Skip to content

Latest commit

 

History

History
83 lines (66 loc) · 4.11 KB

programmers_guide.md

File metadata and controls

83 lines (66 loc) · 4.11 KB

Programmer's Guide

Initialization

Software is expected to configure prescaler and step before activating the timer. These two fields need to be stable to correctly increment the timer value. If software wants to change these fields, it should de-activate the timer and then proceed.

Register Access

The timer IP has 64-bit timer value registers and 64-bit compare registers. The register interface, however, is set to 32-bit data width. The CPU cannot access 64-bit data in a single request. However, when split into two reads, it is possible the timer value can increment between the two requests.

The IP doesn't have a latching or blocking mechanism to avoid this issue. It is the programmer's responsibility to ensure the correct value is read. For instance, if the CPU reads 0xFFFF_FFFF from lower 32-bit timer value (mtime) and 0x0000_0001 from upper 32-bit timer value (mtimeh), there is a chance that rather than having the value 0x1_FFFF_FFFF the timer value has changed from 0x0_FFFF_FFFF to 0x1_0000_0000 between the two reads. If there is the possibility of an interrupt between the two reads then the counter could have advanced even more.

This condition can be detected in a standard way using a third read. Figure 10.1 in the RISC-V unprivileged specification explains how to avoid this.

again:
    rdcycleh  x3
    rdcycle   x2
    rdcycleh  x4
    bne       x3, x4, again

Updating mtimecmp register also follows a similar approach to avoid a spurious interrupt during the register update. Please refer to the mtimecmp section in the RISC-V privileged specification.

# New comparand is in a1:a0.
li t0, -1
li t1, TOP_EARLGREY_RV_TIMER_BASE_ADDR
sw t0, RV_TIMER_COMPARE_LOWER0_0_REG_OFFSET(t1)   # No smaller than old value.
sw a1, RV_TIMER_COMPARE_UPPER0_0_REG_OFFSET(t1)   # No smaller than new value.
sw a0, RV_TIMER_COMPARE_LOWER0_0_REG_OFFSET(t1)   # New value.

Timer behaviour close to 2^64

There are some peculiarities when mtime and mtimecmp get close to the end of the 64-bit integer range. In particular, because an unsigned comparison is done between mtime and mtimecmp care is needed. A couple of cases are:

  1. mtimecmp close to 0xFFFF_FFFF_FFFF_FFFF. In this case the time-out event will be signaled when mtime passes the comparison value, the interrupt will be raised and the source indicated in the corresponding bit of the interrupt status register. However, if there is a delay in servicing the interrupt the mtime value could wrap to zero (and continue to increment) so the value read by the interrupt service routine will be less than the comparison value.

  2. When the timer is setup to trigger a timeout some number of timer ticks into the future, the computation of the comparison value mtime + timeout may overflow. If this value is set in mtimecmp it would make mtime greater than mtimecmp and immediately signal an interrupt. A possible solution is to have an intermediate interrupt by setting the mtimecmp to 64-bit all-ones, 0xFFFF_FFFF_FFFF_FFFF. Then the service routine for that interrupt will need to poll mtime until it wraps (which could take up to a timer clock tick) before scheduling the required interrupt using the originally computed mtimecmp value.

Interrupt Handling

If mtime is greater than or equal to the value of mtimecmp, the interrupt is generated from the RV_TIMER module. If the core enables the timer interrupt in MIE CSR, it jumps into the timer interrupt service routine. Clearing the interrupt can be done by writing 1 into the Interrupt Status register INTR_STATE0. The RV_TIMER module also follows RISC-V Privileged spec that requires the interrupt to be cleared by updating mtimecmp memory-mapped CSRs. In this case both COMPARE_LOWER0_0 and COMPARE_UPPER0_0 can clear the interrupt source.

Device Interface Functions (DIFs)