On 03/07/2018 05:35 PM, Yisheng Xie wrote:
> Hi Mike,
> On 2018/3/8 7:59, Mike Kravetz wrote:
>> A vma with vm_pgoff large enough to overflow a loff_t type when
>> converted to a byte offset can be passed via the remap_file_pages
>> system call. The hugetlbfs mmap routine uses the byte offset to
>> calculate reservations and file size.
>> A sequence such as:
>> mmap(0x20a00000, 0x600000, 0, 0x66033, -1, 0);
>> remap_file_pages(0x20a00000, 0x600000, 0, 0x20000000000000, 0);
>> will result in the following when task exits/file closed,
>> kernel BUG at mm/hugetlb.c:749!
>> Call Trace:
>> The overflowed pgoff value causes hugetlbfs to try to set up a
>> mapping with a negative range (end < start) that leaves invalid
>> state which causes the BUG.
>> Reported-by: Nic Losby <blurbd...@gmail.com>
>> Signed-off-by: Mike Kravetz <mike.krav...@oracle.com>
>> fs/hugetlbfs/inode.c | 7 ++++---
>> 1 file changed, 4 insertions(+), 3 deletions(-)
>> diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c
>> index 8fe1b0aa2896..cb288dec5564 100644
>> --- a/fs/hugetlbfs/inode.c
>> +++ b/fs/hugetlbfs/inode.c
>> @@ -127,12 +127,13 @@ static int hugetlbfs_file_mmap(struct file *file,
>> struct vm_area_struct *vma)
>> vma->vm_ops = &hugetlb_vm_ops;
>> - * Offset passed to mmap (before page shift) could have been
>> - * negative when represented as a (l)off_t.
>> + * page based offset in vm_pgoff could be sufficiently large to
>> + * overflow a (l)off_t when converted to byte offset.
>> - if (((loff_t)vma->vm_pgoff << PAGE_SHIFT) < 0)
>> + if (vma->vm_pgoff && ((loff_t)vma->vm_pgoff << PAGE_SHIFT) <= 0)
>> return -EINVAL;
> This seems still no the right fix, taking the following case as an example:
> mmap(0x20a00000, 0x600000, 0, 0x66033, -1, 0);
> remap_file_pages(0x20a00000, 0x600000, 0, 0x0020001000000000, 0);
> You should just check the highest PAGE_SHIFT+1 bits of pgoff in you want check
> at this point, right?
Yes, thank you!
That would be the correct check and also much simpler. Something like,
unsigned long ovfl_mask;
ovfl_mask = (1UL << (PAGE_SHIFT + 1)) - 1;
ovfl_mask <<= ((sizeof(unsigned long) * BITS_PER_BYTE) -
(PAGE_SHIFT + 1));
if (vma->vm_pgoff & ovfl_mask)
> However, region_chg makes me a litter puzzle that when its return value < 0,
> adds_in_progress is added like this case, while sometime it is not. so why
> not just
> change at the beginning of region_chg ?
> if (f > t)
> return -EINVAL;
If region_chg returns a value < 0, this indicates an error and adds_in_progress
should not be incremented. In the case of this bug, region_chg was passed
values where f > t. Of course, this should never happen. But, because it
assumed f <= t, it returned a negative count needed huge page reservations.
The calling code interpreted the negative value as an error and a subsequent
region_add or region_abort.
I am not opposed to adding the suggested "if (f > t)". However, the
region tracking routines are simple helpers only used by the hugetlbfs
code and the assumption is that they are being called correctly. As
such, I would prefer to leave off the check. But, this is the second
time they have been called incorrectly due to insufficient argument
checking. If we do add this to region_chg, I would also add the check
to all region_* routines for consistency.
I will send out a V2 of this patch tomorrow with the corrected overflow
checking and possibly checks added to the region_* routines.