32-bit physical page table resolution

Question

I'm running a 32 bit system in legacy mode on a 64-bit (x86-64 that is) capable architecture. When a new process is created, the kernel has to decide where in physical memory all of the pages needed at the time of instantiation are to be allocated (assuming a single thread this may include several memory regions such as the stack, the heaps etc).

I'm assuming the kernel keeps some sort of dynamic list of the physical RAM frames that are in use, and also a static list of all the regions of physical memory that have been taken up by devices for systems that use memory-mapped IO. Is this correct?

In addition, I also read that a 32-bit Windows system has a physical memory limit of 4GB (probably due to minimum address bus assumptions) so, even though a system may have more than 4 gigabytes of physical memory installed, a 32 bit kernel will only allocate addresses within the 4GB range.

Specific information regarding low-level operating system implementation for specific cases such as this is quite difficult to find online. Can anyone verify these statements and possibly refer me to a source where I could attain more information?

Thanks for your considerations.

Answer 1

When a new process is created, the kernel has to decide where in physical memory all of the pages needed at the time of instantiation are to be allocated

Why does it have to decide at process creation time? In fact, it only creates them on-demand - it simply creates the PTEs (i.e. "This address range is valid", but the pages are not backed in any way); when the process first starts executing, it immediately page-faults.

What is a page fault though? What happens is, first the CPU reads the TLB to see if it has an address <=> frame mapping. When that fails, it walks the PTEs looking for an entry that matches. If no entry is found, or if the entry indicates that the page isn't backed, a page-fault is generated. This means, that a CPU exception occurs and the CPU immediately jumps to a predefined address. The first thing the kernel then does is save the CPU Context (i.e. the registers at the location of the fault), then dispatches to the page fault handler.

When the page-fault occurs, Mm (the Memory Manager in NT) will read the mapping in its own data structures (remember that all PE images are memory-mapped files) and determine at that time which physical frame (i.e. 'a real piece of memory') which will be used.

Once the page fault is serviced, the page fault restores the saved CPU context, and jumps back to where it was, and retries the instruction that faulted.

You're correct that a 32-bit OS will only use 4GB of address space (not RAM! Don't forget those memory-mapped devices and files!), the processor will operate in 32-bit mode and interpret the PTEs as 32-bit (remember that AMD64 long mode adds an extra level of page tables and extends the address space to 48 bits).

Answer 2

32bit systems can only ever address 4gig directly (2^32 = 4gig). There's PAE hacks, which let the system have more than 4gig of physical ram, but no process can ever have more than 4gig available. As well, even if you have 4gig of ram, you'll never see more than 3.5gig or so actually available - some is reserved for memory mapping hardware devices, such as your video ram.

For one method of dealing with the physical-virtual memory mapping, look at TLB