On Tue, Aug 20, 2019 at 08:27:05PM -0400, Joel Fernandes wrote:
[snip]
> > > > Or is the idea to time the kfree_rcu() loop separately? (I don't see
> > > > any such separate timing, though.)
> > >
> > > The kmalloc() times are included within the kfree loop. The timing of
> > > kfree_rcu() is not separate in my patch.
> >
> > You lost me on this one. What happens when you just interleave the
> > kmalloc() and kfree_rcu(), without looping, compared to the looping
> > above? Does this get more expensive? Cheaper? More vulnerable to OOM?
> > Something else?
>
> You mean pairing a single kmalloc() with a single kfree_rcu() and doing this
> several times? The results are very similar to doing kfree_alloc_num
> kmalloc()s, then do kfree_alloc_num kfree_rcu()s; and repeat the whole thing
> kfree_loops times (as done by this rcuperf patch we are reviewing).
>
> Following are some numbers. One change is the case where we are not at all
> batching does seem to complete even faster when we fully interleave kmalloc()
> with kfree() while the case of batching in the same scenario completes at the
> same time as did the "not fully interleaved" scenario. However, the grace
> period reduction improvements and the chances of OOM'ing are pretty much the
> same in either case.
[snip]
> Not fully interleaved: do kfree_alloc_num kmallocs, then do kfree_alloc_num kfree_rcu()s. And repeat this kfree_loops times.
> =======================
> (1) Batching
> rcuperf.kfree_loops=20000 rcuperf.kfree_alloc_num=8000 rcuperf.kfree_no_batch=0 rcuperf.kfree_rcu_test=1
>
> root@(none):/# free -m
> total used free shared buff/cache available
> Mem: 977 251 686 0 39 684
> Swap: 0 0 0
>
> [ 15.574402] Total time taken by all kfree'ers: 14185970787 ns, loops: 20000, batches: 1548
>
> (2) No Batching
> rcuperf.kfree_loops=20000 rcuperf.kfree_alloc_num=8000 rcuperf.kfree_no_batch=1 rcuperf.kfree_rcu_test=1
>
> root@(none):/# free -m
> total used free shared buff/cache available
> Mem: 977 82 855 0 39 853
> Swap: 0 0 0
>
> [ 13.724554] Total time taken by all kfree'ers: 12246217291 ns, loops: 20000, batches: 7262
And the diff for changing the test to do this case is as follows (I don't
plan to fold this diff in, since I feel the existing test suffices and
results are similar):
diff --git a/kernel/rcu/rcuperf.c b/kernel/rcu/rcuperf.c
index 46f9c4449348..e4e4be4aaf51 100644
--- a/kernel/rcu/rcuperf.c
+++ b/kernel/rcu/rcuperf.c
@@ -618,18 +618,13 @@ kfree_perf_thread(void *arg)
{
int i, loop = 0;
long me = (long)arg;
- struct kfree_obj **alloc_ptrs;
+ struct kfree_obj *alloc_ptr;
u64 start_time, end_time;
VERBOSE_PERFOUT_STRING("kfree_perf_thread task started");
set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
set_user_nice(current, MAX_NICE);
- alloc_ptrs = (struct kfree_obj **)kmalloc(sizeof(struct kfree_obj *) * kfree_alloc_num,
- GFP_KERNEL);
- if (!alloc_ptrs)
- return -ENOMEM;
-
start_time = ktime_get_mono_fast_ns();
if (atomic_inc_return(&n_kfree_perf_thread_started) >= kfree_nrealthreads) {
@@ -646,19 +641,17 @@ kfree_perf_thread(void *arg)
*/
do {
for (i = 0; i < kfree_alloc_num; i++) {
- alloc_ptrs[i] = kmalloc(sizeof(struct kfree_obj), GFP_KERNEL);
- if (!alloc_ptrs[i])
+ alloc_ptr = kmalloc(sizeof(struct kfree_obj), GFP_KERNEL);
+ if (!alloc_ptr)
return -ENOMEM;
- }
- for (i = 0; i < kfree_alloc_num; i++) {
if (!kfree_no_batch) {
- kfree_rcu(alloc_ptrs[i], rh);
+ kfree_rcu(alloc_ptr, rh);
} else {
rcu_callback_t cb;
cb = (rcu_callback_t)(unsigned long)offsetof(struct kfree_obj, rh);
- kfree_call_rcu_nobatch(&(alloc_ptrs[i]->rh), cb);
+ kfree_call_rcu_nobatch(&(alloc_ptr->rh), cb);
}
}
@@ -682,7 +675,6 @@ kfree_perf_thread(void *arg)
}
}
- kfree(alloc_ptrs);
torture_kthread_stopping("kfree_perf_thread");
return 0;
}
