From: Nick Piggin <piggin@cyberone.com.au>
Large TCQ depths can potentially keep a lot or all requests tied up.
This can cause merging and batching to stop working completely. The
patch simply ensures nr_requests is >= 2 * tcq depth.
Jens says "I still greatly prefer the old logic, adjust the depth down
instead."
---
25-akpm/drivers/block/ll_rw_blk.c | 8 ++++----
1 files changed, 4 insertions(+), 4 deletions(-)
diff -puN drivers/block/ll_rw_blk.c~scale-nr_requests drivers/block/ll_rw_blk.c
--- 25/drivers/block/ll_rw_blk.c~scale-nr_requests 2004-03-16 13:17:35.924436152 -0800
+++ 25-akpm/drivers/block/ll_rw_blk.c 2004-03-16 13:17:35.930435240 -0800
@@ -530,10 +530,10 @@ init_tag_map(request_queue_t *q, struct
{
int bits, i;
- if (depth > q->nr_requests * 2) {
- depth = q->nr_requests * 2;
- printk(KERN_ERR "%s: adjusted depth to %d\n",
- __FUNCTION__, depth);
+ if (depth > q->nr_requests / 2) {
+ q->nr_requests = depth * 2;
+ printk(KERN_INFO "%s: large TCQ depth: adjusted nr_requests "
+ "to %lu\n", __FUNCTION__, q->nr_requests);
}
tags->tag_index = kmalloc(depth * sizeof(struct request *), GFP_ATOMIC);
_