From: Victor Kamensky victor.kamensky@linaro.org
Assembler functions defined in sleep44xx.S need to byteswap values after read / before write from h/w register if code compiled in big endian mode. Simple change to do 'rev x, x' before str instruction and after ldr instruction that deals with h/w registers.
Signed-off-by: Victor Kamensky victor.kamensky@linaro.org Signed-off-by: Taras Kondratiuk taras.kondratiuk@linaro.org --- This is a part of RFC series [1]. Based on v3.13-rc8.
[1] http://www.spinics.net/lists/linux-omap/msg99927.html
arch/arm/mach-omap2/sleep44xx.S | 17 +++++++++++++++++ 1 file changed, 17 insertions(+)
diff --git a/arch/arm/mach-omap2/sleep44xx.S b/arch/arm/mach-omap2/sleep44xx.S index 9086ce0..8017016 100644 --- a/arch/arm/mach-omap2/sleep44xx.S +++ b/arch/arm/mach-omap2/sleep44xx.S @@ -12,6 +12,7 @@ #include <linux/linkage.h> #include <asm/smp_scu.h> #include <asm/memory.h> +#include <asm/assembler.h> #include <asm/hardware/cache-l2x0.h>
#include "omap-secure.h" @@ -74,6 +75,7 @@ ENTRY(omap4_finish_suspend) */ bl omap4_get_sar_ram_base ldr r9, [r0, #OMAP_TYPE_OFFSET] +ARM_BE8(rev r9, r9) cmp r9, #0x1 @ Check for HS device bne skip_secure_l1_clean mov r0, #SCU_PM_NORMAL @@ -113,12 +115,14 @@ skip_secure_l1_clean: bl omap4_get_sar_ram_base mov r8, r0 ldr r9, [r8, #OMAP_TYPE_OFFSET] +ARM_BE8(rev r9, r9) cmp r9, #0x1 @ Check for HS device bne scu_gp_set mrc p15, 0, r0, c0, c0, 5 @ Read MPIDR ands r0, r0, #0x0f ldreq r0, [r8, #SCU_OFFSET0] ldrne r0, [r8, #SCU_OFFSET1] +ARM_BE8(rev r0, r0) mov r1, #0x00 stmfd r13!, {r4-r12, r14} ldr r12, =OMAP4_MON_SCU_PWR_INDEX @@ -130,6 +134,7 @@ scu_gp_set: ands r0, r0, #0x0f ldreq r1, [r8, #SCU_OFFSET0] ldrne r1, [r8, #SCU_OFFSET1] +ARM_BE8(rev r1, r1) bl omap4_get_scu_base bl scu_power_mode skip_scu_gp_set: @@ -157,6 +162,7 @@ skip_scu_gp_set: ands r5, r5, #0x0f ldreq r0, [r8, #L2X0_SAVE_OFFSET0] @ Retrieve L2 state from SAR ldrne r0, [r8, #L2X0_SAVE_OFFSET1] @ memory. +ARM_BE8(rev r0, r0) cmp r0, #3 bne do_WFI #ifdef CONFIG_PL310_ERRATA_727915 @@ -167,9 +173,11 @@ skip_scu_gp_set: bl omap4_get_l2cache_base mov r2, r0 ldr r0, =0xffff +ARM_BE8(rev r0, r0) str r0, [r2, #L2X0_CLEAN_INV_WAY] wait: ldr r0, [r2, #L2X0_CLEAN_INV_WAY] +ARM_BE8(rev r0, r0) ldr r1, =0xffff ands r0, r0, r1 bne wait @@ -182,9 +190,11 @@ l2x_sync: bl omap4_get_l2cache_base mov r2, r0 mov r0, #0x0 +ARM_BE8(rev r0, r0) str r0, [r2, #L2X0_CACHE_SYNC] sync: ldr r0, [r2, #L2X0_CACHE_SYNC] +ARM_BE8(rev r0, r0) ands r0, r0, #0x1 bne sync #endif @@ -216,6 +226,7 @@ do_WFI: bl omap4_get_sar_ram_base mov r8, r0 ldr r9, [r8, #OMAP_TYPE_OFFSET] +ARM_BE8(rev r9, r9) cmp r9, #0x1 @ Check for HS device bne scu_gp_clear mov r0, #SCU_PM_NORMAL @@ -258,6 +269,7 @@ ENTRY(omap4_cpu_resume) */ ldr r8, =OMAP44XX_SAR_RAM_BASE ldr r9, [r8, #OMAP_TYPE_OFFSET] +ARM_BE8(rev r9, r9) cmp r9, #0x1 @ Skip if GP device bne skip_ns_smp_enable mrc p15, 0, r0, c0, c0, 5 @@ -292,16 +304,19 @@ skip_ns_smp_enable: */ ldr r2, =OMAP44XX_L2CACHE_BASE ldr r0, [r2, #L2X0_CTRL] +ARM_BE8(rev r0, r0) and r0, #0x0f cmp r0, #1 beq skip_l2en @ Skip if already enabled ldr r3, =OMAP44XX_SAR_RAM_BASE ldr r1, [r3, #OMAP_TYPE_OFFSET] +ARM_BE8(rev r1, r1) cmp r1, #0x1 @ Check for HS device bne set_gp_por ldr r0, =OMAP4_PPA_L2_POR_INDEX ldr r1, =OMAP44XX_SAR_RAM_BASE ldr r4, [r1, #L2X0_PREFETCH_CTRL_OFFSET] +ARM_BE8(rev r4, r4) adr r3, ppa_por_params str r4, [r3, #0x04] mov r1, #0x0 @ Process ID @@ -313,11 +328,13 @@ skip_ns_smp_enable: set_gp_por: ldr r1, =OMAP44XX_SAR_RAM_BASE ldr r0, [r1, #L2X0_PREFETCH_CTRL_OFFSET] +ARM_BE8(rev r0, r0) ldr r12, =OMAP4_MON_L2X0_PREFETCH_INDEX @ Setup L2 PREFETCH DO_SMC set_aux_ctrl: ldr r1, =OMAP44XX_SAR_RAM_BASE ldr r0, [r1, #L2X0_AUXCTRL_OFFSET] +ARM_BE8(rev r0, r0) ldr r12, =OMAP4_MON_L2X0_AUXCTRL_INDEX @ Setup L2 AUXCTRL DO_SMC mov r0, #0x1
On 01/13/2014 09:03 AM, Taras Kondratiuk wrote:
From: Victor Kamensky victor.kamensky@linaro.org
Assembler functions defined in sleep44xx.S need to byteswap values after read / before write from h/w register if code compiled in big endian mode. Simple change to do 'rev x, x' before str instruction and after ldr instruction that deals with h/w registers.
Signed-off-by: Victor Kamensky victor.kamensky@linaro.org Signed-off-by: Taras Kondratiuk taras.kondratiuk@linaro.org
This is a part of RFC series [1]. Based on v3.13-rc8.
[1] http://www.spinics.net/lists/linux-omap/msg99927.html
arch/arm/mach-omap2/sleep44xx.S | 17 +++++++++++++++++ 1 file changed, 17 insertions(+)
OMAP4 is LE, and if there is a gcc flag for the same, is'nt it cleaner to deal with it in Makefile rather than trying to make an assembly meant only for LE by force building it for BE?
diff --git a/arch/arm/mach-omap2/sleep44xx.S b/arch/arm/mach-omap2/sleep44xx.S index 9086ce0..8017016 100644 --- a/arch/arm/mach-omap2/sleep44xx.S +++ b/arch/arm/mach-omap2/sleep44xx.S @@ -12,6 +12,7 @@ #include <linux/linkage.h> #include <asm/smp_scu.h> #include <asm/memory.h> +#include <asm/assembler.h> #include <asm/hardware/cache-l2x0.h> #include "omap-secure.h" @@ -74,6 +75,7 @@ ENTRY(omap4_finish_suspend) */ bl omap4_get_sar_ram_base ldr r9, [r0, #OMAP_TYPE_OFFSET] +ARM_BE8(rev r9, r9) cmp r9, #0x1 @ Check for HS device bne skip_secure_l1_clean mov r0, #SCU_PM_NORMAL @@ -113,12 +115,14 @@ skip_secure_l1_clean: bl omap4_get_sar_ram_base mov r8, r0 ldr r9, [r8, #OMAP_TYPE_OFFSET] +ARM_BE8(rev r9, r9) cmp r9, #0x1 @ Check for HS device bne scu_gp_set mrc p15, 0, r0, c0, c0, 5 @ Read MPIDR ands r0, r0, #0x0f ldreq r0, [r8, #SCU_OFFSET0] ldrne r0, [r8, #SCU_OFFSET1] +ARM_BE8(rev r0, r0) mov r1, #0x00 stmfd r13!, {r4-r12, r14} ldr r12, =OMAP4_MON_SCU_PWR_INDEX @@ -130,6 +134,7 @@ scu_gp_set: ands r0, r0, #0x0f ldreq r1, [r8, #SCU_OFFSET0] ldrne r1, [r8, #SCU_OFFSET1] +ARM_BE8(rev r1, r1) bl omap4_get_scu_base bl scu_power_mode skip_scu_gp_set: @@ -157,6 +162,7 @@ skip_scu_gp_set: ands r5, r5, #0x0f ldreq r0, [r8, #L2X0_SAVE_OFFSET0] @ Retrieve L2 state from SAR ldrne r0, [r8, #L2X0_SAVE_OFFSET1] @ memory. +ARM_BE8(rev r0, r0) cmp r0, #3 bne do_WFI #ifdef CONFIG_PL310_ERRATA_727915 @@ -167,9 +173,11 @@ skip_scu_gp_set: bl omap4_get_l2cache_base mov r2, r0 ldr r0, =0xffff +ARM_BE8(rev r0, r0) str r0, [r2, #L2X0_CLEAN_INV_WAY] wait: ldr r0, [r2, #L2X0_CLEAN_INV_WAY] +ARM_BE8(rev r0, r0) ldr r1, =0xffff ands r0, r0, r1 bne wait @@ -182,9 +190,11 @@ l2x_sync: bl omap4_get_l2cache_base mov r2, r0 mov r0, #0x0 +ARM_BE8(rev r0, r0) str r0, [r2, #L2X0_CACHE_SYNC] sync: ldr r0, [r2, #L2X0_CACHE_SYNC] +ARM_BE8(rev r0, r0) ands r0, r0, #0x1 bne sync #endif @@ -216,6 +226,7 @@ do_WFI: bl omap4_get_sar_ram_base mov r8, r0 ldr r9, [r8, #OMAP_TYPE_OFFSET] +ARM_BE8(rev r9, r9) cmp r9, #0x1 @ Check for HS device bne scu_gp_clear mov r0, #SCU_PM_NORMAL @@ -258,6 +269,7 @@ ENTRY(omap4_cpu_resume) */ ldr r8, =OMAP44XX_SAR_RAM_BASE ldr r9, [r8, #OMAP_TYPE_OFFSET] +ARM_BE8(rev r9, r9) cmp r9, #0x1 @ Skip if GP device bne skip_ns_smp_enable mrc p15, 0, r0, c0, c0, 5 @@ -292,16 +304,19 @@ skip_ns_smp_enable: */ ldr r2, =OMAP44XX_L2CACHE_BASE ldr r0, [r2, #L2X0_CTRL] +ARM_BE8(rev r0, r0) and r0, #0x0f cmp r0, #1 beq skip_l2en @ Skip if already enabled ldr r3, =OMAP44XX_SAR_RAM_BASE ldr r1, [r3, #OMAP_TYPE_OFFSET] +ARM_BE8(rev r1, r1) cmp r1, #0x1 @ Check for HS device bne set_gp_por ldr r0, =OMAP4_PPA_L2_POR_INDEX ldr r1, =OMAP44XX_SAR_RAM_BASE ldr r4, [r1, #L2X0_PREFETCH_CTRL_OFFSET] +ARM_BE8(rev r4, r4) adr r3, ppa_por_params str r4, [r3, #0x04] mov r1, #0x0 @ Process ID @@ -313,11 +328,13 @@ skip_ns_smp_enable: set_gp_por: ldr r1, =OMAP44XX_SAR_RAM_BASE ldr r0, [r1, #L2X0_PREFETCH_CTRL_OFFSET] +ARM_BE8(rev r0, r0) ldr r12, =OMAP4_MON_L2X0_PREFETCH_INDEX @ Setup L2 PREFETCH DO_SMC set_aux_ctrl: ldr r1, =OMAP44XX_SAR_RAM_BASE ldr r0, [r1, #L2X0_AUXCTRL_OFFSET] +ARM_BE8(rev r0, r0) ldr r12, =OMAP4_MON_L2X0_AUXCTRL_INDEX @ Setup L2 AUXCTRL DO_SMC mov r0, #0x1
On 13 January 2014 17:23, Nishanth Menon nm@ti.com wrote:
On 01/13/2014 09:03 AM, Taras Kondratiuk wrote:
From: Victor Kamensky victor.kamensky@linaro.org
Assembler functions defined in sleep44xx.S need to byteswap values after read / before write from h/w register if code compiled in big endian mode. Simple change to do 'rev x, x' before str instruction and after ldr instruction that deals with h/w registers.
Signed-off-by: Victor Kamensky victor.kamensky@linaro.org Signed-off-by: Taras Kondratiuk taras.kondratiuk@linaro.org
This is a part of RFC series [1]. Based on v3.13-rc8.
[1] http://www.spinics.net/lists/linux-omap/msg99927.html
arch/arm/mach-omap2/sleep44xx.S | 17 +++++++++++++++++ 1 file changed, 17 insertions(+)
OMAP4 is LE, and if there is a gcc flag for the same, is'nt it cleaner to deal with it in Makefile rather than trying to make an assembly meant only for LE by force building it for BE?
Hi Nishanth I'm not sure I got your point. Do you propose to build this file as LE while the rest of kernel is BE?
On 01/14/2014 05:14 AM, Taras Kondratiuk wrote:
On 13 January 2014 17:23, Nishanth Menon nm@ti.com wrote:
On 01/13/2014 09:03 AM, Taras Kondratiuk wrote:
From: Victor Kamensky victor.kamensky@linaro.org
Assembler functions defined in sleep44xx.S need to byteswap values after read / before write from h/w register if code compiled in big endian mode. Simple change to do 'rev x, x' before str instruction and after ldr instruction that deals with h/w registers.
Signed-off-by: Victor Kamensky victor.kamensky@linaro.org Signed-off-by: Taras Kondratiuk taras.kondratiuk@linaro.org
This is a part of RFC series [1]. Based on v3.13-rc8.
[1] http://www.spinics.net/lists/linux-omap/msg99927.html
arch/arm/mach-omap2/sleep44xx.S | 17 +++++++++++++++++ 1 file changed, 17 insertions(+)
OMAP4 is LE, and if there is a gcc flag for the same, is'nt it cleaner to deal with it in Makefile rather than trying to make an assembly meant only for LE by force building it for BE?
Hi Nishanth I'm not sure I got your point. Do you propose to build this file as LE while the rest of kernel is BE?
I dont see why I should deal with the BE macro for every code change we have in omap4,am335x assembly. The hardware is LE and wont change just coz you are building it for BE. So I dont get the rationale for changing the assembly here - yes, if the assembly can be maintained as LE only mode and the build handling be adequately handled in Makefile (similar to SMC handling), that would be the best.
is the idea of BE build meant to deal with having a single BE kernel build work for all platforms (including LE ones)?
Hi Nishanth,
On 14 January 2014 07:51, Nishanth Menon nm@ti.com wrote:
On 01/14/2014 05:14 AM, Taras Kondratiuk wrote:
On 13 January 2014 17:23, Nishanth Menon nm@ti.com wrote:
On 01/13/2014 09:03 AM, Taras Kondratiuk wrote:
From: Victor Kamensky victor.kamensky@linaro.org
Assembler functions defined in sleep44xx.S need to byteswap values after read / before write from h/w register if code compiled in big endian mode. Simple change to do 'rev x, x' before str instruction and after ldr instruction that deals with h/w registers.
Signed-off-by: Victor Kamensky victor.kamensky@linaro.org Signed-off-by: Taras Kondratiuk taras.kondratiuk@linaro.org
This is a part of RFC series [1]. Based on v3.13-rc8.
[1] http://www.spinics.net/lists/linux-omap/msg99927.html
arch/arm/mach-omap2/sleep44xx.S | 17 +++++++++++++++++ 1 file changed, 17 insertions(+)
OMAP4 is LE, and if there is a gcc flag for the same, is'nt it cleaner to deal with it in Makefile rather than trying to make an assembly meant only for LE by force building it for BE?
Hi Nishanth I'm not sure I got your point. Do you propose to build this file as LE while the rest of kernel is BE?
I dont see why I should deal with the BE macro for every code change we have in omap4,am335x assembly. The hardware is LE and wont change just coz you are building it for BE. So I dont get the rationale for changing the assembly here - yes, if the assembly can be maintained as LE only mode and the build handling be adequately handled in Makefile (similar to SMC handling), that would be the best.
ARM core is capable of running in LE or BE modes. Yes, OMAP memory mapped periphery gives data in LE form. When core runs in BE mode after reading h/w register it will byteswap it, also before writing h/w register it byteswap it. In such way BE kernel can work with LE periphery.
When it comes to C code that works with LE periphery, if correct access functions are used like readl_relaxed and writel_relaxed (vs __raw_readl and __raw_writel), the functions will take care of the swaps. In case of asm files there is no other way than to insert those byteswaps manually and conditionally - they will be enabled only if kernel compiled in BE mode. The reason why it could be done only manually is that load/store opcodes behavior changes when core runs in BE mode (E bit set) in these case ldr/str treat memory as big endian. So when LE periphery register is read/stored additional byteswaps that compensate for it should be inserted.
When BE kernel is built Makefile does take of compiling code in BE mode. I.e all proper flags like -mbig-endian and -Wl,--be8 will be set.
is the idea of BE build meant to deal with having a single BE kernel build work for all platforms (including LE ones)?
Sort of. The idea here to run BE image on OMAP4 chip, with kernel that would deals with LE periphery correctly, but ARM core run in BE with special kernel that compiled for BE case (i.e CONFIG_CPU_BIG_ENDIAN is set).
Thanks, Victor
-- Regards, Nishanth Menon
On Tue, Jan 14, 2014 at 11:35 AM, Victor Kamensky victor.kamensky@linaro.org wrote:
When BE kernel is built Makefile does take of compiling code in BE mode. I.e all proper flags like -mbig-endian and -Wl,--be8 will be set.
Agreed, and I assume you cannot instead switch to LE mode when entering assembly assuming LE?
The reason I ask this is - most of our development is NOT in BE mode. we will continue to manipulate and add assembly - AM335x, DRA7/OMAP5 etc.. and obviously not every code change will indulge in ensuring right markers will be in place.
by ensuring readl_relaxed handles the variations, you have ensured that I dont need to care about drivers other than to ensure they use _relaxed variants. in the case of assembly, this does not seem long term manageable.
is the idea of BE build meant to deal with having a single BE kernel build work for all platforms (including LE ones)?
Sort of. The idea here to run BE image on OMAP4 chip, with kernel that would deals with LE periphery correctly, but ARM core run in BE with special kernel that compiled for BE case (i.e CONFIG_CPU_BIG_ENDIAN is set).
I still dont get the usecase - other than "hey, we do this coz we can do it!".. I mean, yep, it sounds great and all.. but 4 years down the line, is this still going to work? is this going to be interesting careabout? or we are just maintaining additional code for a passing fancy or proof-of-concept?
Regards, Nishanth Menon
On 14 January 2014 09:56, Nishanth Menon nm@ti.com wrote:
On Tue, Jan 14, 2014 at 11:35 AM, Victor Kamensky victor.kamensky@linaro.org wrote:
When BE kernel is built Makefile does take of compiling code in BE mode. I.e all proper flags like -mbig-endian and -Wl,--be8 will be set.
Agreed, and I assume you cannot instead switch to LE mode when entering assembly assuming LE?
Yes. Note that this asm interacts with other data in kernel that would be in BE form. And after all linker will not allow to put together files that were compiled in different endianity.
The reason I ask this is - most of our development is NOT in BE mode. we will continue to manipulate and add assembly - AM335x, DRA7/OMAP5 etc.. and obviously not every code change will indulge in ensuring right markers will be in place.
by ensuring readl_relaxed handles the variations, you have ensured that I dont need to care about drivers other than to ensure they use _relaxed variants. in the case of assembly, this does not seem long term manageable.
Yes, I agree if it is outside of main use case it will get broken if not attended to. Definitely universe entropy increases :) - if left without attention things will decay. Note we admit that even with ARM CPU core BE changes similar decay can happen eventually ...that is what LNG BE group trying to prevent. I think the deal here is that next user of it can/need to fix things if they decayed.
is the idea of BE build meant to deal with having a single BE kernel build work for all platforms (including LE ones)?
Sort of. The idea here to run BE image on OMAP4 chip, with kernel that would deals with LE periphery correctly, but ARM core run in BE with special kernel that compiled for BE case (i.e CONFIG_CPU_BIG_ENDIAN is set).
I still dont get the usecase - other than "hey, we do this coz we can do it!".. I mean, yep, it sounds great and all.. but 4 years down the line, is this still going to work? is this going to be interesting careabout? or we are just maintaining additional code for a passing fancy or proof-of-concept?
Valid concern. From LNG BE group point of view it is not "we can do it". It is more like we've done it. We have Pandaboard ES running BE kernel for a while. It is in LNG BE tree. We used it as development and testing vehicle for BE work for a while. We are very grateful to the platform for that - it is affordable and easily available! Given, beyond ongoing BE testing on Pandaboard in LNG there may not be valid use case for further things on OMAP4 BE. We had discussion with Santosh Shilimkar from TI during last Linaro connect what to do with LNG BE Pandaboard series. Santosh suggested and we agreed that we would try to contribute them back to community. And that is what Taras is doing. IMHO even though there may not be real product use case for BE OMAP4, it could serve in kernel source as good example that shows how to work with LE periphery from BE kernel. After all, these changes are noop for LE case and they don't introduce a lot of clutter: all BE asm rev and rev16 instructions wrapped around with ARM_BE8 macro.
Thanks, Victor
Regards, Nishanth Menon
On Tue, Jan 14, 2014 at 2:20 PM, Victor Kamensky victor.kamensky@linaro.org wrote:
On 14 January 2014 09:56, Nishanth Menon nm@ti.com wrote:
On Tue, Jan 14, 2014 at 11:35 AM, Victor Kamensky victor.kamensky@linaro.org wrote:
When BE kernel is built Makefile does take of compiling code in BE mode. I.e all proper flags like -mbig-endian and -Wl,--be8 will be set.
Agreed, and I assume you cannot instead switch to LE mode when entering assembly assuming LE?
Yes. Note that this asm interacts with other data in kernel that would be in BE form. And after all linker will not allow to put together files that were compiled in different endianity.
The reason I ask this is - most of our development is NOT in BE mode. we will continue to manipulate and add assembly - AM335x, DRA7/OMAP5 etc.. and obviously not every code change will indulge in ensuring right markers will be in place.
by ensuring readl_relaxed handles the variations, you have ensured that I dont need to care about drivers other than to ensure they use _relaxed variants. in the case of assembly, this does not seem long term manageable.
Yes, I agree if it is outside of main use case it will get broken if not attended to. Definitely universe entropy increases :) - if left without attention things will decay. Note we admit that even with ARM CPU core BE changes similar decay can happen eventually ...that is what LNG BE group trying to prevent. I think the deal here is that next user of it can/need to fix things if they decayed.
Personally, I have no idea what "LNG BE" stands for.. I see the approach we are attempting here is to do any interaction external to ARM boundary to go through the ARM_BE8 macro.
If we want to make this something we can live with, then the platforms will have to ensure memory operations external to ARM must be operated upon by these macros. Instead, an approach that may be used is to introduce accessors macros that will provide right instruction based on BE/LE build and BE/LE SoC peripheral behavior.
is the idea of BE build meant to deal with having a single BE kernel build work for all platforms (including LE ones)?
Sort of. The idea here to run BE image on OMAP4 chip, with kernel that would deals with LE periphery correctly, but ARM core run in BE with special kernel that compiled for BE case (i.e CONFIG_CPU_BIG_ENDIAN is set).
I still dont get the usecase - other than "hey, we do this coz we can do it!".. I mean, yep, it sounds great and all.. but 4 years down the line, is this still going to work? is this going to be interesting careabout? or we are just maintaining additional code for a passing fancy or proof-of-concept?
Valid concern. From LNG BE group point of view it is not "we can do it". It is more like we've done it. We have Pandaboard ES running BE kernel for a while. It is in LNG BE tree. We used it as development and testing vehicle for BE work for a while. We are very grateful to the platform for that - it is affordable and easily available! Given, beyond ongoing BE testing on Pandaboard in LNG there may not be valid use case for further things on OMAP4 BE. We had discussion with Santosh Shilimkar from TI during last Linaro connect what to do with LNG BE Pandaboard series. Santosh suggested and we agreed that we would try to contribute them back to community. And that is what Taras is doing. IMHO even though there may not be real
ok.. some sort of Linaro thing about which I have no background about - but dont really care in this context.
product use case for BE OMAP4, it could serve in kernel source as good example that shows how to work with LE periphery from BE kernel. After all, these changes are noop for LE case and they don't introduce a lot of clutter: all BE asm rev and rev16 instructions wrapped around with ARM_BE8 macro.
So, why are we not doing this for *all* OMAP platforms? we try very hard not to regress on features introduced and wish to ensure all platforms are equivalent in terms of usage. it makes code sharing a little harder if we have to ensure that ARM_BE8 is supported in a mix of reusable codebases.
Personally, here is what I might expect to see: a) if we are going to force usage of BE and LE build options - then we *must* be able to reuse code across platforms without having to worry about breaking platform x etc.. b) this seems like - "take this code in" if something breaks, the guy who needs it later will come and fix - Shrug, I personally dont buy that. if something benefits many folks, lets all use it in upstream, if there is just an exception usage for a short duration which may impact code scalability and hard to maintain, keep it forked. c) considering this is non standard usage, some sort of regular test procedure and owner to ensure this feature is kept sane
With out the above, we are just adding dead code IMHO.
Regards, Nishanth Menon
On Tuesday 14 January 2014 03:56 PM, Nishanth Menon wrote:
On Tue, Jan 14, 2014 at 2:20 PM, Victor Kamensky victor.kamensky@linaro.org wrote:
On 14 January 2014 09:56, Nishanth Menon nm@ti.com wrote:
On Tue, Jan 14, 2014 at 11:35 AM, Victor Kamensky victor.kamensky@linaro.org wrote:
When BE kernel is built Makefile does take of compiling code in BE mode. I.e all proper flags like -mbig-endian and -Wl,--be8 will be set.
Agreed, and I assume you cannot instead switch to LE mode when entering assembly assuming LE?
Yes. Note that this asm interacts with other data in kernel that would be in BE form. And after all linker will not allow to put together files that were compiled in different endianity.
The reason I ask this is - most of our development is NOT in BE mode. we will continue to manipulate and add assembly - AM335x, DRA7/OMAP5 etc.. and obviously not every code change will indulge in ensuring right markers will be in place.
by ensuring readl_relaxed handles the variations, you have ensured that I dont need to care about drivers other than to ensure they use _relaxed variants. in the case of assembly, this does not seem long term manageable.
Yes, I agree if it is outside of main use case it will get broken if not attended to. Definitely universe entropy increases :) - if left without attention things will decay. Note we admit that even with ARM CPU core BE changes similar decay can happen eventually ...that is what LNG BE group trying to prevent. I think the deal here is that next user of it can/need to fix things if they decayed.
Personally, I have no idea what "LNG BE" stands for.. I see the approach we are attempting here is to do any interaction external to ARM boundary to go through the ARM_BE8 macro.
If we want to make this something we can live with, then the platforms will have to ensure memory operations external to ARM must be operated upon by these macros. Instead, an approach that may be used is to introduce accessors macros that will provide right instruction based on BE/LE build and BE/LE SoC peripheral behavior.
is the idea of BE build meant to deal with having a single BE kernel build work for all platforms (including LE ones)?
Sort of. The idea here to run BE image on OMAP4 chip, with kernel that would deals with LE periphery correctly, but ARM core run in BE with special kernel that compiled for BE case (i.e CONFIG_CPU_BIG_ENDIAN is set).
I still dont get the usecase - other than "hey, we do this coz we can do it!".. I mean, yep, it sounds great and all.. but 4 years down the line, is this still going to work? is this going to be interesting careabout? or we are just maintaining additional code for a passing fancy or proof-of-concept?
Valid concern. From LNG BE group point of view it is not "we can do it". It is more like we've done it. We have Pandaboard ES running BE kernel for a while. It is in LNG BE tree. We used it as development and testing vehicle for BE work for a while. We are very grateful to the platform for that - it is affordable and easily available! Given, beyond ongoing BE testing on Pandaboard in LNG there may not be valid use case for further things on OMAP4 BE. We had discussion with Santosh Shilimkar from TI during last Linaro connect what to do with LNG BE Pandaboard series. Santosh suggested and we agreed that we would try to contribute them back to community. And that is what Taras is doing. IMHO even though there may not be real
ok.. some sort of Linaro thing about which I have no background about
- but dont really care in this context.
Nothing related Linaro. Its just that platforms are supporting ARM BE mode and Linaro folks had working patches for Panda. So I suggested to get them on the lists.
Regards, Santosh
On Tue, Jan 14, 2014 at 3:03 PM, Santosh Shilimkar santosh.shilimkar@ti.com wrote:
ok.. some sort of Linaro thing about which I have no background about
- but dont really care in this context.
Nothing related Linaro. Its just that platforms are supporting ARM BE mode and Linaro folks had working patches for Panda. So I suggested to get them on the lists.
I tend to think -> is this with OFF mode and CPUidle completely working? All context save and restore works with this? on HS and GP devices with BE mode builds? works on SDP4430,60 variations, considered reuse with AM43xx which could use parts of that logic?
I mean to indicate that terms like "works on panda" tends always to be relative.
It is nice to see it as a proof of concept, but I'd hate to see some dead code lying around in kernel and folks blindly following suit and introducing macros for new assembly for a feature that in practice just one group of folks care about and creates additional burden for rest of folks trying to keep that functionality going as we jump from one "device tree" style churn to another "framework"? Not to mean that good features should be kept away.. but personally, I could not find convincing arguments in this case..
Regards, Nishanth Menon
On 14 January 2014 23:13, Nishanth Menon nm@ti.com wrote:
On Tue, Jan 14, 2014 at 3:03 PM, Santosh Shilimkar santosh.shilimkar@ti.com wrote:
ok.. some sort of Linaro thing about which I have no background about
- but dont really care in this context.
Nothing related Linaro. Its just that platforms are supporting ARM BE mode and Linaro folks had working patches for Panda. So I suggested to get them on the lists.
I tend to think -> is this with OFF mode and CPUidle completely working? All context save and restore works with this? on HS and GP devices with BE mode builds? works on SDP4430,60 variations, considered reuse with AM43xx which could use parts of that logic?
I mean to indicate that terms like "works on panda" tends always to be relative.
Nishanth, let's be objective here. CPUidle on 4460 does *not* work in mainline for at least two kernel releases even for LE [1]. That fix exists because that "one group of folks" faced the issue during BE testing. Looks like not much people care about CPUIdle on OMAP4.
It is nice to see it as a proof of concept, but I'd hate to see some dead code lying around in kernel and folks blindly following suit and introducing macros for new assembly for a feature that in practice just one group of folks care about and creates additional burden for rest of folks trying to keep that functionality going as we jump from one "device tree" style churn to another "framework"? Not to mean that good features should be kept away.. but personally, I could not find convincing arguments in this case..
In general I understand your concerns from previous e-mails, but I don't see a point to spend time for a generic solution for a single user. If there will be other platforms which need similar changes, then we can think of some generic solution. Let's drop this patch for now. We can just disable CPUidle for BE tasks or keep this patch forked.
[1] https://lkml.org/lkml/2013/10/22/401
On Tuesday 14 January 2014 04:13 PM, Nishanth Menon wrote:
On Tue, Jan 14, 2014 at 3:03 PM, Santosh Shilimkar santosh.shilimkar@ti.com wrote:
ok.. some sort of Linaro thing about which I have no background about
- but dont really care in this context.
Nothing related Linaro. Its just that platforms are supporting ARM BE mode and Linaro folks had working patches for Panda. So I suggested to get them on the lists.
I tend to think -> is this with OFF mode and CPUidle completely working? All context save and restore works with this? on HS and GP devices with BE mode builds? works on SDP4430,60 variations, considered reuse with AM43xx which could use parts of that logic?
I mean to indicate that terms like "works on panda" tends always to be relative.
Fair enough.
It is nice to see it as a proof of concept, but I'd hate to see some dead code lying around in kernel and folks blindly following suit and introducing macros for new assembly for a feature that in practice just one group of folks care about and creates additional burden for rest of folks trying to keep that functionality going as we jump from one "device tree" style churn to another "framework"? Not to mean that good features should be kept away.. but personally, I could not find convincing arguments in this case..
I haven't looked at patch myself but as you pointed out if it adds dead code and makes the code un-readable then probably that something we shouldn't merge.
Regards, Santosh
* Santosh Shilimkar santosh.shilimkar@ti.com [140114 15:46]:
On Tuesday 14 January 2014 04:13 PM, Nishanth Menon wrote:
I haven't looked at patch myself but as you pointed out if it adds dead code and makes the code un-readable then probably that something we shouldn't merge.
Yeah it seems the assembly parts should be done in more generic way using macros so the same setup can then be used for other SoCs. For the other trivial changes, let's try to get them merged to shrink down the patchset.
Regards,
Tony
linaro-kernel@lists.linaro.org
-
Nishanth Menon -
Santosh Shilimkar -
Taras Kondratiuk -
Tony Lindgren -
Victor Kamensky