mm/mmu_notifier.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  linux/mm/mmu_notifier.c
   4  *
   5  *  Copyright (C) 2008  Qumranet, Inc.
   6  *  Copyright (C) 2008  SGI
   7  *             Christoph Lameter <cl@linux.com>
   8  */
   9
  10 #include <linux/rculist.h>
  11 #include <linux/mmu_notifier.h>
  12 #include <linux/export.h>
  13 #include <linux/mm.h>
  14 #include <linux/err.h>
  15 #include <linux/srcu.h>
  16 #include <linux/rcupdate.h>
  17 #include <linux/sched.h>
  18 #include <linux/sched/mm.h>
  19 #include <linux/slab.h>
  20
  21 /* global SRCU for all MMs */
  22 DEFINE_STATIC_SRCU(srcu);
  23
  24 /*
  25  * This function allows mmu_notifier::release callback to delay a call to
  26  * a function that will free appropriate resources. The function must be
  27  * quick and must not block.
  28  */
  29 void mmu_notifier_call_srcu(struct rcu_head *rcu,
  30                             void (*func)(struct rcu_head *rcu))
  31 {
  32         call_srcu(&srcu, rcu, func);
  33 }
  34 EXPORT_SYMBOL_GPL(mmu_notifier_call_srcu);
  35
  36 /*
  37  * This function can't run concurrently against mmu_notifier_register
  38  * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
  39  * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
  40  * in parallel despite there being no task using this mm any more,
  41  * through the vmas outside of the exit_mmap context, such as with
  42  * vmtruncate. This serializes against mmu_notifier_unregister with
  43  * the mmu_notifier_mm->lock in addition to SRCU and it serializes
  44  * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
  45  * can't go away from under us as exit_mmap holds an mm_count pin
  46  * itself.
  47  */
  48 void __mmu_notifier_release(struct mm_struct *mm)
  49 {
  50         struct mmu_notifier *mn;
  51         int id;
  52
  53         /*
  54          * SRCU here will block mmu_notifier_unregister until
  55          * ->release returns.
  56          */
  57         id = srcu_read_lock(&srcu);
  58         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
  59                 /*
  60                  * If ->release runs before mmu_notifier_unregister it must be
  61                  * handled, as it's the only way for the driver to flush all
  62                  * existing sptes and stop the driver from establishing any more
  63                  * sptes before all the pages in the mm are freed.
  64                  */
  65                 if (mn->ops->release)
  66                         mn->ops->release(mn, mm);
  67
  68         spin_lock(&mm->mmu_notifier_mm->lock);
  69         while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
  70                 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
  71                                  struct mmu_notifier,
  72                                  hlist);
  73                 /*
  74                  * We arrived before mmu_notifier_unregister so
  75                  * mmu_notifier_unregister will do nothing other than to wait
  76                  * for ->release to finish and for mmu_notifier_unregister to
  77                  * return.
  78                  */
  79                 hlist_del_init_rcu(&mn->hlist);
  80         }
  81         spin_unlock(&mm->mmu_notifier_mm->lock);
  82         srcu_read_unlock(&srcu, id);
  83
  84         /*
  85          * synchronize_srcu here prevents mmu_notifier_release from returning to
  86          * exit_mmap (which would proceed with freeing all pages in the mm)
  87          * until the ->release method returns, if it was invoked by
  88          * mmu_notifier_unregister.
  89          *
  90          * The mmu_notifier_mm can't go away from under us because one mm_count
  91          * is held by exit_mmap.
  92          */
  93         synchronize_srcu(&srcu);
  94 }
  95
  96 /*
  97  * If no young bitflag is supported by the hardware, ->clear_flush_young can
  98  * unmap the address and return 1 or 0 depending if the mapping previously
  99  * existed or not.
 100  */
 101 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
 102                                         unsigned long start,
 103                                         unsigned long end)
 104 {
 105         struct mmu_notifier *mn;
 106         int young = 0, id;
 107
 108         id = srcu_read_lock(&srcu);
 109         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 110                 if (mn->ops->clear_flush_young)
 111                         young |= mn->ops->clear_flush_young(mn, mm, start, end);
 112         }
 113         srcu_read_unlock(&srcu, id);
 114
 115         return young;
 116 }
 117
 118 int __mmu_notifier_clear_young(struct mm_struct *mm,
 119                                unsigned long start,
 120                                unsigned long end)
 121 {
 122         struct mmu_notifier *mn;
 123         int young = 0, id;
 124
 125         id = srcu_read_lock(&srcu);
 126         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 127                 if (mn->ops->clear_young)
 128                         young |= mn->ops->clear_young(mn, mm, start, end);
 129         }
 130         srcu_read_unlock(&srcu, id);
 131
 132         return young;
 133 }
 134
 135 int __mmu_notifier_test_young(struct mm_struct *mm,
 136                               unsigned long address)
 137 {
 138         struct mmu_notifier *mn;
 139         int young = 0, id;
 140
 141         id = srcu_read_lock(&srcu);
 142         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 143                 if (mn->ops->test_young) {
 144                         young = mn->ops->test_young(mn, mm, address);
 145                         if (young)
 146                                 break;
 147                 }
 148         }
 149         srcu_read_unlock(&srcu, id);
 150
 151         return young;
 152 }
 153
 154 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 155                                pte_t pte)
 156 {
 157         struct mmu_notifier *mn;
 158         int id;
 159
 160         id = srcu_read_lock(&srcu);
 161         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 162                 if (mn->ops->change_pte)
 163                         mn->ops->change_pte(mn, mm, address, pte);
 164         }
 165         srcu_read_unlock(&srcu, id);
 166 }
 167
 168 int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
 169 {
 170         struct mmu_notifier *mn;
 171         int ret = 0;
 172         int id;
 173
 174         id = srcu_read_lock(&srcu);
 175         hlist_for_each_entry_rcu(mn, &range->mm->mmu_notifier_mm->list, hlist) {
 176                 if (mn->ops->invalidate_range_start) {
 177                         int _ret = mn->ops->invalidate_range_start(mn, range);
 178                         if (_ret) {
 179                                 pr_info("%pS callback failed with %d in %sblockable context.\n",
 180                                         mn->ops->invalidate_range_start, _ret,
 181                                         !mmu_notifier_range_blockable(range) ? "non-" : "");
 182                                 ret = _ret;
 183                         }
 184                 }
 185         }
 186         srcu_read_unlock(&srcu, id);
 187
 188         return ret;
 189 }
 190 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
 191
 192 void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range,
 193                                          bool only_end)
 194 {
 195         struct mmu_notifier *mn;
 196         int id;
 197
 198         id = srcu_read_lock(&srcu);
 199         hlist_for_each_entry_rcu(mn, &range->mm->mmu_notifier_mm->list, hlist) {
 200                 /*
 201                  * Call invalidate_range here too to avoid the need for the
 202                  * subsystem of having to register an invalidate_range_end
 203                  * call-back when there is invalidate_range already. Usually a
 204                  * subsystem registers either invalidate_range_start()/end() or
 205                  * invalidate_range(), so this will be no additional overhead
 206                  * (besides the pointer check).
 207                  *
 208                  * We skip call to invalidate_range() if we know it is safe ie
 209                  * call site use mmu_notifier_invalidate_range_only_end() which
 210                  * is safe to do when we know that a call to invalidate_range()
 211                  * already happen under page table lock.
 212                  */
 213                 if (!only_end && mn->ops->invalidate_range)
 214                         mn->ops->invalidate_range(mn, range->mm,
 215                                                   range->start,
 216                                                   range->end);
 217                 if (mn->ops->invalidate_range_end)
 218                         mn->ops->invalidate_range_end(mn, range);
 219         }
 220         srcu_read_unlock(&srcu, id);
 221 }
 222 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
 223
 224 void __mmu_notifier_invalidate_range(struct mm_struct *mm,
 225                                   unsigned long start, unsigned long end)
 226 {
 227         struct mmu_notifier *mn;
 228         int id;
 229
 230         id = srcu_read_lock(&srcu);
 231         hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
 232                 if (mn->ops->invalidate_range)
 233                         mn->ops->invalidate_range(mn, mm, start, end);
 234         }
 235         srcu_read_unlock(&srcu, id);
 236 }
 237 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range);
 238
 239 static int do_mmu_notifier_register(struct mmu_notifier *mn,
 240                                     struct mm_struct *mm,
 241                                     int take_mmap_sem)
 242 {
 243         struct mmu_notifier_mm *mmu_notifier_mm;
 244         int ret;
 245
 246         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 247
 248         ret = -ENOMEM;
 249         mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
 250         if (unlikely(!mmu_notifier_mm))
 251                 goto out;
 252
 253         if (take_mmap_sem)
 254                 down_write(&mm->mmap_sem);
 255         ret = mm_take_all_locks(mm);
 256         if (unlikely(ret))
 257                 goto out_clean;
 258
 259         if (!mm_has_notifiers(mm)) {
 260                 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
 261                 spin_lock_init(&mmu_notifier_mm->lock);
 262
 263                 mm->mmu_notifier_mm = mmu_notifier_mm;
 264                 mmu_notifier_mm = NULL;
 265         }
 266         mmgrab(mm);
 267
 268         /*
 269          * Serialize the update against mmu_notifier_unregister. A
 270          * side note: mmu_notifier_release can't run concurrently with
 271          * us because we hold the mm_users pin (either implicitly as
 272          * current->mm or explicitly with get_task_mm() or similar).
 273          * We can't race against any other mmu notifier method either
 274          * thanks to mm_take_all_locks().
 275          */
 276         spin_lock(&mm->mmu_notifier_mm->lock);
 277         hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
 278         spin_unlock(&mm->mmu_notifier_mm->lock);
 279
 280         mm_drop_all_locks(mm);
 281 out_clean:
 282         if (take_mmap_sem)
 283                 up_write(&mm->mmap_sem);
 284         kfree(mmu_notifier_mm);
 285 out:
 286         BUG_ON(atomic_read(&mm->mm_users) <= 0);
 287         return ret;
 288 }
 289
 290 /*
 291  * Must not hold mmap_sem nor any other VM related lock when calling
 292  * this registration function. Must also ensure mm_users can't go down
 293  * to zero while this runs to avoid races with mmu_notifier_release,
 294  * so mm has to be current->mm or the mm should be pinned safely such
 295  * as with get_task_mm(). If the mm is not current->mm, the mm_users
 296  * pin should be released by calling mmput after mmu_notifier_register
 297  * returns. mmu_notifier_unregister must be always called to
 298  * unregister the notifier. mm_count is automatically pinned to allow
 299  * mmu_notifier_unregister to safely run at any time later, before or
 300  * after exit_mmap. ->release will always be called before exit_mmap
 301  * frees the pages.
 302  */
 303 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 304 {
 305         return do_mmu_notifier_register(mn, mm, 1);
 306 }
 307 EXPORT_SYMBOL_GPL(mmu_notifier_register);
 308
 309 /*
 310  * Same as mmu_notifier_register but here the caller must hold the
 311  * mmap_sem in write mode.
 312  */
 313 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
 314 {
 315         return do_mmu_notifier_register(mn, mm, 0);
 316 }
 317 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
 318
 319 /* this is called after the last mmu_notifier_unregister() returned */
 320 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
 321 {
 322         BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
 323         kfree(mm->mmu_notifier_mm);
 324         mm->mmu_notifier_mm = LIST_POISON1; /* debug */
 325 }
 326
 327 /*
 328  * This releases the mm_count pin automatically and frees the mm
 329  * structure if it was the last user of it. It serializes against
 330  * running mmu notifiers with SRCU and against mmu_notifier_unregister
 331  * with the unregister lock + SRCU. All sptes must be dropped before
 332  * calling mmu_notifier_unregister. ->release or any other notifier
 333  * method may be invoked concurrently with mmu_notifier_unregister,
 334  * and only after mmu_notifier_unregister returned we're guaranteed
 335  * that ->release or any other method can't run anymore.
 336  */
 337 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
 338 {
 339         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 340
 341         if (!hlist_unhashed(&mn->hlist)) {
 342                 /*
 343                  * SRCU here will force exit_mmap to wait for ->release to
 344                  * finish before freeing the pages.
 345                  */
 346                 int id;
 347
 348                 id = srcu_read_lock(&srcu);
 349                 /*
 350                  * exit_mmap will block in mmu_notifier_release to guarantee
 351                  * that ->release is called before freeing the pages.
 352                  */
 353                 if (mn->ops->release)
 354                         mn->ops->release(mn, mm);
 355                 srcu_read_unlock(&srcu, id);
 356
 357                 spin_lock(&mm->mmu_notifier_mm->lock);
 358                 /*
 359                  * Can not use list_del_rcu() since __mmu_notifier_release
 360                  * can delete it before we hold the lock.
 361                  */
 362                 hlist_del_init_rcu(&mn->hlist);
 363                 spin_unlock(&mm->mmu_notifier_mm->lock);
 364         }
 365
 366         /*
 367          * Wait for any running method to finish, of course including
 368          * ->release if it was run by mmu_notifier_release instead of us.
 369          */
 370         synchronize_srcu(&srcu);
 371
 372         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 373
 374         mmdrop(mm);
 375 }
 376 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
 377
 378 /*
 379  * Same as mmu_notifier_unregister but no callback and no srcu synchronization.
 380  */
 381 void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
 382                                         struct mm_struct *mm)
 383 {
 384         spin_lock(&mm->mmu_notifier_mm->lock);
 385         /*
 386          * Can not use list_del_rcu() since __mmu_notifier_release
 387          * can delete it before we hold the lock.
 388          */
 389         hlist_del_init_rcu(&mn->hlist);
 390         spin_unlock(&mm->mmu_notifier_mm->lock);
 391
 392         BUG_ON(atomic_read(&mm->mm_count) <= 0);
 393         mmdrop(mm);
 394 }
 395 EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release);
 396
 397 bool
 398 mmu_notifier_range_update_to_read_only(const struct mmu_notifier_range *range)
 399 {
 400         if (!range->vma || range->event != MMU_NOTIFY_PROTECTION_VMA)
 401                 return false;
 402         /* Return true if the vma still have the read flag set. */
 403         return range->vma->vm_flags & VM_READ;
 404 }
 405 EXPORT_SYMBOL_GPL(mmu_notifier_range_update_to_read_only);