Date:   Wed, 6 May 2020 13:51:46 -0700
From:   Andrii Nakryiko <andrii.nakryiko@...il.com>
To:     Yonghong Song <yhs@...com>
Cc:     Andrii Nakryiko <andriin@...com>, bpf <bpf@...r.kernel.org>,
        Martin KaFai Lau <kafai@...com>,
        Networking <netdev@...r.kernel.org>,
        Alexei Starovoitov <ast@...com>,
        Daniel Borkmann <daniel@...earbox.net>,
        Kernel Team <kernel-team@...com>
Subject: Re: [PATCH bpf-next v2 11/20] bpf: add task and task/file iterator targets

On Wed, May 6, 2020 at 11:24 AM Yonghong Song <yhs@...com> wrote:
>
>
>
> On 5/6/20 12:30 AM, Andrii Nakryiko wrote:
> > On Sun, May 3, 2020 at 11:28 PM Yonghong Song <yhs@...com> wrote:
> >>
> >> Only the tasks belonging to "current" pid namespace
> >> are enumerated.
> >>
> >> For task/file target, the bpf program will have access to
> >>    struct task_struct *task
> >>    u32 fd
> >>    struct file *file
> >> where fd/file is an open file for the task.
> >>
> >> Signed-off-by: Yonghong Song <yhs@...com>
> >> ---
> >
> > I might be missing some subtleties with task refcounting for task_file
> > iterator, asked few questions below...
> >
> >>   kernel/bpf/Makefile    |   2 +-
> >>   kernel/bpf/task_iter.c | 336 +++++++++++++++++++++++++++++++++++++++++
> >>   2 files changed, 337 insertions(+), 1 deletion(-)
> >>   create mode 100644 kernel/bpf/task_iter.c
> >>
> >> diff --git a/kernel/bpf/Makefile b/kernel/bpf/Makefile
> >> index b2b5eefc5254..37b2d8620153 100644
> >> --- a/kernel/bpf/Makefile
> >> +++ b/kernel/bpf/Makefile
> >> @@ -2,7 +2,7 @@
> >>   obj-y := core.o
> >>   CFLAGS_core.o += $(call cc-disable-warning, override-init)
> >>
> >> -obj-$(CONFIG_BPF_SYSCALL) += syscall.o verifier.o inode.o helpers.o tnum.o bpf_iter.o map_iter.o
> >> +obj-$(CONFIG_BPF_SYSCALL) += syscall.o verifier.o inode.o helpers.o tnum.o bpf_iter.o map_iter.o task_iter.o
> >>   obj-$(CONFIG_BPF_SYSCALL) += hashtab.o arraymap.o percpu_freelist.o bpf_lru_list.o lpm_trie.o map_in_map.o
> >>   obj-$(CONFIG_BPF_SYSCALL) += local_storage.o queue_stack_maps.o
> >>   obj-$(CONFIG_BPF_SYSCALL) += disasm.o
> >> diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c
> >> new file mode 100644
> >> index 000000000000..1ca258f6e9f4
> >> --- /dev/null
> >> +++ b/kernel/bpf/task_iter.c
> >> @@ -0,0 +1,336 @@
> >> +// SPDX-License-Identifier: GPL-2.0-only
> >> +/* Copyright (c) 2020 Facebook */
> >> +
> >> +#include <linux/init.h>
> >> +#include <linux/namei.h>
> >> +#include <linux/pid_namespace.h>
> >> +#include <linux/fs.h>
> >> +#include <linux/fdtable.h>
> >> +#include <linux/filter.h>
> >> +
> >> +struct bpf_iter_seq_task_common {
> >> +       struct pid_namespace *ns;
> >> +};
> >> +
> >> +struct bpf_iter_seq_task_info {
> >> +       struct bpf_iter_seq_task_common common;
> >
> > you have comment below in init_seq_pidns() that common is supposed to
> > be the very first field, but I think it's more important and
> > appropriate here, so that whoever adds anything here knows that order
> > of field is important.
>
> I can move the comments here.
>
> >
> >> +       struct task_struct *task;
> >> +       u32 id;
> >> +};
> >> +
> >
> > [...]
> >
> >> +static int __task_seq_show(struct seq_file *seq, void *v, bool in_stop)
> >> +{
> >> +       struct bpf_iter_meta meta;
> >> +       struct bpf_iter__task ctx;
> >> +       struct bpf_prog *prog;
> >> +       int ret = 0;
> >> +
> >> +       meta.seq = seq;
> >> +       prog = bpf_iter_get_info(&meta, in_stop);
> >> +       if (prog) {
> >
> >
> > nit: `if (!prog) return 0;` here would reduce nesting level below
> >
> >> +               meta.seq = seq;
> >> +               ctx.meta = &meta;
> >> +               ctx.task = v;
> >> +               ret = bpf_iter_run_prog(prog, &ctx);
> >> +       }
> >> +
> >> +       return 0;
> >
> > return **ret**; ?
>
> It should return "ret". In task_file show() code is similar but correct.
> I can do early return with !prog too although we do not have
> deep nesting level yet.
>
> >
> >> +}
> >> +
> >
> > [...]
> >
> >> +
> >> +static struct file *task_file_seq_get_next(struct pid_namespace *ns, u32 *id,
> >> +                                          int *fd, struct task_struct **task,
> >> +                                          struct files_struct **fstruct)
> >> +{
> >> +       struct files_struct *files;
> >> +       struct task_struct *tk;
> >> +       u32 sid = *id;
> >> +       int sfd;
> >> +
> >> +       /* If this function returns a non-NULL file object,
> >> +        * it held a reference to the files_struct and file.
> >> +        * Otherwise, it does not hold any reference.
> >> +        */
> >> +again:
> >> +       if (*fstruct) {
> >> +               files = *fstruct;
> >> +               sfd = *fd;
> >> +       } else {
> >> +               tk = task_seq_get_next(ns, &sid);
> >> +               if (!tk)
> >> +                       return NULL;
> >> +
> >> +               files = get_files_struct(tk);
> >> +               put_task_struct(tk);
> >
> > task is put here, but is still used below.. is there some additional
> > hidden refcounting involved?
>
> Good question. I had an impression that we take a reference count
> for task->files so task should not go away. But reading linux
> code again, I do not have sufficient evidence to back my claim.
> So I will reference count task as well, e.g., do not put_task_struct()
> until all files are done here.

All threads within the process share files table. So some threads
might exit, but files will stay, which is why task_struct and
files_struct have separate refcounting, and having refcount on files
doesn't guarantee any particular task will stay alive for long enough.
So I think we need to refcount both files and task in this case.
Reading source code of copy_files() in kernel/fork.c (CLONE_FILES
flags just bumps refcnt on old process' files_struct), seems to
confirm this as well.

>
> >
> >> +               if (!files) {
> >> +                       sid = ++(*id);
> >> +                       *fd = 0;
> >> +                       goto again;
> >> +               }
> >> +               *fstruct = files;
> >> +               *task = tk;
> >> +               if (sid == *id) {
> >> +                       sfd = *fd;
> >> +               } else {
> >> +                       *id = sid;
> >> +                       sfd = 0;
> >> +               }
> >> +       }
> >> +
> >> +       rcu_read_lock();
> >> +       for (; sfd < files_fdtable(files)->max_fds; sfd++) {
> >
> > files_fdtable does rcu_dereference on each iteration, would it be
> > better to just cache files_fdtable(files)->max_fds into local
> > variable? It's unlikely that there will be many iterations, but
> > still...
>
> I borrowed code from fs/proc/fd.c. But I can certainly to avoid
> repeated reading max_fds as suggested.
>
> >
> >> +               struct file *f;
> >> +
> >> +               f = fcheck_files(files, sfd);
> >> +               if (!f)
> >> +                       continue;
> >> +               *fd = sfd;
> >> +               get_file(f);
> >> +               rcu_read_unlock();
> >> +               return f;
> >> +       }
> >> +
> >> +       /* the current task is done, go to the next task */
> >> +       rcu_read_unlock();
> >> +       put_files_struct(files);
> >> +       *fstruct = NULL;
> >
> > *task = NULL; for completeness?
>
> if *fstruct == NULL, will try to get next task, so *task = NULL
> is unnecessary, but I can add it, won't hurt and possibly make
> it easy to understand.
>
> >
> >> +       sid = ++(*id);
> >> +       *fd = 0;
> >> +       goto again;
> >> +}
> >> +
> >> +static void *task_file_seq_start(struct seq_file *seq, loff_t *pos)
> >> +{
> >> +       struct bpf_iter_seq_task_file_info *info = seq->private;
> >> +       struct files_struct *files = NULL;
> >> +       struct task_struct *task = NULL;
> >> +       struct file *file;
> >> +       u32 id = info->id;
> >> +       int fd = info->fd;
> >> +
> >> +       file = task_file_seq_get_next(info->common.ns, &id, &fd, &task, &files);
> >> +       if (!file) {
> >> +               info->files = NULL;
> >
> > what about info->task here?
>
> info->files == NULL indicates the end of iteration, info->task will not
> be checked any more. But I guess, I can assign NULL to task as well to
> avoid confusion.
>
> >
> >> +               return NULL;
> >> +       }
> >> +
> >> +       ++*pos;
> >> +       info->id = id;
> >> +       info->fd = fd;
> >> +       info->task = task;
> >> +       info->files = files;
> >> +
> >> +       return file;
> >> +}
> >> +
> >
> > [...]
> >
> >> +
> >> +struct bpf_iter__task_file {
> >> +       __bpf_md_ptr(struct bpf_iter_meta *, meta);
> >> +       __bpf_md_ptr(struct task_struct *, task);
> >> +       u32 fd;
> >
> > nit: sort of works by accident (due to all other field being 8-byte
> > aligned pointers), shall we add __attribute__((aligned(8)))?
>
> This is what I thought as well. It should work. But I think
> add aligned(8) wont' hurt to expresss the intention.. Will add it.
>
> >
> >> +       __bpf_md_ptr(struct file *, file);
> >> +};
> >> +
> >
> > [...]
> >

Powered by blists - more mailing lists