Re: [PATCH 4/8] scsi: libsas: trigger a new revalidation to discover the device

From: John Garry
Date: Fri Jun 01 2018 - 06:02:56 EST


On 01/06/2018 01:59, Jason Yan wrote:


On 2018/5/31 23:42, John Garry wrote:
On 29/05/2018 03:23, Jason Yan wrote:
Now if a new device replaced a old device, the sas address will change.
We unregister the old device and discover the new device in one
revalidation process. But after we deferred the sas_port_delete(), the
sas port is not deleted when we registering the new port and device.
This will make the sysfs complain of creating duplicate filename.

Fix this by doing the replacement in two steps. The first revalidation
only delete the old device and trigger a new revalidation. The second
revalidation discover the new device.

Signed-off-by: Jason Yan <yanaijie@xxxxxxxxxx>
CC: chenxiang <chenxiang66@xxxxxxxxxxxxx>
CC: John Garry <john.garry@xxxxxxxxxx>
CC: Johannes Thumshirn <jthumshirn@xxxxxxx>
CC: Ewan Milne <emilne@xxxxxxxxxx>
CC: Christoph Hellwig <hch@xxxxxx>
CC: Tomas Henzl <thenzl@xxxxxxxxxx>
CC: Dan Williams <dan.j.williams@xxxxxxxxx>
CC: Hannes Reinecke <hare@xxxxxxxx>
---
drivers/scsi/libsas/sas_expander.c | 11 ++++++++++-
1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/drivers/scsi/libsas/sas_expander.c
b/drivers/scsi/libsas/sas_expander.c
index 629c580d906b..25ad9ef54e6c 100644
--- a/drivers/scsi/libsas/sas_expander.c
+++ b/drivers/scsi/libsas/sas_expander.c
@@ -2013,6 +2013,8 @@ static int sas_rediscover_dev(struct
domain_device *dev, int phy_id, bool last)
{
struct expander_device *ex = &dev->ex_dev;
struct ex_phy *phy = &ex->ex_phy[phy_id];
+ struct asd_sas_port *port = dev->port;
+ struct asd_sas_phy *sas_phy;
enum sas_device_type type = SAS_PHY_UNUSED;
u8 sas_addr[8];
int res;
@@ -2060,7 +2062,14 @@ static int sas_rediscover_dev(struct
domain_device *dev, int phy_id, bool last)
SAS_ADDR(phy->attached_sas_addr));
sas_unregister_devs_sas_addr(dev, phy_id, last);

- return sas_discover_new(dev, phy_id);
+ /* force the next revalidation find this phy and bring it up */
+ phy->phy_change_count = -1;
+ ex->ex_change_count = -1;
+ sas_phy = container_of(port->phy_list.next, struct asd_sas_phy,
+ port_phy_el);
+ port->ha->notify_port_event(sas_phy, PORTE_BROADCAST_RCVD);
+

This is less than ideal: that is, restarting another discovery with this
artifical broadcast event. We do something similar when re-enabling
revalidation.


That will back to what we have discussed before. The sas port
adding/removing is delayed outside the disco_mutex. we can only do the
adding or removing once inside the disco_mutex.

Can we do all the event processing synchronised to the original event?


Actually bcast is a very special event, and what we do in revalidation
at one time is scanning all phy changes, which may include many bcast
events(especially before our first patchset), and the next
revalidations may have nothing to do.

So "do all the event processing synchronised to the original event" is
impossible actually. Maybe if the bcast can indicate which device
originated it, we will achieve this goal.

I mean that since libsas does disocovery/revalidation for all expander PHYS for a single event, than all discovery/revalidation should be synchronised with that same event. I don't mean that for a given expander PHY which originated a broadcast event, the revalidation/discovery for that PHY should be synchronised with that same event. Like you said, I don't think it's possible.

On another point, one of the reasons to synchronise event processing was so events are not lost and are processed in order. In principle, by chaining these bcast events we lose that, since other PHY events may be queued before we queue the new artificial bcast events.


But if you mean we shall do this device removing and rediscovering in
one revalidation if it is not a "flutter", I think we can wrap a new
function for sas_revalidate_domain(), such as:


while (need_to_revalidate_again)
need_to_revalidate_again = sas_revalidate_domain()

In this way the sas_port adding/removing is packed in one loop, we won't
have the annoyance of "duplicate filename" warning. What do you
think?

Something like that, where all the discovery/revalidation and related device + port processing is done before we complete the revalidation event processing. A single revalidation event may defer do device+port processing multiple times.


+ return 0;
}

/**




.



.