[patch 8/9] Add support for filtering out user pages under Xen4

[patch 8/9] Add support for filtering out user pages under Xen4

[Petr Tesarik]

[-- Attachment #1: 08-xen4-exclude-user.patch --]
[-- Type: text/plain, Size: 5014 bytes --]

In Xen-4.0+, usage of a page must be determined from its PGC_xxx
flags, which are stored in the highest bits of the _count field
of struct page_info.

Let's keep the original function for Xen3 only and add a new
function for walking the pages under Xen4. This avoids adding many
new conditionals to the inner loop and also makes the logic easier
to follow for all Xen versions.

Signed-off-by: Norbert Trapp <norbert.trapp@ts.fujitsu.com>
Signed-off-by: Petr Tesarik <ptesarik@suse.cz>

---
 makedumpfile.c |  103 +++++++++++++++++++++++++++++++++++++++++++++++++++++----
 makedumpfile.h |   25 +++++++++++++
 2 files changed, 122 insertions(+), 6 deletions(-)

--- a/makedumpfile.h
+++ b/makedumpfile.h
@@ -78,6 +78,31 @@ int get_mem_type(void);
 #define LSEEKED_PDATA	(3)
 
 /*
+ * Xen page flags
+ */
+#define BITS_PER_BYTE (8)
+#define BITS_PER_LONG (BITS_PER_BYTE * sizeof(long))
+#define PG_shift(idx)	(BITS_PER_LONG - (idx))
+#define PG_mask(x, idx)	(x ## UL << PG_shift(idx))
+ /* Cleared when the owning guest 'frees' this page. */
+#define PGC_allocated       PG_mask(1, 1)
+ /* Page is Xen heap? */
+#define PGC_xen_heap        PG_mask(1, 2)
+ /* Page is broken? */
+#define PGC_broken          PG_mask(1, 7)
+ /* Mutually-exclusive page states: { inuse, offlining, offlined, free }. */
+#define PGC_state           PG_mask(3, 9)
+#define PGC_state_inuse     PG_mask(0, 9)
+#define PGC_state_offlining PG_mask(1, 9)
+#define PGC_state_offlined  PG_mask(2, 9)
+#define PGC_state_free      PG_mask(3, 9)
+#define page_state_is(ci, st) (((ci)&PGC_state) == PGC_state_##st)
+
+ /* Count of references to this frame. */
+#define PGC_count_width   PG_shift(9)
+#define PGC_count_mask    ((1UL<<PGC_count_width)-1)
+
+/*
  * Memory flags
  */
 #define MEMORY_PAGETABLE_4L	(1 << 0)
--- a/makedumpfile.c
+++ b/makedumpfile.c
@@ -5719,7 +5719,7 @@ is_select_domain(unsigned int id)
 }
 
 int
-exclude_xen_user_domain(void)
+exclude_xen3_user_domain(void)
 {
 	int i;
 	unsigned int count_info, _domain;
@@ -5728,9 +5728,6 @@ exclude_xen_user_domain(void)
 	unsigned long long phys_start, phys_end;
 	unsigned long long pfn, pfn_end;
 	unsigned long long j, size;
-	struct timeval tv_start;
-
-	gettimeofday(&tv_start, NULL);
 
 	/*
 	 * NOTE: the first half of bitmap is not used for Xen extraction
@@ -5781,13 +5778,107 @@ exclude_xen_user_domain(void)
 		}
 	}
 
+	return TRUE;
+}
+
+int
+exclude_xen4_user_domain(void)
+{
+	int i;
+	unsigned long count_info;
+	unsigned int  _domain;
+	unsigned int num_pt_loads = get_num_pt_loads();
+	unsigned long page_info_addr;
+	unsigned long long phys_start, phys_end;
+	unsigned long long pfn, pfn_end;
+	unsigned long long j, size;
+
+	/*
+	 * NOTE: the first half of bitmap is not used for Xen extraction
+	 */
+	for (i = 0; get_pt_load(i, &phys_start, &phys_end, NULL, NULL); i++) {
+
+		print_progress(PROGRESS_XEN_DOMAIN, i, num_pt_loads);
+
+		pfn     = paddr_to_pfn(phys_start);
+		pfn_end = paddr_to_pfn(phys_end);
+		size    = pfn_end - pfn;
+
+		for (j = 0; pfn < pfn_end; pfn++, j++) {
+			print_progress(PROGRESS_XEN_DOMAIN, j + (size * i),
+					size * num_pt_loads);
+
+			page_info_addr = info->frame_table_vaddr + pfn * SIZE(page_info);
+			if (!readmem(VADDR_XEN,
+			      page_info_addr + OFFSET(page_info.count_info),
+		 	      &count_info, sizeof(count_info))) {
+				clear_bit_on_2nd_bitmap(pfn);
+				continue;	/* page_info may not exist */
+			}
+
+			/* delete free and offlined pages
+			 * Note: Broken pages also get offlined
+			 */
+			if (page_state_is(count_info, free) ||
+			    page_state_is(count_info, offlined)) {
+				clear_bit_on_2nd_bitmap(pfn);
+				continue;
+			}
+
+			/* always keep Xen heap pages */
+			if (count_info & PGC_xen_heap)
+				continue;
+
+			/* delete pages not allocated to any domain */
+			if (! (count_info & PGC_allocated)) {
+				clear_bit_on_2nd_bitmap(pfn);
+				continue;
+			}
+
+			if (!readmem(VADDR_XEN,
+			      page_info_addr + OFFSET(page_info._domain),
+			      &_domain, sizeof(_domain))) {
+				ERRMSG("Can't get page_info._domain.\n");
+				return FALSE;
+			}
+
+			/*
+			 * keep:
+			 *  - anonymous (_domain == 0), or
+			 *  - selected domain page
+			 */
+			if (_domain == 0)
+				continue;
+			if (is_select_domain(_domain))
+				continue;
+			clear_bit_on_2nd_bitmap(pfn);
+		}
+	}
+
+	return TRUE;
+}
+
+int
+exclude_xen_user_domain(void)
+{
+	struct timeval tv_start;
+	int ret;
+
+	gettimeofday(&tv_start, NULL);
+
+	if (info->xen_crash_info.com &&
+	    info->xen_crash_info.com->xen_major_version >= 4)
+		ret = exclude_xen4_user_domain();
+	else
+		ret = exclude_xen3_user_domain();
+
 	/*
 	 * print [100 %]
 	 */
-	print_progress(PROGRESS_XEN_DOMAIN, num_pt_loads, num_pt_loads);
+	print_progress(PROGRESS_XEN_DOMAIN, 1, 1);
 	print_execution_time(PROGRESS_XEN_DOMAIN, &tv_start);
 
-	return TRUE;
+	return ret;
 }
 
 int



_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [patch 8/9] Add support for filtering out user pages under Xen4

[Petr Tesarik]

Dne Čt 23. srpna 2012 14:10:21 jste napsal(a):
> Dne Čt 23. srpna 2012 12:00:41 jste napsal(a):
> > Hello Petr,
> > 
> > Actually we didn't need to patch the crash utility for the SLES11 SP1
> > version we use. When I call the crash utility with the vmcore.d0 argument
> > 
> > that your current  makedumpfile version produces, I see:
> > 	WARNING: crashing_cpu not found.
> > 	crash: invalid kernel virtual address: 0  type: "opt_sched,"
> > 	
> > 	crash: cannot read opt_sched,.

Hello Norbert once again,

I've investigated the logic once more, and we should probably rethink what 
we're filtering out for "-X". The manual page says:

-X  Exclude  all  the user domain pages from Xen kdump's VMCORE, and
    extracts the part of xen and domain-0.

So, depending on how I read it, it should either:

 a. throw away anything but dom_xen, dom_io and dom0, or
 b. keep everything except domU

I vote for a, because it should give the smallest dump. Now, all such pages 
will have:

1. PGC_allocated (it is allocated to a domain)
2. one of the 3 known values for _domain

All else only adds unnecessary complexity. But even that does not cover all 
cases. For example the "opt_sched" (actually ops.opt_name) is stored at 
physical address 0x3fa38b28 (PFN 0x3fa38), but frame_table[0x3fa38] is:

struct page_info {
  {
    list = {
      next = 0, 
      prev = 0
    }, 
    up = 0, 
    shr_handle = 0
  }, 
  count_info = 0, 
  u = {
    inuse = {
      type_info = 0
    }, 
    sh = {
      type = 0, 
      pinned = 0, 
      count = 0
    }, 
    free = {
      need_tlbflush = 0 '\000'
    }
  }, 
  v = {
    inuse = {
      _domain = 0
    }, 
    sh = {
      back = 0
    }, 
    free = {
      order = 0
    }
  }, 
  {
    tlbflush_timestamp = 0, 
    {
      nr_validated_ptes = 0, 
      partial_pte = 0 '\000'
    }, 
    shadow_flags = 0, 
    next_shadow = 0
  }
}

This looks pretty uninitialized to me. Your patch keeps this page, because you 
don't filter out anything that has PGC_allocated==0 and PGC_state_inuse 
(numerically 0), but I'm not sure that's right.

I'll try asking on the Xen mailing list about the intended hierarchy of the 
various PGC_xxx flags.

Petr Tesarik
SUSE Linux

_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec