Skip to contentSkip to conversation

lazy tlb: consolidate lazy tlb mm switching fix

PATCH In review Nicholas Piggin 2 participants 3 messages Jun 7, 2023 – Nov 15, 2023

Subsystems (2)
lore
Share & tools
lore

Nicholas Piggin

author reviewer
#0 lore raw

PATCH 1/1 lazy tlb: consolidate lazy tlb mm switching fix

Fix an upstream powerpc bug that was discovered with a WARN_ON added in "lazy tlb: consolidate lazy tlb mm switching":

Switching mm and tinkering with current->active_mm should be done with
irqs disabled. There is a path where exit_lazy_flush_tlb can be called with irqs enabled:

exit_lazy_flush_tlb
flush_type_needed
__flush_all_mm
tlb_finish_mmu
exit_mmap

Which results in the switching being incorrectly with irqs enabled.

Fixes a665eec0a22e1 ("powerpc/64s/radix: Fix mm_cpumask trimming race vs k>

I will send that fix upstream, which will end up conflicting with these. The resolution will be to take akpm's side of the merge (with this patch). Or drop the patch when it conflicts and I'll resend it rebased on the fix.

---
 arch/powerpc/mm/book3s64/radix_tlb.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
arch/powerpc/mm/book3s64/radix_tlb.c +9 -1
index 90953cf9f648..8d71fe8f5cdb 100644
@@ -795,8 +795,16 @@ void exit_lazy_flush_tlb(struct mm_struct *mm, bool always_flush)
 		goto out;
 
 	if (current->active_mm == mm) {
-		/* Is a kernel thread and is using mm as the lazy tlb */
+		unsigned long flags;
+
+		/*
+		 * It is a kernel thread and is using mm as the lazy tlb, so
+		 * switch it to init_mm. This is not always called from IPI
+		 * (e.g., flush_type_needed), so must disable irqs.
+		 */
+		local_irq_save(flags);
 		kthread_end_lazy_tlb_mm();
+		local_irq_restore(flags);
 	}
 
 	/*
Signature · 1 line
2.40.1
#1 lore raw

Le 07/06/2023 à 02:56, Nicholas Piggin a écrit :

Quoted 13 lines

Fix an upstream powerpc bug that was discovered with a WARN_ON added in "lazy tlb: consolidate lazy tlb mm switching":

Switching mm and tinkering with current->active_mm should be done with
irqs disabled. There is a path where exit_lazy_flush_tlb can be called with irqs enabled:

exit_lazy_flush_tlb
flush_type_needed
__flush_all_mm
tlb_finish_mmu
exit_mmap

Quoted 8 lines

Which results in the switching being incorrectly with irqs enabled.

Fixes a665eec0a22e1 ("powerpc/64s/radix: Fix mm_cpumask trimming race vs k>

I will send that fix upstream, which will end up conflicting with these. The resolution will be to take akpm's side of the merge (with this patch). Or drop the patch when it conflicts and I'll resend it rebased on the fix.

This seems to badly conflict at the moment, is it still applicable ?

Christophe

Quoted 28 lines
---
  arch/powerpc/mm/book3s64/radix_tlb.c | 10 +++++++++-
  1 file changed, 9 insertions(+), 1 deletion(-)
arch/powerpc/mm/book3s64/radix_tlb.c +9 -1
index 90953cf9f648..8d71fe8f5cdb 100644
@@ -795,8 +795,16 @@ void exit_lazy_flush_tlb(struct mm_struct *mm, bool always_flush)
  		goto out;
 
  	if (current->active_mm == mm) {
-		/* Is a kernel thread and is using mm as the lazy tlb */
+		unsigned long flags;
+
+		/*
+		 * It is a kernel thread and is using mm as the lazy tlb, so
+		 * switch it to init_mm. This is not always called from IPI
+		 * (e.g., flush_type_needed), so must disable irqs.
+		 */
+		local_irq_save(flags);
  		kthread_end_lazy_tlb_mm();
+		local_irq_restore(flags);
  	}
 
  	/*
author reviewer
#2 lore raw
Quoted 18 lines · Christophe Leroy

Le 07/06/2023 à 02:56, Nicholas Piggin a écrit :

Quoted 15 lines

Fix an upstream powerpc bug that was discovered with a WARN_ON added in "lazy tlb: consolidate lazy tlb mm switching":

Switching mm and tinkering with current->active_mm should be done with
irqs disabled. There is a path where exit_lazy_flush_tlb can be called with irqs enabled:

exit_lazy_flush_tlb
flush_type_needed
__flush_all_mm
tlb_finish_mmu
exit_mmap

Which results in the switching being incorrectly with irqs enabled.

Quoted 8 lines
Quoted 6 lines
Fixes a665eec0a22e1 ("powerpc/64s/radix: Fix mm_cpumask trimming race vs k>

I will send that fix upstream, which will end up conflicting with these. The resolution will be to take akpm's side of the merge (with this patch). Or drop the patch when it conflicts and I'll resend it rebased on the fix.

This seems to badly conflict at the moment, is it still applicable ?

Been away, will have to check where all this ended up.

Thanks,
Nick