Patchwork [isci,v3,3/4] libsas: suspend / resume support

login
register
mail settings
Submitter Dan Williams
Date March 14, 2012, 7:14 a.m.
Message ID <20120314071403.7223.82124.stgit@dwillia2-linux.jf.intel.com>
Download mbox | patch
Permalink /patch/146564/
State Not Applicable
Delegated to: David Miller
Headers show

Comments

Dan Williams - March 14, 2012, 7:14 a.m.
libsas power management routines to suspend and recover the sas domain
based on a model where the lldd is allowed and expected to be
"forgetful".

sas_suspend_ha - disable event processing allowing the lldd to take down
                 links without concern for causing hotplug events.
                 Regardless of whether the lldd actually posts link down
                 messages libsas notifies the lldd that all
                 domain_devices are gone.

sas_prep_resume_ha - on the way back up before the lldd starts link
                     training clean out any spurious events that were
                     generated on the way down, and re-enable event
                     processing

sas_resume_ha - after the lldd has started and decided that all phys
		have posted link-up events this routine is called to let
		libsas start it's own timeout of any phys that did not
		resume.  After the timeout an lldd can cancel the
                phy teardown by posting a link-up event.

Storage for ex_change_count (u16) and phy_change_count (u8) are changed
to int so they can be set to -1 to indicate 'invalidated'.

There's a hack added to sas_destruct_devices to workaround a deadlock
when a device is removed across the suspend cycle.  sd_remove is called
under device_lock() and wants to async_synchronize_full(), while the
resume path is running in an async callback and is trying to grab
device_lock()... fun ensues.  So we guarantee that async resume actions
have been flushed before allowing new invocations of sd_remove.

Cc: Alan Stern <stern@rowland.harvard.edu>
Reviewed-by: Jacek Danecki <jacek.danecki@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
---
 drivers/scsi/libsas/sas_ata.c      |   86 ++++++++++++++++++++++++++++++
 drivers/scsi/libsas/sas_discover.c |   81 +++++++++++++++++++++++++----
 drivers/scsi/libsas/sas_dump.c     |    1 
 drivers/scsi/libsas/sas_event.c    |    4 +
 drivers/scsi/libsas/sas_init.c     |  102 ++++++++++++++++++++++++++++++++++++
 drivers/scsi/libsas/sas_internal.h |    1 
 drivers/scsi/libsas/sas_phy.c      |   21 +++++++
 drivers/scsi/libsas/sas_port.c     |   52 ++++++++++++++++++
 include/scsi/libsas.h              |   21 ++++++-
 include/scsi/sas_ata.h             |   10 ++++
 10 files changed, 360 insertions(+), 19 deletions(-)


--
To unsubscribe from this list: send the line "unsubscribe linux-ide" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Alan Stern - March 14, 2012, 2:21 p.m.
On Wed, 14 Mar 2012, Dan Williams wrote:

> libsas power management routines to suspend and recover the sas domain
> based on a model where the lldd is allowed and expected to be
> "forgetful".

What exactly does that mean?  What is there for the lldd to remember?  
Does it maintain some sort of state information about the devices 
attached to the links?

> sas_suspend_ha - disable event processing allowing the lldd to take down
>                  links without concern for causing hotplug events.
>                  Regardless of whether the lldd actually posts link down
>                  messages libsas notifies the lldd that all
>                  domain_devices are gone.
> 
> sas_prep_resume_ha - on the way back up before the lldd starts link
>                      training clean out any spurious events that were
>                      generated on the way down, and re-enable event
>                      processing
> 
> sas_resume_ha - after the lldd has started and decided that all phys
> 		have posted link-up events this routine is called to let
> 		libsas start it's own timeout of any phys that did not
> 		resume.  After the timeout an lldd can cancel the
>                 phy teardown by posting a link-up event.
> 
> Storage for ex_change_count (u16) and phy_change_count (u8) are changed
> to int so they can be set to -1 to indicate 'invalidated'.
> 
> There's a hack added to sas_destruct_devices to workaround a deadlock
> when a device is removed across the suspend cycle.  sd_remove is called
> under device_lock() and wants to async_synchronize_full(), while the
> resume path is running in an async callback and is trying to grab
> device_lock()... fun ensues.  So we guarantee that async resume actions
> have been flushed before allowing new invocations of sd_remove.

Ooh, that doesn't sound like a good way to handle it.  For one thing, 
it's possible for the resume routine to be called while the caller 
holds the device-lock for the device being resumed.  It would be best 
if the resume path defers removal of devices that have disappeared to a 
different thread.

Alan Stern

--
To unsubscribe from this list: send the line "unsubscribe linux-ide" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Dan Williams - March 14, 2012, 8:44 p.m.
On Wed, 2012-03-14 at 10:21 -0400, Alan Stern wrote:
> On Wed, 14 Mar 2012, Dan Williams wrote:
> 
> > libsas power management routines to suspend and recover the sas domain
> > based on a model where the lldd is allowed and expected to be
> > "forgetful".
> 
> What exactly does that mean?  What is there for the lldd to remember?  
> Does it maintain some sort of state information about the devices 
> attached to the links?

The lldd is responsible for reporting link-up events at power on.
Rather than require lldds to remember the state of the links across
suspend (and filter link-up messages for "suspended" links), this
implementation puts all the responsibility on libsas.  The goal being to
try to minimize the need for special case code in the lldd.  The lldd
simply redoes init and posts link-up events as normal.  Then at resume
time libsas reminds the lldd about the topology of devices that was
previously reachable via a given link.

> > sas_suspend_ha - disable event processing allowing the lldd to take down
> >                  links without concern for causing hotplug events.
> >                  Regardless of whether the lldd actually posts link down
> >                  messages libsas notifies the lldd that all
> >                  domain_devices are gone.
> > 
> > sas_prep_resume_ha - on the way back up before the lldd starts link
> >                      training clean out any spurious events that were
> >                      generated on the way down, and re-enable event
> >                      processing
> > 
> > sas_resume_ha - after the lldd has started and decided that all phys
> > 		have posted link-up events this routine is called to let
> > 		libsas start it's own timeout of any phys that did not
> > 		resume.  After the timeout an lldd can cancel the
> >                 phy teardown by posting a link-up event.
> > 
> > Storage for ex_change_count (u16) and phy_change_count (u8) are changed
> > to int so they can be set to -1 to indicate 'invalidated'.
> > 
> > There's a hack added to sas_destruct_devices to workaround a deadlock
> > when a device is removed across the suspend cycle.  sd_remove is called
> > under device_lock() and wants to async_synchronize_full(), while the
> > resume path is running in an async callback and is trying to grab
> > device_lock()... fun ensues.  So we guarantee that async resume actions
> > have been flushed before allowing new invocations of sd_remove.
> 
> Ooh, that doesn't sound like a good way to handle it.  For one thing, 
> it's possible for the resume routine to be called while the caller 
> holds the device-lock for the device being resumed.  It would be best 
> if the resume path defers removal of devices that have disappeared to a 
> different thread.

In fact, that's what the hack does (in Scsi_Host workqueue context), but
that's only part of the solution.  We also need to flush resume
operations otherwise we can trigger the following deadlock (that lockdep
can't catch thanks to lockdep_set_novalidate_class(&dev->mutex); and the
lack of lockdep annotation for async_sychronize_full()):

[  494.237079] INFO: task kworker/u:3:554 blocked for more than 120 seconds.
[  494.294396] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[  494.360809] kworker/u:3     D 0000000000000000     0   554      2 0x00000000
[  494.420739]  ffff88012e4d3af0 0000000000000046 ffff88013200c160 ffff88012e4d3fd8
[  494.484392]  ffff88012e4d3fd8 0000000000012500 ffff8801394ea0b0 ffff88013200c160
[  494.548038]  ffff88012e4d3ae0 00000000000001e3 ffffffff81a249e0 ffff8801321c5398
[  494.611685] Call Trace:
[  494.632649]  [<ffffffff8149dd25>] schedule+0x5a/0x5c
[  494.674687]  [<ffffffff8104b968>] async_synchronize_cookie_domain+0xb6/0x112
[  494.734177]  [<ffffffff810461ff>] ? __init_waitqueue_head+0x50/0x50
[  494.787134]  [<ffffffff8131a224>] ? scsi_remove_target+0x48/0x48
[  494.837900]  [<ffffffff8104b9d9>] async_synchronize_cookie+0x15/0x17
[  494.891567]  [<ffffffff8104ba49>] async_synchronize_full+0x54/0x70  <-- here we wait for async contexts to complete
[  494.943783]  [<ffffffff8104b9f5>] ? async_synchronize_full_domain+0x1a/0x1a
[  495.002547]  [<ffffffffa00114b1>] sd_remove+0x2c/0xa2 [sd_mod]
[  495.051861]  [<ffffffff812fe94f>] __device_release_driver+0x86/0xcf
[  495.104807]  [<ffffffff812fe9bd>] device_release_driver+0x25/0x32  <-- here we take device_lock()

[  853.511341] INFO: task kworker/u:4:549 blocked for more than 120 seconds.
[  853.568693] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[  853.635119] kworker/u:4     D ffff88013097b5d0     0   549      2 0x00000000
[  853.695129]  ffff880132773c40 0000000000000046 ffff880130790000 ffff880132773fd8
[  853.758990]  ffff880132773fd8 0000000000012500 ffff88013288a0b0 ffff880130790000
[  853.822796]  0000000000000246 0000000000000040 ffff88013097b5c8 ffff880130790000
[  853.886633] Call Trace:
[  853.907631]  [<ffffffff8149dd25>] schedule+0x5a/0x5c
[  853.949670]  [<ffffffff8149cc44>] __mutex_lock_common+0x220/0x351
[  854.001225]  [<ffffffff81304bd7>] ? device_resume+0x58/0x1c4
[  854.049082]  [<ffffffff81304bd7>] ? device_resume+0x58/0x1c4
[  854.097011]  [<ffffffff8149ce48>] mutex_lock_nested+0x2f/0x36   <-- here we wait for device_lock()
[  854.145591]  [<ffffffff81304bd7>] device_resume+0x58/0x1c4
[  854.192066]  [<ffffffff81304d61>] async_resume+0x1e/0x45
[  854.237019]  [<ffffffff8104bc93>] async_run_entry_fn+0xc6/0x173  <-- ...while running in async context

--
Dan

--
To unsubscribe from this list: send the line "unsubscribe linux-ide" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Alan Stern - March 14, 2012, 9:11 p.m.
On Wed, 14 Mar 2012, Dan Williams wrote:

> > Ooh, that doesn't sound like a good way to handle it.  For one thing, 
> > it's possible for the resume routine to be called while the caller 
> > holds the device-lock for the device being resumed.  It would be best 
> > if the resume path defers removal of devices that have disappeared to a 
> > different thread.
> 
> In fact, that's what the hack does (in Scsi_Host workqueue context), but
> that's only part of the solution.  We also need to flush resume
> operations otherwise we can trigger the following deadlock (that lockdep
> can't catch thanks to lockdep_set_novalidate_class(&dev->mutex); and the
> lack of lockdep annotation for async_sychronize_full()):

Unfortunately, lockdep isn't able to track dev->mutex accesses 
usefully.  It would be nice if it could.

> [  494.611685] Call Trace:
> [  494.632649]  [<ffffffff8149dd25>] schedule+0x5a/0x5c
> [  494.674687]  [<ffffffff8104b968>] async_synchronize_cookie_domain+0xb6/0x112
> [  494.734177]  [<ffffffff810461ff>] ? __init_waitqueue_head+0x50/0x50
> [  494.787134]  [<ffffffff8131a224>] ? scsi_remove_target+0x48/0x48
> [  494.837900]  [<ffffffff8104b9d9>] async_synchronize_cookie+0x15/0x17
> [  494.891567]  [<ffffffff8104ba49>] async_synchronize_full+0x54/0x70  <-- here we wait for async contexts to complete
> [  494.943783]  [<ffffffff8104b9f5>] ? async_synchronize_full_domain+0x1a/0x1a
> [  495.002547]  [<ffffffffa00114b1>] sd_remove+0x2c/0xa2 [sd_mod]
> [  495.051861]  [<ffffffff812fe94f>] __device_release_driver+0x86/0xcf
> [  495.104807]  [<ffffffff812fe9bd>] device_release_driver+0x25/0x32  <-- here we take device_lock()

> [  853.886633] Call Trace:
> [  853.907631]  [<ffffffff8149dd25>] schedule+0x5a/0x5c
> [  853.949670]  [<ffffffff8149cc44>] __mutex_lock_common+0x220/0x351
> [  854.001225]  [<ffffffff81304bd7>] ? device_resume+0x58/0x1c4
> [  854.049082]  [<ffffffff81304bd7>] ? device_resume+0x58/0x1c4
> [  854.097011]  [<ffffffff8149ce48>] mutex_lock_nested+0x2f/0x36   <-- here we wait for device_lock()
> [  854.145591]  [<ffffffff81304bd7>] device_resume+0x58/0x1c4
> [  854.192066]  [<ffffffff81304d61>] async_resume+0x1e/0x45
> [  854.237019]  [<ffffffff8104bc93>] async_run_entry_fn+0xc6/0x173  <-- ...while running in async context

I see.  It's a nasty situation; I guess the best way to describe it is 
a conflict between the requirements of the PM and SCSI subsystems:

	The PM core runs suspend and resume operations in async 
	threads, and these threads need to acquire the device lock;

	The sd driver needs to insure that async probing is finished
	when starting its remove routine, and it is called with the
	device lock held.

The best solution might be to use a workqueue for sd's async probing 
instead of the "async" threads.  Then the work routine could be 
cancelled without doing async_synchronize_full().

Alan Stern

--
To unsubscribe from this list: send the line "unsubscribe linux-ide" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/drivers/scsi/libsas/sas_ata.c b/drivers/scsi/libsas/sas_ata.c
index d30a093..b161bdd 100644
--- a/drivers/scsi/libsas/sas_ata.c
+++ b/drivers/scsi/libsas/sas_ata.c
@@ -677,6 +677,92 @@  void sas_probe_sata(struct asd_sas_port *port)
 		if (ata_dev_disabled(sas_to_ata_dev(dev)))
 			sas_fail_probe(dev, __func__, -ENODEV);
 	}
+
+}
+
+static bool sas_ata_flush_pm_eh(struct asd_sas_port *port, const char *func)
+{
+	struct domain_device *dev, *n;
+	bool retry = false;
+
+	list_for_each_entry_safe(dev, n, &port->dev_list, dev_list_node) {
+		int rc;
+
+		if (!dev_is_sata(dev))
+			continue;
+
+		sas_ata_wait_eh(dev);
+		rc = dev->sata_dev.pm_result;
+		if (rc == -EAGAIN)
+			retry = true;
+		else if (rc) {
+			/* since we don't have a
+			 * ->port_{suspend|resume} routine in our
+			 *  ata_port ops, and no entanglements with
+			 *  acpi, suspend should just be mechanical trip
+			 *  through eh, catch cases where these
+			 *  assumptions are invalidated
+			 */
+			WARN_ONCE(1, "failed %s %s error: %d\n", func,
+				 dev_name(&dev->rphy->dev), rc);
+		}
+
+		/* if libata failed to power manage the device, tear it down */
+		if (ata_dev_disabled(sas_to_ata_dev(dev)))
+			sas_fail_probe(dev, func, -ENODEV);
+	}
+
+	return retry;
+}
+
+void sas_suspend_sata(struct asd_sas_port *port)
+{
+	struct domain_device *dev;
+
+ retry:
+	mutex_lock(&port->ha->disco_mutex);
+	list_for_each_entry(dev, &port->dev_list, dev_list_node) {
+		struct sata_device *sata;
+
+		if (!dev_is_sata(dev))
+			continue;
+
+		sata = &dev->sata_dev;
+		if (sata->ap->pm_mesg.event == PM_EVENT_SUSPEND)
+			continue;
+
+		sata->pm_result = -EIO;
+		ata_sas_port_async_suspend(sata->ap, &sata->pm_result);
+	}
+	mutex_unlock(&port->ha->disco_mutex);
+
+	if (sas_ata_flush_pm_eh(port, __func__))
+		goto retry;
+}
+
+void sas_resume_sata(struct asd_sas_port *port)
+{
+	struct domain_device *dev;
+
+ retry:
+	mutex_lock(&port->ha->disco_mutex);
+	list_for_each_entry(dev, &port->dev_list, dev_list_node) {
+		struct sata_device *sata;
+
+		if (!dev_is_sata(dev))
+			continue;
+
+		sata = &dev->sata_dev;
+		if (sata->ap->pm_mesg.event == PM_EVENT_ON)
+			continue;
+
+		sata->pm_result = -EIO;
+		ata_sas_port_async_resume(sata->ap, &sata->pm_result);
+	}
+	mutex_unlock(&port->ha->disco_mutex);
+
+	if (sas_ata_flush_pm_eh(port, __func__))
+		goto retry;
 }
 
 /**
diff --git a/drivers/scsi/libsas/sas_discover.c b/drivers/scsi/libsas/sas_discover.c
index 2864b17..007917e 100644
--- a/drivers/scsi/libsas/sas_discover.c
+++ b/drivers/scsi/libsas/sas_discover.c
@@ -24,6 +24,7 @@ 
 
 #include <linux/scatterlist.h>
 #include <linux/slab.h>
+#include <linux/async.h>
 #include <scsi/scsi_host.h>
 #include <scsi/scsi_eh.h>
 #include "sas_internal.h"
@@ -171,16 +172,18 @@  int sas_notify_lldd_dev_found(struct domain_device *dev)
 	struct Scsi_Host *shost = sas_ha->core.shost;
 	struct sas_internal *i = to_sas_internal(shost->transportt);
 
-	if (i->dft->lldd_dev_found) {
-		res = i->dft->lldd_dev_found(dev);
-		if (res) {
-			printk("sas: driver on pcidev %s cannot handle "
-			       "device %llx, error:%d\n",
-			       dev_name(sas_ha->dev),
-			       SAS_ADDR(dev->sas_addr), res);
-		}
-		kref_get(&dev->kref);
+	if (!i->dft->lldd_dev_found)
+		return 0;
+
+	res = i->dft->lldd_dev_found(dev);
+	if (res) {
+		printk("sas: driver on pcidev %s cannot handle "
+		       "device %llx, error:%d\n",
+		       dev_name(sas_ha->dev),
+		       SAS_ADDR(dev->sas_addr), res);
 	}
+	set_bit(SAS_DEV_FOUND, &dev->state);
+	kref_get(&dev->kref);
 	return res;
 }
 
@@ -191,7 +194,10 @@  void sas_notify_lldd_dev_gone(struct domain_device *dev)
 	struct Scsi_Host *shost = sas_ha->core.shost;
 	struct sas_internal *i = to_sas_internal(shost->transportt);
 
-	if (i->dft->lldd_dev_gone) {
+	if (!i->dft->lldd_dev_gone)
+		return;
+
+	if (test_and_clear_bit(SAS_DEV_FOUND, &dev->state)) {
 		i->dft->lldd_dev_gone(dev);
 		sas_put_device(dev);
 	}
@@ -225,6 +231,47 @@  static void sas_probe_devices(struct work_struct *work)
 	}
 }
 
+static void sas_suspend_devices(struct work_struct *work)
+{
+	struct asd_sas_phy *phy;
+	struct domain_device *dev;
+	struct sas_discovery_event *ev = to_sas_discovery_event(work);
+	struct asd_sas_port *port = ev->port;
+	struct Scsi_Host *shost = port->ha->core.shost;
+	struct sas_internal *si = to_sas_internal(shost->transportt);
+
+	clear_bit(DISCE_SUSPEND, &port->disc.pending);
+
+	sas_suspend_sata(port);
+
+	/* lldd is free to forget the domain_device across the
+	 * suspension, we force the issue here to keep the reference
+	 * counts aligned
+	 */
+	list_for_each_entry(dev, &port->dev_list, dev_list_node)
+		sas_notify_lldd_dev_gone(dev);
+
+	/* we are suspending, so we know events are disabled and
+	 * phy_list is not being mutated
+	 */
+	list_for_each_entry(phy, &port->phy_list, port_phy_el) {
+		if (si->dft->lldd_port_formed)
+			si->dft->lldd_port_deformed(phy);
+		phy->suspended = 1;
+		port->suspended = 1;
+	}
+}
+
+static void sas_resume_devices(struct work_struct *work)
+{
+	struct sas_discovery_event *ev = to_sas_discovery_event(work);
+	struct asd_sas_port *port = ev->port;
+
+	clear_bit(DISCE_RESUME, &port->disc.pending);
+
+	sas_resume_sata(port);
+}
+
 /**
  * sas_discover_end_dev -- discover an end device (SSP, etc)
  * @end: pointer to domain device of interest
@@ -302,6 +349,18 @@  static void sas_destruct_devices(struct work_struct *work)
 
 	clear_bit(DISCE_DESTRUCT, &port->disc.pending);
 
+	/* removal wants to synchronize async actions, so we need to
+	 * prevent removals during suspend
+	 */
+	if (test_bit(SAS_HA_SUSPENDED, &port->ha->state))
+		return;
+	/* FIXME: force resume to complete before we start removing
+	 * devices, critically before we take device_lock() and call
+	 * async_synchronize_full in sd_remove (in all its
+	 * lockdep_novalidate glory, sigh)
+	 */
+	async_synchronize_full();
+
 	list_for_each_entry_safe(dev, n, &port->destroy_list, disco_list_node) {
 		list_del_init(&dev->disco_list_node);
 
@@ -521,6 +580,8 @@  void sas_init_disc(struct sas_discovery *disc, struct asd_sas_port *port)
 		[DISCE_DISCOVER_DOMAIN] = sas_discover_domain,
 		[DISCE_REVALIDATE_DOMAIN] = sas_revalidate_domain,
 		[DISCE_PROBE] = sas_probe_devices,
+		[DISCE_SUSPEND] = sas_suspend_devices,
+		[DISCE_RESUME] = sas_resume_devices,
 		[DISCE_DESTRUCT] = sas_destruct_devices,
 	};
 
diff --git a/drivers/scsi/libsas/sas_dump.c b/drivers/scsi/libsas/sas_dump.c
index fc46093..cd6f99c 100644
--- a/drivers/scsi/libsas/sas_dump.c
+++ b/drivers/scsi/libsas/sas_dump.c
@@ -41,6 +41,7 @@  static const char *sas_phye_str[] = {
 	[1] = "PHYE_OOB_DONE",
 	[2] = "PHYE_OOB_ERROR",
 	[3] = "PHYE_SPINUP_HOLD",
+	[4] = "PHYE_RESUME_TIMEOUT",
 };
 
 void sas_dprint_porte(int phyid, enum port_event pe)
diff --git a/drivers/scsi/libsas/sas_event.c b/drivers/scsi/libsas/sas_event.c
index 789c4d8..aadbd53 100644
--- a/drivers/scsi/libsas/sas_event.c
+++ b/drivers/scsi/libsas/sas_event.c
@@ -134,7 +134,7 @@  static void notify_port_event(struct asd_sas_phy *phy, enum port_event event)
 			&phy->port_events[event].work, ha);
 }
 
-static void notify_phy_event(struct asd_sas_phy *phy, enum phy_event event)
+void sas_notify_phy_event(struct asd_sas_phy *phy, enum phy_event event)
 {
 	struct sas_ha_struct *ha = phy->ha;
 
@@ -159,7 +159,7 @@  int sas_init_events(struct sas_ha_struct *sas_ha)
 
 	sas_ha->notify_ha_event = notify_ha_event;
 	sas_ha->notify_port_event = notify_port_event;
-	sas_ha->notify_phy_event = notify_phy_event;
+	sas_ha->notify_phy_event = sas_notify_phy_event;
 
 	return 0;
 }
diff --git a/drivers/scsi/libsas/sas_init.c b/drivers/scsi/libsas/sas_init.c
index 1bbab3d..30dbcbb 100644
--- a/drivers/scsi/libsas/sas_init.c
+++ b/drivers/scsi/libsas/sas_init.c
@@ -159,7 +159,7 @@  Undo_phys:
 	return error;
 }
 
-int sas_unregister_ha(struct sas_ha_struct *sas_ha)
+static void sas_disable_events(struct sas_ha_struct *sas_ha)
 {
 	/* Set the state to unregistered to avoid further unchained
 	 * events to be queued, and flush any in-progress drainers
@@ -170,7 +170,11 @@  int sas_unregister_ha(struct sas_ha_struct *sas_ha)
 	spin_unlock_irq(&sas_ha->lock);
 	__sas_drain_work(sas_ha);
 	mutex_unlock(&sas_ha->drain_mutex);
+}
 
+int sas_unregister_ha(struct sas_ha_struct *sas_ha)
+{
+	sas_disable_events(sas_ha);
 	sas_unregister_ports(sas_ha);
 
 	/* flush unregistration work */
@@ -362,6 +366,102 @@  int sas_set_phy_speed(struct sas_phy *phy,
 	return ret;
 }
 
+void sas_prep_resume_ha(struct sas_ha_struct *ha)
+{
+	int i;
+
+	set_bit(SAS_HA_REGISTERED, &ha->state);
+
+	/* clear out any stale link events/data from the suspension path */
+	for (i = 0; i < ha->num_phys; i++) {
+		struct asd_sas_phy *phy = ha->sas_phy[i];
+
+		memset(phy->attached_sas_addr, 0, SAS_ADDR_SIZE);
+		phy->port_events_pending = 0;
+		phy->phy_events_pending = 0;
+		phy->frame_rcvd_size = 0;
+	}
+}
+EXPORT_SYMBOL(sas_prep_resume_ha);
+
+static int phys_suspended(struct sas_ha_struct *ha)
+{
+	int i, rc = 0;
+
+	for (i = 0; i < ha->num_phys; i++) {
+		struct asd_sas_phy *phy = ha->sas_phy[i];
+
+		if (phy->suspended)
+			rc++;
+	}
+
+	return rc;
+}
+
+void sas_resume_ha(struct sas_ha_struct *ha)
+{
+	const unsigned long tmo = msecs_to_jiffies(25000);
+	int i;
+
+	/* deform ports on phys that did not resume
+	 * at this point we may be racing the phy coming back (as posted
+	 * by the lldd).  So we post the event and once we are in the
+	 * libsas context check that the phy remains suspended before
+	 * tearing it down.
+	 */
+	i = phys_suspended(ha);
+	if (i)
+		dev_info(ha->dev, "waiting up to 25 seconds for %d phy%s to resume\n",
+			 i, i > 1 ? "s" : "");
+	wait_event_timeout(ha->eh_wait_q, phys_suspended(ha) == 0, tmo);
+	for (i = 0; i < ha->num_phys; i++) {
+		struct asd_sas_phy *phy = ha->sas_phy[i];
+
+		if (phy->suspended) {
+			dev_warn(&phy->phy->dev, "resume timeout\n");
+			sas_notify_phy_event(phy, PHYE_RESUME_TIMEOUT);
+		}
+	}
+
+	/* all phys are back up or timed out, turn on i/o so we can
+	 * flush out disks that did not return
+	 */
+	scsi_unblock_requests(ha->core.shost);
+	sas_drain_work(ha);
+
+	/* post the disk removals asynchronously (no drain) since sd_remove
+	 * wants to call async_synchronize_full, and we may be in async
+	 * context for suspend (i.e. deadlock)
+	 */
+	clear_bit(SAS_HA_SUSPENDED, &ha->state);
+	for (i = 0; i < ha->num_phys; i++) {
+		struct asd_sas_port *port = ha->sas_port[i];
+
+		sas_discover_event(port, DISCE_DESTRUCT);
+	}
+}
+EXPORT_SYMBOL(sas_resume_ha);
+
+void sas_suspend_ha(struct sas_ha_struct *ha)
+{
+	int i;
+
+	sas_disable_events(ha);
+	set_bit(SAS_HA_SUSPENDED, &ha->state);
+	scsi_block_requests(ha->core.shost);
+	for (i = 0; i < ha->num_phys; i++) {
+		struct asd_sas_port *port = ha->sas_port[i];
+
+		sas_discover_event(port, DISCE_SUSPEND);
+	}
+
+	/* flush suspend events while unregistered */
+	mutex_lock(&ha->drain_mutex);
+	__sas_drain_work(ha);
+	mutex_unlock(&ha->drain_mutex);
+}
+EXPORT_SYMBOL(sas_suspend_ha);
+
 static void sas_phy_release(struct sas_phy *phy)
 {
 	kfree(phy->hostdata);
diff --git a/drivers/scsi/libsas/sas_internal.h b/drivers/scsi/libsas/sas_internal.h
index 507e4cf..1de6796 100644
--- a/drivers/scsi/libsas/sas_internal.h
+++ b/drivers/scsi/libsas/sas_internal.h
@@ -89,6 +89,7 @@  int sas_smp_phy_control(struct domain_device *dev, int phy_id,
 			enum phy_func phy_func, struct sas_phy_linkrates *);
 int sas_smp_get_phy_events(struct sas_phy *phy);
 
+void sas_notify_phy_event(struct asd_sas_phy *phy, enum phy_event event);
 void sas_device_set_phy(struct domain_device *dev, struct sas_port *port);
 struct domain_device *sas_find_dev_by_rphy(struct sas_rphy *rphy);
 struct domain_device *sas_ex_to_ata(struct domain_device *ex_dev, int phy_id);
diff --git a/drivers/scsi/libsas/sas_phy.c b/drivers/scsi/libsas/sas_phy.c
index 521422e..cdee446 100644
--- a/drivers/scsi/libsas/sas_phy.c
+++ b/drivers/scsi/libsas/sas_phy.c
@@ -94,6 +94,25 @@  static void sas_phye_spinup_hold(struct work_struct *work)
 	i->dft->lldd_control_phy(phy, PHY_FUNC_RELEASE_SPINUP_HOLD, NULL);
 }
 
+static void sas_phye_resume_timeout(struct work_struct *work)
+{
+	struct asd_sas_event *ev = to_asd_sas_event(work);
+	struct asd_sas_phy *phy = ev->phy;
+
+	clear_bit(PHYE_RESUME_TIMEOUT, &phy->phy_events_pending);
+
+	/* phew, lldd got the phy back in the nick of time */
+	if (!phy->suspended) {
+		dev_info(&phy->phy->dev, "resume timeout cancelled\n");
+		return;
+	}
+
+	phy->error = 0;
+	phy->suspended = 0;
+	sas_deform_port(phy, 1);
+}
+
+
 /* ---------- Phy class registration ---------- */
 
 int sas_register_phys(struct sas_ha_struct *sas_ha)
@@ -105,6 +124,8 @@  int sas_register_phys(struct sas_ha_struct *sas_ha)
 		[PHYE_OOB_DONE] = sas_phye_oob_done,
 		[PHYE_OOB_ERROR] = sas_phye_oob_error,
 		[PHYE_SPINUP_HOLD] = sas_phye_spinup_hold,
+		[PHYE_RESUME_TIMEOUT] = sas_phye_resume_timeout,
+
 	};
 
 	static const work_func_t sas_port_event_fns[PORT_NUM_EVENTS] = {
diff --git a/drivers/scsi/libsas/sas_port.c b/drivers/scsi/libsas/sas_port.c
index 1cf7d75..1ac0c8b 100644
--- a/drivers/scsi/libsas/sas_port.c
+++ b/drivers/scsi/libsas/sas_port.c
@@ -39,6 +39,49 @@  static bool phy_is_wideport_member(struct asd_sas_port *port, struct asd_sas_phy
 	return true;
 }
 
+static void sas_resume_port(struct asd_sas_phy *phy)
+{
+	struct domain_device *dev;
+	struct asd_sas_port *port = phy->port;
+	struct sas_ha_struct *sas_ha = phy->ha;
+	struct sas_internal *si = to_sas_internal(sas_ha->core.shost->transportt);
+
+	if (si->dft->lldd_port_formed)
+		si->dft->lldd_port_formed(phy);
+
+	if (port->suspended)
+		port->suspended = 0;
+	else {
+		/* we only need to handle "link returned" actions once */
+		return;
+	}
+
+	/* if the port came back:
+	 * 1/ presume every device came back
+	 * 2/ force the next revalidation to check all expander phys
+	 */
+	list_for_each_entry(dev, &port->dev_list, dev_list_node) {
+		int i, rc;
+
+		rc = sas_notify_lldd_dev_found(dev);
+		if (rc) {
+			sas_unregister_dev(port, dev);
+			continue;
+		}
+
+		if (dev->dev_type == EDGE_DEV || dev->dev_type == FANOUT_DEV) {
+			dev->ex_dev.ex_change_count = -1;
+			for (i = 0; i < dev->ex_dev.num_phys; i++) {
+				struct ex_phy *phy = &dev->ex_dev.ex_phy[i];
+
+				phy->phy_change_count = -1;
+			}
+		}
+	}
+
+	sas_discover_event(port, DISCE_RESUME);
+}
+
 /**
  * sas_form_port -- add this phy to a port
  * @phy: the phy of interest
@@ -58,7 +101,14 @@  static void sas_form_port(struct asd_sas_phy *phy)
 	if (port) {
 		if (!phy_is_wideport_member(port, phy))
 			sas_deform_port(phy, 0);
-		else {
+		else if (phy->suspended) {
+			phy->suspended = 0;
+			sas_resume_port(phy);
+
+			/* phy came back, try to cancel the timeout */
+			wake_up(&sas_ha->eh_wait_q);
+			return;
+		} else {
 			SAS_DPRINTK("%s: phy%d belongs to port%d already(%d)!\n",
 				    __func__, phy->id, phy->port->id,
 				    phy->port->num_phys);
diff --git a/include/scsi/libsas.h b/include/scsi/libsas.h
index ad5229a..8ebd2747 100644
--- a/include/scsi/libsas.h
+++ b/include/scsi/libsas.h
@@ -79,7 +79,8 @@  enum phy_event {
 	PHYE_OOB_DONE         = 1,
 	PHYE_OOB_ERROR        = 2,
 	PHYE_SPINUP_HOLD      = 3, /* hot plug SATA, no COMWAKE sent */
-	PHY_NUM_EVENTS        = 4,
+	PHYE_RESUME_TIMEOUT   = 4,
+	PHY_NUM_EVENTS        = 5,
 };
 
 enum discover_event {
@@ -87,8 +88,10 @@  enum discover_event {
 	DISCE_REVALIDATE_DOMAIN = 1,
 	DISCE_PORT_GONE         = 2,
 	DISCE_PROBE		= 3,
-	DISCE_DESTRUCT		= 4,
-	DISC_NUM_EVENTS		= 5,
+	DISCE_SUSPEND		= 4,
+	DISCE_RESUME		= 5,
+	DISCE_DESTRUCT		= 6,
+	DISC_NUM_EVENTS		= 7,
 };
 
 /* ---------- Expander Devices ---------- */
@@ -128,7 +131,7 @@  struct ex_phy {
 	u8   attached_sas_addr[SAS_ADDR_SIZE];
 	u8   attached_phy_id;
 
-	u8   phy_change_count;
+	int phy_change_count;
 	enum routing_attribute routing_attr;
 	u8   virtual:1;
 
@@ -141,7 +144,7 @@  struct ex_phy {
 struct expander_device {
 	struct list_head children;
 
-	u16    ex_change_count;
+	int    ex_change_count;
 	u16    max_route_indexes;
 	u8     num_phys;
 
@@ -167,6 +170,7 @@  struct sata_device {
         enum   ata_command_set command_set;
         struct smp_resp        rps_resp; /* report_phy_sata_resp */
         u8     port_no;        /* port number, if this is a PM (Port) */
+	int    pm_result;
 
 	struct ata_port *ap;
 	struct ata_host ata_host;
@@ -180,6 +184,7 @@  struct ssp_device {
 
 enum {
 	SAS_DEV_GONE,
+	SAS_DEV_FOUND, /* device notified to lldd */
 	SAS_DEV_DESTROY,
 	SAS_DEV_EH_PENDING,
 	SAS_DEV_LU_RESET,
@@ -271,6 +276,7 @@  struct asd_sas_port {
 	enum   sas_linkrate linkrate;
 
 	struct sas_work work;
+	int suspended;
 
 /* public: */
 	int id;
@@ -319,6 +325,7 @@  struct asd_sas_phy {
 	unsigned long phy_events_pending;
 
 	int error;
+	int suspended;
 
 	struct sas_phy *phy;
 
@@ -382,6 +389,7 @@  enum sas_ha_state {
 	SAS_HA_DRAINING,
 	SAS_HA_ATA_EH_ACTIVE,
 	SAS_HA_FROZEN,
+	SAS_HA_SUSPENDED,
 };
 
 struct sas_ha_struct {
@@ -681,6 +689,9 @@  struct sas_domain_function_template {
 
 extern int sas_register_ha(struct sas_ha_struct *);
 extern int sas_unregister_ha(struct sas_ha_struct *);
+extern void sas_prep_resume_ha(struct sas_ha_struct *sas_ha);
+extern void sas_resume_ha(struct sas_ha_struct *sas_ha);
+extern void sas_suspend_ha(struct sas_ha_struct *sas_ha);
 
 int sas_set_phy_speed(struct sas_phy *phy,
 		      struct sas_phy_linkrates *rates);
diff --git a/include/scsi/sas_ata.h b/include/scsi/sas_ata.h
index 3e13c28..5bb5b1a 100644
--- a/include/scsi/sas_ata.h
+++ b/include/scsi/sas_ata.h
@@ -44,6 +44,8 @@  void sas_ata_eh(struct Scsi_Host *shost, struct list_head *work_q,
 void sas_ata_schedule_reset(struct domain_device *dev);
 void sas_ata_wait_eh(struct domain_device *dev);
 void sas_probe_sata(struct asd_sas_port *port);
+void sas_suspend_sata(struct asd_sas_port *port);
+void sas_resume_sata(struct asd_sas_port *port);
 void sas_ata_end_eh(struct ata_port *ap);
 #else
 
@@ -78,6 +80,14 @@  static inline void sas_probe_sata(struct asd_sas_port *port)
 {
 }
 
+static inline void sas_suspend_sata(struct asd_sas_port *port)
+{
+}
+
+static inline void sas_resume_sata(struct asd_sas_port *port)
+{
+}
+
 static inline int sas_get_ata_info(struct domain_device *dev, struct ex_phy *phy)
 {
 	return 0;