Why does the OS need to copy all RAM into HDD for hibernation?

my question is theoretical in computer OSes, specifically about hibernation.

As far as I know, hibernation (ACPI state S5, is that right?) involves writing into the hard disk all the data needed to restore RAM when OS os booted up again. OK this sounds good to me.

But, I remember the old times of Operating Systems class when the teacher explained the differences in memory management between OSes. Windows and Linux kernel both use virtual memory: all pages are initialized into swap, then loaded into RAM when needed so they can be relocated at any time.

I also know that memory pages are not immediately flushed to swap, but are kept in memory for a certain amount of time, or when there is the need to load a page from swap and all memory pages are dirty

My question is: Why does Windows need to store a separate hiberfil.sys file with all RAM dumped inside since all pages should be already present in swap file (and hibernation would simply mean flushing the caches and stopping the computer) and almost all clean after a prolonged inactivity?

The same, why does Linux explictly say it's copying pages to swap partition? When I simply suspend the system, it takes a couple of seconds of continuous disk access before halt (perhaps kernel flushing caches), but when I want to hibernate it takes the same time it takes to restore.

Am I missing something about kernels?

Solution 1:

Some of what you're describing sounds less like virtual memory and more like a disk cache, where the operating system keeps in-memory copies of the data it reads off of the disk, in case an application wants to read it again. In this scenario it may also delay writing "dirty" pages back to the disk in order to increase efficiency by writing a bunch of pages all at once.

Virtual memory is the other way around -- data is written into RAM first, and only into the swap file if the kernel needs to free up that RAM for some other purpose. (Not every bit of memory necessarily has a corresponding space in the swap file; otherwise you couldn't have a swap file smaller than your RAM size.) Conventional thinking is that this is the way it needs to be; it would be too expensive to keep copying the RAM to disk all the time when you don't need to. Think about how slow your computer runs when you're running lots of memory-hungry programs and it has to start thrashing the swap space.

That said, there are some experimental research operating systems that work more like you describe, with some complex rules for copying the memory back to disk at regular intervals, and the result is that there's no need for hibernate -- you can pull the plug on the computer and lose maybe 30 seconds of work. Check out CoyotOS, CapROS, or EROS if you're interested theoretically. I don't know what their performance is like, or what other downsides there might be to such a strategy, but when you factor in crazy things like memory shared with your video card or other devices, I'd imagine it's difficult to get it working properly for all cases. As far as I know, no production operating system does this sort of thing.

Solution 2:

The misunderstanding here is that not all RAM contents are swappable to disk.

User processes have dedicated swap space, but system does not. That is because many parts of the system are never swapped out, and the ones that are swapped out are simply erased from memory as not needed any more (not written out). System in this case includes the kernel and drivers and all data used to control processes (such as virtual memory segment table).

The hibernation file is written with all the data that is required to duplicate the system state. This is far from being equal in size to the RAM, as only occupied memory is written out.

Saving and restore take about the same time since the same data needs to be read as was written out. On disks where reading is much faster than writing, such as SSD, restore will be much faster.