2 * mpx.c - Memory Protection eXtensions
4 * Copyright (c) 2014, Intel Corporation.
5 * Qiaowei Ren <qiaowei.ren@intel.com>
6 * Dave Hansen <dave.hansen@intel.com>
8 #include <linux/kernel.h>
9 #include <linux/slab.h>
10 #include <linux/syscalls.h>
11 #include <linux/sched/sysctl.h>
15 #include <asm/mmu_context.h>
17 #include <asm/processor.h>
18 #include <asm/fpu/internal.h>
20 #define CREATE_TRACE_POINTS
21 #include <asm/trace/mpx.h>
23 static inline unsigned long mpx_bd_size_bytes(struct mm_struct *mm)
26 return MPX_BD_SIZE_BYTES_64;
28 return MPX_BD_SIZE_BYTES_32;
31 static inline unsigned long mpx_bt_size_bytes(struct mm_struct *mm)
34 return MPX_BT_SIZE_BYTES_64;
36 return MPX_BT_SIZE_BYTES_32;
40 * This is really a simplified "vm_mmap". it only handles MPX
41 * bounds tables (the bounds directory is user-allocated).
43 static unsigned long mpx_mmap(unsigned long len)
46 unsigned long addr, pgoff;
47 struct mm_struct *mm = current->mm;
49 struct vm_area_struct *vma;
51 /* Only bounds table can be allocated here */
52 if (len != mpx_bt_size_bytes(mm))
55 down_write(&mm->mmap_sem);
57 /* Too many mappings? */
58 if (mm->map_count > sysctl_max_map_count) {
63 /* Obtain the address to map to. we verify (or select) it and ensure
64 * that it represents a valid section of the address space.
66 addr = get_unmapped_area(NULL, 0, len, 0, MAP_ANONYMOUS | MAP_PRIVATE);
67 if (addr & ~PAGE_MASK) {
72 vm_flags = VM_READ | VM_WRITE | VM_MPX |
73 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
75 /* Set pgoff according to addr for anon_vma */
76 pgoff = addr >> PAGE_SHIFT;
78 ret = mmap_region(NULL, addr, len, vm_flags, pgoff);
79 if (IS_ERR_VALUE(ret))
82 vma = find_vma(mm, ret);
88 if (vm_flags & VM_LOCKED) {
89 up_write(&mm->mmap_sem);
90 mm_populate(ret, len);
95 up_write(&mm->mmap_sem);
105 static int get_reg_offset(struct insn *insn, struct pt_regs *regs,
110 static const int regoff[] = {
111 offsetof(struct pt_regs, ax),
112 offsetof(struct pt_regs, cx),
113 offsetof(struct pt_regs, dx),
114 offsetof(struct pt_regs, bx),
115 offsetof(struct pt_regs, sp),
116 offsetof(struct pt_regs, bp),
117 offsetof(struct pt_regs, si),
118 offsetof(struct pt_regs, di),
120 offsetof(struct pt_regs, r8),
121 offsetof(struct pt_regs, r9),
122 offsetof(struct pt_regs, r10),
123 offsetof(struct pt_regs, r11),
124 offsetof(struct pt_regs, r12),
125 offsetof(struct pt_regs, r13),
126 offsetof(struct pt_regs, r14),
127 offsetof(struct pt_regs, r15),
130 int nr_registers = ARRAY_SIZE(regoff);
132 * Don't possibly decode a 32-bit instructions as
133 * reading a 64-bit-only register.
135 if (IS_ENABLED(CONFIG_X86_64) && !insn->x86_64)
140 regno = X86_MODRM_RM(insn->modrm.value);
141 if (X86_REX_B(insn->rex_prefix.value) == 1)
146 regno = X86_SIB_INDEX(insn->sib.value);
147 if (X86_REX_X(insn->rex_prefix.value) == 1)
152 regno = X86_SIB_BASE(insn->sib.value);
153 if (X86_REX_B(insn->rex_prefix.value) == 1)
158 pr_err("invalid register type");
163 if (regno > nr_registers) {
164 WARN_ONCE(1, "decoded an instruction with an invalid register");
167 return regoff[regno];
171 * return the address being referenced be instruction
172 * for rm=3 returning the content of the rm reg
173 * for rm!=3 calculates the address using SIB and Disp
175 static void __user *mpx_get_addr_ref(struct insn *insn, struct pt_regs *regs)
177 unsigned long addr, base, indx;
178 int addr_offset, base_offset, indx_offset;
181 insn_get_modrm(insn);
183 sib = insn->sib.value;
185 if (X86_MODRM_MOD(insn->modrm.value) == 3) {
186 addr_offset = get_reg_offset(insn, regs, REG_TYPE_RM);
189 addr = regs_get_register(regs, addr_offset);
191 if (insn->sib.nbytes) {
192 base_offset = get_reg_offset(insn, regs, REG_TYPE_BASE);
196 indx_offset = get_reg_offset(insn, regs, REG_TYPE_INDEX);
200 base = regs_get_register(regs, base_offset);
201 indx = regs_get_register(regs, indx_offset);
202 addr = base + indx * (1 << X86_SIB_SCALE(sib));
204 addr_offset = get_reg_offset(insn, regs, REG_TYPE_RM);
207 addr = regs_get_register(regs, addr_offset);
209 addr += insn->displacement.value;
211 return (void __user *)addr;
213 return (void __user *)-1;
216 static int mpx_insn_decode(struct insn *insn,
217 struct pt_regs *regs)
219 unsigned char buf[MAX_INSN_SIZE];
220 int x86_64 = !test_thread_flag(TIF_IA32);
224 not_copied = copy_from_user(buf, (void __user *)regs->ip, sizeof(buf));
225 nr_copied = sizeof(buf) - not_copied;
227 * The decoder _should_ fail nicely if we pass it a short buffer.
228 * But, let's not depend on that implementation detail. If we
229 * did not get anything, just error out now.
233 insn_init(insn, buf, nr_copied, x86_64);
234 insn_get_length(insn);
236 * copy_from_user() tries to get as many bytes as we could see in
237 * the largest possible instruction. If the instruction we are
238 * after is shorter than that _and_ we attempt to copy from
239 * something unreadable, we might get a short read. This is OK
240 * as long as the read did not stop in the middle of the
241 * instruction. Check to see if we got a partial instruction.
243 if (nr_copied < insn->length)
246 insn_get_opcode(insn);
248 * We only _really_ need to decode bndcl/bndcn/bndcu
249 * Error out on anything else.
251 if (insn->opcode.bytes[0] != 0x0f)
253 if ((insn->opcode.bytes[1] != 0x1a) &&
254 (insn->opcode.bytes[1] != 0x1b))
263 * If a bounds overflow occurs then a #BR is generated. This
264 * function decodes MPX instructions to get violation address
265 * and set this address into extended struct siginfo.
267 * Note that this is not a super precise way of doing this.
268 * Userspace could have, by the time we get here, written
269 * anything it wants in to the instructions. We can not
270 * trust anything about it. They might not be valid
271 * instructions or might encode invalid registers, etc...
273 * The caller is expected to kfree() the returned siginfo_t.
275 siginfo_t *mpx_generate_siginfo(struct pt_regs *regs)
277 const struct bndreg *bndregs, *bndreg;
278 siginfo_t *info = NULL;
283 err = mpx_insn_decode(&insn, regs);
288 * We know at this point that we are only dealing with
291 insn_get_modrm(&insn);
292 bndregno = X86_MODRM_REG(insn.modrm.value);
297 /* get bndregs field from current task's xsave area */
298 bndregs = get_xsave_field_ptr(XSTATE_BNDREGS);
303 /* now go select the individual register in the set of 4 */
304 bndreg = &bndregs[bndregno];
306 info = kzalloc(sizeof(*info), GFP_KERNEL);
312 * The registers are always 64-bit, but the upper 32
313 * bits are ignored in 32-bit mode. Also, note that the
314 * upper bounds are architecturally represented in 1's
317 * The 'unsigned long' cast is because the compiler
318 * complains when casting from integers to different-size
321 info->si_lower = (void __user *)(unsigned long)bndreg->lower_bound;
322 info->si_upper = (void __user *)(unsigned long)~bndreg->upper_bound;
323 info->si_addr_lsb = 0;
324 info->si_signo = SIGSEGV;
326 info->si_code = SEGV_BNDERR;
327 info->si_addr = mpx_get_addr_ref(&insn, regs);
329 * We were not able to extract an address from the instruction,
330 * probably because there was something invalid in it.
332 if (info->si_addr == (void *)-1) {
336 trace_mpx_bounds_register_exception(info->si_addr, bndreg);
339 /* info might be NULL, but kfree() handles that */
344 static __user void *mpx_get_bounds_dir(void)
346 const struct bndcsr *bndcsr;
348 if (!cpu_feature_enabled(X86_FEATURE_MPX))
349 return MPX_INVALID_BOUNDS_DIR;
352 * The bounds directory pointer is stored in a register
353 * only accessible if we first do an xsave.
355 bndcsr = get_xsave_field_ptr(XSTATE_BNDCSR);
357 return MPX_INVALID_BOUNDS_DIR;
360 * Make sure the register looks valid by checking the
363 if (!(bndcsr->bndcfgu & MPX_BNDCFG_ENABLE_FLAG))
364 return MPX_INVALID_BOUNDS_DIR;
367 * Lastly, mask off the low bits used for configuration
368 * flags, and return the address of the bounds table.
370 return (void __user *)(unsigned long)
371 (bndcsr->bndcfgu & MPX_BNDCFG_ADDR_MASK);
374 int mpx_enable_management(void)
376 void __user *bd_base = MPX_INVALID_BOUNDS_DIR;
377 struct mm_struct *mm = current->mm;
381 * runtime in the userspace will be responsible for allocation of
382 * the bounds directory. Then, it will save the base of the bounds
383 * directory into XSAVE/XRSTOR Save Area and enable MPX through
384 * XRSTOR instruction.
386 * The copy_xregs_to_kernel() beneath get_xsave_field_ptr() is
387 * expected to be relatively expensive. Storing the bounds
388 * directory here means that we do not have to do xsave in the
389 * unmap path; we can just use mm->bd_addr instead.
391 bd_base = mpx_get_bounds_dir();
392 down_write(&mm->mmap_sem);
393 mm->bd_addr = bd_base;
394 if (mm->bd_addr == MPX_INVALID_BOUNDS_DIR)
397 up_write(&mm->mmap_sem);
401 int mpx_disable_management(void)
403 struct mm_struct *mm = current->mm;
405 if (!cpu_feature_enabled(X86_FEATURE_MPX))
408 down_write(&mm->mmap_sem);
409 mm->bd_addr = MPX_INVALID_BOUNDS_DIR;
410 up_write(&mm->mmap_sem);
414 static int mpx_cmpxchg_bd_entry(struct mm_struct *mm,
415 unsigned long *curval,
416 unsigned long __user *addr,
417 unsigned long old_val, unsigned long new_val)
421 * user_atomic_cmpxchg_inatomic() actually uses sizeof()
422 * the pointer that we pass to it to figure out how much
423 * data to cmpxchg. We have to be careful here not to
424 * pass a pointer to a 64-bit data type when we only want
427 if (is_64bit_mm(mm)) {
428 ret = user_atomic_cmpxchg_inatomic(curval,
429 addr, old_val, new_val);
431 u32 uninitialized_var(curval_32);
432 u32 old_val_32 = old_val;
433 u32 new_val_32 = new_val;
434 u32 __user *addr_32 = (u32 __user *)addr;
436 ret = user_atomic_cmpxchg_inatomic(&curval_32,
437 addr_32, old_val_32, new_val_32);
444 * With 32-bit mode, a bounds directory is 4MB, and the size of each
445 * bounds table is 16KB. With 64-bit mode, a bounds directory is 2GB,
446 * and the size of each bounds table is 4MB.
448 static int allocate_bt(struct mm_struct *mm, long __user *bd_entry)
450 unsigned long expected_old_val = 0;
451 unsigned long actual_old_val = 0;
452 unsigned long bt_addr;
453 unsigned long bd_new_entry;
457 * Carve the virtual space out of userspace for the new
460 bt_addr = mpx_mmap(mpx_bt_size_bytes(mm));
461 if (IS_ERR((void *)bt_addr))
462 return PTR_ERR((void *)bt_addr);
464 * Set the valid flag (kinda like _PAGE_PRESENT in a pte)
466 bd_new_entry = bt_addr | MPX_BD_ENTRY_VALID_FLAG;
469 * Go poke the address of the new bounds table in to the
470 * bounds directory entry out in userspace memory. Note:
471 * we may race with another CPU instantiating the same table.
472 * In that case the cmpxchg will see an unexpected
475 * This can fault, but that's OK because we do not hold
476 * mmap_sem at this point, unlike some of the other part
477 * of the MPX code that have to pagefault_disable().
479 ret = mpx_cmpxchg_bd_entry(mm, &actual_old_val, bd_entry,
480 expected_old_val, bd_new_entry);
485 * The user_atomic_cmpxchg_inatomic() will only return nonzero
486 * for faults, *not* if the cmpxchg itself fails. Now we must
487 * verify that the cmpxchg itself completed successfully.
490 * We expected an empty 'expected_old_val', but instead found
491 * an apparently valid entry. Assume we raced with another
492 * thread to instantiate this table and desclare succecss.
494 if (actual_old_val & MPX_BD_ENTRY_VALID_FLAG) {
499 * We found a non-empty bd_entry but it did not have the
500 * VALID_FLAG set. Return an error which will result in
501 * a SEGV since this probably means that somebody scribbled
502 * some invalid data in to a bounds table.
504 if (expected_old_val != actual_old_val) {
508 trace_mpx_new_bounds_table(bt_addr);
511 vm_munmap(bt_addr, mpx_bt_size_bytes(mm));
516 * When a BNDSTX instruction attempts to save bounds to a bounds
517 * table, it will first attempt to look up the table in the
518 * first-level bounds directory. If it does not find a table in
519 * the directory, a #BR is generated and we get here in order to
520 * allocate a new table.
522 * With 32-bit mode, the size of BD is 4MB, and the size of each
523 * bound table is 16KB. With 64-bit mode, the size of BD is 2GB,
524 * and the size of each bound table is 4MB.
526 static int do_mpx_bt_fault(void)
528 unsigned long bd_entry, bd_base;
529 const struct bndcsr *bndcsr;
530 struct mm_struct *mm = current->mm;
532 bndcsr = get_xsave_field_ptr(XSTATE_BNDCSR);
536 * Mask off the preserve and enable bits
538 bd_base = bndcsr->bndcfgu & MPX_BNDCFG_ADDR_MASK;
540 * The hardware provides the address of the missing or invalid
541 * entry via BNDSTATUS, so we don't have to go look it up.
543 bd_entry = bndcsr->bndstatus & MPX_BNDSTA_ADDR_MASK;
545 * Make sure the directory entry is within where we think
548 if ((bd_entry < bd_base) ||
549 (bd_entry >= bd_base + mpx_bd_size_bytes(mm)))
552 return allocate_bt(mm, (long __user *)bd_entry);
555 int mpx_handle_bd_fault(void)
558 * Userspace never asked us to manage the bounds tables,
561 if (!kernel_managing_mpx_tables(current->mm))
564 if (do_mpx_bt_fault()) {
565 force_sig(SIGSEGV, current);
567 * The force_sig() is essentially "handling" this
568 * exception, so we do not pass up the error
569 * from do_mpx_bt_fault().
576 * A thin wrapper around get_user_pages(). Returns 0 if the
577 * fault was resolved or -errno if not.
579 static int mpx_resolve_fault(long __user *addr, int write)
585 gup_ret = get_user_pages(current, current->mm, (unsigned long)addr,
586 nr_pages, write, force, NULL, NULL);
588 * get_user_pages() returns number of pages gotten.
589 * 0 means we failed to fault in and get anything,
590 * probably because 'addr' is bad.
594 /* Other error, return it */
597 /* must have gup'd a page and gup_ret>0, success */
601 static unsigned long mpx_bd_entry_to_bt_addr(struct mm_struct *mm,
602 unsigned long bd_entry)
604 unsigned long bt_addr = bd_entry;
607 * Bit 0 in a bt_entry is always the valid bit.
609 bt_addr &= ~MPX_BD_ENTRY_VALID_FLAG;
611 * Tables are naturally aligned at 8-byte boundaries
612 * on 64-bit and 4-byte boundaries on 32-bit. The
613 * documentation makes it appear that the low bits
614 * are ignored by the hardware, so we do the same.
620 bt_addr &= ~(align_to_bytes-1);
625 * Get the base of bounds tables pointed by specific bounds
628 static int get_bt_addr(struct mm_struct *mm,
629 long __user *bd_entry_ptr,
630 unsigned long *bt_addr_result)
634 unsigned long bd_entry;
635 unsigned long bt_addr;
637 if (!access_ok(VERIFY_READ, (bd_entry_ptr), sizeof(*bd_entry_ptr)))
644 ret = get_user(bd_entry, bd_entry_ptr);
649 ret = mpx_resolve_fault(bd_entry_ptr, need_write);
651 * If we could not resolve the fault, consider it
652 * userspace's fault and error out.
658 valid_bit = bd_entry & MPX_BD_ENTRY_VALID_FLAG;
659 bt_addr = mpx_bd_entry_to_bt_addr(mm, bd_entry);
662 * When the kernel is managing bounds tables, a bounds directory
663 * entry will either have a valid address (plus the valid bit)
664 * *OR* be completely empty. If we see a !valid entry *and* some
665 * data in the address field, we know something is wrong. This
666 * -EINVAL return will cause a SIGSEGV.
668 if (!valid_bit && bt_addr)
671 * Do we have an completely zeroed bt entry? That is OK. It
672 * just means there was no bounds table for this memory. Make
673 * sure to distinguish this from -EINVAL, which will cause
679 *bt_addr_result = bt_addr;
683 static inline int bt_entry_size_bytes(struct mm_struct *mm)
686 return MPX_BT_ENTRY_BYTES_64;
688 return MPX_BT_ENTRY_BYTES_32;
692 * Take a virtual address and turns it in to the offset in bytes
693 * inside of the bounds table where the bounds table entry
694 * controlling 'addr' can be found.
696 static unsigned long mpx_get_bt_entry_offset_bytes(struct mm_struct *mm,
699 unsigned long bt_table_nr_entries;
700 unsigned long offset = addr;
702 if (is_64bit_mm(mm)) {
703 /* Bottom 3 bits are ignored on 64-bit */
705 bt_table_nr_entries = MPX_BT_NR_ENTRIES_64;
707 /* Bottom 2 bits are ignored on 32-bit */
709 bt_table_nr_entries = MPX_BT_NR_ENTRIES_32;
712 * We know the size of the table in to which we are
713 * indexing, and we have eliminated all the low bits
714 * which are ignored for indexing.
716 * Mask out all the high bits which we do not need
717 * to index in to the table. Note that the tables
718 * are always powers of two so this gives us a proper
721 offset &= (bt_table_nr_entries-1);
723 * We now have an entry offset in terms of *entries* in
724 * the table. We need to scale it back up to bytes.
726 offset *= bt_entry_size_bytes(mm);
731 * How much virtual address space does a single bounds
732 * directory entry cover?
734 * Note, we need a long long because 4GB doesn't fit in
735 * to a long on 32-bit.
737 static inline unsigned long bd_entry_virt_space(struct mm_struct *mm)
739 unsigned long long virt_space = (1ULL << boot_cpu_data.x86_virt_bits);
741 return virt_space / MPX_BD_NR_ENTRIES_64;
743 return virt_space / MPX_BD_NR_ENTRIES_32;
747 * Free the backing physical pages of bounds table 'bt_addr'.
748 * Assume start...end is within that bounds table.
750 static noinline int zap_bt_entries_mapping(struct mm_struct *mm,
751 unsigned long bt_addr,
752 unsigned long start_mapping, unsigned long end_mapping)
754 struct vm_area_struct *vma;
755 unsigned long addr, len;
760 * if we 'end' on a boundary, the offset will be 0 which
761 * is not what we want. Back it up a byte to get the
762 * last bt entry. Then once we have the entry itself,
763 * move 'end' back up by the table entry size.
765 start = bt_addr + mpx_get_bt_entry_offset_bytes(mm, start_mapping);
766 end = bt_addr + mpx_get_bt_entry_offset_bytes(mm, end_mapping - 1);
768 * Move end back up by one entry. Among other things
769 * this ensures that it remains page-aligned and does
770 * not screw up zap_page_range()
772 end += bt_entry_size_bytes(mm);
775 * Find the first overlapping vma. If vma->vm_start > start, there
776 * will be a hole in the bounds table. This -EINVAL return will
779 vma = find_vma(mm, start);
780 if (!vma || vma->vm_start > start)
784 * A NUMA policy on a VM_MPX VMA could cause this bounds table to
785 * be split. So we need to look across the entire 'start -> end'
786 * range of this bounds table, find all of the VM_MPX VMAs, and
790 while (vma && vma->vm_start < end) {
792 * We followed a bounds directory entry down
793 * here. If we find a non-MPX VMA, that's bad,
794 * so stop immediately and return an error. This
795 * probably results in a SIGSEGV.
797 if (!(vma->vm_flags & VM_MPX))
800 len = min(vma->vm_end, end) - addr;
801 zap_page_range(vma, addr, len, NULL);
802 trace_mpx_unmap_zap(addr, addr+len);
805 addr = vma->vm_start;
810 static unsigned long mpx_get_bd_entry_offset(struct mm_struct *mm,
814 * There are several ways to derive the bd offsets. We
815 * use the following approach here:
816 * 1. We know the size of the virtual address space
817 * 2. We know the number of entries in a bounds table
818 * 3. We know that each entry covers a fixed amount of
819 * virtual address space.
820 * So, we can just divide the virtual address by the
821 * virtual space used by one entry to determine which
822 * entry "controls" the given virtual address.
824 if (is_64bit_mm(mm)) {
825 int bd_entry_size = 8; /* 64-bit pointer */
827 * Take the 64-bit addressing hole in to account.
829 addr &= ((1UL << boot_cpu_data.x86_virt_bits) - 1);
830 return (addr / bd_entry_virt_space(mm)) * bd_entry_size;
832 int bd_entry_size = 4; /* 32-bit pointer */
834 * 32-bit has no hole so this case needs no mask
836 return (addr / bd_entry_virt_space(mm)) * bd_entry_size;
839 * The two return calls above are exact copies. If we
840 * pull out a single copy and put it in here, gcc won't
841 * realize that we're doing a power-of-2 divide and use
842 * shifts. It uses a real divide. If we put them up
843 * there, it manages to figure it out (gcc 4.8.3).
847 static int unmap_entire_bt(struct mm_struct *mm,
848 long __user *bd_entry, unsigned long bt_addr)
850 unsigned long expected_old_val = bt_addr | MPX_BD_ENTRY_VALID_FLAG;
851 unsigned long uninitialized_var(actual_old_val);
856 unsigned long cleared_bd_entry = 0;
859 ret = mpx_cmpxchg_bd_entry(mm, &actual_old_val,
860 bd_entry, expected_old_val, cleared_bd_entry);
865 ret = mpx_resolve_fault(bd_entry, need_write);
867 * If we could not resolve the fault, consider it
868 * userspace's fault and error out.
874 * The cmpxchg was performed, check the results.
876 if (actual_old_val != expected_old_val) {
878 * Someone else raced with us to unmap the table.
879 * That is OK, since we were both trying to do
880 * the same thing. Declare success.
885 * Something messed with the bounds directory
886 * entry. We hold mmap_sem for read or write
887 * here, so it could not be a _new_ bounds table
888 * that someone just allocated. Something is
889 * wrong, so pass up the error and SIGSEGV.
894 * Note, we are likely being called under do_munmap() already. To
895 * avoid recursion, do_munmap() will check whether it comes
896 * from one bounds table through VM_MPX flag.
898 return do_munmap(mm, bt_addr, mpx_bt_size_bytes(mm));
901 static int try_unmap_single_bt(struct mm_struct *mm,
902 unsigned long start, unsigned long end)
904 struct vm_area_struct *next;
905 struct vm_area_struct *prev;
907 * "bta" == Bounds Table Area: the area controlled by the
908 * bounds table that we are unmapping.
910 unsigned long bta_start_vaddr = start & ~(bd_entry_virt_space(mm)-1);
911 unsigned long bta_end_vaddr = bta_start_vaddr + bd_entry_virt_space(mm);
912 unsigned long uninitialized_var(bt_addr);
913 void __user *bde_vaddr;
916 * We already unlinked the VMAs from the mm's rbtree so 'start'
917 * is guaranteed to be in a hole. This gets us the first VMA
918 * before the hole in to 'prev' and the next VMA after the hole
921 next = find_vma_prev(mm, start, &prev);
923 * Do not count other MPX bounds table VMAs as neighbors.
924 * Although theoretically possible, we do not allow bounds
925 * tables for bounds tables so our heads do not explode.
926 * If we count them as neighbors here, we may end up with
927 * lots of tables even though we have no actual table
930 while (next && (next->vm_flags & VM_MPX))
931 next = next->vm_next;
932 while (prev && (prev->vm_flags & VM_MPX))
933 prev = prev->vm_prev;
935 * We know 'start' and 'end' lie within an area controlled
936 * by a single bounds table. See if there are any other
937 * VMAs controlled by that bounds table. If there are not
938 * then we can "expand" the are we are unmapping to possibly
939 * cover the entire table.
941 next = find_vma_prev(mm, start, &prev);
942 if ((!prev || prev->vm_end <= bta_start_vaddr) &&
943 (!next || next->vm_start >= bta_end_vaddr)) {
945 * No neighbor VMAs controlled by same bounds
946 * table. Try to unmap the whole thing
948 start = bta_start_vaddr;
952 bde_vaddr = mm->bd_addr + mpx_get_bd_entry_offset(mm, start);
953 ret = get_bt_addr(mm, bde_vaddr, &bt_addr);
955 * No bounds table there, so nothing to unmap.
957 if (ret == -ENOENT) {
964 * We are unmapping an entire table. Either because the
965 * unmap that started this whole process was large enough
966 * to cover an entire table, or that the unmap was small
967 * but was the area covered by a bounds table.
969 if ((start == bta_start_vaddr) &&
970 (end == bta_end_vaddr))
971 return unmap_entire_bt(mm, bde_vaddr, bt_addr);
972 return zap_bt_entries_mapping(mm, bt_addr, start, end);
975 static int mpx_unmap_tables(struct mm_struct *mm,
976 unsigned long start, unsigned long end)
978 unsigned long one_unmap_start;
979 trace_mpx_unmap_search(start, end);
981 one_unmap_start = start;
982 while (one_unmap_start < end) {
984 unsigned long next_unmap_start = ALIGN(one_unmap_start+1,
985 bd_entry_virt_space(mm));
986 unsigned long one_unmap_end = end;
988 * if the end is beyond the current bounds table,
989 * move it back so we only deal with a single one
992 if (one_unmap_end > next_unmap_start)
993 one_unmap_end = next_unmap_start;
994 ret = try_unmap_single_bt(mm, one_unmap_start, one_unmap_end);
998 one_unmap_start = next_unmap_start;
1004 * Free unused bounds tables covered in a virtual address region being
1005 * munmap()ed. Assume end > start.
1007 * This function will be called by do_munmap(), and the VMAs covering
1008 * the virtual address region start...end have already been split if
1009 * necessary, and the 'vma' is the first vma in this range (start -> end).
1011 void mpx_notify_unmap(struct mm_struct *mm, struct vm_area_struct *vma,
1012 unsigned long start, unsigned long end)
1017 * Refuse to do anything unless userspace has asked
1018 * the kernel to help manage the bounds tables,
1020 if (!kernel_managing_mpx_tables(current->mm))
1023 * This will look across the entire 'start -> end' range,
1024 * and find all of the non-VM_MPX VMAs.
1026 * To avoid recursion, if a VM_MPX vma is found in the range
1027 * (start->end), we will not continue follow-up work. This
1028 * recursion represents having bounds tables for bounds tables,
1029 * which should not occur normally. Being strict about it here
1030 * helps ensure that we do not have an exploitable stack overflow.
1033 if (vma->vm_flags & VM_MPX)
1036 } while (vma && vma->vm_start < end);
1038 ret = mpx_unmap_tables(mm, start, end);
1040 force_sig(SIGSEGV, current);