Showing changes from revision #6 to #7:
Added | Removed | Changed
Hardware vectors, like their software counterparts (Software Vectors), contain an address of a routine that will be called in specific situations. Hardware vectors are called when a privileged mode is entered or when a hardware error occurs. These conditions are known as exceptions.
Each vector will usually hold an address of a routine that will deal with the exception. Each vector has a different priority which is used to determine the order in which exceptions should be handled (if there are simultaneous exceptions).
The ARM processor handles exceptions by:
The following table list each of the different hardware vectors.
Offset | Name |
---|---|
&00 | Reset |
&04 | Undefined instruction |
&08 | SWI |
&0C | Prefetch abort |
&10 | Data abort |
&14 | Address Exception/Hypervisor trap |
&18 | IRQ |
&1C | FIQ |
The offset given in the table above is relative to the vector base address. Under RISC OS this will either be &0 or &FFFF0000, as indicated by the flag returned in bit 20 of OS_PlatformFeatures 0.
This vector is used to specify that the computer is reset. The ARM processor can be reset by pulling its RESET pin HIGH. When RESET goes LOW again, the following will occur:
This vector is called when the ARM processor attempts to execute an instruction that is unknown. If a co-processor (software or hardware) is present on the system such as a floating point emulator, the ARM will pass it onto it (when the co-processor is ready). Any instruction still unknown is passed on, and this vector is called.
The ARM processor will:
This vector is called when a SWI instruction is issued. It contains an address of the routine used by RISC OS to decode the SWI number. Due to the importance of this vector it is strongly recommended not to replace it.
The ARM processor will:
This vector is called when an illegal attempt to prefetch an instruction has been detected. The cause of this could be:
The ARM processor will:
This vector is called when an illegal attempt to fetch data has been detected. The cause of this could be:
The ARM processor will:
On 26-bit only CPUs this was the address exception vector, which would be called when an attempt was made to transfer data from an address outside the range supported by the 26-bit address bus.
With the introduction of 32-bit processor modes in ARMv3 this vector was deprecated and became unused.
With the virtualisation extensions introduced with ARMv7, this vector entry is now being re-used for handling hypervisor traps.
Currently RISC OS does not make use of hypervisor mode, and so the vector has no practical use on any version of RISC OS from 3.5 and above.
This vector is called when an interrupt request is received by the ARM processor.
The ARM processor will:
The FIQ vector is called when a Fast Interrupt Request is received by the ARM processor. The FIQ vector is entered in FIQ mode.
The ARM processor will:
RISC OS provides the owner of the FIQ vector with 228 bytes of workspace to use as they see fit (from vector base +&1C to +&FF, inclusive). If your FIQ handler is small enough you can fit its code within this space, avoiding the need to start the handler with a branch instruction.
Each vector has a different priority which is used to determine the order in which exceptions should be handled (if there are simultaneous exceptions). We list the hardware vectors in order of priority.
Vector | Priority |
---|---|
Reset | 1 (highest priority) |
Data abort, Address exception | 2 |
FIQ | 3 |
IRQ | 4 |
Prefetch abort | 5 |
Undefined instruction, SWI | 6 (lowest priority) |
The Since original the introduction of the full 32bit address space in ARMv3 / ARM RISC processor OS could 3.5, only access 26-bit addresses, but with the introduction of the ARM RISC v3 OS architecture, has full contained 32-bit a addressing set was of made “pre-veneers” possible. which The are executed prior to the abortARMenvironment handlers . processor These included pre-veneers backwards take compatibility on by a supporting number both of 26-bit responsibilities: and 32-bit processor modes.
On these new CPUs the processor vectors are typically entered in 32-bit CPU modes rather than 26-bit. In order to provide compatibility with existing software, 26-bit versions of RISC OS install a set of pre-veneers on the hardware vectors which switch to the relevant 26-bit CPU mode before calling the vector handler. The only exception is the FIQ vector, which cannot support a pre-veneer due to the way the FIQ claim/release mechanism works under RISC OS.
Under RISC OS 5 all software is expected to be running in 32-bit mode, and so these pre-veneers are no longer required. However the system still contains other pre-veneers that are invoked before any user handlers that had been installed by OS_ClaimProcessorVector.
To claim a vector, you use OS_ClaimProcessorVector . You pass it the address of the replacement handler routine, which is installed on the vector. The address of the previous routine will be returned to you, so that your handler routine can use it to pass on a call if necessary. On older versions ofRISC OS (pre-RISC OS 3.5) it is necessary to patch the processor vectors directly. If your code needs to pass on to the previous claimant then care will be needed to make sure you decode the old handler instruction correctly (typically it will either be a branch instruction or a PC-relative LDR of the PC)
Note that the FIQ vector cannot be claimed using OS_ClaimProcessorVector. Instead, Service_ClaimFIQ or Service_ClaimFIQinBackground must be used, on all versions of RISC OS.
The LDREX/STREX instructions that were introduced with ARMv6 rely on a component of the CPU called the ‘local exclusive monitor’. For the load/store exclusive instructions to work correctly, the ARM requires that operating systems manually reset the exclusive monitor to the ‘open’ state before returning to any pre-empted code. Failure to do this can result in the program malfunctioning.
Because taking a hardware vector is often a form of pre-emption, this means that several of the vector handlers may have to reset the exclusive monitor on exit, as outlined below.
For more information on how to reset the exclusive monitor, see the ARMv7 compatibility primer.
Undefined instruction, instruction prefetch abort, and data prefetch abort handlers which (a) return to the running program and (b) make use of load/store exclusive instructions are required to reset the exclusive monitor on exit. exit – whether that exit is by returning to the program, or by passing on to the previous claimant. If the handler does not return exit (i.e. it raises an error instead) then there is no need to reset the exclusive monitor.
The requirements for data abort handlers are the same as for undefined instruction/prefetch abort handlers, with one addition: the exclusive monitor must be reset before any exclusive load/store instruction is used. This is necessary because ARM state that the state of the local exclusive monitor is undefined upon taking a data abort.
This means that the full set of rules are:
ARM recommend that in order to avoid external factors causing the exclusive monitor to spuriously reset, the processing within an exclusive access sequence is kept as short and simple as possible (e.g. only use compare and ALU instructions). A SWI/hypervisor call is not a simple operation, and so they should not be used from within exclusive access sequences. Because they should not be used, there is no point in making sure the exclusive monitor is reset on exit. Therefore RISC OS defines the exclusive monitor state as being unknown on return from a SWI or HVC.
This has implications for other types of handler – if the handler either directly calls a SWI/HVC or calls some other code (e.g. an external callback function) which may use a SWI/HVC then the exclusive monitor state should now be considered as having been modified, and thus it is in need of resetting before the handler returns.
IRQ handlers (as registered via OS_Claim OS_ClaimDeviceVector) do not need to worry about clearing the exclusive monitor on exit; the post-IRQ code within the kernel will clear it for you.
If handlers are installed via OS_ClaimProcessorVector (and they do not return via the kernel) then they must clear the exclusive monitor themselves, as appropriate.
Because FIQ handlers return directly to the interrupted code, any FIQ handler which uses the exclusive access instructions must make sure it resets the exclusive monitor on exit.
Additionally, because FIQs can occur while inside data abort handlers, FIQ code which uses exclusive access instructions must reset the exclusive monitor on entry.
Code which bypasses OS_ClaimProcessorVector and installs itself directly on the processor vectors may have additional requirements. E.g. ARM define the exclusive monitor state as being undefined after taking a data abort; the data abort pre-veneer within the kernel deals with this by resetting the monitor before invoking the installed data abort handler (OS_ClaimProcessorVector). If your code bypasses this then it must reset the exclusive monitor itself.