SharedArea -- share memory pages between uWSGI components ========================================================= .. warning:: SharedArea is a very low-level mechanism. For an easier-to-use alternative, see the :doc:`Caching` and :doc:`Queue` frameworks. .. warning:: This page refers to "new generation" sharedarea introduced in uWSGI 1.9.21, the older API is no longer supported. The sharedarea subsystem allows you to share pages of memory between your uWSGI components (workers, spoolers, mules, etc.) in a very fast (and safe) way. Contrary to the higher-level :doc:`caching framework`, sharedarea operations are way faster (a single copy instead of the double, as required by caching) and offers various optimizations for specific needs. Each sharedarea (yes, you can have multiple areas) has a size (generally specified in the number of pages), so if you need an 8 KiB shared area on a system with 4 KiB pages, you would use ``sharedarea=2``. The sharedarea subsystem is fully thread-safe. Simple option VS keyval *********************** The sharedarea subsystem exposes (for now) a single option: ``--sharedarea``. It takes two kinds of arguments: the number of pages (simple approach) or a keyval arg (for advanced tuning). The following keyval keys are available: * ``pages`` -- set the number of pages * ``file`` -- create the sharedarea from a file that will be ``mmap``\ ed * ``fd`` -- create the sharedarea from a file descriptor that will be ``mmap``\ ed * ``size`` -- mainly useful with the ``fd`` and ``ptr`` keys to specify the size of the map (can be used as a shortcut to avoid calculation of the ``pages`` value too) * ``ptr`` -- directly map the area to the specified memory pointer. The API ******* The API is pretty big, the sharedarea will be the de-facto toy for writing highly optimized web apps (especially for embedded systems). Most of the documented uses make sense on systems with slow CPUs or very small amounts of memory. ``sharedarea_read(id, pos[, len])`` Read ``len`` bytes from the specified sharedarea starting at offset ``pos``. If ``len`` is not specified, the memory will be read til the end (starting from ``pos``). ``sharedarea_write(id, pos, string)`` Write the specified ``string`` (it is language-dependent, obviously) to the specified sharedarea at offset ``pos``. ``sharedarea_read8|16|32|64(id, pos)`` Read a signed integer (8, 16, 32 or 64 bit) from the specified position. ``sharedarea_write8|16|32|64(id, pos)`` Write a signed integer (8, 16, 32 or 64 bit) to the specified position. ``sharedarea_inc8|16|32|64(id, pos)`` Increment the signed integer (8, 16, 32 or 64 bit) at the specified position. ``sharedarea_dec8|16|32|64(id, pos)`` Decrement the signed integer (8, 16, 32 or 64 bit) at the specified position. ``sharedarea_wait(id[, freq, timeout])`` Wait for modifications of the specified sharedarea (see below). ``sharedarea_rlock(id)`` lock a shared area for read (use only if you know what you are doing, generally the sharedarea api functions implement locking by themselves) ``sharedarea_wlock(id)`` lock a shared area for write (use only if you know what you are doing, generally the sharedarea api functions implement locking by themselves) ``sharedarea_unlock(id)`` unlock a shared area (use only if you know what you are doing, generally the sharedarea api functions implement locking by themselves) Waiting for updates ******************* One of the most powerful features of sharedareas (compared to caching) is "waiting for updates". Your worker/thread/async_core can be suspended until a sharedarea is modified. Technically, a millisecond-resolution timer is triggered, constantly checking for updates (the operation is very fast, as the sharedarea object has an update counter, so we only need to check that value for changes). Optional API ************ The following functions require specific features from the language, so not all of the language plugins are able to support them. ``sharedarea_readfast(id, pos, object, [, len])`` Read ``len`` bytes from the specified sharedarea starting at offset ``pos`` to the specified object. If ``len`` is not specified, the memory will be read til the end (starting from ``pos``). Currently is implemented only for Perl. ``sharedarea_memoryview(id)`` returns python memoryview object you can directly manipulate (works only on CPython) ``sharedarea_object(id)`` some plugin exposes an alternative way to create sharedareas from internal objects. This functions returns the original object (currently implemented only on CPython on top of bytearrays using ``--py-sharedarea `` option) Websockets integration API ************************** This is currently supported only in the psgi/perl plugin: ``websocket_send_from_sharedarea(id, pos)`` send a websocket message directly from the specified sharedarea ``websocket_send_binary_from_sharedarea(id, pos)`` send a websocket binary message directly from the specified sharedarea Advanced usage (from C) *********************** Work in progress. Check https://github.com/unbit/uwsgi-capture for an example of sharedarea managed from C