2010-01-14 14:35:03

by Nitin Gupta

Subject: [PATCH] [mmotm] comment on swap notify locking constraints

A block device makes swap_slot_free_notify() callback
when the last reference to a swap slot is dropped.

This callback is made under swap_lock and often with
lock for corresponding swapcache page also held. This
is a note of warning for registered callback function
which must meet these constraints.

Signed-off-by: Nitin Gupta <[email protected]>
---
include/linux/blkdev.h | 1 +
1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 14b95a3..e3dcff6 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -1295,6 +1295,7 @@ struct block_device_operations {
unsigned long long);
int (*revalidate_disk) (struct gendisk *);
int (*getgeo)(struct block_device *, struct hd_geometry *);
+ /* this callback is with swap_lock and often page lock also held */
void (*swap_slot_free_notify) (struct block_device *, unsigned long);
struct module *owner;
};
--
1.6.2.5

2010-01-14 15:10:38

by Hugh Dickins

Subject: Re: [PATCH] [mmotm] comment on swap notify locking constraints

On Thu, 14 Jan 2010, Nitin Gupta wrote:

> A block device makes swap_slot_free_notify() callback
> when the last reference to a swap slot is dropped.
>
> This callback is made under swap_lock and often with
> lock for corresponding swapcache page also held. This
> is a note of warning for registered callback function
> which must meet these constraints.

You're right that the page lock of the corresponding swapcache page
may be held there, but that isn't what worried me enough to ask for
a comment: it's the page table lock (sometimes we call it pte lock),
a spinlock like the swap_lock, being held there that particularly
deserves comment. We don't want someone going to lots of effort
to avoid holding swap_lock there, only to find that effort wasted
because page table lock is also held.

Hugh

>
> Signed-off-by: Nitin Gupta <[email protected]>
> ---
> include/linux/blkdev.h | 1 +
> 1 files changed, 1 insertions(+), 0 deletions(-)
>
> diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
> index 14b95a3..e3dcff6 100644
> --- a/include/linux/blkdev.h
> +++ b/include/linux/blkdev.h
> @@ -1295,6 +1295,7 @@ struct block_device_operations {
> unsigned long long);
> int (*revalidate_disk) (struct gendisk *);
> int (*getgeo)(struct block_device *, struct hd_geometry *);
> + /* this callback is with swap_lock and often page lock also held */

and page table lock

> void (*swap_slot_free_notify) (struct block_device *, unsigned long);
> struct module *owner;
> };
> --
> 1.6.2.5

2010-01-14 18:01:51

by Nitin Gupta

Subject: Re: [PATCH] [mmotm] comment on swap notify locking constraints

On 01/14/2010 08:40 PM, Hugh Dickins wrote:
> On Thu, 14 Jan 2010, Nitin Gupta wrote:
>
>> A block device makes swap_slot_free_notify() callback
>> when the last reference to a swap slot is dropped.
>>
>> This callback is made under swap_lock and often with
>> lock for corresponding swapcache page also held. This
>> is a note of warning for registered callback function
>> which must meet these constraints.
>
> You're right that the page lock of the corresponding swapcache page
> may be held there, but that isn't what worried me enough to ask for
> a comment: it's the page table lock (sometimes we call it pte lock),
> a spinlock like the swap_lock, being held there that particularly
> deserves comment. We don't want someone going to lots of effort
> to avoid holding swap_lock there, only to find that effort wasted
> because page table lock is also held.

Thanks for the review.

I was missing this important point regarding pte lock. I will edit
the comment and resend the patch.

Nitin





>>
>> Signed-off-by: Nitin Gupta <[email protected]>
>> ---
>> include/linux/blkdev.h | 1 +
>> 1 files changed, 1 insertions(+), 0 deletions(-)
>>
>> diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
>> index 14b95a3..e3dcff6 100644
>> --- a/include/linux/blkdev.h
>> +++ b/include/linux/blkdev.h
>> @@ -1295,6 +1295,7 @@ struct block_device_operations {
>> unsigned long long);
>> int (*revalidate_disk) (struct gendisk *);
>> int (*getgeo)(struct block_device *, struct hd_geometry *);
>> + /* this callback is with swap_lock and often page lock also held */
>
> and page table lock
>
>> void (*swap_slot_free_notify) (struct block_device *, unsigned long);
>> struct module *owner;
>> };
>> --
>> 1.6.2.5
>