Hi,
We are running an embedded linux (2.4.19) with no swap
sapce. And we haven't applied any rmap yet.
Currely when the system runs low (reached the
pages_low water-mark), kswapd kicks in. But soon after
kswapd kicks in, the out_of_memory is called due to
swapd not being able to reclaim enough pages from the
page cache from each run of it. I don't understand why
sometimes swapd can't reclaim the SWAP_CLUSTER_MAX
number of pages.
We know there are still some number of cache pages
that can be reclaimed if we run the following c
program:
int main() {
int i, j;
char *tmp[10000]
for (i=0; i<NUMBER_OF_PAGES_LEFT_BEFORE_pages_low;
i++) tmp[i]=malloc(4k);
for (i=0; i<NUMBER_OF_PAGES_LEFT_BEFORE_pages_low;
i++) free(tmp[i]);
}
To solve this problem, I am trying to make a
conditional call to the out_of_memory() -- only if the
pages_min is reached, then the out_of_memory() can be
called.
If we can solve the swapd livelock problem, is this a
feasible solution? If you have a better idea, can you
please share with us?
Any input will be greatly appreciated.
Thanks very much in advance! Please include my email
address in your reply.
-Jason
__________________________________________________
Do you Yahoo!?
Yahoo! Web Hosting - establish your business online
http://webhosting.yahoo.com
Jason Li wrote:
>Hi,
>
>We are running an embedded linux (2.4.19) with no swap
>sapce. And we haven't applied any rmap yet.
>
I believe you may be seeing the same issue I brought up in a previous
email. Have you tried what I suggested? I've appended the email to this
message with out the test code.
Ross
I've verified this in 2.4.21-pre5 by code inspection and can trigger the
problem on 2.4.18. It appears to have been fixed in 2.5.
The folowing code vmscan.c assumes that there is available swap space.
/*
* this is the non-racy check for busy page.
*/
if (!page->mapping || !is_page_cache_freeable(page)) {
spin_unlock(&pagecache_lock);
UnlockPage(page);
page_mapped:
if (--max_mapped >= 0)
continue;
/*
* Alert! We've found too many mapped pages on the
* inactive list, so we start swapping out now!
*/
spin_unlock(&pagemap_lru_lock);
swap_out(priority, gfp_mask, classzone);
return nr_pages;
}
If there is no swap space, then unfreeable pages are left on the
inactive queue and the vmtree is walked rather than going through the
rest of the inactive queue. I believe something like
/*
* this is the non-racy check for busy page.
*/
if (!page->mapping || !is_page_cache_freeable(page)) {
spin_unlock(&pagecache_lock);
UnlockPage(page);
page_mapped:
/* If we don't have any swap space left, there
is no reason to worry about pages that do
not have swap associated with them, there
is nothing we can do about it. */
if (!page->mapping && !swap_avail()) {
/* Let's make the page active since we
cannot swap it out. It get's it off
the inactive list. */
spin_unlock(&pagemap_lru_lock);
activate_page(page);
ClearPageReferenced(page);
spin_lock(&pagemap_lru_lock);
continue;
}
if (--max_mapped >= 0)
continue;
/*
* Alert! We've found too many mapped pages on the
* inactive list, so we start swapping out now!
*/
spin_unlock(&pagemap_lru_lock);
swap_out(priority, gfp_mask, classzone);
return nr_pages;
}
will work better when there is no swap space available. If this change
is made, it may also be necessary to limit refill_inactive to prevent it
from using too much cpu. This bug can be triggered with the attached
code and the correct parameters. In particular on a 3 gigabyte machine
with no swap,
for i in $(seq 0 9); do dd if=/dev/zero of=file$i bs=1024k count=512; done
killmm 1032735283 2 9
Usually causes an out of memory error when there is hundreds of
megabytes of cache.
Ross