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