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/rcupdate.h>
  18 #include <linux/sched.h>
  19 #include <linux/slab.h>
  20
  21 /*
  22  * This function can't run concurrently against mmu_notifier_register
  23  * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
  24  * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
  25  * in parallel despite there being no task using this mm any more,
  26  * through the vmas outside of the exit_mmap context, such as with
  27  * vmtruncate. This serializes against mmu_notifier_unregister with
  28  * the mmu_notifier_mm->lock in addition to RCU and it serializes
  29  * against the other mmu notifiers with RCU. struct mmu_notifier_mm
  30  * can't go away from under us as exit_mmap holds an mm_count pin
  31  * itself.
  32  */
  33 void __mmu_notifier_release(struct mm_struct *mm)
  34 {
  35         struct mmu_notifier *mn;
  36         struct hlist_node *n;
  37
  38         /*
  39          * RCU here will block mmu_notifier_unregister until
  40          * ->release returns.
  41          */
  42         rcu_read_lock();
  43         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist)
  44                 /*
  45                  * if ->release runs before mmu_notifier_unregister it
  46                  * must be handled as it's the only way for the driver
  47                  * to flush all existing sptes and stop the driver
  48                  * from establishing any more sptes before all the
  49                  * pages in the mm are freed.
  50                  */
  51                 if (mn->ops->release)
  52                         mn->ops->release(mn, mm);
  53         rcu_read_unlock();
  54
  55         spin_lock(&mm->mmu_notifier_mm->lock);
  56         while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
  57                 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
  58                                  struct mmu_notifier,
  59                                  hlist);
  60                 /*
  61                  * We arrived before mmu_notifier_unregister so
  62                  * mmu_notifier_unregister will do nothing other than
  63                  * to wait ->release to finish and
  64                  * mmu_notifier_unregister to return.
  65                  */
  66                 hlist_del_init_rcu(&mn->hlist);
  67         }
  68         spin_unlock(&mm->mmu_notifier_mm->lock);
  69
  70         /*
  71          * synchronize_rcu here prevents mmu_notifier_release to
  72          * return to exit_mmap (which would proceed freeing all pages
  73          * in the mm) until the ->release method returns, if it was
  74          * invoked by mmu_notifier_unregister.
  75          *
  76          * The mmu_notifier_mm can't go away from under us because one
  77          * mm_count is hold by exit_mmap.
  78          */
  79         synchronize_rcu();
  80 }
  81
  82 /*
  83  * If no young bitflag is supported by the hardware, ->clear_flush_young can
  84  * unmap the address and return 1 or 0 depending if the mapping previously
  85  * existed or not.
  86  */
  87 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
  88                                         unsigned long address)
  89 {
  90         struct mmu_notifier *mn;
  91         struct hlist_node *n;
  92         int young = 0;
  93
  94         rcu_read_lock();
  95         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
  96                 if (mn->ops->clear_flush_young)
  97                         young |= mn->ops->clear_flush_young(mn, mm, address);
  98         }
  99         rcu_read_unlock();
 100
 101         return young;
 102 }
 103
 104 int __mmu_notifier_test_young(struct mm_struct *mm,
 105                               unsigned long address)
 106 {
 107         struct mmu_notifier *mn;
 108         struct hlist_node *n;
 109         int young = 0;
 110
 111         rcu_read_lock();
 112         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 113                 if (mn->ops->test_young) {
 114                         young = mn->ops->test_young(mn, mm, address);
 115                         if (young)
 116                                 break;
 117                 }
 118         }
 119         rcu_read_unlock();
 120
 121         return young;
 122 }
 123
 124 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 125                                pte_t pte)
 126 {
 127         struct mmu_notifier *mn;
 128         struct hlist_node *n;
 129
 130         rcu_read_lock();
 131         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 132                 if (mn->ops->change_pte)
 133                         mn->ops->change_pte(mn, mm, address, pte);
 134                 /*
 135                  * Some drivers don't have change_pte,
 136                  * so we must call invalidate_page in that case.
 137                  */
 138                 else if (mn->ops->invalidate_page)
 139                         mn->ops->invalidate_page(mn, mm, address);
 140         }
 141         rcu_read_unlock();
 142 }
 143
 144 void __mmu_notifier_invalidate_page(struct mm_struct *mm,
 145                                           unsigned long address)
 146 {
 147         struct mmu_notifier *mn;
 148         struct hlist_node *n;
 149
 150         rcu_read_lock();
 151         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 152                 if (mn->ops->invalidate_page)
 153                         mn->ops->invalidate_page(mn, mm, address);
 154         }
 155         rcu_read_unlock();
 156 }
 157
 158 void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 159                                   unsigned long start, unsigned long end)
 160 {
 161         struct mmu_notifier *mn;
 162         struct hlist_node *n;
 163
 164         rcu_read_lock();
 165         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 166                 if (mn->ops->invalidate_range_start)
 167                         mn->ops->invalidate_range_start(mn, mm, start, end);
 168         }
 169         rcu_read_unlock();
 170 }
 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         struct hlist_node *n;
 177
 178         rcu_read_lock();
 179         hlist_for_each_entry_rcu(mn, n, &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         rcu_read_unlock();
 184 }
 185
 186 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 187                                     struct mm_struct *mm,
 188                                     int take_mmap_sem)
 189 {
 190         struct mmu_notifier_mm *mmu_notifier_mm;
 191         int ret;
 192
 193         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 194
 195         ret = -ENOMEM;
 196         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 197         if (unlikely(!mmu_notifier_mm))
 198                 goto out;
 199
 200         if (take_mmap_sem)
 201                 down_write(&mm->mmap_sem);
 202         ret = mm_take_all_locks(mm);
 203         if (unlikely(ret))
 204                 goto out_cleanup;
 205
 206         if (!mm_has_notifiers(mm)) {
 207                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 208                 spin_lock_init(&mmu_notifier_mm->lock);
 209                 mm->mmu_notifier_mm = mmu_notifier_mm;
 210                 mmu_notifier_mm = NULL;
 211         }
 212         atomic_inc(&mm->mm_count);
 213
 214         /*
 215          * Serialize the update against mmu_notifier_unregister. A
 216          * side note: mmu_notifier_release can't run concurrently with
 217          * us because we hold the mm_users pin (either implicitly as
 218          * current->mm or explicitly with get_task_mm() or similar).
 219          * We can't race against any other mmu notifier method either
 220          * thanks to mm_take_all_locks().
 221          */
 222         spin_lock(&mm->mmu_notifier_mm->lock);
 223         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 224         spin_unlock(&mm->mmu_notifier_mm->lock);
 225
 226         mm_drop_all_locks(mm);
 227 out_cleanup:
 228         if (take_mmap_sem)
 229                 up_write(&mm->mmap_sem);
 230         /* kfree() does nothing if mmu_notifier_mm is NULL */
 231         kfree(mmu_notifier_mm);
 232 out:
 233         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 234         return ret;
 235 }
 236
 237 /*
 238  * Must not hold mmap_sem nor any other VM related lock when calling
 239  * this registration function. Must also ensure mm_users can't go down
 240  * to zero while this runs to avoid races with mmu_notifier_release,
 241  * so mm has to be current->mm or the mm should be pinned safely such
 242  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 243  * pin should be released by calling mmput after mmu_notifier_register
 244  * returns. mmu_notifier_unregister must be always called to
 245  * unregister the notifier. mm_count is automatically pinned to allow
 246  * mmu_notifier_unregister to safely run at any time later, before or
 247  * after exit_mmap. ->release will always be called before exit_mmap
 248  * frees the pages.
 249  */
 250 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 251 {
 252         return do_mmu_notifier_register(mn, mm, 1);
 253 }
 254 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 255
 256 /*
 257  * Same as mmu_notifier_register but here the caller must hold the
 258  * mmap_sem in write mode.
 259  */
 260 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 261 {
 262         return do_mmu_notifier_register(mn, mm, 0);
 263 }
 264 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 265
 266 /* this is called after the last mmu_notifier_unregister() returned */
 267 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 268 {
 269         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 270         kfree(mm->mmu_notifier_mm);
 271         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 272 }
 273
 274 /*
 275  * This releases the mm_count pin automatically and frees the mm
 276  * structure if it was the last user of it. It serializes against
 277  * running mmu notifiers with RCU and against mmu_notifier_unregister
 278  * with the unregister lock + RCU. All sptes must be dropped before
 279  * calling mmu_notifier_unregister. ->release or any other notifier
 280  * method may be invoked concurrently with mmu_notifier_unregister,
 281  * and only after mmu_notifier_unregister returned we're guaranteed
 282  * that ->release or any other method can't run anymore.
 283  */
 284 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 285 {
 286         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 287
 288         if (!hlist_unhashed(&mn->hlist)) {
 289                 /*
 290                  * RCU here will force exit_mmap to wait ->release to finish
 291                  * before freeing the pages.
 292                  */
 293                 rcu_read_lock();
 294
 295                 /*
 296                  * exit_mmap will block in mmu_notifier_release to
 297                  * guarantee ->release is called before freeing the
 298                  * pages.
 299                  */
 300                 if (mn->ops->release)
 301                         mn->ops->release(mn, mm);
 302                 rcu_read_unlock();
 303
 304                 spin_lock(&mm->mmu_notifier_mm->lock);
 305                 hlist_del_rcu(&mn->hlist);
 306                 spin_unlock(&mm->mmu_notifier_mm->lock);
 307         }
 308
 309         /*
 310          * Wait any running method to finish, of course including
 311          * ->release if it was run by mmu_notifier_relase instead of us.
 312          */
 313         synchronize_rcu();
 314
 315         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 316
 317         mmdrop(mm);
 318 }
 319 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);