powerpc/mm: Split mmu_context handling

This splits the mmu_context handling between 32-bit hash based
processors, 64-bit hash based processors and everybody else.  This is
preliminary work for adding SMP support for BookE processors.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Acked-by: Kumar Gala <galak@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>

5 files changed, 277 insertions, 87 deletions

diff --git a/arch/powerpc/mm/Makefile b/arch/powerpc/mm/Makefile
index 148de35c9eef..923bd3fa7d64 100644
--- a/arch/powerpc/mm/Makefile
+++ b/arch/powerpc/mm/Makefile
@@ -8,15 +8,16 @@ endif
 
 obj-y                           := fault.o mem.o pgtable.o \
                                    init_$(CONFIG_WORD_SIZE).o \
-                                   pgtable_$(CONFIG_WORD_SIZE).o \
-                                   mmu_context_$(CONFIG_WORD_SIZE).o
+                                   pgtable_$(CONFIG_WORD_SIZE).o
+obj-$(CONFIG_PPC_MMU_NOHASH)    += mmu_context_nohash.o
 hash-$(CONFIG_PPC_NATIVE)       := hash_native_64.o
 obj-$(CONFIG_PPC64)             += hash_utils_64.o \
                                    slb_low.o slb.o stab.o \
                                    gup.o mmap.o $(hash-y)
 obj-$(CONFIG_PPC_STD_MMU_32)    += ppc_mmu_32.o
 obj-$(CONFIG_PPC_STD_MMU)       += hash_low_$(CONFIG_WORD_SIZE).o \
-                                   tlb_hash$(CONFIG_WORD_SIZE).o
+                                   tlb_hash$(CONFIG_WORD_SIZE).o \
+                                   mmu_context_hash$(CONFIG_WORD_SIZE).o
 obj-$(CONFIG_40x)               += 40x_mmu.o
 obj-$(CONFIG_44x)               += 44x_mmu.o
 obj-$(CONFIG_FSL_BOOKE)         += fsl_booke_mmu.o
diff --git a/arch/powerpc/mm/mmu_context_32.c b/arch/powerpc/mm/mmu_context_32.c
deleted file mode 100644
index cc32ba41d900..000000000000
--- a/arch/powerpc/mm/mmu_context_32.c
+++ /dev/null
@@ -1,84 +0,0 @@
-/*
- * This file contains the routines for handling the MMU on those
- * PowerPC implementations where the MMU substantially follows the
- * architecture specification.  This includes the 6xx, 7xx, 7xxx,
- * 8260, and POWER3 implementations but excludes the 8xx and 4xx.
- *  -- paulus
- *
- *  Derived from arch/ppc/mm/init.c:
- *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
- *
- *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
- *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
- *    Copyright (C) 1996 Paul Mackerras
- *
- *  Derived from "arch/i386/mm/init.c"
- *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
- *
- *  This program is free software; you can redistribute it and/or
- *  modify it under the terms of the GNU General Public License
- *  as published by the Free Software Foundation; either version
- *  2 of the License, or (at your option) any later version.
- *
- */
-
-#include <linux/mm.h>
-#include <linux/init.h>
-
-#include <asm/mmu_context.h>
-#include <asm/tlbflush.h>
-
-unsigned long next_mmu_context;
-unsigned long context_map[LAST_CONTEXT / BITS_PER_LONG + 1];
-#ifdef FEW_CONTEXTS
-atomic_t nr_free_contexts;
-struct mm_struct *context_mm[LAST_CONTEXT+1];
-void steal_context(void);
-#endif /* FEW_CONTEXTS */
-
-/*
- * Initialize the context management stuff.
- */
-void __init
-mmu_context_init(void)
-{
-        /*
-         * Some processors have too few contexts to reserve one for
-         * init_mm, and require using context 0 for a normal task.
-         * Other processors reserve the use of context zero for the kernel.
-         * This code assumes FIRST_CONTEXT < 32.
-         */
-        context_map[0] = (1 << FIRST_CONTEXT) - 1;
-        next_mmu_context = FIRST_CONTEXT;
-#ifdef FEW_CONTEXTS
-        atomic_set(&nr_free_contexts, LAST_CONTEXT - FIRST_CONTEXT + 1);
-#endif /* FEW_CONTEXTS */
-}
-
-#ifdef FEW_CONTEXTS
-/*
- * Steal a context from a task that has one at the moment.
- * This is only used on 8xx and 4xx and we presently assume that
- * they don't do SMP.  If they do then this will have to check
- * whether the MM we steal is in use.
- * We also assume that this is only used on systems that don't
- * use an MMU hash table - this is true for 8xx and 4xx.
- * This isn't an LRU system, it just frees up each context in
- * turn (sort-of pseudo-random replacement :).  This would be the
- * place to implement an LRU scheme if anyone was motivated to do it.
- *  -- paulus
- */
-void
-steal_context(void)
-{
-        struct mm_struct *mm;
-
-        /* free up context `next_mmu_context' */
-        /* if we shouldn't free context 0, don't... */
-        if (next_mmu_context < FIRST_CONTEXT)
-                next_mmu_context = FIRST_CONTEXT;
-        mm = context_mm[next_mmu_context];
-        flush_tlb_mm(mm);
-        destroy_context(mm);
-}
-#endif /* FEW_CONTEXTS */
diff --git a/arch/powerpc/mm/mmu_context_hash32.c b/arch/powerpc/mm/mmu_context_hash32.c
new file mode 100644
index 000000000000..0dfba2bf7f31
--- /dev/null
+++ b/arch/powerpc/mm/mmu_context_hash32.c
@@ -0,0 +1,103 @@
+/*
+ * This file contains the routines for handling the MMU on those
+ * PowerPC implementations where the MMU substantially follows the
+ * architecture specification.  This includes the 6xx, 7xx, 7xxx,
+ * 8260, and POWER3 implementations but excludes the 8xx and 4xx.
+ *  -- paulus
+ *
+ *  Derived from arch/ppc/mm/init.c:
+ *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
+ *
+ *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
+ *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
+ *    Copyright (C) 1996 Paul Mackerras
+ *
+ *  Derived from "arch/i386/mm/init.c"
+ *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
+ *
+ *  This program is free software; you can redistribute it and/or
+ *  modify it under the terms of the GNU General Public License
+ *  as published by the Free Software Foundation; either version
+ *  2 of the License, or (at your option) any later version.
+ *
+ */
+
+#include <linux/mm.h>
+#include <linux/init.h>
+
+#include <asm/mmu_context.h>
+#include <asm/tlbflush.h>
+
+/*
+ * On 32-bit PowerPC 6xx/7xx/7xxx CPUs, we use a set of 16 VSIDs
+ * (virtual segment identifiers) for each context.  Although the
+ * hardware supports 24-bit VSIDs, and thus >1 million contexts,
+ * we only use 32,768 of them.  That is ample, since there can be
+ * at most around 30,000 tasks in the system anyway, and it means
+ * that we can use a bitmap to indicate which contexts are in use.
+ * Using a bitmap means that we entirely avoid all of the problems
+ * that we used to have when the context number overflowed,
+ * particularly on SMP systems.
+ *  -- paulus.
+ */
+#define NO_CONTEXT              ((unsigned long) -1)
+#define LAST_CONTEXT            32767
+#define FIRST_CONTEXT           1
+
+/*
+ * This function defines the mapping from contexts to VSIDs (virtual
+ * segment IDs).  We use a skew on both the context and the high 4 bits
+ * of the 32-bit virtual address (the "effective segment ID") in order
+ * to spread out the entries in the MMU hash table.  Note, if this
+ * function is changed then arch/ppc/mm/hashtable.S will have to be
+ * changed to correspond.
+ *
+ *
+ * CTX_TO_VSID(ctx, va) (((ctx) * (897 * 16) + ((va) >> 28) * 0x111) \
+ *                               & 0xffffff)
+ */
+
+static unsigned long next_mmu_context;
+static unsigned long context_map[LAST_CONTEXT / BITS_PER_LONG + 1];
+
+
+/*
+ * Set up the context for a new address space.
+ */
+int init_new_context(struct task_struct *t, struct mm_struct *mm)
+{
+        unsigned long ctx = next_mmu_context;
+
+        while (test_and_set_bit(ctx, context_map)) {
+                ctx = find_next_zero_bit(context_map, LAST_CONTEXT+1, ctx);
+                if (ctx > LAST_CONTEXT)
+                        ctx = 0;
+        }
+        next_mmu_context = (ctx + 1) & LAST_CONTEXT;
+        mm->context.id = ctx;
+
+        return 0;
+}
+
+/*
+ * We're finished using the context for an address space.
+ */
+void destroy_context(struct mm_struct *mm)
+{
+        preempt_disable();
+        if (mm->context.id != NO_CONTEXT) {
+                clear_bit(mm->context.id, context_map);
+                mm->context.id = NO_CONTEXT;
+        }
+        preempt_enable();
+}
+
+/*
+ * Initialize the context management stuff.
+ */
+void __init mmu_context_init(void)
+{
+        /* Reserve context 0 for kernel use */
+        context_map[0] = (1 << FIRST_CONTEXT) - 1;
+        next_mmu_context = FIRST_CONTEXT;
+}
diff --git a/arch/powerpc/mm/mmu_context_64.c b/arch/powerpc/mm/mmu_context_hash64.c
index 1db38ba1f544..dbeb86ac90cd 100644
--- a/arch/powerpc/mm/mmu_context_64.c
+++ b/arch/powerpc/mm/mmu_context_hash64.c
@@ -24,6 +24,14 @@
 static DEFINE_SPINLOCK(mmu_context_lock);
 static DEFINE_IDR(mmu_context_idr);
 
+/*
+ * The proto-VSID space has 2^35 - 1 segments available for user mappings.
+ * Each segment contains 2^28 bytes.  Each context maps 2^44 bytes,
+ * so we can support 2^19-1 contexts (19 == 35 + 28 - 44).
+ */
+#define NO_CONTEXT      0
+#define MAX_CONTEXT     ((1UL << 19) - 1)
+
 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 {
         int index;
diff --git a/arch/powerpc/mm/mmu_context_nohash.c b/arch/powerpc/mm/mmu_context_nohash.c
new file mode 100644
index 000000000000..00e02150abef
--- /dev/null
+++ b/arch/powerpc/mm/mmu_context_nohash.c
@@ -0,0 +1,162 @@
+/*
+ * This file contains the routines for handling the MMU on those
+ * PowerPC implementations where the MMU is not using the hash
+ * table, such as 8xx, 4xx, BookE's etc...
+ *
+ * Copyright 2008 Ben Herrenschmidt <benh@kernel.crashing.org>
+ *                IBM Corp.
+ *
+ *  Derived from previous arch/powerpc/mm/mmu_context.c
+ *  and arch/powerpc/include/asm/mmu_context.h
+ *
+ *  This program is free software; you can redistribute it and/or
+ *  modify it under the terms of the GNU General Public License
+ *  as published by the Free Software Foundation; either version
+ *  2 of the License, or (at your option) any later version.
+ *
+ */
+
+#include <linux/mm.h>
+#include <linux/init.h>
+
+#include <asm/mmu_context.h>
+#include <asm/tlbflush.h>
+
+/*
+ *   The MPC8xx has only 16 contexts.  We rotate through them on each
+ * task switch.  A better way would be to keep track of tasks that
+ * own contexts, and implement an LRU usage.  That way very active
+ * tasks don't always have to pay the TLB reload overhead.  The
+ * kernel pages are mapped shared, so the kernel can run on behalf
+ * of any task that makes a kernel entry.  Shared does not mean they
+ * are not protected, just that the ASID comparison is not performed.
+ *      -- Dan
+ *
+ * The IBM4xx has 256 contexts, so we can just rotate through these
+ * as a way of "switching" contexts.  If the TID of the TLB is zero,
+ * the PID/TID comparison is disabled, so we can use a TID of zero
+ * to represent all kernel pages as shared among all contexts.
+ *      -- Dan
+ */
+
+#ifdef CONFIG_8xx
+#define NO_CONTEXT              16
+#define LAST_CONTEXT            15
+#define FIRST_CONTEXT           0
+
+#elif defined(CONFIG_4xx)
+#define NO_CONTEXT              256
+#define LAST_CONTEXT            255
+#define FIRST_CONTEXT           1
+
+#elif defined(CONFIG_E200) || defined(CONFIG_E500)
+#define NO_CONTEXT              256
+#define LAST_CONTEXT            255
+#define FIRST_CONTEXT           1
+
+#else
+#error Unsupported processor type
+#endif
+
+static unsigned long next_mmu_context;
+static unsigned long context_map[LAST_CONTEXT / BITS_PER_LONG + 1];
+static atomic_t nr_free_contexts;
+static struct mm_struct *context_mm[LAST_CONTEXT+1];
+static void steal_context(void);
+
+/* Steal a context from a task that has one at the moment.
+ * This is only used on 8xx and 4xx and we presently assume that
+ * they don't do SMP.  If they do then this will have to check
+ * whether the MM we steal is in use.
+ * We also assume that this is only used on systems that don't
+ * use an MMU hash table - this is true for 8xx and 4xx.
+ * This isn't an LRU system, it just frees up each context in
+ * turn (sort-of pseudo-random replacement :).  This would be the
+ * place to implement an LRU scheme if anyone was motivated to do it.
+ *  -- paulus
+ */
+static void steal_context(void)
+{
+        struct mm_struct *mm;
+
+        /* free up context `next_mmu_context' */
+        /* if we shouldn't free context 0, don't... */
+        if (next_mmu_context < FIRST_CONTEXT)
+                next_mmu_context = FIRST_CONTEXT;
+        mm = context_mm[next_mmu_context];
+        flush_tlb_mm(mm);
+        destroy_context(mm);
+}
+
+
+/*
+ * Get a new mmu context for the address space described by `mm'.
+ */
+static inline void get_mmu_context(struct mm_struct *mm)
+{
+        unsigned long ctx;
+
+        if (mm->context.id != NO_CONTEXT)
+                return;
+
+        while (atomic_dec_if_positive(&nr_free_contexts) < 0)
+                steal_context();
+
+        ctx = next_mmu_context;
+        while (test_and_set_bit(ctx, context_map)) {
+                ctx = find_next_zero_bit(context_map, LAST_CONTEXT+1, ctx);
+                if (ctx > LAST_CONTEXT)
+                        ctx = 0;
+        }
+        next_mmu_context = (ctx + 1) & LAST_CONTEXT;
+        mm->context.id = ctx;
+        context_mm[ctx] = mm;
+}
+
+void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
+{
+        get_mmu_context(next);
+
+        set_context(next->context.id, next->pgd);
+}
+
+/*
+ * Set up the context for a new address space.
+ */
+int init_new_context(struct task_struct *t, struct mm_struct *mm)
+{
+        mm->context.id = NO_CONTEXT;
+        return 0;
+}
+
+/*
+ * We're finished using the context for an address space.
+ */
+void destroy_context(struct mm_struct *mm)
+{
+        preempt_disable();
+        if (mm->context.id != NO_CONTEXT) {
+                clear_bit(mm->context.id, context_map);
+                mm->context.id = NO_CONTEXT;
+                atomic_inc(&nr_free_contexts);
+        }
+        preempt_enable();
+}
+
+
+/*
+ * Initialize the context management stuff.
+ */
+void __init mmu_context_init(void)
+{
+        /*
+         * Some processors have too few contexts to reserve one for
+         * init_mm, and require using context 0 for a normal task.
+         * Other processors reserve the use of context zero for the kernel.
+         * This code assumes FIRST_CONTEXT < 32.
+         */
+        context_map[0] = (1 << FIRST_CONTEXT) - 1;
+        next_mmu_context = FIRST_CONTEXT;
+        atomic_set(&nr_free_contexts, LAST_CONTEXT - FIRST_CONTEXT + 1);
+}
+