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/module.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
  37         spin_lock(&mm->mmu_notifier_mm->lock);
  38         while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
  39                 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
  40                                  struct mmu_notifier,
  41                                  hlist);
  42                 /*
  43                  * We arrived before mmu_notifier_unregister so
  44                  * mmu_notifier_unregister will do nothing other than
  45                  * to wait ->release to finish and
  46                  * mmu_notifier_unregister to return.
  47                  */
  48                 hlist_del_init_rcu(&mn->hlist);
  49                 /*
  50                  * RCU here will block mmu_notifier_unregister until
  51                  * ->release returns.
  52                  */
  53                 rcu_read_lock();
  54                 spin_unlock(&mm->mmu_notifier_mm->lock);
  55                 /*
  56                  * if ->release runs before mmu_notifier_unregister it
  57                  * must be handled as it's the only way for the driver
  58                  * to flush all existing sptes and stop the driver
  59                  * from establishing any more sptes before all the
  60                  * pages in the mm are freed.
  61                  */
  62                 if (mn->ops->release)
  63                         mn->ops->release(mn, mm);
  64                 rcu_read_unlock();
  65                 spin_lock(&mm->mmu_notifier_mm->lock);
  66         }
  67         spin_unlock(&mm->mmu_notifier_mm->lock);
  68
  69         /*
  70          * synchronize_rcu here prevents mmu_notifier_release to
  71          * return to exit_mmap (which would proceed freeing all pages
  72          * in the mm) until the ->release method returns, if it was
  73          * invoked by mmu_notifier_unregister.
  74          *
  75          * The mmu_notifier_mm can't go away from under us because one
  76          * mm_count is hold by exit_mmap.
  77          */
  78         synchronize_rcu();
  79 }
  80
  81 /*
  82  * If no young bitflag is supported by the hardware, ->clear_flush_young can
  83  * unmap the address and return 1 or 0 depending if the mapping previously
  84  * existed or not.
  85  */
  86 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
  87                                         unsigned long address)
  88 {
  89         struct mmu_notifier *mn;
  90         struct hlist_node *n;
  91         int young = 0;
  92
  93         rcu_read_lock();
  94         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
  95                 if (mn->ops->clear_flush_young)
  96                         young |= mn->ops->clear_flush_young(mn, mm, address);
  97         }
  98         rcu_read_unlock();
  99
 100         return young;
 101 }
 102
 103 int __mmu_notifier_test_young(struct mm_struct *mm,
 104                               unsigned long address)
 105 {
 106         struct mmu_notifier *mn;
 107         struct hlist_node *n;
 108         int young = 0;
 109
 110         rcu_read_lock();
 111         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 112                 if (mn->ops->test_young) {
 113                         young = mn->ops->test_young(mn, mm, address);
 114                         if (young)
 115                                 break;
 116                 }
 117         }
 118         rcu_read_unlock();
 119
 120         return young;
 121 }
 122
 123 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 124                                pte_t pte)
 125 {
 126         struct mmu_notifier *mn;
 127         struct hlist_node *n;
 128
 129         rcu_read_lock();
 130         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 131                 if (mn->ops->change_pte)
 132                         mn->ops->change_pte(mn, mm, address, pte);
 133                 /*
 134                  * Some drivers don't have change_pte,
 135                  * so we must call invalidate_page in that case.
 136                  */
 137                 else if (mn->ops->invalidate_page)
 138                         mn->ops->invalidate_page(mn, mm, address);
 139         }
 140         rcu_read_unlock();
 141 }
 142
 143 void __mmu_notifier_invalidate_page(struct mm_struct *mm,
 144                                           unsigned long address)
 145 {
 146         struct mmu_notifier *mn;
 147         struct hlist_node *n;
 148
 149         rcu_read_lock();
 150         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 151                 if (mn->ops->invalidate_page)
 152                         mn->ops->invalidate_page(mn, mm, address);
 153         }
 154         rcu_read_unlock();
 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         struct hlist_node *n;
 162
 163         rcu_read_lock();
 164         hlist_for_each_entry_rcu(mn, n, &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         rcu_read_unlock();
 169 }
 170
 171 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 172                                   unsigned long start, unsigned long end)
 173 {
 174         struct mmu_notifier *mn;
 175         struct hlist_node *n;
 176
 177         rcu_read_lock();
 178         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 179                 if (mn->ops->invalidate_range_end)
 180                         mn->ops->invalidate_range_end(mn, mm, start, end);
 181         }
 182         rcu_read_unlock();
 183 }
 184
 185 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 186                                     struct mm_struct *mm,
 187                                     int take_mmap_sem)
 188 {
 189         struct mmu_notifier_mm *mmu_notifier_mm;
 190         int ret;
 191
 192         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 193
 194         ret = -ENOMEM;
 195         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 196         if (unlikely(!mmu_notifier_mm))
 197                 goto out;
 198
 199         if (take_mmap_sem)
 200                 down_write(&mm->mmap_sem);
 201         ret = mm_take_all_locks(mm);
 202         if (unlikely(ret))
 203                 goto out_cleanup;
 204
 205         if (!mm_has_notifiers(mm)) {
 206                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 207                 spin_lock_init(&mmu_notifier_mm->lock);
 208                 mm->mmu_notifier_mm = mmu_notifier_mm;
 209                 mmu_notifier_mm = NULL;
 210         }
 211         atomic_inc(&mm->mm_count);
 212
 213         /*
 214          * Serialize the update against mmu_notifier_unregister. A
 215          * side note: mmu_notifier_release can't run concurrently with
 216          * us because we hold the mm_users pin (either implicitly as
 217          * current->mm or explicitly with get_task_mm() or similar).
 218          * We can't race against any other mmu notifier method either
 219          * thanks to mm_take_all_locks().
 220          */
 221         spin_lock(&mm->mmu_notifier_mm->lock);
 222         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 223         spin_unlock(&mm->mmu_notifier_mm->lock);
 224
 225         mm_drop_all_locks(mm);
 226 out_cleanup:
 227         if (take_mmap_sem)
 228                 up_write(&mm->mmap_sem);
 229         /* kfree() does nothing if mmu_notifier_mm is NULL */
 230         kfree(mmu_notifier_mm);
 231 out:
 232         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 233         return ret;
 234 }
 235
 236 /*
 237  * Must not hold mmap_sem nor any other VM related lock when calling
 238  * this registration function. Must also ensure mm_users can't go down
 239  * to zero while this runs to avoid races with mmu_notifier_release,
 240  * so mm has to be current->mm or the mm should be pinned safely such
 241  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 242  * pin should be released by calling mmput after mmu_notifier_register
 243  * returns. mmu_notifier_unregister must be always called to
 244  * unregister the notifier. mm_count is automatically pinned to allow
 245  * mmu_notifier_unregister to safely run at any time later, before or
 246  * after exit_mmap. ->release will always be called before exit_mmap
 247  * frees the pages.
 248  */
 249 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 250 {
 251         return do_mmu_notifier_register(mn, mm, 1);
 252 }
 253 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 254
 255 /*
 256  * Same as mmu_notifier_register but here the caller must hold the
 257  * mmap_sem in write mode.
 258  */
 259 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 260 {
 261         return do_mmu_notifier_register(mn, mm, 0);
 262 }
 263 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 264
 265 /* this is called after the last mmu_notifier_unregister() returned */
 266 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 267 {
 268         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 269         kfree(mm->mmu_notifier_mm);
 270         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 271 }
 272
 273 /*
 274  * This releases the mm_count pin automatically and frees the mm
 275  * structure if it was the last user of it. It serializes against
 276  * running mmu notifiers with RCU and against mmu_notifier_unregister
 277  * with the unregister lock + RCU. All sptes must be dropped before
 278  * calling mmu_notifier_unregister. ->release or any other notifier
 279  * method may be invoked concurrently with mmu_notifier_unregister,
 280  * and only after mmu_notifier_unregister returned we're guaranteed
 281  * that ->release or any other method can't run anymore.
 282  */
 283 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 284 {
 285         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 286
 287         spin_lock(&mm->mmu_notifier_mm->lock);
 288         if (!hlist_unhashed(&mn->hlist)) {
 289                 hlist_del_rcu(&mn->hlist);
 290
 291                 /*
 292                  * RCU here will force exit_mmap to wait ->release to finish
 293                  * before freeing the pages.
 294                  */
 295                 rcu_read_lock();
 296                 spin_unlock(&mm->mmu_notifier_mm->lock);
 297                 /*
 298                  * exit_mmap will block in mmu_notifier_release to
 299                  * guarantee ->release is called before freeing the
 300                  * pages.
 301                  */
 302                 if (mn->ops->release)
 303                         mn->ops->release(mn, mm);
 304                 rcu_read_unlock();
 305         } else
 306                 spin_unlock(&mm->mmu_notifier_mm->lock);
 307
 308         /*
 309          * Wait any running method to finish, of course including
 310          * ->release if it was run by mmu_notifier_relase instead of us.
 311          */
 312         synchronize_rcu();
 313
 314         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 315
 316         mmdrop(mm);
 317 }
 318 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);