Plasma GitLab Archive
Projects Blog Knowledge

Module Netshm

module Netshm: sig .. end

Shared memory for O'Caml programs using multi-processing


Shared Memory Descriptors

type shm_descr 

A shared memory descriptor refers to a shared memory object. Such a descriptor must only be opened once for every process.

type shm_type = [ `File | `POSIX ] 
val supported_types : shm_type list

The types supported for this OS

type shm_name = [ `File of string | `POSIX of string * string ] 

A shm_name is a name for a shared memory object.

val shm_type_of_name : shm_name -> shm_type
val open_shm : ?unique:bool ->
shm_name -> Unix.open_flag list -> int -> shm_descr

Opens the shared memory object.

For POSIX_shm not all open flags can be specified. The flags are limited to O_RDONLY, O_RDWR, O_CREAT, O_EXCL and O_TRUNC.

unique: Unique file generation as in create_unique_shm.

val create_unique_shm : shm_name -> int -> shm_descr

Create a shared memory object under a name that is derived from the passed shm_name. The actual name can be queried with name_of_shm (below).

For POSIX and File names, deriving works by replacing the 'X' letters in the file name by random digits and letters. For example, if the name is File "/my/directory/ocaml.XXXXXX" six random characters are generated and replace the 'X' letters such that the file is unique.

The integer is the file permission.

val name_of_shm : shm_descr -> shm_name

Returns the name of an object

val close_shm : shm_descr -> unit

Closes the object. The object remains existent and can be opened again.

val unlink_shm : shm_name -> unit

Removes the name permanently from the system

val chmod_shm : shm_descr -> int -> unit
val chown_shm : shm_descr -> int -> int -> unit

Set file permission bits, user and group ownership of the object

type locking_method = [ `No_locking | `Record_locking ] 

The locking method is used to ensure that parallel read and write operations to the memory object do not interfer with each other. If No_locking is selected, such protection is not done - this is ok if only read accesses occur or if the user can ensure that never a write access is done in parallel with another access. The locking method must be compatible with the type of shared memory.

val best_locking_method : shm_type -> locking_method

Return the best locking method other than No_locking

Shared Memory Tables

This is a quite basic data structure implemented for shared memory: hash tables with int32 as key and one-dimensional int32 bigarray as values.

The semantics resembles the Hashtbl of stdlib

type shm_table 
type int32_array = (int32, Bigarray.int32_elt, Bigarray.c_layout) Bigarray.Array1.t 
exception Corrupt_file of string

Raised when a violation of the object format is detected

exception Deadlock

Raised when a deadlock situation was detected. Deadlocks can occur when the group function is used

val manage : ?pagesize:int ->
?init:int -> locking_method -> shm_descr -> shm_table

Starts managing an open shared memory object as shm_table. If the object is empty, it is automatically enlarged to the minimum size and initialized. If the object is non-empty it is expected that it already contains a valid shm_table structure.

The object automatically grows in size when new elements are added to the object. By removing elements, however, the object is never shrinked. Unused memory is held back for later reallocation by the same shm_table.

By default, the table uses a page size of 256 bytes. The page size is the unit of memory allocation. The parameter pagesize overrides this default. The page size must be at least 160 and divisible by 4. The page size must be the same when the table was created.

By passing init it is enforced that the table is re-initialized (deleted and again created). The argument of init is a hint for the number of elements the table will contain. The data structure is then created such that this many elements can be added without needing expensive data reorganization.

Special note for using shm_table with multiple processes: Every process must create its own shm_descr, and every process must call manage to manage it. It is not sufficient to just fork a new process, and to keep using the already existing shm_descr or shm_table in the subprocess. (This doesn't work because the underlying file descriptor would be shared.)

val group : shm_table -> ('a -> 'b) -> 'a -> 'b

Execute a sequence of operations in a group:

 let r =
    group table
      (fun arg ->
         operation1; operation2; ...)
      arg
 

Operations can be any reading or writing functions from below. The effect is that the locking requirements of the operations are merged, so that no operation of another process can interfer with the grouped sequence. Note, however, that this gives no real atomicity as there is no way to roll back half-executed sequences.

Groups can be nested.

An example of group is to set a binding depending on the previous value of the binding. Here, we add 1:

 let add_one table =
    group table
      (fun key ->
        let ba =
          try find table key 
          with Not_found -> 
            Bigarray.Array1.create Bigarray.int32 Bigarray.c_layout 1 in
        ba.{ 0 } <- Int32.succ ba.{ 0 };
        replace table key ba
      )
 

Grouping protects the update from modifications done by other processes at the same time. In particular, without grouping it can happen that another process also modifies the same value between find and replace.

val add : shm_table -> int32 -> int32_array -> unit

add tbl key value: Adds the binding of key to value to the table. Previous bindings are not removed, but simply hidden.

val find : shm_table -> int32 -> int32_array

find tbl key: Finds the current binding of key in tbl or raises Not_found if no such binding exists.

val find_all : shm_table -> int32 -> int32_array list

find_all tbl key returns the list of all data associated with key in tbl. The current binding is returned first, then the previous bindings, in reverse order of introduction in the table.

val mem : shm_table -> int32 -> bool

mem tbl key checks if key is bound in tbl.

val remove : shm_table -> int32 -> unit

remove tbl key removes the current binding of key in tbl, restoring the previous binding if it exists. It does nothing if key is not bound in tbl.

val replace : shm_table -> int32 -> int32_array -> unit

replace tbl key value replaces the current binding of key in tbl by a binding of key to value. If key is unbound in tbl, a binding of key to value is added to tbl.

val iter : (int32 -> int32_array -> unit) -> shm_table -> unit

iter f tbl applies f to all bindings in table tbl. f receives the key as first argument, and the associated value as second argument. Each binding is presented exactly once to f. The order in which the bindings are passed to f is unspecified. However, if the table contains several bindings for the same key, they are passed to f in reverse order of introduction, that is, the most recent binding is passed first.

While the iteration is in progress, the table is read-locked. That means you cannot modify it during the iteration.

val iter_keys : (int32 -> unit) -> shm_table -> unit

iter_keys f tbl applies f to all keys in table tbl. If there are several bindings for a key, f is only called once.

While the iteration is in progress, the table is locked. That means you cannot modify it during the iteration.

val fold : (int32 -> int32_array -> 'a -> 'a) -> shm_table -> 'a -> 'a

fold f tbl init computes (f kN dN ... (f k1 d1 init)...), where k1 ... kN are the keys of all bindings in tbl, and d1 ... dN are the associated values. Each binding is presented exactly once to f. The order in which the bindings are passed to f is unspecified. However, if the table contains several bindings for the same key, they are passed to f in reverse order of introduction, that is, the most recent binding is passed first.

While the iteration is in progress, the table is locked. That means you cannot modify it during the iteration.

val length : shm_table -> int

length tbl returns the number of bindings in tbl. Multiple bindings are counted multiply, so length gives the number of times iter calls its first argument.

Enhanced API to shared memory tables

exception Next
exception Break
val read_blocks : shm_table -> int32 -> (int32_array option -> unit) -> unit

find_blocks tbl key f: The values may be stored in several disk blocks. This interface allows one to access the values block by block. As find_all, all bindings for key in tbl are determined in reverse order, i.e. the newest binding first, the oldest last. For every binding value, the function f is invoked in a sequence f (Some v1), f (Some v2), ..., f (Some vN), f None where value is the array concatenation of v1, v2, ..., vN. The function f may raise the exception Next to go directly to the start of the next binding of key. The exception Break stops the iteration immediately.

Note that the int32_array fragments vK point to shared memory. Any assignment would modify the shared memory object directly! The binding is at that time, however, only read-locked, so this should be avoided.

type write_op = [ `Remove_binding ] 
type ctrl_op = [ `Nop | `Remove_binding ] 
val write_blocks : shm_table ->
write_op list ->
int32 -> (int32_array option -> ctrl_op) -> unit

write_blocks tbl ops key f: Like read_blocks this function iterates over the blocks of all bindings for key. For every binding value, the function f is invoked in a sequence f (Some v1), f (Some v2), ..., f (Some vN), f None. Unlike read_blocks the function f returns a value. The last non-`Nop result value in this sequence determines the modification to carry out for the binding:

  • `Remove_binding: The whole binding is removed from the table.

If all invocations of f just return `Nop, no further modification is done.

The modifications must be announced in the ops argument. It is not allowed that f returns a value not being a member of ops (except `Nop).

It is possible to raise the special exceptions Next and Break just as for read_blocks.

Note that the int32_array fragments vK point to shared memory. Any assignment will modify the shared memory object directly! The binding is at that time write-locked, so such assignments are protected against concurrent writes.

val dump : shm_table -> unit
val bigarray : int array -> int32_array
val memory : shm_table ->
(int32, Bigarray.int32_elt, Bigarray.c_layout) Bigarray.Array2.t
This web site is published by Informatikbüro Gerd Stolpmann
Powered by Caml