| Message ID | 1351028756-22309-1-git-send-email-swarren@wwwdotorg.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Quoting Stephen Warren (2012-10-23 14:45:56) > What do people think of this? Does it sound like a good idea to go ahead > with a reset subsystem? Should we simply add a new API to the common clock > subsystem instead (and assume that reset and clock domains match 1:1). > Should this be implemented as part of the generic power management domains; > see include/linux/pm_domain.h instead? > Hi Stephen, I'm not sure a "reset subsystem" is necessary, but I also do not like using clocks as the keys for IP reset. I think it is more common to map IPs to struct device, no? And of course for clocks shared by multiple users this will not scale. Regards, Mike
On 10/29/2012 12:32 PM, Mike Turquette wrote: > Quoting Stephen Warren (2012-10-23 14:45:56) >> What do people think of this? Does it sound like a good idea to go ahead >> with a reset subsystem? Should we simply add a new API to the common clock >> subsystem instead (and assume that reset and clock domains match 1:1). >> Should this be implemented as part of the generic power management domains; >> see include/linux/pm_domain.h instead? >> > > Hi Stephen, > > I'm not sure a "reset subsystem" is necessary, but I also do not like > using clocks as the keys for IP reset. I'm not sure what you're suggesting as an alternative to a reset subsystem (or API if you want something that sounds smaller!) :-) > I think it is more common to map IPs to struct device, no? It is indeed probably common that there's a 1:1 mapping between IP blocks and struct device. However, I'm sure there are plenty of counter-examples; IP blocks with multiple reset domains (hence struct devices that encompass multiple reset domains, or reset domains that encompass multiple struct devices), just as there are many examples of non-1:1 mappings between struct device and struct clk. Even ignoring that, we'd still need to API say device_reset(struct device *dev) or device_reset(struct device *dev, const char *conid) wouldn't we? That's really all I meant by a reset subsystem. An alternative here would be to simply move Tegra's tegra_periph_reset_{de,}assert() function prototypes into a header in include/linux rather than mach-tegra/include/mach. However, I imagine at least some other SoC needs a similar API, so a common API might be useful?
Quoting Stephen Warren (2012-10-30 11:02:05) > On 10/29/2012 12:32 PM, Mike Turquette wrote: > > Quoting Stephen Warren (2012-10-23 14:45:56) > >> What do people think of this? Does it sound like a good idea to go ahead > >> with a reset subsystem? Should we simply add a new API to the common clock > >> subsystem instead (and assume that reset and clock domains match 1:1). > >> Should this be implemented as part of the generic power management domains; > >> see include/linux/pm_domain.h instead? > >> > > > > Hi Stephen, > > > > I'm not sure a "reset subsystem" is necessary, but I also do not like > > using clocks as the keys for IP reset. > > I'm not sure what you're suggesting as an alternative to a reset > subsystem (or API if you want something that sounds smaller!) :-) > My point was that I do not know if a new subsystem is necessary or not. Your suggestion to "simply add a new API to the common clock subsystem" is an example of an alternative to a whole new subsystem. However I instinctively feel that the clock api is not the right place for reseting devices. > > I think it is more common to map IPs to struct device, no? > > It is indeed probably common that there's a 1:1 mapping between IP > blocks and struct device. However, I'm sure there are plenty of > counter-examples; IP blocks with multiple reset domains (hence struct > devices that encompass multiple reset domains, or reset domains that > encompass multiple struct devices), just as there are many examples of > non-1:1 mappings between struct device and struct clk. > In OMAP code we handle IP resets through the hwmod code and I prefer that IP-centric approach to associating IP resets with a clock node. Perhaps the hwmod approach could serve as inspiration for a new generic way to reset modules. > Even ignoring that, we'd still need to API say device_reset(struct > device *dev) or device_reset(struct device *dev, const char *conid) > wouldn't we? That's really all I meant by a reset subsystem. > Of course. The api must exist. > An alternative here would be to simply move Tegra's > tegra_periph_reset_{de,}assert() function prototypes into a header in > include/linux rather than mach-tegra/include/mach. However, I imagine at > least some other SoC needs a similar API, so a common API might be useful? I also agree. I think that there was simply some confusion about how I responded. To reiterate, I'm not sure whether a new subsystem should be created or if the API can find a home in some existing subsystem, but I don't think the clock framework is the right place for it. Regards, Mike
On 10/31/2012 04:32 AM, Mike Turquette wrote: > Quoting Stephen Warren (2012-10-30 11:02:05) >> On 10/29/2012 12:32 PM, Mike Turquette wrote: >>> Quoting Stephen Warren (2012-10-23 14:45:56) >>>> What do people think of this? Does it sound like a good idea to go ahead >>>> with a reset subsystem? Should we simply add a new API to the common clock >>>> subsystem instead (and assume that reset and clock domains match 1:1). >>>> Should this be implemented as part of the generic power management domains; >>>> see include/linux/pm_domain.h instead? >>>> >>> >>> Hi Stephen, >>> >>> I'm not sure a "reset subsystem" is necessary, but I also do not like >>> using clocks as the keys for IP reset. >> >> I'm not sure what you're suggesting as an alternative to a reset >> subsystem (or API if you want something that sounds smaller!) :-) > > My point was that I do not know if a new subsystem is necessary or not. > Your suggestion to "simply add a new API to the common clock subsystem" > is an example of an alternative to a whole new subsystem. However I > instinctively feel that the clock api is not the right place for > reseting devices. driver/base/power is about the only related place I can think of given a quick look. However, in a similar way to clocks, I don't think there's necessarily a 1:1 relationship between power domains and reset domains either, so driver/base/power doesn't feel like a good fit in just the same way that drivers/clk doesn't. I wonder if a drivers/base/reset/ or drivers/base/reset.c would be appropriate? >>> I think it is more common to map IPs to struct device, no? >> >> It is indeed probably common that there's a 1:1 mapping between IP >> blocks and struct device. However, I'm sure there are plenty of >> counter-examples; IP blocks with multiple reset domains (hence struct >> devices that encompass multiple reset domains, or reset domains that >> encompass multiple struct devices), just as there are many examples of >> non-1:1 mappings between struct device and struct clk. > > In OMAP code we handle IP resets through the hwmod code and I prefer > that IP-centric approach to associating IP resets with a clock node. > Perhaps the hwmod approach could serve as inspiration for a new generic > way to reset modules. OK, I'm not even slightly familiar with the hwmod code, but I keep hearing about it, so I'll go take a quick look.
diff --git a/Documentation/devicetree/bindings/reset/reset.txt b/Documentation/devicetree/bindings/reset/reset.txt new file mode 100644 index 0000000..31db6ff --- /dev/null +++ b/Documentation/devicetree/bindings/reset/reset.txt @@ -0,0 +1,75 @@ += Reset Signal Device Tree Bindings = + +This binding is intended to represent the hardware reset signals present +internally in most IC (SoC, FPGA, ...) designs. Reset signals for whole +standalone chips are most likely better represented as GPIOs, although there +are likely to be exceptions to this rule. + +Hardware blocks typically receive a reset signal. This signal is generated by +a reset provider (e.g. power management or clock module) and received by a +reset consumer (the module being reset, or a module managing when a sub- +ordinate module is reset). This binding exists to represent the provider and +consumer, and provide a way to couple the two together. + +A reset signal is represented by the phandle of the provider, plus a reset +specifier - a list of DT cells that represents the reset signal within the +provider. The length (number of cells) and semantics of the reset specifier +are dictated by the binding of the reset provider, although common schemes +are described below. + +A word on where to place reset signal consumers in device tree: It is possible +in hardware for a reset signal to affect multiple logically separate HW blocks +at once. In this case, it would be unwise to represent this reset signal in +the DT node of each affected HW block, since if activated, an unrelated block +may be reset. Instead, reset signals should be represented in the DT node +where it makes most sense to control it; this may be a bus node if all +children of the bus are affected by the reset signal, or an individual HW +block node for dedicated reset signals. The intent of this binding is to give +appropriate software access to the reset signals in order to manage the HW, +rather than to slavishly enumerate the reset signal that affects each HW +block. + += Reset providers = + +Required properties: +#reset-cells: Number of cells in a reset specifier; Typically 0 for nodes + with a single reset output and 1 for nodes with multiple + reset outputs. + +For example: + + rst: reset-controller { + #reset-cells = <1>; + }; + += Reset consumers = + +Required properties: +resets: List of phandle and reset specifier pairs, one pair + for each reset signal that affects the device, or that the + device manages. Note: if the reset provider specifies '0' for + #reset-cells, then only the phandle portion of the pair will + appear. + +Optional properties: +reset-names: List of reset signal name strings sorted in the same order as + the resets property. Consumers drivers will use reset-names to + match reset signal names with reset specifiers. + +For example: + + device { + resets = <&rst 20>; + reset-names = "reset"; + }; + +This represents a device with a single reset signal named "reset". + + bus { + resets = <&rst 10> <&rst 11> <&rst 12> <&rst 11>; + reset-names = "i2s1", "i2s2", "dma", "mixer"; + }; + +This represents a bus that controls the reset signal of each of four sub- +ordinate devices. Consider for example a bus that fails to operate unless no +child device has reset asserted.