APC_MEMPROTECT macro has no tests, won't work with apc.serializer=default

[TysonAndre]

Currently, I assume it'd be manually set with -DAPC_MEMPROTECT=1, but I wouldn't recommend that since I assume nobody has used this in years?

It may be of use to have a way to mark the returned pointers as read-only (to crash as early as possible instead of having applications misbehave when memory is being used in a way it shouldn't, e.g. due to bugs in pecls or unserializers)

• That type of bug seems really, really unlikely in practice, and apcu always copies arrays and strings when reading from shared memory

[TysonAndre]

The SHM_RDONLY flag is a no-op in TSRM/tsrm_win32.c, so it might also make sense to error and fail with a message saying it does nothing on windows

[TysonAndre]

The SHM_RDONLY flag is a no-op in TSRM/tsrm_win32.c, so it might also make sense to error and fail with a message saying it does nothing on windows

Replacing it with VirtualProtect and mprotect similar to what opcache.protect_memory=1 does seems like a good alternative, which would work on both mmap and shmat with anonymous shared mappings (default, /dev/zero) where APC_MEMPROTECT only worked with named files

Refer to https://github.com/TysonAndre/immutable_cache-pecl/pull/11/files for how this might be implemented (though I expect APCu would be more complicated, e.g. have edge cases for cache clears) - this would be done by moving locks/mutexes into a separate shared memory region (go from n -> n+1 regions) (because acquiring mutexes requires the ability to write to shared memory)