[PATCH v3] powerpc32: memset: only use dcbz once cache is enabled

memset() uses instruction dcbz to speed up clearing by not wasting time
loading cache line with data that will be overwritten.
Some platform like mpc52xx do no have cache active at startup and
can therefore not use memset(). Allthough no part of the code
explicitly uses memset(), GCC may makes calls to it.

This patch modifies memset() such that at startup, memset()
unconditionally jumps to simple_memset() which doesn't use
the dcbz instruction.

Once the initial MMU is set up, in machine_init() we patch memset()
by replacing this inconditional jump by a NOP

Signed-off-by: Christophe Leroy <redacted>
---
This patch goes on to of [v3] powerpc32: memcpy: only use dcbz once cache is enabled

Changes in v2:
   was part of [v2] powerpc32: memcpy/memset: only use dcbz once cache is enabled
changes in v3:
  Not using anymore feature-fixups
  Handling of memcpy() and memset() split in two patches
  
 arch/powerpc/kernel/setup_32.c |  1 +
 arch/powerpc/lib/copy_32.S     | 15 +++++++++++++++
 2 files changed, 16 insertions(+)

diff --git a/arch/powerpc/kernel/setup_32.c b/arch/powerpc/kernel/setup_32.c
index 362495f..345ec3a 100644
--- a/arch/powerpc/kernel/setup_32.c
+++ b/arch/powerpc/kernel/setup_32.c

@@ -124,6 +124,7 @@ notrace void __init machine_init(u64 dt_ptr)
 	udbg_early_init();
 
 	patch_instruction((unsigned int *)&memcpy, PPC_INST_NOP);
+	patch_instruction((unsigned int *)&memset, PPC_INST_NOP);
 
 	/* Do some early initialization based on the flat device tree */
 	early_init_devtree(__va(dt_ptr));

diff --git a/arch/powerpc/lib/copy_32.S b/arch/powerpc/lib/copy_32.S
index da5847d..68a59d4 100644
--- a/arch/powerpc/lib/copy_32.S
+++ b/arch/powerpc/lib/copy_32.S

@@ -73,8 +73,13 @@ CACHELINE_MASK = (L1_CACHE_BYTES-1)
  * Use dcbz on the complete cache lines in the destination
  * to set them to zero.  This requires that the destination
  * area is cacheable.  -- paulus
+ *
+ * During early init, cache might not be active yet, so dcbz cannot be used.
+ * We therefore jump to simple_memset which doesn't use dcbz. This jump is
+ * replaced by a nop once cache is active. This is done in machine_init()
  */
 _GLOBAL(memset)
+	b	simple_memset
 	rlwimi	r4,r4,8,16,23
 	rlwimi	r4,r4,16,0,15
 

@@ -122,6 +127,16 @@ _GLOBAL(memset)
 	bdnz	8b
 	blr
 
+/* Simple version of memset used during early boot until cache is enabled */
+simple_memset:
+	cmplwi	cr0,r5,0
+	addi	r6,r3,-1
+	beqlr
+	mtctr	r5
+1:	stbu	r4,1(r6)
+	bdnz	1b
+	blr
+
 /*
  * This version uses dcbz on the complete cache lines in the
  * destination area to reduce memory traffic.  This requires that

-- 
2.1.0

On Mon, Sep 14 2015, Christophe Leroy wrote:

memset() uses instruction dcbz to speed up clearing by not wasting time
loading cache line with data that will be overwritten.
Some platform like mpc52xx do no have cache active at startup and
can therefore not use memset(). Allthough no part of the code
explicitly uses memset(), GCC may makes calls to it.

This patch modifies memset() such that at startup, memset()
unconditionally jumps to simple_memset() which doesn't use
the dcbz instruction.

I tested both v3 patches (memcpy and memset) and they work for me.

Thanks
-Michal

On Mon, 2015-09-14 at 08:21 +0200, Christophe Leroy wrote:

Instead couldn't you use the generic memset at label 2: and patch the "bne 
2f"?

-Scott

Le 14/09/2015 17:20, Scott Wood a écrit :

On Mon, 2015-09-14 at 08:21 +0200, Christophe Leroy wrote:

quoted

memset() uses instruction dcbz to speed up clearing by not wasting time
loading cache line with data that will be overwritten.
Some platform like mpc52xx do no have cache active at startup and
can therefore not use memset(). Allthough no part of the code
explicitly uses memset(), GCC may makes calls to it.

This patch modifies memset() such that at startup, memset()
unconditionally jumps to simple_memset() which doesn't use
the dcbz instruction.

Once the initial MMU is set up, in machine_init() we patch memset()
by replacing this inconditional jump by a NOP

Signed-off-by: Christophe Leroy <redacted>
---
This patch goes on to of [v3] powerpc32: memcpy: only use dcbz once cache
is enabled

Changes in v2:
    was part of [v2] powerpc32: memcpy/memset: only use dcbz once cache is
enabled
changes in v3:
   Not using anymore feature-fixups
   Handling of memcpy() and memset() split in two patches
   
  arch/powerpc/kernel/setup_32.c |  1 +
  arch/powerpc/lib/copy_32.S     | 15 +++++++++++++++
  2 files changed, 16 insertions(+)

diff --git a/arch/powerpc/kernel/setup_32.c b/arch/powerpc/kernel/setup_32.c
index 362495f..345ec3a 100644
--- a/arch/powerpc/kernel/setup_32.c
+++ b/arch/powerpc/kernel/setup_32.c

@@ -124,6 +124,7 @@ notrace void __init machine_init(u64 dt_ptr)
       udbg_early_init();
  
       patch_instruction((unsigned int *)&memcpy, PPC_INST_NOP);
+     patch_instruction((unsigned int *)&memset, PPC_INST_NOP);
  
       /* Do some early initialization based on the flat device tree */
       early_init_devtree(__va(dt_ptr));

diff --git a/arch/powerpc/lib/copy_32.S b/arch/powerpc/lib/copy_32.S
index da5847d..68a59d4 100644
--- a/arch/powerpc/lib/copy_32.S
+++ b/arch/powerpc/lib/copy_32.S

@@ -73,8 +73,13 @@ CACHELINE_MASK = (L1_CACHE_BYTES-1)
   * Use dcbz on the complete cache lines in the destination
   * to set them to zero.  This requires that the destination
   * area is cacheable.  -- paulus
+ *
+ * During early init, cache might not be active yet, so dcbz cannot be

used.
+ * We therefore jump to simple_memset which doesn't use dcbz. This jump is
+ * replaced by a nop once cache is active. This is done in machine_init()
   */
  _GLOBAL(memset)
+     b       simple_memset
       rlwimi  r4,r4,8,16,23
       rlwimi  r4,r4,16,0,15
  

@@ -122,6 +127,16 @@ _GLOBAL(memset)
       bdnz    8b
       blr
  
+/* Simple version of memset used during early boot until cache is enabled

*/
+simple_memset:
+     cmplwi  cr0,r5,0
+     addi    r6,r3,-1
+     beqlr
+     mtctr   r5
+1:   stbu    r4,1(r6)
+     bdnz    1b
+     blr

Instead couldn't you use the generic memset at label 2: and patch the "bne
2f"?

Yes I could but it means adding a global symbol there at the "bne 2f". I 
thought it was what Michael didn't like in my v1 of memcpy().
What name could I give to that symbol ? Something like 
memcpy_nocache_patch: ?
What would be the best ? Having b 2f, and replacing it with bne 2f ? Or 
have b 2f just above and replace it by nop once cache is up ?

Christophe

On Mon, 2015-09-14 at 17:44 +0200, Christophe LEROY wrote:

Le 14/09/2015 17:20, Scott Wood a écrit :

quoted

On Mon, 2015-09-14 at 08:21 +0200, Christophe Leroy wrote:

quoted

memset() uses instruction dcbz to speed up clearing by not wasting time
loading cache line with data that will be overwritten.
Some platform like mpc52xx do no have cache active at startup and
can therefore not use memset(). Allthough no part of the code
explicitly uses memset(), GCC may makes calls to it.

This patch modifies memset() such that at startup, memset()
unconditionally jumps to simple_memset() which doesn't use
the dcbz instruction.

Once the initial MMU is set up, in machine_init() we patch memset()
by replacing this inconditional jump by a NOP

Signed-off-by: Christophe Leroy <redacted>
---
This patch goes on to of [v3] powerpc32: memcpy: only use dcbz once 
cache
is enabled

Changes in v2:
    was part of [v2] powerpc32: memcpy/memset: only use dcbz once cache 
is
enabled
changes in v3:
   Not using anymore feature-fixups
   Handling of memcpy() and memset() split in two patches
   
  arch/powerpc/kernel/setup_32.c |  1 +
  arch/powerpc/lib/copy_32.S     | 15 +++++++++++++++
  2 files changed, 16 insertions(+)

diff --git a/arch/powerpc/kernel/setup_32.c 

b/arch/powerpc/kernel/setup_32.c
index 362495f..345ec3a 100644

--- a/arch/powerpc/kernel/setup_32.c
+++ b/arch/powerpc/kernel/setup_32.c

@@ -124,6 +124,7 @@ notrace void __init machine_init(u64 dt_ptr)
       udbg_early_init();
  
       patch_instruction((unsigned int *)&memcpy, PPC_INST_NOP);
+     patch_instruction((unsigned int *)&memset, PPC_INST_NOP);
  
       /* Do some early initialization based on the flat device tree */
       early_init_devtree(__va(dt_ptr));

diff --git a/arch/powerpc/lib/copy_32.S b/arch/powerpc/lib/copy_32.S
index da5847d..68a59d4 100644
--- a/arch/powerpc/lib/copy_32.S
+++ b/arch/powerpc/lib/copy_32.S

@@ -73,8 +73,13 @@ CACHELINE_MASK = (L1_CACHE_BYTES-1)
   * Use dcbz on the complete cache lines in the destination
   * to set them to zero.  This requires that the destination
   * area is cacheable.  -- paulus
+ *
+ * During early init, cache might not be active yet, so dcbz cannot be

used.
+ * We therefore jump to simple_memset which doesn't use dcbz. This 
jump is
+ * replaced by a nop once cache is active. This is done in 
machine_init()
   */
  _GLOBAL(memset)
+     b       simple_memset
       rlwimi  r4,r4,8,16,23
       rlwimi  r4,r4,16,0,15
  

@@ -122,6 +127,16 @@ _GLOBAL(memset)
       bdnz    8b
       blr
  
+/* Simple version of memset used during early boot until cache is 

enabled
*/
+simple_memset:
+     cmplwi  cr0,r5,0
+     addi    r6,r3,-1
+     beqlr
+     mtctr   r5
+1:   stbu    r4,1(r6)
+     bdnz    1b
+     blr

Instead couldn't you use the generic memset at label 2: and patch the "bne
2f"?

Yes I could but it means adding a global symbol there at the "bne 2f". I 
thought it was what Michael didn't like in my v1 of memcpy().
What name could I give to that symbol ? Something like 
memcpy_nocache_patch: ?
What would be the best ? Having b 2f, and replacing it with bne 2f ? Or 
have b 2f just above and replace it by nop once cache is up ?

I'm not sure why a global symbol for an instruction to be patched would be a 
big deal (similar to kvm_emul.S)...  I thought Michael just wanted existing 
infrastructure to be used if it's practical to do so.

-Scott

On Mon, 2015-09-14 at 11:13 -0500, Scott Wood wrote:

On Mon, 2015-09-14 at 17:44 +0200, Christophe LEROY wrote:

quoted

Le 14/09/2015 17:20, Scott Wood a écrit :

quoted

On Mon, 2015-09-14 at 08:21 +0200, Christophe Leroy wrote:

quoted

diff --git a/arch/powerpc/lib/copy_32.S b/arch/powerpc/lib/copy_32.S
index da5847d..68a59d4 100644
--- a/arch/powerpc/lib/copy_32.S
+++ b/arch/powerpc/lib/copy_32.S

@@ -73,8 +73,13 @@ CACHELINE_MASK = (L1_CACHE_BYTES-1)

@@ -122,6 +127,16 @@ _GLOBAL(memset)
       bdnz    8b
       blr
  
+/* Simple version of memset used during early boot until cache is 

enabled
*/
+simple_memset:
+     cmplwi  cr0,r5,0
+     addi    r6,r3,-1
+     beqlr
+     mtctr   r5
+1:   stbu    r4,1(r6)
+     bdnz    1b
+     blr

Instead couldn't you use the generic memset at label 2: and patch the "bne
2f"?

Yes I could but it means adding a global symbol there at the "bne 2f". I 
thought it was what Michael didn't like in my v1 of memcpy().
What name could I give to that symbol ? Something like 
memcpy_nocache_patch: ?
What would be the best ? Having b 2f, and replacing it with bne 2f ? Or 
have b 2f just above and replace it by nop once cache is up ?

I'm not sure why a global symbol for an instruction to be patched would be a 
big deal (similar to kvm_emul.S)...  I thought Michael just wanted existing 
infrastructure to be used if it's practical to do so.

Yeah, I don't mind adding a global if you need it.

I'm not wedded to either approach, I'm just trying to find the least intrusive
option that uses the most existing code, and is the most maintainable into the
future.

cheers