Shared heaps of structured values
These heaps live in [root:Netmcore_mempool]-type shared memory pools, and can contain an arbitrary number of Ocaml values. These values can be mutable, but special care has to be taken when modifying them. The first value pushed onto the heap is called the root element. All other values must be (directly or indirectly) reachable from the root element.
Heaps are created with a certain initial size. The heaps remain connected with the memory pool, and they are enlarged if necessary by allocating more blocks in the pool.
As the heaps are shared between processes, it must be taken care that no reference is made from shared heaps to normal process-local memory. These references would only be valid in the process creating them, and following such references from other processes would probably crash the program (or at least return wrong values). In order to ensure this, it is strictly forbidden to directly manipulate mutable data structures. The Netmcore_heap.modify function has to be used, as this function makes it possible to copy more values to the heap. Unfortunately, there is nothing in the type system that would prevent direct mutation. so this can only be ensured by the discipline of the programmer.
The values of the heap are also garbage-collected: If all allocated space is used and more values need to be added, it is first tried to get rid of old unreachable values. The garbarge collection is done by the process that happens to initiate the addition of the value that does no more fit onto the heap. During garbage collection, no other additions can be done, but read accesses are not prevented. The garbage collector does not move values (addresses remain unchanged).
The garabage collector only considers values as reachable that are reachable via the root element. It is not sufficient when a value is only reachable via a process-specific reference.
A heap where the type of the root element is 'a
A descriptor ("address") pointing to the heap. Descriptors can be marshalled.
create_heap pool_id size root: Creates a new heap with size
bytes in the pool identified by pool_id. This ID must refer
to a [root:Netmcore_mempool]-managed pool.
The value root is copied to the new heap. This is done by
deeply duplicating root and all values pointed to by root,
and storing these duplicates in the heap.
The possible types of value root are restricted, see the add
function for more.
Returns the size value one must pass to create_heap at minimum
to put this root element onto the heap.
Mutators allow it to push new values onto the heap.
Caveat: pushed values are not considered as roots, and thus they
need immediately be attached to the existing data structure.
Otherwise, the next push might trigger a garbage collection, and
the new value is deleted. If this is not possible, one can
call pin instead (see below).
modify h mutate: This function locks the heap so that this process
has exclusive write access to it for the duration of the mutate
function. The mutate function is immediately called back, and
the argument of mutate is the mutator m that allows one to push
values onto the heap.
By calling add m x from the body of mutate one can create a copy
of x that is stored in the heap.
Pushes a new value onto the heap. This creates a deep copy of the value.
Not all values can be pushed here. In particular, forbidden are:
in_channel and out_channel.
Supported custom blocks are only int32, int64, nativeint,
and bigarrays.Pushes a new value onto the heap. This function must only be used if the added value is immutable. An important optimization is applied here: if parts of the value are already living on the heap, these parts are not copied, but shared with the output value.
The same value restrictions apply as for add. Note that
add_immutable cannot be used if the value to copy lives in a different
heap of the same pool (as a whole or partially). In this case use
add.
add_uniform_array m n x: Pushes a new value with n elements onto
the heap. Each index position of the array is initialized with
the same copy of x.
You should not call this function with n=0, because this results in
a copied atom, which is an illegal representation in OCaml.
add_init_array m n f: Pushes a new value with n elements onto
the heap. The index position k is inititialized by running
f k and pushing the copy of this onto the heap.
You should not call this function with n=0, because this results in
a copied atom, which is an illegal representation in OCaml.
add_some mut x: Returns Some x where the O'Caml value representing
Some is allocated in the heap using mut. It is assumed that x is
already a resident of the heap. This means x is not copied!
let s = add_string mut len: Adds an uninitialized string of length
len to the heap using mut, and returns the string
pin m x: Pins a shared value x so it cannot be deleted by
the garbage collector. The value remains pinned for the lifetime
of the mutator m (i.e. the runtime of the modify function).
Pinning is relatively expensive if done in masses, and should be avoided if possible.
Creates a deep copy of the input value, and stores the duplicate in normal process memory.
with_value h find process: Logically, this runs
process (find ()) and returns the result. While find is being
executed, the heap is write-locked. This returns a value x.
While process is being executed, the value x is temporarily
added to the set of reachable values, so that a parallely running
garbage collection will not delete it.
Note that x must reside in the heap!
Calling modify from find will cause a deadlock. Calling
it from process is allowed.
Same as with_value, but a tuple of values can be passed down
Same as with_value, but a list of values can be passed down
This example creates a heap that stores a single value. (This is available as [root:Netmcore_ref].)
let shared_ref x =
(* The shm version of [ref x] *)
let r = ref x in
let init_size = minimum_size r in
let hp = create_heap pool_id init_size r in
hp
let deref sref =
(* The shm version of [!] *)
!(root sref)
let assign sref x =
(* The shm version of [:=] - however, a copy of x is done *)
modify sref
(fun add ->
(root sref) := add x
)