Skip to contentSkip to conversation

ns16550: Add support for AUX regs usage on some ARC SoCs

RFC In review Alexey Brodkin 3 participants 11 messages Feb 21, 2018 – Mar 8, 2018

lore
Share & tools
lore
#0 lore raw

RFC 1/1 ns16550: Add support for AUX regs usage on some ARC SoCs

Synopsys Data Fusion subsystem (DFSS) is targeted to deeply built-in use-cases and so to save some silicon area decision was made to escape usage of any busses and use instead directly wired to CPU peripherals. And one of those is DW APB UART.

Later DFSS became a part of larger and more complicated SoCs with some other peripherals connected via common buses but default UART is still used via ARC core's auxulary registers which are not mapped to "normal" address space and we use very special instructions to access them, thus we cannot simply reuse whatever accessor we have in 16550 UART as of now.

Also we cannot just switch inb()/outb() to access ARC AUX regs always for DFSS because other peripherals have normal memory-mapped control registers and we need to use normal accessors for them.

Frankly I don't like a lot what I did here but otherwise if I create a special driver for this I'll need to reimplement ns16550_serial_ops.putc()/getc() which will be pure copy-paseted from ns16550.c because the only difference is only in ns16550_{read|write}b().

As mentioned above we cannot remap those auxiliary registers to normal memory address-space and thus we have to use very special accessors write_aux_reg()/read_aux_reg() that directly use special CPU intructions (namely LR/SR) for dealing with ARC AUX regs.

Also note here I just use a check for a particular SoC being selected (CONFIG_ARCH_DFSS will be introduced shortly) but that is done just for simplicity, otherwise it might be a slecial Kconfig option for NS16550 or anything else.

I'd like to know what people think about possible colutions here. And as always any comments are much appreciated!

Cc Simon Glass <[email protected]>
Cc Tom Rini <[email protected]>
Cc Stefan Roese <[email protected]>
---
 drivers/serial/ns16550.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
drivers/serial/ns16550.c +9 -2
index 6f9ce689cfff..75f342f337f1 100644
@@ -13,6 +13,9 @@
 #include <watchdog.h>
 #include <linux/types.h>
 #include <asm/io.h>
+#ifdef CONFIG_ARCH_DFSS
+#include <asm/arcregs.h>
+#endif
 
 DECLARE_GLOBAL_DATA_PTR;
 
@@ -53,7 +56,9 @@ DECLARE_GLOBAL_DATA_PTR;
 
 static inline void serial_out_shift(void *addr, int shift, int value)
 {
-#ifdef CONFIG_SYS_NS16550_PORT_MAPPED
+#ifdef CONFIG_ARCH_DFSS
+	write_aux_reg((int)addr, value);
+#elif defined(CONFIG_SYS_NS16550_PORT_MAPPED)
 	outb(value, (ulong)addr);
 #elif defined(CONFIG_SYS_NS16550_MEM32) && !defined(CONFIG_SYS_BIG_ENDIAN)
 	out_le32(addr, value);
@@ -70,7 +75,9 @@ static inline void serial_out_shift(void *addr, int shift, int value)
 
 static inline int serial_in_shift(void *addr, int shift)
 {
-#ifdef CONFIG_SYS_NS16550_PORT_MAPPED
+#ifdef CONFIG_ARCH_DFSS
+	return read_aux_reg((int)addr);
+#elif defined(CONFIG_SYS_NS16550_PORT_MAPPED)
 	return inb((ulong)addr);
 #elif defined(CONFIG_SYS_NS16550_MEM32) && !defined(CONFIG_SYS_BIG_ENDIAN)
 	return in_le32(addr);
Signature · 1 line
2.14.3
#1 lore raw
Quoted 69 lines · Alexey Brodkin

Synopsys Data Fusion subsystem (DFSS) is targeted to deeply built-in use-cases and so to save some silicon area decision was made to escape usage of any busses and use instead directly wired to CPU peripherals. And one of those is DW APB UART.

Later DFSS became a part of larger and more complicated SoCs with some other peripherals connected via common buses but default UART is still used via ARC core's auxulary registers which are not mapped to "normal" address space and we use very special instructions to access them, thus we cannot simply reuse whatever accessor we have in 16550 UART as of now.

Also we cannot just switch inb()/outb() to access ARC AUX regs always for DFSS because other peripherals have normal memory-mapped control registers and we need to use normal accessors for them.

Frankly I don't like a lot what I did here but otherwise if I create a special driver for this I'll need to reimplement ns16550_serial_ops.putc()/getc() which will be pure copy-paseted from ns16550.c because the only difference is only in ns16550_{read|write}b().

As mentioned above we cannot remap those auxiliary registers to normal memory address-space and thus we have to use very special accessors write_aux_reg()/read_aux_reg() that directly use special CPU intructions (namely LR/SR) for dealing with ARC AUX regs.

Also note here I just use a check for a particular SoC being selected (CONFIG_ARCH_DFSS will be introduced shortly) but that is done just for simplicity, otherwise it might be a slecial Kconfig option for NS16550 or anything else.

I'd like to know what people think about possible colutions here. And as always any comments are much appreciated!

Cc Simon Glass <[email protected]>
Cc Tom Rini <[email protected]>
Cc Stefan Roese <[email protected]>
---
 drivers/serial/ns16550.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
drivers/serial/ns16550.c +6 -1
index 6f9ce689cfff..75f342f337f1 100644
@@ -13,6 +13,9 @@
 #include <watchdog.h>
 #include <linux/types.h>
 #include <asm/io.h>
+#ifdef CONFIG_ARCH_DFSS
+#include <asm/arcregs.h>
+#endif
 
 DECLARE_GLOBAL_DATA_PTR;
 
@@ -53,7 +56,9 @@ DECLARE_GLOBAL_DATA_PTR;
 
 static inline void serial_out_shift(void *addr, int shift, int value)
 {
-#ifdef CONFIG_SYS_NS16550_PORT_MAPPED
+#ifdef CONFIG_ARCH_DFSS
+	write_aux_reg((int)addr, value);
+#elif defined(CONFIG_SYS_NS16550_PORT_MAPPED)
 	outb(value, (ulong)addr);
 #elif defined(CONFIG_SYS_NS16550_MEM32) && !defined(CONFIG_SYS_BIG_ENDIAN)
 	out_le32(addr, value);

drivers/serial/ns16550.c static inline void serial_out_shift(void *addr, int shift, int value) · line 83

Quoted 10 lines
drivers/serial/ns16550.c +3 -1
@@ -70,7 +75,9 @@ static inline void serial_out_shift(void *addr, int shift, int value)
 
 static inline int serial_in_shift(void *addr, int shift)
 {
-#ifdef CONFIG_SYS_NS16550_PORT_MAPPED
+#ifdef CONFIG_ARCH_DFSS
+	return read_aux_reg((int)addr);
+#elif defined(CONFIG_SYS_NS16550_PORT_MAPPED)
 	return inb((ulong)addr);
 #elif defined(CONFIG_SYS_NS16550_MEM32) && !defined(CONFIG_SYS_BIG_ENDIAN)
 	return in_le32(addr);

As always, thanks for the detailed explanation. Yes, I think that of the options, putting the details in read/write_aux_reg (and please make sure read/write_aux_reg have a good function comment too) is the best choice. Thanks!

Signature · 8 lines
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <http://lists.denx.de/pipermail/u-boot/attachments/20180222/c738c4cb/attachment.sig>
#2 lore raw

Hi Tom,

Quoted 1 line · Tom Rini

On Wed, Feb 21, 2018 at 03:26:05PM +0300, Alexey Brodkin wrote:

[snip]

Quoted 11 lines
Quoted 11 lines

static inline void serial_out_shift(void *addr, int shift, int value) {

drivers/serial/ns16550.c +3 -1
-#ifdef CONFIG_SYS_NS16550_PORT_MAPPED
+#ifdef CONFIG_ARCH_DFSS
+	write_aux_reg((int)addr, value);
+#elif defined(CONFIG_SYS_NS16550_PORT_MAPPED)
 	outb(value, (ulong)addr);
 #elif defined(CONFIG_SYS_NS16550_MEM32) && !defined(CONFIG_SYS_BIG_ENDIAN)
 	out_le32(addr, value);
@@ -70,7 +75,9 @@ static inline void serial_out_shift(void *addr, int shift, int value)
 
Quoted 14 lines

drivers/serial/ns16550.c static inline void serial_out_shift(void *addr, int shift, int value) · line 83

Quoted 9 lines
drivers/serial/ns16550.c +3 -1
 static inline int serial_in_shift(void *addr, int shift)
 {
-#ifdef CONFIG_SYS_NS16550_PORT_MAPPED
+#ifdef CONFIG_ARCH_DFSS
+	return read_aux_reg((int)addr);
+#elif defined(CONFIG_SYS_NS16550_PORT_MAPPED)
 	return inb((ulong)addr);
 #elif defined(CONFIG_SYS_NS16550_MEM32) && !defined(CONFIG_SYS_BIG_ENDIAN)
 	return in_le32(addr);

As always, thanks for the detailed explanation. Yes, I think that of the options, putting the details in read/write_aux_reg (and please make sure read/write_aux_reg have a good function comment too) is the best choice. Thanks!

Frankly I didn't understand your comment well enough :)
Do you suggest to keep proposed implementation (i.e. modification of serial_{in|out}_shift()) but add comments on what do we do and why or you really meant something completely different?

-Alexey

#3 lore raw
Quoted 33 lines · Alexey Brodkin

Hi Tom,

Quoted 1 line · Tom Rini

On Wed, Feb 21, 2018 at 03:26:05PM +0300, Alexey Brodkin wrote:

[snip]

Quoted 11 lines
Quoted 11 lines

static inline void serial_out_shift(void *addr, int shift, int value) {

drivers/serial/ns16550.c +3 -1
-#ifdef CONFIG_SYS_NS16550_PORT_MAPPED
+#ifdef CONFIG_ARCH_DFSS
+	write_aux_reg((int)addr, value);
+#elif defined(CONFIG_SYS_NS16550_PORT_MAPPED)
 	outb(value, (ulong)addr);
 #elif defined(CONFIG_SYS_NS16550_MEM32) && !defined(CONFIG_SYS_BIG_ENDIAN)
 	out_le32(addr, value);
@@ -70,7 +75,9 @@ static inline void serial_out_shift(void *addr, int shift, int value)
 
Quoted 14 lines

drivers/serial/ns16550.c static inline void serial_out_shift(void *addr, int shift, int value) · line 83

Quoted 9 lines
drivers/serial/ns16550.c +3 -1
 static inline int serial_in_shift(void *addr, int shift)
 {
-#ifdef CONFIG_SYS_NS16550_PORT_MAPPED
+#ifdef CONFIG_ARCH_DFSS
+	return read_aux_reg((int)addr);
+#elif defined(CONFIG_SYS_NS16550_PORT_MAPPED)
 	return inb((ulong)addr);
 #elif defined(CONFIG_SYS_NS16550_MEM32) && !defined(CONFIG_SYS_BIG_ENDIAN)
 	return in_le32(addr);

As always, thanks for the detailed explanation. Yes, I think that of the options, putting the details in read/write_aux_reg (and please make sure read/write_aux_reg have a good function comment too) is the best choice. Thanks!

Quoted 4 lines

Frankly I didn't understand your comment well enough :)
Do you suggest to keep proposed implementation (i.e. modification of serial_{in|out}_shift()) but add comments on what do we do and why or you really meant something completely different?

I'm fine with what you have above, in the driver. I missed that read_aux_reg already exists, sorry. So I guess a detailed commit message will have to be enough documentation for future reviewers of the code. Or, we'll see what comes out of Simon's comments.

Signature · 8 lines
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <http://lists.denx.de/pipermail/u-boot/attachments/20180222/c6648a89/attachment.sig>
#4 lore raw
In reply to Alexey Brodkin

Hi,

Quoted 35 lines · Alexey Brodkin

Synopsys Data Fusion subsystem (DFSS) is targeted to deeply built-in use-cases and so to save some silicon area decision was made to escape usage of any busses and use instead directly wired to CPU peripherals. And one of those is DW APB UART.

Later DFSS became a part of larger and more complicated SoCs with some other peripherals connected via common buses but default UART is still used via ARC core's auxulary registers which are not mapped to "normal" address space and we use very special instructions to access them, thus we cannot simply reuse whatever accessor we have in 16550 UART as of now.

Also we cannot just switch inb()/outb() to access ARC AUX regs always for DFSS because other peripherals have normal memory-mapped control registers and we need to use normal accessors for them.

Frankly I don't like a lot what I did here but otherwise if I create a special driver for this I'll need to reimplement ns16550_serial_ops.putc()/getc() which will be pure copy-paseted from ns16550.c because the only difference is only in ns16550_{read|write}b().

As mentioned above we cannot remap those auxiliary registers to normal memory address-space and thus we have to use very special accessors write_aux_reg()/read_aux_reg() that directly use special CPU intructions (namely LR/SR) for dealing with ARC AUX regs.

Also note here I just use a check for a particular SoC being selected (CONFIG_ARCH_DFSS will be introduced shortly) but that is done just for simplicity, otherwise it might be a slecial Kconfig option for NS16550 or anything else.

I'd like to know what people think about possible colutions here. And as always any comments are much appreciated!

Quoted 7 lines
Cc Simon Glass <[email protected]>
Cc Tom Rini <[email protected]>
Cc Stefan Roese <[email protected]>
---
 drivers/serial/ns16550.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

I think a separate driver might be better, unless we want to make the read/write interface go through regmap or similar?

Regards,
Simon

#5 lore raw

Hi Simon,

Quoted 47 lines · Simon Glass

Hi,

Quoted 35 lines · Alexey Brodkin

Synopsys Data Fusion subsystem (DFSS) is targeted to deeply built-in use-cases and so to save some silicon area decision was made to escape usage of any busses and use instead directly wired to CPU peripherals. And one of those is DW APB UART.

Later DFSS became a part of larger and more complicated SoCs with some other peripherals connected via common buses but default UART is still used via ARC core's auxulary registers which are not mapped to "normal" address space and we use very special instructions to access them, thus we cannot simply reuse whatever accessor we have in 16550 UART as of now.

Also we cannot just switch inb()/outb() to access ARC AUX regs always for DFSS because other peripherals have normal memory-mapped control registers and we need to use normal accessors for them.

Frankly I don't like a lot what I did here but otherwise if I create a special driver for this I'll need to reimplement ns16550_serial_ops.putc()/getc() which will be pure copy-paseted from ns16550.c because the only difference is only in ns16550_{read|write}b().

As mentioned above we cannot remap those auxiliary registers to normal memory address-space and thus we have to use very special accessors write_aux_reg()/read_aux_reg() that directly use special CPU intructions (namely LR/SR) for dealing with ARC AUX regs.

Also note here I just use a check for a particular SoC being selected (CONFIG_ARCH_DFSS will be introduced shortly) but that is done just for simplicity, otherwise it might be a slecial Kconfig option for NS16550 or anything else.

I'd like to know what people think about possible colutions here. And as always any comments are much appreciated!

Quoted 7 lines
Cc Simon Glass <[email protected]>
Cc Tom Rini <[email protected]>
Cc Stefan Roese <[email protected]>
---
 drivers/serial/ns16550.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
Quoted 2 lines

I think a separate driver might be better, unless we want to make the read/write interface go through regmap or similar?

But in case of ARC's AUX regs portmap won't help because those AUX regs are couldn't be mapped - that a completely different address space and we may only access them via dedicated instructions (LR vs LD and SR vs ST).

-Alexey

#6 lore raw

Hi Alexey,

Quoted 50 lines · Alexey Brodkin

Hi Simon,

Quoted 47 lines · Simon Glass

Hi,

Quoted 42 lines · Alexey Brodkin

Synopsys Data Fusion subsystem (DFSS) is targeted to deeply built-in use-cases and so to save some silicon area decision was made to escape usage of any busses and use instead directly wired to CPU peripherals. And one of those is DW APB UART.

Later DFSS became a part of larger and more complicated SoCs with some other peripherals connected via common buses but default UART is still used via ARC core's auxulary registers which are not mapped to "normal" address space and we use very special instructions to access them, thus we cannot simply reuse whatever accessor we have in 16550 UART as of now.

Also we cannot just switch inb()/outb() to access ARC AUX regs always for DFSS because other peripherals have normal memory-mapped control registers and we need to use normal accessors for them.

Frankly I don't like a lot what I did here but otherwise if I create a special driver for this I'll need to reimplement ns16550_serial_ops.putc()/getc() which will be pure copy-paseted from ns16550.c because the only difference is only in ns16550_{read|write}b().

As mentioned above we cannot remap those auxiliary registers to normal memory address-space and thus we have to use very special accessors write_aux_reg()/read_aux_reg() that directly use special CPU intructions (namely LR/SR) for dealing with ARC AUX regs.

Also note here I just use a check for a particular SoC being selected (CONFIG_ARCH_DFSS will be introduced shortly) but that is done just for simplicity, otherwise it might be a slecial Kconfig option for NS16550 or anything else.

I'd like to know what people think about possible colutions here. And as always any comments are much appreciated!

Cc Simon Glass <[email protected]>
Cc Tom Rini <[email protected]>
Cc Stefan Roese <[email protected]>
---
 drivers/serial/ns16550.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
Quoted 6 lines
Quoted 2 lines

I think a separate driver might be better, unless we want to make the read/write interface go through regmap or similar?

But in case of ARC's AUX regs portmap won't help because those AUX regs are couldn't be mapped - that a completely different address space and we may only access them via dedicated instructions (LR vs LD and SR vs ST).

Well...

1. With a separate driver, you can do whatever you want :-) I know it introduces code duplication though...

2. With regmap you can add your own regmap driver, and again do whatever you want. I can help with that if it sounds attractive

Regards,
Simon

#7 lore raw

Hi Simon,

Quoted 11 lines · Simon Glass

Hi Alexey,

Quoted 7 lines · Alexey Brodkin

Hi Simon,

Quoted 4 lines · Simon Glass

Hi,

On 21 February 2018 at 05:26, Alexey Brodkin
<[email protected]> wrote:

[snip]

Quoted 11 lines
Quoted 10 lines
Quoted 6 lines
Quoted 3 lines
---
 drivers/serial/ns16550.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

I think a separate driver might be better, unless we want to make the read/write interface go through regmap or similar?

But in case of ARC's AUX regs portmap won't help because those AUX regs are couldn't be mapped - that a completely different address space and we may only access them via dedicated instructions (LR vs LD and SR vs ST).

Quoted 7 lines

Well...

1. With a separate driver, you can do whatever you want :-) I know it introduces code duplication though...

2. With regmap you can add your own regmap driver, and again do whatever you want. I can help with that if it sounds attractive

That sounds definitely interesting!
Are there any pointers on how that usually done?

-Alexey

#8 lore raw

Hi Alexey,

Quoted 30 lines · Alexey Brodkin

Hi Simon,

Quoted 11 lines · Simon Glass

Hi Alexey,

Quoted 7 lines · Alexey Brodkin

Hi Simon,

Quoted 4 lines · Simon Glass

Hi,

On 21 February 2018 at 05:26, Alexey Brodkin
<[email protected]> wrote:

[snip]

Quoted 13 lines
Quoted 10 lines
Quoted 6 lines
Quoted 3 lines
---
 drivers/serial/ns16550.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

I think a separate driver might be better, unless we want to make the read/write interface go through regmap or similar?

But in case of ARC's AUX regs portmap won't help because those AUX regs are couldn't be mapped - that a completely different address space and we may only access them via dedicated instructions (LR vs LD and SR vs ST).

Well...

Quoted 8 lines
Quoted 5 lines

1. With a separate driver, you can do whatever you want :-) I know it introduces code duplication though...

2. With regmap you can add your own regmap driver, and again do whatever you want. I can help with that if it sounds attractive

That sounds definitely interesting!
Are there any pointers on how that usually done?

Not yet. My ideas:

- Add a new regmap_init_xxx() function
- Adjust the regmap structure to allow read()/write() function pointers to be provided
- Adjust the regmap code to call them, instead of the hard-coded memory-mapped ones
- Adjust ns16550 to call regmap instead of readl/writel

This should clean up the mess in ns16550 a bit. But it will require a bit of investigation / thinking.

Regards,
Simon

#9 lore raw
In reply to Simon Glass

Hi Simon,

Quoted 7 lines · Simon Glass

Hi Alexey,

Quoted 3 lines · Alexey Brodkin

Hi Simon,

On Thu, 2018-02-22 at 09:17 -0700, Simon Glass wrote:

[snip]

Quoted 11 lines
Quoted 6 lines
Quoted 2 lines

I think a separate driver might be better, unless we want to make the read/write interface go through regmap or similar?

But in case of ARC's AUX regs portmap won't help because those AUX regs are couldn't be mapped - that a completely different address space and we may only access them via dedicated instructions (LR vs LD and SR vs ST).

Well...

1. With a separate driver, you can do whatever you want :-) I know it introduces code duplication though...

Exactly I hate to introduce another driver which will be 99,9% the same as an existing one and then we'll need to care of it as well.

Quoted 2 lines

2. With regmap you can add your own regmap driver, and again do whatever you want. I can help with that if it sounds attractive

Ok so I took a look at regmap in Linux kernel and indeed that will solve our problem: we'll have a regmap-mmio.c and regmap-arcaux.c with appropriate implementation of accessors but I'm not really sure it worth the trouble. Or your idea was to move all the different #ifdefs from serial_{in|out}_shift() to the corresponding regmap implementations such that we have something like below:
regmap-mem32-le.c
regmap-mem32-be.c
regmap-portmapped.c
etc
?

Regards,
Alexey

#10 lore raw

Hi Alexey.

Quoted 32 lines · Alexey Brodkin

Hi Simon,

Quoted 7 lines · Simon Glass

Hi Alexey,

Quoted 3 lines · Alexey Brodkin

Hi Simon,

On Thu, 2018-02-22 at 09:17 -0700, Simon Glass wrote:

[snip]

Quoted 11 lines
Quoted 6 lines
Quoted 2 lines

I think a separate driver might be better, unless we want to make the read/write interface go through regmap or similar?

But in case of ARC's AUX regs portmap won't help because those AUX regs are couldn't be mapped - that a completely different address space and we may only access them via dedicated instructions (LR vs LD and SR vs ST).

Well...

1. With a separate driver, you can do whatever you want :-) I know it introduces code duplication though...

Exactly I hate to introduce another driver which will be 99,9% the same as an existing one and then we'll need to care of it as well.

Quoted 2 lines

2. With regmap you can add your own regmap driver, and again do whatever you want. I can help with that if it sounds attractive

Quoted 11 lines

Ok so I took a look at regmap in Linux kernel and indeed that will solve our problem: we'll have a regmap-mmio.c and regmap-arcaux.c with appropriate implementation of accessors but I'm not really sure it worth the trouble. Or your idea was to move all the different #ifdefs from serial_{in|out}_shift() to the corresponding regmap implementations such that we have something like below:
regmap-mem32-le.c
regmap-mem32-be.c
regmap-portmapped.c
etc
?

Yes, that's it.

I'm not sure why that driver is so special. Presumably other drivers would have the same problem?

Regards,
Simon