2.6.4-rc1 fix x86 early_printk and make it early
PATCH 1/1 2.6.4-rc1 fix x86 early_printk and make it early
Trying to use early_printk on x86 I found two issues.
- setup_early_printk is currently called much later than necessary.
- VGABASE is using an identity mapped physical address on x86.
o That is problematic with PAE support, because there is a moment during paging_init() when the physical identity mappings do not work.
o Using raw physical addresses is in bad form, and doesn't always work.
o I can't possibly see how Andrew's Changelog that using __pa is more friendly to the 4G/4G split is correct. Unless someone was hard coding a virtual address previously.
The first hunk of this patch moves setup_early_printk as early as possible on x86.
The second hunk in early_printk.c redefines VGABASE as __va(0xb8000). This is correct on both x86 and x86-64 so we don't need any more special cases. The only thing that might be slightly more correct would be to use isa_readw/isa_writew but the only difference is in where PAGE_OFFSET is added in.
Eric
arch/i386/kernel/setup.c +12 -13
@@ -1048,6 +1048,18 @@ {unsigned long max_low_pfn; +#ifdef CONFIG_EARLY_PRINTK +extern void setup_early_printk (char *); +char *str; + +COMMAND_LINE [COMMAND_LINE_SIZE -1 ] ='\0' ; + str =strstr (COMMAND_LINE ,"earlyprintk=" ); +if (str) { +setup_early_printk (str); +printk ("early console enabled\n" ); + } +#endif +memcpy (&boot_cpu_data, &new_cpu_data,sizeof (new_cpu_data));pre_setup_arch_hook ();early_cpu_init ();
@@ -1117,19 +1129,6 @@smp_alloc_memory ();/* AP processor realmode stacks in low memory*/ #endif paging_init (); - -#ifdef CONFIG_EARLY_PRINTK - { -char *s =strstr (*cmdline_p,"earlyprintk=" ); -if (s) { -extern void setup_early_printk (char *); - -setup_early_printk (s); -printk ("early console enabled\n" ); - } - } -#endif -dmi_scan_machine ();
arch/x86_64/kernel/early_printk.c +1 -5
@@ -7,11 +7,7 @@/* Simple VGA output */ -#ifdef __i386__ -#define VGABASE __pa(__PAGE_OFFSET + 0xb8000UL) -#else -#define VGABASE 0xffffffff800b8000UL -#endif +#define VGABASE __va(0xb8000UL)#define MAX_YPOS 25#define MAX_XPOS 80
Quoted 3 lines · On 29 Feb 2004 22:24:12 -0700 [email protected] (Eric W. Biederman) wrote:
o That is problematic with PAE support, because there is a moment during
paging_init()when the physical identity mappings do not work.
Just don't printk in that moment then.
Quoted 1 line
o Using raw physical addresses is in bad form, and doesn't always work.
On x86-64 it's that __pa() doesn't always work very early.
Quoted 3 lines
o I can't possibly see how Andrew's Changelog that using
__pais more friendly to the 4G/4G split is correct. Unless someone was hard coding a virtual address previously.
Yes, it shouldn't make any difference for 4/4.
Quoted 3 lines
The second hunk in
early_printk.credefines VGABASE as__va(0xb8000). This is correct on both x86 and x86-64 so we don't need any more special cases.
Please don't do that on x86-64. On x86-64 there are two ways to reach this address and the previous one is available earlier. Keep the current address for x86-64 please.
-Andi
Quoted 19 lines · Andi Kleen
Quoted 3 lines · On 29 Feb 2004 22:24:12 -0700 [email protected] (Eric W. Biederman) wrote:
o That is problematic with PAE support, because there is a moment during
paging_init()when the physical identity mappings do not work.Just don't printk in that moment then.
Quoted 1 line
o Using raw physical addresses is in bad form, and doesn't always work.
On x86-64 it's that
__pa()doesn't always work very early.Quoted 3 lines
o I can't possibly see how Andrew's Changelog that using
__pais more friendly to the 4G/4G split is correct. Unless someone was hard coding a virtual address previously.Yes, it shouldn't make any difference for 4/4.
Quoted 7 lines
Quoted 3 lines
The second hunk in
early_printk.credefines VGABASE as__va(0xb8000). This is correct on both x86 and x86-64 so we don't need any more special cases.Please don't do that on x86-64. On x86-64 there are two ways to reach this address and the previous one is available earlier. Keep the current address for x86-64 please.
??? I checked the x86-64 early page tables and it does appear to work. That is why I consolidated the change.
In particular from arch/x86_64/head.S:
.org 0x1000ENTRY(init_level4_pgt)
.quad 0x0000000000102007 /* -> level3_ident_pgt */ .fill 255,8,0 .quad 0x000000000010a007 .fill 254,8,0 /* (2^48-(2*1024*1024*1024))/(2^39) = 511 */ .quad 0x0000000000103007 /* -> level3_kernel_pgt */
.org 0x3000ENTRY(level3_kernel_pgt)
.fill 510,8,0 /* (2^48-(2*1024*1024*1024)-((2^39)*511))/(2^30) = 510 */ .quad 0x0000000000105007 /* -> level2_kernel_pgt */ .fill 1,8,0
.org 0xa000ENTRY(level3_physmem_pgt)
.quad 0x0000000000105007 /* -> level2_kernel_pgt (so that __va works even before pagetable_init) */
0x0000010000000000 is handled by the pointer: 0x000000000010a007 0xffffffff80000000 is handled by the pointer: 0x0000000000103007
And then they both point to: 0x0000000000103007 which is the physical mapping.
And of course there is the comment on the 0xa000 entry:
(so that __va works even before pagetable_init)
So I don't see how using __va() early in the code is incorrect. I can buy that __va() adds the wrong constant but that is another matter.
Eric
Quoted 19 lines · Andi Kleen
Quoted 3 lines · On 29 Feb 2004 22:24:12 -0700 [email protected] (Eric W. Biederman) wrote:
o That is problematic with PAE support, because there is a moment during
paging_init()when the physical identity mappings do not work.Just don't printk in that moment then.
Quoted 1 line
o Using raw physical addresses is in bad form, and doesn't always work.
On x86-64 it's that
__pa()doesn't always work very early.Quoted 3 lines
o I can't possibly see how Andrew's Changelog that using
__pais more friendly to the 4G/4G split is correct. Unless someone was hard coding a virtual address previously.Yes, it shouldn't make any difference for 4/4.
Quoted 7 lines
Quoted 3 lines
The second hunk in
early_printk.credefines VGABASE as__va(0xb8000). This is correct on both x86 and x86-64 so we don't need any more special cases.Please don't do that on x86-64. On x86-64 there are two ways to reach this address and the previous one is available earlier. Keep the current address for x86-64 please.
Andi much to my great surprise after reading through the comments
that actually didn't work on x86-64. So here is my corrected patch.
It still uses __va() for everything but it fixes x86-64 so that it works, I just had to move one page table entry...
As for 0xffff000000000000 which the old page table entry applied to I grepped the entire kernel tree and could not find anything that applied to x86_64. So it looks like a left over from an earlier definition of PAGE_OFFSET.
So can we please get this applied to 2.6? Now that it is almost purely mini bug fixes?
Eric
arch/i386/kernel/setup.c +12 -13
@@ -1048,6 +1048,18 @@ {unsigned long max_low_pfn; +#ifdef CONFIG_EARLY_PRINTK +extern void setup_early_printk (char *); +char *str; + +COMMAND_LINE [COMMAND_LINE_SIZE -1 ] ='\0' ; + str =strstr (COMMAND_LINE ,"earlyprintk=" ); +if (str) { +setup_early_printk (str); +printk ("early console enabled\n" ); + } +#endif +memcpy (&boot_cpu_data, &new_cpu_data,sizeof (new_cpu_data));pre_setup_arch_hook ();early_cpu_init ();
@@ -1117,19 +1129,6 @@smp_alloc_memory ();/* AP processor realmode stacks in low memory*/ #endif paging_init (); - -#ifdef CONFIG_EARLY_PRINTK - { -char *s =strstr (*cmdline_p,"earlyprintk=" ); -if (s) { -extern void setup_early_printk (char *); - -setup_early_printk (s); -printk ("early console enabled\n" ); - } - } -#endif -dmi_scan_machine ();
arch/x86_64/kernel/early_printk.c +1 -5
@@ -7,11 +7,7 @@/* Simple VGA output */ -#ifdef __i386__ -#define VGABASE __pa(__PAGE_OFFSET + 0xb8000UL) -#else -#define VGABASE 0xffffffff800b8000UL -#endif +#define VGABASE __va(0xb8000UL)#define MAX_YPOS 25#define MAX_XPOS 80
arch/x86_64/kernel/head.S +3 -2
@@ -206,9 +206,10 @@ .org 0x1000 ENTRY(init_level4_pgt) .quad 0x0000000000102007 /* -> level3_ident_pgt */ - .fill 255,8,0 + .fill 1,8,0 + /* 2^40/2^39 = 2 */ .quad 0x000000000010a007 - .fill 254,8,0 + .fill 508,8,0 /* (2^48-(2*1024*1024*1024))/(2^39) = 511 */ .quad 0x0000000000103007 /* -> level3_kernel_pgt */