Re: [PATCH] x86, mce: Fix rcu splat in drain_mce_log_buffer()

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Wed, Jan 11, 2012 at 07:54:48PM +0530, Srivatsa S. Bhat wrote:
> While booting, the following message is seen:

Hmmm...  This code is interesting.  It looks to me that access to
each entry is gated by the ->finished field.  If so, there should be
no need for any kind of rcu_dereference() on mcelog.next -- ACCESS_ONCE()
should work just fine.

> [   21.665087] ===============================
> [   21.669439] [ INFO: suspicious RCU usage. ]
> [   21.673798] 3.2.0-0.0.0.28.36b5ec9-default #2 Not tainted
> [   21.681353] -------------------------------
> [   21.685864] arch/x86/kernel/cpu/mcheck/mce.c:194 suspicious rcu_dereference_index_check() usage!
> [   21.695013]
> [   21.695014] other info that might help us debug this:
> [   21.695016]
> [   21.703488]
> [   21.703489] rcu_scheduler_active = 1, debug_locks = 1
> [   21.710426] 3 locks held by modprobe/2139:
> [   21.714754]  #0:  (&__lockdep_no_validate__){......}, at: [<ffffffff8133afd3>] __driver_attach+0x53/0xa0
> [   21.725020]  #1:
> [   21.725323] ioatdma: Intel(R) QuickData Technology Driver 4.00
> [   21.733206]  (&__lockdep_no_validate__){......}, at: [<ffffffff8133afe1>] __driver_attach+0x61/0xa0
> [   21.743015]  #2:  (i7core_edac_lock){+.+.+.}, at: [<ffffffffa01cfa5f>] i7core_probe+0x1f/0x5c0 [i7core_edac]
> [   21.753708]
> [   21.753709] stack backtrace:
> [   21.758429] Pid: 2139, comm: modprobe Not tainted 3.2.0-0.0.0.28.36b5ec9-default #2
> [   21.768253] Call Trace:
> [   21.770838]  [<ffffffff810977cd>] lockdep_rcu_suspicious+0xcd/0x100
> [   21.777366]  [<ffffffff8101aa41>] drain_mcelog_buffer+0x191/0x1b0
> [   21.783715]  [<ffffffff8101aa78>] mce_register_decode_chain+0x18/0x20
> [   21.790430]  [<ffffffffa01cf8db>] i7core_register_mci+0x2fb/0x3e4 [i7core_edac]
> [   21.798003]  [<ffffffffa01cfb14>] i7core_probe+0xd4/0x5c0 [i7core_edac]
> [   21.804809]  [<ffffffff8129566b>] local_pci_probe+0x5b/0xe0
> [   21.810631]  [<ffffffff812957c9>] __pci_device_probe+0xd9/0xe0
> [   21.816650]  [<ffffffff813362e4>] ? get_device+0x14/0x20
> [   21.822178]  [<ffffffff81296916>] pci_device_probe+0x36/0x60
> [   21.828061]  [<ffffffff8133ac8a>] really_probe+0x7a/0x2b0
> [   21.833676]  [<ffffffff8133af23>] driver_probe_device+0x63/0xc0
> [   21.839868]  [<ffffffff8133b01b>] __driver_attach+0x9b/0xa0
> [   21.845718]  [<ffffffff8133af80>] ? driver_probe_device+0xc0/0xc0
> [   21.852027]  [<ffffffff81339168>] bus_for_each_dev+0x68/0x90
> [   21.857876]  [<ffffffff8133aa3c>] driver_attach+0x1c/0x20
> [   21.863462]  [<ffffffff8133a64d>] bus_add_driver+0x16d/0x2b0
> [   21.869377]  [<ffffffff8133b6dc>] driver_register+0x7c/0x160
> [   21.875220]  [<ffffffff81296bda>] __pci_register_driver+0x6a/0xf0
> [   21.881494]  [<ffffffffa01fe000>] ? 0xffffffffa01fdfff
> [   21.886846]  [<ffffffffa01fe047>] i7core_init+0x47/0x1000 [i7core_edac]
> [   21.893737]  [<ffffffff810001ce>] do_one_initcall+0x3e/0x180
> [   21.899670]  [<ffffffff810a9b95>] sys_init_module+0xc5/0x220
> [   21.905542]  [<ffffffff8149bc39>] system_call_fastpath+0x16/0x1b
> 
> Fix this by using rcu_access_index() instead of rcu_dereference_check_mce().
> This is similar to what was done in commit a4dd992 (rcu: create new
> rcu_access_index() and use in mce) for a similar case in mce_poll.
> Link to that discussion: http://thread.gmane.org/gmane.linux.kernel/1058460/
> 
> Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
> ---
> I am no MCE expert. Hence requesting a thorough review of the patch.
> 
>  arch/x86/kernel/cpu/mcheck/mce.c |    2 +-
>  1 files changed, 1 insertions(+), 1 deletions(-)
> 
> diff --git a/arch/x86/kernel/cpu/mcheck/mce.c b/arch/x86/kernel/cpu/mcheck/mce.c
> index f22a9f7..c191fec 100644
> --- a/arch/x86/kernel/cpu/mcheck/mce.c
> +++ b/arch/x86/kernel/cpu/mcheck/mce.c
> @@ -191,7 +191,7 @@ static void drain_mcelog_buffer(void)
>  {
>  	unsigned int next, i, prev = 0;
> 
> -	next = rcu_dereference_check_mce(mcelog.next);
> +	next = rcu_access_index(mcelog.next);

So this is not so much a use of RCU as it is a standard mailbox algorithm,
with the ->finished flag controlling access.  So I suggest ACCESS_ONCE()
for the accesses to mcelog.next.

							Thanx, Paul

>  	do {
>  		struct mce *m;
>