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 allows mmu_notifier::release callback to delay a call to
  27  * a function that will free appropriate resources. The function must be
  28  * quick and must not block.
  29  */
  30 void mmu_notifier_call_srcu(struct rcu_head *rcu,
  31                             void (*func)(struct rcu_head *rcu))
  32 {
  33         call_srcu(&srcu, rcu, func);
  34 }
  35 EXPORT_SYMBOL_GPL(mmu_notifier_call_srcu);
  36
  37 void mmu_notifier_synchronize(void)
  38 {
  39         /* Wait for any running method to finish. */
  40         srcu_barrier(&srcu);
  41 }
  42 EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
  43
  44 /*
  45  * This function can't run concurrently against mmu_notifier_register
  46  * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
  47  * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
  48  * in parallel despite there being no task using this mm any more,
  49  * through the vmas outside of the exit_mmap context, such as with
  50  * vmtruncate. This serializes against mmu_notifier_unregister with
  51  * the mmu_notifier_mm->lock in addition to SRCU and it serializes
  52  * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
  53  * can't go away from under us as exit_mmap holds an mm_count pin
  54  * itself.
  55  */
  56 void __mmu_notifier_release(struct mm_struct *mm)
  57 {
  58         struct mmu_notifier *mn;
  59         int id;
  60
  61         /*
  62          * SRCU here will block mmu_notifier_unregister until
  63          * ->release returns.
  64          */
  65         id = srcu_read_lock(&srcu);
  66         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
  67                 /*
  68                  * If ->release runs before mmu_notifier_unregister it must be
  69                  * handled, as it's the only way for the driver to flush all
  70                  * existing sptes and stop the driver from establishing any more
  71                  * sptes before all the pages in the mm are freed.
  72                  */
  73                 if (mn->ops->release)
  74                         mn->ops->release(mn, mm);
  75
  76         spin_lock(&mm->mmu_notifier_mm->lock);
  77         while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
  78                 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
  79                                  struct mmu_notifier,
  80                                  hlist);
  81                 /*
  82                  * We arrived before mmu_notifier_unregister so
  83                  * mmu_notifier_unregister will do nothing other than to wait
  84                  * for ->release to finish and for mmu_notifier_unregister to
  85                  * return.
  86                  */
  87                 hlist_del_init_rcu(&mn->hlist);
  88         }
  89         spin_unlock(&mm->mmu_notifier_mm->lock);
  90         srcu_read_unlock(&srcu, id);
  91
  92         /*
  93          * synchronize_srcu here prevents mmu_notifier_release from returning to
  94          * exit_mmap (which would proceed with freeing all pages in the mm)
  95          * until the ->release method returns, if it was invoked by
  96          * mmu_notifier_unregister.
  97          *
  98          * The mmu_notifier_mm can't go away from under us because one mm_count
  99          * is held by exit_mmap.
 100          */
 101         synchronize_srcu(&srcu);
 102 }
 103
 104 /*
 105  * If no young bitflag is supported by the hardware, ->clear_flush_young can
 106  * unmap the address and return 1 or 0 depending if the mapping previously
 107  * existed or not.
 108  */
 109 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
 110                                         unsigned long address)
 111 {
 112         struct mmu_notifier *mn;
 113         int young = 0, id;
 114
 115         id = srcu_read_lock(&srcu);
 116         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 117                 if (mn->ops->clear_flush_young)
 118                         young |= mn->ops->clear_flush_young(mn, mm, address);
 119         }
 120         srcu_read_unlock(&srcu, id);
 121
 122         return young;
 123 }
 124
 125 int __mmu_notifier_test_young(struct mm_struct *mm,
 126                               unsigned long address)
 127 {
 128         struct mmu_notifier *mn;
 129         int young = 0, id;
 130
 131         id = srcu_read_lock(&srcu);
 132         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 133                 if (mn->ops->test_young) {
 134                         young = mn->ops->test_young(mn, mm, address);
 135                         if (young)
 136                                 break;
 137                 }
 138         }
 139         srcu_read_unlock(&srcu, id);
 140
 141         return young;
 142 }
 143
 144 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 145                                pte_t pte)
 146 {
 147         struct mmu_notifier *mn;
 148         int id;
 149
 150         id = srcu_read_lock(&srcu);
 151         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 152                 if (mn->ops->change_pte)
 153                         mn->ops->change_pte(mn, mm, address, pte);
 154         }
 155         srcu_read_unlock(&srcu, id);
 156 }
 157
 158 void __mmu_notifier_invalidate_page(struct mm_struct *mm,
 159                                           unsigned long address)
 160 {
 161         struct mmu_notifier *mn;
 162         int id;
 163
 164         id = srcu_read_lock(&srcu);
 165         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 166                 if (mn->ops->invalidate_page)
 167                         mn->ops->invalidate_page(mn, mm, address);
 168         }
 169         srcu_read_unlock(&srcu, id);
 170 }
 171
 172 void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 173                                   unsigned long start, unsigned long end)
 174 {
 175         struct mmu_notifier *mn;
 176         int id;
 177
 178         id = srcu_read_lock(&srcu);
 179         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 180                 if (mn->ops->invalidate_range_start)
 181                         mn->ops->invalidate_range_start(mn, mm, start, end);
 182         }
 183         srcu_read_unlock(&srcu, id);
 184 }
 185 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
 186
 187 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 188                                   unsigned long start, unsigned long end)
 189 {
 190         struct mmu_notifier *mn;
 191         int id;
 192
 193         id = srcu_read_lock(&srcu);
 194         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 195                 if (mn->ops->invalidate_range_end)
 196                         mn->ops->invalidate_range_end(mn, mm, start, end);
 197         }
 198         srcu_read_unlock(&srcu, id);
 199 }
 200 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
 201
 202 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 203                                     struct mm_struct *mm,
 204                                     int take_mmap_sem)
 205 {
 206         struct mmu_notifier_mm *mmu_notifier_mm;
 207         int ret;
 208
 209         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 210
 211         /*
 212          * Verify that mmu_notifier_init() already run and the global srcu is
 213          * initialized.
 214          */
 215         BUG_ON(!srcu.per_cpu_ref);
 216
 217         ret = -ENOMEM;
 218         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 219         if (unlikely(!mmu_notifier_mm))
 220                 goto out;
 221
 222         if (take_mmap_sem)
 223                 down_write(&mm->mmap_sem);
 224         ret = mm_take_all_locks(mm);
 225         if (unlikely(ret))
 226                 goto out_clean;
 227
 228         if (!mm_has_notifiers(mm)) {
 229                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 230                 spin_lock_init(&mmu_notifier_mm->lock);
 231
 232                 mm->mmu_notifier_mm = mmu_notifier_mm;
 233                 mmu_notifier_mm = NULL;
 234         }
 235         atomic_inc(&mm->mm_count);
 236
 237         /*
 238          * Serialize the update against mmu_notifier_unregister. A
 239          * side note: mmu_notifier_release can't run concurrently with
 240          * us because we hold the mm_users pin (either implicitly as
 241          * current->mm or explicitly with get_task_mm() or similar).
 242          * We can't race against any other mmu notifier method either
 243          * thanks to mm_take_all_locks().
 244          */
 245         spin_lock(&mm->mmu_notifier_mm->lock);
 246         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 247         spin_unlock(&mm->mmu_notifier_mm->lock);
 248
 249         mm_drop_all_locks(mm);
 250 out_clean:
 251         if (take_mmap_sem)
 252                 up_write(&mm->mmap_sem);
 253         kfree(mmu_notifier_mm);
 254 out:
 255         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 256         return ret;
 257 }
 258
 259 /*
 260  * Must not hold mmap_sem nor any other VM related lock when calling
 261  * this registration function. Must also ensure mm_users can't go down
 262  * to zero while this runs to avoid races with mmu_notifier_release,
 263  * so mm has to be current->mm or the mm should be pinned safely such
 264  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 265  * pin should be released by calling mmput after mmu_notifier_register
 266  * returns. mmu_notifier_unregister must be always called to
 267  * unregister the notifier. mm_count is automatically pinned to allow
 268  * mmu_notifier_unregister to safely run at any time later, before or
 269  * after exit_mmap. ->release will always be called before exit_mmap
 270  * frees the pages.
 271  */
 272 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 273 {
 274         return do_mmu_notifier_register(mn, mm, 1);
 275 }
 276 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 277
 278 /*
 279  * Same as mmu_notifier_register but here the caller must hold the
 280  * mmap_sem in write mode.
 281  */
 282 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 283 {
 284         return do_mmu_notifier_register(mn, mm, 0);
 285 }
 286 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 287
 288 /* this is called after the last mmu_notifier_unregister() returned */
 289 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 290 {
 291         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 292         kfree(mm->mmu_notifier_mm);
 293         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 294 }
 295
 296 /*
 297  * This releases the mm_count pin automatically and frees the mm
 298  * structure if it was the last user of it. It serializes against
 299  * running mmu notifiers with SRCU and against mmu_notifier_unregister
 300  * with the unregister lock + SRCU. All sptes must be dropped before
 301  * calling mmu_notifier_unregister. ->release or any other notifier
 302  * method may be invoked concurrently with mmu_notifier_unregister,
 303  * and only after mmu_notifier_unregister returned we're guaranteed
 304  * that ->release or any other method can't run anymore.
 305  */
 306 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 307 {
 308         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 309
 310         if (!hlist_unhashed(&mn->hlist)) {
 311                 /*
 312                  * SRCU here will force exit_mmap to wait for ->release to
 313                  * finish before freeing the pages.
 314                  */
 315                 int id;
 316
 317                 id = srcu_read_lock(&srcu);
 318                 /*
 319                  * exit_mmap will block in mmu_notifier_release to guarantee
 320                  * that ->release is called before freeing the pages.
 321                  */
 322                 if (mn->ops->release)
 323                         mn->ops->release(mn, mm);
 324                 srcu_read_unlock(&srcu, id);
 325
 326                 spin_lock(&mm->mmu_notifier_mm->lock);
 327                 /*
 328                  * Can not use list_del_rcu() since __mmu_notifier_release
 329                  * can delete it before we hold the lock.
 330                  */
 331                 hlist_del_init_rcu(&mn->hlist);
 332                 spin_unlock(&mm->mmu_notifier_mm->lock);
 333         }
 334
 335         /*
 336          * Wait for any running method to finish, of course including
 337          * ->release if it was run by mmu_notifier_release instead of us.
 338          */
 339         synchronize_srcu(&srcu);
 340
 341         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 342
 343         mmdrop(mm);
 344 }
 345 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
 346
 347 /*
 348  * Same as mmu_notifier_unregister but no callback and no srcu synchronization.
 349  */
 350 void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
 351                                         struct mm_struct *mm)
 352 {
 353         spin_lock(&mm->mmu_notifier_mm->lock);
 354         /*
 355          * Can not use list_del_rcu() since __mmu_notifier_release
 356          * can delete it before we hold the lock.
 357          */
 358         hlist_del_init_rcu(&mn->hlist);
 359         spin_unlock(&mm->mmu_notifier_mm->lock);
 360
 361         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 362         mmdrop(mm);
 363 }
 364 EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release);
 365
 366 static int __init mmu_notifier_init(void)
 367 {
 368         return init_srcu_struct(&srcu);
 369 }
 370 subsys_initcall(mmu_notifier_init);