Skip to contentSkip to conversation

Fix spurious BUG_ON() in mark_bootmem()

PATCH In review Lee Schermerhorn 2 participants 2 messages Jul 9, 2008 Memory management

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Fix spurious BUG_ON() in mark_bootmem()

Against: 2.6.26-rc8-mm1

Fixes problem introduced by patches:

bootmem-factor-out-the-marking-of-a-pfn-range.patch bootmem-replace-node_boot_start-in-struct-bootmem_data.patch

HP ia64 NUMA platform fails to boot 26-rc8-mm1, hitting BUG_ON() in mm/bootmem.c:mark_bootmem().

After linking all bootmem chunks, the 'bdata_list' on HP ia64 numa platforms looks something like this:

        node 4:  0x0-0x8000
        node 0:  0x1c008000-0x1c07ec00
        node 1:  0x1c800000-0x1c87f000
        node 2:  0x1d000000-0x1d07f000
        node 3:  0x1d800000-0x1d87f000

[Node 4 is a pseudo-node generated by the platform firmware to contain a configurable amount of zero-based, hardware interleaved memory. 0x8000 pages or 512M is the minimum that can be configured.]

First call to mark_bootmem() [from free_bootmem()] called with:

	start-end:  0x1c008063-0x1c008262, reserve: 0, flags: 0

I.e, NOT in the first chunk of the list.

However, the "if (pos < bdata->node_min_pfn)" in the loop fails to test the start address of the argument range [in 'pos'] against the end of the chunk. So, it treats the range as being in the node 4 chunk. Second time thru' the loop, pos == 0x8000 is < bdata->node_min_pfn and pos != start, so we trip the BUG_ON().

This patch enhances the if condition to skip chunks that do not overlap the argument range, allowing 26-rc8-mm1 to boot on this platform.

 mm/bootmem.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
mm/bootmem.c +2 -1
Index: linux-2.6.26-rc8-mm1/mm/bootmem.c
===================================================================
@@ -299,7 +299,8 @@ static int __init mark_bootmem(unsigned 
 		int err;
 		unsigned long max;
 
-		if (pos < bdata->node_min_pfn) {
+		if (pos < bdata->node_min_pfn ||
+		    pos >= bdata->node_low_pfn) {
 			BUG_ON(pos != start);
 			continue;
 		}
Acked-by on v1 1/1
#1 lore raw

Hi Lee,

Quoted 49 lines · Lee Schermerhorn

Against: 2.6.26-rc8-mm1

Fixes problem introduced by patches:

bootmem-factor-out-the-marking-of-a-pfn-range.patch bootmem-replace-node_boot_start-in-struct-bootmem_data.patch

HP ia64 NUMA platform fails to boot 26-rc8-mm1, hitting BUG_ON() in mm/bootmem.c:mark_bootmem().

After linking all bootmem chunks, the 'bdata_list' on HP ia64 numa platforms looks something like this:

        node 4:  0x0-0x8000
        node 0:  0x1c008000-0x1c07ec00
        node 1:  0x1c800000-0x1c87f000
        node 2:  0x1d000000-0x1d07f000
        node 3:  0x1d800000-0x1d87f000

[Node 4 is a pseudo-node generated by the platform firmware to contain a configurable amount of zero-based, hardware interleaved memory. 0x8000 pages or 512M is the minimum that can be configured.]

First call to mark_bootmem() [from free_bootmem()] called with:

	start-end:  0x1c008063-0x1c008262, reserve: 0, flags: 0

I.e, NOT in the first chunk of the list.

However, the "if (pos < bdata->node_min_pfn)" in the loop fails to test the start address of the argument range [in 'pos'] against the end of the chunk. So, it treats the range as being in the node 4 chunk. Second time thru' the loop, pos == 0x8000 is < bdata->node_min_pfn and pos != start, so we trip the BUG_ON().

This patch enhances the if condition to skip chunks that do not overlap the argument range, allowing 26-rc8-mm1 to boot on this platform.

 mm/bootmem.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

Index: linux-2.6.26-rc8-mm1/mm/bootmem.c
===================================================================

mm/bootmem.c static int __init mark_bootmem(unsigned · line 306

Quoted 9 lines
mm/bootmem.c +2 -1
@@ -299,7 +299,8 @@ static int __init mark_bootmem(unsigned 
 		int err;
 		unsigned long max;
 
-		if (pos < bdata->node_min_pfn) {
+		if (pos < bdata->node_min_pfn ||
+		    pos >= bdata->node_low_pfn) {
 			BUG_ON(pos != start);
 			continue;
 		}

Ah, yeah, this was obviously wrong. Thanks for the fix!

Nice to know that it boots otherwise on such a setup :)

Hannes