dw_dmac: make driver endianness configurable
PATCH 1/2 dw_dmac: make driver endianness configurable
The dw_dmac was originally developed for avr32 to be used with the Synopsys DesignWare AHB DMA controller. After 2.6.38, device access was done with the little-endian readl/writel functions. This didn't work on the avr32 platform, because it needs native-endian (i.e. big-endian) accessors. This patch makes the endianness configurable using 'DW_DMAC_BE', which will default be true for AVR32
--- drivers/dma/Kconfig | 8 ++++++++ drivers/dma/dw_dmac_regs.h | 23 +++++++++++++++++++++++ 2 files changed, 31 insertions(+), 0 deletions(-)
drivers/dma/Kconfig +8 -0
@@ -89,6 +89,14 @@ config DW_DMAC Support the Synopsys DesignWare AHB DMA controller. This can be integrated in chips such as the Atmel AT32ap7000. +config DW_DMAC_BE + bool "Synopsys DesignWare AHB DMA needs big endian access" + default y if AVR32 + depends on DW_DMAC + help + Say yes if access to the Synopsys DesignWare AHB DMA controller + should be big endian, such as for Atmel AT32ap7000 + config AT_HDMAC tristate "Atmel AHB DMA support" depends on ARCH_AT91
drivers/dma/dw_dmac_regs.h +22 -0
@@ -175,11 +175,22 @@ __dwc_regs(struct dw_dma_chan *dwc)return dwc->ch_regs; } +#ifdef CONFIG_DW_DMAC_BE + +#define channel_readl (dwc, name) \ + ioread32be(&(__dwc_regs(dwc)->name)) +#define channel_writel (dwc, name, val) \ + iowrite32be((val), &(__dwc_regs(dwc)->name)) + +#else +#define channel_readl (dwc, name) \ readl(&(__dwc_regs(dwc)->name))#define channel_writel (dwc, name, val) \ writel((val), &(__dwc_regs(dwc)->name)) +#endif +static inline struct dw_dma_chan *to_dw_dma_chan (struct dma_chan *chan) { return container_of(chan, struct dw_dma_chan, chan);
@@ -201,11 +212,23 @@ static inline struct dw_dma_regs __iomem *__dw_regs(struct dw_dma *dw)return dw->regs; } +#ifdef CONFIG_DW_DMAC_BE + +#define dma_readl (dwc, name) \ + ioread32be(&(__dw_regs(dw)->name)) +#define dma_writel (dwc, name, val) \ + iowrite32be((val), &(__dw_regs(dw)->name)) + +#else +#define dma_readl (dw, name) \ readl(&(__dw_regs(dw)->name))#define dma_writel (dw, name, val) \ writel((val), &(__dw_regs(dw)->name)) +#endif +#define channel_set_bit (dw, reg, mask) \ dma_writel(dw, reg, ((mask) << 8) | (mask))#define channel_clear_bit (dw, reg, mask) \
Signature · 1 line
1.7.8.0
Quoted 6 lines · Hein Tibosch
The
dw_dmacwas originally developed for avr32 to be used with the Synopsys DesignWare AHB DMA controller. After 2.6.38, device access was done with the little-endian readl/writel functions. This didn't work on the avr32 platform, because it needs native-endian (i.e. big-endian) accessors. This patch makes the endianness configurable using 'DW_DMAC_BE', which will default be true for AVR32
I think the English in kconfig could use some brushing up.
Quoted 11 lines
Signed-off-by Hein Tibosch <[email protected]>--- drivers/dma/Kconfig | 8 ++++++++ drivers/dma/dw_dmac_regs.h | 23 +++++++++++++++++++++++ 2 files changed, 31 insertions(+), 0 deletions(-)drivers/dma/Kconfig +0
-0
drivers/dma/Kconfig config DW_DMAC · line 93
Quoted 5 lines
drivers/dma/Kconfig +1
-0@@ -89,6 +89,14 @@ config DW_DMAC Support the Synopsys DesignWare AHB DMA controller. This can be integrated in chips such as the Atmel AT32ap7000. +config DW_DMAC_BE
This name isn't that long, so we could skip the abbreviation of big endian; DW_DMAC_BIG_ENDIAN_IO or something similar?
drivers/dma/Kconfig config DW_DMAC · line 95
Quoted 2 lines
drivers/dma/Kconfig +1
-0+ bool "Synopsys DesignWare AHB DMA needs big endian access"
bool "Use big endian I/O register access"
drivers/dma/Kconfig config DW_DMAC · line 99
Quoted 6 lines
drivers/dma/Kconfig +6
-0+ default y if AVR32 + depends on DW_DMAC + help + Say yes if access to the Synopsys DesignWare AHB DMA controller + should be big endian, such as for Atmel AT32ap7000 +
"Say yes here to use big endian I/O access when reading and writing to the DMA controller registers. This is needed on some platforms, like the Atmel AVR32 architecture.
If unsure, use the default setting."
<snipp patch diff>
Signature · 2 lines
mvh Hans-Christian Egtvedt
Quoted 11 lines · Hans-Christian Egtvedt
I think the English in kconfig could use some brushing up.
Quoted 1 line
+config
DW_DMAC_BEThis name isn't that long, so we could skip the abbreviation of big endian;
DW_DMAC_BIG_ENDIAN_IOor something similar?drivers/dma/Kconfig config DW_DMAC · line 93
Quoted 2 lines
drivers/dma/Kconfig +1
-0+ bool "Synopsys DesignWare AHB DMA needs big endian access"bool "Use big endian I/O register access"
Brushing up the config items:
+config DW_DMAC_BIG_ENDIAN_IO
+ bool "Use big endian I/O register access"
+ default y if AVR32
+ depends on DW_DMAC
+ help
+ Say yes here to use big endian I/O access when reading and writing + to the DMA controller registers. This is needed on some platforms, + like the Atmel AVR32 architecture.
+
+ If unsure, use the default setting.
And as I'd like to define the maximum memory transfer width in the same Kconfig:
+config DW_DMAC_MEM_64_BIT
+ bool "Allow 64-bit memory transfers"
+ default y if !AVR32
+ depends on DW_DMAC
+ help
+ Say yes if the DMA controller may do 64-bit memory transfers + For AVR32, say no because only up to 32-bit transfers are + defined
Would this be better?
Thanks, Hein
Quoted 15 lines · Hein Tibosch
Quoted 11 lines · Hans-Christian Egtvedt
I think the English in kconfig could use some brushing up.
Quoted 1 line
+config
DW_DMAC_BEThis name isn't that long, so we could skip the abbreviation of big endian;
DW_DMAC_BIG_ENDIAN_IOor something similar?drivers/dma/Kconfig config DW_DMAC · line 93
Quoted 2 lines
drivers/dma/Kconfig +1
-0+ bool "Synopsys DesignWare AHB DMA needs big endian access"bool "Use big endian I/O register access"
Brushing up the config items:
Review
Quoted 10 lines
(quoted diff) +10
-0+config DW_DMAC_BIG_ENDIAN_IO + bool "Use big endian I/O register access" + default y if AVR32 + depends on DW_DMAC + help + Say yes here to use big endian I/O access when reading and writing + to the DMA controller registers. This is needed on some platforms, + like the Atmel AVR32 architecture. + + If unsure, use the default setting.
This sounds good in my ears, but I don't speak English natively.
Review
Quoted 11 lines
And as I'd like to define the maximum memory transfer width in the same
Kconfig:(quoted diff) +8
-0+config DW_DMAC_MEM_64_BIT + bool "Allow 64-bit memory transfers" + default y if !AVR32 + depends on DW_DMAC + help + Say yes if the DMA controller may do 64-bit memory transfers + For AVR32, say no because only up to 32-bit transfers are + defined
Is this sane to add? Could some non-AVR32 platforms use 64-bit and 32-bit depending on runtime configuration? E.g. if you build a kernel with support for multiple boards/processors, and there is a mix of 32-bit and 64-bit wide DMA support.
I think it is better to select 32/64-bit at runtime.
Signature · 2 lines
mvh Hans-Christian Egtvedt
Quoted 4 lines · Hans-Christian Egtvedt
Quoted 3 lines · Hein Tibosch
Quoted 2 lines · Hans-Christian Egtvedt
Brushing up the config items:
Quoted 11 lines
Quoted 10 lines
Review
Quoted 10 lines
(quoted diff) +10
-0+config DW_DMAC_BIG_ENDIAN_IO + bool "Use big endian I/O register access" + default y if AVR32 + depends on DW_DMAC + help + Say yes here to use big endian I/O access when reading and writing + to the DMA controller registers. This is needed on some platforms, + like the Atmel AVR32 architecture. + + If unsure, use the default setting.This sounds good in my ears, but I don't speak English natively.
I think you Norwegians are doing very well in English
And btw, I'm not from .es but from .nl, which is very close to England
Quoted 16 lines
Review
Quoted 11 lines
And as I'd like to define the maximum memory transfer width in the same
Kconfig:(quoted diff) +8
-0+config DW_DMAC_MEM_64_BIT + bool "Allow 64-bit memory transfers" + default y if !AVR32 + depends on DW_DMAC + help + Say yes if the DMA controller may do 64-bit memory transfers + For AVR32, say no because only up to 32-bit transfers are + definedIs this sane to add? Could some non-AVR32 platforms use 64-bit and 32-bit depending on runtime configuration? E.g. if you build a kernel with support for multiple boards/processors, and there is a mix of 32-bit and 64-bit wide DMA support.
Quoted 1 line
I think it is better to select 32/64-bit at runtime.
I did that in the first patch, adding a new property to the dw_dma_slave structure. It had the small disadvantage that some arch code had to be adapted (at32ap700x.c).
Viresh, what do you think? Add a property called "mem_64_bit_access" or so?
Or should it be passed as a member of 'dw_dma_platform_data', because it is a property of the (entire) DMA controller?
Hein
Quoted 15 lines · Hein Tibosch
Quoted 8 lines
Review
Quoted 8 lines
(quoted diff) +8
-0+config DW_DMAC_MEM_64_BIT + bool "Allow 64-bit memory transfers" + default y if !AVR32 + depends on DW_DMAC + help + Say yes if the DMA controller may do 64-bit memory transfers + For AVR32, say no because only up to 32-bit transfers are + definedQuoted 6 lines
Is this sane to add? Could some non-AVR32 platforms use 64-bit and 32-bit depending on runtime configuration? E.g. if you build a kernel with support for multiple boards/processors, and there is a mix of 32-bit and 64-bit wide DMA support.
I think it is better to select 32/64-bit at runtime.
Quoted 8 lines
I did that in the first patch, adding a new property to the
dw_dma_slavestructure. It had the small disadvantage that some arch code had to be adapted (at32ap700x.c).Viresh, what do you think? Add a property called "
mem_64_bit_access" or so?Or should it be passed as a member of '
dw_dma_platform_data', because it is a property of the (entire) DMA controller?
I think second option is better. But can there be some supportive scenarios of first option?
We have a system, with two different memory controllers, one supporting 32 bit and other 64 bit?
Or what we can do now is: go with option 2, i.e. update dw_dma_platform_data and if some platform like what i mentioned above comes, then we can move it to slave data.
viresh
Quoted 4 lines · Hein Tibosch
What about this:
In case of AVR32, the arch code will indicate:
Quoted 8 lines
static struct dw_dma_platform_data dw_dmac0_data = { .nr_channels = 3, /* DMAC supports up to 32-bit memory access */ .mem_access_32_bit_only = true, };ARM users won't have to change anything because
mem_access_32_bit_onlywill be false and therefor 'dw->mem_64_bit' will become true
I will go for the earlier solution that you sent: with max-mem-width. 0 = 64 bit, so nothing to be changed for ARM as global structures would be getting initialized to zero
1=32 bit, pass this for AVR32.
--
viresh
Quoted 26 lines · Viresh Kumar
Quoted 14 lines · Hein Tibosch
Quoted 8 lines
Review
Quoted 8 lines
(quoted diff) +8
-0+config DW_DMAC_MEM_64_BIT + bool "Allow 64-bit memory transfers" + default y if !AVR32 + depends on DW_DMAC + help + Say yes if the DMA controller may do 64-bit memory transfers + For AVR32, say no because only up to 32-bit transfers are + definedQuoted 6 lines
Is this sane to add? Could some non-AVR32 platforms use 64-bit and 32-bit depending on runtime configuration? E.g. if you build a kernel with support for multiple boards/processors, and there is a mix of 32-bit and 64-bit wide DMA support.
I think it is better to select 32/64-bit at runtime.
Quoted 8 lines
I did that in the first patch, adding a new property to the
dw_dma_slavestructure. It had the small disadvantage that some arch code had to be adapted (at32ap700x.c).Viresh, what do you think? Add a property called "
mem_64_bit_access" or so?Or should it be passed as a member of '
dw_dma_platform_data', because it is a property of the (entire) DMA controller?I think second option is better. But can there be some supportive scenarios of first option?
Quoted 6 lines
We have a system, with two different memory controllers, one supporting 32 bit and other 64 bit?
Or what we can do now is: go with option 2, i.e. update
dw_dma_platform_dataand if some platform like what i mentioned above comes, then we can move it to slave data.
What about this:
In case of AVR32, the arch code will indicate:
static struct dw_dma_platform_data dw_dmac0_data = {
.nr_channels = 3,
/* DMAC supports up to 32-bit memory access */
.mem_access_32_bit_only = true,
};ARM users won't have to change anything because mem_access_32_bit_only will be false and therefor 'dw->mem_64_bit' will become true
--- drivers/dma/dw_dmac.c | 11 ++++++++--- drivers/dma/dw_dmac_regs.h | 2 ++ include/linux/dw_dmac.h | 2 ++ 3 files changed, 12 insertions(+), 3 deletions(-)
drivers/dma/dw_dmac.c +8 -3
@@ -618,6 +618,7 @@ dwc_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src, size_t len, unsigned long flags) {struct dw_dma_chan *dwc =to_dw_dma_chan (chan); +struct dw_dma *dw =to_dw_dma (dwc->chan.device);struct dw_desc *desc;struct dw_desc *first;struct dw_desc *prev;
@@ -639,7 +640,7 @@ dwc_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,* We can be a lot more clever here, but this should take care * of the most common optimization. */ -if (!((src | dest | len) &7 )) +if (dw->mem_64_bit && !((src | dest | len) &7 )) src_width = dst_width =3 ;else if (!((src | dest | len) &3 )) src_width = dst_width =2 ;
@@ -710,6 +711,7 @@ dwc_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl,struct dw_dma_chan *dwc =to_dw_dma_chan (chan);struct dw_dma_slave *dws = chan->private;struct dma_slave_config *sconfig = &dwc->dma_sconfig; +struct dw_dma *dw =to_dw_dma (dwc->chan.device);struct dw_desc *prev;struct dw_desc *first;u32 ctllo;
@@ -746,7 +748,7 @@ dwc_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl, mem =sg_dma_address (sg); len =sg_dma_len (sg); -if (!((mem | len) &7 )) +if (dw->mem_64_bit && !((mem | len) &7 )) mem_width =3 ;else if (!((mem | len) &3 )) mem_width =2 ;
@@ -813,7 +815,7 @@ slave_sg_todev_fill_desc: mem =sg_dma_address (sg); len =sg_dma_len (sg); -if (!((mem | len) &7 )) +if (dw->mem_64_bit && !((mem | len) &7 )) mem_width =3 ;else if (!((mem | len) &3 )) mem_width =2 ;
@@ -1419,6 +1421,9 @@ static int __init dw_probe(struct platform_device *pdev) goto err_kfree; } +/* Remember if 64-bit access to memory is allowed */ + dw->mem_64_bit = !pdata->mem_access_32_bit_only; + dw->regs =ioremap (io->start,DW_REGLEN );if (!dw->regs) { err = -ENOMEM ;
drivers/dma/dw_dmac_regs.h +2 -0
@@ -199,6 +199,8 @@ struct dw_dma {struct clk *clk;u8 all_chan_mask; +/* 64-bit access to memory is allowed */ +bool mem_64_bit;struct dw_dma_chan chan[0 ]; };
include/linux/dw_dmac.h +2 -0
@@ -29,6 +29,8 @@ struct dw_dma_platform_data {#define CHAN_PRIORITY_ASCENDING 0/* chan0 highest */ #define CHAN_PRIORITY_DESCENDING 1/* chan7 highest */ unsigned char chan_priority; +/* Make true if 64-bit access to memory is not implemented */ +bool mem_access_32_bit_only; };/* bursts size */
Signature · 1 line
1.7.8.0
Quoted 129 lines · Hein Tibosch
Quoted 26 lines · Viresh Kumar
Quoted 14 lines · Hein Tibosch
Quoted 8 lines
Review
Quoted 8 lines
(quoted diff) +8
-0+config DW_DMAC_MEM_64_BIT + bool "Allow 64-bit memory transfers" + default y if !AVR32 + depends on DW_DMAC + help + Say yes if the DMA controller may do 64-bit memory transfers + For AVR32, say no because only up to 32-bit transfers are + definedQuoted 6 lines
Is this sane to add? Could some non-AVR32 platforms use 64-bit and 32-bit depending on runtime configuration? E.g. if you build a kernel with support for multiple boards/processors, and there is a mix of 32-bit and 64-bit wide DMA support.
I think it is better to select 32/64-bit at runtime.
Quoted 8 lines
I did that in the first patch, adding a new property to the
dw_dma_slavestructure. It had the small disadvantage that some arch code had to be adapted (at32ap700x.c).Viresh, what do you think? Add a property called "
mem_64_bit_access" or so?Or should it be passed as a member of '
dw_dma_platform_data', because it is a property of the (entire) DMA controller?I think second option is better. But can there be some supportive scenarios of first option?
Quoted 6 lines
We have a system, with two different memory controllers, one supporting 32 bit and other 64 bit?
Or what we can do now is: go with option 2, i.e. update
dw_dma_platform_dataand if some platform like what i mentioned above comes, then we can move it to slave data.What about this:
In case of AVR32, the arch code will indicate:
static struct dw_dma_platform_data dw_dmac0_data = { .nr_channels = 3, /* DMAC supports up to 32-bit memory access */ .mem_access_32_bit_only = true, };ARM users won't have to change anything because
mem_access_32_bit_onlywill be false and therefor 'dw->mem_64_bit' will become trueSigned-off-by Hein Tibosch <[email protected]>--- drivers/dma/dw_dmac.c | 11 ++++++++--- drivers/dma/dw_dmac_regs.h | 2 ++ include/linux/dw_dmac.h | 2 ++ 3 files changed, 12 insertions(+), 3 deletions(-)drivers/dma/dw_dmac.c +8
-3@@ -618,6 +618,7 @@ dwc_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src, size_t len, unsigned long flags) {struct dw_dma_chan *dwc =to_dw_dma_chan (chan); +struct dw_dma *dw =to_dw_dma (dwc->chan.device);struct dw_desc *desc;struct dw_desc *first;struct dw_desc *prev;@@ -639,7 +640,7 @@ dwc_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,* We can be a lot more clever here, but this should take care * of the most common optimization. */ -if (!((src | dest | len) &7 )) +if (dw->mem_64_bit && !((src | dest | len) &7 )) src_width = dst_width =3 ;else if (!((src | dest | len) &3 )) src_width = dst_width =2 ;@@ -710,6 +711,7 @@ dwc_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl,struct dw_dma_chan *dwc =to_dw_dma_chan (chan);struct dw_dma_slave *dws = chan->private;struct dma_slave_config *sconfig = &dwc->dma_sconfig; +struct dw_dma *dw =to_dw_dma (dwc->chan.device);struct dw_desc *prev;struct dw_desc *first;u32 ctllo;@@ -746,7 +748,7 @@ dwc_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl, mem =sg_dma_address (sg); len =sg_dma_len (sg); -if (!((mem | len) &7 )) +if (dw->mem_64_bit && !((mem | len) &7 )) mem_width =3 ;else if (!((mem | len) &3 )) mem_width =2 ;@@ -813,7 +815,7 @@ slave_sg_todev_fill_desc: mem =sg_dma_address (sg); len =sg_dma_len (sg); -if (!((mem | len) &7 )) +if (dw->mem_64_bit && !((mem | len) &7 )) mem_width =3 ;else if (!((mem | len) &3 )) mem_width =2 ;@@ -1419,6 +1421,9 @@ static int __init dw_probe(struct platform_device *pdev) goto err_kfree; } +/* Remember if 64-bit access to memory is allowed */ + dw->mem_64_bit = !pdata->mem_access_32_bit_only; + dw->regs =ioremap (io->start,DW_REGLEN );if (!dw->regs) { err = -ENOMEM ;drivers/dma/dw_dmac_regs.h +2
-0@@ -199,6 +199,8 @@ struct dw_dma {struct clk *clk;u8 all_chan_mask; +/* 64-bit access to memory is allowed */ +bool mem_64_bit;struct dw_dma_chan chan[0 ]; };include/linux/dw_dmac.h +0
-0
include/linux/dw_dmac.h struct dw_dma_platform_data { · line 34
Quoted 6 lines
include/linux/dw_dmac.h +2
-0@@ -29,6 +29,8 @@ struct dw_dma_platform_data {#define CHAN_PRIORITY_ASCENDING 0/* chan0 highest */ #define CHAN_PRIORITY_DESCENDING 1/* chan7 highest */ unsigned char chan_priority; +/* Make true if 64-bit access to memory is not implemented */ +bool mem_access_32_bit_only; };
Can you not read this from any internal register ? Synopsys generally adds a set of read only registers which we can use to guess which features where enabled in the IP when configuring it with their IP configuration tool.
Signature · 1 line
balbi
Quoted 4 lines · Felipe Balbi
Can you not read this from any internal register ? Synopsys generally adds a set of read only registers which we can use to guess which features where enabled in the IP when configuring it with their IP configuration tool.
You right. I have a patch to support few options like that already. However, I can't share
it yet by some legal reasons.
Signature · 2 lines
With Best Regards, Andy Shevchenko
Quoted 8 lines · Andy Shevchenko
Quoted 4 lines · Felipe Balbi
Can you not read this from any internal register ? Synopsys generally adds a set of read only registers which we can use to guess which features where enabled in the IP when configuring it with their IP configuration tool.
You right. I have a patch to support few options like that already. However, I can't share
it yet by some legal reasons.
fair enough, in that case it'd be best to hold this patch and provide something better later ;-)
Signature · 1 line
balbi