Thread Safety¶
Per libcurl thread safety documentation, libcurl is thread-safe but has no internal thread synchronization.
For Python programs using PycURL, this means:
Accessing the same PycURL object from different threads is OK when this object is not involved in active transfers, as Python internally has a Global Interpreter Lock and only one operating system thread can be executing Python code at a time. On free-threaded CPython, the GIL is no longer present; the same rules apply, but the caller must serialise concurrent access to a shared PycURL object explicitly.
Accessing a PycURL object that is involved in an active transfer from Python code inside a libcurl callback for the PycURL object in question is OK, because PycURL takes out the appropriate locks.
Accessing a PycURL object that is involved in an active transfer from Python code outside of a libcurl callback for the PycURL object in question is unsafe.
CurlShareis thread-safe: methods may be called concurrently from multiple threads, and differentCurlhandles attached to the same share may be used from different threads. Edge cases:close()is not fully thread-safe with respect to associatedCurlhandles — see itsdocsfor the rules it enforces.Not every kind of libcurl shared data is safe to share across threads. See CURLSHOPT_SHARE for the list of supported
CURL_LOCK_DATA_*values and their constraints.
Not every kind of libcurl shared data is safe to share across threads. See CURLSHOPT_SHARE for the list of supported
CURL_LOCK_DATA_*values and their constraints.A WebSocket handle (
CONNECT_ONLY=2plusws_send/ws_recv) follows the same one-handle-one-thread rule: do not callws_sendfrom one thread whilews_recvruns on another. Serialise access with a lock. Calls from another thread whileperform()is running are unsafe unless they happen inside that handle’s activeWRITEFUNCTIONcallback.
PycURL handles the necessary SSL locks for OpenSSL/LibreSSL/BoringSSL, GnuTLS, NSS, mbedTLS and wolfSSL.
A special situation exists when libcurl uses the standard C library
name resolver (i.e., not threaded nor c-ares resolver). By default libcurl
uses signals for timeouts with the C library resolver, and signals do not
work properly in multi-threaded programs. When using PycURL objects from
multiple Python threads NOSIGNAL option must be given.