Re: [PATCH v5 3.1.0-rc4-tip 3/26] Uprobes: register/unregisterprobes.

[Oleg Nesterov]

On 09/20, Srikar Dronamraju wrote:
>
> +static struct vma_info *__find_next_vma_info(struct list_head *head,
> +			loff_t offset, struct address_space *mapping,
> +			struct vma_info *vi)
> +{
> +	struct prio_tree_iter iter;
> +	struct vm_area_struct *vma;
> +	struct vma_info *tmpvi;
> +	loff_t vaddr;
> +	unsigned long pgoff = offset >> PAGE_SHIFT;
> +	int existing_vma;
> +
> +	vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
> +		if (!vma || !valid_vma(vma))
> +			return NULL;

!vma is not possible.

But I can't understand the !valid_vma(vma) check... We shouldn't return,
we should ignore this vma and continue, no? Otherwise, I can't see how
this can work if someone does, say, mmap(PROT_READ).

> +		existing_vma = 0;
> +		vaddr = vma->vm_start + offset;
> +		vaddr -= vma->vm_pgoff << PAGE_SHIFT;
> +		list_for_each_entry(tmpvi, head, probe_list) {
> +			if (tmpvi->mm == vma->vm_mm && tmpvi->vaddr == vaddr) {
> +				existing_vma = 1;
> +				break;
> +			}
> +		}
> +		if (!existing_vma &&
> +				atomic_inc_not_zero(&vma->vm_mm->mm_users)) {

This looks suspicious. If atomic_inc_not_zero() can fail, iow if we can
see ->mm_users == 0, then why it is safe to touch this counter/memory?
How we can know ->mm_count != 0 ?

I _think_ this is probably correct, ->mm_users == 0 means we are racing
mmput(), ->i_mmap_mutex and the fact we found this vma guarantees that
mmput() can't pass unlink_file_vma() and thus mmdrop() is not possible.
May be needs a comment...

> +static struct vma_info *find_next_vma_info(struct list_head *head,
> +			loff_t offset, struct address_space *mapping)
> +{
> +	struct vma_info *vi, *retvi;
> +	vi = kzalloc(sizeof(struct vma_info), GFP_KERNEL);
> +	if (!vi)
> +		return ERR_PTR(-ENOMEM);
> +
> +	INIT_LIST_HEAD(&vi->probe_list);

Looks unneeded.

> +	mutex_lock(&mapping->i_mmap_mutex);
> +	retvi = __find_next_vma_info(head, offset, mapping, vi);
> +	mutex_unlock(&mapping->i_mmap_mutex);

It is not clear why we can't race with mmap() after find_next_vma_info()
returns NULL. I guess this is solved by the next patches.

> +static int __register_uprobe(struct inode *inode, loff_t offset,
> +				struct uprobe *uprobe)
> +{
> +	struct list_head try_list;
> +	struct vm_area_struct *vma;
> +	struct address_space *mapping;
> +	struct vma_info *vi, *tmpvi;
> +	struct mm_struct *mm;
> +	int ret = 0;
> +
> +	mapping = inode->i_mapping;
> +	INIT_LIST_HEAD(&try_list);
> +	while ((vi = find_next_vma_info(&try_list, offset,
> +							mapping)) != NULL) {
> +		if (IS_ERR(vi)) {
> +			ret = -ENOMEM;
> +			break;
> +		}
> +		mm = vi->mm;
> +		down_read(&mm->mmap_sem);
> +		vma = find_vma(mm, (unsigned long) vi->vaddr);

But we can't trust find_vma? The original vma found by find_next_vma_info()
could go away, at least we should verify vi->vaddr >= vm_start.

And worse, I do not understand how we can trust ->vaddr. Can't we race with
sys_mremap() ?

> +static void __unregister_uprobe(struct inode *inode, loff_t offset,
> +						struct uprobe *uprobe)
> +{
> +	struct list_head try_list;
> +	struct address_space *mapping;
> +	struct vma_info *vi, *tmpvi;
> +	struct vm_area_struct *vma;
> +	struct mm_struct *mm;
> +
> +	mapping = inode->i_mapping;
> +	INIT_LIST_HEAD(&try_list);
> +	while ((vi = find_next_vma_info(&try_list, offset,
> +							mapping)) != NULL) {
> +		if (IS_ERR(vi))
> +			break;
> +		mm = vi->mm;
> +		down_read(&mm->mmap_sem);
> +		vma = find_vma(mm, (unsigned long) vi->vaddr);

Same problems...

> +		if (!vma || !valid_vma(vma)) {
> +			list_del(&vi->probe_list);
> +			kfree(vi);
> +			up_read(&mm->mmap_sem);
> +			mmput(mm);
> +			continue;
> +		}

Not sure about !valid_vma() (and note that __find_next_vma_info does() this
check too).

Suppose that register_uprobe() succeeds. After that unregister_ should work
even if user-space does mprotect() which can make valid_vma() == F, right?

> +int register_uprobe(struct inode *inode, loff_t offset,
> +				struct uprobe_consumer *consumer)
> +{
> +	struct uprobe *uprobe;
> +	int ret = 0;
> +
> +	inode = igrab(inode);
> +	if (!inode || !consumer || consumer->next)
> +		return -EINVAL;
> +
> +	if (offset > inode->i_size)
> +		return -EINVAL;

I guess this needs i_size_read().

And every "return" in register/unregister leaks something.

> +
> +	mutex_lock(&inode->i_mutex);
> +	uprobe = alloc_uprobe(inode, offset);

Looks like, alloc_uprobe() doesn't need ->i_mutex.

OTOH,

> +void unregister_uprobe(struct inode *inode, loff_t offset,
> +				struct uprobe_consumer *consumer)
> +{
> +	struct uprobe *uprobe;
> +
> +	inode = igrab(inode);
> +	if (!inode || !consumer)
> +		return;
> +
> +	if (offset > inode->i_size)
> +		return;
> +
> +	uprobe = find_uprobe(inode, offset);
> +	if (!uprobe)
> +		return;
> +
> +	if (!del_consumer(uprobe, consumer)) {
> +		put_uprobe(uprobe);
> +		return;
> +	}
> +
> +	mutex_lock(&inode->i_mutex);
> +	if (!uprobe->consumers)
> +		__unregister_uprobe(inode, offset, uprobe);

It seemes that del_consumer() should be done under ->i_mutex. If it
removes the last consumer, we can race with register_uprobe() which
takes ->i_mutex before us and does another __register_uprobe(), no?

Oleg.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/