include/linux/mmu_notifier.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _LINUX_MMU_NOTIFIER_H
   3 #define _LINUX_MMU_NOTIFIER_H
   4
   5 #include <linux/list.h>
   6 #include <linux/spinlock.h>
   7 #include <linux/mm_types.h>
   8 #include <linux/srcu.h>
   9
  10 struct mmu_notifier;
  11 struct mmu_notifier_ops;
  12
  13 #ifdef CONFIG_MMU_NOTIFIER
  14
  15 /*
  16  * The mmu notifier_mm structure is allocated and installed in
  17  * mm->mmu_notifier_mm inside the mm_take_all_locks() protected
  18  * critical section and it's released only when mm_count reaches zero
  19  * in mmdrop().
  20  */
  21 struct mmu_notifier_mm {
  22         /* all mmu notifiers registerd in this mm are queued in this list */
  23         struct hlist_head list;
  24         /* to serialize the list modifications and hlist_unhashed */
  25         spinlock_t lock;
  26 };
  27
  28 struct mmu_notifier_ops {
  29         /*
  30          * Called either by mmu_notifier_unregister or when the mm is
  31          * being destroyed by exit_mmap, always before all pages are
  32          * freed. This can run concurrently with other mmu notifier
  33          * methods (the ones invoked outside the mm context) and it
  34          * should tear down all secondary mmu mappings and freeze the
  35          * secondary mmu. If this method isn't implemented you've to
  36          * be sure that nothing could possibly write to the pages
  37          * through the secondary mmu by the time the last thread with
  38          * tsk->mm == mm exits.
  39          *
  40          * As side note: the pages freed after ->release returns could
  41          * be immediately reallocated by the gart at an alias physical
  42          * address with a different cache model, so if ->release isn't
  43          * implemented because all _software_ driven memory accesses
  44          * through the secondary mmu are terminated by the time the
  45          * last thread of this mm quits, you've also to be sure that
  46          * speculative _hardware_ operations can't allocate dirty
  47          * cachelines in the cpu that could not be snooped and made
  48          * coherent with the other read and write operations happening
  49          * through the gart alias address, so leading to memory
  50          * corruption.
  51          */
  52         void (*release)(struct mmu_notifier *mn,
  53                         struct mm_struct *mm);
  54
  55         /*
  56          * clear_flush_young is called after the VM is
  57          * test-and-clearing the young/accessed bitflag in the
  58          * pte. This way the VM will provide proper aging to the
  59          * accesses to the page through the secondary MMUs and not
  60          * only to the ones through the Linux pte.
  61          * Start-end is necessary in case the secondary MMU is mapping the page
  62          * at a smaller granularity than the primary MMU.
  63          */
  64         int (*clear_flush_young)(struct mmu_notifier *mn,
  65                                  struct mm_struct *mm,
  66                                  unsigned long start,
  67                                  unsigned long end);
  68
  69         /*
  70          * clear_young is a lightweight version of clear_flush_young. Like the
  71          * latter, it is supposed to test-and-clear the young/accessed bitflag
  72          * in the secondary pte, but it may omit flushing the secondary tlb.
  73          */
  74         int (*clear_young)(struct mmu_notifier *mn,
  75                            struct mm_struct *mm,
  76                            unsigned long start,
  77                            unsigned long end);
  78
  79         /*
  80          * test_young is called to check the young/accessed bitflag in
  81          * the secondary pte. This is used to know if the page is
  82          * frequently used without actually clearing the flag or tearing
  83          * down the secondary mapping on the page.
  84          */
  85         int (*test_young)(struct mmu_notifier *mn,
  86                           struct mm_struct *mm,
  87                           unsigned long address);
  88
  89         /*
  90          * change_pte is called in cases that pte mapping to page is changed:
  91          * for example, when ksm remaps pte to point to a new shared page.
  92          */
  93         void (*change_pte)(struct mmu_notifier *mn,
  94                            struct mm_struct *mm,
  95                            unsigned long address,
  96                            pte_t pte);
  97
  98         /*
  99          * invalidate_range_start() and invalidate_range_end() must be
 100          * paired and are called only when the mmap_sem and/or the
 101          * locks protecting the reverse maps are held. If the subsystem
 102          * can't guarantee that no additional references are taken to
 103          * the pages in the range, it has to implement the
 104          * invalidate_range() notifier to remove any references taken
 105          * after invalidate_range_start().
 106          *
 107          * Invalidation of multiple concurrent ranges may be
 108          * optionally permitted by the driver. Either way the
 109          * establishment of sptes is forbidden in the range passed to
 110          * invalidate_range_begin/end for the whole duration of the
 111          * invalidate_range_begin/end critical section.
 112          *
 113          * invalidate_range_start() is called when all pages in the
 114          * range are still mapped and have at least a refcount of one.
 115          *
 116          * invalidate_range_end() is called when all pages in the
 117          * range have been unmapped and the pages have been freed by
 118          * the VM.
 119          *
 120          * The VM will remove the page table entries and potentially
 121          * the page between invalidate_range_start() and
 122          * invalidate_range_end(). If the page must not be freed
 123          * because of pending I/O or other circumstances then the
 124          * invalidate_range_start() callback (or the initial mapping
 125          * by the driver) must make sure that the refcount is kept
 126          * elevated.
 127          *
 128          * If the driver increases the refcount when the pages are
 129          * initially mapped into an address space then either
 130          * invalidate_range_start() or invalidate_range_end() may
 131          * decrease the refcount. If the refcount is decreased on
 132          * invalidate_range_start() then the VM can free pages as page
 133          * table entries are removed.  If the refcount is only
 134          * droppped on invalidate_range_end() then the driver itself
 135          * will drop the last refcount but it must take care to flush
 136          * any secondary tlb before doing the final free on the
 137          * page. Pages will no longer be referenced by the linux
 138          * address space but may still be referenced by sptes until
 139          * the last refcount is dropped.
 140          */
 141         void (*invalidate_range_start)(struct mmu_notifier *mn,
 142                                        struct mm_struct *mm,
 143                                        unsigned long start, unsigned long end);
 144         void (*invalidate_range_end)(struct mmu_notifier *mn,
 145                                      struct mm_struct *mm,
 146                                      unsigned long start, unsigned long end);
 147
 148         /*
 149          * invalidate_range() is either called between
 150          * invalidate_range_start() and invalidate_range_end() when the
 151          * VM has to free pages that where unmapped, but before the
 152          * pages are actually freed, or outside of _start()/_end() when
 153          * a (remote) TLB is necessary.
 154          *
 155          * If invalidate_range() is used to manage a non-CPU TLB with
 156          * shared page-tables, it not necessary to implement the
 157          * invalidate_range_start()/end() notifiers, as
 158          * invalidate_range() alread catches the points in time when an
 159          * external TLB range needs to be flushed.
 160          *
 161          * The invalidate_range() function is called under the ptl
 162          * spin-lock and not allowed to sleep.
 163          *
 164          * Note that this function might be called with just a sub-range
 165          * of what was passed to invalidate_range_start()/end(), if
 166          * called between those functions.
 167          */
 168         void (*invalidate_range)(struct mmu_notifier *mn, struct mm_struct *mm,
 169                                  unsigned long start, unsigned long end);
 170 };
 171
 172 /*
 173  * The notifier chains are protected by mmap_sem and/or the reverse map
 174  * semaphores. Notifier chains are only changed when all reverse maps and
 175  * the mmap_sem locks are taken.
 176  *
 177  * Therefore notifier chains can only be traversed when either
 178  *
 179  * 1. mmap_sem is held.
 180  * 2. One of the reverse map locks is held (i_mmap_rwsem or anon_vma->rwsem).
 181  * 3. No other concurrent thread can access the list (release)
 182  */
 183 struct mmu_notifier {
 184         struct hlist_node hlist;
 185         const struct mmu_notifier_ops *ops;
 186 };
 187
 188 static inline int mm_has_notifiers(struct mm_struct *mm)
 189 {
 190         return unlikely(mm->mmu_notifier_mm);
 191 }
 192
 193 extern int mmu_notifier_register(struct mmu_notifier *mn,
 194                                  struct mm_struct *mm);
 195 extern int __mmu_notifier_register(struct mmu_notifier *mn,
 196                                    struct mm_struct *mm);
 197 extern void mmu_notifier_unregister(struct mmu_notifier *mn,
 198                                     struct mm_struct *mm);
 199 extern void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
 200                                                struct mm_struct *mm);
 201 extern void __mmu_notifier_mm_destroy(struct mm_struct *mm);
 202 extern void __mmu_notifier_release(struct mm_struct *mm);
 203 extern int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
 204                                           unsigned long start,
 205                                           unsigned long end);
 206 extern int __mmu_notifier_clear_young(struct mm_struct *mm,
 207                                       unsigned long start,
 208                                       unsigned long end);
 209 extern int __mmu_notifier_test_young(struct mm_struct *mm,
 210                                      unsigned long address);
 211 extern void __mmu_notifier_change_pte(struct mm_struct *mm,
 212                                       unsigned long address, pte_t pte);
 213 extern void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 214                                   unsigned long start, unsigned long end);
 215 extern void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 216                                   unsigned long start, unsigned long end);
 217 extern void __mmu_notifier_invalidate_range(struct mm_struct *mm,
 218                                   unsigned long start, unsigned long end);
 219
 220 static inline void mmu_notifier_release(struct mm_struct *mm)
 221 {
 222         if (mm_has_notifiers(mm))
 223                 __mmu_notifier_release(mm);
 224 }
 225
 226 static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
 227                                           unsigned long start,
 228                                           unsigned long end)
 229 {
 230         if (mm_has_notifiers(mm))
 231                 return __mmu_notifier_clear_flush_young(mm, start, end);
 232         return 0;
 233 }
 234
 235 static inline int mmu_notifier_clear_young(struct mm_struct *mm,
 236                                            unsigned long start,
 237                                            unsigned long end)
 238 {
 239         if (mm_has_notifiers(mm))
 240                 return __mmu_notifier_clear_young(mm, start, end);
 241         return 0;
 242 }
 243
 244 static inline int mmu_notifier_test_young(struct mm_struct *mm,
 245                                           unsigned long address)
 246 {
 247         if (mm_has_notifiers(mm))
 248                 return __mmu_notifier_test_young(mm, address);
 249         return 0;
 250 }
 251
 252 static inline void mmu_notifier_change_pte(struct mm_struct *mm,
 253                                            unsigned long address, pte_t pte)
 254 {
 255         if (mm_has_notifiers(mm))
 256                 __mmu_notifier_change_pte(mm, address, pte);
 257 }
 258
 259 static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 260                                   unsigned long start, unsigned long end)
 261 {
 262         if (mm_has_notifiers(mm))
 263                 __mmu_notifier_invalidate_range_start(mm, start, end);
 264 }
 265
 266 static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 267                                   unsigned long start, unsigned long end)
 268 {
 269         if (mm_has_notifiers(mm))
 270                 __mmu_notifier_invalidate_range_end(mm, start, end);
 271 }
 272
 273 static inline void mmu_notifier_invalidate_range(struct mm_struct *mm,
 274                                   unsigned long start, unsigned long end)
 275 {
 276         if (mm_has_notifiers(mm))
 277                 __mmu_notifier_invalidate_range(mm, start, end);
 278 }
 279
 280 static inline void mmu_notifier_mm_init(struct mm_struct *mm)
 281 {
 282         mm->mmu_notifier_mm = NULL;
 283 }
 284
 285 static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
 286 {
 287         if (mm_has_notifiers(mm))
 288                 __mmu_notifier_mm_destroy(mm);
 289 }
 290
 291 #define ptep_clear_flush_young_notify(__vma, __address, __ptep)         \
 292 ({                                                                      \
 293         int __young;                                                    \
 294         struct vm_area_struct *___vma = __vma;                          \
 295         unsigned long ___address = __address;                           \
 296         __young = ptep_clear_flush_young(___vma, ___address, __ptep);   \
 297         __young |= mmu_notifier_clear_flush_young(___vma->vm_mm,        \
 298                                                   ___address,           \
 299                                                   ___address +          \
 300                                                         PAGE_SIZE);     \
 301         __young;                                                        \
 302 })
 303
 304 #define pmdp_clear_flush_young_notify(__vma, __address, __pmdp)         \
 305 ({                                                                      \
 306         int __young;                                                    \
 307         struct vm_area_struct *___vma = __vma;                          \
 308         unsigned long ___address = __address;                           \
 309         __young = pmdp_clear_flush_young(___vma, ___address, __pmdp);   \
 310         __young |= mmu_notifier_clear_flush_young(___vma->vm_mm,        \
 311                                                   ___address,           \
 312                                                   ___address +          \
 313                                                         PMD_SIZE);      \
 314         __young;                                                        \
 315 })
 316
 317 #define ptep_clear_young_notify(__vma, __address, __ptep)               \
 318 ({                                                                      \
 319         int __young;                                                    \
 320         struct vm_area_struct *___vma = __vma;                          \
 321         unsigned long ___address = __address;                           \
 322         __young = ptep_test_and_clear_young(___vma, ___address, __ptep);\
 323         __young |= mmu_notifier_clear_young(___vma->vm_mm, ___address,  \
 324                                             ___address + PAGE_SIZE);    \
 325         __young;                                                        \
 326 })
 327
 328 #define pmdp_clear_young_notify(__vma, __address, __pmdp)               \
 329 ({                                                                      \
 330         int __young;                                                    \
 331         struct vm_area_struct *___vma = __vma;                          \
 332         unsigned long ___address = __address;                           \
 333         __young = pmdp_test_and_clear_young(___vma, ___address, __pmdp);\
 334         __young |= mmu_notifier_clear_young(___vma->vm_mm, ___address,  \
 335                                             ___address + PMD_SIZE);     \
 336         __young;                                                        \
 337 })
 338
 339 #define ptep_clear_flush_notify(__vma, __address, __ptep)               \
 340 ({                                                                      \
 341         unsigned long ___addr = __address & PAGE_MASK;                  \
 342         struct mm_struct *___mm = (__vma)->vm_mm;                       \
 343         pte_t ___pte;                                                   \
 344                                                                         \
 345         ___pte = ptep_clear_flush(__vma, __address, __ptep);            \
 346         mmu_notifier_invalidate_range(___mm, ___addr,                   \
 347                                         ___addr + PAGE_SIZE);           \
 348                                                                         \
 349         ___pte;                                                         \
 350 })
 351
 352 #define pmdp_huge_clear_flush_notify(__vma, __haddr, __pmd)             \
 353 ({                                                                      \
 354         unsigned long ___haddr = __haddr & HPAGE_PMD_MASK;              \
 355         struct mm_struct *___mm = (__vma)->vm_mm;                       \
 356         pmd_t ___pmd;                                                   \
 357                                                                         \
 358         ___pmd = pmdp_huge_clear_flush(__vma, __haddr, __pmd);          \
 359         mmu_notifier_invalidate_range(___mm, ___haddr,                  \
 360                                       ___haddr + HPAGE_PMD_SIZE);       \
 361                                                                         \
 362         ___pmd;                                                         \
 363 })
 364
 365 #define pudp_huge_clear_flush_notify(__vma, __haddr, __pud)             \
 366 ({                                                                      \
 367         unsigned long ___haddr = __haddr & HPAGE_PUD_MASK;              \
 368         struct mm_struct *___mm = (__vma)->vm_mm;                       \
 369         pud_t ___pud;                                                   \
 370                                                                         \
 371         ___pud = pudp_huge_clear_flush(__vma, __haddr, __pud);          \
 372         mmu_notifier_invalidate_range(___mm, ___haddr,                  \
 373                                       ___haddr + HPAGE_PUD_SIZE);       \
 374                                                                         \
 375         ___pud;                                                         \
 376 })
 377
 378 /*
 379  * set_pte_at_notify() sets the pte _after_ running the notifier.
 380  * This is safe to start by updating the secondary MMUs, because the primary MMU
 381  * pte invalidate must have already happened with a ptep_clear_flush() before
 382  * set_pte_at_notify() has been invoked.  Updating the secondary MMUs first is
 383  * required when we change both the protection of the mapping from read-only to
 384  * read-write and the pfn (like during copy on write page faults). Otherwise the
 385  * old page would remain mapped readonly in the secondary MMUs after the new
 386  * page is already writable by some CPU through the primary MMU.
 387  */
 388 #define set_pte_at_notify(__mm, __address, __ptep, __pte)               \
 389 ({                                                                      \
 390         struct mm_struct *___mm = __mm;                                 \
 391         unsigned long ___address = __address;                           \
 392         pte_t ___pte = __pte;                                           \
 393                                                                         \
 394         mmu_notifier_change_pte(___mm, ___address, ___pte);             \
 395         set_pte_at(___mm, ___address, __ptep, ___pte);                  \
 396 })
 397
 398 extern void mmu_notifier_call_srcu(struct rcu_head *rcu,
 399                                    void (*func)(struct rcu_head *rcu));
 400 extern void mmu_notifier_synchronize(void);
 401
 402 #else /* CONFIG_MMU_NOTIFIER */
 403
 404 static inline int mm_has_notifiers(struct mm_struct *mm)
 405 {
 406         return 0;
 407 }
 408
 409 static inline void mmu_notifier_release(struct mm_struct *mm)
 410 {
 411 }
 412
 413 static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
 414                                           unsigned long start,
 415                                           unsigned long end)
 416 {
 417         return 0;
 418 }
 419
 420 static inline int mmu_notifier_test_young(struct mm_struct *mm,
 421                                           unsigned long address)
 422 {
 423         return 0;
 424 }
 425
 426 static inline void mmu_notifier_change_pte(struct mm_struct *mm,
 427                                            unsigned long address, pte_t pte)
 428 {
 429 }
 430
 431 static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
 432                                   unsigned long start, unsigned long end)
 433 {
 434 }
 435
 436 static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
 437                                   unsigned long start, unsigned long end)
 438 {
 439 }
 440
 441 static inline void mmu_notifier_invalidate_range(struct mm_struct *mm,
 442                                   unsigned long start, unsigned long end)
 443 {
 444 }
 445
 446 static inline void mmu_notifier_mm_init(struct mm_struct *mm)
 447 {
 448 }
 449
 450 static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
 451 {
 452 }
 453
 454 #define ptep_clear_flush_young_notify ptep_clear_flush_young
 455 #define pmdp_clear_flush_young_notify pmdp_clear_flush_young
 456 #define ptep_clear_young_notify ptep_test_and_clear_young
 457 #define pmdp_clear_young_notify pmdp_test_and_clear_young
 458 #define ptep_clear_flush_notify ptep_clear_flush
 459 #define pmdp_huge_clear_flush_notify pmdp_huge_clear_flush
 460 #define pudp_huge_clear_flush_notify pudp_huge_clear_flush
 461 #define set_pte_at_notify set_pte_at
 462
 463 #endif /* CONFIG_MMU_NOTIFIER */
 464
 465 #endif /* _LINUX_MMU_NOTIFIER_H */