Re: 64-bit vkernel

[Jordan Gordeev <jgordeev@xxxxxx>]

On 1/25/10 6:10 PM, Matthew Dillon wrote:
>      The issue has to do with the vkernel not having a boot loader.
>      So the vkernel binary is loaded as a normal ELF program by the
>      normal rtelf-elf, libc, etc code.  This portion of the vkernel,
>      along with any incidental malloc()'s done by libc and friends,
>      does not use the mapped vpagetable area.
>
>      I'm not sure what you are proposing in terms of coding.  Are
>      you talking about changing the KernelPTA calculation in the
>      vkernel code and the related pmaps?  If so I think it is a good
>      idea.
I was proposing to change vm_fault() and vm_fault_page() to pass an 
additional argument to vm_fault_vpagetable() - the actual virtual 
address of the fault. This would have allowed vm_fault_vpagetable() to 
always use the actual virtual address and not some value that is offset 
for the vkernel and not offset for userland processes.
I've found a less invasive solution though.
I'll illustrate it on the code of the 32-bit vkernel.
The vkernel does this to map its kernel virtual memory:
        try = (void *)0x40000000;
        base = NULL;
        while ((char *)try + KERNEL_KVA_SIZE < topofstack) {
                base = mmap(try, KERNEL_KVA_SIZE, PROT_READ|PROT_WRITE,
                            MAP_FILE|MAP_SHARED|MAP_VPAGETABLE,
                            MemImageFd, 0);
                if (base == try)
                        break;
                if (base != MAP_FAILED)
                        munmap(base, KERNEL_KVA_SIZE);
                try = (char *)try + 0x10000000;
        }
A simple change does wonders:
--- a/sys/platform/vkernel/platform/init.c
+++ b/sys/platform/vkernel/platform/init.c
@@ -473,7 +473,7 @@ init_kern_memory(void)
        while ((char *)try + KERNEL_KVA_SIZE < topofstack) {
                base = mmap(try, KERNEL_KVA_SIZE, PROT_READ|PROT_WRITE,
                            MAP_FILE|MAP_SHARED|MAP_VPAGETABLE,
-                           MemImageFd, 0);
+                           MemImageFd, try);
                if (base == try)
                        break;
                if (base != MAP_FAILED)
Now when we access address 'va', the vpagetable is asked to map address 
'va - start_of_region + offset' == 'va - try + try' == 'va'.
We can't pass mmap() a suitable offset argument unless we know where it 
will map our region.

About the progress of the 64-bit vkernel: it will be enduring a pkgsrc 
bulk build tomorrow; then I'll have to make the SMP version compile and 
run; after that I have to prepare it for publishing.