[PATCH 21/22] atari_scsi: Allow can_queue to be increased for Falcon

From: Finn Thain
Date: Mon Mar 14 2016 - 00:42:51 EST


The benefit of limiting can_queue to 1 is that atari_scsi shares the
ST DMA chip more fairly with other drivers (e.g. falcon-ide).

Unfortunately, this can limit SCSI bus utilization. On systems without
IDE, atari_scsi should issue SCSI commands whenever it can arbitrate for
the bus. Make that possible by making can_queue configurable.

Signed-off-by: Finn Thain <fthain@xxxxxxxxxxxxxxxxxxx>

---
drivers/scsi/atari_scsi.c | 83 ++++++++++++----------------------------------
1 file changed, 22 insertions(+), 61 deletions(-)

Index: linux/drivers/scsi/atari_scsi.c
===================================================================
--- linux.orig/drivers/scsi/atari_scsi.c 2016-03-14 15:26:55.000000000 +1100
+++ linux/drivers/scsi/atari_scsi.c 2016-03-14 15:26:56.000000000 +1100
@@ -14,55 +14,23 @@
*
*/

-
-/**************************************************************************/
-/* */
-/* Notes for Falcon SCSI: */
-/* ---------------------- */
-/* */
-/* Since the Falcon SCSI uses the ST-DMA chip, that is shared among */
-/* several device drivers, locking and unlocking the access to this */
-/* chip is required. But locking is not possible from an interrupt, */
-/* since it puts the process to sleep if the lock is not available. */
-/* This prevents "late" locking of the DMA chip, i.e. locking it just */
-/* before using it, since in case of disconnection-reconnection */
-/* commands, the DMA is started from the reselection interrupt. */
-/* */
-/* Two possible schemes for ST-DMA-locking would be: */
-/* 1) The lock is taken for each command separately and disconnecting */
-/* is forbidden (i.e. can_queue = 1). */
-/* 2) The DMA chip is locked when the first command comes in and */
-/* released when the last command is finished and all queues are */
-/* empty. */
-/* The first alternative would result in bad performance, since the */
-/* interleaving of commands would not be used. The second is unfair to */
-/* other drivers using the ST-DMA, because the queues will seldom be */
-/* totally empty if there is a lot of disk traffic. */
-/* */
-/* For this reasons I decided to employ a more elaborate scheme: */
-/* - First, we give up the lock every time we can (for fairness), this */
-/* means every time a command finishes and there are no other commands */
-/* on the disconnected queue. */
-/* - If there are others waiting to lock the DMA chip, we stop */
-/* issuing commands, i.e. moving them onto the issue queue. */
-/* Because of that, the disconnected queue will run empty in a */
-/* while. Instead we go to sleep on a 'fairness_queue'. */
-/* - If the lock is released, all processes waiting on the fairness */
-/* queue will be woken. The first of them tries to re-lock the DMA, */
-/* the others wait for the first to finish this task. After that, */
-/* they can all run on and do their commands... */
-/* This sounds complicated (and it is it :-(), but it seems to be a */
-/* good compromise between fairness and performance: As long as no one */
-/* else wants to work with the ST-DMA chip, SCSI can go along as */
-/* usual. If now someone else comes, this behaviour is changed to a */
-/* "fairness mode": just already initiated commands are finished and */
-/* then the lock is released. The other one waiting will probably win */
-/* the race for locking the DMA, since it was waiting for longer. And */
-/* after it has finished, SCSI can go ahead again. Finally: I hope I */
-/* have not produced any deadlock possibilities! */
-/* */
-/**************************************************************************/
-
+/*
+ * Notes for Falcon SCSI DMA
+ *
+ * The 5380 device is one of several that all share the DMA chip. Hence
+ * "locking" and "unlocking" access to this chip is required.
+ *
+ * Two possible schemes for ST DMA acquisition by atari_scsi are:
+ * 1) The lock is taken for each command separately (i.e. can_queue == 1).
+ * 2) The lock is taken when the first command arrives and released
+ * when the last command is finished (i.e. can_queue > 1).
+ *
+ * The first alternative limits SCSI bus utilization, since interleaving
+ * commands is not possible. The second gives better performance but is
+ * unfair to other drivers needing to use the ST DMA chip. In order to
+ * allow the IDE and floppy drivers equal access to the ST DMA chip
+ * the default is can_queue == 1.
+ */

#include <linux/module.h>
#include <linux/types.h>
@@ -441,6 +409,10 @@ static int falcon_get_lock(struct Scsi_H
if (IS_A_TT())
return 1;

+ if (stdma_is_locked_by(scsi_falcon_intr) &&
+ instance->hostt->can_queue > 1)
+ return 1;
+
if (in_interrupt())
return stdma_try_lock(scsi_falcon_intr, instance);

@@ -774,22 +746,11 @@ static int __init atari_scsi_probe(struc
atari_scsi_reg_write = atari_scsi_falcon_reg_write;
}

- /* The values for CMD_PER_LUN and CAN_QUEUE are somehow arbitrary.
- * Higher values should work, too; try it!
- * (But cmd_per_lun costs memory!)
- *
- * But there seems to be a bug somewhere that requires CAN_QUEUE to be
- * 2*CMD_PER_LUN. At least on a TT, no spurious timeouts seen since
- * changed CMD_PER_LUN...
- *
- * Note: The Falcon currently uses 8/1 setting due to unsolved problems
- * with cmd_per_lun != 1
- */
if (ATARIHW_PRESENT(TT_SCSI)) {
atari_scsi_template.can_queue = 16;
atari_scsi_template.sg_tablesize = SG_ALL;
} else {
- atari_scsi_template.can_queue = 8;
+ atari_scsi_template.can_queue = 1;
atari_scsi_template.sg_tablesize = SG_NONE;
}