Hi Krzysztof,
On 5/9/2023 12:41 PM, Krzysztof Kozlowski wrote:
On 08/05/2023 09:10, Mukesh Ojha wrote:
On 5/4/2023 10:04 PM, Krzysztof Kozlowski wrote:
On 04/05/2023 17:21, Krzysztof Kozlowski wrote:
+ ret = qcom_minidump_init_apss_subsystem(md);
+ if (ret) {
+ dev_err(&pdev->dev, "apss minidump initialization failed: %d\n", ret);
+ goto unlock;
+ }
+
+ __md = md;
No. This is a platform device, so it can have multiple instances.
It can have only one instance that is created from SMEM driver probe.
Anyone can instantiate more of them.... how did you solve it?
To clarify - sprinkling more of singletons makes everything tightly
coupled, difficult to debug and non-portable. You cannot have two
instances, you have to control concurrent initialization by yourself in
each of such singletons.
I understand sometimes they are unavoidable, for example when this does
not map to hardware property. However here you have the parent - smem -
which can return you valid instance. Thus you avoid entire problem of
file-scope variables.
I get your point, why one's should avoid file scope variables.
This is infrastructure driver and will not have multiple instances and
even if it happens could be avoided with with the help of global mutex
and protect below function which i am already doing at the moment and
But we do not want global mutexes... so incorrect design is being
improved by more incorrect design.
fail the other probe if it is already initialized with proper logging..e.g
"already initialized..."
ret = qcom_minidump_init_apss_subsystem(md);
And this will be in-lined with
/* Pointer to the one and only smem handle */
static struct qcom_smem *__smem;
Let me know if you still disagree...and have some other way ?
Why the parent - smem - cannot return every consumer the instance it
has? There will be one smem having only one minidump, so all problems
solved?
Sorry, I am extending this discussion but it is needed to avoid rework
in upcoming patches.
I am inline with the thought of each smem has its own minidump instance,
which is basically one at this moment as SMEM has only instance in DT.
In that way, Client driver calling qcom_apss_minidump_region_register()
will also need to know the instance it need to register with right?
However, I do have a use case [1] where SMEM or similar region
supporting memory mapped region could be virtualized and guest vm does
not have direct access to it, that way it will only have one backend at
a time.But even if they exist together that can be done with below approach.
File scope variable is still needed in minidump core but can be avoided
in backend drivers where each backend register with core and get added
itself in the list and for list protection, list mutex would be needed.
#define SMEM 0;
#define MMIO 1;
or enum may be..
And client can call this to the instance it need to register with..
int qcom_apss_minidump_region_register(region, SMEM);
int qcom_apss_minidump_region_register(region, MMIO);
Do you agree with this approach?
[1]
+----------------+
| |
| client A-Z |
+-----+----------+
|
|
|
|
v
+------------------------+
| | other backends
| minidump core +----------------------------+
| | |
+--+---------------------+ |
| | |
| | |
| | |e.g,
| | |gunyah-rm
+--------v------+ +-----v-----------+ +--+---------+
| | | | | |
|minidump_smem | | minidump_mmio | | ..... |
+---------------+ +-----------------+ +------------+
SMEM backend mmio backend where
smem may be virtualized
-- Mukesh
Best regards,
Krzysztof
