Subject: mm, rt: kmap_atomic scheduling
From: Peter Zijlstra <peterz@infradead.org>
Date: Thu, 28 Jul 2011 10:43:51 +0200

In fact, with migrate_disable() existing one could play games with
kmap_atomic. You could save/restore the kmap_atomic slots on context
switch (if there are any in use of course), this should be esp easy now
that we have a kmap_atomic stack.

Something like the below.. it wants replacing all the preempt_disable()
stuff with pagefault_disable() && migrate_disable() of course, but then
you can flip kmaps around like below.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
[dvhart@linux.intel.com: build fix]
Link: http://lkml.kernel.org/r/1311842631.5890.208.camel@twins

[tglx@linutronix.de: Get rid of the per cpu variable and store the idx
		     and the pte content right away in the task struct.
		     Shortens the context switch code. ]
---
 arch/x86/kernel/process_32.c |   32 ++++++++++++++++++++++++++++++++
 arch/x86/mm/highmem_32.c     |   13 ++++++++++---
 arch/x86/mm/iomap_32.c       |    9 ++++++++-
 include/linux/highmem.h      |   27 +++++++++++++++++++++++----
 include/linux/sched.h        |    7 +++++++
 include/linux/uaccess.h      |    2 ++
 mm/highmem.c                 |    6 ++++--
 7 files changed, 86 insertions(+), 10 deletions(-)

--- a/arch/x86/kernel/process_32.c
+++ b/arch/x86/kernel/process_32.c
@@ -35,6 +35,7 @@
 #include <linux/uaccess.h>
 #include <linux/io.h>
 #include <linux/kdebug.h>
+#include <linux/highmem.h>
 
 #include <asm/pgtable.h>
 #include <asm/ldt.h>
@@ -210,6 +211,35 @@ start_thread(struct pt_regs *regs, unsig
 }
 EXPORT_SYMBOL_GPL(start_thread);
 
+#ifdef CONFIG_PREEMPT_RT_FULL
+static void switch_kmaps(struct task_struct *prev_p, struct task_struct *next_p)
+{
+	int i;
+
+	/*
+	 * Clear @prev's kmap_atomic mappings
+	 */
+	for (i = 0; i < prev_p->kmap_idx; i++) {
+		int idx = i + KM_TYPE_NR * smp_processor_id();
+		pte_t *ptep = kmap_pte - idx;
+
+		kpte_clear_flush(ptep, __fix_to_virt(FIX_KMAP_BEGIN + idx));
+	}
+	/*
+	 * Restore @next_p's kmap_atomic mappings
+	 */
+	for (i = 0; i < next_p->kmap_idx; i++) {
+		int idx = i + KM_TYPE_NR * smp_processor_id();
+
+		if (!pte_none(next_p->kmap_pte[i]))
+			set_pte(kmap_pte - idx, next_p->kmap_pte[i]);
+	}
+}
+#else
+static inline void
+switch_kmaps(struct task_struct *prev_p, struct task_struct *next_p) { }
+#endif
+
 
 /*
  *	switch_to(x,y) should switch tasks from x to y.
@@ -286,6 +316,8 @@ EXPORT_SYMBOL_GPL(start_thread);
 		     task_thread_info(next_p)->flags & _TIF_WORK_CTXSW_NEXT))
 		__switch_to_xtra(prev_p, next_p, tss);
 
+	switch_kmaps(prev_p, next_p);
+
 	/*
 	 * Leave lazy mode, flushing any hypercalls made here.
 	 * This must be done before restoring TLS segments so
--- a/arch/x86/mm/highmem_32.c
+++ b/arch/x86/mm/highmem_32.c
@@ -32,10 +32,11 @@ EXPORT_SYMBOL(kunmap);
  */
 void *kmap_atomic_prot(struct page *page, pgprot_t prot)
 {
+	pte_t pte = mk_pte(page, prot);
 	unsigned long vaddr;
 	int idx, type;
 
-	preempt_disable();
+	preempt_disable_nort();
 	pagefault_disable();
 
 	if (!PageHighMem(page))
@@ -45,7 +46,10 @@ void *kmap_atomic_prot(struct page *page
 	idx = type + KM_TYPE_NR*smp_processor_id();
 	vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
 	BUG_ON(!pte_none(*(kmap_pte-idx)));
-	set_pte(kmap_pte-idx, mk_pte(page, prot));
+#ifdef CONFIG_PREEMPT_RT_FULL
+	current->kmap_pte[type] = pte;
+#endif
+	set_pte(kmap_pte-idx, pte);
 	arch_flush_lazy_mmu_mode();
 
 	return (void *)vaddr;
@@ -88,6 +92,9 @@ void __kunmap_atomic(void *kvaddr)
 		 * is a bad idea also, in case the page changes cacheability
 		 * attributes or becomes a protected page in a hypervisor.
 		 */
+#ifdef CONFIG_PREEMPT_RT_FULL
+		current->kmap_pte[type] = __pte(0);
+#endif
 		kpte_clear_flush(kmap_pte-idx, vaddr);
 		kmap_atomic_idx_pop();
 		arch_flush_lazy_mmu_mode();
@@ -100,7 +107,7 @@ void __kunmap_atomic(void *kvaddr)
 #endif
 
 	pagefault_enable();
-	preempt_enable();
+	preempt_enable_nort();
 }
 EXPORT_SYMBOL(__kunmap_atomic);
 
--- a/arch/x86/mm/iomap_32.c
+++ b/arch/x86/mm/iomap_32.c
@@ -56,6 +56,7 @@ EXPORT_SYMBOL_GPL(iomap_free);
 
 void *kmap_atomic_prot_pfn(unsigned long pfn, pgprot_t prot)
 {
+	pte_t pte = pfn_pte(pfn, prot);
 	unsigned long vaddr;
 	int idx, type;
 
@@ -65,7 +66,10 @@ void *kmap_atomic_prot_pfn(unsigned long
 	type = kmap_atomic_idx_push();
 	idx = type + KM_TYPE_NR * smp_processor_id();
 	vaddr = __fix_to_virt(FIX_KMAP_BEGIN + idx);
-	set_pte(kmap_pte - idx, pfn_pte(pfn, prot));
+#ifdef CONFIG_PREEMPT_RT_FULL
+	current->kmap_pte[type] = pte;
+#endif
+	set_pte(kmap_pte - idx, pte);
 	arch_flush_lazy_mmu_mode();
 
 	return (void *)vaddr;
@@ -113,6 +117,9 @@ iounmap_atomic(void __iomem *kvaddr)
 		 * is a bad idea also, in case the page changes cacheability
 		 * attributes or becomes a protected page in a hypervisor.
 		 */
+#ifdef CONFIG_PREEMPT_RT_FULL
+		current->kmap_pte[type] = __pte(0);
+#endif
 		kpte_clear_flush(kmap_pte-idx, vaddr);
 		kmap_atomic_idx_pop();
 	}
--- a/include/linux/highmem.h
+++ b/include/linux/highmem.h
@@ -86,32 +86,51 @@ static inline void __kunmap_atomic(void
 
 #if defined(CONFIG_HIGHMEM) || defined(CONFIG_X86_32)
 
+#ifndef CONFIG_PREEMPT_RT_FULL
 DECLARE_PER_CPU(int, __kmap_atomic_idx);
+#endif
 
 static inline int kmap_atomic_idx_push(void)
 {
+#ifndef CONFIG_PREEMPT_RT_FULL
 	int idx = __this_cpu_inc_return(__kmap_atomic_idx) - 1;
 
-#ifdef CONFIG_DEBUG_HIGHMEM
+# ifdef CONFIG_DEBUG_HIGHMEM
 	WARN_ON_ONCE(in_irq() && !irqs_disabled());
 	BUG_ON(idx >= KM_TYPE_NR);
-#endif
+# endif
 	return idx;
+#else
+	current->kmap_idx++;
+	BUG_ON(current->kmap_idx > KM_TYPE_NR);
+	return current->kmap_idx - 1;
+#endif
 }
 
 static inline int kmap_atomic_idx(void)
 {
+#ifndef CONFIG_PREEMPT_RT_FULL
 	return __this_cpu_read(__kmap_atomic_idx) - 1;
+#else
+	return current->kmap_idx - 1;
+#endif
 }
 
 static inline void kmap_atomic_idx_pop(void)
 {
-#ifdef CONFIG_DEBUG_HIGHMEM
+#ifndef CONFIG_PREEMPT_RT_FULL
+# ifdef CONFIG_DEBUG_HIGHMEM
 	int idx = __this_cpu_dec_return(__kmap_atomic_idx);
 
 	BUG_ON(idx < 0);
-#else
+# else
 	__this_cpu_dec(__kmap_atomic_idx);
+# endif
+#else
+	current->kmap_idx--;
+# ifdef CONFIG_DEBUG_HIGHMEM
+	BUG_ON(current->kmap_idx < 0);
+# endif
 #endif
 }
 
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -26,6 +26,7 @@ struct sched_param {
 #include <linux/nodemask.h>
 #include <linux/mm_types.h>
 #include <linux/preempt.h>
+#include <asm/kmap_types.h>
 
 #include <asm/page.h>
 #include <asm/ptrace.h>
@@ -1954,6 +1955,12 @@ struct task_struct {
 	int softirq_nestcnt;
 	unsigned int softirqs_raised;
 #endif
+#ifdef CONFIG_PREEMPT_RT_FULL
+# if defined CONFIG_HIGHMEM || defined CONFIG_X86_32
+	int kmap_idx;
+	pte_t kmap_pte[KM_TYPE_NR];
+# endif
+#endif
 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
 	unsigned long	task_state_change;
 #endif
--- a/include/linux/uaccess.h
+++ b/include/linux/uaccess.h
@@ -24,6 +24,7 @@ static __always_inline void pagefault_di
  */
 static inline void pagefault_disable(void)
 {
+	migrate_disable();
 	pagefault_disabled_inc();
 	/*
 	 * make sure to have issued the store before a pagefault
@@ -40,6 +41,7 @@ static inline void pagefault_enable(void
 	 */
 	barrier();
 	pagefault_disabled_dec();
+	migrate_enable();
 }
 
 /*
--- a/mm/highmem.c
+++ b/mm/highmem.c
@@ -29,10 +29,11 @@
 #include <linux/kgdb.h>
 #include <asm/tlbflush.h>
 
-
+#ifndef CONFIG_PREEMPT_RT_FULL
 #if defined(CONFIG_HIGHMEM) || defined(CONFIG_X86_32)
 DEFINE_PER_CPU(int, __kmap_atomic_idx);
 #endif
+#endif
 
 /*
  * Virtual_count is not a pure "count".
@@ -107,8 +108,9 @@ static inline wait_queue_head_t *get_pkm
 unsigned long totalhigh_pages __read_mostly;
 EXPORT_SYMBOL(totalhigh_pages);
 
-
+#ifndef CONFIG_PREEMPT_RT_FULL
 EXPORT_PER_CPU_SYMBOL(__kmap_atomic_idx);
+#endif
 
 unsigned int nr_free_highpages (void)
 {