[PATCH] OBP Memory allocation

[Olivier Danet]

When called with a null base address, char *obp_memalloc(va,size,align) 
shall
provide the address itself.

(and remove some trailing whitespaces)

Signed-off-by : Olivier Danet

Index: arch/sparc32/lib.c
===================================================================
--- arch/sparc32/lib.c    (révision 1257)
+++ arch/sparc32/lib.c    (copie de travail)
@@ -212,13 +212,19 @@
      phys_addr_t phys;
      ucell virt;

-    DPRINTF("obp_memalloc: virta 0x%x, sz %d, align %d\n", (unsigned 
int)va, size, align);
-
+    DPRINTF("obp_memalloc: virta 0x%x, sz %d, align %d\n", (unsigned 
int)va, size, align);
+
+    /* Round to page size */
+    size = (size + 0xfff) & 0xFFFFF000;
+
      /* Claim physical memory */
      phys = ofmem_claim_phys(-1, size, align);

/* Claim virtual memory */
-    virt = ofmem_claim_virt(pointer2cell(va), size, 0);
+    if (va)
+        virt = ofmem_claim_virt(pointer2cell(va), size, 0);
+    else
+        virt = ofmem_claim_virt(0, size, 0x1000);

/* Map the memory */
      ofmem_map(phys, virt, size, 
ofmem_arch_default_translation_mode(phys));
@@ -230,7 +236,7 @@
  {
      unsigned long align = size;

-    DPRINTF("obp_dumb_memalloc: virta 0x%x, sz %d\n", (unsigned int)va, 
size);
+    DPRINTF("obp_dumb_memalloc: virta 0x%x, sz %d\n", (unsigned int)va, 
size);

/* Solaris seems to assume that the returned value is physically 
aligned to size.
         e.g. it is used for setting up page tables. Fortunately this is 
now handled by
===================================================================

[Mark Cave-Ayland]

On 09/02/14 22:59, Olivier Danet wrote:
> When called with a null base address, char *obp_memalloc(va,size,align)
> shall
> provide the address itself.
>
> (and remove some trailing whitespaces)
>
> Signed-off-by : Olivier Danet
> Index: arch/sparc32/lib.c
>
> --- arch/sparc32/lib.c (révision 1257)
> +++ arch/sparc32/lib.c (copie de travail)
> @@ -212,13 +212,19 @@
> phys_addr_t phys;
> ucell virt;
> DPRINTF("obp_memalloc: virta 0x%x, sz %d, align %d\n", (unsigned
>
> int)va, size, align);
> DPRINTF("obp_memalloc: virta 0x%x, sz %d, align %d\n", (unsigned
>
> int)va, size, align);
> /* Round to page size */
> size = (size + 0xfff) & 0xFFFFF000;
>
> /* Claim physical memory */
> phys = ofmem_claim_phys(-1, size, align);
>
> /* Claim virtual memory */
> virt = ofmem_claim_virt(pointer2cell(va), size, 0);
> if (va)
> virt = ofmem_claim_virt(pointer2cell(va), size, 0);
> else
> virt = ofmem_claim_virt(0, size, 0x1000);
>
> /* Map the memory */
> ofmem_map(phys, virt, size, ofmem_arch_default_translation_mode(phys));

Hmmmm this doesn't look right at all. The obp_memalloc() function takes 
3 parameters: va, size and align which should correspond with the 
equivalent parameters for the claim word.

The key here is the align parameter; if align == 0 then va is assumed to 
be valid, and the memory is allocated at the given address. If align != 
0 then va is ignored, and OFMEM will follow the specification and return 
the next available memory region from the pool with an alignment 
specified by align.

You'll need to check this using a gdb breakpoint against your NextSTEP 
image, however what you should see is that align should always be != 0 
when va == 0 and so OFMEM will follow the specification and ignore va. 
If this is not the case then NextSTEP is doing something very strange...

ATB,

Mark.