cpu softplug kernel hang

[Prarit Bhargava <prarit@redhat.com>]

The panic below is from an 2.6.32-based kernel, however, AFAICT the same
issue exists with the latest 2.6.35-rc3+ kernel.

I have diagnosed the issue as being identical to the issue that I fixed
with the Intel rngd
driver sometime ago:

http://marc.info/?l=linux-kernel&m=117275119001289&w=2

When doing the following,

while true; do
        for i in `seq 12 23`; do echo 0 >
/sys/devices/system/cpu/cpu$i/online; done
        sleep 5
        for i in `seq 12 23`; do echo 1 >
/sys/devices/system/cpu/cpu$i/online; done
        sleep 5
done

I see (with the nmi_watchdog enabled)

BUG: NMI Watchdog detected LOCKUP on CPU11, ip ffffffff81029e72, registers:
CPU 11 
Modules linked in: nfs lockd fscache nfs_acl auth_rpcgss autofs4 sunrpc
cpufreq_ondemand acpi_cpufreq freq_table ipv6 dm_mirror dm_region_hash dm_log
uinput sg serio_raw i2c_i801 iTCO_wdt iTCO_vendor_support ioatdma i7core_edac
edac_core shpchp igb dca ext4 mbcache jbd2 sr_mod cdrom sd_mod crc_t10dif ahci
pata_acpi ata_generic pata_jmicron radeon ttm drm_kms_helper drm i2c_algo_bit
i2c_core dm_mod [last unloaded: microcode]

Modules linked in: nfs lockd fscache nfs_acl auth_rpcgss autofs4 sunrpc
cpufreq_ondemand acpi_cpufreq freq_table ipv6 dm_mirror dm_region_hash dm_log
uinput sg serio_raw i2c_i801 iTCO_wdt iTCO_vendor_support ioatdma i7core_edac
edac_core shpchp igb dca ext4 mbcache jbd2 sr_mod cdrom sd_mod crc_t10dif ahci
pata_acpi ata_generic pata_jmicron radeon ttm drm_kms_helper drm i2c_algo_bit
i2c_core dm_mod [last unloaded: microcode]
Pid: 704, comm: kexec Not tainted 2.6.32 #1 X8DTN
RIP: 0010:[<ffffffff81029e72>]  [<ffffffff81029e72>] ipi_handler+0x32/0xa0
RSP: 0000:ffff8801474a3f58  EFLAGS: 00000046
RAX: 0000000000000000 RBX: ffff880337393ea8 RCX: ffff88013ae41580
RDX: 00000000ffffffff RSI: 0000000000000000 RDI: ffff880337393ea8
RBP: ffff8801474a3f68 R08: 0000000061c941a6 R09: 00000000578070b9
R10: 0000000080507210 R11: 0000000025410601 R12: 0000000000000086
R13: 00000000ffffffff R14: ffffffff817491d0 R15: 0000000090793245
FS:  00007fefd5f3d700(0000) GS:ffff8801474a0000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
CR2: 000000000040d000 CR3: 0000000316a8e000 CR4: 00000000000006e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Process kexec (pid: 704, threadinfo ffff880314caa000, task ffff880335a0b500)
Stack:
 ffff880147571f40 000000000000000b ffff8801474a3f98 ffffffff810a6d28
<0> 00000000aa149910 00000000f21570f0 000000000008b495 00000000521ebd53
<0> ffff8801474a3fa8 ffffffff8102ea57 ffff880314cabf80 ffffffff81013e53
Call Trace:
 <IRQ> 
 [<ffffffff810a6d28>] generic_smp_call_function_interrupt+0x78/0x130
 [<ffffffff8102ea57>] smp_call_function_interrupt+0x27/0x40
 [<ffffffff81013e53>] call_function_interrupt+0x13/0x20
 <EOI> 
Code: 0f 1f 44 00 00 48 89 fb 9c 58 0f 1f 44 00 00 49 89 c4 fa 66 0f 1f 44 00
00 f0 ff 0f 8b 47 04 85 c0 75 0f 66 0f 1f 44 00 00 f3 90 <8b> 43 04 85 c0 74 f7
8b 7b 18 83 ff ff 74 47 48 8b 05 08 25 a1 

Since this is a panic, I get traces from all other cpus.

CPU 14 is in _write_lock_irq
CPU  2 is in _read_lock
CPU  6 has called smp_call_function() with the ipi_handler to sync mtrr's on
the new cpu

The problem is that ipi_handler does this:

static void ipi_handler(void *info)
{
#ifdef CONFIG_SMP
        struct set_mtrr_data *data = info;
        unsigned long flags;

        local_irq_save(flags);

        atomic_dec(&data->count); << global value that each processor entering
ipi_handler decrements
        while (!atomic_read(&data->gate)) << 1 when data->count != 0
                cpu_relax();

So what happens is that CPU 2 is in _read_lock and has acquired a lock.  CPU 14
is waiting for the release of that lock with IRQs *off*.

CPU 6 launches smp_call_function, and CPU 2 answers and runs the ipi_handler()
and waits (as do all other processors).

CPU 14, however, does not see the IPI because it is waiting with interrupts off
for the lock that CPU 2 is holding.

Boom.  Deadlock.

I'm planning on fixing this by rewriting this code and using stop_machine_run()
to synchronize the cpus instead of doing it in the ipi_handler.  

Unless, of course, someone has a better suggestion.

P.