HOWTO: Map Adapter RAM into Process Address Space

ID: Q189327

The information in this article applies to:

Microsoft Win32 Device Driver Kit (DDK) for Windows NT, version 4.0

SUMMARY

You might run into one of the following common problems while mapping a large amount of adapter memory:

Unavailability of system PTEs (NO_MORE_SYSTEM_PTES).

Not enough contiguous process address space.

This article explains why these problems occur and how to circumvent them. It also explains two ways to map adapter memory into a process address space.

MORE INFORMATION

Introduction to Page Table Entries (PTE)

Every process that runs on a Windows NT system has a 4 GB virtual address ranging from 0x00000000 to 0xFFFFFFFF for its use. Of this, the upper 2 GB address ranging from 0x8000000 to 0xFFFFFFFF is common to all processes running in the system, and it is called kernel or system address space. The lower region ranging from 0x00000000 to 0x7FFFFFFF is called user address space.

From the process perspective, each element of virtual address conceptually refers to a byte of physical memory. It is the responsibility of the Virtual Memory Manager (VMM) in conjunction with processor memory manager unit (MMU) to translate or map each virtual address into a corresponding physical address. The VMM performs the mapping by dividing the RAM into fixed-size page frames, creating page tables to store information about these page frames, and mapping them. Each PTE represents a page frame and contains information necessary for the VMM to locate a page.

On an x86-based system that uses a 4 KB page size, the maximum number of PTEs required to map 2 GB of address space is 524,288 (2 GB/4 KB). On a typical system, this space is used as follows: a maximum of 50,000 PTEs (approximately 195 MB address space) are reserved for general use and the rest is used in mapping system cache, hyperspace, paged pool, nonpaged pool, crash dump area, and so on. This PTE pool size is automatically determined at system startup based on the amount of physical memory in the system. This pool is squeezed in between paged pool and nonpaged pool, which also grows with the amount of physical memory in the system as well.

The system uses these PTEs to create kernel thread stacks, load device drivers (and their DLLs), to map system virtual address space for I/O transfers or callers of MmMapIoSpace/MmMapLockedPages/ MmGetSystemAddressForMdl/MmAllocateNonCachedMemory, and for other miscellaneous purposes. This system PTE pool can become heavily used and heavily fragmented. This means that your driver might not be able to get enough contiguous virtual address space from the system PTE pool at any given time, even though the total address space still remaining in it might be large enough. It also means that if your driver uses system PTE pool all up, other parts of the system will degrade, even resulting in threads not being created, system stalls, and outright bug checks (because some drivers call allocate-system-memory with the MustSucceed parameter set).

The bottom line is that you must be extremely careful when using this pool- only map. You should use the portions of the adapter RAM that you really need system access to from any process context, and only map the amount you need. Do not map the whole adapter range if you do not really need to access it all from system mode.

IMPORTANT: Unmap the memory as soon as you are done with it in the right process context. Otherwise, the system will run out of PTEs, and it will bug check. You can increase the default number of PTEs-calculated based on the total system memory-up to a maximum value by adding a number (equal to the number of pages to be increased) to the registry at:


   HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\Session Manager
   \Memory Management\SystemPages

You can also monitor currently available "Free System Page Table Entries" using Performance Monitor.

Another important point to note is that the adapter memory pages are not used for page replacement. They remain mapped into the process's address space even if the process is idle or entirely swapped out.

Following are the two most common ways to map adapter memory. The first method maps the memory directly into the process's user-space, and the second method maps into process's system-space and optionally into user- space. Both these methods require the physical memory to be contiguous. Because of the PTE limitation, you might not be able to map a large amount of memory in system-space. However, you might be able to map into the user- space if it's not fragmented or used up.

Section Object Method

There is a sample (MAPMEM) in the Windows NT 4.0 DDK (available through MSDN Professional membership) that shows how to perform this mapping. Here is an outline of the technique:

Get the translated physical address of the adapter memory (HalTranslateBusAddress).

Open a section handle (ZwOpenSection) to the physical memory device \Device\PhysicalMemory.

Reference the object handle (ObReferenceObjectByHandle) to prevent it from being deleted.

Map the memory (ZwMapViewOfSection).

The virtual address you get from the ZwMapViewOfSection is only valid in the context of the process it is mapped. If you want to access the memory in your driver's Deferred Procedure Call (DPC) or Interrupt Service Routine (ISR), which runs in an arbitrary process context, you should also map the memory in the system address space (use the next method). You should unmap (ZwUnMapViewOfSection) the memory in the same process context that you mapped before it exits.

MmMapIoSpace Method

This method shows how to map memory in the process's system and user address space.

Get the translated physical address of the adapter memory (HalTranslateBusAddress).

Map the memory into nonpaged system address space as follows:


      SystemVirtualAddress = MmMapIoSpace(PhysicalAddress, SizeofMemory,
         CacheEnable);

Allocate an Mdl:


      Mdl = IoAllocateMdl(SystemVirtualAddress, SizeOfMemory, FALSE, FALSE,
         NULL);

Build the Mdl to describe the memory pages:
```
      MmBuildMdlForNonPagedPool(Mdl); 
```

Map the memory into the process's user-space using MmMapLockedPages. Because there is an inconsistency in the return value of this function between pre-SP4 and post-SP4 releases of Windows NT, the following statement should be used to make this code compatible with all versions of Windows NT:
```
      UserVirtualAddress = (PVOID)(((ULONG)PAGE_ALIGN(MmMapLockedPages(Mdl,
         Mode))) + MmGetMdlByteOffset(Mdl)); 
```

For additional information, click the article number below to view the article in the Microsoft Knowledge Base:

Q199311 INFO: MmMapLockedPages Returns Actual Virtual Address in SP4

The advantage of this method is that you get SystemVirtualAddress, which can be used in any process context (such as DPCs and ISR), and an UserVirtualAddress that can be used by the user-mode application in whose context it is mapped.

If you map into system address space, you should unmap as shown below:


MmUnmapIoSpace(SystemVirtualAddress, SizeofMemory);

If you map into user address space, you should unmap as shown below only while running in the context of the process in which you mapped the memory:


MmUnmapLockedPages(UserVirtualAddress, Mdl); 

Finally you free the MDL by calling:

IoFreeMdl(Mdl);

REFERENCES

Windows NT 4.0 DDK Documentation

"Inside Windows NT" by Helen Custer

Additional query words: MmMapIoSpace MmMapLockedPages ZwMapViewOfSection NO_MORE_SYSTEM_PTES IoFreeMdl(Mdl);

Keywords : kbKMode kbNTOS400
Version : winnt:4.0
Platform : winnt
Issue type : kbhowto