In accordance with [1, 2] memory transactions using CCA=2 (Uncached
Cacheability and Coherency Attribute) are always strongly ordered. This
means the younger memory accesses using CCA=2 are never allowed to be
executed before older memory accesses using CCA=2 (no bypassing is
allowed), and Loads and Stores using CCA=2 are never speculative. It is
expected by the specification that the rest of the system maintains these
properties for processor initiated uncached accesses. So the system IO
interconnect doesn't reorder uncached transactions once they have left the
processor subsystem. Taking into account these properties and what [3]
says about the relaxed IO-accessors we can infer that normal Loads and
Stores from/to CCA=2 memory and without any additional execution barriers
will fully comply with the {read,write}X_relaxed() methods requirements.
Let's convert then currently generated relaxed IO-accessors to being pure
Loads and Stores. Seeing the commit 3d474dacae72 ("MIPS: Enforce strong
ordering for MMIO accessors") and commit 8b656253a7a4 ("MIPS: Provide
actually relaxed MMIO accessors") have already made a preparation in the
corresponding macro, we can do that just by replacing the "barrier"
parameter utilization with the "relax" one. Note the "barrier" macro
argument can be removed, since it isn't fully used anyway other than being
always assigned to 1.
Of course it would be fullish to believe that all the available MIPS-based
CPUs completely follow the denoted specification, especially considering
how old the architecture is. Instead we introduced a dedicated kernel
config, which when enabled will convert the relaxed IO-accessors to being
pure Loads and Stores without any additional barriers around. So if some
CPU supports the strongly ordered UC memory access, it can enable that
config and use a fully optimized relaxed IO-methods. For instance,
Baikal-T1 architecture support code will do that.
[1] MIPS Coherence Protocol Specification, Document Number: MD00605,
Revision 01.01. September 14, 2015, 4.2 Execution Order Behavior,
p. 33
[2] MIPS Coherence Protocol Specification, Document Number: MD00605,
Revision 01.01. September 14, 2015, 4.8.1 IO Device Access, p. 58
[3] "LINUX KERNEL MEMORY BARRIERS", Documentation/memory-barriers.txt,
Section "KERNEL I/O BARRIER EFFECTS"
Signed-off-by: Serge Semin <[email protected]>
Cc: Maciej W. Rozycki <[email protected]>
---
arch/mips/Kconfig | 8 ++++++++
arch/mips/include/asm/io.h | 20 ++++++++++----------
2 files changed, 18 insertions(+), 10 deletions(-)
diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig
index c95fa3a2484c..2c82d927347d 100644
--- a/arch/mips/Kconfig
+++ b/arch/mips/Kconfig
@@ -2066,6 +2066,14 @@ config WEAK_ORDERING
#
config WEAK_REORDERING_BEYOND_LLSC
bool
+
+#
+# CPU may not reorder reads and writes R->R, R->W, W->R, W->W within Uncached
+# Cacheability and Coherency Attribute (CCA=2)
+#
+config STRONG_UC_ORDERING
+ bool
+
endmenu
#
diff --git a/arch/mips/include/asm/io.h b/arch/mips/include/asm/io.h
index 78537aa23500..130c4b6458fc 100644
--- a/arch/mips/include/asm/io.h
+++ b/arch/mips/include/asm/io.h
@@ -213,7 +213,7 @@ void iounmap(const volatile void __iomem *addr);
#define war_io_reorder_wmb() barrier()
#endif
-#define __BUILD_MEMORY_SINGLE(pfx, bwlq, type, barrier, relax, irq) \
+#define __BUILD_MEMORY_SINGLE(pfx, bwlq, type, relax, irq) \
\
static inline void pfx##write##bwlq(type val, \
volatile void __iomem *mem) \
@@ -221,7 +221,7 @@ static inline void pfx##write##bwlq(type val, \
volatile type *__mem; \
type __val; \
\
- if (barrier) \
+ if (!(relax && IS_ENABLED(CONFIG_STRONG_UC_ORDERING))) \
iobarrier_rw(); \
else \
war_io_reorder_wmb(); \
@@ -262,7 +262,7 @@ static inline type pfx##read##bwlq(const volatile void __iomem *mem) \
\
__mem = (void *)__swizzle_addr_##bwlq((unsigned long)(mem)); \
\
- if (barrier) \
+ if (!(relax && IS_ENABLED(CONFIG_STRONG_UC_ORDERING))) \
iobarrier_rw(); \
\
if (sizeof(type) != sizeof(u64) || sizeof(u64) == sizeof(long)) \
@@ -294,14 +294,14 @@ static inline type pfx##read##bwlq(const volatile void __iomem *mem) \
return pfx##ioswab##bwlq(__mem, __val); \
}
-#define __BUILD_IOPORT_SINGLE(pfx, bwlq, type, barrier, relax, p) \
+#define __BUILD_IOPORT_SINGLE(pfx, bwlq, type, relax, p) \
\
static inline void pfx##out##bwlq##p(type val, unsigned long port) \
{ \
volatile type *__addr; \
type __val; \
\
- if (barrier) \
+ if (!(relax && IS_ENABLED(CONFIG_STRONG_UC_ORDERING))) \
iobarrier_rw(); \
else \
war_io_reorder_wmb(); \
@@ -325,7 +325,7 @@ static inline type pfx##in##bwlq##p(unsigned long port) \
\
BUILD_BUG_ON(sizeof(type) > sizeof(unsigned long)); \
\
- if (barrier) \
+ if (!(relax && IS_ENABLED(CONFIG_STRONG_UC_ORDERING))) \
iobarrier_rw(); \
\
__val = *__addr; \
@@ -338,7 +338,7 @@ static inline type pfx##in##bwlq##p(unsigned long port) \
#define __BUILD_MEMORY_PFX(bus, bwlq, type, relax) \
\
-__BUILD_MEMORY_SINGLE(bus, bwlq, type, 1, relax, 1)
+__BUILD_MEMORY_SINGLE(bus, bwlq, type, relax, 1)
#define BUILDIO_MEM(bwlq, type) \
\
@@ -358,8 +358,8 @@ __BUILD_MEMORY_PFX(__mem_, q, u64, 0)
#endif
#define __BUILD_IOPORT_PFX(bus, bwlq, type) \
- __BUILD_IOPORT_SINGLE(bus, bwlq, type, 1, 0,) \
- __BUILD_IOPORT_SINGLE(bus, bwlq, type, 1, 0, _p)
+ __BUILD_IOPORT_SINGLE(bus, bwlq, type, 0,) \
+ __BUILD_IOPORT_SINGLE(bus, bwlq, type, 0, _p)
#define BUILDIO_IOPORT(bwlq, type) \
__BUILD_IOPORT_PFX(, bwlq, type) \
@@ -374,7 +374,7 @@ BUILDIO_IOPORT(q, u64)
#define __BUILDIO(bwlq, type) \
\
-__BUILD_MEMORY_SINGLE(____raw_, bwlq, type, 1, 0, 0)
+__BUILD_MEMORY_SINGLE(____raw_, bwlq, type, 0, 0)
__BUILDIO(q, u64)
--
2.27.0
?? 2020/9/20 19:00, Serge Semin д??:
> In accordance with [1, 2] memory transactions using CCA=2 (Uncached
> Cacheability and Coherency Attribute) are always strongly ordered. This
> means the younger memory accesses using CCA=2 are never allowed to be
> executed before older memory accesses using CCA=2 (no bypassing is
> allowed), and Loads and Stores using CCA=2 are never speculative. It is
> expected by the specification that the rest of the system maintains these
> properties for processor initiated uncached accesses. So the system IO
> interconnect doesn't reorder uncached transactions once they have left the
> processor subsystem. Taking into account these properties and what [3]
> says about the relaxed IO-accessors we can infer that normal Loads and
> Stores from/to CCA=2 memory and without any additional execution barriers
> will fully comply with the {read,write}X_relaxed() methods requirements.
>
> Let's convert then currently generated relaxed IO-accessors to being pure
> Loads and Stores. Seeing the commit 3d474dacae72 ("MIPS: Enforce strong
> ordering for MMIO accessors") and commit 8b656253a7a4 ("MIPS: Provide
> actually relaxed MMIO accessors") have already made a preparation in the
> corresponding macro, we can do that just by replacing the "barrier"
> parameter utilization with the "relax" one. Note the "barrier" macro
> argument can be removed, since it isn't fully used anyway other than being
> always assigned to 1.
>
> Of course it would be fullish to believe that all the available MIPS-based
> CPUs completely follow the denoted specification, especially considering
> how old the architecture is. Instead we introduced a dedicated kernel
> config, which when enabled will convert the relaxed IO-accessors to being
> pure Loads and Stores without any additional barriers around. So if some
> CPU supports the strongly ordered UC memory access, it can enable that
> config and use a fully optimized relaxed IO-methods. For instance,
> Baikal-T1 architecture support code will do that.
>
> [1] MIPS Coherence Protocol Specification, Document Number: MD00605,
> Revision 01.01. September 14, 2015, 4.2 Execution Order Behavior,
> p. 33
>
> [2] MIPS Coherence Protocol Specification, Document Number: MD00605,
> Revision 01.01. September 14, 2015, 4.8.1 IO Device Access, p. 58
>
> [3] "LINUX KERNEL MEMORY BARRIERS", Documentation/memory-barriers.txt,
> Section "KERNEL I/O BARRIER EFFECTS"
>
> Signed-off-by: Serge Semin <[email protected]>
> Cc: Maciej W. Rozycki <[email protected]>
Reviewed-by: Jiaxun Yang <[email protected]>
Based on #mipslinus discussions, I suspect this option can be selected by
most modern MIPS processors including all IMG/MTI cores,
Ingenic and Loongson.
Thanks.
- Jiaxun
> ---
> arch/mips/Kconfig | 8 ++++++++
> arch/mips/include/asm/io.h | 20 ++++++++++----------
> 2 files changed, 18 insertions(+), 10 deletions(-)
>
On Fri, Sep 25, 2020 at 11:54:20AM +0800, Jiaxun Yang wrote:
>
>
> 在 2020/9/20 19:00, Serge Semin 写道:
> > In accordance with [1, 2] memory transactions using CCA=2 (Uncached
> > Cacheability and Coherency Attribute) are always strongly ordered. This
> > means the younger memory accesses using CCA=2 are never allowed to be
> > executed before older memory accesses using CCA=2 (no bypassing is
> > allowed), and Loads and Stores using CCA=2 are never speculative. It is
> > expected by the specification that the rest of the system maintains these
> > properties for processor initiated uncached accesses. So the system IO
> > interconnect doesn't reorder uncached transactions once they have left the
> > processor subsystem. Taking into account these properties and what [3]
> > says about the relaxed IO-accessors we can infer that normal Loads and
> > Stores from/to CCA=2 memory and without any additional execution barriers
> > will fully comply with the {read,write}X_relaxed() methods requirements.
> >
> > Let's convert then currently generated relaxed IO-accessors to being pure
> > Loads and Stores. Seeing the commit 3d474dacae72 ("MIPS: Enforce strong
> > ordering for MMIO accessors") and commit 8b656253a7a4 ("MIPS: Provide
> > actually relaxed MMIO accessors") have already made a preparation in the
> > corresponding macro, we can do that just by replacing the "barrier"
> > parameter utilization with the "relax" one. Note the "barrier" macro
> > argument can be removed, since it isn't fully used anyway other than being
> > always assigned to 1.
> >
> > Of course it would be fullish to believe that all the available MIPS-based
> > CPUs completely follow the denoted specification, especially considering
> > how old the architecture is. Instead we introduced a dedicated kernel
> > config, which when enabled will convert the relaxed IO-accessors to being
> > pure Loads and Stores without any additional barriers around. So if some
> > CPU supports the strongly ordered UC memory access, it can enable that
> > config and use a fully optimized relaxed IO-methods. For instance,
> > Baikal-T1 architecture support code will do that.
> >
> > [1] MIPS Coherence Protocol Specification, Document Number: MD00605,
> > Revision 01.01. September 14, 2015, 4.2 Execution Order Behavior,
> > p. 33
> >
> > [2] MIPS Coherence Protocol Specification, Document Number: MD00605,
> > Revision 01.01. September 14, 2015, 4.8.1 IO Device Access, p. 58
> >
> > [3] "LINUX KERNEL MEMORY BARRIERS", Documentation/memory-barriers.txt,
> > Section "KERNEL I/O BARRIER EFFECTS"
> >
> > Signed-off-by: Serge Semin <[email protected]>
> > Cc: Maciej W. Rozycki <[email protected]>
> Reviewed-by: Jiaxun Yang <[email protected]>
>
>
> Based on #mipslinus discussions, I suspect this option can be selected by
> most modern MIPS processors including all IMG/MTI cores,
> Ingenic and Loongson.
Thanks for reviewing the patch.
Regarding the option. Alas it's not that easy and we must be very careful
with assumption whether some processor supports the denoted feature. Even
if the MIPS cores do imply the strict UC load/stores ordering, the system
interconnects may still perform the out-of-order requests execution. For
instance, the P5600 cores installed into our Baikal-T1 SoC do support the
strong UC ordering, but there is a cascade of the OCP2AXI, AXI2AXI and
AXI2APB bridges behind the CPU memory interface, each of which is equipped
with an internal FIFO and some complicated logic of the traffic routing.
So each platform should be carefully analyzed and tested (if it's
possible) before enabling the suggested feature, otherwise we'll risk to end
up with in general working, but at some point buggy, systems. Needless to
say, that out-of-order exec problems is very hard to track and debug due
to a random nature of impact on the system.
-Sergey
>
> Thanks.
>
> - Jiaxun
>
> > ---
> > arch/mips/Kconfig | 8 ++++++++
> > arch/mips/include/asm/io.h | 20 ++++++++++----------
> > 2 files changed, 18 insertions(+), 10 deletions(-)
> >