2018-05-28 11:06 GMT+02:00 Jann Horn <jannh@xxxxxxxxxx>:
> On Sat, May 26, 2018 at 4:50 PM, Salvatore Mesoraca
> <s.mesoraca16@xxxxxxxxx> wrote:
>> Prevent a task from opening, in "write" mode, any /proc/*/mem
>> file that operates on the task's mm.
>> /proc/*/mem is mainly a debugging means and, as such, it shouldn't
>> be used by the inspected process itself.
>> Current implementation always allow a task to access its own
>> /proc/*/mem file.
>> A process can use it to overwrite read-only memory, making
>> pointless the use of security_file_mprotect() or other ways to
>> enforce RO memory.
>>
>> Signed-off-by: Salvatore Mesoraca <s.mesoraca16@xxxxxxxxx>
>> ---
>> fs/proc/base.c | 25 ++++++++++++++++++-------
>> fs/proc/internal.h | 3 ++-
>> fs/proc/task_mmu.c | 4 ++--
>> fs/proc/task_nommu.c | 2 +-
>> 4 files changed, 23 insertions(+), 11 deletions(-)
>>
>> diff --git a/fs/proc/base.c b/fs/proc/base.c
>> index 1a76d75..01ecfec 100644
>> --- a/fs/proc/base.c
>> +++ b/fs/proc/base.c
>> @@ -762,8 +762,9 @@ static int proc_single_open(struct inode *inode, struct file *filp)
>> .release = single_release,
>> };
>>
>> -
>> -struct mm_struct *proc_mem_open(struct inode *inode, unsigned int mode)
>> +struct mm_struct *proc_mem_open(struct inode *inode,
>> + unsigned int mode,
>> + fmode_t f_mode)
>> {
>> struct task_struct *task = get_proc_task(inode);
>> struct mm_struct *mm = ERR_PTR(-ESRCH);
>> @@ -773,10 +774,20 @@ struct mm_struct *proc_mem_open(struct inode *inode, unsigned int mode)
>> put_task_struct(task);
>>
>> if (!IS_ERR_OR_NULL(mm)) {
>> - /* ensure this mm_struct can't be freed */
>> - mmgrab(mm);
>> - /* but do not pin its memory */
>> - mmput(mm);
>> + /*
>> + * Prevent this interface from being used as a mean
>> + * to bypass memory restrictions, including those
>> + * imposed by LSMs.
>> + */
>> + if (mm == current->mm &&
>> + f_mode & FMODE_WRITE)
>> + mm = ERR_PTR(-EACCES);
>> + else {
>> + /* ensure this mm_struct can't be freed */
>> + mmgrab(mm);
>> + /* but do not pin its memory */
>> + mmput(mm);
>> + }
>> }
>> }
>
> I don't have an opinion on the overall patch, but this part looks
> buggy: In the error path, you set `mm` to an error pointer, but you
> still own the reference that mm_access() took on the old `mm`. The
> error path needs to call `mmput(mm)`.
You are absolutely right,
Thank you,
Salvatore
