[patch] smp-2.3.41-A1, (was: Re: BUG? i386 TLB Flushing)

[Ingo Molnar]

On Mon, 24 Jan 2000, Stephen C. Tweedie wrote:

> > yeah, this is indeed a bug and it also explains the one remaining 64GB
> > mystery: the extra invlpg in highmem.c. Patch (with other SMP fixes) will
> > follow shortly.
> 
> I was really, really hoping that this might lay to rest the boot
> failures with PAE36 for me, but no such luck --- 2.3.40 still dies with
> SIGSEGV in init.  Zach Brown can reproduce the same thing on one of his
> boxes: Zach, does 2.3.40 do anything to help things for you?

ok, how about the attached smp-2.3.41-A1 patch:

	- fixes dumb cr4 flushing typo found by Manfred

	- (tries to fix) current->active_mm == NULL problem by moving the
	  setting of prev->active_mm to after the switch_to(). It happens
	  before has_cpu is set to 0, so it should be safe. It also cannot
	  open up a window to TLB corruption i believe.

-- mingo
------------ próxima parte ------------
--- linux/kernel/sched.c.orig	Mon Jan 24 12:15:41 2000
+++ linux/kernel/sched.c	Mon Jan 24 12:16:43 2000
@@ -412,6 +412,8 @@
 static inline void __schedule_tail(struct task_struct *prev)
 {
 #ifdef __SMP__
+	if (!prev->mm)
+		prev->active_mm = NULL;
 	if ((prev->state == TASK_RUNNING) &&
 			(prev != idle_task(smp_processor_id()))) {
 		unsigned long flags;
@@ -586,10 +588,8 @@
 			switch_mm(oldmm, mm, next, this_cpu);
 		}
 
-		if (!prev->mm) {
-			prev->active_mm = NULL;
+		if (!prev->mm)
 			mmdrop(oldmm);
-		}
 	}
 
 	/*
--- linux/include/asm-i386/pgtable.h.orig	Mon Jan 24 12:10:19 2000
+++ linux/include/asm-i386/pgtable.h	Mon Jan 24 12:10:33 2000
@@ -48,8 +48,8 @@
 			"mov %2, %%cr3;  # flush TLB \n"		\
 			"mov %1, %%cr4;  # turn PGE back on \n"		\
 			:						\
-			: "r" (mmu_cr4_features),			\
-			  "r" (mmu_cr4_features & ~X86_CR4_PGE),	\
+			: "r" (mmu_cr4_features & ~X86_CR4_PGE),	\
+			  "r" (mmu_cr4_features),			\
 			  "r" (__pa(current->active_mm->pgd))		\
 			: "memory");					\
 	} while (0)