arch/arm64/kernel/mte.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2020 ARM Ltd.
   4  */
   5
   6 #include <linux/bitops.h>
   7 #include <linux/kernel.h>
   8 #include <linux/mm.h>
   9 #include <linux/prctl.h>
  10 #include <linux/sched.h>
  11 #include <linux/sched/mm.h>
  12 #include <linux/string.h>
  13 #include <linux/swap.h>
  14 #include <linux/swapops.h>
  15 #include <linux/thread_info.h>
  16 #include <linux/types.h>
  17 #include <linux/uio.h>
  18
  19 #include <asm/barrier.h>
  20 #include <asm/cpufeature.h>
  21 #include <asm/mte.h>
  22 #include <asm/mte-kasan.h>
  23 #include <asm/ptrace.h>
  24 #include <asm/sysreg.h>
  25
  26 u64 gcr_kernel_excl __ro_after_init;
  27
  28 static void mte_sync_page_tags(struct page *page, pte_t *ptep, bool check_swap)
  29 {
  30         pte_t old_pte = READ_ONCE(*ptep);
  31
  32         if (check_swap && is_swap_pte(old_pte)) {
  33                 swp_entry_t entry = pte_to_swp_entry(old_pte);
  34
  35                 if (!non_swap_entry(entry) && mte_restore_tags(entry, page))
  36                         return;
  37         }
  38
  39         page_kasan_tag_reset(page);
  40         /*
  41          * We need smp_wmb() in between setting the flags and clearing the
  42          * tags because if another thread reads page->flags and builds a
  43          * tagged address out of it, there is an actual dependency to the
  44          * memory access, but on the current thread we do not guarantee that
  45          * the new page->flags are visible before the tags were updated.
  46          */
  47         smp_wmb();
  48         mte_clear_page_tags(page_address(page));
  49 }
  50
  51 void mte_sync_tags(pte_t *ptep, pte_t pte)
  52 {
  53         struct page *page = pte_page(pte);
  54         long i, nr_pages = compound_nr(page);
  55         bool check_swap = nr_pages == 1;
  56
  57         /* if PG_mte_tagged is set, tags have already been initialised */
  58         for (i = 0; i < nr_pages; i++, page++) {
  59                 if (!test_and_set_bit(PG_mte_tagged, &page->flags))
  60                         mte_sync_page_tags(page, ptep, check_swap);
  61         }
  62 }
  63
  64 int memcmp_pages(struct page *page1, struct page *page2)
  65 {
  66         char *addr1, *addr2;
  67         int ret;
  68
  69         addr1 = page_address(page1);
  70         addr2 = page_address(page2);
  71         ret = memcmp(addr1, addr2, PAGE_SIZE);
  72
  73         if (!system_supports_mte() || ret)
  74                 return ret;
  75
  76         /*
  77          * If the page content is identical but at least one of the pages is
  78          * tagged, return non-zero to avoid KSM merging. If only one of the
  79          * pages is tagged, set_pte_at() may zero or change the tags of the
  80          * other page via mte_sync_tags().
  81          */
  82         if (test_bit(PG_mte_tagged, &page1->flags) ||
  83             test_bit(PG_mte_tagged, &page2->flags))
  84                 return addr1 != addr2;
  85
  86         return ret;
  87 }
  88
  89 u8 mte_get_mem_tag(void *addr)
  90 {
  91         if (!system_supports_mte())
  92                 return 0xFF;
  93
  94         asm(__MTE_PREAMBLE "ldg %0, [%0]"
  95             : "+r" (addr));
  96
  97         return mte_get_ptr_tag(addr);
  98 }
  99
 100 u8 mte_get_random_tag(void)
 101 {
 102         void *addr;
 103
 104         if (!system_supports_mte())
 105                 return 0xFF;
 106
 107         asm(__MTE_PREAMBLE "irg %0, %0"
 108             : "+r" (addr));
 109
 110         return mte_get_ptr_tag(addr);
 111 }
 112
 113 void *mte_set_mem_tag_range(void *addr, size_t size, u8 tag)
 114 {
 115         void *ptr = addr;
 116
 117         if ((!system_supports_mte()) || (size == 0))
 118                 return addr;
 119
 120         /* Make sure that size is MTE granule aligned. */
 121         WARN_ON(size & (MTE_GRANULE_SIZE - 1));
 122
 123         /* Make sure that the address is MTE granule aligned. */
 124         WARN_ON((u64)addr & (MTE_GRANULE_SIZE - 1));
 125
 126         tag = 0xF0 | tag;
 127         ptr = (void *)__tag_set(ptr, tag);
 128
 129         mte_assign_mem_tag_range(ptr, size);
 130
 131         return ptr;
 132 }
 133
 134 void mte_init_tags(u64 max_tag)
 135 {
 136         static bool gcr_kernel_excl_initialized;
 137
 138         if (!gcr_kernel_excl_initialized) {
 139                 /*
 140                  * The format of the tags in KASAN is 0xFF and in MTE is 0xF.
 141                  * This conversion extracts an MTE tag from a KASAN tag.
 142                  */
 143                 u64 incl = GENMASK(FIELD_GET(MTE_TAG_MASK >> MTE_TAG_SHIFT,
 144                                              max_tag), 0);
 145
 146                 gcr_kernel_excl = ~incl & SYS_GCR_EL1_EXCL_MASK;
 147                 gcr_kernel_excl_initialized = true;
 148         }
 149
 150         /* Enable the kernel exclude mask for random tags generation. */
 151         write_sysreg_s(SYS_GCR_EL1_RRND | gcr_kernel_excl, SYS_GCR_EL1);
 152 }
 153
 154 void mte_enable_kernel(void)
 155 {
 156         /* Enable MTE Sync Mode for EL1. */
 157         sysreg_clear_set(sctlr_el1, SCTLR_ELx_TCF_MASK, SCTLR_ELx_TCF_SYNC);
 158         isb();
 159 }
 160
 161 static void update_sctlr_el1_tcf0(u64 tcf0)
 162 {
 163         /* ISB required for the kernel uaccess routines */
 164         sysreg_clear_set(sctlr_el1, SCTLR_EL1_TCF0_MASK, tcf0);
 165         isb();
 166 }
 167
 168 static void set_sctlr_el1_tcf0(u64 tcf0)
 169 {
 170         /*
 171          * mte_thread_switch() checks current->thread.sctlr_tcf0 as an
 172          * optimisation. Disable preemption so that it does not see
 173          * the variable update before the SCTLR_EL1.TCF0 one.
 174          */
 175         preempt_disable();
 176         current->thread.sctlr_tcf0 = tcf0;
 177         update_sctlr_el1_tcf0(tcf0);
 178         preempt_enable();
 179 }
 180
 181 static void update_gcr_el1_excl(u64 excl)
 182 {
 183
 184         /*
 185          * Note that the mask controlled by the user via prctl() is an
 186          * include while GCR_EL1 accepts an exclude mask.
 187          * No need for ISB since this only affects EL0 currently, implicit
 188          * with ERET.
 189          */
 190         sysreg_clear_set_s(SYS_GCR_EL1, SYS_GCR_EL1_EXCL_MASK, excl);
 191 }
 192
 193 static void set_gcr_el1_excl(u64 excl)
 194 {
 195         current->thread.gcr_user_excl = excl;
 196
 197         /*
 198          * SYS_GCR_EL1 will be set to current->thread.gcr_user_excl value
 199          * by mte_set_user_gcr() in kernel_exit,
 200          */
 201 }
 202
 203 void flush_mte_state(void)
 204 {
 205         if (!system_supports_mte())
 206                 return;
 207
 208         /* clear any pending asynchronous tag fault */
 209         dsb(ish);
 210         write_sysreg_s(0, SYS_TFSRE0_EL1);
 211         clear_thread_flag(TIF_MTE_ASYNC_FAULT);
 212         /* disable tag checking */
 213         set_sctlr_el1_tcf0(SCTLR_EL1_TCF0_NONE);
 214         /* reset tag generation mask */
 215         set_gcr_el1_excl(SYS_GCR_EL1_EXCL_MASK);
 216 }
 217
 218 void mte_thread_switch(struct task_struct *next)
 219 {
 220         if (!system_supports_mte())
 221                 return;
 222
 223         /* avoid expensive SCTLR_EL1 accesses if no change */
 224         if (current->thread.sctlr_tcf0 != next->thread.sctlr_tcf0)
 225                 update_sctlr_el1_tcf0(next->thread.sctlr_tcf0);
 226 }
 227
 228 void mte_suspend_exit(void)
 229 {
 230         if (!system_supports_mte())
 231                 return;
 232
 233         update_gcr_el1_excl(gcr_kernel_excl);
 234 }
 235
 236 long set_mte_ctrl(struct task_struct *task, unsigned long arg)
 237 {
 238         u64 tcf0;
 239         u64 gcr_excl = ~((arg & PR_MTE_TAG_MASK) >> PR_MTE_TAG_SHIFT) &
 240                        SYS_GCR_EL1_EXCL_MASK;
 241
 242         if (!system_supports_mte())
 243                 return 0;
 244
 245         switch (arg & PR_MTE_TCF_MASK) {
 246         case PR_MTE_TCF_NONE:
 247                 tcf0 = SCTLR_EL1_TCF0_NONE;
 248                 break;
 249         case PR_MTE_TCF_SYNC:
 250                 tcf0 = SCTLR_EL1_TCF0_SYNC;
 251                 break;
 252         case PR_MTE_TCF_ASYNC:
 253                 tcf0 = SCTLR_EL1_TCF0_ASYNC;
 254                 break;
 255         default:
 256                 return -EINVAL;
 257         }
 258
 259         if (task != current) {
 260                 task->thread.sctlr_tcf0 = tcf0;
 261                 task->thread.gcr_user_excl = gcr_excl;
 262         } else {
 263                 set_sctlr_el1_tcf0(tcf0);
 264                 set_gcr_el1_excl(gcr_excl);
 265         }
 266
 267         return 0;
 268 }
 269
 270 long get_mte_ctrl(struct task_struct *task)
 271 {
 272         unsigned long ret;
 273         u64 incl = ~task->thread.gcr_user_excl & SYS_GCR_EL1_EXCL_MASK;
 274
 275         if (!system_supports_mte())
 276                 return 0;
 277
 278         ret = incl << PR_MTE_TAG_SHIFT;
 279
 280         switch (task->thread.sctlr_tcf0) {
 281         case SCTLR_EL1_TCF0_NONE:
 282                 ret |= PR_MTE_TCF_NONE;
 283                 break;
 284         case SCTLR_EL1_TCF0_SYNC:
 285                 ret |= PR_MTE_TCF_SYNC;
 286                 break;
 287         case SCTLR_EL1_TCF0_ASYNC:
 288                 ret |= PR_MTE_TCF_ASYNC;
 289                 break;
 290         }
 291
 292         return ret;
 293 }
 294
 295 /*
 296  * Access MTE tags in another process' address space as given in mm. Update
 297  * the number of tags copied. Return 0 if any tags copied, error otherwise.
 298  * Inspired by __access_remote_vm().
 299  */
 300 static int __access_remote_tags(struct mm_struct *mm, unsigned long addr,
 301                                 struct iovec *kiov, unsigned int gup_flags)
 302 {
 303         struct vm_area_struct *vma;
 304         void __user *buf = kiov->iov_base;
 305         size_t len = kiov->iov_len;
 306         int ret;
 307         int write = gup_flags & FOLL_WRITE;
 308
 309         if (!access_ok(buf, len))
 310                 return -EFAULT;
 311
 312         if (mmap_read_lock_killable(mm))
 313                 return -EIO;
 314
 315         while (len) {
 316                 unsigned long tags, offset;
 317                 void *maddr;
 318                 struct page *page = NULL;
 319
 320                 ret = get_user_pages_remote(mm, addr, 1, gup_flags, &page,
 321                                             &vma, NULL);
 322                 if (ret <= 0)
 323                         break;
 324
 325                 /*
 326                  * Only copy tags if the page has been mapped as PROT_MTE
 327                  * (PG_mte_tagged set). Otherwise the tags are not valid and
 328                  * not accessible to user. Moreover, an mprotect(PROT_MTE)
 329                  * would cause the existing tags to be cleared if the page
 330                  * was never mapped with PROT_MTE.
 331                  */
 332                 if (!test_bit(PG_mte_tagged, &page->flags)) {
 333                         ret = -EOPNOTSUPP;
 334                         put_page(page);
 335                         break;
 336                 }
 337
 338                 /* limit access to the end of the page */
 339                 offset = offset_in_page(addr);
 340                 tags = min(len, (PAGE_SIZE - offset) / MTE_GRANULE_SIZE);
 341
 342                 maddr = page_address(page);
 343                 if (write) {
 344                         tags = mte_copy_tags_from_user(maddr + offset, buf, tags);
 345                         set_page_dirty_lock(page);
 346                 } else {
 347                         tags = mte_copy_tags_to_user(buf, maddr + offset, tags);
 348                 }
 349                 put_page(page);
 350
 351                 /* error accessing the tracer's buffer */
 352                 if (!tags)
 353                         break;
 354
 355                 len -= tags;
 356                 buf += tags;
 357                 addr += tags * MTE_GRANULE_SIZE;
 358         }
 359         mmap_read_unlock(mm);
 360
 361         /* return an error if no tags copied */
 362         kiov->iov_len = buf - kiov->iov_base;
 363         if (!kiov->iov_len) {
 364                 /* check for error accessing the tracee's address space */
 365                 if (ret <= 0)
 366                         return -EIO;
 367                 else
 368                         return -EFAULT;
 369         }
 370
 371         return 0;
 372 }
 373
 374 /*
 375  * Copy MTE tags in another process' address space at 'addr' to/from tracer's
 376  * iovec buffer. Return 0 on success. Inspired by ptrace_access_vm().
 377  */
 378 static int access_remote_tags(struct task_struct *tsk, unsigned long addr,
 379                               struct iovec *kiov, unsigned int gup_flags)
 380 {
 381         struct mm_struct *mm;
 382         int ret;
 383
 384         mm = get_task_mm(tsk);
 385         if (!mm)
 386                 return -EPERM;
 387
 388         if (!tsk->ptrace || (current != tsk->parent) ||
 389             ((get_dumpable(mm) != SUID_DUMP_USER) &&
 390              !ptracer_capable(tsk, mm->user_ns))) {
 391                 mmput(mm);
 392                 return -EPERM;
 393         }
 394
 395         ret = __access_remote_tags(mm, addr, kiov, gup_flags);
 396         mmput(mm);
 397
 398         return ret;
 399 }
 400
 401 int mte_ptrace_copy_tags(struct task_struct *child, long request,
 402                          unsigned long addr, unsigned long data)
 403 {
 404         int ret;
 405         struct iovec kiov;
 406         struct iovec __user *uiov = (void __user *)data;
 407         unsigned int gup_flags = FOLL_FORCE;
 408
 409         if (!system_supports_mte())
 410                 return -EIO;
 411
 412         if (get_user(kiov.iov_base, &uiov->iov_base) ||
 413             get_user(kiov.iov_len, &uiov->iov_len))
 414                 return -EFAULT;
 415
 416         if (request == PTRACE_POKEMTETAGS)
 417                 gup_flags |= FOLL_WRITE;
 418
 419         /* align addr to the MTE tag granule */
 420         addr &= MTE_GRANULE_MASK;
 421
 422         ret = access_remote_tags(child, addr, &kiov, gup_flags);
 423         if (!ret)
 424                 ret = put_user(kiov.iov_len, &uiov->iov_len);
 425
 426         return ret;
 427 }