And Steve Jobs quoted him, famously saying in 1996:
"Picasso had a saying -- 'good artists copy; great artists steal' -- and we have always been shameless about stealing great ideas."However, stealing (or sharing, depending on your point of view) is what open source is all about. There are two important things when doing so. One is not to claim other peoples work as your own. Retain any authorship comments and just be honest for crying out loud. (Maybe be thankful too). The second is to retain any licensing and/or copyright headers. Then you are good to go.
For starters, the entire file "arch/arm/arm/cpufunc_asm_armv7.S". This also pulls in five header files, which I also copy. I may be able to prune these down, but if I do that, I will do it after I get everything working. This is done and it now compiles as part of Kyu without error.
Next I am looking at "arch/arm/arm/armv6_start.S". It is quick work to get it using the include files that I sorted out for cpufunc above. I get a variety of unsatisfied references which bear close scrutiny:
in function `generic_savevars': undefined reference to `kern_vtopdiff' undefined reference to `uboot_args' in function `generic_vstartv7': undefined reference to `cpu_info_store' undefined reference to `arm_cpu_topology_set' undefined reference to `start' in function `generic_vprint': in function `generic_prints': in function `generic_printx': undefined reference to `uartputc'The reference to "start" is the easy one, this would be in locore.S on NetBSD and is the label branched to when armv6_start is all done. I can make this whatever I want.
All the references to uartputc are conditional on VERBOSE_INIT_ARM, which is not defined for my build (but easily could be).
kern_vtopdiff is a location that is set by this startup code. The startup code figures out the virtual to physical address offset and posts it here for the kernel to use later. uboot_args is a 4 element array somewhere that the startup code fills with information. cpu_info_store is an array of structures in arm/arm_machdep.c. The startup code is once again placing information into the first of these structures (for core 0) for NetBSD to pick up later.
arm_cpu_topology_set() is a function (the code calls it with a "bl" instruction). It is in arm_cpu_topology.c. The startup code reads the MPIDR register and passes it to this function. The function extracts values and then calls another routine -- cpu_topology_set() -- with the extracted values.
My thoughts about all of this is that all of these can be commented out or made to call stubs (although uartputc would be useful and easy enough to set up).
I will put my comments up front and let the listing below end this page.
First of all note that "uartputc" is defined here and bounces the call
to _platform_early_putchar, which is somewhat like what I would do.
Also note the block of data in .Lstart. They use a multiple register
transfer to cleverly load this into several registers. The first
two items are the start and end addresses for the clearing of BSS.
The r8 register gets a pointer to "cpu_info_store" which we have
seen before. The CI_ARM_CPUID field gets the value of some darn CPU
config register.
MRC p15, 0,Here is what is in the MIDR (specific values for the Cortex-A7 MPCore)
I have no idea what the TPIDRPRW macros are about. A search on this reveals that this is an ARM register described as "PL1 only Thread ID Register". The description indicates that this is a scratch register visible only in PL1 that the OS can use to keep track of which thread is running.
MRC p15, 0,Neither TPI macro is defined in my build and although these macros are also referenced in armv6_start.S, they are only referenced if MULTIPROCESSOR is defined (and it is not). All this is somewhat mysterious, but ultimately the information goes into cpu_info_store[] and is not something I am concerned with.
Notice how they clear fp so backtraces will terminate nicely.
Also notice the call to initarm(). This is a big deal, and there are a multitude of implementations for each board platform. My build uses the one in arch/evbarm/fdt/fdt_machdep.c -- if you want to avoid the project of learning about the fdt, then look at any of the other 30 or so for specific platforms. They do a lot of hardware initialization (which is of course board dependent). This routine returns the value to be used for the sp (but in r0 of course). This code copies that return value into the sp.
ASENTRY_NP(start)
mrs r1, cpsr /* fetch CPSR value */
msr spsr_sx, r1 /* set SPSR[23:8] to known value */
/*
* Get bss bounds (r1, r2), curlwp or curcpu (r8), and set initial
* stack.
*/
adr r1, .Lstart
ldmia r1, {r1, r2, r8, sp}
#if defined(TPIDRPRW_IS_CURCPU) || defined(TPIDRPRW_IS_CURLWP)
mcr p15, 0, r8, c13, c0, 4
#endif
#if defined(TPIDRPRW_IS_CURLWP)
ldr r8, [r8, #L_CPU] /* r8 needs curcpu in it */
#endif
mov r4, #0
#ifdef _ARM_ARCH_DWORD_OK
mov r5, #0
#endif
.L1:
#ifdef _ARM_ARCH_DWORD_OK
strd r4, r5, [r1], #0x0008 /* Zero the bss */
#else
str r4, [r1], #0x0004 /* Zero the bss */
#endif
cmp r1, r2
blt .L1
mrc p15, 0, r3, c0, c0, 0 /* get our cpuid and save it early */
str r3, [r8, #CI_ARM_CPUID]
mov fp, #0x00000000 /* trace back starts here */
bl _C_LABEL(initarm) /* Off we go */
/* initarm will return the new stack pointer. */
mov sp, r0
mov fp, #0x00000000 /* trace back starts here */
mov ip, sp
push {fp, ip, lr, pc}
sub fp, ip, #4
bl _C_LABEL(main) /* call main()! */
adr r0, .Lmainreturned
b _C_LABEL(panic)
ENTRY_NP(uartputc)
#ifdef EARLYCONS
b ___CONCAT(EARLYCONS, _platform_early_putchar)
#endif
RET
ASEND(uartputc)
.Lstart:
.word _edata
.word _end
#if defined(TPIDRPRW_IS_CURLWP)
.word _C_LABEL(lwp0)
#else
.word _C_LABEL(cpu_info_store)
#endif
#if !defined(__HAVE_GENERIC_START)
.word svcstk_end
#else
.word start_stacks_top
#endif
.Lmainreturned:
.asciz "main() returned"
.align 0
ASEND(start)
Kyu / tom@mmto.org