Message-ID: <2adc84c4-2e92-db75-ba27-7e23eb9c95ad@iogearbox.net>
Date:   Thu, 9 Aug 2018 19:02:16 +0200
From:   Daniel Borkmann <daniel@...earbox.net>
To:     Yonghong Song <yhs@...com>,
        Alexei Starovoitov <alexei.starovoitov@...il.com>
Cc:     ast@...com, netdev@...r.kernel.org, kernel-team@...com,
        jakub.kicinski@...ronome.com
Subject: Re: [PATCH bpf] bpf: fix bpffs non-array map seq_show issue

On 08/09/2018 06:55 PM, Yonghong Song wrote:
> On 8/9/18 8:59 AM, Daniel Borkmann wrote:
>> On 08/09/2018 05:15 PM, Yonghong Song wrote:
>>> On 8/9/18 7:24 AM, Daniel Borkmann wrote:
>>>> On 08/09/2018 05:55 AM, Yonghong Song wrote:
>>>>> On 8/8/18 7:25 PM, Alexei Starovoitov wrote:
>>>>>> On Wed, Aug 08, 2018 at 06:25:19PM -0700, Yonghong Song wrote:
>>>>>>> In function map_seq_next() of kernel/bpf/inode.c,
>>>>>>> the first key will be the "0" regardless of the map type.
>>>>>>> This works for array. But for hash type, if it happens
>>>>>>> key "0" is in the map, the bpffs map show will miss
>>>>>>> some items if the key "0" is not the first element of
>>>>>>> the first bucket.
>>>>>>>
>>>>>>> This patch fixed the issue by guaranteeing to get
>>>>>>> the first element, if the seq_show is just started,
>>>>>>> by passing NULL pointer key to map_get_next_key() callback.
>>>>>>> This way, no missing elements will occur for
>>>>>>> bpffs hash table show even if key "0" is in the map.
>>>>>
>>>>> Currently, map_seq_show_elem callback is only implemented
>>>>> for arraymap. So the problem actually is not exposed.
>>>>>
>>>>> The issue is discovered when I tried to implement
>>>>> map_seq_show_elem for hash maps, and I will have followup
>>>>> patches for it.
>>
>> Btw, on that note, I would also prefer if we could decouple
>> BTF from the map_seq_show_elem() as there is really no reason
>> to have it on a per-map. I had a patch below which would enable
>> it for all map types generically, and bpftool works out of the
>> box for it. Also, array doesn't really have to be 'int' type
>> enforced as long as it's some data structure with 4 bytes, it's
>> all fine, so this can be made fully generic (we only eventually
>> care about the match in size).
> 
> I agree with a generic map_check_btf as mostly we only care about size
> and this change should enable btftool btf based pretty print for hash/lru_hash tables.

Yep, agree, the below output from bpftool is from test_xdp_noinline.o
where both work with it.

>>  From 0a8be27cbc2ac0c6fc2632865b5afe37222a1fc7 Mon Sep 17 00:00:00 2001
>> Message-Id: <0a8be27cbc2ac0c6fc2632865b5afe37222a1fc7.1533830053.git.daniel@...earbox.net>
>> From: Daniel Borkmann <daniel@...earbox.net>
>> Date: Thu, 9 Aug 2018 16:50:21 +0200
>> Subject: [PATCH bpf-next] bpf, btf: enable for all maps
>>
>> # bpftool m dump id 19
>> [{
>>          "key": {
>>              "": {
>>                  "vip": 0,
>>                  "vipv6": []
>>              },
>>              "port": 0,
>>              "family": 0,
>>              "proto": 0
>>          },
>>          "value": {
>>              "flags": 0,
>>              "vip_num": 0
>>          }
>>      }
>> ]
>>
>> Signed-off-by: Daniel Borkmann <daniel@...earbox.net>
>> ---
>>   include/linux/bpf.h   |  4 +---
>>   kernel/bpf/arraymap.c | 27 ---------------------------
>>   kernel/bpf/inode.c    |  3 ++-
>>   kernel/bpf/syscall.c  | 24 ++++++++++++++++++++----
>>   4 files changed, 23 insertions(+), 35 deletions(-)
>>
>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>> index cd8790d..91aa4be 100644
>> --- a/include/linux/bpf.h
>> +++ b/include/linux/bpf.h
>> @@ -48,8 +48,6 @@ struct bpf_map_ops {
>>       u32 (*map_fd_sys_lookup_elem)(void *ptr);
>>       void (*map_seq_show_elem)(struct bpf_map *map, void *key,
>>                     struct seq_file *m);
>> -    int (*map_check_btf)(const struct bpf_map *map, const struct btf *btf,
>> -                 u32 key_type_id, u32 value_type_id);
>>   };
>>
>>   struct bpf_map {
>> @@ -118,7 +116,7 @@ static inline bool bpf_map_offload_neutral(const struct bpf_map *map)
>>
>>   static inline bool bpf_map_support_seq_show(const struct bpf_map *map)
>>   {
>> -    return map->ops->map_seq_show_elem && map->ops->map_check_btf;
>> +    return map->ops->map_seq_show_elem;
>>   }
>>
>>   extern const struct bpf_map_ops bpf_map_offload_ops;
>> diff --git a/kernel/bpf/arraymap.c b/kernel/bpf/arraymap.c
>> index 2aa55d030..67f0bdf 100644
>> --- a/kernel/bpf/arraymap.c
>> +++ b/kernel/bpf/arraymap.c
>> @@ -358,32 +358,6 @@ static void array_map_seq_show_elem(struct bpf_map *map, void *key,
>>       rcu_read_unlock();
>>   }
>>
>> -static int array_map_check_btf(const struct bpf_map *map, const struct btf *btf,
>> -                   u32 btf_key_id, u32 btf_value_id)
>> -{
>> -    const struct btf_type *key_type, *value_type;
>> -    u32 key_size, value_size;
>> -    u32 int_data;
>> -
>> -    key_type = btf_type_id_size(btf, &btf_key_id, &key_size);
>> -    if (!key_type || BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
>> -        return -EINVAL;
>> -
>> -    int_data = *(u32 *)(key_type + 1);
>> -    /* bpf array can only take a u32 key.  This check makes
>> -     * sure that the btf matches the attr used during map_create.
>> -     */
>> -    if (BTF_INT_BITS(int_data) != 32 || key_size != 4 ||
>> -        BTF_INT_OFFSET(int_data))
>> -        return -EINVAL;
>> -
>> -    value_type = btf_type_id_size(btf, &btf_value_id, &value_size);
>> -    if (!value_type || value_size != map->value_size)
>> -        return -EINVAL;
>> -
>> -    return 0;
>> -}
>> -
>>   const struct bpf_map_ops array_map_ops = {
>>       .map_alloc_check = array_map_alloc_check,
>>       .map_alloc = array_map_alloc,
>> @@ -394,7 +368,6 @@ const struct bpf_map_ops array_map_ops = {
>>       .map_delete_elem = array_map_delete_elem,
>>       .map_gen_lookup = array_map_gen_lookup,
>>       .map_seq_show_elem = array_map_seq_show_elem,
>> -    .map_check_btf = array_map_check_btf,
>>   };
>>
>>   const struct bpf_map_ops percpu_array_map_ops = {
>> diff --git a/kernel/bpf/inode.c b/kernel/bpf/inode.c
>> index 76efe9a..400f27d 100644
>> --- a/kernel/bpf/inode.c
>> +++ b/kernel/bpf/inode.c
>> @@ -332,7 +332,8 @@ static int bpf_mkmap(struct dentry *dentry, umode_t mode, void *arg)
>>       struct bpf_map *map = arg;
>>
>>       return bpf_mkobj_ops(dentry, mode, arg, &bpf_map_iops,
>> -                 map->btf ? &bpffs_map_fops : &bpffs_obj_fops);
>> +                 bpf_map_support_seq_show(map) ?
>> +                 &bpffs_map_fops : &bpffs_obj_fops);
> 
> There are an issue here, the condition bpf_map_support_seq_show(map) may not be enough since the map specific implementation assumes availability of btf and proper map key/value btf_type_id's.
> We can either use
>     map->btf && bpf_map_support_seq_show(map)
> condition here, or check map->btf in each individual implementation
> of map_support_seq_show().

Good, point, agree. Will fix and cook proper patch later today.
Thanks Yonghong!

>>   }
>>
>>   static struct dentry *
>> diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
>> index 5af4e9e..0b6f6e8 100644
>> --- a/kernel/bpf/syscall.c
>> +++ b/kernel/bpf/syscall.c
>> @@ -455,6 +455,23 @@ static int bpf_obj_name_cpy(char *dst, const char *src)
>>       return 0;
>>   }
>>
>> +static int map_check_btf(const struct bpf_map *map, const struct btf *btf,
>> +             u32 btf_key_id, u32 btf_value_id)
>> +{
>> +    const struct btf_type *key_type, *value_type;
>> +    u32 key_size, value_size;
>> +
>> +    key_type = btf_type_id_size(btf, &btf_key_id, &key_size);
>> +    if (!key_type || key_size != map->key_size)
>> +        return -EINVAL;
>> +
>> +    value_type = btf_type_id_size(btf, &btf_value_id, &value_size);
>> +    if (!value_type || value_size != map->value_size)
>> +        return -EINVAL;
>> +
>> +    return 0;
>> +}
>> +
>>   #define BPF_MAP_CREATE_LAST_FIELD btf_value_type_id
>>   /* called via syscall */
>>   static int map_create(union bpf_attr *attr)
>> @@ -489,8 +506,7 @@ static int map_create(union bpf_attr *attr)
>>       atomic_set(&map->refcnt, 1);
>>       atomic_set(&map->usercnt, 1);
>>
>> -    if (bpf_map_support_seq_show(map) &&
>> -        (attr->btf_key_type_id || attr->btf_value_type_id)) {
>> +    if (attr->btf_key_type_id || attr->btf_value_type_id) {
>>           struct btf *btf;
>>
>>           if (!attr->btf_key_type_id || !attr->btf_value_type_id) {
>> @@ -504,8 +520,8 @@ static int map_create(union bpf_attr *attr)
>>               goto free_map_nouncharge;
>>           }
>>
>> -        err = map->ops->map_check_btf(map, btf, attr->btf_key_type_id,
>> -                          attr->btf_value_type_id);
>> +        err = map_check_btf(map, btf, attr->btf_key_type_id,
>> +                    attr->btf_value_type_id);
>>           if (err) {
>>               btf_put(btf);
>>               goto free_map_nouncharge;
>>

Powered by blists - more mailing lists