hulk inclusion
category: bugfix
bugzilla: 12843
CVE: NA

---------------------------

When I use fio test kernel in the following steps:
1.The sas controller mixes SAS/SATA disks
2.Use fio test all disks
3.Simultaneous enable/disable/link_reset/hard_reset PHY

it will hang in ata_port_wait_eh
Call trace:
 __switch_to+0xb4/0x1b8
 __schedule+0x1e8/0x718
 schedule+0x38/0x90
 ata_port_wait_eh+0x70/0xf8
 sas_ata_wait_eh+0x24/0x30 [libsas]
 transport_sas_phy_reset.isra.3+0x128/0x160 [libsas]
 phy_reset_work+0x20/0x30 [libsas]
 process_one_work+0x1e4/0x460
 worker_thread+0x40/0x450
 kthread+0x12c/0x130
 ret_from_fork+0x10/0x18

The key code process is like this:
scsi_dec_host_busy
	atomic_dec(&shost->host_busy);
	if (unlikely(scsi_host_in_recovery(shost))) {
		spin_lock_irqsave(shost->host_lock, flags);
		...
		scsi_eh_wakeup(shost)
		...
	}

scsi_schedule_eh
	spin_lock_irqsave(shost->host_lock, flags);
	if (scsi_host_set_state(shost, SHOST_RECOVERY) == 0 ||
	    scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY) == 0) {
		...
		scsi_eh_wakeup(shost);
	}

scsi_eh_wakeup
	if (scsi_host_busy(shost) == shost->host_failed)
		wake_up_process(shost->ehandler);

In scsi_dec_host_busy, host_busy & shost_state not in spinlock. Neither
function wakes up the SCSI error handler in the following timing:

CPU 0(call scsi_dec_host_busy)    CPU 1(call scsi_schedule_eh)
LOAD shost_state(!=recovery)
                                  scsi_host_set_state(SHOST_RECOVERY)
                                  scsi_eh_wakeup(host_busy != host_failed)
atomic_dec(&shost->host_busy);
if (scsi_host_in_recovery(shost))

Add a smp_mb between host_busy and shost_state.
Signed-off-by: Nzhengbin <zhengbin13@huawei.com>
[yan: backport from 5.0]
Signed-off-by: NJason Yan <yanaijie@huawei.com>
Reviewed-by: NMiao Xie <miaoxie@huawei.com>
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>