Re: [PATCH RFC] m68k: skip kernel premption if interrupts were disabled

[Finn Thain <fthain@linux-m68k.org>]

On Tue, 26 Mar 2024, Michael Schmitz wrote:

> see below (hopefully not whitespace-damaged).
> 

Following your approach, I added mutual exclusion to the other candidate 
in the same file, namely get_pointer_table(). Together with your patch, 
that seems to be sufficient to make stress-ng happy. I don't know the real 
extent of the data race here so I've probably used too much locking in 
this experiment (?)

diff --git a/arch/m68k/mm/motorola.c b/arch/m68k/mm/motorola.c
index 068ad7c9bb5c..d1ad55ee1b60 100644
--- a/arch/m68k/mm/motorola.c
+++ b/arch/m68k/mm/motorola.c
@@ -140,10 +140,14 @@ void __init init_pointer_table(void *table, int type)
 
 void *get_pointer_table(int type)
 {
-	ptable_desc *dp = ptable_list[type].next;
-	unsigned int mask = list_empty(&ptable_list[type]) ? 0 : PD_MARKBITS(dp);
+	ptable_desc *dp;
+	unsigned int mask;
 	unsigned int tmp, off;
 
+	preempt_disable();
+
+	dp = ptable_list[type].next;
+	mask = list_empty(&ptable_list[type]) ? 0 : PD_MARKBITS(dp);
 	/*
 	 * For a pointer table for a user process address space, a
 	 * table is taken from a page allocated for the purpose.  Each
@@ -154,8 +158,13 @@ void *get_pointer_table(int type)
 		void *page;
 		ptable_desc *new;
 
-		if (!(page = (void *)get_zeroed_page(GFP_KERNEL)))
+		sched_preempt_enable_no_resched();
+
+		if (!(page = (void *)get_zeroed_page(GFP_KERNEL))) {
 			return NULL;
+		}
+
+		preempt_disable();
 
 		if (type == TABLE_PTE) {
 			/*
@@ -171,6 +180,7 @@ void *get_pointer_table(int type)
 		PD_MARKBITS(new) = ptable_mask(type) - 1;
 		list_add_tail(new, dp);
 
+		sched_preempt_enable_no_resched();
 		return (pmd_t *)page;
 	}
 
@@ -181,6 +191,7 @@ void *get_pointer_table(int type)
 		/* move to end of list */
 		list_move_tail(dp, &ptable_list[type]);
 	}
+	sched_preempt_enable_no_resched();
 	return page_address(PD_PAGE(dp)) + off;
 }