Various architectures have almost the same implementations for
__memcpy_{to,from}io and __memset_io functions. So, consolidate them
into the existing lib/iomap_copy.c.
Reviewed-by: Yann Sionneau <ysionneau@kalrayinc.com>
Signed-off-by: Julian Vetter <jvetter@kalrayinc.com>
---
Signed-off-by: Julian Vetter <jvetter@kalrayinc.com>
---
Changes for v6:
- Included linux/aslign.h
- Replaced compile time check by ifdef to remove compiler warning
---
include/asm-generic/io.h | 12 +++++
lib/iomap_copy.c | 109 +++++++++++++++++++++++++++++++++++++++
2 files changed, 121 insertions(+)
diff --git a/include/asm-generic/io.h b/include/asm-generic/io.h
index 80de699bf6af..9b8e0449da28 100644
--- a/include/asm-generic/io.h
+++ b/include/asm-generic/io.h
@@ -102,6 +102,18 @@ static inline void log_post_read_mmio(u64 val, u8 width, const volatile void __i
#endif /* CONFIG_TRACE_MMIO_ACCESS */
+#ifndef __memcpy_fromio
+void __memcpy_fromio(void *to, const volatile void __iomem *from, size_t count);
+#endif
+
+#ifndef __memcpy_toio
+void __memcpy_toio(volatile void __iomem *to, const void *from, size_t count);
+#endif
+
+#ifndef __memset_io
+void __memset_io(volatile void __iomem *dst, int c, size_t count);
+#endif
+
/*
* __raw_{read,write}{b,w,l,q}() access memory in native endianness.
*
diff --git a/lib/iomap_copy.c b/lib/iomap_copy.c
index 2fd5712fb7c0..c2cee6410151 100644
--- a/lib/iomap_copy.c
+++ b/lib/iomap_copy.c
@@ -3,9 +3,15 @@
* Copyright 2006 PathScale, Inc. All Rights Reserved.
*/
+#include <asm/unaligned.h>
+
+#include <linux/align.h>
#include <linux/export.h>
+#include <linux/types.h>
#include <linux/io.h>
+#define NATIVE_STORE_SIZE (BITS_PER_LONG/8)
+
/**
* __iowrite32_copy - copy data to MMIO space, in 32-bit units
* @to: destination, in MMIO space (must be 32-bit aligned)
@@ -76,3 +82,106 @@ void __iowrite64_copy(void __iomem *to, const void *from, size_t count)
}
EXPORT_SYMBOL_GPL(__iowrite64_copy);
#endif
+
+
+#ifndef __memcpy_fromio
+void __memcpy_fromio(void *to, const volatile void __iomem *from, size_t count)
+{
+ while (count && !IS_ALIGNED((unsigned long)from, NATIVE_STORE_SIZE)) {
+ *(u8 *)to = __raw_readb(from);
+ from++;
+ to++;
+ count--;
+ }
+
+ while (count >= NATIVE_STORE_SIZE) {
+#ifdef CONFIG_64BIT
+ put_unaligned(__raw_readq(from), (uintptr_t *)to);
+#else
+ put_unaligned(__raw_readl(from), (uintptr_t *)to);
+#endif
+
+ from += NATIVE_STORE_SIZE;
+ to += NATIVE_STORE_SIZE;
+ count -= NATIVE_STORE_SIZE;
+ }
+
+ while (count) {
+ *(u8 *)to = __raw_readb(from);
+ from++;
+ to++;
+ count--;
+ }
+}
+EXPORT_SYMBOL(__memcpy_fromio);
+#endif
+
+#ifndef __memcpy_toio
+void __memcpy_toio(volatile void __iomem *to, const void *from, size_t count)
+{
+ while (count && !IS_ALIGNED((unsigned long)to, NATIVE_STORE_SIZE)) {
+ __raw_writeb(*(u8 *)from, to);
+ from++;
+ to++;
+ count--;
+ }
+
+ while (count >= NATIVE_STORE_SIZE) {
+#ifdef CONFIG_64BIT
+ __raw_writeq(get_unaligned((uintptr_t *)from), to);
+#else
+ __raw_writel(get_unaligned((uintptr_t *)from), to);
+#endif
+
+ from += NATIVE_STORE_SIZE;
+ to += NATIVE_STORE_SIZE;
+ count -= NATIVE_STORE_SIZE;
+ }
+
+ while (count) {
+ __raw_writeb(*(u8 *)from, to);
+ from++;
+ to++;
+ count--;
+ }
+}
+EXPORT_SYMBOL(__memcpy_toio);
+#endif
+
+#ifndef __memset_io
+void __memset_io(volatile void __iomem *dst, int c, size_t count)
+{
+ uintptr_t qc = (u8)c;
+
+ qc |= qc << 8;
+ qc |= qc << 16;
+
+#ifdef CONFIG_64BIT
+ qc |= qc << 32;
+#endif
+
+ while (count && !IS_ALIGNED((unsigned long)dst, NATIVE_STORE_SIZE)) {
+ __raw_writeb(c, dst);
+ dst++;
+ count--;
+ }
+
+ while (count >= NATIVE_STORE_SIZE) {
+#ifdef CONFIG_64BIT
+ __raw_writeq(qc, dst);
+#else
+ __raw_writel(qc, dst);
+#endif
+
+ dst += NATIVE_STORE_SIZE;
+ count -= NATIVE_STORE_SIZE;
+ }
+
+ while (count) {
+ __raw_writeb(c, dst);
+ dst++;
+ count--;
+ }
+}
+EXPORT_SYMBOL(__memset_io);
+#endif
--
2.34.1From: Julian Vetter
> Sent: 25 September 2024 14:24
>
> Various architectures have almost the same implementations for
> __memcpy_{to,from}io and __memset_io functions. So, consolidate them
> into the existing lib/iomap_copy.c.
>
> Reviewed-by: Yann Sionneau <ysionneau@kalrayinc.com>
> Signed-off-by: Julian Vetter <jvetter@kalrayinc.com>
> ---
> Signed-off-by: Julian Vetter <jvetter@kalrayinc.com>
> ---
> Changes for v6:
> - Included linux/aslign.h
> - Replaced compile time check by ifdef to remove compiler warning
> ---
> include/asm-generic/io.h | 12 +++++
> lib/iomap_copy.c | 109 +++++++++++++++++++++++++++++++++++++++
> 2 files changed, 121 insertions(+)
>
> diff --git a/include/asm-generic/io.h b/include/asm-generic/io.h
> index 80de699bf6af..9b8e0449da28 100644
> --- a/include/asm-generic/io.h
> +++ b/include/asm-generic/io.h
> @@ -102,6 +102,18 @@ static inline void log_post_read_mmio(u64 val, u8 width, const volatile void __i
>
> #endif /* CONFIG_TRACE_MMIO_ACCESS */
>
> +#ifndef __memcpy_fromio
> +void __memcpy_fromio(void *to, const volatile void __iomem *from, size_t count);
> +#endif
> +
> +#ifndef __memcpy_toio
> +void __memcpy_toio(volatile void __iomem *to, const void *from, size_t count);
> +#endif
> +
> +#ifndef __memset_io
> +void __memset_io(volatile void __iomem *dst, int c, size_t count);
> +#endif
> +
> /*
> * __raw_{read,write}{b,w,l,q}() access memory in native endianness.
> *
> diff --git a/lib/iomap_copy.c b/lib/iomap_copy.c
> index 2fd5712fb7c0..c2cee6410151 100644
> --- a/lib/iomap_copy.c
> +++ b/lib/iomap_copy.c
> @@ -3,9 +3,15 @@
> * Copyright 2006 PathScale, Inc. All Rights Reserved.
> */
>
> +#include <asm/unaligned.h>
> +
> +#include <linux/align.h>
> #include <linux/export.h>
> +#include <linux/types.h>
> #include <linux/io.h>
>
> +#define NATIVE_STORE_SIZE (BITS_PER_LONG/8)
(sizeof (long))
> +
> /**
> * __iowrite32_copy - copy data to MMIO space, in 32-bit units
> * @to: destination, in MMIO space (must be 32-bit aligned)
> @@ -76,3 +82,106 @@ void __iowrite64_copy(void __iomem *to, const void *from, size_t count)
> }
> EXPORT_SYMBOL_GPL(__iowrite64_copy);
> #endif
> +
> +
> +#ifndef __memcpy_fromio
> +void __memcpy_fromio(void *to, const volatile void __iomem *from, size_t count)
> +{
> + while (count && !IS_ALIGNED((unsigned long)from, NATIVE_STORE_SIZE)) {
> + *(u8 *)to = __raw_readb(from);
> + from++;
> + to++;
> + count--;
> + }
> +
> + while (count >= NATIVE_STORE_SIZE) {
> +#ifdef CONFIG_64BIT
> + put_unaligned(__raw_readq(from), (uintptr_t *)to);
> +#else
> + put_unaligned(__raw_readl(from), (uintptr_t *)to);
> +#endif
That looks horrid to me.
You seem to be mixing several different types and tests.
NATIVE_STORE_SIZE is based on the 'long ' type (indirectly and by assumption).
CONFIG_64BiIT (probably) implies LP64.
readl() reads 4 bytes and readq() 8 (for both 32bit and 64bit kernels)
uintptr is an unsigned integer large enough to hold a pointer.
The sizes might all happen to match, but there is no need to rely on all of them.
I might be best to just use 'sizeof (long)' except that you might
get a compile error on some 32bit archs for the:
long val = sizeof (val) == 8) ? readq(from) : readl(from);
put_unaligned(val, (long *)to);
(due to there being no declaration readq())
so it might need a #if somewhere.
OTOH there might always be an 'extern' for readq().
If you are using the __raw_readx() functions don't you need the
synchronisation barriers top and bottom?
Also if put_unaligned() is non-trivial the code will be horrid.
An initial test for ((to | from) & (sizeof (long) - 1) == 0) for an
aligned copy may be worthwhile.
There is the question of whether the code is allowed to do full
word reads - valid if the io area behaves like memory.
In which case you don't want to do byte transfers for alignment
and tail transfers - just read the full word that contains the data.
PCIe reads can be horribly slow (writes are 'posted' so much better).
I'm not sure how long they take into a 'normal' target, but back to
back reads into our fpga are about 128 clocks apart on its internal
125Mhz clock - the host cpu will stall for the entire period.
So you definitely want to use the largest register possible.
(Or try very hard to never do non-dma reads in either direction.)
David
> +
> + from += NATIVE_STORE_SIZE;
> + to += NATIVE_STORE_SIZE;
> + count -= NATIVE_STORE_SIZE;
> + }
> +
> + while (count) {
> + *(u8 *)to = __raw_readb(from);
> + from++;
> + to++;
> + count--;
> + }
> +}
> +EXPORT_SYMBOL(__memcpy_fromio);
> +#endif
> +
> +#ifndef __memcpy_toio
> +void __memcpy_toio(volatile void __iomem *to, const void *from, size_t count)
> +{
> + while (count && !IS_ALIGNED((unsigned long)to, NATIVE_STORE_SIZE)) {
> + __raw_writeb(*(u8 *)from, to);
> + from++;
> + to++;
> + count--;
> + }
> +
> + while (count >= NATIVE_STORE_SIZE) {
> +#ifdef CONFIG_64BIT
> + __raw_writeq(get_unaligned((uintptr_t *)from), to);
> +#else
> + __raw_writel(get_unaligned((uintptr_t *)from), to);
> +#endif
> +
> + from += NATIVE_STORE_SIZE;
> + to += NATIVE_STORE_SIZE;
> + count -= NATIVE_STORE_SIZE;
> + }
> +
> + while (count) {
> + __raw_writeb(*(u8 *)from, to);
> + from++;
> + to++;
> + count--;
> + }
> +}
> +EXPORT_SYMBOL(__memcpy_toio);
> +#endif
> +
> +#ifndef __memset_io
> +void __memset_io(volatile void __iomem *dst, int c, size_t count)
> +{
> + uintptr_t qc = (u8)c;
> +
> + qc |= qc << 8;
> + qc |= qc << 16;
> +
> +#ifdef CONFIG_64BIT
> + qc |= qc << 32;
> +#endif
> +
> + while (count && !IS_ALIGNED((unsigned long)dst, NATIVE_STORE_SIZE)) {
> + __raw_writeb(c, dst);
> + dst++;
> + count--;
> + }
> +
> + while (count >= NATIVE_STORE_SIZE) {
> +#ifdef CONFIG_64BIT
> + __raw_writeq(qc, dst);
> +#else
> + __raw_writel(qc, dst);
> +#endif
> +
> + dst += NATIVE_STORE_SIZE;
> + count -= NATIVE_STORE_SIZE;
> + }
> +
> + while (count) {
> + __raw_writeb(c, dst);
> + dst++;
> + count--;
> + }
> +}
> +EXPORT_SYMBOL(__memset_io);
> +#endif
> --
> 2.34.1
>
>
>
>
>
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
On Wed, Sep 25, 2024, at 13:24, Julian Vetter wrote:
> Various architectures have almost the same implementations for
> __memcpy_{to,from}io and __memset_io functions. So, consolidate them
> into the existing lib/iomap_copy.c.
>
> Reviewed-by: Yann Sionneau <ysionneau@kalrayinc.com>
> Signed-off-by: Julian Vetter <jvetter@kalrayinc.com>
> ---
> Signed-off-by: Julian Vetter <jvetter@kalrayinc.com>
You have a duplicated signoff here.
> +#ifndef __memcpy_fromio
> +void __memcpy_fromio(void *to, const volatile void __iomem *from,
> size_t count);
> +#endif
> +
> +#ifndef __memcpy_toio
> +void __memcpy_toio(volatile void __iomem *to, const void *from, size_t
> count);
> +#endif
> +
> +#ifndef __memset_io
> +void __memset_io(volatile void __iomem *dst, int c, size_t count);
> +#endif
I'm not entirely sure about the purpose of the #ifdef here, since
nothing ever overrides the double-underscore versions, both before
and after your patches.
Unless I'm missing something here, I think a more logical
sequence would be:
1. add the definitions in this file without the underscores,
as memcpy_fromio/memcpy_toio/memset_io, with the #ifdef
for that name that is always set at this point
2. replace the default implementation in asm-generic/io.h
with extern prototypes, remove the #define from those
3. convert the other architectures, removing both the
implementations and the prototypes.
Arnd
On 26.09.24 09:14, Arnd Bergmann wrote:
> On Wed, Sep 25, 2024, at 13:24, Julian Vetter wrote:
>> Various architectures have almost the same implementations for
>> __memcpy_{to,from}io and __memset_io functions. So, consolidate them
>> into the existing lib/iomap_copy.c.
>>
>> Reviewed-by: Yann Sionneau <ysionneau@kalrayinc.com>
>> Signed-off-by: Julian Vetter <jvetter@kalrayinc.com>
>> ---
>> Signed-off-by: Julian Vetter <jvetter@kalrayinc.com>
>
> You have a duplicated signoff here.
Yes, thank you. I will remove it in the next patch revision.
>
>
>> +#ifndef __memcpy_fromio
>> +void __memcpy_fromio(void *to, const volatile void __iomem *from,
>> size_t count);
>> +#endif
>> +
>> +#ifndef __memcpy_toio
>> +void __memcpy_toio(volatile void __iomem *to, const void *from, size_t
>> count);
>> +#endif
>> +
>> +#ifndef __memset_io
>> +void __memset_io(volatile void __iomem *dst, int c, size_t count);
>> +#endif
>
> I'm not entirely sure about the purpose of the #ifdef here, since
> nothing ever overrides the double-underscore versions, both before
> and after your patches.
>
> Unless I'm missing something here, I think a more logical
> sequence would be:
>
> 1. add the definitions in this file without the underscores,
by: "...in this file..." you mean the 'lib/iomap_copy.c' file, right?
But what if an architecture does not select 'CONFIG_HAS_IOMEM'. Then
'iomap_copy.c' is not compiled and we don't have an implementation,
right? I tried to compile with ARCH=um, with some MTD chip driver, like
the robot did and it indeed fails, because um has 'NO_IOMEM' set. and
the driver uses memcpy_fromio. I mean it's a strange combination,
because apparently we try to use IO memory? Is this an invalid
combination? But shouldn't the driver then 'depends on HAS_IOMEM'?
> as memcpy_fromio/memcpy_toio/memset_io, with the #ifdef
> for that name that is always set at this point
>
Right. I will remove it in my next patch revision.
> 2. replace the default implementation in asm-generic/io.h
> with extern prototypes, remove the #define from those
>
Yes, I have done this now.
> 3. convert the other architectures, removing both the
> implementations and the prototypes.
>
I have removed the prototypes and have aligned the function arguments in
m68k, alpha, parisc, and sh, which all have their own implementation,
but had slightly different function arguments. Btw, I have not removed
their implementations because some of them seem to have optimized
implementations (e.g., alpha and m68k), that I didn't want to touch. But
you're right others (e.g., sh) just do byte wise accesses and have a
comment "This needs to be optimized." Maybe I should remove these and
let them use the new version?!
> Arnd
>
>
>
>
On Fri, Sep 27, 2024, at 08:19, Julian Vetter wrote:
> On 26.09.24 09:14, Arnd Bergmann wrote:
>>> +#ifndef __memcpy_fromio
>>> +void __memcpy_fromio(void *to, const volatile void __iomem *from,
>>> size_t count);
>>> +#endif
>>> +
>>> +#ifndef __memcpy_toio
>>> +void __memcpy_toio(volatile void __iomem *to, const void *from, size_t
>>> count);
>>> +#endif
>>> +
>>> +#ifndef __memset_io
>>> +void __memset_io(volatile void __iomem *dst, int c, size_t count);
>>> +#endif
>>
>> I'm not entirely sure about the purpose of the #ifdef here, since
>> nothing ever overrides the double-underscore versions, both before
>> and after your patches.
>>
>> Unless I'm missing something here, I think a more logical
>> sequence would be:
>>
>> 1. add the definitions in this file without the underscores,
>
> by: "...in this file..." you mean the 'lib/iomap_copy.c' file, right?
Yes
> But what if an architecture does not select 'CONFIG_HAS_IOMEM'. Then
> 'iomap_copy.c' is not compiled and we don't have an implementation,
> right?
> I tried to compile with ARCH=um, with some MTD chip driver, like
> the robot did and it indeed fails, because um has 'NO_IOMEM' set. and
> the driver uses memcpy_fromio. I mean it's a strange combination,
> because apparently we try to use IO memory? Is this an invalid
> combination? But shouldn't the driver then 'depends on HAS_IOMEM'?
Yes, I think that would be the best way to do it. Alternatively,
arch/um could provide a dummy implementation of these.
>> 3. convert the other architectures, removing both the
>> implementations and the prototypes.
>>
>
> I have removed the prototypes and have aligned the function arguments in
> m68k, alpha, parisc, and sh, which all have their own implementation,
> but had slightly different function arguments.
Sorry for being unclear, I meant only the architectures that
you are already touching.
> Btw, I have not removed
> their implementations because some of them seem to have optimized
> implementations (e.g., alpha and m68k), that I didn't want to touch. But
> you're right others (e.g., sh) just do byte wise accesses and have a
> comment "This needs to be optimized." Maybe I should remove these and
> let them use the new version?!
Ideally we should end up with only one copy, but I'd leave the
rest for a future cleanup. In particular, alpha probably still
needs a custom function.
Arnd
© 2016 - 2026 Red Hat, Inc.