Re: [PATCH] Fix integer overflow at left shift expression

[Maxim Fomin]

------- Original Message -------
On Sunday, December 4th, 2022 at 1:06 PM, Maxim Fomin <maxim@fomin.one> wrote:
> 
> From db82faafba5e7eccd9fd6c0b7314f7322c1aecbd Mon Sep 17 00:00:00 2001
> From: Maxim Fomin maxim@fomin.one
> 
> Date: Sun, 4 Dec 2022 12:05:34 +0000
> Subject: [PATCH] Fix integer overflow at left shift expression.
> 
> In case of large partitions (>1TiB) left shift
> 
> with signed int GRUB_DISK_SECTOR_BITS macro may
> cause integer overflow which results in wrong
> partition size.
> ---
> grub-core/kern/fs.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/grub-core/kern/fs.c b/grub-core/kern/fs.c
> index b9508296d..c196f2bf1 100644
> --- a/grub-core/kern/fs.c
> +++ b/grub-core/kern/fs.c
> @@ -130,7 +130,7 @@ grub_fs_probe (grub_device_t device)
> struct grub_fs_block
> {
> grub_disk_addr_t offset;
> - unsigned long length;
> + grub_disk_addr_t length;
> };
> 
> static grub_err_t
> @@ -195,7 +195,7 @@ grub_fs_blocklist_open (grub_file_t file, const char 
> *name)
> goto fail;
> }
> 
> - file->size += (blocks[i].length << GRUB_DISK_SECTOR_BITS);
> 
> + file->size += (blocks[i].length << (grub_disk_addr_t) 
> GRUB_DISK_SECTOR_BITS);
> 
> p++;
> }
> 
> --
> 2.38.1
> 
> 
> 
> _______________________________________________
> Grub-devel mailing list
> Grub-devel@gnu.org
> https://lists.gnu.org/mailman/listinfo/grub-devel

Some comments. This issue was found during work on plainmount
patch. Previously I tested patch on i386-pc platform only
partially because 'loopback name (disk,part)offset+' command
was giving wrong partition sizes. These partition sizes looked
like there is UINT truncation happening when I boot from my 2
TB drive. I considered this as a BIOS bug because there are
other bugs with this laptop (not related to GRUB).

Recently I discovered that exactly the same bug occurs booting
postUEFI laptop in i386-pc mode on the same 2TB drive. Loopback
module sets size from 'dev->file->size' (loopback.c:166) which
is set in file.c:111 by calling '(file->fs->fs_open) (file, file_name)'
which is blocklist callback (fs.c:137). The root of the bug
is that the expression 'blocks[i].length << GRUB_DISK_SECTOR_BITS'
yields low value in i386-pc mode and correct value in x86_64-efi mode.


Best regards,
Maxim Fomin