Skip to contentSkip to conversation

hv: vmbus_open(): reset the channel state on ENOMEM

PATCH Superseded Dexuan Cui 3 participants 5 messages Jan 29, 2015 – Feb 2, 2015 Hyper-V/Azure core and drivers

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Dexuan Cui

author maintainer
#0 lore raw

PATCH 3/3 hv: vmbus_open(): reset the channel state on ENOMEM

Without this patch, the state is put to CHANNEL_OPENING_STATE, and when the driver is loaded next time, vmbus_open() will fail immediately due to newchannel->state != CHANNEL_OPEN_STATE.

CC "K. Y. Srinivasan" <[email protected]>
---
 drivers/hv/channel.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
drivers/hv/channel.c +5 -3
index 2978f5e..26dcf26 100644
@@ -89,9 +89,10 @@ int vmbus_open(struct vmbus_channel *newchannel, u32 send_ringbuffer_size,
 	out = (void *)__get_free_pages(GFP_KERNEL|__GFP_ZERO,
 		get_order(send_ringbuffer_size + recv_ringbuffer_size));
 
-	if (!out)
-		return -ENOMEM;
-
+	if (!out) {
+		err = -ENOMEM;
+		goto error0;
+	}
 
 	in = (void *)((unsigned long)out + send_ringbuffer_size);
 
@@ -199,6 +200,7 @@ error0:
 	free_pages((unsigned long)out,
 		get_order(send_ringbuffer_size + recv_ringbuffer_size));
 	kfree(open_info);
+	newchannel->state = CHANNEL_OPEN_STATE;
 	return err;
 }
 EXPORT_SYMBOL_GPL(vmbus_open);
Signature · 1 line
1.9.1
#1 lore raw
Quoted 3 lines · Dexuan Cui

Without this patch, the state is put to CHANNEL_OPENING_STATE, and when the driver is loaded next time, vmbus_open() will fail immediately due to newchannel->state != CHANNEL_OPEN_STATE.

The patch makes sense, but I have one small doubt. We call vmbus_open from probe functions of various devices. E.g. in hyperv-keyboard we have:
error = vmbus_open(...)
if (error)

     goto err_free_mem;

and we don't call vmbus_close(...) on this path so no CHANNELMSG_CLOSECHANNEL will be send. Who's gonna retry probe? Wouldn't it be better to close the channel?

Quoted 33 lines
CC "K. Y. Srinivasan" <[email protected]>
---
 drivers/hv/channel.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
drivers/hv/channel.c +5 -3
index 2978f5e..26dcf26 100644
@@ -89,9 +89,10 @@ int vmbus_open(struct vmbus_channel *newchannel, u32 send_ringbuffer_size,
 	out = (void *)__get_free_pages(GFP_KERNEL|__GFP_ZERO,
 		get_order(send_ringbuffer_size + recv_ringbuffer_size));
 
-	if (!out)
-		return -ENOMEM;
-
+	if (!out) {
+		err = -ENOMEM;
+		goto error0;
+	}
 
 	in = (void *)((unsigned long)out + send_ringbuffer_size);
 
@@ -199,6 +200,7 @@ error0:
 	free_pages((unsigned long)out,
 		get_order(send_ringbuffer_size + recv_ringbuffer_size));
 	kfree(open_info);
+	newchannel->state = CHANNEL_OPEN_STATE;
 	return err;
 }
 EXPORT_SYMBOL_GPL(vmbus_open);
Signature · 1 line
  Vitaly
author maintainer
#2 lore raw
Quoted 20 lines

-----Original Message-----
From: Vitaly Kuznetsov [mailto:[email protected]]
Sent: Thursday, January 29, 2015 21:22 PM
To: Dexuan Cui
Cc: [email protected]; [email protected]; driverdev- [email protected]; [email protected]; [email protected]; [email protected]; KY Srinivasan; Haiyang Zhang
Subject: Re: [PATCH 3/3] hv: vmbus_open(): reset the channel state on ENOMEM

Quoted 3 lines · Dexuan Cui

Without this patch, the state is put to CHANNEL_OPENING_STATE, and when the driver is loaded next time, vmbus_open() will fail immediately due to newchannel->state != CHANNEL_OPEN_STATE.

The patch makes sense, but I have one small doubt. We call vmbus_open from probe functions of various devices. E.g. in hyperv-keyboard we have:
error = vmbus_open(...)
if (error)

Quoted 4 lines
     goto err_free_mem;

and we don't call vmbus_close(...) on this path so no CHANNELMSG_CLOSECHANNEL will be send.

Exactly.

Quoted 1 line

Who's gonna retry probe?

The user can try 'rmmod' and 'modprobe' the module to re-probe.

Quoted 1 line

Wouldn't it be better to close the channel?

Good question!

In your example, due to " goto err_free_mem", we don't run vmbus_close(), so the memory allocated for the ringbuffer is actually leaked!

Next time when we reload the module, vmbus_open() will allocate new memory for the ringbuffer.

KY, Haiyang,
Can you please confirm this issue?

Thanks,
-- Dexuan

Quoted 38 lines

drivers/hv/channel.c int vmbus_open(struct vmbus_channel *newchannel,

Quoted 13 lines
CC "K. Y. Srinivasan" <[email protected]>
---
 drivers/hv/channel.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
drivers/hv/channel.c +0 -0
index 2978f5e..26dcf26 100644
@@ -89,9 +89,10 @@ int vmbus_open(struct vmbus_channel *newchannel,

u32 send_ringbuffer_size,

Quoted 14 lines
drivers/hv/channel.c +4 -3
 	out = (void *)__get_free_pages(GFP_KERNEL|__GFP_ZERO,
 		get_order(send_ringbuffer_size + recv_ringbuffer_size));
 
-	if (!out)
-		return -ENOMEM;
-
+	if (!out) {
+		err = -ENOMEM;
+		goto error0;
+	}
 
 	in = (void *)((unsigned long)out + send_ringbuffer_size);
 

drivers/hv/channel.c error0: · line 206

Quoted 7 lines
drivers/hv/channel.c +1 -0
@@ -199,6 +200,7 @@ error0:
 	free_pages((unsigned long)out,
 		get_order(send_ringbuffer_size + recv_ringbuffer_size));
 	kfree(open_info);
+	newchannel->state = CHANNEL_OPEN_STATE;
 	return err;
 }
 EXPORT_SYMBOL_GPL(vmbus_open);

--
Vitaly

#3 lore raw
Quoted 38 lines

-----Original Message-----
From: Dexuan Cui
Sent: Thursday, January 29, 2015 9:03 PM
To: Vitaly Kuznetsov
Cc: [email protected]; [email protected]; driverdev- [email protected]; [email protected]; [email protected]; [email protected]; KY Srinivasan; Haiyang Zhang
Subject: RE: [PATCH 3/3] hv: vmbus_open(): reset the channel state on ENOMEM

Quoted 26 lines

-----Original Message-----
From: Vitaly Kuznetsov [mailto:[email protected]]
Sent: Thursday, January 29, 2015 21:22 PM
To: Dexuan Cui
Cc: [email protected]; [email protected]; driverdev- [email protected]; [email protected]; [email protected]; [email protected]; KY Srinivasan; Haiyang Zhang
Subject: Re: [PATCH 3/3] hv: vmbus_open(): reset the channel state on ENOMEM

Quoted 4 lines · Dexuan Cui

Without this patch, the state is put to CHANNEL_OPENING_STATE, and when the driver is loaded next time, vmbus_open() will fail immediately due to
newchannel->state != CHANNEL_OPEN_STATE.

The patch makes sense, but I have one small doubt. We call vmbus_open from probe functions of various devices. E.g. in hyperv-keyboard we have:
error = vmbus_open(...)
if (error)

     goto err_free_mem;

and we don't call vmbus_close(...) on this path so no CHANNELMSG_CLOSECHANNEL will be send.

If vmbus_open() fails, depending on where the failure occurred, the necessary rollback is expected to have been done in the vmbus_open() function itself. In vmbus_open function the last thing we do is to send the request to the host to open the channel. If this fails, then there is no need to close the channel.

Quoted 7 lines

Exactly.

Quoted 1 line

Who's gonna retry probe?

The user can try 'rmmod' and 'modprobe' the module to re-probe.

Quoted 1 line

Wouldn't it be better to close the channel?

Good question!

We close the channel only if the open() of the channel succeeded.

Quoted 6 lines

In your example, due to " goto err_free_mem", we don't run vmbus_close(), so the memory allocated for the ringbuffer is actually leaked!

Next time when we reload the module, vmbus_open() will allocate new memory for the ringbuffer.

I must be missing something here. When vmbus_open() fails, we will rollback the state and free up the memory allocated for the ring buffer.

As I look at the vmbus_open() code I do see an issue with regards cleaning up the gpadl state and I am sending a patch for this shortly. I will base this on top of the this series from Dexuan.

Regards,

K. Y

author maintainer
#4 lore raw
Quoted 45 lines

-----Original Message-----
From: KY Srinivasan
Sent: Monday, February 2, 2015 3:42 AM
To: Dexuan Cui; Vitaly Kuznetsov
Cc: [email protected]; [email protected]; driverdev- [email protected]; [email protected]; [email protected]; [email protected]; Haiyang Zhang
Subject: RE: [PATCH 3/3] hv: vmbus_open(): reset the channel state on ENOMEM

Quoted 36 lines

-----Original Message-----
From: Dexuan Cui
Sent: Thursday, January 29, 2015 9:03 PM
To: Vitaly Kuznetsov
Cc: [email protected]; [email protected]; driverdev- [email protected]; [email protected]; [email protected]; [email protected]; KY Srinivasan; Haiyang Zhang
Subject: RE: [PATCH 3/3] hv: vmbus_open(): reset the channel state on ENOMEM

Quoted 26 lines

-----Original Message-----
From: Vitaly Kuznetsov [mailto:[email protected]]
Sent: Thursday, January 29, 2015 21:22 PM
To: Dexuan Cui
Cc: [email protected]; [email protected]; driverdev- [email protected]; [email protected]; [email protected]; [email protected]; KY Srinivasan; Haiyang Zhang
Subject: Re: [PATCH 3/3] hv: vmbus_open(): reset the channel state on ENOMEM

Quoted 4 lines · Dexuan Cui

Without this patch, the state is put to CHANNEL_OPENING_STATE, and when the driver is loaded next time, vmbus_open() will fail immediately due to
newchannel->state != CHANNEL_OPEN_STATE.

The patch makes sense, but I have one small doubt. We call vmbus_open from probe functions of various devices. E.g. in hyperv-keyboard we have:
error = vmbus_open(...)
if (error)

     goto err_free_mem;

and we don't call vmbus_close(...) on this path so no CHANNELMSG_CLOSECHANNEL will be send.

Quoted 6 lines

If vmbus_open() fails, depending on where the failure occurred, the necessary rollback is expected to have
been done in the vmbus_open() function itself. In vmbus_open function the last thing we do is to send the request
to the host to open the channel. If this fails, then there is no need to close the channel.

Got it.

Quoted 8 lines
Quoted 7 lines

Exactly.

Quoted 1 line

Who's gonna retry probe?

The user can try 'rmmod' and 'modprobe' the module to re-probe.

Quoted 1 line

Wouldn't it be better to close the channel?

Good question!

We close the channel only if the open() of the channel succeeded.

Got it.

Quoted 10 lines
Quoted 6 lines

In your example, due to " goto err_free_mem", we don't run vmbus_close(), so the memory allocated for the ringbuffer is actually leaked!

Next time when we reload the module, vmbus_open() will allocate new memory for the ringbuffer.

I must be missing something here. When vmbus_open() fails, we will rollback the state and free up the memory
allocated for the ring buffer.

KY,
Sorry, I didn't check the code carefully...
You're correct. There is no issue here.

Quoted 4 lines

As I look at the vmbus_open() code I do see an issue with regards cleaning up the gpadl state and I am sending a patch for this
shortly. I will base this on top of the this series from Dexuan.

Thanks, KY!
I'm not sure you meant the below line at the end of vmbus_open():

        if (err == 0)
                newchannel->state = CHANNEL_OPENED_STATE;
        ELSE
	???

-- Dexuan