If both modules should be used at the same time, the most problematic point would be calling curl_global_init() several times IMO (although the libcurl documentation states that it must be called "at least" once). However, async_http uses the curl_global_init_mem() version (it allows curl to use kamailio's memory API instead of the libc one), and if it is called after curl_global_init(), it will have no effect.
We chose to bind libcurl with the SHM memory API, because the amount of memory needed to handle the HTTP requests and responses under heavy load can be huge, and it made more sense to increase the shared memory pool size than the private ones. The drawback is the cost of the SHM locking for each operation. Another option could be to allocate a new dedicated memory pool to each HTTP worker (or a shared one amongst the workers), this is now easily doable with the new kamailio memory API.

So, in a nutshell, to use both modules at the same time, libcurl initialization should be done in a shared API that would ensure that the initialization is done only once per process. To do this, we also have to agree on the memory API used by libcurl. In your opinion(s), what is the best alternative: libc's malloc, SHM, PKG, dedicated (private / shared) pool?

—
Reply to this email directly or view it on GitHub.