Re: [PATCH v3 3/5] arm64/mm: Create initial page tables in init_pg_dir

2018-08-15 Thread Jun Yao
Hi James,

On Fri, Jul 06, 2018 at 03:41:07PM +0100, James Morse wrote:
> I missed one: head.S has a call to kasan_early_init() before start_kernel(),
> this goes messing with the page tables, and calls pgd_offset_k(), which pulls 
> in
> swapper_pg_dir. This one is enabled by CONFIG_KASAN.
> 
> Something like that same hunk [0] in kasan_early_init() fixes it. This is 
> still
> within arch/arm64, so I still think we should get away without some #ifdeffery
> to override the core-code's initial setup of swapper_pg_dir...

I'm sorry to reply you so late, I missed this email before.

In order to ensure that pgd_offset_k() works properly, I update
init_mm.pgd by introducing set_init_mm_pgd(). And its implementation is
like this:

>diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
>index 65f86271f02b..e4f0868b4cfd 100644
>--- a/arch/arm64/mm/mmu.c
>+++ b/arch/arm64/mm/mmu.c
>@@ -623,6 +623,19 @@ static void __init map_kernel(pgd_t *pgdp)
>kasan_copy_shadow(pgdp);
> }
>
>+void __init set_init_mm_pgd(pgd_t *pgd)
>+{
>+   pgd_t **addr = &(init_mm.pgd);
>+
>+   asm volatile("str %x0, [%1]\n"
>+   : : "r" (pgd), "r" (addr) : "memory");
>+}
> /*
>  * paging_init() sets up the page tables, initialises the zone memory
>  * maps and sets up the zero page.

The purpose of using assembly is to prevent KASAN instrumentation, as
KASAN has not been initialized when this function is called:

>diff --git a/arch/arm64/kernel/head.S b/arch/arm64/kernel/head.S
>index c3e4b1886cde..ede2e964592b 100644
>--- a/arch/arm64/kernel/head.S
>+++ b/arch/arm64/kernel/head.S
>@@ -439,6 +438,9 @@ __primary_switched:
>bl  __pi_memset
>dsb ishst   // Make zero page visible to 
> PTW
>
>+   adrpx0, init_pg_dir
>+   bl  set_init_mm_pgd
>+
> #ifdef CONFIG_KASAN
>bl  kasan_early_init
> #endif

What do you think?


Re: [PATCH v3 3/5] arm64/mm: Create initial page tables in init_pg_dir

2018-08-15 Thread Jun Yao
Hi James,

On Fri, Jul 06, 2018 at 03:41:07PM +0100, James Morse wrote:
> I missed one: head.S has a call to kasan_early_init() before start_kernel(),
> this goes messing with the page tables, and calls pgd_offset_k(), which pulls 
> in
> swapper_pg_dir. This one is enabled by CONFIG_KASAN.
> 
> Something like that same hunk [0] in kasan_early_init() fixes it. This is 
> still
> within arch/arm64, so I still think we should get away without some #ifdeffery
> to override the core-code's initial setup of swapper_pg_dir...

I'm sorry to reply you so late, I missed this email before.

In order to ensure that pgd_offset_k() works properly, I update
init_mm.pgd by introducing set_init_mm_pgd(). And its implementation is
like this:

>diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
>index 65f86271f02b..e4f0868b4cfd 100644
>--- a/arch/arm64/mm/mmu.c
>+++ b/arch/arm64/mm/mmu.c
>@@ -623,6 +623,19 @@ static void __init map_kernel(pgd_t *pgdp)
>kasan_copy_shadow(pgdp);
> }
>
>+void __init set_init_mm_pgd(pgd_t *pgd)
>+{
>+   pgd_t **addr = &(init_mm.pgd);
>+
>+   asm volatile("str %x0, [%1]\n"
>+   : : "r" (pgd), "r" (addr) : "memory");
>+}
> /*
>  * paging_init() sets up the page tables, initialises the zone memory
>  * maps and sets up the zero page.

The purpose of using assembly is to prevent KASAN instrumentation, as
KASAN has not been initialized when this function is called:

>diff --git a/arch/arm64/kernel/head.S b/arch/arm64/kernel/head.S
>index c3e4b1886cde..ede2e964592b 100644
>--- a/arch/arm64/kernel/head.S
>+++ b/arch/arm64/kernel/head.S
>@@ -439,6 +438,9 @@ __primary_switched:
>bl  __pi_memset
>dsb ishst   // Make zero page visible to 
> PTW
>
>+   adrpx0, init_pg_dir
>+   bl  set_init_mm_pgd
>+
> #ifdef CONFIG_KASAN
>bl  kasan_early_init
> #endif

What do you think?


Re: [PATCH v3 3/5] arm64/mm: Create initial page tables in init_pg_dir

2018-07-06 Thread James Morse
Hi Jun,

On 06/07/18 09:58, James Morse wrote:
> The series so far fails to boot from me. This is because the kaslr code tries 
> to
> read the kaslr-seed from the DT, via the fixmap. To do this, it needs the 
> fixmap
> tables installed, which early_fixmap_init() does.
> 
> early_fixmap_init() calls pgd_offset_k(), which assumes init_mm.pgd is in use,
> so we hit a BUG_ON() when trying to setup the fixmap because we added the 
> tables
> to the wrong page tables.
> 
> If you enable 'CONFIG_RANDOMIZE_BASE', even without EFI you should see the 
> same
> thing happen.
> 
> 
> I think we should move this hunk:
>> diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c
>> index 30ad2f085d1f..b065c08d4008 100644
>> --- a/arch/arm64/kernel/setup.c
>> +++ b/arch/arm64/kernel/setup.c
>> @@ -249,6 +249,7 @@ void __init setup_arch(char **cmdline_p)
>>  init_mm.end_code   = (unsigned long) _etext;
>>  init_mm.end_data   = (unsigned long) _edata;
>>  init_mm.brk= (unsigned long) _end;
>> +init_mm.pgd= init_pg_dir;
>>
>>  *cmdline_p = boot_command_line;
>>
> 
> into early_fixmap_init(), which is the first thing setup_arch() calls.
> Something like [0] fixes it.
> 
> This looks to be the only thing that goes near init_mm.pgd before 
> paging_init().

I missed one: head.S has a call to kasan_early_init() before start_kernel(),
this goes messing with the page tables, and calls pgd_offset_k(), which pulls in
swapper_pg_dir. This one is enabled by CONFIG_KASAN.

Something like that same hunk [0] in kasan_early_init() fixes it. This is still
within arch/arm64, so I still think we should get away without some #ifdeffery
to override the core-code's initial setup of swapper_pg_dir...

Thanks,

James


> [0] make sure fixmap tables go in the init page tables
> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index 117d080639b3..e097c78a66f8 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -770,6 +771,13 @@ void __init early_fixmap_init(void)
> pmd_t *pmdp;
> unsigned long addr = FIXADDR_START;
> 
> +   /*
> +* During early setup we use init_pg_dir, update init_mm so its
> +* implicit use by pgd_offset_k() gets the live page tables.
> +* swapper_pg_dir is restored by paging_init().
> +*/
> +   init_mm.pgd = init_pg_dir;
> +
> pgdp = pgd_offset_k(addr);
> pgd = READ_ONCE(*pgdp);
> if (CONFIG_PGTABLE_LEVELS > 3 &&
> 
> 
> 



Re: [PATCH v3 3/5] arm64/mm: Create initial page tables in init_pg_dir

2018-07-06 Thread James Morse
Hi Jun,

On 06/07/18 09:58, James Morse wrote:
> The series so far fails to boot from me. This is because the kaslr code tries 
> to
> read the kaslr-seed from the DT, via the fixmap. To do this, it needs the 
> fixmap
> tables installed, which early_fixmap_init() does.
> 
> early_fixmap_init() calls pgd_offset_k(), which assumes init_mm.pgd is in use,
> so we hit a BUG_ON() when trying to setup the fixmap because we added the 
> tables
> to the wrong page tables.
> 
> If you enable 'CONFIG_RANDOMIZE_BASE', even without EFI you should see the 
> same
> thing happen.
> 
> 
> I think we should move this hunk:
>> diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c
>> index 30ad2f085d1f..b065c08d4008 100644
>> --- a/arch/arm64/kernel/setup.c
>> +++ b/arch/arm64/kernel/setup.c
>> @@ -249,6 +249,7 @@ void __init setup_arch(char **cmdline_p)
>>  init_mm.end_code   = (unsigned long) _etext;
>>  init_mm.end_data   = (unsigned long) _edata;
>>  init_mm.brk= (unsigned long) _end;
>> +init_mm.pgd= init_pg_dir;
>>
>>  *cmdline_p = boot_command_line;
>>
> 
> into early_fixmap_init(), which is the first thing setup_arch() calls.
> Something like [0] fixes it.
> 
> This looks to be the only thing that goes near init_mm.pgd before 
> paging_init().

I missed one: head.S has a call to kasan_early_init() before start_kernel(),
this goes messing with the page tables, and calls pgd_offset_k(), which pulls in
swapper_pg_dir. This one is enabled by CONFIG_KASAN.

Something like that same hunk [0] in kasan_early_init() fixes it. This is still
within arch/arm64, so I still think we should get away without some #ifdeffery
to override the core-code's initial setup of swapper_pg_dir...

Thanks,

James


> [0] make sure fixmap tables go in the init page tables
> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index 117d080639b3..e097c78a66f8 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -770,6 +771,13 @@ void __init early_fixmap_init(void)
> pmd_t *pmdp;
> unsigned long addr = FIXADDR_START;
> 
> +   /*
> +* During early setup we use init_pg_dir, update init_mm so its
> +* implicit use by pgd_offset_k() gets the live page tables.
> +* swapper_pg_dir is restored by paging_init().
> +*/
> +   init_mm.pgd = init_pg_dir;
> +
> pgdp = pgd_offset_k(addr);
> pgd = READ_ONCE(*pgdp);
> if (CONFIG_PGTABLE_LEVELS > 3 &&
> 
> 
> 



Re: [PATCH v3 3/5] arm64/mm: Create initial page tables in init_pg_dir

2018-07-06 Thread James Morse
Hi Jun,

On 02/07/18 12:16, Jun Yao wrote:
> Create initial page tables in init_pg_dir and then create final
> page tables in swapper_pg_dir directly.

This is what the patch does, but doesn't preserve the why for the git-log. Could
you expand it to describe why we're doing this.


The series so far fails to boot from me. This is because the kaslr code tries to
read the kaslr-seed from the DT, via the fixmap. To do this, it needs the fixmap
tables installed, which early_fixmap_init() does.

early_fixmap_init() calls pgd_offset_k(), which assumes init_mm.pgd is in use,
so we hit a BUG_ON() when trying to setup the fixmap because we added the tables
to the wrong page tables.

If you enable 'CONFIG_RANDOMIZE_BASE', even without EFI you should see the same
thing happen.


I think we should move this hunk:
> diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c
> index 30ad2f085d1f..b065c08d4008 100644
> --- a/arch/arm64/kernel/setup.c
> +++ b/arch/arm64/kernel/setup.c
> @@ -249,6 +249,7 @@ void __init setup_arch(char **cmdline_p)
>   init_mm.end_code   = (unsigned long) _etext;
>   init_mm.end_data   = (unsigned long) _edata;
>   init_mm.brk= (unsigned long) _end;
> + init_mm.pgd= init_pg_dir;
>
>   *cmdline_p = boot_command_line;
>

into early_fixmap_init(), which is the first thing setup_arch() calls.
Something like [0] fixes it.

This looks to be the only thing that goes near init_mm.pgd before paging_init().


> diff --git a/arch/arm64/include/asm/pgtable.h 
> b/arch/arm64/include/asm/pgtable.h
> index 7c4c8f318ba9..3b408f21fe2e 100644
> --- a/arch/arm64/include/asm/pgtable.h
> +++ b/arch/arm64/include/asm/pgtable.h
> @@ -718,6 +718,8 @@ static inline pmd_t pmdp_establish(struct vm_area_struct 
> *vma,
>  }
>  #endif
>  
> +extern pgd_t init_pg_dir[PTRS_PER_PGD] __section(.init.data);
> +extern pgd_t init_pg_end[] __section(.init.data);

normally we'd spell this '__initdata', but if we move them out of the
'.init.data' section we shouldn't, otherwise the extern definition doesn't match
where the symbol appears. (Looks like I was wrong to think that tools like
sparse pick this up, that's just the __iomem/__user stuff.)


> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index 2dbb2c9f1ec1..a7ab0010ff80 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -628,26 +628,10 @@ static void __init map_kernel(pgd_t *pgdp)
>   */
>  void __init paging_init(void)
>  {
> - phys_addr_t pgd_phys = early_pgtable_alloc();
> - pgd_t *pgdp = pgd_set_fixmap(pgd_phys);
> -
> - map_kernel(pgdp);
> - map_mem(pgdp);
> -
> - /*
> -  * We want to reuse the original swapper_pg_dir so we don't have to
> -  * communicate the new address to non-coherent secondaries in
> -  * secondary_entry, and so cpu_switch_mm can generate the address with
> -  * adrp+add rather than a load from some global variable.
> -  *
> -  * To do this we need to go via a temporary pgd.
> -  */
> - cpu_replace_ttbr1(__va(pgd_phys));
> - memcpy(swapper_pg_dir, pgdp, PGD_SIZE);
> - cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
> -
> - pgd_clear_fixmap();
> - memblock_free(pgd_phys, PAGE_SIZE);
> + map_kernel(swapper_pg_dir);
> + map_mem(swapper_pg_dir);

> + cpu_replace_ttbr1(swapper_pg_dir);

The lm_alias() here is important: cpu_replace_ttbr1() calls virt_to_phys() on
its argument, virt_to_phys() is only intended for addresses in the linear-map
region of the VA space, as it works by doing some arithmetic with the address.

swapper_pg_dir is the name of the kernel symbol, so its address will be in the
kernel-text region of the VA space.

Today virt_to_phys() catches this happening and fixes it, CONFIG_DEBUG_VIRTUAL
will give you a warning, at some point virt_to_phys()'s safety net will go-away.

The original call that did this was wrapped in lm_alias(), which gives you the
linear-map address of a symbol in the kernel-text.


Thanks,

James


[0] make sure fixmap tables go in the init page tables
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 117d080639b3..e097c78a66f8 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -770,6 +771,13 @@ void __init early_fixmap_init(void)
pmd_t *pmdp;
unsigned long addr = FIXADDR_START;

+   /*
+* During early setup we use init_pg_dir, update init_mm so its
+* implicit use by pgd_offset_k() gets the live page tables.
+* swapper_pg_dir is restored by paging_init().
+*/
+   init_mm.pgd = init_pg_dir;
+
pgdp = pgd_offset_k(addr);
pgd = READ_ONCE(*pgdp);
if (CONFIG_PGTABLE_LEVELS > 3 &&





Re: [PATCH v3 3/5] arm64/mm: Create initial page tables in init_pg_dir

2018-07-06 Thread James Morse
Hi Jun,

On 02/07/18 12:16, Jun Yao wrote:
> Create initial page tables in init_pg_dir and then create final
> page tables in swapper_pg_dir directly.

This is what the patch does, but doesn't preserve the why for the git-log. Could
you expand it to describe why we're doing this.


The series so far fails to boot from me. This is because the kaslr code tries to
read the kaslr-seed from the DT, via the fixmap. To do this, it needs the fixmap
tables installed, which early_fixmap_init() does.

early_fixmap_init() calls pgd_offset_k(), which assumes init_mm.pgd is in use,
so we hit a BUG_ON() when trying to setup the fixmap because we added the tables
to the wrong page tables.

If you enable 'CONFIG_RANDOMIZE_BASE', even without EFI you should see the same
thing happen.


I think we should move this hunk:
> diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c
> index 30ad2f085d1f..b065c08d4008 100644
> --- a/arch/arm64/kernel/setup.c
> +++ b/arch/arm64/kernel/setup.c
> @@ -249,6 +249,7 @@ void __init setup_arch(char **cmdline_p)
>   init_mm.end_code   = (unsigned long) _etext;
>   init_mm.end_data   = (unsigned long) _edata;
>   init_mm.brk= (unsigned long) _end;
> + init_mm.pgd= init_pg_dir;
>
>   *cmdline_p = boot_command_line;
>

into early_fixmap_init(), which is the first thing setup_arch() calls.
Something like [0] fixes it.

This looks to be the only thing that goes near init_mm.pgd before paging_init().


> diff --git a/arch/arm64/include/asm/pgtable.h 
> b/arch/arm64/include/asm/pgtable.h
> index 7c4c8f318ba9..3b408f21fe2e 100644
> --- a/arch/arm64/include/asm/pgtable.h
> +++ b/arch/arm64/include/asm/pgtable.h
> @@ -718,6 +718,8 @@ static inline pmd_t pmdp_establish(struct vm_area_struct 
> *vma,
>  }
>  #endif
>  
> +extern pgd_t init_pg_dir[PTRS_PER_PGD] __section(.init.data);
> +extern pgd_t init_pg_end[] __section(.init.data);

normally we'd spell this '__initdata', but if we move them out of the
'.init.data' section we shouldn't, otherwise the extern definition doesn't match
where the symbol appears. (Looks like I was wrong to think that tools like
sparse pick this up, that's just the __iomem/__user stuff.)


> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index 2dbb2c9f1ec1..a7ab0010ff80 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -628,26 +628,10 @@ static void __init map_kernel(pgd_t *pgdp)
>   */
>  void __init paging_init(void)
>  {
> - phys_addr_t pgd_phys = early_pgtable_alloc();
> - pgd_t *pgdp = pgd_set_fixmap(pgd_phys);
> -
> - map_kernel(pgdp);
> - map_mem(pgdp);
> -
> - /*
> -  * We want to reuse the original swapper_pg_dir so we don't have to
> -  * communicate the new address to non-coherent secondaries in
> -  * secondary_entry, and so cpu_switch_mm can generate the address with
> -  * adrp+add rather than a load from some global variable.
> -  *
> -  * To do this we need to go via a temporary pgd.
> -  */
> - cpu_replace_ttbr1(__va(pgd_phys));
> - memcpy(swapper_pg_dir, pgdp, PGD_SIZE);
> - cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
> -
> - pgd_clear_fixmap();
> - memblock_free(pgd_phys, PAGE_SIZE);
> + map_kernel(swapper_pg_dir);
> + map_mem(swapper_pg_dir);

> + cpu_replace_ttbr1(swapper_pg_dir);

The lm_alias() here is important: cpu_replace_ttbr1() calls virt_to_phys() on
its argument, virt_to_phys() is only intended for addresses in the linear-map
region of the VA space, as it works by doing some arithmetic with the address.

swapper_pg_dir is the name of the kernel symbol, so its address will be in the
kernel-text region of the VA space.

Today virt_to_phys() catches this happening and fixes it, CONFIG_DEBUG_VIRTUAL
will give you a warning, at some point virt_to_phys()'s safety net will go-away.

The original call that did this was wrapped in lm_alias(), which gives you the
linear-map address of a symbol in the kernel-text.


Thanks,

James


[0] make sure fixmap tables go in the init page tables
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 117d080639b3..e097c78a66f8 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -770,6 +771,13 @@ void __init early_fixmap_init(void)
pmd_t *pmdp;
unsigned long addr = FIXADDR_START;

+   /*
+* During early setup we use init_pg_dir, update init_mm so its
+* implicit use by pgd_offset_k() gets the live page tables.
+* swapper_pg_dir is restored by paging_init().
+*/
+   init_mm.pgd = init_pg_dir;
+
pgdp = pgd_offset_k(addr);
pgd = READ_ONCE(*pgdp);
if (CONFIG_PGTABLE_LEVELS > 3 &&





[PATCH v3 3/5] arm64/mm: Create initial page tables in init_pg_dir

2018-07-02 Thread Jun Yao
Create initial page tables in init_pg_dir and then create final
page tables in swapper_pg_dir directly.

Signed-off-by: Jun Yao 
---
 arch/arm64/include/asm/pgtable.h |  2 ++
 arch/arm64/kernel/head.S |  4 ++--
 arch/arm64/kernel/setup.c|  1 +
 arch/arm64/mm/mmu.c  | 24 
 4 files changed, 9 insertions(+), 22 deletions(-)

diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index 7c4c8f318ba9..3b408f21fe2e 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -718,6 +718,8 @@ static inline pmd_t pmdp_establish(struct vm_area_struct 
*vma,
 }
 #endif
 
+extern pgd_t init_pg_dir[PTRS_PER_PGD] __section(.init.data);
+extern pgd_t init_pg_end[] __section(.init.data);
 extern pgd_t swapper_pg_dir[PTRS_PER_PGD];
 extern pgd_t swapper_pg_end[];
 extern pgd_t idmap_pg_dir[PTRS_PER_PGD];
diff --git a/arch/arm64/kernel/head.S b/arch/arm64/kernel/head.S
index a1c7a4d3b9f3..a8e9432ffa8a 100644
--- a/arch/arm64/kernel/head.S
+++ b/arch/arm64/kernel/head.S
@@ -380,7 +380,7 @@ __create_page_tables:
/*
 * Map the kernel image (starting with PHYS_OFFSET).
 */
-   adrpx0, swapper_pg_dir
+   adrpx0, init_pg_dir
mov_q   x5, KIMAGE_VADDR + TEXT_OFFSET  // compile time __va(_text)
add x5, x5, x23 // add KASLR displacement
mov x4, PTRS_PER_PGD
@@ -837,7 +837,7 @@ __primary_switch:
mrs x20, sctlr_el1  // preserve old SCTLR_EL1 value
 #endif
 
-   adrpx26, swapper_pg_dir
+   adrpx26, init_pg_dir
bl  __enable_mmu
 #ifdef CONFIG_RELOCATABLE
bl  __relocate_kernel
diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c
index 30ad2f085d1f..b065c08d4008 100644
--- a/arch/arm64/kernel/setup.c
+++ b/arch/arm64/kernel/setup.c
@@ -249,6 +249,7 @@ void __init setup_arch(char **cmdline_p)
init_mm.end_code   = (unsigned long) _etext;
init_mm.end_data   = (unsigned long) _edata;
init_mm.brk= (unsigned long) _end;
+   init_mm.pgd= init_pg_dir;
 
*cmdline_p = boot_command_line;
 
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 2dbb2c9f1ec1..a7ab0010ff80 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -628,26 +628,10 @@ static void __init map_kernel(pgd_t *pgdp)
  */
 void __init paging_init(void)
 {
-   phys_addr_t pgd_phys = early_pgtable_alloc();
-   pgd_t *pgdp = pgd_set_fixmap(pgd_phys);
-
-   map_kernel(pgdp);
-   map_mem(pgdp);
-
-   /*
-* We want to reuse the original swapper_pg_dir so we don't have to
-* communicate the new address to non-coherent secondaries in
-* secondary_entry, and so cpu_switch_mm can generate the address with
-* adrp+add rather than a load from some global variable.
-*
-* To do this we need to go via a temporary pgd.
-*/
-   cpu_replace_ttbr1(__va(pgd_phys));
-   memcpy(swapper_pg_dir, pgdp, PGD_SIZE);
-   cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
-
-   pgd_clear_fixmap();
-   memblock_free(pgd_phys, PAGE_SIZE);
+   map_kernel(swapper_pg_dir);
+   map_mem(swapper_pg_dir);
+   cpu_replace_ttbr1(swapper_pg_dir);
+   init_mm.pgd = swapper_pg_dir;
 
/*
 * We only reuse the PGD from the swapper_pg_dir, not the pud + pmd
-- 
2.17.1



[PATCH v3 3/5] arm64/mm: Create initial page tables in init_pg_dir

2018-07-02 Thread Jun Yao
Create initial page tables in init_pg_dir and then create final
page tables in swapper_pg_dir directly.

Signed-off-by: Jun Yao 
---
 arch/arm64/include/asm/pgtable.h |  2 ++
 arch/arm64/kernel/head.S |  4 ++--
 arch/arm64/kernel/setup.c|  1 +
 arch/arm64/mm/mmu.c  | 24 
 4 files changed, 9 insertions(+), 22 deletions(-)

diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index 7c4c8f318ba9..3b408f21fe2e 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -718,6 +718,8 @@ static inline pmd_t pmdp_establish(struct vm_area_struct 
*vma,
 }
 #endif
 
+extern pgd_t init_pg_dir[PTRS_PER_PGD] __section(.init.data);
+extern pgd_t init_pg_end[] __section(.init.data);
 extern pgd_t swapper_pg_dir[PTRS_PER_PGD];
 extern pgd_t swapper_pg_end[];
 extern pgd_t idmap_pg_dir[PTRS_PER_PGD];
diff --git a/arch/arm64/kernel/head.S b/arch/arm64/kernel/head.S
index a1c7a4d3b9f3..a8e9432ffa8a 100644
--- a/arch/arm64/kernel/head.S
+++ b/arch/arm64/kernel/head.S
@@ -380,7 +380,7 @@ __create_page_tables:
/*
 * Map the kernel image (starting with PHYS_OFFSET).
 */
-   adrpx0, swapper_pg_dir
+   adrpx0, init_pg_dir
mov_q   x5, KIMAGE_VADDR + TEXT_OFFSET  // compile time __va(_text)
add x5, x5, x23 // add KASLR displacement
mov x4, PTRS_PER_PGD
@@ -837,7 +837,7 @@ __primary_switch:
mrs x20, sctlr_el1  // preserve old SCTLR_EL1 value
 #endif
 
-   adrpx26, swapper_pg_dir
+   adrpx26, init_pg_dir
bl  __enable_mmu
 #ifdef CONFIG_RELOCATABLE
bl  __relocate_kernel
diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c
index 30ad2f085d1f..b065c08d4008 100644
--- a/arch/arm64/kernel/setup.c
+++ b/arch/arm64/kernel/setup.c
@@ -249,6 +249,7 @@ void __init setup_arch(char **cmdline_p)
init_mm.end_code   = (unsigned long) _etext;
init_mm.end_data   = (unsigned long) _edata;
init_mm.brk= (unsigned long) _end;
+   init_mm.pgd= init_pg_dir;
 
*cmdline_p = boot_command_line;
 
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 2dbb2c9f1ec1..a7ab0010ff80 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -628,26 +628,10 @@ static void __init map_kernel(pgd_t *pgdp)
  */
 void __init paging_init(void)
 {
-   phys_addr_t pgd_phys = early_pgtable_alloc();
-   pgd_t *pgdp = pgd_set_fixmap(pgd_phys);
-
-   map_kernel(pgdp);
-   map_mem(pgdp);
-
-   /*
-* We want to reuse the original swapper_pg_dir so we don't have to
-* communicate the new address to non-coherent secondaries in
-* secondary_entry, and so cpu_switch_mm can generate the address with
-* adrp+add rather than a load from some global variable.
-*
-* To do this we need to go via a temporary pgd.
-*/
-   cpu_replace_ttbr1(__va(pgd_phys));
-   memcpy(swapper_pg_dir, pgdp, PGD_SIZE);
-   cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
-
-   pgd_clear_fixmap();
-   memblock_free(pgd_phys, PAGE_SIZE);
+   map_kernel(swapper_pg_dir);
+   map_mem(swapper_pg_dir);
+   cpu_replace_ttbr1(swapper_pg_dir);
+   init_mm.pgd = swapper_pg_dir;
 
/*
 * We only reuse the PGD from the swapper_pg_dir, not the pud + pmd
-- 
2.17.1