[PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[Nitin Gupta]

ramzswap driver creates RAM based block devices which can be
used (only) as swap disks. Pages swapped to these disks are
compressed and stored in memory itself.

However, these devices do not get any notification when a swap
slot is freed (swap_map[i] reaches 0). So, we cannot free memory
allocated corresponding to this swap slot. Such stale data can
quickly accumulate in (compressed) memory defeating the whole
purpose of such devices.

To overcome this problem, we now add a callback in 'struct swap_info_struct'
which is called as soon as a swap slot is freed.

Adding handler for this callback:
	swapon notifier --> set_swap_free_notify(swap_type, fn)

Removing handler:
	swapoff notifier --> set_swap_free_notify(swap_type, NULL)


Alternative approaches:
1) Add callback directly in 'struct block_device_operations' but
that is considered too hacky.
2) Use swap discard mechanism: It involves unncessary overhead of
allocating 'discard bio' requests and its too slow to serve ramzswap
needs.

 drivers/staging/ramzswap/ramzswap_drv.c   |   91 +++++++++++++++++++++++++++++
 drivers/staging/ramzswap/ramzswap_drv.h   |    1 +
 drivers/staging/ramzswap/ramzswap_ioctl.h |    1 +
 include/linux/swap.h                      |   16 +++++-
 mm/swapfile.c                             |   78 ++++++++++++++++++++++++
 5 files changed, 185 insertions(+), 2 deletions(-)

Re: [PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[Nitin Gupta]

On 03/05/2010 03:52 PM, Nitin Gupta wrote:
> ramzswap driver creates RAM based block devices which can be
> used (only) as swap disks. Pages swapped to these disks are
> compressed and stored in memory itself.
> 
> However, these devices do not get any notification when a swap
> slot is freed (swap_map[i] reaches 0). So, we cannot free memory
> allocated corresponding to this swap slot. Such stale data can
> quickly accumulate in (compressed) memory defeating the whole
> purpose of such devices.
> 
> To overcome this problem, we now add a callback in 'struct swap_info_struct'
> which is called as soon as a swap slot is freed.
> 
> Adding handler for this callback:
> 	swapon notifier --> set_swap_free_notify(swap_type, fn)
> 
> Removing handler:
> 	swapoff notifier --> set_swap_free_notify(swap_type, NULL)
> 
> 
> Alternative approaches:
> 1) Add callback directly in 'struct block_device_operations' but
> that is considered too hacky.
> 2) Use swap discard mechanism: It involves unncessary overhead of
> allocating 'discard bio' requests and its too slow to serve ramzswap
> needs.
> 
>  drivers/staging/ramzswap/ramzswap_drv.c   |   91 +++++++++++++++++++++++++++++
>  drivers/staging/ramzswap/ramzswap_drv.h   |    1 +
>  drivers/staging/ramzswap/ramzswap_ioctl.h |    1 +
>  include/linux/swap.h                      |   16 +++++-
>  mm/swapfile.c                             |   78 ++++++++++++++++++++++++
>  5 files changed, 185 insertions(+), 2 deletions(-)
> 


Any chance it can go into 2.6.34? Any comments?


Thanks,
Nitin

Re: [PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[Hugh Dickins]

On Fri, 5 Mar 2010, Nitin Gupta wrote:

> ramzswap driver creates RAM based block devices which can be
> used (only) as swap disks. Pages swapped to these disks are
> compressed and stored in memory itself.
> 
> However, these devices do not get any notification when a swap
> slot is freed (swap_map[i] reaches 0). So, we cannot free memory
> allocated corresponding to this swap slot. Such stale data can
> quickly accumulate in (compressed) memory defeating the whole
> purpose of such devices.
> 
> To overcome this problem, we now add a callback in 'struct swap_info_struct'
> which is called as soon as a swap slot is freed.
> 
> Adding handler for this callback:
> 	swapon notifier --> set_swap_free_notify(swap_type, fn)
> 
> Removing handler:
> 	swapoff notifier --> set_swap_free_notify(swap_type, NULL)
> 
> 
> Alternative approaches:
> 1) Add callback directly in 'struct block_device_operations' but
> that is considered too hacky.
> 2) Use swap discard mechanism: It involves unncessary overhead of
> allocating 'discard bio' requests and its too slow to serve ramzswap
> needs.
> 
>  drivers/staging/ramzswap/ramzswap_drv.c   |   91 +++++++++++++++++++++++++++++
>  drivers/staging/ramzswap/ramzswap_drv.h   |    1 +
>  drivers/staging/ramzswap/ramzswap_ioctl.h |    1 +
>  include/linux/swap.h                      |   16 +++++-
>  mm/swapfile.c                             |   78 ++++++++++++++++++++++++
>  5 files changed, 185 insertions(+), 2 deletions(-)

This is one of the various solutions I disliked already, isn't it?

To me this is just a more convoluted and obscure variant of the
block_device_operations swap_slot_free_notify patch you had in mmotm,
which Linus has rejected.

I admit, I did not understand at all what Linus was proposing with
readpage, writepage and address_space_operations: I kept quiet in
the hope that you'd understand where I didn't!

Hugh

Re: [PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[Nitin Gupta]

On Thu, Mar 11, 2010 at 12:52 PM, Hugh Dickins
<hugh.dickins@tiscali.co.uk> wrote:
> On Fri, 5 Mar 2010, Nitin Gupta wrote:
<snip>
>
>> Adding handler for this callback:
>>       swapon notifier --> set_swap_free_notify(swap_type, fn)
>>
>> Removing handler:
>>       swapoff notifier --> set_swap_free_notify(swap_type, NULL)
>>
>>
>> Alternative approaches:
>> 1) Add callback directly in 'struct block_device_operations' but
>> that is considered too hacky.
>> 2) Use swap discard mechanism: It involves unncessary overhead of
>> allocating 'discard bio' requests and its too slow to serve ramzswap
>> needs.
>>
>>  drivers/staging/ramzswap/ramzswap_drv.c   |   91 +++++++++++++++++++++++++++++
>>  drivers/staging/ramzswap/ramzswap_drv.h   |    1 +
>>  drivers/staging/ramzswap/ramzswap_ioctl.h |    1 +
>>  include/linux/swap.h                      |   16 +++++-
>>  mm/swapfile.c                             |   78 ++++++++++++++++++++++++
>>  5 files changed, 185 insertions(+), 2 deletions(-)
>

> This is one of the various solutions I disliked already, isn't it?
>

This is a kind of hybrid of previous approaches. We now have notifier
for swapon/off events only -- which Linus seems to be okay with (earlier
we had notifier for slot free event also). Handler for swapon is responsible
for setting callback for slot free event  (earlier we were installing it from
"ramzswap_make_request" directly, when first I/O is received).


> To me this is just a more convoluted and obscure variant of the
> block_device_operations swap_slot_free_notify patch you had in mmotm,
> which Linus has rejected.
>

This slot free notification is a sort of black area for which I have
really run out of
ideas. I tried all the approaches I could think of (and all you and
Pekka suggested)
but nothing seems to satisfy everyone. This is so essential for ramzswap but now
I don't know how to get this done.

> I admit, I did not understand at all what Linus was proposing with
> readpage, writepage and address_space_operations: I kept quiet in
> the hope that you'd understand where I didn't!
>

I guess he is suggesting creating a fictional 'address_space' just like
we have 'struct address_space swapper_space'. This way, we can
intercept writepage naturally in pageout() with new mapping->a_ops->writepage().
And then similarly with mapping->a_ops->readpage().

I'm now looking into this approach. This might allow compressing page cache
pages too. However, figuring out all the details can take significant
amount of time
and allowing ramzswap development to happen in parallel should be okay.

Thanks,
Nitin

[PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[Nitin Gupta]

(tested on mainline but should apply to linux-next cleanly)

ramzswap driver creates RAM based block devices which can be
used (only) as swap disks. Pages swapped to these disks are
compressed and stored in memory itself.

However, these devices do not get any notification when a swap
slot is freed (swap_map[i] reaches 0). So, we cannot free memory
allocated corresponding to this swap slot. Such stale data can
quickly accumulate in (compressed) memory defeating the whole
purpose of such devices.

To overcome this problem, we now add a callback in 'struct swap_info_struct'
which is called as soon as a swap slot is freed.

Adding handler for this callback:
swapon notifier --> set_swap_free_notify(swap_type, fn)

Removing handler:
swapoff notifier --> set_swap_free_notify(swap_type, NULL)


Alternative approaches:
1) Add callback directly in 'struct block_device_operations' but
that is considered too hacky (nacked by Linus).
2) Use swap discard mechanism: It involves unncessary overhead of
allocating 'discard bio' requests and its too slow to serve ramzswap
needs.

I also tried an approach similar to what Linus suggested:
Create an address_space like swaper_space, so we can catch
read/writes with a_ops->readpage() etc. However, this approach
turned out to be quite difficult and I could not get it to work.

Nitin Gupta (3):
  Add notifiers for swapon and swapoff events
  Send callback when a swap slot is freed
  ramzswap: Register for swap event notifiers and callback

 drivers/staging/ramzswap/TODO           |    5 --
 drivers/staging/ramzswap/ramzswap_drv.c |   79 ++++++++++++++++++++++++++++++-
 include/linux/swap.h                    |   15 ++++++
 mm/swapfile.c                           |   70 +++++++++++++++++++++++++++
 4 files changed, 163 insertions(+), 6 deletions(-)
 delete mode 100644 drivers/staging/ramzswap/TODO

Re: [PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[Linus Torvalds]

On Wed, 5 May 2010, Nitin Gupta wrote:
> 
> ramzswap driver creates RAM based block devices which can be
> used (only) as swap disks. Pages swapped to these disks are
> compressed and stored in memory itself.

Ok, this patch series looks way better, if only because it looks less 
hacky.

That said, I absolutely _hate_ the f*cking notifier model that takes 
"type" flags. It's a disgrace. It's a horrible horrible model.

I'd much rather bind a nice "swap_operations" structure to the device, and 
have that structure have function pointers for the different operations. 
No stupid "operation type codes". Real, honest-to-goodness function 
pointers.

The notifier layer is a total piece of sh*t. I'm sorry I ever merged it, 
and I'm _doubly_ sorry that it's use is so horribly widespread. It's a 
mistake.

		Linus

Re: [PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[Nitin Gupta]

On 05/05/2010 08:44 PM, Linus Torvalds wrote:
> 
> 
> On Wed, 5 May 2010, Nitin Gupta wrote:
>>
>> ramzswap driver creates RAM based block devices which can be
>> used (only) as swap disks. Pages swapped to these disks are
>> compressed and stored in memory itself.
> 
> Ok, this patch series looks way better, if only because it looks less 
> hacky.
> 
> That said, I absolutely _hate_ the f*cking notifier model that takes 
> "type" flags. It's a disgrace. It's a horrible horrible model.
> 

You mean you didn't like the 'swap type' value passed around by notifier
calls, as here:
 "blocking_notifier_call_chain(&swapon_notify_list, type, swap_file);" ?

> I'd much rather bind a nice "swap_operations" structure to the device, and 
> have that structure have function pointers for the different operations. 
> No stupid "operation type codes". Real, honest-to-goodness function 
> pointers.
> 

I think such 'swap_operations' structure will be have to be part of
block_device_operations, so we may access it from swap_entry_free()
where a swap slot is freed. This will also get rid of all this notifier
stuff.

The patch you nacked did something similar: it add 'swap_slot_free_callback'
directly to block_device_operations. Without such change, I could not think
of any way to do away with notifiers.


> The notifier layer is a total piece of sh*t. I'm sorry I ever merged it, 
> and I'm _doubly_ sorry that it's use is so horribly widespread. It's a 
> mistake.

Thanks,
Nitin

Re: [PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[Linus Torvalds]

On Wed, 5 May 2010, Nitin Gupta wrote:
> 
> I think such 'swap_operations' structure will be have to be part of
> block_device_operations, so we may access it from swap_entry_free()
> where a swap slot is freed. This will also get rid of all this notifier
> stuff.

Yes, I think adding it to block_device_operations would be fine. That 
sounds like a sane layering, and would make it easy for a block device 
driver to say "I want to know about swap events".

In fact, for regular block devices, a swap block free might well translate 
into a TRIM command some day (where "some day" means when the SSD's 
actually get their stuff together and there is real upside and not just 
"most cases will be very slow and the upside is debatable").

> The patch you nacked did something similar: it add 'swap_slot_free_callback'
> directly to block_device_operations. Without such change, I could not think
> of any way to do away with notifiers.

Umm. No. IIRC, the patch I NAK'ed aded it to the 'swap_info_struct', which 
I said was the wrong level. The block device driver level would seem to be 
the _right_ level, since that's what ramzswap is. No?

Also, the patch I NAK'ed also used those nasty notifier chains, making it 
even uglier.

				Linus

Re: [PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[Nitin Gupta]

On 05/05/2010 09:52 PM, Linus Torvalds wrote:
> 
> 
> On Wed, 5 May 2010, Nitin Gupta wrote:
>>
>> I think such 'swap_operations' structure will be have to be part of
>> block_device_operations, so we may access it from swap_entry_free()
>> where a swap slot is freed. This will also get rid of all this notifier
>> stuff.
> 
> Yes, I think adding it to block_device_operations would be fine. That 
> sounds like a sane layering, and would make it easy for a block device 
> driver to say "I want to know about swap events".
> 
> In fact, for regular block devices, a swap block free might well translate 
> into a TRIM command some day (where "some day" means when the SSD's 
> actually get their stuff together and there is real upside and not just 
> "most cases will be very slow and the upside is debatable").
> 

Its great if adding such a callback to block_device_operations is okay. Hugh
suggested this approach and I'm distributing it with compcache for quite some
time now:
http://code.google.com/p/compcache/source/browse/patches/patch_swap_notify_core_support_2.6.33.diff

Can you please have a lot at patch above and see if its acceptable? Then I will
post it to lkml again.


>> The patch you nacked did something similar: it add 'swap_slot_free_callback'
>> directly to block_device_operations. Without such change, I could not think
>> of any way to do away with notifiers.
> 
> Umm. No. IIRC, the patch I NAK'ed aded it to the 'swap_info_struct', which 
> I said was the wrong level. The block device driver level would seem to be 
> the _right_ level, since that's what ramzswap is. No?
> 
> Also, the patch I NAK'ed also used those nasty notifier chains, making it 
> even uglier.

Please see the original mail below (patch you nacked). Maybe, at that time, I didn't
make it clear that ramzswap is really a *block device* :)



-------- Original Message --------
Subject: [nacked] mm-add-swap-slot-free-callback-to-block_device_operations.patch removed from -mm tree
Date: Tue, 09 Mar 2010 14:30:27 -0800
From: akpm@linux-foundation.org
To: <snip/>

The patch titled
     mm: add swap slot free callback to block_device_operations
has been removed from the -mm tree.  Its filename was
     mm-add-swap-slot-free-callback-to-block_device_operations.patch

This patch was dropped because it was nacked

The current -mm tree may be found at http://userweb.kernel.org/~akpm/mmotm/

------------------------------------------------------
Subject: mm: add swap slot free callback to block_device_operations
From: Nitin Gupta <ngupta@vflare.org>

This callback is required when RAM based devices are used as swap disks. 
One such device is ramzswap[1] which is used as compressed in-memory swap
disk.  For such devices, we need a callback as soon as a swap slot is no
longer used to allow freeing memory allocated for this slot.  Without this
callback, stale data can quickly accumulate in memory defeating the whole
purpose of such devices.

Another user of this callback will be "preswap" as introduced by
"Transcendent Memory" patches: http://lwn.net/Articles/367286/ (I intend
to integrade preswap with ramzswap).

[1] ramzswap: http://code.google.com/p/compcache/

Signed-off-by: Nitin Gupta <ngupta@vflare.org>
Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Acked-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Greg KH <greg@kroah.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
---

 include/linux/blkdev.h |    2 ++
 mm/swapfile.c          |    3 +++
 2 files changed, 5 insertions(+)

diff -puN include/linux/blkdev.h~mm-add-swap-slot-free-callback-to-block_device_operations include/linux/blkdev.h
--- a/include/linux/blkdev.h~mm-add-swap-slot-free-callback-to-block_device_operations
+++ a/include/linux/blkdev.h
@@ -1310,6 +1310,8 @@ 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 sometimes page table lock held */
+	void (*swap_slot_free_notify) (struct block_device *, unsigned long);
 	struct module *owner;
 };
 
diff -puN mm/swapfile.c~mm-add-swap-slot-free-callback-to-block_device_operations mm/swapfile.c
--- a/mm/swapfile.c~mm-add-swap-slot-free-callback-to-block_device_operations
+++ a/mm/swapfile.c
@@ -574,6 +574,7 @@ static unsigned char swap_entry_free(str
 
 	/* free if no reference */
 	if (!usage) {
+		struct gendisk *disk = p->bdev->bd_disk;
 		if (offset < p->lowest_bit)
 			p->lowest_bit = offset;
 		if (offset > p->highest_bit)
@@ -583,6 +584,8 @@ static unsigned char swap_entry_free(str
 			swap_list.next = p->type;
 		nr_swap_pages++;
 		p->inuse_pages--;
+		if (disk->fops->swap_slot_free_notify)
+			disk->fops->swap_slot_free_notify(p->bdev, offset);
 	}
 
 	return usage;
_

Patches currently in -mm which might be from ngupta@vflare.org are

linux-next.patch
mm-add-swap-slot-free-callback-to-block_device_operations.patch

Re: [PATCH 0/3] ramzswap: Eliminate stale data in compressed memory

[Linus Torvalds]

On Wed, 5 May 2010, Nitin Gupta wrote:
> 
> Please see the original mail below (patch you nacked). Maybe, at that time, I didn't
> make it clear that ramzswap is really a *block device* :)

Oh, you're right. I looked up another patch of yours that had that 
swap_info" thing, and decided I hated that one with a passion, but it's 
not the one I NAK'ed in the earlier discussion.

Now that I see the block-layer patch, my reaction is (a) it's so much 
nicer than using that horrid nasty 'notifier' crap and (b) it reminds me 
why I wasn't entirely happy: it doesn't work - or even make sense - for 
filesystem-backed swap.

So when you do

	struct gendisk *disk = p->bdev->bd_disk;
	..
	if (disk->fops->swap_slot_free_notify)
		disk->fops->swap_slot_free_notify(p->bdev, offset);

there's nothing that says that 'offset' makes any sense at all, because if 
it's a swap-file on a device, it does all kinds of totally wrong things.

So I don't think that patch works either. I still suspect that the right 
"level" for something like this should be the 'mapping' level (which is 
how we actually do the write), but that seems to not work well with the 
block device layer.

So at a _minimum_, that 'disk->fops' approach needs to verify that the 
swap device is actually the whole bdev, and that the bdev isn't just the 
backing store for a swap _file_.

I dunno how to best check that. Either add a new flag to 
'swap_info_struct' that gets set on 'swapon()' whether it's a full device 
or a file. Or possibly just something like

	static int swap_is_block_device(struct swap_info_struct *p)
	{
		return S_ISBLK(p->swap_file->f_mapping->host);
	}

instead.

Because doing that 'disk->fops' thing _really_ isn't right if it isn't a 
disk.

		Linus