s390/mm,tlb: race of lazy TLB flush vs. recreation of TLB entries

Git commit 050eef364ad70059 "[S390] fix tlb flushing vs. concurrent
/proc accesses" introduced the attach counter to avoid using the
mm_users value to decide between IPTE for every PTE and lazy TLB
flushing with IDTE. That fixed the problem with mm_users but it
introduced another subtle race, fortunately one that is very hard
to hit.
The background is the requirement of the architecture that a valid
PTE may not be changed while it can be used concurrently by another
cpu. The decision between IPTE and lazy TLB flushing needs to be
done while the PTE is still valid. Now if the virtual cpu is
temporarily stopped after the decision to use lazy TLB flushing but
before the invalid bit of the PTE has been set, another cpu can attach
the mm, find that flush_mm is set, do the IDTE, return to userspace,
and recreate a TLB that uses the PTE in question. When the first,
stopped cpu continues it will change the PTE while it is attached on
another cpu. The first cpu will do another IDTE shortly after the
modification of the PTE which makes the race window quite short.

To fix this race the CPU that wants to attach the address space of a
user space thread needs to wait for the end of the PTE modification.
The number of concurrent TLB flushers for an mm is tracked in the
upper 16 bits of the attach_count and finish_arch_post_lock_switch
is used to wait for the end of the flush operation if required.

Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>

arch/s390/include/asm/mmu_context.h

@@ -48,13 +48,42 @@ static inline void update_mm(struct mm_struct *mm, struct task_struct *tsk)
 static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next,
 			     struct task_struct *tsk)
 {
-	cpumask_set_cpu(smp_processor_id(), mm_cpumask(next));
-	update_mm(next, tsk);
+	int cpu = smp_processor_id();
+
+	if (prev == next)
+		return;
+	if (atomic_inc_return(&next->context.attach_count) >> 16) {
+		/* Delay update_mm until all TLB flushes are done. */
+		set_tsk_thread_flag(tsk, TIF_TLB_WAIT);
+	} else {
+		cpumask_set_cpu(cpu, mm_cpumask(next));
+		update_mm(next, tsk);
+		if (next->context.flush_mm)
+			/* Flush pending TLBs */
+			__tlb_flush_mm(next);
+	}
 	atomic_dec(&prev->context.attach_count);
 	WARN_ON(atomic_read(&prev->context.attach_count) < 0);
-	atomic_inc(&next->context.attach_count);
-	/* Check for TLBs not flushed yet */
-	__tlb_flush_mm_lazy(next);
+}
+
+#define finish_arch_post_lock_switch finish_arch_post_lock_switch
+static inline void finish_arch_post_lock_switch(void)
+{
+	struct task_struct *tsk = current;
+	struct mm_struct *mm = tsk->mm;
+
+	if (!test_tsk_thread_flag(tsk, TIF_TLB_WAIT))
+		return;
+	preempt_disable();
+	clear_tsk_thread_flag(tsk, TIF_TLB_WAIT);
+	while (atomic_read(&mm->context.attach_count) >> 16)
+		cpu_relax();
+
+	cpumask_set_cpu(smp_processor_id(), mm_cpumask(mm));
+	update_mm(mm, tsk);
+	if (mm->context.flush_mm)
+		__tlb_flush_mm(mm);
+	preempt_enable();
 }
 
 #define enter_lazy_tlb(mm,tsk)	do { } while (0)

arch/s390/include/asm/pgtable.h

@@ -1034,30 +1034,41 @@ static inline int ptep_test_and_clear_user_young(struct mm_struct *mm,
 
 static inline void __ptep_ipte(unsigned long address, pte_t *ptep)
 {
-	if (!(pte_val(*ptep) & _PAGE_INVALID)) {
+	unsigned long pto = (unsigned long) ptep;
+
 #ifndef CONFIG_64BIT
-		/* pto must point to the start of the segment table */
-		pte_t *pto = (pte_t *) (((unsigned long) ptep) & 0x7ffffc00);
-#else
-		/* ipte in zarch mode can do the math */
-		pte_t *pto = ptep;
+	/* pto in ESA mode must point to the start of the segment table */
+	pto &= 0x7ffffc00;
 #endif
-		asm volatile(
-			"	ipte	%2,%3"
-			: "=m" (*ptep) : "m" (*ptep),
-			  "a" (pto), "a" (address));
-	}
+	/* Invalidation + global TLB flush for the pte */
+	asm volatile(
+		"	ipte	%2,%3"
+		: "=m" (*ptep) : "m" (*ptep), "a" (pto), "a" (address));
+}
+
+static inline void ptep_flush_direct(struct mm_struct *mm,
+				     unsigned long address, pte_t *ptep)
+{
+	if (pte_val(*ptep) & _PAGE_INVALID)
+		return;
+	__ptep_ipte(address, ptep);
 }
 
 static inline void ptep_flush_lazy(struct mm_struct *mm,
 				   unsigned long address, pte_t *ptep)
 {
-	int active = (mm == current->active_mm) ? 1 : 0;
+	int active, count;
 
-	if (atomic_read(&mm->context.attach_count) > active)
-		__ptep_ipte(address, ptep);
-	else
+	if (pte_val(*ptep) & _PAGE_INVALID)
+		return;
+	active = (mm == current->active_mm) ? 1 : 0;
+	count = atomic_add_return(0x10000, &mm->context.attach_count);
+	if ((count & 0xffff) <= active) {
+		pte_val(*ptep) |= _PAGE_INVALID;
 		mm->context.flush_mm = 1;
+	} else
+		__ptep_ipte(address, ptep);
+	atomic_sub(0x10000, &mm->context.attach_count);
 }
 
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
@@ -1074,7 +1085,7 @@ static inline int ptep_test_and_clear_young(struct vm_area_struct *vma,
 	}
 
 	pte = *ptep;
-	__ptep_ipte(addr, ptep);
+	ptep_flush_direct(vma->vm_mm, addr, ptep);
 	young = pte_young(pte);
 	pte = pte_mkold(pte);
 
@@ -1145,7 +1156,6 @@ static inline pte_t ptep_modify_prot_start(struct mm_struct *mm,
 
 	pte = *ptep;
 	ptep_flush_lazy(mm, address, ptep);
-	pte_val(*ptep) |= _PAGE_INVALID;
 
 	if (mm_has_pgste(mm)) {
 		pgste = pgste_update_all(&pte, pgste);
@@ -1182,7 +1192,7 @@ static inline pte_t ptep_clear_flush(struct vm_area_struct *vma,
 	}
 
 	pte = *ptep;
-	__ptep_ipte(address, ptep);
+	ptep_flush_direct(vma->vm_mm, address, ptep);
 	pte_val(*ptep) = _PAGE_INVALID;
 
 	if (mm_has_pgste(vma->vm_mm)) {
@@ -1263,7 +1273,7 @@ static inline int ptep_set_access_flags(struct vm_area_struct *vma,
 		pgste = pgste_ipte_notify(vma->vm_mm, address, ptep, pgste);
 	}
 
-	__ptep_ipte(address, ptep);
+	ptep_flush_direct(vma->vm_mm, address, ptep);
 
 	if (mm_has_pgste(vma->vm_mm)) {
 		pgste_set_pte(ptep, entry);
@@ -1447,12 +1457,16 @@ static inline pmd_t pmd_mkwrite(pmd_t pmd)
 static inline void pmdp_flush_lazy(struct mm_struct *mm,
 				   unsigned long address, pmd_t *pmdp)
 {
-	int active = (mm == current->active_mm) ? 1 : 0;
+	int active, count;
 
-	if ((atomic_read(&mm->context.attach_count) & 0xffff) > active)
-		__pmd_idte(address, pmdp);
-	else
+	active = (mm == current->active_mm) ? 1 : 0;
+	count = atomic_add_return(0x10000, &mm->context.attach_count);
+	if ((count & 0xffff) <= active) {
+		pmd_val(*pmdp) |= _SEGMENT_ENTRY_INVALID;
 		mm->context.flush_mm = 1;
+	} else
+		__pmd_idte(address, pmdp);
+	atomic_sub(0x10000, &mm->context.attach_count);
 }
 
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE

arch/s390/include/asm/thread_info.h

@@ -81,6 +81,7 @@ static inline struct thread_info *current_thread_info(void)
 #define TIF_NOTIFY_RESUME	1	/* callback before returning to user */
 #define TIF_SIGPENDING		2	/* signal pending */
 #define TIF_NEED_RESCHED	3	/* rescheduling necessary */
+#define TIF_TLB_WAIT		4	/* wait for TLB flush completion */
 #define TIF_PER_TRAP		6	/* deliver sigtrap on return to user */
 #define TIF_MCCK_PENDING	7	/* machine check handling is pending */
 #define TIF_SYSCALL_TRACE	8	/* syscall trace active */
@@ -96,6 +97,7 @@ static inline struct thread_info *current_thread_info(void)
 #define _TIF_NOTIFY_RESUME	(1<<TIF_NOTIFY_RESUME)
 #define _TIF_SIGPENDING		(1<<TIF_SIGPENDING)
 #define _TIF_NEED_RESCHED	(1<<TIF_NEED_RESCHED)
+#define _TIF_TLB_WAIT		(1<<TIF_TLB_WAIT)
 #define _TIF_PER_TRAP		(1<<TIF_PER_TRAP)
 #define _TIF_MCCK_PENDING	(1<<TIF_MCCK_PENDING)
 #define _TIF_SYSCALL_TRACE	(1<<TIF_SYSCALL_TRACE)

arch/s390/kernel/entry.S

@@ -43,6 +43,7 @@ _TIF_WORK_INT = (_TIF_SIGPENDING | _TIF_NOTIFY_RESUME | _TIF_NEED_RESCHED | \
 		 _TIF_MCCK_PENDING)
 _TIF_TRACE    = (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_AUDIT | _TIF_SECCOMP | \
 		 _TIF_SYSCALL_TRACEPOINT)
+_TIF_TRANSFER = (_TIF_MCCK_PENDING | _TIF_TLB_WAIT)
 
 STACK_SHIFT = PAGE_SHIFT + THREAD_ORDER
 STACK_SIZE  = 1 << STACK_SHIFT
@@ -159,10 +160,12 @@ ENTRY(__switch_to)
 	lctl	%c4,%c4,__TASK_pid(%r3)		# load pid to control reg. 4
 	mvc	__LC_CURRENT_PID(4,%r0),__TASK_pid(%r3)	# store pid of next
 	l	%r15,__THREAD_ksp(%r3)		# load kernel stack of next
-	tm	__TI_flags+3(%r4),_TIF_MCCK_PENDING # machine check pending?
+	lhi	%r6,_TIF_TRANSFER		# transfer TIF bits
+	n	%r6,__TI_flags(%r4)		# isolate TIF bits
 	jz	0f
-	ni	__TI_flags+3(%r4),255-_TIF_MCCK_PENDING	# clear flag in prev
-	oi	__TI_flags+3(%r5),_TIF_MCCK_PENDING	# set it in next
+	o	%r6,__TI_flags(%r5)		# set TIF bits of next
+	st	%r6,__TI_flags(%r5)
+	ni	__TI_flags+3(%r4),255-_TIF_TRANSFER # clear TIF bits of prev
 0:	lm	%r6,%r15,__SF_GPRS(%r15)	# load gprs of next task
 	br	%r14
 

arch/s390/kernel/entry64.S

@@ -48,6 +48,7 @@ _TIF_WORK_INT = (_TIF_SIGPENDING | _TIF_NOTIFY_RESUME | _TIF_NEED_RESCHED | \
 		 _TIF_MCCK_PENDING)
 _TIF_TRACE    = (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_AUDIT | _TIF_SECCOMP | \
 		 _TIF_SYSCALL_TRACEPOINT)
+_TIF_TRANSFER = (_TIF_MCCK_PENDING | _TIF_TLB_WAIT)
 
 #define BASED(name) name-system_call(%r13)
 
@@ -189,10 +190,12 @@ ENTRY(__switch_to)
 	lctl	%c4,%c4,__TASK_pid(%r3)		# load pid to control reg. 4
 	mvc	__LC_CURRENT_PID+4(4,%r0),__TASK_pid(%r3) # store pid of next
 	lg	%r15,__THREAD_ksp(%r3)		# load kernel stack of next
-	tm	__TI_flags+7(%r4),_TIF_MCCK_PENDING # machine check pending?
+	llill	%r6,_TIF_TRANSFER		# transfer TIF bits
+	ng	%r6,__TI_flags(%r4)		# isolate TIF bits
 	jz	0f
-	ni	__TI_flags+7(%r4),255-_TIF_MCCK_PENDING	# clear flag in prev
-	oi	__TI_flags+7(%r5),_TIF_MCCK_PENDING	# set it in next
+	og	%r6,__TI_flags(%r5)		# set TIF bits of next
+	stg	%r6,__TI_flags(%r5)
+	ni	__TI_flags+7(%r4),255-_TIF_TRANSFER # clear TIF bits of prev
 0:	lmg	%r6,%r15,__SF_GPRS(%r15)	# load gprs of next task
 	br	%r14