[PATCH 0/0] Disable deferred struct page initialisation on Fadump

Fadump kernel reserves large chunks of memory even before the pages are
initialised. This could mean memory that corresponds to several nodes might
fall in memblock reserved regions.

Kernels compiled with CONFIG_DEFERRED_STRUCT_PAGE_INIT will initialise
only certain size memory per node. The certain size takes into account
the dentry and inode cache sizes. However such a kernel when booting a
secondary kernel will not be able to allocate the required amount of
memory to suffice for the dentry and inode caches. This results in
crashes like the below on large systems such as 32 TB systems.

Dentry cache hash table entries: 536870912 (order: 16, 4294967296 bytes)
vmalloc: allocation failure, allocated 4097114112 of 17179934720 bytes
swapper/0: page allocation failure: order:0, mode:0x2080020(GFP_ATOMIC)
CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.6-master+ #3
Call Trace:
[c00000000108fb10] [c0000000007fac88] dump_stack+0xb0/0xf0 (unreliable)
[c00000000108fb50] [c000000000235264] warn_alloc_failed+0x114/0x160
[c00000000108fbf0] [c000000000281484] __vmalloc_node_range+0x304/0x340
[c00000000108fca0] [c00000000028152c] __vmalloc+0x6c/0x90
[c00000000108fd40] [c000000000aecfb0]
alloc_large_system_hash+0x1b8/0x2c0
[c00000000108fe00] [c000000000af7240] inode_init+0x94/0xe4
[c00000000108fe80] [c000000000af6fec] vfs_caches_init+0x8c/0x13c
[c00000000108ff00] [c000000000ac4014] start_kernel+0x50c/0x578
[c00000000108ff90] [c000000000008c6c] start_here_common+0x20/0xa8

This can be solved by two approaches.
1. Disable deferred struct page initialisation on fadump.

2. Detect reserved nodes and allocate accordingly.
 - Detecting nodes whose memblocks are mostly reserved.
 - Allocating extra memory in other nodes in lieu of the nodes whose
   memory is reserved.

This patchset takes the first approach.

Srikar Dronamraju (2):
  mm: Allow disabling deferred struct page initialisation
  fadump: Disable deferred page struct initialisation

 arch/powerpc/kernel/fadump.c |  1 +
 include/linux/mmzone.h       |  2 +-
 mm/page_alloc.c              | 20 ++++++++++++++++++++
 3 files changed, 22 insertions(+), 1 deletion(-)

-- 
1.8.5.6

Kernels compiled with CONFIG_DEFERRED_STRUCT_PAGE_INIT will initialise
only certain size memory per node. The certain size takes into account
the dentry and inode cache sizes. However such a kernel when booting a
secondary kernel will not be able to allocate the required amount of
memory to suffice for the dentry and inode caches. This results in
crashes like the below on large systems such as 32 TB systems.

Dentry cache hash table entries: 536870912 (order: 16, 4294967296 bytes)
vmalloc: allocation failure, allocated 4097114112 of 17179934720 bytes
swapper/0: page allocation failure: order:0, mode:0x2080020(GFP_ATOMIC)
CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.6-master+ #3
Call Trace:
[c00000000108fb10] [c0000000007fac88] dump_stack+0xb0/0xf0 (unreliable)
[c00000000108fb50] [c000000000235264] warn_alloc_failed+0x114/0x160
[c00000000108fbf0] [c000000000281484] __vmalloc_node_range+0x304/0x340
[c00000000108fca0] [c00000000028152c] __vmalloc+0x6c/0x90
[c00000000108fd40] [c000000000aecfb0]
alloc_large_system_hash+0x1b8/0x2c0
[c00000000108fe00] [c000000000af7240] inode_init+0x94/0xe4
[c00000000108fe80] [c000000000af6fec] vfs_caches_init+0x8c/0x13c
[c00000000108ff00] [c000000000ac4014] start_kernel+0x50c/0x578
[c00000000108ff90] [c000000000008c6c] start_here_common+0x20/0xa8

Allow such kernels to disable deferred page struct initialisation.

Signed-off-by: Srikar Dronamraju <redacted>
---
 include/linux/mmzone.h |  2 +-
 mm/page_alloc.c        | 20 ++++++++++++++++++++
 2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index c60df92..1c55200 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h

@@ -1203,7 +1203,7 @@ unsigned long __init node_memmap_size_bytes(int, unsigned long, unsigned long);
 #else
 #define pfn_valid_within(pfn) (1)
 #endif
-
+void disable_deferred_meminit(void);
 #ifdef CONFIG_ARCH_HAS_HOLES_MEMORYMODEL
 /*
  * pfn_valid() is meant to be able to tell if a given PFN has valid memmap

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index c1069ef..dc6ebac 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -301,6 +301,19 @@ static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
 }
 
 /*
+ * Deferred struct page initialisation may not work on a multinode machine,
+ * if a significant amount of memory is reserved at early boot.  Allow apis
+ * that reserve significant memory to disable deferred struct page
+ * initialisation.
+ */
+static bool defer_init_disabled;
+
+void disable_deferred_meminit(void)
+{
+	defer_init_disabled = true;
+}
+
+/*
  * Returns false when the remaining initialisation should be deferred until
  * later in the boot cycle when it can be parallelised.
  */

@@ -313,6 +326,9 @@ static inline bool update_defer_init(pg_data_t *pgdat,
 	/* Always populate low zones for address-contrained allocations */
 	if (zone_end < pgdat_end_pfn(pgdat))
 		return true;
+
+	if (defer_init_disabled)
+		return true;
 	/*
 	 * Initialise at least 2G of a node but also take into account that
 	 * two large system hashes that can take up 1GB for 0.25TB/node.

@@ -350,6 +366,10 @@ static inline bool update_defer_init(pg_data_t *pgdat,
 {
 	return true;
 }
+void disable_deferred_meminit(void)
+{
+}
+
 #endif
 
 

-- 
1.8.5.6

Fadump kernel reserves significant number of memory blocks. On a multi-node
machine, with CONFIG_DEFFERRED_STRUCT_PAGE support, fadump kernel fails to
boot. Fix this by disabling deferred page struct initialisation.

Signed-off-by: Srikar Dronamraju <redacted>
---
 arch/powerpc/kernel/fadump.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/arch/powerpc/kernel/fadump.c b/arch/powerpc/kernel/fadump.c
index 3cb3b02a..117faf2 100644
--- a/arch/powerpc/kernel/fadump.c
+++ b/arch/powerpc/kernel/fadump.c

@@ -318,6 +318,7 @@ int __init fadump_reserve_mem(void)
 				be64_to_cpu(fdm_active->rmr_region.source_len);
 		pr_debug("fadumphdr_addr = %p\n",
 				(void *) fw_dump.fadumphdr_addr);
+		disable_deferred_meminit();
 	} else {
 		/* Reserve the memory at the top of memory. */
 		size = get_fadump_area_size();

-- 
1.8.5.6

On 08/02/2016 06:19 AM, Srikar Dronamraju wrote:

Kernels compiled with CONFIG_DEFERRED_STRUCT_PAGE_INIT will initialise
only certain size memory per node. The certain size takes into account
the dentry and inode cache sizes. However such a kernel when booting a
secondary kernel will not be able to allocate the required amount of
memory to suffice for the dentry and inode caches. This results in
crashes like the below on large systems such as 32 TB systems.

What's a "secondary kernel"?

On Tue, 2016-08-02 at 18:49 +0530, Srikar Dronamraju wrote:

Fadump kernel reserves significant number of memory blocks. On a multi-node
machine, with CONFIG_DEFFERRED_STRUCT_PAGE support, fadump kernel fails to
boot. Fix this by disabling deferred page struct initialisation.

How much memory does a fadump kernel need? Can we bump up the limits depending
on the config. I presume when you say fadump kernel you mean kernel with
FADUMP in the config?

BTW, I would much rather prefer a config based solution that does not select
DEFERRED_INIT if FADUMP is enabled.

Balbir

On 08/03/2016 07:20 AM, Balbir Singh wrote:

On Tue, 2016-08-02 at 18:49 +0530, Srikar Dronamraju wrote:

quoted

Fadump kernel reserves significant number of memory blocks. On a multi-node
machine, with CONFIG_DEFFERRED_STRUCT_PAGE support, fadump kernel fails to
boot. Fix this by disabling deferred page struct initialisation.

How much memory does a fadump kernel need? Can we bump up the limits depending
on the config. I presume when you say fadump kernel you mean kernel with
FADUMP in the config?

BTW, I would much rather prefer a config based solution that does not select
DEFERRED_INIT if FADUMP is enabled.

IIRC the kdump/fadump kernel is typically the same vmlinux as the main 
kernel, just with special initrd and boot params. So if you want 
deferred init for the main kernel, this would be impractical.

Balbir

* Dave Hansen [off-list ref] [2016-08-02 11:09:21]:

On 08/02/2016 06:19 AM, Srikar Dronamraju wrote:

quoted

Kernels compiled with CONFIG_DEFERRED_STRUCT_PAGE_INIT will initialise
only certain size memory per node. The certain size takes into account
the dentry and inode cache sizes. However such a kernel when booting a
secondary kernel will not be able to allocate the required amount of
memory to suffice for the dentry and inode caches. This results in
crashes like the below on large systems such as 32 TB systems.

What's a "secondary kernel"?

I mean the kernel thats booted to collect the crash, On fadump, the
first kernel acts as the secondary kernel i.e the same kernel is booted
to collect the crash.

-- 
Thanks and Regards
Srikar Dronamraju

Vlastimil Babka [off-list ref] writes:

On 08/03/2016 07:20 AM, Balbir Singh wrote:

quoted

On Tue, 2016-08-02 at 18:49 +0530, Srikar Dronamraju wrote:

quoted

Fadump kernel reserves significant number of memory blocks. On a multi-node
machine, with CONFIG_DEFFERRED_STRUCT_PAGE support, fadump kernel fails to
boot. Fix this by disabling deferred page struct initialisation.

How much memory does a fadump kernel need? Can we bump up the limits depending
on the config. I presume when you say fadump kernel you mean kernel with
FADUMP in the config?

BTW, I would much rather prefer a config based solution that does not select
DEFERRED_INIT if FADUMP is enabled.

IIRC the kdump/fadump kernel is typically the same vmlinux as the main 
kernel, just with special initrd and boot params. So if you want 
deferred init for the main kernel, this would be impractical.

Yes. Distros won't build a separate kernel, so it has to work at runtime.

cheers

On 08/02/2016 11:38 PM, Srikar Dronamraju wrote:

* Dave Hansen [off-list ref] [2016-08-02 11:09:21]:

quoted

On 08/02/2016 06:19 AM, Srikar Dronamraju wrote:

quoted

Kernels compiled with CONFIG_DEFERRED_STRUCT_PAGE_INIT will initialise
only certain size memory per node. The certain size takes into account
the dentry and inode cache sizes. However such a kernel when booting a
secondary kernel will not be able to allocate the required amount of
memory to suffice for the dentry and inode caches. This results in
crashes like the below on large systems such as 32 TB systems.

What's a "secondary kernel"?

I mean the kernel thats booted to collect the crash, On fadump, the
first kernel acts as the secondary kernel i.e the same kernel is booted
to collect the crash.

OK, but I'm still not seeing what the problem is.  You've said that it
crashes and that it crashes during inode/dentry cache allocation.

But, *why* does the same kernel image crash in when it is used as a
"secondary kernel"?

* Dave Hansen [off-list ref] [2016-08-03 11:17:43]:

On 08/02/2016 11:38 PM, Srikar Dronamraju wrote:

quoted

* Dave Hansen [off-list ref] [2016-08-02 11:09:21]:

quoted

On 08/02/2016 06:19 AM, Srikar Dronamraju wrote:

quoted

Kernels compiled with CONFIG_DEFERRED_STRUCT_PAGE_INIT will initialise
only certain size memory per node. The certain size takes into account
the dentry and inode cache sizes. However such a kernel when booting a
secondary kernel will not be able to allocate the required amount of
memory to suffice for the dentry and inode caches. This results in
crashes like the below on large systems such as 32 TB systems.

What's a "secondary kernel"?

I mean the kernel thats booted to collect the crash, On fadump, the
first kernel acts as the secondary kernel i.e the same kernel is booted
to collect the crash.

OK, but I'm still not seeing what the problem is.  You've said that it
crashes and that it crashes during inode/dentry cache allocation.

But, *why* does the same kernel image crash in when it is used as a
"secondary kernel"?

I guess you already got it. But let me try to explain it again.

Lets say we have a 32 TB system with 16 nodes each node having 2T of
memory. We are assuming deferred page initialisation is configured.

When the regular kernel boots,
1. It reserves 5% of the memory for fadump.
2. It initializes 8GB per node, i.e 128GB
3. It allocated dentry/inode cache which is around 16GB.
4. It then kicks the parallel page struct initialization.

Now lets say kernel crashed and fadump was triggered.

1. The same kernel boots in the 5% reserved space which is 1600GB
2. It reserves the rest 95% memory.
3. It tries to initialize 8GB per node but can only initialize 8GB.
	(since except for 1st node the rest nodes are all reserved)
4. It tries to allocate dentry/inode cache of 16GB but fails.
	(tries to reclaim but reclaim needs spinlock 
	and spinlock is not yet initialized.)

-- 
Thanks and Regards
Srikar Dronamraju