I am implementing a pool of objects in Delphi. I need to synchronize the threads to get the objects from the pool.
Thread Code:
uClientQueryPool.CLIENT_POOL_GUARD.Acquire();
QueryClient := QUERY_POOL.GetClient();
uClientQueryPool.CLIENT_POOL_GUARD.Release;
Pool Code:
var
CLIENT_POOL_GUARD: TCriticalSection;
type
TClientQueryPool = class
public
function GetClient(): TQueryClient;
end;
The CLIENT_POOL_GUARD is a unit variable. The pool is working well, but can I use "uClientQueryPool.CLIENT_POOL_GUARD.Acquire();" and "uClientQueryPool.CLIENT_POOL_GUARD.Release;" inside the GetClient method?
Like this:
function TClientQueryPool.GetClient: TQueryClient;
begin
CLIENT_POOL_GUARD.Acquire();
...
CLIENT_POOL_GUARD.Release;
end;
Moving the lock inside the get/pop/whatever method is just fine, as is making the CriticalSection instance a private member of the pool class. Use the same CS in the release() call that pushes the objects back onto the pool.
Been doing this for decades, usually with TObjectQueue as the pool queue, a CS to protect it and a semaphore to count the pool contents and something for requesting threads to block on if the pool empties temporarily.
Don't know where that 'double acquire' thread came from. Either the lock is inside the pool class, or outside. I really can't imagine why anyone would code up both!
Example classes:
First, thread-safe P-C queue, for holding the pooled objects:
unit tinySemaphoreQueue;
interface
uses
Windows, Messages, SysUtils, Classes,syncObjs,contnrs;
type
pObject=^Tobject;
TsemaphoreMailbox=class(TobjectQueue)
private
countSema:Thandle;
protected
access:TcriticalSection;
public
property semaHandle:Thandle read countSema;
constructor create; virtual;
procedure push(aObject:Tobject); virtual;
function pop(pResObject:pObject;timeout:DWORD):boolean; virtual;
end;
implementation
{ TsemaphoreMailbox }
constructor TsemaphoreMailbox.create;
begin
inherited Create;
access:=TcriticalSection.create;
countSema:=createSemaphore(nil,0,maxInt,nil);
end;
function TsemaphoreMailbox.pop(pResObject: pObject;
timeout: DWORD): boolean;
begin // wait for a unit from the semaphore
result:=(WAIT_OBJECT_0=waitForSingleObject(countSema,timeout));
if result then // if a unit was supplied before the timeout,
begin
access.acquire;
try
pResObject^:=inherited pop; // get an object from the queue
finally
access.release;
end;
end;
end;
procedure TsemaphoreMailbox.push(aObject: Tobject);
begin
access.acquire;
try
inherited push(aObject); // shove the object onto the queue
finally
access.release;
end;
releaseSemaphore(countSema,1,nil); // release one unit to semaphore
end;
end.
then object pool:
unit tinyObjectPool;
interface
uses
Windows, Messages, SysUtils, Classes,syncObjs,contnrs,
tinySemaphoreQueue;
type
TobjectPool=class;
TpooledObject=class(TObject)
private
FmyPool:TObjectPool;
protected
Fparameter:TObject;
public
procedure release;
constructor create(parameter:TObject); virtual;
end;
TpooledObjectClass=class of TpooledObject;
TobjectPool=class(TsemaphoreMailbox)
private
Fparameter:TObject;
function getPoolLevel: integer;
public
property poolLevel:integer read getPoolLevel;
constructor create(poolDepth:integer;
pooledObjectClass:TpooledObjectClass;parameter:TObject); reintroduce; virtual;
end;
implementation
{ TobjectPool }
constructor TobjectPool.create(poolDepth: integer;
pooledObjectClass: TpooledObjectClass;parameter:TObject);
var objectCount:integer;
thisObject:TpooledObject;
begin
inherited create;
Fparameter:=parameter; // a user parameter passed to all objects
for objectCount:=0 to poolDepth-1 do // fill up the pool with objects
begin
thisObject:=pooledObjectClass.create(parameter);
thisObject.FmyPool:=self;
inherited push(thisObject);
end;
end;
function TobjectPool.getPoolLevel: integer;
begin
access.acquire;
result:=inherited count;
access.release;
end;
{ TpooledObject }
constructor TpooledObject.create(parameter: TObject);
begin
inherited create;
Fparameter:=parameter;
end;
procedure TpooledObject.release;
begin
FmyPool.push(self);
end;
end.
Yes you can. Note, though that although you can pull an object from the pool in a thread-safe manner, it may not be thread-safe to use it if the object itself isn't thread-safe. For instance, in the example below, the pool is thread safe and even makes threads wait if all objects in the pool are in use, but once an object is in use, using it still is not thread safe, because it uses global data.
uses
SyncObjs;
var
GlobalData: Integer = 0;
type
TDataObject = class
Used: Boolean;
procedure UpdateData;
end;
type
TPool = class
FLock: TCriticalSection;
FSemaphore: TSemaphore;
FDataObjects: array[0..9] of TDataObject;
constructor Create;
destructor Destroy; override;
function GetDataObject: TDataObject;
procedure ReleaseDataObject(AObject: TDataObject);
end;
var
Pool: TPool;
type
TDataThread = class(TThread)
constructor Create;
procedure Execute; override;
end;
{ TPool }
constructor TPool.Create;
var
i: Integer;
begin
inherited Create;
FLock := TCriticalSection.Create;
FSemaphore := TSemaphore.Create(nil, Length(FDataObjects), Length(FDataObjects), '', False);
for i := Low(FDataObjects) to High(FDataObjects) do
FDataObjects[i] := TDataObject.Create;
end;
destructor TPool.Destroy;
var
i: Integer;
begin
for i := Low(FDataObjects) to High(FDataObjects) do
FDataObjects[i].Free;
FSemaphore.Free;
FLock.Free;
end;
function TPool.GetDataObject: TDataObject;
var
i: Integer;
begin
Result := nil;
FLock.Acquire;
try
FSemaphore.Acquire;
for i := Low(FDataObjects) to High(FDataObjects) do
if not FDataObjects[i].Used then
begin
Result := FDataObjects[i];
Result.Used := True;
Exit;
end;
Assert(Result <> nil, 'Pool did not return an object');
finally
FLock.Release;
end;
end;
procedure TPool.ReleaseDataObject(AObject: TDataObject);
begin
if not AObject.Used then
raise Exception.Create('Data object cannot be released, because it is not in use.');
AObject.Used := False;
FSemaphore.Release;
end;
{ TDataObject }
procedure TDataObject.UpdateData;
begin
Inc(GlobalData);
end;
procedure TForm1.Button1Click(Sender: TObject);
begin
TDataThread.Create;
end;
{ TDataThread }
constructor TDataThread.Create;
begin
inherited Create(True);
FreeOnTerminate := True;
Resume;
end;
procedure TDataThread.Execute;
var
DataObject: TDataObject;
begin
DataObject := Pool.GetDataObject;
DataObject.UpdateData; // <-- Not thread-safe!
Pool.ReleaseDataObject(DataObject);
end;
initialization
Pool := TPool.Create;
finalization
Pool.Free;
end.
1) I'd remove Acquire/Release code from threads code - it is fragile. In one thread you forget to call it - and ba-bang! Security measures, as a rule of thumb, should be centralized and enforced by server, not distributed in fuzzy way in clients.
2) Acquire/Release calls should be guarded from errors, else any stray exception would forever lock all the threads.
function TClientQueryPool.GetClient: TQueryClient;
begin
CS.Acquire;
try
// actually getting object, preferably just calling
// internal non-public thread-unsafe method for it
finally
CS.Release;
end;
end;
3) Critical section itself should better be a Pool's internal, non-public member. That way you would be allowed in future, when you forget of implementation details, easy refactoring, like:
3.1) implementing several pools
3.2) moving pool code to another unit
3.3) ensuring any stray erroneous code outside pool would not be able to crash the application be randomly acquiring or releasing the CS
4) Double calling of acquire/release over TCriticalSection object puts all your bets over implications from a single note in TCriticalSection documentation, pointed to by The_Fox.
"Each call to Release should be balance by an earlier call to Acquire"
http://docwiki.embarcadero.com/Libraries/en/System.SyncObjs.TCriticalSection.Release
And over the hope that all other Pascal implementations today and tomorrow would not miss it.
That is fragile practice. And multi-threading code is famous for creating Heisenbugs, when there are problems at clients sites, but you can not reproduce and find it in house.
If in future your company would expand to different platform or different language implementation, that puts a potential land mine. And the kind of mine, that would be hard to find by testing in house. Multithreading code is the place where you'd better be over-defeinsive and just do not allow ANY uncertainty to happen.