Skip to contentSkip to conversation

SPARC v9: __tsb_context_switch - demap old tsb_map_vaddr

PATCH In review Kirill Tkhai 2 participants 2 messages Oct 5, 2010

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 SPARC v9: __tsb_context_switch - demap old tsb_map_vaddr

SPARC v9: __tsb_context_switch - demap old tsb_map_vaddr
---
When TSB context switches in the function switch_mm() the following situation could occur. We have old mapping between physical address and tsb_map_vaddr in uTLB, and new mapping is written in fDTLB. Because uDTLB has more priority than fDTLB, all transformations occur through it, and TSB Pointer in tl0_damiss will be formed with the wrong physical address. To prevent this, we need to flush old tsb_map_vaddr (TSB_BASE) in TLB. Kernel version 2.6.35.7.

---

sparc/kernel/tsb.S 2010-10-05 19:48:56.136235329 +0400 renamed from sparc/kernel/tsb.S.orig 2010-10-05 19:46:55.272218523 +0400 +3 -0
@@ -396,6 +396,9 @@ __tsb_context_switch:
 
 2:	ldx	[%o1 + TSB_CONFIG_MAP_VADDR], %o4
 	brz	%o4, 9f
+	or	%o4, 1 << 5, %o5
+	stxa	%g0, [%o5] ASI_DMMU_DEMAP
+	membar	#Sync
 	 ldx	[%o1 + TSB_CONFIG_MAP_PTE], %o5
 
 	sethi	%hi(sparc64_highest_unlocked_tlb_ent), %g2
#1 lore raw

From: Kirill Tkhai <[email protected]>
Date: Tue, 05 Oct 2010 23:57:56 +0400

Quoted 10 lines

When TSB context switches in the function switch_mm() the following situation could occur. We have old mapping between physical address and tsb_map_vaddr in uTLB, and new mapping is written in fDTLB. Because uDTLB has more priority than fDTLB, all transformations occur through it, and TSB Pointer in tl0_damiss will be formed with the wrong physical address. To prevent this, we need to flush old tsb_map_vaddr (TSB_BASE) in TLB. Kernel version 2.6.35.7.

By all pieces of documentation that exist, the uTLB only handles instruction stream accesses.

Therefore, this patch should not be necessary.

Unless you can show me where in the UltraSPARC documentation it states that the uTLB is used for data accesses, I am not applying this patch :)