2.6.24.2(-rt2?) sched_fair issue

[Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>]

Hi Ingo,

I got the following message and the kernel crashed when testing 2.6.24.2-rt2.

Unable to handle kernel NULL pointer dereference at 0000000000000128 RIP:
 [<ffffffff80229805>] pick_next_task_fair+0x2d/0x42
PGD 211db1067 PUD 211c1d067 PMD 0
Oops: 0000 [1] PREEMPT SMP
CPU 2
Modules linked in:
Pid: 898, comm: stress Not tainted 2.6.24.2-rt2 #1
RIP: 0010:[<ffffffff80229805>]  [<ffffffff80229805>] pick_next_task_fair+0x2d/0x42
RSP: 0018:ffff8101ac423948  EFLAGS: 00010046
RAX: 0000000000000000 RBX: 0000000000000000 RCX: ffff81022e562300
RDX: ffff8101ac4239b0 RSI: ffff81022e562300 RDI: ffff8100050196e0
RBP: ffff8101ac423958 R08: ffff810005015680 R09: ffff810005015800
R10: 0000000000000000 R11: 0000000000000001 R12: ffff810005011680
R13: 0000000000000001 R14: ffff810005019680 R15: 00000001002f7a43
FS:  00002ada5c132b00(0000) GS:ffff81022fc057c0(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
CR2: 0000000000000128 CR3: 0000000211dfc000 CR4: 00000000000006e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Process stress (pid: 898, threadinfo ffff8101ac422000, task ffff81022e562300)
Stack:  ffffffff802292e4 ffff810005015ac0 ffff8101ac4239e8 ffffffff804d9685
 ffff8101ac4239b0 ffffffff8022f2bd 00000002ac423998 ffff81022e562300
 ffff8101ac4239a8 ffff81022e562640 00000000000000ff ffffffff804db527
Call Trace:
 [<ffffffff802292e4>] put_prev_task_rt+0xd/0x18
 [<ffffffff804d9685>] __schedule+0x414/0x771
 [<ffffffff8022f2bd>] add_preempt_count+0x18/0xb2
 [<ffffffff804db527>] __spin_unlock+0x14/0x2e
 [<ffffffff804d9cdc>] schedule+0xdf/0xff
 [<ffffffff804da557>] rt_spin_lock_slowlock+0xf9/0x19e
 [<ffffffff804daed8>] __rt_spin_lock+0x6b/0x70
 [<ffffffff804daee6>] rt_spin_lock+0x9/0xb
 [<ffffffff8027b57f>] page_lock_anon_vma+0x2b/0x3b
 [<ffffffff8027c38c>] page_referenced+0x49/0xf5
 [<ffffffff8026f40b>] shrink_active_list+0x222/0x563
 [<ffffffff8022f2bd>] add_preempt_count+0x18/0xb2
 [<ffffffff802705d5>] shrink_zone+0xcc/0x10f
 [<ffffffff80271139>] try_to_free_pages+0x183/0x27a
 [<ffffffff8026b0da>] __alloc_pages+0x1fb/0x344
 [<ffffffff8027c5ca>] anon_vma_prepare+0x29/0xf9
 [<ffffffff80274fc8>] handle_mm_fault+0x251/0x700
 [<ffffffff8020c7e6>] retint_kernel+0x26/0x30
 [<ffffffff8021f522>] do_page_fault+0x315/0x6bf
 [<ffffffff804db527>] __spin_unlock+0x14/0x2e
 [<ffffffff8022f1af>] finish_task_switch+0x2b/0x90
 [<ffffffff804db7c9>] error_exit+0x0/0x51
Code: 48 8b bb 28 01 00 00 48 85 ff 75 dd 48 8d 43 b8 41 58 5b 5d

I've gotten a kernel dump at this time.
The backtrace says;
(gdb) bt
#0  pick_next_task_fair (rq=<value optimized out>) at kernel/sched_fair.c:680
#1  0xffffffff804d9685 in __schedule () at kernel/sched.c:3783
#2  0xffffffff804d9cdc in schedule () at kernel/sched.c:3914
#3  0xffffffff804da557 in rt_spin_lock_slowlock (lock=0xffff81022e8ca618) at kernel/rtmutex.c:735
#4  0xffffffff804daed8 in __rt_spin_lock (lock=0xffff81022e8ca618) at kernel/rtmutex.c:646
#5  0xffffffff804daee6 in rt_spin_lock (lock=0xffff8100050196e0) at kernel/rtmutex.c:799
#6  0xffffffff8027b57f in page_lock_anon_vma (page=<value optimized out>) at mm/rmap.c:172
#7  0xffffffff8027c38c in page_referenced (page=0xffff8100050196e0, is_locked=777396992) at mm/rmap.c:309
(snip)

and here is the kernel/sched_fair.c:680
static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
{
        return grp->my_q;
}

It seems that grp is NULL, so I think pick_next_entity() returns NULL.
I think pick_next_entity() could return NULL when first_fair(cfs_rq) is false.

static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
{
        struct sched_entity *se = NULL;

        if (first_fair(cfs_rq)) {
                se = __pick_next_entity(cfs_rq);
                set_next_entity(cfs_rq, se);
        }

        return se;
}

static struct task_struct *pick_next_task_fair(struct rq *rq)
{
        struct cfs_rq *cfs_rq = &rq->cfs;
        struct sched_entity *se;

        if (unlikely(!cfs_rq->nr_running))
                return NULL;

        do {
                se = pick_next_entity(cfs_rq);
                cfs_rq = group_cfs_rq(se);
        } while (cfs_rq);

        return task_of(se);
}

I'm not sure the actual scenario to cause this panic.
I also don't know how to reproduce it, I was running some benchmarks at
the week end and found it the Monday morning.
I've encountered it on 2.6.24.2-rt2, but 2.6.24.2 and 2.6.25-rc3 also have
the same code pick_next_task_fair().
However latest sched-devel git tree has a modified code.

I'm not so familiar the CFS, is this investigation correct?

Thanks,
Hiroshi Shimamoto