Skip to contentSkip to conversation

SATA powersave patches

DISCUSSION Pavel Machek 3 participants 9 messages Sep 10, 2006 – Nov 21, 2006 4 messages not in the archive SCSI

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

A message that is not in the archive (sent off-list or not yet received): [email protected]

A message that is not in the archive (sent off-list or not yet received): [email protected]

A message that is not in the archive (sent off-list or not yet received): [email protected]

#4 lore raw

Hi!

(cc-ed to the list)

Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Quoted 2 lines

Do you have SATA powersave patches somewhere handy? I found out that SATA eats way too much power.

Hello, Pavel.

Do you mean link powersaving? Or disk spindown?

Quoted 8 lines
Quoted 2 lines

Link powersaving... disk spindown should be handled by laptop_mode + hdparm -y, right?

Yes for disk spindown. Link powersaving is in the following git tree but it's ~2 months old.

http://htj.dyndns.org/git/?p=libata-tj.git;a=shortlog;h=powersave
git://htj.dyndns.org/libata-tj.git powersave

Thanks... I got it to work (on 2 the old tree, I was not able to forward-port it), but power savings were not too big (~0.1W, maybe).

I'm getting huge (~1W) savings by powering down SATA controller, as in ahci_pci_device_suspend(). It would be great to be able to power SATA controller down, then power it back up when it is needed... I tried following hack, but could not get it to work. Any ideas?

Pavel

drivers/scsi/ahci.c +13 -1
index 3e98334..b6e9f55 100644
@@ -222,7 +222,7 @@ static void ahci_thaw(struct ata_port *a
 static void ahci_error_handler(struct ata_port *ap);
 static void ahci_post_internal_cmd(struct ata_queued_cmd *qc);
 static int ahci_pci_device_suspend(struct pci_dev *pdev, pm_message_t state);
-static int ahci_pci_device_resume(struct pci_dev *pdev);
+ int ahci_pci_device_resume(struct pci_dev *pdev);
 static int ahci_port_standby(void __iomem *port_mmio, u32 cap);
 static int ahci_port_spinup(void __iomem *port_mmio, u32 cap);
 static int ahci_port_suspend(struct ata_port *ap, pm_message_t state);
@@ -1082,6 +1082,9 @@ static unsigned int ahci_fill_sg(struct 
 	return n_sg;
 }
 
+int suspended;
+struct pci_dev *mydev;
+
 int ahci_pci_device_suspend(struct pci_dev *pdev, pm_message_t state)
 {
 	struct ata_host_set *host_set = dev_get_drvdata(&pdev->dev);
@@ -1089,6 +1092,11 @@ int ahci_pci_device_suspend(struct pci_d
 	u32 tmp;
 	int i;
 
+	if (suspended)
+		return;
+	suspended = 1;
+	mydev = pdev;
+
 	/* First suspend all ports */
 	for (i = 0; i < host_set->n_ports; i++) {
 		struct ata_port *ap;
@@ -1121,6 +1129,10 @@ int ahci_pci_device_resume(struct pci_de
 	struct ata_port *ap;
 	u32 i, tmp, irq_stat;
 
+	if (!suspended)
+		return;
+	suspended = 0;
+
 	tmp = readl(mmio + HOST_CTL);
 	if (!(tmp & HOST_AHCI_EN)) {
 		tmp |= HOST_AHCI_EN;
drivers/scsi/libata-scsi.c +8 -0
index e92c31d..89fb997 100644
@@ -2822,6 +2822,14 @@ int ata_scsi_queuecmd(struct scsi_cmnd *
 	struct Scsi_Host *shost = scsidev->host;
 	int rc = 0;
 
+	extern int suspended;
+	extern int ahci_pci_device_suspend(struct pci_dev *pdev, pm_message_t state);
+	extern int ahci_pci_device_resume(struct pci_dev *pdev);
+	extern struct pci_dev *mydev;
+
+	if (suspended)
+		ahci_pci_device_resume(mydev);
+
 	ap = ata_shost_to_port(shost);
 
 	spin_unlock(shost->host_lock);
Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#5 lore raw

Hello, Pavel Machek.

Quoted 5 lines · Pavel Machek

Thanks... I got it to work (on 2 the old tree, I was not able to forward-port it), but power savings were not too big (~0.1W, maybe).

I'm getting huge (~1W) savings by powering down SATA controller, as in ahci_pci_device_suspend().

Yeah, it only turns off SATA PHY, so it doesn't result in huge saving. IIRC, it was somewhere around 5 percent on my notebook w/ static linksave mode (turning PHY off on empty port). But link powersaving introduces virtually no recognizable delay, so it's nice to have.

Can you check if there is any difference between [D/H]IPS and static? ICH6M on my notebook can't do DIPS/HIPS, so I couldn't compare them against static.

Quoted 3 lines

It would be great to be able to power SATA
controller down, then power it back up when it is needed... I tried following hack, but could not get it to work. Any ideas?

1. One way to do it would be by dynamic power management. It would be nice to have wake-up mechanism at the block layer. Idle timer can run in the block layer or it can be implemented in the userland.

ATM, this implies that the attached devices are powered down too (spindown). As spinning up takes quite some time, we can implement another level of dynamic PM w/ shorter delay to wake up - drives are not spinned down but controllers are powered down completely.

In any case, channel reset and following revalidation are necessary on wake up - if the device is still spinning, this shouldn't take too long but it will introduce noticeable delay - probably under or around a sec.

2. Another hacky way would be implementing it as an extension of link powersaving. I don't think this is a good idea tho. Waking up a controller usually involves link reset which in turn requires revalidation and reconfiguration of attached device, which should be done from exception handler.

The reason why your hack doesn't work is probably this reason. You need to pass the command to EH and tell it to perform full wake-up sequence and retry the command.

So, I think option #1 is the way to go - implementing leveled dynamic power management infrastructure and adding support in the block layer. What do you think?

Thanks.

Signature · 1 line
tejun
#6 lore raw

Hi!

Quoted 10 lines
Quoted 5 lines

Thanks... I got it to work (on 2 the old tree, I was not able to forward-port it), but power savings were not too big (~0.1W, maybe).

I'm getting huge (~1W) savings by powering down SATA controller, as in ahci_pci_device_suspend().

Yeah, it only turns off SATA PHY, so it doesn't result in huge saving. IIRC, it was somewhere around 5 percent on my notebook w/ static linksave mode (turning PHY off on empty port). But link powersaving introduces virtually no recognizable delay, so it's nice to have.

Yes, any powersavings without cost are a good idea.

Quoted 3 lines

Can you check if there is any difference between [D/H]IPS and static? ICH6M on my notebook can't do DIPS/HIPS, so I couldn't compare them against static.

What is D/HIPS? I could not find anything relevant..

Quoted 10 lines
Quoted 3 lines

It would be great to be able to power SATA
controller down, then power it back up when it is needed... I tried following hack, but could not get it to work. Any ideas?

1. One way to do it would be by dynamic power management. It would be nice to have wake-up mechanism at the block layer. Idle timer can run in the block layer or it can be implemented in the userland.

ATM, this implies that the attached devices are powered down too (spindown). As spinning up takes quite some time, we can implement

For now, powering down controller when disks are spinned down would be very nice first step.

When I forced disk to be spinned down (with power/state file) controller actually survived power down/power up... unfortunately with so long delay (~30 sec) that it is not usable in practice.

Quoted 3 lines

So, I think option #1 is the way to go - implementing leveled dynamic power management infrastructure and adding support in the block layer. What do you think?

Would be nice :-).
Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#7 lore raw

Hello,

Quoted 3 lines · Pavel Machek
Quoted 1 line

Can you check if there is any difference between [D/H]IPS and static? ICH6M on my notebook can't do DIPS/HIPS, so I couldn't compare them against static.

What is D/HIPS? I could not find anything relevant..

D/HIPS stand for device/host initiated power saving. These modes use two SATA link powersaving state (partial and slumber). Static mode simply turns off PHY on unoccupied port using SControl register. So, if you have an access to a notebook which has a SATA dock which support link powersaving, you can test it by...

* set link powersaving mode to HIPS/static. (mode 4)

* w/ device inserted, leave it idle for 15 seconds and record power consumption level (link should be in slumber state).

* pull out the device, wait for libata to detach the device and record power consumption level (libata should have turned off PHY after detaching the device).

I wanna know whether there is any difference in the amount of power saved between slumber and off states.

Quoted 9 lines
Quoted 6 lines
Quoted 3 lines

It would be great to be able to power SATA
controller down, then power it back up when it is needed... I tried following hack, but could not get it to work. Any ideas?

1. One way to do it would be by dynamic power management. It would be nice to have wake-up mechanism at the block layer. Idle timer can run in the block layer or it can be implemented in the userland.

ATM, this implies that the attached devices are powered down too (spindown). As spinning up takes quite some time, we can implement

For now, powering down controller when disks are spinned down would be very nice first step.

Yeap.

Quoted 3 lines

When I forced disk to be spinned down (with power/state file) controller actually survived power down/power up... unfortunately with so long delay (~30 sec) that it is not usable in practice.

Can you describe what you've done in more detail? Do you have dmesg of the 30sec wait?

Quoted 3 lines
Quoted 1 line

So, I think option #1 is the way to go - implementing leveled dynamic power management infrastructure and adding support in the block layer. What do you think?

Would be nice :-).

So, do you think we're ready for another PM infrastructure update? :-P

Signature · 1 line
tejun
#8 lore raw

Hi!

First, sorry for long reply time. I had too many horses and not enough time.

Quoted 17 lines
Quoted 5 lines
Quoted 3 lines

Can you check if there is any difference between [D/H]IPS and static? ICH6M on my notebook can't do DIPS/HIPS, so I couldn't compare them against static.

What is D/HIPS? I could not find anything relevant..

D/HIPS stand for device/host initiated power saving. These modes use two SATA link powersaving state (partial and slumber). Static mode simply turns off PHY on unoccupied port using SControl register. So, if you have an access to a notebook which has a SATA dock which support link powersaving, you can test it by...

* set link powersaving mode to HIPS/static. (mode 4)

* w/ device inserted, leave it idle for 15 seconds and record power consumption level (link should be in slumber state).

Quoted 6 lines

* pull out the device, wait for libata to detach the device and record power consumption level (libata should have turned off PHY after detaching the device).

I wanna know whether there is any difference in the amount of power saved between slumber and off states.

I'm probably doing something wrong, but...

I'm on commit

commit 9a7b050525f7d70d2ed62affb691b9d4ca2b82d2
tree b8195e5625dc5bad6757b0dddec0dacf416a0779
parent 50c3086de212ce56eaa2bf284586fb021615b5e1
author Tejun Heo <[email protected]> Mon, 16 Oct 2006 07:24:57 +0900 committer Tejun Heo <[email protected]> Mon, 16 Oct 2006 07:24:57 +0900

[PATCH] sata_sil24: implement PORT_RST

As DEV_RST (hardreset) sometimes fail to recover the controller (especially after PMP DMA CS errata). In such cases, perform
PORT_RST
prior to DEV_RST.

(2.6.19-rc1)

and I do not see powersave tunable:

root@amd:/sys/module# ls libata/parameters/
ata_probe_timeout         atapi_enabled hotplug_polling_interval
atapi_dmadir              fua

...how do I pull working version?

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines

So, I think option #1 is the way to go - implementing leveled dynamic power management infrastructure and adding support in the block layer. What do you think?

Would be nice :-).

So, do you think we're ready for another PM infrastructure update? :-P

Well... things are pretty quiet in that area just now... So yes.
Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#9 lore raw
Quoted 21 lines · Pavel Machek

I'm probably doing something wrong, but...

I'm on commit

commit 9a7b050525f7d70d2ed62affb691b9d4ca2b82d2
tree b8195e5625dc5bad6757b0dddec0dacf416a0779
parent 50c3086de212ce56eaa2bf284586fb021615b5e1
author Tejun Heo <[email protected]> Mon, 16 Oct 2006 07:24:57 +0900 committer Tejun Heo <[email protected]> Mon, 16 Oct 2006 07:24:57 +0900

[PATCH] sata_sil24: implement PORT_RST

As DEV_RST (hardreset) sometimes fail to recover the controller (especially after PMP DMA CS errata). In such cases, perform
PORT_RST
prior to DEV_RST.

(2.6.19-rc1)

Quoted 7 lines

and I do not see powersave tunable:

root@amd:/sys/module# ls libata/parameters/
ata_probe_timeout         atapi_enabled hotplug_polling_interval
atapi_dmadir              fua

...how do I pull working version?

I haven't updated link powersave patch yet. Core implementation was agreed on but interface hasn't been decided yet. Maybe it's about time to add /sys/class/ata_{host|device}/. So, the patchset is pushed back for the time being.

Quoted 5 lines
Quoted 3 lines
Quoted 2 lines
Quoted 1 line

So, I think option #1 is the way to go - implementing leveled dynamic power management infrastructure and adding support in the block layer. What do you think?

Would be nice :-).

So, do you think we're ready for another PM infrastructure update? :-P

Well... things are pretty quiet in that area just now... So yes.

If I understood correctly, the high power consumption of ahci controller can be solved by dynamically turning off command processing while the controller is idle, which fits nicely into link powersaving, right? So, I think full-fledged leveled dynamic PM would be an overkill for this particular problem, but then again, maybe the correct way to implement link powersaving is via leveled dynamic PM. I'll give it more thoughts.

Thanks.

Signature · 1 line
tejun
#10 lore raw

Hi!

Quoted 7 lines
Quoted 7 lines

and I do not see powersave tunable:

root@amd:/sys/module# ls libata/parameters/
ata_probe_timeout         atapi_enabled 
hotplug_polling_interval
atapi_dmadir              fua
Quoted 7 lines
Quoted 1 line

...how do I pull working version?

I haven't updated link powersave patch yet. Core implementation was agreed on but interface hasn't been decided yet. Maybe it's about time to add /sys/class/ata_{host|device}/. So, the patchset is pushed back for the time being.

/sys/class/ata* would probably be ok.

Quoted 9 lines
Quoted 2 lines

Well... things are pretty quiet in that area just
now... So yes.

If I understood correctly, the high power consumption of ahci controller can be solved by dynamically turning off command processing while the controller is idle, which fits nicely into link powersaving, right? So, I think full-fledged leveled dynamic PM would be an overkill for this particular problem, but then again, maybe the

It is single bit, and it should not even need a timeout, AFAICT, so perhaps we should just fix it (no need for dynamic PM layers). It probably does not even need to be configurable...

Signature · 1 line
Thanks for all the (sleeping) penguins.
#11 lore raw

Hello,

Quoted 5 lines · Pavel Machek
Quoted 1 line

If I understood correctly, the high power consumption of ahci controller can be solved by dynamically turning off command processing while the controller is idle, which fits nicely into link powersaving, right? So, I think full-fledged leveled dynamic PM would be an overkill for this particular problem, but then again, maybe the

It is single bit, and it should not even need a timeout, AFAICT, so perhaps we should just fix it (no need for dynamic PM layers). It probably does not even need to be configurable...

I think this has been discussed in linux-ide recently but just to add my 2 cents. ALPE and ASP can cause quite some problems. Many devices don't implement link powermanagement mode properly and some locks up completely (recent LG dvd-rams, only physical power removal/reapply can recover it) while others (wd raptors) spin down on slumber. If you turn on ALPE and ASP on such devices, all hell will break loose.

Even for devices which implement partial and slumber modes properly, I'm not very comfortable with ALPE. It puts the link into power save mode as soon as there is no command pending. IMHO, that's way too aggressive. It's not like being that aggressive will bring any noticeable difference in power consumption. Disk idle period is too long for that level of aggressiveness to be effective.

Thanks.

Signature · 1 line
tejun
#12 lore raw

Hi!

Quoted 7 lines
Quoted 7 lines
Quoted 6 lines

If I understood correctly, the high power consumption of ahci controller can be solved by dynamically turning off command processing while the controller is idle, which fits nicely into link powersaving, right? So, I think full-fledged leveled dynamic PM would be an overkill for this particular problem, but then again, maybe the

Quoted 7 lines
Quoted 3 lines

It is single bit, and it should not even need a timeout, AFAICT, so perhaps we should just fix it (no need for dynamic PM layers). It probably does not even need to be configurable...

I think this has been discussed in linux-ide recently but just to add my 2 cents. ALPE and ASP can cause quite some problems. Many devices don't implement link powermanagement mode properly and some locks up

Ok, so it needs to be optional :-). I played with the code a little bit, but did not find combination that both worked and saved power :-(.
Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html