Re: [PATCH bpf-next v1 07/19] bpf: create anonymous bpf iterator

[Yonghong Song <yhs@fb.com>]

On 4/28/20 11:56 PM, Andrii Nakryiko wrote:
> On Mon, Apr 27, 2020 at 1:19 PM Yonghong Song <yhs@fb.com> wrote:
>>
>> A new bpf command BPF_ITER_CREATE is added.
>>
>> The anonymous bpf iterator is seq_file based.
>> The seq_file private data are referenced by targets.
>> The bpf_iter infrastructure allocated additional space
>> at seq_file->private after the space used by targets
>> to store some meta data, e.g.,
>>    prog:       prog to run
>>    session_id: an unique id for each opened seq_file
>>    seq_num:    how many times bpf programs are queried in this session
>>    has_last:   indicate whether or not bpf_prog has been called after
>>                all valid objects have been processed
>>
>> A map between file and prog/link is established to help
>> fops->release(). When fops->release() is called, just based on
>> inode and file, bpf program cannot be located since target
>> seq_priv_size not available. This map helps retrieve the prog
>> whose reference count needs to be decremented.
>>
>> Signed-off-by: Yonghong Song <yhs@fb.com>
>> ---
>>   include/linux/bpf.h            |   3 +
>>   include/uapi/linux/bpf.h       |   6 ++
>>   kernel/bpf/bpf_iter.c          | 162 ++++++++++++++++++++++++++++++++-
>>   kernel/bpf/syscall.c           |  27 ++++++
>>   tools/include/uapi/linux/bpf.h |   6 ++
>>   5 files changed, 203 insertions(+), 1 deletion(-)
>>
>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>> index 4fc39d9b5cd0..0f0cafc65a04 100644
>> --- a/include/linux/bpf.h
>> +++ b/include/linux/bpf.h
>> @@ -1112,6 +1112,8 @@ struct bpf_link *bpf_link_get_from_fd(u32 ufd);
>>   int bpf_obj_pin_user(u32 ufd, const char __user *pathname);
>>   int bpf_obj_get_user(const char __user *pathname, int flags);
>>
>> +#define BPF_DUMP_SEQ_NET_PRIVATE       BIT(0)
>> +
>>   struct bpf_iter_reg {
>>          const char *target;
>>          const char *target_func_name;
>> @@ -1133,6 +1135,7 @@ int bpf_iter_run_prog(struct bpf_prog *prog, void *ctx);
>>   int bpf_iter_link_attach(const union bpf_attr *attr, struct bpf_prog *prog);
>>   int bpf_iter_link_replace(struct bpf_link *link, struct bpf_prog *old_prog,
>>                            struct bpf_prog *new_prog);
>> +int bpf_iter_new_fd(struct bpf_link *link);
>>
>>   int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value);
>>   int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value);
>> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
>> index f39b9fec37ab..576651110d16 100644
>> --- a/include/uapi/linux/bpf.h
>> +++ b/include/uapi/linux/bpf.h
>> @@ -113,6 +113,7 @@ enum bpf_cmd {
>>          BPF_MAP_DELETE_BATCH,
>>          BPF_LINK_CREATE,
>>          BPF_LINK_UPDATE,
>> +       BPF_ITER_CREATE,
>>   };
>>
>>   enum bpf_map_type {
>> @@ -590,6 +591,11 @@ union bpf_attr {
>>                  __u32           old_prog_fd;
>>          } link_update;
>>
>> +       struct { /* struct used by BPF_ITER_CREATE command */
>> +               __u32           link_fd;
>> +               __u32           flags;
>> +       } iter_create;
>> +
>>   } __attribute__((aligned(8)));
>>
>>   /* The description below is an attempt at providing documentation to eBPF
>> diff --git a/kernel/bpf/bpf_iter.c b/kernel/bpf/bpf_iter.c
>> index fc1ce5ee5c3f..1f4e778d1814 100644
>> --- a/kernel/bpf/bpf_iter.c
>> +++ b/kernel/bpf/bpf_iter.c
>> @@ -2,6 +2,7 @@
>>   /* Copyright (c) 2020 Facebook */
>>
>>   #include <linux/fs.h>
>> +#include <linux/anon_inodes.h>
>>   #include <linux/filter.h>
>>   #include <linux/bpf.h>
>>
>> @@ -19,6 +20,19 @@ struct bpf_iter_link {
>>          struct bpf_iter_target_info *tinfo;
>>   };
>>
>> +struct extra_priv_data {
>> +       struct bpf_prog *prog;
>> +       u64 session_id;
>> +       u64 seq_num;
>> +       bool has_last;
>> +};
>> +
>> +struct anon_file_prog_assoc {
>> +       struct list_head list;
>> +       struct file *file;
>> +       struct bpf_prog *prog;
>> +};
>> +
>>   static struct list_head targets;
>>   static struct mutex targets_mutex;
>>   static bool bpf_iter_inited = false;
>> @@ -26,6 +40,50 @@ static bool bpf_iter_inited = false;
>>   /* protect bpf_iter_link.link->prog upddate */
>>   static struct mutex bpf_iter_mutex;
>>
>> +/* Since at anon seq_file release function, the prog cannot
>> + * be retrieved since target seq_priv_size is not available.
>> + * Keep a list of <anon_file, prog> mapping, so that
>> + * at file release stage, the prog can be released properly.
>> + */
>> +static struct list_head anon_iter_info;
>> +static struct mutex anon_iter_info_mutex;
>> +
>> +/* incremented on every opened seq_file */
>> +static atomic64_t session_id;
>> +
>> +static u32 get_total_priv_dsize(u32 old_size)
>> +{
>> +       return roundup(old_size, 8) + sizeof(struct extra_priv_data);
>> +}
>> +
>> +static void *get_extra_priv_dptr(void *old_ptr, u32 old_size)
>> +{
>> +       return old_ptr + roundup(old_size, 8);
>> +}
>> +
>> +static int anon_iter_release(struct inode *inode, struct file *file)
>> +{
>> +       struct anon_file_prog_assoc *finfo;
>> +
>> +       mutex_lock(&anon_iter_info_mutex);
>> +       list_for_each_entry(finfo, &anon_iter_info, list) {
>> +               if (finfo->file == file) {
> 
> I'll look at this and other patches more thoroughly tomorrow with
> clear head, but this iteration to find anon_file_prog_assoc is really
> unfortunate.
> 
> I think the problem is that you are allowing seq_file infrastructure
> to call directly into target implementation of seq_operations without
> intercepting them. If you change that and put whatever extra info is
> necessary into seq_file->private in front of target's private state,
> then you shouldn't need this, right?

Yes. This is true. The idea is to minimize the target change.
But maybe this is not a good goal by itself.

You are right, if I intercept all seq_ops(), I do not need the
above change, I can tailor seq_file private_data right before
calling target one and restore after the target call.

Originally I only have one interception, show(), now I have
stop() too to call bpf at the end of iteration. Maybe I can
interpret all four, I think. This way, I can also get ride
of target feature.

> 
> This would also make each target's logic a bit simpler because you can:
> - centralize creation and initialization of bpf_iter_meta (session_id,
> seq, seq_num will be set up once in this generic code);
> - loff_t pos increments;
> - you can extract and centralize bpf_iter_get_prog() call in show()
> implementation as well.
> 
> I think with that each target's logic will be simpler and you won't
> need to maintain anon_file_prog_assocs.
> 
> Are there complications I'm missing?
> 
> [...]
>