「这是我参与11月更文挑战的第17天，活动详情查看：2021最后一次更文挑战」

Traps and system calls

Learning xv6-riscv-book Chapter 4 Traps and system calls

[TOC]

---

Trap: CPU transfer to speical code to handle events

• system call: ecall into the kernel
• exception: something illegal
• interrupt: from device

---

xv6 kernel handles all traps.

code -> trap (handling in kernel) -> resume

---

trap handling proceeds:

1. hardware actions by CPU
2. vector prepares for kernel C code
3. trap handler decides what to do
4. do system call / device service

---

three cases of assembly vectors:

• traps from user space
• traps from kernel space
• timer interrupts

RISC-V trap machinery

Registers

A set of registers:

• kernel reads about a trap
• kernel writes to tell the CPU how to handle traps.

register	description	write by	when
stvec	address of trap handler	kernel
sepc	saved PC when a trap occurs	RISC-V	when a trap occurs
scause	reason of trap	RISC-V
sscratch	places a value that comes in handy	kernel	at the very start of a traphandler
sstatus	SIE bit: device interrupts are enabled? (defer until set) SPP bit: trap from user or supervisor mode?(ctrl what mode sret returns)

These registers are in supervisor mode: cannot be r/w in user mode

Machine mode has an equivalent set of these regs: only for timer interrupts

Each CPU has its own set: can handling traps at a same time

hardware trap handling sequence

trap_occurs:
	if trap is DeviceInterrupt:
        sstatus.SIE = 0
        goto end
    sstatus.SIE = 0  # disable interrupts
    sepc = pc
    sstatus.SPP = current_mode()
    scause = trap.cause
    set_mode(.supervisor)
    pc = stvec(trap)
    execute(pc)
end:

note: CPU does minimal work:

• kernel page table / stack are not switched
• registers other than PC are not saved

(kernel must do these tasks 👆)

Traps from user space

Traps from user space:

1. trap occurs:
   1. uservec (trampoline.S:16)
   2. usertrap (trap.c:37)
2. handle trap
3. returing:
   1. usertrapret (trap.c:90)
   2. userret (trampoline.S:88))

---

Hardware doesn't switch page tables during a trap:

• user page table includes a mapping for uservec
• uservec switch satp to kernel page table
• uservec must be at the same address in k & u. (to continue pc after switch)

So, xv6 use a trampiline page （VA at TRAMPILINE） to contains uservec. The contents of trampiline pages are set in trampiline.S.

---

uservec

1. starts: all registers -> values of interrupted code (sscratch points to p->trapframe);
   • csrrw instruction: swaps(a0, sscratch)
2. save user registers to trapframe (at a0)
3. save a0 to trapframe
4. switch satp to kernel page table (in trapframe)
5. calls usertrap

usertrap

(usertrap handle trap from user space, kernel traps are handled by kerneltrap)

• determine the cause of the trap
• process trap:
  • save sepc (the saved user program counter)
  • if trap is SystemCall: calls syscall()
  • if trap is DeviceIntrrupt: calls devintr
• return by usertrapret

usertrapret

sets up the RISC-V control registers to prepare for a future trap from user space:

• stvec
• tramframe
• sepc
• calls userret

userret

userret(TRAPFRAME, userpagetable);
          (a0)         (a1)

• switches satp to user page table.
• restores saved registers from trapframe.
• sret return to user space.

Calling system calls

e.g. (user/initcode.S:11) calling exec:

# exec(init, argv)
.globl start
start:
        la a0, init
        la a1, argv
        li a7, SYS_exec
        ecall

• a0, a1 <- arguments
• a7 <- system call number
• ecall -> uservec -> usertrap -> syscall
• syscall:
  • retrieves syscall number from saved a7 in the trapframe
  • use syscall number to index into syscalls
• system call implementation function returns:
  • syscall records return value in p->trapframe->a0

System call arguments

system call wrapper functions:

• places arguments in registers
• trap: registers are saved to trapframe
• argint, argaddr, argfd: retrieve arguments as integer, pointer or file descriptor.

---

pointer argument: (e.g. str):

• fetchstr -> copyinstr
• copyinstr: copies bytes from srcva (in user pagetable) to dst.
  • use walkaddr to walk the page table (in software) to get pa for srcva

Traps from kernel space

• kernelvec (kernelvec.S:10)

• kerneltrap (trap.c:134)

  • handle types of trap:

    • devintr
    • exception (kernel error => fatal): calls panic, stops executing
    • timer interrupt: yield to give up CPU.

  • return: restores control registers, return to kernelvec

Page-fault exceptions

Xv6: exception happens:

• exception from user space: kill the faulting process;
• exception from kernel: panic!

Real world: more interesting ways. e.g. use page faults to implement copy-on-write (COW) fork.

---

• Xv6 fork: calling uvmcopy to copy parent's memory into child. (no share)
• COW fork: safely share phyical memory
  • share all physical pages (read-only)
  • when child or patent store: raise a page-fault exception
  • response to this exception:
    • makes a copy of the page that contains the faulted address.
    • both allow to read/write
      • one for child, one for parent
    • resumes the process caused the fault

---

page-fault exception: CPU cannot translate a virtual address to a physical address

• load page faults
• store page faults
• instruction page fault

A page-fault exception happen:

• scause <- PageFault
• stval <- address that couldn't be translated

avoid complete copy & transparent

---

Other page faults feature:

• lazy allocation:
  • sbrk => grows address space, but not marks valid in pgtbl
  • page fault on new address => allocates physical memory, maps into the pgtbl
• paging from disk:
  • app need more memory than physical RAM => write some pages to disk
  • pg fault => if in disk: allocates a page of physical memory, read page from disk. (may evict another page to disk)
• automically extending stacks
• memory-mapped files
• ...

---

EOF

---

# By CDFMLR 2021-03-18
echo "See you."

顶部图片来自于小歪API，系随机选取的图片，仅用于检测屏幕显示的机械、光电性能，与文章的任何内容及观点无关，也并不代表本人局部或全部同意、支持或者反对其中的任何内容及观点。如有侵权，联系删除。