Re: [PATCH v2 bpf-next 1/4] bpf: introduce task_vma bpf_iter

[Song Liu <songliubraving@fb.com>]

> On Dec 17, 2020, at 6:34 PM, Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:
> 
> On Thu, Dec 17, 2020 at 10:08:31PM +0000, Song Liu wrote:
>> 
>> 
>>> On Dec 17, 2020, at 11:03 AM, Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:
>>> 
>>> On Tue, Dec 15, 2020 at 03:36:59PM -0800, Song Liu wrote:
>>>> +/*
>>>> + * Key information from vm_area_struct. We need this because we cannot
>>>> + * assume the vm_area_struct is still valid after each iteration.
>>>> + */
>>>> +struct __vm_area_struct {
>>>> +	__u64 start;
>>>> +	__u64 end;
>>>> +	__u64 flags;
>>>> +	__u64 pgoff;
>>>> +};
>>> 
>>> Where it's inside .c or exposed in uapi/bpf.h it will become uapi
>>> if it's used this way. Let's switch to arbitrary BTF-based access instead.
>>> 
>>>> +static struct __vm_area_struct *
>>>> +task_vma_seq_get_next(struct bpf_iter_seq_task_vma_info *info)
>>>> +{
>>>> +	struct pid_namespace *ns = info->common.ns;
>>>> +	struct task_struct *curr_task;
>>>> +	struct vm_area_struct *vma;
>>>> +	u32 curr_tid = info->tid;
>>>> +	bool new_task = false;
>>>> +
>>>> +	/* If this function returns a non-NULL __vm_area_struct, it held
>>>> +	 * a reference to the task_struct. If info->file is non-NULL, it
>>>> +	 * also holds a reference to the file. If this function returns
>>>> +	 * NULL, it does not hold any reference.
>>>> +	 */
>>>> +again:
>>>> +	if (info->task) {
>>>> +		curr_task = info->task;
>>>> +	} else {
>>>> +		curr_task = task_seq_get_next(ns, &curr_tid, true);
>>>> +		if (!curr_task) {
>>>> +			info->task = NULL;
>>>> +			info->tid++;
>>>> +			return NULL;
>>>> +		}
>>>> +
>>>> +		if (curr_tid != info->tid) {
>>>> +			info->tid = curr_tid;
>>>> +			new_task = true;
>>>> +		}
>>>> +
>>>> +		if (!curr_task->mm)
>>>> +			goto next_task;
>>>> +		info->task = curr_task;
>>>> +	}
>>>> +
>>>> +	mmap_read_lock(curr_task->mm);
>>> 
>>> That will hurt. /proc readers do that and it causes all sorts
>>> of production issues. We cannot take this lock.
>>> There is no need to take it.
>>> Switch the whole thing to probe_read style walking.
>>> And reimplement find_vma with probe_read while omitting vmacache.
>>> It will be short rbtree walk.
>>> bpf prog doesn't need to see a stable struct. It will read it through ptr_to_btf_id
>>> which will use probe_read equivalent underneath.
>> 
>> rw_semaphore is designed to avoid write starvation, so read_lock should not cause
>> problem unless the lock was taken for extended period. [1] was a recent fix that 
>> avoids /proc issue by releasing mmap_lock between iterations. We are using similar
>> mechanism here. BTW: I changed this to mmap_read_lock_killable() in the next version. 
>> 
>> On the other hand, we need a valid vm_file pointer for bpf_d_path. So walking the 
> 
> ahh. I missed that. Makes sense.
> vm_file needs to be accurate, but vm_area_struct should be accessed as ptr_to_btf_id.

Passing pointer of vm_area_struct into BPF will be tricky. For example, shall we 
allow the user to access vma->vm_file? IIUC, with ptr_to_btf_id the verifier will
allow access of vma->vm_file as a valid pointer to struct file. However, since the
vma might be freed, vma->vm_file could point to random data. 

> 
>> rbtree without taking any lock would not work. We can avoid taking the lock when 
>> some SPF like mechanism merged (hopefully soonish). 
>> 
>> Did I miss anything? 
>> 
>> We can improve bpf_iter with some mechanism to specify which task to iterate, so 
>> that we don't have to iterate through all tasks when the user only want to inspect 
>> vmas in one task. 
> 
> yes. let's figure out how to make it parametrizable.
> map_iter runs only for given map_fd.
> Maybe vma_iter should run only for given pidfd?
> I think all_task_all_vmas iter is nice to have, but we don't really need it?
> 
>> Thanks,
>> Song
>> 
>> [1] ff9f47f6f00c ("mm: proc: smaps_rollup: do not stall write attempts on mmap_lock")
> 
> Thanks for this link. With "if (mmap_lock_is_contended())" check it should work indeed.

To make sure we are on the same page: I am using slightly different mechanism in 
task_vma_iter, which doesn't require checking mmap_lock_is_contended(). In the 
smaps_rollup case, the code only unlock mmap_sem when the lock is contended. In 
task_iter, we always unlock mmap_sem between two iterations. This is because we 
don't want to hold mmap_sem while calling the BPF program, which may sleep (calling
bpf_d_path). 

Thanks,
Song