linux/mm/mmu_notifier.c
<<
>>
Prefs
   1/*
   2 *  linux/mm/mmu_notifier.c
   3 *
   4 *  Copyright (C) 2008  Qumranet, Inc.
   5 *  Copyright (C) 2008  SGI
   6 *             Christoph Lameter <clameter@sgi.com>
   7 *
   8 *  This work is licensed under the terms of the GNU GPL, version 2. See
   9 *  the COPYING file in the top-level directory.
  10 */
  11
  12#include <linux/rculist.h>
  13#include <linux/mmu_notifier.h>
  14#include <linux/export.h>
  15#include <linux/mm.h>
  16#include <linux/err.h>
  17#include <linux/srcu.h>
  18#include <linux/rcupdate.h>
  19#include <linux/sched.h>
  20#include <linux/slab.h>
  21
  22/* global SRCU for all MMs */
  23static struct srcu_struct srcu;
  24
  25/*
  26 * This function can't run concurrently against mmu_notifier_register
  27 * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
  28 * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
  29 * in parallel despite there being no task using this mm any more,
  30 * through the vmas outside of the exit_mmap context, such as with
  31 * vmtruncate. This serializes against mmu_notifier_unregister with
  32 * the mmu_notifier_mm->lock in addition to SRCU and it serializes
  33 * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
  34 * can't go away from under us as exit_mmap holds an mm_count pin
  35 * itself.
  36 */
  37void __mmu_notifier_release(struct mm_struct *mm)
  38{
  39        struct mmu_notifier *mn;
  40        int id;
  41
  42        /*
  43         * srcu_read_lock() here will block synchronize_srcu() in
  44         * mmu_notifier_unregister() until all registered
  45         * ->release() callouts this function makes have
  46         * returned.
  47         */
  48        id = srcu_read_lock(&srcu);
  49        spin_lock(&mm->mmu_notifier_mm->lock);
  50        while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
  51                mn = hlist_entry(mm->mmu_notifier_mm->list.first,
  52                                 struct mmu_notifier,
  53                                 hlist);
  54
  55                /*
  56                 * Unlink.  This will prevent mmu_notifier_unregister()
  57                 * from also making the ->release() callout.
  58                 */
  59                hlist_del_init_rcu(&mn->hlist);
  60                spin_unlock(&mm->mmu_notifier_mm->lock);
  61
  62                /*
  63                 * Clear sptes. (see 'release' description in mmu_notifier.h)
  64                 */
  65                if (mn->ops->release)
  66                        mn->ops->release(mn, mm);
  67
  68                spin_lock(&mm->mmu_notifier_mm->lock);
  69        }
  70        spin_unlock(&mm->mmu_notifier_mm->lock);
  71
  72        /*
  73         * All callouts to ->release() which we have done are complete.
  74         * Allow synchronize_srcu() in mmu_notifier_unregister() to complete
  75         */
  76        srcu_read_unlock(&srcu, id);
  77
  78        /*
  79         * mmu_notifier_unregister() may have unlinked a notifier and may
  80         * still be calling out to it.  Additionally, other notifiers
  81         * may have been active via vmtruncate() et. al. Block here
  82         * to ensure that all notifier callouts for this mm have been
  83         * completed and the sptes are really cleaned up before returning
  84         * to exit_mmap().
  85         */
  86        synchronize_srcu(&srcu);
  87}
  88
  89/*
  90 * If no young bitflag is supported by the hardware, ->clear_flush_young can
  91 * unmap the address and return 1 or 0 depending if the mapping previously
  92 * existed or not.
  93 */
  94int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
  95                                        unsigned long address)
  96{
  97        struct mmu_notifier *mn;
  98        struct hlist_node *n;
  99        int young = 0, id;
 100
 101        id = srcu_read_lock(&srcu);
 102        hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 103                if (mn->ops->clear_flush_young)
 104                        young |= mn->ops->clear_flush_young(mn, mm, address);
 105        }
 106        srcu_read_unlock(&srcu, id);
 107
 108        return young;
 109}
 110
 111int __mmu_notifier_test_young(struct mm_struct *mm,
 112                              unsigned long address)
 113{
 114        struct mmu_notifier *mn;
 115        struct hlist_node *n;
 116        int young = 0, id;
 117
 118        id = srcu_read_lock(&srcu);
 119        hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 120                if (mn->ops->test_young) {
 121                        young = mn->ops->test_young(mn, mm, address);
 122                        if (young)
 123                                break;
 124                }
 125        }
 126        srcu_read_unlock(&srcu, id);
 127
 128        return young;
 129}
 130
 131void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 132                               pte_t pte)
 133{
 134        struct mmu_notifier *mn;
 135        struct hlist_node *n;
 136        int id;
 137
 138        id = srcu_read_lock(&srcu);
 139        hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 140                if (mn->ops->change_pte)
 141                        mn->ops->change_pte(mn, mm, address, pte);
 142        }
 143        srcu_read_unlock(&srcu, id);
 144}
 145
 146void __mmu_notifier_invalidate_page(struct mm_struct *mm,
 147                                          unsigned long address)
 148{
 149        struct mmu_notifier *mn;
 150        struct hlist_node *n;
 151        int id;
 152
 153        id = srcu_read_lock(&srcu);
 154        hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 155                if (mn->ops->invalidate_page)
 156                        mn->ops->invalidate_page(mn, mm, address);
 157        }
 158        srcu_read_unlock(&srcu, id);
 159}
 160
 161void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 162                                  unsigned long start, unsigned long end)
 163{
 164        struct mmu_notifier *mn;
 165        struct hlist_node *n;
 166        int id;
 167
 168        id = srcu_read_lock(&srcu);
 169        hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 170                if (mn->ops->invalidate_range_start)
 171                        mn->ops->invalidate_range_start(mn, mm, start, end);
 172        }
 173        srcu_read_unlock(&srcu, id);
 174}
 175
 176void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 177                                  unsigned long start, unsigned long end)
 178{
 179        struct mmu_notifier *mn;
 180        struct hlist_node *n;
 181        int id;
 182
 183        id = srcu_read_lock(&srcu);
 184        hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 185                if (mn->ops->invalidate_range_end)
 186                        mn->ops->invalidate_range_end(mn, mm, start, end);
 187        }
 188        srcu_read_unlock(&srcu, id);
 189}
 190
 191static int do_mmu_notifier_register(struct mmu_notifier *mn,
 192                                    struct mm_struct *mm,
 193                                    int take_mmap_sem)
 194{
 195        struct mmu_notifier_mm *mmu_notifier_mm;
 196        int ret;
 197
 198        BUG_ON(atomic_read(&mm->mm_users) <= 0);
 199
 200        /*
 201         * Verify that mmu_notifier_init() already run and the global srcu is
 202         * initialized.
 203         */
 204        BUG_ON(!srcu.per_cpu_ref);
 205
 206        ret = -ENOMEM;
 207        mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 208        if (unlikely(!mmu_notifier_mm))
 209                goto out;
 210
 211        if (take_mmap_sem)
 212                down_write(&mm->mmap_sem);
 213        ret = mm_take_all_locks(mm);
 214        if (unlikely(ret))
 215                goto out_clean;
 216
 217        if (!mm_has_notifiers(mm)) {
 218                INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 219                spin_lock_init(&mmu_notifier_mm->lock);
 220
 221                mm->mmu_notifier_mm = mmu_notifier_mm;
 222                mmu_notifier_mm = NULL;
 223        }
 224        atomic_inc(&mm->mm_count);
 225
 226        /*
 227         * Serialize the update against mmu_notifier_unregister. A
 228         * side note: mmu_notifier_release can't run concurrently with
 229         * us because we hold the mm_users pin (either implicitly as
 230         * current->mm or explicitly with get_task_mm() or similar).
 231         * We can't race against any other mmu notifier method either
 232         * thanks to mm_take_all_locks().
 233         */
 234        spin_lock(&mm->mmu_notifier_mm->lock);
 235        hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 236        spin_unlock(&mm->mmu_notifier_mm->lock);
 237
 238        mm_drop_all_locks(mm);
 239out_clean:
 240        if (take_mmap_sem)
 241                up_write(&mm->mmap_sem);
 242        kfree(mmu_notifier_mm);
 243out:
 244        BUG_ON(atomic_read(&mm->mm_users) <= 0);
 245        return ret;
 246}
 247
 248/*
 249 * Must not hold mmap_sem nor any other VM related lock when calling
 250 * this registration function. Must also ensure mm_users can't go down
 251 * to zero while this runs to avoid races with mmu_notifier_release,
 252 * so mm has to be current->mm or the mm should be pinned safely such
 253 * as with get_task_mm(). If the mm is not current->mm, the mm_users
 254 * pin should be released by calling mmput after mmu_notifier_register
 255 * returns. mmu_notifier_unregister must be always called to
 256 * unregister the notifier. mm_count is automatically pinned to allow
 257 * mmu_notifier_unregister to safely run at any time later, before or
 258 * after exit_mmap. ->release will always be called before exit_mmap
 259 * frees the pages.
 260 */
 261int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 262{
 263        return do_mmu_notifier_register(mn, mm, 1);
 264}
 265EXPORT_SYMBOL_GPL(mmu_notifier_register);
 266
 267/*
 268 * Same as mmu_notifier_register but here the caller must hold the
 269 * mmap_sem in write mode.
 270 */
 271int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 272{
 273        return do_mmu_notifier_register(mn, mm, 0);
 274}
 275EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 276
 277/* this is called after the last mmu_notifier_unregister() returned */
 278void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 279{
 280        BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 281        kfree(mm->mmu_notifier_mm);
 282        mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 283}
 284
 285/*
 286 * This releases the mm_count pin automatically and frees the mm
 287 * structure if it was the last user of it. It serializes against
 288 * running mmu notifiers with SRCU and against mmu_notifier_unregister
 289 * with the unregister lock + SRCU. All sptes must be dropped before
 290 * calling mmu_notifier_unregister. ->release or any other notifier
 291 * method may be invoked concurrently with mmu_notifier_unregister,
 292 * and only after mmu_notifier_unregister returned we're guaranteed
 293 * that ->release or any other method can't run anymore.
 294 */
 295void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 296{
 297        BUG_ON(atomic_read(&mm->mm_count) <= 0);
 298
 299        spin_lock(&mm->mmu_notifier_mm->lock);
 300        if (!hlist_unhashed(&mn->hlist)) {
 301                int id;
 302
 303                /*
 304                 * Ensure we synchronize up with __mmu_notifier_release().
 305                 */
 306                id = srcu_read_lock(&srcu);
 307
 308                hlist_del_rcu(&mn->hlist);
 309                spin_unlock(&mm->mmu_notifier_mm->lock);
 310
 311                if (mn->ops->release)
 312                        mn->ops->release(mn, mm);
 313
 314                /*
 315                 * Allow __mmu_notifier_release() to complete.
 316                 */
 317                srcu_read_unlock(&srcu, id);
 318        } else
 319                spin_unlock(&mm->mmu_notifier_mm->lock);
 320
 321        /*
 322         * Wait for any running method to finish, including ->release() if it
 323         * was run by __mmu_notifier_release() instead of us.
 324         */
 325        synchronize_srcu(&srcu);
 326
 327        BUG_ON(atomic_read(&mm->mm_count) <= 0);
 328
 329        mmdrop(mm);
 330}
 331EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
 332
 333static int __init mmu_notifier_init(void)
 334{
 335        return init_srcu_struct(&srcu);
 336}
 337
 338module_init(mmu_notifier_init);
 339
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.