Skip to contentSkip to conversation

add execute_in_process_context() API

PATCH Andi Kleen 4 participants 9 messages Feb 8, 2006 1 message not in the archive

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): 1139342419.6065.8.camel@mulgrave.il.steeleye.com.suse.lists.linux.kernel

Andi Kleen

author
#1 lore raw

James Bottomley <[email protected]> writes:

In general this seems like a lot of code for a simple problem. It might be simpler to just put the work structure into the parent object and do the workqueue unconditionally

Review

Quoted 3 lines
(quoted diff) +3 -0
+	if (unlikely(!wqw)) {
+		printk(KERN_ERR "Failed to allocate memory\n");
+		WARN_ON(1);

WARN_ON for GFP_ATOMIC failure is bad. It is not really a bug.

-Andi

#2 lore raw
Quoted 7 lines · Andi Kleen

James Bottomley <[email protected]> writes:

In general this seems like a lot of code for a simple problem. It might be simpler to just put the work structure into the parent object and do the workqueue unconditionally

That apparently would have really bad performance problems. If we're !in_interrupt() we want to do the work synchronously.

But yes, if we can embed the work_struct inside the structure which the callback will operate on

Quoted 5 lines

Review

Quoted 3 lines
(quoted diff) +3 -0
+	if (unlikely(!wqw)) {
+		printk(KERN_ERR "Failed to allocate memory\n");
+		WARN_ON(1);

WARN_ON for GFP_ATOMIC failure is bad. It is not really a bug.

it will solve this problem. (And I think is is a problem: if the allocation fails, we leak things?)

author
#3 lore raw
Quoted 11 lines · Andrew Morton
Quoted 7 lines · Andi Kleen

James Bottomley <[email protected]> writes:

In general this seems like a lot of code for a simple problem. It might be simpler to just put the work structure into the parent object and do the workqueue unconditionally

That apparently would have really bad performance problems. If we're !in_interrupt() we want to do the work synchronously.

It depends if it's common or not. If it's uncommon then simpler code is better.

-Andi

#4 lore raw
In reply to Andi Kleen
Quoted 3 lines · Andi Kleen

In general this seems like a lot of code for a simple problem. It might be simpler to just put the work structure into the parent object and do the workqueue unconditionally

We can't do this. For the target release, there may be multiple calls to the reap function ... if we embed in the structure we have no room for more than one.

Quoted 5 lines

Review

Quoted 3 lines
(quoted diff) +3 -0
+	if (unlikely(!wqw)) {
+		printk(KERN_ERR "Failed to allocate memory\n");
+		WARN_ON(1);

WARN_ON for GFP_ATOMIC failure is bad. It is not really a bug.

Here, it means that the requested work wasn't executed. In SCSI that would mean an object is now in place permanently. The problem is that there's no real way to cope with failure in this case.

James

author
#5 lore raw
Quoted 2 lines · James Bottomley

The problem is that
there's no real way to cope with failure in this case.

Then you can't use GFP_ATOMIC. You have to redesign.

-Andi

#6 lore raw
Quoted 5 lines · Andi Kleen
Quoted 2 lines · James Bottomley

The problem is that
there's no real way to cope with failure in this case.

Then you can't use GFP_ATOMIC. You have to redesign.

My mailer seems to have deleted the part of the email with your redesign suggestion in it...

My initial point is that the current scsi_reap_target() infrastructure is more broken than the new API, so it does represent an improvement. I could also mitigate (but not solve) the problem by adding a wqw slab. However, I really think the redesign would be along the lines I suggested to Jens; do you agree?

James

author
#7 lore raw
Quoted 5 lines · James Bottomley
Quoted 4 lines · Andi Kleen
Quoted 2 lines · James Bottomley

The problem is that
there's no real way to cope with failure in this case.

Quoted 7 lines
Quoted 1 line

Then you can't use GFP_ATOMIC. You have to redesign.

My mailer seems to have deleted the part of the email with your redesign suggestion in it...

My initial point is that the current scsi_reap_target() infrastructure is more broken than the new API, so it does represent an improvement.

An improvement that will randomly fail under high load.

-Andi

#8 lore raw
In reply to James Bottomley
Quoted 19 lines · James Bottomley
Quoted 3 lines · Andi Kleen

In general this seems like a lot of code for a simple problem. It might be simpler to just put the work structure into the parent object and do the workqueue unconditionally

We can't do this. For the target release, there may be multiple calls to the reap function ... if we embed in the structure we have no room for more than one.

Quoted 5 lines

Review

Quoted 3 lines
(quoted diff) +3 -0
+	if (unlikely(!wqw)) {
+		printk(KERN_ERR "Failed to allocate memory\n");
+		WARN_ON(1);

WARN_ON for GFP_ATOMIC failure is bad. It is not really a bug.

Quoted 4 lines

Here, it means that the requested work wasn't executed. In SCSI that would mean an object is now in place permanently. The problem is that there's no real way to cope with failure in this case.

Hi, James.

I haven't really looked at the code carefully, but I think one work struct + atomic counter (say pending_reap_cnt) should do it. queue_work() guarantees the work is run at least once after the call, so bumping pending_reap_cnt and queueing the work in target reap and reaping pending_reap_cnt times in the work should work.

Signature · 1 line
tejun
#9 lore raw
Quoted 5 lines · Tejun

I haven't really looked at the code carefully, but I think one work struct + atomic counter (say pending_reap_cnt) should do it. queue_work() guarantees the work is run at least once after the call, so bumping pending_reap_cnt and queueing the work in target reap and reaping pending_reap_cnt times in the work should work.

Actually, no, unfortunately that doesn't work (and I did actually try it). The problem is that the target_reap actually removes the target from the sysfs namespace. So, if the reap is still pending but you cannot find the target, you'll allocate one, but then you may be racing to add it to the namespace (i.e. a quick succession of scsi remove-single-device followed by scsi add-single-device of the same H/C/T/L shows this). If you lose the race, the add fails because the namespace is already occupied. And, unfortunately, the device_del that does the namespace removal requires user context ...

James