Some content of this application is unavailable at the moment.
If this situation persist, please contact us atFeedback&Contact
1. (WO2019002817) REALM IDENTIFIERS FOR REALMS FOR MEMORY ACCESS CONTROL
Note: Text based on automatic Optical Character Recognition processes. Please use the PDF version for legal matters

REALM IDENTIFIERS FOR REALMS FOR MEMORY ACCESS CONTROL

The present technique relates to the field of data processing.

It is known to provide memory access control techniques for enforcing access rights for particular memory regions. Typically these are based on privilege level, so that a process executing at a higher privilege level can exclude less privileged processes from accessing memory regions.

At least some examples provide an apparatus comprising:

processing circuitry to perform data processing in response to one or more software processes;

memory access circuitry to enforce ownership rights for a plurality of memory regions, wherein a given memory region is associated with an owner realm specified from among a plurality of realms, each realm corresponding to at least a portion of at least one of the software processes, said owner realm having a right to exclude other realms from accessing data stored within said given memory region; and

realm management circuitry to access a realm management tree storing realm management data for at least two realms in a tree structure comprising a variable number of levels;

wherein the realm management circuitry is configured to index the realm management tree using a realm identifier of a given realm, the realm identifier comprising a variable number of variable length bit portions, wherein the realm management circuitry is configured to index into a given level of the realm management tree using a corresponding one of the variable length bit portions of the realm identifier.

At least some examples provide an apparatus comprising:

means for performing data processing in response to one or more software processes;

means for enforcing ownership rights for a plurality of memory regions, wherein a given memory region is associated with an owner realm specified from among a plurality of realms, each realm corresponding to at least a portion of at least one of the software processes, said owner realm having a right to exclude other realms from accessing data stored within said given memory region; and

means for accessing a realm management tree storing realm management data for at least two realms in a tree structure comprising a variable number of levels;

wherein the means for accessing is configured to index the realm management tree using a realm identifier of a given realm, the realm identifier comprising a variable number of variable length bit portions, wherein the means for accessing is configured to index into a given level of the realm management tree using a corresponding one of the variable length bit portions of the realm identifier.

At least some examples provide a data processing method comprising: performing data processing in response to one or more software processes;

enforcing ownership rights for a plurality of memory regions, wherein a given memory region is associated with an owner realm specified from among a plurality of realms, each realm corresponding to at least a portion of at least one of the software processes, said owner realm having a right to exclude other realms from accessing data stored within said given memory region; and

accessing a realm management tree storing realm management data for at least two realms in a tree structure comprising a variable number of levels;

wherein the realm management tree is indexed using a realm identifier of a given realm, the realm identifier comprising a variable number of variable length bit portions, wherein a given level of the realm management tree is indexed using a corresponding one of the variable length bit portions of the realm identifier.

At least some examples provide a computer program for controlling a host data processing apparatus to provide an instruction execution environment comprising:

processing program logic to perform data processing in response to one or more software processes;

memory access program logic to enforce ownership rights for a plurality of memory regions, wherein a given memory region is associated with an owner realm specified from among a plurality of realms, each realm corresponding to at least a portion of at least one of the software processes, said owner realm having a right to exclude other realms from accessing data stored within said given memory region; and

realm management program logic to access a realm management tree storing realm management data for at least two realms in a tree structure comprising a variable number of levels;

wherein the realm management program logic is configured to index the realm management tree using a realm identifier of a given realm, the realm identifier comprising a variable number of variable length bit portions, wherein the realm management program logic is configured to index into a given level of the realm management tree using a corresponding one of the variable length bit portions of the realm identifier.

The computer program may be stored on a storage medium. The storage medium may be a non-transitory storage medium.

Further aspects, features and advantages of the present technique will be apparent from the following description of examples, which is to be read in conjunction with the accompanying drawings, in which:

Figure 1 schematically illustrates a data processing system including a plurality of processing elements utilising memory regions stored within a first memory and a second memory;

Figure 2 schematically illustrates a relationship between a plurality of processes executed, privilege levels associated with those processes, and realms associated with those processes for controlling which process owns a given memory region and accordingly has exclusive rights to control access to that given memory region;

Figure 3 schematically illustrates memory regions under management by a realm management unit and a memory management unit;

Figure 4 schematically illustrates a sequence of program instructions executed to export a given memory region from a first memory to a second memory;

Figure 5 is a flow diagram schematically illustrating page export;

Figure 6 schematically illustrates a plurality of realms and their relationship within a control hierarchy to control which export commands can interrupt which other export commands;

Figure 7 is a flow diagram schematically illustrating page import;

Figure 8 schematically illustrates a first export command source and a second export command source performing overlapping export operations for a given memory region;

Figure 9 illustrates a more detailed example of a processing element and realm management control data stored in memory;

Figure 10 shows an example of a hierarchy of realms in which a parent realm can define realm descriptors describing properties of various child realms;

Figures 1 1 and 12 illustrate two different examples of a realm hierarchy;

Figure 13 shows an example of a realm descriptor tree maintained by a parent realm to record the realm descriptors of its child realms;

Figure 14 shows an example of a local realm identifier constructed from a number of variable length bit portions which each provide an index to a corresponding level of the realm descriptor tree;

Figure 15 shows an example of local and global realm identifiers for each realm in a realm hierarchy;

Figure 16 shows an example of contents of a realm descriptor;

Figure 17 is a table illustrating different realm lifecycle states;

Figure 18 is a state machine diagram indicating changes of lifecycle states of a realm;

Figure 19 is a table showing contents of an entry in an ownership table for a given memory region;

Figure 20 is a table showing visibility attributes which can be set for a given memory region to control which realms other than the owner are allowed to access the region;

Figure 21 shows an example of different lifecycle states for a memory region, including states corresponding to RMU-private memory regions which are reserved for exclusive access by the realm management unit;

Figure 22 is a state machine showing transitions of lifecycle state for a given memory region;

Figure 23 illustrates how ownership of a given memory region can pass between a parent realm and its child realm;

Figure 24 schematically illustrates memory access control provided based on page tables which define memory control attributes which depend on privilege level and realm management unit levels which provide an orthogonal level of control over memory access based on the permissions set by an owner realm;

Figure 25 illustrates an example of a translation lookaside buffer;

Figure 26 is a flow diagram illustrating a method of controlling access to memory based on the page tables and RMU tables;

Figure 27 illustrates state accessible to a process executing at different exception levels;

Figure 28 is a flow diagram illustrating a method of entering a realm or returning from an exception;

Figure 29 is a flow diagram illustrating a method of exiting a realm or taking an exception;

Figure 30 illustrates an example of entry to a child realm and exit back to the parent realm;

Figure 31 shows an example of nested realm exit and nested realm entry;

Figure 32 shows an example of using lazy saving of realm execution context on exit from a realm;

Figure 33 shows an example of use of a flush command to ensure that a subset of state associated with a previously exited child realm is saved to memory before a different child realm is entered;

Figure 34 illustrates the use of sub-realms which correspond to a specific address range within a process associated with the parent realm of the sub-realm; and

Figure 35 illustrates a simulator example that may be used.

Figure 1 schematically illustrates a data processing system 2 comprising a system-on-chip integrated circuit 4 connected to a separate non-volatile memory 6, such as an off-chip flash memory serving as a mass storage device. The system-on-chip integrated circuit 4 comprises a plurality of processing elements in the form of (in this example embodiment) two general purpose processors (CPUs) 8, 10, and a graphics processing unit (GPU) 12. It will be appreciated that in practice, many different forms of processing element may be provided, such as additional general purposes processors, graphics processing units, direct memory access (DMA) units, coprocessors and other processing elements which serve to access memory regions within a memory address space and perform data processing operations upon data stored within those memory regions.

The general purpose processors 8, 10 and the graphics processing unit 12 are coupled to interconnect circuitry 14 via which they perform memory transactions with an on-chip memory 16 and the external memory 6 (via an external memory interface 18). Although the memory 16 is on-chip in Figure 1 , in other embodiments the memory 16 could instead be implemented as an off-chip memory. The on-chip memory 16 stores data corresponding to a plurality of memory regions within an overall memory address space. These memory regions correspond to memory pages and are subject to management operations that control which memory regions (pages) are present within the on-chip memory 16 at a given time, which processes have access to the data stored within those memory regions and other parameters associated with those memory regions. More particularly, in this example embodiment, each of the processing elements 8, 10, 12 includes a realm management unit 20, 22, 24 and a general purpose memory management unit 26, 28, 30. The general purpose memory management units 26, 28, 30 serve to control aspects of the operation of the memory regions such as address mapping (e.g. mapping between virtual addresses and intermediate physical addresses, or physical addresses), privilege level constraints upon processes able to access given memory regions, storage characteristics of data within given memory regions (e.g. cacheability, device memory status, etc.) and other characteristics of regions of memory.

The realm management units 20, 22, 24 manage data serving to enforce ownership rights of the plurality of memory regions whereby a given memory region has a given owning process (or owner "realm") specified from among a plurality of processes (a process or realm being, for example, one of a monitor program, a hypervisor program, a guest operating system program, an application program, or the like, or a specific sub-portion of such a program). The given owning process (owner realm) for a given memory region has the exclusive rights to control access to the given own data stored within that given memory region. In particular, the owner process has the right to prevent access to its owned memory region by processes executed at a greater privilege level than the owner process.

Hence, a plurality of memory regions is divided amongst a plurality of owner realms. Each realm corresponds to at least a portion of at least one software process, and is allocated ownership of a number of memory regions. The owning process/realm has the exclusive rights to control access to the data stored within the memory regions of their realm.

Management and control of which memory regions are memory mapped to each realm is performed by a process other than the owner realm itself. Using this arrangement it is possible for a process, such as a hypervisor, to control which memory regions (pages of memory) are contained within realms owned by respective guest virtual machines (guest operating systems) managed by that hypervisor and yet the hypervisor itself may not have the rights to actually access the data stored within the memory regions which it has allocated to a given realm. Thus, for example, a guest operating system may keep private from its managing hypervisor the data stored within the realm of that guest operating system, i.e. within the memory regions which are owned by that guest operating system.

The division of the memory address space into realms, and the control of ownership of those realms, is managed via the realm management units 20, 22, 24 associated with each of the processing elements 8, 10, 12 and is a control process which is orthogonal to the more conventional forms of control provided by the general purpose memory management units 26, 28, 30. The realm management units 20, 22, 24 accordingly provide memory access circuitry to enforce ownership rights of the memory regions of the memory address space. In some cases the memory access circuitry which enforces realm ownership rights may also include part of the MMU 26, 28, 30 (e.g. a TLB in the MMU 26, 28, 30 may include some control data for controlling accesses based on the realm controls provided by the RMU 20, 22, 24, to avoid needing to access two separate structures). In this example embodiment each of the processing elements 8, 10, 12 contains its own realm management unit 20, 22, 24; this is advantageous for performance purposes. However, more generally the memory access circuitry which enforces ownership rights may comprise a single instance of a realm management unit, the combination of all the realm management units 20, 22, 24 present, or a subset of those realm management units 20, 22, 24 present. Thus, the memory access circuitry for enforcing ownership rights may be distributed across the system-on-chip integrated circuit 4 in association with different processing elements 8, 10, 12, or collected together in a one location or in some other arrangement.

The processing elements comprising the general purpose processors 8, 10 are illustrated as including respective decoding and execution circuitry 32, 34 which decode and execute program instructions. These program instructions include commands which serve to control the management of memory regions within different ownership realms of the memory address space (realm management commands or RMU Commands). As an example, the program instructions executed may include program instructions designated as realm management unit commands and which are directed to the associated realm management unit 20, 22, 24 when they are encountered within the program instruction stream in order that they can be executed (actioned) by the relevant realm management unit 20, 22, 24. Examples of realm management unit commands include commands to initialize

new realms or invalidate existing realms, commands to allocate memory regions to particular realms, remove memory regions from particular realms, export the data contained within a memory region from a first memory 16 to a second memory 6 with encryption and other processes being performed upon the exported data such that it is protected within the second memory 6. Further realm management unit commands are provided to import data back from a second memory 6 to a first memory 16 with associated decryption and validation operations performed upon the imported data.

In the context of such exports and imports of data from memory regions, it will be appreciated that a first memory, such as the on-chip memory 16, is closely managed by the realm management units 20, 22, 24 within the system-on-chip integrated circuit 4 and accordingly those realm management units 20, 22, 24 are able to enforce ownership rights and restrict access to the data within a given memory region to the process owning that memory region, or those processes which the owning process has granted access. However, when the data within that memory region is exported, such as to the external non-volatile memory 6, which is a second memory, then the control over access provided by the realm management units 20, 22, 24 is no longer effective and accordingly the data requires protection in some other way. This is achieved by encrypting the data within a memory region using before it is exported, and then decrypting that data with a secret key when it is imported back to the on-chip memory 16.

The export process may be accompanied by the generation of metadata specifying characteristics of the exported data. Such metadata may be separately stored within a metadata memory region of the first memory (on-chip memory 16) where it is held private to realm management units 20, 22, 24 (i.e. only accessible to such realm management units 20, 22, 24 and not to any of the existing processes) such that when the data is imported back to the on-chip memory 16, the metadata can be read for that imported data and the characteristics of the data represented in the metadata can be checked against the characteristics of the imported data to ensure the integrity of that imported data (e.g. checksums, data size, signatures etc.). It may be that private data of the realm management units 20, 22, 24 (including the above metadata characterising exported regions/pages) needs to be exported from the on-chip memory 16 to the off-chip non-volatile memory 6 (e.g. to make space within the on-chip memory 16) and in this circumstance the RMU-private metadata itself can be encrypted for its protection and new metadata characterising the exported metadata can be retained within the on-chip memory 16 (such retained metadata being significantly smaller in size than the exported metadata) in order that the encrypted and exported metadata can be checked and validated when it is imported back to the on-chip memory 16 for use.

Such metadata describing characteristics of memory regions and the data stored within the memory regions may be arranged as part of a hierarchical structure, such as a metadata memory region tree having a branching pattern. The form of such a metadata memory region tree may be determined under software control as different regions of the memory address space are registered to serve as metadata regions owned by the realm management units 20, 22, 24. It will be appreciated that whilst the software which controls the registering of such memory regions is able to allocate, de-allocate and control the relationships between memory regions serving to store metadata, such software does not itself own the data contained within those memory regions in the sense of being able to control which processes have access to such data. In the case of memory regions which are private to the realm management units 20, 22, 24 (i.e. the memory management circuitry), such access rights may be restricted to only the realm management units 20, 22, 24 themselves and the such RMU-Private data will not be shared with any other process(es).

When the given data stored within a given memory region is exported, then the memory region concerned is made invalid so that the contents are inaccessible. To reuse this page the page is made "Valid", by using a Clean command which overwrites the memory region with other data uncorrelated to the previous content in order that such previous content is not made accessible to another process when that given memory region is released for use by another process. For example, the content of a given memory region may be all written to zero values, or written to a fixed value, or written to random values thereby overwriting the original content of the memory region. In other examples, the overwriting of the contents of the exported memory region could be triggered by an export command itself, rather than a subsequent clean command. Either way, the given owned data being exported may be overwritten with values uncorrelated with the given owned data before the given memory region is made accessible to a process other than the given owning process. When a given memory region owned by a given process is to be exported, as part of the export process, the realm management unit 20, 22, 24 which is executing the realm command to perform the export takes ownership of the memory region concerned from the given process (i.e. makes the region RMU-Private), locks access of that memory region against all other processes (and other realm management units), performs the export operation (including encryption, metadata generation and overwriting), and then unlocks the access to that memory region and releases ownership of that memory region. Thus, a memory region which is in the process of being exported, or imported, may be held private to the realm management unit concerned whilst that command is being performed.

Figure 2 schematically illustrates the relationship between a plurality of processes

(programs/threads), a plurality of exception levels (privilege levels), a secure and a nonsecure processor domain, and a plurality of realms representing ownership of given memory regions. As illustrated, a hierarchy of privilege levels extends from exception level ELO to exception level EL3 (with exception level EL3 having the highest level of privilege). The operating state of the system may be divided between a secure operating state and a nonsecure operating state as represented by the secure domain and the non-secure domain in, for example, processors using the TrustZone® architecture provided by ARM® Limited of Cambridge, UK.

As illustrated in Figure 2, the memory access circuitry (realm management units 20, 22, 24 and associated controlling software (e.g. millicode running one the realm management units)) manage a plurality of realms within the execution environment. A given memory region (memory page) is owned by a particular realm. A realm may have child realms within it, and grandchild realms within those child realms (e.g. see realm A (parent), realm B (child), and realm C (grandchild)). Memory regions for which ownership is given to realm A may have their ownership in turn passed from realm A to realm B under control of processes owned by realm A. Thus, a parent realm is able to give ownership of regions to its own child realm(s). Those child realms in turn may pass ownership of memory regions which they have received from their parent realm to then be owned by their own child realms (e.g. realm C) which is a grandchild realm of the original realm, namely realm A. The processes within a given realm may execute at the same privilege level or at different privilege level. The realm to which a process belongs is accordingly an orthogonal parameter to the privilege level of a process, although in many practical cases the realm and the privilege level may correspond as a convenient mechanism for moving between realms may involve the use of exceptions which themselves move the system between different privilege levels (exception levels).

The relationship between the realms illustrated in Figure 2 shows the child/parent relationship between different realms and this may be used to give rise to a control hierarchy for controlling operation of the system when multiple different sources of commands for memory region management compete with each other. Thus, for example, in the case of export commands for exporting memory regions as discussed above, a first export command may be received by a given realm management unit (memory access circuitry) from a first export command source, such as the operating system kernel 36 within realm B. A second export command may then be received by the given realm management unit from a second command source, such as the hypervisor program 38 executing in realm A. In this example, the hypervisor program 38, which is the second export command source, has a higher priority within the control hierarchy established by the relationship between parent and child realms such that the second export command issued by the hypervisor program 38 interrupts processing of the first export command issued by the operating system kernel 36.

When the second export command, as issued by the hypervisor 38, has completed, the first export command, as issued by the operating system kernel 36, may be resumed.

In this example the second export command has a higher priority and so interrupts operation of the first export command. However, if the second export command had, for example, originated from the application program 40 within realm C, then this has a lower priority position within the control hierarchy established by the relationship between realms and accordingly such a second export command from the application program 40 would not interrupt the operation of the first export command from the operating system kernel 36 and would rather itself be blocked from being performed until the first export command had completed. Thus, paging operations (export and import operations) may be protected from one another in the sense they may or may not interrupt one another in dependence upon a control hierarchy, which may be associated with the realm hierarchy. In other example embodiments the control hierarchy may correspond to the privilege level.

Figure 3 schematically illustrates a realm management unit 20 and a general purpose memory management unit 26 which respectively perform different management operations upon a plurality of memory pages (memory regions) stored within the on-chip memory 16. As illustrated, the realm management unit 24 uses a plurality of realm descriptors 42 with each descriptor specifying properties of a realm. The realm management unit 24 may also maintain a realm granule table (or ownership table) comprising entries indexed by physical address, each entry including information for a corresponding memory region, including an indication of which realm that memory region belongs to, i.e. which realm has the exclusive rights to control access to control data within that memory region even if it does not control whether or not it itself actually owns that memory region. The realm descriptors and realm granule table entries may be stored in memory 16, but could also be cached in the RMU itself. Thus, as illustrated in Figure 3, the different memory regions have different owning realms as indicated by the realm designations RA, RB, RC, RD and RE. Some of the memory regions are also owned by (private to) the realm management unit 20 and are marked as RMU-Private. Such RMU- Private regions may be used to store metadata describing characteristics of other memory regions, temporarily store memory regions being exported or imported, or for other purposes of the realm management unit 20 itself. The RMU-private regions may still be owned by a corresponding owner realm, but may not be accessible to general purpose read/write accesses issued by the owner realm (instead RMU commands issued to the RMU 20 may be required to trigger the RMU 20 to make any changes to RMU-private regions).

The addressing of memory regions may be by virtual, intermediate physical or physical addresses depending upon the particular system concerned. The realm management unit 20, and the general purpose memory management unit 26, may

accordingly store translation data enabling received addresses (whether they be virtual memory addresses or intermediate memory addresses), to be translated to an address, such as a physical address, more directly representing the memory region within the on-chip memory 16 concerned. Such address translation data may be managed and distributed within the system on-chip integrated circuit 4 using translation look aside buffers and other distributed control mechanisms.

Figure 4 schematically illustrates program instructions associated with an export operation of a memory region. These program instructions appear within a program instruction stream and may be executed (actioned) by different elements within the overall circuitry. For example, realm management unit commands are executed by respective realm management units 12, 22, 24. Instructions such as virtual address unmapping instructions (VUMAP) and translation look aside buffer invalidate instructions (TLBI) are broadcast within the system-on-chip integrated circuit 4 and serve to purge the use of translation data as specified by those commands from locations within the system as a whole (although in some examples, a dedicated virtual address unmapping instruction may not be provided, and instead unmapping of a virtual address can be performed by modifying a translation table entry by performing a store to memory, rather than using a special unmapping instruction). A barrier instruction DSB is inserted within the instruction sequence illustrated in Figure 4 and serves to halt processing of that sequence until an acknowledgement has been received that the preceding virtual address unmapping instruction (or equivalent store instruction) and translation look aside buffer invalidate instruction have been completed by all portions of the system. Thus, the purge of the virtual addresses translations for a given memory region within the system other than in realm management systems themselves may be achieved by the sequence of the virtual address unmapping instruction (or equivalent store instruction), the translation look aside buffer invalidate instruction and a corresponding barrier instruction. By unmapping (and thus effectively removing) the virtual address translation data for a given memory region (page), it can be ensured that such a memory region will not be in use elsewhere within the system when an export operation of the data stored in that memory region is to be performed.

Once the barrier instruction DSB has received an acknowledgement confirming that the purging of the virtual address translation data from within the system has completed, then the export command for the realm management unit is executed by the realm management unit. The executing of such an export instruction received from a given process by the realm management unit triggers performance of a command sequence (corresponding to millicode embedded within the realm management units) comprising a plurality of command actions in respect of the specified given memory region. These command aims may include, for example as illustrated in Figure 4, the steps of gathering address translation data, locking the memory region, encrypting the data, storing the data externally, writing the metadata associated with the memory region and then unlocking the memory region.

The address translation gathering step performed as part of the command sequence by the realm management unit gathers to that realm management unit access control data required to complete the access operation concerned. This ensures that once the export operation is underway then there is a reduced likelihood of that export operation being halted, such as may be due to the non availability of parameters or data required to complete that export operation, e.g. address translation data, attribute data or other data required by the export process. As an example of the retrieving to and storing within the memory access circuitry (realm management unit) of access control data, the address translation step serves to retrieve all required address translation data (e.g. virtual to intermediate physical address (or physical address) mapping data) which may be required to complete the export operation.

Once the address translation data has been retrieved, then the realm management unit serves to set a lock flag associated with the region concerned into a locked state. This lock flag may be stored within the region attribute data 42 for the region concerned. Alternatively, the lock flag may be stored within a memory region private to the realm management unit which is performing the export operation such that it cannot be overwritten by any other process or realm management unit. In order to set the lock flag to the lock state, the realm management unit must determine that no other realm management unit is currently holding the memory region concerned in a locked state itself. Thus, a polling of the locked flag values of any region controlling data stored elsewhere is performed and the lock flag is set to a lock state if a result is returned indicating that the region is not locked elsewhere. If the region is locked elsewhere, then the export operation fails and an error is reported to the process which instructed that export operation. Once the lock has been obtained, then the data within the given memory region is encrypted and stored externally of the system-on-chip integrated circuit, such as to the external non-volatile memory 6. As previously discussed, metadata characterising the encrypted data (or the given data before encryption) is then generated and stored within a realm management unit private region such that it can be used to validate the exported data at a later time. Finally the memory region concerned is unlocked by the realm management unit performing the export command by switching the locked flag from a locked state to an unlocked state. The use of a lock enforced by the hardware mechanism of the memory access circuitry (realm management unit) serves to block the progress of any other (second) access command from a further processing element which might be received when the locked flag is in the locked state.

Figure 5 is a flow diagram schematically illustrating page (memory region) export. At step 44 program instructions are executed (VUMAP, TLBI, DSB) which serve to purge use of the page elsewhere in the system other than within the region management units 20, 22, 24. This may be achieved by invalidating and purging translation data pointing to the region which is to be exported. Once this translation data is purged, should another process or processing element wish to access that region, then it will attempt to refetch the translation data. Upon the attempted refetching of the translation data, the process or processing elements seeking to reuse the region will fail to obtain the relevant translation data as the region concerned will have been placed into an RMU private state in which only the region management unit 20, 22, 24 seeking to perform the page export has rights to access that data.

When the purge requests have been issued at step 44, processing waits at step 46 until responses are received from those purge requests which indicate that the address data has been invalidated elsewhere (other than in realm management units) at which point it is safe to continue beyond the barrier instruction DSB within the program sequence (the barrier instruction DSB halts the processing element 8, 10, 12 until the responses have been received). At step 48 a realm management unit export initialisation instruction is executed. This export initialization instruction includes a pointer to a memory region established as RMU private in which a command context buffer (CCB) is established to store context data representing a current partially completed state of the command sequence corresponding to the export operation should that command sequence be interrupted. In alternative example embodiments the realm management unit itself may be responsible for generating a pointer to the command context buffer (CCB). Step 50 determines whether the command context buffer indicated by the pointer within the export command step 48 is empty. If the command context buffer is empty, then step 52 sets this up as an RMU- Private region. If the command context buffer at step 50 is not empty, then this indicates that the export initialisation command being executed at step 48 is attempting to restart a previously interrupted export operation. In this case, processing proceeds to step 54 at which the contents of the command context buffer pointed to by the pointer is validated using associated metadata which was stored when that partially completed state data was stored to the CCB. If the validation is passed, then step 56 serves to use the content of the command context buffer to restore a partially completed state of the export command, e.g. any partially encrypted data, a pointer to the position within the original data to which encryption had progressed, further attributes of the partially completed command etc. After the initialisation of the export operation instructed by the command at step 48, the processing proceeds either via a path including step 52 or a path including steps 54 and 56 to reach step 58 at which the command to execute the realm management unit export command is reached. When this

command is reached, then the region management unit 20, 22, 24 performs encryption upon a portion of the data within the memory region and stores this within the destination (which is also a pointer specified within the RMU export initialisation instruction at step 48). Step 60 determines whether an interrupt from a command source with a higher priority than the command source which issued the instructions being executed at step 48 and 58 has been received. Such a higher priority command would come from a source with a higher priority position within the control hierarchy (e.g. realm hierarchy, priority hierarchy, etc.) as previously discussed. If such a higher priority interrupt is received, then processing proceeds to step 62 at which the export instruction is halted and an error code is returned to the command source which issued the instructions executed at steps 48 and 58. Step 64 serves to save the partially completed state of the command to the command context buffer. Step 66 stores the command context buffer metadata to an RMU-Private memory region for use in validating the partially completed state stored within the command context buffer when this is subsequently retrieved. Step 68 serves to mark the memory region which has been subject to the partially completed export command as in a state of "partially exported" and indicating the process which performs such a partial export. This assists in restarting that export at a later time.

If the determination at step 60 is that there is no interrupt, then processing proceeds to step 70 at which a determination is made as to whether the export of the memory region has been completed. If the export has not been completed, then processing returns to step 58. If export has been completed, then processing proceeds to step 72 at which the memory region which has been emptied (had its stored data exported therefrom) is overwritten with data uncorrelated with the originally stored data (e.g. zeroed, set to some other fixed number, filled with random data etc.). The processing then terminates.

In the above discussed example embodiment the CCB is provided as a separate private memory region specified by an associated pointer, for example, within an initialisation instruction. However, in other example embodiments the CCB may be provided not as a separate memory region, but as part of a memory region already used by the command that may be interrupted, for example a destination memory region into which result data generated by a command is stored. In the case of an export command which may be interrupted, the exported encrypted data is stored within a destination memory region that is an RMU private memory region while the export is being performed. The CCB may be provided, for example, as the end portion of such a destination region while it is being filled with encrypted data. The integrity of the context data stored within the CCB is ensured by the destination region being RMU private while the export operation is being performed.

In another example embodiment the CCB may be provided as a portion of the realm descriptor (RD); in this case the storage space available for the context data may be

constrained by the space available in the RD and so the number of interruptible parallel commands that are supported may be constrained by the storage space available with the RD to serve as respective CCBs. The CCB can be provided separately or as part of a memory region or resource also used for another purpose.

Figure 6 schematically illustrates a relationship between realms and the control hierarchy that determines which commands from differing command sources are permitted to interrupt/block partially completed commands from other sources. The example illustrated includes three levels of nested realms. A parent realm M corresponds to exception level EL3. A child realm N corresponds to exception level EL2. Two grandchild realms within realm N comprise realm O and realm P and are both at exception level EL1 . In this example, both the exception level priority and the relative position within the nested hierarchy of realms, give an ordering in the control hierarchy in which realm M has a higher priority that realm N, and realm N has a higher priority than both realm O and realm P. Realm O and realm P are of an equal priority.

Figure 7 is a flow diagram schematically illustrating a page (memory region) import operation subsequent to an RMU import command. Step 74 serves to obtain and clean an empty page (memory region) into which the data can be imported. Step 76 then verifies the encrypted data which is to be imported using its associated stored metadata (stored in an RMU-Private region). If this verification is not successful, then an error is generated. Subsequent to successful verification, step 78 serves to decrypt the encrypted data and step 80 serves to store that decrypted data into the memory page which his obtained at step 74. Once the memory page has been filled with the decrypted data, it can be released to the owning realm (process). The page which was obtained and then filled is locked so as to be exclusively available to the memory management circuitry (realm management unit 20, 22, 24) during the page importation process.

Figure 8 schematically illustrates two export commands which may arise in parallel from different command sources. One of the sequences of instructions originates from a process corresponding to a virtual machine (e.g. a guest operating system). The other command source is a hypervisor at a higher level of privilege (or potentially higher level within the realm hierarchy) compared with the virtual machine. Accordingly, the export command from the hypervisor is able to interrupt a partially completed export command being performed by the realm management unit 20, 22, 24, on behalf of the virtual machine. When the export on behalf of the hypervisor is completed, then the export on behalf of the virtual machine may be restarted.

In this example, the command to the realm management unit 20, 22, 24 may be a combined initialisation and execute command which is executed repeatedly until the memory access circuitry reports that the command sequence corresponding to the export operation has been completed. In the previously discussed example the export command may be formed of an export initialisation command which specifies a pointer to the command context buffer and other pointers followed by a subsequent export execute command which is repeatedly executed until the memory access circuitry reports that the command sequence has completed. In other example embodiments the export operation may be controlled by a combined initialization and execute command (which is interruptible) and an execute continue command which is issued if the combined initialization and execute command is interrupted.

The command context buffer is used to store partially completed state representing the partially completed command sequence so that this data can be restored at a later time. In this way, the system does not need to wait until a full export operation has been completed before an interrupt can be serviced. Furthermore, as the partially completed state is retained, forward progress through the export operation is ensured even if it is repeatedly interrupted as the export operation will not need to be restarted from its initial point.

Figure 9 illustrates a more detailed example of one of the processing elements 8, 10,

12 of Figure 1 and of the control data stored in the memory 16 for controlling memory accesses. For ease of explanation, Figure 9 shows the CPU 0 as the processing element 8, but it will be appreciated that the processing element could also be CPU 1 10 of the GPU 12 or any other processing elements within a data processing apparatus 2. As shown in Figure 9, the processing element 8 includes processing circuitry 32 (which may comprise the decode and execute logic described above), a memory management unit 26, which may include one or more translation lookaside buffers 100 for caching entries of the translation tables (which may also be appended with realm-based control data from the RMU 20 if shared MMU-RMU TLB structures are used), and a table walk unit 102 for controlling allocation of data to the TLBs 100 and triggering walk accesses to memory to locate the required data used to control whether a given memory access is allowed to be performed. The processing element 8 may also include a cryptographic unit 104 which may perform cryptographic operations for encrypting or decrypting data, for example for use in the paging (export/import) operations discussed above. The processing element 8 also includes a number of caches 1 10 which may cache data or instructions read from the memory 16. If accesses to memory triggered by the processing circuitry 32 or by the table walk unit 102 miss in the caches, the data can be located from main memory 16.

The processing element 8 also includes a realm management unit 20 as discussed above. In some embodiments the realm management unit (RMU) 20 may be provided as a hardware circuit. However, some of the RMU operations discussed below may be relatively complex to implement purely in hardware, for example if they require multiple accesses to different memory regions to be performed. Therefore, in some examples the RMU 20 may

be implemented using program code which may be stored within the data processing apparatus 2 and may be executed using the general purpose processing circuitry 32. Unlike general purpose software which may be written to memory 16 and may be rewritable, the RMU software (millicode) could be installed in a relatively permanent manner on the data processing apparatus so that it cannot be removed, and may be regarded as part of the platform provided by the processing system. For example the RMU program code could be stored within a read only memory (ROM). Hence, the RMU may comprise a hardware unit, or may comprise the processing circuitry 32 executing realm management software, which is triggered to execute by RMU commands included in the general purpose software executed by the processing circuitry 32. In some examples, the RMU 20 may be implemented using a combination of hardware and software, e.g. some simpler functionality may be implemented using hardware circuits for faster processing, but more complex functions may be implemented using the millicode. Hence, it will be appreciated that subsequent references to the RMU may refer to either hardware or software or a combination of both.

As shown in Figure 9, the memory 16 may store a number of pieces of control information used by the MMU 26 and RMU 20 for controlling access to memory. These include translation tables (also known as page tables) 120, which define memory access attributes for controlling which processes are allowed to access a given memory region, as well as address mapping information for translating virtual addresses to physical addresses. The translation tables 120 may be defined based on the exception levels discussed above with respect to Figure 2, so that a process executing at a more privileged exception level may set permissions which govern whether processes executing at less privileged exception levels are allowed to access the corresponding memory regions.

Also, a number of realm management tables or realm control information 122 are provided for controlling memory access in an orthogonal manner to the MMU page tables 120, to allow a less privileged process to control whether a more privileged process is accessed (the realm control is orthogonal to the MMU control in the sense that, for a memory access request to be serviced, it may need to pass both types of access control checking). With the realm management table, an owner process (realm) which owns a given memory region has the right to exclude processes executing at a more privileged exception level from accessing that memory region. The realm management data includes realm descriptors 124 which describe properties of a given realm. Each realm corresponds to at least a portion of at least one software process executed by the processing circuitry 32. Some realms may correspond to two or more processes, while other realms may correspond to only a sub portion of a given software process. A realm can also be viewed as mapping to a given region of the memory address space (with the processing circuitry 32 executing within a given realm when it is executing program instructions which lie within the

corresponding region of the memory address space). Hence, a realm can be seen either as a set of software processes or a portion of a software process, or as an area of the memory address space. These two views are equivalent. For ease of explanation the subsequent description will refer to a realm as at least a portion of at least one software process, but the corresponding view of a realm as a collection of memory regions is equally valid (in this case, "entry" and "exit" to/from a realm may corresponding to program execution reaching/leaving the part of the memory address corresponding to the realm).

The realm management data 122 also includes realm execution context regions 126 which can be used for saving and restoring architectural state associated with a given realm upon realm exit or entry. The realm management data also includes a realm granule table (or ownership table) 128 which defines, for each region of the memory address space which realm is the owner realm for that memory region. The owner realm for a given memory region has the right to exclude other realms (including more privileged processes) from accessing data stored within that memory region. Use of this realm management data is discussed in more detail below. In general, the realm management unit 20 and MMU 26 can be seen as memory access circuitry which enforces the ownership rights defined by an owner realm for the memory regions owned by that realm. This can be particularly useful, for example, for a cloud platform in which a number of virtual machines 36 provided by different parties may be executing under control of a hypervisor 38 provided by the cloud server operator. A party providing one of the virtual machines may not wish their data and code to be accessible to the hypervisor. By introducing the concept of realms where a realm executing at a less privileged exception level can exclude a more privileged exception level from accessing its data or instructions, this enables a blind hypervisor to be provided which can increase the confidence of code developers to install their software on a cloud service where the physical hardware may be shared with code provided by other parties.

As shown in Figure 10, the realms are managed by the RMU 20 according to a realm hierarchy in which each realm other than a root realm 130 is a child realm which has a corresponding parent realm which initialized the child realm by executing an initialization command. The root realm 130 may for example be a realm associated with the monitor code or system firmware executing at the most privileged exception level EL3. For ease of explanation, the example of Figure 10 and the initial examples discussed afterwards show cases where each child realm executes at a lower privileged level than its parent realm. However, as will be discussed below it is also possible to establish a sub-realm which executes at the same exception level as its parent.

In general, for the realm management part of the memory access control provided by the MMU 26, a child realm has default access to any memory regions owned by its parent realm. Similarly, any descendants of a given realm are assumed to have access to the

given realm's owned memory regions. However, as the realm management control is orthogonal to the control provided by the translation tables 120 based on exception level, a process executing at a higher privilege level can still exclude less privileged code from accessing its data, by setting parameters of the translation tables 120 accordingly. Hence, in general a given child realm has a right to exclude its parent realm from accessing data stored in a given memory region owned by the given child realm. Whether the child realm actually excludes the parent realm from accessing a given memory region may be set based on control attributes set in the ownership table 128 (the default may be that the parent realm has no access to the child realm's owned regions, but the child realm may choose to grant access to the parent realm by setting the visibility attributes accordingly). When there are multiple sibling realms (different child realms sharing the same parent realm) then a given child realm may exclude a sibling realm from accessing the data stored in a given memory region owned by the given child realm. Again, visibility attributes set in the ownership table 128 may control the extent to which sibling realms can access each other's data. Alternatively, access by sibling realms may be controlled based on the parent visibility attribute, so that if a child real makes a page visible to its parent realm, the same page also becomes visible to its sibling realms and further descendants of the sibling realms. In some cases the ownership table 128 may have a global visibility attribute which may allow a given owner process to enable any process executing under any realm to access the data within its owned memory region.

As shown in Figure 10, each realm 140 is associated with one or more realm execution context (REC) memory regions 126 which can be used for storing architectural state of the realm, such as register values, when exiting from a given realm. The number of RECs 126 provided for a given realm may depend on how many threads of execution are operating under a given realm. For example a realm when first initialized may be established with a single primary REC region 126, but then the realm may configure other memory regions owned by that realm to act as further RECs as necessary. The REC memory regions are owned by the corresponding realm whose execution state is stored to that REC.

Each realm is associated with a realm descriptor 124 which is stored in a memory region which is owned by the parent realm of the realm whose properties are described in the realm descriptor 124. For flexibility in the number of child realms which can be defined at a given generation of realm, the realm descriptors are managed using a tree structure called the realm descriptor tree (RDT) which will be discussed in more detail later. A realm descriptor 124 can be used to define properties of the realm which can be checked by the RMU 20 on entry or exit from a realm to ensure security. The realm descriptor may also track progression of a realm through various lifecycle states so that execution of certain RMU commands for a realm may be restricted to particular lifecycle states, to ensure that realms are created and invalidated in a secure manner.

Figures 1 1 and 12 show two different examples of possible realm hierarchies. In the example of Figure 1 1 , each of the processes shown in Figure 2 has its own realm defined. Hence, the root realm 130 corresponds to the monitor software or firmware operating at exception level EL3. The root realm defines two child realms 142, one corresponding to the secure operating system operating at secure EL1 and another corresponding to the hypervisor at EL2. The hypervisor defines grandchild realms 144 corresponding to different guest operating systems at EL1 and each of those guest operating systems define further great-grandchild realms 146 corresponding to applications executing at the least privileged exception level EL0. Similarly the secure operating system in realm 142 can define grandchild realms 148 corresponding to different secure applications. A parent realm in the hierarchy can transfer ownership of a memory page that it currently owns to a new child realm (by using a Granule.Add command as discussed below), or can make one of its pages invalid, map it into a child's virtual address space, and allow the child realm to claim ownership of the page by executing a page ownership (Claim) command. A page ownership command may be rejected if the specified page of the memory address space is not already owned by the parent realm which issued the command.

As shown in Figure 12, it is not essential for processes at every level of privilege to have separate realms, and so some of the privilege level boundaries shown in dotted lines in Figure 12 may not correspond to a realm boundary. For example in Figure 12 the application 150 and its operating system execute within the same realm as the hypervisor realm 142 operating at exception level EL2, and so a single realm spans both the EL2 hypervisor code, an operating system operating at EL1 and the application at EL0. On the other hand a different application 152 under the same hypervisor may have its own separate realm defined. In this case the realm boundary is between EL1 and EL0 and there is no EL2-EL1 realm boundary (both the hypervisor and operating system may execute in the same realm). For another operating system, a separate EL1 realm 154 may be defined, which again may have some applications which execute within the same realm as the operating system, and other applications which have their own dedicated realm. Similarly, on the secure side, the secure OS and applications in Figure 12 execute entirely within the EL3 root realm, and so there is no realm boundary when operating on the secure side. Hence, the precise configuration of realms may be determined at run time for a given system depending on the needs of the processes being executed. Software can decide at runtime whether it requires only a small and fixed number of child realms (which might be the case for low-level firmware) or needs many realms or a varying number of realms (which might be useful for a hypervisor on a cloud platform for example, which may manage an unknown number of guest virtual machines).

The realm descriptors 124 for a given parent realm are managed according to a realm descriptor tree (which is an example of a realm management tree which defines the realm management data for a number of child realms of that parent realm). The tree has a variable number of levels. Figure 13 shows an example of such a realm descriptor tree 160 managed by a particular parent realm. The tree 160 comprises a number of realm descriptor tree granules (RDTG) 162 which each comprise a number of realm descriptor tree entries (RDTE) 164. Each RDTE 164 provides a pointer to either a realm descriptor 166 for a given child realm of the parent realm, or a further RDTG 162 for the next level of the realm descriptor tree. The RDTG 162 for the first level of the tree may be identified by a realm descriptor tree pointer 168 which may be stored as part of the data associated with the parent realm (e.g. with the parent realm's realm descriptor). Hence, when the parent realm issues an RMU command associated with a given child realm it can trigger the RMU to traverse the realm descriptor tree in order to locate the required child realm's realm descriptor 166 (if that realm descriptor is not already cached within the RMU 20). Each RDTG 162 may have a variable number of entries 164.

As shown in the table in Figure 13, a given RDTE 164 which provides a pointer to an RDTG 162 at a subsequent level of the tree may include an order value which indicates the maximum number of entries in the pointed to RDTG. For example, the order value may indicate a power of two corresponding to the total number of entries in the pointed to RDTG. Other information which can be included in the RDTE 164 may include a state value indicating the state of the RDTE (e.g. whether the RDTE is free for allocation of realm descriptor tree data, and whether the RDTE provides a pointer to a further RDTG 162 or to a child realm descriptor 166). In addition to the pointer, the RDTE may also include a reference count which may track the number of RDTEs in the pointer to RDTG which are not free, which can be useful for determining whether further RDTEs an be allocated to that RDTG 162. RMU commands triggered by the parent realm may control the RMU 20 to establish further RDTGs of the tree and/or edit the contents of RDTEs within existing RDTGs.

Note that the tree shown in Figure 13 shows the child realms of one particular parent realm. Each other parent realm may have a separate realm descriptor tree tracking its own child realms. The data associated with the tree, including the RDTGs 162 and the child realm descriptors 166, are stored within pages owned by the parent realm and so other realms can be excluded from accessing this data. Therefore, only the parent realm may have visibility of which particular child realms it configured, so that processes executing at higher privilege levels may have no visibility of what realms have been created below any child realms which it has directly created itself.

As shown in Figure 13, each of the child realms of a given parent realm may have a corresponding realm identifier (RID) 168, which is used by that parent realm to identify a particular child realm. The RID is a local realm identifier in that it is specific to a particular parent realm. Child realms of different parent realms may have the same local RID. Although it is possible to use local RIDs which have arbitrary values selected by the parent realm for a given child realm, in the approach shown in Figures 13 and 14, the local RID for a given child realm has a variable number of variable length bit portions and each of the variable length portions is used by the RMU 20 to index into a given level of the realm descriptor tree 160. For example, the realm descriptor of the child realm having local RID = 7 in Figure 13 is accessed by following the realm descriptor pointer in entry 7 of the first level RDTG 162. The realm descriptor of child realm with local RID = 3.3 is accessed by following entry 3 in the first level of the tree and then entry 3 in the second level of the tree. Similarly the realm descriptor for child realm with local RID = 1 .2 is accessed by following entry 1 in the first layer and entry 2 in the second layer.

In Figure 13 the local RIDs are illustrated in decimal form but Figure 14 shows how these can be represented using a binary identifier. The binary identifier may have a number of variable length bit portions 170 and each of those bit portions may specify the index to use at a corresponding level of the RDT. In this example the binary local RID is filled up from the least significant end first, so that the lowest 5 bits correspond to the entry in the first level of the tree, the next 4 bits correspond to the entry in the second level, and so on. In this example the full local RID corresponds to 7.1 .3 as the least significant portion provides value 7, the next least significant bit portion provides value 1 , and so on.

The number of bits to be used within each of the variable length bit portions 170 can be determined by using the order values in the RDTEs which are traversed as the RMU 20 steps through the realm descriptor tree. The order value for the first bit portion of the local RID may be defined within the parent realm's realm descriptor (along with the first level RDTG pointer). In the example of Figure 14, for instance, the order values for the 3 levels are 5, 4 and 2 respectively, indicating that the first 5 bits indicate the index for level 1 , the next 4 bits indicate the index for level 2, and the next 2 bits indicate the index for level 3 of the RDT. Eventually an end of RID (termination) marker 172 is reached, which has a predetermined pattern indicating that there are no further levels of the tree to traverse, and so the next memory region accessed from the pointer in the last level RDTG will provide the realm descriptor for the required child realm. Any unused bits are filled with zeroes. Although Figure 14 shows an example where the RID is filled up from the least significant

end, it would also be possible to construct a RID by concatenating the indices required for stepping through the tree starting from the most significant end of the binary value.

This approach provides a flexible infrastructure for allowing different numbers of child realms to be established by a given parent realm, and for allowing the realm descriptors for those child realms to be accessed efficiently. As the realm identifier explicitly provides the indices required for stepping through the realm descriptor tree, there is no need to maintain a mapping table mapping arbitrary realm numbers to particular routes through the tree. As a tree structure is used, then in comparison to a table structure which would provide a certain fixed number of entries, the tree can be extended as required for the number of child realms by adding additional RDTGs or adding additional RDTEs to a given level of the tree as appropriate. Hence, the structure is scalable to the needs of different software processes. As it is not specified in advance exactly which portions of the RID should map to a given level of the tree, then this enables the available bits of the RID to be allocated flexibly to accommodate different depths/breadths of tree.

The realm management unit 20 may be responsible for accessing the realm descriptor tree when required, e.g. when entering or exiting a realm in order to identify the properties of the new realm and determining whether or not entry to the new realm is permitted, or when deciding whether it is appropriate to execute certain RMU commands. In general the ream management circuitry may index the realm descriptor tree using a realm identifier of a given realm which comprises a variable number of variable length bit portions with a given one of the variable length bit portions providing the index into a given level of the realm descriptor tree. This enables flexible dimensioning of the realm descriptor tree in order to support a wide variety of requirements for different software processes to establish child realms.

The realm management circuitry 20 may determine how many bits are comprised by the variable length bit portion to be used for indexing into a given level of the realm descriptor tree, based on an order value specified by an entry index at a previous level of the realm descriptor tree. By storing the order values within entries of the tree then it is not necessary to predefine which bits of the RID are to be used to be mapped to a given level of the tree and this provides further flexibility.

The RMU may permit variable length bit portions for indexing into different branches at the same level of the realm descriptor tree to have different numbers of bits. That is, although in Figure 13 both of the RDTGs 162 shown in layer 2 have the same order value (and hence the same number of entries), this is not essential and some implementations may have different RDTGs 162 at the same level of the tree having different numbers of entries. Therefore, the corresponding portions of the respective realm's RIDs may have different numbers of bits for the same level of the tree. Hence, the variation in the length of the bit portions used for indexing into given length of the tree may not only vary from parent to parent, but could also vary within different branches of the tree managed by one parent, providing further flexibility in the way in which child realms can be defined.

In general the RID for a given realm may comprise a concatenation of the indices to be used at respective levels of the realm descriptor tree to access the realm management data for the given realm. Although it is not essential for the indices to be concatenated in the same order as the order in which they would be used for stepping through the tree, this can be preferred as it makes management of the tree accesses simpler. It is not important whether the concatenation goes from low to high or from high to low. The concatenation of indices may be followed by a predetermined termination pattern which can allow the RMU 20 to determine when there are no further levels of the tree to be stepped through.

Some implementations could apply this RID construction technique to a global realm descriptor tree which may store realm descriptors for all the realms within the system in a similar tree structure (with each of the RIDs being a globally unique value). However, software development can be made simpler by defining the child realms of a given parent within one tree and then having separate trees for each other parent realm to track its child realms. Hence the realm descriptor tree may be a local realm descriptor tree associated with a given parent realm for storing the realm management data of the child realms which have been initialized by the given parent realm. The realm identifier may therefore be a local realm identifier which identifies a particular child realm used by a given parent realm. Child realms initialized by different parent realms may be permitted to have the same value of the local realm identifier. In this way, a parent realm can select which RIDs are used for its child realms without being aware of what other realms have been established by other parent realms, with the RIDs for the child realms being constructed depending on the way in which the parent realm has configured its realm descriptor tree.

The local realm identifier can be used by realm entry instructions or RMU commands issued by software processes. However, the hardware infrastructure may use an absolute identification of a given child realm to distinguish realms created by different parents. Hence, in addition to the local realm identifiers shown in Figures 13 and 14, a given realm may also have a global realm identifier (or "internal" realm identifier) which is unique to the given realm. At least one hardware structure may identify a given realm using a global realm identifier (GRID) instead of a local realm identifier (LRID). For example, the realm granule table 128 and/or TLBs 100 may identify realms using the global realm identifier.

In some examples an arbitrary binary value could be allocated as the GRID for a given realm, which can be completely unrelated to the LRID used by the parent realm to refer to that child realm. Different micro-architectural implementations of the same realm architecture could use different approaches for allocating the GRID.

However, in one example as shown in Figure 15 the GRID for a given realm may be constructed based on the LRIDs of its ancestor realms. This can be useful because it can be enable simpler determination of whether a given realm is a descendant of another realm or an ancestor of another realm which may be useful for the access control by the MMU 26 and RMU 20. In particular, the realms which share a common ancestor realm may share a common prefix or suffix portion of the GRID. For example, the GRID for a child realm may comprise a concatenation of the GRID of its parent realm and the LRID of that child realm. For example, as shown in Figure 15, if a given realm 180 has a certain local realm identifier X, then its child realms 182 may each have a GRID which is formed by concatenating X with the specific LRIDs formed based on the indices in realm X's realm descriptor tree. For example, the child realm with LRID 1 .2 may have a GRID X.1 .2. Similarly, a grandchild realm of realm X.1 .2 having LRID 7 may have a GRID X.1 .2.7.

In some cases, in the GRID, the LRIDs could be concatenated including the termination marker and zero-filled bits shown in Figure 14. Alternatively, the binary representation of the GRID may exclude such termination markers and zero-filled bits, and instead the meaningful parts of the LRIDs which contain RDT indices could be concatenated directly. As each of the LRIDs may themselves have a variable number of bits depending on the depth and breadth of the RDT for the relevant parent realm, then the number of bits of the global RID which are allocated to representing the local RID of the given generation of realm may be variable. Again, this variation in which parts of the global RID are allocated to a given generation of realm may vary at runtime based on the particular software that is running, but may also vary between different branches of the "family tree" of realms so that one branch of the family tree may use a greater portion of the realm identifier than others. As the common prefix or suffix of the GRID is the same for realms sharing a common ancestor, any subsequent generations can still be distinguished by the remaining portion which is specific to the subsequent generations, regardless of how that remaining portion is divided up among further generations.

By constructing the GRID as a concatenation of LRIDs for a number of ancestor realms, this enables more efficient determination of whether a first realm is an ancestor or descendant of a second realm. Circuitry may be provided (for example within the TLB 100 or RMU 20) to determine whether a global RID of one of the first and second realms matches a prefix or suffix portion of the global RID of the other realm, for example by using bit masking to mask out portions of the global RID corresponding to later generations in order to allow a comparison between the global RIDs of earlier and later realms within the same family to be matched.

It is not essential for all local RIDs to be constructed using the concatenation of tree indices approach shown in Figure 13. In some cases it may be useful for particular values of the local RID to be reserved for referring to certain default realms. RMU commands specifying the current realm or the parent realm of the current realm may be relatively common. Hence, a predetermined RID value may be reserved for referring to the parent realm of the current realm. For example an LRID with all bits set to 1 (indicating a value of 1 ) could be reserved for referring to the parent realm of the current realm. Similarly, a predetermined realm identifier value may be reserved for referring to the current realm itself. For example a LRID value of 0 could be used to refer to the current realm. Note that the use of the RID termination marker 172 in the LRID helps to support a value of zero being used as a predetermined realm ID value for the current realm, as it means that even if a child realm descriptor is identified using RDTE 0 in each level of the RDT, the resulting RID value is still not equal to zero as it may include this termination marker 172 which may have at least one bit set to 1 . Hence, in addition to indicating the end of the RID, the termination marker 172 can also support other functions.

The RMU may support certain query commands which may be triggered by a given realm in order to query constraints which it must satisfy when it is establishing its realm descriptor tree. For example, in response to the query command, the RMU 20 (or the processing circuitry 32) may return a constraint value which indicates at least one of a maximum number of levels of the realm descriptor tree 160 permitted to be defined by the given realm, a maximum number of tree entries permitted at a given level of the tree structure for the given realm, and/or a maximum number of child realms that may be initialized by the given realm. For example, the system may include registers which may indicate properties such as the number of bits available in the LRID or GRIDs used for a particular hardware implementation. The RMU or processing circuitry may, in response to the query command, check the number of bits available for the realm identifiers (or the appropriate response may be hardwired for the particular processor implementation), and could also check information specifying how many bits of the identifiers have been already used up by ancestor realms in the global realm identifier, in order to determine how many bits are available remaining for the current realm to define further children. The parent realm can use the response to a query command to decide how to construct its RDT.

Figure 16 shows an example of contents of a realm descriptor 166 for a given realm.

It will be appreciated that this is just one example and other implementations may not include all of the listed information, or could it include additional information. In this example the realm descriptor includes the following:

• The global RID of the realm. Hence, by traversing the realm descriptor tree based on the local RID, the corresponding global RID can be identified and this can be used to index hardware structures such as the TLB by the given realm, or check ownership tables or other information defined based on the GRID.

A lifecycle state of the given realm, which can be used by the RMU 20 to determine whether to accept a given command triggered by the given realm.

A type of the given realm. For example, the realm type could indicate whether the realm is a full realm or a sub-realm as discussed later.

A boundary exception level (BEL) value, which identifies a boundary exception level for the corresponding realm. The BEL indicates a maximum privilege level at which the realm is permitted to execute. For example, realm 142 in Figure 12 may have a BEL of EL2, realm 152 may have a BEL of ELO, and realm 154 may have a BEL of EL1 . By providing an explicit parameter identifying the BEL in the realm descriptor, this provides flexibility for realms to span multiple exception levels, as the BEL can be used on occurrence of an exception to determine whether the exception can be taken within the current realm or whether a realm exit to a parent realm is required to handle the exception.

A resource count indicating the total number of memory regions (realm protection granules or RPGs) owned by the realm and its descendants. This is used to ensure that all memory pages owned by the realm and its descendants are invalidated (and ultimately wiped of data), before those memory regions can be allocated to a different realm. For example the resource count can be used to track how many regions still need to be scrubbed.

Start and end addresses of a protected address range for the realm. For example the protected address range may define the range of memory address space within which pages can be owned by the corresponding realm. This can be useful for protecting against malicious parent realms reclaiming ownership of a region previously allocated to a child realm in an attempt to access the child realms data, since by comparing the protected address range defined in the realm descriptor with subsequent addresses of memory accesses, cases where a memory region previously owned by the realm is no longer owned by the realm can be identified. One or more encryption keys used by the cryptographic circuitry 104 to encrypt or decrypt data associated with a given realm. In this example two separate encryption keys are provided: a memory key for encrypting/decrypting contents and memory owned by the realm and a paging key for encrypting/decrypting data which is exported/imported between memory 16 to persistent storage 6 as discussed above. However in other examples the same key could be used for both purposes, or further keys could be provided for other specific purposes.

A realm description tree entry (RDTE) which identifies the root of the realm descriptor tree. The RDTE in the realm descriptor provides a pointer for accessing the root

RDTG (and the order value defining how many bits to use as the index for that RDTG).

• A pointer to a primary REC (realm execution context) memory region for saving or restoring architectural state associated with execution of the realm.

Figure 17 shows a set of lifecycle states in which a given realm can exist, which in this example include a Clean state, a New state, an Active state and an Invalid state. Figure 17 summarizes the properties of each state, indicating for each state: whether a realm in the corresponding state can have the parameters of its realm descriptor 166 modified by its parent realm, whether the encryption key specified for that realm can be validly used, whether the realm can own any memory regions (RPGs) and whether or not code associated with that realm is executable. Note that the parameters of the realm descriptor are modifiable in the Clean state but cannot be modified in any of the other states. This prevents a malicious parent realm updating the properties of a given realm after it has become active. Also the realm is executable only in the Active state.

Figure 18 is a state machine diagram showing allowed transitions of lifecycle states of the realm. Each state transition shown in Figure 18 is triggered by the parent realm issuing, to the RMU 20, a realm management command which specifies the local RID of the child target realm (the Realm. Invalidate command 212 can also be issued by the target realm itself). When no previous realm has been defined for that local RID and a realm descriptor register granule command 200 is executed by the parent realm, this triggers the configuration of a given memory region owned by the parent realm as a realm descriptor for the child realm having the specified local RID. The global RID for the child realm may be set based on the concatenation of the parent realm's global RID and the new local RID specified in the realm descriptor register granule command 200. The specified child realm then enters the Clean state 202. In the Clean state, the parent realm can set the properties of the child realm by updating the various parameters of the child realm's realm descriptor. These properties can be modified using further RMU commands issued by the parent realm (such realm descriptor modifying commands may be rejected if the specified child realm is not in the Clean state). When the parent realm has finished setting the parameters of the child realm's realm descriptor, it executes a realm initialize command 204 specifying the child realm's LRID and this triggers a transition of the child realm from the clean state 202 to the New state 206, and at this point the parameters of the realm descriptor are no longer modifiable by the parent realm. A realm initialization command 204 will fail if the specified realm is not currently in the Clean state.

When a realm is in the New state 206, execution of the realm activate command 208 specifying that realm's local RID triggers a transition from the New state 206 to the Active state 210 in which the realm is now executable, and after this point a realm entry into the corresponding realm will no longer trigger a fault. The realm is now fully operational. A subsequent realm invalidation command 212 triggered by the parent realm for a child realm in any of the Clean, New and Active states 202, 206, 210 results in a transition to the Invalid state 214. To leave the invalid state 214 and return to the Clean state 202, the parent realm has to execute a realm scrubbing command 216. The realm scrubbing command 216 is rejected if the resource count tracking the number of pages owned by the realm has a value other than zero. Hence, in order for the realm scrubbing command 216 to be successful, the parent realm first has to issue a Granule.Reclaim command for every page owned by the invalid realm. The Granule.Reclaim command specifies a target memory page and triggers invalidation of the target page to make the page inaccessible, and also reduces the reference count for the page's owner realm by one. It is not necessary to actually overwrite the data in the invalid regions when executing a Granule.Reclaim or realm scrubbing command 216, as the overwriting can be done when a Clean command is subsequently issued to transition the memory page from Invalid to Valid (see Figure 22 discussed below). Also in response to the realm scrubbing command any cached data associated with the invalidated realm may also be invalidated, for example within the TLBs 100 or caches 1 10 of any of the processing elements 8, 10, 12 (not just the processing element executing the RMU command). The global RID may be used to trigger such invalidations of cached data.

Hence, by providing a managed lifecycle of the realm associated with a given realm identifier, this ensures that before the realm can return to the Clean state in which its parameters can be modified (and hence before a given realm identifier can be recycled for use by a different realm), the data associated with a previous realm which used the same realm identifier has to be scrubbed from memory and any caches, to prevent any data associated with the old realm being leaked to other realms through the reuse of the same realm identifier. When a realm is in the Clean state 202, its realm descriptor can also be cancelled by executing a realm descriptor release command 218 which enables the memory regions stored in the realm descriptor to be allocated for other purposes (at this point no scrubbing is required as the realm is already clean).

Figure 19 shows an example of contents of an entry of the realm granule table 128 (or ownership table). Each entry corresponds to a given memory region of the memory address space. The size of a given memory region could be fixed or variable depending on implementation. The particular way in which the ownership table 128 is structured may vary significantly depending on implementation requirements, and so the particular way in which the corresponding memory region for a given entry is identified can vary (e.g. data could be stored in each entry identifying the corresponding region, or alternatively the corresponding entry may be identified at least partially based on the position of the corresponding ownership entry within the table itself). Again, Figure 19 shows a particular example of

parameters which may be specified for a given memory region but other examples could provide more information or could omit some of the types of information shown.

As shown in Figure 19, each ownership table entry may specify, for the corresponding memory region:

• The global RID identifying the owner realm for that memory region. The owner realm may be the realm which has the right to set the attributes which control which other realms are allowed to access the memory region.

• A lifecycle state of the corresponding memory region used to control which RMU commands are allowed to be performed on the memory region.

• A mapped address to which the memory region was mapped by the MMU 26 at the point when the memory region became owned by the owner realm. The mapped address could be a virtual address or an intermediate physical address. By specifying this address in the ownership table, this can guard against potential attempts to circumvent the security provided by the realm infrastructure by remapping the address translation tables after a realm has taken ownership of a given memory region.

• A visibility attribute which specifies which realms in addition to the owner can access the memory region. For example, as shown in Figure 20, the visibility attribute could specify a parent visibility bit which controls whether or not the parent realm of the current realm is allowed to access the region, and a global visibility bit which may specify whether any realm may access a corresponding memory region. In general, the realm protection scheme may assume that descendant realms of the current realm are always allowed to access the memory region owned by its parent or ancestor realm (subject to whether or not accesses are permitted based on the translation tables 120 which provide protection based on privilege level), but a given realm may control whether its parent or any other realm which is not a direct descendant of the given realm can access the memory region. In some embodiments both the parent and global visibility bits could be set by the owner realm itself. Alternatively, while the parent visibility bit may be set by the owner realm, the global visibility bit may be able to be set by the parent realm of the owner realm (provided that the parent visibility bit for the memory region has already been set to give the parent visibility of that memory region). It will be appreciated that this is just one example of how the owner realm could control which other processes could access its data.

Figure 21 is a table showing different lifecycle states in which a given memory region can exist, and Figure 22 is a state machine showing the commands that trigger transitions between the respective lifecycle states. In a similar way to the realm lifecycle states shown in Figure 18, the transitions between memory region lifecycle states are managed to ensure that a memory region passing from ownership by one realm to ownership of another realm must first undergo an invalidation process where data in that region is scrubbed (e.g. set to zero). Hence, in order to transition from an Invalid state 220 to a Valid state 222 in which software can access the memory region, a clean command 224 has to be executed by the RMU 20, triggered by software executing on the processing element 8. The clean command 224 identifies a particular memory region (page) and controls the RMU to step through the memory addresses of the corresponding memory region to invalidate/zero the data in each location within that memory region. The clean command is rejected (e.g. a fault is triggered) if the targeted memory region is in any state other than Invalid.

In some systems it may be sufficient to provide valid and invalid states 222, 220 as the only memory region lifecycle states. However, in the example of Figure 22, a given memory region can also be designated as an 'RMU-private' memory region which is reserved for exclusive access by the RMU 20 itself, so that accesses to an RMU-private memory region which are triggered by software executing on the processing circuitry 32 (other than any RMU software) would be rejected. This could be useful especially for storing the realm management data such as the realm descriptors, realm descriptor tree entries, realm execution context and meta data for paging as discussed above. By providing attributes for designating a given memory region as an RMU-private memory region which is reserved for exclusive access by the RMU, this can prevent software processes, including the owner process of the memory region itself, from being able to access the realm management data which could otherwise allow software processes to circumvent the security protections provided by the realm scheme.

Hence, the clean command 224 may specify, as one of its parameters, a privacy indication specifying whether this is a normal clean command or a private clean command. Alternatively two entirely separate commands could be provided for these purposes. When the clean command is a normal clean command, this triggers the transition to the Valid state 222 as discussed above. However, when the clean command is a private clean command 224, this triggers a transition to the RMUCIean state 226, in which the memory region is designated as an RMU-private memory region. In some examples all types RMU data could be stored within a single type of RMU-private memory region corresponding to the RMUCIean state.

However, robustness can be improved by designating multiple types of RMU-private memory region, each corresponding to a particular form of realm management data. For example, in Figures 21 and 22, a number of RMURegistered states 228 are defined which each correspond to RMU private regions which are designated for a particular purpose. In this example, the RMU Registered states 228 include RMURegisteredRDT (for storing a RDTG of a realm descriptor tree), RMURegisteredRD (for storing a realm descriptor), RMURegisteredREC (for storing realm execution context data) and RMURegisteredMDT (for storing paging metadata used during export/import operations as discussed above). Different forms of registration commands 230 may be executed by the RMU for a memory region in the RMUCIean state, to transition the memory region to a corresponding one of the RMURegistered states 228. A command for storing data to the an RMU-private memory region which does not correspond to the specified purpose (RDT, RD, REC or MDT) may be rejected. Hence, in a first lifecycle state of the RMURegistered states, a first type of RMU command for storing a first type of realm management data may be allowed, and in a second lifecycle state a second type of RMU command for storing a second type of realm management data may be allowed, with the first RMU command being rejected when the targeted memory region is in the second lifecycle state and the second RMU command being rejected when the targeted memory region is in the first lifecycle state. This can enable further security by avoiding a malicious parent realm for example attempting to store realm descriptor entries to a realm execution context region or vice versa in order to attempt to disrupt the operation of the child realm. From each of the RMU register states 228, a corresponding form of release command 232 may return the corresponding memory region to the Invalid state 220. A further clean command may trigger scrubbing of data from a previously defined RMU-private region before the region can be reallocated for general purpose data.

Hence, in summary, at least one RMU-private memory region may be defined which is still owned by a given owner realm, but which has an attribute specified in the ownership table which means that it is reserved for exclusive access by the RMU. In this example the attribute which control the RMU-private status is the lifecycle state specified in the corresponding entry in the ownership table, but it could also be identified in other ways. The MMU may prevent access to a given memory region by one or more software processes when the given memory region is specified as the RMU private memory region by the at least one status attribute. Hence any software-triggered access which is not triggered by the RMU itself may be rejected when it targets an RMU-private memory region. This includes preventing access to an RMU-private memory region by the owner realm itself.

One may question why it is useful to define an owner realm for an RMU-private memory region if the owner realm cannot even access the data in the memory region. For example, an alternative approach for enforcing access to data by the RMU only would be to define a special realm for the RMU, and to allocate the pages of the memory address space for storing the data to be kept private to that special RMU owner realm. However, the inventors recognized that when invalidating realms, there may be a requirement to invalidate all the control data associated with that realm, and if this control data was associated with special RMU owner realm rather than the invalidated realm, this can make scrubbing of the invalidated realm's data more complex.

In contrast, by using the RMU-private attribute, the memory regions which store the control data for a given realm are still owned by that realm, even though the owner cannot access it, which means that it is simpler to identify which memory regions need to be invalidated when that owner realm is cancelled. When a given realm is invalidated, the parent realm may simply perform a sequence of reclaim operations (e.g. by executing a reclaim command which is then acted upon by the RMU) which triggers the memory regions which were owned by the specified invalidated realm (or its descendants) to be invalidated, and made inaccessible, and returned to ownership of the parent realm which triggered the reclaim command. The reclaim operation may not only affect the pages accessible to the invalidated realm but also include the RMU-private memory regions owned by the invalidated realm.

Another advantage of storing the control data for a realm in RMU-private memory regions owned by that realm is when performing export operations. In order to reduce the memory footprint of a Realm to zero, during an export operation, the management structures associated with that Realm may be exported in addition to the normal memory. Requiring that those structures are owned by the Realm simplifies the management of this export operation.

In general any kind of realm management data could be stored in RMU-private regions, but in particular the realm management data may include any of the following: a realm descriptor defining properties of a given realm, a realm descriptor tree entry identifying a memory region storing the realm descriptor for a given realm or a further realm descriptor tree entry, realm execution context data which indicates architectural state associated with at least one thread of execution within a given realm, and temporary working data used at an intermediate point of a predetermined operation associated with a given realm.

Although generally the RMU private regions could be used for storing the specific realm control data associated with a given realm, it could also be used in order to increase security around certain other operations which are performed once a realm is active. For example, when performing the paging export or import operations discussed above where data is encrypted or decrypted, and the checks using the meta data are performed to check that the data is still valid when the data is imported again, such operations may take many cycles and such long running operations are more likely to be interrupted partway through. To avoid needing to start the operation again from the beginning, it can be desirable to allow the meta data or other temporary working data associated with such long running operations to be retained within cache/memory even when interrupted, without making such data

accessible to other processes (including the owner realm itself). By temporarily designating regions of the memory system as RMU-private regions this temporary working data can be protected. Hence, as shown in Figure 21 the page states can also include RMUExporting and RMUImporting states which can be used when this temporary working data is stored to a memory region, and when one of these states is selected then only the RMU can access that data.

Other examples of operations which may benefit from temporarily designating a corresponding memory region as RMU private can include: generation or verification of encrypted or decrypted data during the transfer of data between at least one memory region owned by the given realm and at least one memory region owned by a realm other than the given realm; a transfer of ownership of a memory region to another realm; and a destructive reclaim operation performed to make data stored in an invalidated memory region inaccessible. For example, the reclaim operation for scrubbing all the contents of a given page of the address space may be interrupted partway through and so to ensure that other processes cannot access that page until the scrubbing is complete, the page could temporarily be designated as RMU-private. In general any long latency operation performed by the RMU could benefit from having its temporary working data protected by converting the lifecycle state of some memory regions to an RMU-private state before starting the long running operation and then converting it back again when the long running operation is complete.

When a region is designated as RMU private it is reserved for access by the RMU 20 used to perform realm management operations. The realm management operations may include at least one of: creating a new realm; updating properties of an existing realm; invalidating a realm; allocating memory regions for ownership by a given realm; changing the owner realm for a given memory region; changing a state of a given memory region; updating access control information for controlling access to a given memory region in response to a command triggered by the owner realm for the given memory region; managing transitions between realms during processing of the one or more software processes; managing transfer of data associated with a given realm between a memory region owned by the given realm and a memory region owned by a different realm to the given realm; and encryption or decryption of data associated with a given realm. The RMU could be a hardware unit to perform at least a portion of the realm management operations, or could comprise the processing circuitry 32 executing realm management software to perform at least a portion of the realm management operations, or could be a combination of both.

Figure 22 shows the state transitions which can be triggered by a given realm to clean a given page so it can be validly accessed, or invalidate the corresponding page.

Figure 23 expands this to show further commands which can be used to transfer ownership of a given page from one realm to another. Execution of a region claim command 230 by a parent realm enables the corresponding memory region to be transferred to a specified child realm if that memory region is currently in an Invalid state 220 and is owned by the parent realm. The region claim command 230 is rejected when the target memory region is owned by any realm other than the parent realm of the given child realm, or if the memory region is Valid or in one of the RMU-private lifecycle states 226, 228. This prevents the parent realm arbitrarily assigning ownership of a page which is does not itself have access to or which is in use by the RMU 20. Once the page has been assigned to the child realm, then that child realm can perform the clean command in order to transition to the valid state 222 in the same way as shown in Figure 22. For conciseness the use of RMU-private regions is not shown in Figure 23, but within any given realm a private clean command could alternatively transition the memory region to the RMU clean state 226 as discussed before.

A granule claim command 230 is used to transfer ownership to a child realm which has already been established. In addition, a parent realm can execute a granule add command 232 which triggers the RMU 20 to assign ownership to a new child realm in a New state so that the parent realm can write data to the region assigned to the child. For example this can be used for installing the program code of the new child realm so that the child realm can executed for the first time. Hence the add command 232 differs from the claim command 230 in terms of the lifecycle state in which the corresponding memory region is allocated to the child realm. The add command 232 may be allowed only when the child realm is in the New state 206 shown in Figure 18. A child realm can release ownership of a given memory region back to its parent by executing a granule release command 234, which triggers the RMU to update the corresponding entry of the ownership table 128 as well as updating properties such as the resource count in the realm descriptor of the child realm etc. The granule release command 234 may be rejected if the specified memory region is not owned by the current realm which issued the command, or if the region is in a state other than Invalid (ensuring destructive cleaning of the data is required before it can be returned to ownership by the parent realm).

One advantage of using the hierarchical realm structure discussed above, where parent realms initialize child realms, is that this greatly simplifies invalidation of a realm and its descendants. It is relatively common that if a given virtual machine realm is to be invalidated, then it may also be desirable to invalidate realms for any applications running underneath that virtual machine. However, there may be a significant amount of program code, data and other control information associated with each of the processes to be invalidated. It may be desirable to ensure that such invalidations occur atomically, so that it is not possible to carry on accessing data associated with the invalidated realms when only

part of the data scrubbing has been carried out. If each realm was established entirely independently of other realms, without the realm hierarchy as discussed above, this could make such atomic invalidations difficult as it may be necessary to provide a number of separate commands to separately invalidate each realm identified by a corresponding realm ID.

In contrast, by providing a realm hierarchy in which the RMU manages the realms such that each realm other than a root realm is a child realm initialized in response to a command triggered by a parent realm, then when a command requesting invalidation of a target realm is received, the RMU 20 can make the target realm and any descendant realm of the target realm inaccessible to the processing circuitry, with a more efficient operation.

In particular, in response to invalidation of the target realm, the RMU may update realm management data (e.g. the realm descriptor) associated with the target realm to indicate that the target realm is invalid, but need not update any realm management data associated with any descendant realm of the target realm. The realm management data associated with the descendant realm can remain unchanged. This is because simply invalidating the target realm may make any descendant realms effectively inaccessible even though the realm management data has not changed, because access to a given realm is controlled through its parent and so if the parent realm is invalidated then this means it is also not possible to access the parent realm's descendants. As each of the realms is entered using a realm entry instruction (ERET instruction discussed below) which uses the local RID defined by the parent realm to identify its particular child, and this is used to step through the realm descriptors which are stored in memory regions owned by the parent realm of the given child realm, then no process other than the parent realm can trigger the RMU to access the child realm's realm management data. Hence, if the parent realm is invalidated then the RMU cannot access the realm management data of the given child realm, ensuring that the given child realm becomes inaccessible.

After a realm has been invalidated, the parent realm of that realm may trigger the RMU to perform a reclaim operation for reclaiming each memory region owned by the invalidated target realm. For example as shown in Figure 23, a reclaim command 236 for a memory region owned by a child realm may trigger a return of the memory region to the Invalid state 220 and also transfer ownership of the memory region back to the parent realm. However this reclaim operation can be performed in the background of ongoing processing of other realms and need not be performed instantly in order to allow any descendant realms of the invalidated realm to be made inaccessible. A single action to change a realm state of a given realm from Active to Invalid as shown in Figure 18 is enough to make sure that all the data associated with any descendant realm of that invalidated realm is also inaccessible. Since any parent realm can only assign pages which it owns to its children, and the child can only assign pages which it owns to a grandchild realm, this also means that tracking which pages need to be invalidated and reclaimed upon invalidation of a given realm is relatively straightforward, since the protected address range defined in the realm descriptor of the invalidated realm (see Figure 16) can be used to identify which pages to reclaim, since any further descendant realms of the invalidated realms will also own pages within that range.

Hence, in summary using a realm hierarchy greatly simplifies management of realms and invalidations. On such invalidations, as well as overwriting data in memory, the invalidation may also trigger invalidation of cached realm management data for the target realm and any descendant realm of the target realm held not only in the processing element 8 which triggered the invalidation, but also in other processing elements, such as another CPU or GPU. Hence there may be a broadcast of the invalidation to other processing elements, to ensure that other processing elements do not continue to have access the invalidated realms. When triggering such invalidations, it can be useful for the cached realm management data to be associated with a global realm identifier which identifies the corresponding realm uniquely, and to form the global realm identifier as discussed above so that the global RID of a given child realm shares a common prefix portion with a global RID of its parent realm. This enables bit masking or other similar operations to be used to compare quickly whether a given realm is a descendant of a specified realm ID. If a given realm is made inaccessible through invalidation of an ancestor realm, then an attempt to enter the specified target realm is not possible (as there is no parent realm to execute an ERET instruction for that realm), but even in other implementations which use a different realm entry mechanism, a realm entry could fail and trigger a fault condition if the realm descriptor of the descendant realm is no longer locatable.

Figure 24 shows an example of checks performed by the MMU 26 and RMU 20 for determining whether a given memory access is permitted. The MMU 26 supports two stages of address translation, stage 1 translating a virtual address (VA) to an intermediate physical address (IPA) under control of stage 1 page tables 120-1 which are set by a given guest operating system, and a stage 2 address translation translating the intermediate physical address provided by the stage 1 translation into a physical address (PA) used to access memory 16 based on stage 2 page tables 120-2 which are set by the hypervisor 38. The hypervisor may define multiple sets of stage 2 page tables for different virtual machines, and a virtual machine ID (VMID) 250 provided with the memory access request may identify which particular stage 2 page tables to use. Similarly, the operating system may define multiple sets of stage 1 page tables for different applications and an address space identifier (ASID) 252 may be used to identify which stage 1 page tables to use. The VMID and ASID 250, 252 may collectively be referred to as a translation context identifier 254 which identifies a current translation context associated with the memory access request. The

memory access request also specifies various attributes 256, such as attributes indicating whether the transaction is a read (R) or write (W) request, or indicating an exception level (X) associated with the process which issued the memory access request.

On receipt of a memory access, the MMU 26 may determine based on the information from the stage 1 page tables whether the transaction attributes are valid. For example the stage 1 page tables could specify that only read transactions may be allowed for certain addresses, or could permit both read and write accesses to a given address (some implementations may also permit write only regions of the address space to be defined). Also, the attributes in the stage 1 page tables could limit access to processes operating at a given exception level or higher. If the transaction attributes are valid and the access is permitted by the stage 1 page tables, then the MMU may return the corresponding intermediate physical address (IPA). The IPA together with the VMID 250 then indexes into stage 2 page tables which again validate the attributes of the transaction and, if valid, return a physical address. Note that not all transactions need to undergo two stages of address translation. For example if the input memory transaction was issued at EL3 or EL2, or at EL1 or EL0 in the secure domain, then the output of the stage 1 MMU may be treated as a physical address and the stage 2 MMU could be bypassed.

Having obtained the physical address, the physical address can then be looked up in the RMU tables 128 (Realm Granule Table) to determine whether the realm protections enforced by the MMU allow the memory access to proceed. The realm checks are discussed in more detail in Figure 26 below. If the RMU checks at stage 3 succeed, then the validated physical address is output and the memory access is allowed to proceed. If any of the checks at stage 1 or stage 2 address translations or the RMU enforced realm protection provided at stage 3 fails, then the memory access is rejected. Hence the protections provided by the realm management unit can be seen as an additional layer of checks to be performed, in addition to any existing address translation checks based on page tables 120. The checks shown in Figure 24 can be relatively slow to perform as there may be a number of tables in memory which need to be accessed and compared with parameters of the memory access request or the current translation context or realm from which the access was made. While it would be possible to perform these checks on every memory access, it can be faster to cache data within the TLBs 100 when the checks have been performed successfully for a given memory access request so that the next time a similar memory access request is issued then it can be allowed without repeating all the checks again. Hence, it may be desirable to perform these permission checks only when there is a miss in the TLB 100, and not for a hit.

Figure 25 shows an example of a TLB structure 100 for caching data relating to already validated memory accesses. While Figure 25 shows a single TLB, it will be

appreciated that some systems may include multiple levels of TLB in a cache hierarchy, with a level 1 TLB storing a smaller subset of translation entries for faster access, and level 2 or further level TLBs storing a larger set of translation entries which can be accessed if there is a miss in the level 1 TLB. The TLB 100 (or "translation cache") has a number of entries 260, each entry specifying address translation data for a corresponding memory region. Each entry 260 includes a virtually addressed tag 262 corresponding to the virtual address for which the data provides the corresponding physical address 264. In this example, the TLB is a combined stage 1 and stage 2 TLB so that the virtual address can be translated directly to a physical address using the TLB without having to go via an intermediate physical address (although corresponding stage 1 and stage 2 translations would be performed on a TLB miss in order to locate the correct physical address, the TLB does not need to store the intervening IPA and can map the VA direct to a OA). Other examples could use split stage 1 (S1 ) and stage 2 (S2) TLBs, in which case the VA-PA pair 262, 264 can be replaced with a VA-IPA pair or I PA-PA pair. The TLB entries 260 are also tagged with the translation context identifier 254 (formed of the ASID 252 and VMID 250). Although this example provides two separate translation context identifiers, in other examples a single unified translation context identifier could be used, or in the case of split S1/S2 TLBs, the S1 TLB could use the ASID and the S2 TLB could use the VMID. The translation context identifier allows different operating systems or applications which specify the same virtual address to have their accesses mapped onto different entries of the TLB 100 which provide different physical addresses.

A hit in the TLB 100 requires not only the tag 262 to match a corresponding portion of the address 258 specified for the memory access request, but also the translation context identifiers stored in the same entry should match the current translation context from which the memory access was issued. It may be expected that the comparison of the tag 262 and the translation context identifiers 254 may be enough to locate the correct physical address 264 for a given memory access. However, if these are the only comparisons performed in the lookup, there is a potential security vulnerability if memory accesses hitting in the TLB are accepted without a further check of the realm management unit tables 128. This is because it is possible for a new process to be created which has the same VMID 250 or ASID 252 as a previously executed process, to trick the MMU into accepting a memory access which actually comes from a different realm to the one that was previously accepted for accessing a given memory region.

To address this problem, the TLB 100 may specify within each TLB entry 260 the global RID 270 of the owner realm which owns the corresponding memory region, as well as visibility attributes 272 set by the owner realm for controlling which other realms are allowed to access the corresponding memory region. When a given lookup of the translation cache

100 is performed in response to a memory access to a given target memory region issued from a current translation context and a current realm, if there is a miss in the translation cache 100, TLB control circuitry 280 may trigger the table walk unit 102 to access the relevant page tables 120 and RMU tables 128 in order to check whether the access is allowed. If either the page tables or the RMU tables 128 exclude the current combination of the translation context, exception level and realm from accessing the corresponding memory region, then no data is allocated to the translation cache in response to that memory access. In particular, when the lookup misses and the current realm is excluded from accessing the target memory region by an owner realm of the target memory region, then allocation of address translation data to the translation cache is prevented. Hence, an entry is allocated to the TLB 100 when the corresponding memory access passes the checks of both the MMU page tables 120 and the RMU tables 128.

Subsequently, when looking up the translation cache to check whether it already includes an entry 260 which provides the address translation for a given address, the TLB control circuitry 280 determines whether the memory access matches a given entry of the translation cache 100 in dependence upon a first comparison between the translation context identifier 254 specified in the corresponding entry 260 and the translation context identifier 254 for the current translation context which was received along with the memory access request, and also a second comparison between the global RID 270 specified by that entry 260 and a current global RID associated with the current realm which issued the memory access request. By providing an additional check that the TLB entry is still being accessed from a realm previously verified as allowed to access the memory region, this ensures that even if a malicious supervising process regenerates another process with the same ASID 252 or VMID 250 as a previously existing process which was permitted to access the data by the owner realm, since the global realm identifiers 270 cannot be reallocated to other processors without undergoing the realm scrubbing command 216 as discussed with respect to Figure 18, this means that the current realm's global RID can be trusted as valid and cannot be "faked" as is possible for the ASID or VMID. Therefore, if the global RID of the current realm still satisfies the permissions indicated by the owner GRID 270 and the visibility attributes 272, then this indicates that the previously performed realm table checks are still valid.

If the second comparison of the realm identifier detects a mismatch, then the access request is considered to miss in the TLB even if the tag comparison and translation context comparison match, as it indicates there has been a change in the mapping between the translation context IDs 254 and the realm IDs 270 since the entry was allocated. This does not necessarily imply that the access will be rejected, as another walk of the page tables and RMU tables may be triggered by the table walk unit 102, and if the realm checks are

successful this may lead to allocation of a different entry 260 in the TLB 100 and servicing of the memory access based on the information from the newly allocated entry.

Figure 26 is a flow diagram illustrating a method of determining whether a given memory access is permitted by the MMU 26. At step 300 a memory access request is received and this is looked up in the TLB 100. The memory access request specifies at least a virtual address to be accessed, one or more translation context identifiers indicating a current translation context and a global realm identifier identifying the current realm. For example the global RID may be read from a state register of the processing element 8 which may be written with the global RID of the current realm on entry into the realm.

In response to the memory access request, the TLB control circuitry 280 performs a lookup of the TLB. The lookup accesses at least some entries of the TLB. Some approaches may use a fully associative cache structure, and in this case all the entries of at least a level 1 TLB may be searched and compared with the parameters of the current request in order to identify whether there is a hit or a miss. Other approaches may use a set-associative cache allocation policy and in this case only a subset of the entries of a given level of TLB may need to be looked up, indexed using the target address of the memory access. For each of the accessed set of entries, the TLB control circuitry 280 performs a number of comparisons (either in parallel or sequentially), including:

• a tag comparison 302 for comparing whether the address of the memory access request matches the tag 262 stored in the accessed entry;

• a first (context) comparison 304 for comparing translation context identifiers stored in the accessed entry with the translation context identifier of the memory access request; and

• a second (realm) comparison 306 for comparing the global RID of the memory access request against the owner RID 270 and the visibility attributes 272 for each of the accessed set of entries.

At step 308 the control circuitry 280 determines whether there is an entry in the TLB which returned matches for all of the comparisons 302, 304, 306, and if so then a hit is identified and at step 310 the physical address 264 specified in the matching entry is returned and the memory access is allowed to proceed based on that physical address. In the case of a hit, there is no need to perform any lookup of the page tables or the RMU tables (an ownership table lookup for the memory access can be omitted). The protections provided by the page tables and RMU tables are invoked only on a miss.

It there is no entry which matches all three of the comparisons 302, 304, 306, then a miss is detected. If a further TLB level is provided, a corresponding lookup to steps 300-308 can be performed in a level 2 or subsequent level TLB. If the lookup misses in the last level TLB, then a walk of the various page tables and RMU tables is performed. Hence, at step 31 1 a stage 1 page table walk is performed and at step 312 it is determined whether a stage 1 page table fault has occurred (e.g. because there was no address mapping defined for the specified virtual address or because the current parameters 256 of the access request violate the access permission specified for the targeted virtual address). If a stage 1 fault occurs then at step 314 the memory access is rejected and allocation of address mapping data to the TLB 100 in response the memory access is prevented.

On the other hand, if the access request passed the stage 1 page table checks, then at step 315 a stage 2 page table walk is triggered to obtain mapping data for the intermediate physical address returned by the stage 1 process, and at step 316 it is determined whether a stage 2 page table fault has occurred (again, because the address mapping was not defined or because the access is not permitted by the stage 2 access permissions). If a stage 2 fault occurs then again at step 314 the access request is rejected.

If no stage 2 fault occurs then at step 318 the RMU table lookup is triggered based on the physical address returned by stage 2 and at step 320 it is determined if a realm fault has been detected. A realm fault may be triggered if any of the following events occurs:

• If the lifecycle state for the corresponding memory region is indicated as Invalid in the realm ownership table 128. This ensures that a page of the memory address space which has not undergone the cleaning operation 224 shown in Figure 22 cannot be accessed, to protect any data previously stored by another realm in that memory region from access by a different realm.

• The current realm is not permitted by the owner realm for the corresponding memory region to access that memory region. There could be a number of reasons why a given realm may not be allowed to access a given memory region. If the owner realm has specified that the memory region is visible only to the owner itself and its descendants, then another realm may not be allowed to access that region. Also if the current realm is the parent realm of the owner realm and the owner realm has not defined the parent visibility attributes to permit the parent to access the region, then the memory access may be rejected. Also, the owner realm itself may be prevented from accessing a memory region if that memory region is currently set as RMU- private as discussed above. At the RMU checking stage, a descendant realm of the owner realm may be permitted to access the memory region (as long as the memory region is not an RMU-private region). Hence, this check enforces the access permissions set by the owner realm.

· If the virtual address or intermediate physical address from which the physical address was mapped for the current memory access by the S1 /S2 translations does

not match the mapped address specified in the ownership table 128 for the corresponding memory region as shown in Figure 19, then the memory access is rejected. This protects against situations where a malicious parent realm could allocate ownership of a given memory region to a child realm, but then change the translation mappings in the page tables 120 so that a subsequent memory access triggered by the child realm using the same virtual address it previously used to refer to a page owned by that child realm now maps to a different physical address which is not actually owned by the child realm itself. By providing a reverse mapping in the ownership table from the physical address of the corresponding memory region back to the mapped address used to generate that physical address at the time the ownership was claimed, this allows security breaches caused by changes of address mappings to be detected so that the memory access will fail.

It will be appreciated that other types of checks could also be performed. If the realm checks are successful, at step 322 the physical address is returned, the memory access is allowed to proceed using the physical address, and a new entry is allocated to the TLB indicating the physical address obtained from the page tables 120 and the owner realm and visibility attributes obtained from the ownership table 128 corresponding to the requested virtual address and translation context.

Hence, in summary by requiring a second comparison (comparing the current realm's GRID with a GRID provided in an entry of the translation cache) to match in order to allow a hit to be detected in the translation cache lookup, this ensures that even if there is a change in the translation context identifier associated with a given realm after a TLB entry has been allocated, this cannot be used to circumvent the realm protections even if the realm checks are not repeated again on a TLB hit. This enables performance to be improved, as by making it unnecessary to repeat the realm checks on every memory access (which would be relatively processor intensive given the number of checks to be performed). This allows most memory accesses to be performed faster, since hits are much more common than misses. A mismatch between the memory access and a given entry of the translation cache is detected when the second comparison identifies a mismatch between the realm identifier specified in that entry and the realm identifier of the current realm. This will then trigger a miss and this can trigger page table and RMU table walks in order to find the correct access control data (with the realm checks repeated in case the VMID/ASID has changed).

This approach is secure because the RMU may prevent initialisation of a new realm having a same realm identifier as a previously active realm until after a scrubbing process invalidating information associated with the previously active realm has been performed. This scrubbing process may include invalidation of not only the realm management data and any data stored in memory associated with the invalidated realm, but may also include

invalidation of at least one entry of the translation cache for which the second comparison identifies a match between the realm identifier of that entry and the realm identifier of the invalidated realm. Hence, this means that it is not possible to regenerate a different process using the same realm identifier as a previous process, unless all the data in the translation cache 100 associated with that realm has been invalidated. Hence, a matching realm identifier in the TLB can be trusted to indicate that the previously performed realm checks are still valid. In particular, each realm may be associated with a life cycle state as discussed above and this may be used to enforce that the scrubbing process is performed. Address translation data may only be able to be allocated to the translation cache when the current life cycle state for the current realm is Active. A command for initialising a new realm with a given realm identifier may be rejected in any state other than a Clean state and transitions from the Active state to the Clean state may require a predetermined sequence of command to be executed which includes at least one command for triggering the scrubbing process.

Misses in the translation cache may trigger an ownership table lookup, which accesses an ownership table which specifies, for each of a number of memory regions, the owner realm for the corresponding memory region and access constraints set by the owner realm for controlling which other realms are allowed to access the memory region. By including the additional second comparison for determining a TLB hit, this enables the ownership table lookup to be omitted on lookup hits. The ownership table lookup is performed on a TLB miss.

While Figure 25 shows an approach where the GRID of the owner realm is stored in each TLB entry, there could also be other ways of representing information which enables a determination of whether the GRID of the current realm is suitable for accessing the corresponding memory region. For example, a list of the GRIDs of authorized realms could be maintained in the TLB, or a TLB could maintain a separate list of Active Realms with the TLB entry containing an index into the Active realm list, instead of the full GRID, which could reduce the TLB entry size compared to storing the list in the TLB entry. However, simply representing the owner realm's GRID can be a more efficient way to identify the authorized realms, as it makes the processes of allocating and checking TLB entries less complex by avoiding an additional level of indirection in consulting the Active realm list and also avoids needing to synchronise changes in the active realm list between TLBs.

Note that a match in the second (GRID) comparison performed on looking up the TLB does not necessarily require the current realm identifier to be exactly the same as the global realm identifier 270 specified in both the corresponding TLB entry 260 - some forms of the second comparison may use partial matching. Some implementations could only permit the owner realm to access its owned pages and so in this case an exact match

between the current GRID and owner GRID 270 may be required. However, as it may be useful for data to be shared between realms, visibility attributes 272 can also be provided to allow the owner realm to define what access is permitted by other realms.

Therefore, by caching the visibility attributes 272 in the TLB 100 as well as the owner realms global RID 270, this enables the TLB control circuitry 280 to vary an extent of matching required for the second comparison to determine a match based on the visibility attributes. For example, the visibility attribute 272 could control which portions of the GRIDs should be masked when performing a comparison, so that it does not matter if those masked bits mismatch as they do not affect the overall matching of the comparison. For example, the control circuitry could in some cases determine a mismatch when the current realm indicates a realm other than the owner realm or a descendant realm of the owner realm. The descendant realms can easily be identified using the global RID format discussed above since they will have a prefix or suffix portion which matches the owner realm's GRID.

For at least one value of the visibility attributes, the control circuitry can determine a mismatch when the current realm is a realm other than the owner realm, a descendant realm of the owner realm or the parent realm of the owner realm (e.g. when the parent visibility is set as discussed above). In some cases, at least one value of the visibility attribute may allow the control circuitry 280 to determine a match for the second comparison regardless of which realm is the current realm (e.g. if the global visibility bit is set). Hence, while in general the second comparison based on the realm identifiers does need to be passed, exactly what requirements are to be satisfied by the current realm's GRID may dependent on the visibility bits 272. By constructing realm identifiers of a child realm that include a bit portion which corresponds to the realm identifier of a parent realm which initialised the child realm, this enables such partial matches to be performed efficiently. In implementations which support variable allocations of different variable length bit portions of the global realm identifier to different generations of realms, the TLB entries 260 may also specify some information which identifies the positions of the boundaries between the different local RIDs which are concatenated to form the GRID (in order to allow a parent realm to be distinguished from a grandparent or earlier ancestor realm). This can enable the TLB control circuitry 280 to determine which parts of the realm identifier are to be masked. For other implementations, this may not be necessary (e.g. if any ancestor is permitted to access a memory region for which the owner realm has given visibility to its parent realm). Also, some implementations may have a fixed mapping with each global realm identifier having a fixed number of bits (such as 32-bits) and in this case it may not be necessary to provide any additional boundary defining data.

Figure 27 is a Venn diagram showing an example of architectural state accessible to the processing element 8 at different exception levels. Figure 27 shows an example based

on the ARM® architecture provided by ARM® Limited of Cambridge, UK, but other implementations may be based on other architectures, which may have different state associated with a given process. Also Figure 27 only shows a subset of the state accessible at each exception level, used as an example for conciseness, but it will be appreciated that other registers may also be accessible which are not shown in Figure 27.

When operating at exception level ELO the processing element 8 has access to a subset of architectural state labelled 350, e.g. including the following:

• General purpose registers, including integer registers, floating-point registers and/or vector registers, for storing general purpose data values during data processing operations.

• A program counter (PC) register, which stores a program instruction address representing the current point of execution within the program being executed.

• A saved processor state register (SPSR_EL0), which is used for storing information about the current state of a processor when an exception is taken from a process executed at exception level ELO. The SPSR may include information about the current processor state, e.g. a value PSTATE which represents the current processor mode at the time the exception occurred. The processor mode could specify the exception level as well as other information such as which instruction set was being executed. Also the SPSR register may include a realm entry flag R which is used to control realm entry as will be discussed below.

• An exception link register ELR EL0, which is used to store the current program counter value when an exception is taken, so that the ELR provides the return address to which processing should branch once the exception has been handled.

• A realm identifier register RID EL0 used to store the local RID of a child realm to which a realm entry is made (even though the exception level ELO is the lowest, least privileged, exception level, with the ability to create sub-realms as discussed below, a realm may still be entered from a process operating at the exception level ELO).

• An exception status register (ESR) used by ELO to store information about an exception that occurred (e.g. to allow selection of an appropriate exception handler). When the processor 8 is operating at exception level EL1 , it has access to a subset of state 352 which includes all of the state 350 which was accessible at exception level ELO but also includes additional architectural state. For example at EL1 there are banked versions of the SPSR, ELR, ESR and RID registers which are used by the process operating at EL1 for corresponding purposes to the ELO versions of these registers. The SPSR_EL1 register may in addition to the PSTATE and R values also include an intermediate realm flag Int which is used during nested realm entry and exit as will be discussed below. Although Figure 27 shows an example where the intermediate realm flag is stored in the SPSR, this is not essential and other examples could store the flag in a different register. Another example of state which may be accessible at EL1 but is inaccessible at ELO may be a translation table base register TTBR EL1 which provides an address indicating the base address of the stage one page tables 120-1 used by the MMU during page table walks.

Similarly, when executing instructions at exception level EL2, the processing element 8 has access to a subset of state 354 which includes all the states 352 accessible at EL1 , but also includes additional state, such as further banked versions of the SPSR, ELR and RID registers for EL2 and a virtual translation table base register VTTBR EL2 which provides the base address of the stage two translation tables 120-2.

Finally, when operating at EL3 the processing element has access to a subset of state 356 which includes all of the subset 354 accessible at EL2 but also may include other state such as a further realm identifier register RID EL3 used by processes operating at exception level EL3 to enter realms, and further exception handling registers ELR, SPSR, ESR similar to the corresponding registers used for lower exception levels. Again, Figure 27 is merely an example and other states may also be included in the relevant subsets accessible to particular exception levels.

Hence, each exception level is associated with a corresponding group of registers which is accessible to the processing circuitry when processing a software process at that exception level. For a given exception level other than the least privileged exception level, the group of registers accessible at the given exception level includes the group of registers accessible at a less privileged exception level than the given exception level. This hierarchy of state which is accessible to particular levels can be exploited to reduce the overhead associated with state saving and restoring upon realm entry an exit as will be discussed below.

On entry and exit to or from a realm, the processing element 8 and/or RMU 20 may need to perform a number of operations for ensuring secure handling of the realm entry or exit. For example, on entry to a realm a number of checks may need to be performed to check that the target realm is in the correct life cycle state (to avoid security measures being circumvented by attempting to enter a non-existent realm or a realm which has not undergone scrubbing of data from its owned pages, for example). Also, on exiting a realm, it may be desirable to mask the architectural state stored in registers of the processing element so that a higher privilege level process cannot access the state data used by the realm at a lower privilege level (which would otherwise enable the security measures provided by the realm protection to be circumvented). One approach for handling realm entry and exit could be to provide a dedicated realm entry or a realm exit instruction which triggers the RMU 20 to perform the relevant operations for entering or exiting a realm.

However, this may require significant modification of existing software in order to use the new instruction.

In the technique described below, the realm mechanism re-uses the mechanisms already provided for exception entry and return in order to enter and exit from realms. This reduces the amount of software modification required to support realm entry and exit, and simplifies the architecture and hardware. This is particularly useful because often the realm boundaries may correspond to exception level boundaries anyway, and even if a new instruction was provided to control entry and exit, a behavior for handling exceptions would still be required, so overall it can be less costly to extend the exception mechanisms to also control entry and exit.

Hence, an exception return (ERET) instruction, which would normally return processing from an exception processed in a current realm to another process also processed in the current realm (where the other process could be handled at the same exception level or a less privileged exception level than the exception), may be re-used to trigger realm entry from a current realm to a destination realm. In response to a first variant of the exception return instruction, the processing circuitry may switch processing from a current exception level to a less privileged exception level (without changing realm), while in response to a second variant of the exception return instruction, the processing circuitry may switch processing from a current realm to a destination realm, which could be operating at the same exception level or a lower (less privileged) exception level as the current realm. Using an exception return instruction to trigger realm entry can greatly simplify the architecture and hardware overhead as well as reducing software modification requirements to support use of the realms.

Another advantage of using an exception return instruction is that typically on returning from an exception, the processing circuitry may perform an atomic set of operations in response to an exception return instruction. The set of operations required on returning from an exception may be performed atomically so that these operations cannot be divided partway through, and so either the instruction fails and none of the atomic set of operations are performed or the instruction is successfully executed and all of the atomic set of operations are performed. For the second variant of the exception return instruction, the processing circuitry may similarly perform a second atomic set of operations, which may be different to the first atomic set of operations. The mechanisms already provided in a processor for ensuring that the exception return instruction completes atomically can be reused for realm entry in order to avoid a situation where the realm entry may only have partially have been performed which could lead to security vulnerabilities. For example the second atomic set of operations could include making realm execution context states available, changing the current realm being executed, and controlling branching to a

program counter address at which processing was previously being executed when the last time the same realm was executed.

The first and second variants of the exception return instruction may have the same instruction encoding. Hence, no modification of the exception return instruction itself is necessary in order to trigger the realm entry. This improves compatibility with legacy code. Whether a given exception return instruction is executed as the first variant or the second variant may depend on a control value it stored in a status register (e.g. first and second values of the control value may represent the first and second variants of the exception return instruction respectively). Hence, the current architectural state at the time that the exception return instruction is executed controls whether it returns the processor to a lower privilege level in the same realm, or triggers entry into a new realm.

This approach enables realm entry to be controlled with less software modification, especially as the value in the status register could be set automatically by hardware in response to certain events implying that a realm switch is likely (in addition to allowing voluntary setting of the control value in response to software instructions). For example, when an exception condition occurs which triggers an exit to a given realm, the processing circuitry could set the control value to the second value for the given realm so that a subsequent exception return instruction will automatically return processing to the realm in which the exception occurred, even if the exception handler code for handling the exception is identical to previous legacy code which was not written with realms in mind. Alternatively, in some architectures it may be expected that, when exiting from a realm the control value in the status register will still contain the second value which was set before triggering a realm entry to that realm, and so no explicit setting of the control value in the status register may be required.

In one example, the control value in the status register may be the R flag in the

SPSR register associated with the current exception level as discussed above. It can be useful to use the SPSR because this register would normally be used on exception return to provide the processor mode (including the exception level) and other information about how processing should continue upon return from the exception currently being processed. However, for a realm entry this information may instead be determined from the realm execution context (REC) and so the SPSR may not be needed. By re-using part of the SPSR for storing the R flag which controls whether an exception return instruction is treated as the first or second variant, this avoids the need to provide an extra register for storing this information. Hence, it can be useful to use a status register which, in response to the first variant of the ERET instruction, is used to determine return state information (such as the processing mode) for continuing exception at the less privileged exception level, but in response to the second variant of the exception return instruction, this return state

information would be determined from memory instead so that there is no need to access the status register itself. In particular, the status register used for storing the control value may be a status register associated with the current exception level from which the exception return instruction is executed.

As shown in Figure 27, at least one realm identifier register may be provided, and in response to the second variant of the exception return instruction the processing circuitry may identify the destination realm from a realm identifier stored in the realm identifier register. The realm identifier register may be banked, so that there are multiple realm identifier registers each associated with one of the exception levels, and in response to the second variant of the exception return instruction, the processing circuitry may identify the destination realm from the realm identifier stored in the realm identifier register associated with a current exception level. By using a realm identifier register to store the target realm identifier, there is no need to include this in the instruction encoding of the ERET instruction, which enables an existing format of ERET instruction to be used to trigger realm entry, reducing the amount of software modification required. The realm identifier in the realm identifier register may be a local realm identifier used by a parent realm to refer to its child realms, and so realm entry may be restricted to passing from a parent realm to a child realm, and it is not possible to go from a first realm to another realm which is not a direct child of the first realm. In response to the second variant of the exception return instruction, the processing circuitry may trigger a fault condition when a realm associated with the realm ID identified in the RID register is an invalid realm (a RID for which no realm descriptor has been defined or for which the realm descriptor defines a lifecycle state other than Active).

In response to the second variant of the exception return instruction the processing circuitry may restore architectural state associated with a thread to be processed in the destination realm from a realm execution context (REC) memory region specified for the exception return instruction. The state restoration can be performed immediately in response to the second variant of the exception return instruction (e.g. as part of the atomic set of operations), or could be performed later. For example the state restoration can be performed in a lazy manner so that the state required to commence processing in the destination realm could be restored right away (e.g. the program counter, processing mode information, etc.), but other state, such as general purpose registers, could be restored when required at a later time, or gradually in the background of ongoing processing of the new realm. Hence, the processing circuitry may commence processing of the destination realm before all the required architectural state has been restored from the REC memory region.

In response to the first variant of the exception return instruction the processing circuitry may branch to a program instruction address stored in a link register. For example this may be the ELR of Figure 27 which corresponds to the current exception level at which

the exception return instruction is executed. In contrast, for the second variant of the exception return instruction, the processing circuitry may branch to a program instruction address specified in the realm execution context (REC) memory region. Hence, as the link register would not be used for the second variant of the exception return instruction to directly identify any architectural state for the new realm, the link register can be re-used to instead provide the pointer to the REC memory region from which the new realm's architectural state is to be restored. This avoids the need to provide a further register in order for storing the REC pointer.

Hence, before executing an exception return instruction intended to cause a realm entry to a given realm, some additional instructions can be included in order to set the RID register to the realm identifier of the destination realm and to set the link register to store a pointer of a REC memory region associated with the destination realm. The REC pointer can be obtained by the parent realm from the realm descriptor of the destination realm.

In response to the second variant of the exception return instruction, a fault condition may be triggered by the processing circuitry when either the REC memory region is associated with an owner realm other than the destination realm or the REC memory region specified for the exception return instruction is invalid. The first check prevents a parent realm tricking a child realm into executing with processor state which it did not itself create, since only the memory regions owned by the child realm can store REC memory regions accessible on entering that realm (and as discussed above the REC memory regions will be set as RMU private). The second check of the validity of the REC memory region can be useful for ensuring that a REC memory region can be used only once to enter the realm and after that a subsequent attempt to enter the realm with the same REC data would be rejected. For example, each REC may have a life cycle state which may be either invalid or valid. In response to an exception occurring during processing of a given thread in a current realm, the architectural state of that thread may be saved to a corresponding REC memory region and that corresponding REC memory region may then transition from invalid to valid. In response to successful execution of the second variant of the exception return instruction, the REC memory region may then transition back from valid to invalid. This avoids a parent realm maliciously causing a child realm to behave incorrectly by specifying a pointer of an out of date REC memory region, a REC memory region associated with a different thread, or some other REC which is associated with a destination realm but is not the correct one that was used to store the architectural state upon a previous exit from the realm.

In a corresponding way, exit from a realm may re-use mechanisms provided for exception handling. Hence in response to an exception condition occurring during processing of a first realm which is unable to be handled by the first realm, the processing circuitry may trigger a realm exit to a parent realm that initialised the first realm. On

exception occurrence/realm exit, some additional operations may be performed which would not be performed for an exception occurring which can be handled within the same realm. For example this may include masking or scrubbing of architectural state and triggering of state storage to the REC as will be discussed in more detail below.

However, in some cases an exception may occur which cannot be handled by the parent realm of the first realm in which the exception occurred. Hence in this case it may be necessary to switch to a further ancestor realm beyond the parent. Although it might be possible to provide the ability to directly switch from a given realm to an ancestor realm which is more than one generation older, this may increase the complexity of the state registers required for handling exception entry and return or realm exit and entry.

Instead, a nested realm exit may be performed when the exception condition is to be processed at a target exception level with a greater privilege level than a most privileged exception level at which the parent realm of the first realm is allowed to be processed. The nested realm exit may include two or more successive realm exits from child realm to parent realm, until a second realm is reached that is allowed to be processed at the target exception level of the exception that occurred. Hence, by stepping up the realm hierarchy one level at a time, this can simplify the architecture. At each successive realm exit there may be operations performed to save a subset of processor state to the REC associated with the corresponding realm.

When the exception has been handled, then in response to an exception return instruction of the second variant which is executed in the second realm following the nested realm exit, the processing circuitry may then trigger a nested realm entry to return to the first realm. This can be handled in different ways. In some examples the hardware may trigger the nested realm entry itself without needing any instructions to be executed at any intermediate realms encountered between the first realm and the second realm during the nested realm exit. Alternatively, the hardware can be simplified by providing a nested realm entry process which returns back to each successive realm encountered in the nested realm exit, one level at a time and executes a further ERET instruction of the second variant at each intermediate realm. In this case in order to ensure that an intermediate realm triggers a return to the child realm of the intermediate realm from which the realm exit to the intermediate realm was made during the nested realm exit, an exception status register could be set to indicate that a predetermined type of exception condition occurred in the child realm. For example, a new type of exception condition (e.g. "spurious realm exit") could be defined to handle this intermediate realm case. Hence, when reaching an intermediate realm the processor could then resume processing within the intermediate realm from a program instruction address corresponding to an exception handling routine for handling the predetermined type of exception condition. This exception handling routine

could for example simply determine that the child realm exited for some unknown reason and could then choose to execute another exception return instruction of the second variant to return processing to the further child realm. By doing this at each intermediate realm eventually the original first realm at which the original exception occurred may resume processing.

During this nested realm entry and exit procedure, an intermediate realm flag within a status register can be used to flag which realms are intermediate realms, to either trigger the hardware-triggered immediate realm entry to the relevant child realm or to trigger the setting of the exception status information which will then trigger an exception handler or other code within the intermediate realm to return to the child realm. For example the intermediate realm flag could be the Int flag in the relevant SPSR as discussed in Figure 27.

Figure 28 is a flow diagram illustrating a method of handling realm entry or exception return. At step 400 an exception return (ERET) instruction is executed when the current exception level is ELx. ELx can be any of the exception levels supported by the processing circuitry. Although one might not expect an exception return to occur from the least privileged exception level ELO, the ability to create sub-realms as will be discussed below means that there could still be an ERET instruction executed from ELO in order to trigger entry to the sub-realm also executed at ELO. Also, some types of exception may be able to be handled at the same exception level as the level at which the exception occurred, in which case exception returns could still occur from ELO.

At step 402 the processing element determines the current value of the realm flag R in the SPSR associated with exception level ELx. If the realm flag R is zero then this indicates a conventional exception return, without entering a different realm. At step 404 the processing element 8 determines the target exception level at which to operate following the exception return, based on the PSTATE value within SPSR_ELx and determines the program counter value to branch to from the link register ELR ELx. The new target exception level and program counter values and any other return state associated with the return to processing following the exception are restored to the relevant architectural state registers associated with the target exception level (which would often be a lower exception level which is less privileged than exception level ELx, but could also be the same exception level ELx). Such return state operations are performed atomically. At step 406 processing then resumes at the target exception level, but staying within the same realm as the realm in which the ERET instruction was executed.

If at step 402 the realm flag R was set to 1 , then this indicates realm entry, and so this triggers a second atomic set of operations different to the set performed for a conventional exception return. At step 408 the processing element triggers the realm management unit to carry out a number of realm checks. These include checking that:

• The local RID indicated in the realm identifier register RID ELx associated with exception level ELx indicates a valid child realm. That is, the RMU checks the realm descriptor accessed from the realm descriptor tree 360 for the specified child realm, and checks whether the life cycle state of the child realm's realm descriptor indicates the Active state. If the child realm is in any state other than the active state then the realm check is unsuccessful.

• The RMU 20 also checks that the REC memory region indicated by the pointer in the link register ELR ELx is a memory region which is owned by the child realm indicated in the realm ID register RID ELx. That is, the RMU 20 accesses the realm granule table 128 (or cached information from the RGT 128), locates the relevant entry corresponding to the memory region indicated in the REC pointer, and checks the owner realm specified for that memory region. The owner realm indicated in the ownership table may be specified as a global RID, and this can be compared with the global RID specified in the realm descriptor of the target child realm to determine whether the child realm is the valid owner of the REC. If the REC is owned by any realm other than the specified child realm in the RID register, then this check is unsuccessful.

• The RMU 20 also checks whether the status of the REC memory region defined in ELR ELx is valid. There are different ways in which the validity of REC memory regions can be represented. For example each REC memory region may include a flag specifying whether it is valid. Alternatively, a separate table could define the validity of RECs stored in other memory regions. A REC may be valid if it has been used to store architectural state of the relevant realm on a previous exception exit, but has not yet been used to restore state following a return from an exception. If the REC is invalid then again the realm check is unsuccessful.

• The RMU 20 also checks whether a flush command has been executed since the last exit from any child realm other than the child realm indicated in the RID register RID ELx. The flush command will be discussed in more detail below, but is a command for ensuring that any state still to be saved to a child realm's REC is pushed to memory (this helps to support lazy state saving approaches). If no flush command has been executed and the system is attempting to enter a different child realm to a child realm that was previously exited, then there is a danger that there could still be state left within the processor registers which has not yet been pushed to memory. Enforcing the use of the flush commands ensures that a different child realm can safely be entered without loss (or leakage to other realms) of a previous child realm's state. There may be a number of ways of identifying whether a flush

command has been executed. For example, some status flags may be used to track (a) whether there has been a change to the RID register RID ELx since the last realm exit, and (b) whether a flush command has been executed since the last realm exit. If there has been a change to the RID register and no flush command has been executed since the previous realm exit then this may cause the realm check to be unsuccessful.

If any of the realm checks is unsuccessful, then at step 409 a fault is triggered and the system stays within the current realm associated with the ERET instruction. Hence, it is not possible to reach a child realm unless all of the realm checks are successful.

If all of the realm checks are successful then at step 410 the processing element switches to processing in the child realm indicated in realm ID register RID ELx. For example, the processor may have an internal register which specifies the global RID of the current realm (this internal register may not be visible to software and is different to the banked RID registers shown in Figure 27). The switch to the child realm can be enacted by writing the global RID of the new destination realm to the internal RID register.

At step 412, state associated with the new realm is made available based on the states saved in memory in the REC memory region indicated by the pointer in the ELR ELx register. As the REC region is owned by the new child realm, it is now accessible and so return state information such as the program counter and target exception level can be obtained from the REC. At this point, a selected subset of architectural states may be restored from the REC, or alternatively the architectural state may be restored lazily so that processing may commence without fully restoring all the state, and then the state can be restored as and when required, or gradually over a period of time in order to improve performance by reducing the delay before the processing can resume from the new realm.

At step 414 it is determined whether intermediate realm flag Int is set in the SPSR associated with the new realm. The SPSR contents would be restored from the REC along with the rest of the architectural state. If the intermediate realm flag is not set, then this indicates that the new realm is the realm in which the original exception occurred (or the realm is being entered for the first time without any previous exception occurring in that realm), and so there is no need to trigger any further realm entry into a child realm. At step 416 the program counter is obtained from the REC and then at step 418 processing continues in the new realm at the target exception level obtained from the REC.

Alternatively, if the intermediate realm flag is set then this indicates that a nested realm exit has previously occurred, and the nested realm entry has reached an intermediate realm. Hence there is a need to return processing to a further child realm in order to return to the realm in which the exception originally occurred. There are two alternative techniques for handling this. In a first alternative, at step 420 a spurious realm exit exception is taken, and so the exception status register associated with the new realm may be set by the processing element to a status code associated with this type of exception, and then processing may branch to an exception vector associated with that type of exception which triggers an exception handler to be processed. The exception handler need not perform any actual processing but could simply determine that an unknown type of exception occurred in a given child realm and so could then trigger another ERET instruction to be executed at step 422. The RID and ELR registers associated with the intermediate realm may still have the values which were placed in those registers when the intermediate realm previously entered the further child realm, and so execution of the ERET instruction may then trigger a further realm entry into the further child realm. The method may return to step 408 to check whether the realm checks are successful for that further realm entry and then the method continues in a similar way to the previous realm entry in the nested process.

Alternatively, instead of handling the nested realm entry using another ERET instruction executed in the intermediate realm, at step 424 hardware may detect that the intermediate realm flag is set for the current realm and then may trigger a further realm entry to the child realm without needing to execute any instructions within the intermediate realm and then the method may then return to step 408.

Figure 29 shows a flow diagram illustrating a method of exiting from a realm or taking an exception. At step 430 an exception occurs within a given exception level ELx which targets an exception level ELy (ELy≥ ELx). The target exception level ELy is the exception level at which the exception is to be handled. The target exception level could be the same as ELx, just one exception level higher than ELx, or could be several exception levels higher.

At step 432, the RMU determines whether the target exception level ELy is greater than the boundary exception level (BEL) of the current realm, which can be read from the realm descriptor of the current realm, and whether the current Realm is a sub-realm (see discussion of Sub-Realms below - the Type field of the Realm descriptor shown in Figure 16 specifies whether the Realm is a sub-Realm). If the target exception level ELy is not greater than the boundary exception level then this indicates that the exception can be handled within the current realm and so if the current Realm is a full Realm there is no need to trigger a realm exit. In this case, at step 434 the processing element switches the current exception level to ELy (by updating the current state register for example), or if ELx = ELy then the current exception level remains the same (either way, the current exception level now is ELy). At step 436 the current processing state associated with exception level ELx is saved to registers accessible to target exception level ELy. For example the program counter may be saved to the link register ELR ELy, and the previous exception level ELx at which the exception occurred may be indicated in the PSTATE value in SPSR_ELy. Also, the realm flag R in SPSR_ELy register may be set to 0 to indicate that the exception came from within the same realm, and so any subsequent exception return need not trigger a realm entry. Information about the exception type which occurred can also be saved to an exception status register ESR_ELy. At step 438 the processing element branches to the exception vector associated with the specified exception type and then processing continues in order to handle the exception. Eventually, when the exception handling is complete, an ERET instruction will be executed to trigger a return to the previous processing as discussed above for Figure 28.

On the other hand, if at step 432 the target exception level ELy was greater than the BEL of the current realm, or if the current realm is a Sub-Realm (for which any exception triggers an exit to the Sub-Realm's parent Realm), then a realm exit is required in order to handle the exception. At step 440 it is determined whether the exception corresponds to a voluntary realm exit. Some types of exception may be triggered by unexpected events, such as the user pressing a button on the processing device, or some fault occurring. However, it is also possible for a realm to voluntarily relinquish processing and return to a parent realm. For example, the child realm may have reached the end of some processing routine, or may need to invoke some functionality at a higher exception level. When a child realm deliberately exits to a parent realm, then to allow the child realm to share data with the parent realm, it can be useful to allow some architectural state of the child realm to be retained for access from the parent realm, so that less state is masked and scrubbed compared to an involuntary realm exit. For example, general purpose registers may not be masked and scrubbed, although other control registers could still be masked and scrubbed. This avoids the need for the data to be stored to global visibility memory regions as the state can then be accessed directly from the registers, improving performance. That is, for a voluntary realm exit, the child realm is in control of the realm exit and so it can be assumed that the child realm has already overwritten the visible architectural state which should not be made accessible to the parent realm, and left any state which it does desire to share with the parent realm in the registers. A voluntary realm exit could be triggered for example by a predetermined variant of an exception triggering instruction which corresponds to a particular type of exception for which no masking or scrubbing is performed.

If the realm exit is not a voluntary realm exit, then at step 442 masking of a selected subset of architectural states is performed. The masking ensures that the architectural state to be hidden is made inaccessible to the parent realm. However, the parent realm may still attempt to access the registers which store that architectural state, and so also a scrubbing operation may be performed which ensures that any subsequent access to a scrubbed architectural register will return a predetermined value regardless of what value is actually stored in the physical register associated with that architectural register. That is, when the register scrubbing operation is performed for a given architectural register which stores part of the masked subset of architectural states this ensures that a subsequent read access to the given architectural register will return a predetermined value (e.g. zero) if it is performed by the processing circuitry without any intervening write access to that architectural register between the realms switch and the subsequent read access. The register scrubbing operation can be performed in different ways. For example, a physical register which corresponds to the given architectural register could be set to the predetermined value (e.g. a value of this zero could actually be written to a physical register). Alternatively, the register scrubbing could be performed by register renaming so that the given architectural register is remapped from a first physical register to a second physical register. Another approach could be to set a status value associated with the given architectural register or a physical register mapped to the given architectural register to indicate that a read access to the given architectural registers should return the predetermined value. With these last two approaches, it is possible to retain the masked architectural state within the physical register file of the processing element even though it is not accessible from the new realm, and any access to that state will return a predetermined value. This is useful for supporting lazy state saving as discussed below. Once the masking and scrubbing operations have been performed, at step 444 the processing element triggers saving of the masked subset of architectural states to the REC memory region associated with the current realm in which the exception occurred. This saving could be done immediately, or could be done lazily, overlapped with subsequent processing in the parent realm. Which approach is taken may depend on the particular implementation of the architecture and so is not an essential feature of the architecture. However, lazy state saving can improve performance.

If the realm exit was a voluntary realm exit, then at step 443 a reduced subset of architectural state is masked and scrubbed in the same way as at step 442, except that at least some of the state which would be masked/scrubbed for an involuntary realm exit does not need to be masked/scrubbed for a voluntary realm exit. For example, the reduced subset of architectural state may exclude the general purpose registers. Following step 443, the method continues to step 444 as discussed above for an involuntary realm exit.

Regardless of whether the realm exit was voluntary or involuntary, at step 446 the processing element switches to the parent realm. Since the RID ELx and ELR ELx registers associated with the child realm exception level and controlled by the parent realm exception level will still be set to the RID and REC pointer of the previously exited child realm (since these would have been set before an ERET instruction was executed to enter the child realm in the first place), there is no need to adapt this state, and so any subsequent ERET instruction will return to the same child realm as before. There is only a need to update these registers if the parent realm wishes to switch to a different child realm from the child realm which was last exited. Similarly, the realm flag R in the SPSR of the parent

realm may still be set to 1 following the previous realm entry, and so can retain this value so that an ERET instruction subsequently will be treated as an entry to a new realm unless the R flag is cleared to 0 by an instruction executed in the parent realm.

At step 448 the RMU determines whether the target exception level ELy is greater than the BEL of the new parent realm into which processing has been switched. If not, then this indicates that the exception can be taken in the parent realm and at step 450 the processing element 8 branches to the corresponding exception handler appropriate for handling the type of exception that occurred. If the target exception level is higher than the BEL of the parent realm, then a nested realm exit is required in order to branch to a further ancestor realm and so at step 452 the intermediate realm flag associated with the parent realm is set to 1 in the SPSR and then the method returns to step 442 to trigger the masking and scrubbing of architectural states associated with that realm (this time the realm exit is not a voluntary realm exit since it has been triggered outside the control of the parent realm). The method then loops around again to mask or scrub any appropriate state associated with the current realm and then switch to the parent realm of that realm. The nested realm exit can loop through several times until eventually a realm is reached which can handle the exception.

At steps 412, 442, 444, the subset of architectural state which is masked, scrubbed and saved on exit from a realm, and restored on entry to the realm, may depend on the boundary exception level of the exited/entered realm. For non-nested realm exit and entry, the selected subset may include the subset of architectural state which is accessible to the exception level in which the exception occurred, as indicated in Figure 27 for example. Hence, each exception level is associated with a group of registers, and in response to a realm-exiting exception condition occurring during processing of a given realm which is to be handled by an exception handler at a more privileged exception level than the boundary exception level of the given realm, the subset of registers may be selected depending on the boundary exception level of the given realm, and may include those registers which are accessible to the processing circuitry when processing a software process at that exception level. If the exception occurs at EL1 or higher than this may include all of the registers accessible to that exception level as well as any lower exception levels for a non-nested realm exit. Similarly, on returning to a realm then the corresponding subset of architectural states may be restored.

However, for a nested realm exit, it may be assumed that for an intermediate realm any registers which are accessible at lower exception levels than the boundary exception level of the intermediate realm would have been modified by a child realm at a lower exception level since the intermediate realm triggered entry to the child realm. Therefore, there may be no need to save such registers accessible at a lower exception level to the

REC associated with the intermediate realm during the nested realm exit (no further execution at the intermediate realm has taken place since previously entering the child realm). Conversely, during the nested realm entry there is no need to restore those registers accessible at lower levels during the passage through intermediate realm, as these registers will subsequently be restored by a realm at a lower exception level. Instead the intermediate realm state saving and restoring could simply include the registers accessible to the intermediate realm's boundary exception level, but which are not accessible at lower exception levels. For example an intermediate realm at EL1 , the state saved/restored in nested realm exit/entry may include the subset 352 in Figure 27 but may exclude the subset 350 accessible at EL0. If the intermediate realm is at EL2, the saved/restored subset of state during the nested realm exit/entry may include the subset 354 accessible at EL2 but exclude the subset 352 accessible at EL1 . With this approach, the amount of state saving and restoration required on passing through an intermediate realm can be reduced to improve performance.

Hence, when the exception condition is to be processed by an exception handler at a target exception level which is more privileged than the boundary exception level of the parent realm of the given realm in which the exception occurred, a nested realm exit can be triggered which includes multiple successive realm exits from child realm to parent realm until the target realm is reached which has a boundary exception level corresponding to the target exception level or higher. A respective state masking process (and state saving) may be triggered for each of the successive realm exits and each respective state masking process may mask (and save) the corresponding subset of registers selected based on the boundary exception level. For a realm exit from a given child realm which has a boundary exception level other than the least privileged exception level, the corresponding subset of registers masked/saved during the nested realm exit may include at least one register accessible at the boundary exception level of the given child realm, but may exclude at least one register accessible to the processing circuitry at a less privileged exception level than the boundary exception level of the given child realm (as it may be assumed that such registers would already have been saved when exiting a realm at that less privileged exception level). This reduces the amount of state masking and saving operations required.

Similarly, on realm entry (exception return) the intermediate realm flag may be used to determine whether the realm entered is an intermediate realm. If the intermediate realm status value for a realm having a boundary exception level other than the least privileged exception level is set to a predetermined value (indicating an intermediate realm) then the subset of registers to be restored upon realm entry may include at least one register accessible at the boundary exception level of the intermediate realm, but may exclude at least one register accessible to the processing circuitry at a less privileged exception level than the boundary exception level of the particular realm. If the intermediate status value is set to a value other than the predetermined value, then this realm being entered is the final realm and so the subset of registers to be restored accessed may include all registers accessible to the processing circuitry at the boundary exception level of the particular realm (without excluding any registers from lower levels).

In this way, the state saving and restoring operations during nested realm exit and entry can be made more efficient.

Figure 30 illustrates an example of a non-nested realm entry and exit. In this example, the parent realm is A operating at exception level EL1 and it wishes to enter a child realm B having a BEL of EL0. Clearly, similar realm entry and exit procedures can be performed at other exception levels. At step 460 the parent realm sets the realm identifier register RID EL1 to the local realm ID of child realm B. The local RID has the format obtained by concatenating the tree indices as discussed above. At step 462 the parent realm at EL1 sets the link register ELR EL1 to an address providing the pointer to realm B's realm execution context. The parent realm obtains this pointer from the realm descriptor of the desired child realm B. At step 464 the parent realm sets the realm flag R in SPSR_EL1 to 1 . At step 466 the parent realm then executes an ERET instruction, which as the realm flag is set to 1 is interpreted as triggering a realm entry. The various realm checks described above for step 408 of Figure 28 are performed, and if they are successful then the realm entry is performed to switch to realm B. At step 468 state is restored from the REC indicated in the link register. This state includes the target exception level in which the new realm is to be executed. Note that at the point when the REC is accessed, the processor is still in the previous exception level associated with realm A and so this is why the link register ELR EL1 is still accessible to obtain the REC pointer, even though the switch to realm B has been made and eventually that Realm will resume processing from EL0. The state restoration can be done immediately or can be done lazily over a period of time in parallel with the regular processing 470 of realm B.

When an exception occurs at step 472 during execution of the child realm B, then a set of masking operations is performed to hide state associated with realm B from its parent realm. This includes masking and scrubbing of at least a subset of the architectural states associated with EL0 at step 474 (the subset of architectural state that is masked/scrubbed may depend on whether the exit is a voluntary or involuntary realm exit). The masking makes the state inaccessible and the scrubbing ensures that subsequent accesses to the corresponding registers from the parent realm will trigger a predetermined value to be returned. At step 476 state saving to the REC associated with realm B is performed for the masked subset of architectural state. In this case, since the child realm has a BEL of EL0, the masked subset of state includes at least the subset 350 accessible at EL0. Again the

state saving can be triggered immediately or could be done in a lazy fashion, in parallel with ongoing processing with the exception at the parent realm. At step 478 the realm exit is then triggered to switch to the parent realm A and at step 480 processing of the exception is carried out within the parent realm. Steps 474 and 476 of Figure 30 can be omitted for a voluntary realm exit.

Figure 31 shows a similar example showing a nested realm entry and exit. A grandparent realm A at exception level EL2 executes an ERET instruction at step 500, having already set the link register to point to the desired child realm's REC, set the realm ID register RID EL2 to the child realm's local RID, and set the R flag to 1 in SPSR_EL2. This triggers a realm entry to realm B at step 502, and this triggers state to be restored from realm B's REC at steps 504 which again can be done immediately or lazily. Eventually realm B then executes a further ERET instruction at step 506, again having set the link register, RID register and R flag, to trigger a further realm entry at step 508. Note that the realm entries at steps 502 and 508 may be performed as shown in Figure 28 and so steps such as the realm checks are still performed, but not shown in Figure 31 for conciseness. Following entry to realm C then realm C's architectural state is restored from its REC at step 510 and processing resumes within realm C at step 512.

Subsequently, an exception occurs at step 514. The exception targets exception level EL2. For example, the exception may be a type of exception to be handled by the hypervisor (such as an event associated with a virtualized device or a stage 2 address translation fault). At step 516 the state masking process, including any scrubbing and state saving, is triggered in a similar way as shown in Figure 30, and the realm exit to realm B, which is the parent realm of realm C in which the exception occurred, is performed at step 518.

At step 520 the processing element detects that the target exception level for the exception which occurred is higher than the boundary exception level of realm B, and so realm B is an intermediate realm and a further realm exit to realm B's parent realm is required. Hence the intermediate realm flag is set within SPSR_EL1 at step 522 and at step 524 further state masking, scrubbing and saving of a subset of registers is performed. Again, the saving part of this process could be deferred if the corresponding state can be retained in the process of registers in some way which is inaccessible to the processor following the realm exit to Realm A. The subset of registers which are saved for this intermediate realm include the registers 352 which are accessible at EL1 but excludes the registers 350 which are accessible at EL0. There is no need to save subset of registers 350 because Realm B has not executed any instructions to change these registers since the previous realm entry at step 508 and these registers have already been

masked/scrubbed/saved by Realm C. Hence, this reduces the amount of additional masking and state saving required for the intermediate realm.

A further realm exit 526 is then triggered to the realm A which is the parent of the intermediate realm. In this case, realm A is able to handle the exception as it can operate at EL2, and so at step 528 the exception is trapped and handled at exception level EL2. The exception is processed by the hypervisor and once the exception has been handled an ERET instruction 530 is executed to return to the previous processing. At this point, the values of ELR EL2 and RID EL2 and the R flag in SPSR_EL2 are still the same as they were before the ERET instruction was executed at step 500. Therefore it is not necessary to set these values again. The ERET instruction 530 triggers a realm entry back to the previous child realm B at exception level EL1 . Note that the exception handler code may already have such an ERET instruction in legacy code for returning processing to a lower exception level following handling of the exception. Hence, the re-use of ERET instructions to trigger realm entry as discussed above enables existing exception handler code written without awareness of the realm protection functionality to continue to be used, reducing platform development costs.

On returning to Realm B at EL1 , the processing element detects that the intermediate flag is set in SPSR_EL1 (step 532). Hence, at step 534 the state which is restored from realm B's REC is the state accessible to EL1 (registered 352) but excludes the registers 350 which are accessible at EL0. If the processor implementation uses the software-assisted method of triggered nested realm entry, then at step 536 the processing element sets the exception status register ESR_EL1 for the intermediate realm to a particular value corresponding to a spurious realm exit. The processing element sets the program counter to an exception vector which indicates the exception handler for handling this predetermined type of spurious realm exit exception, and then the exception handler is executed at step 538. The exception handler need not carry out any real functionality but could simply trigger an ERET instruction or alternatively could provide some other operations as well. Eventually an ERET instruction is executed at step 542 which, since the ELR EL1 , RID EL1 and SPSR_EL1 registers are still set as they were before the ERET instruction was executed at step 506, triggers a realm entry 544 back to the previous child realm which was executing before the original exception occurred. At this point state is restored from realm C's REC at step 546, and processing continues at step 548 in the same way as at steps 510 and 512 above.

Alternatively for a hardware assisted nested realm entry, steps 536 to 542 can be omitted and instead having restored the required subset of state at step 534, the hardware of the processing element may detect from the intermediate realm flag that further realm entry is required to realm C and so at step 550 the hardware may directly chain a subsequent

realm entry without requiring a further ERET instruction 542 to be executed. In this case there is no need for any instructions to be executed within the intermediate realm.

By using such a nested realm entry and exit procedure, this avoids the realm at EL2 needing to deal with any realm values or REC pointers associated with ELO. This allows a cleaner and simpler check when entering a realm, as only the parameters of an immediate child of the current realm need to be checked. This greatly simplifies the architecture and hardware implementation.

Figures 32 and 33 show examples of lazy state saving and state restoration to and from the REC on realm exit and realm entry respectively. In general, on exiting a realm to a parent realm it may be desirable to mask state associated with the child realm to hide it from the parent realm, and to perform scrubbing to ensure that the parent realm will see some predetermined value if it attempts to access an architectural register corresponding to scrubbed state. These operations can be performed relatively quickly. However, if there is insufficient space in the physical register file of the processing circuitry for retaining the child realm state indefinitely, then it may be desired to save some of that data to the REC. However, this can take a longer time, and occupies memory bandwidth which could otherwise be used for processing in the parent realm, which can delay processing in the parent realm. Similarly, the corresponding operations to restore state from memory to the registers on entering a realm can take some time. Therefore, for performance reasons it may be desirable to support asynchronous saving/restoring of processing element state to/from the REC. Whether a given processor implementation actually does this lazy state saving is an implementation choice for the particular processor. For example, some processors not aimed at high performance may find it simpler simply to trigger the state saving operation right away, in order to reduce to complexity of managing which state has already been saved and which has not. However, to enable performance improvements if desired, it can be desirable to provide an architecture functionality supporting such asynchronous, lazy state saving, approaches.

Hence, in response to a realm switch from a source realm to a target realm which is to be processed at a more privileged exception level than the source realm, the processing circuitry may perform state masking to make a subset of architectural state data associated with the source realm inaccessible to the target realm. Although it is possible for that masked subset of state to be saved to memory at this point this is not essential. However, the architecture provides a flush command which can be used following a realm switch. When the flush command is executed, the processing circuitry ensures that any of the masked subset of architectural state data not already saved to at least one REC memory region owned by the source realm if saved to the at least one REC memory region). By providing such a flush command, it can be ensured that at a point when it is necessary to

ensure that the subset of architectural state data has definitely been saved, this can be forced through, and this gives the freedom for particular micro-architectural implementations of the architecture to vary exactly when that subset of architectural state data is actually saved to memory in cases when the flush command has not yet been executed.

In addition to the state masking, on a realm switch the processing circuitry may also perform a register scrubbing operation as discussed above which ensures that any subsequent read access to a given architectural register returns the predetermined value (if performed without an intervening write access). This scrubbing could be done by actually writing the predetermined value to a physical register corresponding to the given architectural register, or by register renaming, or by setting other control status values associated with the given architectural register to indicate that a read access should return the predetermined value instead of the actual contents of the corresponding physical register. If the state saving to be performed in response to the realm switch is performed asynchronously, then the processing circuitry may commence processing of the target realm when at least part of the subset of architectural state data made inaccessible in response to the realm switch is still stored in the registers of the processing circuitry. For example, a processor may have a larger physical register file than the number of registers provided as architectural registers in the Instruction Set Architecture, and so some spare physical registers could be used to retain the previously masked states for a period after the target realm has already started processing. This is advantageous because if then the processing returns to the source realm when a given item of the subset of architectural state data is still stored in the registers, the processing circuitry can simply restore access to that given item of architectural state to from the register file, without needing to restore data from the REC. Some types of exceptions may only need a relatively short exception handler to be executed, in which case it is likely that some masked state is still resident in the register file on returning from the exception. Such "shallow" exception entry/return events can benefit from using lazy state saving.

If lazy state saving is used, then once processing of the target realm has commenced following the exception, the processing circuitry may trigger saving of a given item of the REC region(s) in response to occurrence of a predetermined event other than the flush command. Although processing at this point has already switched to the parent realm (which would not usually have access to the REC associated with the previous child realm), as these operations are triggered in hardware by the microarchitecture implementation rather than by software, they may not be subject to the same ownership checks that would be required for a general software-trigger memory accesses (effectively these REC saving operations will already have been authorized by the child realm before exiting).

Many different types of predetermined event could be used to trigger a certain item of subset of architectural state data to be saved to REC including the following:

• A register access to an architectural register corresponding to the given item of the subset of architectural state data. This approach could be useful for less complex processors which do not support register renaming. In this case each architectural register may map to one fixed physical register, and so the first time that code associated with the parent realm tries to access the given architectural register then this may require the old value of that register used by the child realm to be saved to memory.

• Remapping of the physical registers storing the given item of the subset of architectural state data. In systems supporting register renaming, the architectural state could remain in the register file for longer, but eventually it may be necessary to remap the corresponding physical register to store a different value and at this point the corresponding architectural state of the child realm could be saved to the REC.

• A number of available physical registers becoming less than or equal to a predetermined threshold. In this case, rather than waiting for an actual remapping of a given physical register, state saving could start to be pre-emptively performed once the number of free physical registers (which are available for reallocation to a different architectural register) becomes low.

• Elapse of a given number of cycles or a given period of time. Hence it is not necessary for any particular processing event to trigger saving but instead the lazy state saving could simply spread out saving of the child realm's context to the REC over a period of time in order, to reduce the impacts on memory bandwidth available for other memory accesses triggered by processing in the parent realm.

• An event indicative of reduced processor workload, for example a period of idle processor time or some other event which indicates that performing the state saving now would have less impact on the overall performance of the processing of the parent realm. At this point the saving of at least part of the subset of architectural state data could be triggered.

Following the realm switch, if the processing circuitry attempts to enter a further realm other than the source realm from which the realm previously switched to the parent realm, then the processing circuitry may reject a realm entry request when the further realm is to be processed at the same exception level or a less privileged exception level than the target realm of the previous realm exit and no flush command has been received between the realm switch and the realm entry request. Alternatively, the realm entry request could be accepted regardless of whether the flush command has been executed, but if the flush

command has not been executed then the initial child realm REC state may be corrupted so that the REC cannot be used again, preventing valid entry into the child realm. Either way, a flush command is required before a parent realm can successfully direct processing to a different child realm to the one that was previously executed. This ensures that even if the hardware chooses to use the lazy state saving approach, all the necessary state associated with the previous child realm will have been committed to being saved to memory at the point where different child realm is entered. This avoids the need to back up several sets of child realm data to be saved to memory and simplifies the architecture.

Note that the flush command merely needs to ensure that state from the masked registers is committed to being stored to the REC memory region. The store operations triggered by the flush command can be queued within a load/store queue of the processing element 8, or in queues within the interconnect 14, a memory controller or the memory 16 itself, so the actual write to the memory cells may not occur until later, but from the processing element's point of view it is known that the memory writes of the masked state will definitely happen.

The flush command could be a native instruction supported by an instruction decoder of the processing circuitry. Alternatively, the flush command could be a command triggered by a predetermined event to carry on processing of the instructions decoded by the instruction decoder. For example, a flush command could be triggered automatically by some other type of instruction which implies that it should be ensured that state saving operations have been triggered to memory for all the required subset of architectural state associated with the previous child realm.

As discussed above, the particular subset of architectural state which is to be saved during the realm switch may depend on the boundary exception level associated with the source realm (and also may depend on whether the source realm is an intermediate realm in a nested realm exit). The state masking and saving operations may be suppressed if the realm switch is a predetermined type of realm switch (e.g. a realm switch triggered by execution of a voluntary realm switching instruction in the source realm).

Hence, Figure 32 shows an example of lazy state saving and restoring. At step 560 a realm exit is handled in the same way as discussed above. Although the subset of architectural state to be hidden for realm B is masked and scrubbed on the realm exit, the state saving to the REC can be deferred and so a lazy state saving process 562 may be performed in the background of the processing 564 associated with the parent realm A. The particular way in which the lazy state saving is performed may be an implementation choice for a particular processing design. At step 566 the parent realm triggers a realm entry back to the previous child realm (using a ERET instruction as discussed above). In this case, since the realm entry goes back to the same child that was previously exited then there is no need for a flush command for the realm entry to be valid. If a portion 568 of the lazy state saving operations is still to be performed then this can be cancelled following the realm entry, and instead the corresponding state values for realm B can simply be restored from some of the physical registers of the processing element. Hence, for a relatively shallow exception exit and return, using the lazy state saving approach helps to reduce the volume of memory accesses required improving performance.

Figure 33 shows another example in which a realm exit 560 is performed from a realm B1 to its parent realm A in the same way as in Figure 32. However, this time, rather than returning to the same realm, realm A wishes to switch processing to a different child realm B2. Hence, the parent realm executes a flush command at step 570 which triggers the processing element to ensure that any remaining part of the lazy state saving process 562 is completed (i.e. store operations are issued for all the remaining architectural state of the masked subset which is still resident in the processing element's register file). At step 572, the parent realm sets the realm identifier register and link register to indicate the realm ID and REC pointer for the target realm B2, and then executes an ERET instruction at step 574 which triggers a realm entry 566 to realm B2. If the flush command had not been executed prior to executing the ERET instruction, the realm entry would have failed. At step 578 state restoration is performed from realm B2's REC memory region (again this can be done lazily).

Hence, the use of the flush command enables the fast exception exit and trickling out of processor states into the REC of the previously exited realm and also allows for shallow exception exits and returns where state is held within the registers of the processing elements and not stored and reloaded from the REC.

Figure 34 illustrates the concept of sub-realms which can be initialized by a parent realm. As shown in Figure 34, a given parent realm 600 operating at a particular exception level may initialize a sub-realm 602 which operates at the same exception level as its parent. The full realm 600 corresponds to a given software process (or collection of two or more processes) but the sub-realm corresponds to a predetermined address range within the given software process. As the full realm is the parent of the sub-realm, then as discussed above the sub-realm may have a right to access data stored in memory regions owned by the parent full realm, but the sub-realm may have a right to exclude its parent full realm from accessing data stored in a memory region owned by the sub-realm 602. This is useful for allowing a certain portion of a given software process to be made more secure than other parts of the software process. For example a portion of code for checking passwords in a mobile banking application or for processing other sensitive information could be allocated to a sub-realm in order to prevent other portions of the same application or operating system from accessing that sensitive information.

Sub-realms may generally be handled in the same way as full realms, with some differences as explained below. Entry and exit to and from sub-realms may be handled using exception return instructions and exception events in the same way as discussed above. Hence, the sub-realm may have a child realm ID constructed in the same way as for full child realms of the same parent, and may be provided with a realm descriptor within the realm descriptor tree as discussed above. Entry to the sub-realm can be triggered simply by executing the ERET instruction having placed the appropriate child sub-realms RID in the RID register before executing the ERET instruction. Hence, the same type of ERET instruction (of the second variant) can be used to trigger entry to either a full realm or a sub-realm.

One way in which sub-realms may be different to full realms may be that sub-realms may not be allowed to initialize their own child realms. Hence a realm initialization command for initializing a new realm may be rejected if the current realm is a sub-realm. The RMU may determine whether the current realm is a full realm or a sub-realm using the realm type value in the realm descriptor of the current realm. By disabling realm initialization when currently in a sub-realm, this simplifies the architecture since it is not necessary to provide additional state registers for use by the sub-realm in initializing further realms.

Similarly, execution of a realm entry instruction may be disabled when currently in a sub-realm. This simplifies the architecture because it means that some of the banked registers such as the ELR, SPSR, ESR and RID registers discussed above which are used for handling realm entry and exit (and exception entry and return) do not need to be banked a further time for each sub-realm, which would be difficult to manage as it may not be known at design time how many sub-realms a given process will create. Similarly, exception return events which trigger a switch to a process operating at a lower privilege level may be disabled when a current realm is a sub-realm rather than a full realm. Although in the examples discussed above a single type of ERET instruction functions both as a realm entry instruction and an exception return instruction, this is not essential for all embodiments, and in cases where separate instructions are provided, then both the realm entry instruction and the exception return instruction could be disabled when the current realm is a sub-realm.

Similarly, on occurrence of an exception when in a sub-realm, rather than taking the exception from directly from the sub-realm, the processing circuitry may trigger an exit from the sub-realm to a parent full realm which initialized the sub-realm, before handling the exception. Hence, exceptions trigger a return to the parent full realm. The exception return to the parent full realm may include the state masking, scrubbing and saving operations to the REC, but by avoiding exceptions being taken directly from a sub-realm to a realm at a higher exception level, this avoids the need to bank exception control registers such as the ELR, SPSR and ESR a further time for sub-realms, simplifying the architecture.

For a sub-realm, the boundary exception level, which indicates the maximum privilege level allowed for processing of the realm, is equal to the boundary exception level for its parent full realm. In contrast, for a child full realm the boundary exception level is a less privileged exception level than the boundary exception level of its parent realm.

When a realm is initialized by a parent realm, the parent realm may choose whether the new realm is to be a child full realm or a child sub-realm, and may set the appropriate realm type parameter in the realm descriptor accordingly. Once the realm is operational, the parent realm can no longer change the realm type, because modification of the realm descriptor is disabled through the managed realm lifecycle discussed above with respect to Figure 18.

In summary, the ability to introduce sub-realms, which are managed similar to full realms but with exception handling, realm initialization and realm entry functionality disabled within the sub-realm, enables a smaller portion of code corresponding to a given address range within the software process of a full realm to be isolated from other parts of that software, to provide additional security for a certain piece of sensitive code or data.

Figure 35 illustrates a simulator implementation that may be used. Whilst the earlier described embodiments implement the present invention in terms of apparatus and methods for operating specific processing hardware supporting the techniques concerned, it is also possible to provide an instruction execution environment in accordance with the embodiments described herein which is implemented through the use of a computer program. Such computer programs are often referred to as simulators, insofar as they provide a software based implementation of a hardware architecture. Varieties of simulator computer programs include emulators, virtual machines, models, and binary translators, including dynamic binary translators. Typically, a simulator implementation may run on a host processor 730, optionally running a host operating system 720, supporting the simulator program 710. In some arrangements, there may be multiple layers of simulation between the hardware and the provided instruction execution environment, and/or multiple distinct instruction execution environments provided on the same host processor. Historically, powerful processors have been required to provide simulator implementations which execute at a reasonable speed, but such an approach may be justified in certain circumstances, such as when there is a desire to run code native to another processor for compatibility or re-use reasons. For example, the simulator implementation may provide an instruction execution environment with additional functionality which is not supported by the host processor hardware, or provide an instruction execution environment typically associated with a different hardware architecture. An overview of simulation is given in "Some Efficient Architecture Simulation Techniques", Robert Bedichek, Winter 1990 USENIX Conference, Pages 53 - 63.

To the extent that embodiments have previously been described with reference to particular hardware constructs or features, in a simulated embodiment, equivalent functionality may be provided by suitable software constructs or features. For example, particular circuitry may be implemented in a simulated embodiment as computer program logic. Similarly, memory hardware, such as a register or cache, may be implemented in a simulated embodiment as a software data structure. In arrangements where one or more of the hardware elements referenced in the previously described embodiments are present on the host hardware (for example, host processor 730), some simulated embodiments may make use of the host hardware, where suitable.

The simulator program 710 may be stored on a computer-readable storage medium

(which may be a non-transitory medium), and provides a program interface (instruction execution environment) to the target code 700 (which may include the applications, operating systems and hypervisor as shown in Figure 2) which is the same as the application program interface of the hardware architecture being modelled by the simulator program 710. Thus, the program instructions of the target code 700, including the control of memory accesses based on the realm protection functionality described above, may be executed from within the instruction execution environment using the simulator program 710, so that a host computer 730 which does not actually have the hardware features of the apparatus 2 discussed above can emulate these features.

At least some examples provide a virtual machine computer program comprising instructions for controlling a host data processing apparatus to provide an instruction execution environment according to an apparatus comprising: processing circuitry to perform data processing in response to one or more software processes; memory access circuitry to enforce ownership rights for a plurality of memory regions, wherein a given memory region is associated with an owner realm specified from among a plurality of realms, each realm corresponding to at least a portion of at least one of the software processes, said owner realm having a right to exclude other realms from accessing data stored within said given memory region; and realm management circuitry to access a realm management tree storing realm management data for at least two realms in a tree structure comprising a variable number of levels; wherein the realm management circuitry is configured to index the realm management tree using a realm identifier of a given realm, the realm identifier comprising a variable number of variable length bit portions, wherein the realm management circuitry is configured to index into a given level of the realm management tree using a corresponding one of the variable length bit portions of the realm identifier. A storage medium may store the virtual machine computer program. The storage medium may be a non-transitory storage medium.

In the present application, the words "configured to..." are used to mean that an element of an apparatus has a configuration able to carry out the defined operation. In this context, a "configuration" means an arrangement or manner of interconnection of hardware or software. For example, the apparatus may have dedicated hardware which provides the defined operation, or a processor or other processing device may be programmed to perform the function. "Configured to" does not imply that the apparatus element needs to be changed in any way in order to provide the defined operation.

Although illustrative embodiments of the invention have been described in detail herein with reference to the accompanying drawings, it is to be understood that the invention is not limited to those precise embodiments, and that various changes and modifications can be effected therein by one skilled in the art without departing from the scope and spirit of the invention as defined by the appended claims.