Date:   Wed, 29 Apr 2020 13:50:51 -0700
From:   Martin KaFai Lau <kafai@...com>
To:     Yonghong Song <yhs@...com>
CC:     Andrii Nakryiko <andrii.nakryiko@...il.com>,
        Andrii Nakryiko <andriin@...com>, bpf <bpf@...r.kernel.org>,
        Networking <netdev@...r.kernel.org>,
        Alexei Starovoitov <ast@...com>,
        Daniel Borkmann <daniel@...earbox.net>,
        Kernel Team <kernel-team@...com>
Subject: Re: [PATCH bpf-next v1 07/19] bpf: create anonymous bpf iterator

On Wed, Apr 29, 2020 at 12:20:05PM -0700, Yonghong Song wrote:
> 
> 
> 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@...com> wrote:
> > > > 
> > > > 
> > > > 
> > > > On 4/28/20 11:56 PM, Andrii Nakryiko wrote:
> > > > > On Mon, Apr 27, 2020 at 1:19 PM Yonghong Song <yhs@...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@...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.
As long as "seq_file->private = bpf_iter_private->target_private;" as
Andrii also suggested, the existing seq_ops() should not have to be
changed or needed to be intercepted because they are only
accessing it through seq_file->private.

The bpf_iter logic should be the only one needed to access the
bpf_iter_private->iter_private and it can be obtained by, e.g.
"container_of(seq_file->private, struct bpf_iter_private, target_private)"

> 
> > 
> > > 
> > > 
> > > 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.
> > > 

Powered by blists - more mailing lists