Re: [PATCH v4 bpf-next 1/2] bpf: introduce helper bpf_find_vma

On Nov 5, 2021, at 2:16 PM, Alexei Starovoitov [off-list ref] wrote:

On 11/5/21 2:11 PM, Song Liu wrote:

quoted

quoted

On Nov 5, 2021, at 8:23 AM, Hengqi Chen [off-list ref] wrote:

Hi, Song

On 2021/11/5 5:31 AM, Song Liu wrote:

quoted

In some profiler use cases, it is necessary to map an address to the
backing file, e.g., a shared library. bpf_find_vma helper provides a
flexible way to achieve this. bpf_find_vma maps an address of a task to
the vma (vm_area_struct) for this address, and feed the vma to an callback
BPF function. The callback function is necessary here, as we need to
ensure mmap_sem is unlocked.

It is necessary to lock mmap_sem for find_vma. To lock and unlock mmap_sem
safely when irqs are disable, we use the same mechanism as stackmap with
build_id. Specifically, when irqs are disabled, the unlocked is postponed
in an irq_work. Refactor stackmap.c so that the irq_work is shared among
bpf_find_vma and stackmap helpers.

Reported-by: kernel test robot <redacted>
Signed-off-by: Song Liu <redacted>
---

[...]

quoted

-BTF_ID_LIST(btf_task_file_ids)
-BTF_ID(struct, file)
-BTF_ID(struct, vm_area_struct)
-
static const struct bpf_iter_seq_info task_seq_info = {
	.seq_ops		= &task_seq_ops,
	.init_seq_private	= init_seq_pidns,

@@ -586,9 +583,74 @@ static struct bpf_iter_reg task_vma_reg_info = {

	.seq_info		= &task_vma_seq_info,
};

+BPF_CALL_5(bpf_find_vma, struct task_struct *, task, u64, start,
+	   bpf_callback_t, callback_fn, void *, callback_ctx, u64, flags)
+{
+	struct mmap_unlock_irq_work *work = NULL;
+	struct vm_area_struct *vma;
+	bool irq_work_busy = false;
+	struct mm_struct *mm;
+	int ret = -ENOENT;
+
+	if (flags)
+		return -EINVAL;
+
+	if (!task)
+		return -ENOENT;
+
+	mm = task->mm;
+	if (!mm)
+		return -ENOENT;
+
+	irq_work_busy = bpf_mmap_unlock_get_irq_work(&work);
+
+	if (irq_work_busy || !mmap_read_trylock(mm))
+		return -EBUSY;
+
+	vma = find_vma(mm, start);
+

I found that when a BPF program attach to security_file_open which is in
the bpf_d_path helper's allowlist, the bpf_d_path helper is also allowed
to be called inside the callback function. So we can have this in callback
function:

   bpf_d_path(&vma->vm_file->f_path, path, sizeof(path));


I wonder whether there is a guarantee that vma->vm_file will never be null,
as you said in the commit message, a backing file.

I don't think we can guarantee vma->vm_file never be NULL here, so this is
a real problem. Let me see how to fix it.

It's unrelated. There was a separate thread about this.

I see. I will not fix it here then. 

Thanks,
Song