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         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
 111 int __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
 131 void __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
 146 void __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
 161 void __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 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
 176
 177 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 178                                   unsigned long start, unsigned long end)
 179 {
 180         struct mmu_notifier *mn;
 181         struct hlist_node *n;
 182         int id;
 183
 184         id = srcu_read_lock(&srcu);
 185         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 186                 if (mn->ops->invalidate_range_end)
 187                         mn->ops->invalidate_range_end(mn, mm, start, end);
 188         }
 189         srcu_read_unlock(&srcu, id);
 190 }
 191 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
 192
 193 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 194                                     struct mm_struct *mm,
 195                                     int take_mmap_sem)
 196 {
 197         struct mmu_notifier_mm *mmu_notifier_mm;
 198         int ret;
 199
 200         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 201
 202         /*
 203          * Verify that mmu_notifier_init() already run and the global srcu is
 204          * initialized.
 205          */
 206         BUG_ON(!srcu.per_cpu_ref);
 207
 208         ret = -ENOMEM;
 209         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 210         if (unlikely(!mmu_notifier_mm))
 211                 goto out;
 212
 213         if (take_mmap_sem)
 214                 down_write(&mm->mmap_sem);
 215         ret = mm_take_all_locks(mm);
 216         if (unlikely(ret))
 217                 goto out_clean;
 218
 219         if (!mm_has_notifiers(mm)) {
 220                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 221                 spin_lock_init(&mmu_notifier_mm->lock);
 222
 223                 mm->mmu_notifier_mm = mmu_notifier_mm;
 224                 mmu_notifier_mm = NULL;
 225         }
 226         atomic_inc(&mm->mm_count);
 227
 228         /*
 229          * Serialize the update against mmu_notifier_unregister. A
 230          * side note: mmu_notifier_release can't run concurrently with
 231          * us because we hold the mm_users pin (either implicitly as
 232          * current->mm or explicitly with get_task_mm() or similar).
 233          * We can't race against any other mmu notifier method either
 234          * thanks to mm_take_all_locks().
 235          */
 236         spin_lock(&mm->mmu_notifier_mm->lock);
 237         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 238         spin_unlock(&mm->mmu_notifier_mm->lock);
 239
 240         mm_drop_all_locks(mm);
 241 out_clean:
 242         if (take_mmap_sem)
 243                 up_write(&mm->mmap_sem);
 244         kfree(mmu_notifier_mm);
 245 out:
 246         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 247         return ret;
 248 }
 249
 250 /*
 251  * Must not hold mmap_sem nor any other VM related lock when calling
 252  * this registration function. Must also ensure mm_users can't go down
 253  * to zero while this runs to avoid races with mmu_notifier_release,
 254  * so mm has to be current->mm or the mm should be pinned safely such
 255  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 256  * pin should be released by calling mmput after mmu_notifier_register
 257  * returns. mmu_notifier_unregister must be always called to
 258  * unregister the notifier. mm_count is automatically pinned to allow
 259  * mmu_notifier_unregister to safely run at any time later, before or
 260  * after exit_mmap. ->release will always be called before exit_mmap
 261  * frees the pages.
 262  */
 263 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 264 {
 265         return do_mmu_notifier_register(mn, mm, 1);
 266 }
 267 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 268
 269 /*
 270  * Same as mmu_notifier_register but here the caller must hold the
 271  * mmap_sem in write mode.
 272  */
 273 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 274 {
 275         return do_mmu_notifier_register(mn, mm, 0);
 276 }
 277 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 278
 279 /* this is called after the last mmu_notifier_unregister() returned */
 280 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 281 {
 282         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 283         kfree(mm->mmu_notifier_mm);
 284         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 285 }
 286
 287 /*
 288  * This releases the mm_count pin automatically and frees the mm
 289  * structure if it was the last user of it. It serializes against
 290  * running mmu notifiers with SRCU and against mmu_notifier_unregister
 291  * with the unregister lock + SRCU. All sptes must be dropped before
 292  * calling mmu_notifier_unregister. ->release or any other notifier
 293  * method may be invoked concurrently with mmu_notifier_unregister,
 294  * and only after mmu_notifier_unregister returned we're guaranteed
 295  * that ->release or any other method can't run anymore.
 296  */
 297 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 298 {
 299         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 300
 301         spin_lock(&mm->mmu_notifier_mm->lock);
 302         if (!hlist_unhashed(&mn->hlist)) {
 303                 int id;
 304
 305                 /*
 306                  * Ensure we synchronize up with __mmu_notifier_release().
 307                  */
 308                 id = srcu_read_lock(&srcu);
 309
 310                 hlist_del_rcu(&mn->hlist);
 311                 spin_unlock(&mm->mmu_notifier_mm->lock);
 312
 313                 if (mn->ops->release)
 314                         mn->ops->release(mn, mm);
 315
 316                 /*
 317                  * Allow __mmu_notifier_release() to complete.
 318                  */
 319                 srcu_read_unlock(&srcu, id);
 320         } else
 321                 spin_unlock(&mm->mmu_notifier_mm->lock);
 322
 323         /*
 324          * Wait for any running method to finish, including ->release() if it
 325          * was run by __mmu_notifier_release() instead of us.
 326          */
 327         synchronize_srcu(&srcu);
 328
 329         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 330
 331         mmdrop(mm);
 332 }
 333 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
 334
 335 static int __init mmu_notifier_init(void)
 336 {
 337         return init_srcu_struct(&srcu);
 338 }
 339
 340 module_init(mmu_notifier_init);