On Mon 27-11-17 19:53:18, guoxuenan wrote:
> From: chenjie <chenj...@huawei.com>
>
> The madvise() system call supported a set of "conventional" advice values,
> the MADV_WILLNEED parameter has possibility of triggering an infinite loop
> under
> direct access mode(DAX).
>
> Infinite loop situation:
> 1、initial state [ start = vam->vm_start < vam->vm_end < end ].
> 2、madvise_vma() using MADV_WILLNEED parameter;
> madvise_vma() -> madvise_willneed() -> return 0 && the value of [prev] is
> not updated.
>
> In function SYSCALL_DEFINE3(madvise,...)
> When [start = vam->vm_start] the program enters "for" loop,
> find_vma_prev() will set the pointer vma and the pointer prev(prev =
> vam->vm_prev).
> Normally ,madvise_vma() will always move the pointer prev ,but when use DAX
> mode,
> it will never update the value of [prev].
>
> =======================================================================
> SYSCALL_DEFINE3(madvise,...)
> {
> [...]
> //start = vam->start => prev=vma->prev
> vma = find_vma_prev(current->mm, start, &prev);
> [...]
> for(;;)
> {
> update [start = vma->vm_start]
>
> con0: if (start >= end) //false always;
> goto out;
> tmp = vma->vm_end;
>
> //do not update [prev] and always return 0;
> error = madvise_willneed();
>
> con1: if (error) //false always;
> goto out;
>
> //[ vam->vm_start < start = vam->vm_end <end ]
> update [start = tmp ]
>
> con2: if (start >= end) //false always ;
> goto out;
>
> //because of pointer [prev] did not change,[vma] keep as it was;
> update [ vma = prev->vm_next ]
> }
> [...]
> }
> =======================================================================
> After the first cycle ;it will always keep
> vam->vm_start < start = vam->vm_end < end && vma = prev->vm_next;
> since Circulation exit conditions (con{0,1,2}) will never meet ,the
> program stuck in infinite loop.
I find your changelog a bit hard to parse. What would you think about
the following:
"
MADVISE_WILLNEED has always been a noop for DAX (formerly XIP) mappings.
Unfortunatelly madvise_willneed doesn't communicate this information
properly to the generic madvise syscall implementation. The calling
converion is quite subtle there. madvise_vma is supposed to either
return an error or update &prev otherwise the main loop will never
advance to the next vma and it will keep looping for ever without a way
to get out of the kernel.
It seems this has been broken since introduced. Nobody has noticed
because nobody seems to be using MADVISE_WILLNEED on these DAX mappings.
Fixes: fe77ba6f4f97 ("[PATCH] xip: madvice/fadvice: execute in place")
Cc: stable
"
> Signed-off-by: chenjie <chenj...@huawei.com>
> Signed-off-by: guoxuenan <guoxue...@huawei.com>
Other than that
Acked-by: Michal Hocko <mho...@suse.com>
> ---
> mm/madvise.c | 4 +---
> 1 file changed, 1 insertion(+), 3 deletions(-)
>
> diff --git a/mm/madvise.c b/mm/madvise.c
> index 375cf32..751e97a 100644
> --- a/mm/madvise.c
> +++ b/mm/madvise.c
> @@ -276,15 +276,14 @@ static long madvise_willneed(struct vm_area_struct *vma,
> {
> struct file *file = vma->vm_file;
>
> + *prev = vma;
> #ifdef CONFIG_SWAP
> if (!file) {
> - *prev = vma;
> force_swapin_readahead(vma, start, end);
> return 0;
> }
>
> if (shmem_mapping(file->f_mapping)) {
> - *prev = vma;
> force_shm_swapin_readahead(vma, start, end,
> file->f_mapping);
> return 0;
> @@ -299,7 +298,6 @@ static long madvise_willneed(struct vm_area_struct *vma,
> return 0;
> }
>
> - *prev = vma;
> start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
> if (end > vma->vm_end)
> end = vma->vm_end;
> --
> 2.9.5
>
--
Michal Hocko
SUSE Labs
