s390/mm,tlb: optimize TLB flushing for zEC12

The zEC12 machines introduced the local-clearing control for the IDTE
and IPTE instruction. If the control is set only the TLB of the local
CPU is cleared of entries, either all entries of a single address space
for IDTE, or the entry for a single page-table entry for IPTE.
Without the local-clearing control the TLB flush is broadcasted to all
CPUs in the configuration, which is expensive.

The reset of the bit mask of the CPUs that need flushing after a
non-local IDTE is tricky. As TLB entries for an address space remain
in the TLB even if the address space is detached a new bit field is
required to keep track of attached CPUs vs. CPUs in the need of a
flush. After a non-local flush with IDTE the bit-field of attached CPUs
is copied to the bit-field of CPUs in need of a flush. The ordering
of operations on cpu_attach_mask, attach_count and mm_cpumask(mm) is
such that an underindication in mm_cpumask(mm) is prevented but an
overindication in mm_cpumask(mm) is possible.

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

arch/s390/include/asm/mmu.h

@@ -1,9 +1,11 @@
 #ifndef __MMU_H
 #define __MMU_H
 
+#include <linux/cpumask.h>
 #include <linux/errno.h>
 
 typedef struct {
+	cpumask_t cpu_attach_mask;
 	atomic_t attach_count;
 	unsigned int flush_mm;
 	spinlock_t list_lock;

arch/s390/include/asm/mmu_context.h

@@ -15,6 +15,7 @@
 static inline int init_new_context(struct task_struct *tsk,
 				   struct mm_struct *mm)
 {
+	cpumask_clear(&mm->context.cpu_attach_mask);
 	atomic_set(&mm->context.attach_count, 0);
 	mm->context.flush_mm = 0;
 	mm->context.asce_bits = _ASCE_TABLE_LENGTH | _ASCE_USER_BITS;
@@ -59,6 +60,8 @@ static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next,
 
 	if (prev == next)
 		return;
+	if (MACHINE_HAS_TLB_LC)
+		cpumask_set_cpu(cpu, &next->context.cpu_attach_mask);
 	if (atomic_inc_return(&next->context.attach_count) >> 16) {
 		/* Delay update_user_asce until all TLB flushes are done. */
 		set_tsk_thread_flag(tsk, TIF_TLB_WAIT);
@@ -73,6 +76,8 @@ static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next,
 	}
 	atomic_dec(&prev->context.attach_count);
 	WARN_ON(atomic_read(&prev->context.attach_count) < 0);
+	if (MACHINE_HAS_TLB_LC)
+		cpumask_clear_cpu(cpu, &prev->context.cpu_attach_mask);
 }
 
 #define finish_arch_post_lock_switch finish_arch_post_lock_switch

arch/s390/include/asm/pgtable.h

@@ -1068,12 +1068,35 @@ static inline void __ptep_ipte(unsigned long address, pte_t *ptep)
 		: "=m" (*ptep) : "m" (*ptep), "a" (pto), "a" (address));
 }
 
+static inline void __ptep_ipte_local(unsigned long address, pte_t *ptep)
+{
+	unsigned long pto = (unsigned long) ptep;
+
+#ifndef CONFIG_64BIT
+	/* pto in ESA mode must point to the start of the segment table */
+	pto &= 0x7ffffc00;
+#endif
+	/* Invalidation + local TLB flush for the pte */
+	asm volatile(
+		"	.insn rrf,0xb2210000,%2,%3,0,1"
+		: "=m" (*ptep) : "m" (*ptep), "a" (pto), "a" (address));
+}
+
 static inline void ptep_flush_direct(struct mm_struct *mm,
 				     unsigned long address, pte_t *ptep)
 {
+	int active, count;
+
 	if (pte_val(*ptep) & _PAGE_INVALID)
 		return;
-	__ptep_ipte(address, ptep);
+	active = (mm == current->active_mm) ? 1 : 0;
+	count = atomic_add_return(0x10000, &mm->context.attach_count);
+	if (MACHINE_HAS_TLB_LC && (count & 0xffff) <= active &&
+	    cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id())))
+		__ptep_ipte_local(address, ptep);
+	else
+		__ptep_ipte(address, ptep);
+	atomic_sub(0x10000, &mm->context.attach_count);
 }
 
 static inline void ptep_flush_lazy(struct mm_struct *mm,
@@ -1382,35 +1405,6 @@ static inline pmd_t *pmd_offset(pud_t *pud, unsigned long address)
 #define pte_offset_map(pmd, address) pte_offset_kernel(pmd, address)
 #define pte_unmap(pte) do { } while (0)
 
-static inline void __pmd_idte(unsigned long address, pmd_t *pmdp)
-{
-	unsigned long sto = (unsigned long) pmdp -
-			    pmd_index(address) * sizeof(pmd_t);
-
-	if (!(pmd_val(*pmdp) & _SEGMENT_ENTRY_INVALID)) {
-		asm volatile(
-			"	.insn	rrf,0xb98e0000,%2,%3,0,0"
-			: "=m" (*pmdp)
-			: "m" (*pmdp), "a" (sto),
-			  "a" ((address & HPAGE_MASK))
-			: "cc"
-		);
-	}
-}
-
-static inline void __pmd_csp(pmd_t *pmdp)
-{
-	register unsigned long reg2 asm("2") = pmd_val(*pmdp);
-	register unsigned long reg3 asm("3") = pmd_val(*pmdp) |
-					       _SEGMENT_ENTRY_INVALID;
-	register unsigned long reg4 asm("4") = ((unsigned long) pmdp) + 5;
-
-	asm volatile(
-		"	csp %1,%3"
-		: "=m" (*pmdp)
-		: "d" (reg2), "d" (reg3), "d" (reg4), "m" (*pmdp) : "cc");
-}
-
 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLB_PAGE)
 static inline unsigned long massage_pgprot_pmd(pgprot_t pgprot)
 {
@@ -1479,18 +1473,80 @@ static inline pmd_t pmd_mkwrite(pmd_t pmd)
 }
 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLB_PAGE */
 
+static inline void __pmdp_csp(pmd_t *pmdp)
+{
+	register unsigned long reg2 asm("2") = pmd_val(*pmdp);
+	register unsigned long reg3 asm("3") = pmd_val(*pmdp) |
+					       _SEGMENT_ENTRY_INVALID;
+	register unsigned long reg4 asm("4") = ((unsigned long) pmdp) + 5;
+
+	asm volatile(
+		"	csp %1,%3"
+		: "=m" (*pmdp)
+		: "d" (reg2), "d" (reg3), "d" (reg4), "m" (*pmdp) : "cc");
+}
+
+static inline void __pmdp_idte(unsigned long address, pmd_t *pmdp)
+{
+	unsigned long sto;
+
+	sto = (unsigned long) pmdp - pmd_index(address) * sizeof(pmd_t);
+	asm volatile(
+		"	.insn	rrf,0xb98e0000,%2,%3,0,0"
+		: "=m" (*pmdp)
+		: "m" (*pmdp), "a" (sto), "a" ((address & HPAGE_MASK))
+		: "cc" );
+}
+
+static inline void __pmdp_idte_local(unsigned long address, pmd_t *pmdp)
+{
+	unsigned long sto;
+
+	sto = (unsigned long) pmdp - pmd_index(address) * sizeof(pmd_t);
+	asm volatile(
+		"	.insn	rrf,0xb98e0000,%2,%3,0,1"
+		: "=m" (*pmdp)
+		: "m" (*pmdp), "a" (sto), "a" ((address & HPAGE_MASK))
+		: "cc" );
+}
+
+static inline void pmdp_flush_direct(struct mm_struct *mm,
+				     unsigned long address, pmd_t *pmdp)
+{
+	int active, count;
+
+	if (pmd_val(*pmdp) & _SEGMENT_ENTRY_INVALID)
+		return;
+	if (!MACHINE_HAS_IDTE) {
+		__pmdp_csp(pmdp);
+		return;
+	}
+	active = (mm == current->active_mm) ? 1 : 0;
+	count = atomic_add_return(0x10000, &mm->context.attach_count);
+	if (MACHINE_HAS_TLB_LC && (count & 0xffff) <= active &&
+	    cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id())))
+		__pmdp_idte_local(address, pmdp);
+	else
+		__pmdp_idte(address, pmdp);
+	atomic_sub(0x10000, &mm->context.attach_count);
+}
+
 static inline void pmdp_flush_lazy(struct mm_struct *mm,
 				   unsigned long address, pmd_t *pmdp)
 {
 	int active, count;
 
+	if (pmd_val(*pmdp) & _SEGMENT_ENTRY_INVALID)
+		return;
 	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);
+	} else if (MACHINE_HAS_IDTE)
+		__pmdp_idte(address, pmdp);
+	else
+		__pmdp_csp(pmdp);
 	atomic_sub(0x10000, &mm->context.attach_count);
 }
 
@@ -1543,7 +1599,7 @@ static inline int pmdp_test_and_clear_young(struct vm_area_struct *vma,
 	pmd_t pmd;
 
 	pmd = *pmdp;
-	__pmd_idte(address, pmdp);
+	pmdp_flush_direct(vma->vm_mm, address, pmdp);
 	*pmdp = pmd_mkold(pmd);
 	return pmd_young(pmd);
 }
@@ -1554,7 +1610,7 @@ static inline pmd_t pmdp_get_and_clear(struct mm_struct *mm,
 {
 	pmd_t pmd = *pmdp;
 
-	__pmd_idte(address, pmdp);
+	pmdp_flush_direct(mm, address, pmdp);
 	pmd_clear(pmdp);
 	return pmd;
 }
@@ -1570,7 +1626,7 @@ static inline pmd_t pmdp_clear_flush(struct vm_area_struct *vma,
 static inline void pmdp_invalidate(struct vm_area_struct *vma,
 				   unsigned long address, pmd_t *pmdp)
 {
-	__pmd_idte(address, pmdp);
+	pmdp_flush_direct(vma->vm_mm, address, pmdp);
 }
 
 #define __HAVE_ARCH_PMDP_SET_WRPROTECT
@@ -1580,7 +1636,7 @@ static inline void pmdp_set_wrprotect(struct mm_struct *mm,
 	pmd_t pmd = *pmdp;
 
 	if (pmd_write(pmd)) {
-		__pmd_idte(address, pmdp);
+		pmdp_flush_direct(mm, address, pmdp);
 		set_pmd_at(mm, address, pmdp, pmd_wrprotect(pmd));
 	}
 }

arch/s390/include/asm/setup.h

@@ -68,6 +68,7 @@ void create_mem_hole(struct mem_chunk mem_chunk[], unsigned long addr,
 #define MACHINE_FLAG_TOPOLOGY	(1UL << 14)
 #define MACHINE_FLAG_TE		(1UL << 15)
 #define MACHINE_FLAG_RRBM	(1UL << 16)
+#define MACHINE_FLAG_TLB_LC	(1UL << 17)
 
 #define MACHINE_IS_VM		(S390_lowcore.machine_flags & MACHINE_FLAG_VM)
 #define MACHINE_IS_KVM		(S390_lowcore.machine_flags & MACHINE_FLAG_KVM)
@@ -90,6 +91,7 @@ void create_mem_hole(struct mem_chunk mem_chunk[], unsigned long addr,
 #define MACHINE_HAS_TOPOLOGY	(0)
 #define MACHINE_HAS_TE		(0)
 #define MACHINE_HAS_RRBM	(0)
+#define MACHINE_HAS_TLB_LC	(0)
 #else /* CONFIG_64BIT */
 #define MACHINE_HAS_IEEE	(1)
 #define MACHINE_HAS_CSP		(1)
@@ -102,6 +104,7 @@ void create_mem_hole(struct mem_chunk mem_chunk[], unsigned long addr,
 #define MACHINE_HAS_TOPOLOGY	(S390_lowcore.machine_flags & MACHINE_FLAG_TOPOLOGY)
 #define MACHINE_HAS_TE		(S390_lowcore.machine_flags & MACHINE_FLAG_TE)
 #define MACHINE_HAS_RRBM	(S390_lowcore.machine_flags & MACHINE_FLAG_RRBM)
+#define MACHINE_HAS_TLB_LC	(S390_lowcore.machine_flags & MACHINE_FLAG_TLB_LC)
 #endif /* CONFIG_64BIT */
 
 /*

arch/s390/include/asm/tlbflush.h

@@ -7,19 +7,41 @@
 #include <asm/pgalloc.h>
 
 /*
- * Flush all tlb entries on the local cpu.
+ * Flush all TLB entries on the local CPU.
  */
 static inline void __tlb_flush_local(void)
 {
 	asm volatile("ptlb" : : : "memory");
 }
 
-#ifdef CONFIG_SMP
 /*
- * Flush all tlb entries on all cpus.
+ * Flush TLB entries for a specific ASCE on all CPUs
  */
+static inline void __tlb_flush_idte(unsigned long asce)
+{
+	/* Global TLB flush for the mm */
+	asm volatile(
+		"	.insn	rrf,0xb98e0000,0,%0,%1,0"
+		: : "a" (2048), "a" (asce) : "cc");
+}
+
+/*
+ * Flush TLB entries for a specific ASCE on the local CPU
+ */
+static inline void __tlb_flush_idte_local(unsigned long asce)
+{
+	/* Local TLB flush for the mm */
+	asm volatile(
+		"	.insn	rrf,0xb98e0000,0,%0,%1,1"
+		: : "a" (2048), "a" (asce) : "cc");
+}
+
+#ifdef CONFIG_SMP
 void smp_ptlb_all(void);
 
+/*
+ * Flush all TLB entries on all CPUs.
+ */
 static inline void __tlb_flush_global(void)
 {
 	register unsigned long reg2 asm("2");
@@ -42,36 +64,89 @@ static inline void __tlb_flush_global(void)
 		: : "d" (reg2), "d" (reg3), "d" (reg4), "m" (dummy) : "cc" );
 }
 
+/*
+ * Flush TLB entries for a specific mm on all CPUs (in case gmap is used
+ * this implicates multiple ASCEs!).
+ */
 static inline void __tlb_flush_full(struct mm_struct *mm)
 {
-	cpumask_t local_cpumask;
-
 	preempt_disable();
-	/*
-	 * If the process only ran on the local cpu, do a local flush.
-	 */
-	cpumask_copy(&local_cpumask, cpumask_of(smp_processor_id()));
-	if (cpumask_equal(mm_cpumask(mm), &local_cpumask))
+	atomic_add(0x10000, &mm->context.attach_count);
+	if (cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id()))) {
+		/* Local TLB flush */
 		__tlb_flush_local();
-	else
+	} else {
+		/* Global TLB flush */
 		__tlb_flush_global();
+		/* Reset TLB flush mask */
+		if (MACHINE_HAS_TLB_LC)
+			cpumask_copy(mm_cpumask(mm),
+				     &mm->context.cpu_attach_mask);
+	}
+	atomic_sub(0x10000, &mm->context.attach_count);
 	preempt_enable();
 }
+
+/*
+ * Flush TLB entries for a specific ASCE on all CPUs.
+ */
+static inline void __tlb_flush_asce(struct mm_struct *mm, unsigned long asce)
+{
+	int active, count;
+
+	preempt_disable();
+	active = (mm == current->active_mm) ? 1 : 0;
+	count = atomic_add_return(0x10000, &mm->context.attach_count);
+	if (MACHINE_HAS_TLB_LC && (count & 0xffff) <= active &&
+	    cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id()))) {
+		__tlb_flush_idte_local(asce);
+	} else {
+		if (MACHINE_HAS_IDTE)
+			__tlb_flush_idte(asce);
+		else
+			__tlb_flush_global();
+		/* Reset TLB flush mask */
+		if (MACHINE_HAS_TLB_LC)
+			cpumask_copy(mm_cpumask(mm),
+				     &mm->context.cpu_attach_mask);
+	}
+	atomic_sub(0x10000, &mm->context.attach_count);
+	preempt_enable();
+}
+
+static inline void __tlb_flush_kernel(void)
+{
+	if (MACHINE_HAS_IDTE)
+		__tlb_flush_idte((unsigned long) init_mm.pgd |
+				 init_mm.context.asce_bits);
+	else
+		__tlb_flush_global();
+}
 #else
-#define __tlb_flush_full(mm)	__tlb_flush_local()
 #define __tlb_flush_global()	__tlb_flush_local()
-#endif
+#define __tlb_flush_full(mm)	__tlb_flush_local()
 
 /*
- * Flush all tlb entries of a page table on all cpus.
+ * Flush TLB entries for a specific ASCE on all CPUs.
  */
-static inline void __tlb_flush_idte(unsigned long asce)
+static inline void __tlb_flush_asce(struct mm_struct *mm, unsigned long asce)
 {
-	asm volatile(
-		"	.insn	rrf,0xb98e0000,0,%0,%1,0"
-		: : "a" (2048), "a" (asce) : "cc" );
+	if (MACHINE_HAS_TLB_LC)
+		__tlb_flush_idte_local(asce);
+	else
+		__tlb_flush_local();
 }
 
+static inline void __tlb_flush_kernel(void)
+{
+	if (MACHINE_HAS_TLB_LC)
+		__tlb_flush_idte_local((unsigned long) init_mm.pgd |
+				       init_mm.context.asce_bits);
+	else
+		__tlb_flush_local();
+}
+#endif
+
 static inline void __tlb_flush_mm(struct mm_struct * mm)
 {
 	/*
@@ -80,7 +155,7 @@ static inline void __tlb_flush_mm(struct mm_struct * mm)
 	 * only ran on the local cpu.
 	 */
 	if (MACHINE_HAS_IDTE && list_empty(&mm->context.gmap_list))
-		__tlb_flush_idte((unsigned long) mm->pgd |
+		__tlb_flush_asce(mm, (unsigned long) mm->pgd |
 				 mm->context.asce_bits);
 	else
 		__tlb_flush_full(mm);
@@ -130,7 +205,7 @@ static inline void flush_tlb_range(struct vm_area_struct *vma,
 static inline void flush_tlb_kernel_range(unsigned long start,
 					  unsigned long end)
 {
-	__tlb_flush_mm(&init_mm);
+	__tlb_flush_kernel();
 }
 
 #endif /* _S390_TLBFLUSH_H */

arch/s390/kernel/early.c

@@ -386,6 +386,8 @@ static __init void detect_machine_facilities(void)
 		S390_lowcore.machine_flags |= MACHINE_FLAG_TE;
 	if (test_facility(66))
 		S390_lowcore.machine_flags |= MACHINE_FLAG_RRBM;
+	if (test_facility(51))
+		S390_lowcore.machine_flags |= MACHINE_FLAG_TLB_LC;
 #endif
 }
 

arch/s390/kernel/smp.c

@@ -236,6 +236,9 @@ static void pcpu_prepare_secondary(struct pcpu *pcpu, int cpu)
 {
 	struct _lowcore *lc = pcpu->lowcore;
 
+	if (MACHINE_HAS_TLB_LC)
+		cpumask_set_cpu(cpu, &init_mm.context.cpu_attach_mask);
+	cpumask_set_cpu(cpu, mm_cpumask(&init_mm));
 	atomic_inc(&init_mm.context.attach_count);
 	lc->cpu_nr = cpu;
 	lc->percpu_offset = __per_cpu_offset[cpu];
@@ -760,6 +763,9 @@ void __cpu_die(unsigned int cpu)
 		cpu_relax();
 	pcpu_free_lowcore(pcpu);
 	atomic_dec(&init_mm.context.attach_count);
+	cpumask_clear_cpu(cpu, mm_cpumask(&init_mm));
+	if (MACHINE_HAS_TLB_LC)
+		cpumask_clear_cpu(cpu, &init_mm.context.cpu_attach_mask);
 }
 
 void __noreturn cpu_die(void)

arch/s390/mm/hugetlbpage.c

@@ -123,10 +123,7 @@ pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
 	pmd_t *pmdp = (pmd_t *) ptep;
 	pte_t pte = huge_ptep_get(ptep);
 
-	if (MACHINE_HAS_IDTE)
-		__pmd_idte(addr, pmdp);
-	else
-		__pmd_csp(pmdp);
+	pmdp_flush_direct(mm, addr, pmdp);
 	pmd_val(*pmdp) = _SEGMENT_ENTRY_EMPTY;
 	return pte;
 }

arch/s390/mm/init.c

@@ -124,8 +124,6 @@ void __init paging_init(void)
 	__ctl_load(S390_lowcore.kernel_asce, 13, 13);
 	arch_local_irq_restore(4UL << (BITS_PER_LONG - 8));
 
-	atomic_set(&init_mm.context.attach_count, 1);
-
 	sparse_memory_present_with_active_regions(MAX_NUMNODES);
 	sparse_init();
 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
@@ -136,6 +134,11 @@ void __init paging_init(void)
 
 void __init mem_init(void)
 {
+	if (MACHINE_HAS_TLB_LC)
+		cpumask_set_cpu(0, &init_mm.context.cpu_attach_mask);
+	cpumask_set_cpu(0, mm_cpumask(&init_mm));
+	atomic_set(&init_mm.context.attach_count, 1);
+
         max_mapnr = max_low_pfn;
         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE);
 

arch/s390/mm/pgtable.c

@@ -200,7 +200,7 @@ static int gmap_unlink_segment(struct gmap *gmap, unsigned long *table)
 static void gmap_flush_tlb(struct gmap *gmap)
 {
 	if (MACHINE_HAS_IDTE)
-		__tlb_flush_idte((unsigned long) gmap->table |
+		__tlb_flush_asce(gmap->mm, (unsigned long) gmap->table |
 				 _ASCE_TYPE_REGION1);
 	else
 		__tlb_flush_global();
@@ -219,7 +219,7 @@ void gmap_free(struct gmap *gmap)
 
 	/* Flush tlb. */
 	if (MACHINE_HAS_IDTE)
-		__tlb_flush_idte((unsigned long) gmap->table |
+		__tlb_flush_asce(gmap->mm, (unsigned long) gmap->table |
 				 _ASCE_TYPE_REGION1);
 	else
 		__tlb_flush_global();

arch/s390/mm/vmem.c

@@ -138,7 +138,6 @@ static int vmem_add_mem(unsigned long start, unsigned long size, int ro)
 	}
 	ret = 0;
 out:
-	flush_tlb_kernel_range(start, end);
 	return ret;
 }
 
@@ -265,7 +264,6 @@ int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 	memset((void *)start, 0, end - start);
 	ret = 0;
 out:
-	flush_tlb_kernel_range(start, end);
 	return ret;
 }