powerpc/tlb: Implement book3s/32/tlbflush.h local_flush_tlb_page_psize
PATCH 1/1 powerpc/tlb: Implement book3s/32/tlbflush.h local_flush_tlb_page_psize
The commit introducing this function implemented it as a build bug on this platform to make the compiler happy, as the only use in the code is guarded behind a radix_enabled() conditional.
GCC recognises that cpu_has_feature(MMU_FTR_TYPE_RADIX) returns false on this platform and eliminates the build bug as dead code. However, when CONFIG_JUMP_LABEL_FEATURE_CHECK_DEBUG is enabled, the assertion and possible call to early_cpu_... prevents Clang from eliminating any code that's conditional on the return value. So Clang triggers the build bug as reported by the kernel test robot:
https://lore.kernel.org/llvm/[email protected]
---
Supersedes https://patchwork.ozlabs.org/project/linuxppc-dev/patch/[email protected]/
--- arch/powerpc/include/asm/book3s/32/tlbflush.h | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-)
arch/powerpc/include/asm/book3s/32/tlbflush.h +1 -3
@@ -2,8 +2,6 @@#ifndef _ASM_POWERPC_BOOK3S_32_TLBFLUSH_H#define _ASM_POWERPC_BOOK3S_32_TLBFLUSH_H -#include <linux/build_bug.h> -#define MMU_NO_CONTEXT (0)/* * TLB flushing for "classic" hash-MMU 32-bit CPUs, 6xx, 7xx, 7xxx
@@ -80,7 +78,7 @@ static inline void local_flush_tlb_page(struct vm_area_struct *vma,static inline void local_flush_tlb_page_psize (struct mm_struct *mm,unsigned long vmaddr,int psize) { -BUILD_BUG (); +flush_range (mm, vmaddr, vmaddr +PAGE_SIZE ); }static inline void local_flush_tlb_mm (struct mm_struct *mm)
base-commit: ca272751ba18ca8f137af631cbc9f3f987fab6e3 --
2.39.1
Le 31/01/2023 à 03:58, Benjamin Gray a écrit :
Quoted 10 lines
The commit introducing this function implemented it as a build bug on this platform to make the compiler happy, as the only use in the code is guarded behind a
radix_enabled()conditional.GCC recognises that
cpu_has_feature(MMU_FTR_TYPE_RADIX)returns false on this platform and eliminates the build bug as dead code. However, whenCONFIG_JUMP_LABEL_FEATURE_CHECK_DEBUGis enabled, the assertion and possible call toearly_cpu_... prevents Clang from eliminating any code that's conditional on the return value. So Clang triggers the build bug as reported by the kernel test robot:
I still think it is not the correct fix. You are putting the problem under the carpet instead of fixing it. There are many other places where radix_enabled() or other mmu_has_feature() are used with the expectation that one leg will be eliminated at build time.
As written in previous thread, have you considered reworking mmu_has_feature() to move the CONFIG_JUMP_LABEL_FEATURE_CHECK_DEBUG after the below block:
if (MMU_FTRS_ALWAYS & feature) return true; if (!(MMU_FTRS_POSSIBLE & feature)) return false;
And as this looks like a Clang bug or limitation, can you provide us with a link to the Clang issue you have opened for it ?
Looking into it in more details, I'm even more puzzled. As far as I can see, local_flush_tlb_page_psize() is used only at one place, that is function __do_patch_instruction_mm(). So if Clang fails to identify it as a dead leg, it is the full __do_patch_instruction_mm() which is kept for no reason.
On the other hand, I see that local_flush_tlb_page_psize() implemented for nohash/32, so yes we can also implement it for book3s/32. But then the commit log should explain things differently.
By the way, I also see that local_flush_tlb_page_psize() for book3s/64 does just nothing at all for non Radix. Is that correct ?
Thanks
Christophe
Quoted 40 lines
https://lore.kernel.org/llvm/[email protected]
Fixes 274d842fa1ef ("powerpc/tlb: Add local flush for page given mm_struct and psize")Reported-by kernel test robot <[email protected]>Signed-off-by Benjamin Gray <[email protected]>---
Supersedes https://patchwork.ozlabs.org/project/linuxppc-dev/patch/[email protected]/
--- arch/powerpc/include/asm/book3s/32/tlbflush.h | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-)arch/powerpc/include/asm/book3s/32/tlbflush.h +1
-3@@ -2,8 +2,6 @@#ifndef _ASM_POWERPC_BOOK3S_32_TLBFLUSH_H#define _ASM_POWERPC_BOOK3S_32_TLBFLUSH_H -#include <linux/build_bug.h> -#define MMU_NO_CONTEXT (0)/* * TLB flushing for "classic" hash-MMU 32-bit CPUs, 6xx, 7xx, 7xxx @@ -80,7 +78,7 @@ static inline void local_flush_tlb_page(struct vm_area_struct *vma,static inline void local_flush_tlb_page_psize (struct mm_struct *mm,unsigned long vmaddr,int psize) { -BUILD_BUG (); +flush_range (mm, vmaddr, vmaddr +PAGE_SIZE ); }static inline void local_flush_tlb_mm (struct mm_struct *mm)base-commit:
ca272751ba18ca8f137af631cbc9f3f987fab6e3--
2.39.1
Quoted 6 lines · Christophe Leroy
I still think it is not the correct fix. You are putting the problem under the carpet instead of fixing it. There are many other places where
radix_enabled()or othermmu_has_feature()are used with the expectation
that one leg will be eliminated at build time.
And none of them are actively causing build failures AFAIK. I agree that there may be a pre-existing optimisation problem, but I'm not trying to address it in this patch. I'm just trying to fix the build I broke. As such I haven't opened an issue with Clang yet either.
Quoted 10 lines
As written in previous thread, have you considered reworking
mmu_has_feature()to move theCONFIG_JUMP_LABEL_FEATURE_CHECK_DEBUGafter the below block:if (
MMU_FTRS_ALWAYS& feature)
return true;if (!(MMU_FTRS_POSSIBLE & feature))
return false;
Yes, I did. I also discussed with Michael Ellerman what he would prefer, and he indicated he still would still like to just implement the function.
Quoted 8 lines
Looking into it in more details, I'm even more puzzled. As far as I can
see,local_flush_tlb_page_psize()is used only at one place, that is function__do_patch_instruction_mm(). So if Clang fails to identify it
as a dead leg, it is the full__do_patch_instruction_mm()which is kept
for no reason.
Right, because that is the function that's guarded behind radix_enabled().
Quoted 3 lines
By the way, I also see that
local_flush_tlb_page_psize()for book3s/64
does just nothing at all for non Radix. Is that correct ?
That is how the other local page flushes are implemented. If there's some undocumented exception here I'd be relying on review on the list from people who have experience with details of how Hash mmu is handled on this platform.
Le 31/01/2023 à 22:58, Benjamin Gray a écrit :
Quoted 20 lines · Christophe Leroy
Quoted 6 lines
I still think it is not the correct fix. You are putting the problem under the carpet instead of fixing it. There are many other places where
radix_enabled()or othermmu_has_feature()are used with the expectation
that one leg will be eliminated at build time.And none of them are actively causing build failures AFAIK. I agree that there may be a pre-existing optimisation problem, but I'm not trying to address it in this patch. I'm just trying to fix the build I broke. As such I haven't opened an issue with Clang yet either.
Quoted 7 lines
As written in previous thread, have you considered reworking
mmu_has_feature()to move theCONFIG_JUMP_LABEL_FEATURE_CHECK_DEBUGafter the below block:if (
MMU_FTRS_ALWAYS& feature)
return true;
Quoted 7 lines
Quoted 3 lines
if (!(MMU_FTRS_POSSIBLE & feature))
return false;Yes, I did. I also discussed with Michael Ellerman what he would prefer, and he indicated he still would still like to just implement the function.
I'm fine with that. But if it is the intention, don't focus on the bug it fixes, but more on what it implements and how. That's what I would like to read in the commit message. It is nice to explain that there is a problem with CLANG and what the problem is, but only as side benefit of the commit.
For instance, explain why you can use flush_range() to implement it, and why you use PAGE_SIZE and not psize, etc ...
Christophe