AHCI: Remove an unnecessary flush from ahci_qc_issue
Matthew Wilcox [Sun, 6 Jul 2008 13:23:20 +0000 (09:23 -0400)]
In an I/O heavy workload (IOZone), ahci_qc_issue is the second-highest
consumer of CPU cycles.  Removing the flush gets us approximately 10%
bandwidth improvement.  I believe this to be because the CPU can start
queueing the next request instead of waiting for the readl() to flush the
writes to the device.  The flush isn't necessary because we're using a
'queue' metaphor; we don't guarantee the command has got to the device,
nor do we need to guarantee the command has got to the controller.

Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>

drivers/ata/ahci.c

index 4ff3f03..dc7596f 100644 (file)
@@ -2136,7 +2136,6 @@ static unsigned int ahci_qc_issue(struct ata_queued_cmd *qc)
        if (qc->tf.protocol == ATA_PROT_NCQ)
                writel(1 << qc->tag, port_mmio + PORT_SCR_ACT);
        writel(1 << qc->tag, port_mmio + PORT_CMD_ISSUE);
-       readl(port_mmio + PORT_CMD_ISSUE);      /* flush */
 
        ahci_sw_activity(qc->dev->link);