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