Mel Gorman wrote:
On Fri, Mar 20, 2009 at 05:48:25PM -0700, Avantika Mathur wrote:
Mel Gorman wrote:
Fix it up so that --hard can be specified anywhere and I'll be happy.
I've made it so --hard can be specified anywhere with your suggestion, and also
updated man/hugeadm.8. Hopefully this is the final version! thanks for the
comments
thanks Avantika
Growing the hugepage pool with --pool-pages-min can fail due to the
fragmentation within the system but one failure is not a guarantee of
future failure, particularly if a very large amount of memory has to be
reclaimed to satisfy the allocation. This patch adds a --hard flag to
retry allocations multiple times. It initially tries to resize the pool
up to 5 times. It continues to try if progress is being made towards
the resize.
Signed-off-by: Avantika Mathur <mat...@us.ibm.com>
---
Index: libhugetlbfs/hugeadm.c
===================================================================
--- libhugetlbfs.orig/hugeadm.c 2009-03-20 17:10:02.000000000 -0700
+++ libhugetlbfs/hugeadm.c 2009-03-23 22:56:53.000000000 -0700
@@ -67,6 +67,9 @@
OPTION("--list-all-mounts", "List all current hugetlbfs mount points");
OPTION("--pool-list", "List all pools");
+ OPTION("--hard", "specified with --pool-pages-min to make");
+ CONT("multiple attempts at adjusting the pool size to the");
+ CONT("specified count on failure");
OPTION("--pool-pages-min <size>:[+|-]<count>", "");
CONT("Adjust pool 'size' lower bound");
OPTION("--pool-pages-max <size>:[+|-]<count>", "");
@@ -97,6 +100,7 @@
}
int opt_dry_run = 0;
+int opt_hard = 0;
/*
* getopts return values for options which are long only.
@@ -106,6 +110,8 @@
#define LONG_POOL_MIN_ADJ (LONG_POOL|'m')
#define LONG_POOL_MAX_ADJ (LONG_POOL|'M')
+#define LONG_HARD ('h' << 8)
+
#define LONG_PAGE ('P' << 8)
#define LONG_PAGE_SIZES (LONG_PAGE|'s')
#define LONG_PAGE_AVAIL (LONG_PAGE|'a')
@@ -465,6 +471,7 @@
unsigned long min;
unsigned long max;
+ unsigned long last_pool_value;
/* Extract the pagesize and adjustment. */
page_size_str = strtok_r(cmd, ":", &iter);
@@ -518,16 +525,38 @@
INFO("setting HUGEPAGES_OC to %ld\n", (max - min));
set_huge_page_counter(page_size, HUGEPAGES_OC, (max - min));
}
- if (pools[pos].minimum != min) {
- INFO("setting HUGEPAGES_TOTAL to %ld\n", min);
+
+ if (opt_hard)
+ cnt = 5;
+ else
+ cnt = -1;
+
+ INFO("setting HUGEPAGES_TOTAL to %ld\n", min);
+ set_huge_page_counter(page_size, HUGEPAGES_TOTAL, min);
+ get_pool_size(page_size, &pools[pos]);
+
+ /* If we fail to make an allocation, retry if user requests */
+ last_pool_value = pools[pos].minimum;
+ while ((pools[pos].minimum != min) && (cnt > 0)) {
+ /* Make note if progress is being made and sleep for IO */
+ if (last_pool_value == pools[pos].minimum)
+ cnt--;
+ else
+ cnt = 5;
+ sleep(6);
+
+ last_pool_value = pools[pos].minimum;
+ INFO("Retrying allocation HUGEPAGES_TOTAL to %ld current %ld\n",
+ min, pools[pos].minimum);
set_huge_page_counter(page_size, HUGEPAGES_TOTAL, min);
+ get_pool_size(page_size, &pools[pos]);
}
+
/*
* HUGEPAGES_TOTAL is not guarenteed to check to exactly the figure
* requested should there be insufficient pages. Check the new
* value and adjust HUGEPAGES_OC accordingly.
*/
- get_pool_size(page_size, &pools[pos]);
if (pools[pos].minimum != min) {
WARNING("failed to set pool minimum to %ld became %ld\n",
min, pools[pos].minimum);
@@ -566,6 +595,7 @@
char opts[] = "+hd";
char base[PATH_MAX];
+ char * opt_min_adj = NULL;
int ret = 0, index = 0;
struct option long_opts[] = {
{"help", no_argument, NULL, 'h'},
@@ -574,6 +604,7 @@
{"pool-list", no_argument, NULL, LONG_POOL_LIST},
{"pool-pages-min", required_argument, NULL, LONG_POOL_MIN_ADJ},
{"pool-pages-max", required_argument, NULL, LONG_POOL_MAX_ADJ},
+ {"hard", no_argument, NULL, LONG_HARD},
{"create-mounts", no_argument, NULL, LONG_CREATE_MOUNTS},
{"create-user-mounts", required_argument, NULL, LONG_CREATE_USER_MOUNTS},
{"create-group-mounts", required_argument, NULL, LONG_CREATE_GROUP_MOUNTS},
@@ -619,6 +650,10 @@
case -1:
break;
+ case LONG_HARD:
+ opt_hard = 1;
+ continue;
+
case LONG_LIST_ALL_MOUNTS:
mounts_list_all();
break;
@@ -628,10 +663,7 @@
break;
case LONG_POOL_MIN_ADJ:
- if (! kernel_has_overcommit())
- pool_adjust(optarg, POOL_BOTH);
- else
- pool_adjust(optarg, POOL_MIN);
+ opt_min_adj = optarg;
break;
case LONG_POOL_MAX_ADJ:
@@ -679,6 +711,14 @@
if (ret != -1)
ops++;
}
+
+ if (opt_min_adj != NULL) {
+ if (! kernel_has_overcommit())
+ pool_adjust(opt_min_adj, POOL_BOTH);
+ else
+ pool_adjust(opt_min_adj, POOL_MIN);
+ }
+
index = optind;
if ((argc - index) != 0 || ops == 0) {
Index: libhugetlbfs/man/hugeadm.8
===================================================================
--- libhugetlbfs.orig/man/hugeadm.8 2009-03-23 22:41:06.000000000 -0700
+++ libhugetlbfs/man/hugeadm.8 2009-03-23 22:57:15.000000000 -0700
@@ -117,6 +117,15 @@
the number of huge pages requested by applications is between the Minimum and
Maximum pool sizes.
+.TP
+.B --hard
+
+
+This option is specified with --pool-pages-min to retry allocations multiple
+times on failure to allocate the desired count of pages. It initially tries
+to resize the pool up to 5 times and continues to try if progress is being
+made towards the resize.
+
.SH SEE ALSO
.I oprofile(1),
.I pagesize(1),
------------------------------------------------------------------------------
Apps built with the Adobe(R) Flex(R) framework and Flex Builder(TM) are
powering Web 2.0 with engaging, cross-platform capabilities. Quickly and
easily build your RIAs with Flex Builder, the Eclipse(TM)based development
software that enables intelligent coding and step-through debugging.
Download the free 60 day trial. http://p.sf.net/sfu/www-adobe-com
_______________________________________________
Libhugetlbfs-devel mailing list
Libhugetlbfs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libhugetlbfs-devel
