On Monday, September 24, 2012, Aaron Lu wrote: > On Fri, Sep 21, 2012 at 11:18:27PM +0200, Rafael J. Wysocki wrote: > > On Friday, September 21, 2012, Aaron Lu wrote: > > > On Thu, Sep 20, 2012 at 10:00:51PM +0200, Rafael J. Wysocki wrote: > > > > On Wednesday, September 19, 2012, Aaron Lu wrote: > > > > > Thanks Rafael, and if there is any question/problem, > > > > > please kindly let me know. > > > > > > > > Well, unfortunately my initial review indicates that the patchset is not > > > > quite ready to go upstream yet. > > > > > > > > I'll send comments in replies to the individual patches, but overall I > > > > can > > > > say that at this stage of development, when I look at the patches, it > > > > should > > > > be clear to me not only what is being changed, but _why_ it is being > > > > changed > > > > in the first place and, secondly, why it is being changed in this > > > > particular > > > > way. It's far from that, though. > > > > > > I'm adding zero power support for optical disk drive(ZPODD), which is > > > made possible with the newly defined device attention(DA) pin introduced > > > in SATA 3.1 spec. > > > > > > The idea here is to use runtime pm to achieve this, so I basically did 2 > > > things: > > > 1 Add runtime pm support for ODD; > > > 2 Add power off support for ODD after it is runtime suspended. > > > > > > Patch 2 is runtime pm support for ODD, the reason it is done this way is > > > discussed here: > > > http://www.spinics.net/lists/linux-scsi/msg61551.html > > > > Why isn't it explained in the patch changelog, then? People should be able > > to learn why things are done the way they are done from git logs. > > > > > The basic idea is, the ODD will be runtime suspended as long as there is > > > nobody using it, that is, no programs opening the block device. > > > > > > The ODD will be polled periodically, so it will be runtime resumed > > > before checking if there is any events pending and suspended when done. > > > > OK. So what happens if we power off the drive via runtime PM. Does it > > it really make sense to resumie it through polling in that case? > > No, this is the reason I introduced the powered_off flag. If set, the > poll will simply return without touching the device. > > I've tried to do a disk_block_events call on its suspend callback when > it is ready to be powered off, but there is a race that I don't know how > to solve: > pm_runtime_suspend disk_events_workfn > scsi_dev_type_suspend sr_block_check_events > sr_suspend cdrom_check_events > disk_block_events cdrom_update_events > (this call waits for all sr_check_events > running events_checking function scsi_autopm_get_device > to return) > > Suppose sr_suspend runs first, and then sr_check_events comes in. > sr_suspend calls disk_block_events, which waits for sr_check_events, > while scsi_autopm_get_device wait for suspend callback to finish, > deadlock.
I need some more time to think about this, stay tuned. > > > The only exception is, if we found a disc is just inserted, we will not > > > idle it immediately at the end of the poll, reason explained in another > > > mail. > > > > > > This is the rational I wrote patch 2, and patch 1 is used by patch 2. > > > > > > Patch 3 is adding power off support for ODD after it is runtime > > > suspended, the condition is specified in section 15: > > > ftp://ftp.seagate.com/sff/INF-8090.PDF > > > > > > That is, for tray type ODD: no media inside and door closed; for slot > > > type ODD: no media inside. > > > > > > The is the reason sr_suspend is written, for non-ZPODD capable devices, > > > it does nothing; for ZPODD devices, it will check the above condition to > > > see if it is ready to be powered off. The ready_to_power_off flag will be > > > used by ATA layer to decide if power can be removed. > > > > Now, James says he doesn't like the way ready_to_power_off is used. Sure > > enough, it is totally irrelevant to the majority of SCSI devices. It > > actually > > is totally irrelevant to everything in the SCSI subsystem except for the sr > > driver and libata. So I wonder if you have considered any alternative > > way to address the use case at hand? > > > > > When in powered off state, if user presses the eject button or insert a > > > disc, an ACPI event will be generated and our acpi wake handler will > > > pm_runtime_resume the ODD. And if it is a tray type ODD, its tray should > > > be ejected(need_eject flag) after powered on. This is patch 3. > > > > That sounds reasonable enough, but the role of the powered_off and > > need_eject flags could be explained a bit better. In particular, it would > > powered_off: set when the device is powered off, clear otherwise. That's pretty clear, but I think this flag should be called no_polling or something like this, because that's what it means to the SCSI layer. > need_eject: > First consider how the device will be runtime resumed: > 1 Some program opens the block device; > 2 Events checking poll when it's not powered off yet; > 3 User presses the eject button or inserts a disc into the slot when the > device is in powered off state. > And the need_eject flag is for case 3, when the device is in powered off > state and user presses the eject button, it will be powered on(through > acpi wake notification function) and runtime resumed. In its runtime > resume callback, its tray needs to be ejected since user just presses > the eject button. The whole process of ZPODD is opaque to the user, > he/she doesn't know the ODD lost power so the ODD has to behave exactly > like it doesn't lose power. Do you think it can be useful for other types of devices, not necessarily handled through ACPI? > Hi Oliver, > This flag is really to say the tray needs to be ejected after runtime > resumed, it's not that media change detected. It is possible that user > ejects the tray without putting any disc inside and simply close the > tray, which doesn't qualify a media change event. And if user does > put a disc in, the sr_check_events will find that and report the media > change event to user space. Agree? > > > be nice to have explained why they have to be present in struct scsi_device, > > because they don't seem to be particularly useful for many SCSI devices > > that aren't CD drives (the need_eject one in particular). > > With your suggestion of pm_platform_power_off_allowed, I suppose > powered_off can be eliminated similarly with something like > pm_platform_powered_off returning true or false(for ACPI platform, > return true when device is in D3 cold state). I'm currently thinking that using PM QoS may be a better approach here. > And for the need_eject flag, I don't know if there is a better place for > it. The acpi wake notification code resides in libata(where we need to > record that this resume is due to user presses the eject button and the > tray needs to be ejected after resumed), and the runtime resume callback > resides in scsi driver(where we actually eject the tray). Ideally, this > flag should sit in scsi_cd structure, but libata does not have access to > it. Yes, that's the problem that James mentioned. Thanks, Rafael -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
