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