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 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 104                                pte_t pte)
 105 {
 106         struct mmu_notifier *mn;
 107         struct hlist_node *n;
 108
 109         rcu_read_lock();
 110         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 111                 if (mn->ops->change_pte)
 112                         mn->ops->change_pte(mn, mm, address, pte);
 113                 /*
 114                  * Some drivers don't have change_pte,
 115                  * so we must call invalidate_page in that case.
 116                  */
 117                 else if (mn->ops->invalidate_page)
 118                         mn->ops->invalidate_page(mn, mm, address);
 119         }
 120         rcu_read_unlock();
 121 }
 122
 123 void __mmu_notifier_invalidate_page(struct mm_struct *mm,
 124                                           unsigned long address)
 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->invalidate_page)
 132                         mn->ops->invalidate_page(mn, mm, address);
 133         }
 134         rcu_read_unlock();
 135 }
 136
 137 void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 138                                   unsigned long start, unsigned long end)
 139 {
 140         struct mmu_notifier *mn;
 141         struct hlist_node *n;
 142
 143         rcu_read_lock();
 144         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 145                 if (mn->ops->invalidate_range_start)
 146                         mn->ops->invalidate_range_start(mn, mm, start, end);
 147         }
 148         rcu_read_unlock();
 149 }
 150
 151 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 152                                   unsigned long start, unsigned long end)
 153 {
 154         struct mmu_notifier *mn;
 155         struct hlist_node *n;
 156
 157         rcu_read_lock();
 158         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 159                 if (mn->ops->invalidate_range_end)
 160                         mn->ops->invalidate_range_end(mn, mm, start, end);
 161         }
 162         rcu_read_unlock();
 163 }
 164
 165 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 166                                     struct mm_struct *mm,
 167                                     int take_mmap_sem)
 168 {
 169         struct mmu_notifier_mm *mmu_notifier_mm;
 170         int ret;
 171
 172         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 173
 174         ret = -ENOMEM;
 175         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 176         if (unlikely(!mmu_notifier_mm))
 177                 goto out;
 178
 179         if (take_mmap_sem)
 180                 down_write(&mm->mmap_sem);
 181         ret = mm_take_all_locks(mm);
 182         if (unlikely(ret))
 183                 goto out_cleanup;
 184
 185         if (!mm_has_notifiers(mm)) {
 186                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 187                 spin_lock_init(&mmu_notifier_mm->lock);
 188                 mm->mmu_notifier_mm = mmu_notifier_mm;
 189                 mmu_notifier_mm = NULL;
 190         }
 191         atomic_inc(&mm->mm_count);
 192
 193         /*
 194          * Serialize the update against mmu_notifier_unregister. A
 195          * side note: mmu_notifier_release can't run concurrently with
 196          * us because we hold the mm_users pin (either implicitly as
 197          * current->mm or explicitly with get_task_mm() or similar).
 198          * We can't race against any other mmu notifier method either
 199          * thanks to mm_take_all_locks().
 200          */
 201         spin_lock(&mm->mmu_notifier_mm->lock);
 202         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 203         spin_unlock(&mm->mmu_notifier_mm->lock);
 204
 205         mm_drop_all_locks(mm);
 206 out_cleanup:
 207         if (take_mmap_sem)
 208                 up_write(&mm->mmap_sem);
 209         /* kfree() does nothing if mmu_notifier_mm is NULL */
 210         kfree(mmu_notifier_mm);
 211 out:
 212         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 213         return ret;
 214 }
 215
 216 /*
 217  * Must not hold mmap_sem nor any other VM related lock when calling
 218  * this registration function. Must also ensure mm_users can't go down
 219  * to zero while this runs to avoid races with mmu_notifier_release,
 220  * so mm has to be current->mm or the mm should be pinned safely such
 221  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 222  * pin should be released by calling mmput after mmu_notifier_register
 223  * returns. mmu_notifier_unregister must be always called to
 224  * unregister the notifier. mm_count is automatically pinned to allow
 225  * mmu_notifier_unregister to safely run at any time later, before or
 226  * after exit_mmap. ->release will always be called before exit_mmap
 227  * frees the pages.
 228  */
 229 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 230 {
 231         return do_mmu_notifier_register(mn, mm, 1);
 232 }
 233 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 234
 235 /*
 236  * Same as mmu_notifier_register but here the caller must hold the
 237  * mmap_sem in write mode.
 238  */
 239 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 240 {
 241         return do_mmu_notifier_register(mn, mm, 0);
 242 }
 243 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 244
 245 /* this is called after the last mmu_notifier_unregister() returned */
 246 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 247 {
 248         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 249         kfree(mm->mmu_notifier_mm);
 250         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 251 }
 252
 253 /*
 254  * This releases the mm_count pin automatically and frees the mm
 255  * structure if it was the last user of it. It serializes against
 256  * running mmu notifiers with RCU and against mmu_notifier_unregister
 257  * with the unregister lock + RCU. All sptes must be dropped before
 258  * calling mmu_notifier_unregister. ->release or any other notifier
 259  * method may be invoked concurrently with mmu_notifier_unregister,
 260  * and only after mmu_notifier_unregister returned we're guaranteed
 261  * that ->release or any other method can't run anymore.
 262  */
 263 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 264 {
 265         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 266
 267         spin_lock(&mm->mmu_notifier_mm->lock);
 268         if (!hlist_unhashed(&mn->hlist)) {
 269                 hlist_del_rcu(&mn->hlist);
 270
 271                 /*
 272                  * RCU here will force exit_mmap to wait ->release to finish
 273                  * before freeing the pages.
 274                  */
 275                 rcu_read_lock();
 276                 spin_unlock(&mm->mmu_notifier_mm->lock);
 277                 /*
 278                  * exit_mmap will block in mmu_notifier_release to
 279                  * guarantee ->release is called before freeing the
 280                  * pages.
 281                  */
 282                 if (mn->ops->release)
 283                         mn->ops->release(mn, mm);
 284                 rcu_read_unlock();
 285         } else
 286                 spin_unlock(&mm->mmu_notifier_mm->lock);
 287
 288         /*
 289          * Wait any running method to finish, of course including
 290          * ->release if it was run by mmu_notifier_relase instead of us.
 291          */
 292         synchronize_rcu();
 293
 294         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 295
 296         mmdrop(mm);
 297 }
 298 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);