forked from jserv/xv6-x86_64
-
Notifications
You must be signed in to change notification settings - Fork 0
/
README.64bit
98 lines (89 loc) · 4.43 KB
/
README.64bit
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
While porting Xv6 to 64bit Intel/AMD platforms, the goal is to keep changes
to the codebase as non-intrusive as possible, leaving it as short, clean,
and readable as it was in the original 32bit version. When possible,
#ifdefs will be avoided, and the existing style will be maintained. Also,
it should continue to be possible to build, boot, and run Xv6 on 32bit
platforms from the same codebase.
DONE
* Quick hack-and-slash modifications to get a buildable 64bit kernel.elf
* stubbed out or quickly patched up assembly glue where it didn't build
* temporarily turned off int-to-pointer and pointer-to-int warnings
to make finding build breakages faster
* chase down link errors due to too-large relocations
* switch to cross-x86-64 compiler to avoid weirdness from host compiler
* https://github.com/travisg/toolchains
* make printf() and cprintf() use stdarg.h instead of rolling their own
* x86-64's calling conventions for var-args are more complex
* notably the first six arguments may be passed in registers
* implement 32bit-to-64bit kernel entry glue (entry64.S)
* qemu cannot load 64bit ELF kernels
* sounds like most linux bootloaders don't support that yet either
* implement a 32bit multiboot header and shim
* setup an initial identity-mapped and kernel-address-mapped
pagetable since you cannot enter 64bit mode without paging on
* adjust syscall assembly for 64bit
* implement 64bit task switch glue
* introduce new type (currently 'uintp' for unsigned integer the size
of a pointer) to use in all the places where Xv6 assumes pointers and
integers are interchangeable.
* almost all warnings squashed
* consider a better name. addr_t?
* implement 64bit GDT and TSS setup
* implement 64bit interrupt support and IDT setup
* adjust process entry code for 64bit
* 64bit pointers and ABI (pass args in regs)
* implement 64bit mmu support (vm64.c)
* verify usermode works
* turn back on pre-emption
* verify usertests work
* update boot sector to support multiboot extended header
* the ELF header no longer handles the entrypoint
* get 32bit build working again
* tidy up exec() for 64bit (argument passing, stack alignment)
* test on real hardware
* enable SMP and verify it
* tidy up vm64.c
* fix procdump for 64bit mode
* figure out how to enumerate CPUs without BIOS MP table
* had to implement minimal ACPI table parsing
IN PROGRESS
TODO
* detect actual memory size, make PHYSTOP dynamic
* handle >2GB ram
* will have to move the linear map region below the kernel
since we can only map 2GB from the kernel base to end of vram
* fix validateint() of usertests for 64bit
THINGS TO FIX LATER / NICE TO HAVE
* I wish qemu had a "halt-on-exception" mode. Would help early bringup
* gdb pukes when qemu switches from 32bit to 64bit mode
* this made debugging the mode change entertaining
* for now attach gdb after the switch
* move userspace entry code to ulib
LESSONS LEARNED / UNRESOLVED
* using 1GB pages in the initial page table did not work
* In qemu you can do cheesy debugging by writing bytes to port 0x3f8
(the first UART) without any normal UART setup.
* Xv6:32 creates kernel mappings (separate copies) in every process's
page table, roughly 70k pages worth, consuming about 280MB of ram
for page tables.
* Xv6:32's trick for cpu-local storage does not work on gcc x86-64
- see README.CLS for an exploration of alternate options
* Not sure why you're double/triple-faulting?
drop some instrumentation code in qemu/target-i386/seg_helper.c's
do_interrupt64() -- this is where all the validation on irq/trap
entry happens
IMPLEMENTATION NOTES
* use "kernel" addressing mode for kernel compilation
* kernel virtual address space 0xFFFFFFFF80000000:0xFFFFFFFFFF000000
* linear-map first 1GB of RAM to kernel virtual address space
* linear-map first 1GB of RAM 1:1 during boot
* x64 has 4 levels of page tables instead of just 2
* continue to return a 2nd-level page table from setupkvm()
but actually create 4th, 3rd, and 2nd level tables
* use the top two entries of the 2nd level table for backpointers
to the 4th and 3rd level tables so that switchkvm(), switchuvm(),
and freevm() can get at these when needed
* use the top entry of the 4th level table to point to a shared
(among all processes) 3rd level table which handles the kernel
memory mapping (which is the same everywhere)
* userland will have 4GB - 8KB address space