Re: [PATCH v4] debugobjects: scale the static pool size
From: Qian Cai
Date: Sun Nov 25 2018 - 23:54:30 EST
On 11/25/18 8:31 PM, Waiman Long wrote:
On 11/25/2018 03:42 PM, Qian Cai wrote:
On 11/23/18 10:01 PM, Qian Cai wrote:
On Nov 22, 2018, at 4:56 PM, Thomas Gleixner <tglx@xxxxxxxxxxxxx>
wrote:
On Tue, 20 Nov 2018, Qian Cai wrote:
Looking deeper at that.
diff --git a/lib/debugobjects.c b/lib/debugobjects.c
index 70935ed91125..140571aa483c 100644
--- a/lib/debugobjects.c
+++ b/lib/debugobjects.c
@@ -23,9 +23,81 @@
#define ODEBUG_HASH_BITSÂÂÂ 14
#define ODEBUG_HASH_SIZEÂÂÂ (1 << ODEBUG_HASH_BITS)
-#define ODEBUG_POOL_SIZEÂÂÂ 1024
+#define ODEBUG_DEFAULT_POOLÂÂÂ 512
#define ODEBUG_POOL_MIN_LEVELÂÂÂ 256
+/*
+ * Some debug objects are allocated during the early boot.
Enabling some options
+ * like timers or workqueue objects may increase the size required
significantly
+ * with large number of CPUs. For example (as today, 20 Nov. 2018),
+ *
+ * No. CPUs x 2 (worker pool) objects:
+ *
+ * start_kernel
+ *ÂÂ workqueue_init_early
+ *ÂÂÂÂ init_worker_pool
+ *ÂÂÂÂÂÂ init_timer_key
+ *ÂÂÂÂÂÂÂÂ debug_object_init
+ *
+ * No. CPUs objects (CONFIG_HIGH_RES_TIMERS):
+ *
+ * sched_init
+ *ÂÂ hrtick_rq_init
+ *ÂÂÂÂ hrtimer_init
+ *
+ * CONFIG_DEBUG_OBJECTS_WORK:
+ * No. CPUs x 6 (workqueue) objects:
+ *
+ * workqueue_init_early
+ *ÂÂ alloc_workqueue
+ *ÂÂÂÂ __alloc_workqueue_key
+ *ÂÂÂÂÂÂ alloc_and_link_pwqs
+ *ÂÂÂÂÂÂÂÂ init_pwq
+ *
+ * Also, plus No. CPUs objects:
+ *
+ * perf_event_init
+ *ÂÂÂ __init_srcu_struct
+ *ÂÂÂÂÂ init_srcu_struct_fields
+ *ÂÂÂÂÂÂÂ init_srcu_struct_nodes
+ *ÂÂÂÂÂÂÂÂÂ __init_work
None of the things are actually used or required _BEFORE_
debug_objects_mem_init() is invoked.
The reason why the call is at this place in start_kernel() is
historical. It's because back in the days when debugobjects were
added the
memory allocator was enabled way later than today. So we can just
move the
debug_objects_mem_init() call right before sched_init() I think.
Well, now that kmemleak_init() seems complains that
debug_objects_mem_init()
is called before it.
[ÂÂÂ 0.078805] kmemleak: Cannot insert 0xc000000dff930000 into the
object search tree (overlaps existing)
[ÂÂÂ 0.078860] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.20.0-rc3+ #3
[ÂÂÂ 0.078883] Call Trace:
[ÂÂÂ 0.078904] [c000000001c8fcd0] [c000000000c96b34]
dump_stack+0xe8/0x164 (unreliable)
[ÂÂÂ 0.078935] [c000000001c8fd20] [c000000000486e84]
create_object+0x344/0x380
[ÂÂÂ 0.078962] [c000000001c8fde0] [c000000000489544]
early_alloc+0x108/0x1f8
[ÂÂÂ 0.078989] [c000000001c8fe20] [c00000000109738c]
kmemleak_init+0x1d8/0x3d4
[ÂÂÂ 0.079016] [c000000001c8ff00] [c000000001054028]
start_kernel+0x5c0/0x6f8
[ÂÂÂ 0.079043] [c000000001c8ff90] [c00000000000ae7c]
start_here_common+0x1c/0x520
[ÂÂÂ 0.079070] kmemleak: Kernel memory leak detector disabled
[ÂÂÂ 0.079091] kmemleak: Object 0xc000000ffd587b68 (size 40):
[ÂÂÂ 0.079112] kmemleak:ÂÂ comm "swapper/0", pid 0, jiffies 4294937299
[ÂÂÂ 0.079135] kmemleak:ÂÂ min_count = -1
[ÂÂÂ 0.079153] kmemleak:ÂÂ count = 0
[ÂÂÂ 0.079170] kmemleak:ÂÂ flags = 0x5
[ÂÂÂ 0.079188] kmemleak:ÂÂ checksum = 0
[ÂÂÂ 0.079206] kmemleak:ÂÂ backtrace:
[ÂÂÂ 0.079227]ÂÂÂÂÂ __debug_object_init+0x688/0x700
[ÂÂÂ 0.079250]ÂÂÂÂÂ debug_object_activate+0x1e0/0x350
[ÂÂÂ 0.079272]ÂÂÂÂÂ __call_rcu+0x60/0x430
[ÂÂÂ 0.079292]ÂÂÂÂÂ put_object+0x60/0x80
[ÂÂÂ 0.079311]ÂÂÂÂÂ kmemleak_init+0x2cc/0x3d4
[ÂÂÂ 0.079331]ÂÂÂÂÂ start_kernel+0x5c0/0x6f8
[ÂÂÂ 0.079351]ÂÂÂÂÂ start_here_common+0x1c/0x520
[ÂÂÂ 0.079380] kmemleak: Early log backtrace:
[ÂÂÂ 0.079399]ÂÂÂ memblock_alloc_try_nid_raw+0x90/0xcc
[ÂÂÂ 0.079421]ÂÂÂ sparse_init_nid+0x144/0x51c
[ÂÂÂ 0.079440]ÂÂÂ sparse_init+0x1a0/0x238
[ÂÂÂ 0.079459]ÂÂÂ initmem_init+0x1d8/0x25c
[ÂÂÂ 0.079498]ÂÂÂ setup_arch+0x3e0/0x464
[ÂÂÂ 0.079517]ÂÂÂ start_kernel+0xa4/0x6f8
[ÂÂÂ 0.079536]ÂÂÂ start_here_common+0x1c/0x520
So this is an chicken-egg problem. Debug objects need kmemleak_init()
first, so it can make use of kmemleak_ignore() for all debug objects
in order to avoid the overlapping like the above.
while (obj_pool_free < debug_objects_pool_min_level) {
ÂÂÂÂnew = kmem_cache_zalloc(obj_cache, gfp);
ÂÂÂÂif (!new)
ÂÂÂÂÂÂÂ return;
ÂÂÂÂkmemleak_ignore(new);
However, there seems no way to move kmemleak_init() together this
early in start_kernel() just before vmalloc_init() [1] because it
looks like it depends on things like workqueue
(schedule_work(&cleanup_work)) and rcu. Hence, it needs to be after
workqueue_init_early() and rcu_init()
Given that, maybe the best outcome is to stick to the alternative
approach that works [1] rather messing up with the order of
debug_objects_mem_init() in start_kernel() which seems tricky. What do
you think?
[1] https://goo.gl/18N78g
[2] https://goo.gl/My6ig6
Could you move kmemleak_init() and debug_objects_mem_init() as far up as
possible, like before the hrtimer_init() to at least make static count
calculation as simple as possible?
Well, there is only 2 x NR_CPUS difference after moved both calls just after
rcu_init().
Before After
64-CPU: 1114 974
160-CPU: 2774 2429
256-CPU: 3853 4378
I suppose it is possible that the timers only need the scale factor 5 instead of
10. However, it needs to be retested for all the configurations to be sure, and
likely need to remove all irqs calls in kmemleak_init() and subroutines because
it is now called with irq disabled. Given the initdata will be freed anyway,
does it really worth doing?
BTW, calling debug_objects_mem_init() before kmemleak_init() actually could
trigger a loop on machines with 160+ CPUs until the pool is filled up,
debug_objects_pool_min_level += num_possible_cpus() * 4;
[1] while (obj_pool_free < debug_objects_pool_min_level)
kmemleak_init
kmemleak_ignore (from replaced static debug objects)
make_black_object
put_object
__call_rcu (kernel/rcu/tree.c)
debug_rcu_head_queue
debug_object_activate
debug_object_init
fill_pool
kmemleak_ignore (looping in [1])
make_black_object
...
I think until this is resolved, there is no way to move debug_objects_mem_init()
before kmemleak_init().