Date:	Tue, 16 Dec 2008 14:43:32 -0800
From:	Mike Waychison <mikew@...gle.com>
To:	Dave Hansen <dave@...ux.vnet.ibm.com>
CC:	Oren Laadan <orenl@...columbia.edu>, jeremy@...p.org,
	arnd@...db.de, linux-api@...r.kernel.org,
	containers@...ts.linux-foundation.org,
	linux-kernel@...r.kernel.org, linux-mm@...ck.org,
	Linux Torvalds <torvalds@...l.org>,
	Alexander Viro <viro@...iv.linux.org.uk>,
	"H. Peter Anvin" <hpa@...or.com>,
	Thomas Gleixner <tglx@...utronix.de>,
	Ingo Molnar <mingo@...e.hu>
Subject: Re: [RFC v11][PATCH 03/13] General infrastructure for checkpoint
 restart

Dave Hansen wrote:
> On Tue, 2008-12-16 at 13:54 -0800, Mike Waychison wrote:
>> Oren Laadan wrote:
>>> diff --git a/checkpoint/sys.c b/checkpoint/sys.c
>>> index 375129c..bd14ef9 100644
>>> --- a/checkpoint/sys.c
>>> +++ b/checkpoint/sys.c
>>> +/*
>>> + * During checkpoint and restart the code writes outs/reads in data
>>> + * to/from the checkpoint image from/to a temporary buffer (ctx->hbuf).
>>> + * Because operations can be nested, use cr_hbuf_get() to reserve space
>>> + * in the buffer, then cr_hbuf_put() when you no longer need that space.
>>> + */
>> This seems a bit over-kill for buffer management no?  The only large 
>> header seems to be cr_hdr_head and the blowup comes from utsinfo string 
>> data (which could easily be moved out to be in it's own CR_HDR_STRING 
>> blocks).
>>
>> Wouldn't it be easier to use stack-local storage than balancing the 
>> cr_hbuf_get/put routines?
> 
> I've asked the same question, so I'll give you Oren's response that I
> remember:
> 
> cr_hbuf_get/put() are more of an API that we can use later.  For now,
> those buffers really are temporary.  But, in a case where we want to do
> a really fast checkpoint (to reduce "downtime" during the checkpoint) we
> store the image entirely in kernel memory to be written out later.
> 

Hmm, if I'm understanding you correctly, adding ref counts explicitly 
(like you suggest below)  would be used to let a lower layer defer 
writes.  Seems like this could be just as easily done with explicits 
kmallocs and transferring ownership of the allocated memory to the 
in-kernel representation handling layer below (which in turn queues the 
data structures for writes).

Any such layer would probably need to hold references to objects 
enqueued for write-out, so they will still a full cleanup path in case 
of success/error/abort (which means that any advantage of creating a 
pool of allocations for O(1) cleanup disappears).

Reference counting these guys doesn't have a clear advantage to me. 
They seem to have a pretty linear lifetime.

> In that case, cr_hbuf_put() stops doing anything at all because we keep
> the memory around.
> 
> cr_hbuf_get() becomes, "I need some memory to write some checkpointy
> things into".
> 
> cr_hbuf_put() becomes, "I'm done with this for now, only keep it if
> someone else needs it."
> 
> This might all be a lot clearer if we just kept some more explicit
> accounting around about who is using the objects.  Something like:
> 
> struct cr_buf {
> 	struct kref ref;
> 	int size;
> 	char buf[0];
> };
> 
> /* replaces cr_hbuf_get() */
> struct cr_buf *alloc_cr_buf(int size, gfp_t flags)
> {
> 	struct cr_buf *buf;
> 
> 	buf = kmalloc(sizeof(cr_buf) + size, flags);
> 	if (!buf)
> 		return NULL;
> 	buf->ref = 1; /* or whatever */
> 	buf->size = size;
> 	return buf;
> }
> 
> int cr_kwrite(struct cr_buf *buf)
> {
> 	if (writing_checkpoint_now) {
> 		// or whatever this write call was...
> 		vfs_write(&buf->buf[0], buf->size);
> 	} else if (deferring_write) {		
> 		kref_get(buf->kref);
> 	}
> }
> 
> -- Dave
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists