| Message ID | 1422978684-4826-1-git-send-email-grygorii.strashko@linaro.org |
|---|---|
| State | New |
| Headers | show |
Hi, On Tuesday 03 February 2015 09:21 PM, grygorii.strashko@linaro.org wrote: > From: Grygorii Strashko <grygorii.strashko@linaro.org> > > Now DRA7xx pcie1/2 hwmods define PRCM configuration as following: > .clkctrl_offs = DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET, > .rstctrl_offs = DRA7XX_RM_L3INIT_RSTCTRL_OFFSET, > .modulemode = MODULEMODE_SWCTRL, > which is completely wrong because DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET > is clockdomain ctrl register and NOT module ctrl register. > And they have diffrent allowed values for bits[0,1]: > CLKTRCTRL MODULEMODE > 0x0: NO_SLEEP 0x0: Module is disabled by SW. > 0x1: SW_SLEEP 0x1: Module is managed automatically by HW > 0x2: SW_WKUP 0x2: Module is explicitly enabled. > 0x3: HW_AUTO 0x3: Reserved > > As result, following message can be seen during suspend: > "omap_hwmod: pcie1: _wait_target_disable failed" > > Fix it by removing .modulemode from pcie1/2 hwmods and, in that > way, prevent clockdomain ctrl register writing from HWMOD core. Looks correct except for one change. Acked-by: Kishon Vijay Abraham I <kishon@ti.com> > > Signed-off-by: Grygorii Strashko <Grygorii.Strashko@linaro.org> > --- > > More over, it looks like pcie1/2 hwmods are fake and have to be dropped at all. > The real HWMODs are PCIESS1/2. Not sure I get this. You mean "dra7xx_pcie1_hwmod" should be replaced with "dra7xx_pciess1_hwmod"? Or you mean an entire new hwmod is missing? Please note we still have to enable the clock domain and main clock. We've also purposefully omitted sysconfig from hwmod data since pcie reset (RM_PCIESS_RSTCTRL) should be done before accessing the syconfig register and the infrastructure for that is currently not present. > Unfortunatelly, not all information on PCIE is public, so > I could be wrong here. > --- > arch/arm/mach-omap2/omap_hwmod_7xx_data.c | 2 -- > 1 file changed, 2 deletions(-) > > diff --git a/arch/arm/mach-omap2/omap_hwmod_7xx_data.c b/arch/arm/mach-omap2/omap_hwmod_7xx_data.c > index ffd6604..a428b2d 100644 > --- a/arch/arm/mach-omap2/omap_hwmod_7xx_data.c > +++ b/arch/arm/mach-omap2/omap_hwmod_7xx_data.c > @@ -1478,7 +1478,6 @@ static struct omap_hwmod dra7xx_pcie1_hwmod = { > .prcm = { > .omap4 = { > .clkctrl_offs = DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET, > - .modulemode = MODULEMODE_SWCTRL, I think the entire .prcm can be removed here. > }, > }, > }; > @@ -1492,7 +1491,6 @@ static struct omap_hwmod dra7xx_pcie2_hwmod = { > .prcm = { > .omap4 = { > .clkctrl_offs = DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET, > - .modulemode = MODULEMODE_SWCTRL, > }, > }, > }; > Thanks Kishon
Hi Kishon, On 02/09/2015 04:50 PM, Kishon Vijay Abraham I wrote: > On Tuesday 03 February 2015 09:21 PM, grygorii.strashko@linaro.org wrote: >> From: Grygorii Strashko <grygorii.strashko@linaro.org> >> >> Now DRA7xx pcie1/2 hwmods define PRCM configuration as following: >> .clkctrl_offs = DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET, >> .rstctrl_offs = DRA7XX_RM_L3INIT_RSTCTRL_OFFSET, >> .modulemode = MODULEMODE_SWCTRL, >> which is completely wrong because DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET >> is clockdomain ctrl register and NOT module ctrl register. >> And they have diffrent allowed values for bits[0,1]: >> CLKTRCTRL MODULEMODE >> 0x0: NO_SLEEP 0x0: Module is disabled by SW. >> 0x1: SW_SLEEP 0x1: Module is managed automatically by HW >> 0x2: SW_WKUP 0x2: Module is explicitly enabled. >> 0x3: HW_AUTO 0x3: Reserved >> >> As result, following message can be seen during suspend: >> "omap_hwmod: pcie1: _wait_target_disable failed" >> >> Fix it by removing .modulemode from pcie1/2 hwmods and, in that >> way, prevent clockdomain ctrl register writing from HWMOD core. > > Looks correct except for one change. > > Acked-by: Kishon Vijay Abraham I <kishon@ti.com> >> >> Signed-off-by: Grygorii Strashko <Grygorii.Strashko@linaro.org> >> --- >> >> More over, it looks like pcie1/2 hwmods are fake and have to be dropped at all. >> The real HWMODs are PCIESS1/2. > > Not sure I get this. You mean "dra7xx_pcie1_hwmod" should be replaced with > "dra7xx_pciess1_hwmod"? Or you mean an entire new hwmod is missing? > > Please note we still have to enable the clock domain and main clock. We've also > purposefully omitted sysconfig from hwmod data since pcie reset > (RM_PCIESS_RSTCTRL) should be done before accessing the syconfig register and > the infrastructure for that is currently not present. What I'm trying to say is that now PM control data mixed between "pcieX" and "pcieX-phy" hwmods. After this patch "pcieX" hwmods will actually do nothing (I assume that "pciex-phy" will be enabled before "pcieX"), and probably can be removed if "pcie_clkdm" could be attached to "pcieX-phy" hwmod instead. More over, now, "pcie_clkdm" is connected to "pcieX" hwmod while MODULEMODE register is controlled by "pciex-phy" hwmod, so when pciess is going to be enabled the "l3init_clkdm" will be waken-up by hwmode core and not "pcie_clkdm" - as I can remember this is not good (we should alway wake-up clockdomain and keep it in SWSUP mode while changing MODULEMODE and SYSC registers). static struct omap_hwmod dra7xx_pcie1_hwmod = { .name = "pcie1", .class = &dra7xx_pcie_hwmod_class, .clkdm_name = "pcie_clkdm", .main_clk = "l4_root_clk_div", static struct omap_hwmod dra7xx_pcie1_phy_hwmod = { .name = "pcie1-phy", .class = &dra7xx_pcie_phy_hwmod_class, .clkdm_name = "l3init_clkdm", .main_clk = "l4_root_clk_div", So, in my opinion, some rework may be needed here. Am I right? > >> Unfortunatelly, not all information on PCIE is public, so >> I could be wrong here. >> --- >> arch/arm/mach-omap2/omap_hwmod_7xx_data.c | 2 -- >> 1 file changed, 2 deletions(-) >> >> diff --git a/arch/arm/mach-omap2/omap_hwmod_7xx_data.c b/arch/arm/mach-omap2/omap_hwmod_7xx_data.c >> index ffd6604..a428b2d 100644 >> --- a/arch/arm/mach-omap2/omap_hwmod_7xx_data.c >> +++ b/arch/arm/mach-omap2/omap_hwmod_7xx_data.c >> @@ -1478,7 +1478,6 @@ static struct omap_hwmod dra7xx_pcie1_hwmod = { >> .prcm = { >> .omap4 = { >> .clkctrl_offs = DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET, >> - .modulemode = MODULEMODE_SWCTRL, > > I think the entire .prcm can be removed here. not sure. I've tried it and Kernel boot failed (on 3.14) >> }, >> }, >> }; >> @@ -1492,7 +1491,6 @@ static struct omap_hwmod dra7xx_pcie2_hwmod = { >> .prcm = { >> .omap4 = { >> .clkctrl_offs = DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET, >> - .modulemode = MODULEMODE_SWCTRL, >> }, >> }, >> }; >> > > Thanks > Kishon >
Hi, On Monday 09 February 2015 03:58 PM, Grygorii.Strashko@linaro.org wrote: > Hi Kishon, > On 02/09/2015 04:50 PM, Kishon Vijay Abraham I wrote: >> On Tuesday 03 February 2015 09:21 PM, grygorii.strashko@linaro.org wrote: >>> From: Grygorii Strashko <grygorii.strashko@linaro.org> >>> >>> Now DRA7xx pcie1/2 hwmods define PRCM configuration as following: >>> .clkctrl_offs = DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET, >>> .rstctrl_offs = DRA7XX_RM_L3INIT_RSTCTRL_OFFSET, >>> .modulemode = MODULEMODE_SWCTRL, >>> which is completely wrong because DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET >>> is clockdomain ctrl register and NOT module ctrl register. >>> And they have diffrent allowed values for bits[0,1]: >>> CLKTRCTRL MODULEMODE >>> 0x0: NO_SLEEP 0x0: Module is disabled by SW. >>> 0x1: SW_SLEEP 0x1: Module is managed automatically by HW >>> 0x2: SW_WKUP 0x2: Module is explicitly enabled. >>> 0x3: HW_AUTO 0x3: Reserved >>> >>> As result, following message can be seen during suspend: >>> "omap_hwmod: pcie1: _wait_target_disable failed" >>> >>> Fix it by removing .modulemode from pcie1/2 hwmods and, in that >>> way, prevent clockdomain ctrl register writing from HWMOD core. >> >> Looks correct except for one change. >> >> Acked-by: Kishon Vijay Abraham I <kishon@ti.com> >>> >>> Signed-off-by: Grygorii Strashko <Grygorii.Strashko@linaro.org> >>> --- >>> >>> More over, it looks like pcie1/2 hwmods are fake and have to be dropped at all. >>> The real HWMODs are PCIESS1/2. >> >> Not sure I get this. You mean "dra7xx_pcie1_hwmod" should be replaced with >> "dra7xx_pciess1_hwmod"? Or you mean an entire new hwmod is missing? >> >> Please note we still have to enable the clock domain and main clock. We've also >> purposefully omitted sysconfig from hwmod data since pcie reset >> (RM_PCIESS_RSTCTRL) should be done before accessing the syconfig register and >> the infrastructure for that is currently not present. > > What I'm trying to say is that now PM control data mixed between "pcieX" and "pcieX-phy" hwmods. > After this patch "pcieX" hwmods will actually do nothing (I assume that "pciex-phy" will be > enabled before "pcieX"), and probably can be removed if "pcie_clkdm" could be attached to "pcieX-phy" hwmod > instead. > > More over, now, "pcie_clkdm" is connected to "pcieX" hwmod while MODULEMODE register is controlled > by "pciex-phy" hwmod, so when pciess is going to be enabled the "l3init_clkdm" will be waken-up by > hwmode core and not "pcie_clkdm" - as I can remember this is not good (we should alway wake-up clockdomain > and keep it in SWSUP mode while changing MODULEMODE and SYSC registers). > > static struct omap_hwmod dra7xx_pcie1_hwmod = { > .name = "pcie1", > .class = &dra7xx_pcie_hwmod_class, > .clkdm_name = "pcie_clkdm", > .main_clk = "l4_root_clk_div", > > static struct omap_hwmod dra7xx_pcie1_phy_hwmod = { > .name = "pcie1-phy", > .class = &dra7xx_pcie_phy_hwmod_class, > .clkdm_name = "l3init_clkdm", > .main_clk = "l4_root_clk_div", > > So, in my opinion, some rework may be needed here. > Am I right? you are right. We should have a single hwmod like dra7xx_pciess1_hwmod whose clkdm should be "pcie_clkdm" and whose clkctrl_offs should be DRA7XX_CM_L3INIT_PCIESS1_CLKCTRL_OFFSET (for controlling MODULEMODE). PCIE PHY shouldn't have a hwmod entry at all. Thanks Kishon
On 02/09/2015 09:24 PM, Kishon Vijay Abraham I wrote: > Hi, > > On Monday 09 February 2015 03:58 PM, Grygorii.Strashko@linaro.org wrote: >> Hi Kishon, >> On 02/09/2015 04:50 PM, Kishon Vijay Abraham I wrote: >>> On Tuesday 03 February 2015 09:21 PM, grygorii.strashko@linaro.org wrote: >>>> From: Grygorii Strashko <grygorii.strashko@linaro.org> >>>> >>>> Now DRA7xx pcie1/2 hwmods define PRCM configuration as following: >>>> .clkctrl_offs = DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET, >>>> .rstctrl_offs = DRA7XX_RM_L3INIT_RSTCTRL_OFFSET, >>>> .modulemode = MODULEMODE_SWCTRL, >>>> which is completely wrong because DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET >>>> is clockdomain ctrl register and NOT module ctrl register. >>>> And they have diffrent allowed values for bits[0,1]: >>>> CLKTRCTRL MODULEMODE >>>> 0x0: NO_SLEEP 0x0: Module is disabled by SW. >>>> 0x1: SW_SLEEP 0x1: Module is managed automatically by HW >>>> 0x2: SW_WKUP 0x2: Module is explicitly enabled. >>>> 0x3: HW_AUTO 0x3: Reserved >>>> >>>> As result, following message can be seen during suspend: >>>> "omap_hwmod: pcie1: _wait_target_disable failed" >>>> >>>> Fix it by removing .modulemode from pcie1/2 hwmods and, in that >>>> way, prevent clockdomain ctrl register writing from HWMOD core. >>> >>> Looks correct except for one change. >>> >>> Acked-by: Kishon Vijay Abraham I <kishon@ti.com> >>>> >>>> Signed-off-by: Grygorii Strashko <Grygorii.Strashko@linaro.org> >>>> --- >>>> >>>> More over, it looks like pcie1/2 hwmods are fake and have to be dropped at all. >>>> The real HWMODs are PCIESS1/2. >>> >>> Not sure I get this. You mean "dra7xx_pcie1_hwmod" should be replaced with >>> "dra7xx_pciess1_hwmod"? Or you mean an entire new hwmod is missing? >>> >>> Please note we still have to enable the clock domain and main clock. We've also >>> purposefully omitted sysconfig from hwmod data since pcie reset >>> (RM_PCIESS_RSTCTRL) should be done before accessing the syconfig register and >>> the infrastructure for that is currently not present. >> >> What I'm trying to say is that now PM control data mixed between "pcieX" and "pcieX-phy" hwmods. >> After this patch "pcieX" hwmods will actually do nothing (I assume that "pciex-phy" will be >> enabled before "pcieX"), and probably can be removed if "pcie_clkdm" could be attached to "pcieX-phy" hwmod >> instead. >> >> More over, now, "pcie_clkdm" is connected to "pcieX" hwmod while MODULEMODE register is controlled >> by "pciex-phy" hwmod, so when pciess is going to be enabled the "l3init_clkdm" will be waken-up by >> hwmode core and not "pcie_clkdm" - as I can remember this is not good (we should alway wake-up clockdomain >> and keep it in SWSUP mode while changing MODULEMODE and SYSC registers). >> >> static struct omap_hwmod dra7xx_pcie1_hwmod = { >> .name = "pcie1", >> .class = &dra7xx_pcie_hwmod_class, >> .clkdm_name = "pcie_clkdm", >> .main_clk = "l4_root_clk_div", >> >> static struct omap_hwmod dra7xx_pcie1_phy_hwmod = { >> .name = "pcie1-phy", >> .class = &dra7xx_pcie_phy_hwmod_class, >> .clkdm_name = "l3init_clkdm", >> .main_clk = "l4_root_clk_div", >> >> So, in my opinion, some rework may be needed here. >> Am I right? > > you are right. We should have a single hwmod like dra7xx_pciess1_hwmod whose > clkdm should be "pcie_clkdm" and whose clkctrl_offs should be > DRA7XX_CM_L3INIT_PCIESS1_CLKCTRL_OFFSET (for controlling MODULEMODE). PCIE PHY > shouldn't have a hwmod entry at all. > > Thanks > Kishon > Could this patch be applied any way? It fixes real issue for me.
diff --git a/arch/arm/mach-omap2/omap_hwmod_7xx_data.c b/arch/arm/mach-omap2/omap_hwmod_7xx_data.c index ffd6604..a428b2d 100644 --- a/arch/arm/mach-omap2/omap_hwmod_7xx_data.c +++ b/arch/arm/mach-omap2/omap_hwmod_7xx_data.c @@ -1478,7 +1478,6 @@ static struct omap_hwmod dra7xx_pcie1_hwmod = { .prcm = { .omap4 = { .clkctrl_offs = DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET, - .modulemode = MODULEMODE_SWCTRL, }, }, }; @@ -1492,7 +1491,6 @@ static struct omap_hwmod dra7xx_pcie2_hwmod = { .prcm = { .omap4 = { .clkctrl_offs = DRA7XX_CM_PCIE_CLKSTCTRL_OFFSET, - .modulemode = MODULEMODE_SWCTRL, }, }, };