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
  20 /*
  21  * This function can't run concurrently against mmu_notifier_register
  22  * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
  23  * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
  24  * in parallel despite there being no task using this mm any more,
  25  * through the vmas outside of the exit_mmap context, such as with
  26  * vmtruncate. This serializes against mmu_notifier_unregister with
  27  * the mmu_notifier_mm->lock in addition to RCU and it serializes
  28  * against the other mmu notifiers with RCU. struct mmu_notifier_mm
  29  * can't go away from under us as exit_mmap holds an mm_count pin
  30  * itself.
  31  */
  32 void __mmu_notifier_release(struct mm_struct *mm)
  33 {
  34         struct mmu_notifier *mn;
  35
  36         spin_lock(&mm->mmu_notifier_mm->lock);
  37         while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
  38                 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
  39                                  struct mmu_notifier,
  40                                  hlist);
  41                 /*
  42                  * We arrived before mmu_notifier_unregister so
  43                  * mmu_notifier_unregister will do nothing other than
  44                  * to wait ->release to finish and
  45                  * mmu_notifier_unregister to return.
  46                  */
  47                 hlist_del_init_rcu(&mn->hlist);
  48                 /*
  49                  * RCU here will block mmu_notifier_unregister until
  50                  * ->release returns.
  51                  */
  52                 rcu_read_lock();
  53                 spin_unlock(&mm->mmu_notifier_mm->lock);
  54                 /*
  55                  * if ->release runs before mmu_notifier_unregister it
  56                  * must be handled as it's the only way for the driver
  57                  * to flush all existing sptes and stop the driver
  58                  * from establishing any more sptes before all the
  59                  * pages in the mm are freed.
  60                  */
  61                 if (mn->ops->release)
  62                         mn->ops->release(mn, mm);
  63                 rcu_read_unlock();
  64                 spin_lock(&mm->mmu_notifier_mm->lock);
  65         }
  66         spin_unlock(&mm->mmu_notifier_mm->lock);
  67
  68         /*
  69          * synchronize_rcu here prevents mmu_notifier_release to
  70          * return to exit_mmap (which would proceed freeing all pages
  71          * in the mm) until the ->release method returns, if it was
  72          * invoked by mmu_notifier_unregister.
  73          *
  74          * The mmu_notifier_mm can't go away from under us because one
  75          * mm_count is hold by exit_mmap.
  76          */
  77         synchronize_rcu();
  78 }
  79
  80 /*
  81  * If no young bitflag is supported by the hardware, ->clear_flush_young can
  82  * unmap the address and return 1 or 0 depending if the mapping previously
  83  * existed or not.
  84  */
  85 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
  86                                         unsigned long address)
  87 {
  88         struct mmu_notifier *mn;
  89         struct hlist_node *n;
  90         int young = 0;
  91
  92         rcu_read_lock();
  93         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
  94                 if (mn->ops->clear_flush_young)
  95                         young |= mn->ops->clear_flush_young(mn, mm, address);
  96         }
  97         rcu_read_unlock();
  98
  99         return young;
 100 }
 101
 102 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 103                                pte_t pte)
 104 {
 105         struct mmu_notifier *mn;
 106         struct hlist_node *n;
 107
 108         rcu_read_lock();
 109         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 110                 if (mn->ops->change_pte)
 111                         mn->ops->change_pte(mn, mm, address, pte);
 112                 /*
 113                  * Some drivers don't have change_pte,
 114                  * so we must call invalidate_page in that case.
 115                  */
 116                 else if (mn->ops->invalidate_page)
 117                         mn->ops->invalidate_page(mn, mm, address);
 118         }
 119         rcu_read_unlock();
 120 }
 121
 122 void __mmu_notifier_invalidate_page(struct mm_struct *mm,
 123                                           unsigned long address)
 124 {
 125         struct mmu_notifier *mn;
 126         struct hlist_node *n;
 127
 128         rcu_read_lock();
 129         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 130                 if (mn->ops->invalidate_page)
 131                         mn->ops->invalidate_page(mn, mm, address);
 132         }
 133         rcu_read_unlock();
 134 }
 135
 136 void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 137                                   unsigned long start, unsigned long end)
 138 {
 139         struct mmu_notifier *mn;
 140         struct hlist_node *n;
 141
 142         rcu_read_lock();
 143         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 144                 if (mn->ops->invalidate_range_start)
 145                         mn->ops->invalidate_range_start(mn, mm, start, end);
 146         }
 147         rcu_read_unlock();
 148 }
 149
 150 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 151                                   unsigned long start, unsigned long end)
 152 {
 153         struct mmu_notifier *mn;
 154         struct hlist_node *n;
 155
 156         rcu_read_lock();
 157         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 158                 if (mn->ops->invalidate_range_end)
 159                         mn->ops->invalidate_range_end(mn, mm, start, end);
 160         }
 161         rcu_read_unlock();
 162 }
 163
 164 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 165                                     struct mm_struct *mm,
 166                                     int take_mmap_sem)
 167 {
 168         struct mmu_notifier_mm *mmu_notifier_mm;
 169         int ret;
 170
 171         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 172
 173         ret = -ENOMEM;
 174         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 175         if (unlikely(!mmu_notifier_mm))
 176                 goto out;
 177
 178         if (take_mmap_sem)
 179                 down_write(&mm->mmap_sem);
 180         ret = mm_take_all_locks(mm);
 181         if (unlikely(ret))
 182                 goto out_cleanup;
 183
 184         if (!mm_has_notifiers(mm)) {
 185                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 186                 spin_lock_init(&mmu_notifier_mm->lock);
 187                 mm->mmu_notifier_mm = mmu_notifier_mm;
 188                 mmu_notifier_mm = NULL;
 189         }
 190         atomic_inc(&mm->mm_count);
 191
 192         /*
 193          * Serialize the update against mmu_notifier_unregister. A
 194          * side note: mmu_notifier_release can't run concurrently with
 195          * us because we hold the mm_users pin (either implicitly as
 196          * current->mm or explicitly with get_task_mm() or similar).
 197          * We can't race against any other mmu notifier method either
 198          * thanks to mm_take_all_locks().
 199          */
 200         spin_lock(&mm->mmu_notifier_mm->lock);
 201         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 202         spin_unlock(&mm->mmu_notifier_mm->lock);
 203
 204         mm_drop_all_locks(mm);
 205 out_cleanup:
 206         if (take_mmap_sem)
 207                 up_write(&mm->mmap_sem);
 208         /* kfree() does nothing if mmu_notifier_mm is NULL */
 209         kfree(mmu_notifier_mm);
 210 out:
 211         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 212         return ret;
 213 }
 214
 215 /*
 216  * Must not hold mmap_sem nor any other VM related lock when calling
 217  * this registration function. Must also ensure mm_users can't go down
 218  * to zero while this runs to avoid races with mmu_notifier_release,
 219  * so mm has to be current->mm or the mm should be pinned safely such
 220  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 221  * pin should be released by calling mmput after mmu_notifier_register
 222  * returns. mmu_notifier_unregister must be always called to
 223  * unregister the notifier. mm_count is automatically pinned to allow
 224  * mmu_notifier_unregister to safely run at any time later, before or
 225  * after exit_mmap. ->release will always be called before exit_mmap
 226  * frees the pages.
 227  */
 228 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 229 {
 230         return do_mmu_notifier_register(mn, mm, 1);
 231 }
 232 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 233
 234 /*
 235  * Same as mmu_notifier_register but here the caller must hold the
 236  * mmap_sem in write mode.
 237  */
 238 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 239 {
 240         return do_mmu_notifier_register(mn, mm, 0);
 241 }
 242 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 243
 244 /* this is called after the last mmu_notifier_unregister() returned */
 245 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 246 {
 247         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 248         kfree(mm->mmu_notifier_mm);
 249         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 250 }
 251
 252 /*
 253  * This releases the mm_count pin automatically and frees the mm
 254  * structure if it was the last user of it. It serializes against
 255  * running mmu notifiers with RCU and against mmu_notifier_unregister
 256  * with the unregister lock + RCU. All sptes must be dropped before
 257  * calling mmu_notifier_unregister. ->release or any other notifier
 258  * method may be invoked concurrently with mmu_notifier_unregister,
 259  * and only after mmu_notifier_unregister returned we're guaranteed
 260  * that ->release or any other method can't run anymore.
 261  */
 262 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 263 {
 264         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 265
 266         spin_lock(&mm->mmu_notifier_mm->lock);
 267         if (!hlist_unhashed(&mn->hlist)) {
 268                 hlist_del_rcu(&mn->hlist);
 269
 270                 /*
 271                  * RCU here will force exit_mmap to wait ->release to finish
 272                  * before freeing the pages.
 273                  */
 274                 rcu_read_lock();
 275                 spin_unlock(&mm->mmu_notifier_mm->lock);
 276                 /*
 277                  * exit_mmap will block in mmu_notifier_release to
 278                  * guarantee ->release is called before freeing the
 279                  * pages.
 280                  */
 281                 if (mn->ops->release)
 282                         mn->ops->release(mn, mm);
 283                 rcu_read_unlock();
 284         } else
 285                 spin_unlock(&mm->mmu_notifier_mm->lock);
 286
 287         /*
 288          * Wait any running method to finish, of course including
 289          * ->release if it was run by mmu_notifier_relase instead of us.
 290          */
 291         synchronize_rcu();
 292
 293         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 294
 295         mmdrop(mm);
 296 }
 297 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);