UEFI News and Commentary

Saturday, August 11, 2012

UEFI Application Programming: System Object Library (Part 1)

This article is the first in a series on writing UEFI applications. There are examples available on www.tianocore.org, but very little in terms of explaining what is going on. I tried to do a better job of this in the book that I co-wrote, Harnessing The UEFI Shell, where I took real, working programs and dissected them line-by-line. But that book suffered from working with an unfinished Shell 2.0 and I had to make a number of compromises and assumptions.

One of the key pieces of working in UEFI is having a base set of libraries that provide basic functionality. The EDK2 has done a good job of creating functions that give easy access to some UEFI features. And the StdLib is a fully-functioning version of the C standard library. And the Nt32Pkg provides a decent Windows-hosted UEFI environment that is really useful for debugging the applications.
My irritation is that I learned C++ before I learned C. So I grew used to having a set of container classes, whether MFC from Microsoft or the Standard Template Library (STL) that (now) comes with every C++ compiler. In C++, container classes like strings, lists, arrays, buffers/streams and maps/dictionaries manage collections of data structures, including all of the associated memory management.

But now I work with UEFI. UEFI is tied closely to C. The open-source implementations don’t support C++, including the necessary library to support to handle operators like new and delete. I like containers, and I have used them throughout my programming career.
So the first thing I did when it came to writing UEFI applications was to create a library to manage real containers. The library, called SysLib, was designed with a few goals in mind. And yes, I’ll publish the source code.

1)      Easily be converted to C++. This meant that the functions all used the pointer to the container as the first (‘this’) parameter. This allowed my C containers to be easily converted into C++ classes.
2)      Non-invasive. For many C container implementations, putting a data structure in the container requires the data structure be modified, usually with container’s book-keeping information. Then there is often weird pointer math to deduce the pointer to the beginning of the data structure from the pointer of that book-keeping data. An example of this is the EDK2’s LIST_ENTRY data structure from BaseLib.
3)      Constructor/Destructor. In C++, a class can provide a constructor, which initializes the class’ members to some default value. In C++, a class can also provide a destructor, which is responsible for releasing all resources allocated during the lifetime that object.
4)      Run-Time Type Information. In C++, run-time type information can be used to query whether an object is of a specific type. This allows the user to take advantage of specific capabilities of that object, including verifying that the pointer is a pointer to a valid container object.
5)      Debug Support. I wanted to be able to query any container to see if it was valid and to dump out the container’s contents.
6)      Virtual Functions. In C++, this is handled through a hidden data structure member that points to a table of pointers.  In UEFI, this has been simulated through (non-hidden) function pointers.
7)      Inheritance. All containers in this library derive from a common base class. It is possible to extend a class via a sort of “single-inheritance” by placing the data structure for a container at the beginning of a new container’s data structure and then pointing the virtual function pointer members to a new function. Since the C compiler doesn’t support inheritance, there is always some sort of type-casting involved (to get the parent type) but it works. Primitive, but effective.
8)      Light-Weight. Hey, let’s not take all of our valuable ROM resources or single-threaded UEFI computing time in making this work, ok?
Pretty nice list.  Now let’s look at some features that are not supported:
1)      Data hiding. This is C. So all structure members are available for inspection. Of course, you can always use the PIMPL idiom if data-hiding is important, but I don’t use it.
2)      Multiple Inheritance. I never used it in C++. When I first disassembled C++ code and looked at how it was really implemented, I started feeling queasy and decided I’d use aggregation instead.

SysObj

So the first step is to introduce you to SysObj, which is the parent of all of the container classes that I will introduce in future articles. This structure is designed to sit as the first member of a container class structure.
typedef struct _SYS_OBJ {
  UINT32 Signature;                     // object-specific signature.
  UINT32 Size;                          // size of object, in bytes.
  SYS_OBJ_INIT Init;                    // initialize new object.
  SYS_OBJ_EMPTY Empty;                  // free all allocated resources.
  SYS_OBJ_COPY Copy;                    // create copy at new location.
  SYS_OBJ_VALID IsValid;                // check object validity.
  SYS_OBJ_DUMP Dump;                    // debug dump of object contents.
} SYS_OBJ;
 
So, let’s walk through the data structure, piece-by-piece and then I’ll introduce you to the handy library functions and macros that make life easy.

·         Signature. The signature is a unique 32-bit integer that identifies the object type. Each of the classes that I will introduce later has a signature and uses that number of verify that the object pointer that is passed is of the correct type. This is primitive run-time type information. I considered using a pointer to a string, but I wanted the type verification to be low-overhead.
·         Size. The size of the object, in bytes. This includes the size of the SYS_OBJ structure. While not strictly necessary, it provides a convenient way to do pointer validation and the default implementation of Init() and Dump().
·         Init. A pointer to the virtual initialization function, which initializes a block of memory so to be a valid, empty object. All of the members of this structure must also be set. It does not consider what was in the memory already.
·         Empty. A pointer to the virtual empty function, which frees all resources allocated within the object, but not the object itself. The default version simply returns.
·         Copy. A pointer to the virtual copy function, which copies the contents of an object to a new memory location. The default version simply copies the entire object, byte-for-byte.
·         IsValid. A pointer to the virtual validity-checking function, which checks whether a pointer in fact points to a valid object. The default version just checks to see if the pointer is non-NULL and that Size is, at least, the size of the SYS_OBJ structure.  Most objects implement a version which uses the helper function SysObjIsValid(), which also checks the signature against a known value.
·         Dump. A pointer to the virtual debug-dump function, which prints out the contents of the object as text to the console. When I’m debugging, it is often useful to just dump out an object so I can see what it contained at a specific point in time. By providing this as a fundamental part of SysObj, I get, at least, some debug dump capabilities, because the default version, SysObjDump(), will print out the signature and a hex dump of the object’s contents. More advanced versions can add more details.
SysObj Macros
Now, let’s take a look at some of the useful macros that automate using some of the typical uses of objects:

·         SysInit(type, ptr). Initialize a previously allocated block of memory (ptr) as an empty object of type type. This macro is used, for example, to initialize global or local (stack) objects. Since there is no way to automatically invoke a “constructor” in C when an object is declared, we have to do it manually. And this is the macro that does it.
·         SysInitWith(ptr1, ptr2). Initialize a previously allocated block (ptr1) with the contents of another object (ptr2).
·         SysNew(type). Return a pointer to a newly allocated object of type type, or NULL if initialization failed or memory could not be allocated.
·         SysDelete(ptr). Free all resources associated with the object (ptr) and free the object. This is used with objects which were allocated with SysNew. Global or local objects initialized with SysInit should use SysEmpty instead.
·         SysEmpty(ptr). Free all resources associated with the object (ptr) and make it appear “empty” but do not free the object. This is used with objects which were initialized with SysInit().
·         SysCopy(ptr1,ptr2). Copy the contents of object ptr2 into object ptr1. This uses the virtual Copy function.
·         SysDup(ptr).Return a pointer to a newly allocated copy of object ptr.
·         SysIsValid(ptr). Return whether the object ptr is a valid object.
·         SysDump(ptr). Dump the contents of the object ptr to the console.
·         SysDebugDump(ptr). Dump the contents of the object ptr to the console if library debugging is enabled. The global flag (SYSLIB_DEBUG) controls whether library debugging is enabled. I haven’t updated this to use EDK2’s PCDs yet. Yes, it is on my todo list.

SysObj Functions

This section walks through key support functions provided by the library for new objects:

SysObjInit

This function is used by Init() virtual member functions to set up the SysObj members to correct values. It also performs some basis error checking. Two important notes: the SYS_INFO is a debug macro used for informational output. It can be separately turned on or off. Also, four of the virtual functions can be NULL, wherein they will provide a default function which assumes no allocated resources and byte-by-byte copy.
SYS_OBJ *
SysObjInit (
  OUT SYS_OBJ *Obj,
  IN UINT32 Signature,
  IN UINT32 Size,
  IN SYS_OBJ_INIT Init,
  IN SYS_OBJ_EMPTY Empty OPTIONAL,
  IN SYS_OBJ_COPY Copy OPTIONAL,
  IN SYS_OBJ_VALID IsValid OPTIONAL,
  IN SYS_OBJ_DUMP Dump OPTIONAL
  )
{
  if (Obj == NULL) {
    SYSINFO ("invalid object pointer. pointer is NULL.\n");
    return NULL;
  }

  if (Init == NULL) {
    SYSINFO ("invalid object Init function pointer.\n");
    return NULL;
  }

  if (Size < sizeof (SYS_OBJ)) {
    SYSINFO ("invalid object size. must be at least %d bytes (%d specified).\n", \
      sizeof (SYS_OBJ), Size);
    return NULL;
  }

  Obj->Signature = Signature;
  Obj->Size = Size;
  Obj->Init = Init;
  Obj->Empty = (Empty == NULL) ? SysObjEmptyDummy : Empty;
  Obj->Copy = (Copy == NULL) ? SysObjCopy : Copy;
  Obj->IsValid = (IsValid == NULL) ? SysObjIsValidDummy : IsValid;
  Obj->Dump = (Dump == NULL) ? SysObjDump : Dump;
  return Obj;
}

SysObjDump

This is the default version of the Dump() virtual member function, which prints out the signature and then a hex dump of any object data.

VOID
SysObjDump (IN CONST SYS_OBJ *p)
{
  UINT32 i;
  UINT8 *b;
 
  if (!p->IsValid (p)) {
    printf ("invalid object pointer.\n");
  }
  printf ("Object: %c%c%c%c\n",
    (char) (p->Signature & 0xff),
    (char) ((p->Signature & 0xff00) >> 8),
    (char) ((p->Signature & 0xff0000) >> 16),
    (char) ((p->Signature & 0xff000000) >> 24));
  printf ("Size  : %d bytes\n", p->Size);
 
  b = (UINT8 *) (p + 1);                // points just after end of SYS_OBJ structure.
  for (i = 0; i < p->Size - sizeof (SYS_OBJ); i++) {
    printf ("0x%02x ", b[i]);
    if (i % 16 == 15) {
      printf ("\n");
    }
  }
}

SysObjIsValid

This is a helper function that can be used by objects to implement the IsValid() member functions. It verifies that there is a valid pointer, that the size is at least the minimum size and that the signature matches.

BOOLEAN
SysObjIsValid (
  IN CONST SYS_OBJ *Obj,
  IN UINT32 Signature,
  IN UINT32 Size
  )
{
  if (Obj == NULL) {
    SYSINFO ("invalid object pointer. pointer is NULL.\n");
    return FALSE;
  }
 
  if (Obj->Size < sizeof (SYS_OBJ) || Obj->Size != Size) {
    SYSINFO ("invalid object pointer. invalid object size. must be %d bytes \
     (%d found).\n", Size, Obj->Size);
    return FALSE;
  }
 
  if (Obj->Signature != Signature) {
    SYSINFO ("invalid object pointer. invalid signature.\n");
    return FALSE;
  }
 
  return TRUE;
} 

SysObjCopy

This is the default version of the Copy() member function, which performs a simple byte-by-byte copy.

SYS_OBJ *
SysObjCopy (
  OUT SYS_OBJ *ObjDest,
  IN CONST SYS_OBJ *ObjSrc
  )
{
  return (SYS_OBJ *) memcpy (ObjDest, ObjSrc, ObjDest->Size);
}

Conclusion

Some of you are probably wondering what this has to do with UEFI. Good question. In fact, these libraries could work just as well under Windows, because they only depend on the C library. But I have packaged them for EDK2 build and they are the building blocks for more serious UEFI applications to come, including some fun HII code.

In the next article, we’ll look at a few actual container classes built on top of SysObj, including SysList, a basic double-linked list container and then SysOList, which extends this to manage (“own”) the objects in the list.
P.S. Some of you sharp-eyed readers will note that I said, “allowed my C++ containers” implying that I had done the work to make a C++ wrapper. Well, I have, but it involved a few hacks to EDK2, including the build tools and libraries. If I have time, I’ll try and point out how I did it.

The Nitty-Gritty EDK2 Library Details

Most of you can stop reading now, since most of what follows gives the low-level details of how I got this to build in EDK2. Don’t worry, you won’t miss anything, since I’ll show it all later, but I thought I’d give a quick tour.

SysLib.dec

The whole library is in a package called SysLib. Every package in EDK2 needs a .DEC file. This one is about as simple as it gets.

[Defines]
  DEC_SPECIFICATION              = 0x00010005
  PACKAGE_NAME                   = SysLib
  PACKAGE_GUID                   = 1842ace0-5d82-11e1-b86c-0800200c9a66
  PACKAGE_VERSION                = 0.01

[Includes]
  Include

SysLib.inf

Here’s the actual INF file used to build the library.

[defines]
  INF_VERSION                    = 0x00010005
  BASE_NAME                      = SysLib
  FILE_GUID                      = 4847bcc0-5d85-11e1-b86c-0800200c9a66
  MODULE_TYPE                    = UEFI_APPLICATION
  VERSION_STRING                 = 1.0
  LIBRARY_CLASS                  = SysLib
 
#
#  VALID_ARCHITECTURES           = IA32 X64 IPF
#
 
[LibraryClasses]
  LibC
  DebugLib
  PrintLib
 
[Packages]
  MdePkg/MdePkg.dec
  StdLib/StdLib.dec
  SysLib/SysLib.dec
 
[Sources]
  SysLib.c
  Debug.c
  CLibSupplement.c
  Uefi.c 

Nt32Pkg.dsc

This is the master DSC file for building the Windows-hosted UEFI environment. Most of this is pretty standard: add the library class (SysLib) to the [LibraryClasses] section.

  SysLib|SysLib/Source/SysLib.inf
 
But when writing a new library which both works with Nt32 and uses the StdLib, you have to tell the EDK2 build environment to ignore the standard Visual Studio build paths, as follows (in the [Components] section) using the /X command-line option.

  SysLib/Source/SysLib.inf {
   
      MSFT:*_*_*_CC_FLAGS    = /X /Zc:wchar_t /GL-
  } 


3 comments:

Finnbarr P. Murphy said...

Interesting post. However I am not sure what value such a C++ library adds to coding a UEFI application.

Tim Lewis said...

Good question. I needed a good library to handle the data structures that my later UEFI applications that deal with HII use. For example, representing the HII database, or HII strings and fonts, etc. I had a lot of experience doing this in C++ before, but EDK2 doesn't support that. So I created C libs that did the same sort of thing. Along the way, I also figured out how to make EDK2 support C++, but decided to focus on getting my HII stuff blogged about first.

süli said...

A lot of very helpful information on this blog, thanks a lot!!

And I'm very curious about this:
"Along the way, I also figured out how to make EDK2 support C++"