On Friday, 20 July 2007 18:08, Milton Miller wrote: > On Jul 19, 2007, at 3:28 PM, Rafael J. Wysocki wrote: > > On Thursday, 19 July 2007 17:46, Milton Miller wrote: > >> The currently identified problems under discussion include: > >> (1) how to interact with acpi to enter into S4. > >> (2) how to identify which memory needs to be saved > >> (3) how to communicate where to save the memory > >> (4) what state should devices be in when switching kernels > >> (5) the complicated setup required with the current patch > >> (6) what code restores the image > > > > (7) how to avoid corrupting filesystems mounted by the hibernated > > kernel > > > > Ok I talked on this too. > > >> I'll now start with quotes from several articles in this thread and my > >> responses. > >> > >> Message-ID: <200707172217.01890.rjw@xxxxxxx> > >> On Tue Jul 17 13:10:00 2007, Rafael J. Wysocki wrote: > >>> (1) Upon entering the sleep state, which IMO can be done _after_ the > >>> image > >>> has been saved: > >>> * figure out which devices can wake up > >>> * put devices into low power states (wake-up devices are placed in > >>> the Dx > >>> states compatible with the wake capability, the others are > >>> powered > >>> off) > >>> * execute the _PTS global control method > >>> * switch off the nonlocal CPUs (eg. nonboot CPUs on x86) > >>> * execute the _GTS global control method > >>> * set the GPE enable registers corresponding to the wake-up > >>> devices) > >>> * make the platform enter S4 (there's a well defined procedure for > >>> that) > >>> I think that this should be done by the image-saving kernel. > >> > >> Message-ID: <87odiag45q.fsf@xxxxxxxxxxx> > >> On Tue Jul 17 13:35:52 2007, Jeremy Maitin-Shepard > >> expressed his agreement with this block but also confusion on the > >> other > >> blocks. > >> > >> > >> I strongly disagree. > >> > >> (1) as has been pointed out, this requires the new kernel to > >> understand > >> all io devices in the first kernel. > >> (2) it requires both kernels to talk to ACPI. This is doomed to > >> failure. How can the second kernel initialize ACPI? The platform > >> thinks it has already been initialized. Do we plan to always undo all > >> acpi initialization? > > > > Good question. I don't know. > > > >>> (2) Upon start-up (by which I mean what happens after the user has > >>> pressed > >>> the power button or something like that): > >>> * check if the image is present (and valid) _without_ enabling ACPI > >>> (we don't > >>> do that now, but I see no reason for not doing it in the new > >>> framework) > >>> * if the image is present (and valid), load it > >>> * turn on ACPI (unless already turned on by the BIOS, that is) > >>> * execute the _BFS global control method > >>> * execute the _WAK global control method > >>> * continue > >>> Here, the first two things should be done by the image-loading > >>> kernel, but > >>> the remaining operations have to be carried out by the restored > >>> kernel. > >> > >> Here I agree. > >> > >> Here is my proposal. Instead of trying to both write the image and > >> suspend, I think this all becomes much simpler if we limit the scope > >> the work of the second kernel. Its purpose is to write the image. > >> After that its done. The platform can be powered off if we are going > >> to S5. However, to support suspend to ram and suspend to disk, we > >> return to the first kernel. > > > > We can't do this unless we have frozen tasks (this way, or another) > > before > > carrying out the entire operation. > > What can't we do? We've already worked with the drivers to quesce the > hardware and put any information to resume the device in ram. Now we > ask them to put their device in low power mode so we can go to sleep. For that to work, we have to require the image-saving kernel to leave devices in the same state, or in a state compatible with the state, in which they were when it got control. > Even if we schedule, the only thing userspace could touch is memory. > If we resume, they just run those computations again. > > > In that case, however, the kexec-based > > approach would have only one advantage over the current one. Namely, > > it > > would allow us to create bigger images. > > The advantage is we don't have to come up with a way to teach drivers > "wake up to run these requests, but no other requests". We don't have > to figure out what we need to resume to allow them to process a > request. I'm not sure what you mean here. Please explain. > >> This means that the first kernel will need to know why it got resumed. > >> Was the system powered off, and this is the resume from the user? Or > >> was it restarted because the image has been saved, and its now time to > >> actually suspend until woken up? If you look at it, this is the same > >> interface we have with the magic arch_suspend hook -- did we just > >> suspend and its time to write the image, or did we just resume and its > >> time to wake everything up. > >> > >> I think this can be easily solved by giving the image saving kernel > >> two > >> resume points: one for the image has been written, and one for we > >> rebooted and have restored the image. I'm not familiar with ACPI. > >> Perhaps we need a third to differentiate we read the image from S4 > >> instead of from S5, but that information must be available to the OS > >> because it needs that to know if it should resume from hibernate. > >> > >> By making the split at image save and restore we have several > >> advantages: > >> > >> (1) the kernel always initializes with devices in the init or quiesced > >> but active state. > >> > >> (2) the kernel always resumes with devices in the init or quiesced but > >> active state. > >> > >> (3) the kjump save and restore kernel does not need to know how to > >> suspend all devices in the platform. > >> > >> (4) we have a merged path for suspend to disk, suspend to ram, and > >> suspend to both. > >> > >> (5) because of (4), we can implement sleep policys where we save the > >> image to disk but try to stay in ram based on expected remaining > >> battery life. > >> > >> (6) we confine all platform (acpi) interaction to the main kernel > >> > >> (7) we limit the knowledge needed in the second kernel. It needs to > >> know how to do its job and then put the hardware back how it found it. > >> Nothing more. > > > > This would have been nice if we had been able to do it. > > I don't understand this comment. "if we had been able"? I don't > think we have tried yet. That's related to the discussion above. If we are unable to do (3) and (6) without the freezing of tasks, which I'm not sure is not the case, the entire scheme won't be viable. Well, we might be able to do it provided that drivers will block the tasks on I/O effectively, but I see a big 'if' here ... > >> For the suspend to ram and then woken up case, we simply need to > >> invalidate the image before restarting normal kernel operation. > >> > >> People have worried about how to boot and restore the kernel, and what > >> to do if reading the image fails. They worry about needing memory > >> hotplug or delayed acpi parsing. They are forgetting one thing. This > >> kernel has support for kexec. > >> > >> This is all easily solved by having the bootloader from the bios > >> always > >> boot the restore kernel. > > > > Well, I think this is not generally acceptable, although I agree that > > it would > > be simpler. > > For those that don't find it acceptable they can teach their bootloader > when they may have a image to resume. Yes, and I think we need to seriously consider this possibility. > >> It will boot with limited useable memory and > >> no acpi support. If the restore kernel userspace detects that there > >> is > >> no restore image, it simply loads the normal main kernel and initrd / > >> initramfs and calls the normal kexec. The cost is the time to init > >> the > >> restore kernel, read the kernel with full drivers (vs reading it from > >> the bootloader). If you want a boot menu, use kboot (on sourceforge). > > > > Well, I'm afraid of adding more and more infrastructure to the mix. > > Requiring the hibernated kernel to be able to start from kexec should > not be bad. If you were referring to adding kboot, that is just an > option. Yes, I was. > One can still use bootloaders menus to select alternate kernels. > However, as you said, you want to boot differently for resume (no acpi > until after image loaded) from full boot. That's correct and I think some kind of cooperation with the bootloader is needed for that. > >> On Jul 17, 2007, at 2:13 PM, Rafael J. Wysocki wrote: > >>> On Tuesday, 17 July 2007 22:27, david@xxxxxxx wrote: > >>>> On Tue, 17 Jul 2007, Alan Stern wrote: > >>>>> But what about the freezer? The original reason for using kexec > >>>>> was > >>>>> to > >>>>> avoid the need for the freezer. With no freezer, while the > >>>>> original > >>>>> kernel is busy powering down its devices, user tasks will be free > >>>>> to > >>>>> carry out I/O -- which will make the memory snapshot inconsistent > >>>>> with > >>>>> the on-disk data structures. > >>>> > >>>> no, user tasks just don't get scheduled during shutdown. > >>>> > >>>> the big problem with the freezer isn't stopping anything from > >>>> happening, > >>>> it's _selectivly_ stopping things. > >> > >> Agreed. Or rather, selectively not stopping and resuming things. > > > > I don't quite understant this statement. Can you please elaborate? > > Feel free to list other problems with the freezer, but I'm saying that > the problems are stemming from trying to freeze most of userspace and > some selection of kernel threads so that new requests to the outside > are not made, but then turning around and saying "ok now do some io, > but only what this thread of execution originates". We're _not_ doing anything like this. > Its originates not generates so we are trying to teach the whole stack > these limits, including going back to userspace for FUSE. Again, I don't understand what you're talking about. This is not like things work right now, that's for sure. :-) The problem with FUSE is related to the fact that the freezer can't freeze uninterruptible tasks and we said that perhaps we might avoid it if FUSE was made freezing-aware. Still, no one has gone in this direction and I don't know of any plans to do that. Please, stop trying to blame the freezer for all evil. Also, it's better if you know how the things that you want to improve really work. > >>> It's selectively stopping kernel threads, which is just about right. > >>> If you > >>> that _this_ is a main problem with the freezer, then think again. > >>> > >>>> with kexec you don't need to let any portion of the origional kernel > >>>> or userspace operate so you don't have a problem. > >>> > >>> In fact, the main problem with the freezer is that it is a > >>> coarse-grained > >>> solution. Therefore, what I believe we should do is to evolve in the > >>> directoin > >>> of more fine-grained solutions and gradually phase out the freezer. > >>> > >>> The kexec-based approach is an attempt to replace one coarse-grained > >>> solution > >>> (the freezer) with even more coarse-grained solution (stopping the > >>> entire > >>> kernel with everything), which IMO doesn't address the main problem. > >>> > >> > >> I think this addresses teh problem. Its probably a bit harder than > >> powermac because we have to fully quiesce devices; we can't cheat by > >> leaving interrupts off. But once the drivers save the state of their > >> devices and stop their queues, it should be easy to audit the paths to > >> powerdown devices and call the platform suspend and ram wakeup paths. > > In other words, I'm replacing a course-grained solution with an > absolute solution. "From this point on you can only write to ram." Which means that we need to take care of the drivers _before_ doing anything else. I agree with that, of course. :-) > >> Going back to the requirements document that started this thread: > >> > >> Message-ID: <200707151433.34625.rjw@xxxxxxx> > >> On Sun Jul 15 05:27:03 2007, Rafael J. Wysocki wrote: > >>> (1) Filesystems mounted before the hibernation are untouchable > >> > >> This is because some file systems do a fsck or other activity even > >> when > >> mounted read only. For the kexec case, however, this should be "file > >> systems mounted by the hibernated system must not be written". As > >> has > >> been mentioned in the past, we should be able to use something like dm > >> snapshot to allow fsck and the file system to see the cleaned copy > >> while not actually writing the media. > > > > We can't _require_ users to use the dm snapshot in order for the > > hibernation > > to work, sorry. > > I actually listed three ways to start. Not all of them required > dm-snapshot. I was proposing "if you need to read ext3, then use > dm-snapshot". I don't think we should differentiate filesystems this way. We should just do the same thing with all of them. > > And by _reading_ from a filesystem you generally update metadata. > > not on ones mounted read-only. I'll reply more later in the thread. OK > >> The kjump kernel must not have any knowledge retained if we reuse it. > >> > >>> (2) Swap space in use before the hibernation must be handled with > >>> care > >> > >> Yes. Actually, even though they have been used by the write-in-the > >> kernel users, they will be among the most difficult devices to use for > >> snapshots by a userspace second kernel. > >> > >>> (3) There are memory regions that must not be saved or restored > >> > >> because they may not exist. This means that we must identify the > >> memory to be saved and restored in a format to be passed between the > >> kernel. > >> > >>> (4) The user should be able to limit the size of a hibernation image > >> > >> This means the suspending kernel must arrange to reduce its active > >> memory. The limited save can be done by providing a limited list in > >> (3). > > > > It seems to me that you don't understand the problem here. > > > > Assume you have 90% of RAM allocated before the hibernation and the > > user has > > requested the image to be not greater than 50% of RAM. In that case > > you have > > to free some memory _before_ identifying memory to save and you must > > not > > race with applications that attempt to allocate memory while you're > > doing it. > > Hmm... I didn't say how to reduce the memory or identify it, did I? > > Ok fine. I'll allocate a bunch of memory and put it on a list. And you cause the OOM killer to show up. Not good. > Normal memory pressure will swap things out or drop filesystem pages. > When I build the list of memory to backup, I filter out this list. > After resume, I'll free it back. > > We can arrange for this "task" to be preferred by the oom killer, if > case the user is trying to suspend into less than memory than can be > freed. > > >>> (5) Hibernation should be transparent from the applications' point of > >>> view > >> > >> People have pointed out they may want userspace to be aware of the > >> suspend. I believed this can be done with /proc/apm emulation today > >> or by other means; it seems that should be hooked up to dbus in some > >> fashion. > > > > Not a solution, because there still will be programs not needing to > > know > > anything about hibernation. After all, we don't require all > > applications to > > know anything about SMP, even if they are executed on an SMP system. > > How do any of those methods require userpsace to know anything about > hibernation? I was talking about a general framework consistent with > todays kernel to user communication for those parts of userspace that > *want* to know about suspend and hibernation. OK, I didn't understand, then. > >>> (6) State of devices from before hibernation should be restored, if > >>> possible > >> > >> related to suspend should be transparent ... yes. > >> > >>> (7) On ACPI systems special platform-related actions have to be > >>> carried out at > >>> the right points, so that the platform works correctly after the > >>> restore > >> > >> I believe I have explained my suggestion. > >> > >>> (8) Hibernation and restore should not be too slow > >> > >> We control the added code. We are using full runtime drivers and > >> will > >> run at hardware speeds. > > > > That may not be enough. If you're going to save, say, 80% of RAM on a > > 2 GB > > machine, then you'll have to be using image compression. > > Yea, so? We have a full kernel and userspace, adding compression > before writing should be easy. The is no struct page for memory in the > old kernel, so we likely need to be copying them in userspace anyways. > Adding compression should be easy. Yes, it's not that difficult. > >>> (9) Hibernation framework should not be too difficult to set up > >> > >> Ok the current patch is presently too difficult. But I think it will > >> be much simpler with a few small changes. > >> > >> As noted in the thread > >> > >> Message-ID: <873azxwqhr.fsf@xxxxxxxxxxx> > >> Subject: Re: hibernation/snapshot design > >> on Mon Jul 9 08:23:53 2007, Jeremy Maitin-Shepard wrote: > >>>> Both would work. One would eat 8-64MB of your RAM, permanently; > >>> > >>> As I have stated in other messages, the kdump approach would not > >>> waste > >>> any RAM permanently. > >> ... > >>> Immediately before jumping to the new kernel, the first X bytes > >>> (where > >>> X > >>> is the amount of memory the new kernel will get, typically 16MB or > >>> 64MB) > >>> of physical memory are backed up into the arbitrary discontiguous > >>> pages > >>> that are made available. This will not take very long, because > >>> copying > >>> even 64MB of memory is extremely fast. Then the new kernel is free > >>> to > >>> use the first X bytes of contiguous physical memory. Problem solved. > >> > >> > >> Ok, now let's look at my list again: > >> > >>> (1) how to interact with acpi to enter into S4. > >> > >> This was discussed. > >> > >>> (2) how to identify which memory needs to be saved > >> > >> We need to generate a list. We need it to fit in a compuatable size > >> so > >> that we can free and allocate the pages before suspending IO in the > >> first kernel. > >> > >> One possibility is to use something like the kexec copy list. If we > >> are imaging a small fraction of ram this is appropriate, but if we are > >> doing dense saves we need something extent based. We should be able > >> to > >> extend the list. > >> > >>> (3) how to communicate where to save the memory > >> > >> This is an intresting topic. The suspended kernel has most IO and > >> disk > >> space. It also knows how much space is to be occupied by the kernel. > >> So communicating a block map to the second kernel would be the obvious > >> choice. But the second kernel must be able to find the image to > >> restore it, and it must have drivers for the media. Also, this is not > >> feasible for storing to nfs. > >> > >> I think we will end up with several methods. > >> > >> One would be supply a list of blocks, and implement a file system that > >> reads the file by reading the scatter list from media. The restore > >> kernel then only needs to read an anchor, and can build upon that > >> until > >> the image is read into memory. Or do this in userspace. > >> > >> I don't know how this compares to the current restore path. I wasn't > >> able to identify the code that creates the on disk structure in my 10 > >> minute perusal of kernel/power/. > > > > The structure is created at two levels. > > > > First, the code in snapshot.c makes the image available to the code in > > swap.c > > as a stream of pages. The first page is the header, followed by some > > pages > > containing the PFNs of the page frames to which the image data pages > > are to be > > restored, followed by the image data pages themselves (the ordering of > > the PFNs > > must be the same as the ordering of data pages that correspond to > > them). > > Still, the low-level image format only needs to be known by the > > restore code in > > snapshot.c . > > Ok sounds like this code could be reused. I'll look into it. > > > Second, the code in swap.c writes the image pages to a storage adding > > some > > metadata making it possible to reproduce their original ordering > > during the > > restore. > > So you are allocating the blocks as you go ... and adding meta data > along the way? Something like this. I can't say how Nigel does it, though. > > The fact that we use swap spaces as the storage is related to > > implementation > > simplicity rather than anything else. > > Ok ... this only supports uncompressed hibernation? The in-kernel version doesn't support compression (again, this was a choice made to keep the code relatively simple), but the userland version supports compression (and image encryption). > The first kernel is going to specify (1) what to backup. It can > specify (2) where to backup, although we have to be careful identify > the device in a persistent way. Yes, that seems doable. > >> A second method will be to supply a device and file that will be > >> mounted by the save kernel, then unmounted and restored. This would > >> require a partition that is not mounted or open by the suspended > >> kernel > >> (or use nfs or a similar protocol that is designed for multiple client > >> concurrent access). > >> > >> A third method would be to allocate a file with the first kernel, and > >> make sure the blocks are flushed to disk. The save and restore > >> kernels > >> map the file system using a snapshot device. Writing would map the > >> blocks and use the block offset to write to the real device using the > >> method from the first option; reading could be done directly from the > >> snapshot device. > >> > >> The first and third option are dead on log based file systems (where > >> the data is stored in the log). > > > > All in all, we have three different and working implementation of the > > image-writing and image-reading code at our disposal. Why would you > > want to > > break the open doors? > > The problem I'm saying kexec solves is how to get the data to the > device while most of the kernel is trying not do anything permanent. > > If we can reuse existing code, great. I think we can. > >>> (4) what state should devices be in when switching kernels > >> > >> My proposal is either initialized and untouched or quiesced. > > > > This is reasonable, but in general we also need to save some > > information > > about the pre-hibernation state of devices, so that we can put them > > into the > > same state, if reasonably possible, during the restore. > > What state are you referring to? > > Yes, there is state that the drivers have to store to ram, but this the > same state they need to store when suspending to ram if the device can > be powered off. Yes. > Maybe we need to teach drivers to store more state, like remember that > a hard drive was spun down. I'm not sure about that. > So we may need a flag saying "we powered off", "we resumed from > suspend". There already is something like this. Generally, we're going to have a special callback that will be used by the core after the restore, so the driver will always know what it's supposed to do. > >>> (5) the complicated setup required with the current patch > >> > >> I think a few simple changes to kjump will make this much simpler. > >> See > >> below. > >> > >>> (6) what code restores the image > >> > >> The save kernel, loaded at boot. People have suggested booting the > >> first kernel, and using current restore code. However, I think that > >> ignores that (1) we saved from a different kernel, so the backed up > >> region will be restored to its backed up random pages, > > > > This problem has already been solved. > > > >> (2) the code was written to restore the same kernel, > > > > Not exactly. In fact, the current implementation only relies on the > > tiny > > portion of the restore code being in the same place in both kernels, > > but > > we can change the code not to make this assumption (it'll be more > > complicated, > > but that's perfectly doable). > > If the save kernel is different from the run kernel (to make it > smaller), its likely the image saving code will move. I view restoring > from a different kernel than saving as an advanced feature. > > Lets get resuming from the save kernel working first. > > >> so the text and data will be replaced by identical text. Its much > >> simpler > >> conceptually to use the same kernel to save and restore the image. > > > > Here I agree. :-) > > > >> Simplifying kjump: the proposal for v3. > >> > >> The current code is trying to use crash dump area as a safe, reserved > >> area to run the second kernel. However, that means that the kernel > >> has to be linked specially to run in the reserved area. I think we > >> need to finish separating kexec_jump from the other code paths. > >> > >> (1) add a new command line argument that specifies the kexec_jump > >> target area. > >> > >> (2) add a kjump flag to the flags parameter, used by kexec_load. > >> When > >> loading a jump kernel, it is loaded like a normal kernel, however, > >> additional control pages are allocated to (a) save the kexec_jump > >> target area (b) save the backed up region that is used by all kernels > >> like crash dump, and (c) space for invoking relocate_new_kernel that > >> will get its args from the execution entry point and will restore the > >> kernel then call resume and suspend. > >> > >> (3) replace jump_huf_pfn with two command line addresses that specify > >> the (a) return point for after resume, and (b) the return point for > >> after image save. Actually these can be done in userspace; the > >> second > >> restore kernel can just specify the null copy list and the entry > >> points > >> supplied by the suspended kernel. To do resume we also need (c) where > >> to store resume address for the save kernel. > >> > >> > >> As a first stage of suspend and resume, we can save to dedicated > >> partitions all memory (as supplied to crash_dump) that is not marked > >> nosave and not part of the save kernel's image. > > > > A little problem here: there are "nosave" areas that are not marked as > > nosave. > > If crash_dump is going work the memory must exist. > > >> The fancy block lists and memory lists can be added later. > > > > On the majority of systems that will work. On some of them it won't. > > Ok .... well, my point is we can get started while we workout what the > list format is. If we decide to reuse the pfn lists above that may > come quickly. I think it's generally reasonable (a) not to save the entire memory (like free RAM areas etc.) and (b) include the information of the original location of each data page in the image, this way or another. This doesn't complicate things all that much. Greetings, Rafael -- "Premature optimization is the root of all evil." - Donald Knuth _______________________________________________ linux-pm mailing list linux-pm@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/linux-pm
