Re: [PATCH v3 bpf-next 02/11] bpf: add generic support for lookup and lookup_and_delete batch ops

[Yonghong Song <yhs@fb.com>]

On 1/6/20 10:50 PM, Brian Vazquez wrote:
> On Fri, Dec 13, 2019 at 11:26 AM Yonghong Song <yhs@fb.com> wrote:
>>
>>
>>
>> On 12/11/19 2:33 PM, Brian Vazquez wrote:
>>> This commit introduces generic support for the bpf_map_lookup_batch and
>>> bpf_map_lookup_and_delete_batch ops. This implementation can be used by
>>> almost all the bpf maps since its core implementation is relying on the
>>> existing map_get_next_key, map_lookup_elem and map_delete_elem
>>> functions. The bpf syscall subcommands introduced are:
>>>
>>>     BPF_MAP_LOOKUP_BATCH
>>>     BPF_MAP_LOOKUP_AND_DELETE_BATCH
>>>
>>> The UAPI attribute is:
>>>
>>>     struct { /* struct used by BPF_MAP_*_BATCH commands */
>>>            __aligned_u64   in_batch;       /* start batch,
>>>                                             * NULL to start from beginning
>>>                                             */
>>>            __aligned_u64   out_batch;      /* output: next start batch */
>>>            __aligned_u64   keys;
>>>            __aligned_u64   values;
>>>            __u32           count;          /* input/output:
>>>                                             * input: # of key/value
>>>                                             * elements
>>>                                             * output: # of filled elements
>>>                                             */
>>>            __u32           map_fd;
>>>            __u64           elem_flags;
>>>            __u64           flags;
>>>     } batch;
>>>
>>> in_batch/out_batch are opaque values use to communicate between
>>> user/kernel space, in_batch/out_batch must be of key_size length.
>>>
>>> To start iterating from the beginning in_batch must be null,
>>> count is the # of key/value elements to retrieve. Note that the 'keys'
>>> buffer must be a buffer of key_size * count size and the 'values' buffer
>>> must be value_size * count, where value_size must be aligned to 8 bytes
>>> by userspace if it's dealing with percpu maps. 'count' will contain the
>>> number of keys/values successfully retrieved. Note that 'count' is an
>>> input/output variable and it can contain a lower value after a call.
>>>
>>> If there's no more entries to retrieve, ENOENT will be returned. If error
>>> is ENOENT, count might be > 0 in case it copied some values but there were
>>> no more entries to retrieve.
>>>
>>> Note that if the return code is an error and not -EFAULT,
>>> count indicates the number of elements successfully processed.
>>>
>>> Suggested-by: Stanislav Fomichev <sdf@google.com>
>>> Signed-off-by: Brian Vazquez <brianvv@google.com>
>>> Signed-off-by: Yonghong Song <yhs@fb.com>
>>> ---
>>>    include/linux/bpf.h      |  11 +++
>>>    include/uapi/linux/bpf.h |  19 +++++
>>>    kernel/bpf/syscall.c     | 172 +++++++++++++++++++++++++++++++++++++++
>>>    3 files changed, 202 insertions(+)
>> [...]
>>> diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
>>> index 2530266fa6477..708aa89fe2308 100644
>>> --- a/kernel/bpf/syscall.c
>>> +++ b/kernel/bpf/syscall.c
>>> @@ -1206,6 +1206,120 @@ static int map_get_next_key(union bpf_attr *attr)
>>>        return err;
>>>    }
>>>
>>> +#define MAP_LOOKUP_RETRIES 3
>>> +
>>> +static int __generic_map_lookup_batch(struct bpf_map *map,
>>> +                                   const union bpf_attr *attr,
>>> +                                   union bpf_attr __user *uattr,
>>> +                                   bool do_delete)
>>> +{
>>> +     void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
>>> +     void __user *uobatch = u64_to_user_ptr(attr->batch.out_batch);
>>> +     void __user *values = u64_to_user_ptr(attr->batch.values);
>>> +     void __user *keys = u64_to_user_ptr(attr->batch.keys);
>>> +     void *buf, *prev_key, *key, *value;
>>> +     u32 value_size, cp, max_count;
>>> +     bool first_key = false;
>>> +     int err, retry = MAP_LOOKUP_RETRIES;
>>
>> Could you try to use reverse Christmas tree style declaration here?
> 
> ACK
>>
>>> +
>>> +     if (attr->batch.elem_flags & ~BPF_F_LOCK)
>>> +             return -EINVAL;
>>> +
>>> +     if ((attr->batch.elem_flags & BPF_F_LOCK) &&
>>> +         !map_value_has_spin_lock(map))
>>> +             return -EINVAL;
>>> +
>>> +     value_size = bpf_map_value_size(map);
>>> +
>>> +     max_count = attr->batch.count;
>>> +     if (!max_count)
>>> +             return 0;
>>> +
>>> +     buf = kmalloc(map->key_size + value_size, GFP_USER | __GFP_NOWARN);
>>> +     if (!buf)
>>> +             return -ENOMEM;
>>> +
>>> +     err = -EFAULT;
>>> +     first_key = false;
>>> +     if (ubatch && copy_from_user(buf, ubatch, map->key_size))
>>> +             goto free_buf;
>>> +     key = buf;
>>> +     value = key + map->key_size;
>>> +     if (!ubatch) {
>>> +             prev_key = NULL;
>>> +             first_key = true;
>>> +     }
>>> +
>>> +     for (cp = 0; cp < max_count;) {
>>> +             if (cp || first_key) {
>>> +                     rcu_read_lock();
>>> +                     err = map->ops->map_get_next_key(map, prev_key, key);
>>> +                     rcu_read_unlock();
>>> +                     if (err)
>>> +                             break;
>>> +             }
>>> +             err = bpf_map_copy_value(map, key, value,
>>> +                                      attr->batch.elem_flags, do_delete);
>>> +
>>> +             if (err == -ENOENT) {
>>> +                     if (retry) {
>>> +                             retry--;
>>> +                             continue;
>>> +                     }
>>> +                     err = -EINTR;
>>> +                     break;
>>> +             }
>>> +
>>> +             if (err)
>>> +                     goto free_buf;
>>> +
>>> +             if (copy_to_user(keys + cp * map->key_size, key,
>>> +                              map->key_size)) {
>>> +                     err = -EFAULT;
>>> +                     goto free_buf;
>>> +             }
>>> +             if (copy_to_user(values + cp * value_size, value, value_size)) {
>>> +                     err = -EFAULT;
>>> +                     goto free_buf;
>>> +             }
>>> +
>>> +             prev_key = key;
>>> +             retry = MAP_LOOKUP_RETRIES;
>>> +             cp++;
>>> +     }
>>> +
>>> +     if (!err) {
>>> +             rcu_read_lock();
>>> +             err = map->ops->map_get_next_key(map, prev_key, key);
>>> +             rcu_read_unlock();
>>> +     }
>>> +
>>> +     if (err)
>>> +             memset(key, 0, map->key_size);
>>
>> So if any error happens due to above map_get_next_key() or earlier
>> error, the next "batch" returned to user could be "0". What should
>> user space handle this? Ultimately, the user space needs to start
>> from the beginning again?
>>
>> What I mean is here how we could design an interface so user
>> space, if no -EFAULT error, can successfully get all elements
>> without duplication.
>>
>> One way to do here is just return -EFAULT if we cannot get
>> proper next key. But maybe we could have better mechanism
>> when we try to implement what user space codes will look like.
> 
> I was thinking that instead of using the "next key" as a token we
> could use the last value successfully copied as the token, that way
> user space code would always be able to start/retry from the last
> processed entry. Do you think this would work?

Yes, this should work.


>>
>>> +
>>> +     if ((copy_to_user(&uattr->batch.count, &cp, sizeof(cp)) ||
>>> +                 (copy_to_user(uobatch, key, map->key_size))))
>>> +             err = -EFAULT;
>>> +
>>> +free_buf:
>>> +     kfree(buf);
>>> +     return err;
>>> +}
>>> +
>> [...]