Re: [PATCH v5 bpf-next 1/4] bpf: introduce task_vma bpf_iter

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




> On Feb 9, 2021, at 7:00 PM, Alexei Starovoitov <ast@xxxxxx> wrote:
> 
> On 2/9/21 2:08 PM, Song Liu wrote:
>>> On Feb 9, 2021, at 1:30 PM, Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote:
>>> 
>>> On Mon, Feb 08, 2021 at 02:52:52PM -0800, Song Liu wrote:
>>>> Introduce task_vma bpf_iter to print memory information of a process. It
>>>> can be used to print customized information similar to /proc/<pid>/maps.
>>>> 
>>>> Current /proc/<pid>/maps and /proc/<pid>/smaps provide information of
>>>> vma's of a process. However, these information are not flexible enough to
>>>> cover all use cases. For example, if a vma cover mixed 2MB pages and 4kB
>>>> pages (x86_64), there is no easy way to tell which address ranges are
>>>> backed by 2MB pages. task_vma solves the problem by enabling the user to
>>>> generate customize information based on the vma (and vma->vm_mm,
>>>> vma->vm_file, etc.).
>>>> 
>>>> To access the vma safely in the BPF program, task_vma iterator holds
>>>> target mmap_lock while calling the BPF program. If the mmap_lock is
>>>> contended, task_vma unlocks mmap_lock between iterations to unblock the
>>>> writer(s). This lock contention avoidance mechanism is similar to the one
>>>> used in show_smaps_rollup().
>>>> 
>>>> Signed-off-by: Song Liu <songliubraving@xxxxxx>
>>>> ---
>>>> kernel/bpf/task_iter.c | 217 ++++++++++++++++++++++++++++++++++++++++-
>>>> 1 file changed, 216 insertions(+), 1 deletion(-)
>>>> 
>>>> diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c
>>>> index 175b7b42bfc46..a0d469f0f481c 100644
>>>> --- a/kernel/bpf/task_iter.c
>>>> +++ b/kernel/bpf/task_iter.c
>>>> @@ -286,9 +286,198 @@ static const struct seq_operations task_file_seq_ops = {
>>>> 	.show	= task_file_seq_show,
>>>> };
>>>> 
>>>> +struct bpf_iter_seq_task_vma_info {
>>>> +	/* The first field must be struct bpf_iter_seq_task_common.
>>>> +	 * this is assumed by {init, fini}_seq_pidns() callback functions.
>>>> +	 */
>>>> +	struct bpf_iter_seq_task_common common;
>>>> +	struct task_struct *task;
>>>> +	struct vm_area_struct *vma;
>>>> +	u32 tid;
>>>> +	unsigned long prev_vm_start;
>>>> +	unsigned long prev_vm_end;
>>>> +};
>>>> +
>>>> +enum bpf_task_vma_iter_find_op {
>>>> +	task_vma_iter_first_vma,   /* use mm->mmap */
>>>> +	task_vma_iter_next_vma,    /* use curr_vma->vm_next */
>>>> +	task_vma_iter_find_vma,    /* use find_vma() to find next vma */
>>>> +};
>>>> +
>>>> +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;
>>>> +	enum bpf_task_vma_iter_find_op op;
>>>> +	struct vm_area_struct *curr_vma;
>>>> +	struct task_struct *curr_task;
>>>> +	u32 curr_tid = info->tid;
>>>> +
>>>> +	/* If this function returns a non-NULL vma, it holds a reference to
>>>> +	 * the task_struct, and holds read lock on vma->mm->mmap_lock.
>>>> +	 * If this function returns NULL, it does not hold any reference or
>>>> +	 * lock.
>>>> +	 */
>>>> +	if (info->task) {
>>>> +		curr_task = info->task;
>>>> +		curr_vma = info->vma;
>>>> +		/* In case of lock contention, drop mmap_lock to unblock
>>>> +		 * the writer.
>>>> +		 */
>>>> +		if (mmap_lock_is_contended(curr_task->mm)) {
>>>> +			info->prev_vm_start = curr_vma->vm_start;
>>>> +			info->prev_vm_end = curr_vma->vm_end;
>>>> +			op = task_vma_iter_find_vma;
>>>> +			mmap_read_unlock(curr_task->mm);
>>>> +			if (mmap_read_lock_killable(curr_task->mm))
>>>> +				goto finish;
>>> 
>>> in case of contention the vma will be seen by bpf prog again?
>>> It looks like the 4 cases of overlaping vmas (after newly acquired lock)
>>> that show_smaps_rollup() is dealing with are not handled here?
>> I am not sure I am following here. The logic below should avoid showing
>> the same vma again:
>>  	curr_vma = find_vma(curr_task->mm, info->prev_vm_end - 1);
>> 	if (curr_vma && (curr_vma->vm_start == info->prev_vm_start))
>> 		curr_vma = curr_vma->vm_next;
>> This logic handles case 1, 2, 3 same as show_smaps_rollup(). For case 4,
>> this logic skips the changed vma (from [prev_vm_start, prev_vm_end] to
>> [prev_vm_start, prev_vm_end + something]); while show_smaps_rollup() will
>> process the new vma.  I think skipping or processing the new vma are both
>> correct, as we already processed part of it [prev_vm_start, prev_vm_end]
>> once.
> 
> Got it. Yeah, if there is a new vma that has extra range after
> prem_vm_end while prev_vm_start(s) are the same, arguably,
> bpf prog shouldn't process the same range again,
> but it will miss the upper part of the range.
> In other words there is no equivalent here to 'start'
> argument that smap_gather_stats has.
> It's fine, but lets document this subtle difference.

Make sense. I will add information in the comment.  

> 
>>> 
>>>> +		} else {
>>>> +			op = task_vma_iter_next_vma;
>>>> +		}
>>>> +	} else {
>>>> +again:
>>>> +		curr_task = task_seq_get_next(ns, &curr_tid, true);
>>>> +		if (!curr_task) {
>>>> +			info->tid = curr_tid + 1;
>>>> +			goto finish;
>>>> +		}
>>>> +
>>>> +		if (curr_tid != info->tid) {
>>>> +			info->tid = curr_tid;
>>>> +			op = task_vma_iter_first_vma;
>>>> +		} else {
>>>> +			op = task_vma_iter_find_vma;
>>> 
>>> what will happen if there was no contetion on the lock and no seq_stop
>>> when this line was hit and set op = find_vma; ?
>>> If I'm reading this correctly prev_vm_start/end could still
>>> belong to some previous task.
>> In that case, we should be in "curr_tid != info->tid" path, no?
>>> My understanding that if read buffer is big the bpf_seq_read()
>>> will keep doing while(space in buffer) {seq->op->show(), seq->op->next();}
>>> and task_vma_seq_get_next() will iterate over all vmas of one task and
>>> will proceed into the next task, but if there was no contention and no stop
>>> then prev_vm_end will either be still zero (so find_vma(mm, 0 - 1) will be lucky
>>> and will go into first vma of the new task) or perf_vm_end is some address
>>> of some previous task's vma. In this case find_vma may return wrong vma
>>> for the new task.
>>> It seems to me prev_vm_end/start should be set by this task_vma_seq_get_next()
>>> function instead of relying on stop callback.
> 
> Yeah. I misread where the 'op' goes.
> But I think the problem still exists. Consider the loop of
> show;next;show;next;...
> Here it will be: case first_vma; case next_vma; case next_vma;
> Now it goes into new task and 'curr_tid != info->tid',
> so it does op = first_vma and info->tid = curr_tid.
> But we got unlucky and the process got suspended (with ctrl-z)
> and mmap_read_lock_killable returned eintr.
> The 'if' below will jump to finish.
> It will set info->task = NULL
> The process suppose to continue sys_read after resume.
> It will come back here to 'again:', but now it will do 'case find_vma'
> and will search for wrong prev_vm_end.

You are right. We will hit the issue in this case. Let me fix it in 
the next version. 

> 
> Maybe I'm missing something again.
> It's hard for me to follow the code.
> Could you please add diagrams like show_smaps_rollup() does and
> document what happens with all the 'op's.

Will also add more documents here. 

Thanks,
Song




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux