Proposal: U-Boot memory management
Tom Rini
trini at konsulko.com
Fri Dec 29 18:51:15 CET 2023
On Fri, Dec 29, 2023 at 06:30:43PM +0100, Heinrich Schuchardt wrote:
> On 12/29/23 18:21, Tom Rini wrote:
> > On Fri, Dec 29, 2023 at 06:09:44PM +0100, Heinrich Schuchardt wrote:
> > > On 12/29/23 17:47, Tom Rini wrote:
> > > > On Fri, Dec 29, 2023 at 05:42:17PM +0100, Heinrich Schuchardt wrote:
> > > > > On 12/20/23 20:12, Tom Rini wrote:
> > > > > > On Tue, Dec 19, 2023 at 09:15:21PM -0700, Simon Glass wrote:
> > > > > > > Hi,
> > > > > > >
> > > > > > > On Tue, 19 Dec 2023 at 05:46, Tom Rini <trini at konsulko.com> wrote:
> > > > > > > >
> > > > > > > > On Tue, Dec 19, 2023 at 03:15:38AM +0100, Heinrich Schuchardt wrote:
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > Am 19. Dezember 2023 02:26:00 MEZ schrieb Tom Rini <trini at konsulko.com>:
> > > > > > > > > > On Tue, Dec 19, 2023 at 01:01:51AM +0100, Heinrich Schuchardt wrote:
> > > > > > > > > > >
> > > > > > > > > > >
> > > > > > > > > > > Am 19. Dezember 2023 00:31:30 MEZ schrieb Tom Rini <trini at konsulko.com>:
> > > > > > > > > > > > On Tue, Dec 19, 2023 at 12:29:19AM +0100, Heinrich Schuchardt wrote:
> > > > > > > > > > > > >
> > > > > > > > > > > > >
> > > > > > > > > > > > > Am 19. Dezember 2023 00:16:40 MEZ schrieb Tom Rini <trini at konsulko.com>:
> > > > > > > > > > > > > > On Tue, Dec 19, 2023 at 12:08:31AM +0100, Heinrich Schuchardt wrote:
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > Am 18. Dezember 2023 23:41:08 MEZ schrieb Tom Rini <trini at konsulko.com>:
> > > > > > > > > > > > > > > > On Mon, Dec 18, 2023 at 11:34:16PM +0100, Heinrich Schuchardt wrote:
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > [snip]
> > > > > > > > > > > > > > > > > Or take:
> > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > > load host 0:1 $c kernel.efi
> > > > > > > > > > > > > > > > > load host 0:1 $d initrd.img
> > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > > How could we ensure that initrd.img is not overwriting a part of kernel.efi without memory allocation?
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > Today, invalid checksum as part of some part of the kernel fails. But
> > > > > > > > > > > > > > > > how do we do this tomorrow, are you suggesting that "load" perform
> > > > > > > > > > > > > > > > malloc() in some predefined size? If $c is below $d and $c + kernel.efi
> > > > > > > > > > > > > > > > is now above $d we can throw an error before trying to load, yes. But
> > > > > > > > > > > > > > > > what about:
> > > > > > > > > > > > > > > > load host 0:1 $d initrd.img
> > > > > > > > > > > > > > > > load host 0:1 $c kernel.efi
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > In that case (which is only marginally contrived, the more real case is
> > > > > > > > > > > > > > > > loading device tree in to unexpectedly large ramdisk because someone
> > > > > > > > > > > > > > > > didn't understand the general advice on why device tree is lower than
> > > > > > > > > > > > > > > > ramdisk address) I'm fine with an error that amounts to "you just
> > > > > > > > > > > > > > > > corrupted another allocation" and then "fail, reset the board" or so.
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > Our current malloc library cannot manage the complete memory. We need a library like lmb which should also cover the memory management that we currently have in lib/efi/efi_memory.c. This must include a memory type attribute for usage in the GetMemoryMap() service. A management on page level seems sufficient.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > The load command should permanently allocate memory in that lmb+ library.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > We need an unload command to free the memory if we want to reuse the memory or we might let the load comand free the memory if exactly the same start address is reused.
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > Our current way of loading things in to memory does not handle the case
> > > > > > > > > > > > > > I described, yes. How would what you're proposing handle it?
> > > > > > > > > > > > >
> > > > > > > > > > > > > If the load command has to allocate memory for the image and that allocation is kept, any attempt to allocate overlapping memory would fail.
> > > > > > > > > > > >
> > > > > > > > > > > > So you're saying that the load command has to pre-allocate memory? Or as
> > > > > > > > > > > > it goes? If the latter, in what size chunks? This starts to get at what
> > > > > > > > > > > > Simon was talking about with respect to memory fragmentation. Which to
> > > > > > > > > > > > be clear is a problem we have today, we just let things overlap and hope
> > > > > > > > > > > > something later catches an incorrect checksum.
> > > > > > > > > > > >
> > > > > > > > > > >
> > > > > > > > > > > I don't want to replace the malloc library which handles large numbets of allocations.
> > > > > > > > > >
> > > > > > > > > > I'm confused. The normal malloc library is not involved with current
> > > > > > > > > > image loading, it's direct to memory (with some attempts at sanity
> > > > > > > > > > checking by lmb). Are you proposing a different allocator with
> > > > > > > > > > malloc/free like behavior? If so, please outline how it will determine
> > > > > > > > > > pool size, and how we'll use it to load thing to memory.
> > > > > > > > >
> > > > > > > > > All memory below the stack needs to be managed. Malloc uses a small memory area (a few MiB) above the stack.
> > > > > > > >
> > > > > > > > That's a rather huge change for how U-Boot works.
> > > > > > > >
> > > > > > > > > > > Closing the eyes when the user loads multiple files does not solve the fragmentation problem.
> > > > > > > > > >
> > > > > > > > > > Yes. I'm only noting that today we just ignore the problem and sometimes
> > > > > > > > > > catch it via checksums.
> > > > > > > > > >
> > > > > > > > > > > Fragmentation only happens if we have many concurrent allocations. In EFI we are allocating top down. The number of concurrent allocations is low. Typically a few dozen at most. After terminating an application these should be freed again.
> > > > > > > > > >
> > > > > > > > > > OK, so are you saying that we would no longer be loading _to_ a location
> > > > > > > > > > in memory and instead just be saying "load this thing" and picking where
> > > > > > > > > > dynamically?
> > > > > > > > >
> > > > > > > > > Both preassigned and allocator assigned adresses are compatible with memory management.
> > > > > > > > >
> > > > > > > > > Architectures and binaries have different requirements. On riscv64 you can load Linux kernel, initrd, fdt anywhere. We don't need predefined addresses there. Other architectures have restrictions.
> > > > > > > >
> > > > > > > > Yes, 64 bit architecture tend to only have alignment requirements while
> > > > > > > > 32bit architectures have both alignment requirements and some memory
> > > > > > > > window requirement. Whatever we implement here needs to handle both
> > > > > > > > cases.
> > > > > > > >
> > > > > > > > > > > When loading a file from a file system we know the filesize beforehand. So allocation is trivial.
> > > > > > > > > > >
> > > > > > > > > > > The loady command currently does not use the offered size information but could do so.
> > > > > > > > > >
> > > > > > > > > > We should be using that information to make sure we don't overwrite
> > > > > > > > > > U-Boot itself, but I don't recall how exactly we handle it today
> > > > > > > > > > off-hand.
> > > > > > > > >
> > > > > > > > > If the user issues multiple load commands, he can overwrite previous files.
> > > > > > > >
> > > > > > > > Then it sounds like we lost one benefit of all of this overhead.
> > > > > > > >
> > > > > > > > > During boot command execution I guess the different allocations respect each other.
> > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > > TFTP is problematic because it does not transfer the filesize. We would probably try to allocate a large chunk of memory and then downsize the allocation after reading the whole file.
> > > > > > > > > >
> > > > > > > > > > Reading from non-filesystem flash also has this problem, but we at least
> > > > > > > > > > specify the amount to read too. But yes, it gets back to what I was
> > > > > > > > > > asking about on how you're proposing to handle network load cases.
> > > > > > > > > >
> > > > > > > > >
> > > > > > > > > It depends on the protocol. Http conveys the size before the data. Tftp does not.
> > > > > > > > >
> > > > > > > > > If you don't know the size, you must preallocate a big chunk, check that the download does not exceed it, and downsize the allocation afterwards. This is not a new problem but exists already with current lmb usage.
> > > > > > > >
> > > > > > > > Yes, and what I'm trying to find out is if what you're suggesting would
> > > > > > > > do anything about it, since previous statements you made implied to me
> > > > > > > > that we would prevent it.
> > > > > > > >
> > > > > > > > To me, at this point it sounds like what we need is more like persistent
> > > > > > > > memory blocks and a hook that can be called in to for both "give me all
> > > > > > > > known memory blocks" and "add this memory block to the list", so that
> > > > > > > > EFI can do whatever it needs to do upon starting an application and then
> > > > > > > > upon return to U-Boot. Both malloc/free allocations and "load this blob
> > > > > > > > to memory from whatever" allocations would call the appropriate hook for
> > > > > > > > tracking.
> > > > > > >
> > > > > > > In my mind the solution to this entire problem is fairly minor changes
> > > > > > > to how memory is allocated and only for EFI.
> > > > > > >
> > > > > > > I tried to map out what that would look like and we have IMO got lost
> > > > > > > in the weeds a bit.
> > > > > > >
> > > > > > > I am not trying to solve the problem of the 'load' command doing an
> > > > > > > allocation and throwing it away. To be that is WAI, at least until we
> > > > > > > come up with another type of command. This is one of the reasons for
> > > > > > > standard boot, allowing a more cohesive approach to booting.
> > > > > > >
> > > > > > > I will think about this some more...
> > > > > >
> > > > > > OK, but please keep in mind that lmb not being at all persistent is a
> > > > > > problem for everyone, not just EFI. That really needs to be addressed,
> > > > > > maybe with some flags for dis-allowing overwrites to the area. For
> > > > > > example, the apple-m1 code to use lmb to find locations for the
> > > > > > kernel/etc can be written to more than once (allocate the address, then
> > > > > > write to it to start with, even) but the range that covers U-Boot itself
> > > > > > (malloc pool and so forth) need to be stopped.
> > > > > >
> > > > >
> > > > > The range managed by the EFI sub-system extends over all of RAM
> > > > > including the addresses used by load commands. Hence, "the range that
> > > > > covers U-Boot itself (malloc pool and so forth)" includes all memory.
> > > > >
> > > > > Maybe we could pre-allocate a memory area for file loading. This just
> > > > > requires to define a device-specific high file memory address above
> > > > > which no file loads will be allowed and below which EFI, bootm, etc will
> > > > > not be allowed to allocate memory for further uses.
> > > >
> > > > We're not just / only an EFI runtime. When we load something persistent
> > > > to memory outside of EFI, we need to note that, so that when EFI_LOADER
> > > > beings work it can see this and do whatever it needs to do. And if/when
> > > > EFI_LOADER returns control back to U-Boot itself, it needs to update
> > > > that list with any new allocations that were done and mark them
> > > > appropriately.
> > >
> > > The idea of EFI loader returning control is misleading.
> > >
> > > You can load an EFI driver binary which provides a device to U-Boot.
> > > Whenever U-Boot accesses that driver allocations might occur.
> >
> > We cannot, today, load an EFI driver and then make use of it outside of
> > the context of having bootefi'd something, yes?
>
> We already can use iPXE to access iSCSI which provides a U-Boot block
> device. But it is an app. A similar binary could be compiled as EFI
> driver return to U-Boot and provide a block device to be directly
> addressed by the user via the CLI.
To be clear:
Today we can, or cannot, use iPXE to access an iSCSI device and then
"ext4load" off of it?
Today we can, or cannot, use iPXE to access an iSCSI device and then
"bootefi" another EFI application such as grub which can then load off
if it?
Today we can, or cannot, use iPXE to access an iSCSI device and have it
place a Linux Kernel Image at $loadaddr which we can then "md" or any
other standard U-Boot command with?
My assumption is that we can do the last two, today, and the first one
is something that yes, the specification allows, but not something we
have today and requires some discussion on how much we want to allow
modules-by-another-name to work. I very much don't want to start that
discussion right now either.
> > Or is this about that separate runtime area from another thread?
> >
> > But even then, if something is making an allocation it needs to inform
> > the rest of the world.
>
> We should have one memory management. When EFI's AllocatPages() or
> AllocatePool() is called it should use the same backend as any other
> allocation (except for the small region managed by malloc()).
>
> >
> > > I thought that we already agreed to have only one memory management and
> > > get rid of the LMB/EFI memory management duplication.
> >
> > We've agreed that today, CONFIG_LMB does not fit the use cases that we
> > have in a modern system, for anyones usage. It has been suggested and I
> > know it's somewhere on Ilias' TODO list, to try and take what we have
> > in the EFI_LOADER & co area and make it usable and useful for all
> > contexts. We also have what in some ways feels to me to be the inverse
> > of that, which is Simon's proposal.
>
> LMB and efi/loader/efi_memory.c have a big overlap in functionality.
>
> But we use LMB in a unfavorable way discarding all allocations.
Yes, we all agree that what we do with CONFIG_LMB today is not good nor
what we want long term. In some ways that feels like the single point of
agreement.
> > We have not agreed that the EFI model and restrictions are what
> > everything will now be using. What's on Ilias' TODD list might bring us
> > in that direction, and I want to see what that looks like in practice.
> > But it's not agreed to as the end goal here.
> >
> > And to be clear, I also would like to see what Simon's proposed, in
> > practice.
Because this part here. We've probably all talked this out a lot.
Without an implementation to examine we're probably going to continue to
not agree and possibly unintentionally misunderstand each other.
--
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20231229/74e9c4a1/attachment.sig>
More information about the U-Boot
mailing list