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

[Yonghong Song <yhs@fb.com>]

On 4/29/20 11:46 AM, Martin KaFai Lau wrote:
> On Wed, Apr 29, 2020 at 11:16:35AM -0700, Andrii Nakryiko wrote:
>> On Wed, Apr 29, 2020 at 12:07 AM Yonghong Song <yhs@fb.com> wrote:
>>>
>>>
>>>
>>> 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.
>>
>> If the main goal is to minimize target changes and make them exactly
>> seq_operations implementation, then one easier way to get easy access
>> to our own metadata in seq_file->private is to set it to point
>> **after** our metadata, but before target's metadata. Roughly in
>> pseudo code:
>>
>> struct bpf_iter_seq_file_meta {} __attribute((aligned(8)));
>>
>> void *meta = kmalloc(sizeof(struct bpf_iter_seq_file_meta) +
>> target_private_size);
>> seq_file->private = meta + sizeof(struct bpf_iter_seq_file_meta);
> I have suggested the same thing earlier.  Good to know that we think alike ;)
> 
> May be put them in a struct such that container_of...etc can be used:
> struct bpf_iter_private {
>          struct extra_priv_data iter_private;
> 	u8 target_private[] __aligned(8);
> };

This should work, but need to intercept all seq_ops() operations
because target expects private data is `target_private` only.
Let me experiment what is the best way to do this.

> 
>>
>>
>> Then to recover bpf_iter_Seq_file_meta:
>>
>> struct bpf_iter_seq_file_meta *meta = seq_file->private - sizeof(*meta);
>>
>> /* voila! */
>>
>> This doesn't have a benefit of making targets simpler, but will
>> require no changes to them at all. Plus less indirect calls, so less
>> performance penalty.
>>