From: Xiongfeng Wang <hidden> Date: 2021-10-15 08:51:57
DGH (Data Gathering Hint) prohibits merging memory accesses with
Normal-NC or Device-GRE attributes before the hint instruction with any
memory accesses appearing after the hint instruction. It can be used for
performance benefit to prevent the prior access waiting too long just to
be merged. This patch provide a macro to expose it to the arch code.
Signed-off-by: Xiongfeng Wang <redacted>
Signed-off-by: Cheng Jian <redacted>
Signed-off-by: Yufeng Mo <redacted>
---
v1->v2:
extract this single patch from the origin patchset.
remove the macro in the assemble code.
add comments for macro dgh().
modify the commit message a little bit.
---
arch/arm64/include/asm/barrier.h | 8 ++++++++
1 file changed, 8 insertions(+)
On its own, this patch doesn't do anything. It's more interesting to see
how it will be used and maybe come up with a common name that other
architectures would share (or just implement as no-opp). I'm not sure
there was any conclusion last time we discussed this.
--
Catalin
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
On its own, this patch doesn't do anything. It's more interesting to see
how it will be used and maybe come up with a common name that other
architectures would share (or just implement as no-opp). I'm not sure
there was any conclusion last time we discussed this.
In the last mail, I was suggested to investigate the code in other architecture
to find if there exists similar interface. I searched 'merg' in the code and
didn't find similar interface.
The only thing similar I found is in Intel Software Developer's Manual. It says
"Write Combining (WC) ... Writes may be delayed and combined in the write
combining buffer (WC buffer) to reduce memory accesses. If the WC buffer is
partially filled, the writes may be delayed until the next occurrence of a
serializing event; such as an SFENCE or MFENCE instruction, CPUID or other
serializing instruction, a read or write to uncached memory, an interrupt
occurrence, or an execution of a LOCK instruction (including one with an
XACQUIRE or XRELEASE prefix)."
Maybe this is more like the write combine buffer flushing, not prevent merging.
Sorry I still didn't understand the difference clearly.
How about a common name called 'merge_prohibit_hint()'? Could you give me some
suggestions ?
Thanks,
Xiongfeng
On its own, this patch doesn't do anything. It's more interesting to see
how it will be used and maybe come up with a common name that other
architectures would share (or just implement as no-opp). I'm not sure
there was any conclusion last time we discussed this.
In the last mail, I was suggested to investigate the code in other architecture
to find if there exists similar interface. I searched 'merg' in the code and
didn't find similar interface.
Maybe no other architecture has such hint. They have write buffer
draining but that's more expensive.
The only thing similar I found is in Intel Software Developer's Manual. It says
"Write Combining (WC) ... Writes may be delayed and combined in the write
combining buffer (WC buffer) to reduce memory accesses. If the WC buffer is
partially filled, the writes may be delayed until the next occurrence of a
serializing event; such as an SFENCE or MFENCE instruction, CPUID or other
serializing instruction, a read or write to uncached memory, an interrupt
occurrence, or an execution of a LOCK instruction (including one with an
XACQUIRE or XRELEASE prefix)."
Maybe this is more like the write combine buffer flushing, not prevent merging.
Sorry I still didn't understand the difference clearly.
IIUC those drivers on x86 just rely on the microarchitecture aspects of
the draining (e.g. fill 64 bytes). On Arm we don't have such guarantee
as there's a wide variation in implementations, hence the DGH
instruction.
How about a common name called 'merge_prohibit_hint()'? Could you give me some
suggestions ?
I think "prohibit" looks more like not allowing any write-buffer merge.
Maybe stop_merge_hint(), stop_write_buffer_merge(),
stop_write_combining() (any other ideas?). It would be a NOP on all
other architectures.
--
Catalin
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
On its own, this patch doesn't do anything. It's more interesting to see
how it will be used and maybe come up with a common name that other
architectures would share (or just implement as no-opp). I'm not sure
there was any conclusion last time we discussed this.
In the last mail, I was suggested to investigate the code in other architecture
to find if there exists similar interface. I searched 'merg' in the code and
didn't find similar interface.
Maybe no other architecture has such hint. They have write buffer
draining but that's more expensive.
quoted
The only thing similar I found is in Intel Software Developer's Manual. It says
"Write Combining (WC) ... Writes may be delayed and combined in the write
combining buffer (WC buffer) to reduce memory accesses. If the WC buffer is
partially filled, the writes may be delayed until the next occurrence of a
serializing event; such as an SFENCE or MFENCE instruction, CPUID or other
serializing instruction, a read or write to uncached memory, an interrupt
occurrence, or an execution of a LOCK instruction (including one with an
XACQUIRE or XRELEASE prefix)."
Maybe this is more like the write combine buffer flushing, not prevent merging.
Sorry I still didn't understand the difference clearly.
IIUC those drivers on x86 just rely on the microarchitecture aspects of
the draining (e.g. fill 64 bytes). On Arm we don't have such guarantee
as there's a wide variation in implementations, hence the DGH
instruction.
quoted
How about a common name called 'merge_prohibit_hint()'? Could you give me some
suggestions ?
I think "prohibit" looks more like not allowing any write-buffer merge.
Maybe stop_merge_hint(), stop_write_buffer_merge(),
stop_write_combining() (any other ideas?). It would be a NOP on all
other architectures.
Given that DGH only affects prior memory accesses to Normal-NC or Device-GRE
memory, we probably want to avoid making this too broad. We don't use GRE in
Linux and I think Normal-NC is only exposed as an explicit part of the I/O
accessors in ioremap_wc(). So I'd be inclined to add something like:
io_stop_wc()
with corresponding documentation in Documentation/memory-barriers.txt.
Will
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
On its own, this patch doesn't do anything. It's more interesting to see
how it will be used and maybe come up with a common name that other
architectures would share (or just implement as no-opp). I'm not sure
there was any conclusion last time we discussed this.
In the last mail, I was suggested to investigate the code in other architecture
to find if there exists similar interface. I searched 'merg' in the code and
didn't find similar interface.
Maybe no other architecture has such hint. They have write buffer
draining but that's more expensive.
quoted
The only thing similar I found is in Intel Software Developer's Manual. It says
"Write Combining (WC) ... Writes may be delayed and combined in the write
combining buffer (WC buffer) to reduce memory accesses. If the WC buffer is
partially filled, the writes may be delayed until the next occurrence of a
serializing event; such as an SFENCE or MFENCE instruction, CPUID or other
serializing instruction, a read or write to uncached memory, an interrupt
occurrence, or an execution of a LOCK instruction (including one with an
XACQUIRE or XRELEASE prefix)."
Maybe this is more like the write combine buffer flushing, not prevent merging.
Sorry I still didn't understand the difference clearly.
IIUC those drivers on x86 just rely on the microarchitecture aspects of
the draining (e.g. fill 64 bytes). On Arm we don't have such guarantee
as there's a wide variation in implementations, hence the DGH
instruction.
quoted
How about a common name called 'merge_prohibit_hint()'? Could you give me some
suggestions ?
I think "prohibit" looks more like not allowing any write-buffer merge.
Maybe stop_merge_hint(), stop_write_buffer_merge(),
stop_write_combining() (any other ideas?). It would be a NOP on all
other architectures.
Given that DGH only affects prior memory accesses to Normal-NC or Device-GRE
memory, we probably want to avoid making this too broad. We don't use GRE in
Linux and I think Normal-NC is only exposed as an explicit part of the I/O
accessors in ioremap_wc(). So I'd be inclined to add something like:
io_stop_wc()
with corresponding documentation in Documentation/memory-barriers.txt.