@@ -2262,6 +2262,54 @@ static const struct seq_operations extfrag_sops = {
DEFINE_SEQ_ATTRIBUTE(extfrag);
+static ssize_t read_extfrag_pct(struct file *file, char __user *user_buf,
+ size_t count, loff_t *ppos)
+{
+ char buf[32];
+ unsigned int len;
+ pg_data_t *pgdat;
+ struct zone *zone;
+ unsigned long flags;
+ unsigned int order;
+ unsigned int num_pgdats = 0;
+ int index_total = 0;
+
+ for_each_online_pgdat(pgdat) {
+ int index_pgt = 0;
+ int num_zones = 0;
+ num_pgdats++;
+ for_each_populated_zone_pgdat(zone, pgdat) {
+ num_zones++;
+ int index = 0;
+
+ spin_lock_irqsave(&zone->lock, flags);
+ for (order = 0; order < NR_PAGE_ORDERS; ++order) {
+ index += fragmentation_index(zone, order);
+ }
+ spin_unlock_irqrestore(&zone->lock, flags);
+
+ index_pgt += index / NR_PAGE_ORDERS;
+ }
+
+ BUG_ON(!num_zones);
+
+ index_total += index_pgt / num_zones;
+ }
+
+ index_total = index_total / num_pgdats;
+
+ len = sprintf(buf, "%d.%02d\n", index_total / 10, index_total % 10);
+
+ return simple_read_from_buffer(user_buf, count, ppos, buf, len);
+}
+
+static const struct file_operations extfrag_pct_fops = {
+ .read = read_extfrag_pct,
+ .open = simple_open,
+ .owner = THIS_MODULE,
+ .llseek = default_llseek,
+};
+
static int __init extfrag_debug_init(void)
{
struct dentry *extfrag_debug_root;
@@ -2274,6 +2322,9 @@ static int __init extfrag_debug_init(void)
debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
&extfrag_fops);
+ debugfs_create_file("extfrag_pct", 0444, extfrag_debug_root, NULL,
+ &extfrag_pct_fops);
+
return 0;
}
In considering the general impact to memory fragmentation over new features, or enhancehments I had asked what metric we could use which is a single digit value. John Hubbard provided one [0] however we'd need to tally up used folios per order as well, and tallying this up would be expensive today. The value would also only tell us how memory fragmented a system is. We can instead just use the existing fragmentation index but generalize a single value from it. This tells us more, when generalized to one value it can tell us both how likely memory allocations might fail due to external fragmention and how likely we are to fail allocations due to low memory. Today we expose an external fragmentation index per node and per zone, per each supported order. This value is useful to tell us how externally fragmented a system might be with the full scope of the CPU topology. Obviously, the CPU topology can vary per system and per architecture, for instance two separate x86 systems may have: cat /sys/kernel/debug/extfrag/extfrag_index Node 0, zone DMA -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 Node 0, zone DMA32 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 Node 0, zone Normal -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 Node 1, zone Normal -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 cat /sys/kernel/debug/extfrag/extfrag_index Node 0, zone DMA -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 Node 0, zone DMA32 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 Node 0, zone Normal -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 -1.000 The number of zones we have may also change over time. This puts a bit of onus on userspace if all it wants is a general sense of how externally fragmented a system is, overall. Provide a simple one unit average for the fragmentation index to allow to simplify measurements in userspace. To make it easier for humans to grok, adjust it to be a value between -100 (allocations failing due to lack of memory) to 100 (super fragmented). [0] https://lore.kernel.org/all/5ac6a387-0ca7-45ca-bebc-c3bdd48452cb@nvidia.com/T/#u Signed-off-by: Luis Chamberlain <mcgrof@kernel.org> --- mm/vmstat.c | 51 +++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 51 insertions(+)