On Saturday 06 January 2007 3:35 am, Zhang Rui wrote: > This patch set is against acpi-test sysfs branch, and will duplicate > several ACPI procfs functions in sysfs. > As struct subsystem power_subsys is referred and the interface under > /sys/power/ is changed by this patch series, > I wish to receive some comments from linux-pm too. First, please be sure your mailer wraps TX lines well before the end of an 80 character line. (I had to do that manually in this reply.) Second, I'm glad to see attention paid to those /proc/acpi files, which AFAICT have been mostly left alone for many years now; that neglect has been puzzling, given the driver model and sysfs plus the general deprecation of procfs for operational tasks. Third, stuff in /sys/power should be generic, and not depend on ACPI infrastructure or models ... it should for example work for embedded platforms where Linux runs on the bare metal. Moreover, in general an ACPI-specific interface should be the very last (and non-portable, worst) choice, not the first/default/preferred one. Which means in particular that: > Patch 03-05: > add ACPI sleep attributes in sysfs. > /proc/acpi/sleep is already deprecated by /sys/power/state. > /proc/acpi/alarm is deprecated by /sys/power/alarm. A /sys/power/alarm thing conflicts with ongoing RTC framework work. You may have missed the patches I've sent making RTCs be "just another wakeup-capable device" -- an interface that already works with multiple RTCs, is not in the least bit coupled to ACPI, and is on track to work with on more platforms -- or the proposal to provide sysfs support for activiting alarms within the RTC framework. The latest patches relevant to the RTC used by ACPI are: http://marc.theaimsgroup.com/?l=linux-kernel&m=116802017116195&w=2 http://marc.theaimsgroup.com/?l=linux-kernel&m=116802044406838&w=2 http://marc.theaimsgroup.com/?l=linux-kernel&m=116802044531547&w=2 http://marc.theaimsgroup.com/?l=linux-kernel&m=116802044532646&w=2 (Only that last one touches ACPI directly.) Note that like it says in the first message, this version doesn't have the "remove /proc/acpi/alarm" stuff, mostly to simplify merging of the core driver code; that will be done later, and in any case the old drivers/char/rtc.c code didn't interact with it either. It worked in earlier versions of the patch. (Modulo the fact that ACPI wakeup from STR seems to usually be a total lose.) One such version: http://marc.theaimsgroup.com/?l=linux-kernel&m=116406177605472&w=2 You'll observe those ACPI bits are kind of messy. The clean approach I have in mind is to have callbacks in platform_data, with which ACPI -- or EFI etc -- can do whatever nastiness it needs to let the RTC alarm trigger wakeup from system sleep state. A "bare metal" platform would just enable_irq_wake(); ACPI would set a magic PM1 bit or whatever; etc. (In terms of userspace tools, I posted an "rtcwake.c" tool working on several platforms, including x86 PCs with the above patch plus some ARM systems, notably AT91rm9200, with rtc wakeup irqs.) > For those devices that support "wake" attribute, Please tell me you mean "devices with a /sys/devices/.../power/wakeup" attribute. And that ACPI is finally going to start working with those attributes ... > two files, > "sleep_state" and "wakeup", are created for these devices. I'm going to be assuming your "wakeup" attribute has the same semantics as that existing one. If it doesn't, you should either fix your model, or justify changing the Linux-wide model. As for sleep_state: > "sleep_state" indicates the lowest sleeping state that can be > entered while still providing wake functionality. I understand that /proc/acpi/wakeup exposes this information. I don't understand how userspace is expected to use it. Or, for that matter, how this model ought to work on non-ACPI systems. But most especially, if it's useful, I don't understand why it would be ACPI-only. (Yes, there are huge gaps in Linux-PM wakeup support. Weak ACPI suppport for it, especially from STR -- without swsusp -- is a big factor. I was pleased to see RTC wakeup from ACPI S4 actually work ... first time I've ever seen ACPI wakeup ever work correctly.) > "wakeup" can be used to enable/disable the device's > ability to wake a sleeping system. We've found a need to nuance that a bit. So for example a USB root hub sure ought to support the USB "remote wakeup" mechanism in that way, and by and large the BIOS will declare that it does. (Not that I've ever observed that to work from ACPI STR, mind you!!) However, it appears out that on some boards remote wakeup doesn't work right ... either there's a chipset bug, or some board wiring problem, or something of that ilk. This impacts _runtime_ power management as well as system-wide sleep states. So "wakeup" also gets used to disable runtime PM techniques like suspending USB devices to reduce system power drain (since that relies on remote wakeup). The hardware mechanism isn't used only to wake sleeping systems; so the sysfs attribute isn't limited to that role either. > So /proc/acpi/wakeup is deprecated by > /sys/devices/acpi_system/.../xxx/sleep_state && wakeup. Why is ACPI still not coupling such information to the REAL device nodes? On my laptop, right now without any wakeup-capable USB devices attached, the appended script produces: % pm-wake lan on pci0000:00/0000:00:0a.0/0000:01:01.0 on pci0000:00/0000:00:06.1 hub on pci0000:00/0000:00:02.2/usb3 usb_host on pci0000:00/0000:00:02.2 hub on pci0000:00/0000:00:02.1/usb2 usb_host on pci0000:00/0000:00:02.1 hub on pci0000:00/0000:00:02.0/usb1 usb_host on pci0000:00/0000:00:02.0 on pci0000:00/0000:00:01.1 rtc on pnp0/00:06 % That's with various patches applied, notably one to transfer the PCI PM "can issue PME#" flags to the device's wakeup flag. (Much the same patch circulated when the wakeup flag was added. Works fine on x86, but there was an PCI init sequencing issue on PPC that kept it from getting merged. And of course, ACPI wants some additional out-of-band signaling, so only the RTC device is truly a wakeup source on this system; sigh.) That 00:06.1 device is a Winmodem, which /proc/acpi/wakeup doesn't list as a wakeup event source ... maybe for good reason. The 00:01.1 device is SMBUS, not listed for what I'm guessing is a less good reason (i.e. it really IS a wakeup device, but managed internally to ACPI). I'm guessing RTC is omitted from /proc/acpi/wakeup since that's really just a dump of one table, and it's not actually a list of the system's wakeup-capable ACPI devices. Note that /proc/acpi/wakeup says everything supports wake from S3; but that's never worked on this system. (Yes, even after enabling everything in /proc/acpi/wakeup.) RTC wakeup from S4 works, but that's a mode that /proc/acpi/wakeup says doesn't exist, so I might have been imagining that. :) - Dave #!/bin/bash # show the wakeup-capable devices and what's enabled/disabled # class_label *:* ==> $type class_label () { case $1 in # recognize common types of wakeup-capable devices i2c-dev:*) type="smbus "; return 0;; input:*) type="input "; return 0;; mmc_host:*) type="mmc_host "; return 0;; net:eth*) type="lan "; return 0;; net:*) type="net "; return 0;; pcmcia_socket:*)type="pcmcia "; return 0;; rtc:*) type="rtc "; return 0;; sound:*) type="modem "; return 0;; tty:*) type="tty "; return 0;; usb_host:*) type="usb_host "; return 0;; esac return 1 } # interface_label $PATH ==> $type interface_label () { for F in $(cd $1 >/dev/null 2>&1 ; echo *:*) do class_label $F && return done } # devtype $PATH ==> $type devtype () { local F T # fixed length, currently ten spaces type="" for F in $(cd $1 >/dev/null 2>&1 ; echo *:*) do if [ ! -d "$1/$F" ] then break; fi # is this a usb interface? if [ -f $1/$F/bInterfaceClass ] then interface_label $1/$F && return fi case $F in # use interface's label if possible, else generic usb_device:*) read T < $1/maxchild if [ 0 -lt $T ] then type="hub " return fi type="(usb) " continue;; *:*) class_label $F && return ;; esac done if [ "$type" = "" ] then for T in $(cd $1 >/dev/null 2>&1 ; echo fw-host*/ieee1394_host:*) do if [ ! -L "$1/$T" ] then break; fi type="firewire " return done fi if [ "$type" = "" ] then type=" " fi } cd /sys/devices for F in $(find * -name 'wakeup') do # F=.../power/wakeup read value < $F if [ "$value" = "" ] then continue fi # F=... F=$(dirname $(dirname $F)) devtype $F # for each entry that actually supports wakeup, one line with: # - device type (if recognized) # - wake on/OFF # - /sys/devices/... path case "$value" in "disabled") echo "$type OFF $F" ;; "enabled") echo "$type on $F" ;; esac done
