[PATCH 3/3] powerpc: Rework context management for CPUs with no hash table
Benjamin Herrenschmidt
benh at kernel.crashing.org
Thu Dec 4 16:41:53 EST 2008
This reworks the context management code used by 4xx,8xx and
freescale BookE. It adds support for SMP by implementing a
concept of stale context map to lazily flush the TLB on
processors where a context may have been invalidated. This
also contains the ground work for generalizing such lazy TLB
flushing by just picking up a new PID and marking the old one
stale. This will be implemented later.
This is a first implementation that uses a global spinlock.
Ideally, we should try to get at least the fast path (context ID
already assigned) lockless or limited to a per context lock,
but for now this will do.
I tried to keep the UP case reasonably simple to avoid adding
too much overhead to 8xx which does a lot of context stealing
since it effectively has only 16 PIDs available.
Signed-off-by: Benjamin Herrenschmidt <benh at kernel.crashing.org>
---
arch/powerpc/include/asm/mmu-40x.h | 5
arch/powerpc/include/asm/mmu-44x.h | 5
arch/powerpc/include/asm/mmu-fsl-booke.h | 5
arch/powerpc/mm/mmu_context_nohash.c | 258 ++++++++++++++++++++++++-------
4 files changed, 217 insertions(+), 56 deletions(-)
--- linux-work.orig/arch/powerpc/include/asm/mmu-40x.h 2008-12-04 15:02:19.000000000 +1100
+++ linux-work/arch/powerpc/include/asm/mmu-40x.h 2008-12-04 15:04:03.000000000 +1100
@@ -54,8 +54,9 @@
#ifndef __ASSEMBLY__
typedef struct {
- unsigned long id;
- unsigned long vdso_base;
+ unsigned short id;
+ unsigned short active;
+ unsigned long vdso_base;
} mm_context_t;
#endif /* !__ASSEMBLY__ */
Index: linux-work/arch/powerpc/include/asm/mmu-44x.h
===================================================================
--- linux-work.orig/arch/powerpc/include/asm/mmu-44x.h 2008-12-04 15:02:19.000000000 +1100
+++ linux-work/arch/powerpc/include/asm/mmu-44x.h 2008-12-04 15:04:03.000000000 +1100
@@ -56,8 +56,9 @@
extern unsigned int tlb_44x_hwater;
typedef struct {
- unsigned long id;
- unsigned long vdso_base;
+ unsigned short id;
+ unsigned short active;
+ unsigned long vdso_base;
} mm_context_t;
#endif /* !__ASSEMBLY__ */
Index: linux-work/arch/powerpc/include/asm/mmu-fsl-booke.h
===================================================================
--- linux-work.orig/arch/powerpc/include/asm/mmu-fsl-booke.h 2008-12-04 15:02:19.000000000 +1100
+++ linux-work/arch/powerpc/include/asm/mmu-fsl-booke.h 2008-12-04 15:04:03.000000000 +1100
@@ -74,8 +74,9 @@
#ifndef __ASSEMBLY__
typedef struct {
- unsigned long id;
- unsigned long vdso_base;
+ unsigned short id;
+ unsigned short active;
+ unsigned long vdso_base;
} mm_context_t;
#endif /* !__ASSEMBLY__ */
Index: linux-work/arch/powerpc/mm/mmu_context_nohash.c
===================================================================
--- linux-work.orig/arch/powerpc/mm/mmu_context_nohash.c 2008-12-04 15:04:03.000000000 +1100
+++ linux-work/arch/powerpc/mm/mmu_context_nohash.c 2008-12-04 15:18:46.000000000 +1100
@@ -1,32 +1,43 @@
/*
* 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 at linuxppc.org)
+ * PowerPC implementations where the MMU is not using the hash
+ * table, such as 8xx, 4xx, BookE's etc...
*
- * Modifications by Paul Mackerras (PowerMac) (paulus at cs.anu.edu.au)
- * and Cort Dougan (PReP) (cort at cs.nmt.edu)
- * Copyright (C) 1996 Paul Mackerras
+ * Copyright 2008 Ben Herrenschmidt <benh at kernel.crashing.org>
+ * IBM Corp.
*
- * Derived from "arch/i386/mm/init.c"
- * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
+ * 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.
*
+ * TODO:
+ *
+ * - The global context lock will not scale very well
+ * - The maps should be dynamically allocated to allow for processors
+ * that support more PID bits at runtime
+ * - Implement flush_tlb_mm() by making the context stale and picking
+ * a new one
+ * - More aggressively clear stale map bits and maybe find some way to
+ * also clear mm->cpu_vm_mask bits when processes are migrated
+ *
*/
+#undef DEBUG
+#define DEBUG_STEAL_ONLY
+
+#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/init.h>
#include <asm/mmu_context.h>
#include <asm/tlbflush.h>
+#include <linux/spinlock.h>
+
+#undef DEBUG_MAP_CONSISTENCY
/*
* The MPC8xx has only 16 contexts. We rotate through them on each
@@ -64,11 +75,11 @@
#error Unsupported processor type
#endif
-static unsigned long next_mmu_context;
+static unsigned int next_context, nr_free_contexts;
static unsigned long context_map[LAST_CONTEXT / BITS_PER_LONG + 1];
-static atomic_t nr_free_contexts;
+static unsigned long stale_map[NR_CPUS][LAST_CONTEXT / BITS_PER_LONG + 1];
static struct mm_struct *context_mm[LAST_CONTEXT+1];
-static void steal_context(void);
+static spinlock_t context_lock = SPIN_LOCK_UNLOCKED;
/* 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
@@ -81,49 +92,187 @@ static void steal_context(void);
* place to implement an LRU scheme if anyone was motivated to do it.
* -- paulus
*/
-static void steal_context(void)
+#include <asm/udbg.h>
+
+/*
+ * For context stealing, we use a slightly different approach for
+ * SMP and UP. Basically, the UP one is simpler and doesn't use
+ * the stale map as we can just flush the local CPU
+ */
+#ifdef CONFIG_SMP
+static unsigned int steal_context_smp(unsigned int id)
{
struct mm_struct *mm;
+ unsigned int cpu, max;
- /* 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);
-}
+ again:
+ max = LAST_CONTEXT - FIRST_CONTEXT;
+ /* Attempt to free next_context first and then loop until we manage */
+ while (max--) {
+ /* Pick up the victim mm */
+ mm = context_mm[id];
+
+ /* We have a candidate victim, check if it's active, on SMP
+ * we cannot steal active contexts
+ */
+ if (mm->context.active) {
+ id ++;
+ if (id > LAST_CONTEXT)
+ id = FIRST_CONTEXT;
+ continue;
+ }
+ pr_debug("[%d] steal context %d from mm @%p\n",
+ smp_processor_id(), id, mm);
-/*
- * Get a new mmu context for the address space described by `mm'.
+ /* Mark this mm has having no context anymore */
+ mm->context.id = NO_CONTEXT;
+
+ /* Mark it stale on all CPUs that used this mm */
+ for_each_cpu_mask_nr(cpu, mm->cpu_vm_mask)
+ __set_bit(id, stale_map[cpu]);
+ return id;
+ }
+
+ /* This will happen if you have more CPUs than available contexts,
+ * all we can do here is wait a bit and try again
+ */
+ spin_unlock(&context_lock);
+ cpu_relax();
+ spin_lock(&context_lock);
+ goto again;
+}
+#endif /* CONFIG_SMP */
+
+/* Note that this will also be called on SMP if all other CPUs are
+ * offlined, which means that it may be called for cpu != 0. For
+ * this to work, we somewhat assume that CPUs that are onlined
+ * come up with a fully clean TLB (or are cleaned when offlined)
*/
-static inline void get_mmu_context(struct mm_struct *mm)
+static unsigned int steal_context_up(unsigned int id)
{
- unsigned long ctx;
+ struct mm_struct *mm;
+ int cpu = smp_processor_id();
- if (mm->context.id != NO_CONTEXT)
- return;
+ /* Pick up the victim mm */
+ mm = context_mm[id];
+
+ pr_debug("[%d] steal context %d from mm @%p\n", cpu, id, mm);
+
+ /* Mark this mm has having no context anymore */
+ mm->context.id = NO_CONTEXT;
+
+ /* Flush the TLB for that context */
+ local_flush_tlb_mm(mm);
- while (atomic_dec_if_positive(&nr_free_contexts) < 0)
- steal_context();
+ /* XXX This clear should ultimately be part of local_flush_tlb_mm */
+ __clear_bit(id, stale_map[cpu]);
- 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;
+ return id;
}
+#ifdef DEBUG_MAP_CONSISTENCY
+static void context_check_map(void)
+{
+ unsigned int id, nrf;
+
+ nrf = 0;
+ for (id = FIRST_CONTEXT; id <= LAST_CONTEXT; id++) {
+ int used = test_bit(id, context_map);
+ if (!used)
+ nrf++;
+ if (used != (context_mm[id] != NULL))
+ pr_err("MMU: Context %d is %s and MM is %p !\n",
+ id, used ? "used" : "free", context_mm[id]);
+ }
+ if (nrf != nr_free_contexts) {
+ pr_err("MMU: Free context count out of sync ! (%d vs %d)\n",
+ nr_free_contexts, nrf);
+ nr_free_contexts = nrf;
+ }
+}
+#else
+static void context_check_map(void) { }
+#endif
+
void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
{
- get_mmu_context(next);
+ unsigned int id, cpu = smp_processor_id();
+ unsigned long *map;
+
+ /* No lockless fast path .. yet */
+ spin_lock(&context_lock);
+
+#ifndef DEBUG_STEAL_ONLY
+ pr_debug("[%d] activating context for mm @%p, active=%d, id=%d\n",
+ cpu, next, next->context.active, next->context.id);
+#endif
+
+#ifdef CONFIG_SMP
+ /* Mark us active and the previous one not anymore */
+ next->context.active++;
+ if (prev)
+ prev->context.active--;
+#endif /* CONFIG_SMP */
+
+ /* If we already have a valid assigned context, skip all that */
+ id = next->context.id;
+ if (likely(id != NO_CONTEXT))
+ goto ctxt_ok;
+
+ /* We really don't have a context, let's try to acquire one */
+ id = next_context;
+ if (id > LAST_CONTEXT)
+ id = FIRST_CONTEXT;
+ map = context_map;
+
+ /* No more free contexts, let's try to steal one */
+ if (nr_free_contexts == 0) {
+#ifdef CONFIF_SMP
+ if (num_online_cpus() > 1) {
+ id = steal_context_smp(id);
+ goto stolen;
+ }
+#endif /* CONFIG_SMP */
+ id = steal_context_up(id);
+ goto stolen;
+ }
+ nr_free_contexts--;
- set_context(next->context.id, next->pgd);
+ /* We know there's at least one free context, try to find it */
+ while (__test_and_set_bit(id, map)) {
+ id = find_next_zero_bit(map, LAST_CONTEXT+1, id);
+ if (id > LAST_CONTEXT)
+ id = FIRST_CONTEXT;
+ }
+ stolen:
+ next_context = id + 1;
+ context_mm[id] = next;
+ next->context.id = id;
+
+#ifndef DEBUG_STEAL_ONLY
+ pr_debug("[%d] picked up new id %d, nrf is now %d\n",
+ cpu, id, nr_free_contexts);
+#endif
+
+ context_check_map();
+ ctxt_ok:
+
+ /* If that context got marked stale on this CPU, then flush the
+ * local TLB for it and unmark it before we use it
+ */
+ if (test_bit(id, stale_map[cpu])) {
+ pr_debug("[%d] flushing stale context %d for mm @%p !\n",
+ cpu, id, next);
+ local_flush_tlb_mm(next);
+
+ /* XXX This clear should ultimately be part of local_flush_tlb_mm */
+ __clear_bit(id, stale_map[cpu]);
+ }
+
+ /* Flick the MMU and release lock */
+ set_context(id, next->pgd);
+ spin_unlock(&context_lock);
}
/*
@@ -140,13 +289,22 @@ int init_new_context(struct task_struct
*/
void destroy_context(struct mm_struct *mm)
{
- preempt_disable();
- if (mm->context.id != NO_CONTEXT) {
- clear_bit(mm->context.id, context_map);
+ unsigned int id;
+
+ if (mm->context.id == NO_CONTEXT)
+ return;
+
+ spin_lock(&context_lock);
+ id = mm->context.id;
+ if (id != NO_CONTEXT) {
+ __clear_bit(id, context_map);
mm->context.id = NO_CONTEXT;
- atomic_inc(&nr_free_contexts);
+#ifdef DEBUG_MAP_CONSISTENCY
+ context_mm[id] = NULL;
+#endif
+ nr_free_contexts++;
}
- preempt_enable();
+ spin_unlock(&context_lock);
}
@@ -162,7 +320,7 @@ void __init mmu_context_init(void)
* 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);
+ next_context = FIRST_CONTEXT;
+ nr_free_contexts = LAST_CONTEXT - FIRST_CONTEXT + 1;
}
More information about the Linuxppc-dev
mailing list