RFC: change swap_map to be 32bits varible instead of 16

Hi,

When i backported Hugh patches into the rhel6 kernel today, I noticed during my testing that at very high load of swap tests i get the following error:

Dec 16 17:06:25 dhcp-1-211 kernel: swap_dup: swap entry overflow
Dec 16 17:06:25 dhcp-1-211 kernel: swap_dup: swap entry overflow
Dec 16 17:06:25 dhcp-1-211 kernel: swap_dup: swap entry overflow
Dec 16 17:06:25 dhcp-1-211 kernel: swap_dup: swap entry overflow

The problem probably happen due to the swap_map limitation of being able to address just ~128mb of memory, and with the zero_page mapped when using ksm much more than this amount of memory it was triggered

There may be many soultions to this problem, and I send for RFC the easiest one (just increase the map_count to be unsiged int and allow ~8terabyte of memory)

Thanks.

From 9c661a87c6583531560aaac6a4724df254a6e49b Mon Sep 17 00:00:00 2001
From: Izik Eidus <[email protected]>
Date: Wed, 16 Dec 2009 19:48:43 +0200
Subject: [PATCH] RFC: change swap_map to be 32bits varible instead of 16

Right now after 15bits of usage ~128mb the swap_map will overflow.

While it might never been a problem before with KSM it might just happen due to pages such the zero_page that can be mapped many times.

This patch address this problem by increasing the swap_map to be 32bits and effectivly allow usage of 31bits of that varible (allow ~8Terabyte)

---
 include/linux/swap.h |    8 ++++----
 mm/swapfile.c        |   14 +++++++-------
 2 files changed, 11 insertions(+), 11 deletions(-)

index 4ec9001..34ac29a 100644
@@ -151,9 +151,9 @@ enum {
 
 #define SWAP_CLUSTER_MAX 32
 
-#define SWAP_MAP_MAX	0x7ffe
-#define SWAP_MAP_BAD	0x7fff
-#define SWAP_HAS_CACHE  0x8000		/* There is a swap cache

of entry. */ +#define SWAP_MAP_MAX 0x7ffffffe
+#define SWAP_MAP_BAD 0x7fffffff
+#define SWAP_HAS_CACHE 0x80000000 /* There is a swap cache of
entry. */ #define SWAP_COUNT_MASK (~SWAP_HAS_CACHE)
/*
* The in-memory structure used to track swap areas.
@@ -166,7 +166,7 @@ struct swap_info_struct {
struct block_device *bdev;
struct list_head extent_list;
struct swap_extent *curr_swap_extent;
- unsigned short *swap_map;
+ unsigned int *swap_map;
unsigned int lowest_bit;
unsigned int highest_bit;
unsigned int lowest_alloc; /* while preparing discard
cluster */ diff --git a/mm/swapfile.c b/mm/swapfile.c index c6d5bfd..b6e9b8b 100644

@@ -61,19 +61,19 @@ enum {
 	SWAP_CACHE,	/* ops for reference from swap cache */
 };
 
-static inline int swap_count(unsigned short ent)
+static inline int swap_count(unsigned int ent)
 {
 	return ent & SWAP_COUNT_MASK;
 }
 
-static inline bool swap_has_cache(unsigned short ent)
+static inline bool swap_has_cache(unsigned int ent)
 {
 	return !!(ent & SWAP_HAS_CACHE);
 }
 
-static inline unsigned short encode_swapmap(int count, bool has_cache)
+static inline unsigned int encode_swapmap(unsigned int count, bool

has_cache) {
- unsigned short ret = count;
+ unsigned int ret = count;

if (has_cache)
return SWAP_HAS_CACHE | ret;
@@ -1519,7 +1519,7 @@ out:
SYSCALL_DEFINE1(swapoff, const char __user *, specialfile) {
struct swap_info_struct * p = NULL;
- unsigned short *swap_map;
+ unsigned int *swap_map;
struct file *swap_file, *victim;
struct address_space *mapping;
struct inode *inode;
@@ -1941,13 +1941,13 @@ SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags) goto bad_swap;

/* OK, set up the swap map and apply the bad block list */
- swap_map = vmalloc(maxpages * sizeof(short));
+ swap_map = vmalloc(maxpages * sizeof(unsigned int));
if (!swap_map) {
error = -ENOMEM;
goto bad_swap;
}

- memset(swap_map, 0, maxpages * sizeof(short));
+ memset(swap_map, 0, maxpages * sizeof(unsigned int));
for (i = 0; i < swap_header->info.nr_badpages; i++) {
int page_nr = swap_header->info.badpages[i]; if (page_nr <= 0 || page_nr >=
swap_header->info.last_page) {

1.6.5.2

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

Hi Izik,

The problem here is that you've backported too little: there's a group of 9 "swap_info" patches, before the "mm" patches which prepare for ksm swapping, and the "ksm" swapping patches themselves.

I did include the swap_info patches in the latter rollup I sent you privately, and did highlight this issue when I sent an earlier rollup:
it was an amusing surprise to me that KSM suddenly required our years old bad assumptions in swapoff to be fixed in a hurry.

But I didn't Cc you on them when I sent to Andrew, mistakenly thinking that they weren't "KSM enough" to be of interest you - I hadn't realized that you were planning a backport, sorry.

Maybe you should also include the set of patches which reintroduce the zero page (which won't be swapped and won't be inspected by KSM, being not PageAnon); but that wouldn't be sufficient in itself, since I found it very easy for KSM to overflow the unsigned short *swap_map even with non-zero pages.

Or, dare I say it, maybe you should just use 2.6.33?

The patch you sent as RFC, changing from unsigned short to unsigned int *swap_map: that may be sufficient - I admit it's a very much smaller patch than my lot - I'm not certain. But it's not the way I wanted mainline to go, since most people will never use more than one byte of your 32-bit swap map elements, and vmalloc space may be at a premium on 32-bit architectures.

Hugh

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

Ok, that expline.

Yes I have backported by mails I got from the final summbit...

Not think is possible (Rhel 6 schdule...) :)

No problem, I will just backport your patch`s, no point in making rhel 6 bheave diffrently than what in mainline...

Thanks.

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