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 /*
  39  * This function can't run concurrently against mmu_notifier_register
  40  * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
  41  * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
  42  * in parallel despite there being no task using this mm any more,
  43  * through the vmas outside of the exit_mmap context, such as with
  44  * vmtruncate. This serializes against mmu_notifier_unregister with
  45  * the mmu_notifier_mm->lock in addition to SRCU and it serializes
  46  * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
  47  * can't go away from under us as exit_mmap holds an mm_count pin
  48  * itself.
  49  */
  50 void __mmu_notifier_release(struct mm_struct *mm)
  51 {
  52         struct mmu_notifier *mn;
  53         int id;
  54
  55         /*
  56          * SRCU here will block mmu_notifier_unregister until
  57          * ->release returns.
  58          */
  59         id = srcu_read_lock(&srcu);
  60         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
  61                 /*
  62                  * If ->release runs before mmu_notifier_unregister it must be
  63                  * handled, as it's the only way for the driver to flush all
  64                  * existing sptes and stop the driver from establishing any more
  65                  * sptes before all the pages in the mm are freed.
  66                  */
  67                 if (mn->ops->release)
  68                         mn->ops->release(mn, mm);
  69
  70         spin_lock(&mm->mmu_notifier_mm->lock);
  71         while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
  72                 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
  73                                  struct mmu_notifier,
  74                                  hlist);
  75                 /*
  76                  * We arrived before mmu_notifier_unregister so
  77                  * mmu_notifier_unregister will do nothing other than to wait
  78                  * for ->release to finish and for mmu_notifier_unregister to
  79                  * return.
  80                  */
  81                 hlist_del_init_rcu(&mn->hlist);
  82         }
  83         spin_unlock(&mm->mmu_notifier_mm->lock);
  84         srcu_read_unlock(&srcu, id);
  85
  86         /*
  87          * synchronize_srcu here prevents mmu_notifier_release from returning to
  88          * exit_mmap (which would proceed with freeing all pages in the mm)
  89          * until the ->release method returns, if it was invoked by
  90          * mmu_notifier_unregister.
  91          *
  92          * The mmu_notifier_mm can't go away from under us because one mm_count
  93          * is held by exit_mmap.
  94          */
  95         synchronize_srcu(&srcu);
  96 }
  97
  98 /*
  99  * If no young bitflag is supported by the hardware, ->clear_flush_young can
 100  * unmap the address and return 1 or 0 depending if the mapping previously
 101  * existed or not.
 102  */
 103 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
 104                                         unsigned long start,
 105                                         unsigned long end)
 106 {
 107         struct mmu_notifier *mn;
 108         int young = 0, id;
 109
 110         id = srcu_read_lock(&srcu);
 111         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 112                 if (mn->ops->clear_flush_young)
 113                         young |= mn->ops->clear_flush_young(mn, mm, start, end);
 114         }
 115         srcu_read_unlock(&srcu, id);
 116
 117         return young;
 118 }
 119
 120 int __mmu_notifier_clear_young(struct mm_struct *mm,
 121                                unsigned long start,
 122                                unsigned long end)
 123 {
 124         struct mmu_notifier *mn;
 125         int young = 0, id;
 126
 127         id = srcu_read_lock(&srcu);
 128         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 129                 if (mn->ops->clear_young)
 130                         young |= mn->ops->clear_young(mn, mm, start, end);
 131         }
 132         srcu_read_unlock(&srcu, id);
 133
 134         return young;
 135 }
 136
 137 int __mmu_notifier_test_young(struct mm_struct *mm,
 138                               unsigned long address)
 139 {
 140         struct mmu_notifier *mn;
 141         int young = 0, id;
 142
 143         id = srcu_read_lock(&srcu);
 144         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 145                 if (mn->ops->test_young) {
 146                         young = mn->ops->test_young(mn, mm, address);
 147                         if (young)
 148                                 break;
 149                 }
 150         }
 151         srcu_read_unlock(&srcu, id);
 152
 153         return young;
 154 }
 155
 156 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 157                                pte_t pte)
 158 {
 159         struct mmu_notifier *mn;
 160         int id;
 161
 162         id = srcu_read_lock(&srcu);
 163         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 164                 if (mn->ops->change_pte)
 165                         mn->ops->change_pte(mn, mm, address, pte);
 166         }
 167         srcu_read_unlock(&srcu, id);
 168 }
 169
 170 int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
 171 {
 172         struct mmu_notifier *mn;
 173         int ret = 0;
 174         int id;
 175
 176         id = srcu_read_lock(&srcu);
 177         hlist_for_each_entry_rcu(mn, &range->mm->mmu_notifier_mm->list, hlist) {
 178                 if (mn->ops->invalidate_range_start) {
 179                         int _ret = mn->ops->invalidate_range_start(mn, range);
 180                         if (_ret) {
 181                                 pr_info("%pS callback failed with %d in %sblockable context.\n",
 182                                         mn->ops->invalidate_range_start, _ret,
 183                                         !range->blockable ? "non-" : "");
 184                                 ret = _ret;
 185                         }
 186                 }
 187         }
 188         srcu_read_unlock(&srcu, id);
 189
 190         return ret;
 191 }
 192 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
 193
 194 void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range,
 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, &range->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, range->mm,
 217                                                   range->start,
 218                                                   range->end);
 219                 if (mn->ops->invalidate_range_end)
 220                         mn->ops->invalidate_range_end(mn, range);
 221         }
 222         srcu_read_unlock(&srcu, id);
 223 }
 224 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
 225
 226 void __mmu_notifier_invalidate_range(struct mm_struct *mm,
 227                                   unsigned long start, unsigned long end)
 228 {
 229         struct mmu_notifier *mn;
 230         int id;
 231
 232         id = srcu_read_lock(&srcu);
 233         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 234                 if (mn->ops->invalidate_range)
 235                         mn->ops->invalidate_range(mn, mm, start, end);
 236         }
 237         srcu_read_unlock(&srcu, id);
 238 }
 239 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range);
 240
 241 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 242                                     struct mm_struct *mm,
 243                                     int take_mmap_sem)
 244 {
 245         struct mmu_notifier_mm *mmu_notifier_mm;
 246         int ret;
 247
 248         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 249
 250         ret = -ENOMEM;
 251         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 252         if (unlikely(!mmu_notifier_mm))
 253                 goto out;
 254
 255         if (take_mmap_sem)
 256                 down_write(&mm->mmap_sem);
 257         ret = mm_take_all_locks(mm);
 258         if (unlikely(ret))
 259                 goto out_clean;
 260
 261         if (!mm_has_notifiers(mm)) {
 262                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 263                 spin_lock_init(&mmu_notifier_mm->lock);
 264
 265                 mm->mmu_notifier_mm = mmu_notifier_mm;
 266                 mmu_notifier_mm = NULL;
 267         }
 268         mmgrab(mm);
 269
 270         /*
 271          * Serialize the update against mmu_notifier_unregister. A
 272          * side note: mmu_notifier_release can't run concurrently with
 273          * us because we hold the mm_users pin (either implicitly as
 274          * current->mm or explicitly with get_task_mm() or similar).
 275          * We can't race against any other mmu notifier method either
 276          * thanks to mm_take_all_locks().
 277          */
 278         spin_lock(&mm->mmu_notifier_mm->lock);
 279         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 280         spin_unlock(&mm->mmu_notifier_mm->lock);
 281
 282         mm_drop_all_locks(mm);
 283 out_clean:
 284         if (take_mmap_sem)
 285                 up_write(&mm->mmap_sem);
 286         kfree(mmu_notifier_mm);
 287 out:
 288         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 289         return ret;
 290 }
 291
 292 /*
 293  * Must not hold mmap_sem nor any other VM related lock when calling
 294  * this registration function. Must also ensure mm_users can't go down
 295  * to zero while this runs to avoid races with mmu_notifier_release,
 296  * so mm has to be current->mm or the mm should be pinned safely such
 297  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 298  * pin should be released by calling mmput after mmu_notifier_register
 299  * returns. mmu_notifier_unregister must be always called to
 300  * unregister the notifier. mm_count is automatically pinned to allow
 301  * mmu_notifier_unregister to safely run at any time later, before or
 302  * after exit_mmap. ->release will always be called before exit_mmap
 303  * frees the pages.
 304  */
 305 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 306 {
 307         return do_mmu_notifier_register(mn, mm, 1);
 308 }
 309 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 310
 311 /*
 312  * Same as mmu_notifier_register but here the caller must hold the
 313  * mmap_sem in write mode.
 314  */
 315 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 316 {
 317         return do_mmu_notifier_register(mn, mm, 0);
 318 }
 319 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 320
 321 /* this is called after the last mmu_notifier_unregister() returned */
 322 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 323 {
 324         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 325         kfree(mm->mmu_notifier_mm);
 326         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 327 }
 328
 329 /*
 330  * This releases the mm_count pin automatically and frees the mm
 331  * structure if it was the last user of it. It serializes against
 332  * running mmu notifiers with SRCU and against mmu_notifier_unregister
 333  * with the unregister lock + SRCU. All sptes must be dropped before
 334  * calling mmu_notifier_unregister. ->release or any other notifier
 335  * method may be invoked concurrently with mmu_notifier_unregister,
 336  * and only after mmu_notifier_unregister returned we're guaranteed
 337  * that ->release or any other method can't run anymore.
 338  */
 339 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 340 {
 341         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 342
 343         if (!hlist_unhashed(&mn->hlist)) {
 344                 /*
 345                  * SRCU here will force exit_mmap to wait for ->release to
 346                  * finish before freeing the pages.
 347                  */
 348                 int id;
 349
 350                 id = srcu_read_lock(&srcu);
 351                 /*
 352                  * exit_mmap will block in mmu_notifier_release to guarantee
 353                  * that ->release is called before freeing the pages.
 354                  */
 355                 if (mn->ops->release)
 356                         mn->ops->release(mn, mm);
 357                 srcu_read_unlock(&srcu, id);
 358
 359                 spin_lock(&mm->mmu_notifier_mm->lock);
 360                 /*
 361                  * Can not use list_del_rcu() since __mmu_notifier_release
 362                  * can delete it before we hold the lock.
 363                  */
 364                 hlist_del_init_rcu(&mn->hlist);
 365                 spin_unlock(&mm->mmu_notifier_mm->lock);
 366         }
 367
 368         /*
 369          * Wait for any running method to finish, of course including
 370          * ->release if it was run by mmu_notifier_release instead of us.
 371          */
 372         synchronize_srcu(&srcu);
 373
 374         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 375
 376         mmdrop(mm);
 377 }
 378 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
 379
 380 /*
 381  * Same as mmu_notifier_unregister but no callback and no srcu synchronization.
 382  */
 383 void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
 384                                         struct mm_struct *mm)
 385 {
 386         spin_lock(&mm->mmu_notifier_mm->lock);
 387         /*
 388          * Can not use list_del_rcu() since __mmu_notifier_release
 389          * can delete it before we hold the lock.
 390          */
 391         hlist_del_init_rcu(&mn->hlist);
 392         spin_unlock(&mm->mmu_notifier_mm->lock);
 393
 394         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 395         mmdrop(mm);
 396 }
 397 EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release);