diff mbox series

[U-Boot,RESEND,2/3] arm: introduce _relaxed MMIO accessors

Message ID 20190210161726.5454-3-andre.przywara@arm.com
State Changes Requested
Delegated to: Jagannadha Sutradharudu Teki
Headers show
Series arm: Introduce writel/readl_relaxed accessors | expand

Commit Message

Andre Przywara Feb. 10, 2019, 4:17 p.m. UTC
The normal MMIO accessor macros (readX/writeX) guarantee a strong ordering,
even with normal memory accesses: https://lwn.net/Articles/698014/
For some MMIO operations (framebuffers being a prominent example) this is
not needed, and the rather costly barrier inserted on weakly ordered
systems like ARM costs some performance.
To mitigate this issue, Linux introduced readX_relaxed and
writeX_relaxed primitives, which only guarantee ordering between each
other, so are typically faster due to the missing barrier.

We probably do not care so much about performance in U-Boot, but want to
have those primitives for two other reasons:
- Being able to use the _relaxed macros simplifies porting drivers from
  Linux.
- The missing barrier typically allows to generate smaller code, which can
  relieve some chronically tight SPL builds.

Add those macros definitions by using the __raw versions of the
accessors, which matches an earlier (and less complicated) version of
the Linux implementation.

Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Reviewed-by: Philipp Tomsich <philipp.tomsich@theobroma-systems.com>
[On my experimental RK3399 after modifying a few drivers:]
Tested-by: Philipp Tomsich <philipp.tomsich@theobroma-systems.com>
---
 arch/arm/include/asm/io.h | 10 ++++++++++
 1 file changed, 10 insertions(+)

Comments

Alexander Graf Feb. 20, 2019, 12:24 p.m. UTC | #1
On 02/10/2019 05:17 PM, Andre Przywara wrote:
> The normal MMIO accessor macros (readX/writeX) guarantee a strong ordering,
> even with normal memory accesses: https://lwn.net/Articles/698014/
> For some MMIO operations (framebuffers being a prominent example) this is
> not needed, and the rather costly barrier inserted on weakly ordered
> systems like ARM costs some performance.
> To mitigate this issue, Linux introduced readX_relaxed and
> writeX_relaxed primitives, which only guarantee ordering between each
> other, so are typically faster due to the missing barrier.
>
> We probably do not care so much about performance in U-Boot, but want to
> have those primitives for two other reasons:
> - Being able to use the _relaxed macros simplifies porting drivers from
>    Linux.
> - The missing barrier typically allows to generate smaller code, which can
>    relieve some chronically tight SPL builds.
>
> Add those macros definitions by using the __raw versions of the
> accessors, which matches an earlier (and less complicated) version of
> the Linux implementation.
>
> Signed-off-by: Andre Przywara <andre.przywara@arm.com>
> Reviewed-by: Philipp Tomsich <philipp.tomsich@theobroma-systems.com>
> [On my experimental RK3399 after modifying a few drivers:]
> Tested-by: Philipp Tomsich <philipp.tomsich@theobroma-systems.com>

Reviewed-by: Alexander Graf <agraf@suse.de>


Alex
diff mbox series

Patch

diff --git a/arch/arm/include/asm/io.h b/arch/arm/include/asm/io.h
index bcbaf0d83c..5b24b88efc 100644
--- a/arch/arm/include/asm/io.h
+++ b/arch/arm/include/asm/io.h
@@ -98,11 +98,21 @@  static inline void __raw_readsl(unsigned long addr, void *data, int longlen)
 #define writel(v,c)	({ u32 __v = v; __iowmb(); __arch_putl(__v,c); __v; })
 #define writeq(v,c)	({ u64 __v = v; __iowmb(); __arch_putq(__v,c); __v; })
 
+#define writeb_relaxed(v,c)	__raw_writeb(v, c)
+#define writew_relaxed(v,c)	__raw_writew(v, c)
+#define writel_relaxed(v,c)	__raw_writel(v, c)
+#define writeq_relaxed(v,c)	__raw_writeq(v, c)
+
 #define readb(c)	({ u8  __v = __arch_getb(c); __iormb(); __v; })
 #define readw(c)	({ u16 __v = __arch_getw(c); __iormb(); __v; })
 #define readl(c)	({ u32 __v = __arch_getl(c); __iormb(); __v; })
 #define readq(c)	({ u64 __v = __arch_getq(c); __iormb(); __v; })
 
+#define readb_relaxed(c)	__raw_readb(c)
+#define readw_relaxed(c)	__raw_readw(c)
+#define readl_relaxed(c)	__raw_readl(c)
+#define readq_relaxed(c)	__raw_readq(c)
+
 /*
  * The compiler seems to be incapable of optimising constants
  * properly.  Spell it out to the compiler in some cases.