diff mbox series

[v4,08/21] dt-bindings: reserved-memory: Add qcom,ramoops binding

Message ID 1687955688-20809-9-git-send-email-quic_mojha@quicinc.com
State New
Headers show
Series [v4,01/21] docs: qcom: Add qualcomm minidump guide | expand

Commit Message

Mukesh Ojha June 28, 2023, 12:34 p.m. UTC
Qualcomm ramoops minidump logger provide a means of storing
the ramoops data to some dynamically reserved memory instead
of traditionally implemented ramoops where the region should
be statically fixed ram region. Its device tree binding
would be exactly same as ramoops device tree binding and is
going to contain traditional ramoops frontend data and this
content will be collected via Qualcomm minidump infrastructure
provided from the boot firmware.

Signed-off-by: Mukesh Ojha <quic_mojha@quicinc.com>
---
 .../devicetree/bindings/soc/qcom/qcom,ramoops.yaml | 126 +++++++++++++++++++++
 1 file changed, 126 insertions(+)
 create mode 100644 Documentation/devicetree/bindings/soc/qcom/qcom,ramoops.yaml

Comments

Krzysztof Kozlowski July 2, 2023, 8:12 a.m. UTC | #1
On 28/06/2023 17:01, Mukesh Ojha wrote:
> 
> 
> On 6/28/2023 8:17 PM, Rob Herring wrote:
>> On Wed, Jun 28, 2023 at 6:36 AM Mukesh Ojha <quic_mojha@quicinc.com> wrote:
>>>
>>> Qualcomm ramoops minidump logger provide a means of storing
>>> the ramoops data to some dynamically reserved memory instead
>>> of traditionally implemented ramoops where the region should
>>> be statically fixed ram region. Its device tree binding
>>> would be exactly same as ramoops device tree binding and is
>>> going to contain traditional ramoops frontend data and this
>>> content will be collected via Qualcomm minidump infrastructure
>>> provided from the boot firmware.
>>
>> The big difference is if firmware is not deciding where this log
>> lives, then it doesn't need to be in DT. How does anything except the
>> kernel that allocates the log find the logs?
> 
> Yes, you are correct, firmware is not deciding where the logs lives
> instead here, Kernel has reserved the region where the ramoops region
> lives and later with the minidump registration where, physical
> address/size/virtual address(for parsing) are passed and that is how
> firmware is able to know and dump those region before triggering system
> reset.

Your explanation does not justify storing all this in DT. Kernel can
allocate any memory it wishes, store there logs and pass the address to
the firmware. That's it, no need for DT.

> 
> A part of this registration code you can find in 11/21
> 
>> I'm pretty sure I already said all this before.
> 
> Yes, you said this before but that's the reason i came up with vendor
> ramoops instead of changing traditional ramoops binding.

That's unexpected conclusion. Adding more bindings is not the answer to
comment that it should not be in the DTS in the first place.

Best regards,
Krzysztof
Mukesh Ojha July 3, 2023, 6:21 a.m. UTC | #2
On 7/2/2023 1:42 PM, Krzysztof Kozlowski wrote:
> On 28/06/2023 17:01, Mukesh Ojha wrote:
>>
>>
>> On 6/28/2023 8:17 PM, Rob Herring wrote:
>>> On Wed, Jun 28, 2023 at 6:36 AM Mukesh Ojha <quic_mojha@quicinc.com> wrote:
>>>>
>>>> Qualcomm ramoops minidump logger provide a means of storing
>>>> the ramoops data to some dynamically reserved memory instead
>>>> of traditionally implemented ramoops where the region should
>>>> be statically fixed ram region. Its device tree binding
>>>> would be exactly same as ramoops device tree binding and is
>>>> going to contain traditional ramoops frontend data and this
>>>> content will be collected via Qualcomm minidump infrastructure
>>>> provided from the boot firmware.
>>>
>>> The big difference is if firmware is not deciding where this log
>>> lives, then it doesn't need to be in DT. How does anything except the
>>> kernel that allocates the log find the logs?
>>
>> Yes, you are correct, firmware is not deciding where the logs lives
>> instead here, Kernel has reserved the region where the ramoops region
>> lives and later with the minidump registration where, physical
>> address/size/virtual address(for parsing) are passed and that is how
>> firmware is able to know and dump those region before triggering system
>> reset.
> 
> Your explanation does not justify storing all this in DT. Kernel can
> allocate any memory it wishes, store there logs and pass the address to
> the firmware. That's it, no need for DT.

If you go through the driver, you will know that what it does, is
just create platform device for actual ramoops driver to probe and to
provide this it needs exact set of parameters of input what original 
ramoops DT provides, we need to keep it in DT as maintaining this in
driver will not scale well with different size/parameter size
requirement for different targets.

> 
>>
>> A part of this registration code you can find in 11/21
>>
>>> I'm pretty sure I already said all this before.
>>
>> Yes, you said this before but that's the reason i came up with vendor
>> ramoops instead of changing traditional ramoops binding.
> 
> That's unexpected conclusion. Adding more bindings is not the answer to
> comment that it should not be in the DTS in the first place.

Please suggest, what is the other way being above text as requirement..

-- Mukesh
> 
> Best regards,
> Krzysztof
>
Krzysztof Kozlowski July 3, 2023, 7:20 a.m. UTC | #3
On Mon, 3 Jul 2023 at 08:22, Mukesh Ojha <quic_mojha@quicinc.com> wrote:
> On 7/2/2023 1:42 PM, Krzysztof Kozlowski wrote:
> >>> The big difference is if firmware is not deciding where this log
> >>> lives, then it doesn't need to be in DT. How does anything except the
> >>> kernel that allocates the log find the logs?
> >>
> >> Yes, you are correct, firmware is not deciding where the logs lives
> >> instead here, Kernel has reserved the region where the ramoops region
> >> lives and later with the minidump registration where, physical
> >> address/size/virtual address(for parsing) are passed and that is how
> >> firmware is able to know and dump those region before triggering system
> >> reset.
> >
> > Your explanation does not justify storing all this in DT. Kernel can
> > allocate any memory it wishes, store there logs and pass the address to
> > the firmware. That's it, no need for DT.
>
> If you go through the driver, you will know that what it does, is

We talk about bindings and I should not be forced to look at the
driver to be able to understand them. Bindings should stand on their
own.

> just create platform device for actual ramoops driver to probe and to

Not really justification for Devicetree anyway. Whatever your driver
is doing, is driver's business, not bindings.

> provide this it needs exact set of parameters of input what original
> ramoops DT provides, we need to keep it in DT as maintaining this in
> driver will not scale well with different size/parameter size
> requirement for different targets.

Really? Why? I don't see a problem in scaling. At all.

>
> >
> >>
> >> A part of this registration code you can find in 11/21
> >>
> >>> I'm pretty sure I already said all this before.
> >>
> >> Yes, you said this before but that's the reason i came up with vendor
> >> ramoops instead of changing traditional ramoops binding.
> >
> > That's unexpected conclusion. Adding more bindings is not the answer to
> > comment that it should not be in the DTS in the first place.
>
> Please suggest, what is the other way being above text as requirement..

I do not see any requirement for us there. Forcing me to figure out
how to add non-hardware property to DT is not the way to convince
reviewers. But if you insist - we have ABI for this, called sysfs. If
it is debugging feature, then debugfs.

Best regards,
Krzysztof
Mukesh Ojha July 3, 2023, 3:55 p.m. UTC | #4
On 7/3/2023 12:50 PM, Krzysztof Kozlowski wrote:
> On Mon, 3 Jul 2023 at 08:22, Mukesh Ojha <quic_mojha@quicinc.com> wrote:
>> On 7/2/2023 1:42 PM, Krzysztof Kozlowski wrote:
>>>>> The big difference is if firmware is not deciding where this log
>>>>> lives, then it doesn't need to be in DT. How does anything except the
>>>>> kernel that allocates the log find the logs?
>>>>
>>>> Yes, you are correct, firmware is not deciding where the logs lives
>>>> instead here, Kernel has reserved the region where the ramoops region
>>>> lives and later with the minidump registration where, physical
>>>> address/size/virtual address(for parsing) are passed and that is how
>>>> firmware is able to know and dump those region before triggering system
>>>> reset.
>>>
>>> Your explanation does not justify storing all this in DT. Kernel can
>>> allocate any memory it wishes, store there logs and pass the address to
>>> the firmware. That's it, no need for DT.
>>
>> If you go through the driver, you will know that what it does, is
> 
> We talk about bindings and I should not be forced to look at the
> driver to be able to understand them. Bindings should stand on their
> own.

Why can't ramoops binding have one more feature where it can add a flag 
*dynamic* to indicate the regions are dynamic and it is for platforms
where there is another entity 'minidump' who is interested in these
regions.

> 
>> just create platform device for actual ramoops driver to probe and to
> 
> Not really justification for Devicetree anyway. Whatever your driver
> is doing, is driver's business, not bindings.
> 
>> provide this it needs exact set of parameters of input what original
>> ramoops DT provides, we need to keep it in DT as maintaining this in
>> driver will not scale well with different size/parameter size
>> requirement for different targets.
> 
> Really? Why? I don't see a problem in scaling. At all.

I had attempted it here,

https://lore.kernel.org/lkml/1683133352-10046-10-git-send-email-quic_mojha@quicinc.com/

but got comments related to hard coding and some in favor of having
the same set of properties what ramoops has/provides

https://lore.kernel.org/lkml/e25723bf-be85-b458-a84c-1a45392683bb@gmail.com/

https://lore.kernel.org/lkml/202305161347.80204C1A0E@keescook/
> 
>>
>>>
>>>>
>>>> A part of this registration code you can find in 11/21
>>>>
>>>>> I'm pretty sure I already said all this before.
>>>>
>>>> Yes, you said this before but that's the reason i came up with vendor
>>>> ramoops instead of changing traditional ramoops binding.
>>>
>>> That's unexpected conclusion. Adding more bindings is not the answer to
>>> comment that it should not be in the DTS in the first place.
>>
>> Please suggest, what is the other way being above text as requirement..
> 
> I do not see any requirement for us there. Forcing me to figure out
> how to add non-hardware property to DT is not the way to convince
> reviewers. But if you insist - we have ABI for this, called sysfs. If
> it is debugging feature, then debugfs.

ramoops already support module params and a way to pass these parameters
from bootargs but it also need to know the hard-codes addresses, so, 
doing something in sysfs will be again duplication with ramoops driver..

If this can be accommodated under ramoops, this will be very small 
change, like this

https://lore.kernel.org/lkml/20230622005213.458236-1-isaacmanjarres@google.com/

-- Mukesh
> 
> Best regards,
> Krzysztof
Krzysztof Kozlowski July 4, 2023, 5:57 a.m. UTC | #5
On 03/07/2023 17:55, Mukesh Ojha wrote:
> 
> 
> On 7/3/2023 12:50 PM, Krzysztof Kozlowski wrote:
>> On Mon, 3 Jul 2023 at 08:22, Mukesh Ojha <quic_mojha@quicinc.com> wrote:
>>> On 7/2/2023 1:42 PM, Krzysztof Kozlowski wrote:
>>>>>> The big difference is if firmware is not deciding where this log
>>>>>> lives, then it doesn't need to be in DT. How does anything except the
>>>>>> kernel that allocates the log find the logs?
>>>>>
>>>>> Yes, you are correct, firmware is not deciding where the logs lives
>>>>> instead here, Kernel has reserved the region where the ramoops region
>>>>> lives and later with the minidump registration where, physical
>>>>> address/size/virtual address(for parsing) are passed and that is how
>>>>> firmware is able to know and dump those region before triggering system
>>>>> reset.
>>>>
>>>> Your explanation does not justify storing all this in DT. Kernel can
>>>> allocate any memory it wishes, store there logs and pass the address to
>>>> the firmware. That's it, no need for DT.
>>>
>>> If you go through the driver, you will know that what it does, is
>>
>> We talk about bindings and I should not be forced to look at the
>> driver to be able to understand them. Bindings should stand on their
>> own.
> 
> Why can't ramoops binding have one more feature where it can add a flag 
> *dynamic* to indicate the regions are dynamic and it is for platforms
> where there is another entity 'minidump' who is interested in these
> regions.

Because we do not define dynamic stuff in Devicetree. Dynamic means
defined by SW or runtime configurable. It is against the entire idea of
Devicetree which is for non-discoverable hardware.

> 
>>
>>> just create platform device for actual ramoops driver to probe and to
>>
>> Not really justification for Devicetree anyway. Whatever your driver
>> is doing, is driver's business, not bindings.
>>
>>> provide this it needs exact set of parameters of input what original
>>> ramoops DT provides, we need to keep it in DT as maintaining this in
>>> driver will not scale well with different size/parameter size
>>> requirement for different targets.
>>
>> Really? Why? I don't see a problem in scaling. At all.
> 
> I had attempted it here,
> 
> https://lore.kernel.org/lkml/1683133352-10046-10-git-send-email-quic_mojha@quicinc.com/
> 
> but got comments related to hard coding and some in favor of having
> the same set of properties what ramoops has/provides
> 
> https://lore.kernel.org/lkml/e25723bf-be85-b458-a84c-1a45392683bb@gmail.com/
> 
> https://lore.kernel.org/lkml/202305161347.80204C1A0E@keescook/

Then you were tricked. I don't get why someone else suggests that
non-hardware property should be part of Devicetree, but anyway it's the
call of Devicetree binding maintainers, not someone else. DT is not
dumping ground for all the system configuration variables.


>>
>>>
>>>>
>>>>>
>>>>> A part of this registration code you can find in 11/21
>>>>>
>>>>>> I'm pretty sure I already said all this before.
>>>>>
>>>>> Yes, you said this before but that's the reason i came up with vendor
>>>>> ramoops instead of changing traditional ramoops binding.
>>>>
>>>> That's unexpected conclusion. Adding more bindings is not the answer to
>>>> comment that it should not be in the DTS in the first place.
>>>
>>> Please suggest, what is the other way being above text as requirement..
>>
>> I do not see any requirement for us there. Forcing me to figure out
>> how to add non-hardware property to DT is not the way to convince
>> reviewers. But if you insist - we have ABI for this, called sysfs. If
>> it is debugging feature, then debugfs.
> 
> ramoops already support module params and a way to pass these parameters
> from bootargs but it also need to know the hard-codes addresses, so, 
> doing something in sysfs will be again duplication with ramoops driver..

Why do you need hard-coded addresses?

> 
> If this can be accommodated under ramoops, this will be very small 
> change, like this
> 
> https://lore.kernel.org/lkml/20230622005213.458236-1-isaacmanjarres@google.com/

That's also funny patch - missing bindings updated, missing CC DT
maintainers.

Best regards,
Krzysztof
Luca Stefani July 19, 2023, 10:29 a.m. UTC | #6
On 04/07/23 07:57, Krzysztof Kozlowski wrote:
> On 03/07/2023 17:55, Mukesh Ojha wrote:
>>
>> On 7/3/2023 12:50 PM, Krzysztof Kozlowski wrote:
>>> On Mon, 3 Jul 2023 at 08:22, Mukesh Ojha <quic_mojha@quicinc.com> wrote:
>>>> On 7/2/2023 1:42 PM, Krzysztof Kozlowski wrote:
>>>>>>> The big difference is if firmware is not deciding where this log
>>>>>>> lives, then it doesn't need to be in DT. How does anything except the
>>>>>>> kernel that allocates the log find the logs?
>>>>>> Yes, you are correct, firmware is not deciding where the logs lives
>>>>>> instead here, Kernel has reserved the region where the ramoops region
>>>>>> lives and later with the minidump registration where, physical
>>>>>> address/size/virtual address(for parsing) are passed and that is how
>>>>>> firmware is able to know and dump those region before triggering system
>>>>>> reset.
>>>>> Your explanation does not justify storing all this in DT. Kernel can
>>>>> allocate any memory it wishes, store there logs and pass the address to
>>>>> the firmware. That's it, no need for DT.
>>>> If you go through the driver, you will know that what it does, is
>>> We talk about bindings and I should not be forced to look at the
>>> driver to be able to understand them. Bindings should stand on their
>>> own.
>> Why can't ramoops binding have one more feature where it can add a flag
>> *dynamic* to indicate the regions are dynamic and it is for platforms
>> where there is another entity 'minidump' who is interested in these
>> regions.
> Because we do not define dynamic stuff in Devicetree. Dynamic means
> defined by SW or runtime configurable. It is against the entire idea of
> Devicetree which is for non-discoverable hardware.
>
>>>> just create platform device for actual ramoops driver to probe and to
>>> Not really justification for Devicetree anyway. Whatever your driver
>>> is doing, is driver's business, not bindings.
>>>
>>>> provide this it needs exact set of parameters of input what original
>>>> ramoops DT provides, we need to keep it in DT as maintaining this in
>>>> driver will not scale well with different size/parameter size
>>>> requirement for different targets.
>>> Really? Why? I don't see a problem in scaling. At all.
>> I had attempted it here,
>>
>> https://lore.kernel.org/lkml/1683133352-10046-10-git-send-email-quic_mojha@quicinc.com/
>>
>> but got comments related to hard coding and some in favor of having
>> the same set of properties what ramoops has/provides
>>
>> https://lore.kernel.org/lkml/e25723bf-be85-b458-a84c-1a45392683bb@gmail.com/
>>
>> https://lore.kernel.org/lkml/202305161347.80204C1A0E@keescook/
> Then you were tricked. I don't get why someone else suggests that
> non-hardware property should be part of Devicetree, but anyway it's the
> call of Devicetree binding maintainers, not someone else. DT is not
> dumping ground for all the system configuration variables.

Sorry for that, I assumed the interface should be as close as possible 
to pstore, but apparently that's not the case. Why does it have to be 
different from it? It provides the same functionality and is 
configurable even if it doesn't explicitly configure non discoverable 
hardware properties.

Assuming we make the driver picks the values, how would it do that?  
Hardcoding a configuration could lead to a few problems, such as the 
allocated region being smaller than the driver defaults or driver 
defaults not fully utilizing the allocated region, possibly wasting more 
memory than it'll ever use. On top of that what happens if we want to 
configure it differently than the hardcoded default values? Via cmdline 
options? For example in the previous version it allocated the whole 
region for the console alone, while other entries, such as pmsg that 
could be useful on devices using minidump to store Android logs, was 
zero-sized.

>
>>>>>> A part of this registration code you can find in 11/21
>>>>>>
>>>>>>> I'm pretty sure I already said all this before.
>>>>>> Yes, you said this before but that's the reason i came up with vendor
>>>>>> ramoops instead of changing traditional ramoops binding.
>>>>> That's unexpected conclusion. Adding more bindings is not the answer to
>>>>> comment that it should not be in the DTS in the first place.
>>>> Please suggest, what is the other way being above text as requirement..
>>> I do not see any requirement for us there. Forcing me to figure out
>>> how to add non-hardware property to DT is not the way to convince
>>> reviewers. But if you insist - we have ABI for this, called sysfs. If
>>> it is debugging feature, then debugfs.
>> ramoops already support module params and a way to pass these parameters
>> from bootargs but it also need to know the hard-codes addresses, so,
>> doing something in sysfs will be again duplication with ramoops driver..
> Why do you need hard-coded addresses?
>
>> If this can be accommodated under ramoops, this will be very small
>> change, like this
>>
>> https://lore.kernel.org/lkml/20230622005213.458236-1-isaacmanjarres@google.com/
> That's also funny patch - missing bindings updated, missing CC DT
> maintainers.
>
> Best regards,
> Krzysztof
>
>
>
diff mbox series

Patch

diff --git a/Documentation/devicetree/bindings/soc/qcom/qcom,ramoops.yaml b/Documentation/devicetree/bindings/soc/qcom/qcom,ramoops.yaml
new file mode 100644
index 000000000000..b1fdcf3f8ad4
--- /dev/null
+++ b/Documentation/devicetree/bindings/soc/qcom/qcom,ramoops.yaml
@@ -0,0 +1,126 @@ 
+# SPDX-License-Identifier: (GPL-2.0 OR BSD-2-Clause)
+%YAML 1.2
+---
+$id: "http://devicetree.org/schemas/soc/qcom/qcom,ramoops.yaml#"
+$schema: "http://devicetree.org/meta-schemas/core.yaml#"
+
+title: Qualcomm Ramoops minidump logger
+
+description: |
+  Qualcomm ramoops minidump logger provide a means of storing the ramoops
+  data to some dynamically reserved memory instead of traditionally
+  implemented ramoops where the region should be statically fixed ram
+  region. Because of its similarity with ramoops it will also have same
+  set of property what ramoops have it in its schema and is going to
+  contain traditional ramoops frontend data and this region will be
+  collected via Qualcomm minidump infrastructure provided from the
+  boot firmware.
+
+maintainers:
+  - Mukesh Ojha <quic_mojha@quicinc.com>
+
+properties:
+  compatible:
+    items:
+      - enum:
+          - qcom,sm8450-ramoops
+      - const: qcom,ramoops
+
+  memory-region:
+    maxItems: 1
+    description: handle to memory reservation for qcom,ramoops region.
+
+  ecc-size:
+    $ref: /schemas/types.yaml#/definitions/uint32
+    description: enables ECC support and specifies ECC buffer size in bytes
+    default: 0 # no ECC
+
+  record-size:
+    $ref: /schemas/types.yaml#/definitions/uint32
+    description: maximum size in bytes of each kmsg dump
+    default: 0
+
+  console-size:
+    $ref: /schemas/types.yaml#/definitions/uint32
+    description: size in bytes of log buffer reserved for kernel messages
+    default: 0
+
+  ftrace-size:
+    $ref: /schemas/types.yaml#/definitions/uint32
+    description: size in bytes of log buffer reserved for function tracing and profiling
+    default: 0
+
+  pmsg-size:
+    $ref: /schemas/types.yaml#/definitions/uint32
+    description: size in bytes of log buffer reserved for userspace messages
+    default: 0
+
+  mem-type:
+    $ref: /schemas/types.yaml#/definitions/uint32
+    description: if present, sets the type of mapping is to be used to map the reserved region.
+    default: 0
+    oneOf:
+      - const: 0
+        description: write-combined
+      - const: 1
+        description: unbuffered
+      - const: 2
+        description: cached
+
+  max-reason:
+    $ref: /schemas/types.yaml#/definitions/uint32
+    default: 2 # log oopses and panics
+    maximum: 0x7fffffff
+    description: |
+      If present, sets maximum type of kmsg dump reasons to store.
+      This can be set to INT_MAX to store all kmsg dumps.
+      See include/linux/kmsg_dump.h KMSG_DUMP_* for other kmsg dump reason values.
+      Setting this to 0 (KMSG_DUMP_UNDEF), means the reason filtering will be
+      controlled by the printk.always_kmsg_dump boot param.
+      If unset, it will be 2 (KMSG_DUMP_OOPS), otherwise 5 (KMSG_DUMP_MAX).
+
+  flags:
+    $ref: /schemas/types.yaml#/definitions/uint32
+    default: 0
+    description: |
+      If present, pass ramoops behavioral flags
+      (see include/linux/pstore_ram.h RAMOOPS_FLAG_* for flag values).
+
+  no-dump-oops:
+    deprecated: true
+    type: boolean
+    description: |
+      Use max_reason instead. If present, and max_reason is not specified,
+      it is equivalent to max_reason = 1 (KMSG_DUMP_PANIC).
+
+  unbuffered:
+    deprecated: true
+    type: boolean
+    description: |
+      Use mem_type instead. If present, and mem_type is not specified,
+      it is equivalent to mem_type = 1 and uses unbuffered mappings to map
+      the reserved region (defaults to buffered mappings mem_type = 0).
+      If both are specified -- "mem_type" overrides "unbuffered".
+
+unevaluatedProperties: false
+
+required:
+  - compatible
+  - memory-region
+
+anyOf:
+  - required: [record-size]
+  - required: [console-size]
+  - required: [ftrace-size]
+  - required: [pmsg-size]
+
+examples:
+  - |
+
+    qcom_ramoops {
+        compatible = "qcom,sm8450-ramoops", "qcom,ramoops";
+        memory-region = <&qcom_ramoops_region>;
+        console-size = <0x8000>;    /* 32kB */
+        record-size = <0x400>;      /*  1kB */
+        ecc-size = <16>;
+    };