Re: [PATCH] srcu: Fix flush sup work warning in cleanup_srcu_struct()
From: Paul E. McKenney
Date: Thu Mar 23 2023 - 01:07:50 EST
On Thu, Mar 23, 2023 at 03:39:12AM +0000, Zhang, Qiang1 wrote:
> > >On Wed, Mar 22, 2023 at 10:08:54PM +0000, Zhang, Qiang1 wrote:
> > > > > > insmod rcutorture.ko
> > > > > > rmmod rcutorture.ko
> > > > > >
> > > > > > [ 209.437327] WARNING: CPU: 0 PID: 508 at kernel/workqueue.c:3167
> > > > > > __flush_work+0x50a/0x540 [ 209.437346] Modules linked in:
> > > > > > rcutorture(-) torture [last unloaded: rcutorture] [ 209.437382]
> > > > > > CPU: 0 PID: 508 Comm: rmmod Tainted: G W 6.3.0-rc1-yocto-standard+
> > > > > > [ 209.437406] RIP: 0010:__flush_work+0x50a/0x540 .....
> > > > > > [ 209.437758] flush_delayed_work+0x36/0x90 [ 209.437776]
> > > > > > cleanup_srcu_struct+0x68/0x2e0 [ 209.437817]
> > > > > > srcu_module_notify+0x71/0x140 [ 209.437854]
> > > > > > blocking_notifier_call_chain+0x9d/0xd0
> > > > > > [ 209.437880] __x64_sys_delete_module+0x223/0x2e0
> > > > > > [ 209.438046] do_syscall_64+0x43/0x90 [ 209.438062]
> > > > > > entry_SYSCALL_64_after_hwframe+0x72/0xdc
> > > > > >
> > > > > > For srcu objects defined with DEFINE_SRCU() or DEFINE_STATIC_SRCU(),
> > > > > > when compiling and loading as modules, the srcu_module_coming() is
> > > > > > invoked, allocate memory for srcu structure's->sda and initialize
> > > > > > sda structure, due to not fully initialize srcu structure's->sup, so
> > > > > > at this time the sup structure's->delaywork.func is null, if not
> > > > > > invoke init_srcu_struct_fields() before unloading modules, in
> > > > > > srcu_module_going() the __flush_work() find
> > > > > > work->func is empty, will raise the warning above.
> > > > > >
> > > > > > This commit add init_srcu_struct_fields() to initialize srcu
> > > > > > structure's->sup, in srcu_module_coming().
> > > > > >
> > > > > > Signed-off-by: Zqiang <qiang1.zhang@xxxxxxxxx>
> > > > > >
> > > > > >Good catch, and thank you for testing the in-module case!
> > > > > >
> > > > > >One question below...
> > > > > >
> > > > > > Thanx, Paul
> > > > > >
> > > > > > ---
> > > > > > kernel/rcu/srcutree.c | 11 ++++++++---
> > > > > > 1 file changed, 8 insertions(+), 3 deletions(-)
> > > > > >
> > > > > > diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c index
> > > > > > 1fb078abbdc9..42d8720e016c 100644
> > > > > > --- a/kernel/rcu/srcutree.c
> > > > > > +++ b/kernel/rcu/srcutree.c
> > > > > > @@ -1921,7 +1921,8 @@ static int srcu_module_coming(struct module *mod)
> > > > > > ssp->sda = alloc_percpu(struct srcu_data);
> > > > > > if (WARN_ON_ONCE(!ssp->sda))
> > > > > > return -ENOMEM;
> > > > > > - init_srcu_struct_data(ssp);
> > > > > > + if (WARN_ON_ONCE(init_srcu_struct_fields(ssp, true)))
> > > > > > + return -ENOMEM;
> > > > > >
> > > > > >Wouldn't it be better to simply delete the init_srcu_struct_data()?
> > > > > >
> > > > > >Then the first call to check_init_srcu_struct() would take care of
> > > > > >the initialization, just as for the non-module case. Or am I missing
> > > > > >something subtle?
> > > > >
> > > > > Hi Paul
> > > > >
> > > > > Maybe the check_init_srcu_struct() is always not invoked, for example,
> > > > > In rcutorture.c, here is such a definition DEFINE_STATIC_SRCU(srcu_ctl),
> > > > > but we use torture_type=rcu to test, there will not be any interface
> > > > > calling
> > > > > check_init_srcu_struct() to initialize srcu_ctl and set
> > > > > structure's->delaywork.func is process_srcu().
> > > > > when we unload the rcutorture module, invoke cleanup_srcu_struct() to
> > > > > flush sup structure's->delaywork.func, due to the func pointer is not
> > > > > initialize, it's null, will trigger warning.
> > > > >
> > > > > About kernel/workqueue.c:3167
> > > > >
> > > > > __flush_work
> > > > > if (WARN_ON(!work->func)) <---------trigger waning
> > > > > return false;
> > > > >
> > > > >
> > > > > and in init_srcu_struct_fields(ssp, true), wil set
> > > > > srcu_sup->sda_is_static is true and set srcu_sup-> srcu_gp_seq_needed
> > > > > is zero, after that when we call
> > > > > check_init_srcu_struct() again, it not be initialized again.
> > > > >
> > > > >
> > > > >Good point! In the non-module statically allocated case there is never a call to cleanup_srcu_struct().
> > > > >
> > > > >So suppose the code in srcu_module_coming() only did the alloc_percpu(), and then the
> > > > >code in srcu_module_going() only did the the matching
> > > > >free_percpu() instead of the current cleanup_srcu_struct()?
> > > >
> > > > But in modules, for srcu objects defined with DEFINE_SRCU() or DEFINE_STATIC_SRCU(),
> > > > when a module is unloaded, we usually don't call cleanup_srcu_struct() in the module
> > > > unload function.
> > > > If in srcu_module_going() only do free_percpu(), the srcu_sup->node memory maybe
> > > > can not free and also lost the opportunity to refresh the running work.
> > > >
> > > >
> > > >But in the module case, isn't the srcu_sup->node also statically
> > > >allocated via the "static struct srcu_usage" declaration?
> > >
> > > static bool init_srcu_struct_nodes(struct srcu_struct *ssp, gfp_t gfp_flags)
> > > {
> > > sp->srcu_sup->node = kcalloc(rcu_num_nodes, sizeof(*ssp->srcu_sup->node), gfp_flags);
> > > ...
> > > }
> > >
> > > Regardless of whether the srcu object is declared in the module or not, sup->node is dynamically allocated.
> > > right?
> > >
> > >You are absolutely right, thank you!
> > >
> > >There are a couple of ways to resolve this. One is to simply add
> > >a check_init_srcu_struct() before the call to cleanup_srcu_struct()
> > >from srcu_module_going(), as shown below. This seems a bit silly,
> > >potentially initializing fields for no good reason.
> > >
> > >Another way is to make cleanup_srcu_struct() do the same check
> > >that check_init_srcu_struct() does:
> > >
> > > rcu_seq_state(smp_load_acquire(&ssp->srcu_sup->srcu_gp_seq_needed))
> > >
> > >If the value is non-zero, then cleanup_srcu_struct() should skip
> > >consistency checks that complain about things that cannot happen if
> > >there never was an RCU grace period. Maybe something as shown after
> > >the second line of dashes.
> > >
> > >Thoughts?
> > >
> > > Thanx, Paul
> > >
> > >------------------------------------------------------------------------
> > >
> > >
> > >/* Initialize any global-scope srcu_struct structures used by this module. */
> > >static int srcu_module_coming(struct module *mod)
> > >{
> > > int i;
> > > struct srcu_struct **sspp = mod->srcu_struct_ptrs;
> > > struct srcu_struct *ssp;
> > >
> > > for (i = 0; i < mod->num_srcu_structs; i++) {
> > > ssp = *(sspp++);
> > > ssp->sda = alloc_percpu(struct srcu_data);
> > > if (WARN_ON_ONCE(!ssp->sda))
> > > return -ENOMEM;
> > > init_srcu_struct_data(ssp);
> > > }
> > > return 0;
> > >}
> > >
> > >/* Clean up any global-scope srcu_struct structures used by this module. */
> > >static void srcu_module_going(struct module *mod)
> > >{
> > > int i;
> > > struct srcu_struct *ssp;
> > > struct srcu_struct **sspp = mod->srcu_struct_ptrs;
> > >
> > > for (i = 0; i < mod->num_srcu_structs; i++) {
> > > ssp = *(sspp++);
> > > check_init_srcu_struct(ssp);
> > > cleanup_srcu_struct(ssp);
> > > }
> > >}
> > >
> > >------------------------------------------------------------------------
> > >
> > >void cleanup_srcu_struct(struct srcu_struct *ssp)
> > >{
> > > int cpu;
> > > struct srcu_usage *sup = ssp->srcu_sup;
> > > bool wasused = !rcu_seq_state(smp_load_acquire(&ssp->srcu_sup->srcu_gp_seq_needed));
> > >
> > > if (WARN_ON(wasused && !srcu_get_delay(ssp)))
> > > return; /* Just leak it! */
> > > if (WARN_ON(srcu_readers_active(ssp)))
> > > return; /* Just leak it! */
> > > flush_delayed_work(&sup->work);
> > > if (wasused) {
> >
> > If wasused=false It not need to invoke flush_delayed_work(&sup->work);
> > this trigger WARN_ON(!work->func)) .
> >
> >Again, good catch! I will pull that flush_delayed_work() into this
> >"if" statement.
> >
> > > for_each_possible_cpu(cpu) {
> > > struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu);
> > >
> > > del_timer_sync(&sdp->delay_work);
> > > flush_work(&sdp->work);
> > > if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist)))
> > > return; /* Forgot srcu_barrier(), so just leak it! */
> > > }
> > > }
> > > if (WARN_ON(wasused && rcu_seq_state(READ_ONCE(sup->srcu_gp_seq)) != SRCU_STATE_IDLE) ||
> > > WARN_ON(wasused && rcu_seq_current(&sup->srcu_gp_seq) != sup->srcu_gp_seq_needed) ||
> > > WARN_ON(srcu_readers_active(ssp))) {
> > > pr_info("%s: Active srcu_struct %p read state: %d gp state: %lu/%lu\n",
> > > __func__, ssp, rcu_seq_state(READ_ONCE(sup->srcu_gp_seq)),
> > > rcu_seq_current(&sup->srcu_gp_seq), sup->srcu_gp_seq_needed);
> > > return; /* Caller forgot to stop doing call_srcu()? */
> > > }
> > > kfree(sup->node);
> > > sup->node = NULL;
> > > sup->srcu_size_state = SRCU_SIZE_SMALL;
> > > if (!sup->sda_is_static) {
> > > free_percpu(ssp->sda);
> > > ssp->sda = NULL;
> > > kfree(sup);
> > > ssp->srcu_sup = NULL;
> > > }
> > >}
> >
> >
> > If we have not invoke check_init_srcu_struct() , that means call_srcu(), synchronize_srcu(), srcu_barrier(), start_poll_synchronize_srcu() are also not invoke, so Is there no need to check
> > srcu_readers_active()?
> >
> >The module might have had lots of SRCU readers, but no updaters, and
> >a bug in that module might mean that that readers are still active.
> >For example, the module might have passed the srcu_struct structure
> >to some function in the main kernel, and then forgotten to tell that
> >function to stop doing srcu_read_lock() and srcu_read_unlock() on it.
> >
> >Or the module might have created a kthread that did SRCU readers, and
> >then have forgotten to stop that kthread.
> >
> >
> >fully understand.
> >
> >
> >Please see below for an untested patch.
> >
> >I will test the following modification.
> >
> >Thanks
> >Zqiang
> >
> >
> >And yet again, thoughts? ;-)
>
> Maybe add the following modification, otherwise, in cleanup_srcu_struct()
> Kfree(sup) will release ssp->srcu_sup which is statically allocated.
>
> @@ -1926,6 +1926,7 @@ static int srcu_module_coming(struct module *mod)
> if (WARN_ON_ONCE(!ssp->sda))
> return -ENOMEM;
> init_srcu_struct_data(ssp);
> + ssp->srcu_sup->sda_is_static = true;
> }
Good catch and good point!
But the underlying problem is that I am still making things too complex.
How about the following? The idea is to skip the cleanup_srcu_struct()
unless there was a call to check_init_srcu_struct(), and to free the
per-CPU data either way.
Thanx, Paul
------------------------------------------------------------------------
diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 1fb078abbdc9..06f8ed1ce272 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -1921,7 +1921,6 @@ static int srcu_module_coming(struct module *mod)
ssp->sda = alloc_percpu(struct srcu_data);
if (WARN_ON_ONCE(!ssp->sda))
return -ENOMEM;
- init_srcu_struct_data(ssp);
}
return 0;
}
@@ -1930,10 +1929,17 @@ static int srcu_module_coming(struct module *mod)
static void srcu_module_going(struct module *mod)
{
int i;
+ struct srcu_data __percpu *sda;
+ struct srcu_struct *ssp;
struct srcu_struct **sspp = mod->srcu_struct_ptrs;
- for (i = 0; i < mod->num_srcu_structs; i++)
- cleanup_srcu_struct(*(sspp++));
+ for (i = 0; i < mod->num_srcu_structs; i++) {
+ ssp = *(sspp++);
+ sda = ssp->sda;
+ if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_sup->srcu_gp_seq_needed)))
+ cleanup_srcu_struct(ssp);
+ free_percpu(sda);
+ }
}
/* Handle one module, either coming or going. */