Skip to contentSkip to conversation

AHCI prefetch

PATCH In review Jeff Garzik 2 participants 5 messages Mar 4, 2006 – Mar 12, 2006

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 AHCI prefetch

This patch has been sitting in my tmp directory for ages.

We should probably turn this on, though the practical difference is probably minimal.

drivers/scsi/ahci.c +3 -1
index 4e96ec5..973a794 100644
@@ -64,8 +64,10 @@ enum {
 	AHCI_PORT_PRIV_DMA_SZ	= AHCI_CMD_SLOT_SZ + AHCI_CMD_TBL_SZ +
 				  AHCI_RX_FIS_SZ,
 	AHCI_IRQ_ON_SG		= (1 << 31),
+
 	AHCI_CMD_ATAPI		= (1 << 5),
 	AHCI_CMD_WRITE		= (1 << 6),
+	AHCI_CMD_PREFETCH	= (1 << 7),
 
 	RX_FIS_D2H_REG		= 0x40,	/* offset of D2H Register FIS data */
 
@@ -533,7 +535,7 @@ static void ahci_qc_prep(struct ata_queu
 	if (qc->tf.flags & ATA_TFLAG_WRITE)
 		opts |= AHCI_CMD_WRITE;
 	if (is_atapi_taskfile(&qc->tf))
-		opts |= AHCI_CMD_ATAPI;
+		opts |= AHCI_CMD_ATAPI | AHCI_CMD_PREFETCH;
 
 	pp->cmd_slot[0].opts = cpu_to_le32(opts);
 	pp->cmd_slot[0].status = 0;
#1 lore raw
Quoted 6 lines · Jeff Garzik

This patch has been sitting in my tmp directory for ages.

We should probably turn this on, though the practical difference is probably minimal.

The patch works okay on my machine (ICH7R) although the patch didn't apply to #upstream. I'm not very sure about this change though.

1. Why apply it only to ATAPI devices? ATA devices can benefit to. If it's because this bit shouldn't be turned on for NCQ, we can turn it on conditionally. We'll probably need similar condition for ATAPI devices too if we support FIS-based PM switching.

2. I'm a bit skeptical whether this change will bring any noticeable performance improvement. OTOH, this seems to be a good source for obscure problems on some controllers which might not implement/test this feature properly. As more controllers implement AHCI spec, the possibility grows.

Anyways, here's the patch regenerated against #upstream.

drivers/scsi/ahci.c +2 -1
index 1c2ab3d..cfa6eaf 100644
@@ -66,6 +66,7 @@ enum {
 	AHCI_IRQ_ON_SG		= (1 << 31),
 	AHCI_CMD_ATAPI		= (1 << 5),
 	AHCI_CMD_WRITE		= (1 << 6),
+	AHCI_CMD_PREFETCH	= (1 << 7),
 	AHCI_CMD_RESET		= (1 << 8),
 	AHCI_CMD_CLR_BUSY	= (1 << 10),
 
@@ -631,7 +632,7 @@ static void ahci_qc_prep(struct ata_queu
 	if (qc->tf.flags & ATA_TFLAG_WRITE)
 		opts |= AHCI_CMD_WRITE;
 	if (is_atapi)
-		opts |= AHCI_CMD_ATAPI;
+		opts |= AHCI_CMD_ATAPI | AHCI_CMD_PREFETCH;
 
 	ahci_fill_cmd_slot(pp, opts);
 }
#2 lore raw
Quoted 9 lines · Tejun Heo
Quoted 5 lines · Jeff Garzik

This patch has been sitting in my tmp directory for ages.

We should probably turn this on, though the practical difference is probably minimal.

Quoted 7 lines

The patch works okay on my machine (ICH7R) although the patch didn't apply to #upstream. I'm not very sure about this change though.

1. Why apply it only to ATAPI devices? ATA devices can benefit to. If it's because this bit shouldn't be turned on for NCQ, we can turn it on conditionally. We'll probably need similar condition for ATAPI devices too if we support FIS-based PM switching.

Main reason is that it will largely only have benefits on ATAPI devices, and I've only tested it on ATAPI devices.

Quoted 2 lines

2. I'm a bit skeptical whether this change will bring any noticeable performance improvement. OTOH, this seems to be a good source for

Agreed.

Quoted 3 lines

obscure problems on some controllers which might not implement/test this feature properly. As more controllers implement AHCI spec, the possibility grows.

Agreed.

Quoted 1 line

Anyways, here's the patch regenerated against #upstream.

Could I trouble you for a resend, with a proper signed-off-by and patch description?

Jeff

#3 lore raw

PATCH 1/1 ahci: enable prefetching for PACKET commands

Turn on AHCI_CMD_PREFETCH for PACKET commands. This hints the controller that it can prefetch the CDB and the PRD entries. This patch is originally from Jeff Garzik.

---

Jeff, I'll send two flavors of this patch. The other one will enable AHCI_CMD_PREFETCH for both ATA and ATAPI devices. I've tested both cases on ICH7R.

 ahci.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/scsi/ahci.c +2 -1
index 00dfdef..e97ab3e 100644
@@ -66,6 +66,7 @@ enum {
 	AHCI_IRQ_ON_SG		= (1 << 31),
 	AHCI_CMD_ATAPI		= (1 << 5),
 	AHCI_CMD_WRITE		= (1 << 6),
+	AHCI_CMD_PREFETCH	= (1 << 7),
 	AHCI_CMD_RESET		= (1 << 8),
 	AHCI_CMD_CLR_BUSY	= (1 << 10),
 
@@ -631,7 +632,7 @@ static void ahci_qc_prep(struct ata_queu
 	if (qc->tf.flags & ATA_TFLAG_WRITE)
 		opts |= AHCI_CMD_WRITE;
 	if (is_atapi)
-		opts |= AHCI_CMD_ATAPI;
+		opts |= AHCI_CMD_ATAPI | AHCI_CMD_PREFETCH;
 
 	ahci_fill_cmd_slot(pp, opts);
 }
#4 lore raw

PATCH 1/1 ahci: enable prefetching

In reply to Jeff Garzik

Enable prefetching. This hints AHCI controller that it can prefetch PRT entries and CDB (for PACKET commands). This is a hint. AHCI controller is free to ignore it or prefetch regardless of this bit.

Note that prefetching bit should be turned off for NCQ commands or FIS-based switching PM support.

---

And this is the second version.

 ahci.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/scsi/ahci.c +2 -1
index 00dfdef..454a393 100644
@@ -66,6 +66,7 @@ enum {
 	AHCI_IRQ_ON_SG		= (1 << 31),
 	AHCI_CMD_ATAPI		= (1 << 5),
 	AHCI_CMD_WRITE		= (1 << 6),
+	AHCI_CMD_PREFETCH	= (1 << 7),
 	AHCI_CMD_RESET		= (1 << 8),
 	AHCI_CMD_CLR_BUSY	= (1 << 10),
 
@@ -627,7 +628,7 @@ static void ahci_qc_prep(struct ata_queu
 	/*
 	 * Fill in command slot information.
 	 */
-	opts = cmd_fis_len | n_elem << 16;
+	opts = AHCI_CMD_PREFETCH | cmd_fis_len | n_elem << 16;
 	if (qc->tf.flags & ATA_TFLAG_WRITE)
 		opts |= AHCI_CMD_WRITE;
 	if (is_atapi)