: Staging: hv: Use native wait primitives

In preperation for getting rid of the osd layer; change the code to use native wait interfaces. As part of this, fixed the buggy implementation in the osd_wait_primitive where the condition was cleared potentially after the condition was signalled.

drivers/staging/hv/channel_mgmt.c +2 -1

@@ -689,7 +693,8 @@ static void vmbus_ongpadl_torndown(
 				memcpy(&msginfo->response.gpadl_torndown,
 				       gpadl_torndown,
 				       sizeof(struct vmbus_channel_gpadl_torndown));
-				osd_waitevent_set(msginfo->waitevent);
+				msginfo->wait_condition = 1;
+				wake_up(&msginfo->waitevent);
 				break;
 			}
 		}

-----Original Message-----
From: Jiri Slaby [mailto:[email protected]]
Sent: Tuesday, February 15, 2011 4:21 AM
To: KY Srinivasan
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives

Quoted 5 lines · K. Y. Srinivasan

In preperation for getting rid of the osd layer; change the code to use native wait interfaces. As part of this, fixed the buggy implementation in the osd_wait_primitive where the condition was cleared potentially after the condition was signalled.

...

Review int vmbus_establish_gpadl(struct vmbus_channel

Quoted 1 line

(quoted diff) +0 -0

@@ -566,7 +567,11 @@ int vmbus_establish_gpadl(struct vmbus_channel

*channel, void *kbuffer,

Quoted 1 line

drivers/staging/hv/channel.c int vmbus_establish_gpadl(struct vmbus_channel *channel, void *kbuffer, · line 573

Quoted 7 lines

drivers/staging/hv/channel.c +4 -1

 		}
 	}
-	osd_waitevent_wait(msginfo->waitevent);
+	wait_event_timeout(msginfo->waitevent,
+				msginfo->wait_condition,
+				msecs_to_jiffies(1000));
+	BUG_ON(msginfo->wait_condition == 0);

The added BUG_ONs all over the code look scary. These shouldn't be BUG_ONs at all. You should maybe warn and bail out, but not kill the whole machine.

And looking at the code, more appropriate would be completion instead of wait events.

And msecs_to_jiffies(1000) == HZ.

drivers/staging/hv/channel_mgmt.c static void vmbus_ongpadl_torndown( · line 695

Quoted 5 lines

drivers/staging/hv/channel_mgmt.c +0 -0

@@ -689,7 +693,8 @@ static void vmbus_ongpadl_torndown(
 				memcpy(&msginfo->response.gpadl_torndown,
 				       gpadl_torndown,
 				       sizeof(struct

vmbus_channel_gpadl_torndown));

Quoted 7 lines

drivers/staging/hv/channel_mgmt.c +2 -1

-				osd_waitevent_set(msginfo->waitevent);
+				msginfo->wait_condition = 1;
+				wake_up(&msginfo->waitevent);
 				break;
 			}
 		}

drivers/staging/hv/channel_mgmt.c static void vmbus_onversion_response( · line 742

Quoted 9 lines

drivers/staging/hv/channel_mgmt.c +2 -1

@@ -730,7 +735,8 @@ static void vmbus_onversion_response(
 			memcpy(&msginfo->response.version_response,
 			      version_response,
 			      sizeof(struct vmbus_channel_version_response));
-			osd_waitevent_set(msginfo->waitevent);
+			msginfo->wait_condition = 1;
+			wake_up(&msginfo->waitevent);
 		}
 	}
 	spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);

regards,
--
js

Quoted 20 lines

-----Original Message-----
From: Jiri Slaby [mailto:[email protected]]
Sent: Tuesday, February 15, 2011 4:21 AM
To: KY Srinivasan
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives

Quoted 5 lines · K. Y. Srinivasan

In preperation for getting rid of the osd layer; change the code to use native wait interfaces. As part of this, fixed the buggy implementation in the osd_wait_primitive where the condition was cleared potentially after the condition was signalled.

...

Review int vmbus_establish_gpadl(struct vmbus_channel

Quoted 1 line

(quoted diff) +0 -0

@@ -566,7 +567,11 @@ int vmbus_establish_gpadl(struct vmbus_channel

*channel, void *kbuffer,

Quoted 1 line

Quoted 11 lines

drivers/staging/hv/channel.c int vmbus_establish_gpadl(struct vmbus_channel *channel, void *kbuffer, · line 573

Quoted 7 lines

drivers/staging/hv/channel.c +4 -1

 		}
 	}
-	osd_waitevent_wait(msginfo->waitevent);
+	wait_event_timeout(msginfo->waitevent,
+				msginfo->wait_condition,
+				msecs_to_jiffies(1000));
+	BUG_ON(msginfo->wait_condition == 0);

The added BUG_ONs all over the code look scary. These shouldn't be BUG_ONs at all. You should maybe warn and bail out, but not kill the whole machine.

This is Linux code running as a guest on a Windows host; and so the guest cannot tolerate a failure of the host. In the cases where I have chosen to BUG_ON, there is no reasonable recovery possible when the host is non-functional (as determined by a non-responsive host).

Quoted 4 lines

And looking at the code, more appropriate would be completion instead of wait events.

And msecs_to_jiffies(1000) == HZ.

Agreed. In this first round of cleanup, I chose to keep the primitives as they were in osd.c. Greg, if it is ok with you, I will send you a patch that fixes these issues on top of the patches I have already sent.

-----Original Message-----
From: Greg KH [mailto:[email protected]]
Sent: Tuesday, February 15, 2011 9:03 AM
To: KY Srinivasan
Cc: Jiri Slaby; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives

Quoted 32 lines · KY Srinivasan

Quoted 20 lines

-----Original Message-----
From: Jiri Slaby [mailto:[email protected]]
Sent: Tuesday, February 15, 2011 4:21 AM
To: KY Srinivasan
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives

Quoted 5 lines · K. Y. Srinivasan

In preperation for getting rid of the osd layer; change the code to use native wait interfaces. As part of this, fixed the buggy implementation in the osd_wait_primitive where the condition was cleared potentially after the condition was signalled.

...

Review int vmbus_establish_gpadl(struct vmbus_channel

Quoted 1 line

(quoted diff) +0 -0

@@ -566,7 +567,11 @@ int vmbus_establish_gpadl(struct vmbus_channel

*channel, void *kbuffer,

Quoted 1 line

Quoted 11 lines

drivers/staging/hv/channel.c int vmbus_establish_gpadl(struct vmbus_channel *channel, void *kbuffer, · line 573

Quoted 7 lines

drivers/staging/hv/channel.c +4 -1

 		}
 	}
-	osd_waitevent_wait(msginfo->waitevent);
+	wait_event_timeout(msginfo->waitevent,
+				msginfo->wait_condition,
+				msecs_to_jiffies(1000));
+	BUG_ON(msginfo->wait_condition == 0);

The added BUG_ONs all over the code look scary. These shouldn't be BUG_ONs at all. You should maybe warn and bail out, but not kill the whole machine.

Quoted 1 line

This is Linux code running as a guest on a Windows host; and so the guest

cannot

Quoted 1 line

tolerate a failure of the host. In the cases where I have chosen to BUG_ON,

there

Quoted 1 line

is no reasonable recovery possible when the host is non-functional (as

determined

Quoted 1 line

by a non-responsive host).

If you have a non-responsive host, wouldn't that imply that this guest code wouldn't run at all? :)

Having BUG_ON() in drivers is not a good idea either way. Please remove these in future patches.

Quoted 10 lines

Quoted 5 lines

And looking at the code, more appropriate would be completion instead of wait events.

And msecs_to_jiffies(1000) == HZ.

Agreed. In this first round of cleanup, I chose to keep the primitives as they were in osd.c. Greg, if it is ok with you, I will send you a patch that fixes these issues on top of the patches I have already sent.

Yes, that is fine.

thanks,

greg k-h

Quoted 45 lines

-----Original Message-----
From: Greg KH [mailto:[email protected]]
Sent: Tuesday, February 15, 2011 9:03 AM
To: KY Srinivasan
Cc: Jiri Slaby; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives

Quoted 32 lines · KY Srinivasan

Quoted 20 lines

-----Original Message-----
From: Jiri Slaby [mailto:[email protected]]
Sent: Tuesday, February 15, 2011 4:21 AM
To: KY Srinivasan
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives

Quoted 5 lines · K. Y. Srinivasan

In preperation for getting rid of the osd layer; change the code to use native wait interfaces. As part of this, fixed the buggy implementation in the osd_wait_primitive where the condition was cleared potentially after the condition was signalled.

...

Review int vmbus_establish_gpadl(struct vmbus_channel

Quoted 1 line

(quoted diff) +0 -0

@@ -566,7 +567,11 @@ int vmbus_establish_gpadl(struct vmbus_channel

*channel, void *kbuffer,

Quoted 1 line

Quoted 11 lines

drivers/staging/hv/channel.c int vmbus_establish_gpadl(struct vmbus_channel *channel, void *kbuffer, · line 573

Quoted 7 lines

drivers/staging/hv/channel.c +4 -1

 		}
 	}
-	osd_waitevent_wait(msginfo->waitevent);
+	wait_event_timeout(msginfo->waitevent,
+				msginfo->wait_condition,
+				msecs_to_jiffies(1000));
+	BUG_ON(msginfo->wait_condition == 0);

The added BUG_ONs all over the code look scary. These shouldn't be BUG_ONs at all. You should maybe warn and bail out, but not kill the whole machine.

Quoted 1 line

This is Linux code running as a guest on a Windows host; and so the guest

cannot

Quoted 8 lines

Quoted 1 line

tolerate a failure of the host. In the cases where I have chosen to BUG_ON,

there

Quoted 1 line

is no reasonable recovery possible when the host is non-functional (as

determined

Quoted 1 line

by a non-responsive host).

If you have a non-responsive host, wouldn't that imply that this guest code wouldn't run at all? :)

The fact that on a particular transaction the host has not responded within an expected time interval does not necessarily mean that the guest code would not be running. There may be issues on the host side that may be either transient or permanent that may cause problems like this. Keep in mind, HyperV is a type 1 hypervisor that would schedule all VMs including the host and so, guest would get scheduled.

Quoted 3 lines

Having BUG_ON() in drivers is not a good idea either way. Please remove these in future patches.

In situations where there is not a reasonable rollback strategy (for instance in one of the cases, we are granting access to the guest physical pages to the host) we really have only 2 options:

1) Wait until the host responds. This wait could potentially be unbounded and in fact this was the way the code was to begin with. One of the reviewers had suggested that unbounded wait was to be corrected.
2) Wait for a specific period and if the host does not respond within a reasonable period, kill the guest since there is no recovery possible.

I chose option 2, as part of addressing some of the prior review comments. If the consensus now is to go back to option 1, I am fine with that;

-----Original Message-----
From: Greg KH [mailto:[email protected]]
Sent: Tuesday, February 15, 2011 11:30 AM
To: KY Srinivasan
Cc: Jiri Slaby; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives

Quoted 21 lines · KY Srinivasan

Quoted 21 lines

-----Original Message-----
From: Greg KH [mailto:[email protected]]
Sent: Tuesday, February 15, 2011 9:03 AM
To: KY Srinivasan
Cc: Jiri Slaby; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives

Quoted 10 lines · KY Srinivasan

Quoted 10 lines

-----Original Message-----
From: Jiri Slaby [mailto:[email protected]]
Sent: Tuesday, February 15, 2011 4:21 AM
To: KY Srinivasan
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives

Quoted 8 lines

Quoted 8 lines

Quoted 8 lines

Quoted 8 lines · K. Y. Srinivasan

Quoted 5 lines

In preperation for getting rid of the osd layer; change the code to use native wait interfaces. As part of this, fixed the buggy implementation in the osd_wait_primitive where the condition was cleared potentially after the condition was signalled.

...

Review int vmbus_establish_gpadl(struct

Quoted 1 line

(quoted diff) +0 -0

@@ -566,7 +567,11 @@ int vmbus_establish_gpadl(struct

vmbus_channel

Quoted 18 lines

Quoted 18 lines

Quoted 14 lines

Quoted 13 lines

*channel, void *kbuffer,

drivers/staging/hv/channel.c int vmbus_establish_gpadl(struct vmbus_channel *channel, void *kbuffer, · line 573

Quoted 8 lines

drivers/staging/hv/channel.c +4 -1

 		}
 	}
-	osd_waitevent_wait(msginfo->waitevent);
+	wait_event_timeout(msginfo->waitevent,
+				msginfo->wait_condition,
+				msecs_to_jiffies(1000));
+	BUG_ON(msginfo->wait_condition == 0);

The added BUG_ONs all over the code look scary. These shouldn't be BUG_ONs at all. You should maybe warn and bail out, but not kill the whole machine.

Quoted 1 line

This is Linux code running as a guest on a Windows host; and so the guest

cannot

Quoted 1 line

tolerate a failure of the host. In the cases where I have chosen to BUG_ON,

there

Quoted 8 lines

Quoted 6 lines

Quoted 1 line

is no reasonable recovery possible when the host is non-functional (as

determined

Quoted 1 line

by a non-responsive host).

If you have a non-responsive host, wouldn't that imply that this guest code wouldn't run at all? :)

The fact that on a particular transaction the host has not responded within an

expected

Quoted 1 line

time interval does not necessarily mean that the guest code would not be

running. There may be

Quoted 1 line

issues on the host side that may be either transient or permanent that may

cause problems like

Quoted 1 line

this. Keep in mind, HyperV is a type 1 hypervisor that would schedule all VMs

including the host

Quoted 10 lines

and so, guest would get scheduled.

Quoted 3 lines

Having BUG_ON() in drivers is not a good idea either way. Please remove these in future patches.

In situations where there is not a reasonable rollback strategy (for instance in one of the cases, we are granting access to the guest physical pages to the host) we really have only 2 options:

Quoted 6 lines

1) Wait until the host responds. This wait could potentially be unbounded and in fact this was the way the code was to begin with. One of the reviewers had suggested that unbounded wait was to be corrected.
2) Wait for a specific period and if the host does not respond within a reasonable period, kill the guest since there is no recovery possible.

Killing the guest is a very serious thing, causing all sorts of possible problems with it, right?

Quoted 3 lines

I chose option 2, as part of addressing some of the prior review comments. If the consensus now is to go back to option 1, I am fine with that;

Unbounded waits aren't ok either, you need some sort of timeout.

But, as this is a bit preferable to dieing, I suggest doing this, and comment the heck out of it to explain all of this for anyone who reads it.

thanks,

greg k-h

If I understand you correctly, you would be prefer to have unbounded waiting with comments justifying why we cannot have timeouts. I will roll out a patch once the tree stabilizes.