Skip to contentSkip to conversation

sata_mv.c :: three bug fixes

PATCH In review Mark Lord 4 participants 4 messages Mar 29, 2006 – Mar 30, 2006

Subsystems (2)
lore
Share & tools
lore

Mark Lord

author
#0 lore raw

PATCH 1/1 sata_mv.c :: three bug fixes

This patch is against 2.6.16, and fixes three serious bugs in sata_mv.c. The same fixes are applicable to 2.6.16-git*.

(1) A DMA transfer size of 0x10000 was not being written as 0x0000 in the PRDs. Fixed.

(1) The DEV_IRQ interrupt cause bit happens spuriously during EDMA operation, and was not being ignored by the driver. This led to various "drive busy" errors being reported, with associated unpredictable behaviour. Fixed.

(2) If a SATA or PCI interrupt was received with no outstanding command, the interrupt handler still attempted to invoke ata_qc_complete(), triggering assert()/BUG_ON() behaviour elsewhere in libata. Fixed.

The driver still has issues with confusion after error-recovery, but should now be reliable in the absence of drive errors. I will be looking more into the error-handling bugs next.

---

drivers/scsi/sata_mv.c +23 -18
@@ -906,25 +906,25 @@
 
 		addr = sg_dma_address(sg);
 		sg_len = sg_dma_len(sg);
 
 		while (sg_len) {
 			offset = addr & MV_DMA_BOUNDARY;
 			len = sg_len;
 			if ((offset + sg_len) > 0x10000)
 				len = 0x10000 - offset;
 
 			pp->sg_tbl[i].addr = cpu_to_le32(addr & 0xffffffff);
 			pp->sg_tbl[i].addr_hi = cpu_to_le32((addr >> 16) >> 16);
-			pp->sg_tbl[i].flags_size = cpu_to_le32(len);
+			pp->sg_tbl[i].flags_size = cpu_to_le32(len & 0xffff);
 
 			sg_len -= len;
 			addr += len;
 
 			if (!sg_len && ata_sg_is_last(sg, qc))
 				pp->sg_tbl[i].flags_size |= cpu_to_le32(EPRD_FLAG_END_OF_TBL);
 
 			i++;
 		}
 	}
 }
 
@@ -1178,83 +1178,88 @@
  *      port error ints are reported in the higher level main
  *      interrupt status register and thus are passed in via the
  *      'relevant' argument.
  *
  *      LOCKING:
  *      Inherited from caller.
  */
 static void mv_host_intr(struct ata_host_set *host_set, u32 relevant,
 			 unsigned int hc)
 {
 	void __iomem *mmio = host_set->mmio_base;
 	void __iomem *hc_mmio = mv_hc_base(mmio, hc);
-	struct ata_port *ap;
 	struct ata_queued_cmd *qc;
 	u32 hc_irq_cause;
 	int shift, port, port0, hard_port, handled;
 	unsigned int err_mask;
 	u8 ata_status = 0;
 
 	if (hc == 0) {
 		port0 = 0;
 	} else {
 		port0 = MV_PORTS_PER_HC;
 	}
 
 	/* we'll need the HC success int register in most cases */
 	hc_irq_cause = readl(hc_mmio + HC_IRQ_CAUSE_OFS);
 	if (hc_irq_cause) {
 		writelfl(~hc_irq_cause, hc_mmio + HC_IRQ_CAUSE_OFS);
 	}
 
 	VPRINTK("ENTER, hc%u relevant=0x%08x HC IRQ cause=0x%08x\n",
 		hc,relevant,hc_irq_cause);
 
 	for (port = port0; port < port0 + MV_PORTS_PER_HC; port++) {
-		ap = host_set->ports[port];
+		struct ata_port *ap = host_set->ports[port];
+		struct mv_port_priv *pp = ap->private_data;
 		hard_port = port & MV_PORT_MASK;	/* range 0-3 */
 		handled = 0;	/* ensure ata_status is set if handled++ */
 
-		if ((CRPB_DMA_DONE << hard_port) & hc_irq_cause) {
-			/* new CRPB on the queue; just one at a time until NCQ
-			 */
-			ata_status = mv_get_crpb_status(ap);
-			handled++;
-		} else if ((DEV_IRQ << hard_port) & hc_irq_cause) {
-			/* received ATA IRQ; read the status reg to clear INTRQ
-			 */
-			ata_status = readb((void __iomem *)
+		/* Note that DEV_IRQ might happen spuriously during EDMA,
+		 * and should be ignored in such cases.  We could mask it,
+		 * but it's pretty rare and may not be worth the overhead.
+		 */ 
+		if (pp->pp_flags & MV_PP_FLAG_EDMA_EN) {
+			/* EDMA: check for response queue interrupt */
+			if ((CRPB_DMA_DONE << hard_port) & hc_irq_cause) {
+				ata_status = mv_get_crpb_status(ap);
+				handled = 1;
+			}
+		} else {
+			/* PIO: check for device (drive) interrupt */
+			if ((DEV_IRQ << hard_port) & hc_irq_cause) {
+				ata_status = readb((void __iomem *)
 					   ap->ioaddr.status_addr);
-			handled++;
+				handled = 1;
+			}
 		}
 
-		if (ap &&
-		    (ap->flags & (ATA_FLAG_PORT_DISABLED | ATA_FLAG_NOINTR)))
+		if (ap->flags & (ATA_FLAG_PORT_DISABLED | ATA_FLAG_NOINTR))
 			continue;
 
 		err_mask = ac_err_mask(ata_status);
 
 		shift = port << 1;		/* (port * 2) */
 		if (port >= MV_PORTS_PER_HC) {
 			shift++;	/* skip bit 8 in the HC Main IRQ reg */
 		}
 		if ((PORT0_ERR << shift) & relevant) {
 			mv_err_intr(ap);
 			err_mask |= AC_ERR_OTHER;
-			handled++;
+			handled = 1;
 		}
 
-		if (handled && ap) {
+		if (handled) {
 			qc = ata_qc_from_tag(ap, ap->active_tag);
-			if (NULL != qc) {
+			if (qc && (qc->flags & ATA_QCFLAG_ACTIVE)) {
 				VPRINTK("port %u IRQ found for qc, "
 					"ata_status 0x%x\n", port,ata_status);
 				/* mark qc status appropriately */
 				if (!(qc->tf.ctl & ATA_NIEN)) {
 					qc->err_mask |= err_mask;
 					ata_qc_complete(qc);
 				}
 			}
 		}
 	}
 	VPRINTK("EXIT\n");
 }
#1 lore raw
Quoted 10 lines · Mark Lord

This patch is against 2.6.16, and fixes three serious bugs in sata_mv.c. The same fixes are applicable to 2.6.16-git*.

(1) A DMA transfer size of 0x10000 was not being written as 0x0000 in the PRDs. Fixed.

(1) The DEV_IRQ interrupt cause bit happens spuriously during EDMA operation, and was not being ignored by the driver. This led to various "drive busy" errors being reported, with associated unpredictable behaviour. Fixed.

I saw that happening too, thanks for the fix.

Quoted 4 lines

(2) If a SATA or PCI interrupt was received with no outstanding command, the interrupt handler still attempted to invoke ata_qc_complete(), triggering assert()/BUG_ON() behaviour elsewhere in libata. Fixed.

Dito. Thanks again.

Quoted 10 lines

The driver still has issues with confusion after error-recovery, but should now be reliable in the absence of drive errors. I will be looking more into the error-handling bugs next.

---
[...]

drivers/scsi/sata_mv.c +0 -0
 	/* we'll need the HC success int register in most cases */
 	hc_irq_cause = readl(hc_mmio + HC_IRQ_CAUSE_OFS);
 	if (hc_irq_cause) {

drivers/scsi/sata_mv.c line 1213

Quoted 10 lines
drivers/scsi/sata_mv.c +2 -1
 		writelfl(~hc_irq_cause, hc_mmio + HC_IRQ_CAUSE_OFS);
 	}
 
 	VPRINTK("ENTER, hc%u relevant=0x%08x HC IRQ cause=0x%08x\n",
 		hc,relevant,hc_irq_cause);
 
 	for (port = port0; port < port0 + MV_PORTS_PER_HC; port++) {
-		ap = host_set->ports[port];
+		struct ata_port *ap = host_set->ports[port];
+		struct mv_port_priv *pp = ap->private_data;

The original code checks ap for NULL, are you sure it is safe to remove this?

Signature · 2 lines
Dan Aloni
[email protected], [email protected], [email protected], [email protected]
#2 lore raw
Quoted 2 lines · Dan Aloni

The original code checks ap for NULL, are you sure it is safe to remove this?

"ap" is used above in the same function, long before the check for NULL, so the checks lower down are worthless.

Cheers

#3 lore raw
In reply to Mark Lord
Quoted 1 line · Mark Lord

u8 ata_status = 0;

[...]

drivers/scsi/sata_mv.c line 1214

Quoted 5 lines
drivers/scsi/sata_mv.c +2 -1
 	for (port = port0; port < port0 + MV_PORTS_PER_HC; port++) {
-		ap = host_set->ports[port];
+		struct ata_port *ap = host_set->ports[port];
+		struct mv_port_priv *pp = ap->private_data;
 		hard_port = port & MV_PORT_MASK;	/* range 0-3 */

Applied, even though it should be quite obvious that the patch does not apply to 2.6.16 at all; its missing your previous ata_status patch, which the upstream kernel has already included.

Jeff