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         struct hlist_node *n;
  41         int id;
  42
  43         /*
  44          * SRCU here will block mmu_notifier_unregister until
  45          * ->release returns.
  46          */
  47         id = srcu_read_lock(&srcu);
  48         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist)
  49                 /*
  50                  * if ->release runs before mmu_notifier_unregister it
  51                  * must be handled as it's the only way for the driver
  52                  * to flush all existing sptes and stop the driver
  53                  * from establishing any more sptes before all the
  54                  * pages in the mm are freed.
  55                  */
  56                 if (mn->ops->release)
  57                         mn->ops->release(mn, mm);
  58         srcu_read_unlock(&srcu, id);
  59
  60         spin_lock(&mm->mmu_notifier_mm->lock);
  61         while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
  62                 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
  63                                  struct mmu_notifier,
  64                                  hlist);
  65                 /*
  66                  * We arrived before mmu_notifier_unregister so
  67                  * mmu_notifier_unregister will do nothing other than
  68                  * to wait ->release to finish and
  69                  * mmu_notifier_unregister to return.
  70                  */
  71                 hlist_del_init_rcu(&mn->hlist);
  72         }
  73         spin_unlock(&mm->mmu_notifier_mm->lock);
  74
  75         /*
  76          * synchronize_srcu here prevents mmu_notifier_release to
  77          * return to exit_mmap (which would proceed freeing all pages
  78          * in the mm) until the ->release method returns, if it was
  79          * invoked by mmu_notifier_unregister.
  80          *
  81          * The mmu_notifier_mm can't go away from under us because one
  82          * mm_count is hold by exit_mmap.
  83          */
  84         synchronize_srcu(&srcu);
  85 }
  86
  87 /*
  88  * If no young bitflag is supported by the hardware, ->clear_flush_young can
  89  * unmap the address and return 1 or 0 depending if the mapping previously
  90  * existed or not.
  91  */
  92 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
  93                                         unsigned long address)
  94 {
  95         struct mmu_notifier *mn;
  96         struct hlist_node *n;
  97         int young = 0, id;
  98
  99         id = srcu_read_lock(&srcu);
 100         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 101                 if (mn->ops->clear_flush_young)
 102                         young |= mn->ops->clear_flush_young(mn, mm, address);
 103         }
 104         srcu_read_unlock(&srcu, id);
 105
 106         return young;
 107 }
 108
 109 int __mmu_notifier_test_young(struct mm_struct *mm,
 110                               unsigned long address)
 111 {
 112         struct mmu_notifier *mn;
 113         struct hlist_node *n;
 114         int young = 0, id;
 115
 116         id = srcu_read_lock(&srcu);
 117         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 118                 if (mn->ops->test_young) {
 119                         young = mn->ops->test_young(mn, mm, address);
 120                         if (young)
 121                                 break;
 122                 }
 123         }
 124         srcu_read_unlock(&srcu, id);
 125
 126         return young;
 127 }
 128
 129 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 130                                pte_t pte)
 131 {
 132         struct mmu_notifier *mn;
 133         struct hlist_node *n;
 134         int id;
 135
 136         id = srcu_read_lock(&srcu);
 137         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 138                 if (mn->ops->change_pte)
 139                         mn->ops->change_pte(mn, mm, address, pte);
 140         }
 141         srcu_read_unlock(&srcu, id);
 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         int id;
 150
 151         id = srcu_read_lock(&srcu);
 152         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 153                 if (mn->ops->invalidate_page)
 154                         mn->ops->invalidate_page(mn, mm, address);
 155         }
 156         srcu_read_unlock(&srcu, id);
 157 }
 158
 159 void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 160                                   unsigned long start, unsigned long end)
 161 {
 162         struct mmu_notifier *mn;
 163         struct hlist_node *n;
 164         int id;
 165
 166         id = srcu_read_lock(&srcu);
 167         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 168                 if (mn->ops->invalidate_range_start)
 169                         mn->ops->invalidate_range_start(mn, mm, start, end);
 170         }
 171         srcu_read_unlock(&srcu, id);
 172 }
 173
 174 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 175                                   unsigned long start, unsigned long end)
 176 {
 177         struct mmu_notifier *mn;
 178         struct hlist_node *n;
 179         int id;
 180
 181         id = srcu_read_lock(&srcu);
 182         hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 183                 if (mn->ops->invalidate_range_end)
 184                         mn->ops->invalidate_range_end(mn, mm, start, end);
 185         }
 186         srcu_read_unlock(&srcu, id);
 187 }
 188
 189 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 190                                     struct mm_struct *mm,
 191                                     int take_mmap_sem)
 192 {
 193         struct mmu_notifier_mm *mmu_notifier_mm;
 194         int ret;
 195
 196         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 197
 198         /*
 199          * Verify that mmu_notifier_init() already run and the global srcu is
 200          * initialized.
 201          */
 202         BUG_ON(!srcu.per_cpu_ref);
 203
 204         ret = -ENOMEM;
 205         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 206         if (unlikely(!mmu_notifier_mm))
 207                 goto out;
 208
 209         if (take_mmap_sem)
 210                 down_write(&mm->mmap_sem);
 211         ret = mm_take_all_locks(mm);
 212         if (unlikely(ret))
 213                 goto out_clean;
 214
 215         if (!mm_has_notifiers(mm)) {
 216                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 217                 spin_lock_init(&mmu_notifier_mm->lock);
 218
 219                 mm->mmu_notifier_mm = mmu_notifier_mm;
 220                 mmu_notifier_mm = NULL;
 221         }
 222         atomic_inc(&mm->mm_count);
 223
 224         /*
 225          * Serialize the update against mmu_notifier_unregister. A
 226          * side note: mmu_notifier_release can't run concurrently with
 227          * us because we hold the mm_users pin (either implicitly as
 228          * current->mm or explicitly with get_task_mm() or similar).
 229          * We can't race against any other mmu notifier method either
 230          * thanks to mm_take_all_locks().
 231          */
 232         spin_lock(&mm->mmu_notifier_mm->lock);
 233         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 234         spin_unlock(&mm->mmu_notifier_mm->lock);
 235
 236         mm_drop_all_locks(mm);
 237 out_clean:
 238         if (take_mmap_sem)
 239                 up_write(&mm->mmap_sem);
 240         kfree(mmu_notifier_mm);
 241 out:
 242         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 243         return ret;
 244 }
 245
 246 /*
 247  * Must not hold mmap_sem nor any other VM related lock when calling
 248  * this registration function. Must also ensure mm_users can't go down
 249  * to zero while this runs to avoid races with mmu_notifier_release,
 250  * so mm has to be current->mm or the mm should be pinned safely such
 251  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 252  * pin should be released by calling mmput after mmu_notifier_register
 253  * returns. mmu_notifier_unregister must be always called to
 254  * unregister the notifier. mm_count is automatically pinned to allow
 255  * mmu_notifier_unregister to safely run at any time later, before or
 256  * after exit_mmap. ->release will always be called before exit_mmap
 257  * frees the pages.
 258  */
 259 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 260 {
 261         return do_mmu_notifier_register(mn, mm, 1);
 262 }
 263 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 264
 265 /*
 266  * Same as mmu_notifier_register but here the caller must hold the
 267  * mmap_sem in write mode.
 268  */
 269 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 270 {
 271         return do_mmu_notifier_register(mn, mm, 0);
 272 }
 273 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 274
 275 /* this is called after the last mmu_notifier_unregister() returned */
 276 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 277 {
 278         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 279         kfree(mm->mmu_notifier_mm);
 280         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 281 }
 282
 283 /*
 284  * This releases the mm_count pin automatically and frees the mm
 285  * structure if it was the last user of it. It serializes against
 286  * running mmu notifiers with SRCU and against mmu_notifier_unregister
 287  * with the unregister lock + SRCU. All sptes must be dropped before
 288  * calling mmu_notifier_unregister. ->release or any other notifier
 289  * method may be invoked concurrently with mmu_notifier_unregister,
 290  * and only after mmu_notifier_unregister returned we're guaranteed
 291  * that ->release or any other method can't run anymore.
 292  */
 293 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 294 {
 295         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 296
 297         if (!hlist_unhashed(&mn->hlist)) {
 298                 /*
 299                  * SRCU here will force exit_mmap to wait ->release to finish
 300                  * before freeing the pages.
 301                  */
 302                 int id;
 303
 304                 id = srcu_read_lock(&srcu);
 305                 /*
 306                  * exit_mmap will block in mmu_notifier_release to
 307                  * guarantee ->release is called before freeing the
 308                  * pages.
 309                  */
 310                 if (mn->ops->release)
 311                         mn->ops->release(mn, mm);
 312                 srcu_read_unlock(&srcu, id);
 313
 314                 spin_lock(&mm->mmu_notifier_mm->lock);
 315                 hlist_del_rcu(&mn->hlist);
 316                 spin_unlock(&mm->mmu_notifier_mm->lock);
 317         }
 318
 319         /*
 320          * Wait any running method to finish, of course including
 321          * ->release if it was run by mmu_notifier_relase instead of us.
 322          */
 323         synchronize_srcu(&srcu);
 324
 325         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 326
 327         mmdrop(mm);
 328 }
 329 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
 330
 331 static int __init mmu_notifier_init(void)
 332 {
 333         return init_srcu_struct(&srcu);
 334 }
 335
 336 module_init(mmu_notifier_init);