Andrew,
Your less-requests patch signaled a way-too-many magic numbers alarm
(not the patches fault, of course, but it pointed it out).
Attached patch removes the minimum queue length, maximum queue length,
factor of queue length that is number of batch requests, and the maximum
number of batch request magic numbers and replaces them with defines and
some comments.
Look OK?
Robert Love
Replace magic numbers in block queue init with sexy defines.
drivers/block/ll_rw_blk.c | 32 ++++++++++++++++++++++++--------
1 files changed, 24 insertions(+), 8 deletions(-)
diff -urN linux-2.5.48/drivers/block/ll_rw_blk.c linux/drivers/block/ll_rw_blk.c
--- linux-2.5.48/drivers/block/ll_rw_blk.c 2002-11-17 23:29:22.000000000 -0500
+++ linux/drivers/block/ll_rw_blk.c 2002-11-19 17:59:07.000000000 -0500
@@ -2109,6 +2109,22 @@
__blk_put_request(req->q, req);
}
+/*
+ * The maximum and minimum free requests slots in the queue are
+ * dynamically calculated as a function of total memory. Below is the
+ * upper and lower bound to those calculations. We do not want the
+ * queue too large, as more memory than desired can be under writeback.
+ */
+#define MAX_QUEUE_REQUESTS 128
+#define MIN_QUEUE_REQUESTS 16
+
+/*
+ * Number of requests to batch together is calculated as the queue size
+ * over BATCH_QUEUE_FACTOR. This number is capped at MAX_BATCH_REQUESTS
+ */
+#define BATCH_QUEUE_FACTOR 8
+#define MAX_BATCH_REQUESTS 8
+
int __init blk_dev_init(void)
{
int total_ram = nr_free_pages() << (PAGE_SHIFT - 10);
@@ -2125,14 +2141,14 @@
* We use this many requests for reads, and this many for writes.
*/
queue_nr_requests = (total_ram >> 9) & ~7;
- if (queue_nr_requests < 16)
- queue_nr_requests = 16;
- if (queue_nr_requests > 128)
- queue_nr_requests = 128;
-
- batch_requests = queue_nr_requests / 8;
- if (batch_requests > 8)
- batch_requests = 8;
+ if (queue_nr_requests < MIN_QUEUE_REQUESTS)
+ queue_nr_requests = MIN_QUEUE_REQUESTS;
+ if (queue_nr_requests > MAX_QUEUE_REQUESTS)
+ queue_nr_requests = MAX_QUEUE_REQUESTS;
+
+ batch_requests = queue_nr_requests / BATCH_QUEUE_FACTOR;
+ if (batch_requests > MAX_BATCH_REQUESTS)
+ batch_requests = MAX_BATCH_REQUESTS;
printk("block request queues:\n");
printk(" %d requests per read queue\n", queue_nr_requests);
Robert Love wrote:
>
> Andrew,
>
> Your less-requests patch signaled a way-too-many magic numbers alarm
> (not the patches fault, of course, but it pointed it out).
>
> Attached patch removes the minimum queue length, maximum queue length,
> factor of queue length that is number of batch requests, and the maximum
> number of batch request magic numbers and replaces them with defines and
> some comments.
>
> Look OK?
>
Spose so. Sometime soon these need to be per-queue rather than global,
and set to intelligent defaults by the driver and runtime tunable
via files in /whereveritsmounted.
Joe Perches just pointed out we can make the second `if' an `else if'.
Probably does not affect the object code, but since we are editing this
here now, why not...
Updated patch applied.
Robert Love
Replace magic numbers in block queue init with sexy defines.
Also change an if to an else if in the name of micro-ops.
drivers/block/ll_rw_blk.c | 32 ++++++++++++++++++++++++--------
1 files changed, 24 insertions(+), 8 deletions(-)
diff -urN linux-2.5.48/drivers/block/ll_rw_blk.c linux/drivers/block/ll_rw_blk.c
--- linux-2.5.48/drivers/block/ll_rw_blk.c 2002-11-17 23:29:22.000000000 -0500
+++ linux/drivers/block/ll_rw_blk.c 2002-11-19 17:59:07.000000000 -0500
@@ -2109,6 +2109,22 @@
__blk_put_request(req->q, req);
}
+/*
+ * The maximum and minimum free requests slots in the queue are
+ * dynamically calculated as a function of total memory. Below is the
+ * upper and lower bound to those calculations. We do not want the
+ * queue too large, as more memory than desired can be under writeback.
+ */
+#define MAX_QUEUE_REQUESTS 128
+#define MIN_QUEUE_REQUESTS 16
+
+/*
+ * Number of requests to batch together is calculated as the queue size
+ * over BATCH_QUEUE_FACTOR. This number is capped at MAX_BATCH_REQUESTS
+ */
+#define BATCH_QUEUE_FACTOR 8
+#define MAX_BATCH_REQUESTS 8
+
int __init blk_dev_init(void)
{
int total_ram = nr_free_pages() << (PAGE_SHIFT - 10);
@@ -2125,14 +2141,14 @@
* We use this many requests for reads, and this many for writes.
*/
queue_nr_requests = (total_ram >> 9) & ~7;
- if (queue_nr_requests < 16)
- queue_nr_requests = 16;
- if (queue_nr_requests > 128)
- queue_nr_requests = 128;
-
- batch_requests = queue_nr_requests / 8;
- if (batch_requests > 8)
- batch_requests = 8;
+ if (queue_nr_requests < MIN_QUEUE_REQUESTS)
+ queue_nr_requests = MIN_QUEUE_REQUESTS;
+ else if (queue_nr_requests > MAX_QUEUE_REQUESTS)
+ queue_nr_requests = MAX_QUEUE_REQUESTS;
+
+ batch_requests = queue_nr_requests / BATCH_QUEUE_FACTOR;
+ if (batch_requests > MAX_BATCH_REQUESTS)
+ batch_requests = MAX_BATCH_REQUESTS;
printk("block request queues:\n");
printk(" %d requests per read queue\n", queue_nr_requests);
On Tue, 2002-11-19 at 18:18, Andrew Morton wrote:
> Spose so. Sometime soon these need to be per-queue rather than global,
> and set to intelligent defaults by the driver and runtime tunable
> via files in /whereveritsmounted.
Yah I was thinking of making them sysctl^Wsysfs files.
That can be done in the future, I guess.
Thanks,
Robert Love
On Tue, Nov 19 2002, Robert Love wrote:
> Andrew,
>
> Your less-requests patch signaled a way-too-many magic numbers alarm
> (not the patches fault, of course, but it pointed it out).
>
> Attached patch removes the minimum queue length, maximum queue length,
> factor of queue length that is number of batch requests, and the maximum
> number of batch request magic numbers and replaces them with defines and
> some comments.
>
> Look OK?
No, please leave these alone, testing is on-going in these parts right
now.
--
Jens Axboe
On Wed, 2002-11-20 at 03:44, Jens Axboe wrote:
> No, please leave these alone, testing is on-going in these parts right
> now.
Did you look at the patch? It just pulls the magic numbers out and uses
defines. Should make testing easier.
Robert Love
On Wed, Nov 20 2002, Robert Love wrote:
> On Wed, 2002-11-20 at 03:44, Jens Axboe wrote:
>
> > No, please leave these alone, testing is on-going in these parts right
> > now.
>
> Did you look at the patch? It just pulls the magic numbers out and uses
> defines. Should make testing easier.
Yes I looked at the patch. Considering that the magic numbers are most
likely going away, it's of no use.
--
Jens Axboe