2.4: export the symbol "mmu_cr4_features" for XFree86 DRM kernel drivers

I'd like to request comments on the appended; proposed patch for 2.4 to export the symbol "mmu_cr4_features".

This is needed by the XFree86 DRM kernel drivers, since Christoph's backport of vmap() in 2.4.22-pre7. [some] DRM kernel drivers (e.g. radeon] refuse to load without this fix, from 2.4.22-pre7 onwards.

cheers,
Calum.

[diffs -u against 2.4.22-pre9]

@@ -128,6 +128,7 @@
 struct cpuinfo_x86 boot_cpu_data = { 0, 0, 0, 0, -1, 1, 0, 0, -1 };
 
 unsigned long mmu_cr4_features;
+EXPORT_SYMBOL(mmu_cr4_features);
 
 /*
  * Bus types ..

@@ -30,6 +30,7 @@
 #include <linux/blk.h>
 #include <linux/highmem.h>
 #include <linux/bootmem.h>
+#include <linux/module.h>
 #include <asm/processor.h>
 #include <linux/console.h>
 #include <linux/seq_file.h>

@@ -58,6 +59,7 @@
 };
 
 unsigned long mmu_cr4_features;
+EXPORT_SYMBOL(mmu_cr4_features);
 
 /* For PCI or other memory-mapped resources */
 unsigned long pci_mem_start = 0x10000000;

Can you explain why they need it and why they magically need it because of vmap() ?

Sorry, of course.

[as Mitch has already pointed out privately] e.g the XFree86 DRM radeon Makefile includes this:

[xc/programs/Xserver/hw/xfree86/os-support/linux/drm/kernel/Makefile.linux]

	# Check for 4-argument vmap() in some 2.5.x and 2.4.x kernels
	VMAP := $(shell grep -A1 'vmap.*count,$$' 

$(LINUXDIR)/include/linux/vmalloc.h | grep -c prot)

	ifneq ($(VMAP),0)
	EXTRA_CFLAGS += -DVMAP_4_ARGS
	endif

Christoph's vmap() backport introduced this reference in linux/vmalloc.h, so the DRM module is now compiled with VMAP_4_ARGS defined.

The radeon kernel driver source includes (via drm_memory.h) linux/vmalloc.h, iff VMAP_4_ARGS is defined. The latter includes (on i386) asm/asm-i386/pgtable.h, which includes asm/asm-i386/processor.h.

processor.h defines some inline functions [set_in_cr4() & clear_in_cr4()] which reference mmu_cr4_features.

mmu_cr4_features is currently declared in arch/i386/kernel/setup.c [and also arch/x86_64/kernel/setup.c] but not currently exported.

[The problem is not restricted to the radeon kernel driver; I've just used it as an example]

cheers,
Calum.

Sadly, this isn't the full story, since the radeon code doesn't reference *_in_cr4.

The radeon code calls io/agp remap functions - again only ifdef VMAP_4_ARGS - which do however call __flush_tlb_all(). This manipulates the PGE bit via mmu_cr4_features.

apols for the spam...

cheers,
Calum.