On Sat, 11 Oct 2025 at 21:20, Matthew Lugg <[email protected]> wrote:
>
> This typo meant that calls to `mremap` which shrink a mapping by some N
> bytes would, when the virtual address space was pre-reserved (e.g.
> 32-bit guest on 64-bit host), unmap the N bytes following the *original*
> mapping.
>
> Signed-off-by: Matthew Lugg <[email protected]>
> ---
>  linux-user/mmap.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/linux-user/mmap.c b/linux-user/mmap.c
> index 847092a28a..ec8392b35b 100644
> --- a/linux-user/mmap.c
> +++ b/linux-user/mmap.c
> @@ -1164,7 +1164,8 @@ abi_long target_mremap(abi_ulong old_addr, abi_ulong 
> old_size,
>                      errno = ENOMEM;
>                      host_addr = MAP_FAILED;
>                  } else if (reserved_va && old_size > new_size) {
> -                    mmap_reserve_or_unmap(old_addr + old_size,
> +                    /* Re-reserve pages we just shrunk out of the mapping */
> +                    mmap_reserve_or_unmap(old_addr + new_size,
>                                            old_size - new_size);
>                  }
>              }
> --

Reviewed-by: Peter Maydell <[email protected]>

I agree with your cover letter:

Cc: [email protected]

I think this has been broken for a long time, right back to
the introduction of pre-allocated guest address space in
commit 68a1c8168, where the code was clearly confused between
old_size and new_size:

+            if (host_addr != MAP_FAILED && reserved_va && old_size >
new_size) {
+                mmap_reserve(old_addr + old_size, new_size - old_size);
+            }

In 2020 commit 257a7e212d5e5 fixed half of this problem
(swapping the two sizes in the second argument to
mmap_reserve()) but we didn't notice that the first
argument was also wrong.

thanks
-- PMM

Reply via email to