Re: [PATCH 3/8] bpf: Add bpf_cookie support to fprobe

[Andrii Nakryiko]

On Tue, Feb 8, 2022 at 1:07 AM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
>
> On Mon, Feb 07, 2022 at 10:59:21AM -0800, Andrii Nakryiko wrote:
> > On Wed, Feb 2, 2022 at 5:54 AM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> > >
> > > Adding support to call bpf_get_attach_cookie helper from
> > > kprobe program attached by fprobe link.
> > >
> > > The bpf_cookie is provided by array of u64 values, where
> > > each value is paired with provided function address with
> > > the same array index.
> > >
> > > Suggested-by: Andrii Nakryiko <andrii@xxxxxxxxxx>
> > > Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> > > ---
> > >  include/linux/bpf.h            |  2 +
> > >  include/uapi/linux/bpf.h       |  1 +
> > >  kernel/bpf/syscall.c           | 83 +++++++++++++++++++++++++++++++++-
> > >  kernel/trace/bpf_trace.c       | 16 ++++++-
> > >  tools/include/uapi/linux/bpf.h |  1 +
> > >  5 files changed, 100 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > index 6eb0b180d33b..7b65f05c0487 100644
> > > --- a/include/linux/bpf.h
> > > +++ b/include/linux/bpf.h
> > > @@ -1301,6 +1301,8 @@ static inline void bpf_reset_run_ctx(struct bpf_run_ctx *old_ctx)
> > >  #endif
> > >  }
> > >
> > > +u64 bpf_fprobe_cookie(struct bpf_run_ctx *ctx, u64 ip);
> > > +
> > >  /* BPF program asks to bypass CAP_NET_BIND_SERVICE in bind. */
> > >  #define BPF_RET_BIND_NO_CAP_NET_BIND_SERVICE                   (1 << 0)
> > >  /* BPF program asks to set CN on the packet. */
> > > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > > index c0912f0a3dfe..0dc6aa4f9683 100644
> > > --- a/include/uapi/linux/bpf.h
> > > +++ b/include/uapi/linux/bpf.h
> > > @@ -1484,6 +1484,7 @@ union bpf_attr {
> > >                                 __aligned_u64   addrs;
> > >                                 __u32           cnt;
> > >                                 __u32           flags;
> > > +                               __aligned_u64   bpf_cookies;
> >
> > maybe put it right after addrs, they are closely related and cnt
> > describes all of syms/addrs/cookies.
>
> ok
>
> >
> > >                         } fprobe;
> > >                 };
> > >         } link_create;
> > > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> > > index 0cfbb112c8e1..6c5e74bc43b6 100644
> > > --- a/kernel/bpf/syscall.c
> > > +++ b/kernel/bpf/syscall.c
> > > @@ -33,6 +33,8 @@
> > >  #include <linux/rcupdate_trace.h>
> > >  #include <linux/memcontrol.h>
> > >  #include <linux/fprobe.h>
> > > +#include <linux/bsearch.h>
> > > +#include <linux/sort.h>
> > >
> > >  #define IS_FD_ARRAY(map) ((map)->map_type == BPF_MAP_TYPE_PERF_EVENT_ARRAY || \
> > >                           (map)->map_type == BPF_MAP_TYPE_CGROUP_ARRAY || \
> > > @@ -3025,10 +3027,18 @@ static int bpf_perf_link_attach(const union bpf_attr *attr, struct bpf_prog *pro
> > >
> > >  #ifdef CONFIG_FPROBE
> > >
> > > +struct bpf_fprobe_cookie {
> > > +       unsigned long addr;
> > > +       u64 bpf_cookie;
> > > +};
> > > +
> > >  struct bpf_fprobe_link {
> > >         struct bpf_link link;
> > >         struct fprobe fp;
> > >         unsigned long *addrs;
> > > +       struct bpf_run_ctx run_ctx;
> > > +       struct bpf_fprobe_cookie *bpf_cookies;
> >
> > you already have all the addrs above, why keeping a second copy of
> > each addrs in bpf_fprobe_cookie. Let's have two arrays: addrs
> > (unsigned long) and cookies (u64) and make sure that they are sorted
> > together. Then lookup addrs, calculate index, use that index to fetch
> > cookie.
> >
> > Seems like sort_r() provides exactly the interface you'd need to do
> > this very easily. Having addrs separate from cookies also a bit
> > advantageous in terms of TLB misses (if you need any more persuasion
> > ;)
>
> no persuation needed, I actually tried that but it turned out sort_r
> is not ready yet ;-)
>
> because you can't pass priv pointer to the swap callback, so we can't
> swap the other array.. I did a change to allow that, but it's not trivial
> and will need some bigger testing/review because the original sort
> calls sort_r, and of course there are many 'sort' users ;-)

Big sigh... :( Did you do something similar to _CMP_WRAPPER? You don't
need to change the interface of sort(), so it shouldn't require
extensive code refactoring. You'll just need to adjust priv to be not
just cmp_func, but cmp_func + swap_fun (need a small struct on the
stack in sort, probably). Or you did something else?

>
> >
> > > +       u32 cnt;
> > >  };
> > >
> > >  static void bpf_fprobe_link_release(struct bpf_link *link)
> > > @@ -3045,6 +3055,7 @@ static void bpf_fprobe_link_dealloc(struct bpf_link *link)
> > >
> > >         fprobe_link = container_of(link, struct bpf_fprobe_link, link);
> > >         kfree(fprobe_link->addrs);
> > > +       kfree(fprobe_link->bpf_cookies);
> > >         kfree(fprobe_link);
> > >  }
> > >
> > > @@ -3053,9 +3064,37 @@ static const struct bpf_link_ops bpf_fprobe_link_lops = {
> > >         .dealloc = bpf_fprobe_link_dealloc,
> > >  };
> > >
> > > +static int bpf_fprobe_cookie_cmp(const void *_a, const void *_b)
> > > +{
> > > +       const struct bpf_fprobe_cookie *a = _a;
> > > +       const struct bpf_fprobe_cookie *b = _b;
> > > +
> > > +       if (a->addr == b->addr)
> > > +               return 0;
> > > +       return a->addr < b->addr ? -1 : 1;
> > > +}
> > > +
> > > +u64 bpf_fprobe_cookie(struct bpf_run_ctx *ctx, u64 ip)
> > > +{
> > > +       struct bpf_fprobe_link *fprobe_link;
> > > +       struct bpf_fprobe_cookie *val, key = {
> > > +               .addr = (unsigned long) ip,
> > > +       };
> > > +
> > > +       if (!ctx)
> > > +               return 0;
> >
> > is it allowed to have ctx == NULL?
>
> nope, I was also thinking this is more 'WARN_ON[_ONCE]' check
>
> >
> > > +       fprobe_link = container_of(ctx, struct bpf_fprobe_link, run_ctx);
> > > +       if (!fprobe_link->bpf_cookies)
> > > +               return 0;
> > > +       val = bsearch(&key, fprobe_link->bpf_cookies, fprobe_link->cnt,
> > > +                     sizeof(key), bpf_fprobe_cookie_cmp);
> > > +       return val ? val->bpf_cookie : 0;
> > > +}
> > > +
> > >  static int fprobe_link_prog_run(struct bpf_fprobe_link *fprobe_link,
> > >                                 struct pt_regs *regs)
> > >  {
> > > +       struct bpf_run_ctx *old_run_ctx;
> > >         int err;
> > >
> > >         if (unlikely(__this_cpu_inc_return(bpf_prog_active) != 1)) {
> > > @@ -3063,12 +3102,16 @@ static int fprobe_link_prog_run(struct bpf_fprobe_link *fprobe_link,
> > >                 goto out;
> > >         }
> > >
> > > +       old_run_ctx = bpf_set_run_ctx(&fprobe_link->run_ctx);
> > > +
> > >         rcu_read_lock();
> > >         migrate_disable();
> > >         err = bpf_prog_run(fprobe_link->link.prog, regs);
> > >         migrate_enable();
> > >         rcu_read_unlock();
> > >
> > > +       bpf_reset_run_ctx(old_run_ctx);
> > > +
> > >   out:
> > >         __this_cpu_dec(bpf_prog_active);
> > >         return err;
> > > @@ -3161,10 +3204,12 @@ static int fprobe_resolve_syms(const void *usyms, u32 cnt,
> > >
> > >  static int bpf_fprobe_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
> > >  {
> > > +       struct bpf_fprobe_cookie *bpf_cookies = NULL;
> > >         struct bpf_fprobe_link *link = NULL;
> > >         struct bpf_link_primer link_primer;
> > > +       void __user *ubpf_cookies;
> > > +       u32 flags, cnt, i, size;
> > >         unsigned long *addrs;
> > > -       u32 flags, cnt, size;
> > >         void __user *uaddrs;
> > >         void __user *usyms;
> > >         int err;
> > > @@ -3205,6 +3250,37 @@ static int bpf_fprobe_link_attach(const union bpf_attr *attr, struct bpf_prog *p
> > >                         goto error;
> > >         }
> > >
> > > +       ubpf_cookies = u64_to_user_ptr(attr->link_create.fprobe.bpf_cookies);
> >
> > nit: let's call all this "cookies", this bpf_ prefix feels a bit
> > redundant (I know about perf_event.bpf_cookie, but still).
>
> ok
>
> >
> > > +       if (ubpf_cookies) {
> > > +               u64 *tmp;
> > > +
> > > +               err = -ENOMEM;
> > > +               tmp = kzalloc(size, GFP_KERNEL);
> >
> > kvmalloc?
>
> ok
>
> thanks,
> jirka
>