Message-ID: <40e802eb-3764-47af-8b4f-9f7c8b5b60c1@linaro.org>
Date: Mon, 1 Sep 2025 11:57:06 +0300
From: Eugen Hristev <eugen.hristev@...aro.org>
To: David Hildenbrand <david@...hat.com>, Michal Hocko <mhocko@...e.com>
Cc: linux-kernel@...r.kernel.org, linux-arm-msm@...r.kernel.org,
 linux-arch@...r.kernel.org, linux-mm@...ck.org, tglx@...utronix.de,
 andersson@...nel.org, pmladek@...e.com,
 linux-arm-kernel@...ts.infradead.org, linux-hardening@...r.kernel.org,
 corbet@....net, mojha@....qualcomm.com, rostedt@...dmis.org,
 jonechou@...gle.com, tudor.ambarus@...aro.org,
 Christoph Hellwig <hch@...radead.org>,
 Sergey Senozhatsky <senozhatsky@...omium.org>
Subject: Re: [RFC][PATCH v2 22/29] mm/numa: Register information into Kmemdump



On 8/27/25 23:06, David Hildenbrand wrote:
> On 27.08.25 16:08, Eugen Hristev wrote:
>>
>>
>> On 8/27/25 15:18, David Hildenbrand wrote:
>>> On 27.08.25 13:59, Eugen Hristev wrote:
>>>>
>>>>
>>>> On 8/25/25 16:58, David Hildenbrand wrote:
>>>>> On 25.08.25 15:36, Eugen Hristev wrote:
>>>>>>
>>>>>>
>>>>>> On 8/25/25 16:20, David Hildenbrand wrote:
>>>>>>>
>>>>>>>>>
>>>>>>>>> IIRC, kernel/vmcore_info.c is never built as a module, as it also
>>>>>>>>> accesses non-exported symbols.
>>>>>>>>
>>>>>>>> Hello David,
>>>>>>>>
>>>>>>>> I am looking again into this, and there are some things which in my
>>>>>>>> opinion would be difficult to achieve.
>>>>>>>> For example I looked into my patch #11 , which adds the `runqueues` into
>>>>>>>> kmemdump.
>>>>>>>>
>>>>>>>> The runqueues is a variable of `struct rq` which is defined in
>>>>>>>> kernel/sched/sched.h , which is not supposed to be included outside of
>>>>>>>> sched.
>>>>>>>> Now moving all the struct definition outside of sched.h into another
>>>>>>>> public header would be rather painful and I don't think it's a really
>>>>>>>> good option (The struct would be needed to compute the sizeof inside
>>>>>>>> vmcoreinfo). Secondly, it would also imply moving all the nested struct
>>>>>>>> definitions outside as well. I doubt this is something that we want for
>>>>>>>> the sched subsys. How the subsys is designed, out of my understanding,
>>>>>>>> is to keep these internal structs opaque outside of it.
>>>>>>>
>>>>>>> All the kmemdump module needs is a start and a length, correct? So the
>>>>>>> only tricky part is getting the length.
>>>>>>
>>>>>> I also have in mind the kernel user case. How would a kernel programmer
>>>>>> want to add some kernel structs/info/buffers into kmemdump such that the
>>>>>> dump would contain their data ? Having "KMEMDUMP_VAR(...)" looks simple
>>>>>> enough.
>>>>>
>>>>> The other way around, why should anybody have a saying in adding their
>>>>> data to kmemdump? Why do we have that all over the kernel?
>>>>>
>>>>> Is your mechanism really so special?
>>>>>
>>>>> A single composer should take care of that, and it's really just start +
>>>>> len of physical memory areas.
>>>>>
>>>>>> Otherwise maybe the programmer has to write helpers to compute lengths
>>>>>> etc, and stitch them into kmemdump core.
>>>>>> I am not saying it's impossible, but just tiresome perhaps.
>>>>>
>>>>> In your patch set, how many of these instances did you encounter where
>>>>> that was a problem?
>>>>>
>>>>>>>
>>>>>>> One could just add a const variable that holds this information, or even
>>>>>>> better, a simple helper function to calculate that.
>>>>>>>
>>>>>>> Maybe someone else reading along has a better idea.
>>>>>>
>>>>>> This could work, but it requires again adding some code into the
>>>>>> specific subsystem. E.g. struct_rq_get_size()
>>>>>> I am open to ideas , and thank you very much for your thoughts.
>>>>>>
>>>>>>>
>>>>>>> Interestingly, runqueues is a percpu variable, which makes me wonder if
>>>>>>> what you had would work as intended (maybe it does, not sure).
>>>>>>
>>>>>> I would not really need to dump the runqueues. But the crash tool which
>>>>>> I am using for testing, requires it. Without the runqueues it will not
>>>>>> progress further to load the kernel dump.
>>>>>> So I am not really sure what it does with the runqueues, but it works.
>>>>>> Perhaps using crash/gdb more, to actually do something with this data,
>>>>>> would give more insight about its utility.
>>>>>> For me, it is a prerequisite to run crash, and then to be able to
>>>>>> extract the log buffer from the dump.
>>>>>
>>>>> I have the faint recollection that percpu vars might not be stored in a
>>>>> single contiguous physical memory area, but maybe my memory is just
>>>>> wrong, that's why I was raising it.
>>>>>
>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>>     From my perspective it's much simpler and cleaner to just add the
>>>>>>>> kmemdump annotation macro inside the sched/core.c as it's done in my
>>>>>>>> patch. This macro translates to a noop if kmemdump is not selected.
>>>>>>>
>>>>>>> I really don't like how we are spreading kmemdump all over the kernel,
>>>>>>> and adding complexity with __section when really, all we need is a place
>>>>>>> to obtain a start and a length.
>>>>>>>
>>>>>>
>>>>>> I understand. The section idea was suggested by Thomas. Initially I was
>>>>>> skeptic, but I like how it turned out.
>>>>>
>>>>> Yeah, I don't like it. Taste differs ;)
>>>>>
>>>>> I am in particular unhappy about custom memblock wrappers.
>>>>>
>>>>> [...]
>>>>>
>>>>>>>>
>>>>>>>> To have this working outside of printk, it would be required to walk
>>>>>>>> through all the printk structs/allocations and select the required info.
>>>>>>>> Is this something that we want to do outside of printk ?
>>>>>>>
>>>>>>> I don't follow, please elaborate.
>>>>>>>
>>>>>>> How is e.g., log_buf_len_get() + log_buf_addr_get() not sufficient,
>>>>>>> given that you run your initialization after setup_log_buf() ?
>>>>>>>
>>>>>>>
>>>>>>
>>>>>> My initial thought was the same. However I got some feedback from Petr
>>>>>> Mladek here :
>>>>>>
>>>>>> https://lore.kernel.org/lkml/aBm5QH2p6p9Wxe_M@localhost.localdomain/
>>>>>>
>>>>>> Where he explained how to register the structs correctly.
>>>>>> It can be that setup_log_buf is called again at a later time perhaps.
>>>>>>
>>>>>
>>>>> setup_log_buf() is a __init function, so there is only a certain time
>>>>> frame where it can be called.
>>>>>
>>>>> In particular, once the buddy is up, memblock allocations are impossible
>>>>> and it would be deeply flawed to call this function again.
>>>>>
>>>>> Let's not over-engineer this.
>>>>>
>>>>> Peter is on CC, so hopefully he can share his thoughts.
>>>>>
>>>>
>>>> Hello David,
>>>>
>>>> I tested out this snippet (on top of my series, so you can see what I
>>>> changed):
>>>>
>>>>
>>>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>>>> index 18ba6c1e174f..7ac4248a00e5 100644
>>>> --- a/kernel/sched/core.c
>>>> +++ b/kernel/sched/core.c
>>>> @@ -67,7 +67,6 @@
>>>>    #include <linux/wait_api.h>
>>>>    #include <linux/workqueue_api.h>
>>>>    #include <linux/livepatch_sched.h>
>>>> -#include <linux/kmemdump.h>
>>>>
>>>>    #ifdef CONFIG_PREEMPT_DYNAMIC
>>>>    # ifdef CONFIG_GENERIC_IRQ_ENTRY
>>>> @@ -120,7 +119,12 @@
>>>> EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
>>>>    EXPORT_TRACEPOINT_SYMBOL_GPL(sched_compute_energy_tp);
>>>>
>>>>    DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
>>>> -KMEMDUMP_VAR_CORE(runqueues, sizeof(runqueues));
>>>> +
>>>> +size_t runqueues_get_size(void);
>>>> +size_t runqueues_get_size(void)
>>>> +{
>>>> +       return sizeof(runqueues);
>>>> +}
>>>>
>>>>    #ifdef CONFIG_SCHED_PROXY_EXEC
>>>>    DEFINE_STATIC_KEY_TRUE(__sched_proxy_exec);
>>>> diff --git a/kernel/vmcore_info.c b/kernel/vmcore_info.c
>>>> index d808c5e67f35..c6dd2d6e96dd 100644
>>>> --- a/kernel/vmcore_info.c
>>>> +++ b/kernel/vmcore_info.c
>>>> @@ -24,6 +24,12 @@
>>>>    #include "kallsyms_internal.h"
>>>>    #include "kexec_internal.h"
>>>>
>>>> +typedef void* kmemdump_opaque_t;
>>>> +
>>>> +size_t runqueues_get_size(void);
>>>> +
>>>> +extern kmemdump_opaque_t runqueues;
>>>
>>> I would have tried that through:
>>>
>>> struct rq;
>>> extern struct rq runqueues;
>>>
>>> But the whole PER_CPU_SHARED_ALIGNED makes this all weird, and likely
>>> not the way we would want to handle that.
>>>
>>>>    /* vmcoreinfo stuff */
>>>>    unsigned char *vmcoreinfo_data;
>>>>    size_t vmcoreinfo_size;
>>>> @@ -230,6 +236,9 @@ static int __init crash_save_vmcoreinfo_init(void)
>>>>
>>>>           kmemdump_register_id(KMEMDUMP_ID_COREIMAGE_VMCOREINFO,
>>>>                                (void *)vmcoreinfo_data, vmcoreinfo_size);
>>>> +       kmemdump_register_id(KMEMDUMP_ID_COREIMAGE_runqueues,
>>>> +                            (void *)&runqueues, runqueues_get_size());
>>>> +
>>>>           return 0;
>>>>    }
>>>>
>>>> With this, no more .section, no kmemdump code into sched, however, there
>>>> are few things :
>>>
>>> I would really just do here something like the following:
>>>
>>> /**
>>>    * sched_get_runqueues_area - obtain the runqueues area for dumping
>>>    * @start: ...
>>>    * @size: ...
>>>    *
>>>    * The obtained area is only to be used for dumping purposes.
>>>    */
>>> void sched_get_runqueues_area(void *start, size_t size)
>>> {
>>> 	start = &runqueues;
>>> 	size = sizeof(runqueues);
>>> }
>>>
>>> might be cleaner.
>>>
>>
>> How about this in the header:
>>
>> #define DECLARE_DUMP_AREA_FUNC(subsys, symbol) \
>>
>> void subsys ## _get_ ## symbol ##_area(void **start, size_t *size);
>>
>>
>>
>> #define DEFINE_DUMP_AREA_FUNC(subsys, symbol) \
>>
>> void subsys ## _get_ ## symbol ##_area(void **start, size_t *size)\
>>
>> {\
>>
>>          *start = &symbol;\
>>
>>          *size = sizeof(symbol);\
>>
>> }
>>
>>
>> then, in sched just
>>
>> DECLARE_DUMP_AREA_FUNC(sched, runqueues);
>>
>> DEFINE_DUMP_AREA_FUNC(sched, runqueues);
>>
>> or a single macro that wraps both.
>>
>> would make it shorter and neater.
>>
>> What do you think ?
> 
> Looks a bit over-engineered, and will require us to import a header 
> (likely kmemdump.h) in these files, which I don't really enjoy.
> 
> I would start simple, without any such macro-magic. It's a very simple 
> function after all, and likely you won't end up having many of these?
> 

Thanks David, I will do it as you suggested and see what comes out of it.

I have one side question you might know much better to answer:
As we have a start and a size for each region, this start is a virtual
address. The firmware/coprocessor that reads the memory and dumps it,
requires physical addresses. What do you suggest to use to retrieve that
address ? virt_to_phys might be problematic, __pa or __pa_symbol? or
better lm_alias ?
As kmemdump is agnostic of the region of the memory the `start` comes
from, and it should be portable and platform independent.

Thanks again,
Eugen

Powered by blists - more mailing lists