[ Upstream commit 23c3828a ]

With commit 09c2f95a ("scsi: mpt3sas: Swap I/O memory read value back
to cpu endianness"), 64bit writes in _base_writeq() were rewritten to use
__raw_writeq() instad of writeq().

This introduced a bug apparent on powerpc64 systems such as the Raptor
Talos II that causes the HBA to drop from the PCIe bus under heavy load and
being reinitialized after a couple of seconds.

It can easily be triggered on affacted systems by using something like

  fio --name=random-write --iodepth=4 --rw=randwrite --bs=4k --direct=0 \
    --size=128M --numjobs=64 --end_fsync=1
  fio --name=random-write --iodepth=4 --rw=randwrite --bs=64k --direct=0 \
    --size=128M --numjobs=64 --end_fsync=1

a couple of times. In my case I tested it on both a ZFS raidz2 and a btrfs
raid6 using LSI 9300-8i and 9400-8i controllers.

The fix consists in resembling the write ordering of writeq() by adding a
mandatory write memory barrier before device access and a compiler barrier
afterwards. The additional MMIO barrier is superfluous.
Signed-off-by: NStephan Günther <moepi@moepi.net>
Reported-by: NMatt Corallo <linux@bluematt.me>
Acked-by: NSreekanth Reddy <Sreekanth.Reddy@broadcom.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: NSasha Levin <sashal@kernel.org>
Signed-off-by: NYang Yingliang <yangyingliang@huawei.com>