]> err.no Git - linux-2.6/commitdiff
sata_mv fix mv_host_intr bug for hc_irq_cause
authorMark Lord <liml@rtr.ca>
Fri, 2 May 2008 06:13:27 +0000 (02:13 -0400)
committerJeff Garzik <jgarzik@redhat.com>
Tue, 6 May 2008 15:37:41 +0000 (11:37 -0400)
Remove the unwanted reads of hc_irq_cause from mv_host_intr(),
thereby removing a bug whereby we were not always reading it when needed..

Signed-off-by: Mark Lord <mlord@pobox.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
drivers/ata/sata_mv.c

index d995e0e15d873265c1cc33b318b145195ff75a92..31e42deb746f60b1b08311bb865b162d396b57c0 100644 (file)
@@ -1818,48 +1818,61 @@ static void mv_process_crpb_entries(struct ata_port *ap, struct mv_port_priv *pp
 static int mv_host_intr(struct ata_host *host, u32 main_irq_cause)
 {
        struct mv_host_priv *hpriv = host->private_data;
-       void __iomem *mmio = hpriv->base, *hc_mmio = NULL;
-       u32 hc_irq_cause = 0;
+       void __iomem *mmio = hpriv->base, *hc_mmio;
        unsigned int handled = 0, port;
 
        for (port = 0; port < hpriv->n_ports; port++) {
                struct ata_port *ap = host->ports[port];
                struct mv_port_priv *pp;
-               unsigned int shift, hardport, port_cause;
-               /*
-                * When we move to the second hc, flag our cached
-                * copies of hc_mmio (and hc_irq_cause) as invalid again.
-                */
-               if (port == MV_PORTS_PER_HC)
-                       hc_mmio = NULL;
-               /*
-                * Do nothing if port is not interrupting or is disabled:
-                */
+               unsigned int p, shift, hardport, port_cause;
+
                MV_PORT_TO_SHIFT_AND_HARDPORT(port, shift, hardport);
-               port_cause = (main_irq_cause >> shift) & (DONE_IRQ | ERR_IRQ);
-               if (!port_cause || !ap || (ap->flags & ATA_FLAG_DISABLED))
-                       continue;
                /*
-                * Each hc within the host has its own hc_irq_cause register.
-                * We defer reading it until we know we need it, right now:
-                *
-                * FIXME later: we don't really need to read this register
-                * (some logic changes required below if we go that way),
-                * because it doesn't tell us anything new.  But we do need
-                * to write to it, outside the top of this loop,
-                * to reset the interrupt triggers for next time.
+                * Each hc within the host has its own hc_irq_cause register,
+                * where the interrupting ports bits get ack'd.
                 */
-               if (!hc_mmio) {
+               if (hardport == 0) {    /* first port on this hc ? */
+                       u32 hc_cause = (main_irq_cause >> shift) & HC0_IRQ_PEND;
+                       u32 port_mask, ack_irqs;
+                       /*
+                        * Skip this entire hc if nothing pending for any ports
+                        */
+                       if (!hc_cause) {
+                               port += MV_PORTS_PER_HC - 1;
+                               continue;
+                       }
+                       /*
+                        * We don't need/want to read the hc_irq_cause register,
+                        * because doing so hurts performance, and
+                        * main_irq_cause already gives us everything we need.
+                        *
+                        * But we do have to *write* to the hc_irq_cause to ack
+                        * the ports that we are handling this time through.
+                        *
+                        * This requires that we create a bitmap for those
+                        * ports which interrupted us, and use that bitmap
+                        * to ack (only) those ports via hc_irq_cause.
+                        */
+                       ack_irqs = 0;
+                       for (p = 0; p < MV_PORTS_PER_HC; ++p) {
+                               if ((port + p) >= hpriv->n_ports)
+                                       break;
+                               port_mask = (DONE_IRQ | ERR_IRQ) << (p * 2);
+                               if (hc_cause & port_mask)
+                                       ack_irqs |= (DMA_IRQ | DEV_IRQ) << p;
+                       }
                        hc_mmio = mv_hc_base_from_port(mmio, port);
-                       hc_irq_cause = readl(hc_mmio + HC_IRQ_CAUSE_OFS);
-                       writelfl(~hc_irq_cause, hc_mmio + HC_IRQ_CAUSE_OFS);
+                       writelfl(~ack_irqs, hc_mmio + HC_IRQ_CAUSE_OFS);
                        handled = 1;
                }
+               port_cause = (main_irq_cause >> shift) & (DONE_IRQ | ERR_IRQ);
+               if (!port_cause)
+                       continue;
                /*
                 * Process completed CRPB response(s) before other events.
                 */
                pp = ap->private_data;
-               if (hc_irq_cause & (DMA_IRQ << hardport)) {
+               if (port_cause & DONE_IRQ) {
                        if (pp->pp_flags & MV_PP_FLAG_EDMA_EN)
                                mv_process_crpb_entries(ap, pp);
                }
@@ -1868,15 +1881,15 @@ static int mv_host_intr(struct ata_host *host, u32 main_irq_cause)
                 */
                if (unlikely(port_cause & ERR_IRQ)) {
                        mv_err_intr(ap);
-               } else if (hc_irq_cause & (DEV_IRQ << hardport)) {
+               } else {
                        if (!(pp->pp_flags & MV_PP_FLAG_EDMA_EN)) {
                                struct ata_queued_cmd *qc = mv_get_active_qc(ap);
                                if (qc) {
                                        ata_sff_host_intr(ap, qc);
                                        continue;
                                }
+                               mv_unexpected_intr(ap);
                        }
-                       mv_unexpected_intr(ap);
                }
        }
        return handled;