }
    upstream upstream {
        server 127.0.0.1:8001;
    }

    server {
        listen 0.0.0.0:8080;

        access_log access.log;

        location / {
            body_filter_by_lua_block {
            }

            proxy_cache cache;
            proxy_pass http://upstream;
        }
    }
}

Run with:

openresty -p ./ -e error.log -c nginx.conf -g 'pid nginx.pid;'

Test with:

curl -sv host:8080 | wc -m

Expected to receive exactly 4 MB, but sometimes (~50% of time) curl complains with e.g.:
transfer closed with 1489665 bytes remaining to read.

Start with the supplied nginx.conf and reproduction command, then run curl against the physical interface while checking error.log for the “task already active” alert. Compare results after removing body_filter_by_lua_block, sendfile, or aio threads, as described. Done means the configured response consistently delivers exactly 4 MB without truncation or the alert.