On Thu, Oct 13, 2022 at 11:50:54AM -0700, Andrii Nakryiko wrote:
> On Sun, Oct 9, 2022 at 3:00 PM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> >
> > Currently we allow to create kprobe multi link on function from kernel
> > module, but we don't take the module reference to ensure it's not
> > unloaded while we are tracing it.
> >
> > The multi kprobe link is based on fprobe/ftrace layer which takes
> > different approach and releases ftrace hooks when module is unloaded
> > even if there's tracer registered on top of it.
> >
> > Adding code that gathers all the related modules for the link and takes
> > their references before it's attached. All kernel module references are
> > released after link is unregistered.
> >
> > Note that we do it the same way already for trampoline probes
> > (but for single address).
> >
> > Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> > ---
> > kernel/trace/bpf_trace.c | 100 +++++++++++++++++++++++++++++++++++++++
> > 1 file changed, 100 insertions(+)
> >
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index 9be1a2b6b53b..f3d7565fee79 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -2447,6 +2447,8 @@ struct bpf_kprobe_multi_link {
> > unsigned long *addrs;
> > u64 *cookies;
> > u32 cnt;
> > + struct module **mods;
> > + u32 mods_cnt;
> > };
> >
> > struct bpf_kprobe_multi_run_ctx {
> > @@ -2502,6 +2504,14 @@ static int copy_user_syms(struct user_syms *us, unsigned long __user *usyms, u32
> > return err;
> > }
> >
> > +static void kprobe_multi_put_modules(struct module **mods, u32 cnt)
> > +{
> > + u32 i;
> > +
> > + for (i = 0; i < cnt; i++)
> > + module_put(mods[i]);
> > +}
> > +
> > static void free_user_syms(struct user_syms *us)
> > {
> > kvfree(us->syms);
> > @@ -2514,6 +2524,7 @@ static void bpf_kprobe_multi_link_release(struct bpf_link *link)
> >
> > kmulti_link = container_of(link, struct bpf_kprobe_multi_link, link);
> > unregister_fprobe(&kmulti_link->fp);
> > + kprobe_multi_put_modules(kmulti_link->mods, kmulti_link->mods_cnt);
> > }
> >
> > static void bpf_kprobe_multi_link_dealloc(struct bpf_link *link)
> > @@ -2523,6 +2534,7 @@ static void bpf_kprobe_multi_link_dealloc(struct bpf_link *link)
> > kmulti_link = container_of(link, struct bpf_kprobe_multi_link, link);
> > kvfree(kmulti_link->addrs);
> > kvfree(kmulti_link->cookies);
> > + kfree(kmulti_link->mods);
> > kfree(kmulti_link);
> > }
> >
> > @@ -2658,6 +2670,80 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
> > }
> > }
> >
> > +struct module_addr_args {
> > + unsigned long *addrs;
> > + u32 addrs_cnt;
> > + struct module **mods;
> > + int mods_cnt;
> > + int mods_alloc;
> > +};
> > +
> > +static int module_callback(void *data, const char *name,
> > + struct module *mod, unsigned long addr)
> > +{
> > + struct module_addr_args *args = data;
> > + bool realloc = !args->mods;
> > + struct module **mods;
> > +
> > + /* We iterate all modules symbols and for each we:
> > + * - search for it in provided addresses array
> > + * - if found we check if we already have the module pointer stored
> > + * (we iterate modules sequentially, so we can check just the last
> > + * module pointer)
> > + * - take module reference and store it
> > + */
> > + if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(unsigned long),
>
> nit: sizeof(addr) is shorter and will stay in sync with addr variable?
ok
>
> > + bpf_kprobe_multi_addrs_cmp))
> > + return 0;
> > +
> > + if (args->mods) {
> > + struct module *prev = NULL;
> > +
> > + if (args->mods_cnt > 1)
> > + prev = args->mods[args->mods_cnt - 1];
>
> doesn't args->mods != NULL imply that args->mods_cnt > 1?
>
> > + if (prev == mod)
> > + return 0;
> > + if (args->mods_cnt == args->mods_alloc)
>
> nit: in libbpf we consistently use the cnt and cap (capacity)
> terminology for this, "mods_alloc" reads like a bool flag or something
ok
>
> > + realloc = true;
> > + }
> > +
> > + if (realloc) {
> > + args->mods_alloc += 100;
>
> agree with Song, this looks pretty arbitrary and quite large. Again,
> from libbpf experience, we do something like:
>
> mods_alloc = max(16, mods_alloc * 3 / 2);
>
> so grow by 50%, but start of with reasonable 16-element array. We can
> use similar approach here.
ok
>
> > + mods = krealloc_array(args->mods, args->mods_alloc, sizeof(*mods), GFP_KERNEL);
> > + if (!mods)
> > + return -ENOMEM;
> > + args->mods = mods;
> > + }
>
> Previous two blocks read pretty convoluted. Isn't it equivalent to simpler:
>
> if (args->mods && args->mods[args->mods_cnt - 1] == mod)
> return 0;
>
> if (args->mods_cnt == args->mods_alloc /* but I'd use mods_cap */) {
> /* realloc here */
> }
sure, I can chage to that
thanks,
jirka
