powerpc: mm: radix_tlb: rearrange the if-else block

Clang warns:

arch/powerpc/mm/book3s64/radix_tlb.c:1191:23: error: variable 'hstart' is uninitialized when used here [-Werror,-Wuninitialized]
                                __tlbiel_va_range(hstart, hend, pid,
                                                  ^~~~~~
arch/powerpc/mm/book3s64/radix_tlb.c:1175:23: note: initialize the variable 'hstart' to silence this warning
                unsigned long hstart, hend;
                                    ^
                                     = 0
arch/powerpc/mm/book3s64/radix_tlb.c:1191:31: error: variable 'hend' is uninitialized when used here [-Werror,-Wuninitialized]
                                __tlbiel_va_range(hstart, hend, pid,
                                                          ^~~~
arch/powerpc/mm/book3s64/radix_tlb.c:1175:29: note: initialize the variable 'hend' to silence this warning
                unsigned long hstart, hend;
                                          ^
                                           = 0
2 errors generated.

Rework the if-else to pull the 'IS_ENABLE(CONFIG_TRANSPARENT_HUGEPAGE)'
check one level up, this will silent the warnings. That will also
simplify the 'else' path. Clang is getting confused with these warnings,
but the warnings is a false-positive.

Suggested-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Anders Roxell <redacted>
---
 arch/powerpc/mm/book3s64/radix_tlb.c | 31 +++++++++++++++++++++-------
 1 file changed, 24 insertions(+), 7 deletions(-)

diff --git a/arch/powerpc/mm/book3s64/radix_tlb.c b/arch/powerpc/mm/book3s64/radix_tlb.c
index 7724af19ed7e..e494a45ce1b4 100644
--- a/arch/powerpc/mm/book3s64/radix_tlb.c
+++ b/arch/powerpc/mm/book3s64/radix_tlb.c

@@ -1170,16 +1170,14 @@ static inline void __radix__flush_tlb_range(struct mm_struct *mm,
 				_tlbiel_pid_multicast(mm, pid, RIC_FLUSH_ALL);
 			}
 		}
-	} else {
+	} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
 		bool hflush = false;
 		unsigned long hstart, hend;
 
-		if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
-			hstart = (start + PMD_SIZE - 1) & PMD_MASK;
-			hend = end & PMD_MASK;
-			if (hstart < hend)
-				hflush = true;
-		}
+		hstart = (start + PMD_SIZE - 1) & PMD_MASK;
+		hend = end & PMD_MASK;
+		if (hstart < hend)
+			hflush = true;
 
 		if (type == FLUSH_TYPE_LOCAL) {
 			asm volatile("ptesync": : :"memory");

@@ -1207,6 +1205,25 @@ static inline void __radix__flush_tlb_range(struct mm_struct *mm,
 				_tlbiel_va_range_multicast(mm,
 					hstart, hend, pid, PMD_SIZE, MMU_PAGE_2M, flush_pwc);
 		}
+	} else {
+
+		if (type == FLUSH_TYPE_LOCAL) {
+			asm volatile("ptesync" : : : "memory");
+			if (flush_pwc)
+				/* For PWC, only one flush is needed */
+				__tlbiel_pid(pid, 0, RIC_FLUSH_PWC);
+			__tlbiel_va_range(start, end, pid, page_size, mmu_virtual_psize);
+			ppc_after_tlbiel_barrier();
+		} else if (cputlb_use_tlbie()) {
+			asm volatile("ptesync" : : : "memory");
+			if (flush_pwc)
+				__tlbie_pid(pid, RIC_FLUSH_PWC);
+			__tlbie_va_range(start, end, pid, page_size, mmu_virtual_psize);
+			asm volatile("eieio; tlbsync; ptesync" : : : "memory");
+		} else {
+			_tlbiel_va_range_multicast(mm,
+					start, end, pid, page_size, mmu_virtual_psize, flush_pwc);
+		}
 	}
 out:
 	preempt_enable();

-- 
2.33.0

Le 25/11/2021 à 16:44, Anders Roxell a écrit :

Clang warns:

arch/powerpc/mm/book3s64/radix_tlb.c:1191:23: error: variable 'hstart' is uninitialized when used here [-Werror,-Wuninitialized]
                                 __tlbiel_va_range(hstart, hend, pid,
                                                   ^~~~~~
arch/powerpc/mm/book3s64/radix_tlb.c:1175:23: note: initialize the variable 'hstart' to silence this warning
                 unsigned long hstart, hend;
                                     ^
                                      = 0
arch/powerpc/mm/book3s64/radix_tlb.c:1191:31: error: variable 'hend' is uninitialized when used here [-Werror,-Wuninitialized]
                                 __tlbiel_va_range(hstart, hend, pid,
                                                           ^~~~
arch/powerpc/mm/book3s64/radix_tlb.c:1175:29: note: initialize the variable 'hend' to silence this warning
                 unsigned long hstart, hend;
                                           ^
                                            = 0
2 errors generated.

Rework the if-else to pull the 'IS_ENABLE(CONFIG_TRANSPARENT_HUGEPAGE)'
check one level up, this will silent the warnings. That will also
simplify the 'else' path. Clang is getting confused with these warnings,
but the warnings is a false-positive.

But you are duplicating a significant part of the code by doing that, and duplicated code generaly leads to bugs.

And we already have redundant stuff between FLUSH_TYPE_LOCAL leg and cputlb_use_tlbie() leg.

Can't you fix CLANG instead :) ?

Or just add an else to the IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) that sets hstart and hend to 0 ?

Or just put hstart and hend calculation outside the IS_ENABLED() ? After all GCC should drop the calculation when not used.

20 more quoted lines

On Fri, Nov 26, 2021 at 2:43 PM Christophe Leroy
[off-list ref] wrote:

Le 25/11/2021 à 16:44, Anders Roxell a écrit :
Can't you fix CLANG instead :) ?

Or just add an else to the IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) that
sets hstart and hend to 0 ?

That doesn't sound any less risky than duplicating the code, it can lead to
incorrect changes just as easily if a patch ends up actually flushing at the
wrong address, and the compiler fails to complain because of the bogus
initialization.

Or just put hstart and hend calculation outside the IS_ENABLED() ? After
all GCC should drop the calculation when not used.

I like this one. I'm still unsure how clang can get so confused about whether
the variables are initialized or not, usually it handles this much better than
gcc. My best guess is that one of the memory clobbers makes it conclude
that 'hflush' can be true when it gets written to by an inline asm.

        Arnd

On Fri, Nov 26, 2021 at 02:59:29PM +0100, Arnd Bergmann wrote:

On Fri, Nov 26, 2021 at 2:43 PM Christophe Leroy
[off-list ref] wrote:

quoted

Le 25/11/2021 à 16:44, Anders Roxell a écrit :
Can't you fix CLANG instead :) ?

Or just add an else to the IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) that
sets hstart and hend to 0 ?

That doesn't sound any less risky than duplicating the code, it can lead to
incorrect changes just as easily if a patch ends up actually flushing at the
wrong address, and the compiler fails to complain because of the bogus
initialization.

quoted

Or just put hstart and hend calculation outside the IS_ENABLED() ? After
all GCC should drop the calculation when not used.

I like this one. I'm still unsure how clang can get so confused about whether
the variables are initialized or not, usually it handles this much better than
gcc. My best guess is that one of the memory clobbers makes it conclude
that 'hflush' can be true when it gets written to by an inline asm.

As far as I am aware, clang's analysis does not evaluate variables when
generating a control flow graph and using that for static analysis:

https://godbolt.org/z/PdGxoq9j7

Based on the control flow graph, it knows that hstart and hend are
uninitialized because IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) gets
expanded to 0 by the preprocessor but it does not seem like it can piece
together that hflush's value of false is only changed to true under the
now 'if (0) {' branch, meaning that all the calls to __tlbiel_va_range()
never get evaluated. That may or may not be easy to fix in clang but we
run into issues like this so infrequently.

At any rate, the below diff works for me.

Cheers,
Nathan

diff --git a/arch/powerpc/mm/book3s64/radix_tlb.c b/arch/powerpc/mm/book3s64/radix_tlb.c
index 7724af19ed7e..156a631df976 100644
--- a/arch/powerpc/mm/book3s64/radix_tlb.c
+++ b/arch/powerpc/mm/book3s64/radix_tlb.c

@@ -1174,12 +1174,10 @@ static inline void __radix__flush_tlb_range(struct mm_struct *mm,
 		bool hflush = false;
 		unsigned long hstart, hend;
 
-		if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
-			hstart = (start + PMD_SIZE - 1) & PMD_MASK;
-			hend = end & PMD_MASK;
-			if (hstart < hend)
-				hflush = true;
-		}
+		hstart = (start + PMD_SIZE - 1) & PMD_MASK;
+		hend = end & PMD_MASK;
+		if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hstart < hend)
+			hflush = true;
 
 		if (type == FLUSH_TYPE_LOCAL) {
 			asm volatile("ptesync": : :"memory");

Le 26/11/2021 à 16:46, Nathan Chancellor a écrit :

On Fri, Nov 26, 2021 at 02:59:29PM +0100, Arnd Bergmann wrote:

quoted

On Fri, Nov 26, 2021 at 2:43 PM Christophe Leroy
[off-list ref] wrote:

quoted

Le 25/11/2021 à 16:44, Anders Roxell a écrit :
Can't you fix CLANG instead :) ?

Or just add an else to the IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) that
sets hstart and hend to 0 ?

That doesn't sound any less risky than duplicating the code, it can lead to
incorrect changes just as easily if a patch ends up actually flushing at the
wrong address, and the compiler fails to complain because of the bogus
initialization.

quoted

Or just put hstart and hend calculation outside the IS_ENABLED() ? After
all GCC should drop the calculation when not used.

I like this one. I'm still unsure how clang can get so confused about whether
the variables are initialized or not, usually it handles this much better than
gcc. My best guess is that one of the memory clobbers makes it conclude
that 'hflush' can be true when it gets written to by an inline asm.

As far as I am aware, clang's analysis does not evaluate variables when
generating a control flow graph and using that for static analysis:

https://godbolt.org/z/PdGxoq9j7

Based on the control flow graph, it knows that hstart and hend are
uninitialized because IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) gets
expanded to 0 by the preprocessor but it does not seem like it can piece
together that hflush's value of false is only changed to true under the
now 'if (0) {' branch, meaning that all the calls to __tlbiel_va_range()
never get evaluated. That may or may not be easy to fix in clang but we
run into issues like this so infrequently.

At any rate, the below diff works for me.

Cheers,
Nathan

18 more quoted lines

Yes I like that much better.

Maybe even better with

	hflush = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hstart < hend;

(And remove default false value at declaration).

  
  		if (type == FLUSH_TYPE_LOCAL) {
  			asm volatile("ptesync": : :"memory");