mm/mmu_notifier.c

   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 */
  23 static 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  */
  37 void __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  */
  94 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
  95                                         unsigned long address)
  96 {
  97         struct mmu_notifier *mn;
  98         int young = 0, id;
  99
 100         id = srcu_read_lock(&srcu);
 101         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 102                 if (mn->ops->clear_flush_young)
 103                         young |= mn->ops->clear_flush_young(mn, mm, address);
 104         }
 105         srcu_read_unlock(&srcu, id);
 106
 107         return young;
 108 }
 109
 110 int __mmu_notifier_test_young(struct mm_struct *mm,
 111                               unsigned long address)
 112 {
 113         struct mmu_notifier *mn;
 114         int young = 0, id;
 115
 116         id = srcu_read_lock(&srcu);
 117         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 118                 if (mn->ops->test_young) {
 119                         young = mn->ops->test_young(mn, mm, address);
 120                         if (young)
 121                                 break;
 122                 }
 123         }
 124         srcu_read_unlock(&srcu, id);
 125
 126         return young;
 127 }
 128
 129 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 130                                pte_t pte)
 131 {
 132         struct mmu_notifier *mn;
 133         int id;
 134
 135         id = srcu_read_lock(&srcu);
 136         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 137                 if (mn->ops->change_pte)
 138                         mn->ops->change_pte(mn, mm, address, pte);
 139         }
 140         srcu_read_unlock(&srcu, id);
 141 }
 142
 143 void __mmu_notifier_invalidate_page(struct mm_struct *mm,
 144                                           unsigned long address)
 145 {
 146         struct mmu_notifier *mn;
 147         int id;
 148
 149         id = srcu_read_lock(&srcu);
 150         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 151                 if (mn->ops->invalidate_page)
 152                         mn->ops->invalidate_page(mn, mm, address);
 153         }
 154         srcu_read_unlock(&srcu, id);
 155 }
 156
 157 void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 158                                   unsigned long start, unsigned long end)
 159 {
 160         struct mmu_notifier *mn;
 161         int id;
 162
 163         id = srcu_read_lock(&srcu);
 164         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 165                 if (mn->ops->invalidate_range_start)
 166                         mn->ops->invalidate_range_start(mn, mm, start, end);
 167         }
 168         srcu_read_unlock(&srcu, id);
 169 }
 170 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
 171
 172 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 173                                   unsigned long start, unsigned long end)
 174 {
 175         struct mmu_notifier *mn;
 176         int id;
 177
 178         id = srcu_read_lock(&srcu);
 179         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 180                 if (mn->ops->invalidate_range_end)
 181                         mn->ops->invalidate_range_end(mn, mm, start, end);
 182         }
 183         srcu_read_unlock(&srcu, id);
 184 }
 185 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
 186
 187 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 188                                     struct mm_struct *mm,
 189                                     int take_mmap_sem)
 190 {
 191         struct mmu_notifier_mm *mmu_notifier_mm;
 192         int ret;
 193
 194         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 195
 196         /*
 197          * Verify that mmu_notifier_init() already run and the global srcu is
 198          * initialized.
 199          */
 200         BUG_ON(!srcu.per_cpu_ref);
 201
 202         ret = -ENOMEM;
 203         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 204         if (unlikely(!mmu_notifier_mm))
 205                 goto out;
 206
 207         if (take_mmap_sem)
 208                 down_write(&mm->mmap_sem);
 209         ret = mm_take_all_locks(mm);
 210         if (unlikely(ret))
 211                 goto out_clean;
 212
 213         if (!mm_has_notifiers(mm)) {
 214                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 215                 spin_lock_init(&mmu_notifier_mm->lock);
 216
 217                 mm->mmu_notifier_mm = mmu_notifier_mm;
 218                 mmu_notifier_mm = NULL;
 219         }
 220         atomic_inc(&mm->mm_count);
 221
 222         /*
 223          * Serialize the update against mmu_notifier_unregister. A
 224          * side note: mmu_notifier_release can't run concurrently with
 225          * us because we hold the mm_users pin (either implicitly as
 226          * current->mm or explicitly with get_task_mm() or similar).
 227          * We can't race against any other mmu notifier method either
 228          * thanks to mm_take_all_locks().
 229          */
 230         spin_lock(&mm->mmu_notifier_mm->lock);
 231         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 232         spin_unlock(&mm->mmu_notifier_mm->lock);
 233
 234         mm_drop_all_locks(mm);
 235 out_clean:
 236         if (take_mmap_sem)
 237                 up_write(&mm->mmap_sem);
 238         kfree(mmu_notifier_mm);
 239 out:
 240         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 241         return ret;
 242 }
 243
 244 /*
 245  * Must not hold mmap_sem nor any other VM related lock when calling
 246  * this registration function. Must also ensure mm_users can't go down
 247  * to zero while this runs to avoid races with mmu_notifier_release,
 248  * so mm has to be current->mm or the mm should be pinned safely such
 249  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 250  * pin should be released by calling mmput after mmu_notifier_register
 251  * returns. mmu_notifier_unregister must be always called to
 252  * unregister the notifier. mm_count is automatically pinned to allow
 253  * mmu_notifier_unregister to safely run at any time later, before or
 254  * after exit_mmap. ->release will always be called before exit_mmap
 255  * frees the pages.
 256  */
 257 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 258 {
 259         return do_mmu_notifier_register(mn, mm, 1);
 260 }
 261 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 262
 263 /*
 264  * Same as mmu_notifier_register but here the caller must hold the
 265  * mmap_sem in write mode.
 266  */
 267 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 268 {
 269         return do_mmu_notifier_register(mn, mm, 0);
 270 }
 271 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 272
 273 /* this is called after the last mmu_notifier_unregister() returned */
 274 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 275 {
 276         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 277         kfree(mm->mmu_notifier_mm);
 278         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 279 }
 280
 281 /*
 282  * This releases the mm_count pin automatically and frees the mm
 283  * structure if it was the last user of it. It serializes against
 284  * running mmu notifiers with SRCU and against mmu_notifier_unregister
 285  * with the unregister lock + SRCU. All sptes must be dropped before
 286  * calling mmu_notifier_unregister. ->release or any other notifier
 287  * method may be invoked concurrently with mmu_notifier_unregister,
 288  * and only after mmu_notifier_unregister returned we're guaranteed
 289  * that ->release or any other method can't run anymore.
 290  */
 291 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 292 {
 293         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 294
 295         spin_lock(&mm->mmu_notifier_mm->lock);
 296         if (!hlist_unhashed(&mn->hlist)) {
 297                 int id;
 298
 299                 /*
 300                  * Ensure we synchronize up with __mmu_notifier_release().
 301                  */
 302                 id = srcu_read_lock(&srcu);
 303
 304                 hlist_del_rcu(&mn->hlist);
 305                 spin_unlock(&mm->mmu_notifier_mm->lock);
 306
 307                 if (mn->ops->release)
 308                         mn->ops->release(mn, mm);
 309
 310                 /*
 311                  * Allow __mmu_notifier_release() to complete.
 312                  */
 313                 srcu_read_unlock(&srcu, id);
 314         } else
 315                 spin_unlock(&mm->mmu_notifier_mm->lock);
 316
 317         /*
 318          * Wait for any running method to finish, including ->release() if it
 319          * was run by __mmu_notifier_release() instead of us.
 320          */
 321         synchronize_srcu(&srcu);
 322
 323         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 324
 325         mmdrop(mm);
 326 }
 327 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
 328
 329 static int __init mmu_notifier_init(void)
 330 {
 331         return init_srcu_struct(&srcu);
 332 }
 333
 334 module_init(mmu_notifier_init);