Prep: Turning the MMU On — 39k

Session: Fri Oct 30, 1h30 · Exercises: 39k_virtual_memory · Prep time: ~20 min · Lecture: Week 10 · Locks, Semaphores, and Turning the MMU On

What you will build

The kernel's address space, handed to the hardware. Using mappages from 33k_paging, you will identity-map everything the kernel touches after the switch: the UART page, the test-finisher page, and all of RAM from KERNBASE to PHYSTOP, page tables included. You will also pack the satp value that names the root table, for the given csrw satp and sfence.vma sequence. With the MMU still off, the harness walks each region to confirm it is present, identity-mapped, and correctly permissioned, checks that satp carries MODE 8 and the root's page number, then flips the switch and prints OSLINGS:PASS.

Concepts you need

• The bootstrap paradox: the fetch after csrw satp is translated — Week 10 · The next fetch is translated
• Identity mapping, va == pa — Week 10 · The next fetch is translated
• If the kernel will touch it, map it; the third argument is a size — Week 10 · The next fetch is translated · Sv39 Paging guide: The kernel page table on a fresh boot
• satp: MODE 8, ASID 0, root PPN; machine mode ignores it — Week 10 · satp by hand, Silence, and checking first · Sv39 Paging guide: The satp register
• The TLB is not coherent; sfence.vma invalidates and orders — Week 10 · Two instructions, and the TLB · Sv39 Paging guide: sfence.vma and the TLB
• Silence is the default failure; read satp, pc, scause, stval in GDB — Week 10 · Silence, and checking first, Reading the silence · QEMU and GDB guide: Diagnostic playbook

Read before class

What	Time
Week 10 · Friday · 39k Turning the MMU on through satp by hand, with its recap of 33k	3 min
Week 10 · Two instructions, and the TLB and Silence, and checking first	3 min
Week 10 · For the exam: Megapages, and 66 pages. Not needed today; on Midterm 2	1 min
Sv39 Paging guide: The satp register and sfence.vma and the TLB	4 min
QEMU and GDB guide: Diagnostic playbook, the "It worked until I turned the MMU on" and "QEMU hangs" rows	2 min

Mental model

A toy kernel, root table at 0x8700_0000, next instruction at 0x8000_1234:

satp = (8 << 60) | (0x8700_0000 >> 12)
     = 0x8000_0000_0008_7000

csrw satp, t0          executes with translation OFF
fetch 0x8000_1234      translation ON: root[2] -> L1[0] -> L0[1]
                       leaf needs V=1, X=1, PPN<<12 == 0x8000_1000
sfence.vma zero, zero  runs only if that fetch succeeded

Nothing in RAM moved; the meaning of every register changed between two adjacent instructions. Under an identity map 0x8000_1234 translates to itself and the kernel does not notice. If that leaf is missing, the fetch faults, stvec is still zero, and the machine loops silently at address 0, because printing itself needs a fetch, a stack store, and the UART page. That is why the harness verifies with walk first, and why a silent kernel means p/x $satp in GDB, not another print.

Check yourself

1. GDB shows satp = 0x8000_0000_0008_7FFF. Is paging on, and where is the root table?
   Answer

   Top nibble 8 is Sv39, so yes; ASID 0. PPN 0x87FFF shifted left by 12 puts the root at 0x87FF_F000, the highest page in RAM and the allocator's first.
2. Your kernel survives the switch, prints OSLINGS:PASS, then QEMU never exits and oslings waits out its 10-second timeout. Which region is missing?
   Answer

   The test-finisher page at 0x10_0000. Text and the UART must be mapped, since it ran and printed; the exit store is the first access outside them. oslings saw PASS, so the run still goes green: the wait is the only symptom. (In 39k itself, which runs in machine mode, the check before the switch reports a missing region as a [fail] line instead.)
3. Right after the switch, scause reads 1, not 12. Do you suspect a mapping or satp?
   Answer

   satp. A page fault (12) means your table refused the fetch; an access fault (1) means the hardware could not even read a PTE, classically a PPN field holding the root's address unshifted. On QEMU 10, though, that unshifted root reports 12 too, so read satp before blaming a mapping.

What "done" looks like

oslings run is green, then oslings submit before you leave. Not green? Submit anyway — the tests that pass earn their share — then finish it at a make-up session (office hours, on the class network) for three quarters, or anywhere else for half, and submit again.

If you finish early

Work Week 10 · Problem 4: satp both ways and Problem 5: Count the tables on paper, read Week 10 · How xv6 and Linux do it beside xv6 book sections 3.3–3.4, or start next Thursday's prep page, Prep: Filesystem.