linux/arch/powerpc/mm/mmu_context_nohash.c
<<
>>
Prefs
   1/*
   2 * This file contains the routines for handling the MMU on those
   3 * PowerPC implementations where the MMU is not using the hash
   4 * table, such as 8xx, 4xx, BookE's etc...
   5 *
   6 * Copyright 2008 Ben Herrenschmidt <benh@kernel.crashing.org>
   7 *                IBM Corp.
   8 *
   9 *  Derived from previous arch/powerpc/mm/mmu_context.c
  10 *  and arch/powerpc/include/asm/mmu_context.h
  11 *
  12 *  This program is free software; you can redistribute it and/or
  13 *  modify it under the terms of the GNU General Public License
  14 *  as published by the Free Software Foundation; either version
  15 *  2 of the License, or (at your option) any later version.
  16 *
  17 * TODO:
  18 *
  19 *   - The global context lock will not scale very well
  20 *   - The maps should be dynamically allocated to allow for processors
  21 *     that support more PID bits at runtime
  22 *   - Implement flush_tlb_mm() by making the context stale and picking
  23 *     a new one
  24 *   - More aggressively clear stale map bits and maybe find some way to
  25 *     also clear mm->cpu_vm_mask bits when processes are migrated
  26 */
  27
  28//#define DEBUG_MAP_CONSISTENCY
  29//#define DEBUG_CLAMP_LAST_CONTEXT   31
  30//#define DEBUG_HARDER
  31
  32/* We don't use DEBUG because it tends to be compiled in always nowadays
  33 * and this would generate way too much output
  34 */
  35#ifdef DEBUG_HARDER
  36#define pr_hard(args...)        printk(KERN_DEBUG args)
  37#define pr_hardcont(args...)    printk(KERN_CONT args)
  38#else
  39#define pr_hard(args...)        do { } while(0)
  40#define pr_hardcont(args...)    do { } while(0)
  41#endif
  42
  43#include <linux/kernel.h>
  44#include <linux/mm.h>
  45#include <linux/init.h>
  46#include <linux/spinlock.h>
  47#include <linux/bootmem.h>
  48#include <linux/notifier.h>
  49#include <linux/cpu.h>
  50#include <linux/slab.h>
  51
  52#include <asm/mmu_context.h>
  53#include <asm/tlbflush.h>
  54
  55static unsigned int first_context, last_context;
  56static unsigned int next_context, nr_free_contexts;
  57static unsigned long *context_map;
  58static unsigned long *stale_map[NR_CPUS];
  59static struct mm_struct **context_mm;
  60static DEFINE_RAW_SPINLOCK(context_lock);
  61
  62#define CTX_MAP_SIZE    \
  63        (sizeof(unsigned long) * (last_context / BITS_PER_LONG + 1))
  64
  65
  66/* Steal a context from a task that has one at the moment.
  67 *
  68 * This is used when we are running out of available PID numbers
  69 * on the processors.
  70 *
  71 * This isn't an LRU system, it just frees up each context in
  72 * turn (sort-of pseudo-random replacement :).  This would be the
  73 * place to implement an LRU scheme if anyone was motivated to do it.
  74 *  -- paulus
  75 *
  76 * For context stealing, we use a slightly different approach for
  77 * SMP and UP. Basically, the UP one is simpler and doesn't use
  78 * the stale map as we can just flush the local CPU
  79 *  -- benh
  80 */
  81#ifdef CONFIG_SMP
  82static unsigned int steal_context_smp(unsigned int id)
  83{
  84        struct mm_struct *mm;
  85        unsigned int cpu, max, i;
  86
  87        max = last_context - first_context;
  88
  89        /* Attempt to free next_context first and then loop until we manage */
  90        while (max--) {
  91                /* Pick up the victim mm */
  92                mm = context_mm[id];
  93
  94                /* We have a candidate victim, check if it's active, on SMP
  95                 * we cannot steal active contexts
  96                 */
  97                if (mm->context.active) {
  98                        id++;
  99                        if (id > last_context)
 100                                id = first_context;
 101                        continue;
 102                }
 103                pr_hardcont(" | steal %d from 0x%p", id, mm);
 104
 105                /* Mark this mm has having no context anymore */
 106                mm->context.id = MMU_NO_CONTEXT;
 107
 108                /* Mark it stale on all CPUs that used this mm. For threaded
 109                 * implementations, we set it on all threads on each core
 110                 * represented in the mask. A future implementation will use
 111                 * a core map instead but this will do for now.
 112                 */
 113                for_each_cpu(cpu, mm_cpumask(mm)) {
 114                        for (i = cpu_first_thread_sibling(cpu);
 115                             i <= cpu_last_thread_sibling(cpu); i++)
 116                                __set_bit(id, stale_map[i]);
 117                        cpu = i - 1;
 118                }
 119                return id;
 120        }
 121
 122        /* This will happen if you have more CPUs than available contexts,
 123         * all we can do here is wait a bit and try again
 124         */
 125        raw_spin_unlock(&context_lock);
 126        cpu_relax();
 127        raw_spin_lock(&context_lock);
 128
 129        /* This will cause the caller to try again */
 130        return MMU_NO_CONTEXT;
 131}
 132#endif  /* CONFIG_SMP */
 133
 134/* Note that this will also be called on SMP if all other CPUs are
 135 * offlined, which means that it may be called for cpu != 0. For
 136 * this to work, we somewhat assume that CPUs that are onlined
 137 * come up with a fully clean TLB (or are cleaned when offlined)
 138 */
 139static unsigned int steal_context_up(unsigned int id)
 140{
 141        struct mm_struct *mm;
 142        int cpu = smp_processor_id();
 143
 144        /* Pick up the victim mm */
 145        mm = context_mm[id];
 146
 147        pr_hardcont(" | steal %d from 0x%p", id, mm);
 148
 149        /* Flush the TLB for that context */
 150        local_flush_tlb_mm(mm);
 151
 152        /* Mark this mm has having no context anymore */
 153        mm->context.id = MMU_NO_CONTEXT;
 154
 155        /* XXX This clear should ultimately be part of local_flush_tlb_mm */
 156        __clear_bit(id, stale_map[cpu]);
 157
 158        return id;
 159}
 160
 161#ifdef DEBUG_MAP_CONSISTENCY
 162static void context_check_map(void)
 163{
 164        unsigned int id, nrf, nact;
 165
 166        nrf = nact = 0;
 167        for (id = first_context; id <= last_context; id++) {
 168                int used = test_bit(id, context_map);
 169                if (!used)
 170                        nrf++;
 171                if (used != (context_mm[id] != NULL))
 172                        pr_err("MMU: Context %d is %s and MM is %p !\n",
 173                               id, used ? "used" : "free", context_mm[id]);
 174                if (context_mm[id] != NULL)
 175                        nact += context_mm[id]->context.active;
 176        }
 177        if (nrf != nr_free_contexts) {
 178                pr_err("MMU: Free context count out of sync ! (%d vs %d)\n",
 179                       nr_free_contexts, nrf);
 180                nr_free_contexts = nrf;
 181        }
 182        if (nact > num_online_cpus())
 183                pr_err("MMU: More active contexts than CPUs ! (%d vs %d)\n",
 184                       nact, num_online_cpus());
 185        if (first_context > 0 && !test_bit(0, context_map))
 186                pr_err("MMU: Context 0 has been freed !!!\n");
 187}
 188#else
 189static void context_check_map(void) { }
 190#endif
 191
 192void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
 193{
 194        unsigned int i, id, cpu = smp_processor_id();
 195        unsigned long *map;
 196
 197        /* No lockless fast path .. yet */
 198        raw_spin_lock(&context_lock);
 199
 200        pr_hard("[%d] activating context for mm @%p, active=%d, id=%d",
 201                cpu, next, next->context.active, next->context.id);
 202
 203#ifdef CONFIG_SMP
 204        /* Mark us active and the previous one not anymore */
 205        next->context.active++;
 206        if (prev) {
 207                pr_hardcont(" (old=0x%p a=%d)", prev, prev->context.active);
 208                WARN_ON(prev->context.active < 1);
 209                prev->context.active--;
 210        }
 211
 212 again:
 213#endif /* CONFIG_SMP */
 214
 215        /* If we already have a valid assigned context, skip all that */
 216        id = next->context.id;
 217        if (likely(id != MMU_NO_CONTEXT)) {
 218#ifdef DEBUG_MAP_CONSISTENCY
 219                if (context_mm[id] != next)
 220                        pr_err("MMU: mm 0x%p has id %d but context_mm[%d] says 0x%p\n",
 221                               next, id, id, context_mm[id]);
 222#endif
 223                goto ctxt_ok;
 224        }
 225
 226        /* We really don't have a context, let's try to acquire one */
 227        id = next_context;
 228        if (id > last_context)
 229                id = first_context;
 230        map = context_map;
 231
 232        /* No more free contexts, let's try to steal one */
 233        if (nr_free_contexts == 0) {
 234#ifdef CONFIG_SMP
 235                if (num_online_cpus() > 1) {
 236                        id = steal_context_smp(id);
 237                        if (id == MMU_NO_CONTEXT)
 238                                goto again;
 239                        goto stolen;
 240                }
 241#endif /* CONFIG_SMP */
 242                id = steal_context_up(id);
 243                goto stolen;
 244        }
 245        nr_free_contexts--;
 246
 247        /* We know there's at least one free context, try to find it */
 248        while (__test_and_set_bit(id, map)) {
 249                id = find_next_zero_bit(map, last_context+1, id);
 250                if (id > last_context)
 251                        id = first_context;
 252        }
 253 stolen:
 254        next_context = id + 1;
 255        context_mm[id] = next;
 256        next->context.id = id;
 257        pr_hardcont(" | new id=%d,nrf=%d", id, nr_free_contexts);
 258
 259        context_check_map();
 260 ctxt_ok:
 261
 262        /* If that context got marked stale on this CPU, then flush the
 263         * local TLB for it and unmark it before we use it
 264         */
 265        if (test_bit(id, stale_map[cpu])) {
 266                pr_hardcont(" | stale flush %d [%d..%d]",
 267                            id, cpu_first_thread_sibling(cpu),
 268                            cpu_last_thread_sibling(cpu));
 269
 270                local_flush_tlb_mm(next);
 271
 272                /* XXX This clear should ultimately be part of local_flush_tlb_mm */
 273                for (i = cpu_first_thread_sibling(cpu);
 274                     i <= cpu_last_thread_sibling(cpu); i++) {
 275                        __clear_bit(id, stale_map[i]);
 276                }
 277        }
 278
 279        /* Flick the MMU and release lock */
 280        pr_hardcont(" -> %d\n", id);
 281        set_context(id, next->pgd);
 282        raw_spin_unlock(&context_lock);
 283}
 284
 285/*
 286 * Set up the context for a new address space.
 287 */
 288int init_new_context(struct task_struct *t, struct mm_struct *mm)
 289{
 290        pr_hard("initing context for mm @%p\n", mm);
 291
 292        mm->context.id = MMU_NO_CONTEXT;
 293        mm->context.active = 0;
 294
 295        return 0;
 296}
 297
 298/*
 299 * We're finished using the context for an address space.
 300 */
 301void destroy_context(struct mm_struct *mm)
 302{
 303        unsigned long flags;
 304        unsigned int id;
 305
 306        if (mm->context.id == MMU_NO_CONTEXT)
 307                return;
 308
 309        WARN_ON(mm->context.active != 0);
 310
 311        raw_spin_lock_irqsave(&context_lock, flags);
 312        id = mm->context.id;
 313        if (id != MMU_NO_CONTEXT) {
 314                __clear_bit(id, context_map);
 315                mm->context.id = MMU_NO_CONTEXT;
 316#ifdef DEBUG_MAP_CONSISTENCY
 317                mm->context.active = 0;
 318#endif
 319                context_mm[id] = NULL;
 320                nr_free_contexts++;
 321        }
 322        raw_spin_unlock_irqrestore(&context_lock, flags);
 323}
 324
 325#ifdef CONFIG_SMP
 326
 327static int __cpuinit mmu_context_cpu_notify(struct notifier_block *self,
 328                                            unsigned long action, void *hcpu)
 329{
 330        unsigned int cpu = (unsigned int)(long)hcpu;
 331#ifdef CONFIG_HOTPLUG_CPU
 332        struct task_struct *p;
 333#endif
 334        /* We don't touch CPU 0 map, it's allocated at aboot and kept
 335         * around forever
 336         */
 337        if (cpu == boot_cpuid)
 338                return NOTIFY_OK;
 339
 340        switch (action) {
 341        case CPU_ONLINE:
 342        case CPU_ONLINE_FROZEN:
 343                pr_devel("MMU: Allocating stale context map for CPU %d\n", cpu);
 344                stale_map[cpu] = kzalloc(CTX_MAP_SIZE, GFP_KERNEL);
 345                break;
 346#ifdef CONFIG_HOTPLUG_CPU
 347        case CPU_DEAD:
 348        case CPU_DEAD_FROZEN:
 349                pr_devel("MMU: Freeing stale context map for CPU %d\n", cpu);
 350                kfree(stale_map[cpu]);
 351                stale_map[cpu] = NULL;
 352
 353                /* We also clear the cpu_vm_mask bits of CPUs going away */
 354                read_lock(&tasklist_lock);
 355                for_each_process(p) {
 356                        if (p->mm)
 357                                cpumask_clear_cpu(cpu, mm_cpumask(p->mm));
 358                }
 359                read_unlock(&tasklist_lock);
 360        break;
 361#endif /* CONFIG_HOTPLUG_CPU */
 362        }
 363        return NOTIFY_OK;
 364}
 365
 366static struct notifier_block __cpuinitdata mmu_context_cpu_nb = {
 367        .notifier_call  = mmu_context_cpu_notify,
 368};
 369
 370#endif /* CONFIG_SMP */
 371
 372/*
 373 * Initialize the context management stuff.
 374 */
 375void __init mmu_context_init(void)
 376{
 377        /* Mark init_mm as being active on all possible CPUs since
 378         * we'll get called with prev == init_mm the first time
 379         * we schedule on a given CPU
 380         */
 381        init_mm.context.active = NR_CPUS;
 382
 383        /*
 384         *   The MPC8xx has only 16 contexts.  We rotate through them on each
 385         * task switch.  A better way would be to keep track of tasks that
 386         * own contexts, and implement an LRU usage.  That way very active
 387         * tasks don't always have to pay the TLB reload overhead.  The
 388         * kernel pages are mapped shared, so the kernel can run on behalf
 389         * of any task that makes a kernel entry.  Shared does not mean they
 390         * are not protected, just that the ASID comparison is not performed.
 391         *      -- Dan
 392         *
 393         * The IBM4xx has 256 contexts, so we can just rotate through these
 394         * as a way of "switching" contexts.  If the TID of the TLB is zero,
 395         * the PID/TID comparison is disabled, so we can use a TID of zero
 396         * to represent all kernel pages as shared among all contexts.
 397         *      -- Dan
 398         *
 399         * The IBM 47x core supports 16-bit PIDs, thus 65535 contexts. We
 400         * should normally never have to steal though the facility is
 401         * present if needed.
 402         *      -- BenH
 403         */
 404        if (mmu_has_feature(MMU_FTR_TYPE_8xx)) {
 405                first_context = 0;
 406                last_context = 15;
 407        } else if (mmu_has_feature(MMU_FTR_TYPE_47x)) {
 408                first_context = 1;
 409                last_context = 65535;
 410        } else {
 411                first_context = 1;
 412                last_context = 255;
 413        }
 414
 415#ifdef DEBUG_CLAMP_LAST_CONTEXT
 416        last_context = DEBUG_CLAMP_LAST_CONTEXT;
 417#endif
 418        /*
 419         * Allocate the maps used by context management
 420         */
 421        context_map = alloc_bootmem(CTX_MAP_SIZE);
 422        context_mm = alloc_bootmem(sizeof(void *) * (last_context + 1));
 423#ifndef CONFIG_SMP
 424        stale_map[0] = alloc_bootmem(CTX_MAP_SIZE);
 425#else
 426        stale_map[boot_cpuid] = alloc_bootmem(CTX_MAP_SIZE);
 427
 428        register_cpu_notifier(&mmu_context_cpu_nb);
 429#endif
 430
 431        printk(KERN_INFO
 432               "MMU: Allocated %zu bytes of context maps for %d contexts\n",
 433               2 * CTX_MAP_SIZE + (sizeof(void *) * (last_context + 1)),
 434               last_context - first_context + 1);
 435
 436        /*
 437         * Some processors have too few contexts to reserve one for
 438         * init_mm, and require using context 0 for a normal task.
 439         * Other processors reserve the use of context zero for the kernel.
 440         * This code assumes first_context < 32.
 441         */
 442        context_map[0] = (1 << first_context) - 1;
 443        next_context = first_context;
 444        nr_free_contexts = last_context - first_context + 1;
 445}
 446
 447