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/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 start,
 111                                         unsigned long end)
 112 {
 113         struct mmu_notifier *mn;
 114         int young = 0, id;
 115
 116         id = srcu_read_lock(&srcu);
 117         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 118                 if (mn->ops->clear_flush_young)
 119                         young |= mn->ops->clear_flush_young(mn, mm, start, end);
 120         }
 121         srcu_read_unlock(&srcu, id);
 122
 123         return young;
 124 }
 125
 126 int __mmu_notifier_clear_young(struct mm_struct *mm,
 127                                unsigned long start,
 128                                unsigned long end)
 129 {
 130         struct mmu_notifier *mn;
 131         int young = 0, id;
 132
 133         id = srcu_read_lock(&srcu);
 134         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 135                 if (mn->ops->clear_young)
 136                         young |= mn->ops->clear_young(mn, mm, start, end);
 137         }
 138         srcu_read_unlock(&srcu, id);
 139
 140         return young;
 141 }
 142
 143 int __mmu_notifier_test_young(struct mm_struct *mm,
 144                               unsigned long address)
 145 {
 146         struct mmu_notifier *mn;
 147         int young = 0, id;
 148
 149         id = srcu_read_lock(&srcu);
 150         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 151                 if (mn->ops->test_young) {
 152                         young = mn->ops->test_young(mn, mm, address);
 153                         if (young)
 154                                 break;
 155                 }
 156         }
 157         srcu_read_unlock(&srcu, id);
 158
 159         return young;
 160 }
 161
 162 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 163                                pte_t pte)
 164 {
 165         struct mmu_notifier *mn;
 166         int id;
 167
 168         id = srcu_read_lock(&srcu);
 169         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 170                 if (mn->ops->change_pte)
 171                         mn->ops->change_pte(mn, mm, address, pte);
 172         }
 173         srcu_read_unlock(&srcu, id);
 174 }
 175
 176 void __mmu_notifier_invalidate_page(struct mm_struct *mm,
 177                                           unsigned long address)
 178 {
 179         struct mmu_notifier *mn;
 180         int id;
 181
 182         id = srcu_read_lock(&srcu);
 183         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 184                 if (mn->ops->invalidate_page)
 185                         mn->ops->invalidate_page(mn, mm, address);
 186         }
 187         srcu_read_unlock(&srcu, id);
 188 }
 189
 190 void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 191                                   unsigned long start, unsigned long end)
 192 {
 193         struct mmu_notifier *mn;
 194         int id;
 195
 196         id = srcu_read_lock(&srcu);
 197         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 198                 if (mn->ops->invalidate_range_start)
 199                         mn->ops->invalidate_range_start(mn, mm, start, end);
 200         }
 201         srcu_read_unlock(&srcu, id);
 202 }
 203 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
 204
 205 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 206                                   unsigned long start, unsigned long end)
 207 {
 208         struct mmu_notifier *mn;
 209         int id;
 210
 211         id = srcu_read_lock(&srcu);
 212         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 213                 /*
 214                  * Call invalidate_range here too to avoid the need for the
 215                  * subsystem of having to register an invalidate_range_end
 216                  * call-back when there is invalidate_range already. Usually a
 217                  * subsystem registers either invalidate_range_start()/end() or
 218                  * invalidate_range(), so this will be no additional overhead
 219                  * (besides the pointer check).
 220                  */
 221                 if (mn->ops->invalidate_range)
 222                         mn->ops->invalidate_range(mn, mm, start, end);
 223                 if (mn->ops->invalidate_range_end)
 224                         mn->ops->invalidate_range_end(mn, mm, start, end);
 225         }
 226         srcu_read_unlock(&srcu, id);
 227 }
 228 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
 229
 230 void __mmu_notifier_invalidate_range(struct mm_struct *mm,
 231                                   unsigned long start, unsigned long end)
 232 {
 233         struct mmu_notifier *mn;
 234         int id;
 235
 236         id = srcu_read_lock(&srcu);
 237         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 238                 if (mn->ops->invalidate_range)
 239                         mn->ops->invalidate_range(mn, mm, start, end);
 240         }
 241         srcu_read_unlock(&srcu, id);
 242 }
 243 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range);
 244
 245 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 246                                     struct mm_struct *mm,
 247                                     int take_mmap_sem)
 248 {
 249         struct mmu_notifier_mm *mmu_notifier_mm;
 250         int ret;
 251
 252         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 253
 254         /*
 255          * Verify that mmu_notifier_init() already run and the global srcu is
 256          * initialized.
 257          */
 258         BUG_ON(!srcu.per_cpu_ref);
 259
 260         ret = -ENOMEM;
 261         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 262         if (unlikely(!mmu_notifier_mm))
 263                 goto out;
 264
 265         if (take_mmap_sem)
 266                 down_write(&mm->mmap_sem);
 267         ret = mm_take_all_locks(mm);
 268         if (unlikely(ret))
 269                 goto out_clean;
 270
 271         if (!mm_has_notifiers(mm)) {
 272                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 273                 spin_lock_init(&mmu_notifier_mm->lock);
 274
 275                 mm->mmu_notifier_mm = mmu_notifier_mm;
 276                 mmu_notifier_mm = NULL;
 277         }
 278         atomic_inc(&mm->mm_count);
 279
 280         /*
 281          * Serialize the update against mmu_notifier_unregister. A
 282          * side note: mmu_notifier_release can't run concurrently with
 283          * us because we hold the mm_users pin (either implicitly as
 284          * current->mm or explicitly with get_task_mm() or similar).
 285          * We can't race against any other mmu notifier method either
 286          * thanks to mm_take_all_locks().
 287          */
 288         spin_lock(&mm->mmu_notifier_mm->lock);
 289         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 290         spin_unlock(&mm->mmu_notifier_mm->lock);
 291
 292         mm_drop_all_locks(mm);
 293 out_clean:
 294         if (take_mmap_sem)
 295                 up_write(&mm->mmap_sem);
 296         kfree(mmu_notifier_mm);
 297 out:
 298         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 299         return ret;
 300 }
 301
 302 /*
 303  * Must not hold mmap_sem nor any other VM related lock when calling
 304  * this registration function. Must also ensure mm_users can't go down
 305  * to zero while this runs to avoid races with mmu_notifier_release,
 306  * so mm has to be current->mm or the mm should be pinned safely such
 307  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 308  * pin should be released by calling mmput after mmu_notifier_register
 309  * returns. mmu_notifier_unregister must be always called to
 310  * unregister the notifier. mm_count is automatically pinned to allow
 311  * mmu_notifier_unregister to safely run at any time later, before or
 312  * after exit_mmap. ->release will always be called before exit_mmap
 313  * frees the pages.
 314  */
 315 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 316 {
 317         return do_mmu_notifier_register(mn, mm, 1);
 318 }
 319 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 320
 321 /*
 322  * Same as mmu_notifier_register but here the caller must hold the
 323  * mmap_sem in write mode.
 324  */
 325 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 326 {
 327         return do_mmu_notifier_register(mn, mm, 0);
 328 }
 329 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 330
 331 /* this is called after the last mmu_notifier_unregister() returned */
 332 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 333 {
 334         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 335         kfree(mm->mmu_notifier_mm);
 336         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 337 }
 338
 339 /*
 340  * This releases the mm_count pin automatically and frees the mm
 341  * structure if it was the last user of it. It serializes against
 342  * running mmu notifiers with SRCU and against mmu_notifier_unregister
 343  * with the unregister lock + SRCU. All sptes must be dropped before
 344  * calling mmu_notifier_unregister. ->release or any other notifier
 345  * method may be invoked concurrently with mmu_notifier_unregister,
 346  * and only after mmu_notifier_unregister returned we're guaranteed
 347  * that ->release or any other method can't run anymore.
 348  */
 349 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 350 {
 351         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 352
 353         if (!hlist_unhashed(&mn->hlist)) {
 354                 /*
 355                  * SRCU here will force exit_mmap to wait for ->release to
 356                  * finish before freeing the pages.
 357                  */
 358                 int id;
 359
 360                 id = srcu_read_lock(&srcu);
 361                 /*
 362                  * exit_mmap will block in mmu_notifier_release to guarantee
 363                  * that ->release is called before freeing the pages.
 364                  */
 365                 if (mn->ops->release)
 366                         mn->ops->release(mn, mm);
 367                 srcu_read_unlock(&srcu, id);
 368
 369                 spin_lock(&mm->mmu_notifier_mm->lock);
 370                 /*
 371                  * Can not use list_del_rcu() since __mmu_notifier_release
 372                  * can delete it before we hold the lock.
 373                  */
 374                 hlist_del_init_rcu(&mn->hlist);
 375                 spin_unlock(&mm->mmu_notifier_mm->lock);
 376         }
 377
 378         /*
 379          * Wait for any running method to finish, of course including
 380          * ->release if it was run by mmu_notifier_release instead of us.
 381          */
 382         synchronize_srcu(&srcu);
 383
 384         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 385
 386         mmdrop(mm);
 387 }
 388 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
 389
 390 /*
 391  * Same as mmu_notifier_unregister but no callback and no srcu synchronization.
 392  */
 393 void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
 394                                         struct mm_struct *mm)
 395 {
 396         spin_lock(&mm->mmu_notifier_mm->lock);
 397         /*
 398          * Can not use list_del_rcu() since __mmu_notifier_release
 399          * can delete it before we hold the lock.
 400          */
 401         hlist_del_init_rcu(&mn->hlist);
 402         spin_unlock(&mm->mmu_notifier_mm->lock);
 403
 404         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 405         mmdrop(mm);
 406 }
 407 EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release);
 408
 409 static int __init mmu_notifier_init(void)
 410 {
 411         return init_srcu_struct(&srcu);
 412 }
 413 subsys_initcall(mmu_notifier_init);