Hardware support for virtual memory

As covered in the section called “The TLB”, the processor hardware provides a lookup-table that links virtual addresses to physical addresses. Each processor architecture defines different ways to manage the TLB with various advantages and disadvantages.

The part of the processor that deals with virtual memory is generally referred to as the Memory Management Unit or MMU

XXX

The Itanium MMU provides many interesting features for the operating system to work with virtual memory.

the section called “Flushing the TLB” introduced the concept of the address-space ID to reduce the overheads of flushing the TLB when context switching. However, programmers often use threads to allow execution contexts to share an address space. Each thread has the same ASID and hence shares TLB entries, leading to increased performance. However, a single ASID prevents the TLB from enforcing protection; sharing becomes an "all or nothing" approach. To share even a few bytes, threads must forgo all protection from each other (see also the section called “Protection”).

Figure 6.7. Illustration Itanium regions and protection keys
Itanium regions and protection keys. In this example the processes alias region 1. Each process has a private mapping and they share a key for another.

The Itanium MMU considers these problems and provides the ability to share an address space (and hence translation entries) at a much lower granularity whilst still maintaining protection within the hardware. The Itanium divides the 64-bit address space up into 8 regions, as illustrated in Figure 6.7, “Illustration Itanium regions and protection keys”. Each process has eight 24-bit region registers as part of its state, which each hold a region ID (RID) for each of the eight regions of the process address space. TLB translations are tagged with the RID and thus will only match if the process also holds this RID, as illustrated in Figure 6.8, “Illustration of Itanium TLB translation”.

Figure 6.8. Illustration of Itanium TLB translation
Illustration of the Itanium translation process (Mosberger).

Further to this, the top three bits (the region bits) are not considered in virtual address translation. Therefore, if two processes share a RID (i.e., hold the same value in one of their region registers) then they have an aliased view of that region. For example, if process-A holds RID 0x100 in region-register 3 and process-B holds the same RID 0x100 in region-register 5 then process-A, region 3 is aliased to process-B, region 5. This limited sharing means both processes receive the benefits of shared TLB entries without having to grant access to their entire address space.

Protection Keys

To allow for even finer grained sharing, each TLB entry on the Itanium is also tagged with a protection key. Each process has an additional number of protection key registers under operating-system control.

When a series of pages is to be shared (e.g., code for a shared system library), each page is tagged with a unique key and the OS grants any processes allowed to access the pages that key. When a page is referenced the TLB will check the key associated with the translation entry against the keys the process holds in its protection key registers, allowing the access if the key is present or otherwise raising a protection fault to the operating system.

The key can also enforce permissions; for example, one process may have a key which grants write permissions and another may have a read-only key. This allows for sharing of translation entries in a much wider range of situations with granularity right down to a single-page level, leading to large potential improvements in TLB performance.

Itanium Hardware Page-Table Walker

Switching context to the OS when resolving a TLB miss adds significant overhead to the fault processing path. To combat this, Itanium allows the option of using built-in hardware to read the page-table and automatically load virtual-to-physical translations into the TLB. The hardware page-table walker (HPW) avoids the expensive transition to the OS, but requires translations to be in a fixed format suitable for the hardware to understand.

The Itanium HPW is referred to in Intel's documentation as the virtually hashed page-table walker or VHPT walker, for reasons which should become clear. Itanium gives developers the option of two mutually exclusive HPW implementations; one based on a virtual linear page-table and the other based on a hash table.

It should be noted it is possible to operate with no hardware page-table walker; in this case each TLB miss is resolved by the OS and the processor becomes a software-loaded architecture. However, the performance impact of disabling the HPW is so considerable it is very unlikely any benefit could be gained from doing so

Virtual Linear Page-Table

The virtual linear page-table implementation is referred to in documentation as the short format virtually hashed page-table (SF-VHPT). It is the default HPW model used by Linux on Itanium.

The usual solution is a multi-level or hierarchical page-table, where the bits comprising the virtual page number are used as an index into intermediate levels of the page-table (see the section called “Three Level Page Table”). Empty regions of the virtual address space simply do not exist in the hierarchical page-table. Compared to a linear page-table, for the (realistic) case of a tightly-clustered and sparsely-filled address space, relatively little space is wasted in overheads. The major disadvantage is the multiple memory references required for lookup.

Figure 6.9. Illustration of a hierarchical page-table
The hierarchical page-table

With a 64-bit address space, even a 512~GiB linear table identified in the section called “Virtual Address Translation” takes only 0.003% of the 16-exabytes available. Thus a virtual linear page-table (VLPT) can be created in a contiguous area of virtual address space.

Just as for a physically linear page-table, on a TLB miss the hardware uses the virtual page number to offset from the page-table base. If this entry is valid, the translation is read and inserted directly into the TLB. However, with a VLPT the address of the translation entry is itself a virtual address and thus there is the possibility that the virtual page which it resides in is not present in the TLB. In this case a nested fault is raised to the operating system. The software must then correct this fault by mapping the page holding the translation entry into the VLPT.

Figure 6.10. Itanium short-format VHPT implementation
Operation of the Itanium short-format VHPT

This process can be made quite straight forward if the operating system keeps a hierarchical page-table. The leaf page of a hierarchical page-table holds translation entries for a virtually contiguous region of addresses and can thus be mapped by the TLB to create the VLPT as described in Figure 6.10, “Itanium short-format VHPT implementation”.

Figure 6.11. Itanium PTE entry formats
Itanium PTE entry formats

The major advantage of a VLPT occurs when an application makes repeated or contiguous accesses to memory. Consider that for a walk of virtually contiguous memory, the first fault will map a page full of translation entries into the virtual linear page-table. A subsequent access to the next virtual page will require the next translation entry to be loaded into the TLB, which is now available in the VLPT and thus loaded very quickly and without invoking the operating system. Overall, this will be an advantage if the cost of the initial nested fault is amortised over subsequent HPW hits.

The major drawback is that the VLPT now requires TLB entries which causes an increase on TLB pressure. Since each address space requires its own page table the overheads become greater as the system becomes more active. However, any increase in TLB capacity misses should be more than regained in lower refill costs from the efficient hardware walker. Note that a pathological case could skip over page_size ÷ translation_size entries, causing repeated nested faults, but this is a very unlikely access pattern.

The hardware walker expects translation entries in a specific format as illustrated on the left of Figure 6.11, “Itanium PTE entry formats”. The VLPT requires translations in the so-called 8-byte short format. If the operating system is to use its page-table as backing for the VLPT (as in Figure 6.10, “Itanium short-format VHPT implementation”) it must use this translation format. The architecture describes a limited number of bits in this format as ignored and thus available for use by software, but significant modification is not possible.

A linear page-table is premised on the idea of a fixed page size. Multiple page-size support is problematic since it means the translation for a given virtual page is no longer at a constant offset. To combat this, each of the 8-regions of the address space (Figure 6.7, “Illustration Itanium regions and protection keys”) has a separate VLPT which only maps addresses for that region. A default page-size can be given for each region (indeed, with Linux HugeTLB, discussed below, one region is dedicated to larger pages). However, page sizes can not be mixed within a region.

Virtual Hash Table

Using TLB entries in an effort to reduce TLB refill costs, as done with the SF-VHPT, may or may not be an effective trade-off. Itanium also implements a hashed page-table with the potential to lower TLB overheads. In this scheme, the processor hashes a virtual address to find an offset into a contiguous table.

The previously described physically linear page-table can be considered a hash page-table with a perfect hash function which will never produce a collision. However, as explained, this requires an impractical trade-off of huge areas of contiguous physical memory. However, constraining the memory requirements of the page table raises the possibility of collisions when two virtual addresses hash to the same offset. Colliding translations require a chain pointer to build a linked-list of alternative possible entries. To distinguish which entry in the linked-list is the correct one requires a tag derived from the incoming virtual address.

The extra information required for each translation entry gives rise to the moniker long-format~VHPT (LF-VHPT). Translation entries grow to 32-bytes as illustrated on the right hand side of Figure 6.11, “Itanium PTE entry formats”.

The main advantage of this approach is the global hash table can be pinned with a single TLB entry. Since all processes share the table it should scale better than the SF-VHPT, where each process requires increasing numbers of TLB entries for VLPT pages. However, the larger entries are less cache friendly; consider we can fit four 8-byte short-format entries for every 32-byte long-format entry. The very large caches on the Itanium processor may help mitigate this impact, however.

One advantage of the SF-VHPT is that the operating system can keep translations in a hierarchical page-table and, as long as the hardware translation format is maintained, can map leaf pages directly to the VLPT. With the LF-VHPT the OS must either use the hash table as the primary source of translation entries or otherwise keep the hash table as a cache of its own translation information. Keeping the LF-VHPT hash table as a cache is somewhat sub-optimal because of increased overheads on time critical fault paths, however advantages are gained from the table requiring only a single TLB entry.