Hello Andrew,

On Wed, 19 Dec 2012 16:18:56 -0800
Andrew Morton [off-list ref] wrote:

On Mon, 10 Dec 2012 10:39:13 +0900
Atsushi Kumagai [off-list ref] wrote:

quoted

This patch adds the values related to buddy system to vmcoreinfo data
so that makedumpfile (dump filtering command) can filter out all free
pages with the new logic.
It's faster than the current logic because it can distinguish free page
by analyzing page structure at the same time as filtering for other
unnecessary pages (e.g. anonymous page).
OTOH, the current logic has to trace free_list to distinguish free 
pages while analyzing page structure to filter out other unnecessary
pages.

The new logic uses the fact that buddy page is marked by _mapcount == 
PAGE_BUDDY_MAPCOUNT_VALUE. But, _mapcount shares its memory with other
fields for SLAB/SLUB when PG_slab is set, so we need to check if PG_slab
is set or not before looking up _mapcount value.
And we can get the order of buddy system from private field.
To sum it up, the values below are required for this logic.

Required values:
  - OFFSET(page._mapcount)
  - OFFSET(page.private)
  - NUMBER(PG_slab)
  - NUMBER(PAGE_BUDDY_MAPCOUNT_VALUE)

Changelog from v1 to v2:
1. remove SIZE(pageflags)
  The new logic was changed after I sent v1 patch.  
  Accordingly, SIZE(pageflags) has been unnecessary for makedumpfile.

What's makedumpfile:
  makedumpfile creates a small dumpfile by excluding unnecessary pages
  for the analysis. To distinguish unnecessary pages, makedumpfile gets
  the vmcoreinfo data which has the minimum debugging information only
  for dump filtering.

Gee, this info is getting highly dependent upon deep internal kernel
behaviour.

Yes. makedumpfile should be changed depend on kernel version and we did it.

quoted

index 5e4bd78..b27efe4 100644

--- a/kernel/kexec.c
+++ b/kernel/kexec.c

@@ -1490,6 +1490,8 @@ static int __init crash_save_vmcoreinfo_init(void)

	VMCOREINFO_OFFSET(page, _count);
	VMCOREINFO_OFFSET(page, mapping);
	VMCOREINFO_OFFSET(page, lru);
+	VMCOREINFO_OFFSET(page, _mapcount);
+	VMCOREINFO_OFFSET(page, private);
	VMCOREINFO_OFFSET(pglist_data, node_zones);
	VMCOREINFO_OFFSET(pglist_data, nr_zones);
 #ifdef CONFIG_FLAT_NODE_MEM_MAP

@@ -1512,6 +1514,8 @@ static int __init crash_save_vmcoreinfo_init(void)

	VMCOREINFO_NUMBER(PG_lru);
	VMCOREINFO_NUMBER(PG_private);
	VMCOREINFO_NUMBER(PG_swapcache);
+	VMCOREINFO_NUMBER(PG_slab);
+	VMCOREINFO_NUMBER(PAGE_BUDDY_MAPCOUNT_VALUE);

We might change the PageBuddy() implementation at any time, and
makedumpfile will break.  Or in this case, become less efficient.

Is there any way in which we can move some of this logic into the
kernel?  In this case, add some kernel code which uses PageBuddy() on
behalf of makedumpfile, rather than replicating the PageBuddy() logic
in userspace?

In last month, Cliff Wickman proposed such idea:

  [PATCH v2] makedumpfile: request the kernel do page scans
  http://lists.infradead.org/pipermail/kexec/2012-November/007318.html

  [PATCH] scan page tables for makedumpfile, 3.0.13 kernel
  http://lists.infradead.org/pipermail/kexec/2012-November/007319.html

In his idea, the kernel does page scans to distinguish unnecessary pages
(free pages and others) and returns the list of PFN's which should be
excluded for makedumpfile.
As a result, makedumpfile doesn't need to consider internal kernel
behavior.

I think it's a good idea from the viewpoint of maintainability and
performance.


Thanks
Atsushi Kumagai

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>