Hi Loc, Grant,
On Thu, 14 Nov 2013 16:01:13 -0800, Loc Ho lho@xxxxxxx wrote:
Hi Mike,
I will be looking into covering the clock driver to support UEFI. Before I do this, can someone explain to me how x86 system handle each IP clock as I don't see such thing?
UEFI is completely unrelated to the clock drivers, it is only a firmware interface for the OS loader. UEFI mostly disappears after Linux is loaded. Are you perhaps asking about ACPI?
There isn't such a thing on x86. This is entirely new territory. On x86 as far as I know any clock manipulation that does need to be done would be performed by an ACPI method, but there is no concept of an IP clock in the ACPI namespace, and so no standard way of describing them.
Yes, there is no standard way (ACPI device object or ACPI table) to describe clocks for x86 in ACPI spec, but there is a table called GTDT (Generic Timer Description Table) for ARM which contains information for arch timer initialization.
you can refer to ACPI spec 5.0 chapter 5.2.24: http://www.acpi.info/spec50.htm
You'd be much better of talking to Al Stone and Grame Gregory about what their plans are for clock support in ACPI and to post your question to the ACPI mailing list.
We already finished the implementation of convert fixed clock and arch timer to ACPI, and I had already post the RFC patch set for review now:
http://marc.info/?l=linaro-acpi&m=138450991609818&w=2 http://marc.info/?l=linaro-acpi&m=138198249622451&w=2
Thanks Hanjun
Hi,
I will be looking into covering the clock driver to support UEFI. Before I do this, can someone explain to me how x86 system handle each IP clock as I don't see such thing?
UEFI is completely unrelated to the clock drivers, it is only a firmware interface for the OS loader. UEFI mostly disappears after Linux is loaded. Are you perhaps asking about ACPI?
There isn't such a thing on x86. This is entirely new territory. On x86 as far as I know any clock manipulation that does need to be done would be performed by an ACPI method, but there is no concept of an IP clock in the ACPI namespace, and so no standard way of describing them.
Yes, there is no standard way (ACPI device object or ACPI table) to describe clocks for x86 in ACPI spec, but there is a table called GTDT (Generic Timer Description Table) for ARM which contains information for arch timer initialization.
you can refer to ACPI spec 5.0 chapter 5.2.24: http://www.acpi.info/spec50.htm
[Loc Ho] Do you know how an IP (such as SATA controller) clock is enabled for x86? Or x86 is mostly PCIe and the clock is handled inside the driver itself. Another related question would be how x86 enables the PCIe controller clock? Is that too handled inside the host controller driver itself or by some standard method within ACPI?
You'd be much better of talking to Al Stone and Grame Gregory about what their plans are for clock support in ACPI and to post your question to the ACPI mailing list.
We already finished the implementation of convert fixed clock and arch timer to ACPI, and I had already post the RFC patch set for review now:
[Loc Ho] Looked over this one. We will take this patch into our kernel.
[Loc Ho]
From what I can tell with DT, the only reason that an IP would need
this would be an common clock interface for framework to enable an IP clock. If the framework expected an clock instance and one isn't available, it will fail. Other than that, if an IP can handle its clock initialization, do we really need this at all. For reference clock and a few others, this is required if an IP clock is scaled from an reference clock for say. I would like an agreement on the IP clock itself. Is it required or they will be handled by the IP driver itself?
Please be aware of the following limitation that we (at APM) ran into and looking for an fix:
1. The fixed 32-bit memory resource can NOT be overlapped with another memory resource with kernel version below 3.12. Don't know if this limitation still existed in 3.12 kernel. May be this limit is removed if one use 64-bit resource address but hasn't tried yet. 2. Overlap memory resource may be required if the same clock resource combined other reset fields into the same registers. For this, one can NOT use the standard _CRS as the same region may be defined with the IP driver. Don't see an solution besides to use an non-standard address. Then, what do one put in the _CRS as the ACPI expect one. If you use an empty entry with 0 for address and 0 for length, other OS vendor will complaint as well.
-Loc
On 2013-11-18 13:57, Loc Ho wrote:
Hi,
I will be looking into covering the clock driver to support UEFI. Before I do this, can someone explain to me how x86 system handle each IP clock as I don't see such thing?
UEFI is completely unrelated to the clock drivers, it is only a firmware interface for the OS loader. UEFI mostly disappears after Linux is loaded. Are you perhaps asking about ACPI?
There isn't such a thing on x86. This is entirely new territory. On x86 as far as I know any clock manipulation that does need to be done would be performed by an ACPI method, but there is no concept of an IP clock in the ACPI namespace, and so no standard way of describing them.
Yes, there is no standard way (ACPI device object or ACPI table) to describe clocks for x86 in ACPI spec, but there is a table called GTDT (Generic Timer Description Table) for ARM which contains information for arch timer initialization.
you can refer to ACPI spec 5.0 chapter 5.2.24: http://www.acpi.info/spec50.htm
[Loc Ho] Do you know how an IP (such as SATA controller) clock is enabled for x86? Or x86 is mostly PCIe and the clock is handled inside the driver
Sorry, I have no idea about this :(
itself. Another related question would be how x86 enables the PCIe controller clock? Is that too handled inside the host controller driver itself or by some standard method within ACPI?
AFAIK, there is a HPET table for x86 for timer init, but there is no standard method within ACPI to get timer for devices in DSDT.
You'd be much better of talking to Al Stone and Grame Gregory about what their plans are for clock support in ACPI and to post your question to the ACPI mailing list.
We already finished the implementation of convert fixed clock and arch timer to ACPI, and I had already post the RFC patch set for review now:
[Loc Ho] Looked over this one. We will take this patch into our kernel.
It is RFC, still needs some update, I will send another version soon. I will send to linaro-acpi@lists.linaro.org, it is a public mailing list.
[Loc Ho] From what I can tell with DT, the only reason that an IP would need this would be an common clock interface for framework to enable an IP clock. If the framework expected an clock instance and one isn't available, it will fail. Other than that, if an IP can handle its clock initialization, do we really need this at all. For reference clock and a few others, this is required if an IP clock is scaled from an reference clock for say. I would like an agreement on the IP clock itself. Is it required or they will be handled by the IP driver itself?
Please be aware of the following limitation that we (at APM) ran into and looking for an fix:
- The fixed 32-bit memory resource can NOT be overlapped with another
memory resource with kernel version below 3.12. Don't know if this limitation still existed in 3.12 kernel. May be this limit is removed if one use 64-bit resource address but hasn't tried yet. 2. Overlap memory resource may be required if the same clock resource combined other reset fields into the same registers. For this, one can NOT use the standard _CRS as the same region may be defined with the IP driver. Don't see an solution besides to use an non-standard address. Then, what do one put in the _CRS as the ACPI expect one. If you use an empty entry with 0 for address and 0 for length, other OS vendor will complaint as well.
I didn't catch up with the limitation clearly, if any specific example, that would be great.
Thanks Hanjun