I'm familiar with MIPS architecture, and I've known that MIPS has memory section such as kseg0, kseg1. Which determine whether the segment can be cached or mapped. For example, you should locate some I/O devices(like UART) to the uncached segment.
But I didn't find anything related in RISCV arch. So how does the RISCV OS know the address should be mapped or not?
By the way: I know the value in satp CSR desrcibes the translation mode. When OS is running, the value must set other than "Bare(disabled MMU)" so that the OS can support the virtual memory. So if CPU access UART address, the value in satp is still not "Bare"? But it should be "Bare"?
RISC-V is a family of instruction sets, ranging from MCU style processors that have no memory-mapping and no memory protection mechanisms (Physical Memory Protection is optional).
From your question, I assume you are talking about processors that support User and Supervisor level ISA, as documented in the RISC-V privileged spec.
It sounds like you want a spec describing which physical addresses are cacheable. Looking at the list of CSRs, I believe this information is not in the CSRs because it is platform specific. In systems I've worked with, it is either hard-coded in platform drivers or passed via device-tree.
For Linux, the device-tree entries are not RISC-V specific: there are device tree entries specifying the physical address range of memory. Additionally, each I/O device would have a device tree entry specifying its physical address range.