Skip to contentSkip to conversation

sched.h: always_inline alloc_tag_{save|restore} to fix modpost warnings

PATCH Superseded Suren Baghdasaryan 2 participants 10 messages Jul 3, 2024 – Jul 4, 2024 Scheduler

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sched.h: always_inline alloc_tag_{save|restore} to fix modpost warnings

Mark alloc_tag_{save|restore} as always_inline to fix the following modpost warnings:

WARNING: modpost: vmlinux: section mismatch in reference: alloc_tag_save+0x1c (section: .text.unlikely) -> initcall_level_names (section: .init.data)
WARNING: modpost: vmlinux: section mismatch in reference: alloc_tag_restore+0x3c (section: .text.unlikely) -> initcall_level_names (section: .init.data)

---
 include/linux/sched.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
include/linux/sched.h +2 -2
index 61591ac6eab6..a5f4b48fca18 100644
@@ -2192,13 +2192,13 @@ static inline int sched_core_idle_cpu(int cpu) { return idle_cpu(cpu); }
 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
 
 #ifdef CONFIG_MEM_ALLOC_PROFILING
-static inline struct alloc_tag *alloc_tag_save(struct alloc_tag *tag)
+static __always_inline struct alloc_tag *alloc_tag_save(struct alloc_tag *tag)
 {
 	swap(current->alloc_tag, tag);
 	return tag;
 }
 
-static inline void alloc_tag_restore(struct alloc_tag *tag, struct alloc_tag *old)
+static __always_inline void alloc_tag_restore(struct alloc_tag *tag, struct alloc_tag *old)
 {
 #ifdef CONFIG_MEM_ALLOC_PROFILING_DEBUG
 	WARN(current->alloc_tag != tag, "current->alloc_tag was changed:\n");

base-commit: 795c58e4c7fc6163d8fb9f2baa86cfe898fa4b19

Signature · 1 line
2.45.2.803.g4e1b14247a-goog
#1 lore raw
Quoted 5 lines · On Wed, 3 Jul 2024 15:15:20 -0700 Suren Baghdasaryan <[email protected]> wrote:

Mark alloc_tag_{save|restore} as always_inline to fix the following modpost warnings:

WARNING: modpost: vmlinux: section mismatch in reference: alloc_tag_save+0x1c (section: .text.unlikely) -> initcall_level_names (section: .init.data)
WARNING: modpost: vmlinux: section mismatch in reference: alloc_tag_restore+0x3c (section: .text.unlikely) -> initcall_level_names (section: .init.data)

Well, is it only about fixing warnings? If the warning is correct then this might be fixing kernel crashes.

Do you know where these references are coming from?

I'm curious about the .text.unlikely. Makes me wonder if we should also have .init.unlikely for unlikely() calls which happen from __init code. Maybe we already handle that.

Quoted 4 lines
Fixes 22d407b164ff ("lib: add allocation tagging support for memory allocation profiling")
Cc stable

yes?

#2 lore raw
Quoted 6 lines · Andrew Morton
Quoted 3 lines · On Wed, 3 Jul 2024 15:15:20 -0700 Suren Baghdasaryan <[email protected]> wrote:

Mark alloc_tag_{save|restore} as always_inline to fix the following modpost warnings:

Quoted 7 lines
Quoted 2 lines

WARNING: modpost: vmlinux: section mismatch in reference: alloc_tag_save+0x1c (section: .text.unlikely) -> initcall_level_names (section: .init.data)
WARNING: modpost: vmlinux: section mismatch in reference: alloc_tag_restore+0x3c (section: .text.unlikely) -> initcall_level_names (section: .init.data)

Well, is it only about fixing warnings? If the warning is correct then this might be fixing kernel crashes.

Do you know where these references are coming from?

I *think* this happens when alloc_tag_save()/alloc_tag_restore() are not inlined and are called from an __init function. They access the tag parameter passed to them and since that tag is a static local variable inside an __init function, I assume it gets allocated inside __initdata. If so, an example of such case is cma_activate_area() which is an __init function and allocates memory using bitmap_zalloc():
https://elixir.bootlin.com/linux/v6.10-rc6/source/mm/cma.c#L97. There are likely more cases like that.

Quoted 4 lines

I'm curious about the .text.unlikely. Makes me wonder if we should also have .init.unlikely for unlikely() calls which happen from __init code. Maybe we already handle that.

I don't really know.

Quoted 7 lines
Quoted 5 lines
Fixes 22d407b164ff ("lib: add allocation tagging support for memory allocation profiling")

Yes. Do you want me to post a v2 or will handle that locally?

Quoted 1 line
Cc stable

I don't think so. This feature was introduced in 6.10, so no backports needed, right?

Quoted 4 lines

yes?

#3 lore raw
Quoted 16 lines · Suren Baghdasaryan
Quoted 6 lines · Andrew Morton
Quoted 3 lines · On Wed, 3 Jul 2024 15:15:20 -0700 Suren Baghdasaryan <[email protected]> wrote:

Mark alloc_tag_{save|restore} as always_inline to fix the following modpost warnings:

Quoted 7 lines
Quoted 2 lines

WARNING: modpost: vmlinux: section mismatch in reference: alloc_tag_save+0x1c (section: .text.unlikely) -> initcall_level_names (section: .init.data)
WARNING: modpost: vmlinux: section mismatch in reference: alloc_tag_restore+0x3c (section: .text.unlikely) -> initcall_level_names (section: .init.data)

Well, is it only about fixing warnings? If the warning is correct then this might be fixing kernel crashes.

Do you know where these references are coming from?

Quoted 9 lines

I *think* this happens when alloc_tag_save()/alloc_tag_restore() are not inlined and are called from an __init function. They access the tag parameter passed to them and since that tag is a static local variable inside an __init function, I assume it gets allocated inside __initdata. If so, an example of such case is cma_activate_area() which is an __init function and allocates memory using bitmap_zalloc():
https://elixir.bootlin.com/linux/v6.10-rc6/source/mm/cma.c#L97. There are likely more cases like that.

Actually, my theory is wrong. Allocation tags are always allocated from the "alloc_tags" section. This makes me think that it's the access to current->alloc_tag that causes these warnings. After my change I actually see another warning like this:

WARNING: modpost: vmlinux: section mismatch in reference:
get_current+0xc (section: .text.unlikely) -> initcall_level_names (section: .init.data)

get_current()/current_thread_info() for xtensa arch are inline functions:
https://elixir.bootlin.com/linux/v6.10-rc6/source/arch/xtensa/include/asm/current.h#L22
https://elixir.bootlin.com/linux/v6.10-rc6/source/arch/xtensa/include/asm/thread_info.h#L94

and they return a local variable thread_info.
Let me dig a bit more to understand what's really happening here.

Quoted 27 lines
Quoted 5 lines

I'm curious about the .text.unlikely. Makes me wonder if we should also have .init.unlikely for unlikely() calls which happen from __init code. Maybe we already handle that.

I don't really know.

Quoted 7 lines
Quoted 5 lines
Fixes 22d407b164ff ("lib: add allocation tagging support for memory allocation profiling")

Yes. Do you want me to post a v2 or will handle that locally?

Quoted 1 line
Cc stable

I don't think so. This feature was introduced in 6.10, so no backports needed, right?

Quoted 4 lines

yes?

#4 lore raw
Quoted 37 lines · Suren Baghdasaryan
Quoted 16 lines · Suren Baghdasaryan
Quoted 6 lines · Andrew Morton
Quoted 3 lines · On Wed, 3 Jul 2024 15:15:20 -0700 Suren Baghdasaryan <[email protected]> wrote:

Mark alloc_tag_{save|restore} as always_inline to fix the following modpost warnings:

Quoted 7 lines
Quoted 2 lines

WARNING: modpost: vmlinux: section mismatch in reference: alloc_tag_save+0x1c (section: .text.unlikely) -> initcall_level_names (section: .init.data)
WARNING: modpost: vmlinux: section mismatch in reference: alloc_tag_restore+0x3c (section: .text.unlikely) -> initcall_level_names (section: .init.data)

Well, is it only about fixing warnings? If the warning is correct then this might be fixing kernel crashes.

Do you know where these references are coming from?

Quoted 9 lines

I *think* this happens when alloc_tag_save()/alloc_tag_restore() are not inlined and are called from an __init function. They access the tag parameter passed to them and since that tag is a static local variable inside an __init function, I assume it gets allocated inside __initdata. If so, an example of such case is cma_activate_area() which is an __init function and allocates memory using bitmap_zalloc():
https://elixir.bootlin.com/linux/v6.10-rc6/source/mm/cma.c#L97. There are likely more cases like that.

Actually, my theory is wrong. Allocation tags are always allocated from the "alloc_tags" section. This makes me think that it's the access to current->alloc_tag that causes these warnings. After my change I actually see another warning like this:

WARNING: modpost: vmlinux: section mismatch in reference:
get_current+0xc (section: .text.unlikely) -> initcall_level_names (section: .init.data)

Quoted 6 lines

get_current()/current_thread_info() for xtensa arch are inline functions:
https://elixir.bootlin.com/linux/v6.10-rc6/source/arch/xtensa/include/asm/current.h#L22
https://elixir.bootlin.com/linux/v6.10-rc6/source/arch/xtensa/include/asm/thread_info.h#L94

and they return a local variable thread_info.
Let me dig a bit more to understand what's really happening here.

Ok, I confirmed that the warning is happening due to the access to "current" from alloc_tag_save()/alloc_tag_restore() functions. I guess when these functions access "thread_info" variable:
https://elixir.bootlin.com/linux/v6.10-rc6/source/arch/xtensa/include/asm/thread_info.h#L96,
compiler flags that because the variable is on the stack of an __init function while alloc_tag_save()/alloc_tag_restore() when not inlined are from .text section.
To fix this warning completely I also need to change get_current() and current_thread_info() for xtensa to be __always_inline. I confirmed that after changing that the warnings caused by memory allocation profiling are gone. Note that there are more similar warnings when building this architecture but they happen even when CONFIG_MEM_ALLOC_PROFILING=n. I tried fixing them all but one fix leads to inner functions needing the same fix, so the patch grows exponentially, so I left it for now.

If there are no objections, I'll post a v2 as two separate patches. One changing the sched.h and the second one modifying xtensa-specific get_current() and current_thread_info(). Will post tomorrow unless someone replies with a better suggestion by then.
Thanks,
Suren.

Quoted 28 lines
Quoted 28 lines
Quoted 6 lines

I'm curious about the .text.unlikely. Makes me wonder if we should also have .init.unlikely for unlikely() calls which happen from __init code. Maybe we already handle that.

I don't really know.

Quoted 7 lines
Quoted 5 lines
Fixes 22d407b164ff ("lib: add allocation tagging support for memory allocation profiling")

Yes. Do you want me to post a v2 or will handle that locally?

Quoted 1 line
Cc stable

I don't think so. This feature was introduced in 6.10, so no backports needed, right?

Quoted 4 lines

yes?

#5 lore raw
Quoted 7 lines · On Wed, 3 Jul 2024 20:46:11 -0700 Suren Baghdasaryan <[email protected]> wrote:

Ok, I confirmed that the warning is happening due to the access to "current" from alloc_tag_save()/alloc_tag_restore() functions. I guess when these functions access "thread_info" variable:
https://elixir.bootlin.com/linux/v6.10-rc6/source/arch/xtensa/include/asm/thread_info.h#L96,
compiler flags that because the variable is on the stack of an __init function while alloc_tag_save()/alloc_tag_restore() when not inlined are from .text section.

Well, is the warning legitimate? I don't see why an automatic variable of an __init function should be considered to be .init storage - we can assume it won't become an invalid reference while the .init function is executing?

#6 lore raw
Quoted 11 lines · Andrew Morton
Quoted 7 lines · On Wed, 3 Jul 2024 20:46:11 -0700 Suren Baghdasaryan <[email protected]> wrote:

Ok, I confirmed that the warning is happening due to the access to "current" from alloc_tag_save()/alloc_tag_restore() functions. I guess when these functions access "thread_info" variable:
https://elixir.bootlin.com/linux/v6.10-rc6/source/arch/xtensa/include/asm/thread_info.h#L96,
compiler flags that because the variable is on the stack of an __init function while alloc_tag_save()/alloc_tag_restore() when not inlined are from .text section.

Quoted 4 lines

Well, is the warning legitimate? I don't see why an automatic variable of an __init function should be considered to be .init storage - we can assume it won't become an invalid reference while the .init function is executing?

I don't think it's really a problem. __init function is executing, it calls a function from .text (say alloc_tag_save() that was not inlined) which in turn calls get_current(), which returns a pointer somewhere inside __initdata. That should be fine since this can only happen during init stage. If this call happens after init, get_current() can't return a pointer from __initdata. If it does then we have a much bigger problem.

Quoted 1 line
#7 lore raw
Quoted 17 lines · On Wed, 3 Jul 2024 21:07:56 -0700 Suren Baghdasaryan <[email protected]> wrote:
Quoted 11 lines · Andrew Morton
Quoted 7 lines · On Wed, 3 Jul 2024 20:46:11 -0700 Suren Baghdasaryan <[email protected]> wrote:

Ok, I confirmed that the warning is happening due to the access to "current" from alloc_tag_save()/alloc_tag_restore() functions. I guess when these functions access "thread_info" variable:
https://elixir.bootlin.com/linux/v6.10-rc6/source/arch/xtensa/include/asm/thread_info.h#L96,
compiler flags that because the variable is on the stack of an __init function while alloc_tag_save()/alloc_tag_restore() when not inlined are from .text section.

Quoted 4 lines

Well, is the warning legitimate? I don't see why an automatic variable of an __init function should be considered to be .init storage - we can assume it won't become an invalid reference while the .init function is executing?

Quoted 7 lines

I don't think it's really a problem. __init function is executing, it calls a function from .text (say alloc_tag_save() that was not inlined) which in turn calls get_current(), which returns a pointer somewhere inside __initdata. That should be fine since this can only happen during init stage. If this call happens after init, get_current() can't return a pointer from __initdata. If it does then we have a much bigger problem.

So I think you're saying "yes, the warning is legitimate and this might be a problem, but it isn't in this case, so the checking code is OK so let's just work around it"?

#8 lore raw
Quoted 28 lines · Andrew Morton
Quoted 17 lines · On Wed, 3 Jul 2024 21:07:56 -0700 Suren Baghdasaryan <[email protected]> wrote:
Quoted 11 lines · Andrew Morton
Quoted 7 lines · On Wed, 3 Jul 2024 20:46:11 -0700 Suren Baghdasaryan <[email protected]> wrote:

Ok, I confirmed that the warning is happening due to the access to "current" from alloc_tag_save()/alloc_tag_restore() functions. I guess when these functions access "thread_info" variable:
https://elixir.bootlin.com/linux/v6.10-rc6/source/arch/xtensa/include/asm/thread_info.h#L96,
compiler flags that because the variable is on the stack of an __init function while alloc_tag_save()/alloc_tag_restore() when not inlined are from .text section.

Quoted 4 lines

Well, is the warning legitimate? I don't see why an automatic variable of an __init function should be considered to be .init storage - we can assume it won't become an invalid reference while the .init function is executing?

Quoted 7 lines

I don't think it's really a problem. __init function is executing, it calls a function from .text (say alloc_tag_save() that was not inlined) which in turn calls get_current(), which returns a pointer somewhere inside __initdata. That should be fine since this can only happen during init stage. If this call happens after init, get_current() can't return a pointer from __initdata. If it does then we have a much bigger problem.

Quoted 3 lines

So I think you're saying "yes, the warning is legitimate and this might be a problem, but it isn't in this case, so the checking code is OK so let's just work around it"?

Yes, that's my understanding. If get_current() always returns a valid pointer for the given execution stage (during init/after init) then it should be safe to call it even from a function residing in the .text section during init.

Quoted 1 line
#9 lore raw
Quoted 8 lines · On Wed, 3 Jul 2024 21:25:30 -0700 Suren Baghdasaryan <[email protected]> wrote:
Quoted 8 lines
Quoted 7 lines

I don't think it's really a problem. __init function is executing, it calls a function from .text (say alloc_tag_save() that was not inlined) which in turn calls get_current(), which returns a pointer somewhere inside __initdata. That should be fine since this can only happen during init stage. If this call happens after init, get_current() can't return a pointer from __initdata. If it does then we have a much bigger problem.

Quoted 8 lines
Quoted 3 lines

So I think you're saying "yes, the warning is legitimate and this might be a problem, but it isn't in this case, so the checking code is OK so let's just work around it"?

Yes, that's my understanding. If get_current() always returns a valid pointer for the given execution stage (during init/after init) then it should be safe to call it even from a function residing in the .text section during init.

Make sense, thanks for confirming. I'll await the v2 series, if you think it remains appropriate?