在 2022/5/10 上午9:04, Yosry Ahmed 写道:
> On Mon, May 9, 2022 at 5:34 PM Andrii Nakryiko
> <[email protected]> wrote:
>> On Fri, May 6, 2022 at 7:49 PM Feng zhou <[email protected]> wrote:
>>> From: Feng Zhou <[email protected]>
>>>
>>> Trace some functions, such as enqueue_task_fair, need to access the
>>> corresponding cpu, not the current cpu, and bpf_map_lookup_elem percpu map
>>> cannot do it. So add bpf_map_lookup_percpu_elem to accomplish it for
>>> percpu_array_map, percpu_hash_map, lru_percpu_hash_map.
>>>
>>> The implementation method is relatively simple, refer to the implementation
>>> method of map_lookup_elem of percpu map, increase the parameters of cpu, and
>>> obtain it according to the specified cpu.
>>>
>> I don't think it's safe in general to access per-cpu data from another
>> CPU. I'd suggest just having either a ARRAY_OF_MAPS or adding CPU ID
>> as part of the key, if you need such a custom access pattern.
> I actually just sent an RFC patch series containing a similar patch
> for the exact same purpose. There are instances in the kernel where
> per-cpu data is accessed from other cpus (e.g.
> mem_cgroup_css_rstat_flush()). I believe, like any other variable,
> percpu data can be safe or not safe to access, based on the access
> pattern. It is up to the user to coordinate accesses to the variable.
>
> For example, in my use case, one of the accessors only reads percpu
> values of different cpus, so it should be safe. If a user accesses
> percpu data of another cpu without guaranteeing safety, they corrupt
> their own data. I understand that the main purpose of percpu data is
> lockless (and therefore fast) access, but in some use cases the user
> may be able to safely (and locklessly) access the data concurrently.
>
Regarding data security, I think users need to consider before using it,
such
as hook enqueue_task_fair, the function itself takes the rq lock of the
corresponding cpu, there is no problem, and the kernel only provides a
method,
like bpf_per_cpu_ptr and bpf_this_cpu_ptr, data security needs to be
guaranteed
by users in different scenarios, such as using bpf_spin_lock.
>>> Signed-off-by: Feng Zhou <[email protected]>
>>> ---
>>> include/linux/bpf.h | 2 ++
>>> include/uapi/linux/bpf.h | 9 +++++++++
>>> kernel/bpf/arraymap.c | 15 +++++++++++++++
>>> kernel/bpf/core.c | 1 +
>>> kernel/bpf/hashtab.c | 32 ++++++++++++++++++++++++++++++++
>>> kernel/bpf/helpers.c | 18 ++++++++++++++++++
>>> kernel/bpf/verifier.c | 17 +++++++++++++++--
>>> kernel/trace/bpf_trace.c | 2 ++
>>> tools/include/uapi/linux/bpf.h | 9 +++++++++
>>> 9 files changed, 103 insertions(+), 2 deletions(-)
>>>
>> [...]
On Mon, May 9, 2022 at 7:41 PM Feng Zhou <[email protected]> wrote:
>
> 在 2022/5/10 上午9:04, Yosry Ahmed 写道:
> > On Mon, May 9, 2022 at 5:34 PM Andrii Nakryiko
> > <[email protected]> wrote:
> >> On Fri, May 6, 2022 at 7:49 PM Feng zhou <[email protected]> wrote:
> >>> From: Feng Zhou <[email protected]>
> >>>
> >>> Trace some functions, such as enqueue_task_fair, need to access the
> >>> corresponding cpu, not the current cpu, and bpf_map_lookup_elem percpu map
> >>> cannot do it. So add bpf_map_lookup_percpu_elem to accomplish it for
> >>> percpu_array_map, percpu_hash_map, lru_percpu_hash_map.
> >>>
> >>> The implementation method is relatively simple, refer to the implementation
> >>> method of map_lookup_elem of percpu map, increase the parameters of cpu, and
> >>> obtain it according to the specified cpu.
> >>>
> >> I don't think it's safe in general to access per-cpu data from another
> >> CPU. I'd suggest just having either a ARRAY_OF_MAPS or adding CPU ID
> >> as part of the key, if you need such a custom access pattern.
> > I actually just sent an RFC patch series containing a similar patch
> > for the exact same purpose. There are instances in the kernel where
> > per-cpu data is accessed from other cpus (e.g.
> > mem_cgroup_css_rstat_flush()). I believe, like any other variable,
> > percpu data can be safe or not safe to access, based on the access
> > pattern. It is up to the user to coordinate accesses to the variable.
> >
> > For example, in my use case, one of the accessors only reads percpu
> > values of different cpus, so it should be safe. If a user accesses
> > percpu data of another cpu without guaranteeing safety, they corrupt
> > their own data. I understand that the main purpose of percpu data is
> > lockless (and therefore fast) access, but in some use cases the user
> > may be able to safely (and locklessly) access the data concurrently.
> >
>
> Regarding data security, I think users need to consider before using it,
> such
> as hook enqueue_task_fair, the function itself takes the rq lock of the
> corresponding cpu, there is no problem, and the kernel only provides a
> method,
> like bpf_per_cpu_ptr and bpf_this_cpu_ptr, data security needs to be
> guaranteed
> by users in different scenarios, such as using bpf_spin_lock.
Right. The new helper looks useful and is safe.
Please add a selftest and respin.
在 2022/5/10 上午11:15, Alexei Starovoitov 写道:
> On Mon, May 9, 2022 at 7:41 PM Feng Zhou <[email protected]> wrote:
>> 在 2022/5/10 上午9:04, Yosry Ahmed 写道:
>>> On Mon, May 9, 2022 at 5:34 PM Andrii Nakryiko
>>> <[email protected]> wrote:
>>>> On Fri, May 6, 2022 at 7:49 PM Feng zhou <[email protected]> wrote:
>>>>> From: Feng Zhou <[email protected]>
>>>>>
>>>>> Trace some functions, such as enqueue_task_fair, need to access the
>>>>> corresponding cpu, not the current cpu, and bpf_map_lookup_elem percpu map
>>>>> cannot do it. So add bpf_map_lookup_percpu_elem to accomplish it for
>>>>> percpu_array_map, percpu_hash_map, lru_percpu_hash_map.
>>>>>
>>>>> The implementation method is relatively simple, refer to the implementation
>>>>> method of map_lookup_elem of percpu map, increase the parameters of cpu, and
>>>>> obtain it according to the specified cpu.
>>>>>
>>>> I don't think it's safe in general to access per-cpu data from another
>>>> CPU. I'd suggest just having either a ARRAY_OF_MAPS or adding CPU ID
>>>> as part of the key, if you need such a custom access pattern.
>>> I actually just sent an RFC patch series containing a similar patch
>>> for the exact same purpose. There are instances in the kernel where
>>> per-cpu data is accessed from other cpus (e.g.
>>> mem_cgroup_css_rstat_flush()). I believe, like any other variable,
>>> percpu data can be safe or not safe to access, based on the access
>>> pattern. It is up to the user to coordinate accesses to the variable.
>>>
>>> For example, in my use case, one of the accessors only reads percpu
>>> values of different cpus, so it should be safe. If a user accesses
>>> percpu data of another cpu without guaranteeing safety, they corrupt
>>> their own data. I understand that the main purpose of percpu data is
>>> lockless (and therefore fast) access, but in some use cases the user
>>> may be able to safely (and locklessly) access the data concurrently.
>>>
>> Regarding data security, I think users need to consider before using it,
>> such
>> as hook enqueue_task_fair, the function itself takes the rq lock of the
>> corresponding cpu, there is no problem, and the kernel only provides a
>> method,
>> like bpf_per_cpu_ptr and bpf_this_cpu_ptr, data security needs to be
>> guaranteed
>> by users in different scenarios, such as using bpf_spin_lock.
> Right. The new helper looks useful and is safe.
> Please add a selftest and respin.
Ok, will do. Thanks.