objtool warning in ice_free_prof_mask

objtool warning in ice_free_prof_mask

[Oleksandr Natalenko]

[-- Attachment #1: Type: text/plain, Size: 502 bytes --]

Hello.

With v6.15-rc1, CONFIG_OBJTOOL_WERROR=y and gcc 14.2.1 the following happens:

```
drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0() falls through to next function ice_free_flow_profs.cold()
drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0.cold() is missing an ELF size annotation
```

If I mark ice_write_prof_mask_reg() as noinline, this warning disappears.

Any idea what's going wrong?

Thank you.

-- 
Oleksandr Natalenko, MSE

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: objtool warning in ice_free_prof_mask

[Przemek Kitszel]

On 4/7/25 08:20, Oleksandr Natalenko wrote:
> Hello.
> 
> With v6.15-rc1, CONFIG_OBJTOOL_WERROR=y and gcc 14.2.1 the following happens:

have you COMPILE_TEST'ed whole kernel and this is the only (new) error?

> 
> ```
> drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0() falls through to next function ice_free_flow_profs.cold()
> drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0.cold() is missing an ELF size annotation
> ```
> 
> If I mark ice_write_prof_mask_reg() as noinline, this warning disappears.
> 
> Any idea what's going wrong?

sorry, no idea

> 
> Thank you.
> 

Re: objtool warning in ice_free_prof_mask

[Oleksandr Natalenko]

[-- Attachment #1: Type: text/plain, Size: 1183 bytes --]

Hello.

On pondělí 7. dubna 2025 11:03:31, středoevropský letní čas Przemek Kitszel wrote:
> On 4/7/25 08:20, Oleksandr Natalenko wrote:
> > Hello.
> > 
> > With v6.15-rc1, CONFIG_OBJTOOL_WERROR=y and gcc 14.2.1 the following happens:
> 
> have you COMPILE_TEST'ed whole kernel and this is the only (new) error?

It's not a new warning, I've observe it for several recent major kernel releases already.

I do not build with CONFIG_COMPILE_TEST.

I've also realised I see this warning with -O3 only. I know this is unsupported, so feel free to ignore me, but I do -O3 builds for finding out possible loose ends in the code, and this is the only place where it breaks.

> > ```
> > drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0() falls through to next function ice_free_flow_profs.cold()
> > drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0.cold() is missing an ELF size annotation
> > ```
> > 
> > If I mark ice_write_prof_mask_reg() as noinline, this warning disappears.
> > 
> > Any idea what's going wrong?
> 
> sorry, no idea
> 
> > 
> > Thank you.

-- 
Oleksandr Natalenko, MSE

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: objtool warning in ice_free_prof_mask

[Josh Poimboeuf]

On Mon, Apr 07, 2025 at 11:21:27AM +0200, Oleksandr Natalenko wrote:
> It's not a new warning, I've observe it for several recent major kernel releases already.
> 
> I do not build with CONFIG_COMPILE_TEST.
> 
> I've also realised I see this warning with -O3 only. I know this is
> unsupported, so feel free to ignore me, but I do -O3 builds for
> finding out possible loose ends in the code, and this is the only
> place where it breaks.
> 
> > > ```
> > > drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0() falls through to next function ice_free_flow_profs.cold()
> > > drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0.cold() is missing an ELF size annotation
> > > ```
> > > 
> > > If I mark ice_write_prof_mask_reg() as noinline, this warning disappears.
> > > 
> > > Any idea what's going wrong?

This type of error usually means some type of undefined behavior.  Can
you share your config?  No guarantees since it is -O3 after all, but I
can still take a look to see if it's pointing to a bug of some kind.

-- 
Josh

Re: objtool warning in ice_free_prof_mask

[Oleksandr Natalenko]

[-- Attachment #1: Type: text/plain, Size: 2918 bytes --]

Hello.

On pondělí 7. dubna 2025 23:42:51, středoevropský letní čas Josh Poimboeuf wrote:
> On Mon, Apr 07, 2025 at 11:21:27AM +0200, Oleksandr Natalenko wrote:
> > It's not a new warning, I've observe it for several recent major kernel releases already.
> > 
> > I do not build with CONFIG_COMPILE_TEST.
> > 
> > I've also realised I see this warning with -O3 only. I know this is
> > unsupported, so feel free to ignore me, but I do -O3 builds for
> > finding out possible loose ends in the code, and this is the only
> > place where it breaks.
> > 
> > > > ```
> > > > drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0() falls through to next function ice_free_flow_profs.cold()
> > > > drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0.cold() is missing an ELF size annotation
> > > > ```
> > > > 
> > > > If I mark ice_write_prof_mask_reg() as noinline, this warning disappears.
> > > > 
> > > > Any idea what's going wrong?
> 
> This type of error usually means some type of undefined behavior.  Can
> you share your config?  No guarantees since it is -O3 after all, but I
> can still take a look to see if it's pointing to a bug of some kind.

Sure, thank you for looking into this.

Here's my recipe to reproduce the issue:

1. expose -O3

```
diff --git a/Makefile b/Makefile
index 38689a0c36052..5ce5b44fa1496 100644
--- a/Makefile
+++ b/Makefile
@@ -858,6 +858,9 @@ KBUILD_CFLAGS	+= -fno-delete-null-pointer-checks
 ifdef CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE
 KBUILD_CFLAGS += -O2
 KBUILD_RUSTFLAGS += -Copt-level=2
+else ifdef CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE_O3
+KBUILD_CFLAGS += -O3
+KBUILD_RUSTFLAGS += -Copt-level=3
 else ifdef CONFIG_CC_OPTIMIZE_FOR_SIZE
 KBUILD_CFLAGS += -Os
 KBUILD_RUSTFLAGS += -Copt-level=s
diff --git a/init/Kconfig b/init/Kconfig
index dd2ea3b9a7992..03b1d768d1a55 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1481,6 +1481,12 @@ config CC_OPTIMIZE_FOR_PERFORMANCE
 	  with the "-O2" compiler flag for best performance and most
 	  helpful compile-time warnings.
 
+config CC_OPTIMIZE_FOR_PERFORMANCE_O3
+	bool "Optimize more for performance (-O3)"
+	help
+	  Choosing this option will pass "-O3" to your compiler to optimize
+	  the kernel yet more for performance.
+
 config CC_OPTIMIZE_FOR_SIZE
 	bool "Optimize for size (-Os)"
 	help
```

2. use this config with v6.15-rc1: https://paste.voidband.net/2BVaYDQS.txt

3. compile:

```
$ make drivers/net/ethernet/intel/ice/ice.o
…
  LD [M]  drivers/net/ethernet/intel/ice/ice.o
drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0() falls through to next function ice_free_flow_profs.cold()
drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0.cold() is missing an ELF size annotation
…
```

-- 
Oleksandr Natalenko, MSE

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: [Intel-wired-lan] objtool warning in ice_free_prof_mask

[Josh Poimboeuf]

On Mon, Apr 07, 2025 at 11:49:35PM +0200, Oleksandr Natalenko wrote:
> $ make drivers/net/ethernet/intel/ice/ice.o
> …
>   LD [M]  drivers/net/ethernet/intel/ice/ice.o
> drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0() falls through to next function ice_free_flow_profs.cold()
> drivers/net/ethernet/intel/ice/ice.o: error: objtool: ice_free_prof_mask.isra.0.cold() is missing an ELF size annotation

Thanks, I was able to recreate.

This is the -O3 optimizer noticing that ice_write_prof_mask_reg() is
only called with ICE_BLK_RSS or ICE_BLK_FD.  So it optimizes out the
impossible 'default' case in this switch statement:

	switch (blk) {
	case ICE_BLK_RSS:
		offset = GLQF_HMASK(mask_idx);
		val = FIELD_PREP(GLQF_HMASK_MSK_INDEX_M, idx);
		val |= FIELD_PREP(GLQF_HMASK_MASK_M, mask);
		break;
	case ICE_BLK_FD:
		offset = GLQF_FDMASK(mask_idx);
		val = FIELD_PREP(GLQF_FDMASK_MSK_INDEX_M, idx);
		val |= FIELD_PREP(GLQF_FDMASK_MASK_M, mask);
		break;
	default:
		ice_debug(hw, ICE_DBG_PKG, "No profile masks for block %d\n",
			  blk);
		return;
	}

Unfortunately, instead of finishing the optimization, it inserts
undefined behavior for the 'default' case by branching off to some
random code.

So there doesn't seem to be any underlying bug, it's just that objtool
doesn't like undefined behavior.

So for building with -O3 I'd recommend just disabling
CONFIG_OBJTOOL_WERROR and ignoring any objtool warnings.

-- 
Josh