Re: [PATCH] mm/mmap: Fix memory leak in mmap_region()

[Liam Howlett <liam.howlett@oracle.com>]

* Li Zetao <lizetao1@huawei.com> [221026 21:55]:
> There is a memory leak reported by kmemleak:
> 
>   unreferenced object 0xffff88817231ce40 (size 224):
>     comm "mount.cifs", pid 19308, jiffies 4295917571 (age 405.880s)
>     hex dump (first 32 bytes):
>       00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
>       60 c0 b2 00 81 88 ff ff 98 83 01 42 81 88 ff ff  `..........B....
>     backtrace:
>       [<ffffffff81936171>] __alloc_file+0x21/0x250
>       [<ffffffff81937051>] alloc_empty_file+0x41/0xf0
>       [<ffffffff81937159>] alloc_file+0x59/0x710
>       [<ffffffff81937964>] alloc_file_pseudo+0x154/0x210
>       [<ffffffff81741dbf>] __shmem_file_setup+0xff/0x2a0
>       [<ffffffff817502cd>] shmem_zero_setup+0x8d/0x160
>       [<ffffffff817cc1d5>] mmap_region+0x1075/0x19d0
>       [<ffffffff817cd257>] do_mmap+0x727/0x1110
>       [<ffffffff817518b2>] vm_mmap_pgoff+0x112/0x1e0
>       [<ffffffff83adf955>] do_syscall_64+0x35/0x80
>       [<ffffffff83c0006a>] entry_SYSCALL_64_after_hwframe+0x46/0xb0
> 
> The root cause was traced to an error handing path in mmap_region()
> when arch_validate_flags() or mas_preallocate() fails. In the shared
> anonymous mapping sence, vma will be setuped and mapped with a new
> shared anonymous file via shmem_zero_setup(). So in this case, the
> file resource needs to be released.
> 
> Fix it by calling fput(vma->vm_file) when arch_validate_flags() or
> mas_preallocate() returns an error. And for the beauty of the code,
> put fput() under mapping_unmap_writable().

It does look like the unrolling is in the wrong order in that section.

> 
> Fixes: d4af56c5c7c6 ("mm: start tracking VMAs with maple tree")
> Fixes: c462ac288f2c ("mm: Introduce arch_validate_flags()")
> Signed-off-by: Li Zetao <lizetao1@huawei.com>
> ---
>  mm/mmap.c | 10 +++++++---
>  1 file changed, 7 insertions(+), 3 deletions(-)
> 
> diff --git a/mm/mmap.c b/mm/mmap.c
> index e270057ed04e..8530195b3ec5 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -2674,6 +2674,8 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
>  		error = -EINVAL;
>  		if (file)
>  			goto close_and_free_vma;
> +		else if (vm_flags & VM_SHARED)
> +			goto put_vma_file;
>  		else
>  			goto free_vma;
>  	}
> @@ -2682,6 +2684,8 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
>  		error = -ENOMEM;
>  		if (file)
>  			goto close_and_free_vma;
> +		else if (vm_flags & VM_SHARED)
> +			goto put_vma_file;
>  		else
>  			goto free_vma;
>  	}

I am not happy about this getting more complex as it already duplicates
too much code.

> @@ -2746,13 +2750,13 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
>  	if (vma->vm_ops && vma->vm_ops->close)
>  		vma->vm_ops->close(vma);
>  unmap_and_free_vma:
> -	fput(vma->vm_file);
> -	vma->vm_file = NULL;
> -
>  	/* Undo any partial mapping done by a device driver. */
>  	unmap_region(mm, mas.tree, vma, prev, next, vma->vm_start, vma->vm_end);
>  	if (vm_flags & VM_SHARED)

Could we change the above if to "if (file && vm_flags & VM_SHARED)" and
jump to unmap_and_free_vma "if (vma->vm_file)"?  We could then drop your
new goto label. I still think the reodering is correct and worth while.

Or am I missing something?

>  		mapping_unmap_writable(file->f_mapping);
> +put_vma_file:
> +	fput(vma->vm_file);
> +	vma->vm_file = NULL;
>  free_vma:
>  	vm_area_free(vma);
>  unacct_error:
> -- 
> 2.31.1
>