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 <cl@linux.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/sched/mm.h>
  21 #include <linux/slab.h>
  22
  23 /* global SRCU for all MMs */
  24 DEFINE_STATIC_SRCU(srcu);
  25
  26 /*
  27  * This function allows mmu_notifier::release callback to delay a call to
  28  * a function that will free appropriate resources. The function must be
  29  * quick and must not block.
  30  */
  31 void mmu_notifier_call_srcu(struct rcu_head *rcu,
  32                             void (*func)(struct rcu_head *rcu))
  33 {
  34         call_srcu(&srcu, rcu, func);
  35 }
  36 EXPORT_SYMBOL_GPL(mmu_notifier_call_srcu);
  37
  38 void mmu_notifier_synchronize(void)
  39 {
  40         /* Wait for any running method to finish. */
  41         srcu_barrier(&srcu);
  42 }
  43 EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
  44
  45 /*
  46  * This function can't run concurrently against mmu_notifier_register
  47  * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
  48  * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
  49  * in parallel despite there being no task using this mm any more,
  50  * through the vmas outside of the exit_mmap context, such as with
  51  * vmtruncate. This serializes against mmu_notifier_unregister with
  52  * the mmu_notifier_mm->lock in addition to SRCU and it serializes
  53  * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
  54  * can't go away from under us as exit_mmap holds an mm_count pin
  55  * itself.
  56  */
  57 void __mmu_notifier_release(struct mm_struct *mm)
  58 {
  59         struct mmu_notifier *mn;
  60         int id;
  61
  62         /*
  63          * SRCU here will block mmu_notifier_unregister until
  64          * ->release returns.
  65          */
  66         id = srcu_read_lock(&srcu);
  67         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
  68                 /*
  69                  * If ->release runs before mmu_notifier_unregister it must be
  70                  * handled, as it's the only way for the driver to flush all
  71                  * existing sptes and stop the driver from establishing any more
  72                  * sptes before all the pages in the mm are freed.
  73                  */
  74                 if (mn->ops->release)
  75                         mn->ops->release(mn, mm);
  76
  77         spin_lock(&mm->mmu_notifier_mm->lock);
  78         while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
  79                 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
  80                                  struct mmu_notifier,
  81                                  hlist);
  82                 /*
  83                  * We arrived before mmu_notifier_unregister so
  84                  * mmu_notifier_unregister will do nothing other than to wait
  85                  * for ->release to finish and for mmu_notifier_unregister to
  86                  * return.
  87                  */
  88                 hlist_del_init_rcu(&mn->hlist);
  89         }
  90         spin_unlock(&mm->mmu_notifier_mm->lock);
  91         srcu_read_unlock(&srcu, id);
  92
  93         /*
  94          * synchronize_srcu here prevents mmu_notifier_release from returning to
  95          * exit_mmap (which would proceed with freeing all pages in the mm)
  96          * until the ->release method returns, if it was invoked by
  97          * mmu_notifier_unregister.
  98          *
  99          * The mmu_notifier_mm can't go away from under us because one mm_count
 100          * is held by exit_mmap.
 101          */
 102         synchronize_srcu(&srcu);
 103 }
 104
 105 /*
 106  * If no young bitflag is supported by the hardware, ->clear_flush_young can
 107  * unmap the address and return 1 or 0 depending if the mapping previously
 108  * existed or not.
 109  */
 110 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
 111                                         unsigned long start,
 112                                         unsigned long end)
 113 {
 114         struct mmu_notifier *mn;
 115         int young = 0, id;
 116
 117         id = srcu_read_lock(&srcu);
 118         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 119                 if (mn->ops->clear_flush_young)
 120                         young |= mn->ops->clear_flush_young(mn, mm, start, end);
 121         }
 122         srcu_read_unlock(&srcu, id);
 123
 124         return young;
 125 }
 126
 127 int __mmu_notifier_clear_young(struct mm_struct *mm,
 128                                unsigned long start,
 129                                unsigned long end)
 130 {
 131         struct mmu_notifier *mn;
 132         int young = 0, id;
 133
 134         id = srcu_read_lock(&srcu);
 135         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 136                 if (mn->ops->clear_young)
 137                         young |= mn->ops->clear_young(mn, mm, start, end);
 138         }
 139         srcu_read_unlock(&srcu, id);
 140
 141         return young;
 142 }
 143
 144 int __mmu_notifier_test_young(struct mm_struct *mm,
 145                               unsigned long address)
 146 {
 147         struct mmu_notifier *mn;
 148         int young = 0, 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->test_young) {
 153                         young = mn->ops->test_young(mn, mm, address);
 154                         if (young)
 155                                 break;
 156                 }
 157         }
 158         srcu_read_unlock(&srcu, id);
 159
 160         return young;
 161 }
 162
 163 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 164                                pte_t pte)
 165 {
 166         struct mmu_notifier *mn;
 167         int id;
 168
 169         id = srcu_read_lock(&srcu);
 170         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 171                 if (mn->ops->change_pte)
 172                         mn->ops->change_pte(mn, mm, address, pte);
 173         }
 174         srcu_read_unlock(&srcu, id);
 175 }
 176
 177 void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 178                                   unsigned long start, unsigned long end)
 179 {
 180         struct mmu_notifier *mn;
 181         int id;
 182
 183         id = srcu_read_lock(&srcu);
 184         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 185                 if (mn->ops->invalidate_range_start)
 186                         mn->ops->invalidate_range_start(mn, mm, start, end);
 187         }
 188         srcu_read_unlock(&srcu, id);
 189 }
 190 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
 191
 192 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 193                                          unsigned long start,
 194                                          unsigned long end,
 195                                          bool only_end)
 196 {
 197         struct mmu_notifier *mn;
 198         int id;
 199
 200         id = srcu_read_lock(&srcu);
 201         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 202                 /*
 203                  * Call invalidate_range here too to avoid the need for the
 204                  * subsystem of having to register an invalidate_range_end
 205                  * call-back when there is invalidate_range already. Usually a
 206                  * subsystem registers either invalidate_range_start()/end() or
 207                  * invalidate_range(), so this will be no additional overhead
 208                  * (besides the pointer check).
 209                  *
 210                  * We skip call to invalidate_range() if we know it is safe ie
 211                  * call site use mmu_notifier_invalidate_range_only_end() which
 212                  * is safe to do when we know that a call to invalidate_range()
 213                  * already happen under page table lock.
 214                  */
 215                 if (!only_end && mn->ops->invalidate_range)
 216                         mn->ops->invalidate_range(mn, mm, start, end);
 217                 if (mn->ops->invalidate_range_end)
 218                         mn->ops->invalidate_range_end(mn, mm, start, end);
 219         }
 220         srcu_read_unlock(&srcu, id);
 221 }
 222 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
 223
 224 void __mmu_notifier_invalidate_range(struct mm_struct *mm,
 225                                   unsigned long start, unsigned long end)
 226 {
 227         struct mmu_notifier *mn;
 228         int id;
 229
 230         id = srcu_read_lock(&srcu);
 231         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 232                 if (mn->ops->invalidate_range)
 233                         mn->ops->invalidate_range(mn, mm, start, end);
 234         }
 235         srcu_read_unlock(&srcu, id);
 236 }
 237 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range);
 238
 239 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 240                                     struct mm_struct *mm,
 241                                     int take_mmap_sem)
 242 {
 243         struct mmu_notifier_mm *mmu_notifier_mm;
 244         int ret;
 245
 246         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 247
 248         ret = -ENOMEM;
 249         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 250         if (unlikely(!mmu_notifier_mm))
 251                 goto out;
 252
 253         if (take_mmap_sem)
 254                 down_write(&mm->mmap_sem);
 255         ret = mm_take_all_locks(mm);
 256         if (unlikely(ret))
 257                 goto out_clean;
 258
 259         if (!mm_has_notifiers(mm)) {
 260                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 261                 spin_lock_init(&mmu_notifier_mm->lock);
 262
 263                 mm->mmu_notifier_mm = mmu_notifier_mm;
 264                 mmu_notifier_mm = NULL;
 265         }
 266         mmgrab(mm);
 267
 268         /*
 269          * Serialize the update against mmu_notifier_unregister. A
 270          * side note: mmu_notifier_release can't run concurrently with
 271          * us because we hold the mm_users pin (either implicitly as
 272          * current->mm or explicitly with get_task_mm() or similar).
 273          * We can't race against any other mmu notifier method either
 274          * thanks to mm_take_all_locks().
 275          */
 276         spin_lock(&mm->mmu_notifier_mm->lock);
 277         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 278         spin_unlock(&mm->mmu_notifier_mm->lock);
 279
 280         mm_drop_all_locks(mm);
 281 out_clean:
 282         if (take_mmap_sem)
 283                 up_write(&mm->mmap_sem);
 284         kfree(mmu_notifier_mm);
 285 out:
 286         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 287         return ret;
 288 }
 289
 290 /*
 291  * Must not hold mmap_sem nor any other VM related lock when calling
 292  * this registration function. Must also ensure mm_users can't go down
 293  * to zero while this runs to avoid races with mmu_notifier_release,
 294  * so mm has to be current->mm or the mm should be pinned safely such
 295  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 296  * pin should be released by calling mmput after mmu_notifier_register
 297  * returns. mmu_notifier_unregister must be always called to
 298  * unregister the notifier. mm_count is automatically pinned to allow
 299  * mmu_notifier_unregister to safely run at any time later, before or
 300  * after exit_mmap. ->release will always be called before exit_mmap
 301  * frees the pages.
 302  */
 303 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 304 {
 305         return do_mmu_notifier_register(mn, mm, 1);
 306 }
 307 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 308
 309 /*
 310  * Same as mmu_notifier_register but here the caller must hold the
 311  * mmap_sem in write mode.
 312  */
 313 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 314 {
 315         return do_mmu_notifier_register(mn, mm, 0);
 316 }
 317 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 318
 319 /* this is called after the last mmu_notifier_unregister() returned */
 320 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 321 {
 322         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 323         kfree(mm->mmu_notifier_mm);
 324         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 325 }
 326
 327 /*
 328  * This releases the mm_count pin automatically and frees the mm
 329  * structure if it was the last user of it. It serializes against
 330  * running mmu notifiers with SRCU and against mmu_notifier_unregister
 331  * with the unregister lock + SRCU. All sptes must be dropped before
 332  * calling mmu_notifier_unregister. ->release or any other notifier
 333  * method may be invoked concurrently with mmu_notifier_unregister,
 334  * and only after mmu_notifier_unregister returned we're guaranteed
 335  * that ->release or any other method can't run anymore.
 336  */
 337 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 338 {
 339         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 340
 341         if (!hlist_unhashed(&mn->hlist)) {
 342                 /*
 343                  * SRCU here will force exit_mmap to wait for ->release to
 344                  * finish before freeing the pages.
 345                  */
 346                 int id;
 347
 348                 id = srcu_read_lock(&srcu);
 349                 /*
 350                  * exit_mmap will block in mmu_notifier_release to guarantee
 351                  * that ->release is called before freeing the pages.
 352                  */
 353                 if (mn->ops->release)
 354                         mn->ops->release(mn, mm);
 355                 srcu_read_unlock(&srcu, id);
 356
 357                 spin_lock(&mm->mmu_notifier_mm->lock);
 358                 /*
 359                  * Can not use list_del_rcu() since __mmu_notifier_release
 360                  * can delete it before we hold the lock.
 361                  */
 362                 hlist_del_init_rcu(&mn->hlist);
 363                 spin_unlock(&mm->mmu_notifier_mm->lock);
 364         }
 365
 366         /*
 367          * Wait for any running method to finish, of course including
 368          * ->release if it was run by mmu_notifier_release instead of us.
 369          */
 370         synchronize_srcu(&srcu);
 371
 372         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 373
 374         mmdrop(mm);
 375 }
 376 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
 377
 378 /*
 379  * Same as mmu_notifier_unregister but no callback and no srcu synchronization.
 380  */
 381 void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
 382                                         struct mm_struct *mm)
 383 {
 384         spin_lock(&mm->mmu_notifier_mm->lock);
 385         /*
 386          * Can not use list_del_rcu() since __mmu_notifier_release
 387          * can delete it before we hold the lock.
 388          */
 389         hlist_del_init_rcu(&mn->hlist);
 390         spin_unlock(&mm->mmu_notifier_mm->lock);
 391
 392         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 393         mmdrop(mm);
 394 }
 395 EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release);