Re: [PATCH bpf-next v6 1/2] bpf: Fix attaching fentry/fexit/fmod_ret/lsm to modules

[Jiri Olsa <olsajiri@gmail.com>]

On Fri, Mar 31, 2023 at 01:08:42PM +0200, Petr Mladek wrote:
> On Fri 2023-03-31 17:15:56, Leizhen (ThunderTown) wrote:
> > 
> > 
> > On 2023/3/31 16:31, Petr Mladek wrote:
> > > On Thu 2023-03-30 22:59:12, Jiri Olsa wrote:
> > >> On Thu, Mar 30, 2023 at 08:26:41PM +0800, Leizhen (ThunderTown) wrote:
> > >>>
> > >>>
> > >>> On 2023/3/30 15:29, Jiri Olsa wrote:
> > >>>> ping,
> > >>>>
> > >>>> Petr, Zhen, any comment on discussion below?
> > >>>>
> > >>>> thanks,
> > >>>> jirka
> > >>>>
> > >>>> On Thu, Mar 23, 2023 at 03:00:25PM +0100, Jiri Olsa wrote:
> > >>>>> On Wed, Mar 22, 2023 at 09:03:46AM -0700, Alexei Starovoitov wrote:
> > >>>>>> On Wed, Mar 22, 2023 at 5:14 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> > >>>>>>>
> > >>>>>>> On Wed, Mar 22, 2023 at 10:49:38AM +0100, Artem Savkov wrote:
> > >>>>>>>
> > >>>>>>> SNIP
> > >>>>>>>
> > >>>>>>>>>> Hm, do we even need to preempt_disable? IIUC, preempt_disable is used
> > >>>>>>>>>> in module kallsyms to prevent taking the module lock b/c kallsyms are
> > >>>>>>>>>> used in the oops path. That shouldn't be an issue here, is that correct?
> > >>>>>>>>>
> > >>>>>>>>> btf_try_get_module calls try_module_get which disables the preemption,
> > >>>>>>>>> so no need to call it in here
> > >>>>>>>>
> > >>>>>>>> It does, but it reenables preemption right away so it is enabled by the
> > >>>>>>>> time we call find_kallsyms_symbol_value(). I am getting the following
> > >>>>>>>> lockdep splat while running module_fentry_shadow test from test_progs.
> > >>>>>>>>
> > >>>>>>>> [   12.017973][  T488] =============================
> > >>>>>>>> [   12.018529][  T488] WARNING: suspicious RCU usage
> > >>>>>>>> [   12.018987][  T488] 6.2.0.bpf-test-13063-g6a9f5cdba3c5 #804 Tainted: G           OE
> > >>>>>>>> [   12.019898][  T488] -----------------------------
> > >>>>>>>> [   12.020391][  T488] kernel/module/kallsyms.c:448 suspicious rcu_dereference_check() usage!
> > >>>>>>>> [   12.021335][  T488]
> > >>>>>>>> [   12.021335][  T488] other info that might help us debug this:
> > >>>>>>>> [   12.021335][  T488]
> > >>>>>>>> [   12.022416][  T488]
> > >>>>>>>> [   12.022416][  T488] rcu_scheduler_active = 2, debug_locks = 1
> > >>>>>>>> [   12.023297][  T488] no locks held by test_progs/488.
> > >>>>>>>> [   12.023854][  T488]
> > >>>>>>>> [   12.023854][  T488] stack backtrace:
> > >>>>>>>> [   12.024336][  T488] CPU: 0 PID: 488 Comm: test_progs Tainted: G           OE      6.2.0.bpf-test-13063-g6a9f5cdba3c5 #804
> > >>>>>>>> [   12.025290][  T488] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.1-2.fc37 04/01/2014
> > >>>>>>>> [   12.026108][  T488] Call Trace:
> > >>>>>>>> [   12.026381][  T488]  <TASK>
> > >>>>>>>> [   12.026649][  T488]  dump_stack_lvl+0xb4/0x110
> > >>>>>>>> [   12.027060][  T488]  lockdep_rcu_suspicious+0x158/0x1f0
> > >>>>>>>> [   12.027541][  T488]  find_kallsyms_symbol_value+0xe8/0x110
> > >>>>>>>> [   12.028028][  T488]  bpf_check_attach_target+0x838/0xa20
> > >>>>>>>> [   12.028511][  T488]  check_attach_btf_id+0x144/0x3f0
> > >>>>>>>> [   12.028957][  T488]  ? __pfx_cmp_subprogs+0x10/0x10
> > >>>>>>>> [   12.029408][  T488]  bpf_check+0xeec/0x1850
> > >>>>>>>> [   12.029799][  T488]  ? ktime_get_with_offset+0x124/0x1d0
> > >>>>>>>> [   12.030247][  T488]  bpf_prog_load+0x87a/0xed0
> > >>>>>>>> [   12.030627][  T488]  ? __lock_release+0x5f/0x160
> > >>>>>>>> [   12.031010][  T488]  ? __might_fault+0x53/0xb0
> > >>>>>>>> [   12.031394][  T488]  ? selinux_bpf+0x6c/0xa0
> > >>>>>>>> [   12.031756][  T488]  __sys_bpf+0x53c/0x1240
> > >>>>>>>> [   12.032115][  T488]  __x64_sys_bpf+0x27/0x40
> > >>>>>>>> [   12.032476][  T488]  do_syscall_64+0x3e/0x90
> > >>>>>>>> [   12.032835][  T488]  entry_SYSCALL_64_after_hwframe+0x72/0xdc
> > >>>>>>>
> > >>>>>>> --- a/kernel/module/kallsyms.c
> > >>>>>>> +++ b/kernel/module/kallsyms.c
> > >>> Commit 91fb02f31505 ("module: Move kallsyms support into a separate file") hides
> > >>> the answer. find_kallsyms_symbol_value() was originally a static function, and it
> > >>> is only called by module_kallsyms_lookup_name() and is preemptive-protected.
> > >>>
> > >>> Now that we've added a call to function find_kallsyms_symbol_value(), it seems like
> > >>> we should do the same thing as function module_kallsyms_lookup_name().
> > >>>
> > >>> Like this?
> > >>> +				mod = btf_try_get_module(btf);
> > >>> +				if (mod) {
> > >>> +					preempt_disable();
> > >>> +					addr = find_kallsyms_symbol_value(mod, tname);
> > >>> +					preempt_enable();
> > >>> +				} else
> > >>> +					addr = 0;
> > >>
> > >> yes, that's what I did above, but I was just curious about the strange
> > >> RCU usage Alexei commented on earlier:
> > >>
> > >> 	>>> +unsigned long find_kallsyms_symbol_value(struct module *mod, const char *name)
> > >> 	>>> +{
> > >> 	>>> +       unsigned long ret;
> > >> 	>>> +
> > >> 	>>> +       preempt_disable();
> > >> 	>>> +       ret = __find_kallsyms_symbol_value(mod, name);
> > >> 	>>> +       preempt_enable();
> > >> 	>>> +       return ret;
> > >> 	>>> +}
> > >> 	>>
> > >> 	>> That doesn't look right.
> > >> 	>> I think the issue is misuse of rcu_dereference_sched in
> > >> 	>> find_kallsyms_symbol_value.
> > >> 	>
> > >> 	> it seems to be using rcu pointer to keep symbols for module init time and
> > >> 	> then core symbols for after init.. and switch between them when module is
> > >> 	> loaded, hence the strange rcu usage I think
> > 
> > load_module
> > 	post_relocation
> > 		add_kallsyms
> > 			mod->kallsyms = (void __rcu *)mod->init_layout.base + info->mod_kallsyms_init_off;   (1)
> > 	do_init_module
> > 		freeinit->module_init = mod->init_layout.base;
> > 		rcu_assign_pointer(mod->kallsyms, &mod->core_kallsyms);                                      (2)
> > 		if (llist_add(&freeinit->node, &init_free_list))
> > 			schedule_work(&init_free_wq);
> > 
> > do_free_init
> > 	synchronize_rcu();
> > 	module_memfree(initfree->module_init);
> > 
> > IIUC, the RCU can help synchronize_rcu() in do_free_init() to make sure that no one
> > is still using the first mod->kallsyms (1). If find_kallsyms_symbol_value() is executed
> > between (1) and (2).
> 
> Yes, this seems to be another scenario where the RCU synchronization/access
> is needed.

thanks for the details

still curious.. confusing part for me is the use of rcu_dereference in
add_kallsyms IIUC there's no need for that because mod->kallsyms is not
exposed at that time? we could do without it like in patch below?

thanks,
jirka


---
diff --git a/kernel/module/kallsyms.c b/kernel/module/kallsyms.c
index bdc911dbcde5..bc1e748a1357 100644
--- a/kernel/module/kallsyms.c
+++ b/kernel/module/kallsyms.c
@@ -170,20 +170,18 @@ void add_kallsyms(struct module *mod, const struct load_info *info)
 	Elf_Sym *dst;
 	char *s;
 	Elf_Shdr *symsec = &info->sechdrs[info->index.sym];
+	struct mod_kallsyms *kallsyms;
 	unsigned long strtab_size;
 
-	/* Set up to point into init section. */
-	mod->kallsyms = (void __rcu *)mod->init_layout.base +
-		info->mod_kallsyms_init_off;
+	kallsyms = mod->init_layout.base + info->mod_kallsyms_init_off;
 
-	rcu_read_lock();
 	/* The following is safe since this pointer cannot change */
-	rcu_dereference(mod->kallsyms)->symtab = (void *)symsec->sh_addr;
-	rcu_dereference(mod->kallsyms)->num_symtab = symsec->sh_size / sizeof(Elf_Sym);
+	kallsyms->symtab = (void *)symsec->sh_addr;
+	kallsyms->num_symtab = symsec->sh_size / sizeof(Elf_Sym);
 	/* Make sure we get permanent strtab: don't use info->strtab. */
-	rcu_dereference(mod->kallsyms)->strtab =
+	kallsyms->strtab =
 		(void *)info->sechdrs[info->index.str].sh_addr;
-	rcu_dereference(mod->kallsyms)->typetab = mod->init_layout.base + info->init_typeoffs;
+	kallsyms->typetab = mod->init_layout.base + info->init_typeoffs;
 
 	/*
 	 * Now populate the cut down core kallsyms for after init
@@ -193,20 +191,20 @@ void add_kallsyms(struct module *mod, const struct load_info *info)
 	mod->core_kallsyms.strtab = s = mod->data_layout.base + info->stroffs;
 	mod->core_kallsyms.typetab = mod->data_layout.base + info->core_typeoffs;
 	strtab_size = info->core_typeoffs - info->stroffs;
-	src = rcu_dereference(mod->kallsyms)->symtab;
-	for (ndst = i = 0; i < rcu_dereference(mod->kallsyms)->num_symtab; i++) {
-		rcu_dereference(mod->kallsyms)->typetab[i] = elf_type(src + i, info);
+	src = kallsyms->symtab;
+	for (ndst = i = 0; i < kallsyms->num_symtab; i++) {
+		kallsyms->typetab[i] = elf_type(src + i, info);
 		if (i == 0 || is_livepatch_module(mod) ||
 		    is_core_symbol(src + i, info->sechdrs, info->hdr->e_shnum,
 				   info->index.pcpu)) {
 			ssize_t ret;
 
 			mod->core_kallsyms.typetab[ndst] =
-			    rcu_dereference(mod->kallsyms)->typetab[i];
+			    kallsyms->typetab[i];
 			dst[ndst] = src[i];
 			dst[ndst++].st_name = s - mod->core_kallsyms.strtab;
 			ret = strscpy(s,
-				      &rcu_dereference(mod->kallsyms)->strtab[src[i].st_name],
+				      &kallsyms->strtab[src[i].st_name],
 				      strtab_size);
 			if (ret < 0)
 				break;
@@ -214,8 +212,10 @@ void add_kallsyms(struct module *mod, const struct load_info *info)
 			strtab_size -= ret + 1;
 		}
 	}
-	rcu_read_unlock();
 	mod->core_kallsyms.num_symtab = ndst;
+
+	/* Set up to point into init section. */
+	rcu_assign_pointer(mod->kallsyms, kallsyms);
 }
 
 #if IS_ENABLED(CONFIG_STACKTRACE_BUILD_ID)