Mahmoud Mandour <ma.mandourr@gmail.com> writes:

> On Mon, Jul 19, 2021 at 11:48 AM Alex Bennée <alex.bennee@linaro.org> wrote:
>
>  Mahmoud Mandour <ma.mandourr@gmail.com> writes:
>
>  > Since callbacks may be interleaved because of multithreaded execution,
>  > we should not make assumptions about `plugin_exit` either. The problem
>  > with `plugin_exit` is that it frees shared data structures (caches and
>  > `miss_ht` hash table). It should not be assumed that callbacks will not
>  > be called after it and hence use already-freed data structures.
>
>  What was your test case for this because I wonder if it would be worth
>  coming up with one for check-tcg?
>
> I think just any ad-hoc multithreaded execution will evoke the race pretty much
> consistently.

I haven't managed to trigger it with testthread but of course my
libcache is trying to to defend against it.

> 
>  From what I remember the race is
>  in between preexit_cleanup and the eventual _exit/exit_group which nixes
>  all other child threads. Maybe we should be triggering a more graceful
>  unload here?
>
> I think so. This remedies the bug for this particular plugin and I think there
> would be a better solution of course. However, I just can't ever get plugin_exit
> callback to be called more than once so I think that's probably not the problem?
>
> The problem is that we *use* the data in translation/mem_access/exec callbacks
> after a plugin_exit call is already called (this can be easily verified by having a
> boolean set to true once plugin_exit is called and then g_assert this boolean is
> false in the callbacks)

We have mechanisms for safely unloading plugins during running so I
think we should be able to do something cleanly here. I'll cook up an
RFC.

>
>  > This is mitigated in this commit by synchronizing the call to
>  > `plugin_exit` through locking to ensure execlusive access to data
>  > structures, and NULL-ifying those data structures so that subsequent
>  > callbacks can check whether the data strucutres are freed, and if so,
>  > immediately exit.
>  >
>  > It's okay to immediately exit and don't account for those callbacks
>  > since they won't be accounted for anyway since `plugin_exit` is already
>  > called once and reported the statistics.
>  >
>  > Signed-off-by: Mahmoud Mandour <ma.mandourr@gmail.com>
>  > ---
>  >  contrib/plugins/cache.c | 31 ++++++++++++++++++++++++++++++-
>  >  1 file changed, 30 insertions(+), 1 deletion(-)
>  >
>  > diff --git a/contrib/plugins/cache.c b/contrib/plugins/cache.c
>  > index 695fb969dc..a452aba01c 100644
>  > --- a/contrib/plugins/cache.c
>  > +++ b/contrib/plugins/cache.c
>  > @@ -363,6 +363,11 @@ static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
>  >      effective_addr = hwaddr ? qemu_plugin_hwaddr_phys_addr(hwaddr) : vaddr;
>  > 
>  >      g_mutex_lock(&mtx);
>  > +    if (dcache == NULL) {
>  > +        g_mutex_unlock(&mtx);
>  > +        return;
>  > +    }
>  > +
>  >      if (!access_cache(dcache, effective_addr)) {
>  >          insn = (InsnData *) userdata;
>  >          insn->dmisses++;
>  > @@ -380,6 +385,11 @@ static void vcpu_insn_exec(unsigned int vcpu_index, void *userdata)
>  >      g_mutex_lock(&mtx);
>  >      insn_addr = ((InsnData *) userdata)->addr;
>  > 
>  > +    if (icache == NULL) {
>  > +        g_mutex_unlock(&mtx);
>  > +        return;
>  > +    }
>  > +
>  >      if (!access_cache(icache, insn_addr)) {
>  >          insn = (InsnData *) userdata;
>  >          insn->imisses++;
>  > @@ -406,12 +416,24 @@ static void vcpu_tb_trans(qemu_plugin_id_t id, struct qemu_plugin_tb *tb)
>  >              effective_addr = (uint64_t) qemu_plugin_insn_vaddr(insn);
>  >          }
>  > 
>  > +        g_mutex_lock(&mtx);
>  > +
>  > +        /*
>  > +         * is the plugin_exit callback called? If so, any further callback
>  > +         * registration is useless as it won't get accounted for after calling
>  > +         * plugin_exit once already, and also will use miss_ht after it's freed
>  > +         */
>  > +        if (miss_ht == NULL) {
>  > +            g_mutex_unlock(&mtx);
>  > +            return;
>  > +        }
>  > +
>  >          /*
>  >           * Instructions might get translated multiple times, we do not create
>  >           * new entries for those instructions. Instead, we fetch the same
>  >           * entry from the hash table and register it for the callback again.
>  >           */
>  > -        g_mutex_lock(&mtx);
>  > +
>  >          data = g_hash_table_lookup(miss_ht, GUINT_TO_POINTER(effective_addr));
>  >          if (data == NULL) {
>  >              data = g_new0(InsnData, 1);
>  > @@ -527,13 +549,20 @@ static void log_top_insns()
>  > 
>  >  static void plugin_exit(qemu_plugin_id_t id, void *p)
>  >  {
>  > +    g_mutex_lock(&mtx);
>  >      log_stats();
>  >      log_top_insns();
>  > 
>  >      cache_free(dcache);
>  > +    dcache = NULL;
>  > +
>  >      cache_free(icache);
>  > +    icache = NULL;
>  > 
>  >      g_hash_table_destroy(miss_ht);
>  > +    miss_ht = NULL;
>  > +
>  > +    g_mutex_unlock(&mtx);
>  >  }
>  > 
>  >  static void policy_init()
>
>  --
>  Alex Bennée


--
Alex Bennée