Hi Miles,
Could you please cc me and Andrey Konovalov on future versions of this
patch (if any)?
On Mon, 23 Nov 2020 at 08:47, Miles Chen <miles.chen@xxxxxxxxxxxx> wrote:
> When we try to visit the pagemap of a tagged userspace pointer, we find
> that the start_vaddr is not correct because of the tag.
> To fix it, we should untag the usespace pointers in pagemap_read().
>
> I tested with 5.10-rc4 and the issue remains.
>
> My test code is baed on [1]:
>
> A userspace pointer which has been tagged by 0xb4: 0xb400007662f541c8
>
> === userspace program ===
>
> uint64 OsLayer::VirtualToPhysical(void *vaddr) {
> uint64 frame, paddr, pfnmask, pagemask;
> int pagesize = sysconf(_SC_PAGESIZE);
> off64_t off = ((uintptr_t)vaddr) / pagesize * 8; // off = 0xb400007662f541c8 / pagesize * 8 = 0x5a00003b317aa0
Arguably, that's a user-space bug since tagged file offsets were never
supported. In this case it's not even a tag at bit 56 as per the arm64
tagged address ABI but rather down to bit 47. You could say that the
problem is caused by the C library (malloc()) or whoever created the
tagged vaddr and passed it to this function. It's not a kernel
regression as we've never supported it.
Now, pagemap is a special case where the offset is usually not
generated as a classic file offset but rather derived by shifting a
user virtual address. I guess we can make a concession for pagemap
(only) and allow such offset with the tag at bit (56 - PAGE_SHIFT +
3).
Please fix the patch as per Eric's suggestion on avoiding the
overflow. You should also add a Cc: stable v5.4- as that's when we
enabled the tagged address ABI on arm64 and when it's more likely for
the C library/malloc() to start generating such pointers.
If the problem is only limited to this test, I'd rather fix the user
but I can't tell how widespread the /proc/pid/pagemap usage is.
Thanks.
--
Catalin
