* also clear mm->cpu_vm_mask bits when processes are migrated
*/
-#undef DEBUG
-#define DEBUG_STEAL_ONLY
-#undef DEBUG_MAP_CONSISTENCY
+//#define DEBUG_MAP_CONSISTENCY
+//#define DEBUG_CLAMP_LAST_CONTEXT 31
+//#define DEBUG_HARDER
+
+/* We don't use DEBUG because it tends to be compiled in always nowadays
+ * and this would generate way too much output
+ */
+#ifdef DEBUG_HARDER
+#define pr_hard(args...) printk(KERN_DEBUG args)
+#define pr_hardcont(args...) printk(KERN_CONT args)
+#else
+#define pr_hard(args...) do { } while(0)
+#define pr_hardcont(args...) do { } while(0)
+#endif
#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/init.h>
+#include <linux/spinlock.h>
+#include <linux/bootmem.h>
+#include <linux/notifier.h>
+#include <linux/cpu.h>
+#include <linux/slab.h>
#include <asm/mmu_context.h>
#include <asm/tlbflush.h>
-#include <linux/spinlock.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 LAST_CONTEXT 15
-#define FIRST_CONTEXT 0
-
-#elif defined(CONFIG_4xx)
-#define LAST_CONTEXT 255
-#define FIRST_CONTEXT 1
-#elif defined(CONFIG_E200) || defined(CONFIG_E500)
-#define LAST_CONTEXT 255
-#define FIRST_CONTEXT 1
+static unsigned int first_context, last_context;
+static unsigned int next_context, nr_free_contexts;
+static unsigned long *context_map;
+static unsigned long *stale_map[NR_CPUS];
+static struct mm_struct **context_mm;
+static DEFINE_RAW_SPINLOCK(context_lock);
-#else
-#error Unsupported processor type
-#endif
+#define CTX_MAP_SIZE \
+ (sizeof(unsigned long) * (last_context / BITS_PER_LONG + 1))
-static unsigned int next_context, nr_free_contexts;
-static unsigned long context_map[LAST_CONTEXT / BITS_PER_LONG + 1];
-static unsigned long stale_map[NR_CPUS][LAST_CONTEXT / BITS_PER_LONG + 1];
-static struct mm_struct *context_mm[LAST_CONTEXT+1];
-static spinlock_t context_lock = SPIN_LOCK_UNLOCKED;
/* Steal a context from a task that has one at the moment.
*
static unsigned int steal_context_smp(unsigned int id)
{
struct mm_struct *mm;
- unsigned int cpu, max;
+ unsigned int cpu, max, i;
- again:
- max = LAST_CONTEXT - FIRST_CONTEXT;
+ max = last_context - first_context;
/* Attempt to free next_context first and then loop until we manage */
while (max--) {
*/
if (mm->context.active) {
id++;
- if (id > LAST_CONTEXT)
- id = FIRST_CONTEXT;
+ if (id > last_context)
+ id = first_context;
continue;
}
- pr_debug("[%d] steal context %d from mm @%p\n",
- smp_processor_id(), id, mm);
+ pr_hardcont(" | steal %d from 0x%p", id, mm);
/* Mark this mm has having no context anymore */
mm->context.id = MMU_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]);
+ /* Mark it stale on all CPUs that used this mm. For threaded
+ * implementations, we set it on all threads on each core
+ * represented in the mask. A future implementation will use
+ * a core map instead but this will do for now.
+ */
+ for_each_cpu(cpu, mm_cpumask(mm)) {
+ for (i = cpu_first_thread_in_core(cpu);
+ i <= cpu_last_thread_in_core(cpu); i++)
+ __set_bit(id, stale_map[i]);
+ cpu = i - 1;
+ }
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);
+ raw_spin_unlock(&context_lock);
cpu_relax();
- spin_lock(&context_lock);
- goto again;
+ raw_spin_lock(&context_lock);
+
+ /* This will cause the caller to try again */
+ return MMU_NO_CONTEXT;
}
#endif /* CONFIG_SMP */
/* 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 = MMU_NO_CONTEXT;
+ pr_hardcont(" | steal %d from 0x%p", id, mm);
/* Flush the TLB for that context */
local_flush_tlb_mm(mm);
+ /* Mark this mm has having no context anymore */
+ mm->context.id = MMU_NO_CONTEXT;
+
/* XXX This clear should ultimately be part of local_flush_tlb_mm */
__clear_bit(id, stale_map[cpu]);
unsigned int id, nrf, nact;
nrf = nact = 0;
- for (id = FIRST_CONTEXT; id <= LAST_CONTEXT; id++) {
+ for (id = first_context; id <= last_context; id++) {
int used = test_bit(id, context_map);
if (!used)
nrf++;
if (nact > num_online_cpus())
pr_err("MMU: More active contexts than CPUs ! (%d vs %d)\n",
nact, num_online_cpus());
+ if (first_context > 0 && !test_bit(0, context_map))
+ pr_err("MMU: Context 0 has been freed !!!\n");
}
#else
static void context_check_map(void) { }
void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
{
- unsigned int id, cpu = smp_processor_id();
+ unsigned int i, id, cpu = smp_processor_id();
unsigned long *map;
/* No lockless fast path .. yet */
- spin_lock(&context_lock);
+ raw_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
+ pr_hard("[%d] activating context for mm @%p, active=%d, id=%d",
+ cpu, next, next->context.active, next->context.id);
#ifdef CONFIG_SMP
/* Mark us active and the previous one not anymore */
next->context.active++;
if (prev) {
+ pr_hardcont(" (old=0x%p a=%d)", prev, prev->context.active);
WARN_ON(prev->context.active < 1);
prev->context.active--;
}
+
+ again:
#endif /* CONFIG_SMP */
/* If we already have a valid assigned context, skip all that */
id = next->context.id;
- if (likely(id != MMU_NO_CONTEXT))
+ if (likely(id != MMU_NO_CONTEXT)) {
+#ifdef DEBUG_MAP_CONSISTENCY
+ if (context_mm[id] != next)
+ pr_err("MMU: mm 0x%p has id %d but context_mm[%d] says 0x%p\n",
+ next, id, id, context_mm[id]);
+#endif
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;
+ if (id > last_context)
+ id = first_context;
map = context_map;
/* No more free contexts, let's try to steal one */
#ifdef CONFIG_SMP
if (num_online_cpus() > 1) {
id = steal_context_smp(id);
+ if (id == MMU_NO_CONTEXT)
+ goto again;
goto stolen;
}
#endif /* CONFIG_SMP */
/* 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;
+ 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
+ pr_hardcont(" | new id=%d,nrf=%d", id, nr_free_contexts);
context_check_map();
ctxt_ok:
* 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);
+ pr_hardcont(" | stale flush %d [%d..%d]",
+ id, cpu_first_thread_in_core(cpu),
+ cpu_last_thread_in_core(cpu));
+
local_flush_tlb_mm(next);
/* XXX This clear should ultimately be part of local_flush_tlb_mm */
- __clear_bit(id, stale_map[cpu]);
+ for (i = cpu_first_thread_in_core(cpu);
+ i <= cpu_last_thread_in_core(cpu); i++) {
+ __clear_bit(id, stale_map[i]);
+ }
}
/* Flick the MMU and release lock */
+ pr_hardcont(" -> %d\n", id);
set_context(id, next->pgd);
- spin_unlock(&context_lock);
+ raw_spin_unlock(&context_lock);
}
/*
*/
int init_new_context(struct task_struct *t, struct mm_struct *mm)
{
+ pr_hard("initing context for mm @%p\n", mm);
+
mm->context.id = MMU_NO_CONTEXT;
mm->context.active = 0;
*/
void destroy_context(struct mm_struct *mm)
{
+ unsigned long flags;
unsigned int id;
if (mm->context.id == MMU_NO_CONTEXT)
WARN_ON(mm->context.active != 0);
- spin_lock(&context_lock);
+ raw_spin_lock_irqsave(&context_lock, flags);
id = mm->context.id;
if (id != MMU_NO_CONTEXT) {
__clear_bit(id, context_map);
mm->context.id = MMU_NO_CONTEXT;
#ifdef DEBUG_MAP_CONSISTENCY
mm->context.active = 0;
- context_mm[id] = NULL;
#endif
+ context_mm[id] = NULL;
nr_free_contexts++;
}
- spin_unlock(&context_lock);
+ raw_spin_unlock_irqrestore(&context_lock, flags);
}
+#ifdef CONFIG_SMP
+
+static int __cpuinit mmu_context_cpu_notify(struct notifier_block *self,
+ unsigned long action, void *hcpu)
+{
+ unsigned int cpu = (unsigned int)(long)hcpu;
+#ifdef CONFIG_HOTPLUG_CPU
+ struct task_struct *p;
+#endif
+ /* We don't touch CPU 0 map, it's allocated at aboot and kept
+ * around forever
+ */
+ if (cpu == 0)
+ return NOTIFY_OK;
+
+ switch (action) {
+ case CPU_ONLINE:
+ case CPU_ONLINE_FROZEN:
+ pr_devel("MMU: Allocating stale context map for CPU %d\n", cpu);
+ stale_map[cpu] = kzalloc(CTX_MAP_SIZE, GFP_KERNEL);
+ break;
+#ifdef CONFIG_HOTPLUG_CPU
+ case CPU_DEAD:
+ case CPU_DEAD_FROZEN:
+ pr_devel("MMU: Freeing stale context map for CPU %d\n", cpu);
+ kfree(stale_map[cpu]);
+ stale_map[cpu] = NULL;
+
+ /* We also clear the cpu_vm_mask bits of CPUs going away */
+ read_lock(&tasklist_lock);
+ for_each_process(p) {
+ if (p->mm)
+ cpumask_clear_cpu(cpu, mm_cpumask(p->mm));
+ }
+ read_unlock(&tasklist_lock);
+ break;
+#endif /* CONFIG_HOTPLUG_CPU */
+ }
+ return NOTIFY_OK;
+}
+
+static struct notifier_block __cpuinitdata mmu_context_cpu_nb = {
+ .notifier_call = mmu_context_cpu_notify,
+};
+
+#endif /* CONFIG_SMP */
/*
* Initialize the context management stuff.
init_mm.context.active = NR_CPUS;
/*
+ * 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
+ */
+ if (mmu_has_feature(MMU_FTR_TYPE_8xx)) {
+ first_context = 0;
+ last_context = 15;
+ } else {
+ first_context = 1;
+ last_context = 255;
+ }
+
+#ifdef DEBUG_CLAMP_LAST_CONTEXT
+ last_context = DEBUG_CLAMP_LAST_CONTEXT;
+#endif
+ /*
+ * Allocate the maps used by context management
+ */
+ context_map = alloc_bootmem(CTX_MAP_SIZE);
+ context_mm = alloc_bootmem(sizeof(void *) * (last_context + 1));
+ stale_map[0] = alloc_bootmem(CTX_MAP_SIZE);
+
+#ifdef CONFIG_SMP
+ register_cpu_notifier(&mmu_context_cpu_nb);
+#endif
+
+ printk(KERN_INFO
+ "MMU: Allocated %zu bytes of context maps for %d contexts\n",
+ 2 * CTX_MAP_SIZE + (sizeof(void *) * (last_context + 1)),
+ last_context - first_context + 1);
+
+ /*
* 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.
+ * This code assumes first_context < 32.
*/
- context_map[0] = (1 << FIRST_CONTEXT) - 1;
- next_context = FIRST_CONTEXT;
- nr_free_contexts = LAST_CONTEXT - FIRST_CONTEXT + 1;
+ context_map[0] = (1 << first_context) - 1;
+ next_context = first_context;
+ nr_free_contexts = last_context - first_context + 1;
}