ARM: OMAP3: hwmod data: disable MIDLEMODE control for musb
Subsystems (4)
PATCH 1/1 ARM: OMAP3: hwmod data: disable MIDLEMODE control for musb
For some unknown reason, allowing hwmod to control MIDLEMODE causes core_pwrdm to not hit idle states for musb in DM3730 at least. I've verified that setting any MIDLEMODE value other than "force standby" before enabling the device causes subsequent suspend attempts to fail with core_pwrdm not entering idle states, even if the driver is unloaded and "force standby" is restored before suspend attempt.
Keeping the register set at force standby (reset value) makes it work and device still functions properly. musb has driver-controlled OTG_FORCESTDBY register that controls MSTANDBY signal, so perhaps MIDLEMODE is not even needed? Note that TI PSP kernels also have similar workarounds.
--- arch/arm/mach-omap2/omap_hwmod_3xxx_data.c | 14 +++++++++----- 1 file changed, 9 insertions(+), 5 deletions(-)
arch/arm/mach-omap2/omap_hwmod_3xxx_data.c +9 -5
@@ -1666,11 +1666,15 @@ static struct omap_hwmod_class_sysconfig omap3xxx_usbhsotg_sysc = { .rev_offs = 0x0400, .sysc_offs = 0x0404, .syss_offs = 0x0408, - .sysc_flags = (SYSC_HAS_SIDLEMODE |SYSC_HAS_MIDLEMODE | -SYSC_HAS_ENAWAKEUP |SYSC_HAS_SOFTRESET | - SYSC_HAS_AUTOIDLE), - .idlemodes = (SIDLE_FORCE |SIDLE_NO |SIDLE_SMART | - MSTANDBY_FORCE | MSTANDBY_NO | MSTANDBY_SMART), +/* +* musb has MMIDLEMODE, but if we ever enable the device not in force +* idle mode, core_pwrdm does not enter idle states at least on 36xx. +* Note that musb also has OTG_FORCESTDBY register that the driver +* uses to control MSTANDBY signal manually. +*/ + .sysc_flags = (SYSC_HAS_SIDLEMODE |SYSC_HAS_ENAWAKEUP | + SYSC_HAS_SOFTRESET | SYSC_HAS_AUTOIDLE), + .idlemodes = (SIDLE_FORCE | SIDLE_NO | SIDLE_SMART), .sysc_fields = &omap_hwmod_sysc_type1, };
Signature · 1 line
1.7.9.5
Paul Walmsley
maintainerPATCH 1/1 ARM: OMAP3: hwmod data: disable MIDLEMODE control for musb
Hi Gra?vydas,
Quoted 7 lines · Grazvydas Ignotas
For some unknown reason, allowing hwmod to control MIDLEMODE causes
core_pwrdmto not hit idle states for musb in DM3730 at least. I've verified that setting any MIDLEMODE value other than "force standby" before enabling the device causes subsequent suspend attempts to fail withcore_pwrdmnot entering idle states, even if the driver is unloaded and "force standby" is restored before suspend attempt.
Ugh sounds like a broken bootloader/previous OS could easily block full chip idle in this case :-( Does that match your understanding? That, even if the new kernel does everything right in terms of initialization and reset, the PRCM's perception of whether the device is in STANDBY is permanently stuck?
Quoted 5 lines
Keeping the register set at force standby (reset value) makes it work and device still functions properly. musb has driver-controlled
OTG_FORCESTDBYregister that controls MSTANDBY signal, so perhaps MIDLEMODE is not even needed? Note that TI PSP kernels also have similar workarounds.
Would like to get your opinion on a different implementation. For other situations where IP blocks don't work in the standard, expected way, we've defined hwmod flags for those situations, like HWMOD_SWSUP_*, and HWMOD_NO_OCP_AUTOIDLE. The motivation is to affirmatively mark IP blocks that don't work as expected, rather than changing the existing, documented hardware bits, which are ideally autogenerated.
So what do you think about adding a hwmod flag, HWMOD_FORCE_MSTDBY, and using that in the hwmod code to program the MSTDBYMODE to FORCE_STANDBY and then skipping all other attempts to write to it?
- Paul
PATCH 1/1 ARM: OMAP3: hwmod data: disable MIDLEMODE control for musb
Quoted 12 lines · Paul Walmsley
Hi Gra?vydas,
Quoted 7 lines · Grazvydas Ignotas
For some unknown reason, allowing hwmod to control MIDLEMODE causes
core_pwrdmto not hit idle states for musb in DM3730 at least. I've verified that setting any MIDLEMODE value other than "force standby" before enabling the device causes subsequent suspend attempts to fail withcore_pwrdmnot entering idle states, even if the driver is unloaded and "force standby" is restored before suspend attempt.
Quoted 5 lines
Ugh sounds like a broken bootloader/previous OS could easily block full chip idle in this case :-( Does that match your understanding? That, even if the new kernel does everything right in terms of initialization and reset, the PRCM's perception of whether the device is in STANDBY is permanently stuck?
Soft reset seems to recover from this so there is no problem, but you can't reset before every suspend so a workaround is still needed..
Quoted 13 lines
Quoted 5 lines
Keeping the register set at force standby (reset value) makes it work and device still functions properly. musb has driver-controlled
OTG_FORCESTDBYregister that controls MSTANDBY signal, so perhaps MIDLEMODE is not even needed? Note that TI PSP kernels also have similar workarounds.Would like to get your opinion on a different implementation. For other situations where IP blocks don't work in the standard, expected way, we've defined hwmod flags for those situations, like HWMOD_SWSUP_*, and
HWMOD_NO_OCP_AUTOIDLE. The motivation is to affirmatively mark IP blocks that don't work as expected, rather than changing the existing, documented hardware bits, which are ideally autogenerated.
Quoted 3 lines
So what do you think about adding a hwmod flag,
HWMOD_FORCE_MSTDBY, and using that in the hwmod code to program the MSTDBYMODE toFORCE_STANDBYand then skipping all other attempts to write to it?
Well as long as it works it's good for me, although it'll bloat the code a bit compared to just changing the data. Should I attempt an implementation?
Signature · 1 line
Gra?vydas
Paul Walmsley
maintainerPATCH 1/1 ARM: OMAP3: hwmod data: disable MIDLEMODE control for musb
Hello Gra?vydas,
Quoted 2 lines · Grazvydas Ignotas
Soft reset seems to recover from this so there is no problem, but you can't reset before every suspend so a workaround is still needed..
OK, that's good.
Quoted 3 lines
Well as long as it works it's good for me, although it'll bloat the code a bit compared to just changing the data. Should I attempt an implementation?
Please do. I'd be shocked if it comes out to more than four or five extra ARM instructions...
- Paul